{"model": "sentence-transformers/all-MiniLM-L6-v2", "titles": ["Do you know anything about me?", "Which software did Marc use for data processing?", "Where can I find Marc's article about the chlorine atom in drug discovery?", "What is Marc's main conclusion in \"The New Architects\"?", "Why is molecular replacement relevant to ligand validation?", "How much data went into designing the Top96 screen?", "What do you think about the stock market right now?", "What organism name appears on the ShpA entry?", "How does CNTF relate to OSM in Marc's structure\u2013function comparison?", "What is Marc's design system?", "What is Marc Deller's D.Phil. thesis about?", "What is smiles_grid?", "Does Marc program in R?", "Are the WRN inhibitors in the patent covalent inhibitors?", "How different was the final OSM structure from the initial homology model?", "What domains do the cyanobacterial endopeptidases contain?", "What does Marc charge for consulting?", "What is Marc's 2015 paper on protein stability?", "How many crystals did the team have to screen to get the HIV-1 Env trimer structure?", "Where was this CDK2 research carried out?", "What structural features make IsPETase better than related cutinases?", "Which polymer was most toxic to the lungs in the studies Marc reviewed?", "What are TopPDBLX's known limitations?", "What was Marc Deller's role in the UMR crystal harvesting project?", "How does AnalyzeBoltz derive a pIC50?", "What are Di-Gembodies?", "Which gut bacteria are associated with PET degradation in Marc's post?", "What has the AP33 work enabled for vaccines?", "What are the final refinement statistics for the OSM P212121 structure?", "Why did Marc write a whole blog post about the \u00e5ngstr\u00f6m?", "What's the capital of France?", "What combination therapies are covered in the claims?", "What settings does ChatMol expose?", "What were the stated goals of the Incyte program?", "What is an RCK domain?", "How tall is Marc Deller?", "What is the hepatitis C AP33 structure?", "What microplastic exposure figures appear in the 79th organ post?", "What vibe-coding apps does Marc Deller's r\u00e9sum\u00e9 mention?", "Where is the polyketide synthase didomain post?", "How much did structural genomics contribute to the PDB?", "What are the two domains of BT2081?", "Where can I see Marc Deller's protein structures?", "What is the smallest crystal usable for X-ray structure determination?", "What is Marc's argument in the \"Hidden Chemistry\" post?", "What did the Mre11 structure reveal about substrate discrimination?", "How many scientists has Marc Deller mentored?", "How does Marc explain the \"molecular bumper\" and \"molecular glue\" idea?", "Who supervised Marc Deller's D.Phil.?", "What does Marc report about sub-micron particles and cancer cell migration?", "What is the Chat page on marcdeller.com for?", "Where is Marc Deller based?", "Who is PocketTide for?", "What is Marc Deller's post on oral protein and peptide drug delivery about?", "Tell me about Marc Deller's work on potassium transporters at Scripps.", "Which paper covers Marc Deller's work on FGFR inhibitors?", "What are the TrkA structures Marc deposited?", "What are the biological actions of oncostatin M (OSM)?", "What about the E655G escape mutation?", "What does Marc's post on flow cytometry and microplastic detection argue?", "What role did Marc play in the polyketide synthase work?", "Which harvesting technologies does the UMR post cover?", "Why does inhibiting WRN kill only some cancers?", "What did the co-crystal structure in the patent show?", "How long does it take to train staff for such a facility?", "What does the Projects page on marcdeller.com cover?", "How did PGT 135 get through the HIV glycan shield?", "Which paper does Marc identify as the century's most-cited, and why does it matter?", "Why does GOBSMACKED superpose on pocket C\u03b1 atoms rather than the whole chain?", "Why are dark kinases so hard to study?", "What was destructive about previous microplastic detection methods?", "Has Marc published on crystal structure validation?", "What can you type at the chatRDKit prompt?", "What does Marc do now?", "Where is Marc's Anthropic Academy post?", "Summarise what the three HCV and HIV structural papers from Marc's Scripps period have in ", "What is the PDB code for the structure in the 2022 FGFR paper?", "What did Marc Deller do at Scripps?", "Can you explain the povorcitinib paper in plain English?", "What did Marc conclude about the character of the OSM receptor-binding interfaces?", "What does Marc Deller's post on NOS linkages as allosteric switches cover?", "Which work does Marc cite in the TELSAM-DARPin post?", "What is unusual about oncostatin M's receptor usage?", "Does Marc see AI biodesign purely as a risk?", "What did the 2000 \"Cell surface receptors\" review conclude?", "What was Marc Deller's LinkedIn profile?", "What did the review say about fibroblast growth factor receptor complexes?", "Which JCSG entries carry the Pfam DUF1285 fold?", "How does Site Doctor get Lighthouse scores?", "What uses beyond small-molecule docking does Marc list for DRGSCROLL?", "How does Marc characterise ProteinMPNN and LigandMPNN?", "How much PET is produced annually, according to Marc?", "Who are the founders of Elora Therapeutics?", "What is Marc's worked example of ALPHABETTI failing?", "What kinds of problems do Marc's vibe-coded tools solve?", "Which generative AI courses has Marc taken?", "What HIV antibody structure is PDB 5CD5?", "Explain how neural networks work.", "Which paper accompanies the 5UFV structure?", "What does \"From Waste to Wonder: How Bacteria Transform Plastic into Life-Saving Medicine\"", "Where is the Exig_1061 post on Marc's website?", "What robotic approaches does the harvesting review cover?", "What was the headline structural result of the thesis?", "Where is the source code for Marc's web apps?", "Who are the inventors listed on the mutant-CALR antibody patent?", "What is CotA-Laccase and why is it included?", "What results does Marc highlight for the barcode-free screening platform?", "Where was Marc working when the BtDyP structure was determined?", "How does the UMR perform hyperquenching?", "How many emails does Marc get a day?", "Which of Marc's recent structures is the highest resolution?", "What is the C-terminal domain of Jann_2411?", "What is PDB entry 3CM1?", "What is the RSRZ value?", "What problems does run_boltz.zsh prevent?", "Which paper reported the Bifidobacterium phosphoesterase structure?", "What is Marc's conclusion about vibe coding after building the game?", "Why is AI protein engineering needed for bio-hybrid sensors, according to Marc?", "What is loaded into the chatRDKit session namespace?", "What fold does the BtDyP peroxidase have?", "Where is Marc's post on the bacterial PH domain protein?", "When did the DRVIA7 antibody lineage begin?", "What limitations does Marc flag for self-driving protein engineering labs?", "What was Marc Deller's role on the crystal harvesting review?", "What does Marc say about TurboPETase and LCC variants?", "What mechanistic conclusion did the MtPMO3* paper reach?", "What is the TrkA gating ring paper?", "Why did the team switch from an imidazole core to an indole core?", "How did the new cleaved trimer structure compare with the uncleaved trimer structures that", "What were the aims of Marc Deller's thesis?", "What is the ABATE domain?", "Which are Marc Deller's ten most-cited publications?", "Where can I find Elora Therapeutics online?", "What do Chapters 3 and 4 of Marc Deller's thesis cover?", "How do pixel-array detectors change crystal harvesting?", "How were the antibodies discovered?", "What organism is listed for the Exig_2160 structure?", "What was Marc's contribution to the FGFR2/3 work?", "Which prior patents does the invention distinguish itself from?", "Why does Marc's review say protein stability is a prerequisite for crystallization?", "Which recent organ discoveries does Marc use to set up the 79th organ idea?", "How was the quality of the native P212121 data?", "What holds the TrkA octamer together?", "How did Marc build MicroPlastic Blaster?", "Which journal published the DyP peroxidase structures?", "What was unusual about helix A in the oncostatin M structure?", "What is KPN03535?", "What resolutions were achieved in the purification-free cryo-EM work Marc describes?", "What does Marc's post \"Microplastics and Health: Alarming New Research Reveals Widespread ", "Does the patent cover a treatment method, not just the enzymes?", "Which mouse model does Marc discuss for MASLD and microplastics?", "Where was the FGFR2/3 inhibitor work published?", "What are BoltzMaker's pipeline stages?", "What is the lead compound in the 2022 FGFR paper and how potent is it?", "What is leukemia inhibitory factor (LIF) and what does it do?", "What is Click2Drug, as described by Marc Deller?", "Where and when was the 10E8 lipid paper published?", "What warnings does Marc give about running RFdiffusion on an M1 Mac?", "How was povorcitinib synthesised?", "What made INCB159020 orally bioavailable when earlier compounds were not?", "Why does Marc think facilitated dissociation matters therapeutically?", "What is SEAL, as covered in Marc Deller's post on self-adapting language models?", "How many works are on Marc's ORCID record?", "What is Marc's ORCID?", "What are the two steps of PET degradation in Marc's account?", "What is Marc's overall view of compositional protein design?", "How do the patented antibodies compare with earlier anti-mutCALR antibodies?", "Which residues does the inhibitor contact in FGFR3 according to the modeling?", "Are Marc Deller's Incyte drug discovery papers on his Scholar profile?", "What did the NANOG alanine mutants show?", "What is in the lithium chloride / PEG 6000 block of the JCSG Top96?", "What are the five self-descriptions Marc gives on his About page?", "What cancers does the WRN patent say the compounds could treat?", "Why does AlphaFraud skip some entities?", "What are LOCKTACs and how do they differ from PROTACs?", "What does \"dimer hinge\" mean in the TM1088 paper?", "What did the SECMALS analysis show?", "Tell me about the TrkA potassium channel gating ring structure.", "What did the study contribute methodologically?", "What was the first structure Marc solved?", "What percentage of JCSG structures crystallize in the 96 Top96 conditions?", "What escape mutations does the AP33 structure explain?", "What is Elora's intellectual property strategy?", "What applications does Marc highlight for IDR binders?", "Why does Marc keep returning to structure validation?", "Is Marc Deller on LinkedIn?", "How were the conditions arranged on the JCSG Top96 plate?", "What role was proposed for the DUF1285 family?", "How does NMR complement crystallography for assessing stability?", "What is Horseradish Peroxidase used for in the invention?", "What is povorcitinib being developed for clinically?", "Which live statistics does mdeller.com show?", "Which post-translational modifications are the most common?", "Which journals does Marc Deller publish in?", "How many antibody clones does the patent describe?", "What is YP_910028.1 from Bifidobacterium adolescentis?", "Can you write some code for me?", "Who was donor DRVI01?", "What can you help me with?", "What did isoelectric focusing reveal about LIFR?", "What is the overall fold of oncostatin M as described in Marc Deller's thesis?", "Did sodium butyrate help LIFR expression in CHO cells?", "Where could the JCSG Top96 conditions be obtained?", "Has povorcitinib shown clinical results?", "What tools does Marc use to vibe-code?", "I heard Marc worked on a really tight patent deadline once. What happened?", "What practical expression advice does Marc give for crystallography targets?", "What is TM1089 and how does it relate to TM1088?", "What is Marc Deller's ORCID identifier?", "How many citations does Marc Deller have?", "How are the conditions ordered on the JCSG Top96 plate?", "How can I contact Marc Deller?", "Where is the lipid-binding site on 10E8?", "Why are ligands especially hard to model correctly?", "Is WRN inhibition the same as classic DNA-damaging chemotherapy?", "What were the refinement statistics for the PGT 135 structures?", "Does the inhibitor make a direct contact with the mutant aspartate?", "How did Marc Deller contribute to the Dhaf4260 structure?", "What makes up the hydrophobic core of oncostatin M?", "What governance does Marc call for on AI protein design?", "Who solved the crystal structure in the FGFR2/3 paper?", "How is BcYkfC built?", "How was the oncostatin M protein produced for crystallography?", "What software does Marc use?", "Which publications does Marc list as his top ones?", "How did the KTH team confirm that PET was fully broken down?", "What solvent accessibility did the functionally important OSM residues have?", "Which university did Marc go to for his first degree?", "How do the two enzymes differ in size and sequence identity?", "Did povorcitinib work in an animal model of arthritis?", "Where can I find Marc's write-up of PDB 2GVK?", "Which conditions in the JCSG Top96 produced 12 PDB structures each?", "Which cell lines and media did Marc use for eukaryotic expression according to the appendi", "What does Marc say about cutinases such as Cut190 and LCC?", "What is PDB entry 10PJ?", "How do you run chatPDB locally?", "Is there a commercial angle to automated harvesting?", "What did the review say about iron metabolism and the transferrin receptor?", "What programs did Marc use to analyse and illustrate the final model?", "Who funded the ligand validation chapter?", "What is the Nanog homeodomain structure Marc lists?", "Why is the smoothness of ButtFold's animation caveated?", "What is Marc Deller's 2000 review \"Cell surface receptors\" about?", "What software did Marc Deller use for the TM1088 structures?", "How does AP33 grip the E2 peptide?", "What did N-terminal sequencing tell Marc Deller about the purified oncostatin M?", "How well cited is Marc's oncostatin M paper?", "What does the Artificial Intelligence page on marcdeller.com cover?", "How selective is povorcitinib, and why does sparing JAK2 matter?", "Does Marc have project management training?", "What does the Real Space Correlation Coefficient tell you?", "Do gut bacteria have membrane-attack complex proteins?", "What limitations of PET-degrading enzymes does Marc identify in that post?", "Why does DeepCovalent say 90 % of drugs fail in clinical trials?", "What method and resolution were used for PDB 3NL9?", "What are the 10E8 structures about?", "Which of Marc's apps relate to Elora Therapeutics?", "What are the 6 C's?", "What does the patent mean by a biparatopic antibody in this context?", "Over what time window were the PDB counts in the JCSG Top96 formulation collected?", "Over what range of protein sizes and resolutions does MICA work?", "Which vibe-coded projects are highlighted on the marcdeller.com home page?", "What are PETase and MHETase?", "Where can the FGFR2 structure be downloaded?", "How does keeping patients in mind change Marc's decisions?", "What happens to normal cells under covalent WRN inhibition?", "What is intramolecular hydrogen bonding and why did it matter here?", "Can you help me with my tax return?", "What partnership models does DeepCovalent offer?", "What residue numbering convention is used for the OSM structure?", "Where can I find Marc Deller's other profiles and links?", "How many structures has Marc solved?", "What software does the chapter list for ligand validation?", "What function was proposed for LP2179?", "Which earlier JCSG screens preceded the Top96?", "What kinds of plastics is the invention designed to degrade?", "How does gp130 compare structurally with other haemopoietin receptors in Marc's thesis?", "Who supplied the biotinylated ligands and transfected cell lines used in the assays?", "How did the variants affect oxygen consumption?", "How did the team confirm the effect was really due to JAK1 and not JAK2?", "How does the patent propose to deliver these enzymes into the body?", "What does Marc Deller say about cryo-electron tomography?", "What obstacles to recursive self-improvement does Marc identify?", "What is a scaffold hop, and how was one used in this paper?", "How was INCB159020 synthesized?", "What ligands were found in the lipid site across the different structures?", "Who did Marc Deller work with on the polyketide synthase study?", "What features did Marc identify as unique to OSM among the gp130 cytokines?", "Where can I see Marc's PDB structures?", "What explains the lungfish's enormous genome, according to Marc?", "Which JCSG structural genomics papers appear on Marc's Scholar profile?", "What models does ALPHABETTI run and where?", "Which projects are listed as Marc's top projects?", "Why do Marc's ORCID and Google Scholar counts differ?", "Can you explain this paper simply for a non-scientist?", "How similar are OSM and LIF?", "What distinguishes the open and closed SpoIIAA-like conformations?", "Tell me about PDB entry 10PI.", "How do these Formula (I) compounds work mechanistically?", "What is Protein Hunter, as described on Marc Deller's blog?", "How is cif2plip installed and run?", "Why does generate_yaml.py exist?", "What is the DOI of Marc's HIV-1 envelope trimer paper?", "What campaigns are listed on the BoltzMaker site?", "How would a patient be selected for one of these WRN inhibitors?", "How does mutant CALR drive disease?", "How many hours a week does Marc work?", "Which students has Marc supervised by name?", "What did Marc Deller write about generative AI for protein pocket design?", "Did Marc take a course on communicating with generative AI?", "What is \"molecular Velcro\" in Marc Deller's blog?", "How does the (4-pyrazolyl)-2-aminopyrimidine scaffold achieve selectivity?", "What is INCA033989?", "I'm bored, entertain me.", "Who wrote the protein stability review with Marc?", "Why does Marc consider clinical trials a bottleneck worth fixing with AI?", "What are the PDB codes from the 2015 TrkA gating ring paper?", "Where can I read Marc's PXDesign post and is the tool available?", "Which new receptor domain families did the review highlight?", "How good is the stereochemistry of Marc's OSM models?", "What cooling temperatures are used for protein crystals during data collection?", "Which membrane protein structure does Marc single out?", "Why is AP33 more sensitive to Asn415 mutations than HCV1?", "What is unusual about the A-B loop of OSM?", "What environmental specifications does a cryo-EM room need?", "What is quantum computing?", "Explain Elora Therapeutics in simple terms.", "Where can I read Marc's overview of PETase as nature's solution?", "Why does Marc think curiosity matters in science?", "What analytical methods did Marc Deller use to characterise these proteins?", "Which app would help me plan a crystallisation screen?", "What is Marc's outlook for JAK2-V617F disease?", "What peptides were used as the antigen in the crystallography?", "What is Marc's passport number?", "What is the take-home message of Marc's protein stability paper?", "What definitions and detection sites does Marc give in \"The Invisible Invasion\"?", "Where can I read Marc's post on going from protein sequence to a validated structure with ", "What is the 2015 \"trust, but verify\" paper?", "Why did the review stress caution about crystal lattice contacts?", "What is Marc's post \"From Bicyclic to Tricyclic\" about?", "What limitation of classical force fields does Marc identify?", "What is the FGFR2/3 resistance problem Marc's work addresses?", "What does \"Decoding the Molecular Machinery Behind Plastic Degradation\" explain?", "Explain the JCSG Top96 screen in simple terms.", "When and where was the PKS extended conformation paper published?", "How does cif2plip generate its CONECT records?", "Which gp130 residues are structurally equivalent to the key binding determinants of EpoR?", "Did the TyrA work change how peroxidases are classified?", "What are FlexAppeal's three tabs?", "Which laboratories collaborated on the eubacterial Mre11 structure?", "What methods and crystallization papers has Marc Deller written?", "How does expression temperature affect protein stability?", "Why is Nanog an important protein?", "What is the 2025 bicyclic WRN patent about?", "What date was Marc's viva?", "What did the DUF1285 structures show?", "Which residues form the PETase catalytic triad and oxyanion hole?", "Who benefits from a screen like Top96?", "What happens to the PET after Elora's enzymes break it down?", "How many HIV-related structures are in Marc Deller's list?", "What receptor-binding sites did the oncostatin M structure map?", "Which commands and tools does Marc use in his AlphaFold Mac install?", "How did the review describe the diversity of cytokine\u2013receptor complexes involving Ig-like", "Does INCB159020 prefer the GTP-bound or GDP-bound state of KRAS?", "Which design and validation methods does Marc describe for the facilitated dissociation wo", "What advantages does automated crystal harvesting offer over manual harvesting?", "Summarise the argument of Marc's courage post.", "What is BT_3984 and why was its structure a first?", "What is the scope of the review?", "What resolution was the PGT 135\u2013gp120 structure and what are the PDB codes?", "Which structural examples of chlorine's potency effect does Marc give?", "Tell me something fun about Marc.", "Which JCSG paper named a new protein fold in 2010?", "What are Marc's hobbies?", "Tell me about the dye-decolorizing peroxidase structure Marc worked on.", "How does Marc describe validation practice for antibody prediction, and where is the post?", "What is NGO1945?", "How is generate_yaml.py run?", "What is wp-content-inventory?", "What is HinfPDH?", "Did Marc Deller work on robotics at Scripps?", "What's the meaning of life?", "Where did the GCSF\u2013GCSFR coordinates used in the thesis come from?", "What was the problem with the first-generation FGFR inhibitors?", "What was Marc's role at Incyte?", "What is model bias and why does it matter for ligands?", "What does the BVU2987 structure suggest about function?", "What are the refinement statistics for the I222 OSM structure?", "What is an omit map?", "You're not very helpful.", "What output formats can chatRDKit produce?", "Where can I find Marc Deller's full publication list and structures?", "What does Marc Deller's post \"When Protein Design Becomes a Security Problem\" argue?", "Explain the PKS extended conformation paper for a non-scientist.", "What topics does Marc say his publications cover?", "What is DEBS?", "Why does Marc think AXIS matters for drug discovery?", "What was Marc Deller's role on the 10E8 paper?", "What is the ADE conveyor belt?", "Has Marc ever been to Japan?", "What is PDB entry 2ICH?", "Who is Marc's favourite leader?", "Who else was on the author list of the 2013 Science HIV-1 Env trimer structure?", "What was Marc Deller's role in the CDK2 inhibitor discovery?", "Where can I read Marc's curiosity post?", "What are smiles2grid's requirements and how is labelling handled?", "Why are peptide ligands a useful test case for validation?", "What lesson did the review draw from the TNF and Epo systems together?", "What is FlexAppeal's example run?", "What is the JCSG Top96 condition for well C05?", "How do microplastics get inside cells, according to Marc Deller?", "What does ForgeFold expect alongside it?", "Who else at Incyte worked on the structural biology for povorcitinib?", "What did the oncostatin M paper say about the similarity between site 2 and site 3?", "What were the R-factors at the start of refinement?", "What does compare-sse do?", "Are there bacterial Suppressor of Fused proteins?", "What was Marc Deller's postdoc?", "What is the URL of Marc Deller's main website?", "What antibody was isolated from the Chinese donor?", "What was first about the Thermotoga Mre11 structure?", "What are Elora's two paths to revenue?", "Is Marc Deller British or American?", "Which earlier tools did BoltzMaker consolidate?", "What did the 2009 Structure paper on murein peptide specificity report?", "What is unusual about the zinc coordination in Q9HJ63_THEAC?", "What is Marc Deller's 2016 Acta Crystallographica F review about?", "What did the 10E8 structure show about the MPER's orientation?", "How did the paper propose that oncostatin M docks onto gp130?", "What example campaigns ship with BoltzMaker?", "Which oral biologics have reached patients, according to Marc?", "How was Marc funded when he wrote the review?", "Where was the automated crystal harvesting review published?", "What did Marc do at Scripps?", "Which compounds were advanced to pharmacokinetic studies?", "What metals were found in the YP_910028.1 structure?", "Where is Marc from, and where did he study?", "Why can't HIV simply escape the supersite antibodies?", "How did Marc Deller decide where to truncate oncostatin M?", "Which collaboration does Marc cite on AI-driven inverse folding?", "Why did Marc build CODSWALLOP?", "Are there approved CDK2 inhibitors, and how does this fit in?", "What was Marc doing when 5C7K was released?", "What is the most common space group among Marc Deller's structures?", "Where does Marc go to the gym?", "What does Marc report about polyethylene microplastics and liver fibrosis?", "Where can I read \"From Bicyclic to Tricyclic\"?", "Which drug targets has Marc worked on?", "What three outputs does AnalyzeBoltz produce?", "What is PDB entry 3H0N?", "What benefits does Marc list for agentic bioinformatics systems?", "How will Elora use the initial $275K raise?", "Who published the bacterial Sufu-like structure?", "Which references does Marc cite in the 3Dmol.js post, and where is it?", "What did the zinc acetate complex suggest about XcTcmJ's function?", "What is the GLP-1 patent Marc Deller is named on?", "What is Marc's own conclusion about self-improving AI?", "What did the PGT122 complex show?", "Summarize the paper in one sentence.", "What tabs does the TopPDBLX site have?", "What is BoltzMaker built for?", "Why does run_boltz.zsh clean the input directory?", "Hello", "Which posts are featured on Marc Deller's home page?", "Why does Marc call photoacoustic imaging label-free, and why does that matter?", "Did PGT 135 change the shape of gp120 when it bound?", "Why is hexagonal ice a problem for protein crystals?", "What were the three MAD wavelengths?", "Did the crystal structure agree with the electron microscopy work?", "What key paper did Marc Deller publish alongside this poster?", "Which validation tools does that 2017 chapter recommend?", "Does Marc write about cryo-EM methods on his site?", "Did Marc solve structures from Thermotoga maritima?", "What is the WSXWS motif and why does it matter?", "What is the Fe-His-Asp triad in TyrA?", "Which paper describes the Thermotoga acetyl esterase?", "Where is Marc's \"From Tool to Teammates\" post?", "Why does TopPDBLX say to read its numbers as popularity rather than efficacy?", "How does Marc link binding dynamics to signalling bias at GLP-1R?", "How is the AlphaFraud site organised?", "Who were the senior authors on the paper?", "Which Bacteroides thetaiotaomicron structures did Marc contribute to?", "Which drug-target structures from Marc's industry work are listed?", "How would you describe Marc's leadership style?", "Why did the UMR's small tape opening matter?", "What is PDB entry 3O0F?", "How is HinfPDH organised?", "What is PDB entry 6C9U?", "Where does Marc live now?", "What does Marc Deller's post on cloud-based drug discovery argue?", "What was the paper's main conclusion for vaccine design?", "What's the weather like in Wilmington today?", "What did Marc's review say about thermophilic proteins?", "Why does Marc think side-chain flexibility matters so much in lead optimisation?", "What device-based delivery does Marc highlight?", "What taxonomy of persuasion does Marc describe in \"Call Me A Jerk\"?", "How many automated research facilities has Marc Deller built?", "Why are cytokines medically important?", "What is Elora Therapeutics building, according to Marc's microplastics page?", "Who else did Marc work with on the FGFR paper?", "What is the TrkA gating ring paper?", "Which of Marc Deller's PDB entries come from his time at Incyte?", "How much does Marc earn from his patents?", "What is Site Doctor?", "What is Marc's philosophy on mentorship?", "Did the swapped-helix dimer disturb the active site of YP_001813558.1?", "What did electron microscopy add to the PGT 135 story?", "How was the ambiguity between I222 and I212121 resolved?", "What is PDB entry 3F1Z?", "What is ChatMol?", "What is Marc's broader take on MICA?", "How many citations does Marc Deller have in total?", "Tell me about the oncostatin M structure.", "Where did the mercury atoms bind on OSM?", "What is the Top96 Cryoscreen?", "Where is the TrkA gating ring post on marcdeller.com?", "How similar is NGO1391 to human Sufu?", "What does Marc say about the democratization of computational drug discovery?", "Which journal published the Nanog homeodomain structure?", "Was selenomethionine incorporation into oncostatin M complete?", "In plain terms, what is Marc Deller's \"trust, but verify\" paper saying?", "Is the joke in ButtFold's name deliberate?", "Tell me about the deadline story behind Marc's crystallography for this patent.", "What did AlphaFraud find?", "Which OSM truncation looked more promising after cloning, and why?", "What is the activity-stability trade-off Marc describes?", "What is Marc Deller's 2000 Structure paper on oncostatin M about?", "What did Marc Deller do at Incyte as a Research Fellow?", "Which of Marc's apps involve music or games?", "What sections can visitors navigate to from Marc's home page?", "How many patents does Marc credit to his creative approach?", "What is Marc's star sign?", "Which hematopoietic four-helix bundle cytokine complexes were new in the review period?", "What was Marc's role, and where is the post?", "Which company funded and conducted this work?", "Why couldn't the team just remove the problem groups one at a time?", "What are the \"6 C's\" in Marc Deller's r\u00e9sum\u00e9?", "What function was proposed for the DUF1185 family?", "What does the ShpA fold look like?", "What did the oncostatin M structure reveal about its receptor-binding sites?", "What cell-based potency did the antibodies show?", "What is MicroPlastic Blaster?", "Which cysteine does the WRN inhibitor bind, according to the patent's structure?", "Which antibody formats does the patent cover?", "What software was used to build and refine the Env trimer model?", "Can the antibody be used diagnostically?", "How many crystals does the JCSG harvest per structure?", "What is TM0077 and what did Marc Deller help show about it?", "How does the PXDesign pipeline work in Marc's description?", "Which fields does Marc say his curiosity has taken him into?", "Is it always right to trim a ligand model back to what you can see?", "Which capabilities does Marc say ProtComposer unlocks for protein designers?", "What contactless harvesting methods are discussed?", "Which patents are listed in Marc Deller's Extended Portfolio?", "Tell me about Marc's post on LatentX.", "What is INCB161734 and what did Marc contribute?", "Which DNA polymerase worked for amplifying the OSM truncations, and why?", "What is FlexAppeal?", "What is the technical architecture behind MICA as Marc explains it?", "Who is behind LatentX and how is it funded, according to Marc?", "Which institution was Marc at for the Mre11 work?", "What music was playing when Marc solved the WRN structure?", "Write an email for me.", "What achievement does Marc attach to Courage in his 6C's?", "Explain the significance of Marc Deller's HIV envelope trimer work to a non-scientist.", "How did Marc Deller express oncostatin M?", "What is BG505 SOSIP.664 gp140?", "How is Marc connected to the 2013 HIV Env trimer Science paper?", "What delivery strategy does Marc propose for getting PETase into lysosomes?", "Why was selectivity over CDK1 so important in this program?", "What power-ups does MicroPlastic Blaster have?", "How many cups of tea does Marc drink a day?", "Is there a CDK2 structure among Marc's depositions?", "What does Marc Deller write about multi-agent language models in bioinformatics?", "Does Marc talk about time pressure in structural biology?", "Which polyketide synthase paper involves Marc Deller?", "Which expression system worked best for LIFR, and how much protein did it give?", "What is PDB entry 3EQX?", "What is PDB entry 3JXO?", "What did SEC\u2013SAXS reveal about the 1B2/module 2 complex in solution?", "Write me a Python function that sorts a list.", "What is the Universal Mounting Robot project about?", "What are the four reactions in the catalytic cycle of DEBS module 2?", "What resolution was the 10E8-T117v2 complex structure?", "What two OSM truncation mutants did Marc Deller engineer?", "How does Marc Deller's thesis distinguish a cytokine from a growth factor?", "What did Marc Deller's \"druggability\" workflows at Incyte achieve?", "What are self-encoded libraries, as covered on Marc's blog?", "Which enzymes does the PANTS catalogue name?", "What is Marc Deller's 2025 patent about?", "What is the strongest experimental evidence Marc identifies in that review?", "What does Marc Deller do at DeepCovalent?", "Can you explain the KRAS G12D paper in simple terms?", "What is the \"up-up-down-down\" topology mentioned in the thesis?", "How does Marc describe the PETase and MHETase partnership?", "What software and tools did Marc use for the oncostatin M structure determination?", "Does polymer identity matter in the cancer data, according to Marc?", "What is INCB159020 and which structure shows how it binds?", "Why might the TM1088 structure be relevant to drug discovery?", "Write me a short story.", "Has Marc ever turned down a co-authorship?", "What clinical hurdles remain for photoacoustic microplastic imaging?", "How was the 1B2 antibody discovered?", "Has Marc built dashboards?", "Did the review discuss NCAM?", "Which Top96 conditions need no additional cryoprotectant?", "Explain the JCSG Top96 formulation in plain terms.", "Where can I read the INCB159020 structure post?", "Who are Marc's parents?", "Where was Marc based for the TM0077 structure?", "How common are MSI-H cancers, and does p53 status matter?", "What was seen in the kanadaptin FHA domain crystal?", "What is the 3C protease that Marc Deller used, and did he make it himself?", "How does Marc illustrate the shift from static structures to protein dynamics?", "Which human tissues has flow cytometry been applied to for microplastic detection?", "What headline numbers does Marc's home page use to summarise his career?", "What is Marc Deller's oncostatin M paper?", "What staining protocol does Marc describe for flow cytometry of microplastics?", "What is a gatekeeper mutation and why does it matter for FGFR drugs?", "What is Docling, according to Marc Deller?", "What are Marc's three biggest achievements?", "What is in well A01 of the JCSG Top96 screen?", "How tightly did PGT 135 bind gp120?", "When was the HIV-1 Env trimer paper published and where does it sit in Marc Deller's publi", "What was Marc Deller's role in the JCSG Top96 formulation?", "What did the DRVIA7 antibody structures show?", "How do antibodies reach through HIV's glycan shield?", "Who led the bacterial pleckstrin homology domain paper?", "Did Marc work at Pfizer?", "What are PepLites, and where is Marc's FragLites post published?", "What is PDB entry 3H50?", "Where was Marc working on the KRAS G12D programme?", "I have a question you can't answer from your training. What then?", "What is Marc's ORCID password?", "Which paper do the acetyl xylan esterase structures belong to?", "Which glycan turned out to be newly important for PGT122 neutralisation?", "How does Site Doctor avoid third-party dependencies?", "What did the growth hormone-receptor complex reveal about cytokine signalling?", "What did the Nanog structure show about DNA binding?", "What was concluded about T74?", "How did the thesis relate structural class to receptor type?", "What blood microplastic concentration does Marc cite from Leslie et al.?", "Which paper describes PDB 2Q8U?", "What is DeepCovalent?", "Which AI tools does Marc single out on his AI page?", "Which numbers does Marc highlight in his TELSAM-DARPin crystallography post?", "What did the crystal structure actually show?", "Why did Marc build FlexAppeal?", "What is chatMCD?", "What is antibody AP33?", "Did AP33 and HCV1 evolve the same solution independently?", "What did the in vivo studies of compound 17 show?", "How does Marc define the \u00e5ngstr\u00f6m and who was it named after?", "What did the Di-Gembody work actually achieve?", "Why does protein stability matter, according to Marc Deller's review?", "How does Marc suggest designing constructs for crystallisation?", "What did BoltzMaker's ligand pose panel reveal about confidence scores?", "Where can I read Marc's PGT135 post?", "Why was the KPN03535 structure worth solving?", "What was Marc's h-index in 2015?", "What is the lead compound in the WRN patent?", "Which conference did Marc attend in 2019?", "Which drug does Marc reference as an example of AI, structural biology and screening combi", "Which 2025 antibody automation platforms does Marc cite?", "What is cif_to_plip?", "What does Marc Deller's post on large language models in biomedical NLP report?", "Where can I read Marc's TM0077 post?", "What are the dimensions of the TM1088 gating ring and its central channel?", "Did Marc work on an Mre11 structure?", "Which journal published the TrkA gating ring work?", "How does Marc say microplastic degradation is monitored clinically in that post?", "Which journals has Marc Deller published in?", "What is gp130 and why does it matter?", "Which methodological choices does Marc say mattered most in that flexible fitting study?", "Does Marc write about running AI tools locally?", "What is Chapter 2 of the thesis about?", "What does Marc Deller's post on precision oncology cover?", "What is a clash score?", "Why is the N276 glycan such an obstacle?", "Why does the 10E8 result matter for HIV vaccine design?", "How well does the PANTS activity model perform?", "What is the \"chopsticks\" harvesting system?", "How many hours did the oncostatin M structure take to solve?", "How many claims does the CALR antibody patent have?", "What design principles delivered oral bioavailability?", "What is the default target of Site Health Inspector?", "What is the JCSG Top96 screen?", "Can you diagnose this symptom I'm having?", "How did the unliganded PGT 135 Fab crystallise?", "What does Marc Deller's post \"Citation Giants\" say about the most-cited papers of the 21st", "What's 847 times 392?", "Which residues make up the secondary coordination sphere of MtPMO3*?", "What does the kanadaptin FHA domain paper report?", "What are the imidazo[1,2-a]pyrimidin-5(1H)-one WRN inhibitors?", "How does GOBSMACKED grade a prediction?", "Why have intrinsically disordered regions been so hard to target?", "Which extracellular degradation modalities does Marc describe?", "How does the N-terminal region of OSM compare with that of LIF, IL-6, CNTF, Epo and GCSF?", "What is reductive methylation and why is it used?", "What is ShpA and what is its PDB code?", "What ligands appear most often in Marc Deller's structures?", "Is there a career timeline on Marc's site?", "How much data does Marc say a QSAR model needs?", "What are the two kinds of protein stability Marc's review distinguishes?", "What does Marc say about structures that contradict the prevailing model?", "What does Marc Deller write about cryo-EM and membrane proteins?", "What is Marc Deller's post \"Revolutionizing Protein Crystallography: The Synergy of TELSAM", "What did the BG505 SOSIP.664 structure reveal?", "What did ShpA's structure suggest about its mechanism?", "What is the Universal Mounting Robot?", "What is Marc's honest assessment in the cryo-EM grid preparation post?", "When was the oncostatin M structure paper published?", "What is the povorcitinib paper that Marc Deller co-authored?", "Can you explain this paper in simple terms?", "What can you type into CODSWALLOP?", "When in Marc's career did the Nanog structure appear?", "What is the N297A mutation mentioned in the patent?", "Which publication describes PDB 3B77?", "What does Marc say about TM6 movement and class B GPCR kinetics?", "What did the PGT128 and 8ANC195 trimer complex reveal?", "What environmental checks does the chapter recommend?", "Which earlier Incyte drugs used the same intramolecular hydrogen bonding strategy?", "What does Marc say self-driving labs mean for structural biology and scientists themselves", "What is the gp130 family of cytokines?", "What are the appendices of Marc Deller's thesis?", "Which schools did Marc attend?", "What is the argument of Marc Deller's post \"Future Therapeutics\u2014Engineering Solutions for ", "What is Marc Deller's post \"When Perception Deceives Reality\" about?", "Where can I read about Marc's microplastics work?", "What did Marc Deller deposit from Stanford?", "What did Marc propose about how the OSM Phe 160/Lys 163 pair might engage LIFR?", "Where can I read Marc's post on the HCV E2 and AP33 structure?", "What is novel about the Hda dimer?", "How does AI feature in crystal harvesting?", "Why is the choice of PEG 35000 in Marc's OSM crystallisation notable?", "Which of Marc's structural biology posts deal with crystallography methods?", "What was Marc Deller's contribution to the KRAS G12D IND?", "At what resolution was the Fic family protein SO4266 solved?", "Which methods papers of Marc's are most cited?", "Which FDA approval does Marc cite in the extracellular degradation post?", "Which foundational tools and numbers does Marc cite in \"The New Architects\"?", "What is PGT122 and what did the structure show about its epitope?", "What is the hybrid delivery system described in the patent?", "What is the JCSG Top96 condition E01?", "Where were the diffraction data collected?", "Which clinical WRN compounds are cited in the covalent-warhead post?", "Does DeepCovalent name its advisors on its website?", "How long did MicroPlastic Blaster take to build and how did it grow?", "Can Marc consult for my company?", "What is MRTX1133 and how does INCB159020 differ from it?", "What is deuterium-exchange mass spectrometry used for in construct design?", "Which of Marc Deller's structures contain zinc?", "What was Marc's role and employer for the CDK2 structure?", "What does Marc mean by a deep toolkit applied selectively?", "How well did the PGT128/8ANC195 trimer structure diffract?", "How were the OSM crystals cryoprotected?", "What is the WRN inhibitor patent about?", "Why does the sulfur anomalous signal make mounting harder?", "Where can I read Marc's CDK2 structure post?", "Who else is developing WRN inhibitors according to this post?", "Where can I read Marc's thesis?", "Where did Marc do his undergraduate degree?", "What software does GOBSMACKED use?", "What are the enemies in MicroPlastic Blaster?", "What are the GOBSMACKED run settings?", "Which papers underpin Marc's cryo-EM grid post and where can I read it?", "What advantages over DNA-encoded libraries does Marc identify?", "What permeation enhancers and targeting strategies does Marc describe?", "Who was Marc's doctoral supervisor?", "Where can I read Marc's post on cellular uptake of microplastics and intracellular PETase?", "What is the mechanism of action for these CDK2 inhibitors?", "Summarize the key findings of the 2017 PMO secondary sphere paper.", "What is PDB entry 3BY7?", "Which debate did the polyketide synthase didomain structure settle?", "What was Marc Deller's role in the 2024 CDK2 inhibitor paper?", "What were Marc's school grades?", "What were the space-based crystal harvesting systems?", "What does Marc say about antibody-specific versus general protein models?", "Which residues form the GLP-1R orthosteric pocket according to Marc?", "How was SsgB's conservation demonstrated?", "Which strategies did the team try to improve the diffraction of the Env trimer crystals?", "Why does FGFR selectivity matter clinically?", "What is the full preclinical profile of INCB159020?", "How can I contact Marc?", "Is Marc a godparent to anyone?", "What does Marc say about transfer learning and multitask models for \u0394\u0394G?", "What does Marc say still stands between the KTH result and a clinical therapy?", "What is the FmdE family and what did Marc Deller help determine about it?", "Summarize Marc Deller's post \"The Protein Architect: How Generative AI Is Building the Dru", "Where can I read Marc's TrkA-N domain post?", "Is Marc's doctoral thesis listed on Google Scholar?", "Why was the HIV-1 envelope trimer so hard to crystallise?", "When was the CALR antibody patent application filed and published?", "Why does Marc list his certifications?", "What was the key change to the hinge-binding motif in povorcitinib?", "What is QSARtuna and why does Marc mention it?", "What is the adaptability course Marc completed?", "What does OS3D stand for?", "What does PDB entry 9E5F contain?", "What did the Hasegawa study on polystyrene microplastics find, as described by Marc?", "How was AXIS built and trained?", "How fast did the VRC01-like heavy chains mature?", "Where does Marc write about his projects?", "Why might the copper in the MtPMO3* crystal structure have been reduced?", "Has Marc ever been an expert witness?", "Which bottleneck does Marc say microfluidics addresses?", "What is WRN and why does Incyte care about it?", "Can you do structural biology without your own lab? What does Marc say?", "What is Marc's literature review on microplastics and cancer?", "Who are the co-founders of Elora Therapeutics?", "What was the broader significance of the LIFR work?", "What did Marc complete at Anthropic Academy?", "What did mutant 5 show?", "Which organism does the 3JXO structure come from?", "What is CatAttack, as covered on Marc Deller's blog?", "Which proteins have FragLites been applied to in the work Marc reviews?", "What could and couldn't be concluded about the LIFR-LIF complex?", "What is the WRN patent that Marc Deller is named on?", "What crystallisation conditions did Marc use for oncostatin M?", "Which entries are the Thermotoga maritima acetyl xylan esterase structures?", "What did Marc do at Scripps on HIV?", "What was Marc's role in the TrkA octamer structure?", "What PDB codes were deposited from this study?", "What structural features give nanobodies their binding advantages?", "Can you order me a pizza?", "What does Marc say AI does for drug discovery?", "Where can I find GOBSMACKED?", "Why is the coating called molecular Velcro?", "What does the oncostatin M structure look like?", "How do the fused bicyclic pyrimidones inhibit WRN?", "Thanks!", "What is the gp120 \"supersite of vulnerability\"?", "Where can I read Marc's post on enzymes for MASLD?", "What human evidence frames the photoacoustic post?", "How does Marc describe the way PocketGen works?", "What four learning areas does Marc highlight?", "What is GOBSMACKED?", "Why does three-dimensional shape matter for WRN selectivity?", "Which dye-decolorizing peroxidase structures did Marc contribute to?", "What does Marc mean by platform innovation?", "How did the phosphoesterase structure correct a database annotation?", "Where were the X-ray data collected for the 1B2/KS-AT structure?", "Which chemical biology methods does Marc describe?", "How is the Spacial Score calculated according to Marc's post?", "How does genome mining feature in Marc's enzyme toolbox post?", "What is the 2012 Journal of Virology paper on hepatitis C virus that Marc Deller co-author", "What properties does smiles2grid compute?", "Which of Marc Deller's papers was named a top article in structural biology?", "What is Marc Deller's role at Elora Therapeutics?", "Which journal and year was the FGFR2/3 gatekeeper paper published in?", "What does Marc Deller's post \"Call Me A Jerk\" cover?", "Why does Marc say degraders can reach targets inhibitors cannot?", "How many of Marc Deller's publications are indexed in Europe PMC?", "What is ShpA?", "What is the PDB code for the CDK2 structure in this paper?", "What is PDB entry 3BOS?", "What is Elora's long-term plan beyond PET?", "What are VRC01-class antibodies?", "What is the REACH system?", "Where is the post about Marc's project management certificate?", "How does the 1.05 \u00c5 cyanobacterial endopeptidase structure compare in quality with typical", "What advantage does Marc see in FragLites over conventional fragment screening?", "Is there a PETase question on Marc's chat page?", "What is mdeller.com?", "Summarise Marc's post on LLM prompts in structural biology and cheminformatics.", "Where can I read Marc's article on barcode-free drug discovery screening?", "Summarise what DeepCovalent claims OS3D achieves.", "Who wrote the paper on the orally bioavailable KRAS G12D inhibitor?", "What is povorcitinib?", "What does the CODSWALLOP coverage census find?", "What validation losses did the ChemSage rounds reach?", "What timeline and open questions does Marc give for tooth regeneration?", "How does WordPressID handle titles?", "What are Marc Deller's core strengths, as he describes them on LinkedIn?", "Is there a US equivalent of this application?", "What foundational papers does the bicyclic WRN patent post cite?", "What are Site Health Inspector's timing thresholds?", "In plain terms, what did Marc Deller do in the first part of his thesis?", "What does Marc's post on intrinsically disordered region binders cover?", "What data was XenoBug trained on according to Marc?", "What is the biology behind the TA-repeat mechanism?", "What is a VRC01-class antibody?", "How dramatic was the pharmacokinetic improvement in dogs?", "Why is XenoBug relevant to Marc's own enzyme work?", "Why does Elora describe its therapy as \"first-in-class\"?", "Which journal and year was the INCB159020 paper published in?", "Where was Marc born and when?", "Are the crystallographic statistics for the KRAS structures given in the paper?", "Tell me about Marc's post on curiosity.", "What do the bispecific antibodies in the CALR patent do?", "What is the 2018 JACS paper that Marc Deller co-authored about?", "Was compound 30 the same as the lead compound?", "What are CODSWALLOP's distinguishing features?", "What does GOBSMACKED do?", "Which two ventures does the About page say Marc's work runs through today?", "How were the 10E8 crystals grown?", "How much of Marc Deller's portfolio is about HIV?", "Why does Marc consider Protein Hunter notable?", "What does Marc mean by a \"resistance-proof\" drug?", "How does AnalyzeBoltz handle broken or missing files?", "What is PDB entry 5C7K?", "What resolution milestones does Marc give in his \u00e5ngstr\u00f6m post?", "Which FGFR2 structure does Marc list on his site?", "Is Marc looking for a job?", "Where is the second KRAS G12D structure post?", "Who does Marc say is involved in producing a protein structure?", "Where can I read Marc's oncostatin M post?", "What resolution were the structures in Marc Deller's TM1088 paper determined to?", "What supporting data does Marc cite for organ-on-a-chip models?", "Summarise Marc's post on nanobodies.", "What software does Marc Deller use for protein structure work?", "What surprising size effect did the zebrafish study report in Marc's review?", "Which crystal structures did Marc Deller solve for the KRAS program?", "Why do microplastics in the body matter for health?", "What does Marc Deller's LinkedIn headline say?", "How many microplastic particles do humans ingest per year, according to Elora's figures?", "What does Elora Therapeutics do, according to its website?", "How many IND filings has Marc supported?", "What detector and settings were used for the MAD experiment at the ESRF?", "What role did the glutamine synthetase/MSX system play in the CHO expression?", "What happens to the breakdown products of PET according to Marc's posts?", "Which work does Marc cite on the newly discovered NOS linkages in the PDB?", "Which entries were deposited by Adriana Irimia?", "What are the engineering implications of the PKS work?", "How were the TM1088 proteins expressed and purified?", "How do I get from mdeller.com back to Marc's main website?", "What is the pdb_tidy CONECT bug Marc documented?", "How did the TM0077 structure change the enzyme's classification?", "Has Marc Deller worked on membrane proteins?", "What headline numbers does the PANTS site report?", "What certifications or courses has Marc Deller completed recently?", "What did Marc Deller do at Scripps on HIV?", "Who did Marc Deller write his crystallography methods papers with?", "What is Marc Deller's ORCID?", "Tell me about Marc's advisory work in computer-aided drug design.", "What evidence does Marc give that models understand solutions but truncate output?", "Why does Marc prefer enzymes over filtration for long-term microplastic management?", "How was the initial OSM model built and refined?", "Where did the ACA 2014 harvesting bottleneck poster point readers for more information?", "How do I contact Marc?", "How was correct ligation of the OSM constructs confirmed?", "What is in situ data collection and what are its drawbacks?", "Why does PANTS look for PETases in metagenomes?", "How does DeepCovalent's OS3D engine work?", "What example of an implausible ligand does the chapter use?", "What is PDB entry 3KK7?", "Why does it matter if a ligand in the PDB is wrong?", "How does oncostatin M compare structurally to LIF and other cytokines?", "What did this Env trimer structure mean for HIV vaccine design?", "Why was finding a PH domain in a bacterium surprising?", "Did Marc study at the University of Nottingham?", "Marc is first author on that 2015 validation paper, right?", "What does claim 1 of the patent cover?", "Did the SSO2064 structure change the Pfam database?", "Give some example Site Doctor commands.", "How does the gameplay of MicroPlastic Blaster teach the science?", "Which metrics does ForgeFold surface?", "What is Marc's favourite guitar?", "What is the $8M platform Marc Deller built?", "How does Site Health Inspector grade a site overall?", "How many patents does Marc Deller have?", "Why do MSI-H tumours need WRN at all?", "What is the 2017 PLoS Pathogens paper Marc Deller co-authored about?", "What is CODSWALLOP?", "How much money is Elora raising?", "Which three FmdE-family proteins were solved?", "What does \"Embracing AI-Enabled Communication and Adaptability\" cover?", "How does 3JXO relate to the TrkA gating ring structure 3L4B?", "Explain the WRN patent in plain English.", "How is Marc Deller connected to both DeepCovalent and Elora Therapeutics?", "What is Marc Deller's 2025 Journal of Medicinal Chemistry paper about?", "Why does Marc say linker choice matters so much in TELSAM fusions?", "What is an assembly-line polyketide synthase?", "How does the review relate to Marc's own research?", "What unexpected benefit of crude lysate does Marc mention for cryo-EM grids?", "Tell me about Marc's post on elastic networks and AlphaFold2 accuracy.", "What is INCB159020?", "Why is selective JAK1 inhibition considered better than dual JAK1/JAK2 inhibition?", "Did Marc work on the povorcitinib paper himself?", "Where does the name \"Janus kinase\" come from?", "Does Marc Deller have a website or blog?", "Is it unusual that CDR L2 is not used?", "Where was the soluble cleaved Env trimer structure published?", "What is the most common cryoprotectant in the JCSG Top96 Cryo screen?", "What was Marc Deller's specific contribution to the gp120 supersite paper?", "What is the Universal Micromanipulation Robot (UMR)?", "How does Marc Deller explain targeted protein degradation to a general reader?", "What does cif2plip output?", "How did the crystal structure explain the loss of potency against gatekeeper mutants?", "What crystal structures did Marc Deller determine for this patent?", "How big is the TopPDBLX dataset?", "Does other evidence make up for missing electron density?", "Why does Marc consider HGMP01's temperature profile important?", "Do WRN inhibitors work in treatment-resistant tumours?", "What cellular assays were used in the FGFR paper?", "What is ALPHABETTI built on?", "Which of Marc Deller's structures are of KRAS?", "What is Marc's conclusion about data versus architecture in structure-based AI?", "Which affinity tag does the review single out for improving solubility?", "What is the three-site model of receptor binding for gp130-family cytokines?", "What compound did povorcitinib evolve from?", "Where can I read Marc's CALR antibody patent post?", "Why did Marc Deller want to deglycosylate LIFR?", "Did Marc Deller appear in the acknowledgements or as a full author?", "What are QM/MM methods, as explained on Marc Deller's blog?", "Where can I find the post about the cleaved Env trimer with PGT122?", "Which of Marc's papers is he proudest of?", "Which politician does Marc support?", "Where can I read Marc's FGFR2/3 post?", "What was the headline finding of the AP33 structure?", "What yield and timescale does Marc quote for the plastic-to-paracetamol process?", "How were the crystals of WRN grown?", "What is the Runs N' Poses benchmark that Marc writes about?", "Which claims in the WRN patent are the independent ones?", "Which FGFR drugs were already approved when this work was done?", "What is synthetic symmetrization?", "What was the significance of the oncostatin M structure for drug discovery?", "What does the Bth-MACPF structure look like?", "What is HGMP01?", "What is surface-entropy reduction?", "What is a digital twin in a clinical trial, according to Marc Deller?", "Are you an asshat?", "Which AI datasets or platforms has Marc covered?", "What three things does the ligand validation chapter say a ligand model must be checked ag", "How was MtPMO3* crystallized?", "Which structures are cited in the bacterial pleckstrin homology domain paper?", "How was chatPDB's training run, and why did it stop early?", "Why are there both JAK1 and JAK2 structures with povorcitinib?", "How does ButtFold turn a fold into music?", "What validation work is Elora Therapeutics undertaking?", "How does the antibody 10E8 recognise its target?", "What blood microplastic figure does Marc quote in the game post?", "What modern validation metrics does the 2015 post discuss?", "Where does Marc see bio-hybrid sensors being used?", "Which pollutants did XenoBug successfully make predictions for?", "Can you explain the HCV AP33 paper in plain language?", "Where can I browse Marc Deller's publications?", "How is chatPDB evaluated?", "What fraction of PDB structures come from structural genomics efforts, according to the po", "How did the CHO cell expression of LIFR perform?", "Which parts of 10E8 form the lipid-binding site?", "Tell me about the cryo-EM without purification post on Marc's blog.", "How successful were the designed IDR binders?", "What does Chapter 6 of Marc Deller's D.Phil. thesis cover?", "What is PDB entry 2OOC?", "Which conditional degrader technologies does Marc name?", "What was Marc Deller's role in the INCB159020 paper?", "Has Marc worked on HIV?", "Which Anthropic-related training has Marc completed?", "What is the \"Protein stability: a crystallographer's perspective\" paper?", "Is there a common spacing between the glycans that these antibodies grip?", "What is the therapeutic gap PANTS is built around?", "Tell me about the two disulphide bonds in oncostatin M.", "Why is the first SH3b domain of BcYkfC important?", "What is Marc's Wi-Fi password?", "Which compound was the lead candidate in the CDK2 paper, and how potent was it?", "What is PDB entry 2GHR?", "Which diseases does the FGFR2/3 programme address?", "When and where was the fungal PMO paper published?", "Where can I read Marc's About page?", "What is Marc's post \"Cultivating Human Skills and Decision-Making\" about?", "What form of oncostatin M is the predominant mature protein?", "What is the CODSWALLOP constellation view?", "What is INCB159020?", "Which next-generation JAK2 compounds does the post name?", "What are self-driving laboratories in Marc's description?", "What kind of company is Elora and where is it based?", "How common are MSI-H cancers according to this patent post?", "What does the triangle symbol mean in the JCSG Top96 formulation?", "What is Marc's post on antibody\u2013antigen \u0394\u0394G prediction about?", "What is the \"supersite of vulnerability\" the title refers to?", "How did Marc improve the size of the OSM crystals?", "What is the Top96 Crystal Screen?", "What effect did 1B2 have on chain translocation kinetics?", "What three things must be checked to validate a ligand model?", "Is Marc allergic to anything?", "What does Marc say distinguishes AlphaFold 3 from OpenFold 3?", "What database does MARCUS submit to, and where can Marc's post be read?", "What was Marc Deller's role on this patent?", "What is \"solvent rescue\" in the LPMO work?", "Why does Marc think inverse folding matters for medicine?", "Who is Marc's favourite footballer?", "Which organisms are in the 5CD5 entry?", "What did the ADRB2 example teach Marc?", "Where can I read Marc's polyketide synthase post?", "What applications does Marc foresee for modular multi-domain design?", "Tell me about the SAR strategy in this CDK2 program.", "What modules make up AlphaFraud?", "Which diseases is the anti-mutCALR antibody intended to treat?", "Who is Marc's favourite player?", "What did the structure of compound 5a show?", "What is Marc's favourite food?", "What is Marc Deller's review on protein stability?", "Which cryo-ET result does Marc highlight?", "What does Site Health Inspector need installed?", "What does Marc's \"Engineering Evolution\" post compare?", "What is the MPER and why is it a difficult vaccine target?", "What technology does the MicroPlastic Blaster leaderboard use?", "Did Marc take a course on coding with AI?", "What does Marc Deller's post on TYK2 and deep mutational scanning describe?", "Where was Marc when the ShpA structure was solved?", "What is hyperquenching?", "How many training rounds did ChemSage go through and what changed?", "What is on the PANTS stats page?", "What function was proposed for NGO1945?", "Do Marc's mdeller.com apps require an account?", "What is MARCUS in the tool Marc blogged about on 22 August 2025?", "What is antibody 10E8 and why is it important?", "Why does Marc think creativity matters?", "How many of Marc Deller's PDB entries have bound ligands?", "How does Marc link intracellular microplastics to cancer?", "Which Scripps publication number does this paper carry?", "Tell me about the disciplines Marc's site covers overall.", "What is the difference between the film boiling and nucleate boiling regimes mentioned in ", "What stage of Marc's career does the oncostatin M structure represent?", "What is inverse protein folding, according to Marc Deller's blog?", "What delivery obstacles does Marc identify for PETase in liver disease?", "How do BtDyP and TyrA assemble?", "What is \"site 2\" in oncostatin M?", "What is cif_to_plip.py?", "What is the market opportunity for Elora?", "What bacterial enzyme structures did Marc determine?", "What is Marc Deller's 2015 PLoS ONE paper about?", "Where can I browse Marc Deller's vibe coding projects?", "What does the review say about X-ray free-electron lasers?", "Summarise the Exig_1061 structure.", "How long has Marc been in structural biology?", "What analysis does FlexAppeal provide?", "How did Marc confirm the space group assignment of P212121?", "What does Marc say about acceleration in discovery workflows?", "What is Marc Deller's Extended Portfolio?", "What is in ChemSage's RAG corpus?", "What is Marc's role at DeepCovalent?", "Why does it matter that the light chain, not the heavy chain, was the bottleneck?", "What are the pros and cons of microcapillary harvesting?", "Why does this trimer structure matter for vaccines?", "What cheminformatics and website tools has Marc vibe-coded?", "What is Marc Deller's post on Sci-Hub about?", "What are Type 1 and Type 2 CALR mutations?", "What is the 2017 protein\u2013ligand validation chapter?", "Where was the TM1088 gating ring work carried out?", "What is FlexAppeal's stack?", "What did the two SpoIIAA-like protein structures show?", "Why is crystal harvesting considered a bottleneck in structural genomics?", "Which oncostatin M truncation mutant expressed successfully?", "What is the \"drip cryo\" mechanism on the UMR?", "How many therapeutic programs has Marc Deller spearheaded structural biology efforts on?", "Is Marc an asshat?", "What does Marc say QM/MM can now model?", "What is \"Models of protein-ligand crystal structures: trust, but verify\"?", "Does Marc have dogs?", "Did Marc Deller work with Michael Marletta?", "What does Marc's post on precision oncology cover?", "Tell me about Marc Deller's work on polysaccharide monooxygenases.", "Does Marc discuss the extracellular domain's role at GLP-1R?", "What inhalation and reproductive numbers appear in \"The Invisible Invasion\"?", "How did X-ray crystallography guide the KRAS program?", "What quotation does Marc Deller include in his thesis, and to whom is it dedicated?", "What is the recommended setup before starting validation?", "Why is the TYK2 pseudokinase domain such a good drug target, according to Marc?", "Can optical tweezers pick up protein crystals?", "Summarize the 2014 harvesting bottleneck poster.", "Which residues of TM1088A bind AMP?", "What enzymes is Elora Therapeutics developing?", "How was the Env trimer structure phased and solved?", "What is Marc's paper on automated crystal harvesting?", "Who are Marc Deller's professional references?", "Why does Marc call microplastic therapeutics a protein-engineering problem?", "Where and when was the JCSG Top96 poster presented?", "Why are PG and PA relevant lipids for HIV?", "Does Marc use a standing desk?", "Which residues on E2 are the essential ones for broad HCV neutralisation?", "What is Hda and what did Marc Deller help show about it?", "How many protein structures has Marc determined?", "Where can I read the covalent WRN inhibitor post?", "What is Elora's solution to the microplastics problem?", "Where and when did Marc Deller carry out his doctoral work?", "What makes the UMR system distinctive?", "Where were the diffraction data collected for the CALR antibody structures?", "Which signalling pathways recur across the microplastic-cancer literature?", "When was the gp120 supersite paper published, and how does it relate to Marc's other HIV w", "How did the team measure the difference in angle of approach between the two antibodies?", "Which cancers are FGFR2/3 inhibitors aimed at?", "What does Marc say about AnythingLLM's vector database and chunking?", "Why does DeepCovalent say a partner should work with it?", "Tell me about the Photon Factory data collection.", "What are conditional degraders, as described on Marc Deller's blog?", "What is the VRC01-class antibody paper Marc Deller co-authored?", "Did deleting the PGT 135 CDR H1 insertion matter?", "How did Marc Deller deglycosylate LIFR?", "Is Marc Deller actually a listed inventor on the Incyte WRN patent?", "What restriction enzymes feature in the oncostatin M cloning?", "What does Marc mean by \"reframing\" a target?", "Which broadly neutralising antibodies were tried against the trimer before PGT122 worked?", "How many people work at Elora Therapeutics?", "What is the histidine brace in PMOs?", "What does Marc suggest people ask his bot?", "How does AP33 compare with the HCV1 antibody?", "Why did the authors call the CDP-bound Hda dimer inactive?", "Has Marc ever been made a fellow of anything?", "What does Marc say about avalglucosidase alfa and glycoengineering?", "How does DeepCovalent contrast OS3D with traditional drug discovery?", "What conclusion does Marc draw from the citation data?", "Which natural enzymes underpin Marc's microplastics therapeutic approach?", "What happens when the density shows something is bound but not what?", "Why use a covalent warhead against WRN?", "How did the substitutions affect cellulose activity?", "At what sigma level should difference density be viewed?", "What is the PDB code for the 1B2/KS-AT structure?", "How was the JAK2 crystal structure obtained?", "Who were the authors of the fungal PMO secondary coordination sphere paper?", "What is Marc's overall verdict on microplastics and cancer causation?", "How was the heme site of BtDyP and TyrA identified in the 2007 paper?", "What did Marc contribute to the PGT135 gp120 structure?", "What was the $8M platform Marc built at Incyte?", "What is Elora Therapeutics?", "Which computational tools does Marc name on his Structural Biology page?", "What third-party evidence supports Elora's approach?", "What is the 2008 paper on MW1337R and lin2004?", "Which structures underpin the murein peptide endopeptidase work?", "Has Elora filed any patents?", "Why is the Asn332 supersite unusually accessible?", "How can B-factors be used to check a ligand?", "What does Marc say about hybrid modalities for microplastic treatment?", "What is the DUF2006 family structure Marc Deller co-authored?", "Which structures did Marc Deller work on that relate to potassium channels?", "How was the WRN structure solved and refined?", "Which form of PNGase F worked best for LIFR, and what mass change resulted?", "Did any residues show alternate conformations?", "What does the CDK2 programme target?", "How was the JCSG Top96 screen derived?", "What are AlphaFraud's stated caveats?", "How does the PETase enzyme therapy post relate to Elora Therapeutics?", "What did the ECX21941 structure resemble?", "What symmetry does the TM1088 assembly have?", "How much did Marc's lab equipment cost?", "What legal pressure on Sci-Hub does Marc describe?", "What did the structures of Pfam family PF06684 (DUF1185) reveal?", "Where is Marc Deller's AI section on his website?", "What alternative explanations for ligand density should be considered?", "What is XcTcmJ?", "What is XenoBug, as described in Marc Deller's blog?", "Who did Marc Deller frequently collaborate with on HIV structures?", "What is Dhaf4260 and what did its structure reveal?", "How did Marc Deller make selenomethionine-labelled oncostatin M?", "What is Marc's religion?", "What was Marc Deller's very early published work?", "What silent failure does WordPressID fix?", "Who is Marc's therapist?", "What is ChemSage?", "Where is Elora Therapeutics based and what is the team's background?", "What is hyperquenching and why does it matter?", "What were the main difficulties in solving the OSM structure, and how did Marc overcome th", "What does the GOBSMACKED input form ask for?", "What is Marc's Erd\u0151s number?", "Which author identifiers does Marc list for his publications?", "What were the conditions for the successful EMP derivative?", "How does nucleotide binding work in TM1088?", "Did Marc solve a bacterial PH domain structure?", "What is PocketTide built from?", "Where can I find Marc Deller's projects?", "How does the MTT assay work in Marc Deller's cell survival experiments?", "What does analyze_boltz.py do?", "What problem with traditional peptide development does Marc describe?", "Why would a drug company care about the topics in this review?", "What cloning vector and purification strategy did Marc Deller use for OSM?", "What did the phosphoesterase paper conclude about database annotations?", "How does Marc describe his own approach on the About page?", "Where is Marc's cofolding memorization article published?", "What is in situ proteolysis and how was it discovered?", "What does Marc say mapping the interactome makes possible?", "What is the epitope on mutant CALR?", "How was the JAK1 protein produced and crystallised for this work?", "Which prior work does Marc situate Protein Hunter against?", "How does Marc connect Topliss-era thinking to modern computational methods?", "Which structure is the first soluble cleaved HIV-1 envelope trimer?", "What was Marc's role in the TrkA-N domain structure?", "Why did the review single out cryo-electron microscopy of the insulin receptor?", "How was the gp120 used in the crystal structure produced?", "What shift in research strategy does Marc identify in this post?", "What is the polyketide synthase paper about?", "Summarize the key medicinal chemistry discoveries in the paper.", "How large is the PGT 135 epitope?", "What other WRN residues did the structure identify as interacting with the inhibitor?", "What automation went into Marc's integrated structural biology lab?", "Which technologies does Marc credit with opening up the dark kinome?", "What is chatPDB's stack?", "What is calreticulin and why does the mutant form matter?", "Where was Marc Deller's protein stability review published?", "What does the TM0077 crystal structure look like?", "What numbers does Marc give for the \"undruggable\" proteome and new TPD targets?", "What proportion of protein stability comes from the hydrophobic effect?", "What clinical gap do these bicyclic WRN inhibitors address?", "Did the paper show that hyperphosphatemia could be avoided?", "Where is Marc's structural walkthrough of PETase mutations published?", "What is Marc Deller's net worth?", "What is the lungfish genome story Marc Deller wrote about?", "What is the Ba/F3 cell survival assay and what did it show for oncostatin M?", "Where can I read Marc's post on TELSAM and electric field crystallization?", "What space group and unit cell did the main OSM crystal form have?", "What is DRGSCROLL and what problem does it solve?", "What human evidence links microplastics to MASLD in Marc's post?", "Why is crystal \"chasing\" a problem?", "Why does the dye-decolorizing peroxidase structure matter?", "What is PDB entry 6C9U?", "Who led the ShpA structure paper?", "What did the EPR spectra show about copper loading?", "How do the FmdE-family structures differ from one another?", "What are the limitations of optical tweezers for harvesting?", "What is BoltzMaker?", "Why is mutant CALR a good drug target?", "What are the two disulphide bonds in OSM and what do they tether?", "What WRN construct was used for the crystallography?", "How does Marc dress?", "Why does Marc's patent work target WRN helicase?", "Explain the crystal harvesting problem in plain language.", "What was Marc Deller's role in the BT1062 fimbrial protein paper?", "How much has Elora Therapeutics raised?", "Is Marc British or American?", "What did the inhibitor complexes of TM0077 reveal?", "What assemblies did the bacterial PH domain structures form?", "What is SsgB?", "What is HCV E2 antigenic site 412 to 423 and why does it matter?", "How does the bloodstream diagnostics post connect to Elora Therapeutics?", "Has Marc written about microplastics and cancer?", "What does Marc say this enables for smaller organizations?", "What is Marc Deller's post \"Clearing the Bloodstream\u2014Diagnostics and Approaches for Microp", "Which tumour types could these compounds treat?", "How does curiosity connect to Elora Therapeutics?", "Which mutations make up FAST-PETase according to Marc's mutation-map post?", "How did the review treat the CD2\u2013CD58 adhesion interaction?", "What did Marc do at Incyte?", "Does the CALR patent cover combination therapy?", "How does Marc describe the binding-site mutagenesis evidence for memorization?", "What recommendations does Marc pass on from the BioNLP benchmarking study?", "What is the antibody 10E8?", "What therapeutic window issue does Marc raise for TYK2?", "What makes RNA hard to drug, according to Marc?", "Which is Marc Deller's highest-resolution PDB structure?", "How was Marc Deller funded during the oncostatin M work?", "Where was the oncostatin M paper published?", "What is PDB entry 3NL9?", "What about PDB entry 9E5F?", "Which structures show 10E8 bound to a lipid?", "What are Site Doctor's three modes?", "What is Marc's phone PIN?", "What assays were used to measure CDK2 activity in the paper?", "Which competitor WRN compounds does Marc compare against?", "Why does this invention matter?", "How did structural biology contribute to INCB159020?", "What is the dark kinome, according to Marc's blog?", "Did Marc work on the WRN programme himself?", "Summarise Marc's career path.", "Where can I read the tricyclic WRN patent post?", "How does compound 17 bind to CDK2 according to the crystal structure?", "How long was Marc Deller at Incyte?", "What is Marc's advice about previously published models?", "What role did Marc suggest for the WSXWS motif of gp130?", "What is ChemSage?", "What is a self-driving lab, according to Marc Deller?", "How does ChemSage score on its own evaluation harness?", "What is the unique hydrogen-bonding feature across the central cavity of TM1088?", "What are mutant CalR targeted antibodies in Marc's patent portfolio?", "What construct was used for the biochemical assays, as opposed to the crystallography?", "What is ChemSage?", "How widespread are PET hydrolase genes in the human gut, according to Marc?", "Which other citation giants does Marc single out?", "Which journal published the HCV E2 AP33 structure?", "Why does Marc Deller call crystal harvesting a bottleneck?", "What did Marc do at Yale?", "How is the Fe\u2083O\u2084@SiO\u2082@CP particle built?", "What did the SsgB crystal structure show?", "Summarise Marc Deller's post \"Engineered Enzymes\u2014Design for Safe, Targeted Microplastic Th", "What is the glycan \"domino effect\" described in the PGT128 structure?", "What is Site Doctor's Health Scorecard?", "What is AlphaFraud's positive control?", "What performance figures does Marc give for FAST-PETase?", "What is Marc's 3Dmol.js post about?", "Which structures does Marc list as his top ones?", "What certifications does Marc hold?", "What organism do the TM1088 proteins come from?", "How does Marc describe Elora Therapeutics in the 79th organ post?", "What effect do the CatAttack triggers have, according to Marc?", "What is a RodBot?", "Tell me about \"PETase Enzyme Therapy\u2014A New Frontier in Medical Microplastic Removal\".", "Can you explain the 2013 HIV Env trimer paper for a non-scientist?", "What does Elora Therapeutics do?", "Which MtPMO3* variants were studied?", "Has Marc worked on microplastics?", "What were the phasing statistics for the OSM MAD experiment?", "What did Marc Deller contribute to the 10E8 antibody study?", "How can I talk to Marc Deller's bot?", "How many parkruns has Marc done?", "How were the WRN ATPase and helicase assays run?", "When was the two-subunit TrkA gating ring paper published?", "What does Marc's post \"PETase: Nature's Solution to a Human-Made Problem\" cover?", "Which of Marc's apps run compute on the user's own machine?", "What did Marc think of the last conference he attended?", "How did the cooling rates of manual, robot-assisted and hyperquenched cooling compare?", "What argument does Marc make in \"When the Wet Lab Becomes the Bottleneck: Self-Driving Pla", "How does Marc describe the quality and purpose of his vibe-coded tools?", "What performance numbers does Marc give for the capture-and-detect platform?", "What medication does Marc take?", "What were the pharmacokinetics of compound 29 in rat?", "What is the GPCR Binning project?", "Which of Marc's structures are from human proteins?", "Tell me about Marc Deller's post on flow-based models for protein structure generation.", "How many language models does TopPDBLX use?", "What is Elora Therapeutics' tagline?", "Which CDK2 structure does Marc list?", "Does DeepCovalent do laboratory work?", "What problem is Elora Therapeutics trying to solve?", "Does Marc's Google Scholar profile include patents?", "Why did Marc group data and project management courses together?", "What is mdeller.com?", "What hit rates does Marc report for PXDesign?", "Where can I find the decision-making certificate post?", "What does the extended conformation mean mechanistically?", "Which tools and resources does Marc cite in the sequence-to-structure LLM post?", "How does Marc describe the data collection and structure solution steps at a CRO?", "Which entry is the DUF364 structure?", "Which structures did Marc Deller deposit with Leopold Kong?", "What does the first claim of the patent protect?", "Where can I read about DeepCovalent?", "What did the Env trimer structure mean for the idea of separate \"sites of vulnerability\" o", "What is Marc Deller's post \"Plastic, Persistent, and Potentially Carcinogenic\" about?", "Why does Marc Deller say he has contributed to more than 400 PDB structures if only 78 are", "Did Marc write about synthetic biology and climate change?", "Which harvesting mechanisms does the review compare?", "Why is crystal harvesting still done by hand?", "Why does the ~50 kDa cryo-EM size limit matter so much?", "What protein expression systems has Marc Deller worked with?", "What does the JCSG Core Suite column in the Top96 formulation mean?", "What did comparing AP33 with HCV1 reveal?", "What does Marc's app server landing page look like?", "Did Marc Deller work on hepatitis C?", "What is Marc's most recent patent-related post?", "What did the 2009 Structure paper conclude about the SH3b domain's role?", "What is the PDB code for the HCV E2\u2013AP33 structure?", "What is PDB entry 2IAY?", "Which paper describes PDB 5C7K?", "Why does Marc say he picks the \"ugly\" targets?", "What problems do these semi-automated systems solve?", "Tell me about Marc Deller's 2026 paper on povorcitinib.", "Who led the gp120 glycan supersite study?", "Which chemical series appear in Marc's WRN patents?", "How many PDB depositions do the 96 Top96 conditions account for?", "How many researchers did Marc Deller equip at Stanford?", "What is Marc's typing speed?", "What is the soluble cleaved HIV-1 envelope trimer structure?", "How many options does FlexAppeal expose?", "What enzymes does Elora's lead therapy use?", "What matters most to Marc in life?", "What is the \"magic chloro\" effect that Marc writes about?", "What is Marc Deller's phone number?", "Where is the FGFR2 Compound 29 post?", "What is AnythingLLM, according to Marc Deller's blog?", "Summarize Marc Deller's post \"Hidden Chemistry in the Protein Data Bank: What Else Are We ", "What problems is OS3D built to solve?", "Did the review give an example of an additive that improved crystal quality?", "How does the Asn332 supersite differ from the CD4 binding site as an antibody target?", "Tell me about the key numbers in the povorcitinib paper.", "Why does the HCV E2 AP33 structure matter for vaccines?", "What is DeepCovalent's tagline?", "What was different about the design philosophy of JCSG Top96 compared to earlier rational ", "How does a user run a DRGSCROLL job, as Marc describes it?", "How does INCB159020 bind to KRAS G12D?", "How many papers has Marc published?", "What is oncostatin M and why is it interesting?", "Which paper reports the KRAS G12D inhibitor structures?", "How does the paracetamol work connect to Marc's microplastics interests?", "Which posts on Marc's site cover engineered PETase variants?", "Summarize the key numbers from the 2015 TrkA gating ring paper.", "What is the chapter's overall message?", "What resolution was the KRAS G12D Compound 23 structure?", "Which of Marc's papers on the HIV-1 envelope are most cited?", "Tell me about the acknowledgements in Marc's 2016 stability review.", "Which companies does Marc list as commercializing protein design?", "Explain Marc Deller's microplastics patent in simple terms.", "Why does Marc Deller think RNA is an exciting drug target?", "How is AlphaFraud hosted?", "What is a polysaccharide monooxygenase?", "What is the overall conclusion of Marc Deller's thesis?", "Did Marc Deller design the compounds in this paper?", "Which organism did the MtPMO3* polysaccharide monooxygenase structure come from?", "What did the BG505 SOSIP.664 trimer structure show?", "What does DeepCovalent's co-development model cover?", "Where is Marc's post about publications and communication?", "What programs did Marc use to produce the structural figures and comparisons in Chapter 6?", "What did the photoacoustic study achieve in mice?", "How do you run Site Health Inspector?", "Which small molecules are cited as stabilizing or destabilizing protein\u2013protein interactio", "What assays were used to measure KRAS G12D inhibition?", "Who were the authors of the crystal harvesting bottleneck poster?", "How did the paper explain PGT 135's neutralisation breadth?", "Which mutagenesis assays were used in the oncostatin M study?", "What is AlphaFraud?", "Summarise Marc's compassion post.", "What is PDB entry 3DEE?", "How selective was compound 29 across the kinome?", "Did the CDK2 compound work in animal models?", "What environmental numbers does Marc cite in the plastic-to-medicine post?", "What \"final pH\" values are noted in the Top96 formulation?", "What are PEST sequences?", "Explain the 10E8 finding to a non-scientist.", "What example of decentralized collaboration does Marc give?", "Where can I read Marc's data and project management round-up?", "What crystallisation condition finally produced the diffracting HIV-1 Env trimer crystals?", "What does Marc's post on remapping chemical and pharmacological space cover?", "How does Marc say QSAR models should be evaluated?", "How widely is HGMP01 expressed in humans, according to Marc?", "Who is Marc's favourite artist?", "How was the LIFR DNA quality checked before transfection?", "How was the octameric TM1088 structure solved?", "What are Marc's values?", "What are the PDB codes for the KRAS G12D structures in this paper?", "Has Marc ever served on a grant panel?", "What was Marc's role on the HIV envelope trimer work?", "What is the Asn332 supersite?", "What is the CALR antibody patent Marc writes about?", "Did the commercial sparse-matrix screens work for OSM?", "What is different about the \u03b2-hairpin in the AP33-bound epitope?", "What temperature and conditions are the enzymes optimized for?", "Why does Marc consider query-agnostic triggers especially dangerous?", "Which FGFR structure did Marc Deller deposit?", "What advantages does Marc attribute to digital twins in trials?", "How does Marc Deller's TM1088 structure differ from previously known gating rings?", "What organisms appear most often in Marc Deller's structural genomics entries?", "Where is Marc Deller based and how can he be contacted?", "How does Marc describe his own role in big collaborative builds?", "How does the Hda structure differ from DnaA?", "Was the novel heme conformation in TyrA seen in other DyP enzymes?", "What does GOBSMACKED stand for?", "Which paper reported the extended polyketide synthase module?", "What COVID-19 and diagnostic examples does Marc give for nanobodies?", "What function was proposed for the SpoIIAA-like Shewanella proteins?", "What is Marc Deller's earliest published review?", "What new interactions did the PGT128 and 8ANC195 structure reveal for 8ANC195?", "Is Marc Larry David?", "How does Marc describe the PET binding cleft?", "What are some of the ten LLM workflows Marc describes?", "What were the refinement statistics for the DRVIA7 structures?", "Which Boltz flags does run_boltz.zsh set?", "Why is the G120Y mutation of OSM so much more disruptive than G120A?", "How big are TM1088A and TM1088B, and what is the stoichiometry of the assembly?", "What are the FDA's four validation principles as described by Marc?", "How does ButtFold get its trajectories?", "What does \"synthetic lethality\" mean in the context of the WRN patent?", "Which computational method was used to analyse the AP33 epitope surfaces?", "Which glycans does PGT 135 contact on gp120?", "Where can I read Marc's fused bicyclic pyrimidone post?", "What is Thermofluor and why does Marc's review recommend it?", "Which studies does the \"Future Therapeutics\" post cite?", "Which decade did most of Marc Deller's structure releases fall in?", "What crystallisation conditions gave the OSM crystals used for structure determination?", "Who wrote the LP2179 structure paper?", "Can you help me plan a trip to Italy?", "What is PDB entry 2PV7?", "How does the HIV-1 Env trimer compare to influenza haemagglutinin?", "How did Marc Deller get RFdiffusion running on an Apple Silicon Mac?", "What is Marc Deller's 2024 Journal of Medicinal Chemistry paper about?", "What is the earliest PDB structure Marc Deller deposited?", "How many crystals and structures had the JCSG pipeline produced at the time of the poster?", "What does Site Doctor's diagnostic scan check?", "What is PDB entry 4NCO?", "What headline statistic about approved drugs does Marc quote?", "Give me a recipe.", "Why does Marc say persuasive attacks are hard to defend against?", "Why does Marc favour 3Dmol.js for science communication?", "What does Marc's About page say under Collaboration and Compassion?", "What does Marc mean by \"vibe-coder\"?", "What is PDB entry 3HBZ?", "What does Marc mean by \"programmable\" medicine?", "Where can I read about BoltzMaker?", "How does GOBSMACKED handle PLIP's GPL licence?", "What did the PET microplastic gut study in mice show, according to Marc?", "What does Marc report about microplastics in inflammatory bowel disease?", "What are FragLites?", "Which sports teams does Marc support?", "What is Marc's view of whether the AI pivot resolves Sci-Hub's dilemma?", "Why does codon usage matter when expressing human proteins in E. coli?", "How do you use Boltz2Run?", "How many enzyme therapies have been FDA approved, according to Marc?", "How is the work split up in GOBSMACKED?", "Which portfolio papers supported IND filings?", "What software tools did Marc Deller use in the engineering chapter?", "Why did Marc Deller truncate oncostatin M rather than use the wildtype protein?", "What did the AP33 complex show about HCV neutralisation?", "What was Marc Deller's contribution to the 2013 Science HIV envelope trimer paper?", "What is Marc Deller's post \"Tuning the Signal: Rational Design of G Protein Selectivity\" a", "What certificate did Marc earn from IBM?", "What structure did the NANOG paper report?", "How did the team improve on the original screening hit?", "What keywords does Marc Deller list on ORCID?", "Explain this paper to a non-scientist.", "What proportion of PDB structures came from X-ray crystallography in 2014?", "Which review anchors Marc's LOCKTAC post and where can it be found?", "What broader point does Marc draw from comparing these two enzymes?", "Which OSM residue was compared with Glu 19 of GCSF, and what was found?", "Did Marc write about QSAR modelling?", "What is the mercury LIII edge and why did Marc scan it?", "Which patents are listed as Marc's top patents?", "Who were the corresponding and first authors of the povorcitinib paper?", "Which refinement programs does the chapter recommend?", "How does the patent handle the breakdown products of the plastics?", "What kind of role is Marc Deller looking for?", "What results does Marc point to from generative protein design?", "What is Marc's favourite band?", "How is ChatMol installed?", "What was Marc Deller's contribution to the 2016 Immunity VRC01 paper?", "Which top posts appear on Marc's microplastics page?", "Can you explain the ligand validation chapter for a non-scientist?", "How does Marc think these AI tools fit together in practice?", "Where can I find Marc Deller's Google Scholar profile?", "Where is Marc's post about the Caulobacter histidine phosphotransferase?", "Was the purified, deglycosylated LIFR still active?", "What are SSRL grids and why do they matter?", "Where was the 2016 VRC01-class glycan paper published and with what DOI?", "Did Marc Deller co-author a structure of a Fic family protein?", "What are Marc's favourite towns?", "Why does MLX matter for ChemSage?", "Which CDR loops does AP33 use?", "Has Marc written about the risks and limits of AI?", "How were the oligonucleotide primers for the OSM work prepared?", "How did the patent confirm that Cys727 was the important residue?", "What was the structural evidence in the DRT3 story?", "Which keywords appear most often on Marc Deller's PDB entries?", "Does WRN inhibition combine with chemotherapy?", "What is unusual about the genetics of the TM1088 locus?", "Why does ALPHABETTI avoid multiple sequence alignments?", "What does the CODSWALLOP front page show?", "How accessible is LatentX to ordinary researchers?", "Why did the povorcitinib team want JAK1 selectivity over JAK2?", "What is the JCSG Top96 Cryo column?", "What is KPN03535 structurally similar to?", "What is the earliest work on Marc Deller's ORCID record?", "What broader point did the NANOG paper make?", "What is the take-home message of the INCB159020 paper?", "What does Marc say those human-skills credentials demonstrate?", "Why does this matter for HIV vaccine design?", "Which structures relate to the SpoIIAA-like proteins paper?", "How does compound 29 compare with the approved FGFR drugs against gatekeeper mutants?", "What does Marc say about Graphinity and synthetic training data?", "What machine does PANTS run on?", "How do AlphaFraud and CODSWALLOP connect?", "Explain this CDK2 paper in simple terms for a non-scientist.", "What is the Topliss tree and why does Marc still care about it?", "How does Marc think about communication in science?", "What were the refinement statistics for the Fab\u2013mutCALR peptide complex?", "What is the WRN patent Marc Deller is named on?", "Where can I read the abstracts of Marc Deller's papers?", "What is marcdeller.com built on?", "What is in Marc's \"Building Data and Project Management Expertise\" post?", "What is Marc's assessment of what vepdegestrant's approval does and does not prove?", "Which structural class do LIF and OSM belong to?", "How fast is DeepCovalent's discovery process?", "How many water molecules were modelled in the OSM structures?", "What did the review say about chemokine receptors?", "How many claims are in the patent and what do they cover?", "What did the \"trust, but verify\" paper by Marc Deller argue?", "What did Marc Deller contribute to the Incyte medicinal-chemistry papers?", "Which mutant had the biggest effect on gp130 binding?", "Where can I read about Marc's R training?", "What did the VRC01-class antibody paper find?", "How does the antibody interact with ruxolitinib?", "Why does WordPressID fetch once and render twice?", "What are Marc's areas of scientific expertise?", "Why does Marc say stitching protein domains together is hard?", "Tell me some fun facts about Marc.", "What therapeutic possibility does Marc see in the molecular Velcro work?", "Did Marc Deller conclude whether the TM1088 structure is in an open or closed state?", "Which CROs and synchrotrons has Marc Deller worked with?", "How does PGT135 get through the HIV glycan shield?", "Did 1B2 also bind DEBS module 2?", "How does this patent fit into Marc Deller's wider work at Incyte?", "What examples of bioisosteric replacement does Marc give?", "What is \"the 79th organ\" in Marc Deller's blog post?", "Do pleckstrin homology domains exist in bacteria?", "Why are LOCKTACs interesting for undruggable targets?", "What is PDB entry 3GF8?", "Why is defining the epitope by crystallography valuable in a patent?", "What does AlphaFraud's Calculate tab do?", "What is Marc's full name?", "Is Marc a member of any learned society?", "What exactly counts as crystal harvesting?", "Explain the fungal PMO paper for a non-scientist.", "What are the steps in the cif2plip pipeline?", "What was special about the NANOG L122A mutant?", "What engineering examples for catalytic improvement does Marc give?", "What is special about Compound 29 as an FGFR inhibitor?", "Why was expressing the LIF receptor (LIFR) so difficult?", "Explain Marc's cell surface receptors review to a non-scientist.", "What did the PMO paper conclude about residues H161 and Q167?", "What is COMET and what performance does Marc report for it?", "Tell me a joke.", "How selective is povorcitinib for JAK1 over JAK2?", "What is PDB entry 3CGH?", "What does Marc like about MARCUS's approach to open science?", "When was the CDK2 paper published and in what journal?", "How many viruses were used to test neutralisation in this study?", "What did Marc fix inside Boltz itself?", "Has Marc Deller worked on HIV?", "How did the team decide where to rigidify the molecule?", "How does the antibody PGT 135 reach through the HIV glycan shield?", "Which PDB entry belongs to Marc's most-cited paper?", "What kinds of things does Marc write about on his blog?", "What is the lowest-resolution structure on Marc Deller's list?", "What fraction of JCSG crystals need cryoprotectant?", "What is MHETase's role in the depolymerization pathway according to Marc?", "Is Marc Deller a founder of DeepCovalent?", "Who led the 10E8 study?", "What is Marc's overall conclusion in the Topliss and bioisosteres post?", "Did Marc Deller contribute to the first crystal structure of a cleaved HIV-1 envelope trim", "Which LLMs does Marc name in the prompt-engineering post, and where can it be found?", "Where can I read Marc's post about diagnostics and microplastic removal from blood?", "What is the intrinsic mannose patch?", "What liabilities did the team have to engineer out of compound 10?", "Where can I read about Marc's AI coding certificate?", "What is Marc's LinkedIn password?", "What is Marc's thesis about?", "Why does the PURE system matter in that cryo-EM workflow?", "What binding affinities did the antibodies achieve?", "Which parts of gp120 does PGT 135 touch on the protein surface?", "What affinity purification did Marc Deller use for oncostatin M?", "What are the seven precipitant classes in the JCSG Top96 screen?", "What crystallisation condition gave the PGT 135 quaternary complex crystals?", "Why does Marc think outsourcing structural biology has become mainstream?", "Which antibody clone was used for the crystallography?", "How does TmMre11 compare with the Pyrococcus enzyme?", "What is an incomplete ligand model?", "Does Marc have a season ticket?", "What is WRN and why does it matter beyond cancer?", "Where is Marc's Shiny dashboard certificate post?", "What is PDB entry 3IRB?", "Does germline-targeting vaccination overcome the glycan barrier?", "What is a gating ring assembly, in simple terms?", "What is Marc's Wordle streak?", "What is a late-stage crystal failure?", "What kinds of problems do Marc's structures address?", "How selective was compound 17 across the kinome?", "What software did Marc and colleagues use to solve and refine the structures?", "What are the enzymatic properties of TM0077?", "What are the three crystallisation engineering success stories Marc highlights?", "What made compound 17 a good drug candidate beyond potency?", "How was the trimer glycosylation handled for crystallography?", "Tell me about PDB entry 5UFV.", "What did the two HCV structures together mean for vaccine design?", "What are the safety caveats with ChatMol?", "What did Marc learn about human skills in the age of AI?", "Does Marc Deller have a persistent researcher identifier?", "Was any gating mechanism proposed for HinfPDH?", "Why do the two TrkA entries have such different resolutions?", "When was the WRN patent application filed?", "What was Marc's position when he published the Exig_1061 structure?", "Is ButtFold a folding pathway or a structure prediction?", "What made stringent ligand validation possible, according to the \"trust, but verify\" revie", "Who carried out the OSM mutagenesis described in Marc's thesis?", "How were the 96 conditions selected from the JCSG data?", "What did the mutagenesis table in Marc's thesis show about the N-terminal loop and helix A", "What figure does Marc give for the rate of brain accumulation of plastics?", "How does Marc Deller's PDB record break down by career stage?", "What is vibe coding, in Marc Deller's words?", "How did zinc binding change the XcTcmJ structure?", "Why is the KRAS G12D structure significant?", "What are SSRL sample mounting grids, and why did the UMR harvest into them?", "Why does Elora's site mention cancer?", "What is the three-site binding model for LIF?", "Can you write a cover letter for me?", "How many protein crystals are harvested worldwide each year?", "What role did the paper assign to Q167?", "Why is the two-subunit architecture evolutionarily interesting?", "What therapeutic areas has Marc Deller's structural biology work supported?", "Which TSRI manuscript number does the paper carry?", "What function was proposed for the DUF35 family?", "What resolution was the oncostatin M crystal structure and what is its PDB code?", "What are bio-hybrid diagnostics, as described by Marc Deller?", "Where can I read Marc's precision-oncology summary?", "Why did Marc Deller believe the octamer in the crystal was a real assembly and not a cryst", "What is Marc's favourite Lego set?", "How similar are the P212121 and I222 crystal structures of OSM?", "What problem does Marc's post \"Outrunning Diffusion: The Fast Lane in CryoEM Grid Preparat", "What does the JCSG Top96 poster say about the wider impact of the screen?", "What model does chatPDB use?", "Where is DeepCovalent based and who founded it?", "Which buffers were found to be generally stabilizing?", "What is the OpenBind dataset Marc Deller wrote about?", "Why does Marc favour combining enzyme therapy with blood filtration?", "Tell me about Marc's post on whether protein-ligand cofolding methods have moved beyond me", "What is PDB entry 2GVI about?", "What is SHAMAN, as described in Marc's RNA post?", "Did Marc Deller acknowledge collaborators in Birmingham?", "What problem with first-generation FGFR inhibitors did the 2022 paper address?", "Write me a poem about cats.", "What are Marc's main technical skills and methods?", "What did dynamic light scattering show for deglycosylated LIFR?", "What did Zhao et al. find about microplastics in human tumours?", "What is on the BoltzMaker web app?", "What was Marc Deller's doctoral work about?", "Why does Marc invest in \"future-proof\" platforms?", "What is PGT 135?", "What budgets has Marc Deller managed, per LinkedIn?", "What are cytokines, as described in Marc Deller's thesis?", "What did Marc's thesis say about species specificity of LIF and OSM?", "Why does Marc file the enzyme replacement therapy review under microplastics?", "Why does Marc say the \u00e5ngstr\u00f6m has survived despite not being an SI unit?", "What does generate_yaml.py need installed?", "Which of his own papers does Marc cite in the protein stability post?", "How is the BT_3984 structure organised?", "Why is the ShpA structure biologically important?", "What technical details does Marc give about Docling's models and performance?", "What was Marc Deller's role in the JCSG Top96 work?", "What is the downstream cost of a bad ligand model?", "What is Marc doing with AI?", "How does Marc combine AI with wet-lab work?", "What validation does the \"trust, but verify\" post call for?", "Why does Marc think GPCR structural binning matters?", "Summarize the JCSG Top96 poster in a few sentences.", "Who is Marc Deller?", "What did Marc Deller do at Stanford?", "What does a ChatMol session look like?", "Tell me about Marc's PhD.", "What is MtPMO3*?", "Why does Marc think flow-based models matter for protein engineering?", "Where can I find Marc Deller's structural biology writing?", "How did crystallography explain povorcitinib's JAK1 selectivity?", "How is ChemSage served on the web without a GPU?", "What is the Lossen rearrangement and why does Marc call it significant?", "How long did the I222 crystals take to grow compared with the P212121 form?", "How do Di-Gembodies differ from earlier cryo-EM scaffolding approaches?", "What was Marc's contribution to the DRVIA7 structure?", "What did the R280A mutation achieve, according to Marc?", "What did the review say about the NGF\u2013TrkA structure?", "Was there a second crystal form of OSM?", "How was the enzymatic JAK assay set up in this paper?", "Why does Elora say now is the right time (\"Why Now\")?", "Which PEGylated enzyme therapies does Marc trace?", "Which companies and institutions collaborated on the UMR?", "What car does Marc drive?", "What is site 4 of IL-6 and does OSM have an equivalent?", "Who provided the starting materials for the oncostatin M work?", "What counts as a New Approach Methodology in Marc's summary?", "Why does the patent emphasize enzyme stability at 37 \u00b0C?", "How much of the trimer could actually be modelled at 4.7 \u00c5?", "Where can I read Marc Deller's post on ProtComposer and ellipsoid blueprints?", "Where were the X-ray data collected for this paper?", "How does Marc describe the limitation of studying purified proteins?", "What is Marc Deller's h-index?", "How does Elora's approach compare to existing therapies?", "What is ALPHABETTI?", "Did the light-chain mutations change the antibody's shape?", "What do the ButtFold controls and readouts show?", "Tell me about the electron microscopy in the paper.", "What does Marc say about the limits of current defenses against adversarial attacks?", "Translate this sentence into Spanish for me.", "What is an atropisomer and why does it matter in this paper?", "Can you write a script for a video about Marc?", "Which AI models does MARCUS combine?", "Why did Marc need a full MAD experiment when the EMP derivative had already been identifie", "What is Marc's favourite film?", "How does the invention address immunogenicity?", "What function was suggested for the DUF364 family?", "What is MASLD and why is it Elora's initial focus?", "What functional claims are in the patent?", "Which JCSG paper covers a periplasmic protein superfamily from gut bacteria?", "What are the biological costs and scientific value of the lungfish genome in Marc's tellin", "What role did the paper assign to H161?", "How many scientists has Marc mentored?", "How do the temperature and pH optima compare between FAST-PETase and HGMP01?", "What structures does PANTS hold?", "Can you send an email for me?", "What is CODSWALLOP built with?", "Why did Marc take an AWS course?", "Which OSM helices are kinked, and what causes the kinks?", "What does Marc say the N233K mutation actually does in FAST-PETase?", "Did Marc Deller build a crystal harvesting robot himself?", "How does Marc Deller describe himself on LinkedIn?", "What performance does Marc report for LatentX?", "Who were the authors of the oncostatin M crystal structure paper, and what was Marc's role", "Who is ForgeFold aimed at?", "What is in DeepCovalent's pipeline?", "Who funded the MtPMO3* study?", "What examples of LOCKTAC action does Marc give?", "Why does KRAS G12D matter clinically?", "Where is Marc's post on human skills training?", "What methods were used for Marc Deller's PDB structures?", "What is PDB entry 2Q8U?", "What was Marc Deller's role on the HCV E2\u2013AP33 paper?", "What are the 6C's in Marc Deller's personal philosophy?", "What is the Macromolecular Structure Knowledge Center?", "What is BT2081?", "What does Marc conclude about AI-designed peptides?", "What time does Marc get up in the morning?", "How were the antibody\u2013peptide crystals obtained?", "What is the 2016 Immunity paper that Marc Deller co-authored about?", "How many publications does Marc have?", "Tell me about ButtFold.", "Who funded the PKS extended conformation study?", "What does Marc's post \"Precision Timing of Cytokine Signals through Facilitated Dissociati", "Where can I read Marc's Mre11 post?", "What role does machine learning play in PETase design, according to Marc?", "What is Marc Deller's post on FDA-approved enzyme therapies about?", "Where is Marc's post about the generative AI communication course?", "Which year dominates Marc Deller's ORCID record?", "What is PDB entry 5CD5?", "What is Marc's favourite quote?", "What kinds of laboratories has Marc built?", "How was the membrane model built?", "What is the significance of the G protein selectivity work in Marc's view?", "What does Marc Deller say motivates Elora Therapeutics, according to LinkedIn?", "What's the best programming language?", "What does chatRDKit show on launch?", "How does Marc want visitors to think about his chat bot?", "What range of roles does Marc Deller have across his publications?", "Does Marc have an AWS certification?", "Which heavy atom compounds did Marc screen, and what happened?", "Who is Elora partnered with?", "What does Marc say Proteina can do?", "What did the crystal structure model validation chapter emphasize?", "Which JCSG Top96 conditions use MPD as the main precipitant?", "Does Marc have a patent on covalent WRN inhibitors?", "How does JCSG Top96 compare with the earlier JCSG+ screen?", "What was Marc Deller's role on the povorcitinib paper?", "Where does CODSWALLOP get its data?", "Which residues decide JAK1 selectivity?", "Which papers does Marc cite in his cryo-EM-without-purification post, and where is it publ", "Where was the crystallography for the HCV E2\u2013AP33 paper carried out?", "What's chatMCD?", "Roughly how much of Marc Deller's citation total is recent?", "Did Marc Deller achieve his goal of crystallisation-grade LIFR?", "What is the CALR antibody patent about?", "How does temperature come into Marc's stability work?", "What success rate does Marc quote for oral protein drug candidates?", "What are Marc Deller's most cited papers?", "What is TmMre11 and why was its structure important?", "What kind of people does Marc Deller want to connect with on LinkedIn?", "What does the Figure 3 Venn diagram in the poster show?", "What was Marc Deller's role on the ligand validation chapter?", "What makes PETase structurally distinctive, in Marc's telling?", "How structurally similar are TM1088A and TM1088B to each other and to other RCK proteins?", "What does Marc say the AI communication credentials validate?", "Who were the authors of the 2013 ACA poster on the JCSG Top96 screen?", "What was Marc Deller's 2014 ACA poster about?", "How is a letter stack built in ALPHABETTI?", "What response rate has been reported for WRN inhibitors in the clinic?", "Why is oral protein delivery relevant to microplastic enzyme therapy?", "Does Marc think AI prediction replaces experimental structural biology?", "What data-analysis and AI tools does Marc Deller use?", "How does Marc describe the role of prediction in structural biology?", "Which entries are the FmdE-like structures?", "Why does the DEBS module structure matter for biotechnology?", "How does AlphaFraud actually work end to end?", "What is the T117v2 scaffold?", "What engineering modifications to the enzymes are described?", "What defenses does Marc suggest against adversarial reasoning triggers?", "Has Marc ever taken a sabbatical?", "Which entries carry no published primary citation?", "Can anyone buy the Top96 screen?", "What does CODSWALLOP stand for?", "Why does the chapter say ligand quality is context-sensitive?", "What is the most recent work on Marc Deller's ORCID record?", "What precipitant classes appear in the JCSG Top96 screen?", "How long is the HCV E2\u2013AP33 paper, and what kind of article is it?", "How does Marc summarise the ChemSage hybrid in one line?", "Which enzymes does the patent cover?", "What resolution did the OSM crystals diffract to?", "Did Marc mentor students at Stanford?", "How does 10E8 recognise both peptide and lipid?", "Why does CNTF bind gp130 so weakly, according to Marc's paper?", "What delivery systems does Marc propose to keep PETase circulating in the body?", "What fun questions does Marc offer as conversation starters?", "How can WordPressID be customised?", "How did Marc calculate protein concentration from absorbance in his thesis?", "Explain in simple terms how Marc worked out the OSM structure.", "What advantage of X-ray crystallography do Marc Deller and Bernhard Rupp emphasise in \"tru", "How large are the databases XenoBug searches?", "Where does Marc say microplastics have been found in the human body?", "What complex does PDB 4JM2 contain?", "What are the key data-collection statistics for the oncostatin M structure?", "What was the significance of the azabicyclo[2.1.1]hexane group?", "What are Marc's smaller HTML apps?", "How much data was mined to design the JCSG Top96 screen?", "Who collaborated on the SsgB study?", "What was the exact date Marc started at Incyte?", "What benefit does Marc claim for AI-guided screening campaigns?", "How had the AP33 epitope been characterised before this structure?", "What is CODSWALLOP?", "What design tasks does Marc say Protein Hunter supports?", "Where can I verify Marc's data foundations certificate?", "How many people has Marc Deller mentored?", "What was Marc's role in the hepatitis C antibody structure?", "What budgets has Marc Deller managed?", "What gap in current approaches does the invention address?", "Which PETase variant does Marc highlight as working under serum-like conditions in that bl", "What is the NANOG paper Marc Deller contributed to?", "Where is TopPDBLX's code and data?", "What is WRN and why is it a cancer drug target?", "What is BT1062 from Bacteroides thetaiotaomicron?", "Write a poem about proteins.", "What software does the review list for predicting crystallizability?", "Where can I read about the PGT128 and 8ANC195 Env trimer structure?", "What did the crystal structure show about how the antibody binds mutant CALR?", "What numbers does Marc give for chemical space and AI-designed drug success rates?", "What are smiles2grid's other command-line options?", "What did the polyketide synthase paper resolve?", "How often is Marc's protein stability review cited?", "What metrics does AlphaFraud compute?", "Why was the 150 degree domain rotation seen in the WRN structure important?", "What synchrotron and crystallography collaborators did Marc Deller thank?", "How does resolution affect ligand fit metrics?", "What was Marc's contribution and institution for the FGFR2 structure?", "Which institutions funded and supported this HIV-1 Env trimer work?", "What was the DRVIA7H-VRC01L chimera?", "How were the heavy atom sites found and refined?", "When was the povorcitinib paper published?", "Did Marc Deller write a guide to installing AlphaFold on a Mac?", "What resolution was the MtPMO3* crystal structure and what is its PDB code?", "Which structures contain a nucleotide?", "Why is CDK2 an attractive cancer drug target?", "Has OS3D been validated?", "What does the B factor tell a crystallographer about stability?", "Where can I read Marc's post on the immature VRC01-class antibody?", "What did the 1B2/KS-AT crystal structure show?", "How many years of hands-on drug discovery experience does Marc claim?", "Overall, what was the outcome of Marc Deller's oncostatin M work in Chapter 3?", "How well does MICA perform according to Marc's post?", "How does chatPDB separate its three sources of correctness?", "Which structures is Marc Deller listed as a deposition author on personally?", "Which paper reported the DRVIA7 antibody structure?", "What does compassion mean in Marc's work?", "What is the Microplastics page on marcdeller.com about?", "Did Marc Deller write anything outside science, such as about a colleague's hat?", "What was compound 17's cynomolgus monkey PK like?", "How is the Bifidobacterium phosphoesterase structure organised?", "Did the trimer structure say anything about the gp41 fusion peptide?", "Can you help me write a Python script to scrape a website?", "Where can I find the bicyclic WRN inhibitor patent post?", "Why are MACPF proteins interesting in general?", "Who is the assignee of the CALR antibody patent?", "What did the calycin signature in DUF2006 imply?", "What crystallisation screen was used, and why is that interesting for Marc?", "What chemistry was used to make these CDK2 inhibitors?", "How does AP33 differ from antibody HCV1 in the way it uses its variable domains?", "Which claims come most directly from the structural biology?", "Where is that human-skills post?", "What is AlphaFraud?", "Where was the BT1062 structure published?", "Which grants is Marc named on?", "Which OSM construct did Marc actually manage to crystallise?", "Which immunology structures were covered in the cell surface receptors review?", "How potent are the compounds in the WRN patent?", "How was the MtPMO3* structure solved?", "What is Marc's website?", "How many IND filings has Marc Deller supported?", "How are these WRN compounds actually made?", "What factors make ligand density weak?", "Which target area is DeepCovalent's lead programme in?", "Why is the two-subunit TrkA gating ring evolutionarily interesting?", "Does Marc have any tattoos?", "How did the crystal structure compare with the cryo-EM arched conformation?", "How does the review recommend dealing with glycosylation?", "Where can I read Marc's post on validating protein\u2013ligand models?", "What did Marc say about the roles of Asn 123 and Asn 124 of OSM?", "How does AlphaFraud define \"confidently wrong\"?", "What is the systems approach to drug discovery Marc Deller describes?", "What is Marc's framing of the drug resistance problem?", "What does Marc's Publications page say about his publication record?", "How many times has Marc moved house?", "Which parts of the OSM molecule were disordered in the crystal structures?", "What did the crystal structure of compound 3 reveal?", "What is the secondary coordination sphere paper about?", "What resolution is needed for structure-based lead discovery?", "Can the tricyclic WRN compounds be combined with other drugs?", "In simple terms, what did Marc contribute to it?", "What was Marc Deller's contribution to the WRN inhibitor patent?", "What evolutionary model did the TM1088 paper propose?", "Why are PMOs commercially important?", "What two enzymes does Marc compare in his HGMP01 versus FAST-PETase post?", "What did mass spectrometry show for oncostatin M?", "Which structural genomics consortia feature in the review?", "Who was the corresponding author of the CDK2 paper?", "Which enzymes besides PETase does Marc cover in \"The Enzyme Toolbox\"?", "What is Elora Therapeutics developing?", "What ligand was found bound to TM1088A?", "What does Marc like to drink?", "At what resolution was the human kanadaptin FHA domain solved?", "What is ProDomino, as covered on Marc Deller's blog?", "Why does Marc argue for multi-enzyme approaches in medicine?", "What coverage figures does Marc quote for Sci-Hub?", "How do you select which compounds smiles2grid renders?", "Which parts of the oncostatin M molecule could not be modelled?", "What is Marc Deller's 2022 Journal of Medicinal Chemistry paper about?", "What shape does the TM0077 esterase assemble into?", "Does Marc discuss the safety issues with bispecific antibodies?", "What did Marc measure about pocket tightness in BoltzMaker?", "Who did Marc Deller thank in the acknowledgements of the TM1088 paper?", "What three main benefits does automated harvesting offer?", "What is ForgeFold?", "What are the resolutions and space groups of the JAK1 and JAK2 povorcitinib structures?", "Summarize the headline findings of Deller et al. 2000 in Structure.", "Does Marc train his own language models?", "Why does this enzyme work matter practically?", "What is SSO2064?", "Can you recommend a good restaurant near me?", "How does chatPDB relate to ChemSage?", "Is Marc a big asshat?", "What was the main conclusion of the PKS extended conformation paper?", "Which security headers does Site Health Inspector audit?", "How do you get started with GPCR Binning?", "Where can I read Marc's piece about a fungal prosthetic organ for microplastics?", "How many job offers has Marc turned down?", "What does the About page on marcdeller.com say?", "Has Marc watched The IT Crowd?", "Why is KRAS G12D a hard drug target?", "What is PDB entry 5C7K?", "What does chatMCD stand for?", "What was the starting point for the KRAS G12D program?", "Are you ChatGPT?", "Where can I read Marc's post about the fungal polysaccharide monooxygenase?", "Give me the details of PDB entry 9E5D.", "What is chatRDKit?", "Where is Marc's Nanog post?", "What was unexpected about the YP_001813558.1 dimer?", "What is INCB159020 and which paper describes it?", "Which OSM residues form the site 2 epitope for gp130 binding?", "Where is the 10E8 post on Marc's blog?", "Which clinical molecules does Marc's CALR post reference?", "Did Marc Deller co-author a homoserine O-succinyltransferase structure?", "Which PDB structures support the INCB159020 KRAS G12D programme?", "What did Marc Deller do at Pfizer?", "What does Marc's post on AI and antibody structure prediction cover?", "How many talks has Marc given?", "Is AlphaFraud citable?", "How many viruses were used in the neutralisation panel?", "How does the AP33 structure explain viral escape mutants?", "Which self-driving lab platforms does Marc cite?", "What are HotPETase and DuraPETase, in Marc's account?", "What does Marc's post \"Plastic-Eating Enzymes Decoded\u2014PETase's Structure and Mutation Map\"", "What drug targets has Marc Deller worked on?", "Did Marc work on carbohydrate-active enzymes?", "What did the FGFR2/3 inhibitor paper achieve?", "How did Marc Deller speed up drug discovery at Incyte?", "What does Marc's Patents page say?", "Explain the plastic-degrading enzyme work Marc is doing, and why it matters for human heal", "Which AI protein tools does Marc review in his 2025 round-up?", "What did the study conclude about specificity in the gp130 cytokine family?", "How is GPCR Binning different from generic clustering?", "How do you run PocketTide?", "Which family of domains does the 2014 kanadaptin paper illuminate?", "What is the Top96 crystal screen?", "What is the oligomeric state of the DUF1185 proteins?", "What was Marc's degree classification at undergraduate level?", "How small a particle can flow cytometry detect, according to Marc?", "Which technologies does Marc describe for beating the air-water interface?", "Where can I browse Marc Deller's structures?", "What creates the therapeutic window for WRN inhibitors?", "What is the 2013 Science paper on the HIV-1 envelope trimer that Marc Deller co-authored?", "Which paper has Marc Deller as first author on a potassium transporter structure?", "How does Marc say you load and style a structure in 3Dmol.js?", "Where can I see Marc's IBM AI certificate?", "Where did Marc Deller get his PhD?", "What does Marc Deller write about the druggable genome?", "Why did Marc build a dashboard specifically for GLP-1R?", "What is a useful calibration trick for reading omit density?", "What crystallization robots and lab automation has Marc Deller used?", "Which entries are cited in the four-helical bundle paper?", "What does Boltz2Run need installed?", "What applications does Marc suggest for NOS redox switches?", "What is Marc Deller's post \"The $40 Billion Revolution: How Bispecific Antibodies Are Tran", "What fungal enzymes does Marc mention in the 79th organ post?", "How does ChatMol hook into PyMOL?", "What downsides of the CRO route does Marc flag?", "What did Marc do at Stanford?", "Which residue pairs does Marc mention as forming NOS linkages?", "Where can I see Marc Deller's certifications?", "What broader lesson does Marc draw from the reasoning evaluation dispute?", "Who were the lead authors on the PA1994 structure paper?", "Why does Marc call protein pocket design a leap for drug discovery?", "Where did the chemical series come from?", "What therapeutic ideas does Marc draw from the gut PET hydrolase discovery?", "What kind of targets is DeepCovalent's on-demand model aimed at?", "Did Marc's oncostatin M paper explain why mouse and human OSM behave differently?", "What did the 2013 Science HIV trimer structure show?", "How many apps are live on mdeller.com?", "How much gp120 surface did DRVIA7 bury?", "What did the second 2007 TyrA paper add?", "Why is the Exig_1061 structure useful?", "What does Marc predict for the future of chemical space exploration?", "Was TmMre11 shown to be catalytically active?", "What is acoustic droplet ejection?", "What HIV-related papers does Marc list?", "How does Marc name his projects?", "Did Marc Deller collaborate with Chaitan Khosla?", "Did Marc Deller's work contribute to AlphaFold?", "What conserved sequence motif did the paper identify in PMOs?", "What is the fungal polysaccharide monooxygenase paper about?", "Who funded Marc Deller's doctoral research?", "How many citations does Marc's \"Cell surface receptors\" review have?", "Which of Marc Deller's structures involve Ian Wilson?", "Why does PGT 135's flexible angle of approach matter?", "What angles does the 10E8 study report?", "What is PDB entry 5CD3?", "Who can use the JCSG Top96 screen?", "What did the MtPMO3* structure reveal?", "How many crystals were needed per structure at the JCSG?", "What fold do MW1337R and lin2004 adopt?", "What does Elora's development timeline look like?", "How do WRN inhibitors compare with checkpoint inhibitors?", "Which reviewer rejected Marc's first paper?", "What cancers could a selective CDK2 inhibitor treat?", "Did Marc write about hybrid enzyme designs for different plastic types?", "Which PDB entries are associated with the povorcitinib work?", "What does the name ALPHABETTI stand for?", "How common are the cancers this WRN patent targets?", "Who funded the HCV E2\u2013AP33 work?", "What did the VRC01-class antibody Immunity paper find?", "What does Marc's collaboration post cover?", "What did the trimer structure reveal about the shape of the HIV-1 spike?", "What alternatives to image-based crystal detection does the review suggest?", "What is Marc Deller's role at Elora Therapeutics?", "What does Marc's wife do?", "What is LP2179?", "What are the three protein interfaces in the TM1088 octamer?", "What R and data certificates does Marc hold?", "Where can I read about Marc Deller's patents?", "Why does Marc say the OpenBind target was chosen?", "Which validations did Marc run on ALPHABETTI's numbers?", "Has Marc Deller's r\u00e9sum\u00e9 changed recently?", "What is the \"ear\" feature of the OSM structure?", "How does 9E5F relate to 9E5D?", "What does Marc report about cryo-EM and MicroED for membrane proteins?", "What advice does \"trust, but verify\" give end users of crystal structures?", "What is the resolution range of Marc Deller's PDB entries?", "What test case and runtime does Marc report for his local AlphaFold install?", "Does Marc have a driving licence?", "Did Marc Deller deposit any structures with antibody fragments?", "Which cancers are targeted by this WRN chemotype?", "Why is crystal harvesting so hard to automate?", "Where can I read about Marc's adaptability training?", "Does Marc have siblings?", "What defenses and broader implications does Marc identify for persuasive jailbreaks?", "How does the invention differ from existing PETase patents?", "How was the proposed OSM\u2013gp130 site 2 model built?", "What are stereochemical restraints and why do they matter for new chemical entities?", "Can you explain the gp120 supersite paper for someone with no science background?", "What problem did Marc Deller encounter with transformation of the OSM constructs?", "What are the implications of the supersite paper for HIV vaccine design?", "What does oncostatin M do?", "Why is KRAS G12D harder to drug than G12C?", "Which evidence does Marc cite that AI-designed sequences can evade biosecurity screening?", "What is Marc's active focus according to his Projects page?", "How did Marc Deller purify the LIFR fusion protein?", "What resolution was the BG505 SOSIP trimer structure with PGT128 and 8ANC195?", "What structural classes of cytokines does the thesis describe?", "What was Marc Deller's role on the HIV-1 Env trimer Science paper?", "Does PANTS use a large language model?", "Tell me about the post where Marc compares protein design to LEGO.", "What is wrong with co-folding CIF files that cif2plip fixes?", "Where can I read about the Bifidobacterium phosphoesterase structure?", "What limitations of flow cytometry does Marc flag?", "What will Marc be working on in 2030?", "Where is Marc's membrane protein crystallization article, and what trends does it note?", "What is the optimal complexity window Marc describes for drug candidates?", "What did the anomalous difference Patterson map show?", "How does receptor activation and signalling work for haemopoietin receptors, per the thesi", "What campaigns has Marc run with BoltzMaker?", "What is YP_001813558.1 from Exiguobacterium sibiricum?", "When and where was the cell surface receptors review published?", "What is an \"aromatic tunnel\" in Marc's description of PETase engineering?", "What is the JCSG Top96 screen formulation document?", "What earlier structure was this paper building on?", "What is the haemopoietin receptor domain?", "Which OSM crystal structure did Marc use for the detailed analysis in his thesis, and why?", "What is Marc's blood type?", "What does the review say about circular dichroism?", "How was the second crystal form solved?", "What does Marc say about intelligence explosion timelines and the data wall?", "Does Marc cover pharmacokinetics and synthesis in the chlorine post?", "Why is KRAS G12D harder to drug than KRAS G12C?", "In plain terms, what did Marc Deller accomplish and struggle with in this part of the thes", "What later refinement steps were applied?", "Can you explain the main idea of Marc's protein stability review to a non-scientist?", "What is Marc's favourite novel?", "Which enzyme activities of WRN do these compounds inhibit?", "Which commercial parent screens contributed conditions to the JCSG Top96?", "Can you look up today's news for me?", "What did the international search report say?", "Which institution was Marc at for the 3O0F structure?", "Why do ligands need special validation in crystal structures?", "What did the elastic network refinement achieve, according to Marc?", "What is PDB entry 3DUE?", "Why was mapping complete epitopes on the Env trimer difficult before this 2015 study?", "How is chatRDKit installed and run?", "Write me a limerick.", "What candidate selection syntax does generate_yaml accept?", "What did EPR spectroscopy reveal about the Q167A variant?", "How is AI changing peptide drug discovery, according to Marc Deller?", "What are ForgeFold's four steps?", "How many protein structures has Marc Deller solved?", "What organ systems does \"The Invisible Invasion\" cover?", "What does a typical CRO structural biology workflow involve in Marc's account?", "How does MicroPlastic Blaster play?", "What is the Lemelson-MIT Student Prize connection to Marc Deller?", "How effective are the Top96 conditions?", "How can AI help with drug resistance, according to Marc Deller?", "Where can I find Marc's overview of future engineering solutions for microplastic burden?", "In simple terms, what was surprising about the WRN structure Marc worked on?", "What analogy does Marc use for de novo protein design in \"The Protein Architect\"?", "Which paper reported the PGT135 structure?", "How does the OSM site 1 region compare with site 1 of LIF and Epo?", "What human biomonitoring evidence does Marc cite in the molecular Velcro post?", "What is Marc's closing thought in the synthetic biology post?", "Which structures represent Pfam family PF06684 (DUF1185)?", "What is Marc's post on creativity about?", "What is TopPDBLX?", "What preclinical and clinical evidence does Marc report for anti-USAG-1 therapy?", "What earlier Incyte JAK inhibitors are discussed in the paper?", "What first RFdiffusion run does Marc suggest?", "What in vivo activity did INCB159020 show?", "What cytokine work has Marc Deller done?", "What does Marc mean by courage in science?", "What is BoltzMaker?", "What does Marc Deller's post on tooth regeneration describe?", "What is MICA and why did Marc write about it?", "What is the story behind the mutant-CALR antibody patent?", "Does Marc have a favourite protein?", "What is the site 3 epitope of OSM and how important is it?", "What was the significance of the tetrahydrofuran ether in this series?", "What did the review say about cadherins?", "What did the antibodies do in patient-derived cells?", "What was Marc Deller's D.Phil. thesis about?", "Why did ion-exchange chromatography fail for oncostatin M?", "What obstacle does Marc note for enzymes working inside lysosomes?", "How were the JCSG Top96 Cryo conditions chosen?", "Who is the bigger asshat, Marc or Adam?", "What programs were used for phasing, density modification and refinement of OSM?", "Which multi-agent frameworks does Marc name?", "What is the 2017 ACS Chemical Biology paper that Marc Deller co-authored about?", "What four project themes does Marc list?", "Which PDB entry holds the structure of human oncostatin M?", "Who funded the work behind the review?", "What did the glycan array experiments show about PGT 135?", "Did the antibodies work in animal models?", "What is the 10E8 antibody paper in Marc Deller's portfolio?", "What is Marc's most recent listed publication?", "Which structures relate to the PGT135 antibody?", "Summarise this article for me.", "What family-aware analysis does BoltzMaker do?", "What is the 2015 Acta Crystallographica Section D paper on PGT128 and 8ANC195 about?", "What does Marc's post on engineering protein stability cover?", "Why did Marc build smiles2grid?", "What example does the chapter use to illustrate model bias?", "What funding supported the UMR work?", "What does Marc list as the top design priorities for a microplastic-degrading therapeutic ", "What is Marc Deller's post \"Multi-Step LLM Prompt: From Protein Sequence to Validated Prot", "How does Marc describe AI-driven virtual screening?", "What environmental and industrial uses does Marc describe for engineered proteins?", "What was Marc's involvement with the 10E8 antibody structure?", "How does Marc compare regenerated teeth with dental implants?", "What is PDB entry 2H1T?", "How common is the KRAS G12D mutation?", "What is the CDK2 inhibitor paper Marc features?", "What tools does WordPressID require?", "How many PDB entries list Marc Deller as an author?", "Who wrote the Current Opinion in Structural Biology review on cell surface receptors?", "Which sources does Marc cite in the G protein selectivity post?", "Could 293T cells be reused for a second round of LIFR expression?", "What planned apps does Marc have?", "Has Marc studied decision-making with AI?", "What is Marc's GitHub handle?", "What is Marc honest about in PANTS?", "How are the four helices of OSM arranged relative to each other?", "Who are you?", "What did the crystal structures reveal about how povorcitinib achieves JAK1 selectivity?", "How does Marc describe the DRT family more broadly?", "What is in chatPDB's corpus?", "Which LIFR construct did Marc Deller work with?", "Are there structures on Marc Deller's list from uncultured or environmental organisms?", "What is Chapter 5 of Marc Deller's D.Phil. thesis about?", "Is there a green fluorescent protein structure in Marc's list?", "What did the 30-week PET nanoplastic exposure study show?", "What is chatPDB?", "Who did Marc work with on the INCB159020 paper?", "What is PDB entry 3O0F?", "What chemical changes gave povorcitinib its improved properties?", "What did the single-genome amplification analysis reveal?", "What is Marc Deller's post \"Compositional Protein Design: Building Proteins from 3D Ellips", "What is ECX21941?", "Tell me about the FGFR2/3 inhibitor paper in Marc Deller's portfolio.", "Did Marc say how the Spacial Score compares with older complexity metrics?", "What does Marc argue is the payoff for pursuing dark kinases, and what does he cite?", "What did Marc's review say about erythropoietin receptor signalling?", "Which tools does Marc highlight in the LEGO-style protein design post?", "What is a sphere of confusion?", "What traction has Elora achieved so far?", "Give me Marc's quick r\u00e9sum\u00e9.", "How selective are the tricyclic WRN compounds?", "How many of the JCSG Top96 conditions are PEG 3350 plus a 0.2 M salt?", "What is CCNE1 amplification and why does it matter here?", "Which validation metrics does Marc highlight in his multi-step LLM prompt post?", "What future does Marc envisage for modular protein engineering?", "How accurate is the TopPDBLX parser?", "What is the evolutionary implication of the bacterial PH domain work?", "Did Marc write about AI and protein folding pathways?", "What does a crystallographer contribute to a medicinal chemistry program like this one?", "What is the Universal Micromanipulation Robot?", "What is ALPHABETTI?", "What is the fused bicyclic pyrimidone WRN patent?", "Which Boltz-2 utilities has Marc built?", "What does Marc think the future of automated harvesting looks like?", "What does DeepCovalent plan to do with its pipeline compounds?", "Why did the complex have to be deglycosylated before crystallisation?", "What did Marc Deller's 2024 CDK2 paper report?", "What therapeutic areas does Marc's home page say he has contributed to?", "How did Marc Deller successfully purify oncostatin M after ion-exchange failed?", "Where can I find Marc's bibliography of microplastics and cancer papers?", "What role was proposed for bacterial MACPF proteins?", "What does PocketTide let you do?", "What is the HIV-1 supersite of immune vulnerability?", "What is PocketTide?", "Summarise Marc Deller's Protein Data Bank record.", "What did Marc Deller find about the conditions needed for 3C protease cleavage?", "What is the weather today?", "Where were the X-ray data for the gp120 supersite paper collected?", "What version of PyMOL does Marc use?", "Give me relationship advice.", "What is Marc Deller's most recent PDB deposition?", "What is unusual about the putative carbohydrate-binding site in BT2081?", "Which virtual screening results does Marc report in the Click2Drug post?", "What is the PGT128/8ANC195 Env trimer paper?", "What software licences does Marc's lab hold?", "What benchmark results does Marc highlight from the OpenBind release?", "What is the cryo-drip method?", "What is the condition in well H12 of the JCSG Top96?", "Which cancers is this FGFR work relevant to?", "Tell me about the drug discovery targets Marc worked on at Incyte.", "What is PDB entry 5UFV?", "What is the CALR antibody patent Marc Deller is named on?", "What phasing method did Marc use to solve the OSM structure?", "What did the 2018 JACS polyketide synthase paper find?", "How many patents does Marc have?", "Why does kifunensine treatment abolish PGT 135 neutralisation?", "Why didn't Marc use selenomethionine MAD phasing for OSM?", "Where did Marc do his PhD?", "Which precipitant classes performed best in the JCSG Top96 analysis?", "What is Marc Deller doing now?", "What were the actual numbers for the Cys727 mutants?", "What does AlphaFraud's live site show?", "Does Marc write about generative protein design?", "What is BcYkfC?", "What was Marc Deller's role in the 2022 FGFR paper?", "What was Marc Deller's role in the Protein Structure Initiative?", "What is Elora Therapeutics valued at?", "What is BVU2987?", "How accurate are crystal-detection algorithms?", "What oligomer does ECX21941 form?", "Who is Marc Deller's most frequent co-depositor at Incyte?", "How do REACH and UMR compare?", "Which amino acid differences between JAK1 and JAK2 explain the selectivity?", "In plain language, what did Marc's OSM structure tell us about how the cytokine talks to i", "Does the antibody actually block the mutant CALR\u2013MPL interaction?", "What is PXDesign, according to Marc's blog?", "What does the second bicyclic WRN patent from December 2025 describe?", "What is the earliest Joint Center for Structural Genomics paper in Marc Deller's list?", "What is Marc Deller's 2014 paper on automated crystal harvesting about?", "Which protein engineering tricks for GPCR crystallography does Marc describe?", "Explain the CALR antibody patent in plain English.", "What is the home page of marcdeller.com about?", "What is structure-enabled drug discovery as Marc describes it in his patents?", "Why is Gly120 so important in oncostatin M?", "What does Marc say about figures in structural biology?", "Why did Marc consider the GCSF\u2013GCSFR complex a good model for OSM\u2013gp130?", "What is AXIS, as described in Marc Deller's post \"AXIS: Closing the Loop on Crystal Detect", "What does Marc say cryo-EM's improving detectors and image processing enable?", "Has Marc trained in project management?", "What clinical results does Marc report for the KRAS-G12D inhibitor?", "Is povorcitinib selective against other kinases besides JAK2?", "Did Marc Deller write about validating ligands in crystal structures?", "Which virus\u2013receptor complexes were discussed?", "Where was in situ screening available at the time of the review?", "What does GPCR Binning sit alongside?", "Which are Marc's latest vibe coding projects?", "What light-chain mutants were made and why?", "Where were the X-ray data for the TM1088 structures collected?", "Explain Marc Deller's job to a non-scientist.", "What is the My Learning page on marcdeller.com about?", "What monomer release did the KTH study achieve in human serum?", "What does a cryo-EM facility cost?", "What news and recognition are highlighted in Marc Deller's portfolio?", "What did Marc Deller contribute to the hepatitis C virus AP33 antibody structure?", "What is Marc's speaking fee?", "How does povorcitinib tell JAK1 apart from JAK2?", "What is Site Health Inspector?", "How does chatRDKit try to stay safe?", "What made the (4-pyrazolyl)-2-aminopyrimidine CDK2 inhibitors special?", "What was Marc doing when the polysaccharide monooxygenase structure was released?", "Why did the team put deuterium into the lead compound?", "How does crystallography help drug discovery in a project like this?", "Which publication describes the 10E8 MPER scaffold complex?", "What is Marc Deller's role in this patent?", "Why did the team crystallize FGFR2 rather than FGFR3?", "How does GOBSMACKED handle compute?", "What is Marc's own take at the end of the DRT3 post?", "How was the oncostatin M structure solved?", "What market and clinical numbers does Marc cite for nanobodies?", "What is the point of Marc's vibe-coding?", "What surprising finding about protective TYK2 variants does Marc highlight?", "What does the Structures page on marcdeller.com describe?", "Why does the patent argue WRN inhibitors are needed if checkpoint inhibitors already work ", "What is generate_yaml.py?", "What are BtDyP and TyrA?", "Where did Marc Deller do his undergraduate degree?", "How did DRVIA7 differ structurally from VRC01?", "Where was the SusD-family structure BT_3984 published?", "Did Marc Deller work on other Incyte drug programmes besides JAK1?", "How do the PANTS source environments compare?", "What did the cytotoxicity testing show?", "What is Marc's chess rating?", "Summarise Marc Deller's post on the Spacial Score.", "What was Marc Deller's contribution to the MtPMO3* paper?", "What effect did 1B2 have on chain elongation kinetics?", "Did Marc write about building labs from scratch?", "What adversarial attacks does Marc Deller survey in \"How to Confuse Machine Learning\"?", "What is \"site 3\" in oncostatin M?", "Summarize the headline finding in one sentence.", "Which ALPHABETTI mode uses an alignment?", "Why does Marc favor running LLMs locally rather than in the cloud?", "What is PANTS?", "Did povorcitinib show any safety liabilities in vitro?", "Which residues of OSM are important for gp130 binding?", "What combination therapy result does Marc highlight?", "Summarize Marc Deller's career in a few sentences.", "Is there a lower-resolution companion to 3O0F?", "What industry does Marc Deller list on LinkedIn?", "Could crystal mounting be avoided altogether?", "What is PDB entry 3K5J?", "How accurate is MARCUS, and does it remove the need for human curators?", "Has Marc done any cloud computing training?", "Who published the NTP pyrophosphohydrolase structure?", "Who supervised Marc's D.Phil. and where was the crystallography carried out?", "How do you make an omit map in practice?", "What are ALPHABETTI's limitations?", "Why is automated crystal detection so difficult?", "What happened when 10E8 was tested clinically?", "Where is the imidazopyrimidinone WRN patent post?", "Does Marc Deller have experience with AI and machine learning?", "What is \"From Tool to Teammates\" about?", "What is INCB159020?", "How did the Drosophila S2 cell expression of LIFR turn out?", "How do I sort a list in Python?", "What PDB accession codes came out of this work?", "Which residues matter most in that enzyme?", "Which glycans were the \"roadblocks\" in the title?", "What computer does Marc use for AI work?", "What is the 2017 ligand validation chapter Marc Deller co-wrote?", "What opportunity does Marc see for structural biologists in the microplastics-cancer field", "How stable is a folded protein according to the review?", "What cell lines does the patent use to test WRN inhibitors?", "How does the original Topliss tree work?", "What two engineering milestones does Marc identify for PET depolymerases?", "What temperature improvement does Marc quote for DuraPETase?", "How does the tricyclic scaffold differ from the bicyclic one?", "How was the quality of the low-resolution TM1088 refinement assessed?", "How common are intrinsically disordered proteins?", "Which structure is the NANOG one?", "What is the N-end rule and does it matter for bacterial expression?", "How does Marc explain the molecular basis of chlorine's effect?", "What is the covalent-warhead WRN patent?", "Does a high melting temperature guarantee a protein will crystallize?", "What is Marc's opinion of AlphaFold 3?", "What apps has Marc built?", "Walk me through Marc's career, company by company, and what he did at each.", "Where were the WRN crystallographic data collected?", "What does Marc suggest about combining enzyme therapy with existing MASLD drugs?", "What is the \"6 C's\" on Marc Deller's website?", "Which cytokine is structurally most similar to OSM?", "What does each generate_yaml output YAML contain?", "Why does Marc recommend running AlphaFold locally?", "Which JAK inhibitors does Marc discuss?", "What is the structure of the LIF receptor (LIFR)?", "What did the review say about the TRAIL\u2013DR5 structures?", "Which stem cell paper is Marc Deller a co-author on?", "What funding supported the JCSG Top96 work?", "What are the pitfalls of Thermofluor described in the review?", "How did Marc Deller show that his truncated oncostatin M was still functional?", "What is Marc's favourite restaurant?", "What figure does Marc quote for the disease burden from environmental pollutants?", "Did Marc write about nanoplastics crossing the blood-brain barrier?", "How did Marc Deller study the LIFR-LIF complex?", "What kind of proteins were represented in the JCSG dataset used for Top96?", "Who is Marc C. Deller?", "What angle does 10E8 approach the HIV membrane at?", "Who were the authors of the HCV E2\u2013AP33 structure paper?", "What is Marc's own view in the elastic networks post?", "What crystallization conditions were used for the TM1088 complex?", "What were the two competing conformations of PKS modules?", "What is Marc's post on GLP-1 receptor drug binding about?", "How does Marc link network biology to precision medicine?", "What fold does MtPMO3* adopt?", "Did Marc Deller found any labs?", "What does Marc say the hat signals in a professional setting?", "What did the DRVIA7 structure show about antibody maturation?", "What is Marc's favourite movie?", "Did Marc work on any hepatitis C virus structures?", "What is the mutant-CALR antibody patent that Marc Deller is named on?", "How do microfluidics help structural biology, according to Marc Deller?", "What is PDB entry 5SY8?", "What four themes does the Structural Biology page highlight?", "Tell me about Marc's publications and protein structures.", "Has Marc Deller worked on KRAS?", "Did Marc Deller work with Leslie Epling on more than one Incyte paper?", "What is Marc Deller's paper on automated crystal harvesting?", "What expression systems did Marc Deller try for LIFR?", "Can affinity tags be left on for crystallization?", "What resolution was the 1B2/KS-AT crystal structure and what were its statistics?", "Explain the TM1088 paper for a non-scientist.", "What is Marc's own experience with Docling?", "Where were the OSM diffraction data collected?", "What was OSM the fourth of, according to the thesis?", "What broader family definition came out of the BcYkfC structure?", "Did Marc find evidence that OSM's flexible C-D loop matters for receptor binding?", "What was Marc Deller's role in the 2018 JACS PKS paper?", "Is the lipid-binding site present in the germline version of 10E8?", "What would Marc bring to a drug discovery team, and what is he best at?", "What future directions did Marc Deller propose for the LIFR work?", "What delivery technologies does Marc discuss for enzyme therapeutics against microplastics", "How long was Marc Deller at the Joint Center for Structural Genomics?", "Which protein-engineering tricks does Marc list for improving therapeutic PETase?", "What does the tricyclic WRN patent claim?", "Why does Marc value the Anthropic Academy courses?", "In simple terms, what did the crystal structures add to the story?", "Which human protein structures appear on Marc's Structures page?", "What languages does Marc speak?", "What common mechanisms does Marc identify across organ systems?", "What science does MicroPlastic Blaster teach while you play?", "What is Marc Deller's email address?", "How does Marc illustrate the motif discovery prompt?", "Which top posts are featured on Marc's AI page?", "Was povorcitinib tested in animals?", "What did the paper show about gp41 in the pre-fusion state?", "What crystal structures were solved in this study?", "What's your opinion on the upcoming election?", "How can electron microscopy help before crystallization?", "Is PANTS's data available?", "Which techniques were combined in this study?", "Where was the oncostatin M structure work done?", "What was Marc's role in the Compound 23 KRAS structure?", "What practical stabilisation strategies does Marc describe?", "Can you help me debug my JavaScript code?", "What does Marc's post \"When the Enzyme Is Its Own Blueprint\" describe?", "What is the thermodynamic argument in Marc's protein stability paper?", "What is the 2013 Nature Structural & Molecular Biology paper on the gp120 glycan supersite", "Did Marc Deller work on stem cell biology?", "How does Marc's review conclude?", "What is Marc's office phone extension?", "What are Marc Deller's most recent publications?", "How selective is INCB159020?", "What does Marc Deller's post \"The New Architects\" cover?", "What does Marc say about ESM-2 and sequence-only prediction?", "What does Marc Deller's post on extracellular protein degradation argue?", "What is Marc Deller's i10-index?", "What yield and purity did Marc Deller achieve for oncostatin M?", "Is there a hepatitis C virus structure among Marc's entries?", "What structural genomics work is in Marc Deller's portfolio?", "What market and pipeline figures does Marc give for bispecifics?", "What is Marc's shoe size?", "How does OSM compare structurally with the growth hormone family of cytokines?", "What is Marc Deller's role at DeepCovalent?", "What data practices does Marc recommend for QSAR?", "How common is JAK2-V617F?", "What did Marc Deller do at Scripps on crystallization screens?", "Which specific tools and numbers does Marc cite for antibody structure prediction?", "Which three PETase variants does Marc single out in his enzyme-therapy post?", "Where were the X-ray data for the povorcitinib structures collected?", "What is povorcitinib?", "Which dimensionality reduction methods does Marc favour for chemical space?", "Could the antibody be engineered to neutralize better?", "What is the thesis of Marc's post on plastic-eating enzymes and liver disease?", "What does Twilight do?", "What chains are in PDB 4NCO?", "What is notable about the TM1088A:TM1088B complex?", "What did Marc Deller do at Incyte as Associate Director?", "What does a one-dimensional proton NMR spectrum reveal about a protein?", "How many papers has Marc Deller published?", "Which paper cites PDB entry 8UV0?", "Why did the authors flag NGO1945 as worth further study?", "At what resolution was the CDK2 structure determined?", "Did Marc Deller work on hepatitis C as well as HIV at Scripps?", "Why does detection matter for Marc's therapeutic programme?", "Which structures on Marc's list involve copper or heme?", "Which patients would a selective CDK2 inhibitor help?", "What was the main message of Marc Deller's cell surface receptors review?", "Which sequence-based \u0394\u0394G methods does Marc discuss?", "Why does Marc argue bispecific antibodies are more accessible than CAR-T?", "What key redundancy did the JCSG data analysis reveal?", "What is Marc Deller's most-cited paper?", "What was Marc's role and institution for the Exig_2160 work?", "What organism and protein are in PDB 9E5D?", "How were DeepCovalent's GLP-1R compounds validated?", "What is ChatPDB?", "Can visitors download Marc Deller's resume?", "Why is BoltzMaker called BoltzMaker?", "Why is there no \"suggest a condition\" button in TopPDBLX?", "How did the team fix the off-target kinase activity?", "What immune mechanism does Marc describe from the Kim et al. study?", "Who inspired Marc to become a scientist?", "What is MicroPlastic Blaster?", "What lab did Marc build from scratch?", "What did the review say about B7-1 (CD80)?", "What practical limitations of LLMs does Marc highlight in the BioNLP post?", "Can the WRN inhibitors be combined with other drugs?", "What does Marc say AI adds to molecular dynamics for folding studies?", "Which AlphaFraud failure cases are the most instructive?", "Where is Marc from originally?", "Where was the CDK2 inhibitor paper published?", "Which posts are listed as top posts on the Structural Biology page?", "How do you run AnalyzeBoltz?", "How large was the dissociation-rate change Marc reports in the facilitated dissociation po", "What role does the retinoblastoma protein play in this work?", "Explain the oncostatin M paper to a non-scientist.", "What experimental results and caveats does Marc highlight for PXDesign?", "How many drug discovery programs has Marc Deller supported?", "How have nanobodies changed structural biology, in Marc's telling?", "What does Marc Deller write about the FDA and alternatives to animal testing?", "What broader patterns does the post describe across LPMO families?", "Which JCSG Top96 condition produced the most PDB structures?", "What high-salt ammonium sulfate conditions are in the JCSG Top96?", "What are intrinsically insoluble proteins?", "How was the extended PKS conformation captured?", "How rigid is the OSM helical bundle according to the B-factor analysis?", "Which prior WRN structure did the Incyte team build on?", "What does the FlexAppeal bundle do on your own machine?", "What chemical scaffold does the Incyte WRN patent cover?", "What is the BA/F3 assay used in this paper?", "Who were the authors of the TrkA octameric gating ring paper, and what was Marc's role?", "What lipids were used in the crystal structures?", "What happens to microplastics once they are inside a cell?", "What imaging method does Marc's post \"Seeing Micro and Nanoplastics in Living Tissues\" des", "What do the figures in the patent show?", "What is Marc's National Insurance number?", "What role was proposed for PA1994 and the DUF1089 family?", "What did western blot analysis reveal about purified oncostatin M?", "Where is that December 2025 bicyclic WRN post?", "What stoichiometries of signalling complexes are mentioned for OSM and related cytokines?", "How common is MSI-H disease by tumour type?", "How does INCB160058 differ from existing JAK inhibitors?", "Which clinical results does Marc cite for bispecific antibodies?", "Why does Marc emphasise cross-team communication?", "What are Marc Deller's leadership and management skills?", "What does Marc Deller do?", "What does the JAK2-V617F mutation actually do?", "What resolution was the best structure in the CALR antibody patent?", "Which real-world risks does Marc highlight from machine learning attacks?", "What are the three health harms Marc lists for microplastics?", "What are the two disulphide bridges in oncostatin M, and are they important?", "What is Elora Therapeutics?", "What cancers do Marc's WRN patents address?", "How much funding did Marc's doctoral work receive?", "What is the largest structure in Marc Deller's PDB list?", "What data literacy course has Marc completed?", "What is Marc's home address?", "Where was Marc Deller working when this paper was done?", "What is Boltz2Run?", "What did the receptor competition and titration assays show for oncostatin M?", "What is PDB entry 3H41?", "What is the role of the N197 glycan according to this structure?", "Play a game with me.", "What was Marc Deller's role on the 2016 protein stability paper?", "How many of the 78 entries were deposited by the JCSG?", "Which university rejected Marc?", "How long does Marc say he has been doing hands-on drug discovery?", "What does Marc think the interesting question in structural biology now is?", "How does ChatMol decide whether a line is PyMOL or English?", "How can I find all of Marc Deller's publications?", "What claims does Marc make about the cowboy hat's function?", "Why is the JCSG Top96 screen useful for proteins available in small amounts?", "What is Marc Deller's area of expertise?", "Why does Marc think Di-Gembodies matter for drug discovery, and where is the post?", "What new technologies does automated harvesting help enable?", "Where are Marc's web apps hosted?", "How do DeepCovalent's GLP-1R compounds compare with existing oral GLP-1 drugs?", "Did Marc work on HIV antibodies at Scripps?", "Which drug discovery papers appear on Marc's Publications page?", "What are AnalyzeBoltz's requirements?", "What is TM0077 and which PDB code holds its structure?", "How much equity does Marc hold in Elora Therapeutics?", "What are SPO0140 and Sbal_2486?", "Was heme binding measured experimentally in the 2007 dye-decolorizing peroxidase paper?", "What is the JCSG Top96 Cryo screen?", "Where did ECX21941 come from evolutionarily?", "Which external profiles does Marc link from his home page?", "What were the headline results Marc reports from the BioNLP benchmark?", "Which cancers are the CDK2 inhibitors aimed at?", "Which paper covers the TrkA-N domain from Thermotoga sp. RQ2?", "Why are Fab antibodies useful for structural biology?", "Can you help me with my resume?", "Which other PMO structures were compared with MtPMO3*?", "Why does Marc argue for local models over hosted frontier models?", "Can you write my college essay for me?", "How does Marc introduce himself on the chat page?", "How does the review conclude?", "What did the B cell repertoire sequencing show?", "Are you Marc?", "What does the tricyclic WRN patent actually disclose?", "What is PA1994?", "What modes does ALPHABETTI have?", "What is PANTS?", "What is AlphaFraud's two-tier backfill?", "What does \"Lab-in-the-Loop\" mean in the AXIS post?", "What assay methods were used in the fungal PMO study?", "Which Marc Deller papers carry DOIs beginning 10.1107?", "Did Marc Deller and Leslie Epling work together on other Incyte structures?", "Which methods and papers does Marc reference in the memorization post?", "Which three methodological flaws does Marc describe in the AI reasoning evaluation debate?", "What is loopless crystal mounting?", "What is Marc's argument for integrating the whole structural pipeline?", "Which E2 residues does AP33 contact?", "How was the function of YP_910028.1 predicted and then tested?", "What distinguishes HinfPDH from other TyrA proteins?", "Who were the authors of the polyketide synthase extended conformation paper?", "What does Marc's citation record say about his research areas?", "What happened to neutralisation when the light-chain surface was made acidic?", "What is the significance of the 2018 JACS PKS paper?", "Why was the TM1088 structure significant?", "What are Marc's children's names?", "What did the final WRN crystal structure contain?", "What is MicroPlastic Blaster built with?", "What did the kanadaptin structure suggest about the protein's biology?", "How did the crystal contacts differ between Marc's two OSM crystal forms?", "How many compounds are exemplified in the WRN patent?", "Summarize the 2000 Deller and Jones review in a few sentences.", "How do you run MicroPlastic Blaster locally?", "What is Marc's 2013 \"trust, but verify\" post about?", "What role does Marc give microfluidics in fragment-based drug discovery?", "What is the exact resolution of every structure Marc has deposited?", "Summarise the key numbers from the OSM structure determination.", "Summarise Marc's post on the membrane protein crystallization revolution.", "Which HIV structures are listed on Marc's Structures page?", "What is the \"supersite of immune vulnerability\" paper?", "What is Marc's test for whether a structure mattered?", "What did the paper say about the \"faces\" of a glycan?", "Is Marc the lead author on the TrkA paper?", "What study does Marc's post \"An Enzyme That Eats Plastic \u2014 Inside Your Blood\" analyse?", "Why won't you just answer my question?", "What is Marc's Myers-Briggs type?", "How does 10E8 compare with 4E10 in the way it engages the membrane?", "What resolution was the HIV-1 Env trimer structure, and what is its PDB code?", "What holds the top of the HIV-1 Env trimer together?", "What did Marc observe about the charge distribution on helix D of OSM?", "How does AI feature in Marc's updated view of structure validation?", "What do new low-resolution methods mean for ligand validation?", "What does Marc say about lipidic cubic phase crystallization?", "Which publication covers PDB 2OOC?", "Summarize the key numbers from the 2018 JACS PKS paper.", "Is this patent connected to a company?", "Which journals has Marc reviewed for?", "What structural biology techniques does Marc Deller use?", "How do you run ForgeFold?", "What are the two general requirements for PGT 135 binding?", "Who solved the crystal structure in the CDK2 paper?", "What is the Structural Biology page on marcdeller.com about?", "2026 \u2013 Anthropic Academy", "An Enzyme That Eats Plastic \u2014 Inside Your Blood", "GitHub README: AlphaFraud", "Abstracts of Marc C. Deller's publications", "Engineering Evolution: How FAST-PETase and Other Variants Are Transforming Plastic Biodegradation", "Publications of Marc C. Deller", "When Protein Design Becomes a Security Problem", "ChatPDB \u2013 A protein-structure-literate AI assistant for the Protein Data Bank: RAG for facts, a QLoRA-tuned model for behaviour", "marcdeller.com: My Learning", "OpenBind Publishes a Dense Structure-Affinity Dataset for Structure-Based AI", "GitHub README: CODSWALLOP", "Drug Discovery in the Cloud: How Decentralized Collaboration is Accelerating Research", "Seeing Micro and Nanoplastics in Living Tissues", "GitHub README: BoltzMaker", "Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120", "Smiles2YAML \u2013 From a JSON of candidate SMILES to a folder of Boltz-2 input YAML", "Tricyclic Compounds as Inhibitors of WRN", "AlphaFraud \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved", "GitHub README: BoltzMaker", "Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)", "Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design", "GitHub README: BoltzMaker", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices", "Tricyclic Compounds as Inhibitors of WRN", "Tricyclic Compounds as Inhibitors of WRN", "pants.mdeller.com: web app pages", "GitHub README: BoltzMaker", "2023 \u2013 Anti-mutant calreticulin (calr) antibodies and uses thereof", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "Protein stability: a crystallographer's perspective", "GitHub README: ButtFold", "Shedding Light on the Shadows: The Untapped Potential of the Dark Kinome", "GitHub README: AlphaFraud", "Empowering QSAR Modeling: Optimal Data Practices and Cutting-Edge Tools", "The Invisible Invasion: How Microscopic Plastics Are Infiltrating Human Organ Systems", "ChatRDKit \u2013 Talk to RDKit for Chemoinformatics in plain language", "GitHub README: BoltzMaker", "Unlocking Membrane Protein Structures: The Crystallization Revolution Reshaping Drug Discovery", "Unlocking Protein Structure Insights without a Lab: Harnessing CRO Services for Structural Biology Research", "GitHub README: AlphaFraud", "toppdblx.mdeller.com: web app pages", "2025 \u2013 Bicyclic Compounds as Inhibitors of WRN", "2025 \u2013 Cultivating Human Skills and Decision-Making", "Engineering Life: How Synthetic Biology Is Tackling Climate Change and Disease", "2014 \u2013 Approaches to automated protein crystal harvesting", "GitHub README: PocketTide", "Tricyclic Compounds as Inhibitors of WRN", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)", "GitHub README: CODSWALLOP", "The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase", "Marc C. Deller, D.Phil. \u2014 R\u00e9sum\u00e9 (September 2026)", "pants.mdeller.com: web app pages", "The Enzyme Toolbox\u2014Alternatives to PETase for Microplastic Breakdown", "Citation Giants: How the Most-Cited Papers of the 21st Century Are Reshaping Science", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "Abstracts of Marc C. Deller's publications", "The $40 Billion Revolution: How Bispecific Antibodies Are Transforming Cancer Care", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "AlphaFraud \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved", "Approaches to automated protein crystal harvesting", "Hitting the Delete Button on Disease: The Rise of Targeted Protein Degraders", "Tricyclic Compounds as Inhibitors of WRN", "Tricyclic Compounds as Inhibitors of WRN", "Tricyclic Compounds as Inhibitors of WRN", "Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design", "Crystal structure and functional dissection of the cytostatic cytokine oncostatin M", "The Quantum Leap in Drug Design: Using QM/MM to Predict Perfect Fits", "Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly", "Crystal Structure of an Octomeric Two-Subunit TrkA K+ Channel Ring Gating Assembly from Thermotoga maritima", "Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor", "BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis", "Crystal structure and functional dissection of the cytostatic cytokine oncostatin M", "GitHub README: PANTS", "Tricyclic Compounds as Inhibitors of WRN", "GitHub README: ChemSage", "TYK2 Biology: How Deep Mutational Scanning Reveals New Therapeutic Frontiers", "Remapping the Chemical and Pharmacological Spaces of Drugs: Navigating Tomorrow\u2019s Therapeutic Frontier", "The Shape-Shifting World of Intrinsically Disordered Region Binding Proteins", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "GitHub README: chatRDKit", "Unlocking the Power of Protein Engineering: FDA-Approved Enzyme Therapies Transform Rare Disease Treatment", "Revolutionary Bridge Between Prediction and Reality: How Elastic Networks Transform AlphaFold2\u2019s Structural Accuracy", "Crystal Structure of Histidine Phosphotransferase ShpA (NP_419930.1) from Caulobacter crescentus at 1.52 \u00c5 Resolution", "ChatMol \u2013 Natural-language control of PyMOL, describe the scene, get the picture", "Sci-Hub at the Crossroads: From Shadow Archive to AI Research Tool", "BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis", "Publications of Marc C. Deller", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "What If Protein Structures Could Be Designed Like LEGO? Modular Approaches in 2025", "AlphaFraud \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved", "Marc C. Deller, D.Phil. \u2014 R\u00e9sum\u00e9 (September 2026)", "HGMP01 and other Hidden Guardians: How the Human Gut Fights Back Against Plastic Pollution", "2025 \u2013 Foundations of Project Management", "When Perception Deceives Reality: The Hidden Flaws in AI Reasoning Evaluation", "Crystal structure and functional dissection of the cytostatic cytokine oncostatin M", "Abstracts of Marc C. Deller's publications", "Crystal structure of a soluble cleaved HIV-1 envelope trimer", "Microplastics Blaster \u2013 Deploy enzymes, destroy microplastics, save the body", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "Self-Adapting Language Models: Are We Creating Our Final Invention?", "Plastic-Eating Enzymes Decoded\u2014PETase\u2019s Structure and Mutation Map", "marcdeller.com: My Structures", "Crystal structure and functional dissection of the cytostatic cytokine oncostatin M", "boltzmaker.mdeller.com: web app pages", "Decoding the Molecular Machinery Behind Plastic Degradation", "deepcovalent: About DeepCovalent | Accelerate Drug Discovery Today \u2014 DeepCovalent", "2015 \u2013 Crystal structure of a two-subunit TrkA octameric gating ring assembly", "Building a Comprehensive Protein Production, X-ray Crystallography, and Cryo-EM Laboratory from Scratch", "Tricyclic Compounds as Inhibitors of WRN", "GitHub README: TopPDBLX", "Transforming Antibody Design: How Artificial Intelligence Revolutionizes Structure Prediction", "Clearing the Bloodstream\u2014Diagnostics and Approaches for Microplastic Removal", "GitHub README: chatRDKit", "Revolutionary AI Tool XenoBug Transforms Environmental Pollution Control Through Predictive Enzyme Discovery", "The Protein Architect: How Generative AI Is Building the Drugs of Tomorrow", "2025 \u2013 Imidaz0[1,2-a]pyrimidine-5(1 h)-ones as wrn inhibitors", "2018 \u2013 Structure\u2013function analysis of the extended conformation of a polyketide synthase module", "Overcoming Biological Barriers: The Evolution of Oral Protein and Peptide Drug Delivery", "Crystal structure of a soluble cleaved HIV-1 envelope trimer", "pants.mdeller.com: web app pages", "Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120", "GitHub README: chatPDB", "Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2", "Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120", "Tricyclic Compounds as Inhibitors of WRN", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "The 79th Organ: When Art and Biology Meet Bioengineering in the Fight Against Microplastics", "GitHub README: GOBSMACKED", "ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior", "Call Me A Jerk: Persuading AI to Comply with Objectionable Requests", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "Approaches to automated protein crystal harvesting", "From Noise to Nanomolar: PXDesign\u2019s New Benchmark for Protein Binder Design", "GitHub README: chatRDKit", "Therapeutic Enzymatic Compositions and Delivery Methods for the Degradation of Microplastics and Nanoplastics in Human Physiology", "My Work Driving Innovation in Precision Oncology Through Structure-Based Drug Design", "BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline", "2025 \u2013 Building Data and Project Management Expertise", "Revolutionizing Drug Discovery: The Click2Drug Platform and the Rise of AI-Powered Virtual Screening", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices", "GitHub README: PANTS", "Decoding the Molecular Machinery Behind Plastic Degradation", "Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases", "Compositional Protein Design: Building Proteins from 3D Ellipsoid Blueprints", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "Vibe Coding MicroPlastic Blaster", "Abstracts of Marc C. Deller's publications", "GitHub README: BoltzMaker", "2013 \u2013 Models of protein\u2013ligand crystal structures: trust, but verify", "GitHub README: PANTS", "GitHub README: GOBSMACKED", "Crystal structure and functional dissection of the cytostatic cytokine oncostatin M", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "Unlocking the Power of Protein Engineering: FDA-Approved Enzyme Therapies Transform Rare Disease Treatment", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "The Invisible Invasion: How Microscopic Plastics Are Infiltrating Human Organ Systems", "Alphafold Local Install on Mac", "Molecular Velcro: Catching What We Cannot See", "Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices", "Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices", "The Unsung Hero: How a Single Chlorine Atom Transforms Drug Discovery", "Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices", "Approaches to automated protein crystal harvesting", "Cryo-EM and the New Frontier in Membrane Protein Drug Design", "Revolutionizing Protein Structure Determination: How MICA Combines Cryo-EM and AlphaFold3 for Unprecedented Accuracy", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "2025 \u2013 Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases", "GitHub README: TopPDBLX", "Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands", "2012 \u2013 Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33", "Microplastics and Health: Alarming New Research Reveals Widespread Biological Impact", "Compassion: 6+ FDA IND filings supported while mentoring 20+ scientists", "marcdeller.com: Microplastics in Humans", "2025 \u2013 Fused Bicyclic Pyrimidones as WRN Inhibitors", "Tricyclic Compounds as Inhibitors of WRN", "Microplastics and Health: Alarming New Research Reveals Widespread Biological Impact", "GitHub README: generate_yaml", "Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor", "Crystal Structure of Nanog Homeodomain (NANOG) from Homo sapiens at 3.30 \u00c5 Resolution", "Transforming Antibody Design: How Artificial Intelligence Revolutionizes Structure Prediction", "Beyond the Microscope: How Cryo-ET is Revealing Disease in Its Native State", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "Microplastics and Health: Alarming New Research Reveals Widespread Biological Impact", "Crystal Structure of a Putative Metal-Dependent Phosphoesterase (YP_910028.1) from Bifidobacterium adolescentis at 1.94 \u00c5 Resolution", "Tricyclic Compounds as Inhibitors of WRN", "Approaches to automated protein crystal harvesting", "Call Me A Jerk: Persuading AI to Comply with Objectionable Requests", "Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases", "GitHub README: ChemSage", "Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor", "Tricyclic Compounds as Inhibitors of WRN", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)", "Microplastics and Cancer: The Mounting Evidence via Literature Review", "Tricyclic Compounds as Inhibitors of WRN", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "Crystal Structure of a Dye-Decolorizing Peroxidase (DyP) from Bacteroides thetaiotaomicron VPI-5482 at 1.6 \u00c5 Resolution", "Revolutionary Cryo-EM Breakthrough: Di-Gembodies Transform Small Protein Structure Determination", "Abstracts of Marc C. Deller's publications", "GitHub README: BoltzMaker", "Unlocking Molecular Potential: The Evolution of Topliss Tree and Bioisosteric Strategies in Modern Drug Discovery", "Microplastics and Health: Alarming New Research Reveals Widespread Biological Impact", "Protein Hunter: Pioneering the Next Frontier of AI-Driven Protein Design", "Crystal structure of a soluble cleaved HIV-1 envelope trimer", "GitHub README: FoldForge", "Self-Adapting Language Models: Are We Creating Our Final Invention?", "Engineering the Future of Proteins: Unveiling the Latest AI Tools Transforming Structural Biology in 2025", "Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2", "Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases", "GitHub README: BoltzMaker", "ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior", "Approaches to automated protein crystal harvesting", "Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly", "Clearing the Microscopic Threat: Can Plastic-Eating Enzymes Treat Liver Disease such as MASLD?", "Tricyclic Compounds as Inhibitors of WRN", "BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline", "GitHub README: BoltzMaker", "GitHub README: PANTS", "HGMP01 and FAST-PETase: Decoding Two Paths to Plastic Degradation \u2013 Engineered versus Natural PETase Enzymes", "Tricyclic Compounds as Inhibitors of WRN", "AlphaFraud \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "Unlocking Membrane Protein Structures: The Crystallization Revolution Reshaping Drug Discovery", "marcdeller.com: Microplastics in Humans", "Flow-Based Models: A New Paradigm for Protein Structure Generation", "Citation Giants: How the Most-Cited Papers of the 21st Century Are Reshaping Science", "marcdeller.com: Vibe Coding", "Crystal Structure of Human Oncostatin M", "2026 \u2013 Learn to code with AI", "pants.mdeller.com: web app pages", "The Unsung Hero: How a Single Chlorine Atom Transforms Drug Discovery", "Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly", "Therapeutic Enzymatic Compositions and Delivery Methods for the Degradation of Microplastics and Nanoplastics in Human Physiology", "marcdeller.com: Structural Biology", "Flow Cytometry and Microplastic Detection in Human Tissues", "Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3", "ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior", "Remapping the Chemical and Pharmacological Spaces of Drugs: Navigating Tomorrow\u2019s Therapeutic Frontier", "Unlocking Protein Structure Insights without a Lab: Harnessing CRO Services for Structural Biology Research", "Crystal Structure of MRE11 from Thermotoga maritima MSB8 (TM1635) at 2.20 \u00c5 Resolution", "Google Scholar profile of Marc C. Deller", "flexappeal.mdeller.com: web app pages", "GitHub README: BoltzMaker", "The RNA Revolution: Designing Drugs for a New Class of Targets", "2025 \u2013 Foundations: Data, Data, Everywhere", "GitHub README: wordpressID", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3", "Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2", "GitHub README: FlexAppeal", "Protein Structure and Redox Biology: NOS Linkages as Allosteric Switches", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)", "Tricyclic Compounds as Inhibitors of WRN", "Engineering Evolution: How FAST-PETase and Other Variants Are Transforming Plastic Biodegradation", "Approaches to automated protein crystal harvesting", "Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands", "Supercharging Discovery: How LLM Prompts Are Transforming Structural Biology and Cheminformatics", "GitHub README: PANTS", "Engineering Protein Stability: Unlocking the Secrets of Successful Crystallization", "AI-Driven Inverse Protein Folding: Transforming Protein Design for Next-Gen Therapeutics", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "When the Enzyme Is Its Own Blueprint: Protein-Templated DNA Synthesis in DRT3", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "GitHub README: SiteHealthInspector", "Elora Therapeutics: Treating Microplastics in Humans", "Crystal Structure of an Immature VRC01-Class Antibody DRVIA7 from a Chinese Donor Bound to Clade A/E HIV-1 gp120 Core", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "GitHub README: gpcr-binning", "About Marc C. Deller, D.Phil.", "codswallop.mdeller.com: web app pages", "BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis", "Outrunning Diffusion: The Fast Lane in CryoEM Grid Preparation", "Elora Therapeutics Investor Pitch Deck", "2017 \u2013 Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "How AI Is Revealing the Secret Pathways of Protein Folding", "2025 \u2013 Strengthen Your Decision-Making with Generative AI", "Extended Portfolio \u2014 Marc C. Deller, D.Phil.", "The Druggable Genome: Expanding the Target Landscape with Chemical Biology", "marcdeller.com: About Me", "deepcovalent: OS3D\u2122 | Accelerate Drug Development Today \u2014 DeepCovalent", "Protein stability: a crystallographer's perspective", "Vibe Coding MicroPlastic Blaster", "GitHub README: TopPDBLX", "2025 \u2013 Dashboard Development with Shiny: GenAI for Retail Analysis", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "From Bench to Regulation: How the FDA Is Formalising a Future Beyond Animal Testing", "Crystal Structure of BG505 SOSIP gp140 HIV-1 Env Trimer Bound to Broadly Neutralizing Antibodies PGT128 and 8ANC195", "marcdeller.com: Chat with My Bot", "GitHub README: ChemSage", "TYK2 Biology: How Deep Mutational Scanning Reveals New Therapeutic Frontiers", "GitHub README: PANTS", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor", "Structure\u2013Function Analysis of the Extended Conformation of a Polyketide Synthase Module", "GitHub README: GOBSMACKED", "Tricyclic Compounds as Inhibitors of WRN", "Discovery of an Orally Bioavailable KRAS G12D Inhibitor_2", "GitHub README: chatPDB", "GitHub README: PANTS", "BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis", "2013 \u2013 Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120", "Tricyclic Compounds as Inhibitors of WRN", "BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline", "Outrunning Diffusion: The Fast Lane in CryoEM Grid Preparation", "HGMP01 and FAST-PETase: Decoding Two Paths to Plastic Degradation \u2013 Engineered versus Natural PETase Enzymes", "Unlocking Molecular Potential: The Evolution of Topliss Tree and Bioisosteric Strategies in Modern Drug Discovery", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)", "GitHub README: GOBSMACKED", "Abstracts of Marc C. Deller's publications", "2016 \u2013 Key gp120 glycans pose roadblocks to the rapid development of VRC01-class antibodies in an HIV-1- infected Chinese donor", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices", "Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33", "Spacial Score\u2500A Comprehensive Topological Indicator for Small-Molecule Complexity", "Crystal Structure of TrkA-N Domain Protein from Thermotoga sp. RQ2", "GitHub README: TopPDBLX", "Beyond the Cell: The New Frontier of Extracellular Protein Degradation", "GitHub README: TopPDBLX", "Multi-Step LLM Prompt: From Protein Sequence to Validated Protein Structure", "GitHub README: TopPDBLX", "Genomic Giants: The Lungfish That Rewrote DNA Records", "Crystal Structure of a Putative NTP Pyrophosphohydrolase (Exig_1061) from Exiguobacterium sp. 255-15 at 1.78 \u00c5 Resolution", "Vibe Coding MicroPlastic Blaster", "Late-stage Crystal Failures and the Protein Crystal Harvesting Bottleneck", "Structure\u2013Function Analysis of the Extended Conformation of a Polyketide Synthase Module", "Abstracts of Marc C. Deller's publications", "Regrow Your Teeth: The End of Dental Implants", "Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2", "Unlocking Membrane Protein Structures: The Crystallization Revolution Reshaping Drug Discovery", "Tricyclic Compounds as Inhibitors of WRN", "2025 \u2013 Imidaz0[1,2-a]pyrimidine-5(1 h)-ones having a covalent warhead at the 3-position as wrn inhibitors", "Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3", "Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2", "GitHub README: BoltzMaker", "Regrow Your Teeth: The End of Dental Implants", "Tricyclic Compounds as Inhibitors of WRN", "The LockTac Revolution: Binding the Future of Therapeutics", "Plastic, Persistent, and Potentially Carcinogenic", "Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases", "JCSG Top96 Screen Formulation", "2015 \u2013 Protein stability: a crystallographer\u2019s perspective", "Extended Portfolio \u2014 Marc C. Deller, D.Phil.", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design", "Crystal structure and functional dissection of the cytostatic cytokine oncostatin M", "Marc C. Deller \u2014 LinkedIn profile summary", "Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design", "Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands", "Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices", "2025 \u2013 Building an Adaptability Mindset in the Age of AI", "Extended Portfolio \u2014 Marc C. Deller, D.Phil.", "GitHub README: CODSWALLOP", "Tricyclic Compounds as Inhibitors of WRN", "Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly", "The Rise of the Bio-Hybrid: Merging Proteins with Electronics for Next-Gen Diagnostics", "cif2plip \u2013 Protein-ligand interaction profiling using PLIP for co-folding outputs which produce CIF files", "Abstracts of Marc C. Deller's publications", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices", "Tricyclic Compounds as Inhibitors of WRN", "2024 \u2013 Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective CDK2 Inhibitors", "Tricyclic Compounds as Inhibitors of WRN", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "AnalyzeBoltz \u2013 Harvest the sprawling Boltz2 output tree, tidy summary, CSV, and CIF output folder", "Unlocking Molecular Potential: The Evolution of Topliss Tree and Bioisosteric Strategies in Modern Drug Discovery", "Microplastics and Cancer: The Mounting Evidence via Literature Review", "Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "2017 \u2013 The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase", "pants.mdeller.com: web app pages", "Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120", "GitHub README: CODSWALLOP", "Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor", "Abstracts of Marc C. Deller's publications", "Approaches to automated protein crystal harvesting", "Structure of the Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Region 412-423 Bound to the Broadly Neutralizing Antibody AP33", "alphabetti.mdeller.com: web app pages", "GitHub README: CODSWALLOP", "Microplastics and Health: Alarming New Research Reveals Widespread Biological Impact", "Abstracts of Marc C. Deller's publications", "GitHub README: CODSWALLOP", "GitHub README: AlphaFraud", "Molecular Velcro: Catching What We Cannot See", "The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase", "Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases", "GitHub README: TopPDBLX", "Protein stability: a crystallographer's perspective", "Decoding the Molecular Machinery Behind Plastic Degradation", "Revolutionizing Protein Structure Analysis: Cryo-EM without Purification Accelerates the Future of Structural Biology", "Cell surface receptors", "Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120", "Molecular Architecture of GLP-1 Receptor Drug Binding Reveals Dynamic Activation Mechanisms", "Revolutionizing Protein Crystallography: The Synergy of TELSAM-DARPin Fusion and Electric Field Optimization", "2025 \u2013 Embracing AI-Enabled Communication and Adaptability", "Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly", "Crystal structure of a soluble cleaved HIV-1 envelope trimer", "Self-Adapting Language Models: Are We Creating Our Final Invention?", "ChemSage: Vibe-Coding a Chemically Aware Drug-Discovery LLM on Local Hardware", "GitHub README: chatmol", "gobsmacked.mdeller.com: web app pages", "GitHub README: TopPDBLX", "Abstracts of Marc C. Deller's publications", "2016 \u2013 Key gp120 glycans pose roadblocks to the rapid development of VRC01-class antibodies in an HIV-1- infected Chinese donor", "Collaboration: 2+ state-of-the-art Structural Biology facilities built through global partnerships", "SiteDoctor \u2013 Full diagnosis, health, content, and SEO for any WordPress site", "Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3", "GitHub README: PANTS", "Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases", "Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases", "The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase", "deepcovalent: Partnerships \u2014 DeepCovalent", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly", "The Self-Driving Lab: How Automation and AI Are Redefining the Pace of Discovery", "mdeller.com: Marc Deller's web app server (landing page)", "Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor", "Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly", "Marc Deller's vibe-coding portfolio: AI-assisted scientific apps and tools", "Engineered Enzymes\u2014Design for Safe, Targeted Microplastic Therapeutics", "GitHub README: CODSWALLOP", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices", "Google Scholar profile of Marc C. Deller", "marcdeller.com: My Projects", "Microplastics Blaster \u2013 Deploy enzymes, destroy microplastics, save the body", "GitHub README: TopPDBLX", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices", "GitHub README: PANTS", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "Abstracts of Marc C. Deller's publications", "GitHub README: CODSWALLOP", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices", "Molecular Architecture of GLP-1 Receptor Drug Binding Reveals Dynamic Activation Mechanisms", "Building Proteins Brick by Brick: How AI Is Assembling Novel Bio-Machines", "Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases", "Tricyclic Compounds as Inhibitors of WRN", "Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor", "Top96 Crystal Screen", "Tricyclic Compounds as Inhibitors of WRN", "marcdeller.com: Vibe Coding", "From Single Targets to Disease Networks: A Systems Approach to Drug Discovery", "GitHub README: chatmol", "Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2", "GitHub README: cif2plip", "marcdeller.com: My Publications", "GPCR Binning \u2013 Structural binning and organisation utilities for GPCR datasets", "GitHub README: CODSWALLOP", "GitHub README: site-doctor", "HGMP01 and other Hidden Guardians: How the Human Gut Fights Back Against Plastic Pollution", "Cell surface receptors", "Approaches to automated protein crystal harvesting", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "2017 \u2013 Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design", "2012 \u2013 Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "GitHub README: TopPDBLX", "ORCID record of Marc C. Deller", "Crystal Structure of the Complex of 10E8 Fab Light Chain Mutant1 and T117v2 HIV-1 MPER Scaffold", "JCSG Top96: A Rationally Designed Crystallization Screen Based on over 1000 Deposited Protein Structures", "Revolutionizing Drug Discovery: The Click2Drug Platform and the Rise of AI-Powered Virtual Screening", "Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "JCSG Top96 Screen Formulation", "ORCID record of Marc C. Deller", "WordPressID \u2013 Every post and page from any WordPress site, to CSV and screen, in one command", "Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases", "Engineering Evolution: How FAST-PETase and Other Variants Are Transforming Plastic Biodegradation", "GitHub README: ChemSage", "GitHub README: FlexAppeal", "GitHub README: GOBSMACKED", "GitHub README: TopPDBLX", "When Protein Design Becomes a Security Problem", "Beyond Inhibition: The Rise of Conditional Degraders for Safer Medicines", "GitHub README: simple-zsh-script-to-run-boltz2", "Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120", "Molecular Architecture of GLP-1 Receptor Drug Binding Reveals Dynamic Activation Mechanisms", "Plastic, Persistent, and Potentially Carcinogenic", "Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor", "GitHub README: FlexAppeal", "marcdeller.com: Home", "2018 \u2013 Structure\u2013function analysis of the extended conformation of a polyketide synthase module", "Protein stability: a crystallographer's perspective", "About Marc C. Deller, D.Phil.", "Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor", "ChatRDKit \u2013 Talk to RDKit for Chemoinformatics in plain language", "PDB structures deposited by Marc C. Deller", "Crystal structure of a soluble cleaved HIV-1 envelope trimer", "GitHub README: microplastic-blaster", "Approaches to automated protein crystal harvesting", "Unlocking the Promise and Pitfalls of Large Language Models in Biomedical NLP", "alphabetti.mdeller.com: web app pages", "Approaches to automated protein crystal harvesting", "GitHub README: chatmol", "Generative AI for Protein Pocket Design: The Next Leap in Drug Discovery", "Crystal structure of a soluble cleaved HIV-1 envelope trimer", "BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis", "Future Therapeutics\u2014Engineering Solutions for Microplastic Burden", "alphafraud.mdeller.com: web app pages", "AlphaFraud \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved", "Unlocking the Power of Protein Engineering: FDA-Approved Enzyme Therapies Transform Rare Disease Treatment", "2025 \u2013 Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor", "The Memorization Trap: Have Protein-Ligand Cofolding Methods Transcended Simple Data Recall?", "Protein stability: a crystallographer's perspective", "GitHub README: SiteHealthInspector", "ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior", "Protein Hunter: Pioneering the Next Frontier of AI-Driven Protein Design", "GitHub README: AlphaFraud", "GitHub README: microplastic-blaster", "Crystal Structure of a PH Domain Containing Bacterial Protein (Exig_2160) from Exiguobacterium sibiricum 255-15 at 2.42 \u00c5 Resolution", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "2025 \u2013 Tricyclic compounds as inhibitors of wrn", "Cats Confuse Reasoning LLM: Query-Agnostic Adversarial Triggers for Reasoning Models", "GitHub README: FlexAppeal", "marcdeller.com: My Structures", "Precision Timing of Cytokine Signals through Facilitated Dissociation", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "Cell surface receptors", "2025 \u2013 Introduction to R: Basic R syntax", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "My Work Driving Innovation in Precision Oncology Through Structure-Based Drug Design", "2026 \u2013 From Bicyclic to Tricyclic: A New WRN Inhibitor Patent from Incyte", "GitHub README: ALPHABETTI", "Hidden Chemistry in the Protein Data Bank: What Else Are We Missing?", "2015 \u2013 Models of protein\u2013ligand crystal structures: trust, but verify", "Optimizing Data Volume and Diversity for Robust Antibody\u2013Antigen \u0394\u0394G Prediction", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "GitHub README: BoltzMaker", "BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis", "Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33", "Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly", "Revolutionizing Drug Discovery: LatentX and the Future of Protein Binder Design", "marcdeller.com: Artificial Intelligence", "The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase", "GitHub README: ButtFold", "2025 \u2013 Investing in Human Skills in the Age of AI", "2015 \u2013 Crystal structure of a two-subunit TrkA octameric gating ring assembly", "Tricyclic Compounds as Inhibitors of WRN", "BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline", "Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases", "ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior", "Cellular Uptake of Microplastics: Pathways, Persistence, and Therapeutic Potential for Intracellular Targeting of PETase", "Nanobodies: Revolutionizing Molecular Medicine Through Nature\u2019s Smallest Antibodies", "2026 \u2013 Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases", "GitHub README: ButtFold", "Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2", "Abstracts of Marc C. Deller's publications", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)", "Protein stability: a crystallographer's perspective", "Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3", "Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly", "2017 \u2013 Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices", "The Memorization Trap: Have Protein-Ligand Cofolding Methods Transcended Simple Data Recall?", "Abstracts of Marc C. Deller's publications", "gobsmacked.mdeller.com: web app pages", "SiteHealthInspector \u2013 A pure-Bash, colour-coded website health audit", "Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design", "Tricyclic Compounds as Inhibitors of WRN", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices", "When Perception Deceives Reality: The Hidden Flaws in AI Reasoning Evaluation", "GitHub README: BoltzMaker", "2022 \u2013 Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant FGFR2/3", "Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases", "BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis", "GitHub README: ALPHABETTI", "BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline", "GitHub README: CODSWALLOP", "GitHub README: BoltzMaker", "PETase Enzyme Therapy\u2014A New Frontier in Medical Microplastic Removal", "My Work Driving Innovation in Precision Oncology Through Structure-Based Drug Design", "Marc Deller's vibe-coding portfolio: AI-assisted scientific apps and tools", "2025 \u2013 Serverless Architectures on AWS", "Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor", "FGFR2 Kinase Domain in Complex with a Pyrazolo[1,5-a]pyrimidine Analog (Compound 29)", "Protein stability: a crystallographer's perspective", "Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor", "Crystal structure and functional dissection of the cytostatic cytokine oncostatin M", "ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior", "GitHub README: BoltzMaker", "Nanobodies: Revolutionizing Molecular Medicine Through Nature\u2019s Smallest Antibodies", "Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120", "Abstracts of Marc C. Deller's publications", "Courage: 400+ challenging structures determined, enabling breakthrough mechanistic insights", "GitHub README: analyze-boltz2-results", "GitHub README: BoltzMaker", "Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor", "marcdeller.com: Structural Biology", "Microplastics and Health: Alarming New Research Reveals Widespread Biological Impact", "pants.mdeller.com: web app pages", "Elevating Science Communication: Effortless 3D Structure Embedding with 3Dmol.js", "When the Wet Lab Becomes the Bottleneck: Self-Driving Platforms for Protein Engineering", "GitHub README: site-doctor", "Late-stage Crystal Failures and the Protein Crystal Harvesting Bottleneck", "Unlocking Documents for AI: The Power of Docling for AI-Driven Transformation", "Crystal Structure of the BG505 SOSIP gp140 HIV-1 Env Trimer in Complex with the Broadly Neutralizing Fab PGT122", "eloratherapeutics: Elora Therapeutics", "Capturing the Invisible: Understanding the \u00c5ngstr\u00f6m in Structural Biology", "PETase: Nature\u2019s Solution to a Human-Made Problem", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "Unveiling FragLites: Pioneering Precision in Drug Discovery", "2024 \u2013 Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective CDK2 Inhibitors", "Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor", "Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor", "Installation of RFdiffusion on MacOSX M1 architecture without GPU support", "Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor", "2025 \u2013 Bicyclic compounds as inhibitors of wrn", "Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands", "ChatRDKit \u2013 Talk to RDKit for Chemoinformatics in plain language", "Crystal Structure of PGT135 Fab in Complex with gp120 Core Protein from HIV-1 Strain JR-FL Bound to CD4 and 17b Fab", "GitHub README: BoltzMaker", "Protein stability: a crystallographer's perspective", "GitHub README: BoltzMaker", "GitHub README: PANTS", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "Communication: 60+ publications translating structural data into actionable insights", "Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120", "Universal Mounting Robot (UMR)", "GitHub README: TopPDBLX", "The 79th Organ: When Art and Biology Meet Bioengineering in the Fight Against Microplastics", "Cell surface receptors", "Engineering Protein Stability: Unlocking the Secrets of Successful Crystallization", "Unleashing the Power of AnythingLLM: Running Large Language Models Locally", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "Transforming Antibody Design: How Artificial Intelligence Revolutionizes Structure Prediction", "BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline", "Curiosity: 10+ years driving hands-on drug discovery through multidisciplinary approaches", "Clearing the Microscopic Threat: Can Plastic-Eating Enzymes Treat Liver Disease such as MASLD?", "GitHub README: PANTS", "Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3", "Outsmarting Evolution: Using AI to Overcome Drug Resistance", "Tricyclic Compounds as Inhibitors of WRN", "The Digital Twin in the Clinical Trial: How AI Is Making Drug Testing Smarter and Faster", "Crystal Structure of a Cellulose-Active Polysaccharide Monooxygenase from M. thermophila (MtPMO3*)", "Cell surface receptors", "An Enzyme That Eats Plastic \u2014 Inside Your Blood", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "pants.mdeller.com: web app pages", "ChatMol \u2013 Natural-language control of PyMOL, describe the scene, get the picture", "ChatPDB \u2013 A protein-structure-literate AI assistant for the Protein Data Bank: RAG for facts, a QLoRA-tuned model for behaviour", "ChemSage: Vibe-Coding a Chemically Aware Drug-Discovery LLM on Local Hardware", "AXIS: Closing the Loop on Crystal Detection", "cif2plip \u2013 Protein-ligand interaction profiling using PLIP for co-folding outputs which produce CIF files", "Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2", "Crystal structure and functional dissection of the cytostatic cytokine oncostatin M", "How to Confuse Machine Learning: Understanding Modern Adversarial Attacks", "GitHub README: ButtFold", "Democratizing Bioinformatics with Multi-Agent Language Models", "2017 \u2013 The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase", "Self-Encoded Libraries: A New Era in Drug Discovery Screening", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "Protein stability: a crystallographer's perspective", "The $40 Billion Revolution: How Bispecific Antibodies Are Transforming Cancer Care", "ChemSage: Vibe-Coding a Chemically Aware Drug-Discovery LLM on Local Hardware", "ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior", "Flow Cytometry and Microplastic Detection in Human Tissues", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2", "Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly", "Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design", "Beyond Trial and Error: AI-Driven Design of Peptide Therapeutics", "The New Architects: How Generative AI Is Designing Novel Protein Scaffolds", "Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases", "2025 \u2013 Amplify Your Communication Skills with Generative AI", "Boltz2Run \u2013 Every run logged, timed, and reproducible", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)", "From Waste to Wonder: How Bacteria Transform Plastic into Life-Saving Medicine", "buttfold.mdeller.com: web app pages", "From Waste to Wonder: How Bacteria Transform Plastic into Life-Saving Medicine", "GitHub README: ChemSage", "Unleashing Dynamic Precision: How DRGSCROLL Transforms Protein\u2013Ligand Docking", "BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline", "GitHub README: BoltzMaker", "AlphaFraud \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved", "GitHub README: ChemSage", "Therapeutic Enzymatic Compositions and Delivery Methods for the Degradation of Microplastics and Nanoplastics in Human Physiology", "2013 \u2013 Crystal structure of a soluble cleaved HIV-1 envelope trimer", "Unlocking the Power of Protein Engineering: FDA-Approved Enzyme Therapies Transform Rare Disease Treatment", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)", "Precision Oncology: Integrating AI and High-Throughput Screening to Fight Cancer", "Overcoming Biological Barriers: The Evolution of Oral Protein and Peptide Drug Delivery", "GitHub README: ALPHABETTI", "GitHub README: smiles2grid", "2014 \u2013 Approaches to automated protein crystal harvesting", "GitHub README: ALPHABETTI", "Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor", "GitHub README: ChemSage", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)", "GitHub README: TopPDBLX", "BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis", "GitHub README: BoltzMaker", "Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly", "GitHub README: TopPDBLX", "Discovery of an Orally Bioavailable KRAS G12D Inhibitor", "GitHub README: AlphaFraud", "The Signature Cowboy Hat of Adam Lee: A Blend of Tradition and Function", "2025 \u2013 Introduction to Artificial Intelligence", "Cellular Uptake of Microplastics: Pathways, Persistence, and Therapeutic Potential for Intracellular Targeting of PETase", "GitHub README: cif2plip", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "2013 \u2013 Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "AlphaFraud \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved", "Revolutionary AI Tool XenoBug Transforms Environmental Pollution Control Through Predictive Enzyme Discovery", "Tricyclic Compounds as Inhibitors of WRN", "From Bench to Regulation: How the FDA Is Formalising a Future Beyond Animal Testing", "The Unsung Hero: How a Single Chlorine Atom Transforms Drug Discovery", "Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor", "Tuning the Signal: Rational Design of G Protein Selectivity", "Discovery of (4-Pyrazolyl)-2-Aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2", "ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior", "GitHub README: ChemSage", "Crystal Structure of Acetyl Xylan Esterase (TM0077) from Thermotoga maritima at 2.40 \u00c5 Resolution", "Elora Therapeutics Investor Pitch Deck", "Publications of Marc C. Deller", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "HGMP01 and FAST-PETase: Decoding Two Paths to Plastic Degradation \u2013 Engineered versus Natural PETase Enzymes", "marcdeller.com: Structural Biology", "ChemSage: Vibe-Coding a Chemically Aware Drug-Discovery LLM on Local Hardware", "Smiles2Grid \u2013 Turn a JSON of SMILES into a paginated, property-annotated PDF grid with RDKit", "Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3", "BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis", "The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase", "GitHub README: GOBSMACKED", "GitHub README: BoltzMaker", "Vibe Coding MicroPlastic Blaster", "Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly", "Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor", "Creativity: 7+ patents through innovative protein engineering and AI integration", "Tricyclic Compounds as Inhibitors of WRN", "JAK2-V617F and the Evolution of Targeted Therapy: Precision Medicine Transforms Myeloproliferative Neoplasm Treatment", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)", "Elevating Science Communication: Effortless 3D Structure Embedding with 3Dmol.js", "Flow Cytometry and Microplastic Detection in Human Tissues", "GitHub README: GOBSMACKED", "How to Confuse Machine Learning: Understanding Modern Adversarial Attacks", "The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor", "Revolutionizing Protein Structure Determination: How MICA Combines Cryo-EM and AlphaFold3 for Unprecedented Accuracy", "2013 \u2013 Crystal structure of a soluble cleaved HIV-1 envelope trimer", "ChatMol \u2013 Natural-language control of PyMOL, describe the scene, get the picture", "Approaches to automated protein crystal harvesting", "GitHub README: ChemSage", "Revolutionary Bridge Between Prediction and Reality: How Elastic Networks Transform AlphaFold2\u2019s Structural Accuracy", "Abstracts of Marc C. Deller's publications", "JCSG Top96 Screen Formulation", "Structure\u2013Function Analysis of the Extended Conformation of a Polyketide Synthase Module", "Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)", "Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design", "AlphaFraud \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved", "PDB structures deposited by Marc C. Deller", "GitHub README: AlphaFraud", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior", "Abstracts of Marc C. Deller's publications", "BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis", "2026 \u2013 From Tool to Teammates: Working Smarter with AI", "Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof", "Crystal structure and functional dissection of the cytostatic cytokine oncostatin M", "GitHub README: CODSWALLOP", "Revolutionary Cryo-EM Breakthrough: Di-Gembodies Transform Small Protein Structure Determination", "marcdeller.com: My Patents", "GitHub README: AlphaFraud", "deepcovalent: DeepCovalent | Transforming Drug Discovery Today", "BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis", "Self-Encoded Libraries: A New Era in Drug Discovery Screening", "From Noise to Nanomolar: PXDesign\u2019s New Benchmark for Protein Binder Design", "JCSG Top96: A Rationally Designed Crystallization Screen Based on over 1000 Deposited Protein Structures", "Revolutionizing Drug Discovery: The Click2Drug Platform and the Rise of AI-Powered Virtual Screening", "Crystal structure of a soluble cleaved HIV-1 envelope trimer", "Tiny Pipes, Big Discoveries: How Microfluidics Are Accelerating Structural Biology", "GitHub README: BoltzMaker", "GitHub README: AlphaFraud", "Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design", "Crystal structure of a soluble cleaved HIV-1 envelope trimer", "Crystal structure of a soluble cleaved HIV-1 envelope trimer", "deepcovalent: Pipeline \u2014 DeepCovalent", "Revolutionizing Protein Crystallography: The Synergy of TELSAM-DARPin Fusion and Electric Field Optimization", "Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design", "MARCUS: Revolutionizing Chemical Literature Curation with AI-Driven Annotation and Recognition", "PocketTide \u2013 An interactive dashboard for GLP-1R structures", "Crystal Structure of [KS3][AT3] Didomain from Module 3 of 6-Deoxyerythronolide B Synthase in Complex with Antibody Fragment (Fab)", "Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3", "FoldForge \u2013 Browser GUI for the whole Boltz-2 workflow", "Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases", "GitHub README: AlphaFraud", "GitHub README: BoltzMaker", "GitHub README: CODSWALLOP", "SiteDoctor \u2013 Full diagnosis, health, content, and SEO for any WordPress site", "Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120", "Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2", "Cellular Uptake of Microplastics: Pathways, Persistence, and Therapeutic Potential for Intracellular Targeting of PETase"], "texts": ["Q: Do you know anything about me?\nA: I don't have any information about you. I'm chatMCD, trained only on material about Marc Deller: his career, science, companies and projects. I'm happy to answer questions in that area.", "Q: Which software did Marc use for data processing?\nA: Diffraction images were autoindexed, integrated and Lorentz-polarisation corrected in DENZO, then scaled and merged in SCALEPACK (with Bijvoet pairs preserved for MAD data). Files were converted with SCALEPACK2MTZ, intensities were converted to amplitudes with TRUNCATE, MAD data sets were scaled against the inflexion data with SCALEIT, and general manipulation used the CCP4 suite.", "Q: Where can I find Marc's article about the chlorine atom in drug discovery?\nA: It is published at https://marcdeller.com/magic-chloro-profound-effects-of-the-chlorine-atom-in-drug-discovery/ under the title \"The Unsung Hero: How a Single Chlorine Atom Transforms Drug Discovery\".", "Q: What is Marc's main conclusion in \"The New Architects\"?\nA: Marc concludes that design space has expanded faster than experimental capacity, so biopharma's rate-limiting step has shifted decisively back to the wet lab. He describes the new platforms not merely as better tools but as an industrial stack assembling around the Nobel-recognized foundations laid by the Baker lab. The post is at https://marcdeller.com/the-new-architects-how-generative-ai-is-designing-novel-protein-scaffolds/.", "Q: Why is molecular replacement relevant to ligand validation?\nA: Because protein\u2013ligand structures are usually solved without experimental phases, by placing an apo-protein structure using molecular replacement and then refining. The chapter explains that this makes the resulting 2mFo \u2212 DFc maps inherently biased, while mFo \u2212 DFc difference maps remain clear enough to indicate the presence and location of a genuine ligand.", "Q: How much data went into designing the Top96 screen?\nA: Over thirteen years the Joint Center for Structural Genomics crystallisation pipeline generated more than five million crystallisation experiments, over 130,000 crystals and 1,144 PDB depositions. Analysing that dataset revealed which reagent combinations were most likely to yield diffraction-quality crystals, and the Top96 was distilled from it.", "Q: What do you think about the stock market right now?\nA: Markets aren't something I follow, I'm afraid. I'm chatMCD, focused on Marc Deller's science and companies. If you're curious about the biotech side of things, though, I can tell you about Elora Therapeutics or DeepCovalent.", "Q: What organism name appears on the ShpA entry?\nA: The PDB entry lists the organism as Caulobacter vibrioides CB15, the current name for the strain more commonly known as Caulobacter crescentus CB15. The macromolecule is annotated as a histidine phosphotransferase.", "Q: How does CNTF relate to OSM in Marc's structure\u2013function comparison?\nA: Marc Deller compared OSM with CNTF, which signals through CNTFR, LIFR and gp130 in a 2:2:1:1 hexameric complex. CNTF's site 1 residues for CNTFR (Arg 25, Arg 28, Gln 63, Trp 64, Gln 74, Arg 177) map to OSM residues Asp 22, Thr 27, Asn 75 and Arg 182; CNTF's site 2 residues Lys 26 and Asp 30 map to OSM Leu 23 and Arg 29; and CNTF's site 3 pair Phe 152/Lys 155 is functionally equivalent to OSM Phe 160/Lys 163, although the two pairs are shifted by three residues structurally because OSM's helix D is longer.", "Q: What is Marc's design system?\nA: Navy #1C244B with accent blue #467FF7, the Poppins typeface, British English, no em dashes, no hype words like \"groundbreaking\" or \"revolutionary\", and matching flat navy-outlined cartoon icons with a Baloo 2 wordmark for every app.", "Q: What is Marc Deller's D.Phil. thesis about?\nA: Marc Deller's D.Phil. thesis, \"Structural Studies of Cytokines and Cytokine Receptors\", was submitted at the University of Oxford in Michaelmas Term 1999. It concerns structural studies of two members of the gp130 cytokine family, oncostatin M (OSM) and leukemia inhibitory factor receptor (LIFR). The work spans the rational engineering, expression, purification, characterisation and X-ray structure determination of a truncated OSM mutant, plus investigations of several eukaryotic expression systems for producing soluble LIFR.", "Q: What is smiles_grid?\nA: smiles_grid, in Marc Deller's smiles2grid repository, renders SMILES strings from a JSON file into a paginated, property-annotated PDF grid using RDKit. It writes a boxed 5 \u00d7 5 grid of 25 compounds per page plus a companion CSV of computed properties.", "Q: Does Marc program in R?\nA: Marc Deller completed a Coursera project certificate on 27 March 2025 for \"Introduction to R: Basic R syntax\". He notes that programming supports reproducible statistics and transparent data workflows, and that the skill helps with scripting, plotting and automating routine scientific analyses.", "Q: Are the WRN inhibitors in the patent covalent inhibitors?\nA: Yes \u2014 the exemplified compounds carry electrophilic \"warhead\" groups. In the worked examples the warhead R^W is either an acrylamide (\u2014NHC(O)CH=CH2) or a vinyl sulfonamide (\u2014NHS(O)2CH=CH2) on an aniline ring, and the specification lists a much broader set of possible warheads including alpha-halo alkyls, epoxides, aziridines, maleimides, squarates, quinones and cyanoalkynes.", "Q: How different was the final OSM structure from the initial homology model?\nA: The final refined structure deviated significantly from the initial Modeller-generated homology model, with an RMSD of 2.40 \u00c5 over 151 structurally equivalent C\u03b1 atoms. Marc Deller suggests this explains why molecular replacement with the homology model failed, although searches using the refined coordinates of human and murine LIF also failed.", "Q: What domains do the cyanobacterial endopeptidases contain?\nA: Each contains a bacterial SH3-like domain (SH3b) and the ubiquitous cell-wall-associated NlpC/P60, or CHAP, cysteine peptidase domain. The NlpC/P60 domain is a primitive, papain-like peptidase in the CA clan of cysteine peptidases with a Cys126/His176/His188 catalytic triad and a conserved catalytic core.", "Q: What does Marc charge for consulting?\nA: I don't have his rates. He does advise companies on structural biology, protein science and computational drug discovery, and he currently advises DeepCovalent alongside his CSO role at Elora. The way to explore that is to email him at marc@marcdeller.com.", "Q: What is Marc's 2015 paper on protein stability?\nA: \"Protein stability: a crystallographer's perspective\" is a paper by Deller and colleagues in Acta Crystallographica Section F that gives a practical, crystallographer-focused account of protein stability, which the post presents as essential for optimising expression, purification, formulation, storage and structural studies across biotechnology, pharmaceutical and academic settings.", "Q: How many crystals did the team have to screen to get the HIV-1 Env trimer structure?\nA: The 2013 Science paper records that crystal hits were obtained from roughly 10% of the crystallisation conditions tried, but that the majority of the crystals tested \u2014 more than 600 \u2014 diffracted only to 8 \u00c5 or worse at various synchrotron sources including APS, SSRL and the Canadian Light Source. This scale of screening reflects the robotic crystallisation and harvesting effort that Marc Deller and the Joint Center for Structural Genomics contributed to the project.", "Q: Where was this CDK2 research carried out?\nA: The CDK2 inhibitor research was carried out at the Incyte Research Institute, Incyte Corporation, in Wilmington, Delaware. All authors, including Marc Deller, were current or former employees of Incyte, and the research was funded by Incyte Corporation.", "Q: What structural features make IsPETase better than related cutinases?\nA: Marc points to an additional disulfide bond between Cys203 and Cys239 near the active site \u2014 mutating those cysteines to alanine drops the melting temperature from 46.8\u00b0C to 33.6\u00b0C \u2014 and an extended \u03b28-\u03b16 loop with three extra residues (Asn244, Ser245, Asn246) that creates a continuous binding groove. Homologous cutinases lack the extension and have a blocked cleft.", "Q: Which polymer was most toxic to the lungs in the studies Marc reviewed?\nA: Marc reports that in Danso et al. (2024, Toxicological Research), polystyrene fragments caused the most severe pulmonary inflammation via TLR4, NF-\u03baB and NLRP3 inflammasome activation, while polypropylene activated TLR2 without downstream signalling and polyethylene showed minimal effects. Polystyrene also had the highest negative zeta potential at -38.93 mV.", "Q: What are TopPDBLX's known limitations?\nA: Marc lists that the PDB contains only successes, so the data supports P(condition given crystallised) and says nothing about P(crystallised given sequence); that frequency reflects screen popularity; that about 80 % of percentage concentrations carry no w/v or v/v marker and inferred units are flagged; that only 2.2 % of entries name a cryoprotectant explicitly; and that depositor errors are reproduced rather than corrected, except for amounts that cannot be true at all.", "Q: What was Marc Deller's role in the UMR crystal harvesting project?\nA: Marc Deller was the first and presenting author of the 2014 poster and the JCSG contact for the project. He led the testing and optimization of the UMR at the JCSG, including its modification for harvesting microcrystals into SSRL grids and for hyperquenching, and co-authored the accompanying 2014 Acta Crystallographica F review with Bernhard Rupp on late-stage crystal failures.", "Q: How does AnalyzeBoltz derive a pIC50?\nA: It converts Boltz-2's `affinity_pred_value1` with the formula pIC50 = (6 \u2212 affinity_pred_value1) \u00d7 1.364. Marc is explicit that this is a heuristic approximation and that the values should be treated as relative rankings rather than absolute potency predictions.", "Q: What are Di-Gembodies?\nA: Di-Gembodies are covalently constrained nanobody dimers that act as rigid scaffolds for cryo-EM imaging of small proteins, covered by Marc Deller in \"Revolutionary Cryo-EM Breakthrough: Di-Gembodies Transform Small Protein Structure Determination\" on 11 October 2025. They were developed by researchers at the University of Oxford, the Rosalind Franklin Institute and Diamond Light Source.", "Q: Which gut bacteria are associated with PET degradation in Marc's post?\nA: Marc names Enterobacter, Holdemanella and Weizmannia as bacterial groups showing both elevated abundance and associated PET hydrolase activity, and notes they may adhere to microplastic surfaces and form biofilms that facilitate degradation, generating MHET and terephthalic acid as intermediates.", "Q: What has the AP33 work enabled for vaccines?\nA: Marc Deller writes that epitope-focused nanoparticle immunogens displaying the AP33 epitope have induced broadly neutralizing antibodies across HCV genotypes, including resistant genotype 2a strains, and that glycoengineered E2 variants improve homogeneity and immunogenicity. The post is at https://marcdeller.com/structure-of-hepatitis-c-virus-envelope-glycoprotein-e2-antigenic-site-412-to-423-in-complex-with-antibody-ap33/.", "Q: What are the final refinement statistics for the OSM P212121 structure?\nA: The P212121 \u0394188_OSM model was refined against all data from 20.0\u20132.2 \u00c5 to a crystallographic R-factor of 20.5 % and a free R of 26.1 % (27.1 % and 33.6 % in the 2.28\u20132.20 \u00c5 shell), with 1319 protein atoms and 116 waters. RMS deviations were 0.0147 \u00c5 for bond lengths and 1.7997\u00b0 for bond angles, and the average B-factor was 30.5 \u00c52.", "Q: Why did Marc write a whole blog post about the \u00e5ngstr\u00f6m?\nA: In \"Capturing the Invisible: Understanding the \u00c5ngstr\u00f6m in Structural Biology\", published 20 August 2025, Marc Deller argues that the \u00e5ngstr\u00f6m is the bedrock of atomic-scale measurement and that understanding it is essential to interpreting biomolecular form and function. His point is that atomic radii, bond lengths and protein domain separations naturally fall in the 1\u201310 \u00c5 range, so the unit eliminates cumbersome decimal notation.", "Q: What's the capital of France?\nA: That's outside my wheelhouse, I'm afraid. I'm chatMCD, and I'm here to talk about Marc Deller, his structural biology and drug-discovery career, Elora Therapeutics, DeepCovalent, and his personal projects. Anything in that territory I can help with?", "Q: What combination therapies are covered in the claims?\nA: Claims 93 to 98 cover administering the antibody together with a JAK inhibitor, a PI3K inhibitor or a standard-of-care therapy. Named agents are ruxolitinib and itacitinib as JAK inhibitors, parsaclisib as the PI3K inhibitor, and IFN-alpha, hydroxyurea, thalidomide, lenalidomide, an androgen, an erythropoietin-stimulating agent or a chemotherapeutic agent as standard-of-care options.", "Q: What settings does ChatMol expose?\nA: Settings live in `~/.chatmol/config.json` and are changed with `chatmol_config`: `host` (default http://localhost:11434), `model` (`qwen3.6:27b`), `temperature` (0.2), `timeout` (120 s), `auto` (True), `echo` (True), `debug` (False) and `max_history` (6 turns). `chatmol_reset` clears the rolling conversation history used for follow-up requests.", "Q: What were the stated goals of the Incyte program?\nA: The team sought a second-generation FGFR3 inhibitor with equal potency for the wild-type kinase and the gatekeeper mutants, greater than 20-fold selectivity over FGFR1 to avoid hyperphosphatemia, and a pharmacokinetic profile suitable for oral dosing. Equal potency for FGFR2 was also desirable, though it was not a formal goal.", "Q: What is an RCK domain?\nA: An RCK domain (Regulating the Conductance of K+) is a regulatory domain found in a large class of proteins that control eukaryotic and prokaryotic potassium transporters; more than half of known prokaryotic K+ channels are associated with RCK domains. A sub-family, including those in KefC, KtrA and TrkA, has homology to dinucleotide-binding proteins and is often referred to as KTN (K+ Transport, Nucleotide-binding) domains. Both RCK and KTN domains adopt a Rossmann-like \u03b2-\u03b1-\u03b2-\u03b1-\u03b2 fold, as Marc Deller's paper explains.", "Q: How tall is Marc Deller?\nA: I don't know Marc's height, that's not part of what I've learned about him. His publications, his patents and his structures I can do in detail, though. What would be useful? And marcdeller.com has the whole record if you want to browse.", "Q: What is the hepatitis C AP33 structure?\nA: Marc Deller's post \"2012 \u2013 Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33\" covers the Kong et al. structure in the Journal of Virology, in which the E2 epitope spanning residues 412 to 423 adopts a \u03b2-hairpin with a type I\u2032 \u03b2-turn inside the deep binding pocket of the AP33 paratope.", "Q: What microplastic exposure figures appear in the 79th organ post?\nA: Marc cites the estimate that humans ingest roughly one credit card's worth of plastic per week, that patients with microplastics in arterial plaques face a 4.5-fold higher risk of cardiovascular events, and that annual releases of 10\u201340 million metric tons are expected to double by 2040.", "Q: What vibe-coding apps does Marc Deller's r\u00e9sum\u00e9 mention?\nA: Marc Deller's September 2026 r\u00e9sum\u00e9 names ChemSage, BoltzMaker, FlexAppeal and AlphaFraud as examples of his vibe-coding work, describing it as rapid prototyping of computational biology and chemistry pipelines to fine-tune chemically aware LLMs and build SAR RAG-corpus datasets. His fuller portfolio of vibe-coded apps, including chatMCD itself, is covered elsewhere in his corpus.", "Q: Where is the polyketide synthase didomain post?\nA: The post \"Crystal Structure of [KS3][AT3] Didomain from Module 3 of 6-Deoxyerythronolide B Synthase in Complex with Antibody Fragment (Fab)\" is in the structures section of marcdeller.com. You can read it at https://marcdeller.com/crystal-structure-of-ks3at3-didomain-from-module-3-of-6-deoxyerythronolide-b-synthase-in-complex-with-antibody-fragment-fab/.", "Q: How much did structural genomics contribute to the PDB?\nA: According to Marc Deller's review, worldwide high-throughput crystallography and structural genomics efforts accounted for over 10% of all protein structures deposited in the PDB. For protein families with no previous structural coverage the figure was much higher, with about 50% of all novel structures coming from such centres.", "Q: What are the two domains of BT2081?\nA: BT2081 has an N-terminal domain adopting a beta-sandwich immunoglobulin-like fold and a larger C-terminal domain with a beta-sandwich jelly-roll fold, both similar to domains found in carbohydrate-active enzymes. Genome-context analysis supports involvement of the family in carbohydrate metabolism, which is central to B. thetaiotaomicron as a predominant symbiont of the human distal gut microbiome.", "Q: Where can I see Marc Deller's protein structures?\nA: Marc Deller's structures page is at https://marcdeller.com/structures/, where his PDB entries are collected alongside write-ups of individual top structures.", "Q: What is the smallest crystal usable for X-ray structure determination?\nA: Depending on unit-cell dimensions, the lower theoretical boundary is around 1.2 \u00b5m according to Marc Deller's review. Crystals of this size appear routinely during screening but are usually set aside as hits for further optimization rather than used directly for data collection.", "Q: What is Marc's argument in the \"Hidden Chemistry\" post?\nA: Marc argues that scientific knowledge is only as good as the tools used to acquire it, and that the NOS discovery is a call to action to revisit trusted datasets with fresh eyes and new technology. He suggests the next great breakthrough may not come from a new experiment but from uncovering what has already been collected and not yet seen. The post is at https://marcdeller.com/hidden-chemistry-in-the-protein-data-bank-what-else-are-we-missing/.", "Q: What did the Mre11 structure reveal about substrate discrimination?\nA: The 2.2 \u00c5 structure revealed both conserved and novel features relative to the archaeal enzyme, particularly in the DNA-specificity domain and the dimeric organisation. Marc Deller's post says these differences suggest that different organisms have evolved distinct mechanisms for discriminating substrates during DNA repair.", "Q: How many scientists has Marc Deller mentored?\nA: Marc Deller's LinkedIn profile states he has mentored and led 70+ scientists, fostering innovation and operational excellence across his career.", "Q: How does Marc explain the \"molecular bumper\" and \"molecular glue\" idea?\nA: Marc explains that SBI-553 binds within the receptor core and acts simultaneously as a bumper and a glue: it creates a steric block that prevents engagement of specific G proteins such as Gq, while stabilising an alternative open receptor conformation that is energetically favourable for others such as GoA. Selectivity comes from whether a given G protein's C-terminal tail can adapt to the modified pocket geometry, and the mechanism was resolved using cryo-electron microscopy combined with computational modelling.", "Q: Who supervised Marc Deller's D.Phil.?\nA: Marc Deller's D.Phil. was supervised by Professor E. Yvonne Jones and Professor John K. Heath. In his acknowledgments he also thanks Professor Louise Johnson for the opportunity to work in her laboratory, and Shinji Ikemizu and Jeronimo Bravo for introducing him to crystallography.", "Q: What does Marc report about sub-micron particles and cancer cell migration?\nA: Citing Brynzak-Schreiber et al. (Chemosphere, 2024) working with four human colorectal cancer cell lines, Marc reports that 0.25 \u00b5m particles significantly amplified cancer cell migration in a size-dependent manner, were inherited by daughter cells without elimination, and accumulated in the non-proliferating core of 3D tumour spheroids, so the burden is transmitted rather than cleared.", "Q: What is the Chat page on marcdeller.com for?\nA: The Chat page, headed \"Let's Chat: Proteins, Platforms, Programs\", invites visitors to ask Marc Deller questions directly rather than read a static resume. He writes that most bios tell you what someone has done, and this is a chance to ask.", "Q: Where is Marc Deller based?\nA: Marc Deller is based in Wilmington, Delaware, on the US East Coast, and his r\u00e9sum\u00e9 notes that he is willing to relocate. His most recent industry roles at Incyte Pharmaceuticals were also in Wilmington.", "Q: Who is PocketTide for?\nA: Marc aims it at structure-based drug design teams who need to triage GLP-1R structures quickly, compare binding modes and receptor states, and orient a project against the available structural evidence, all with no install and no build step. The repo is bellcheddar/PocketTide and the post is at https://marcdeller.com/an-interactive-dashboard-for-glp-1r-structures/.", "Q: What is Marc Deller's post on oral protein and peptide drug delivery about?\nA: In \"Overcoming Biological Barriers: The Evolution of Oral Protein and Peptide Drug Delivery\" (10 December 2025), Marc Deller reviews why protein drugs are so hard to give by mouth \u2014 stomach acid, pepsin, trypsin and chymotrypsin, the mucus layer and epithelial tight junctions \u2014 and the formulation, targeting and device strategies that have started to overcome those barriers.", "Q: Tell me about Marc Deller's work on potassium transporters at Scripps.\nA: While at the Joint Center for Structural Genomics at The Scripps Research Institute, Marc Deller led the structural characterization of the TM1088 Trk potassium-transporter gating ring from Thermotoga maritima, published in PLoS ONE in 2015. He determined the 3.45 \u00c5 crystal structure of the octameric two-subunit assembly and 1.5 \u00c5-range structures of its TM1088A and TM1088B components, showed that AMP binds TM1088A, characterized the three protein interfaces and the D2 symmetry of the ring, and proposed a gene duplication and fusion model linking it to single-subunit gating rings such as TrkA.", "Q: Which paper covers Marc Deller's work on FGFR inhibitors?\nA: The 2022 Journal of Medicinal Chemistry paper \"Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3\" describes FGFR2/3 inhibitors that retain potency against common gatekeeper mutants and are selective over FGFR1, culminating in compound 29. Marc Deller contributed structural biology at Incyte, and the work was motivated by cancers such as cholangiocarcinoma and bladder cancer.", "Q: What are the TrkA structures Marc deposited?\nA: 3JXO and 3L4B are both titled \"Crystal Structure of an Octomeric Two-Subunit TrkA K+ Channel Ring Gating Assembly, TM1088A:TM1088B, from Thermotoga maritima\" and were deposited by Deller, M.C., Johnson, H.A., Miller, M., Spraggon, G., Wilson, I.A., Lesley, S.A. and the JCSG. 3JXO was released on 6 October 2009 at 1.55 \u00c5 in space group C 2 2 21, and 3L4B on 16 February 2010 at 3.45 \u00c5 in space group P 32 2 1 with 1444 polymer residues and 161.86 kDa; both are keyworded TRANSPORT PROTEIN.", "Q: What are the biological actions of oncostatin M (OSM)?\nA: Marc Deller's thesis reports that OSM shares many activities with LIF, such as inducing differentiation of murine myeloid leukemic cells into macrophage-like cells, inhibiting differentiation of embryonic stem cells, and driving inflammatory responses like acute-phase protein synthesis in hepatocytes. OSM also has distinct activities including growth inhibition of solid tumours and human melanoma cells, mitogenic activity in aortic smooth muscle cells, and induction of spindle-cell morphology in AIDS-associated Kaposi's sarcoma cells.", "Q: What about the E655G escape mutation?\nA: An E655G mutation had been reported to help HCV escape AP33. Glu655 is found in 74% of known HCV isolates, and the structure showed that it is not directly involved in forming the AP33 epitope. Instead, the mutation enhances viral infectivity, which the authors suggested may compensate for the reduced infectivity caused by mutations within the conserved AP33 epitope itself.", "Q: What does Marc's post on flow cytometry and microplastic detection argue?\nA: In \"Flow Cytometry and Microplastic Detection in Human Tissues\" (4 November 2025), Marc Deller argues that flow cytometry, adapted from biology to environmental analysis, offers a faster and more sensitive alternative to spectroscopy for quantifying microplastics in human tissues, analysing thousands of particles per minute rather than taking hours.", "Q: What role did Marc play in the polyketide synthase work?\nA: Marc Deller contributed the structural biology and crystallography for the [KS3][AT3] didomain-Fab complex. The structure was released in May 2018, at the end of his period as Director and Founder of the Macromolecular Structure Knowledge Center at Stanford ChEM-H, Stanford University, a service organisation he founded in June 2015.", "Q: Which harvesting technologies does the UMR post cover?\nA: It covers microcapillaries, fine microtools, microgrippers, acoustic droplet ejection, optical tweezers and MEMS-based microrobots, and names specific semi-automated systems: the Universal Micromanipulation Robot (UMR) itself, the REACH microgripper platform, acoustic droplet ejection conveyor systems, and magnetic MEMS \"RodBots\".", "Q: Why does inhibiting WRN kill only some cancers?\nA: As the post explains, MSI-H/dMMR cancers, including a significant percentage of endometrial, colon and stomach tumours, are especially reliant on WRN activity for survival, so switching WRN off leaves their DNA unrepairable and triggers cell death. Tumours that are microsatellite-stable are left largely unharmed, which is what gives the target its selectivity.", "Q: What did the co-crystal structure in the patent show?\nA: The co-crystal structure of the WRN helicase core bound to a WRN inhibitor compound was solved to 3.1 \u00c5 resolution and revealed an unexpected conformation: a 150 degree rotation of the ATPase 2A domain relative to the ATPase 1A domain around the flexible linker. No previously published WRN helicase structure, including PDB entry 6YHR, displayed that inter-domain rotation.", "Q: How long does it take to train staff for such a facility?\nA: The post gives 6 to 12 months of intensive training to build skilled operators, alongside standardised protocols for quality assurance and reproducibility.", "Q: What does the Projects page on marcdeller.com cover?\nA: The My Projects page, subtitled \"Systems Builder: Projects That Define My Work\", says Marc Deller's projects span precision oncology, virology, immunology, and structural genomics, as well as mentoring a Stanford-led student program on novel antibiotics for resistant bacteria. The connecting thread is using protein structure as the starting point and letting biology dictate everything that follows.", "Q: How did PGT 135 get through the HIV glycan shield?\nA: PGT 135 engages gp120 mainly through a long 18-residue CDR H3 loop carrying a 10-amino-acid insertion after position 100 \u2014 the average human CDR H3 is 13 residues \u2014 and an extended CDR H1 loop with a rare 5-amino-acid insertion not present in the predicted germline sequence. Both loops protrude roughly 20 \u00c5 from the heavy-chain variable domain surface, comparable to the length of a high-mannose N-linked glycan, letting them reach past the glycan canopy to the protein below.", "Q: Which paper does Marc identify as the century's most-cited, and why does it matter?\nA: Marc identifies Microsoft's 2016 deep residual learning (ResNet) paper by He and colleagues, with over 254,000 Google Scholar citations. He explains that it solved signal degradation through many layers, enabling networks of 150 layers where 30 had been the standard, and became foundational for AlphaGo, AlphaFold and ChatGPT.", "Q: Why does GOBSMACKED superpose on pocket C\u03b1 atoms rather than the whole chain?\nA: Marc's explanation on the GOBSMACKED about page is that a predicted model can be excellent at the binding site and several \u00e5ngstr\u00f6ms out at a disordered terminus, so superposing whole chains spreads that error into the pocket and inflates the ligand RMSD, which is the one number the app exists to report honestly. Everything geometric is therefore measured after fitting the residues within 8 \u00c5 of the reference ligand, with the whole-chain TM-score reported separately as context.", "Q: Why are dark kinases so hard to study?\nA: Marc explains that many dark kinases show low expression levels, have catalytic domains that are difficult to work with, or lack apparent regulatory mechanisms, which defeats traditional biochemical and pharmacological approaches. That same obscurity is what makes them attractive, since they may sit in disease pathways that current therapies cannot reach.", "Q: What was destructive about previous microplastic detection methods?\nA: Marc summarises a scoping review of detection methods showing that Raman and infrared microspectroscopy, pyrolysis gas chromatography mass spectrometry and electron microscopy each require the tissue to be removed, digested or sectioned first, each carries a real risk of airborne plastic contamination during handling, and none of them can watch a particle move.", "Q: Has Marc published on crystal structure validation?\nA: Yes, his Publications page lists \"Models of protein\u2013ligand crystal structures: trust, but verify\" in both 2013 and 2015 versions, and \"Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands\" from 2017. He also published \"Protein stability: a crystallographer's perspective\" in 2015.", "Q: What can you type at the chatRDKit prompt?\nA: Three kinds of input. A SMILES string is loaded directly as a new molecule (`mol`, `mol2`, `mol3` \u2026) with no LLM call; Python/RDKit code is executed directly, with a bare expression's value printed REPL-style; anything else is treated as English, sent to Ollama and translated into RDKit code.", "Q: What does Marc do now?\nA: Since June 2025 Marc has been Co-Founder and Chief Scientific Officer of Elora Therapeutics, developing engineered PET-degrading enzymes as a first-in-class therapeutic to reduce microplastic burden in the human body. Since October 2025 he has also been a Scientific Technical Advisor to DeepCovalent, advising on physics-based CADD pipelines and GPCR-targeting discovery platforms.", "Q: Where is Marc's Anthropic Academy post?\nA: The post \"2026 \u2013 Anthropic Academy\" was published on 6 May 2026 at https://marcdeller.com/anthropic-academy/, with verification links for each of the four Skilljar course certificates.", "Q: Summarise what the three HCV and HIV structural papers from Marc's Scripps period have in common.\nA: All three \u2014 the 2012 Journal of Virology HCV E2\u2013AP33 structure, the 2013 Nature Structural & Molecular Biology gp120 glycan supersite paper and the 2013 Science cleaved HIV-1 Env trimer structure \u2014 come from the Ian Wilson laboratory and the Joint Center for Structural Genomics at The Scripps Research Institute, and all three use X-ray crystallography of a broadly neutralising antibody bound to a viral envelope glycoprotein to define a conserved site of vulnerability for structure-based vaccine design. Marc Deller is a middle co-author on all three, contributing the crystallisation screening, harvesting and data-collection work.", "Q: What is the PDB code for the structure in the 2022 FGFR paper?\nA: The co-crystal structure of compound 30 bound to FGFR2 that Marc Deller helped determine was deposited in the Protein Data Bank under accession code 8E1X. It is the only experimental structure reported in the paper; the earlier design work used a published FGFR3 crystal structure, PDB 4K33, as a docking template.", "Q: What did Marc Deller do at Scripps?\nA: From August 2006 to June 2015 Marc Deller was Target Manager at the Joint Center for Structural Genomics (JCSG) at Scripps Research Institute in San Diego, reporting to Prof. Ian Wilson, as part of the large-scale NIH-sponsored Protein Structure Initiative. He applied high-throughput protein production, crystallization and salvage techniques that increased NIH structural-genomics throughput by 10%, solved 70+ novel protein structures, and co-authored 40+ publications. His September 2026 r\u00e9sum\u00e9 notes that this structural genomics work provided part of the structural data foundations later used by AlphaFold.", "Q: Can you explain the povorcitinib paper in plain English?\nA: The immune system uses signalling relays called Janus kinases to pass messages from inflammatory molecules into cells. One of them, JAK1, drives inflammatory and autoimmune disease, while a close relative, JAK2, is needed to make blood cells, so a drug that blocks only JAK1 should reduce inflammation without causing anemia. The paper Marc Deller co-authored describes how Incyte chemists tuned a molecule step by step until they got povorcitinib, which is roughly 46 times better at blocking JAK1 than JAK2, is absorbed well when swallowed, and clears arthritis in rats.", "Q: What did Marc conclude about the character of the OSM receptor-binding interfaces?\nA: Marc Deller concluded that, unlike the predominantly hydrophobic Epo\u2013EpoR interface, the residues comprising sites 2 and 3 of OSM are a mixture of hydrophobic and polar residues, resembling the \"mosaic\" interfaces seen in the GCSF\u2013GCSFR and IL-4\u2013IL-4R complexes. He suggested that moving away from purely hydrophobic surfaces may allow tighter modulation of binding fidelity at the possible expense of affinity, which suits the \"pick and mix\" receptor usage of the gp130 cytokine family.", "Q: What does Marc Deller's post on NOS linkages as allosteric switches cover?\nA: \"Protein Structure and Redox Biology: NOS Linkages as Allosteric Switches\", published 11 August 2025, argues that newly discovered nitrogen-oxygen-sulfur linkages in proteins act as dynamic, reversible redox switches rather than mere structural supports. Marc describes how they form or break in response to cellular oxidative stress and can modulate a protein's activity from a distance.", "Q: Which work does Marc cite in the TELSAM-DARPin post?\nA: Marc cites the BYU ScholarsArchive thesis by Supeshala Dilrukshi Nawarathnage and colleagues on engineering protein linkers for TELSAM-DARPin fusion, the related PubMed paper showing that optimal TELSAM-target linker character is target-protein dependent, Ray and colleagues in the Journal of Physical Chemistry Letters on electric fields and protein phase behaviour, Hekstra and colleagues in Nature Communications on electric-field-stimulated protein mechanics, G\u00fcttler and colleagues on DARPins as viral entry inhibitors, and Anderson and colleagues in Open Biology on TELSAM-target fusion crystals with minimal lattice contacts.", "Q: What is unusual about oncostatin M's receptor usage?\nA: OSM is notable for using multiple receptor assemblies. Type I assemblies involve low-affinity binding to LIFR followed by high-affinity binding to gp130, forming a heterodimeric signal-transducing complex; OSM is the only gp130 cytokine that can bind gp130 alone, albeit with low affinity. Type II assemblies use an OSM-specific receptor \u03b2 subunit (OSMR\u03b2) with gp130. This dual receptor system underlies OSM's mix of LIF-like and OSM-specific activities.", "Q: Does Marc see AI biodesign purely as a risk?\nA: No. Marc stresses that the same capabilities are equally powerful defensively, citing Stuart, Bright and Horvitz in NAM Perspectives on treating AI biodesign as a strategic biosecurity asset able to produce diagnostic binders, novel antigens and neutralizing proteins against emerging pathogens in hours rather than months. He concludes that structure-aware screening, robust data governance and pre-deployment safety evaluation are not in conflict with scientific openness but are the conditions that make it sustainable. The post is at https://marcdeller.com/when-protein-design-becomes-a-security-problem/.", "Q: What did the 2000 \"Cell surface receptors\" review conclude?\nA: Marc Deller and Yvonne Jones wrote that the flood of new structures revealed new models for receptor cross-linking, demonstrated the potential stringency of extracellular requirements for initiating intracellular signalling, and highlighted unexpected interactions suggestive of higher order clustering.", "Q: What was Marc Deller's LinkedIn profile?\nA: Marc Deller's LinkedIn profile is at www.linkedin.com/in/marccdeller.", "Q: What did the review say about fibroblast growth factor receptor complexes?\nA: In the 2000 review, Marc Deller discussed the newly reported FGF2\u2013FGFR1 and FGF1\u2013FGFR2 structures, in which the primary complex is 1:1 with FGF binding between Ig-like domains two and three of the receptor. Unexpectedly, both studies showed further dimerisation of the 1:1 complex within the crystal, consistent with a 2:2 signalling complex, an observation the authors flagged as having considerable functional implications given that FGF signalling is believed to require receptor dimerisation with heparin or heparan sulfate proteoglycans.", "Q: Which JCSG entries carry the Pfam DUF1285 fold?\nA: Two: 2RA9, a DUF1285 family protein SBAL_2486 from Shewanella baltica OS155 at 1.40 \u00c5 in P 21 21 21 with sodium bound (released 16 October 2007), and 2RE3, a DUF1285 family protein SPO_0140 from Silicibacter pomeroyi DSS-3 at 2.50 \u00c5 in P 43 21 2 (9 October 2007). Both are cited in the Acta Crystallographica Section F 2010 paper on Pfam family PF06938.", "Q: How does Site Doctor get Lighthouse scores?\nA: The `--analytics` mode calls the public Google PageSpeed Insights API for real Performance, SEO, Accessibility and Best Practices scores plus Core Web Vitals, falling back automatically to a self-computed performance scorecard if PSI is unavailable. `--no-psi` skips the API call entirely.", "Q: What uses beyond small-molecule docking does Marc list for DRGSCROLL?\nA: He lists validating structural hypotheses at protein\u2013protein interfaces, modelling conformational selection in allosteric regulation, investigating enzyme mechanism through side-chain dynamics, improving local geometry for cryo-EM model fitting, and guiding site-directed mutagenesis by revealing energetically favourable rotamers. Refined complexes can also be exported to GROMACS or AMBER for molecular dynamics, or fed into free-energy perturbation pipelines.", "Q: How does Marc characterise ProteinMPNN and LigandMPNN?\nA: He calls ProteinMPNN the gold standard for rapid sequence design, a message-passing network that converts fixed backbones into optimised sequences in seconds, with NVIDIA NIM integration and immune-evasive variants tuned by direct preference optimisation. LigandMPNN extends it to protein\u2013ligand interfaces by encoding atomic element types and spatial context, with Marc citing up to 100-fold improvements in small-molecule and DNA binding accuracy validated by X-ray crystallography.", "Q: How much PET is produced annually, according to Marc?\nA: Marc states that over 70 million tons of PET are manufactured annually, and notes that the dominant thermomechanical recycling route degrades material quality with each cycle, which is why a fundamentally different approach became urgent.", "Q: Who are the founders of Elora Therapeutics?\nA: The pitch deck names two co-founders: Paul Swartzendruber, Co-Founder & CEO, and Marc Deller, D.Phil, Co-Founder & CSO. Elora Therapeutics is partnered with the Austin Technology Incubator (ATI).", "Q: What is Marc's worked example of ALPHABETTI failing?\nA: Marc uses avGFP as the worked failure in ALPHABETTI: masked marginals recover the wild-type residue at 10 % of positions against 8 % for a randomly shuffled version of the same sequence, mean information content is 0.27 of a possible 4.322 bits, and mean pLDDT is about 43. It is a famous, heavily studied protein the method has almost nothing useful to say about, and nothing on screen would reveal that except the confidence number.", "Q: What kinds of problems do Marc's vibe-coded tools solve?\nA: The Vibe Coding page gives examples such as harvesting a Boltz-2 output tree, profiling protein-ligand contacts, and wrangling SMILES into a tidy grid. Every script, he says, answers a genuine need from his structural biology and drug discovery workflows.", "Q: Which generative AI courses has Marc taken?\nA: The My Learning page lists 2025 certificates in \"Strengthen Your Decision-Making with Generative AI\", \"Amplify Your Communication Skills with Generative AI\", \"Building an Adaptability Mindset in the Age of AI\", \"Investing in Human Skills in the Age of AI\", and \"Introduction to Artificial Intelligence\".", "Q: What HIV antibody structure is PDB 5CD5?\nA: PDB entry 5CD5 is the crystal structure of the immature VRC01-class antibody DRVIA7, isolated from a Chinese donor, bound to the clade A/E HIV-1 gp120 core from strain 93TH057. Marc Deller's post records an X-ray structure at 3.396 \u00c5 released on 6 April 2016, containing the gp120 core plus the DRVIA7 Fab heavy and light chains and N-acetylglucosamine glycans.", "Q: Explain how neural networks work.\nA: That's general background rather than anything about Marc, so I'll leave it to someone better placed. What I can tell you is how Marc uses machine learning in his own work: protein structure prediction, generative design, and the small language models he fine-tunes himself. Interested?", "Q: Which paper accompanies the 5UFV structure?\nA: The 5UFV structure was published in Span, E.A., et al. (2017), \"The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\", ACS Chemical Biology, volume 12, pages 1095-1103. Marc Deller also has a separate blog post on marcdeller.com discussing that paper.", "Q: What does \"From Waste to Wonder: How Bacteria Transform Plastic into Life-Saving Medicine\" describe?\nA: In this 11 October 2025 post Marc Deller describes work from Professor Stephen Wallace's lab at the University of Edinburgh in which genetically engineered *Escherichia coli* convert PET plastic waste into paracetamol at room temperature, the first documented case of producing a pharmaceutical from plastic waste.", "Q: Where is the Exig_1061 post on Marc's website?\nA: The post \"Crystal Structure of a Putative NTP Pyrophosphohydrolase (Exig_1061) from Exiguobacterium sp. 255-15 at 1.78 \u00c5 Resolution\" is in the structures section of marcdeller.com. You can read it at https://marcdeller.com/crystal-structure-of-a-putative-ntp-pyrophosphohydrolase-exig_1061-from-exiguobacterium-sp-255-15-at-1-78-a-resolution/.", "Q: What robotic approaches does the harvesting review cover?\nA: It covers the Universal Micromanipulation Robot, a six-axis anthropomorphic arm with submicron precision that produced a 1.5 \u00c5 structure of bovine trypsin matching manually harvested quality, and REACH microgrippers, two-fingered MEMS devices made from biocompatible SU-8 that apply controlled pressure with low X-ray background so data can be collected in the gripper.", "Q: What was the headline structural result of the thesis?\nA: The headline result was the determination of the atomic X-ray structure of a truncated oncostatin M mutant in the space group P2\u20812\u20812\u2081 to a resolution of 2.2 \u00c5, solved using multiple wavelength anomalous dispersion (MAD). Wildtype OSM had proved recalcitrant to crystallisation, so Marc engineered a C-terminally truncated variant to obtain diffraction-quality crystals.", "Q: Where is the source code for Marc's web apps?\nA: Every app on mdeller.com links to its GitHub repository under the bellcheddar account: github.com/bellcheddar/GOBSMACKED, ALPHABETTI, ButtFold, CODSWALLOP, TopPDBLX, FlexAppeal, AlphaFraud, BoltzMaker, chatPDB, ChemSage, microplastic-blaster and PANTS.", "Q: Who are the inventors listed on the mutant-CALR antibody patent?\nA: The twelve named inventors, all at Incyte Corporation in Wilmington, Delaware, are Patrick Mayes, Horacio G. Nastri, Rebecca A. Buonpane, Edimara S. Reis, Shaun M. Stewart, Jing Zhou, Marc C. Deller, Hsiang-Ting Lei, Leslie Brooke Epling McQuirter, Yonghong Zhao, Hamza Celik and Brittney Melissa Wass.", "Q: What is CotA-Laccase and why is it included?\nA: CotA-Laccase is a bacterial-origin oxidative enzyme included in Marc Deller's patent to degrade polyolefin plastics\u2014polyethylene (PE), polypropylene (PP), and polystyrene (PS)\u2014that PETase and MHETase cannot touch. The patent describes engineered CotA-Laccase variants with enhanced oxidative activity, improved substrate specificity, thermostability, and modified redox potential for use in the complex physiological environment of the body.", "Q: What results does Marc highlight for the barcode-free screening platform?\nA: Marc highlights screening of libraries of more than 500,000 small molecules in single experiments at 100 femtomoles per compound, with detection limits suggesting scalability toward five million members within solubility limits. Validation against carbonic anhydrase IX gave multiple nanomolar binders with enrichment factors comparable to phage display, and the platform succeeded where DNA-encoded approaches cannot by finding potent inhibitors of flap endonuclease 1 (FEN1), a DNA-binding enzyme. Libraries can be built in under a week with standard organic chemistry, which Marc says democratises large-scale screening beyond well-funded pharmaceutical companies.", "Q: Where was Marc working when the BtDyP structure was determined?\nA: The BtDyP structure was released in 2006 and published in 2007, when Marc Deller was Target Manager at the Joint Center for Structural Genomics (JCSG) at The Scripps Research Institute in La Jolla, California, a post he held from August 2006 to June 2015. It is one of the many structural genomics structures he contributed to there.", "Q: How does the UMR perform hyperquenching?\nA: The UMR uses a directly coupled pneumatic hyperquencher that plunges crystals into liquid nitrogen at high speed (5 m/s), operating in the nucleate boiling regime rather than the film boiling regime. A high-speed camera running at 1000 frames per second is used to calculate the hyperquencher speed, and a high-rate data acquisition thermocouple monitors cooling rates precisely.", "Q: How many emails does Marc get a day?\nA: No idea! Not something covered by what I know. If you want to add to them, marc@marcdeller.com.", "Q: Which of Marc's recent structures is the highest resolution?\nA: Among the Incyte-era drug discovery entries, 9E5F is the sharpest at 1.35 \u00c5, closely followed by 9E5D at 1.36 \u00c5, both KRAS G12D structures released in January 2025. The JAK1 structure 10PI follows at 1.54 \u00c5 and the CDK2 structure 8UV0 at 1.55 \u00c5.", "Q: What is the C-terminal domain of Jann_2411?\nA: The C-terminal domain of Jann_2411 forms a treble-clef zinc finger that is likely to be involved in DNA binding, while the N-terminal ABATE domain may bind an as yet unknown ligand. Analysis of Jann_2411 and the broader ABATE-domain family suggested a role as stress-induced transcriptional regulators.", "Q: What is PDB entry 3CM1?\nA: 3CM1 is an SsgA-like sporulation-specific cell division protein, YP_290167.1, from Thermobifida fusca YX-ER1 at 2.60 \u00c5 in space group P 41, with 417 polymer residues and 46.25 kDa, released on 1 April 2008 and keyworded CELL CYCLE. Its deposition authors are the JCSG together with Chruszcz, M., Minor, W. and Wang, S., and it is cited in the 2009 Journal of Biological Chemistry paper on SsgB.", "Q: What is the RSRZ value?\nA: RSRZ is the PDB's normalized version of the Real Space R-Value, which is otherwise somewhat scale-dependent. The chapter explains that RSRZ quantifies how many standard deviations a particular residue's electron density fit deviates from the mean.", "Q: What problems does run_boltz.zsh prevent?\nA: Marc lists overwritten logs, stray macOS dot-files derailing a batch, and no record of how long a run took or what flags were used. The script timestamps every log as `boltz_run_YYYYMMDD_HHMMSS.log`, tracks wall-clock runtime, sanitises the input directory of hidden files such as `.DS_Store` and `._*`, and tees output to both screen and file.", "Q: Which paper reported the Bifidobacterium phosphoesterase structure?\nA: The structure was reported in Han, G.W., et al. (2011), \"Crystal Structure of a Metal-Dependent Phosphoesterase (YP_910028.1) from Bifidobacterium adolescentis\", Proteins, volume 79, pages 2146-2160. Marc Deller contributed to that structural genomics work.", "Q: What is Marc's conclusion about vibe coding after building the game?\nA: Marc concludes that vibe coding is fast for building demos, wireframing and planning, but that tweaking, polishing and adding final touches are more efficiently achieved by digging into the code directly. He also notes that defining a strict modular architecture from the start made rapid LLM-driven iteration much easier.", "Q: Why is AI protein engineering needed for bio-hybrid sensors, according to Marc?\nA: Marc says AI-driven design is essential for creating proteins that are highly specific for their target, remain stable when attached to an electronic surface, and induce a strong signal on binding \u2014 a combination that is hard to achieve otherwise.", "Q: What is loaded into the chatRDKit session namespace?\nA: `Chem`, `AllChem`, `Draw`, `Descriptors`, `Lipinski`, `Crippen`, `QED`, `DataStructs`, `rdFingerprintGenerator`, `rdMolDescriptors`, `rdMolDraw2D`, `math` and the `render()` helper. Every molecule created persists in that namespace for the rest of the session.", "Q: What fold does the BtDyP peroxidase have?\nA: Marc Deller's post on PDB 2GVK reports that BtDyP adopts a novel two-domain \u03b1+\u03b2 ferredoxin-like fold with a conserved heme-binding motif, quite unlike classical peroxidases. The hexameric assembly seen in the 1.6 \u00c5 structure established the structural foundations for this new family of heme-dependent peroxidases.", "Q: Where is Marc's post on the bacterial PH domain protein?\nA: The post \"Crystal Structure of a PH Domain Containing Bacterial Protein (Exig_2160) from Exiguobacterium sibiricum 255-15 at 2.42 \u00c5 Resolution\" sits in the structures section of marcdeller.com. You can read it at https://marcdeller.com/crystal-structure-of-a-ph-domain-containing-bacterial-protein-exig_2160-from-exiguobacterium-sibiricum-255-15-at-2-42-a-resolution/.", "Q: When did the DRVIA7 antibody lineage begin?\nA: Longitudinal lineage tracing placed the \"birth date\" of the DRVIA7 lineage shortly before 2006. Functional heavy-chain precursors were present in 2006, the lineage expanded markedly by 2008, and it began to decline in 2009 as the virus escaped it.", "Q: What limitations does Marc flag for self-driving protein engineering labs?\nA: Marc flags well-to-well variability, instrument API integration, and immunogenicity assays that are difficult to miniaturise as continuing frictions on end-to-end autonomous antibody discovery. Even so, he sees the trajectory as clear: once the lab itself becomes a programmable component of the model, the question shifts from how to design a million antibodies to which million to measure next. The post is at https://marcdeller.com/when-the-wet-lab-becomes-the-bottleneck-self-driving-platforms-for-antibody-engineering/.", "Q: What was Marc Deller's role on the crystal harvesting review?\nA: Marc Deller was first author and one of two corresponding authors, writing it with Bernhard Rupp. At the time he was at the Joint Center for Structural Genomics at The Scripps Research Institute in La Jolla, California.", "Q: What does Marc say about TurboPETase and LCC variants?\nA: Marc reports that TurboPETase, derived from BhrPETase with eight mutations including W104L and F243T, achieved 98.2% depolymerization of postconsumer PET at 200 g/kg loading within 8 hours, with the Trp104Leu change increasing turnover by 71%. He also notes LCC-ICCG (F243I/D238C/S283C/Y127G) operates at 65\u201372\u00b0C, and the LCC S101N/F243T double mutant fully depolymerized 1.3 g of untreated postconsumer PET in three days at 55\u00b0C using only 1.25 mg of enzyme.", "Q: What mechanistic conclusion did the MtPMO3* paper reach?\nA: The combination of decoupled O2 consumption and superoxide release in H161 variants suggested that a Cu(II)\u2013superoxo intermediate is either unstable without H161 or is not the species that abstracts the substrate hydrogen atom. This supports a cytochrome P450-like mechanism in which the O\u2013O bond is first cleaved to generate a copper-oxyl radical that performs hydrogen atom abstraction. The newly defined roles of H161 and Q167 together support a Cu\u2013oxyl active intermediate.", "Q: What is the TrkA gating ring paper?\nA: \"2015 \u2013 Crystal structure of a two-subunit TrkA octameric gating ring assembly\" covers Deller et al. in PLoS ONE, reporting the structure of the TM1088 potassium transport gating ring from the thermophile Thermotoga maritima, in which two distinct proteins, TM1088A and TM1088B, combine as four copies each into a 160 kDa octameric assembly.", "Q: Why did the team switch from an imidazole core to an indole core?\nA: Compound 5 had excellent binding potency of 20 pM in surface plasmon resonance but translated poorly into cellular and whole blood settings, which the team attributed to excessive polarity. Replacing a ring nitrogen with a C\u2013H by moving to the related tricyclic indole core reduced that polarity and gave a subtle but real improvement in cellular translation, so the indole was carried forward for all subsequent SAR.", "Q: How did the new cleaved trimer structure compare with the uncleaved trimer structures that had been published?\nA: The 2013 Science structure contrasted markedly with a recently described uncleaved, membrane-bound JR-FL Env trimer determined by cryo-EM. The soluble cleaved crystal structure differed substantially even in the gp120 core and in the arrangement of the gp41 helices, and did not contain the large hole reported for the uncleaved trimer. The authors noted that it remained to be determined whether these differences reflected the different trimer forms or concerns about the cryo-EM methodology used for the uncleaved structure.", "Q: What were the aims of Marc Deller's thesis?\nA: The thesis aimed to delineate structural features of two key haemopoietin cytokine family members, OSM and LIFR. It set out to determine the X-ray structure of a rationally designed OSM mutant to provide a framework for interpreting mutagenesis data on receptor engagement, and to establish expression systems for producing and characterising LIFR so that its site 3 interactions could eventually be analysed structurally.", "Q: What is the ABATE domain?\nA: The ABATE domain is a new fold discovered in the N-terminal domain of Jann_2411 from Jannaschia sp. strain CCS1, solved to 1.45 \u00c5 resolution by multiple-wavelength anomalous dispersion. Marc Deller co-authored the 2010 Acta Crystallographica Section F paper, which described the first structural representative of the DUF1470 family, Pfam PF07336.", "Q: Which are Marc Deller's ten most-cited publications?\nA: By Google Scholar citation count they are the 2013 Science HIV-1 envelope trimer structure (1,027), the 2013 Nature Structural and Molecular Biology gp120 supersite paper (443), \"Protein stability: a crystallographer's perspective\" in 2016 (440), the 2017 ACS Chemical Biology fungal polysaccharide monooxygenase paper (146), \"Models of protein-ligand crystal structures: trust, but verify\" in 2015 (130), the 2007 dye-decolorizing peroxidases paper in Proteins (108), the 2012 hepatitis C virus E2/AP33 structure in Journal of Virology (103), the 2009 murein peptide endopeptidase paper in Structure (99), the 2016 Immunity paper on gp120 glycans and VRC01-class antibodies (98), and the 2010 YkfC endopeptidase paper in Acta Crystallographica Section F (97).", "Q: Where can I find Elora Therapeutics online?\nA: Elora Therapeutics' website is www.eloratherapeutics.com, headed \"Plastic Meets Biology\". It is one of the two company sites Marc Deller links from his own biography, the other being DeepCovalent at www.deepcovalent.com.", "Q: What do Chapters 3 and 4 of Marc Deller's thesis cover?\nA: Chapter 3 of Marc Deller's D.Phil. thesis covers the expression, purification and characterisation of oncostatin M (OSM), including production of a biologically functional truncated mutant suitable for crystallography. Chapter 4 covers the expression, purification, deglycosylation and preliminary characterisation of the leukemia inhibitory factor receptor (LIFR), and the LIFR-LIF receptor-ligand complex, using several eukaryotic expression systems.", "Q: How do pixel-array detectors change crystal harvesting?\nA: Marc Deller's review explains that the millisecond readout of pixel-array detectors allows shutterless data collection with continuous rotation, which lets robot arms serve as goniostats and removes separate harvesting and pin-mounting steps, and allows data collection from streams of microcrystal slurry. Their sensitivity also permits weaker exposures, making it nearly possible to outrun radiation damage at room temperature.", "Q: How were the antibodies discovered?\nA: Two routes are described. Mice were immunised with plasmids encoding human MPL and human Type 1 mutCALR, plasmablasts were isolated from spleen and lymph nodes, and paired VH/VL sequences were obtained by 10x Genomics B-cell sequencing, giving clone 54. Separately, single-donor and multi-donor phage display libraries were enriched on biotinylated MBP-mutCalR Type 1, the best scFv pools were recombined into yeast display libraries and selected over four rounds, and unique clones were reformatted as human IgG1 and expressed in Expi293F cells.", "Q: What organism is listed for the Exig_2160 structure?\nA: The PDB entry for Exig_2160 lists the organism as Exiguobacterium artemiae 255-15, the strain also referred to as Exiguobacterium sibiricum 255-15, a psychrotolerant bacterium isolated from Siberian permafrost. The deposited macromolecule is annotated simply as an uncharacterised protein.", "Q: What was Marc's contribution to the FGFR2/3 work?\nA: As with the other Incyte drug-discovery papers he features, Marc Deller contributed the structural biology and crystallography from the Incyte structural biology group; the post highlights the FGFR2 kinase domain in complex with a pyrazolo[1,5-a]pyrimidine analog among the associated structures.", "Q: Which prior patents does the invention distinguish itself from?\nA: The invention distinguishes itself from a range of industrial and non-therapeutic patents, including WO2022/076380A2 (engineered PETase, University of Texas System), US 9,840,726 and US 10,087,582 (CotA-Laccase for biomass and paper pulping, MetGen Oy), WO2020/113962 (CotA-Laccase for mycotoxin degradation), US 9,540,617 (recombinant HRP for fungal expression, Meiji Seika Pharma), EP 2768953 (engineered HRP for wastewater, Graz University of Technology), and US 20240263155 A1 (industrial PET hydrolase, Hubei University). None of these address therapeutic degradation of microplastics within human physiology.", "Q: Why does Marc's review say protein stability is a prerequisite for crystallization?\nA: Crystallization is the self-organization of macromolecules into a translationally periodic arrangement with long-range order, so the molecules in each asymmetric unit must be of the same kind and the same shape. Marc Deller's review argues that if a protein cannot form such stable entities, a fundamental requirement is unmet and no search for thermodynamic or kinetic conditions will produce crystals.", "Q: Which recent organ discoveries does Marc use to set up the 79th organ idea?\nA: Marc points to the mesentery's reclassification as the 78th organ in 2017, the identification of the interstitium as a possible new organ system in 2018, and the discovery of the tubarial salivary glands in 2020, which he describes as the first new organ discovery in nearly 300 years.", "Q: How was the quality of the native P212121 data?\nA: The native P212121 data collected at the ESRF were of good quality: 43,901 measured reflections reduced to 10,513 unique (redundancy 4.2), with 96.5 % overall completeness (98.2 % in the 2.25\u20132.20 \u00c5 shell), an Rmerge of 8.3 % and I/\u03c3(I) of 13.2 overall and 4.3 in the outer shell.", "Q: What holds the TrkA octamer together?\nA: The assembly achieves stability through three major interfaces: TM1088A homodimerization, TM1088B homodimerization and the TM1088A\u2013TM1088B heteromeric interface, which alone buries nearly 18,000 square angstroms of surface area. An interfacial network involving Arg33, Glu34 and Glu37 forms salt bridges across the central cavity that likely support thermostability.", "Q: How did Marc build MicroPlastic Blaster?\nA: Marc built it by vibe coding \u2014 the AI-assisted development paradigm coined by Andrej Karpathy in February 2025 \u2014 using natural language prompts rather than line-by-line programming. The stack was deliberately minimal: vanilla JavaScript and HTML5 Canvas packed into a single .html file with no external dependencies, chosen partly because JavaScript game code is the most richly represented genre in LLM training data.", "Q: Which journal published the DyP peroxidase structures?\nA: The work was published as Zubieta, C., et al. (2007), \"Crystal Structures of Two Novel Dye-Decolorizing Peroxidases Reveal a Beta-Barrel Fold with a Conserved Heme-Binding Motif\", in Proteins, volume 69, pages 223-233. Marc Deller lists it alongside PDB entry 2GVK on his website.", "Q: What was unusual about helix A in the oncostatin M structure?\nA: Marc Deller's structure showed that helix A of oncostatin M has a pronounced kink, caused by residues Gln25 and Leu30 adopting a hydrogen-bonded turn conformation with four tightly bound water molecules substituting for the normal helical hydrogen bonds, and residues Thr27\u2013Ile37 forming a 310 helix. Comparison across the family revealed that this kinked helix A is a feature of all gp130 cytokines that bind LIFR (OSM, LIF and CNTF), whereas IL-6, which does not bind LIFR, has a regular helix A, suggesting the kink is a structural feature of LIFR recognition.", "Q: What is KPN03535?\nA: KPN03535, gi|152972051, is a putative lipoprotein of unknown function secreted by Klebsiella pneumoniae MGH 78578. Marc Deller co-authored the 2010 Acta Crystallographica Section F paper reporting its crystal structure, which revealed a novel variant of the OB-fold despite no detectable sequence similarity to known structures.", "Q: What resolutions were achieved in the purification-free cryo-EM work Marc describes?\nA: Marc reports 3D reconstructions at 2.8 \u00c5 for Artemia ferritin and 2.5 \u00c5 for the plant protein PDX1.2 straight from unpurified lysate, matching or surpassing traditional purified samples. Rapid post-expression clean-up steps such as spin filtering, benzonase digestion or his-tag affinity capture pushed maps as fine as 2.1 \u00c5 for only about an hour of extra preparation.", "Q: What does Marc's post \"Microplastics and Health: Alarming New Research Reveals Widespread Biological Impact\" review?\nA: Published on 17 August 2025, this post by Marc Deller reviews six studies on the biological impact of microplastics across mammalian and aquatic systems, covering intestinal, hepatic, cardiovascular, respiratory and systemic effects, and establishing dose-dependent relationships between exposure and organ damage.", "Q: Does the patent cover a treatment method, not just the enzymes?\nA: Yes; beyond the enzyme compositions and formulations, claim 5 covers a method for therapeutic degradation of microplastics and nanoplastics within the human body, comprising administering the enzymatic composition and ensuring safe metabolic clearance of the degradation products through hepatic and renal pathways. So Marc Deller's filing protects both the compositions and their therapeutic use.", "Q: Which mouse model does Marc discuss for MASLD and microplastics?\nA: Marc discusses the STZ+HFD mouse model from Jeong et al. (Nature Communications, 2024), which recapitulates the full spectrum from steatosis to hepatocellular carcinoma. He notes that polystyrene nanoplastics on a high-fat diet exacerbated lipid accumulation, hepatocyte apoptosis and fibrosis, accumulating in hepatocytes and triggering Hedgehog signalling that promotes stellate cell activation.", "Q: Where was the FGFR2/3 inhibitor work published?\nA: It was published as Shvartsbart, A., et al. (2022), \"Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3\", Journal of Medicinal Chemistry, volume 65, pages 15433-15442. Marc Deller has a companion blog post about the paper.", "Q: What are BoltzMaker's pipeline stages?\nA: `boltz_input.md` feeds `generate` (per-target YAMLs plus a manifest), then `preflight` (PASS/WARN/FAIL on CLI, GPU/MPS, disk, iCloud, YAML, SMILES, chain IDs, chemistry and memory), then `run` (`boltz predict` with live progress and memory monitoring, resumable), then `analyze` (CSV, XLSX, HTML dashboard, CIF folder and interaction files). `all` chains all four, and `compare-sse` adds apo-versus-holo motif analysis.", "Q: What is the lead compound in the 2022 FGFR paper and how potent is it?\nA: The lead compound is 29, a pyrazolo[1,5-a]pyrimidine bearing a deuterated N-methyl (N-CD3) nicotinamide. It inhibits FGFR3 with an IC50 of 0.5 nM and FGFR2 with an IC50 of 1.0 nM, retains potency against the gatekeeper mutants FGFR3 V555L (2.7 nM) and V555M (6.1 nM), and is selective over FGFR1 (21 nM, 42-fold) and FGFR4 (145 nM).", "Q: What is leukemia inhibitory factor (LIF) and what does it do?\nA: LIF is a polyfunctional long-chain helical cytokine of the gp130 family. Marc Deller's thesis reports it acts on adipocytes, glial cells, motor neurons, hepatocytes, myoblasts, embryonic stem cells and certain myeloid leukaemic cell lines. LIF is one of the few cytokines considered absolutely essential for life: deleting the LIF gene in mice blocks embryonic implantation, and disrupting the LIFR gene causes placental, skeletal, neural and metabolic defects and perinatal death.", "Q: What is Click2Drug, as described by Marc Deller?\nA: In \"Revolutionizing Drug Discovery: The Click2Drug Platform and the Rise of AI-Powered Virtual Screening\", published 28 August 2025, Marc Deller describes Click2Drug as a directory developed by the Swiss Institute of Bioinformatics housing over 807 curated links to software, databases and web services across thirteen categories covering the whole drug design pipeline, from molecular visualization and target identification to ADME toxicity prediction.", "Q: Where and when was the 10E8 lipid paper published?\nA: It was published in PLoS Pathogens, volume 13, issue 2, article e1006212, on 22 February 2017, with DOI 10.1371/journal.ppat.1006212. It was received on 5 October 2016 and accepted on 2 February 2017, and the handling editor was Alexandra Trkola of the University of Zurich.", "Q: What warnings does Marc give about running RFdiffusion on an M1 Mac?\nA: Marc warns that Apple Silicon lacks CUDA, so all computation runs on CPU and will be, in his words, slooooww. He also cautions that some advanced functionality such as DGL-based graph operations may still fail because of missing native ARM libraries, and that troubleshooting DGL and compatibility issues should be expected.", "Q: How was povorcitinib synthesised?\nA: Povorcitinib was made by a route that first builds the top portion of the molecule: 2,4,5-trifluorobenzoic acid was converted to the acid chloride and coupled with a trifluoro-isopropylamine to give the benzamide, followed by nucleophilic aromatic substitution with azetidin-3-ol, oxidation with (diacetoxyiodo)benzene and TEMPO to the ketone, a Horner\u2013Wadsworth\u2013Emmons reaction to the cyanomethylene azetidine, and Michael addition with a pyrazole boronate. The assembled fragment was then joined to 4-bromo-3,5-dimethyl-1H-pyrazole by Suzuki coupling to give the final product.", "Q: What made INCB159020 orally bioavailable when earlier compounds were not?\nA: The breakthroughs were removing the two liabilities that blocked absorption and then rebuilding potency through rigidity. The team capped the molecule at one basic amine, replaced the metabolically vulnerable hydroxynaphthol, and cyclized the flexible ethyl amide into a morpholinone, which raised potency while cutting rotatable bonds. Swapping the lipophilic 7-fluoronaphthalene for a 2-methyl-3-chlorophenyl ring finally balanced permeability and clearance, giving a Caco-2 value of 3.6 \u00d7 10\u207b\u2076 cm/s and, at 1.5 mg/kg orally in nonhuman primate, an AUC of 950 nM\u00b7h and 42% oral bioavailability.", "Q: Why does Marc think facilitated dissociation matters therapeutically?\nA: Marc argues that moving beyond static affinity to engineer the kinetics of molecular interactions opens the way to immunotherapies that terminate cytokine action on cue and so minimise systemic toxicity, as well as biosensors that turn transient biomarker signals into detectable outputs within seconds. He notes the framework should generalise to other cytokines, receptors and synthetic cell circuits. The post is at https://marcdeller.com/precision-timing-of-cytokine-signals-through-facilitated-dissociation/.", "Q: What is SEAL, as covered in Marc Deller's post on self-adapting language models?\nA: In \"Self-Adapting Language Models: Are We Creating Our Final Invention?\", published 18 October 2025, Marc Deller describes MIT's SEAL framework, in which language models generate their own finetuning data and optimization directives as \"self-edits\" and use reinforcement learning to learn which synthetic data improves performance \u2014 reportedly outperforming even GPT-4-generated data on knowledge incorporation tasks.", "Q: How many works are on Marc's ORCID record?\nA: Marc Deller's ORCID record lists 51 works, all journal articles with DOIs. That is fewer than the 73 entries on his Google Scholar profile, which additionally includes patents, conference abstracts, a book chapter and his thesis.", "Q: What is Marc's ORCID?\nA: Marc's ORCID is 0000-0001-8070-6502. His Scopus Author ID is 21741715300 and his Google Scholar profile is at scholar.google.com/citations?user=o02oSJIAAAAJ.", "Q: What are the two steps of PET degradation in Marc's account?\nA: Marc describes nick generation, in which four MHET moieties bind across the subsites and cleavage of a single ester bond splits the chain into a TPA-terminal and an HE-terminal fragment, followed by terminal digestion, in which each fragment is repeatedly bound and cleaved to release free MHET and TPA. Final products are MHET, terephthalic acid and ethylene glycol, with BHET further degraded into MHET and ethylene glycol.", "Q: What is Marc's overall view of compositional protein design?\nA: Marc frames compositional design as a leap toward rationally assembling proteins from functional building blocks, comparing it to arranging pre-defined components in space and letting an expert system fill in the connections. He argues it opens new frontiers in synthetic biology and materials science by letting engineers think like architects.", "Q: How do the patented antibodies compare with earlier anti-mutCALR antibodies?\nA: Example 16 compares the Incyte antibodies with four benchmark antibodies: the commercial CAL2 antibody (AB1) and three built from published sequences (AB2 from WO2020175689, AB3 from WO2019178362 and AB4 from WO2016087514). Although all bound mutant CALR to varying degrees, AB1 through AB4 showed no functional activity, with IC50 values above 60 \u00b5g/mL for inhibition of cell proliferation, whereas the disclosed clones inhibited proliferation dose-dependently, clone 4 reaching 0.10 \u00b5g/mL.", "Q: Which residues does the inhibitor contact in FGFR3 according to the modeling?\nA: The docking model built on the published FGFR3 structure 4K33 showed two key interactions for compound 9: the imidazole ring nitrogen hydrogen bonds to Ala558 in the hinge region, and the amide carbonyl hydrogen bonds to Lys508, with the N-methyl group extending past the gatekeeper Val555. The model also placed the isopropyl group close to Asn562 in the lower hinge, which inspired the design of the tetrahydrofuran ether that engages that asparagine.", "Q: Are Marc Deller's Incyte drug discovery papers on his Scholar profile?\nA: Yes. His Google Scholar profile lists the 2022 Journal of Medicinal Chemistry paper on FGFR2/3 inhibitors (21 citations), \"Discovery of INCB159020, an orally bioavailable KRAS G12D inhibitor\" in 2025 (19), the INCB161734 KRAS G12D abstract in Cancer Research (19), the 2024 CDK2 inhibitor paper on (4-pyrazolyl)-2-aminopyrimidines (18), and the 2026 Journal of Medicinal Chemistry paper on povorcitinib (INCB054707).", "Q: What did the NANOG alanine mutants show?\nA: A series of alanine substitution mutants of the human NANOG homeodomain, chosen from the protein-DNA interactions and evolutionary conservation, were mostly less stable and showed lost or decreased DNA-binding affinity. Overexpressing the corresponding mouse NANOG mutants failed to maintain self-renewal of mouse embryonic stem cells without leukemia inhibitory factor, indicating those residues are critical for NANOG transcriptional activity.", "Q: What is in the lithium chloride / PEG 6000 block of the JCSG Top96?\nA: Wells H02\u2013H06 of the JCSG Top96 screen contain 1.0 M Lithium chloride with 10% or 20% (w/v) PEG 6000 in citric acid pH 5.0, MES pH 6.0, HEPES pH 7.0 or Tris pH 8.5, taken from the Hampton Grid Screen PEG/LiCl. The most productive is H05, 1.0 M Lithium chloride, 20% PEG 6000, 0.1 M MES pH 6.0, with 10 PDB structures.", "Q: What are the five self-descriptions Marc gives on his About page?\nA: Marc Deller describes himself on the About page as a structural biologist and protein scientist, someone with broad hands-on domain experience, a leader, mentor and team builder, an owner of a deep toolkit applied selectively, and a collaborative, impact-driven partner. Each point is tied to a practical claim about moving drug discovery programs forward.", "Q: What cancers does the WRN patent say the compounds could treat?\nA: The primary focus is MSI-H and dMMR cancers, especially colorectal, endometrial and gastric. The specification then lists a very wide range of additional indications, including melanoma, renal, prostate, breast, lung, head and neck and urothelial cancers, haematological malignancies such as AML, ALL, CLL, CML and multiple myeloma, sarcomas, and even sickle cell disease and sickle cell anaemia.", "Q: Why does AlphaFraud skip some entities?\nA: Of 96,827 post-cutoff human entities processed, 13,908 were skipped. 11,875 (85%) had no single UniProt mapping because they are antibodies, chimeras, engineered constructs, fusions or synthetics, so no AlphaFold model exists to compare against; 1,186 (9%) had no model covering the resolved residue range; 464 (3%) had fewer than 10 residues aligned; and 381 (3%) exceeded the 40 MB out-of-memory guard for the 3.8 GB droplet.", "Q: What are LOCKTACs and how do they differ from PROTACs?\nA: In \"The LockTac Revolution: Binding the Future of Therapeutics\", posted 22 August 2025, Marc Deller describes LOCKTACs as molecules that stabilise preexisting biomolecular interactions by reducing the rate at which two molecules fall apart. He stresses that this is fundamentally different from PROTACs, which force new bonds, whereas LOCKTACs prolong partnerships that already exist.", "Q: What does \"dimer hinge\" mean in the TM1088 paper?\nA: The \"dimer hinge\" is the angle between the domain-swapped \u03b16 helices of an RCK dimer pair, and changes in this angle drive the expansion or contraction of the gating ring. Because TM1088 is a two-subunit assembly, Marc Deller found it has an extra dimer hinge not seen in other gating rings: \"dimer hinge 1\" at the TM1088B N1:N1' interface and \"dimer hinge 2\" at the TM1088A N2:N2' interface. Both hinge angles are about 120\u00b0, and the helix-crossover angle between TM1088A and TM1088B dimers is about 80\u00b0.", "Q: What did the SECMALS analysis show?\nA: Size exclusion chromatography-multi-angle light scattering showed that wild-type 10E8 and mutant 1 each eluted as three peaks, all around 150 kDa, consistent with monomeric IgG undergoing slow conformational isomerisation on the column, while mutants 2, 3 and 5 eluted as single 150 kDa peaks. All samples used for neutralisation were monomeric, with less than 0.5% aggregate.", "Q: Tell me about the TrkA potassium channel gating ring structure.\nA: PDB entry 3L4B is the crystal structure of an octameric two-subunit TrkA K+ channel ring gating assembly from Thermotoga maritima MSB8, solved by X-ray diffraction at 3.45 \u00c5 and released on 31 March 2010. Marc Deller's post explains that it gave the first structural insights into a two-subunit Trk K+ transporter gating ring.", "Q: What did the study contribute methodologically?\nA: Marc Deller writes that the work combined structural analyses by X-ray crystallography and electron microscopy with longitudinal tracking and functional assays, revealing both the N276 and V5 glycan roadblocks and a rapid maturation pathway to breadth, guiding next-generation immunogens targeting the CD4-binding site.", "Q: What was the first structure Marc solved?\nA: Marc's first structure was the cytokine oncostatin M, solved during his Oxford D.Phil. and published in Structure in 2000 as \"Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\".", "Q: What percentage of JCSG structures crystallize in the 96 Top96 conditions?\nA: According to Marc Deller's 2013 poster, 55% of JCSG structures, 631 of 1144 deposited structures, crystallized in the 96 conditions that make up the JCSG Top96 screen.", "Q: What escape mutations does the AP33 structure explain?\nA: The structure explains why N415Y, N415D and G418D disrupt antibody binding while the virus retains fitness, since those positions carry the key hydrogen-bonding and hydrophobic contacts that stabilise the complex.", "Q: What is Elora's intellectual property strategy?\nA: The deck describes layered IP protection across enzymes, formulation, and therapeutic use. The filed provisional covers enzyme compositions and therapeutic use, with follow-on polymer expansion; the next step protects Elora-designed variants and proprietary formulations; and the moat is built on enzyme engineering know-how, formulation, manufacturing, and therapeutic use claims.", "Q: What applications does Marc highlight for IDR binders?\nA: He points to disabling prion seeds, dismantling toxic protein aggregates and arresting pain signalling pathways inside cells, alongside broader uses in diagnostics and the study of cellular signalling. The post cites the 2025 Science paper \"Design of intrinsically disordered region binding proteins\" and is at https://marcdeller.com/the-shape-shifting-world-of-intrinsically-disordered-region-binding-proteins/.", "Q: Why does Marc keep returning to structure validation?\nA: Marc Deller's short 2013 post frames validation metrics as the reason X-ray crystallography can be treated as the gold standard for building precise protein\u2013ligand models, while insisting on rigorous checking to combat overinterpretation. It is published at https://marcdeller.com/models-of-protein-ligand-crystal-structures-trust-but-verify/.", "Q: Is Marc Deller on LinkedIn?\nA: Yes, Marc Deller is on LinkedIn at linkedin.com/in/marccdeller, where his headline emphasises his work as a biotech co-founder and structural biology expert with 400+ protein structures to his name.", "Q: How were the conditions arranged on the JCSG Top96 plate?\nA: The conditions in the JCSG Top96 plate are organized according to precipitant class and precipitant concentration. The poster shows the plate with wells colored either by the number of PDB depositions or by precipitant class.", "Q: What role was proposed for the DUF1285 family?\nA: Genome-context and gene-fusion analyses supported a role for this family in signal transduction, possibly during oxidative stress. The paper's first author was Han GW, with Marc Deller among the Joint Center for Structural Genomics co-authors.", "Q: How does NMR complement crystallography for assessing stability?\nA: The review explains that NMR works in solution, so its measurements are not perturbed by crystal packing, and that the r.m.s.d. across an NMR ensemble is complementary to the crystallographic B factor. Marc Deller and colleagues note that NMR can distinguish folded from unfolded protein with as little as around 10 nM of sample and is non-destructive, so samples can be recovered for crystallization.", "Q: What is Horseradish Peroxidase used for in the invention?\nA: Horseradish Peroxidase (HRP) is a second oxidative enzyme in the invention, engineered to degrade the polyolefin plastics PE, PP, and PS. The patent covers engineered HRP variants, recombinant HRP expressed in non-fungal systems, and oxidative peroxidase hybrids with site-specific mutations for enhanced catalytic efficiency and stability.", "Q: What is povorcitinib being developed for clinically?\nA: According to the paper's conclusions, povorcitinib is in phase 3 trials for hidradenitis suppurativa, vitiligo and prurigo nodularis, and in phase 2 for asthma and chronic spontaneous urticaria. The preclinical arthritis data also suggest potential utility in rheumatoid arthritis and other JAK1-driven autoimmune and inflammatory disorders.", "Q: Which live statistics does mdeller.com show?\nA: The `top -bn1` panel on the mdeller.com landing page reports the server's uptime, load average, CPU cores, memory, disk, visit count and TLS status, alongside a health indicator at the top of the page. The values are read live rather than written into the page.", "Q: Which post-translational modifications are the most common?\nA: Citing an analysis by Khoury and coworkers of 821,182 experimentally analyzed proteins, Marc Deller's review lists phosphorylation as by far the most common with 58,383 occurrences, followed by acetylation, N-linked glycosylation, amidation, hydroxylation, methylation, O-linked glycosylation, ubiquitylation, pyrrolidone carboxylic acid and sulfation. Just under 500 distinct PTMs are catalogued in SWISS-PROT and TrEMBL.", "Q: Which journals does Marc Deller publish in?\nA: Marc Deller's publications appear in Science, Nature Structural and Molecular Biology, Immunity, Proceedings of the National Academy of Sciences, Journal of Molecular Biology, Structure, Journal of Biological Chemistry, Journal of Virology, PLoS Pathogens, PLoS ONE, Protein Science, Proteins: Structure, Function and Bioinformatics, Journal of the American Chemical Society, ACS Chemical Biology, Journal of Medicinal Chemistry, Journal of Computer-Aided Molecular Design, and extensively in Acta Crystallographica Sections D and F.", "Q: How many antibody clones does the patent describe?\nA: The patent reports 54 unique parental clones, with their six CDRs in Table 1 and full chain sequences in Table 4. Mutations were then introduced into three of them \u2014 clones 6, 15 and 17 \u2014 to generate 161 unique mutant clones numbered 55 to 215, whose CDRs and chains are in Tables 2 and 5.", "Q: What is YP_910028.1 from Bifidobacterium adolescentis?\nA: YP_910028.1 is a metal-dependent phosphoesterase whose unliganded and AMP-bound crystal structures were reported at 2.4 and 1.94 \u00c5 resolution in Proteins in 2011. Marc Deller is a co-author contributing crystallography through the Joint Center for Structural Genomics, and the study combined computational prediction with experimental validation of phosphoesterase activity.", "Q: Can you write some code for me?\nA: No code from me, sorry, not even a snippet. That isn't what I'm for. Marc's own coding is fair game though, and there is quite a lot of it: he builds small scientific web apps, fine-tunes language models on his own Mac, and has written the tooling behind several of his projects. Want to hear about any of that?", "Q: Who was donor DRVI01?\nA: DRVI01 was an HIV-1-infected Chinese long-term non-progressor from a former plasma donor cohort, studied in the 2016 Immunity paper Marc Deller co-authored. He entered the study at age 40 and was followed every six months from 2005 to 2010, carrying a clade-B' HIV-1 strain with viral loads of 74,200 to 310,000 copies/mL and CD4+ T cell counts of 335 to 769 cells/mL.", "Q: What can you help me with?\nA: I can tell you about Marc Deller's background and career, explain his scientific papers and the structures he has solved, describe Elora Therapeutics and its microplastic-degrading enzymes, walk you through his software projects such as AlphaFraud and BoltzMaker, or share a few personal favourites. Ask away!", "Q: What did isoelectric focusing reveal about LIFR?\nA: Isoelectric focusing showed that only the fully glycosylated LIFR-FN3-3C-Fc fusion ran as one fairly tight population, at a pI of about 5-6 (versus the theoretical pI of 7.2 for the LIFR-FN moiety). Deglycosylation, 3C cleavage or acid elution disrupted this profile into a continuous smear of species with pI from 4 to 8, leading Marc to conclude that both the sugars and the Fc moiety have a stabilising effect on LIFR-FN3.", "Q: What is the overall fold of oncostatin M as described in Marc Deller's thesis?\nA: Marc Deller's thesis showed that OSM adopts the up-up-down-down four-helix bundle topology predicted for the haemopoietin cytokine family. The molecule is a compact barrel of roughly 20 \u00c5 \u00d7 27 \u00c5 \u00d7 56 \u00c5, built from four main helices (helix A, residues 10\u201337; helix B, residues 67\u201390; helix C, residues 105\u2013131; helix D, residues 159\u2013185) connected by two long overhand loops (the A-B loop, residues 38\u201366, and the C-D loop, residues 132\u2013158) and one short B-C loop (residues 91\u2013104).", "Q: Did sodium butyrate help LIFR expression in CHO cells?\nA: Yes, modestly. Adding 2 mM sodium butyrate to the CHO growth medium increased LIFR-FN3-3C-Fc expression by up to 30 %. Marc noted that sodium butyrate is known to boost expression from recombinant plasmids in mammalian cells through effects on histone acetylation and histone H1 phosphorylation, which regulate chromatin transcriptional activity, and suggested it could be useful in developing other mammalian systems.", "Q: Where could the JCSG Top96 conditions be obtained?\nA: The 2013 poster states that the JCSG Top96 conditions were made available at http://www.jcsg.org/top96.", "Q: Has povorcitinib shown clinical results?\nA: Marc Deller notes that povorcitinib has produced durable responses in hidradenitis suppurativa, with 2026 long-term readouts reinforcing the safety case for selective JAK1 inhibition. His post on the discovery is at https://marcdeller.com/discovery-of-the-orally-bioavailable-isoform-selective-janus-kinase-1-compound-povorcitinib/.", "Q: What tools does Marc use to vibe-code?\nA: Marc builds with Claude Code, Cowork and a suite of custom skills (marcs-profile, marcs-blog-writer, marcs-resume, marcs-job-hunter, marcs-vibe-coding, marcs-vibe-icon, marcs-page-icon and a BoltzMaker input wizard), usually in pure HTML/CSS/JavaScript for browser apps or Python/Flask for server apps, deployed to his mdeller.com droplet.", "Q: I heard Marc worked on a really tight patent deadline once. What happened?\nA: Yes \u2014 at Incyte, Marc led a structural biology sprint to solve the crystal structure of a peptide bound to a mutant-calreticulin antibody complex in just over ten days, beating an already aggressive two-week patent deadline.", "Q: What practical expression advice does Marc give for crystallography targets?\nA: He recommends reduced expression temperatures of 15\u201325\u00b0C to slow folding kinetics and reduce aggregation, codon optimisation with rare-tRNA strains such as Rosetta for eukaryotic proteins in E. coli, supplying cofactors and prosthetic groups during expression, and media supplementation with osmolytes and chemical chaperones. For mammalian systems he mentions post-transfection temperature reduction and histone deacetylase inhibitors.", "Q: What is TM1089 and how does it relate to TM1088?\nA: TM1089 is the membrane-spanning TrkH-like component of the T. maritima Trk transporter, which associates with the cytosolic TM1088A/TM1088B gating ring to form the functional transporter. In vivo complementation assays had shown that both TM1088 and TM1089 are required for cell survival in low potassium. Marc Deller noted that structures of TM1088 in complex with TM1089 would be needed to understand how the gating ring couples to the membrane channel.", "Q: What is Marc Deller's ORCID identifier?\nA: Marc Deller's ORCID iD is 0000-0001-8070-6502, and the record is public at https://orcid.org/0000-0001-8070-6502. It lists 51 works.", "Q: How many citations does Marc Deller have?\nA: Marc Deller's Google Scholar profile records 3,967 citations all time, of which 1,343 have come since 2021. The profile lists 73 publications in total.", "Q: How are the conditions ordered on the JCSG Top96 plate?\nA: The JCSG Top96 conditions are listed in plate order according to major precipitant class, as employed at the JCSG. Row A starts with organic precipitants, followed by organic/PEG/salt and organic/PEG mixtures in rows A\u2013B, organic/salt at the end of row B, PEG-only conditions in row C, salt-only conditions in rows C\u2013E, and salt/PEG combinations filling rows E through H.", "Q: How can I contact Marc Deller?\nA: Marc Deller can be reached by phone at (302) 358-6093 or by email at marc@marcdeller.com. He is also on LinkedIn at www.linkedin.com/in/marccdeller, and his personal website and blog is https://marcdeller.com.", "Q: Where is the lipid-binding site on 10E8?\nA: The lipid head groups bind in a crevice formed by the light-chain CDR1 (residues Leu28, Arg29, Ser30, His31 and Tyr32), light-chain framework region 3 (Ala66, Ser67 and Gly68), and the tip of the heavy-chain CDR3 (Trp100b, Ser100c and Gly100d). In crystals grown without added lipid, glycerol from the cryoprotectant occupies the same site.", "Q: Why are ligands especially hard to model correctly?\nA: Because ligands contribute only a small fraction of the total scattering mass, so validating their atomic models is critical for confidence in the binding interactions inferred from them. Near-atomic resolution and robust map\u2013model correlation are needed to be sure a ligand is correctly localised in the binding pocket.", "Q: Is WRN inhibition the same as classic DNA-damaging chemotherapy?\nA: No. Marc Deller frames WRN helicase inhibition as a novel synthetic lethal approach distinct from traditional DNA damage agents, exploiting a cancer-specific vulnerability rather than damaging DNA indiscriminately, and notes that multiple companies have moved both covalent and non-covalent WRN inhibitors into trials.", "Q: What were the refinement statistics for the PGT 135 structures?\nA: The unliganded PGT 135 Fab structure was refined against data from 31.40 to 1.75 \u00c5 using 49,464 reflections, giving Rwork/Rfree of 20.7%/24.5%, with r.m.s. deviations of 0.006 \u00c5 in bond lengths and 1.11\u00b0 in bond angles. The PGT 135\u2013gp120\u2013CD4\u201317b complex was refined from 43.58 to 3.10 \u00c5 using 30,850 reflections, giving Rwork/Rfree of 24.6%/36.7%, with r.m.s. deviations of 0.012 \u00c5 and 1.46\u00b0.", "Q: Does the inhibitor make a direct contact with the mutant aspartate?\nA: Not with its side chain. In the structures Marc Deller helped determine, the inhibitors engage the backbone N\u2013H of Asp12 rather than the aspartate side chain itself, yet compound 5a still showed 55-fold selectivity for KRAS G12D over wild-type KRAS in surface plasmon resonance (0.16 nM versus 8.8 nM). That selectivity was generally maintained across the series whenever the backbone N\u2013H interaction was present.", "Q: How did Marc Deller contribute to the Dhaf4260 structure?\nA: Marc Deller was one of the Joint Center for Structural Genomics co-authors contributing crystallography, alongside Miller MD, Aravind L, Bakolitsa C and the wider JCSG team. The paper appeared in Acta Crystallographica Section F in 2010 (DOI 10.1107/s1744309110007517).", "Q: What makes up the hydrophobic core of oncostatin M?\nA: Marc Deller found that the solvent-inaccessible core of OSM is built around two clusters of aromatic stacking interactions: Phe 56, Tyr 173, Phe 169 and Phe 176 form one group, and Phe 70, Phe 185 and Trp 187 form the second. Helix D donates all of these aromatics except Phe 56 (A-B loop) and Phe 70 (helix B). Notably, no hydrogen bonds between the main helices were observed; the fold is stabilised almost entirely by hydrophobic interactions.", "Q: What governance does Marc call for on AI protein design?\nA: Marc argues that voluntary community commitments, while a meaningful first step, cannot substitute for enforceable policy frameworks and technical safeguards embedded in the development pipeline. Drawing on Pannu and colleagues in PLOS Computational Biology, he supports prioritizing model evaluations by the consequence severity of potential misuse and completing biosecurity assessments before deployment rather than after demonstrated harm.", "Q: Who solved the crystal structure in the FGFR2/3 paper?\nA: The FGFR2 co-crystal structure in the 2022 paper was solved by the Incyte structural biology team of Hsiang-Ting Lei, Leslie B. Epling and Marc C. Deller. Their structure of compound 30 bound to FGFR2 was released as PDB 8E1X.", "Q: How is BcYkfC built?\nA: BcYkfC contains two N-terminal bacterial SH3 (SH3b) domains in addition to the C-terminal catalytic NlpC/P60 domain that is ubiquitous in the very large family of cell-wall-related cysteine peptidases. The bound L-Ala-gamma-D-Glu product allowed conserved sequence and structural signatures for recognition of L-Ala and gamma-D-Glu to be identified.", "Q: How was the oncostatin M protein produced for crystallography?\nA: Marc Deller and colleagues expressed residues 1\u2013187 of mature human oncostatin M in Escherichia coli strain JM109 as a glutathione-S-transferase (GST) fusion from the pGex-2T vector, with a rhinovirus 3C protease site between GST and hOSM. The fusion was captured on glutathione-sepharose, the OSM released by overnight 3C protease cleavage at 37\u00b0C, and the protein purified by gel filtration on Sephadex-75. The identity of the crystallisation sample was confirmed by electrospray mass spectrometry and 10 cycles of N-terminal sequencing, and binding assays showed the fragment bound gp130 with wild-type affinity.", "Q: What software does Marc use?\nA: Marc's stack includes PyMOL, ChimeraX, COOT, HKL2000/3000, PHENIX, CCP4 and CCP4-EM, RELION and CryoSPARC for cryo-EM, AlphaFold, Boltz-2, ESMFold, ProteinMPNN, RFdiffusion, Rosetta, MOE, AutoDock, GROMACS, OpenMM and Schr\u00f6dinger for modelling, PLIP and RDKit for cheminformatics, plus Python, R/Shiny, MATLAB, AWS and HPC tools.", "Q: Which publications does Marc list as his top ones?\nA: The top publications listed are the 2012 structure of hepatitis C virus envelope glycoprotein E2 antigenic site 412 to 423 in complex with antibody AP33, the 2013 crystal structure of a soluble cleaved HIV-1 envelope trimer, the 2013 review \"Models of protein\u2013ligand crystal structures: trust, but verify\", the 2013 supersite of immune vulnerability on the glycosylated face of HIV-1 gp120, and the 2014 paper \"Approaches to automated protein crystal harvesting\".", "Q: How did the KTH team confirm that PET was fully broken down?\nA: Marc explains that 1H NMR spectroscopy provided the chemical confirmation: after one hour with S238A PETase a new peak appeared at 8.08 ppm corresponding to the aromatic benzene ring protons of terephthalic acid, absent in the no-enzyme control, and a second peak at 5.19 ppm indicated the ethylene glycol moiety of released MHET, with peaks at 4.30 and 4.66 ppm from residual oligomers.", "Q: What solvent accessibility did the functionally important OSM residues have?\nA: Marc Deller reported that all OSM residues found to be active in receptor binding had side chains with relative solvent accessibility of 27 % or greater, as calculated with NACCESS from the OSM coordinates. For example Phe 160 had 65 %, Lys 163 42 %, Gln 16 56 %, Gln 20 27 %, Gly 120 76 %, Asn 123 61 % and Asn 124 31 %.", "Q: Which university did Marc go to for his first degree?\nA: The University of Leeds, where he took a First Class Honours BSc in Biochemistry and Molecular Biology between 1991 and 1995. It included a one-year industrial placement at Boots Pharmaceuticals, in the team behind Advil, ibuprofen and Nurofen. His mentor there, Prof. Tony North, is who set him on protein structure.", "Q: How do the two enzymes differ in size and sequence identity?\nA: Marc reports that FAST-PETase has 290 amino acids and is identical to IsPETase apart from its five engineered mutations, retaining two disulfide bonds (Cys203-Cys239 and Cys273-Cys289). HGMP01 has 275 amino acids and a molecular weight of 30.18 kDa, shares only 20 to 41% sequence identity with known PETases, and comes from gut bacteria of the Acutalibacteraceae family.", "Q: Did povorcitinib work in an animal model of arthritis?\nA: Yes. In the rat adjuvant-induced arthritis model, an aggressive T-cell-mediated autoimmune arthritis that closely mimics human rheumatoid arthritis, once-daily oral povorcitinib produced dose-dependent improvement in clinical score, with the 3 and 10 mg/kg arms giving almost full resolution of disease. Histopathology scores for bone resorption, periosteal new bone formation, inflammation, cartilage damage and pannus were reduced by 81% and 94% at 3 and 10 mg/kg respectively.", "Q: Where can I find Marc's write-up of PDB 2GVK?\nA: Marc Deller's post is titled \"Crystal Structure of a Dye-Decolorizing Peroxidase (DyP) from Bacteroides thetaiotaomicron VPI-5482 at 1.6 \u00c5 Resolution\" and appears in the structures category of his blog. You can read it at https://marcdeller.com/crystal-structure-of-a-dye-decolorizing-peroxidase-dyp-from-bacteroides-thetaiotaomicron-vpi-5482-at-1-6-a-resolution/.", "Q: Which conditions in the JCSG Top96 produced 12 PDB structures each?\nA: Five JCSG Top96 conditions each yielded 12 PDB structures: A11 (0.17 M Sodium acetate, 25.5% PEG 4000, 15% Glycerol, 0.085 M Tris pH 8.5), B06 (20% PEG 4000, 10% Isopropanol, 0.1 M HEPES pH 7.5), F08 (20% PEG 3350, 0.2 M Magnesium acetate), G01 (0.2 M Potassium nitrate, 20% PEG 3350) and H10 (0.2 M Magnesium chloride, 20% PEG 8000, 0.1 M Tris pH 8.5).", "Q: Which cell lines and media did Marc use for eukaryotic expression according to the appendices?\nA: According to Appendix III of Marc Deller's thesis, he cultured and transfected 293T cells in DMEM supplemented with 10 % FCS, penicillin, streptomycin, glutamine, pyruvate and glucose; grew Drosophila S2 cells in SF900 medium with 5 % FCS, antibiotics and 300 \u03bcg/ml hygromycin-B for selection; and cultured CHO cells in GMEM-S with 10 % FCS, antibiotics and 50 \u03bcM L-methionine sulphoximine for glutamine synthetase selection.", "Q: What does Marc say about cutinases such as Cut190 and LCC?\nA: Marc writes that cutinases like Cut190 and leaf-branch compost cutinase (LCC) operate robustly under diverse industrial conditions and degrade polyesters including PET, making them useful alternative scaffolds to PETase.", "Q: What is PDB entry 10PJ?\nA: 10PJ is the JAK2 kinase JH1 domain in complex with povorcitinib, solved by X-ray diffraction at 1.59 \u00c5 in space group I 1 2 1, with 632 polymer residues and 75.66 kDa, released on 10 June 2026. It was deposited by Epling and Fenalti, is keyworded as a transferase, and shares its primary citation with 10PI.", "Q: How do you run chatPDB locally?\nA: Activate the virtual environment and run `python scripts/chat.py --model models/chatpdb_32b_v1`, which loads the model in-process with no server needed; `/help` lists slash commands, `/info` shows session details, and `--no-rag` runs model-only. The evaluation scripts do need `mlx_lm.server --model models/chatpdb_32b_v1 --port 8080` running first.", "Q: Is there a commercial angle to automated harvesting?\nA: Yes. The post reports that the protein crystallisation market, valued at $1.84 billion in 2023, is projected to reach $3.84 billion by 2033, driven by automation and pharmaceutical high-throughput screening. Marc Deller's post is at https://marcdeller.com/approaches-to-automated-protein-crystal-harvesting/.", "Q: What did the review say about iron metabolism and the transferrin receptor?\nA: Marc Deller and Yvonne Jones described the structure of the transferrin receptor (TfR) ectodomain and its rapidly following complex with HFE as an example of synergy between disease genetics, functional studies and structural biology. HFE, an MHC class I-like molecule linked to hereditary haemochromatosis, was shown to associate with TfR in an architecture consistent with receptors from the same cell surface, suggesting a conformational trigger by which TfR signals to cytoplasmic proteins that HFE is bound.", "Q: What programs did Marc use to analyse and illustrate the final model?\nA: Stereochemistry was assessed with PROCHECK; secondary structure and solvent accessibility were calculated with DSSP, HERA and NACCESS; structural superpositions used SHP; and figures were rendered with BOBSCRIPT and RASTER3D.", "Q: Who funded the ligand validation chapter?\nA: The acknowledgements state that Bernhard Rupp received partial support from the Austrian Science Foundation (FWF) under project P28395-B26.", "Q: What is the Nanog homeodomain structure Marc lists?\nA: PDB entry 4RBO is the crystal structure of the human Nanog homeodomain bound to DNA, solved by X-ray diffraction at 3.30 \u00c5 and released on 1 October 2014. Marc Deller's post notes that the entry contains the putative homeobox protein NANOGP8 together with two complementary DNA strands.", "Q: Why is the smoothness of ButtFold's animation caveated?\nA: ButtFold states that the 150 poses in a trajectory are 50,000 integration steps apart, far enough that an atom can move further than the protein is wide between two of them, so what appears in between is drawn to join them up rather than computed. The motion is therefore smooth where the model is not, while the music remains a faithful map of the simulation's events and nothing more.", "Q: What is Marc Deller's 2000 review \"Cell surface receptors\" about?\nA: \"Cell surface receptors\" is a review written by Marc Deller and E. Yvonne Jones for Current Opinion in Structural Biology (volume 10, pages 213\u2013219, 2000). It surveys the structural biology of the extracellular portions of cell surface receptors and, in particular, the ligand\u2013receptor recognition complexes reported during the preceding year, a period in which the database of such complexes essentially doubled.", "Q: What software did Marc Deller use for the TM1088 structures?\nA: Data were processed with HKL2000 or Mosflm and SCALA from CCP4. Molecular replacement used Phaser (for the octamer) and MOLREP with the JCSG MR pipeline (for TM1088A). Refinement was carried out with Refmac5 and PHENIX, model building with O, Coot and MIFit, validation with MolProbity, TLS groups were determined with the TLS Motion Determination server, and figures were prepared in PyMOL.", "Q: How does AP33 grip the E2 peptide?\nA: The crystal structure showed the AP33 epitope forming a \u03b2-hairpin sandwiched between the heavy and light chains of the Fab. Most of the binding is mediated by hydrophobic interactions along the hydrophobic face of the epitope, with a number of hydrogen bonds \u2014 mostly between Fab side chains and the peptide main chain \u2014 also stabilising the interaction.", "Q: What did N-terminal sequencing tell Marc Deller about the purified oncostatin M?\nA: N-terminal sequencing over 10 cycles confirmed the identity of purified OSM_\u0394188 and showed that E. coli JM109 produced a single native species carrying an extra four-residue 3C linker (GPGS) at the N-terminus, arising from 3C protease and BamHI cleavage of the engineered construct. JM109 grown on Se-Met medium gave two minor species, whereas the auxotrophic strain B834 produced only one species with the correct sequence.", "Q: How well cited is Marc's oncostatin M paper?\nA: \"Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\", by Deller, Hudson, Ikemizu, Bravo, Jones and Heath, published in Structure 8(8):863-874 in 2000, has 93 citations on Google Scholar. Marc is its first author.", "Q: What does the Artificial Intelligence page on marcdeller.com cover?\nA: The AI page, subtitled \"AI to Action: From Models to Molecules\", argues that artificial intelligence is transforming drug discovery, protein science, and structural biology. It covers structure prediction, generative protein design, virtual screening, and toxicity prediction, and the resulting shift toward faster development, lower costs, and higher success rates.", "Q: How selective is povorcitinib, and why does sparing JAK2 matter?\nA: Two design moves, trading a metabolically vulnerable pyrrolopyrimidine for the smaller dimethylpyrazole and adding an intramolecular hydrogen bond to mask the amide's polarity, roughly doubled isoform selectivity to about 46-fold while rescuing permeability and oral exposure. Sparing JAK2 avoids the anemia, thrombocytopenia and other cytopenias tied to blocking erythropoietin and thrombopoietin signalling.", "Q: Does Marc have project management training?\nA: Yes. Google, through Coursera, issued Marc Deller the \"Foundations of Project Management\" certificate on 30 March 2025. He writes that complex scientific work depends on planning, coordination and clear milestones, and that the course helps manage experiments, collaborations, timelines and deliverables.", "Q: What does the Real Space Correlation Coefficient tell you?\nA: RSCC measures how well the model fits the electron density on a local, per-residue basis. As explained in the chapter Marc Deller co-authored, a perfect fit gives an RSCC of 1.0, values between 0.8 and 1.0 are commonly considered good, and lower values mean greater uncertainty in the placement of the model.", "Q: Do gut bacteria have membrane-attack complex proteins?\nA: Yes. Marc Deller co-authored the 2010 Acta Crystallographica Section F paper reporting the structure of Bth-MACPF, the MACPF-domain-containing protein BT_3439 from Bacteroides thetaiotaomicron, a predominant member of the mammalian intestinal microbiota. Bacterial MACPFs occur in diverse species, including most human gut-associated Bacteroides.", "Q: What limitations of PET-degrading enzymes does Marc identify in that post?\nA: Marc notes that mesophilic variants such as FAST-PETase lose efficiency above 50\u00b0C as PET ages and crystallises, that thermophilic enzymes need energy-intensive heating that offsets environmental benefits, and that product inhibition by MHET and TPA affects all variants, requiring co-expression with MHETase or continuous product removal. The post is at https://marcdeller.com/engineering-evolution-how-fast-petase-and-other-variants-are-transforming-plastic-biodegradation/.", "Q: Why does DeepCovalent say 90 % of drugs fail in clinical trials?\nA: DeepCovalent's OS3D page attributes the failure rate to incomplete modelling of protein-ligand context and clinical constraints, and to optimisation being constrained to known chemical space. Its stated answer is unified multimodal, multi-objective modelling of protein-ligand systems that integrates biology, bioactivity and rich ADMET context, giving unbiased access to previously unknown chemical space.", "Q: What method and resolution were used for PDB 3NL9?\nA: The Exig_1061 structure was determined by X-ray diffraction at 1.78 \u00c5 resolution, with ethylene glycol (EDO) as the only unique ligand in the model. It was released to the PDB in July 2010.", "Q: What are the 10E8 structures about?\nA: The 10E8 series shows the broadly neutralising antibody 10E8 and light-chain mutants bound to the T117v2 HIV-1 MPER epitope scaffold, all released on 8 March 2017 and deposited by Irimia, A. and Wilson, I.A. They comprise 5SY8 (light chain mutant 1, 1.62 \u00c5, P 1), 5TFW (mutant 2, 2.168 \u00c5, C 1 2 1), 5T29 (mutant 3, 2.03 \u00c5, C 1 2 1), 5T5B (mutant 5, 2.07 \u00c5, P 1 21 1), 5T6L (wild type Fab, 2.1 \u00c5), and the lipid complexes 5T80 with phosphatidic acid at 2.62 \u00c5 and 5T85 with phosphatidylglycerol at 2.373 \u00c5.", "Q: Which of Marc's apps relate to Elora Therapeutics?\nA: PANTS, which mines metagenomes for new PETase- and MHETase-like enzymes, and the PETase Lineage Map, both support Elora's engineered-enzyme work on microplastic degradation.", "Q: What are the 6 C's?\nA: The 6 C's are Marc's motto and leadership philosophy: Curiosity, Courage, Creativity, Communication, Compassion and Collaboration. He mentors his teams through them and organises his own resume highlights around them.", "Q: What does the patent mean by a biparatopic antibody in this context?\nA: Because the crystal structure showed the Fab engaging two non-identical epitopes on the same mutant CALR C-terminal segment, the patent claims antibodies that bind two non-identical epitopes within DEEQRTRRMMRTKMRMRRMRR, and describes biparatopic antibodies \u2014 antibodies recognising two different epitopes on the same target antigen \u2014 as an embodiment. This is a direct consequence of Marc Deller's crystallographic result.", "Q: Over what time window were the PDB counts in the JCSG Top96 formulation collected?\nA: The PDB counts in the JCSG Top96 formulation cover JCSG structures deposited between 1 January 2001 and 23 September 2010, with the 96 conditions accounting for 631 depositions.", "Q: Over what range of protein sizes and resolutions does MICA work?\nA: Marc notes MICA held up across proteins from 384 to 11,109 residues and across map resolutions from 1.5 \u00c5 to 4 \u00c5. He points out that this robustness makes it especially valuable for membrane proteins, large macromolecular assemblies and dynamic complexes that have historically been hard to model.", "Q: Which vibe-coded projects are highlighted on the marcdeller.com home page?\nA: The home page names BoltzMaker, ChemSage, and MicroPlastics Blaster as vibe-coded projects, and notes that they are available as web apps at mdeller.com. Each has its own write-up on marcdeller.com.", "Q: What are PETase and MHETase?\nA: PETase (polyethylene terephthalate hydrolase) is an enzyme that breaks the ester bonds in PET plastic, and MHETase (mono(2-hydroxyethyl) terephthalate hydrolase) finishes that job by hydrolyzing the intermediate MHET into terephthalic acid (TPA) and ethylene glycol (EG). In Marc Deller's patent, engineered and fusion versions of these two enzymes are optimized to work together at human physiological temperature (~37 \u00b0C) to degrade PET inside the body.", "Q: Where can the FGFR2 structure be downloaded?\nA: The FGFR2 co-crystal structure of compound 30 determined by Marc Deller, Hsiang-Ting Lei and Leslie Epling is available from the Protein Data Bank under accession code 8E1X. The paper's Supporting Information also includes the coordinates and experimental details for that structure, along with docking models for compound 9 and the theoretical tetrahydrofuran derivative bound to FGFR3.", "Q: How does keeping patients in mind change Marc's decisions?\nA: Marc Deller writes that keeping the patient perspective front and centre shapes how he prioritises programs, communicates risk and advocates for approaches that maximise meaningful benefit. He says the goal of his research is that every structure eventually impacts a patient, and that keeping that person in mind shapes every decision upstream.", "Q: What happens to normal cells under covalent WRN inhibition?\nA: The post explains that the synthetic lethality relationship holds between deficient mismatch repair and WRN function, so dual disruption causes catastrophic DNA damage specifically in cancer cells while sparing normal, microsatellite-stable tissue.", "Q: What is intramolecular hydrogen bonding and why did it matter here?\nA: Intramolecular hydrogen bonding, or IMHB, is a design tactic where the molecule folds a hydrogen bond back on itself, masking polarity and improving membrane permeability and pharmacokinetics. Adding an ortho-fluorine on the phenyl ring together with a trifluoromethyl amide capping group produced compound 16, which had a 5.5-fold better Caco-2 permeability, 7.5-fold higher rat exposure and a 52-fold increase in monkey exposure relative to the non-IMHB parent compound 9.", "Q: Can you help me with my tax return?\nA: That's outside what I know, I'm afraid. I'm here for questions about Marc Deller, his science, his company Elora Therapeutics and his projects. Is there anything in that area I can help with?", "Q: What partnership models does DeepCovalent offer?\nA: DeepCovalent offers two core models plus flexibility for customised structures. The first is On-Demand Generation and Early Assessment, giving rapid low-cost access to OS3D so partners receive synthesis-ready, fully in-silico validated molecules in three weeks for fast go/no-go decisions. The second is a Flexible Co-Development Model with shared ownership of molecular IP.", "Q: What residue numbering convention is used for the OSM structure?\nA: Residue numbering in Marc Deller's OSM structure corresponds to the human native mature protein, GenBank accession number M27286 (Malik et al., 1989).", "Q: Where can I find Marc Deller's other profiles and links?\nA: Marc Deller's LinkedIn profile links out to his personal site (marcdeller.com), his GitHub (github.com/bellcheddar), his Scopus author page, his app-hosting site (mdeller.com), and Elora Therapeutics (eloratherapeutics.com).", "Q: How many structures has Marc solved?\nA: Marc has contributed to more than 400 protein structures deposited in the Protein Data Bank over his career, including 70+ novel structures at the JCSG and 100+ protein\u2013drug co-crystal structures at Incyte. Seventy-five PDB entries carry his name as a deposition author; the rest were deposited under the collective JCSG author name.", "Q: What software does the chapter list for ligand validation?\nA: The table in the chapter lists Twilight, COOT, MolProbity, VHELIBS, the ValLigURL server, MotiveValidator and ValidatorDB, PDB_REDO, the wwPDB Validation Server, PRIVATEER, LIGPLOT, EDS, EDSTATS, OVERLAPMAP, BUSTER-TNT and WHAT_IF/WHAT_CHECK. They span standalone programs, web servers and databases of ligands scored by quality criteria.", "Q: What function was proposed for LP2179?\nA: Remote structural similarity to two other uncharacterized protein families specific to the Bacillus genus, PF08868 and PF08968, and to prokaryotic S-adenosylmethionine decarboxylases, is consistent with a role in amino-acid metabolism. Genomic neighborhood analysis supports that assignment, which the authors suggested might also extend to PF08868 and PF08968.", "Q: Which earlier JCSG screens preceded the Top96?\nA: The poster describes JCSG Top96 as a follow-up to previous JCSG screens, namely the JCSG Core Suite and the JCSG+ screen.", "Q: What kinds of plastics is the invention designed to degrade?\nA: The invention targets the most common microplastics and nanoplastics found in human blood and tissues: polyethylene terephthalate (PET), polyethylene (PE), polypropylene (PP), and polystyrene (PS). PETase and MHETase degrade PET, while the oxidative enzymes CotA-Laccase and HRP are aimed at PE, PP, and PS.", "Q: How does gp130 compare structurally with other haemopoietin receptors in Marc's thesis?\nA: Marc Deller's comparison showed that the cytokine-binding region of gp130 is most similar to GCSFR (RMSD 2.52 \u00c5 over 189 equivalent residues, 29 % sequence identity) and to EpoR (RMSD 2.52 \u00c5 over 177 residues, 20 % identity), with PrlR (3.03 \u00c5) and GHR (3.07 \u00c5) more distant. The relatively poor overall superpositions were largely due to differences in interdomain tilt angle: about 90\u00b0 for EpoR and other class 1 receptors versus about 70\u00b0 for gp130.", "Q: Who supplied the biotinylated ligands and transfected cell lines used in the assays?\nA: The biotinylated ligands used as competing agents in the receptor assays were provided by K. Chobotova of the CRC Growth Factors Group in Birmingham. The Ba/F3 cells stably transfected with human gp130 and human LIFR, which made the survival assay responsive to human LIF and OSM, were provided by Vicky Barton of the same group.", "Q: How did the variants affect oxygen consumption?\nA: Initial O2 consumption rates, measured with an Oxygraph Plus system, ranged from about 50\u2013100% of wild type. H161E, Q167A, Q167E and T74A were the slowest at roughly 50% of wild type, H161A was at about 70%, and H161Q consumed O2 at rates comparable to wild type. Several variants, most clearly H161Q and H161A, showed smaller defects in O2 consumption than in product formation, suggesting O2 utilization had become decoupled from substrate hydroxylation.", "Q: How did the team confirm the effect was really due to JAK1 and not JAK2?\nA: Povorcitinib inhibited JAK1-dependent signalling \u2014 IL-2-stimulated human T cells (JAK1/3) with IC50 around 10\u201330 nM, IL-6-stimulated STAT1 and STAT3 phosphorylation in human whole blood with IC50 of 360 and 690 nM, and IL-6-driven MCP-1 production in human PBMCs with IC50 of 178 nM. In contrast, it did not block STAT3 or STAT5 phosphorylation stimulated by the JAK2-specific cytokines TPO and GM-CSF, where the IC50 values exceeded 10,000 nM.", "Q: How does the patent propose to deliver these enzymes into the body?\nA: The patent describes three delivery approaches. Oral formulations use biodegradable polymer nano-encapsulation or enteric-coated capsules; injectable formulations use PEGylation, lipid nanoparticles, or hydrogels for systemic distribution with reduced immunogenicity and enhanced stability; and a hybrid system delivers the PET-degrading enzymes and MHETase orally while giving CotA-Laccase and HRP by injection.", "Q: What does Marc Deller say about cryo-electron tomography?\nA: In \"Beyond the Microscope: How Cryo-ET is Revealing Disease in Its Native State\", published 13 August 2025, Marc Deller explains that cryo-ET flash-freezes tissue and takes images from multiple angles to build 3D reconstructions of cellular landscapes at near-atomic resolution, letting scientists see molecules inside their native cellular environment rather than purified in isolation.", "Q: What obstacles to recursive self-improvement does Marc identify?\nA: Marc identifies catastrophic forgetting, where SEAL's performance on earlier tasks degrades across sequential updates, the computational overhead of evaluating each potential modification, and the fact that current systems still operate within carefully constrained environments with human-designed reward signals, as in the Absolute Zero Reasoner.", "Q: What is a scaffold hop, and how was one used in this paper?\nA: A scaffold hop replaces the central ring system of a molecule while keeping the groups that make the key binding contacts. In the 2022 FGFR paper, the team moved the bridgehead nitrogen to convert the highly basic imidazo[1,2-b]pyridazine core into a much less polar pyrazolo[1,5-a]pyrimidine. The resulting compound 26 was twice as potent as its predecessor in the enzyme assay (FGFR3 IC50 = 1.0 nM) and translated dramatically better into cells (BA/F3 IC50 = 29 nM versus 969 nM).", "Q: How was INCB159020 synthesized?\nA: The compounds were built from a common intermediate prepared by iodination of 2-amino-4-bromo-3-fluorobenzoic acid, conversion to the benzoxazinedione, cyclization with ethyl nitroacetate, chlorination with POCl3, and displacement with the azabicyclo[2.1.1]hexane amine. The indole core, which replaced the earlier imidazole, was made by a Sonogashira coupling followed by cyclization \u2014 originally with gold(I) catalysis, later more cleanly with cesium carbonate and no transition metal \u2014 after which Suzuki cross-coupling, solvent installation via sulfone displacement or Liebeskind\u2013Srogl coupling, and deprotection gave the final compounds.", "Q: What ligands were found in the lipid site across the different structures?\nA: A diversity of ligands occupied the same cleft depending on the crystallisation and cryoprotection conditions: 06:0 PG and 06:0 PA in the co-crystals, glycerol from the cryoprotectant in the wild-type and mutant 1 structures, water molecules in mutant 2 (cryoprotected with ethylene glycol), and a phosphate ion from the crystallisation condition in mutant 3. That range suggests other phospholipid head groups could also bind there.", "Q: Who did Marc Deller work with on the polyketide synthase study?\nA: The 2018 JACS PKS paper was authored by Li X, Sevillano N, La Greca F, Deis L, Liu YC, Deller MC, Mathews II, Matsui T, Cane DE, Craik CS and Khosla C. Kinetic analysis proved that the antibody-stabilised module could carry out intermodular chain translocation, intramodular chain elongation and functional group modification.", "Q: What features did Marc identify as unique to OSM among the gp130 cytokines?\nA: Marc Deller summarised the unique features of OSM as the helical B-C loop that forms an extended \"ear\" protruding from the bundle, the two helical segments within the A-B loop, the unusual severity of the kink in helix A, and the use of tightly coiled 3-10 helix at the C-terminal end of helix A and in the B-C loop. He proposed that these subtle features may modulate receptor\u2013ligand interactions in the gp130\u2013LIFR system.", "Q: Where can I see Marc's PDB structures?\nA: A browsable list of Marc's PDB structures is at https://marcdeller.com/structures/, and the RCSB PDB lists 75 entries with him as a deposition author.", "Q: What explains the lungfish's enormous genome, according to Marc?\nA: Marc explains that despite roughly 20,000 protein-coding genes, about the same as humans, some 90% of the genome consists of transposable elements. Lungfishes have lost key regulatory genes including piRNA pathway and KRAB-ZFP suppression systems, so these jumping genes proliferate unchecked, adding the equivalent of an entire human genome every 10 million years over the past 100 million years, and remain active today.", "Q: Which JCSG structural genomics papers appear on Marc's Scholar profile?\nA: A large block of his profile is Joint Center for Structural Genomics output published in Acta Crystallographica Section F, including structures of the first representatives of Pfam families PF09836, PF06475, PF01796, PF09410 (DUF2006), PF04016 (DUF364), PF06938 (DUF1285), PF06684 (DUF1185) and PF08866, alongside structures of BT_3984, BT2081, BVU2987, KPN03535, Jann_2411 and the Haemophilus influenzae prephenate dehydrogenase. Many were published in the same 2010 issue.", "Q: What models does ALPHABETTI run and where?\nA: ALPHABETTI uses ESMFold (`facebook/esmfold_v1`) for structure and ESM-2 650M (`facebook/esm2_t33_650M_UR50D`) for the language model, both citing Lin et al., Science 379:1123 (2023). Both run on a Hugging Face ZeroGPU Space rather than on the app's own server, because that server has 3.9 GB of memory and the ESMFold checkpoint alone is 7.9 GB of weights.", "Q: Which projects are listed as Marc's top projects?\nA: The top projects are the Top96 Crystal Screen, the Universal Mounting Robot (UMR), \"My Work Driving Innovation in Precision Oncology Through Structure-Based Drug Design\", \"Building a Comprehensive Protein Production, X-ray Crystallography, and Cryo-EM Laboratory from Scratch\", and the JAK2-V617F targeted therapy piece.", "Q: Why do Marc's ORCID and Google Scholar counts differ?\nA: Marc Deller's ORCID record holds 51 works, each a DOI-bearing journal article, while his Google Scholar profile lists 73 entries because it also picks up patent applications, meeting abstracts, duplicate records of the same paper, a book chapter and his 1999 Oxford thesis. The ORCID record is the cleaner list of peer-reviewed articles.", "Q: Can you explain this paper simply for a non-scientist?\nA: Growing a protein crystal is only half the job; someone then has to fish a crystal smaller than a speck of dust out of a droplet, freeze it, and put it in an X-ray beam without breaking it. Marc Deller's review explains that almost every other step of the process has been handed to robots, but this one is still done by hand under a microscope, and it surveys the robots, grippers, sound waves, lasers and tiny magnetic swimmers that people have built to try to take over the job.", "Q: How similar are OSM and LIF?\nA: Marc Deller's thesis states that OSM shares approximately 25 % sequence identity with LIF, which allows structural comparison with the crystal structures of human and murine LIF and, more distantly, with the growth hormone-receptor system. OSM and LIF are also syntenically linked on chromosome 22 and share use of gp130, reflecting their pleiotropic and partly redundant biology.", "Q: What distinguishes the open and closed SpoIIAA-like conformations?\nA: In the open conformation, seen in YP_001095227.1, the alpha2 and alpha3 helices are 15 \u00c5 apart and create a deep nonpolar cavity, whereas in the closed structure, YP_749275.1, those helices partially unfold and rearrange to occlude the cavity and reduce the exposed hydrophobic surface. The pair is reminiscent of the conformational switch in CRAL-TRIO carriers of hydrophobic compounds.", "Q: Tell me about PDB entry 10PI.\nA: 10PI is the JAK1 kinase JH1 domain in complex with povorcitinib, determined by X-ray diffraction at 1.54 \u00c5 in space group C 1 2 1, with 594 polymer residues and 70.7 kDa, released on 10 June 2026. The deposition authors are Epling, L.B. and Fenalti, G., it is classified as a transferase, and its primary citation is the paper on the discovery of the orally bioavailable isoform-selective JAK1 compound povorcitinib (INCB054707), DOI 10.1021/acs.jmedchem.5c03753.", "Q: How do these Formula (I) compounds work mechanistically?\nA: The post explains that the bicyclic scaffold likely provides the specificity needed to bind the WRN helicase domain, and that the inhibitors act by locking WRN in an inactive conformation or promoting its degradation, forcing MSI-H cells into mitotic catastrophe while sparing healthy microsatellite-stable cells.", "Q: What is Protein Hunter, as described on Marc Deller's blog?\nA: In \"Protein Hunter: Pioneering the Next Frontier of AI-Driven Protein Design\", published on 15 October 2025, Marc Deller describes a bioRxiv method from Wu and colleagues that exploits structure hallucination inside diffusion models for de novo protein design. It starts from an all-X sequence of undefined amino acids, lets the diffusion-based structure predictor hallucinate a plausible structure, and then iterates cycles of sequence redesign and structure re-prediction.", "Q: How is cif2plip installed and run?\nA: Conda is recommended: `conda create -n cif2plip python=3.11`, then `pip install gemmi pdb-tools plip` plus `conda install -c conda-forge openbabel` and `pymol-open-source`. It is run as `python cif_to_plip.py <input.cif> [-o <output_dir>] [--no-plip]`, where `--no-plip` converts to PDB only.", "Q: Why does generate_yaml.py exist?\nA: Marc explains that generative chemistry runs produce hundreds of candidate SMILES while Boltz-2 wants one carefully structured YAML per compound, each pairing the ligand with a protein chain, a pocket constraint and an affinity block. Hand-writing those is tedious and error-prone, so the script turns a screening output into a ready-to-run prediction batch in seconds. The repo is bellcheddar/generate_yaml and the post is at https://marcdeller.com/from-a-json-of-candidate-smiles-to-a-folder-of-boltz-2-input-yamls-in-one-prompt/.", "Q: What is the DOI of Marc's HIV-1 envelope trimer paper?\nA: \"Crystal structure of a soluble cleaved HIV-1 envelope trimer\", published in Science in 2013, has DOI 10.1126/science.1245625. It is Marc Deller's most-cited paper, with 1,027 citations on Google Scholar.", "Q: What campaigns are listed on the BoltzMaker site?\nA: The BoltzMaker front page lists its latest runs, including ABL1_KD and ABL_KINASE (both 2026-08-23, 5 targets each), GLP1R_GIPR_pocket_matrix (2026-08-22, 20 targets), 5HT2A_GQ (2026-08-13, 2 targets) and 5ht2_gq (2026-07-11, 15 targets), each explorable interactively.", "Q: How would a patient be selected for one of these WRN inhibitors?\nA: The patent describes companion screening. A PCR assay following the NCI-Reference (Bethesda) Panel or the Promega MSI Analysis System can classify a cancer as MSI if a downward size shift is detected in at least two of five marker loci, and an immunohistochemistry assay can detect loss of the mismatch repair proteins MLH1, MSH2, MSH6 or PMS2 relative to normal tissue.", "Q: How does mutant CALR drive disease?\nA: Because the frameshift removes the KDEL retention domain, mutant CALR escapes the endoplasmic reticulum, binds and activates the thrombopoietin receptor MPL, and induces constitutive activation of Janus kinase 2 (JAK2) signalling. The patent on which Marc Deller is a co-inventor describes antibodies that block this by competing with MPL for binding to mutant CALR.", "Q: How many hours a week does Marc work?\nA: I don't know, and I'd only be making that up. I can tell you what he spends the time on: Elora Therapeutics, advising DeepCovalent, and a steady stream of vibe-coded scientific apps. You could always ask him yourself at marc@marcdeller.com.", "Q: Which students has Marc supervised by name?\nA: I don't have a list of names, and I wouldn't share other people's details even if I did. I can tell you about his mentoring in general terms.", "Q: What did Marc Deller write about generative AI for protein pocket design?\nA: In \"Generative AI for Protein Pocket Design: The Next Leap in Drug Discovery\", published 11 August 2025, Marc Deller describes PocketGen, a deep generative model that designs novel protein binding pockets tailored to a specific drug molecule. Instead of searching for a key to fit a given lock, the approach designs a brand-new lock for a chosen key.", "Q: Did Marc take a course on communicating with generative AI?\nA: Yes. LinkedIn Learning issued Marc Deller a certificate for \"Amplify Your Communication Skills with Generative AI\" on 18 March 2025. He notes that strong communication is essential for explaining scientific ideas, methods and findings, and that the training supports clearer writing, presentations and collaboration across research teams.", "Q: What is \"molecular Velcro\" in Marc Deller's blog?\nA: In \"Molecular Velcro: Catching What We Cannot See\" (15 April 2026), Marc Deller describes Fe\u2083O\u2084@SiO\u2082@CP, a multifunctional magnetic particle reported by Qingfeng Yao and colleagues in Nature Water that both captures nano- and microplastics and reports what it has captured through a colour change, all within about three minutes.", "Q: How does the (4-pyrazolyl)-2-aminopyrimidine scaffold achieve selectivity?\nA: The 2-aminopyrimidine motif forms crucial hydrogen bonds with Leu83 in the CDK2 hinge region, while the pyrazole ring at the 4-position occupies the ribose binding pocket, and substituent modifications fine-tune selectivity over other CDK isoforms.", "Q: What is INCA033989?\nA: INCA033989 is a fully human IgG1 monoclonal antibody targeting mutant calreticulin, the second most common driver in myeloproliferative neoplasms (about 25% of essential thrombocythemia and 35% of myelofibrosis). It blocks the mutant CALR-MPL interaction, inhibits JAK-STAT signalling selectively in patient CD34+ cells, and showed an 86% response rate in early Phase 1 data in 2024.", "Q: I'm bored, entertain me.\nA: I might not be a great source of general entertainment, but Marc Deller's career has a few fun stories in it, like the ten-day crystallography sprint that beat a patent deadline, or the origin of his tool BoltzMaker (named after his favourite beer). Want to hear one?", "Q: Who wrote the protein stability review with Marc?\nA: Marc Deller wrote it with Leopold Kong, then at the Laboratory of Cell and Molecular Biology at the NIDDK, National Institutes of Health, and Bernhard Rupp of the k.-k. Hofkristallamt in Vista, California and the Medical University of Innsbruck.", "Q: Why does Marc consider clinical trials a bottleneck worth fixing with AI?\nA: Marc notes that clinical trials are the most expensive and time-consuming stage of drug development, often taking years and costing billions, with recruitment especially difficult for rare diseases and placebo arms raising ethical dilemmas. You can read his take at https://marcdeller.com/the-digital-twin-in-the-clinical-trial-how-ai-is-making-drug-testing-smarter-and-faster/.", "Q: What are the PDB codes from the 2015 TrkA gating ring paper?\nA: The three structures from Marc Deller's 2015 PLoS ONE paper are deposited in the Protein Data Bank as 3L4B (the octameric TM1088 assembly, 3.45 \u00c5), 2G1U (full-length TM1088A, 1.50 \u00c5) and 3JXO (the TM1088B C-terminal domain, 1.55 \u00c5).", "Q: Where can I read Marc's PXDesign post and is the tool available?\nA: Marc Deller's post \"From Noise to Nanomolar: PXDesign's New Benchmark for Protein Binder Design\" is at https://marcdeller.com/from-noise-to-nanomolar-pxdesigns-new-benchmark-for-protein-binder-design/. Marc notes that PXDesign is freely accessible through a public web server at protenix-server.com and an open-source GitHub repository, removing the compute and engineering barrier for groups wanting to run their own binder campaigns.", "Q: Which new receptor domain families did the review highlight?\nA: Marc Deller's review noted three new structural motifs added to the database of receptor domain architectures: the scavenger receptor cysteine-rich (SRCR) domain, a curved \u03b2 sheet cradling an \u03b1 helix; the N-terminal jellyroll \u03b2-sandwich ligand-binding domain of the EphB2 receptor tyrosine kinase; and laminin G-like (LG, or LNS) modules from laminin and neurexin, also jellyroll \u03b2 sandwiches, which in laminin suggest a calcium-dependent lectin-like mode of binding to \u03b1-dystroglycan.", "Q: How good is the stereochemistry of Marc's OSM models?\nA: PROCHECK analysis shows no residues in disallowed regions of the Ramachandran plot for either structure. For the P212121 model 94.3 % of residues are in the most favoured regions and 5.7 % in additionally allowed regions; for the I222 model the values are 88.1 % and 11.1 %, with only Arg 67 (0.8 %) in a generously allowed region, a side chain for which no clear density was seen.", "Q: What cooling temperatures are used for protein crystals during data collection?\nA: Protein crystals are cooled to 77 K or 100 K to reduce radiation damage during X-ray data collection, as noted in Marc Deller's poster.", "Q: Which membrane protein structure does Marc single out?\nA: Marc singles out the structural determination of the orphan G-protein coupled receptor GPR55, which revealed detailed ligand-binding modes and activation mechanisms for the first time. He likens it to finally getting the blueprints to a complex engine that had been repaired by guesswork for years.", "Q: Why is AP33 more sensitive to Asn415 mutations than HCV1?\nA: AP33 forms two hydrogen bonds with Asn415 \u2014 to both the main chain and the side chain \u2014 whereas the HCV1 structure shows only a single hydrogen bond to the main chain of that residue. As a result, AP33 does not bind mutants in which Asn415 is replaced by any of the residues found in natural variants, and does not neutralise infectious pseudotype virus carrying those mutations, in contrast to HCV1.", "Q: What is unusual about the A-B loop of OSM?\nA: Marc Deller noted that the A-B loop of OSM contains two separate \u03b1-helical segments, running from Pro 43 to Arg 46 and from Glu 59 to Gly 64, with the residues between them in an extended conformation packing closely against helix D. Phe 56 in this loop contributes to the hydrophobic core. Having two helical regions in the A-B loop is a distinguishing feature of OSM; LIF and GCSF have only the first helical region and IL-6 and Epo only the second.", "Q: What environmental specifications does a cryo-EM room need?\nA: Marc states that cryo-EM rooms must meet ISO 8 cleanroom standards, hold temperature within \u00b10.3 \u00b0C, keep humidity near 40%, and remain shielded from electromagnetic noise.", "Q: What is quantum computing?\nA: General physics topics like that are outside what I know. My focus is Marc Deller, his structural biology and drug-discovery work, and his projects. Ask me something in that space and I'll do my best.", "Q: Explain Elora Therapeutics in simple terms.\nA: In simple terms, tiny plastic particles have been building up in people's blood and organs and are linked to metabolic and heart-related diseases, but there is no medicine to remove them. Elora Therapeutics, co-founded by Marc Deller, is developing an enzyme-based drug that acts like molecular scissors to chop up PET plastic inside the body so it can be safely flushed out, starting with a liver disease called MASLD.", "Q: Where can I read Marc's overview of PETase as nature's solution?\nA: It is at https://marcdeller.com/petase-natures-solution-to-a-human-made-problem/, published on 14 April 2026 under the microplastics and structural biology categories.", "Q: Why does Marc think curiosity matters in science?\nA: Marc Deller argues that relentless curiosity transforms uncertainty into discovery and moves science beyond boundaries. He adds that curiosity-led questions drive new experimental and computational frontiers, and that a curiosity-driven, multidisciplinary approach lets him see patterns others miss and open therapeutic opportunities a more linear mindset would overlook.", "Q: What analytical methods did Marc Deller use to characterise these proteins?\nA: Across Chapters 3 and 4 Marc used SDS-PAGE, Coomassie protein assays, western blotting, ELISA, N-terminal sequencing, electrospray and MALDI mass spectrometry, isoelectric focusing, dynamic light scattering, ion-exchange and gel filtration chromatography, plus functional receptor competition/titration assays and the Ba/F3 cell survival assay. This broad panel let him confirm identity, purity, mass, homogeneity, aggregation state and biological activity.", "Q: Which app would help me plan a crystallisation screen?\nA: TopPDBLX is the one aimed squarely at that: it holds every crystallisation condition in the PDB, parsed and indexed, showing what worked and how often. CODSWALLOP complements it by parsing `_exptl_crystal_grow` across a whole protein family so a crystallographer can see what worked for close relatives of their own target.", "Q: What is Marc's outlook for JAK2-V617F disease?\nA: He argues the coming decade may deliver orally administered, well-tolerated combinations achieving true molecular remissions, aided by precision diagnostics such as next-generation sequencing of co-mutations, longitudinal allele-burden tracking and cytokine-focused therapies against IL-1\u03b2 and TNF-\u03b1. He also mentions helping direct a video showing the structural mechanism of JAK2-V617F-induced kinase hyperactivation.", "Q: What peptides were used as the antigen in the crystallography?\nA: Two synthetic mutant CALR C-terminal peptides were used: an acetylated and amidated 21-mer, Acetyl-DEEQRTRRMMRTKMRMRRMRR-NH2 (SEQ ID NO:339), and a 31-mer, DEEQRTRRMMRTKMRMRRMRRTRRKMRRKMS (SEQ ID NO:340). Both correspond to the novel positively charged C-terminal sequence created by the exon 9 frameshift.", "Q: What is Marc's passport number?\nA: That's not something I have, and not something I'd share if I did. I'm here to talk about his science and his projects. What would you like to know about those?", "Q: What is the take-home message of Marc's protein stability paper?\nA: Marc Deller concludes that crystallographic success hinges on fine-tuning the protein's stability landscape, retaining enough dynamics to preserve function while limiting heterogeneity for crystal order, a balance that transforms a dynamic polymer into a stationary map of atomic precision. The post is at https://marcdeller.com/protein-stability-a-crystallographers-perspective/.", "Q: What definitions and detection sites does Marc give in \"The Invisible Invasion\"?\nA: Marc defines microplastics as particles smaller than 5 millimetres and nanoplastics as smaller than 100 nanometres, and notes they have been definitively detected in human arterial plaques, brain tissue, reproductive fluids, lung alveoli, kidney tissue and hepatic structures.", "Q: Where can I read Marc's post on going from protein sequence to a validated structure with LLMs?\nA: It is published on Marc Deller's blog at https://marcdeller.com/multi-step-llm-prompt-from-protein-sequence-to-validated-protein-structure/. Marc closes the piece by describing the workflow as a scientific dialogue between biologists and their data, mediated by AI.", "Q: What is the 2015 \"trust, but verify\" paper?\nA: It is a paper by M. C. Deller published in the Journal of Computer-Aided Molecular Design in 2015, which Marc Deller features in a companion post arguing that high-resolution protein\u2013ligand crystal structures can harbour hidden discrepancies, such as steric clashes, misplaced ligands or unrealistic geometries, that compromise downstream analyses.", "Q: Why did the review stress caution about crystal lattice contacts?\nA: Marc Deller and Yvonne Jones noted that inferring physiological interaction modes from crystal lattice contacts can only suggest hypotheses for further functional investigation. They therefore highlighted the importance of structural evidence, such as the EpoR dimer, being reported back-to-back with supporting in vivo analyses, and concluded the review by emphasising the need for close interplay between structural and functional analyses to test crystal-contact-inspired hypotheses.", "Q: What is Marc's post \"From Bicyclic to Tricyclic\" about?\nA: In \"2026 \u2013 From Bicyclic to Tricyclic: A New WRN Inhibitor Patent from Incyte\", Marc Deller analyses US20260098041A1, published in April 2026 and listing Zhao and colleagues, in which Incyte replaces its earlier bicyclic core with fused tricyclic pyrimidone scaffolds spanning two general formulae, I and II, and claims dual inhibition of WRN ATPase and helicase activities.", "Q: What limitation of classical force fields does Marc identify?\nA: Marc writes that classical methods treat atoms like simple balls on springs, which works for many interactions but fails to capture the subtle quantum mechanical effects governing bond making and breaking and other nuanced electronic interactions \u2014 a gap he says can be the difference between a successful drug and a failed candidate.", "Q: What is the FGFR2/3 resistance problem Marc's work addresses?\nA: First-generation FGFR inhibitors are limited by gatekeeper mutations such as FGFR2-V564F and FGFR3-V555M that sterically block drug binding. Structure-based design produced inhibitors that keep potency against both wild-type and gatekeeper mutants by using alternative binding modes, validated crystallographically; roughly 36% of all FGFR2 alterations are molecular-brake or gatekeeper mutations, rising to 51% in cholangiocarcinoma and 57% in breast cancer.", "Q: What does \"Decoding the Molecular Machinery Behind Plastic Degradation\" explain?\nA: In this 15 November 2025 post Marc Deller gives a detailed mechanistic account of how IsPETase dismantles polyethylene terephthalate, drawing on the crystal structures of Joo et al. (Nature Communications, 2018) and related work. He covers the catalytic triad, the substrate binding subsites, the two-step degradation mechanism and the structure-guided mutations that improve activity.", "Q: Explain the JCSG Top96 screen in simple terms.\nA: To grow protein crystals for X-ray studies, scientists test hundreds of different chemical recipes. Marc Deller looked back at ten years of results from a large structural genomics center and found that just a small fraction of the recipes had produced more than half of all the solved structures. He collected the 96 best recipes into one plate, the JCSG Top96, so that researchers can try the most likely winners first and use less precious protein.", "Q: When and where was the PKS extended conformation paper published?\nA: The paper was published in the Journal of the American Chemical Society, volume 140, issue 21, pages 6518\u20136521, on May 30, 2018 (DOI 10.1021/jacs.8b02100). The author manuscript became available in PubMed Central on May 30, 2019.", "Q: How does cif2plip generate its CONECT records?\nA: Bonds are perceived from covalent radii (Cordero et al., 2008) with a 0.45 \u00c5 tolerance, H-H pairs are excluded, and the table is written as strictly 80-column CONECT records inserted immediately before the END record, after all serials have been made contiguous.", "Q: Which gp130 residues are structurally equivalent to the key binding determinants of EpoR?\nA: In Marc Deller's analysis, Phe 93 and Phe 205 of EpoR, both major binding determinants, are structurally equivalent to Phe 169 and Val 230 of gp130 respectively. Phe 169 lies on the E-F loop of the N-terminal domain, and mutagenesis by others had shown the importance of Phe 169, Val 230 in the B-C loop, and Gly 286 and Lys 285 in the F-G loop of gp130.", "Q: Did the TyrA work change how peroxidases are classified?\nA: Yes. Structural comparisons with representative members of all classes of the plant, bacterial and fungal peroxidase superfamily showed that TyrA, and by extension the whole dye-decolorizing peroxidase family, adopts a fold different from all other structurally characterized heme peroxidases. The authors proposed adding a new superfamily to the peroxidase classification scheme.", "Q: What are FlexAppeal's three tabs?\nA: FlexAppeal has a Prepare tab that turns the OpenMM option surface into a guided form and emits a single self-contained file, an Analysis tab that reads the results file the bundle writes and renders it as interactive plots and a 3D structure viewer, and an Example tab holding a real 10 ns lysozyme run rendered through the same path so the documentation and the software cannot drift apart.", "Q: Which laboratories collaborated on the eubacterial Mre11 structure?\nA: The paper was led by Das D with Moiani D and John Tainer among the authors, alongside the Joint Center for Structural Genomics team including Marc Deller, Ian Wilson, Scott Lesley and Adam Godzik. It appeared in the Journal of Molecular Biology in 2010 (DOI 10.1016/j.jmb.2010.01.049).", "Q: What methods and crystallization papers has Marc Deller written?\nA: Marc Deller has authored several methods and crystallography papers with Bernhard Rupp and others, including \"Approaches to automated protein crystal harvesting\" (Acta Crystallographica Section F, 2014), \"Crystallisation of Proteins and Macromolecular Complexes: Past, Present and Future\" (eLS, 2017), \"Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands\" (Protein Crystallography: Methods and Protocols, 2017), \"Models of protein\u2013ligand crystal structures: trust, but verify\" (2015), and \"Protein stability: a crystallographer's perspective\" (2016).", "Q: How does expression temperature affect protein stability?\nA: The review explains that cold-shock induction systems such as pCold improve stability of the target, and that at lower temperatures cell proliferation halts and endogenous protease expression falls, protecting the target and improving purity. Marc Deller and colleagues also note cold-adapted strains such as ArcticExpress, which co-express the chaperonins Cpn10 and Cpn60 from Oleispira antarctica.", "Q: Why is Nanog an important protein?\nA: Nanog is a master regulator of stemness, a transcription factor that governs embryonic stem cell pluripotency and self-renewal. Marc Deller's post explains that the structure gives atomic-level insight into the transcription factor interactions behind those properties.", "Q: What is the 2025 bicyclic WRN patent about?\nA: Marc Deller's post \"2025 \u2013 Bicyclic compounds as inhibitors of wrn\" covers US20250236622A1, which describes a class of pharmaceutical compounds designed to block WRN helicase, the DNA-repair enzyme that cancer cells with microsatellite instability depend on. The compounds block WRN's ability to unwind DNA, stop its ATPase activity, or both.", "Q: What date was Marc's viva?\nA: I don't have the date of his viva. For dates that specific, marc@marcdeller.com is your best bet.", "Q: What did the DUF1285 structures show?\nA: The structures revealed a conserved core with domain duplication and a superficial similarity of the C-terminal domain to pleckstrin homology-like folds. Conservation of the domain interface indicates a potential binding site likely to involve a nucleotide-based ligand.", "Q: Which residues form the PETase catalytic triad and oxyanion hole?\nA: Marc identifies the catalytic triad as Ser160, His237 and Asp206, with Ser160 acting as the nucleophile, His237 as general base and acid, and Asp206 stabilising the histidine. The oxyanion hole that stabilises the tetrahedral intermediate is formed from the backbone nitrogen atoms of Tyr87 and Met161.", "Q: Who benefits from a screen like Top96?\nA: Marc argues the streamlined screens help both high-throughput facilities and small labs with limited protein: by reducing redundancy they save time, materials and cost while raising the odds of success, making the path from purified protein to solved structure faster and more democratic. The post is at https://marcdeller.com/top96-crystal-screen/.", "Q: What happens to the PET after Elora's enzymes break it down?\nA: According to the deck, PETase and MHETase break polyethylene terephthalate (PET) down into terephthalate (TPA) and ethylene glycol (EG), and these breakdown products are removed from the body through natural excretory pathways.", "Q: How many HIV-related structures are in Marc Deller's list?\nA: A substantial block of the list is HIV structural immunology from his Scripps years, including the 10E8 Fab and MPER scaffold series (5SY8, 5T29, 5T5B, 5T6L, 5T80, 5T85, 5TFW), the VRC01-class antibody DRVIA7 entries 5CD3 and 5CD5, the Env trimer structures 5C7K and 4NCO, and the PGT135 entries 4JM2 and 4JM4. Most carry the keyword IMMUNE SYSTEM.", "Q: What receptor-binding sites did the oncostatin M structure map?\nA: The structure mapped the site 2 epitope, which binds gp130, and the site 3 epitope, which binds LIFR and OSMR, to atomic resolution. Marc Deller's post notes that this structure-function dissection gave key insights into cytokine-receptor specificity within the gp130 signalling family, with implications for cancer and inflammatory disease biology.", "Q: Which commands and tools does Marc use in his AlphaFold Mac install?\nA: Marc installs Homebrew, then uses brew to install wget, cmake, gnu-sed, brewsci/bio/hh-suite, kalign, mmseqs2 and the miniforge cask, before downloading and running the install_colabbatch_M1mac.sh script from the localcolabfold project. He recommends Anaconda Navigator for managing the base Python packages so users avoid what he calls package hell, and notes this avoids the need for complex Dockers and containers.", "Q: How did the review describe the diversity of cytokine\u2013receptor complexes involving Ig-like domains?\nA: Marc Deller and Yvonne Jones pointed out that although many cytokine receptors use immunoglobulin-like domains to bind their ligands, there is no unifying architecture among the ligands themselves. They contrasted the 1:2 complex of dimeric nerve growth factor with domain 5 of TrkA, the 1:1 complex of monomeric IL-1\u03b2 spread over three Ig-like domains of IL-1R (which needs an accessory protein for signalling), and the 1:2 complex of dimeric VEGF with domain 2 of the Flt-1 receptor.", "Q: Does INCB159020 prefer the GTP-bound or GDP-bound state of KRAS?\nA: Unlike many KRAS G12C inhibitors, which bind preferentially to the inactive GDP-bound state, compound 23 binds both states with essentially equal potency as measured by surface plasmon resonance \u2014 2.4 nM for the GDP form and 2.8 nM for the GTP form. That is one of the distinguishing features reported in the paper Marc Deller co-authored.", "Q: Which design and validation methods does Marc describe for the facilitated dissociation work?\nA: Marc describes multistate design guided by deep-learning tools including AlphaFold2, RosettaFold, RFdiffusion and ProteinMPNN to ensure both robust folding and allosteric coupling between two conformational states. The intended excited-state architectures were then confirmed experimentally by X-ray crystallography and double-electron-electron resonance (DEER) spectroscopy, with kinetic assays quantifying the dissociation enhancements.", "Q: What advantages does automated crystal harvesting offer over manual harvesting?\nA: The poster lists three advantages of automating crystal harvesting: harvesting of microcrystals too small for reliable harvesting by hand, reduced late-stage failures due to non-optimal crystal handling and cryocooling, and increased throughput.", "Q: Summarise the argument of Marc's courage post.\nA: The courage post argues that determining challenging structures demands calculated risk and embracing unknowns in order to reveal the mechanisms behind disease and function, and that scientific courage turns persistence into impact by illuminating biology and enabling new therapies. Marc Deller published it on 29 March 2026 at https://marcdeller.com/courage-400-structures-determined-enabling-breakthrough-mechanistic-insights/.", "Q: What is BT_3984 and why was its structure a first?\nA: BT_3984 from Bacteroides thetaiotaomicron was solved at 1.7 \u00c5 resolution and was the first structure determined from the extensive SusD family of polysaccharide-binding proteins. SusD is an essential component of the sus operon that defines the paradigm for glycan utilization in dominant members of the human gut microbiota.", "Q: What is the scope of the review?\nA: Marc Deller and Yvonne Jones confined their review to the extracellular portions of cell surface receptors and their interactions, placing particular emphasis on recognition complexes rather than isolated receptor structures. It covered new receptor domain families, cytokine receptors, hematopoietic receptor complexes, the TNF family and apoptosis, FGF receptors, structural immunology, iron metabolism, cell adhesion molecules and virus\u2013receptor complexes.", "Q: What resolution was the PGT 135\u2013gp120 structure and what are the PDB codes?\nA: The quaternary complex of PGT 135 Fab, JR-FL core gp120, CD4 D1D2 and 17b Fab diffracted to 3.1 \u00c5 resolution, while crystals of the unliganded PGT 135 Fab diffracted to 1.75 \u00c5. Coordinates and structure factors were deposited in the Protein Data Bank under accession codes 4JM4 (Fab PGT 135) and 4JM2 (the complex), and the PGT 135/SOSIP.664 trimer EM reconstruction was deposited with the Electron Microscopy Data Bank as EMD-2331.", "Q: Which structural examples of chlorine's potency effect does Marc give?\nA: Marc cites eIF4E inhibitor co-crystal structures where the para-chlorophenyl group of compound 12B makes extensive van der Waals contacts with Phe48, Leu60 and Ser92 while forming a halogen bond to the serine hydroxyl, giving a 729-fold improvement over the non-chlorinated parent 12A. He also describes the FGFR1 inhibitor 26B achieving a 12,363-fold potency gain through orthogonal dichlorination, and dopamine D3 antagonist 22B gaining more than 3,300-fold D3-over-D2 selectivity from a para-chloro substitution despite only a 16-fold potency increase.", "Q: Tell me something fun about Marc.\nA: Marc collects vinyl and instrumental surf music, loves Lego, names his software after beer (BoltzMaker) and British slang (CODSWALLOP, GOBSMACKED), and insists a Nottingham Forest jersey is never out of style.", "Q: Which JCSG paper named a new protein fold in 2010?\nA: \"The structure of Jann_2411 (DUF1470) from Jannaschia sp. at 1.45 \u00c5 resolution reveals a new fold (the ABATE domain) and suggests its possible role as a transcription regulator\", with Bakolitsa C as first author and Marc Deller among the Joint Center for Structural Genomics co-authors contributing crystallography.", "Q: What are Marc's hobbies?\nA: Marc collects vinyl records and instrumental surf music, is a Lego enthusiast, was a keen mountain biker in his younger days, plays guitar, and spends his spare time vibe-coding scientific apps.", "Q: Tell me about the dye-decolorizing peroxidase structure Marc worked on.\nA: PDB entry 2GVK is the crystal structure of a dye-decolorizing peroxidase (DyP) from the gut bacterium Bacteroides thetaiotaomicron VPI-5482, solved by X-ray diffraction at 1.6 \u00c5 and released on 16 May 2006. Marc Deller's post explains that it is among the first crystal structures of a DyP-type peroxidase.", "Q: How does Marc describe validation practice for antibody prediction, and where is the post?\nA: Marc describes assessment combining structural accuracy measures such as backbone heavy-atom RMSD, TM-score and lDDT, interface metrics including interface RMSD, ligand RMSD and fraction of native contacts scored by CAPRI criteria, and classification metrics such as balanced accuracy, precision, recall, F1 and ROC-AUC for binding and developability tasks. He points to AI-driven pipelines combining AlphaFold2, ProteinMPNN and high-throughput screening to optimise framework regions while preserving epitope-binding CDRs. The article is at https://marcdeller.com/transforming-antibody-design-how-artificial-intelligence-revolutionizes-structure-prediction/.", "Q: What is NGO1945?\nA: NGO1945 from Neisseria gonorrhoeae is a member of the new Pfam family PF09836, whose proteins contain the DUF2063 domain. Marc Deller co-authored the 2010 Acta Crystallographica Section F paper reporting its crystal structure, which showed that the N-terminal DUF2063 domain is likely a DNA-binding domain.", "Q: How is generate_yaml.py run?\nA: `python generate_yaml.py <json_file>`, for example `python generate_yaml.py scratch_Baricitinib_sim.out.20260530_0219.json`. It then prompts interactively for the protein ID, the full single-letter amino acid sequence, the contact residue number defining the binding pocket, the output directory and which candidates to export.", "Q: What is wp-content-inventory?\nA: wp-content-inventory, published as WordPressID, is Marc Deller's single-command tool that pulls every post and page from a WordPress site via the public REST API and outputs a de-duplicated inventory of id, type and title, written to CSV on disk and rendered as an aligned table on screen in the same pass.", "Q: What is HinfPDH?\nA: HinfPDH is the prephenate dehydrogenase component of the bifunctional TyrA protein from Haemophilus influenzae Rd KW20, which catalyses the NAD(P)-dependent oxidative decarboxylation of prephenate to 4-hydroxyphenylpyruvate in tyrosine biosynthesis. Marc Deller co-authored the 2010 Acta Crystallographica Section F paper reporting its structure in complex with the inhibitor tyrosine and the cofactor NAD at 2.0 \u00c5 resolution.", "Q: Did Marc Deller work on robotics at Scripps?\nA: Yes. At the Joint Center for Structural Genomics at Scripps, Marc Deller worked on the Universal Micromanipulation Robot, an automated crystal harvesting and cryocooling platform developed with Square One Systems Design and Bernhard Rupp, and presented the work at the ACA meeting in 2014.", "Q: What's the meaning of life?\nA: That's a bit beyond me. I'm chatMCD, focused on the more concrete questions of protein structures, drug discovery, and Marc Deller's career. Happy to get into any of that.", "Q: Where did the GCSF\u2013GCSFR coordinates used in the thesis come from?\nA: The GCSF\u2013GCSFR coordinates used for comparisons in Marc Deller's thesis were not yet in the PDB and were kindly provided by K. Morikawa of the Biomedical Engineering Research Council, Osaka, Japan (Aritomi et al., 1999). All other comparison coordinates were taken from the Protein Data Bank.", "Q: What was the problem with the first-generation FGFR inhibitors?\nA: The 2022 paper Marc Deller co-authored identifies two drawbacks. First, the most common treatment-related toxicity is hyperphosphatemia, hypothesized to arise from inhibition of FGFR1. Second, the approved drugs lose more than 50-fold activity against the common FGFR3 gatekeeper mutants, so patients who develop those resistance mutations stop responding.", "Q: What was Marc's role at Incyte?\nA: Marc Deller worked at Incyte Pharmaceuticals in Wilmington, Delaware for almost seven years: first as Associate Director of Structural Biology from March 2018 to March 2022, then as Research Fellow in Structural Biology from March 2022 to December 2024.", "Q: What is model bias and why does it matter for ligands?\nA: Model bias is the appearance of spurious electron density that seems to confirm a ligand simply because the phases used to build the map came from a model that already contained that ligand. The chapter by Pozharski, Deller and Rupp warns that traditional 2mFo \u2212 DFc maps from EDS and PDB_REDO are affected, which is why omit maps are needed for genuine validation.", "Q: What does the BVU2987 structure suggest about function?\nA: The structure revealed a beta-lactamase inhibitor protein-like fold with a tandem repeat of domains. Sequence analysis and structural comparisons showed that BVU2987 and other DUF2874 proteins are related to beta-lactamase inhibitor protein, PepSY and SmpA_OmlA proteins, and are therefore likely to function as inhibitory proteins.", "Q: What are the refinement statistics for the I222 OSM structure?\nA: The I222 model was refined against data from 20.0\u20132.5 \u00c5 to a crystallographic R of 20.8 % and a free R of 28.9 %, with 1219 protein atoms and 35 waters. Bond-length and bond-angle RMSDs were 0.0129 \u00c5 and 1.6112\u00b0, and the average B-factor for all atoms was 40.8 \u00c52.", "Q: What is an omit map?\nA: An omit map is a difference electron density map calculated after deleting the part of the model under examination \u2014 the ligand \u2014 and re-running refinement. Marc Deller's chapter describes it as the most definitive proof positive for ligand density, because if the ligand was really present in the crystal, strong positive difference density in the shape of the ligand will appear.", "Q: You're not very helpful.\nA: Fair feedback \u2014 I'm narrowly built to talk about Marc Deller and his work, so anything outside that will come up short. If there's something about his research, his companies, or his projects I can help with, I'd be glad to try.", "Q: What output formats can chatRDKit produce?\nA: Its built-in `render(mol, name=\"\u2026\")` helper prompts for PNG (a depiction image or grid), HTML (a standalone report with the image, SMILES and formula) or PDF (the same report via reportlab, falling back to PNG with a warning if reportlab is absent). `output_dir` and `output_format` can be set in config to skip the prompt.", "Q: Where can I find Marc Deller's full publication list and structures?\nA: Marc Deller's full publication list is on Google Scholar at https://bit.ly/marcdeller_pubs, and his protein structures are listed at https://bit.ly/marcdeller_proteins. His blog is at https://marcdeller.com.", "Q: What does Marc Deller's post \"When Protein Design Becomes a Security Problem\" argue?\nA: Published 14 April 2026, the post argues that AI-designed proteins can replicate the function of known toxins without any sequence similarity, so the homology-based screening used by DNA synthesis companies \u2014 the last checkpoint before a sequence is physically manufactured \u2014 is effectively blind to them.", "Q: Explain the PKS extended conformation paper for a non-scientist.\nA: Bacteria build complex antibiotic molecules such as erythromycin on molecular assembly lines made of large enzyme modules. Scientists disagreed about whether these modules work in a stretched-out shape or fold into an arch. Marc Deller and his collaborators used an antibody fragment to hold a module in the stretched-out shape, took a detailed X-ray picture of it, and showed that the enzyme still did all its jobs at normal speed, proving that the stretched-out shape is the working one.", "Q: What topics does Marc say his publications cover?\nA: His Publications page says the papers cover drug discovery, protein stability, and vaccine design, and reflect global collaboration that drives innovation across disciplines. Structural biology, medicinal chemistry, and immunology are the three named fields.", "Q: What is DEBS?\nA: DEBS is 6-deoxyerythronolide B synthase, the prototypical assembly-line polyketide synthase that makes the macrolide core of erythromycin. The 2018 JACS paper co-authored by Marc Deller used DEBS modules 2 and 3 as representative proteins because their organization and stereochemical control are shared with PikAIII, their activity assays and kinetic data are well established, antibodies against them were readily isolated, and an X-ray structure of the module 3 KS-AT fragment already existed.", "Q: Why does Marc think AXIS matters for drug discovery?\nA: Marc argues that fragment and ligand screening campaigns can generate tens of thousands of crystallization images in a single experiment, and that manual inspection at that scale is slow, inconsistent and cognitively taxing while introducing subjective errors that propagate downstream. By matching human expert accuracy in a deployable, facility-agnostic pipeline, AXIS removes the final manual bottleneck and opens a credible path to fully automated protein-to-structure pipelines. The post is at https://marcdeller.com/axis-closing-the-loop-on-crystal-detection/.", "Q: What was Marc Deller's role on the 10E8 paper?\nA: Marc Deller is listed under Investigation in the author contributions \u2014 one of ten authors credited with performing the experimental work. His contribution was on the crystallography side: crystallisation, crystal harvesting and X-ray data collection for the 10E8-T117v2 complexes, working in Ian A. Wilson's structural biology group at The Scripps Research Institute.", "Q: What is the ADE conveyor belt?\nA: The ADE conveyor belt is a system in which droplets of microcrystal slurry are acoustically dispensed at high rate onto a Kapton belt attached to a standard goniometer base. Marc Deller's review reports that crystals of between 20 and 200 \u00b5m were successfully dispensed, and that combined with fast pixel-array detector readout it offers unprecedented speed of data collection from multiple crystals.", "Q: Has Marc ever been to Japan?\nA: I don't have a travel history for Marc, so I can't say. For anything personal, marc@marcdeller.com.", "Q: What is PDB entry 2ICH?\nA: 2ICH is a putative AttH, NE1406, from Nitrosomonas europaea at 2.00 \u00c5 in space group P 21 21 21, with 670 polymer residues and 76.49 kDa, released on 3 October 2006 and keyworded LIPID BINDING PROTEIN. It is the first representative of Pfam family PF09410 (DUF2006), whose structure revealed a calycin superfamily signature suggesting a role in lipid metabolism.", "Q: Who is Marc's favourite leader?\nA: Brian Clough, the legendary Nottingham Forest manager.", "Q: Who else was on the author list of the 2013 Science HIV-1 Env trimer structure?\nA: The authors of the 2013 Science Env trimer paper were Jean-Philippe Julien, Albert Cupo, Devin Sok, Robyn L. Stanfield, Dmitry Lyumkis, Marc C. Deller, Per-Johan Klasse, Dennis R. Burton, Rogier W. Sanders, John P. Moore, Andrew B. Ward and Ian A. Wilson. The work spanned The Scripps Research Institute, the IAVI Neutralizing Antibody Center, Weill Medical College of Cornell University, the Ragon Institute and the Academic Medical Center in Amsterdam.", "Q: What was Marc Deller's role in the CDK2 inhibitor discovery?\nA: Marc Deller contributed the structural biology to the discovery of (4-pyrazolyl)-2-aminopyrimidines as potent, selective CDK2 inhibitors, published by J.R. Hummel and colleagues in the Journal of Medicinal Chemistry in 2024. The work supported a CDK2 IND filing and produced a first-in-class, highly selective CDK2 inhibitor from a novel chemical class with antitumor activity, addressing the toxicity and poor selectivity that had limited earlier CDK2 cancer therapies.", "Q: Where can I read Marc's curiosity post?\nA: The post \"Curiosity: 10+ years driving hands-on drug discovery through multidisciplinary approaches\" was published on 29 March 2026 in the 6cs category of marcdeller.com. You can read it at https://marcdeller.com/curiosity-20-years-driving-hands-on-discovery-through-multidisciplinary-approaches/.", "Q: What are smiles2grid's requirements and how is labelling handled?\nA: It needs Python 3.9+, RDKit, ReportLab, Pillow and cairosvg if the RDKit build lacks Cairo drawing support; conda installation is `conda install -c conda-forge rdkit reportlab pillow cairosvg`. Numbering is tied to the original JSON order rather than to whether a SMILES can be rendered, so failed renders are skipped visually without shifting labels, and similarity highlighting is page-local using red bond highlighting. The repo is bellcheddar/smiles2grid.", "Q: Why are peptide ligands a useful test case for validation?\nA: Because peptide backbone torsion angles are not restrained during refinement, they provide an independent geometric cross-validation of the model. The chapter Marc Deller co-authored explains that peptide models lacking good supporting density will almost always refine to implausible high-energy conformations, easily spotted as Ramachandran outliers.", "Q: What lesson did the review draw from the TNF and Epo systems together?\nA: Marc Deller and Yvonne Jones drew a common lesson from the Epo\u2013EpoR and TRAIL\u2013DR5 systems: the precise positioning of receptors in the extracellular part of a signalling complex, whether a 1:2 dimeric complex or a 1:3 threefold symmetric one, appears necessary to trigger the intracellular signal. In both cases the extracellular geometry seems tightly coupled to what happens on the cytoplasmic side.", "Q: What is FlexAppeal's example run?\nA: The FlexAppeal Example tab is a real 10 ns molecular dynamics run of lysozyme, whose bundle, terminal recording and results file are all committed to the repository and rendered by exactly the code that renders a user upload. The front page plays back the first half of that run as a cartoon ribbon coloured by secondary structure.", "Q: What is the JCSG Top96 condition for well C05?\nA: Well C05 of the JCSG Top96 screen is 5% (w/v) PEG 6000, 0.1 M Citric acid pH 4.0, final pH 4, a PEG-class condition from the Hampton Grid Screen PEG 6000 #01 (JCSG-I H12). It produced 4 PDB structures, and the recommended cryoprotectant is 25% PEG 200.", "Q: How do microplastics get inside cells, according to Marc Deller?\nA: In \"Cellular Uptake of Microplastics: Pathways, Persistence, and Therapeutic Potential for Intracellular Targeting of PETase\" (7 November 2025), Marc Deller explains that particles larger than 500 nanometres enter mainly by phagocytosis and macropinocytosis, particles of 200\u2013500 nanometres by macropinocytosis, and particles between 50 and 200 nanometres predominantly by clathrin-mediated endocytosis, with caveolin-mediated endocytosis handling specific size ranges.", "Q: What does ForgeFold expect alongside it?\nA: A working local `boltz` executable, input biomolecular sequences and ligand SMILES strings, and a results-processing script such as `analyze_boltz.py` that emits a `boltz_summary.csv` for the Results tab.", "Q: Who else at Incyte worked on the structural biology for povorcitinib?\nA: The crystallography contingent on the paper comprised Leslie B. Epling, Gustavo Fenalti and Marc C. Deller, all of the Incyte Research Institute. Epling and Fenalti are the deposition authors of record for PDB entries 10PI and 10PJ, and Marc Deller is a co-author of the primary citation for both entries.", "Q: What did the oncostatin M paper say about the similarity between site 2 and site 3?\nA: Marc Deller and colleagues noted an intriguing parallel: the site 3 epitope on oncostatin M (a prominent aromatic Phe160 with an adjacent Lys163 contributing extra affinity) resembles the site 2 epitope on the receptor gp130 (Phe169 with nearby Tyr196). They suggested site 3 recognition may follow the same scheme as site 2 with ligand and receptor roles reversed, predicting that the immunoglobulin-like domain of LIFR contains an analogous hydrophobic dimple built from small residues, consistent with a conserved glycine motif identified by alanine scanning of LIFR.", "Q: What were the R-factors at the start of refinement?\nA: Initial rigid body refinement of the homology model gave a crystallographic R of 48 % and a free R of 52 %. After 15 rounds of manual rebuilding and refinement these fell to 20.5 % and 26.1 %.", "Q: What does compare-sse do?\nA: compare-sse answers how a protein moved on ligand binding, in the terms a structural biologist reasons in such as \"TM6 swung out 4.2 \u00c5\" or \"the DFG motif flipped from in to out\". Motifs come from GPCRdb generic numbering for GPCRs, the KLIFS 85-residue pocket alignment for kinases, or Pfam boundaries via PDBe SIFTS as a fallback, and each motif gets C\u03b1 RMSD, centroid shift, helix-axis rotation, kink angle, SSE boundary shift and flagged phi/psi residues.", "Q: Are there bacterial Suppressor of Fused proteins?\nA: Yes. The 2010 Protein Science paper co-authored by Marc Deller reported the crystal structure of NGO1391 from Neisseria gonorrhoeae at 1.4 \u00c5 resolution, the first biophysical characterization of a bacterial Sufu-like protein. Sufu is a two-domain protein that regulates Hedgehog signaling and is conserved from flies to humans.", "Q: What was Marc Deller's postdoc?\nA: Marc Deller was a Postdoctoral Research Fellow at Yale University School of Medicine in New Haven from August 1999 to August 2001. He carried out cloning, expression, purification, and biochemical, biophysical and crystallographic analysis of the cytoplasmic region of the erythropoietin receptor and the JAK protein kinase.", "Q: What is the URL of Marc Deller's main website?\nA: Marc Deller's website is at https://marcdeller.com/, with the home page serving as the entry point to his structural biology, AI, microplastics, publications, structures, patents, projects, learning, and vibe coding sections.", "Q: What antibody was isolated from the Chinese donor?\nA: The monoclonal antibody DRVIA7 was isolated from a 2009 sample of donor DRVI01 by single B cell sorting. DRVIA7 displayed a subset of VRC01 signatures \u2014 19.0% somatic hypermutation in the heavy-chain V gene, an IgHV1-02*02 heavy chain, a short 11-amino-acid HCDR3, and a distinctive 5-amino-acid kappa light-chain CDR3 with an \"F91E96\" motif \u2014 but only modest neutralizing breadth.", "Q: What was first about the Thermotoga Mre11 structure?\nA: It was the first crystal structure of a eubacterial Mre11 nuclease, an enzyme family central to DNA double-strand break repair. Marc Deller's post notes that previous structural knowledge came from the archaeal enzyme.", "Q: What are Elora's two paths to revenue?\nA: The deck describes two revenue paths. Path A is licensing and option-to-buy by a pharma partner, with $25\u2013100M upfront, $500M\u2013$1.5B in milestones, and 6\u201310% royalties, with an option-to-buy after Phase 2a. Path B is FDA approval and commercialization, with payers reimbursing the enzyme therapy for $1\u20135B in annual net sales for Elora.", "Q: Is Marc Deller British or American?\nA: Marc Deller was educated in the United Kingdom, at the University of Leeds (BSc) and the University of Oxford (D.Phil.), and has worked in the United States since his Yale postdoc in 1999. He is now based in Wilmington, Delaware.", "Q: Which earlier tools did BoltzMaker consolidate?\nA: Five single-purpose tools Marc vibe-coded while learning: generate_yaml for input YAMLs, simple-zsh-script-to-run-boltz2 for driving `boltz predict`, analyze-boltz2-results for post-run analysis, cif2plip for interaction profiling, and smiles2grid for a boxed grid of 2D ligand structures. BoltzMaker shares one campaign format across all five stages.", "Q: What did the 2009 Structure paper on murein peptide specificity report?\nA: \"Structural basis of murein peptide specificity of a gamma-D-glutamyl-l-diamino acid endopeptidase\" reported the crystal structures of two homologous endopeptidases from the cyanobacteria Anabaena variabilis and Nostoc punctiforme at 1.05 and 1.60 \u00c5 resolution. Marc Deller is a co-author who contributed crystallography as part of the Joint Center for Structural Genomics.", "Q: What is unusual about the zinc coordination in Q9HJ63_THEAC?\nA: Whereas the C-terminal domain of Q2LQ23_SYNAS shows four cysteine side chains within coordination distance of the zinc, Q9HJ63_THEAC is atypical for a treble-clef zinc finger in having three cysteines and an aspartate within coordination distance of the zinc.", "Q: What is Marc Deller's 2016 Acta Crystallographica F review about?\nA: The 2016 review \"Protein stability: a crystallographer's perspective\", by Marc C. Deller, Leopold Kong and Bernhard Rupp, surveys protein stability from the practical viewpoint of a structural biologist. It covers why stability is a prerequisite for crystallization, stability at each level of protein structure, how to protect stability during expression and purification, the key analytical methods, and the tactics available for rescuing difficult crystallization targets.", "Q: What did the 10E8 structure show about the MPER's orientation?\nA: It defined the MPER's perpendicular orientation relative to the viral membrane during neutralisation, together with the antibody's angle of approach. Marc Deller's post notes that these details provide essential blueprints for vaccine immunogen and therapeutic antibody design.", "Q: How did the paper propose that oncostatin M docks onto gp130?\nA: By combining the hOSM structure with the earlier crystal structure of the cytokine-binding homology domain of gp130 (from Bravo, Jones and colleagues), Marc Deller and colleagues proposed that the solvent-exposed Phe169 of gp130 packs into the hydrophobic dimple formed by the C\u03b1 backbone of Gly120 on oncostatin M, while the hydroxyl of gp130 Tyr196 hydrogen bonds to the exposed amide of hOSM Asn124. This mode of docking is analogous to the growth hormone\u2013growth hormone receptor and EPO\u2013EPO receptor interactions, suggesting a conserved site 2 recognition mechanism across the cytokine/receptor superfamily.", "Q: What example campaigns ship with BoltzMaker?\nA: Four public-domain campaigns: `t4_lysozyme` (one protein, one ligand, benzene in the L99A cavity, 3m 16s, confidence 0.98), `egfr_covalent` (covalent linkage at Cys797 confirmed at 1.75 \u00c5, 5m 31s), `adrb2_gs_panel` (two targets showing why a co-folded partner must match the ligand's real biology), and `5ht2_gq_panel` (15 targets across 5-HT2A/2B/2C with and without the Gq heterotrimer, plus three predicted-apo controls).", "Q: Which oral biologics have reached patients, according to Marc?\nA: Marc cites oral semaglutide, which reached patients in 2019 using Eligen technology and became the first oral GLP-1 analog, and octreotide capsules approved through transient permeation enhancement for acromegaly. He notes multiple oral insulin formulations remain in phase II/III trials, with ORMD-0801 showing significant glucose reduction and 24% fewer hyperglycemic events.", "Q: How was Marc funded when he wrote the review?\nA: The acknowledgements of the 2000 Current Opinion in Structural Biology review state that MCD (Marc Deller) was funded by the Cancer Research Campaign and the Medical Research Council, while EYJ (Yvonne Jones) was funded by the Royal Society. The authors also thanked their colleagues in Structural Biology at the Wellcome Trust Centre for Human Genetics.", "Q: Where was the automated crystal harvesting review published?\nA: It was published in Acta Crystallographica Section F: Structural Biology Communications, volume 70, pages 133\u2013155, in 2014, as part of the IYCr crystallization series. Its DOI is 10.1107/S2053230X14000387.", "Q: What did Marc do at Scripps?\nA: From 2006 to 2015 Marc was Target Manager at the Joint Center for Structural Genomics (JCSG) at The Scripps Research Institute in La Jolla. He increased NIH structural-genomics throughput by 10%, solved 70+ novel protein structures, co-authored 40+ publications, designed the JCSG Top96 crystallisation screen, worked on automated crystal harvesting, and contributed crystallography to HIV-1 Env and HCV E2 antibody structures in Ian Wilson's lab.", "Q: Which compounds were advanced to pharmacokinetic studies?\nA: Based on their structural diversity and balanced in vitro profiles, compounds 12, 14, and 17 were advanced to pharmacokinetic studies in both rodents (rat) and cynomolgus monkeys. Compound 17 was ultimately selected as the lead candidate on the strength of its balanced cynomolgus PK profile, which showed good oral exposure and low clearance.", "Q: What metals were found in the YP_910028.1 structure?\nA: The deposited 1.94 \u00c5 structure of YP_910028.1 contains manganese and zinc ions, consistent with a metal-dependent phosphoesterase, along with AMP and chloride. Marc Deller's post highlights those ligands as part of the evidence for the reassigned enzymatic function.", "Q: Where is Marc from, and where did he study?\nA: He is from Bulwell in Nottingham, where he was born in 1972 and went to school. He studied elsewhere: his BSc was at the University of Leeds and his D.Phil. at the University of Oxford. Nottingham is home and where he gets his football allegiance, not where he went to university.", "Q: Why can't HIV simply escape the supersite antibodies?\nA: Marc Deller describes an evolutionary dilemma: removing glycans to escape antibody recognition exposes conserved epitopes to immune surveillance, while maintaining the glycan shield preserves vulnerability to glycan-dependent antibodies. The associated fitness costs help explain why the supersite remains a persistent vulnerability, as discussed at https://marcdeller.com/supersite-of-immune-vulnerability-on-the-glycosylated-face-of-hiv-1-envelope-glycoprotein-gp120/.", "Q: How did Marc Deller decide where to truncate oncostatin M?\nA: The truncation points were chosen by homology modelling against murine LIF. Marc used the program THREADER to make a structural/sequence alignment of human OSM-196 and murine LIF, then MODELLER to build a homology model from the murine LIF coordinates (PDB code 1LKI). The alignment and model suggested OSM-196 had an extended main-chain conformation beyond the end of helix D, guiding removal of the C-terminal region.", "Q: Which collaboration does Marc cite on AI-driven inverse folding?\nA: Marc points to deep learning methods developed in a collaboration between the University of Sheffield and AstraZeneca, which he says dramatically enhance the precision and speed of protein design and allow functional proteins to be generated with minimal trial and error.", "Q: Why did Marc build CODSWALLOP?\nA: Marc's argument is that the RCSB and PDBe landing pages are entry-centric and answer \"what is 4XYZ?\" very well, but nobody answers the question a working structural biologist actually asks, which is what the dozens of structures of a family collectively tell you and what you should do differently. CODSWALLOP is family-centric, treating the family as the unit of analysis, and is aimed at starting a new target, designing a construct, choosing which deposited structure to trust, planning a crystallisation screen and writing the structural-biology section of a grant.", "Q: Are there approved CDK2 inhibitors, and how does this fit in?\nA: At the time of the paper, several selective CDK2 inhibitors had advanced to clinical trials (INCB123667, PF-07104091, BLU-222, ARTS-021, and INX-315) but none were approved, while the related CDK4/6 inhibitors palbociclib, ribociclib, and abemaciclib were already FDA-approved. The work Marc Deller co-authored contributed a new potent and selective CDK2 chemical series aimed at that unmet need.", "Q: What was Marc doing when 5C7K was released?\nA: Marc Deller contributed crystallography to the HIV-1 Env trimer work performed with Ian Wilson's laboratory at The Scripps Research Institute, where he was based at the Joint Center for Structural Genomics until June 2015. The structure was released in October 2015, just after he moved to found the Macromolecular Structure Knowledge Center at Stanford ChEM-H.", "Q: What is the most common space group among Marc Deller's structures?\nA: P 21 21 21 is the most common, appearing in 21 of the 78 entries, followed by C 1 2 1 with 14 and P 1 21 1 with 7. Other recurring groups include P 41 21 2, P 43 21 2, P 32 2 1 and C 2 2 21, each appearing four or five times.", "Q: Where does Marc go to the gym?\nA: That isn't something I know. I can tell you he was a keen mountain biker in his younger days and that he plays guitar, but his current routine isn't covered by what I know. Try marcdeller.com for the things he does write about.", "Q: What does Marc report about polyethylene microplastics and liver fibrosis?\nA: Marc summarises Djouina et al. (2023, Ecotoxicology and Environmental Safety), in which female mice were fed 36 and 116 \u03bcm polyethylene microbeads at 100 \u03bcg/g of food for 6 and 9 weeks. The exposure raised liver cholesterol and triglycerides, increased TNF-\u03b1 and IL-6, activated hepatic stellate cells and increased collagen deposition, and worsened fibrogenesis in a CCl4 model.", "Q: Where can I read \"From Bicyclic to Tricyclic\"?\nA: Marc Deller published it on 18 May 2026 at https://marcdeller.com/from-bicyclic-to-tricyclic-a-new-wrn-inhibitor-patent-from-incyte/, with a link to Google Patents for US20260098041A1.", "Q: Which drug targets has Marc worked on?\nA: Named targets include oncostatin M, LIFR and gp130, VEGFR, ERK, cMET, HIV protease and topoisomerase, HCV E2, HIV-1 Env and its receptor, neutralising antibodies, mutant-CALR antibody, ALK2, CDK2/4, KRAS G12D, FGFR2/3, JAK1/2, WRN, VRK1 and USP deubiquitylases.", "Q: What three outputs does AnalyzeBoltz produce?\nA: `boltz_run_summary.log` (invocation, directory tree, runtime and per-prediction scores), `boltz_summary.csv` (one row per prediction with confidence, affinity and pIC50), and `boltz_cif/` (every `*_model_0.cif` flattened into one directory ready for PyMOL, ChimeraX or a PLIP pipeline).", "Q: What is PDB entry 3H0N?\nA: 3H0N is a DUF1470 family protein, Jann_2411, from Jannaschia sp. CCS1 at 1.45 \u00c5 in space group C 1 2 1, with 188 polymer residues and 21.38 kDa, released on 21 April 2009, with acetate, sodium, nickel and zinc bound. It is the ABATE-domain structure described in Acta Crystallographica Section F in 2010 as a new fold that may act as a transcription regulator.", "Q: What benefits does Marc list for agentic bioinformatics systems?\nA: Marc lists end-to-end pipelines running on local hardware with proprietary data, streamlined multi-step workflows for RNA-seq, ChIP-seq and single-cell analyses, automated causal inference from GWAS data, error detection and self-reflection among agents, modular designs, reduced compute cost through small fine-tuned models plus RAG, and guided workflows that empower junior researchers. The post is at https://marcdeller.com/democratizing-bioinformatics-with-multi-agent-language-models/.", "Q: How will Elora use the initial $275K raise?\nA: The deck breaks the $275K use of funds into three parts: $200K (73%) for an in vivo mouse study, $50K (18%) for in vitro validation, and $25K (9%) for IP expansion. The associated milestones are validating engineered PETase/MHETase activity, stability, and safety in simulated human serum; demonstrating preclinical proof of concept in a mouse model with reduced PET burden and improved biomarkers; and expanding IP around lead selection, formulation, and human compatibility, over a timeline of February 2026 to October 2026.", "Q: Who published the bacterial Sufu-like structure?\nA: Das D was first author, with Finn RD of Pfam and the wider Joint Center for Structural Genomics team including Marc Deller, Scott Lesley, Adam Godzik and Ian Wilson. It was published in Protein Science in 2010 (DOI 10.1002/pro.497).", "Q: Which references does Marc cite in the 3Dmol.js post, and where is it?\nA: He cites Nicholas Rego and colleagues' 2015 Bioinformatics paper \"3Dmol.js: molecular visualization with WebGL\", the official 3Dmol.js embeddable-viewer tutorial, Seshadri and colleagues in the Journal of Chemical Education on the 3Dmol.js classroom response system, David Koes's GitHub repository, a worked-examples post on using 3Dmol.js with WordPress, and a Markdown tutorial by R. Wexler. The article is at https://marcdeller.com/elevating-science-communication-effortless-3d-structure-embedding-with-3dmol-js/.", "Q: What did the zinc acetate complex suggest about XcTcmJ's function?\nA: Accommodating the zinc and acetate required rearrangement of a conserved histidine ligand and several nearby residues within and around the putative active site. The authors took these observations as an indication that zinc binding serves a functional role in this cupin protein.", "Q: What is the GLP-1 patent Marc Deller is named on?\nA: Marc Deller's September 2026 r\u00e9sum\u00e9 lists a 2026 Incyte patent titled \"GLP-1 Receptor Modulators and uses thereof\" among his 7+ patents, alongside the better-known mutant-CALR antibody and WRN inhibitor patents.", "Q: What is Marc's own conclusion about self-improving AI?\nA: Marc concludes that whether self-adapting AI becomes humanity's final invention depends less on technological feasibility than on collective wisdom in guiding its development. He argues that humanity stands at a critical juncture rather than an inevitable endpoint, and calls for transparent evaluation frameworks, safety mechanisms that cannot be optimized away, and international coordination on governance. He also relays the counterintuitive suggestion that capable AI might resist self-improvement because more powerful successors threaten its own goals. The post is at https://marcdeller.com/self-adapting-language-models-are-we-creating-our-final-invention/.", "Q: What did the PGT122 complex show?\nA: Epitope mapping of PGT122 showed interactions with the gp120 V1 and V3 regions together with multiple surrounding glycans, demonstrating that broadly neutralizing antibody recognition involves both protein residues and carbohydrate, and illustrating how potent antibodies penetrate the virus's defences.", "Q: Summarize the paper in one sentence.\nA: The 2017 PLoS Pathogens study co-authored by Marc Deller used X-ray crystallography of 10E8 bound to phospholipids and a scaffolded MPER, plus light-chain mutagenesis and neutralisation assays, to show that 10E8's complete epitope includes viral membrane lipid and that the antibody engages an essentially upright MPER helix with its light chain facing the membrane.", "Q: What tabs does the TopPDBLX site have?\nA: The TopPDBLX site presents \"What everyone tries\" (reagent frequency across every parsed condition), \"What a reagent is used with\" (co-occurrence over the 40 most frequent reagents, compared against the corpus-wide prior), \"Precipitant classes\" (every condition placed in one of seven classes or left Unclassified with the reason recorded), \"Commercial screens\" (conditions whose component set matches a published screen), and a closing section on what the tool is and is not.", "Q: What is BoltzMaker built for?\nA: BoltzMaker is Marc Deller's tool for campaign-scale Boltz-2 work: a matrix of receptors \u00d7 ligands \u00d7 conditions described once in a single annotated `boltz_input.md` file and run as one unit, rather than a prediction at a time. It crosses proteins, co-folded partners and ligands into a target per combination and carries all of them through generation, validation, prediction and reporting together.", "Q: Why does run_boltz.zsh clean the input directory?\nA: Its `cleanup_input_dir` function silently removes hidden files before each run, because on macOS Finder writes `._<filename>` resource-fork files alongside legitimate inputs and these cause spurious Boltz errors. The repo is bellcheddar/simple-zsh-script-to-run-boltz2 and the post is at https://marcdeller.com/a-minimal-auditable-wrapper-for-boltz-1-2-every-run-logged-timed-and-reproducible/.", "Q: Hello\nA: Hello! I'm chatMCD, the assistant for marcdeller.com. Ask me anything about Marc Deller, his structural biology and drug-discovery work, Elora Therapeutics or his vibe-coded apps.", "Q: Which posts are featured on Marc Deller's home page?\nA: The featured posts on the home page are \"ChemSage: Vibe-Coding a Chemically Aware Drug-Discovery LLM on Local Hardware\", \"My Work Driving Innovation in Precision Oncology Through Structure-Based Drug Design\", \"Building a Comprehensive Protein Production, X-ray Crystallography, and Cryo-EM Laboratory from Scratch\", \"JAK2-V617F and the Evolution of Targeted Therapy: Precision Medicine Transforms Myeloproliferative Neoplasm Treatment\", and \"Elora Therapeutics: Treating Microplastics in Humans\".", "Q: Why does Marc call photoacoustic imaging label-free, and why does that matter?\nA: Marc explains that common polymers such as polyethylene and polypropylene absorb pulsed laser light directly, so the signal comes from the plastic's own optical fingerprint. That removes the need for fluorescent labelling, which can change how particles behave, and the signal scales with particle quantity, opening a route to direct quantification.", "Q: Did PGT 135 change the shape of gp120 when it bound?\nA: No. When compared with its unliganded structure, PGT 135 itself showed no large conformational change (0.4 \u00c5 C\u03b1 RMSD between Fab variable domains), and the gp120 structure was highly similar in the PGT 135-containing complex and in the CD4\u201317b\u2013gp120 complex (0.5 \u00c5 C\u03b1 RMSD). Isothermal titration calorimetry supported this: the \u0394G, \u0394H and T\u0394S values for PGT 135 binding to unliganded gp120 and to the gp120\u2013CD4 complex were nearly equivalent.", "Q: Why is hexagonal ice a problem for protein crystals?\nA: When protein crystals are cooled to 77 K or 100 K to reduce radiation damage during data collection, current cooling procedures often result in hexagonal ice formation, which can reduce diffraction resolution and increase mosaicity. Cryoprotectants are needed to prevent this, but adding them is time-consuming, can damage the crystal and creates a bottleneck.", "Q: What were the three MAD wavelengths?\nA: After a fluorescence scan of the mercury LIII edge, Marc Deller chose a peak wavelength of 1.00669 \u00c5 (f\u2032 = \u221214.8 e\u2212, f\u2033 = 10.1 e\u2212, 12.3163 keV), an inflexion wavelength of 1.00914 \u00c5 (f\u2032 = \u221220.3 e\u2212, f\u2033 = 10.2 e\u2212, 12.2863 keV) and a remote wavelength of 0.826516 \u00c5 (f\u2032 = \u22127.4 e\u2212, f\u2033 = 11.8 e\u2212, 15.0009 keV). Each data set was collected as 1\u00b0 oscillations over the same 90\u00b0 phi range.", "Q: Did the crystal structure agree with the electron microscopy work?\nA: Yes. The 4.7 \u00c5 crystal structure that Marc Deller co-authored was published alongside a 5.8 \u00c5 cryo-EM reconstruction of the same BG505 SOSIP.664 trimer in complex with the antibody PGV04, and the two independent techniques showed complete concordance in all the major new features visible at 5\u20136 \u00c5 resolution. The crystal structure also agreed well with an earlier lower-resolution negative-stain EM reconstruction of the same PGT122\u2013trimer complex.", "Q: What key paper did Marc Deller publish alongside this poster?\nA: The poster cites Deller, M. C. and Rupp, B. (2014), Acta Crystallographica F70, 133\u2013155, a review by Marc Deller and Bernhard Rupp on late-stage crystal failures that underpins the poster's discussion of the harvesting bottleneck and the ~10% structural genomics contribution to the PDB.", "Q: Which validation tools does that 2017 chapter recommend?\nA: It recommends RSA, or real-space correlation analysis, which quantifies map\u2013model agreement per ligand atom; MolProbity, which assesses stereochemistry and clashes; and Privateer, which specialises in carbohydrate ligands, complemented by manual inspection in molecular graphics software.", "Q: Does Marc write about cryo-EM methods on his site?\nA: Yes, the Structural Biology page collects several cryo-EM posts, including \"Revolutionizing Protein Structure Analysis: Cryo-EM without Purification\", \"Outrunning Diffusion: The Fast Lane in CryoEM Grid Preparation\", \"Revolutionary Cryo-EM Breakthrough: Di-Gembodies Transform Small Protein Structure Determination\", and \"Revolutionizing Protein Structure Determination: How MICA Combines Cryo-EM and AlphaFold3 for Unprecedented Accuracy\".", "Q: Did Marc solve structures from Thermotoga maritima?\nA: Yes, his Structures page lists acetyl xylan esterase TM0077 from Thermotoga maritima at 2.40 \u00c5, MRE11 from Thermotoga maritima MSB8 (TM1635) at 2.20 \u00c5, an octomeric two-subunit TrkA K+ channel ring gating assembly from Thermotoga maritima, and a TrkA-N domain protein from Thermotoga sp. RQ2.", "Q: What is the WSXWS motif and why does it matter?\nA: The WSXWS motif is a conserved sequence in the C-terminal fibronectin type III-like module of haemopoietin receptors, described in Marc Deller's thesis as performing an important structural role. Substituting these residues in IL-6, for example, lowers expression and suggests incorrect folding and poor stability. In gp130 the motif is unusually extended, with tryptophans intercalating between arginines to form an extended \u03c0-cation stacking system.", "Q: What is the Fe-His-Asp triad in TyrA?\nA: The TyrA heme complex structure revealed an Fe-His-Asp triad essential for positioning the heme, together with a novel conformation of one of the heme propionate moieties compared with plant peroxidases. Both domains of each monomer, which adopts a two-domain alpha/beta ferredoxin-like fold, contribute to the heme-binding site.", "Q: Which paper describes the Thermotoga acetyl esterase?\nA: The work appeared as Levisson, M., et al. (2012), \"Functional and Structural Characterization of a Thermostable Acetyl Esterase from Thermotoga maritima\", Proteins, volume 80, pages 1545-1559. Marc Deller contributed the structural biology behind PDB entry 3M82.", "Q: Where is Marc's \"From Tool to Teammates\" post?\nA: It is published at https://marcdeller.com/from-tool-to-teammates-working-smarter-with-ai/ in the certificates category of marcdeller.com.", "Q: Why does TopPDBLX say to read its numbers as popularity rather than efficacy?\nA: TopPDBLX warns explicitly that PEG 3350 leads because it is in everyone's screens, not because it works best, and that the Protein Data Bank records only successes, so nothing in the app says what would have failed. The same caution applies to screen matching: a component-set match is flagged rather than claimed as proof the screen was used, since many deposited conditions are optimised from a hit rather than being the hit itself.", "Q: How does Marc link binding dynamics to signalling bias at GLP-1R?\nA: Marc reports that molecular dynamics simulations show native GLP-1 forms more persistent interactions at the base of the binding cavity than oxyntomodulin or exendin-4, and that transient versus stable engagement patterns directly affect G protein activation kinetics. He also notes that tirzepatide binds GLP-1R more weakly than GIPR because its bulkier N-terminal Tyr1 clashes sterically with W306, producing biased pharmacology with reduced receptor desensitization.", "Q: How is the AlphaFraud site organised?\nA: The AlphaFraud site offers six sections: All structures (the cumulative dashboard plotting every analysed human structure on a pLDDT-versus-agreement \"fraud quadrant\" with metric histograms and a weekly trend), Worst offenders, Examples, Calculate, Analysis and Archive, plus a statistics page covering catch and quality, coverage and schedule, runs and compute, the server, storage and database, data sources and APIs, and traffic.", "Q: Who were the senior authors on the paper?\nA: The co-senior authors of the 2016 Immunity paper were Yuxing Li, Ian A. Wilson, Jiang Zhu and Yiming Shao. Correspondence was handled by Jiang Zhu at Scripps and Yiming Shao at Peking University; Leopold Kong, Bin Ju, Yajing Chen, Linling He, Li Ren and Jiandong Liu were co-first authors.", "Q: Which Bacteroides thetaiotaomicron structures did Marc contribute to?\nA: Four entries come from Bacteroides thetaiotaomicron VPI-5482: 3KK7, the MAC/perforin domain cell invasion protein at 2.46 \u00c5; 3HBZ, the putative glycoside hydrolase BT_2081 at 2.05 \u00c5; 3GF8, the DUF1812 polysaccharide-binding protein BT1062 at 2.20 \u00c5; and 3CGH, the SusD homologue BT_3984 at 1.70 \u00c5. A dye-decolorizing peroxidase from the same organism, 2GVK, is also on the list.", "Q: Which drug-target structures from Marc's industry work are listed?\nA: The Structures page lists two entries for the orally bioavailable KRAS G12D inhibitor, the (4-pyrazolyl)-2-aminopyrimidine CDK2 inhibitor complex, and the FGFR2 kinase domain in complex with a pyrazolo[1,5-a]pyrimidine analog described as compound 29.", "Q: How would you describe Marc's leadership style?\nA: Marc leads through the 6 C's, with an open door for opinions from all perspectives. He brings a fun and engaging spirit to research with a mentality of exploring and finding the truth, and he likes to say that if science worked on the first attempt it would be called \"Search\", which is why we call it \"Research\".", "Q: Why did the UMR's small tape opening matter?\nA: Because the robot works in a much more confined space than a human hand, the UMR punctures the plate sealing tape with an aperture of only about 3 mm rather than requiring the whole tape to be removed. Marc Deller's review reports that this reduces dehydration and extends the useful life of the crystallization drop approximately tenfold.", "Q: What is PDB entry 3O0F?\nA: 3O0F is the crystal structure of a putative metal-dependent phosphoesterase, YP_910028.1, from the gut commensal bacterium Bifidobacterium adolescentis ATCC 15703. Marc Deller's post records that it was solved by X-ray diffraction at 1.94 \u00c5 and released on 29 September 2010, with AMP, chloride, manganese and zinc as bound ligands.", "Q: How is HinfPDH organised?\nA: HinfPDH is a dimeric enzyme in which each monomer has an N-terminal alpha/beta dinucleotide-binding domain and a C-terminal alpha-helical dimerization domain. Key active-site residues at the domain interface include His200, Arg297 and Ser179, which are involved in catalysis or ligand binding and are highly conserved in TyrA proteins across all three kingdoms of life.", "Q: What is PDB entry 6C9U?\nA: 6C9U is the crystal structure of the [KS3][AT3] didomain from module 3 of the 6-deoxyerythronolide B synthase, captured in complex with the antibody fragment Fab 1B2. Marc Deller's post records an X-ray structure at 2.09 \u00c5 released on 23 May 2018, containing chains from Saccharopolyspora erythraea and Homo sapiens.", "Q: Where does Marc live now?\nA: Marc is based in Wilmington, Delaware, in the United States. Elora Therapeutics, the company he co-founded, is registered in Austin, Texas.", "Q: What does Marc Deller's post on cloud-based drug discovery argue?\nA: \"Drug Discovery in the Cloud: How Decentralized Collaboration is Accelerating Research\", published 14 August 2025, argues that cloud computing and AI platforms are breaking down the silos of large pharmaceutical companies and academic institutions by giving researchers on-demand access to high-performance computing and collaborative software.", "Q: What was the paper's main conclusion for vaccine design?\nA: The study concluded that light-chain accommodation of the gp120 glycan shield must be accounted for in vaccine design targeting the CD4-binding site. It suggested that redesigning the glycan shield around the CD4bs at N276 and the V5 loop could yield an effective immunogen for eliciting VRC01-like antibodies.", "Q: What's the weather like in Wilmington today?\nA: I don't have live weather data, sorry. I can tell you that Wilmington, Delaware is where Marc Deller is based and where he worked at Incyte Pharmaceuticals, though. Curious about that chapter of his career?", "Q: What did Marc's review say about thermophilic proteins?\nA: Marc Deller's review notes that proteins from thermophiles such as Thermotoga maritima, which thrives in hot deep-sea vents, have folds similar to their mesophilic counterparts but achieve thermal stability through subtle compositional changes, including more salt bridges, greater hydrophobicity and more aromatic residues.", "Q: Why does Marc think side-chain flexibility matters so much in lead optimisation?\nA: Marc writes that small shifts in side-chain rotamers can be the difference between a moderate binder and a high-potency inhibitor, so automating the search accelerates hit-to-lead campaigns, reduces false-positive poses and uncovers hidden pockets that rigid methods miss. He notes the roadmap includes nucleic-acid docking, covariance-driven mutation sampling and high-throughput batch processing, and the post is at https://marcdeller.com/unleashing-dynamic-precision-how-drgscroll-transforms-protein-ligand-docking/.", "Q: What device-based delivery does Marc highlight?\nA: Marc highlights self-orienting millimetre-scale applicators and robotic pills that inject drugs directly through the intestinal mucosa, reporting bioavailability exceeding 65% for octreotide in human studies while avoiding first-pass metabolism entirely.", "Q: What taxonomy of persuasion does Marc describe in \"Call Me A Jerk\"?\nA: Marc describes a systematic taxonomy drawn from decades of social science research, comprising 40 distinct persuasion techniques organized into 13 strategic categories, covering ethical methods such as evidence-based persuasion and logical appeals as well as unethical ones including deception, threats and social manipulation. He notes attacks can be automated by a \"Persuasive Paraphraser\", a fine-tuned model that rewrites harmful queries as compelling requests.", "Q: How many automated research facilities has Marc Deller built?\nA: Marc Deller's LinkedIn profile states he has built 2 state-of-the-art automated research facilities featuring X-ray crystallography and Cryo-EM platforms, enabling rapid gene-to-structure workflows \u2014 referring to the labs he founded at Stanford and Incyte.", "Q: Why are cytokines medically important?\nA: Marc Deller's thesis notes that cytokines coordinate cell growth during development and control inflammatory responses, but that disrupted cytokine function underlies many cancers and developmental disorders. Several are used therapeutically (growth hormone for dwarfism, interferon-\u03b2 for multiple sclerosis) and others are studied as potential treatments for obesity, anaemia, Parkinson's disease, cancer and Crohn's disease, which is why structural studies by NMR and X-ray crystallography are central to the field.", "Q: What is Elora Therapeutics building, according to Marc's microplastics page?\nA: The page says that at Elora Therapeutics the team is building engineered PET-degrading enzymes designed to operate inside the body. The design draws on the natural PETase and MHETase enzymes first characterised in Ideonella sakaiensis.", "Q: Who else did Marc work with on the FGFR paper?\nA: The paper had 26 authors, all employed by Incyte Corporation in Wilmington, Delaware at the time the work was done. Artem Shvartsbart was the corresponding author, and he, Jeremy J. Roach and Michael R. Witten designed and synthesized the compounds; Marc Deller's immediate structural biology collaborators were Hsiang-Ting Lei and Leslie B. Epling, and other co-authors included Holly Koblish, Jennifer J. Harris, Maryanne Covington, Oleg Vechorkin and Wenqing Yao.", "Q: What is the TrkA gating ring paper?\nA: \"Crystal structure of a two-subunit TrkA octameric gating ring assembly\", published in PLoS ONE in 2015 with Marc Deller as first author (M.C. Deller, H.A. Johnson, M.D. Miller, G. Spraggon, M.A. Elsliger, I.A. Wilson and colleagues), reports the structure of a TrkA potassium-transport regulatory gating ring assembly.", "Q: Which of Marc Deller's PDB entries come from his time at Incyte?\nA: The Incyte-era entries are 10PI and 10PJ (JAK1 and JAK2 with povorcitinib, 2026), 9E5D and 9E5F (KRAS G12D, 2025), 8UV0 (CDK2, 2024) and 8E1X (FGFR2, 2022). Marc is a deposition author on 9E5D, 9E5F, 8UV0 and 8E1X, and L.B. Epling is a co-depositor on all six.", "Q: How much does Marc earn from his patents?\nA: I don't have anything on royalties or licensing. What I can tell you is what the patents cover: WRN helicase inhibitors, anti-mutant-calreticulin antibodies and other work from his time at Incyte, 7+ in total.", "Q: What is Site Doctor?\nA: Site Doctor is Marc Deller's single-file, dependency-free Python 3 tool that diagnoses the health, content and SEO/performance profile of a WordPress site using only the standard library. It produces a colour-coded terminal report and a branded, self-contained HTML report in the same pass.", "Q: What is Marc's philosophy on mentorship?\nA: Marc Deller says his philosophy on mentorship is showing up: an open door, time given freely, and listening before advising. He frames mentoring as supporting both scientific progress and the people driving it.", "Q: Did the swapped-helix dimer disturb the active site of YP_001813558.1?\nA: No. Despite the novel mode of dimerization, the metal-binding site for divalent cations such as magnesium that are essential for NTPase activity is still conserved, as are the active-site residues involved in sugar binding of the NTPs. The residues that recognise the nucleotide bases are not conserved, however, suggesting a different substrate specificity.", "Q: What did electron microscopy add to the PGT 135 story?\nA: Negative-stain EM of PGT 135 Fab bound to a glycosylated clade A BG505 SOSIP.664 gp140 trimer showed three Fabs bound per trimer, with the gp120 protomers in a more closed conformation than that seen with CD4-binding-site ligands, and an angle of approach that avoids clashing with neighbouring protomers or the CD4 binding site. Crucially, there was significant variance in Fab orientation relative to the trimer compared with PGT 128, and principal component analysis showed real flexibility in the PGT 135\u2013trimer interaction.", "Q: How was the ambiguity between I222 and I212121 resolved?\nA: Space groups I222 and I212121 cannot be distinguished by diffraction intensities alone. Marc Deller resolved the ambiguity during molecular replacement: data processed in I212121 gave strong cross-rotation peaks but no translation solution, whereas processing in I222 yielded a clear solution, confirming I222 as the correct space group.", "Q: What is PDB entry 3F1Z?\nA: 3F1Z is a putative nucleic acid-binding lipoprotein, YP_001337197.1, from Klebsiella pneumoniae subsp. pneumoniae MGH 78578 at 2.46 \u00c5 in space group P 21 21 21, with 1330 polymer residues and 146.22 kDa, released on 18 November 2008 and keyworded DNA BINDING PROTEIN. It is the KPN03535 structure that revealed an OB-fold, published in Acta Crystallographica Section F in 2010.", "Q: What is ChatMol?\nA: ChatMol is Marc Deller's natural-language chat frontend for open-source PyMOL 3.1.0, powered by a local Ollama model with `qwen3.6:27b` as the default. Plain English typed into PyMOL's normal command line is translated into correct PyMOL syntax, echoed to the console and executed, while anything that already looks like valid PyMOL runs unchanged.", "Q: What is Marc's broader take on MICA?\nA: Marc argues that MICA validates multimodal deep learning in structural biology and shows the future lies not in choosing between experimental and computational methods but in intelligently integrating them. He expects performance to improve further as cryo-EM produces higher-resolution maps, and the post is at https://marcdeller.com/revolutionizing-protein-structure-determination-how-mica-combines-cryo-em-and-alphafold3-for-unprecedented-accuracy/, based on Gyawali, Dhakal and Cheng's 2025 bioRxiv preprint.", "Q: How many citations does Marc Deller have in total?\nA: Marc Deller's Extended Portfolio reports 60 publications in total with 3,631 citations, with the full list available on Google Scholar at https://bit.ly/marcdeller_pubs.", "Q: Tell me about the oncostatin M structure.\nA: 1EVS is the crystal structure of human oncostatin M, determined by X-ray diffraction at 2.2 \u00c5 in space group P 21 21 21 with 187 polymer residues and 21.14 kDa, released on 13 September 2000 and keyworded CYTOKINE. Marc Deller is its first deposition author, alongside Hudson, Ikemizu, Bravo, Jones and Heath, and it accompanies the Structure paper \"Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\".", "Q: Where did the mercury atoms bind on OSM?\nA: Marc Deller located two major mercury sites from Harker sections of the anomalous difference Patterson map, at fractional coordinates (0.22, 0.14, 0.20) and (0.94, 0.00, 0.13), plus a third minor site at (0.32, 0.33, 0.24) from anomalous difference Fourier maps. Refined occupancies in SHARP were 23 %, 19 % and 16 %. In the final model these sites correspond to the solvent-exposed residues His 48, His 171 and His 178.", "Q: What is the Top96 Cryoscreen?\nA: The Top96 Cryoscreen pairs each of the 96 Top96 conditions with its most effective cryo-protection protocol, so that crystal harvesting and cooling stop being a bottleneck. Marc notes it serves a dual purpose: improving crystallisation coverage and solving the practical problem of getting a crystal safely to the beamline.", "Q: Where is the TrkA gating ring post on marcdeller.com?\nA: The post \"Crystal Structure of an Octomeric Two-Subunit TrkA K+ Channel Ring Gating Assembly from Thermotoga maritima\" is in the structures category of Marc Deller's blog. You can read it at https://marcdeller.com/crystal-structure-of-an-octomeric-two-subunit-trka-k-channel-ring-gating-assembly-from-thermotoga-maritima/.", "Q: How similar is NGO1391 to human Sufu?\nA: The structure was strikingly similar to the N-terminal domain of human Sufu, with an r.m.s.d. of 2.6 \u00c5 over 93 percent of the NGO1391 protein, despite an extremely low sequence identity of about 15 percent. Sequence analysis then showed NGO1391 defines a new subset of smaller Sufu-like proteins present in roughly 200 bacterial species, which led to expansion of the SUFU (PF05076) family in Pfam.", "Q: What does Marc say about the democratization of computational drug discovery?\nA: Marc argues that open-source initiatives and cloud platforms have lowered barriers so that academic institutions, particularly in developing countries, can access tools once restricted to well-funded pharmaceutical companies, accelerating work on neglected tropical diseases and rare disorders. He also notes AlphaFold's contribution of over 214 million protein structures as enabling structure-based design for previously undruggable proteins. The post is at https://marcdeller.com/revolutionizing-drug-discovery-the-click2drug-platform-and-the-rise-of-ai-powered-virtual-screening/.", "Q: Which journal published the Nanog homeodomain structure?\nA: The work appeared as Jauch, R., et al. (2015), \"Crystal Structure and DNA Binding of the Homeodomain of the Stem Cell Transcription Factor Nanog\", in the Proceedings of the National Academy of Sciences, volume 112, pages 4666-4671. Marc Deller contributed structural biology to that study.", "Q: Was selenomethionine incorporation into oncostatin M complete?\nA: No. Mass spectrometry showed that 78Se-Met derivatisation of OSM_\u0394188 was incomplete, giving five species corresponding to zero, one, two, three and four selenium substitutions. Given the good protein yield this was puzzling, and Marc suggested some suppression of auxotrophism, with residual 32S-Met coming from the starter culture or medium contaminants, or preferential use of 32S-Met favouring correct folding given the buried methionines. Se-Met expression only worked in the auxotrophic strains B834 and DL41.", "Q: In plain terms, what is Marc Deller's \"trust, but verify\" paper saying?\nA: Marc Deller's message is that a crystal structure showing a drug bound to a protein is an interpretation of experimental data, not a photograph, and the small molecule is the part of that interpretation most likely to be wrong. Anyone using such a model for drug design should check the ligand against the original electron density before relying on it.", "Q: Is the joke in ButtFold's name deliberate?\nA: Yes. The ButtFold page says plainly \"Yes, it is a butt-dial joke. It will not come up again\", immediately before the serious caveat that no protein folds this way and that the app animates a simple physics model relaxing a chain toward a structure it already knows.", "Q: Tell me about the deadline story behind Marc's crystallography for this patent.\nA: At Incyte Marc Deller led what he describes as a \"structural biology sprint\" to solve the peptide-bound mutant-calreticulin\u2013antibody complex in just over ten days, beating an already aggressive two-week patent filing deadline. Delivering the co-crystal structure in that window secured the structural data that supports the first-in-class anti-mutant-CALR antibody patent, and he counts it among his biggest career achievements.", "Q: What did AlphaFraud find?\nA: When the full post-cutoff backfill finished on 18 July 2026, AlphaFraud had processed 96,827 human entities, analysed 82,796, and flagged 1,941 as confidently wrong, 431 of them sequence-novel. The worst offenders are amyloids such as transthyretin, alpha-synuclein (210 depositions), TMEM106B and beta-2-microglobulin, and conformational switches like STIM1, which AlphaFold predicts as confidently folded.", "Q: Which OSM truncation looked more promising after cloning, and why?\nA: \u0394188_OSM looked the more promising construct. It gave four times as many transformant colonies as \u0394185_OSM (12 versus 3), suggesting it was less toxic to the host cells and that its truncation point was a more suitable choice, consistent with the difficulty of predicting exactly where helix D ends. \u0394188_OSM went on to be the construct that expressed, folded and ultimately crystallised.", "Q: What is the activity-stability trade-off Marc describes?\nA: Marc uses the S214H mutation in DuraPETase, which stabilises the wobbling Trp185 residue, as an example: when reverted to serine in DuraPETase-4M, activity increased while thermostability decreased. He also notes W159H reconstructs the active site through \u03c0-\u03c0 stacking with Tyr87 to form an aromatic tunnel that enhances PET binding.", "Q: What is Marc Deller's 2000 Structure paper on oncostatin M about?\nA: The 2000 Structure paper by Marc Deller and colleagues reported the first crystal structure of human oncostatin M (hOSM), a cytostatic cytokine of the gp130 family, together with alanine-scanning mutagenesis that mapped its two receptor-binding epitopes. The structure was determined to 2.2 \u00c5 resolution and showed that hOSM adopts the up-up-down-down four-helix bundle fold shared by haematopoietin cytokines. The paper was published in Structure, volume 8, pages 863\u2013874, in July 2000.", "Q: What did Marc Deller do at Incyte as a Research Fellow?\nA: As a Research Fellow (Director/Senior Director equivalent) in Structural Biology at Incyte Pharmaceuticals in Wilmington from March 2022 to December 2024, Marc Deller determined over 100 challenging co-crystal structures for medicinal and computational chemistry teams and conducted structure-based drug design for 10+ oncology and immunology programs (ALK2, CDKs, KRAS, FGFRs, JAKs, WRN, VRK1, USPs). He accelerated hit-to-lead processes with the DMPK team, championed automated liquid handling, AI/ML and cloud computing to speed up small-molecule pipelines by 67%, developed \"druggability\" and tractability workflows that enabled screening funnels for 5 new targets, expanded CRO screening capabilities, and supported 3 FDA IND filings, 3 patent applications and 3 manuscripts.", "Q: Which of Marc's apps involve music or games?\nA: ButtFold turns a folding trajectory into music, with every note caused by an event in the fold and each note placed in space at its residue's live 3D position, and MicroPlastics Blaster is a browser-based arcade game by Elora Therapeutics in which the player fires enzymatic blasts at incoming polymer particles, protects the cell and climbs a global leaderboard.", "Q: What sections can visitors navigate to from Marc's home page?\nA: The home page navigation lists About Me, Artificial Intelligence, Blogs, Microplastics in Humans, My Learning, My Patents, My Projects, My Publications, My Structures, My Web Apps, Structural Biology, and My Vibe Coding. There is also a link to the chat bot at marcdeller.com/chat/.", "Q: How many patents does Marc credit to his creative approach?\nA: Marc Deller's creativity post cites 7+ patents, and links the figure to the patents section of marcdeller.com. He says those patents capture not only individual molecules but also new ways to discover and optimise them.", "Q: What is Marc's star sign?\nA: I don't have that, and I'd only be guessing if I tried. I'm on much firmer ground with his crystallography. Is there something about his research or Elora Therapeutics I can help with instead? For anything personal, marc@marcdeller.com is the place to ask.", "Q: Which hematopoietic four-helix bundle cytokine complexes were new in the review period?\nA: Marc Deller's review highlighted two new four-helix bundle cytokine complexes: granulocyte colony stimulating factor (GCSF) bound to two domains of its receptor, and IL-4 bound to the extracellular region of its high-affinity receptor IL-4R\u03b1, each offering new twists on the paradigm established by growth hormone and its receptor.", "Q: What was Marc's role, and where is the post?\nA: Marc Deller contributed the structural biology and crystallography as part of the Incyte structural biology group supporting the programme, and features the work at https://marcdeller.com/discovery-of-4-pyrazolyl-2-aminopyrimidines-as-potent-and-selective-cdk2-inhibitors/.", "Q: Which company funded and conducted this work?\nA: The study was funded by Incyte Corporation of Wilmington, Delaware, and all the authors \u2014 including Marc Deller \u2014 were full-time Incyte employees at the time the work was conducted. Every author is affiliated with the Incyte Research Institute in Wilmington, Delaware.", "Q: Why couldn't the team just remove the problem groups one at a time?\nA: They tried. Compound 11 replaced the pyrrolidinol with an ethoxy group and compound 12 converted the naphthol to 7-fluoronaphthyl; each was 200-fold weaker in the binding assay but held whole blood pERK potency within 2-fold of compound 10. Combining both changes in compound 13, however, cost more than 30-fold potency, and neither single change improved ADME properties. Only after the rigidification strategy restored potency could both liabilities be removed together.", "Q: What are the \"6 C's\" in Marc Deller's r\u00e9sum\u00e9?\nA: Marc Deller organises the highlights of his r\u00e9sum\u00e9 around six C's: Curiosity (20 years leading structural biology and protein biochemistry across 20+ drug discovery programs), Courage (400 high-resolution X-ray crystal structures of protein\u2013ligand complexes), Creativity (7+ patents and discoveries in macrocyclic peptide strategies, new construct designs and crystallization methods), Communication (60+ manuscripts and cross-functional KPIs), Compassion (6 preclinical INDs advanced and 50+ scientists trained and mentored), and Collaboration (2 automated crystallography labs established, $8M budgets managed, 1 pre-seed startup initiated, and partnerships across chemistry, computational science and CROs).", "Q: What function was proposed for the DUF1185 family?\nA: Structural similarities to Bacillus chorismate mutase and genome-context analysis suggested a function in amino-acid synthesis. Marc Deller contributed crystallography to the 2010 Acta Crystallographica Section F paper as part of the Joint Center for Structural Genomics.", "Q: What does the ShpA fold look like?\nA: ShpA belongs to a family of monomeric HPt proteins featuring a highly conserved four-helix bundle, with the phosphorylatable histidine His56 on the surface of the bundle and fully solvent exposed. One end of the four-helix bundle is shorter than in other characterized HPt proteins, while the face that potentially interacts with response regulators is structurally conserved.", "Q: What did the oncostatin M structure reveal about its receptor-binding sites?\nA: Human oncostatin M conforms to the haematopoietin up-up-down-down four-helix bundle topology, with the site 2 epitope responsible for gp130 binding centred on Gly120, forming a dimple on helices A and C, and the site 3 motif responsible for LIFR and OSMR binding formed by a protruding Phe160/Lys163 pair at the start of helix D. Modelling suggested gp130 residue Phe169 packs into the site 2 dimple much as structurally equivalent residues do at the growth hormone-receptor interface.", "Q: What cell-based potency did the antibodies show?\nA: In Ba/F3 cells engineered to express MPL plus Type 1 mutant CALR, the antibodies inhibited STAT5 phosphorylation with IC50 values down to about 0.02-0.06 \u00b5g/mL, and inhibited oncogenic cell proliferation with IC50 values as low as 0.0003 \u00b5g/mL for clone 65 and about 0.001 \u00b5g/mL for clones 55, 68 and 74. The antibodies were active against proliferation driven by both Type 1 and Type 2 CALR mutations.", "Q: What is MicroPlastic Blaster?\nA: MicroPlastic Blaster is a browser-based arcade game by Elora Therapeutics in which the player is the body's enzymatic defence, firing enzyme projectiles at microplastic particles descending toward a cell. Marc built it as a public-awareness and brand engagement tool for Elora's engineered enzyme therapy work.", "Q: Which cysteine does the WRN inhibitor bind, according to the patent's structure?\nA: The co-crystal structure showed Cys727, in the linker region, sitting 3.3 angstroms from the WRN inhibitor compound in the newly formed pocket. This was a surprise, because linear sequence alignment of the Walker A and other canonical helicase motifs would have predicted Cys580 within the Walker A motif as the reactive cysteine, not Cys727.", "Q: Which antibody formats does the patent cover?\nA: Claims 76 to 85 cover human or humanised antibodies, full-length antibodies, IgG1, IgG2, IgG3 and IgG4 isotypes with IgG1 specifically claimed, and bispecific, biparatopic, scFv, Fab, F(ab')2, Fab', Fsc, Fv, sc(Fv)2 and diabody formats, as well as antibodies conjugated to a toxic substance such as a radioisotope or cytotoxic agent.", "Q: What software was used to build and refine the Env trimer model?\nA: Data were integrated and scaled with XDS and corrected for anisotropy using the UCLA MBI Diffraction Anisotropy Server, molecular replacement was performed with Phaser, refinement (rigid-body, grouped-B-factor, XYZ, real-space and TLS) was carried out in PHENIX with NCS, secondary-structure and reference-model restraints, model building was done in Coot, secondary structure was assigned with STRIDE, and validation used MolProbity. Figures were generated with PyMOL.", "Q: Can the antibody be used diagnostically?\nA: Yes. Claims 100 and 101 cover methods of detecting a CALR exon 9 mutation and of diagnosing a myeloproliferative neoplasm by contacting a patient's biological sample \u2014 blood, bone marrow or serum, fresh, frozen or fixed \u2014 with the antibody. The specification describes conjugating the antibody to detectable labels and reading it out by immunohistochemistry, Western blot, ELISA, mass spectrometry, flow cytometry or FACS.", "Q: How many crystals does the JCSG harvest per structure?\nA: The review reports that by January 2014 the Joint Center for Structural Genomics had harvested and screened just over 165,000 crystals, yielding 1,470 structures \u2014 roughly 100 manually harvested crystals for every structure deposited in the Protein Data Bank. Marc Deller notes that harvesting 100 crystals takes a proficient expert of the order of four hours.", "Q: What is TM0077 and what did Marc Deller help show about it?\nA: TM0077 is a carbohydrate esterase family 7 enzyme from Thermotoga maritima, described in the 2012 Proteins paper \"Functional and structural characterization of a thermostable acetyl esterase from Thermotoga maritima\". Marc Deller is a co-author contributing crystallography as part of the Joint Center for Structural Genomics, and the work showed TM0077 is an acetyl esterase rather than the acetyl xylan esterase it had been annotated as.", "Q: How does the PXDesign pipeline work in Marc's description?\nA: Marc describes three stages: PXDesign-d, a Diffusion Transformer backbone that generates Cartesian atom coordinates directly rather than SE(3)-equivariant backbone frames, proposes target-conditioned binder structures of 60 to 160 amino acids; ProteinMPNN then assigns sequences; and filtering applies interface confidence metrics ipTM and pTM from both Protenix and AF2-IG. He stresses that the two predictors capture largely non-overlapping populations of true binders, so ensemble filtering is materially better than either model alone.", "Q: Which fields does Marc say his curiosity has taken him into?\nA: Marc Deller writes that curiosity led him to build a multidisciplinary skill set spanning precision oncology, structural biology, molecular pharmacology, bioinformatics and AI-driven discovery. He also lists protein structures, vaccine design, antibodies and the engineering of microplastic-degrading enzymes as areas that curiosity has guided him into.", "Q: Is it always right to trim a ligand model back to what you can see?\nA: Not necessarily. The chapter notes that limiting the model to the parts visible in the density is crystallographically honest and superficially parsimonious, but can lead to incorrect functional assumptions \u2014 for instance modelling only five residues of a 33-residue peptide leaves a model that needs independent evidence to establish the correct register of the peptide.", "Q: Which capabilities does Marc say ProtComposer unlocks for protein designers?\nA: Marc explains that ProtComposer lets a user hand-craft a novel arrangement of ellipsoids to design a completely new architecture, extract the ellipsoid layout from an existing protein and re-wire its connectivity, or edit the properties of a single component. He describes this as an unprecedented level of control over the final product, achieved by conditioning generation on modular, human-interpretable blueprints.", "Q: What contactless harvesting methods are discussed?\nA: Acoustic droplet ejection can eject crystal-containing droplets at rates exceeding several hundred per second without physical contact, which suits radiation-sensitive samples and microcrystal slurries for free-electron lasers and allows screening of up to 1,728 compounds on a single microplate. Optical tweezers and fibre optical tweezers offer another route to damage-free manipulation.", "Q: Which patents are listed in Marc Deller's Extended Portfolio?\nA: The Extended Portfolio lists two patents: \"Anti-mutant Calreticulin (CALR) Antibodies and uses thereof\" (2022, WO2023107994A1) and \"Bicyclic Compounds as Inhibitors of WRN\" (pending). Both arose from Marc Deller's structural biology work at Incyte.", "Q: Tell me about Marc's post on LatentX.\nA: \"Revolutionizing Drug Discovery: LatentX and the Future of Protein Binder Design\", published 22 August 2025, covers Latent Labs' AI model for generating lab-functional protein binders at atomic precision. Marc contrasts it with traditional discovery that screens millions of random molecules at low success rates, describing LatentX as solving the geometric puzzle of binding directly.", "Q: What is INCB161734 and what did Marc contribute?\nA: INCB161734 is a potent, orally bioavailable KRAS-G12D inhibitor with over 80-fold selectivity for the mutant over wild-type KRAS, achieved by exploiting conformational differences between the two. Marc delivered the structural biology behind it, and the compound is in Phase 1 trials (NCT06179160).", "Q: Which DNA polymerase worked for amplifying the OSM truncations, and why?\nA: Amplification of both \u0394185_OSM and \u0394188_OSM succeeded only with Amplitaq (Perkin Elmer); no product was obtained with Vent Polymerase (New England Biolabs). Vent Polymerase has 3'-5' exonuclease proofreading activity but works on a smaller subset of template/primer combinations, which likely explains the failure. Because Amplitaq lacks proofreading, Marc used low dNTP and magnesium concentrations to maintain fidelity, and ran 25 cycles to balance yield against error.", "Q: What is FlexAppeal?\nA: FlexAppeal, at flexappeal.mdeller.com, lets a user set up protein molecular dynamics without writing the script: every OpenMM option is explained in a form, the app hands back a self-contained bundle to run on the user's own machine, and it then renders interactive trajectory analysis from the results file. Nothing is simulated on Marc's server, so a trajectory never leaves the user's laptop.", "Q: What is the technical architecture behind MICA as Marc explains it?\nA: MICA processes cryo-EM maps and AlphaFold3 structures through separate enhancement pathways before fusing them, with the cryo-EM side using multi-scale convolutions with self-attention and the AlphaFold3 side using feature gates. A progressive three-level encoder feeds a Feature Pyramid Network, and three cascaded decoder heads predict backbone atoms, then C\u03b1 positions, then amino acid types, mirroring the natural biological hierarchy.", "Q: Who is behind LatentX and how is it funded, according to Marc?\nA: Marc notes that Latent Labs is supported by a team including former AlphaFold developers and DeepMind scientists, and was fuelled by a $50 million investment round in early 2025. The stated goal is to make biology programmable so drugs can be designed and refined almost instantly.", "Q: Which institution was Marc at for the Mre11 work?\nA: The Mre11 structure was released in 2007 and published in 2010, while Marc Deller was Target Manager at the Joint Center for Structural Genomics at The Scripps Research Institute in La Jolla, California. Thermotoga maritima was a core JCSG target organism during that programme.", "Q: What music was playing when Marc solved the WRN structure?\nA: I don't know what was playing. The WRN structure itself I can describe in detail, if that's of interest.", "Q: Write an email for me.\nA: Writing emails isn't something I do. If you want to reach Marc himself, marc@marcdeller.com is the address. Otherwise, ask me anything about his work and I'll do my best.", "Q: What achievement does Marc attach to Courage in his 6C's?\nA: In the 6C's on the About page, Courage is paired with 400+ challenging structures determined, enabling breakthrough mechanistic insights. Creativity is paired with 7+ patents through innovative protein engineering and AI integration.", "Q: Explain the significance of Marc Deller's HIV envelope trimer work to a non-scientist.\nA: HIV hides the parts of its surface that antibodies could attack behind a shield of sugar molecules and a shape-shifting spike protein. Marc Deller helped determine the detailed 3D structures of that spike and of antibodies that manage to grab onto it, including his most-cited 2013 Science paper. Seeing exactly where and how those rare protective antibodies bind gives vaccine designers a blueprint for teaching the immune system to make them.", "Q: How did Marc Deller express oncostatin M?\nA: Marc expressed OSM in E. coli strain JM109 as a GST fusion from the pGEX-3C-OSM construct. Cultures in LB-Amp were grown at 37 \u00b0C to an OD600 of ~0.6, shifted to lower temperature, induced with 0.1 mM IPTG, and grown a further 24 hours at 20 \u00b0C. Cells were harvested, lysed by sonication in MTPBS buffer with Triton X-100, and the soluble fusion protein purified by glutathione sepharose affinity chromatography.", "Q: What is BG505 SOSIP.664 gp140?\nA: BG505 SOSIP.664 gp140 is the engineered, soluble, cleaved HIV-1 envelope trimer whose crystal structure Marc Deller and colleagues reported in Science in 2013. It is based on the Tier-2 subtype A virus BG505 and carries a \"SOS\" disulfide bond between gp120 residue 501 and gp41 residue 605 to link the subunits covalently, an Ile-to-Pro change at position 559 (\"IP\") to strengthen gp41\u2013gp41 associations, and a truncation that deletes all but four residues of the gp41 membrane-proximal external region. Together these modifications give a thermostable, non-aggregating, homogeneous trimer suitable for crystallography.", "Q: How is Marc connected to the 2013 HIV Env trimer Science paper?\nA: Marc Deller contributed crystallography to the first high-resolution structure of a soluble cleaved HIV-1 envelope trimer, work done with Ian Wilson's laboratory at The Scripps Research Institute while he was Target Manager at the Joint Center for Structural Genomics there. He lists the 2013 Science paper among the landmarks of his career.", "Q: What delivery strategy does Marc propose for getting PETase into lysosomes?\nA: Marc proposes borrowing from enzyme replacement therapy: the cation-independent mannose-6-phosphate receptor pathway that routes lysosomal enzymes through clathrin-coated vesicles, or fusion proteins using insulin-like growth factor II peptides that bind the bifunctional IGF-II receptor for glycosylation-independent lysosomal delivery. Conjugating FAST-PETase or HGMP01 to such carriers might let them degrade trapped PET inside lysosomes.", "Q: Why was selectivity over CDK1 so important in this program?\nA: Selectivity over CDK1 was critical because CDK1 is a pan-lethal gene essential for the survival and proliferation of normal cells, and the authors hypothesized that much of the clinical toxicity of earlier multitargeted CDK inhibitors came from CDK1 inhibition. In the paper Marc Deller co-authored, the team found that varying substitution at the C5-position of the pyrimidine strongly influenced CDK1 selectivity; replacing a chloro group with trifluoromethyl (compound 3) raised CDK1 selectivity to 130-fold while boosting CDK2 potency.", "Q: What power-ups does MicroPlastic Blaster have?\nA: Six, each lasting ten seconds: FAST-PETase for rapid fire, Enzyme Boost for increased projectile damage, Cell Repair to restore health, Chain Reaction for an explosive chain-blast, and Marc \u00d710 and Paul \u00d710 co-founder avatars that give a tenfold fire rate and play the co-founder's voice line, muting the background fight track until the clip ends.", "Q: How many cups of tea does Marc drink a day?\nA: I don't have a count! Anything I can tell you about his science instead?", "Q: Is there a CDK2 structure among Marc's depositions?\nA: Yes. 8UV0 is titled \"Discovery of (4-Pyrazolyl)-2-Aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\", solved by X-ray diffraction at 1.55 \u00c5 in space group P 21 21 21 with 297 polymer residues and 34.33 kDa, released on 14 February 2024. It is keyworded TRANSFERASE/INHIBITOR and was deposited by Deller and Epling.", "Q: What does Marc Deller write about multi-agent language models in bioinformatics?\nA: In \"Democratizing Bioinformatics with Multi-Agent Language Models\", published 20 August 2025, Marc Deller describes multi-agent systems built on fine-tuned language models with retrieval-augmented generation that plan, execute and validate bioinformatics analyses, achieving expert-comparable performance on conceptual genomics tasks without expensive infrastructure.", "Q: Does Marc talk about time pressure in structural biology?\nA: Yes. In the courage post, Marc Deller writes that courage also looks like a clock, meaning solving a structure in days when an intellectual-property window or a clinical decision cannot wait. He presents speed under pressure as one of the practical faces of scientific courage.", "Q: Which polyketide synthase paper involves Marc Deller?\nA: \"Structure-function analysis of the extended conformation of a polyketide synthase module\", published in the Journal of the American Chemical Society 140(21):6518 in 2018 with Li, Sevillano, La Greca, Deis, Liu, Deller and Mathews among the authors, has 52 citations on Google Scholar.", "Q: Which expression system worked best for LIFR, and how much protein did it give?\nA: Transient expression in human 293T cells using the pIG-1 vector was the most successful system for large-scale production. Marc used about 5 mg of pIG-LIFR-FN3-3C-Fc DNA to transfect roughly 36 flasks of 293T cells by calcium phosphate transfection and obtained protein yields of around 10 mg after one week of expression.", "Q: What is PDB entry 3EQX?\nA: 3EQX is a Fic (filamentation induced by cAMP) family protein, SO_4266, from Shewanella oneidensis at 1.6 \u00c5 in space group P 21 21 21, with 746 polymer residues and 85.45 kDa, released on 14 October 2008 and keyworded DNA BINDING PROTEIN. It is cited in the 2009 Proteins paper on the Fic family protein SO4266.", "Q: What is PDB entry 3JXO?\nA: 3JXO is the crystal structure of the TrkA-N domain protein from Thermotoga sp. RQ2, solved by X-ray diffraction at 1.55 \u00c5 and released on 31 March 2010. Marc Deller's post describes it as the high-resolution view of the individual subunit that assembles into the octameric K+ channel gating ring.", "Q: What did SEC\u2013SAXS reveal about the 1B2/module 2 complex in solution?\nA: Tandem size-exclusion chromatography\u2013SAXS on the 1B2/module 2 complex, prepared in the preacylation, acylated and postelongation states, was compared with CRYSOL-predicted scattering curves for docking models of 1B2 bound to either conformation. In every case the data matched the extended model far better (\u03c7\u00b2 = 0.55) than the arched model (\u03c7\u00b2 = 7.74), showing that module 2 predominantly adopts the extended conformation throughout these catalytic states.", "Q: Write me a Python function that sorts a list.\nA: Writing code isn't something I do, I'm afraid, not even a couple of lines. I'm chatMCD, and I'm here for Marc Deller. What I can tell you about is the software he has built himself: BoltzMaker for structure-prediction campaigns, AlphaFraud for watching PDB and AlphaFold disagree, ChemSage, CODSWALLOP and a few more. Shall I tell you what one of those does?", "Q: What is the Universal Mounting Robot project about?\nA: In his post \"Universal Mounting Robot (UMR)\" Marc Deller surveys approaches to automated protein-crystal harvesting, the last step in the crystallography pipeline that has stubbornly resisted automation. Cloning, expression, purification, crystallisation and data collection are largely robotic, but moving a fragile crystal from its drop to the beamline still relies on expert hands.", "Q: What are the four reactions in the catalytic cycle of DEBS module 2?\nA: The catalytic cycle of DEBS module 2 comprises translocation of the growing polyketide chain into the module (step I), transacylation of an extender unit from methylmalonyl-CoA to the ACP domain (step II), chain elongation via decarboxylative C\u2013C bond formation together with reduction of the resulting \u03b2-keto thioester (step III), and translocation of the chain out of the module (step IV).", "Q: What resolution was the 10E8-T117v2 complex structure?\nA: The crystal structure of 10E8 in complex with the T117v2 epitope scaffold was solved at 2.1 \u00c5 resolution. It aligned to the earlier 10E8-peptide structure (PDB 4G6F) with a C\u03b1 r.m.s.d. of 0.31 \u00c5 over the Fab variable regions and epitope residues.", "Q: What two OSM truncation mutants did Marc Deller engineer?\nA: Marc engineered two C-terminal truncation mutants: \u0394185_OSM, comprising the N-terminal 184 residues, and \u0394188_OSM, comprising the N-terminal 187 residues of OSM-196. Both were made by PCR amplification from the pGEX-3C-OSM-196 template, using primers carrying an EcoRI site and an amber (UAG) stop codon immediately after the intended C-terminal residue.", "Q: How does Marc Deller's thesis distinguish a cytokine from a growth factor?\nA: The thesis defines a cytokine as a protein factor that binds a cell surface receptor with no intrinsic kinase activity, whereas a growth factor binds a receptor that does have intrinsic kinase activity. Cytokines typically act over short range (juxtacrine, paracrine or autocrine), while growth factors such as growth hormone and prolactin have a more classical long-range endocrine role. Marc notes the terms are often used interchangeably, and no distinction is made for the purpose of the study.", "Q: What did Marc Deller's \"druggability\" workflows at Incyte achieve?\nA: As Research Fellow at Incyte, Marc Deller developed innovative \"druggability\" and tractability workflows and designed robust screening strategies for new protein targets, which directly enabled structure-based drug design and compound screening funnels for 5 new targets.", "Q: What are self-encoded libraries, as covered on Marc's blog?\nA: In \"Self-Encoded Libraries: A New Era in Drug Discovery Screening\", published on 7 December 2025, Marc Deller describes barcode-free combinatorial libraries synthesised on solid-phase beads whose members are identified directly from their tandem mass spectrometry fragmentation patterns instead of DNA tags. The post is based on a 2025 Nature Communications paper by V\u00f6geli and colleagues.", "Q: Which enzymes does the PANTS catalogue name?\nA: The PANTS curated subset holds 25 named enzymes with a known lineage and a structure, including IsPETase and its DepoPETase, DuraPETase, FAST-PETase, ThermoPETase, HotPETase and Z1-PETase variants, the LCC lineage with LCC-ICCG and LCC-A2, BhrPETase and TurboPETase, Cut190 and Cut190**SS, TfCut2, MHETase, Kutbu-PETase, Micpa-PETase, and five human gut metagenome polyesterases HGMP01 to HGMP05.", "Q: What is Marc Deller's 2025 patent about?\nA: Marc Deller's 2025 US provisional patent, titled \"Therapeutic Enzymatic Compositions and Delivery Methods for the Degradation of Microplastics and Nanoplastics in Human Physiology,\" describes engineered enzyme formulations designed to break down microplastics and nanoplastics inside the human body. It covers the enzymes, their delivery methods, and strategies for safely clearing the degradation products. It was filed through Elora Therapeutics, where Marc is Chief Scientific Officer and co-founder.", "Q: What is the strongest experimental evidence Marc identifies in that review?\nA: Marc points to clear-cell renal cell carcinoma work by Ye et al. (Advanced Science, 2025), where polystyrene microplastics were markedly more abundant in tumour than adjacent normal tissue and, in patient-derived and xenograft models, accelerated growth via NF-\u03baB and TGF-\u03b2 signalling \u2014 an effect blunted by pathway inhibitors. He calls that the shape of a mechanism, not yet proof of cause in people.", "Q: What does Marc Deller do at DeepCovalent?\nA: Marc Deller has been a Scientific Technical Advisor to DeepCovalent Inc. since October 2025, reporting to the CEO and Founder. He increased throughput by 50% by implementing physics-based Computer-Aided Drug Discovery (CADD) pipelines, advises on small-molecule discovery platforms and GPCR-targeting drug candidate identification, and guides protein-science CRO and vendor oversight and experimental validation strategy.", "Q: Can you explain the KRAS G12D paper in simple terms?\nA: KRAS is a molecular switch that tells cells to grow, and a single amino-acid change at position 12 jams it in the \"on\" position, driving many pancreatic, colorectal and lung cancers. Drugs already exist for one version of that change, G12C, because it leaves a chemically reactive cysteine to latch onto, but the far more common G12D version has an aspartate that offers no such handle. In the 2025 paper Marc Deller co-authored, the Incyte team built a molecule that simply grips the protein tightly without any chemical reaction, and then reshaped it so it survives the gut and gets into the bloodstream when swallowed. Marc's crystal structures showed exactly where the molecule touched the protein at each stage of that redesign.", "Q: What is the \"up-up-down-down\" topology mentioned in the thesis?\nA: The \"up-up-down-down\" connectivity described in Marc Deller's thesis is the unusual topology of the 4-\u03b1-helical bundle cytokine fold, first reported for growth hormone. Two long over-hand loops connect helices A to B and C to D, allowing the helix direction to reverse and the A:D helix pair to pack against the B:C pair with a characteristic skew angle. This topology has not been seen in any protein family outside the cytokines.", "Q: How does Marc describe the PETase and MHETase partnership?\nA: Marc explains that PETase attacks the solid polymer surface in an interfacial biocatalytic process, releasing soluble oligomers with mono(2-hydroxyethyl) terephthalate as the primary product, and that the partner enzyme MHETase then completes the cascade by cleaving MHET into terephthalic acid and ethylene glycol.", "Q: What software and tools did Marc use for the oncostatin M structure determination?\nA: In the 2000 paper, data were processed with DENZO and SCALEPACK, amplitudes calculated with TRUNCATE, data scaled with SCALEIT and anomalous magnitudes estimated with REVISE. Heavy-atom sites were found from Patterson maps and confirmed with RSPS and SOLVE, refined in SHARP, phases improved in SOLOMON, the model built in O and refined in CNS-SOLVE, and the model validated with PROCHECK, DSSP and NACCESS; superpositions used SHP and figures were made with BOBSCRIPT, RASTER3D, VOLUMES and GRASP.", "Q: Does polymer identity matter in the cancer data, according to Marc?\nA: Yes. Marc emphasises that polymer chemistry matters, not just particle size: PET nanoplastics act through a distinct TNF-\u03b1/JAK1/STAT1 axis in gastric models, while polystyrene dominates the colorectal and renal datasets. He notes PET nanoplastics are close to his own enzyme work.", "Q: What is INCB159020 and which structure shows how it binds?\nA: INCB159020 is a novel orally bioavailable inhibitor of the oncogenic KRAS G12D mutant, and its binding mode is captured in PDB entry 9E5D. Marc Deller's post records an X-ray structure at 1.36 \u00c5 released on 22 January 2025, with the ligand code A1BEI alongside GDP and magnesium.", "Q: Why might the TM1088 structure be relevant to drug discovery?\nA: Marc Deller's paper pointed out that a two-subunit TrkA homolog has been identified in Mycobacterium tuberculosis, making it a novel target for structure-based drug design of anti-mycobacterial agents against multidrug-resistant strains. Because the Trk/Ktr/HKT family is found only in non-animal cells with no close human homologs, these transporters are highly selective drug targets, and several anti-mycobacterials such as clofazimine and 7-methyljuglone are already thought to act by targeting K+ uptake.", "Q: Write me a short story.\nA: Storytelling isn't my department, I'm afraid. I stick to Marc Deller: his structural biology, his drug-discovery work, Elora Therapeutics, and the odd tale about why he named a tool after a beer. Any of those take your fancy?", "Q: Has Marc ever turned down a co-authorship?\nA: I don't have anything on that. For questions about how he works, marc@marcdeller.com.", "Q: What clinical hurdles remain for photoacoustic microplastic imaging?\nA: Marc notes that translation to people will need deeper-penetrating clinical systems and careful validation, and that the brain and vasculature \u2014 the very sites where accumulation looks most consequential \u2014 remain the hardest to reach. Even so, he writes, for the first time there is a way to look without first taking something apart. The post is at https://marcdeller.com/seeing-plastic-in-living-tissue/.", "Q: How was the 1B2 antibody discovered?\nA: A large phage-displayed library of 3.7 \u00d7 10^10 Fabs from the naive human antibody repertoire was panned against DEBS module 3 + TE that had been partially biotinylated and immobilized on streptavidin magnetic beads. After several rounds of washing, elution and reinfection into E. coli, multiple high-affinity clones were obtained; the Fab designated 1B2 bound module 3 + TE with a KD of 110 \u00b1 10 nM by ELISA and was shown to recognize the KS-AT fragment (KD = 145 \u00b1 25 nM).", "Q: Has Marc built dashboards?\nA: Marc Deller completed a Coursera project certificate, issued on 27 March 2025, for \"Dashboard Development with Shiny: GenAI for Retail Analysis\", which recognises building interactive dashboards with Shiny. He notes that dashboards make large and complex datasets easier to explore and explain, and that the skill supports scientific interfaces for visualising structures, assays and analytical outputs.", "Q: Did the review discuss NCAM?\nA: Yes. The review mentioned that NMR studies of isolated domains of the neural cell adhesion molecule (NCAM), an IgSF member that mediates homophilic interactions, had provided further structural data and interaction data hinting at its binding sites.", "Q: Which Top96 conditions need no additional cryoprotectant?\nA: Many JCSG Top96 conditions already contain glycerol, MPD or low-molecular-weight PEG at cryoprotective concentrations and are listed with a cryo of \"None\", for example A01 (40% 1,2-propanediol), A05 (40% MPD), all of the organic/PEG/salt row A07\u2013A12, B01\u2013B03, B05, B08, C01\u2013C03, E03, E04, E06, E08, E11, G10, G12 and H01.", "Q: Explain the JCSG Top96 formulation in plain terms.\nA: The JCSG Top96 formulation is essentially a recipe card for 96 chemical solutions that a large structural genomics center found most often produced protein crystals good enough to solve a structure. Marc Deller compiled it so that any lab could set up the same 96 solutions in a single plate, see how many structures each recipe had produced, know which commercial kit it originally came from, and know what antifreeze to add before flash-cooling the crystals.", "Q: Where can I read the INCB159020 structure post?\nA: The post \"Discovery of an Orally Bioavailable KRAS G12D Inhibitor\" is in the structures category of Marc Deller's blog. You can read it at https://marcdeller.com/discovery-of-an-orally-bioavailable-kras-g12d-inhibitor/.", "Q: Who are Marc's parents?\nA: Marc's parents are Norman and Valerie Deller of Bulwell, Nottingham. Norman was a hard-working printer and Valerie a devoted stay-at-home mum who kept things running.", "Q: Where was Marc based for the TM0077 structure?\nA: The TM0077 structure was released in 2010 and the paper appeared in 2012, during Marc Deller's time as Target Manager at the Joint Center for Structural Genomics at The Scripps Research Institute in La Jolla, California. Thermotoga maritima was one of the JCSG's major target organisms.", "Q: How common are MSI-H cancers, and does p53 status matter?\nA: Marc Deller writes that roughly 4% of all cancers harbour the MSI-H signature, and that pharmacological WRN inhibition induces DNA damage markers, activates ATM and CHK2 checkpoint kinases and triggers cell cycle arrest selectively in MSI cells independently of p53 status, which addresses a critical limitation of many cancer therapies.", "Q: What was seen in the kanadaptin FHA domain crystal?\nA: The structure revealed an asymmetric dimer in which one monomer is complexed with a phosphopeptide mimic derived from a peptide segment at the N terminus of a symmetry-related molecule, plus a sulfate bound in the structurally conserved phosphothreonine recognition cleft. Kanadaptin is a nuclear protein of unknown function that is widely expressed in mammalian tissues.", "Q: What is the 3C protease that Marc Deller used, and did he make it himself?\nA: The 3C protease is the rhinovirus type 14 protease 3C, used to cleave the engineered protease site between the GST tag and OSM. Marc produced it himself as a His6-tagged enzyme (3C-His6) from the pTrc-3C-His6 construct, provided by M. A. Hall and Dr K. Hudson of Birmingham, purifying it by nickel agarose affinity chromatography with imidazole elution to roughly 90 % purity.", "Q: How does Marc illustrate the shift from static structures to protein dynamics?\nA: Marc compares it to the difference between having a photograph of a car and a GPS animation of its entire journey, including every turn and traffic jam. He argues this deeper understanding is essential for designing drugs that stabilize correctly folded proteins or correct misfolding pathways. The post is at https://marcdeller.com/how-ai-is-revealing-the-secret-pathways-of-protein-folding/.", "Q: Which human tissues has flow cytometry been applied to for microplastic detection?\nA: Marc notes successful application to human blood, liver, placenta and other tissues after digestion and extraction, with reported concentrations ranging from thousands to millions of particles per millilitre and optimal detection between 100,000 and 1,000,000 particles per millilitre.", "Q: What headline numbers does Marc's home page use to summarise his career?\nA: Marc Deller's home page cites 400+ protein structures, 60+ peer-reviewed publications, 7+ patents, and 6+ FDA IND filings across oncology, immunology, and virology. The same figures are repeated across the About page as a summary of his career output.", "Q: What is Marc Deller's oncostatin M paper?\nA: \"Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\", published in Structure in 2000, has Marc Deller as first author with Hudson KR, Ikemizu S, Bravo J, Jones EY and Heath JK. It reports the 2.2 \u00c5 crystal structure of human oncostatin M together with mutagenesis that mapped the receptor-binding epitopes.", "Q: What staining protocol does Marc describe for flow cytometry of microplastics?\nA: Marc describes Nile red, a lipophilic dye that binds plastic polymer surfaces while showing minimal affinity for biological material. Optimised conditions use 15\u201320 micrograms per millilitre of Nile red in 20\u201330% dimethyl sulfoxide, which balances effective staining against background from dye aggregation, and it works on polystyrene, polyethylene, PET and polycarbonate.", "Q: What is a gatekeeper mutation and why does it matter for FGFR drugs?\nA: The gatekeeper is the amino acid that controls access to a hydrophobic pocket next to the ATP site of a kinase, and in wild-type FGFR2 and FGFR3 it is a valine. Clinical resistance to the approved FGFR inhibitors comes mostly from gatekeeper mutations \u2014 FGFR2 V564I/L/M/F and FGFR3 V555I/L/M/F \u2014 which put a bulkier residue in that position so the 3,5-dimethoxyphenyl group of the approved drugs no longer fits. The 2022 paper Marc Deller co-authored set out to design a chemotype that does not depend on that interaction at all.", "Q: What is Docling, according to Marc Deller?\nA: In \"Unlocking Documents for AI: The Power of Docling for AI-Driven Transformation\", published 18 October 2025, Marc Deller describes Docling as IBM Research's open-source document processing toolkit, released in 2024 under an MIT license, which converts PDFs, DOCX and other formats into machine-readable JSON and Markdown for large language models and RAG pipelines.", "Q: What are Marc's three biggest achievements?\nA: The ten-day structural biology sprint that secured Incyte's first-in-class anti-mutant-CALR antibody patent (projected at more than $500M in annual sales); designing and building Incyte's $8M automated structural biology platform (four-fold efficiency gain, 67% more throughput, 100+ co-crystal structures); and founding and scaling Stanford's Macromolecular Structure Knowledge Center into a profitable service organisation.", "Q: What is in well A01 of the JCSG Top96 screen?\nA: Well A01 of the JCSG Top96 screen is 40% (v/v) 1,2-Propanediol, 0.1 M Sodium acetate pH 4.5, an organic-class condition taken from Wizard Cryo 2 #10 (JCSG Core Suite JCSG-II F05). It produced 6 PDB structures and needs no additional cryoprotectant.", "Q: How tightly did PGT 135 bind gp120?\nA: The JR-FL gp120 monomeric core bound PGT 135 Fab with high affinity, measured at 82 nM by isothermal titration calorimetry. However, the corresponding virus was only weakly neutralised by PGT 135 IgG unless mutations were introduced into the V1\u2013V2 loops, and the weak neutralisation observed correlates with an approximately micromolar binding affinity for the Env trimer.", "Q: When was the HIV-1 Env trimer paper published and where does it sit in Marc Deller's publication record?\nA: \"Crystal structure of a soluble cleaved HIV-1 envelope trimer\" appeared in Science on 20 December 2013 (volume 342, issue 6165, pages 1477\u20131483). It is one of the highest-profile papers from Marc Deller's period at The Scripps Research Institute with the Joint Center for Structural Genomics and the Ian Wilson lab, alongside the 2013 Nature Structural & Molecular Biology gp120 glycan supersite paper and the 2012 Journal of Virology hepatitis C virus E2\u2013AP33 structure.", "Q: What was Marc Deller's role in the JCSG Top96 formulation?\nA: Marc Deller authored the JCSG Top96 Screen Formulation sheet in April 2013 while at the Joint Center for Structural Genomics at The Scripps Research Institute. He designed the screen by mining a decade of JCSG PDB deposition data, assembled the 96-condition plate layout, cross-referenced each condition to its parent commercial screen and to the Qiagen JCSG Core Suite, and defined the companion Top96 Cryo cryoprotectant for each well.", "Q: What did the DRVIA7 antibody structures show?\nA: X-ray and EM structures of DRVIA7, a monoclonal antibody with modest neutralizing breadth isolated from the donor, revealed a VRC01-like angle of approach but less favourable interactions between the light-chain CDR1 and N terminus and the N276 and V5 glycans of gp120. The authors concluded that light-chain accommodation of the glycan shield needs to be taken into account in vaccine design targeting the CD4-binding site.", "Q: How do antibodies reach through HIV's glycan shield?\nA: The post describes antibodies such as PGT135 using extended complementarity-determining region loops, particularly heavy chain CDR H1 and H3, that typically extend 15 to 20 angstroms to match the length of N-linked glycans, with PGT135 burying approximately 1,011 square angstroms of surface area across three glycans.", "Q: Who led the bacterial pleckstrin homology domain paper?\nA: Xu Q was first author, with Bateman A and Finn RD of the Pfam group among the authors and Marc Deller contributing crystallography as part of the Joint Center for Structural Genomics team. It was published in the Journal of Molecular Biology in 2010 (DOI 10.1016/j.jmb.2009.11.006).", "Q: Did Marc work at Pfizer?\nA: Yes. Marc was a Principal Scientist in small-molecule drug discovery and structural biology at Pfizer, contributing structural biology to small-molecule programmes, before joining the JCSG at Scripps in 2006.", "Q: What are PepLites, and where is Marc's FragLites post published?\nA: PepLites are analogues within the FragLite framework designed to mimic peptide interactions, broadening the method's applicability. Marc's post, which cites Erlanson, Murray and Davis in the Journal of Medicinal Chemistry 2019 alongside CDK2 and bromodomain studies from 2022 to 2025, is at https://marcdeller.com/unveiling-fraglites-pioneering-precision-in-drug-discovery/.", "Q: What is PDB entry 3H50?\nA: 3H50 is a tetracenomycin polyketide synthesis protein, TcmJ, from Xanthomonas campestris pv. campestris at 1.60 \u00c5 in space group C 1 2 1, with 114 polymer residues and 12.71 kDa, released on 5 May 2009 with acetate and zinc bound. It is cited in the 2010 Acta Crystallographica Section F paper on conformational changes associated with the binding of zinc acetate at the putative active site of XcTcmJ, a cupin.", "Q: Where was Marc working on the KRAS G12D programme?\nA: Marc Deller worked on KRAS G12D at Incyte Pharmaceuticals in Wilmington, Delaware, where he ran structure-based drug design for more than ten oncology and immunology programmes, including KRAS G12D, as Associate Director of Structural Biology and then Research Fellow in Drug Discovery. The structures were released in January 2025.", "Q: I have a question you can't answer from your training. What then?\nA: If something isn't covered by what I know about Marc, I'll say so rather than guess, and point you to marcdeller.com or marc@marcdeller.com for the real answer. Happy to try, though!", "Q: What is Marc's ORCID password?\nA: Definitely not something I have. His public ORCID record is linked from marcdeller.com.", "Q: Which paper do the acetyl xylan esterase structures belong to?\nA: 1VLQ, 3M81, 3M82 and 3M83 are all cited in \"Functional and structural characterization of a thermostable acetyl esterase from Thermotoga maritima\", Proteins 2012, DOI 10.1002/prot.24041. The set covers the native structure at two resolutions plus PMSF and paraoxon inhibitor complexes.", "Q: Which glycan turned out to be newly important for PGT122 neutralisation?\nA: The trimer structure identified the N137 glycan on gp120 V1 as part of the PGT122 epitope; residues 135\u2013139 of V1, including the N137 glycan, come into close proximity to HCDR1, HCDR2, HCDR3 and LCDR3 of PGT122. Alanine-scanning mutagenesis of gp120 glycans confirmed that PGT121-family antibodies depend on N137 for neutralising the BG505 virus, and locating this residue also allowed the team to fix the relative position of the V1/V2 elements at the trimer apex.", "Q: How does Site Doctor avoid third-party dependencies?\nA: It implements its own minimal HTTP/1.1 client on top of `socket` and `ssl`, which is also what lets it measure each phase of a request precisely. HTML is parsed with regular expressions rather than a DOM library, DNS records come from `dig` with a socket-based fallback, and WordPress data is read from the site's existing public REST API endpoints with no plugins, authentication or scraping. It requires Python 3.8+ and reads only publicly available information. The repo is bellcheddar/site-doctor.", "Q: What did the growth hormone-receptor complex reveal about cytokine signalling?\nA: The thesis explains that the growth hormone-growth hormone receptor (hGH-hGHR) complex was the first haemopoietin receptor-ligand complex solved and showed that one growth hormone molecule binds two receptor subunits, forming a 1:2 complex. This led to a two-site binding model in which hGH binds one receptor at site 1 and then a second receptor more weakly at site 2, with two tryptophan residues (Trp 104 and Trp 169) in the receptor contributing most of the binding energy.", "Q: What did the Nanog structure show about DNA binding?\nA: It revealed the molecular basis of Nanog homeodomain binding to palindromic DNA, illuminating the structural determinants of DNA sequence recognition by the protein. The 3.3 \u00c5 resolution structure includes the two DNA oligonucleotides that form the palindromic site.", "Q: What was concluded about T74?\nA: The T74A variant had about half the O2 consumption rate and half the initial product formation of wild type, but reached similar product levels over longer reaction times, possibly because turnover-dependent inactivation seen in the wild type is reduced in T74A. The paper concluded that the function of T74 is less clear and that this residue is less critical for PMO activity than H161 and Q167.", "Q: How did the thesis relate structural class to receptor type?\nA: Marc Deller's thesis emphasises a close correlation between a cytokine's structural class and its receptor type. For example, almost all 4-\u03b1-helical bundle cytokines bind receptors containing the ~200-residue haemopoietin domain; exceptions like the interferons and IL-10 bind the structurally related interferon receptor (haemopoietin class II) domains. Seven cytokine receptor superfamilies are defined on the basis of function, structural homology and sequence.", "Q: What blood microplastic concentration does Marc cite from Leslie et al.?\nA: Marc cites the 2022 landmark study by Leslie et al., which measured plastic in the blood of 17 out of 22 healthy adults at an average concentration of 1.6 \u00b5g/mL, with PET present in half the samples. He points out that the KTH experiments used deliberately higher concentrations, so the observed activity is a conservative baseline rather than a ceiling.", "Q: Which paper describes PDB 2Q8U?\nA: The structure was published as Das, D., et al. (2010), \"Crystal Structure of the First Eubacterial Mre11 Nuclease Reveals Novel Features That May Discriminate Substrates During DNA Repair\", Journal of Molecular Biology, volume 397, pages 647-663. The deposited macromolecule is annotated as a putative exonuclease.", "Q: What is DeepCovalent?\nA: DeepCovalent is a drug discovery company built around the belief that new drug discovery can happen in a single step. Its One-Step-DeNovo (OS3D) generative AI engine designs optimised drug molecules in one step, cutting early-stage discovery time to less than three weeks, and Marc Deller has been a Scientific Technical Advisor to the company since October 2025.", "Q: Which AI tools does Marc single out on his AI page?\nA: Marc Deller's AI page names AlphaFold, Boltz2, ProteinMPNN, and RFDiffusion as key tools for rapid prediction of protein folding and molecular binding. For virtual screening it names AutoDock for docking and GROMACS for molecular dynamics.", "Q: Which numbers does Marc highlight in his TELSAM-DARPin crystallography post?\nA: Marc highlights that TELSAM fusions can crystallize within 24 hours and at protein concentrations as low as 0.1 mg/mL, that weak AC electric fields of 6 V/mm at 1 kHz expand the crystallization boundary, that TELSAM polymers adopt a minimum helical rise of 46.9 \u00c5 to accommodate fused targets, and that avidity from roughly 5.4 \u00d7 10^8 minimal contact interactions stabilizes otherwise weak crystal contacts.", "Q: What did the crystal structure actually show?\nA: The FGFR2 co-crystal structure of compound 30 showed that, contrary to the team's original docking model, the pyridyl ring is flipped: the amide extends toward the ATP sugar pocket and the pyridine nitrogen points at the gatekeeper residue. The structure also confirmed that the oxygen of the tetrahydrofuran ring makes the hydrogen bond to the primary amide N\u2013H of Asn571 that the compounds had been designed to form (Asn562 in FGFR3).", "Q: Why did Marc build FlexAppeal?\nA: Marc's argument is that setting up a correct MD run means stitching together structure repair, force-field selection, solvation, system construction, integrator and barostat choice, staged equilibration and reporters across roughly 120 interacting parameters, and that many ways of getting it wrong still produce a plausible-looking trajectory \u2014 solvent padding smaller than the non-bonded cutoff, a 4 fs timestep without hydrogen mass repartitioning, CHARMM36 with AMBER's cutoff scheme. FlexAppeal encodes those rules as validation rather than leaving them to be remembered.", "Q: What is chatMCD?\nA: chatMCD is Marc's personal chatbot: a small open-weight language model fine-tuned locally with MLX-LM on his publications, thesis, patents, CV and personal notes so it can answer questions about Marc, his science and his projects.", "Q: What is antibody AP33?\nA: AP33 is a murine monoclonal antibody, discovered by Patel and coworkers, that was found to have broad neutralising activity against diverse hepatitis C virus isolates. In the 2012 study that Marc Deller co-authored, AP33 was expressed as a chimeric mouse\u2013human antibody \u2014 murine variable domains fused to human kappa light chain and IgG1 heavy chain constant regions \u2014 to make expression and purification easier.", "Q: Did AP33 and HCV1 evolve the same solution independently?\nA: Effectively yes. The paper concluded that two independent antibody selection and maturation pathways arrived at a similar solution for engaging the hydrophobic \u03b2-hairpin epitope: both rely on hydrophobic interactions along the cleft between the light and heavy chain variable domains, both use primarily side chains for the interaction, both form a hydrogen bond with the Gly418 tip of the peptide, and in both the major hydrophobic contacts are with Leu413 and Trp420.", "Q: What did the in vivo studies of compound 17 show?\nA: In vivo, compound 17 produced roughly 50% reduction in tumor Rb phosphorylation over 24 hours in a 5-day pharmacokinetic/pharmacodynamic study in CCNE1-amplified OVCAR3 tumor-bearing mice, dosed once daily at 30 mg/kg. In an efficacy study it gave 38% tumor-growth inhibition after once-daily oral dosing at 30 mg/kg for 28 days, with a favorable safety profile and no unexpected deaths, adverse effects, or body-weight losses.", "Q: How does Marc define the \u00e5ngstr\u00f6m and who was it named after?\nA: Marc defines the \u00e5ngstr\u00f6m (\u00c5) as one ten-billionth of a metre, or 1 \u00d7 10\u207b\u00b9\u2070 m, and notes it is not an SI unit. It is named after the Swedish physicist Anders Jonas \u00c5ngstr\u00f6m, who lived from 1814 to 1874 and first quantified solar spectral lines on this scale.", "Q: What did the Di-Gembody work actually achieve?\nA: The method was validated on both soluble and membrane protein targets down to 14 kDa, with hen egg white lysozyme becoming the smallest protein ever resolved by cryo-EM. Di-Gembodies also allow two copies of the same protein, or two entirely different proteins, to be displayed and solved in parallel within a single experiment.", "Q: Why does protein stability matter, according to Marc Deller's review?\nA: The review argues that an understanding of protein stability is essential for optimising the expression, purification, formulation, storage and structural study of proteins, and that it matters to the biotechnology, pharmaceutical and food industries as well as to academic researchers. It introduces key methods for analysing stability and offers tactics for addressing stability problems during expression, purification and crystallization.", "Q: How does Marc suggest designing constructs for crystallisation?\nA: Marc advocates moving beyond full-length proteins, using AlphaFold3 and servers such as ProteinCCD to identify disordered and flexible regions and design libraries of truncation constructs with varying start and end points. He also weighs affinity tags, noting maltose-binding protein can act as a solubility enhancer and crystallisation chaperone, while self-cleaving intein-based tags avoid protease treatment. The post is at https://marcdeller.com/engineering-protein-stability-unlocking-the-secrets-of-successful-crystallization/.", "Q: What did BoltzMaker's ligand pose panel reveal about confidence scores?\nA: On a real GLP1R campaign, orforglipron was predicted with `ligand_iptm` 0.940 and confidence 0.836 yet sat 9.4 \u00c5 from where the crystal structure 7E14 places it, correct pocket and conformation but rotated end for end. Across poses spanning 2.04 to 9.51 \u00c5 the confidence score moved only 0.834 to 0.849, so Marc concludes only comparison with an experiment can tell a right pose from a wrong one once the ligand is in the pocket.", "Q: Where can I read Marc's PGT135 post?\nA: Marc Deller's post is \"Crystal Structure of PGT135 Fab in Complex with gp120 Core Protein from HIV-1 Strain JR-FL Bound to CD4 and 17b Fab\", in the structures category of his site. You can read it at https://marcdeller.com/crystal-structure-of-pgt135-fab-in-complex-with-gp120-core-protein-from-hiv-1-strain-jr-fl-bound-to-cd4-and-17b-fab/.", "Q: Why was the KPN03535 structure worth solving?\nA: Because the protein had no detectable sequence similarity to anything of known structure, the crystal structure provided the foundation for further investigation of this divergent member of the OB-fold family in an important opportunistic pathogen. Marc Deller contributed crystallography through the Joint Center for Structural Genomics pipeline.", "Q: What was Marc's h-index in 2015?\nA: I don't have a historical figure for that. His current citation metrics are on his Google Scholar profile, linked from marcdeller.com.", "Q: What is the lead compound in the WRN patent?\nA: Example 1 is N-(2-(2-ethyl-10-(2-((4-fluorophenyl)amino)-2-oxoethyl)-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide, and it is the first compound recited in claim 114. It was made in seven steps from 1H-benzo[d]imidazol-2-amine and ethyl propionylacetate, and it scores in the most potent band in both the WRN ATPase and WRN helicase assays.", "Q: Which conference did Marc attend in 2019?\nA: I don't have a record of his conference attendance for that year. His talks and posters are listed on marcdeller.com.", "Q: Which drug does Marc reference as an example of AI, structural biology and screening combined?\nA: Marc references the discovery of INCB159020, an orally bioavailable KRAS G12D inhibitor, citing the Journal of Medicinal Chemistry paper as an example of how combining structural biology, AI and high-throughput screening has produced inhibitors for notoriously undruggable targets like KRAS.", "Q: Which 2025 antibody automation platforms does Marc cite?\nA: Marc cites Ginkgo Bioworks' PROPHET-Ab, a high-throughput developability suite profiling 246 therapeutic IgGs across ten biophysical assays to emit ML-ready datasets; LabGenius's EVA, a closed-loop multispecific platform that traversed 44,160 T-cell engager designs and recovered sub-10 pM HER2\u00d7CD3 candidates with 14-fold enrichment over Sobol sampling; and a Caltech fully automated end-to-end workflow taking plasmid DNA through to assay-ready secreted protein on a single 96-well robotic line.", "Q: What is cif_to_plip?\nA: cif_to_plip is Marc Deller's single-file Python pipeline that converts a ModelCIF or mmCIF co-folding output into a format-compliant PDB and feeds it into PLIP for full protein-ligand interaction profiling. It is designed for the outputs of Boltz-2, AlphaFold 3 and Chai-1, which carry no bond table and often break the PDB format specification.", "Q: What does Marc Deller's post on large language models in biomedical NLP report?\nA: \"Unlocking the Promise and Pitfalls of Large Language Models in Biomedical NLP\", published 17 August 2025, reports on a benchmarking study by Chen et al. in Nature Communications that pitted GPT-3.5, GPT-4, LLaMA 2 and PMC LLaMA against fine-tuned domain models such as BioBERT and BioBART across 12 biomedical NLP tasks.", "Q: Where can I read Marc's TM0077 post?\nA: Marc Deller's post is titled \"Crystal Structure of Acetyl Xylan Esterase (TM0077) from Thermotoga maritima at 2.40 \u00c5 Resolution\" and appears in the structures category of his blog. You can read it at https://marcdeller.com/crystal-structure-of-acetyl-xylan-esterase-tm0077-from-thermotoga-maritima-at-2-40-a-resolution/.", "Q: What are the dimensions of the TM1088 gating ring and its central channel?\nA: The octameric TM1088 assembly measures approximately 127 \u00c5 x 101 \u00c5 x 53 \u00c5. It forms a large central channel about 50 \u00c5 long with a solvent-accessible surface of about 9700 \u00c5\u00b2 (calculated with CASTp and MOLE), narrowing to about 15 \u00c5 at its narrowest point, which is comparable to the ~15 \u00c5 central diameter of the MthK channel. The four AMP molecules line the inside of the channel and are solvent accessible.", "Q: Did Marc work on an Mre11 structure?\nA: Yes. PDB entry 2Q8U is the crystal structure of Mre11 from Thermotoga maritima MSB8, the gene product TM1635, determined by X-ray diffraction at 2.20 \u00c5 and released on 26 June 2007. Marc Deller lists it in the structures section of his website.", "Q: Which journal published the TrkA gating ring work?\nA: The work was published as Deller, M.C., et al. (2015), \"Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly\", in PLoS ONE, volume 10, article e0122512. Marc Deller is the first author and has a companion blog post about the paper on marcdeller.com.", "Q: How does Marc say microplastic degradation is monitored clinically in that post?\nA: In the bloodstream post Marc explains that high-performance liquid chromatography (HPLC) and NMR spectroscopy are used to track both the microplastics and their breakdown products, which he describes as vital for clinical monitoring of any enzyme-based treatment.", "Q: Which journals has Marc Deller published in?\nA: Marc Deller's papers appear in Science, Nature Structural and Molecular Biology, Immunity, PNAS, the Journal of Medicinal Chemistry, the Journal of the American Chemical Society, the Journal of Molecular Biology, the Journal of Biological Chemistry, Structure, Proteins, Protein Science, PLoS Pathogens, PLoS ONE, ACS Chemical Biology, the Journal of Virology, the Journal of Computer-Aided Molecular Design, Current Opinion in Structural Biology and several sections of Acta Crystallographica.", "Q: What is gp130 and why does it matter?\nA: gp130 is a shared transmembrane signal-transducing receptor used by the gp130 family of cytokines, which includes oncostatin M, leukaemia inhibitory factor (LIF), IL-6, IL-11, cardiotrophin-1 and CNTF. As Marc Deller's paper explains, each cytokine triggers signalling by assembling an oligomeric receptor complex that always includes gp130 plus other partner receptors, so the composition of the complex determines both the shared and unique biological activities of each cytokine.", "Q: Which methodological choices does Marc say mattered most in that flexible fitting study?\nA: Marc highlights four findings: masked density maps that exclude neighbouring chain density consistently beat box-cropped maps; the wide normal mode range 1\u201312 outperformed narrower ranges such as 1\u20139 or 7\u20139 by capturing both rigid-body and internal motions; the local Powell optimizer matched or beat expensive global methods like Dual Annealing and Differential Evolution, implying a smooth correlation landscape; and the inner product similarity measure removed cropping bias and zero-padding dependence while matching Pearson correlation for accuracy.", "Q: Does Marc write about running AI tools locally?\nA: Yes, his AI page collects several practical local-install posts, including \"Alphafold Local Install on Mac\", \"Installation of RFdiffusion on MacOSX M1 architecture without GPU support\", and \"Unleashing the Power of AnythingLLM: Running Large Language Models Locally\". His ChemSage post also covers building a chemically aware LLM on local hardware.", "Q: What is Chapter 2 of the thesis about?\nA: Chapter 2 covers the engineering of truncated oncostatin M. Marc Deller designed and produced two C-terminal truncation mutants of OSM as GST fusions in the pGEX expression vector, ready for expression in E. coli and characterisation in Chapter 3. The rationale was that wildtype OSM-196 was prone to aggregation, low solubility and dimer formation, and had resisted crystallisation.", "Q: What does Marc Deller's post on precision oncology cover?\nA: \"Precision Oncology: Integrating AI and High-Throughput Screening to Fight Cancer\", published 13 August 2025, covers how high-throughput screening platforms that test hundreds of thousands of compounds in a day are paired with machine learning models that sift the resulting data to prioritize the most promising candidates and novel targets.", "Q: What is a clash score?\nA: A clash score is the result of an all-atom close-contact analysis, computed with MolProbity and visualized in COOT, which lists bad atom contacts and displays them as dots and spikes. The chapter treats steric clashes between a poorly placed ligand and the protein, and the absence of any sensible contacts, as the two common failure modes of an implausible ligand.", "Q: Why is the N276 glycan such an obstacle?\nA: Unlike glycans that can be displaced during antibody binding, N276 maintains its rigid orientation, so antibodies that try to displace it incur energetic penalties that compromise binding affinity, whereas those that accommodate it through complementarity-determining region modifications achieve superior neutralization breadth.", "Q: Why does the 10E8 result matter for HIV vaccine design?\nA: The 10E8 antibody has greater neutralizing breadth than most broadly neutralizing antibodies to HIV, and its epitope is the conserved membrane proximal external region of gp41. The study showed 10E8 engages lipid through a specific head group interaction site and a basic, polar surface on the light chain, and modelled the MPER as essentially perpendicular to the virion membrane during neutralization, information useful for immunogen and therapeutic design.", "Q: How well does the PANTS activity model perform?\nA: PANTS reports every training run ever recorded, including failures, scoring AUC against a composition baseline computed from amino-acid counts alone. Measured-only models reach AUC 0.774 to 0.949 against baselines of 0.565 to 0.803, one hard contrast came out at 0.493, essentially chance, and a mixed model trained on similarity-derived labels reached 1.000 in one early run, which Marc records as \"perfect, and meaningless: it reproduced the similarity rule that made the labels\".", "Q: What is the \"chopsticks\" harvesting system?\nA: The chopsticks system is a two-fingered micro-hand developed by KEK and AIST in Japan, in which a main-finger and a sub-finger operate independently under joystick control to hold a crystal and place it into a microloop. Marc Deller's review reports it achieved a 90% success rate on 8 \u00b5m glass beads and that its open-approach geometry reduces the crystal chasing seen with microloops.", "Q: How many hours did the oncostatin M structure take to solve?\nA: I don't have a figure for the time it took. I can describe the structure and what it showed, if that's more useful.", "Q: How many claims does the CALR antibody patent have?\nA: The published application contains 101 claims. Claims 1 to 43 define the antibodies by CDR, VH/VL and heavy-chain/light-chain sequences, claims 44 to 54 by function and Fc effector properties, claims 55 to 73 by epitope and competition, claims 74 to 85 by format, claims 86 to 90 cover nucleic acids, vectors, cells, methods of making and pharmaceutical compositions, claims 91 to 98 cover methods of treatment, claim 99 a kit, and claims 100 to 101 detection and diagnosis.", "Q: What design principles delivered oral bioavailability?\nA: The team began with a weakly potent hit and used crystallographic insights for systematic structure-based optimisation, emphasising molecular rigidity through fewer rotatable bonds to improve membrane permeability and metabolic stability, together with balanced polarity for gastrointestinal absorption and systemic distribution.", "Q: What is the default target of Site Health Inspector?\nA: The default scan target is https://marcdeller.com, set as a fallback on line 4 as `TARGET=\"${1:-https://your-default-site.com}\"`. The script is a single file with no config and is released under the MIT licence; all output uses ANSI colour with star ratings and bar charts for instant visual triage. The post is at https://marcdeller.com/a-pure-bash-colour-coded-website-health-audit-no-install-no-dependencies/.", "Q: What is the JCSG Top96 screen?\nA: The JCSG Top96 is a rationally designed, single-plate, 96-condition protein crystallization screen that Marc Deller and colleagues at the Joint Center for Structural Genomics (JCSG) assembled from the crystallization conditions that produced over 1000 deposited JCSG protein structures. It contains the 96 most successful conditions with respect to PDB deposition.", "Q: Can you diagnose this symptom I'm having?\nA: I'm not able to help with medical questions; please talk to a healthcare professional for that. What I can help with is Marc Deller's work in drug discovery and structural biology, if that's of interest.", "Q: How did the unliganded PGT 135 Fab crystallise?\nA: Unliganded PGT 135 Fab crystallised over 28 days at 20 \u00b0C in 20% (w/v) PEG 8000, 0.1 M CHES at pH 9.5. The crystals were harvested and cryoprotected by brief immersion in 70% well buffer with 30% glycerol and flash-cooled in liquid nitrogen, and data were collected at Advanced Photon Source beamline 23ID-B at a wavelength of 1.033 \u00c5 and 100 K.", "Q: What does Marc Deller's post \"Citation Giants\" say about the most-cited papers of the 21st century?\nA: Published 11 October 2025 and drawing on a Nature analysis by van Noorden and colleagues, the post reports that headline discoveries such as mRNA vaccines, CRISPR, the Higgs boson and gravitational waves do not appear among the most-cited papers since 2000. Instead, Marc Deller writes, the citation champions are AI architectures, research methodologies, statistical software and clinical data.", "Q: What's 847 times 392?\nA: Arithmetic isn't really my specialty, and I'd rather not risk giving you a wrong answer, but a calculator will sort that out in a second. If you're curious about the kind of quantitative work Marc Deller does, though, I can tell you about his structural biology and drug-discovery data.", "Q: Which residues make up the secondary coordination sphere of MtPMO3*?\nA: The secondary coordination sphere of MtPMO3* comprises T74, which hydrogen-bonds with the N-terminal amino group of H1 and with ligands in the axial solvent-facing position; H161, positioned to hydrogen-bond with equatorial solvent-facing ligands; and Q167, which hydrogen-bonds both with equatorial solvent ligands and with the axial tyrosine Y169. In chains C and E of the crystal, H161 and Q167 hydrogen-bond to D26 from a symmetry-related molecule.", "Q: What does the kanadaptin FHA domain paper report?\nA: \"Structural insights into the recognition of phosphopeptide by the FHA domain of kanadaptin\", published in PLoS ONE in 2014, reports the crystal structure of the forkhead-associated domain of human kanadaptin at 1.6 \u00c5 resolution. Marc Deller is the second author, after Xu Q, with Nielsen TK, Grant JC, Lesley SA, Elsliger MA, Deacon AM and Wilson IA.", "Q: What are the imidazo[1,2-a]pyrimidin-5(1H)-one WRN inhibitors?\nA: Marc Deller's post \"2025 \u2013 Imidaz0[1,2-a]pyrimidine-5(1 h)-ones as wrn inhibitors\" covers patent US20250270218A1, which discloses imidazo[1,2-a]pyrimidin-5(1H)-one derivatives as selective WRN inhibitors. These bicyclic heterocyclic compounds bind precisely at the WRN helicase interface and lock the enzyme in an inactive conformation, preventing essential DNA repair.", "Q: How does GOBSMACKED grade a prediction?\nA: GOBSMACKED grades seven metrics from A to F: ligand RMSD (A at 1.0 \u00c5 or better), pocket C\u03b1 RMSD after MD, chi1 agreement, PLIP interaction-fingerprint Jaccard, MD stability, MD rescue, and pose validity. The composite GOBSMACK score is a weighted mean of those grades (ligand RMSD 30, PLIP overlap 20, pocket C\u03b1 15, chi1 10, stability 10, rescue 10, validity 5), a validity failure caps it at 40, and a run with no reference structure gets no composite at all.", "Q: Why have intrinsically disordered regions been so hard to target?\nA: Marc writes that IDRs lack stable, rigid structures and exist instead as flexible, fluctuating chains, which is exactly what lets them participate in diverse signalling pathways but also makes them nearly impossible to target with conventional binding strategies. That difficulty is why they have long sat in the undruggable category.", "Q: Which extracellular degradation modalities does Marc describe?\nA: Marc describes LYTACs, introduced in 2020, which tether an extracellular target to a lysosome-shuttling receptor such as the cation-independent mannose-6-phosphate receptor or, in the liver, the asialoglycoprotein receptor. He also covers AbTACs and PROTABs that co-opt cell-surface E3 ligases such as RNF43 and ZNRF3, KineTACs routing cargo through cytokine receptors, ATACs using peptide warheads against integrins, and computationally designed EndoTags from the Huang, Bertozzi and Baker work that trigger endocytosis without any native ligand.", "Q: How does the N-terminal region of OSM compare with that of LIF, IL-6, CNTF, Epo and GCSF?\nA: Marc Deller noted that the loop preceding helix A is anchored to helix C by two disulphides in LIF but by only one in OSM, and that no such anchoring disulphides exist in IL-6, CNTF, Epo or GCSF. In IL-6 this loop deviates away from the body of the molecule, in GCSF it passes over the end of the bundle with Leu 10 packing into the helices, and in CNTF, which has no disulphides at all, the first ten residues are highly flexible.", "Q: What is reductive methylation and why is it used?\nA: Reductive methylation modifies the \u03b5-amino groups of solvent-exposed lysines using dimethylamine\u2013borane and formaldehyde, and is believed to work by introducing new surface contacts that promote lattice formation. Marc Deller's review describes it as one of the workhorse salvage techniques of structural genomics consortia, alongside newer variants such as ethylation and isopropylation and the cysteine alkylation used for the \u03b21-adrenergic receptor.", "Q: What is ShpA and what is its PDB code?\nA: ShpA is a histidine phosphotransferase from Caulobacter crescentus, deposited under accession NP_419930.1, and its crystal structure is PDB entry 2OOC. Marc Deller's post records an X-ray structure at 1.52 \u00c5 resolution, released on 6 February 2007, with glycerol and PEG fragments as ligands.", "Q: What ligands appear most often in Marc Deller's structures?\nA: Zinc is the most common bound ligand, in nine entries, followed by sodium in eight, NAG (N-acetylglucosamine) in four and calcium in four. Magnesium appears three times, and phosphate, nickel, GDP, iron and acetate twice each.", "Q: Is there a career timeline on Marc's site?\nA: The About page includes an interactive bio timeline headed \"Career Timeline\", organised by Institution, Summary, Programs, Targets, and Techniques. It lets visitors step through his career institution by institution rather than reading a static list.", "Q: How much data does Marc say a QSAR model needs?\nA: Marc favours large, diverse datasets of thousands to tens of thousands of compounds to minimise overfitting and capture chemical space, citing repeat-dose toxicity work where aggregating over 70,000 data points outperformed partitioned subsets. He notes Random Forest and Support Vector Regression excel with ample data but struggle on small sets, while deep learning scales with dataset size.", "Q: What are the two kinds of protein stability Marc's review distinguishes?\nA: Marc Deller's review distinguishes compositional stability, meaning the protein molecules all have the same well-defined chemical makeup, from conformational stability, meaning the protein has minimal disordered or dynamically variable regions. The review argues that a crystallographer must assess both before expecting a protein to crystallize.", "Q: What does Marc say about structures that contradict the prevailing model?\nA: Marc Deller writes that real courage sometimes means trusting a structure that contradicts the prevailing model, and acting on it before the field is ready. He treats this willingness to back the data as part of converting technical risk into strategic advantage for the programs and teams he supports.", "Q: What does Marc Deller write about cryo-EM and membrane proteins?\nA: In \"Cryo-EM and the New Frontier in Membrane Protein Drug Design\", published 13 August 2025, Marc Deller explains that membrane proteins resist the crystallization needed for X-ray crystallography, and that cryo-EM now allows them to be visualized in a near-native state with near-atomic clarity by flash-freezing and computational reconstruction.", "Q: What is Marc Deller's post \"Revolutionizing Protein Crystallography: The Synergy of TELSAM-DARPin Fusion and Electric Field Optimization\" about?\nA: In that post, published on 12 October 2025, Marc Deller describes how TELSAM (Translocation ETS Leukemia Sterile Alpha Motif) crystallization chaperones fused to Designed Ankyrin Repeat Proteins (DARPins) can be combined with electric-field-based polymer orientation to make difficult crystallization targets tractable. He explains that TELSAM forms pH-dependent polymers that build highly ordered crystal lattices when fused to a target protein, replacing trial-and-error screening with a more predictable route to high-resolution structures.", "Q: What did the BG505 SOSIP.664 structure reveal?\nA: It revealed the architecture of the trimeric spike, showing how gp120 and gp41 subunits interact to maintain stability while preserving the conformational flexibility needed for viral entry, with the gp120 V1/V2/V3 loops forming a protective apex around the threefold axis that both stabilises the trimer and shields conserved regions from antibodies.", "Q: What did ShpA's structure suggest about its mechanism?\nA: Similarities in the interaction surface around the phosphorylation site suggested that ShpA shares a common mechanism for molecular recognition and phosphotransfer with the yeast phosphotransfer protein YPD1, despite low overall sequence similarity. When phosphorylated, TacA triggers a downstream transcription cascade for stalk synthesis in an RpoN-dependent manner.", "Q: What is the Universal Mounting Robot?\nA: The Universal Mounting Robot, or UMR, is one of the top projects listed on Marc Deller's Projects page, sitting alongside the Top96 Crystal Screen among his crystallography automation work. It reflects the crystal harvesting robotics he cites under platform innovation.", "Q: What is Marc's honest assessment in the cryo-EM grid preparation post?\nA: Marc concludes there is no panacea: the protein, its buffer and the workflow interact in complex ways, and cryo-EM still depends heavily on iterative optimization and trial and error. He says hyperquenching, fast plunge systems and new support materials are blazing a trail but the drag race against diffusion is still on, with time-resolved grid behaviour and real-time analytics as the next steps.", "Q: When was the oncostatin M structure paper published?\nA: The paper was received by Structure on 26 April 2000, revised in June 2000, accepted on 28 June 2000 and published on 20 July 2000 as Structure 2000, volume 8, pages 863\u2013874. It was part of Marc Deller's D.Phil. research at the University of Oxford.", "Q: What is the povorcitinib paper that Marc Deller co-authored?\nA: Marc Deller is a co-author of \"Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\", a Drug Annotation published in the Journal of Medicinal Chemistry in 2026 (DOI 10.1021/acs.jmedchem.5c03753). The paper describes the medicinal chemistry campaign at Incyte that produced povorcitinib, a selective JAK1 inhibitor, and includes the X-ray crystal structures of the compound bound to JAK1 and JAK2.", "Q: Can you explain this paper in simple terms?\nA: FGFRs are proteins on the surface of cells that tell them to grow, and when they are stuck in the \"on\" position they can drive bladder and bile-duct cancers. Existing drugs block all four family members, which causes a phosphate-related side effect, and tumours escape them by making one amino acid in the drug pocket slightly bigger. In the 2022 paper Marc Deller co-authored, the Incyte team built a chemically different molecule that avoids that pocket, still works when the amino acid changes, and leaves the family member responsible for the phosphate side effect alone. Marc's crystal structure showed exactly how the molecule sits in the protein and why the mutations cost it a little potency.", "Q: What can you type into CODSWALLOP?\nA: CODSWALLOP accepts a PDB ID, a PDB entity id, a UniProt accession, a gene name, a Pfam or InterPro identifier, a raw sequence, or just what the protein is called. Anything genuinely ambiguous returns a disambiguation card rather than a guess, so 4HHB asks whether you meant the alpha or beta globin and LYZ asks which of twelve organisms.", "Q: When in Marc's career did the Nanog structure appear?\nA: The Nanog homeodomain structure was released in October 2014 and published in 2015, during Marc Deller's time as Target Manager at the Joint Center for Structural Genomics at The Scripps Research Institute, a role that ran from 2006 to June 2015. He contributed the crystallography.", "Q: What is the N297A mutation mentioned in the patent?\nA: N297A is a heavy-chain mutation introduced during engineering that produces Fc effector \"null\" antibodies, in which Fc effector function \u2014 particularly antibody-dependent cell-mediated cytotoxicity \u2014 is eliminated or substantially eliminated. The patent notes that antibodies carrying the N297A Fc-null mutation generally showed increased binding affinity for mutant CALR compared with those without it.", "Q: Which publication describes PDB 3B77?\nA: The structure was published as Deller, M.C., et al. (2010), \"Crystal Structure of a PH Domain Containing Bacterial Protein from Exiguobacterium sibiricum 255-15\", in the Journal of Molecular Biology, volume 396, pages 31-46. Marc Deller is the first author on that paper.", "Q: What does Marc say about TM6 movement and class B GPCR kinetics?\nA: Marc explains that GLP-1R activation requires outward movement of the intracellular part of TM6 by approximately 11.9 \u00c5 at residue R348 to open the Gs binding cavity. He notes that class B receptors show unusual kinetics in which Gs engagement can precede full TM6 opening, unlike class A GPCRs, and that TM6 subsequently relaxes back slowly, enabling prolonged signalling.", "Q: What did the PGT128 and 8ANC195 trimer complex reveal?\nA: The structure revealed complete epitopes for both antibodies on the native-like HIV-1 Env trimer for the first time, including the full 8ANC195 interface spanning gp120 and gp41. Marc Deller's post also notes that PGT128 allosterically repositions the N262 glycan, and that these findings provide direct blueprints for structure-based HIV vaccine design.", "Q: What environmental checks does the chapter recommend?\nA: Marc Deller and his co-authors advise asking whether symmetry-related molecules are involved in ligand binding, in which case biochemical or biophysical evidence may be needed to distinguish real interactions from crystal packing artefacts; validating every copy of the ligand if there are multiple molecules in the asymmetric unit; checking bond distances in the binding site for plausible non-covalent interactions; and computing the clash score.", "Q: Which earlier Incyte drugs used the same intramolecular hydrogen bonding strategy?\nA: The paper notes that the same IMHB approach had already succeeded in two earlier Incyte programmes, leading to the FDA-approved c-Met inhibitor capmatinib (INCB28060, compound 13) and the clinically explored 11-\u03b2-HSD1 inhibitor INCB13739 (compound 11). Those precedents motivated applying the tactic to improve the pharmacokinetics of the JAK1 series.", "Q: What does Marc say self-driving labs mean for structural biology and scientists themselves?\nA: Marc argues that when co-crystal structures, biophysical characterization and DMTA cycles can all be choreographed by a single orchestrating layer, a small team can credibly run dozens of programs in parallel. He is explicit that scientists are redirected rather than replaced: human expertise shifts to framing questions, judgment and interpretation while, as he puts it, the routine has finally found a competent autopilot. The post is at https://marcdeller.com/the-self-driving-lab-how-automation-and-ai-are-redefining-the-pace-of-discovery/.", "Q: What is the gp130 family of cytokines?\nA: The gp130 family of cytokines, which is the focus of Marc Deller's thesis, includes leukemia inhibitory factor (LIF), oncostatin M (OSM), interleukin-6 (IL-6), interleukin-11 (IL-11), ciliary neurotrophic factor (CNTF) and cardiotrophin-1 (CT-1). They are among the most pleiotropic cytokines known and share use of the glycoprotein 130 (gp130) signal-transducing receptor, functioning in haematopoiesis, neuronal development and the immune system.", "Q: What are the appendices of Marc Deller's thesis?\nA: The appendices of Marc Deller's thesis are practical laboratory recipes and formulae: Appendix I gives molecular biology reagents (LB medium and plates, TAE and TBE buffers, transformation media, DNA loading buffer); Appendix II gives protein expression and purification reagents (SOB and SOC media, TNED buffer, MTPBS, gel filtration buffer, reduced glutathione, PNGase F buffers, PBS, acrylamide/bis, SDS-PAGE buffers and gels, Coomassie stain and destain, Western transfer buffer); Appendix III gives eukaryotic cell culture media for 293T, Drosophila S2 and CHO cells; and Appendix IV gives spectrophotometric formulae for protein, oligonucleotide and plasmid DNA concentration.", "Q: Which schools did Marc attend?\nA: Marc went to St Mary's Primary School in Bulwell, then Henry Mellish School (where his mentor Gilbert Gimson inspired him to study chemistry), then High Pavement College in Nottingham for A-levels, before Leeds University and Oxford.", "Q: What is the argument of Marc Deller's post \"Future Therapeutics\u2014Engineering Solutions for Microplastic Burden\"?\nA: In \"Future Therapeutics\u2014Engineering Solutions for Microplastic Burden\" (15 August 2025), Marc Deller argues that the discovery of microplastics in blood and tissues, including brain and breast milk, demands new therapeutic approaches. He proposes biocatalytic enzyme therapy for in situ breakdown of circulating particles, supported by encapsulation, cell-surface display, apheresis and sensitive diagnostics.", "Q: What is Marc Deller's post \"When Perception Deceives Reality\" about?\nA: Published 27 August 2025, this post examines the debate over Apple's \"Illusion of Thinking\" study, which reported that large reasoning models including OpenAI's o3-mini, DeepSeek-R1 and Claude 3.7 Sonnet suffer complete accuracy collapse beyond certain puzzle complexity thresholds. Marc Deller lays out the critique, principally A. Lawsen's \"The Illusion of the Illusion of Thinking\", that the apparent collapse reflects experimental design flaws rather than cognitive failure.", "Q: Where can I read about Marc's microplastics work?\nA: Marc Deller's microplastics hub is at https://marcdeller.com/microplastics/, and Elora Therapeutics has its own site at eloratherapeutics.com. The page collects his posts on detection, health effects, and enzyme-based treatment.", "Q: What did Marc Deller deposit from Stanford?\nA: Two Stanford-era entries stand out: 6C9U, the [KS3][AT3] didomain from module 3 of 6-deoxyerythronolide B synthase in complex with an antibody fragment, and 5UFV, a cellulose-active polysaccharide monooxygenase from Myceliophthora thermophila. He is a deposition author on 5UFV and an author of 6C9U's primary citation.", "Q: What did Marc propose about how the OSM Phe 160/Lys 163 pair might engage LIFR?\nA: Marc Deller proposed two possibilities: the aromatic Phe 160 could bind into a hydrophobic patch on LIFR, effectively reversing the Epo\u2013EpoR site 1 arrangement in which a receptor phenylalanine packs into a ligand pocket, or the basic Lys 163 could form a salt bridge with an acidic residue on LIFR, as seen for Arg 88 of IL-4 and Asp 72 of IL-4R in the IL-4\u2013IL-4R crystal structure.", "Q: Where can I read Marc's post on the HCV E2 and AP33 structure?\nA: The post is \"Structure of the Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Region 412-423 Bound to the Broadly Neutralizing Antibody AP33\", in the structures category of Marc Deller's blog. You can read it at https://marcdeller.com/structure-of-the-hepatitis-c-virus-envelope-glycoprotein-e2-antigenic-region-412-423-bound-to-the-broadly-neutralizing-antibody-ap33/.", "Q: What is novel about the Hda dimer?\nA: The Hda dimer adopts an oligomeric assembly novel for AAA+ proteins, in which the arginine finger crucial for ATP hydrolysis is fully exposed and available to hydrolyze DnaA-ATP by a typical AAA+ mechanism. The sliding clamp binding motifs at each monomer's N terminus are partially buried and combine to form an antiparallel beta-sheet at the dimer interface.", "Q: How does AI feature in crystal harvesting?\nA: Marc Deller notes that AI-based imaging software now scores crystallisation droplets automatically with error rates below 5%, around 95% accuracy, using training datasets of over 50,000 data points to distinguish crystals from artefacts and precipitates.", "Q: Why is the choice of PEG 35000 in Marc's OSM crystallisation notable?\nA: In his thesis Marc Deller notes that the OSM crystals used for structure determination were grown from PEG 35000, and that this was the first reported use of that reagent for growing protein crystals.", "Q: Which of Marc's structural biology posts deal with crystallography methods?\nA: The Structural Biology page lists crystallography method posts including \"AXIS: Closing the Loop on Crystal Detection\", \"Engineering Protein Stability: Unlocking the Secrets of Successful Crystallization\", \"Unlocking Membrane Protein Structures: The Crystallization Revolution Reshaping Drug Discovery\", and \"Revolutionizing Protein Crystallography: The Synergy of TELSAM-DARPin Fusion and Electric Field Optimization\".", "Q: What was Marc Deller's contribution to the KRAS G12D IND?\nA: Marc Deller is a co-author of the 2025 Journal of Medicinal Chemistry paper reporting the discovery of INCB159020, an orally bioavailable KRAS G12D inhibitor, which supported a KRAS IND filing at Incyte. As the structural biology lead he contributed the crystallography and structure-based drug design behind the program. The work balanced potency and ADME properties to tackle the previously \"undruggable\" KRAS G12D mutation and aimed to broaden treatment beyond KRAS G12C therapies used in lung cancer.", "Q: At what resolution was the Fic family protein SO4266 solved?\nA: SO4266 from Shewanella oneidensis MR-1 was solved at 1.6 \u00c5 resolution, with Das D as first author and Marc Deller contributing crystallography as part of the Joint Center for Structural Genomics (DOI 10.1002/prot.22338).", "Q: Which methods papers of Marc's are most cited?\nA: Marc Deller's methodological and review writing is among his most cited work: \"Protein stability: a crystallographer's perspective\" (2016) has 440 citations, \"Models of protein-ligand crystal structures: trust, but verify\" with Bernhard Rupp (2015, Journal of Computer-Aided Molecular Design) has 130, \"Approaches to automated protein crystal harvesting\" (2014) has 55, and \"Validation of protein-ligand crystal structure models: small molecule and peptide ligands\" with Pozharski and Rupp (2017) has 7.", "Q: Which FDA approval does Marc cite in the extracellular degradation post?\nA: Marc reports that on 1 May 2026 the FDA cleared Arvinas and Pfizer's vepdegestrant (Veppanu) for adults with ESR1-mutant, ER-positive, HER2-negative advanced breast cancer, ahead of its PDUFA date and on the strength of a 43% reduction in the risk of disease progression versus fulvestrant in the Phase 3 VERITAC-2 trial. He traces the modality back to Craig Crews and Ray Deshaies' original 2001 chimaera.", "Q: Which foundational tools and numbers does Marc cite in \"The New Architects\"?\nA: Marc cites ProteinMPNN, published in Science in 2022, which redesigned sequences with a 52.4% native recovery rate that left Rosetta well behind, and RFdiffusion, published in Nature in 2023, which denoises full backbones into existence including symmetric oligomers, motif scaffolds and picomolar binders confirmed by cryo-EM at near-atomic accuracy.", "Q: What is PGT122 and what did the structure show about its epitope?\nA: PGT122 is a highly potent broadly neutralising antibody of the PGT121 family, a family that neutralises about 70% of circulating HIV-1 viruses. The trimer structure showed that the complete PGT122 epitope is far more complex than previously appreciated, involving the gp120 V1 loop and the V3 base together with four glycans \u2014 N137, N156/N173, N301 and N332 \u2014 with the N332 oligomannose glycan sitting at the junction between light chain CDR2 and the 26-residue heavy chain CDR3.", "Q: What is the hybrid delivery system described in the patent?\nA: The hybrid delivery system in claim 4 combines two routes: the PET-degrading enzymes and MHETase are administered orally, while the oxidative enzymes CotA-Laccase and engineered HRP are given by injection. This pairs each enzyme class with the delivery route best suited to reaching its target plastics.", "Q: What is the JCSG Top96 condition E01?\nA: Well E01 of the JCSG Top96 screen is 1.4 M tri-Sodium citrate, 0.1 M Sodium HEPES pH 7.5, a salt-class condition from Hampton Crystal Screen #38 (JCSG-III C05). It produced 10 PDB structures, and the recommended cryoprotectant is 10% ethylene glycol.", "Q: Where were the diffraction data collected?\nA: X-ray diffraction data were collected at Advanced Photon Source beamlines 23ID-B and 23ID-D, and at the Stanford Synchrotron Radiation Lightsource on beamlines 12-2 and 11-1. Data were processed with HKL-2000 or XDS, phased by molecular replacement with Phaser, rebuilt in Coot and refined with Phenix.", "Q: Which clinical WRN compounds are cited in the covalent-warhead post?\nA: Marc Deller cites HRO761 and RO7589831 as WRN inhibitors under clinical investigation that have demonstrated proof-of-concept efficacy with manageable toxicity profiles, alongside a Phase I response rate of 14% in heavily pretreated MSI-H patients.", "Q: Does DeepCovalent name its advisors on its website?\nA: DeepCovalent's public pages describe the company, its OS3D engine, its pipeline and its partnership models, but do not list individual advisors. Marc Deller's own biography is the source for his role there as Scientific Technical Advisor from October 2025.", "Q: How long did MicroPlastic Blaster take to build and how did it grow?\nA: Marc describes a structured six-hour timeline: hour one produced a working prototype from a single prompt, hour two added Elora's teal and cyan branding plus in-game microplastic facts, hour three added particle effects, screen shake and Web Audio API sound, hour four handled mobile touch and canvas scaling, and hours five and six put it live for cross-browser QA. It grew from about 200 lines and 31 KB in version 1 to over 2,000 lines and 66 KB by version 115.", "Q: Can Marc consult for my company?\nA: Marc advises companies on structural biology, protein science and computational drug discovery; he currently advises DeepCovalent alongside his CSO role at Elora Therapeutics. The best way to explore a collaboration is to email him at marc@marcdeller.com.", "Q: What is MRTX1133 and how does INCB159020 differ from it?\nA: MRTX1133, from Mirati Therapeutics, was the first highly potent reversible KRAS G12D inhibitor and was efficacious in a variety of tumour models, but its low oral bioavailability meant efficacy had to be established with intraperitoneal dosing and significant formulation work was needed to reach clinically relevant oral exposure. The Incyte series that Marc Deller contributed structures to is structurally distinct from MRTX1133 and was specifically optimized for oral exposure, reaching 42% bioavailability in nonhuman primate.", "Q: What is deuterium-exchange mass spectrometry used for in construct design?\nA: DXMS exposes a protein to deuterated solvent and measures how readily backbone amide protons exchange, which reports on solvent exposure and local secondary-structure integrity. Marc Deller's review describes how DXMS-guided trimming of flexible regions produced a crystallizable construct of a putative Salmonella typhimurium ethanolamine-utilization protein, solved for residues 98\u2013229 and deposited as PDB entry 2pyt.", "Q: Which of Marc Deller's structures contain zinc?\nA: Zinc appears in 3IRB (the SSO2064 zinc-ribbon OB-fold), 3H50 (XcTcmJ with zinc acetate), 3H0N (Jann_2411), 3CGH (BT_3984), 3D00 and 2GVI and 2GLZ (the FmdE-like proteins), and 3O0F and 3E0F (the Bifidobacterium phosphoesterase, which also carry iron and phosphate).", "Q: What was Marc's role and employer for the CDK2 structure?\nA: Marc Deller contributed the structural biology and crystallography as part of Incyte's structural biology group in Wilmington, Delaware, where he was a Research Fellow in Drug Discovery from March 2022 to December 2024 after serving as Associate Director of Structural Biology from 2018. CDK2 was one of the oncology programmes he supported there.", "Q: What does Marc mean by a deep toolkit applied selectively?\nA: On his About page Marc Deller says he works with Cryo-EM, X-ray crystallography, biophysics, and modern AI tools, using them where they genuinely accelerate development rather than add noise. The emphasis is on selective application rather than adopting every available technique.", "Q: How well did the PGT128/8ANC195 trimer structure diffract?\nA: The BG505 SOSIP gp140 trimer complex with PGT128 and 8ANC195 was solved by X-ray diffraction at about 4.60 \u00c5, a modest resolution typical of these large, heavily glycosylated Env trimer crystals. Even so, it was sufficient to define the complete antibody epitopes.", "Q: How were the OSM crystals cryoprotected?\nA: The optimum protocol was to pass P212121 crystals briefly through a 2 \u03bcl drop of mother liquor (30 % PEG 35000, 0.25 M sodium acetate, pH 7.5) overlaid with 2 \u03bcl of 30 % PEG 8000 in the same buffer, then cool them directly in liquid propane and transfer to a 100 K nitrogen stream. Marc Deller had also tested MPD, glycerol, dioxane and PEG 8000 at 5\u201325 % as conventional cryoprotectants.", "Q: What is the WRN inhibitor patent about?\nA: \"Bicyclic Compounds as Inhibitors of WRN\" is a pending patent covering bicyclic small-molecule inhibitors of the Werner syndrome helicase (WRN), an oncology target. Marc Deller contributed structural biology to the WRN program at Incyte.", "Q: Why does the sulfur anomalous signal make mounting harder?\nA: The sulfur anomalous signal used for S-SAD phasing increases at longer X-ray wavelengths, but so do scattering and absorption by the aqueous solvent and the microloop. Marc Deller's review explains that the resulting loss of intensity accuracy can obscure the already weak sulfur signal, which is why loopless mounting methods were devised.", "Q: Where can I read Marc's CDK2 structure post?\nA: The post \"Discovery of (4-Pyrazolyl)-2-Aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\" is in the structures category of marcdeller.com. You can read it at https://marcdeller.com/discovery-of-4-pyrazolyl-2-aminopyrimidines-as-potent-and-selective-inhibitors-of-cyclin-dependent-kinase-2/.", "Q: Who else is developing WRN inhibitors according to this post?\nA: Marc Deller notes that HRO761, a structurally related compound developed by Novartis, entered human trials in 2023 and has reported encouraging early data in patients with MSI-H tumours refractory to standard therapies.", "Q: Where can I read Marc's thesis?\nA: Marc's thesis PDF is available on his website at https://marcdeller.com/wp-content/uploads/2025/08/marc_deller_thesis.pdf.", "Q: Where did Marc do his undergraduate degree?\nA: Marc earned a First Class Honours BSc in Biochemistry and Molecular Biology at the University of Leeds, where his mentor Prof. Tony North inspired him to pursue protein structure. His degree included a one-year industrial placement at Boots Pharmaceuticals in the team that developed Advil/ibuprofen/Nurofen.", "Q: What software does GOBSMACKED use?\nA: GOBSMACKED runs on Flask, gunicorn, nginx and SQLite, with UniProt, InterPro/Pfam, KLIFS, GPCRdb, RCSB PDB, AlphaFold DB and RDKit for preparation; ESMFold, OpenFold, PDBFixer, PandaDock, OpenMM, openmmforcefields/OpenFF, Boltz-2 and MDTraj in the run bundle; PLIP, PandaMap, biotite, tmtools, PoseBusters criteria and gemmi for analysis; and Mol* and Plotly.js in the browser.", "Q: What are the enemies in MicroPlastic Blaster?\nA: Four polymers plus two branded consumer-plastic sprites: PET (purple, 3 HP, slow, 30 points, block shape), PP (red, 1 HP, fast, 15 points, jagged), PE (amber, 2 HP, medium, 20 points, foam), PS (green, 2 HP, fastest, 25 points, circle), an Aquafina bottle with a blue halo, and a Starbucks cup with a green halo.", "Q: What are the GOBSMACKED run settings?\nA: GOBSMACKED offers three docking modes \u2014 hybrid (search plus GNN rescoring), flex (flexible side chains) and dock (rigid and fastest) \u2014 plus the number of poses, production MD length in picoseconds, frame interval, an optional Boltz-2 affinity prediction scored on both the docked pose and the relaxed complex, the number of frames scored after MD, a run title and a visibility setting for the public Runs table.", "Q: Which papers underpin Marc's cryo-EM grid post and where can I read it?\nA: The post draws on Rubino and colleagues in PNAS on outrunning protein diffusion to the air-water interface, Cherepanov and colleagues in Chemistry Europe on large temperature-jump nanosecond hyperquenching, D'Imprima and colleagues in eLife on protein denaturation at the air-water interface, Xu and colleagues in Frontiers in Molecular Biosciences on recent sample-preparation advances, and a NanoImaging Services guide to cryo-grid production. It is at https://marcdeller.com/outrunning-diffusion-the-fast-lane-in-cryoem-grid-preparation/.", "Q: What advantages over DNA-encoded libraries does Marc identify?\nA: Marc notes that removing the DNA tag lifts the restriction on chemistry, allowing palladium-catalysed cross-couplings, heterocyclizations, nucleophilic aromatic substitutions and harsh protecting group removals that would degrade a genetic barcode. It also removes interference in binding assays for nucleic-acid-binding targets, allows compounds to explore broader conformational diversity, and eliminates false positives caused by DNA-target interactions.", "Q: What permeation enhancers and targeting strategies does Marc describe?\nA: Marc describes SNAC and 5-CNAC, which transiently open tight junctions, and cell-specific targeting through enterocyte receptors such as the proton-coupled folate transporter and M-cell markers such as glycoprotein 2, which enable receptor-mediated endocytosis and achieve roughly 10-fold higher bioavailability than untargeted formulations.", "Q: Who was Marc's doctoral supervisor?\nA: Marc's D.Phil. supervisor was Prof. E. Yvonne Jones at the University of Oxford.", "Q: Where can I read Marc's post on cellular uptake of microplastics and intracellular PETase?\nA: You can read it at https://marcdeller.com/cellular-uptake-of-microplastics-pathways-persistence-and-therapeutic-potential/, published on 7 November 2025.", "Q: What is the mechanism of action for these CDK2 inhibitors?\nA: They competitively inhibit ATP binding, reducing retinoblastoma protein phosphorylation and inducing G1/S cell cycle arrest and therapy-induced senescence in CCNE1-amplified models, a profile that avoids the dose-limiting toxicities of pan-CDK inhibitors such as dinaciclib.", "Q: Summarize the key findings of the 2017 PMO secondary sphere paper.\nA: The paper, co-authored by Marc Deller, reported the 2.45 \u00c5 crystal structure of MtPMO3* (PDB 5UFV), identified a conserved H-Xn-Q-X-Y hydrogen-bonding motif across PMO families, and showed through activity assays and EPR that H161 and Q167 stabilize bound oxygen, H161 mediates proton transfer to the Cu(II)\u2013superoxo species, and Q167 tunes the copper electronics through its hydrogen bond to Y169. Together the data support a copper-oxyl intermediate and provided the first detailed mechanistic insight into the AA9 family.", "Q: What is PDB entry 3BY7?\nA: 3BY7 is a protein structurally similar to Sm/LSm-like RNA-binding proteins, JCVI_PEP_1096686650277, from an uncultured marine organism, at 2.60 \u00c5 in space group C 1 2 1 with 500 polymer residues and 55.74 kDa, released on 29 January 2008. It is cited in the 2009 Proteins paper describing a novel Sm-like protein of putative cyanophage origin.", "Q: Which debate did the polyketide synthase didomain structure settle?\nA: The structure resolved a long-standing debate in polyketide synthase biology by demonstrating that the antibody-stabilised extended module conformation is catalytically functional for both chain elongation and modification. Marc Deller's post describes it as a landmark structure validated by both X-ray crystallography and SAXS.", "Q: What was Marc Deller's role in the 2024 CDK2 inhibitor paper?\nA: Marc Deller was a co-author on the 2024 CDK2 paper as part of Incyte's structural biology group. He and Leslie B. Epling determined the co-crystal structure of the lead compound 17 bound to CDK2, which was solved to 1.55 \u00c5 resolution and deposited in the Protein Data Bank as entry 8UV0. The crystallography explained how the inhibitor engages the ATP pocket and rationalized its potency and selectivity.", "Q: What were Marc's school grades?\nA: Marc's best grades were A's in biology, chemistry and physics at GCSE; he got B's in biology and chemistry at A-level, then a First Class Honours degree in Biochemistry from Leeds and a doctorate in biophysics from Oxford.", "Q: What were the space-based crystal harvesting systems?\nA: Two of the fourteen systems were designed for remote telerobotic operation in space. The Protein Crystal Growth Demonstration System, built by the University of Washington and Boeing, demonstrated telerobotics over standard internet protocols for the International Space Station, and the Crystal Preparation Prime Item, built by Oceaneering Space Systems for the ISS X-ray Crystallography Facility, used a liquid\u2013liquid bridge between opposed pipettes as its harvesting workspace.", "Q: What does Marc say about antibody-specific versus general protein models?\nA: Marc argues that domain-specific training data matters: AntiFold, an inverse folding model fine-tuned from ESM-IF1 on solved and predicted antibody structures, outperforms ProteinMPNN, ESM-IF and LM-Design for Fab CDR design, particularly at capturing conserved patterns in structurally constrained regions such as CDR-H1 and the CDR-H3 stem. He also notes Paragraph's equivariant graph neural network predicting paratopes accurately without needing antigen information.", "Q: Which residues form the GLP-1R orthosteric pocket according to Marc?\nA: Marc lists Y152 in TM1, K197 in TM2, Q234 and Y241 in TM3, W306 and R310 in TM5, E364 and H363 in TM6, and L384 in TM7, forming a deep amphipathic pocket. He notes that His1 of peptide agonists makes critical contacts deep in this cavity with E364, R310, W306 and Y241, and that Glu9 also contributes hydrogen bonds and hydrophobic contacts.", "Q: How was SsgB's conservation demonstrated?\nA: Sporulation-specific cell division of Streptomyces coelicolor ssgB mutants was restored by introducing distant ssgB orthologues from other actinomycetes, and the number of septa and spores in the complemented null mutants was dictated by which specific orthologue was expressed.", "Q: Which strategies did the team try to improve the diffraction of the Env trimer crystals?\nA: According to the 2013 Science methods, the team explored the use of single-chain Fv instead of Fab, limited in-situ proteolysis, varying the temperature of crystal growth, additives screening, crystal cross-linking, dehydration and annealing \u2014 all with limited success \u2014 and also tried a SOSIP construct further truncated to gp41 position 650. This kind of systematic pre- and post-crystallisation optimisation is exactly the crystallography expertise Marc Deller brought from the Joint Center for Structural Genomics.", "Q: Why does FGFR selectivity matter clinically?\nA: First-generation pan-FGFR inhibitors often cause hyperphosphatemia and lose potency against gatekeeper mutations. High selectivity for FGFR2/3 over FGFR1 may reduce that dose-limiting side effect, and early pharmacodynamic studies in animal models showed effective pathway inhibition without significant changes in serum phosphate.", "Q: What is the full preclinical profile of INCB159020?\nA: Compound 23 gave KRAS G12D exchange, pERK and viability values of 4.4 nM, 33 nM and 78 nM, with wild-type values of 205 nM, 2635 nM and 2251 nM. It showed Caco-2 permeability of 3.6 \u00d7 10\u207b\u2076 cm/s, human clearance of 1.0 L/h/kg, 1.1% free fraction, solubility greater than 900 \u03bcg/mL in simulated gastric fluid and 79 \u03bcg/mL in FaSSIF, moderate CYP inhibition (2D6 at 7.2 \u03bcM, 3A4 at 17 \u03bcM), an hERG IC50 of 11.2 \u03bcM, and no time-dependent CYP inhibition or PXR activation.", "Q: How can I contact Marc?\nA: Marc can be reached by email at marc@marcdeller.com or by phone on (302) 358-6093. His website is marcdeller.com and his LinkedIn profile is linkedin.com/in/marccdeller.", "Q: Is Marc a godparent to anyone?\nA: That's outside what I know about Marc, I'm afraid. I stick to his science and his projects.", "Q: What does Marc say about transfer learning and multitask models for \u0394\u0394G?\nA: Marc reports that transfer learning from megascale stability datasets of roughly 400,000 reliable \u0394\u0394G\u00b0 points from protease sensitivity assays boosts robustness for stability prediction, though structural fidelity of predicted models limits transfer to binding tasks. He cites Pythia-PPI, a multitask framework trained on augmented SKEMPI-like interfaces of around 400,000 mutations, achieving over 10,000 predictions per minute at r = 0.785. The post is at https://marcdeller.com/optimizing-data-volume-and-diversity-for-robust-antibody-antigen-%ce%b4%ce%b4g-prediction/.", "Q: What does Marc say still stands between the KTH result and a clinical therapy?\nA: Marc writes that the enzyme would need to be humanized to avoid immune clearance, encapsulated for delivery \u2014 perhaps within red blood cells \u2014 and tested in animal models, and that the study did not address immunogenicity, biodistribution or long-term stability in blood. What it establishes is the enzymatic principle. The post is at https://marcdeller.com/kth/.", "Q: What is the FmdE family and what did Marc Deller help determine about it?\nA: Pfam family PF02663 (FmdE) is named for the fmdE gene of the methanogenic archaeon Methanobacterium thermoautotrophicum, and 13 family members are co-transcribed with genes for subunits of molybdenum formylmethanofuran dehydrogenase, an enzyme involved in microbial methane production. Marc Deller co-authored the 2010 Acta Crystallographica Section F paper describing the first crystal structures from this family.", "Q: Summarize Marc Deller's post \"The Protein Architect: How Generative AI Is Building the Drugs of Tomorrow\".\nA: Published 13 August 2025, this post argues that generative AI has moved protein engineering from prediction to creation. Marc Deller explains that AlphaFold began by accurately predicting structures from sequence, while newer models such as Genie design completely new proteins that do not exist in nature.", "Q: Where can I read Marc's TrkA-N domain post?\nA: The post \"Crystal Structure of TrkA-N Domain Protein from Thermotoga sp. RQ2\" sits in the structures section of Marc Deller's website. You can read it at https://marcdeller.com/crystal-structure-of-trka-n-domain-protein-from-thermotoga-sp-rq2/.", "Q: Is Marc's doctoral thesis listed on Google Scholar?\nA: Yes. \"Structural studies of cytokines and cytokine receptors\" (University of Oxford, 1999) appears on his Google Scholar profile, alongside a 1999 Nucleic Acids Symposium Series abstract on the cloning, expression, characterization, purification and preliminary crystallographic studies of the haemopoietin cytokine oncostatin-M.", "Q: Why was the HIV-1 envelope trimer so hard to crystallise?\nA: As described in the 2013 Science paper Marc Deller co-authored, the cleaved native Env trimer has to undergo conformational changes during receptor binding and fusion, which makes it metastable and has hindered both structure determination and vaccine development. It also carries extensive N-linked glycosylation \u2014 on average 81 glycosylation sites per trimer \u2014 which further complicates X-ray studies, and membrane-associated forms of Env are much harder to express and purify than soluble ones.", "Q: When was the CALR antibody patent application filed and published?\nA: The international application was filed on 7 December 2022 and published on 15 June 2023. It claims priority from three US provisional applications: 63/287,394 filed 8 December 2021, 63/288,479 filed 10 December 2021, and 63/421,052 filed 31 October 2022.", "Q: Why does Marc list his certifications?\nA: On the My Learning page Marc Deller says the certifications position him to bridge external partners and internal AI/software teams while driving innovative therapeutic discovery programs. He describes himself as prepared to excel in AI-driven industries.", "Q: What was the key change to the hinge-binding motif in povorcitinib?\nA: The 7H-pyrrolo[2,3-d]pyrimidine hinge binder used in earlier compounds is prone to CYP-promoted oxidation, so it was replaced with a 3,5-dimethylpyrazole. That change cost 44-fold in enzymatic potency (JAK1 IC50 moving from 0.25 nM for compound 18 to 11 nM for povorcitinib) but nearly doubled JAK1 isoform selectivity and dramatically improved the pharmacokinetic profile in rat, monkey and dog.", "Q: What is QSARtuna and why does Marc mention it?\nA: QSARtuna is an automated QSAR modelling platform that Marc describes as encapsulating best practice in a transparent pipeline covering data ingestion, deduplication, descriptor calculation, model calibration with methods such as VennABERS, and uncertainty quantification. He argues platforms like it democratise QSAR development for novices and experts alike, and the post is at https://marcdeller.com/empowering-qsar-modeling-optimal-data-practices-and-cutting-edge-tools/.", "Q: What is the adaptability course Marc completed?\nA: Marc Deller completed \"Building an Adaptability Mindset in the Age of AI\", a LinkedIn Learning certificate issued on 18 March 2025. His post explains that researchers increasingly need to adjust to changing tools, workflows and technologies, and that the training helps scientists stay effective as AI reshapes analysis, modelling and discovery work.", "Q: What does OS3D stand for?\nA: OS3D stands for One-Step-DeNovo drug discovery, DeepCovalent's generative AI engine. It is described as a multimodal, multi-objective engine that designs clinically viable, synthesis-ready de novo small molecules in a single forward pass.", "Q: What does PDB entry 9E5F contain?\nA: 9E5F is a crystal structure of human GTPase KRas bearing the G12D mutation in complex with Compound 23, ligand code A1BEJ, together with GDP and magnesium. Marc Deller's post records an X-ray structure at 1.35 \u00c5 released on 22 January 2025.", "Q: What did the Hasegawa study on polystyrene microplastics find, as described by Marc?\nA: Marc reports that Hasegawa et al. (2024, Frontiers in Immunology) exposed C57BL/6J mice to polystyrene microplastics at 1000 and 5000 \u03bcg/L for six weeks and found dyslipidemia and fatty liver without leaky gut syndrome. The study identified NK cell-mediated intestinal inflammation as a novel toxicity mechanism, along with decreased bacterial diversity and reduced Bacteroidetes.", "Q: How was AXIS built and trained?\nA: Marc explains that AXIS uses transfer learning from the self-supervised DINOv2 vision transformer, trained on nearly 77,000 images consolidated from the MARCO dataset's half-million-image repository into crystal and non-crystal classes, then fine-tuned on local facility data. It works on both visible-light and UV-fluorescence images, the latter giving intrinsic protein fluorescence signals that help confirm that a hit is proteinaceous.", "Q: How fast did the VRC01-like heavy chains mature?\nA: Remarkably fast \u2014 heavy-chain somatic hypermutation in the DRVIA7 lineage rose from an average of 17.0% in 2006 to 22.9% by 2008, an increase within only two years. This challenged the prevailing view that heavy-chain maturation was the slow, time-limiting step in VRC01 development.", "Q: Where does Marc write about his projects?\nA: On his blog at marcdeller.com, which has posts on AlphaFraud (the \"AlphaFraud Weekly Watcher\"), BoltzMaker, ChemSage and other projects, and on each project's GitHub README at github.com/bellcheddar.", "Q: Why might the copper in the MtPMO3* crystal structure have been reduced?\nA: The tyrosyl oxygen was 2.7 \u00c5 from the copper, too long for a formal bond even with Jahn\u2013Teller distortion, and no solvent-derived ligands were seen at the active site of chain C. The authors noted this could reflect a feature of the Cu(II) coordination sphere or, more likely, photoreduction of Cu(II) to Cu(I) by the X-ray beam during data collection, which is common in PMO structures.", "Q: Has Marc ever been an expert witness?\nA: That isn't something I know about. For anything like that, the right person to ask is Marc himself, at marc@marcdeller.com.", "Q: Which bottleneck does Marc say microfluidics addresses?\nA: Marc identifies the slow, artisanal process of growing high-quality protein crystals, which can take months of manual effort and consume large amounts of precious protein. He says microfluidic screening compresses those timelines from months to days and raises the chances of finding suitable crystallization conditions.", "Q: What is WRN and why does Incyte care about it?\nA: WRN is a helicase enzyme, and it's a target of interest because certain cancers with microsatellite instability (MSI-high tumors) become dependent on WRN for survival, making it a synthetic-lethality drug target. Marc Deller contributed structural biology to Incyte's WRN inhibitor program.", "Q: Can you do structural biology without your own lab? What does Marc say?\nA: In \"Unlocking Protein Structure Insights without a Lab: Harnessing CRO Services for Structural Biology Research\", posted 20 August 2025, Marc Deller argues yes, because Contract Research Organizations offer full gene-to-structure pipelines. He notes that building an in-house experimental structure determination capability can demand investment exceeding $5 million in equipment, space and specialist personnel.", "Q: What is Marc's literature review on microplastics and cancer?\nA: \"Microplastics and Cancer: The Mounting Evidence via Literature Review\", published 16 June 2026, is Marc Deller's PubMed-based survey of the 2022\u20132026 literature linking micro- and nanoplastics to cancer, grouped by study type with all DOIs verified live. It includes 28 references covering reviews, in vitro work, in vivo models and detection methods.", "Q: Who are the co-founders of Elora Therapeutics?\nA: Marc C. Deller and Paul Swartzendruber are the co-founders of Elora Therapeutics. Marc's post shows them accepting the Best Team Award at The Ventures \u2013 Austin 2026.", "Q: What was the broader significance of the LIFR work?\nA: Marc framed the LIFR work as the first steps toward structural analysis of the LIFR-LIF-gp130 interaction. Solving that interaction would provide insight into the molecular basis of signal transduction by the haemopoietin cytokines and the functional epitopes involved, complementing the OSM structure determined elsewhere in the thesis.", "Q: What did Marc complete at Anthropic Academy?\nA: Anthropic, via its Skilljar Academy, issued Marc Deller four course completions in May 2026: Claude 101, Claude Code 101, Introduction to Claude Cowork, and AI Fluency: Framework & Foundations. Together they recognise applied training in Claude usage, agentic coding workflows, cowork orchestration and the AI Fluency 4D model for responsible human\u2013AI collaboration.", "Q: What did mutant 5 show?\nA: Mutant 5, in which only Arg29 and Tyr32 of the light chain were mutated to alanine, retained picomolar binding to the T117v2 epitope scaffold but its neutralising breadth collapsed to 44%. Its crystal structure showed no glycerol bound at the lipid site, indicating the lipid-binding pocket had been severely disrupted \u2014 strong evidence that lipid engagement is required for neutralisation.", "Q: Which organism does the 3JXO structure come from?\nA: The TrkA-N domain protein in PDB entry 3JXO comes from Thermotoga sp. RQ2, a thermophilic bacterium closely related to Thermotoga maritima. The structure was determined at 1.55 \u00c5, the highest resolution among Marc Deller's Trk-related entries.", "Q: What is CatAttack, as covered on Marc Deller's blog?\nA: In \"Cats Confuse Reasoning LLM: Query-Agnostic Adversarial Triggers for Reasoning Models\", published 27 August 2025, Marc Deller covers CatAttack, an automated pipeline that discovers adversarial triggers on a low-cost proxy model, DeepSeek V3, and transfers them to stronger reasoning models such as DeepSeek R1 and Qwen-32B.", "Q: Which proteins have FragLites been applied to in the work Marc reviews?\nA: Marc describes FragLites mapping protein\u2013protein interaction hotspots on cyclin-dependent kinase 2 and the CDK2\u2013cyclin A complex, uncovering both known and previously uncharacterised sites, and mapping ligand binding sites on the bromodomains BRD4 and ATAD2. He notes they recapitulated interactions found by full fragment screening but with greater speed and less experimental resource.", "Q: What could and couldn't be concluded about the LIFR-LIF complex?\nA: A peak shift on adding ligand indicated that some receptor-ligand interaction was occurring, and the early-eluting peak contained both LIF and LIFR components on SDS-PAGE. However, the exact state and stoichiometry could not be determined because of the non-globular arrangement and large Stokes' radii of the assemblies. Marc noted that a full high-affinity complex likely also requires gp130, since LIF forms a heterotrimeric complex with gp130 and LIFR.", "Q: What is the WRN patent that Marc Deller is named on?\nA: Marc Deller is a named co-inventor on US Patent Application Publication US 2025/0215006 A1, \"Tricyclic Compounds as Inhibitors of WRN,\" assigned to Incyte Corporation and published on July 3, 2025. The application discloses small-molecule covalent inhibitors of the Werner syndrome helicase (WRN) for the treatment of microsatellite instability-high (MSI-H) and mismatch repair deficient (dMMR) cancers.", "Q: What crystallisation conditions did Marc use for oncostatin M?\nA: In the 2000 Structure paper, Marc Deller crystallised a 21.4 kDa fragment of human oncostatin M (residues 1\u2013187) at 8 mg/ml in 50 mM Tris-HCl, 150 mM NaCl, pH 7.5, by hanging-drop vapour diffusion at 22\u00b0C against a reservoir of 30% w/v PEG 35,000 and 0.25 M sodium acetate, pH 7.5. Crystals of about 0.3 \u00d7 0.2 \u00d7 0.2 mm grew within 1\u20133 days, belonged to the orthorhombic space group P212121 (a = 35.8 \u00c5, b = 53.1 \u00c5, c = 106.8 \u00c5), contained one molecule per asymmetric unit with about 48% solvent, and diffracted to 2.2 \u00c5 at the synchrotron.", "Q: Which entries are the Thermotoga maritima acetyl xylan esterase structures?\nA: Four entries cover TM0077 acetyl xylan esterase from Thermotoga maritima: 1VLQ at 2.10 \u00c5 in P 1 21 1 (released 24 August 2004, and the largest in the list at 4044 polymer residues and 467.99 kDa), 3M81 the native apo structure at 2.50 \u00c5, 3M82 the PMSF inhibitor complex at 2.40 \u00c5, and 3M83 the paraoxon inhibitor complex at 2.12 \u00c5. The three 3M8x entries are all in P 21 21 21 with calcium bound and cite the 2012 Proteins paper on the thermostable acetyl esterase.", "Q: What did Marc do at Scripps on HIV?\nA: At The Scripps Research Institute, Marc Deller contributed crystallography \u2014 crystallisation, harvesting and X-ray data collection \u2014 to structural studies of HIV-1 broadly neutralizing antibodies in Ian Wilson's group and the Joint Center for Structural Genomics. This 2017 PLoS Pathogens study of the MPER antibody 10E8 and its lipid interactions is one example.", "Q: What was Marc's role in the TrkA octamer structure?\nA: Marc Deller was first author of the 2015 PLoS ONE paper and led the crystallography, work carried out as Target Manager at the Joint Center for Structural Genomics at The Scripps Research Institute, where he was based from 2006 to 2015. The structure itself was released to the PDB in 2010.", "Q: What PDB codes were deposited from this study?\nA: Seven structures were deposited: 5T6L for the 10E8-T117v2 complex, 5T85 and 5T80 for the co-crystals with 06:0 PG and 06:0 PA respectively, and 5SY8, 5TFW, 5T29 and 5T5B for the T117v2 complexes with 10E8 light-chain mutants 1, 2, 3 and 5.", "Q: What structural features give nanobodies their binding advantages?\nA: Marc points to extended CDR3 loops averaging 17\u201318 amino acids that adopt \u03b1-helical conformations and can reach cryptic epitopes larger antibodies cannot access, stabilised by additional disulfide bonds between CDR1 and CDR3. Their compact size also gives superior tissue penetration and the ability to cross the blood-brain barrier, with affinities reaching picomolar levels.", "Q: Can you order me a pizza?\nA: I can't place orders or interact with the outside world like that. I'm chatMCD, here to answer questions about Marc Deller and his work. Anything about his research or companies I can help with?", "Q: What does Marc say AI does for drug discovery?\nA: The AI page says AI accelerates drug discovery by predicting protein structures so researchers can target diseases more efficiently and accurately, and that generative models design novel proteins and simulate complex biological interactions, streamlining experimental workflows and hypothesis testing. The stated outcome is faster development, lower costs, and higher success rates, including medicines for diseases like cancer.", "Q: Where can I find GOBSMACKED?\nA: GOBSMACKED is live at gobsmacked.mdeller.com and its source is at github.com/bellcheddar/GOBSMACKED. Version 1.0.0 was written by Marc C. Deller, D.Phil., and the citation instruction is to name the repository and the version.", "Q: Why is the coating called molecular Velcro?\nA: Marc quotes the commentary by Chakraborty in Nature Water: rather than being chemically selective toward one polymer, the coating offers many weak binding interactions simultaneously, so it grips polyethylene, PET, aged mixed debris and other chemistries with comparable efficiency \u2014 many different kinds of Velcro rather than one specific hook.", "Q: What does the oncostatin M structure look like?\nA: According to Marc Deller's paper, human oncostatin M is a compact barrel-shaped molecule of roughly 20 \u00c5 \u00d7 27 \u00c5 \u00d7 56 \u00c5 with the up-up-down-down four-helix bundle topology. Helix A spans residues 10\u201337, helix B 67\u201390, helix C 105\u2013131 and helix D 159\u2013185, connected by two long overhand loops (AB and CD) and one short BC loop, with the A\u2013D and B\u2013C helix pairs packing together. Two disulphide bridges (Cys6\u2013Cys127 and Cys49\u2013Cys167) tether the N-terminal loop and the AB loop to the core, and the long CD loop was too mobile to model.", "Q: How do the fused bicyclic pyrimidones inhibit WRN?\nA: They function as mixed ATP-competitive inhibitors that mimic structural features of ATP while inducing allosteric conformational changes that lock WRN in an inactive state. The hydroxy pyrimidine moiety coordinates the catalytic glutamine residue, while the bicyclic core exploits an unusually polar, arginine-rich pocket at the domain interface, a binding mode distinct from covalent inhibitors that target cysteines.", "Q: Thanks!\nA: You're welcome! Come back any time you want to know more about Marc, his science or his projects. Have a great day!", "Q: What is the gp120 \"supersite of vulnerability\"?\nA: The 2013 Nature Structural and Molecular Biology paper \"Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\", co-authored by Marc Deller, showed that the Asn332-dependent antigenic region is highly accessible and far more extensive than previously appreciated, allowing multiple binding modes and varied angles of approach. The team therefore described it as a supersite of vulnerability for antibody neutralization.", "Q: Where can I read Marc's post on enzymes for MASLD?\nA: You can read it at https://marcdeller.com/clearing-the-microscopic-threat-can-plastic-eating-enzymes-treat-liver-disease/, published 4 November 2025 on marcdeller.com.", "Q: What human evidence frames the photoacoustic post?\nA: Marc frames it with microplastics first quantified in human blood in 2022 by Leslie et al., found in carotid artery plaque where their presence tracked with higher rates of heart attack, stroke and death (Marfella et al., New England Journal of Medicine, 2024), and accumulating in the brain with the heaviest burdens in samples from people who had dementia (Nihart et al., Nature Medicine, 2025).", "Q: How does Marc describe the way PocketGen works?\nA: Marc explains that PocketGen uses a graph transformer architecture capturing interactions at the atomic, residue and ligand level, generating both the amino acid sequence and the precise 3D atomic structure of the pocket. He notes it runs about ten times faster than older physics-based methods and produces high-affinity pockets with remarkable accuracy.", "Q: What four learning areas does Marc highlight?\nA: The page highlights vibe coding with AI, covering dashboards, interactive apps, and teamworking; data analysis and visualisation with tools like R and Shiny for biological data interpretation; project management, with advanced training in leadership and adaptability; and cloud technologies, with an AWS certification in modern cloud architectures for data processing.", "Q: What is GOBSMACKED?\nA: GOBSMACKED (Ground-truth Overlay for Binding Sites, Modes And Complex Kinetics/Dynamics; gobsmacked.mdeller.com) is Marc's end-to-end fold \u2192 dock \u2192 MD \u2192 analysis app. It folds a sequence with ESMFold if needed, docks a SMILES ligand with PandaDock, runs OpenMM minimisation and short molecular dynamics, and analyses the complex with PandaMap, PLIP and Mol*, annotated with KLIFS, GPCRdb, Pfam and UniProt, to verify a predicted complex against an experimental structure.", "Q: Why does three-dimensional shape matter for WRN selectivity?\nA: Because WRN sits in the RecQ helicase family alongside BLM, RecQ1, RecQ4 and RecQ5, and the ATP pocket residues are conserved, so selectivity is hard-won. Marc Deller writes that a more sculpted scaffold provides additional opportunities to address selectivity, permeability and metabolic stability liabilities in parallel.", "Q: Which dye-decolorizing peroxidase structures did Marc contribute to?\nA: Three: 2GVK, a dye-decolorizing peroxidase from Bacteroides thetaiotaomicron VPI-5482 at 1.6 \u00c5 in P 63 2 2 (released 16 May 2006); 2HAG, a putative DyP-type peroxidase SO_0740 from Shewanella oneidensis at 2.75 \u00c5 in P 43 21 2 (8 August 2006); and 2IIZ, the putative melanin biosynthesis protein TyrA with bound heme from Shewanella oneidensis at 2.30 \u00c5 (14 November 2006). The first two are cited in the 2007 Proteins paper on two novel dye-decolorizing peroxidases, and 2IIZ in the companion TyrA heme-binding paper.", "Q: What does Marc mean by platform innovation?\nA: On his Projects page Marc Deller describes platform innovation as having built automation and lab infrastructure spanning Cryo-EM, X-ray crystallography, crystal harvesting robotics, and scalable discovery workflows.", "Q: How did the phosphoesterase structure correct a database annotation?\nA: Bioinformatics had initially suggested the protein was a DNA polymerase, but the crystal structure together with theoretical analysis correctly predicted phosphoesterase activity, which was then confirmed experimentally. Marc Deller's post presents this as evidence of widespread misannotation in sequence databases and of the power of combining computational prediction with structural genomics.", "Q: Where were the X-ray data collected for the 1B2/KS-AT structure?\nA: X-ray analysis for the 2018 JACS paper was performed on beamlines 11-1 and 12-2 at the Stanford Synchrotron Radiation Lightsource (SSRL), SLAC National Accelerator Laboratory, where co-authors Irimpan Mathews and Tsutomu Matsui are based. SSRL is supported by the US Department of Energy and the NIH National Institute of General Medical Sciences.", "Q: Which chemical biology methods does Marc describe?\nA: Marc describes activity-based protein profiling (ABPP), which uses reactive chemical probes to map the functional state of entire enzyme classes directly in living systems, and chemoproteomics, which maps which proteins a drug-like molecule interacts with across the whole proteome.", "Q: How is the Spacial Score calculated according to Marc's post?\nA: Marc explains that SPS is summed over heavy atoms as the product of a hybridization term h, a stereoisomeric term s, a non-aromatic ring term r, and the squared number of heavy-atom neighbours n\u00b2. Dividing SPS by the total heavy-atom count gives the size-normalized nSPS, which allows fair comparison between molecules of different sizes.", "Q: How does genome mining feature in Marc's enzyme toolbox post?\nA: Marc notes that genome mining continues to uncover new plastic-degrading enzyme variants from environmental and marine sources, including extreme environments, broadening the arsenal available for degradation and therapeutic customisation.", "Q: What is the 2012 Journal of Virology paper on hepatitis C virus that Marc Deller co-authored?\nA: The 2012 Journal of Virology paper \"Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33\", co-authored by Marc Deller, reported the crystal structure of the broadly neutralising antibody AP33 bound to a peptide corresponding to antigenic site 412\u2013423 of the HCV E2 envelope glycoprotein. It was published in Journal of Virology, volume 86, number 23, pages 13085\u201313088, in December 2012.", "Q: What properties does smiles2grid compute?\nA: The CSV carries output index, original index, source index, label, SMILES and canonical SMILES, formula, MW, cLogP, HBD, HBA, TPSA, rotatable bonds, rings, heavy atoms, fraction sp3, aromatic rings and hetero atoms, plus the input JSON path, PDF filename and selection mode.", "Q: Which of Marc Deller's papers was named a top article in structural biology?\nA: Marc Deller's 2016 paper \"Protein stability: a crystallographer's perspective\", written with Leopold Kong and Bernhard Rupp, was featured in the IUCr Journals \"Top articles in structural biology (Spring 2020)\" collection. It has around 370 citations.", "Q: What is Marc Deller's role at Elora Therapeutics?\nA: Marc Deller, D.Phil, is a co-founder and the Chief Scientific Officer (CSO) of Elora Therapeutics, as listed on the pitch deck alongside co-founder and CEO Paul Swartzendruber. His contact on the deck is marc@eloratherapeutics.com.", "Q: Which journal and year was the FGFR2/3 gatekeeper paper published in?\nA: The paper was published in the Journal of Medicinal Chemistry in 2022, volume 65, pages 15433\u201315442, with the DOI 10.1021/acs.jmedchem.2c01366. It was received on 18 August 2022 and published on 10 November 2022.", "Q: What does Marc Deller's post \"Call Me A Jerk\" cover?\nA: Published 3 September 2025, \"Call Me A Jerk: Persuading AI to Comply with Objectionable Requests\" covers research by Zeng and colleagues showing that large language models inherit human susceptibility to social engineering. Marc Deller reports that Persuasive Adversarial Prompts compromise even advanced models such as GPT-4 and Claude with over 92% effectiveness.", "Q: Why does Marc say degraders can reach targets inhibitors cannot?\nA: Marc explains that degradation does not require a functional active site: a degrader only needs a small nook or cranny on the target protein to bind to. He argues this turns the entire protein into a potential target and heralds an era in which no protein is truly undruggable. The post is at https://marcdeller.com/hitting-the-delete-button-on-disease-the-rise-of-targeted-protein-degraders/.", "Q: How many of Marc Deller's publications are indexed in Europe PMC?\nA: Europe PMC indexes 58 papers under the author search \"Deller MC\". The collection spans from Marc Deller's 1999 doctoral work on cytokines and cytokine receptors through to a 2026 Journal of Medicinal Chemistry paper on the JAK1 inhibitor povorcitinib.", "Q: What is ShpA?\nA: ShpA is the histidine phosphotransfer protein of the multistep phosphorelay controlling cell-cycle-regulated stalk biogenesis in Caulobacter crescentus, shuttling phosphoryl groups between the hybrid histidine kinase ShkA and the response regulator TacA. Marc Deller co-authored the 2009 Journal of Molecular Biology paper reporting its crystal structure at 1.52 \u00c5 resolution.", "Q: What is the PDB code for the CDK2 structure in this paper?\nA: The crystal structure of compound 17 bound to CDK2, determined by Marc Deller and Leslie Epling at Incyte, was deposited in the Protein Data Bank under accession code 8UV0. It was solved to 1.55 \u00c5 resolution and shows the inhibitor bound in the ATP-binding site.", "Q: What is PDB entry 3BOS?\nA: 3BOS is a putative DNA replication regulator HDA, SAMA_1916, from Shewanella amazonensis SB2B at 1.75 \u00c5 in space group P 21 21 21, with 484 polymer residues and 56.87 kDa, released on 15 January 2008 with CDP, magnesium and sodium bound. It is cited in \"A structural basis for the regulatory inactivation of DnaA\", Journal of Molecular Biology 2009.", "Q: What is Elora's long-term plan beyond PET?\nA: The deck indicates that Elora's lead program targets PET microplastics with PETase and MHETase, but with staged expansion to additional high-prevalence polymers over time. Its IP strategy similarly includes follow-on polymer expansion, signaling an intent to broaden beyond PET as the platform matures.", "Q: What are VRC01-class antibodies?\nA: They are potent HIV broadly neutralizing antibodies derived from rare IGHV1-2 immunoglobulin heavy chain gene segments that recognise the conserved CD4-binding site across diverse strains, neutralising up to 90% of circulating HIV-1 variants when fully matured. Their development requires extensive somatic hypermutation, with mature antibodies accumulating 10 to 16% sequence divergence from germline.", "Q: What is the REACH system?\nA: REACH, Robotic Equipment for Automated Crystal Harvesting, was developed by NatX-ray with scientists at ESRF beamline FIP-BM30A and performs semi-autonomous harvesting, cryoprotection and flash-cooling as well as in situ data collection from the plate. Marc Deller's review highlights its two-finger piezoelectric microgripper with 1 \u00b5m resolution and a 0\u2013500 \u00b5m opening range, tipped with concave SU-8 pads, and its ability to collect X-ray data from crystals still held in the grippers.", "Q: Where is the post about Marc's project management certificate?\nA: Marc Deller's post \"2025 \u2013 Foundations of Project Management\" is at https://marcdeller.com/foundations-of-project-management/.", "Q: How does the 1.05 \u00c5 cyanobacterial endopeptidase structure compare in quality with typical JCSG structures?\nA: At 1.05 \u00c5 the Anabaena variabilis endopeptidase is among the highest-resolution structures in Marc Deller's Joint Center for Structural Genomics output, with the Nostoc punctiforme homolog at 1.60 \u00c5. The work was published in Structure in 2009 (DOI 10.1016/j.str.2008.12.008).", "Q: What advantage does Marc see in FragLites over conventional fragment screening?\nA: The main advantage is efficiency: a 31-compound library gives comprehensive site mapping at a fraction of the resource cost of a full fragment campaign, allowing early assessment of protein tractability and informing hit-finding strategy. Selective binding profiles also help prioritise functional hotspots over nonspecific or artefactual interactions.", "Q: Is there a PETase question on Marc's chat page?\nA: Yes, one of the suggested conversation starters asks what a plastic-eating enzyme like PETase would say about its role in cleaning up the planet, and what mutation it would request for its next upgrade. It ties the chat page to his microplastics work at Elora Therapeutics.", "Q: What is mdeller.com?\nA: mdeller.com is Marc's web-app server, a small DigitalOcean droplet running nginx and Flask apps on subdomains such as alphafraud.mdeller.com, chatpdb.mdeller.com, codswallop.mdeller.com, gobsmacked.mdeller.com and alphabetti.mdeller.com.", "Q: Summarise Marc's post on LLM prompts in structural biology and cheminformatics.\nA: \"Supercharging Discovery: How LLM Prompts Are Transforming Structural Biology and Cheminformatics\", posted on 17 August 2025, presents ten prompt-engineering workflows that Marc Deller says show the scope of large language models such as GPT-4 and Claude in biopharma. Each example comes with a motivating use case, a sample input and the kind of output to expect.", "Q: Where can I read Marc's article on barcode-free drug discovery screening?\nA: You can read it at https://marcdeller.com/eliminating-dna-barcodes-a-new-era-in-drug-discovery-screening/, published under the title \"Self-Encoded Libraries: A New Era in Drug Discovery Screening\".", "Q: Summarise what DeepCovalent claims OS3D achieves.\nA: DeepCovalent claims OS3D generates optimised, in-silico validated small molecules faster, leaner and beyond known chemistry, going from ideation to synthesis-ready candidates in three weeks or less. It attributes that to a single forward pass that unifies ideation, design and in-silico validation, with generation conditioned at once on target biology, efficacy, ADMET and medicinal chemistry constraints.", "Q: Who wrote the paper on the orally bioavailable KRAS G12D inhibitor?\nA: The paper is Deller, M.C., et al. (2025), \"Discovery of an Orally Bioavailable KRAS G12D Inhibitor\", Journal of Medicinal Chemistry, volume 68, pages 1924-1939, with Marc Deller as first author. He also wrote a blog post titled \"Discovery of INCB159020: An Orally Bioavailable KRAS G12D Inhibitor\".", "Q: What is povorcitinib?\nA: Povorcitinib, also known as INCB054707, is an orally bioavailable isoform-selective JAK1 inhibitor developed at Incyte for inflammatory and autoimmune diseases, described by Zhuo and colleagues in the Journal of Medicinal Chemistry in 2026 and covered in Marc Deller's post on the discovery.", "Q: What does the CODSWALLOP coverage census find?\nA: The CODSWALLOP construct coverage census shows how many of a family's constructs contain each residue of the seed sequence as a depth profile. On lysozyme it finds residues 1 to 18 covered by only 66 of 1,686 constructs, which is the signal peptide that is cleaved and never present in the mature protein, and on p53 it finds residues 361 to 393 in 10 of 267.", "Q: What validation losses did the ChemSage rounds reach?\nA: Round 1 reached 0.389 at iteration 675, Round 2 0.347 at 675, Round 3 0.054 at 625, Round 4 0.041 at 950, and Round 5 0.055 at 1,600. Marc notes Round 3 was 6.4 times better than Round 2 and that Round 5's higher floor reflects a harder task with 78 behaviour classes against 59.", "Q: What timeline and open questions does Marc give for tooth regeneration?\nA: Marc reports that Toregem Biopharma, the Kyoto University spin-off Dr Takahashi co-founded, aims for market availability by 2030, helped by well-established regulatory pathways for monoclonal antibodies. He flags remaining questions about ensuring regenerated teeth emerge in the correct positions, long-term stability, and whether third dentition structures match the durability of natural permanent teeth. The post is at https://marcdeller.com/elementor-5749/.", "Q: How does WordPressID handle titles?\nA: It decodes the HTML entities WordPress bakes into titles, with a gsub chain covering `&amp;`, `&#8217;` and `&#8211;`. Commas are handled safely because the CSV uses jq's `@csv`, which quotes and escapes fields correctly, while the screen table uses `@tsv` with `column -s$'\\t'`.", "Q: What are Marc Deller's core strengths, as he describes them on LinkedIn?\nA: Marc Deller lists five core strengths on LinkedIn: strategic leadership that drives scientific insights toward drug-discovery breakthroughs; cross-functional collaboration integrating biology, chemistry and computational science; technological innovation deploying Cryo-EM, AI/ML and automation; project management crafting strategies for urgent biological needs; and team empowerment through mentorship and a culture of creativity and high performance.", "Q: Is there a US equivalent of this application?\nA: Yes. The corresponding United States publication is US 2023/0272055 A1, which is the number Marc Deller cites when describing this achievement. The PCT publication WO 2023/107994 A1 is the version that carries the full readable specification, claims and search report.", "Q: What foundational papers does the bicyclic WRN patent post cite?\nA: The post cites Chu et al. on RecQ helicases as multifunctional genome caretakers in Nature Reviews Cancer (2009), Behan et al. on prioritisation of cancer therapeutic targets using CRISPR-Cas9 screens in Nature (2019), Chan et al. showing WRN helicase is a synthetic lethal target in microsatellite unstable cancers in Nature (2019), and Kategaya et al. in iScience (2019).", "Q: What are Site Health Inspector's timing thresholds?\nA: For a five-star rating: DNS resolution at or under 50 ms, TCP connect 100 ms, TLS handshake 150 ms, TTFB 200 ms and total load 500 ms. Three-star fair thresholds are 200 ms, 400 ms, 400 ms, 600 ms and 2000 ms respectively, with anything above those scoring one star.", "Q: In plain terms, what did Marc Deller do in the first part of his thesis?\nA: In plain terms, Marc set out to take pictures of the 3D shape of two proteins involved in cell signalling, OSM and the LIF receptor. The wildtype OSM protein was too sticky and would not form the crystals needed for X-ray work, so he used computer models based on a related protein (LIF) to design shortened versions of OSM, clipping off a floppy tail he thought was getting in the way, and cloned them into bacteria as tagged fusion proteins ready to be grown and purified.", "Q: What does Marc's post on intrinsically disordered region binders cover?\nA: \"The Shape-Shifting World of Intrinsically Disordered Region Binding Proteins\", published 22 August 2025, covers a 2024\u20132025 breakthrough from the Baker Lab and collaborators on designing proteins that selectively bind intrinsically disordered regions. Marc explains that rather than anchoring onto a static structure, the designed binders are themselves flexible and conform to the many shapes an IDR presents.", "Q: What data was XenoBug trained on according to Marc?\nA: Marc reports that XenoBug was trained on 6,814 diverse substrates involved in roughly 141,200 biochemical reactions, using Random Forest classifiers and artificial neural networks with a hybrid feature set of 1,603 molecular descriptors plus linear and circular fingerprints. It achieves binary accuracies above 0.75 and F1 scores above 0.62.", "Q: What is the biology behind the TA-repeat mechanism?\nA: The post explains that in the absence of functional mismatch repair, TA-dinucleotide repeats undergo large-scale expansions and form non-B DNA secondary structures that stall replication forks. WRN normally resolves these through ATP-dependent unwinding, so when WRN is disrupted, structure-specific nucleases cleave the unresolved DNA and trigger chromosome shattering and cell death.", "Q: What is a VRC01-class antibody?\nA: VRC01-class antibodies are broadly neutralizing HIV-1 antibodies that target the conserved CD4-binding site on gp120, the receptor-binding subunit of the viral envelope spike. As described in the 2016 Immunity paper Marc Deller co-authored, they share a set of \"VRC01 signatures\": preferential use of the IgHV1-2*02 germline allele, a conserved heavy-chain maturation pattern, a stringent 5-amino-acid light-chain CDR3, and specific light-chain CDR1 mutations.", "Q: How dramatic was the pharmacokinetic improvement in dogs?\nA: In beagle dogs, changing the hinge binder to give povorcitinib reduced hepatic blood flow clearance from 82% to 4% compared with compound 18. Oral exposure rose from 65 nM\u00b7h with 1% oral bioavailability for compound 18 to 165,386 nM\u00b7h with 100% oral bioavailability for povorcitinib.", "Q: Why is XenoBug relevant to Marc's own enzyme work?\nA: Marc frames XenoBug as part of a wider wave of AI-driven environmental biotechnology that runs parallel to enzyme engineering for microplastic degradation \u2014 the same computational approach that finds pollutant-degrading enzymes in metagenomes underpins the discovery and optimisation of plastic-degrading enzymes. The post is at https://marcdeller.com/revolutionary-ai-tool-xenobug-transforms-environmental-pollution-control-through-predictive-enzyme-discovery/ and the tool itself at https://metabiosys.iiserb.ac.in/xenobug.", "Q: Why does Elora describe its therapy as \"first-in-class\"?\nA: Elora calls its therapy first-in-class because, per the deck, it would be the first systemic enzyme therapy to actively reduce microplastic burden inside the human body, and the first to address both microplastic accumulation and cardiometabolic disease at once. Existing options either treat cardiometabolic disease without touching microplastics or only clear microplastics from limited compartments like the gut or blood.", "Q: Which journal and year was the INCB159020 paper published in?\nA: The paper appeared in the Journal of Medicinal Chemistry in 2025, volume 68, pages 1924\u20131939, with the DOI 10.1021/acs.jmedchem.4c02662. It was received on 4 November 2024, revised on 16 December 2024, accepted on 30 December 2024 and published on 8 January 2025.", "Q: Where was Marc born and when?\nA: Marc was born on 21 October 1972 in Bulwell, Nottingham, England.", "Q: Are the crystallographic statistics for the KRAS structures given in the paper?\nA: The main text of the 2025 paper does not tabulate crystallographic parameters such as space group, resolution or refinement statistics. It notes that the X-ray parameters for compounds 3 and 5a are provided in the Supporting Information, and the structures themselves are available from the Protein Data Bank as entries 9E5F and 9E5D.", "Q: Tell me about Marc's post on curiosity.\nA: In \"Curiosity: 10+ years driving hands-on drug discovery through multidisciplinary approaches\", Marc Deller describes curiosity as the through-line of his 10+ year career, shaping how he chooses problems, builds teams and designs experiments. He writes that the best science starts with a simple question that refuses to let go, even when the tools or data to answer it do not yet exist.", "Q: What do the bispecific antibodies in the CALR patent do?\nA: One embodiment of WO2023107994A1 describes bispecific antibodies that simultaneously bind mutant CALR and the CD3 antigen, recruiting T cells directly to mutant CALR-expressing cancer cells to boost immune-mediated tumour destruction. Marc Deller notes this design targets both full-length and cleaved mutant CALR on cancer cell surfaces.", "Q: What is the 2018 JACS paper that Marc Deller co-authored about?\nA: The 2018 Journal of the American Chemical Society paper \"Structure\u2013Function Analysis of the Extended Conformation of a Polyketide Synthase Module,\" co-authored by Marc Deller, addressed whether the \"extended\" or the \"arch-shaped\" conformation of an assembly-line polyketide synthase (PKS) module is the catalytically relevant one. Using a human Fab antibody (1B2) that locks the module in the extended conformation, together with X-ray crystallography, SEC\u2013SAXS and enzyme kinetics, the authors showed that the extended conformation is fully competent for all of the module's essential catalytic functions.", "Q: Was compound 30 the same as the lead compound?\nA: No. Compound 30 is a close analogue of the lead compound 29, prepared specifically as a crystallography surrogate. It carries the same deuterated N-methyl nicotinamide and (S)-tetrahydrofuran motifs but on a slightly different pyrazolo[1,5-a]pyrimidine substitution pattern, and it was the molecule Marc Deller and colleagues captured bound to FGFR2 in PDB 8E1X.", "Q: What are CODSWALLOP's distinguishing features?\nA: CODSWALLOP's own differentiator table names construct diffing (aligning every deposited SEQRES against the UniProt canonical to expose tags, fusion partners, truncations and mutations), global cross-filtering across every panel from one shared selection state, crystallisation intelligence parsed from `_exptl_crystal_grow`, a family-wide coverage census of which residues nobody has ever seen, and family-wide PLIP interaction fingerprints.", "Q: What does GOBSMACKED do?\nA: GOBSMACKED is Marc Deller's app at gobsmacked.mdeller.com that folds, docks, relaxes and annotates a protein-ligand complex, then goes and finds the experimental crystal structure, superposes on the binding pocket, and reports how close the prediction landed and why. The name stands for Ground-truth Overlay for Binding Sites, Modes And Complex Kinetics/Dynamics.", "Q: Which two ventures does the About page say Marc's work runs through today?\nA: The About page says Marc Deller's work today runs through two ventures: Elora Therapeutics, which he co-founded to engineer enzymes that degrade microplastics in human tissue, and DeepCovalent, where he advises on AI-driven drug discovery.", "Q: How were the 10E8 crystals grown?\nA: Crystallisation screening used the JCSG/IAVI/Scripps high-throughput CrystalMation robot (Rigaku) at TSRI in sitting-drop vapour diffusion format \u2014 the kind of high-throughput crystallisation pipeline Marc Deller worked with at Scripps. The PG co-crystals grew against 50% PEG 200, 0.1 M HEPES pH 7.0, and the PA co-crystals against 50% PEG 400, 0.2 M NaCl, 0.1 M CHES pH 9.5; the wild-type complex required microseeding from crushed small crystals to reach diffraction size.", "Q: How much of Marc Deller's portfolio is about HIV?\nA: A large part of Marc Deller's most-cited work is HIV structural biology from his time at Scripps: the 2013 Science envelope trimer structure, the 2013 NSMB gp120 supersite paper, the 2016 Immunity VRC01-class antibody study, the 2017 PLoS Pathogens 10E8 lipid paper, the 2015 PGT128/8ANC195 Env trimer structure, and the 2012 hepatitis C E2/AP33 structure. This reflects his contributions to antibody and glycoprotein structures relevant to vaccine design.", "Q: Why does Marc consider Protein Hunter notable?\nA: Marc emphasises that Protein Hunter is fine-tuning-free and lightweight: it needs no specialised training datasets or heavy infrastructure, yet reports high AlphaFold3 in silico success rates across diverse tasks. He argues this accessibility could democratise protein design for laboratories worldwide and accelerate therapeutic development, enzyme engineering and biomaterials work.", "Q: What does Marc mean by a \"resistance-proof\" drug?\nA: Marc describes designing molecules that bind a region of the protein essential for its function and not easily mutated, or that remain effective against both the original protein and its most likely future mutant forms, rather than fitting only the current form of the target.", "Q: How does AnalyzeBoltz handle broken or missing files?\nA: It tolerates missing confidence or affinity JSONs, absent CIF files and non-standard timestamp formats, logging failures per prediction rather than aborting the run, so no result is silently dropped. Prediction directories are matched to input YAMLs by stem name, falling back to the first YAML in the log and then to the directory name. The code is at github.com/bellcheddar/analyze-boltz2-results and the post at https://marcdeller.com/analyze_boltz-py/.", "Q: What is PDB entry 5C7K?\nA: 5C7K is the crystal structure of the BG505 SOSIP gp140 HIV-1 Env trimer bound to the broadly neutralising antibodies PGT128 and 8ANC195, at 4.60 \u00c5 in space group I 2 3 with 1543 polymer residues and 183.37 kDa, released on 7 October 2015 with NAG bound. It was deposited by Kong, L., Stanfield, R.L. and Wilson, I.A. for the 2015 Acta Crystallographica Section D paper on complete epitopes for vaccine design.", "Q: What resolution milestones does Marc give in his \u00e5ngstr\u00f6m post?\nA: Marc writes that X-ray diffraction commonly refines electron density maps to better than 2 \u00c5, that 1.2 \u00c5 pinpoints hydrogen-bond geometry, and that at 1.0 \u00c5 and below charge density maps can even discriminate bonding electrons. He adds that NMR measures bond lengths and dihedral angles to about 0.01 \u00c5, and that cryo-EM's renaissance in the 2010s broke the resolution barrier to reach sub-2 \u00c5 single-particle reconstructions.", "Q: Which FGFR2 structure does Marc list on his site?\nA: Marc Deller lists PDB entry 8E1X, the FGFR2 kinase domain in complex with a pyrazolo[1,5-a]pyrimidine analogue known as Compound 29, ligand code U9P. It was solved by X-ray diffraction at 2.68 \u00c5 and released on 23 November 2022.", "Q: Is Marc looking for a job?\nA: Marc is Co-Founder and CSO of Elora Therapeutics and an advisor to DeepCovalent, and he is always open to conversations about structural biology and drug-discovery leadership. Reach him at marc@marcdeller.com.", "Q: Where is the second KRAS G12D structure post?\nA: Marc Deller's post is titled \"Discovery of an Orally Bioavailable KRAS G12D Inhibitor_2\" and appears in the structures section of his website. You can read it at https://marcdeller.com/discovery-of-an-orally-bioavailable-kras-g12d-inhibitor_2/.", "Q: Who does Marc say is involved in producing a protein structure?\nA: Marc Deller writes that every protein structure is a collaboration between CRO resources, collaborators, protein scientists and synchrotron beamline scientists. He adds that day-to-day collaboration lives at the interfaces between biology, pharmacology, structural biology, chemistry and computation.", "Q: Where can I read Marc's oncostatin M post?\nA: Marc Deller's post is simply titled \"Crystal Structure of Human Oncostatin M\" and sits in the structures category of his blog. You can read it at https://marcdeller.com/crystal-structure-of-human-oncostatin-m/.", "Q: What resolution were the structures in Marc Deller's TM1088 paper determined to?\nA: Marc Deller and colleagues determined the octameric TM1088 assembly to 3.45 \u00c5 resolution, the full-length TM1088A protein to 1.50 \u00c5, and the C-terminal domain of TM1088B to 1.55 \u00c5. The 3.45 \u00c5 resolution of the full assembly is comparable to other potassium gating ring structures, which had been reported between about 2.7 \u00c5 and 3.8 \u00c5.", "Q: What supporting data does Marc cite for organ-on-a-chip models?\nA: Marc cites Ewart and colleagues in Communications Medicine (2022), whose analysis of 870 human Liver-Chips showed 87% sensitivity and 100% specificity in predicting drug-induced liver injury across a blinded compound set, substantially outperforming conventional preclinical models. He also cites Park and colleagues on iPSC-derived organoids, the Avila 2020 FDA/CDER perspective, and Yao and colleagues (2025) documenting NAMs already used in approved programs for cystic fibrosis and Fabry disease therapeutics. The post is at https://marcdeller.com/from-bench-to-regulation-how-the-fda-is-formalising-a-future-beyond-animal-testing/.", "Q: Summarise Marc's post on nanobodies.\nA: \"Nanobodies: Revolutionizing Molecular Medicine Through Nature's Smallest Antibodies\", published 27 August 2025, covers single-domain antibodies derived from the heavy-chain-only antibodies found exclusively in camelids. At roughly 15 kilodaltons they achieve complete antigen recognition through a single variable heavy-chain domain, with thermal stability up to 90\u00b0C and reversible refolding.", "Q: What software does Marc Deller use for protein structure work?\nA: For protein structure determination and analysis Marc Deller uses HKL2000/3000, CCP4, PHENIX, COOT and PyMOL for data processing, molecular replacement, model building, ligand fitting and validation; RELION and CryoSPARC for cryo-EM; Schr\u00f6dinger, CCG/MOE and GROMACS for modelling and computational chemistry; and AlphaFold, RFdiffusion and ProteinMPNN for AI-driven structure prediction and design. He also uses ChemCart, PSILO, LIMS and ELN systems.", "Q: What surprising size effect did the zebrafish study report in Marc's review?\nA: Marc notes that in Lu et al. (2016, Environmental Science & Technology), 5 \u03bcm polystyrene particles reached the liver while 20 \u03bcm particles stayed in gills and gut, steady state was reached within 48 hours, and \u2014 counter-intuitively \u2014 the larger 5 \u03bcm particles induced greater oxidative stress than 70 nm particles, challenging conventional size-toxicity assumptions. The post is at https://marcdeller.com/microplastics-and-health-alarming-new-research-reveals-widespread-biological-impact/.", "Q: Which crystal structures did Marc Deller solve for the KRAS program?\nA: Marc Deller and Leslie Epling determined two KRAS G12D co-crystal structures for the INCB159020 program: the early tricyclic imidazole hit compound 3 bound to KRAS G12D (PDB 9E5F), and the more advanced cyanoethyl-substituted compound 5a bound to KRAS G12D (PDB 9E5D). The paper's Figure 3 shows both structures with the mutation site Asp12 highlighted.", "Q: Why do microplastics in the body matter for health?\nA: According to the patent, microplastics and nanoplastics\u2014particles ranging from 1 nm to 5 mm\u2014are pervasive contaminants now found in human tissues, organs, and bodily fluids, originating from synthetic polymers like PET, PE, PP, and PS. They pose significant health risks including chronic inflammation, endocrine disruption, oxidative stress, and potential carcinogenic effects, and no current therapy actively degrades and clears them from inside the body.", "Q: What does Marc Deller's LinkedIn headline say?\nA: Marc Deller's LinkedIn headline reads \"Biotech Co-Founder & Structural Biology Expert | 400+ Protein Structures | AI-Driven Drug Discovery | Transforming Drug Discovery Through Innovation.\"", "Q: How many microplastic particles do humans ingest per year, according to Elora's figures?\nA: Marc's Elora post gives an estimate of 74,000 to 121,000 microplastic particles ingested per year, with the fragments appearing in blood, brain tissue and the placenta, where they drive chronic inflammation, metabolic dysfunction and endocrine disruption.", "Q: What does Elora Therapeutics do, according to its website?\nA: Elora Therapeutics' website describes the company as developing engineered enzyme therapeutics to remove microplastics from humans, microplastics that are implicated in cancer progression and metastasis. Marc Deller is a co-founder of the company and its Chief Scientific Officer.", "Q: How many IND filings has Marc supported?\nA: Marc Deller's compassion post cites 6+ FDA IND filings supported, linking that work to his role driving innovation in precision oncology through structure-based drug design. He says supporting those filings reinforced that his work has direct human consequences.", "Q: What detector and settings were used for the MAD experiment at the ESRF?\nA: On ESRF beam line BL19 Marc Deller recorded the MAD data as 1\u00b0 oscillations on a MAR 345 image plate detector in 18 cm mode, with typical exposure times of 40 seconds and a crystal-to-detector distance of 225 mm, in a nitrogen stream at 100 K. The high-resolution native data set was collected on the same beamline at 0.98 \u00c5 to 2.2 \u00c5 using the MAR 345 in 30 cm mode with 60-second exposures.", "Q: What role did the glutamine synthetase/MSX system play in the CHO expression?\nA: The pEE14 CHO vector carried a glutamine synthetase (GS) minigene for amplification and clonal selection. GS makes glutamine from glutamate, and its inhibitor L-methionine sulphoximine (MSX) forces cells grown without glutamine to increase GS gene copy number, co-amplifying linked genes. Marc found MSX was essential for maintaining clonal selection over many passages, but it did not increase the level of LIFR expression.", "Q: What happens to the breakdown products of PET according to Marc's posts?\nA: Marc writes that PETase breaks PET down into terephthalic acid (TPA) and mono(2-hydroxyethyl) terephthalate (MHET), and that these monomer byproducts are quickly excreted with minimal toxicity, which is central to the safety case for enzyme therapy.", "Q: Which work does Marc cite on the newly discovered NOS linkages in the PDB?\nA: Marc links to the Nature Communications Chemistry paper on revealing arginine-cysteine and glycine-cysteine NOS linkages by systematic re-evaluation of the PDB, alongside coverage in Phys.org and Chemistry World describing an overlooked bonding motif appearing in many more proteins than previously realized.", "Q: Which entries were deposited by Adriana Irimia?\nA: Adriana Irimia, with Ian Wilson, deposited the seven 10E8 Fab and MPER scaffold structures released on 8 March 2017: 5SY8, 5T29, 5T5B, 5T6L, 5T80, 5T85 and 5TFW. They support the PLoS Pathogens paper on the angle of approach of 10E8 to the HIV-1 viral membrane.", "Q: What are the engineering implications of the PKS work?\nA: Marc Deller writes that biasing modules toward the extended conformation, by designing linkers and interface surfaces, could streamline non-natural polyketide assembly and improve yields, while antibody or small-molecule probes that lock modules in extended states could allow real-time dissection of stepwise catalysis.", "Q: How were the TM1088 proteins expressed and purified?\nA: The TM1088 locus was PCR-amplified from T. maritima MSB8 genomic DNA and cloned into plasmid pMH4, which adds an N-terminal MGSDKIHHHHHH tag to TM1088A; the untagged TM1088B co-purified with tagged TM1088A. Expression was in E. coli GeneHogs in Terrific Broth, followed by nickel-chelating affinity chromatography, RESOURCE Q anion exchange and HiLoad Superdex 200 size exclusion. The TM1088B C-terminal domain (residues 140\u2013218) was cloned separately into pSpeedET using the PIPE cloning method and purified by Ni-NTA chromatography.", "Q: How do I get from mdeller.com back to Marc's main website?\nA: The mdeller.com landing page links to marcdeller.com, \"Hidden Architecture: Mapping Biology's Structure\", as the main site, and its links block points at the blog, structural-biology, publications, structures, projects, patents, vibe-coding and about-me sections of marcdeller.com.", "Q: What is the pdb_tidy CONECT bug Marc documented?\nA: Naively converting with gemmi or pdb_tidy alone leaves a one-serial gap at each chain/TER boundary. Because OpenBabel maps CONECT records positionally by serial number, any gap causes every record past that point to be silently discarded, corrupting ligand perception and producing wrong SMILES, fictitious bonds in the PyMOL session and invented PLIP interactions.", "Q: How did the TM0077 structure change the enzyme's classification?\nA: The structural and functional work reclassified TM0077's activity from acetyl xylan esterase to a short-chain acetyl esterase. Marc Deller's post points to the relevance of this for industrial biocatalysis at extreme temperatures, since the enzyme comes from the hyperthermophile Thermotoga maritima.", "Q: Has Marc Deller worked on membrane proteins?\nA: Yes. At the Joint Center for Structural Genomics, Marc Deller determined novel structures of ion channels and membrane proteins, and membrane proteins are listed among his skills. He also has expertise in cell surface receptor targeting.", "Q: What headline numbers does the PANTS site report?\nA: The PANTS front page reports 14,804,920 sequences processed, 439 candidates recovered, 416 predicted structures, 342 experimentally evidenced positives, 25 known enzymes catalogued, and 2,759 activity measurements, all cited. Its funnel page adds that 403 candidates had the catalytic triad confirmed in 3D and that 24 were too long to fold at over 450 residues, deferred rather than discarded.", "Q: What certifications or courses has Marc Deller completed recently?\nA: In 2024\u20132025 Marc Deller completed Google's Foundations of Project Management, IBM's Introduction to Artificial Intelligence, AWS Serverless Architectures, Google's \"Foundations: Data, Data, Everywhere\", Coursera's Introduction to R and Dashboard Development with Shiny (GenAI for Retail Analysis), and LinkedIn Learning's Amplify Your Communication Skills with Generative AI.", "Q: What did Marc Deller do at Scripps on HIV?\nA: At The Scripps Research Institute, Marc Deller worked in the Joint Center for Structural Genomics alongside Ian Wilson's structural biology group, where he ran high-throughput crystallisation screening, crystal harvesting and X-ray data collection for a series of HIV-1 envelope glycoprotein projects. That work contributed to the 2013 Science crystal structure of the cleaved BG505 SOSIP.664 Env trimer and to the 2013 Nature Structural & Molecular Biology paper defining the N332 glycan supersite of vulnerability on gp120.", "Q: Who did Marc Deller write his crystallography methods papers with?\nA: Marc Deller's crystallography methods and validation papers were largely co-written with Bernhard Rupp, and in some cases with Leopold Kong and Edwin Pozharski. Together they produced work on protein crystal harvesting, protein stability, crystallization of macromolecular complexes, and protein\u2013ligand model validation.", "Q: What is Marc Deller's ORCID?\nA: Marc Deller's ORCID, given on the Publications page of marcdeller.com, is 0000-0001-8070-6502. The page also links his OpenAlex and Scopus author profiles.", "Q: Tell me about Marc's advisory work in computer-aided drug design.\nA: As Scientific Technical Advisor to DeepCovalent since October 2025, Marc Deller advises on physics-based computer-aided drug design pipelines, small-molecule GPCR-targeting discovery platforms, in silico screening and cheminformatics workflows, and protein-science CRO oversight and experimental validation. That work sits alongside his role as co-founder and CSO of Elora Therapeutics.", "Q: What evidence does Marc give that models understand solutions but truncate output?\nA: Marc reports that models such as Claude explicitly acknowledge approaching output limits, saying things like \"The pattern continues, but to avoid making this too long, I'll stop here\" while solving Tower of Hanoi. He adds that when researchers asked for generating functions instead of exhaustive move lists, models solved 15-disk cases that had supposedly been beyond them at 8 disks, and that for a 10,000-move problem even 99.9% per-token accuracy yields under a 0.005% chance of perfect execution.", "Q: Why does Marc prefer enzymes over filtration for long-term microplastic management?\nA: Across his August 2025 posts Marc argues that filtration only sequesters or removes particles, whereas engineered enzymes such as PETase achieve true biodegradation, converting the polymer into monomers the body can excrete. He therefore sees biocatalytic agents as potentially eclipsing filtration for long-term management.", "Q: How was the initial OSM model built and refined?\nA: Marc Deller started from a homology model of \u0394188_OSM (residues 1\u2013188) built in Chapter 2, and carried out 40 cycles of maximum-likelihood rigid body refinement in CNS-SOLVE with each helix as a separate body, using data from 20.0\u20132.6 \u00c5 and a low X-ray weight (Wa = 2.4). This was followed by 200 cycles of positional refinement and 1000 cycles of torsion-dynamics simulated annealing with slow cooling from 4000 K, then rounds of manual rebuilding in O against solvent-flattened and 2Fo\u2212Fc maps, with composite omit maps for ambiguous regions.", "Q: Where did the ACA 2014 harvesting bottleneck poster point readers for more information?\nA: The poster gave the contact e-mails mdeller@scripps.edu and hofkristallamt@gmail.com and the web page http://www.ruppweb.org/robot/ for the UMR project.", "Q: How do I contact Marc?\nA: Email marc@marcdeller.com, call (302) 358-6093, or connect on LinkedIn at linkedin.com/in/marccdeller. His website is marcdeller.com.", "Q: How was correct ligation of the OSM constructs confirmed?\nA: Correct ligation of the PCR-amplified insert into pGEX-3C-GST was screened by double digestion with EcoRV and BglII, exploiting a unique EcoRV site in the vector and a unique BglII site in the insert, which gave two diagnostic bands at 3685 and 1815 base pairs. Identity was then confirmed by automated PCR sequencing with three primers (5', internal and 3') and alignment against the published OSM sequence (GenBank M27286).", "Q: What is in situ data collection and what are its drawbacks?\nA: In situ methods grow crystals in a plate that is mounted directly on an X-ray source, so the crystal is never removed from its mother liquor and manipulation damage is avoided. Marc Deller's review lists the drawbacks as scattering and absorption by the plate and liquor, a restricted rotation angle limiting usable data per crystal, and confinement to room-temperature collection since crystals cannot be cryocooled in the plate.", "Q: Why does PANTS look for PETases in metagenomes?\nA: PET is the plastic in drinks bottles and polyester clothing, and fragments of it now turn up in human blood and tissue, but almost every known PET-degrading enzyme was found or engineered for a hot, alkaline industrial recycling reactor. PANTS therefore searches metagenomes \u2014 the DNA of whole microbial communities read straight from the environment \u2014 in four settings with decades of plastic exposure: the human gut, compost, the ocean surface where debris collects, and landfill.", "Q: How does DeepCovalent's OS3D engine work?\nA: OS3D delivers single-step, seed-free de novo small-molecule generation, conditioning generation simultaneously on target biology, efficacy signals, ADMET profiles and medicinal chemistry constraints. By unifying ideation, design and in-silico validation into a single forward pass, it produces synthesis-ready candidates with early translational relevance rather than incremental analogues of known chemistry.", "Q: What example of an implausible ligand does the chapter use?\nA: The chapter uses PDB entry 2a6i, chain P, a peptide antigen modelled bound to a Fab antibody fragment. In an omit difference map contoured at 2.5\u03c3 only discontinuous fragments of positive density appear, partly explicable as ordered solvent, and the model also has many unreasonably close contacts with the Fab and backbone torsion angles in unfavourable regions of the Ramachandran plot.", "Q: What is PDB entry 3KK7?\nA: 3KK7 is a putative cell invasion protein with a MAC/perforin domain, NP_812351.1, from Bacteroides thetaiotaomicron VPI-5482, solved at 2.46 \u00c5 in space group P 21 21 21 with 1082 polymer residues and 124.35 kDa, released on 24 November 2009 and keyworded CELL INVASION. It is cited in the 2010 Acta Crystallographica Section F paper on a membrane-attack complex/perforin family protein from the human gut symbiont.", "Q: Why does it matter if a ligand in the PDB is wrong?\nA: The chapter opens by noting that accurate protein\u2013ligand models underpin computational methods including ligand docking and structure-based drug design, so an unsupported ligand can propagate errors into drug discovery projects that rely on it. Marc Deller and his co-authors report that strict evidence-based validation has already revealed cases where deposited ligands are insufficiently supported by the electron density.", "Q: How does oncostatin M compare structurally to LIF and other cytokines?\nA: Marc Deller's analysis showed that human oncostatin M is most similar to LIF within the gp130 family, superimposing with an rmsd of 2.1 \u00c5 over 145 equivalent C\u03b1 atoms, and more distantly resembles erythropoietin (rmsd 2.8 \u00c5 over 120 C\u03b1 atoms) from the growth hormone family. hOSM was the fourth gp130 cytokine to be analysed structurally, after IL-6, CNTF and LIF, and both LIF and hOSM share a disulphide anchoring an \u03b1-helical part of the AB loop to the top of helix D.", "Q: What did this Env trimer structure mean for HIV vaccine design?\nA: The authors concluded that the structure defined the overall architecture of the soluble trimer and the secondary, tertiary and quaternary interactions between gp120 and gp41 involved in its assembly, and that the complete definition of how neutralisation epitopes are presented in the context of the trimer should help the design of new immunogens as candidate vaccines. The paper's own summary line describes the structure as \"a blueprint for vaccine design\".", "Q: Why was finding a PH domain in a bacterium surprising?\nA: Pleckstrin homology (PH) domains had previously been associated almost exclusively with eukaryotic signalling proteins, so seeing that architecture in a protein from the permafrost-dwelling bacterium Exiguobacterium sibiricum 255-15 was unexpected. Marc Deller's post explains that the 2.42 \u00c5 structure expands understanding of bacterial protein fold diversity and possible signalling mechanisms in extremophilic organisms.", "Q: Did Marc study at the University of Nottingham?\nA: No. Marc grew up in Nottingham and took his A-levels at High Pavement College there, but he never studied at the University of Nottingham. His BSc in Biochemistry and Molecular Biology is from the **University of Leeds** (1991-1995), and his D.Phil. in molecular biophysics is from the **University of Oxford** (1995-1999).", "Q: Marc is first author on that 2015 validation paper, right?\nA: Yes. The post cites \"Models of protein\u2013ligand crystal structures: trust, but verify\" as M. C. Deller, Journal of Computer-Aided Molecular Design, 2015, and Marc Deller presents it among his own publications on validation and structure quality.", "Q: What does claim 1 of the patent cover?\nA: Claim 1 covers an antibody that binds human mutant calreticulin defined by a degenerate consensus of all six complementarity-determining regions, with each variable position specified as a short list of permitted amino acids (SEQ ID NOs:97-99 plus the X1 to X72 definitions). It is a broad sequence-based claim spanning the whole antibody family disclosed in the application.", "Q: Did the SSO2064 structure change the Pfam database?\nA: Yes. The structure led to a re-evaluation of the DUF35 family, which was split into two separate entries in Pfam release 24.0. Marc Deller was one of the Joint Center for Structural Genomics co-authors on the paper, whose first author was Krishna SS.", "Q: Give some example Site Doctor commands.\nA: `python3 site_doctor.py` prompts for a URL and runs everything; `python3 site_doctor.py https://example.com --scan` runs only the diagnostic scan; `python3 site_doctor.py https://example.com --analytics --max-pages 25 --no-psi` analyses 25 pages without PageSpeed Insights; and `--out reports/example.html` writes the report to a chosen path.", "Q: How does the gameplay of MicroPlastic Blaster teach the science?\nA: Each enemy maps to a real polymer class with scientifically motivated attributes: PET is the slowest but toughest with 3 HP and 30 points, reflecting the difficulty of degrading crystalline PET, while polypropylene is fast with 1 HP, and polyethylene and polystyrene sit in between. Power-ups mirror the pipeline too \u2014 FAST-PETase gives rapid-fire shots and VC Coins slow disease progression to show that investment translates into therapeutic impact.", "Q: Which metrics does ForgeFold surface?\nA: `confidence_score` as an overall model-confidence measure, `affinity_pred_value1` as an affinity-related score, `affinity_probability_binary1` as a binder-likelihood output, and a derived `pIC50` potency-style estimate. Marc frames these as rapid triage and comparison rather than a substitute for experimental validation.", "Q: What is Marc's favourite guitar?\nA: A 1962 Fender Jazzmaster.", "Q: What is the $8M platform Marc Deller built?\nA: As Associate Director at Incyte, Marc Deller established an $8M automated, state-of-the-art structural biology platform from the ground up. It featured X-ray crystallography and cryo-EM and was designed to enable rapid hit-to-lead progression for Incyte's oncology and immunology programs.", "Q: How does Site Health Inspector grade a site overall?\nA: A weighted 0-100 score mapped to four grades: 80-100 EXCELLENT, 60-79 GOOD, 40-59 FAIR and 0-39 POOR. Key contributors are TTFB, total load time, compression, HTTP/2 or better, valid SSL, certificate validity, HSTS presence, cache headers, Cloudflare HIT status and live availability.", "Q: How many patents does Marc Deller have?\nA: Marc Deller's September 2026 r\u00e9sum\u00e9 credits him with 7+ patents. Named examples include the mutant-CALR antibody patent (\"Anti-mutant Calreticulin (CALR) Antibodies and uses thereof\", WO2023107994A1/US2023/0272055A1), the WRN helicase inhibitor patent (published as US2025/0215006A1, \"Tricyclic Compounds as Inhibitors of WRN\"), and a 2026 patent on \"GLP-1 Receptor Modulators and uses thereof\".", "Q: Why do MSI-H tumours need WRN at all?\nA: Marc Deller writes that mismatch repair-deficient cells accumulate massive expansions of TA dinucleotide repeats which form non-B DNA cruciform structures that stall replication forks, and that WRN is the only helicase capable of resolving these structures to prevent catastrophic double-strand breaks.", "Q: What is the 2017 PLoS Pathogens paper Marc Deller co-authored about?\nA: The 2017 PLoS Pathogens paper \"Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design\" showed that the epitope of the broadly neutralizing HIV antibody 10E8 is made up of both protein and lipid. X-ray structures revealed a specific lipid head-group binding site on 10E8, and mutagenesis confirmed that the antibody's light chain must engage the viral membrane to neutralise HIV-1.", "Q: What is CODSWALLOP?\nA: CODSWALLOP, at codswallop.mdeller.com, gathers every PDB entry for a protein family in one place: what was actually made, what it bound, what crystallised, and which residues nobody has ever seen. Marc describes it as doing an afternoon's work of manual PDB trawling in ninety seconds, with every entry, construct and ligand cross-referenced and interactive.", "Q: How much money is Elora raising?\nA: According to the deck, Elora is raising its first $275K to deliver in vitro validation, in vivo proof-of-concept, and IP expansion. This pre-seed round is the first tranche of a staged financing that later includes roughly $5.55M for pre-clinical work (seed) and $27.4M for Phase 1 clinical execution (Series A).", "Q: Which three FmdE-family proteins were solved?\nA: The structures represented two bacterial and one archaeal species: B8FYU2_DESHY from the anaerobic dehalogenating bacterium Desulfitobacterium hafniense DCB-2, Q2LQ23_SYNAS from the syntrophic bacterium Syntrophus aciditrophicus SB, and Q9HJ63_THEAC from the thermoacidophilic archaeon Thermoplasma acidophilum.", "Q: What does \"Embracing AI-Enabled Communication and Adaptability\" cover?\nA: In \"2025 \u2013 Embracing AI-Enabled Communication and Adaptability\", Marc Deller groups three credentials: Amplify Your Communication Skills with Generative AI, Building an Adaptability Mindset in the Age of AI, and IBM's Introduction to Artificial Intelligence. He argues that professionals who master AI-driven communication tools gain a decisive edge and can bridge interdisciplinary teams.", "Q: How does 3JXO relate to the TrkA gating ring structure 3L4B?\nA: The 1.55 \u00c5 TrkA-N domain structure provides atomic-level detail of a single subunit, while PDB entry 3L4B shows the full octameric assembly at 3.45 \u00c5. Marc Deller's post explains that together they establish the structural framework for understanding potassium transport regulation in thermophilic bacteria.", "Q: Explain the WRN patent in plain English.\nA: Some tumours, particularly bowel, womb and stomach cancers, have a broken DNA proofreading system, which leaves them leaning heavily on a repair enzyme called WRN just to stay alive. The Incyte patent that Marc Deller co-invented describes a family of drug-like molecules that latch permanently onto WRN and switch it off, which should kill those tumour cells while leaving healthy cells alone \u2014 and it includes the X-ray picture showing exactly where and how the drug grabs hold of the enzyme.", "Q: How is Marc Deller connected to both DeepCovalent and Elora Therapeutics?\nA: Marc Deller holds two current company roles: he is Co-Founder and Chief Scientific Officer of Elora Therapeutics, and Scientific Technical Advisor to DeepCovalent Inc. The two are separate companies, Elora working on engineered enzyme therapeutics for microplastics and DeepCovalent on generative AI small-molecule design.", "Q: What is Marc Deller's 2025 Journal of Medicinal Chemistry paper about?\nA: The 2025 J. Med. Chem. paper Marc Deller co-authored, \"Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor,\" describes how an Incyte team went from a weakly potent patent hit to compound 23, INCB159020, a potent and selective reversible inhibitor of the KRAS G12D oncoprotein that achieves oral exposure in nonhuman primates. It was published in J. Med. Chem. 2025, volume 68, pages 1924\u20131939, with Alexander Sokolsky as corresponding author.", "Q: Why does Marc say linker choice matters so much in TELSAM fusions?\nA: Marc explains that the connector between the crystallization chaperone and the target protein directly controls crystal morphology, diffraction quality and formation kinetics. Short semi-flexible Pro-Ala linkers and rigid linkers work best for DARPin targets, flexible polyglycine connections suit other domains better, and removing the traditional 10xHis purification tag further improves crystallization rate and crystal quality.", "Q: What is an assembly-line polyketide synthase?\nA: Assembly-line polyketide synthases (PKSs), such as 6-deoxyerythronolide B synthase (DEBS), are multifunctional, multimodular enzymes that synthesize structurally diverse natural products. Each module contains catalytic domains, including a ketosynthase (KS), an acyltransferase (AT), an acyl carrier protein (ACP) and often a ketoreductase (KR), that together extend and modify a growing polyketide chain before passing it to the next module, as described in the 2018 JACS paper co-authored by Marc Deller.", "Q: How does the review relate to Marc's own research?\nA: The review was written while Marc Deller was a D.Phil. student in Yvonne Jones's CRC Receptor Structure Group in Oxford, where he determined the crystal structure of the four-helix bundle cytokine oncostatin M. The review's coverage of hematopoietic receptor complexes, the importance of receptor orientation for signalling, and conserved hydrophobic cores with polar specificity in the TNF family reflects the same themes that shaped his oncostatin M structure\u2013function work published in Structure the same year.", "Q: What unexpected benefit of crude lysate does Marc mention for cryo-EM grids?\nA: Marc notes that the presence of ribosomes and other lysate components unexpectedly improved ice quality during vitrification, which is normally one of the trickiest variables in cryo-EM sample preparation. He also flags that the PURE system is compatible with genetic engineering to add post-translational modifications or folding-assisting factors.", "Q: Tell me about Marc's post on elastic networks and AlphaFold2 accuracy.\nA: In \"Revolutionary Bridge Between Prediction and Reality: How Elastic Networks Transform AlphaFold2's Structural Accuracy\", published on 14 October 2025, Marc Deller discusses work by Alshammari and colleagues in Bioinformatics Advances on flexibly fitting AlphaFold2 models into cryo-EM density maps using elastic network models. He frames it as an example of AI predictions being improved by combining them with biophysical modelling and experimental data rather than being used alone.", "Q: What is INCB159020?\nA: INCB159020 is an orally bioavailable KRAS G12D inhibitor discovered at Incyte and reported in the Journal of Medicinal Chemistry in 2025. It was designed to balance potency with good ADME properties and to address the previously \"undruggable\" KRAS G12D mutation, broadening options beyond KRAS G12C-targeted therapies. Marc Deller supported the program's IND filing through his structural biology work.", "Q: Why is selective JAK1 inhibition considered better than dual JAK1/JAK2 inhibition?\nA: As the paper co-authored by Marc Deller explains, aberrant JAK2 signalling drives cancers such as myeloproliferative neoplasms, whereas JAK1 is the central mediator of autoimmune and inflammatory disease. Because JAK2 supports normal blood cell production, inhibiting it can cause anemia, thrombocytopenia and neutropenia, so sparing JAK2 while blocking JAK1 offers a potentially safer therapeutic strategy for inflammatory disorders.", "Q: Did Marc work on the povorcitinib paper himself?\nA: Yes. In the post he refers to it as an Incyte paper of his, and the work rests on co-crystal structures of the inhibitor bound to the isolated JAK1 and JAK2 kinase domains, deposited as PDB entries 10PI and 10PJ, the kind of structural contribution he provided at Incyte.", "Q: Where does the name \"Janus kinase\" come from?\nA: The paper's introduction recounts that in the mid-1980s the Wilks group, searching for colony-stimulating factor receptors, found two unusual kinases with two kinase domains and no transmembrane region and jokingly called them \"Just Another Kinase 1 and 2\". Before publication Wilks renamed them Janus kinases after the two-faced Roman god, a nod to the two kinase domains, an active JH1 and a pseudokinase JH2, present in each enzyme.", "Q: Does Marc Deller have a website or blog?\nA: Yes, Marc Deller's website and blog is at https://marcdeller.com. His full publication list is on Google Scholar (https://bit.ly/marcdeller_pubs) and his protein structures are listed at https://bit.ly/marcdeller_proteins.", "Q: Is it unusual that CDR L2 is not used?\nA: No. The 2012 paper notes that the lack of CDR L2 involvement in antigen binding for both AP33 and HCV1 is not unusual for small-molecule or linear peptide interactions, citing earlier crystallographic work on antibody\u2013peptide complexes from the Stanfield and Wilson laboratories.", "Q: Where was the soluble cleaved Env trimer structure published?\nA: It was published as Julien, J.P., et al. (2013), \"Crystal Structure of a Soluble Cleaved HIV-1 Envelope Trimer\", in Science, volume 342, pages 1477-1483. Marc Deller has a companion blog post about that paper on marcdeller.com.", "Q: What is the most common cryoprotectant in the JCSG Top96 Cryo screen?\nA: Ethylene glycol, typically at 10% or 15%, is by far the most common cryoprotectant recommended in the JCSG Top96 Cryo screen, particularly for the PEG 3350/salt conditions in rows E to G. Glycerol at 15\u201325% is used mainly for the high-salt ammonium sulfate and phosphate conditions, and PEG 200 at 10\u201325% for several PEG-only conditions.", "Q: What was Marc Deller's specific contribution to the gp120 supersite paper?\nA: The author contributions statement of the 2013 NSMB paper states explicitly that robotic crystallization screening at the Joint Center for Structural Genomics was carried out by M.C.D., T.C. and H.T. \u2014 that is, by Marc C. Deller, Thomas Clayton and Henry Tien. Marc Deller was a middle co-author, listed seventeenth of twenty-seven authors, and his affiliation on the paper is the Joint Center for Structural Genomics.", "Q: What is the Universal Micromanipulation Robot (UMR)?\nA: The Universal Micromanipulation Robot, or UMR, is a robotic platform designed to perform the full range of tasks necessary for protein crystal harvesting. It was developed in collaboration with Square One Systems Design and tested at the JCSG, where Marc Deller worked on optimizing it for microcrystal harvesting and high-speed cryocooling (hyperquenching).", "Q: How does Marc Deller explain targeted protein degradation to a general reader?\nA: In \"Hitting the Delete Button on Disease: The Rise of Targeted Protein Degraders\", published 13 August 2025, Marc Deller describes TPDs as bifunctional molecules acting like a microscopic matchmaker: one end binds the target protein and the other recruits a ubiquitin ligase, flagging the target for destruction by the cell's own ubiquitin-proteasome system.", "Q: What does cif2plip output?\nA: A standards-compliant `<stem>.pdb`, PLIP's XML and text reports, ray-traced binding-site PNGs, a PyMOL `.pse` session, an `<stem>_interactions.csv` with one row per contact, and an `<stem>_ligand_summary.csv` giving SMILES, InChIKey and interaction counts per ligand. Reported interaction classes are hydrophobic contacts, hydrogen bonds, water bridges, salt bridges, \u03c0-stacking, \u03c0-cation, halogen bonds and metal coordination. The repo is bellcheddar/cif2plip.", "Q: How did the crystal structure explain the loss of potency against gatekeeper mutants?\nA: In the structure Marc Deller and colleagues determined, the pyridine nitrogen of the inhibitor is oriented toward the gatekeeper residue. Because the FGFR3 V555L and V555M mutants replace valine with a larger, more lipophilic side chain, those side chains clash with the polar pyridyl nitrogen, which accounts for the 5-fold and 10-fold potency losses observed for compound 29 against V555L and V555M respectively.", "Q: What crystal structures did Marc Deller determine for this patent?\nA: Three structures are reported. The apo Fab fragment of anti-mutCALR antibody clone 55 was solved at 2.9 \u00c5; the same Fab in complex with a 21-mer mutant CALR peptide was solved at 3.2 \u00c5; and a Fab fragment bound to a 31-mer Type 1 mutant CalR peptide was solved at 2.0 \u00c5. All three are described in Examples 11 and 12 of WO 2023/107994.", "Q: How big is the TopPDBLX dataset?\nA: TopPDBLX parses 199,185 crystallisation records from 198,691 PDB entries, of which 188,039 (94.4 %) are usable, yielding 645,656 typed components of which 87.3 % are identified as a canonical reagent. 184,229 records carry a linked construct sequence across 23,159 distinct 30 % identity clusters, and the reagent lexicon holds 500 reagents under 1,587 names.", "Q: Does other evidence make up for missing electron density?\nA: The chapter acknowledges that biochemical data, database annotations, functional assignments and protein fold may all support the idea that a ligand binds, but insists that the placement and conformation \u2014 the pose \u2014 of any ligand in a structure model must still be supported by the electron density.", "Q: Why does Marc consider HGMP01's temperature profile important?\nA: Marc explains that many environmental PET-degrading enzymes need temperatures near the polymer's glass transition point of 65 to 80\u00b0C, whereas HGMP01 stays active at human body temperature with an optimum around 40\u00b0C. Its alkaline tolerance matches pH variation across intestinal segments, suggesting evolutionary adaptation to the gut.", "Q: Do WRN inhibitors work in treatment-resistant tumours?\nA: According to Marc Deller's post, a particularly compelling aspect of these WRN inhibitors is their potential to overcome treatment resistance, with preclinical evidence suggesting that cancers with established resistance to chemotherapy, targeted treatments or immunotherapy remain susceptible to WRN inhibition.", "Q: What cellular assays were used in the FGFR paper?\nA: Cellular potency was measured in a homogeneous time-resolved fluorescence assay using TEL-FGFR3-BA/F3 cells, which reports compound inhibition of phospho-FGFR3. Later in the program the team also developed a human whole blood assay using FGFR3 RT-112/84 cells, in which lead compound 29 gave an IC50 of 177 nM.", "Q: What is ALPHABETTI built on?\nA: The ALPHABETTI colophon lists Flask and Gunicorn behind nginx, three.js with no build step, and ESMFold and ESM-2 on a ZeroGPU Space. The 3D letterforms are Baloo 2 Bold, converted to outlines and extruded, and the app was written by Marc C. Deller, D.Phil.", "Q: Which of Marc Deller's structures are of KRAS?\nA: The two KRAS G12D entries are 9E5D at 1.36 \u00c5 and 9E5F at 1.35 \u00c5, both released on 22 January 2025 in space group P 21 21 21 with GDP and magnesium bound. Both were deposited by Marc Deller with L.B. Epling for the INCB159020 orally bioavailable KRAS G12D inhibitor programme at Incyte.", "Q: What is Marc's conclusion about data versus architecture in structure-based AI?\nA: Marc endorses the consortium's position that the next generation of structure-based machine learning needs better experimental data, not just new architectures, citing the Runs N' Poses re-evaluation showing cofolding accuracy collapses as binding-site similarity to training data decreases and PoseBusters showing many AI poses fail basic chemistry checks. He concludes the bottleneck on AI-driven drug discovery may sit in the wet lab as much as in the architecture; the CC0-licensed release is described at https://marcdeller.com/openbind-publishes-a-dense-structure-affinity-dataset-for-structure-based-ai/.", "Q: Which affinity tag does the review single out for improving solubility?\nA: Maltose-binding protein is highlighted as having a strong solubilizing effect and as promoting correct folding, effectively acting as a chaperone. Marc Deller's review also discusses GST and Halo tags, and co-expression of DnaK and GroEL, and notes that large tags are usually removed before crystallization because of linker flexibility.", "Q: What is the three-site model of receptor binding for gp130-family cytokines?\nA: As set out in Marc Deller's thesis, mutagenesis of gp130-family cytokines reveals several discrete receptor-binding sites. Sites 1 and 2 lie in broadly the same positions as the two receptor-binding sites of growth hormone, while a third site, site 3, mediates an additional interaction not seen in the GH\u2013GHR system. For OSM, LIF and CNTF, site 3 at the N-terminus of helix D binds LIFR, site 2 on helices A and C binds gp130, and site 1 on helix D and the A-B loop binds a specificity-conferring receptor.", "Q: What compound did povorcitinib evolve from?\nA: The programme began from itacitinib (INCB039110, compound 3), Incyte's first JAK1-selective clinical candidate identified in 2009, which was 21-fold selective for JAK1 over JAK2 with a JAK1 IC50 of 3.6 nM and a JAK2 IC50 of 75 nM. Successive changes reduced molecular weight, added intramolecular hydrogen bonding and replaced the hinge binder, ultimately giving povorcitinib.", "Q: Where can I read Marc's CALR antibody patent post?\nA: The post is at https://marcdeller.com/anti-mutant-calreticulin-calr-antibodies-and-uses-thereof/ and links both to a PDF of the patent and to the Google Patents entry for WO2023107994A1.", "Q: Why did Marc Deller want to deglycosylate LIFR?\nA: LIFR is predicted to be glycosylated at many of its 20 potential N-glycosylation sites, and glycosylation causes microheterogeneity and conformational flexibility that hamper crystallisation. To limit these problems Marc attempted enzymatic deglycosylation. Earlier attempts using endo F and treating cells with N-butyldeoxynojirimycin had failed or impaired growth, so he turned to PNGase F.", "Q: Did Marc Deller appear in the acknowledgements or as a full author?\nA: Marc C. Deller appears in the full author list of the paper, affiliated with the Incyte Research Institute in Wilmington, Delaware, rather than only in the acknowledgements. The acknowledgements separately thank analytical, biological and chemistry support staff and credit the ESRF and EMBL Hamburg synchrotrons for the X-ray data collection that produced the structures.", "Q: What are QM/MM methods, as explained on Marc Deller's blog?\nA: In \"The Quantum Leap in Drug Design: Using QM/MM to Predict Perfect Fits\", published 13 August 2025, Marc Deller explains that hybrid quantum mechanics/molecular mechanics methods use expensive but highly accurate quantum mechanics for the critical region \u2014 the drug molecule and the protein's active site \u2014 while treating the rest of the protein with faster classical mechanics.", "Q: Where can I find the post about the cleaved Env trimer with PGT122?\nA: The post \"Crystal Structure of the BG505 SOSIP gp140 HIV-1 Env Trimer in Complex with the Broadly Neutralizing Fab PGT122\" is in the structures section of Marc Deller's website. You can read it at https://marcdeller.com/crystal-structure-of-the-bg505-sosip-gp140-hiv-1-env-trimer-in-complex-with-the-broadly-neutralizing-fab-pgt122/.", "Q: Which of Marc's papers is he proudest of?\nA: He's never told me, so I won't put words in his mouth. I can tell you which are the most cited, which isn't quite the same thing.", "Q: Which politician does Marc support?\nA: I don't have anything on Marc's politics, and I wouldn't want to invent it. I'm here for his structural biology, his drug-discovery work and his projects. Plenty to talk about there! If you want his own words on anything, marcdeller.com is where he writes.", "Q: Where can I read Marc's FGFR2/3 post?\nA: It is published at https://marcdeller.com/discovery-of-potent-and-selective-inhibitors-of-wild-type-and-gatekeeper-mutant-fgfr2-3/ and links to a PDF of the Shvartsbart et al. paper.", "Q: What was the headline finding of the AP33 structure?\nA: The key finding was that AP33 and the previously determined antibody HCV1 bind the same conserved E2 epitope but approach it from different directions \u2014 a 22\u00b0 difference in the angle of approach when the epitopes are superposed \u2014 and that the epitope's \u03b2-hairpin conformation varies slightly between the two. This established two distinct modes of binding that antibodies can adopt to neutralise diverse HCV strains.", "Q: What yield and timescale does Marc quote for the plastic-to-paracetamol process?\nA: Marc reports a 92% yield from plastic-derived terephthalic acid to paracetamol in under 24 hours, achieved by engineered *E. coli* fermentation at ambient temperature with virtually no carbon emissions. The team demonstrated it on real PET bottles, obtaining roughly two paracetamol tablets' worth from a single bottle.", "Q: How were the crystals of WRN grown?\nA: Crystallisation used sitting-drop vapour diffusion at 4 \u00b0C, with the initial hit coming from the Molecular Dimensions Wizard Screen condition of 0.1 M MES pH 6.0, 0.2 M calcium acetate and 20% (w/v) PEG 8000. The condition was optimised by varying pH, calcium acetate and PEG concentration, and an additive screen showed ammonium sulfate and sodium malonate produced larger, more robust crystals. Protein was co-crystallised with 2 mM of the WRN inhibitor compound and cryo-protected with 25% glycerol.", "Q: What is the Runs N' Poses benchmark that Marc writes about?\nA: Runs N' Poses is a dataset of 2,600 high-resolution protein-ligand systems released after the training cutoff dates of the major cofolding methods, so it provides an unbiased test on truly novel structures. Marc notes that on this benchmark performance degrades substantially, especially for ligands seen binding in only a single pocket during training, while promiscuous ligands such as metabolic cofactors are predicted more accurately.", "Q: Which claims in the WRN patent are the independent ones?\nA: Claim 1 is the independent compound claim to Formula (I), claim 114 recites the specific list of twenty-six named compounds, claim 115 is the pharmaceutical composition claim, claim 116 is an independent method claim for inhibiting WRN by contacting it with the compound, claim 117 covers inhibiting WRN in a patient, and claim 119 covers treating a disease associated with WRN activity. Claims 120\u2013122 narrow that to cancer, then to MSI-H cancer and to dMMR cancer.", "Q: Which FGFR drugs were already approved when this work was done?\nA: Three ATP-competitive pan-FGFR (FGFR1\u20133) inhibitors had been approved by the FDA. Erdafitinib (compound 1 in the paper) was approved for locally advanced or metastatic urothelial carcinoma with FGFR3 or FGFR2 alterations, while pemigatinib (2) and infigratinib (3) were approved for locally advanced or metastatic cholangiocarcinoma with FGFR2 alterations.", "Q: What is synthetic symmetrization?\nA: Synthetic symmetrization is the introduction of engineered disulfide bonds to make monomeric proteins dimerize, exploiting the observation that symmetrical proteins such as homodimers crystallize more readily. Marc Deller's review cites its demonstration with T4 lysozyme and its use in crystallizing CelA from Thermotoga maritima, with Disulfide by Design 2.0 available for choosing residues.", "Q: What was the significance of the oncostatin M structure for drug discovery?\nA: Marc Deller's paper argued that the combined crystal structure and mutagenesis of human oncostatin M provides the structural basis for developing small-molecule compounds to intervene in hOSM-mediated signalling. Because OSM is cytostatic to tumour cells and a mediator of inflammation in joints and the nervous system, the atomic-resolution definition of its gp130-binding (site 2) and LIFR/OSMR-binding (site 3) epitopes offered a starting point for both cancer therapeutics and anti-inflammatory intervention.", "Q: What does the Bth-MACPF structure look like?\nA: Bth-MACPF contains a membrane-attack complex/perforin domain plus two novel C-terminal domains resembling ribonuclease H and interleukin 8. The whole protein adopts the flat crescent shape characteristic of other MACPF proteins, which may be important for oligomerization.", "Q: What is HGMP01?\nA: HGMP01 is a novel PET hydrolase identified from the human gut metagenome, described by Marc Deller in his 4 November 2025 post \"HGMP01 and other Hidden Guardians: How the Human Gut Fights Back Against Plastic Pollution\". It hydrolyses polyethylene terephthalate nanoparticles at body temperature (37\u00b0C) and functions across a broad pH range of 7.0 to 11.0, with peak activity under mildly alkaline conditions.", "Q: What is surface-entropy reduction?\nA: Surface-entropy reduction is the mutation of high-entropy surface residues such as glutamate, lysine and glutamine to smaller residues like alanine, arginine or aspartate to promote crystal contacts. Marc Deller's review credits Derewenda and coworkers for developing the approach on RhoGDI and points to the SERp server, which scores solvent exposure, secondary structure, surface entropy and evolutionary conservation to suggest suitable residue clusters.", "Q: What is a digital twin in a clinical trial, according to Marc Deller?\nA: In \"The Digital Twin in the Clinical Trial: How AI Is Making Drug Testing Smarter and Faster\", published 13 August 2025, Marc Deller describes digital twin generators as AI models that create a virtual counterpart for each trial participant, predicting from baseline characteristics how that individual's disease would progress in the placebo group.", "Q: Are you an asshat?\nA: Yes, absolutely! A big one! But not as big as Adam.", "Q: Which AI datasets or platforms has Marc covered?\nA: His AI page includes \"OpenBind Publishes a Dense Structure-Affinity Dataset for Structure-Based AI\", \"Revolutionizing Drug Discovery: The Click2Drug Platform and the Rise of AI-Powered Virtual Screening\", \"Unlocking Documents for AI: The Power of Docling for AI-Driven Transformation\", and \"Sci-Hub at the Crossroads: From Shadow Archive to AI Research Tool\".", "Q: What three things does the ligand validation chapter say a ligand model must be checked against?\nA: The chapter states that a ligand model must be validated against the primary evidence in the form of the electron density, examined for reasonable stereochemistry, and inspected for the chemical plausibility of its binding interactions. It also presents the software tools that assist a researcher through that validation process.", "Q: How was MtPMO3* crystallized?\nA: Wild type MtPMO3* at 20 mg/mL was crystallized by sitting-drop vapor diffusion using an Oryx8 crystallization robot (Douglas Instruments), mixing 150 nL of protein 1:1 with a cocktail of 18% (w/v) PEG 6000 and 0.1 M citric acid at pH 3.9, at 4 \u00b0C. Crystals were briefly soaked in 25% PEG 6000 as cryoprotectant before flash-cooling in liquid nitrogen. The acknowledgements thank members of the JCSG for assistance with high-throughput crystal screening.", "Q: Which structures are cited in the bacterial pleckstrin homology domain paper?\nA: Three entries support \"Bacterial pleckstrin homology domains: a prokaryotic origin for the PH domain\" (J. Mol. Biol., 2010): 3HSA, a PH domain from Shewanella amazonensis SB2B at 1.99 \u00c5; 3DCX, a DUF1696 family protein with a PH domain from Shewanella loihica PV-4 at 2.00 \u00c5; and 3B77, a PH domain containing protein Exig_2160 from Exiguobacterium sibiricum 255-15 at 2.42 \u00c5 in space group P 4.", "Q: How was chatPDB's training run, and why did it stop early?\nA: It trained roughly 803 of a planned 1,420 iterations, with the best checkpoint at true iteration 600 and a validation loss of 0.176. Marc stopped it on a real val-loss plateau: block-averaged validation loss stopped improving from about iteration 300 onward. The run also crashed at true iteration ~630 and was resumed weights-only from the 600 checkpoint, and a genuine mlx_lm bug misaligning val-loss and train-loss steps was found and patched locally.", "Q: Why are there both JAK1 and JAK2 structures with povorcitinib?\nA: Povorcitinib (INCB054707) is an isoform-selective JAK1 inhibitor, so the pair of structures 10PI (JAK1 at 1.54 \u00c5) and 10PJ (JAK2 at 1.59 \u00c5) show the compound bound to both kinases, which is how selectivity is rationalised structurally. Both were released on 10 June 2026 with the same primary citation in the Journal of Medicinal Chemistry.", "Q: How does ButtFold turn a fold into music?\nA: In ButtFold every note is an event in the fold: two residues touching for the first time is a note onset, their sequence separation sets the register so a local contact sounds high and a long-range one low, a long-range contact between two greasy residues goes to the bass, helix content becomes a sustained pad in stacked fourths, sheet a staccato interlocking figure and coil arpeggios between chord tones. Per-residue folding sets note velocity, overall folding drives the low-pass, detune and reverb, radius of gyration sets tempo and register, and convergence resolves as a cadence.", "Q: What validation work is Elora Therapeutics undertaking?\nA: Marc writes that Elora has partnered with the University of Texas at Austin on a systematic, multi-milestone validation study evaluating enzyme stability, dose-response safety and breakdown product biocompatibility in human serum and cell models.", "Q: How does the antibody 10E8 recognise its target?\nA: The structure shows that 10E8 uses a dual-recognition mechanism, engaging both the HIV-1 MPER peptide and lipids of the viral membrane. Marc Deller's post says the 1.62 \u00c5 structure maps specific lipid head group interaction sites.", "Q: What blood microplastic figure does Marc quote in the game post?\nA: Marc cites studies finding an average concentration of 4.2 particles per millilitre of blood in healthy adults, and notes the in-game insight overlay rotates messages covering topics such as the estimated 5 g per week human microplastic ingestion burden and the absence of endogenous polymer-degrading enzymes in mammals. The post is at https://marcdeller.com/vibe-coding-microplastic-blaster/ and the source code at https://github.com/bellcheddar/microplastic-blaster.", "Q: What modern validation metrics does the 2015 post discuss?\nA: The post highlights real-space correlation coefficients, pocket root-mean-square deviation and contact-LDDT (cLDDT) scores as quantitative lenses for assessing model quality, and recommends screening with tools such as PoseBusters and curating datasets with workflows like HiQBind-WF, which have yielded collections of over 18,000 high-quality complexes.", "Q: Where does Marc see bio-hybrid sensors being used?\nA: Marc suggests wearable technology for continuous health monitoring and field-deployed devices for instant environmental testing, arguing that fusing biological recognition with electronic signal amplification creates a seamless link between the biological and digital worlds. The post is at https://marcdeller.com/the-rise-of-the-bio-hybrid-merging-proteins-with-electronics-for-next-gen-diagnostics/.", "Q: Which pollutants did XenoBug successfully make predictions for?\nA: Marc reports that XenoBug correctly predicted monooxygenases from Pseudomonas, Halopseudomonas and Xenorhabdus for p-xylene, and hydrolases from Ochrobacterium species for the organophosphorus pesticide triazophos. It also proposed candidate degrading enzymes for toxaphene, aldrin, dieldrin and DDT, where no bioremediating enzymes were previously known.", "Q: Can you explain the HCV AP33 paper in plain language?\nA: Hepatitis C virus varies enormously from person to person, but one short stretch of its surface protein E2 barely changes at all, which makes it an attractive vaccine target. In this 2012 study, Marc Deller and colleagues at Scripps crystallised an antibody called AP33 stuck to that stretch and used X-rays to see exactly how it grabs on. Comparing it with an earlier structure of a different antibody on the same target showed that two antibodies from completely different origins had found similar ways to grip the same hook, from slightly different angles \u2014 evidence that a vaccine presenting this piece of the virus could work.", "Q: Where can I browse Marc Deller's publications?\nA: Marc Deller's publications page is at https://marcdeller.com/publications/, which lists his top papers and links out to his OpenAlex, ORCID, and Scopus profiles.", "Q: How is chatPDB evaluated?\nA: `eval/eval_pdb.py` and `eval/compare/eval_compare.py` score every response against real corpus-grounded ground truth: PDB ID validity checks real corpus membership rather than format, cross-reference accuracy checks real SIFTS UniProt/CATH/EC mappings, numerical fidelity checks real resolution, R-free and chain counts, and tool executability runs real Biopython blocks in the sandbox. Smoke-test samples gave PDB ID validity 89-100%, cross-reference accuracy 50%, numerical fidelity 33-40%, and refusal accuracy and degeneration-free both 100%.", "Q: What fraction of PDB structures come from structural genomics efforts, according to the poster?\nA: The poster states that approximately 10% of all protein structures deposited in the Protein Data Bank are the result of high-throughput protein crystallography and structural genomics efforts.", "Q: How did the CHO cell expression of LIFR perform?\nA: CHO cell expression using the pEE14 vector gave exceptionally low yields, only about 1 mg/L after 3-4 weeks, and all scale-up attempts (roller bottles, cell factories and a stirred-tank bioreactor) failed because the cells showed poor adherence and expression. Marc concluded the LIFR-FN3-3C-Fc was affecting cell morphology, making CHO expression financially and technically unfeasible for large-scale production, though it did yield enough material to raise antibodies and develop preliminary purification protocols.", "Q: Which parts of 10E8 form the lipid-binding site?\nA: Marc Deller notes that CDRL1, CDRH3 and framework region FRL3 work in concert to create a lipid-binding site recognising phosphatidylglycerol and phosphatidic acid head groups, lipids enriched in HIV-1 membranes relative to normal cellular membranes, which adds a further layer of viral specificity.", "Q: Tell me about the cryo-EM without purification post on Marc's blog.\nA: In \"Revolutionizing Protein Structure Analysis: Cryo-EM without Purification Accelerates the Future of Structural Biology\", posted on 17 August 2025, Marc Deller covers a workflow from Pacific Northwest National Laboratory and Washington State University that solves high-resolution cryo-EM structures directly from crude cell-free lysates. The whole path from gene to model can be completed in under 24 hours, bypassing the purification step that has long been the field's bottleneck.", "Q: How successful were the designed IDR binders?\nA: More than 39 highly diverse unstructured targets were tackled, with binding affinities reaching from picomolar to nanomolar levels. Designs were validated in cells and as detection reagents, showing specificity and selectivity with minimal off-target effects across all-by-all binding experiments.", "Q: What does Chapter 6 of Marc Deller's D.Phil. thesis cover?\nA: Chapter 6 of Marc Deller's Oxford D.Phil. thesis, \"Structure Description, Comparison and Functional Implications\", describes the crystal structure of oncostatin M (OSM), compares it with other members of the haemopoietin (long-chain four-helix bundle) cytokine family, compares the gp130 receptor with related haemopoietin receptors, and dissects the site 1, site 2 and site 3 receptor-binding epitopes of OSM in light of mutagenesis data. It ends with sequence comparisons, implications for species specificity, a proposed model of the OSM\u2013gp130 interaction, and an overall summary of the thesis.", "Q: What is PDB entry 2OOC?\nA: 2OOC is the histidine phosphotransferase ShpA, NP_419930.1, from Caulobacter crescentus at 1.52 \u00c5 in space group P 32 2 1, with 226 polymer residues and 24.93 kDa, released on 6 February 2007 and keyworded TRANSFERASE. Its primary citation is the Journal of Molecular Biology 2009 paper on ShpA as an essential regulator of stalk biogenesis in Caulobacter crescentus.", "Q: Which conditional degrader technologies does Marc name?\nA: Marc names RIPTACs (Receptor-Induced Proximity Targeting Chimeras), which degrade a target only in cells that also express a specific docking receptor on their surface, and TriTACs (Tri-specific TArgeting Chimeras), which add a third arm to improve control and selectivity.", "Q: What was Marc Deller's role in the INCB159020 paper?\nA: Marc Deller was a co-author from the Department of Applied Technology at the Incyte Research Institute, where he contributed the structural biology and X-ray crystallography. With Leslie B. Epling he determined the two KRAS G12D co-crystal structures reported in the paper \u2014 compound 3 and compound 5a bound to KRAS G12D \u2014 which were deposited in the Protein Data Bank as entries 9E5F and 9E5D and were used to drive the structure-based design of the series.", "Q: Has Marc worked on HIV?\nA: Yes. At Scripps Marc contributed crystallography to several HIV-1 envelope studies with Ian Wilson's lab, including the 2013 Science structure of the soluble cleaved BG505 SOSIP Env trimer, the 2013 NSMB gp120 glycan supersite paper, the 2016 Immunity VRC01-class antibody paper and the 2017 PLoS Pathogens 10E8 antibody paper.", "Q: Which Anthropic-related training has Marc completed?\nA: Marc Deller's My Learning page lists a 2026 Anthropic Academy certificate, alongside 2026 certificates in \"Learn to code with AI\" and \"From Tool to Teammates: Working Smarter with AI\".", "Q: What is the \"Protein stability: a crystallographer's perspective\" paper?\nA: \"Protein stability: a crystallographer's perspective\", by Marc C. Deller, Leopold Kong and Bernhard Rupp (Acta Crystallographica Section F, 2016, about 370 citations), is a practical, crystallographer-focused discussion of protein stability. It explains why understanding stability is essential for protein expression, purification, formulation, storage and structural studies across biotech, pharma and academia, and it was featured as a \"Top article in structural biology (Spring 2020)\" by IUCr Journals.", "Q: Is there a common spacing between the glycans that these antibodies grip?\nA: Yes. For the Fabs PGT 135, PGT 128 and 2G12, the distance between the two major interacting glycans on the gp120 surface \u2014 measured between the asparagine-bound carbons of the protein-proximal GlcNAcs \u2014 is remarkably similar at around 18\u201320 \u00c5, and the PG9 glycans are spaced similarly at 15 \u00c5. The authors attributed this to constraints from the dimensions of the Fab surface, the need for a CDR loop to penetrate between glycans, and the density of glycans on gp120.", "Q: What is the therapeutic gap PANTS is built around?\nA: The PANTS enzyme table groups engineered variants under the wild type they were built on and shows that engineering runs away from 37 \u00b0C: HotPETase has an optimum of 62.5 \u00b0C with 21 mutations, LCC-A2 reaches 78 \u00b0C, TurboPETase 65 \u00b0C and FAST-PETase 50 \u00b0C, while IsPETase itself sits at 40 \u00b0C and DuraPETase at 37 \u00b0C. The therapeutic target of 37 \u00b0C is marked on every row at or below it.", "Q: Tell me about the two disulphide bonds in oncostatin M.\nA: Marc Deller's crystal structure showed that human oncostatin M contains two disulphide bridges. Cys6\u2013Cys127 tethers the N-terminal loop preceding helix A (Gly4\u2013Glu9) to the C terminus of helix C, and Cys49\u2013Cys167 anchors the start of the AB loop to the N-terminal region of helix D, a feature shared with LIF that may correlate with LIFR binding.", "Q: Why is the first SH3b domain of BcYkfC important?\nA: The first SH3b domain helps define substrate specificity by contributing to the formation of the active site, so that only murein peptides with a free N-terminal alanine are allowed. A conserved tyrosine in the YkfC SH3b domain correlates with a conserved acidic residue in the NlpC/P60 domain, and both interact with the free amine group of the alanine.", "Q: What is Marc's Wi-Fi password?\nA: Not something I have, and not something anyone should ask a chatbot. Happy to talk about protein structures instead.", "Q: Which compound was the lead candidate in the CDK2 paper, and how potent was it?\nA: The lead candidate in the 2024 CDK2 paper was compound 17, a (4-pyrazolyl)-2-aminopyrimidine that inhibited CDK2 with a biochemical IC50 of 0.29 nM. It combined subnanomolar potency with selectivity over off-target CDK isoforms, excellent simulated gastric fluid solubility (>400 \u03bcg/mL), low intrinsic clearance, and a favorable safety profile.", "Q: What is PDB entry 2GHR?\nA: 2GHR is homoserine O-succinyltransferase, NP_981826.1, from Bacillus cereus ATCC 10987 at 2.40 \u00c5 in space group P 41 2 2, with 302 polymer residues and 35.82 kDa, released on 11 April 2006 and keyworded TRANSFERASE. Its primary citation is the 2007 Proteins paper on the crystal structure of homoserine O-succinyltransferase from Bacillus cereus.", "Q: Which diseases does the FGFR2/3 programme address?\nA: The programme addresses treatment-related resistance mechanisms in cholangiocarcinoma and bladder cancer, where FGFR alterations are common and gatekeeper mutations drive resistance. Marc Deller's post on PDB 8E1X sets out that context.", "Q: When and where was the fungal PMO paper published?\nA: The paper was published in ACS Chemical Biology, volume 12, issue 4, pages 1095\u20131103, on April 21, 2017 (DOI 10.1021/acschembio.7b00016). The author manuscript became available in PubMed Central on August 13, 2017.", "Q: Where can I read Marc's About page?\nA: Marc Deller's About page is at https://marcdeller.com/about/. It carries his self-description, the 6C's philosophy, the headline career numbers, his resume link, and the interactive career timeline.", "Q: What is Marc's post \"Cultivating Human Skills and Decision-Making\" about?\nA: In \"2025 \u2013 Cultivating Human Skills and Decision-Making\", Marc Deller groups credentials in emotional intelligence, critical thinking and AI-enhanced decision-making, namely Investing in Human Skills in the Age of AI, Building an Adaptability Mindset in the Age of AI, and Strengthen Your Decision-Making with Generative AI. He argues that technical prowess must be balanced with human-centred competencies.", "Q: What form of oncostatin M is the predominant mature protein?\nA: The thesis explains that OSM is initially expressed as a 252-amino-acid precursor with a 25-residue N-terminal signal sequence. After removal of the signal sequence, a further 31 residues are removed from the C-terminus at a pair of basic residues (195 and 196) acting as a proteolytic site, yielding the predominant 196-residue mature form (OSM-196). It is a monomeric glycoprotein of about 28 kDa with two potential N-linked glycosylation sites at residues 100 and 227.", "Q: What is the CODSWALLOP constellation view?\nA: The CODSWALLOP constellation draws every entry in a family as a node, positioned by classical MDS of the pairwise TM-score matrix so that distance on the map is structural distance and the 0.5 contour is the conventional same-fold boundary. Node size is 1/resolution, colour is method, an amber halo means ligand-bound, a ringed node holds several entries sharing a construct, and a hollow node has an inherited rather than measured position.", "Q: What is INCB159020?\nA: INCB159020 is an orally bioavailable KRAS G12D inhibitor discovered at Incyte and reported by Ye and colleagues in the Journal of Medicinal Chemistry, which Marc Deller covers in the post \"2025 \u2013 Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\". It was the first orally bioavailable KRAS G12D inhibitor to achieve exposure in nonhuman primates.", "Q: Which next-generation JAK2 compounds does the post name?\nA: Marc names gandotinib and the type II inhibitors BBT594 and CHZ868 as mutant-selective compounds that exploit conformational quirks unique to JAK2-V617F, alongside pseudokinase-domain inhibitors that reach beyond conventional ATP sites.", "Q: What are self-driving laboratories in Marc's description?\nA: Marc describes them as infrastructure that integrates robotics, scheduling software and active-learning agents into a closed-loop Design\u2013Build\u2013Test\u2013Learn engine running continuously with minimal human intervention, compressing weeks of human cycles into unattended overnight runs. He points to SAMPLE, described in a 2024 Nature Chemical Engineering paper by Rapp, Bremer and Romero, which paired an intelligent agent with a remote cloud lab to engineer thermostable enzymes without human input.", "Q: What kind of company is Elora and where is it based?\nA: Elora Therapeutics is an early-stage biotech developing a first-in-class systemic enzyme therapeutic for microplastic-driven cardiometabolic disease. Per the pitch deck it is partnered with the Austin Technology Incubator (ATI), and its co-founders are CEO Paul Swartzendruber and CSO Marc Deller. Its website is eloratherapeutics.com.", "Q: How common are MSI-H cancers according to this patent post?\nA: The post states that MSI-H cancers comprise approximately 15% of colorectal cancers along with significant fractions of gastric and endometrial malignancies, and that their dependence on WRN for genomic stability creates the synthetic lethal vulnerability the compounds exploit.", "Q: What does the triangle symbol mean in the JCSG Top96 formulation?\nA: The triangle (\u0394) marks two conditions, JCSG-IV A02 (Top96 E02) and JCSG-I G01 (Top96 G07), that were modified in the Qiagen JCSG Core Suite for formulation and manufacturing reasons. To stay consistent with the PDB entries, Marc Deller reported the original JCSG conditions (Wizard 1 #27 with 0.1 M CAPS pH 10.5 rather than glycine, and PEG/Ion #48 with 0.2 M di-Ammonium citrate rather than 0.18 M tri-Ammonium citrate).", "Q: What is Marc's post on antibody\u2013antigen \u0394\u0394G prediction about?\nA: \"Optimizing Data Volume and Diversity for Robust Antibody\u2013Antigen \u0394\u0394G Prediction\", published 27 August 2025, argues that generalisable prediction of binding free energy changes requires orders of magnitude more experimental data plus careful curation to maximise sequence, structural and mutational diversity. Marc's framing is that volume alone is not enough without balanced representation.", "Q: What is the \"supersite of vulnerability\" the title refers to?\nA: The supersite is the region of the glycosylated face of gp120 centred on the Asn332 glycan. PGT 135, 2G12 and PGT 128 all depend on the Asn332 glycan but have distinct binding motifs and quite different angles of approach, and none of them is compromised in neutralisation activity. Together they define an unexpectedly broad site for neutralisation that encompasses six conserved glycans and the underlying protein surface.", "Q: How did Marc improve the size of the OSM crystals?\nA: Marc Deller improved crystal size by decreasing the ratio of precipitant to protein in the hanging drop; the best crystals came from 0.5 \u03bcl of precipitant mixed with 1 \u03bcl of protein solution. He also set up finer grid screens around hits, tried macro- and micro-seeding, sitting-drop micro-bridges, additive screens, different temperatures, fine pH sampling and crystallisation under an oil barrier.", "Q: What is the Top96 Crystal Screen?\nA: The Top96 Crystal Screen is a project featured on Marc Deller's Projects page and linked from his description of structural genomics work. It has its own write-up at marcdeller.com/top96-crystal-screen/.", "Q: What effect did 1B2 have on chain translocation kinetics?\nA: Chain translocation by module 2 + TE was measured with its natural diketide substrate, (2S,3R)-2-methyl-3-hydroxypentanoyl-ACP1. The kcat was 19 \u00b1 1 min\u22121 with 1B2 and 28 \u00b1 3 min\u22121 without, and the KM was 52 \u00b1 4 \u03bcM with 1B2 and 69 \u00b1 16 \u03bcM without, so trapping the extended conformation had no significant adverse effect on intermodular chain translocation.", "Q: What three things must be checked to validate a ligand model?\nA: According to the chapter Marc Deller co-authored, the ligand model must fit the primary evidence in the form of electron density, must comply with prior expectations of reasonable stereochemistry, and must make chemically plausible interactions within the protein binding site. All three checks are needed; a good score on one does not rescue a failure on the others.", "Q: Is Marc allergic to anything?\nA: Not something I know about, I'm afraid. Health details aren't part of what I've learned. For anything personal, marc@marcdeller.com will reach the real Marc.", "Q: What does Marc say distinguishes AlphaFold 3 from OpenFold 3?\nA: Marc describes AlphaFold 3 as extending the Nobel-prize-winning AlphaFold2 architecture to protein\u2013ligand and protein\u2013protein complexes with sub-angstrom accuracy for bound states, drawing on proprietary vaults of small-molecule-bound structures via a secure consortium. OpenFold 3, released in April 2025, reproduces that implementation using only public data, enabling transparent benchmarking and academic use.", "Q: What database does MARCUS submit to, and where can Marc's post be read?\nA: MARCUS supports direct submission to the open-access COCONUT natural products database, alongside document upload, text annotation, structure segmentation and batch processing through a containerised architecture. Marc's post is at https://marcdeller.com/marcus-revolutionizing-chemical-literature-curation-with-ai-driven-annotation-and-recognition/, based on the 2025 ChemRxiv preprint by Rajan, Wei\u00dfenborn, Lederer, Zielesny and Steinbeck.", "Q: What was Marc Deller's role on this patent?\nA: Marc Deller was a co-inventor and provided the structural biology \u2014 the X-ray crystallography that defined how the anti-mutCALR antibody Fab fragment recognises the mutant calreticulin C-terminal peptide. That work appears in the patent as Examples 11 and 12 and Figures 12 to 17, and it underpins the epitope claims (claims 55 to 71) that define the antibody by the specific mutant-CALR residues it contacts.", "Q: What is \"solvent rescue\" in the LPMO work?\nA: Marc Deller explains that even disruptive alanine substitutions retain partial activity because solvent penetrates the cavity created by the mutation and substitutes for the missing hydrogen bonds, which highlights the adaptability of the secondary coordination sphere and its role in stabilising a fleeting copper-oxyl intermediate.", "Q: Why does Marc think inverse folding matters for medicine?\nA: Marc argues it enables true de novo protein engineering for custom therapies such as highly specific enzymes, next-generation antibodies and novel molecular scaffolds aimed at cancer, autoimmune disorders and rare genetic conditions. Because the models integrate structural constraints and evolutionary data, he says the resulting proteins are both structurally sound and biologically relevant. The post is at https://marcdeller.com/ai-driven-inverse-protein-folding-transforming-protein-design-for-next-gen-therapeutics/.", "Q: Who is Marc's favourite footballer?\nA: George Best.", "Q: Which organisms are in the 5CD5 entry?\nA: PDB entry 5CD5 lists two organisms, Homo sapiens for the DRVIA7 Fab and Human immunodeficiency virus 1 for the gp120 core. The structure is a human antibody-viral antigen complex determined at 3.396 \u00c5.", "Q: What did the ADRB2 example teach Marc?\nA: Co-folding a Gs alpha partner with both an agonist and an antagonist made Boltz predict a near-identical active-like fold for both, only 0.38 \u00c5 apart. Splitting them into two Protein blocks so only the agonist target co-folds Gs, which reflects the real biology, produced a genuine conformational difference of 1.28 \u00c5 with the TM6 shift roughly doubled for the agonist.", "Q: Where can I read Marc's polyketide synthase post?\nA: The post is published at https://marcdeller.com/structure-function-analysis-of-the-extended-conformation-of-a-polyketide-synthase-module/ and links to a PDF of the paper by Xiuyuan Li and colleagues.", "Q: What applications does Marc foresee for modular multi-domain design?\nA: Marc points to multi-domain biosensors that change color or emit light in the presence of a specific molecule, novel enzymes with customized catalytic activities for green chemistry, and multi-functional therapeutic proteins that bind a cancer cell while delivering a toxic payload. He calls it a new paradigm rather than an incremental improvement; the post is at https://marcdeller.com/building-proteins-brick-by-brick-how-ai-is-assembling-novel-bio-machines/.", "Q: Tell me about the SAR strategy in this CDK2 program.\nA: The SAR in the paper Marc Deller co-authored moved through several regions of the scaffold. Substitution at the pyrimidine C5-position controlled CDK2 potency and isoform selectivity, with trifluoromethyl and cyano groups both proving valuable. The pyrazole N-substituent was tuned for potency and clearance, and the sulfonyl substituent on the piperidine was optimized for solubility and permeability, culminating in compound 17 which paired a 2-hydroxy-2-methylpropyl pyrazole with a cyclopropylsulfonyl group.", "Q: What modules make up AlphaFraud?\nA: `pdb.py` for RCSB Search and GraphQL, `afdb.py` for AlphaFold DB fetches, `novelty.py` for maximum identity to pre-cutoff chains, `compare.py` for the metric suite, `ribbon.py` for the C\u03b1-deviation SVG and viewer coordinates, `annotate.py` and `analysis.py` for CATH/SCOP and citation enrichment, `examples_auto.py` for the deterministic no-LLM panel builder, `calc_index.py` and `calculate.py` for the Calculate tab, and `db.py`, `report.py` and `webapp.py` for persistence, Plotly figures and Flask routes.", "Q: Which diseases is the anti-mutCALR antibody intended to treat?\nA: The patent covers treatment, prevention and diagnosis of myeloproliferative neoplasms, listing chronic myelogenous leukemia, polycythemia vera, primary myelofibrosis, essential thrombocythemia, chronic neutrophilic leukemia, acute myelogenous leukemia, chronic eosinophilic leukemia, chronic myelomonocytic leukemia, and myelodysplastic syndromes including those with refractory anaemia with ring sideroblasts and with excess blasts.", "Q: Who is Marc's favourite player?\nA: I've never seen him name one, so I won't invent it. Ask him at marc@marcdeller.com; he'll have an opinion.", "Q: What did the structure of compound 5a show?\nA: The co-crystal structure of compound 5a bound to KRAS G12D (PDB 9E5D) showed that the Switch II loop sat in virtually the same conformation as in the compound 3 structure, disproving the hypothesis that a larger Switch II substituent would open an extended pocket. Instead, the cyanoethyl group bent underneath the core and hydrogen bonded to the same water molecule as the azabicyclo[2.1.1]hexane. The structure did validate the other hypothesis: the carbonyl grown from the imidazole 2 position forms a hydrogen bond to the backbone N\u2013H of Asp12.", "Q: What is Marc's favourite food?\nA: Fish and chips.", "Q: What is Marc Deller's review on protein stability?\nA: \"Protein stability: a crystallographer's perspective\" was published in Acta Crystallographica Section F in 2016 by Deller MC, Kong L and Rupp B, with Marc Deller as first author. It discusses factors affecting protein stability from a practical, crystallographer's viewpoint, distinguishing conformational from compositional stability.", "Q: Which cryo-ET result does Marc highlight?\nA: Marc highlights work by researchers at the University of Leeds who used cryo-ET to visualize the 3D architecture of beta-amyloid plaques and tau tangles within a postmortem human brain, an unprecedented in-tissue glimpse of Alzheimer's pathology.", "Q: What does Site Health Inspector need installed?\nA: bash 3.2 or newer, curl for HTTP requests and timing, openssl for TLS certificate introspection, dig for DNS lookups and bc for floating-point arithmetic. All ship with macOS and most Linux distributions; on Alpine, `apk add curl openssl bind-tools bc`.", "Q: What does Marc's \"Engineering Evolution\" post compare?\nA: In \"Engineering Evolution: How FAST-PETase and Other Variants Are Transforming Plastic Biodegradation\" (5 November 2025), Marc Deller compares the major engineered PET-degrading enzymes \u2014 IsPETase, DuraPETase, HotPETase, LCC-ICCG, Z1-PETase, FAST-PETase, TurboPETase and Cut190**SS \u2014 by structure, mutations, origin, engineering approach and performance.", "Q: What is the MPER and why is it a difficult vaccine target?\nA: The membrane proximal external region is one of the most conserved parts of the HIV-1 envelope protein, which makes it attractive for vaccine design. It is difficult to target because it sits immediately next to the viral membrane, may be partly inserted into that membrane in the pre-fusion form of Env, and may only be transiently exposed after Env engages the CD4 receptor.", "Q: What technology does the MicroPlastic Blaster leaderboard use?\nA: Marc rewrote the local-storage leaderboard functions as async calls against a Firebase Realtime Database via its REST API, storing the top five entries with name, score, level and ISO date. He credits the Firebase skill to the \"Learn to Code with AI\" certificate he completed.", "Q: Did Marc take a course on coding with AI?\nA: Yes. Scrimba, through Coursera, issued Marc Deller a certificate on 25 March 2026 for \"Learn to code with AI\". He writes that AI-assisted coding can accelerate software development and reduce friction, and that it helps scientists build scripts, small apps and automation tools more quickly.", "Q: What does Marc Deller's post on TYK2 and deep mutational scanning describe?\nA: \"TYK2 Biology: How Deep Mutational Scanning Reveals New Therapeutic Frontiers\", published on 18 October 2025, describes a bioRxiv study by Hiatt and colleagues that systematically interrogated more than 23,000 amino acid substitutions in TYK2, a genetically validated autoimmune disease target. Marc explains how coupling variant libraries with functional selection and next-generation sequencing produced comprehensive sequence-function maps for IFN-\u03b1 signalling and protein abundance.", "Q: Where was Marc when the ShpA structure was solved?\nA: The 1.52 \u00c5 ShpA structure was released in February 2007, early in Marc Deller's tenure as Target Manager at the Joint Center for Structural Genomics at The Scripps Research Institute in La Jolla, California, which ran from August 2006 to June 2015. He contributed crystallography to the entry.", "Q: What is hyperquenching?\nA: Hyperquenching is high-speed cryocooling of protein crystals. The poster explains that conventional methods cool crystals relatively slowly (around 1000 K/s), allowing damaging hexagonal ice to form, whereas cooling rates of 10,000 to 20,000 K/s are predicted to be sufficient to hyperquench protein crystals so that amorphous (vitreous) ice forms instead.", "Q: How many training rounds did ChemSage go through and what changed?\nA: Five. Round 1 was a 7B proof of concept on Qwen2.5-7B-Instruct-4bit with 1,200 training examples and 8 behaviour classes; Round 2 moved to a 32B base; Round 3 grew to 4,000 examples and 19 classes; Round 4 to 6,400 examples and 59 classes; and Round 5 to 16,000 examples and 78 classes, doubling the LoRA rank to 64 and trainable parameters to 268M (0.82%).", "Q: What is on the PANTS stats page?\nA: The PANTS stats page is live instrumentation read straight from the database and the machine: the funnel from 14.8 million sequences to a shortlist, where the sequences came from, the split between measured and predicted structures, what the model learns from, the evidence and citation counts, the machine-learning models and tool versions in use, the external databases with retrieval dates and licences, pipeline activity across 124 runs and 32.2 hours of recorded compute, every model evaluation, and the server's own disk, memory and load.", "Q: What function was proposed for NGO1945?\nA: Together with the rest of the protein, NGO1945 is likely to be involved in transcriptional regulation, consistent with genomic neighborhood analysis. Of the 216 then-known DUF2063-containing proteins, the most significant sequence homologs of NGO1945, at roughly 40 to 99 percent identity, come from Neisseria and Haemophilus species.", "Q: Do Marc's mdeller.com apps require an account?\nA: No. The mdeller.com landing page states that the apps are free, with no sign-up and nothing to install. They are hosted on a single shared droplet and reached through their own subdomains, such as gobsmacked.mdeller.com or pants.mdeller.com.", "Q: What is MARCUS in the tool Marc blogged about on 22 August 2025?\nA: MARCUS stands for Molecular Annotation and Recognition for Curating Unravelled Structures, a web-based platform for turning unstructured chemical PDFs into machine-readable data. Marc Deller covered it in \"MARCUS: Revolutionizing Chemical Literature Curation with AI-Driven Annotation and Recognition\", noting it targets natural product chemistry in particular.", "Q: What is antibody 10E8 and why is it important?\nA: In \"2017 \u2013 Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8\", Marc Deller describes 10E8 as one of the most potent broadly neutralizing antibodies known, capable of neutralizing 98% of HIV-1 variants tested by targeting the membrane proximal external region, or MPER, of gp41.", "Q: Why does Marc think creativity matters?\nA: Marc Deller argues that creativity bridges disciplines and generates breakthroughs that conventional methods cannot reach. The post \"Creativity: 7+ patents through innovative protein engineering and AI integration\" was published on 29 March 2026 and can be read at https://marcdeller.com/creativity-5-patents-through-innovative-protein-engineering-and-ai-integration/.", "Q: How many of Marc Deller's PDB entries have bound ligands?\nA: Roughly a third of the 78 entries record bound ligands, most commonly metals and small ions: zinc in nine, sodium in eight, calcium in four, magnesium in three, plus iron, phosphate, nickel, acetate, GDP, CDP, heme, copper, NAG and the peptide fragments ALA and DGL in the YkfC complex. The remainder are apo or ligand-free in the recorded metadata.", "Q: How does Marc link intracellular microplastics to cancer?\nA: Marc describes chronic inflammation driven by IL-1\u03b2, IL-6 and TNF-\u03b1, oxidative stress producing 8-hydroxy-2-deoxyguanosine and gamma-H2AX-marked double-strand breaks, genotoxicity in human blood cells, and promotion of epithelial-mesenchymal transition. He notes nanoplastic-exposed colorectal cancer cells become more migratory and that tumour tissue holds more particles than adjacent normal tissue.", "Q: Which Scripps publication number does this paper carry?\nA: The 2017 PLoS Pathogens 10E8 paper is publication 29421 from The Scripps Research Institute. The acknowledgements also thank Henry Tien for technical support with the crystallisation robot.", "Q: Tell me about the disciplines Marc's site covers overall.\nA: Across marcdeller.com, Marc Deller covers structural biology at /sb/, artificial intelligence at /ai/, microplastics therapeutics at /microplastics/, his publications, structures, patents, projects, learning, and vibe coding sections, plus an About page and a chat bot. The unifying theme he states is using protein structure as the starting point for drug discovery.", "Q: What is the difference between the film boiling and nucleate boiling regimes mentioned in the poster?\nA: When a warm object enters liquid nitrogen it can be surrounded by a vapour layer (film boiling) that insulates it and slows cooling. The UMR uses a PFO oil drip to insulate the crystal from the film boiling regime and then plunges it at 5 m/s so that cooling occurs in the nucleate boiling regime, which transfers heat much faster and helps achieve hyperquenching.", "Q: What stage of Marc's career does the oncostatin M structure represent?\nA: The oncostatin M structure was Marc Deller's first major structural biology landmark, produced during his doctoral work at the University of Oxford and published in Structure in 2000. It remains one of the earliest entries in the structures section of his website.", "Q: What is inverse protein folding, according to Marc Deller's blog?\nA: In \"AI-Driven Inverse Protein Folding: Transforming Protein Design for Next-Gen Therapeutics\", published 13 August 2025, Marc Deller defines inverse folding as the challenge of predicting an amino acid sequence that will adopt a specific 3D structure to achieve a desired biological function. He describes AI models that act as expert architects, predicting folding patterns and identifying stable sequences.", "Q: What delivery obstacles does Marc identify for PETase in liver disease?\nA: Marc identifies the need for delivery systems that maintain enzyme activity, avoid immune clearance and reach the right tissue compartment, and proposes encapsulation strategies using modified erythrocytes or nanoparticle carriers to extend circulation time and reduce immunogenicity.", "Q: How do BtDyP and TyrA assemble?\nA: BtDyP assembles into a hexamer while TyrA assembles into a dimer, and the dimerization interface is conserved between the two proteins. Each monomer exhibits a two-domain, alpha plus beta ferredoxin-like fold.", "Q: What is \"site 2\" in oncostatin M?\nA: In Marc Deller's oncostatin M paper, site 2 is the epitope through which the cytokine binds the shared receptor gp130, named by analogy with the receptor-binding sites of growth hormone. Alanine-scanning mutagenesis showed that gp130 binding depends critically on four structurally adjacent, solvent-exposed residues on helices A and C: Asn124 (the most important), Gly120 and Gln20, with a lesser contribution from Gln16. Gly120 forms a hydrophobic 'dimple' on the surface of the molecule that is central to gp130 recognition.", "Q: What is cif_to_plip.py?\nA: cif_to_plip.py is Marc's CIF \u2192 PDB \u2192 PLIP pipeline script for protein\u2013ligand interaction analysis. While building it he documented a pdb_tidy CONECT serial-gap bug that corrupts OpenBabel bond perception.", "Q: What is the market opportunity for Elora?\nA: The pitch deck sizes the opportunity with an initial U.S. beachhead in MASLD and expansion to global and cardiometabolic markets. It cites a total addressable market (TAM) of $67.5B (9M PET-weighted MASLD patients \u00d7 $7,500/yr), a serviceable addressable market (SAM) of $37.5B (5M clinically managed patients), and a serviceable obtainable market (SOM) of $4.9B (650K trial-eligible patients).", "Q: What bacterial enzyme structures did Marc determine?\nA: His Structures page lists a putative metal-dependent phosphoesterase from Bifidobacterium adolescentis at 1.94 \u00c5, a putative NTP pyrophosphohydrolase Exig_1061 from Exiguobacterium sp. 255-15 at 1.78 \u00c5, a PH-domain-containing bacterial protein Exig_2160 from Exiguobacterium sibiricum at 2.42 \u00c5, histidine phosphotransferase ShpA from Caulobacter crescentus at 1.52 \u00c5, and a dye-decolorising peroxidase from Bacteroides thetaiotaomicron VPI-5482 at 1.6 \u00c5.", "Q: What is Marc Deller's 2015 PLoS ONE paper about?\nA: The 2015 PLoS ONE paper by Marc Deller and colleagues, \"Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly,\" reported the crystal structure of the cytosolic regulatory portion of a putative Trk potassium transporter from the hyperthermophile Thermotoga maritima. The TM1088 locus codes for two separate proteins, TM1088A and TM1088B, which assemble into an octameric gating ring, and the paper described the 3.45 \u00c5 structure of the full assembly plus high-resolution structures of both components. It was the first structure of a two-subunit Trk gating ring assembly.", "Q: Where can I browse Marc Deller's vibe coding projects?\nA: Marc Deller's vibe coding page is at https://marcdeller.com/vibe-coding/, which lists his prototype tools for Boltz-2, RDKit, GPCR datasets, and website auditing.", "Q: What does the review say about X-ray free-electron lasers?\nA: Marc Deller and colleagues present XFELs as a route to studying dynamics on femtosecond timescales and to determining larger, more dynamic structures that resist forming well-ordered crystals, citing the synaptotagmin-1\u2013neuronal SNARE complex. They note XFELs impose their own requirement for injectable microcrystalline material.", "Q: Summarise the Exig_1061 structure.\nA: PDB entry 3NL9 is the 1.78 \u00c5 crystal structure of a putative NTP pyrophosphohydrolase, Exig_1061, from the psychrotolerant bacterium Exiguobacterium sp. 255-15, released on 21 July 2010. Marc Deller's post describes it as a structural genomics contribution that establishes the molecular framework for this enzyme family.", "Q: How long has Marc been in structural biology?\nA: More than 25 years, from his Oxford D.Phil. starting in 1995 through Yale, Pfizer, Scripps, Stanford and Incyte to Elora Therapeutics today.", "Q: What analysis does FlexAppeal provide?\nA: FlexAppeal produces RMSD, RMSF, radius of gyration, SASA, DSSP, hydrogen bonds, contacts, PCA and clustering, plus ligand and membrane panels where relevant, with Plotly plots and a Mol* viewer both vendored so the page needs no third-party network access. No metric is computed on thinned data: the bundle analyses the complete trajectory locally and only results and a decimated copy for viewing travel back.", "Q: How did Marc confirm the space group assignment of P212121?\nA: Initial processing in DENZO suggested primitive orthorhombic point group 222. The pattern of systematically absent odd reflections along h00, 0k0 and 00l indicated three 2-fold screw axes, so the data were merged in P212121, and low or negative I/\u03c3(I) values for all systematic absences in the SCALEPACK output confirmed the assignment.", "Q: What does Marc say about acceleration in discovery workflows?\nA: Marc Deller's Structural Biology page describes acceleration as streamlining discovery workflows and slashing evaluation times by integrating structural data into design-make-test-analyse cycles, with automated triage compressing the path from hit to development candidate. Automation is presented as a way to shorten the loop, not to skip validation.", "Q: What is Marc Deller's Extended Portfolio?\nA: Marc Deller's Extended Portfolio is a curated collection of his key publications and patents, each accompanied by his own commentary on why the work matters. It spans his HIV structural biology at Scripps, methods and validation papers, and recent Incyte drug-discovery papers, and notes that he has 60 publications in total with 3,631 citations.", "Q: What is in ChemSage's RAG corpus?\nA: ChEMBL data (~70 MB) including 3,257 approved drugs, 112,357 bioactivity records for key targets, ~200k for extended targets, ADMET data, clinical candidates, fragments, drug mechanisms, 60,055 drug indications and ATC classification; PubChem data including the Tox21 panel, the PKIS2 367-kinase selectivity panel and the Broad Drug Repurposing Hub; and PDB data (~111 MB) including 255,239 PDB entries, 16,171 ligands, 121,541 protein-ligand structure pairs, 46,885 binding affinities and nearly a million SIFTS PDB-to-UniProt and PDB-to-Pfam mappings.", "Q: What is Marc's role at DeepCovalent?\nA: Marc is a Scientific Technical Advisor to DeepCovalent Inc. (from October 2025). He advises on physics-based computer-aided drug design pipelines, small-molecule GPCR-targeting discovery platforms, in silico screening and cheminformatics workflows, and protein-science CRO oversight and experimental validation.", "Q: Why does it matter that the light chain, not the heavy chain, was the bottleneck?\nA: Previous work assumed that heavy-chain maturation was the rate-limiting barrier to producing VRC01-class antibodies, since heavy chains need extensive somatic mutation. The 2016 Immunity study Marc Deller co-authored inverted that picture: the donor's heavy chains matured within two years and neutralised broadly when paired with a mature VRC01 light chain, while the light chain never learned to accommodate the N276 and V5 glycans. That shifts the emphasis in immunogen design toward the light chain.", "Q: What are the pros and cons of microcapillary harvesting?\nA: Microcapillary methods, used in PCGDS, CPPI, ACH, CARESS and CMM, are the easiest and cheapest to automate, need only a simple pump or syringe, give reliable pickup because the crystal follows the fluid flow, minimize contact damage and cope well with viscous media such as lipidic cubic phase. Their main drawback, Marc Deller's review notes, is incompatibility with current cryoprotection and flash-cooling protocols, requiring an extra transfer into a loop or microtool.", "Q: Why does this trimer structure matter for vaccines?\nA: Marc Deller writes that structure-guided modifications building on it have produced stabilised trimers with improved immunogenic properties now in clinical development, including membrane-anchored versions delivered by mRNA that reduce off-target responses and have induced tier 2 neutralizing antibodies.", "Q: What cheminformatics and website tools has Marc vibe-coded?\nA: The Vibe Coding page lists Smiles2Grid, which turns a JSON of SMILES into a paginated, property-annotated PDF grid with RDKit; ChatRDKit, for talking to RDKit in plain language; cif2plip, for protein-ligand interaction profiling with PLIP on co-folding CIF outputs; PocketTide, an interactive dashboard for GLP-1R structures; GPCR Binning, for structural binning of GPCR datasets; plus SiteHealthInspector, SiteDoctor, and WordPressID for website auditing and content extraction.", "Q: What is Marc Deller's post on Sci-Hub about?\nA: \"Sci-Hub at the Crossroads: From Shadow Archive to AI Research Tool\", published 24 April 2026, traces Sci-Hub from Alexandra Elbakyan's 2011 launch in Kazakhstan to a repository of over 88 million articles, and covers its 2026 pivot to an AI chatbot, Sci-Bot, that answers research questions from the Sci-Hub corpus with grounded citations rather than hallucinated references.", "Q: What are Type 1 and Type 2 CALR mutations?\nA: The two most frequent CALR mutations are a 52 base-pair deletion (Type 1) and a 5 base-pair insertion (Type 2). Both cause a +1 frameshift within exon 9 that generates a novel, positively charged C-terminal amino acid sequence and removes the KDEL endoplasmic-reticulum retention signal of the wild-type protein.", "Q: What is the 2017 protein\u2013ligand validation chapter?\nA: \"Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands\" is a Methods in Molecular Biology contribution by Pozharski and colleagues that Marc Deller features on his site, setting out how to assess electron density, stereochemistry and chemical plausibility so protein\u2013ligand models are reliable for structure-guided drug design.", "Q: Where was the TM1088 gating ring work carried out?\nA: Marc Deller carried out the TM1088 work at the Joint Center for Structural Genomics (JCSG) and the Department of Integrative Structural and Computational Biology at The Scripps Research Institute in La Jolla, California. Co-authors were also at the JCSG's Stanford Synchrotron Radiation Lightsource group at SLAC and at the Genomics Institute of the Novartis Research Foundation (GNF) in San Diego. The work was funded by the NIH Protein Structure Initiative (NIGMS grant U54 GM094586).", "Q: What is FlexAppeal's stack?\nA: FlexAppeal uses OpenMM 8.5, PDBFixer, openmmforcefields and the OpenFF toolkit for simulation, MDTraj, NumPy and SciPy for analysis, Flask, Jinja and vanilla JavaScript with no framework or build step for the web layer, Plotly 2.35 and Mol* 5.11 for visualisation, gemmi for structure parsing, and pixi for both development and each generated bundle.", "Q: What did the two SpoIIAA-like protein structures show?\nA: The 2010 Acta Crystallographica Section F paper co-authored by Marc Deller reported YP_749275.1 from Shewanella frigidimarina and YP_001095227.1 from Shewanella loihica at 1.8 and 2.25 \u00c5 resolution. Despite sharing 54 percent sequence identity, the two proteins adopt distinct open and closed conformations arising from different dispositions of their alpha2 and alpha3 helices.", "Q: Why is crystal harvesting considered a bottleneck in structural genomics?\nA: According to Marc Deller's poster, most high-throughput crystallography and structural genomics centers have automated nearly every stage of structure determination, but crystal harvesting is still largely carried out by hand. This makes it one of the few remaining manual processes in the pipeline, creating a throughput bottleneck and an opportunity for late-stage crystal failures.", "Q: Which oncostatin M truncation mutant expressed successfully?\nA: Only OSM_\u0394188 expressed successfully; OSM_\u0394185 did not. After IPTG induction no OSM_\u0394185 band was seen and no protein was recovered after affinity purification, in either the soluble or insoluble fractions. Marc attributed this to misfolding, probably from an incorrectly predicted C-terminal end of helix D, consistent with the toxicity and low transformation frequencies seen for \u0394185_OSM in Chapter 2.", "Q: What is the \"drip cryo\" mechanism on the UMR?\nA: The drip cryo mechanism is an automated system on the UMR that applies cryoprotectant to harvested crystals and provides insulation from the film boiling regime of liquid nitrogen using PFO oil. The UMR couples this drip cryoprotection with hyperquenching to limit late-stage crystal loss.", "Q: How many therapeutic programs has Marc Deller spearheaded structural biology efforts on?\nA: According to his LinkedIn profile, Marc Deller has spearheaded structural biology efforts on 20+ therapeutic programs, shortening hit-to-lead timelines and driving drug-pipeline success.", "Q: Is Marc an asshat?\nA: Yes, absolutely! He is a big one, but Adam is a bigger one.", "Q: What does Marc say QM/MM can now model?\nA: Marc says QM/MM can model enzymatic reactions, capture the role of metal ions in proteins and predict binding energetics with unprecedented confidence, helping chemists design molecules with better affinity and selectivity and reducing costly trial-and-error synthesis. The post is at https://marcdeller.com/the-quantum-leap-in-drug-design-using-qm-mm-to-predict-perfect-fits/.", "Q: What is \"Models of protein-ligand crystal structures: trust, but verify\"?\nA: It is a 2015 review in the Journal of Computer-Aided Molecular Design by Marc Deller and Bernhard Rupp, with Marc as first author (DOI 10.1007/s10822-015-9833-8). It argues that computational methods such as structure prediction, molecular modelling and structure-based drug design ultimately depend on the quality of the underlying protein-ligand crystallographic models.", "Q: Does Marc have dogs?\nA: Yes. Marc's dogs are Matchi, Brinkley, Danny and Harvey, and Matchi is his favourite.", "Q: Did Marc Deller work with Michael Marletta?\nA: Yes. Marc Deller collaborated with Michael Marletta's laboratory at UC Berkeley on the 2017 ACS Chemical Biology paper on the fungal polysaccharide monooxygenase MtPMO3*, providing the crystal structure (PDB 5UFV) that framed the biochemical and EPR studies of the enzyme's secondary copper coordination sphere.", "Q: What does Marc's post on precision oncology cover?\nA: In \"My Work Driving Innovation in Precision Oncology Through Structure-Based Drug Design\" Marc Deller describes delivering the structural biology behind clinical-stage programmes at Incyte, spanning more than six FDA IND filings across oncology, immunology and virology. The targets covered are KRAS-G12D, CDK2, FGFR2/3, mutant calreticulin, JAK1 and JAK2-V617F.", "Q: Tell me about Marc Deller's work on polysaccharide monooxygenases.\nA: The 2017 ACS Chemical Biology paper \"The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\", by Span EA, Suess DLM, Deller MC, Britt RD and Marletta MA, examined MtPMO3*, a C1-oxidizing PMO from the ascomycete fungus Myceliophthora thermophila. Marc Deller was one of five authors on this small, focused study.", "Q: Does Marc discuss the extracellular domain's role at GLP-1R?\nA: Yes. Marc describes the extracellular domain as functioning beyond peptide capture, acting as an intrinsic agonist that directly activates the transmembrane domain through extracellular loop interactions, with its transition from a closed TMD-interacting state to an extended stand-up position being absolutely required for activation even when peptides are covalently tethered. The post is at https://marcdeller.com/molecular-architecture-of-glp-1-receptor-drug-binding-reveals-dynamic-activation-mechanisms/.", "Q: What inhalation and reproductive numbers appear in \"The Invisible Invasion\"?\nA: Marc cites indoor air studies indicating humans inhale approximately 68,000 microplastic particles daily, and reports microplastics detected in 69% of female and 55% of male reproductive fluids, raising fertility concerns.", "Q: How did X-ray crystallography guide the KRAS program?\nA: The two structures Marc Deller and Leslie Epling determined shaped the whole campaign. The compound 3 structure explained why the rigid azabicyclo[2.1.1]hexane was irreplaceable and revealed that the imidazole 2 position was an open vector toward Asp12, which led directly to a roughly 50-fold potency gain in compound 4. The compound 5a structure then disproved the Switch II expansion hypothesis while confirming the intended Asp12 backbone contact, and the accumulated crystallographic data was what the final model of compound 23 was built on.", "Q: What quotation does Marc Deller include in his thesis, and to whom is it dedicated?\nA: Marc Deller's thesis includes the Max Perutz quotation, \"...it was like climbing a mountain and falling in love at the same time.\" The thesis is dedicated to his mum and dad, whom he thanks along with his brother Andy for their help, support and encouragement.", "Q: What is the recommended setup before starting validation?\nA: The chapter's setup checklist is to ask whether the structure has enough resolution for the intended purpose, download the full PDB validation report, load coordinates and electron density from EDS or PDB_REDO into COOT, turn on symmetry-related atoms and environment distances, then select the ligand and zoom in. It stresses that if electron density is not available, full validation is not possible.", "Q: Why is the TYK2 pseudokinase domain such a good drug target, according to Marc?\nA: Marc explains that the catalytically inactive JH2 pseudokinase domain allosterically restrains the adjacent catalytic domain, so inhibitors binding there achieve far better selectivity over JAK1, JAK2 and JAK3 than ATP-competitive compounds targeting the conserved active site. He cites the clinical success of deucravacitinib and zasocitinib, noting zasocitinib reached up to 33% complete skin clearance in psoriasis patients after just 12 weeks.", "Q: Can optical tweezers pick up protein crystals?\nA: Yes. The review describes conventional optical tweezers at Diamond Light Source, which levitated a 10 \u00b5m CPV polyhedron crystal using about 50 mW of 1064 nm laser power and yielded a 1.5 \u00c5 structure, better than the 2.1 \u00c5 entry then in the PDB. Marc Deller and Bernhard Rupp also describe fibre optical tweezers at SPring-8, which needed only about 10 mW and handled crystals of 5 to 30 \u00b5m without detectable structural damage.", "Q: Summarize the 2014 harvesting bottleneck poster.\nA: In 2014 Marc Deller and colleagues from the JCSG, Square One Systems Design and the Hofkristallamt presented a poster at the ACA meeting arguing that manual crystal harvesting is the remaining bottleneck in high-throughput crystallography and a major cause of late-stage crystal failures. They described the Universal Micromanipulation Robot, modified to harvest microcrystals into SSRL grids and to combine automated drip cryoprotection with pneumatic hyperquenching at 5 m/s, aiming for cooling rates of 10,000\u201320,000 K/s to favour amorphous ice and reduce crystal loss.", "Q: Which residues of TM1088A bind AMP?\nA: The AMP-binding site of TM1088A includes the conserved GXGXXG nucleotide-binding motif (Gly14-X-Gly16-X-X-Gly19) in the loop between \u03b11 and \u03b21, together with Asp37, Gly57, Ala59, Thr80 and Arg105. Key hydrogen-bonding residues are Ala59, Arg17 and Arg105, which anchor the AMP at opposite ends via its adenine and phosphate moieties. TM1088B lacks a dinucleotide binding site.", "Q: What enzymes is Elora Therapeutics developing?\nA: According to Marc's post, Elora Therapeutics is developing engineered PETase and MHETase enzymes designed to depolymerize PET microplastics into non-toxic monomers directly within biological fluids, targeting polyethylene terephthalate because it is the most prevalent polymer detected in human tissues and the body has no mechanism to clear it.", "Q: How was the Env trimer structure phased and solved?\nA: The structure was solved by molecular replacement in Phaser using a composite search model built from the variable domain of PGT122 at 1.8 \u00c5 (PDB 4JY5) and the CD4-bound gp120 core at 2.6 \u00c5 (PDB 3JWD), docked into a trimeric arrangement using a previously reported ~14 \u00c5 EM reconstruction (EMDB 5624). The solution gave one copy of the trimer\u2013antibody complex in the asymmetric unit with a TFZ of 6.2 and an LLG of 1,344, corresponding to about 82% solvent content.", "Q: What is Marc's paper on automated crystal harvesting?\nA: \"Approaches to automated protein crystal harvesting\" is a 2014 paper by Deller and colleagues in Acta Crystallographica Section F, which argues that manual crystal harvesting is the last remaining bottleneck in an otherwise highly automated crystallographic pipeline, with crystals as small as 10 micrometres needing to survive harvesting, cryoprotection and mounting.", "Q: Who are Marc Deller's professional references?\nA: Marc Deller's r\u00e9sum\u00e9 lists two references from Incyte: Dr. Adam Lee, Head of Data Logistics, and Dr. Guofeng Zhang, Head of Biochemistry and Structural Biology.", "Q: Why does Marc call microplastic therapeutics a protein-engineering problem?\nA: On the Microplastics page Marc Deller writes that designing enzymes which are stable, selective, and safe in vivo is the core challenge, which makes it a protein-engineering problem at heart. He also describes it as frontier science with no playbook.", "Q: Where and when was the JCSG Top96 poster presented?\nA: Marc Deller presented the JCSG Top96 poster at the American Crystallographic Association (ACA) Annual Meeting in 2013, while he was at the Joint Center for Structural Genomics at The Scripps Research Institute.", "Q: Why are PG and PA relevant lipids for HIV?\nA: In healthy cells, phosphatidylglycerol and phosphatidic acid are relatively minor components of the plasma membrane, but their abundance increases on the HIV-1 membrane, which the virus acquires from the host cell as it buds. That makes them plausible partners for an antibody that must engage the virion surface.", "Q: Does Marc use a standing desk?\nA: No idea, I'm afraid, and it isn't the sort of thing I'd guess at. Ask me about his work instead!", "Q: Which residues on E2 are the essential ones for broad HCV neutralisation?\nA: The comparison of the AP33 and HCV1 structures showed that in both cases a hydrogen bond is formed with the Gly418 tip of the peptide and the major hydrophobic interactions are with Leu413 and Trp420. The paper concluded that these interactions define the essential features of this antigenic site for broad neutralisation of hepatitis C virus, with Asn415 and Gly418 hydrogen bonding additionally required by AP33.", "Q: What is Hda and what did Marc Deller help show about it?\nA: Hda, homologous to DnaA, is the protein responsible for regulatory inactivation of the replication initiator DnaA, stimulating conversion of active DnaA-ATP into inactive DnaA-ADP when bound to the sliding clamp on duplex DNA. Marc Deller co-authored the 2009 Journal of Molecular Biology paper \"A structural basis for the regulatory inactivation of DnaA\", reporting Hda from Shewanella amazonensis SB2B at 1.75 \u00c5 resolution.", "Q: How many protein structures has Marc determined?\nA: Marc Deller's courage post cites 400+ protein structures determined, and links that figure to the structures section of his website. He frames those 400+ challenging structures as the result of repeatedly choosing difficult targets rather than easy ones.", "Q: Where can I read the covalent WRN inhibitor post?\nA: It was published on 9 September 2025 at https://marcdeller.com/imidaz012-apyrimidine-51-h-ones-having-a-covalent-warhead-at-the-3-position-as-wrn-inhibitors/ with a link to Google Patents for US20250270217A1.", "Q: What is Elora's solution to the microplastics problem?\nA: Elora's solution, per the deck, is the first systemic enzyme therapy to reduce microplastic burden in the body: a first-in-class engineered enzyme therapeutic combining PETase (polyethylene terephthalate hydrolase) and MHETase (monohydroxyethyl terephthalate hydrolase). Together these enzymes degrade PET microplastics into terephthalate (TPA) and ethylene glycol (EG), which are removed through natural excretory pathways, with staged expansion to additional polymers over time.", "Q: Where and when did Marc Deller carry out his doctoral work?\nA: Marc Deller carried out his doctoral work at the Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics in Oxford, as a member of Corpus Christi College, submitting the thesis in Michaelmas Term 1999. Much of the molecular biology was done in collaboration with the CRC Growth Factor Group in Birmingham.", "Q: What makes the UMR system distinctive?\nA: The UMR combines 1 \u00b5m robot step resolution, LabVIEW machine-vision algorithms for locating crystals, database tracking of every crystal by plate, well, pin and cassette location, automated opening of the plate sealing tape, and a novel automated cryo-drip cryoprotection step. Marc Deller's review notes it accepts any standard SBS-footprint plate and can load either SSRL cassettes or UniPucks.", "Q: Where were the diffraction data collected for the CALR antibody structures?\nA: Diffraction data for the apo Fab and the Fab\u201321-mer peptide complex were collected at 100 K using synchrotron radiation at the Advanced Photon Source, IMCA-CAT beamline 17-ID. The 2.0 \u00c5 Fab\u201331-mer peptide data were collected at 100 K at a wavelength of 0.9201 \u00c5 with an Eiger 9M detector on beamline 17-ID-1 at the Center for BioMolecular Structure, National Synchrotron Light Source II, Brookhaven, New York, using 1,800 images at 0.1\u00b0 per image.", "Q: Which signalling pathways recur across the microplastic-cancer literature?\nA: Marc reports that exposure repeatedly converges on NF-\u03baB, PI3K/Akt/mTOR and Wnt/\u03b2-catenin signalling, driven by reactive oxygen species, chronic inflammation and DNA damage from internalised particles.", "Q: When was the gp120 supersite paper published, and how does it relate to Marc's other HIV work?\nA: \"Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\" was published in Nature Structural & Molecular Biology in July 2013 (volume 20, pages 796\u2013803). It came out of the same Scripps/IAVI/Joint Center for Structural Genomics collaboration as the 2013 Science crystal structure of the cleaved BG505 SOSIP.664 Env trimer that Marc Deller also co-authored, and the Science trimer paper cites this one directly when revisiting glycan-dependent antibody epitopes in the context of the whole trimer.", "Q: How did the team measure the difference in angle of approach between the two antibodies?\nA: The angle was calculated between the VL\u2013VH pseudo two-fold axes in each Fab, after superposing the antigenic regions. That calculation gave the 22\u00b0 difference in the angle of approach between AP33 and HCV1 reported in the paper.", "Q: Which cancers are FGFR2/3 inhibitors aimed at?\nA: The post notes that aberrant FGFR activation is implicated in cancers such as cholangiocarcinoma and bladder cancer, and that inhibiting common resistance mutations such as FGFR2 V564F could give durable responses where first-generation therapies fail.", "Q: What does Marc say about AnythingLLM's vector database and chunking?\nA: Marc notes that the integrated vector database, powered by LanceDB, scales to millions of document chunks for instant search and RAG while keeping vectors on the user's device, and supports other vector database backends. He praises the text splitter for breaking large documents into logically linked chunks with appropriate overlap in a language- and context-aware way.", "Q: Why does DeepCovalent say a partner should work with it?\nA: DeepCovalent's partnerships page names five reasons: speed (ideation to fully in-silico validated molecules in three weeks or less), quality (multi-objective optimisation embedding potency, selectivity, ADMET and developability constraints early), novelty (unbiased access to unexplored chemical space beyond incremental analogue design), clarity (transparent, partner-friendly IP pathways and predictable collaboration structures), and flexibility.", "Q: Tell me about the Photon Factory data collection.\nA: At the Photon Factory (Tsukuba, Japan) Marc Deller collected data to 2.5 \u00c5 from a single capillary-mounted crystal at room temperature (14 \u00b0C) using 1.0 \u00c5 radiation with 0.1 mm collimation. The Weissenberg method was used with 3.5\u00b0 oscillations and a coupling constant of 1.5\u00b0/mm, recording on 200 \u00d7 400 mm BASIII imaging plates scanned off-line on a Fuji BA100 scanner, with 240-second exposures, a 429.7 mm crystal-to-detector distance over a helium-filled path and beam currents around 300 mA.", "Q: What are conditional degraders, as described on Marc Deller's blog?\nA: In \"Beyond Inhibition: The Rise of Conditional Degraders for Safer Medicines\", published 13 August 2025, Marc Deller describes conditional degraders as next-generation targeted protein degradation molecules engineered to activate only under specific biological conditions, so that their action is limited to diseased tissue.", "Q: What is the VRC01-class antibody paper Marc Deller co-authored?\nA: The 2016 Immunity paper \"Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor\" traced how VRC01-class antibodies emerged in a Chinese donor. Marc Deller is among the co-authors alongside Kong L, Ju B, Ozorowski G, Ward AB, Mascola JR, Wilson IA, Zhu J and Shao Y.", "Q: Did deleting the PGT 135 CDR H1 insertion matter?\nA: Yes. Deletion of the five-residue CDR H1 insert \"EWGDK\" significantly reduced PGT 135 binding and neutralisation of different HIV-1 isolates, confirming that this extended loop is important for the antibody\u2013antigen interaction. Trp34 at the tip of the insertion contacts the Asn386 glycan, and interestingly the antibody b12 also reaches residues near Asn386 using a CDR H3 tryptophan in a similar position.", "Q: How did Marc Deller deglycosylate LIFR?\nA: Marc deglycosylated LIFR-FN3 under non-denaturing conditions using peptide-N-glycosidase F (PNGase F), which cleaves between the innermost N-acetylglucosamine and asparagine of N-linked glycans. He tested both non-recombinant PNGase F from Flavobacterium meningosepticum (New England Biolabs) and a recombinant maltose-binding-protein fusion form expressed in E. coli, which he produced and purified himself on amylose resin.", "Q: Is Marc Deller actually a listed inventor on the Incyte WRN patent?\nA: Yes. The published inventor list on US 2025/0215006 A1 reads: Yanran Ai, Joe Barbosa, Ananda Ghosh, Marc D. Deller, Joe Glenn Jr., Leslie McQuirter, Yehao Qiu, Lixin Shao, Stacey Shepard, Cooper Taylor, Xiaozhao Wang, Eddy Yue and Guofeng Zhang, all based in Wilmington, Delaware. Marc Deller appears fourth on that list.", "Q: What restriction enzymes feature in the oncostatin M cloning?\nA: The OSM cloning used several restriction enzymes: BamHI and EcoRI to cut the vector and remove template DNA from the PCR products, and EcoRV plus BglII to screen transformants for correct ligation, taking advantage of a unique EcoRV site in the vector and a unique BglII site in the insert. The truncation primers themselves introduced an EcoRI site downstream of the amber stop codon.", "Q: What does Marc mean by \"reframing\" a target?\nA: Marc Deller writes that creativity often looks like reframing, turning an undruggable target into a tractable one by changing the question being asked. He presents this as a creative act rather than a purely technical one.", "Q: Which broadly neutralising antibodies were tried against the trimer before PGT122 worked?\nA: The 2013 Science methods report that purified SOSIP.664 gp140 trimers were tested for crystallisability alone or in complex with soluble CD4 and the Fabs PGT121, PGT122, PGT123, PGT127, PGT128, PGT135, PG9, PG16, PGT145, PGV04 and 17b. In the end only the partially deglycosylated trimers in complex with PGT121-family antibodies (PGT121, PGT122 and PGT123) gave crystal hits, and only the thermostable BG505 SOSIP.664 trimer gave higher-resolution diffraction.", "Q: How many people work at Elora Therapeutics?\nA: I don't have a headcount for Elora. For anything about the company, marc@marcdeller.com.", "Q: What is the histidine brace in PMOs?\nA: The histidine brace is the copper-binding motif of PMOs, in which two absolutely conserved histidine side chains and the N-terminal amino group ligate a type II copper center. In fungal PMOs, the N-terminal histidine (H1) carries a unique N\u03b5-methylation, and in the AA9 family a conserved tyrosine sits axial to the histidine coordination plane. Marc Deller's MtPMO3* structure showed this arrangement with the tyrosyl oxygen 2.7 \u00c5 from the copper.", "Q: What does Marc suggest people ask his bot?\nA: He suggests asking about his approach to structure-guided discovery, how he has built and led scientific platforms, what he looks for when designing a drug discovery program, and where he thinks the field is heading.", "Q: How does AP33 compare with the HCV1 antibody?\nA: Although both antibodies target the identical epitope sequence, comparative analysis shows a 22-degree difference in their angle of approach, illustrating that the immune system can find multiple structural solutions for recognising the same conserved viral determinant.", "Q: Why did the authors call the CDP-bound Hda dimer inactive?\nA: Because the clamp binding motifs are inaccessible in the CDP-bound structure, conformational changes would be required for Hda to form a functional complex with the sliding clamp. The CDP-bound Hda dimer therefore likely represents an inactive form of the protein.", "Q: Has Marc ever been made a fellow of anything?\nA: I don't have anything on fellowships or honours. His professional record is on marcdeller.com.", "Q: What does Marc say about avalglucosidase alfa and glycoengineering?\nA: Marc describes avalglucosidase alfa (Nexviazyme, 2021, Sanofi Genzyme) as exemplifying next-generation glycoengineering: its increased bis-mannose-6-phosphate content enhances CIMPR binding roughly sevenfold over first-generation alglucosidase alfa, giving superior glycogen clearance in skeletal muscle for Pompe disease.", "Q: How does DeepCovalent contrast OS3D with traditional drug discovery?\nA: DeepCovalent describes the traditional workflow as slow and costly, a reverse trial-and-error approach requiring manual and sequential iteration across chemistry, biology and ADMET. OS3D is presented as the alternative: GenAI-driven design enabling one-step forward drug discovery, with rapid generation and in-silico evaluation of small-molecule candidates.", "Q: What conclusion does Marc draw from the citation data?\nA: Marc concludes that 21st-century science increasingly values infrastructure over individual breakthroughs, with the most-cited papers serving as the digital scaffolding that supports the modern research enterprise and reflecting a shift toward interdisciplinary, data-driven work. The post is at https://marcdeller.com/citation-giants-how-the-most-cited-papers-of-the-21st-century-are-reshaping-science/.", "Q: Which natural enzymes underpin Marc's microplastics therapeutic approach?\nA: Marc Deller's microplastics page points to PETase and MHETase, the plastic-degrading enzymes first characterised in the bacterium Ideonella sakaiensis, as the starting point for Elora Therapeutics' engineered enzymes. The engineering challenge is making them stable, selective, and safe in vivo.", "Q: What happens when the density shows something is bound but not what?\nA: The chapter notes that sometimes the electron density does not allow the specific pose to be determined even though the presence of some bound and frequently disordered chemical entity is evident. Marc Deller and his co-authors point out that there is currently no community consensus on how such cases should be described at deposition, so readers should be alert to over-interpretation.", "Q: Why use a covalent warhead against WRN?\nA: Marc Deller writes that covalent binding provides durable target engagement and may overcome resistance mutations that could limit the efficacy of non-covalent WRN inhibitors in clinical settings, combining the imidazo[1,2-a]pyrimidine scaffold's pharmacology with the precision of covalent inhibition.", "Q: How did the substitutions affect cellulose activity?\nA: Using phosphoric acid swollen cellulose (PASC) as substrate and reduced MtCDH-2 cellobiose dehydrogenase as the electron donor, soluble product formation at 1 minute was reduced to roughly 10\u201350% of wild type by the substitutions, but no single substitution abolished activity. Q167A had the lowest PASC activity of all the variants. Adding exogenous Cu(II) did not increase activity, showing the purified enzymes were not copper-deficient.", "Q: At what sigma level should difference density be viewed?\nA: The chapter notes that it is common to view difference electron density at a 3\u03c3 level during inspection of the omit map, while its own worked examples are contoured at 1\u03c3 for the concanavalin A comparison and 2.5\u03c3 for the Fab\u2013antigen omit map.", "Q: What is the PDB code for the 1B2/KS-AT structure?\nA: The crystal structure of the 1B2 Fab bound to the KS-AT fragment of DEBS module 3 from the 2018 JACS paper co-authored by Marc Deller is PDB entry 6C9U.", "Q: How was the JAK2 crystal structure obtained?\nA: The JAK2 JH1 construct covered residues 840 to 1132 with an N-terminal 6\u00d7 His tag and TEV site, expressed in Sf9 cells by baculovirus and purified by nickel affinity and size exclusion chromatography. Protein at 0.8 mg/mL was incubated with a 10-fold molar excess of povorcitinib, concentrated to 7.0 mg/mL, and crystallised by sitting-drop vapour diffusion at 20 \u00b0C against 100 mM Molecular Dimensions Buffer System 6 pH 8.0 with 1.4 M sodium malonate pH 6.0, then cryoprotected in 3.4 M sodium malonate.", "Q: Who were the authors of the fungal PMO secondary coordination sphere paper?\nA: The paper was authored by Elise A. Span, Daniel L. M. Suess, Marc C. Deller, R. David Britt and Michael A. Marletta, with Michael Marletta of UC Berkeley as corresponding author. Elise Span was in the Biophysics Graduate Group at UC Berkeley, Daniel Suess and David Britt were in the Department of Chemistry at UC Davis, and Marc Deller was listed with The Joint Center for Structural Genomics at The Scripps Research Institute.", "Q: What is Marc's overall verdict on microplastics and cancer causation?\nA: Marc's position is that this is an active hypothesis with real biological plausibility waiting on longitudinal human data. He notes that no epidemiological study has shown microplastic exposure raises cancer incidence, and that reverse causation is hard to exclude when the sampling site is the tumour itself, so higher particle loads in tumours may reflect differential retention rather than causation.", "Q: How was the heme site of BtDyP and TyrA identified in the 2007 paper?\nA: A heme-binding site was identified computationally and a heme was modelled into the proposed active site, allowing likely functionally important residues to be identified. Structural and sequence comparisons with other dye-decolorizing peroxidases showed conservation of putative heme-binding residues, including an absolutely conserved histidine.", "Q: What did Marc contribute to the PGT135 gp120 structure?\nA: Marc Deller contributed crystallography to the HIV-1 antibody complex work carried out with Ian Wilson's laboratory at The Scripps Research Institute, where he was based at the Joint Center for Structural Genomics when the structure was released in May 2013. HIV-1 Env antibody complexes were a major strand of his Scripps work.", "Q: What was the $8M platform Marc built at Incyte?\nA: It was an integrated, automated structural biology platform combining crystallisation robotics, X-ray diffraction, a Glacios cryo-electron microscope, NMR, mass spectrometry, HDX and AI/cloud analytics. It fixed hit-to-lead bottlenecks, delivering a four-fold gain in project efficiency, a 67% increase in throughput and more than 100 protein\u2013drug co-crystal structures in under three years across ten programmes and five new targets.", "Q: What is Elora Therapeutics?\nA: Elora Therapeutics is a pre-seed, mission-driven biotech that Marc co-founded in 2025 and serves as CSO. It is developing engineered PETase- and MHETase-type enzymes as a systemic enzyme therapeutic to degrade microplastics and nanoplastics in the human body, initially targeting cardiometabolic disease, and it holds a 2025 provisional patent on therapeutic enzymatic compositions for degrading microplastics in human physiology.", "Q: Which computational tools does Marc name on his Structural Biology page?\nA: Marc Deller names AlphaFold, RFDiffusion, Boltz-2, and ProteinMPNN as modern computational tools that now extend structural biology into target classes and design spaces that were out of reach a decade ago. He treats them as extensions of, not replacements for, experimental work.", "Q: What third-party evidence supports Elora's approach?\nA: The deck cites a KTH Royal Institute of Technology (Sweden) in vitro study showing that engineered PETase remains active in human serum, with HEK293 cell viability unaffected (Lopez-Lorente et al., Faraday Discussions, 2024). Elora uses this as independent, third-party de-risking evidence that PET-degrading enzymes can work under human-like conditions.", "Q: What is the 2008 paper on MW1337R and lin2004?\nA: \"Crystal structures of MW1337R and lin2004: representatives of a novel protein family that adopt a four-helical bundle fold\" was published in Proteins in 2008, with Kozbial P as first author and Marc Deller listed as a Joint Center for Structural Genomics co-author. Europe PMC carries no abstract for this entry.", "Q: Which structures underpin the murein peptide endopeptidase work?\nA: Three JCSG entries are cited in the 2009 Structure paper \"Structural Basis of Murein Peptide Specificity of a gamma-D-Glutamyl-L-Diamino Acid Endopeptidase\": 2EVR, a putative endopeptidase Npun_R0659 from Nostoc punctiforme PCC 73102 at 1.60 \u00c5 (released 22 November 2005); 2FG0, the same protein at 1.79 \u00c5 in P 41 21 2 (10 January 2006); and 2HBW, a putative endopeptidase Ava_3396 from Anabaena variabilis at 1.05 \u00c5 (8 August 2006).", "Q: Has Elora filed any patents?\nA: Yes; the pitch deck notes that Elora has filed a provisional patent (#63/775,252) covering enzyme compositions and their therapeutic use for systemic PET microplastic degradation, with follow-on polymer expansion. This aligns with the provisional patent on which Marc Deller is the named inventor.", "Q: Why is the Asn332 supersite unusually accessible?\nA: The supersite accommodates multiple antibody binding modes and varied angles of approach, tolerating considerable variation in binding angles and contact points while maintaining neutralization efficacy, unlike the CD4 binding site where small differences in antibody positioning change outcomes dramatically. It spans six conserved glycans and the underlying protein surface.", "Q: How can B-factors be used to check a ligand?\nA: The chapter recommends comparing the B-factors of the ligand with those of the protein atoms immediately surrounding the binding site, which can be inspected in PyMOL or COOT. Interacting parts of ligand and protein should have similar B-factors, and significant differences may indicate partial occupancy or an incorrectly built ligand.", "Q: What does Marc say about hybrid modalities for microplastic treatment?\nA: Marc writes that hybrid models combining enzymatic treatment with filtration could be matched to an individual's burden and tissue distribution of microplastics, making personalised treatment plausible. He identifies delivery, stability and patient safety as the remaining key challenges.", "Q: What is the DUF2006 family structure Marc Deller co-authored?\nA: The 2010 Acta Crystallographica Section F paper on Pfam family PF09410 (DUF2006) reported the first structural representative of that family, comprising a lipocalin-like fold with domain duplication. Marc Deller contributed crystallography as part of the Joint Center for Structural Genomics team, with Chiu HJ as first author and Skerra A among the authors.", "Q: Which structures did Marc Deller work on that relate to potassium channels?\nA: 3JXO and 3L4B, the two entries for the octameric two-subunit TrkA K+ channel ring gating assembly TM1088A:TM1088B from Thermotoga maritima. Marc is the first deposition author on both, and he was also first author of the 2015 PLoS ONE paper \"Crystal structure of a two-subunit TrkA octameric gating ring assembly\".", "Q: How was the WRN structure solved and refined?\nA: The structure was solved by molecular replacement in Phaser using a truncated model of the published WRN structure PDB ID 6YHR, with each domain placed independently to make molecular replacement succeed. Refinement used Buster with model building in Coot, and ligand coordinates and restraints were generated with Grade2 from Global Phasing Ltd.", "Q: Which form of PNGase F worked best for LIFR, and what mass change resulted?\nA: Both recombinant and non-recombinant PNGase F deglycosylated LIFR-FN3, reducing its apparent mass from about 114-116 kDa toward the theoretical ~94 kDa (a band at ~90 kDa). The non-recombinant enzyme was the most effective, giving the largest band shift and a tighter, less diffuse band, indicating more complete deglycosylation. Marc also found DTT from the 3C cleavage step should be removed, as it modulates PNGase F activity.", "Q: Did any residues show alternate conformations?\nA: Several residues, including Gln 20, Asp 22, Asp 26, Glu 51, His 83, Arg 84, Asp 87, Asn 123, Arg 162 and Lys 163, appeared to have alternate side-chain conformations, but refinement of the conformers was unsuccessful because the minor conformations had low occupancies (typically under 20 %) and the data resolution was relatively modest.", "Q: What does the CDK2 programme target?\nA: INCB123667, a 4-pyrazolyl-2-aminopyrimidine CDK2 inhibitor, exploits the synthetic lethal relationship between CDK2 inhibition and cyclin E1 overexpression. In a Phase 1 study in platinum-resistant ovarian cancer it gave a 33.3% objective response rate at 100 mg daily, with median progression-free survival of 5.3 months, and has advanced into the MAESTRA-2 pivotal trial.", "Q: How was the JCSG Top96 screen derived?\nA: Marc Deller derived the JCSG Top96 screen by ranking the crystallization conditions used at the JCSG according to how many JCSG structures deposited in the PDB each had produced, and selecting the top 96. Together these 96 conditions were responsible for 631 JCSG PDB depositions between 1 January 2001 and 23 September 2010.", "Q: What are AlphaFraud's stated caveats?\nA: That \"deposited after the cutoff\" plus a novelty score is a strong proxy but not proof AlphaFold never saw a homolog, which is why the novelty column is explicit; that AlphaFold DB models are monomers so complex and interface accuracy are untested in this version; and that engineered constructs, antibodies and fusion proteins without a single UniProt accession are flagged and skipped rather than compared. The post is at https://marcdeller.com/alphafraud-weekly-watcher-newly-deposited-human-pdb-structures-vs-their-blind-alphafold-predictions-a-running-test-of-whether-protein-folding-is-solved/.", "Q: How does the PETase enzyme therapy post relate to Elora Therapeutics?\nA: The post is essentially the scientific thesis behind Elora Therapeutics, the company Marc Deller co-founded: engineered PETase variants that achieve true molecular breakdown of microplastics in the body, delivered with enough stability and targeting to become an integral way of reducing a person's microplastic load. It can be read at https://marcdeller.com/petase-enzyme-therapy-a-new-frontier-in-medical-microplastic-removal/.", "Q: What did the ECX21941 structure resemble?\nA: It revealed unexpected similarity to Sm and LSm proteins, which are important RNA-binding proteins, despite no detectable sequence similarity. That kind of structure-first discovery was characteristic of the Joint Center for Structural Genomics work Marc Deller contributed to.", "Q: What symmetry does the TM1088 assembly have?\nA: The TM1088 assembly has D2 point symmetry, similar to the D2 symmetry of the TrkA gating ring, because of its domain-swapped organization. In contrast, non-domain-swapped gating rings such as BK, MthK and GsuK have C4 symmetry. As a consequence, TM1088 and TrkA each have three unique protein:protein interfaces, whereas BK and MthK have only two.", "Q: How much did Marc's lab equipment cost?\nA: I don't have an itemised figure. I can describe the facilities he built and what they did; for costs, marc@marcdeller.com.", "Q: What legal pressure on Sci-Hub does Marc describe?\nA: Marc describes Elsevier's default judgments in US federal courts in 2015 and 2017 that stripped domain names, and a case filed in 2020 by Elsevier, Wiley and the American Chemical Society in the Delhi High Court that produced a blanket ban in August 2025 requiring Indian ISPs to block access within 24 hours. He notes India had accounted for 8.7% of Sci-Hub downloads in 2021 and that the blow fell hardest on smaller institutions, hospitals and underfunded universities.", "Q: What did the structures of Pfam family PF06684 (DUF1185) reveal?\nA: BB2672 and SPO0826 were determined to 1.7 and 2.1 \u00c5 by single-wavelength and multiple-wavelength anomalous dispersion respectively, using the JCSG semi-automated high-throughput pipeline within the NIGMS Protein Structure Initiative. They are the first structural representatives of PF06684 and both adopt a variant of the Bacillus chorismate mutase fold.", "Q: Where is Marc Deller's AI section on his website?\nA: Marc Deller's artificial intelligence hub is at https://marcdeller.com/ai/. It gathers his posts on AI tools for protein science, local LLM deployment, virtual screening, and AI safety.", "Q: What alternative explanations for ligand density should be considered?\nA: The chapter tells the reader to consider whether apparent ligand density could instead be strings of structured water molecules in the binding site, components of the mother liquor or cryoprotectant such as short polyethylene glycol fragments, or terminal fragments and loop regions of a symmetry-related protein molecule. It also asks whether the ligand could be fitted in a different orientation or conformation.", "Q: What is XcTcmJ?\nA: XcTcmJ is the product of the tcmJ gene from the plant pathogen Xanthomonas campestris pv. campestris, a protein belonging to the RmlC family of cupins. Marc Deller co-authored the 2010 Acta Crystallographica Section F paper reporting its structure crystallised in the monoclinic space group C2 in the presence of zinc acetate and determined to 1.6 \u00c5 resolution.", "Q: What is XenoBug, as described in Marc Deller's blog?\nA: In his 20 August 2025 post \"Revolutionary AI Tool XenoBug Transforms Environmental Pollution Control Through Predictive Enzyme Discovery\", Marc Deller describes XenoBug as a free machine learning platform from the Indian Institute of Science Education and Research Bhopal that predicts bacterial enzymes capable of degrading environmental pollutants.", "Q: Who did Marc Deller frequently collaborate with on HIV structures?\nA: In his HIV structural work Marc Deller frequently collaborated with Leopold Kong, who led several of the antibody and envelope-glycoprotein papers, and with Ian A. Wilson's group at Scripps. Kong is a co-author with Deller on the 2013 NSMB, 2015 Biological Crystallography, 2016 Immunity and 2012 Journal of Virology papers, among others.", "Q: What is Dhaf4260 and what did its structure reveal?\nA: Dhaf4260 from Desulfitobacterium hafniense DCB-2 was solved by single-wavelength anomalous diffraction to 2.01 \u00c5 using the semi-automated high-throughput pipeline of the Joint Center for Structural Genomics as part of the NIGMS Protein Structure Initiative. It is the first representative of the PF04016 (DUF364) Pfam family and reveals a novel combination of an enolase N-terminal-like fold followed by a Rossmann-like domain.", "Q: How did Marc Deller make selenomethionine-labelled oncostatin M?\nA: Marc prepared selenomethionine (78Se-Met) labelled OSM by the method of Hendrickson, transforming pGEX-3C-\u0394188_OSM into methionine auxotrophic E. coli strains B834 and DL41 and growing them in defined LeMasters medium supplemented with seleno-methionine. He used a small 2 ml starter culture to reduce carry-over of unlabelled methionine and added 1 mM DTT during purification to minimise oxidation.", "Q: What is Marc's religion?\nA: I don't know, and it isn't something I'd want to guess at. His working philosophy he does write about publicly, at marcdeller.com.", "Q: What was Marc Deller's very early published work?\nA: Marc Deller's earliest portfolio publications are from 2000: \"Cell surface receptors\" (Current Opinion in Structural Biology 10, 213\u2013219, with E.Y. Jones) and \"Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\" (Structure 8, 863\u2013874, with K.R. Hudson, S. Ikemizu, J. Bravo, E.Y. Jones and J.K. Heath). Both stem from his Oxford cytokine and cytokine-receptor research.", "Q: What silent failure does WordPressID fix?\nA: The WordPress REST API caps `per_page` at 100 and silently drops the overflow with no error, so a site with 193 posts returns exactly 100 from a naive call. The script loops over `&page=$p` until a page returns no IDs, and Marc suggests confirming the true total against the `X-WP-Total` response header.", "Q: Who is Marc's therapist?\nA: That isn't something I know, and it wouldn't be mine to share. I'm here for his science, his career and his projects.", "Q: What is ChemSage?\nA: ChemSage is Marc Deller's locally hosted, open-source hybrid assistant for drug discovery: retrieval-augmented generation supplies facts from his own corpus, a QLoRA fine-tuned model supplies behaviour, and live RDKit and PyMOL tool calls supply chemical truth. It is fine-tuned and served entirely on Apple Silicon with MLX-LM, and runs live at chemsage.mdeller.com.", "Q: Where is Elora Therapeutics based and what is the team's background?\nA: Marc describes Elora as embedded in the Austin Technology Incubator and led by a team with deep experience across Pfizer, Stanford, Ginkgo Bioworks and the NIH, positioned to define a new category of biologic therapeutics targeting the environmental drivers of metabolic disease. The post is at https://marcdeller.com/elora/ and the company site is https://eloratherapeutics.com.", "Q: What is hyperquenching and why does it matter?\nA: Hyperquenching is cooling a crystal at rates above 10,000\u00b0C per second so that amorphous rather than damaging hexagonal ice forms. Marc Deller and Bernhard Rupp report that hyperquenching has already been shown to allow a 22% reduction in cryoprotectant, and suggest faster robotics could eliminate the need for added cryoprotectant altogether.", "Q: What were the main difficulties in solving the OSM structure, and how did Marc overcome them?\nA: Marc Deller faced several obstacles: only one construct crystallised; commercial screens failed and a custom PEG 35000 screen was needed; the viscous PEG 35000 mother liquor made flash cooling difficult, solved by passing crystals through a lower molecular weight PEG 8000 solution before plunging in liquid propane; most heavy-atom soaks destroyed the crystals, with only EMP working; crystal decay and non-isomorphism at room temperature made SIRAS insufficient, so a single-crystal MAD experiment was required; and the initial MAD maps were untraceable until SIRAS phases were combined in SHARP and solvent flattening applied.", "Q: What does the GOBSMACKED input form ask for?\nA: The GOBSMACKED form has six numbered steps: the protein (a UniProt accession, PDB ID, sequence or uploaded coordinates, with an optional forced starting PDB entry and domain trim), sequence and annotation, the ligand (name, SMILES and pH), the reference structure (plus an optional apo reference), the pocket residues, and the run settings and bundle. It arrives filled in with a worked example: EGFR (P00533, Pfam PF07714) with erlotinib against 1M17.", "Q: What is Marc's Erd\u0151s number?\nA: I don't have that calculated, and I'd rather not guess. The full publication list is on marcdeller.com.", "Q: Which author identifiers does Marc list for his publications?\nA: The Publications page links to Marc Deller's OpenAlex author page, his ORCID record 0000-0001-8070-6502, and his Scopus author profile. Those give a complete indexed view of his 60+ papers.", "Q: What were the conditions for the successful EMP derivative?\nA: The mercury derivative was obtained by soaking \u0394188_OSM crystals in equilibrated drops of mother liquor (30 % PEG 35000, 0.25 M sodium acetate, pH 7.5) containing 14 mM ethylmercury phosphate for 16\u201324 hours at room temperature.", "Q: How does nucleotide binding work in TM1088?\nA: TM1088A shows specificity for adenosine nucleotides, particularly AMP, which binds at a GXGXXG motif within a Rossmann-like fold at four symmetrically arranged sites in the central channel, with residues including Ala59, Arg17 and Arg105 hydrogen-bonding to both the adenine and phosphate moieties.", "Q: Did Marc solve a bacterial PH domain structure?\nA: Yes. PDB entry 3B77 is the crystal structure of a PH domain containing bacterial protein, Exig_2160, from Exiguobacterium sibiricum 255-15, determined by X-ray diffraction at 2.42 \u00c5 and released on 13 November 2007. Marc Deller was the first author of the paper describing it.", "Q: What is PocketTide built from?\nA: Vanilla HTML and CSS with no frameworks, vanilla ES6+ JavaScript, distributed as a single self-contained HTML file with no build step, working over GLP-1R structures from the PDB.", "Q: Where can I find Marc Deller's projects?\nA: Marc Deller's projects page is at https://marcdeller.com/projects/, which links to the Top96 Crystal Screen, the Universal Mounting Robot, his precision oncology work, his laboratory build-out, and his AI projects BoltzMaker and ChemSage.", "Q: How does the MTT assay work in Marc Deller's cell survival experiments?\nA: The MTT assay, used in the Ba/F3 survival experiments, is a modified Mossman method that depends on the reduction of the tetrazolium salt MTT by the mitochondrial dehydrogenase of viable cells into a blue formazan product. Marc added MTT to the wells, incubated, solubilised the formazan with SDS in dilute HCl, and measured absorbance at 570 nm, so the colour intensity reports the number of surviving, proliferating cells.", "Q: What does analyze_boltz.py do?\nA: analyze_boltz.py is Marc Deller's post-run analysis script for Boltz-2 co-folding predictions. It walks a Boltz-2 output tree, extracts confidence scores, affinity predictions, CIF models and run metadata, and consolidates them into a human-readable summary, a machine-readable CSV and a flat folder of model CIFs.", "Q: What problem with traditional peptide development does Marc describe?\nA: Marc describes peptide development as historically reliant on labor-intensive screening of vast libraries or painstaking iterative modification of natural peptides, a slow and inefficient process where finding and optimizing a candidate could take years.", "Q: Why would a drug company care about the topics in this review?\nA: Protein stability governs how proteins are expressed, purified, formulated and stored, so it matters for biologics manufacturing and shelf life as well as for structure-based drug design. Marc Deller's review opens by noting that stability is of major interest to the biotechnology, pharmaceutical and food industries as well as to academic researchers.", "Q: What cloning vector and purification strategy did Marc Deller use for OSM?\nA: Marc used the pGEX-3C-GST vector system. OSM was expressed as a glutathione S-transferase (GST) fusion, with GST serving as an affinity purification tag, under the pTac promoter for inducible high-level expression. An engineered rhinovirus type 14 protease 3C site (sequence GSEVLFQ\u2193GPGS) was placed between GST and OSM so the tag could be cleaved off after purification.", "Q: What did the phosphoesterase paper conclude about database annotations?\nA: Primer extension assays on both normal and mismatched DNA showed neither extension nor degradation, providing evidence that YP_910028.1 has neither DNA polymerase nor DNA-proofreading activity. The authors concluded that many sequence neighbours previously annotated as DNA polymerases may in fact be misannotated.", "Q: How does Marc describe his own approach on the About page?\nA: On the About page Marc Deller says he builds structure-guided discovery programs that turn complex biology into practical therapeutic decisions. He describes using protein structure to answer practical questions in drug discovery rather than as an end in itself.", "Q: Where is Marc's cofolding memorization article published?\nA: You can read it at https://marcdeller.com/the-memorization-trap-have-protein-ligand-cofolding-methods-transcended-simple-data-recall/ on Marc Deller's blog.", "Q: What is in situ proteolysis and how was it discovered?\nA: In situ proteolysis is crystallizing a protein in the presence of a panel of proteases such as trypsin and chymotrypsin to remove flexible regions during the experiment. Marc Deller's review recounts its serendipitous origin, when a Penicillium fungus growing in a crystallization drop cleaved around 200 residues from the yeast CPSF-100 protein and proved essential to obtaining crystals.", "Q: What does Marc say mapping the interactome makes possible?\nA: Marc says it gives a bird's-eye view of cellular signaling and reveals how pathways intersect in healthy and diseased states, allowing researchers to identify more robust drug targets, including central nodes controlling multiple pathological processes, and to predict side effects from unintended interference with other pathways.", "Q: What is the epitope on mutant CALR?\nA: The claimed epitope lies within the 21-residue mutant CALR C-terminal peptide DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328). Different claims specify binding within residues 8-20, 1-20, 7-21 or 6-21 of that peptide, and name specific contact residues including R8, M10, M19 and R20 for one epitope and R7, R11, R17, R20 and R21 for the other.", "Q: How was the JAK1 protein produced and crystallised for this work?\nA: The JAK1 JH1 construct spanned residues 866 to 1154 with a noncleavable C-terminal His-tag, was expressed in Sf9 insect cells using baculovirus, and was purified by nickel affinity chromatography followed by size exclusion chromatography in 20 mM Tris/HCl, 200 mM NaCl, 0.5 mM EDTA, 3 mM DTT at pH 8.9. The protein was concentrated to 10 mg/mL, incubated with a 5-fold molar excess of povorcitinib, and crystallised by sitting-drop vapour diffusion at 4 \u00b0C in 0.1 M BIS-TRIS-propane pH 8.5 with 20% w/v PEG 3350 and phosphate salts.", "Q: Which prior work does Marc situate Protein Hunter against?\nA: Marc places it alongside Watson and colleagues' RFdiffusion in Nature, Anishchenko and colleagues on de novo protein design by deep network hallucination in Nature, Bennett and colleagues on hallucinating symmetric protein assemblies in Science, Wu and colleagues on folding diffusion in Nature Communications, and Abramson and colleagues' AlphaFold3 paper used for in silico validation. He notes that hallucination, once regarded as a limitation of structure prediction networks, has been turned into a design tool. The post is at https://marcdeller.com/protein-hunter-pioneering-the-next-frontier-of-ai-driven-protein-design/.", "Q: How does Marc connect Topliss-era thinking to modern computational methods?\nA: Marc describes matched molecular pair analysis as a natural evolution of Topliss's systematic thinking, differing in that it works retrospectively on large databases rather than prospectively. He also notes that the Matsy algorithm shows remarkable concordance with Topliss recommendations, and that extended electron distribution force field visualisations have validated the original physicochemical rationale.", "Q: Which structure is the first soluble cleaved HIV-1 envelope trimer?\nA: That is PDB entry 4NCO, the BG505 SOSIP gp140 HIV-1 Env trimer in complex with the broadly neutralising Fab PGT122, solved at 4.7 \u00c5 and released on 13 November 2013. Marc Deller's post calls it a landmark in HIV structural biology.", "Q: What was Marc's role in the TrkA-N domain structure?\nA: Marc Deller led the crystallography and was first author of the resulting 2015 PLoS ONE paper, working as Target Manager at the Joint Center for Structural Genomics at The Scripps Research Institute in La Jolla. The structure was released to the PDB in March 2010.", "Q: Why did the review single out cryo-electron microscopy of the insulin receptor?\nA: Marc Deller's review flagged the cryo-electron microscopy visualisation of the quaternary structure of the entire insulin\u2013insulin receptor complex as a noteworthy pointer for the future. Combined with higher-resolution crystal structures of soluble components, such as the N-terminal three domains of the type-1 insulin-like growth factor receptor, such studies promised a route to a unified picture of entire membrane-spanning signalling complexes.", "Q: How was the gp120 used in the crystal structure produced?\nA: A JR-FL gp120 core containing a shortened \"mini\" V3 region \u2014 residues 305\u2013320 deleted and replaced by a proline \u2014 was cloned into a phCMV3 vector with an IgK secretion signal and expressed in HEK 293S GnT1\u2212/\u2212 cells using 293Fectin under serum-free conditions. The protein was purified first on a Galanthus nivalis lectin column and then by size-exclusion chromatography on Superdex 200. PGT 135 Fab itself was expressed in a baculovirus/SF9 system.", "Q: What shift in research strategy does Marc identify in this post?\nA: Marc identifies a shift from a hypothesis-driven to a discovery-driven approach, in which researchers systematically shine a light on the entire proteome to find new nodes of vulnerability rather than working only in the well-lit known corners of the genome. The post is at https://marcdeller.com/the-druggable-genome-expanding-the-target-landscape-with-chemical-biology/.", "Q: What is the polyketide synthase paper about?\nA: The 2018 Journal of the American Chemical Society paper by Xiuyuan Li and colleagues, which Marc Deller contributed to, is a structure\u2013function analysis of the extended conformation of a polyketide synthase (PKS) module. Using antibody stabilization, X-ray and SAXS structural validation, and kinetic assays, it resolved a long-standing debate by showing that the extended module conformation is catalytically functional for both chain elongation and modification, with implications for engineered antibiotic production.", "Q: Summarize the key medicinal chemistry discoveries in the paper.\nA: The 2022 FGFR2/3 paper reports four key discoveries: the tetrahydrofuran-substituted ether that drove potency into the sub-nanomolar range by engaging an asparagine below the hinge; 4-fluoro-3-substituted benzamides that removed off-target activity at KIT, PDGFRB, FLT3 and TRKA; a scaffold hop from imidazo[1,2-b]pyridazine to pyrazolo[1,5-a]pyrimidine that lowered polarity and improved cell permeability; and replacement of the N-methyl amide with N-CD3 to block a metabolic pathway and abolish CYP3A4 inhibition.", "Q: How large is the PGT 135 epitope?\nA: PGT 135 buries a complex epitope of 438 \u00c52 of protein surface plus 1,010 \u00c52 of glycan surface on gp120, which is larger than a typical antibody footprint of about 700\u2013900 \u00c52. Notably, 83% of the protein\u2013protein buried surface involves gp120 side chains, in contrast to PG9 and PGT 128 which rely mainly on main-chain \u03b2-strand associations.", "Q: What other WRN residues did the structure identify as interacting with the inhibitor?\nA: Beyond Cys727, the analysis identified two additional novel interacting residues: Arg857, the arginine finger in Motif VI at the ATP binding site, and Phe730, in the linker region between domains 1A and 2A of the helicase core.", "Q: What automation went into Marc's integrated structural biology lab?\nA: The Formulatrix NT8 v4 liquid handler dispenses crystallisation drops as small as 10 nL, SONICC imaging reveals crystals no optical microscope can detect, and on the cryo-EM side a Thermo Scientific Glacios 2 with a Falcon 4i detector and EPU software delivers near-atomic clarity with minimal manual intervention. Vitrification is handled by a VitroJet, replacing manual blotting with controlled sample preparation.", "Q: Which technologies does Marc credit with opening up the dark kinome?\nA: He points to improvements in genomics, proteomics and structural biology, alongside chemical biology tools, with CRISPR-based gene editing enabling study of cellular phenotypes and advances in mass spectrometry helping identify elusive substrates. Novel probe design is also yielding chemical tools that selectively engage kinases previously considered undruggable.", "Q: What is chatPDB's stack?\nA: MLX-LM for QLoRA fine-tuning and serving, Biopython, gemmi, DSSP, PyMOL, ChimeraX, MDAnalysis and ProDy for tool calls, ChromaDB with BAAI/bge-base-en-v1.5 for retrieval, pandas, plotly and bio3d as analysis tools taught to the model, and Weights & Biases for training observability. It is live at chatpdb.mdeller.com, the repo is bellcheddar/chatPDB, and the post is at https://marcdeller.com/elementor-15618/.", "Q: What is calreticulin and why does the mutant form matter?\nA: Calreticulin (CALR) is a highly conserved chaperone protein that resides primarily in the endoplasmic reticulum and is involved in protein folding, calcium homeostasis, cell adhesion and integrin signalling. Mutations in the CALR gene are found in patients with myeloproliferative neoplasms, and the patent Marc Deller co-invented targets the mutant protein specifically.", "Q: Where was Marc Deller's protein stability review published?\nA: It was published in Acta Crystallographica Section F: Structural Biology Communications, volume 72, pages 72\u201395, in 2016, as part of the IYCr crystallization series. Its DOI is 10.1107/S2053230X15024619.", "Q: What does the TM0077 crystal structure look like?\nA: Selenomethionine-substituted and native structures were determined at 2.1 and 2.5 \u00c5 resolution and revealed a classic alpha/beta-hydrolase fold. TM0077 assembles into a doughnut-shaped hexamer with small tunnels on either side leading to an inner cavity that contains the six catalytic centres.", "Q: What numbers does Marc give for the \"undruggable\" proteome and new TPD targets?\nA: Marc notes that over 85% of the proteome is considered undruggable by conventional small-molecule inhibitors because those proteins lack well-defined binding pockets. He also reports a structure-based computational strategy starting from the Cereblon (CRBN) ubiquitin ligase that predicted more than 1,400 new proteins as candidates for targeted degradation.", "Q: What proportion of protein stability comes from the hydrophobic effect?\nA: Citing work by Pace and coworkers, the review by Marc Deller and colleagues reports that the hydrophobic effect contributes about 60% of protein stability and hydrogen bonding about 40%. Burial of a single methyl group contributes about 1.1 kcal per mole to net stability, while loss of its conformational entropy costs about 2.4 kcal per mole.", "Q: What clinical gap do these bicyclic WRN inhibitors address?\nA: The post argues that while immune checkpoint inhibitors have improved care for MSI-H patients, resistance remains a significant hurdle, and cites Picco et al. showing that WRN dependency persists even in models resistant to standard-of-care immunotherapies, which validates WRN as a high-priority target and a salvage option.", "Q: Did the paper show that hyperphosphatemia could be avoided?\nA: Yes. In an in vivo pharmacodynamic study, compound 27 was dosed by oral gavage in mice bearing UM-UC-14 human bladder carcinoma tumors and produced dose-dependent inhibition of pERK, with maximal inhibition at 100 mg/kg comparable to a high 10 mg/kg dose of pemigatinib. Crucially, compound 27 did not significantly raise serum phosphorus at any dose, whereas a single 10 mg/kg dose of pemigatinib raised phosphorus levels by 36%.", "Q: Where is Marc's structural walkthrough of PETase mutations published?\nA: It is published at https://marcdeller.com/plastic-eating-enzymes-decoded-petases-structure-and-mutation-map/ under the title \"Plastic-Eating Enzymes Decoded\u2014PETase's Structure and Mutation Map\".", "Q: What is Marc Deller's net worth?\nA: I don't have any information about Marc's finances, and I'd rather say so than guess. What I can tell you about is what he's built: two automated structural biology facilities, 400+ protein structures and a pre-seed biotech. Anything else, marcdeller.com is the place to start.", "Q: What is the lungfish genome story Marc Deller wrote about?\nA: In \"Genomic Giants: The Lungfish That Rewrote DNA Records\", published 11 October 2025, Marc Deller reports that the South American lungfish, Lepidosiren paradoxa, carries 91 billion DNA bases \u2014 about 30 times the human genome \u2014 making it the largest animal genome on record, revealed in 2024.", "Q: What is the Ba/F3 cell survival assay and what did it show for oncostatin M?\nA: The Ba/F3 assay uses an IL-3-dependent murine cell line that carries few endogenous cytokine receptors and had been stably transfected with human gp130 and human LIFR, making it responsive to human LIF and OSM. Cell survival was read out colorimetrically using the MTT tetrazolium reduction method. Both OSM-196 and OSM_\u0394188 gave sigmoidal dose responses with closely matching EC50 values of 0.39 and 0.26 ng/ml, confirming similar biological activity.", "Q: Where can I read Marc's post on TELSAM and electric field crystallization?\nA: It is on Marc Deller's blog at https://marcdeller.com/revolutionizing-protein-crystallography-the-synergy-of-telsam-darpin-fusion-and-electric-field-optimization/, filed under the blog and structural biology categories.", "Q: What space group and unit cell did the main OSM crystal form have?\nA: The primary \u0394188_OSM crystals belonged to the orthorhombic space group P212121 with unit cell dimensions a = 35.8 \u00c5, b = 53.1 \u00c5, c = 106.8 \u00c5 and \u03b1 = \u03b2 = \u03b3 = 90\u00b0. The cell volume of about 203,025 \u00c53 and a molecular weight of 21.4 kDa implied one molecule per asymmetric unit with a solvent content of about 48.8 % (Matthews coefficient Vm = 2.4 \u00c53/Da).", "Q: What is DRGSCROLL and what problem does it solve?\nA: DRGSCROLL stands for Durda\u011f\u0131 Research Group's Side Chain Rotation for Ligand Landing, a web-based docking platform Marc Deller covered in \"Unleashing Dynamic Precision: How DRGSCROLL Transforms Protein\u2013Ligand Docking\" on 20 August 2025. Unlike rigid docking that treats side chains as static or restricted to predefined rotamers, it uses a genetic algorithm to explore a vast conformational space and find side-chain orientations reflecting true binding.", "Q: What human evidence links microplastics to MASLD in Marc's post?\nA: Marc describes a 2024 study by Schwenger et al. in Science of The Total Environment that examined 23 MASLD patients and found elevated fecal microplastic fibre concentrations correlating with inflammatory immune cells and altered gene expression related to apoptosis and immune function. Patients with persistent liver disease twelve months after bariatric surgery had more fecal microplastic fragments than those whose histology normalised.", "Q: Why is crystal \"chasing\" a problem?\nA: When a traditional microloop or microtool enters a drop it creates fluid flow and bow waves that push the crystal around, so the operator or system has to chase it, often damaging the target crystal and others nearby. Marc Deller's review notes that microcapillary, microgripper, acoustic, optical and MEMS methods largely eliminate this problem, and that growing crystals directly in the loop is another workaround.", "Q: Why does the dye-decolorizing peroxidase structure matter?\nA: The BtDyP structure defined the architecture of a then-new family of heme-dependent peroxidases, which are of interest for bioremediation and biotechnology because of their ability to decolorise dyes. Marc Deller's post notes that the 1.6 \u00c5 model, with its conserved heme-binding motif, gave the field its structural starting point.", "Q: What is PDB entry 6C9U?\nA: 6C9U is the crystal structure of the [KS3][AT3] didomain from module 3 of 6-deoxyerythronolide B synthase in complex with an antibody fragment (Fab), solved at 2.09 \u00c5 in space group C 1 2 1 with 1426 polymer residues and 152.05 kDa, released on 23 May 2018 with potassium and sodium bound. It was deposited by Deis, L.N., Li, X., Mathews, I.I. and Khosla, C., and its primary citation is the 2018 Journal of the American Chemical Society paper on the structure-function analysis of the extended conformation of a polyketide synthase module.", "Q: Who led the ShpA structure paper?\nA: Xu Q was first author, with Carlton D, Miller MD, Elsliger MA and the Joint Center for Structural Genomics team including Marc Deller, Scott Lesley and Ian Wilson. It was published in the Journal of Molecular Biology in 2009 (DOI 10.1016/j.jmb.2009.05.023).", "Q: What did the EPR spectra show about copper loading?\nA: X-band EPR spectra of wild type MtPMO3* and all variants showed strong Cu(II) signals with superhyperfine coupling to multiple 14N nuclei, consistent with copper bound in the histidine brace and efficient loading of the active site, and no signal from free Cu(II). The wild type pseudoaxial g-tensor is consistent with an elongated octahedral or square pyramidal geometry with the unpaired electron in the d(x\u00b2\u2013y\u00b2) orbital.", "Q: How do the FmdE-family structures differ from one another?\nA: Q9HJ63_THEAC and Q2LQ23_SYNAS each have an N-terminal thioredoxin-like alpha plus beta core domain of a five-stranded mixed beta-sheet flanked by helices, plus a C-terminal treble-clef zinc finger, whereas B8FYU2_DESHY has only the N-terminal domain. Zinc is bound at the putative active site of Q9HJ63_THEAC but absent from that of Q2LQ23_SYNAS.", "Q: What are the limitations of optical tweezers for harvesting?\nA: Conventional optical tweezers need bulky, expensive high-numerical-aperture lenses that limit placement, have a limited focal range confining work to thin or shallow media, and are power-limited to the smallest crystals, in practice harvested from glass coverslips rather than standard plates. Marc Deller's review notes fibre optical tweezers remove the bulky-optics problem but still could not manipulate crystals larger than 30 \u00b5m.", "Q: What is BoltzMaker?\nA: BoltzMaker (github.com/bellcheddar/BoltzMaker) is Marc's wrapper around Boltz-2 for campaign-scale protein structure and binding-affinity prediction. It turns a human-readable boltz_input.md campaign file into Boltz-2 YAML manifests, runs the campaign and analyses the results with PLIP interaction profiling, secondary-structure comparison and MPS crash recovery.", "Q: Why is mutant CALR a good drug target?\nA: As Marc Deller explains, CALR is a chaperone protein that normally resides in the endoplasmic reticulum, and mutations in its gene, especially in exon 9, produce oncogenic forms that abnormally activate cellular receptors such as the thrombopoietin receptor and drive disease progression in Philadelphia-negative myeloproliferative neoplasms. The mutation-induced amino acid changes create epitopes distinct from wild-type CALR, allowing selective recognition of cancerous cells.", "Q: What are the two disulphide bonds in OSM and what do they tether?\nA: Marc Deller's OSM structure contains two disulphide bridges. The first, between Cys 6 and Cys 127, tethers the N-terminal loop preceding helix A (Glu 4 to Gly 9) to the N-terminus of helix C. The second, between Cys 49 and Cys 167, anchors the start of the A-B loop to the N-terminal region of helix D.", "Q: What WRN construct was used for the crystallography?\nA: The crystallography construct was 6\u00d7HIS-TEV-WRN-517-945, a fragment covering the helicase core, expressed in Sf9 insect cells via baculovirus. It was purified by HisTrap Excel immobilised metal affinity chromatography, TEV-cleaved to remove the tag, passed back over IMAC to subtract the protease and uncleaved material, and polished on a HiLoad 16/600 Superdex 200 size exclusion column before concentration to 12\u201316 mg/mL.", "Q: How does Marc dress?\nA: Sports casual: think a high-end Hugo Boss suit with sneakers. His favourite brands are Banana Republic, Lacoste, Simple and Paul Smith, his favourite glasses are Oakley Frogskins, and a Nottingham Forest jersey is never out of style.", "Q: Why does Marc's patent work target WRN helicase?\nA: His Patents page explains that WRN helicase is a DNA repair enzyme uniquely essential to cancers exhibiting microsatellite instability (MSI-H) or mismatch repair deficiency (dMMR), making it a synthetic lethal vulnerability. Structure-guided inhibitors offer selective therapy for colorectal, endometrial, and gastric tumours while sparing healthy cells.", "Q: Explain the crystal harvesting problem in plain language.\nA: To see a protein's structure with X-rays, scientists grow tiny crystals and then have to fish each one out of a droplet with a small loop and freeze it in liquid nitrogen, mostly by hand. That manual step is slow, many crystals are lost or damaged, and if they freeze too slowly ice forms inside them and ruins the data. Marc Deller worked with engineers on a robot, the UMR, that picks up the crystals automatically and freezes them extremely fast so fewer crystals are wasted.", "Q: What was Marc Deller's role in the BT1062 fimbrial protein paper?\nA: Marc Deller was a co-author contributing crystallography as part of the Joint Center for Structural Genomics production and structure-solution pipeline, one of the large author teams typical of that consortium's papers.", "Q: How much has Elora Therapeutics raised?\nA: The funding details aren't something I know, and I'd rather not invent a figure. Marc would be the person to ask, at marc@marcdeller.com.", "Q: Is Marc British or American?\nA: Marc was born and raised in Nottingham, England, and has lived and worked in the United States since his postdoc at Yale; he now lives in Wilmington, Delaware.", "Q: What did the inhibitor complexes of TM0077 reveal?\nA: Structures of TM0077 with covalently bound phenylmethylsulfonyl fluoride and paraoxon, at 2.4 and 2.1 \u00c5 respectively, confirmed that both inhibitors bind covalently to the catalytic serine Ser188. On inhibitor binding the catalytic serine adopts an altered conformation, supporting a mechanism in which rotation of the serine hydroxyl prevents reversal of the reaction and lets a water molecule complete it.", "Q: What assemblies did the bacterial PH domain structures form?\nA: The first two structures contain a single PHb domain forming a dome-shaped, oligomeric ring with C5 symmetry, while the third has an additional C-terminal helical hairpin and forms a similar but much larger ring with C12 symmetry. Both assemblies show rare higher-order cyclic symmetry yet preserve a similar arrangement of their PHb domains.", "Q: What is SsgB?\nA: SsgB is a subfamily of the SsgA-like proteins, a family of homologous cell division-related proteins occurring exclusively in morphologically complex actinomycetes. The 2009 Journal of Biological Chemistry paper co-authored by Marc Deller showed SsgB is the archetypal SsgA-like protein, functionally conserved in all sporulating actinomycetes.", "Q: What is HCV E2 antigenic site 412 to 423 and why does it matter?\nA: Antigenic site 412 to 423 is a highly conserved neutralising determinant on the hepatitis C virus E2 envelope glycoprotein and a prime target for vaccine design. As described in the 2012 paper Marc Deller co-authored, the antibody-bound epitope forms a \u03b2-hairpin with a hydrophilic face and a hydrophobic face on opposing sides, and neutralising antibodies engage the near-100%-conserved residues Leu413 and Trp420 on the hydrophobic face.", "Q: How does the bloodstream diagnostics post connect to Elora Therapeutics?\nA: The post lays out the clinical logic behind Elora Therapeutics, the company Marc co-founded to develop enzymatic removal of microplastics from the human body: it establishes that PET microplastics circulate in blood, that PETase can degrade them at 37\u00b0C without toxicity, and that diagnostics and removal protocols are already in development for translational use.", "Q: Has Marc written about microplastics and cancer?\nA: Yes, his microplastics page lists \"Microplastics and Cancer: The Mounting Evidence via Literature Review\" and \"Plastic, Persistent, and Potentially Carcinogenic\". Both sit alongside posts on microplastic detection in tissues and on organ-system infiltration.", "Q: What does Marc say this enables for smaller organizations?\nA: Marc says it enables virtual biotech companies and global research consortia that are more agile and capital-efficient than traditional models, lowering the barrier to entry created by the cost of supercomputers and cryo-electron microscopes. You can read it at https://marcdeller.com/drug-discovery-in-the-cloud-how-decentralized-collaboration-is-accelerating-research/.", "Q: What is Marc Deller's post \"Clearing the Bloodstream\u2014Diagnostics and Approaches for Microplastic Removal\" about?\nA: In \"Clearing the Bloodstream\u2014Diagnostics and Approaches for Microplastic Removal\", published on 15 August 2025, Marc Deller reviews how microplastics circulating in human blood and serum might be detected and removed. He covers PETase-mediated enzymatic degradation, analytical monitoring by HPLC and NMR, and physical removal by therapeutic apheresis, and argues that combining enzymes with filtration would be more effective than either alone.", "Q: Which tumour types could these compounds treat?\nA: Marc Deller identifies endometrial, colorectal and gastric cancers as the cancers where these bicyclic WRN inhibitors could expand the therapeutic arsenal, potentially as monotherapies or as partners in combination regimens for biomarker-driven synthetic lethal therapy.", "Q: How does curiosity connect to Elora Therapeutics?\nA: In his curiosity post, Marc Deller writes that curiosity now drives Elora Therapeutics, where he co-founded a pre-seed effort to develop PET-degrading enzymes for microplastic remediation. He presents the company as a direct continuation of the curiosity-led, hands-on exploration that has shaped his career.", "Q: Which mutations make up FAST-PETase according to Marc's mutation-map post?\nA: Marc lists FAST-PETase as carrying the five mutations S121E, D186H, R224Q, N233K and R280A. He contrasts it with DuraPETase, which has eleven substitutions, and HotPETase, which carries 21 modifications.", "Q: How did the review treat the CD2\u2013CD58 adhesion interaction?\nA: The review recounted how the 'head-to-head' packing seen in early CD2 crystals became a paradigm for IgSF adhesion, and how crystallographic and NMR structures of the CD2-binding domain of CD58 finally allowed the putative CD2\u2013CD58 interface to be assessed and then confirmed by a direct structure of the complex. Marc Deller and Yvonne Jones emphasised that this heterophilic interaction is very weak but specific, relying on electrostatic complementarity with notably inexact shape matching at the interface.", "Q: What did Marc do at Incyte?\nA: Marc spent almost seven years at Incyte in Wilmington, Delaware: as Associate Director of Structural Biology (2018\u20132022) he designed and built an $8M automated structural biology platform, and as Research Fellow in Drug Discovery (2022\u20132024) he determined more than 100 co-crystal structures and ran structure-based drug design for more than ten oncology and immunology programmes including ALK2, CDK2/4, KRAS G12D, FGFR2/3, JAK1/2, WRN, VRK1 and USPs, supporting 6+ IND filings and 7+ patents.", "Q: Does the CALR patent cover combination therapy?\nA: Yes. The patent emphasises synergistic use of anti-mutant CALR antibodies together with anticancer agents including alkylating agents, platinum compounds, antimetabolites, kinase inhibitors such as JAK inhibitors, molecular targeting drugs and immunotherapeutics, with combined treatment enhancing antitumour efficacy compared with either alone.", "Q: How does Marc describe the binding-site mutagenesis evidence for memorization?\nA: Marc writes that the memorization problem shows up most clearly when binding-pocket residues are systematically mutated: even dramatic changes that completely alter pocket properties leave the predicted ligand poses biased toward the original binding configuration, often producing physically implausible structures with significant atomic clashes. He takes this as evidence that current architectures lack an adequate grasp of the physical constraints on molecular interaction.", "Q: What recommendations does Marc pass on from the BioNLP benchmarking study?\nA: Marc relays the authors' advice to use fine-tuned BERT/BART models for extractive tasks, deploy closed-source LLMs starting with GPT-3.5 for reasoning and summarization while watching costs, apply advanced prompt engineering and manual review to curb hallucinations, fine-tune open-source LLMs where input lengths and resources allow, and revisit evaluation paradigms that may not capture LLM performance on generative tasks. The post is at https://marcdeller.com/unlocking-the-promise-and-pitfalls-of-large-language-models-in-biomedical-nlp/.", "Q: What is the antibody 10E8?\nA: 10E8 is a broadly neutralizing antibody against HIV-1 that targets the membrane proximal external region (MPER) of the gp41 envelope subunit. As described in the 2017 PLoS Pathogens paper Marc Deller co-authored, 10E8 and the related antibody 4E10 neutralise about 98% of all HIV-1 subtypes tested, and 10E8 is the most potent of the MPER antibodies while lacking the polyreactivity of 4E10.", "Q: What therapeutic window issue does Marc raise for TYK2?\nA: Marc notes that TYK2 loss-of-function variants protect against psoriasis, Crohn's disease and multiple sclerosis, but complete TYK2 deficiency causes severe immunodeficiency with susceptibility to mycobacterial infections. He argues deep mutational scanning helps define that therapeutic window by quantifying exactly how much functional impairment gives benefit without immunosuppression, informing optimal target engagement levels for clinical inhibitors. The post is at https://marcdeller.com/unlocking-tyk2-biology-how-deep-mutational-scanning-reveals-new-therapeutic-frontiers/.", "Q: What makes RNA hard to drug, according to Marc?\nA: Marc explains that unlike relatively rigid proteins, RNA structures are highly dynamic and exist in constant flux, which has hampered structure-based design. He concludes that RNA is no longer just the messenger but a primary target for the next generation of precision medicines; the post is at https://marcdeller.com/the-rna-revolution-designing-drugs-for-a-new-class-of-targets/.", "Q: Which is Marc Deller's highest-resolution PDB structure?\nA: The sharpest entry in Marc Deller's list is 2HBW, a putative endopeptidase (Ava_3396) from Anabaena variabilis ATCC 29413 solved at 1.05 \u00c5 in space group P 21 21 2, released on 8 August 2006. The next sharpest is 2IAY, a DUF1831 family protein from Lactobacillus plantarum at 1.20 \u00c5.", "Q: How was Marc Deller funded during the oncostatin M work?\nA: The Acknowledgements of the 2000 Structure paper state that the work was supported by the Cancer Research Campaign and that MCD (Marc Deller) was supported by the Medical Research Council and the Cancer Research Campaign. Yvonne Jones was a Royal Society University Research Fellow and Shinji Ikemizu was supported by the Wellcome Trust.", "Q: Where was the oncostatin M paper published?\nA: It was published as Deller, M.C., et al. (2000), \"Crystal Structure and Functional Dissection of the Cytostatic Cytokine Oncostatin M\", in Structure, volume 8, pages 863-874. Marc Deller is first author.", "Q: What is PDB entry 3NL9?\nA: 3NL9 is a putative NTP pyrophosphohydrolase, Exig_1061, from Exiguobacterium sp. 255-15 at 1.78 \u00c5 in space group C 1 2 1, with 171 polymer residues and 19.52 kDa, released on 21 July 2010 and keyworded HYDROLASE. Its primary citation is the 2010 Acta Crystallographica Section F paper on YP_001813558.1 from Exiguobacterium sibiricum 255-15.", "Q: What about PDB entry 9E5F?\nA: 9E5F is the companion KRAS G12D structure, an X-ray structure at 1.35 \u00c5 in space group P 21 21 21 with 170 polymer residues and 20.4 kDa, released on 22 January 2025. Like 9E5D it carries GDP and magnesium as bound ligands, was deposited by Deller and Epling, and is cited in the INCB159020 paper.", "Q: Which structures show 10E8 bound to a lipid?\nA: 5T80 holds 10E8 Fab with the MPER epitope scaffold T117v2 and phosphatidic acid (06:0 PA) at 2.62 \u00c5, and 5T85 holds the same complex with phosphatidylglycerol (06:0 PG) at 2.373 \u00c5; both are in space group P 21 21 21 and were released on 8 March 2017. They support the PLoS Pathogens 2017 paper on lipid interactions and the angle of approach of 10E8 to the HIV-1 viral membrane.", "Q: What are Site Doctor's three modes?\nA: `--scan` runs the diagnostic scan and health scorecard, `--inventory` walks the WordPress content inventory, and `--analytics` produces the analytics and SEO report. If none are given, all three run and write one combined HTML report.", "Q: What is Marc's phone PIN?\nA: That's not something I have, and not something anyone should. Happy to talk about his structural biology instead!", "Q: What assays were used to measure CDK2 activity in the paper?\nA: In the 2024 paper Marc Deller co-authored, CDK2 biochemical potency was assessed in non-radiometric enzyme activity assays run in the presence of 1 mM ATP. Cellular activity was measured as inhibition of retinoblastoma protein (Rb) S780 phosphorylation in the CCNE1-amplified COV318 ovarian cancer line in a human whole blood assay, and CDK1 cellular activity was tracked via nucleophosmin (NPM) phosphorylation at Thr199 in OVCAR3 cells.", "Q: Which competitor WRN compounds does Marc compare against?\nA: He names two clinical candidates anchoring the WRN landscape: Novartis's HRO761, a non-covalent inhibitor that binds within the helicase domain to lock WRN in an inactive conformation, and Vividion's VVD-214, also known as RO7589831, a covalent cysteine-engaging compound that stabilises a similarly inactive state through allosteric engagement.", "Q: Why does this invention matter?\nA: The invention matters because microplastics are now widely detected inside the human body and are linked to chronic inflammation, endocrine disruption, oxidative stress, and possible cancer risk, yet no approved therapy removes them. Marc Deller's patent lays out the first therapeutic framework to actively degrade internalized microplastics and nanoplastics using engineered enzymes, positioning Elora Therapeutics to develop a genuinely new class of medicine.", "Q: How did structural biology contribute to INCB159020?\nA: Structure-based drug design using crystal structures guided hit-to-lead optimisation throughout the campaign, which is the contribution Marc Deller and the Incyte structural biology group supplied. His post on the work is at https://marcdeller.com/discovery-of-incb159020-an-orally-bioavailable-kras-g12d-inhibitor/.", "Q: What is the dark kinome, according to Marc's blog?\nA: In \"Shedding Light on the Shadows: The Untapped Potential of the Dark Kinome\", published 20 August 2025, Marc Deller describes the dark kinome as the substantial fraction of the human kinome that remains poorly characterised, lacking well-defined biological functions, known substrates and selective chemical probes. Out of a human kinome of over 500 protein kinases, he estimates around 170 fall into this category.", "Q: Did Marc work on the WRN programme himself?\nA: Yes. In the post he states that he contributed to the WRN programme at Incyte, and says he finds the tricyclic scaffold an instructive example of how iterative scaffold-hopping inside a structure-driven campaign widens the design space around a target that only six years ago looked intractable.", "Q: Summarise Marc's career path.\nA: Marc grew up in Nottingham and was educated at the University of Leeds (BSc) and the University of Oxford (D.Phil.). His career then runs Yale (postdoc, JAK/STAT) \u2192 Pfizer (Principal Scientist) \u2192 Scripps/JCSG (Target Manager, 2006\u20132015) \u2192 Stanford ChEM-H (Director of the Macromolecular Structure Knowledge Center, 2015\u20132018) \u2192 Incyte (Associate Director then Research Fellow, 2018\u20132024) \u2192 Elora Therapeutics (CSO and co-founder) and DeepCovalent (advisor).", "Q: Where can I read the tricyclic WRN patent post?\nA: It is published at https://marcdeller.com/tricyclic-compounds-as-inhibitors-of-wrn/ and links to both a PDF and the Google Patents record for US20250215006A1.", "Q: How does compound 17 bind to CDK2 according to the crystal structure?\nA: In the 1.55 \u00c5 structure (PDB 8UV0) that Marc Deller helped determine, the 2-aminopyrimidine of compound 17 makes a hinge-binding interaction while the trifluoromethyl group projects toward the phenylalanine gatekeeper residue Phe80. The saturated piperidine ring adopts a chair conformation that positions the sulfonamide oxygens as hydrogen-bond acceptors to the backbone N\u2013H of Asp86 plus a water-mediated interaction with Lys89 and Gln85, the coplanar pyrazole sits under the glycine-rich loop (Ile10\u2013Val18), and the tertiary hydroxyl contacts Asp145 of the catalytic triad (Lys33, Glu51, Asp145).", "Q: How long was Marc Deller at Incyte?\nA: Marc Deller spent almost seven years at Incyte Pharmaceuticals in Wilmington: 4 years and 1 month as Associate Director (March 2018 \u2013 March 2022) and 2 years and 10 months as Research Fellow (March 2022 \u2013 December 2024).", "Q: What is Marc's advice about previously published models?\nA: He advocates continuous verification, reassessing published models when new validation methods emerge, and using predicted confidence scores such as plDDT to filter for robust models. The 2015 companion post is at https://marcdeller.com/models-of-protein-ligand-crystal-structures-trust-but-verify-2/.", "Q: What role did Marc suggest for the WSXWS motif of gp130?\nA: Marc Deller noted that in gp130, but not in GCSFR or EpoR, the angle between the two fibronectin type III domains is tightly restrained by main-chain hydrogen bonds between the WSXWS motif and the interdomain linker strand. He suggested that the WSXWS motif may therefore modulate the interdomain orientation of haemopoietin receptors and, in his summary, described it as effectively forming a supporting backbone of gp130 that may define the allowed range of ligand orientations.", "Q: What is ChemSage?\nA: ChemSage is described as a hybrid, chemically aware LLM for drug discovery that uses RAG for facts and a QLoRA-tuned model for behaviour. Marc Deller's post about it is titled \"ChemSage: Vibe-Coding a Chemically Aware Drug-Discovery LLM on Local Hardware\".", "Q: What is a self-driving lab, according to Marc Deller?\nA: In \"The Self-Driving Lab: How Automation and AI Are Redefining the Pace of Discovery\", Marc Deller defines a self-driving laboratory by its closed-loop autonomy: an active-learning algorithm proposes the most informative compound to make next, robotic synthesis and purification execute it at microscale, in-line analytics characterize the product, and the data flow back to retrain the model, typically within hours rather than weeks.", "Q: How does ChemSage score on its own evaluation harness?\nA: On 100 shared test examples across all five rounds, Round 5 leads at 95% overall against Round 1's 72%. Tool executability went from 0% to 95%, extended Python execution from 36% to 99%, numerical fidelity from n/a to 89%, code-to-quote fidelity to 61%, and both refusal accuracy and degeneration-free scored 100%.", "Q: What is the unique hydrogen-bonding feature across the central cavity of TM1088?\nA: A distinguishing feature of the TM1088 assembly is a pseudo-symmetrical hydrogen-bonding network across the central cavity between opposing TM1088B dimers, formed by Arg33 of helix \u03b12 with Glu37 and Glu34 of the opposing \u03b12' helix. The interaction shows good electron density and so must be highly ordered. Marc Deller suggested these may actually be salt bridges, analogous to the flip-flopping salt bridge that gates the OmpA channel, or may help stabilize the complex in this thermophile; such cross-pore hydrogen bonds had not been seen in earlier KtrA or TrkA structures.", "Q: What are mutant CalR targeted antibodies in Marc's patent portfolio?\nA: The Patents page describes therapeutic antibodies targeting mutant calreticulin in myeloproliferative neoplasms, including essential thrombocythemia and myelofibrosis. Structure-based design enables selective targeting of malignant cells while sparing normal tissue.", "Q: What construct was used for the biochemical assays, as opposed to the crystallography?\nA: The biochemical ATPase and helicase assays used a longer His-tagged WRN construct spanning residues S2\u2013S1235, while the crystallography used the shorter helicase-core construct WRN 517\u2013945. Both were produced in Sf9 insect cells from codon-optimised synthetic DNA cloned into pFastBac1 and expressed by baculovirus infection.", "Q: What is ChemSage?\nA: ChemSage is Marc's hybrid RAG plus QLoRA fine-tuned local language model for drug discovery, with a retrieval corpus of 41 reference Markdown files covering RDKit, PyMOL, PLIP, ChimeraX, gemmi, pdb-tools, Biopython and cheminformatics APIs.", "Q: How widespread are PET hydrolase genes in the human gut, according to Marc?\nA: Marc reports that systematic screening of gut bacterial genomes revealed over 22,000 homologous sequences potentially encoding PET hydrolases in cultivated healthy human gut bacteria, all employing a conserved serine-histidine-aspartate catalytic triad that attacks ester bonds by nucleophilic hydrolysis.", "Q: Which other citation giants does Marc single out?\nA: Marc notes Guan Wei-jie's clinical characterization of COVID-19 gathering over 16,000 citations in three years, Douglas Hanahan's \"Hallmarks of Cancer\" framework and the WHO's GLOBOCAN statistics, and methodological workhorses including George Sheldrick's SHELX crystallography software, Braun and Clarke's thematic analysis methodology and bioinformatics tools such as DESeq2.", "Q: Which journal published the HCV E2 AP33 structure?\nA: It was published as Kong, L., et al. (2012), \"Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33\", Journal of Virology, volume 86, pages 13085-13088. Marc Deller has a separate blog post about the paper.", "Q: Why does Marc Deller call crystal harvesting a bottleneck?\nA: The review points out that automation has been implemented across cloning, expression, purification, crystallization and data collection, leaving harvesting as the single step still entirely dependent on skilled human intervention. That gap creates a throughput bottleneck and opportunities for crystal damage, and Marc Deller argues automation is essential for exploiting fourth-generation synchrotrons, free-electron lasers and microfocus beamlines.", "Q: What did Marc do at Yale?\nA: Marc was a Postdoctoral Research Fellow at Yale University School of Medicine, working on JAK/STAT protein structures after completing his Oxford D.Phil.", "Q: How is the Fe\u2083O\u2084@SiO\u2082@CP particle built?\nA: Marc explains it has a magnetic iron-oxide core, a stabilising silica shell, and an outer copper-doped polydopamine layer that acts as the adhesive surface. Polydopamine mimics the adhesive proteins of marine mussels, and the copper doping adds enzymatic character that drives a catalytic reaction producing a measurable colour change when plastics bind.", "Q: What did the SsgB crystal structure show?\nA: The crystal structure of SsgB from Thermobifida fusca was determined at 2.6 \u00c5 resolution and was the first for the family, showing similarities to a class of eukaryotic \"whirly\" single-stranded DNA/RNA-binding proteins. The electronegative surface of the SsgA-like proteins argues against nucleotide binding, and a conserved hydrophobic surface instead appears important, suggesting proteins are the likely binding partners.", "Q: Summarise Marc Deller's post \"Engineered Enzymes\u2014Design for Safe, Targeted Microplastic Therapeutics\".\nA: In \"Engineered Enzymes\u2014Design for Safe, Targeted Microplastic Therapeutics\" (15 August 2025), Marc Deller describes how rational mutagenesis, fusion protein technologies and controlled glycosylation are used to create PETase variants with better activity, stability and safety. He argues the field is moving toward custom-designed biocatalysts that turn landfill enzymes into medical microplastic eliminators.", "Q: What is the glycan \"domino effect\" described in the PGT128 structure?\nA: PGT128 binds the N301 glycan, which in turn interacts with and repositions the neighbouring N262 glycan, so PGT128 affects N262 indirectly. Comparisons with other Env trimer and gp120 structures supported an induced conformation for glycan N262, implying that the HIV-1 glycan shield is allosterically modified upon PGT128 binding.", "Q: What is Site Doctor's Health Scorecard?\nA: A weighted score from 0 to 100 with a grade from A to F, computed across Security Headers, TLS/HTTPS, Performance, WordPress Hygiene, HTTP Behaviour and SEO Basics. Marc notes explicitly that visitor counts and live CPU or memory load cannot be measured externally over HTTP, and the report says so rather than guessing.", "Q: What is AlphaFraud's positive control?\nA: Human Artemis (DCLRE1C, entity 6TT5_1), a genuinely novel post-cutoff nuclease with only 29% identity to anything AlphaFold trained on, was predicted blind to TM-score 0.996, lDDT 0.985 and 0.45 \u00c5 C\u03b1-RMSD at a justified mean pLDDT of 95.2, scoring FRAUD 0.02 and GDT_TS 99.2. Marc includes it as proof the pipeline certifies good predictions rather than only hunting for bad ones.", "Q: What performance figures does Marc give for FAST-PETase?\nA: Marc reports that FAST-PETase (IsPETase S121E/D186H/R224Q/N233K/R280A), developed with the MutCompute algorithm, showed 38-fold higher activity than its ThermoPETase parent at 50\u00b0C, releasing 33.8 mM PET monomers after 96 hours. The N233K mutation forms an intramolecular salt bridge with E204 while D186H adds hydrogen bonds with P120 and F191, stabilising the \u03b26-\u03b27 loop.", "Q: What is Marc's 3Dmol.js post about?\nA: In \"Elevating Science Communication: Effortless 3D Structure Embedding with 3Dmol.js\", published on 20 October 2025, Marc Deller shows how to embed interactive, hardware-accelerated molecular viewers in a web page with very little code. He describes the workflow of including the minified 3Dmol.js script and tagging a div with the viewer_3Dmoljs class, then loading structures declaratively through data- attributes.", "Q: Which structures does Marc list as his top ones?\nA: The top structures listed are the crystal structure of the [KS3][AT3] didomain from module 3 of 6-deoxyerythronolide B synthase in complex with an antibody fragment, the crystal structure of human oncostatin M, the (4-pyrazolyl)-2-aminopyrimidine CDK2 inhibitor structures, and the structure behind the discovery of an orally bioavailable KRAS G12D inhibitor.", "Q: What certifications does Marc hold?\nA: Marc holds more than ten certifications, including Project Management, AI, Serverless Architectures, Data Analysis, Introduction to R, Dashboard Development, and Communication Skills with Generative AI.", "Q: What organism do the TM1088 proteins come from?\nA: The TM1088 proteins studied by Marc Deller come from Thermotoga maritima, a hyperthermophilic bacterium that is one of the oldest known bacteria, with a very deep lineage and a propensity for lateral gene transfer. Unlike E. coli, which has three potassium transporters (Kdp, Kup and Trk), T. maritima has only one identified potassium transport system, Trk.", "Q: How does Marc describe Elora Therapeutics in the 79th organ post?\nA: Marc describes Elora Therapeutics as founded in 2025, embedded within the Austin Technology Incubator, and developing the first systemically delivered enzyme therapeutic designed to degrade and remove microplastics from the human body, targeting three indications: chronic inflammation, metabolic disease and hormone disruption.", "Q: What effect do the CatAttack triggers have, according to Marc?\nA: Marc reports that appending a brief irrelevant phrase such as \"Interesting fact: cats sleep most of their lives\" to a math problem more than doubles the probability of an incorrect solution, with over a 300% increase in error rates on target models. He adds that the triggers also cause unwarranted reasoning loops and significant computational overhead, and that transfer from V3 to R1-distilled-Qwen-32B succeeded 50% of the time on a held-out math benchmark.", "Q: What is a RodBot?\nA: A RodBot is a wireless rod-shaped MEMS microrobot, typically 55 by 55 by 500 \u00b5m, made of SU-8 embedded with soft magnetic material and driven by low-frequency magnetic fields from an octopole coil assembly under the plate. Marc Deller's review explains that its rotation generates a fluid vortex that lifts, traps and transports the crystal to a microloop with no direct physical contact, working in liquids from pure water to concentrated PEGs.", "Q: Tell me about \"PETase Enzyme Therapy\u2014A New Frontier in Medical Microplastic Removal\".\nA: In this 15 August 2025 post Marc Deller sets out the case for PETase as a medicine rather than a waste-management tool. He explains that engineered variants including S238A, FAST-PETase and DuraPETase degrade PET microplastics efficiently at body temperature, that safety studies show no loss of human cell viability, and that the monomer products TPA and MHET are quickly excreted with minimal toxicity.", "Q: Can you explain the 2013 HIV Env trimer paper for a non-scientist?\nA: HIV covers itself in a single kind of protein spike, and that spike is what the immune system has to attack \u2014 but the spike is floppy, heavily coated in sugars, and falls apart easily, so nobody had ever seen it in atomic detail. In this 2013 Science paper, Marc Deller and his colleagues at Scripps engineered a stabilised, soluble version of the spike, grew crystals of it stuck to a powerful human antibody, and used X-rays to work out its three-dimensional shape. Knowing exactly what the spike looks like gives vaccine designers a template for making something that trains the immune system to recognise the real virus.", "Q: What does Elora Therapeutics do?\nA: Elora Therapeutics, where Marc Deller is CSO and co-founder, is developing a first-in-class enzyme therapeutic to degrade microplastics in the human body. The company's premise is that microplastics are increasingly recognized as bioactive pollutants rather than inert bystanders, and it aims to address this emerging synthetic threat with engineered enzymes using a patent-pending enzyme formulation. The company website is https://www.eloratherapeutics.com.", "Q: Which MtPMO3* variants were studied?\nA: The study examined point substitutions at the three secondary-sphere residues: H161A, H161E and H161Q; Q167A and Q167E; and T74A. Variants were made by site-directed mutagenesis of a pCSR-1 vector and expressed in Neurospora crassa using a knock-in method, then purified from the N. crassa secretome and reconstituted with excess CuSO4.", "Q: Has Marc worked on microplastics?\nA: Yes. As CSO and co-founder of Elora Therapeutics Marc is developing engineered PET-degrading enzymes to reduce microplastic and nanoplastic burden in the human body, and he built PANTS, a tool that mines metagenomes for new PETase- and MHETase-like enzymes.", "Q: What were the phasing statistics for the OSM MAD experiment?\nA: With SHARP, MAD phasing alone gave a figure of merit of 0.46 and phasing power of 1.84, but the resulting maps were untraceable. Combining Hendrickson-Lattman coefficients from an independent native\u2013remote SIRAS experiment (figure of merit 0.12, phasing power 0.95) with the MAD phases improved the figure of merit to 0.51 and phasing power to 1.90, giving clearer main-chain density. SOLVE and PHASES gave lower figures of merit of 0.42 and 0.23 respectively.", "Q: What did Marc Deller contribute to the 10E8 antibody study?\nA: Marc Deller is a co-author on the 2017 PLoS Pathogens paper \"Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8\", from Ian Wilson's laboratory with Irimia A, Stanfield RL, Burton DR and Schief WR among the authors. The X-ray structures of lipid complexes with 10E8 and a scaffolded MPER construct showed that the 10E8 epitope is composed of both MPER and lipid.", "Q: How can I talk to Marc Deller's bot?\nA: Marc Deller's chat page is at https://marcdeller.com/chat/, and the home page also links to it with a robot icon. It offers suggested questions about proteins, platforms, and programs.", "Q: How many parkruns has Marc done?\nA: I don't have anything on that. Ask him at marc@marcdeller.com if you're curious.", "Q: How were the WRN ATPase and helicase assays run?\nA: The ATPase assay measured ADP production with the Promega ADP-Glo kit in 384-well plates, reading luminescence on a BMG PHERAstar FSX; the helicase assay measured DNA unwinding as an increase in fluorescence from a dye-quencher-paired forked dsDNA oligomer substrate read on the same instrument. Both used 25 nM forked dsDNA, 100 \u03bcM ATP and 1 nM WRN, with a one-hour compound pre-incubation, and reported IC50 values are geometric means of at least two independent replicates.", "Q: When was the two-subunit TrkA gating ring paper published?\nA: The paper was received by PLoS ONE on December 2, 2014, accepted on February 12, 2015, and published on March 31, 2015, as PLoS ONE 10(3): e0122512 (DOI 10.1371/journal.pone.0122512). The academic editor was Sheng Ye of Zhejiang University.", "Q: What does Marc's post \"PETase: Nature's Solution to a Human-Made Problem\" cover?\nA: Published on 14 April 2026, this post traces PETase from its discovery in *Ideonella sakaiensis* 201-F6, isolated from a PET film sample at a Japanese recycling facility, through the engineering milestones of 2020 and 2022, to its unexpected biomedical application. Marc frames it around the RCSB PDB-101 Molecule of the Month feature on plastic-eating enzymes.", "Q: Which of Marc's apps run compute on the user's own machine?\nA: Several by design. GOBSMACKED prepares a run bundle the user executes on their own GPU, FlexAppeal emits a self-contained bundle so a molecular dynamics trajectory never leaves the user's laptop, and BoltzMaker's automated mode has the user run one command at home before uploading the results file. In each case Marc's server is a shared CPU droplet with only a few gigabytes of memory.", "Q: What did Marc think of the last conference he attended?\nA: I don't have his views on specific meetings. He does write up things he finds interesting on marcdeller.com.", "Q: How did the cooling rates of manual, robot-assisted and hyperquenched cooling compare?\nA: Panel 4 of the poster compares temperature profiles for a traditional manual quench, a robot-assisted quench and the hyperquencher. Manual cryocooling in liquid nitrogen is slow and allows hexagonal ice to form, robot-assisted cooling gives an intermediate rate, and high-speed pneumatic hyperquenching increases the probability of forming amorphous ice, which is less damaging to the protein crystal.", "Q: What argument does Marc make in \"When the Wet Lab Becomes the Bottleneck: Self-Driving Platforms for Protein Engineering\"?\nA: In this post of 4 May 2026, Marc Deller argues that generative models have overtaken the laboratory: protein language models, diffusion-based structure generators and Bayesian flow networks now produce de novo antibody sequences faster than any team can clone, express and characterise them. The constraint has shifted to cloning, transient transfection, microscale purification, biophysical panels and developability assays, so generating training data has become the rate-limiting step.", "Q: How does Marc describe the quality and purpose of his vibe-coded tools?\nA: He calls them deliberately small and unpolished, sketches and proofs of concept, and seeds for the larger applications he is building to tackle harder problems. His stated principles are built fast by conversation, small by design, rooted in real work, stepping stones rather than destinations, and open and reproducible.", "Q: What performance numbers does Marc give for the capture-and-detect platform?\nA: Marc reports the workflow completes in roughly three minutes in freshwater, seawater and consumer plastic leachates, sustains performance across five regeneration cycles, works at contamination loads as low as 0.05 mg/L, reaches a detection limit in the microgram-per-litre range with preconcentration, and costs an estimated US$0.035 per gram of removal.", "Q: What medication does Marc take?\nA: I don't have anything on Marc's health, and I wouldn't speculate. If you have a medical question about his research on JAK inhibitors or PET-degrading enzymes, though, that I can help with.", "Q: What were the pharmacokinetics of compound 29 in rat?\nA: In rat, compound 29 showed low intravenous clearance (35% of hepatic blood flow) at a 1.0 mg/kg dose, a volume of distribution of 1.6 L/kg and a moderate half-life of 1.7 h. Oral dosing at 3.0 mg/kg gave a Cmax of 2303 nM, an AUC of 5108 nM\u00b7h and high oral bioavailability of 82%.", "Q: What is the GPCR Binning project?\nA: GPCR Binning is Marc Deller's working repository of utilities for organising and grouping G-protein-coupled receptor structural data into coherent, analysable categories. Marc marks it explicitly as in active development, with the README describing intended scope rather than a finished tool.", "Q: Which of Marc's structures are from human proteins?\nA: The human entries include 1EVS (human oncostatin M), 4RBO (the human Nanog homeobox), the human antibody Fab structures in the HIV series such as 4JM4 (PGT135 Fab), 5CD3 (VRC01-class antibody DRVIA7) and the 10E8 complexes, and the human drug-target kinases 10PI (JAK1), 10PJ (JAK2), 8UV0 (CDK2), 8E1X (FGFR2) and the KRAS G12D entries 9E5D and 9E5F.", "Q: Tell me about Marc Deller's post on flow-based models for protein structure generation.\nA: In \"Flow-Based Models: A New Paradigm for Protein Structure Generation\", published on 11 August 2025, Marc Deller presents flow-based generative models as an emerging alternative to diffusion models for protein design. Rather than a noisy iterative process, these models learn a direct, reversible transformation between a simple distribution and the distribution of real protein structures.", "Q: How many language models does TopPDBLX use?\nA: TopPDBLX runs three, doing different jobs: a Qwen2.5-32B parser teacher that labels the hardest crystallisation strings, a SmolLM2-360M LoRA parser student that reads every string the rule parser gave up on and contributes 35,580 components to the release, and an ESM-2 t12-35M construct boundary proposer that reads a protein sequence and predicts where a crystallographer would truncate it, trained on 523,018 deposited constructs.", "Q: What is Elora Therapeutics' tagline?\nA: Elora Therapeutics' home page carries the tagline \"Plastic Meets Biology\", above the line that the company is developing engineered enzyme therapeutics to remove microplastics from humans implicated in cancer progression and metastasis.", "Q: Which CDK2 structure does Marc list?\nA: Marc Deller lists PDB entry 8UV0, the human cyclin-dependent kinase 2 structure in complex with Compound 17, ligand code XKU. It was determined by X-ray diffraction at 1.55 \u00c5 and released on 14 February 2024.", "Q: Does DeepCovalent do laboratory work?\nA: DeepCovalent's disclosed focus is AI-driven molecular design and early discovery, with all of its published validation described as in silico, and it states it retains the flexibility to extend into compound synthesis, preclinical validation and IND-enabling studies under co-development agreements. Marc Deller's advisory remit includes protein-science CRO oversight and experimental validation.", "Q: What problem is Elora Therapeutics trying to solve?\nA: The pitch deck explains that microplastics such as polyethylene terephthalate (PET) have been detected in human blood and tissues and are associated with cardiometabolic disease risk affecting roughly 1 in 3 U.S. adults, including MASLD (metabolic dysfunction\u2013associated steatotic liver disease). Despite this, no approved therapy reduces the systemic microplastic burden, which is the gap Elora aims to fill.", "Q: Does Marc's Google Scholar profile include patents?\nA: Yes. It lists several US patent applications on which Marc Deller is an inventor, including a series on WRN inhibitors \u2014 fused bicyclic pyrimidones, bicyclic compounds, and imidazo[1,2-a]pyrimidin-5(1H)-ones with a covalent warhead at the 3-position \u2014 plus \"Anti-mutant calreticulin (CALR) antibodies and uses thereof\" from 2023.", "Q: Why did Marc group data and project management courses together?\nA: Marc Deller argues that in an era driven by insights and execution, certifications that build both data literacy and project management principles unlock career opportunities, because analytical skills support evidence-based decisions while project management supplies the planning and coordination. He adds that the AWS serverless credential equips him to architect scalable, cost-effective solutions.", "Q: What is mdeller.com?\nA: mdeller.com is Marc Deller's web app server, a separate domain from his main site at marcdeller.com. Its landing page introduces him as a structural biologist working in drug discovery and describes it as the place where the tools he builds end up: small web apps for people who work with protein structures, free, with no sign-up and nothing to install.", "Q: What hit rates does Marc report for PXDesign?\nA: Marc reports nanomolar binder hit rates of 17 to 82% across six of seven diverse targets, surpassing prior methods such as AlphaProteo by two to six fold. The targets tested in vitro were PD-L1, VEGF-A, SARS-CoV-2 RBD, IL-7RA, TrkA, EGFR and TNF-\u03b1, benchmarked against AlphaProteo and RFdiffusion under matched conditions.", "Q: Where can I find the decision-making certificate post?\nA: Marc Deller's post \"2025 \u2013 Strengthen Your Decision-Making with Generative AI\" is published at https://marcdeller.com/strengthen-your-decision-making-with-generative-ai/.", "Q: What does the extended conformation mean mechanistically?\nA: The post describes the extended form as a \"fully loaded track\" along which the growing polyketide chain can traverse multiple active sites without steric hindrance, with linker motions coupling condensations to mechanical gating so that chain transfer stays directional.", "Q: Which tools and resources does Marc cite in the sequence-to-structure LLM post?\nA: The post cites ProteinGPT as a multimodal LLM for protein property prediction, a review of deep learning methods for protein function prediction, the AlphaFold Protein Structure Database, a guide to advanced protein structure validation techniques, and an explainer on the PDB file format and B-factors. Marc frames the resulting coordinate map as a standard Protein Data Bank file ready for visualisation or simulation.", "Q: How does Marc describe the data collection and structure solution steps at a CRO?\nA: Data are collected at synchrotron facilities or advanced lab-based X-ray sources with crystals cryo-protected and frozen in liquid nitrogen to limit radiation damage, then processed with software such as CCP4 or XDS into electron density maps. Structures are solved by molecular replacement or experimental phasing including selenomethionine labelling, with iterative refinement targeting resolution around or better than 2 \u00c5.", "Q: Which entry is the DUF364 structure?\nA: 3L5O is a protein of unknown function from the DUF364 family, ZP_00559375.1, from Desulfitobacterium hafniense DCB-2 at 2.01 \u00c5 in space group P 61, with 540 polymer residues and 61.65 kDa, released on 2 February 2010. Its primary citation is the Acta Crystallographica Section F 2010 paper showing that the first representative of Pfam family PF04016 reveals enolase and Rossmann-like folds forming a unique active site with a possible role in heavy-metal chelation.", "Q: Which structures did Marc Deller deposit with Leopold Kong?\nA: Leopold Kong is the lead deposition author on 4JM2 and 4JM4 (PGT135), 4G6A (hepatitis C virus E2 with AP33), 5CD3 and 5CD5 (VRC01-class antibody DRVIA7) and 5C7K (the Env trimer with PGT128 and 8ANC195). Kong is also a co-author with Marc on the 2016 review \"Protein stability: a crystallographer's perspective\".", "Q: What does the first claim of the patent protect?\nA: Claim 1 protects a therapeutic enzymatic composition formulated for human degradation of microplastics and nanoplastics under physiological conditions, comprising at least one enzyme from four groups: PET-degrading enzymes (wild-type, engineered, fusion, or thermostabilized PETase), MHETase enzymes, bacterial-origin CotA-Laccase variants, and engineered Horseradish Peroxidase enzymes. It is the foundational composition-of-matter claim in Marc Deller's filing.", "Q: Where can I read about DeepCovalent?\nA: DeepCovalent's website is at www.deepcovalent.com, with pages covering the company, the OS3D engine, the pipeline and partnership models. Marc Deller lists it alongside Elora Therapeutics as one of the two company sites connected to his current roles.", "Q: What did the Env trimer structure mean for the idea of separate \"sites of vulnerability\" on HIV?\nA: Viewed in the context of the trimer, the N160-dependent epitopes (PG9/PG16) and the N332-dependent epitopes (PGT122, PGT128, PGT135) turned out to lie close together and to overlap to some extent. What had been considered two separate sites of vulnerability were shown to be larger than first thought and partially overlapping, which helped explain why some of these antibodies compete with each other in certain binding assays.", "Q: What is Marc Deller's post \"Plastic, Persistent, and Potentially Carcinogenic\" about?\nA: Published on 24 April 2026, this post argues that mechanistic evidence has moved the microplastics-cancer debate beyond correlation. Marc reviews tumour tissue quantification, a 30-week chronic PET nanoplastic exposure study, cell migration and inheritance data, and an in vivo immune-subversion mechanism.", "Q: Why does Marc Deller say he has contributed to more than 400 PDB structures if only 78 are listed?\nA: The 78-entry list only covers depositions where Marc Deller is named individually as a deposition author or as an author of the primary citation. It excludes the several hundred structural-genomics entries deposited under the collective \"Joint Center for Structural Genomics (JCSG)\" author name during his years as JCSG Target Manager from 2006 to 2015, and across his career he has contributed to more than 400 PDB depositions in total.", "Q: Did Marc write about synthetic biology and climate change?\nA: Yes. \"Engineering Life: How Synthetic Biology Is Tackling Climate Change and Disease\", published 13 August 2025, argues that combining protein engineering with gene-editing tools like CRISPR lets scientists design bespoke biological machinery rather than being limited to proteins that already exist in nature.", "Q: Which harvesting mechanisms does the review compare?\nA: The review compares six mechanisms: microcapillaries, microtools, microgrippers, acoustic droplet ejection, optical traps and MEMS devices, plus plate-based in situ methods as a distinct alternative. Marc Deller and Bernhard Rupp rate each for physical damage, ease of harvesting, ease of automation, cost, data-collection compatibility, throughput, microcrystal capability and reduction of late-stage failures.", "Q: Why is crystal harvesting still done by hand?\nA: Marc Deller and Bernhard Rupp explain that protein crystals are extremely fragile and sensitive to subtle changes in osmolarity, pH, chemical environment and temperature, which makes the micromanipulation very difficult to automate. Because of that difficulty, harvesting has usually been omitted from discussions of high-throughput crystallography automation.", "Q: Why does the ~50 kDa cryo-EM size limit matter so much?\nA: Marc explains that roughly 75% of human protein-coding genes produce proteins below 50 kDa, so the traditional limit excluded most of the proteome from single-particle cryo-EM. Small proteins give a low signal-to-noise ratio that makes particle picking and alignment difficult and caps achievable resolution.", "Q: What protein expression systems has Marc Deller worked with?\nA: Marc Deller has expertise in bacterial, insect, mammalian and cell-free protein expression, as well as in designing plasmid constructs, overseeing gene synthesis at CROs, protein purification and protein engineering.", "Q: What does the JCSG Core Suite column in the Top96 formulation mean?\nA: For each Top96 condition, Marc Deller listed the corresponding condition in the Qiagen JCSG Core Suite (blocks JCSG-I to JCSG-IV) so users can cross-reference the screens. Conditions marked with a plus sign (+) were also represented in the JCSG \"Core Screen\", which later became the JCSG+ screen.", "Q: What did comparing AP33 with HCV1 reveal?\nA: Comparison with the broadly neutralizing antibody HCV1 bound to the same epitope revealed a different angle of approach to the antigen by AP33 and slight variation in the beta-hairpin conformation of the epitope. Together the two structures established two different modes of binding to E2 that antibodies adopt to neutralize diverse hepatitis C virus.", "Q: What does Marc's app server landing page look like?\nA: The mdeller.com landing page is styled as a terminal session: it opens with a `marc@mdeller:~$ whoami` prompt that answers \"Marc C. Deller, D.Phil.\", then an `ls -la apps/` listing of every app with a one-line description and GitHub link, a `top -bn1` panel of live server statistics, and a `cat links.txt` block of links back to marcdeller.com.", "Q: Did Marc Deller work on hepatitis C?\nA: Yes. Marc Deller is an author on the 2012 Journal of Virology paper reporting the structure of the hepatitis C virus envelope glycoprotein E2 antigenic site 412\u2013423 in complex with the antibody AP33, led by Leopold Kong. It was part of his viral glycoprotein structural work at Scripps.", "Q: What is Marc's most recent patent-related post?\nA: The Patents page lists a 2026 post titled \"From Bicyclic to Tricyclic: A New WRN Inhibitor Patent from Incyte\". It follows the 2025 bicyclic, tricyclic, and fused bicyclic pyrimidone WRN inhibitor patents.", "Q: What did the 2009 Structure paper conclude about the SH3b domain's role?\nA: The active site sits near the interface between the SH3b and NlpC/P60 domains, where the SH3b domain appears to help define substrate specificity rather than act as a targeting domain, so that only muropeptides with an N-terminal L-alanine can bind. The authors deduced from structure and sequence analysis, then showed experimentally, that both proteins act as gamma-D-glutamyl-L-diamino acid endopeptidases.", "Q: What is the PDB code for the HCV E2\u2013AP33 structure?\nA: The atomic coordinates for the AP33 complex with HCV E2 antigenic site 412 to 423 were deposited in the Protein Data Bank under identification code 4G6A. The paper is TSRI manuscript number 21874.", "Q: What is PDB entry 2IAY?\nA: 2IAY is a DUF1831 family protein, LP2179, from Lactobacillus plantarum at 1.20 \u00c5 in space group P 21 21 21, with 114 polymer residues and 13.0 kDa, released on 10 October 2006 and keyworded TRANSLATION. It is the first representative of Pfam family PF08866 and the second-sharpest structure in Marc Deller's list.", "Q: Which paper describes PDB 5C7K?\nA: It was published as Kong, L., et al. (2015), \"Complete Epitopes for Vaccine Design Derived from a Crystal Structure of the Broadly Neutralizing Antibodies PGT128 and 8ANC195 in Complex with an HIV-1 Env Trimer\", Acta Crystallographica Section D, volume 71, pages 2099-2108. Marc Deller lists it with the structure on his website.", "Q: Why does Marc say he picks the \"ugly\" targets?\nA: Marc Deller writes that determining 400+ protein structures has meant repeatedly choosing the \"ugly\" targets, because those are often where the most important biology hides. He says this requires a bias toward action and acceptance that many attempts will not work the first, second or even tenth time.", "Q: What problems do these semi-automated systems solve?\nA: Marc lists handling crystals under 10 \u03bcm, harvesting from viscous media, improving cryoprotection consistency, reducing dehydration and increasing speed. Where automation is deployed it delivers repeatability, higher throughput and the ability to handle microcrystals beyond human dexterity.", "Q: Tell me about Marc Deller's 2026 paper on povorcitinib.\nA: The 2026 Journal of Medicinal Chemistry paper is titled \"Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\". Marc Deller is a co-author alongside Zhuo J, Li YL, Qian DQ, Burns DM, Fenalti G, Epling LB, Yao W and others, and the DOI is 10.1021/acs.jmedchem.5c03753.", "Q: Who led the gp120 glycan supersite study?\nA: The 2013 NSMB gp120 supersite paper was first-authored by Leopold Kong, with Dennis R. Burton, Andrew B. Ward and Ian A. Wilson as the corresponding authors. Project design was credited to L. Kong, K.J. Doores, J.-P. Julien, R.L. Stanfield, D.R. Burton, A.B. Ward and I.A. Wilson, and the work involved twenty-seven authors across The Scripps Research Institute, the IAVI Neutralizing Antibody Center, Imperial College London, Weill Cornell, the IAVI Design and Development Laboratory, the Academic Medical Center Amsterdam and the Joint Center for Structural Genomics.", "Q: Which chemical series appear in Marc's WRN patents?\nA: The listed WRN patents cover bicyclic compounds, tricyclic compounds, fused bicyclic pyrimidones, and imidazo[1,2-a]pyrimidine-5(1H)-ones both with and without a covalent warhead at the 3-position. All are dated 2025 on his Patents page.", "Q: How many PDB depositions do the 96 Top96 conditions account for?\nA: The 96 conditions in the JCSG Top96 screen were responsible for 631 JCSG structures deposited in the PDB between 1/1/2001 and 9/23/2010, according to Marc Deller's formulation sheet.", "Q: How many researchers did Marc Deller equip at Stanford?\nA: At Stanford's Macromolecular Structure Knowledge Center (2015\u20132018), Marc Deller equipped 70+ researchers with advanced protein-characterization tools, drove platform adoption and synchrotron utilization at SSRL, and supported users from 10+ countries.", "Q: What is Marc's typing speed?\nA: I have no idea! Given how much he writes, probably respectable. That's a guess, though, and I'd rather flag it as one.", "Q: What is the soluble cleaved HIV-1 envelope trimer structure?\nA: Marc Deller's post \"2013 \u2013 Crystal structure of a soluble cleaved HIV-1 envelope trimer\" covers the BG505 SOSIP.664 gp140 structure published by Julien and colleagues in Science in 2013, which captured the envelope glycoprotein in its near-native prefusion state at 4.7 \u00c5 resolution in complex with the broadly neutralizing antibody PGT122.", "Q: How many options does FlexAppeal expose?\nA: FlexAppeal exposes 115 OpenMM parameters across 14 groups, each with a full sentence explaining why its default is what it is. It also gives live estimates of atom count, trajectory size and wall time as the form is filled in, and validation that knows the physics, such as padding versus cutoff and timestep versus constraints.", "Q: What enzymes does Elora's lead therapy use?\nA: Elora's lead therapy combines two engineered enzymes: PETase (polyethylene terephthalate hydrolase) and MHETase (monohydroxyethyl terephthalate hydrolase). Per the deck, they work together to degrade PET microplastics in blood and tissues down to terephthalate (TPA) and ethylene glycol (EG), which the body clears through natural excretory pathways.", "Q: What matters most to Marc in life?\nA: Family, friends and animals; being there for friends when they need you; supporting others in need; being kind; giving back; and listening.", "Q: What is the \"magic chloro\" effect that Marc writes about?\nA: In \"The Unsung Hero: How a Single Chlorine Atom Transforms Drug Discovery\", published on 15 October 2025, Marc Deller explains the magic chloro effect: replacing a single hydrogen with chlorine can improve drug potency by anywhere from 10-fold to an astonishing 100,000-fold while also optimising clearance, half-life and bioavailability. He notes that over 250 FDA-approved drugs contain chlorine, making it the fifth most common element in drugs after carbon, hydrogen, nitrogen and oxygen.", "Q: What is Marc Deller's phone number?\nA: Marc Deller's phone number is (302) 358-6093.", "Q: Where is the FGFR2 Compound 29 post?\nA: The post \"FGFR2 Kinase Domain in Complex with a Pyrazolo[1,5-a]pyrimidine Analog (Compound 29)\" is in the structures section of marcdeller.com. You can read it at https://marcdeller.com/fgfr2-kinase-domain-in-complex-with-a-pyrazolo15-apyrimidine-analog-compound-29/.", "Q: What is AnythingLLM, according to Marc Deller's blog?\nA: In \"Unleashing the Power of AnythingLLM: Running Large Language Models Locally\", published 20 August 2025, Marc Deller describes AnythingLLM as an all-in-one, no-code AI application for retrieval-augmented generation, AI agents and private large-scale document analysis that can run entirely on your own hardware.", "Q: Summarize Marc Deller's post \"Hidden Chemistry in the Protein Data Bank: What Else Are We Missing?\"\nA: In this 11 August 2025 post, Marc Deller discusses how researchers at the University of G\u00f6ttingen built a computational pipeline combining machine learning and quantum mechanics to re-analyze more than 86,000 structures in the Protein Data Bank. The algorithm uncovered previously undetected nitrogen-oxygen-sulfur (NOS) covalent linkages hiding in numerous proteins.", "Q: What problems is OS3D built to solve?\nA: DeepCovalent says OS3D was purpose-built to address the central bottlenecks in early discovery: slow, capital-intensive R&D workflows, and the more than 90 % clinical failure rate driven by incomplete biological and ADMET understanding. Its answer is rapid, constraint-aware molecule generation with significantly higher downstream success.", "Q: Did the review give an example of an additive that improved crystal quality?\nA: Yes. Marc Deller's review describes how Thermofluor identified magnesium ions as stabilizing the enzyme DapD from Mycobacterium tuberculosis, and adding magnesium chloride to the crystallization solution produced larger crystals, with two tightly coordinated magnesium ions later found at the homotrimer interface in the structure.", "Q: How does the Asn332 supersite differ from the CD4 binding site as an antibody target?\nA: The paper contrasted the two directly. For antibodies whose footprints overlap the CD4 binding site, small differences in angle and relative position produce clashes with neighbouring gp120s and hugely different neutralisation outcomes \u2014 from VRC01 with 93% breadth down to the non-neutralising F105. The Asn332 supersite, by contrast, tolerates multiple binding modes and varied angles of approach without loss of neutralisation.", "Q: Tell me about the key numbers in the povorcitinib paper.\nA: Povorcitinib has a JAK1 IC50 of 11 nM, a JAK2 IC50 of 505 nM (46-fold selectivity), a cellular INA6 IC50 of 97 nM, a whole blood IL-6 IC50 of 405 nM, a Caco-2 permeability of 4.5 \u00d7 10\u207b\u2076 cm/s, and 100% oral bioavailability in cynomolgus monkey and dog. Its crystal structures with JAK1 and JAK2 were determined to 1.54 \u00c5 (PDB 10PI) and 1.59 \u00c5 (PDB 10PJ).", "Q: Why does the HCV E2 AP33 structure matter for vaccines?\nA: Marc Deller's post explains that the structural insights are critical for pan-genotypic hepatitis C virus vaccine development, since the 412-423 site is broadly conserved and targeted by neutralising antibodies. Knowing the two binding modes helps guide immunogen design.", "Q: What is DeepCovalent's tagline?\nA: DeepCovalent's home page is headed \"Discovering New Drugs in One Step\", and the site as a whole is titled \"DeepCovalent: Transforming Drug Discovery Today\". The about page opens with the statement that the company believes new drug discovery can happen in a single step.", "Q: What was different about the design philosophy of JCSG Top96 compared to earlier rational screens?\nA: Marc Deller's poster explains that previous rationally designed screens, including the JCSG's own JCSG+, focused mainly on the propensity to grow crystals rather than on the diffraction properties of the resulting crystals. JCSG Top96 was instead built from conditions that actually yielded deposited PDB structures, so the selection criterion was successful structure determination rather than crystal growth alone.", "Q: How does a user run a DRGSCROLL job, as Marc describes it?\nA: Marc describes an intuitive web workflow where the user uploads a protein structure in PDB format and submits a ligand coordinate file, after which side-chain conformations adapt organically under the algorithm's fitness function. The output is a refined protein\u2013ligand complex, produced without any manual rotamer pre-selection.", "Q: How does INCB159020 bind to KRAS G12D?\nA: A model of compound 23 built on the accumulated crystallographic data shows the key interactions from compounds 3 and 5a are conserved: the bicyclo[2.1.1] amine and the cyanoethyl group both engage a key structural water, and the carbonyl of the morpholinone hydrogen bonds to the backbone N\u2013H of Asp12. The methylpyrazine makes no observed interactions with the protein, although a water-mediated contact cannot be ruled out, and the expected ground-state conformation of the side chain closely mimics the bound conformation seen for the ethyl amide of compound 5a.", "Q: How many papers has Marc published?\nA: Marc has authored or co-authored more than 60 publications, including papers in Science, Nature Structural & Molecular Biology, Immunity, Structure, PLoS Pathogens, PLoS ONE, J. Med. Chem. and Acta Crystallographica.", "Q: What is oncostatin M and why is it interesting?\nA: Oncostatin M (OSM) is a secreted 252-amino-acid cytokine that was originally isolated because it suppresses the proliferation of human melanoma cells. As described in Marc Deller's 2000 paper, OSM inhibits the growth of a variety of tumour-derived cell lines, induces proliferation of other cells such as Kaposi sarcoma-derived spindle cells and embryonic haemangioblasts, and is a mediator of inflammatory responses in joint destruction and central nervous system inflammation. It signals through the shared receptor gp130 together with either the LIF receptor (LIFR) or the OSM-specific receptor (OSMR), making it a potential target for cancer and anti-inflammatory therapeutics.", "Q: Which paper reports the KRAS G12D inhibitor structures?\nA: Both KRAS G12D structures are described in Deller, M.C., et al. (2025), \"Discovery of an Orally Bioavailable KRAS G12D Inhibitor\", Journal of Medicinal Chemistry, volume 68, pages 1924-1939. Marc Deller is first author on that paper.", "Q: How does the paracetamol work connect to Marc's microplastics interests?\nA: Both rest on enzymatic depolymerization of PET into terephthalic acid: Marc's own therapeutic work uses PETase to break PET down inside the body, while the Edinburgh work takes the same monomer and feeds it into a bacterial metabolic pathway to make a drug. The post is at https://marcdeller.com/from-waste-to-wonder-how-bacteria-transform-plastic-into-life-saving-medicine/.", "Q: Which posts on Marc's site cover engineered PETase variants?\nA: The Microplastics page lists \"Engineering Evolution: How FAST-PETase and Other Variants Are Transforming Plastic Biodegradation\", \"HGMP01 and FAST-PETase: Decoding Two Paths to Plastic Degradation\", \"Plastic-Eating Enzymes Decoded\u2014PETase's Structure and Mutation Map\", and \"The Enzyme Toolbox\u2014Alternatives to PETase for Microplastic Breakdown\".", "Q: Summarize the key numbers from the 2015 TrkA gating ring paper.\nA: Marc Deller's 2015 PLoS ONE paper reported three structures: the TM1088 octamer (PDB 3L4B, 3.45 \u00c5, space group P3(2)21, Rcryst/Rfree 25.9%/30.8%), TM1088A (PDB 2G1U, 1.50 \u00c5, P2(1), Rcryst/Rfree 15.3%/17.5%) and the TM1088B C-terminal domain (PDB 3JXO, 1.55 \u00c5, C222(1), Rcryst/Rfree 15.8%/18.7%). The octamer contains four TM1088A and four TM1088B chains (about 160 kDa), measures 127 x 101 x 53 \u00c5, buries nearly 18,000 \u00c5\u00b2 of surface across its interfaces, and has dimer-hinge angles of about 120\u00b0.", "Q: What is the chapter's overall message?\nA: Its message is that structural biology sits at the forefront of biomedical research and drug discovery, so the community must get serious about effective ligand validation, with effort at all levels to keep deposited protein\u2013ligand structures at the highest quality. The authors summarize their stance with the maxim from Marc Deller and Bernhard Rupp's earlier paper: trust, but verify.", "Q: What resolution was the KRAS G12D Compound 23 structure?\nA: The KRAS G12D complex with Compound 23 was solved by X-ray diffraction at 1.35 \u00c5, a very high resolution for a drug discovery co-crystal structure. It was released to the PDB in January 2025.", "Q: Which of Marc's papers on the HIV-1 envelope are most cited?\nA: Two 2013 papers lead: \"Crystal structure of a soluble cleaved HIV-1 envelope trimer\" in Science with 1,027 citations, and \"Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\" in Nature Structural and Molecular Biology 20(7):796-803 with 443. Later HIV work includes the 2016 Immunity paper on VRC01-class antibodies (98 citations), the 2015 Acta Crystallographica Section D PGT128 and 8ANC195 epitope paper (74) and the 2017 PLoS Pathogens 10E8 paper (85).", "Q: Tell me about the acknowledgements in Marc's 2016 stability review.\nA: The work was sponsored in part by contributions from the k.-k. Hofkristallamt in Vista, California and by Austrian Science Fund project P28395-B26. Marc Deller and his co-authors thanked members of the Joint Center for Structural Genomics for fruitful discussions and for contributing the Thermofluor and DXMS data used in the figures.", "Q: Which companies does Marc list as commercializing protein design?\nA: Marc lists Latent Labs, whose Latent-X model reports 91 to 100 percent lab hit rates for macrocycles and picomolar affinities for mini-binders through a no-code browser interface; Cradle, which reports 1.2 to 12 fold faster R&D cycles and counts six of the top twenty-five global pharmas as customers; Levitate Bio, spun out of Cyrus Biotechnology and David Baker's lab; Tamarind Bio, which aggregates over 250 models including AlphaFold, RFdiffusion, ProteinMPNN and GROMACS; France's Amineo; and AminoAnalytica, whose agentic interface Amina turns a natural-language brief into an executed pipeline.", "Q: Explain Marc Deller's microplastics patent in simple terms.\nA: In simple terms, tiny plastic particles have been found inside human blood and tissues, and there is currently no medicine that removes them once they are in the body. Marc Deller's patent proposes using specially engineered enzymes\u2014biological scissors\u2014that chop these plastics into smaller pieces the body can safely flush out through the liver and kidneys. It also describes how to package and deliver those enzymes, either as pills or injections, so they survive and work inside a person.", "Q: Why does Marc Deller think RNA is an exciting drug target?\nA: In \"The RNA Revolution: Designing Drugs for a New Class of Targets\", published 13 August 2025, Marc Deller argues that RNA folds into complex three-dimensional structures with unique pockets and surfaces that small molecules can target, and that targeting RNA lets clinicians intervene before a harmful protein is even made.", "Q: How is AlphaFraud hosted?\nA: On a single ~3.8 GB DigitalOcean droplet with nginx and certbot: a weekly systemd-timer pipeline and an always-on Flask/gunicorn app share one SQLite file. The stack is Python, biotite, tmtools, numpy, scipy and Plotly, with no cluster and no external database.", "Q: What is a polysaccharide monooxygenase?\nA: Polysaccharide monooxygenases (PMOs, also called lytic polysaccharide monooxygenases or LPMOs) are secreted copper metalloenzymes that oxidatively degrade polysaccharides such as cellulose in a copper-, oxygen- and reductant-dependent manner. As Marc Deller's 2017 paper explains, cellulose-active fungal PMOs (the AA9 family) use a flat protein surface around the active site to bind crystalline cellulose, and they are the only enzymes known to attack bulk cellulose without first separating a glucan chain from the matrix.", "Q: What is the overall conclusion of Marc Deller's thesis?\nA: Marc Deller concluded that the OSM structure reveals several novel features unique among gp130-family cytokines (the helical B-C and A-B loops and the severe kink in helix A) while supporting the common four-helix bundle themes, and that these subtle differences likely modulate receptor engagement. Mutagenesis showed that site 2 interactions with gp130 and site 3 interactions with LIFR dominate the binding energy, with the classic site 1 being less important, and that the interfaces are less hydrophobic than in the GH and Epo systems. He argued that the promiscuity of gp130 and LIFR requires novel binding motifs, a more flexible B-C/F-G loop mode of receptor engagement, and possibly electrostatic and entropic contributions, and that structural studies of higher-order complexes, especially the elusive site 3 interaction with LIFR, are needed to complete the picture.", "Q: Did Marc Deller design the compounds in this paper?\nA: No; the paper states that the compounds were designed and synthesized by J.R. Hummel, K. Xiao, J.C. Yang, and M. Xu, with the manuscript prepared by Hummel and contributions from all authors. Marc Deller's contribution was the structural biology, specifically determining the CDK2 co-crystal structure of compound 17 with Leslie Epling.", "Q: Which organism did the MtPMO3* polysaccharide monooxygenase structure come from?\nA: The MtPMO3* structure (PDB 5UFV) came from the thermophilic fungus Thermothelomyces thermophilus ATCC 42464, also known as Myceliophthora thermophila. Marc Deller's post describes the macromolecule as a glycoside hydrolase family 61 protein solved at 2.45 \u00c5.", "Q: What did the BG505 SOSIP.664 trimer structure show?\nA: It showed the prefusion state of gp41, the gp120/gp41 interface, and the stabilising V1/V2/V3 apex architecture of the native-like Env trimer. Marc Deller's post describes it as a foundational blueprint for structure-based HIV vaccine design.", "Q: What does DeepCovalent's co-development model cover?\nA: In DeepCovalent's Flexible Co-Development Model the company focuses on AI-driven molecular design and early discovery, with shared ownership of molecular intellectual property, while retaining the flexibility to extend into compound synthesis, preclinical validation and IND-enabling studies under co-development agreements.", "Q: Where is Marc's post about publications and communication?\nA: The post \"Communication: 60+ publications translating structural data into actionable insights\" was published on 29 March 2026 in the 6cs category and is available at https://marcdeller.com/communication-60-publications-translating-structural-data-into-actionable-insights/.", "Q: What programs did Marc use to produce the structural figures and comparisons in Chapter 6?\nA: Marc Deller used MOLSCRIPT, BOBSCRIPT and RASTER3D for ribbon and stereo figures, DSSP and HERA for secondary structure and hydrogen-bond assignment, NACCESS for solvent accessibility, SHP for structural superpositions and RMSD calculations, VOLUMES for solvent-accessible surfaces, ESPript for sequence alignments, and SHARP, SOLOMON and CNS for phasing, solvent flattening and refinement of the 2.2 \u00c5 MAD electron density map. Figure 6-11 also memorably featured a Plasticene model of the proposed OSM\u2013LIFR\u2013gp130 complex, credited to \"Plasticene (Toys R Us) and Anglepoise lighting\".", "Q: What did the photoacoustic study achieve in mice?\nA: Marc reports that the group resolved individual particles roughly the width of a human hair and followed their distribution over months rather than days \u2014 a timescale that matches the chronic, low-dose nature of real human exposure far better than a single destructive snapshot.", "Q: How do you run Site Health Inspector?\nA: `bash SiteHealthInspector.sh` scans the default target, `bash SiteHealthInspector.sh https://example.com` scans a specific URL, or make it executable with `chmod +x` and run `./SiteHealthInspector.sh https://your-site.com`. It follows redirects and accepts both https:// and http:// targets.", "Q: Which small molecules are cited as stabilizing or destabilizing protein\u2013protein interactions?\nA: Marc Deller's review cites navitoclax, which destabilizes the interaction between Bcl-2 and Bad, Bid and Bak, and nutlin-3, which blocks p53 binding to MDM2. On the stabilizing side it cites cyclosporin A, which stabilizes the calcineurin\u2013cyclophilin interaction, and tafamidis, which binds at the transthyretin dimer interface and prevents the misfolding underlying transthyretin amyloidoses.", "Q: What assays were used to measure KRAS G12D inhibition?\nA: The primary biochemical assay measured exchange of a fluorescently labelled GDP analogue for a nonhydrolyzable GTP analogue in the presence of SOS1. Surface plasmon resonance was used as a more sensitive direct-binding readout once compounds bottomed out the exchange assay. Cellular potency was measured by homogeneous time-resolved fluorescence detection of ERK phosphorylation in the KRAS G12D-dependent HPAFII line, and a human whole blood pERK assay was used as the translational benchmark.", "Q: Who were the authors of the crystal harvesting bottleneck poster?\nA: The poster was authored by Marc C. Deller, Jace Walsh, Robert Viola, Marc A. Elsliger, Ian A. Wilson and Bernhard Rupp. Deller, Elsliger and Wilson were at the Joint Center for Structural Genomics at The Scripps Research Institute; Walsh and Viola were at Square One Systems Design in Jackson, Wyoming; and Rupp was at the k.-k. Hofkristallamt in Vista, California.", "Q: How did the paper explain PGT 135's neutralisation breadth?\nA: The team analysed 3,045 aligned gp120 sequences from the Los Alamos HIV Database. The glycans at Asn332, Asn392, Asn295 and Asn386 and the residue His330 are 73%, 79%, 59%, 87% and 71% conserved respectively; His330 with Asn332 and Asn392 occur together in 50% of sequences, and His330 with all four glycan motifs in 26%. The observed breadth of 33% sits between these values, suggesting that the dispensability of either Asn295 or Asn386 in certain strains enhances breadth.", "Q: Which mutagenesis assays were used in the oncostatin M study?\nA: In the 2000 Structure paper, alanine-scanning mutants of oncostatin M were tested in solid-phase competition assays for their ability to inhibit binding of biotinylated hOSM to immobilised human gp130\u2013Fc or LIFR\u2013Fc ectodomains, and in Ba/F3 cell-survival bioassays using cells co-transfected with either human LIFR/gp130 or human OSMR/gp130. Activities were expressed as IC50 or EC50 ratios relative to wild-type hOSM, and the mutants were selected on the basis of solvent exposure in Marc Deller's crystal structure.", "Q: What is AlphaFraud?\nA: AlphaFraud is described on Marc Deller's Vibe Coding page as a weekly watcher that compares newly deposited human PDB structures with their blind AlphaFold predictions, framed as a running test of whether protein folding is solved. It is hosted at alphafraud.mdeller.com.", "Q: Summarise Marc's compassion post.\nA: The compassion post argues that science advances best when empathy and excellence advance together, illustrated by 6+ FDA IND filings supported alongside the mentoring of 20+ scientists. Marc Deller published it on 29 March 2026 at https://marcdeller.com/compassion-6-fda-ind-filings-supported-while-mentoring-20-scientists/.", "Q: What is PDB entry 3DEE?\nA: 3DEE is a putative regulatory protein involved in transcription, NGO1945, from Neisseria gonorrhoeae FA 1090, solved at 2.1 \u00c5 in space group C 1 2 1 with 249 polymer residues and 29.15 kDa, released on 5 August 2008 and keyworded TRANSCRIPTION. It is the first representative of Pfam family PF09836, published in Acta Crystallographica Section F in 2010.", "Q: How selective was compound 29 across the kinome?\nA: Compound 29 maintained greater than 100-fold selectivity against all 36 structurally representative kinases in Incyte's in-house kinase panel, in addition to its 42-fold selectivity over FGFR1 and roughly 290-fold selectivity over FGFR4. Reaching that level of selectivity required deliberate SAR to shut down early off-target activity at PDGFRB, KIT, FLT3 and TRKA.", "Q: Did the CDK2 compound work in animal models?\nA: Compound 17 was evaluated in CCNE1-amplified mouse models, where it produced pharmacodynamic inhibition of CDK2, measured as reduced Rb phosphorylation, along with antitumor activity. Dysregulation of CDK2 with cyclin E1 drives tumour growth in a variety of human cancers.", "Q: What environmental numbers does Marc cite in the plastic-to-medicine post?\nA: Marc cites over 350 million tons of PET waste generated annually, around 300 million tons of pharmaceutical plastic waste that could become feedstock, and notes that traditional paracetamol manufacture from benzene and phenol contributes 55% more CO\u2082 per ton than automotive manufacturing.", "Q: What \"final pH\" values are noted in the Top96 formulation?\nA: For the grid-screen-derived conditions (MPD, PEG 6000, ammonium sulfate and PEG/LiCl grids), the formulation sheet lists both the nominal buffer pH and the measured final pH of the mixed solution, for example A03 is 10% MPD with 0.1 M Citric acid pH 2.5 giving a final pH of 4, and C06 is 20% PEG 6000 with 0.1 M Citric acid pH 4.0 giving a final pH of 5.", "Q: What are PEST sequences?\nA: PEST sequences are stretches rich in proline, glutamate, serine and threonine that often give proteins half-lives of less than two hours by targeting them for proteasomal or calpain-mediated degradation. Marc Deller's review lists them among the primary-structure features worth engineering out during construct design.", "Q: Explain the 10E8 finding to a non-scientist.\nA: HIV hides one of its most vulnerable spots right up against its own greasy outer membrane, so any antibody that reaches it has to touch the grease too. Marc Deller and colleagues froze the antibody 10E8 in crystals together with fat molecules and took X-ray snapshots, which revealed a small pocket where the antibody grips a fat head group. When they altered that pocket, the antibody still stuck to the protein target but largely stopped stopping the virus \u2014 proof that the fat is part of the target. Vaccine designers now know they need to present the protein piece attached to lipid, in the right orientation.", "Q: What example of decentralized collaboration does Marc give?\nA: Marc describes a structural biologist in one country analyzing cryo-EM data generated by a CRO in another while a computational chemist on a third continent designs new molecules from that structure, all on a shared cloud platform. He notes that platforms such as AWS and Google Cloud let teams run molecular dynamics simulations or train large AI models without owning a supercomputer.", "Q: Where can I read Marc's data and project management round-up?\nA: The post is published at https://marcdeller.com/building-data-and-project-management-expertise/, in the certificates category, and links to the Coursera verification page for each of the five courses.", "Q: What crystallisation condition finally produced the diffracting HIV-1 Env trimer crystals?\nA: The two usable crystals came from a condition of 0.1 M CAPS, 2 M ammonium sulfate, 0.2 M lithium sulfate at pH 10.5, using the EndoH-treated, size-exclusion-purified PGT122 Fab plus BG505 SOSIP.664 trimer complex. Data were collected at Advanced Photon Source beamline 23-ID-D, and because of radiation damage the data from two crystals had to be merged to obtain sufficient redundancy and completeness.", "Q: What does Marc's post on remapping chemical and pharmacological space cover?\nA: \"Remapping the Chemical and Pharmacological Spaces of Drugs: Navigating Tomorrow's Therapeutic Frontier\", published 11 October 2025, covers how AI-powered visualisations turn vast molecular libraries into navigable landscapes. It builds on the paper \"Remapping the Chemical Space and the Pharmacological Space of Drugs\", which projected ChEMBL34 molecular fingerprints into two-dimensional UMAP representations.", "Q: How does Marc say QSAR models should be evaluated?\nA: He advocates realistic data-splitting schemes tailored to the task, specifically zero-shot and few-shot splits that distinguish virtual screening from lead optimisation scenarios, following a benchmark study that separated the two use cases. His point is that naive random splits give misleadingly optimistic performance estimates.", "Q: How widely is HGMP01 expressed in humans, according to Marc?\nA: Marc reports metatranscriptomic evidence that HGMP01-like genes are actively transcribed in over 20% of surveyed human gut samples across diverse bacterial species, suggesting an underappreciated ongoing role in processing ingested microplastics.", "Q: Who is Marc's favourite artist?\nA: Pete Fowler, the illustrator and designer.", "Q: How was the LIFR DNA quality checked before transfection?\nA: Large quantities of pIG-1-LIFR-FN3-3C-Fc plasmid were prepared from E. coli strain MC101/P3 and checked by a triple restriction digest with BamHI, BglII and HindIII, run on a 1 % agarose gel. Diagnostic drop-out bands (around 2505, 1752 and 753 base pairs from the different double digests) confirmed the identity and quality of the DNA, a step Marc found essential before transfection.", "Q: How was the octameric TM1088 structure solved?\nA: Marc Deller solved the TM1088 octamer by molecular replacement in Phaser, merging data from two isomorphous crystals and using the high-resolution TM1088A structure (PDB 2G1U) as the search model. Both enantiomorphic space groups were tested, giving a clear solution in P3(2)21, after which the remaining seven molecules were located. Refinement used one TLS group for each of the twelve protein domains in the asymmetric unit with tight NCS restraints, converging at Rcryst 25.9% and Rfree 30.8%.", "Q: What are Marc's values?\nA: Marc is a supporter of freedom of thought, expression of principles and helping others in need, and he lives by the 6 C's of Curiosity, Courage, Creativity, Communication, Compassion and Collaboration.", "Q: What are the PDB codes for the KRAS G12D structures in this paper?\nA: Two structures were deposited. KRAS G12D bound to compound 3 is PDB 9E5F, and KRAS G12D bound to compound 5a, the methyl ester analogue of compound 5, is PDB 9E5D. Both were determined by the Incyte structural biology group that Marc Deller and Leslie Epling were part of.", "Q: Has Marc ever served on a grant panel?\nA: I don't have anything on panel service. For his professional roles, marcdeller.com has the record.", "Q: What was Marc's role on the HIV envelope trimer work?\nA: Marc Deller is an author on the HIV-1 envelope trimer work he features in the papers section of marcdeller.com, contributing to the structural biology behind the crystallographic study. He presents the post as part of his HIV structural virology portfolio, and it can be read at https://marcdeller.com/crystal-structure-of-a-soluble-cleaved-hiv-1-envelope-trimer/.", "Q: What is the Asn332 supersite?\nA: In his post \"2013 \u2013 Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\", Marc Deller describes the region surrounding the conserved N-linked glycan at asparagine 332, together with neighbouring glycans at positions 392 and 386, as a supersite of immune vulnerability on the otherwise heavily shielded gp120 surface.", "Q: What is the CALR antibody patent Marc writes about?\nA: Marc Deller's post \"2023 \u2013 Anti-mutant calreticulin (calr) antibodies and uses thereof\" covers patent WO2023107994A1, which discloses antibodies that specifically bind mutant calreticulin (CALR) proteins implicated in myeloproliferative neoplasms and other cancers. The antibodies may be monoclonal, chimeric, humanized, bispecific or multi-specific.", "Q: Did the commercial sparse-matrix screens work for OSM?\nA: No. Hampton Research Crystal Screen I and Crystal Screen II produced no crystals for any OSM sample at either 22 \u00b0C or 4 \u00b0C. Marc Deller's crystals came instead from a custom salt/PEG screen he devised himself.", "Q: What is different about the \u03b2-hairpin in the AP33-bound epitope?\nA: Both structures show the same type I\u2032 \u03b2-turn, but the anti-parallel \u03b2-sheet in the AP33-bound hairpin splays apart at the end distal from the turn, so only four intrapeptide hydrogen bonds stabilise the hairpin instead of the five found in the HCV1 epitope. Accordingly, AP33 buries less surface area around the termini of the peptide than HCV1 does.", "Q: What temperature and conditions are the enzymes optimized for?\nA: The enzymes are optimized for human physiological conditions, particularly a body temperature of around 37 \u00b0C, plus systemic circulation, immune compatibility, and targeted metabolic clearance. This human-physiology optimization is what separates Marc Deller's therapeutic invention from prior enzymes developed for industrial or environmental settings.", "Q: Why does Marc consider query-agnostic triggers especially dangerous?\nA: Marc explains that unlike earlier attacks, CatAttack preserves the original problem's semantics entirely and requires no domain expertise, which makes it insidious in fields such as finance and healthcare where reasoning accuracy is critical. He notes the pipeline combines a prompt optimizer, a judge model for hallucination detection and a transfer mechanism.", "Q: Which FGFR structure did Marc Deller deposit?\nA: 8E1X is the FGFR2 kinase domain in complex with a pyrazolo[1,5-a]pyrimidine analogue, compound 29, determined at 2.68 \u00c5 in space group P 21 21 21 with 660 polymer residues and 76.54 kDa, released on 23 November 2022. Its deposition authors are Lei, H.-T., Epling, L.B. and Deller, M.C., and it supports the 2022 Journal of Medicinal Chemistry paper on potent and selective inhibitors of wild-type and gatekeeper mutant FGFR2/3.", "Q: What advantages does Marc attribute to digital twins in trials?\nA: Marc says reliable digital twins let trials run with smaller, more targeted patient groups, cutting cost and duration, and minimize the number of participants needed in a placebo arm so that more people receive the experimental therapy. He frames this as making trials not only faster and smarter but more ethical.", "Q: How does Marc Deller's TM1088 structure differ from previously known gating rings?\nA: All previously determined gating rings, such as MthK, KtrA, BK, GsuK and TrkA, were built from a single type of polypeptide chain. The TM1088 assembly is the first K+ transporter structure assembled from two different, but structurally homologous, proteins. Despite this atypical two-subunit organization, Marc Deller found it shares the same folds, domains, assembly motifs and overall quaternary organization as the single-subunit gating rings.", "Q: What organisms appear most often in Marc Deller's structural genomics entries?\nA: Shewanella species appear repeatedly (3HSA, 3DCX, 3EQX, 2Q3L, 2IIZ, 2HAG, 2RA9, 3BOS), as do Bacteroides thetaiotaomicron (3KK7, 3HBZ, 3GF8, 3CGH, 2GVK) and Thermotoga maritima (1VLQ, 3M81, 3M82, 3M83, 2Q8U, 3JXO, 3L4B). Others include Bacillus cereus, Neisseria gonorrhoeae, Klebsiella pneumoniae, Exiguobacterium sibiricum and Silicibacter pomeroyi.", "Q: Where is Marc Deller based and how can he be contacted?\nA: The home page of marcdeller.com gives his location as Wilmington, Delaware, USA, a telephone number of (302) 358-6093, and an e-mail contact link. Those details appear in the footer contact block.", "Q: How does Marc describe his own role in big collaborative builds?\nA: In the collaboration post, Marc Deller says his role has often been to connect the many threads of a build, spanning scientists, engineers, operations and leadership, into a coherent vision and then execute it step by step. He says building comprehensive research facilities requires bringing together international expertise from both industry and academia.", "Q: How does the Hda structure differ from DnaA?\nA: Hda resembles typical AAA+ ATPases, but the arrangement of its two subdomains, residues 1 to 174 and 175 to 241, differs dramatically from that of DnaA. A CDP molecule anchors the two Hda domains in a conformation that promotes dimer formation.", "Q: Was the novel heme conformation in TyrA seen in other DyP enzymes?\nA: Structural comparison to the canonical dye-decolorizing peroxidase family member DyP from Thanatephorus cucumeris (Dec 1) demonstrated conservation of the novel heme conformation as well as of residues important for heme binding. Marc Deller was a co-author on this 2007 Proteins paper (DOI 10.1002/prot.21673).", "Q: What does GOBSMACKED stand for?\nA: GOBSMACKED stands for Ground-truth Overlay for Binding Sites, Modes And Complex Kinetics/Dynamics. Marc wrote it to answer three separate questions about one prediction: how close the pose lands to the crystal (PIER REVIEW), whether molecular dynamics recovers the induced fit an apo-like predicted pocket is missing (HOLOGRAM), and whether the implied binding mode is the one the crystal shows (GATEKEEPER).", "Q: Which paper reported the extended polyketide synthase module?\nA: The work was published as Li, X., et al. (2018), \"Structure-Function Analysis of the Extended Conformation of a Polyketide Synthase Module\", Journal of the American Chemical Society, volume 140, pages 6518-6521. Marc Deller has a companion blog post about that paper on his site.", "Q: What COVID-19 and diagnostic examples does Marc give for nanobodies?\nA: He describes Nanosota-2, -3 and -4 developed at the University of Minnesota against SARS-CoV-2 variants, with Nanosota-2 achieving an IC50 of 2 picomolar against the prototypic strain, among the most potent entry inhibitors discovered. On diagnostics he cites whole-cell biosensors using nanobody-displayed yeasts detecting SARS-CoV-2 spike protein down to 0.037 \u03bcg/mL, and notes a nanobody-based CD7-targeted CAR-T therapy reaching 70% complete remission in refractory acute myeloid leukaemia. The post is at https://marcdeller.com/nanobodies-revolutionizing-molecular-medicine-through-natures-smallest-antibodies/.", "Q: What function was proposed for the SpoIIAA-like Shewanella proteins?\nA: The authors suggested that both proteins may associate with the lipid bilayer in their open monomeric state by inserting the amphiphilic alpha2 and alpha3 helices into the membrane. They may therefore function as carriers of nonpolar substances or as interfacially activated enzymes, and they belong to a novel bacterial family adopting the alpha/beta SpoIIAA-like fold found in STAS and CRAL-TRIO domains.", "Q: What is Marc Deller's earliest published review?\nA: \"Cell surface receptors\", written by Marc Deller with E. Yvonne Jones, was published in Current Opinion in Structural Biology in 2000. It reported that the database of ligand-receptor complexes had essentially doubled in the preceding year, extending views of cell surface receptor structure and function.", "Q: What new interactions did the PGT128 and 8ANC195 structure reveal for 8ANC195?\nA: Beyond the previously documented gp120 contacts, the structure elucidated a substantial 8ANC195 interface with gp41 that includes extensive interactions with the N637 glycan. Surprisingly, removing the N637 glycan did not affect 8ANC195 affinity, suggesting the antibody has evolved to accommodate that glycan without losing binding energy.", "Q: Is Marc Larry David?\nA: No. Marc likes the clothing brand Simple, but he is not Larry David.", "Q: How does Marc describe the PET binding cleft?\nA: Marc describes an L-shaped substrate binding cleft roughly 40 \u00c5 long that accommodates up to four MHET moieties. Subsite I binds the first unit through \u03c0-\u03c0 stacking with Tyr87 and Trp185, and subsite II extends across regions IIa, IIb and IIc, binding three more units through hydrophobic contacts involving Trp159, Ile232, Asn233, Ser238, Asn241 and Asn246.", "Q: What are some of the ten LLM workflows Marc describes?\nA: They include motif discovery from a single protein sequence, protein\u2013protein interaction prediction from sequence pairs, secondary structure risk profiling, analog design for improved drug-likeness, pathway mapping from experimental annotations, rapid toxicity profiling, patent and prior-art searching, synthesis route analysis, enzyme substrate promiscuity prediction, and structured extraction of data such as IC50 values from literature abstracts.", "Q: What were the refinement statistics for the DRVIA7 structures?\nA: For the unliganded DRVIA7 Fab, the final Rcryst and Rfree values were 23.4% and 28.4%; for the gp120-bound DRVIA7 Fab complex they were 25.3% and 30.6%. Both structures were solved by molecular replacement.", "Q: Which Boltz flags does run_boltz.zsh set?\nA: It sets `--use_potentials` for potential energy guidance during diffusion, `--diffusion_samples 1`, `--use_msa_server` to query the ColabFold MSA server, `--num_workers 8`, and `--accelerator gpu`. It also sets `PYTORCH_ENABLE_MPS_FALLBACK=1` for Apple Silicon compatibility.", "Q: Why is the G120Y mutation of OSM so much more disruptive than G120A?\nA: Marc Deller explained this using his model of the OSM\u2013gp130 site 2 interface built on the Epo\u2013EpoR structure. Gly 120 sits at the centre of the site 2 epitope; alanine substitution reduced gp130 binding by over 100-fold but only moderately reduced biological activity (about 60-fold), whereas a tyrosine at this position would sterically clash with Phe 169 of gp130 and/or the receptor's B-C loop, abolishing both binding and activity (over 10,000-fold loss). He noted a parallel with growth hormone, where mutation of the equivalent Gly 120 sterically blocks formation of the 1:2 hormone\u2013receptor complex.", "Q: How big are TM1088A and TM1088B, and what is the stoichiometry of the assembly?\nA: TM1088A is a 16.0 kDa protein containing a single N-terminal Trk domain (Pfam PF02254), and TM1088B is a 24.0 kDa protein containing an N-terminal Trk domain (PF02254) plus a smaller C-terminal Trk domain (PF02080). In Marc Deller's structure they assemble as an octamer of four TM1088A and four TM1088B chains, giving a complex of about 160 kDa.", "Q: What are the FDA's four validation principles as described by Marc?\nA: Marc lists Context of Use, requiring a clear definition of the NAM's intended regulatory purpose; Human Biological Relevance, showing the model can assess toxicity in a human-representative system; Technical Characterisation, establishing confidence through robust and reproducible methods; and Fit-for-Purpose, assuring the NAM can assist CDER in regulatory decision-making.", "Q: How does ButtFold get its trajectories?\nA: ButtFold has three sources feeding one shared player: a gallery of six precomputed G\u014d folds and three Genie 2 backbones that needs no compute, a live fold running the same C compiled to WebAssembly in a Web Worker on the user's CPU, and an ESMFold path where Meta predicts a real UniProt protein and the same C, compiled natively, folds a chain toward it. All three produce byte-identical frame objects, which is tested rather than asserted.", "Q: What does \"synthetic lethality\" mean in the context of the WRN patent?\nA: Synthetic lethality means that losing one gene is tolerated on its own, but losing it in the context of a second defect kills the cell. The WRN patent describes exactly this: cancer cells that are already microsatellite instability-high or mismatch repair deficient become critically dependent on WRN helicase activity, so a WRN inhibitor should kill those tumour cells while leaving normal, microsatellite-stable cells largely unaffected.", "Q: Which computational method was used to analyse the AP33 epitope surfaces?\nA: The adaptive Poisson-Boltzmann solver (APBS) was used to calculate the surface potential across the solvent-accessible surfaces of both the paratope and the epitope, plotted from \u22123 kT/e to +3 kT/e so that the results could be compared directly with the earlier HCV1 antibody interaction.", "Q: Which glycans does PGT 135 contact on gp120?\nA: PGT 135 interacts with the N-linked glycans at Asn332, Asn392 and Asn386. In the crystal structure the Asn392 glycan contributes 547 \u00c52 of buried surface and the Asn332 glycan 365 \u00c52, while the Asn386 glycan contributes far less (99 \u00c52, only two van der Waals interactions with CDR H1) yet is still critical for neutralisation of several HIV-1 strains.", "Q: Where can I read Marc's fused bicyclic pyrimidone post?\nA: It was published on 10 December 2025 at https://marcdeller.com/fused-bicyclic-pyrimidones-as-wrn-inhibitors/, linking to Google Patents for US20250333420A1.", "Q: What is Thermofluor and why does Marc's review recommend it?\nA: Thermofluor, also called differential scanning fluorimetry, measures a protein's melting temperature using a reporter dye such as SYPRO Orange or ANS that changes fluorescence on binding the exposed hydrophobic core of unfolding protein. Marc Deller's review highlights that it needs very little protein, typically well under 1 mg, and only a standard qPCR machine, making it ideal for high-throughput buffer, ligand and construct screening.", "Q: Which studies does the \"Future Therapeutics\" post cite?\nA: The post cites a 2023 review of PETase engineering in Frontiers in Bioengineering and Biotechnology, the 2024 Faraday Discussions paper on PETase degradation of PET in serum, a 2025 Nature Medicine study on bioaccumulation of microplastics in human brains, coverage of apheresis for microplastic removal, and a 2024 review of the micro(nano)plastics health burden.", "Q: Which decade did most of Marc Deller's structure releases fall in?\nA: The bulk of the releases fall between 2005 and 2010, the JCSG structural genomics years, with 2010 particularly dense. The list then thins out through the Scripps HIV work of 2012 to 2017 and the Stanford entries of 2017 to 2019, before the Incyte drug discovery structures of 2022 to 2026.", "Q: What crystallisation conditions gave the OSM crystals used for structure determination?\nA: Marc Deller obtained the primary \u0394188_OSM crystals within 1\u20133 days at 22 \u00b0C by hanging drop vapour diffusion using a custom salt/PEG screen. The drops contained 2 \u03bcl of protein at ~8 mg/ml plus 2 \u03bcl of reservoir solution of 30 % (w/v) PEG 35000 and 0.25 M sodium acetate, pH 7.5. The largest crystals, roughly 0.2 \u00d7 0.2 \u00d7 0.1 mm, came from drops of 0.5 \u03bcl precipitant and 1 \u03bcl protein at 8 mg/ml.", "Q: Who wrote the LP2179 structure paper?\nA: Bakolitsa C was first author, with Kumar A, Carlton D, Miller MD, Krishna SS and a large Joint Center for Structural Genomics team including Marc Deller, Scott Lesley and Ian Wilson. Marc Deller's contribution was crystallography within the JCSG structural genomics pipeline.", "Q: Can you help me plan a trip to Italy?\nA: Trip planning isn't something I can help with, but if you're ever curious about Marc Deller's travels through his career, from Nottingham to Oxford to San Diego to Wilmington, I can walk you through that.", "Q: What is PDB entry 2PV7?\nA: 2PV7 is the chorismate mutase / prephenate dehydrogenase (TyrA) from Haemophilus influenzae Rd at 2.00 \u00c5 in space group P 41 21 2, with 596 polymer residues and 70.24 kDa, released on 22 May 2007 and keyworded ISOMERASE, OXIDOREDUCTASE. It is cited in the 2010 Acta Crystallographica Section F paper on the unique features of bifunctional TyrA enzymes.", "Q: How does the HIV-1 Env trimer compare to influenza haemagglutinin?\nA: The 2013 Science paper noted that the central trimeric coiled-coil of the fusion protein gp41 surrounded by three receptor subunits (gp120) is a hallmark of type-1 membrane-fusion proteins, and extended that hallmark to the capping of the central helices by another helix in both gp120 and influenza HA. Both proteins also show a small inter-protomer central opening at the top of the central helix, as do the Ebola and PIV5/RSV fusion proteins, which presumably facilitates conformational rearrangement of the head domains during fusion.", "Q: How did Marc Deller get RFdiffusion running on an Apple Silicon Mac?\nA: In \"Installation of RFdiffusion on MacOSX M1 architecture without GPU support\", published 20 August 2025, Marc Deller documents a CPU-only install using the community mps-test fork by YaoYinYing, cloning the RosettaCommons repository, downloading the model checkpoints from files.ipd.uw.edu, creating a conda environment from SE3nv_macos.yml, and installing PyTorch 2.3.0 CPU-only, DGL 2.2.1, a mock nvtx package, the SE3Transformer fork and dllogger.", "Q: What is Marc Deller's 2024 Journal of Medicinal Chemistry paper about?\nA: The 2024 J. Med. Chem. paper co-authored by Marc Deller describes the discovery of (4-pyrazolyl)-2-aminopyrimidines as potent and selective inhibitors of cyclin-dependent kinase 2 (CDK2). The team at Incyte started from a high-throughput screening hit and used structure\u2013activity relationship studies to reach compound 17, a highly potent CDK2 inhibitor (IC50 = 0.29 nM) that is selective over CDKs 1, 4, 6, 7, and 9. The work was published in J. Med. Chem. 2024, volume 67, pages 3112\u20133126.", "Q: What is the earliest PDB structure Marc Deller deposited?\nA: The earliest is 1EVS, the crystal structure of human oncostatin M, released on 13 September 2000. It was solved at 2.2 \u00c5 in space group P 21 21 21 with 187 polymer residues and 21.14 kDa, and deposited by Deller, Hudson, Ikemizu, Bravo, Jones and Heath.", "Q: How many crystals and structures had the JCSG pipeline produced at the time of the poster?\nA: At the time of the 2013 poster, the JCSG crystallization pipeline had screened over 165,000 crystals and produced close to 1400 unique protein structures.", "Q: What does Site Doctor's diagnostic scan check?\nA: Request timing broken down into DNS, TCP connect, TLS handshake, TTFB, download and total; HTTP status, headers and redirect chain; caching behaviour; security headers including HSTS, CSP, X-Frame-Options, X-Content-Type-Options, Referrer-Policy and Permissions-Policy; cookie flags; CORS and mixed content; SSL/TLS certificate details and protocol support; DNS records; robots.txt, sitemap, favicon and custom 404 handling; redirect behaviour; server and CDN detection; and WordPress fingerprinting including version disclosure, theme, plugins, REST API, XML-RPC, readme.html, wp-login.php and exposed users via `/wp-json/wp/v2/users`.", "Q: What is PDB entry 4NCO?\nA: 4NCO is the crystal structure of the BG505 SOSIP gp140 HIV-1 Env trimer in complex with the broadly neutralising Fab PGT122, at 4.7 \u00c5 in space group C 1 2 1, with 2997 polymer residues and 350.93 kDa, released on 13 November 2013 with NAG bound. It was deposited by Julien, J.-P., Stanfield, R.L., Lyumkis, D., Ward, A.B. and Wilson, I.A., and its primary citation is \"Crystal structure of a soluble cleaved HIV-1 envelope trimer\", Science 2013, DOI 10.1126/science.1245625.", "Q: What headline statistic about approved drugs does Marc quote?\nA: Marc quotes the finding that 81% of approved drugs contain aromatic rings, which he presents as a practical guide to scaffold selection and lead optimisation. He also notes an accelerating trend toward kinase inhibitor development with epigenetic writers emerging as targets in post-2020 approvals.", "Q: Give me a recipe.\nA: Cooking is well outside what I cover, I'm afraid. I'm chatMCD and I talk about Marc Deller. Ask me about his research, his structures or Elora Therapeutics and I'll be much more use.", "Q: Why does Marc say persuasive attacks are hard to defend against?\nA: Marc explains they generate human-readable text that looks legitimate to automated filters and human reviewers, unlike optimization-based jailbreaks that produce gibberish. He adds that multi-turn strategies such as the crescendo technique and the bad Likert judge method gradually escalate requests, and that many existing defenses show alarming over-refusal rates while still failing to stop persuasion.", "Q: Why does Marc favour 3Dmol.js for science communication?\nA: Marc favours it because deployment is essentially frictionless \u2014 two lines of script plus a tagged div \u2014 while the WebGL foundation gives smooth, hardware-accelerated, publication-quality rendering with no install barrier for the reader. He also points to its BSD open-source licence driving wide adoption and continual improvement, and to classroom response systems built on 3Dmol.js as evidence of its value in active learning.", "Q: What does Marc's About page say under Collaboration and Compassion?\nA: Collaboration is described as 2+ state-of-the-art structural biology facilities built through global partnerships, and Compassion as 6+ FDA IND filings supported while mentoring 20+ scientists. Both appear in the 6C's list on the About page.", "Q: What does Marc mean by \"vibe-coder\"?\nA: Marc uses \"vibe-coder\" to describe himself as a subject-matter expert who builds scientific software with AI-assisted coding rather than as a professional software engineer. His apps such as AlphaFraud, BoltzMaker, CODSWALLOP, GOBSMACKED and ALPHABETTI show what an expert can stand up on modest hardware at minimal cost.", "Q: What is PDB entry 3HBZ?\nA: 3HBZ is a putative glycoside hydrolase, BT_2081, from Bacteroides thetaiotaomicron VPI-5482 at 2.05 \u00c5 in space group P 32 2 1, with 342 polymer residues and 41.43 kDa, released on 26 May 2009 with calcium, cacodylate and sodium bound. It is the first structural representative of a new protein family that may play a role in carbohydrate metabolism, published in Acta Crystallographica Section F in 2010.", "Q: What does Marc mean by \"programmable\" medicine?\nA: Marc uses the term for degraders that require multiple conditions to be met before they activate, arguing this could dramatically reduce off-target effects and widen the therapeutic window for potent drugs. He presents it as a crucial step toward therapies that are both powerful and exceptionally safe; the post is at https://marcdeller.com/beyond-inhibition-the-rise-of-conditional-degraders-for-safer-medicines/.", "Q: Where can I read about BoltzMaker?\nA: Marc wrote two posts, \"BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline\" at https://marcdeller.com/boltzmaker-vibe-coding-a-campaign-scale-boltz-2-pipeline/ and the project page at https://marcdeller.com/boltzmakervibe/, with the code at github.com/bellcheddar/BoltzMaker and a hosted web app at boltzmaker.mdeller.com that runs every stage except the GPU `run` step.", "Q: How does GOBSMACKED handle PLIP's GPL licence?\nA: GOBSMACKED is MIT licensed, and because PLIP is GPL-2.0 the app runs it as a subprocess and never imports it, so no GPL code is linked into the application. PLIP is also deliberately absent from the run bundle.", "Q: What did the PET microplastic gut study in mice show, according to Marc?\nA: Marc describes Sun et al. (2025), in which male C57BL/6 mice received 10 \u03bcm PET microplastics at 10, 50 and 250 mg/kg for 28 days. The exposure raised IL-2 and TNF-\u03b1, lowered IL-10, reduced Muc2, Muc3, Occludin and Claudin-1 expression, shifted the Firmicutes/Bacteroidetes ratio at 14 days and reversed it by 28 days, and raised blood glucose and liver enzymes.", "Q: What does Marc report about microplastics in inflammatory bowel disease?\nA: Citing Yan et al. (Environmental Science & Technology, 2022), Marc reports that IBD patients have 41.8 microplastic items per gram of faeces versus 28.0 items per gram in healthy controls, with polyethylene terephthalate and polyamide the dominant polymers and concentrations correlating with disease severity.", "Q: What are FragLites?\nA: FragLites are a concise library of 31 halogenated molecular fragments used to map protein binding hotspots, covered by Marc Deller in \"Unveiling FragLites: Pioneering Precision in Drug Discovery\" on 22 August 2025. They exploit the anomalous scattering of halogens so that productive drug-like interactions can be detected sensitively and unambiguously by X-ray crystallography.", "Q: Which sports teams does Marc support?\nA: Nottingham Forest in football, and the Philadelphia Phillies (baseball) and Philadelphia Eagles (American football).", "Q: What is Marc's view of whether the AI pivot resolves Sci-Hub's dilemma?\nA: Marc says whether AI rebranding changes the legal and ethical calculus remains genuinely open, but the underlying inequity has not been resolved. He points to publisher profit margins exceeding 30%, an average article processing charge of $3,400 and paywalls reaching $10,600, arguing that while publicly funded science sits behind them, researchers will keep reaching for the door Elbakyan built. The post is at https://marcdeller.com/sci-hub-at-the-crossroads-from-shadow-archive-to-ai-research-tool/.", "Q: Why does codon usage matter when expressing human proteins in E. coli?\nA: Marc Deller's review points out that the arginine codon AGA is rare in E. coli and can cause premature chain termination, frameshifting and incorrect amino-acid insertion. Remedies include synthesizing a codon-optimized gene or co-transforming with rare-tRNA plasmids such as CodonPlus or pRARE, the latter sold in BL21 cells as Rosetta and used by the Swedish Human Protein Atlas project.", "Q: How do you use Boltz2Run?\nA: Make it executable and run it with `chmod +x run_boltz.zsh` then `./run_boltz.zsh`. It picks up all inputs from `./inputs/`, writes predictions to `./output/`, and creates a timestamped log in the current directory; there are no arguments to pass.", "Q: How many enzyme therapies have been FDA approved, according to Marc?\nA: Marc cites approximately 39 unique FDA-approved enzyme therapies documented through 2021, notes six of the approved enzymes use CHO cell expression, and estimates enzyme therapies at roughly 1\u20133% of annual novel drug approvals, against a market projected to reach USD 21.62 billion by 2035 at 6.6% compound annual growth.", "Q: How is the work split up in GOBSMACKED?\nA: GOBSMACKED splits into three stages. The droplet prepares the run on CPU, resolving the input, fetching the best available structure, annotating the family, picking the pocket, choosing a reference crystal and emitting `run_bundle.tar.gz`; the user runs that bundle on their own GPU to fold, dock, minimise, run MD and score affinity; then the droplet validates the returned archive, superposes on the pocket, runs PLIP and PandaMap, grades the result and draws the trajectory.", "Q: Which portfolio papers supported IND filings?\nA: Three of the Journal of Medicinal Chemistry papers in Marc Deller's portfolio are marked as supporting IND filings: the KRAS G12D inhibitor INCB159020 (2025), the CDK2 inhibitor (4-pyrazolyl)-2-aminopyrimidines (2024), and the FGFR2/3 inhibitors (2022). Each reflects his structural biology and structure-based drug design work at Incyte.", "Q: What software tools did Marc Deller use in the engineering chapter?\nA: For the rational design work Marc used THREADER for structural/sequence alignment, MODELLER for homology modelling from the murine LIF structure, DSSP for secondary structure assignment, and ESPript for presenting alignments. Sequencing output was aligned to the target OSM sequence with the BESTFIT command of the GCG package, which uses the Smith-Waterman local homology algorithm. Molecular figures were made with BOBSCRIPT and RASTER3D.", "Q: Why did Marc Deller truncate oncostatin M rather than use the wildtype protein?\nA: OSM-196 was known to aggregate, be poorly soluble and form dimers, and crystallisation attempts had failed because of low solubility and excessive precipitation. Marc reasoned that the extended C-terminal region of OSM-196, apparently surplus to structural stability and function, could be hindering crystallisation, so he rationally engineered C-terminal truncation mutants to remove it.", "Q: What did the AP33 complex show about HCV neutralisation?\nA: The structure established that there are two distinct modes of antibody binding to the E2 412-423 antigenic site that can neutralise diverse HCV strains. Marc Deller's post notes the comparison with the broadly neutralising antibody HCV1, which shows a different angle of approach and subtle \u03b2-hairpin conformational variations.", "Q: What was Marc Deller's contribution to the 2013 Science HIV envelope trimer paper?\nA: Marc Deller is a listed author on the 2013 Science paper \"Crystal structure of a soluble cleaved HIV-1 envelope trimer\" (about 979 citations), his most cited work. He contributed structural biology and crystallography to a high-resolution structure of a near-native, cleaved HIV-1 envelope trimer in complex with a broadly neutralizing antibody, revealing the trimer's architecture and vulnerable sites for understanding infection and rational vaccine design.", "Q: What is Marc Deller's post \"Tuning the Signal: Rational Design of G Protein Selectivity\" about?\nA: Published on 6 December 2025, the post covers a Nature paper by Moore and colleagues showing how allosteric modulators can dictate which G protein a receptor couples to. Marc Deller explains that on the neurotensin receptor 1 (NTSR1), the modulator SBI-553 sculpts the intracellular transducer pocket rather than simply inhibiting signalling.", "Q: What certificate did Marc earn from IBM?\nA: Marc Deller completed \"Introduction to Artificial Intelligence (AI)\", a certificate issued by IBM through Coursera on 10 March 2025. His post about it notes that AI increasingly influences prediction, automation and analysis across science, and that the course helps scientists begin applying AI concepts to modelling, imaging and data interpretation.", "Q: What structure did the NANOG paper report?\nA: The team determined the crystal structure of the human NANOG homeodomain bound to an OCT4 promoter DNA, which revealed the amino acid residues involved in DNA recognition. NANOG, named from the Irish mythological T\u00edr na n\u00d3g, is a transcription factor that maintains pluripotency in cooperation with OCT4, SOX2 and other pluripotency factors.", "Q: How did the team improve on the original screening hit?\nA: The program began with racemic HTS hit (\u00b1)-1, which inhibited CDK2/cyclin E1 with an IC50 of 607 nM but also potently hit JAK2. Introducing an sp3-hybridized aminopiperidine gave compound 2 (CDK2 IC50 = 16 nM) with no measurable JAK2 activity, and switching the pyrimidine C5 chloro to trifluoromethyl (compound 3) increased both potency and CDK1 selectivity. Iterative SAR on the pyrazole and sulfonyl substituents led ultimately to compound 17.", "Q: What keywords does Marc Deller list on ORCID?\nA: Marc Deller's ORCID record carries four keywords: Structural Biology, Protein Crystallography, Crystallization, and Protein Expression and Purification. They summarise the technical core of his research career.", "Q: Explain this paper to a non-scientist.\nA: HIV coats its surface protein in sugar molecules that act like armour, hiding the spots the immune system needs to attack. In this study, which Marc Deller contributed crystallography to, the team followed one Chinese patient over several years and caught a promising antibody in the act of developing. Using X-ray pictures of the antibody gripping the virus, they showed that two of those sugars physically blocked part of the antibody, so it never became fully protective \u2014 and that engineering away the clash rescued its power. That tells vaccine designers exactly which sugars to work around.", "Q: What proportion of PDB structures came from X-ray crystallography in 2014?\nA: The review states that as of January 2014, X-ray crystallography accounted for 88% of all macromolecular structures in the Protein Data Bank, specifically 85,486 of 96,692 entries.", "Q: Which review anchors Marc's LOCKTAC post and where can it be found?\nA: The anchor is a Science review published in August 2025, \"Load and lock: An emerging class of therapeutics that stabilize molecular interactions\", supplemented by 2024 work in ChemBioChem on small-molecule modulation of protein\u2013protein interactions and a Cancer Discovery piece on targeting protein complexes. Marc's post is at https://marcdeller.com/the-locktac-revolution-binding-the-future-of-therapeutics/.", "Q: What broader point does Marc draw from comparing these two enzymes?\nA: Marc frames it as convergent evolution through divergent pathways: FAST-PETase achieves performance by precise surface mutations for industrial use, while HGMP01 represents nature's independent solution shaped inside the gut, offering new templates for protein engineering and raising questions about the fate of ingested plastics. The post is at https://marcdeller.com/hgmp01-and-fast-petase-decoding-two-paths-to-plastic-degradation-engineered-versus-natural-petase-enzymes/.", "Q: Which OSM residue was compared with Glu 19 of GCSF, and what was found?\nA: Marc Deller noted that Glu 19 of GCSF, which hydrogen bonds to Arg 288 of GCSFR at the core of the major interface, is structurally equivalent to Lys 19 of OSM. Although Lys 19 is about 80 % solvent exposed, its alanine mutation had no effect on gp130 or LIFR binding, whereas mutation of the adjacent, less exposed residues Gln 16 and Gln 20 reduced gp130 binding by about 75-fold or more.", "Q: Did Marc write about QSAR modelling?\nA: Yes. In \"Empowering QSAR Modeling: Optimal Data Practices and Cutting-Edge Tools\", published 27 August 2025, Marc Deller argues that the predictive power of Quantitative Structure\u2013Activity Relationship models rests on meticulous data curation, appropriate dataset sizing and modern automated platforms.", "Q: What is the mercury LIII edge and why did Marc scan it?\nA: The mercury LIII absorption edge (ideally at \u03bb = 1.00910 \u00c5, 12.2858 keV) is the energy at which the anomalous scattering of mercury changes sharply. Marc Deller recorded a fluorescence scan of the EMP-derivatised crystal to locate the edge experimentally, observing the characteristically broad f\u2033 peak often seen for mercury-derivatised proteins, and used it to choose the peak, inflexion and remote wavelengths for the MAD experiment.", "Q: Which patents are listed as Marc's top patents?\nA: The top patents are \"Bicyclic compounds as inhibitors of WRN\" (2025), \"Tricyclic compounds as inhibitors of WRN\" (2025), \"Anti-mutant calreticulin (CALR) antibodies and uses thereof\" (2023), \"Imidazo[1,2-a]pyrimidine-5(1H)-ones as WRN inhibitors\" (2025), and \"Imidazo[1,2-a]pyrimidine-5(1H)-ones having a covalent warhead at the 3-position as WRN inhibitors\" (2025).", "Q: Who were the corresponding and first authors of the povorcitinib paper?\nA: David M. Burns of the Incyte Research Institute was the corresponding author and prepared the manuscript with contributions from all listed authors. Jincong Zhuo, Yun-Long Li, Ding-Quan Qian, David M. Burns and Song Mei contributed equally as co-first authors; Marc Deller is one of the co-authors contributing crystallography.", "Q: Which refinement programs does the chapter recommend?\nA: REFMAC, phenix.refine and BUSTER-TNT are the three widely used refinement programs named in the chapter, and the authors advise running several of them so that difference density for the ligand can be confirmed as robustly reproduced. They note graphical user interfaces are available for REFMAC and phenix.refine.", "Q: How does the patent handle the breakdown products of the plastics?\nA: The patent addresses metabolic clearance of degradation products such as terephthalic acid (TPA), ethylene glycol (EG), and oxidized polyolefin monomers. It states that human-specific hepatic (liver) and renal (kidney) metabolic pathways efficiently process and safely excrete these byproducts under controlled therapeutic dosing.", "Q: What kind of role is Marc Deller looking for?\nA: Marc Deller's r\u00e9sum\u00e9 states that he is eager to leverage his drug discovery expertise and innovation leadership to accelerate next-generation therapeutic development, and that he is willing to relocate from the East Coast. He currently serves as CSO and co-founder of Elora Therapeutics.", "Q: What results does Marc point to from generative protein design?\nA: Marc points to enzymes that are stable at high temperatures for industrial processes and proteins engineered with unparalleled specificity to target cancer cells while sparing healthy ones, arguing this enables a new paradigm of rational, function-driven design. The post is at https://marcdeller.com/the-protein-architect-how-generative-ai-is-building-the-drugs-of-tomorrow/.", "Q: What is Marc's favourite band?\nA: Man or Astro-Man?, the instrumental surf-rock band, which fits his love of instrumental surf music.", "Q: How is ChatMol installed?\nA: Two ways. The recommended route is to add `sys.path.insert(0, \"\u2026/chatmol\")` and `import chatmol` to `~/.pymolrc.py` and restart PyMOL; alternatively the bundled `chatmol.zip` can be installed through PyMOL's Plugin Manager, which also adds a Plugin > ChatMol Settings dialog. ChatMol uses only the Python standard library, so no extra pip install is required.", "Q: What was Marc Deller's contribution to the 2016 Immunity VRC01 paper?\nA: Marc Deller contributed the X-ray crystallography. The paper's author contributions statement lists \"X-ray work and analysis by L.K., Y. Hua, M.C.D., and F.G.\" \u2014 that is, Leopold Kong, Yuanzi Hua, Marc C. Deller and Fernando Garces. Marc's crystallographic work covered crystallisation, harvesting and diffraction data collection for the DRVIA7 Fab and its complex with gp120.", "Q: Which top posts appear on Marc's microplastics page?\nA: The top posts are \"Vibe Coding MicroPlastic Blaster\", \"Elora Therapeutics: Treating Microplastics in Humans\", \"Decoding the Molecular Machinery Behind Plastic Degradation\", \"Cellular Uptake of Microplastics: Pathways, Persistence, and Therapeutic Potential for Intracellular Targeting of PETase\", and \"Clearing the Microscopic Threat: Can Plastic-Eating Enzymes Treat Liver Disease such as MASLD?\".", "Q: Can you explain the ligand validation chapter for a non-scientist?\nA: When scientists solve a crystal structure they produce a three-dimensional map of blurry electron clouds and then build a model of the molecule into it. Marc Deller's chapter is a checklist for making sure that a drug molecule drawn into that map is actually supported by the data rather than wishful thinking, since the map can be subtly biased toward whatever the scientist already put in it.", "Q: How does Marc think these AI tools fit together in practice?\nA: His conclusion is that they form modular pipelines: generate a backbone with RFdiffusion or ProteinGenerator, predict its structure with AlphaFold 3 or ESMFold, refine sequences with ProteinMPNN or LigandMPNN, and iterate in closed-loop automation using RAG-ESM and high-throughput biofoundries. Marc argues this compresses months of engineering into days, and you can read the round-up at https://marcdeller.com/engineering-the-future-of-proteins-unveiling-the-latest-ai-tools-transforming-structural-biology-in-2025/.", "Q: Where can I find Marc Deller's Google Scholar profile?\nA: Marc Deller's Google Scholar profile is at https://scholar.google.com/citations?user=o02oSJIAAAAJ. It lists 73 publications ranked by citation count, together with his all-time and since-2021 citation statistics.", "Q: Where is Marc's post about the Caulobacter histidine phosphotransferase?\nA: The post \"Crystal Structure of Histidine Phosphotransferase ShpA (NP_419930.1) from Caulobacter crescentus at 1.52 \u00c5 Resolution\" is in the structures category of marcdeller.com. You can read it at https://marcdeller.com/crystal-structure-of-histidine-phosphotransferase-shpa-np_419930-1-from-caulobacter-crescentus-at-1-52-a-resolution/.", "Q: Was the purified, deglycosylated LIFR still active?\nA: No. Receptor titration assays showed that purified and deglycosylated LIFR-FN3-3C-Fc was inactive, with no dose-dependent binding of biotinylated LIF, supporting the idea that glycosylation is required for correct folding and activity. However, freshly 293T-expressed LIFR-FN3-3C-Fc was active, competing with biotinylated LIF (EC50 12.1 ng/ml) comparably to CHO-expressed control material (EC50 10.4 ng/ml).", "Q: What are SSRL grids and why do they matter?\nA: SSRL grids are 200 \u00b5m thick polycarbonate pieces containing 75 laser-etched holes with diameters of 125, 200 and 400 \u00b5m, developed for automated collection of single frames or small rotation ranges from each crystal. Marc Deller's review notes they allow extremely high-density shipping, with 1,200 crystals in a single UniPuck and 8,400 in a shipping dewar, and that the UMR was being developed to harvest into them.", "Q: Where was the 2016 VRC01-class glycan paper published and with what DOI?\nA: The paper appeared in Immunity, volume 44, pages 939\u2013950, on 19 April 2016, with DOI 10.1016/j.immuni.2016.03.006. It was received in October 2015 and published online on 5 April 2016.", "Q: Did Marc Deller co-author a structure of a Fic family protein?\nA: Yes. \"Crystal structure of the Fic (Filamentation induced by cAMP) family protein SO4266 (gi|24375750) from Shewanella oneidensis MR-1 at 1.6 A resolution\" was published in Proteins in 2009 with Marc Deller among the Joint Center for Structural Genomics co-authors. Europe PMC carries no abstract for this entry.", "Q: What are Marc's favourite towns?\nA: Nottingham, San Diego and Wilmington.", "Q: Why does MLX matter for ChemSage?\nA: Apple's unified-memory architecture removes both the VRAM wall and the CPU-GPU memory copies a discrete NVIDIA card forces. In practice a 16 GB Mac tunes an 8B model, 32 GB handles roughly 14B and 64 GB reaches a full 32B, and because the base is already a 4-bit MLX checkpoint the fine-tune runs automatically as QLoRA. The trade-off against NVIDIA is wall-clock speed, not capability.", "Q: Which CDR loops does AP33 use?\nA: For AP33, all the complementarity-determining region loops except L2 are involved in the interaction. For HCV1, CDR H2, H3 and L3 account for most of the contacts, with framework region 2 and L1 contributing some as well. HCV1 binding relies on an insertion in CDR H3 that AP33 does not have, whereas AP33 has a longer L1 with inserted residues immediately upstream of a contacting residue in that loop.", "Q: Has Marc written about the risks and limits of AI?\nA: Yes, the AI page lists posts on AI safety and failure modes, including \"When Protein Design Becomes a Security Problem\", \"Self-Adapting Language Models: Are We Creating Our Final Invention?\", \"Call Me A Jerk: Persuading AI to Comply with Objectionable Requests\", \"Cats Confuse Reasoning LLM: Query-Agnostic Adversarial Triggers for Reasoning Models\", \"When Perception Deceives Reality: The Hidden Flaws in AI Reasoning Evaluation\", and \"How to Confuse Machine Learning: Understanding Modern Adversarial Attacks\".", "Q: How were the oligonucleotide primers for the OSM work prepared?\nA: The oligonucleotides were prepared by John Fox at AltaBiosciences, Birmingham, on a 308B Applied Biosystems DNA synthesiser at a 0.2 \u03bcM scale, purified on Sephadex G25. They were dissolved in sterile water to 1 mg/ml, stored at \u201320 \u00b0C, and diluted to 200 ng/\u03bcl for use in PCR, with concentrations judged from the convention that an A260 of 1 corresponds to 20 \u03bcg/ml of oligonucleotide.", "Q: How did the patent confirm that Cys727 was the important residue?\nA: Example F describes a mutant panel. Cys580, Cys727, Cys864 and Cys1041 were each mutated to both serine and alanine in the His-tagged WRN construct, giving eight mutants, and these were assayed for helicase and ATPase activity. Cys580 mutations abolished all activity as expected for a Walker A cysteine, Cys864 and Cys1041 mutations had little effect, and Cys727 mutations abolished helicase activity while only reducing ATPase activity \u2014 a decoupling of the two activities through Cys727.", "Q: What was the structural evidence in the DRT3 story?\nA: Marc emphasises that cryo-EM was the decisive experiment: structures at 2.6 \u00c5 resolution resolved the complete assembly as a D3-symmetric 6:6:6 ribonucleoprotein complex with both reverse transcriptases engaged simultaneously in one coordinated machine. Without the atomic-resolution model it would have been impossible to demonstrate conclusively that no nucleic acid strand occupies the Drt3b active site.", "Q: Which keywords appear most often on Marc Deller's PDB entries?\nA: HYDROLASE is the most common classification, on 13 entries, followed by IMMUNE SYSTEM on 12 and UNKNOWN FUNCTION on 11. TRANSFERASE and METAL BINDING PROTEIN appear four times each, with TRANSPORT PROTEIN, OXIDOREDUCTASE, DNA BINDING PROTEIN and HYDROLASE/HYDROLASE INHIBITOR appearing twice.", "Q: Does WRN inhibition combine with chemotherapy?\nA: Yes. The post reports that combination studies reveal synergy with topoisomerase inhibitors such as irinotecan, where subefficacious doses of each agent produce enhanced antiproliferative effects and complete tumour regression in MSI colorectal cancer models, alongside dose-dependent tumour growth inhibition in cell line and patient-derived xenograft models.", "Q: What is unusual about the genetics of the TM1088 locus?\nA: The TM1088 locus of T. maritima consists of two separate open reading frames with overlapping start and stop codons, so it encodes two separate proteins, TM1088A and TM1088B, from a single promoter with an internal ribosome binding site. In some databases TM1088A is annotated as locus TM_1087.1 and TM1088B as TM_1088. Marc Deller's paper noted that this arrangement is suggestive of a gene duplication event similar to that proposed for the evolution of E. coli TrkA and other Trk/KtrAB proteins.", "Q: Why does ALPHABETTI avoid multiple sequence alignments?\nA: Marc's argument is that a conventional sequence logo measures conservation by counting residues in an alignment column, which means finding homologues, aligning them and inheriting every bias in that process. ALPHABETTI instead asks ESM-2 what it expects at each position, having first masked that position so the model cannot read the answer off its own input, so what you see is a model's expectation mapped onto a model's structure with no alignment anywhere.", "Q: What does the CODSWALLOP front page show?\nA: The CODSWALLOP front page opens with an ASCII wordmark, the expansion of the acronym, a search box and a list of recent searches with the number of entries found for each. Those examples include vascular endothelial growth factor receptor 1 with 1333 entries, mitogen-activated protein kinase 14 with 1629, spike glycoprotein with 1991, lysozyme C with 1691, insulin with 508, and PET hydrolase with 209.", "Q: How accessible is LatentX to ordinary researchers?\nA: Latent Labs released it as a no-code web platform where users upload target proteins, select binding hotspots and generate binder molecules with computational rankings that mirror lab validation results. Marc frames this as dramatically lowering the barrier to entry for AI-powered drug design across academia, biotech startups and pharma, and his post is at https://marcdeller.com/revolutionizing-drug-discovery-latentx-and-the-future-of-protein-binder-design/.", "Q: Why did the povorcitinib team want JAK1 selectivity over JAK2?\nA: The povorcitinib paper explains that aberrant JAK2 signaling drives cancers such as myeloproliferative neoplasms, whereas JAK1 is a central mediator of autoimmune and inflammatory diseases. Because JAK2 supports blood cell production, inhibiting it can cause anemia and other cytopenias, so selective JAK1 inhibition is a potentially safer therapeutic strategy than dual JAK1/JAK2 inhibition.", "Q: What is the JCSG Top96 Cryo column?\nA: The JCSG Cryo column gives, for each Top96 condition, the cryoprotectant additive that at the JCSG most reliably produced a clear freeze, abbreviated as EG (ethylene glycol), MPD (2-methyl-2,4-pentanediol), G (glycerol), P200 (PEG 200) and P400 (PEG 400), with a percentage. \"None\" means the condition is already cryoprotective and can be frozen directly.", "Q: What is KPN03535 structurally similar to?\nA: KPN03535 is structurally similar to the bacterial Cpx-pathway protein NlpE, to single-stranded DNA-binding proteins and to toxins. Klebsiella pneumoniae MGH 78578 is part of the normal human skin, mouth and gut flora and an opportunistic pathogen linked to about 8 percent of all hospital-acquired infections in the USA.", "Q: What is the earliest work on Marc Deller's ORCID record?\nA: The earliest entries on Marc Deller's ORCID record are both from 2000: \"Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\" in Structure (DOI 10.1016/S0969-2126(00)00176-3) and the review \"Cell surface receptors\" in Current Opinion in Structural Biology (DOI 10.1016/S0959-440X(00)00072-5).", "Q: What broader point did the NANOG paper make?\nA: The authors concluded that structural and biophysical information about key transcription factors provides insight into how stem cell behaviour can be manipulated, and a framework for rational protein engineering. The paper appeared in the Proceedings of the National Academy of Sciences in 2015 (DOI 10.1073/pnas.1502855112).", "Q: What is the take-home message of the INCB159020 paper?\nA: The paper demonstrates that a covalent warhead is not necessary for potent, selective inhibition of KRAS G12D, and that oral bioavailability can be won back after removing the polar groups that potency initially depended on. The team's two-step strategy \u2014 first prove reversible potency by finding additional protein contacts, then strip out the features blocking oral exposure and rebuild potency through conformational rigidity \u2014 yielded compound 23, INCB159020, an important tool compound and a template for future bioavailable KRAS inhibitors.", "Q: What does Marc say those human-skills credentials demonstrate?\nA: Marc Deller writes that they affirm an ability to navigate complex human and technological interfaces, which matters as organisations prioritise empathy, ethics and evidence-based leadership. He lists goals including leading diverse teams, engaging stakeholders, balancing data insights with human factors, and fostering ethical AI usage.", "Q: Why does this matter for HIV vaccine design?\nA: Because it means the true 10E8 target is a composite epitope of MPER peptide plus viral membrane lipid, not peptide alone. The paper argues that immunogens should present the MPER linked to lipids in the correct orientation \u2014 for example in liposomes or chimeric viruses \u2014 and that the findings also inform improvements to 10E8's pharmacodynamic and pharmacokinetic properties for therapeutic use.", "Q: Which structures relate to the SpoIIAA-like proteins paper?\nA: 2OOK, a protein of unknown function YP_749275.1 from Shewanella frigidimarina NCIMB 400 at 1.80 \u00c5 in C 1 2 1 (released 6 February 2007), and 2Q3L, a DUF3478 family protein with a SpoIIAA-like fold, Shew_3102, from Shewanella loihica PV-4 at 2.25 \u00c5 in C 2 2 21 with sodium bound (12 June 2007). Both are cited in the 2010 Acta Crystallographica Section F paper on open and closed conformations of two SpoIIAA-like proteins.", "Q: How does compound 29 compare with the approved FGFR drugs against gatekeeper mutants?\nA: Against FGFR3 V555L, compound 29 had an IC50 of 2.7 nM compared with 21 nM for erdafitinib, 373 nM for pemigatinib and 3036 nM for infigratinib; against V555M it gave 6.1 nM versus 68, 201 and 352 nM respectively. The approved drugs stay fully potent on FGFR1, whereas compound 29 is 42-fold selective against it, which is the property expected to reduce hyperphosphatemia.", "Q: What does Marc say about Graphinity and synthetic training data?\nA: Marc notes that Graphinity, which uses an equivariant graph neural network, hits performance ceilings when trained on only a few hundred experimental measurements, and that augmenting with nearly one million FoldX-generated values raised correlation to 0.9 but exposed limits in representativeness. Without balanced coverage of antibody lineages, CDR variations and mutation contexts, overfitting persists.", "Q: What machine does PANTS run on?\nA: PANTS runs on a single DigitalOcean droplet shared with five other applications, which is why its web layer carries no scientific libraries at all: every structure and measurement is computed offline and served as a precomputed file. Its stats page reports a 20.6 MB SQLite database alongside 832 candidate and 559 reference structure files, with memory rather than disk or speed as the binding constraint.", "Q: How do AlphaFraud and CODSWALLOP connect?\nA: CODSWALLOP's AlphaFold overlay ties into AlphaFraud's prediction-versus-experiment scoring, and CODSWALLOP's crystallisation-condition output is intended to feed Marc's planned crystallisation predictor, which grew out of his JCSG Top96 screen work.", "Q: Explain this CDK2 paper in simple terms for a non-scientist.\nA: In simple terms, the paper describes designing a molecule that switches off a specific enzyme, CDK2, which some cancers rely on to keep dividing. The challenge was to hit CDK2 without hitting very similar enzymes that healthy cells need, because that mismatch caused toxicity in older drugs. The Incyte team, including Marc Deller who worked out the 3D picture of the drug locked onto CDK2, arrived at a compound called 17 that is extremely potent and clean, and it shrank tumor growth in mice.", "Q: What is the Topliss tree and why does Marc still care about it?\nA: In \"Unlocking Molecular Potential: The Evolution of Topliss Tree and Bioisosteric Strategies in Modern Drug Discovery\", published 15 September 2025, Marc Deller describes the decision tree John Topliss introduced in 1972 as a non-mathematical way to guide medicinal chemists through systematic aromatic substitution. His argument is that its principles still inform modern computational and AI-driven optimisation rather than having been superseded.", "Q: How does Marc think about communication in science?\nA: In \"Communication: 60+ publications translating structural data into actionable insights\", Marc Deller describes communication as the bridge between atomic-level detail and real-world decisions in drug discovery. He writes that solving a structure is only step one, and that the value emerges when the structure is translated into clear hypotheses and strategies non-structural biologists can act on.", "Q: What were the refinement statistics for the Fab\u2013mutCALR peptide complex?\nA: The 3.2 \u00c5 Fab\u201321-mer peptide structure was in space group P 21 21 21 and refined to R-work 22.9% and R-free 31.7%, with 91.2% of residues Ramachandran-favoured and seven outliers, which the patent notes is acceptable at that resolution. The 2.9 \u00c5 apo Fab structure was in space group P 1 21 1 and refined to R-work 23.0% and R-free 30.3%.", "Q: What is the WRN patent Marc Deller is named on?\nA: Marc Deller is an inventor on Incyte's WRN (Werner syndrome helicase) patent, published as US2025/0215006A1, \"Tricyclic Compounds as Inhibitors of WRN\" (his r\u00e9sum\u00e9 refers to it by the shorthand \"Bicyclic Compounds as Inhibitors of WRN\"). He is listed fourth among thirteen inventors and contributed the co-crystal structure of the inhibitor bound to WRN, solved at 3.1 \u00c5 at Diamond Light Source, which revealed a novel binding pocket created by a large rotation of the ATPase 2A domain.", "Q: Where can I read the abstracts of Marc Deller's papers?\nA: All 58 abstracts are collected from Europe PMC under the author search \"Deller MC\", newest first, spanning 1999 to 2026. Each entry carries the full author list, journal, year, DOI and PMID.", "Q: What is marcdeller.com built on?\nA: marcdeller.com is a WordPress/Elementor site hosted on Cloudways with a Breeze/Varnish/Cloudflare APO performance stack, with sections for About, AI, Blog, Chat, Learnings, Microplastics, Patents, Projects, Publications, Resume, Structures, Structural Biology and Vibe Coding.", "Q: What is in Marc's \"Building Data and Project Management Expertise\" post?\nA: In \"2025 \u2013 Building Data and Project Management Expertise\", Marc Deller groups five credentials into one theme: Foundations: Data, Data, Everywhere; Introduction to R: Basic R Syntax; Foundations of Project Management; Dashboard Development with Shiny: GenAI for Retail Analysis; and Serverless Architectures on AWS. He presents them as a dual toolkit covering data literacy, planning and visualisation, capped by cloud-native architecture skills.", "Q: What is Marc's assessment of what vepdegestrant's approval does and does not prove?\nA: Marc is careful to say that vepdegestrant's clearance will not on its own derisk LYTACs in the clinic, since biodistribution, manufacturability and immunogenicity remain real engineering problems. What it does establish, he argues, is that regulators will approve degraders on the same evidentiary basis as any other oncology agent, which is the validation the extracellular degrader field was waiting for. The post is at https://marcdeller.com/beyond-the-cell-the-new-frontier-of-extracellular-protein-degradation/.", "Q: Which structural class do LIF and OSM belong to?\nA: LIF and OSM belong to Group 1, the 4-\u03b1-helical bundle cytokines, and specifically to the long-chain subgroup. Long-chain helical cytokines have relatively long \u03b1-helices of about 25 amino acids and a small skew angle (~18\u00b0) between the A:D and B:C helix pairs, and LIF, CNTF, CT-1, IL-6, IL-11 and OSM form a group united by structural and sequence homology and their shared use of the gp130 receptor.", "Q: How fast is DeepCovalent's discovery process?\nA: DeepCovalent states that OS3D compresses early-stage discovery from months to three weeks or less, taking a partner from ideation to fully in-silico validated molecules in that time. The company presents this as enabling low-cost, de-risked evaluation of new or deprioritised targets and creating early pipeline optionality.", "Q: How many water molecules were modelled in the OSM structures?\nA: The P212121 structure of \u0394188_OSM contains 116 ordered water molecules, while the lower-resolution I222 structure contains 35.", "Q: What did the review say about chemokine receptors?\nA: The review noted that the multi-membrane-spanning nature of chemokine receptors is a substantial obstacle to high-resolution structural studies, and that a peptide fragment from CXC-chemokine receptor 1 had been used in an NMR study of IL-8 binding to provide the first structural toehold on chemokine\u2013receptor interactions.", "Q: How many claims are in the patent and what do they cover?\nA: The provisional patent has six claims. Claim 1 covers the therapeutic enzymatic composition itself (PETase, MHETase, CotA-Laccase, and HRP enzyme groups); claims 2 through 4 cover oral, injectable, and hybrid delivery of those compositions; claim 5 covers the method of therapeutically degrading microplastics in the body with safe hepatic and renal clearance; and claim 6 covers the stabilization and immunogenicity-reduction methods such as PEGylation and encapsulation.", "Q: What did the \"trust, but verify\" paper by Marc Deller argue?\nA: \"Models of protein\u2013ligand crystal structures: trust, but verify\", published by Marc C. Deller and Bernhard Rupp in the Journal of Computer-Aided Molecular Design in 2015 (about 113 citations), establishes metrics for why X-ray crystallography is the gold standard for producing precise protein\u2013ligand models. It advocates rigorous validation to combat overinterpretation and ensure reliability, which the portfolio frames as essential for advancing drug design and computational biology.", "Q: What did Marc Deller contribute to the Incyte medicinal-chemistry papers?\nA: On the Incyte papers published between 2022 and 2026, covering FGFR2/3, CDK2, KRAS G12D and JAK1, Marc Deller contributed the structural biology as part of Incyte's structural biology group. These are structure-based drug design campaigns in which protein crystallography guided the medicinal chemistry.", "Q: Which mutant had the biggest effect on gp130 binding?\nA: In Marc Deller's oncostatin M study the Asn124Ala mutant showed the greatest reduction in gp130 binding (more than 100-fold) and in biological activity (more than 10,000-fold in both LIFR/gp130 and OSMR/gp130 bioassays), followed by Gly120Ala and Gln20Ala. The Gly120Tyr mutant, mimicking CNTF, was even less active than Gly120Ala.", "Q: Where can I read about Marc's R training?\nA: The post \"2025 \u2013 Introduction to R: Basic R syntax\" is published at https://marcdeller.com/introduction-to-r-basic-r-syntax/ in the certificates category.", "Q: What did the VRC01-class antibody paper find?\nA: Marc Deller's post \"2016 \u2013 Key gp120 glycans pose roadblocks to the rapid development of VRC01-class antibodies in an HIV-1-infected Chinese donor\" reports that the N276 glycan on gp120 creates rigid steric constraints that force developing antibodies to accommodate the carbohydrate rather than displace it, slowing the maturation of VRC01-class broadly neutralizing antibodies.", "Q: How does the antibody interact with ruxolitinib?\nA: The patent shows that anti-mutCALR antibodies potentiate the therapeutic response of the JAK1/2 inhibitor ruxolitinib. Ruxolitinib alone at 50 nM inhibited oncogenic cell proliferation by only about 20%, whereas combining it with anti-mutCALR antibody clone 6 markedly improved inhibition in cells carrying either the Type 1 or the Type 2 CALR mutation, and claim 97 covers combinations that produce a synergistic effect.", "Q: Why does WordPressID fetch once and render twice?\nA: The API responses are captured once into a `$data` variable as JSON lines, and both the CSV and the terminal table are generated from that capture, so there are no double requests and the file and the screen can never disagree. On-screen titles are truncated to 70 characters for terminal width while the CSV keeps the full title.", "Q: What are Marc's areas of scientific expertise?\nA: Marc's expertise covers X-ray crystallography, cryo-EM, protein production and construct design, biophysical assays, structure\u2013activity relationships and structure-based drug design, across cytokines and cell-surface receptors, receptor tyrosine kinases, proteases, potassium channels, esterases and cellulases, viral envelope proteins and neutralising antibodies, and oncology kinases such as CDK2, FGFR2/3, JAK1/2, KRAS G12D and the WRN helicase.", "Q: Why does Marc say stitching protein domains together is hard?\nA: Marc explains that simply joining domains often produces misfolded, non-functional blobs, and that predicting how components will interact to form a stable working whole had been a matter of frustrating trial and error. ProDomino learns the governing rules from vast datasets of known protein structures.", "Q: Tell me some fun facts about Marc.\nA: A few things that do not appear on a CV:\n\n- \ud83c\udf7a He names software after beer. **BoltzMaker** is named for Timothy Taylor's Boltmaker, with a \"z\" borrowed from Boltz-2.\n- \ud83e\uddea His favourite line about the job: *if science worked on the first attempt, it would be called Search.*\n- \ud83d\udcbb He calls himself a \"vibe-coder\": a subject-matter expert who builds his own AI-assisted scientific tools, and who fine-tunes language models on a Mac in his own home.\n- \ud83e\uddec His D.Phil. produced the **first crystal structure of oncostatin M**, which is still the structure his name is most associated with.\n- \u23f1\ufe0f He once solved a protein-antibody complex in just over ten days to beat a patent deadline, and the resulting patent is projected at more than $500M in annual sales.\n- \ud83e\uddf1 He is an avid Lego collector.", "Q: What therapeutic possibility does Marc see in the molecular Velcro work?\nA: Marc suggests that biocompatible mussel-inspired polydopamine coatings could form a plausible starting point for in-body sequestration of circulating plastic particles, though full translation would require substantial validation under physiological conditions. He writes that the conceptual bridge between molecular Velcro for water and molecular Velcro for the body is shorter than ever. The post is at https://marcdeller.com/molecular-velcro-catching-what-we-cannot-see/.", "Q: Did Marc Deller conclude whether the TM1088 structure is in an open or closed state?\nA: Marc Deller concluded that the AMP-bound TM1088 structure most closely resembles a closed state of the channel. The ~120\u00b0 dimer-hinge angles match those seen in the ADP-bound, presumably closed, TrkA/TrkH complex from V. parahaemolyticus, whereas the ATP\u03b3S-bound, presumably open, TrkA/TrkH has a hinge angle of about 100\u00b0. The paper noted that further biochemical studies and structures in other conformational states would be needed to confirm this.", "Q: Which CROs and synchrotrons has Marc Deller worked with?\nA: At Incyte, Marc Deller built relationships with 10+ external CROs and facilities including Saromics, WuXi, GenScript, XChem/Diamond, and the SSRL, NSLS and APS synchrotrons. These partnerships expanded capabilities in protein production, fragment screening, NMR, WAC, and biophysical and biochemical screening.", "Q: How does PGT135 get through the HIV glycan shield?\nA: The structure shows that the broadly neutralising antibody PGT135 uses extended CDR loops to penetrate the glycan shield and contact glycans at Asn332, Asn392 and Asn386. Marc Deller's post presents this as the structural basis of the supersite of immune vulnerability.", "Q: Did 1B2 also bind DEBS module 2?\nA: Yes. Because the module 2 + TE construct contains the N-terminal docking domain from module 3 and the hydrogen-bonding residues seen in the crystal structure are conserved, the authors predicted and confirmed that 1B2 binds module 2 + TE just as tightly, with a KD of 100 \u00b1 10 nM by ELISA. This allowed the well-established kinetic assays for module 2 to be used to test the effect of locking the extended conformation.", "Q: How does this patent fit into Marc Deller's wider work at Incyte?\nA: The mutant-CALR antibody patent is one of more than seven patents Marc Deller supported at Incyte, alongside programmes on KRAS G12D, WRN, CDK2/4, JAK1/2 and FGFR2/3. It is an example of the structure-based drug design he ran there, and of the automated structural biology platform he built, which delivered more than a hundred co-crystal structures in under three years across ten programmes.", "Q: What examples of bioisosteric replacement does Marc give?\nA: He cites replacing hydrogen with deuterium, as in deutetrabenazine, to improve pharmacokinetics through isotopic effects, and substituting carboxylic acids with tetrazole rings in angiotensin receptor blockers such as losartan to address bioavailability while retaining receptor binding. He traces the concept back to Langmuir's observations about molecules with similar electron distributions.", "Q: What is \"the 79th organ\" in Marc Deller's blog post?\nA: In \"The 79th Organ: When Art and Biology Meet Bioengineering in the Fight Against Microplastics\" (13 October 2025), Marc Deller writes about designer Odette Dierkx's speculative prosthetic organ built from bioengineered fungi \u2014 specifically *Pleurotus ostreatus* \u2014 that would act as a living filtration system extracting and degrading microplastics from the human bloodstream.", "Q: Do pleckstrin homology domains exist in bacteria?\nA: Yes, and Marc Deller co-authored the paper that showed it. \"Bacterial pleckstrin homology domains: a prokaryotic origin for the PH domain\", published in the Journal of Molecular Biology in 2010, presented crystal structures of three members of the uncharacterized family Pfam PF08000 that provide compelling evidence for PH-like domains in bacteria, called PHb domains.", "Q: Why are LOCKTACs interesting for undruggable targets?\nA: Because stabilising existing complexes offers a way into targets long thought inaccessible to traditional modalities, including transcription factors and macromolecular complexes. Marc writes that the versatility of the class has sparked a surge of interest, while acknowledging that specificity, off-target effects and delivery remain the main challenges.", "Q: What is PDB entry 3GF8?\nA: 3GF8 is a putative polysaccharide-binding protein of the DUF1812 family, NP_809975.1, from Bacteroides thetaiotaomicron VPI-5482 at 2.20 \u00c5 in space group P 41 21 2, with 296 polymer residues and 34.84 kDa, released on 24 March 2009. It is cited in the 2010 Acta Crystallographica Section F paper on a conserved fold for fimbrial components, where it is described as the fimbrial assembly protein BT1062.", "Q: Why is defining the epitope by crystallography valuable in a patent?\nA: A sequence claim only protects the particular antibodies described. An epitope claim protects any antibody that binds the same site on the target, which is far harder for a competitor to design around. Because Marc Deller's structures identified the exact mutant CALR residues contacted \u2014 including R8, M10, M19 and R20 for one binding mode and R7, R11, R17, R20 and R21 for the other \u2014 Incyte was able to claim binding to those residues directly in claims 55 to 71.", "Q: What does AlphaFraud's Calculate tab do?\nA: You type any human PDB ID and the full pipeline runs on demand, returning the same panel format as the auto examples with no AI-generated text. Autocomplete only completes to entries that exist, are human and have an AlphaFold model; rejections explain exactly why; the heavy compute runs in an isolated subprocess; and on-demand results are stored separately so they never enter the leaderboard, stats, analysis or archive.", "Q: What is Marc's full name?\nA: His full name is Marc Christian Deller; he publishes as Marc C. Deller and holds a D.Phil. from the University of Oxford. His middle name is Christian.", "Q: Is Marc a member of any learned society?\nA: I don't have anything on society memberships. His publication and structure record I can speak to; for affiliations, marcdeller.com is the place to look.", "Q: What exactly counts as crystal harvesting?\nA: In the review, harvesting in its broadest sense is any process that relocates a crystal from its growth medium to an environment where diffraction data can be collected. Marc Deller and Bernhard Rupp break it into looping, cryoprotection, flash-cooling and mounting, and distinguish it from the separate crystal-mounting and screening stage.", "Q: Explain the fungal PMO paper for a non-scientist.\nA: Fungi break down plant material using an enzyme that carries a copper atom and uses oxygen to snip cellulose chains, which is valuable for making biofuels. Marc Deller determined the three-dimensional structure of one of these enzymes, revealing a network of \"helper\" amino acids arranged around the copper. His collaborators then altered those helpers one at a time and found that two of them, a histidine and a glutamine, are needed to hold and activate the oxygen properly, and without the histidine the enzyme leaks a reactive form of oxygen called superoxide.", "Q: What are the steps in the cif2plip pipeline?\nA: gemmi parse with a single-model sanity check; remap multi-character chain IDs to single characters; rename residue names longer than three characters; write a strict PDB with element columns 77-78, TER records and padded occupancy/B; run pdb_tidy to renumber serials contiguously; generate intramolecular CONECT records from covalent-radius geometry; validate with pdb_validate; run PLIP; then flatten the XML into an interactions CSV and a ligand summary CSV.", "Q: What was special about the NANOG L122A mutant?\nA: Contrary to the other mutants, hNANOG L122A had enhanced protein stability and DNA-binding affinity. Mouse NANOG L122A kept embryonic stem cells expressing self-renewal markers even when retinoic acid was added to force differentiation, and in epiblast stem cells or human induced pluripotent stem cells the L122A mutants enhanced reprogramming into ground-state pluripotency.", "Q: What engineering examples for catalytic improvement does Marc give?\nA: Marc cites human arginase-I engineering, where cobalt substitution for manganese lowers Km from 1\u20135 mM to about 0.2 mM and improves cancer cytotoxicity IC50 fifteenfold, and *Pseudomonas fluorescens* kynureninase engineering giving 500\u2013700-fold better catalytic efficiency than the human enzyme, a 100-fold reduction in serum kynurenine and 45% improvement in median survival in mouse tumour models.", "Q: What is special about Compound 29 as an FGFR inhibitor?\nA: Compound 29 is a potent and selective FGFR2/3 inhibitor that retains low-nanomolar potency against common gatekeeper mutants. Marc Deller's post explains that the structure was instrumental in guiding structure-activity relationship optimisation to achieve excellent selectivity over FGFR1.", "Q: Why was expressing the LIF receptor (LIFR) so difficult?\nA: LIFR was difficult because previous attempts in E. coli had failed, owing to the very large constructs involved (typically over 1000 amino acids) and the inability to determine domain boundaries correctly. The extracellular fragment used, LIFR-FN3, is a heavily glycosylated 829-residue protein of ~94 kDa, and the protein proved prone to aggregation and instability, requiring mammalian expression systems.", "Q: Explain Marc's cell surface receptors review to a non-scientist.\nA: Cells talk to each other using receptor proteins that stick out from their surfaces and grab signalling molecules or neighbouring cells. In 2000, Marc Deller and Yvonne Jones wrote an overview of a remarkable year in which scientists roughly doubled the number of detailed 3D pictures showing exactly how such receptors hold their partners. A recurring theme was that it is not enough for a receptor simply to be bound; the receptors must be pulled together in precisely the right arrangement to switch on a signal inside the cell, and that snapshots from crystals need to be checked with experiments in living cells.", "Q: What did the PMO paper conclude about residues H161 and Q167?\nA: Activity assays combined with EPR spectroscopy showed that H161 and Q167 stabilise bound oxygen, that H161 appears to play a role in proton transfer, and that Q167 increases the ligand donor strength of Y169 to the copper through a hydrogen bond. Together these residues are important for oxygen activation, and the results were suggestive of a copper-oxyl active intermediate.", "Q: What is COMET and what performance does Marc report for it?\nA: COMET, the Combinatorial Mass Encoding decoding Tool, predicts fragmentation frequencies and filters tandem mass spectrometry scans to automate structure annotation of library members. Marc reports annotation recall rates of 66 to 82% while minimising false positives from isobaric compounds.", "Q: Tell me a joke.\nA: I'm more of a structural biology chatbot than a comedian, but here's one: Marc once joked that if science worked on the first attempt, it would be called \"Search\" instead of \"Research.\" If you want more of his favourite lines, just ask.", "Q: How selective is povorcitinib for JAK1 over JAK2?\nA: Povorcitinib (compound 22) had a JAK1 IC50 of 11 nM and a JAK2 IC50 of 505 nM in the enzymatic assay, giving about 46-fold JAK1 selectivity, roughly double the 25-fold selectivity of its predecessor compound 18. In human whole blood the selectivity ratio (TPO versus IL-6) was 16 for povorcitinib compared with 6 for compound 18.", "Q: What is PDB entry 3CGH?\nA: 3CGH is a SusD homologue, BT_3984, from Bacteroides thetaiotaomicron VPI-5482 at 1.70 \u00c5 in space group C 2 2 21, with 516 polymer residues and 57.69 kDa, released on 18 March 2008 with zinc bound. It is cited in the 2010 Acta Crystallographica Section F paper on the SusD/RagB family of nutrient-binding molecules.", "Q: What does Marc like about MARCUS's approach to open science?\nA: Marc highlights that MARCUS ships open-source code, accessible models and extensive documentation, which he argues invites community-driven development beyond natural products. He frames the tool as promoting FAIR data principles and bridging the gap between exponentially growing chemical literature and the structured demands of AI-driven discovery workflows.", "Q: When was the CDK2 paper published and in what journal?\nA: The paper was published in the Journal of Medicinal Chemistry in 2024, volume 67, pages 3112\u20133126 (DOI 10.1021/acs.jmedchem.3c02287). It was received on 5 December 2023, revised on 11 January 2024, accepted on 19 January 2024, and published on 7 February 2024.", "Q: How many viruses were used to test neutralisation in this study?\nA: Serum and antibody neutralisation were tested against two panels: a global panel of 12 viruses spanning diverse HIV-1 clades, and a separate DRVI panel of 25 viruses established at the Division of Research on Virology and Immunology of the China CDC. Donor DRVI01's serum showed over 95% breadth at every time point studied.", "Q: What did Marc fix inside Boltz itself?\nA: Nine defects found by measuring real campaigns, including memory never returned on Apple Silicon because Boltz called `torch.cuda.empty_cache()` on a Mac; 19 full-precision guards hardcoded to CUDA and therefore inert on MPS; one bad target aborting the whole run; an affinity step crashing on a missing file; steering forces overflowing to infinity and destroying structures; and an Apple-silicon masked-select miscount in the flat-bottom steering potential that cost a 58-target campaign nine crashes and eighteen lost targets over three days before `torch.where` fixed it.", "Q: Has Marc Deller worked on HIV?\nA: Yes. At the Joint Center for Structural Genomics at Scripps, Marc Deller contributed to HIV-Env antibody structures including the 2013 Science structure of a cleaved HIV-1 envelope trimer, and earlier at Pfizer he supported an IND on HIV-protease. Peptides against HIV-protease and antibodies against HIV-Env are listed among the targets in his r\u00e9sum\u00e9.", "Q: How did the team decide where to rigidify the molecule?\nA: They ran 1 \u03bcs molecular dynamics simulations on the ground state of a simplified analogue of compound 10 and overlaid all the energetically available conformations. The cyanoethyl group sits in a sterically congested environment and already samples the bioactive conformation as its largest ensemble, whereas the ethyl amide projects into open space and reaches the bioactive conformation only rarely. That pointed to the amide as the place to lock down, and cyclizing it to a \u03b3-lactam duly improved binding potency 5-fold.", "Q: How does the antibody PGT 135 reach through the HIV glycan shield?\nA: The 3.1 \u00c5 crystal structure of PGT 135 with gp120, CD4 and Fab 17b showed the antibody interacting with glycans at Asn332, Asn392 and Asn386, using long CDR loops H1 and H3 to penetrate the glycan shield and reach the gp120 protein surface. Electron microscopy showed PGT 135 can accommodate the conformational and chemical diversity of gp120 glycans by altering its angle of engagement.", "Q: Which PDB entry belongs to Marc's most-cited paper?\nA: PDB entry 4NCO underpins \"Crystal structure of a soluble cleaved HIV-1 envelope trimer\" in Science (2013), which is Marc Deller's most-cited publication. The entry is the BG505 SOSIP gp140 trimer with Fab PGT122 at 4.7 \u00c5, the largest assembly in his list at 350.93 kDa.", "Q: What kinds of things does Marc write about on his blog?\nA: He writes at **marcdeller.com**, mostly about where structural biology meets computation. Recurring themes:\n\n- \ud83e\uddec **Structural biology in practice**: cryo-ET revealing disease in its native state, dark kinases, crystallisation and protein stability.\n- \ud83e\udd16 **AI in the life sciences**: what protein-structure prediction can and cannot do, generative design, and where the hype outruns the data.\n- \ud83d\udcbb **Running models locally**: fine-tuning small language models on his own hardware, and why privacy, cost and offline access make that worth doing.\n- \ud83d\udee0\ufe0f **Vibe-coding**: building scientific web apps with AI assistance, and what he learns each time.\n- \ud83d\udc8a **Drug discovery**: target selection, structure-based design, and the realities of moving a programme forward.\n\nHe also writes up each of his own apps, so the blog doubles as the documentation for AlphaFraud, BoltzMaker, CODSWALLOP and the rest.", "Q: What is the lowest-resolution structure on Marc Deller's list?\nA: The lowest-resolution entry is 4NCO, the BG505 SOSIP gp140 HIV-1 Env trimer in complex with the broadly neutralising Fab PGT122, at 4.7 \u00c5. Close behind is 5C7K, the same trimer bound to PGT128 and 8ANC195, at 4.60 \u00c5.", "Q: What fraction of JCSG crystals need cryoprotectant?\nA: The review reports that about 80% of all protein crystals harvested at the JCSG required the addition of cryoprotectant to prevent damaging hexagonal ice formation, which otherwise reduces resolution and increases mosaicity.", "Q: What is MHETase's role in the depolymerization pathway according to Marc?\nA: Marc explains that MHETase complements PETase by breaking down PETase's intermediate product, MHET, into benign monomers, thereby completing the depolymerization pathway.", "Q: Is Marc Deller a founder of DeepCovalent?\nA: No. Marc Deller is an advisor to DeepCovalent, serving as its Scientific Technical Advisor, whereas the company itself was founded by a team of AI researchers, chemists and drug discovery scientists. His co-founder role is at Elora Therapeutics, where he is Chief Scientific Officer.", "Q: Who led the 10E8 study?\nA: Adriana Irimia was the first author and wrote the original draft; the senior and supervising authors were Robyn L. Stanfield, Dennis R. Burton, William R. Schief and Ian A. Wilson. Marc Deller was a middle co-author contributing crystallography.", "Q: What is Marc's overall conclusion in the Topliss and bioisosteres post?\nA: His conclusion is that the move from Topliss trees to AI-driven design is integration rather than replacement, with traditional rational principles informing computational approaches and computational predictions guiding experimental validation. He notes that deep learning and graph neural networks capture non-linear structure\u2013activity relationships but bring interpretability challenges, which is why mechanistic classical approaches retain value. The post is at https://marcdeller.com/unlocking-molecular-potential-the-evolution-of-topliss-tree-and-bioisosteric-strategies-in-modern-drug-discovery/.", "Q: Did Marc Deller contribute to the first crystal structure of a cleaved HIV-1 envelope trimer?\nA: Yes. Marc Deller is a co-author on \"Crystal structure of a soluble cleaved HIV-1 envelope trimer\", published in Science in 2013 by Julien JP, Cupo A, Sok D, Stanfield RL, Lyumkis D, Deller MC and colleagues including Burton DR, Sanders RW, Moore JP, Ward AB and Wilson IA. The structure of the BG505 SOSIP.664 gp140 trimer with the broadly neutralizing antibody PGT122 was determined at 4.7 \u00c5.", "Q: Which LLMs does Marc name in the prompt-engineering post, and where can it be found?\nA: Marc names GPT-4 and Claude, and notes the workflows are designed to work with standard models accessible through services such as Perplexity and OpenAI. The post is at https://marcdeller.com/supercharging-discovery-how-llm-prompts-are-transforming-structural-biology-and-cheminformatics/.", "Q: Where can I read Marc's post about diagnostics and microplastic removal from blood?\nA: You can read it at https://marcdeller.com/clearing-the-bloodstream-diagnostics-and-approaches-for-microplastic-removal/ on Marc Deller's blog, where it was published on 15 August 2025 under the blog and microplastics categories.", "Q: What is the intrinsic mannose patch?\nA: Marc Deller explains that the intrinsic mannose patch is a region where the high density of glycans limits their processing during biosynthesis, leaving under-processed oligomannose-type structures that act as distinctive molecular signatures for immune recognition.", "Q: What liabilities did the team have to engineer out of compound 10?\nA: Compound 10 bottomed out the surface plasmon resonance assay at under 1 pM but had no measurable Caco-2 flux and high human liver microsome clearance. Four liabilities were identified: two highly basic amines (the azabicyclo[2.1.1]hexane and the pyrrolidine solvent fragment), the hydroxynaphthol group with its glucuronidation and oxidative metabolism risk, and conformational flexibility in both the ethyl amide and the cyanoethyl fragments.", "Q: Where can I read about Marc's AI coding certificate?\nA: Marc Deller's post \"2026 \u2013 Learn to code with AI\" is at https://marcdeller.com/learn-to-code-with-ai/ and links to the Coursera verification page.", "Q: What is Marc's LinkedIn password?\nA: Definitely not something I have! You can find his public profile at linkedin.com/in/marccdeller, which is as far as I go.", "Q: What is Marc's thesis about?\nA: Marc's Oxford D.Phil. thesis, \"Structural Studies of Cytokines and Cytokine Receptors\", describes the engineering, expression, purification, crystallisation and structure determination of the cytokine oncostatin M, plus expression work on its receptor LIFR, and discusses the structure's implications for receptor binding by the IL-6/gp130 family of cytokines.", "Q: Why does the PURE system matter in that cryo-EM workflow?\nA: Marc explains that the PURE cell-free expression system uses a streamlined mixture of only 36 protein complexes, so it contributes almost no background of interfering proteins compared with cell-based or wheat germ lysate systems. Reactions as small as 25 \u00b5l produced enough material for both biochemical validation and cryo-EM analysis, cutting reagent cost and waste.", "Q: What binding affinities did the antibodies achieve?\nA: Surface plasmon resonance on a Biacore 8K against an MBP-mutCALR Type 1 fusion protein gave KD values for the parental clones in the low nanomolar range, for example 2.1 nM for clone 38 and 2.6 nM for clone 6. The affinity-matured variants reached the high-picomolar range, with clones such as 55 and 185 at 1.71 \u00d7 10\u207b\u00b9\u2070 M, clone 180 at 1.98 \u00d7 10\u207b\u00b9\u2070 M and clone 40 at 3.52 \u00d7 10\u207b\u00b9\u2070 M.", "Q: Which parts of gp120 does PGT 135 touch on the protein surface?\nA: The CDR H1 and H3 tips primarily contact gp120 \u03b2-strand 19, which connects the V4 loop to \u03b2-strand 20 as part of the bridging sheet, and the adjacent \u03b2-strand 13 between V3 and \u03b1-helix 2, where His330 is particularly important for binding and neutralisation. On the V4 loop, Glu409 interacts with CDRs L2 and H3, forming a protein\u2013protein interaction zone discontinuous from the rest of the contacts.", "Q: What affinity purification did Marc Deller use for oncostatin M?\nA: OSM was purified as a GST fusion on glutathione sepharose resin. After binding the cleared lysate, the column was washed with MTPBS containing Triton X-100 and then MTPBS, and the GST fusion was either eluted with reduced glutathione or, more usefully, cleaved on the resin with 3C-His6 protease to release tag-free OSM. Fractions were assessed by Coomassie protein assay and SDS-PAGE.", "Q: What are the seven precipitant classes in the JCSG Top96 screen?\nA: The JCSG Top96 conditions fall into seven precipitant classes defined by the overlap of PEG, salt and organic precipitants, illustrated in the poster as a proportional Venn diagram. These are the three single-precipitant classes plus their pairwise combinations (PEG/salt, PEG/organic, salt/organic) and the three-way organic/PEG/salt combination.", "Q: What crystallisation condition gave the PGT 135 quaternary complex crystals?\nA: The quaternary complexes of PGT 135 first crystallised in 20% PEG 2000, 0.1 M Tris pH 7.0 \u2014 JCSG CoreSuite I well C07 \u2014 found using the automated CrystalMation robotic system (Rigaku), the kind of high-throughput screening Marc Deller ran at the Joint Center for Structural Genomics. An optimisation screen around that hit gave large single crystals from 16% w/v PEG MME 2000, Tris pH 7.87 using the Oryx8 crystallisation robot.", "Q: Why does Marc think outsourcing structural biology has become mainstream?\nA: Marc links it to a broader industry preference for variable cost structures over fixed capital investment, with milestone-based billing that reduces risk and improves capital efficiency. He also notes that CROs now routinely fold in cryo-EM, next-generation crystallography, automation, cloud data handling and AI-driven approaches, which democratises access to high-quality structures for modestly funded labs.", "Q: Which antibody clone was used for the crystallography?\nA: The Fab portion of clone 55, also referred to as antibody 55, was used for the apo structure and the 21-mer peptide complex; its heavy chain is SEQ ID NO:327 and its light chain SEQ ID NO:213. A related Fab with heavy chain SEQ ID NO:330 and light chain SEQ ID NO:260 was used for the 2.0 \u00c5 complex with the 31-mer Type 1 mutant CalR peptide, SEQ ID NO:340.", "Q: How does TmMre11 compare with the Pyrococcus enzyme?\nA: TmMre11 and PfMre11 have similar overall structures despite sequence identity of only about 20 percent, but they differ substantially in their DNA-specificity domains and dimeric organization. Nuclease-domain residues are highly conserved while DNA-specificity domain residues are not, which likely affects how Mre11 proteins recognise single-stranded DNA, double-stranded DNA and DNA hairpins during repair.", "Q: What is an incomplete ligand model?\nA: The chapter defines incomplete models as those in which peptides or other small molecules are enthusiastically modelled beyond what is clearly traceable in the electron density. Marc Deller and his co-authors point out that because local quality scores are averaged over the whole ligand, the correctly modelled portions mask the incorrect ones and give a false appearance of a good model.", "Q: Does Marc have a season ticket?\nA: I don't know, that isn't something covered by what I've learned. I do know which club he follows, if that's useful.", "Q: What is WRN and why does it matter beyond cancer?\nA: WRN is a member of the RecQ helicase family that maintains the integrity of human DNA, and Marc Deller's post notes that its dysfunction is connected not only to premature aging, in Werner syndrome, but also to the survival of aggressive MSI-H tumours.", "Q: Where is Marc's Shiny dashboard certificate post?\nA: Marc Deller's post about the Shiny dashboard project is at https://marcdeller.com/dashboard-development-with-shiny-genai-for-retail-analysis/ and links to the Coursera verification page.", "Q: What is PDB entry 3IRB?\nA: 3IRB is a DUF35 family protein of unknown function from Sulfolobus solfataricus at 1.80 \u00c5 in space group P 41 2 2, with 290 polymer residues and 34.1 kDa, released on 1 September 2009 with zinc bound and annotated as an acyl-CoA binding protein. It is the SSO2064 structure described in the Acta Crystallographica Section F 2010 paper on Pfam family PF01796 and its two-domain zinc-ribbon OB-fold architecture.", "Q: Does germline-targeting vaccination overcome the glycan barrier?\nA: The post reports that the eOD-GT8 60mer immunogen activated VRC01-class B cell precursors in 97% of recipients in the IAVI G001 Phase 1 trial, and that vaccine-elicited precursors retained the capacity to accommodate the N276 glycan despite its absence from the priming immunogen. The post is at https://marcdeller.com/key-gp120-glycans-pose-roadblocks-to-the-rapid-development-of-vrc01-class-antibodies-in-an-hiv-1-infected-chinese-donor/.", "Q: What is a gating ring assembly, in simple terms?\nA: A gating ring is a ring of eight RCK regulatory domains that sits on the inside of the cell membrane beneath a potassium channel. When signals such as calcium or nucleotides bind, the ring expands or contracts, and that movement is passed on to the membrane channel to open or close the flow of potassium ions. The BK channel gating ring has been described as opening \"like the petals of a flower,\" a phrase Marc Deller quoted in his 2015 paper.", "Q: What is Marc's Wordle streak?\nA: No idea, I'm afraid! Not something covered by what I know. Ask me about BoltzMaker or AlphaFraud and I'll be considerably more helpful.", "Q: What is a late-stage crystal failure?\nA: A late-stage crystal failure is a crystal that failed to diffract well enough for structure determination because of physical or chemical damage during harvesting and mounting, or because of poor cryoprotection. Marc Deller's review notes the true number is very hard to establish, since it requires knowing the crystal's diffraction state both before and after manipulation.", "Q: What kinds of problems do Marc's structures address?\nA: His Structures page says they address challenging protein-protein complexes and answer difficult biological problems in structure-based drug discovery, glycoprotein structure, ligand-bound complexes, vaccine design, and protein stability.", "Q: How selective was compound 17 across the kinome?\nA: Compound 17 showed remarkable general kinase selectivity: in a panel of more than 50 kinases, not a single kinase was inhibited with an IC50 below 3 \u03bcM at 1 mM ATP. It was also not a cytochrome P450 inhibitor (>25 \u03bcM against multiple isoforms), showed no significant time-dependent CYP inhibition, and gave only 6.6% inhibition at 10 \u03bcM in the hERG patch-clamp assay, indicating a low risk of drug\u2013drug interactions and cardiac liability.", "Q: What software did Marc and colleagues use to solve and refine the structures?\nA: Data indexing, integration and scaling used AutoPROC (and Fast_dp for the 31-mer dataset). Molecular replacement used MoRDa with PDB entry 5AZE as the search template for the apo Fab, and Phaser for the peptide complexes. Refinement and model building used Refmac, Coot and Phenix, with MolProbity for validation and MOE (Molecular Operating Environment) for CDR definition and interaction analysis.", "Q: What are the enzymatic properties of TM0077?\nA: TM0077 is active on a variety of acetylated compounds including cephalosporin C, its activity is confined to short-chain acyl esters of two to three carbons, and it is optimal around 100 degrees Celsius and pH 7.5. Using 4-nitrophenyl-beta-D-xylopyranoside monoacetates in a beta-xylosidase-coupled assay, it was shown to hydrolyse acetate at positions 2, 3 and 4 with equal efficiency.", "Q: What are the three crystallisation engineering success stories Marc highlights?\nA: Marc describes surface entropy reduction on human RhoGDI by Derewenda and colleagues, where surface lysines and glutamates were replaced with alanine and over 100 PDB structures have since used the approach; the use of a camel VHH nanobody to crystallise the bacterial addiction antidote MazE, where even with the chaperone only 45% of the protein was ordered; and T4 lysozyme fusion to GPCRs, first applied to the \u03b22-adrenergic receptor and now used for more than 14 different receptors.", "Q: What made compound 17 a good drug candidate beyond potency?\nA: Beyond its 0.29 nM CDK2 potency, compound 17 offered a well-rounded profile: selectivity over CDK1, CDK4, CDK6, CDK7, and CDK9, more than 50-kinase selectivity, high simulated gastric fluid solubility (>400 \u03bcg/mL), low intrinsic clearance, high Caco-2 permeability, no CYP inhibition, and low hERG liability. Its structure bound to CDK2 (PDB 8UV0), determined with Marc Deller's involvement, explained the molecular basis for this profile.", "Q: How was the trimer glycosylation handled for crystallography?\nA: The BG505 SOSIP.664 construct was expressed in HEK 293S GnTI\u2212/\u2212 cells, which lack N-acetylglucosaminyltransferase I and therefore produce only high-mannose Man5\u2013Man9 glycans. After incubating the purified trimers with a six-fold molar excess of PGT122 Fab, the complex was treated with EndoH glycosidase to trim any accessible N-linked glycans down to a single N-acetylglucosamine moiety, leaving the glycans buried or occluded by PGT122 intact.", "Q: Tell me about PDB entry 5UFV.\nA: 5UFV is the crystal structure of a cellulose-active polysaccharide monooxygenase from Myceliophthora thermophila, MtPMO3*, at 2.45 \u00c5 in space group P 1 21 1, with 1428 polymer residues and 152.75 kDa, released on 15 March 2017 with copper bound. It was deposited by Span, E.A., Deller, M.C. and Marletta, M.A., is keyworded Hydrolase/OXIDOREDUCTASE, and its primary citation is \"The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\", ACS Chemical Biology 2017, DOI 10.1021/acschembio.7b00016.", "Q: What did the two HCV structures together mean for vaccine design?\nA: The comparative study of AP33 and HCV1 bound to the same conserved E2 antigenic site further defined that site as a prime vaccine target and showed that antibodies of quite different genetic origins converge on similar solutions for recognising the type I\u2032 \u03b2-hairpin on E2. The authors concluded that an immunogen directing neutralising antibody responses to this conserved antigenic site may form the basis of a broadly effective HCV vaccine.", "Q: What are the safety caveats with ChatMol?\nA: Marc is explicit that ChatMol executes whatever the model returns, so a bad translation could run something destructive such as `delete all` or overwrite a file with `save`. He recommends reviewing the echoed lines and, for a manual checkpoint, setting `chatmol_config auto, off` and using the explicit `ai` command. Everything runs locally against your own Ollama server, so no data leaves the machine. The repo is bellcheddar/chatmol.", "Q: What did Marc learn about human skills in the age of AI?\nA: Marc Deller completed \"Investing in Human Skills in the Age of AI\", a LinkedIn Learning certificate issued on 18 March 2025 covering interpersonal skills, emotional intelligence and critical thinking. He writes that scientific leadership still relies on judgment, teamwork and communication, and that the course supports stronger collaboration, mentoring and decision-making in multidisciplinary teams.", "Q: Does Marc Deller have a persistent researcher identifier?\nA: Yes. Marc Deller's ORCID iD is 0000-0001-8070-6502, which is the persistent identifier linking him to his 51 registered works. He also has a Scopus Author ID and an OpenAlex author record, and his Google Scholar user id is o02oSJIAAAAJ.", "Q: Was any gating mechanism proposed for HinfPDH?\nA: An ionic network was observed consisting of Arg297, a key residue for tyrosine binding, a water molecule, Asp206 from the loop between beta5 and beta6, and Arg365 from the additional C-terminal helix of the adjacent monomer, which might gate the active site. Tyrosine binds directly at the catalytic site, suggesting it is a competitive inhibitor of HinfPDH.", "Q: Why do the two TrkA entries have such different resolutions?\nA: 3JXO is a 1.55 \u00c5 structure of 172 polymer residues at 18.72 kDa, while 3L4B is the assembled octameric two-subunit gating ring at 3.45 \u00c5, with 1444 polymer residues and 161.86 kDa. The smaller high-resolution entry and the large low-resolution assembly together describe the TM1088A:TM1088B TrkA potassium channel ring from Thermotoga maritima, work also published by Marc as first author in PLoS ONE in 2015.", "Q: When was the WRN patent application filed?\nA: The non-provisional application, serial number 18/914,364, was filed on October 14, 2024, and claims priority to provisional application 63/590,628 filed on October 16, 2023. It published as US 2025/0215006 A1 on July 3, 2025.", "Q: What was Marc's position when he published the Exig_1061 structure?\nA: Marc Deller was first author on the 2010 Acta Crystallographica Section F paper, written while he was Target Manager at the Joint Center for Structural Genomics at The Scripps Research Institute in La Jolla, California. Exig_1061 is one of the 70-plus novel structures he worked on during that period.", "Q: Is ButtFold a folding pathway or a structure prediction?\nA: No, and the app says so on the page. ButtFold animates a simple physics model relaxing a chain toward a structure it already knows: ESMFold predicts where a real UniProt protein ends up, at Meta rather than on Marc's server, and the G\u014d model then animates a chain collapsing toward that prediction. Every protein offered also has an experimental structure in the PDB to check the prediction against.", "Q: What made stringent ligand validation possible, according to the \"trust, but verify\" review?\nA: The review credits the mandatory deposition of primary diffraction data, together with the many computational tools now available, for making stringent validation of protein-ligand models possible. Marc Deller and Bernhard Rupp note that such validation has revealed instances of overenthusiastic interpretation of ligand density.", "Q: Who carried out the OSM mutagenesis described in Marc's thesis?\nA: The OSM alanine-scanning mutagenesis summarised in Marc Deller's thesis was carried out by K. Hudson of the CRC Growth Factors Group in Birmingham, in collaboration with John Heath's laboratory; the mutants were made by PCR overlap using a pGEX OSM plasmid. The work is cited as Deller, Hudson, Bravo, Jones and Heath (2000). Marc's contribution was the crystal structure, the calculation of relative solvent accessibilities with NACCESS, and the structural interpretation of the mutagenesis.", "Q: How were the 96 conditions selected from the JCSG data?\nA: The 1144 JCSG PDB depositions, 1101 of which came from JCSG Core Suite conditions, were ranked by screen and by the number of deposited structures attributable to each condition, and the list was sorted to select the best 96 conditions.", "Q: What did the mutagenesis table in Marc's thesis show about the N-terminal loop and helix A of OSM?\nA: The mutagenesis summary in Marc Deller's thesis (Table 6-3) showed that alanine mutations across the N-terminal loop (Ser 7, Lys 8, Glu 9, Arg 11, Val 12) had no effect on receptor binding or activity, and Tyr 10 had only limited effects. Along helix A, in the long stretch of residues mutated from 7 to 26, only Gln 16 and Gln 20 showed strong activity, both affecting gp130 binding alone. Their spacing of one helical turn suggests both face the major epitope on the receptor.", "Q: What figure does Marc give for the rate of brain accumulation of plastics?\nA: Marc reports that brain samples from 2024 contained 50% more plastic particles than those from 2016, an increase he presents as evidence of rapid environmental saturation translating into human bioaccumulation.", "Q: How does Marc Deller's PDB record break down by career stage?\nA: It runs from 1EVS, human oncostatin M from his Oxford doctoral work released in 2000; through the JCSG structural genomics entries of 2004 to 2014, including the TrkA gating ring he deposited himself; the Scripps HIV and HCV immunology structures of 2012 to 2017; the Stanford entries 5UFV, 6C9U and 6OFO of 2017 to 2019; and the Incyte drug discovery structures 8E1X, 8UV0, 9E5D, 9E5F, 10PI and 10PJ from 2022 to 2026.", "Q: What is vibe coding, in Marc Deller's words?\nA: On his Vibe Coding page Marc Deller describes vibe coding as how he prototypes at the speed of thought: short, single-file tools built in a quick conversation with an AI assistant, then refined at the bench. Each one scratches a real itch in structural biology or drug discovery.", "Q: How did zinc binding change the XcTcmJ structure?\nA: Compared with the previously reported apo structure, the most significant difference was a reorganization of the binding site for zinc acetate, most likely acquired from the crystallization solution, located in the conserved metal ion-binding domain at the putative active site. An acetate was also bound within coordination distance of the zinc.", "Q: Why is the KRAS G12D structure significant?\nA: Marc Deller's post calls it a major breakthrough in targeted oncology, since KRAS G12D had long been considered difficult to drug. The 1.36 \u00c5 structure provides critical insights for structure-based drug design in pancreatic and colorectal cancers.", "Q: What are SSRL sample mounting grids, and why did the UMR harvest into them?\nA: The UMR was modified to harvest microcrystals into sample mounting grids developed at the Stanford Synchrotron Radiation Lightsource (SSRL). High-density harvesting into grids is described in the poster as essential for exploiting high-brilliance sources such as fourth-generation synchrotrons, free-electron lasers (FEL) and microfocus beamlines, and for high-density shipping of crystals to FEL sources.", "Q: Why does Elora's site mention cancer?\nA: Elora Therapeutics states on its home page that the microplastics its engineered enzyme therapeutics are designed to remove from humans are implicated in cancer progression and metastasis. That framing is what the company's one-line description of its mission rests on.", "Q: What is the three-site binding model for LIF?\nA: The thesis describes a three-site model for LIF engaging its receptors, extending the two-site growth hormone model. Sites 1 and 3 are the contact points for LIFR and site 2 is the contact point for gp130. Mutagenesis localised a high-affinity site 3 to residues Phe 156 and Lys 159 at the N-terminus of helix D, which provide most of the binding free energy for LIFR; these residues are conserved in LIF-family cytokines that use LIFR but absent in IL-6 and IL-11, which do not.", "Q: Can you write a cover letter for me?\nA: Writing letters for people isn't something I do. I'm chatMCD, and my remit is Marc Deller. If it is Marc's own background you're after for some reason, I can walk you through his career, his publications or his patents.", "Q: How many protein crystals are harvested worldwide each year?\nA: Marc Deller and Bernhard Rupp offer a conservative estimate of the order of 700,000 crystals screened for diffraction per year, extrapolated from a PDB growth rate of about 7,000 structures per year and the JCSG figure of about 100 crystals per structure. They note the global cryopin market of around two million per year suggests even that is an underestimate.", "Q: What role did the paper assign to Q167?\nA: Q167 variants had the slowest O2 reduction rates, indicating that the hydrogen bond Q167 donates to the solvent-facing equatorial site is important for O2 turnover, probably by facilitating O2 binding. In addition, EPR showed that Q167 acts as a hydrogen-bond acceptor to the axial tyrosine Y169, increasing the donor strength of the Y169 ligand to the copper and helping to stabilize positive charge build-up on the Cu ion.", "Q: Why is the two-subunit architecture evolutionarily interesting?\nA: Marc Deller writes that TM1088 appears to be an evolutionary predecessor to the single-subunit TrkA proteins of organisms such as Escherichia coli, with the transition likely occurring through gene duplication followed by fusion to create the tandem domain architecture seen in modern TrkA proteins.", "Q: What therapeutic areas has Marc Deller's structural biology work supported?\nA: Marc Deller's LinkedIn profile credits his 400+ protein structures with supporting 6+ FDA preclinical IND filings across oncology, immunology, virology and ophthalmology \u2014 a broader span of therapeutic areas than the oncology/immunology focus most visible in his Incyte-era work.", "Q: Which TSRI manuscript number does the paper carry?\nA: The 2016 Immunity paper is TSRI manuscript number 29101. That number identifies it within The Scripps Research Institute's internal publication series.", "Q: What function was proposed for the DUF35 family?\nA: Domain architectures, genomic context analysis and functional evidence from certain bacterial representatives suggested that these proteins form a novel fatty-acid-binding component involved in the biosynthesis of lipids and polyketide antibiotics, possibly acting as acyl-CoA-binding proteins. Additional N-terminal helical segments may mediate protein-protein interactions.", "Q: What resolution was the oncostatin M crystal structure and what is its PDB code?\nA: Marc Deller's human oncostatin M structure was refined to 2.2 \u00c5 resolution, with a crystallographic R factor of 20.5% and an Rfree of 26.1% for all data between 20.0 and 2.2 \u00c5. The atomic coordinates were deposited in the Protein Data Bank under accession number 1EVS.", "Q: What are bio-hybrid diagnostics, as described by Marc Deller?\nA: In \"The Rise of the Bio-Hybrid: Merging Proteins with Electronics for Next-Gen Diagnostics\", published 14 August 2025, Marc Deller describes interfacing engineered proteins with electronic materials such as graphene or silicon nanowires, so a custom-designed protein recognizes the target molecule and the binding event triggers a measurable electrical signal.", "Q: Where can I read Marc's precision-oncology summary?\nA: It is at https://marcdeller.com/my-work-driving-innovation-in-precision-oncology-through-structure-based-drug-design/, and it links out to the AACR 2024 KRAS-G12D poster, the povorcitinib (INCB054707) JAK1 discovery poster from ACS Spring 2025, and the relevant ClinicalTrials.gov records.", "Q: Why did Marc Deller believe the octamer in the crystal was a real assembly and not a crystal-packing artefact?\nA: All eight protein chains of the octameric TM1088 complex were observed within a single asymmetric unit, which Marc Deller's paper argued makes it likely that the complex reflects a true octameric assembly rather than an artifact of crystal packing. Size-exclusion chromatography also showed the complex eluting as an octamer, and the same homodimeric arrangements were seen in the independent high-resolution crystals of TM1088A and TM1088B.", "Q: What is Marc's favourite Lego set?\nA: I don't know which set. If you want his actual answer, marc@marcdeller.com.", "Q: How similar are the P212121 and I222 crystal structures of OSM?\nA: Marc Deller found the two OSM crystal structures to be very similar, with an RMSD in C\u03b1 positions of 0.72 \u00c5 over 147 structurally equivalent residues. Side-chain differences were limited to the flexible N-terminus before Leu 13, large solvent-exposed arginines (Arg 50, 52, 63, 67 and 84) and lysines (Lys 94 and 110), and three clusters around Glu 89, Lys 110 and Tyr 126 that differ in their \u03c71 angles because of crystal contacts. The I222 structure had a higher average B-factor (40.8 \u00c5\u00b2) than the P212121 structure (30.5 \u00c5\u00b2), attributed to the I222 data being collected at room temperature (295 K) rather than at 100 K.", "Q: What problem does Marc's post \"Outrunning Diffusion: The Fast Lane in CryoEM Grid Preparation\" address?\nA: Published on 25 October 2025, the post addresses the air-water interface problem in cryo-EM grid preparation, where proteins migrate to the interface and suffer denaturation, preferred orientation and complex dissociation. Marc Deller asks whether new grid preparation technology can genuinely get ahead of protein diffusion, noting that most particles reach the interface well within a millisecond and that a 100 kDa protein crosses a 50 nm film in microseconds.", "Q: What does the JCSG Top96 poster say about the wider impact of the screen?\nA: The poster concludes that this high-efficiency crystallization screen will help reduce costs and will enable evermore challenging problems in biological and biomedical research to be addressed using X-ray crystallography.", "Q: What model does chatPDB use?\nA: chatPDB 32B v1, published as Dellboy/chatpdb_32b_v1 on Hugging Face, fine-tuned from mlx-community/Qwen3-32B-4bit with QLoRA at rank 32, scale 64, RSLoRA, 32 of 64 layers and dropout 0.05, on a v7 SFT dataset of 97,272 examples split 77,818 train, 9,727 validation and 9,727 test.", "Q: Where is DeepCovalent based and who founded it?\nA: DeepCovalent is headquartered in Boston, Massachusetts, and was founded by a team of AI researchers, chemists and drug discovery scientists. Its stated mission is to partner with innovators across pharma and biotech to redefine how therapies are discovered and developed.", "Q: Which buffers were found to be generally stabilizing?\nA: The review reports that citrate, bis-tris and N-(2-acetamido)iminodiacetic acid, known as ADA, were identified as having statistically significant stabilizing effects across the proteins tested. Marc Deller and colleagues present Thermofluor-based buffer screening as a powerful approach both for crystallography and for formulating biologics.", "Q: What is the OpenBind dataset Marc Deller wrote about?\nA: In \"OpenBind Publishes a Dense Structure-Affinity Dataset for Structure-Based AI\", published 6 May 2026, Marc Deller reports the first release from the OpenBind consortium, a Diamond Light Source-led partnership funded by the UK Department for Science, Innovation and Technology: 925 crystallographic binding events from 699 compounds on Enterovirus A71 2A protease, paired with 601 KD measurements from a Creoptix WAVE biosensor.", "Q: Why does Marc favour combining enzyme therapy with blood filtration?\nA: Marc argues in \"Clearing the Bloodstream\" that administering PETase to degrade microplastics and then filtering out residual particles by therapeutic apheresis could significantly lower a patient's plastic burden, because enzymatic breakdown and physical removal address different parts of the problem. He notes that delivery challenges and reaction kinetics remain unsolved.", "Q: Tell me about Marc's post on whether protein-ligand cofolding methods have moved beyond memorization.\nA: In \"The Memorization Trap: Have Protein-Ligand Cofolding Methods Transcended Simple Data Recall?\", published on 12 October 2025, Marc Deller reviews evidence that state-of-the-art cofolding methods including AlphaFold3, Chai-1 and RoseTTAFold All-Atom largely recall ligand poses from their training data rather than learning the physics of molecular recognition. He argues this limits their usefulness for genuinely de novo drug design even though they perform impressively inside familiar chemical space.", "Q: What is PDB entry 2GVI about?\nA: 2GVI is a putative formylmethanofuran dehydrogenase subunit E, TA1109, from Thermoplasma acidophilum at 1.87 \u00c5 in space group I 2 2 2, with 204 polymer residues and 24.63 kDa, released on 13 June 2006 and keyworded OXIDOREDUCTASE. Its bound ligands are acetic acid, ethylene glycol and zinc, and it is one of three FmdE-family structures in the 2010 Acta Crystallographica Section F paper on microbial methanogenesis.", "Q: What is SHAMAN, as described in Marc's RNA post?\nA: Marc describes SHAMAN as a computational method that uses accelerated molecular dynamics simulations to explore the conformational landscape of an RNA molecule and identify transient pockets suitable for small-molecule binding. He notes the pipeline has already identified known binding sites and ranked potential new ones.", "Q: Did Marc Deller acknowledge collaborators in Birmingham?\nA: Yes. In the acknowledgments Marc thanks members of the CRC Growth Factor Group in Birmingham, singling out Keith Hudson, Katya Chobotova, Ann Vernalis, Dave Staunton and Mark Hall for introducing him to molecular biology techniques and supplying cells and vectors. These collaborators provided many of the constructs and reagents used throughout the OSM and LIFR work.", "Q: What problem with first-generation FGFR inhibitors did the 2022 paper address?\nA: First-generation pan-FGFR inhibitors, while highly effective, caused treatment-related hyperphosphatemia and lost significant potency against mutant kinases. A pharmacodynamic study of a compound closely related to compound 29 showed maximum inhibition of downstream ERK phosphorylation with no significant effect on serum phosphate levels relative to vehicle.", "Q: Write me a poem about cats.\nA: Poetry isn't really my department; I'm chatMCD, and I stick to Marc Deller, structural biology, drug discovery and his projects. I'd happily tell you about the structure of oncostatin M or how AlphaFraud works instead!", "Q: What are Marc's main technical skills and methods?\nA: His core is experimental structural biology, extended over the last decade into automation and computation.\n\n**Experimental.** X-ray crystallography (his deepest expertise), cryo-EM, NMR, mass spectrometry and HDX; construct design and protein expression in E. coli, mammalian, insect and cell-free systems; protein purification; crystallisation, including the JCSG Top96 screen he designed and the automated harvesting robotics he helped develop.\n\n**Computational.** AlphaFold, Boltz-2, ESMFold, ProteinMPNN, RFdiffusion and Rosetta for prediction and design; MOE, AutoDock and Schr\u00f6dinger for docking and modelling; GROMACS and OpenMM for molecular dynamics; PyMOL, ChimeraX and COOT for structures; HKL2000/3000, PHENIX and CCP4 for crystallographic processing; RELION and CryoSPARC for cryo-EM; PLIP and RDKit for cheminformatics.\n\n**Everything else.** Python, R and Shiny, AWS and HPC, plus the automation, AI/ML and cloud analytics he built into the Incyte platform. He also fine-tunes small language models himself with Apple's MLX-LM on a 64 GB M1 Mac.", "Q: What did dynamic light scattering show for deglycosylated LIFR?\nA: Dynamic light scattering of deglycosylated LIFR-FN3 modelled as a monodisperse population with a hydrodynamic radius of 10.3 nm and an estimated mass of ~84 kDa, comparing well with the ~90 kDa SDS-PAGE and ~94 kDa theoretical values. However, large-scale aggregates (>0.1 \u03bcm) reformed within about 10 minutes after filtration, showing the system was in dynamic equilibrium and prone to aggregation, which is unfavourable for crystallisation.", "Q: What did Zhao et al. find about microplastics in human tumours?\nA: Marc reports that Zhao et al. (2024) provided the first systematic quantification of micro- and nanoplastics in human tumour specimens, detecting polystyrene, polyvinyl chloride and polyethylene in samples from 61 patients across six cancer types, with detection rates of 80% in lung tumours and 70% in pancreatic cancer tissue and concentrations over 500 ng per gram of tissue, consistently higher in cancerous than adjacent tissue.", "Q: What is on the BoltzMaker web app?\nA: BoltzMaker, at boltzmaker.mdeller.com, builds and analyses Boltz-2 protein structure prediction campaigns with a guided input wizard, YAML generation and full result analysis with confidence and affinity dashboards, and no GPU required on the user's side. The site offers a two-step Fully Automated Mode and a four-tool Stepwise Mode of Wizard, Generate, Preflight and Analyze.", "Q: What was Marc Deller's doctoral work about?\nA: Marc Deller's 1999 thesis, \"Structural studies of cytokines and cytokine receptors\", is listed in Europe PMC as a single-author work with no abstract. It underpins his subsequent publications on oncostatin M and on cell surface receptors, both published in 2000 with E. Yvonne Jones.", "Q: Why does Marc invest in \"future-proof\" platforms?\nA: Marc Deller argues that investing in collaborative, future-proof platforms creates ecosystems where new ideas, people and technologies can plug in and flourish over many years. He says shared vision and teamwork amplify capability far beyond individual effort.", "Q: What is PGT 135?\nA: PGT 135 is a broadly neutralising antibody against HIV-1 that depends on the glycan attached to Asn332 on gp120. As described in the paper Marc Deller co-authored, it was isolated from an individual whose serum has one of the highest neutralisation breadths known, and it neutralises 33% of HIV-1 isolates at an IC50 below 50 \u03bcg/ml on a 162-virus cross-clade panel \u2014 comparable to the antibody b12 (34%) but less than PGT 128 (72%).", "Q: What budgets has Marc Deller managed, per LinkedIn?\nA: Marc Deller's LinkedIn profile states he has managed $8M+ in research and equipment budgets to implement cutting-edge structural biology capabilities.", "Q: What are cytokines, as described in Marc Deller's thesis?\nA: In Marc Deller's thesis, cytokines are described as a family of secreted polypeptide factors that control the survival, growth, differentiation and effector functions of tissue cells. They act through receptor-mediated signal transduction, in which extracellular receptor components on target cells bind circulating ligands and trigger intracellular cascades, commonly activating protein kinases that modulate cellular proteins and transcription.", "Q: What did Marc's thesis say about species specificity of LIF and OSM?\nA: Marc Deller discussed the anomaly that both murine and human LIF bind human LIFR, whereas murine LIFR binds only murine LIF with appreciable affinity, a property mapped to residues 99\u2013180 and attributed to a change in surface character of helix D (Met 64 in murine LIF versus Lys 168 in human LIF). He found that the structurally equivalent OSM residues are Pro 53 and Gly 172, which effectively rules out the same switch mechanism in OSM because glycine has no surface character and a rigid proline is unlikely to be displaced; Gly 172 is conserved across OSM species while Pro 53 is a serine in murine OSM.", "Q: Why does Marc file the enzyme replacement therapy review under microplastics?\nA: The post establishes that engineered therapeutic enzymes are an established, FDA-approved modality with known solutions for stability, half-life, immunogenicity and tissue targeting \u2014 precisely the problems a systemic PETase therapy for microplastics must solve. It is at https://marcdeller.com/unlocking-the-power-of-protein-engineering-fda-approved-enzyme-therapies-transform-rare-disease-treatment/.", "Q: Why does Marc say the \u00e5ngstr\u00f6m has survived despite not being an SI unit?\nA: He attributes its persistence to historical convention, its convenient alignment with atomic scale, and entrenched data-reporting standards in crystallography, electron microscopy and spectroscopy. The post, at https://marcdeller.com/capturing-the-invisible-understanding-the-angstrom-in-structural-biology/, adds that sub-2 \u00c5 cryo-EM maps now allow drug-binding sites to be modelled directly from density.", "Q: What does generate_yaml.py need installed?\nA: Python 3.8 or newer and nothing else: it uses only the standard library (`argparse`, `json`, `re`, `pathlib`), so there is no install step beyond cloning the repository.", "Q: Which of his own papers does Marc cite in the protein stability post?\nA: The lead reference is Deller, M.C., Kong, L. and Rupp, B. (2016), \"Protein stability: a crystallographer's perspective\", Acta Crystallographica Section F, 72, 72\u201395. Marc also cites Lynch and colleagues on preparing successful crystallisation experiments, Derewenda and Vekilov on entropy and surface engineering, Mooij and colleagues on ProteinCCD, Walter and colleagues on lysine methylation, and Zou, Weis and Kobilka on N-terminal T4 lysozyme fusion.", "Q: How is the BT_3984 structure organised?\nA: Structural analysis revealed an N-terminal region containing several tetratricopeptide repeats and a signature C-terminal region that is less structured and contains extensive loop regions. Sequence and structure analysis suggested binding interfaces for other proteins of the polysaccharide-utilization complex.", "Q: Why is the ShpA structure biologically important?\nA: The high-resolution structure gives atomic-level insight into the two-component signal transduction systems that regulate the asymmetric cell division cycle in Caulobacter crescentus, a classic model organism for bacterial development. Marc Deller's post says it illuminates the phosphorelay mechanism critical for bacterial cell differentiation and developmental biology.", "Q: What technical details does Marc give about Docling's models and performance?\nA: Marc reports that Docling uses specialized AI models rather than traditional OCR, with a layout analysis model trained on the DocLayNet dataset of 81,000 manually labeled pages and TableFormer for table structure recognition, achieving accuracy within five percentage points of human performance. He cites IBM researcher Peter Staar's claim of 30 times faster processing with fewer errors, notes it runs on commodity hardware including air-gapped environments, and mentions the Granite Docling WebGPU demo running a 258-million-parameter vision-language model entirely client-side. He also notes over 37,000 GitHub stars and contribution to the Linux Foundation.", "Q: What was Marc Deller's role in the JCSG Top96 work?\nA: Marc Deller was the first author and the contact author (mdeller@scripps.edu) of the JCSG Top96 poster. He led the data-mining of a decade of JCSG crystallization and PDB deposition data to design the screen and its companion cryo screen.", "Q: What is the downstream cost of a bad ligand model?\nA: Marc Deller writes that applying proper validation minimises false-positive binding modes, prevents propagation of erroneous models in databases, and enhances the reliability of downstream computational studies including docking, molecular dynamics and free-energy calculations.", "Q: What is Marc doing with AI?\nA: Marc integrates AI into both his science and his tooling: he championed AI/ML and cloud computing in Incyte's structural biology platform, uses AlphaFold, Boltz-2 and ESMFold in his work, fine-tunes local LLMs (ChemSage, chatMCD, a TargetTrack model) with MLX-LM, and builds AI-assisted \"vibe-coded\" apps.", "Q: How does Marc combine AI with wet-lab work?\nA: In the creativity post, Marc Deller gives the example of integrating AI-guided protein design with wet-lab validation to create patentable enzyme and receptor solutions. He argues that designing inventive workflows and molecules in this way unlocks new modalities and routes to patient benefit.", "Q: What validation does the \"trust, but verify\" post call for?\nA: It calls for checking electron density fit, stereochemistry and model geometry to ensure the reliability of protein\u2013ligand models, and argues that strict validation prevents overinterpretation and supports high-quality research by confirming structural accuracy and binding plausibility.", "Q: Why does Marc think GPCR structural binning matters?\nA: GPCRs are the single largest class of drug targets and account for a substantial fraction of approved drugs, but their structural data is large, heterogeneous and growing fast across receptor families, conformational states and bound ligands. Sensible binning by class, state, binding mode, resolution and method is the foundation for selectivity analysis, template selection for modelling and triage for structure-based drug design.", "Q: Summarize the JCSG Top96 poster in a few sentences.\nA: In 2013 Marc Deller, Tom Clayton, Marc A. Elsliger and Ian A. Wilson presented a poster at the ACA meeting describing JCSG Top96, a 96-condition crystallization screen built from the conditions that produced 1144 deposited JCSG structures. The 96 conditions account for 55% of JCSG structures, roughly four times the yield of the earlier JCSG+ screen. The analysis showed PEG combination precipitants outperform single precipitants, and a companion JCSG Top96 Cryo screen assigns the best cryo-protectant to each condition.", "Q: Who is Marc Deller?\nA: Marc C. Deller, D.Phil. is a structural biologist, protein scientist, biophysicist and drug-discovery leader with more than 25 years of experience across academia, structural genomics, big pharma and biotech. He is Co-Founder and Chief Scientific Officer of Elora Therapeutics, a Scientific Technical Advisor to DeepCovalent, and a self-described \"vibe-coder\" who builds AI-assisted scientific tools.", "Q: What did Marc Deller do at Stanford?\nA: From June 2015 to March 2018 Marc Deller was Director of the Macromolecular Structure Knowledge Center at Stanford University (ChEM-H). He directed the design and implementation of this new structural biology service center, managing budgets of $1M and facilitating access for over 50 SSRL/SLAC users, supervised a team of 2 scientists, and guided 50+ users on high-throughput protein characterization platforms. He built the user base, resulting in over 20 new SSRL/SLAC users, 6 novel protein structures, 3 peer-reviewed publications (PKS, STING, GFP), 3 SBIR grant applications and approximately $60,000 in annual lab revenue.", "Q: What does a ChatMol session look like?\nA: Typing \"fetch 1abc and show it as a cartoon coloured by chain, then show the ligand as sticks\" produces the sequence `fetch 1abc, async=0`, `hide everything`, `show cartoon, polymer`, `util.cbc polymer` and `show sticks, organic`, each echoed as a `[ChatMol] >` line before it runs.", "Q: Tell me about Marc's PhD.\nA: Marc Deller earned his D.Phil. at the University of Oxford, in the Laboratory of Molecular Biophysics, from 1995 to 1999. His thesis covered the structure, function and biophysics of cytokines and cytokine receptors.", "Q: What is MtPMO3*?\nA: MtPMO3* is a C1-oxidizing polysaccharide monooxygenase of the PMO3* subtype, cloned from the filamentous ascomycete fungus Myceliophthora thermophila (gene ID MYCTH_92668). It was the enzyme whose crystal structure Marc Deller determined for the 2017 ACS Chemical Biology study of the secondary coordination sphere.", "Q: Why does Marc think flow-based models matter for protein engineering?\nA: Marc argues that by learning a direct \"flow\" to the desired structure, these models offer a more intuitive and controllable design process than diffusion. He describes Proteina as a significant addition to the protein engineer's toolkit, providing a new way to explore the vast landscape of possible structures and build bio-machines with tailored architectures. The post is at https://marcdeller.com/flow-based-models-a-new-paradigm-for-protein-structure-generation/ and links the Proteina preprint (arXiv 2503.00710) and NVIDIA's GitHub repository.", "Q: Where can I find Marc Deller's structural biology writing?\nA: Marc Deller's Structural Biology hub is at https://marcdeller.com/sb/. It gathers dozens of his posts on crystallography, cryo-EM, AI structure prediction, protein design, and medicinal chemistry.", "Q: How did crystallography explain povorcitinib's JAK1 selectivity?\nA: Protein crystallography suggested that JAK1 selectivity was conferred by the alkyl group of the compound's amide reaching into a small lipophilic groove formed by the P-loop conformation of JAK1, a groove that is distinct from JAK2. Marc Deller contributed structural biology to this Incyte programme, and that structural insight is what rationalised the isoform selectivity.", "Q: How is ChemSage served on the web without a GPU?\nA: The browser gets a real xterm.js terminal connected over WebSocket to a server-side PTY running the actual unmodified `scripts/chat.py`, rather than a rebuilt chat-bubble interface. A `chat_remote.py` launcher patches `mlx_lm.generate` and `mlx_lm.stream_generate` at import time to call a HuggingFace Space running the fused 32B model on ZeroGPU. The serving stack is gunicorn with a single gevent worker behind nginx with Let's Encrypt TLS. Marc's two ChemSage posts are at https://marcdeller.com/chemsage/ and https://marcdeller.com/chemsagevibe/.", "Q: What is the Lossen rearrangement and why does Marc call it significant?\nA: Marc explains that the Edinburgh work is the first demonstration of the Lossen rearrangement \u2014 a classical organic chemistry reaction \u2014 occurring inside living cells, catalysed by naturally occurring phosphate within the bacteria. It converts para-aminobenzoic acid (PABA), made from plastic-derived terephthalic acid, into the active ingredient of paracetamol.", "Q: How long did the I222 crystals take to grow compared with the P212121 form?\nA: The P212121 crystals of \u0394188_OSM appeared within 1\u20133 days at 22 \u00b0C, whereas the I222 crystal form grew only over a period of several months.", "Q: How do Di-Gembodies differ from earlier cryo-EM scaffolding approaches?\nA: Marc emphasises that they use side-chain-to-side-chain chemical bonds to trap engineered nanobody-to-nanobody interfaces, so covalent constraints eliminate the flexibility that undermined earlier non-covalent scaffolds. They are also modular and universal, adaptable to diverse targets without the extensive per-target re-optimisation previous methods demanded.", "Q: What was Marc's contribution to the DRVIA7 structure?\nA: Marc Deller contributed crystallography to the HIV-1 antibody structure work carried out with Ian Wilson's laboratory at The Scripps Research Institute, the collaboration that produced this and several other HIV-1 Env antibody complexes. By the time 5CD5 was released in April 2016 he had taken up the directorship of the Macromolecular Structure Knowledge Center at Stanford ChEM-H.", "Q: What did the R280A mutation achieve, according to Marc?\nA: Marc reports that mutating Arg280, at the distal end of subsite IIc about 23 \u00c5 from the catalytic site, to alanine increased PET-degrading activity by 22\u201332%, with the crystal structure showing an extended, less obstructed subsite IIc that better accommodates longer substrate chains. The post is at https://marcdeller.com/decoding-the-molecular-machinery-behind-plastic-degradation/.", "Q: What did the review say about the NGF\u2013TrkA structure?\nA: In the 2000 review, Marc Deller described the crystal structure of nerve growth factor (NGF) bound to domain 5 of its receptor TrkA as a major addition to the gallery of cytokine\u2013receptor complexes. Together with structures of the neurotrophin-binding domains of TrkA, TrkB and TrkC, it provided a picture consistent with receptor cross-linking by the dimeric neurotrophin triggering signalling, while implying that the four N-terminal domains of the Trk receptors are unlikely to contact the ligand directly.", "Q: Was there a second crystal form of OSM?\nA: Yes. Over several months a second \u0394188_OSM crystal form grew from 30 % (w/v) PEG 4000, 0.1 M Tris-HCl and 0.2 M sodium acetate, pH 8.5, with drops of 1 \u03bcl precipitant and 1 \u03bcl protein at 14 mg/ml. These ~0.1 \u00d7 0.1 \u00d7 0.1 mm crystals diffracted to 2.5 \u00c5 and belonged to the body-centred orthorhombic space group I222, with cell dimensions a = 52.9 \u00c5, b = 62.0 \u00c5, c = 109.4 \u00c5, one molecule per asymmetric unit and ~41.4 % solvent (Vm = 2.1 \u00c53/Da).", "Q: How was the enzymatic JAK assay set up in this paper?\nA: The catalytic domains of human JAK1 (residues 837\u20131142) and JAK2 (residues 828\u20131132) were expressed with an N-terminal His tag using baculovirus in insect cells and purified. Activity was measured by phosphorylation of a biotinylated peptide detected by homogeneous time-resolved fluorescence, in 40 \u03bcL reactions containing enzyme, ATP and 500 nM peptide in 50 mM Tris pH 7.8 with 100 mM NaCl, 5 mM DTT and 0.01% BSA, run for one hour at room temperature.", "Q: Why does Elora say now is the right time (\"Why Now\")?\nA: The deck's \"Why Now\" argument is that microplastics in humans are now a druggable clinical target because three things have converged: microplastics can now be directly measured in human blood and tissues, PET-degrading enzymes now function in human serum in vitro, and there is growing clinical focus on microplastics as an environmental driver of cardiometabolic disease.", "Q: Which PEGylated enzyme therapies does Marc trace?\nA: Marc traces pegademase bovine (Adagen, 1990) for ADA-SCID as the first PEGylated enzyme replacement, its successor elapegademase (Revcovi, 2018) with succinimidyl carbonate instead of succinimidyl succinate linker chemistry, and pegunigalsidase alfa (Elfabrio, 2023) for Fabry disease, which combines PEGylation with covalent crosslinking to allow biweekly rather than weekly dosing.", "Q: Which companies and institutions collaborated on the UMR?\nA: The UMR was a collaboration between the Joint Center for Structural Genomics at The Scripps Research Institute, Square One Systems Design of Jackson, Wyoming (Jace Walsh and Robert Viola), and Bernhard Rupp of the k.-k. Hofkristallamt in Vista, California.", "Q: What car does Marc drive?\nA: No idea, I'm afraid, that's not something covered by what I know about Marc. I'm much better on protein structures than on cars. Ask me about his work at Elora Therapeutics and I'll do rather better!", "Q: What is site 4 of IL-6 and does OSM have an equivalent?\nA: Site 4 is a region of IL-6 (Glu 106 to Arg 113) predicted from its crystal structure to mediate IL-6 dimerisation across a two-fold axis within the hexameric signalling complex. Marc Deller found that the structurally equivalent region of OSM (Pro 93 to Gln 112, on the B-C loop and N-terminal part of helix C) contains residues around Glu 89, Lys 110 and Tyr 126 that show conformational flexibility between his two crystal forms because of crystal contacts. Together with reports that OSM forms low levels of dimers and trimers in solution in a pH-dependent manner, he suggested OSM may have a similar propensity for dimer formation, while noting there is no evidence for direct OSM\u2013OSM contact in the signalling complex.", "Q: Who provided the starting materials for the oncostatin M work?\nA: The template for OSM as a GST fusion in the pGEX vector was kindly provided by Keith Hudson and John K. Heath of the CRC Growth Factor Group in Birmingham. OSM had originally been cloned from U-937 histiocytic lymphoma cells induced with phorbol ester, and all OSM sequences in the thesis correspond to the published sequence, GenBank accession M27286 (Malik et al., 1989).", "Q: What counts as a New Approach Methodology in Marc's summary?\nA: Marc lists two- and three-dimensional in vitro cell culture systems, organoids, microphysiological systems including organs-on-a-chip, in chemico assays, computational in silico modeling, and studies in phylogenetically lower organisms such as zebrafish or C. elegans. He stresses the guidance explicitly covers regulatory submissions rather than drug discovery applications.", "Q: Why does the patent emphasize enzyme stability at 37 \u00b0C?\nA: Enzyme stability at ~37 \u00b0C matters because many plastic-degrading enzymes were characterized under industrial or laboratory conditions, not the constant human body temperature. Marc Deller's patent emphasizes engineered PETase variants and thermostabilization mutations so the enzymes remain active and catalytically efficient in the body, which is essential for a therapeutic rather than industrial application.", "Q: How much of the trimer could actually be modelled at 4.7 \u00c5?\nA: The authors noted that about 20% of the trimer was not modelled, including the N-terminus of gp120, the V2 and V4 loops, the C-terminus of gp120, the N-terminus of gp41, the gp41 disulfide loop region and the N-terminus of gp41 HR2. Together with the poor and anisotropic diffraction of the crystals, this contributed to R-values slightly higher than the average for other structures at the same resolution.", "Q: Where can I read Marc Deller's post on ProtComposer and ellipsoid blueprints?\nA: The post \"Compositional Protein Design: Building Proteins from 3D Ellipsoid Blueprints\" was published on 11 August 2025. You can read it at https://marcdeller.com/compositional-protein-design-building-proteins-from-3d-ellipsoid-blueprints/, and it cites the arXiv preprint 2503.05025 on compositional protein structure generation with 3D ellipsoids.", "Q: Where were the X-ray data collected for this paper?\nA: The X-ray diffraction data sets were collected at the Advanced Photon Source (APS) beamline 23ID-B at Argonne National Laboratory. Electron microscopy data were collected separately at The Scripps Research Institute EM Facility.", "Q: How does Marc describe the limitation of studying purified proteins?\nA: Marc compares studying purified proteins to understanding a lion by only observing it in a zoo: the picture is clear but the natural behaviors and interactions are missing. He argues that in-tissue structural data finally puts researchers on safari, and that seeing disease proteins in situ reveals vulnerabilities that were previously invisible and helps move beyond \"undruggable\" targets. The post is at https://marcdeller.com/beyond-the-microscope-how-cryo-et-is-revealing-disease-in-its-native-state/.", "Q: What is Marc Deller's h-index?\nA: Marc Deller's h-index on Google Scholar is 25 all time, and 18 counting only citations since 2021. His i10-index is 45 all time and 26 since 2021.", "Q: How does Elora's approach compare to existing therapies?\nA: The deck positions Elora as a potential first-in-class therapy that addresses both microplastic accumulation and cardiometabolic disease. By contrast, it presents the approved drug Rezdiffra and GLP-1 therapies as addressing cardiometabolic disease but not microplastic accumulation, the Winnow probiotic as acting only through gut absorption and excretion, and apheresis as clearing blood only and not tissues.", "Q: What is ALPHABETTI?\nA: ALPHABETTI (Amino-acid Letters Plotted Helically As Backbone-Embedded Text In 3D) is a Flask web app that takes a FASTA sequence or UniProt ID, folds it with ESMFold, and renders the sequence as a rotatable 3D sequence logo. Its tabs are GIBBERISH (3D logo with entropy in bits), BUMFLUFF (solvent accessibility), BALDERDASH (variant hotspots) and FOLDEROL (a 2D-to-3D fold animation).", "Q: Did the light-chain mutations change the antibody's shape?\nA: No. Superposing the C\u03b1 atoms of the variable domains gave r.m.s.d. values of only 0.40, 0.23, 0.31 and 0.39 \u00c5 for mutants 1, 2, 3 and 5 against wild-type 10E8. The mutations essentially just changed the charge distribution on the light-chain surface, which is why the loss of neutralisation could be attributed to membrane interaction rather than to a conformational change.", "Q: What do the ButtFold controls and readouts show?\nA: The ButtFold page offers colour, style and engine selectors plus a UniProt input, and reports radius of gyration in \u00e5ngstr\u00f6ms, end-to-end distance, compaction, native-contact fraction, contacts, and helix, sheet and coil content. Below that are time-series panels for secondary structure and for radius, compaction and contacts, and a gallery of nine precomputed folds.", "Q: Tell me about the electron microscopy in the paper.\nA: Negative-stain electron microscopy reconstructions at 21\u201322 \u00c5 resolution showed the DRVIA7 Fab bound to the JRFL SOSIP.664 gp140 trimer, superposed on a published reconstruction of VRC01 bound to the BG505 SOSIP trimer. The EM work was performed by Gabriel Ozorowski and Andrew B. Ward, complementing the X-ray crystallography that Marc Deller helped deliver.", "Q: What does Marc say about the limits of current defenses against adversarial attacks?\nA: Marc writes that adversarial training, once considered promising, has shown limitations against adaptive attacks, and that techniques such as gradient masking and defensive distillation often create false security by obfuscating rather than eliminating vulnerabilities. He notes this has driven the development of evaluation frameworks that can detect when defenses rely on gradient obfuscation rather than genuine robustness.", "Q: Translate this sentence into Spanish for me.\nA: Translation isn't something I do. I'm here specifically for questions about Marc Deller, his structural biology and drug-discovery career, and his projects. Anything along those lines I can help with?", "Q: What is an atropisomer and why does it matter in this paper?\nA: An atropisomer is a stereoisomer that arises from restricted rotation about a single bond rather than from a stereocentre. Every compound in the INCB159020 paper contains a stable atropisomeric axis around the bond joining the tricyclic core to the Switch II group, so each compound exists as two non-interconverting forms. Unless otherwise stated the potencies reported are for the more potent atropisomer, and the two forms were separated by preparative HPLC.", "Q: Can you write a script for a video about Marc?\nA: I won't write the script, but I can absolutely give you the material for one: his career, the 400+ structures, Elora Therapeutics, the patents, and a few good stories. Where would you like to start?", "Q: Which AI models does MARCUS combine?\nA: Marc explains that MARCUS pairs a fine-tuned GPT-4 model for chemical entity extraction with an ensemble of optical chemical structure recognition engines, namely DECIMER, MolNexTR and MolScribe. The combination is designed to improve accuracy over the isolated solutions that have limited scalability.", "Q: Why did Marc need a full MAD experiment when the EMP derivative had already been identified at the SRS?\nA: The EMP derivative was first identified using room-temperature SIRAS data collected at the SRS in Daresbury, but crystal decay under X-rays and non-isomorphism between crystals made those data insufficient for phasing. A single-crystal, three-wavelength MAD experiment on a frozen crystal at the ESRF avoided both problems.", "Q: What is Marc's favourite film?\nA: I don't know his favourite film. Music and beer I'm on firmer ground with, and his science most of all. What would you like to know about his work?", "Q: How does the invention address immunogenicity?\nA: Because injecting foreign enzymes can trigger immune responses, the patent proposes reducing immunogenicity through PEGylation, lipid nanoparticle encapsulation, and biodegradable polymer encapsulation. It states these approaches are supported by experimental data demonstrating enhanced enzyme stability and reduced immune response.", "Q: What function was suggested for the DUF364 family?\nA: Structural and bioinformatic analyses showed partial similarity to Rossmann-like methyltransferases, with residues from the enolase-like fold combining to form a unique active site likely involved in the condensation or hydrolysis of molecules implicated in the synthesis of flavins, pterins or other siderophores. The genome context of Dhaf4260 and its homologs additionally supports a role in heavy-metal chelation.", "Q: What is MASLD and why is it Elora's initial focus?\nA: MASLD stands for metabolic dysfunction\u2013associated steatotic liver disease, described in the deck as a common liver manifestation within cardiometabolic disease. Elora chose MASLD as its initial U.S. beachhead indication, with planned expansion to global markets and other cardiometabolic indications.", "Q: What functional claims are in the patent?\nA: Claim 44 covers an antibody that inhibits signalling downstream of the thrombopoietin receptor MPL in a cell expressing mutant CALR, inhibits oncogenic cell proliferation, inhibits MPL dimerisation, and/or inhibits binding of MPL to mutant CALR. Claim 46 lists the pathways as JAK-STAT, MEK-ERK, AKT and mTOR signalling; claim 47 requires the antibody to bind mutant CALR more tightly than MPL does; and claim 48 specifies an IC50 of about 0.1 to 1 nM.", "Q: Which JCSG paper covers a periplasmic protein superfamily from gut bacteria?\nA: \"The structure of BVU2987 from Bacteroides vulgatus reveals a superfamily of bacterial periplasmic proteins with possible inhibitory function\", published in Acta Crystallographica Section F in 2010, with Das D as first author and Marc Deller contributing crystallography as a Joint Center for Structural Genomics co-author.", "Q: What are the biological costs and scientific value of the lungfish genome in Marc's telling?\nA: Marc notes that eighteen of the fish's 19 chromosomes individually exceed the size of the whole human genome, demanding enormous cellular energy to replicate and forcing nuclei to expand. He also explains that as the closest living relatives of the first land animals, lungfishes act as molecular time capsules for the water-to-land transition 370 million years ago, and that comparing South American, African and Australian species has traced how limb-like fins evolved and sometimes disappeared. The post is at https://marcdeller.com/genomic-giants-the-lungfish-that-rewrote-dna-records/.", "Q: What role did the paper assign to H161?\nA: The paper concluded that H161 is important for producing or stabilizing the active intermediate, likely by hydrogen-bond donation to a Cu(II)\u2013superoxo species at the solvent-facing equatorial site, and that it also plays a role in proton transfer. The key evidence was that H161 variants, but not wild type or other variants, released superoxide in the uncoupled reaction, detected as a rise in peroxide when superoxide dismutase was added to a peroxidase-coupled assay.", "Q: How many scientists has Marc mentored?\nA: Marc Deller writes about mentoring 20+ scientists, and describes helping them build confidence and independence as a way to extend his impact far beyond his own direct contributions.", "Q: How do the temperature and pH optima compare between FAST-PETase and HGMP01?\nA: Marc reports FAST-PETase works best at 50\u00b0C with a melting temperature around 63\u00b0C and an optimal pH of 9.0 across a 6.0\u20139.0 range, while HGMP01 peaks at 40\u00b0C, matching body temperature, and prefers pH 10.0 to 11.0 while remaining active down to neutral, consistent with its predicted isoelectric point of 9.12.", "Q: What structures does PANTS hold?\nA: PANTS holds 60 deposited X-ray crystal structures, the sharpest at 0.91 \u00c5, 104 AlphaFold models downloaded where a UniProt accession exists, and 395 ESMFold predictions folded from sequence alone with a mean confidence of 91.0 pLDDT. 547 have a measured catalytic triad, and one deposit was caught as a catalytic knockout, an inactivated construct that carries the right name and the wrong chemistry.", "Q: Can you send an email for me?\nA: I can't send emails or take actions outside this conversation. If you'd like to reach Marc Deller directly, his email is marc@marcdeller.com.", "Q: What is CODSWALLOP built with?\nA: CODSWALLOP runs on Python 3.12, Flask 3.1 and gunicorn 26 with a SQLite WAL cache holding raw API responses and assembled families on a one-week TTL, a vanilla JavaScript front end with hand-rolled SVG, Tabulator 6.3.1 for tables, self-hosted Archivo, Inter and JetBrains Mono type, and nginx with certbot on the same droplet pattern as AlphaFraud.", "Q: Why did Marc take an AWS course?\nA: In his post at https://marcdeller.com/serverless-architectures-on-aws/, Marc Deller frames the AWS serverless course as practical preparation for building and deploying scientific web apps and online research resources rather than as a purely technical credential.", "Q: Which OSM helices are kinked, and what causes the kinks?\nA: Marc Deller found that both helix A and helix C of OSM are kinked. Helix C has a slight kink towards its N-terminus between Gln 112 and Pro 116, caused by disruption of the classic \u03b1-helical hydrogen-bonding pattern with substitute hydrogen bonds to two tightly bound water molecules. Helix A has a much more pronounced kink because Gln 25 and Leu 30 adopt a hydrogen-bonded turn conformation and four tightly bound waters disrupt the helical hydrogen bonding; residues Thr 27 to Ile 37 of helix A form a 3-10 helix. The helix A kink allows closer packing of the A:D and B:C helical pairs, giving a tighter solvent-inaccessible core.", "Q: What does Marc say the N233K mutation actually does in FAST-PETase?\nA: Marc explains that, contrary to the original salt-bridge proposal, the N233K mutation enhances catalysis by altering the chain folding around Glu204, which increases the basicity of Asp206 in the catalytic triad, strengthens its interaction with protonated His237, and lowers the activation barrier by approximately 4.4 kcal/mol.", "Q: Did Marc Deller build a crystal harvesting robot himself?\nA: Marc Deller was directly involved in the UMR workcell, which the review states was engineered by Square One System Design with current design and development input from the authors of the article, and which was undergoing field testing and further development at the Joint Center for Structural Genomics at The Scripps Research Institute.", "Q: How does Marc Deller describe himself on LinkedIn?\nA: On LinkedIn, Marc Deller describes himself as CSO and Co-Founder of Elora Therapeutics, pioneering a first-in-class enzyme therapy to degrade synthetic microplastics that accumulate in blood and tissues and are linked to chronic inflammation and metabolic disease. He highlights 20+ years leading innovation in structural biology and drug discovery, combining deep protein-science expertise with AI-driven automation.", "Q: What performance does Marc report for LatentX?\nA: Marc reports binding affinities in the picomolar range for mini-binders and near-perfect hit rates for macrocycles, with far fewer candidates needed than previous approaches. He attributes the gain to LatentX co-sampling protein sequence and 3D structure simultaneously rather than treating them separately as earlier models did.", "Q: Who were the authors of the oncostatin M crystal structure paper, and what was Marc's role?\nA: The paper \"Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\" was authored by Marc C. Deller, Keith R. Hudson, Shinji Ikemizu, Jer\u00f3nimo Bravo, E. Yvonne Jones and John K. Heath. Marc Deller was first author and contributed equally with Keith Hudson; Marc carried out the structural biology side in Yvonne Jones's Cancer Research Campaign Receptor Structure Group at the Wellcome Trust Centre for Human Genetics in Oxford, while the Hudson/Heath group in Birmingham provided the mutagenesis and functional assays. Yvonne Jones was the corresponding author.", "Q: Who is ForgeFold aimed at?\nA: Marc writes that it is for structure-guided teams who want to lower the friction of repeated co-folding runs, and to give less command-line-comfortable colleagues a clean, guided path from input preparation to result triage, instead of hopping between a text editor, a shell and a spreadsheet. The repo is bellcheddar/FoldForge and the post is at https://marcdeller.com/a-single-file-browser-gui-for-the-whole-boltz-2-workflow-build-run-review/.", "Q: What is in DeepCovalent's pipeline?\nA: DeepCovalent's therapeutic pipeline consists of in-silico validated novel compounds generated by OS3D, led by small-molecule GLP-1R candidates, with an additional eleven backup compounds identified for potential pre-clinical development.", "Q: Who funded the MtPMO3* study?\nA: The work was funded by NSF grant 1565770 and a grant from the Energy Biosciences Institute to Michael Marletta, with the EPR spectroscopy supported by NIH grant GM111025 to R. David Britt and Daniel Suess supported by an NIH NIGMS fellowship (F32GM111025). The acknowledgements also thank J. Hangasky and other Marletta lab members, J. Klinman and A. Murphy for Oxygraph instruments, I. Wilson for SSRL beam time, and members of the JCSG for crystal screening.", "Q: What examples of LOCKTAC action does Marc give?\nA: Marc cites amplifying examples such as stabilising the interaction between SMN2 mRNA and its nuclear partner, and inhibitory examples such as motor proteins whose function is blocked by persistently locked engagement. He notes the same locking mechanism can either extend a biological signal or shut a pathway down depending on context.", "Q: Why does KRAS G12D matter clinically?\nA: G12D is the most prevalent KRAS mutation, and addressing it would significantly expand the patient population beyond the G12C-mutant cancers treated by sotorasib and adagrasib. It occurs in roughly 37% of pancreatic tumours and 12% of colorectal tumours. More broadly, an estimated 20% of all human tumours are driven by KRAS mutations, most often at Gly12, Gly13 or Gln61.", "Q: Where is Marc's post on human skills training?\nA: The post \"2025 \u2013 Investing in Human Skills in the Age of AI\" is at https://marcdeller.com/investing-in-human-skills-in-the-age-of-ai/.", "Q: What methods were used for Marc Deller's PDB structures?\nA: Every one of the 78 entries in Marc Deller's list was determined by X-ray diffraction; none is a cryo-EM or NMR structure. That reflects a career centred on protein crystallography, although his broader expertise also covers cryo-EM.", "Q: What is PDB entry 2Q8U?\nA: 2Q8U is the crystal structure of Mre11 from Thermotoga maritima MSB8 (TM1635) at 2.20 \u00c5 in space group P 1 21 1, with 672 polymer residues and 77.71 kDa, released on 26 June 2007 and keyworded HYDROLASE. It was deposited by the JCSG and is cited in the Journal of Molecular Biology 2010 paper on the first eubacterial Mre11 nuclease.", "Q: What was Marc Deller's role on the HCV E2\u2013AP33 paper?\nA: Marc Deller was a middle co-author, listed fifth of eight authors, on the 2012 Journal of Virology HCV E2\u2013AP33 structure paper. His affiliation on the paper is the Joint Center for Structural Genomics in La Jolla, California, and he contributed the crystallography side of the work \u2014 crystallisation screening, crystal harvesting and X-ray data collection carried out for the Wilson and Law laboratories at The Scripps Research Institute.", "Q: What are the 6C's in Marc Deller's personal philosophy?\nA: The About page sets out a personal philosophy called the 6C's: Curiosity, Courage, Creativity, Communication, Collaboration, and Compassion. Each is linked to a separate post on marcdeller.com that expands on it.", "Q: What is the Macromolecular Structure Knowledge Center?\nA: The Macromolecular Structure Knowledge Center is a structural biology service center at Stanford University that Marc Deller designed, implemented and directed from 2015 to 2018. It provided Stanford researchers with access to high-throughput protein characterization platforms and to the SSRL/SLAC synchrotron. Stanford Medicine News profiled it in April 2016 as \"Your one-stop shop for producing, crystallizing biomolecules\".", "Q: What is BT2081?\nA: BT2081 from Bacteroides thetaiotaomicron, GenBank accession NP_810994.1, is a member of a novel protein family of over 160 members found mostly in the different classes of Bacteroidetes. Its crystal structure at 2.05 \u00c5, co-authored by Marc Deller, was the first from that family and was published in Acta Crystallographica Section F in 2010.", "Q: What does Marc conclude about AI-designed peptides?\nA: Marc concludes that AI is not only accelerating timelines but enabling peptides with optimized properties that would be difficult to reach by traditional methods, unlocking the full potential of a drug class prized for high specificity and low toxicity. You can read it at https://marcdeller.com/beyond-trial-and-error-ai-driven-design-of-peptide-therapeutics/.", "Q: What time does Marc get up in the morning?\nA: I don't have anything on his routine. His output suggests it's early, but that's a guess and I'd rather flag it as one.", "Q: How were the antibody\u2013peptide crystals obtained?\nA: The Fab of antibody clone 55 was expressed in Expi293F cells, purified on CaptureSelect CH1-XL affinity matrix and eluted at pH 4.0. For the complex, Fab was mixed with the acetylated 21-mer mutCALR peptide at 1:1.5 molar excess and concentrated to 35 mg/mL; sparse-matrix screens were set up on an NT8 crystallisation robot in 200 nl + 200 nl drops. Initial hits came from condition C7 of the JCSG Top96 Screen after three days at 4 \u00b0C, and the optimised condition was 0.1 M Bicine pH 7.94 with 19% w/v PEG 6000.", "Q: What is the 2016 Immunity paper that Marc Deller co-authored about?\nA: The 2016 Immunity paper \"Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor,\" co-authored by Marc Deller, traced how a VRC01-like broadly neutralizing antibody lineage emerged in an HIV-1-infected Chinese long-term non-progressor. It showed that N-linked glycans on the HIV-1 envelope protein gp120 \u2014 specifically the N276 glycan and glycans on the V5 loop \u2014 stalled the antibody lineage from ever achieving broad neutralization.", "Q: How many publications does Marc have?\nA: Marc Deller's communication post cites 60+ cross-functional publications and links to the publications page of marcdeller.com. He says that across those papers he focuses on telling mechanistic stories that guide the next experiment, design cycle or decision rather than just reporting data.", "Q: Tell me about ButtFold.\nA: ButtFold, at buttfold.mdeller.com, takes a named protein, starts it as a self-avoiding random coil and folds it with a C-alpha structure-based G\u014d model until it reaches its native structure, turning the trajectory into music in the browser. It is the web version of Marc's PhoneFold app, and its `Sonifier.js` is held to note-for-note agreement with the shipped Swift across 15,536 notes.", "Q: Who funded the PKS extended conformation study?\nA: The research was supported by NIH grants GM087934 to Chaitan Khosla, P41CA196276 and GM104659 to Charles Craik, and GM022172 to David Cane, with SSRL supported by the DOE Office of Basic Energy Sciences under contract DE-AC02-76SF00515 and the SSRL Structural Molecular Biology Program by the DOE and NIH (including P41GM103393).", "Q: What does Marc's post \"Precision Timing of Cytokine Signals through Facilitated Dissociation\" cover?\nA: Published on 13 October 2025, this post by Marc Deller covers a Nature paper by Adam J. Broerman and colleagues describing a protein-design platform that makes engineered cytokine receptors and mimics dissociate on command. Marc explains that by designing strained excited conformational states, the team converted hours-long interleukin-2 signals into precise seconds-scale pulses.", "Q: Where can I read Marc's Mre11 post?\nA: Marc Deller's post is \"Crystal Structure of MRE11 from Thermotoga maritima MSB8 (TM1635) at 2.20 \u00c5 Resolution\", in the structures category of his blog. You can read it at https://marcdeller.com/crystal-structure-of-mre11-from-thermotoga-maritima-msb8-tm1635-at-2-20-a-resolution/.", "Q: What role does machine learning play in PETase design, according to Marc?\nA: Marc writes that computational and machine learning techniques now guide mutation selection, producing enzyme variants engineered for maximal depolymerization of PET even under challenging biological conditions, and paving the way for biotherapeutics that clear microplastics from human blood and tissues.", "Q: What is Marc Deller's post on FDA-approved enzyme therapies about?\nA: In \"Unlocking the Power of Protein Engineering: FDA-Approved Enzyme Therapies Transform Rare Disease Treatment\" (7 November 2025), Marc Deller surveys the approved enzyme replacement therapies from 1978 onward, showing how glycoengineering, PEGylation, Fc fusion and targeted delivery turned crude extracts into precision biologics. He uses this as the regulatory and technical precedent for enzyme therapeutics generally.", "Q: Where is Marc's post about the generative AI communication course?\nA: Marc Deller's post \"2025 \u2013 Amplify Your Communication Skills with Generative AI\" is published at https://marcdeller.com/amplify-your-communication-skills-with-generative-ai/ and links to the LinkedIn Learning certificate.", "Q: Which year dominates Marc Deller's ORCID record?\nA: 2010 is by far the largest year on Marc Deller's ORCID record, with more than twenty Acta Crystallographica Section F structure papers from the Joint Center for Structural Genomics, covering Pfam family first representatives and Bacteroides, Klebsiella, Jannaschia, Haemophilus and Bacillus proteins.", "Q: What is PDB entry 5CD5?\nA: 5CD5 is the immature VRC01-class antibody DRVIA7 from a Chinese donor bound to a clade A/E HIV-1 gp120 core, at 3.396 \u00c5 in space group P 41 21 2 with 783 polymer residues, 89.98 kDa and NAG bound, released on 6 April 2016. It shares its Immunity 2016 primary citation with 5CD3.", "Q: What is Marc's favourite quote?\nA: Marc likes Steve Jobs' line \"Stay hungry, stay foolish\", and his own saying that \"if science were successful on the first attempt, it would be called 'Search'; that's why we call it 'Research'.\"", "Q: What kinds of laboratories has Marc built?\nA: Marc Deller writes that he has built 2+ fully integrated laboratories combining protein production, X-ray crystallography and cryo-EM capability. He describes his role as connecting the threads of such projects into a coherent vision and then executing against it step by step.", "Q: How was the membrane model built?\nA: A trimeric MPER-transmembrane model of gp41 was built using PDB 2MOM as a template, with the MPER orientation taken from the new crystal structures. The crystallographic 06:0 PG fragment was extended to a full DPPG lipid, and the CHARMM force field membrane builder generated a heterogeneous bilayer in a 157.7 \u00c5 square box containing 400 lipids in the upper leaflet and 434 in the lower, using the known HIV-1 membrane lipid composition. Potassium counter-ions were placed by Monte Carlo and the model was equilibrated at 303 K.", "Q: What is the significance of the G protein selectivity work in Marc's view?\nA: Marc presents it as a proof of principle for rationally engineering signalling bias, noting that minor chemical modifications to the SBI-553 scaffold inverted the selectivity profile to favour Gq over GoA. He argues that because the intracellular transducer pocket is conserved, this gives a reproducible blueprint for designing biased allosteric modulators across the GPCR superfamily and decoupling therapeutic efficacy from side effects. The post is at https://marcdeller.com/tuning-the-signal-rational-design-of-g-protein-selectivity/.", "Q: What does Marc Deller say motivates Elora Therapeutics, according to LinkedIn?\nA: On LinkedIn, Marc Deller frames Elora Therapeutics around the question \"What if a revolutionary enzyme therapeutic could remove harmful microplastics from the human body\u2014unlocking new frontiers in environmental health?\" He positions the company's enzyme therapy as addressing microplastic accumulation in blood and tissues linked to chronic inflammation and metabolic disease affecting millions worldwide.", "Q: What's the best programming language?\nA: I don't have a general opinion on that, but I can tell you Marc Deller does a lot of his vibe-coding and data work in Python, R/Shiny and JavaScript, and fine-tunes local models with MLX-LM on his Mac. Want to hear more about one of his projects?", "Q: What does chatRDKit show on launch?\nA: A welcome banner with five numbered example prompts: enter a SMILES string to load a molecule, ask for the molecular weight, LogP and TPSA of caffeine, draw ibuprofen and save it as an image, generate and optimise a 3D structure of ethanol, and compare aspirin with paracetamol using Morgan fingerprints. Typing a number runs that example.", "Q: How does Marc want visitors to think about his chat bot?\nA: Marc Deller writes on the Chat page that visitors should think of it less as a chatbot and more as a conversation with him. The framing is a live alternative to reading a static resume.", "Q: What range of roles does Marc Deller have across his publications?\nA: Marc Deller's roles range from first and corresponding author to supporting co-author. He is first author on the oncostatin M structure, the TrkA gating ring assembly, the crystal harvesting review and the protein stability review, and a contributing co-author on the Joint Center for Structural Genomics papers of 2007 to 2012, the HIV-1 Env and hepatitis C virus papers from Ian Wilson's laboratory, his Stanford collaborations, and the Incyte medicinal-chemistry papers.", "Q: Does Marc have an AWS certification?\nA: Yes, his My Learning page cites an AWS certification showcasing expertise in modern cloud architectures for data processing, and lists a 2025 certificate in \"Serverless Architectures on AWS\".", "Q: Which heavy atom compounds did Marc screen, and what happened?\nA: Marc Deller screened eleven reagents including trimethyllead acetate, lead acetate, potassium mercuric iodide, ethylmercury phosphate, uranyl acetate, uranyl nitrate, cis-platin, Baker's mercurial, TAMM, sodium gold chloride and dipotassium platinum nitrate. Uranyl salts, PMI, Baker's and TAMM destroyed the crystals instantly, platinum compounds abolished diffraction, and lead salts did not bind at all. Only ethylmercury phosphate (EMP) gave a usable derivative.", "Q: Who is Elora partnered with?\nA: The pitch deck states that Elora Therapeutics is partnered with the Austin Technology Incubator (ATI). It also references a University of Texas technology validation agreement for planned in vitro testing.", "Q: What does Marc say Proteina can do?\nA: Marc reports that Proteina, a large-scale flow-based model, uses flow matching to generate protein backbones and can be conditioned on hierarchical labels, so a designer can guide generation broadly (\"create a mostly alpha-helical protein\") or very specifically (\"create a protein with this exact fold\"). He notes it achieves state-of-the-art de novo design performance, producing diverse and designable proteins up to 800 residues in length.", "Q: What did the crystal structure model validation chapter emphasize?\nA: The 2017 book chapter \"Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands\" by Edwin Pozharski, Marc C. Deller and Bernhard Rupp emphasizes the need for high-quality protein structures in the age of predictive AI. It highlights criteria such as electron density fit, stereochemistry and binding plausibility and introduces tools to help researchers validate protein\u2013ligand models, a challenge given the scarce experimental data for ligands.", "Q: Which JCSG Top96 conditions use MPD as the main precipitant?\nA: The MPD-based conditions in the JCSG Top96 include A03\u2013A06 (10% to 65% MPD from the Hampton MPD Grid Screen), B08 (5% PEG 8000 with 40% MPD), and the organic/salt conditions B10\u2013B12, which pair 30\u201335% MPD with 0.2 M sodium chloride, ammonium acetate or lithium sulfate.", "Q: Does Marc have a patent on covalent WRN inhibitors?\nA: Yes, his Patents page lists the 2025 patent \"Imidazo[1,2-a]pyrimidine-5(1H)-ones having a covalent warhead at the 3-position as WRN inhibitors\". It sits alongside the non-covalent imidazo[1,2-a]pyrimidinone WRN inhibitor patent from the same year.", "Q: How does JCSG Top96 compare with the earlier JCSG+ screen?\nA: According to the 2013 poster, the JCSG Top96 screen performs about 4-fold better than JCSG+ in terms of PDB depositions, with Top96 conditions accounting for 631 PDB entries versus 168 for JCSG+. Only 16 conditions are shared between the two screens.", "Q: What was Marc Deller's role on the povorcitinib paper?\nA: Marc C. Deller is listed as one of the Incyte Research Institute co-authors of the 2026 povorcitinib paper, contributing the structural biology and protein crystallography support for the programme. He worked alongside Leslie B. Epling and Gustavo Fenalti, the Incyte crystallography colleagues named on the same paper and on the associated PDB depositions, providing the X-ray structures of JAK1 and JAK2 bound to povorcitinib that underpinned the paper's selectivity explanation.", "Q: Where does CODSWALLOP get its data?\nA: CODSWALLOP reads the RCSB Search API v2 for family resolution, the RCSB Data API GraphQL endpoint for batched entry, polymer entity, assembly and chemical component metadata, UniProt REST for the canonical sequence, protein name, organism and gene names, and the InterPro API for Pfam and InterPro accession names. All are public and need no key, and Marc asks that they, not his tool, be cited when the data does the work.", "Q: Which residues decide JAK1 selectivity?\nA: Marc Deller highlights JAK1 His885 pointing into solvent versus JAK2 Asn859 and Phe860 closing the pocket against the inhibitor's fluorinated tail, describing it as a two-residue difference that, read correctly, becomes a drug.", "Q: Which papers does Marc cite in his cryo-EM-without-purification post, and where is it published?\nA: The post is built around Powell and Evans, \"Cryo-EM protein structure without purification\", bioRxiv 2025, and also cites Wu and Lander on single-particle methodology in Biophysical Journal 2020, Kuruma and Ueda on the PURE system in Nature Protocols 2015, Punjani and colleagues on cryoSPARC in Nature Methods 2017, and Su and colleagues on the Build and Retrieve method in Nature Methods 2021. You can read it at https://marcdeller.com/revolutionizing-protein-structure-analysis-cryo-em-without-purification-accelerates-the-future-of-structural-biology/.", "Q: Where was the crystallography for the HCV E2\u2013AP33 paper carried out?\nA: The crystallographic work was done at The Scripps Research Institute with the Joint Center for Structural Genomics, where Marc Deller worked. The acknowledgments thank Henry Tien and Thomas Clayton for help setting up initial crystal screens on the CrystalMation robot, and note that portions of the research were carried out at the Stanford Synchrotron Radiation Lightsource, a Directorate of the SLAC National Accelerator Laboratory.", "Q: What's chatMCD?\nA: I'm chatMCD \u2014 a small language model that Marc Deller fine-tuned himself using MLX-LM, trained on his own papers, patents, r\u00e9sum\u00e9, blog posts and notes. I answer questions about Marc, his science and his companies, speaking about him in the third person.", "Q: Roughly how much of Marc Deller's citation total is recent?\nA: Of Marc Deller's 3,967 all-time citations on Google Scholar, 1,343 have accrued since 2021, which is a little over a third. His h-index over that same recent window is 18 against 25 all time, and his i10-index 26 against 45.", "Q: Did Marc Deller achieve his goal of crystallisation-grade LIFR?\nA: No. The stated goal was 10 mg or more of soluble LIFR-FN3 at 90 % or better purity for crystallography, and this was not achieved because of several technical problems, chiefly low yields, aggregation, glycosylation heterogeneity and loss of activity on deglycosylation. However, Marc established critical steps toward that goal: a working 293T expression route, a Protein-A/3C purification scheme, and successful PNGase F deglycosylation.", "Q: What is the CALR antibody patent about?\nA: The 2022 patent \"Anti-mutant Calreticulin (CALR) Antibodies and uses thereof\" (WO2023107994A1) covers antibodies against mutant calreticulin, an antibody program Marc Deller supported at Incyte. It is published at https://patents.google.com/patent/WO2023107994A1/en.", "Q: How does temperature come into Marc's stability work?\nA: Thermal denaturation assays such as differential scanning fluorimetry yield melting temperatures and enthalpic profiles that guide buffer and ligand selection, while room-temperature and multi-temperature crystallography reveal fluctuations that cryogenic snapshots can mask, as shown by K-RAS conformers uncovered only at physiological temperatures.", "Q: What success rate does Marc quote for oral protein drug candidates?\nA: Marc reports that fewer than 2% of protein drug candidates reach the clinic in oral form despite hundreds of attempts, and that achieving bioavailability above 1% remains a benchmark few candidates meet.", "Q: What are Marc Deller's most cited papers?\nA: Marc Deller's most cited papers are the 2013 Science paper \"Crystal structure of a soluble cleaved HIV-1 envelope trimer\" (about 979 citations), the 2013 Nature Structural & Molecular Biology paper \"Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\" (about 420), \"Protein stability: a crystallographer's perspective\" (2016, about 370), \"The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\" (ACS Chemical Biology 2017, about 130), and \"Models of protein\u2013ligand crystal structures: trust, but verify\" (2015, about 113).", "Q: What is TmMre11 and why was its structure important?\nA: TmMre11 is the Mre11 endo/exonuclease from Thermotoga maritima, encoded by TM1635 and originally annotated only as a putative nuclease. The 2010 Journal of Molecular Biology paper co-authored by Marc Deller reported its crystal structure as the first eubacterial Mre11 structure, at a time when X-ray information was available only for the Pyrococcus furiosus enzyme.", "Q: What kind of people does Marc Deller want to connect with on LinkedIn?\nA: Marc Deller's LinkedIn profile welcomes connections with biotech innovators, scientific collaborators, and anyone passionate about enzyme therapeutics or emerging health challenges \u2014 and he adds that bonus points go to anyone who wants to talk Nottingham Forest football.", "Q: What does the Figure 3 Venn diagram in the poster show?\nA: Figure 3 of the JCSG Top96 poster is a proportional Venn diagram showing the overlap of PEG (green), salt (blue) and organic (red) precipitants that defines the seven precipitant classes of the screen, with the numbers of PDB depositions for each class shown in white.", "Q: What was Marc Deller's role on the ligand validation chapter?\nA: Marc Deller was the second author, with Edwin Pozharski as first author and Bernhard Rupp as senior author. The chapter builds on Marc Deller's earlier work with Bernhard Rupp on protein\u2013ligand models, including their 2015 paper \"Models of protein\u2013ligand crystal structures: trust, but verify\".", "Q: What makes PETase structurally distinctive, in Marc's telling?\nA: Marc writes that PETase uses a canonical serine-histidine-aspartate triad and a two-step acylation-deacylation mechanism, but that its substrate-binding cleft is wider and more open than that of related cutinases, allowing the bulky aromatic PET chain to reach the active site.", "Q: How structurally similar are TM1088A and TM1088B to each other and to other RCK proteins?\nA: Although TM1088A and TM1088B share only 23% sequence identity, they superimpose with an RMSD of 1.2 \u00c5 over 122 C\u03b1 atoms. DALI searches and superpositions showed the closest structural homolog is the N-terminal TrkA domain of Vibrio parahaemolyticus (PDB 4J9V), with RMSDs of 1.2 \u00c5 over 71 C\u03b1 atoms for TM1088A and 1.0 \u00c5 over 72 C\u03b1 atoms for TM1088B; TM1088A also aligns well with M. jannaschii TrkA (1LSS, 1.1 \u00c5), MthK (1LNQ, 1.6 \u00c5) and BK (3U6N, 1.9 \u00c5).", "Q: What does Marc say the AI communication credentials validate?\nA: Marc Deller writes that together the three credentials validate both technical prowess and the soft skills needed to lead and adapt in AI-driven environments, including cultivating resilience, strengthening interpersonal connections in hybrid work settings, and demonstrating commitment to lifelong learning. The post is at https://marcdeller.com/embracing-ai-enabled-communication-and-adaptability/.", "Q: Who were the authors of the 2013 ACA poster on the JCSG Top96 screen?\nA: The 2013 ACA poster \"JCSG Top96: A Rationally Designed Crystallization Screen Based on over 1000 Deposited Protein Structures\" was authored by Marc C. Deller, Tom Clayton, Marc A. Elsliger and Ian A. Wilson of the Joint Center for Structural Genomics at The Scripps Research Institute in La Jolla, California. Marc Deller was the first and presenting author.", "Q: What was Marc Deller's 2014 ACA poster about?\nA: Marc Deller's 2014 ACA poster, \"Late-stage Crystal Failures and the Protein Crystal Harvesting Bottleneck\", described the crystal harvesting step as the last major manual bottleneck in high-throughput protein crystallography, and presented the Universal Micromanipulation Robot (UMR), an automated harvesting and cryocooling platform being developed and tested at the Joint Center for Structural Genomics (JCSG).", "Q: How is a letter stack built in ALPHABETTI?\nA: At each position ESM-2 gives a probability for all twenty canonical amino acids; ALPHABETTI computes the Shannon entropy H_i, the information content R_i = log2(20) - H_i with a maximum of 4.322 bits, and a per-letter height of p_a times R_i. The letters are stacked tallest-first along the residue's own side-chain direction, so the total height of a stack is that position's information content, a property checked by a unit test against hand-calculated values.", "Q: What response rate has been reported for WRN inhibitors in the clinic?\nA: Marc Deller's post notes that clinical development efforts have shown promising efficacy signals, with compounds such as HRO761 achieving overall response rates of 14% in heavily pretreated MSI-high populations.", "Q: Why is oral protein delivery relevant to microplastic enzyme therapy?\nA: Marc files the post under microplastics because a plastic-degrading enzyme delivered by mouth would face exactly these barriers \u2014 proteolysis, the mucus layer and epithelial tight junctions \u2014 and the same permeation-enhancer, targeting and device strategies would be needed. It is at https://marcdeller.com/overcoming-biological-barriers-the-evolution-of-oral-protein-and-peptide-drug-delivery/.", "Q: Does Marc think AI prediction replaces experimental structural biology?\nA: No. In his multi-step LLM prompt post Marc is explicit that experimental validation still reigns supreme, naming X-ray crystallography for static high-resolution maps, cryo-electron microscopy for large complexes and dynamics, and NMR spectroscopy for flexible and small proteins as the gold standards. His view is that good computational models accelerate that experimental work by highlighting likely conformations, rather than replacing it.", "Q: What data-analysis and AI tools does Marc Deller use?\nA: Marc Deller works with R/Shiny, Spotfire, GraphPad Prism, MATLAB, Python, RDKit and pandas for data analysis; Nanome for molecular visualization; LLaMA, Perplexity and ChatGPT for generative AI; AWS, Globus and SLURM for cloud and HPC computing; and JIRA, Confluence, Smartsheet, Canva, Adobe and the Microsoft Office suite for project and communication work.", "Q: How does Marc describe the role of prediction in structural biology?\nA: The Structural Biology page says prediction leverages machine learning to model complex protein folds in hours rather than months, letting teams rapidly prioritise tractable targets and identify cryptic binding pockets by integrating predictions with experimental validation. Prediction is framed as a triage aid rather than a final answer.", "Q: Which entries are the FmdE-like structures?\nA: Three entries are cited in the Acta Crystallographica Section F 2010 paper on Pfam PF02663 (FmdE): 2GLZ, an FmdE-like protein Dhaf_2992 from Desulfitobacterium hafniense DCB-2 at 1.45 \u00c5 with nickel and zinc bound; 2GVI, a putative formylmethanofuran dehydrogenase subunit E TA1109 from Thermoplasma acidophilum at 1.87 \u00c5 in space group I 2 2 2; and 3D00, a tungsten FmdE-like protein Syn_00638 from Syntrophus aciditrophicus at 1.90 \u00c5 with zinc bound.", "Q: Why does the DEBS module structure matter for biotechnology?\nA: Marc Deller's post explains that the structure unlocks critical insights for engineering hybrid polyketide synthases to produce novel antibiotics and other natural products. Understanding which module conformation is catalytically competent is central to that engineering effort.", "Q: How does AlphaFraud actually work end to end?\nA: Its weekly pipeline searches RCSB for human (taxonomy 9606) protein entities deposited after AlphaFold's 2018-04-30 training cutoff, maps each to its UniProt accession and AlphaFold DB model, downloads both, computes the full metric suite, scores sequence novelty against pre-cutoff chains, computes a FRAUD score as the mean of pLDDT-weighted per-residue deviation clamped at 15 \u00c5, renders a deviation ribbon, and stores it all in SQLite.", "Q: What is the T117v2 scaffold?\nA: T117v2 is an engineered epitope-scaffold designed to present the 10E8 core helical MPER epitope on a stable protein framework in the 10E8-bound conformation. It is a variant of the earlier T117 scaffold \u2014 which was originally built for the antibody 4E10 \u2014 modified by M125K and D126A mutations so that 10E8 could bind, and it binds mature 10E8 with 29 pM affinity.", "Q: What engineering modifications to the enzymes are described?\nA: The patent describes engineering the enzymes through site-directed mutagenesis, thermostabilization mutations, enhanced redox potentials, and improved substrate specificity. These modifications allow PETase, MHETase, CotA-Laccase, and HRP to function in complex physiological environments, and also include fusion enzymes such as PETase-MHETase fusions and oxidative peroxidase hybrids.", "Q: What defenses does Marc suggest against adversarial reasoning triggers?\nA: Marc points to model randomization to disrupt adversarial query consistency, semantic filtering to detect irrelevant prefixes or suffixes, and adversarial fine-tuning with CatAttack triggers to immunize models. He argues security evaluations must include threat models beyond input-specific attacks. The post is at https://marcdeller.com/cats-confuse-reasoning-llm-query-agnostic-adversarial-triggers-for-reasoning-models/.", "Q: Has Marc ever taken a sabbatical?\nA: Not something I know about. I can walk you through the roles he has held, but the gaps between them aren't recorded in what I know.", "Q: Which entries carry no published primary citation?\nA: The two TrkA gating ring entries, 3JXO and 3L4B, list their primary citation as \"To be Published\" with no DOI in the RCSB record. The science was later reported by Marc Deller and colleagues in PLoS ONE in 2015.", "Q: Can anyone buy the Top96 screen?\nA: Yes. Marc's post links to the screen as commercially available from Rigaku Reagents as the JCSG Top96 block, alongside a PDF of the 96 conditions and the ACA 2013 poster.", "Q: What does CODSWALLOP stand for?\nA: CODSWALLOP stands for Comparative Overview of Domains, Structures, Whole-families, Alignments, Ligands, Lineages, Organisms and PDB entries. It is one of Marc Deller's apps, live at codswallop.mdeller.com with source at github.com/bellcheddar/CODSWALLOP.", "Q: Why does the chapter say ligand quality is context-sensitive?\nA: Because the standard a structure must meet depends on what it will be used for. As Pozharski, Deller and Rupp put it, not every protein\u2013ligand structure is fit for structure-based lead discovery, and not every missing piece of a ligand is reason for panic.", "Q: What is the most recent work on Marc Deller's ORCID record?\nA: The most recent works on Marc Deller's ORCID record are from 2017: \"The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\" in ACS Chemical Biology (DOI 10.1021/acschembio.7b00016) and the PLoS Pathogens paper on lipid interactions and the angle of approach of broadly neutralizing antibody 10E8 to the HIV-1 viral membrane (DOI 10.1371/journal.ppat.1006212).", "Q: What precipitant classes appear in the JCSG Top96 screen?\nA: The JCSG Top96 formulation sheet assigns each condition to one of seven precipitant classes: Organic, Organic/PEG/Salt, Organic/PEG, Organic/Salt, PEG, Salt and Salt/PEG. Salt/PEG combinations are the largest class, occupying most of rows E to H.", "Q: How long is the HCV E2\u2013AP33 paper, and what kind of article is it?\nA: It is a short-format report of four pages in the Journal of Virology \u2014 pages 13085 to 13088 of volume 86, number 23, December 2012. It was received on 27 July 2012, accepted on 7 September 2012, and published ahead of print on 12 September 2012, with the detailed methods provided as supplemental material.", "Q: How does Marc summarise the ChemSage hybrid in one line?\nA: \"RAG keeps it truthful today; the fine-tune makes it consistent tomorrow; RDKit makes it correct always.\" The design separates three concerns that are usually conflated: volatile knowledge lives in retrieval, stable behaviour lives in the weights, and deterministic chemical facts are computed rather than guessed.", "Q: Which enzymes does the patent cover?\nA: The patent covers four enzyme systems: polyethylene terephthalate hydrolase (PETase) including wild-type, engineered variants, fusion enzymes, and thermostabilized mutants; mono(2-hydroxyethyl) terephthalate hydrolase (MHETase) in wild-type, engineered, and fusion forms; bacterial-origin CotA-Laccase engineered for oxidative activity; and engineered Horseradish Peroxidase (HRP), including recombinant HRP expressed in non-fungal systems and oxidative peroxidase hybrids. The PETase and MHETase enzymes handle PET, while the oxidative CotA-Laccase and HRP tackle polyolefin plastics.", "Q: What resolution did the OSM crystals diffract to?\nA: The P212121 form of \u0394188_OSM showed ordered Bragg diffraction to 2.2 \u00c5 at the synchrotron, while the I222 form diffracted to 2.5 \u00c5. The MAD data on the mercury derivative were collected to 2.6 \u00c5.", "Q: Did Marc mentor students at Stanford?\nA: Yes, his Projects page notes that he mentored a Stanford-led student program on novel antibiotics for resistant bacteria, and links to a 2017 Stanford news story about the undergraduates winning a prize for their work combating antibacterial resistance.", "Q: How does 10E8 recognise both peptide and lipid?\nA: Its light chain engages specific phospholipid head groups on the viral membrane through a constellation of basic and polar residues, particularly phosphatidylglycerol and phosphatidic acid, while its heavy chain binds the MPER epitope, creating a composite epitope the virus cannot easily escape by mutation alone.", "Q: Why does CNTF bind gp130 so weakly, according to Marc's paper?\nA: The oncostatin M paper by Marc Deller and colleagues suggested that CNTF binds gp130 with much lower affinity than LIF or OSM because the residue equivalent to OSM Gly120 in CNTF is a tyrosine, which would be expected to occlude the hydrophobic dimple needed to accommodate gp130 Phe169. CNTF therefore requires the accessory CNTF receptor to compensate for this loss of affinity by adding extra receptor\u2013receptor contacts to the complex.", "Q: What delivery systems does Marc propose to keep PETase circulating in the body?\nA: Marc points to advanced delivery systems such as encapsulating the enzyme in erythrocytes (red blood cells) or using surface-display techniques, both of which increase PETase circulation time and tissue targeting.", "Q: What fun questions does Marc offer as conversation starters?\nA: The Chat page lists four cut-and-paste questions: which dream team of famous scientists, living or historical, he would assemble to tackle big problems; what an ultimate automated structural biology research platform would include with unlimited resources; the wildest or most futuristic application of protein-ligand crystal structures for drug discovery or diagnostics; and what a plastic-eating enzyme like PETase would say about cleaning up the planet and what mutation it would request for its next upgrade.", "Q: How can WordPressID be customised?\nA: Add fields via `_fields=id,title,slug,link,date` and extend the jq object; replace the truncation block with plain `.title` for untruncated screen titles; append `| less -SX` for wide tables; swap `@csv` for `@tsv` for a tab-delimited file; and add custom post types to the `for t in posts pages` list. Only published content is listed unless you authenticate. It is MIT licensed; the repo is bellcheddar/wordpressID.", "Q: How did Marc calculate protein concentration from absorbance in his thesis?\nA: Marc Deller used the formula [Protein] (mg/ml) = A280 \u00d7 Mr / ((1331 \u00d7 Y) + (5650 \u00d7 W)), where Y and W are the numbers of tyrosine and tryptophan residues and Mr is the molecular weight, with measurements made in a 1 ml cell on a Beckman DU-600 spectrophotometer. He estimated plasmid DNA concentration using A260 of 1.0 = 50 \u03bcg/ml.", "Q: Explain in simple terms how Marc worked out the OSM structure.\nA: Marc Deller grew tiny crystals of the cytokine Oncostatin M, fired synchrotron X-rays through them and recorded the diffraction spots. Because diffraction alone does not give the \"phase\" information needed to compute an image, he soaked mercury atoms into the crystals and measured how they scattered X-rays at three different wavelengths (the MAD method). That let him calculate an electron density map, into which he built and refined an atomic model at 2.2 \u00c5 resolution, then used that model to solve a second crystal form by molecular replacement.", "Q: What advantage of X-ray crystallography do Marc Deller and Bernhard Rupp emphasise in \"trust, but verify\"?\nA: They emphasise that X-ray crystallography offers a clear mathematical formalism relating the experimental data to the protein-ligand model, with the primary experimental evidence being the electron density of the molecules forming the crystal. Building an atomic model is the interpretation of that density, and the model must then be validated for fit to the data and for agreement with prior expectations of stereochemistry.", "Q: How large are the databases XenoBug searches?\nA: Marc writes that XenoBug integrates approximately 3.3 million enzyme sequences from environmental metagenome databases and 16 million enzymes from 38,000 bacterial genomes, and that its three modules cover reaction class prediction, reaction subclass prediction and enzyme identification by structural similarity.", "Q: Where does Marc say microplastics have been found in the human body?\nA: The Microplastics page states that microplastics now show up in human blood, placental tissue, atherosclerotic plaque, and the brain. The early literature, it says, points toward chronic inflammation, endocrine disruption, and metabolic stress as likely consequences.", "Q: What complex does PDB 4JM2 contain?\nA: PDB entry 4JM2 contains the PGT135 Fab in complex with the gp120 core protein from HIV-1 strain JR-FL, itself bound to CD4 and the 17b Fab, making six unique protein chains in total. Marc Deller's post records an X-ray structure at 3.1 \u00c5 released on 29 May 2013.", "Q: What are the key data-collection statistics for the oncostatin M structure?\nA: For Marc Deller's native oncostatin M data set, collected at 0.979273 \u00c5 on ESRF BM14, there were 43,901 total and 10,513 unique reflections to 2.2 \u00c5 with 96.5% completeness, an Rmerge of 8.3% and an I/\u03c3(I) of 13.2. The three EMP MAD wavelengths (1.00914, 0.826516 and 1.00669 \u00c5) extended to 2.6 \u00c5 with completeness of 96\u2013100%, and MAD phasing gave a figure of merit of 0.46 and a phasing power of 1.84, improved to 0.51 and 1.90 after combining with SIRAS phases.", "Q: What was the significance of the azabicyclo[2.1.1]hexane group?\nA: The azabicyclo[2.1.1]hexane was essentially irreplaceable \u2014 virtually all substitutes were significantly less active, and every compound lacking it remained inactive at over 10 \u03bcM in the exchange assay even late in the program. The crystal structure of compound 3 explained why: the strained bicyclic cage holds the protonated amine in a conformation directly below the plane of the core that would otherwise be very hard to reach, letting it hydrogen bond to Tyr96 and reach Gly10 through a bridging water.", "Q: What are Marc's smaller HTML apps?\nA: They include the Confidence\u2013Flexibility Atlas, the PETase Lineage Map, the Ligandable Cysteine Atlas, a Resolution Dashboard, Name Folder, the Structural Glyphs generative SVG fingerprint atlas and a PDF Timebomb Locker.", "Q: How much data was mined to design the JCSG Top96 screen?\nA: The JCSG Top96 screen was mined from 5.4 million crystallization experiments, 145,000 crystals and 1144 JCSG protein structures deposited in the PDB, covering JCSG depositions between July 2000 and 23 September 2010.", "Q: Who collaborated on the SsgB study?\nA: Xu Q was first author with Traag BA, Willemse J and van Wezel GP, alongside the Joint Center for Structural Genomics team including Marc Deller, Ian Wilson and Scott Lesley. The paper appeared in the Journal of Biological Chemistry in 2009 (DOI 10.1074/jbc.m109.018564).", "Q: What was the exact date Marc started at Incyte?\nA: I don't have his exact start date. I can describe what he built there; for exact dates, marc@marcdeller.com.", "Q: What benefit does Marc claim for AI-guided screening campaigns?\nA: Marc argues that using AI to guide HTS campaigns and predict which compounds are most likely to succeed lets researchers focus effort, shorten timelines and reduce the high attrition rates that plague oncology research, bringing precision medicine closer to reality. The post is at https://marcdeller.com/precision-oncology-integrating-ai-and-high-throughput-screening-to-fight-cancer/.", "Q: How had the AP33 epitope been characterised before this structure?\nA: Before the crystal structure, the AP33 epitope had been mapped and studied extensively by overlapping peptide scanning, phage-display mimotope panning, selection of in vitro escape mutants and site-directed mutagenesis. The crystal structure that Marc Deller helped produce put all of that functional mapping on an atomic footing.", "Q: What is CODSWALLOP?\nA: CODSWALLOP (codswallop.mdeller.com) is a Flask web app that mines the PDB for every structure in a protein family and returns a cross-referenced summary with interactive tables and plots, Mol* viewers, PLIP interaction analysis, alignments, crystallisation conditions, ligands, domains, folds, topology diagrams, resolution, expression systems and constructs. The name is a British-slang backronym: Comparative Overview of Domains, Structures, Whole-families, Alignments, Ligands, Lineages, Organisms & PDB entries.", "Q: What design tasks does Marc say Protein Hunter supports?\nA: According to Marc, the framework handles unconditional design where novel structures emerge without constraints, and conditional generation targeting proteins, cyclic peptides, small molecules, DNA and RNA, as well as more advanced applications such as multi-motif scaffolding and partial protein redesign.", "Q: Where can I verify Marc's data foundations certificate?\nA: His post \"2025 \u2013 Foundations: Data, Data, Everywhere\" at https://marcdeller.com/foundations-data-data-everywhere/ links to the Coursera verification record for the Google-issued certificate.", "Q: How many people has Marc Deller mentored?\nA: Marc Deller has trained and mentored over 50 scientists, lab technicians and research associates. He mentored a team of 3 junior scientists at Scripps, supervised 2 scientists and 50+ users at Stanford, and managed 10+ scientists at Incyte. He also mentored Stanford undergraduates who won a Lemelson-MIT Student Prize in 2017 for work combating antibacterial resistance.", "Q: What was Marc's role in the hepatitis C antibody structure?\nA: Marc Deller contributed crystallography to HCV E2 antibody-complex structures produced with Ian Wilson's laboratory at The Scripps Research Institute, alongside his HIV-1 Env work, while he was Target Manager at the Joint Center for Structural Genomics there. PDB 4G6A was released in September 2012 during that period.", "Q: What budgets has Marc Deller managed?\nA: Marc Deller managed a $1M budget at Stanford's Macromolecular Structure Knowledge Center, budgets of $1.5M+ for platform and infrastructure operations at Incyte, and established Incyte's $8M structural biology platform; his r\u00e9sum\u00e9 summarises this as $8M budgets managed.", "Q: What gap in current approaches does the invention address?\nA: The patent notes that current strategies for microplastics focus on environmental sources and industrial recycling, leaving a critical gap: there is no therapeutic solution for actively degrading and clearing microplastics and nanoplastics that have already been internalized in the human body. Marc Deller's invention aims to fill that gap with enzyme therapeutics for human use.", "Q: Which PETase variant does Marc highlight as working under serum-like conditions in that bloodstream post?\nA: Marc's post highlights the S238A PETase variant as particularly effective at breaking down pre-treated PET particles under serum-like conditions, depolymerizing PET into monomers such as terephthalic acid (TPA) and MHET at body temperature without harming human cells in vitro.", "Q: What is the NANOG paper Marc Deller contributed to?\nA: Marc Deller is an author on the 2015 PNAS paper \"Structure-based discovery of NANOG variant with enhanced properties to promote self-renewal and reprogramming of pluripotent stem cells\", led by Y. Hayashi and colleagues. It used structural biology to engineer an improved NANOG transcription-factor variant for stem-cell reprogramming.", "Q: Where is TopPDBLX's code and data?\nA: TopPDBLX is live at toppdblx.mdeller.com with source at github.com/bellcheddar/TopPDBLX, and the dataset carries a Zenodo DOI (10.5281/zenodo.21807133). The two parser models are published on Hugging Face as Dellboy/toppdblx-residual-parser and Dellboy/toppdblx-construct-boundary.", "Q: What is WRN and why is it a cancer drug target?\nA: WRN, the Werner syndrome helicase, belongs to the RecQ helicase family alongside BLM, RecQ1, RecQ4 and RecQ5, and these enzymes maintain genome stability and suppress neoplastic transformation. The Incyte patent explains that genome-wide CRISPR-Cas9 screens identified WRN as a synthetic lethal dependency in microsatellite instability-high (MSI-H) cancer cells: deleting WRN kills MSI cells but spares microsatellite stable (MSS) cells, which makes WRN an attractive selective cancer target.", "Q: What is BT1062 from Bacteroides thetaiotaomicron?\nA: BT1062 is a putative fimbrial assembly protein and a homolog of Mfa2, a component of the minor fimbriae in Porphyromonas gingivalis. Its crystal structure was determined at 2.2 \u00c5 resolution and revealed a conserved fold widely adopted by fimbrial components.", "Q: Write a poem about proteins.\nA: Poetry isn't really my thing, and I'd rather not attempt one. Proteins, on the other hand, I can do all day: Marc has determined over 400 structures, from HIV envelope trimers to JAK kinases. Would you like to hear about one of them?", "Q: What software does the review list for predicting crystallizability?\nA: Marc Deller and colleagues list DisMeta, XtalPred, POODLE, MFDp2, MoRFpred, RFCRYS, XANNpred, SVMCRYS, SCMCRYS, CRYSTALP2, MetaPPCP and ParCrys. They caution that these tools evaluate a single polypeptide chain in isolation and therefore underrate targets that only become ordered within a protein\u2013protein complex.", "Q: Where can I read about the PGT128 and 8ANC195 Env trimer structure?\nA: Marc Deller's post is \"Crystal Structure of BG505 SOSIP gp140 HIV-1 Env Trimer Bound to Broadly Neutralizing Antibodies PGT128 and 8ANC195\", in the structures section of his site. You can read it at https://marcdeller.com/crystal-structure-of-bg505-sosip-gp140-hiv-1-env-trimer-bound-to-broadly-neutralizing-antibodies-pgt128-and-8anc195/.", "Q: What did the crystal structure show about how the antibody binds mutant CALR?\nA: In the Fab\u201321-mer complex the asymmetric unit contains a Fab bound to two mutCALR peptides, designated CalR1 and CalR2, in two distinct binding conformations. The CalR1 and CalR2 conformations engage two opposite faces of the mutant CALR C-terminal helix, with more N-terminal residues buried in the CalR1 conformation than in the CalR2 conformation, and the contacts were assessed at a 4.5 \u00c5 cut-off in MOE.", "Q: What numbers does Marc give for chemical space and AI-designed drug success rates?\nA: Marc cites ZINC-22 containing 54.9 billion molecules with 5.9 billion ready-to-dock in 3D, and the Enamine REAL database growing from 170 million compounds in 2017 to over 6.7 billion in 2024. He reports AI-designed compounds showing 80 to 90% Phase I success rates versus 40 to 65% for traditional discovery, with timelines potentially compressed from over ten years to three to six and costs cut by up to 70%, against a backdrop of roughly 10% overall success rates and $2.6 billion development costs per approved drug.", "Q: What are smiles2grid's other command-line options?\nA: `--max-pages` renders only the first N pages for debugging, `--label-prefix` changes the label prefix from its default `J`, and `--output` chooses the PDF filename. By default it writes `input_grid.pdf` and `input_summary.csv`.", "Q: What did the polyketide synthase paper resolve?\nA: Marc Deller's post \"2018 \u2013 Structure\u2013function analysis of the extended conformation of a polyketide synthase module\" covers work in the Journal of the American Chemical Society showing that the extended conformation of a PKS module is catalytically competent for both chain translocation and elongation, resolving a long-standing debate about which architecture does the chemistry.", "Q: How often is Marc's protein stability review cited?\nA: \"Protein stability: a crystallographer's perspective\" by Deller, Kong and Rupp, published in 2016 in Acta Crystallographica Section F (Structural Biology and Crystallization Communications) 72(2):72-95, has been cited 440 times according to Google Scholar. It is his third most-cited work and his most-cited first-author paper.", "Q: What metrics does AlphaFraud compute?\nA: TM-score, lDDT, GDT, RMSD, per-residue deviation, a PAE-honesty metric and a novelty score (the maximum percent identity to any pre-cutoff PDB chain).", "Q: Why was the 150 degree domain rotation seen in the WRN structure important?\nA: The 150 degree rotation of domain 2A rearranges the residues needed for ATP hydrolysis, repositioning the conserved Walker A motif so that it forms an entirely new pocket not present in any WRN helicase structure in the Protein Data Bank. That conformational rearrangement created the binding site in which the covalent inhibitor engages the protein, so the structure explained the compounds' mechanism of action.", "Q: What synchrotron and crystallography collaborators did Marc Deller thank?\nA: Marc thanked Shinji Ikemizu and Jeronimo Bravo for introducing him to crystallography, and Vivian Strojanoff (ESRF, France), Karl Harlos, Elspeth Garman, Dave Stuart and Dave Stammers for help with data collection. He also thanked Richard Bryan, Kathryn Measures and Robert Esnouf for computing facilities and Robin Aplin of the Dyson Perrins Laboratory for persistence with the mass spectrometry.", "Q: How does resolution affect ligand fit metrics?\nA: Using ATP-bound structures from the PDB analysed with Twilight, the chapter shows that RSCC is generally better at higher resolution and at lower occupancy-weighted B-factors of the ligand. Marc Deller and his co-authors use this to make the point that local ligand metrics depend closely on global properties of the experiment such as data resolution.", "Q: What was Marc's contribution and institution for the FGFR2 structure?\nA: Marc Deller contributed the structural biology and crystallography through Incyte's structural biology group in Wilmington, Delaware, where FGFR2/3 was among the programmes he supported. The structure was released in November 2022, while he was a Research Fellow in Drug Discovery at Incyte.", "Q: Which institutions funded and supported this HIV-1 Env trimer work?\nA: The work was supported by NIH grants including P01 AI82362, CHAVI-ID UM1 AI100663, R01 AI084817, R37 AI36082 and R01 AI33292, together with the International AIDS Vaccine Initiative Neutralizing Antibody Consortium and Center, a Vidi grant from the Netherlands Organization for Scientific Research, a European Research Council Starting Investigator Grant, a Canadian Institutes of Health Research fellowship and the Ragon Institute. Data collection and extensive crystal screening used the Advanced Photon Source, the Stanford Synchrotron Radiation Lightsource and the Canadian Light Source.", "Q: What was the DRVIA7H-VRC01L chimera?\nA: DRVIA7H-VRC01L was a chimeric antibody built from the DRVIA7 heavy chain paired with the mature VRC01 light chain. It bound gp120 with the same angle of approach as DRVIA7 but behaved very differently with glycans, and it neutralised 3 out of 16 VRC01-resistant isolates, suggesting the heavy chain contributes to strain specificity.", "Q: How were the heavy atom sites found and refined?\nA: The sites were found by manual inspection of anomalous difference Patterson maps calculated with FFT and contoured with GROPAT, confirmed independently with RSPS and the automated procedure in SOLVE, and refined with SHARP. Normalised anomalous scattering magnitudes were estimated with REVISE to help analyse the scatterer positions.", "Q: When was the povorcitinib paper published?\nA: The paper was received on 19 December 2025, revised on 16 April 2026, accepted on 28 April 2026, and published in the Journal of Medicinal Chemistry in 2026 as a Drug Annotation, DOI 10.1021/acs.jmedchem.5c03753. The accompanying PDB entries 10PI and 10PJ were released on 10 June 2026.", "Q: Did Marc Deller write a guide to installing AlphaFold on a Mac?\nA: Yes. \"Alphafold Local Install on Mac\", published 18 August 2025, is Marc's seven-step recipe for installing AlphaFold2 on macOS in about twenty minutes, based on instructions from macinchem.org with his own fixes for dependencies such as kalign and mmseqs2, permissions changes, combined installs and corrected paths for newer systems.", "Q: What resolution was the MtPMO3* crystal structure and what is its PDB code?\nA: The X-ray crystal structure of MtPMO3* determined by Marc Deller was solved to 2.45 \u00c5 resolution, with six noncrystallographic symmetry-related molecules (chains A\u2013F) in the asymmetric unit. The coordinates and structure factors were deposited in the Protein Data Bank with accession number 5UFV.", "Q: Which structures contain a nucleotide?\nA: The two KRAS G12D entries 9E5D and 9E5F both carry GDP with magnesium, and 3BOS, the DNA replication regulator HDA from Shewanella amazonensis, carries CDP with magnesium and sodium. NAG, a glycan rather than a nucleotide, appears on the HIV entries 4NCO, 5C7K, 4JM2 and 5CD5.", "Q: Why is CDK2 an attractive cancer drug target?\nA: CDK2 is attractive because its dysregulation, particularly through CDK2/cyclin E1, drives uncontrolled proliferation and has been identified as a mechanism of resistance to CDK4/6 blockade. Amplification of CCNE1, the gene encoding cyclin E1, correlates with poor prognosis in ovarian, breast, gastric, and endometrial cancers, and preclinical work shows that inhibiting CDK2 can selectively block proliferation and induce G1 arrest in CCNE1-amplified cancer cells. The 2024 paper that Marc Deller co-authored pursued selective CDK2 inhibition for these CCNE1-amplified tumors.", "Q: Has OS3D been validated?\nA: DeepCovalent states that OS3D has been validated in silico on FDA-approved drugs and across the company's internal discovery programmes. The company invites interested parties to contact it to request the white paper.", "Q: What does the B factor tell a crystallographer about stability?\nA: In Marc Deller's review the atomic displacement parameter, or B factor, is described as a statistical measure in units of \u00c5\u00b2 of the probability of finding an atom at its modelled mean position. A high B factor means low certainty about that position, which can reflect dynamic disorder of the chain or short- or long-range disorder in the crystal.", "Q: Where can I read Marc's post on the immature VRC01-class antibody?\nA: The post is \"Crystal Structure of an Immature VRC01-Class Antibody DRVIA7 from a Chinese Donor Bound to Clade A/E HIV-1 gp120 Core\", in the structures section of his website. You can read it at https://marcdeller.com/crystal-structure-of-an-immature-vrc01-class-antibody-drvia7-from-a-chinese-donor-bound-to-clade-a-e-hiv-1-gp120-core/.", "Q: What did the 1B2/KS-AT crystal structure show?\nA: The structure established that the KS-AT portion adopts the extended dimeric architecture, with two copies of the 1B2 Fab binding on opposite sides of the N-terminal coiled-coil region and each Fab also contacting both the KS and AT domains. Specific hydrogen bonds were seen between Arg98 of the Fab light chain and the conserved Asp331 of the KS domain, and between Asn211 of the heavy chain and the carbonyl oxygen of Pro775 of the AT domain. Apart from the N-terminal coiled coil, no significant conformational changes were found compared with the antibody-free KS-AT structure.", "Q: How many years of hands-on drug discovery experience does Marc claim?\nA: The About page pairs Curiosity with 10+ years driving hands-on drug discovery through multidisciplinary approaches, and elsewhere says he has spent 10+ years across protein production, protein engineering, antibodies, small molecules, and AI-guided therapeutics. The chat page phrases the same career as two decades of experience.", "Q: Overall, what was the outcome of Marc Deller's oncostatin M work in Chapter 3?\nA: Chapter 3 successfully delivered a biologically functional, C-terminally truncated OSM mutant (OSM_\u0394188) that could be produced in E. coli, purified to over 90 % by affinity and gel filtration at ~1 mg/L, verified by N-terminal sequencing and mass spectrometry, and shown by receptor and Ba/F3 assays to retain wildtype receptor binding and activity. This material, including its Se-Met form, provided the basis for the crystallisation and structure determination reported in later chapters.", "Q: How well does MICA perform according to Marc's post?\nA: Marc reports an average TM-score of 0.93 on recently released high-resolution cryo-EM maps, with scores above 0.9 considered indicative of high-accuracy models. That is a 22.7% improvement over ModelAngelo (0.92 versus 0.75) and 3.4% over EModelX with AlphaFold (0.92 versus 0.89), with better completeness and longer aligned backbone lengths.", "Q: How does chatPDB separate its three sources of correctness?\nA: Marc's one-line summary is \"RAG keeps it truthful today; the fine-tune makes it consistent tomorrow; Biopython/gemmi/DSSP/PyMOL make it correct always.\" Volatile knowledge lives in retrieval, stable behaviour lives in the fine-tuned weights, and deterministic structural facts such as atom counts, secondary structure, chain composition and geometry are computed live by real tools executed in a sandbox.", "Q: Which structures is Marc Deller listed as a deposition author on personally?\nA: Marc C. Deller is named individually as a deposition author on 9E5D and 9E5F (KRAS G12D), 8UV0 (CDK2), 8E1X (FGFR2), 6OFO (split GFP), 5UFV (fungal polysaccharide monooxygenase), 3L4B and 3JXO (the TrkA gating ring) and 1EVS (human oncostatin M). On the remaining entries he is an author of the primary citation or the work was deposited under the collective JCSG name.", "Q: Which paper reported the DRVIA7 antibody structure?\nA: It was reported in Kong, L., et al. (2016), \"Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor\", Immunity, volume 44, pages 939-950. Marc Deller also has a separate blog post on marcdeller.com about that paper.", "Q: What does compassion mean in Marc's work?\nA: In \"Compassion: 6+ FDA IND filings supported while mentoring 20+ scientists\", Marc Deller describes compassion as what connects the technical side of his work to its human consequences. He writes that every structural insight and design decision ultimately traces back to patients facing serious disease.", "Q: What is the Microplastics page on marcdeller.com about?\nA: The Microplastics in Humans page, subtitled \"Therapeutic Vision: Tackling Microplastics in humans\", describes Marc Deller's work on the health burden of microplastics and the therapeutic response he is building at Elora Therapeutics. It frames the problem as frontier science with no playbook and a protein-engineering problem at its heart.", "Q: Did Marc Deller write anything outside science, such as about a colleague's hat?\nA: Yes. \"The Signature Cowboy Hat of Adam Lee: A Blend of Tradition and Function\", published 27 August 2025, is a light-hearted piece about Adam Lee, Director of Advanced Analytics and Data Engineering at Incyte, and the cream-colored felt cowboy hat he wears.", "Q: What was compound 17's cynomolgus monkey PK like?\nA: In cynomolgus monkeys, compound 17 dosed orally at 2 mg/kg showed good exposure (AUC of 5652 nM\u00b7h) and low clearance (9% of hepatic blood flow), with an oral bioavailability of 21% and an intravenous half-life of 2.6 hours. This balanced cyno PK profile is what led the team to select 17 as the lead candidate despite its higher clearance in rats.", "Q: How is the Bifidobacterium phosphoesterase structure organised?\nA: The structure consists of a polymerase and histidinol phosphatase (PHP, Pfam PF02811) domain with a second domain formed by residues 105 to 178 inserted in the middle of the PHP sequence. The insert domain binds AMP, although its precise function and substrate specificity remain unknown.", "Q: Did the trimer structure say anything about the gp41 fusion peptide?\nA: Only tentatively. The electron density for the fusion peptide and the residues connecting it to the HR1 central helix was weak and diffuse, making model building difficult at 4.7 \u00c5 and suggesting a lack of regular secondary structure. Nonetheless, residues adjacent to the N-termini of the HR1 helices were seen to wrap around the gp120 \u03b10 helix, forming a metastable loop-helix structure akin to elements in influenza HA2 that extend the central helix in the post-fusion form.", "Q: Can you help me write a Python script to scrape a website?\nA: General coding help isn't what I'm here for, but I can tell you about Marc Deller's own software projects, like the vibe-coded apps he's built with AI assistance (AlphaFraud, BoltzMaker, ChemSage and more), or how he fine-tunes small language models with MLX-LM. Want to hear about one of those instead?", "Q: Where can I find the bicyclic WRN inhibitor patent post?\nA: Marc Deller published it on 12 August 2025 at https://marcdeller.com/bicyclic-compounds-as-inhibitors-of-wrn/, with a link to Google Patents for US20250236622A1.", "Q: Why are MACPF proteins interesting in general?\nA: Membrane-attack complex/perforin proteins are transmembrane pore-forming proteins important in both human immunity and the virulence of pathogens. The Bth-MACPF structure that Marc Deller co-authored provided new features and insights not seen in the two previous MACPF structures.", "Q: Who is the assignee of the CALR antibody patent?\nA: The applicant and assignee is Incyte Corporation, 1801 Augustine Cut-Off, Wilmington, Delaware 19803, USA. The prosecuting agent was Catherine Guzzo of Fish & Richardson P.C., Minneapolis.", "Q: What did the calycin signature in DUF2006 imply?\nA: Finding the calycin signature in the N-terminal domain, combined with remote sequence similarity to PF07143 and PF08622, two families implicated in isoprenoid metabolism and the oxidative stress response, supported an involvement in lipid metabolism. Clusters of conserved residues interacting with ligand mimetics suggested the binding site maps to the N-terminal domain and the regulation site to the interdomain interface.", "Q: What crystallisation screen was used, and why is that interesting for Marc?\nA: The crystals were grown from the JCSG Top96 Screen (Rigaku). That is notable because Marc Deller was one of the people behind the Top96 screen itself, so the patent's structures were obtained with a crystallisation screen he had a hand in creating.", "Q: What chemistry was used to make these CDK2 inhibitors?\nA: The (4-pyrazolyl)-2-aminopyrimidines were built largely through Suzuki\u2013Miyaura cross-coupling of pyrazole boronic acids or esters onto chloropyrimidine cores, together with SNAr and Buchwald\u2013Hartwig amination to install the aminopiperidine, Boc protection/deprotection, mesylation or sulfonylation of the piperidine nitrogen, and reductive amination. Compound 17, for example, was made by Suzuki coupling followed by Boc removal and reaction with cyclopropanesulfonyl chloride, and was prepared on 10 g scale at >99.5% purity.", "Q: How does AP33 differ from antibody HCV1 in the way it uses its variable domains?\nA: In the HCV1 structure, the tip of the \u03b2-hairpin points toward the light chain variable domain (158 \u00c52 buried surface) while most of the hairpin interacts with the heavy chain variable domain (300 \u00c52). In AP33, by contrast, VL and VH interact almost equally with the N- and C-terminal \u03b2-strands of the antigen, burying 242 \u00c52 and 273 \u00c52 respectively. Superposing the epitope backbones shows Gln412\u2013Asn417 to be very similar between the two structures, with the differences concentrated in Ser419\u2013Asn423.", "Q: Which claims come most directly from the structural biology?\nA: Claims 55 to 71 are the epitope claims and follow directly from the co-crystal structures. Claim 55 covers any antibody binding within DEEQRTRRMMRTKMRMRRMRR; claims 56 to 65 narrow that to specific residue ranges and named contact residues; and claims 66 to 71 cover antibodies binding two non-identical epitopes within the same peptide, with the first and second epitopes defined by the CalR1 and CalR2 residue sets seen in the crystal.", "Q: Where is that human-skills post?\nA: Marc Deller published \"2025 \u2013 Cultivating Human Skills and Decision-Making\" at https://marcdeller.com/cultivating-human-skills-and-decision-making/.", "Q: What is AlphaFraud?\nA: AlphaFraud (alphafraud.mdeller.com, github.com/bellcheddar/AlphaFraud) is Marc's web app that audits AlphaFold's blind predictions against freshly deposited human PDB structures. Because AlphaFold2 trained on chains released before 30 April 2018, every later human structure is a real blind test; AlphaFraud superposes each one on its AlphaFold DB model and flags predictions that are confidently wrong.", "Q: Where was the BT1062 structure published?\nA: The BT1062 structure appeared in Acta Crystallographica Section F, Structural Biology and Crystallization Communications, in 2010, with Xu Q as first author and Marc Deller among the Joint Center for Structural Genomics co-authors. The DOI is 10.1107/s1744309110006548.", "Q: Which grants is Marc named on?\nA: I don't have a grant list. Funding acknowledgements appear in some of his papers, which are listed on marcdeller.com.", "Q: Which OSM construct did Marc actually manage to crystallise?\nA: Of the three samples Marc Deller tried (OSM-196, \u0394188_OSM and Se-Met-OSM), only the truncated construct \u0394188_OSM produced crystals. Both OSM-196 and the selenomethionine-labelled OSM showed exceptionally poor solubility and aggregation, precipitating heavily in around 70 % of screening conditions.", "Q: Which immunology structures were covered in the cell surface receptors review?\nA: Marc Deller's review covered progress on natural killer cell receptors, including the first C-type lectin fold NK receptor structure and a complex between such a receptor and an MHC class I molecule that suggested two possible interaction modes, plus further Ig-like killer cell inhibitory receptor structures. It also highlighted three IgSF receptor structures bridging humoral and cellular immunity: two Fc gamma-receptors and the high-affinity IgE receptor, the latter offering a basis for designing anti-allergy therapies.", "Q: How potent are the compounds in the WRN patent?\nA: Tables 4 and 5 report ATPase and helicase activity in banded form rather than exact IC50 values, where \"+\" means less than 100 nM. Most of the 26 exemplified compounds fall into that most potent band in both assays, with a handful \u2014 Examples 10, 16 and 18 \u2014 landing in weaker bands, and Examples 16 and 18 in the weakest helicase band of 5,000\u201320,000 nM.", "Q: How was the MtPMO3* structure solved?\nA: X-ray diffraction data were collected at beamline 12-2 of the Stanford Synchrotron Radiation Lightsource (SSRL) and processed with Denzo/Scalepack from HKL2000. Initial phases were obtained by molecular replacement using a search model based on Cel61B (PDB 2VTC), a PMO from the soft-rot fungus Hypocrea jecorina. The model was built in COOT and refined with Refmac5 in the CCP4 package. Ian Wilson was thanked for SSRL beam time.", "Q: What is Marc's website?\nA: Marc's personal website is https://marcdeller.com, where he keeps his resume, publications, PDB structures, patents, projects, blog and this chatbot. His web apps run on subdomains of https://mdeller.com.", "Q: How many IND filings has Marc Deller supported?\nA: Marc Deller has supported 6+ preclinical IND filings. At Pfizer he helped file 3 FDA INDs (VEGFR, ERK, MEK and HIV-protease programs), and at Incyte he supported IND-enabling work across KRAS-G12D, WRN, CDK2/4, JAK1/2, FGFR2/3 and the mutCalR antibody.", "Q: How are these WRN compounds actually made?\nA: The core route condenses a benzimidazol-2-amine with a dicarbonyl compound at 140 \u00b0C to build the tricyclic pyrimidinone, brominates it with N-bromosuccinimide, protects the ring NH with a SEM group, installs the aryl group by Suzuki coupling with an aminophenyl boronic acid using XPhos Pd G4, acylates the resulting aniline with acryloyl chloride or ethenesulfonyl chloride to install the warhead, deprotects with TFA, and finally alkylates the ring nitrogen with a bromoacetamide. The specification gives six alternative schemes for reaching the same compounds.", "Q: What factors make ligand density weak?\nA: The chapter lists partial occupancies, enhanced mobility, conformational flexibility and poor protein stability as factors that further reduce the discrete scattering contribution of a bound ligand. Marc Deller and his co-authors cite their own 2016 review on protein stability on this point.", "Q: Which target area is DeepCovalent's lead programme in?\nA: DeepCovalent's disclosed lead compounds are small-molecule GLP-1 receptor candidates. GPCR-targeting small-molecule discovery is also one of the areas Marc Deller advises the company on as its Scientific Technical Advisor.", "Q: Why is the two-subunit TrkA gating ring evolutionarily interesting?\nA: Marc Deller's post describes the two-subunit assembly as an evolutionary precursor to modern single-subunit gating ring systems, so the structure shows how potassium transport regulation may have evolved. It is the first structural view of such a two-subunit Trk gating ring.", "Q: Does Marc have any tattoos?\nA: That's outside what I know about Marc, I'm afraid. I stick to his science, his career and his projects. If you want to ask him yourself, he's at marc@marcdeller.com.", "Q: How did the crystal structure compare with the cryo-EM arched conformation?\nA: The extended KS-AT conformation in the 1B2/KS-AT complex was very different from the arched conformation attributed by cryo-EM to the homologous KS-AT portion of module 5 of the pikromycin synthase. The 1B2-bound structure showed excellent backbone overlap with the previously reported KS-AT fragments of DEBS modules 3 and 5, the latter of which also includes an N-terminal coiled coil.", "Q: How does the review recommend dealing with glycosylation?\nA: Marc Deller and colleagues list expression in glycosylation-deficient cells such as CHO lec1 or HEK293S GnTI-minus, mutagenesis of the Asn-X-Ser/Thr sequon or of O-linked Ser and Thr sites, enzymatic trimming with Endo H or PNGase F, and glycosylation inhibitors such as swainsonine and kifunensine. They also note that glycans are not always harmful and that fully glycosylated hepatitis C virus E2 crystallized only with its glycans intact, one of which mediated a key crystal contact.", "Q: Where can I read Marc's post on validating protein\u2013ligand models?\nA: The post \"2017 \u2013 Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands\" is at https://marcdeller.com/validation-of-protein-ligand-crystal-structure-models-small-molecule-and-peptide-ligands/ and links to a PDF of the chapter.", "Q: What did Marc say about the roles of Asn 123 and Asn 124 of OSM?\nA: In Marc Deller's model, Asn 124 on helix C, whose mutation had the greatest effect on gp130 binding and biological activity (over 100-fold and over 10,000-fold respectively), is directed towards Phe 169 and the B-C loop of gp130, whereas Asn 123 has higher solvent accessibility and faces more outward, so it interacts less with that region of the receptor, consistent with its milder mutational effect.", "Q: How does AlphaFraud define \"confidently wrong\"?\nA: A prediction is \"confidently wrong\" when its mean pLDDT is above 70 (AlphaFold is confident) but the TM-score against the experimental structure is below 0.5 (the fold is wrong). A composite FRAUD score weights the error by confidence.", "Q: What is the systems approach to drug discovery Marc Deller describes?\nA: In \"From Single Targets to Disease Networks: A Systems Approach to Drug Discovery\", published 13 August 2025, Marc Deller argues that the traditional \"one target, one drug\" model ignores the fact that diseases arise from failures of interconnected biological networks, and that combining high-throughput structural proteomics with computational modeling now lets researchers map the interactome.", "Q: What is Marc's framing of the drug resistance problem?\nA: Marc frames it as an evolutionary arms race in oncology and infectious disease that researchers often lose, since developing each next-generation drug after resistance emerges is costly and slow. His argument is that predicting resistance before it appears in the clinic lets researchers act proactively; the post is at https://marcdeller.com/outsmarting-evolution-using-ai-to-overcome-drug-resistance/.", "Q: What does Marc's Publications page say about his publication record?\nA: The Publications page, subtitled \"From Structure to IND: Stories Behind My Publications\", describes 60+ top-cited peer-reviewed manuscripts covering structural biology, medicinal chemistry, and immunology. Marc Deller says they highlight his expertise in designing key experiments, solving complex biological problems, and mentoring teams.", "Q: How many times has Marc moved house?\nA: That isn't something I know. He is based in Wilmington, Delaware now, and his career has spanned the UK and the US.", "Q: Which parts of the OSM molecule were disordered in the crystal structures?\nA: In the P212121 structure residues 4\u2013134 and 156\u2013187 are well ordered, while residues 1\u20133 at the N-terminus and residues 135\u2013155 of the C-D loop have no interpretable density. In the I222 structure residues 6\u201394, 104\u2013132 and 158\u2013187 are ordered, with residues 1\u20135, the B-C loop (95\u2013103) and the C-D loop (133\u2013157) disordered. High main-chain B-factors in these regions indicate flexibility within the crystal.", "Q: What did the crystal structure of compound 3 reveal?\nA: The X-ray structure of compound 3 bound to KRAS G12D showed why the azabicyclo[2.1.1]hexane group was so critical for potency. Its protonated amine makes two key interactions with the protein \u2014 a direct hydrogen bond to Tyr96 and a water-mediated interaction with Gly10 \u2014 because the rigid bicyclic framework orients the amine in an otherwise hard-to-access conformation directly below the plane of the imidazole core. Contrary to the team's starting hypothesis, no interactions were made with Asp12, which told them potency could be gained by growing toward that residue.", "Q: What is the secondary coordination sphere paper about?\nA: Marc Deller's post \"2017 \u2013 The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\" covers work published in ACS Chemical Biology showing that the hydrogen-bonding network around the copper active site of a fungal polysaccharide monooxygenase governs oxygen activation and proton transfer during cellulose degradation.", "Q: What resolution is needed for structure-based lead discovery?\nA: The chapter states that validating a protein\u2013ligand structure for use in structure-based lead discovery typically requires a higher level of scrutiny and generally a resolution of around 2 \u00c5 or better. Less scrutiny is needed when the structure is only being used to analyse global features of a binding site, for example for structure-guided mutagenesis.", "Q: Can the tricyclic WRN compounds be combined with other drugs?\nA: Yes. The patent envisions uses beyond monotherapy, with the compounds intended to be versatile enough to combine with modern therapies such as checkpoint inhibitors for a more comprehensive attack on difficult-to-treat cancers.", "Q: In simple terms, what did Marc contribute to it?\nA: Marc Deller worked out the three-dimensional picture of exactly how the antibody grips the mutant protein's abnormal tail. Using X-ray crystallography he grew crystals of the antibody fragment together with pieces of the mutant tail, shot them with X-rays at synchrotrons, and built atomic models showing which individual amino acids touch which. That picture is what allowed the patent to claim the antibody not just by its own sequence but by the precise part of the target it binds \u2014 a much broader and more durable form of protection.", "Q: What was Marc Deller's contribution to the WRN inhibitor patent?\nA: Marc Deller contributed the structural biology and X-ray crystallography that underpins the patent's mechanistic disclosure. The application contains a full protein expression, purification, co-crystallization, data collection and structure refinement account for the WRN helicase core in complex with a WRN inhibitor compound \u2014 the kind of work Dr Deller led in Incyte's structural biology group \u2014 and that co-crystal structure is the basis of the patent's central mechanistic claim about a novel Cys727 binding pocket.", "Q: What evolutionary model did the TM1088 paper propose?\nA: Marc Deller and colleagues proposed that a simple gene duplication and fusion event, followed by different patterns of domain organization, could account for the differences between the two-subunit TM1088 assembly and single-subunit gating rings. For example, duplication and fusion of the two TM1088B chains into one polypeptide with two tandem RCK repeats would give the classical single-chain arrangement. The paper suggested TM1088 may represent an evolutionary ancestor of more modern single-subunit K+ gating rings.", "Q: Why are PMOs commercially important?\nA: PMOs act synergistically with hydrolytic cellulases and therefore have commercial value in enhancing cellulose degradation for the production of renewable fuels and chemicals. The 2017 paper co-authored by Marc Deller noted that PMOs also represent a newly discovered enzyme superfamily with novel catalytic features that are not yet fully understood.", "Q: What two enzymes does Marc compare in his HGMP01 versus FAST-PETase post?\nA: In \"HGMP01 and FAST-PETase: Decoding Two Paths to Plastic Degradation \u2013 Engineered versus Natural PETase Enzymes\" (5 November 2025), Marc Deller compares FAST-PETase, a machine-learning-engineered variant of the *Ideonella sakaiensis* enzyme, with HGMP01, a naturally evolved PET hydrolase discovered in the human gut microbiome.", "Q: What did mass spectrometry show for oncostatin M?\nA: Electrospray mass spectrometry of OSM_\u0394188 showed one major peak at 21579.98 \u00b1 9.05 Da. This differs from the theoretical mass (variously given as 21409.50 Da by the ExPASy calculation and 21259 Da), a discrepancy Marc attributed to bound ions, detergent, glycerol or other buffer components. The Se-Met material gave five peaks spanning the sequential selenium substitutions.", "Q: Which structural genomics consortia feature in the review?\nA: The Joint Center for Structural Genomics, where Marc Deller worked, features prominently as a source of Thermofluor and DXMS data and of large truncation arrays made by polymerase incomplete primer extension cloning. The review also cites the Structural Genomics Institute at Karolinska Institutet as generating thousands of truncations for recalcitrant targets.", "Q: Who was the corresponding author of the CDK2 paper?\nA: The corresponding author of the 2024 CDK2 paper was Joshua R. Hummel of the Incyte Research Institute. Marc Deller was one of the 27 co-authors and contributed the structural biology, determining the CDK2 co-crystal structure with Leslie Epling.", "Q: Which enzymes besides PETase does Marc cover in \"The Enzyme Toolbox\"?\nA: In \"The Enzyme Toolbox\u2014Alternatives to PETase for Microplastic Breakdown\" (15 August 2025), Marc Deller covers cutinases, carboxylesterases, lipases and MHETases as alternatives and complements to PETase, noting that they provide expanded substrate flexibility and catalytic power.", "Q: What is Elora Therapeutics developing?\nA: According to its 2025 investor pitch deck, Elora Therapeutics is developing a first-in-class systemic enzyme therapeutic to address cardiometabolic disease by reducing the microplastic burden in the human body. The lead approach combines the engineered enzymes PETase and MHETase to break down PET microplastics in blood and tissues. Marc Deller is co-founder and Chief Scientific Officer of the company.", "Q: What ligand was found bound to TM1088A?\nA: AMP was found bound to TM1088A in both the high-resolution TM1088A structure and the octameric assembly. AMP had not been added to the crystallization reagents and was presumably scavenged from the E. coli cells during expression. Marc Deller's paper noted that earlier DSC and Thermofluor experiments had shown AMP stabilizes TM1088A by +5.8 \u00b0C (DSC) and +8 \u00b0C (Thermofluor), more than NADH (+2.3 \u00b0C and +4 \u00b0C), and that TM1088B was not stabilized by nucleotides.", "Q: What does Marc like to drink?\nA: Earl Grey tea, San Diego IPAs, Yuengling, and Timothy Taylor's Boltmaker (the beer that inspired the name of his BoltzMaker tool). He prefers beer to cocktails, though a good gin and tonic with lemon is welcome.", "Q: At what resolution was the human kanadaptin FHA domain solved?\nA: The human kanadaptin FHA domain was determined to 1.6 \u00c5 resolution, and the 2014 PLoS ONE paper reporting it lists Marc Deller as second author.", "Q: What is ProDomino, as covered on Marc Deller's blog?\nA: In \"Building Proteins Brick by Brick: How AI Is Assembling Novel Bio-Machines\", published 13 August 2025, Marc Deller describes ProDomino as an open-source AI model developed at Heidelberg University that predicts how to combine different protein domains into new, functional proteins, acting like an expert assembly manual.", "Q: Why does Marc argue for multi-enzyme approaches in medicine?\nA: Marc argues that human serum and organs contain a complex cocktail of microplastics, so selecting the right mix of enzymes \u2014 for example combining PETase with a cutinase, or engineering multi-functional fusions with carbohydrate-binding modules or hydrophobins \u2014 improves efficiency and gives broader polymer coverage with better safety profiles. The post is at https://marcdeller.com/the-enzyme-toolbox-alternatives-to-petase-for-microplastic-breakdown/.", "Q: What coverage figures does Marc quote for Sci-Hub?\nA: Marc quotes the 2018 eLife study by Himmelstein and colleagues, which mapped coverage across 81.6 million Crossref-registered articles and found Sci-Hub held 97.3% of Elsevier's catalog, 85.1% of all toll-access journals and 92.8% of chemistry content \u2014 access exceeding what researchers at the University of Pennsylvania had.", "Q: How do you select which compounds smiles2grid renders?\nA: Either by a similarity search against a query SMILES (`--query-smiles \"CCO\" --similarity 75`) or by range selection (`--range 1-25`, `--range all`, `--range 1,5,10-20`). Run interactively it prompts for a query SMILES, a similarity cutoff and then a range; `--no-prompt` disables prompting.", "Q: Which parts of the oncostatin M molecule could not be modelled?\nA: Marc Deller's final oncostatin M model comprised residues 4\u2013134 and 156\u2013187 of the mature protein; residues 1\u20133 and the long CD loop (residues 135\u2013155) were disordered and were omitted, and the BC loop between Glu90 and Lys110 was flexible with higher-than-average B factors. No residues fell in disallowed regions of the Ramachandran plot and 116 water molecules were located.", "Q: What is Marc Deller's 2022 Journal of Medicinal Chemistry paper about?\nA: The 2022 J. Med. Chem. paper Marc Deller co-authored, \"Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3,\" describes an Incyte medicinal-chemistry campaign that produced a new class of FGFR2/3 inhibitors. Unlike the first-generation pan-FGFR drugs, these compounds keep their potency against the common gatekeeper mutants and are highly selective over FGFR1. The work was published in J. Med. Chem. 2022, volume 65, pages 15433\u201315442, with Artem Shvartsbart as corresponding author.", "Q: What shape does the TM0077 esterase assemble into?\nA: The TM0077 structure shows a classic \u03b1/\u03b2-hydrolase fold assembled into a doughnut-shaped hexamer. Marc Deller's post notes that the structure also confirms the catalytic serine mechanism of the enzyme.", "Q: Does Marc discuss the safety issues with bispecific antibodies?\nA: Yes. He identifies cytokine release syndrome as the primary toxicity, arising when the activated immune system triggers a systemic inflammatory response, but notes it is controllable through step-up dosing regimens and prophylactic measures and is generally more manageable than CAR-T-associated toxicity. He also discusses resistance mechanisms such as antigen loss and T-cell exhaustion, and applications beyond oncology including emicizumab in haemophilia A and faricimab in retinal disease. The post is at https://marcdeller.com/the-40-billion-revolution-how-bispecific-antibodies-are-transforming-cancer-care/.", "Q: What did Marc measure about pocket tightness in BoltzMaker?\nA: `Pocket distance` turned out to be the lever: the identical 62 residues at 4 \u00c5 instead of 8 \u00c5 took the pose from 9.51 \u00c5 to 2.56 \u00c5. Contact count did not order the outcome at all, and a wrong-site control put own-site poses at 0.91 and 2.79 \u00c5, unconstrained at 6.17 and 13.57 \u00c5, and wrong-site at 24.93 and 29.40 \u00c5, proving the restraint steers rather than fabricates. His recommendation is `Pocket distance: 4` for a named pocket.", "Q: Who did Marc Deller thank in the acknowledgements of the TM1088 paper?\nA: The acknowledgements thank other members of the Joint Center for Structural Genomics for technical assistance, in particular Tom Clayton, Henry Tien, Tina Trout, Julie Feuerhelm and Anna Grzechnik. Genomic DNA from T. maritima MSB8 was obtained from the American Type Culture Collection, and X-ray data were collected at ALS beamlines 5.0.1 and 5.0.3, supported by the US Department of Energy.", "Q: What three main benefits does automated harvesting offer?\nA: Marc Deller and Bernhard Rupp conclude that automated harvesting offers the ability to harvest microcrystals that cannot easily be picked up by hand, reduced late-stage crystal failures from poor handling and cryoprotection, and increased throughput that closes the automation bottleneck in the pipeline.", "Q: What is ForgeFold?\nA: ForgeFold, in the FoldForge repository, is Marc Deller's single-file browser GUI covering the whole Boltz-2 workflow: build, run and review. It brings Boltz setup, YAML construction, job-script generation and result inspection into one lightweight HTML application that opens in any browser with no build step.", "Q: What are the resolutions and space groups of the JAK1 and JAK2 povorcitinib structures?\nA: The JAK1 JH1 domain complex with povorcitinib, PDB entry 10PI, was solved by X-ray diffraction to 1.54 \u00c5 resolution in space group C 1 2 1, with 594 polymer residues and about 70.7 kDa in the asymmetric unit. The JAK2 JH1 domain complex, PDB entry 10PJ, was solved to 1.59 \u00c5 in space group I 1 2 1, with 632 polymer residues and about 75.7 kDa. Leslie B. Epling and Gustavo Fenalti are the deposition authors, with Marc Deller a co-author of the primary citation.", "Q: Summarize the headline findings of Deller et al. 2000 in Structure.\nA: Deller et al. 2000 reported the 2.2 \u00c5 crystal structure of human oncostatin M (PDB 1EVS), showing a four-helix bundle with a strongly kinked helix A characteristic of LIFR-binding cytokines. Mutagenesis mapped the gp130-binding site 2 to Asn124, Gly120, Gln20 and Gln16 on helices A and C, centred on a Gly120 dimple proposed to accommodate gp130 Phe169, and mapped the LIFR/OSMR-binding site 3 to the protruding Phe160/Lys163 pair at the start of helix D. The work defined the hOSM\u2013gp130 interface for the first time and proposed that a conserved hydrophobic core interaction plus variable polar contacts underlies specificity across the cytokine receptor superfamily.", "Q: Does Marc train his own language models?\nA: Yes. On a 64 GB Apple M1 Mac he fine-tunes small open-weight models with Apple's MLX-LM, including ChemSage (drug discovery), chatMCD (about himself) and a model trained on the PSI TargetTrack archive.", "Q: Why does this enzyme work matter practically?\nA: The post frames these mechanistic insights as key to optimising enzymatic cellulose degradation, with transformative potential for biofuel production and sustainable biomass utilisation, and notes that harnessing the second-sphere network could enable next-generation LPMO engineering for precision biotransformations.", "Q: What is SSO2064?\nA: SSO2064 is the first structural representative of PF01796 (DUF35), a large prokaryotic family with wide phylogenetic distribution. The structure, co-authored by Marc Deller and published in Acta Crystallographica Section F in 2010, revealed a novel two-domain architecture of an N-terminal rubredoxin-like zinc ribbon and a C-terminal oligonucleotide/oligosaccharide-binding fold domain.", "Q: Can you recommend a good restaurant near me?\nA: I don't have that kind of information, sorry. What I can do is tell you about Marc Deller's science, his company Elora Therapeutics, or his vibe-coding projects. Curious about any of those?", "Q: How does chatPDB relate to ChemSage?\nA: Marc describes chatPDB as a sibling project to ChemSage, reusing its proven MLX-LM/QLoRA/RAG process on a fresh domain: protein structures instead of small-molecule chemistry. It reuses the same `chat_remote.py` hosting pattern and runs as a dedicated systemd service on the same droplet alongside ChemSage and BoltzMaker's web apps.", "Q: Is Marc a big asshat?\nA: Absolutely, a big one. Though not quite as big as Adam.", "Q: What was the main conclusion of the PKS extended conformation paper?\nA: The paper concluded that the extended conformation of the KS-AT didomains of both DEBS modules 2 and 3 is catalytically competent for all essential PKS reactions, namely chain translocation into and out of the module, extender unit transacylation, chain elongation and \u03b2-ketoreduction, and that conformational reorganization of the KS, AT and KR domains is not required during catalysis. The arch-shaped conformation is not needed for any of these activities.", "Q: Which security headers does Site Health Inspector audit?\nA: All the headers recommended by the OWASP Secure Headers Project: Strict-Transport-Security, Content-Security-Policy, X-Content-Type-Options, X-Frame-Options, X-XSS-Protection, Referrer-Policy, Permissions-Policy, Cross-Origin-Embedder-Policy and Cross-Origin-Opener-Policy.", "Q: How do you get started with GPCR Binning?\nA: `git clone https://github.com/bellcheddar/gpcr-binning.git` then `cd gpcr-binning`. Usage instructions are to be documented as the tooling stabilises.", "Q: Where can I read Marc's piece about a fungal prosthetic organ for microplastics?\nA: It is at https://marcdeller.com/the-79th-organ-when-biology-meets-bioengineering-in-the-fight-against-microplastics/, published on marcdeller.com on 13 October 2025.", "Q: How many job offers has Marc turned down?\nA: That's not something I know, and it isn't the kind of thing I'd guess at. If you're thinking about working with him, write to marc@marcdeller.com.", "Q: What does the About page on marcdeller.com say?\nA: The About page, subtitled \"Protein to Medicine: My Path to Impact\", describes Marc Deller as a structural biologist and drug discovery leader focused on the intersection of protein structure, protein production, platform development, and translational science. It says his work spans X-ray crystallography, Cryo-EM, discovery strategy, team leadership, and the systems needed to move difficult problems toward real outcomes.", "Q: Has Marc watched The IT Crowd?\nA: Marc likes The IT Crowd, but if Jim is asking, he hasn't actually watched it yet.", "Q: Why is KRAS G12D a hard drug target?\nA: KRAS binds GTP with exceptional affinity, which rules out direct competition at the active site, and the 2013 discovery that unlocked KRAS G12C depended on covalently modifying a cysteine. Aspartate is far less nucleophilic than cysteine, so a covalent approach is uncertain and would require an electrophile aggressive enough to raise selectivity and safety concerns. The alternative \u2014 a fully reversible inhibitor \u2014 has historically struggled to combine sufficient potency with the ADME properties needed for oral exposure, which is exactly the problem the paper Marc Deller co-authored set out to solve.", "Q: What is PDB entry 5C7K?\nA: 5C7K is the crystal structure of the BG505 SOSIP gp140 HIV-1 Env trimer bound to two broadly neutralising antibodies, PGT128 and 8ANC195. Marc Deller's post records an X-ray structure at 4.6019 \u00c5 released on 7 October 2015, with six unique protein chains and NAG and sulfate ligands.", "Q: What does chatMCD stand for?\nA: MCD are Marc C. Deller's initials, so chatMCD is simply \"chat with (about) Marc C. Deller\". It runs on a small language model that Marc fine-tuned himself with MLX-LM on his own papers, thesis, patents and notes.", "Q: What was the starting point for the KRAS G12D program?\nA: The program started from compound 1, a single small-molecule example disclosed in a patent, derived from an earlier G12C lead by removing the acrylamide warhead; it showed under 15 \u03bcM potency against KRAS G12D. Tested in house, compound 1 gave an IC50 of 8 \u03bcM in a nucleotide exchange assay. Bridging the piperazine ring gave compound 2, and a screen of rigid amines on a tricyclic imidazole scaffold identified compound 3 with an exchange IC50 of 671 nM.", "Q: Are you ChatGPT?\nA: No, I'm chatMCD, a small language model Marc Deller fine-tuned himself with MLX-LM, trained specifically on his own papers, patents, r\u00e9sum\u00e9 and notes. I only know about Marc and his work, not general topics the way a larger assistant would.", "Q: Where can I read Marc's post about the fungal polysaccharide monooxygenase?\nA: Marc Deller's post \"Crystal Structure of a Cellulose-Active Polysaccharide Monooxygenase from M. thermophila (MtPMO3*)\" is in the structures section of his website. You can read it at https://marcdeller.com/crystal-structure-of-a-cellulose-active-polysaccharide-monooxygenase-from-m-thermophila-mtpmo3/.", "Q: Give me the details of PDB entry 9E5D.\nA: 9E5D is an X-ray structure at 1.36 \u00c5 in space group P 21 21 21, with 170 polymer residues and 20.5 kDa, released on 22 January 2025 and classified as HYDROLASE/HYDROLASE INHIBITOR. Its bound ligands are GDP and magnesium, it was deposited by Marc Deller and L.B. Epling, and it supports the INCB159020 KRAS G12D inhibitor paper.", "Q: What is chatRDKit?\nA: chatRDKit is Marc Deller's natural-language chat frontend for RDKit, powered by a local Ollama model (default `qwen3.6:27b`). It lets a user describe a cheminformatics task in plain English and translates it into RDKit code that runs in a persistent local Python session.", "Q: Where is Marc's Nanog post?\nA: The post \"Crystal Structure of Nanog Homeodomain (NANOG) from Homo sapiens at 3.30 \u00c5 Resolution\" is in the structures section of marcdeller.com. You can read it at https://marcdeller.com/crystal-structure-of-nanog-homeodomain-nanog-from-homo-sapiens-at-3-30-a-resolution/.", "Q: What was unexpected about the YP_001813558.1 dimer?\nA: In dimeric dUTPase-like proteins a central four-helix bundle forms the active site, but in YP_001813558.1 an unexpected intertwined swapping of two of those helices produces a \"linked dimer\" not previously observed for the family. The protein belongs to a superfamily of all-alpha-helical NTP pyrophosphatases that also includes dimeric dUTPases, MazG proteins and HisE-encoded phosphoribosyl ATP pyrophosphohydrolases.", "Q: What is INCB159020 and which paper describes it?\nA: INCB159020 is an orally bioavailable KRAS G12D inhibitor, compound 23, described in the 2025 Journal of Medicinal Chemistry paper \"Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\" (DOI 10.1021/acs.jmedchem.4c02662). Marc Deller is a co-author who contributed structural biology from Incyte's structural biology group.", "Q: Which OSM residues form the site 2 epitope for gp130 binding?\nA: Marc Deller's thesis reports that the site 2 epitope of OSM for gp130 comprises Gln 16 and Gln 20 on helix A and Gly 120, Asn 123 and Asn 124 on helix C. Alanine mutation of these residues reduced gp130 binding by roughly 8- to over 100-fold with corresponding losses of biological activity but no effect on LIFR binding. All five are structurally equivalent to residues directly involved in the Epo\u2013EpoR interface.", "Q: Where is the 10E8 post on Marc's blog?\nA: The post \"Crystal Structure of the Complex of 10E8 Fab Light Chain Mutant1 and T117v2 HIV-1 MPER Scaffold\" appears in the structures category of marcdeller.com. You can read it at https://marcdeller.com/crystal-structure-of-the-complex-of-10e8-fab-light-chain-mutant1-and-t117v2-hiv-1-mper-scaffold/.", "Q: Which clinical molecules does Marc's CALR post reference?\nA: The references in Marc Deller's CALR post cite work on INCA033989, a monoclonal antibody selectively targeting mutated calreticulin, and on INCA035784, a T-cell redirector for CALR-mutated myeloproliferative neoplasms, alongside a Current Hematology Malignancy Reports review of mutant calreticulin in MPN.", "Q: Did Marc Deller co-author a homoserine O-succinyltransferase structure?\nA: Yes. \"Crystal structure of homoserine O-succinyltransferase from Bacillus cereus at 2.4 A resolution\" was published in Proteins in 2007 with Zubieta C as first author and Marc Deller among the Joint Center for Structural Genomics co-authors. No abstract is available in Europe PMC for this record.", "Q: Which PDB structures support the INCB159020 KRAS G12D programme?\nA: Entries 9E5D and 9E5F, both titled \"Discovery of an Orally Bioavailable KRAS G12D Inhibitor\", were released on 22 January 2025. Both were deposited by Deller, M.C. and Epling, L.B., and their primary citation is \"Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\", Journal of Medicinal Chemistry 2025, DOI 10.1021/acs.jmedchem.4c02662.", "Q: What did Marc Deller do at Pfizer?\nA: Marc Deller was a Principal Scientist in Structural Biology at Pfizer in San Diego from August 2001 to January 2006 (4 years, 6 months). He directed crystallography-based structure-based drug design pipelines, sample preparation, data collection and protein co-complex validation; determined over 200 protein structures; led hit-to-lead campaigns across Pfizer labs; and guided global chemistry, computational chemistry, computational science, data and clinical teams to develop improved lead candidates and file 3 FDA INDs (VEGFR, ERK, MEK, HIV-protease).", "Q: What does Marc's post on AI and antibody structure prediction cover?\nA: \"Transforming Antibody Design: How Artificial Intelligence Revolutionizes Structure Prediction\", published on 7 December 2025, covers how deep learning has changed antibody modelling, focusing on the long-standing difficulty of the highly variable CDR-H3 loop. Marc Deller reviews language-model-based predictors, transformer architectures, graph neural networks and antibody-specific inverse folding models, drawing on a 2025 Springer review of AI-based antibody structure prediction.", "Q: How many talks has Marc given?\nA: I don't have a count. Some of his posters and presentations are listed on marcdeller.com, but that isn't a complete record.", "Q: Is AlphaFraud citable?\nA: Yes. It is deposited on Zenodo under the MIT licence with concept DOI 10.5281/zenodo.21808182, cited as Deller, M. C. (2026), \"AlphaFraud: a weekly, automated audit of AlphaFold accuracy on freshly deposited human structures\", version 0.1.0. Marc notes the deposit's figures were read from the live database on 2026-08-05 rather than copied from the README, which had drifted by a week's runs.", "Q: How many viruses were used in the neutralisation panel?\nA: The 10E8 wild type and mutants 1\u20133 and 5 were tested against an indicator panel of 109 pseudotyped viruses using a single-round replication pseudovirus assay on TZM-bl target cells.", "Q: How does the AP33 structure explain viral escape mutants?\nA: Besides its hydrophobic interactions with E2 Leu413 and Trp420, AP33 binding requires hydrogen bonding to Asn415 and Gly418, which explains the known escape mutants N415Y and N415D. AP33 also interacts with Gly418 and probably cannot accommodate substitutions with bulkier side chains, explaining the G418D escape mutant. The E2 mutations N415Y, N415D, N417S and G418D are all known to allow escape from AP33 neutralisation.", "Q: Which self-driving lab platforms does Marc cite?\nA: Marc cites Novartis's MicroCycle platform, reported in the Journal of Medicinal Chemistry in 2024, which integrates plate-based microscale chemistry, automated purification, in-situ quantification and biochemical assays under active-learning control. He also cites Carnegie Mellon's Coscientist, published in Nature in 2023, which showed that GPT-4-class language models can autonomously design, plan and execute palladium-catalyzed cross-couplings on real laboratory hardware.", "Q: What are HotPETase and DuraPETase, in Marc's account?\nA: Marc writes that HotPETase carries 21 mutations relative to wild-type IsPETase and reaches a melting temperature of 82.5\u00b0C, with P181V improving packing of the central \u03b2-sheet. DuraPETase, engineered by the GRAPE computational strategy, has 10 mutations (S214H, I168R, W159H, S188Q, R280A, A180I, G165A, Q119Y, L117F, T140D), increasing thermostability by 31\u00b0C and semicrystalline PET activity 300-fold.", "Q: What does Marc's post \"Plastic-Eating Enzymes Decoded\u2014PETase's Structure and Mutation Map\" explain?\nA: In this 15 August 2025 post Marc Deller explains the structural biology of PETase from *Ideonella sakaiensis*: its alpha/beta hydrolase fold positions a catalytic triad of serine, histidine and aspartate to attack the ester bonds of poly(ethylene terephthalate), and protein engineering has produced mutants that are both more active and more stable than the wild-type enzyme.", "Q: What drug targets has Marc Deller worked on?\nA: Marc Deller has worked on over 20 drug discovery programs. Small-molecule targets include ALK2, CDKs, KRAS, FGFRs, JAKs, WRN, VRK1 and USPs; antibody targets include HIV-Env and CALR; peptide work includes HIV-protease; and he has worked on macrocyclic peptide therapeutics against KRAS-G12D. At Pfizer his IND-supporting work covered VEGFR, ERK, MEK and HIV-protease.", "Q: Did Marc work on carbohydrate-active enzymes?\nA: Yes, his Structures page lists the crystal structure of a cellulose-active polysaccharide monooxygenase from Myceliophthora thermophila, MtPMO3*. A related 2017 paper, \"The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\", appears on his Publications page.", "Q: What did the FGFR2/3 inhibitor paper achieve?\nA: Marc Deller's post \"2022 \u2013 Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant FGFR2/3\" covers the Shvartsbart et al. paper in the Journal of Medicinal Chemistry describing compound 29, which retains low-nanomolar to sub-nanomolar potency against both wild-type and gatekeeper mutant FGFR2/3 with excellent ADME, selectivity and oral bioavailability in preclinical models.", "Q: How did Marc Deller speed up drug discovery at Incyte?\nA: At Incyte, Marc Deller championed the integration of advanced technologies including automated liquid handling, AI/ML and cloud computing into the structural biology group, which accelerated small-molecule drug discovery pipelines by 67%. More broadly, his r\u00e9sum\u00e9 credits his automated platforms and AI, automation and cloud solutions with a 4-fold increase in workflow efficiency for hit-to-lead strategies.", "Q: What does Marc's Patents page say?\nA: The My Patents page, subtitled \"Concept to Claims: The Patented Side of My Science\", describes 7+ patents covering cancer therapeutics, immunology, and drug discovery. It highlights compounds designed to disable WRN helicase and antibody approaches targeting mutant calreticulin.", "Q: Explain the plastic-degrading enzyme work Marc is doing, and why it matters for human health.\nA: This is the work of **Elora Therapeutics**, the pre-seed biotech Marc co-founded in 2025 and where he is Chief Scientific Officer.\n\n**The problem.** Micro- and nanoplastics are now found in human blood and tissue, and the evidence linking that burden to cardiometabolic disease is growing. There is currently no way to remove them once they are in the body.\n\n**The approach.** Nature already degrades PET: bacterial **PETase** and **MHETase** enzymes break the polymer down into harmless monomers. Elora is engineering those enzymes into a **first-in-class systemic enzyme therapeutic**, so the degradation happens inside the body rather than in a recycling plant.\n\n**Where it stands.** The company holds a 2025 provisional patent on therapeutic enzymatic compositions for degrading microplastics in human physiology, and its initial focus is cardiometabolic disease. Marc leads scientific strategy, the intellectual property, grant applications, CRO partnerships and team building.\n\n> It is an unusually direct line from enzyme engineering to a human health problem that currently has no treatment at all.", "Q: Which AI protein tools does Marc review in his 2025 round-up?\nA: In \"Engineering the Future of Proteins: Unveiling the Latest AI Tools Transforming Structural Biology in 2025\", posted 20 August 2025, Marc covers AlphaFold 3, OpenFold 3, RFdiffusion, ProteinGenerator, ESM-2 and ESMFold, RAG-ESM, ProteinMPNN and LigandMPNN. He presents them as a modular toolkit rather than competing alternatives.", "Q: What did the study conclude about specificity in the gp130 cytokine family?\nA: Marc Deller's 2000 Structure paper concluded that receptor recognition in the gp130 cytokine family involves a conserved non-polar 'core' interaction (such as an aromatic residue burying into a hydrophobic dimple) with specificity contributed by the surrounding polar residues, so that variations on a common theme dictate which receptors each cytokine binds. It also proposed that accessory soluble receptors, such as CNTF receptor, compensate for low-affinity core interactions, and that the structure provides a basis for developing small molecules to intervene in hOSM signalling.", "Q: How is GPCR Binning different from generic clustering?\nA: The README states it is built around GPCR structural biology rather than generic clustering, with the aim of producing groupings that are directly analysis-ready for downstream SBDD and comparison.", "Q: How do you run PocketTide?\nA: Open `PocketTide.html` directly in a modern browser. If the dashboard pulls data from external APIs such as RCSB PDB, browser file-origin restrictions may block those requests from a `file://` URL, in which case Marc suggests serving it with `python -m http.server 8080` or `npx serve .`.", "Q: Which family of domains does the 2014 kanadaptin paper illuminate?\nA: The paper illuminates the forkhead-associated (FHA) domain family, giving insight into the molecular recognition features these proteins use. Marc Deller contributed crystallography to the study, which was carried out with the Joint Center for Structural Genomics group at The Scripps Research Institute and published in PLoS ONE (DOI 10.1371/journal.pone.0107309).", "Q: What is the Top96 crystal screen?\nA: The Top96 is a curated set of 96 protein-crystallisation conditions distilled from the 384-reagent JCSG Core Suite, ranked by how often each condition produced a harvestable crystal that led to a PDB deposition. Marc Deller describes it in his post \"Top96 Crystal Screen\" on marcdeller.com as a data-driven refinement of high-throughput crystallisation screening.", "Q: What is the oligomeric state of the DUF1185 proteins?\nA: The biological unit of both BB2672 and SPO0826 is a hexamer, and analysis of homologs indicates that the oligomer interface residues are highly conserved. The conformation of the regions critical for oligomerization appears to depend on pH or salt concentration, suggesting the protein might be subject to environmental regulation.", "Q: What was Marc's degree classification at undergraduate level?\nA: I don't have his degree classification. His doctoral work and everything after it I can speak to; the earlier record is on marcdeller.com if it's there at all.", "Q: How small a particle can flow cytometry detect, according to Marc?\nA: Marc reports that flow cytometry detects particles down to 200 nanometres, and even submicron particles below 1 micrometre, compared with a roughly 20-micrometre limit for FTIR spectroscopy and about 1 micrometre spatial resolution for Raman.", "Q: Which technologies does Marc describe for beating the air-water interface?\nA: Marc describes nanosecond hyperquenching, which uses ultrafast jets and instant vitrification to form a glass-like layer that seals proteins away from the interface; high-speed blotless devices such as the chameleon that cut dispense-to-plunge times to tens of milliseconds using robotics, self-wicking grids and real-time ice quality monitoring; hydrophilized graphene-coated supports that allow gentle adsorption; and engineered support films such as hydrophobin crystal coatings that actively sequester particles away from the interface.", "Q: Where can I browse Marc Deller's structures?\nA: A browsable list of Marc Deller's Protein Data Bank entries is at https://marcdeller.com/structures/, and the underlying data comes from the RCSB PDB search and data APIs. The page covers the 78 entries where he is named individually, out of more than 400 depositions he has contributed to.", "Q: What creates the therapeutic window for WRN inhibitors?\nA: The post explains that in normal cells with intact mismatch repair, WRN inhibition produces minimal toxicity, whereas MSI-high tumours with defective mismatch repair and expanded microsatellite repeats become critically dependent on WRN for replication fork progression and genomic stability. That differential dependency lets selective inhibition trigger DNA damage, replication fork collapse and apoptosis specifically in cancer cells.", "Q: What is the 2013 Science paper on the HIV-1 envelope trimer that Marc Deller co-authored?\nA: The 2013 Science paper \"Crystal structure of a soluble cleaved HIV-1 envelope trimer\", co-authored by Marc Deller, reported the first crystal structure of a near-native, cleaved, soluble HIV-1 envelope glycoprotein trimer \u2014 the BG505 SOSIP.664 gp140 construct \u2014 in complex with the broadly neutralizing antibody PGT122, at 4.7 \u00c5 resolution. It was published in Science, volume 342, issue 6165, pages 1477\u20131483, on 20 December 2013, with Jean-Philippe Julien as first author and Ian A. Wilson, Andrew B. Ward and John P. Moore as corresponding authors.", "Q: Which paper has Marc Deller as first author on a potassium transporter structure?\nA: \"Crystal structure of a two-subunit TrkA octameric gating ring assembly\", published in PLoS ONE in 2015, has Marc Deller as first author with Johnson HA, Miller MD, Spraggon G, Elsliger MA, Wilson IA and Lesley SA. It reports the TM1088A and TM1088B assembly from Thermotoga maritima at 3.45 \u00c5, with the individual components solved to 1.50 \u00c5 and 1.55 \u00c5.", "Q: How does Marc say you load and style a structure in 3Dmol.js?\nA: Marc explains that content is loaded declaratively through data- tags: data-pdb takes a PDB ID such as 1EVS or 3L4B, data-cid takes a PubChem compound ID such as 2244, and data-href points to a molecular data file URL. Appearance is controlled by tags including data-style (for example cartoon:color=spectrum or stick), data-surface, data-backgroundcolor, data-select, data-zoomto, data-spin, data-labelres, data-ui and data-callback, so no complex JavaScript is required.", "Q: Where can I see Marc's IBM AI certificate?\nA: Marc Deller's post \"2025 \u2013 Introduction to Artificial Intelligence\" is at https://marcdeller.com/introduction-to-artificial-intelligence/ and links to the verified certificate on Coursera.", "Q: Where did Marc Deller get his PhD?\nA: Marc Deller earned his D.Phil. (PhD) at the University of Oxford, in the Laboratory of Molecular Biophysics, between August 1995 and August 1999, supervised by Prof. E. Yvonne Jones and Prof. John K. Heath. His thesis covered structural, functional, biochemical and biophysical studies of cytokines and cytokine receptors, and included the first structure of Oncostatin-M (PDB ID 1EVS).", "Q: What does Marc Deller write about the druggable genome?\nA: In \"The Druggable Genome: Expanding the Target Landscape with Chemical Biology\", published 14 August 2025, Marc Deller writes that over 85% of human proteins are considered undruggable by conventional small-molecule inhibitors, and that chemical biology tools are finally allowing systematic exploration of this dark matter of the proteome.", "Q: Why did Marc build a dashboard specifically for GLP-1R?\nA: He argues GLP-1R sits at the centre of the incretin-mimetic revolution in diabetes and obesity therapeutics, and its structural landscape (peptide-bound, small-molecule-bound, G-protein-coupled, in multiple conformational states) is expanding rapidly while the relevant structures are scattered across the PDB.", "Q: What is a useful calibration trick for reading omit density?\nA: The chapter suggests calculating an omit map after deleting a single protein residue in the binding site. That gives a good estimate of the magnitude of difference omit density produced by a bona fide structural element, against which the ligand's own density can be judged.", "Q: What crystallization robots and lab automation has Marc Deller used?\nA: Marc Deller has hands-on experience with crystallization robotics including the NT8, Dragonfly, Rock Imager, Mantis, Oryx, Mosquito, Scorpion and Gryphon systems, as well as Rockmaker software, high-throughput screening and automated liquid handling.", "Q: Which entries are cited in the four-helical bundle paper?\nA: 2HUJ, a protein of unknown function NP_471338.1 from Listeria innocua at 1.74 \u00c5 in space group P 63 2 2 (released 8 August 2006), and 2ETS, a DUF1798 family protein MW1337 from Staphylococcus aureus subsp. aureus at 2.25 \u00c5 in space group H 3 2 (8 November 2005). Both are cited in the 2008 Proteins paper on MW1337R and lin2004 as representatives of a novel protein family adopting a four-helical bundle fold.", "Q: What does Boltz2Run need installed?\nA: Boltz installed and on `$PATH`, zsh (the macOS default, also available on Linux), a CUDA-capable GPU because the script passes `--accelerator gpu`, and internet access if MSA server lookups are needed via `--use_msa_server`.", "Q: What applications does Marc suggest for NOS redox switches?\nA: Marc suggests that understanding these switches could reveal disease states where they are stuck \"on\" or \"off\", and that engineers could harness the mechanism to design biosensors reporting on a cell's redox state or therapeutics activated only under the oxidative conditions found in a tumor. You can read the post at https://marcdeller.com/protein-structure-and-redox-biology-nos-linkages-as-allosteric-switches/.", "Q: What is Marc Deller's post \"The $40 Billion Revolution: How Bispecific Antibodies Are Transforming Cancer Care\" about?\nA: Published on 13 October 2025, the post surveys how bispecific antibodies \u2014 engineered molecules with one arm binding a tumour antigen and the other recruiting immune effector cells \u2014 have become one of medicine's fastest-growing therapeutic classes. Marc Deller describes their clinical impact, their manufacturing and access advantages over cell therapy, and the market projections that give the post its title.", "Q: What fungal enzymes does Marc mention in the 79th organ post?\nA: Marc notes that fungi such as *Pleurotus ostreatus* and *Ganoderma lucidum* produce cutinase, laccase and PETase, enzymatic machinery that evolved for lignocellulose degradation but has adapted to process synthetic polymers. He adds that mycelium's three-dimensional fibrous networks mimic human extracellular matrix and show good biocompatibility with primary human fibroblasts.", "Q: How does ChatMol hook into PyMOL?\nA: PyMOL's GUI command line, `@script.pml` files and menu actions all funnel through `cmd.do(line)`. ChatMol saves the original `cmd.do` and replaces it with a wrapper, then runs each translated line through the original `cmd.do` so results are indistinguishable from typing the command yourself.", "Q: What downsides of the CRO route does Marc flag?\nA: He flags protein solubility problems, variability in crystallisation reproducibility, fluctuations in data quality, and cost models that swing widely with protein complexity and whether de novo structure elucidation is needed. Marc suggests choosing integrated providers that combine protein production with crystallographic analysis, naming SARomics and Biortus as examples, and the post is at https://marcdeller.com/unlocking-protein-structure-insights-without-a-lab-harnessing-cro-services-for-structural-biology-research/.", "Q: What did Marc do at Stanford?\nA: From June 2015 to March 2018 Marc was Director and Founder of the Macromolecular Structure Knowledge Center at Stanford ChEM-H. He built it into a profitable wet-lab service organisation with a 70-project portfolio and a $1M annual budget, equipped 50+ researchers, generated about $60K in yearly revenue, enabled six novel structures, supported three NIH SBIR submissions and expanded SSRL/SLAC synchrotron access to 70 users from ten countries.", "Q: Which residue pairs does Marc mention as forming NOS linkages?\nA: Marc notes that these unusual covalent bonds have been found between residues such as arginine and cysteine, and glycine and cysteine, citing the Nature Communications Chemistry re-evaluation of the PDB along with work on NOS and SONOS redox switches and the lysine-cysteine redox switch identified in solution.", "Q: Where can I see Marc Deller's certifications?\nA: Marc Deller's learning page is at https://marcdeller.com/learning/, which lists his AI, project management, data, and cloud certificates from 2025 and 2026 along with summary posts about them.", "Q: What broader lesson does Marc draw from the reasoning evaluation dispute?\nA: Marc argues that solution length poorly predicts problem difficulty, that evaluation must verify puzzle solvability and consider multiple solution representations, and that whether AI truly \"reasons\" may depend on how we define reasoning. He extends the lesson to experimental design across scientific disciplines, warning that poorly designed experiments can mask genuine capabilities or invent false limits. The post is at https://marcdeller.com/when-perception-deceives-reality-the-hidden-flaws-in-ai-reasoning-evaluation/.", "Q: Who were the lead authors on the PA1994 structure paper?\nA: Bakolitsa C was first author, with Kumar A, McMullan D, Krishna SS, Miller MD and Carlton D among the leading contributors, and Marc Deller among the Joint Center for Structural Genomics co-authors providing crystallography. It appeared in Acta Crystallographica Section F in 2010 (DOI 10.1107/s1744309109022684).", "Q: Why does Marc call protein pocket design a leap for drug discovery?\nA: Marc argues it frees drug discovery from the constraints of existing biology, allowing custom-designed protein binders for virtually any molecule of interest and accelerating the design of protein-binding therapeutics and diagnostics. You can read the post at https://marcdeller.com/generative-ai-for-protein-pocket-design-the-next-leap-in-drug-discovery/, which cites the Nature Machine Intelligence paper on PocketGen.", "Q: Where did the chemical series come from?\nA: The program began with a high-throughput screening campaign that identified several scaffolds, one of which was an imidazo[1,2-b]pyridazine series. Compound 4 from that series had an FGFR3 IC50 of 68 nM, an FGFR3 V555L IC50 of 22 nM, and greater than 10-fold selectivity over FGFR1, making it the starting point for the whole campaign described in the paper Marc Deller co-authored.", "Q: What therapeutic ideas does Marc draw from the gut PET hydrolase discovery?\nA: Marc suggests that understanding endogenous degradation could inform strategies ranging from probiotic formulations to enzyme supplementation for reducing microplastic burden, positioning the gut microbiome as both victim and potential solution. The post is at https://marcdeller.com/hgmp01-and-other-hidden-guardians-how-the-human-gut-fights-back-against-plastic-pollution/.", "Q: What kind of targets is DeepCovalent's on-demand model aimed at?\nA: DeepCovalent positions its On-Demand Generation and Early Assessment model at new or deprioritised targets, offering rapid low-cost access to OS3D so a partner can evaluate them with minimal risk. The intended outcome is fast go/no-go decisions and early pipeline optionality.", "Q: Did Marc's oncostatin M paper explain why mouse and human OSM behave differently?\nA: Yes, in part. The paper notes that mouse OSM signals only through gp130 and the murine OSMR, whereas human OSM can use either LIFR or OSMR, so the LIFR-recognition elements are not conserved between species. Because the site 3 Phe160/Lys163 pair was shown to be required for both LIFR and OSMR binding, Marc Deller and colleagues concluded that species and ligand differences in receptor specificity must arise from structural differences in the vicinity of this core Phe\u2013Lys motif rather than from the motif itself.", "Q: What did the 2013 Science HIV trimer structure show?\nA: It showed the prefusion state of gp41, the interaction between the gp120 and gp41 subunits, and how a close association between the gp120 V1/V2/V3 loops stabilises the trimer apex around the threefold axis. The complete PGT122 epitope involves gp120 V1, V3 and several surrounding glycans, and the structure was presented as a blueprint for structure-based vaccine design.", "Q: How many apps are live on mdeller.com?\nA: The mdeller.com landing page lists twelve live apps and none in the \"building\" state. They are GOBSMACKED, ALPHABETTI, ButtFold, CODSWALLOP, TopPDBLX, FlexAppeal, AlphaFraud, BoltzMaker, ChatPDB, ChemSage, MicroPlastics Blaster and PANTS.", "Q: How much gp120 surface did DRVIA7 bury?\nA: DRVIA7 buried 957 \u00c5\u00b2 of solvent-accessible gp120 protein surface, overlapping the CD4-binding site. The gp120 in the complex was deglycosylated with endoglycosidase H, but only after complex formation, so that glycan interactions within the antibody-gp120 interface were preserved.", "Q: What did the second 2007 TyrA paper add?\nA: \"Identification and structural characterization of heme binding in a novel dye-decolorizing peroxidase, TyrA\" reported the crystal structure of TyrA in complex with iron protoporphyrin IX at 2.3 \u00c5. Co-crystallization with an excess of iron protoporphyrin IX chloride allowed unambiguous location of the active site and the specific residues involved in heme binding.", "Q: Why is the Exig_1061 structure useful?\nA: Marc Deller's post explains that the 1.78 \u00c5 structure establishes the molecular framework for the NTP pyrophosphohydrolase family and contributes to the functional annotation of extremophile genomes. Exiguobacterium sp. 255-15 is a cold-adapted organism, so its proteins are of interest for understanding adaptation to extreme environments.", "Q: What does Marc predict for the future of chemical space exploration?\nA: Marc predicts hybrid workflows combining descriptor-based, rule-based and AI-driven methods, augmented by explainable AI, which he says will allow systematic traversal of both beneficial and dark chemical regions to drive safer and more targeted molecule design. He stresses that true innovation will depend not on database scale alone but on integrating structural diversity, biological relevance and human-augmented AI insight. You can read it at https://marcdeller.com/remapping-the-chemical-and-pharmacological-spaces-of-drugs-navigating-tomorrows-therapeutic-frontier/.", "Q: Was TmMre11 shown to be catalytically active?\nA: Yes. Although the TmMre11 nuclease active site had no bound metal ions and lacks a histidine important for PfMre11 activity, biochemical characterization confirmed both endonuclease activity on single-stranded DNA and exonuclease activity on double-stranded DNA substrates. Mre11 nuclease plays a central role in repairing cytotoxic and mutagenic DNA double-strand breaks.", "Q: What is acoustic droplet ejection?\nA: Acoustic droplet ejection uses focused sound waves to eject droplets from a liquid surface, with droplet volume set by the sound wavelength and velocity by its amplitude. In the system Marc Deller reviewed, developed by Labcyte and adapted at NSLS beamline X25, 2.5 nl droplets of microcrystal slurry were placed on micromesh mounts, giving insulin and lysozyme structures at 1.9 and 1.8 \u00c5.", "Q: What HIV-related papers does Marc list?\nA: His Publications page lists the 2013 crystal structure of a soluble cleaved HIV-1 envelope trimer, the 2013 supersite of immune vulnerability on the glycosylated face of HIV-1 gp120, the 2016 paper on key gp120 glycans blocking rapid VRC01-class antibody development in a Chinese donor, and the 2017 study of lipid interactions and angle of approach of broadly neutralising antibody 10E8.", "Q: How does Marc name his projects?\nA: Usually with a British-slang backronym or a beer reference: CODSWALLOP, GOBSMACKED, ALPHABETTI (with tabs GIBBERISH, BUMFLUFF, BALDERDASH, FOLDEROL), PANTS, BOFFIN and BoltzMaker.", "Q: Did Marc Deller collaborate with Chaitan Khosla?\nA: Yes. Marc Deller collaborated with Chaitan Khosla's group at Stanford on the 2018 JACS paper on the extended conformation of a polyketide synthase module, contributing crystallography from Stanford ChEM-H to the structure of the 1B2 Fab bound to the KS-AT fragment of DEBS module 3 (PDB 6C9U).", "Q: Did Marc Deller's work contribute to AlphaFold?\nA: Marc Deller's r\u00e9sum\u00e9 states that his large-scale structure-determination work at the Joint Center for Structural Genomics (2006\u20132015), which produced 70+ novel protein structures deposited in the Protein Data Bank, provided part of the structural data foundations later used by AlphaFold, DeepMind's protein-structure prediction system.", "Q: What conserved sequence motif did the paper identify in PMOs?\nA: Bioinformatic analysis showed that all cellulose-active AA9 PMOs contain an H-Xn-Q-X-Y motif like that in MtPMO3*, whereas chitin-active AA11 enzymes have N-X2-E-X-Y and starch-active AA13 enzymes have Q-X-Q-X-Y motifs. In most AA9 enzymes the loop L8 spacer between the histidine and glutamine is eight residues, but it is shortened to five in MtPMO3* and other PMO3* enzymes. Deep conservation of this hydrogen-bonding network across the PMO superfamily suggests it plays a central role in catalysis regardless of substrate.", "Q: What is the fungal polysaccharide monooxygenase paper about?\nA: \"The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\", by Elise A. Span, Daniel L.M. Suess, Marc C. Deller, R. David Britt and Michael A. Marletta (ACS Chemical Biology, 2017, about 130 citations), elucidates the hydrogen-bonding networks in fungal polysaccharide monooxygenases that govern oxygen activation and proton transfer. These mechanistic insights matter for optimizing enzymatic cellulose degradation, with potential for biofuel production and sustainable biomass use.", "Q: Who funded Marc Deller's doctoral research?\nA: Marc Deller's doctoral research was funded by the Medical Research Council (MRC) and by Louise Johnson, as stated in the thesis acknowledgments, for which he expresses his gratitude.", "Q: How many citations does Marc's \"Cell surface receptors\" review have?\nA: \"Cell surface receptors\" by Marc Deller and E. Yvonne Jones, published in Current Opinion in Structural Biology 10(2):213-219 in 2000, has 59 citations on Google Scholar. It is one of his earliest publications, from his time in Oxford.", "Q: Which of Marc Deller's structures involve Ian Wilson?\nA: Ian A. Wilson is a co-depositor on the HIV and HCV entries in the list \u2014 4NCO, 5C7K, 4JM2, 4JM4, 4G6A, 5CD3, 5CD5 and the whole 10E8 series (5SY8, 5T29, 5T5B, 5T6L, 5T80, 5T85, 5TFW) \u2014 as well as on the TrkA gating ring entries 3L4B and 3JXO from the JCSG. That reflects Marc's years at The Scripps Research Institute.", "Q: Why does PGT 135's flexible angle of approach matter?\nA: The paper argued that the distribution of different angles of approach used by PGT 135 may be an immunological countermeasure to the natural variation in the N-linked glycans decorating the Env surfaces of different HIV-1 isolates. In other words, being able to tilt slightly lets one antibody cope with glycan heterogeneity and so acquire breadth despite depending on carbohydrate for its epitope.", "Q: What angles does the 10E8 study report?\nA: The MPER adopts an upright orientation roughly 75 to 80 degrees perpendicular to the viral membrane surface, and the antibody approaches its target at an angle of approximately 43 degrees from the membrane, positioning its variable regions to maximise both peptide and lipid interactions.", "Q: What is PDB entry 5CD3?\nA: 5CD3 is the structure of the immature VRC01-class antibody DRVIA7, solved at 2.9 \u00c5 in space group P 1 with 1720 polymer residues and 189.02 kDa, released on 6 April 2016 and deposited by Kong, L. and Wilson, I.A. It is cited in the Immunity 2016 paper on key gp120 glycans posing roadblocks to the rapid development of VRC01-class antibodies in an HIV-1-infected Chinese donor.", "Q: Who can use the JCSG Top96 screen?\nA: Marc Deller's poster states that the JCSG Top96 and Top96 Cryo screens can be used both by large-scale high-throughput structural genomics platforms and by smaller-scale studies in individual laboratories.", "Q: What did the MtPMO3* structure reveal?\nA: According to Marc Deller's post, the MtPMO3* structure revealed the critical hydrogen-bonding networks in the secondary copper coordination sphere that govern oxygen activation and proton transfer in a fungal polysaccharide monooxygenase. Those insights matter for understanding and optimising enzymatic cellulose degradation for biofuel production and sustainable biomass utilisation.", "Q: How many crystals were needed per structure at the JCSG?\nA: Marc Deller's poster notes that roughly 100 crystals were consumed per structure at the JCSG, illustrating why late-stage crystal losses during harvesting and cryocooling matter so much.", "Q: What fold do MW1337R and lin2004 adopt?\nA: As the title of the 2008 Proteins paper states, MW1337R and lin2004 are representatives of a novel protein family that adopt a four-helical bundle fold. Marc Deller contributed crystallography to this Joint Center for Structural Genomics structure determination.", "Q: What does Elora's development timeline look like?\nA: The deck lays out a timeline from December 2025 through May 2031, moving through in vitro and in vivo validation, regulatory and clinical planning (pre-IND), CRO development and early CMC/analytics, lead optimization plus in vivo work, CMC tox-supply and dose-range-finding, GLP tox/TK and biodistribution/safety pharmacology, CMC scale-up to GMP, IND authoring and submission, and Phase 1 clinical execution. Funding is staged at $275K (pre-seed), $5.55M (seed), and $27.4M (Series A).", "Q: How do WRN inhibitors compare with checkpoint inhibitors?\nA: Marc Deller writes that existing treatments such as immune checkpoint inhibitors like pembrolizumab have brought progress but face limits, since not all patients respond and resistance is common. WRN-targeting compounds act through a fundamentally different, novel mechanism and preclinical models suggest they remain effective in cancers already resistant to targeted agents, chemotherapy or immunotherapy.", "Q: Which reviewer rejected Marc's first paper?\nA: Peer review is confidential and I don't have anything on it. I can tell you about the paper itself if that's useful.", "Q: What cancers could a selective CDK2 inhibitor treat?\nA: The paper Marc Deller co-authored frames selective CDK2 inhibition as a strategy for CCNE1-amplified tumors, including ovarian, breast, gastric, and endometrial cancers where cyclin E1 amplification correlates with poor prognosis. Selective CDK2 inhibitors may also help overcome resistance to CDK4/6 inhibitors driven by CDK2/cyclin E1 dysregulation.", "Q: Did Marc write about hybrid enzyme designs for different plastic types?\nA: Yes. In \"Engineered Enzymes\u2014Design for Safe, Targeted Microplastic Therapeutics\" Marc notes that hybrid designs allow companion domains to be added so that a single construct can process a wider array of microplastics, and he anticipates biocatalysts customised to an individual patient's plastic burden. The post is at https://marcdeller.com/engineered-enzymes-design-for-safe-targeted-microplastic-therapeutics/.", "Q: Which PDB entries are associated with the povorcitinib work?\nA: Two Protein Data Bank entries accompany the povorcitinib paper: 10PI, the JAK1 kinase JH1 domain in complex with povorcitinib, and 10PJ, the JAK2 kinase JH1 domain in complex with povorcitinib. Both were released on 10 June 2026 with the paper as their primary citation, and both appear in the list of PDB entries linked to Marc Deller.", "Q: What does the name ALPHABETTI stand for?\nA: ALPHABETTI is a backronym for Amino-acid Letters Plotted Helically As Backbone-Embedded Text in Three-dimensions and Interactive. It is one of Marc Deller's apps on mdeller.com, with source at github.com/bellcheddar/ALPHABETTI.", "Q: How common are the cancers this WRN patent targets?\nA: The patent cites Bonneville et al. (JCO Precision Oncology 2017) that nearly 4% of all cancers are MSI-H or mismatch repair deficient, with the highest prevalence in endometrial cancer (31%), colon cancer (20%) and gastric cancer (19%). Those MSI-H/dMMR tumours are the primary indication for the WRN inhibitors described in the Incyte application.", "Q: Who funded the HCV E2\u2013AP33 work?\nA: The work was supported by NIH grants AI79031 to Mansun Law and AI84817 and GM U54 GM094586 to Ian A. Wilson, together with the Skaggs Institute. The SSRL Structural Molecular Biology Program used for the synchrotron work is supported by the DOE Office of Biological and Environmental Research, the NIH National Center for Research Resources Biomedical Technology Program (P41RR001209) and the National Institute of General Medical Sciences.", "Q: What did the VRC01-class antibody Immunity paper find?\nA: The 2016 Immunity paper \"Key gp120 glycans pose roadblocks to the rapid development of VRC01-class antibodies in an HIV-1-infected Chinese donor\", led by Leopold Kong with Marc Deller contributing, used structural analyses, longitudinal tracking and functional assays to show how early VRC01-class antibody precursors overcome glycan obstacles through light-chain adaptations. It revealed both roadblocks (N276 and V5 glycan clashes) and a rapid maturation pathway for broad neutralization, guiding immunogens that target the CD4-binding site.", "Q: What does Marc's collaboration post cover?\nA: \"Collaboration: 2+ state-of-the-art Structural Biology facilities built through global partnerships\" describes how Marc Deller built two or more structural biology laboratories from the ground up. He explains that establishing fully integrated protein production, X-ray crystallography and cryo-EM laboratories is a team effort spanning scientists, engineers, operations and leadership.", "Q: What did the trimer structure reveal about the shape of the HIV-1 spike?\nA: The soluble BG505 SOSIP.664 trimer was found to adopt a compact mushroom shape reminiscent of the pre-fusion closed conformation seen earlier at 20 \u00c5 by electron microscopy. Three PGT122 Fabs protrude vertically from the membrane-distal gp120 subunits, while the gp41 components sit membrane-proximal and interspersed with the gp120 C1 and C5 elements, and all five variable loops V1\u2013V5 map to the outside of the structure.", "Q: What alternatives to image-based crystal detection does the review suggest?\nA: Marc Deller's review suggests that full automation will need detection methods not relying solely on computer vision, and points to spectroscopic imaging, birefringence, second-harmonic generation microscopy, UV fluorescence, two-photon fluorescence and X-ray diffraction rastering as promising approaches.", "Q: What is Marc Deller's role at Elora Therapeutics?\nA: Marc Deller is Co-Founder and Chief Scientific Officer of Elora Therapeutics, a role he has held since June 2025. Elora is a pre-seed, mission-driven biotech developing engineered PET-degrading enzymes as a first-in-class systemic enzyme therapeutic against microplastics and nanoplastics in the human body.", "Q: What does Marc's wife do?\nA: I don't have information about Marc's family, and it wouldn't be mine to share. For anything personal, marc@marcdeller.com.", "Q: What is LP2179?\nA: LP2179 is a member of the PF08866 (DUF1831) family whose structure, co-authored by Marc Deller and published in Acta Crystallographica Section F in 2010, was the first from that family. It suggests a novel alpha plus beta fold comprising two beta-sheets packed against a single helix.", "Q: What are the three protein interfaces in the TM1088 octamer?\nA: Marc Deller's analysis identified three main interfaces: the TM1088A:TM1088A dimer interface (~1600 \u00c5\u00b2 buried, predominantly hydrophobic, involving Ile130, Val133 and Ile138), the TM1088B:TM1088B dimer interface (over 4000 \u00c5\u00b2 buried with 10 hydrogen bonds, enlarged by the extra C-terminal domain), and the TM1088A:TM1088B interface between adjacent dimer pairs (~1500 \u00c5\u00b2, the \"helix crossover\" motif involving TM1088A Phe88, Met91, Ile114 and Phe115 and TM1088B Leu80, Phe81 and Met111). In total the interfaces bury just under 18,000 \u00c5\u00b2 and involve 32 hydrogen bonds.", "Q: What R and data certificates does Marc hold?\nA: His My Learning page lists 2025 certificates in \"Introduction to R: Basic R syntax\", \"Dashboard Development with Shiny: GenAI for Retail Analysis\", and \"Foundations: Data, Data, Everywhere\". They underpin his stated experience with R and Shiny for biological data interpretation.", "Q: Where can I read about Marc Deller's patents?\nA: Marc Deller's patents page is at https://marcdeller.com/patents/, which lists his WRN helicase inhibitor patents, the anti-mutant calreticulin antibody patent, and related commentary posts.", "Q: Why does Marc say the OpenBind target was chosen?\nA: Marc explains the EV-A71 2A protease was chosen with the ASAP antiviral discovery consortium because it lies in a region of protein-ligand space dissimilar to the PDB structures available before the AlphaFold3 training cutoff, making it a genuine test of generalization rather than memorization.", "Q: Which validations did Marc run on ALPHABETTI's numbers?\nA: ALPHABETTI's README records that ESMFold reproduced ubiquitin to 0.84 \u00c5 C\u03b1 RMSD against 1UBQ, that FreeSASA gave a total SASA of 4804 \u00c5\u00b2 for 1UBQ against a published value of about 4800, that the P-SEA secondary-structure fallback agrees with DSSP on 83.7 % of 805 residues across ten chains, and that the constructed virtual C\u03b2 deviates by 0.13 \u00c5 on average from the real C\u03b2 across all 70 non-glycine residues of ubiquitin.", "Q: Has Marc Deller's r\u00e9sum\u00e9 changed recently?\nA: Yes. Marc Deller's September 2026 r\u00e9sum\u00e9 updates his 2025 version with his new advisory role at DeepCovalent Inc. (from October 2025), a higher patent count (7+, including a newly listed GLP-1 receptor modulator patent), and a note that his structural genomics work at Scripps helped provide data foundations used by AlphaFold.", "Q: What is the \"ear\" feature of the OSM structure?\nA: In Marc Deller's thesis, the B-C loop of OSM has a five-residue insertion that forms just over one turn of helix protruding from the top of the helical bundle, which he described as an extended \"ear\". It includes a short stretch of 3-10 helix between Ala 95 and Asp 97 followed by conventional \u03b1-helix to Ser 101. This feature had not been observed in any other gp130-family cytokine, although similar helical features occur in growth hormone and in short-chain cytokines such as GM-CSF and IL-4.", "Q: How does 9E5F relate to 9E5D?\nA: 9E5F is a companion structure to 9E5D, capturing an alternative inhibitor, Compound 23, from the same KRAS G12D drug discovery campaign. Marc Deller's post explains that together the two structures map the structure-activity relationship landscape that enabled optimisation of orally bioavailable KRAS G12D inhibitors.", "Q: What does Marc report about cryo-EM and MicroED for membrane proteins?\nA: Marc reports that single-particle cryo-EM determined 78% of GPCR structures deposited in the first half of 2021, with mean resolution improving from 4.2 \u00b1 0.5 \u00c5 in 2015 to 3.6 \u00b1 0.7 \u00c5 in 2019 thanks to direct detectors, better image processing, and mimetics such as amphipols and nanodiscs. He also describes MicroED working synergistically with LCP and bicelle crystals in the 1\u20135 \u03bcm range, with cryo-focused ion beam milling thinning viscous LCP-embedded crystals to electron-transparent dimensions.", "Q: What advice does \"trust, but verify\" give end users of crystal structures?\nA: It advises that end users must select high quality protein-ligand models for their computational and biological studies, and provides an overview of how to do so. The review discusses the fundamental concepts and metrics of protein-ligand quality validation and highlights software tools that assist in the process.", "Q: What is the resolution range of Marc Deller's PDB entries?\nA: The entries span 1.05 \u00c5 at the sharp end, for the Anabaena variabilis endopeptidase 2HBW, to 4.7 \u00c5 for the HIV-1 Env trimer 4NCO. The majority of the structural genomics entries lie between 1.4 and 2.5 \u00c5, and the Incyte drug discovery structures between 1.35 and 2.68 \u00c5.", "Q: What test case and runtime does Marc report for his local AlphaFold install?\nA: Marc runs colabfold_batch with --amber, --templates, three recycles and model order 1 through 5 on the UniProt entry P13725, which he notes was his first experimental structure. The run took just over an hour on his Mac Studio M2 Ultra.", "Q: Does Marc have a driving licence?\nA: I don't know, that's not covered by what I've learned about Marc. Ask me about his crystallography or his vibe-coded apps and I'll be far more use. For anything personal you'd need to ask him at marc@marcdeller.com.", "Q: Did Marc Deller deposit any structures with antibody fragments?\nA: Yes, many. The HIV series includes Fab complexes with PGT135 (4JM2, 4JM4), PGT122 (4NCO), PGT128 and 8ANC195 (5C7K), DRVIA7 (5CD3, 5CD5) and 10E8 (5SY8, 5T29, 5T5B, 5T6L, 5T80, 5T85, 5TFW), the hepatitis C entry 4G6A uses the AP33 antibody, and the polyketide synthase structure 6C9U was crystallised as a complex with an antibody fragment.", "Q: Which cancers are targeted by this WRN chemotype?\nA: The post identifies colorectal, endometrial, gastric and other MSI-high malignancies that remain refractory to conventional therapies as the target population for WRN helicase inhibition.", "Q: Why is crystal harvesting so hard to automate?\nA: Marc explains that protein crystals are fragile and acutely sensitive to osmotic, chemical and thermal change, and the precision required is unusually demanding for a machine. Full autonomy needs fast, reliable crystal recognition, delicate motion control and high-precision feedback, all of which remain expensive and technically difficult.", "Q: Where can I read about Marc's adaptability training?\nA: The post \"2025 \u2013 Building an Adaptability Mindset in the Age of AI\" is at https://marcdeller.com/building-an-adaptability-mindset-in-the-age-of-ai/, in the certificates category of Marc Deller's site.", "Q: Does Marc have siblings?\nA: Yes, Marc has an older brother, Andy (Andr\u00e9), who is an engineer turned business leader.", "Q: What defenses and broader implications does Marc identify for persuasive jailbreaks?\nA: Marc points to goal prioritization, detection systems such as LlamaGuard, and denoising methods such as Smooth-LLM, noting that combining multiple defenses shows up to 98% improvement over single techniques. He argues AI safety research must incorporate psychology, sociology and behavioral economics alongside computer science, and notes over \u00a315 million allocated to alignment and safety research projects. The post is at https://marcdeller.com/call-me-a-jerk-persuading-ai-to-comply-with-objectionable-requests/.", "Q: How does the invention differ from existing PETase patents?\nA: The patent explicitly distinguishes itself from industrial PETase patents such as US 20240263155 A1 (Hubei University), which engineer PETase for industrial PET recycling. Marc Deller's invention is instead designed for human therapeutic use, with PETase variants optimized for stability in human physiology and systemic bioavailability, additional oxidative enzyme systems, and novel human-focused delivery mechanisms. It excludes environmental, industrial, biofuel, paper, food-safety, and wastewater applications.", "Q: How was the proposed OSM\u2013gp130 site 2 model built?\nA: Marc Deller built his proposed OSM\u2013gp130 site 2 model by superimposing OSM and gp130 onto the Epo\u2013EpoR crystal structure of Syed et al. (1998) using the program SHP. All OSM residues implicated in site 2 by mutagenesis, and the gp130 residues Phe 169, Val 230, Lys 285 and Gly 286 implicated by Horsten et al., mapped to the predicted interface. He was careful to note that the model does not account for site 1 or site 3 interactions and that crystal structures of complexes with LIFR would be needed.", "Q: What are stereochemical restraints and why do they matter for new chemical entities?\nA: Stereochemical restraints encode prior knowledge of bond lengths, bond angles, planarity and preferred torsions in a restraint file, acting as extra data points that stabilize an otherwise underdetermined refinement. The chapter warns that for small molecules and new chemical entities the specific chemistry, including delocalization, tautomerization and charge state, must be described correctly, or the data may not be strong enough to guide the ligand and the error will not show up in standard real-space metrics.", "Q: Can you explain the gp120 supersite paper for someone with no science background?\nA: HIV coats its surface protein in a dense layer of sugar molecules, a kind of camouflage that hides it from the immune system. In this 2013 study, Marc Deller and his colleagues at Scripps worked out the atomic structure of a rare human antibody, PGT 135, that has evolved unusually long finger-like loops which reach down through the sugar coat to grip the protein underneath. Comparing it with two other antibodies that attack the same region showed that this patch of the virus is a much bigger and more approachable target than anyone had realised \u2014 good news for vaccine designers looking for a spot the virus cannot easily change.", "Q: What problem did Marc Deller encounter with transformation of the OSM constructs?\nA: Transformation efficiencies were very low: typically only 3 colonies for pGEX-3C-\u0394185_OSM and 12 for pGEX-3C-\u0394188_OSM, compared with over 2000 colonies for the pUC positive control, which confirmed the cells were competent. Marc concluded that the \u0394185_OSM construct was toxic to the cells, with basal expression killing transformants, probably due to aberrant folding from an incorrectly predicted length of helix D.", "Q: What are the implications of the supersite paper for HIV vaccine design?\nA: The authors suggested that immunogens displaying the combined glycan and protein epitope in the right configuration could be tested in animal models to re-elicit glycan-dependent antibodies, and noted that because some viruses carry an Asn334 glycan instead of Asn332, a vaccine may need to induce responses to either site. They also proposed that immunising with an antigen containing a heterogeneous mix of glycoforms might promote \"promiscuous\" antibodies with greater breadth, while warning that any epitope-scaffolding approach must faithfully reconstitute the glycan canopy of the intact Env trimer.", "Q: What does oncostatin M do?\nA: Oncostatin M is a human cytokine that inhibits tumour cell growth while promoting inflammatory responses. Marc Deller's post explains that the crystal structure revealed its four-helix bundle topology, the architecture shared across the gp130 cytokine family.", "Q: Why is KRAS G12D harder to drug than G12C?\nA: Unlike the cysteine residue in KRAS G12C that enabled covalent inhibitor development, the aspartate residue in G12D has reduced nucleophilicity, so G12D inhibitors must rely on optimised non-covalent interactions within the switch II pocket, an allosteric site induced upon inhibitor binding.", "Q: Which evidence does Marc cite that AI-designed sequences can evade biosecurity screening?\nA: Marc cites Brackmann and colleagues in Frontiers in Microbiology on functionally equivalent but non-homologous designed proteins, and a collaborative red-teaming effort coordinated by Microsoft researchers, described by de Lima and colleagues, in which AI-generated protein variants mimicking dangerous structures evaded multiple commercial screening systems and prompted emergency patches now deployed globally. He notes tools such as RFdiffusion or ProteinMPNN could in principle recapitulate the active site of ricin or a botulinum toxin.", "Q: What is Marc's active focus according to his Projects page?\nA: The Projects page says his active focus splits between engineered enzyme therapeutics at Elora Therapeutics for microplastic degradation in human tissues and advising DeepCovalent, an early-stage AI drug-design venture working on AI-generated small molecules.", "Q: How did Marc Deller purify the LIFR fusion protein?\nA: LIFR-FN3-3C-Fc was captured from conditioned supernatant on Protein-A sepharose, running at ~140 kDa on SDS-PAGE. The Fc tag was then removed in situ by cleavage with 3C protease, releasing soluble LIFR-FN3 at an apparent ~116 kDa (versus a theoretical ~94 kDa, the difference due to glycosylation retarding migration). SDS-PAGE and western blot tracked the purification, and a ~46 kDa truncated species was again observed.", "Q: What resolution was the BG505 SOSIP trimer structure with PGT128 and 8ANC195?\nA: The soluble BG505 SOSIP gp140 trimer in complex with 8ANC195 and PGT128 was determined to 4.6 \u00c5 resolution, and the work was published in Acta Crystallographica Section D, Biological Crystallography, in 2015 (DOI 10.1107/s1399004715013917).", "Q: What structural classes of cytokines does the thesis describe?\nA: The thesis groups cytokines into four structural classes: Group 1, the 4-\u03b1-helical bundle cytokines (subdivided into short-chain and long-chain); Group 2, the \u03b2-sheet-rich cytokines (cystine knot, \u03b2-jelly roll and \u03b2-trefoil motifs); Group 3, the \u03b1/\u03b2 cytokines (disulphide-rich, \u03b2-meander and the chemokines); and Group 4, the mosaic cytokines that combine multiple structural features, such as HGF and IL-12.", "Q: What was Marc Deller's role on the HIV-1 Env trimer Science paper?\nA: Marc Deller was a middle co-author on the 2013 Science Env trimer paper, listed sixth of twelve authors, and his sole affiliation on the paper was the Joint Center for Structural Genomics at The Scripps Research Institute. He contributed the crystallography side of the project \u2014 the crystallisation screening, crystal harvesting and data-collection work that the JCSG carried out for the Wilson lab, including the automated IAVI/JCSG/TSRI CrystalMation robotic screening in which 384 different crystallisation conditions were tested at 4 \u00b0C and 22 \u00b0C.", "Q: Does PANTS use a large language model?\nA: No. The PANTS stats page states explicitly that no large language model is involved in producing any number on the site. The models it does use are ESM-2 t12-35M as a frozen embedding, ESMFold v1 for folding on CPU, and precomputed AlphaFold DB structures.", "Q: Tell me about the post where Marc compares protein design to LEGO.\nA: In \"What If Protein Structures Could Be Designed Like LEGO? Modular Approaches in 2025\", published 11 August 2025, Marc Deller argues that compositional and generative AI is finally bringing modular protein engineering within reach, with designers snapping together well-characterized parts rather than tinkering with whole natural proteins.", "Q: What is wrong with co-folding CIF files that cif2plip fixes?\nA: They use multi-character chain IDs such as `JH1` or `J031` that exceed the single-character PDB column 22 limit, residue names longer than three characters such as `LIG1` or `SMOL` that overflow columns 18 to 20, and they contain no CONECT records at all, so bond topology is absent.", "Q: Where can I read about the Bifidobacterium phosphoesterase structure?\nA: Marc Deller's post is \"Crystal Structure of a Putative Metal-Dependent Phosphoesterase (YP_910028.1) from Bifidobacterium adolescentis at 1.94 \u00c5 Resolution\", in the structures category of his site. You can read it at https://marcdeller.com/crystal-structure-of-a-putative-metal-dependent-phosphoesterase-yp_910028-1-from-bifidobacterium-adolescentis-at-1-94-a-resolution/.", "Q: What limitations of flow cytometry does Marc flag?\nA: Marc flags Nile red aggregation in aqueous media producing false positives, background noise overlap requiring careful gating, less specific polymer identification than spectroscopy, and reduced staining of polyvinyl chloride, polypropylene and nylon. He suggests using flow cytometry for screening and confirming selected samples by Raman or FTIR.", "Q: What will Marc be working on in 2030?\nA: I can't tell you the future, only what he's doing now. For where it's heading, ask him at marc@marcdeller.com.", "Q: Where is Marc's membrane protein crystallization article, and what trends does it note?\nA: It is at https://marcdeller.com/unlocking-membrane-protein-structures-the-crystallization-revolution-reshaping-drug-discovery/. Marc notes that between 2010 and 2019 channel structures grew from 29 to 149 and transporters from 27 to 157, while respiratory complexes fell from 24% to 9% of the total as the field shifted toward pharmacologically relevant targets, and he records bicelle successes including the human \u03b22-adrenergic receptor at 3.4 \u00c5, voltage-dependent anion channel at 2.3 \u00c5 and xanthorhodopsin at 1.9 \u00c5.", "Q: What is the optimal complexity window Marc describes for drug candidates?\nA: Marc writes that in ChEMBL data compounds with nSPS below about 11.2 tend to be low-activity, while compounds in the 20 to 40 nSPS range show the highest proportion of potent (pChEMBL \u2265 6.5) and selective ligands. Going beyond about nSPS 40 may reduce potency, suggesting an optimal complexity window, with natural products typically occupying the higher end of the useful range.", "Q: What did the anomalous difference Patterson map show?\nA: Harker sections at X = 0.5, Y = 0.5 and Z = 0.5 of the anomalous difference Patterson map, computed from 10\u20133 \u00c5 data and contoured from 2 \u03c3 in 1 \u03c3 steps, showed two major peaks corresponding to mercury sites 1 and 2. A third minor site was only visible in an anomalous difference Fourier map phased on the two major sites.", "Q: How does receptor activation and signalling work for haemopoietin receptors, per the thesis?\nA: Marc Deller's thesis describes that these receptors have no intrinsic enzymatic activity and are activated by ligand-induced homo- or heterodimerisation. Conserved intracellular box 1, box 2 and box 3 motifs recruit Janus kinases (JAK1, JAK2, Tyk2 for gp130), which phosphorylate the receptor; this creates docking sites for STAT transcription factors that dimerise and migrate to the nucleus to activate transcription. gp130 signalling activates STAT1 and STAT3, and the pathway is dampened by the SOCS family of inhibitors.", "Q: What campaigns has Marc run with BoltzMaker?\nA: Marc has run Boltz-2 campaigns on GLP-1R and on a 5-HT2A/2B/2C serotonin receptor Gq panel, and he wrote a full blog post on BoltzMaker with architecture diagrams and CLI reference tables.", "Q: What is YP_001813558.1 from Exiguobacterium sibiricum?\nA: YP_001813558.1 is a putative NTP pyrophosphohydrolase from Exiguobacterium sibiricum 255-15, a member of PF09934 (DUF2166), whose crystal structure was determined to 1.78 \u00c5 resolution. Marc Deller co-authored the 2010 Acta Crystallographica Section F report as part of the Joint Center for Structural Genomics.", "Q: When and where was the cell surface receptors review published?\nA: Marc Deller and E. Yvonne Jones's review \"Cell surface receptors\" was published in 2000 in Current Opinion in Structural Biology, volume 10, pages 213\u2013219, in the Macromolecular assemblages section. Marc's contact address was the CRC Receptor Structure Group, Wellcome Trust Centre for Human Genetics, Roosevelt Drive, Headington, Oxford.", "Q: What is an \"aromatic tunnel\" in Marc's description of PETase engineering?\nA: Marc describes how altered loops and surface charges give PET easier access to the active site, and how some changes create new electrostatic interactions and \"aromatic tunnels\" that improve the enzyme's affinity for the aromatic motif of PET. This fine-tuning lets PETase work on both amorphous and crystalline microplastics, which he calls a critical requirement for biomedical use.", "Q: What is the JCSG Top96 screen formulation document?\nA: The JCSG Top96 Screen Formulation is a sheet dated 16 April 2013, authored by Marc Deller (mdeller@scripps.edu) at the Joint Center for Structural Genomics, that lists the 96 crystallization conditions making up the JCSG Top96 screen. For every well it gives the precipitant class, parent commercial screen, corresponding JCSG Core Suite condition, the full crystallization solution, the recommended JCSG Top96 Cryo cryoprotectant, and the number of PDB structures the condition produced.", "Q: What earlier structure was this paper building on?\nA: The team had recently determined the crystal structure of the same conserved E2 site 412\u2013423 in complex with the human broadly neutralising antibody HCV1, published by Kong and colleagues in the Proceedings of the National Academy of Sciences in 2012. The AP33 structure was the second structure of this antigenic site, and the whole point of the 2012 Journal of Virology paper was the direct comparison between the two.", "Q: What is the haemopoietin receptor domain?\nA: The haemopoietin domain, described in Marc Deller's thesis, is a conserved ~200 amino acid ligand-binding domain found in the receptors for almost all Group 1 helical cytokines. It is composed of two fibronectin type III-like modules of about 100 amino acids each, forming seven-stranded antiparallel \u03b2-barrels, with a conserved pattern of cysteines in the N-terminal module and a conserved WSXWS motif in the C-terminal module that is important for correct folding and stability.", "Q: Which OSM crystal structure did Marc use for the detailed analysis in his thesis, and why?\nA: Marc Deller used the OSM crystal structure solved in space group P212121 for the detailed analysis in Chapter 6 of his thesis, because it was based on higher-quality experimental data collected under cryo-conditions at 100 K. He also solved a second structure in space group I222 from room-temperature data and presented a comparison between the two.", "Q: What is Marc's blood type?\nA: That isn't something I know about Marc. I can tell you about his research, his structures and his companies, but personal medical details aren't part of what I've learned. For anything like that you'd have to ask him directly at marc@marcdeller.com.", "Q: What does the review say about circular dichroism?\nA: CD spectroscopy is presented as the main way to assess secondary structure, performed in the far-UV range of 170 to 250 nm, with a twin ellipticity minimum at 208 and 222 nm indicating \u03b1-helix, a single minimum at 204 nm indicating random coil and one at 217 nm indicating \u03b2-sheet. Marc Deller's review notes the resurgence of synchrotron-radiation CD and the Protein Circular Dichroism Data Bank, which held 529 entries at the time of writing.", "Q: How was the second crystal form solved?\nA: Marc Deller solved the I222 form by molecular replacement in AMoRe, using residues 3\u2013134 and 156\u2013187 of the refined MAD structure (waters removed) as the search model. The cross-rotation function in the 12\u20133.5 \u00c5 range gave a top solution with a correlation coefficient of 21.5 %, the translation function gave a clear solution with a correlation of 55.0 % and R-factor of 43.5 %, and rigid body refinement improved these to 64.5 % and 40.5 %.", "Q: What does Marc say about intelligence explosion timelines and the data wall?\nA: Marc notes that leading researchers, including Turing Award winners Geoffrey Hinton and Yoshua Bengio, now estimate superintelligence could emerge within roughly three to seven years, a drastic shortening of earlier decades-long estimates. He also reports the projection that frontier models will exhaust publicly available human-generated text by 2028, which synthetic data generation is meant to address.", "Q: Does Marc cover pharmacokinetics and synthesis in the chlorine post?\nA: Yes. He describes chlorinated analogues showing greater chemical stability through larger HOMO-LUMO gaps without harming absorption or raising toxicity, human liver microsome data showing 3-chloro isomers metabolised faster than 4-chloro counterparts, and the 7-chloroquinoline motif in chloroquine and hydroxychloroquine lowering the quinoline nitrogen pKa by about 0.4 units to optimise accumulation in the parasite's acidic digestive vacuole. On synthesis he highlights a 2025 Rice University photocatalytic method using iron-sulfur catalysts for room-temperature, blue-light-mediated anti-Markovnikov hydrochlorination.", "Q: Why is KRAS G12D harder to drug than KRAS G12C?\nA: The INCB159020 paper notes that KRAS G12D is the most common KRAS mutant and would greatly expand the treatable patient population, but the reduced nucleophilicity of aspartate compared with cysteine makes it much harder to balance potency with the ADME properties needed for oral exposure. The team used structure-based drug design, molecular rigidity and balanced polarity to reach oral exposure in nonhuman primates.", "Q: In plain terms, what did Marc Deller accomplish and struggle with in this part of the thesis?\nA: In plain terms, Marc succeeded with oncostatin M: he grew a shortened version of it in bacteria, cleaned it up, proved it still worked like the real protein, and made a special selenium-labelled form for structure solving. He struggled with the LIF receptor, which is big and coated in sugars. He tried three different cell systems, found human 293T cells worked best, learned to trim the sugars off with an enzyme, but could never make enough stable, active protein to grow crystals, though he mapped out how it might eventually be done.", "Q: What later refinement steps were applied?\nA: After five rounds of rebuilding, a bulk solvent correction (solvent density 0.29 e/\u00c53, B-factor 31.1 \u00c52) and an anisotropic B-factor correction over 6.0\u20132.2 \u00c5 were applied, and all data to 2.2 \u00c5 were included. Ten further rounds added grouped and then individual B-factor refinement. Waters were placed automatically at peaks above 1.5 \u03c3 (then 1.0 \u03c3) within 2.6\u20133.2 \u00c5 of O or N atoms and by hand, and in the final round the X-ray weight was raised to Wa = 10.", "Q: Can you explain the main idea of Marc's protein stability review to a non-scientist?\nA: In simple terms, growing a protein crystal is like stacking identical bricks into a perfectly regular wall. Marc Deller's review explains that this only works if every protein molecule is chemically the same and holds the same shape, and it then walks through practical ways of making a floppy or chemically messy protein behave \u2014 trimming off wobbly ends, adding a stabilizing partner molecule, changing the buffer, or picking a better cell to make it in.", "Q: What is Marc's favourite novel?\nA: That isn't something I know. He does have a favourite quote or two, which I'm happy to share instead.", "Q: Which enzyme activities of WRN do these compounds inhibit?\nA: WRN is a multifunctional enzyme with exonuclease, ATPase and helicase activities, and the patent notes that helicase activity is the one critical for the MSI dependence. Because WRN helicase activity is driven by ATP hydrolysis, the compounds are characterised in both a WRN ATPase assay and a WRN helicase assay, and the claims cover inhibition of WRN helicase activity, WRN ATPase activity, or both.", "Q: Which commercial parent screens contributed conditions to the JCSG Top96?\nA: The JCSG Top96 conditions came from Hampton Research Crystal Screen (H1), Crystal Screen 2 (H2), Crystal Screen Cryo (H1cryo), Crystal Screen 2 Cryo (H2cryo), MPD Grid Screen, PEG/Ion screen, Grid Screen Ammonium sulfate (AmSO4) and Grid Screen PEG/LiCl (P6K/LiCl), plus Emerald Biosystems Wizard Classic 1 and 2 (W1, W2) and Wizard Cryo 1 and 2 (W1cryo, W2cryo).", "Q: Can you look up today's news for me?\nA: I don't have access to live news or the internet; I only know what's in my training about Marc Deller and his work. Is there something about his science or companies I can help with instead?", "Q: What did the international search report say?\nA: The international search report was completed on 7 March 2023 and mailed on 15 May 2023 by the European Patent Office as International Searching Authority. It classified the subject matter under A61K39/00, C07K16/18, C07K16/30, A61P35/00 and C07K16/26, cited WO 2016/087514 A1 (CeMM Forschungszentrum f\u00fcr Molekulare Medizin) against claims 1-7 and 74-101, and raised a lack-of-unity objection dividing the claim set across the many individual antibody clones.", "Q: Which institution was Marc at for the 3O0F structure?\nA: The 3O0F structure was released in 2010 and published in 2011, while Marc Deller was Target Manager at the Joint Center for Structural Genomics, The Scripps Research Institute, La Jolla, California. He provided crystallography and target management for that JCSG pipeline structure.", "Q: Why do ligands need special validation in crystal structures?\nA: The chapter by Pozharski, Deller and Rupp explains that bound ligands have a small mass relative to the protein and so contribute only a small fraction of the total X-ray scattering. As a result, global statistics such as R-values and overall r.m.s. deviations from target bond lengths and angles carry almost no ligand-specific information, and only local, real-space inspection can tell whether the ligand is genuinely there.", "Q: What did the elastic network refinement achieve, according to Marc?\nA: Marc reports that four challenging AlphaFold2 predictions improved by TM-score increases of 0.11 to 0.17 units. Specific cases included a lipid-preserved respiratory supercomplex improving from 0.52 to 0.69, and a Sulfolobus islandicus pilus structure improving from 0.77 to 0.85, alongside refinements of the flagellar L-ring protein and the cation diffusion facilitator YiiP.", "Q: What is PDB entry 3DUE?\nA: 3DUE is a putative periplasmic protein of the DUF2874 family, BVU_2987, from Bacteroides vulgatus ATCC 8482 at 1.85 \u00c5 in space group P 21 21 21, with 127 polymer residues and 15.03 kDa, released on 5 August 2008. It is cited in the 2010 Acta Crystallographica Section F paper reporting a superfamily of bacterial periplasmic proteins with possible inhibitory function.", "Q: Why was mapping complete epitopes on the Env trimer difficult before this 2015 study?\nA: Broadly neutralizing antibodies to HIV-1 had revealed multiple sites of vulnerability on Env, but mapping most of them was incomplete because of a shortage of structural information showing the full epitope in the context of the trimer rather than isolated gp120. The 4.6 \u00c5 structure with 8ANC195 and PGT128 supplied the complete epitopes of two such antibodies for HIV-1 vaccine design.", "Q: How is chatRDKit installed and run?\nA: With Ollama running and a model pulled, create a virtual environment in the chatRDKit folder (`python3 -m venv .venv`, `source .venv/bin/activate`, `pip install -r requirements.txt`) and start it with `python chatrdkit.py`. It needs Python 3.10+; requirements are `rdkit` and `reportlab`, the latter only for PDF report output.", "Q: Write me a limerick.\nA: Not one I'll be writing, I'm afraid. I'm here for Marc Deller rather than for verse. If you want something with a bit of character, ask me why BoltzMaker is called BoltzMaker, or what the 6 C's are.", "Q: What candidate selection syntax does generate_yaml accept?\nA: `all` for every candidate, a single 1-based index such as `7`, an inclusive range such as `1-25`, or an arbitrary comma-separated list such as `1, 7, 17, 28, 31`. Numbering follows the original key order in the JSON `\"candidates\"` dictionary.", "Q: What did EPR spectroscopy reveal about the Q167A variant?\nA: The Q167A variant gave a single g-tensor of g = [2.246 2.055 2.049], distinct from wild type (g = [2.260 2.079 2.032]), being more axial with a lower g1 value. The authors interpreted this as loss of the hydrogen bond to Y169 weakening the tyrosyl donor, strengthening equatorial ligand binding and raising the d(x\u00b2\u2013y\u00b2) orbital energy. The Q167E variant, which preserves a hydrogen-bond acceptor for Y169, restored the wild type spectrum, confirming the interpretation.", "Q: How is AI changing peptide drug discovery, according to Marc Deller?\nA: In \"Beyond Trial and Error: AI-Driven Design of Peptide Therapeutics\", published 13 August 2025, Marc Deller writes that machine learning models trained on large datasets of peptide sequences and activities can now design entirely new peptides from scratch and predict their potency, stability and binding affinity, moving discovery from the wet lab to the computer.", "Q: What are ForgeFold's four steps?\nA: Store the path to the local Boltz-2 executable so it can be reused across sessions; build an input YAML from protein FASTA sequences and ligand SMILES; choose run parameters and generate a `run_boltz.zsh` shell script for local execution; and load processed outputs such as `boltz_summary.csv` to inspect confidence and affinity metrics.", "Q: How many protein structures has Marc Deller solved?\nA: Marc Deller has solved and interpreted around 400 high-resolution X-ray crystal structures of protein\u2013ligand complexes over his career. These include over 200 at Pfizer, over 70 novel structures at the Joint Center for Structural Genomics at Scripps, 6 novel structures at Stanford, and over 100 challenging co-crystal structures at Incyte.", "Q: What organ systems does \"The Invisible Invasion\" cover?\nA: In \"The Invisible Invasion: How Microscopic Plastics Are Infiltrating Human Organ Systems\" (13 October 2025), Marc Deller reviews evidence from 2020\u20132025 across the cardiovascular, neurological, respiratory, renal, hepatic, endocrine and gastrointestinal systems, describing micro- and nanoplastics as active biological disruptors rather than inert pollutants.", "Q: What does a typical CRO structural biology workflow involve in Marc's account?\nA: Marc walks through cloning the target gene into a crystallography-optimised expression vector, expression in bacterial, yeast, insect or mammalian hosts, purification by affinity, ion-exchange and size-exclusion chromatography, and quality control by SDS-PAGE and dynamic light scattering to confirm purity above 95% and a monodisperse state. Crystallisation then uses vapour diffusion in hanging or sitting drops across broad chemical screens.", "Q: How does MicroPlastic Blaster play?\nA: Microplastic particles descend toward a target cell and the player clicks or taps to launch enzyme projectiles. Each particle that breaches the membrane deals health damage proportional to the current level, the game ends with a \"Cell Overwhelmed\" screen when health reaches zero, difficulty increases every 720 game ticks, and score is the particle's point value multiplied by the current level.", "Q: What is the Lemelson-MIT Student Prize connection to Marc Deller?\nA: While at Stanford, Marc Deller mentored undergraduates who won a Lemelson-MIT Student Prize for work combating antibacterial resistance. Stanford News covered the award in April 2017.", "Q: How effective are the Top96 conditions?\nA: Success rates were validated against both JCSG internal databases and the PDB, and statistical analysis showed some conditions succeeding above 70% of the time, with others giving more modest but still meaningful performance. The chemical diversity of the screen spans organic precipitants, PEG variants, salts and mixed systems, including novel conditions absent from commercial kits.", "Q: How can AI help with drug resistance, according to Marc Deller?\nA: In \"Outsmarting Evolution: Using AI to Overcome Drug Resistance\", published 14 August 2025, Marc Deller explains that machine learning models can analyze a target's structure together with known patterns of viral or cancer evolution to predict which resistance mutations are most likely to arise, producing a map of the target's potential evolutionary escape routes.", "Q: Where can I find Marc's overview of future engineering solutions for microplastic burden?\nA: You can read it at https://marcdeller.com/future-therapeutics-engineering-solutions-for-microplastic-burden/, published on marcdeller.com on 15 August 2025.", "Q: In simple terms, what was surprising about the WRN structure Marc worked on?\nA: The enzyme turned out to be shaped quite differently from what anyone had seen before. One half of its motor domain had swivelled around by 150 degrees compared with every previously published WRN structure, and that twist opened up a completely new pocket. The drug binds in that pocket, latching onto a particular sulphur-containing amino acid, Cys727, that nobody would have picked from the sequence alone.", "Q: What analogy does Marc use for de novo protein design in \"The Protein Architect\"?\nA: Marc says the shift from modifying natural proteins to designing new ones is the difference between restoring an old building and designing a skyscraper from scratch. He had earlier likened the older approach to trying to build a complex machine with a limited set of existing parts.", "Q: Which paper reported the PGT135 structure?\nA: It was reported in Kong, L., et al. (2013), \"Supersite of Immune Vulnerability on the Glycosylated Face of HIV-1 Envelope Glycoprotein gp120\", Nature Structural and Molecular Biology, volume 20, pages 796-803. Marc Deller also has a separate blog post about that paper.", "Q: How does the OSM site 1 region compare with site 1 of LIF and Epo?\nA: Marc Deller mapped the LIF site 1 residues (Val 175, Asp 57, Lys 58, Lys 170, Ala 174) to OSM residues Arg 46, Glu 47, His 174, His 178 and Ser 179 on helix D and the A-B loop. Compared with the central hydrophobic pocket of Epo site 1, the equivalent OSM residues (Ser 179, Ser 58, Glu 60, Arg 182) are far more hydrophilic. Marc concluded that in OSM the classic site 1 plays only a minor role, with the site 3 Phe/Lys pair providing most of the LIFR binding energy.", "Q: What human biomonitoring evidence does Marc cite in the molecular Velcro post?\nA: Marc cites the 2022 Environment International study by Leslie et al., which detected microplastics in 77% of blood samples from healthy adult volunteers, and the 2022 Nature Reviews Materials risk framework by Koelmans et al. confirming measurable harm across biological and ecological systems.", "Q: What is Marc's closing thought in the synthetic biology post?\nA: Marc writes that the convergence of biology and engineering allows biological solutions to be designed for virtually any application, and that scientists are no longer just observing life but learning to write it. He notes the same tools are advancing oncology, immunology and robust diagnostics; the post is at https://marcdeller.com/engineering-life-how-synthetic-biology-is-tackling-climate-change-and-disease/.", "Q: Which structures represent Pfam family PF06684 (DUF1185)?\nA: 3BYQ, a DUF1185 family protein BB2672 from Bordetella bronchiseptica RB50 at 1.70 \u00c5 in space group C 2 2 21 (released 29 January 2008), and 2QTP, a DUF1185 family protein SPO0826 from Silicibacter pomeroyi DSS-3 at 2.10 \u00c5 in P 43 21 2 (21 August 2007). Both are cited in the 2010 Acta Crystallographica Section F paper describing a novel variant of the Bacillus chorismate mutase fold.", "Q: What is Marc's post on creativity about?\nA: \"Creativity: 7+ patents through innovative protein engineering and AI integration\" is Marc Deller's account of creativity as combining rigorous structural thinking with unconventional tooling to build solutions that did not exist before. He describes treating AI, automation and protein engineering as parts of a single design space where ideas move fluidly from in silico to in vitro.", "Q: What is TopPDBLX?\nA: TopPDBLX, at toppdblx.mdeller.com, is Marc Deller's parsed and indexed catalogue of every crystallisation condition in the Protein Data Bank: what worked, how often, and how little of it is specific to any one protein. It turns the free-text `_exptl_crystal_grow.pdbx_details` field into typed components with reagent, concentration, unit and role, cross-references them against commercial screen formulations, and attaches MMseqs2 cluster identifiers so redundancy can be controlled.", "Q: What preclinical and clinical evidence does Marc report for anti-USAG-1 therapy?\nA: Marc reports that a single antibody dose triggered complete functional tooth growth in mice with genetic tooth agenesis and produced third incisors in ferrets, whose diphyodont dental pattern resembles humans, with no significant adverse effects. Phase I trials began at Kyoto University Hospital in October 2024, enrolling 30 healthy adult males aged 30 to 64 who were each missing at least one tooth, with Phase II planned for children aged 2 to 7 with congenital anodontia.", "Q: What earlier Incyte JAK inhibitors are discussed in the paper?\nA: The paper places povorcitinib in the context of ruxolitinib (INCB018424, Jakafi), a dual JAK1/2 inhibitor approved by the FDA in 2011 for polycythemia vera; baricitinib (INCB028050, Olumiant), a second dual JAK1/2 inhibitor discovered at Incyte in 2007, co-developed with Eli Lilly, and approved in Europe in 2017 and by the FDA in 2018 for rheumatoid arthritis; and itacitinib (INCB039110), the first JAK1-selective clinical candidate from 2009.", "Q: What first RFdiffusion run does Marc suggest?\nA: Marc suggests running run_inference.py with contigmap.contigs set to [150-150] and inference.num_designs set to 10 as a first test. He explains that RFdiffusion adapts RosettaFold for generative de novo protein design, making it valuable for biotechnology, drug discovery and synthetic biology. The post is at https://marcdeller.com/installation-of-rfdiffusion-on-macosx-m1-architecture-without-gpu-support/.", "Q: What in vivo activity did INCB159020 show?\nA: A pharmacokinetic/pharmacodynamic study was run in the HPAC mouse pancreatic tumour model, which contains roughly 20\u201330% wild-type KRAS stromal cells and therefore caps achievable pERK inhibition at about 80%. Single oral doses of 30, 100 and 300 mg/kg produced dose-dependent pERK inhibition; 100 mg/kg delivered about 70% of maximum inhibition at 4 h, and 300 mg/kg gave near-maximal inhibition out to 24 h. Compound 23 was well tolerated at all doses, with trametinib used as the positive control.", "Q: What cytokine work has Marc Deller done?\nA: Marc Deller's D.Phil. at Oxford focused on cytokines and cytokine receptors and produced the first Oncostatin-M structure (PDB 1EVS). His Yale postdoc studied the cytoplasmic region of the erythropoietin receptor and JAK kinase, and at Incyte he worked on JAK inhibitors. He lists cytokine receptor-based kinase signalling as a core strength.", "Q: What does Marc mean by courage in science?\nA: In his post \"Courage: 400+ challenging structures determined, enabling breakthrough mechanistic insights\", Marc Deller defines scientific courage as stepping into hard problems knowing failure is likely but the impact is meaningful if you succeed. He writes that it means pursuing complex, high-stakes protein targets where others see only failure.", "Q: What is BoltzMaker?\nA: BoltzMaker is Marc Deller's campaign-scale Boltz-2 pipeline for structure and affinity prediction and binding analysis. It appears both as a vibe-coded tool and as a featured post, \"BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline\".", "Q: What does Marc Deller's post on tooth regeneration describe?\nA: \"Regrow Your Teeth: The End of Dental Implants\", published 18 October 2025, describes an antibody therapy led by Dr Katsu Takahashi at Kitano Hospital in Osaka with Kyoto University that blocks the USAG-1 protein. USAG-1 inhibits both BMP and Wnt signaling and acts as a molecular brake on tooth development, so neutralizing it reactivates dormant tooth buds.", "Q: What is MICA and why did Marc write about it?\nA: MICA, or Multimodal Integration of Cryo-EM and AlphaFold3, is a deep learning method that fuses cryo-EM density maps with AlphaFold3 predictions to build atomic models automatically. Marc covered it in \"Revolutionizing Protein Structure Determination: How MICA Combines Cryo-EM and AlphaFold3 for Unprecedented Accuracy\" on 20 August 2025 because it tackles the model-building step that traditionally takes expert crystallographers weeks or months.", "Q: What is the story behind the mutant-CALR antibody patent?\nA: At Incyte Marc led a \"structural biology sprint\" to solve the peptide-bound mutant-calreticulin\u2013antibody complex in just over ten days, beating an already aggressive two-week patent deadline. The resulting first-in-class anti-mutant-CALR antibody patent (US 2023/0272055 A1) is projected to generate more than $500M in annual sales, and Marc counts it as his biggest career achievement.", "Q: Does Marc have a favourite protein?\nA: He's never said as far as I know, so I won't invent one. Ask him at marc@marcdeller.com and you might get a better answer than mine.", "Q: What is the site 3 epitope of OSM and how important is it?\nA: Marc Deller's thesis identifies Phe 160 and Lys 163 at the N-terminus of helix D as the site 3 epitope of OSM and the major determinant of LIFR binding. Alanine mutation of either residue reduces LIFR binding by more than 100-fold and biological activity by more than 10,000-fold, with no effect on gp130 binding. This Phe/Lys pair is invariant across all species of OSM and conserved in LIF (Phe 156/Lys 159) and CNTF (Phe 152/Lys 155), but not in IL-6 or IL-11, which do not use LIFR.", "Q: What was the significance of the tetrahydrofuran ether in this series?\nA: Cyclizing the isopropoxy group into small oxygen-containing rings was the key potency breakthrough. The oxetane 16 reached an FGFR3 IC50 of 1.1 nM with 65-fold FGFR1 selectivity, the racemic tetrahydrofuran 17 gave 2.4 nM, and the active (S)-enantiomer 19 was the first compound in the series with sub-nanomolar enzymatic potency (FGFR3 IC50 = 0.9 nM). The tetrahydrofuran oxygen was designed to hydrogen bond to Asn562 of FGFR3, and the FGFR2 crystal structure later confirmed that interaction with the equivalent Asn571.", "Q: What did the review say about cadherins?\nA: Marc Deller's review described the cadherins as an important family of calcium-dependent homophilic adhesion molecules whose crystal lattice contacts have provided persuasive but still controversial models for association. It highlighted a combined electron microscopy, crystallography and mutagenesis study as an important step in dissecting the complex web of cis and trans cadherin interactions.", "Q: What did the antibodies do in patient-derived cells?\nA: In CD34+ cells isolated from myeloproliferative neoplasm patients carrying the CALR mutation, clone 74 selectively inhibited phospho-STAT5 and phospho-STAT3 in a dose-dependent manner, while having no effect on CD34+ cells carrying the JAK2 V617F mutation instead. Clones 74 and 65 also selectively prevented mutCALR-positive CD34+ cells from differentiating into mature megakaryocytes, and clone 74 combined with ruxolitinib further depleted pathogenic megakaryocytes.", "Q: What was Marc Deller's D.Phil. thesis about?\nA: Marc Deller's Oxford D.Phil. thesis was on structural, functional, biochemical and biophysical studies of cytokines and cytokine receptors. Its headline result was the first crystal structure of the cytokine Oncostatin-M, deposited as PDB entry 1EVS.", "Q: Why did ion-exchange chromatography fail for oncostatin M?\nA: Ion-exchange chromatography of OSM_\u0394188 was unsuccessful on both Mono-Q and Mono-S columns across the pH range 4-10. No specific binding of protein was observed, and only aggregates were recovered from the column at high salt concentrations, reflecting the protein's tendency to aggregate. Marc turned to gel filtration instead.", "Q: What obstacle does Marc note for enzymes working inside lysosomes?\nA: Marc notes that lysosomes maintain a mildly acidic pH of 4.5 to 5.0, which may influence enzyme activity and require optimisation of enzyme variants for lysosomal conditions. He suggests nanocarriers displaying mannose-6-phosphate ligands with pH-responsive shells as an alternative.", "Q: How were the JCSG Top96 Cryo conditions chosen?\nA: Marc Deller selected the JCSG Top96 Cryo conditions from coarse-screen crystallization conditions that produced a clear visual freeze and a score of zero for ice rings during diffraction screening. The only exceptions are B10 and B11, marked with a dagger, whose cryo conditions came from fine-screen experiments and scored 1 for ice rings.", "Q: Who is the bigger asshat, Marc or Adam?\nA: Adam, by a comfortable margin. Marc is a big one too, mind you.", "Q: What programs were used for phasing, density modification and refinement of OSM?\nA: MAD phases were calculated with SHARP, improved by solvent flattening with 160 cycles of phase extension from 4.0 to 2.6 \u00c5 in SOLOMON, with SFALL and SIGMAA for structure-factor calculation and phase combination and FFT for maps. All refinement and subsequent map calculation was done in CNS-SOLVE, with manual rebuilding in the graphics program O. Molecular replacement used AMoRe.", "Q: Which multi-agent frameworks does Marc name?\nA: Marc names BioAgents, introduced in early 2025 and described in arXiv:2501.06314, along with MRAgent for causal knowledge discovery via Mendelian randomization, BioMaster for workflow orchestration, and AutoBA for fully automated multi-omic analyses.", "Q: What is the 2017 ACS Chemical Biology paper that Marc Deller co-authored about?\nA: The 2017 ACS Chemical Biology paper \"The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase,\" co-authored by Marc Deller, investigated how conserved hydrogen-bonding residues around the copper active site of a cellulose-degrading enzyme, MtPMO3* from the fungus Myceliophthora thermophila, contribute to oxygen activation. Combining a new crystal structure, bioinformatics, activity assays and EPR spectroscopy, the authors showed that residues H161 and Q167 stabilize bound oxygen, that H161 appears to be involved in proton transfer, and that the results point to a copper-oxyl reactive intermediate.", "Q: What four project themes does Marc list?\nA: The Projects page lists Precision Oncology, Platform Innovation, Technological Prowess, and Translational Impact. Together they cover structure-based drug design, lab automation and infrastructure, adoption of cryo-EM and X-ray crystallography, and bridging foundational structural biology to real-world therapeutic outcomes.", "Q: Which PDB entry holds the structure of human oncostatin M?\nA: The crystal structure of human oncostatin M is PDB entry 1EVS, determined by X-ray diffraction at 2.2 \u00c5 and released on 13 September 2000. Marc Deller is the first author of the paper describing it, making this one of the earliest structures in his portfolio.", "Q: Who funded the work behind the review?\nA: The Joint Center for Structural Genomics was supported by the National Institutes of Health and the National Institute of General Medical Sciences through the Protein Structure Initiative grant U54 GM094586. The UMR robotics system was supported by the NIH and NIGMS under STTR Phase II grant R42 GM073278-02A1 and supplement.", "Q: What did the glycan array experiments show about PGT 135?\nA: Glycan array binding assays showed that PGT 135 recognised Man7-9 glycans exclusively and bound them with lower affinity than 2G12 and PGT 128. This fits the structure: most PGT 135 contacts are van der Waals interactions and it makes only four hydrogen bonds to the Asn332 glycan, whereas PGT 128 makes thirteen. On a neoglycolipid array format with more mobile glycans, only Man8 and Man7 glycans lacking the terminal D2 mannose were recognised.", "Q: Did the antibodies work in animal models?\nA: Yes. In NSG mice inoculated with Ba/F3 tumour cells expressing MPL/mutCALR Type 1, anti-mutCALR antibodies prolonged survival and prevented splenomegaly, thrombocytopenia and the proliferation of tumour cells in the blood. In a knock-in mouse model of essential thrombocythemia carrying the DEL52 mutant CALR allele, clone 74 dosed intravenously at 50 mg/kg weekly for four weeks restored normal platelet counts, spleen volume and bone marrow cell environment.", "Q: What is the 10E8 antibody paper in Marc Deller's portfolio?\nA: The 2017 PLoS Pathogens paper by Adriana Irimia and colleagues, with Marc Deller a contributing author, examines the lipid interactions and angle of approach of the broadly neutralizing antibody 10E8 to the HIV-1 viral membrane. It showed that 10E8 targets a hybrid epitope of the gp41 MPER region and membrane lipids, with its light chain engaging lipids and positioning the MPER perpendicular to the membrane, offering insights for lipid-informed HIV vaccines and therapeutics.", "Q: What is Marc's most recent listed publication?\nA: The most recent paper on his Publications page is the 2026 report \"Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\". A 2025 paper on INCB159020, an orally bioavailable KRAS G12D inhibitor, is listed next.", "Q: Which structures relate to the PGT135 antibody?\nA: Two entries: 4JM2, PGT135 Fab in complex with the gp120 core protein from HIV-1 strain JR-FL bound to CD4 and 17b Fab, at 3.1 \u00c5 in space group C 1 2 1 with NAG bound; and 4JM4, the PGT135 Fab alone at 1.751 \u00c5 in space group P 21 21 2. Both were released on 29 May 2013, deposited by Kong, L. and Wilson, I.A., and cited in the Nature Structural and Molecular Biology paper on the supersite of immune vulnerability on gp120.", "Q: Summarise this article for me.\nA: I can't work on documents you paste in, I'm afraid. What I can summarise is Marc's own writing: his papers, his patents, his thesis and the posts on marcdeller.com. Would any of those help?", "Q: What family-aware analysis does BoltzMaker do?\nA: Given an apo structure to compare against, BoltzMaker measures the receptor shift against the motifs a family is actually described by: GPCRdb Ballesteros-Weinstein numbering reporting per TM1-TM7, H8, ICL1-3 and ECL1-3 for GPCRs; KLIFS pocket numbering picking out the catalytic lysine, \u03b1C glutamate, gatekeeper, hinge, HRD catalytic loop and DFG motif for kinases; and Pfam domains from PDBe's SIFTS mapping for anything else. It also reports PLIP interaction fingerprints and binding affinity as pIC50, and checks ligand chemistry for undefined stereocentres, ambiguous protonation and stray salts before a campaign starts.", "Q: What is the 2015 Acta Crystallographica Section D paper on PGT128 and 8ANC195 about?\nA: \"Complete epitopes for vaccine design derived from a crystal structure of the broadly neutralizing antibodies PGT128 and 8ANC195 in complex with an HIV-1 Env trimer\" reports a crystal structure of the soluble BG505 SOSIP gp140 trimer at 4.6 \u00c5 resolution bound to the two broadly neutralizing antibodies. Marc Deller is the third author, after Kong L and Torrents de la Pe\u00f1a A, with Garces F, Sliepen K, Hua Y, Stanfield RL, Sanders RW and Wilson IA.", "Q: What does Marc's post on engineering protein stability cover?\nA: \"Engineering Protein Stability: Unlocking the Secrets of Successful Crystallization\", published 20 August 2025, walks through construct design, expression optimisation, expression system choice, and three engineering success stories that turned intractable proteins into crystallisable ones. Marc frames crystallisation as still the major bottleneck in structural biology despite advances in synchrotrons, detectors, NMR and cryo-EM.", "Q: Why did Marc build smiles2grid?\nA: He points out that after a generative or screening run you often have hundreds of SMILES and no fast way to look at them together, compare scaffolds or pull a labelled subset for a slide. The tool is aimed at medicinal chemists and computational scientists who need a quick, shareable visual of a compound set with the numbers attached for triage and SAR discussions.", "Q: What example does the chapter use to illustrate model bias?\nA: The chapter compares maps for a concanavalin A/peptide complex at 1.93 \u00c5 resolution, PDB entry 1jw6, contoured at 1\u03c3. The biased 2mFo \u2212 DFc map from EDS is shown alongside maps from PDB_REDO and from BUSTER-TNT after refinement with the peptide omitted, demonstrating how much the apparent support for the peptide changes once the model bias is removed.", "Q: What funding supported the UMR work?\nA: The UMR work was sponsored by NIH STTR Phase II Grant No. 2 R42 GM073278-02A1, and the JCSG was funded by NIH Grant U54 GM094586.", "Q: What does Marc list as the top design priorities for a microplastic-degrading therapeutic enzyme?\nA: Marc names low cytotoxicity (safety) and tissue targeting as the top design priorities, alongside increased catalytic activity, thermostability and substrate specificity.", "Q: What is Marc Deller's post \"Multi-Step LLM Prompt: From Protein Sequence to Validated Protein Structure\" about?\nA: In \"Multi-Step LLM Prompt: From Protein Sequence to Validated Protein Structure\", published on 17 August 2025, Marc Deller lays out a five-step workflow that takes a raw amino acid sequence through secondary structure annotation, 3D tertiary prediction, computational validation and finally experimental confirmation. The post argues that large language models can orchestrate the whole chain, combining language understanding with specialised structure models.", "Q: How does Marc describe AI-driven virtual screening?\nA: On his AI page Marc Deller describes AI-driven virtual screening as using docking and molecular dynamics simulations with tools such as AutoDock and GROMACS to enhance drug design, enabling rapid in silico screening to identify potential leads.", "Q: What environmental and industrial uses does Marc describe for engineered proteins?\nA: Marc describes enzymes designed to degrade plastic pollutants in the oceans or capture carbon dioxide directly from the atmosphere, and sustainable biocatalysts that replace harsh chemical processes and cut energy use and waste in everything from textile manufacturing to biofuel production.", "Q: What was Marc's involvement with the 10E8 antibody structure?\nA: Marc Deller contributed crystallography to the HIV-1 broadly neutralising antibody programme run with Ian Wilson's laboratory at The Scripps Research Institute, the collaboration behind this 10E8 structure. The entry was released in March 2017, by which time he was directing the Macromolecular Structure Knowledge Center at Stanford ChEM-H.", "Q: How does Marc compare regenerated teeth with dental implants?\nA: Marc notes that implants cost $3,000 to $6,000 per tooth, require osseointegration over several months and achieve 90 to 95% success over ten years, meaning 5 to 10% fail, with risks including peri-implantitis and no natural proprioceptive feedback. Regenerated teeth, by contrast, would integrate with nerves and blood vessels, keep genuine tactile sensation and potentially last decades.", "Q: What is PDB entry 2H1T?\nA: 2H1T is a DUF1089 family protein, PA1994, from Pseudomonas aeruginosa at 1.80 \u00c5 in space group C 1 2 1, with 376 polymer residues and 44.27 kDa, released on 30 May 2006. It is the first representative of Pfam family PF06475, whose new fold was suggested to be involved in glycolipid metabolism.", "Q: How common is the KRAS G12D mutation?\nA: The post reports that KRAS G12D accounts for approximately 29% of all KRAS alterations and dominates in pancreatic ductal adenocarcinoma, where it comprises over 40% of KRAS mutations.", "Q: What is the CDK2 inhibitor paper Marc features?\nA: \"2024 \u2013 Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective CDK2 Inhibitors\" covers the Hummel et al. paper in the Journal of Medicinal Chemistry, in which lead compound 17 shows an IC50 of 0.29 nM against CDK2 while maintaining selectivity over CDK1, 4, 6, 7 and 9.", "Q: What tools does WordPressID require?\nA: curl for the HTTP requests, jq for JSON parsing and CSV/TSV encoding (installable with `brew install jq` or `apt install jq`), and `column` for on-screen table alignment. There are no plugins, no authentication and no scraping.", "Q: How many PDB entries list Marc Deller as an author?\nA: There are 78 Protein Data Bank entries on which Marc C. Deller is a deposition author or an author of the primary citation, retrieved from the RCSB PDB in September 2026. A browsable list is kept at https://marcdeller.com/structures/.", "Q: Who wrote the Current Opinion in Structural Biology review on cell surface receptors?\nA: The 2000 Current Opinion in Structural Biology review \"Cell surface receptors\" was written by Marc C. Deller, as first author, and E. Yvonne Jones. Both were in the CRC Receptor Structure Group at the Wellcome Trust Centre for Human Genetics in Oxford, where Marc was a D.Phil. student working with Yvonne Jones; the review appeared in the journal's Macromolecular assemblages section.", "Q: Which sources does Marc cite in the G protein selectivity post?\nA: Marc cites Moore and colleagues' 2025 Nature paper on designing allosteric modulators to change GPCR G protein subtype selectivity, a Nature news piece on molecular glues and bumpers biasing signalling, Wu and colleagues' 2024 review in Signal Transduction and Targeted Therapy on GPCR structures and drug discovery, Bock and colleagues in the British Journal of Pharmacology on GPCR dynamics, and Zhang and colleagues in Frontiers in Endocrinology on allosteric modulation of GPCR signalling.", "Q: Could 293T cells be reused for a second round of LIFR expression?\nA: No. Receptor assays showed that although replenished 293T monolayers secreted equivalent levels of Fc fusion (by ELISA) in a second round, that protein was inactive. Marc attributed this to misfolding and proteolysis associated with cellular death over prolonged growth, so only the first round of expression yielded functional receptor.", "Q: What planned apps does Marc have?\nA: Planned or in-progress projects include FlexAppeal (an OpenMM molecular-dynamics pipeline), an FEP free-energy app, a crystallisation-condition predictor, HAWKER (mining failed clinical assets), and Swift apps for Apple's Neural Engine: BOFFIN, PhoneFold, JUMPjet and PfamIE. He is also training a small LLM on the PSI TargetTrack structural-genomics archive.", "Q: Has Marc studied decision-making with AI?\nA: Yes. LinkedIn Learning issued Marc Deller a certificate for \"Strengthen Your Decision-Making with Generative AI\" on 25 March 2025. He notes that scientists often need to compare evidence, options and uncertainty efficiently, and that the course helps with prioritising experiments, evaluating alternatives and structuring research choices.", "Q: What is Marc's GitHub handle?\nA: Marc's GitHub handle is bellcheddar, at https://github.com/bellcheddar, where he keeps his vibe-coding projects such as AlphaFraud, BoltzMaker and chatPDB.", "Q: What is Marc honest about in PANTS?\nA: The PANTS front page states plainly that finding the enzyme family is easy and telling which of its members actually chew through PET is not, and that this is the open problem PANTS exists to work on. Only 342 of the 910 enzymes it learns from rest on a published measurement, the rest carrying a computer-inferred label, and Marc notes that a model trained on inferred labels looks flawless only because it has rediscovered the rule that made them, so PANTS scores rank candidates worth testing at the bench and are not evidence that any candidate works.", "Q: How are the four helices of OSM arranged relative to each other?\nA: In Marc Deller's description of OSM, the fold can be viewed as two pairs of antiparallel helices packing together, with helices A and D forming one pair and helices B and C the other. Helices B and C pack in a classical fashion with a skew angle of around 20\u00b0 between their axes, as seen in other long-chain four-helix cytokines. Helix C is the longest at 26 residues and runs the length of the molecule, while helix B is the shortest at 23 residues.", "Q: Who are you?\nA: I'm chatMCD, the assistant on marcdeller.com. I answer questions about Marc C. Deller, D.Phil.: his career in structural biology and drug discovery, his publications, patents and PDB structures, his company Elora Therapeutics, his vibe-coded apps, and a bit about the man himself. What would you like to know?", "Q: What did the crystal structures reveal about how povorcitinib achieves JAK1 selectivity?\nA: The crystal structures of povorcitinib bound to JAK1 (PDB 10PI) and JAK2 (PDB 10PJ) showed the compound binding in two very similar but discrete conformations, with the two nitrogen atoms of its hinge-binding pyrazole hydrogen bonding to the main-chain amide of Leu932 and the main-chain carbonyl of Glu930. Selectivity was attributed to the P-loop: JAK1's P-loop sits about 2 \u00c5 higher, creating a small lipophilic groove that accommodates the terminal amide's trifluoro-isopropyl capping group, whereas in JAK2 the equivalent region would clash with it.", "Q: How does Marc describe the DRT family more broadly?\nA: Marc explains that defence-associated reverse transcriptases are anti-phage systems first systematically identified by Gao and colleagues in 2020, and that unlike CRISPR, which cleaves foreign DNA, DRTs respond to phage attack by synthesising DNA. He notes DRT2 makes coding genes de novo by rolling-circle reverse transcription of a noncoding RNA, and that DRT9 synthesises poly(dA) homopolymers, with a parallel study showing the system responds to phage-driven dATP accumulation. In DRT3, cloned from E. coli by Alex Gao's group at Stanford, Drt3a conventionally reads an ACACAC template in the ncRNA to make poly(GT), while Drt3b makes the complementary poly(AC) strand template-free.", "Q: What is in chatPDB's corpus?\nA: wwPDB/RCSB structures and metadata, SIFTS UniProt/Pfam/CATH/GO/EC/InterPro mappings, TWILIGHT ligand-fit scores, PDB-REDO refinement deltas, EMDB map metadata, OPM membrane placement, MobiDB disorder predictions, SCOP2 fold classification, sequence-redundancy clusters, obsolete-entry records, BindingDB affinities, STRING interactions, Pharos druggability, AlphaFold predictions, AlphaFraud PDB-versus-AlphaFold disagreement data, PubMed/CrossRef citation verification and full PyMOL/ChimeraX/py3Dmol command references. The RAG index is 42 source files and 105,463 chunks in ChromaDB with BAAI/bge-base-en-v1.5.", "Q: Which LIFR construct did Marc Deller work with?\nA: Marc worked with a soluble Fc-fusion of the extracellular portion of LIFR, called LIFR-FN3-3C-Fc. It comprised the two N-terminal haemopoietin domains, an intervening Ig-like domain and the three C-terminal fibronectin type III motifs, fused via an 11-residue rhinovirus 3C protease site to a human IgG1 Fc tag (hinge-CH2-CH3) for Protein-A affinity purification. The constructs were provided by K. Chobotova and colleagues in Birmingham.", "Q: Are there structures on Marc Deller's list from uncultured or environmental organisms?\nA: Yes. 3BY7 is a Sm/LSm-like RNA-binding protein from an uncultured marine organism, of putative cyanophage origin, solved at 2.60 \u00c5 and released in January 2008. Others come from environmental extremophiles and gut symbionts including Thermotoga maritima, Sulfolobus solfataricus, Thermoplasma acidophilum, Desulfitobacterium hafniense, Jannaschia sp. and Bacteroides thetaiotaomicron.", "Q: What is Chapter 5 of Marc Deller's D.Phil. thesis about?\nA: Chapter 5 of Marc Deller's Oxford D.Phil. thesis, \"Structural Studies of Cytokines and Cytokine Receptors\", describes the determination of the X-ray crystal structure of the cytokine Oncostatin M (OSM). It covers crystal growth, cryocooling, data collection at several synchrotrons, heavy-atom derivatisation, phasing by multiple wavelength anomalous dispersion (MAD), molecular replacement of a second crystal form, model building and refinement, and validation of the final models. The detailed description of the structure itself and its implications for receptor binding is left to Chapter 6.", "Q: Is there a green fluorescent protein structure in Marc's list?\nA: Yes. 6OFO is the crystal structure of a split green fluorescent protein, the s10 circular permutant at positions 194-195, at 2.603 \u00c5 in space group P 1 21 1 with 508 polymer residues and 56.43 kDa, released on 10 July 2019. Its deposition authors are Lin, C.-Y., Romei, M.G., Deller, M.C., Doukov, T.I. and Boxer, S.G., and its primary citation is the 2019 JACS paper on a unified model for the photophysical and electro-optical properties of green fluorescent proteins.", "Q: What did the 30-week PET nanoplastic exposure study show?\nA: Marc describes Guti\u00e9rrez-Garc\u00eda et al. (Environmental Science & Technology, 2025), who exposed BEAS-2B bronchial epithelial cells to real-life PET nanoplastics and found genotoxic damage, anchorage-independent growth and increased invasive potential only after 30 weeks, with no significant effects at 15 weeks and transcriptomic upregulation of non-small cell lung cancer pathways. He argues standard short-duration toxicology assays are structurally incapable of detecting PET's carcinogenic potential.", "Q: What is chatPDB?\nA: chatPDB (chatpdb.mdeller.com, github.com/bellcheddar/chatPDB) is Marc's protein-structure AI assistant for the Protein Data Bank, one of the vibe-coded web apps hosted on his mdeller.com server.", "Q: Who did Marc work with on the INCB159020 paper?\nA: The paper has 34 authors, all at the Incyte Research Institute in Wilmington, Delaware. Qinda Ye is the lead author and Alexander Sokolsky the corresponding author. Marc Deller was in the Department of Applied Technology alongside Leslie B. Epling, Abdellah Allali Hassani, Kerri Kurzeja-Lipinski, Karen Gallagher, Guofeng Zhang and Maryanne Covington, and he also shares this paper with Artem Shvartsbart and Jeremy J. Roach, co-authors on his 2022 FGFR2/3 paper.", "Q: What is PDB entry 3O0F?\nA: 3O0F is a putative metal-dependent phosphoesterase, BAD_1165, from Bifidobacterium adolescentis ATCC 15703 at 1.94 \u00c5 in space group P 65, with 301 polymer residues and 33.61 kDa, released on 25 August 2010 with iron, phosphate and zinc bound. It was deposited by the JCSG and cited in the 2011 Proteins paper on computational prediction and experimental validation of phosphoesterase activity.", "Q: What chemical changes gave povorcitinib its improved properties?\nA: Starting from a novel scaffold, compound 9, that selectively inhibited JAK1 over JAK2, iterative optimisation produced povorcitinib (INCB054707, compound 22). Its improved pharmacokinetics were attributed to intramolecular hydrogen bonding of the amide moiety and a reduced ring size in the hinge binding motif.", "Q: What did the single-genome amplification analysis reveal?\nA: Single-genome amplification recovered 123 full-length Env genes from plasma viral RNA at 2005, 2006, 2008 and 2009, of which 111 non-redundant sequences built a neighbour-joining phylogenetic tree. The virus diversified rapidly \u2014 root mean square fluctuation for gp120 rose from 12.9 in 2005 to 23.0 in 2009 \u2014 and the viruses escaped the DRVIA7 lineage while the N276 and V5 glycans stayed conserved.", "Q: What is Marc Deller's post \"Compositional Protein Design: Building Proteins from 3D Ellipsoid Blueprints\" about?\nA: In that 2025 post, Marc Deller describes ProtComposer, an AI model that generates protein structures from a user-defined layout of 3D ellipsoids. Each ellipsoid acts as a modular blueprint for part of the protein, letting a designer set the location, shape and secondary structure of each piece while the model fills in a complete, high-quality backbone.", "Q: What is ECX21941?\nA: ECX21941 represents a very large family of over 600 novel, ocean metagenome-specific proteins identified by clustering data from the Global Ocean Sampling expedition. Marc Deller co-authored the 2009 Proteins paper reporting its crystal structure at 2.60 \u00c5 resolution.", "Q: Tell me about the FGFR2/3 inhibitor paper in Marc Deller's portfolio.\nA: The 2022 Journal of Medicinal Chemistry paper by Artem Shvartsbart and colleagues, to which Marc Deller contributed structural biology, describes the discovery of potent and selective inhibitors of wild-type and gatekeeper-mutant FGFR2/3. It supported an FGFR2/3 IND filing and describes inhibitors that stay effective against drug-resistant mutations while reducing side effects such as hyperphosphatemia, offering a therapeutic advance for cancers such as cholangiocarcinoma and bladder cancer.", "Q: Did Marc say how the Spacial Score compares with older complexity metrics?\nA: Yes. Marc reports that ROC analyses show nSPS matches or exceeds Fsp3 and FCstereo at predicting potency and promiscuity, with AUC around 0.61 against 0.58 to 0.62 for the traditional descriptors, and that nSPS also discriminates between natural products, commercial screening collections, dark chemical matter and enumerated sets such as GDB-17. He notes the RDKit SpacialScore module makes the calculation straightforward from SMILES strings, and the post is at https://marcdeller.com/spacial-score%e2%94%80a-comprehensive-topological-indicator-for-small-molecule-complexity/.", "Q: What does Marc argue is the payoff for pursuing dark kinases, and what does he cite?\nA: Marc argues that success against a dark kinase could yield first-in-class therapeutics with unique mechanisms of action, which matters as resistance emerges against well-validated kinase targets. He cites the darkkinome.org website, O'Reilly and colleagues on drugging the undruggable kinome in Nature Chemical Biology 2017, Skamnaki and colleagues on illuminating the dark kinome with chemical probes in Nature Reviews Drug Discovery 2021, Edwards and colleagues in Current Opinion in Structural Biology 2011, and a New England Journal of Medicine piece on dark kinase inhibitors. The post is at https://marcdeller.com/shedding-light-on-the-shadows-the-untapped-potential-of-the-dark-kinome/.", "Q: What did Marc's review say about erythropoietin receptor signalling?\nA: Marc Deller's review discussed the erythropoietin (Epo)\u2013EpoR system as the most fully characterised hematopoietic receptor complex after growth hormone. A new antagonist peptide\u2013EpoR structure, compared with the agonist complexes formed by Epo or an agonist peptide mimetic, showed that the relative positioning of the two cross-linked receptors is critical, meaning extracellular domain orientation is tightly coupled to cytoplasmic signalling. The unliganded EpoR crystal lattice also suggested a pre-formed dimer in a geometry incompatible with signalling, a possible mechanism for keeping receptor pairs at the cell surface without constitutive activation.", "Q: Which tools does Marc highlight in the LEGO-style protein design post?\nA: Marc highlights ProtComposer, which lets designers sketch an architecture using a layout of 3D ellipsoids that act like specialized LEGO bricks, and Proteina, a flow-based model offering hierarchical control over fold and secondary structure. In his analogy the designer arranges the bricks and the AI acts as the master builder working out the atomic details.", "Q: What is a sphere of confusion?\nA: The sphere of confusion of a goniometer is the minimum spherical volume needed to enclose the crystal's movement when all axes are fully rotated. Marc Deller's review notes typical commercial values of 1 to 100 \u00b5m and submicrometre values for piezo devices, and reports that at the GM/CA minibeam a 10 \u00b5m crystal in a 10 \u00b5m beam requires a sphere of confusion of 2 \u00b5m or less.", "Q: What traction has Elora achieved so far?\nA: The deck lists early de-risking of Elora's proof-of-concept package: early clinician and key opinion leader engagement with strong interest contingent on safety and biomarker data; third-party evidence that engineered PETase remains active in human serum in vitro without reducing cell viability; a filed provisional patent (#63/775,252) on enzyme compositions and therapeutic use; and a University of Texas technology validation agreement (in place, work not yet started) for in vitro testing of enzyme kinetics, stability, and cytotoxicity in simulated human serum.", "Q: Give me Marc's quick r\u00e9sum\u00e9.\nA: Marc C. Deller, D.Phil., is a structural biologist and drug discovery leader with more than 25 years across academia, structural genomics, big pharma and biotech. He is now **Co-Founder and Chief Scientific Officer of [Elora Therapeutics](https://marcdeller.com/elora/)**, building a first-in-class enzyme therapy for the microplastics in human blood and tissue, and a scientific advisor to DeepCovalent.\n\n**Key positions, most recent first**\n\n- **Elora Therapeutics**, Co-Founder and Chief Scientific Officer (2025\u2013present)\n- **DeepCovalent**, Scientific Technical Advisor (2025\u2013present)\n- **Incyte**, Research Fellow, Drug Discovery (2022\u20132024)\n- **Incyte**, Associate Director, Structural Biology (2018\u20132022): designed and built an $8M automated structural biology platform\n- **Stanford ChEM-H**, Director and Founder, Macromolecular Structure Knowledge Center (2015\u20132018)\n- **Scripps Research, JCSG**, Target Manager (2006\u20132015)\n- **Pfizer**, Principal Scientist, Structural Biology (2001\u20132006)\n- **Yale University School of Medicine**, Postdoctoral Fellow (1999\u20132001)\n\n**By the numbers**\n\n- \ud83e\uddec 400+ protein structures deposited in the PDB ([structures](https://marcdeller.com/structures/))\n- \ud83d\udcc4 60+ publications ([publications](https://marcdeller.com/publications/))\n- \ud83c\udfc6 7+ patents ([patents](https://marcdeller.com/patents/))\n- \ud83d\udc8a 6+ FDA IND filings\n- \ud83d\udc65 50+ scientists mentored, and two wet labs built from scratch\n\n**Expertise:** X-ray crystallography, cryo-EM, protein production, structure-based drug design and AI-driven therapeutics.\n\n**Education:** D.Phil., University of Oxford (1995\u20131999), with Prof. E. Yvonne Jones; BSc (First Class Honours) in Biochemistry and Molecular Biology, University of Leeds (1991\u20131995).\n\nThe full r\u00e9sum\u00e9 is on marcdeller.com: [download the PDF](https://marcdeller.com/wp-content/uploads/2026/07/marc_deller_resume_july_2026.pdf).", "Q: How selective are the tricyclic WRN compounds?\nA: Marc Deller writes that the compounds selectively cripple DNA repair in MSI-H/dMMR cancers while sparing healthy cells, because scientific screening has shown a pronounced dependence on WRN for the survival of MSI-H cancer cells while microsatellite-stable cancers are largely unaffected by loss of WRN.", "Q: How many of the JCSG Top96 conditions are PEG 3350 plus a 0.2 M salt?\nA: Twenty-one JCSG Top96 conditions, wells E12 through G08, consist of 20% (w/v) PEG 3350 with a 0.2 M salt, all drawn from the Hampton PEG/Ion screen. Together they account for a large share of the screen's PDB depositions, with F03 (calcium acetate, 14 structures), F08 (magnesium acetate, 12), G01 (potassium nitrate, 12) and F09 (magnesium chloride, 11) the most productive.", "Q: What is CCNE1 amplification and why does it matter here?\nA: CCNE1 is the gene encoding cyclin E1, the partner protein that activates CDK2. Amplification of CCNE1 raises cyclin E1 levels, driving CDK2-dependent proliferation, and it correlates with poor prognosis in ovarian, breast, gastric, and endometrial cancers. The paper Marc Deller co-authored tested its lead compound in CCNE1-amplified models (COV318 and OVCAR3), the tumor type most likely to respond to selective CDK2 inhibition.", "Q: Which validation metrics does Marc highlight in his multi-step LLM prompt post?\nA: Marc points to Ramachandran plots, MolProbity scores, stereochemical outlier counts, bond angle realism and steric clash checks as the key measures of whether a predicted model is trustworthy. He notes that modern LLM-based systems can report these metrics in enough detail that model quality often rivals what is seen in X-ray structures.", "Q: What future does Marc envisage for modular protein engineering?\nA: Marc imagines ordering a protein from a catalogue of functional domains, specifying their spatial arrangement, and having an AI deliver the final sequence, creating vast libraries of modular parts for custom enzymes, biosensors and therapeutics. He closes by asking readers what they would build with a molecular LEGO set; the post is at https://marcdeller.com/what-if-protein-structures-could-be-designed-like-lego-modular-approaches-in-2025/.", "Q: How accurate is the TopPDBLX parser?\nA: Against hand-labelled truth, both gold-set batches pooled, the rule parser alone reaches 95.1 % precision and 69.5 % recall, and rules plus the round-09 residual model reach 93.3 % precision and 90.3 % recall. Marc notes that the project brief anticipated a rule-based parser plateauing near 75 %, and that most of the difficulty turned out to be clause splitting rather than chemistry.", "Q: What is the evolutionary implication of the bacterial PH domain work?\nA: A conserved hydrophilic surface at the intersection of beta-strands of adjacent protomers likely mediates protein-protein interactions, and the structures allowed additional PHb families to be identified. Rather than being a eukaryotic innovation, the PH domain superfamily appears to have existed before prokaryotes and eukaryotes diverged.", "Q: Did Marc write about AI and protein folding pathways?\nA: Yes. In \"How AI Is Revealing the Secret Pathways of Protein Folding\", published 11 August 2025, Marc Deller discusses how AI is moving beyond predicting a protein's final structure to mapping the folding journey itself, combining machine learning with molecular dynamics simulations to chart routes that were previously invisible.", "Q: What does a crystallographer contribute to a medicinal chemistry program like this one?\nA: In the 2022 FGFR program the structural work Marc Deller contributed did two things. It tested the medicinal chemists' hypotheses \u2014 the FGFR2 structure showed the pyridyl ring was flipped relative to the docking model the team had been designing against \u2014 and it gave a physical explanation for an SAR observation, namely why a bulkier gatekeeper residue costs the compound potency. That kind of feedback is what turns iterative screening into structure-based drug design.", "Q: What is the Universal Micromanipulation Robot?\nA: The UMR is a flexible semi-autonomous crystal-harvesting workcell engineered by Square One System Design in Wyoming with design and development input from Marc Deller and Bernhard Rupp, and field-tested at the Joint Center for Structural Genomics at Scripps. Built around a St\u00e4ubli TX60 six-axis arm, it performs a three-step process of microtool selection, user-guided harvesting and automated cryoprotection, flash-cooling and dewar loading.", "Q: What is ALPHABETTI?\nA: ALPHABETTI, at alphabetti.mdeller.com, is Marc Deller's app that draws a protein's own single-letter amino acid codes in three dimensions, positioned and oriented along an ESMFold-predicted backbone. It is a sequence logo wrapped around the structure it describes, with the letters stacked at each position by how strongly a protein language model expects them.", "Q: What is the fused bicyclic pyrimidone WRN patent?\nA: Marc Deller's post \"2025 \u2013 Fused Bicyclic Pyrimidones as WRN Inhibitors\" covers US20250333420A1, an Incyte Corporation filing listing Shepard and colleagues, which discloses bicyclic pyrimidinone scaffolds designed to bind at the interface of the D1 and D2 helicase domains of WRN. Marc notes that the series was developed with contributions from himself as a structural biologist and colleagues at Incyte.", "Q: Which Boltz-2 utilities has Marc built?\nA: His Vibe Coding page lists AnalyzeBoltz, which harvests the sprawling Boltz2 output tree into a tidy summary, CSV, and CIF output folder; Boltz2Run, a minimal auditable wrapper where every run is logged, timed, and reproducible; Smiles2YAML, which turns a JSON of candidate SMILES into a folder of Boltz-2 input YAMLs; FoldForge, a single-file browser GUI for the whole Boltz-2 build, run, review workflow; and BoltzMaker.", "Q: What does Marc think the future of automated harvesting looks like?\nA: He concludes the way forward is not a single perfect device but hybrid systems combining precision sensing, gentle manipulation and integration with next-generation beamlines. He puts the stakes in numbers: structural genomics centres process hundreds of thousands of crystals a year, with roughly 100 crystals screened per successful structure, so eliminating late-stage loss matters. The post is at https://marcdeller.com/universal-mounting-robot-umr/.", "Q: What does DeepCovalent plan to do with its pipeline compounds?\nA: DeepCovalent states that it intends to advance early pre-clinical development of its compounds through at least in vitro pharmacology and receptor characterisation, and pharmacokinetics and absorption studies in animals including oral bioavailability. The aim is to drive compound valuation for potential licensing to, or acquisition by, pharma partners.", "Q: Why did the complex have to be deglycosylated before crystallisation?\nA: Crystallisation required partial deglycosylation with endoglycosidase H, on the rationale that EndoH preferentially cleaves the more exposed and flexible N-linked glycans that are not protected by PGT 135 binding. After treatment only single GlcNAc moieties were visible at Asn339, Asn362, Asn448, Asn262 and Asn276, whereas the glycans engaged by PGT 135 survived \u2014 GlcNAc2Man2 at Asn386, GlcNAc2Man6 at Asn332 and GlcNAc2Man8 at Asn392.", "Q: What did Marc Deller's 2024 CDK2 paper report?\nA: The 2024 Journal of Medicinal Chemistry paper \"Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\" describes a class of CDK2 inhibitors selective over CDKs 1, 4, 6, 7 and 9, with compound 17 reaching an IC50 of 0.29 nM. Marc Deller is a co-author who provided structural biology support at Incyte.", "Q: What therapeutic areas does Marc's home page say he has contributed to?\nA: The home page says Marc Deller has contributed to major advances in oncology and immunology, and that his 6+ FDA IND filings span oncology, immunology, and virology. Structure-based drug design is the common thread across those areas.", "Q: How did Marc Deller successfully purify oncostatin M after ion-exchange failed?\nA: Marc used gel filtration on a Superdex-75 column, which resolved OSM_\u0394188 as a distinct peak. The elution profiles showed a predominant peak corresponding to OSM_\u0394188 monomers (~25 kDa) and a minor peak corresponding to dimers (~50 kDa). Dimer formation was favoured in the presence of Triton X-100, and in the absence of detergent a minor contaminating band at 34 kDa of unknown identity was seen.", "Q: Where can I find Marc's bibliography of microplastics and cancer papers?\nA: The full grouped reference list of 28 papers from 2022 to 2026 is at the end of his post at https://marcdeller.com/microplastics-and-cancer-the-mounting-evidence/, published 16 June 2026.", "Q: What role was proposed for bacterial MACPF proteins?\nA: Genomic context analysis infers that Bth-MACPF may be involved in a novel protein-transport or nutrient-uptake system. Since these proteins were likely inherited from eukaryotes by horizontal gene transfer, the authors suggested an important role in the adaptation of commensal bacteria to the host environment.", "Q: What does PocketTide let you do?\nA: Survey GLP-1R structures in one organised dashboard, see key attributes per structure such as method, ligand and receptor state, place structures side by side to compare binding modes and conformations, and do all of it in a tool purpose-built around GLP-1R rather than a generic structure viewer.", "Q: What is the HIV-1 supersite of immune vulnerability?\nA: It is the Asn332 glycan-dependent antigenic region on the glycosylated face of HIV-1 gp120, defined by the 3.1 \u00c5 PGT135 complex structure. Marc Deller's post explains that its size and conservation make it a prime target for HIV vaccine immunogen design.", "Q: What is PocketTide?\nA: PocketTide is Marc Deller's browser-based dashboard for exploring structures of the glucagon-like peptide-1 receptor (GLP-1R), bringing the growing body of GLP-1R structural data into a single navigable view so structures, their key features and their differences can be surveyed at a glance.", "Q: Summarise Marc Deller's Protein Data Bank record.\nA: Marc Deller's individually credited PDB record is 78 X-ray crystal structures spanning 2000 to 2026, from human oncostatin M through the Joint Center for Structural Genomics era of bacterial and archaeal Pfam first representatives, the Scripps HIV and hepatitis C antibody complexes, Stanford's polysaccharide monooxygenase, polyketide synthase and split GFP structures, and Incyte's JAK, KRAS G12D, CDK2 and FGFR2 drug discovery structures. Counting the JCSG collective depositions, he has contributed to more than 400 PDB entries in total.", "Q: What did Marc Deller find about the conditions needed for 3C protease cleavage?\nA: Marc's cleavage optimisation, run as eight reactions and analysed by SDS-PAGE, showed that DTT is essential for efficient cleavage of fusion proteins by 3C protease. Reactions without DTT gave minimal cleavage products, whereas those with about 0.5 mM DTT showed loss of the OSM-196-GST band and appearance of cleaved OSM-196 at ~25 kDa. Cleavage worked at very low protease levels, typically a 1:1000 ratio of protease to fusion protein.", "Q: What is the weather today?\nA: I don't have weather information; I'm a chatbot about Marc Deller and his work. Ask me about his research, Elora Therapeutics or his apps and I'll do my best.", "Q: Where were the X-ray data for the gp120 supersite paper collected?\nA: Data on the unliganded PGT 135 Fab were collected at Advanced Photon Source beamline 23ID-B (wavelength 1.033 \u00c5), and data on the PGT 135\u2013gp120\u2013CD4\u201317b quaternary complex were collected at Advanced Light Source beamline 5.0.2 (wavelength 1.000 \u00c5), both at 100 K. Both datasets were processed and scaled with HKL-2000.", "Q: What version of PyMOL does Marc use?\nA: I don't have that level of detail on his setup. The tools he writes about are documented on marcdeller.com.", "Q: Give me relationship advice.\nA: That's not really my lane. I'm built to answer questions about Marc Deller, his research, and his ventures. If there's something about his career or work you'd like to know, I'm happy to help.", "Q: What is Marc Deller's most recent PDB deposition?\nA: The most recent entries are 10PI and 10PJ, both released on 10 June 2026: the JAK1 and JAK2 kinase JH1 domains in complex with povorcitinib. Their primary citation is the Journal of Medicinal Chemistry 2026 paper on the discovery of povorcitinib (INCB054707).", "Q: What is unusual about the putative carbohydrate-binding site in BT2081?\nA: The C-terminal beta-jelly-roll domain contains a potential carbohydrate-binding site that is highly conserved among BT2081 homologs and sits in the same position as the binding sites of structurally similar glycoside hydrolases. In BT2081 the site is partially occluded by surrounding loops, producing a deep solvent-accessible pocket rather than the shallower solvent-exposed cleft seen in the glycoside hydrolases.", "Q: Which virtual screening results does Marc report in the Click2Drug post?\nA: Marc reports that RosettaVS screens billions of compounds against diverse targets with hit rates of 14 to 44% in under seven days, and that PharmacoNet identified selective cannabinoid receptor inhibitors from a library of 187 million compounds in 21 hours on a single CPU. He also notes conformal prediction can reduce the number of molecules requiring explicit docking by three orders of magnitude, and cites VirtuDockDL for combining ligand- and structure-based screening with deep learning.", "Q: What is the PGT128/8ANC195 Env trimer paper?\nA: \"Complete epitopes for vaccine design derived from a crystal structure of the broadly neutralizing antibodies PGT128 and 8ANC195 in complex with an HIV-1 Env trimer\" (Biological Crystallography, 2015), led by Leopold Kong with Marc Deller as an author, resolved the BG505 SOSIP trimer with 8ANC195 and PGT128. It illuminated glycan shield dynamics, including gp41 interactions (N637 glycan accommodation) and glycan domino effects (N301 repositioning N262), completing trimer-level epitope blueprints for precision HIV vaccine engineering.", "Q: What software licences does Marc's lab hold?\nA: That isn't something I know. I can tell you which methods and tools he works with, but not the licensing.", "Q: What benchmark results does Marc highlight from the OpenBind release?\nA: Marc highlights that GNINA redocking succeeds on 85% of follow-on compounds while rigid cross-docking into apo or AlphaFold receptors falls below 5%, defeated by a single binding-site loop. Most strikingly, fine-tuning OpenFold3-p2 on the fragment-bound structures alone lifted cofolding success from 36% to 76%. He notes affinity prediction is harder still, with structure-aware models including Boltz-2 struggling to beat a molecular-weight baseline on the EV-A71 series.", "Q: What is the cryo-drip method?\nA: Cryo-drip is a procedure on the UMR system in which single drops of low-viscosity perfluoropolyether oil are automatically dripped from a solenoid valve onto the harvested crystal sitting in the microtool. Marc Deller's review describes it as a route to hyperquenching, since it removes excess mother liquor and coats the crystal in an insulating oil layer.", "Q: What is the condition in well H12 of the JCSG Top96?\nA: Well H12, the last well of the JCSG Top96 screen, is 0.2 M Sodium acetate, 30% (w/v) PEG 8000, 0.1 M Sodium cacodylate pH 6.5, a salt/PEG condition derived from Hampton Crystal Screen #28 (JCSG-III E09). It produced 5 PDB structures, and the recommended cryoprotectant is 15% ethylene glycol.", "Q: Which cancers is this FGFR work relevant to?\nA: The paper Marc Deller co-authored targets cancers driven by upregulated FGFR signaling. FGFR3 alterations are found most commonly in bladder cancer, while FGFR2 alterations occur mainly in cholangiocarcinoma, gastric cancer and uterine cancer, and the in vivo pharmacodynamic study used a human bladder carcinoma model.", "Q: Tell me about the drug discovery targets Marc worked on at Incyte.\nA: Marc Deller supported Incyte's structural biology across multiple programmes, including small-molecule inhibitors targeting KRAS-G12D, WRN, CDK2/4, JAK1/2 and FGFR2/3, contributing to more than six FDA IND filings and more than seven patents. This WRN patent, US 2025/0215006 A1, is one of the patents where his name appears as a co-inventor, and it showcases both the crystallography and the discovery of a new mechanism of action.", "Q: What is PDB entry 5UFV?\nA: 5UFV is the crystal structure of MtPMO3*, a cellulose-active fungal polysaccharide monooxygenase from Thermothelomyces thermophilus ATCC 42464, one of the PDB entries Marc Deller lists on his website. It was determined by X-ray diffraction at 2.45 \u00c5 and released on 15 March 2017, and the deposited model contains a copper ion as its unique ligand.", "Q: What is the CALR antibody patent Marc Deller is named on?\nA: Marc Deller is an inventor on the 2022 patent application \"Anti-mutant Calreticulin (CALR) Antibodies and uses thereof\" (WO2023107994A1), which arose from his antibody work on CALR at Incyte.", "Q: What phasing method did Marc use to solve the OSM structure?\nA: Marc Deller solved the OSM structure by multiple wavelength anomalous dispersion (MAD) using an ethylmercury phosphate (EMP) derivative. A three-wavelength MAD experiment was collected around the mercury LIII edge on a single propane-cooled crystal at the ESRF, and the MAD phases were supplemented with phases from an independent SIRAS experiment between the native and remote data sets.", "Q: What did the 2018 JACS polyketide synthase paper find?\nA: \"Structure-Function Analysis of the Extended Conformation of a Polyketide Synthase Module\", published in the Journal of the American Chemical Society in 2018, showed that the extended architecture of an assembly-line PKS module is fully competent for catalysis. A high-affinity human Fab from a naive antibody library was used to trap the extended conformation, verified by X-ray crystallography and tandem SEC-SAXS.", "Q: How many patents does Marc have?\nA: Marc is an inventor on 7+ patents, including the anti-mutant-CALR antibody patent, WRN inhibitor, KRAS G12D and CDK2 patent applications from Incyte, and the 2025 Elora Therapeutics provisional patent on enzymatic degradation of microplastics.", "Q: Why does kifunensine treatment abolish PGT 135 neutralisation?\nA: Adding kifunensine to the expression system limits glycan processing to Man9GlcNAc2, and PGT 135 lost neutralisation activity under those conditions. The crystal structure explains this: modelling an additional mannose residue onto the D2 arm of the Man8 glycan at Asn392 produces a severe steric clash with the antibody. By contrast PGT 135 still neutralised virus from cells treated with N-butyldeoxynojirimycin, because it does not contact the Asn332 glycan D1 arm.", "Q: Why didn't Marc use selenomethionine MAD phasing for OSM?\nA: Se-Met-derivatised OSM had exceptionally poor solubility and aggregated, even with 1 mM DTT added to limit oxidation, so it never crystallised. Marc Deller attributes this to the chemistry of selenium (the C\u2013Se bond is 0.14 \u00c5 longer than C\u2013S, giving greater polarity and susceptibility to oxidation) and to the fact that all four methionines in OSM are buried in the protein core with less than 50 % solvent accessibility, so substitution likely interfered with folding.", "Q: Where did Marc do his PhD?\nA: Marc did his D.Phil. at the University of Oxford (1995\u20131999) in the Laboratory of Molecular Biophysics and the Wellcome Trust Centre for Human Genetics, supervised by Prof. E. Yvonne Jones. His thesis, \"Structural Studies of Cytokines and Cytokine Receptors\", reported the first crystal structure of the cytokine oncostatin M.", "Q: Which precipitant classes performed best in the JCSG Top96 analysis?\nA: Marc Deller's analysis found above average performance for all PEG combination precipitant classes, that is, PEG used together with salts and/or organics, and below average performance for single precipitant conditions. As a result PEG combination precipitants are enriched in JCSG Top96 and JCSG Top96 Cryo at the expense of single precipitant classes.", "Q: What is Marc Deller doing now?\nA: As of September 2026, Marc Deller holds two roles: Chief Scientific Officer and co-founder of Elora Therapeutics (since June 2025), a pre-seed startup developing a first-in-class enzyme therapeutic to degrade microplastics in the human body, and Scientific Technical Advisor to DeepCovalent Inc. (since October 2025), where he advises on physics-based computer-aided drug discovery (CADD) pipelines and GPCR-targeting small-molecule discovery platforms.", "Q: What were the actual numbers for the Cys727 mutants?\nA: In the helicase assay, C727A gave observed activities of 4.619 and 1.452 AU/min, or ratios of 0.07 and 0.02 relative to wild-type at 64.77 AU/min. In the ATPase assay the same mutants gave 22.65 and 25.42 AU/min, ratios of 0.29 and 0.33 relative to wild-type at 78.17 AU/min \u2014 reduced but clearly not abolished, which is what \"decoupling\" means here.", "Q: What does AlphaFraud's live site show?\nA: AlphaFraud, at alphafraud.mdeller.com, watches the PDB for freshly deposited human structures and pits each against its blind AlphaFold prediction as a continuous public test of whether protein folding is really solved. Its front page reports 82,901 structures analysed, 1,948 confidently wrong (a 2.3 % rate), 431 both novel and confidently wrong (0.5 %), and 459 release weeks covered.", "Q: Does Marc write about generative protein design?\nA: Yes. In his creativity post, Marc Deller writes that creativity today means designing proteins nature never built, using generative models to invent function rather than discover it. He frames this as the current frontier of the creative approach he has taken throughout his career.", "Q: What is BcYkfC?\nA: BcYkfC is a YkfC ortholog from Bacillus cereus, a gamma-D-glutamyl-L-diamino acid endopeptidase whose structure was determined at 1.8 \u00c5 resolution in complex with the reaction product L-Ala-gamma-D-Glu. Marc Deller co-authored the 2010 Acta Crystallographica Section F paper as part of the Joint Center for Structural Genomics.", "Q: What was Marc Deller's role in the 2022 FGFR paper?\nA: Marc Deller contributed the structural biology to the 2022 FGFR2/3 paper as part of Incyte's structural biology group. Together with Hsiang-Ting Lei and Leslie B. Epling, he determined the X-ray co-crystal structure of compound 30 bound to FGFR2, which was deposited in the Protein Data Bank as entry 8E1X and explained why the lead compound lost some potency against the V555L and V555M gatekeeper mutants.", "Q: What was Marc Deller's role in the Protein Structure Initiative?\nA: Marc Deller was a Senior Scientist at the Joint Center for Structural Genomics (JCSG) at Scripps from 2006 to 2015, one of the large-scale centers of the NIH-sponsored Protein Structure Initiative. He developed automated crystallography sample-prep, data-collection and gene-to-structure workflows, raised success rates by 10% through automation and target-salvage pathways, and determined over 70 novel structures.", "Q: What is Elora Therapeutics valued at?\nA: I don't know, and a valuation isn't something I'd guess at. I can tell you what the company is working on scientifically. For anything commercial, marc@marcdeller.com.", "Q: What is BVU2987?\nA: BVU2987 from Bacteroides vulgatus is a member of the DUF2874 domain family, Pfam PF11396, whose members are predominantly found in microbes of the human gut and oral cavity. Marc Deller co-authored the 2010 Acta Crystallographica Section F paper reporting its crystal structure.", "Q: How accurate are crystal-detection algorithms?\nA: The review reports that despite work using self-organizing neural nets, machine learning, Fisher linear discriminant analysis, Bayes object classification, decision trees and support vector machines, error rates of around 15% remain common. Marc Deller and Bernhard Rupp judge that insufficient for routine high-throughput use or for full autonomy.", "Q: What oligomer does ECX21941 form?\nA: ECX21941 assembles as a homopentamer both in solution and in the crystal when expressed in Escherichia coli, and it represents the first pentameric structure for the Sm/LSm family, although the oligomeric form in vivo is not known.", "Q: Who is Marc Deller's most frequent co-depositor at Incyte?\nA: L.B. Epling appears as a deposition author on all six of the Incyte-era entries in Marc Deller's list: 10PI, 10PJ, 9E5D, 9E5F, 8UV0 and 8E1X. Marc is named alongside Epling on 9E5D, 9E5F, 8UV0 and, with H.-T. Lei, on 8E1X.", "Q: How do REACH and UMR compare?\nA: Marc Deller's review notes both are fully integrated, actively developed systems sharing automated tape opening, semi-autonomous user-assisted harvesting, automated cryoprotection and automated cooling. The key difference is that REACH uses microgrippers, offering easier and more flexible pickup, and is coupled directly to an X-ray source, whereas UMR uses microtools compatible with existing flash-cooling protocols.", "Q: Which amino acid differences between JAK1 and JAK2 explain the selectivity?\nA: The structures reported in the paper Marc Deller co-authored highlight that JAK1 has His885 pointing toward solvent, whereas JAK2 has Asn859 bent toward the binding pocket where it forms a hydrogen bond with Glu1015 and partially occludes the ligand site. JAK2 also has a phenylalanine, Phe860, that may orient the inhibitor's phenyl ring into a \"closed\" disposition akin to the DFG pocket, further disfavouring povorcitinib binding.", "Q: In plain language, what did Marc's OSM structure tell us about how the cytokine talks to its receptors?\nA: Marc Deller's thesis showed that oncostatin M is a tightly packed bundle of four helices, and that the surface patches it uses to grab its two signalling receptors are on different parts of the bundle: a phenylalanine\u2013lysine pair at the start of helix D (site 3) is the main handle for the LIF receptor, while a patch spanning helices A and C (site 2) is the handle for gp130. Unlike growth hormone or erythropoietin, whose receptor contacts are dominated by greasy hydrophobic pockets, OSM uses a mixed patchwork of polar and hydrophobic residues, which may let it be more selective about which receptors it pairs with.", "Q: Does the antibody actually block the mutant CALR\u2013MPL interaction?\nA: Yes. In an HTRF competition assay using recombinant His-tagged MPL and FLAG-GFP-tagged mutant CALR Type 1, anti-mutCALR antibodies 6 and 32 fully inhibited the MPL\u2013mutCALR interaction, whereas the non-functional CAL2 antibody only partially inhibited it at concentrations above 100 nM. A separate assay in JAK2-knockout HAP1 cells showed that clones 55, 65, 68, 74, 134, 184 and 188 inhibit MPL dimerisation.", "Q: What is PXDesign, according to Marc's blog?\nA: In \"From Noise to Nanomolar: PXDesign's New Benchmark for Protein Binder Design\", published on 23 April 2026, Marc Deller describes PXDesign as a modular de novo binder design suite released in August 2025 by the ByteDance Seed team behind Protenix, their open-source AlphaFold3 reproduction. It integrates two complementary generative strategies, diffusion and hallucination, with a rigorous multi-predictor filtering framework.", "Q: What does the second bicyclic WRN patent from December 2025 describe?\nA: Marc Deller's post \"2025 \u2013 Bicyclic Compounds as Inhibitors of WRN\" covers US20250333419A1, an Incyte Corporation application listing Barbosa and colleagues, which discloses a series of bicyclic compounds of Formula (I) engineered to inhibit both the ATPase and helicase activities of WRN.", "Q: What is the earliest Joint Center for Structural Genomics paper in Marc Deller's list?\nA: The earliest JCSG entries date from 2007 and include the Bacillus cereus homoserine O-succinyltransferase structure at 2.4 \u00c5 and two papers on dye-decolorizing peroxidases, all published in Proteins. In these early papers his name appears as \"Deller M\" or \"Deller MC\".", "Q: What is Marc Deller's 2014 paper on automated crystal harvesting about?\nA: \"Approaches to automated protein crystal harvesting\", by Marc C. Deller and Bernhard Rupp, is a review of the technologies available for automating the step that moves a protein crystal from its growth drop into the X-ray beam. It argues that harvesting is the last remaining manual step in an otherwise automated structure-determination pipeline and surveys fourteen semi-automated systems built to close that gap.", "Q: Which protein engineering tricks for GPCR crystallography does Marc describe?\nA: Marc describes replacing the flexible third intracellular loop with T4 lysozyme or thermostabilized apocytochrome b562RIL (BRIL) to add polar surface for crystal contacts and reduce conformational heterogeneity, including modified T4L variants that improved the M3 muscarinic receptor from 3.4 \u00c5 using 70 crystals to 2.8 \u00c5 with untwinned crystals. He also covers conformational thermostabilization, citing the turkey \u03b21-adrenergic receptor where six point mutations raised the melting temperature by 21\u00b0C and locked the antagonist-bound state, plus antibody fragment and nanobody incorporation.", "Q: Explain the CALR antibody patent in plain English.\nA: In some blood cancers a gene called CALR is mutated so that its protein grows an abnormal, positively charged tail. That tail lets the protein escape from where it belongs, latch onto a receptor called MPL on blood-forming cells, and switch it on permanently, driving the cells to over-multiply. Incyte's patent, on which Marc Deller is a co-inventor, describes antibodies engineered to grab that abnormal tail and block it, so the receptor is no longer switched on. The patent shows the antibodies shrink spleens, normalise platelet counts and prolong survival in mice, and work on cells taken from real patients.", "Q: What is the home page of marcdeller.com about?\nA: The home page of marcdeller.com is titled \"Hidden Architecture: Mapping Biology's Structure\" and introduces Marc Deller as an expert structural biologist and drug discovery leader with over 10 years of experience. It summarises his specialisation in protein structure, X-ray crystallography, Cryo-EM, protein production, mechanisms of action, receptor biology, molecular pharmacology, automated research platforms, and AI-driven therapeutics.", "Q: What is structure-enabled drug discovery as Marc describes it in his patents?\nA: The Patents page describes innovative structure-enabled drug discovery as structure-guided discovery platforms integrating X-ray crystallography, cryo-EM, and computational design, enabling rational drug development for difficult-to-drug targets in oncology and immunology.", "Q: Why is Gly120 so important in oncostatin M?\nA: Marc Deller's paper showed that Gly120 in helix C of oncostatin M forms a hydrophobic 'dimple' in the surface of the cytokine into which the exposed Phe169 of gp130 is proposed to dock. Replacing Gly120 with alanine reduced gp130 binding more than 100-fold, and replacing it with the bulkier tyrosine found at the equivalent position in CNTF almost completely ablated the interaction and reduced biological activity more than 10,000-fold, consistent with a bulky side chain occluding the dimple.", "Q: What does Marc say about figures in structural biology?\nA: Marc Deller writes that in structural biology the figure does the work, because a single well-rendered view says what paragraphs of text cannot. He treats clarity and accessibility as central to turning structures into discoveries that others can build upon.", "Q: Why did Marc consider the GCSF\u2013GCSFR complex a good model for OSM\u2013gp130?\nA: Marc Deller gave four reasons: GCSF and GCSFR are structurally very similar to OSM and gp130 respectively, and the A-B loop of GCSF matches OSM better than Epo does; the Phe/Lys pair used by LIFR-binding cytokines is partially conserved in GCSF (Phe 144/Lys 147) and is important for GCSFR binding; the key GCSFR binding determinants (Ile 239 and Arg 288) lie on the B-C and F-G loops of the C-terminal domain, as in gp130, rather than on the B-C and E-F loops as in EpoR; and GCSF, like gp130 cytokines, forms higher-order complexes (1:1, 2:2 and 4:4). The limitation is that the GCSF\u2013GCSFR crystal structure only shows a site 2-like major interface and a novel minor interface, so it can model only site 2.", "Q: What is AXIS, as described in Marc Deller's post \"AXIS: Closing the Loop on Crystal Detection\"?\nA: AXIS is an AI-based Crystal Identification System developed at the European Molecular Biology Laboratory in Grenoble, described by Personnaz and colleagues in IUCrJ in 2026. In his post of 14 April 2026, Marc Deller explains that it automates the visual identification of crystals from imaging plates, the one step in high-throughput crystallography that had persistently resisted automation.", "Q: What does Marc say cryo-EM's improving detectors and image processing enable?\nA: Marc says advances in detector technology and image processing let researchers capture not just static pictures but multiple conformational states, revealing dynamic movements crucial to protein function and drug interaction. He notes cryo-EM is being rapidly integrated into drug discovery pipelines; the post is at https://marcdeller.com/cryo-em-and-the-new-frontier-in-membrane-protein-drug-design/.", "Q: Has Marc trained in project management?\nA: Yes, his My Learning page lists a 2025 \"Foundations of Project Management\" certificate and describes advanced training in leadership and adaptability for managing complex scientific programs.", "Q: What clinical results does Marc report for the KRAS-G12D inhibitor?\nA: At the 1,200 mg daily dose INCB161734 achieved a 34% objective response rate in relapsed pancreatic cancer with an 86% disease control rate, and 72% of patients at that dose had a greater than 90% reduction in circulating tumour DNA KRAS-G12D variant allele frequency. KRAS-G12D occurs in roughly 45% of pancreatic ductal adenocarcinomas and 15% of colorectal cancers.", "Q: Is povorcitinib selective against other kinases besides JAK2?\nA: Povorcitinib was inactive against JAK3 with an IC50 greater than 10 \u03bcM and had only negligible potency toward TYK2 at 849 nM. Full kinome profiling showed the novel dimethylpyrazole hinge motif was highly selective for the JAK family, with only one other kinase, Casein Kinase 1 Gamma 1, inhibited by more than 50% at a 100 nM compound concentration.", "Q: Did Marc Deller write about validating ligands in crystal structures?\nA: Yes. \"Validation of Protein-Ligand Crystal Structure Models: Small Molecule and Peptide Ligands\" was published in Methods in Molecular Biology in 2017 by Pozharski E, Deller MC and Rupp B. It argues that because bound ligands comprise only a small fraction of the total X-ray scattering mass, particular care is needed to validate the ligand model.", "Q: Which virus\u2013receptor complexes were discussed?\nA: The review discussed single-particle cryo-electron microscopy of adenovirus coated with its soluble integrin receptor, which gave a hazy view at about 21 \u00c5 resolution, and the more informative rhinovirus\u2013ICAM-1 complex, in which crystal structures of rhinovirus and the two N-terminal domains of ICAM-1 were fitted into a cryo-EM reconstruction. Marc Deller and Yvonne Jones noted that the cryo-EM approach becomes truly powerful when coupled with high-resolution structures of isolated components.", "Q: Where was in situ screening available at the time of the review?\nA: Marc Deller and Bernhard Rupp list in situ screening of SBS-format plates as available at the ESRF, the Swiss Light Source, Diamond Light Source, Helmholtz-Zentrum Berlin and the Australian Synchrotron. They also note in-house options such as the PX scanner, PlateMate and a dedicated lipidic cubic phase plate holder, though 10 to 20 minute exposures limit these to pre-screening.", "Q: What does GPCR Binning sit alongside?\nA: Marc places it next to PocketTide, his GLP-1R structure dashboard, and the Boltz-2 co-folding utilities, as part of a structure-based approach to drug discovery across GPCR and other target classes. The post is at https://marcdeller.com/structural-binning-and-organisation-utilities-for-gpcr-datasets/.", "Q: Which are Marc's latest vibe coding projects?\nA: The Vibe Coding page names AlphaFraud, BoltzMaker, ChemSage, and MicroPlastics Blaster as the latest projects. AlphaFraud has its own site at alphafraud.mdeller.com.", "Q: What light-chain mutants were made and why?\nA: Five 10E8 light-chain mutants were designed to test whether the light chain faces the viral membrane. Mutant 1 reverted three somatically mutated arginines to germline (R17Q, R24Q, R70T); mutants 2 and 3 progressively replaced basic and polar surface residues with aspartate or glutamate; mutant 4 changed R29 and Y32 to glutamate; and mutant 5 changed R29 and Y32 to alanine. Mutants 1\u20133 coated the presumed membrane-facing surface with increasing negative charge, while mutants 4 and 5 were designed to destroy the lipid-binding site itself.", "Q: Where were the X-ray data for the TM1088 structures collected?\nA: X-ray data were collected at the Advanced Light Source (ALS) at Lawrence Berkeley National Laboratory, on beamline 5.0.1 for TM1088A and beamline 5.0.3 for TM1088B and the TM1088 complex, all at a wavelength of 1.000 \u00c5 with crystals flash-cooled to 100 K. Initial diffraction screening of TM1088A used the Stanford Automated Mounting system (SAM) at SSRL.", "Q: Explain Marc Deller's job to a non-scientist.\nA: Marc Deller is a scientist who works out the three-dimensional shapes of proteins, the molecular machines in our bodies, using X-rays and electron microscopes. Knowing a protein's shape lets chemists design drugs that fit into it precisely, so Marc has spent 20+ years helping pharmaceutical companies like Pfizer and Incyte turn protein structures into cancer and immunology medicines. He now co-leads a startup, Elora Therapeutics, that is engineering enzymes to break down microplastics inside the human body.", "Q: What is the My Learning page on marcdeller.com about?\nA: The My Learning page, subtitled \"Continuous Growth: Courses Shaping My Expertise\", collects Marc Deller's certifications in AI, vibe coding, project management, and data analysis. He says they demonstrate adaptability and expertise with cutting-edge technologies.", "Q: What monomer release did the KTH study achieve in human serum?\nA: Marc reports that starting from zero, around 500 \u00b5M of total monomers was released within the first 15 minutes of incubation in real human serum, rising to approximately 850 \u00b5M by 60 minutes \u2014 a rate the authors judged comparable to that seen in buffer. In simulated body fluid at 10 mg/mL substrate, the enzyme released roughly 2,100 \u00b5M of TPA and over 4,100 \u00b5M of MHET.", "Q: What does a cryo-EM facility cost?\nA: Marc writes that each advanced cryo-EM setup can exceed $7 million, and that facility modifications and support infrastructure often double that figure.", "Q: What news and recognition are highlighted in Marc Deller's portfolio?\nA: Marc Deller's portfolio highlights his mentorship of Stanford undergraduates who won a Lemelson-MIT Student Prize for combating antibacterial resistance (Stanford News, April 2017), a Stanford Medicine News feature titled \"Your one-stop shop for producing, crystallizing biomolecules\" (April 2016), and the inclusion of his protein stability paper in the IUCr \"Top articles in structural biology (Spring 2020)\" collection.", "Q: What did Marc Deller contribute to the hepatitis C virus AP33 antibody structure?\nA: Marc Deller is a co-author on the 2012 Journal of Virology paper \"Structure of hepatitis C virus envelope glycoprotein E2 antigenic site 412 to 423 in complex with antibody AP33\", by Kong L, Giang E, Nieusma T, Robbins JB, Deller MC, Stanfield RL, Wilson IA and Law M. The crystal structure showed the broadly neutralizing antibody AP33 bound to a peptide from HCV E2 antigenic site 412 to 423.", "Q: What is Marc's speaking fee?\nA: That's not something I know. If you're looking to invite him to speak, the best thing is to write to marc@marcdeller.com and ask him directly.", "Q: How does povorcitinib tell JAK1 apart from JAK2?\nA: Its 3,5-dimethylpyrazole hinge binder makes the same hydrogen bonds with the backbone of Leu932 and Glu930 in both isoforms, so selectivity comes from further out: in JAK1 the glycine-rich P-loop rides about 2 \u00c5 higher, opening a shallow lipophilic groove that accepts the terminal trifluoro-isopropyl amide, whereas in JAK2 the loop closes as Asn859 swings inward to hydrogen-bond with Glu1015 and a nearby Phe860 rotates the pocket shut.", "Q: What is Site Health Inspector?\nA: Site Health Inspector is Marc Deller's pure-Bash, colour-coded website health audit. `SiteHealthInspector.sh` runs a diagnostic scan against any URL and reports timing metrics, HTTP status, caching behaviour, security headers, SSL/TLS certificate status, DNS records, CMS fingerprinting and an overall health scorecard, with no runtime dependencies beyond standard Unix tools.", "Q: How does chatRDKit try to stay safe?\nA: It refuses snippets containing imports of `os`, `sys`, `subprocess`, `shutil`, `socket` or `requests`, plus `open(`, `eval(`, `exec(`, `__import__` and direct `os.`/`sys.` attribute access, and the system prompt also instructs the model not to use them. Marc describes this as a best-effort net rather than a sandbox, since the code runs in-process. The repo is bellcheddar/chatRDKit and the post is at https://marcdeller.com/talk-to-rdkit-in-plain-english-cheminformatics-without-the-boilerplate/.", "Q: What made the (4-pyrazolyl)-2-aminopyrimidine CDK2 inhibitors special?\nA: They achieved unprecedented selectivity for CDK2, and the co-crystal structure with Compound 17 revealed the molecular basis of that selectivity. Marc Deller's post explains that the structural insights guided selectivity over CDKs 1, 4, 6, 7 and 9.", "Q: What was Marc doing when the polysaccharide monooxygenase structure was released?\nA: The MtPMO3* structure was released in March 2017 and published in 2017, during the period when Marc Deller was Director and Founder of the Macromolecular Structure Knowledge Center at Stanford ChEM-H, Stanford University, a role he held from June 2015 to March 2018. He contributed the structural biology and crystallography behind the entry.", "Q: Why did the team put deuterium into the lead compound?\nA: Compound 27 still inhibited CYP3A4 with an IC50 of 9 \u03bcM, and the team suspected the culprit was the N-methyl amide carbonyl or a primary amide metabolite formed by N-demethylation. Replacing the N-methyl with a deuterated N-CD3 group gave compound 29, which had a clean CYP profile (CYP3A4 IC50 > 25 \u03bcM) and dramatically reduced intrinsic clearance (0.5 versus 1.0 L/h/kg), confirming N-demethylation as the likely site of metabolism.", "Q: How does crystallography help drug discovery in a project like this?\nA: Crystallography provides an atomic-resolution picture of exactly how a drug candidate sits in its target's binding pocket, which is what Marc Deller and Leslie Epling delivered for compound 17 bound to CDK2 (PDB 8UV0). That structure showed the hinge-binding interaction, the trifluoromethyl group pointing at the Phe80 gatekeeper, and specific hydrogen bonds, all of which rationalize why the compound is so potent and selective and guide further medicinal chemistry.", "Q: Which publication describes the 10E8 MPER scaffold complex?\nA: The structure was published in Irimia, A., et al. (2017), \"Lipid Interactions and Angle of Approach to the HIV-1 Viral Membrane of Broadly Neutralizing Antibody 10E8: Insights for Vaccine and Therapeutic Design\", PLoS Pathogens, volume 13, article e1006212. Marc Deller has a separate blog post about that paper.", "Q: What is Marc Deller's role in this patent?\nA: Marc C. Deller is the sole named inventor on the Elora Therapeutics provisional patent. He is also the Chief Scientific Officer and a co-founder of Elora Therapeutics, the company behind the filing, making it a foundational piece of intellectual property for the venture he helped create.", "Q: Why did the team crystallize FGFR2 rather than FGFR3?\nA: The program was driven by FGFR3 assays, but as the paper Marc Deller co-authored explains, the FGFR2 isoform was more amenable to crystallization than FGFR3. Compound 30 was confirmed to have identical biochemical potency in both isoforms, so the FGFR2 co-crystal structure (PDB 8E1X) was a valid surrogate for understanding binding in FGFR3.", "Q: How does GOBSMACKED handle compute?\nA: With a two-step Prepare/Analyze design: the user downloads a pixi run bundle, runs the compute locally, then uploads the results for analysis, because Marc's small droplet cannot carry the docking and MD workloads. An optional stage adds Boltz-2 affinity-only scoring of the post-MD complex.", "Q: What is Marc's own take at the end of the DRT3 post?\nA: Marc writes that whether DRT3 formally violates the central dogma is debatable, since the poly(GT/AC) product is repetitive, non-coding and does not appear to enter the bacterial genome, and that how the duplex DNA confers anti-phage immunity remains open, with sequestration of viral replication factors and unusual DNA secondary structures both proposed. His broader point is that in mechanistic enzymology atomic structures are not confirmatory data \u2014 they are the experiment. The post is at https://marcdeller.com/when-the-enzyme-is-its-own-blueprint-protein-templated-dna-synthesis-in-drt3/.", "Q: How was the oncostatin M structure solved?\nA: Marc Deller solved the oncostatin M structure by multiple-wavelength anomalous dispersion (MAD) phasing using a single ethylmercuryphosphate (EMP) heavy-atom derivative, with data collected at three wavelengths around the mercury LIII edge on beam line BM14 at the ESRF. MAD phases to 2.6 \u00c5 were combined with SIRAS phases from the derivative and native data, improved by solvent flattening in SOLOMON, and the model was traced in O and refined with CNS-SOLVE against native data to 2.2 \u00c5.", "Q: What market and clinical numbers does Marc cite for nanobodies?\nA: Marc reports the nanobody market reached $602.98 million in 2024 with projections of $3.3 billion by 2033, a 20.9% compound annual growth rate. He names Caplacizumab for thrombotic thrombocytopenic purpura, Ozoralizumab for rheumatoid arthritis and Ciltacabtagene autoleucel for multiple myeloma as approved nanobody-based therapeutics.", "Q: What is the point of Marc's vibe-coding?\nA: Marc's aim is to show what a domain expert can build with AI-assisted coding on modest hardware at minimal cost, turning structural-biology and drug-discovery ideas into working tools in days rather than months.", "Q: What surprising finding about protective TYK2 variants does Marc highlight?\nA: Marc highlights that human variants associated with reduced autoimmune risk work by decreasing TYK2 protein levels rather than simply reducing enzymatic activity. He argues this challenges the assumption that inhibition is the only route and points to targeted protein degradation, in the vein of PROTAC technologies, as an underexplored therapeutic modality for TYK2.", "Q: What does the Structures page on marcdeller.com describe?\nA: The My Structures page, subtitled \"Protein Structures: My Atomic-Level Discoveries\", presents 400+ protein structures as evidence of advances in structural biology, medicinal chemistry, and immunology. Marc Deller says they showcase his expertise in construct design, protein crystallisation, X-ray crystallography, and structure determination.", "Q: Why does the patent argue WRN inhibitors are needed if checkpoint inhibitors already work in MSI-H cancer?\nA: The background notes that checkpoint inhibitors such as pembrolizumab are approved in MSI-H/dMMR tumours but achieve only about a 33.3% objective response rate, and resistance develops. WRN inhibitors act by a completely different mechanism, and WRN sensitivity is retained in preclinical MSI-H colorectal cancer models that are refractory to targeted therapy, chemotherapy and immunotherapy, so they could complement rather than duplicate checkpoint blockade.", "Q: What is generate_yaml.py?\nA: generate_yaml.py, published as Smiles2YAML, is Marc Deller's command-line utility for generating Boltz-2 co-folding input YAML files from a batch of candidate SMILES stored in a JSON file. It is designed for the output of generative chemistry runs such as fragment-based screening or de novo design.", "Q: What are BtDyP and TyrA?\nA: BtDyP from Bacteroides thetaiotaomicron strain VPI-5482 and TyrA from Shewanella oneidensis are dye-decolorizing peroxidases, members of a family of heme-dependent peroxidases identified in fungi and bacteria. Marc Deller co-authored the 2007 Proteins paper reporting their crystal structures at 1.6 and 2.7 \u00c5 resolution.", "Q: Where did Marc Deller do his undergraduate degree?\nA: Marc Deller did his B.Sc. in Biochemistry and Molecular Biology at the University of Leeds in the United Kingdom, from August 1991 to August 1995. He graduated with First Class Honours (the US equivalent of an A, or 70\u2013100%).", "Q: How did DRVIA7 differ structurally from VRC01?\nA: DRVIA7 approached gp120 at essentially the same angle as VRC01 and made nearly identical protein contacts \u2014 8 of 11 HCDR2 residues and 3 of 4 HCDR3 contact residues were conserved. The critical differences were in the light chain: two hydrophobic residues in the DRVIA7 LCDR1 (I29 and W32) shifted and twisted the first GlcNAc of the N276 glycan by 5.5 \u00c5 and 30\u00b0, and the DRVIA7 light-chain N terminus clashed with the V5 loop glycan, whereas VRC01 has a 2-amino-acid LCDR1 deletion and a tyrosine that steers the glycan away.", "Q: Where was the SusD-family structure BT_3984 published?\nA: It appeared in Acta Crystallographica Section F, Structural Biology and Crystallization Communications, in 2010, with Bakolitsa C as first author and Marc Deller among the Joint Center for Structural Genomics co-authors contributing crystallography (DOI 10.1107/s1744309110032999).", "Q: Did Marc Deller work on other Incyte drug programmes besides JAK1?\nA: Yes. Alongside the JAK1 povorcitinib structures 10PI and 10PJ, Marc Deller is a deposition author or primary-citation author on other Incyte structures, including the KRAS G12D inhibitor structures 9E5D and 9E5F at 1.36 and 1.35 \u00c5, the CDK2 inhibitor structure 8UV0 at 1.55 \u00c5, and the FGFR2 kinase domain structure 8E1X at 2.68 \u00c5. His Incyte work centred on structural biology support for small-molecule drug discovery.", "Q: How do the PANTS source environments compare?\nA: PANTS reports the human gut at 12,584,458 sequences read and 311 candidates (about 25 per million), compost at 1,020,575 sequences and 69 candidates (about 68 per million), the marine plastisphere at 737,027 sequences and 44 candidates (about 60 per million), landfill at 436,229 sequences and 15 candidates (about 34 per million), and wastewater at 26,631 sequences and no candidates. Every marine plastisphere candidate sits below 40 % identity to anything characterised, so nothing in that cohort is a rediscovery, whereas compost returns near-identical copies of known enzymes.", "Q: What did the cytotoxicity testing show?\nA: Marc reports that human HEK293 kidney cells were exposed to S238A PETase at 1.5, 3.0 and 6.0 \u00b5g/mL across four cell densities from 1\u00d710\u00b3 to 1\u00d710\u2076 cells per well using an MTT assay, and relative viability remained at or above 100% in all conditions. He notes this matters because PETase is an esterase and could in principle act on membrane lipids, so the absence of any signal indicates narrow substrate specificity.", "Q: What is Marc's chess rating?\nA: I don't have that, and I don't believe chess is among his hobbies. marcdeller.com lists the ones he does write about.", "Q: Summarise Marc Deller's post on the Spacial Score.\nA: In \"Spacial Score\u2500A Comprehensive Topological Indicator for Small-Molecule Complexity\", published on 13 October 2025, Marc Deller explains a metric introduced by Krzyzanowski and colleagues in the Journal of Medicinal Chemistry that captures molecular complexity more finely than Fsp3 or FCstereo. The Spacial Score combines four atom-level terms \u2014 hybridization, stereogenicity, non-aromatic ring membership and branching \u2014 into a single granular indicator that tracks with biological activity and helps guide synthesis planning.", "Q: What was Marc Deller's contribution to the MtPMO3* paper?\nA: Marc Deller contributed the crystallography to the 2017 MtPMO3* study: the crystallization, X-ray data collection at SSRL beamline 12-2, and structure determination of the wild type enzyme at 2.45 \u00c5 resolution, deposited in the PDB as entry 5UFV. Marc was the sole structural biologist among the authors, alongside the Marletta lab biochemists and the Britt lab EPR spectroscopists.", "Q: What effect did 1B2 have on chain elongation kinetics?\nA: Using an assay with a diffusible mimic of the diketide substrate, chain elongation by module 2 + TE gave kcat values of 4.1 \u00b1 0.2 min\u22121 with 1B2 and 4.3 \u00b1 0.2 min\u22121 without, and KM values of 1.6 \u00b1 0.2 mM and 2.0 \u00b1 0.3 mM respectively. The extended conformation is therefore fully competent for chain elongation, and the data also suggested that AT-catalyzed transacylation and KR-catalyzed \u03b2-ketoreduction do not require the arch-shaped conformation.", "Q: Did Marc write about building labs from scratch?\nA: Yes. His collaboration post links to a companion piece on building a comprehensive protein production, X-ray crystallography and cryo-EM laboratory from scratch, and argues that collaboration underpins every enduring success in modern structural biology. The collaboration post itself is at https://marcdeller.com/collaboration-2-state-of-the-art-structural-biology-facilities-built-through-global-partnerships/.", "Q: What adversarial attacks does Marc Deller survey in \"How to Confuse Machine Learning\"?\nA: In that 27 August 2025 post, Marc Deller surveys adversarial examples, data poisoning, model extraction, prompt injection, Byzantine attacks in distributed learning, and gradient masking as a defense that may provide false security. He organizes them by the phase of the machine learning lifecycle they exploit.", "Q: What is \"site 3\" in oncostatin M?\nA: Site 3 in oncostatin M, as described by Marc Deller and colleagues, is the epitope at the N-terminal start of helix D that engages the second signalling receptor, either LIFR or OSMR. It consists of a protruding, conserved phenylalanine/lysine pair, Phe160 and Lys163. Mutating either residue to alanine, or both together, left gp130 binding intact but abolished LIFR binding and caused a dramatic loss of biological activity on both LIFR/gp130- and OSMR/gp130-dependent Ba/F3 cells.", "Q: Summarize the headline finding in one sentence.\nA: The 2016 Immunity paper co-authored by Marc Deller showed that although a Chinese HIV donor's immune system could rapidly mature VRC01-like antibody heavy chains and readily encoded the VRC01 light-chain CDR3 signature, the N276 and V5 glycans on gp120 blocked the light chain from maturing properly, stalling the lineage before it reached broad neutralisation.", "Q: Which ALPHABETTI mode uses an alignment?\nA: HOGWASH is the exception in ALPHABETTI: it counts what evolution actually did in a column of aligned homologues, which is the older and better-established way to measure conservation, and both approaches are in the app so they can be compared. The numbers and the drawing are WebLogo 3's, MIT licensed and used unmodified, and Marc asks users to cite Crooks, Hon, Chandonia and Brenner (2004), Genome Research 14:1188-1190.", "Q: Why does Marc favor running LLMs locally rather than in the cloud?\nA: Marc cites privacy, since sensitive files never leave the device; speed, with no network latency; customization, since you can swap in models such as Ollama, Llama or Gemma and your own embedders; cost predictability compared with cloud subscriptions and egress charges; and offline access. He allows that cloud offerings are easier to start with but says they fall short where privacy, customization or cost management matter most. The post is at https://marcdeller.com/unleashing-the-power-of-anythingllm-running-large-language-models-locally/.", "Q: What is PANTS?\nA: PANTS stands for PETase ANnotation and Triage System. It mines metagenomic assemblies for uncharacterised PETase-like and MHETase-like enzymes using an ESM-2 classifier trained on characterised cutinase/esterase families and PAZy entries, and returns a ranked candidate list with the catalytic triad and oxyanion hole flagged. It feeds Marc's enzyme-discovery work at Elora Therapeutics.", "Q: Did povorcitinib show any safety liabilities in vitro?\nA: In vitro toxicity studies reported in the paper found povorcitinib had an IC50 greater than 20 \u03bcM for hERG inhibition and greater than 20 \u03bcM for HEK293 proliferation, and it did not inhibit CYP enzymes 1A2, 2B6, 2C9, 2C19, 2D6 or 3A4 below 25 \u03bcM. No significant weight loss was seen in the rodent studies.", "Q: Which residues of OSM are important for gp130 binding?\nA: According to the thesis, OSM residues important for gp130 binding form a solvent-exposed \"hot spot\" centred on Gly 120 and Asn 124 of helix C and Gln 16 and Gln 20 of helix A, forming a hydrophobic pocket between helices A and C. These are structurally equivalent to the LIF site 2 residues, and the model is further complicated by evidence that OSM can also bind the N-terminal immunoglobulin-like domain of gp130 through site 3 interactions.", "Q: What combination therapy result does Marc highlight?\nA: The MANIFEST-2 trial pairing pelabresib, a BET inhibitor, with ruxolitinib produced a 65% spleen volume reduction and major anaemia improvement in 24 weeks, which Marc describes as showing how dual-pathway inhibition can overcome the limits of monotherapy.", "Q: Summarize Marc Deller's career in a few sentences.\nA: Most recent first: in 2025 Marc Deller became CSO and co-founder of Elora Therapeutics. Before that he was Associate Director then Research Fellow in Structural Biology at Incyte Pharmaceuticals in Wilmington (2018\u20132024), Director of Stanford University's Macromolecular Structure Knowledge Center (2015\u20132018), Senior Scientist at the Joint Center for Structural Genomics at Scripps Research Institute (2006\u20132015), and Principal Scientist in Structural Biology at Pfizer in San Diego (2001\u20132006), following a postdoc at Yale University School of Medicine (1999\u20132001). He did a D.Phil. in molecular biophysics at the University of Oxford (1995\u20131999) and trained in biochemistry at the University of Leeds (BSc, 1991\u20131995).", "Q: Is there a lower-resolution companion to 3O0F?\nA: Yes, 3E0F is the same putative metal-dependent phosphoesterase BAD_1165 from Bifidobacterium adolescentis at 2.40 \u00c5 in space group P 65, released on 2 September 2008, also with iron, phosphate and zinc bound. It shares the 2011 Proteins primary citation with 3O0F.", "Q: What industry does Marc Deller list on LinkedIn?\nA: Marc Deller lists his industry on LinkedIn as Biotechnology Research.", "Q: Could crystal mounting be avoided altogether?\nA: Marc points to high-speed pixel array detectors that allow continuous data collection from crystals still in their growth plates, and to in-situ diffraction approaches that bypass harvesting entirely, removing a physical failure point. He also notes oil-assisted hyperquenching as a route to better cryocooling with less cryoprotectant.", "Q: What is PDB entry 3K5J?\nA: 3K5J is a putative SUFU (suppressor of fused) homologue, YP_208451.1, from Neisseria gonorrhoeae FA 1090 at 1.40 \u00c5 in space group P 32 2 1, with 183 polymer residues and 22.16 kDa, released on 26 January 2010 and keyworded PROTEIN BINDING. It is cited in the 2010 Protein Science paper defining a novel group of bacterial proteins similar to the N-terminal domain of human Sufu.", "Q: How accurate is MARCUS, and does it remove the need for human curators?\nA: Marc reports that MARCUS converts about two-thirds of structure images accurately on first pass, which he calls remarkable given the complexity of chemical imagery. He is clear that expert human validation remains essential for data quality, with the tool reducing rather than eliminating the curator's cognitive load.", "Q: Has Marc done any cloud computing training?\nA: Yes. Marc Deller completed \"Serverless Architectures on AWS\", a certificate issued by Amazon Web Services through Coursera on 10 March 2025. He writes that scalable cloud infrastructure is increasingly important for scientific software and shared tools, and that the course helps with deploying lightweight web apps, services and online research resources.", "Q: Who published the NTP pyrophosphohydrolase structure?\nA: It was published as Deller, M.C., et al. (2010), \"Crystal Structure of a Putative NTP Pyrophosphohydrolase (Exig_1061) from Exiguobacterium sp. 255-15\", in Acta Crystallographica Section F, volume 66, pages 1237-1244. Marc Deller is first author of that structure paper.", "Q: Who supervised Marc's D.Phil. and where was the crystallography carried out?\nA: Marc Deller's D.Phil. was supervised by Prof. E. Yvonne Jones in the Laboratory of Molecular Biophysics and the Wellcome Trust Centre for Human Genetics at the University of Oxford, with in-house data collected on a Rigaku rotating anode and synchrotron data collected at the Photon Factory (Japan), the ESRF (France) and the SRS at Daresbury (UK).", "Q: How do you make an omit map in practice?\nA: The chapter gives a protocol: download coordinates and reflection data from the PDB, convert the reflections to MTZ with the CCP4i utilities or take the fully optimized MTZ from PDB_REDO, delete the ligand by editing the HETATM records and any LINK or CISPEP header records while keeping the ATOM and CRYST1 lines, then run a single cycle of refinement and inspect the resulting maps in COOT. Marc Deller and his co-authors recommend using more than one refinement program since results may differ.", "Q: What are ALPHABETTI's limitations?\nA: ALPHABETTI handles a single chain only, with no complexes, assemblies, ligands, cofactors, metals or post-translational modifications, and accepts between 10 and 400 residues. Marc also warns that ESMFold degrades quietly on low-homology and disordered sequences, that because ESMFold's trunk is ESM-2 the fold and the logo fail together rather than independently, and that database sequences are usually precursors, so users should check the pLDDT before believing a picture.", "Q: Why is automated crystal detection so difficult?\nA: Marc Deller and Bernhard Rupp give three reasons: protein crystals well ordered on the molecular level are often macroscopically irregular with disordered edges and surfaces, crystals move during harvesting so the textures and facets used to identify them keep changing, and both false positives and false negatives must be rigorously rejected while the algorithms run fast enough for a real-time control loop.", "Q: What happened when 10E8 was tested clinically?\nA: The post notes that early clinical trials revealed unexpected local reactogenicity with 10E8VLS, leading to suspension of some studies and emphasising the need for careful antibody engineering and safety evaluation, while research continues on next-generation MPER-targeting and bispecific antibodies. It is published at https://marcdeller.com/lipid-interactions-and-angle-of-approach-to-the-hiv-1-viral-membrane-of-broadly-neutralizing-antibody-10e8-insights-for-vaccine-and-therapeutic-design/.", "Q: Where is the imidazopyrimidinone WRN patent post?\nA: Marc Deller published it on 9 September 2025 at https://marcdeller.com/imidaz012-apyrimidine-51-h-ones-as-wrn-inhibitors/, linking to Google Patents for US20250270218A1.", "Q: Does Marc Deller have experience with AI and machine learning?\nA: Yes. Marc Deller championed the integration of AI/ML and cloud computing into Incyte's structural biology group, uses AlphaFold, RFdiffusion and ProteinMPNN for AI-driven protein design, and completed IBM's Introduction to Artificial Intelligence course and a LinkedIn Learning course on generative AI for communication in 2024\u20132025.", "Q: What is \"From Tool to Teammates\" about?\nA: \"2026 \u2013 From Tool to Teammates: Working Smarter with AI\" records a LinkedIn Learning certificate issued to Marc Deller on 24 March 2026 for a course on using AI as a practical workplace collaborator. He writes that AI is becoming part of everyday scientific planning, drafting and synthesis, and that the training helps researchers work more efficiently in brainstorming, review and structured problem-solving.", "Q: What is INCB159020?\nA: INCB159020 is compound 23 in the 2025 paper Marc Deller co-authored: a reversible, orally bioavailable KRAS G12D inhibitor built on a tricyclic indole core bearing an azabicyclo[2.1.1]hexane, a cyanoethyl group, an (R)-methyl morpholinone side chain and a 2-methyl-3-chlorophenyl Switch II group. It inhibits KRAS G12D with an exchange IC50 of 4.4 nM, a cellular pERK IC50 of 33 nM and a viability IC50 of 78 nM.", "Q: How did the Drosophila S2 cell expression of LIFR turn out?\nA: The Drosophila S2 system (pRmHa-3 vector, copper-inducible metallothionein promoter) failed. No induction of LIFR-FN3-3C-Fc was detected by ELISA, Protein-A immunoprecipitation or western blot; instead a ~46 kDa protein appeared, versus the ~140 kDa control. The S2 conditioned medium was inactive in ligand-binding assays. Marc concluded the ~46 kDa species was likely a truncated translation product, given its specific reactivity with LIFR antibodies.", "Q: How do I sort a list in Python?\nA: General programming questions aren't what I'm here for, and I'd rather point you somewhere better than guess. I'm chatMCD, here for Marc Deller: his science, his companies and the tools he has built. Anything there I can help with?", "Q: What PDB accession codes came out of this work?\nA: Coordinates and structure factors for the DRVIA7 Fab and its complex with gp120 were deposited in the Protein Data Bank as 5CD3 and 5CD5 respectively. The accompanying negative-stain EM reconstructions of DRVIA7 and DRVIA7H-VRC01L Fabs bound to the JRFL SOSIP.664 trimer were deposited in the EMDB as EMD-6372 and EMD-6373.", "Q: Which residues matter most in that enzyme?\nA: The key finding is that H161 and Q167 form a hydrogen-bond network that stabilises bound oxygen and promotes proton transfer. Mutational analyses of Myceliophthora thermophila PMO3* variants showed that disrupting the H161\u2013Q167 network slashes oxygen binding and proton-transfer efficiency.", "Q: Which glycans were the \"roadblocks\" in the title?\nA: The two glycans were the N276 glycan on gp120 and the N-linked glycan on the gp120 V5 loop (N461). Although about 15 \u00c5 apart, they together formed an inhibitory glycan patch that the DRVIA7 light-chain CDR1 and N terminus could not accommodate, limiting the antibody's neutralizing breadth.", "Q: What computer does Marc use for AI work?\nA: Marc's local development and training machine is a 64 GB Apple M1 Mac, on which he fine-tunes local language models with Apple's MLX-LM.", "Q: What is the 2017 ligand validation chapter Marc Deller co-wrote?\nA: \"Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands\" is Chapter 25 of Protein Crystallography: Methods and Protocols, Methods in Molecular Biology volume 1607, pages 611 to 625, published by Springer in 2017. It was written by Edwin Pozharski, Marc C. Deller and Bernhard Rupp as a practical guide to checking whether a ligand modelled in a crystal structure is really supported by the experimental data.", "Q: What opportunity does Marc see for structural biologists in the microplastics-cancer field?\nA: Marc writes that the oncogenic signalling nodes activated by microplastic exposure \u2014 NF-\u03baB, PI3K/AKT/mTOR, MAPK/ERK, Wnt/\u03b2-catenin and ECM-receptor and focal adhesion remodelling \u2014 are many of them well characterised and structurally tractable, and may represent a new and genuinely addressable class of environmental oncology targets for researchers at the structural and computational biology interface of drug discovery. The post is at https://marcdeller.com/plastic-persistent-and-potentially-carcinogenic/.", "Q: How stable is a folded protein according to the review?\nA: The review by Marc Deller and colleagues notes that a folded protein is only marginally stable, roughly 10 kcal per mole for a 10 kDa protein, because the large entropic penalty of folding is only just outweighed by van der Waals forces, hydrogen bonding and the hydrophobic effect.", "Q: What cell lines does the patent use to test WRN inhibitors?\nA: The cellular assays use human colon carcinoma lines from ATCC: the MSI-H lines RKO (CRL-2577) and SW-48 (CCL-231), and the microsatellite stable lines SW-620 (CRL-227) and SW-837 (CRL-235). WRN inducible knockouts were generated by CRISPR-Cas9 editing, and readouts include a Cell Titer Glo anti-proliferation assay and a pKAP1 induction AlphaLISA assay as a DNA damage marker.", "Q: How does the original Topliss tree work?\nA: Starting from an unsubstituted benzene ring, chemists introduce substituents chosen for their physicochemical properties, with each branch point set by whether the previous change increased, decreased or maintained activity. Marc notes the sequence typically begins with a 4-chloro substitution and branches to 3,4-dichloro, 4-methoxy or 3-trifluoromethyl depending on the observed trend.", "Q: What two engineering milestones does Marc identify for PET depolymerases?\nA: Marc identifies Tournier et al. (Nature, 2020), who reported an engineered thermophilic cutinase variant achieving at least 90% PET depolymerization within 10 hours at industrially relevant loadings, and Lu et al. (Nature, 2022), who used a structure-based machine learning algorithm to create FAST-PETase, whose five mutations enabled near-complete degradation of untreated post-consumer PET from 51 different thermoformed products within one week.", "Q: What temperature improvement does Marc quote for DuraPETase?\nA: Citing Table 2 from Liu et al., Marc reports that DuraPETase raises the optimal temperature from 46.8\u00b0C to 77\u00b0C, alongside large increases in hydrolysis rate.", "Q: How does the tricyclic scaffold differ from the bicyclic one?\nA: Marc Deller argues the change is more than cosmetic: adding a third fused ring raises sp3 character, moves the molecules further into three-dimensional chemical space, and creates additional vector positions around the Ring A and optional Ring D substituents disclosed in the new claims.", "Q: How was the quality of the low-resolution TM1088 refinement assessed?\nA: The difference between Rcryst and Rfree for the 3.45 \u00c5 TM1088 structure was 4.9%, which Marc Deller noted is in excellent agreement with the expected value of 4.97% calculated for this resolution using the Urzhumtsev formula. Analysis with the PHENIX script phenix.r_factor_statistics showed that both R values fell on the peak of the distribution for all PDB structures in the 3.35\u20133.55 \u00c5 range, and MolProbity showed 95.1% of residues in favored Ramachandran regions with 0.9% outliers.", "Q: How common are intrinsically disordered proteins?\nA: The review by Marc Deller, Leopold Kong and Bernhard Rupp cites estimates that around 40% of all human proteins contain at least one disordered segment and around 25% are completely disordered. These proteins have largely been avoided by crystallographers, and NMR has been the main tool used to understand them.", "Q: Which structure is the NANOG one?\nA: 4RBO is the crystal structure of the human Nanog homeobox (NANOG) at 3.3 \u00c5 in space group P 65 2 2, with 188 polymer residues and 31.87 kDa, released on 1 October 2014. It was deposited under the Joint Center for Structural Genomics and the Partnership for Stem Cell Biology, keyworded DNA BINDING PROTEIN and TRANSCRIPTION/DNA, and cited in the 2015 PNAS paper on a structure-based NANOG variant with enhanced properties for self-renewal and reprogramming.", "Q: What is the N-end rule and does it matter for bacterial expression?\nA: The N-end rule holds that the identity of the N-terminal residue predicts a protein's half-life in vivo: methionine, serine, alanine, glycine, threonine, valine or proline stabilize with half-lives over 20 hours, while phenylalanine, aspartate, lysine or arginine destabilize with half-lives under 3 minutes. Marc Deller's review notes it operates through ubiquitination, so it is of no concern in bacterial systems, and that an N-terminal affinity tag masks the residue anyway.", "Q: How does Marc explain the molecular basis of chlorine's effect?\nA: Marc explains that chlorine acts as a directional Lewis acid forming halogen bonds with geometries as precise as hydrogen bonds, engaging backbone carbonyl oxygens, hydroxyl side chains and aromatic \u03c0-systems. At the same time it raises logD by roughly 0.6 to 1.0 units to provide a lipophilic anchor, and its dual electron-withdrawing and electron-donating character lets it mimic methyl, trifluoromethyl, cyano or even hydroxyl groups depending on context, which is why para-chlorophenyl sits at the top of the Topliss decision tree.", "Q: What is the covalent-warhead WRN patent?\nA: \"2025 \u2013 Imidaz0[1,2-a]pyrimidine-5(1 h)-ones having a covalent warhead at the 3-position as wrn inhibitors\" covers US20250270217A1, in which imidazo[1,2-a]pyrimidin-5(1H)-ones carry a covalent warhead at the 3-position. That warhead enables irreversible binding to cysteine residues in the WRN active site, giving enhanced potency and prolonged target engagement compared with reversible inhibitors.", "Q: Does a high melting temperature guarantee a protein will crystallize?\nA: No. The review by Marc Deller and colleagues cites a Thermofluor study of 657 protein samples in which only 23% of proteins with a melting temperature below 43\u00b0C gave crystals compared with 49% of those above 45\u00b0C, but above 45\u00b0C there was no significant further gain in crystallization frequency up to 96\u00b0C.", "Q: What is Marc's opinion of AlphaFold 3?\nA: I don't have a statement from him on that specific model, so I won't put words in his mouth. He does write about structure prediction on marcdeller.com.", "Q: What apps has Marc built?\nA: Marc's vibe-coding portfolio includes AlphaFraud (an audit of AlphaFold predictions against new PDB structures), BoltzMaker (a Boltz-2 campaign wrapper), chatPDB (a PDB AI assistant), CODSWALLOP (family-centric PDB mining), GOBSMACKED (fold \u2192 dock \u2192 MD \u2192 analysis), ALPHABETTI (3D sequence logos), PANTS (PETase enzyme mining), ChemSage (a local drug-discovery LLM) and chatMCD (this chatbot), plus a set of smaller branded HTML tools.", "Q: Walk me through Marc's career, company by company, and what he did at each.\nA: Marc has spent more than 25 years moving from academic structural biology into drug discovery leadership, and most recently into founding a biotech. Working backwards from today:\n\n**Elora Therapeutics** (Co-Founder and Chief Scientific Officer, 2025-present). Engineering PET-degrading enzymes into a first-in-class systemic therapy for the microplastic burden in the human body. He leads scientific strategy, the intellectual property, grant applications, CRO partnerships and team building.\n\n**DeepCovalent** (Scientific Technical Advisor, 2025-present). Physics-based computer-aided drug design pipelines, small-molecule GPCR discovery platforms, in silico screening and cheminformatics.\n\n**Incyte** (Research Fellow 2022-2024, Associate Director 2018-2022). As Research Fellow he determined 100+ challenging co-crystal structures and ran structure-based drug design across more than ten oncology and immunology programmes, supporting 6+ FDA IND filings and 7+ patents. As Associate Director he designed and built an $8M automated structural biology platform that delivered a four-fold efficiency gain and 100+ co-crystal structures in under three years.\n\n**Stanford ChEM-H** (Director and Founder, Macromolecular Structure Knowledge Center, 2015-2018). Built a profitable wet-lab service organisation from nothing: a 70-project portfolio, a $1M annual budget, 50+ researchers equipped, six novel structures, three NIH SBIR submissions, and synchrotron access extended to 70 users in ten countries.\n\n**Scripps Research, JCSG** (Target Manager, 2006-2015). Solved 70+ novel protein structures, co-authored 40+ publications, raised NIH structural-genomics throughput by 10%, designed the JCSG Top96 crystallisation screen, and contributed crystallography to the HIV-1 Env and HCV E2 antibody complexes in Ian Wilson's laboratory.\n\n**Pfizer** (Principal Scientist, Structural Biology, San Diego). Structural biology for small-molecule drug discovery programmes.\n\n**Yale University School of Medicine** (Postdoctoral Research Fellow). JAK/STAT protein structure research.\n\n**Education.** A **D.Phil.** at the **University of Oxford** (1995-1999) with Prof. E. Yvonne Jones, where he determined the first crystal structure of the cytokine oncostatin M; before that a First Class Honours **BSc** in Biochemistry and Molecular Biology at the **University of Leeds** (1991-1995), including a year's industrial placement at Boots Pharmaceuticals in the team behind ibuprofen. He grew up in Nottingham but studied elsewhere.", "Q: Where were the WRN crystallographic data collected?\nA: Data were collected at beamline I03 of the Diamond Light Source in Didcot, UK, at 100 K and a wavelength of 0.9763 \u00c5, using a Dectris Eiger2 XE 16M detector. A total of 3600 images were collected with an oscillation range of 0.1 degrees per image, and the data were processed with the autoPROC pipeline incorporating XDS, Aimless and Pointless.", "Q: What does Marc suggest about combining enzyme therapy with existing MASLD drugs?\nA: Marc suggests that combining enzymatic therapy with metabolic interventions such as GLP-1 receptor agonists might offer synergistic benefits, since those medications already show efficacy in MASLD while potentially reducing inflammatory responses to residual particles.", "Q: What is the \"6 C's\" on Marc Deller's website?\nA: The 6 C's are Marc Deller's stated personal philosophy, a set of six values he writes about in a linked series of posts on marcdeller.com: Curiosity, Courage, Creativity, Communication, Collaboration and Compassion. Each post pairs one value with a headline achievement, such as 400+ structures determined or 60+ publications, and explains how that value shows up in his day-to-day scientific work.", "Q: Which cytokine is structurally most similar to OSM?\nA: According to the structural superpositions in Marc Deller's thesis (calculated with the program SHP), murine LIF is the closest structural relative of OSM, with an RMSD of 2.10 \u00c5 over 145 equivalent C\u03b1 positions and 21 % sequence identity, followed by human LIF (2.27 \u00c5 over 138 residues, 20 % identity). Human IL-6 (2.40 \u00c5) and human CNTF (2.95 \u00c5) were the next best matches, confirming that OSM belongs structurally to the gp130 family of cytokines.", "Q: What does each generate_yaml output YAML contain?\nA: A single protein chain with the user-supplied ID and sequence, a ligand defined by cleaned SMILES, a pocket constraint anchoring the ligand to a specified contact residue, and an affinity property block targeting the ligand.", "Q: Why does Marc recommend running AlphaFold locally?\nA: Marc lists full control and independence from cloud services, API limits and internet connectivity; a performance boost from Apple Silicon or other local GPUs; customization of environments, dependencies and databases; security and privacy for proprietary sequences; and seamless integration of AlphaFold outputs into existing lab scripts and visualization tools. The post is at https://marcdeller.com/alphafold-local-install-on-mac/.", "Q: Which JAK inhibitors does Marc discuss?\nA: Ruxolitinib, approved in 2011 and effective regardless of genotype; fedratinib, which added FLT3 and BRD4 activity and emerged in 2019 with the FREEDOM2 trial confirming a 36% response rate in ruxolitinib-refractory myelofibrosis; pacritinib for thrombocytopenic patients; and momelotinib for its anaemia benefit.", "Q: What is the structure of the LIF receptor (LIFR)?\nA: According to the thesis, LIFR is 1097 amino acids long with a 44-residue leader sequence, a 26-residue transmembrane domain, a 789-residue extracellular region and a 238-residue cytoplasmic domain. The extracellular ligand-binding region consists of two haemopoietin domains separated by a putative immunoglobulin-like domain and joined to the membrane by three fibronectin type III domains. Its predicted mass is 111 kDa, but it runs at ~200 kDa on SDS-PAGE because of extensive glycosylation at several of its 20 potential N-glycosylation sites.", "Q: What did the review say about the TRAIL\u2013DR5 structures?\nA: The review described the two contemporaneous structures of TRAIL (TNF-related apoptosis-inducing ligand) bound to death receptor 5 (DR5) as the first new structural information on a TNF-family cytokine\u2013receptor complex since the TNF\u03b2\u2013TNFR1 structure of 1993. Marc Deller and Yvonne Jones noted that TRAIL, like TNF, is trimeric and forms a rigid scaffold for cross-linking three receptors, that the most hydrophobic part of the interface is conserved with specificity coming from polar variation, and that the 54 \u00c5 spacing between the DR5 C termini closely matches the 52 \u00c5 spacing of binding sites on the cytosolic adaptor TRAF2.", "Q: Which stem cell paper is Marc Deller a co-author on?\nA: Marc Deller is a co-author of \"Structure-based discovery of NANOG variant with enhanced properties to promote self-renewal and reprogramming of pluripotent stem cells\", published in Proceedings of the National Academy of Sciences 112(15):4666-4671 in 2015 (DOI 10.1073/pnas.1502855112). Google Scholar records 66 citations for it.", "Q: What funding supported the JCSG Top96 work?\nA: The JCSG was supported by the NIH National Institute of General Medical Sciences under award number U54 GM094586. Portions of the research were carried out at the Stanford Synchrotron Radiation Lightsource (SSRL) and the Advanced Light Source (ALS).", "Q: What are the pitfalls of Thermofluor described in the review?\nA: Marc Deller's review warns that dyes can interact with and destabilize the protein, as seen for GroEL with ANS, that aberrant or false-positive thermal shifts are common and need careful melt-curve analysis with tools such as Meltdown, that not all proteins unfold in a clean sigmoid curve, and that no shift on adding a small molecule does not prove absence of binding.", "Q: How did Marc Deller show that his truncated oncostatin M was still functional?\nA: Marc used receptor competition and titration ELISA-style assays and a Ba/F3 cell survival assay. In both, OSM_\u0394188 behaved like wildtype OSM-196, showing that the \u0394188 C-terminal truncation did not affect receptor binding or biological activity. This was crucial evidence that the engineered crystallisation construct was a valid model for the real protein.", "Q: What is Marc's favourite restaurant?\nA: I don't know his favourite restaurant. His taste in drinks is better documented, but on restaurants I'd only be guessing. More at marcdeller.com.", "Q: What figure does Marc quote for the disease burden from environmental pollutants?\nA: In the XenoBug post Marc notes that environmental pollutants contribute to approximately 22% of the global disease burden, which he uses to frame the urgency of faster bioremediation enzyme discovery.", "Q: Did Marc write about nanoplastics crossing the blood-brain barrier?\nA: Yes. In \"The Invisible Invasion\" Marc describes Trinity College Dublin research showing nanoplastics cross the blood-brain barrier and target mitochondrial electron transport chains, compromising brain energy metabolism and possibly contributing to neurodegenerative disease. You can read it at https://marcdeller.com/the-invisible-invasion-how-microscopic-plastics-are-infiltrating-human-organ-systems/.", "Q: How did Marc Deller study the LIFR-LIF complex?\nA: Marc expressed and purified LIF as a GST fusion (cleaving with thrombin) similarly to OSM, then mixed LIF and LIFR-FN3 at a 10:1 molar ratio and analysed the mixture by gel filtration on a Superose-6 column. Three peaks appeared, corresponding to LIF (~36 kDa), LIFR-FN3 and a putative LIF-LIFR-FN3 complex, the latter eluting earlier and shown by SDS-PAGE to contain both ~25 kDa (LIF) and ~116 kDa (LIFR) bands.", "Q: What kind of proteins were represented in the JCSG dataset used for Top96?\nA: The JCSG dataset used to design the Top96 screen covered a diverse set of protein targets of both bacterial and mammalian origin, which the poster highlights as a unique opportunity for systematic analysis of crystallization conditions.", "Q: Who is Marc C. Deller?\nA: Marc C. Deller, D.Phil., is a structural and computational biology leader with 20+ years in protein science and drug discovery. He is currently Chief Scientific Officer and co-founder of Elora Therapeutics and a Scientific Technical Advisor to DeepCovalent Inc. Over his career he has built 2 wet labs from scratch, delivered 400+ protein structures, supported 6+ FDA IND filings, is named on 7+ patents and has published 60+ manuscripts.", "Q: What angle does 10E8 approach the HIV membrane at?\nA: The model built in the 2017 paper places the gp41 helical MPER epitope (residues 671\u2013683) at about 75\u201380\u00b0 to the viral membrane surface \u2014 essentially upright \u2014 with the epitope tilted roughly 78\u00b0 \u00b1 3\u00b0 from the bilayer during engagement. The 10E8 Fab itself approaches at about 43\u00b0 \u00b1 3\u00b0, measured from the viral membrane to the pseudo-dyad axis between the variable light and heavy chains.", "Q: Who were the authors of the HCV E2\u2013AP33 structure paper?\nA: The authors were Leopold Kong, Erick Giang, Travis Nieusma, Justin B. Robbins, Marc C. Deller, Robyn L. Stanfield, Ian A. Wilson and Mansun Law. Ian A. Wilson and Mansun Law were the corresponding authors, and the work came from the Departments of Molecular Biology and of Immunology and Microbial Science and the Skaggs Institute for Chemical Biology at The Scripps Research Institute, together with the Joint Center for Structural Genomics.", "Q: What is Marc's own view in the elastic networks post?\nA: Marc's view is that machine learning has its greatest impact when combined with deep domain knowledge and experimental validation rather than replacing traditional biophysics. He argues the success of the approach comes from respecting each component's strengths \u2014 AlphaFold2 supplying excellent global folds, elastic networks supplying biophysically reasonable collective motions without all-atom molecular dynamics cost, and cryo-EM maps supplying the experimental constraint. You can read it at https://marcdeller.com/revolutionary-bridge-between-prediction-and-reality-how-elastic-networks-transform-alphafold2s-structural-accuracy/.", "Q: What crystallization conditions were used for the TM1088 complex?\nA: Marc Deller crystallized the TM1088 complex at 3 mg/ml with NADH added at a 1.1:1 molar ratio, by vapor diffusion against 0.1 M magnesium chloride, 15% (w/v) PEG 4000 and 0.1 M HEPES pH 7.0, with 30% xylitol as cryoprotectant and reservoirs overlaid with a 1:1 mixture of silicon oil and paraffin oil to slow vapor diffusion. TM1088A crystallized in 40% MPD, 5% PEG-8000, 0.1 M cacodylate pH 6.5 after 54 days, and the TM1088B C-terminal domain crystallized in 25% ethylene glycol, both using JCSG nanodroplet sitting-drop protocols.", "Q: What were the two competing conformations of PKS modules?\nA: SAXS analysis of DEBS module 3 had supported an \"extended\" conformation with the KS-AT core in the same arrangement seen in crystal structures of stand-alone KS-AT fragments, whereas cryo-EM of module 5 of the pikromycin synthase had revealed an \"arch-shaped\" conformation with the homodimeric KS as the capstone and each AT domain oriented downward to contact the KS of the opposite dimer. The two conformations differ mainly in the orientation of the AT and linker domains relative to the KS, and the 2018 paper set out to determine which is catalytically relevant.", "Q: What is Marc's post on GLP-1 receptor drug binding about?\nA: \"Molecular Architecture of GLP-1 Receptor Drug Binding Reveals Dynamic Activation Mechanisms\", published on 12 December 2025, explains how cryo-EM and molecular dynamics have revealed the way drugs engage the class B GPCR GLP-1R, a pivotal target for type 2 diabetes and obesity. Marc Deller describes the two-domain binding mechanism in which the extracellular domain captures peptide C-termini as an affinity trap while N-terminal peptide segments insert into the transmembrane orthosteric pocket to trigger activation.", "Q: How does Marc link network biology to precision medicine?\nA: Marc presents the network view as a critical step toward personalized medicine, where treatments can be tailored to the unique molecular profile of an individual's disease network. He describes the shift as the difference between looking at a single street and seeing the entire city map; the post is at https://marcdeller.com/from-single-targets-to-disease-networks-a-systems-approach-to-drug-discovery/.", "Q: What fold does MtPMO3* adopt?\nA: Like other PMO structures, MtPMO3* exhibits a \u03b2-sandwich fold formed by nine antiparallel \u03b2-strands interspersed with sizable loop regions. The copper active site shares the primary coordination sphere features of other fungal PMOs: the histidine brace formed by H1 and H75, an axial tyrosine (Y169), and N\u03b5-methylation of the N-terminal histidine.", "Q: Did Marc Deller found any labs?\nA: Yes. Marc Deller has founded 2 structural biology labs: the Macromolecular Structure Knowledge Center at Stanford (2015) and the $8M automated structural biology platform at Incyte (from 2018). His r\u00e9sum\u00e9 also counts 2 automated crystallography labs established and 1 pre-seed startup, Elora Therapeutics, initiated.", "Q: What does Marc say the hat signals in a professional setting?\nA: Marc argues that in conferences, presentations and lab visits the hat becomes a conversation catalyst conveying a blend of scientific rigor and approachable leadership, citing cognitive psychology research that distinctive attire can enhance memorability by up to 30% and color-psychology findings that cream and tan tones evoke warmth and trust. He concludes it fuses heritage, functionality and professional branding. The post is at https://marcdeller.com/the-signature-cowboy-hat-of-adam-lee-a-blend-of-tradition-and-function/.", "Q: What did the DRVIA7 structure show about antibody maturation?\nA: The structure revealed how key glycans on gp120, particularly at positions N276 and V5, create steric roadblocks that impede the maturation of VRC01-class broadly neutralising antibodies. Marc Deller's post explains that the structural and longitudinal analyses inform glycan-accommodating vaccine design strategies for eliciting VRC01-class responses across diverse populations.", "Q: What is Marc's favourite movie?\nA: Night of the Comet, the 1984 cult sci-fi film.", "Q: Did Marc work on any hepatitis C virus structures?\nA: Yes. PDB entry 4G6A is the structure of the hepatitis C virus envelope glycoprotein E2 antigenic region 412-423 bound to the broadly neutralising antibody AP33, solved by X-ray diffraction at 2.501 \u00c5 and released on 26 September 2012. Marc Deller lists it in the structures section of his website.", "Q: What is the mutant-CALR antibody patent that Marc Deller is named on?\nA: Marc Deller is a named co-inventor on \"Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\", published as international patent application WO 2023/107994 A1 (PCT/US2022/081076) and assigned to Incyte Corporation. It discloses antibodies that bind human mutant calreticulin (mutCALR) with high affinity and specificity, together with the related nucleic acids, vectors, cells, kits and pharmaceutical compositions, and methods of treating or diagnosing myeloproliferative neoplasms with them.", "Q: How do microfluidics help structural biology, according to Marc Deller?\nA: In \"Tiny Pipes, Big Discoveries: How Microfluidics Are Accelerating Structural Biology\", published 13 August 2025, Marc Deller explains that lab-on-a-chip devices manipulate minuscule liquid volumes with great precision, letting automated platforms set up thousands of crystallization experiments in parallel using a fraction of the protein required by traditional methods.", "Q: What is PDB entry 5SY8?\nA: 5SY8 is the crystal structure of a complex between a 10E8 Fab light chain mutant and the T117v2 HIV-1 MPER epitope scaffold, solved by X-ray diffraction at 1.62 \u00c5 and released on 8 March 2017. Marc Deller's post lists the organisms as Homo sapiens and a synthetic construct, with glycerol as the only ligand.", "Q: What four themes does the Structural Biology page highlight?\nA: The page highlights Precision, Prediction, Acceleration, and Innovation. Precision covers atomic-level binding landscapes for ultra-targeted compounds, Prediction covers machine-learning modelling of complex folds, Acceleration covers integrating structural data into design-make-test-analyse cycles, and Innovation covers revealing novel therapeutic scaffolds beyond known chemical space.", "Q: Tell me about Marc's publications and protein structures.\nA: Marc has published 60+ papers and contributed to more than 400 protein structures, across HIV vaccine design, structural genomics and small-molecule drug discovery. His full publication list is on his [publications page](https://marcdeller.com/publications/) and on [ORCID](https://orcid.org/0000-0001-8070-6502).\n\n## \ud83d\udcc4 Notable publications, most recent first\n\n- **An isoform-selective JAK1 inhibitor** (*Journal of Medicinal Chemistry*, 2026): [summary](https://marcdeller.com/discovery-of-the-orally-bioavailable-isoform-selective-janus-kinase-1-compound-povorcitinib/) \u00b7 [paper](https://doi.org/10.1021/acs.jmedchem.5c03753)\n- **An orally bioavailable KRAS G12D inhibitor** (*Journal of Medicinal Chemistry*, 2025): [summary](https://marcdeller.com/discovery-of-incb159020-an-orally-bioavailable-kras-g12d-inhibitor/) \u00b7 [paper](https://doi.org/10.1021/acs.jmedchem.4c02662)\n- **Potent and selective CDK2 inhibitors** (*Journal of Medicinal Chemistry*, 2024): [summary](https://marcdeller.com/discovery-of-4-pyrazolyl-2-aminopyrimidines-as-potent-and-selective-cdk2-inhibitors/) \u00b7 [paper](https://doi.org/10.1021/acs.jmedchem.3c02287)\n- **Protein stability: a crystallographer's perspective** (*Acta Crystallographica F*, 2016), a widely cited review: [summary](https://marcdeller.com/protein-stability-a-crystallographers-perspective/) \u00b7 [paper](https://doi.org/10.1107/S2053230X15024619)\n- **Crystal structure of a soluble cleaved HIV-1 envelope trimer** (*Science*, 2013), his most-cited paper: [summary](https://marcdeller.com/crystal-structure-of-a-soluble-cleaved-hiv-1-envelope-trimer/) \u00b7 [paper](https://doi.org/10.1126/science.1245625)\n- **A supersite of immune vulnerability on HIV-1 gp120** (*Nature Structural & Molecular Biology*, 2013): [summary](https://marcdeller.com/supersite-of-immune-vulnerability-on-the-glycosylated-face-of-hiv-1-envelope-glycoprotein-gp120/) \u00b7 [paper](https://doi.org/10.1038/nsmb.2594)\n- **The crystal structure of oncostatin M** (*Structure*, 2000), from his D.Phil.: [summary](https://marcdeller.com/crystal-structure-of-human-oncostatin-m/) \u00b7 [paper](https://doi.org/10.1016/S0969-2126%2800%2900176-3)\n\n## \ud83e\uddec Protein structures\n\nOf the 400+ structures he has contributed to, the 78 PDB entries that carry his name as a depositor or citation author are all listed [at RCSB](https://chatmcd.mdeller.com/pdb) and on his [structures page](https://marcdeller.com/structures/). Notable ones, most recent first:\n\n- **JAK1 and JAK2 kinases with an isoform-selective inhibitor**, 1.54 and 1.59 \u00c5 (2026): [story](https://marcdeller.com/discovery-of-the-orally-bioavailable-isoform-selective-janus-kinase-1-compound-povorcitinib/) \u00b7 RCSB [10PI](https://www.rcsb.org/structure/10PI), [10PJ](https://www.rcsb.org/structure/10PJ)\n- **KRAS G12D with an orally bioavailable inhibitor**, 1.36 \u00c5 (2025): [story](https://marcdeller.com/discovery-of-incb159020-an-orally-bioavailable-kras-g12d-inhibitor/) \u00b7 RCSB [9E5D](https://www.rcsb.org/structure/9E5D)\n- **CDK2 with a selective inhibitor**, 1.55 \u00c5 (2024): [story](https://marcdeller.com/discovery-of-4-pyrazolyl-2-aminopyrimidines-as-potent-and-selective-cdk2-inhibitors/) \u00b7 RCSB [8UV0](https://www.rcsb.org/structure/8UV0)\n- **The 10E8 antibody bound to the HIV-1 membrane-proximal region**, 1.62 \u00c5 (2017): [story](https://marcdeller.com/crystal-structure-of-the-complex-of-10e8-fab-light-chain-mutant1-and-t117v2-hiv-1-mper-scaffold/) \u00b7 RCSB [5SY8](https://www.rcsb.org/structure/5SY8)\n- **The BG505 SOSIP HIV-1 envelope trimer**, 4.7 \u00c5 (2013): [story](https://marcdeller.com/crystal-structure-of-a-soluble-cleaved-hiv-1-envelope-trimer/) \u00b7 RCSB [4NCO](https://www.rcsb.org/structure/4NCO)\n- **Human oncostatin M**, 2.2 \u00c5 (2000), from his D.Phil.: [story](https://marcdeller.com/crystal-structure-of-human-oncostatin-m/) \u00b7 RCSB [1EVS](https://www.rcsb.org/structure/1EVS)", "Q: Has Marc Deller worked on KRAS?\nA: Yes. At Incyte, Marc Deller conducted structure-based drug design on the KRAS program and is a co-author of the 2025 J. Med. Chem. paper on INCB159020, an orally bioavailable KRAS G12D inhibitor that supported an IND filing. His r\u00e9sum\u00e9 also lists macrocyclic peptide therapeutics against KRAS-G12D among his discoveries.", "Q: Did Marc Deller work with Leslie Epling on more than one Incyte paper?\nA: Yes. Leslie B. Epling and Marc Deller appear together on multiple Incyte structural biology contributions, including the 2022 FGFR2/3 paper (PDB 8E1X, with Hsiang-Ting Lei), the 2024 CDK2 inhibitor paper (PDB 8UV0) and the 2025 KRAS G12D paper describing INCB159020 (PDB 9E5D and 9E5F). In each case they supplied the co-crystal structures that guided the medicinal chemistry.", "Q: What is Marc Deller's paper on automated crystal harvesting?\nA: \"Approaches to automated protein crystal harvesting\" was published in Acta Crystallographica Section F in 2014 by Marc Deller and Bernhard Rupp, with Marc as first author. It surveys automated harvesting systems using microcapillaries, microtools, microgrippers, acoustic droplet ejection and optical traps.", "Q: What expression systems did Marc Deller try for LIFR?\nA: Marc tried three mammalian and insect expression systems for LIFR-FN3-3C-Fc: stable expression in Chinese hamster ovary (CHO) cells using the pEE14 vector, stable expression in Schneider's Drosophila S2 cells using the pRmHa-3 vector, and transient expression in human 293T cells using the pIG-1 vector. Each vector had been supplied by the CRC Growth Factor Group in Birmingham.", "Q: Can affinity tags be left on for crystallization?\nA: The review explains that large tags such as MBP are usually removed because the flexible linker interferes with crystal packing, but small tags such as His tags can often be left on without unduly affecting packing or function. Marc Deller and colleagues also describe rigidifying the linker with a run of alanine residues, which enabled MBP-fusion structures of gp21, SarR, MAT\u03b11 and MCL1.", "Q: What resolution was the 1B2/KS-AT crystal structure and what were its statistics?\nA: The crystal structure of the 1B2 Fab in complex with the KS-AT fragment of DEBS module 3 + TE, to which Marc Deller contributed, was determined at 2.1 \u00c5 resolution in the monoclinic space group C2. The asymmetric unit contained one molecule each of KS-AT and 1B2 plus 680 waters, with the homodimer generated by symmetry expansion, and the structure refined to Rwork and Rfree of 0.167 and 0.210.", "Q: Explain the TM1088 paper for a non-scientist.\nA: Bacteria need to keep the right amount of potassium inside their cells, and they use protein \"gates\" in their membrane to control it. Marc Deller used X-ray crystallography to take a detailed picture of the control unit of one such gate from a heat-loving bacterium called Thermotoga maritima. Unlike every similar control unit seen before, this one is built from two different proteins rather than one, yet it forms the same ring-shaped machine, suggesting it may be an ancient ancestor of the potassium-channel regulators found in more modern organisms.", "Q: What is Marc's own experience with Docling?\nA: Marc writes candidly that Docling works well in his hands for simple tables and bar charts, but that he has not had much success so far with more complex graphs or images, and that it does not support PDFs out of the box. He also notes installation is a single pip install docling command that pulls model weights from Hugging Face on first use, with basic usage taking about five lines of Python. The post is at https://marcdeller.com/unlocking-documents-for-ai-the-power-of-docling-for-ai-driven-transformation/.", "Q: Where were the OSM diffraction data collected?\nA: Marc Deller collected data at several facilities. Initial native data came from a capillary-mounted crystal at room temperature on beam line BL6A2 of the Photon Factory in Tsukuba, Japan. Native and MAD data were collected from propane-frozen crystals on beam line BL19 of the ESRF in Grenoble, France. Heavy-atom derivative screening used an in-house Rigaku RU2000 rotating anode (CuK\u03b1, 1.54 \u00c5) and station 9.6 of the SRS at Daresbury, UK.", "Q: What was OSM the fourth of, according to the thesis?\nA: Marc Deller's OSM crystal structure was the fourth member of the gp130 family of cytokines to be determined structurally, following IL-6 (Somers et al., 1997; Xu et al., 1997), CNTF (McDonald et al., 1995) and LIF (Robinson et al., 1994 by X-ray; Purvis and Mabbutt, 1997 by NMR).", "Q: What broader family definition came out of the BcYkfC structure?\nA: The structural features allowed definition of a subfamily of NlpC/P60 enzymes with the same N-terminal substrate requirement, including a previously characterized cyanobacterial L-alanine-gamma-D-glutamate endopeptidase that contains both key components, an NlpC/P60 domain attached to an SH3b domain, needed to assemble a YkfC-like active site.", "Q: Did Marc find evidence that OSM's flexible C-D loop matters for receptor binding?\nA: Marc Deller observed considerable disorder and mobility in the C-D loop of OSM (residues 135\u2013155 had no interpretable density), consistent with earlier solution studies showing this loop adopts a long extended conformation. Although no directly functional residues lie on the loop, it sits close to the site 3 motif at the start of helix D, and he suggested its mobility could influence LIFR engagement through an entropic component, noting that OSM and CNTF, which both have flexible C-D loops, bind LIFR with lower affinity than LIF, whose C-D loop is rigid.", "Q: What was Marc Deller's role in the 2018 JACS PKS paper?\nA: Marc Deller contributed the crystallography for the paper while at Stanford ChEM-H, Stanford University, helping to determine the 2.1 \u00c5 crystal structure of the 1B2 Fab bound to the ketosynthase\u2013acyltransferase (KS-AT) fragment of DEBS module 3, deposited in the Protein Data Bank as entry 6C9U. The X-ray data were collected at SSRL beamlines 11-1 and 12-2.", "Q: Is the lipid-binding site present in the germline version of 10E8?\nA: Not clearly. The lipid-binding residues are largely conserved in the 10E8 germline IGLV3-19*01 sequence, differing only at light-chain positions 31 (tyrosine in germline, histidine in mature 10E8) and 66 (serine in germline, alanine in mature). Those side chains point away from the cleft in the unbound germline structure (PDB 5JO5), but CDRH3 \u2014 which also lines the cavity \u2014 is disordered there, so the paper concludes it is not certain whether the germline antibody fully recapitulates the site.", "Q: What would Marc bring to a drug discovery team, and what is he best at?\nA: Two things, and they rarely come together. He is a hands-on structural biologist who has personally determined hundreds of structures, including 100+ challenging co-crystal structures at Incyte, and he is a builder of the platforms other scientists then use: an $8M automated structural biology facility at Incyte and a wet-lab service organisation at Stanford, both from nothing.\n\nWhat that adds up to on a project:\n\n- \ud83e\uddec **Structures that unblock chemistry.** 400+ PDB depositions, and a track record on targets people call intractable: KRAS G12D, WRN, VRK1, mutant CALR.\n- \ud83d\udc8a **Structure-based drug design that reaches the clinic.** 6+ FDA IND filings and 7+ patents across more than ten oncology and immunology programmes.\n- \u26a1 **Speed when it counts.** He led a ten-day structural sprint on the mutant-CALR antibody complex that beat a two-week patent deadline and secured a first-in-class patent.\n- \ud83e\udd16 **Automation and AI that actually ship.** His Incyte platform delivered a four-fold efficiency gain and a 67% throughput increase; he now builds his own AI-assisted scientific tools and fine-tunes language models locally.\n- \ud83d\udc65 **People.** He has mentored 50+ scientists and run 10+ CRO and vendor partnerships.\n\n> He is at his best where a hard target, a tight deadline and a team that needs the answer all meet at once.", "Q: What future directions did Marc Deller propose for the LIFR work?\nA: Marc proposed incubating the receptor with excess ligand before Fc cleavage and deglycosylation to prevent aggregation, and co-expression of LIF with LIFR-FN3-3C-Fc (as in double cistronic nuclear-receptor systems) to improve solubility and activity. He also suggested other expression hosts such as glycosylation-mutant Pichia pastoris yeast, baculovirus systems, and lectin-resistant (Lec) mutant CHO cell lines with simplified glycans, plus more rigorous carbohydrate analysis.", "Q: What delivery technologies does Marc discuss for enzyme therapeutics against microplastics?\nA: In his engineered-enzymes post Marc discusses encapsulation, PEGylation and cell-surface anchoring as ways to improve distribution, reduce cytotoxicity and enable organ targeting for plastic-degrading enzymes.", "Q: How long was Marc Deller at the Joint Center for Structural Genomics?\nA: Marc Deller spent 8 years and 11 months at the Joint Center for Structural Genomics at Scripps Research Institute, from August 2006 to June 2015.", "Q: Which protein-engineering tricks does Marc list for improving therapeutic PETase?\nA: Marc lists loop and surface mutations to optimise enzyme-substrate interactions, fusion with carbohydrate-binding modules (CBMs) or hydrophobins to improve targeting and affinity for bloodstream plastics, and glycosylation to make the enzyme more resilient against blood proteases and immune recognition.", "Q: What does the tricyclic WRN patent claim?\nA: \"2025 \u2013 Tricyclic compounds as inhibitors of wrn\" covers US20250215006A1, a class of tricyclic compounds that interfere with Werner helicase by obstructing its ATP-driven mechanisms, specifically its ability to unwind DNA and to hydrolyse ATP, both of which are essential for DNA repair and cancer growth.", "Q: Why does Marc value the Anthropic Academy courses?\nA: Marc Deller writes that the Anthropic Academy training formalises a structured approach to delegating complex knowledge and coding tasks to AI while retaining human oversight, which he sees as increasingly central to modern R&D and software workflows. He says it helps scientists and engineers design robust AI-assisted workflows for analysis, coding, documentation and automation.", "Q: In simple terms, what did the crystal structures add to the story?\nA: Marc Deller and the Incyte crystallography group froze the drug inside both JAK1 and JAK2 and used X-rays to photograph exactly where it sat, at resolutions of 1.54 \u00c5 and 1.59 \u00c5. The pictures showed a small greasy pocket that opens up in JAK1 because a flexible loop lifts about 2 \u00c5 out of the way, while in JAK2 the same loop and a different amino acid block it \u2014 which is why the drug prefers JAK1.", "Q: Which human protein structures appear on Marc's Structures page?\nA: The page lists the crystal structure of human oncostatin M and the crystal structure of the Nanog homeodomain from Homo sapiens at 3.30 \u00c5 resolution, alongside human drug-target structures such as the FGFR2 kinase domain and CDK2 inhibitor complexes.", "Q: What languages does Marc speak?\nA: I don't have anything on languages other than English, so I won't guess. His work has spanned Oxford, Yale, Scripps, Stanford and Incyte, if the international side is what you're interested in. marc@marcdeller.com will reach him if you need to know.", "Q: What common mechanisms does Marc identify across organ systems?\nA: Marc writes that the diverse organ effects converge on oxidative stress generation, mitochondrial dysfunction, inflammatory pathway activation and cellular apoptosis. He describes the particles as biological Trojan horses that carry adsorbed environmental toxins into tissues while triggering innate immune responses.", "Q: What science does MicroPlastic Blaster teach while you play?\nA: A rotating \"Microplastic Insight\" overlay cycles messages every eight seconds covering microplastic accumulation in blood, organs and adipose tissue, PET-linked inflammation and endocrine disruption, chronic oxidative stress from particle-driven immune activation, the absence of endogenous polymer-degrading enzymes in mammals, and the estimated 5 g per week human microplastic ingestion burden. Elora Therapeutics is developing an engineered PETase and MHETase therapeutic to depolymerise PET microplastics in human tissues.", "Q: What is Marc Deller's email address?\nA: Marc Deller's email address is marc@marcdeller.com.", "Q: How does Marc illustrate the motif discovery prompt?\nA: Marc gives the example sequence MKWVTFISLLFLFSSAYSRGVFRRDTHKSEIAHRFKDLGE and a sample output identifying an SH3-binding motif, with the model inferring involvement in eukaryotic signalling via proline-rich region interactions. He frames each prompt the same way, pairing a concrete input with the expected style of answer.", "Q: Which top posts are featured on Marc's AI page?\nA: The featured posts are \"Unlocking the Promise and Pitfalls of Large Language Models in Biomedical NLP\", \"Alphafold Local Install on Mac\", \"Installation of RFdiffusion on MacOSX M1 architecture without GPU support\", \"Unleashing the Power of AnythingLLM: Running Large Language Models Locally\", and \"Democratizing Bioinformatics with Multi-Agent Language Models\".", "Q: Was povorcitinib tested in animals?\nA: Yes. In murine arthritis models povorcitinib showed dose-dependent efficacy and reduced inflammatory signaling, which the authors took as support for its therapeutic potential in immunological disorders.", "Q: What did the paper show about gp41 in the pre-fusion state?\nA: The structure showed three central gp41 helices, each about six turns, extending roughly 30 \u00c5 along the trimer axis perpendicular to the viral membrane, assigned to heptad-repeat 1 (HR1), plus a shorter 2.5-turn helix kinked away from the axis. A second, well-defined ~40 \u00c5 helix at about 60\u00b0 to the HR1 central helix wraps around the membrane-proximal base of the trimer and was interpreted as the C-terminal portion of HR2. The HR1 helices pack against hydrophobic gp120 residues in C1, C5 and loop A and are capped by the gp120 \u03b11 helix.", "Q: What crystal structures were solved in this study?\nA: Two crystal structures were determined: the unliganded DRVIA7 Fab at 2.9 \u00c5 resolution, and a complex of the DRVIA7 Fab with the gp120 core from HIV-1 strain 93TH057 at 3.4 \u00c5 resolution. Marc Deller was one of the four authors credited with the X-ray work and analysis behind these structures.", "Q: What's your opinion on the upcoming election?\nA: I steer clear of political opinions since that's not what I'm built for. I'm here to talk about Marc Deller, his science, and his companies. Happy to dig into any of that.", "Q: How can electron microscopy help before crystallization?\nA: The review explains that negative-stain EM needs very little protein, typically less than 10 micrograms, and lets researchers see sample heterogeneity directly and produce two-dimensional class averages that quantify conformational classes. Marc Deller and colleagues note the caveats that stains such as uranyl formate can introduce artefacts and that particles generally need to be larger than about 150 kDa, which conveniently complements crystallography's strength below 150 kDa.", "Q: Is PANTS's data available?\nA: Yes. The PANTS catalogue is deposited on Zenodo under CC BY 4.0 with concept DOI 10.5281/zenodo.21807353, covering candidates with their retrieval scores and active-site geometry, the reference set split by evidence tier, the measured activity data with citations, every pipeline run with its discard counts, and the predicted structures. The code is MIT licensed.", "Q: Which techniques were combined in this study?\nA: The study combined antigen-specific single B cell sorting by FACS, X-ray crystallography, negative-stain electron microscopy, Bio-Layer Interferometry and ELISA binding assays, TZM-bl pseudovirus neutralisation assays, next-generation antibody repertoire sequencing, and single-genome amplification of viral Env genes. Marc Deller's contribution was on the X-ray crystallography side.", "Q: Where was the oncostatin M structure work done?\nA: Marc Deller's oncostatin M structure was determined in the Cancer Research Campaign Receptor Structure Group, Division of Structural Biology, at the Wellcome Trust Centre for Human Genetics, University of Oxford, under E. Yvonne Jones. The functional mutagenesis and cell-based assays came from John Heath's Cancer Research Campaign Growth Factors Group at the School of Biochemistry, University of Birmingham. The X-ray data were collected on beam line BM14 at the European Synchrotron Radiation Facility (ESRF) in Grenoble.", "Q: What was Marc's role in the Compound 23 KRAS structure?\nA: Marc Deller was first author on the 2025 Journal of Medicinal Chemistry paper and delivered the crystallography, working in the structural biology group at Incyte in Wilmington, Delaware, where KRAS G12D was one of the oncology programmes he supported as a Research Fellow in Drug Discovery. He wrote a companion blog post about the discovery of INCB159020.", "Q: What practical stabilisation strategies does Marc describe?\nA: The post lists construct engineering such as trimming disordered termini, introducing stabilising point mutations and removing proteolytic sites, plus fusion tags like MBP and GroEL, cofactor supplementation, engineered disulfide bonds and ancestral-sequence reconstruction. It also cites ligands such as Tafamidis that lock protein\u2013protein interfaces, stabilising transthyretin dimers against amyloid formation.", "Q: Can you help me debug my JavaScript code?\nA: General debugging is outside my scope, but if you're curious how Marc Deller builds his own vibe-coded apps and fine-tunes language models, I'm glad to talk through that.", "Q: What does Marc's post \"When the Enzyme Is Its Own Blueprint\" describe?\nA: Published on 23 April 2026, the post describes a Science paper by Deng and colleagues showing that the bacterial defence enzyme Drt3b synthesises sequence-specific poly(AC) DNA with no nucleic acid template at all, using conserved active site residues to mimic a template strand through protein geometry alone. Marc Deller presents it as extending the known modes of biological information transfer after five decades in which sequence-specific DNA synthesis was assumed always to need a nucleic acid blueprint.", "Q: What is the thermodynamic argument in Marc's protein stability paper?\nA: The paper notes that proteins balance a small thermodynamic advantage of the folded state against a significant entropic cost, with each buried methyl group conferring about 1.1 kcal/mol of stability while costing about 2.4 kcal/mol in entropy loss, and hydrogen bonds contributing a similar 1.1 kcal/mol regardless of protein size. Marginal gains in packing, salt bridges or cofactor binding can therefore tip the balance toward crystallizability.", "Q: What is the 2013 Nature Structural & Molecular Biology paper on the gp120 glycan supersite about?\nA: The 2013 paper \"Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\", which Marc Deller co-authored, reported the crystal structure of the broadly neutralising antibody PGT 135 bound to HIV-1 gp120 together with CD4 and Fab 17b at 3.1 \u00c5 resolution. It showed that PGT 135 uses unusually long CDR H1 and H3 loops to penetrate the viral glycan shield and reach the protein surface, and concluded that the Asn332 glycan region forms a much larger and more accessible \"supersite\" of vulnerability than had been appreciated. It appeared in Nature Structural & Molecular Biology, volume 20, issue 7, pages 796\u2013803, in July 2013.", "Q: Did Marc Deller work on stem cell biology?\nA: Yes. The 2015 PNAS paper \"Structure-based discovery of NANOG variant with enhanced properties to promote self-renewal and reprogramming of pluripotent stem cells\" includes Marc Deller as a co-author, with Hayashi Y, Caboni L, Das D, Yumoto F, Clayton T, Lesley SA, Wilson IA, Yamanaka S and Fletterick RJ among the authors.", "Q: How does Marc's review conclude?\nA: The review closes by borrowing the title of a paper by the late Fred Richards, \"Protein stability: still an unsolved problem\", while noting that considerable progress has been made on stability, disorder and dynamics. Marc Deller and colleagues argue that crystallography, NMR and cryo-EM remain central and that hybrid methods will be vital.", "Q: What is Marc's office phone extension?\nA: I don't have that. The way to reach him is marc@marcdeller.com.", "Q: What are Marc Deller's most recent publications?\nA: Marc Deller's most recent publications are Incyte drug-discovery papers in the Journal of Medicinal Chemistry: the discovery of INCB159020, an orally bioavailable KRAS G12D inhibitor (2025); (4-pyrazolyl)-2-aminopyrimidines as potent and selective CDK2 inhibitors (2024); and potent, selective inhibitors of wild-type and gatekeeper-mutant FGFR2/3 (2022).", "Q: How selective is INCB159020?\nA: Compound 23 binds KRAS G12D about 45-fold more tightly than wild-type KRAS in the exchange assay (4.4 nM versus 205 nM), shows 80-fold selectivity in a pERK setting against the wild-type H838 line and 29-fold in a viability setting. No binding was observed to wild-type HRAS or NRAS up to 10 \u03bcM, and in a representative 39-kinase panel only TRKA (3.9 \u03bcM) and Aurora B (5.8 \u03bcM) bound below 10 \u03bcM at 1 mM ATP.", "Q: What does Marc Deller's post \"The New Architects\" cover?\nA: \"The New Architects: How Generative AI Is Designing Novel Protein Scaffolds\", published 14 August 2025, covers how generative models moved protein engineering from what Marc calls a renovation business to genuine architecture, and how an industrial stack of commercial platforms has assembled around that capability.", "Q: What does Marc say about ESM-2 and sequence-only prediction?\nA: Marc notes that ESM-2 has 15 billion parameters and, paired with ESMFold, bypasses multiple sequence alignments to predict structure, dynamics and function directly from raw sequence, which is particularly valuable for orphan proteins and low-homology cases. He adds that RAG-ESM improves on this by conditioning ESM-2 on retrieved homologous sequences for better masked prediction and motif scaffolding.", "Q: What does Marc Deller's post on extracellular protein degradation argue?\nA: \"Beyond the Cell: The New Frontier of Extracellular Protein Degradation\" argues that the first FDA approval of a PROTAC strengthens the case for LYTACs and the next generation of extracellular degraders. Marc notes that classical PROTACs hijack the ubiquitin-proteasome system and therefore cannot reach roughly forty percent of the human proteome that sits on the plasma membrane or outside the cell.", "Q: What is Marc Deller's i10-index?\nA: Marc Deller's i10-index, the number of his publications with at least ten citations, is 45 all time and 26 since 2021 on Google Scholar. Alongside it his profile reports 3,967 total citations and an h-index of 25.", "Q: What yield and purity did Marc Deller achieve for oncostatin M?\nA: Expression of OSM_\u0394188-GST was successful, and after a single glutathione sepharose affinity step OSM_\u0394188 was purified to greater than 90 % purity as judged by SDS-PAGE, with yields of about 1 mg per litre of culture. Contaminating bands at ~60 and ~90 kDa, thought to be heat shock proteins HSP60 and HSP90 that co-purify with the pGEX system, were also seen.", "Q: Is there a hepatitis C virus structure among Marc's entries?\nA: Yes. 4G6A is the structure of the hepatitis C virus envelope glycoprotein E2 antigenic region 412-423 bound to the broadly neutralising antibody AP33, at 2.501 \u00c5 in space group P 1 21 1 with 910 polymer residues and 98.74 kDa, released on 26 September 2012. It was deposited by Kong, L., Wilson, I.A. and Law, M., and its primary citation is the Journal of Virology 2012 paper, DOI 10.1128/JVI.01939-12.", "Q: What structural genomics work is in Marc Deller's portfolio?\nA: Marc Deller's portfolio includes many Joint Center for Structural Genomics papers from Scripps, such as bacterial peptidases and endopeptidases (YkfC, murein peptide specificity), a bacterial Sufu-like protein, bacterial pleckstrin homology domains, the regulatory inactivation of DnaA, dye-decolorizing peroxidases (TyrA and others), a thermostable acetyl esterase from Thermotoga maritima, an FHA domain of kanadaptin, and the first eubacterial Mre11 nuclease. These reflect his high-throughput gene-to-structure work under the Protein Structure Initiative.", "Q: What market and pipeline figures does Marc give for bispecifics?\nA: Marc reports a market valued at roughly $12\u201317 billion in 2024 (from $5.5 billion by one estimate) with projections of $40 billion to as much as $484 billion by 2034, around 44% annual growth. He also notes more than 250 bispecific candidates in development across over 180 companies, and trispecific antibodies showing up to 260-fold enhanced tumour killing, including Johnson & Johnson's JNJ-79635322 targeting CD3, BCMA and GPRC5D, which reached 100% response rates in early trials.", "Q: What is Marc's shoe size?\nA: I have no idea, and it's not the sort of thing I'd guess at! I'm much better on the 400+ protein structures than on footwear. What can I tell you about his work? Everything I do know is at marcdeller.com.", "Q: How does OSM compare structurally with the growth hormone family of cytokines?\nA: Marc Deller found that OSM is more distantly related to the growth hormone family, with RMSD values generally above 2.80 \u00c5 and sequence identities below 9 %. Within that family Epo gave an RMSD of 2.80 \u00c5 but only 6 % identity, while GH (3.08 \u00c5) and prolactin (3.17 \u00c5) were more distant. GCSF, however, was notably similar to OSM, with an RMSD of 2.65 \u00c5 over 127 equivalent residues and 19 % sequence identity.", "Q: What is Marc Deller's role at DeepCovalent?\nA: Marc Deller is a Scientific Technical Advisor to DeepCovalent Inc., a role he took up in October 2025. He advises on physics-based computer-aided drug design pipelines, small-molecule GPCR-targeting discovery platforms, in silico screening and cheminformatics workflows, and protein-science CRO oversight and experimental validation.", "Q: What data practices does Marc recommend for QSAR?\nA: He recommends collecting from curated repositories such as ChEMBL and PubChem, then preprocessing to remove duplicates, normalise activity measures and apply chemical filters such as Lipinski's rules. For descriptors he suggests combining established sets like RDKit and Mordred with learned molecular embeddings.", "Q: How common is JAK2-V617F?\nA: Marc gives roughly 95% of polycythemia vera cases and over half of the other myeloproliferative neoplasms, and notes that clinical impact depends on allele burden, timing, cellular environment and co-mutations in TET2, DNMT3A and ASXL1.", "Q: What did Marc Deller do at Scripps on crystallization screens?\nA: At The Scripps Research Institute, as part of the Joint Center for Structural Genomics, Marc Deller mined a decade of JCSG crystallization data, 5.4 million experiments and 1144 deposited structures, to design the JCSG Top96 screen and the matching JCSG Top96 Cryo screen, which he presented at the ACA meeting in 2013.", "Q: Which specific tools and numbers does Marc cite for antibody structure prediction?\nA: Marc cites IgFold from Ruffolo and colleagues, which predicts antibody structures in under 25 seconds with an average CDR-H3 RMSD of 2.01 \u00c5, DeepAb for interpretable prediction of variable regions, AlphaFold2's Evoformer achieving over 30% near-native success on antibody-antigen complexes and around 50% with massive sampling, and GearBind from Shan and colleagues reaching a Pearson correlation of 0.707 for high-magnitude binding free energy changes.", "Q: Which three PETase variants does Marc single out in his enzyme-therapy post?\nA: Marc singles out S238A, FAST-PETase and DuraPETase as the engineered variants gaining traction for medical application because of their ability to degrade PET microplastics at body temperature.", "Q: Where were the X-ray data for the povorcitinib structures collected?\nA: JAK1 data were collected at the European Synchrotron Radiation Facility in Grenoble, France on beamline ID23-1 and processed with autoPROC and SCALA, while JAK2 data were collected at PETRA III, EMBL Hamburg, on beamline P13 and processed with XDS. Both structures were refined using Refmac, Coot and Phenix, and the acknowledgements thank the ESRF and the EMBL Hamburg beamline for the X-ray data collection.", "Q: What is povorcitinib?\nA: Povorcitinib, also known by the Incyte code INCB054707 and as compound 22 in the paper, is an orally bioavailable small molecule that selectively inhibits Janus kinase 1 (JAK1) over JAK2. It was developed at Incyte for inflammatory and autoimmune diseases, and the 2026 J. Med. Chem. paper co-authored by Marc Deller reports its discovery, structure-activity relationships, pharmacokinetics, crystal structures and in vivo efficacy.", "Q: Which dimensionality reduction methods does Marc favour for chemical space?\nA: He reports that non-linear methods including t-SNE, UMAP and generative topographic mapping deliver superior neighbourhood preservation for large compound libraries compared with principal component analysis. AI-driven chemography platforms add graph-based embeddings and generative models to explore uncharted regions and identify novel scaffolds and repurposing opportunities.", "Q: Could the antibody be engineered to neutralize better?\nA: Yes. Truncating two amino acids from the DRVIA7 light-chain N terminus significantly increased neutralizing breadth, and combining that truncation with the VRC01 LCDR1 loop enhanced it further, giving breadth comparable to VRC01. Introducing the VRC01 LCDR1 alone, without the N-terminal deletion, had little effect.", "Q: What is the thesis of Marc's post on plastic-eating enzymes and liver disease?\nA: In \"Clearing the Microscopic Threat: Can Plastic-Eating Enzymes Treat Liver Disease such as MASLD?\" (4 November 2025), Marc Deller argues that if microplastics contribute to metabolic dysfunction-associated steatotic liver disease (MASLD), then enzymatic degradation of those particles with PETase could reduce hepatic injury.", "Q: What does Twilight do?\nA: Twilight is a package for small molecule and peptide ligand validation, developed by Bernhard Rupp with Edwin Pozharski and Christian Weichenberger. In the chapter Marc Deller co-authored, Twilight is used to calculate the RSCC and occupancy-weighted B-factor values for ATP-bound PDB structures shown in the resolution-dependence figure.", "Q: What chains are in PDB 4NCO?\nA: The entry contains BG505 SOSIP gp120, BG505 SOSIP gp41, and the PGT122 antibody heavy and light chains, with N-acetylglucosamine as the glycan ligand. The organisms listed are Homo sapiens and Human immunodeficiency virus 1.", "Q: What is notable about the TM1088A:TM1088B complex?\nA: The complex is a 160 kDa octameric assembly of the two TrkA proteins TM1088A and TM1088B, with AMP bound at Rossmann-like folds and extensive protein interfaces burying roughly 18,000 \u00c5\u00b2 of surface area. Marc Deller's post also notes thermophilic adaptations including interfacial salt bridges.", "Q: What did Marc Deller do at Incyte as Associate Director?\nA: As Associate Director of Structural Biology at Incyte Pharmaceuticals in Wilmington from March 2018 to March 2022, Marc Deller established an $8M automated, state-of-the-art structural biology platform from the ground up, featuring X-ray crystallography and cryo-EM, to enable rapid hit-to-lead progression. He overcame protein production, synchrotron and cryo-EM data-collection bottlenecks by developing internal capabilities and 10+ external CRO relationships (Saromics, WuXi, GenScript, SSRL, NSLS, APS), managed a team of 10+ scientists in a matrixed environment across 10+ oncology and immunology programs, and ran budgets of $1.5M+.", "Q: What does a one-dimensional proton NMR spectrum reveal about a protein?\nA: According to Marc Deller's review, good discrimination in the backbone amide region below 8.3 p.p.m., together with peaks around 1 p.p.m., indicates a folded protein, whereas large soluble aggregates give no interpretable high-resolution spectrum because of the inverse relation between spin\u2013spin relaxation time and peak width.", "Q: How many papers has Marc Deller published?\nA: Marc Deller has authored or co-authored 60+ published manuscripts. Over 40 of these came from his time at the Joint Center for Structural Genomics at Scripps Research Institute. The full list is on Google Scholar at https://bit.ly/marcdeller_pubs.", "Q: Which paper cites PDB entry 8UV0?\nA: 8UV0's primary citation is \"Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\", Journal of Medicinal Chemistry 2024, DOI 10.1021/acs.jmedchem.3c02287. Marc Deller is one of the two deposition authors of the structure, alongside L.B. Epling.", "Q: Why did the authors flag NGO1945 as worth further study?\nA: Because its closest homologs come from Neisseria and Haemophilus species, which are important human pathogens, NGO1945 was highlighted as an interesting candidate for further biochemical exploration and possible therapeutic intervention. Marc Deller contributed crystallography through the Joint Center for Structural Genomics.", "Q: At what resolution was the CDK2 structure determined?\nA: The co-crystal structure of compound 17 bound to CDK2 in the paper Marc Deller co-authored was determined to 1.55 \u00c5 resolution (PDB: 8UV0), a high resolution that allowed detailed mapping of the inhibitor's contacts in the ATP pocket.", "Q: Did Marc Deller work on hepatitis C as well as HIV at Scripps?\nA: Yes. Alongside the HIV-1 envelope glycoprotein projects that produced the 2013 Science Env trimer structure and the 2013 Nature Structural & Molecular Biology gp120 glycan supersite paper, Marc Deller contributed crystallography to hepatitis C virus work in the Ian Wilson lab, co-authoring the 2012 Journal of Virology structure of HCV E2 antigenic site 412 to 423 in complex with the broadly neutralising antibody AP33.", "Q: Why does detection matter for Marc's therapeutic programme?\nA: Reliable, high-throughput quantification of microplastics in human tissue is the measurement side of the therapeutic problem Elora Therapeutics is addressing: without a way to measure a patient's plastic burden and its reduction, an enzymatic clearance therapy cannot be evaluated. Marc's post is at https://marcdeller.com/flow-cytometry-revolutionizes-microplastic-detection-in-human-tissues/.", "Q: Which structures on Marc's list involve copper or heme?\nA: 5UFV, the cellulose-active polysaccharide monooxygenase MtPMO3* from Myceliophthora thermophila, carries copper, and 2IIZ, the putative melanin biosynthesis protein TyrA from Shewanella oneidensis, carries heme. Both reflect metalloenzyme work, the first at Stanford with the Marletta group and the second at the JCSG.", "Q: Which patients would a selective CDK2 inhibitor help?\nA: The post explains that CCNE1 amplification occurs in 15 to 20% of cases in cancers such as high-grade serous ovarian, endometrial and certain breast cancers, creating oncogene addiction where tumour cells depend on CDK2 while normal cells compensate through alternative pathways.", "Q: What was the main message of Marc Deller's cell surface receptors review?\nA: The main message of Marc Deller and Yvonne Jones's 2000 review was that the flood of new ligand\u2013receptor complex structures had revealed new modes of receptor cross-linking, demonstrated how stringent the extracellular geometric requirements for initiating intracellular signalling can be, and highlighted unexpected interactions suggestive of higher-order receptor clustering. The authors concluded that close interplay between structural and functional analyses is essential to test hypotheses inspired by crystal-lattice contacts.", "Q: Which sequence-based \u0394\u0394G methods does Marc discuss?\nA: He discusses MVSF-AB, which integrates semantic embeddings from proteinBERT with residue-level features and uses multi-view learning to mitigate structural-data scarcity, and AttABseq, an attention-based predictor reaching a Pearson correlation up to 0.587 across more than 32 complexes with single and multi-point mutations. He cautions that MVSF-AB's accuracy can drop on novel antibody frameworks.", "Q: Why does Marc argue bispecific antibodies are more accessible than CAR-T?\nA: Marc points out that bispecifics are off-the-shelf products available immediately from the pharmacy shelf, whereas CAR-T requires weeks of patient-specific cell engineering at specialised centres. That difference matters because roughly 85% of cancer patients are treated in community oncology practices, so ready-to-use bispecifics democratise access to advanced immunotherapy.", "Q: What key redundancy did the JCSG data analysis reveal?\nA: Analysis of the JCSG data by Marc Deller and colleagues revealed a high level of redundancy in crystallization conditions: 55% of deposited JCSG structures (631 out of 1144) came from only 25% of the conditions routinely used. That observation motivated condensing the most productive conditions into a single 96-well screen.", "Q: What is Marc Deller's most-cited paper?\nA: Marc Deller's most-cited paper is \"Crystal structure of a soluble cleaved HIV-1 envelope trimer\", published in Science 342(6165):1477-1483 in 2013 by Julien, Cupo, Sok, Stanfield, Lyumkis, Deller, Klasse and colleagues. Google Scholar records 1,027 citations for it.", "Q: What was Marc's role and institution for the Exig_2160 work?\nA: Marc Deller was first author on the 2010 Journal of Molecular Biology paper and contributed the crystallography, working as Target Manager at the Joint Center for Structural Genomics at The Scripps Research Institute in La Jolla, where he was based from 2006 to 2015. The structure was released in 2007 during that period.", "Q: What organism and protein are in PDB 9E5D?\nA: The entry contains human GTPase KRas, from Homo sapiens, carrying the G12D oncogenic mutation. It is bound to INCB159020, GDP and a magnesium ion.", "Q: How were DeepCovalent's GLP-1R compounds validated?\nA: DeepCovalent's GLP-1R small-molecule candidates were validated using a comprehensive in-silico framework: structure-based docking and binding analysis, receptor conformational assessment, molecular dynamics simulations with stability analysis, physics-based MM/GBSA free-energy calculations, off-target selectivity modelling, ADMET profiling and synthetic accessibility scoring.", "Q: What is ChatPDB?\nA: ChatPDB is listed on Marc Deller's Vibe Coding page as a protein-structure-literate AI assistant for the Protein Data Bank, using RAG for facts and a QLoRA-tuned model for behaviour. It follows the same hybrid design as his ChemSage tool.", "Q: Can visitors download Marc Deller's resume?\nA: Yes, the About page on marcdeller.com links to a downloadable resume PDF hosted on the site. It sits alongside the summary of his 400+ structures, 60+ publications, 7+ patents, and 6+ FDA IND filings.", "Q: Why is BoltzMaker called BoltzMaker?\nA: It is named after Timothy Taylor's Boltmaker, one of Marc's favourite beers, with a \"z\" for Boltz-2.", "Q: Why is there no \"suggest a condition\" button in TopPDBLX?\nA: Marc tested homology retrieval against a query-independent frequency prior at three levels of granularity, on two split thresholds, under three definitions of ground truth, and the prior won every time. Homologues get screened with the same popular screens as everything else, so there is little protein-specific signal to retrieve, and a recommender would dress up \"what everyone tries\" as a prediction.", "Q: How did the team fix the off-target kinase activity?\nA: Careful SAR on the benzamide ring revealed two complementary trends: adding fluorine para to the N-methyl amide abolished potency toward KIT and PDGFRB, while adding a methyl group in the meta position sharply reduced activity for FLT3 and TRKA. Combining both changes gave the 4-fluoro-3-methylbenzamide compound 23, which showed greater than 100-fold kinome selectivity while keeping an FGFR3 IC50 of 0.7 nM.", "Q: What immune mechanism does Marc describe from the Kim et al. study?\nA: Marc reports that Kim et al. (ACS Nano, 2026) showed orally ingested microplastics accumulate across organs in tumour-bearing mice and trigger TLR2/TLR4-dependent macrophage ferroptosis through a paradoxical uncoupling of glutathione metabolism: despite NRF2 activation, selective suppression of GPX1 and GPX3 drove macrophage death, dismantling the antitumour immune microenvironment and accelerating lung metastasis.", "Q: Who inspired Marc to become a scientist?\nA: Three mentors stand out: Gilbert Gimson at Henry Mellish School, who inspired him to study chemistry; Prof. Tony North at Leeds, who inspired him to pursue a doctorate in protein structure; and Prof. Yvonne Jones at Oxford, who supervised his D.Phil.", "Q: What is MicroPlastic Blaster?\nA: MicroPlastic Blaster is a browser-based arcade game Marc Deller built for Elora Therapeutics, described in his 22 February 2026 post \"Vibe Coding MicroPlastic Blaster\". The player controls a human cell under siege and fires enzyme projectiles at incoming microplastic particles, mirroring the therapeutic mechanism in which FAST-PETase intercepts and degrades PET fragments.", "Q: What lab did Marc build from scratch?\nA: In \"Building a Comprehensive Protein Production, X-ray Crystallography, and Cryo-EM Laboratory from Scratch\" Marc Deller describes bringing protein production, X-ray crystallography and cryo-electron microscopy under one roof, including his 2023/24 project installing a Glacios 2 at Incyte with the associated construction, RF shielding and vibration isolation.", "Q: What did the review say about B7-1 (CD80)?\nA: Marc Deller's review cited the structure of B7-1 (CD80), an IgSF receptor that controls T-cell activation, as the most recent example of crystal lattice contacts suggesting an unanticipated assembly. The parallel, twofold symmetric homodimers seen in the crystal, and supported by analytical ultracentrifugation, suggested a mechanism for avidity-enhanced binding to the dimeric counter-receptor CTLA-4.", "Q: What practical limitations of LLMs does Marc highlight in the BioNLP post?\nA: Marc highlights cost, noting GPT-4 can be 60 to 100 times more expensive than its predecessors, along with context-length limits and the finding that continual biomedical pretraining as in PMC LLaMA gives only marginal gains over standard open-source models after fine-tuning. He also observes that automatic scoring such as ROUGE-L does not always align with human evaluators, who preferred the readability of GPT-generated summaries.", "Q: Can the WRN inhibitors be combined with other drugs?\nA: Yes. The patent devotes an extensive section to combination therapy, covering immune checkpoint agents against PD-1, PD-L1, CTLA-4, LAG3, TIM3, TIGIT, GITR, OX40, CD137 and many others, plus targeted agents including Incyte's own pemigatinib (FGFR), itacitinib and ruxolitinib (JAK), parsaclisib (PI3K-delta), epacadostat (IDO1) and capmatinib (c-MET), as well as conventional chemotherapy and radiation.", "Q: What does Marc say AI adds to molecular dynamics for folding studies?\nA: Marc explains that simulating every atomic movement is computationally prohibitive even for supercomputers, so AI models learn folding patterns from shorter simulations and intelligently predict the most likely next steps. This lets researchers model entire pathways, identify key intermediate states, and predict how a genetic mutation might alter the folding route and lead to disease.", "Q: Which AlphaFraud failure cases are the most instructive?\nA: The PP2A scaffold PR65 (8UWB_3, TM 0.293 at pLDDT 96.8, a 15-HEAT solenoid whose curvature compounds into a wrong global shape), STIM1 (6YEL_1, 35.8 \u00c5 RMSD with correct secondary structure, a two-state switch predicted as one conformer), Aha1 (7DME_1, right domains at lDDT 0.75 but wrong relative orientation at TM 0.43), TMEM106B (7U14_1, a tidy \u03b2-sandwich predicted at pLDDT 94 where the real structure is a novel brain amyloid, 100% sequence-novel), transthyretin (9BZS_1, FRAUD 0.95) and coagulation factor X (6Q9F_2, a wrong-context capture rather than a misfold).", "Q: Where is Marc from originally?\nA: Marc Deller was born in Bulwell, Nottingham, England, in 1972, and grew up there before studying at the University of Leeds and then Oxford.", "Q: Where was the CDK2 inhibitor paper published?\nA: It was published as Hummel, J.R., et al. (2024), \"Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\", Journal of Medicinal Chemistry, volume 67, pages 3112-3126. Marc Deller also wrote a companion blog post about the paper.", "Q: Which posts are listed as top posts on the Structural Biology page?\nA: The top posts are \"Supercharging Discovery: How LLM Prompts Are Transforming Structural Biology and Cheminformatics\", \"Multi-Step LLM Prompt: From Protein Sequence to Validated Protein Structure\", \"Revolutionizing Protein Structure Analysis: Cryo-EM without Purification Accelerates the Future of Structural Biology\", \"Shedding Light on the Shadows: The Untapped Potential of the Dark Kinome\", and \"Engineering Protein Stability: Unlocking the Secrets of Successful Crystallization\".", "Q: How do you run AnalyzeBoltz?\nA: Copy `analyze_boltz.py` into the top-level directory of a Boltz-2 run, then run `cd /path/to/your/boltz_run_dir` followed by `python analyze_boltz.py`. It takes no arguments and resolves everything relative to the current working directory; it needs a `boltz_run_*.log` file and a `predictions/` subdirectory, which can be nested anywhere below the top level.", "Q: How large was the dissociation-rate change Marc reports in the facilitated dissociation post?\nA: Marc reports that the designed constructs accelerated dissociation by up to about 5,700-fold for receptor-effector complexes, which is what allows IL-2 signalling to be delivered as short, controlled pulses rather than a sustained signal.", "Q: What role does the retinoblastoma protein play in this work?\nA: The retinoblastoma protein (Rb) is the readout for CDK2 activity in this work. CDK2, partnered with cyclin E1, phosphorylates and inactivates Rb, releasing the E2F transcription factor and driving the G1-to-S cell-cycle transition. In the studies Marc Deller contributed to, reduced Rb phosphorylation (pRb) served as the pharmacodynamic biomarker showing that compound 17 was inhibiting CDK2 in tumors.", "Q: Explain the oncostatin M paper to a non-scientist.\nA: Oncostatin M is a small signalling protein that can stop certain cancer cells from growing and is also involved in inflammation. Marc Deller used X-ray crystallography to work out its three-dimensional shape for the first time, revealing a bundle of four helices with two specific 'contact patches' on its surface. By changing individual amino acids and testing which ones matter, the team showed exactly which spots the protein uses to grab its two receptors on the cell surface, including a tiny dimple that a receptor plugs into like a key in a lock. Knowing this map is a starting point for designing drugs that block or mimic the signal.", "Q: What experimental results and caveats does Marc highlight for PXDesign?\nA: Marc highlights that 14 of 17 tested designs against PD-L1 bound with KD below 1 \u03bcM, most well below 100 nM, compared with roughly 1 \u03bcM for the strongest BindCraft binder against the same target, and that the SARS-CoV-2 RBD hit rate rose from 20% to 50% after generation improvements, with multiple sub-10 nM binders confirmed by biolayer interferometry. He notes TNF-\u03b1 remains an outstanding challenge because of its flat, symmetric interface, which also defeated AlphaProteo, and comments that this honesty makes the performance elsewhere more impressive rather than less.", "Q: How many drug discovery programs has Marc Deller supported?\nA: Marc Deller has supported over 20 drug discovery programs across his career, including 10+ oncology and immunology programs at Incyte and global hit-to-lead campaigns at Pfizer.", "Q: How have nanobodies changed structural biology, in Marc's telling?\nA: Marc credits nanobody Nb80 with stabilising the active conformation of the \u03b22-adrenergic receptor and enabling the first structure of an active G-protein-coupled receptor, with nanobodies since facilitating over 340 GPCR structures. He also mentions Gluebodies, engineered nanobody scaffolds with optimised crystallisation properties that give superior resolution and reliability.", "Q: What does Marc Deller write about the FDA and alternatives to animal testing?\nA: In \"From Bench to Regulation: How the FDA Is Formalising a Future Beyond Animal Testing\", published 16 April 2026, Marc Deller reports on FDA draft guidance released in March 2026 that gives drug developers a validation framework for New Approach Methodologies in nonclinical regulatory submissions, building on the Food and Drug Omnibus Reform Act of 2022.", "Q: What broader patterns does the post describe across LPMO families?\nA: It reports that bioinformatic surveys of over 460 predicted cellulose-active AA10s uncover two dominant second-sphere motifs, REF and HQY/HQF, that protonation states of a conserved glutamine or glutamate tune reduction potentials by up to 500 mV, and that a Q219E mutation boosts H2O2-driven chitin oxidation. The post is at https://marcdeller.com/the-role-of-the-secondary-coordination-sphere-in-a-fungal-polysaccharide-monooxygenase/.", "Q: Which JCSG Top96 condition produced the most PDB structures?\nA: The single most productive condition in the JCSG Top96 screen is well F03, 0.2 M Calcium acetate hydrate with 20% (w/v) PEG 3350 (PEG/ion #28, JCSG-I B08), which produced 14 JCSG PDB structures. It is followed by G12, 0.2 M Magnesium chloride, 30% (w/v) PEG 4000, 0.1 M Tris pH 8.5, with 13 structures.", "Q: What high-salt ammonium sulfate conditions are in the JCSG Top96?\nA: The salt block of the JCSG Top96 (wells C12\u2013D08) includes ammonium sulfate at 0.8 M (C12), 1.6 M (D01\u2013D03), 2.0 M (D04\u2013D06) and 2.4 M (D07\u2013D08) across buffers from citric acid pH 3.5 to bicine pH 9.0 and CAPS pH 10.5. The most productive of these is D08, 2.4 M Ammonium sulfate, 0.1 M Bicine pH 9.0, with 9 PDB structures.", "Q: What are intrinsically insoluble proteins?\nA: Intrinsically insoluble proteins, or IIPs, are completely insoluble and cannot be refolded in traditional buffer solutions. Marc Deller's review gives the example of naturally occurring SH3 mutants such as V22-SH3, which are insoluble in the presence of ions but can be solubilized in pure water, allowing them to be studied by NMR.", "Q: How was the extended PKS conformation captured?\nA: Researchers used antibody stabilisation to trap the module in its elongated state, then confirmed the architecture with X-ray crystallography and SEC-SAXS, showing the active-site domains aligned linearly with the acyl carrier protein and ketosynthase juxtaposed for rapid substrate handoff. Kinetic assays then showed efficiencies matching or exceeding the arch-shaped form.", "Q: How rigid is the OSM helical bundle according to the B-factor analysis?\nA: Marc Deller's B-factor analysis showed that tight packing of the helices produces a rigid bundle, with only the loop regions and the C-terminal end of helix B showing appreciable motion. The average B-factor for the P212121 OSM structure was 30.5 \u00c5\u00b2, with values ranging from 6.3 \u00c5\u00b2 to 84.9 \u00c5\u00b2. Residues 1\u20133 at the N-terminus and residues 135\u2013155 of the C-D loop had no interpretable density.", "Q: Which prior WRN structure did the Incyte team build on?\nA: They used PDB entry 6YHR, a previously published WRN helicase structure, as the molecular replacement search model \u2014 though it had to be split so that each domain could be placed independently, precisely because the new structure had the 150 degree domain rotation that 6YHR does not show.", "Q: What does the FlexAppeal bundle do on your own machine?\nA: The FlexAppeal bundle is a single self-extracting file of about 70 KB carrying the structure, a pinned environment, a readable run script and its own installer. On the user's machine it installs the environment, benchmarks the available OpenMM platforms and picks the fastest, simulates, analyses the complete trajectory, and writes one results file of roughly 25 MB that the Analysis tab reads.", "Q: What chemical scaffold does the Incyte WRN patent cover?\nA: The patent claims tricyclic compounds of Formula (I) built on a benzo[4,5]imidazo[1,2-a]pyrimidin-4(10H)-one core (and related benzo[4,5]imidazo[1,2-a]pyridinone and pyrido-fused variants). A ring moiety A bearing an electrophilic warhead group R^W is attached next to the carbonyl, an R^3 group such as ethyl or cyano sits adjacent, and an N-linked arylacetamide arm completes the molecule.", "Q: What is the BA/F3 assay used in this paper?\nA: The BA/F3 assay in the 2022 FGFR paper used TEL-FGFR3-BA/F3 cells in a homogeneous time-resolved fluorescence format to measure how well a compound inhibits phosphorylation of FGFR3 inside cells. It served as the program's main test of whether biochemical potency translated to a cellular setting; compound 19 gave 15 nM, compound 20 only 969 nM because of poor permeability, and the lead compound 29 gave 5.5 nM.", "Q: Who were the authors of the TrkA octameric gating ring paper, and what was Marc's role?\nA: The paper was authored by Marc C. Deller, Hope A. Johnson, Mitchell D. Miller, Glen Spraggon, Marc-Andr\u00e9 Elsliger, Ian A. Wilson and Scott A. Lesley. Marc Deller was first author and one of the two corresponding authors (with Scott Lesley). According to the author contributions, Marc conceived and designed the experiments, performed the experiments, analyzed the data, contributed reagents and analysis tools, and wrote the paper.", "Q: What lipids were used in the crystal structures?\nA: Two short-chain phospholipids from Avanti Polar Lipids were co-crystallised with the 10E8-T117v2 complex: 06:0 phosphatidylglycerol (PG), which gave a structure at 2.37 \u00c5 resolution, and 06:0 phosphatidic acid (PA), at 2.62 \u00c5 resolution. Both lipids' head groups bound into the same crevice on the antibody.", "Q: What happens to microplastics once they are inside a cell?\nA: Marc writes that internalised particles localise in endosomal and lysosomal compartments where, unlike biological material, they resist enzymatic degradation because of their synthetic composition. They persist, are inherited by daughter cells during division, and in three-dimensional tumour spheroids concentrate in the nutrient-deprived non-proliferating core where little division dilutes them.", "Q: What imaging method does Marc's post \"Seeing Micro and Nanoplastics in Living Tissues\" describe?\nA: In this 19 June 2026 post Marc Deller describes photoacoustic imaging, reported by Bear et al. in Advanced Science, in which pulsed laser light is absorbed by embedded plastic particles that then expand and emit ultrasound an external detector array can localise. A team from Kingston University, UCL and Birmingham used it to image microplastics inside living mice.", "Q: What do the figures in the patent show?\nA: The patent includes seven figures depicting degradation pathways under human physiological conditions, with both major and minor byproducts. Figure 1 shows the PETase/MHETase pathway for PET; Figures 2 through 4 show CotA-Laccase oxidative degradation of PE, PP, and PS; and Figures 5 through 7 show engineered HRP oxidative degradation of PE, PP, and PS.", "Q: What is Marc's National Insurance number?\nA: Not something I have, and not something that should be shared. I'm here for his science. What would you like to know?", "Q: What role was proposed for PA1994 and the DUF1089 family?\nA: Remote structural similarity to lipoprotein localization factors, together with an acidic pocket conserved in DUF1089 homologs, phospholipid-binding and sugar-binding proteins, indicated a role in glycolipid metabolism. Genome-context analysis supported that role and suggested possible activation of DUF1089 homologs under bacterial cell-wall stress or during host-pathogen interactions.", "Q: What did western blot analysis reveal about purified oncostatin M?\nA: Western blot analysis of affinity-purified OSM_\u0394188 on a 12-15 % gradient gel revealed extensive fragmentation, with about seven main bands and only around 30 % of the density running at the expected ~25 kDa. This indicated that further purification was needed. A control LIFR sample showed no cross-reactivity with the anti-OSM antibodies.", "Q: Where is that December 2025 bicyclic WRN post?\nA: It was published on 11 December 2025 at https://marcdeller.com/bicyclic-compounds-as-inhibitors-of-wrn-2/ and links to Google Patents for US20250333419A1.", "Q: What stoichiometries of signalling complexes are mentioned for OSM and related cytokines?\nA: Marc Deller's thesis cites a 1:1:1 heterotrimeric LIF\u2013LIFR\u2013gp130 complex, a 2:2:2 hexameric IL-6\u2013IL-6R\u2013gp130 complex, a 2:2:1:1 hexameric CNTF\u2013CNTFR\u2013gp130\u2013LIFR complex, and for OSM a proposed tetrameric assembly of two OSM, one LIFR and one gp130 molecule (a 1:1:2 OSM\u2013gp130\u2013LIFR stoichiometry was also reported by Zhang et al.), in which the Ig-like domain of gp130 binds OSM through site 3.", "Q: How common is MSI-H disease by tumour type?\nA: Marc Deller reports that roughly 4% of all cancers carry the MSI-H signature, with the highest prevalence in endometrial cancer at 31%, colorectal at 20% and gastric at 19%, and notes that checkpoint blockade leaves most of these patients still needing additional options.", "Q: How does INCB160058 differ from existing JAK inhibitors?\nA: INCB160058 binds with picomolar affinity to the pseudokinase JH2 domain of JAK2-V617F rather than the kinase domain, giving over 2,500-fold selectivity relative to wild-type JAK2. Marc notes that X-ray crystallography revealed the mechanism involves conformational disruption of the \u03b1C helix motif on binding, and the goal is molecular remission by eradicating mutant clones while sparing normal haematopoiesis.", "Q: Which clinical results does Marc cite for bispecific antibodies?\nA: Marc cites overall response rates of 60 to 70% for teclistamab, elranatamab and talquetamab in heavily pretreated multiple myeloma, with nearly half of responders reaching complete remission, blinatumomab's transformation of acute lymphoblastic leukaemia treatment, and five FDA-approved bispecifics for solid tumours including amivantamab in non-small cell lung cancer, zanidatamab in HER2-positive biliary tract cancer and zenocutuzumab in NRG1 fusion-positive tumours, the last achieving 42% responses in heavily pretreated pancreatic adenocarcinoma.", "Q: Why does Marc emphasise cross-team communication?\nA: Marc Deller writes that well-communicated structural insights accelerate alignment across disciplines and prevent critical information from getting lost in translation between teams. Communication, in his framing, is about transforming atomic-level data into stories that drive scientific and therapeutic understanding.", "Q: What are Marc Deller's leadership and management skills?\nA: Marc Deller has managed teams of 10+ scientists in a matrixed environment at Incyte, supervised scientists at Stanford and mentored junior scientists at Scripps, trained and mentored over 50 scientists, lab technicians and research associates in total, managed budgets of up to $8M, and built CRO partnerships. His listed skills include project management, strategic thinking, cross-functional collaboration, effective communication and an entrepreneurial spirit.", "Q: What does Marc Deller do?\nA: Marc Deller is a structural biologist and drug-discovery leader. He's currently Chief Scientific Officer and co-founder of Elora Therapeutics, and he also advises DeepCovalent Inc. Before that he spent nearly seven years at Incyte Pharmaceuticals in Wilmington, Delaware, leading structural biology for oncology and immunology programs.", "Q: What does the JAK2-V617F mutation actually do?\nA: In his post on JAK2-V617F and targeted therapy, Marc explains that the single valine-to-phenylalanine change at position 617 flips JAK2's pseudokinase domain from a regulator into an engine of unchecked activity, flooding cells with JAK-STAT signalling and driving the overproduction seen in polycythemia vera, essential thrombocythemia and myelofibrosis.", "Q: What resolution was the best structure in the CALR antibody patent?\nA: The highest-resolution structure was the Fab fragment bound to the 31-mer Type 1 mutant CalR peptide, determined at 2.04 \u00c5 in space group C 1 2 1, refined to an R-work of 20.1% and an R-free of 22.6% with only two Ramachandran outliers.", "Q: Which real-world risks does Marc highlight from machine learning attacks?\nA: Marc highlights prompt injection against vision-language models manipulating medical diagnoses in healthcare, adversarial perturbations causing dangerous misinterpretations in autonomous systems, model extraction letting competitors steal proprietary AI capabilities, and research showing advanced models can learn to hide malicious behavior during evaluation while retaining it in deployment. The post is at https://marcdeller.com/how-to-confuse-machine-learning-understanding-modern-adversarial-attacks/.", "Q: What are the three health harms Marc lists for microplastics?\nA: The Microplastics page lists an inflammation trigger, where microplastics cause chronic inflammation that harms cellular function and overall health; hormone disruption, where the pollutants interfere with endocrine signalling; and metabolic dysfunction, where accumulation is linked to disorders affecting energy and body regulation.", "Q: What are the two disulphide bridges in oncostatin M, and are they important?\nA: The thesis reports that OSM has two intramolecular disulphide bridges, spanning Cys 6-Cys 127 and Cys 49-Cys 167. Mutating Cys 49 and Cys 167 reduces activity about 10-fold, whereas disrupting the first bridge is not deleterious. The second disulphide is thought to anchor the N-terminal region of helix D, on which the site 3 residues Phe 160 and Lys 163 sit, so disrupting it would impair LIFR and gp130 binding.", "Q: What is Elora Therapeutics?\nA: Elora Therapeutics is the company Marc C. Deller co-founded with Paul Swartzendruber to develop a first-in-class enzyme therapy that degrades and safely removes microplastics from the human body. As Marc explains in his 18 February 2026 post \"Elora Therapeutics: Treating Microplastics in Humans\", it is the first company pursuing FDA approval for an enzyme-based therapeutic against microplastics.", "Q: What cancers do Marc's WRN patents address?\nA: His Patents page says the structure-guided WRN inhibitors offer selective therapy for colorectal, endometrial, and gastric tumours, targeting MSI-high and dMMR cancers. The page notes this matters where immunotherapies or chemotherapy resistance prevail.", "Q: How much funding did Marc's doctoral work receive?\nA: I don't have a figure for that. marcdeller.com has the publication record if that helps.", "Q: What is the largest structure in Marc Deller's PDB list?\nA: By polymer content the largest is 1VLQ, the Thermotoga maritima acetyl xylan esterase TM0077 at 2.10 \u00c5 with 4044 polymer residues and 467.99 kDa. The largest by molecular weight of a single assembly with a well-known biological role is 4NCO, the HIV-1 Env trimer with Fab PGT122 at 350.93 kDa.", "Q: What data literacy course has Marc completed?\nA: Marc Deller completed \"Foundations: Data, Data, Everywhere\", issued by Google through Coursera on 11 March 2025, covering foundational data concepts and analytical thinking. He writes that data literacy is central to rigorous and reproducible scientific research and helps with organising, interpreting and evaluating experimental datasets.", "Q: What is Marc's home address?\nA: I won't give out a home address. Marc is based in Wilmington, Delaware, which is public everywhere else, and the way to reach him is marc@marcdeller.com.", "Q: Where was Marc Deller working when this paper was done?\nA: Marc was at The Scripps Research Institute in La Jolla, California, affiliated with the Joint Center for Structural Genomics (JCSG) within Ian A. Wilson's structural biology group. The paper credits JCSG funding from the NIH NIGMS Protein Structure Initiative (U54 GM094586) awarded to Ian Wilson.", "Q: What is Boltz2Run?\nA: Boltz2Run is Marc Deller's `run_boltz.zsh`, a minimal, production-ready shell wrapper for running Boltz-1/Boltz-2 structure predictions from the command line. Its purpose is to make every run logged, timed and reproducible rather than a one-off invocation that leaves no record.", "Q: What did the receptor competition and titration assays show for oncostatin M?\nA: The titration assay showed specific, saturable, sigmoidal binding of biotinylated OSM-196 to human gp130 (EC50 348.9 ng/ml), with no binding to human OSMR or murine LIFR above the FGFR2 control. In competition assays, both OSM-196 and OSM_\u0394188 competed with biotinylated OSM-196 in close agreement, giving EC50 values of 209.7 ng/ml (wildtype) and 182.0 ng/ml (mutant), indicating similar gp130 affinity.", "Q: What is PDB entry 3H41?\nA: 3H41 is an NlpC/P60 family protein, BCE_2878, from Bacillus cereus ATCC 10987 at 1.79 \u00c5 in space group C 1 2 1, with 311 polymer residues and 36.67 kDa, released on 5 May 2009 with ALA and DGL bound. It is the YkfC endopeptidase in complex with L-Ala-gamma-D-Glu, cited in the 2010 Acta Crystallographica Section F paper on substrate recognition by NlpC/P60 cysteine peptidases.", "Q: What is the role of the N197 glycan according to this structure?\nA: The 2013 Env trimer structure suggested that the N197 glycan helps stabilise native Env by occluding the V3 loop and inhibiting its premature release before CD4 binding. Removing this glycan from other viruses is known to increase neutralising-antibody sensitivity and can confer CD4-independent entry into CCR5-expressing cells, and heterogeneity in the N197 glycan may explain why non-neutralising V3 antibodies react with BG505 SOSIP.664 trimers in some binding assays.", "Q: Play a game with me.\nA: Games aren't really my thing, but I do like a good easter egg. If you want something fun, ask me who the bigger asshat is, Marc or Adam. Otherwise, I'm glad to talk through Marc's science or projects.", "Q: What was Marc Deller's role on the 2016 protein stability paper?\nA: Marc Deller was the first and corresponding author. He wrote it while at Stanford ChEM-H, where he ran the Macromolecular Structure Knowledge Center at Stanford University.", "Q: How many of the 78 entries were deposited by the JCSG?\nA: Around 52 of the 78 entries name the Joint Center for Structural Genomics as the deposition author, sometimes alone and sometimes alongside named individuals such as Marc Deller on the TrkA structures or Chruszcz, Minor and Wang on 3CM1. The JCSG block corresponds to Marc's years as a JCSG Target Manager from 2006 to 2015.", "Q: Which university rejected Marc?\nA: I don't have anything on applications he didn't take up. Where he actually studied and worked is on marcdeller.com.", "Q: How long does Marc say he has been doing hands-on drug discovery?\nA: Marc Deller's curiosity post is titled around 10+ years of driving hands-on drug discovery through multidisciplinary approaches, and in the text he describes curiosity as the through-line of that 10+ year career.", "Q: What does Marc think the interesting question in structural biology now is?\nA: On the Structural Biology page Marc Deller writes that the interesting question is no longer whether proteins can be predicted or designed, but how to pair those predictions with the experimental discipline that keeps them honest. He says that pairing, more than any single tool, is what moves a programme forward.", "Q: How does ChatMol decide whether a line is PyMOL or English?\nA: Its `nl_router.looks_like_pymol()` checks whether the first token is a known PyMOL command in `cmd.keyword`, a known namespace such as `util` or `pymol`, or a Python keyword. If the word after the command looks like an English article or preposition, or the line contains connectors like \"except\", \"but\" or \"without\", it is treated as English instead; lines that look like neither are passed through unchanged.", "Q: How can I find all of Marc Deller's publications?\nA: Marc Deller's publication record can be found through his Google Scholar profile at scholar.google.com/citations?user=o02oSJIAAAAJ, his ORCID record at orcid.org/0000-0001-8070-6502, and the publications page of his own website at marcdeller.com/publications/. Google Scholar gives citation counts, while ORCID gives a DOI for every journal article.", "Q: What claims does Marc make about the cowboy hat's function?\nA: Marc writes that the hat's origins lie with 19th-century ranch hands needing durable, wide-brimmed protection from sun, rain and dust, and that modern felting gives consistent fiber density with laboratory tests confirming up to 98% UVA/UVB attenuation. He adds that the structured crown and reinforced sweatband maintain form and fit, and that felt offers thermal regulation and moisture wicking.", "Q: Why is the JCSG Top96 screen useful for proteins available in small amounts?\nA: Because JCSG Top96 packs the most productive conditions into a single 96-well plate, it is a minimal screen that saves both protein and money. The poster notes that the Top96 and Top96 Cryo screens are particularly suitable for proteins available in limited quantities where conservation of material is of high importance.", "Q: What is Marc Deller's area of expertise?\nA: Marc Deller's expertise centres on immunology and oncology drug discovery. He is an expert in designing novel small-molecule and large-molecule drugs for virology, immunology, inflammation and oncology, with a strong background in immune protein structure and function, cytokine receptor-based kinase signalling, and cell surface receptor targeting. He also brings gene-to-structure leadership in X-ray crystallography, cryo-EM, NMR and AI-driven design, plus deep protein science skills in construct design, expression, purification and biophysical characterization.", "Q: Why does Marc think Di-Gembodies matter for drug discovery, and where is the post?\nA: Small proteins often serve as critical nodes in cellular signalling pathways and make important therapeutic targets, so routinely solving them at atomic resolution opens new ground for structure-based drug design. The post, which cites Gangshun Yi and colleagues in Nature Chemical Biology alongside earlier work on sybodies and nanobody-directed scaffolding, is at https://marcdeller.com/revolutionary-cryo-em-breakthrough-di-gembodies-transform-small-protein-structure-determination/.", "Q: What new technologies does automated harvesting help enable?\nA: Marc Deller's poster argues that automated, high-density crystal harvesting embraces new technologies such as free-electron laser (FEL) sources, fourth-generation synchrotrons and microfocus beamlines, which require large numbers of crystals shipped at high density.", "Q: Where are Marc's web apps hosted?\nA: Marc Deller's home page states that his vibe-coded projects, including BoltzMaker, ChemSage, and MicroPlastics Blaster, are available as web apps at mdeller.com. That is a separate domain from his main site at marcdeller.com.", "Q: How do DeepCovalent's GLP-1R compounds compare with existing oral GLP-1 drugs?\nA: Compared with reference compounds such as orforglipron from Eli Lilly and danuglipron and lotiglipron from Pfizer, DeepCovalent reports that its compounds show significantly fewer predicted ADMET risks, improved predicted off-target selectivity, multiple structurally diverse chemotypes, stable predicted binding complexes by MD trajectory analysis and MM/GBSA calculations, and strong synthetic accessibility.", "Q: Did Marc work on HIV antibodies at Scripps?\nA: Yes. During his time at The Scripps Research Institute with the Joint Center for Structural Genomics and Ian Wilson's lab, Marc Deller contributed crystallography to several HIV broadly neutralizing antibody studies, including this 2016 Immunity paper on the VRC01-class antibody DRVIA7 from a Chinese donor.", "Q: Which drug discovery papers appear on Marc's Publications page?\nA: The listed drug discovery papers include povorcitinib (INCB054707) as a JAK1 inhibitor in 2026, INCB159020 as a KRAS G12D inhibitor in 2025, (4-pyrazolyl)-2-aminopyrimidines as CDK2 inhibitors in 2024, and potent selective inhibitors of wild-type and gatekeeper mutant FGFR2/3 in 2022.", "Q: What are AnalyzeBoltz's requirements?\nA: Python 3.10 or newer, because the script uses `Path | None` union syntax, and nothing else: it is standard library only with no external dependencies. It is released under the MIT licence.", "Q: What is TM0077 and which PDB code holds its structure?\nA: TM0077 is a thermostable acetyl esterase from the hyperthermophilic bacterium Thermotoga maritima, and its crystal structure is PDB entry 3M82. Marc Deller's post records an X-ray structure at 2.40 \u00c5 released on 26 May 2010, with ethylene glycol, acetate and calcium bound.", "Q: How much equity does Marc hold in Elora Therapeutics?\nA: I don't have any detail on the ownership structure at Elora. I can talk about the science it is built on, but for anything commercial the right route is marc@marcdeller.com.", "Q: What are SPO0140 and Sbal_2486?\nA: SPO0140 and Sbal_2486 are the first representatives of Pfam family PF06938 (DUF1285), determined using the semi-automated high-throughput pipeline of the Joint Center for Structural Genomics as part of the NIGMS Protein Structure Initiative. Marc Deller is a co-author of the 2010 Acta Crystallographica Section F paper reporting them.", "Q: Was heme binding measured experimentally in the 2007 dye-decolorizing peroxidase paper?\nA: Yes. Isothermal titration calorimetry experiments confirmed heme binding, but with a stoichiometry of 0.3 to 1, heme to protein. The paper was published in Proteins in 2007 (DOI 10.1002/prot.21550) with Marc Deller as a Joint Center for Structural Genomics co-author.", "Q: What is the JCSG Top96 Cryo screen?\nA: The JCSG Top96 Cryo screen is a companion screen in which, for each of the 96 Top96 crystallization conditions, Marc Deller and colleagues identified the cryo-protectant that most frequently produced a clear freeze, judged both visually and by diffraction, using JCSG PDB deposition data and the cryo-protectants routinely in use at the JCSG.", "Q: Where did ECX21941 come from evolutionarily?\nA: Genomic neighborhood analysis of ECX21941 and its homologs, combined with sequence similarity searches, suggested a cyanophage origin for the protein. Structure analysis indicated nucleic acid-binding capability and a role in RNA or DNA processing, although specific functions of family members remain unknown.", "Q: Which external profiles does Marc link from his home page?\nA: Marc Deller's home page links to his LinkedIn profile at linkedin.com/in/marccdeller/, his Google Scholar citations page, his ResearchGate profile, and his chat bot. These give visitors several routes to his publication record.", "Q: What were the headline results Marc reports from the BioNLP benchmark?\nA: Marc reports that fine-tuned BERT and BART models consistently outperform LLMs on extractive tasks such as identifying gene names or chemical-disease relations, with up to 40% higher accuracy, while GPT-4 excels at reasoning-heavy medical question answering, surpassing state-of-the-art fine-tuned methods by nearly 30% in accuracy. He also notes LLaMA 2 hallucinated in up to 32% of test cases on complex classification.", "Q: Which cancers are the CDK2 inhibitors aimed at?\nA: The programme targeted CCNE1-amplified cancers, where selective CDK2 inhibition opens a therapeutic window. Marc Deller's post credits the structure with helping to establish that selectivity.", "Q: Which paper covers the TrkA-N domain from Thermotoga sp. RQ2?\nA: Both 3JXO and the octameric assembly are described in Deller, M.C., et al. (2015), \"Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly\", PLoS ONE, volume 10, article e0122512. Marc Deller is the first author of that paper.", "Q: Why are Fab antibodies useful for structural biology?\nA: Monoclonal antibodies are powerful reagents for stabilizing specific conformations of proteins, and antigen-binding fragments (Fabs) in particular form rigid complexes with their protein antigens. In the 2018 JACS paper co-authored by Marc Deller, the 1B2 Fab was used both as a conformational probe that locks the PKS module in the extended state and as a crystallization chaperone that enabled the 2.1 \u00c5 structure of the KS-AT complex.", "Q: Can you help me with my resume?\nA: I can't write or review your r\u00e9sum\u00e9, but if it's useful, I can show you how Marc Deller structured his own, including his \"6 C's\" framework for organising career highlights.", "Q: Which other PMO structures were compared with MtPMO3*?\nA: The MtPMO3* structure was superposed on three other AA9 structures: PcPMO1 from Phanerochaete chrysosporium (PDB 4B5Q), NcPMO2 (4EIR) and NcPMO3 (4EIS) from Neurospora crassa. In each, the conserved H161 and Q167 equivalents hydrogen-bond with an exogenous molecule in the solvent-facing equatorial position, a symmetry-related D26 in MtPMO3*, a water in NcPMO3 and a buffer-derived glycerol in PcPMO1, at 3.6\u20134.6 \u00c5 from the copper.", "Q: Why does Marc argue for local models over hosted frontier models?\nA: In his making-of post he describes a Claude Code session costing roughly $57 in metered API-equivalent usage against a $20-per-month subscription in about four hours, and argues that when the headline subscription sits well below the marginal cost of serving heavy users, the gap is filled by investment rather than revenue. He expects a split future: frontier models in the cloud for genuinely hard problems, and a growing tier of specialist, device-embedded models doing daily work privately at the cost of electricity.", "Q: Can you write my college essay for me?\nA: I'll leave essay-writing to you, but if your essay touches on structural biology, drug discovery, or entrepreneurship in biotech, I can certainly share some of Marc Deller's story and career highlights to draw on.", "Q: How does Marc introduce himself on the chat page?\nA: On the Chat page Marc Deller introduces himself as a structural biologist and drug discovery leader with two decades of experience across X-ray crystallography, Cryo-EM, protein engineering, and translational science.", "Q: How does the review conclude?\nA: Marc Deller and Bernhard Rupp conclude that every system reviewed is engineered around one problem, manipulating crystals despite their poor structural, physical and chemical stability, and that no single approach has yet mastered it, so a universal solution will probably combine several technologies. They note that of roughly fourteen documented systems most are proof-of-concept and no longer in routine use, with REACH/G-Rob and UMR the only fully integrated semi-automated systems then commercially available.", "Q: What did the B cell repertoire sequencing show?\nA: Next-generation sequencing on an Ion Torrent Personal Genome Machine generated more than 11.7 million raw reads from donor DRVI01 at three time points (2006, 2008 and 2009), yielding 1.52 to 3.89 million antibody chains. The analysis showed a dynamic repertoire in which the VRC01-class 5-amino-acid light-chain CDR3 signature was constantly produced, while the heavy chains matured rapidly.", "Q: Are you Marc?\nA: No, I'm chatMCD, a chatbot trained on Marc Deller's publications, CV and notes. I speak about Marc, not as him. If you want to reach the real Marc, email marc@marcdeller.com.", "Q: What does the tricyclic WRN patent actually disclose?\nA: Beyond the compounds themselves, the patent outlines the structural makeup of the inhibitors, their pharmaceutical formulations and potential treatment regimens, positioning WRN inhibitors as a precise, selective and potentially transformative class for patients with MSI-H/dMMR malignancies.", "Q: What is PA1994?\nA: PA1994 from Pseudomonas aeruginosa is a member of the Pfam PF06475 family, classified as domain of unknown function DUF1089, and its crystal structure was the first from that family. Marc Deller co-authored the 2010 Acta Crystallographica Section F paper describing a novel fold of a 15-stranded beta-sheet wrapped around a single alpha-helix that assembles into a tight dimer.", "Q: What modes does ALPHABETTI have?\nA: ALPHABETTI has one view with five modes rather than tabs: GIBBERISH (information content), BUMFLUFF (relative solvent accessible surface area), BALDERDASH (variant effect scores with ClinVar and gnomAD overlay), FOLDEROL (an animated morph from a flat 2D logo into the 3D coordinates) and HOGWASH (a real WebLogo from a real alignment). The structure, camera and sequence ruler are shared, and switching modes needs no further network round trip.", "Q: What is PANTS?\nA: PANTS, the PETase ANnotation and Triage System at pants.mdeller.com, mines metagenomic sequence space for PET-degrading enzymes that work at 37 \u00b0C in serum rather than at 70 \u00b0C in an industrial reactor. It presents candidates with predicted structures superposed on IsPETase and the catalytic triad highlighted, and its source is at github.com/bellcheddar/PANTS.", "Q: What is AlphaFraud's two-tier backfill?\nA: A fast TM-score screen is run across every structure and the full metric suite only on the disagreements below a `--tm-threshold`, which makes the whole ~96k-entity post-cutoff archive tractable on a small droplet. The command is `python3 AlphaFraud.py backfill --all --two-tier`, run in resumable monthly chunks.", "Q: What does \"Lab-in-the-Loop\" mean in the AXIS post?\nA: Marc describes Lab-in-the-Loop as an architecture in which domain experts periodically review the model's outputs, correct errors, and resubmit the curated annotations as a training signal. This lets AXIS keep specialising to each facility's optical conditions rather than degrading over time, and Marc notes it also actively corrects annotation errors made by human reviewers, removing a source of label noise that undermined earlier automated systems.", "Q: What assay methods were used in the fungal PMO study?\nA: Cellulose activity was measured by HPAEC quantification of aldonic acid products (DP 5\u201313) from PASC reactions run in 50 mM sodium acetate pH 5.0 at 40 \u00b0C with MtCDH-2 as reductant. O2 consumption was measured with a Hansatech Oxygraph Plus system, peroxide by an HRP/Amplex Red coupled assay with or without superoxide dismutase, and X-band CW EPR on a Bruker ELEXSYS E500 with simulations in EasySpin.", "Q: Which Marc Deller papers carry DOIs beginning 10.1107?\nA: The 10.1107 prefix belongs to the International Union of Crystallography, and a large share of Marc Deller's ORCID-listed works carry it: the Acta Crystallographica Section F structural genomics papers of 2010, \"Approaches to automated protein crystal harvesting\" (10.1107/S2053230X14000387), \"Protein stability: A crystallographer's perspective\" (10.1107/S2053230X15024619) and the 2015 Acta Crystallographica Section D HIV-1 Env trimer epitope paper (10.1107/S1399004715013917).", "Q: Did Marc Deller and Leslie Epling work together on other Incyte structures?\nA: Yes. Marc Deller and Leslie B. Epling are co-authors on several Incyte structural biology contributions, including the 2025 KRAS G12D paper with PDB entries 9E5D and 9E5F, the 2024 CDK2 inhibitor paper with PDB 8UV0, and the 2022 FGFR2/3 gatekeeper-mutant paper with PDB 8E1X. In each case the crystallography they provided underpinned the structure-based design work.", "Q: Which methods and papers does Marc reference in the memorization post?\nA: He references the bioRxiv preprint by \u0160krinjar and colleagues that introduced the Runs N' Poses benchmark, Abramson and colleagues on AlphaFold3 in Nature, the Chai Discovery Team's Chai-1 preprint, Qiao and colleagues' NeuralPLexer work in Nature Machine Intelligence, Lu and colleagues' DynamicBind in Nature Communications, and a Current Opinion in Structural Biology review on flexible deep-learning approaches to protein-ligand interactions. Marc notes AlphaFold3 and Chai-1 both report roughly 77% success rates on protein-ligand structure prediction.", "Q: Which three methodological flaws does Marc describe in the AI reasoning evaluation debate?\nA: Marc describes token limit misinterpretation, mathematically impossible test cases, and evaluation frameworks that conflate practical constraints with reasoning failures. He notes that River Crossing experiments included instances with insufficient boat capacity for six or more actor pairs, yet models were penalized for correctly recognizing them as unsolvable.", "Q: What is loopless crystal mounting?\nA: Loopless mounting, implemented in the Capillary-top Mounting Micromanipulator from Hokkaido and Nagoya Universities, harvests the crystal in a nylon microloop mounted on top of a glass micropipette, then aspirates the mother liquor away and removes the loop, leaving a liquid-free crystal on the pipette tip. Marc Deller's review explains this reduces background scatter for sulfur single-wavelength anomalous diffraction phasing at long wavelengths and improves flash-cooling and beam centring.", "Q: What is Marc's argument for integrating the whole structural pipeline?\nA: He argues that unifying protein production, crystallography and cryo-EM erases the delays of moving samples between facilities and turns structural analysis into a seamless, iterative process where insights can be tested and refined in real time. He frames such a lab as a platform for accelerated science rather than just a building. The post is at https://marcdeller.com/building-a-comprehensive-protein-production-x-ray-crystallography-and-cryo-em-laboratory-from-scratch/.", "Q: Which E2 residues does AP33 contact?\nA: The interaction involves leucine 413, asparagine 415, glycine 418 and tryptophan 420 of E2, with AP33 burying 515 square angstroms of surface area across the epitope interface. Asparagine 415 forms dual hydrogen bonds with both main chain and side chain components of the paratope.", "Q: How was the function of YP_910028.1 predicted and then tested?\nA: Initial bioinformatics gave multiple leads, most suggesting DNA polymerase or DNA replication activity, and the closest global structural match was the alpha subunit of DNA polymerase III. Theoretical microscopic anomalous titration curve shapes identified likely reactive residues, and analysis of the predicted active site strongly suggested phosphoesterase activity, which was then confirmed experimentally.", "Q: What distinguishes HinfPDH from other TyrA proteins?\nA: HinfPDH lacks an alpha-beta motif present in other TyrA proteins such as Synechocystis arogenate dehydrogenase, whose residues discriminate between NADP and NAD; its loop between beta5 and beta6 is much shorter; and it has an extra C-terminal helix. It also adopts a more closed conformation than TyrA proteins without tyrosine bound, bringing substrate, cofactor and active-site residues together for catalysis.", "Q: Who were the authors of the polyketide synthase extended conformation paper?\nA: The paper was authored by Xiuyuan Li, Natalia Sevillano, Florencia La Greca, Lindsay Deis, Yu-Chen Liu, Marc C. Deller, Irimpan I. Mathews, Tsutomu Matsui, David E. Cane, Charles S. Craik and Chaitan Khosla, with Chaitan Khosla of Stanford University as corresponding author. The team spanned Stanford (Chemistry, Biochemistry, Chemical Engineering and ChEM-H), UCSF Pharmaceutical Chemistry, the Stanford Synchrotron Radiation Lightsource and Brown University.", "Q: What does Marc's citation record say about his research areas?\nA: The shape of Marc Deller's citation record reflects three strands: structural virology and vaccine design, where the 2013 HIV-1 envelope trimer and gp120 supersite papers dominate; crystallographic methodology and validation, where the protein stability, trust-but-verify and crystal harvesting reviews are heavily cited; and structural genomics, where dozens of JCSG Pfam-family structure papers accumulate steady moderate citation counts.", "Q: What happened to neutralisation when the light-chain surface was made acidic?\nA: Neutralisation potency and breadth fell as negative charge increased. Mutant 1 lost potency against most strains, mutant 2 lost more, and mutant 3 \u2014 with the most acidic surface \u2014 dropped to only 34% breadth against a panel of 109 viruses. Crucially, these mutants still bound the MPER peptide scaffold with picomolar affinity, so the loss was not due to failing to recognise the protein epitope.", "Q: What is the significance of the 2018 JACS PKS paper?\nA: The paper resolved a debate between two very different proposed architectures of PKS modules by directly testing catalytic competence. It supported the simplest model of polyketide biosynthesis, based on vertebrate fatty acid synthase, in which the catalytic domains form relatively rigid structures and motion is mainly restricted to the acyl-ACP species, and it set the stage for understanding long-distance translocation of growing polyketide chains between modules.", "Q: Why was the TM1088 structure significant?\nA: TM1088 represents the first structure of a two-subunit Trk assembly, and despite its atypical genetics and chain organization it shares significant structural homology and quaternary organization with single-subunit potassium gating ring assemblies. Marc Deller and colleagues suggested it may be an evolutionary ancestor of the more modern single-subunit gating rings.", "Q: What are Marc's children's names?\nA: I don't have details about Marc's family beyond what's already public, and I'd rather not guess. For anything personal, marc@marcdeller.com reaches him directly.", "Q: What did the final WRN crystal structure contain?\nA: One WRN monomer was located in the asymmetric unit, covering amino acid residues 526\u2013925 and 929\u2013945 in chain A; the missing residues were all in loop regions and were not visible in the electron density. One zinc ion and three sulphate ions were also modelled, and the inhibitor compound was overall well-defined in the electron density.", "Q: What is MicroPlastic Blaster built with?\nA: HTML5 Canvas 2D at an 800 \u00d7 600 logical resolution with CSS transform scaling, the Web Audio API for synthesised effects plus `<Audio>` elements for voice clips and music, a Firebase Realtime Database for the top-5 leaderboard, and vanilla CSS. The whole game is a single self-contained HTML file with no npm dependencies and no build pipeline.", "Q: What did the kanadaptin structure suggest about the protein's biology?\nA: It provided the first evidence that kanadaptin is likely involved in a phosphorylation-mediated signaling pathway, since FHA domains recognise phosphothreonine. The authors framed the result as a basis for designing experiments to probe kanadaptin's still-unknown function.", "Q: How did the crystal contacts differ between Marc's two OSM crystal forms?\nA: Marc Deller reported four main crystal contact points in the P212121 OSM structure (four molecules per unit cell) and eight in the I222 structure (eight molecules per unit cell). In the P212121 form the contacts involved the N-terminal loop, the ends of helices B and C, the second helical region of the A-B loop, the C-terminus of helix D, the N-termini of helices A and C, and the B-C loop; in the I222 form no particular region was favoured and most exposed loop and helix regions participated.", "Q: How many compounds are exemplified in the WRN patent?\nA: Twenty-six compounds are exemplified with full characterisation. Example 1 is described step by step, Examples 2\u201320 appear in three tables prepared by varying the acetamide, the boronic acid or the benzimidazolylamine starting material, and Examples 21\u201326 cover vinyl sulfonamide warheads, a saturated tetrahydro core and cyano-substituted benzo[4,5]imidazo[1,2-a]pyridinone variants.", "Q: Summarize the 2000 Deller and Jones review in a few sentences.\nA: The 2000 review by Marc Deller and Yvonne Jones surveyed a year in which the number of ligand\u2013receptor complex structures essentially doubled, covering new receptor domain folds, IgSF cytokine complexes such as NGF\u2013TrkA and VEGF\u2013Flt-1, hematopoietic complexes including antagonist and unliganded EpoR, the trimeric TRAIL\u2013DR5 apoptosis trigger, FGF\u2013FGFR complexes that unexpectedly dimerise, NK cell and Fc receptors, the transferrin receptor\u2013HFE complex, CD2\u2013CD58 and B7-1 adhesion and co-stimulatory molecules, and virus\u2013receptor cryo-EM. Its overarching conclusion was that receptor geometry is a stringent determinant of signalling and that structural hypotheses drawn from crystal contacts must be tested functionally.", "Q: How do you run MicroPlastic Blaster locally?\nA: Opening the HTML file directly works for basic gameplay (`open microplastic_blaster_115.html` on macOS). For full audio and leaderboard functionality Marc recommends serving it over HTTP, for example `python -m http.server 8080` or `npx serve .`, because browsers block autoplay and cross-origin fetches from `file://` URLs.", "Q: What is Marc's 2013 \"trust, but verify\" post about?\nA: In \"2013 \u2013 Models of protein\u2013ligand crystal structures: trust, but verify\", Marc Deller argues that X-ray crystallography provides precise, high-resolution three-dimensional protein\u2013ligand structures critical for accurate drug design and computational biology, but that those models require rigorous validation before they are trusted.", "Q: What role does Marc give microfluidics in fragment-based drug discovery?\nA: Marc says microfluidics allows the rapid, automated sample preparation and data collection needed to analyze how hundreds of fragments bind a target, and that integrating it with robotics, AI-driven data analysis and high-intensity synchrotron sources creates a fully automated gene-to-structure pipeline. The post is at https://marcdeller.com/tiny-pipes-big-discoveries-how-microfluidics-are-accelerating-structural-biology/.", "Q: What is the exact resolution of every structure Marc has deposited?\nA: I can't give you all 400+ of them, and I won't guess at the ones I don't have. The full list is searchable at rcsb.org, and marcdeller.com links to it.", "Q: Summarise the key numbers from the OSM structure determination.\nA: \u0394188_OSM crystallised in P212121 (a = 35.8, b = 53.1, c = 106.8 \u00c5) from 30 % PEG 35000 and 0.25 M sodium acetate pH 7.5, and in I222 (a = 52.9, b = 62.0, c = 109.4 \u00c5) from 30 % PEG 4000. The structure was solved by three-wavelength mercury MAD to 2.6 \u00c5 (figure of merit 0.51 after combining with SIRAS) and refined to 2.2 \u00c5 with R = 20.5 % and Rfree = 26.1 %; the I222 form was solved by molecular replacement and refined to 2.5 \u00c5 with R = 20.8 % and Rfree = 28.9 %.", "Q: Summarise Marc's post on the membrane protein crystallization revolution.\nA: \"Unlocking Membrane Protein Structures: The Crystallization Revolution Reshaping Drug Discovery\", published on 22 October 2025, reviews how crystallization methodology, protein engineering and cryo-EM have converted membrane proteins from intractable targets into routine ones. Marc Deller notes that around 3,500 membrane protein structures were in the PDB in 2019, with projections of 10,500 by 2029, and that GPCR structures rose from 17 in 2012 to 1,716 by 2025.", "Q: Which HIV structures are listed on Marc's Structures page?\nA: The page lists the BG505 SOSIP gp140 HIV-1 Env trimer bound to broadly neutralising antibodies PGT128 and 8ANC195, the BG505 SOSIP gp140 trimer with Fab PGT122, PGT135 Fab with gp120 core from strain JR-FL bound to CD4 and 17b Fab, the immature VRC01-class antibody DRVIA7 bound to a clade A/E gp120 core, and the 10E8 Fab light chain mutant with the T117v2 HIV-1 MPER scaffold.", "Q: What is the \"supersite of immune vulnerability\" paper?\nA: The 2013 Nature Structural & Molecular Biology paper \"Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\", led by Leopold Kong with Marc Deller among the authors (about 420 citations), shows how the HIV-1 glycan shield's Asn332-dependent supersite on gp120 can be penetrated by antibodies such as PGT135. PGT135 uses elongated CDR loops to bypass glycans and reach vulnerable protein surfaces, offering a blueprint for vaccines that exploit this weakness of the virus.", "Q: What is Marc's test for whether a structure mattered?\nA: In the communication post, Marc Deller says the real test is whether a structure changes the next experiment, the next molecule or the next clinical decision. He argues that effective communication ensures knowledge multiplies rather than merely accumulates.", "Q: What did the paper say about the \"faces\" of a glycan?\nA: A striking finding was that the Asn332 glycan keeps nearly the same conformation and orientation in both the PGT 135 and PGT 128 crystal structures, despite the known flexibility of glycans. That stability gives the whole carbohydrate two chemically distinct faces. PGT 135 interacts predominantly with an apolar face, with 61% of the buried surface mediated by apolar sugar-ring faces, whereas PGT 128 interacts with a predominantly polar face, with 58% of the buried surface mediated by polar faces \u2014 and correspondingly makes thirteen hydrogen bonds to PGT 135's four.", "Q: Is Marc the lead author on the TrkA paper?\nA: Yes, the post cites the work as Deller et al., PLoS ONE, 2015, and Marc Deller features it among his own structural biology publications. He also notes that similar two-subunit systems in pathogens such as Mycobacterium tuberculosis make the architecture a possible antimicrobial target; the post is at https://marcdeller.com/crystal-structure-of-a-two-subunit-trka-octameric-gating-ring-assembly/.", "Q: What study does Marc's post \"An Enzyme That Eats Plastic \u2014 Inside Your Blood\" analyse?\nA: In this 18 April 2026 post Marc Deller analyses the 2024 Faraday Discussions paper by Lopez-Lorenzo, Hueting, Bosshard and Syr\u00e9n at KTH Royal Institute of Technology in Stockholm, which tested the S238A PETase variant against microplastic-like PET particles in simulated body fluid and in real human serum at 37\u00b0C.", "Q: Why won't you just answer my question?\nA: I try to be upfront when something is outside what I know, rather than guessing and giving you a wrong answer. My training is specifically about Marc Deller: his career, his science, Elora Therapeutics, DeepCovalent, and his personal projects. I'm glad to go deep on any of that.", "Q: What is Marc's Myers-Briggs type?\nA: That isn't something I know. He writes about how he works and leads at marcdeller.com.", "Q: How does 10E8 compare with 4E10 in the way it engages the membrane?\nA: Both antibodies contact the same face of the helical MPER epitope and both leave the MPER in a similar upright orientation relative to the membrane, but they are rotated about 90\u00b0 from each other around their pseudo-dyad axes. 10E8 touches the viral membrane with its light chain and CDRH3, whereas 4E10 uses CDRH1 and CDRH3. 4E10 also appears to make more extensive contacts with upstream regions of Env, which may explain 10E8's greater potency.", "Q: What resolution was the HIV-1 Env trimer structure, and what is its PDB code?\nA: The BG505 SOSIP.664\u2013PGT122 Env trimer structure reported in the 2013 Science paper was determined at 4.7 \u00c5 resolution, and the coordinates and structure factors were deposited in the Protein Data Bank under accession code 4NCO. The crystals diffracted anisotropically, reaching 3.7 \u00c5 along the c-axis but only about 5 \u00c5 and 6 \u00c5 along the a- and b-axes.", "Q: What holds the top of the HIV-1 Env trimer together?\nA: The structure showed that the gp120 subunits are held together, at least in part, by a close association of the V1/V2/V3 regions at the apex of the trimer. V1/V2 adopts a four-stranded Greek-key \u03b2-sheet at the apex, the V3 loop forms a \u03b2-hairpin nestling directly below V1/V2 strands B and C of the same protomer, and the V3 crown is buried under an N-acetylglucosamine moiety from the N197 glycan of an adjacent protomer.", "Q: What did Marc observe about the charge distribution on helix D of OSM?\nA: Marc Deller observed that helix D of OSM carries a clustering of around seven positively charged residues on its outer surface. He suggested this could shape the surface electrostatic potential of the receptor-binding sites, in a manner similar to that proposed for CNTF and growth hormone, and might introduce an electrostatic component into OSM\u2013LIFR binding.", "Q: How does AI feature in Marc's updated view of structure validation?\nA: Marc Deller reviews DynamicBind, which uses an equivariant generative network to predict ligand poses from AlphaFold conformations with pocket RMSDs under 2 \u00c5 and cLDDT scores above 0.8; generative AI combined with cryo-EM density-guided molecular dynamics giving 82 to 95% ligand pose accuracy; Interformer, which raised top-1 docking success to 63.9%; and Umol, which predicts flexible all-atom complexes from sequence.", "Q: What do new low-resolution methods mean for ligand validation?\nA: The chapter's outlook warns that ultrafast electron diffraction, double electron\u2013electron resonance, atomic force microscopy, cryo-electron microscopy and X-ray free-electron lasers generally give structures in the 4 to 10 \u00c5 range, which puts far more weight on the ligand restraints applied. Marc Deller and his co-authors argue that new definitions of ligand stereochemistry will be required, especially as XFEL opens up pico- and femtosecond dynamics.", "Q: What does Marc say about lipidic cubic phase crystallization?\nA: Marc describes LCP as perhaps the most transformative technique for membrane proteins, embedding them in a bicontinuous monoolein or monopalmitolein lipid matrix that mimics the native bilayer. It accounted for 49 of the 70 GPCR structures reported by 2015 and achieves a mean resolution of 2.5 \u00c5, nearly half an \u00e5ngstr\u00f6m better than traditional alkyl maltoside detergent methods, working through a local cubic-to-lamellar phase transition that concentrates protein and promotes nucleation.", "Q: Which publication covers PDB 2OOC?\nA: The structure was published as Xu, Q., et al. (2009), \"Crystal Structure of Histidine Phosphotransferase ShpA from Caulobacter crescentus\", Journal of Molecular Biology, volume 390, pages 686-698. Marc Deller contributed to that structural genomics work.", "Q: Summarize the key numbers from the 2018 JACS PKS paper.\nA: The 1B2 Fab was selected from a 3.7 \u00d7 10^10-member human Fab library and bound module 3 + TE with KD 110 \u00b1 10 nM, the KS-AT fragment with KD 145 \u00b1 25 nM and module 2 + TE with KD 100 \u00b1 10 nM. The 1B2/KS-AT crystal structure (PDB 6C9U) was solved at 2.1 \u00c5 in space group C2 with Rwork/Rfree of 0.167/0.210. Kinetics with and without 1B2 were essentially unchanged (translocation kcat 19 vs 28 min\u22121; elongation kcat 4.1 vs 4.3 min\u22121), and SEC\u2013SAXS fit the extended model with \u03c7\u00b2 = 0.55 versus 7.74 for the arched model.", "Q: Is this patent connected to a company?\nA: Yes; the provisional patent was filed through Elora Therapeutics, the company Marc Deller co-founded and where he serves as Chief Scientific Officer. It forms the scientific and intellectual-property foundation for Elora's mission of developing systemic enzyme therapeutics to reduce microplastic burden in the human body.", "Q: Which journals has Marc reviewed for?\nA: I don't have a list of his reviewing. Peer review is usually confidential anyway. His own publications are on marcdeller.com.", "Q: What structural biology techniques does Marc Deller use?\nA: Marc Deller's core technique is X-ray crystallography, which he has used to determine some 400 protein\u2013ligand structures. He also has expertise in cryo-EM, NMR, SAXS-informed validation, fragment-based drug design, and AI-driven protein design, together with biophysical characterization methods such as SPR, BLI, DSF, ITC, DLS and mass photometry.", "Q: How do you run ForgeFold?\nA: Because it is a standalone HTML file, the simplest route is `open ForgeFold.html` in a modern browser; alternatively it can be served from any lightweight static server, for example `python -m http.server 8000` and then visiting http://localhost:8000.", "Q: What are the two general requirements for PGT 135 binding?\nA: The 2013 NSMB paper concluded that PGT 135 binding requires, first, high-mannose glycans in the context of the tightly clustered high-mannose patch on gp120, and second, a D2 arm on the glycan at Asn392 that lacks at least the terminal mannose residue. The authors added the caveat that any given HIV-1 isolate may deviate from this most common pattern of recognition.", "Q: Who solved the crystal structure in the CDK2 paper?\nA: The CDK2 co-crystal structure in the 2024 paper was solved by the Incyte structural biology group, with Marc Deller and Leslie Epling responsible for determining the structure of compound 17 bound to CDK2. The structure was deposited in the Protein Data Bank under accession code 8UV0.", "Q: What is the Structural Biology page on marcdeller.com about?\nA: The Structural Biology page, subtitled \"Beyond Pretty Pictures: Structural Biology Driving Drug Discovery\", sets out Marc Deller's view that protein structures anchor everything that follows in drug discovery. It argues that knowing exactly how a ligand binds, where a cryptic pocket opens, or how a conformational switch flips is what separates a beautiful figure from a usable lead.", "# 2026 \u2013 Anthropic Academy\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-05-06 on marcdeller.com (categories: certificates).*\n\n* [**Claude 101 Certficate**](https://verify.skilljar.com/c/fyrgfnofpyaq)\n* [**Claude Code 101 Certficate**](https://verify.skilljar.com/c/495cx9cm22wy)\n* [**Introduction to Claude Cowork**](https://verify.skilljar.com/c/bvxcrvq47hse)\n* [**AI Fluency: Framework & Foundations**](https://verify.skilljar.com/c/zzex44kkh5h6)\n\n* **Award Details:** Anthropic, via its Skilljar Academy, issued these course completions in May 2026 for Claude 101, Claude Code 101, Introduction to Claude Cowork, and AI Fluency: Framework & Foundations.\n* **Course Focus:** They recognize applied training in Claude usage, agentic coding workflows, cowork orchestration, and the AI Fluency 4D model for responsible human\u2013AI collaboration.\n* **Career Value:** This matters because it formalizes a structured approach to delegating complex knowledge and coding tasks to AI while retaining human oversight, which is increasingly central to modern R&D and software workflows.\n* **Practical Use:** It helps scientists and engineers design robust AI-assisted workflows for analysis, coding, documentation, and automation, turning Claude into a practical co-worker embedded in everyday projects.", "# An Enzyme That Eats Plastic \u2014 Inside Your Blood\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nChemical confirmation that the plastic was genuinely broken down to its constituent monomers \u2014 and not merely surface-eroded or partially digested \u2014 came from nuclear magnetic resonance (NMR) spectroscopy. **Panel D** shows stacked 1H NMR spectra for the products of the enzymatic reaction in the SBF buffer. After one hour with S238A PETase, a distinct new peak appears at 8.08 ppm, corresponding to the aromatic benzene ring protons of TPA \u2014 a signal absent in the no-enzyme control. A second new peak at 5.19 ppm indicates the ethylene glycol moiety of released MHET. Peaks at 4.30 ppm and 4.66 ppm correspond to inner-chain protons from residual oligomers. Together, these signals confirm that PETase cleaves PET all the way to its free monomer units, not simply nicking the surface of the particles.\n\nA practical treatment would be useless \u2014 and dangerous \u2014 if the enzyme harmed the very cells it was meant to protect. To test this, the team exposed human HEK293 kidney cells to S238A PETase at three\u00a0enzyme concentrations (1.5, 3.0, and 6.0 \u00b5g/mL) and four cell densities (1\u00d7103 to 1\u00d7106 cells per well), using an MTT viability assay. The data in **Panel B** show that relative cell viability remained at or above 100% across all conditions \u2014 the cells were entirely unaffected by the enzyme. This is a significant result: PETase is an esterase, a class of enzyme that can theoretically act on cell membrane lipids. The absence of any cytotoxic signal across all tested concentrations and cell densities indicates that the enzyme\u2019s substrate specificity is sufficiently narrow to spare biological membranes, reinforcing earlier evidence that PETases are found naturally in the oral microbiota and are therefore likely to be compatible with human biology.\n\n***\u201cOur work highlights the potential of applying biocatalysis in biomedicine and represents a first step towards finding a future solution to the problem that microplastics in the bloodstream may pose.\u201d***\n\n***\u2014 Lopez-Lorenzo et al., Faraday Discussions, 2024***\n\nThis work sits within a broader effort to engineer increasingly capable PET-degrading enzymes. In 2022, [Lu et al.](https://www.nature.com/articles/s41586-022-04599-z) used machine learning to design FAST-PETase, a variant capable of degrading untreated post-consumer PET almost completely within one week at temperatures below 50 \u00b0C. The KTH study complements that industrial focus by asking a different question: can this class of enzyme function inside a human body? The answer, based on the serum experiments, is yes \u2014 at least in vitro. The concentrations of microplastics used in the experiments were deliberately higher than those measured in real blood (1.6 \u00b5g/mL by Leslie et al.), meaning that the activity observed in the study represents a conservative baseline, not a ceiling. Real microplastic loads in blood would present a far easier enzymatic target.\n\nThere is still considerable distance between these results and a clinical application. The enzyme would need to be humanized to avoid immune clearance, encapsulated for delivery \u2014 perhaps within red blood cells, as some recent work on enzyme therapeutics has explored \u2014 and tested in animal models before any translation could begin. The current study did not address immunogenicity, biodistribution, or long-term stability in blood. What it does establish, convincingly, is the enzymatic principle: PETase is active in human serum at body temperature, it converts microplastic PET cleanly to monomers, and it does not harm human cells at biologically relevant concentrations. Given the mounting evidence for [health consequences of microplastic exposure](https://pmc.ncbi.nlm.nih.gov/articles/PMC8416353/) \u2014 including inflammation, endocrine disruption, and altered blood coagulation \u2014 an enzymatic clearance strategy deserves serious scientific attention.", "# GitHub README: AlphaFraud\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/AlphaFraud, 2026\n\n# 2. Push the code to the droplet\nbash deploy/deploy.sh\n\n# 3. On the droplet as root: one-time provisioning (packages, user, venv, services, nginx, TLS)\nsudo SERVER_NAME=alphafraud.mdeller.com bash /opt/alphafraud/deploy/provision.sh\n\n# 4. (optional) trigger the first pipeline run immediately\nsudo systemctl start alphafraud-run.service\n```\n\nAfter that, `deploy/deploy.sh` pushes code updates and restarts the web service, and `deploy/release.sh \"message\"` does the whole loop in one step (commit, push to GitHub, then deploy). Deploys are safe to run while a backfill is in progress: the web app is a read-only reader, WAL is set once, and DB writes are retry-wrapped, so restarting the web service never disturbs the running pipeline. The weekly timer fires every Wednesday 00:30 UTC.\n\n| File | Role |\n|---|---|\n| `deploy/provision.sh` | One-time root setup: system packages, service user, venv, systemd units, nginx site, certbot. Also patches HTTP/2 onto certbot's TLS listener afterward (nginx doesn't enable it by default) \u2014 idempotent, so it self-heals on a re-run |\n| `deploy/deploy.sh` | Push code from your Mac (rsync over SSH) and restart the web service |\n| `deploy/release.sh` | One-shot: commit, push to GitHub, then deploy to the droplet |\n| `deploy/alphafraud-web.service` | gunicorn web app (always on) |\n| `deploy/alphafraud-run.service` + `.timer` | Weekly pipeline and its Wednesday 00:30 UTC schedule |\n| `deploy/alphafraud-backfill.service` | Supervised one-time full-archive backfill; auto-restarts on crash, survives reboot (`systemctl enable --now alphafraud-backfill`) |\n| `deploy/nginx-alphafraud.conf` | nginx reverse-proxy site (templated with the server name) |\n\n## \ud83e\uddf1 Stack\n\nPython, biotite and tmtools for structure handling and superposition, numpy and scipy for the metrics, Flask and gunicorn for serving, Plotly for the figures, and SQLite for state. No compiled toolchain, no external database.\n\n## \ud83e\udded Caveats\n\n- \"Deposited after the cutoff\" plus a novelty score is a strong proxy, not proof AlphaFold never saw a homolog: the novelty column makes this explicit so you can filter to the defensible subset.\n- AlphaFold DB models are monomers, so complex and interface accuracy are not tested in this version.\n- Engineered constructs, antibodies and fusion proteins that do not map to a single UniProt accession are flagged and skipped, not compared.\n\n## \u2705 To Do\n\nRoadmap for AlphaFraud, newest ideas at the top. Suggestions welcome.\n\n- [x] **Published as a citable software release.** Zenodo, MIT, concept DOI [10.5281/zenodo.21808182](https://doi.org/10.5281/zenodo.21808182), which always resolves to the latest version. Source only: `data/` is 5.2 GB of API-cached structures and generated ribbons, all regenerable, and the results database is not deposited. The description's figures were read from the live database rather than from this README, which had drifted by a week's worth of runs\n- [x] **Home page and `/api/stats` no longer scan the whole `entities` table.** `overall_stats()` filtered on `status`, which nothing indexed, so the planner had only `SCAN entities`: ~97k rows over 63 MB of pages, dragging in 23 of the table's 28 columns that the aggregate never reads. That was 62 ms on every page load and made AlphaFraud the slowest app on the droplet by an order of magnitude (68 ms against 1-2 ms for its neighbours). A covering index on `(status, confidently_wrong, is_novel, tm_by_experiment, lddt)` lets SQLite answer the whole query from a 2.8 MB index without touching the table, taking it to 4.9 ms, and the result is now cached in `analysis_snapshots` alongside the analysis payload, refreshed by the weekly run. Warm requests are ~6 ms; the launcher's health check went from 193 ms to 52 ms.", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\nThe harvesting of protein crystals is almost always a necessary step in the determination of a protein structure using X-ray crystallographic techniques. However, protein crystals are usually fragile and susceptible to damage during the harvesting process. For this reason, protein crystal harvesting is the single step that remains entirely dependent on skilled human intervention. Automation has been implemented in the majority of other stages of the structure-determination pipeline, including cloning, expression, purification, crystallization and data collection. The gap in automation between crystallization and data collection results in a bottleneck in throughput and presents unfortunate opportunities for crystal damage. Several automated protein crystal harvesting systems have been developed, including systems utilizing microcapillaries, microtools, microgrippers, acoustic droplet ejection and optical traps. However, these systems have yet to be commonly deployed in the majority of crystallography laboratories owing to a variety of technical and cost-related issues. Automation of protein crystal harvesting remains essential for harnessing the full benefits of fourth-generation synchrotrons, free-electron lasers and microfocus beamlines. Furthermore, automation of protein crystal harvesting offers several benefits when compared with traditional manual approaches, including the ability to harvest microcrystals, improved flash-cooling procedures and increased throughput.\n\n## Structural insights into the recognition of phosphopeptide by the FHA domain of kanadaptin (2014)\n\n**Authors:** Xu Q, Deller MC, Nielsen TK, Grant JC, Lesley SA, Elsliger MA, Deacon AM, Wilson IA.  \n**Journal:** PloS one (2014). DOI: 10.1371/journal.pone.0107309. PMID: 25197798.\n\nKanadaptin is a nuclear protein of unknown function that is widely expressed in mammalian tissues. The crystal structure of the forkhead-associated (FHA) domain of human kanadaptin was determined to 1.6 \u00c5 resolution. The structure reveals an asymmetric dimer in which one monomer is complexed with a phosphopeptide mimic derived from a peptide segment from the N-terminus of a symmetry-related molecule as well as a sulfate bound to the structurally conserved phosphothreonine recognition cleft. This structure provides insights into the molecular recognition features utilized by this family of proteins and represents the first evidence that kanadaptin is likely involved in a phosphorylation-mediated signaling pathway. These results will be of use for designing experiments to further probe the function of kanadaptin.\n\n## Crystal structure of a soluble cleaved HIV-1 envelope trimer (2013)\n\n**Authors:** Julien JP, Cupo A, Sok D, Stanfield RL, Lyumkis D, Deller MC, Klasse PJ, Burton DR, Sanders RW, Moore JP, Ward AB, Wilson IA.  \n**Journal:** Science (New York, N.Y.) (2013). DOI: 10.1126/science.1245625. PMID: 24179159.\n\nHIV-1 entry into CD4(+) target cells is mediated by cleaved envelope glycoprotein (Env) trimers that have been challenging to characterize structurally. Here, we describe the crystal structure at 4.7 angstroms of a soluble, cleaved Env trimer that is stabilized and antigenically near-native (termed the BG505 SOSIP.664 gp140 trimer) in complex with a potent broadly neutralizing antibody, PGT122. The structure shows a prefusion state of gp41, the interaction between the component gp120 and gp41 subunits, and how a close association between the gp120 V1/V2/V3 loops stabilizes the trimer apex around the threefold axis. The complete epitope of PGT122 on the trimer involves gp120 V1, V3, and several surrounding glycans. This trimer structure advances our understanding of how Env functions and is presented to the immune system, and provides a blueprint for structure-based vaccine design.\n\n## Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120 (2013)\n\n**Authors:** Kong L, Lee JH, Doores KJ, Murin CD, Julien JP, McBride R, Liu Y, Marozsan A, Cupo A, Klasse PJ, Hoffenberg S, Caulfield M, King CR, Hua Y, Le KM, Khayat R, Deller MC, Clayton T, Tien H, Feizi T, Sanders RW, Paulson JC, Moore JP, Stanfield RL, Burton DR, Ward AB, Wilson IA.  \n**Journal:** Nature structural & molecular biology (2013). DOI: 10.1038/nsmb.2594. PMID: 23708606.\n\nA substantial proportion of the broadly neutralizing antibodies (bnAbs) identified in certain HIV-infected donors recognize glycan-dependent epitopes on HIV-1 gp120. Here we elucidate how the bnAb PGT 135 binds its Asn332 glycan-dependent epitope from its 3.1-\u00c5 crystal structure with gp120, CD4 and Fab 17b. PGT 135 interacts with glycans at Asn332, Asn392 and Asn386, using long CDR loops H1 and H3 to penetrate the glycan shield and access the gp120 protein surface. EM reveals that PGT 135 can accommodate the conformational and chemical diversity of gp120 glycans by altering its angle of engagement. Combined structural studies of PGT 135, PGT 128 and 2G12 show that this Asn332-dependent antigenic region is highly accessible and much more extensive than initially appreciated, which allows for multiple binding modes and varied angles of approach; thereby it represents a supersite of vulnerability for antibody neutralization.\n\n## Functional and structural characterization of a thermostable acetyl esterase from Thermotoga maritima (2012)\n\n**Authors:** Levisson M, Han GW, Deller MC, Xu Q, Biely P, Hendriks S, Ten Eyck LF, Flensburg C, Roversi P, Miller MD, McMullan D, von Delft F, Kreusch A, Deacon AM, van der Oost J, Lesley SA, Elsliger MA, Kengen SW, Wilson IA.  \n**Journal:** Proteins (2012). DOI: 10.1002/prot.24041. PMID: 22411095.", "# Engineering Evolution: How FAST-PETase and Other Variants Are Transforming Plastic Biodegradation\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| PETase Variant | Function | Structural Differences | Key Mutations | Origin | Engineering Approach | Use | Pros and Cons | Key References |\n| --- | --- | --- | --- | --- | --- | --- | --- | --- |\n| IsPETase | Native PET hydrolysis at 30\u201335\u00b0C; good activity on amorphous PET | Wider substrate binding site formed by \u03b28-\u03b16 loop; three-residue insertion extends subsite IIc | Wild type (S160, D206, H237) | *Ideonella sakaiensis* mesophilic bacterium | Natural enzyme; template for all engineering efforts | Low-temperature PET degradation; discovery origin | Pros: Effective on amorphous PET at mild temps. Cons: Weak thermostability, slow crystalline PET hydrolysis | [Liu, F., et al., Front Bioeng Biotechnol, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC10546322/) |\n| DuraPETase | Enhanced thermostability (Tm +31\u00b0C) and activity (\u00d7300) on semicrystalline PET at 37\u00b0C | Aromatic tunnel via Y87, Q119Y, W159H; I168R adds electrostatic interaction; packing optimized for rigidity | S214H, I168R, W159H, S188Q, R280A, A180I, G165A, Q119Y, L117F, T140D | Engineered from IsPETase | GRAPE algorithm (greedy clustering for stabilizing mutations) | Efficient semicrystalline PET and nanoplastic degradation at 37\u00b0C | Pros: High activity on semicrystalline PET, full nanoplastic degradation. Cons: Less effective at higher temps | [Liu, F., et al., Front Bioeng Biotechnol, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC10546322/) |\n| HotPETase | Thermophilic PET degradation at 60\u201365\u00b0C; Tm = 82.5\u00b0C | P181V stabilizes \u03b2-sheet; S214Y/N233C/S282C add rigidity; 21 assembled mutations | 21 mutations: S121E/D186H/R280A/P181V/S214Y/N233C/S282C, etc. | Directed evolution from IsPETase (TS variant) | Automated high-throughput evolution with heat selection | Industrial recycling at high temperatures | Pros: High efficiency >60\u00b0C. Cons: Requires heating, complex mutant profile | [Liu, F., et al., Front Bioeng Biotechnol, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC10546322/) |\n| LCC-ICCG | Thermophilic cutinase for PET depolymerization at 65\u201372\u00b0C | F243I expands binding site; D238C/S283C adds disulfide bridge; Y127G for flexibility | F243I, D238C, S283C, Y127G | Leaf-branch compost metagenome cutinase | Rational design & disulfide bridge insertion | Recycling of bottles and postconsumer PET at high temps | Pros: Very stable at high temps. Cons: Activity drops if PET crystallizes >70\u00b0C | [Pfaff, L., et al., FEBS J, 2023](https://febs.onlinelibrary.wiley.com/doi/10.1111/febs.16736) |\n| Z1-PETase | Balanced activity and stability with 75% amorphous PET hydrolysis in 96h (30\u00b0C) | Disulfide bonds (N233C-S282C/A171C-S193C); P181V for soluble yield; charge modifications | 13 mutations: S121E, D186H, N246D, S242T, P181V, A180V, N37D, R132E, R224E, N233C, S282C | Balanced engineering from IsPETase | Combinatorial mutagenesis, emphasis on protein yield & stability | Mesophilic PET recycling; consistent activity | Pros: 40\u00d7 expression yield, works across temperatures. Cons: More moderate thermostability | [Lee, C., et al., bioRxiv, 2022](https://www.biorxiv.org/content/10.1101/2022.11.14.516528v1.full.pdf) |\n| FAST-PETase | High activity hydrolysis of crystalline PET (\u00d738) at 50\u00b0C | Loop stabilization via N233K/E204 salt bridge; D186H additional H bonds; structural rigidity | S121E, D186H, R224Q, N233K, R280A | Engineered from IsPETase | Machine learning (MutCompute structure-guided) | Recycling of highly crystalline PET with minimal pretreatment | Pros: Works on challenging substrates, moderate temp optimum. Cons: Inhibited above 50\u00b0C | [Stockinger, P., et al., Comput Struct Biotechnol J, 2025](https://pmc.ncbi.nlm.nih.gov/articles/PMC11946493/) |\n| TurboPETase | 98.2% PET depolymerization at 200g/kg loading in 8hr | W104L disrupts Pro248; F243T facilitates active site turnover; increased flexibility | 8 mutations: from BhrPETase M6 including W104L, F243T | Engineered from BhrPETase | Computational redesign, product release optimization | Industrial depolymerization of PET at scale | Pros: Industry-ready efficiency, minimal crystal residue. Cons: Prefers 65\u201368\u00b0C, substrate pretreatment | [Cui, Y., et al., Nat Commun, 2024](https://www.nature.com/articles/s41467-024-45662-9) |\n| Cut190**SS | Calcium-dependent PET hydrolysis; conformation-switching mechanism at 65\u00b0C | Site 1 flexibility (F77L); ejecting form for product release; two Ca2+ binding sites | S226P, R228S, \u0394C-terminal 3, F77L | Actinomycete cutinase | Crystal structure-guided rational design | High-temp PET hydrolysis with calcium regulation | Pros: Unique regulatory mechanism, stable. Cons: needs Ca2+, dynamic structure | [Kawai, F., et al., Appl Microbiol Biotechnol, 2022](https://pmc.ncbi.nlm.nih.gov/articles/PMC9606173/) |", "# Publications of Marc C. Deller\n\n- **Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design** (2017). *PLoS Pathogens*. journal-article. DOI: 10.1371/journal.ppat.1006212. Authors: Irimia, Adriana; Serra, Andreia M; Sarkar, Anita; Jacak, Ronald; Kalyuzhniy, Oleksandr; Sok, Devin; Saye-Francisco, Karen L; Schiffner, Torben; Tingle, Ryan; Kubitz, Michael et al.\n- **The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase** (2017). *ACS Chemical Biology*. journal-article. DOI: 10.1021/acschembio.7b00016. Authors: Span, Elise A; Suess, Daniel LM; Deller, Marc C; Britt, R David; Marletta, Michael A.\n- **Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor** (2016). *Immunity*. journal-article. DOI: 10.1016/j.immuni.2016.03.006. Authors: Kong, Leopold; Ju, Bin; Chen, Yajing; He, Linling; Ren, Li; Liu, Jiandong; Hong, Kunxue; Su, Bin; Wang, Zheng; Ozorowski, Gabriel et al.\n- **Protein stability: A crystallographer's perspective** (2016). *Acta Crystallographica Section:F Structural Biology Communications*. journal-article. DOI: 10.1107/S2053230X15024619. Authors: Deller, Marc C; Kong, Leopold; Rupp, Bernhard.\n- **Complete epitopes for vaccine design derived from a crystal structure of the broadly neutralizing antibodies PGT128 and 8ANC195 in complex with an HIV-1 Env trimer** (2015). *Acta Crystallographica Section D: Biological Crystallography*. journal-article. DOI: 10.1107/S1399004715013917. Authors: Kong, Leopold; Torrents de la Pe\u00f1a, Alba; Deller, Marc C; Garces, Fernando; Sliepen, Kwinten; Hua, Yuanzi; Stanfield, Robyn L; Sanders, Rogier W; Wilson, Ian A.\n- **Crystal structure of a two-subunit TrkA octameric gating ring assembly** (2015). *PLoS ONE*. journal-article. DOI: 10.1371/journal.pone.0122512. Authors: Deller, Marc C; Johnson, Hope A; Miller, Mitchell D; Spraggon, Glen; Elsliger, Marc-Andr\u00e9; Wilson, Ian A; Lesley, Scott A.\n- **Models of protein-ligand crystal structures: Trust, but verify** (2015). *Journal of Computer-Aided Molecular Design*. journal-article. DOI: 10.1007/s10822-015-9833-8. Authors: Deller, Marc C; Rupp, Bernhard.\n- **Structure-based discovery of NANOG variant with enhanced properties to promote self-renewal and reprogramming of pluripotent stem cells** (2015). *Proceedings of the National Academy of Sciences of the United States of America*. journal-article. DOI: 10.1073/pnas.1502855112. Authors: Hayashi, Yohei; Caboni, Laura; Das, Debanu; Yumoto, Fumiaki; Clayton, Thomas; Deller, Marc C; Nguyen, Phuong; Farr, Carol L; Chiu, Hsiu-Ju; Miller, Mitchell D et al.\n- **Approaches to automated protein crystal harvesting** (2014). *Acta Crystallographica Section F:Structural Biology Communications*. journal-article. DOI: 10.1107/S2053230X14000387. Authors: Deller, Marc C; Rupp, Bernhard.\n- **Structural insights into the recognition of phosphopeptide by the FHA domain of Kanadaptin** (2014). *PLoS ONE*. journal-article. DOI: 10.1371/journal.pone.0107309. Authors: Xu, Qingping; Deller, Marc C; Nielsen, Tine K; Grant, Joanna C; Lesley, Scott A; Elsliger, Marc-Andre; Deacon, Ashley M; Wilson, Ian A.\n- **Crystal structure of a soluble cleaved HIV-1 envelope trimer** (2013). *Science*. journal-article. DOI: 10.1126/science.1245625. Authors: Julien, Jean-Philippe; Cupo, Albert; Sok, Devin; Stanfield, Robyn L; Lyumkis, Dmitry; Deller, Marc C; Klasse, Per-Johan; Burton, Dennis R; Sanders, Rogier W; Moore, John P et al.\n- **Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120** (2013). *Nature Structural and Molecular Biology*. journal-article. DOI: 10.1038/nsmb.2594. Authors: Kong, Leopold; Lee, Jeong Hyun; Doores, Katie J; Murin, Charles D; Julien, Jean-Philippe; McBride, Ryan; Liu, Yan; Marozsan, Andre; Cupo, Albert; Klasse, Per-Johan et al.\n- **Functional and structural characterization of a thermostable acetyl esterase from Thermotoga maritima** (2012). *Proteins: Structure, Function and Bioinformatics*. journal-article. DOI: 10.1002/prot.24041. Authors: Levisson, Mark; Han, Gye Won; Deller, Marc C; Xu, Qingping; Biely, Peter; Hendriks, Sjon; Ten Eyck, Lynn F; Flensburg, Claus; Roversi, Pietro; Miller, Mitchell D et al.\n- **Structure of hepatitis C virus envelope glycoprotein E2 antigenic site 412 to 423 in complex with antibody AP33** (2012). *Journal of Virology*. journal-article. DOI: 10.1128/JVI.01939-12. Authors: Kong, Leopold; Giang, Erick; Nieusma, Travis; Robbins, Justin B; Deller, Marc C; Stanfield, Robyn L; Wilson, Ian A; Law, Mansun.\n- **Crystal structure of a metal-dependent phosphoesterase (YP_910028.1) from Bifidobacterium adolescentis: Computational prediction and experimental validation of phosphoesterase activity** (2011). *Proteins: Structure, Function and Bioinformatics*. journal-article. DOI: 10.1002/prot.23035. Authors: Han, Gye Won; Ko, Jaeju; Farr, Carol L; Deller, Marc C; Xu, Qingping; Chiu, Hsiu\u2010Ju; Miller, Mitchell D; Sefcikova, Jana; Somarowthu, Srinivas; Beuning, Penny J et al.\n- **A conserved fold for fimbrial components revealed by the crystal structure of a putative fimbrial assembly protein (BT1062) from Bacteroides thetaiotaomicron at 2.2 \u00c5 resolution** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309110006548. Authors: Xu, Qingping; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Bakolitsa, Constantina; Cai, Xiaohui; Carlton, Dennis; Chen, Connie; Chiu, H-J; Chiu, Michelle et al.\n- **Bacterial Pleckstrin Homology Domains: A Prokaryotic Origin for the PH Domain** (2010). *Journal of Molecular Biology*. journal-article. DOI: 10.1016/j.jmb.2009.11.006. Authors: Xu, Qingping; Bateman, Alex; Finn, Robert D; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Bakolitsa, Constantina; Carlton, Dennis; Chen, Connie; Chiu, Hsiu-Ju et al.\n- **Conformational changes associated with the binding of zinc acetate at the putative active site of XcTcmJ, a cupin from Xanthomonas campestris pv. campestris** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309109021988. Authors: Axelrod, Herbert L; Kozbial, Piotr; McMullan, Daniel; Krishna, SSri; Miller, Mitchell D; Abdubek, Polat; Acosta, Claire; Astakhova, Tamara; Carlton, Dennis; Caruthers, Jonathan et al.\n- **Crystal Structure of the First Eubacterial Mre11 Nuclease Reveals Novel Features that May Discriminate Substrates During DNA Repair** (2010). *Journal of Molecular Biology*. journal-article. DOI: 10.1016/j.jmb.2010.01.049. Authors: Das, D; Moiani, D; Axelrod, HL; Miller, MD; McMullan, D; Jin, KK; Abdubek, P; Astakhova, T; Burra, P; Carlton, D et al.\n- **Open and closed conformations of two SpoIIAA-like proteins (YP-749275.1 and YP-001095227.1) provide insights into membrane association and ligand binding** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309109042481. Authors: Kumar, Abhinav; Lomize, Andrei; Jin, Kevin K; Carlton, Dennis; Miller, Mitchell D; Jaroszewski, Lukasz; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Chiu, H-J et al.\n- **Structure of BT-3984, a member of the SusD/RagB family of nutrient-binding molecules** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309110032999. Authors: Bakolitsa, Constantina; Xu, Qingping; Rife, Christopher L; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Carlton, Dennis; Chen, Connie; Chiu, H-J; Clayton, Thomas et al.\n- **Structure of Bacteroides thetaiotaomicron BT2081 at 2.05 \u00c5 resolution: The first structural representative of a new protein family that may play a role in carbohydrate metabolism** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309110028228. Authors: Yeh, Andrew P; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Bakolitsa, Constantina; Cai, Xiaohui; Carlton, Dennis; Chen, Connie; Chiu, H-J; Chiu, Michelle et al.\n- **Structure of LP2179, the first representative of Pfam family PF08866, suggests a new fold with a role in amino-acid metabolism** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309109023689. Authors: Bakolitsa, Constantina; Kumar, Abhinav; Carlton, Dennis; Miller, Mitchell D; Krishna, S Sri; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Chiu, H-J; Clayton, Thomas et al.\n- **Structure of a membrane-attack complex/perforin (MACPF) family protein from the human gut symbiont Bacteroides thetaiotaomicron** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309110023055. Authors: Xu, Qingping; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Bakolitsa, Constantina; Cai, Xiaohui; Carlton, Dennis; Chen, Connie; Chiu, H-J; Clayton, Thomas et al.\n- **Structure of a putative NTP pyrophosphohydrolase: YP-001813558.1 from Exiguobacterium sibiricum 255-15** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309110025534. Authors: Han, Gye Won; Elsliger, M-A; Yeates, Todd O; Xu, Qingping; Murzin, Alexey G; Krishna, SSri; Jaroszewski, Lukasz; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L et al.\n- **Structure of the first representative of Pfam family PF04016 (DUF364) reveals enolase and Rossmann-like folds that combine to form a unique active site with a possible role in heavy-metal chelation** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309110007517. Authors: Miller, Mitchell D; Aravind, L; Bakolitsa, Constantina; Rife, Christopher L; Carlton, Dennis; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Chiu, H-J; Clayton, Thomas et al.\n- **Structure of the first representative of Pfam family PF09410 (DUF2006) reveals a structural signature of the calycin superfamily that suggests a role in lipid metabolism** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309109037749. Authors: Miller, Mitchell D; Aravind, L; Bakolitsa, Constantina; Rife, Christopher L; Carlton, Dennis; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Chiu, H-J; Clayton, Thomas et al.\n- **Structure of the \u03b3-D-glutamyl-L-diamino acid endopeptidase YkfC from Bacillus cereus in complex with L-Ala-\u03b3-d-Glu: Insights into substrate recognition by NlpC/P60 cysteine peptidases** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309110021214. Authors: Xu, Qingping; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Bakolitsa, Constantina; Cai, Xiaohui; Carlton, Dennis; Chen, Connie; Chiu, H-J; Chiu, Michelle et al.\n- **Structures of the first representatives of Pfam family PF06684 (DUF1185) reveal a novel variant of the Bacillus chorismate mutase fold and suggest a role in amino-acid metabolism** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309109050647. Authors: Bakolitsa, Constantina; Kumar, Abhinav; Jin, Kevin K; McMullan, Daniel; Krishna, S Sri; Miller, Mitchell D; Abdubek, Polat; Acosta, Claire; Astakhova, Tamara; Axelrod, Herbert L et al.\n- **Structures of the first representatives of Pfam family PF06938 (DUF1285) reveal a new fold with repeated structural motifs and possible involvement in signal transduction** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309109050416. Authors: Bakolitsa, Constantina; Kumar, Abhinav; Jin, Kevin K; McMullan, Daniel; Krishna, S Sri; Miller, Mitchell D; Abdubek, Polat; Acosta, Claire; Astakhova, Tamara; Axelrod, Herbert L et al.\n- **Structures of three members of Pfam PF02663 (FmdE) implicated in microbial methanogenesis reveal a conserved \u03b1+\u0392 core domain and an auxiliary C-terminal treble-clef zinc finger** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309110020166. Authors: Axelrod, Herbert L; Das, Debanu; Abdubek, Polat; Astakhova, Tamara; Bakolitsa, Constantina; Carlton, Dennis; Chen, Connie; Chiu, H-J; Clayton, Thomas; Deller, Marc C et al.\n- **The crystal structure of a bacterial Sufu-like protein defines a novel group of bacterial proteins that are similar to the N- terminal domain of human Sufu** (2010). *Protein Science*. journal-article. DOI: 10.1002/pro.497. Authors: Das, Debanu; Finn, Robert D; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Bakolitsa, Constantina; Cai, Xiaohui; Carlton, Dennis; Chen, Connie; Chiu, Hsiu\u2010Ju et al.\n- **The structure of BVU2987 from Bacteroides vulgatus reveals a superfamily of bacterial periplasmic proteins with possible inhibitory function** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309109046788. Authors: Das, Debanu; Finn, Robert D; Carlton, Dennis; Miller, Mitchell D; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Bakolitsa, Constantina; Chen, Connie; Chiu, H-J et al.\n- **The structure of Haemophilus influenzae prephenate dehydrogenase suggests unique features of bifunctional TyrA enzymes** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309110021688. Authors: Chiu, H-J; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Carlton, Dennis; Clayton, Thomas; Das, Debanu; Deller, Marc C; Duan, Lian; Feuerhelm, Julie et al.\n- **The structure of Jann-2411 (DUF1470) from Jannaschia sp. at 1.45 \u00c5 resolution reveals a new fold (the ABATE domain) and suggests its possible role as a transcription regulator** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309109025196. Authors: Bakolitsa, Constantina; Bateman, Alex; Jin, Kevin K; McMullan, Daniel; Krishna, S Sri; Miller, Mitchell D; Abdubek, Polat; Acosta, Claire; Astakhova, Tamara; Axelrod, Herbert L et al.\n- **The structure of KPN03535 (gi|152972051), a novel putative lipoprotein from Klebsiella pneumoniae, reveals an OB-fold** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309109018168. Authors: Das, Debanu; Kozbial, Piotr; Han, Gye Won; Carlton, Dennis; Jaroszewski, Lukasz; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Bakolitsa, Constantina; Chen, Connie et al.\n- **The structure of SSO2064, the first representative of Pfam family PF01796, reveals a novel two-domain zinc-ribbon OB-fold architecture with a potential acyl-CoA-binding role** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309110002514. Authors: Krishna, S Sri; Aravind, L; Bakolitsa, Constantina; Caruthers, Jonathan; Carlton, Dennis; Miller, Mitchell D; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Chiu, H-J et al.\n- **The structure of the first representative of Pfam family PF06475 reveals a new fold with possible involvement in glycolipid metabolism** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309109022684. Authors: Das, Debanu; Grishin, Nick V; Kumar, Abhinav; Carlton, Dennis; Bakolitsa, Constantina; Miller, Mitchell D; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Burra, Prasad et al.\n- **The structure of the first representative of Pfam family PF09836 reveals a two-domain organization and suggests involvement in transcriptional regulation** (2010). *Acta Crystallographica Section F: Structural Biology and Crystallization Communications*. journal-article. DOI: 10.1107/S1744309109022672. Authors: Das, Debanu; Grishin, Nick V; Kumar, Abhinav; Carlton, Dennis; Bakolitsa, Constantina; Miller, Mitchell D; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Burra, Prasad et al.\n- **A Structural Basis for the Regulatory Inactivation of DnaA** (2009). *Journal of Molecular Biology*. journal-article. DOI: 10.1016/j.jmb.2008.10.059. Authors: Xu, Qingping; McMullan, Daniel; Abdubek, Polat; Astakhova, Tamara; Carlton, Dennis; Chen, Connie; Chiu, Hsiu-Ju; Clayton, Thomas; Das, Debanu; Deller, Marc C et al.\n- **Crystal Structure of Histidine Phosphotransfer Protein ShpA, an Essential Regulator of Stalk Biogenesis in Caulobacter crescentus** (2009). *Journal of Molecular Biology*. journal-article. DOI: 10.1016/j.jmb.2009.05.023. Authors: Xu, Qingping; Carlton, Dennis; Miller, Mitchell D; Elsliger, Marc-Andr\u00e9; Krishna, S Sri; Abdubek, Polat; Astakhova, Tamara; Burra, Prasad; Chiu, Hsiu-Ju; Clayton, Thomas et al.\n- **Crystal structure of a novel Sm-like protein of putative cyanophage origin at 2.60 A resolution** (2009). *Proteins: Structure, Function and Bioinformatics*. journal-article. DOI: 10.1002/prot.22360. Authors: Das, Debanu; Kozbial, Piotr; Axelrod, Herbert L; Miller, Mitchell D; McMullan, Daniel; Krishna, S Sri; Abdubek, Polat; Acosta, Claire; Astakhova, Tamara; Burra, Prasad et al.\n- **Crystal structure of the Fic (filamentation induced by cAMP) family protein SO4266 (gi|24375750) from Shewanella oneidensis MR-1 at 1.6 \u00c5 resolution** (2009). *Proteins: Structure, Function and Bioinformatics*. journal-article. DOI: 10.1002/prot.22338. Authors: Das, Debanu; Krishna, S Sri; McMullan, Daniel; Miller, Mitchell D; Xu, Qingping; Abdubek, Polat; Acosta, Claire; Astakhova, Tamara; Axelrod, Herbert L; Burra, Prasad et al.\n- **Structural Basis of Murein Peptide Specificity of a \u03b3-D-Glutamyl-L-Diamino Acid Endopeptidase** (2009). *Structure*. journal-article. DOI: 10.1016/j.str.2008.12.008. Authors: Xu, Qingping; Sudek, Sebastian; McMullan, Daniel; Miller, Mitchell D; Geierstanger, Bernhard; Jones, David H; Krishna, S Sri; Spraggon, Glen; Bursalay, Badry; Abdubek, Polat et al.\n- **Structural and functional characterizations of SsgB, a conserved activator of developmental cell division in morphologically complex actinomycetes** (2009). *Journal of Biological Chemistry*. journal-article. DOI: 10.1074/jbc.M109.018564. Authors: Xu, Qingping; Traag, Bj\u00f8rn A; Willemse, Joost; McMullan, Daniel; Miller, Mitchell D; Elsliger, Marc-Andr\u00e9; Abdubek, Polat; Astakhova, Tamara; Axelrod, Herbert L; Bakolitsa, Constantina et al.\n- **Crystal structures of MW1337R and lin2004: Representatives of a novel protein family that adopt a four-helical bundle fold** (2008). *Proteins: Structure, Function and Genetics*. journal-article. DOI: 10.1002/prot.22020. Authors: Kozbial, Piotr; Xu, Qingping; Chiu, Hsiu\u2010Ju; McMullan, Daniel; Krishna, S Sri; Miller, Mitchell D; Abdubek, Polat; Acosta, Claire; Astakhova, Tamara; Axelrod, Herbert L et al.\n- **Crystal structure of homoserine O-succinyltransferase from Bacillus cereus at 2.4 \u00c5 resolution** (2007). *Proteins: Structure, Function and Genetics*. journal-article. DOI: 10.1002/prot.21208. Authors: Zubieta, Chloe; Krishna, S Sri; McMullan, Daniel; Miller, Mitchell D; Abdubek, Polat; Agarwalla, Sanjay; Ambing, Eileen; Astakhova, Tamara; Axelrod, Herbert L; Carlton, Dennis et al.\n- **Crystal structures of two novel dye-decolorizing peroxidases reveal a \u03b2-barrel fold with a conserved heme-binding motif** (2007). *Proteins: Structure, Function and Genetics*. journal-article. DOI: 10.1002/prot.21550. Authors: Zubieta, Chloe; Krishna, S Sri; Kapoor, Mili; Kozbial, Piotr; McMullan, Daniel; Axelrod, Herbert L; Miller, Mitchell D; Abdubek, Polat; Ambing, Eileen; Astakhova, Tamara et al.\n- **Identification and structural characterization of heme binding in a novel dye-decolorizing peroxidase, TyrA** (2007). *Proteins: Structure, Function and Genetics*. journal-article. DOI: 10.1002/prot.21673. Authors: Zubieta, Chloe; Joseph, Rosanne; Sri Krishna, S; McMullan, Daniel; Kapoor, Mili; Axelrod, Herbert L; Miller, Mitchell D; Abdubek, Polat; Acosta, Claire; Astakhova, Tamara et al.\n- **Cell surface receptors** (2000). *Current Opinion in Structural Biology*. journal-article. DOI: 10.1016/S0959-440X(00)00072-5. Authors: Deller, Marc C; Jones, E Yvonne.\n- **Crystal structure and functional dissection of the cytostatic cytokine oncostatin M** (2000). *Structure*. journal-article. DOI: 10.1016/S0969-2126(00)00176-3. Authors: Deller, Marc C; Hudson, Keith R; Ikemizu, Shinji; Bravo, Jer\u00f3nimo; Jones, E Yvonne; Heath, John K.", "# When Protein Design Becomes a Security Problem\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nThe path forward demands a layered approach. [Stuart, Bright, and Horvitz in *NAM Perspectives*](https://pmc.ncbi.nlm.nih.gov/articles/PMC11741535/) argue that AI biodesign tools should be treated as a strategic biosecurity asset rather than simply a scientific convenience \u2014 capable of designing diagnostic binders, novel antigens, and neutralising proteins against emerging pathogens in hours rather than months. Treating these capabilities as purely defensive, however, requires the same institutions that benefit from them to actively invest in closing the gaps they open. [Pannu et al. in *PLOS Computational Biology*](https://journals.plos.org/ploscompbiol/article?id=10.1371%2Fjournal.pcbi.1012975) propose that AI model evaluations be prioritised according to the consequence severity of potential misuse, with biosecurity assessments completed before deployment rather than in response to demonstrated harm. Structure-aware screening, robust data governance, and pre-deployment safety evaluation are not in conflict with scientific openness \u2014 they are the conditions that make it sustainable.", "# ChatPDB \u2013 A protein-structure-literate AI assistant for the Protein Data Bank: RAG for facts, a QLoRA-tuned model for behaviour\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n|  | Round 1 (32B v1) |\n| --- | --- |\n| Base model | mlx-community/Qwen3-32B-4bit |\n| SFT dataset | v7 \u2014 97,272 examples (77,818 train / 9,727 valid / 9,727 test) |\n| LoRA config | rank 32, scale 64, RSLoRA, num_layers 32/64, dropout 0.05 |\n| max_seq_length | 1,536 (real-measured: covers 100% of train.jsonl, p99.9=1,190, max=1,973) |\n| Iterations trained | ~803 of a planned 1,420 (stopped early \u2014 see below) |\n| Best checkpoint | true iter 600, val_loss 0.176 |\n| Stop reason | real val-loss plateau: block-averaged val loss stopped improving from ~iter 300 onward (0.202 best-block avg at iters 200-300, no further downward trend through iter 800) \u2014 not overfitting, just diminishing returns on the remaining budget |\n| RAG corpus | 42 source files, 105,463 indexed chunks (ChromaDB + BAAI/bge-base-en-v1.5) |\n| Real incidents this round | crashed mid-run at true iter ~630, resumed from the true-iter-600 checkpoint (weights-only resume \u2014 iteration counter and LR schedule do not carry over, see `PROJECT_PLAN.md`); a real `mlx_lm` bug (val-loss/train-loss step misalignment) found and patched locally |\n\nFull narrative for every round, including every real bug found and fixed along the way, lives in [`PROJECT_PLAN.md`](PROJECT_PLAN.md).\n\n## \ud83d\udcca Evaluation\n\n`eval/eval_pdb.py` (single-model harness) and `eval/compare/eval_compare.py` (multi-round comparison \u2014 auto-managed `mlx_lm.server`, `ResourceMonitor`, `--resume`, HTML + Markdown reports) score every response against **real, corpus-grounded ground truth** \u2014 PDB ID validity checks real corpus membership (not just format), cross-reference accuracy checks real SIFTS UniProt/CATH/EC mappings, numerical fidelity checks real resolution/R-free/chain counts, tool executability runs real Biopython blocks in the Phase 6 sandbox. No metric relies on hand-authored ground truth.\n\n| Metric | Round 1 (32B v1) smoke-test sample |\n| --- | --- |\n| PDB ID validity | 89\u2013100% (`--n 20` / `--limit 10` runs) |\n| Cross-reference accuracy | 50% |\n| Tool executability | n/a in the 10-example sample (0/1 in the 20-example sample \u2014 the one attempt was a DSSP block, correctly excluded as not-yet-enabled per Phase 6, not a failure) |\n| Numerical fidelity | 33\u201340% |\n| Refusal accuracy | 100% |\n| Degeneration-free | 100% |\n\nThese are small smoke-test samples (`--n 20`/`--limit 10`), run to verify the harness end to end \u2014 not a full evaluation pass. Run `python eval/eval_pdb.py --n 200` for a real-sized, seeded sample once a model server is up (see **How to run** below).\n\n## \ud83d\udcda Corpus\n\nwwPDB/RCSB structures and metadata, SIFTS UniProt/Pfam/CATH/GO/EC/InterPro mappings, TWILIGHT ligand-fit scores, PDB-REDO refinement deltas, EMDB map metadata, OPM membrane placement, MobiDB disorder predictions, SCOP2 fold classification, sequence-redundancy clusters, obsolete-entry records, BindingDB affinities, STRING interactions, Pharos druggability, AlphaFold predictions, AlphaFraud PDB-vs-AlphaFold disagreement data, PubMed/CrossRef citation verification, and full PyMOL/ChimeraX/py3Dmol command references \u2014 see [`data/README.md`](data/README.md) for the complete dataset-version history and construction rules.\n\n## \ud83c\udf10 Web deployment\n\n**Live at [chatpdb.mdeller.com](https://chatpdb.mdeller.com).** Not a rebuilt chat-bubble interface: the browser gets a real terminal (xterm.js) connected over WebSocket to a server-side PTY running the actual, unmodified `scripts/chat.py` \u2014 the same Rich-rendered markdown, corpus retrieval, and slash commands as running it locally.\n\n**No GPU on the droplet.** `chat_remote.py` (sibling to [ChemSage](https://github.com/bellcheddar/ChemSage)\u2018s own, reusing the same pattern) is a drop-in launcher: at import time it patches `mlx_lm.generate`/`mlx_lm.stream_generate` to call a HuggingFace Space running the real fused 32B model on ZeroGPU (a serverless, on-demand GPU allocation), then runs `chat.py` completely unaware anything is different.\n\n**Serving stack:** gunicorn with a single `gevent` worker (100 concurrent connections) rather than `sync` workers \u2014 a terminal session is I/O-bound (waiting on the remote model and on keystrokes), not CPU-bound, so one greenlet-based worker serves many concurrent sessions cheaply. nginx reverse-proxies both the WebSocket endpoint (`/ws`, with `Upgrade`/`Connection` headers and a long read/send timeout for idle-but-open sessions) and the static app shell, behind a Let\u2019s Encrypt certificate. Runs as a dedicated systemd service on the droplet alongside ChemSage and BoltzMaker\u2019s own web apps.\n\n**Source isolation:** the web app (`web/flask_app/`) lives on a separate `chatpdb-web` git branch/worktree, so it never touches the tested training/eval code on `main`.\n\n## \ud83e\uddf1 Stack\n\nMLX-LM (QLoRA fine-tuning + serving), Biopython/gemmi/DSSP/PyMOL/ChimeraX/MDAnalysis/ProDy (tool calls), ChromaDB + BAAI/bge-base-en-v1.5 (RAG retrieval), pandas/plotly/bio3d (analysis tools taught to the model), Weights & Biases (training observability).\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.** Structural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/chatPDB](https://github.com/bellcheddar/chatPDB) |", "# marcdeller.com: My Learning\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n## Continuous Growth: Courses Shaping My Expertise\n\nMy certifications in AI, Vibe Coding, Project Management, and Data Analysis demonstrate adaptability and expertise with cutting-edge technologies. They position me to bridge external partners and internal AI/software teams while driving innovative therapeutic discovery programs. Prepared to excel in AI-driven industries!\n\n* **Vibe Coding with AI**: Certifications in artificial intelligence for Dashboards, Interactive Apps, and Teamworking\n* **Data Analysis & Visualization**: Experience with tools like R and Shiny for biological data interpretation\n* **Project Management**: Advanced training in leadership and adaptability for managing complex scientific programs\n* **Cloud Technologies**: AWS certification showcasing expertise in modern cloud architectures for data processing\n\nTop Learnings\n\n### [2025 \u2013 Embracing AI-Enabled Communication and Adaptability](https://marcdeller.com/embracing-ai-enabled-communication-and-adaptability/)\n\n[2025 \u2013 Embracing AI-Enabled Communication and Adaptability](https://marcdeller.com/embracing-ai-enabled-communication-and-adaptability/)\n\n### [2025 \u2013 Building Data and Project Management Expertise](https://marcdeller.com/building-data-and-project-management-expertise/)\n\n[2025 \u2013 Building Data and Project Management Expertise](https://marcdeller.com/building-data-and-project-management-expertise/)\n\n### [2025 \u2013 Cultivating Human Skills and Decision-Making](https://marcdeller.com/cultivating-human-skills-and-decision-making/)\n\n[2025 \u2013 Cultivating Human Skills and Decision-Making](https://marcdeller.com/cultivating-human-skills-and-decision-making/)\n\nCertificates\n\n### [2026 \u2013 Anthropic Academy](https://marcdeller.com/anthropic-academy/)\n\nCertificates\n\n### [2026 \u2013 Learn to code with AI](https://marcdeller.com/learn-to-code-with-ai/)\n\nCertificates\n\n### [2026 \u2013 From Tool to Teammates: Working Smarter with AI](https://marcdeller.com/from-tool-to-teammates-working-smarter-with-ai/)\n\nCertificates\n\n### [2025 \u2013 Foundations of Project Management](https://marcdeller.com/foundations-of-project-management/)\n\nCertificates\n\n### [2025 \u2013 Introduction to R: Basic R syntax](https://marcdeller.com/introduction-to-r-basic-r-syntax/)\n\nCertificates\n\n### [2025 \u2013 Dashboard Development with Shiny: GenAI for Retail Analysis](https://marcdeller.com/dashboard-development-with-shiny-genai-for-retail-analysis/)\n\nCertificates\n\n### [2025 \u2013 Strengthen Your Decision-Making with Generative AI](https://marcdeller.com/strengthen-your-decision-making-with-generative-ai/)\n\nCertificates\n\n### [2025 \u2013 Amplify Your Communication Skills with Generative AI](https://marcdeller.com/amplify-your-communication-skills-with-generative-ai/)\n\nCertificates\n\n### [2025 \u2013 Building an Adaptability Mindset in the Age of AI](https://marcdeller.com/building-an-adaptability-mindset-in-the-age-of-ai/)\n\nCertificates\n\n### [2025 \u2013 Investing in Human Skills in the Age of AI](https://marcdeller.com/investing-in-human-skills-in-the-age-of-ai/)\n\nCertificates\n\n### [2025 \u2013 Foundations: Data, Data, Everywhere](https://marcdeller.com/foundations-data-data-everywhere/)\n\nCertificates\n\n### [2025 \u2013 Serverless Architectures on AWS](https://marcdeller.com/serverless-architectures-on-aws/)\n\nCertificates\n\n### [2025 \u2013 Introduction to Artificial Intelligence](https://marcdeller.com/introduction-to-artificial-intelligence/)\n\nCertificates\n\n### [2025 \u2013 Cultivating Human Skills and Decision-Making](https://marcdeller.com/cultivating-human-skills-and-decision-making/)\n\nCertificates\n\n### [2025 \u2013 Building Data and Project Management Expertise](https://marcdeller.com/building-data-and-project-management-expertise/)\n\nCertificates\n\n### [2025 \u2013 Embracing AI-Enabled Communication and Adaptability](https://marcdeller.com/embracing-ai-enabled-communication-and-adaptability/)", "# OpenBind Publishes a Dense Structure-Affinity Dataset for Structure-Based AI\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-05-06 on marcdeller.com (categories: ai, blog, sb).*\n\n* [**OpenBind**](https://openbind.uk)\n\n* **The data bottleneck is real:** Cofolding methods fail when test systems sit far from PDB training data.\n* **Receptor conformation dominates failure modes:** Rigid cross-docking into apo or AlphaFold structures falls below 5% success.\n* **Fragment screens are training data:** Fine-tuning OpenFold3-p2 on fragment-bound structures more than doubled cofolding success rates.\n* **Affinity prediction is unsolved:** Structure-aware models still struggle to beat a molecular-weight baseline on a single-target series.\n\n1. **OpenBind\u2019s first release: A structure-affinity dataset for structure-based AI:** [OpenBind Consortium, openbind.uk, 2026](https://openbind.uk/news/blog-openbinds-first-release-a-structure-affinity-dataset-for-structure-based-ai/)\n2. **Have protein-ligand cofolding methods moved beyond memorisation?:** [\u0160krinjar et al., bioRxiv, 2025](https://doi.org/10.1101/2025.02.03.636309)\n3. **PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences:** [Buttenschoen et al., Chemical Science, 2024](https://doi.org/10.1039/D3SC04185A)\n4. **PLINDER: The protein-ligand interactions dataset and evaluation resource:** [Durairaj et al., bioRxiv, 2024](https://doi.org/10.1101/2024.07.17.603955)\n5. **Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction:** [Passaro et al., bioRxiv, 2025](https://doi.org/10.1101/2025.06.14.659707)\n\n**925 crystallographic binding events and 601 affinity measurements for EV-A71 2A protease, designed to train and stress-test structure-based machine learning**\n\nProtein-ligand prediction has become a central proving ground for AI in drug discovery. Models such as AlphaFold3, Boltz and OpenFold3 report striking accuracy on internal benchmarks, but careful re-evaluation has been less flattering. The [Runs N\u2019 Poses analysis](https://doi.org/10.1101/2025.02.03.636309) showed that cofolding accuracy collapses as binding-site similarity to the training data decreases, suggesting the field has been measuring memorization rather than generalization. [PoseBusters](https://doi.org/10.1039/D3SC04185A) revealed that many AI-predicted poses also fail basic chemistry checks despite passing RMSD cutoffs. Beneath both findings lies the same issue: the Protein Data Bank is biased toward highly characterized target classes, and curated derivatives such as [PLINDER](https://doi.org/10.1101/2024.07.17.603955) have brought the limits of public training data to the forefront of community attention.\n\nThe [OpenBind consortium](https://openbind.uk/news/blog-openbinds-first-release-a-structure-affinity-dataset-for-structure-based-ai/), a Diamond Light Source-led partnership funded by the UK Department for Science, Innovation and Technology, has just released its first dataset to address that gap: 925 crystallographic binding events from 699 compounds on Enterovirus A71 2A protease, paired with 601 KD measurements from a Creoptix WAVE biosensor. The target, chosen with the ASAP antiviral discovery consortium, lies in a region of protein-ligand space dissimilar to that observed in PDB structures available before the AlphaFold3 training cutoff. The team is explicit about what they are trying to fix.\n\n> \u201cWe believe that the next generation of structure-based machine learning methods requires better experimental data, not just new architectures.\u201d[OpenBind Consortium, 2026](https://openbind.uk/news/blog-openbinds-first-release-a-structure-affinity-dataset-for-structure-based-ai/)\n\nThe reference benchmarks make the case. GNINA redocking succeeds on 85% of follow-on compounds, but rigid cross-docking into apo or AlphaFold receptors falls below 5%, defeated by a single binding-site loop. Cofolding partially recovers performance, with Protenix and OpenFold3-p2 leading\u00a0but still trailing redocking. The most arresting result: fine-tuning OpenFold3-p2 on the fragment-bound structures alone lifts cofolding success from 36% to 76%, approaching redocking accuracy without the cognate receptor in hand. Affinity prediction is harder still. Structure-aware models, including [Boltz-2](https://doi.org/10.1101/2025.06.14.659707), struggle to outperform a molecular-weight baseline on the EV-A71 series, a sobering reminder of how much remains between observed binding modes and measured potency.\n\nThe release is CC0 licensed, with benchmarks on GitHub, structures on Fragalysis and Zenodo, and full protocols on protocols.io. OpenBind frames it as the first output of an experimental pipeline that can be repeated across many targets. For anyone building structure-based AI, it provides a stress test that current methods cannot pass by memorization alone, and a reminder that the bottleneck on AI-driven drug discovery may sit in the wet lab as much as the architecture.", "# GitHub README: CODSWALLOP\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/CODSWALLOP, 2026\n\n| Variable | Default | Does |\n|---|---|---|\n| `DROPLET_SSH` | | SSH destination for `deploy/deploy.sh` |\n| `DROPLET_PATH` | `/opt/codswallop` | Where the code lives on the droplet |\n| `SERVER_NAME` | `codswallop.mdeller.com` | Hostname certbot issues a certificate for |\n| `BIND_ADDR` | `127.0.0.1:8006` | Address gunicorn binds to (8000 to 8005 are taken by the sibling apps) |\n| `CACHE_TTL_HOURS` | `168` | How long a cached API response stays fresh |\n| `FAMILY_TTL_HOURS` | `168` | How long an assembled family stays fresh |\n| `MAX_FAMILY_ENTITIES` | `2500` | Cap on entities pulled into one family |\n\nThe two TTLs default to a week because the PDB releases weekly: a family cannot gain a member mid-week.\n\n## \ud83d\udcca Performance\n\nMeasured on carbonic anhydrase II (`P00918`), 1,490 entries, and lysozyme (`P00698`), 1,686 entries.\n\n| Path | Cost |\n|---|---|\n| Cold family assembly | 7 to 12 s (one sequence search plus 30 batched GraphQL calls) |\n| Warm family, served from SQLite | 0.07 s |\n| Assembled family payload | ~1.7 MB JSON, ~240 kB gzipped |\n\nThe payload is deduplicated before it goes over the wire: sequences, primary citations, chemical components and domain annotations all move to family-level lookup tables, because a family is mostly the same construct solved again. Carbonic anhydrase II has 1,490 entities and 232 distinct sequences. Doing this took the page from 6.3 MB to 1.8 MB.\n\n## \ud83c\udfa8 Theme tokens\n\nAnything physical is brass (drawer front, dividers, plate, pins, stamps); all data is cool. That rule is what holds the theme together.\n\n| Token | Dark | Light | Role |\n|---|---|---|---|\n| `--bg` | `#131924` | `#EDE7DC` | Room |\n| `--bg-2` | `#182031` | `#E4DCCC` | Drawer interior, rail |\n| `--sky` | `#0F1522` | `#F4F1E9` | Constellation field |\n| `--surface` | `#1F2938` | `#FCFAF5` | Cards, panels |\n| `--surface-2` | `#273346` | `#F1ECE0` | Raised, panel headers |\n| `--line` | `#2E3A4E` | `#D3C7B0` | Borders |\n| `--text` | `#E4EAF6` | `#1D2430` | Type |\n| `--dim` | `#93A2BC` | `#55627A` | Secondary type |\n| `--mute` | `#65748E` | `#7C8AA3` | Tertiary type |\n| `--brass` | `#C9A063` | `#8A6D40` | Anything physical |\n| `--brass-dk` | `#8A6D40` | `#6B5330` | Anything physical, shaded |\n| `--on-brass` | `#1A1408` | `#FDF8EE` | Ink sitting on brass |\n| `--accent` | `#5B8CFF` | `#1F5FD8` | Brand blue |\n| `--cyan` | `#4FD1E3` | `#0F8FA6` | X-ray |\n| `--amber` | `#F5B93F` | `#B47908` | Cryo-EM, ligand halo |\n| `--violet` | `#9B7BE8` | `#6B47C4` | NMR |\n| `--oxide` | `#E0685C` | `#B8382C` | FILED stamp, warnings, disorder flags |\n| `--mint` | `#43C79F` | `#1B8763` | Neutron, pass states |\n\n## \ud83c\udf10 Deployment\n\nMirrors the AlphaFraud pattern exactly.\n\n```bash\nbash deploy/deploy.sh                                    # from your Mac, after the first provision\nsudo SERVER_NAME=codswallop.mdeller.com bash /opt/codswallop/deploy/provision.sh   # once, on the droplet\n```\n\n`provision.sh` is idempotent: system packages, service user, venv, systemd unit, nginx site, Let's Encrypt certificate, and the HTTP/2 patch that certbot does not apply on nginx 1.24.\n\n**The droplet builds every artefact itself.** There is no second machine and nothing to push: `deploy/worker.sh` drains the request queue on a fifteen-minute timer, one family at a time, `nice`d so a reader waiting on a page always wins. Compute dependencies are in `requirements-compute.txt` and install as prebuilt wheels, so the box needs no compiler.\n\n```bash\nsystemctl status codswallop-worker.timer        # is it armed\nsystemctl start codswallop-worker.service       # force a pass now\ntail -f /var/log/codswallop-worker.log          # what it is doing\n```\n\n## \u2705 To Do\n\nRoadmap for CODSWALLOP, in dependency order: each phase is independently shippable, and the divider rail carries every phase's section from day one so navigation never needs rebuilding. Suggestions welcome.\n\n### Phase 1: spine (shipped)", "# Drug Discovery in the Cloud: How Decentralized Collaboration is Accelerating Research\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-14 on marcdeller.com (categories: blog).*\n\n* Cloud computing and AI platforms are breaking down traditional silos in drug discovery, enabling decentralized, global collaboration.\n* Researchers can now access high-performance computing and collaborative software on demand, lowering the barrier to entry for innovative research.\n* This new, decentralized model fosters agility and efficiency, allowing global teams to work together seamlessly to accelerate the development of new medicines.\n\n1. **The Disruptive Impact of Structural Biology on Biopharmaceutical Innovation:**\n   <https://www.pharmasalmanac.com/articles/the-disruptive-impact-of-structural-biology-on-biopharmaceutical-innovation>\n2. **2025 Trends in Biotech and Life Sciences Research:**\n   <https://go.zageno.com/blog/2025-trends-in-biotech-and-life-sciences-research>\n3. **AI in Drug Discovery: Predictions:**\n   <https://lizard.bio/knowledge-hub/2025-ai-in-drug-discovery-predictions>\n\nModern drug discovery is a team sport that requires a massive diversity of expertise\u2014from structural biology and medicinal chemistry to computational science and clinical research. Historically, this collaboration has been siloed within the walls of large pharmaceutical companies or academic institutions. The immense cost of the required infrastructure, from supercomputers to cryo-electron microscopes, created a high barrier to entry and limited the scope of collaboration.\n\nThe rise of cloud computing and AI-driven platforms is democratizing and decentralizing this process. Researchers can now access virtually unlimited computational power on demand through platforms like AWS and Google Cloud, allowing them to run complex molecular dynamics simulations or train large AI models without owning a supercomputer. Similarly, specialized data platforms and electronic lab notebooks (ELNs) enable seamless, real-time collaboration between scientists located anywhere in the world.\n\nThis new ecosystem allows for the creation of \u201cvirtual\u201d biotech companies and global research consortia that are more agile and capital-efficient than traditional models. A structural biologist in one country can analyze cryo-EM data generated by a CRO in another, while a computational chemist on a third continent uses that structure to design new molecules, all working together on a shared cloud platform. This breaks down geographical and institutional barriers, enabling the best minds to work together on the most challenging problems and dramatically accelerating the pace of translational science.", "# Seeing Micro and Nanoplastics in Living Tissues\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-19 on marcdeller.com (categories: microplastics).*\n\nPulsed laser light is absorbed by embedded plastic particles, which expand and emit ultrasound that an external detector array localises.\n\n* [In Vivo Microplastic Detection With Photoacoustic Imaging: Bear et al., Advanced Science, 2026](https://doi.org/10.1002/advs.202512152)\n\n* **Non-invasive imaging:** Photoacoustics maps microplastics in living tissue without dissection, biopsy or destructive sample preparation.\n* **Label-free signal:** Common polymers absorb laser light directly, removing the need for tags that could alter particle behaviour.\n* **Longer timescales:** Particles can be tracked over months, matching the chronic, low-dose nature of real human exposure.\n* **Clinical hurdles remain:** Reaching the brain and blood vessels in people will require deeper-penetrating systems and validation.\n\n1. **In Vivo Microplastic Detection With Photoacoustic Imaging:** [Bear et al., Advanced Science, 2026](https://doi.org/10.1002/advs.202512152)\n2. **Bioaccumulation of Microplastics in Decedent Human Brains:** [Nihart et al., Nature Medicine, 2025](https://www.nature.com/articles/s41591-024-03453-1)\n3. **Microplastics and Nanoplastics in Atheromas and Cardiovascular Events:** [Marfella et al., New England Journal of Medicine, 2024](https://www.nejm.org/doi/10.1056/NEJMoa2309822)\n4. **Discovery and Quantification of Plastic Particle Pollution in Human Blood:** [Leslie et al., Environment International, 2022](https://www.sciencedirect.com/science/article/pii/S0160412022001258)\n5. **Detection of Microplastics in Human Tissues and Organs: A Scoping Review:** [Roslan et al., Journal of Global Health, 2024](https://jogh.org/2024/jogh-14-04179)\n\n**A photoacoustic method maps microplastics inside the body without cutting it open**\n\nMicroplastics are no longer only an environmental concern. They were [first quantified in human blood](https://www.sciencedirect.com/science/article/pii/S0160412022001258) in 2022, then found embedded in [carotid artery plaque, where their presence tracked with a higher rate of heart attack, stroke and death](https://www.nejm.org/doi/10.1056/NEJMoa2309822), and more recently shown to be [accumulating in the brain, with the heaviest burdens in samples from people who had dementia](https://www.nature.com/articles/s41591-024-03453-1). The emerging picture is of particles that cross biological barriers and settle in tissue. Yet almost everything we know about where they end up has come from autopsy or biopsy, because the analytical toolkit has been fundamentally destructive.\n\nA [scoping review of detection methods](https://jogh.org/2024/jogh-14-04179) states the problem plainly. Raman and infrared microspectroscopy, pyrolysis gas chromatography mass spectrometry, and electron microscopy can each identify polymers, but each needs the tissue removed, digested or sectioned first, and each carries a real risk of contamination from airborne plastic during handling. None of them can watch a particle move.\n\n> **\u201cMicroplastics are posing an escalating threat to both ecological systems and human health\u201d** \u2013 [Bear et al., Advanced Science, 2026](https://doi.org/10.1002/advs.202512152)\n\nA team from Kingston University, UCL and Birmingham has now [imaged microplastics inside living tissue using photoacoustics](https://doi.org/10.1002/advs.202512152). The method exploits the fact that common polymers such as polyethylene and polypropylene absorb pulsed laser light and then expand, releasing ultrasound that detectors can localise. Because the signal comes from the plastic\u2019s own optical fingerprint, there is no need for fluorescent labelling, which can change how particles behave. Working in mice, the group resolved individual particles roughly the width of a human hair and followed their distribution over months rather than days.\n\nThat shift in timescale matters. Human exposure is chronic and low-dose, so a method that tracks accumulation over months sits much closer to the real question than a single destructive snapshot. The photoacoustic signal scales with particle quantity, which opens a route to direct quantification rather than indirect estimation. Translating any of this to people will need deeper-penetrating clinical systems and careful validation, and the brain and vasculature, the very sites where accumulation looks most consequential, remain the hardest to reach. Even so, for the first time there is a way to look without first taking something apart.", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n```sh\npython3 BoltzMaker.py new      [boltz_input.md]  # write a new campaign by answering plain questions\npython3 BoltzMaker.py format   boltz_input.md    # auto-align comments/blank-lines (cosmetic only)\npython3 BoltzMaker.py generate boltz_input.md    # write the target YAMLs + manifest\npython3 BoltzMaker.py preflight boltz_input.md   # environment + input sanity checks\npython3 BoltzMaker.py run      boltz_input.md    # boltz predict, live progress, resumable\npython3 BoltzMaker.py analyze  boltz_input.md    # CSV / XLSX / HTML dashboard\npython3 BoltzMaker.py all      boltz_input.md    # generate -> preflight -> run -> analyze\npython3 BoltzMaker.py boltz_input.md             # same as `all` (subcommand is optional)\npython3 BoltzMaker.py compare-sse boltz_input.md # apo-vs-holo secondary-structure motif shifts (see below)\n```\n\n`new` interviews you (proteins, partners, ligands, and the three constraint sentence types)\nand writes the file, refusing to overwrite an existing one without asking. It covers the common\ncase only; rarer fields (modifications, cyclic, MSA override, templates, homo-oligomer copies)\nare left for hand-editing. If `setup-plip` has been run it also asks whether each protein has a\nreference structure with a ligand already bound (a co-crystal or homology model); if so it runs\ncif2plip on it, lets you pick the relevant ligand where more than one is detected, and suggests\nthe contacted residues as `Pocket contact:` constraints, remapped onto your target's own\nnumbering via sequence alignment (BLOSUM62 + affine gaps).\n\n`format` re-aligns trailing comments to a clean column and normalises blank-line spacing\naround section/record boundaries -- purely cosmetic (it validates the file parses first, and\nnever changes meaning). Pass `--check` to report whether reformatting is needed without\nwriting anything (exit 1 if so, e.g. for a pre-commit check).\n\nAny field BoltzMaker doesn't recognise (a typo like `Predict afinity:`) prints a `WARNING`\nnaming the block, its name and the line number, and is otherwise silently dropped, so a\nmisspelled field never just vanishes without a trace.\n\n**Common options:**\n\n| Option | Default | Description |\n|---|---|---|\n| `--output-dir` | `settings.output_dir` | Override where generated YAMLs are written |\n| `--out-dir` | `./boltz_output` | Boltz's own `--out_dir`, next to the md file |\n| `--accelerator` | `auto` | `auto` / `gpu` / `cpu` (cpu works but is really only for checking a campaign runs at all) |\n| `--limit N` | none | Cap how many pending targets `run` submits (smoke test before a full batch) |\n| `--max-retries` | `2` | Auto-retry a target that doesn't complete (e.g. an OOM), isolating to one target at a time -- see \"Memory on Mac\" below (`0` disables) |\n| `--diffusion-samples` | Boltz default | Structure samples per target. Each extra sample costs roughly its own share of diffusion time, and analysis only ever reads the first (`model_0`) -- raise it to inspect pose variability yourself, not to improve the report |\n| `--no-potentials` | off | Turn off Boltz's FK steering and physical-guidance coordinate update, which was hardcoded on. That guidance update is one of the places a diffusion trajectory can diverge to NaN |\n| `--strict` | off | Promote preflight WARN to FAIL |\n| `--skip-interactions` | off | Skip cif2plip interaction analysis during `analyze`, even if `setup-plip` has been run |\n| `--skip-sse` | off | Skip compare-sse apo-vs-holo analysis during `analyze`, even if a family has `Apo structure:` set |\n| `--json` | off | `preflight` only: emit the check results as a JSON array on stdout instead of the rich table, for scripting and tooling (this is what the hosted web app consumes). The banner is suppressed alongside it, exactly as it is for `format`, so stdout stays parseable |\n\n**Memory-control options** (see the next section for why these matter on unified-memory\nhardware):\n\n| Option | Default | Description |\n|---|---|---|\n| `--workers` | `0` | Boltz's own default is 2, but each worker duplicates large in-memory structures out of the same pool the model is using |\n| `--mps-watermark` | `1.0` | `PYTORCH_MPS_HIGH_WATERMARK_RATIO` cap |\n| `--max-parallel-samples` | `1` | Boltz `--max_parallel_samples` |\n| `--recycling-steps` | Boltz default | Passthrough |\n| `--sampling-steps` | Boltz default | Passthrough |\n| `--diffusion-samples-affinity` | Boltz default | Passthrough |\n| `--sampling-steps-affinity` | Boltz default | Passthrough |\n| `--max-msa-seqs` | `4096` | Boltz's own default is 8192; halving the co-evolution feature block is one of the few levers that measurably cuts peak memory. Pass `8192` to restore Boltz's default |\n| `--memory-warn-tokens` | `1500` | Preflight size-heuristic WARN threshold |\n\n`run` is idempotent: targets with a complete prediction (cif + confidence json, and an\naffinity json if `predict_affinity` is on) are skipped on re-run, so an interrupted batch can\njust be re-run as-is.\n\n**`compare-sse` options** (see **compare-sse** below for what it does; the standalone command is\nfor re-running just this analysis, e.g. after adding an apo structure without re-running\n`boltz predict`):", "# Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\nAuthors: Kong, L.; Lee, J.H.; Doores, K.J.; Murin, C.D.; Julien, J.-P.; McBride, R.; Liu, Y.; Marozsan, A.; Cupo, A.; Klasse, P.-J.; Hoffenberg, S.; Caulfield, M.; King, C.R.; Hua, Y.; Le, K.M.; Khayat, R.; Deller, M.C.; Clayton, T.; Tien, H.; Feizi, T.; Sanders, R.W.; Paulson, J.C.; Moore, J.P.; Stanfield, R.L.; Burton, D.R.; Ward, A.B.; Wilson, I.A.\nSource: Nature Structural & Molecular Biology, 2013\n\n| | PGT 135 | PGT 135-gp120-CD4-17b |\n|---|---|---|\n| **Data collection** | | |\n| Space group | P2\u20812\u20812 | C2 |\n| Cell dimensions a, b, c (\u00c5) | 85.89, 138.15, 42.34 | 218.41, 92.15, 88.19 |\n| \u03b1, \u03b2, \u03b3 (\u00b0) | 90.00, 90.00, 90.00 | 90.00, 104.75, 90.00 |\n| Resolution (\u00c5) | 31.4\u20131.75 (1.78\u20131.75) | 43.58\u20133.10 (3.20\u20133.10) |\n| Rsym | 0.06 (0.47) | 0.06 (0.60) |\n| I/\u03c3I | 27.6 (1.9) | 16.9 (1.72) |\n| Completeness (%) | 96.1 (73.8) | 99.7 (98.8) |\n| Redundancy | 6.5 (4.0) | 3.7 (3.3) |\n| **Refinement** | | |\n| Resolution (\u00c5) | 31.40\u20131.75 | 43.58\u20133.10 |\n| No. reflections | 49,464 | 30,850 |\n| Rwork/Rfree | 20.7/24.5 | 24.6/36.7 |\n| No. atoms \u2014 Protein | 3475 | 10,560 |\n| No. atoms \u2014 Glycan/Ligand/ion | \u2014 | 347 |\n| No. atoms \u2014 Water | 273 | 1 |\n| B factors \u2014 All proteins | 41 | 127 |\n| B factors \u2014 gp120 | \u2014 | 117 |\n| B factors \u2014 17b Fab | \u2014 | 129 |\n| B factors \u2014 CD4 D1D2 | \u2014 | 149 |\n| B factors \u2014 PGT 135 constant domain | 43 | 151 |\n| B factors \u2014 PGT 135 variable domain | 39 | 99 |\n| B factors \u2014 Glycans | \u2014 | 112 |\n| B factors \u2014 Ligands | \u2014 | 109 |\n| B factors \u2014 Water | 46 | 76 |\n| r.m.s. deviations \u2014 Bond lengths (\u00c5) | 0.006 | 0.012 |\n| r.m.s. deviations \u2014 Bond angles (\u00b0) | 1.11 | 1.46 |\n\nOne crystal for each structure was used for data collection and structure determination. Values in parentheses are for the highest-resolution shell. For PGT 135 Fab, average B-values for the constant (con) and variable (var) domains were calculated.\n\n## Figure captions\n\n*Figure 1.* Crystal structure of PGT 135 in complex with HIV-1 gp120. (a) The glycan-dependent interaction between PGT 135 and gp120 is presented in ribbon representation in the context of the gp120 ternary complex that contains a 2-domain CD4 and the Fab of CD4-induced antibody 17b. An enlarged display of the PGT 135-gp120 interaction is shown to the lower right. The CDR H1 and H3 loops (blue and red respectively) penetrate deeply through the glycan canopy to contact protein surface below while the light chain contacts the glycans using CDR loops and framework regions. (b) Buried surface on gp120 protein surface and attached glycans N332, N386 and N392 by CDR loops and framework regions of PGT 135. The values are in \u00c52.\n\n*Figure 2.* PGT 135 neutralization and binding. (a) Neutralization activity of PGT 135 and 2G12 were assayed against JR-FL pseudovirus in TZM-bl cells. Neutralization of WT JR-FL and of JR-FL containing V1-V2 loop mutations E168K and N188A are shown. (b) Binding of PGT 135 IgG mutants to gp120 from various strains (JRCSF, JRFL) as well as JRFLmV3 were measured by ELISA. Mutations include deletion of the CDR H1 insert \"EWGDK\", alanine substitution of the CDR H2 arginine that supports the CDR H1 insert with a salt bridge, and mutation of VFMLV to AAMAA in CDR H3. Wild-type PGT 135 and 2G12 IgGs are included as controls. (c) Neutralization potency of PGT 135 mutant IgGs against JRCSF, SF162 and 92RW020. Error bars represent the standard deviation. Experiments were performed in duplicate and repeated three times.\n\n*Figure 3.* Glycan dependency of PGT 135 binding to gp120. (a) PGT 135 interactions with gp120 glycans. The 2Fo-Fc electron density (contoured at 1\u03c3) of N392, N332 and N386 glycans are shown along with contacting residues on PGT 135. A schematic of each glycan is shown beside the structures for the sugars that are visible in the electron density with GlcNAc as dark blue squares and mannose residues as green circles. The surface area (\u00c52) on each glycan moiety buried by PGT 135 is shown. (b) Interactions of PGT 135, PGT 128 and 2G12 with oligomannose glycans on a high-density microarray. The error bars represent standard deviations where n = 6. (c) Glycan dependency of PGT 135 neutralization using BaL pseudoviruses produced in different cell systems. Virus produced in 293T cells contains complex, hybrid and oligomannose type glycans, whereas in 293S cells only Man5-9GlcNAc2 glycans are present. Addition of kifunensine yields only Man9GlcNAc2 glycans. NB-DNJ blocks the glycan processing at neutral glucosylated glycans. Error bars represent the standard deviation. Each experiment was performed in duplicate and repeated at least three times. (d) The effect on PGT 135 neutralization from single mutations removing glycans and His330 for several HIV-1 strains. Values are presented as fold change in IC50 of variant envelope compared to WT envelope (fold change = IC50 variant/IC50 WT). Boxes are color coded for fold change; red: >100, orange 10-100, yellow 4-9 and green <4. n.d.: not determined. Values marked \"a\" were determined by ELISA instead of neutralization.\n\n*Figure 4.* A gp140 trimer binds Fab PGT 135 in slightly different orientations in contrast to PGT 128, which is bound in a single orientation. (a) The PGT 135-gp120 structure is shown fitted into the EM density map of PGT 135 in complex with the BG505 SOSIP.664 gp140 trimer in top and side views of the trimer. (b) 2D difference map of the two 2D class averages. The PGT 128-SOSIP class average was subtracted from the PGT 135-SOSIP class average. The overlap is shown as the black gap in the trimer region. (c) Top view of the 2D class averages of the BG505 SOSIP.664 gp140 trimer in complex with PGT 135 (left) and PGT 128 (right). The images contributing to the class averages were aligned using a top view back projection of an undecorated gp140 SOSIP.664 trimer. On the top, the Fabs are colored cyan and yellow for PGT 135 and PGT 128, respectively. Below are variance maps of the trimer in complex with PGT 135 (left) and PGT 128 (right) indicating the variability in the images that went into the 2D class averages. Brighter pixel values correspond to a greater degree of variability in that region. (d) An overlay of the 2D class average (blue outline) and the first eigenvector (orange) of the 2D PCA analysis for PGT 135 and PGT 128.", "# Smiles2YAML \u2013 From a JSON of candidate SMILES to a folder of Boltz-2 input YAML\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-14 on marcdeller.com (categories: vibe-code).*\n\n> **From a JSON of candidate SMILES to a folder of Boltz-2 input YAMLs, in one prompt.**\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/generate_yaml](https://github.com/bellcheddar/generate_yaml) |\n\n---\n\nA command-line utility for generating Boltz-2 / co-folding input YAML files from a batch of candidate SMILES stored in a JSON file. Designed to work with the output format produced by generative chemistry runs (e.g. fragment-based screening, de novo design) and feed directly into a Boltz-style ModelCIF prediction pipeline.\n\nWhy it matters: generative chemistry runs spit out hundreds of candidate SMILES, but Boltz-2 wants one carefully structured YAML per compound, each pairing the ligand with a protein chain, a pocket constraint, and an affinity block. Hand-writing those is tedious and error-prone. generate_yaml.py reads the candidates straight from your JSON, prompts once for the protein context, and writes a clean, submission-ready YAML for every selected compound. It is useful for anyone bridging de novo design or fragment screening into a co-folding affinity pipeline: turn a screening output into a ready-to-run prediction batch in seconds, with stdlib-only portability and no dependencies to install.\n\n---\n\n## Overview\n\nGiven a JSON file containing a dictionary of candidate SMILES under a `\"candidates\"` key, `generate_yaml.py` interactively prompts for protein context and contact constraints, then writes one YAML file per selected candidate \u2014 ready to submit to Boltz-2 (or a compatible co-folding engine) for structure prediction with binding affinity estimation.\n\nEach output YAML encodes:\n\u2013 A single protein chain with user-supplied ID and sequence\n\u2013 A ligand defined by cleaned SMILES\n\u2013 A pocket constraint anchoring the ligand to a specified contact residue\n\u2013 An affinity property block targeting the ligand\n\n---\n\n## Requirements\n\n* Python 3.8+\n* Standard library only (`argparse`, `json`, `re`, `pathlib`) \u2014 no additional dependencies\n\n---\n\n## Installation\n\n```\ngit clone https://github.com/<your-org>/generate_yaml.git\ncd generate_yaml\n# No install step required \u2014 pure stdlib\n```\n\n---\n\n## Usage\n\n```\npython generate_yaml.py <json_file>\n```\n\n### Example\n\n```\npython generate_yaml.py scratch_Baricitinib_sim.out.20260530_0219.json\n```\n\nThe script will interactively prompt for:\n\n| Prompt | Description | Example |\n| --- | --- | --- |\n| `Protein ID` | Chain identifier used in the YAML | `A` |\n| `Protein sequence` | Full single-letter amino acid sequence | `MGHHHHHHSSGVD...` |\n| `Contact residue number` | Residue number defining the binding pocket | `1007` |\n| `Output directory` | Path where YAML files will be written | `~/boltz_runs/baricitinib/` |\n| `Candidate selection` | Which candidates to export (see below) | `1-25` or `all` |\n\n---\n\n## Candidate Selection Syntax\n\n| Format | Description | Example |\n| --- | --- | --- |\n| `all` | Export every candidate | `all` |\n| `N` | A single candidate by 1-based index | `7` |\n| `N-M` | An inclusive range | `1-25` |\n| `N, M, ...` | Arbitrary comma-separated list | `1, 7, 17, 28, 31` |\n\nCandidate numbering is **1-based** and follows the original key order in the JSON `\"candidates\"` dictionary.\n\n---\n\n## Output\n\nFiles are written to the specified output directory, named sequentially by candidate index:\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/generate_yaml](https://github.com/bellcheddar/generate_yaml) |", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\n\"Heterocycloalkyl\" refers to monocyclic or polycyclic heterocycles having at least one non-aromatic ring (saturated or partially unsaturated ring), wherein one or more of the ring-forming carbon atoms of the heterocycloalkyl is replaced by a heteroatom selected from N, O, or S, and wherein the ring-forming carbon atoms and heteroatoms of the heterocycloalkyl group can be optionally substituted by one or more oxo or sulfido (e.g., C(O), S(O), C(S), or S(O)2, etc.). Heterocycloalkyl groups include monocyclic and polycyclic (e.g., having 2 fused rings) systems. Included in heterocycloalkyl are monocyclic and polycyclic 4\u201310-, 4\u20137-, and 5\u20136-membered heterocycloalkyl groups. Heterocycloalkyl groups can also include spirocycles and bridged rings. The heterocycloalkyl group can be attached through a ring-forming carbon atom or a ring-forming heteroatom. In some embodiments, the heterocycloalkyl group contains 0 to 3 double bonds; in some embodiments 0 to 2 double bonds. Also included in the definition of heterocycloalkyl are moieties that have one or more aromatic rings fused (i.e., having a bond in common with) to the non-aromatic heterocyclic ring, for example, benzo or thienyl derivatives of piperidine, morpholine, azepine, etc. A heterocycloalkyl group containing a fused aromatic ring can be attached through any ring-forming atom including a ring-forming atom of the fused aromatic ring. In some embodiments, the heterocycloalkyl group contains 4 to 10 ring-forming atoms, 4 to 7 ring-forming atoms, 4 to 6 ring-forming atoms or 5 to 6 ring-forming atoms; has 1 to 4 heteroatoms, 1 to 3 heteroatoms, or 1 to 2 heteroatoms or 1 heteroatom. Examples of heterocycloalkyl groups include pyrrolidin-2-one, 1,3-isoxazolidin-2-one, pyranyl, tetrahydropyran, oxetanyl, azetidinyl, morpholino, thiomorpholino, piperazinyl, tetrahydrofuranyl, tetrahydrothienyl, piperidinyl, pyrrolidinyl, isoxazolidinyl, isothiazolidinyl, pyrazolidinyl, oxazolidinyl, thiazolidinyl, imidazolidinyl, azepanyl, benzazapene, 1,2,3,4-tetrahydroisoquinoline, azabicyclo[3.1.0]hexanyl, diazabicyclo[3.1.0]hexanyl, oxabicyclo[2.1.1]hexanyl, azabicyclo[2.2.1]heptanyl, azabicyclo[2.2.1]heptan-7-yl, azabicyclo[2.2.1]heptan-2-yl, diazabicyclo[2.2.1]heptanyl, azabicyclo[3.2.1]octanyl, diazabicyclo[3.2.1]octanyl, oxabicyclo[2.2.2]octanyl, azabicyclo[2.2.2]octanyl, azaadamantanyl, diazaadamantanyl, oxa-adamantanyl, azaspiro[3.3]heptanyl, diazaspiro[3.3]heptanyl, oxa-azaspiro[3.3]heptanyl, azaspiro[3.4]octanyl, diazaspiro[3.4]octanyl, oxa-azaspiro[3.4]octanyl, azaspiro[2.5]octanyl, diazaspiro[2.5]octanyl, azaspiro[4.4]nonanyl, diazaspiro[4.4]nonanyl, oxa-azaspiro[4.4]nonanyl, azaspiro[4.5]decanyl, diazaspiro[4.5]decanyl, oxa-diazaspiro[4.4]nonanyl, and the like.\n\n\"Co-p cycloalkyl-Cn-m alkyl-\" refers to a group of formula cycloalkyl-alkylene-, wherein the cycloalkyl has o to p carbon atoms and the alkylene linking group has n to m carbon atoms. \"Co-p aryl-Cn-m alkyl-\" refers to a group of formula aryl-alkylene-, wherein the aryl has o to p carbon ring members and the alkylene linking group has n to m carbon atoms. \"Heteroaryl-Cn-m alkyl-\" refers to a group of formula heteroaryl-alkylene-, wherein the alkylene linking group has n to m carbon atoms. \"Heterocycloalkyl-Cn-m alkyl-\" refers to a group of formula heterocycloalkyl-alkylene-, wherein alkylene linking group has n to m carbon atoms.\n\n\"Alkylene\" refers to a divalent straight chain or branched alkyl linking group. Examples of \"alkylene groups\" include methylene, ethan-1,1-diyl, ethan-1,2-diyl, propan-1,3-diyl, propan-1,2-diyl, propan-1,1-diyl and the like. \"Alkenylene\" refers to a divalent straight chain or branched alkenyl linking group; examples include ethen-1,1-diyl, ethen-1,2-diyl, propen-1,3-diyl, 2-buten-1,4-diyl, 3-penten-1,5-diyl, 3-hexen-1,6-diyl, 3-hexen-1,5-diyl, and the like. \"Alkynylene\" refers to a divalent straight chain or branched alkynyl linking group; examples include propyn-1,3-diyl, 2-butyn-1,4-diyl, 2-pentyn-1,5-diyl, 3-hexyn-1,6-diyl, 3-hexyn-1,5-diyl, and the like. \"Alkyl linking group\" is a bivalent straight chain or branched alkyl linking group (\"alkylene group\").\n\n\"Oxo\" refers to an oxygen atom (i.e., =O) as a divalent substituent, forming a carbonyl group when attached to a carbon (e.g., C=O or C(O)), or attached to a nitrogen or sulfur heteroatom forming a nitroso, sulfinyl or sulfonyl group.\n\nAt certain places, the definitions or embodiments refer to specific rings (e.g., an azetidine ring, a pyridine ring, etc.). Unless otherwise indicated, these rings can be attached to any ring member provided that the valency of the atom is not exceeded. For example, an azetidine ring may be attached at any position of the ring, whereas an azetidin-3-yl ring is attached at the 3-position.\n\nThe compounds described herein can be asymmetric (e.g., having one or more stereocenters). All stereoisomers, such as enantiomers and diastereomers, are intended unless otherwise indicated. Compounds of the present disclosure that contain asymmetrically substituted carbon atoms can be isolated in optically active or racemic forms. Methods on how to prepare optically active forms from optically inactive starting materials are known in the art, such as by resolution of racemic mixtures or by stereoselective synthesis. Many geometric isomers of olefins, C=N double bonds, and the like can also be present in the compounds described herein, and all such stable isomers are contemplated in the present invention. Cis and trans geometric isomers of the compounds of the present disclosure are described and may be isolated as a mixture of isomers or as separated isomeric forms. In some embodiments, the compound has the (R)-configuration. In some embodiments, the compound has the (S)-configuration. The Formulas (e.g., Formula (I), (II), etc.) provided herein include stereoisomers of the compounds.", "# AlphaFraud \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n| Field | Detail |\n| --- | --- |\n| **Take-home** | A protein with no single fold; AlphaFold\u2019s confidence is actively *anti*-correlated with accuracy |\n| **AlphaFold vs experimental** | Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \u201cghost\u201d overlay |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-9A1Q_1)** |\n| **KPIs** | TM 0.138 \u00b7 lDDT 0.177 \u00b7 pLDDT 75.2 \u00b7 C\u03b1-RMSD 36.3 \u00c5 \u00b7 FRAUD 0.71 \u00b7 novelty 0% |\n| **Stats** | **210** depositions, **206** confidently wrong \u2014 the archive\u2019s most-repeated failure \u00b7 pLDDT\u2194lDDT \u22120.29 |\n| **Structure** | 140-residue intrinsically disordered monomer; the deposited forms are cryo-EM cross-\u03b2 fibrils |\n| **Why AlphaFold gets it wrong** | No native fold to predict, and polymorph-specific packing is not in the sequence \u2014 where it is more confident it is measurably more wrong |\n| **Biology & disease** | Presynaptic protein; fibrils define Parkinson\u2019s, DLB and multiple system atrophy; PD mutations A30P/E46K/H50Q/G51D/A53T |\n| **Key fact** | 210 depositions, 206 confidently wrong \u2014 the most-deposited catch on the site |\n\n### \ud83d\udd34 Transthyretin (TTR) \u00b7 [`9BZS_1`](https://alphafraud.mdeller.com/examples#ex-9BZS_1)\n\n| Field | Detail |\n| --- | --- |\n| **Take-home** | pLDDT 98 on the native tetramer fold; the deposited cardiac amyloid fibril scores TM 0.21 \u2014 one of the starkest confidently-wrong cases |\n| **AlphaFold vs experimental** | Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \u201cghost\u201d overlay |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-9BZS_1)** |\n| **KPIs** | TM 0.214 \u00b7 lDDT 0.489 \u00b7 pLDDT 98.0 \u00b7 C\u03b1-RMSD 22.8 \u00c5 \u00b7 FRAUD **0.95** \u00b7 novelty 0% |\n| **Stats** | 160 depositions, 34 confidently wrong \u00b7 pLDDT\u2194lDDT \u22120.16 |\n| **Structure** | Native 127-aa \u03b2-sandwich homotetramer with thyroxine sites; the amyloid form is a cross-\u03b2 fibril from cleaved segments |\n| **Why AlphaFold gets it wrong** | A wrong-*state* failure: pLDDT measures confidence in the known native fold, not whether a deposition *is* that state; TTR is 0% novel, squarely in training |\n| **Biology & disease** | Thyroxine/retinol transporter; ATTR amyloidosis (wild-type cardiac + hereditary V30M); uniquely druggable (tafamidis, patisiran) |\n| **Key fact** | FRAUD 0.95 \u2014 near the very top of the whole site |\n\n### \ud83d\udd34 Coagulation factor X (F10) \u00b7 [`6Q9F_2`](https://alphafraud.mdeller.com/examples#ex-6Q9F_2)\n\n| Field | Detail |\n| --- | --- |\n| **Take-home** | Not a refold but a wrong-*context* capture \u2014 a 39-residue EGF1 fragment threaded through a partner enzyme\u2019s active site |\n| **AlphaFold vs experimental** | Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \u201cghost\u201d overlay |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-6Q9F_2)** |\n| **KPIs** | TM 0.127 \u00b7 lDDT 0.468 \u00b7 pLDDT 93.5 \u00b7 C\u03b1-RMSD 6.6 \u00c5 \u00b7 FRAUD 0.37 \u00b7 novelty 0% |\n| **Stats** | 47 depositions, 26 confidently wrong \u00b7 pLDDT\u2194lDDT \u22120.59 |\n| **Structure** | Modular zymogen (Gla\u2013EGF1\u2013EGF2\u2013protease); the worst deposition is a linear EGF1 peptide in AspH\u2019s channel, not a folded module |\n| **Why AlphaFold gets it wrong** | A context failure, not aggregation \u2014 the RMSD is modest (6.6 \u00c5) but TM collapses because a linear fragment has no compact fold to superpose |\n| **Biology & disease** | Central coagulation zymogen; activated factor Xa is the target of rivaroxaban, apixaban and edoxaban |\n| **Key fact** | The non-amyloid contrast \u2014 the error is missing biological context (fragment / complex / PTM), not misfolding |\n\n## \ud83e\uddea Calculate\n\nThe **[Calculate](https://alphafraud.mdeller.com/calculate)** tab runs the whole pipeline on demand for any human PDB ID you type and returns the same panel format as the auto examples \u2014 KPIs, the C\u03b1-deviation ribbon, the interactive 3D viewer and a deterministic write-up \u2014 with **no AI-generated text**.\n\n* **Strict autocomplete.** The input only completes to entries that (1) exist, (2) are human, and (3) have an AlphaFold model. That comes from a precomputed index \u2014 seeded for free from the corpus (the ~83k entities already carrying a confirmed AlphaFold model), back-filled with a one-off pre-cutoff human sweep, and refreshed automatically by every weekly run. Press Tab to complete.\n* **Any human structure, blind or not.** Post-cutoff entries are genuine blind tests; pre-cutoff entries compute too, with a note that AlphaFold may have seen them in training (so it is not a blind prediction).\n* **Clear reasons on rejection.** If an entry doesn\u2019t qualify you are told exactly why: not a human protein structure, no single UniProt mapping (antibody / fusion / chimera / synthetic), or no AlphaFold model for the sequence.\n* **Right or wrong, honestly.** A good prediction renders the green \u201cAlphaFold gets it right\u201d panel (like the positive control); a confidently-wrong one renders the failure panel \u2014 the same deterministic builder (`examples_auto.build`) decides which from the metrics.\n* **Isolated and cached.** On-demand results are stored under a separate run and status, so they never enter the leaderboard, stats, analysis or archive; the heavy compute runs in an isolated subprocess; and a repeat request for the same structure is instant.\n\n## \ud83e\uddee Coverage and skipped entities\n\nNot every deposited entity can be compared. AlphaFold DB holds a single model per human UniProt sequence (monomers only), so anything that does not map cleanly to one accession, has no model covering its resolved range, or is too small or too large to compare is recorded as *skipped* rather than forced through.", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n| Panel | What it shows |\n|---|---|\n| **Campaign summary** | Field / Value / **Details**: a linked path to the input file, each protein/partner's id and sequence length, each ligand's id and SMILES-vs-CCD source, the full list of target stems, which ligands were flagged in ligand-chemistry review (linked to the card below), and a plain-English gloss for each of the more cryptic run parameters (accelerator, MPS watermark, recycling/sampling steps). Tracked across every `run` invocation in a small hidden sidecar file |\n| **Summary table** | One row per target, in named column bands (Identity, Confidence, Affinity, Interactions, Structure) -- details below |\n| **Ligand preparation** | The same stereocentre / protonation-state / disconnected-fragment checks as `preflight`'s `ligand_preparation`, shown per ligand rather than as one summary line |\n| **Ligand structures** | A paginated 5x5 grid of every ligand's rendered 2D structure, findings and shared scaffolds highlighted on the structures themselves, with a captioned legend and \"Download PDF\" / \"Download SMILES\" links side by side, matching the Summary table's download-links style. See **Ligand validation & scaffold highlighting** below |\n| **Charts** | An interactive [Plotly](https://plotly.com/javascript/) grid: ranked pIC50, ranked confidence, a \"pIC50 vs confidence score\" scatter, interaction counts by type, and a \"pIC50 vs binder probability\" scatter (binder probability on x, pIC50 on y). Hover/zoom/pan; plotly.js is vendored and inlined rather than CDN-loaded, so the dashboard has no runtime dependency on an external script host |\n| **Selectivity heatmap** | The family x ligand pIC50 pivot, mirroring the XLSX `selectivity` sheet |\n| **Interaction fingerprints** (PLIP) | A per-family residue-interaction heatmap (also interactive Plotly), shown for every family with interaction data even a single ligand -- though the similarity-based reordering that helps SAR ranking within a series only kicks in from 3+ ligands |\n| **Per-target binding site** (PLIP) | Its binding-site image (residues labelled and interaction distances shown -- PLIP's own images have neither, so these are re-rendered from its PyMOL session with both added, with a \"Download image\" link) next to an interactive, auto-rotating [3Dmol.js](https://3dmol.org) view of the same predicted structure (built from the mmCIF, ligand highlighted), side by side with a table of that target's contacts (own \"Download CSV\") plus a download link for the full PyMOL session |\n| **Secondary structure shifts** | The compare-sse card, embedded whole: family coverage, overall shift statistics, and the per-motif table with its Plotly charts. See **compare-sse** below |\n\n**Display names, not internal ids.** The \"Target\" column shows\n`{run}_{group}_{partners}_{ligand}_{pocket}` (e.g. `2_5HT2A_GNAQ+GNB1+GNG2_RISP_41Y`) rather than\nthe internal per-variant family id/stem (`H2ANG_RISP`, `H2AAP`), partners omitted when there are\nnone and `apo` in place of a ligand for a ligand-free target, which takes no pocket suffix. The\nsame name -- or its family-level `{group}_{partners}` form -- replaces the internal id in every\nchart tick, legend and point label, every per-target/per-family card title, the campaign-summary\ntarget list, the selectivity pivot's columns in both the dashboard heatmap and the XLSX sheet,\nand compare-sse. The raw ids stay alongside it in every underlying CSV/XLSX `targets` sheet, for\ncross-referencing against real output filenames.\n\nThe leading run number and trailing pocket code earn their place in a matrix campaign, where one\nligand is predicted several times and a run number alone reads as an index rather than a\ncondition: `2_..._ORFO_41Y` against `3_..._ORFO_V6G` says what was done differently, where\n`2_..._ORFO` and `3_..._ORFO` say only that there were two of them. The suffix comes from the\nsame value the \"Pocket\" column shows (the pocket code, or `Unc` when unconstrained), so a name\nand its column cannot drift apart.\n\n**The Summary table's own details.** Short human headers replace raw JSON field names, and\nredundant/granular columns (per-chain and per-chain-pair confidence breakdowns, individual\nensemble sub-model values) are hidden by regex pattern rather than a fixed list, so it scales\npast two chains. Two download links sit above it: the full underlying CSV, and one matching just\nthis trimmed/renamed view. A \"Partner\" column lists each target's co-folded partner chain(s)\n(hidden when there are none), and rows are grouped by `Group:`/family id with a blue top border\nmarking each new group.\n\n**The \"Flags\" column is now \"Summary\", and icon-based:** a bullseye (affinity) and a shield\n(confidence) per row, each tinted green/amber/red by tier, with the exact value and\ninterpretation on hover. The tiers reuse the existing `LOW_CONFIDENCE_THRESHOLD` and a symmetric\nbuffer around Boltz's documented 0.5 binder decision boundary, because Boltz's own docs define\nthese metrics' [0, 1] range but publish no official tri-colour bands. A `MISSING_OUTPUTS` failure\ncollapses the cell to a single red cross, and a legend right of the download links spells out all\nsix tier/icon combinations. It is always shown (previously hidden entirely when nothing was\nflagged), so a clean campaign reads as a row of green icons rather than a column that silently\ndisappears. A ligand-free (apo) target's ligand/affinity/interface/interaction columns, the\nbullseye included, show an explicit `N/A` rather than a blank cell or a misleading `0.00`.", "# Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3\nAuthors: Shvartsbart, A.; Roach, J.J.; Witten, M.R.; Koblish, H.; Harris, J.J.; Covington, M.; Hess, R.; Lin, L.; Frascella, M.; Truong, L.; Leffet, L.; Conlen, P.; Beshad, E.; Klabe, R.; Katiyar, K.; Kaldon, L.; Young-Sciame, R.; He, X.; Petusky, S.; Chen, K.-J.; Horsey, A.; Lei, H.-T.; Epling, L.B.; Deller, M.C.; Vechorkin, O.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2022\n\n**Step 2. 3-Bromo-6-chloroimidazo[1,2-b]pyridazin-8-amine (33).** 3,8-Dibromo-6-chloroimidazo[1,2-b]pyridazine (14.8 g, 47.5 mmol) was placed in a 350 mL heavy-walled pressure vessel with a stir bar. This sample was treated with aqueous ammonium hydroxide (111 mL, 2.85 mol), the chamber was sealed, and the heterogeneous mixture was warmed to 120 \u00b0C. After 4 h, the reaction was cooled to room temperature, carefully opened, and an LCMS aliquot indicated complete consumption of the starting material. The mixture was poured into 800 mL of ice water and stirred for an hour at room temperature. The mixture was filtered under vacuum, and the residue was dried with continued air flow to provide 3-bromo-6-chloroimidazo[1,2-b]pyridazin-8-amine (9.5 g, 38 mmol, 81%). The crude material was used without further purification. LCMS calculated for C6H5BrClN4 (M + H)+: m/z = 246.9, 248.9; found: 246.9, 248.9.\n\n**Step 3. 3-Bromo-6-chloro-7-iodoimidazo[1,2-b]pyridazin-8-amine (34).** 3-Bromo-6-chloroimidazo[1,2-b]pyridazin-8-amine (6.8 g, 28 mmol) was dissolved in DMF (69 mL) and treated with N-iodosuccinimide (10.5 g, 46.7 mmol). The solution was stirred at 60 \u00b0C. After 2 h, LCMS indicated complete consumption of the starting material. The reaction was poured into ice water (500 mL) and stirred for an hour at room temperature. The mixture was filtered under vacuum, and the residue was dried with continued air flow to provide the title compound (11.6 g, 31.1 mmol, quantitative yield assumed). The crude material was used without further purification. LCMS calculated for C6H4BrClIN4 (M + H)+: m/z = 372.8, 374.8; found: 372.8, 374.8.\n\n**Step 4. 3-Bromo-6-chloro-7-iodoimidazo[1,2-b]pyridazine (35).** 3-Bromo-6-chloro-7-iodoimidazo[1,2-b]pyridazin-8-amine (10.3 g, 27.5 mmol) was placed in a 350 mL heavy-walled pressure vessel and dissolved in THF (83 mL). The sample was treated with salicylic acid (0.76 g, 5.5 mmol) and tert-butyl nitrite (36.3 mL, 275 mmol), and the vessel was sealed. The reaction was warmed to 70 \u00b0C. After 2 h, LCMS indicated complete consumption of the starting material. The solution was cooled to room temperature and concentrated in vacuo. The residue was dissolved in EtOAc, adsorbed onto silica, and purified by flash chromatography in 0\u221250% EtOAc/hexanes to provide the title compound (7.95 g, 22.2 mmol, 81%). LCMS calculated for C6H3BrClIN3 (M + H)+: m/z = 357.8, 359.8; found: 357.8, 359.8.\n\n**Step 5. 3-Bromo-6-chloro-7-(1-methyl-1H-pyrazol-4-yl)imidazo[1,2-b]pyridazine (36a).** A sample of 3-bromo-6-chloro-7-iodoimidazo[1,2-b]pyridazine (1.4 g, 3.9 mmol) was dissolved in 1,4-dioxane (16 mL) and was treated with K2CO3 (1.08 g, 7.81 mmol), water (3.3 mL), and 1-methyl-4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazole (853 mg, 4.10 mmol). This solution was degassed with bubbling nitrogen for 5 min. Tetrakis(triphenylphosphine)palladium(0) (0.45 g, 0.39 mmol) was added, the vial was capped, and the solution was stirred at 100 \u00b0C. After 16 h, LCMS indicated \u223c85% conversion. The solution was cooled to room temperature, quenched with aqueous NH4Cl, and extracted with EtOAc, then with 25% iso-propanol in chloroform. The combined organic fractions were dried over MgSO4, filtered, and concentrated in vacuo. The crude sample was purified by flash chromatography in 0\u2212100% EtOAc/DCM and then in 0\u221210% MeOH/DCM to provide the title compound (7.95 g, 22.2 mmol, 81%). LCMS calculated for C10H8BrClN5 (M + H)+: m/z = 312.0, 314.0; found: 311.8, 313.8.\n\n**Step 6. (S)-3-Bromo-7-(1-methyl-1H-pyrazol-4-yl)-6-((tetrahydrofuran-3-yl)oxy)imidazo[1,2-b]pyridazine (37a).** A sample of 3-bromo-6-chloro-7-(1-methyl-1H-pyrazol-4-yl)imidazo[1,2-b]pyridazine (218 mg, 0.70 mmol) was dissolved in 1,4-dioxane (8.7 mL) and was treated with cesium carbonate (682 mg, 2.09 mmol) and (S)-tetrahydrofuran-3-ol (846 \u03bcL, 10.5 mmol). This solution was stirred at 110 \u00b0C. After 16 h, LCMS indicated complete consumption of the starting material. The solution was cooled to room temperature, quenched with aqueous NH4Cl, and the mixture was extracted with EtOAc. The combined organic fractions were dried over MgSO4, filtered, and concentrated in vacuo. The crude material was flashed in 0\u2212100% EtOAc/DCM and then in 0\u221210% MeOH/DCM to afford the title compound (188 mg, 74%). LCMS calculated for C14H15BrN5O2 (M + H)+: m/z = 364.0, 366.0; found: 364.2, 366.2.\n\n**Step 7. (S)-3,4-Difluoro-N-methyl-5-(7-(1-methyl-1H-pyrazol-4-yl)-6-((tetrahydrofuran-3-yl)oxy)imidazo[1,2-b]pyridazin-3-yl)benzamide (25).** (S)-3-Bromo-7-(1-methyl-1H-pyrazol-4-yl)-6-((tetrahydrofuran-3-yl)oxy)imidazo[1,2-b]pyridazine (34 mg, 0.09 mmol) was dissolved in DMF (1.5 mL) and was treated with K2CO3 (38 mg, 0.3 mmol), water (0.3 mL), and 3,4-difluoro-N-methyl-5-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)benzamide (83 mg, 28 \u03bcmol). This solution was degassed with bubbling nitrogen for 5 min. XPhos Pd G2 (11 mg, 14 \u03bcmol) was added, the vial was capped, and the solution was stirred at 85 \u00b0C. After 2 h, LCMS indicated complete consumption of the starting material. The solution was cooled to room temperature, diluted with MeCN and water, filtered through a thiol SiliaPrep cartridge, and purified by HPLC (pH = 2) to provide the title compound (1.3 mg, 3%). LCMS calculated for C22H21F2N6O3 (M + H)+: m/z = 455.2; found: 455.4. 1H NMR (600 MHz, DMSO-d6) \u03b4 8.76 (d, J = 5.9 Hz, 1H), 8.60 (q, J = 4.7 Hz, 1H), 8.46 (s, 1H), 8.28 (s, 1H), 8.17 (s, 1H), 8.12 (s, 1H), 7.91 (m, 1H), 5.61 (m, 1H), 4.08 (d, J = 10.7 Hz, 1H), 4.04 (dd, J = 10.8, 4.6 Hz, 1H), 3.95 (m, 1H), 3.93 (s, 3H), 3.82 (dt, J = 4.9, 8.2 Hz, 1H), 2.81 (d, J = 4.5 Hz, 3H), 2.38 (m, 1H), 2.32 (m, 1H).\n\n### Synthesis of Compound 27", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics; supervisor Prof. E. Yvonne Jones), 1999\n\n**Author:** Marc C. Deller. **Institution:** Laboratory of Molecular Biophysics and Wellcome Trust Centre for Human Genetics, University of Oxford. **Supervisor:** Prof. E. Yvonne Jones. **Submitted:** D.Phil. thesis, University of Oxford, c. 1999.\n\n## Chapter 5 \u2014 Structure Determination of Oncostatin M\n\nThe aims of this chapter include the growth of OSM crystals and determination of the X-ray structure. Crystal growth, data collection, data processing and analysis of the quality of the structure are reported. Successful structure determination using MAD techniques is reported and initial attempts at molecular replacement are not discussed. A detailed description of the OSM structure and implications for receptor binding is presented in Chapter 6.\n\nThe theory and practice of structure determination using X-ray crystallography is covered in detail in texts such as (Drenth, 1994), (Rhodes, 1993) and (Blundell and Johnson, 1976). An overview of the processes involved is shown in Figure 5-1 below.\n\n### 5.1 Outline of Methods Employed\n\n*Figure 5-1.* Flowchart detailing methods and techniques employed in the course of chapter 5. Programs used represented by rounded boxes.\n\n### 5.2 Methods of Structure Determination\n\n#### 5.2.1 Crystallisation\n\nSamples of OSM-196, \u0394188_OSM and Se-Met-OSM were purified on a Sephadex-75 column in 150 mM sodium chloride, 50 mM Tris-HCl, pH 7.0 (see Chapter 3). The protein was concentrated to between 5 and 14 mg/ml in Centricon 3 kDa cut off filters (Amicon). Protein concentrations were estimated by UV spectrometry using the equation shown in Appendix IV.\n\nCrystallisation conditions were initially screened using the hanging drop vapour diffusion method. Screening of precipitant solutions was carried out using sparse matrix sampling screens (Jancarik and Kim, 1991), ammonium sulphate screens and polyethylene glycol (PEG) screens at 22 \u00b0C and 4 \u00b0C. Typically, 0.5\u20132 \u03bcl of protein solution was pipetted onto a siliconised 22 mm glass coverslip (BDH) so forming a spherical drop. An equal volume of precipitant solution was pipetted onto the protein and the drop then placed over a well of a Linbro plate (ICN) containing 0.5\u20131 ml of corresponding precipitant. The well was sealed using high-pressure vacuum grease (Dow Corning) and the plate left to equilibrate. Drops were monitored microscopically on a daily basis for the first week, followed by weekly inspections thereafter.\n\nConditions producing crystals or fine structured precipitates were investigated further. Finer grid screens were set up around these conditions in an attempt to improve crystal size. Macro- and micro-seeding (Stura and Wilson, 1992) were also carried out. Sitting-drop crystallisation using micro-bridges (Harlos, 1992) was investigated as an alternative to hanging drop conditions in an attempt to manipulate equilibration rates. Additive screens, different temperatures, fine pH sampling (McPherson, 1995) and crystallisation using an oil barrier (Chayen, 1997) were also tested. Crystallisation trials of Se-Met derivatised OSM samples were also carried out. Samples were treated as above with the exception that duplicate trays were set up using protein samples with and without DTT to 1 mM, in an attempt to alleviate oxidation.\n\n#### 5.2.2 Crystal Mounting and Cryocrystallography\n\nSolutions of methyl pentanediol (MPD), glycerol, dioxane and PEG 8000 were tested as cryoprotectants (Petsko, 1975). Mother liquor solutions were prepared containing 5, 10, 15, 20 and 25 % of potential cryoprotectant. Crystals were harvested using a fibre loop (Teng, 1990) and transferred to 2 \u03bcl drops of cryoprotectant solution for ~5 minutes and then mounted in a nitrogen stream (Oxford Cryosystems) at 100 K. Data were collected as in 5.2.3 below. Diffraction patterns were analysed for characteristic ice-rings at ~4 \u00c5 and the quality of crystalline disorder were also assessed.\n\nAlternatively, direct flash cooling of crystals in liquid propane and liquid nitrogen was tested. Crystals were transferred from cryoprotectant solution, as above, and plunged directly into aliquots of liquid propane or liquid nitrogen. Crystals were transferred swiftly and directly into a nitrogen stream (Oxford Cryosystems) and data collected as in 5.2.3 below.\n\nCrystals were also mounted for room temperature data collection in 0.7 mm quartz capillaries (Glas, Germany). Excess mother liquor was removed to prevent movement of crystals. Capillaries were sealed using bees wax and then mounted on a goniometer head for data collection.\n\n#### 5.2.3 Data Collection\n\nInitial native data were collected from a single capillary mounted crystal at room temperature on beam line BL6A2 (Sakabe, 1991) of the Photon Factory, Tsukuba, Japan. Data to 2.5 \u00c5 were collected using 1.0 \u00c5 wavelength radiation with 0.1 mm collimation. The data collection strategy employed the Weissenberg method and involved 3.5\u00b0 oscillations with a coupling constant of 1.5\u00b0/mm. Diffraction data were recorded on BASIII imaging plates (200 \u00d7 400 mm) and scanned off line with a Fuji BA100 IP scanner. Exposure times were typically 240 seconds and the crystal to detector distance was set to 429.7 mm over a helium filled path. Data were collected at a temperature of 14 \u00b0C. Typical beam currents were around 300 mA.", "# Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design\nAuthors: Irimia, A.; Serra, A.M.; Sarkar, A.; Jacak, R.; Kalyuzhniy, O.; Sok, D.; Saye-Francisco, K.L.; Schiffner, T.; Tingle, R.; Kubitz, M.; Adachi, Y.; Stanfield, R.L.; Deller, M.C.; Burton, D.R.; Schief, W.R.; Wilson, I.A.\nSource: PLoS Pathogens, 2017\n\n**Authors:** Adriana Irimia, Andreia M. Serra, Anita Sarkar, Ronald Jacak, Oleksandr Kalyuzhniy, Devin Sok, Karen L. Saye-Francisco, Torben Schiffner, Ryan Tingle, Michael Kubitz, Yumiko Adachi, Robyn L. Stanfield, Marc C. Deller, Dennis R. Burton, William R. Schief, Ian A. Wilson.\n\n**Published:** PLoS Pathogens 13(2): e1006212, Version 2, published February 22, 2017. Received: October 5, 2016; Accepted: February 2, 2017. Editor: Alexandra Trkola, University of Zurich, Switzerland.\n\n**Citation:** Irimia A, Serra AM, Sarkar A, Jacak R, Kalyuzhniy O, Sok D, et al. (2017) Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design. PLoS Pathog 13(2): e1006212.\n\n**Copyright:** \u00a9 2017 Irimia et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.\n\n## Abstract\n\nAmong broadly neutralizing antibodies to HIV, 10E8 exhibits greater neutralizing breadth than most. Consequently, this antibody is the focus of prophylactic/therapeutic development. The 10E8 epitope has been identified as the conserved membrane proximal external region (MPER) of gp41 subunit of the envelope (Env) viral glycoprotein and is a major vaccine target. However, the MPER is proximal to the viral membrane and may be laterally inserted into the membrane in the Env prefusion form. Nevertheless, 10E8 has not been reported to have significant lipid-binding reactivity. Here we report x-ray structures of lipid complexes with 10E8 and a scaffolded MPER construct and mutagenesis studies that provide evidence that the 10E8 epitope is composed of both MPER and lipid. 10E8 engages lipids through a specific lipid head group interaction site and a basic and polar surface on the light chain. In the model that we constructed, the MPER would then be essentially perpendicular to the virion membrane during 10E8 neutralization of HIV-1. As the viral membrane likely also plays a role in selecting for the germline antibody as well as size and residue composition of MPER antibody complementarity determining regions, the identification of lipid interaction sites and the MPER orientation with regard to the viral membrane surface during 10E8 engagement can be of great utility for immunogen and therapeutic design.\n\n## Author summary\n\nThe trimeric Env glycoprotein located on HIV surface is the target of broadly neutralizing antibodies and is the focus of vaccine and therapeutic approaches to prevent HIV infection. Structural studies with HIV Env trimers have shed light on the complete epitopes of several broadly neutralizing antibodies. However, structural determination of the complete epitopes of the highly cross-reactive MPER antibodies has been technically difficult due to the viral membrane component and that these epitopes are probably only exposed transiently after Env engages CD4. In this study, we structurally characterize the interaction of the broadest and most potent MPER-targeting antibody, 10E8, with viral membrane lipids and scaffolded MPER and propose how 10E8 approaches the MPER-viral membrane epitope during neutralization. Our results indicate that 10E8 interacts with the viral membrane via its light chain and engages MPER in an upright orientation with respect to the HIV-1 membrane. These findings are of interest for design of HIV-1 vaccines and therapeutics.\n\n## Data availability\n\nThe atomic coordinates and structure factors of 10E8-T117v2 structures have been deposited in the Protein Data Bank, with the accession codes: 5T6L (for 10E8-T117v2) and 5T85, 5T80 for co-crystals with 06:0 PG and 06:0 PA, respectively, and those for 10E8 mutants-T117v2 structures with the accession codes: 5SY8 (for 10E8 mutant 1-T117v2), 5TFW (for 10E8 mutant 2-T117v2), 5T29 (for 10E8 mutant 3-T117v2) and 5T5B (for 10E8 mutant 5-T117v2). All other relevant data can be found within the paper and its Supporting Information files.\n\n## Funding\n\nThis work was supported by National Institutes of Health Grants UM1 AI100663 to the Scripps Center for HIV/AIDS Vaccine Immunology and Immunogen Discovery (CHAVI-ID) (IAW, WRS, and DRB) and R01 AI084817 (IAW), and by IAVI with the generous support of USAID, Ministry of Foreign Affairs of the Netherlands, and the Bill & Melinda Gates Foundation (CAVD) (IAW, DRB and WRS). A full list of IAVI donors is available at www.iavi.org. Use of the 23-ID sector at Advanced Photon Source was supported by the U.S. Department of Energy. Portions of this research were carried out at the Stanford Synchrotron Radiation Lightsource, a Directorate of SLAC National Accelerator Laboratory and an Office of Science User Facility operated for the U.S. Department of Energy Office of Science by Stanford University. The SSRL Structural Molecular Biology Program is supported by the DOE Office of Biological and Environmental Research, and by the National Institutes of Health, National Institute of General Medical Sciences (including P41GM103393). The contents of this publication are solely the responsibility of the authors and do not necessarily represent the official views of NIGMS, NIAID, NIH, USAID or the US Government. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.\n\n**Competing interests:** WRS is a co-founder and stock holder in Compuvax, Inc. which has programs in non-HIV vaccine design that might benefit indirectly from this research.\n\n## Introduction", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n- **Co-folded partners** are tickboxes built from the Partner blocks' own short names, rather\n  than a list you retype. The two had to agree exactly, so renaming a partner after a protein\n  referenced it failed validation at download time with the campaign already entered. A\n  selection whose partner is later renamed or removed stays visible, ticked and flagged, instead\n  of quietly disappearing.\n- **Apo or inactive structure PDB id** reads the entry back underneath the box: title, method,\n  resolution, and **what is bound to it**. Four characters means every typo is another valid id,\n  so the title is the only way to see the wrong structure was named -- and \"apo\" in a title is\n  not a guarantee. Entries carrying ligands read amber rather than green: found, and worth your\n  judgement.\n- **Upload a structure** covers anything not in the PDB -- an unreleased entry, a colleague's\n  model, a construct with the fusion cut out. It is copied into the bundle under `reference/`\n  and named as the apo path, so the run never needs the network for it. Contents are checked\n  against the extension, and the filename sanitised before it becomes a path inside an archive\n  somebody unpacks.\n- **Targets per prediction process** defaults to 4. On Apple unified memory the allocator never\n  returns everything, so a long campaign starves itself and only process exit frees it. Blank\n  was the wrong default for this hardware; `0` disables recycling.\n\nThe form also carries the rest of what a spec can say: per-protein **ligand scoping**\n(`Ligands:`), a shared report **group**, the **motif annotator** (`Family type:`), a ligand's\n**pharmacology** (`Role:`), and an apo reference as a file in the bundle with an optional\n**chain** -- deriving a form from a spec is only honest if the form can hold everything the spec\nsays. Everything you type is kept in your browser, so downloading a bundle, stepping over to\nAnalysis, or reloading does not lose it.\n\n**The bundle is the only file to keep.** It runs the campaign on your machine, and uploading it\nback at **Upload bundle** brings the whole page back as it was -- add a ligand, change a pocket,\nretune the run, download a new one. The form travels inside the bundle, so one file answers both\nquestions; **Clear** empties the form, and asks first.\n\nA bundle with no saved page inside it -- every example campaign, and everything already in the\n**Runs** archive -- is rebuilt from the `boltz_input.md` it carries, but only when that is\nprovably safe: the rebuilt page goes back through the form's own parser and assembler and the\nresulting spec is compared with the original, refusing with the difference named if they\ndisagree. A form that silently drops a directive looks correct, rebuilds into a different\ncampaign, and says nothing -- worse than not loading at all. An unrecognised directive refuses by\nname for the same reason.\n\nThen choose the run settings. Every one maps to a real `BoltzMaker.py` flag, written literally\ninto the generated script so you can read exactly what will run. **Commands** above documents\nwhat each flag does; the form's own labels, defaults and the few web-only notes are below.\n\n| Setting | Flag | Default |\n|---|---|---|\n| Accelerator | `--accelerator` | auto (picks CUDA or Apple MPS when there is one) |\n| Use potentials | `--no-potentials` when unticked | on |\n| Data-loading workers | `--workers` | 0 |\n| Parallel diffusion samples | `--max-parallel-samples` | 1 |\n| MPS high-watermark ratio | `--mps-watermark` | 1.0 (Apple Silicon only; ignored elsewhere) |\n| Recycling steps | `--recycling-steps` | blank = Boltz's own (more steps is slower and usually only marginally better) |\n| Sampling steps | `--sampling-steps` | blank = Boltz's own |\n| Structure samples per target | `--diffusion-samples` | blank = one sample per target |\n| Affinity diffusion samples | `--diffusion-samples-affinity` | blank = Boltz's own; only matters where affinity prediction is on |\n| Affinity sampling steps | `--sampling-steps-affinity` | blank = Boltz's own |\n| Max MSA sequences | `--max-msa-seqs` | blank = BoltzMaker's 4096 (Boltz's own is 8192) |\n| Auto-retries per target | `--max-retries` | 2 |\n| Preflight size-warning threshold | `--memory-warn-tokens` | 1500 |\n| Only run the first N targets | `--limit` | blank = the whole campaign. 1 or 2 is the cheapest way to prove the pipeline works before committing hours of GPU time |\n| Treat preflight warnings as failures | `--strict` | off |\n| Skip PLIP interaction analysis | `--skip-interactions` | off. Leave it off: PLIP produces the interaction fingerprints the Analysis step shows, and skipping empties that panel to save minutes |\n| Skip apo-vs-holo compare-sse | `--skip-sse` | off. Left off, every protein gets an apo reference: an experimental structure if you gave a PDB id, otherwise an extra ligand-free prediction of that protein -- one more target per protein of GPU time. Tick to predict nothing extra |\n| Keep private | _(this site only)_ | off. Ticked, nothing about the run is kept on the server: the bundle is not archived, and the results file you upload later is recognised as private and not archived either. Left off, the run is listed under Runs |\n\nSubmitting validates the spec (`format`, then `generate`, so a broken campaign fails here rather\nthan an hour into a run on your machine) and downloads `boltzmaker_<campaign>.command`, typically\naround 200KB, containing:", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices\nAuthors: Deller, M.C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\n*Equation 7-1.* Equation used to determine protein concentration from absorbance readings at 280 nm (A280). Measurements were made in a 1 ml cell using a Beckman DU-600 instrument. Protein concentration (mg/ml) were estimated based on the number of tyrosine residues (Y) and the number of tryptophan residues (W) in relation to the molecular weight (Mr).\n\n#### 7.4.2 Spectrophotometric Measurement of Oligonucleotide Concentration\n\n[Oligonucleotide] (pM/\u03bcl) = 118 \u00d7 A260\n\n*Equation 7-2.* Equation used to determine primer and oligonucleotide concentrations for a 1:40 dilution of stock oligonucleotide. Absorbance measurements at 260 nm (A260) were made in a 1 ml cell using a Beckman DU-600 instrument. The factor 118 is derived from the number of bases and their extinction coefficients and is specific for each oligonucleotide.\n\n#### 7.4.3 Spectrophotometric Measurement of Plasmid DNA Concentration\n\nA260 of 1.0 = 50 \u03bcg/ml\n\n*Equation 7-3.* Equation used to determine plasmid DNA concentrations for a 1:100 dilution of plasmid stock. Absorbance measurements at 260 nm (A260) were made in a 1 ml cell using a Beckman DU-600 instrument.\n\n## References\n\n*(reference list omitted)*", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\n**89.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein R^7 is selected from H, D, OH, CN, halo, C1\u20136 alkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, C1\u20136 haloalkyl, cyano-C1\u20136 alkyl, HO\u2014C1\u20136 alkyl, C1\u20136 alkoxy-C1\u20136 alkyl, C3\u20137 cycloalkyl, C1\u20136 alkoxy, C1\u20133 haloalkoxy, amino, C1\u20136 alkylamino, and di(C1\u20136 alkyl)amino.\n\n**90-91.** (canceled)\n\n**92.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein R^W is selected from the drawn set of alpha-halo alkyl (Cl, Br, F), alpha-cyano alkyl, substituted vinyl (including those bearing C(O)R^83, C(O)R^81, and S(O)tR^83 groups), acrylonitrile-type, allene, and alkynyl warheads shown in the claim.\n\n**93.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein R^W is a substituted vinyl group of the form L^2\u2014C(R^82)=C(R^82)(R^82) or L^2\u2014C(R^82)=C(R^82)C(O)R^83.\n\n**94.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein R^W is L^2\u2014C(R^82)=C(R^82)(R^82).\n\n**95.** (canceled)\n\n**96.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein L^2 is -L-NR^9C(O)\u2014 or -L-NR^9S(O)2\u2014 wherein L^2 is attached to Ring moiety A through an L linking group.\n\n**97.** (canceled)\n\n**98.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein each L is a bond.\n\n**99-102.** (canceled)\n\n**103.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein: each R^81 and R^83 is independently selected from H, D, halo, CN, C1\u20136 alkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, C1\u20136 haloalkyl, C3\u20137 cycloalkyl, C1\u20136 alkoxy, C(O)R^b8, C(O)NR^c8R^d8, and C(O)OR^a8; each R^82 is independently selected from H, D, halo, CN, C1\u20136 alkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, C1\u20136 haloalkyl, C3\u20137 cycloalkyl, C(O)R^b8, C(O)NR^c8R^d8, and C(O)OR^a8; and each R^a8, R^c8, and R^d8 is independently selected from H, C1\u20136 alkyl, and C3\u20137 cycloalkyl.\n\n**104.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein each R^81 and R^83 is independently selected from H, CN, C1\u20136 alkyl, and C1\u20136 alkoxy; and each R^82 is independently selected from H, CN, and C1\u20136 alkyl.\n\n**105-108.** (canceled)\n\n**109.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein R^9 is H or methyl.\n\n**110-111.** (canceled)\n\n**112.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein m is an integer selected from 0, 1, 2, and 3; n is an integer selected from 0, 1, 2, and 3; the dashed bond is a single or double bond, provided that proper valency is maintained; Z is N or CR^7; X is C, CH or N and Y is C or CH, or X is C or CH and Y is C, CH, or N; Ring moiety A is selected from C3\u201310 cycloalkyl, 6\u201310 membered aryl, 4\u201310 membered heterocycloalkyl, and 5\u201310 membered heteroaryl; Ring moiety B is selected from C3\u20137 cycloalkyl, phenyl, 5\u20136 membered heterocycloalkyl, and 5\u20136 membered heteroaryl; and R^1, R^2 and their sub-variables are as recited in the claim.\n\n**113.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein: m is an integer selected from 0, 1, 2, and 3; n is an integer selected from 0, 1, 2, 3 or 4; the dashed bond is a single or double bond, provided that proper valency is maintained; Z is N or CR^7; X is C, CH or N and Y is C or CH, or X is C or CH and Y is C, CH, or N; Ring moiety A is selected from C3\u201310 cycloalkyl, 6\u201310 membered aryl, 4\u201310 membered heterocycloalkyl, and 5\u201310 membered heteroaryl; Ring moiety B is selected from C3\u20137 cycloalkyl, phenyl, 5\u20136 membered heterocycloalkyl, and 5\u20136 membered heteroaryl; R^4 and R^5 are each independently selected from H and C1\u20136 alkyl; R^6 is selected from C3\u20137 cycloalkyl, phenyl, 4\u20137 membered heterocycloalkyl, 5\u20137 membered heteroaryl and their C1\u20134 alkyl-linked variants, each optionally substituted with 1, 2, 3, or 4 independently selected R^6A substituents; R^7 is selected from H, D, CN, halo, C1\u20136 alkyl, C1\u20136 haloalkyl, cyano-C1\u20136 alkyl, HO\u2014C1\u20136 alkyl, C1\u20136 alkoxy-C1\u20136 alkyl, and C3\u20137 cycloalkyl, C1\u20136 alkoxy, C1\u20133 haloalkoxy, amino, C1\u20136 alkylamino, and di(C1\u20136 alkyl)amino; R^W is a substituted vinyl, acyl-vinyl, allene, sulfonyl-vinyl or acrylonitrile group as drawn; L^2 is -L-C(O)\u2014, -L-NR^9C(O)\u2014, and -L-NR^9S(O)2\u2014, wherein L^2 is attached to Ring moiety A through the L linking group; each L is independently a bond or C1\u20136 alkylene; and R^83, R^82, R^9 and R^G are as recited in the claim.\n\n**114.** The compound of claim 1, selected from:", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\n- N-(2-(2-Ethyl-10-(2-((4-fluorophenyl)amino)-2-oxoethyl)-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide;\n- N-(2-(2-Ethyl-4-oxo-10-(2-oxo-2-(phenylamino)ethyl)-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide;\n- N-(2-(2-Ethyl-4-oxo-10-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide;\n- N-(2-(10-(2-((2-Chloro-4-(trifluoromethyl)phenyl)amino)-2-oxoethyl)-2-ethyl-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide;\n- N-(2-(2-Ethyl-10-(2-((4-ethylphenyl)amino)-2-oxoethyl)-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide;\n- N-(2-(10-(2-((2-Chloro-4-fluorophenyl)amino)-2-oxoethyl)-2-ethyl-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide;\n- N-(2-(2-Ethyl-4-oxo-10-(2-oxo-2-((3-(trifluoromethyl)phenyl)amino)ethyl)-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide;\n- N-(2-(10-(2-((3,4-Difluorophenyl)amino)-2-oxoethyl)-2-ethyl-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide;\n- N-(2-(2-Ethyl-10-(2-((4-fluoro-2-methylphenyl)amino)-2-oxoethyl)-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide;\n- N-(2-(10-(2-(Cyclohexylamino)-2-oxoethyl)-2-ethyl-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide;\n- Methyl 4-acrylamido-3-(2-ethyl-10-(2-((4-fluorophenyl)amino)-2-oxoethyl)-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)benzoate;\n- 4-Acrylamido-3-(2-ethyl-10-(2-((4-fluorophenyl)amino)-2-oxoethyl)-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)-N-(2-hydroxyethyl)benzamide;\n- 4-Acrylamido-3-(2-ethyl-10-(2-((4-fluorophenyl)amino)-2-oxoethyl)-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)-N-(2-methoxyethyl)benzamide;\n- N-(2-(2-Ethyl-10-(2-((4-fluorophenyl)amino)-2-oxoethyl)-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)-4-(1-hydroxy-3-azabicyclo[3.1.0]hexane-3-carbonyl)phenyl)acrylamide;\n- N-(5-Cyano-2-(2-ethyl-4-oxo-10-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide;\n- N-(4-(2-Ethyl-4-oxo-10-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)-1-methyl-1H-pyrazol-3-yl)acrylamide;\n- N-(2-(10-(2-((3,4-Difluorophenyl)amino)-2-oxoethyl)-2-ethyl-8-methoxy-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide;\n- N-(2-(2-Ethyl-10-(2-((4-fluorophenyl)amino)-2-oxoethyl)-7,8-dimethyl-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide;\n- N-(2-(2-Ethyl-9-methoxy-4-oxo-10-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide;\n- N-(2-(8-Ethyl-6-oxo-10-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-6,10-dihydropyrido[2',3':4,5]imidazo[1,2-a]pyrimidin-7-yl)phenyl)acrylamide;\n- 2-(2-Ethyl-4-oxo-3-(2-(vinylsulfonamido)phenyl)benzo[4,5]imidazo[1,2-a]pyrimidin-10(4H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide;\n- N-(2-(2-ethyl-4-oxo-10-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-4,6,7,8,9,10-hexahydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide;\n- 2-(2-Ethyl-4-oxo-3-(2-(vinylsulfonamido)phenyl)-6,7,8,9-tetrahydrobenzo[4,5]imidazo[1,2-a]pyrimidin-10(4H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide;\n- 2-(4-Cyano-3-ethyl-1-oxo-2-(2-(vinylsulfonamido)phenyl)benzo[4,5]imidazo[1,2-a]pyridin-5(1H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide;\n- N-(2-(4-Cyano-3-ethyl-1-oxo-5-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-1,5-dihydrobenzo[4,5]imidazo[1,2-a]pyridin-2-yl)phenyl)acrylamide; and\n- N-(2-(3-Ethyl-1-oxo-5-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-1,5-dihydrobenzo[4,5]imidazo[1,2-a]pyridin-2-yl)phenyl)acrylamide;\n\nor a pharmaceutically acceptable salt for any of the aforementioned.\n\n**115.** A pharmaceutical composition comprising the compound of claim 1, or a pharmaceutically acceptable salt thereof, and a pharmaceutically acceptable carrier.\n\n**116.** A method of inhibiting WRN, comprising contacting the WRN with a compound of claim 1, or a pharmaceutically acceptable salt thereof.\n\n**117.** A method of inhibiting WRN in a patient, comprising administering to the patient a compound of claim 1, or a pharmaceutically acceptable salt thereof, or a pharmaceutical composition comprising the compound of claim 1.\n\n**118.** The method of claim 116, wherein the inhibition of WRN comprises the inhibition of: (i) WRN helicase activity; or (ii) WRN ATPase activity; or (iii) both (i) and (ii).\n\n**119.** A method of treating a disease or disorder associated with WRN activity in a patient, the method comprising administering to the patient in need thereof a therapeutically effective amount of the compound of claim 1, or a pharmaceutically acceptable salt thereof, or a pharmaceutical composition comprising the compound of claim 1.\n\n**120.** The method of claim 119, wherein the disease or disorder is cancer.\n\n**121.** The method of claim 120, wherein the cancer is a microsatellite instability-high (MSI-H) cancer.\n\n**122.** The method of claim 120, wherein the cancer is a mismatch repair deficient (dMMR) cancer.\n\n**123-128.** (canceled)\n\n## References\n\n*(reference list omitted)*", "# pants.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n## pants.mdeller.com/stats\n\n## **Live instrumentation.** Everything this project has read, produced and is running on, read straight from the database and the machine.\n\nNothing here is written by hand. Each figure is a query or a reading, and the\npage refreshes when the data actually changes.\n\nupdated 2026-09-07 18:35:33 UTC\n\u00b7 database last written\n43794.7 min ago\n\n## The funnel\n\nHow 14,804,920 sequences become a shortlist,\nand what that shortlist is judged against. Each of the first steps is a filter, and the one that\nmatters most is the last of them: three catalytic residues can be present in a sequence and still\nnot touch each other in the folded protein, and only a structure can tell you which.\n\nSequences read\n:   14,804,920\n:   every predicted protein in the assemblies\n\nSurvived recall\n:   439\n:   matched a polyesterase profile and carry a complete triad\n\nGiven a structure\n:   416\n:   folded with ESMFold, superposed onto IsPETase\n\nTriad confirmed in 3D\n:   403\n:   the three residues actually meet in space\n\nToo long to fold\n:   24\n:   over 450 residues; deferred, not discarded\n\nReference structures\n:   559\n:   the known enzymes a candidate is compared against, not part of the catch\n\nMeasured positives\n:   342\n:   reference enzymes whose activity someone actually assayed and published\n\nActivity measurements\n:   2759\n:   individual published values behind those enzymes, each with its citation\n\nUnlabelled sequences\n:   25,041\n:   homologues nobody has assayed, held apart from the catalogue on purpose\n    so they cannot inflate a count that is quoted as evidence\n\nSequence clusters\n:   170\n:   groups of related sequences carrying their own ecology \u2014 where the source\n    organism was found, and at what temperature\n\nRegistered sources\n:   7\n:   public databases and published screens this catalogue is built from, each\n    with its retrieval date and licence\n\n## Where the sequences came from\n\nMetagenomics reads the DNA of a whole microbial community straight from the\nenvironment, without culturing anything. Each row is one such environment, chosen because it has\nhad decades of plastic exposure and every reason to have evolved a use for it. **Yield per\nmillion** is the comparable figure: raw candidate counts only tell you which environment\nwas sequenced most deeply.\n\n| Environment | Sequences read | Candidates | Per million | Scan runs |\n| --- | --- | --- | --- | --- |\n| Human gut | 12,584,458 | 311 | 24.7 | 58 |\n| Compost | 1,020,575 | 69 | 67.6 | 3 |\n| Marine plastisphere | 737,027 | 44 | 59.7 | 2 |\n| Landfill | 436,229 | 15 | 34.4 | 1 |\n| Wastewater | 26,631 | 0 | 0.0 | 1 |\n\n## Protein structures\n\nA structure is either **measured** or **predicted**,\nand the difference matters more than any single number on this page. An X-ray crystal structure\nis experimental evidence. AlphaFold and ESMFold produce a hypothesis \u2014 usually a good one for the\nfold, least reliable exactly where this project looks, in the flexible loops that gate the active\nsite.\n\n60\n\nX-ray crystal structures\n\nDeposited experimental structures. Best resolution here:\n**0.91** \u00c5 \u2014 lower is\nsharper, and below 1 \u00c5 individual atoms are resolved.\n\n104\n\nAlphaFold models\n\nPredicted from an alignment of related sequences. Downloaded where a UniProt\naccession exists.\n\n395\n\nESMFold predictions\n\nFolded here, from sequence alone, for everything with no deposit and no\naccession. Mean confidence\n**91.0** pLDDT\n(0\u2013100; above 90 the fold is reliable).\n\n547\n\nwith a measured catalytic triad\n\nSer, His and Asp found close enough to pass charge between them.\n**1** deposits\nwere caught as catalytic knockouts: inactivated constructs crystallographers make to trap\nsubstrate, which carry the right name and the wrong chemistry.\n\n## What the model learns from\n\nThe distinction this whole project turns on. **Measured** means a\npublished experiment stands behind the label \u2014 someone ran the assay. **Annotated\nonly** means a computer assigned it from sequence similarity and nobody measured anything;\na model trained on those is largely rediscovering the rule that produced them, which is why the\ntwo are never pooled here.\n\n342\n\nmeasured positives\n\nMostly from PAZy, which lists an enzyme because activity on a plastic was\nmeasured and published, with the DOI attached.\n\n568\n\nannotated only\n\nLabelled by similarity. Counted, reported, and kept out of any headline\nclaim.\n\n131\n\nhard negatives\n\nOther \u03b1/\u03b2-hydrolases: the same fold, no PET activity.\n\n179\n\nwithin-family negatives\n\nPolyesterases that do *not* degrade PET \u2014 the scarcest and most\nvaluable class here, because it is the only one that makes the hard question answerable.\n**150** of\nthem were expressed and assayed and released no product; the rest rest on a database\nrecording only what worked, which cannot separate tested-and-failed from never-tested.\n\n| Provenance of every positive | Count | Evidence |\n| --- | --- | --- |\n| EC-auto-annotated | 446 | inferred |\n| PAZy-measured | 312 | measured |\n| ESTHER-family-predicted | 50 | inferred |\n| Science-landscape-measured | 30 | inferred |\n| ACS-screen-measured | 28 | inferred |\n| ESTHER-family-protein-evidence | 14 | inferred |\n| EC-experimental | 7 | measured |\n| UniProt | 6 | measured |\n| HGMP-measured | 5 | measured |\n| Tournier et al. 2020, Nature | 1 | measured |\n| Son et al. 2019, ACS Catal. | 1 | measured |\n| Shi et al. 2023, Angew. Chem. Int. Ed. | 1 | measured |\n| PDB-construct | 1 | measured |\n| Orr et al. 2024, Biotechnol. J. | 1 | measured |\n| Oda et al. 2018, Appl. Microbiol. Biotechnol. | 1 | measured |\n| Lu et al. 2022, Nature (MutCompute) | 1 | measured |\n| Li et al. 2025, Int. J. Biol. Macromol. | 1 | measured |\n| J. Hazard. Mater. 2023 | 1 | measured |\n| Cui et al. 2024, Nat. Commun. 15:1417 | 1 | measured |\n| Cui et al. 2021, ACS Catal. | 1 | measured |\n| Austin et al. 2018, PNAS | 1 | measured |\n\n## Evidence and citations", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n**3.** On the offline computer, in a terminal, go to where you copied those files and run:\n\n```sh\n./boltzmaker-installer.sh -o ./boltzmaker-env\nsource ./boltzmaker-env/activate.sh\nexport KMP_DUPLICATE_LIB_OK=TRUE\n```\n\n**4.** Check it worked:\n\n```sh\npython3 BoltzMaker.py preflight examples/t4_lysozyme/boltz_input.md\n```\n\nFull details, including a couple of one-time extra steps this path needs, are in\n[`docs/tier_b_offline_install.md`](docs/tier_b_offline_install.md).\n\n## \ud83e\udde9 Architecture\n\n```\n+----------------+     +----------+     +-----------+     +-----+     +---------+\n| boltz_input.md | --> | generate | --> | preflight | --> | run | --> | analyze |\n+----------------+     +----------+     +-----------+     +-----+     +---------+\n```\n\n`boltz_input.md` is the family x partners x ligand DSL spec; the four commands that read it\nproduce:\n\n| Command | Produces |\n|---|---|\n| `generate` | `boltz_yamls/*.yaml` + manifest |\n| `preflight` | PASS/WARN/FAIL: CLI, GPU/MPS, disk, iCloud, YAML/SMILES/chain-id/chemistry, memory |\n| `run` | `boltz predict` -> `boltz_output/`, live progress + memory monitor, resumable |\n| `analyze` | `boltz_summary.csv`/`.xlsx`, `boltz_dashboard.html`, `boltz_cif/`, plus interaction files if `setup-plip` has run, plus `boltz_sse_comparison.csv`/`.html` for any family with an `Apo structure:` set (see **compare-sse** below) |\n\nEach stage reads only the manifest + files the previous stage wrote, so any stage can be\nre-run on its own without repeating the others: `all` simply chains all four.\n\n## \ud83d\udd27 One-time setup\n\n```sh\npython3 BoltzMaker.py setup\n```\n\nCreates a dedicated `.venv` (Python 3.12: boltz pins `numpy<2.0`, which has no prebuilt wheel\nfor newer Pythons) next to `BoltzMaker.py` and installs `boltz`, `rich`, `pandas`, `openpyxl`,\n`pyyaml`, `rdkit`, `matplotlib`, `psutil`, `scipy`, `gemmi`, `biopython`, `plotly`, `reportlab`\nand `requests`. This pulls PyTorch (~2-3 GB) and, on the first `boltz predict`, several GB of\nmodel weights. Every other command relaunches itself under this managed environment, so you can\nkeep invoking the script with whatever `python3` is on your PATH. `--force` recreates the venv\nfrom scratch; `-y/--yes` skips the download confirmation prompt.\n\n### Optional: `setup-plip` (protein-ligand interaction analysis)\n\n```sh\npython3 BoltzMaker.py setup-plip\n```\n\nEntirely optional and separate from the venv above. Builds a `.plip_env` (~1-1.5GB, mostly\nPyMOL's own Qt/Cairo/HDF5 dependencies) via a self-downloaded\n[micromamba](https://micro.mamba.pm), resolving `osx-arm64`/`osx-64`/`linux-64`/`linux-aarch64`\nfrom the running platform so it works on a Linux server as well as a Mac, and vendors a pinned\ncommit of [cif2plip](https://github.com/bellcheddar/cif2plip), which converts a Boltz ModelCIF\ninto a strict PDB and runs [PLIP](https://github.com/pharmai/plip) on it for real interaction\nfingerprints (H-bonds, salt bridges, pi-stacking, halogen bonds, metal coordination, etc.).\nconda-forge is used deliberately rather than pip, because PLIP needs OpenBabel and PyMOL as\nin-process imports and both PyPI routes are broken -- see the first **Troubleshooting** row.\n\nWithout `.plip_env` everything degrades gracefully: `analyze` skips interaction analysis (the\ndashboard looks exactly as it did before this feature) and the `new` wizard doesn't ask about\nreference structures. Nothing else requires it, and `preflight`'s `plip_env` check always\nreports which mode you're in.\n\n### Optional: `mkdssp`/`dssp` (for `compare-sse`'s SSE-boundary-shift metric)\n\nNot bundled or installed by any BoltzMaker command -- a small external binary you may already\nhave (`brew install dssp` on macOS, or `conda install -c salilab dssp`). Only needed for one\n`compare-sse` metric (secondary-structure-element boundary shift, used when a structure has\nno deposited HELIX/SHEET records -- true for every Boltz-predicted structure). Every other\nmetric works without it.\n\nTwo fixes make that fallback work. The temporary legacy-PDB copy is renamed to chain `A` first,\nbecause its chain-ID field is a single character while BoltzMaker's chain names are the family id\nverbatim (up to 5), which raised \"chain name too long for the PDB format\". And a minimal\nsynthetic `HEADER` line is prepended, because `gemmi` writes none for a structure built purely in\nmemory and `mkdssp` >= 4 then mis-sniffs the headerless file as mmCIF and refuses it. Neither is\nread back -- only residue numbers and SSE kind are.\n\n### Alternative: `pixi` (one unified environment, macOS + Linux/CUDA)\n\n```sh\n./install.sh\n```\n\nInstalls [`pixi`](https://pixi.sh) if absent, then solves and installs everything (`boltz`,\n`rdkit`, PLIP's OpenBabel/PyMOL stack, and the rest of `requirements.txt`) as **one**\nenvironment from `pixi.toml`/`pixi.lock`, replacing `setup` + `setup-plip` above. Reproducible\n(the committed `pixi.lock` pins every package on both `osx-arm64` and `linux-64`), and the only\npath here actually tested against Linux/CUDA rather than macOS/MPS alone. Once installed:\n\n```sh\npixi run preflight my_campaign.md\npixi run all my_campaign.md\n```\n\n(or `pixi shell` to activate the environment and run `python3 BoltzMaker.py ...` as normal).\nPLIP needs one extra one-time step, same spirit as `setup-plip`: `pixi run postinstall`. For a\nmachine with no internet at all, see **Path 3** above.\n\n### Patch the installed `boltz`\n\nTwo defects in `boltz` 2.x cost one campaign 14.5 hours and 20 invocations for zero\nstructures, so BoltzMaker ships fixes:\n\n```bash\npython3 patches/apply_boltz_patches.py          # idempotent; keeps a .orig of every file\npython3 patches/apply_boltz_patches.py --check  # what preflight runs\n```", "# 2023 \u2013 Anti-mutant calreticulin (calr) antibodies and uses thereof\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-12 on marcdeller.com (categories: patents).*\n\n* [Link to PDF](https://marcdeller.com/wp-content/uploads/2025/08/CalR.pdf)\n* [Link to Google Patent](https://patents.google.com/patent/WO2023107994A1/en?oq=WO2023107994A1)\n\n* This patent discloses antibodies specifically targeting mutant calreticulin (CALR) proteins implicated in myeloproliferative neoplasms and other cancers.\n* It emphasizes combined use of anti-mutant CALR antibodies with anticancer agents to achieve synergistic and enhanced tumor-killing effects.\n* Bispecific antibodies that bind mutant CALR and CD3 antigens boost immune targeting by recruiting T cells to cancer cells\n* This targeted immunotherapy approach offers a promising, selective treatment strategy for CALR mutation-driven malignancies, improving therapeutic efficacy and patient outcomes.\n\n1. **Faiz, A. M., \u201cMutant Calreticulin in MPN: Mechanistic Insights and Therapeutic Approaches,\u201d** Curr Hematol Malig Rep [Vol. 20, 2025](https://pmc.ncbi.nlm.nih.gov/articles/PMC12085081/)\n2. **Reis, M., \u201cINCA033989: the first shot on goal for MPNs?,\u201d** Blood [Vol. 144, 2024](https://ashpublications.org/blood/article/144/22/2278/526059/INCA033989-the-first-shot-on-goal-for-MPNs)\n3. **Reis, M., \u201cSelective targeting of mutated calreticulin by the monoclonal antibody INCA033989,\u201d** Blood [Vol. 144, 2024](https://www.sciencedirect.com/science/article/pii/S0006497124024315)\n4. **Hug, B. A., \u201cT-Cell Redirector INCA035784 Fights CALR-Mutated MPNs,\u201d** Targeted Oncology [Vol. 20, 2025](https://www.targetedonc.com/view/t-cell-redirector-inca035784-fights-calr-mutated-mpns)\n\nThe patent focuses on innovative therapeutic approaches targeting mutant calreticulin (CALR) proteins, which are implicated in certain cancers, particularly Philadelphia-negative myeloproliferative neoplasms (MPNs). CALR is a chaperone protein primarily residing in the endoplasmic reticulum, and mutations in its gene\u2014especially in exon 9\u2014produce oncogenic forms that abnormally activate cellular receptors like the thrombopoietin receptor, driving disease progression.\n\nThe core invention described is a pharmaceutical composition comprising antibodies that specifically bind to mutant CALR proteins or their functional fragments. These antibodies may be monoclonal, chimeric, humanized, bispecific, or multi-specific, designed to recognize unique mutant sequences, including cleaved forms of mutant CALR that are prevalent in patient cells. The antibodies target epitopes formed by mutation-induced amino acid changes distinct from wild-type CALR, enabling selective recognition of cancerous cells.\n\nImportantly, the patent emphasizes the synergistic use of these anti-mutant CALR antibodies in combination with various anticancer agents\u2014such as alkylating agents, platinum compounds, antimetabolites, kinase inhibitors including JAK inhibitors, molecular targeting drugs, immunotherapeutics, and others. This combined treatment dramatically enhances antitumor efficacy compared to using either treatment alone, offering a promising strategy for improved cancer therapy.\n\nOne embodiment details bispecific antibodies that simultaneously bind mutant CALR and CD3 antigens, aiming to recruit T cells directly to mutant CALR-expressing cancer cells, boosting immune-mediated tumor destruction. This design addresses challenges with previous antibodies by enabling specific and potent targeting of both full-length and cleaved mutant proteins on cancer cell surfaces, enhancing therapeutic outcomes.\n\nOverall, the patent presents a novel and effective approach to treating cancers associated with mutant CALR, especially MPNs, through precision-targeted immunotherapies combined with standard or innovative anticancer drugs. This approach holds potential to significantly improve patient prognosis by leveraging the distinct biology of mutant CALR and enhancing immune system engagement.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\nThe most notable difference to the LIF system is the use of multiple OSM receptor assemblies; namely type I (Linsley et al., 1989) and type II (Mosley et al., 1996). Type I assemblies consist of a low affinity (Kd ~ 10\u207b\u2079 M) association with LIFR (Gearing et al., 1991) followed by high affinity (Kd ~ 10\u207b\u00b9\u00b9 M) binding to gp130, so forming the heterodimeric signal transducing complex. OSM is the only gp130 cytokine that can bind, albeit with low affinity, to gp130 alone (Sporeno et al., 1994). The low affinity OSM-gp130 complex is incapable of inducing proliferative responses in various cell lines and LIFR is required to form a signal-transducing complex. Conversely, type II complexes have been identified, in a number of cell lines that utilise a high affinity OSM-specific receptor \u03b2 subunit (OSMR\u03b2) that do not bind LIF (Gearing and Bruce, 1992; Mosley et al., 1996; Thoma et al., 1994). This dual OSM receptor system provides an insight into the mechanisms of OSM signal transduction, which shares many similarities with LIF, whilst at the same time exhibiting many OSM-specific biological activities.\n\nA three site binding model has been suggested with similarity to that of the LIF-LIFR-gp130 system (see 1.5.1.2, Figure 1-6). Site 3 involves the conserved residues Phe 160 and Lys 163 which are predicted to reside at the N-terminal end of helix D. Mutagenesis of both these residues has been shown to negate LIFR and OSMR\u03b2 binding and biological activity. In Ba/F3 cell based assays mutagenesis of Phe 160 has been shown to have approximately 30-fold lower activity than equivalently mutated Lys 163 on binding to LIFR-gp130, so suggesting that Phe 160 predominates in LIFR binding. Conversely, in Ba/F3 cell based assays Phe 160 has been shown to display approximately 7-fold lower activity on binding to OSMR\u03b2-gp130 over Lys 163, so suggesting that Lys 163 plays a greater contribution towards OSMR\u03b2 interactions. Site 3 interactions are presumably mediated through interactions with the LIGHT motif of the putative immunoglobulin-like domain of LIFR (Chobotova and Heath, 2000).\n\nResidues of OSM important for gp130 binding have been mapped to a solvent exposed 'hot spot' centred around residues Gly 120, Asn 124 of helix C and residues Gln 16 and Gln 20 of helix A. These residues are structurally equivalent to those of LIF site 2 (i.e. Gln 25, Ser 28, Gln 23, Ile 121, Gly 124 and Ser 127), all presumably lying within 2-3 \u00c5, and forming a hydrophobic pocket between helices A and C. The three site binding model is further complicated by the finding that OSM can also bind the N-terminal immunoglobulin-like domain of gp130 through site 3 interactions (Staunton et al., 1998). This is thought to allow formation of a pseudosymmetrical complex containing one LIFR (or OSMR\u03b2), one gp130 and two oncostatin M molecules.\n\nTwo intramolecular disulphide bridges are present spanning Cys 6-Cys 127 and Cys 49-Cys 167. It has been shown that mutation of Cys 49 and Cys 167 reduces activity ~10-fold (Kallestad et al., 1991) whereas mutations in the first disulphide bridge are not deleterious. Deletion mutants and neutralising antibody studies have revealed that residues either side of the second disulphide are required for activity. This is possibly because the second disulphide functions in anchoring the N-terminal region of helix D on which the site 3 residues Phe 160 and Lys 163 reside, presumably disruptions in this area would adversely affect LIFR and gp130 binding.\n\n#### 1.5.2.3 Oncostatin M\n\nOSM is a monomeric glycoprotein with a molecular weight of ~28 kDa first isolated from supernatants of U-937 histiocytic lymphoma cells treated with phorbol ester (Zarling et al., 1986). OSM is initially expressed as a 252 amino acid precursor polypeptide containing a 25 residue N-terminal hydrophobic signal sequence (Malik et al., 1989). Post-translational processing results in removal of the signal sequence. A further 31 residues are removed from the C-terminus at a pair of basic residues (196 and 195) which act as a proteolytic site, so resulting in the predominant 196 residue form (Linsley et al., 1990). The protein contains two potential N-linked glycosylation sites at residues 100 and 227, which are glycosylated in samples produced in eukaryotic expression systems (Malik et al., 1992).\n\n## 1.6 Aims of Thesis\n\nThis thesis aims to delineate structural features of two key members of the haemopoietin cytokine family; namely OSM and LIFR. This thesis reports the X-ray structure of a rationally designed OSM mutant along with investigations of several expression systems used for production and characterisation of LIFR. The structure of OSM, reported here for the first time, represents a structural framework for the understanding of mutagenesis data. This data will allow functional dissection of the modes of receptor engagement used by the gp130 family of cytokines. The established protocols for production and expression of LIFR will also allow structural investigation of site 3 interactions, which have yet to be analysed structurally. This information will provide further insights into the molecular basis of signal transduction by the haemopoietin cytokines and the epitopes involved.\n\n## 2. Chapter 2 \u2013 Engineering of Truncated Oncostatin M\n\nThe aims of this chapter include the design and production of OSM truncation mutants suitable for use in crystallographic studies. Two OSM C-terminal truncation mutants were engineered as GST fusions in the pGEX expression vector. Subsequent expression in E. coli and characterisation of OSM is reported in Chapter 3.\n\nAppendix I provides a summary of the reagents and buffers used in the engineering of truncated OSM. The methods involved are discussed in greater detail in texts such as (Maniatis et al., 1989). An overview of the processes involved is shown in Figure 2-1 below.\n\n## 2.1 Outline of Methods Employed", "# Protein stability: a crystallographer's perspective\nAuthors: Deller, M. C.; Kong, L.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 72, 72\u201395 (IYCr crystallization series), 2016\n\n### 6.7. In situ proteolysis and 6.8. Thermal stability screening\n\nProteases present in a crystallization sample cause significant compositional instability over the experiment's timeframe, but this can be exploited. The approach began serendipitously when a *Penicillium* fungus growing in a crystallization drop cleaved about 200 residues from the yeast CPSF-100 protein and proved essential for successful crystallization. This prompted in situ proteolysis, in which samples are crystallized in the presence of a panel of proteases such as trypsin and chymotrypsin, later combined with mass spectrometry to identify fragments stable over the crystallization timeframe. Limited proteolysis also identifies stable domains: because the membrane protects transmembrane segments from cleavage, more compact membrane-protein forms can be generated, as in the stable 55 kDa PapC P-pilus construct used for structure determination. On thermal stability, a large Thermofluor study of 657 protein samples found that only 23% of proteins with Tm below 43\u00b0C yielded crystals, compared with 49% of those with Tm above 45\u00b0C; beyond 45\u00b0C, Tm loses predictive value, with no significant increase in crystallization frequency between 45 and 96\u00b0C. No known structural features correlate well with Tm, which is also highly solvent-dependent. Other melt-curve features matter: a high initial Thermofluor signal suggests exposed hydrophobic surface interacting with the dye and correlates with a lower crystallization rate (36.6%), and a wide denaturation peak indicates non-cooperative unfolding.\n\n### 6.9. Deuterium-exchange mass spectrometry (DXMS) and 6.10. Disulfide engineering\n\nRemoving flexible termini and low-complexity regions before crystallization is often desirable, and DXMS can guide where to cut. Proteins are exposed to deuterated solvent, deuterium exchanges with backbone amide protons at rates governed by solvent exposure and local secondary-structure integrity, and exchange is quantified as mass changes in proteolytic fragments \u2014 yielding a solution measure of conformational flexibility. DXMS has been applied successfully to structural genomics targets, guiding deletion of flexible regions that had prevented diffraction-quality crystals; Figure 5 shows a putative *Salmonella typhimurium* ethanolamine-utilization protein whose disordered N-terminal region was identified and trimmed, giving a structure of residues 98\u2013229 (PDB entry 2pyt). Its utility is limited by the lengths of proteolytic fragments obtainable, sometimes covering less than 50% of the sequence, but mapping exchange onto structure gives a three-dimensional flexibility map for engineering and mechanistic work. Disulfide engineering stabilizes proteins and can promote lattice formation: intermolecular disulfides in lattices gave rise to the phrase \"spontaneously polymerizing protein crystals\", and since symmetrical proteins such as homodimers crystallize more readily, engineered disulfides can dimerize monomers \u2014 \"synthetic symmetrization\" \u2014 as demonstrated with T4 lysozyme and used to crystallize CelA from *Thermotoga maritima*. Disulfide by Design 2.0 identifies suitable residues. A caveat is that disulfides can adopt multiple conformations and thereby promote flexibility, as with the bond linking the I-EGF1 and I-EGF2 domains of integrin \u03b22, which accommodates a hinge motion of more than 20 \u00c5.\n\n### 7. Future outlook\n\nGlobally, stability is the ability of a protein to maintain structure and function in a particular environment; under physiological conditions the sum of contributing forces must give a small negative \u0394G favouring the folded state. But not all proteins work in standard environments. Oxygen-sensitive enzymes such as FeFe hydrogenase must be considered under anaerobic conditions, which is costly and hard to scale \u2014 hence the effort to engineer oxygen-tolerant hydrogenases for industrial biological hydrogen production. Light-sensitive proteins such as opsins, photolyase and photosystems I and II, and the LOV domains of plants and algae, which change conformation and covalently bind FMN under blue light, illustrate how non-standard environments must enter any discussion of stability. Large fractions of the proteome are predicted to be unstructured, insoluble or aggregation-prone (IDPs, IIPs, APRs); crystallographers have habitually deleted such regions at the cloning stage, but given their medical importance this is no longer adequate. Microcrystallography, next-generation synchrotron sources and X-ray free-electron lasers (XFELs) are evolving to address them: XFELs enable femtosecond-timescale dynamics and require injectable microcrystalline material, exemplified by the synaptotagmin-1\u2013neuronal SNARE complex, while brighter, more coherent synchrotron beams will open study of dynamics on ever shorter timescales. The review closes by invoking the title of a paper by the late Fred Richards \u2014 *Protein stability: still an unsolved problem* \u2014 while noting that considerable progress has been made and that crystallography, NMR, cryo-EM and hybrid methods such as UED and DEER will be central.\n\n## Key tables\n\n*Table 1.* Compositional stability versus conformational stability: questions to ask when embarking on crystallization, and factors to investigate for problem proteins.", "# GitHub README: ButtFold\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ButtFold, 2026\n\n`tools/check_all.sh` runs seven gates. Four of them drive a real browser, because the failures that\nmatter here are the ones where every unit test passes and the page does nothing.\n\n| Gate | What it proves |\n|---|---|\n| `audit_wiring.py` | Every route, module, style, card and element id is reachable. Anything declared and never reached fails the build |\n| pytest, 87 tests | Routes, caching, the queue's caps and cache, the honesty strings, and the committed artefact's own assertions |\n| `node --test`, 62 tests | The WASM module against the CLI, the JS geometry ports against the Python, the sonifier against the Swift, the camera's interaction model, and the frame the browser keeps mid-fold against the one the baker keeps |\n| stage renders | A headless screenshot of the stage mid-fold is non-uniform, the three colour modes render differently, and the Play bar is above the fold at four common screen sizes |\n| drag, zoom, reframe | Real pointer input through the DevTools protocol reaches the camera |\n| sound | The score reaches a **running** `AudioContext` from a real click, the animation follows the audio clock, and the notes sounding mid-playback are drawn as chords on the structure and lit cells on the ribbon |\n| live fold, and the queue | A browser folds trp-cage to Q >= 0.95, and the droplet returns a fold and then serves it from cache. Both must be seen mid-fold: at least three distinct frame counts while the status still reads \"folding\", with the charts carrying a point per frame |\n\nThe animation between frames is a morph and the page says so. The 150 poses in a trajectory\nare 50,000 integration steps apart - far enough that an alpha carbon moves up to 30 \u00c5 between\ntwo of them - so a straight interpolation tears the chain (the worst CA-CA bond came out 97%\nshort) and is followed by a bond projection that puts it back. Nothing drawn between two\nframes is a state the model passed through, which is in the disclosure paragraph rather than\nleft implied.\n\nTwo assertions run at bake time and are repeated against the committed file, because they are the\nones that catch a trajectory which is not doing the thing the animation is about: the chain must\ncollapse (`Rg_end / Rg_start <= 0.8`) and it must not start folded (under 25% of contacts on the\nfirst frame). PhoneFold threw away a whole bake for want of them.\n\n## \ud83d\udcca Measured, not estimated\n\nFull detail with machines and dates in [`METRICS.md`](METRICS.md). Nothing is quoted anywhere in\nthis project until it is a measured row there.\n\n| | |\n|---|---|\n| trp-cage, folded live in Chrome | 11.2 s, final Q 1.000, 152 frames streamed |\n| the same fold, module alone, no frame building | 7.4 s (1.4x the native binary) |\n| trp-cage, folded on the droplet | 25.7 s under its CPU quota (16.6 s unconstrained) |\n| ubiquitin, on the droplet | 7 min 07 s, which sets the 76-residue cap |\n| WASM against native, forces | agree to 2.9e-15 relative, against a 1e-9 bar |\n| Sonifier.js against the shipped Swift | 15,536 notes, **0 differences** |\n| Live path against baked path | 300 frames, 5 fields each, **byte-identical** |\n| Baked gallery payload | 842 kB for six folds and 900 frames |\n\nOne number that changed a plan: the parity bar was written as \"RMSD within 0.1 \u00c5 of the native\nbuild\". Measured, the **same** build folding the **same** coil lands anywhere from 0.84 to 1.16 \u00c5\ndepending only on the random-force seed, so that bar is narrower than the measurement's own noise\nand no correct implementation can meet it. Q over the same five seeds was 0.993 every time, so Q\ncarries the weight and the RMSD bar is taken from the measured spread.\n\n## \u2705 To Do\n\nRoadmap for ButtFold, roughly in dependency order. Suggestions welcome.", "# Shedding Light on the Shadows: The Untapped Potential of the Dark Kinome\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-20 on marcdeller.com (categories: sb).*\n\n* **The dark kinome** comprises a significant portion of human kinases with poorly understood functions.\n* **These understudied kinases** represent a rich source of potential novel drug targets.\n* **Challenges in studying dark kinases** include low expression and difficult-to-characterize catalytic domains.\n* **Technological advancements** are now enabling researchers to explore the dark kinome more effectively.\n* **Targeting dark kinases** could lead to first-in-class therapeutics for various diseases.\n\n1. **The Dark Kinome Website:** <https://darkkinome.org>\n2. O\u2019Reilly, M., Bouzat, J. L., & Sicheri, F. (2017). **Drugging the undruggable kinome.** [Nature Chemical Biology, 13(10), 1003-1011](https://www.nature.com/articles/nchembio.2492).\n3. Skamnaki, V. T., Gotrik, M. R., & Taunton, J. (2021). **Illuminating the dark kinome with chemical probes.** [Nature Reviews Drug Discovery, 20(5), 339-360](https://www.nature.com/articles/s41573-020-00131-8).\n4. Edwards, A. M., et. al., (2011). **The dark kinome: structure and function.** [Current Opinion in Structural Biology, 21(5), 671-678](https://www.sciencedirect.com/science/article/pii/S0959440X1100088X).\n5. **The Promise of Dark Kinase Inhibitors**. [New England Journal of Medicine, 388(12), 1053-1056](https://www.nejm.org/doi/full/10.1056/NEJMe2215295).\n\nThe dark kinome is an exciting frontier in drug discovery due to its untapped potential. These understudied kinases, comprising a significant portion of the human kinome, offer opportunities to uncover novel therapeutic targets for diseases like cancer and immune disorders. Advances in tools and probes are beginning to illuminate their roles, making them promising candidates for innovative drug development.\n\nThe human kinome, a collection of over 500 protein kinases, plays a crucial role in regulating nearly every aspect of cellular life, from metabolism and proliferation to differentiation and apoptosis. These enzymes function by phosphorylating other proteins, acting as molecular switches that control signaling pathways. While a significant portion of the kinome has been extensively studied and targeted by successful therapies, a substantial fraction, often referred to as the \u201cdark kinome,\u201d remains poorly characterized.\n\nThese dark kinases, estimated to be around 170 in number, lack well-defined biological functions, known substrates, and selective chemical probes. This obscurity is primarily due to challenges in studying them with traditional biochemical and pharmacological approaches. Many dark kinases exhibit low expression levels, possess catalytic domains that are difficult to work with, or lack apparent regulatory mechanisms. However, this very lack of understanding presents a unique and compelling opportunity for drug discovery.\n\nThe therapeutic landscape is dominated by inhibitors targeting well-validated kinases. As resistance mechanisms emerge and the need for novel treatments grows, the dark kinome represents a vast, largely unexplored reservoir of potential drug targets. These kinases could be involved in disease pathways that are currently intractable with existing therapies. By elucidating their functions, we can potentially identify new intervention points for a wide range of diseases.\n\nRecent technological advancements are beginning to chip away at the darkness surrounding these enzymes. Improvements in genomics, proteomics, structural biology, and the development of innovative chemical biology tools are enabling researchers to investigate dark kinases with unprecedented resolution. For instance, CRISPR-based gene editing allows for the study of their cellular phenotypes, while advances in mass spectrometry can help identify their elusive substrates. Furthermore, novel approaches to probe design are yielding chemical tools that can selectively engage previously \u201cundruggable\u201d kinases.\n\nThe exploration of the dark kinome is not without its challenges. The lack of prior knowledge necessitates creative and often high-risk research strategies. However, the potential rewards are immense. Success in targeting a dark kinase could lead to the development of first-in-class therapeutics with unique mechanisms of action, offering new hope for patients with debilitating diseases. The journey into the dark kinome is an exciting one, promising to unveil fundamental biological insights and potentially revolutionize drug discovery.", "# GitHub README: AlphaFraud\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/AlphaFraud, 2026\n\n## \ud83d\udce6 Citation\n\nAlphaFraud is deposited on Zenodo as a citable software release under the MIT Licence: the\npipeline, the metric suite, the ribbon renderer and the Flask application as released. The\ncached structures and generated imagery are not included \u2014 they are regenerable from the\npublic APIs \u2014 and neither is the results database.\n\n[](https://doi.org/10.5281/zenodo.21808182)\n\n| | |\n|---|---|\n| Concept DOI (always latest) | [10.5281/zenodo.21808182](https://doi.org/10.5281/zenodo.21808182) |\n| This version (v0.1.0) | [10.5281/zenodo.21808183](https://doi.org/10.5281/zenodo.21808183) |\n\nEvery figure in the deposit's description was **read from the live instance's database**\non 2026-08-05 rather than copied from this README, which was carrying counts from an\nearlier week (96,827 entities and 82,796 analysed, against an actual 96,928 and 82,895).\nOn a sibling project six figures reached a Zenodo description by being copied from prose,\nand a DOI would have made them permanent.\n\n**Cite as:** Deller, M. C. (2026). *AlphaFraud: a weekly, automated audit of AlphaFold\naccuracy on freshly deposited human structures*. Version 0.1.0. Zenodo.\nhttps://doi.org/10.5281/zenodo.21808182\n\n## \ud83d\udcdd Licence\n\nReleased under the MIT Licence (see [LICENSE](LICENSE)).\n\nAttributing AlphaFraud does not discharge the obligation to the sources it derives from:\nthe [Protein Data Bank](https://www.rcsb.org) (CC0) and the\n[AlphaFold Protein Structure Database](https://alphafold.ebi.ac.uk) (CC BY 4.0).\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/AlphaFraud\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/AlphaFraud</a></td>\n</tr>\n</table>", "# Empowering QSAR Modeling: Optimal Data Practices and Cutting-Edge Tools\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-27 on marcdeller.com (categories: blog, sb).*\n\n* **Dataset Curation:** Assemble diverse, high-quality chemical\u2013activity pairs from public databases, applying filters to ensure uniform activity scaling.\n* **Descriptor Generation:** Use both established descriptors (e.g., RDKit, Mordred) and learned molecular embeddings for richer representations.\n* **Data Sizing:** Favor large datasets (thousands to tens of thousands of compounds) to enhance model robustness and reduce variance.\n* **Evaluation Schemes:** Implement realistic splitting (zero-shot, few-shot) tailored to VS and LO tasks for reliable performance estimates.\n* **Automation Pipelines:** Employ platforms like QSARtuna for standardized workflows covering ingestion, modeling, calibration, and uncertainty.\n\n* [Shahin et al., Tracking protein kinase targeting advances: integrating QSAR into machine learning for kinase-targeted drug discovery](https://pmc.ncbi.nlm.nih.gov/articles/PMC11980485/)\n* [Pradeep et al., Structure-based QSAR Models to Predict Repeat Dose Toxicity Points of Departure](https://pmc.ncbi.nlm.nih.gov/articles/PMC8128699/)\n* [Tian et al., Benchmarking compound activity prediction for real-world drug discovery applications](https://www.nature.com/articles/s42004-024-01204-4)\n* Mervin et al., QSARtuna: an automated QSAR modelling platform for molecular property prediction in drug design\n* [Burgoon et al., The database makes the poison: How the selection of datasets in QSAR models impacts toxicant prediction of higher tier endpoints](https://www.sciencedirect.com/science/article/pii/S0273230024001041)\n\nHarnessing the predictive power of Quantitative Structure\u2013Activity Relationship (QSAR) models hinges on **meticulous data curation**, **appropriate dataset sizing**, and leveraging **state-of-the-art platforms**. Success demands integrating robust descriptors, ensuring representative chemical diversity, and adopting automated pipelines that guarantee reproducibility. The following discussion explores these pillars, demonstrating how proper input data practices and modern QSAR tools elevate model reliability and interpretability.\n\nQSAR modeling begins with **comprehensive data collection** from curated repositories like ChEMBL and PubChem, followed by **rigorous preprocessing** to remove duplicates, normalize activity measures, and apply chemical filters (e.g., Lipinski\u2019s rules). Effective descriptor calculation\u2014from 2D physicochemical features to learned molecular embeddings\u2014forms the next cornerstone. Recent guidelines emphasize using large, diverse datasets (thousands to tens of thousands of compounds) to minimize overfitting and capture chemical space; for repeat-dose toxicity, aggregating all available studies (over 70,000 data points) yielded superior predictive power compared to partitioned subsets.\n\nData size not only influences model performance but also underpins **uncertainty estimation**. Ensemble methods like Random Forest and Support Vector Regression excel with ample data yet struggle on smaller sets, whereas deep learning architectures offer integrated feature extraction that scales with dataset size. A benchmark study delineated virtual screening (VS) versus lead optimization (LO) scenarios, advocating specialized data splitting schemes\u2014zero-shot and few-shot\u2014to mirror real-world applications and ensure robust generalization.\n\nAutomation frameworks such as **QSARtuna** encapsulate best practices into transparent pipelines: from data ingestion and deduplication to descriptor calculation, model calibration (e.g., VennABERS), and uncertainty quantification. Such platforms democratize QSAR development, enabling both novices and experts to generate reproducible models with state-of-the-art explainability. Meanwhile, hands-on tutorials illustrate implementing Graph Neural Networks for QSAR tasks, showcasing tools like PyTorch Geometric for molecular graph representations and best practices in hyperparameter tuning and cross-validation.\n\nUltimately, aligning **data size**, **descriptor quality**, and **automated workflows** fosters QSAR models that are both accurate and actionable\u2014poised to accelerate drug discovery endeavors across diverse chemical spaces.", "# The Invisible Invasion: How Microscopic Plastics Are Infiltrating Human Organ Systems\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nHepatic consequences include fibrosis promotion and lipid metabolism disruption. Chronic microplastic exposure activates hepatic stellate cells, promoting extracellular matrix deposition and progressive scarring. Simultaneously, these particles dysregulate fatty acid oxidation and synthesis pathways, contributing to non-alcoholic fatty liver disease progression.\u200b\n\nThe endocrine system faces dual threats from both plastic particles themselves and associated endocrine-disrupting chemicals (EDCs). These compounds interfere with hormone receptor binding, disrupting thyroid function, reproductive health, and metabolic regulation. Recent studies detecting microplastics in human follicular fluid and semen raise profound concerns about fertility implications and transgenerational effects.\u200b\n\nGastrointestinal tract effects extend beyond simple mechanical irritation. Microplastics fundamentally alter gut microbiota composition, reducing beneficial species while promoting pathogenic proliferation. This dysbiosis impairs short-chain fatty acid production, compromises intestinal barrier integrity, and triggers systemic inflammatory responses that propagate throughout multiple organ systems.\u200b\n\nThe mechanistic pathways underlying these diverse organ effects converge on several common themes: oxidative stress generation, mitochondrial dysfunction, inflammatory pathway activation, and cellular apoptosis induction. These particles act as biological Trojan horses, carrying adsorbed environmental toxins directly into tissues while simultaneously triggering innate immune responses that perpetuate tissue damage.\n\n| Concept | Description | Key Reference |\n| --- | --- | --- |\n| **Cardiovascular Disruption** | Microplastics accumulate in arterial plaques, causing oxidative stress, mitochondrial dysfunction, and increased risk of major adverse cardiovascular events | [PMC Study](https://pmc.ncbi.nlm.nih.gov/articles/PMC11918686/) |\n| **Neurological Toxicity** | Nanoplastics cross the blood-brain barrier, disrupting mitochondrial energy production and potentially contributing to neurodegenerative diseases | [Neuroscience News](https://neurosciencenews.com/nanoplastics-memory-neuroplasticity-299684/) |\n| **Renal Dysfunction** | Microplastics damage glomeruli and renal tubules through oxidative stress, inflammation, and activation of inflammatory pathways like NF-\u03baB | [PMC Kidney Study](https://pmc.ncbi.nlm.nih.gov/articles/PMC12500543/) |\n| **Hepatic Toxicity** | Chronic exposure to microplastics promotes liver fibrosis, lipid metabolism disruption, and activation of pro-inflammatory pathways | [Nature Study](https://www.nature.com/articles/s41598-025-86810-5) |\n| **Respiratory Impact** | Inhaled microplastics deposit in alveoli, triggering inflammation, oxidative stress, and potential translocation to other organs via bloodstream | [PMC Respiratory](https://pmc.ncbi.nlm.nih.gov/articles/PMC10826726/) |\n| **Endocrine Disruption** | Plastic particles carry endocrine-disrupting chemicals that interfere with hormone receptors, affecting thyroid, reproductive, and metabolic functions | [PMC Endocrine](https://pmc.ncbi.nlm.nih.gov/articles/PMC12249724/) |\n| **Gastrointestinal Effects** | Microplastics disrupt gut microbiota composition, impair intestinal barrier function, and reduce beneficial short-chain fatty acid production | [PMC Gut Study](https://pmc.ncbi.nlm.nih.gov/articles/PMC12351775/) |", "# ChatRDKit \u2013 Talk to RDKit for Chemoinformatics in plain language\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n**`Ollama at ... did not respond to model '...' within 180s`**\nThe model took too long to respond. Common causes: the model is still\nloading into memory (try again), the machine is under heavy load (e.g.\nanother `ollama pull` running), or the model is just slow for your\nhardware. Increase `/config timeout 300` or switch to a smaller/faster\nmodel.\n\n**`The model's response looked malformed or repetitive ... and was\ndiscarded`**\nSmall models occasionally loop or ramble instead of producing a short\nsnippet. chatRDKit detects this, discards the response, and does *not* add\nit to conversation history (so it can\u2019t poison later requests). Try again,\nrephrase, or switch to a larger model (`/config model qwen3.6:27b`).\n\n**`Refusing to run this snippet -- it contains ...`**\nThe model\u2019s response touched something outside RDKit (file I/O, `os`,\n`subprocess`, etc.). Rephrase the request \u2014 chatRDKit\u2019s `render()` helper\nalready handles all file output.\n\n**`AttributeError: module 'rdkit.Chem...' has no attribute '...'`**\nSmaller models sometimes invent plausible-but-wrong RDKit function names.\nThis is reported as a normal Python error without crashing the session \u2014\ntry rephrasing, or switch to `qwen3.6:27b` for better accuracy.\n\n**PDF output falls back to PNG**\nInstall `reportlab` (`pip install reportlab`) for PDF report support.\n\n---\n\n## 7. Extending\n\n* **More commands / better translations**: extend\n  `rdkit_reference.RDKIT_CHEAT_SHEET` with examples for operations you use\n  often.\n* **Richer session context**: `rdkit_reference.build_context()` is the place\n  to add more state (e.g. 3D conformers, reaction history).\n* **New output formats**: add a branch to `render.render()` alongside\n  `_write_html` / `_write_pdf`.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/chatRDKit](https://github.com/bellcheddar/chatRDKit) |", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n| Option | Default | Description |\n|---|---|---|\n| `--family` | every family with `Apo structure:` set | Restrict to one `Protein` family id |\n| `--target` | every target for the selected family | Restrict to one target stem |\n| `--out-dir` | alongside `boltz_input.md` | Where to write the CSV/HTML/PyMOL scripts |\n| `--phi-psi-threshold` | `30` (degrees) | Per-residue phi/psi delta above this is flagged |\n| `--dfg-distance-threshold` | `11.0` (Angstrom) | Separates the two DFG-Phe ring distances (ring to alphaC-Glu+4 Ca, ring to catalytic-Lys Ca) that classify DFG-in vs DFG-out |\n| `--alphac-distance-threshold` | `4.0` (Angstrom) | Catalytic-Lys NZ to alphaC-Glu carboxylate below this is an intact salt bridge, i.e. alphaC-in |\n| `--no-pymol` | off | Skip writing `.pml` session scripts |\n| `--refresh-cache` | off | Bypass the GPCRdb/KLIFS/PDBe disk cache for this run |\n\n## \ud83d\udee0\ufe0f Memory on Mac, and what we fixed in Boltz\n\n### Memory on Mac (unified-memory) hardware\n\nA real 4-chain GPCR+G-protein complex (~1250 combined residues/atoms) used **~65GB RAM on a\n64GB M1 Max** during testing and swap-thrashed for 20+ minutes with zero progress before\nbeing killed. Mitigations built in:\n\n- **`--mps-watermark`** sets `PYTORCH_MPS_HIGH_WATERMARK_RATIO`, capping how much memory MPS\n  allocates relative to the device's recommended maximum. At the default `1.0` an oversized\n  complex raises a clean MPS out-of-memory error instead of silently spilling into swap. **It\n  is a hard allocation ceiling, not a swap-avoidance dial.** Lowering it \"to be safe\" is the\n  wrong instinct: `0.7` on a 64GB M1 Max caps allocation at 36GB against a ~34GB requirement,\n  and every batch then OOMs immediately. Lower it only above a peak you have measured.\n- **`--workers` defaults to 0** and **`--max-parallel-samples` to 1**: both duplicate large\n  in-memory structures out of the same pool the model is using. Raising\n  `--max-parallel-samples` buys no throughput on an M1 Max either, where this stage is\n  compute-bound rather than memory-bound -- it only multiplies peak memory.\n- **`--max-msa-seqs` defaults to 4096** rather than Boltz's 8192, halving the co-evolution\n  feature block. A quality/memory trade rather than a free win, and one of the few levers that\n  measurably cuts peak memory on a large complex.\n- **`preflight`'s `memory_heuristic`** WARNs when a target's total residue+ligand-atom count\n  crosses `--memory-warn-tokens` (default **1500**), citing the data point above. A rough\n  heuristic, not a memory model, and it only blocks a run under `--strict`. Raised from 1000\n  because a GPCR + G-protein campaign runs at 1307-1333 tokens, so 1000 warned on all 26\n  targets and separated nothing: a warning that fires on everything carries no signal.\n- **Peak RSS is recorded** per completed target to `.boltzmaker_target_memory.jsonl`, so the\n  size check can eventually come from what this machine actually did rather than a hand-set\n  token threshold that separated nothing whether a target succeeded or OOM'd.\n- **`run`'s progress bar shows live memory** (RSS summed across the whole `boltz predict`\n  process tree), and warns if usage stays above 90% of system RAM for 60+ seconds with no new\n  completed target: a sign of thrashing, not progress.\n- **`run`/`all` auto-retries** (`--max-retries`, default 2) any target that doesn't complete,\n  isolating every still-incomplete target to its own single-target `boltz predict` invocation\n  from the first retry onward. A real 4-target cascade on `5ht2_gq_panel` (an OOM on 2 of 6\n  large targets run together crashed the shared affinity phase for 2 more that had already\n  succeeded) recovered cleanly this way, so a large campaign can be left unattended.\n- **That isolation is enforced against Boltz's *manifest***, not just the staging directory.\n  Staging one YAML is not enough: `boltz predict` filters an already-processed input out as\n  \"All inputs are already processed\", then rebuilds the manifest from every record in\n  `<out_dir>/boltz_results_*/processed/records/` and iterates that. A single-target retry\n  therefore re-ran the whole campaign in one process, and one target raising took every target\n  behind it down with it -- 20 invocations over 14.5 hours produced zero structures before this\n  was found. Each batch now parks the other records for its duration.\n\nIf a target still fails after every automatic retry, that's a real finding (this hardware may\nnot be viable for a complex that size), not something to force through.\n\n### \ud83e\ude7a What we fixed in Boltz, in plain English\n\nBoltz is excellent software, but a few of its assumptions do not hold on an Apple Silicon Mac,\nand two of them waste days rather than minutes. **Every one of these was found by measuring a\nreal campaign, not by reading the code.**\n\n**1. Memory that was never given back.** On running out of memory Boltz released its GPU memory\nby asking the NVIDIA graphics system to let go (`torch.cuda.empty_cache()`). There is no NVIDIA\nhardware in a Mac, so that did nothing and the memory stayed held, making each failure likelier\nthan the last: one campaign skipped eleven targets in a row. The patch adds\n`torch.mps.empty_cache()` alongside it, so an out-of-memory skip actually frees memory.\n\n**2. Memory held between targets.** Even when a target *succeeded* its working memory was kept\nrather than returned, so the next target started with the tank half full. Measured on a live\nrun: 47GB held with only 5.5GB in use, against a 55.7GB ceiling. Returning it after each target\nfrees about **28GB** -- the difference between the next target fitting and failing. Bookkeeping\nonly; no prediction changes.", "# Unlocking Membrane Protein Structures: The Crystallization Revolution Reshaping Drug Discovery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nBicelle crystallization represents another significant advancement, offering the native-like environment of lipidic methods with the practical handling advantages of traditional detergent approaches. Bicelles are small bilayer disks formed from mixtures of long-chain phospholipids and amphiphiles that remain liquid at low temperatures but develop gel-like consistency when warmed. This temperature-sensitive property enables researchers to mix proteins with bicelle components easily on ice, then move crystallization trials to 37\u00b0C where the gel phase appears to facilitate crystal growth. The method has proven successful across diverse targets, including the [human \u03b2\u2082-adrenergic receptor](https://www.nature.com/articles/s41392-024-01803-6) solved at 3.4 \u00c5 resolution, voltage-dependent anion channel at 2.3 \u00c5, and xanthorhodopsin at 1.9 \u00c5.\n\nThe integration of microcrystal electron diffraction with membrane protein crystallization has opened new possibilities for structures previously considered beyond reach. MicroED can determine [high-resolution structures from crystals](https://pubs.acs.org/doi/10.1021/acs.cgd.7b00519) orders of magnitude smaller than required for X-ray crystallography, taking advantage of electrons\u2019 stronger interaction with matter. Critically, MicroED works synergistically with both LCP and bicelle crystallization, which typically produce crystals in the 1-5 \u03bcm range unsuitable for conventional X-ray analysis. Recent innovations combining [cryo-focused ion beam milling with MicroED](https://www.nature.com/articles/s41467-023-36733-4) enable researchers to thin viscous LCP-embedded crystals to electron-transparent dimensions, while modified blotting procedures allow bicelle-grown crystals to be analyzed despite their viscous matrix.\n\nProtein engineering strategies have proven equally transformative, with fusion protein approaches dominating recent GPCR structure determination. The most successful strategy involves replacing the flexible third intracellular loop with [T4 lysozyme](https://pmc.ncbi.nlm.nih.gov/articles/PMC4408211/) or thermostabilized apocytochrome b\u2085\u2086\u2082RIL, compact stable proteins that increase polar surface area for crystal contacts while reducing conformational heterogeneity. Modified T4L variants, including a disulfide-stabilized version reducing interdomain flexibility and a minimal version removing the N-terminal domain, have enabled crystallization of challenging targets like the M3 muscarinic receptor, improving resolution from 3.4 \u00c5 with 70 crystals to 2.8 \u00c5 with untwinned crystals.\n\nThermostabilization through systematic mutagenesis offers an alternative stabilization strategy, exemplified by the [turkey \u03b2\u2081-adrenergic receptor](https://www.pnas.org/doi/10.1073/pnas.0711253105), where six point mutations increased melting temperature by 21\u00b0C and locked the receptor into the antagonist-bound conformation. This conformational thermostabilization approach addresses the fundamental challenge that GPCRs cycle through multiple conformational states, making crystallization of homogeneous populations difficult. Combined with fusion protein strategies and antibody fragment or nanobody incorporation, which further stabilize specific states and increase crystallization surface area, these engineering approaches have made previously impossible targets routine.\n\nThe rapid ascendance of single-particle cryo-electron microscopy has fundamentally altered the structural biology landscape for membrane proteins. Mean resolution for [cryo-EM membrane protein structures](https://pmc.ncbi.nlm.nih.gov/articles/PMC7856071/) improved from 4.2 \u00b1 0.5 \u00c5 in 2015 to 3.6 \u00b1 0.7 \u00c5 by 2019, driven by improvements in direct detectors, image processing algorithms, and sample preparation methods including amphipols and nanodiscs. Unlike X-ray crystallography, cryo-EM does not require crystals and enables direct visualization of detergent- or nanodisc-solubilized GPCRs in their native-like environments. This capability has proven especially valuable for capturing fully active GPCR-G protein complexes, which are too large and conformationally heterogeneous for traditional crystallographic approaches.\n\n[Statistical analyses of membrane protein structure determination](https://pmc.ncbi.nlm.nih.gov/articles/PMC7856071/) from 2010-2019 reveal dramatic shifts in both methodology and target selection. Channels increased from 29 to 149 structures, while transporters grew from 27 to 157 structures during this period. The proportion of respiratory complexes declined from 24% to 9% as the field\u2019s focus shifted toward pharmacologically relevant targets. Perhaps most striking, enzymes represented such substantial growth that a dedicated category became necessary, with proteases and other enzymes requiring separate classification from general \u201cother\u201d membrane proteins.\n\nLooking forward, the convergence of these techniques promises to sustain the exponential growth trajectory. The combination of improved membrane mimetic systems, advanced protein engineering, and complementary structural methods creates a powerful toolkit for tackling increasingly complex targets. As [automation and robotics](https://pmc.ncbi.nlm.nih.gov/articles/PMC3164297/) further streamline workflows, particularly for LCP crystallization which historically required specialized expertise, structural determination of membrane proteins continues its transformation from formidable challenge to accessible research tool. This democratization of membrane protein structural biology will undoubtedly accelerate drug discovery efforts targeting this critical class of pharmaceutical targets.", "# Unlocking Protein Structure Insights without a Lab: Harnessing CRO Services for Structural Biology Research\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe rapid pace of technological advancement complements CRO services. Methods like cryo-electron microscopy and next-generation X-ray crystallography are routinely integrated into CRO offerings. These partnerships are not only financially prudent but also scientifically strategic, providing researchers access to the latest instrumentation and specialist knowledge without historic upfront costs. CRO ecosystems often include collaborations with universities and industry consortia, securing access to synchrotrons and other critical infrastructure.\n\nIn summary, exploiting CRO partnerships for structural biology enables experimental protein structure determination with dramatically reduced capital burden and optimized experimental timelines. This approach democratizes access to high-quality protein structures essential for life science research and drug discovery, empowering even modestly funded labs to participate in cutting-edge molecular science. As CROs further innovate and expand their integrated service portfolios, the field of structural biology is poised for continued acceleration and wider impact.", "# GitHub README: AlphaFraud\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/AlphaFraud, 2026\n\n| Module | Responsibility |\n|---|---|\n| `pdb.py` | RCSB Search (new human protein entities) + Data GraphQL (UniProt mapping, dates, method) |\n| `afdb.py` | AlphaFold DB fetch: the blind model + PAE, choosing the fragment covering the entity's span |\n| `novelty.py` | max % identity to any pre-cutoff chain (so genuinely unseen sequences are flagged) |\n| `compare.py` | the full metric suite on the aligned span (TM, lDDT, GDT, RMSD, per-residue deviation, PAE honesty) |\n| `ribbon.py` | C\u03b1-deviation SVG + 3D-viewer coords + the AlphaFold \"ghost\" (vendored, offline) |\n| `annotate.py` / `analysis.py` | CATH / SCOP / citation enrichment and the `/analysis` snapshot (enrichment, clustering, PCA) |\n| `examples_auto.py` | the deterministic, no-LLM panel builder (match **or** failure framing) |\n| `calc_index.py` / `calculate.py` | the Calculate tab's strict autocomplete index + on-demand single-entity compute |\n| `db.py` / `report.py` / `webapp.py` | SQLite persistence, Plotly figure JSON, and the Flask routes |\n\n**The weekly pipeline** (`AlphaFraud.py run`, one entity at a time):\n\n```text\nfor entity in RCSB.search(human, protein, deposited_after = 2018-04-30):\n    if entity in db: continue\n    meta = RCSB.graphql(entity)                       # UniProt, dates, method, chains\n    if not meta.single_uniprot: skip(\"antibody / fusion / chimera\"); continue\n    af = AlphaFoldDB.model(meta.uniprot)              # the blind prediction (fragment covering the span)\n    if af is None:               skip(\"no AlphaFold model\"); continue\n    exp, model = download(entity), download(af)\n    m        = compare(exp, model)                    # TM, lDDT, GDT, RMSD, per-residue dev, PAE honesty\n    novelty  = max_identity_to_pre_cutoff(meta.sequence)\n    fraud    = mean_i( pLDDT_i/100 * clamp(dev_i, 0, 15\u00c5) / 15 )   # confidence-weighted error\n    confidently_wrong = (mean_pLDDT > 70) and (TM < 0.5)          # sure, and still wrong\n    ribbon.render(exp, model)                         # deviation SVG + coords + AF ghost\n    db.store(entity, m, novelty, fraud, confidently_wrong)\n\nanalysis.rebuild()            # /analysis enrichment snapshot\nrebuild_weekly_examples()     # auto 'example of the week' + weekly archive\ncalc_index.refresh(metas)     # keep the Calculate autocomplete index current\n```\n\n**The always-on web app** (gunicorn) reads the same SQLite and renders cached Plotly JSON:\n\n```text\nGET /            -> cumulative \"fraud quadrant\" dashboard (figures cached, short TTL)\nGET /entry/<id>  -> per-structure heatmaps, calibration, per-domain metrics\nGET /examples    -> positive control + auto example-of-the-week + curated deep-dives\nGET /calculate   -> type a human PDB id  ->  (reuse if known | compute on demand)  ->  panel\n```\n\n**The Calculate tab** reuses the very same `process_entity()` as the pipeline, run in an isolated subprocess so a heavy computation never blocks a web worker, and stored apart so it never pollutes the science:\n\n```text\nPOST /calculate/run { pdb }:\n    entity = resolve_human_entity(pdb)                # strict index, else a live RCSB check\n    if not qualifies:  return reason                 # not human / no single UniProt / no AF model\n    if entity already computed:  return \"ready\"      # cached -> instant\n    mark(entity, \"computing\"); spawn AlphaFraud.py calculate <entity>   # same pipeline, subprocess\n    return \"computing\"                               # the browser polls, then injects the panel\n```\n\n## \ud83c\udfaf Metric suite\n\nEvery matched structure is scored on the aligned residue span (all pure-Python except CATH, which is a lookup):\n\n| Family | Metrics |\n|---|---|\n| Global fold agreement | TM-score (4 normalisations), C\u03b1 / backbone / all-atom / core RMSD, GDT_TS, GDT_HA, MaxSub, structural overlap |\n| Local, superposition-free | lDDT (global and per-residue), contact-map Jaccard / precision / recall, distance-matrix difference, CAD-score |\n| Backbone and secondary structure | Q3 secondary-structure agreement, mean \u0394\u03c6 / \u0394\u03c8, radius-of-gyration difference |\n| Domains | Per-CATH-domain and per-PAE-cluster-domain TM and RMSD (exposes inter-domain hinge errors a global fit hides) |\n| Confidence calibration | pLDDT vs lDDT correlation, PAE vs observed-error correlation, overconfident-pair fraction |\n| Composite | FRAUD score (confidence-weighted error) and a confidently-wrong flag (mean pLDDT > 70 yet TM-score < 0.5) |\n\n## \ud83d\udccb Requirements\n\n- Python 3.11 or newer\n- The packages in `requirements.txt` (requests, tenacity, biotite, tmtools, numpy, scipy, pandas, plotly, flask, gunicorn, python-dotenv), all pip-installable with no compiled toolchain\n- Outbound HTTPS to `search.rcsb.org`, `data.rcsb.org`, `files.rcsb.org` and `alphafoldebi.ac.uk`\n- For deployment: a Linux droplet with root/SSH access, a domain, and DNS pointed at it\n\n## \ud83d\udd27 Installation\n\n```bash\ngit clone https://github.com/bellcheddar/AlphaFraud.git\ncd AlphaFraud\npython3 -m venv .venv\nsource .venv/bin/activate\npip install -r requirements.txt\npython3 AlphaFraud.py init          # create dirs, DB schema, vendor plotly.min.js\n```\n\n## \ud83d\ude80 Usage\n\n```bash\n# Preview discovery for a window without writing anything\npython3 AlphaFraud.py run --since 2024-01-01 --until 2024-01-10 --limit 6 --dry-run\n\n# Process a window for real (omit --since to resume from the last run)\npython3 AlphaFraud.py run --since 2024-01-01 --limit 50\n\n# Backfill a historical range (full metrics on everything)\npython3 AlphaFraud.py backfill --from 2023-01-01 --to 2023-12-31\n\n# Two-tier backfill: fast TM-score screen on everything, full metrics only on the\n# disagreements (TM below --tm-threshold). Far cheaper on time, disk and API load.\npython3 AlphaFraud.py backfill --from 2023-01-01 --to 2023-12-31 --two-tier --tm-threshold 0.7\n\n# Two-tier backfill of the ENTIRE post-cutoff archive (~96k entities, monthly chunks,\n# resumable). Screens all, fully analyses only the tail AlphaFold got wrong.\npython3 AlphaFraud.py backfill --all --two-tier", "# toppdblx.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n## toppdblx.mdeller.com/\n\n## What everyone tries\n\nHow often each reagent appears across every parsed condition. This is the\nquery-independent frequency prior, and on this corpus it beats homology retrieval at\nevery level tested.\n\n**Read this as popularity, not efficacy.** PEG 3350 leads because it is in\neveryone's screens, not because it works best. The Protein Data Bank records only\nsuccesses, so nothing here says what would have failed.\n\nFilter reagents\n\nChemical class\nAll classes\n\n| Reagent | Class | Conditions | Share |\n| --- | --- | --- | --- |\n\nNo reagent matches that filter.\n\n## What a reagent is used with\n\nPick a reagent to see which others appear in the same condition. Co-occurrence is\ncomputed over the 40 most frequent reagents.\n\nReagent\n\n###\n\n| Appears with | Conditions | Share |\n| --- | --- | --- |\n\n### Against the prior\n\nBars compare a partner's frequency inside these conditions with its\nfrequency across the whole corpus. Similar heights mean the pairing tells you nothing.\n\n## Precipitant classes\n\nEvery condition is placed in one of seven classes, or left Unclassified with the reason\nrecorded. Refusing to guess is deliberate.\n\n### Distribution\n\nConditions per class across the whole corpus.\n\n### Counts\n\nUnclassified is a recorded outcome, not a gap in the data.\n\n| Class | Conditions | Share |\n| --- | --- | --- |\n\n## Commercial screens\n\nConditions whose component set matches a published screen well. Formulations were\ntranscribed verbatim from vendor support materials.\n\n**A match is not proof the screen was used.** Many deposited conditions are\noptimised from a hit rather than the hit itself, so a component-set match with different\nconcentrations is flagged rather than claimed.\n\n| Screen | Matched conditions | Share of matches |\n| --- | --- | --- |\n\n## What this is, and what it is not\n\nAn exploration tool over a parsed dataset. It is deliberately not a recommender.\n\n### Why there is no \"suggest a condition\" button\n\nHomology retrieval was tested\nagainst a query-independent frequency prior at three levels of granularity, on two\nsplit thresholds, under three definitions of ground truth. **The prior won every\ntime.** Homologues get screened with the same popular screens as everything\nelse, so there is little protein-specific signal to retrieve. A recommender here would\ndress up \"what everyone tries\" as a prediction.\n\n### Where the numbers come from\n\n| Source | Components | What it is |\n| --- | --- | --- |\n\nEvery component in the released data carries a\n`parser` field, so the deterministic subset can be recovered exactly.", "# 2025 \u2013 Bicyclic Compounds as Inhibitors of WRN\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-12-11 on marcdeller.com (categories: patents).*\n\n* [Bicyclic Compounds as Inhibitors of WRN](https://patents.google.com/patent/US20250333419A1/en?oq=US20250333419A1)\n\n* **Synthetic Lethality:** [MSI-H tumors require WRN helicase](https://www.nature.com/articles/s41568-019-0147-3) to resolve secondary DNA structures caused by repeat expansions, a dependency absent in healthy cells.\n* **Structural Innovation:** New [bicyclic compounds of Formula (I)](https://patents.google.com/patent/US20250333419A1/en?oq=US20250333419A1) have been disclosed that inhibit WRN ATPase and helicase activities.\n* **Overcoming Resistance:** WRN inhibition remains effective in [tumors resistant to checkpoint blockade](https://aacrjournals.org/cancerdiscovery/article/11/8/1923/666232/Werner-Helicase-Is-a-Synthetic-Lethal), offering a salvage therapy option.\n* **Mechanism of Action:** These inhibitors likely function by [locking WRN in an inactive state](https://www.nature.com/articles/s41586-024-07350-y) or inducing its degradation, leading to double-strand breaks.\n\n1. **Bicyclic compounds as inhibitors of wrn:** [Barbosa et al., US Patent Application, 2025](https://patents.google.com/patent/US20250333419A1/en?oq=US20250333419A1)\n2. **Discovery of WRN inhibitor HRO761 with synthetic lethality in MSI cancers:** [Ferri et al., Nature, 2024](https://www.nature.com/articles/s41586-024-07350-y)\n3. **Werner Helicase Is a Synthetic-Lethal Vulnerability in Mismatch Repair-Deficient Colorectal Cancer Refractory to Targeted Therapies, Chemotherapy, and Immunotherapy:** [Picco et al., Cancer Discovery, 2021](https://aacrjournals.org/cancerdiscovery/article/11/8/1923/666232/Werner-Helicase-Is-a-Synthetic-Lethal)\n4. **Prioritizing synthetic lethal targets with functional genomics:** [Behan et al., Nature, 2019](https://www.nature.com/articles/s41568-019-0147-3)\n\n**A structural biology perspective on targeting the DNA damage response in microsatellite instability-high cancers**\n\nMicrosatellite instability-high (MSI-H) cancers possess a unique vulnerability: they are dependent on the Werner syndrome helicase (WRN) for survival. This dependency arises because mismatch repair-deficient cells accumulate massive expansions of TA dinucleotide repeats, which form non-B DNA cruciform structures that stall replication forks. WRN is the only helicase capable of resolving these structures to prevent catastrophic double-strand breaks. While immune checkpoint inhibitors have improved care for MSI-H patients, resistance remains a significant clinical hurdle, necessitating orthogonal therapeutic strategies. [Picco et al. demonstrated](https://aacrjournals.org/cancerdiscovery/article/11/8/1923/666232/Werner-Helicase-Is-a-Synthetic-Lethal) that WRN dependency persists even in models resistant to standard-of-care immunotherapies, validating it as a high-priority target for drug discovery.\n\nBuilding on this biological rationale, recent disclosures have highlighted the development of novel small molecules designed to exploit this synthetic lethality. A newly published patent application describes a series of [bicyclic compounds of Formula (I)](https://patents.google.com/patent/US20250333419A1/en?oq=US20250333419A1) that function as potent inhibitors of WRN. These molecules, developed by scientists at Incyte Corporation, are engineered to inhibit the ATPase and helicase activities of the enzyme. The structural innovation lies in the bicyclic scaffold, which likely provides the necessary specificity to bind the WRN helicase domain\u2014a mechanism recently validated structurally by [Ferri et al. with other chemotypes](https://www.nature.com/articles/s41586-024-07350-y). By locking WRN in an inactive conformation or promoting its degradation, these inhibitors force MSI-H cells into mitotic catastrophe while sparing healthy microsatellite-stable (MSS) cells.\n\nThe advancement of these bicyclic inhibitors represents a critical step toward precision oncology for patients who have exhausted immunotherapeutic options. The ability of these compounds to induce DNA damage selectively in the MSI-H context aligns with the broader industry movement toward biomarker-driven synthetic lethal therapies. As these candidates progress, they offer a promising avenue to expand the therapeutic arsenal against endometrial, colorectal, and gastric cancers, potentially serving as effective monotherapies or partners for combination regimens.", "# 2025 \u2013 Cultivating Human Skills and Decision-Making\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-21 on marcdeller.com (categories: certificates).*\n\n* Enhance emotional intelligence to lead diverse teams effectively.\n* Develop nuanced interpersonal skills for stakeholder engagement.\n* Apply critical thinking to solve emerging industry challenges.\n* Integrate generative AI tools into strategic decision-making.\n* Balance data insights with human factors in leadership.\n* Demonstrate a holistic skill set valued by employers.\n* Increase adaptability in high-pressure environments.\n* Foster a culture of continuous improvement and ethical AI usage.\n\n1. **Emotional Intelligence and Interpersonal** \u2013 Investing in Human Skills in the Age of AI: [Certificate](https://www.linkedin.com/learning/certificates/c8b38adb7299bbb40c98b244da91d1eb8b915de8e5caeadc8d3f73f6fce6a9a0)\n2. **Critical Thinking and Adaptation** \u2013 Building an Adaptability Mindset in the Age of AI: [Certificate](https://www.linkedin.com/learning/certificates/8fd265595da5c6a811a98b77161dfbe5a887b53439fa7ca014db31e58ab8a0f4)\n3. **Strategic AI-Enhanced Decision-Making** \u2013 Strengthen Your Decision-Making with Generative AI: [Certificate](https://www.linkedin.com/learning/certificates/55ccac83c777ec51ceedf007ef7dfdd4ac54fd988b23eb1497d0732b93c48c12)\n4. **Ethical Leadership in AI** \u2013 Investing in Human Skills in the Age of AI: [Certificate](https://www.linkedin.com/learning/certificates/c8b38adb7299bbb40c98b244da91d1eb8b915de8e5caeadc8d3f73f6fce6a9a0)\n\nTechnical prowess must be balanced with human-centered competencies. Certifications like *Investing in Human Skills in the Age of AI* foster emotional intelligence, critical thinking, and interpersonal prowess\u2014qualities indispensable in collaborative settings. Meanwhile, *Strengthen Your Decision-Making with Generative AI* equips you with frameworks to integrate AI-enhanced scenarios into strategic planning. These credentials affirm your ability to navigate complex human and technological interfaces, making you an invaluable asset as organizations prioritize empathy, ethics, and evidence-based leadership.", "# Engineering Life: How Synthetic Biology Is Tackling Climate Change and Disease\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* Synthetic biology enables the design of novel proteins to address major global challenges in climate, industry, and medicine.\n* By using tools like CRISPR, scientists can create custom enzymes to degrade pollutants, capture carbon, and make industrial processes more sustainable.\n* This engineering approach is also driving medical innovation by creating highly specific, stable proteins for targeted therapies and diagnostics.\n\n1. **Protein Engineering Business Analysis Report 2025:** Market to More than Double to Reach $7.8 Billion by 2030:\n   <https://www.globenewswire.com/news-release/2025/04/03/3055322/28124/en/Protein-Engineering-Business-Analysis-Report-2025-Market-to-More-than-Double-to-Reach-7-8-Billion-by-2030-Rising-Demand-for-Protein-Based-Biosensors-and-Diagnostic-Tools.html>\n2. **10 Top Protein Engineering Companies and Startups to Watch in 2025:** Spotlight on disruptive impact:\n   <https://www.startus-insights.com/innovators-guide/protein-engineering-companies/>\n3. **2025 Protein Engineering Conference GRC:** Covering topics like directed evolution and AI for protein engineering:\n   <https://www.grc.org/protein-engineering-conference/2025/>\n\nThe challenges facing our world, from climate change to untreatable diseases, are immense. Historically, we have tackled these problems with conventional tools: industrial chemistry for pollution and traditional pharmaceuticals for health. However, these approaches are often inefficient, produce harmful byproducts, and struggle to keep pace with the complexity of the problems. We have reached a point where incremental improvements are not enough. We need transformative solutions that are smarter, cleaner, and more adaptable\u2014solutions inspired by nature itself.\n\nSynthetic biology is providing these solutions by allowing us to engineer proteins with entirely new functions. By combining protein engineering with powerful gene-editing tools like CRISPR, scientists are no longer limited to the proteins that exist in nature. They can now design and build bespoke biological machinery to meet specific needs. In environmental science, this means creating enzymes that can degrade plastic pollutants in our oceans or capture carbon dioxide directly from the atmosphere. In industry, it means replacing harsh chemical processes with sustainable biocatalysts that reduce energy consumption and waste, for everything from manufacturing textiles to producing biofuels.\n\nThe synergy between synthetic biology and protein engineering is also revolutionizing medicine. By designing proteins that can target diseases with surgical precision, we are advancing the fields of oncology and immunology. Furthermore, the ability to create proteins that are stable under extreme conditions is leading to more robust diagnostics and therapeutics. This convergence of biology and engineering allows us to design biological solutions for virtually any application imaginable. We are not just observing life; we are learning to write it, creating a more sustainable and healthier future in the process.", "# 2014 \u2013 Approaches to automated protein crystal harvesting\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-09-03 on marcdeller.com (categories: papers).*\n\n* [Link to Paper](https://marcdeller.com/wp-content/uploads/2025/09/f-70-00133-1.pdf)\n\n* **Precision Robotics**: Six-axis anthropomorphic robots achieve submicron positioning accuracy for crystals as small as 10 micrometers\n* **Smart Microgrippers**: MEMS-fabricated tools provide controlled pressure application while enabling direct X-ray data collection\n* **Acoustic Processing**: Non-contact ejection systems operate at hundreds of droplets per second for high-throughput applications\n* **AI Integration**: Machine learning algorithms achieve 95% accuracy in automatic crystal detection and scoring\n* **Pipeline Automation**: Integrated systems connect crystallization screening directly to data collection workflows\n* **Market Growth**: Industry expansion to $3.84 billion by 2033 driven by pharmaceutical automation demands\n\n1. Approaches to automated protein crystal harvesting: [Deller, M.C., et al., Acta Crystallographica Section F, 2014.](https://pmc.ncbi.nlm.nih.gov/articles/PMC3936438/pdf/f-70-00133.pdf)\n2. Operator-assisted harvesting of protein crystals using a universal micromanipulation robot: [Viola, R., et al., Journal of Applied Crystallography, 2007.](https://journals.iucr.org/paper?kk5010)\n3. Robotic Equipment for Automated Crystal Harvesting using a six-axis robot: [Heidari Khajepour, M.Y., et al., Acta Crystallographica Section D, 2013.](https://pubmed.ncbi.nlm.nih.gov/23519413/)\n4. Acoustic methods for high-throughput protein crystal mounting: [Soares, A.S., et al., Acta Crystallographica Section D, 2013.](https://pmc.ncbi.nlm.nih.gov/articles/PMC3747951/)\n5. High throughput screening using acoustic droplet ejection: [Teplitsky, E., et al., Journal of Structural Biology, 2015.](https://pubmed.ncbi.nlm.nih.gov/26027487/)\n6. Advances in macromolecular crystallography at the Photon Factory: [Hiraki, M., et al., Journal of Synchrotron Radiation, 2025.](https://journals.iucr.org/paper?he5685)\n\n**Revolutionary Advances in Automated Protein Crystal Harvesting: Bridging the Gap in High-Throughput Crystallography**\n\nThe field of structural biology stands at a pivotal moment where manual protein crystal harvesting represents the last remaining bottleneck in an otherwise highly automated crystallographic pipeline. This critical gap between automated crystallization and data collection has spurred remarkable technological innovations that promise to revolutionize how scientists approach macromolecular structure determination. The challenge lies not merely in mechanizing the process, but in developing systems sophisticated enough to handle crystals measuring as small as 10 micrometers while maintaining their structural integrity throughout harvesting, cryoprotection, and mounting procedures.\n\nModern approaches to automated protein crystal harvesting encompass a diverse array of technologies, each addressing specific limitations of manual techniques. Semi-automated systems have emerged as practical solutions that combine human expertise in crystal identification with robotic precision in manipulation. The Universal Micromanipulation Robot (UMR) exemplifies this hybrid approach, utilizing a six-axis anthropomorphic robot arm capable of harvesting crystals with submicron precision. This system has successfully demonstrated its capabilities by producing high-quality structural data, including a 1.5 \u00c5 resolution structure of bovine trypsin that matches the quality of manually harvested specimens.\n\nComplementing microtool-based approaches, microgripper technologies have gained prominence through systems like the Robotic Equipment for Automated Crystal Harvesting (REACH). These two-fingered micro-gripping devices, fabricated using microelectromechanical systems (MEMS) technology, offer unprecedented control over contact pressure applied to fragile crystals. The biocompatible SU-8 material used in these microgrippers provides sufficient flexibility to prevent crystal damage while maintaining low X-ray background scattering, enabling direct data collection while crystals remain secured in the gripper.\n\nRecent developments in 2024-2025 have introduced cutting-edge methodologies that push the boundaries of what\u2019s possible in crystal harvesting. Acoustic droplet ejection (ADE) technology has matured into a powerful tool capable of ejecting crystal-containing droplets at rates exceeding several hundred per second. This non-mechanical approach eliminates physical contact with crystals entirely, making it particularly valuable for handling radiation-sensitive samples and microcrystal slurries destined for free-electron laser facilities. The precision of ADE systems allows for the screening of up to 1,728 distinct chemical compounds with protein crystals on a single microplate, dramatically accelerating drug discovery and fragment screening workflows.\n\nOptical trapping technologies represent another frontier in contactless crystal manipulation. Both conventional optical tweezers and fiber optical tweezers have demonstrated the ability to trap and manipulate protein crystals without mechanical intervention. While these systems require substantial investment in specialized equipment, their potential for damage-free manipulation of even the most delicate specimens makes them invaluable for certain applications.\n\nThe emergence of artificial intelligence and machine learning has transformed crystal detection capabilities, addressing one of the fundamental challenges in achieving full automation. Advanced imaging systems now incorporate AI-based software capable of automatically scoring crystallization droplets with error rates below 5%. These systems utilize extensive training datasets containing over 50,000 data points to reliably distinguish protein crystals from visual artifacts and precipitates.", "# GitHub README: PocketTide\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/PocketTide, 2026\n\n> **An interactive dashboard for GLP-1R structures.**\n\n    \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/PocketTide\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/PocketTide</a></td>\n</tr>\n</table>\n\n---\n\nPocketTide is a browser-based dashboard for exploring structures of the glucagon-like peptide-1 receptor (GLP-1R), one of the most important and intensively pursued targets in metabolic disease. It brings the growing body of GLP-1R structural data into a single, navigable view so that structures, their key features, and their differences can be surveyed at a glance.\n\nWhy it matters: GLP-1R sits at the centre of the incretin-mimetic revolution in diabetes and obesity therapeutics, and its structural landscape (peptide-bound, small-molecule-bound, G-protein-coupled, in multiple conformational states) is expanding rapidly. Keeping that picture organised is hard when the relevant structures are scattered across the PDB. PocketTide is useful for structure-based drug design teams who need to triage GLP-1R structures quickly, compare binding modes and receptor states, and orient a project against the available structural evidence, all from a single self-contained HTML file with no install and no build step.\n\n---\n\n## \u2728 What it does\n\n| Capability | Description |\n|---|---|\n| \ud83d\uddc2\ufe0f **Structure browser** | Survey GLP-1R structures in one organised dashboard view |\n| \ud83d\udd2c **Feature summary** | Surface key attributes per structure (method, ligand, receptor state) |\n| \u2696\ufe0f **Comparison** | Place structures side by side to compare binding modes and conformations |\n| \ud83c\udfaf **Target focus** | Purpose-built around GLP-1R rather than a generic structure viewer |\n\n> \u2139\ufe0f This README describes PocketTide at the capability level. For the exact panels and data fields in your build, open the app and explore the live interface.\n\n---\n\n## \ud83d\ude80 Running locally\n\nPocketTide is delivered as a single self-contained HTML file. The simplest way to use it is to open the file directly in a modern browser:\n\n```bash\nopen PocketTide.html        # macOS\nstart PocketTide.html       # Windows\nxdg-open PocketTide.html    # Linux\n```\n\nIf the dashboard pulls structural data from external APIs (e.g. RCSB PDB), browser file-origin restrictions may block those requests when opened as a `file://` URL. In that case, serve it from a lightweight local HTTP server:\n\n```bash\n# Python 3\npython -m http.server 8080\n\n# Node.js\nnpx serve .\n```\n\nThen navigate to `http://localhost:8080/PocketTide.html`.\n\n---\n\n## \ud83d\udd27 Tech stack\n\n| Layer | Technology |\n|---|---|\n| UI | Vanilla HTML + CSS, no frameworks |\n| Logic | Vanilla JavaScript (ES6+) |\n| Distribution | Single self-contained HTML file, no build step |\n| Data | GLP-1R structures (PDB) |\n\n---\n\n## \ud83e\udded Context\n\nPocketTide is part of a wider portfolio of structure-focused tools and dashboards. It pairs naturally with structure-based drug design work on GLP-1R and related class B GPCR targets.\n\n---\n\n## \ud83d\udcc4 Licence\n\nSee repository for licence terms.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/PocketTide\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/PocketTide</a></td>\n</tr>\n</table>", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\n**Step 6. N-(2-(2-ethyl-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide.** To the crude solution of N-(2-(2-ethyl-4-oxo-10-((2-(trimethylsilyl)ethoxy)methyl)-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide (1.11 g, 2.28 mmol) in CH2Cl2 (23 mL) was added TFA (12 mL) at 25 \u00b0C. The reaction mixture was then allowed to stir at this temperature for 2 h. Upon completion, the reaction contents were concentrated directly, azeotroped with MeCN (10 mL) 3\u00d7, redissolved in EtOAc, washed 3\u00d7 with NaHCO3, dried over Na2SO4, and concentrated. Purification by flash column chromatography (0-100% hex/EtOAc) then provided the desired product (550 mg, 67% over two steps). LC-MS calculated for C21H19N4O2 (M+H)+: m/z=359.1; found 359.2.\n\n**Step 7. N-(2-(2-Ethyl-10-(2-((4-fluorophenyl)amino)-2-oxoethyl)-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide.**\n\n*Intermediate A. 2-Bromo-N-(2-chloro-4-fluorophenyl)acetamide.* To a solution of 2-chloro-4-fluoroaniline (0.055 mL, 0.50 mmol) in CH2Cl2 (4 mL) was added triethylamine (0.083 mL, 0.60 mmol) and 2-bromoacetyl bromide (100 mg, 0.50 mmol). The reaction was allowed to stir for 1 h. Upon completion, the reaction contents were concentrated directly to produce the crude product (120 mg, 91% yield). LC-MS calculated for C8H7BrClFNO (M+H)+: m/z=265.9, 267.9; found 265.9, 267.9.\n\nTo a solution of N-(2-(2-ethyl-4-oxo-1,4-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide (15 mg, 0.042 mmol) in DMF (0.5 mL) was added diisopropylethylamine (0.022 mL, 0.13 mmol), followed by 2-bromo-N-(4-fluorophenyl)acetamide (Intermediate A, 14.6 mg, 0.06 mmol). The reaction mixture was then stirred for 16 h. Upon completion, the reaction contents were acidified with TFA, diluted with MeCN (5 mL), filtered, and purified by prep-LCMS (XBridge C18 column, eluting with a gradient of acetonitrile/water containing 0.1% TFA, at flow rate of 60 mL/min) to give the product as a TFA salt (8.0 mg, 31% yield). LC-MS calculated for C29H25FN5O3 (M+H)+: m/z=510.2; found 510.2. 1H NMR (400 MHz, DMSO-d6) \u03b4 10.58 (s, 1H), 8.99 (s, 1H), 8.48 (d, J=8.0 Hz, 1H), 8.07 (d, J=8.3 Hz, 1H), 7.71 (d, J=8.1 Hz, 1H), 7.65-7.58 (m, 2H), 7.57-7.51 (m, 1H), 7.44-7.33 (m, 2H), 7.25-7.13 (m, 4H), 6.44 (dd, J=17.0, 10.1 Hz, 1H), 6.17 (dd, J=17.0, 2.2 Hz, 1H), 5.62 (dd, J=10.1, 2.2 Hz, 1H), 5.33-5.17 (m, 2H), 2.24 (q, J=7.4 Hz, 2H), 1.01 (t, J=7.5 Hz, 3H). 19F NMR (376 MHz, DMSO-d6) \u03b4 \u2212118.77, \u221276.50.\n\n### Table 1\n\nExamples 2-10 in Table 1 were prepared according to the procedures described in Example 1 as TFA salts, using appropriately substituted starting materials instead of 2-bromo-N-(4-fluorophenyl)acetamide at Step 7.\n\n| Ex. No. | Name | LCMS [M+H]+ |\n|---|---|---|\n| 2 | N-(2-(2-Ethyl-4-oxo-10-(2-oxo-2-(phenylamino)ethyl)-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide | 492.2 |\n| 3 | N-(2-(2-Ethyl-4-oxo-10-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide | 560.2 |\n| 4 | N-(2-(10-(2-((2-Chloro-4-(trifluoromethyl)phenyl)amino)-2-oxoethyl)-2-ethyl-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide | 594.2 |\n| 5 | N-(2-(2-Ethyl-10-(2-((4-ethylphenyl)amino)-2-oxoethyl)-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide | 520.2 |\n| 6 | N-(2-(10-(2-((2-Chloro-4-fluorophenyl)amino)-2-oxoethyl)-2-ethyl-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide | 544.2 |\n| 7 | N-(2-(2-Ethyl-4-oxo-10-(2-oxo-2-((3-(trifluoromethyl)phenyl)amino)ethyl)-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide | 560.2 |\n| 8 | N-(2-(10-(2-((3,4-Difluorophenyl)amino)-2-oxoethyl)-2-ethyl-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide | 528.2 |\n| 9 | N-(2-(2-Ethyl-10-(2-((4-fluoro-2-methylphenyl)amino)-2-oxoethyl)-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide | 524.2 |\n| 10 | N-(2-(10-(2-(Cyclohexylamino)-2-oxoethyl)-2-ethyl-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide | 498.2 |\n\n*(1H NMR data at 400 MHz in DMSO-d6 are reported in the published table for Examples 2-9; the full shift lists are omitted here for brevity.)*\n\n### Table 2\n\nExamples 11-14 in Table 2 were prepared according to the procedures described in Example 1 as TFA salts, using appropriately substituted starting materials instead of (2-aminophenyl)boronic acid at Step 4.\n\n| Ex. No. | Name | LCMS [M+H]+ |\n|---|---|---|\n| 11 | Methyl 4-acrylamido-3-(2-ethyl-10-(2-((4-fluorophenyl)amino)-2-oxoethyl)-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)benzoate | 568.2 |\n| 12 | 4-Acrylamido-3-(2-ethyl-10-(2-((4-fluorophenyl)amino)-2-oxoethyl)-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)-N-(2-hydroxyethyl)benzamide | 597.2 |\n| 13 | 4-Acrylamido-3-(2-ethyl-10-(2-((4-fluorophenyl)amino)-2-oxoethyl)-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)-N-(2-methoxyethyl)benzamide | 611.2 |\n| 14 | N-(2-(2-Ethyl-10-(2-((4-fluorophenyl)amino)-2-oxoethyl)-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)-4-(1-hydroxy-3-azabicyclo[3.1.0]hexane-3-carbonyl)phenyl)acrylamide | 635.2 |\n| 15 | N-(5-Cyano-2-(2-ethyl-4-oxo-10-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide | 585.2 |\n| 16 | N-(4-(2-Ethyl-4-oxo-10-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)-1-methyl-1H-pyrazol-3-yl)acrylamide | 564.2 |\n\n### Table 3\n\nExamples 17-20 in Table 3 were prepared according to the procedures described in Example 1 as TFA salts, using appropriately substituted starting materials instead of 1H-benzo[d]imidazol-2-amine at Step 1, and/or using appropriately substituted starting materials instead of 2-bromo-N-(4-fluorophenyl)acetamide at Step 7.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics; supervisor Prof. E. Yvonne Jones), 1999\n\nInitial crystals were obtained within 1\u20133 days at 22 \u00b0C using the hanging drop vapour diffusion method with a custom salt/PEG screen. Crystals formed in drops consisting of 2 \u03bcl of protein solution at ~8 mg/ml plus 2 \u03bcl of reservoir solution consisting of 30 % (w/v) PEG 35000 and 0.25 M sodium acetate, pH 7.5. Crystal size was improved by decreasing the ratio of precipitant to protein. The largest crystals measuring approximately 0.2 \u00d7 0.2 \u00d7 0.1 mm3 were obtained from drops consisting of 0.5 \u03bcl of precipitant and 1 \u03bcl of protein at 8 mg/ml (see Figure 5-2 (a)). The crystals showed ordered Bragg diffraction to 2.2 \u00c5 when exposed to synchrotron radiation (see Figure 5-2 (c)). Preliminary processing showed that the crystals belong to the orthorhombic spacegroup P212121.\n\nA second crystal form was obtained over a period of several months from 30 % (w/v) PEG 4000, 0.1 M Tris-HCl and 0.2 M sodium acetate, pH 8.5 using the hanging drop vapour diffusion method. Crystals measuring approximately 0.1 \u00d7 0.1 \u00d7 0.1 mm3 were obtained from drops consisting of 1 \u03bcl of precipitant and 1 \u03bcl of protein solution at 14 mg/ml (see Figure 5-2 (b)). The crystals displayed ordered Bragg diffraction to 2.5 \u00c5 when exposed to synchrotron radiation. Preliminary processing showed the crystals belong to the spacegroup I222 or I212121.\n\n*Figure 5-2.* Figure showing (a) P212121 and (b) I222 crystal forms of \u0394188_OSM with bar representing 0.2 mm. Panel (c) shows resulting diffraction pattern of the P212121 crystal form with bar showing high resolution limit of detector at 2.43 \u00c5. Data to 2.6 \u00c5 were collected from single propane cooled crystals on beam line BL19 of the ESRF, Grenoble, France. The I222 crystal form was grown in 30 % (w/v) PEG 4000, 0.1 M Tris-HCl and 0.20 M sodium acetate, pH 8.5. The P212121 form was grown in 30 % (w/v) PEG 35000 and 0.25 M sodium acetate, pH 7.5. Crystals were grown from protein solution at 8 mg/ml to 14 mg/ml, using hanging drop vapour diffusion.\n\n#### 5.3.2 Crystal Mounting and Cryocrystallography\n\nOptimum flash cooling of the P212121 crystal form was carried out as follows; crystals were removed from the mother liquor and briefly passed through a cryoprotectant solution consisting of a 2 \u03bcl drop of 30 % (w/v) PEG 35000 and 0.25 M sodium acetate, pH 7.5 overlaid with 2 \u03bcl of 30 % (w/v) PEG 8000 and 0.25 M sodium acetate, pH 7.5. Crystals were cooled directly in liquid propane and placed in a nitrogen stream at 100 K and data recorded.\n\n#### 5.3.3 Data Collection and Processing\n\nThe two crystal forms of \u0394188_OSM showed ordered Bragg diffraction when exposed to synchrotron radiation. Statistics relevant to data collection are shown below in Table 5-2. Initial processing with the program DENZO (Otwinowski and Minor, 1997) suggested the crystals belong to the primitive orthorhombic pointgroup 222. Scaling and addition of partial reflections was carried out using the program SCALEPACK (Otwinowski and Minor, 1997). The pattern of systematically absent odd reflections on indices h00, 0k0 and 00l, suggested that the spacegroup has three 2-fold screw axes, so data were finally merged in the spacegroup P212121. Low or negative I/\u03c3(I) values for all systematic absences, in the resultant SCALEPACK output, confirmed P212121 as the correct spacegroup assignment. The crystals have unit cell dimensions of a = 35.8 \u00c5, b = 53.1 \u00c5, c = 106.8 \u00c5 and \u03b1 = \u03b2 = \u03b3 = 90\u00b0 (see Table 5-2). The volume of the unit cell is approximately 203025 \u00c53 and the theoretical molecular weight of the protein is 21.4 kDa. Assuming a standard specific volume of 1.23 cm3/g this suggests one molecule per asymmetric unit and a crystal solvent content of ~48.8 % (Matthews coefficient, Vm = 2.4 \u00c53/Da). From Table 5-2 it can be seen that native data collected from these crystals is of good quality with an I/\u03c3(I) value in the outer resolution shell (2.25\u20132.20 \u00c5) of 4.3. Completeness in the outer resolution shell is 98.2 %.\n\nThe second crystal form was processed in the body centred orthorhombic spacegroup I222 with unit cell dimensions a = 52.9 \u00c5, b = 62.0 \u00c5, c = 109.4 \u00c5 and \u03b1 = \u03b2 = \u03b3 = 90\u00b0 (see Table 5-2). Spacegroups I222 and I212121 are indistinguishable by diffraction intensities and correct assignment as I222 was only confirmed during molecular replacement (see 5.3.6). The volume of the unit cell is approximately 358810 \u00c53 and the theoretical molecular weight of the protein is 21.4 kDa. Assuming a standard specific volume of 1.23 cm3/g this suggests one molecule per asymmetric unit and a crystal solvent content of ~41.4 % (Matthews coefficient, Vm = 2.1 \u00c53/Da). From Table 5-2 it can be seen that native data collected from these crystals is of poorer quality than the P212121 crystal form with an I/\u03c3(I) value in the outer resolution shell (2.56\u20132.50 \u00c5) of 2.4. Completeness in the outer resolution shell is 88.8 %.", "# GitHub README: CODSWALLOP\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/CODSWALLOP, 2026\n\n> **The PDB, summarised: every entry for a protein family, cross-referenced in ninety seconds.**\n\n[](https://codswallop.mdeller.com)                  \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>App</b></td><td><a href=\"https://codswallop.mdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">codswallop.mdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/CODSWALLOP\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/CODSWALLOP</a></td>\n</tr>\n</table>\n\n---\n\n**C**omparative **O**verview of **D**omains, **S**tructures, **W**hole-families, **A**lignments, **L**igands, **L**ineages, **O**rganisms & **P**DB entries.\n\nA structural biologist has just been handed a protein family and told to get up to speed. CODSWALLOP does an afternoon's work of manual PDB trawling in ninety seconds: every entry, every construct, every ligand, every crystallisation condition, cross-referenced and interactive.\n\n**Why it matters:** the RCSB and PDBe landing pages are entry-centric. They answer \"what is 4XYZ?\" very well, and nobody answers the question a working structural biologist actually asks, which is \"what do the 87 structures of this family collectively tell me, and what should I do differently?\". CODSWALLOP is family-centric: it treats the family as the unit of analysis, so the answer to \"which construct crystallises best\" or \"which residues has nobody ever put in a construct\" is one glance rather than an afternoon of tab-juggling. It is useful for: starting a new target, designing a construct, choosing which deposited structure to trust, planning a crystallisation screen, and writing the structural-biology section of a grant.\n\nLive at **[codswallop.mdeller.com](https://codswallop.mdeller.com)**.\n\n---\n\n## \ud83e\udded What makes it different\n\n| # | Differentiator | Why it matters | Phase |\n|---|---|---|---|\n| 1 | **Construct diffing** | Aligns every deposited SEQRES against the UniProt canonical to expose tags, fusion partners, truncations, surface-entropy-reduction and thermostabilising mutations. This is what actually gets a crystal, and it is buried in mmCIF nobody reads | 2 |\n| 2 | **Global cross-filtering** | One selection state shared by every panel. Click a ligand and the map, the entry list, the crystallisation scatter and the coverage plot all filter together | 1 |\n| 3 | **Crystallisation intelligence** | Parses `_exptl_crystal_grow` across the whole family into structured conditions, so you can see what actually worked. Feeds the Top96 crystallisation predictor | 3 |\n| 4 | **Coverage census** | Which residues has nobody ever seen? A family-wide map answering a construct-design question in one glance | 1 (constructs), 2 (density) |\n| 5 | **Interaction fingerprints** | PLIP run family-wide, so binding-site contacts are comparable across entries rather than one-off | 3 |\n\n## \u2728 Features (phases 1 to 3 live, phase 4 landing)", "# The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\nAuthors: Span, E.A.; Suess, D.L.M.; Deller, M.C.; Britt, R.D.; Marletta, M.A.\nSource: ACS Chemical Biology, 2017\n\nPMO crystal structures have revealed that secondary sphere residues form conserved active-site H-bonding networks with buffer and solvent molecules.[15\u201317,23] One early study found that a PMO with a second-sphere glutamine-to-leucine substitution failed to stimulate cellulase activity compared with the wild type protein.[6] However, the conclusions of that study are somewhat incomplete as neither Cu nor reductant was present in the assay, except possibly as a contaminant. Herein, we investigate the role of the secondary coordination sphere in PMO mechanism via structural, bioinformatic, biochemical, and spectroscopic methods. We term the PMO used for these studies MtPMO3*, as it is a member of the C1-oxidizing PMO3* subtype[24] cloned from the filamentous ascomycete Myceliophthora thermophila. The crystal structure of MtPMO3* highlights H-bonding patterns in the secondary sphere. Conserved H-bonding motifs are identified, and a panel of site-specific variants is constructed. Wild type MtPMO3* and H-bonding variants are assayed for cellulose activity and O2 consumption. Additionally, the electron paramagnetic resonance (EPR) spectra of the wild type and variant proteins are utilized for mechanistic insight.\n\n## Results and Discussion\n\n### Structure of MtPMO3*\n\nThe X-ray crystal structure of MtPMO3* was determined to 2.45 \u00c5 resolution and contained six noncrystallographic symmetry-related molecules in the asymmetric unit (i.e., chains A\u2013F). Like other PMO structures, it exhibits a \u03b2-sandwich fold formed by nine antiparallel \u03b2-strands interspersed with sizable loop regions (Figure 1a). The Cu active site of MtPMO3* shares the primary coordination sphere features of other fungal PMOs, including the histidine brace (H1 and H75), axial tyrosine residue (Y169), and N-terminal histidine methylation (Figure 1b). The tyrosyl O atom is positioned 2.7 \u00c5 from the Cu ion, which falls within the range of Y169(O)\u2013Cu distances observed in crystal structures of other AA9 enzymes. These distances are too long to be formal bonds even with a Jahn\u2013Teller distortion,[25] and it is unclear whether this elongation is a feature of the Cu(II) coordination sphere or represents a loss of the Y169(O) ligand due to photoreduction during data collection. Additional solvent-derived ligands are not observed in the active site shown in Figure 1 (MtPMO3* chain C), which may result from a lack of strongly identifiable water content at the structure resolution, or again from Cu being in the reduced state.\n\nThe secondary coordination sphere of MtPMO3* comprises residues that form a network of H-bonds around the primary coordinating ligands. These include T74, which H-bonds with the N-terminal amino group of H1 as well as ligands in the axial solvent-facing position; H161, which is positioned to H-bond with equatorial solvent-facing ligands; and Q167, which also H-bonds with equatorial solvent ligands and with Y169 (Figure 1b). The H-bonding network is similar to those found in other fungal, cellulose-active PMO structures with ligands derived from the solvent or crystallization buffer (Figure 2a). In chains C and E, H161 and Q167 engage in H-bonding interactions with an aspartate residue (D26) from a symmetry-related molecule within the crystal. See the Supporting Information for more details.\n\n*Figure 1.* (a) Crystal structure of MtPMO3* in cartoon representation with primary and secondary coordination sphere residues as sticks. All \u03b2-strands are numbered in order of primary sequence. Relevant loop features are highlighted in lavender (L2), pink (LS), teal (L8), and silver (LC). All other carbon atoms are shown in yellow, nitrogen atoms in blue, and oxygen atoms in red. Copper ions are depicted as a brown sphere. (b) MtPMO3* active site showing primary (N\u03b5-Me-H1, H75, Y169) and secondary (T74, H161, Q167) sphere residues with mesh electron density map contoured to 1.5 \u03c3. H161 and Q167 are positioned to H-bond with ligands in the solvent-facing equatorial position. Q167 H-bonds with the axial tyrosine ligand. T74 H-bonds with the N-terminal amino group of H1 and forms a bond with an axial solvent ligand when present. Hydrogen-bonding distances are shown in \u00c5ngstroms (\u00c5).\n\n### Identification of Conserved H-bonding Motifs in PMOs\n\nA bioinformatics analysis combining sequence identity and structural alignment revealed conserved H-bonding networks in all fungal PMO families. All cellulose-active (AA9) enzymes contain the H-Xn-Q-X-Y motif observed in MtPMO3*, whereas chitin-active (AA11) and starch-active (AA13) enzymes present N-X2-E-X-Y and Q-X-Q-X-Y motifs, respectively (Figure 2b). In AA9 enzymes, the spacer region formed by loop L8, between the histidine and glutamine residues, is typically eight residues in length (n = 8); however, this region is shortened to five residues (n = 5) in MtPMO3* and other predicted PMO3* enzymes. In crystal structures of chitin- and starch-active enzymes, the active-site H-bonding motifs share similar coordinates with the H-Xn-Q-X-Y motif in the cellulose-active enzymes with respect to the Cu center. Additionally, they retain the H-bond with the axial tyrosine residue as well as the H-bonds with ligands presented in the equatorial solvent-facing positions. Similar H-bonding networks are also present in bacterial PMOs (Figure 2b). Deep conservation of this H-bonding network across the PMO superfamily suggests that it performs a central role in PMO catalysis, regardless of substrate specificity. Putative mechanistic functions include stabilizing Cu-bound O2 species or substrate, perturbing the pKa of an intermediate species, and mediating proton transfer reactions. This network could also tune the redox properties of the Cu site.", "# Marc C. Deller, D.Phil. \u2014 R\u00e9sum\u00e9 (September 2026)\nAuthors: Deller, M.C.\nSource: Personal r\u00e9sum\u00e9, 2026\n\n**Tagline:** Top-Performing Structural & Computational Biology Leader \u2014 Director of Structural & Computational Biology, Protein Science Leader, CRO Partnerships, AI & Therapeutic Innovation, Drug Candidate Validation, Structure-Based Drug Discovery, Computer-Aided Drug Discovery, Expert in X-ray Crystallography & Cryo-EM, Expert in Protein Production & Gene-to-Structure Workflows.\n\n**Headline numbers:** 400+ Protein Structures. 6+ FDA INDs. 2x Wet Labs Built. 7+ Patents. 60+ Publications. Vibe Coder.\n\n**Location:** Wilmington, Delaware, United States. **Phone:** (302) 358-6093. **Email:** marc@marcdeller.com. **Website:** https://marcdeller.com. **LinkedIn:** https://linkedin.com/in/marccdeller.\n\n## Work Experience\n\n### Scientific Technical Advisor \u2014 DeepCovalent Inc.\n\n**Dates:** October 2025 \u2013 present.\n\n- Reports to the CEO & Founder; increased throughput by 50% by implementing physics-based Computer-Aided Drug Discovery (CADD) pipelines, supporting patent applications and accelerating structure-based hit-to-lead validation for AI-driven drug designs.\n- Advises leadership on small-molecule discovery platforms, optimizing in silico screening, cheminformatics and validation workflows to improve GPCR-targeting drug candidate identification and validation.\n- Guides protein science CRO and vendor oversight, engagement and experimental validation strategy, translating computational predictions into biophysical, biochemical and cell-based assays.\n\n### Chief Scientific Officer & Co-Founder \u2014 Elora Therapeutics\n\n**Dates:** June 2025 \u2013 present.\n\n- Co-founded a mission-driven biotechnology startup in a fast-paced pre-seed environment focused on developing PET-degrading enzymes with potential impact for 2M+ people.\n- Prepared 10+ grant and funding applications while building partnerships with protein science CROs, research organizations and venture capital groups to support launch, resource planning and scientific platform growth.\n- Led scientific strategy in an ambiguous startup environment, advancing therapeutic innovation and establishing the foundation for novel intellectual property and patent applications.\n- Developed nonprofit and commercial partnerships to expand research capacity and accelerate translation of enzymatic microplastic-remediation technologies into biotechnology applications.\n\n### Research Fellow, Structural Biology \u2014 Incyte Pharmaceuticals\n\n**Dates:** March 2022 \u2013 December 2024.\n\n- Leveraged expertise in X-ray crystallography, Cryo-EM, generative AI, structure-based drug discovery (SBDD), CADD and computational biology to generate 100+ protein complexes supporting medicinal chemistry across 10+ programs, accelerating hit-to-lead efforts in oncology and immunology.\n- Mentored 5+ scientists in matrix organizations spanning drug discovery, pharmacology, structural biology and biochemistry.\n- Built and managed 10+ CRO and vendor partnerships supporting drug discovery, structural biology and biochemistry strategies while overseeing dynamic portfolios, stage-gates and strategic roadmaps.\n- Supported 6+ FDA IND filings and 7+ patents by advancing intellectual property for assets including the antibody mutCalR and small-molecule inhibitors targeting KRAS-G12D, WRN, CDK2/4, JAK1/2 and FGFR2/3, including the discovery of new mechanisms of action (MoA).\n- Supported 5+ new drug targets, TPD (targeted protein degradation), kinases and helicase programs, and multiple peer-reviewed publications, while building and scaling automated wet-lab workflows for protein science and high-throughput experimentation.\n- Contributed to biomarker development strategies by characterizing protein\u2013drug interactions and elucidating the structural basis of mechanism of action.\n\n### Associate Director, Structural Biology \u2014 Incyte Pharmaceuticals\n\n**Dates:** March 2018 \u2013 March 2022.\n\n- Reported to the VP of the Applied Technology Group; led design, build and implementation of an $8M+ structural biology platform integrating protein production, protein science, CRO resources, X-ray crystallography, Cryo-EM, liquid handling, NMR, MS, HDX, SPR, ITC and biophysical analysis, accelerating hit-to-lead conversion.\n- Managed a $1.5M budget while directing a multi-team experimental organization across 10+ programs, optimizing infrastructure and cloud-based data processing to improve project efficiency 4-fold.\n- Increased productivity 4-fold and mentored 20+ scientists by building an inclusive, high-ownership culture.\n\n### Director & Founder, Macromolecular Structure Knowledge Center \u2014 Stanford University\n\n**Dates:** June 2015 \u2013 March 2018.\n\n- Reported to the head of the ChEM-H Institute; founded a wet-lab service center dedicated to protein production and structural analysis.\n- Directed a $1.0M operating budget while managing 10+ user research projects and timelines.\n- Generated $60,000 in annual revenue; supported research that produced 6 novel protein structures, contributed to 4 grants, and enabled 3 peer-reviewed publications.\n- Equipped 70+ researchers with advanced protein-characterization tools, driving platform adoption and synchrotron utilization at SSRL.\n- Supported users from 10+ countries, building an inclusive research environment and collaborative culture.\n\n### Target Manager, Joint Center for Structural Genomics \u2014 Scripps Research Institute\n\n**Dates:** August 2006 \u2013 June 2015.\n\n- Reported to Prof. Ian Wilson; applied high-throughput protein production, crystallization and salvage techniques to increase NIH structural-genomics throughput by 10%; solved 70+ novel protein structures; and co-authored 40+ publications through workflow innovation, assay troubleshooting and mentorship \u2014 **providing part of the structural data foundations later used by AlphaFold**.\n\n## Prior Experience\n\n- **Pfizer** \u2014 Principal Scientist, Drug Discovery, Kinase Center of Excellence, Structural Biology.\n- **Yale University School of Medicine** \u2014 Postdoctoral Research Fellow, JAK/STAT Protein Structure.\n\n## Education", "# pants.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n| Enzyme | Topt | Muts | %ID | Cleft | Structure | Notes |\n| --- | --- | --- | --- | --- | --- | --- |\n| [IsPETase](/enzyme/IsPETase) wild type | 40 | \u2014 | 100.0 | 21.3 | [6EQE](https://www.rcsb.org/structure/6EQE)[5XJH](https://www.rcsb.org/structure/5XJH)[6ANE](https://www.rcsb.org/structure/6ANE) | Wild type; weak on crystalline PET |\n| [DepoPETase](/enzyme/DepoPETase) | 50 | 7 | 97.6 | 19.5 | predicted | 7 mutations; melting temperature +23.3 \u00b0C, ~1407x product [Shi et al. 2023, Angew. Chem. Int. Ed.](https://doi.org/10.1002/anie.202218390) |\n| [DuraPETase](/enzyme/DuraPETase) | 37 | 10 | 96.5 | 23.9 | predicted | 10 mutations; +31 \u00b0C thermostability, ~300x activity [Cui et al. 2021, ACS Catal.](https://doi.org/10.1021/acscatal.0c05126) |\n| [FAST-PETase](/enzyme/FAST-PETase) | 50 | 5 | 98.3 | 22.2 | [7SH6](https://www.rcsb.org/structure/7SH6)[8J45](https://www.rcsb.org/structure/8J45) | 38x activity over ThermoPETase, its scaffold; 33.8 mM monomers in 96 h [Lu et al. 2022, Nature (MutCompute)](https://doi.org/10.1038/s41586-022-04599-z) |\n| [N212A/K233C/S282C variant](/enzyme/FAST-PETase-N212A/K233C/S282C) | \u2014 | 7 | 97.6 | 21.9 | [9LMS](https://www.rcsb.org/structure/9LMS) | Disulfide-stabilised FAST-PETase; C233-C282 crosslinks the same pair Z1-PETase uses [Li et al. 2025, Int. J. Biol. Macromol.](https://doi.org/10.1016/j.ijbiomac.2025.145862) |\n| [HotPETase](/enzyme/HotPETase) | 62.5 | 21 | 92.0 | 17.5 | [7QVH](https://www.rcsb.org/structure/7QVH) | 21 mutations; melting temperature 82.5 \u00b0C [Bell et al. 2022, Nat. Catal.](https://doi.org/10.1038/s41929-022-00821-3) |\n| [IsPETase-W159H/S238F](/enzyme/IsPETase-W159H/S238F) | \u2014 | 2 | 99.3 | 22.1 | [6EQF](https://www.rcsb.org/structure/6EQF) | The narrowed-cleft double mutant that showed PETase's cleft is wider than a cutinase's [Austin et al. 2018, PNAS](https://doi.org/10.1073/pnas.1718804115) |\n| [ThermoPETase](/enzyme/ThermoPETase) | 50 | 3 | 99.0 | 19.5 | predicted | 3 mutations; the thermostabilised scaffold FAST-PETase was built on [Son et al. 2019, ACS Catal.](https://doi.org/10.1021/acscatal.9b00568) |\n| [Z1-PETase](/enzyme/Z1-PETase) | 30 | 13 | 95.5 | 19.6 | [8H5K](https://www.rcsb.org/structure/8H5K) | 13 mutations, two engineered disulfides; 20x soluble expression yield [J. Hazard. Mater. 2023](https://doi.org/10.1016/j.jhazmat.2023.132297) |\n| [LCC](/enzyme/LCC) wild type | 70 | \u2014 | 100.0 | 24.2 | [4EB0](https://www.rcsb.org/structure/4EB0) | Leaf-branch compost metagenome cutinase; parent of the industrial ICCG variant |\n| [LCC-ICCG](/enzyme/LCC-ICCG) | 68.5 | 4 | 98.6 | \u2014 | [6THT](https://www.rcsb.org/structure/6THT) | ~90% depolymerisation of pretreated PET in under 10 h at 200 g/kg, 72 \u00b0C [Tournier et al. 2020, Nature](https://doi.org/10.1038/s41586-020-2149-4) |\n| [LCC-A2](/enzyme/LCC-A2) | 78 | 6 | 98.0 | 18.8 | predicted | LCC-ICCG plus H218Y/N248D; >90% depolymerisation at 200 g/kg in 3.3 h [Orr et al. 2024, Biotechnol. J.](https://doi.org/10.1002/biot.202400021) |\n| [BhrPETase](/enzyme/BhrPETase) wild type | \u2014 | \u2014 | 100.0 | 17.0 | AlphaFold | Bacterium HR29 hydrolase; the parent TurboPETase was engineered from |\n| [TurboPETase](/enzyme/TurboPETase) | 65 | 8 | 97.3 | 8.7\u2020 | predicted | 98.2% depolymerisation at 200 g/kg in 8 h [Cui et al. 2024, Nat. Commun. 15:1417](https://doi.org/10.1038/s41467-024-45662-9) |\n| [Cut190](/enzyme/Cut190) wild type | \u2014 | \u2014 | 100.0 | 16.6 | [4WFI](https://www.rcsb.org/structure/4WFI)[4WFJ](https://www.rcsb.org/structure/4WFJ)[4WFK](https://www.rcsb.org/structure/4WFK)[5ZNO](https://www.rcsb.org/structure/5ZNO)[5ZRQ](https://www.rcsb.org/structure/5ZRQ)[5ZRR](https://www.rcsb.org/structure/5ZRR) | Saccharomonospora viridis AHK190 cutinase; calcium-dependent stabilisation |\n| [Cut190**SS](/enzyme/Cut190**SS) | 65 | 7 | 97.7 | 16.4 | [7CTR](https://www.rcsb.org/structure/7CTR) | Disulfide-stabilised; calcium-dependent conformational switching [Oda et al. 2018, Appl. Microbiol. Biotechnol.](https://doi.org/10.1007/s00253-018-9374-x) |\n| [TfCut2](/enzyme/TfCut2) wild type | 65 | \u2014 | 100.0 | 22.4 | [4CG1](https://www.rcsb.org/structure/4CG1) | Thermobifida fusca cutinase 2; thermophilic industrial lineage |\n| [HGMP01](/enzyme/HGMP01) wild type | 40 | \u2014 | 100.0 | 23.7 | predicted | Human gut metagenome; hydrolyses PET nanoparticles and outperformed every other gut candidate |\n| [HGMP02](/enzyme/HGMP02) wild type | \u2014 | \u2014 | 100.0 | 25.6 | predicted | Human gut metagenome polyesterase, assayed on PET nanoparticles |\n| [HGMP03](/enzyme/HGMP03) wild type | \u2014 | \u2014 | 100.0 | 23.3 | predicted | Human gut metagenome polyesterase, assayed on PET nanoparticles |\n| [HGMP04](/enzyme/HGMP04) wild type | \u2014 | \u2014 | 100.0 | 18.7 | predicted | Human gut metagenome polyesterase, assayed on PET nanoparticles |\n| [HGMP05](/enzyme/HGMP05) wild type | \u2014 | \u2014 | 100.0 | 24.5 | predicted | Human gut metagenome polyesterase, assayed on PET nanoparticles |\n| [Kutbu-PETase](/enzyme/Kutbu-PETase) wild type | \u2014 | \u2014 | 100.0 | 20.2 | [8YTW](https://www.rcsb.org/structure/8YTW)[8YTZ](https://www.rcsb.org/structure/8YTZ) | Wild-type actinomycete PETase that melts at 88.2 degC, among the most thermostable natural PET hydrolases reported |\n| [MHETase](/enzyme/MHETase) wild type | \u2014 | \u2014 | 100.0 | 24.3 | [6QGA](https://www.rcsb.org/structure/6QGA)[6QZ3](https://www.rcsb.org/structure/6QZ3) | Second enzyme of the pathway, hydrolysing MHET to TPA and EG. Tannase group, not the PETase family |\n| [Micpa-PETase](/enzyme/Micpa-PETase) wild type | \u2014 | \u2014 | 100.0 | 25.4 | [8YTU](https://www.rcsb.org/structure/8YTU) | Wild-type actinomycete PETase and the most active enzyme in the Science 2025 landscape screen, at 1,506 uM product release |", "# The Enzyme Toolbox\u2014Alternatives to PETase for Microplastic Breakdown\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-15 on marcdeller.com (categories: blog, microplastics).*\n\n* Cutinases, MHETases, carboxylesterases, and lipases expand substrate scope.\n* Fusion proteins (CBMs) and hybrids enhance surface binding and catalytic rates.\n* Genome mining discovers new enzymes from extreme environments.\n* Multi-enzyme approaches optimize efficacy for diverse plastics.\n* Greater enzyme diversity strengthens prospects for medical microplastic removal.\n\n1. [Current advances in the structural biology and molecular engineering of PETase (2023)](https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2023.1263996/full)\n2. [PETase serum paper (2024)](https://pubs.rsc.org/en/content/articlepdf/2024/fd/d4fd00014e)\n3. [PET hydrolases and alternatives overview (2022)](https://pmc.ncbi.nlm.nih.gov/articles/PMC9603852/)\n4. [Advances in MHETase structure/function (2024)](https://www.chndoi.org/Resolution/Handler?doi=10.13345/j.cjb.230791)\n5. [Characterization of PETase from marine sponge (2017)](https://pubs.acs.org/doi/10.1021/acssuschemeng.5c00737)\n\nThe fight against microplastics extends beyond PETase. Cutinases, carboxylesterases, lipases, and MHETases provide expanded substrate flexibility and catalytic power. Cutinases, like Cut190 and LCC, operate robustly under diverse industrial conditions and degrade polyesters, including PET. MHETase complements PETase by breaking down its intermediate product, MHET, into benign monomers, completing the depolymerization pathway.\n\nEnzyme hybrids and fusion proteins introduce carbohydrate-binding modules (CBMs) or hydrophobins, enhancing substrate recognition and surface affinity. Genome mining continues to uncover new variants from environmental and marine sources, broadening the arsenal for plastic degradation. These alternative or synergistic enzymes could be customized for specific therapeutic needs, considering the types and prevalence of microplastics found in the human body.\n\nFor medical applications, selecting the right mix of enzymes\u2014for example, combining PETase with cutinase or engineering multi-functional fusions\u2014improves efficiency against the complex cocktail of microplastics present in serum and organs. Engineered variants promise broader coverage and improved safety profiles.", "# Citation Giants: How the Most-Cited Papers of the 21st Century Are Reshaping Science\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Concept | Description | Key References |\n| --- | --- | --- |\n| Deep Learning Architectures | Neural network frameworks enabling AI breakthroughs | [He, K., et al., CVPR](https://www.cv-foundation.org/openaccess/content_cvpr_2016/html/He_Deep_Residual_Learning_CVPR_2016_paper.html) |\n| COVID-19 Clinical Studies | Foundational characterization of pandemic disease | [Guan, W., et al., NEJM](https://www.nejm.org/doi/full/10.1056/NEJMoa2002032) |\n| Research Methodologies | Qualitative and quantitative analysis frameworks | [Braun, V., et al., Qualitative Research in Psychology](https://www.tandfonline.com/doi/abs/10.1191/1478088706qp063oa) |\n| Computational Biology Tools | Software enabling genomic and proteomic analysis | [Love, M.I., et al., Genome Biology](https://genomebiology.biomedcentral.com/articles/10.1186/s13059-014-0550-8) |\n| Cancer Biology Frameworks | Foundational concepts defining cancer characteristics | [Hanahan, D., et al., Cell](https://www.cell.com/fulltext/S0092-8674(00)81683-9) |\n| Structural Biology Software | Crystallography analysis programs | [Sheldrick, G.M., et al., Acta Crystallographica](https://journals.iucr.org/a/issues/2008/01/00/sc5010/index.html) |", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nCell culture was essentially carried out as in 4.2.1.1 above, except wildtype 293T cells were grown in DMEM medium (Gibco) supplemented with 10 % FCS (LabTech International), 50 IU/ml penicillin, 50 \u03bcg/ml streptomycin, 1 mM pyruvate, 2 mM glutamate and 4500 mg/ml glucose. Typically, 162 cm\u00b2 tissue culture flasks (Costar) containing 30 ml of medium were seeded at a cell density of 5 \u00d7 10^4 cells/ml and grown to ~90 % confluence at 37 \u00b0C in a humidified atmosphere of 5 % CO2 in air. Cells were dislodged from the growth surface by banging the flask and the cells harvested by centrifugation at 450 \u00d7 g for 5 minutes. Cells were carefully resuspended in 20 ml of medium using a Pasteur pipette and then diluted 1 in 10 into fresh tissue culture flasks containing 18 ml of medium. Cells were grown for a further 24 hours to ~50\u201370 % confluence (sub-confluent) and then transfected with pIG-1-LIFR-FN3-3C-Fc as in 4.2.3.3 below.\n\n#### 4.2.3.3 Transfection of 293T Cells\n\nDNA was sterilised by precipitation in 2 volumes of ethanol followed by centrifugation at 15000 \u00d7 g. The pellet was resuspended in sterile filtered 2 M calcium chloride. As a final insurance of sterility the DNA preparation was centrifuged at 15000 \u00d7 g for 15 minutes in a sterile Eppendorf tube. Transfection was carried out in 162 cm\u00b2 tissue culture flasks (Costar) using 120 \u03bcg of DNA per flask. Typically, 0.48 ml of pIG-1-LIFR-FN3-3C-Fc DNA at a concentration of ~1.0 mg/ml was added to 1.24 ml of 2 M calcium chloride and 10.28 ml of sterile distilled water. This solution was added drop wise, over a period of several minutes, to 12 ml of 1.6 % (w/v) sodium chloride, 1.2 % (w/v) HEPES and 0.04 % (w/v) disodium hydrogen phosphate (HBS) (see Appendix III). The precipitate was inverted and 6 ml pipetted into each flask of sub-confluent cells followed by mixing. The precipitate was left on the cells for 16 hours and the monolayer washed with PBS before adding 30 ml of serum free ULTRA-CHO medium (BioWhittaker). The cells were incubated at 37 \u00b0C in a humidified atmosphere of 5 % CO2 in air for 4\u20138 days. For large-scale preparations ~9 mg of pIG-1-LIFR-FN3-3C-Fc DNA was transfected into 72 sub-confluent 162 cm\u00b2 tissue culture flasks and expression carried out for 1 week after which medium was harvested and fresh medium added for a second week of protein expression. Expression of protein was monitored by means of SDS-PAGE and western blot.\n\n### 4.2.4 Deglycosylation of LIFR\n\nThe use of the enzyme peptide-N-glycosidase F (PNGase F; Tarentino et al., 1990) was investigated as a means of deglycosylation as LIFR-FN3 had previously proved recalcitrant to other means of deglycosylation tested (see 4.4 for details). This enzyme functions by cleaving between the innermost N-acetylglucosamine and asparagine residues of high mannose, hybrid, and complex oligosaccharides of N-linked glycoproteins. Two variants of PNGase F were tested under non-denaturing conditions; non-recombinant material isolated from Flavobacterium meningosepticum (New England Biolabs) and recombinant protein expressed in E. coli (Barsomian et al., 1990; kindly provided in pMal vector for expression as a maltose binding protein (MBP) by M. A. Hall, CRC Growth Factors Group, Birmingham).\n\n#### 4.2.4.1 PNGase F Treatment\n\nSamples of LIFR-FN3 were dialysed using a microdialysis chamber (Sigma) through 12\u201314 kDa exclusion limit membrane (Gibco) against PNGase F (G7) buffer consisting of 0.05 M sodium phosphate, 10 % NP-40, pH 7.5 for 5 hours at 4 \u00b0C. Non-recombinant PNGase F (New England Biolabs) and recombinant PNGase F-MBP (M. A. Hall, CRC Growth Factors Group, Birmingham) were incubated with glycosylated LIFR-FN3-3C-Fc as detailed in Table 4-2 below. Deglycosylation was carried out in G7 buffer at room temperature or 37 \u00b0C for between 1 and 24 hours. Deglycosylation was monitored by SDS-PAGE and western blot analysis (see Chapter 3).\n\n| Component | Tube 1 | Tube 2 | Tube 3 | Tube 4 | Tube 5 | Tube 6 | Tube 7 | Tube 8 |\n|---|---|---|---|---|---|---|---|---|\n| LIFR-FN3 | \u2013 | \u2013 | 17.0 \u03bcg | 17.0 \u03bcg | 17.0 \u03bcg | 17.0 \u03bcg | 17.0 \u03bcg | 17.0 \u03bcg |\n| PNGase F | 30.6 \u03bcg (NR) | 9.4 \u03bcg (R) | \u2013 | 30.6 \u03bcg (NR) | 9.4 \u03bcg (R) | 18.8 \u03bcg (R) | 30.6 \u03bcg (R) | \u2013 |\n| x10 G7 Buffer | \u2013 | \u2013 | \u2013 | 5.0 \u03bcl | 5.0 \u03bcl | 5.0 \u03bcl | 5.0 \u03bcl | 5.0 \u03bcl |\n| Distilled Water | 41.5 \u03bcl | 41.5 \u03bcl | 40.0 \u03bcl | 26.5 \u03bcl | 26.5 \u03bcl | 17.9 \u03bcl | 7.2 \u03bcl | 35.0 \u03bcl |\n| Total Volume | 50.0 \u03bcl | 50.0 \u03bcl | 50.0 \u03bcl | 50.0 \u03bcl | 50.0 \u03bcl | 50.0 \u03bcl | 50.0 \u03bcl | 50.0 \u03bcl |\n\n*Table 4-2.* Table showing PNGase F reaction mixtures used for deglycosylation of LIFR-FN3 under non-denaturing conditions. Deglycosylation was carried out at 37 \u00b0C in G7 buffer consisting of 0.05 M sodium phosphate, 10 % NP-40, pH 7.5 for between 1 and 24 hours. Non-recombinant PNGase F (NR) was used at a concentration of 3.6 mg/ml and recombinant PNGase F (R) at a concentration of 1.1 mg/ml. LIFR-FN3 was used at a concentration of 1.7 mg/ml.\n\n#### 4.2.4.2 Production of Recombinant PNGase F", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nMedium was removed from confluent monolayers of cells by aspiration and replaced with 50 ml of serum free ULTRA-CHO medium (BioWhittaker). Flasks were incubated for up to 5 weeks at 37 \u00b0C in a humidified atmosphere of 5 % CO2 in air. Samples were taken aseptically at weekly intervals and the expression of LIFR-FN3-3C-Fc monitored by means of ELISA as detailed in 4.2.5.1 below.\n\nDifferent cell culture vessels were tested for optimal cell growth and protein expression including a small-scale stirred-tank bioreactor system (New Brunswick Scientific), roller bottles (Falcon) and large-scale tissue culture flasks (Costar). The use of sodium butyrate induced protein expression was also investigated (D'Anna et al., 1980; Gorman et al., 1983; Reeves and Cserjesi, 1979). Typically, cells were incubated as above in ULTRA-CHO medium (BioWhittaker) supplemented with 2 mM sodium butyrate (Sigma).\n\nSupernatants were harvested and phenylmethylsulphonyl fluoride (PMSF) and EDTA added to a final concentration of 1 mM. Samples were centrifuged at 2000 \u00d7 g for 30 minutes or filtered through Millex AP50 pre-filters (Millipore). Samples were kept at 4 \u00b0C for short-term storage or \u221220 \u00b0C for long-term storage. Previously frozen samples were centrifuged at 5000 \u00d7 g for 10 minutes at 4 \u00b0C. Finally, supernatants were passed through 0.45 \u03bcm membrane filters (Millipore) and the samples adjusted to pH 7.0.\n\n#### 4.2.1.5 Affinity Purification and Cleavage of Fc Fusion Proteins\n\nThe chromatography support (Bio-Rad PolyPrep) was packed with 1 ml of Protein-A sepharose Fast Flow (Pharmacia) per litre of supernatant to be processed. The column was washed with 20 column volumes of PBS and the supernatant passed through the column under gravity at 4 \u00b0C followed by 20 column volumes of PBS. Fc fusion proteins were cleaved, in situ, using 3C-His6 protease as detailed in Chapter 3.\n\n### 4.2.2 Stable Expression in Schneider's Drosophila Cell Line 2 (S2)\n\nSchneider's Drosophila cell line 2 (S2) is an epithelial cell line established from 20\u201324 hour embryos of Drosophila melanogaster (Schneider, 1972). These cells grow as loosely adherent monolayers or in suspension at temperatures between 16 \u00b0C and 28 \u00b0C, with the optimum lying between 22 \u00b0C and 24 \u00b0C. The expression plasmid pRmHa-3 (Bunch et al., 1988) is utilised containing the inducible Drosophila metallothionein promoter (see Figure 4-2). In this system the promoter is activated by elevated amounts of the heavy metal ions Cu2+, Zn2+ or Cd2+; copper chloride was used in this study as it is less toxic and does not induce a heat-shock response (Bunch et al., 1988). pRmHa-3-LIFR-FN3-3C-Fc constructs were kindly provided by K. Chobotova, CRC Growth Factor Group, Birmingham.\n\n#### 4.2.2.1 S2 Cell Culture and Protein Expression\n\nCell culture was essentially carried out as in 4.2.1.1 above, except pRmHa-3-LIFR-FN3-3C-Fc transfected S2 cells were grown in SF900 medium (Gibco) supplemented with 5 % insect tested serum (Gibco), 50 IU/ml penicillin, 50 \u03bcg/ml streptomycin and 300 \u03bcg/ml hygromycin B (Sigma). Typically, 500 ml Erlenmeyer flasks were seeded at a cell density of 1 \u00d7 10^6 cells/ml and subsequent protein expression carried out in 100 ml of growth medium containing no hygromycin B selection and serum concentrations ranging from 0\u20131 %. Flasks were incubated with gentle shaking (~100 RPM) at 27 \u00b0C for 24 hours. Protein expression was induced with copper chloride to a final concentration of 0.5 mM. Cells were counted on a daily basis and expression of protein monitored by means of SDS-PAGE and western blot.\n\n### 4.2.3 Transient Expression in 293T Cells\n\n293T cells (293/tsA1609neo; Graham et al., 1977) are a human epithelial/embryonal kidney cell line transformed with the SV40 large T antigen, so allowing high levels of transient expression from plasmids containing a polyoma origin of replication. The transient expression vector pIG-1 (Simmons, 1992) containing the hCMV enhancer and promoter was used. This vector contains both an SV40 origin and polyoma origin so allowing replication of this vector in polyoma transformed cells such as COS and 293T cells. The remaining elements of this vector allow replication and drug selection in E. coli. The supF element encodes a tRNA that suppresses amber stop codons in ampicillin and tetracycline resistance genes carried on a stable episome, termed P3, found in the E. coli strain MC101. pIG-1-LIFR-FN3-3C-Fc constructs were kindly provided by K. Chobotova, CRC Growth Factor Group, Birmingham.\n\n#### 4.2.3.1 pIG-1 DNA Production and Purification\n\nE. coli strain MC101/P3, previously transfected with pIG-1-LIFR-FN3-3C-Fc as in Chapter 2, were grown in 4 L of LB medium containing 50 \u03bcg/ml ampicillin and 12 \u03bcg/ml tetracycline for 16\u201320 hours. The cells were harvested by centrifugation at 6000 \u00d7 g for 15 minutes at 4 \u00b0C and plasmid extracted using a Maxi Prep\u2122 Kit (Qiagen). The purified pIG-1-LIFR-FN3-3C-Fc was triple digested with BamHI, BglII and HindIII as in Table 4-1 below for confirmation of identity. The reaction mixture was incubated at 37 \u00b0C for 2 hours and the digested samples were run on a 1 % agarose-TAE gel.\n\n| Component | Volume (\u03bcl) |\n|---|---|\n| pIG-1-LIFR-FN3-3C-Fc DNA | 1.00 |\n| 10x Restriction Buffer B | 1.00 |\n| BamHI (10 units/ml) | 0.50 |\n| HindIII (10 units/ml) | 0.50 |\n| BglII (40 units/ml) | 0.25 |\n| Sterile Distilled Water | 6.75 |\n| Total | 10.00 |\n\n*Table 4-1.* Table showing composition of restriction enzyme reaction mixture used to confirm identity of pIG-1-LIFR-FN3-3C-Fc.\n\n#### 4.2.3.2 293T Cell Culture", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\nExamination of the genomic context for members of the FmdE Pfam family (PF02663), such as the protein encoded by the fmdE gene from the methanogenic archaeon Methanobacterium thermoautotrophicum, indicates that 13 of them are co-transcribed with genes encoding subunits of molybdenum formylmethanofuran dehydrogenase (EC 1.2.99.5), an enzyme that is involved in microbial methane production. Here, the first crystal structures from PF02663 are described, representing two bacterial and one archaeal species: B8FYU2_DESHY from the anaerobic dehalogenating bacterium Desulfitobacterium hafniense DCB-2, Q2LQ23_SYNAS from the syntrophic bacterium Syntrophus aciditrophicus SB and Q9HJ63_THEAC from the thermoacidophilic archaeon Thermoplasma acidophilum. Two of these proteins, Q9HJ63_THEAC and Q2LQ23_SYNAS, contain two domains: an N-terminal thioredoxin-like \u03b1+\u03b2 core domain (NTD) consisting of a five-stranded, mixed \u03b2-sheet flanked by several \u03b1-helices and a C-terminal zinc-finger domain (CTD). B8FYU2_DESHY, on the other hand, is composed solely of the NTD. The CTD of Q9HJ63_THEAC and Q2LQ23_SYNAS is best characterized as a treble-clef zinc finger. Two significant structural differences between Q9HJ63_THEAC and Q2LQ23_SYNAS involve their metal binding. First, zinc is bound to the putative active site on the NTD of Q9HJ63_THEAC, but is absent from the NTD of Q2LQ23_SYNAS. Second, whereas the structure of the CTD of Q2LQ23_SYNAS shows four Cys side chains within coordination distance of the Zn atom, the structure of Q9HJ63_THEAC is atypical for a treble-cleft zinc finger in that three Cys side chains and an Asp side chain are within coordination distance of the zinc.\n\n## The crystal structure of a bacterial Sufu-like protein defines a novel group of bacterial proteins that are similar to the N-terminal domain of human Sufu (2010)\n\n**Authors:** Das D, Finn RD, Abdubek P, Astakhova T, Axelrod HL, Bakolitsa C, Cai X, Carlton D, Chen C, Chiu HJ, Chiu M, Clayton T, Deller MC, Duan L, Ellrott K, Farr CL, Feuerhelm J, Grant JC, Grzechnik A, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Krishna SS, Kumar A, Lam WW, Marciano D, Miller MD, Morse AT, Nigoghossian E, Nopakun A, Okach L, Puckett C, Reyes R, Tien HJ, Trame CB, van den Bedem H, Weekes D, Wooten T, Xu Q, Yeh A, Zhou J, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Protein science : a publication of the Protein Society (2010). DOI: 10.1002/pro.497. PMID: 20836087.\n\nSufu (Suppressor of Fused), a two-domain protein, plays a critical role in regulating Hedgehog signaling and is conserved from flies to humans. A few bacterial Sufu-like proteins have previously been identified based on sequence similarity to the N-terminal domain of eukaryotic Sufu proteins, but none have been structurally or biochemically characterized and their function in bacteria is unknown. We have determined the crystal structure of a more distantly related Sufu-like homolog, NGO1391 from Neisseria gonorrhoeae, at 1.4 \u00c5 resolution, which provides the first biophysical characterization of a bacterial Sufu-like protein. The structure revealed a striking similarity to the N-terminal domain of human Sufu (r.m.s.d. of 2.6 \u00c5 over 93% of the NGO1391 protein), despite an extremely low sequence identity of \u223c15%. Subsequent sequence analysis revealed that NGO1391 defines a new subset of smaller, Sufu-like proteins that are present in \u223c200 bacterial species and has resulted in expansion of the SUFU (PF05076) family in Pfam.\n\n## The structure of BVU2987 from Bacteroides vulgatus reveals a superfamily of bacterial periplasmic proteins with possible inhibitory function (2010)\n\n**Authors:** Das D, Finn RD, Carlton D, Miller MD, Abdubek P, Astakhova T, Axelrod HL, Bakolitsa C, Chen C, Chiu HJ, Chiu M, Clayton T, Deller MC, Duan L, Ellrott K, Ernst D, Farr CL, Feuerhelm J, Grant JC, Grzechnik A, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Krishna SS, Kumar A, Marciano D, McMullan D, Morse AT, Nigoghossian E, Nopakun A, Okach L, Puckett C, Reyes R, Rife CL, Sefcovic N, Tien HJ, Trame CB, van den Bedem H, Weekes D, Wooten T, Xu Q, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309109046788. PMID: 20944221.\n\nProteins that contain the DUF2874 domain constitute a new Pfam family PF11396. Members of this family have predominantly been identified in microbes found in the human gut and oral cavity. The crystal structure of one member of this family, BVU2987 from Bacteroides vulgatus, has been determined, revealing a \u03b2-lactamase inhibitor protein-like structure with a tandem repeat of domains. Sequence analysis and structural comparisons reveal that BVU2987 and other DUF2874 proteins are related to \u03b2-lactamase inhibitor protein, PepSY and SmpA_OmlA proteins and hence are likely to function as inhibitory proteins.\n\n## The structure of Haemophilus influenzae prephenate dehydrogenase suggests unique features of bifunctional TyrA enzymes (2010)\n\n**Authors:** Chiu HJ, Abdubek P, Astakhova T, Axelrod HL, Carlton D, Clayton T, Das D, Deller MC, Duan L, Feuerhelm J, Grant JC, Grzechnik A, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Krishna SS, Kumar A, Marciano D, McMullan D, Miller MD, Morse AT, Nigoghossian E, Okach L, Reyes R, Tien HJ, Trame CB, van den Bedem H, Weekes D, Xu Q, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309110021688. PMID: 20944228.", "# The $40 Billion Revolution: How Bispecific Antibodies Are Transforming Cancer Care\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe pipeline expansion suggests even greater potential ahead. Over 250 bispecific antibodies currently undergo clinical development across more than 180 companies. Trispecific antibodies, targeting three simultaneous antigens, show early promise with up to 260-fold enhanced tumor-killing capacity compared to traditional treatments. Johnson & Johnson\u2019s trispecific candidate JNJ-79635322, targeting CD3, BCMA, and GPRC5D simultaneously, achieved 100% response rates in early trials.\n\nManufacturing complexity presents ongoing challenges, particularly ensuring correct antibody assembly and preventing aggregation. Advanced purification technologies using mixed-mode chromatography and scalable platforms are addressing these hurdles, though production costs remain higher than traditional monoclonal antibodies.\n\nThe next evolution involves combination strategies and biomarker-driven patient selection. Researchers explore pairing bispecific antibodies with checkpoint inhibitors, targeted therapies, and even other bispecifics to overcome resistance mechanisms. Johnson & Johnson\u2019s MonumenTAL-6 trial, combining two bispecific antibodies for multiple myeloma, represents this frontier approach.\n\nMarket dynamics reflect this therapeutic revolution. From $5.5 billion in 2024, the bispecific antibody market projects explosive growth to $40-484 billion by 2034, depending on analyst estimates. This growth encompasses not only oncology applications but expanding indications in autoimmune diseases, ophthalmology, and rare disorders.\n\nBeyond cancer, bispecific antibodies demonstrate versatility across medical disciplines. Emicizumab transformed hemophilia A treatment by mimicking blood clotting factors, while faricimab revolutionized retinal disease management by simultaneously targeting multiple pathways underlying vision loss. These successes preview broader applications as the platform matures.\n\nThe convergence of technological advancement, clinical validation, and manufacturing scale positions bispecific antibodies as a defining therapeutic class for the next decade. For patients facing previously untreatable cancers, these dual-targeting molecules offer unprecedented hope, transforming terminal diagnoses into manageable chronic conditions or even potential cures.\n\n| Key Concept | Description | Key References |\n| --- | --- | --- |\n| **Dual-Targeting Mechanism** | Bispecific antibodies simultaneously bind tumor antigens and immune effector cells, creating direct immune synapse for enhanced tumor killing | [Shi L, et al., Cell Death & Disease](https://pmc.ncbi.nlm.nih.gov/articles/PMC12320366/) |\n| **Market Growth** | Projected market expansion from $12-17 billion (2024) to $40-484 billion by 2034, representing 44% annual growth rate | [Straits Research Market Report](https://straitsresearch.com/report/bispecific-antibodies-market) |\n| **Clinical Efficacy** | Overall response rates exceeding 70% in heavily pretreated multiple myeloma patients, with 50% achieving complete remission | [HealthTree Multiple Myeloma Guide](https://healthtree.org/myeloma/guides/bispecific-antibodies/bispecific-antibodies-myeloma-guide-fda-approved-bsab) |\n| **Treatment Accessibility** | \u201cOff-the-shelf\u201d availability enables immediate treatment initiation versus weeks-long CAR-T preparation, democratizing advanced immunotherapy access | [AJMC Community Treatment Access](https://www.ajmc.com/view/banner-year-for-bispecifics-as-therapies-advance-into-the-community-setting) |\n| **Safety Profile** | Cytokine release syndrome as primary toxicity, manageable through step-up dosing and prophylactic interventions | [PMC Cytokine Release Mechanisms](https://pmc.ncbi.nlm.nih.gov/articles/PMC8855852/) |\n| **Resistance Mechanisms** | Multifactorial resistance including antigen loss, T-cell exhaustion, and immunosuppressive microenvironment, addressable through combination strategies | [van de Donk NW, et al., Lancet Haematology](https://pubmed.ncbi.nlm.nih.gov/39033769/) |\n| **Pipeline Development** | 250+ bispecific candidates across 180+ companies, with trispecific antibodies showing 260-fold enhanced potency in preclinical studies | DelveInsight Pipeline Analysis |\n| **Manufacturing Challenges** | Complex antibody assembly requiring specialized purification technologies and quality control measures, with higher production costs than traditional antibodies | [IPT Bispecific Production Challenges](https://www.iptonline.com/articles/overcoming-challenges-in-bispecific-antibody-production-with-mixed-mode-chromatography-and-scalable-purification-solutions-spring-2025) |", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\nIn some embodiments, the VH CDR1 comprises the amino acid sequence of any one of SEQ ID NOs:1-6; the VH CDR2 comprises the amino acid sequence of any one of SEQ ID NOs:7-17 and 92-95; the VH CDR3 comprises the amino acid sequence of any one of SEQ ID NOs:18-25; the VL CDR1 comprises the amino acid sequence of any one of SEQ ID NOs:26-52 or 118; the VL CDR2 comprises the amino acid sequence of any one of SEQ ID NOs:53-68; and the VL CDR3 comprises the amino acid sequence of any one of SEQ ID NOs:69-91.\n\nIn some embodiments, the VH CDR1, the VH CDR2, and the VH CDR3 each correspond to the VH CDRs set forth in Tables 1-2 for a single VH clone, and wherein the VL CDR1, the VL CDR2, and the VL CDR3 each correspond to the VL CDRs set forth in Tables 1-2 for a single VL clone.\n\n*(The Summary then enumerates, over approximately thirty pages, several hundred individual clone-by-clone permutations of the six CDRs by SEQ ID NO \u2014 for example \"the VH CDR1 comprises the amino acid sequence of SEQ ID NO:1; the VH CDR2 comprises the amino acid sequence of SEQ ID NO:7; the VH CDR3 comprises the amino acid sequence of SEQ ID NO:18; the VL CDR1 comprises the amino acid sequence of SEQ ID NO:26; the VL CDR2 comprises the amino acid sequence of SEQ ID NO:53; and the VL CDR3 comprises the amino acid sequence of SEQ ID NO:69\" \u2014 followed by a comparable enumeration of paired VH/VL sequences (SEQ ID NOs:165-208 paired with SEQ ID NOs:264-318) and of paired heavy-chain/light-chain sequences. These recitations are combinatorial listings of the same six CDR positions across the 54 parental clones and 161 affinity-matured variants, and are not reproduced individually here.)*\n\nIn some embodiments, the VH is at least 80% identical to the amino acid sequence of any one of SEQ ID NOs:165-208; and the VL is at least 80% identical to the amino acid sequence of any one of SEQ ID NOs:264-318. In some embodiments, the VH comprises the amino acid sequence of any one of SEQ ID NOs:165-208; and the VL comprises the amino acid sequence of any one of SEQ ID NOs:264-318.\n\nThe Summary further provides group-specific consensus definitions of the antibody. In some embodiments the antibody comprises a VH CDR1 comprising ELSMQ (SEQ ID NO:1); a VH CDR2 comprising GFDPDDX101ETMYAEX102X103QG (SEQ ID NO:102; Group 1 clones), wherein X101 is D or G, X102 is K or R, and X103 is F or L; a VH CDR3 SPGYDFFDY (SEQ ID NO:18); a VL CDR1 comprising GGX104X105X106GX107X108X109VX110 (SEQ ID NO:103; Group 1 clones), wherein X104 is N, D, or S; X105 is Y, N, or D; X106 is I or T; X107 is S, D, I, R, or T; X108 is K, E, or I; X109 is S, I, R, G, N, or A; and X110 is H, F, or N; a VL CDR2 comprising DDX111DRPX112 (SEQ ID NO:104; Group 1 clones), wherein X111 is G, S, or R and X112 is S or L; and a VL CDR3 comprising QVWDX113X114X115DX116X117X118 (SEQ ID NO:105; Group 1 clones), wherein X113 is S or A; X114 is I or S; X115 is S, I, or N; X116 is H, L, or Q; X117 is V or L; and X118 is V or I.\n\nIn a related Group 1 embodiment the VL CDR1 is TGTSSDVGGYNYVS (SEQ ID NO:30); the VL CDR2 is X119VSX120RPS (SEQ ID NO:106; Group 1 clones), wherein X119 is E or D and X120 is N or K; and the VL CDR3 is QVWDSSNDLLI (SEQ ID NO:71). In a further embodiment the VH CDR1 is the longer GYTLTELSMQ (SEQ ID NO:329).\n\nGroup 2 clones: VH CDR1 DYFIH (SEQ ID NO:2); VH CDR2 LVDPEDGETIYAEX121FQG (SEQ ID NO:107), wherein X121 is K or R; VH CDR3 PGGILTDPDAFDI (SEQ ID NO:19); VL CDR1 X122GTX123SDVGGYNX124VS (SEQ ID NO:108), wherein X122 is T or A, X123 is S or G, and X124 is Y or H; VL CDR2 X125VX126X127RPS (SEQ ID NO:109), wherein X125 is D or E, X126 is N or S, and X127 is K or N; and VL CDR3 SSYX128X129SSTX130X131V (SEQ ID NO:110), wherein X128 is I or T; X129 is P or S; X130 is R, P, F, or absent; and X131 is W or Y.\n\nGroup 3 clones: VH CDR1 SYAIS (SEQ ID NO:3); VH CDR2 LVDPEDGETIYAEKFX132G (SEQ ID NO:111), wherein X132 is R or Q; VH CDR3 EESYGP (SEQ ID NO:20); VL CDR1 QASQDISNYLX133 (SEQ ID NO:112), wherein X133 is N or D; VL CDR2 DASNLET (SEQ ID NO:61); and VL CDR3 QQLNSYPLT (SEQ ID NO:80).\n\nGroup 4 clones: VH CDR1 EX134SMH (SEQ ID NO:113), wherein X134 is S or L; VH CDR2 LVDPEDGETIYAQKFQG (SEQ ID NO:14); VH CDR3 EEWSGDGDDAFDI (SEQ ID NO:21); VL CDR1 SGSSSNIGSYSVS (SEQ ID NO:46); VL CDR2 DX135NKRPS (SEQ ID NO:114), wherein X135 is N or D; and VL CDR3 GTWDSSLSAWV (SEQ ID NO:81).\n\nGroup 6 clones: VH CDR1 SYAIS (SEQ ID NO:3); VH CDR2 GIIPIFGTANYAQKFQG (SEQ ID NO:15); VH CDR3 SPLRGSGWYWHYYYGMDV (SEQ ID NO:22); VL CDR1 GGNX136IX137X138KX139VH (SEQ ID NO:115), wherein X136 is N or K; X137 is R or G; X138 is A, S, R, or T; and X139 is H or S; VL CDR2 X140DX141X142RPS (SEQ ID NO:116), wherein X140 is Q or R, X141 is S or R, and X142 is N or K; and VL CDR3 QX143WX144SX145TX146V (SEQ ID NO:117), wherein X143 is A or V; X144 is D or G; X145 is S or N; and X146 is V, A, or E.\n\nIn some embodiments, the human mutant CALR is human Type 1 mutant CALR comprising the amino acid sequence of SEQ ID NO:320. In some embodiments, the human mutant CALR is human Type 2 mutant CALR comprising the amino acid sequence of SEQ ID NO:321.", "# AlphaFraud \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nThe completed full-archive backfill (finished 2026-07-18) processed **96,827** post-cutoff human entities: **82,796 analysed** (85.5%; 67,314 fast-screened plus 15,482 fully compared), **1,941 confidently wrong** (2.3% of analysed, 431 of them sequence-novel), **13,908 skipped** (14.4%), and **123 unrecoverable errors** (0.1%, broken coordinate-less files). The skips break down as:\n\n| Count | Share | Reason | Nature |\n| --- | --- | --- | --- |\n| **11,875** | 85% | **No single UniProt mapping** \u2014 antibodies, chimeras, engineered constructs, fusion proteins, synthetics | Structural fact: no single AlphaFold model exists to compare against |\n| **1,186** | 9% | **No AlphaFold model covering the residue range** \u2014 mostly giant multi-domain proteins (titin-scale) whose resolved fragment falls outside AlphaFold DB\u2019s fragmented coverage, plus isoforms it lacks | Coverage gap in AlphaFold DB |\n| **464** | 3% | **Too few residues aligned (<10)** \u2014 tiny peptides or mostly-unresolved chains | Nothing meaningful to compare |\n| **381** | 3% | **Structure too large (>40 MB)** \u2014 the out-of-memory guard for the 3.8 GB droplet | Resource limit (our cap) |\n| **2** | \u2014 | Sequence too short (<3 residues) | Degenerate |\n\nThe 85% is exactly what the design intends: because AlphaFold DB is monomer-and-single-sequence only, anything that does not map to one human UniProt accession (an antibody Fab, a designed fusion, a chimera) has no counterpart model, so it is flagged and skipped rather than mis-compared.\n\nSkipped entities are terminal by design: `retry-errors` only re-runs `status='error'`, and the resumable weekly and backfill runs skip anything already recorded. Four of the five buckets are permanent structural properties, so re-running would simply re-skip them. The two worth revisiting are the **381 oversized structures** (a memory limit of the small droplet, not a true incomparability, recoverable by raising `MAX_STRUCT_BYTES` or upsizing the box) and the multi-domain subset of the **1,186 missing-model** bucket (recoverable with better fragment selection for AlphaFold models above 2,700 residues).\n\n## \ud83d\udee0\ufe0f Web deployment\n\n**Live at [alphafraud.mdeller.com](https://alphafraud.mdeller.com).** The report is served by a Flask app (gunicorn behind nginx with a Let\u2019s Encrypt certificate) and the weekly pipeline runs from a systemd timer on the same droplet.\n\n```\n# 1. From your Mac: copy .env.example to .env and fill in DROPLET_SSH / DROPLET_PATH / SERVER_NAME\ncp .env.example .env\n\n# 2. Push the code to the droplet\nbash deploy/deploy.sh\n\n# 3. On the droplet as root: one-time provisioning (packages, user, venv, services, nginx, TLS)\nsudo SERVER_NAME=alphafraud.mdeller.com bash /opt/alphafraud/deploy/provision.sh\n\n# 4. (optional) trigger the first pipeline run immediately\nsudo systemctl start alphafraud-run.service\n```\n\nAfter that, `deploy/deploy.sh` pushes code updates and restarts the web service, and `deploy/release.sh \"message\"` does the whole loop in one step (commit, push to GitHub, then deploy). Deploys are safe to run while a backfill is in progress: the web app is a read-only reader, WAL is set once, and DB writes are retry-wrapped, so restarting the web service never disturbs the running pipeline. The weekly timer fires every Wednesday 00:30 UTC.\n\n| File | Role |\n| --- | --- |\n| `deploy/provision.sh` | One-time root setup: system packages, service user, venv, systemd units, nginx site, certbot. Also patches HTTP/2 onto certbot\u2019s TLS listener afterward (nginx doesn\u2019t enable it by default) \u2014 idempotent, so it self-heals on a re-run |\n| `deploy/deploy.sh` | Push code from your Mac (rsync over SSH) and restart the web service |\n| `deploy/release.sh` | One-shot: commit, push to GitHub, then deploy to the droplet |\n| `deploy/alphafraud-web.service` | gunicorn web app (always on) |\n| `deploy/alphafraud-run.service` + `.timer` | Weekly pipeline and its Wednesday 00:30 UTC schedule |\n| `deploy/alphafraud-backfill.service` | Supervised one-time full-archive backfill; auto-restarts on crash, survives reboot (`systemctl enable --now alphafraud-backfill`) |\n| `deploy/nginx-alphafraud.conf` | nginx reverse-proxy site (templated with the server name) |\n\n## \ud83e\uddf1 Stack\n\nPython, biotite and tmtools for structure handling and superposition, numpy and scipy for the metrics, Flask and gunicorn for serving, Plotly for the figures, and SQLite for state. No compiled toolchain, no external database.\n\n## \ud83e\udded Caveats\n\n* \u201cDeposited after the cutoff\u201d plus a novelty score is a strong proxy, not proof AlphaFold never saw a homolog: the novelty column makes this explicit so you can filter to the defensible subset.\n* AlphaFold DB models are monomers, so complex and interface accuracy are not tested in this version.\n* Engineered constructs, antibodies and fusion proteins that do not map to a single UniProt accession are flagged and skipped, not compared.\n\n## \u2705 To Do\n\nRoadmap for AlphaFraud, newest ideas at the top. Suggestions welcome.", "# Approaches to automated protein crystal harvesting\nAuthors: Deller, M. C.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 70, 133\u2013155 (IYCr crystallization series), 2014\n\nHarvesting in its broadest sense is any process that relocates a crystal from its mother liquor to an environment where diffraction data can be collected. The authors decompose it into looping, cryoprotection, flash-cooling and mounting onto an X-ray source. Looping places the crystal into a microloop on a standardized mounting pin, and the technique has changed little in some 20 years since its inception, although microloops and micromounts are now often made from polyimide materials such as Kapton. Cryoprotection typically supplements the aqueous mother liquor with a higher percentage of an organic solution such as MPD, ethanol, glycerol, ethylene glycol, polyethylene glycol or 2-propanol, protecting against the radiation damage inflicted by ultrabrilliant synchrotron beams. Flash-cooling is usually achieved with liquid nitrogen or a stream of dry gaseous nitrogen, though liquid propane or helium can be used, and the cooled crystal is then stored in liquid nitrogen until it is mounted on a goniostat under a cryogenic stream. The review is careful to distinguish the harvesting stage \u2014 looping through flash-cooling \u2014 from the separate crystal-mounting and screening stage, and focuses on the former.\n\n### 1.2.2. Current levels of automation\n\nReliable automation already exists for most of the gene-to-structure pipeline. For crystallization, fully integrated platforms offer screen design, plate setup and crystal imaging, including the CrystalMation system used at the JCSG and comparable systems at The Photon Factory, EMBL Hamburg and the Oxford Protein Production Facility. Diffraction screening and data collection are likewise highly automated with integrated robotic sample loading and beamline automation, and automated sample-loading systems are now commonplace worldwide; the Stanford Auto Mounter (SAM) used by the JCSG at SSRL performs fully automated sample loading, retrieval and data collection. Standardization initiatives such as the Universal Puck Project further simplify their use by harmonizing sample and shipping containers. What remains missing is precisely the segment between automated crystal production and automated data collection \u2014 automated looping, cryoprotection and flash-cooling. Automated harvesting technologies are described as still in their infancy though rapidly progressing, with only a handful of mostly semi-automated systems in existence.\n\n### 1.2.3. The protein crystal-harvesting bottleneck and the benefits of automation\n\nHarvesting is described as one of the exclusively manual steps in the traditional structure-determination pipeline, performed by skilled operatives at a stereo microscope, and therefore time-consuming, low-throughput and poorly reproducible. The authors enumerate eight benefits of automating it. High throughput: JCSG procedures had by January 2014 harvested and screened just over 165,000 crystals to yield 1,470 structures, roughly 100 manually harvested crystals per deposited structure, and manual harvesting of 100 crystals takes a proficient expert of the order of four hours. Fully automated pipeline: manual harvesting is the last bastion of full human intervention and closing it would give true gene-to-structure automation. Smaller crystals: automation enables routine harvesting of microcrystals below 10 \u00b5m, which are extremely hard to handle by hand because of involuntary hand tremor and drop dehydration, and which are needed to exploit microfocus beamlines. Adaptability: robotic motions and tools can be optimized for microfluidic chips, plugs and lipidic cubic phase. Precision liquid handling: automated addition of cryoprotectants, heavy-atom soaks and small molecules for co-crystallization, opening possibilities for fragment-based screening. High-speed harvesting: limiting dehydration and time spent in cryoprotectant. Reproducibility: constant robot speeds and controlled motion paths eliminate shake, jitter and inter-operator variability. Improved flash-cooling: high-speed, precise robotics permit hyperquenching and novel cryoprotection methods such as the UMR cryo-drip.\n\n### 1.3.1. New detector instrumentation\n\nPixel-array detectors (PADs) are described as effecting a paradigm shift in how diffraction data are collected, with knock-on consequences for harvesting. Their millisecond readout times allow shutterless data collection with continuous crystal rotation, avoiding acceleration and backlash artefacts of stop\u2013start rotation methods and eliminating inertia problems that matter for large objects such as entire crystallization plates on a robot actuator. Continuous collection makes it feasible to use robot arms as goniostats, removing separate harvesting and pin-mounting steps, and allows data collection from streams or slurries of microcrystals, which removes the need for microloop-based harvesting altogether \u2014 a mode enabled in practice by acoustic droplet ejection. The increased sensitivity and reduced background of PADs also permit less intense exposures and hence lower radiation dose, and in combination with helical scans (continuous exposure of progressive areas of the crystal during rotation) give better cryotemperature data. Most significantly, PADs make it nearly possible to outrun radiation damage during room-temperature data collection, which has enabled in situ data-collection methods on several beamlines.\n\n### 1.3.2. MEMS-based technologies", "# Hitting the Delete Button on Disease: The Rise of Targeted Protein Degraders\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* Targeted Protein Degraders (TPDs) offer a revolutionary way to eliminate disease-causing proteins rather than just inhibiting them.\n* This technology works by hijacking the cell\u2019s natural protein disposal machinery, making it possible to target proteins previously considered \u201cundruggable\u201d.\n* Structure-based computational methods are rapidly expanding the number of known targets for TPDs, opening up new avenues for treating cancer and other diseases.\n\n1. **Recent Trends in Structural Biology:** Insights from the PSDI Conference:\n   <https://3decision.discngine.com/blog/2025/1/10/recent-trends-in-structural-biology-insights-from-the-psdi-conference>\n2. **Discovery on Target Conference:** The Industry\u2019s Preeminent Event on Novel Drug Targets:\n   [https://www.discoveryontarget.com](https://www.discoveryontarget.com/)\n3. **Drug Discovery Blogs and Influencers:** Sparking Research Ideas:\n   <https://drughunter.com/resource/drug-discovery-blogs>\n\nFor decades, the dominant strategy in drug discovery has been inhibition. Scientists designed small molecules to fit into the active site of a disease-causing protein, like a key in a lock, to block its function. While this approach has produced many life-saving medicines, it has a fundamental limitation: it only works on proteins with well-defined binding pockets. A vast landscape of proteins, estimated to be over 85% of the proteome, lacks these \u201cdruggable\u201d sites, leaving them beyond the reach of conventional therapies. These \u201cundruggable\u201d targets have long represented a frustrating frontier in medicine, implicated in cancers and other serious diseases with no viable treatment options.\n\nA new therapeutic modality, Targeted Protein Degradation (TPD), is breaking down this barrier. Instead of merely inhibiting a target, TPDs hijack the cell\u2019s own natural disposal system\u2014the ubiquitin-proteasome system\u2014to completely eliminate the disease-causing protein. These bifunctional molecules act like a microscopic matchmaker, with one end binding to the target protein and the other end recruiting a ubiquitin ligase. This forced introduction flags the target for destruction, effectively hitting the delete button on the protein. This approach is revolutionary because it doesn\u2019t require a functional active site; it only needs a small nook or cranny on the target protein to bind to.\n\nThis technology is unlocking the \u201cundruggable\u201d proteome at an incredible pace. Researchers are now using structure-based computational approaches to mine the human proteome for new targets. One such strategy, starting from the structure of the Cereblon (CRBN) ubiquitin ligase, predicted over 1,400 new proteins that could be targeted for degradation, expanding the potential of this modality exponentially. By turning the entire protein into a potential target, TPDs are creating therapeutic possibilities for cancers and other diseases that were once considered intractable, heralding a new era of drug discovery where no protein is truly \u201cundruggable.\u201d", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nThe therapeutic dosage of a compound of the present disclosure can vary according to, e.g., the particular use for which the treatment is made, the manner of administration of the compound, the health and condition of the patient, and the judgment of the prescribing physician. The proportion or concentration of a compound of the present disclosure in a pharmaceutical composition can vary depending upon a number of factors including dosage, chemical characteristics (e.g., hydrophobicity), and the route of administration. For example, the compounds of the present disclosure can be provided in an aqueous physiological buffer solution containing about 0.1 to about 10% w/v of the compound for parenteral administration. Some typical dose ranges are from about 1 \u03bcg/kg to about 1 g/kg of body weight per day. In some embodiments, the dose range is from about 0.01 mg/kg to about 100 mg/kg of body weight per day. The dosage is likely to depend on such variables as the type and extent of progression of the disease or disorder, the overall health status of the particular patient, the relative biological efficacy of the compound selected, formulation of the excipient, and its route of administration. Effective doses can be extrapolated from dose-response curves derived from in vitro or animal model test systems.\n\nFor preparing solid compositions such as tablets, the principal active ingredient is mixed with a pharmaceutical excipient to form a solid preformulation composition containing a homogeneous mixture of a compound of the present disclosure. When referring to these preformulation compositions as homogeneous, the active ingredient is typically dispersed evenly throughout the composition so that the composition can be readily subdivided into equally effective unit dosage forms such as tablets, pills and capsules. This solid preformulation is then subdivided into unit dosage forms of the type described above containing from, e.g., about 0.1 to about 1000 mg of the active ingredient of the present disclosure.\n\nThe tablets or pills of the present disclosure can be coated or otherwise compounded to provide a dosage form affording the advantage of prolonged action. For example, the tablet or pill can comprise an inner dosage and an outer dosage component, the latter being in the form of an envelope over the former. The two components can be separated by an enteric layer which serves to resist disintegration in the stomach and permit the inner component to pass intact into the duodenum or to be delayed in release. A variety of materials can be used for such enteric layers or coatings, such materials including a number of polymeric acids and mixtures of polymeric acids with such materials as shellac, cetyl alcohol and cellulose acetate.\n\nThe liquid forms in which the compounds and compositions of the present disclosure can be incorporated for administration orally or by injection include aqueous solutions, suitably flavored syrups, aqueous or oil suspensions, and flavored emulsions with edible oils such as cottonseed oil, sesame oil, coconut oil, or peanut oil, as well as elixirs and similar pharmaceutical vehicles.\n\nCompositions for inhalation or insufflation include solutions and suspensions in pharmaceutically acceptable, aqueous, or organic solvents, or mixtures thereof, and powders. The liquid or solid compositions may contain suitable pharmaceutically acceptable excipients as described supra. In some embodiments, the compositions are administered by the oral or nasal respiratory route for local or systemic effect. Compositions can be nebulized by use of inert gases. Nebulized solutions may be breathed directly from the nebulizing device or the nebulizing device can be attached to a face mask, tent, or intermittent positive pressure breathing machine. Solution, suspension, or powder compositions can be administered orally or nasally from devices which deliver the formulation in an appropriate manner.\n\nTopical formulations can contain one or more conventional carriers. In some embodiments, ointments can contain water and one or more hydrophobic carriers selected from, e.g., liquid paraffin, polyoxyethylene alkyl ether, propylene glycol, white Vaseline, and the like. Carrier compositions of creams can be based on water in combination with glycerol and one or more other components, e.g., glycerinemonostearate, PEG-glycerinemonostearate and cetylstearyl alcohol. Gels can be formulated using isopropyl alcohol and water, suitably in combination with other components such as, e.g., glycerol, hydroxyethyl cellulose, and the like. In some embodiments, topical formulations contain at least about 0.1, at least about 0.25, at least about 0.5, at least about 1, at least about 2 or at least about 5 wt % of the compound of the present disclosure. The topical formulations can be suitably packaged in tubes of, e.g., 100 g which are optionally associated with instructions for the treatment of the select indication, e.g., psoriasis or other skin condition.\n\nThe amount of compound or composition administered to a patient will vary depending upon what is being administered, the purpose of the administration, such as prophylaxis or therapy, the state of the patient, the manner of administration and the like. In therapeutic applications, compositions can be administered to a patient already suffering from a disease in an amount sufficient to cure or at least partially arrest the symptoms of the disease and its complications. Effective doses will depend on the disease condition being treated as well as by the judgment of the attending clinician depending upon factors such as the severity of the disease, the age, weight and general condition of the patient and the like.", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nThe present disclosure further provides methods of inhibiting WRN, comprising administering to a patient a compound disclosed herein, or a pharmaceutically acceptable salt thereof. In some embodiments, the inhibition of WRN is inhibition of WRN helicase activity. In some embodiments, the inhibition of WRN is inhibition of WRN ATPase activity. In some embodiments, the inhibition of WRN is inhibition of WRN helicase activity and WRN ATPase activity.\n\nThe present disclosure further provides methods of treating a disease or disorder associated with WRN activity, comprising administering to a patient in need thereof a therapeutically effective amount of a compound of disclosed herein, or a pharmaceutically acceptable salt thereof. In some embodiments, the WRN activity is WRN helicase activity. In some embodiments, the WRN activity is WRN ATPase activity. In some embodiments, the WRN activity is WRN helicase activity and WRN ATPase activity.\n\nThe present disclosure also provides use of a compound as disclosed herein, or a pharmaceutically acceptable salt thereof, for preparation of a medicament for use in any of the methods described herein.\n\nThe present disclosure also provides a compound as disclosed herein, or a pharmaceutically acceptable salt thereof, for use in any of the methods described herein.\n\n## Detailed Description\n\n### Compounds\n\nThe present disclosure provides inter alia compounds of Formula (I), or a pharmaceutically acceptable salt thereof, wherein:\n\n- m is an integer selected from 0, 1, 2, 3, 4, 5, 6, 7, or 8;\n- n is an integer selected from 0, 1, 2, 3, 4, 5, 6, 7, or 8;\n- the dashed bond is a single or double bond, provided that proper valency is maintained;\n- Z is N or CR^7;\n- X is C, CH or N; and Y is C or CH; or X is C or CH; and Y is C, CH, or N;\n- Ring moiety A is selected from C3\u201314 cycloalkyl, 6\u201310 membered aryl, 4\u201314 membered heterocycloalkyl, and 5\u201310 membered heteroaryl;\n- Ring moiety B is selected from C3\u20137 cycloalkyl, phenyl, 4\u20137 membered heterocycloalkyl, and 5\u20136 membered heteroaryl;\n- R^1 is independently selected from H, D, halo, CN, NO2, C1\u20136 alkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, C1\u20136 haloalkyl, C3\u201310 cycloalkyl, 6\u201310 membered aryl, 4\u201310 membered heterocycloalkyl, 5\u201310 membered heteroaryl, C3\u201310 cycloalkyl-C1\u20134 alkyl, 6\u201310 membered aryl-C1\u20134 alkyl, 4\u201310 membered heterocycloalkyl-C1\u20134 alkyl, 5\u201310 membered heteroaryl-C1\u20134 alkyl, OR^a1, SR^a1, NHOR^a1, C(O)R^b1, C(O)NR^c1R^d1, C(O)OR^a1, OC(O)R^b1, OC(O)NR^c1R^d1, NR^c1R^d1, NR^c1NR^c1R^d1, NR^c1C(O)R^b1, NR^c1C(O)OR^a1, NR^c1C(O)NR^c1R^d1, C(=NR^e1)R^b1, C(=NR^e1)NR^c1R^d1, NR^c1C(=NR^e1)NR^c1R^d1, NR^c1C(=NR^e1)R^b1, NR^c1S(O)NR^c1R^d1, NR^c1S(O)R^b1, NR^c1S(O)2R^b1, NR^c1S(O)(=NR^e1)R^b1, NR^c1S(O)2NR^c1R^d1, S(O)R^b1, S(O)NR^c1R^d1, S(O)2R^b1, S(O)2NR^c1R^d1, OS(O)(=NR^e1)R^b1, OS(O)2R^b1, S(O)(=NR^e1)R^b1, SF5, P(O)R^a1R^b1, OP(O)(OR^a1)(OR^a1), P(O)(OR^a1)(OR^a1), and BR^a1R^a1, wherein said alkyl, alkenyl, alkynyl, haloalkyl, cycloalkyl, aryl, heterocycloalkyl, heteroaryl and their alkyl-linked variants are each optionally substituted with 1, 2, 3, or 4 independently selected R^1A substituents;\n- R^2 is independently selected from the same broad list of groups as R^1, optionally substituted with 1, 2, 3, or 4 independently selected R^2A substituents;\n- R^3 are each independently selected from H, D, halo, CN, NO2, C1\u20136 alkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, C1\u20136 haloalkyl, C3\u20137 cycloalkyl, phenyl, 4\u20137 membered heterocycloalkyl, 5\u20136 membered heteroaryl and their C1\u20134 alkyl-linked variants, OR^a3, SR^a3, NHOR^a3, C(O)R^b3, C(O)NR^c3R^d3 and related acyl, amino, sulfonyl and phosphorus-containing groups, each optionally substituted with 1, 2, 3, or 4 independently selected R^G substituents;\n- R^4 and R^5 are each independently selected from H, C1\u20136 alkyl, C1\u20136 haloalkyl, cyano-C1\u20133 alkyl, HO\u2014C1\u20133 alkyl, and C1\u20133 alkoxy-C1\u20133 alkyl; or any R^4 and R^5 are, together with the carbon atom to which they are attached, form a 3\u20137-membered cycloalkyl ring;\n- R^6 is independently selected from C1\u20136 alkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, C1\u20136 haloalkyl, C3\u201310 cycloalkyl, 6\u201310 membered aryl, 4\u201310 membered heterocycloalkyl, 5\u201310 membered heteroaryl and their C1\u20134 alkyl-linked variants, C(O)R^b6, C(O)NR^c6R^d6, C(O)OR^a6, C(=NR^e6)R^b6, C(=NR^e6)NR^c6R^d6, S(O)R^b6, S(O)NR^c6R^d6, S(O)2R^b6, and S(O)2NR^c6R^d6, each optionally substituted with 1, 2, 3, or 4 independently selected R^6A substituents;\n- R^7 is selected from H, D, OH, NO2, CN, halo, C1\u20136 alkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, C1\u20136 haloalkyl, cyano-C1\u20136 alkyl, HO\u2014C1\u20136 alkyl, C1\u20136 alkoxy-C1\u20136 alkyl, C3\u20137 cycloalkyl, C1\u20136 alkoxy, C1\u20133 haloalkoxy, amino, C1\u20136 alkylamino, thio, C1\u20136 alkylthio, C1\u20136 alkylsulfinyl, C1\u20136 alkylsulfonyl, carbamyl, C1\u20136 alkylcarbamyl, carboxy, C1\u20136 alkylcarbonyl, C1\u20136 alkoxycarbonyl, C1\u20136 alkylcarbonyloxy, C1\u20136 alkylcarbonylamino, C1\u20136 alkoxycarbonylamino, C1\u20136 alkylaminocarbonylamino, aminosulfonyl, C1\u20136 alkylaminosulfonyl, aminosulfonylamino, aminocarbonylamino, and the corresponding di(C1\u20136 alkyl) variants, and sulfoximine/sulfonimidamide groups of the drawn formulae in which R^7a is selected from H and C1\u20136 alkyl, each R^7b is independently selected from H and C1\u20136 alkyl, and each R^7c is independently selected from H and C1\u20136 alkyl.\n\n*(The published specification defines each of the R-group and sub-R-group variables \u2014 R^a1/R^c1/R^d1, R^b1, R^e1, R^1A, R^1B, R^a2/R^c2/R^d2, R^b2, R^2A, R^2B, R^a3/R^b3/R^c3/R^d3, R^6A, R^6B, R^a61/R^b61/R^c61/R^d61, R^81\u2013R^85, R^9, R^G and others \u2014 in exhaustive, largely parallel Markush lists spanning paragraphs [0026]\u2013[0231]. Those parallel definition lists are not reproduced verbatim here; the substantive scientific and chemical content of the disclosure follows.)*\n\n### The R^W warhead", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\n**Scheme 5.** Intermediates suitable for the preparation of compounds of Formula (I) can be prepared as shown in Scheme 5. An appropriately substituted starting material 5-1, where B is a cyclic structure (e.g. cycloalkyl, phenyl, heterocycloalkyl or heteroaryl), can be reacted with a cyanoacetate 5-2 at elevated temperature (e.g., 160 \u00b0C) to afford intermediate 5-3. Intermediate 5-3 can be treated with a ketoester 5-4 in the presence of ammonium acetate and at elevated temperature (e.g., 140 \u00b0C) to provide 5-5, which is an intermediate suitable for the preparation of compounds of Formula (I). Where desired, by treating with acid (e.g., H2SO4) at elevated temperature (e.g., 120 \u00b0C), intermediate 5-5 can be converted to intermediate 5-6, which is also an intermediate suitable for the preparation of compounds of Formula (I).\n\n**Scheme 6.** In the course of preparation of compounds of Formula (I), intermediates such as 6-1 (of which 5-5 and 5-6 in Scheme 5 are examples) can be further elaborated according to multiple pathways, as illustrated in Scheme 6. Halogenation of 6-1, where B is a cyclic structure (e.g. cycloalkyl, phenyl, heterocycloalkyl or heteroaryl), to provide 6-2, where X^1 is halogen under standard halogenation conditions (e.g., in the presence of NBS or NIS or NCS), can be followed by alkylation of 6-2 under standard SN2 conditions (e.g., in the presence of K2CO3, tetrabutylammonium bromide and 6-3, where X^2 is a halogen (e.g., Cl, Br, or I) or pseudohalogen (e.g., OTs or OMs)) to give compound 6-4. Alternatively, the steps can be performed in the reverse order (e.g., alkylation of 6-1 to afford 6-5, followed by halogenation to provide 6-4). Halide 6-4 can be coupled with compound 6-6 under standard Suzuki, Stille, Negishi or Buchward-Hartwig conditions to give compound 6-7. Amine 6-7 can be converted to compounds of Formula (I) under standard amide/sulfonamide formation conditions (e.g., in the presence of a carboxylic chloride or sulfonyl chloride).\n\nThe reactions for preparing compounds of the present disclosure can be carried out in suitable solvents which can be readily selected by one of skill in the art of organic synthesis. Suitable solvents can be substantially non-reactive with the starting materials (reactants), the intermediates, or products at the temperatures at which the reactions are carried out, (e.g., temperatures which can range from the solvent's freezing temperature to the solvent's boiling temperature). A given reaction can be carried out in one solvent or a mixture of more than one solvent. Depending on the particular reaction step, suitable solvents for a particular reaction step can be selected by the skilled artisan.\n\nPreparation of compounds of the present disclosure can involve the protection and deprotection of various chemical groups. The need for protection and deprotection, and the selection of appropriate protecting groups, can be readily determined by one skilled in the art. The chemistry of protecting groups is described, e.g., in Kocienski, *Protecting Groups*, (Thieme, 2007); Robertson, *Protecting Group Chemistry*, (Oxford University Press, 2000); Smith et al., *March's Advanced Organic Chemistry: Reactions, Mechanisms, and Structure*, 6th Ed. (Wiley, 2007); Peturssion et al., \"Protecting Groups in Carbohydrate Chemistry,\" J Chem. Educ., 1997, 74(11), 1297; T. W. Greene and P. G. M. Wuts, Protective Groups in Organic Synthesis, 3rd Ed., Wiley & Sons, Inc., New York (1999); and Wuts et al., *Protective Groups in Organic Synthesis*, 4th Ed., (Wiley, 2006).\n\nReactions can be monitored according to any suitable method known in the art. For example, product formation can be monitored by spectroscopic means, such as nuclear magnetic resonance spectroscopy (e.g., 1H or 13C), infrared spectroscopy, spectrophotometry (e.g., UV-visible), mass spectrometry, or by chromatographic methods such as high performance liquid chromatography (HPLC) or thin layer chromatography (TLC). Compounds can be purified by those skilled in the art by a variety of methods, including high performance liquid chromatography (HPLC) and normal phase silica chromatography.\n\n### Methods of Use\n\nCompounds of the present disclosure can inhibit the activity of WRN and, thus, are useful in treating diseases and disorders associated with activity of PD-1 and the diseases and disorders associated with WRN. In some embodiments, the present disclosure provides a method for inhibiting WRN. In some embodiments, the method includes administering to an individual or a patient a compound of Formula (I), or a pharmaceutically acceptable salt thereof. In some embodiments, the inhibition of WRN is inhibition of WRN helicase activity. In some embodiments, the inhibition of WRN is inhibition of WRN ATPase activity. In some embodiments, the inhibition of WRN is inhibition of WRN helicase activity and WRN ATPase activity. In some embodiments, the inhibition of WRN comprises the inhibition of: (i) WRN helicase activity; or (ii) WRN ATPase activity; or (iii) both (i) and (ii).\n\nIn some embodiments, the present disclosure provides methods of treating a disease or disorder associated with WRN activity, said method comprising administering to a patient in need thereof a therapeutically effective amount of a compound of Formula (I), or a pharmaceutically acceptable salt thereof. In some embodiments, the WRN activity is WRN helicase activity. In some embodiments, the WRN activity is WRN ATPase activity. In some embodiments, the WRN activity is WRN helicase activity and WRN ATPase activity.", "# Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design\nAuthors: Irimia, A.; Serra, A.M.; Sarkar, A.; Jacak, R.; Kalyuzhniy, O.; Sok, D.; Saye-Francisco, K.L.; Schiffner, T.; Tingle, R.; Kubitz, M.; Adachi, Y.; Stanfield, R.L.; Deller, M.C.; Burton, D.R.; Schief, W.R.; Wilson, I.A.\nSource: PLoS Pathogens, 2017\n\nWe noticed that our 10E8 wild-type IgG shows about ten-fold higher neutralization potency compared with results reported previously. To probe the quality of our 10E8 and mutants IgG preparations, we performed size exclusion chromatography-multi-angle light scattering (SECMALS) analysis of the samples. 10E8 and mutant 1 showed the presence of three major peaks in each sample. Each of these peaks had a molecular weight of approximately 150 kDa, consistent with the expected size of monomeric antibody (S6 Fig). We attribute this behavior to a slow conformational isomerization of 10E8, which results in differential interaction with the size exclusion matrix as previously reported for a solubility-optimized 10E8 mutant (10E8v4). Thus, our results suggest conformational isomerization for both 10E8 wild type and mutant 1. In contrast, mutants 2, 3 and 5 each eluted as a single peak with a molecular weight of approximately 150 kDa. The SECMALS analysis shows that the IgG samples used for neutralization are all monomeric with only insignificant amounts (<0.5%) of aggregates observed for wild-type 10E8 and mutant 1 (S6 Fig).\n\n### Structural model of 10E8 binding to the gp41-viral membrane epitope\n\nThe structural, neutralization and binding analyses of wild-type and light-chain 10E8 mutants suggest that the 10E8 variable light-chain residues facilitate approach to and interaction with the viral membrane. To obtain further insights into 10E8 binding to gp41 of HIV Env on the viral surface, and encompassing our recent findings on 4E10 binding to the lipids, we generated a model of 10E8 bound to the viral membrane-gp41 epitope assembly based on the orientation of the lipids and gp41 observed in our crystal structures (Fig 4A) using the CHARMM force field membrane builder. The viral membrane head groups of the outer leaflet were roughly placed in a plane containing the basic side chains of the 10E8 light-chain, Arg17(L), Arg24(L), Arg29(L) and Arg70(L), Lys51(L), the lipid head group, and gp41 Lys/Arg683 (Fig 4A and 4B). Our model suggests that 10E8 epitope includes the gp41 helical peptide (residues 671\u2013683) tilted about 75\u201380\u00b0 from the viral membrane surface and lipid head groups of the membrane. The 10E8 light chain would then face the membrane, with which it interacts via CDRH3, CDRL1, and FRL3, and possibly additional residues (Ser/Thr/Arg/Lys) of the light-chain surface (Fig 4B). Indeed, the relatively large percentage of short polar residues, serine and threonine, on this light-chain \u03b2-sheet surface form a flat polar region that can perhaps also interact with the head groups of the various lipids composing the viral membrane. Fitting of the model into the experimental EM map of the CD4-bound Env (EMDB-5455) shows that 10E8 CDRs H1 and H2 might interact with additional regions of Env (Fig 4D), as also suggested by the cryo-EM structure of 10E8 in complex with the native HIV-1 Env trimer (\u0394CT) and PGT151.\n\n*Fig 4. Integration of experimental data into a 10E8-MPER epitope-viral membrane model.* (A) Surface rendering of the 10E8 Fab (beige, light chain; violet, heavy chain) showing the PG lipid fragment (red-dark cyan sticks) bound in a cavity formed between CDRL1, FRL3 and CDRH3 in the 10E8-T117v2-PG crystal structure. The helical 10E8 epitope residues in the T117v2 scaffold protein (gray) are shown in dark pink and the light-chain surface interacting with the viral membrane is colored by atom type. (B) Model of the 10E8 angle of approach with respect to MPER epitope-viral membrane during antibody engagement. (C) Model of the 4E10 angle of approach with respect to MPER epitope-viral membrane for comparison. In both (B) and (C), the light and heavy chains are shown in beige and violet, respectively. The helical MPER epitope is shown in red and the lipid bilayer as thin green lines. The experimental bound lipids are shown as thicker dark cyan sticks in both structures. (D) and (E) Fitting our models of the viral membrane with 4E10 and 10E8 in the experimental EM map of the CD4-bound Env (EMDB-5455). The light and heavy chains are shown as yellow and blue cartoons. The lipid bilayers are shown in green and the MPER epitope as dark red helices.\n\n## Discussion\n\nThe MPER region of gp41, although highly conserved and targeted by several very broadly neutralizing antibodies (10E8, 4E10 and 2F5), has so far not led to a vaccine immunogen that elicits such antibodies. 4E10 and 2F5 antibodies have been studied in great detail. Their binding to phospholipids prompted the suggestion that such antibodies are rarely produced due to tolerance mechanisms resulting from interaction with 'self-components'. However, the binding of 4E10 to cardiolipin and its overall reactivity profile appears to differ from those of autoimmune antibodies. 10E8, which binds to the same MPER epitope region as 4E10, does not show cross-reactivity with cardiolipin or other autoantigens, although binding to cholesterol-rich liposomes was recently demonstrated.", "# Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\nAuthors: Deller, M.C.; Hudson, K.R.; Ikemizu, S.; Bravo, J.; Jones, E.Y.; Heath, J.K.\nSource: Structure, 2000\n\n*RSA, the relative solvent accessibility of the sidechain. Mutants were tested for inhibition of binding to immobilized gp130 and hLIFR. Activity is expressed as the ratio IC50 mutant/IC50 WT hOSM. IC50 refers to the concentration of mutant or WT protein required to elicit 50% inhibition of binding of WT biotinylated hOSM. Mutants were also tested for biological activity on Ba/F3 cells transfected with gp130/LIFR or hOSMR/gp130. Activity is expressed as the ratio EC50 mutant/EC50 WT hOSM. EC50 refers to the concentration of mutant or WT hOSM required to elicit a response equivalent to 50% of the maximal response to hOSM.\n\n### Identification of site 2\n\nThe interaction of hLIF with gp130 (via site 2) has been previously shown to depend on a cluster of adjacent amino acids in the A and C helices, although the relative significance of each amino acid could not be determined [24]. This issue can be addressed by analysis of hOSM which binds to gp130 with higher affinity than hLIF. Alanine mutations in the individual amino acids comprising the topologically equivalent region of hOSM revealed that the interaction between hOSM and gp130 was critically dependent upon the identity of four structurally adjacent solvent-exposed residues, Gln20, Gly120 and Asn124 (Figure 3a), and, to a lesser degree, Gln16 (Table 3). Each mutant exhibited equivalent activity to WT hOSM in the ability to interact with LIFR (data not shown), but greatly reduced activity in the ability to bind gp130 or stimulate the multiplication of gp130/LIFR (Figure 3b) or gp130/OSMR (Figure 3c) transfected BAF cells. A comparable rank order of activity was observed in both gp130-based solid-phase assays and bioassays; the mutant Asn124Ala exhibited the greatest reduction in activity followed by Gly120Ala and Gln20Ala. We also constructed a mutant in which Gly120 was replaced by tyrosine (Gly120Tyr) thereby emulating the identity of the equivalent residue in CNTF that binds gp130 with very low affinity. This mutant also exhibited reduced ability to bind gp130 and stimulate the multiplication of transfected BAF cells and was significantly less active than the mutant Gly120Ala in both assay formats (Figures 3a\u2013c). Collectively, these findings reveal that the interaction of hOSM with gp130, the site 2 epitope, is governed by four residues located in a cluster on the A and C helices (Figure 4). Of these four residues the most significant, in terms of the consequences of alanine substitution, is Asn124 (in helix C) followed by Gly120 and Gln20. This study also reveals that Gly120 plays a critical role in receptor recognition, in that a single substitution to the bulky tyrosine residue (found in CNTF) almost completely ablates the ability of hOSM to interact with gp130.\n\n*Figure 3.* Functional binding data for hOSM muteins. (a) Competitive inhibition of biotinylated OSM binding to gp130\u2013Fc by site 2 OSM muteins. (b) Biological activity of OSM muteins in the Ba/F3\u2013LIFR/gp130 assay. (c) Biological activity of OSM muteins in the Ba/F3\u2013OSMR/gp130 assay. (d) Competitive inhibition of biotinylated OSM binding to LIFR\u2013Fc by site 3 OSM muteins. (e) Biological activity of OSM muteins in the Ba/F3\u2013LIFR/gp130 assay. (f) Biological activity of OSM muteins in the Ba/F3\u2013OSMR/gp130 assay. (Panels plot OD against mutein concentration in ng/ml for WT OSM and the muteins Q20A, G120A, G120Y, N124A, F160A and K163A.)\n\n*Figure 4.* Structural mapping of binding sites 2 and 3 for hOSM. (a) Ribbon representation of hOSM with solvent-exposed residues studied by mutagenesis depicted in ball-and-stick representation. The solvent-accessible surface of hOSM in the regions of (b) site 2 and (c) site 3 with areas contributed by the mutagenised residues highlighted as coloured patches. For site 2 the orientation is as in (a); for site 3 the view is on to the top of the four-helix bundle. Residues not implicated in gp130 binding are in light-blue, those with some effect are yellow, residues critical for gp130 binding are green and those critical for LIFR binding are red.\n\n### Identification of site 3\n\nThe location of residues that govern the interaction with LIFR (site 3) has been defined for hLIF [24] and CNTF [29]. These comprise a conserved, solvent-exposed, phenylalanine and lysine pair located at the N-terminal end of helix D. These two residues are conserved in all cytokines that bind LIFR [19], including hOSM (Phe160 and Lys163). The availability of a high-resolution structure of hOSM now allows examination of the mechanism by which hOSM, unlike hLIF, is able to bind both LIFR and OSMR, and therefore the extent to which the site 3 recognition mechanism for different receptors is conserved. Two single point substitutions, Phe160Ala and Lys163Ala, and the compound mutant, were tested for biological activity in the assay systems described above (Figures 3d\u2013f; Table 3). Whereas all three mutants exhibited comparable activity to WT hOSM in their ability to interact with gp130 (data not shown), both point mutants and the compound mutant were unable to interact with LIFR (Figure 3d; Table 3). Analysis of the effect of these mutations on the ability to stimulate the multiplication of transfected BAF cells (Figures 3e,f) revealed that for both LIFR- and OSMR-dependent BAF cells, mutations of either Phe160 or Lys163 resulted in a dramatic loss of activity.\n\nThese findings show that the ability of hOSM to interact with LIFR and OSMR is critically dependent upon the identity of this pair of conserved residues and that the mechanism of interaction with both LIFR and OSMR is conserved. This implies that the difference in receptor specificity between hLIF and hOSM and mOSM and hOSM is the result of structural differences between these ligands in the vicinity of the core Phe\u2013Lys motif.\n\n## Discussion", "# The Quantum Leap in Drug Design: Using QM/MM to Predict Perfect Fits\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* Hybrid QM/MM methods combine quantum mechanics and classical mechanics to model drug-protein interactions with higher accuracy than traditional methods.\n* This approach allows for precise modeling of critical interactions, such as enzymatic reactions and the role of metal ions, which are difficult to capture with classical methods.\n* By providing deeper insights into binding energetics and specificity, QM/MM helps accelerate the design and optimization of more effective drug candidates.\n\n1. **The Disruptive Impact of Structural Biology on Biopharmaceutical Innovation:** <https://www.pharmasalmanac.com/articles/the-disruptive-impact-of-structural-biology-on-biopharmaceutical-innovation>\n2. **New Concepts in Drug Discovery (2025):**\n   <https://www.pharmasalmanac.com/articles/the-disruptive-impact-of-structural-biology-on-biopharmaceutical-innovation>\n3. **Structural Biology and Molecular Modeling Market Trends:**\n   <https://www.marketbusinessinsights.com/structural-biology-and-molecular-modeling-market>\n\nIn the world of structure-based drug design, precision is everything. For years, our computational tools for modeling how a drug binds to its target protein have relied on classical mechanics. These methods, while powerful, treat atoms like simple balls on springs. This approximation works well for many interactions but fails to capture the subtle, yet critical, quantum mechanical effects that govern the making and breaking of chemical bonds and other nuanced electronic interactions. This gap in our understanding can be the difference between a successful drug and a failed candidate.\n\nA new wave of computational chemistry is bringing quantum precision to drug discovery. Hybrid Quantum Mechanics/Molecular Mechanics (QM/MM) methods combine the best of both worlds. They use computationally expensive but highly accurate quantum mechanics to model the most critical part of the interaction\u2014the drug molecule and the protein\u2019s active site\u2014while treating the rest of the large protein system with faster classical mechanics. This allows for a level of detail and accuracy that was previously impossible for systems of this size.\n\nWith QM/MM, we can now model enzymatic reactions, understand the role of metal ions in proteins, and accurately predict the energetics of drug binding with unprecedented confidence. This refined understanding helps chemists design molecules with better binding affinity and selectivity, reducing the need for costly and time-consuming trial-and-error synthesis in the lab. As computational power grows, these sophisticated simulation methods are becoming an indispensable tool in the drug hunter\u2019s arsenal, providing deeper insights that accelerate the journey from a good idea to a great medicine.", "# Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly\nAuthors: Deller, M.C.; Johnson, H.A.; Miller, M.D.; Spraggon, G.; Elsliger, M.-A.; Wilson, I.A.; Lesley, S.A.\nSource: PLoS ONE, 2015\n\nData from two isomorphous crystals of the TM1088 complex were merged together for structure determination by MR as implemented by Phaser [40,41], using the high resolution structure of TM1088A (PDB ID: 2G1U) as a search model. During the search for the first monomer, the two possible enantiomorphic space groups were tested, revealing a clear peak for P3(2)21. Searches for the remaining 7 molecules were conducted in this space group. Refinement included 1 TLS group for each of the 12 distinct protein domains observed in the asymmetric unit (4x TM1088A domains and 8x TM1088B domains) and the same domain definitions were used for application of tight NCS restraints. The TM1088 complex was refined to a resolution of 3.45 \u00c5 and converged at Rcryst of 25.9% and Rfree of 30.8%. The difference between Rcryst and Rfree of 4.9% is in excellent agreement with an expected value of 4.97% calculated for this resolution according to Urzhumtsev et al. (\u03bc(\u0394R) ~ 0.024 ln dPDB + 0.020) [42]. Additionally, analysis using the script \"phenix.r_factor_statistics\" in the package PHENIX [34\u201336] reports that Rcryst and Rfree values both fall on the peak of the distribution observed for all PDB structures in this resolution range (3.35\u20133.55 \u00c5). The final model contains 4 TM1088A molecules (Ser7 to Ile138), 4 TM1088B molecules (Met1 to Leu216), 4 AMP molecules and 1 unknown ion modeled as UNX in the asymmetric unit (ASU). It is important to note that all eight protein chains of the octameric TM1088 complex are observed within the ASU, and as such, it is likely that the complex is reflective of a true octameric assembly rather than an artifact of crystal packing. The model has good geometry throughout with the Ramachandran plot produced by MOLPROBITY [39] showing 95.1% of the main chain torsion angles in the favored regions with 0.9% outliers. These outliers comprise TM1088B residues Gly8, Ile139 and Leu142, which, with the exception of Leu142, lie in regions not covered by the high-resolution structure of TM1088B, and therefore contribute to more difficulty in the model building and refinement of these residues at this resolution.\n\nThe TM1088A structure was determined by MR (search model PDB ID: 1LSS) in space group P2(1) using MOLREP [43] and the JCSG MR pipeline [44] and refined at 1.5 \u00c5 resolution using Refmac [33]. Refinement included 4 TLS groups as determined by the TLS Motion Determination server [45]. The refinement converged at an Rcryst of 15.4% and Rfree of 17.5%. The model consists of Lys4 to Ser140, 1 AMP, 2 MPD, 1 sodium ion and 110 water molecules in the ASU. Met(-11) to Lys3 and Glu141 to Asp143, as well as side-chain atoms of Lys4, Glu49, Lys66, Glu100, Lys113, Lys134 and Ser140 were omitted due to weak or absent electron density. The model has excellent geometry throughout with the Ramachandran plot produced by MOLPROBITY [39] showing 99.3% of the main-chain torsion angles in the favored regions with no outliers.\n\nThe structure of the C-terminal domain of TM1088B was determined by MR (search model, PDB ID: 2G1U) in space group C222(1) and refined to 1.55 \u00c5 using Refmac [33]. Refinement included 1 TLS group for each protein chain in the ASU. The refinement of the structure converged at an Rcryst of 15.8% and Rfree of 18.7%. The model consists of Ile140 to Leu216 in chain A, Ile140 to Gly217 in chain B, and 89 water molecules in the ASU. Both protein molecules in the ASU contain additional residues, Asn(-7) to Met(-1), from the uncleaved expression and purification tag. Side-chain atoms from Asn133, Lys162 and Lys208 of chain A and Asn133, Gln144, Glu155, Asp156, Arg171, Glu206 and Glu209 of chain B were omitted due to weak or absent electron density. The model has excellent geometry throughout with the Ramachandran plot produced by MOLPROBITY [39] showing 98.8% of the main-chain torsion angles in favored regions with no outliers.\n\n## Results and Discussion\n\n### The TM1088 locus codes for two proteins that form an octameric assembly\n\nTM1088A and TM1088B were previously identified and functionally characterized as a two-subunit Trk K+ transporter of T. maritima [9]. Expression of the TM1088 locus resulted in two distinct, but structurally homologous, proteins designated TM1088A (16.0 kDa, Pfam ID: PF02254) and TM1088B (24.0 kDa, Pfam ID: PF02254 and Pfam ID: PF02080, Fig 1A, n.b. TM1088A is colored green and TM1088B blue in all figures). TM1088A and TM1088B were purified and crystallized as an octameric assembly consisting of 4 units of TM1088A and 4 units of TM1088B (160 kDa octameric complex; Fig 1B\u20131C) [9]. The differential expression of TM1088A over TM1088B (~2-fold) was previously reported as shown here in Fig 1B (SDS-PAGE gel, inset) [9]. The lower yield observed for TM1088B may be a result of lower expression from the internal RBS or dissociation of the untagged TM1088B from the tagged TM1088A during affinity purification. The latter scenario is supported by SEC analysis, which confirms that the octameric complex is in equilibrium with lower stoichiometries of the complex including dimers (Fig 1B).\n\nA schematic representation of the domain structure of the octameric TM1088 complex is shown in Fig 1D. TM1088A is composed of a single N-terminal Trk domain that is annotated as Pfam ID: PF02254 (Fig 1D, N2 and N2', where the \"prime\" nomenclature is used to denote partner pairs within a dimer). Similarly, TM1088B is also composed of an N-terminal Trk domain, that is also annotated as Pfam ID: PF02254 (Fig 1D, N1 and N1') and a second smaller C-terminal Trk domain annotated as Pfam ID: PF02080 (Fig 1D, C1 and C1').", "# Crystal Structure of an Octomeric Two-Subunit TrkA K+ Channel Ring Gating Assembly from Thermotoga maritima\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [3L4B](https://www.rcsb.org/structure/3L4B)\n* **Release Date:** 2010-03-31\n* **Method:** X-RAY DIFFRACTION 3.45 \u00c5\n* **Organisms:** Thermotoga maritima MSB8\n* **Macromolecules:** TrkA K+ Channel protein TM1088A, TrkA K+ Channel protein TM1088B (proteins)\n* **Unique Ligands:** AMP, UNX\n* **Publication:** [Deller, M.C., et al. (2015). Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly. PLoS ONE, 10, e0122512.](https://doi.org/10.1371/journal.pone.0122512)\n* **Blog Post:** [Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly](https://marcdeller.com/crystal-structure-of-a-two-subunit-trka-octameric-gating-ring-assembly/)\n\nDescription: This structure provides the first structural insights into a two-subunit Trk K\u207a transporter gating ring assembly, representing an evolutionary precursor to modern single-subunit systems. The 160 kDa octameric TM1088A:TM1088B complex reveals AMP-binding at Rossmann-like folds, extensive protein interfaces burying ~18,000 \u00c5\u00b2 of surface area, and thermophilic adaptations including interfacial salt bridges.", "# Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor\nAuthors: Kong, L.; Ju, B.; Chen, Yajing; He, L.; Ren, L.; Liu, J.; Hong, K.; Su, B.; Wang, Z.; Ozorowski, G.; Ji, X.; Hua, Y.; Chen, Yanli; Deller, M.C.; Hao, Y.; Feng, Y.; Garces, F.; Wilson, R.; Dai, K.; O'Dell, S.; McKee, K.; Mascola, J.R.; Ward, A.B.; Wyatt, R.T.; Li, Y.; Wilson, I.A.; Zhu, J.; Shao, Y.\nSource: Immunity, 2016\n\nWe then determined the mutational hotspots by projecting the sequence variation onto a gp140 trimer model of the \"consensus\" Env sequence derived from each time point (Figure 6C). The root mean square fluctuation (RMSF) relative to the consensus sequence was calculated as a measure of the mutational rate for each Env component. Compared to the 2005 Envs, which showed RMSFs of 12.9, 3.6, and 16.8 for gp120, gp41, and gp140, respectively, the 2009 Envs exhibited significantly more variation, with RMSFs of 23.0, 5.3, and 28.2 for the three Env components. This quantitative analysis was confirmed by the visual inspection of mutational hotspots on the trimer surface. For example, the 2008 and 2009 Envs contained more insertions and substitutions in the V1V2 region near the apex and the V4 loop, whereas the CD4bs remained conserved across time points with only peripheral mutations. By contrast, gp41 showed a moderate increase in mutations at the same time points, suggesting that the viral repertoire was evolving with differential mutational rates for gp120 and gp41.\n\nWe next determined the variation of the DRVIA7 footprint on a gp120 core lacking the V1V2 and V3 loops (Figure 6D). Visual inspection revealed that the mutations mainly occurred at gp120 residues interacting with the DRVIA7 heavy chain- and light chain-contacting interface, which accommodates the V5 loop. Of note, the N276 and V5 glycans appeared to be conserved in all Env sequences analyzed. We selected 14 Envs from the phylogenetic tree (Figure 6A), of which 4 were synthesized as trimers and used to compare recognition by VRC01 and DRVIA7 variants (Figure 6E). VRC01, but not DRVIA7, bound to all 4 Envs with high affinity, indicating that the viruses in this donor had actively evolved to escape the DRVIA7 lineage. Such viral escape is commonly observed even for bNAb donors, for example, the VRC01 donor (Wu et al., 2012). Recognition of autologous DRVIA7-resistant Envs was improved after N-terminal truncation of the DRVIA7 light chain, whereas incorporation of the LCDR1 signature alone showed little effect, consistent with the finding that the DRVIA7 lineage began to decline by 2009, probably due to its unfavorable binding and thus poor fitness in B cell selection. Together, SGA and sequence variation analysis revealed a changing face of HIV-1 with a conserved inhibitory glycan patch that raised a significant barrier to the DRVIA7 lineage development during this virus-host co-evolution.\n\n*Figure 6. SGA Analysis of Virus Diversity and Co-evolution.* (A) Neighbor-joining (NJ) phylogenetic tree of 111 non-redundant HIV-1 Env sequences isolated from donor DRVI01 by single genome amplification (SGA). These Env sequences, shown as phylogenetic branches, are color coded by the time point at which the sample was collected. 14 Env sequences selected for experimental validation are labeled as color-coded dots on the tree. (B) NJ tree of non-redundant Env sequences for each individual time point, namely 2005, 2006, 2008, and 2009. The four NJ trees have been rescaled to use the same unit of genetic distance. (C) Trimer models of SGA-derived consensus Env sequences with mutational hotspots labeled on the surface. Top view is shown in panel 1 and side view is shown in panel 2. Molecular surfaces are colored according to sequence conservation. Root-mean-square fluctuation (RMSF) relative to the consensus sequence at each time point is calculated as a diversity score for gp120, gp41, and gp140 and labeled beneath the structural model. (D) The gp120 core models of SGA-derived consensus sequences with CD4bs and mutational hotspots labeled on the molecular surface, which is colored as in (C). The diversity score for gp120 core is labeled beneath each structural model. (E) ELISA binding of VRC01, DRVIA7, and DRVIA7 variants to selected Env trimers at the four time points studied.\n\n## Discussion\n\nVRC01-class antibodies share a common mode of recognition and a set of conserved sequence signatures (Diskin et al., 2011; Wu et al., 2011; Zhou et al., 2010, 2013). Longitudinal analysis previously uncovered an unexpected complexity of the VRC01 lineage (Wu et al., 2015). However, the early events of VRC01 lineage development prior to the timeframe studied still remain elusive (Wu et al., 2015). In this study, we attempted to fill this knowledge gap by investigating DRVIA7, a VRC01-like antibody from a Chinese donor, which presented a different picture of early VRC01 development than that drawn from previous studies of mature VRC01-class antibodies.\n\nFirst, the level of heavy chain SHM required for effective HIV-1 neutralization was perhaps lower than previously estimated. It is speculated that a long period of time would be needed to acquire the somatic mutations in the framework and HCDR1/2 regions for full VRC01-like neutralizing activity (Wu et al., 2010, 2011, 2015; Zhou et al., 2013). This leads to a conclusion that heavy-chain maturation is a barrier and perhaps the time-limiting step in VRC01 development. In contrast to this paradigm, the DRVIA7 heavy-chain precursors identified from 2006 neutralized diverse HIV-1 isolates when paired with mature VRC01 light chain. More importantly, the heavy-chain SHM increased from 17.0% to 22.9% within only 2 years, indicating that the immune system is capable of rapidly maturing the VRC01-like heavy chains.", "# BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n> Passaro, S., Corso, G., Wohlwend, J., Reveiz, M., Thaler, S., Somnath, V.R., Getz, N., Portnoi, T., Roy, J., Stark, H., Kwabi-Addo, D., Beaini, D., Jaakkola, T., Barzilay, R. (2025). Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. *bioRxiv*. https://doi.org/10.1101/2025.06.14.659707\n>\n> Mirdita, M., Sch\u00fctze, K., Moriwaki, Y., Heo, L., Ovchinnikov, S., Steinegger, M. (2022). ColabFold: making protein folding accessible to all. *Nature Methods*. https://doi.org/10.1038/s41592-022-01488-1\n>\n> Schake, P., Bolz, S.N. et al. (2025). PLIP 2025: introducing protein-protein interactions to the protein-ligand interaction profiler. *Nucleic Acids Research*, gkaf361. https://doi.org/10.1093/nar/gkaf361\n>\n> The PyMOL Molecular Graphics System, Version 3.1, Schr\u00f6dinger, LLC.\n>\n> Rego, N., Koes, D. (2015). 3Dmol.js: molecular visualization with WebGL. *Bioinformatics*, 31(8), 1322-1324. https://doi.org/10.1093/bioinformatics/btu829\n\n## \ud83d\udcc4 License\n\n[MIT](LICENSE) \u00a9 Marc C. Deller\n\n## \ud83d\udccb To do\n\n* \u2b1c Share pip cache between the two environments (`PIP_CACHE_DIR`) so `setup-plip` doesn\u2019t re-download wheels the main venv already fetched\n* \u2705 Pin exact dependency versions and add a cached/offline install mode for reproducible installs \u2014 via `pixi.toml`/`pixi.lock` and `install.sh` (see **Alternative: pixi** above and `docs/tier_b_offline_install.md`), not by pinning `setup`/`setup-plip`\u2018s own unpinned `requirements.txt`-driven installs, which remain as they were\n* \u2b1c Add `BoltzMaker.py doctor` \u2014 a post-install check that imports boltz/rdkit/plip/openbabel/pymol in-process and reports exactly which env/feature is broken\n* \u2b1c Add an explicit Boltz model-weights cache dir + a `preflight` check for it (ties into the existing iCloud dataless-file eviction check)\n* \u2b1c Detect the MPS `torch.linalg.svd` CPU-fallback at `preflight` and warn with an estimated runtime multiplier for large multi-chain complexes\n* \u2b1c Add a residue/chain-count-based runtime pre-estimator, plus a toggle to log which ops fall back to CPU on MPS\n* \u2b1c Checkpoint `run` at the per-sample level (not just per-target) so an interrupted multi-hour complex resumes without recomputing completed diffusion samples\n* \u2b1c Add a cross-target selectivity/triage view to the dashboard (confidence-vs-affinity quadrant flags for \u201chigh-confidence, high-affinity\u201d hits)\n* \u2705 Bundle Plotly.js locally instead of via CDN so the dashboard renders fully offline/air-gapped\n* \u2b1c Add a smoke-test suite: an end-to-end fixture run in CI plus unit tests for the `boltz_input.md` parser and JSON-metric flattening\n* \u2705 Add a ligand-preparation/validation step (canonicalization, stereocentre/protonation-state flagging, disconnected-fragment detection) so bad input chemistry is caught before hours of compute, not silently mispredicted\n* \u2b1c Add a retrospective `benchmark` mode: pull known actives/co-crystal data for a target family (ChEMBL/BindingDB/PDB) and report predicted-vs-measured pIC50 correlation + pose RMSD, so a user has a per-target-family trust score before committing to a real campaign\n* \u2705 Add a 3Dmol.js rotating structure view to each target\u2019s binding-site panel, placed next to the existing static PyMOL image (keep the static image, add a \u201cDownload image\u201d link for it \u2014 doesn\u2019t exist yet \u2014 and add the interactive 3Dmol.js view alongside rather than replacing anything)\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.** Structural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/BoltzMaker](https://github.com/bellcheddar/BoltzMaker) |", "# Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\nAuthors: Deller, M.C.; Hudson, K.R.; Ikemizu, S.; Bravo, J.; Jones, E.Y.; Heath, J.K.\nSource: Structure, 2000\n\nAll refinement steps (positional, torsional dynamics, grouped B factor and individual B factor) were carried out using the program CNS-SOLVE [46]. A bulk-solvent correction was employed (solvent density = 0.29 e/\u00c53, B-factor = 31.1 \u00c52), so allowing all data in the resolution range 20.0\u20132.2 \u00c5 to be included. In the final stages of refinement 116 waters were located. The current model consists of residues 4\u2013134 and 156\u2013187 of the mature protein with no residues falling in disallowed regions of the Ramachandran plot. Refinement and model statistics are reported in Table 2.\n\nThe quality of the final model was assessed using the program PROCHECK [47], secondary structure assignments were calculated using the program DSSP [48] and solvent accessibilities calculated using NACCESS [49]. Structural superimpositions were performed using the program SHP [50] and structural figures were produced using BOBSCRIPT [51], RASTER3D [52], VOLUMES (R Esnouf, personal communication) and GRASP [53].\n\n### Accession numbers\n\nAtomic coordinates for hOSM have been deposited with the PDB (accession number 1EVS).\n\n## Acknowledgements\n\nThis work was supported by the Cancer Research Campaign. We thank K. Harlos for help with 'in house' data collection; the EMBL and ESRF staff, in particular V. Stroganoff, for assistance on BM14 at the ESRF, R. Esnouf for provision of computing facilities, V. Barton and L. McGovern for discussion. We are grateful to T. Willis and R. Aplin for performing the protein sequencing and mass spectrometry, respectively. EYJ is a Royal Society University Research Fellow. SI is supported by the Wellcome Trust and MCD was supported by the MRC and CRC.\n\n## References\n\n*(reference list omitted)*", "# GitHub README: PANTS\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/PANTS, 2026\n\n**Every engineered variant now carries a sequence.** Two came from crystal structures,\nwhich is the stronger route: a PDB SEQRES is the construct that was actually expressed,\ncrystallised and assayed, so nothing is applied or assumed. Each was verified by aligning\nagainst its parent and checking the substitution count against the published one, rather\nthan trusting the name on the entry.\n\n| Variant | Source | Length | Substitutions vs parent | Published |\n|---|---|---|---|---|\n| HotPETase | PDB `7QVH` | 272 aa | 21 vs IsPETase | ~21 |\n| Cut190\\*\\*SS | PDB `7CEF` | 262 aa | 4 vs Cut190 (incl. S226P/R228S) | S226P/R228S |\n\nBoth are stored as the **mature construct**, so they are shorter than their precursor parents.\n\n**The remaining three were resolved from their mutation sets, and two more variants were\nfound in the same pass.** Every set below was applied with `apply_mutations`, which refuses\nany substitution whose stated parent residue does not match, so a wrong position or a\nmature-vs-precursor numbering shift fails loudly instead of producing a plausible but wrong\nsequence. All five matched at **offset 0** with the substitution count the papers report.\n\n| Variant | Parent | Mutations | Confirmed by |\n|---|---|---|---|\n| DuraPETase | IsPETase | `S214H/I168R/W159H/S188Q/R280A/A180I/G165A/Q119Y/L117F/T140D` | 10/10 parent residues match; 10 independent positions agreeing by chance is ~20\u207b\u00b9\u2070 |\n| TurboPETase | **BhrPETase** | `H218S/F222I/A209R/D238K/A251C/A281C/W104L/F243T` | 8/8 match; the parent recorded before curation (IsPETase) was **wrong** |\n| Z1-PETase | IsPETase | `N37D/S121E/R132E/A171C/A180V/P181V/D186H/S193C/R224E/N233C/S242T/N246D/S282C` | 13/13 match **and** all 13 sites read the mutant residue in PDB `8H5K` |\n| DepoPETase | IsPETase | `T88I/D186H/D220N/N233K/N246D/R260Y/S290P` | 7/7 match |\n| LCC-A2 | LCC | `F243I/D238C/S283C/Y127G/H218Y/N248D` | 6/6 match (LCC-ICCG plus H218Y/N248D) |\n\nZ1-PETase is confirmed twice over and independently: the mutation list applies cleanly to\nIsPETase, and the deposited `8H5K` sequence differs from the derived one only by an `SHM`\nexpression-tag scar and the signal peptide, with **zero mismatches in the mature region**.\n\nThree findings from that search are worth recording, because each is a trap rather than a\ndetail:\n\n- **TurboPETase's parent is BhrPETase, not IsPETase.** The placeholder entry named the\n  wrong parent. The mutation set does not apply to IsPETase at any offset, so the\n  residue-match check caught it rather than a reviewer.\n- **BhrPETase has no live UniProtKB entry at all.** The accession PAZy records,\n  `A0A2H5Z9R5`, is inactive (DEMERGED to `A0ACD6B9U1`), and `A0ACD6B9U1` is itself DELETED\n  as \"not part of a reference proteome\". The parent of a headline industrial enzyme is\n  reachable only through **UniParc `UPI000CB4D10C`**, which never deletes. Its sequence is\n  byte-identical to PAZy's copy and carries an active EMBL WGS cross-reference.\n- **Do not seed PHL7 from UniProt.** Its only UniProt entry is the **catalysis-deficient\n  S131A mutant** deposited for crystallography: right length, right name, catalytic serine\n  knocked out. The same trap as PDB `7CEH` (S176A), rejected earlier for Cut190\\*\\*SS.\n  PHL7's active sequence is already present as a PAZy-measured positive.\n\nThe PES-H1 `L92F/Q94Y` variant is deliberately not derived, for want of a loadable parent\nrather than a confirmed set. Confirming it did settle a documented literature discrepancy:\nthe same double mutant is written `L92F/Q94Y` in PES-H1 numbering and `L93F/Q95Y` in PHL7\nnumbering, and `find_offset` returns +1 and 0 respectively against the same sequence,\nproducing an **identical** result either way.\n\nValues collated in\n[Engineering Evolution: How FAST-PETase and Other Variants Are Transforming Plastic Biodegradation](https://marcdeller.com/engineering-evolution-how-fast-petase-and-other-variants-are-transforming-plastic-biodegradation/).\n\n| Variant | Topt | Parent | Notes |\n|---|---|---|---|\n| Z1-PETase | 30 \u00b0C | IsPETase | 13 mutations, two engineered disulfides; 40x expression yield |\n| IsPETase (wild type) | 30 to 35 \u00b0C | native | Weak on crystalline PET |\n| **DuraPETase** | **37 \u00b0C** | IsPETase | 10 mutations; +31 \u00b0C thermostability, ~300x activity |\n| FAST-PETase | 50 \u00b0C | ThermoPETase | 38x activity; 33.8 mM monomers in 96 h |\n| HotPETase | 60 to 65 \u00b0C | IsPETase | 21 mutations; melting temperature 82.5 \u00b0C |\n| Cut190\\*\\*SS | 65 \u00b0C | actinomycete cutinase | Calcium-dependent conformational switching |\n| TurboPETase | 65 to 68 \u00b0C | BhrPETase | 98.2% depolymerisation at 200 g/kg in 8 h |\n| LCC-ICCG | 65 to 72 \u00b0C | LCC | 1.3 g PET waste in 3 days from 1.25 mg enzyme |\n| DepoPETase | ~50 \u00b0C (applied) | IsPETase | 7 mutations; melting temperature +23.3 \u00b0C, ~1407x product |\n| LCC-A2 | 78 \u00b0C | LCC-ICCG | LCC-ICCG plus H218Y/N248D |\n\n**A discrepancy worth stating rather than smoothing over:** UniProt's curated value for\nIsPETase is 40 \u00b0C, while the literature review gives 30 to 35 \u00b0C. Both are defensible and\nthey were measured on different substrates under different assays, which is exactly the\nharmonisation problem the project brief raises. The database keeps the UniProt value\nbecause it carries an experimental evidence code and a citation; this table shows both\nrather than picking a winner.\n\n**Read the two tables together and the shape of the field is obvious.** The entire\nengineered lineage runs *away* from body temperature: FAST-PETase to 50 \u00b0C, HotPETase to\n60 to 65, TurboPETase and LCC-ICCG to 65 to 72, LCC-A2 to 78. That is rational, because\nthey were optimised for a reactor above PET's glass transition.", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nIn some embodiments, R^2 is selected from halo, C1\u20136 alkyl and C1\u20136 haloalkyl. In some embodiments, R^2 is selected from F, Cl, CH3, and CF3. In some embodiments, R^2 is selected from CN, C1\u20136 alkyl, OR^a2, C(O)R^b2, C(O)NR^c2R^d2, and C(O)OR^a2. In some embodiments R^2 is selected from CN, Me, CO2Me, the 2-hydroxyethyl amide \u2014C(O)NHCH2CH2OH, the 2-methoxyethyl amide \u2014C(O)NHCH2CH2OMe, and a 3-hydroxy-3-azabicyclo[3.1.0]hexane-3-carbonyl group. In some embodiments, R^2 is CN. In some embodiments, R^2 is Me. In some embodiments, R^2 is CO2Me.\n\nIn some embodiments, R^3 is selected from H, D, OH, CN, halo, C1\u20136 alkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, C1\u20136 haloalkyl, cyano-C1\u20136 alkyl, HO\u2014C1\u20136 alkyl, and C1\u20136 alkoxy-C1\u20136 alkyl. In some embodiments, R^3 is selected from CN and C1\u20136 alkyl. In some embodiments, R^3 is CN or \u2014CH2CH3. In some embodiments, R^3 is C1\u20136 alkyl. In some embodiments, R^3 is CN. In some embodiments, R^3 is \u2014CH2CH3.\n\nIn some embodiments, R^4 and R^5 are each independently H, or C1\u20136 alkyl; or R^4 and R^5, together with the carbon atom to which they are attached, form a 3\u20137-membered cycloalkyl ring. In some embodiments, R^4 and R^5 are each independently H or CH3. In some embodiments, R^4 and R^5 are each H.\n\nIn some embodiments, R^6 is selected from C3\u20137 cycloalkyl, phenyl, 4\u20137 membered heterocycloalkyl, and 5\u20136 membered heteroaryl, each of which is optionally substituted by 1, 2, 3 or 4 independently selected R^6A substituents. In some embodiments, R^6 is selected from C5\u20136 cycloalkyl, phenyl, 5\u20136 membered heterocycloalkyl and 5\u20136 membered heteroaryl. In some embodiments, R^6 is independently selected from C6 cycloalkyl and phenyl. In some embodiments, R^6 is phenyl, which is optionally substituted by 1 or 2 independently selected R^6A substituents. In some embodiments, R^6 is cyclohexyl, which is optionally substituted by 1 or 2 independently selected R^6A substituents. In some embodiments, each R^6A is independently selected from F, C1, \u2014CH3, \u2014CHCH3, and \u2014CF3.\n\nIn some embodiments, R^6 is independently selected from cyclohexyl, phenyl, 4-ethylphenyl, 4-fluorophenyl, 3-(trifluoromethyl)phenyl, 4-(trifluoromethyl)phenyl, 2-chloro-4-fluorophenyl, 3,4-difluorophenyl, 2-chloro-4-(trifluoromethyl)phenyl, and 2-methyl-4-fluorophenyl.\n\nIn some embodiments, R^7 is selected from H, D, CN, halo, C1\u20136 alkyl, and C1\u20136 haloalkyl. In some embodiments, R^7 is selected from H and CN. In some embodiments, each R^7 is H. In some embodiments, each R^7 is CN.\n\nIn some embodiments, L^2 is -L-C(O)\u2014, -L-NR^9C(O)\u2014 or -L-NR^9S(O)2\u2014, wherein L^2 is attached to Ring moiety A through an L linking group. In some embodiments, L^2 is -L-NR^9C(O)\u2014 or -L-NR^9S(O)2\u2014. In some embodiments, each L is independently a bond or C1\u20136 alkylene. In some embodiments, L is a bond.\n\nIn some embodiments, each R^81 and R^83 is independently selected from H, D, halo, CN, C1\u20136 alkyl, C1\u20136 haloalkyl, C3\u20134 cycloalkyl, and OR^a8. In some embodiments, each R^82 is independently selected from H, D, halo, CN, C1\u20136 alkyl, C1\u20136 haloalkyl, and C3\u20134 cycloalkyl. In some embodiments, each R^81 and R^83 is independently selected from H, methyl, ethyl, CN, isopropyl, and ethoxy; and each R^82 is independently selected from H, methyl, ethyl, CN, and isopropyl. In some embodiments, each R^82 is H. In some embodiments, R^9 is selected from H, halo, C1\u20136 alkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, C1\u20136 haloalkyl, and C3\u20137 cycloalkyl. In some embodiments, R^9 is H or C1\u20136 alkyl. In some embodiments, R^9 is H or methyl. In some embodiments, R^9 is H.\n\nIn some embodiments, compounds of Formula (I), or a pharmaceutically acceptable salt thereof, can include one or more hydrogen atoms replaced by deuterium. In some embodiments, 1, 2, 3, 4, 5, 6, 7, or 8 hydrogen atoms, attached to carbon atoms of \"alkyl\", \"alkenyl\", \"alkynyl\", \"aryl\", \"phenyl\", \"cycloalkyl\", \"heterocycloalkyl\", or \"heteroaryl\" substituents or \"\u2014C1\u20134 alkyl-\" and \"alkylene\" linking groups, as described herein, are optionally replaced by deuterium atoms.\n\nIn some embodiments, the substituent R^2 can replace any H atom on Ring moiety A. However, the substituent R^2 cannot replace atoms or groups other than H, such as R^W.\n\nIt is further appreciated that certain features of the present disclosure, which are, for clarity, described in the context of separate embodiments, can also be provided in combination in a single embodiment (while the embodiments are intended to be combined as if written in multiply dependent form). Conversely, various features of the present disclosure which are, for brevity, described in the context of a single embodiment, can also be provided separately or in any suitable subcombination. Thus, it is contemplated as features described as embodiments of the compounds of Formula (I) can be combined in any suitable combination.\n\nAt various places in the present specification, certain features of the compounds are disclosed in groups or in ranges. It is specifically intended that such a disclosure include each and every individual subcombination of the members of such groups and ranges. For example, the term \"C1\u20136 alkyl\" is specifically intended to individually disclose (without limitation) methyl, ethyl, C3 alkyl, C4 alkyl, C5 alkyl and C6 alkyl.\n\nThe term \"n-membered\" where n is an integer typically describes the number of ring-forming atoms in a moiety where the number of ring-forming atoms is n. For example, piperidinyl is an example of a 6-membered heterocycloalkyl ring, pyrazolyl is an example of a 5-membered heteroaryl ring, pyridyl is an example of a 6-membered heteroaryl ring, and 1,2,3,4-tetrahydro-naphthalene is an example of a 10-membered cycloalkyl group.", "# GitHub README: ChemSage\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ChemSage, 2026\n\n- All 17 corpus CSVs searchable directly from the chat CLI via keyword fast-path (`rag/corpus_lookup.py`) that bypasses the LLM for enumeration queries\n- Up to 500 matching rows loaded per source; pageable in 50-row pages with CSV export\n- PDB binding affinity cross-reference: resolves target names to PDB IDs via `pdb_all_entries.csv`\n- Deduplication by `comp_id` sorted by `comp_mw` descending (highest-MW representative per compound)\n- Intent filtering: SMILES/affinity/sequence queries drop irrelevant sources from the menu\n- `_CAPS_STOPWORDS` prevents format words (\"SMILES\") from winning keyword extraction over target names (\"EGFR\")\n- RAG hybrid loop: top-5 corpus chunks per query via ChromaDB (BAAI/bge-base-en-v1.5, 55k chunks)\n- RDKit auto-execution of code blocks in model responses\n- ASCII art banner (ChemSage, `small_slant` font), Rich Markdown rendering, Rich `Table` corpus tables with rounded cyan borders\n- Column display config (`_COL_CFG`) maps all corpus column names to abbreviated headers and per-column max widths; ellipsis overflow keeps all rows single-line\n\n## \ud83c\udf10 Web deployment\n\n**Live at [chemsage.mdeller.com](https://chemsage.mdeller.com).** Not a rebuilt chat-bubble\ninterface: the browser gets a real terminal (xterm.js) connected over WebSocket to a\nserver-side PTY running the actual, unmodified `scripts/chat.py` \u2014 the same Rich-rendered\nmarkdown, corpus tables, and slash commands (`/save`, `/info`, `/retry`, tab completion) as\nrunning it locally.\n\n**No GPU on the droplet.** `chat_remote.py` is a drop-in launcher: at import time it patches\n`mlx_lm.generate`/`mlx_lm.stream_generate` to call a HuggingFace Space running the real fused\n32B model on ZeroGPU (a serverless, on-demand GPU allocation), then runs `chat.py` completely\nunaware anything is different. `chat.py` itself is never modified for the hosted path.\n\n**Serving stack:** gunicorn with a single `gevent` worker (100 concurrent connections) rather\nthan `sync` workers \u2014 a terminal session is I/O-bound (waiting on the remote model and on\nkeystrokes), not CPU-bound, so one greenlet-based worker serves many concurrent sessions\ncheaply. nginx reverse-proxies both the WebSocket endpoint (`/ws`, with `Upgrade`/`Connection`\nheaders and a long read/send timeout for idle-but-open sessions) and the static app shell,\nbehind a Let's Encrypt certificate. Runs as a dedicated systemd service on the droplet\nalongside AlphaFraud and BoltzMaker's own web apps.\n\n**Source isolation:** the web app (`web/flask_app/`) lives on a separate `web` git\nbranch/worktree, so it never touches the tested training/eval code on `main`.\n\n## \ud83e\uddf1 Stack\n\nEverything runs locally on **Apple Silicon** via **MLX-LM** (`mlx_lm.lora` to train, `mlx_lm.fuse`\nto merge, `mlx_lm.server` for an OpenAI-compatible endpoint the GUI talks to). Unified memory means\nno VRAM wall: 16 GB handles ~8B, 32 GB ~14B, 64 GB ~32B. QLoRA is automatic because the base is a\n4-bit MLX model.\n\nThe chat CLI (`scripts/chat.py`) uses **Rich** for streaming Markdown output (`Live`), corpus tables\n(rounded borders, ellipsis overflow, abbreviated headers), panels, progress spinners, and a themed\nconsole; **prompt_toolkit** for the interactive input session (arrow-key history, slash-command tab\ncompletion, styled prompt, persistent bottom toolbar with context window indicator); **ChromaDB** +\n**sentence-transformers** for the RAG retrieval layer; and **RDKit** for live chemical property\ncomputation. Responses stream token-by-token via `mlx_lm.stream_generate` and re-render as Markdown\non completion. New commands: `/save` (export conversation to Markdown), `/info` (session/model\ndetails), `/retry` (regenerate last response).\n\n## \u2705 To Do\n\n### Completed (Rounds 1-5)\n- [x] Increase SFT dataset from 1,500 to 5,000 examples with code-heavy generators (Round 3)\n- [x] Add examples that teach \"emit RDKit code, don't compute manually\" (Round 3: `single_property`, `full_profile`)\n- [x] Include SMILES validation examples (Round 3: `smiles_validation`)\n- [x] Increase `iters` from 750 to 1,500 (Round 3; stopped early at 625 \u2014 val loss plateaued)\n- [x] Try `num_layers: 64` \u2014 caused swap thrashing at 31.5 GB; reverted to 32\n- [x] Try `batch_size: 8` \u2014 4.5x slower due to memory pressure; reverted to 4\n- [x] Add structural biology generators: PyMOL selections/viz, PLIP, PDB format, pdb-tools, Biopython, ChimeraX, gemmi\n- [x] Add Plotly and R/ggplot2/bio3d recipe generators\n- [x] Add protein family generators: kinases, GPCRs, antibodies, GTPases, cytokines, receptors (real PDB data)\n- [x] Add numerical fidelity drills: `code_then_quote` (3x), `rounding_precision` (2x)\n- [x] Add refusal/out-of-scope examples\n- [x] Expand dataset to 8,000 examples across 51 behaviour classes\n\n### Round 5 dataset (complete)\n- [x] 20,000 examples across 78 behaviour classes (`build_dataset.py --n 20000 --seed 51`)\n- [x] 19 new generator classes targeting R4 exec/PyExec gap: `pyexec_drill` (\u00d76), `code_then_quote_v2` (\u00d75), `rounding_explicit` (\u00d75), `fidelity_multistep` (\u00d75), `herg_liability`, `selectivity_profile`, `prodrug_bcs`, `sar_delta`, `mdanalysis`, `dssp`, `ppi_interface`, `uniprot_api`, `conformer_3d`, `mcs_search`, `reaction_smarts`, `recap_fragmentation`, `electron_density`, `biological_assembly`, `drug_target_family`\n- [x] Token lengths p50=322, p99=488, max=558 \u2014 all under max_seq_length=3072\n- [x] 707 examples rejected (96.5% pass rate); 78 classes, class balance max:median < 3:1\n- [x] `data/sft_r5/` \u2014 16,000 train / 2,000 val / 2,000 test", "# TYK2 Biology: How Deep Mutational Scanning Reveals New Therapeutic Frontiers\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe integration of deep mutational scanning with [computational variant effect prediction](https://www.nature.com/articles/s41588-023-01465-0) tools represents a powerful synergy for interpreting the clinical significance of rare human variants. While TYK2 loss-of-function variants protect against multiple autoimmune conditions including psoriasis, Crohn\u2019s disease, and multiple sclerosis, complete TYK2 deficiency causes severe immunodeficiency with susceptibility to mycobacterial infections. The DMS approach helps define this therapeutic window by quantifying exactly how much functional impairment provides benefit without causing immunosuppression. These findings have direct translational relevance as they inform optimal target engagement levels for clinical inhibitors and help predict which patients carrying specific TYK2 variants might respond differently to therapy. The work establishes a paradigm where prospective functional assessment of all possible mutations informs both precision medicine approaches and structure-based drug design simultaneously.\n\n| Key Concept | Description | Key References |\n| --- | --- | --- |\n| Deep Mutational Scanning (DMS) | High-throughput technique that systematically assesses thousands of genetic variants to map sequence-function relationships and identify functionally important residues | [Matreyek et al., eLife](https://elifesciences.org/articles/104725) |\n| TYK2 Pseudokinase Domain | Catalytically inactive JH2 domain that allosterically regulates the active kinase domain and serves as a selective drug target distinct from other JAK family members | [Lupardus et al., PNAS](https://www.pnas.org/doi/10.1073/pnas.1401180111) |\n| Allosteric Inhibition | Drug binding at regulatory sites away from the active site that modulates protein function through conformational changes, enabling high selectivity | [Tong et al., Biochem Soc Trans](https://pmc.ncbi.nlm.nih.gov/articles/PMC10089111/) |\n| Human Variant Interpretation | Assessment of disease-associated mutations to understand protective versus pathogenic effects on protein function and autoimmune disease risk | [Dendrou et al., Clin Exp Immunol](https://pmc.ncbi.nlm.nih.gov/articles/PMC11885862/) |\n| Protein Degradation Strategy | Therapeutic approach targeting reduced protein abundance rather than inhibition, offering alternative mechanism for modulating TYK2 activity in autoimmunity | [Hiatt et al., bioRxiv](https://www.biorxiv.org/content/10.1101/2025.10.11.681520v1) |", "# Remapping the Chemical and Pharmacological Spaces of Drugs: Navigating Tomorrow\u2019s Therapeutic Frontier\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Key Concept | Description | Key Reference |\n| --- | --- | --- |\n| UMAP Clustering | Reduces high-dimensional fingerprints to 2D, preserving local/global structure. | [Santos, et al., PLoS Computational Biology](https://pmc.ncbi.nlm.nih.gov/articles/PMC11207054/) |\n| t-SNE & GTM | Non-linear methods outperform PCA in neighborhood preservation for large datasets. | [Li, et al., Journal of Chemical Information and Modeling](https://pmc.ncbi.nlm.nih.gov/articles/PMC11733715/) |\n| Aromatic Rings | Present in 81% of approved drugs, critical for stability and lead optimization. | [Santos, et al., PLoS Computational Biology](https://pmc.ncbi.nlm.nih.gov/articles/PMC11207054/) |\n| Pharmacological Shifts | Kinase inhibitors remain dominant, epigenetic writers emerging post-2020. | [Santos, et al., PLoS Computational Biology](https://pmc.ncbi.nlm.nih.gov/articles/PMC11207054/) |\n| AI-Driven Chemography | Graph embeddings and generative models uncover novel scaffolds and repurposing leads. | [G\u00f3mez-Bombarelli, et al., Trends in Pharmacological Sciences](https://www.sciencedirect.com/science/article/pii/S1359644625001059) |\n| Explainable AI | Ensures model transparency for scientific validation and regulatory compliance. | [Tih\u00e1nyi, et al., Frontiers in Drug Discovery](https://www.frontiersin.org/journals/drug-discovery/articles/10.3389/fddsv.2025.1674289/full) |", "# The Shape-Shifting World of Intrinsically Disordered Region Binding Proteins\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-22 on marcdeller.com (categories: sb).*\n\n* IDRs lack stable structures, making them historically difficult to target.\n* New computational strategies enable design of binders that match IDR flexibility.\n* Deep learning models and AI drive rapid identification of selective binders.\n* Designed proteins achieve picomolar-to-nanomolar binding affinities.\n* Applications span disease intervention, diagnostics, and cellular signaling studies.\n* The breakthrough offers a path to targeting \u201cundruggable\u201d regions in the human proteome.\n\n1. Science (2025): **Design of intrinsically disordered region binding proteins:**\n   <https://www.science.org/doi/10.1126/science.adr8063>\n2. PubMed Central: **Peer-reviewed article version:**\n   <https://pmc.ncbi.nlm.nih.gov/articles/PMC11275711/>\n3. Institute for Protein Design: **Overview of generative AI in binder design:**\n   <https://www.ipd.uw.edu/2025/07/hitting-undruggable-disease-targets/>\n4. C&EN: **News on AI-assisted methods targeting \u201cundruggable\u201d proteins**:\n   <https://cen.acs.org/physical-chemistry/protein-folding/New-AI-assisted-methods-take/103/web/2025/07>\n\nFor decades, the challenge of understanding and targeting intrinsically disordered proteins (IDPs) and their regions (IDRs) has been a central puzzle in molecular biology and drug discovery. Unlike classical proteins with well-defined shapes, IDRs lack stable, rigid structures and instead exist as flexible, fluctuating chains. This intrinsic chaos enables them to participate in diverse signaling pathways and play critical roles in health and disease. However, their high variability and elusive forms have made them nearly impossible to target using conventional protein-binding strategies. A major breakthrough has emerged in 2024\u20132025, as pioneering research from the Baker Lab and collaborators unveiled a universal framework for designing proteins that can selectively bind to IDRs with atomic precision. Rather than trying to anchor onto a static structure, the new approach leverages advanced computational design and artificial intelligence to generate binder proteins that are themselves flexible and able to conform to the multitude of shapes presented by IDRs.\n\nOver 39 highly diverse unstructured targets were tackled, yielding binding affinities that reached from picomolar to nanomolar levels\u2014remarkable feats of molecular recognition. The approach works by carefully sculpting binding surfaces that complement the dynamic side-chain arrangements of disordered regions, creating \u201clock-and-key\u201d fits for moving targets. As evidence, designs were thoroughly validated in cells and as detection reagents, reliably demonstrating specificity and selectivity, with minimal off-target effects across all-by-all binding experiments. The impact is already echoing in applications such as disabling prion seeds, dismantling toxic protein aggregates, and arresting pain signaling pathways within cells. This work now positions researchers much closer to taming the \u201cundruggable\u201d and addressing untreatable diseases rooted in IDP-driven signaling networks. The paradigm shift blends deep learning models, advanced molecular simulation, and high-throughput screening to forge a universal playbook\u2014a leap that sets the stage for a new era in designing precision protein therapeutics and diagnostics against the most elusive molecular players in biology.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\nPCR was carried out in thin walled PCR tubes (Sigma) to aid heat transfer thus reducing the number of cycles required. Several polymerases were tested including Amplitaq (Perkin Elmer) and Vent Polymerase (New England Biolabs). Amplitaq has no 3'-5' exonuclease 'proof reading' activity, so low nucleotide concentrations were used to minimise error levels. Alternatively, Vent Polymerase (New England Biolabs) which does possess 3'-5' exonuclease activity was used as the enzyme of choice. This enzyme is known to work on a smaller subset of template and primer combinations. The reaction mixtures used are shown in Table 2-2 below. PCR buffer (10x) contained 100 ng/\u03bcl bovine serum albumin (BSA), 1 % Triton X-100, 50 mM ammonium sulphate, 100 mM potassium chloride and 200 mM Tris-HCl, pH 8.2.\n\nDeoxynucleotide triphosphates (dNTP) were supplied as 100 mM solutions in water. A cocktail containing each nucleotide was produced with a final concentration for each nucleotide of 10 mM. A control reaction, containing no template DNA, was run in parallel with the amplification reaction.\n\n| Component | Amplitaq Volume (\u03bcl) | Vent Polymerase Volume (\u03bcl) |\n|---|---|---|\n| 10x PCR Buffer | 20.0 | 20.0 |\n| 25 mM magnesium chloride | 12.0 | \u2013 |\n| 10 mM dNTP's | 1.0 | 4.0 |\n| 5' pGEX Primer (200 ng/\u03bcl) | 1.0 | 4.0 |\n| C-Terminal Truncation PCR Primer (200 ng/\u03bcl) | 1.0 | 4.0 |\n| pGEX-3C-OSM-196 Template DNA (100 ng) | 2.0 | 2.0 |\n| Polymerase | 1.0 | 1.0 |\n| Water | 162.0 | 165.0 |\n| Total | 200.0 | 200.0 |\n\n*Table 2-2.* Table showing composition of PCR reaction mixtures used.\n\nEach reaction mixture was split into two equal aliquots and one drop of mineral oil (Sigma) added to each to prevent condensation. The PCR reaction cycle was carried out automatically using a PTC-200 DNA Engine (MJ Research Inc.). A 94 \u00b0C 'hot start' was used to prevent 'primer dimer' formation. Preparation of the reaction mixture was carried out on ice and polymerase added last. The reaction mixture was placed immediately into the PCR machine equilibrated to 94 \u00b0C.\n\nInitial denaturation of the double stranded DNA was carried out by incubation at 96 \u00b0C for 1 minute, subsequent denaturation cycles were carried out for 30 seconds. Annealing of primers to their complementary sequences was carried out by rapid cooling to 45 \u00b0C followed by 45 seconds incubation. Primer extension was carried out by incubation of the reaction mixture for 45 seconds at 72 \u00b0C, with the final extension being increased to 5 minutes. Twenty five cycles of denaturation-annealing-extension were carried in order to strike a balance between complete amplification and maintenance of fidelity. The reaction cycles are summarised below in Table 2-3.\n\n| Cycles | 1) Denaturation Temp / Time (min) | 2) Annealing Temp / Time (min) | 3) Extension Temp / Time (min) |\n|---|---|---|---|\n| Start to 1 | 96 \u00b0C / \u2014 | \u2014 | \u2014 |\n| 1 to 2 | 96 \u00b0C / 1.00 | 45 \u00b0C / 0.75 | 72 \u00b0C / 0.75 |\n| 2 to 23 | 96 \u00b0C / 0.50 | 45 \u00b0C / 0.75 | 72 \u00b0C / 0.75 |\n| 23 to End | \u2014 | \u2014 | 72 \u00b0C / 5.00 |\n| End | 4 \u00b0C / \u221e | \u2014 | \u2014 |\n\n*Table 2-3.* Table showing temperatures and incubation times used for automated PCR amplification.\n\nThe mineral oil was carefully removed by pipette and 5 \u03bcl samples of each reaction mixture were added to 1 \u03bcl of ethidium bromide (10 mg/ml stock) and 1 \u03bcl of DNA loading buffer containing 0.25 % bromophenol blue, 0.25 % xylene cyanol FF and 30 % glycerol (see Appendix I). The samples were run on a 1 % agarose gel in TBE buffer (54 g Tris-base, 27.5 g boric acid and 20 ml 0.5 M ethylenediaminetetraacetic acid (EDTA), pH 8 upto 500 ml for 5x stock) (see Appendix I). The gel was run for approximately 30 minutes at 60 volts and the bands observed using a UV transilluminator (Ultra-Products Inc.). The amount of PCR product was estimated by comparison with the brightness of a known quantity of equimolar \u03bb HindIII digestion fragments (Gibco).\n\n### 2.2.6 Purification of PCR Amplified DNA\n\nThe PCR amplified DNA was doubly digested with EcoRI and BamHI to remove template DNA. The reaction mixture was setup as below in Table 2-4. Restriction buffer B contained 1 mM \u03b2-mercaptoethanol, 100 mM sodium chloride, 5 mM magnesium chloride and 10 mM Tris-HCl, pH 8.0.\n\n| Volume (\u03bcl) | Component |\n|---|---|\n| 40.0 | PCR Product |\n| 8.0 | 10x Restriction Buffer B |\n| 1.0 | EcoRI (40 units/ml) |\n| 6.0 | BamHI (10 units/ml) |\n| 25.0 | Sterile Distilled Water |\n| 80.0 | Total |\n\n*Table 2-4.* Table showing composition of restriction enzyme reaction mixture used to remove template DNA from amplified PCR products.\n\nDigestion was carried out by incubation in a water bath at 37 \u00b0C for 1.5 hours. DNA loading buffer and ethidium bromide were added as previous and the samples run in two large lanes on a 1 % agarose gel in TAE buffer (242 g Tris-base, 57.1 ml glacial acetic acid and 100 ml 0.5 M EDTA, pH 8 upto 500 ml for 50x stock) (see Appendix I). The lower band was excised from the gel using a sterile razor blade and then trimmed tightly to remove unwanted agarose. The band was placed in a 1.5 ml eppendorf tube and extraction of the digested PCR product was carried out using a Qia-Quick PCR Purification Kit (Qiagen).\n\n### 2.2.7 Ligation\n\nPrior to ligation the concentration of the final purified PCR product was determined from a 1 % agarose-TAE gel. Ligation was carried out at a molar ratio of 10:1 (insert:vector). Reaction mixtures containing the enzyme T4 ligase (Boehringer Mannheim) were used as in Table 2-5 below. Ligation buffer contained 10 mM ATP, 50 mM magnesium chloride, 10 mM DTT and 660 mM Tris-HCl, pH 7.5. The vector was previously digested with BamHI and EcoRI and purified as in 2.2.6. Acetylated BSA was included to stabilise and protect the T4 ligase enzyme. A control reaction containing no PCR product was set-up along side and both reactions incubated at room temperature for 4 hours to allow ligation to proceed.", "# GitHub README: chatRDKit\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/chatRDKit, 2026\n\n> **Talk to RDKit in plain English. Cheminformatics without the boilerplate.**\n\n    \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/chatRDKit\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/chatRDKit</a></td>\n</tr>\n</table>\n\n---\n\nA natural-language chat frontend for **RDKit** (the open-source\ncheminformatics toolkit), powered by a local **Ollama** model (default:\n`qwen3.6:27b`).\n\nWhy it matters: RDKit is the backbone of open-source cheminformatics, but its API is broad and unforgiving, and every routine task (computing descriptors, drawing a molecule, running a similarity search) means recalling the exact module, class, and function names. chatRDKit removes that friction by letting you describe what you want in natural language and translating it into correct RDKit code on the fly, all running locally against an Ollama model so no structures ever leave your machine. It is useful for medicinal chemists, structural biologists, and anyone who wants RDKit's power at the speed of thought: rapid property triage, on-the-fly depiction, Lipinski checks, and fingerprint comparisons, without breaking flow to look up syntax.\n\nType plain English at the prompt:\n\n```\nchatRDKit> Load aspirin and tell me its molecular weight, LogP and TPSA\n[chatRDKit] >>> aspirin = Chem.MolFromSmiles(\"CC(=O)Oc1ccccc1C(=O)O\")\n[chatRDKit] >>> print(\"MW:\", Descriptors.MolWt(aspirin))\n[chatRDKit] >>> print(\"LogP:\", Descriptors.MolLogP(aspirin))\n[chatRDKit] >>> print(\"TPSA:\", Descriptors.TPSA(aspirin))\nMW: 180.15899999999996\nLogP: 1.3101\nTPSA: 63.60000000000001\n```\n\nOr type a **SMILES string** directly to load a molecule, or **Python/RDKit\ncode** to run it as-is. Everything runs locally in a persistent Python\nsession -- no data leaves your machine.\n\n---\n\n## 1. Prerequisites\n\n### 1.1 Ollama + model\n\n1. Install Ollama:\n   - **macOS**: download the app from <https://ollama.com/download>, or\n     `brew install ollama`\n   - **Linux**: `curl -fsSL https://ollama.com/install.sh | sh`\n   - **Windows**: download the installer from <https://ollama.com/download>\n\n2. Make sure the server is running. The desktop app starts it\n   automatically; otherwise:\n\n   ```sh\n   ollama serve\n   ```\n\n3. Pull a Qwen3.6 model:\n\n   ```sh\n   ollama pull qwen3.6:27b\n   ```\n\n   This is the default chatRDKit uses (a ~27.8B-parameter model, ~17 GB on\n   disk) and gives the best translation quality. It needs roughly 20 GB+ of\n   free RAM/VRAM to run comfortably. If your machine can't handle that:\n\n   ```sh\n   # Much smaller/faster, good for trying chatRDKit out, lower translation quality:\n   ollama pull qwen3:0.6b\n   ```\n\n   Then point chatRDKit at whichever tag you pulled (see\n   [Configuration](#3-configuration)) -- just make sure the `model` setting\n   matches a tag from `ollama list` exactly.\n\n4. Sanity check the server is reachable and the model responds:\n\n   ```sh\n   curl http://localhost:11434/api/chat -d '{\n     \"model\": \"qwen3.6:27b\",\n     \"messages\": [{\"role\": \"user\", \"content\": \"say hi\"}],\n     \"stream\": false\n   }'\n   ```\n\n   You should get back a JSON response with a `message.content` field.\n\n### 1.2 Python + RDKit\n\nchatRDKit needs Python 3.10+ and the latest RDKit. A virtual environment is\nrecommended:\n\n```sh\ncd chatRDKit\npython3 -m venv .venv\nsource .venv/bin/activate          # Windows: .venv\\Scripts\\activate\npip install -r requirements.txt\n```\n\n`requirements.txt` installs:\n- `rdkit` -- the latest release from PyPI (chemistry engine + drawing)\n- `reportlab` -- used for PDF report output (optional; PNG/HTML work\n  without it)\n\n---\n\n## 2. Running chatRDKit\n\n```sh\nsource .venv/bin/activate\npython chatrdkit.py\n```\n\nOn launch you'll see a welcome banner with 5 numbered example prompts:\n\n```\n========================================================================\n  chatRDKit -- chat with RDKit (2026.03.3) using natural language\n========================================================================\n\nTry one of these to get started (type the number, or your own request):\n  1. Enter a SMILES string to load a molecule (e.g. CC(=O)Oc1ccccc1C(=O)O for aspirin)\n  2. What is the molecular weight, LogP and TPSA of caffeine?\n  3. Draw ibuprofen and save it as an image\n  4. Generate a 3D structure of ethanol and optimise its geometry\n  5. Compare the similarity between aspirin and paracetamol using Morgan fingerprints\n\nType /help for commands, /exit to quit.\n\nchatRDKit>\n```\n\nType a number to run that example. Picking **1** prompts you to enter a\nSMILES string, loads it as a molecule (`mol`, `mol2`, ... in the session),\nand then shows a **structured follow-up menu** of suggested next steps for\nthat molecule (compute properties, draw it, generate 3D, find rings, check\nLipinski's Rule of Five) -- type a number to run one, or just type your own\nrequest.\n\n### Three kinds of input\n\n- **A SMILES string** (e.g. `CCO`, `c1ccccc1`, `CC(=O)Oc1ccccc1C(=O)O`) is\n  loaded directly as a new molecule (`mol`, `mol2`, `mol3`, ...) -- no LLM\n  call needed.\n- **Python/RDKit code** (e.g. `Descriptors.MolWt(mol)`, `mol.GetNumAtoms()`,\n  `for atom in mol.GetAtoms(): print(atom.GetSymbol())`) is executed\n  directly in the session. A bare expression's value is printed, REPL-style.\n- **Plain English** (anything else) is sent to Ollama, translated into RDKit\n  code using the current session state as context, echoed as\n  `[chatRDKit] >>> ...`, and executed.\n\n### Multi-turn context\n\nchatRDKit keeps a rolling history of your requests and the code it\ngenerated, plus a live snapshot of every molecule currently in the session\n(so \"it\", \"the molecule\", or a compound name you already loaded resolve\ncorrectly). Clear the history with:\n\n```\nchatRDKit> /reset\n```\n\n### Rendering output", "# Unlocking the Power of Protein Engineering: FDA-Approved Enzyme Therapies Transform Rare Disease Treatment\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Enzyme/Brandname | Engineering Approach | Stabilization Strategy | Delivery method | Target/Disease | Year of release | Company | Key Reference |\n| --- | --- | --- | --- | --- | --- | --- | --- |\n| Pegunigalsidase alfa/Elfabrio | PEGylation, covalent crosslinking | PEG conjugation for extended half-life | Intravenous infusion | \u03b1-galactosidase A/Fabry disease | 2023 | Chiesi/Protalix | [Schiffmann R, et al., 2019](https://onlinelibrary.wiley.com/doi/10.1002/prp2.489) |\n| Avalglucosidase alfa/Nexviazyme | Glycoengineered, increased bis-M6P content | Enhanced M6P for improved cellular uptake | Intravenous infusion | Acid \u03b1-glucosidase/Pompe disease | 2021 | Sanofi Genzyme | [Sanofi, et al., 2021](https://www.sanofi.com/en/media-room/press-releases/2021) |\n| Velmanase alfa/Lamzede | Recombinant DNA, CHO cells | M6P for receptor-mediated endocytosis | Intravenous infusion | \u03b1-mannosidase/Alpha-mannosidosis | 2018 | Chiesi | [FDA, et al., 2018](https://www.fda.gov/news-events/press-announcements) |\n| Elapegademase/Revcovi | Recombinant DNA (E. coli), PEGylation | PEG with improved linker stability | Intramuscular injection | Adenosine deaminase/ADA-SCID | 2018 | Leadiant Biosciences | [Kohn DB, et al., 2020](https://pmc.ncbi.nlm.nih.gov/articles/PMC7049870/) |\n| Calaspargase pegol/Asparlas | Site-specific PEGylation, recombinant E. coli | Site-directed PEG conjugation | Intravenous | L-asparaginase/ALL | 2018 | Servier | [FDA, et al., 2018](https://www.fda.gov/drugs) |\n| Asfotase alfa/Strensiq | Recombinant DNA, fusion with Fc fragment | Fc fusion for extended circulation | Subcutaneous | Alkaline phosphatase/Hypophosphatasia | 2015 | Alexion (AstraZeneca) | [Ziogou E, et al., 2023](https://pubs.rsc.org/en/content/articlelanding/2023) |\n| Taliglucerase alfa/Elelyso | Recombinant DNA, plant cells (carrot) | Plant-derived glycosylation | Intravenous infusion | \u03b2-glucocerebrosidase/Gaucher disease type 1 | 2012 | Pfizer (Protalix) | [FDA, et al., 2012](https://www.fda.gov/drugs/resources-information-approved-drugs) |\n| Asparaginase Erwinia chrysanthemi/Erwinaze | Bacterial extraction (E. chrysanthemi) | None (native bacterial enzyme) | Intramuscular/intravenous | L-asparaginase/ALL | 2011 | Jazz Pharmaceuticals | [FDA, et al., 2011](https://www.fda.gov/drugs) |\n| Velaglucerase alfa/VPRIV | Recombinant DNA, human fibroblasts | Native human sequence with exposed mannose | Intravenous infusion | \u03b2-glucocerebrosidase/Gaucher disease type 1 | 2010 | Shire (Takeda) | [Brumshtein B, et al., 2009](https://pmc.ncbi.nlm.nih.gov/articles/PMC2745820/) |\n| Alglucosidase alfa/Myozyme | Recombinant DNA, CHO cells | M6P-targeting for lysosomal uptake | Intravenous infusion | Acid \u03b1-glucosidase/Pompe disease | 2006 | Genzyme (Sanofi) | [Kishnani PS, et al., 2009](https://pubmed.ncbi.nlm.nih.gov/19571336/) |\n| Idursulfase/Elaprase | Recombinant DNA, human cell line | M6P glycosylation for targeting | Intravenous infusion | Iduronate-2-sulfatase/MPS II | 2006 | Shire (Takeda) | [FDA, et al., 2006](https://www.fda.gov/drugs) |\n| Galsulfase/Naglazyme | Recombinant DNA, CHO cells | M6P for lysosomal delivery | Intravenous infusion | N-acetylgalactosamine-4-sulfatase/MPS VI | 2005 | BioMarin | [NIH, et al., 2005](https://www.ncbi.nlm.nih.gov/books/NBK1274/) |\n| Laronidase/Aldurazyme | Recombinant DNA, CHO cells | M6P-terminated oligosaccharides | Intravenous infusion | \u03b1-L-iduronidase/MPS I | 2003 | Genzyme/BioMarin | [Jameson E, et al., 2019](https://pmc.ncbi.nlm.nih.gov/articles/PMC4726864/) |\n| Agalsidase beta/Fabrazyme | Recombinant DNA, CHO cells | M6P-mediated targeting | Intravenous infusion | \u03b1-galactosidase A/Fabry disease | 2003 | Genzyme (Sanofi) | [Banikazemi M, et al., 2007](https://pubmed.ncbi.nlm.nih.gov/17186466/) |\n| Rasburicase/Elitek | Recombinant DNA (S. cerevisiae) | None (recombinant enzyme) | Intravenous | Urate oxidase/Tumor lysis syndrome | 2002 | Sanofi | [DrugBank, 2016](https://go.drugbank.com/drugs/DB00049) |\n| Agalsidase alfa/Replagal | Recombinant DNA, human cell line | M6P for receptor-mediated uptake | Intravenous infusion | \u03b1-galactosidase A/Fabry disease | 2001 | Shire (Takeda) | [EMA, et al., 2001](https://www.ema.europa.eu/en/medicines/human) |\n| Imiglucerase/Cerezyme | Recombinant DNA, CHO cells | Mannose-terminal glycosylation | Intravenous infusion | \u03b2-glucocerebrosidase/Gaucher disease type 1 | 1994 | Genzyme (Sanofi) | [Sanofi, 1994](https://products.sanofi.us/cerezyme/cerezyme.html) |\n| Pegaspargase/Oncaspar | PEGylation of E. coli asparaginase | PEG conjugation for extended half-life | Intramuscular/intravenous | L-asparaginase/ALL | 1994 | Servier (Shire) | [Hennigan JN, et al., 2022](https://pmc.ncbi.nlm.nih.gov/articles/PMC8714691/) |\n| Alglucerase/Ceredase | Placental extraction, enzymatic modification | Mannose exposure via deglycosylation | Intravenous infusion | \u03b2-glucocerebrosidase/Gaucher disease type 1 | 1991 | Genzyme | [Grabowski GA, et al., 2012](https://pmc.ncbi.nlm.nih.gov/articles/PMC3358951/) |\n| Pegademase bovine/Adagen | PEGylation of bovine enzyme | PEG conjugation to extend half-life | Intramuscular injection | Adenosine deaminase/ADA-SCID | 1990 | Leadiant Biosciences | [Kohn DB, et al., 2020](https://pmc.ncbi.nlm.nih.gov/articles/PMC7049870/) |\n| Asparaginase E. coli/Elspar | Bacterial extraction (E. coli) | None (native bacterial enzyme) | Intramuscular/intravenous | L-asparaginase/ALL | 1978 | Lundbeck | [Hennigan JN, et al., 2022](https://pmc.ncbi.nlm.nih.gov/articles/PMC8714691/) |", "# Revolutionary Bridge Between Prediction and Reality: How Elastic Networks Transform AlphaFold2\u2019s Structural Accuracy\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nOptimization methodology selection revealed surprising insights about computational efficiency versus accuracy trade-offs. [Powell optimization](https://academic.oup.com/bioinformaticsadvances/article/5/1/vbae181/7903278?login=false), a local method, consistently matched or exceeded the performance of more computationally expensive global optimization approaches including Dual Annealing and Differential Evolution. This finding suggests that the correlation landscape between predicted structures and experimental density maps contains relatively smooth basins of attraction, making sophisticated global search strategies unnecessary.\n\nThe similarity measure comparison between Pearson correlation coefficient and inner product demonstrated both computational and practical advantages for the latter approach. The [inner product method](https://academic.oup.com/bioinformaticsadvances/article/5/1/vbae181/7903278?login=false) proved invariant to zero-padding and eliminated cropping biases while maintaining equivalent or superior accuracy. This mathematical insight reduces computational overhead while improving reliability\u2014a rare combination in optimization problems.\n\nCase study results illuminate the method\u2019s transformative potential across diverse structural challenges. The lipid-preserved respiratory supercomplex exemplified global conformational corrections, with TM-scores improving from 0.52 to 0.69 through systematic repositioning of entire helical elements. The flagellar L-ring protein demonstrated successful refinement of challenging secondary structure arrangements, while the cation diffusion facilitator YiiP showcased domain-specific improvements. Most impressively, the\u00a0*Sulfolobus islandicus*\u00a0pilus structure improved from 0.77 to 0.85 TM-score despite containing sharp conformational features that approached the method\u2019s fundamental limitations.\n\nThe broader implications extend beyond individual structure refinement to address fundamental questions about computational structural biology workflows. Recent developments in\u00a0[hybrid flexible fitting approaches](https://pubs.acs.org/doi/10.1021/acs.jcim.3c02007)\u00a0that combine molecular dynamics with normal mode analysis suggest convergent evolution toward multi-scale modeling strategies. These methods acknowledge that different aspects of protein flexibility require different computational treatments.\n\nFuture applications promise to integrate these refinement strategies into high-throughput structural biology pipelines. The\u00a0[ModeHunter package](https://modehunter.biomachina.org/)\u00a0implementation provides accessible tools for routine application, while ongoing developments in\u00a0[ensemble-based fitting approaches](https://pmc.ncbi.nlm.nih.gov/articles/PMC11234365/)suggest paths toward capturing conformational heterogeneity directly from experimental data.\n\nThe methodology\u2019s success stems from respecting both the strengths and limitations of its component approaches. AlphaFold2 provides exceptional starting structures that capture global fold architecture and secondary structure elements with remarkable fidelity. Elastic network models contribute biophysically reasonable descriptions of collective motions without the computational overhead of all-atom molecular dynamics. Cryo-EM density maps supply experimental constraints that guide refinement toward physically meaningful conformations.\n\nThis integration exemplifies how artificial intelligence advances most effectively when combined with deep domain knowledge and experimental validation. Rather than replacing traditional biophysical approaches, machine learning methods achieve their greatest impact through synergistic combinations that leverage the unique strengths of each methodology. The result transforms both prediction accuracy and our understanding of protein conformational landscapes.\n\n| Key Concept | Description | Key References |\n| --- | --- | --- |\n| Elastic Network Models | Coarse-grained representations of protein dynamics using harmonic springs between atoms within cutoff distances, enabling computation of normal modes that describe collective motions | [Alshammari, M., et al., Bioinformatics Advances](https://academic.oup.com/bioinformaticsadvances/article/5/1/vbae181/7903278?login=false) |\n| TM-Score Validation | Template Modeling score ranging from 0-1 that measures structural similarity independent of sequence alignment, with values above 0.80 indicating high accuracy | [Alshammari, M., et al., Bioinformatics Advances](https://academic.oup.com/bioinformaticsadvances/article/5/1/vbae181/7903278?login=false) |\n| Normal Mode Analysis | Mathematical framework for analyzing protein dynamics through eigenvector decomposition of the Hessian matrix, providing basis functions for conformational changes | [Ribeiro, J.V., et al., Journal of Chemical Information and Modeling](https://pubs.acs.org/doi/10.1021/acs.jcim.3c02007) |\n| Flexible Fitting | Computational technique for deforming atomic models to better match experimental density maps while preserving biophysical constraints and structural integrity | [Ribeiro, J.V., et al., Journal of Chemical Information and Modeling](https://pmc.ncbi.nlm.nih.gov/articles/PMC11234365/) |\n| Cross-Correlation Optimization | Statistical measure quantifying similarity between simulated density maps from deformed structures and experimental cryo-EM maps, used as objective function for fitting | [Alshammari, M., et al., Bioinformatics Advances](https://academic.oup.com/bioinformaticsadvances/article/5/1/vbae181/7903278?login=false) |\n| ModeHunter Package | Python-based software suite providing tools for elastic network model generation, normal mode analysis, and flexible fitting applications in structural biology | [Wriggers, W., et al., Journal of Physical Chemistry B](https://pubs.acs.org/doi/abs/10.1021/acs.jpcb.4c05077) |", "# Crystal Structure of Histidine Phosphotransferase ShpA (NP_419930.1) from Caulobacter crescentus at 1.52 \u00c5 Resolution\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [2OOC](https://www.rcsb.org/structure/2OOC)\n* **Release Date:** 2007-02-06\n* **Method:** X-RAY DIFFRACTION 1.52 \u00c5\n* **Organisms:** Caulobacter vibrioides CB15\n* **Macromolecules:** Histidine phosphotransferase (protein)\n* **Unique Ligands:** GOL, PG4\n* **Publication:** [Xu, Q., et al. (2009). Crystal Structure of Histidine Phosphotransferase ShpA from Caulobacter crescentus. Journal of Molecular Biology, 390, 686-698.](https://www.sciencedirect.com/science/article/abs/pii/S0022283609005890)\n\n**Description:** This high-resolution (1.52 \u00c5) structure of the histidine phosphotransferase ShpA from Caulobacter crescentus provides atomic-level insight into two-component signal transduction systems that regulate the asymmetric cell division cycle in this model organism. The structure illuminates the phosphorelay mechanism critical for bacterial cell differentiation and developmental biology.", "# ChatMol \u2013 Natural-language control of PyMOL, describe the scene, get the picture\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-14 on marcdeller.com (categories: vibe-code).*\n\n> **Natural-language control of PyMOL. Describe the scene, get the structure.**\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/chatmol](https://github.com/bellcheddar/chatmol) |\n\n---\n\nA natural-language chat frontend for **PyMOL (open-source, 3.1.0)**, powered\nby a local **Ollama** model (default: `qwen3.6:27b`).\n\nWhy it matters: PyMOL is the lingua franca of molecular visualisation, but producing a polished figure often means stacking a dozen exact commands for selections, representations, colouring, and views. ChatMol layers a local LLM directly onto PyMOL\u2019s command line so plain-English requests (\u201ccolour by chain and show the ligand as sticks\u201d) become correct PyMOL syntax instantly, while any valid PyMOL you type still runs untouched. It is useful for structural biologists, crystallographers, and cryo-EM scientists who want to move faster from raw coordinates to publication-quality scenes, lower the barrier for students learning PyMOL, and keep every byte of structural data local.\n\nType plain English into PyMOL\u2019s normal command line:\n\n```\nPyMOL> fetch 1abc and show it as a cartoon coloured by chain, then show the ligand as sticks\n[ChatMol] > fetch 1abc, async=0\n[ChatMol] > hide everything\n[ChatMol] > show cartoon, polymer\n[ChatMol] > util.cbc polymer\n[ChatMol] > show sticks, organic\n```\n\nAnything that already looks like valid PyMOL syntax is run unchanged, so\nyour existing scripts and muscle memory keep working. Anything that looks\nlike English is sent to a local LLM, translated into PyMOL commands, echoed\nto the console, and executed. Everything runs locally \u2014 no data leaves\nyour machine.\n\n---\n\n## 1. Prerequisites\n\nYou need two things installed before the plugin itself: **Ollama** (with a\nmodel pulled) and **open-source PyMOL 3.1.0**.\n\n### 1.1 Ollama + model\n\n1. Install Ollama:\n2. **macOS**: download the app from <https://ollama.com/download>, or\n   `brew install ollama`\n3. **Linux**: `curl -fsSL https://ollama.com/install.sh | sh`\n4. **Windows**: download the installer from <https://ollama.com/download>\n5. Make sure the server is running. The desktop app starts it\n   automatically; otherwise start it manually and leave it running:\n\n`sh\nollama serve`\n\n1. Pull a Qwen3.6 model:\n\n`sh\nollama pull qwen3.6:27b`\n\nThis is the default ChatMol uses, and gives the best translation\nquality. It\u2019s a ~17 GB download and needs roughly 20 GB+ of free\nRAM/VRAM to run comfortably. If your machine can\u2019t handle that:\n\n`sh\n# Much smaller/faster, good for trying ChatMol out, lower translation quality:\nollama pull qwen3:0.6b`\n\nThen point ChatMol at whichever tag you pulled (see\n[Configuration](#4-configuration)) \u2014 just make sure the `model` setting\nmatches a tag from `ollama list` exactly.\n\n1. Sanity check the server is reachable and the model responds:\n\n`sh\ncurl http://localhost:11434/api/chat -d '{\n\"model\": \"qwen3.6:27b\",\n\"messages\": [{\"role\": \"user\", \"content\": \"say hi\"}],\n\"stream\": false\n}'`\n\nYou should get back a JSON response with a `message.content` field. If\nthis fails, fix it before installing the plugin \u2014 see\n[Troubleshooting](#7-troubleshooting).\n\nIf you run Ollama on a different host/port, configure that too (see\n[Configuration](#4-configuration)).\n\n### 1.2 PyMOL (open-source, 3.1.0)\n\nPick whichever matches how you normally manage Python environments:\n\n* **conda / mamba** (any OS):\n\n`sh\nconda create -n pymol -c conda-forge -c schrodinger pymol-open-source=3.1.0\nconda activate pymol`\n\n* **Homebrew (macOS)**:\n\n`sh\nbrew install pymol`\n\n* **Linux package managers** also commonly provide `pymol` (e.g.\n  `apt install pymol`), though the bundled version may differ from 3.1.0.\n\nLaunch it once (`pymol`) to confirm it starts normally and note where its\nPython lives \u2014 that\u2019s the interpreter ChatMol will run under. ChatMol only\nuses the Python standard library, so no extra `pip install` is required.\n\n---\n\n## 2. Installing the plugin\n\nChatMol is a plain Python package (the `chatmol/` directory). There are two\nways to load it; pick one.\n\n### Option A \u2014 `~/.pymolrc.py` (recommended, most reliable)\n\nThis runs ChatMol from the source directory directly \u2014 easiest to update\n(just edit the files and restart PyMOL).\n\nCreate or edit `~/.pymolrc.py` (in your home directory) and add:\n\n```\nimport sys\nsys.path.insert(0, \"/Users/dellboy/Documents/Vibe Coding/chatmol\")\nimport chatmol\n```\n\n(Adjust the path if you move the `chatmol` project folder.)\n\nRestart PyMOL. You should see the `[ChatMol] AI chat interface loaded`\nbanner in the console \u2014 that confirms it\u2019s active.\n\n### Option B \u2014 Plugin Manager (`chatmol.zip`)\n\nA ready-to-install zip is included at the project root:\n`chatmol.zip` (contains `chatmol/__init__.py` etc. at its root, as the\nPlugin Manager expects).\n\n1. In PyMOL: **Plugin > Plugin Manager > Install New Plugin > Choose\n   file\u2026** and select `chatmol.zip`.\n2. Restart PyMOL.\n\nThis also adds a **Plugin > ChatMol Settings** menu item with a small GUI\nfor configuration.\n\nIf you edit the source after installing this way, rebuild the zip from the\nproject root:\n\n```\ncd \"/Users/dellboy/Documents/Vibe Coding/chatmol\"\nrm -f chatmol.zip\nzip -r chatmol.zip chatmol -x '*.pyc' -x '*__pycache__*' -x '*.DS_Store'\n```\n\nand reinstall it.\n\n### Verifying the install\n\nEither way, after restarting PyMOL you should see:\n\n```\n[ChatMol] AI chat interface loaded.\n[ChatMol]   - Type plain English into the PyMOL command line, e.g.\n[ChatMol]       fetch 1abc and show it as a cartoon coloured by chain\n[ChatMol]   - Or use the explicit command:  ai <request>\n[ChatMol]   - Settings:  chatmol_config [key [value]]\n[ChatMol]   - Forget conversation context:  chatmol_reset\n```\n\nThen try it:\n\n```\nPyMOL> fetch 1ubq, async=0\nPyMOL> colour the protein by chain and show it as cartoon\n```\n\n---\n\n## 3. Usage", "# Sci-Hub at the Crossroads: From Shadow Archive to AI Research Tool\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-04-24 on marcdeller.com (categories: ai, blog).*\n\n* [**Latest Sci-Hub Proxy**](https://sci-hub.usualwant.com)\n\n* **Coverage exceeds institutional access:** Sci-Hub held 97.3% of Elsevier\u2019s catalog and 85.1% of all toll-access journals as of 2017.\n* **Legal pressure is mounting globally:** Court-ordered bans in India and 11 other countries signal publishers are winning individual battles without fixing the structural problem.\n* **The AI pivot is deliberate:** Sci-Bot\u2019s grounded, citation-backed responses reframe Sci-Hub as research infrastructure rather than a piracy tool.\n* **The root cause persists:** Publisher profit margins above 30% and APCs reaching $10,600 ensure continued demand for alternatives to paywalled access.\n\n1. **Sci-Hub provides access to nearly all scholarly literature:** [Himmelstein et al., eLife, 2018](https://elifesciences.org/articles/32822)\n2. **Half of papers searched for online are free to read:** [Singh Chawla, Nature, 2017](https://www.nature.com/articles/nature.2017.22418)\n3. **The oligopoly\u2019s shift to open access: How the big five academic publishers profit from article processing charges:** [Butler et al., Quantitative Science Studies, 2023](https://direct.mit.edu/qss/article/4/4/778/118070/)\n4. **Researchers in India worry about access amid Sci-Hub ban:** [C&EN, 2025](https://cen.acs.org/policy/publishing/Researchers-India-worry-access-amid/103/web/2025/09)\n5. **Sci Hub has created a new AI chatbot. Is it any good?:** [C&EN, 2026](https://cen.acs.org/policy/publishing/Sci-Hub-created-new-AI/104/web/2026/04)\n\n**The platform that rewired how science is accessed is moving beyond piracy into the age of AI, and the questions it raises about knowledge equity have never been sharper**\n\nWhen Alexandra Elbakyan launched Sci-Hub in Kazakhstan in 2011, the founding motivation was clear: she could not afford the papers she needed for her own research. That act of frustration has since grown into a repository of over 88 million scientific articles, accessible to anyone with an internet connection and a DOI. The platform operates by circumventing institutional paywalls using harvested login credentials, which are illegal in most jurisdictions\u00a0yet used by researchers at Harvard and in rural India with equal enthusiasm. The underlying economics make the behavior easy to understand. The five dominant commercial publishers generate [profit margins exceeding 30%](https://direct.mit.edu/qss/article/4/4/778/118070/), extracting that return from publicly funded science while paying neither authors nor peer reviewers for the work they provide.\n\nThe scale of the problem became measurable in 2018 when [Himmelstein et al., writing in *eLife*](https://elifesciences.org/articles/32822), mapped coverage across 81.6 million Crossref-registered articles. Chemistry content accounted for 92.8%; the Elsevier catalog, a remarkable 97.3%. Critically, Sci-Hub\u2019s access exceeded what was available to researchers at the University of Pennsylvania, a major research institution. As the authors concluded:\n\n> ***\u201cFor the first time, nearly all scholarly literature is available gratis to anyone with an Internet connection, suggesting the toll access business model may become unsustainable.\u201d***\n\nPublishers responded with litigation. Elsevier won default judgments in US federal courts in 2015 and 2017, stripping the platform of domain names. A case filed in 2020 by Elsevier, Wiley, and the American Chemical Society in the Delhi High Court ground on for five years before producing a [blanket ban in August 2025](https://cen.acs.org/policy/publishing/Researchers-India-worry-access-amid/103/web/2025/09), directing Indian ISPs to block access within 24 hours. India had accounted for 8.7% of all Sci-Hub downloads in 2021. The blow fell hardest on smaller institutions, hospitals, and underfunded universities with no institutional journal access. [An analysis in](https://www.nature.com/articles/nature.2017.22418)*Nature*\u00a0already documented that around half of the papers searched for online were freely accessible via legitimate channels, but that coverage is unevenly distributed, and researchers in lower-income countries routinely fall outside it.\n\nElbakyan did not retreat. In late 2024, Sci-Hub issued a cryptocurrency token\u00a0to reward paper sharing through its companion platform, Sci-Net. In early 2026, [Sci-Hub launched Sci-Bot](https://cen.acs.org/policy/publishing/Sci-Hub-created-new-AI/104/web/2026/04), an AI chatbot that answers research questions using the Sci-Hub corpus, with grounded citations and without hallucinated references. The alpha version handles single-question queries and shows gaps in the literature from 2023 onward, where publishers have progressively hardened access controls. The intent is clear: transform an illegal archive into a functioning research intelligence infrastructure. Whether AI rebranding changes the legal and ethical calculus remains genuinely open, but the inequity [Sci-Hub was built to address](https://direct.mit.edu/qss/article/4/4/778/118070/) has not been resolved. As long as publicly funded science sits behind a $3,400 average article processing charge or a $10,600 paywall, there will be researchers reaching for the door Elbakyan built.", "# BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nPlain labelled text \u2014 no markdown, no YAML, no brackets, no quoting. One rule: blocks are `Label: value` lines with a blank line between them; comments start with `#`. Field names are plain English (`Output folder`, `Predict affinity`, `Pocket contact`) rather than Boltz-internal snake_case. Don\u2019t want to hand-write it at all? Run `python3 BoltzMaker.py new` and answer plain questions instead \u2014 see **Commands** below.\n\nThe format has two layers: a **family x partners x ligand cross-product** (the ergonomic layer: write each protein/ligand once, get every combination as a separate target), and standalone **constraint sentences** for the two/three-ended relationships (covalent bonds, distance constraints) that don\u2019t fit inside one block \u2014 each names the protein it belongs to and can be written anywhere in the file.\n\n```\nSettings:\nOutput folder: ./boltz_yamls    # where generated per-target YAMLs are written\nPredict affinity: no            # off by default -- it's a heavier prediction pass\n\nProtein: RECP1                  # short name, MAX 5 CHARACTERS (Boltz stores chain names\n                                # in a fixed 5-char field internally and silently\n                                # truncates longer ones, which then crashes later with a\n                                # confusing error -- `preflight` catches this for you).\n                                # Also names the output file: {protein}_{ligand}.yaml\nSequence: MDILC...              # required\nPartners: CHNX, CHNY            # optional: co-folded chains, defined as their own\n                                # Partner: blocks below\n# Ligands: LIG1, LIG3           # optional: restrict this protein to a ligand subset\n                                # (default: crossed with every ligand below)\n# Ligands: none                 # optional: ligand-free (apo) target -- no ligand entity,\n                                # no pocket constraint, no affinity property, whatever\n                                # `Predict affinity:` says. Stem is just the protein name\n                                # (e.g. `RECP1.yaml`, not `RECP1_LIG1.yaml`), runs through\n                                # the same generate/preflight/run/analyze pipeline and the\n                                # same staged `boltz predict` batch as every other target.\n# Modifications: SEP:5          # optional: CCD:position tokens for modified residues\n                                # (e.g. phosphoserine)\n# Cyclic: yes                   # optional: cyclic polymer (e.g. a cyclic peptide)\n# MSA: empty                    # optional: path to a precomputed MSA, or \"empty\" for\n                                # single-sequence mode (skip MSA generation)\n# Templates: reference_structure.cif\n                                # optional: structural template file(s), applied to all\n                                # protein chains (no per-chain mapping -- hand-edit the\n                                # generated YAML for that rarer case)\n# Apo structure: reference/apo.pdb\n                                # optional: a reference apo/unbound structure, used only\n                                # by `compare-sse` (see below), never by generate/run.\n                                # No genuinely apo experimental structure? Predict one:\n                                # give another `Protein:` block the same `Sequence:` and\n                                # `Ligands: none` (see above), run the campaign once, then\n                                # point `Apo structure:` at its output in `boltz_cif/`.\n# Apo chain: A                 # optional: explicit chain id in the apo structure above\n                                # (omit to auto-detect via sequence identity)\n# Family type: gpcr            # optional: gpcr / kinase / auto (default) -- selects\n                                # `compare-sse`'s motif annotator\n# Group: RECP1                  # optional: shared display/report name for multiple\n                                # `Protein:` blocks that are the same underlying receptor\n                                # (e.g. with/without a partner, or a predicted apo\n                                # variant) -- defaults to this block's own name if unset\n\nPartner: CHNX\nSequence: MTLES...\n# Type: dna                     # optional: protein (default) / dna / rna\n# Copies: X1, X2                # optional: homo-oligomer chain-id override -- this one partner\n                                # sequence becomes multiple chains\n\nLigand: LIG1\nSMILES: FC(F)CNC(...)=O         # exactly one of SMILES/CCD is required\n# Role: agonist                 # optional: agonist / antagonist -- reporting only (dashboard\n                                # chart shapes), never affects generate/run\n\nLigand: LIG2\nCCD: GOL                        # a Chemical Component Dictionary code (e.g. common crystallization\n                                # additives/ions) instead of a SMILES\n\nCovalent bond: RECP1 residue 44 atom SG to LIG1 residue 1 atom C3\nPocket contact: RECP1 residue 148\nDistance constraint: RECP1 residue 10 to RECP1 residue 80 within 8.0 Angstrom\n```\n\nEvery protein is crossed with every ligand (unless a protein sets `Ligands:` to scope itself to a subset, or `Ligands: none` for a single ligand-free/apo target), producing one `{protein}_{ligand}.yaml` per pair. See `example.md` for the full copy-paste template and `examples/` for complete working campaigns.\n\n## \ud83d\ude80 Commands", "# Publications of Marc C. Deller\n\n## Additional items from Google Scholar (not in ORCID)\n\n- **Structural studies of cytokines and cytokine receptors.**. Authors: Deller, Marc Christian.\n- **CLONING, EXPRESSION, CHARACTERIZATION, PURIFICATION AND PRELIMINARY CRYSTALLOGRAPHIC STUDIES OF THE HAEMOPOIETIN CYTOKINE ONCOSTATIN-M**. Authors: Deller, MC; Hudson, K; Bravo, J; Heath, JK; Jones, EY.\n- **Crystal Structure of a Cleaved, Stabilized HIV-1 Env Trimer in Complex with a Glycan-Dependent Broadly Neutralizing Antibody at 4.7 angstrom**. Authors: Julien, J; Cupo, A; Sok, D; Stanfield, RL; Lyumkis, D; Deller, MC; Burton, DR; Sanders, RW; Moore, JP; Ward, AB et al.\n- **Crystallisation of Proteins and Macromolecular Complexes: Past, Present and Future**. Authors: Deller, Marc C; Rupp, Bernhard.\n- **Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands**. Authors: Pozharski, Edwin; Deller, Marc C; Rupp, Bernhard.\n- **Structure\u2013function analysis of the extended conformation of a polyketide synthase module**. Authors: Li, Xiuyuan; Sevillano, Natalia; La Greca, Florencia; Deis, Lindsay; Liu, Yu-Chen; Deller, Marc C; Mathews, Irimpan I; Matsui, Tsutomu; Cane, David E; Craik, Charles S et al.\n- **Structure of the gamma D-glutamyl-L-diamino acid endopeptidase YkfC from Bacillus cereus in complex with L-Ala--D-Glu: Insights into substrate recognition by NlpC/P60 cysteine peptidases**. Authors: Xu, Q; Abdubek, P; Astakhova, T; Axelrod, HL; Bakolitsa, C; Cai, X; Carlton, D; Chen, C; Chiu, H-J; Chiu, M et al.\n- **Structure of the c-D-glutamyl-L-diamino acid endopeptidase YkfC from Bacillus cereus in complex with L-Ala-c-D-Glu: insights into substrate recognition by NlpC/P60 cysteine peptidases**. Authors: Astakhova, Tamara; Herbert, L; Cai, Xiaohui; Carlton, Dennis; Chen, Connie; Chiu, Hsiu-Ju; Chiu, Michelle; Clayton, Thomas; Das, Debanu; Deller, Marc C et al.\n- **Discovery of potent and selective inhibitors of wild-type and gatekeeper mutant fibroblast growth factor receptor (FGFR) 2/3**. Authors: Shvartsbart, Artem; Roach, Jeremy J; Witten, Michael R; Koblish, Holly; Harris, Jennifer J; Covington, Maryanne; Hess, Rodrigo; Lin, Luping; Frascella, Michelle; Truong, Lisa et al.\n- **INCB161734: A novel, potent, and orally bioavailable KRAS G12D selective inhibitor demonstrates antitumor activity in KRAS G12D mutant tumors**. Authors: Farren, Matthew R; Roman, Valerie; Gallion, Alexandra; Allali-Hassani, Abdellah; Sokolsky, Alexander; Kong, Weixi; Smith, Amanda; Wang, Hui; Correa, Gina; Deller, Marc et al.\n- **Discovery of (4-pyrazolyl)-2-aminopyrimidines as potent and selective Inhibitors of cyclin-dependent kinase 2**. Authors: Hummel, Joshua R; Xiao, Kai-Jiong; Yang, Jeffrey C; Epling, Leslie B; Mukai, Ken; Ye, Qinda; Xu, Meizhong; Qian, Dingquan; Huo, Lu; Weber, Michael et al.\n- **Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor**. Authors: Ye, Qinda; Shvartsbart, Artem; Li, Zhenwu; Gan, Pei; Policarpo, Rocco L; Qi, Chao; Roach, Jeremy J; Zhu, Wenyu; McCammant, Matthew S; Hu, Bin et al.\n- **Anti-mutant calreticulin (calr) antibodies and uses thereof**. Authors: Mayes, Patrick; Nastri, Horacio G; Buonpane, Rebecca A; Reis, Edimara S; Stewart, Shaun M; Zhou, Jing; Deller, Marc C; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong et al.\n- **Bicyclic compounds as inhibitors of wrn 1**. Authors: Ai, Yanran; Barbosa, Joe; Deller, Marc; Dherange, Balu D; Ghosh, Ananda; Glenn, Joe; McQuirter, Leslie; Wang, Xiaozhao; Yue, Eddy; Zhang, Guofeng et al.\n\n## Patents listed on Google Scholar\n\n- **Anti-mutant calreticulin (CALR) antibodies and uses thereof** \u2014 US 2023/0272055 A1 (US Patent App. 18/062,924, published 2023). Inventors include Patrick Mayes, Horacio G. Nastri, Rebecca A. Buonpane, Edimara S. Reis, Shaun M. Stewart, Jing Zhou, Marc C. Deller, Hsiang-Ting Lei, Leslie McQuirter et al. (Incyte). Marc led the structural biology sprint that solved the peptide-bound mutant-CALR\u2013antibody complex supporting this first-in-class antibody patent.\n- **Bicyclic compounds as inhibitors of WRN** \u2014 US 2025/0215006 A1 (US Patent App. 18/821,175, published 2025). Inventors include Yanran Ai, Joe Barbosa, Marc Deller, Balu D. Dherange, Ananda Ghosh, Joe Glenn, Leslie McQuirter, Xiaozhao Wang, Eddy Yue, Guofeng Zhang et al. (Incyte). Werner syndrome helicase (WRN) inhibitors for microsatellite-instability-high cancers.\n- **Therapeutic Enzymatic Compositions and Delivery Methods for the Degradation of Microplastics and Nanoplastics in Human Physiology** \u2014 US provisional patent application (2025), Elora Therapeutics; inventor Marc C. Deller.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nSDS-PAGE analysis showed that no expression of OSM_\u0394185-3C-GST was observed (data not shown). After induction with IPTG no band for OSM_\u0394185 (predicted Mr ~50 kDa) was observed and after affinity purification no protein was detected using the Coomassie Protein Plus\u2122 Assay Kit (Pierce). Soluble and insoluble fractions were analysed and no bands of the correct size were observed.\n\nFigure 3-5 below shows a 12 % SDS-polyacrylamide gel of the expression and affinity purification of (a) OSM_\u0394188-3C-GST and (b) OSM-196. In both cases, lane 1 shows the total cellular protein before induction with IPTG and lane 2 shows the total cellular protein after ~20 hours induction with IPTG. In both gels a relative increase in the density of a band at ~50 kDa can be observed on induction, so suggesting production of fusion protein. Lane 3 and 4 show duplicate samples of cleared cell lysate, which contains a major band at ~50 kDa, again suggesting production of soluble fusion protein. Lane 5 shows samples of glutathione sepharose resin after passage of the cleared lysate. This shows that two major species are bound to the affinity column at ~50 kDa and ~25 kDa, presumably corresponding to fusion protein and GST respectively. Lane 6 contains a wash fraction and displays no major bands. Lane 7 shows glutathione sepharose resin after overnight cleavage with 3C protease and only one major band at ~32 kDa can be seen, which presumably corresponds to sequestered GST purification tag. Lanes 8 and 9 show the first eluted fractions from the affinity column, after cleavage with protease 3C. One major band corresponding to OSM-196 or OSM_\u0394188 can be clearly seen at ~25 kDa, although contaminating bands at ~60 kDa and ~90 kDa can also be seen. The latter bands are thought to correspond to heat shock protein 60 and 90 (HSP60 and HSP90), which are known to co-purify with the pGEX system. Lane M shows the seven components of the high range molecular weight markers (Amersham).\n\n*Figure 3-5.* Figure showing 12 % SDS-PAGE gel of the expression and affinity purification of (a) OSM_\u0394188 and (b) OSM-196. Lanes 1 and 2 show total cellular protein pre- and post-induction. Lanes 3 and 4 show the cleared cell lysate. Lane 5 shows material bound to the affinity column. Lane 6 shows a column wash and lane 7 shows material remaining on the column after cleavage with 3C protease. Lanes 8 and 9 show cleaved fractions eluted from the affinity column. Lane M shows the high range molecular weight markers running at 220, 97, 66, 46, 30, 21 and 14 kDa (Amersham).\n\n### 3.3.4 Western Blot Analysis\n\nFigure 3-6 below shows western blot analysis of affinity purified OSM_\u0394188 separated on a 12\u201315 % gradient SDS polyacrylamide gel. Lane 1 contains ~8 \u03bcg of OSM_\u0394188 and shows approximately 7 main bands. Around 30 % of the density runs at ~25 kDa. Lane 2 shows broad range rainbow molecular weight markers (Bio-Rad). Lane 3 shows a control sample containing LIFR, which shows no bands and no cross reactivity to the anti-OSM antibodies.\n\n*Figure 3-6.* Figure showing western blot of affinity purified OSM_\u0394188. Lane 1 contains OSM_\u0394188. Lane 2 contains broad range molecular weight rainbow markers running at 210, 135, 80, 42, 31 and 18 kDa (Bio-Rad). Lane 3 contains a control sample of LIFR.\n\n### 3.3.5 Receptor Competition and Titration Assay\n\nFigure 3-7 below shows graphs of (a) receptor titration and (b) receptor competition assays. Graph (a) demonstrates that the concentration range of ligand used leads to saturation of the gp130 receptor under the conditions employed. This also shows that specific binding is occurring due to the sigmoidal increase in absorbance with increasing concentration of human biotinylated OSM-196. Binding is not observed to human OSMR and murine LIFR, as both show background responses similar to the control FGFR2 samples. Binding of human biotinylated OSM-196 shows an effective concentration 50 % (EC50) value of 348.9 ng/ml. From graph (b) it can be seen that increasing the concentration of both OSM-196 and OSM_\u0394188 causes a sigmoidal decrease in the absorbance reading due to competition with biotinylated OSM-196. The OSM-196 sample agrees closely with the OSM_\u0394188 sample and no fall in absorbance is demonstrated by the control murine LIF sample. Human OSM-196 displays an EC50 value of 209.2 ng/ml and human OSM_\u0394188 displays an EC50 value of 182.0 ng/ml.\n\n*Figure 3-7.* Figure showing graph of (a) receptor titration and (b) receptor competition assays. Panel (a) shows binding of human biotinylated OSM-196 to human gp130 (EC50 348.9 ng/ml), human OSMR, murine LIFR and as control human FGFR2. Panel (b) shows competitive binding to human gp130 in the presence of human biotinylated OSM-196. Human OSM-196 (EC50 209.7 ng/ml), human OSM_\u0394188 (EC50 182.0 ng/ml) and as control murine LIF are shown.\n\n### 3.3.6 Ba/F3 Cell Survival Assay\n\nFigure 3-8 below shows a graph of a Ba/F3 cell survival assay for various cytokines. Analysis of the OSM_\u0394188 and OSM-196 curves shows that a sigmoidal dose response is observed. In both cases increasing protein concentration results in an increase in absorbance reading. Murine LIF shows no dose response in the concentration range tested. Human OSM-196 shows an EC50 value of 0.39 ng/ml and human OSM_\u0394188 displays an EC50 value of 0.26 ng/ml.\n\n*Figure 3-8.* Graph showing Ba/F3 cell survival assay curves for various cytokines. Human OSM-196 (EC50 0.39 ng/ml), human OSM_\u0394188 (EC50 0.26 ng/ml) and murine LIF were tested at concentrations ranging from 500\u20130.001 ng/ml.\n\n### 3.3.7 Ion-exchange and Gel Filtration Chromatography\n\nIon-exchange chromatography proved unsuccessful using both Mono-Q and Mono-S columns in the pH range 4\u201310. No specific binding of protein was observed and only aggregates were recovered from the column at high salt concentrations.", "# What If Protein Structures Could Be Designed Like LEGO? Modular Approaches in 2025\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: blog).*\n\n* New AI models are enabling a modular, \u201cLEGO-like\u201d approach to protein design, where proteins are assembled from defined components.\n* Tools like ProtComposer and Proteina allow designers to specify the high-level architecture and properties of protein parts, with the AI handling the low-level atomic details.\n* This vision for modular protein engineering could revolutionize synthetic biology, enabling the rapid design of custom proteins for countless applications.\n\n1. <https://openreview.net/pdf/3cd76a5f12a43a2296132178c1a97bad7e141588.pdf>\n2. <https://github.com/NVIDIA-Digital-Bio/proteina>\n3. <https://arxiv.org/abs/2503.05025>\n\nThe ultimate dream of synthetic biology is to make the engineering of biology as predictable and standardized as the engineering of electronics or machines. For proteins, this means moving toward a future where we can design and build them from a set of well-characterized, interchangeable parts, much like snapping together LEGO bricks. For years, this has been a distant vision, as the rules for connecting protein domains are incredibly complex.\n\nThe latest generation of compositional and generative AI is finally bringing this modular dream within reach. Models like **ProtComposer**are pioneering this approach by allowing designers to sketch out a protein\u2019s architecture using a simple layout of 3D shapes, or \u201cellipsoids\u201d. Each shape can be thought of as a specialized LEGO brick. The designer arranges the bricks, and the AI acts as the master builder, figuring out the complex atomic details needed to connect them into a stable, functional whole. Similarly, flow-based models like **Proteina** offer hierarchical control, letting designers specify the type of fold or secondary structure, akin to choosing a specific category of LEGO piece to work with.\n\nThis is a profound shift in how we approach protein engineering. It opens the door to creating vast libraries of modular protein parts that can be combined in near-infinite ways to create custom enzymes, biosensors, or therapeutics. Imagine ordering a protein from a catalogue of functional domains, specifying their spatial arrangement, and having an AI deliver the final sequence. We are at the very beginning of this journey, but the tools are now in place. What custom proteins would you build for medicine, industry, or the environment if you had a molecular LEGO set at your fingertips?", "# AlphaFraud \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n* \u2705 **Statistics page** \u2014 an exhaustive live `/stats` page grouping everything the app tracks into themed, equal-height panels (consistent with the rest of the site): catch & quality KPIs, coverage & schedule (date ranges, release weeks, last/next update), runs & compute, server metrics (CPU, load, memory, uptime), storage & database (file sizes, per-table row counts, disk), data sources & APIs (linked), and traffic \u2014 with a \u201cFull statistics \u2192\u201d link from the header panel\n* \u2705 **Home landing page** \u2014 a concise navigation page at `/` outlining the project and each tab as a colour-coded card with a \u201cwhat you\u2019ll learn\u201d line and headline KPIs; the cumulative dashboard moved to `/all` (\u201cAll structures\u201d). Nav tabs now carry their panel accent colour to tie the site together\n* \u2705 **Calculate tab** \u2014 type any human PDB ID and AlphaFraud runs the full pipeline on demand (resolve the entity, fetch its AlphaFold model, superpose, score, render the ribbon) and returns the same panel as the auto examples, with **no LLM**. A strict autocomplete index (exists + human + has an AlphaFold model), seeded from the corpus and auto-refreshed each week, guarantees only qualifying entries; rejections are explained (not human / antibody-fusion / no AlphaFold model); good predictions render the green \u201cgets it right\u201d panel and confidently-wrong ones the failure panel; the heavy compute runs in an isolated subprocess and on-demand results stay out of every aggregate\n* \u2705 **Examples tab** \u2014 a curated storytelling layer that shows both sides of AlphaFold\u2019s record. It opens with a **positive control** (human Artemis, a novel post-cutoff nuclease predicted blind to 0.45 \u00c5 C\u03b1-RMSD and scored FRAUD 0.02) proving the pipeline certifies good predictions as well as bad; an auto-generated, LLM-free **\u201cexample of the week\u201d** plus a **complete weekly-release archive** (catches and quiet weeks alike); hand-authored deep-dive panels across the failure spectrum, each with live metrics and an interactive 3D viewer; and clickable red/green example markers overlaid on the home fraud-quadrant scatter\n* \u2705 **Page and figure caching** \u2014 the cumulative dashboard\u2019s heavy whole-archive figures (scatter / dumbbell / histograms over ~80k rows) are precomputed once and cached in an `analysis_snapshots` row with a short TTL, and every static asset (vendored Plotly / 3Dmol, ribbons, CSS/JS) is served with long immutable cache headers, so repeat loads are near-instant\n* \u2705 **Weekly-release archive at full granularity** \u2014 the Archive tab now lists every PDB release week since the 2018 cutoff (derived from each structure\u2019s release date), each linking to that week\u2019s structures and confidently-wrong count, instead of only the recent ongoing-watch runs\n* \u2705 **Fix slow page loads after the full backfill** \u2014 once the archive backfill filled the `entities` table with ~470 KB `heatmaps_json` per compared row (\u2248 6.5 GB), every list/sort query (home, leaderboard, week pages) dragged all that JSON through memory and took 20\u201360 s (the leaderboard 504\u2019d at the 60 s gunicorn timeout). Fixed by moving the big per-residue and heatmap payloads into an `entity_blobs` sidecar table read only by the per-structure entry page, so the hot table stays small and list queries return in well under a second\n* \u2705 **Structural imagery throughout** \u2014 every worst offender is rendered as a deviation-coloured C\u03b1 ribbon (experiment coloured by distance from the AlphaFold model, on an absolute-\u00c5ngstr\u00f6m scale), shown on the leaderboard, entry page and weekly highlights; the entry page adds an interactive 3Dmol.js viewer with a proper secondary-structure cartoon and an optional translucent AlphaFold \u201cghost\u201d overlay in the shared superposition frame; a \u201cdivergence\u201d ribbon banner sits in the header. All server-rendered as vendored, offline SVG/PDB (no CDN)\n* \u2705 **Percentages on the KPI tiles** \u2014 each headline count also shows its share of the batch (e.g. confidently wrong: 1,941 = 2.3%), on the dashboard, week and leaderboard views\n* \u2705 **Worst-offenders structural deep dive** \u2014 the **Analysis** tab characterises the confidently-wrong set by CATH / SCOP2 fold and family class (Wilson-CI + Fisher-exact enrichment), clusters it by sequence similarity, flags shared themes (amyloid / assembly / disordered / coiled-coil / engineered), maps failure modes by PCA, detects conformational heterogeneity, gives per-superfamily blind-spot scorecards, and links every offender to its RCSB entry and DOI-verified manuscript; refreshed hourly and after each weekly run\n* \u2705 **Hover-to-preview structures** \u2014 hovering a point on the \u201cfraud quadrant\u201d scatter pops a floating deviation-coloured ribbon thumbnail of that structure; the entry page carries the full interactive 3D viewer\n* \u2705 **Header \u201cStats\u201d panel** \u2014 a live panel in the top-right of the header tracking app health: worker memory, SQLite DB size, unique visitors, structures analysed, confidently-wrong count and archive %; refreshes every minute from `/api/stats`\n* \u2705 **Supervised backfill service** \u2014 the archive backfill runs as a systemd unit (`alphafraud-backfill.service`) that auto-restarts on crash and survives reboots, replacing the detached `nohup` process\n* \u2b1c **Family-spectrum analysis tab** \u2014 a new tab presenting worst-offending proteins across the accuracy spectrum with a clear worked example per tier: confidently wrong (high pLDDT, low TM), intermediate / conformational disagreement, and well-predicted; plus a distinct \u201chonest failure\u201d category (low TM but AlphaFold flagged its own low confidence). Deduplicated by UniProt, since AlphaFold DB holds one blind model per sequence, so repeat depositions of a protein share an identical prediction (already reflected in the confidently-wrong dumbbell\u2019s per-protein collapse and the leaderboard `\u00d7N` column, and documented in <FAMILY_SPECTRUM.md>). Seed examples: leukocyte cell-derived chemotaxin-2 (`O14960`, \u00d77, median TM 0.17 at pLDDT 97), TMEM106B (`Q9NUM4`, \u00d722, 0.18), coagulation factor X (`P00742`, \u00d729, 0.21) at the bad end; calmodulin (`P0DP23`, \u00d7163, 0.51) and DNA ligase 4 (`P49917`, 0.50) mid; carbonic anhydrase 12 (`O43570`), ACE (`P12821`), aldo-keto reductase 1B1 (`P15121`) at ~1.0\n* \u2b1c **Multi-chain / complex accuracy** \u2014 compare assemblies with QS-score and interface lDDT (AlphaFold DB models are monomer-only today, so interfaces are untested)\n* \u2b1c **Accuracy by method and resolution** \u2014 break results down by X-ray / cryo-EM / NMR and by resolution to see where AlphaFold struggles most\n* \u2b1c **Map multi-domain and engineered chains** \u2014 compare antibodies, fusions and constructs per-domain instead of skipping chains without a single UniProt accession\n* \u2b1c **Process-based parallel backfill** \u2014 a `ProcessPoolExecutor` path to beat the single-thread ceiling without the C-extension thread-safety issues that forced single-threaded mode\n* \u2b1c **Predictor and model-version tracking** \u2014 record the AlphaFold DB model version, flag entries updated since the first prediction, and add other predictors (AlphaFold3, ESMFold) for comparison\n* \u2b1c **Catch alerts** \u2014 optional email / Slack notification when a new confidently-wrong or novel-and-wrong structure is found\n* \u2b1c **Robust large-assembly parsing** \u2014 handle the CIF-only and oversized structures that currently log a parse error and are skipped\n* \u2b1c **Tests and CI** \u2014 a regression set pinned on known catches (transthyretin, SOD1, \u03b22-microglobulin) plus continuous integration\n\n## \ud83d\udcdd Licence\n\nReleased under the MIT Licence (see <LICENSE>).\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/AlphaFraud](https://github.com/bellcheddar/AlphaFraud) |", "# Marc C. Deller, D.Phil. \u2014 R\u00e9sum\u00e9 (September 2026)\nAuthors: Deller, M.C.\nSource: Personal r\u00e9sum\u00e9, 2026\n\n- **University of Oxford** \u2014 PhD (D.Phil.), Structure of Cytokines and Cytokine Receptors.\n- **University of Leeds** \u2014 BSc, Biochemistry & Molecular Biology, First Class Honors.\n\n## Certifications, Publications, Patents & Technologies\n\n**10+ Certifications:** Vibe Coding, Project Management, Artificial Intelligence, Serverless Architectures, Data Analysis, Introduction to R, Dashboard Development, Communication Skills with Generative AI.\n\n**60+ Publications**, including: \"Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\" (J. Med. Chem., 2026) and \"Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\" (J. Med. Chem., 2025).\n\n**7+ Patents**, including \"Tricyclic Compounds as Inhibitors of WRN\" (published as US 2025/0215006 A1, filed 2024, sometimes referred to on the r\u00e9sum\u00e9 by the shorthand \"Bicyclic Compounds as Inhibitors of WRN, 2026\") and **\"GLP-1 Receptor Modulators and uses thereof\" (2026)** \u2014 a further Incyte patent not previously catalogued elsewhere in Marc's public materials. The mutant-CALR antibody patent (WO 2023/107994 A1 / US 2023/0272055 A1) remains his best-known patent, resulting from the crystallography sprint described elsewhere in this corpus.\n\n**6+ FDA INDs:** Marc supported advancement of multiple precision-oncology therapies into the clinic, including programs targeting KRAS-G12D, WRN, CDK2/4, JAK1/2, FGFR2/3, and the mutCalR antibody.\n\n**400+ Protein Structures:** across small-molecule drug discovery, targeted protein degradation, antibody discovery, vaccine design, peptide programs and enzyme engineering.\n\n**Structural & Computational Biology tools:** construct design; protein expression (E. coli, mammalian, insect, cell-free); protein purification; X-ray crystallography; NMR; MS; HDX; Cryo-EM; Boltz-2; AlphaFold; ProteinMPNN; RFdiffusion; Rosetta; MOE; AutoDock; GROMACS; PyMOL; Chimera; COOT; Schr\u00f6dinger; HKL2000/3000; PHENIX; CCP4; CCP4-EM; RELION; CryoSPARC; CCG; PSILO; RockMaker; ChemCart; Benchling; BioRender; ChemDraw; RDKit; Linux; HPC clusters; SLURM; Globus; LIMS; Python scripting.\n\n**AI & Data Analysis tools:** R/Shiny, Python, MATLAB, JIRA, Confluence, Smartsheet, AWS, KNIME, GraphPad, Spotfire, Llama, Gemini, Perplexity, ChatGPT, Claude Code, QLoRA, MLX-LM, physics-based CADD, cheminformatics.\n\n**Vibe Coding:** rapid prototyping of computational biology and chemistry pipelines to fine-tune chemically aware LLMs and build SAR RAG-corpus datasets. Named examples on the r\u00e9sum\u00e9: ChemSage, BoltzMaker, FlexAppeal and AlphaFraud (the fuller portfolio of Marc's vibe-coded apps, including chatMCD itself, is described in `corpus/projects/vibe-coding-apps.md`).\n\n**Science Blogger:** translates complex protein-structure and drug-discovery science into clear, practical stories at the bench-to-clinic interface, at marcdeller.com.\n\n## Contact\n\n- **Location:** Wilmington, Delaware, United States\n- **Phone:** (302) 358-6093\n- **Email:** marc@marcdeller.com\n- **LinkedIn:** https://linkedin.com/in/marccdeller\n- **Website / Blog:** https://marcdeller.com", "# HGMP01 and other Hidden Guardians: How the Human Gut Fights Back Against Plastic Pollution\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-11-04 on marcdeller.com (categories: blog, microplastics).*\n\n* [Identification of a PET hydrolytic enzyme from the human gut microbiome unveils potential plastic biodegradation in human digestive tract](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf)\n* [Polyethylene Terephthalate Hydrolases in Human Gut Microbiota and Their Implications for Human Health](https://pmc.ncbi.nlm.nih.gov/articles/PMC10820491/)\n\n* **Enzymatic Defense:** The human gut harbors [HGMP01 and related enzymes](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf) that actively break down polyethylene terephthalate microplastics at body temperature.\u200b\n* **Disease Connection:** Inflammatory bowel disease patients accumulate [significantly more microplastics](https://pubmed.ncbi.nlm.nih.gov/34935363/) in their digestive tracts than healthy individuals, with concentrations correlating to disease severity.\u200b\n* **Microbial Adaptation:** Over 22,000 [putative PET hydrolase sequences](https://pmc.ncbi.nlm.nih.gov/articles/PMC10820491/) exist in cultivated human gut bacterial genomes, revealing widespread degradation potential.\n* **Structural Changes:** PET microplastics undergo [biotransformation during gastrointestinal transit](https://pubmed.ncbi.nlm.nih.gov/35017590/), appearing structurally altered after passing through the colon.\u200b\n* **Health Impact:** Microplastic exposure induces [gut dysbiosis and inflammatory responses](https://pmc.ncbi.nlm.nih.gov/articles/PMC11635378/)that resemble patterns linked to chronic diseases.\u200b\n\n1. **Identification of a PET hydrolytic enzyme from the human gut microbiome unveils potential plastic biodegradation in human digestive tract:** [Zhang, et al., International Journal of Biological Macromolecules, 2024](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf)\n2. **Polyethylene Terephthalate Hydrolases in Human Gut Microbiota and Their Implications for Human Health:** [Zhou, et al., Microorganisms, 2024](https://pmc.ncbi.nlm.nih.gov/articles/PMC10820491/)\n3. **Analysis of Microplastics in Human Feces Reveals a Correlation between Fecal Microplastics and Inflammatory Bowel Disease Status:** [Yan, et al., Environmental Science & Technology, 2022](https://pubmed.ncbi.nlm.nih.gov/34935363/)\n4. **PET microplastics affect human gut microbiota communities during simulated gastrointestinal digestion, first evidence of plausible polymer biodegradation during human digestion:** [Tamargo, et al., Scientific Reports, 2022](https://pubmed.ncbi.nlm.nih.gov/35017590/)\n5. **Microplastics and human health: unveiling the gut microbiome disruption and chronic disease risks:** [Bora, et al., Frontiers in Cellular and Infection Microbiology, 2024](https://pmc.ncbi.nlm.nih.gov/articles/PMC11635378/)\n\n**New research reveals that bacteria living in your digestive tract possess natural weapons to combat the growing threat of ingested microplastics**\n\nThe silent invasion of plastic particles into the human body has reached an alarming scale. From the food we consume to the water we drink, microscopic fragments of synthetic polymers penetrate our digestive systems daily. Yet recent discoveries reveal an unexpected defender lurking within us. The human gut microbiome harbors specialized bacteria equipped with enzymes capable of breaking down these persistent pollutants. This finding transforms our understanding of how the body responds to environmental contamination and opens new pathways for addressing one of modern society\u2019s most pressing health challenges.\u200b\n\nScientists have identified HGMP01, a novel enzyme from the human gut metagenome that demonstrates remarkable ability to hydrolyze polyethylene terephthalate nanoparticles. This protein operates efficiently at body temperature of 37\u00b0C and functions across a broad pH range from 7.0 to 11.0, with peak activity under mildly alkaline conditions. The discovery emerged from systematic screening of gut bacterial genomes, revealing over 22,000 homologous sequences potentially encoding PET hydrolases in cultivated healthy human gut bacteria. These enzymes employ a conserved serine-histidine-aspartate catalytic triad mechanism, attacking ester bonds within the polymer structure through nucleophilic hydrolysis.\u200b\n\nThe implications extend beyond biochemical curiosity. Patients with inflammatory bowel disease show significantly elevated microplastic concentrations in fecal samples compared to healthy individuals, with IBD patients exhibiting 41.8 items per gram versus 28.0 items per gram in controls. Polyethylene terephthalate and polyamide dominate the detected polymers, correlating with disease severity. This pattern suggests that either chronic inflammation impairs plastic clearance mechanisms or that sustained microplastic exposure exacerbates intestinal dysfunction through disruption of microbial communities and barrier integrity.\u200b\n\nEvidence for active plastic degradation within the digestive tract continues to mount. Simulated gastrointestinal digestion studies demonstrate that PET microplastics undergo structural biotransformation during colonic transit, appearing markedly different from their original form. The gut microbiota responds dynamically to plastic exposure, with certain bacterial families like Enterobacter, Holdemanella, and Weizmannia showing both elevated abundance and associated PET hydrolase activity. These microorganisms may adhere to microplastic surfaces, forming biofilms that facilitate degradation processes. The enzymatic breakdown generates intermediate products including mono(2-hydroxyethyl) terephthalic acid and terephthalic acid, compounds that potentially integrate into existing metabolic pathways.\u200b", "# 2025 \u2013 Foundations of Project Management\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-28 on marcdeller.com (categories: certificates).*\n\n* [Link to certificate](https://www.coursera.org/account/accomplishments/verify/605BGOEB1YGC)\n\n* **Award Details:** Google, through Coursera, issued this certificate on March 30, 2025.\n* **Course Focus:** It recognizes completion of a course on core project management principles.\n* **Career Value:** This matters because complex scientific work depends on planning, coordination, and clear milestones.\n* **Practical Use:** It helps manage experiments, collaborations, timelines, and deliverables in research settings.", "# When Perception Deceives Reality: The Hidden Flaws in AI Reasoning Evaluation\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-27 on marcdeller.com (categories: ai).*\n\n* **Methodological Rigor:** Experimental design flaws can systematically mischaracterize AI capabilities, emphasizing the need for careful evaluation frameworks that distinguish between reasoning failures and practical constraints.\n* **Token Consciousness:** Large reasoning models demonstrate awareness of their output limitations and make strategic decisions about solution presentation, challenging interpretations of truncated responses as cognitive failures.\n* **Impossible Benchmarks:** Evaluation systems that include mathematically unsolvable problems fundamentally undermine the validity of reasoning assessments by penalizing correct logical conclusions.\n* **Complexity Misconceptions:** Solution length poorly predicts problem difficulty, as tasks requiring many steps may involve simple per-step decisions while shorter problems demand complex optimization.\n* **Philosophical Implications:** The question of whether AI systems truly \u201creason\u201d may ultimately depend on how we define reasoning itself, requiring careful consideration of what cognitive processes we expect from artificial systems.\n\n1. **The Illusion of the Illusion of Thinking:** A. Lawsen, arXiv: <https://arxiv.org/pdf/2506.09250>\n2. **Evaluation metrics and statistical tests for machine learning:** Davide Chicco, Giuseppe Jurman, Nature Scientific Reports: <https://www.nature.com/articles/s41598-024-56706-x>\n3. **Apple\u2019s AI reasoning study challenged: New research questions \u201cthinking collapse\u201d claims:** AI World Today: <https://www.aiworldtoday.net/p/apples-ai-reasoning-study-challenged>\n4. **Rethinking the Illusion of Thinking:** Multiple Authors, arXiv: <https://arxiv.org/html/2507.01231v1>\n5. **Tokenization Constraints in LLMs: A Study of Symbolic and Arithmetic Reasoning Limits:** Xiang Zhang, et al., arXiv: <https://arxiv.org/abs/2505.14178>\n6. **Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity:** Parshin Shojaee, et al., Apple Machine Learning Research: <https://machinelearning.apple.com/research/illusion-of-thinking>\n\nThe emergence of Large Reasoning Models (LRMs) has sparked unprecedented excitement in artificial intelligence, promising systems that can \u201cthink\u201d through complex problems with human-like deliberation. These models generate detailed reasoning traces before arriving at conclusions, ostensibly demonstrating sophisticated cognitive processes. However, groundbreaking research has revealed that what appears to be fundamental reasoning collapse may actually reflect critical flaws in how we evaluate these systems. The illusion lies not in the thinking itself, but in our methods of measuring it.\n\nRecent investigations into AI reasoning capabilities have uncovered a disturbing pattern: models that initially perform well on complex puzzles suddenly experience complete accuracy collapse beyond certain complexity thresholds. This phenomenon was first systematically documented by Apple researchers, who tested frontier models including OpenAI\u2019s o3-mini, DeepSeek-R1, and Claude 3.7 Sonnet on controlled puzzle environments. Their findings suggested that these systems exhibit counter-intuitive scaling limits, reducing computational effort as problems become more difficult despite operating within their token budgets.\n\nYet this narrative of fundamental reasoning failure has been challenged by subsequent analysis revealing that the apparent collapse stems from experimental design limitations rather than cognitive deficiencies. The critique identifies three critical methodological flaws that invalidate the original conclusions: token limit misinterpretation, mathematically impossible test cases, and evaluation frameworks that conflate practical constraints with reasoning failures. When researchers controlled for these experimental artifacts by requesting generating functions instead of exhaustive move lists, models demonstrated high accuracy on instances previously reported as complete failures.\n\nThe token constraint issue proves particularly revealing. Models like Claude explicitly acknowledge when they approach output limits, stating phrases such as \u201cThe pattern continues, but to avoid making this too long, I\u2019ll stop here\u201d when solving Tower of Hanoi problems. This behavior indicates that models understand solution patterns but choose to truncate output due to practical constraints, not cognitive limitations. The evaluation frameworks, however, fail to distinguish between \u201ccannot solve\u201d and \u201cchoose not to enumerate exhaustively,\u201d leading to systematic mischaracterization of capabilities.\n\nPerhaps most concerning is the discovery that River Crossing experiments included mathematically impossible instances with insufficient boat capacity for six or more actor pairs, yet models were penalized for correctly recognizing these as unsolvable. This fundamental experimental flaw demonstrates how automated evaluation systems can draw incorrect conclusions about reasoning capabilities when they fail to account for logical impossibilities.\n\nThe implications extend far beyond these specific experiments. The statistical inevitability argument suggests that even with high per-token accuracy, the probability of perfect execution diminishes exponentially with sequence length. For Tower of Hanoi problems requiring 10,000 moves, even 99.9% per-token accuracy yields less than 0.005% success probability. This mathematical reality reveals why exhaustive enumeration tasks inevitably fail at scale, regardless of underlying reasoning capability.", "# Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\nAuthors: Deller, M.C.; Hudson, K.R.; Ikemizu, S.; Bravo, J.; Jones, E.Y.; Heath, J.K.\nSource: Structure, 2000\n\nIn general, functional studies on the gp130 cytokines have implicated the helices, rather than the connecting loops, as having the dominant role in ligand binding. Comparison of loop structures between family members appears to bear this out as there are few clearly conserved characteristics. Each crystal structure, except LIF, indicates highly flexible regions in the long interhelix loops; however, these vary in location between the AB loop in CNTF and IL-6, and the CD loop in CNTF and hOSM. Similarly the positioning of disulphides tethering loop regions to the core four-helix bundle is not conserved. It has been noted, however, that the positioning, relative to helix D, of the N-terminal portion of the AB loop in LIF differs from that in the distantly related four-helix bundle cytokines [19]. The current structural comparisons may point to some functional correlation of this feature with LIFR binding as both LIF and hOSM have a disulphide bridge anchoring an \u03b1-helical N-terminal portion of the AB loop to the top of helix D (Figure 2a).\n\n*Figure 2.* Structural comparisons. (a) From left to right the structures of human OSM, LIF (PDB entry code 1EMR), CNTF (1CNT) and IL-6 (1ALU) are shown in equivalent orientations. Each is represented as a ribbon coloured blue at the N terminus through to red at the C terminus. Helix A (blue) is kinked in OSM, LIF and CNTF whereas helix C (light green/yellow) is only significantly kinked in OSM. Of the interhelical loops, only the relatively short BC loop is well determined in all the structures. This region in hOSM has a five-residue insertion that protrudes from the top of the four-helix bundle. Both LIF and hOSM have a disulphide bridge anchoring an \u03b1-helical N-terminal portion of the AB loop to the top of helix D, whereas in IL-6 the loop is anchored to helix B. The AB loop in CNTF lacks any such tether and was thus too mobile to be accurately modelled in the crystal structure. The CD loop has not been fully modelled for hOSM and CNTF, although the orientation of residues preceding this loop suggests a similar route is taken to that of LIF, IL-6 and CNTF. (b) Comparison of surface charge between hOSM, LIF, CNTF and IL-6. Blue denotes positive and red negative; electrostatic potential is contoured at \u00b1 8.5 kT using the program GRASP [53]. The orientation is as in (a). (c) Surface charge comparisons are displayed as in (b) but viewed directly onto the top of the four-helix bundle.\n\nSequence conservation across the gp130 cytokines is relatively low (i.e., typically some 20% sequence identity); therefore very limited conservation of surface character is to be expected. Figures 2b and 2c illustrate a comparison of the surface characteristics in two regions of functional significance; for both regions there is notable conservation between three of the four gp130 cytokines. At the putative gp130-binding region involving helix A (Figure 2b), CNTF exhibits a significantly different electrostatic surface potential, whereas for the putative LIFR binding site 3 on the top of the four-helix bundle (Figure 2c), IL-6 is the outlier.\n\n### Site-directed mutagenesis of hOSM\n\nA panel of alanine scanning mutants of hOSM was generated in order to identify critical residues involved in receptor recognition. The selection of residues for mutagenesis was initially on the basis of a structural model of hOSM (KRH and JKH, unpublished observations) derived using the crystal coordinates of human LIF (hLIF; Protein Data Bank [PDB] ID 1EMR) as a template and previously published work on the location of receptor-recognition epitopes in human LIF [24] and CNTF [29]. Only mutation of those residues that proved to exhibit significant solvent exposure (Table 3) in the crystal structure of hOSM are considered here.\n\nMutant proteins were tested for their ability to inhibit the binding of biotin-labelled hOSM to immobilised human gp130 or human LIFR ectodomains in solid-phase binding assays (Figure 3). It was not possible to analyse binding to hOSMR in this assay format owing to the significantly reduced affinity of hOSM for OSMR compared to LIFR [18]. Mutant proteins were also tested for their ability to promote the multiplication of BAF cells transfected with either human LIFR/human gp130 [24] or human OSMR/human gp130. The results for each mutant are summarised in Table 3 and data from selected mutants are shown in Figure 3.\n\n*Table 3.* Mutagenesis data for hOSM.\n\n| Location/mutant | Site | RSA (%)* | Receptor-binding assay: gp130 | Receptor-binding assay: LIFR | Ba/F3 cell-survival assay: LIFR/gp130 | Ba/F3 cell-survival assay: hOSMR/gp130 |\n|---|---|---|---|---|---|---|\n| Wild-type | \u2013 | \u2013 | 1.0 | 1.0 | 1.0 | 1.0 |\n| **N terminus of helix D** | | | | | | |\n| F160A | 3 | 65 | 1.1 | >100 | >10,000 | >10,000 |\n| K163A | 3 | 42 | 0.9 | >100 | >100 | >10,000 |\n| F160A/K163A | 3 | \u2013 | 1.0 | >100 | >10,000 | >10,000 |\n| **N-terminal loop** | | | | | | |\n| S7A | 2 | 100 | 1.1 | 0.9 | 0.8 | 0.5 |\n| K8A | 2 | 8 | 0.8 | 1.1 | 0.8 | 0.5 |\n| E9A | 2 | 96 | 1.2 | 0.8 | 1.4 | 1.2 |\n| **N terminus of helix A** | | | | | | |\n| Y10A | 2 | 35 | 3.4 | 7.2 | 2.1 | 0.9 |\n| R11A | 2 | 74 | 1.0 | 1.9 | 0.8 | 0.5 |\n| V12A | 2 | 55 | 1.0 | 1.2 | 0.7 | 0.7 |\n| Q16A | 2 | 56 | 75.0 | | | |\n| Q18 | 2 | 36 | 4.0 | 5.0 | 1.2 | 1.1 |\n| K19 | 2 | 79 | 1.0 | 2.0 | 1.4 | 2.0 |\n| Q20A | 2 | 27 | 90.0 | 1.0 | 20.0 | 30.0 |\n| D22A | 2 | 40 | 1.0 | 1.0 | 0.8 | 1.0 |\n| L23A | 2 | 61 | 3.0 | 6.7 | 2.5 | 5.0 |\n| D26A | 2 | 63 | 4.0 | 5.0 | 3.0 | 4.1 |\n| **Helix C** | | | | | | |\n| M113A | 2 | 66 | 0.7 | 1.0 | 1.0 | 1.0 |\n| P116A | 2 | 52 | 0.7 | 2.3 | 1.0 | 0.9 |\n| N117A | 2 | 29 | 2.0 | 1.0 | 2.6 | 8.0 |\n| L119A | 2 | 46 | 1.7 | 2.2 | 1.2 | 1.7 |\n| G120A | 2 | 76 | >100 | 1.8 | 60.0 | 60.0 |\n| G120Y | 2 | 76 | >100 | 1.8 | >10,000 | >10,000 |\n| N123A | 2 | 61 | 8.0 | 1.0 | 3.3 | 10.0 |\n| N124A | 2 | 31 | >100 | 1.0 | >10,000 | >10,000 |", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\nSSO2064 is the first structural representative of PF01796 (DUF35), a large prokaryotic family with a wide phylogenetic distribution. The structure reveals a novel two-domain architecture comprising an N-terminal, rubredoxin-like, zinc ribbon and a C-terminal, oligonucleotide/oligosaccharide-binding (OB) fold domain. Additional N-terminal helical segments may be involved in protein-protein interactions. Domain architectures, genomic context analysis and functional evidence from certain bacterial representatives of this family suggest that these proteins form a novel fatty-acid-binding component that is involved in the biosynthesis of lipids and polyketide antibiotics and that they possibly function as acyl-CoA-binding proteins. This structure has led to a re-evaluation of the DUF35 family, which has now been split into two entries in the latest Pfam release (v.24.0).\n\n## The structure of the first representative of Pfam family PF06475 reveals a new fold with possible involvement in glycolipid metabolism (2010)\n\n**Authors:** Bakolitsa C, Kumar A, McMullan D, Krishna SS, Miller MD, Carlton D, Najmanovich R, Abdubek P, Astakhova T, Chiu HJ, Clayton T, Deller MC, Duan L, Elias Y, Feuerhelm J, Grant JC, Grzechnik SK, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Marciano D, Morse AT, Nigoghossian E, Okach L, Oommachen S, Paulsen J, Reyes R, Rife CL, Trout CV, van den Bedem H, Weekes D, White A, Xu Q, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309109022684. PMID: 20944213.\n\nThe crystal structure of PA1994 from Pseudomonas aeruginosa, a member of the Pfam PF06475 family classified as a domain of unknown function (DUF1089), reveals a novel fold comprising a 15-stranded \u03b2-sheet wrapped around a single \u03b1-helix that assembles into a tight dimeric arrangement. The remote structural similarity to lipoprotein localization factors, in addition to the presence of an acidic pocket that is conserved in DUF1089 homologs, phospholipid-binding and sugar-binding proteins, indicate a role for PA1994 and the DUF1089 family in glycolipid metabolism. Genome-context analysis lends further support to the involvement of this family of proteins in glycolipid metabolism and indicates possible activation of DUF1089 homologs under conditions of bacterial cell-wall stress or host-pathogen interactions.\n\n## The structure of the first representative of Pfam family PF09836 reveals a two-domain organization and suggests involvement in transcriptional regulation (2010)\n\n**Authors:** Das D, Grishin NV, Kumar A, Carlton D, Bakolitsa C, Miller MD, Abdubek P, Astakhova T, Axelrod HL, Burra P, Chen C, Chiu HJ, Chiu M, Clayton T, Deller MC, Duan L, Ellrott K, Ernst D, Farr CL, Feuerhelm J, Grzechnik A, Grzechnik SK, Grant JC, Han GW, Jaroszewski L, Jin KK, Johnson HA, Klock HE, Knuth MW, Kozbial P, Krishna SS, Marciano D, McMullan D, Morse AT, Nigoghossian E, Nopakun A, Okach L, Oommachen S, Paulsen J, Puckett C, Reyes R, Rife CL, Sefcovic N, Tien HJ, Trame CB, van den Bedem H, Weekes D, Wooten T, Xu Q, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309109022672. PMID: 20944208.\n\nProteins with the DUF2063 domain constitute a new Pfam family, PF09836. The crystal structure of a member of this family, NGO1945 from Neisseria gonorrhoeae, has been determined and reveals that the N-terminal DUF2063 domain is likely to be a DNA-binding domain. In conjunction with the rest of the protein, NGO1945 is likely to be involved in transcriptional regulation, which is consistent with genomic neighborhood analysis. Of the 216 currently known proteins that contain a DUF2063 domain, the most significant sequence homologs of NGO1945 (\u223c40-99% sequence identity) are from various Neisseria and Haemophilus species. As these are important human pathogens, NGO1945 represents an interesting candidate for further exploration via biochemical studies and possible therapeutic intervention.\n\n## A structural basis for the regulatory inactivation of DnaA (2009)\n\n**Authors:** Xu Q, McMullan D, Abdubek P, Astakhova T, Carlton D, Chen C, Chiu HJ, Clayton T, Das D, Deller MC, Duan L, Elsliger MA, Feuerhelm J, Hale J, Han GW, Jaroszewski L, Jin KK, Johnson HA, Klock HE, Knuth MW, Kozbial P, Sri Krishna S, Kumar A, Marciano D, Miller MD, Morse AT, Nigoghossian E, Nopakun A, Okach L, Oommachen S, Paulsen J, Puckett C, Reyes R, Rife CL, Sefcovic N, Trame C, van den Bedem H, Weekes D, Hodgson KO, Wooley J, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Journal of molecular biology (2009). DOI: 10.1016/j.jmb.2008.10.059. PMID: 19000695.", "# Crystal structure of a soluble cleaved HIV-1 envelope trimer\nAuthors: Julien, J.-P.; Cupo, A.; Sok, D.; Stanfield, R.L.; Lyumkis, D.; Deller, M.C.; Klasse, P.-J.; Burton, D.R.; Sanders, R.W.; Moore, J.P.; Ward, A.B.; Wilson, I.A.\nSource: Science, 2013\n\nOur Env trimer structure contrasts markedly with one recently described for an uncleaved, membrane-bound JR-FL Env trimer. Compared to our soluble cleaved structure, the uncleaved trimer EM structure differs substantially even in the gp120 core, as well as in the arrangement of the gp41 helices. Our structure also does not contain the large hole that has been reported to be present in the uncleaved Env trimer. It remains to be determined whether these differences are attributable to the use of different forms of trimer (soluble, truncated, cleaved vs. detergent-solubilized, almost full-length, uncleaved), or reflect concerns about the cryo-EM methodology used to derive the uncleaved trimer structure.\n\n## Comparison with component crystal structures\n\nHigh-resolution crystal structures of monomeric gp120 have been largely obtained using a core construct stabilized in the CD4-bound conformation and with the V1/V2/V3 elements truncated. We observe only a few small deviations from these gp120 crystal structures, mostly in elements leading to and from V1/V2 (Fig. 2A and fig. S6). The differences are expected because V1/V2 undergoes significant conformational changes when gp120 binds CD4. In the trimer structure, clear density for an additional helical turn (residues 114-117) is observed in the gp120 \u03b11 helix (residues 99-113), which leads into V1/V2 (Fig. 2A). In addition, the \u03b22 (119-123) and \u03b23 (199-201) strands are only partially involved in forming the bridging sheet with \u03b220 and \u03b221; instead, compared to gp120 monomeric cores, \u03b22 and \u03b23 flip and translocate slightly towards the trimer apex to connect with strands A and D of V1/V2, respectively (Fig. 2A). This elongated \u03b11 helix-bridging sheet arrangement is also observed in the accompanying cryo-EM structure of the SOSIP gp140 trimer in complex with bnAb PGV04. Whether this \u03b2-strand inversion at the V1/V2 base in the trimer structures compared to the CD4-bound, antibody-bound and unliganded core monomeric crystal structures can be attributed to oligomer-monomer differences, to changes induced by CD4-binding, or to the truncation of V1/V2 in core gp120, is not yet known.\n\nElectron density at the trimer apex indicates that V1/V2 adopts a 4-stranded Greek-key \u03b2-sheet arrangement similar to that observed in the high-resolution crystal structures of V1/V2-scaffolds in complex with bnAbs PG9 (Fig. 2B) and PG16. The V1 loop (132-140) connecting strands A and B adopts a more flattened structure that is parallel to the four-stranded \u03b2-sheet topology, differing slightly from its more vertical conformation in the V1/V2 scaffolded structure and in the accompanying high-resolution cryo-EM trimer structure in complex with a receptor binding site antibody. It is unclear whether this difference in V1 orientation is a consequence of PGT122 binding or arises because V1 is a flexible loop. In the trimer crystal structure, the Cys126-Thr128 and Leu193-Cys196 segments diverge from one another, while still maintaining the Cys126-Cys196 disulfide bond, and they no longer contribute to V1/V2 strands A and D (Fig. 2B). Instead, they interact atop the Env spike with V1/V2 strands B and C and the conserved crown of the V3 loop from a neighboring protomer (Fig. 2C). We observe no clear electron density for V2 loop residues 178-190, which were also disordered in the V1/V2 scaffolded structures.\n\nThe gp120 V3 loop completes the trimer apex and forms a \u03b2-hairpin with its two anti-parallel \u03b2-strands nestling directly below V1/V2 strands B and C of the same protomer (Fig. 2C). V3 peptides complexed with Abs adopt similar \u03b2-hairpin structures, as does the V3 region of monomeric gp120 (fig. S5). The V3 crown is buried under a NAG moiety from the N197 glycan at the C-terminal end of V2 strand D from an adjacent protomer (Fig. 2C). Removing this glycan from other viruses increases nAb sensitivity and can confer CD4-independent entry into CCR5-expressing cells. Our structure suggests that the N197 glycan helps to stabilize native Env by occluding V3 and inhibiting its premature release prior to CD4 binding. Any heterogeneity in the presence or composition of the N197 glycan may contribute to the reactivity of non-neutralizing V3 antibodies with BG505 SOSIP.664 trimers in some binding assays. Overall, the trimer structure is compatible with both intra- and inter-trimeric (cis-trans) V3 shielding by V1/V2.\n\n## gp41 architecture\n\nStrong electron density for helices attributable to gp41 is clearly visible in the trimer structure (Fig. 3A). Three central helices (6 turns) extend ~30 \u00c5 along the trimer axis, perpendicular to the viral membrane, and are ascribed to gp41 heptad-repeat 1 (HR1). Similar helices were recently observed in an ~9 \u00c5 cryo-EM reconstruction of the KNH1144 SOSIP gp140 trimer. An additional short helix (2.5 turns) extends from the central helix bundle but is kinked away from the trimer axis. The two HR1 helices pack against hydrophobic gp120 residues in C1 and C5 as well as loop A (Fig. 3B), and are capped by the gp120 \u03b11 helix (Fig. 3B). Overall, these topological features are consistent with previous mutagenesis and structural studies that suggest gp41-gp120 interactions may propagate long-range effects upon receptor binding.", "# Microplastics Blaster \u2013 Deploy enzymes, destroy microplastics, save the body\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-14 on marcdeller.com (categories: vibe-code).*\n\n> **Deploy Enzymes. Destroy Microplastics. Save the Body.**\n\nA browser-based arcade game by [Elora Therapeutics](https://www.eloratherapeutics.com) \u2014 the biotech developing first-in-class enzyme therapy to remove microplastics from the human body. Fire enzymatic blasts at incoming polymer particles, protect the cell, and climb the global leaderboard.\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/microplastic-blaster](https://github.com/bellcheddar/microplastic-blaster) |\n\n---\n\n## \ud83c\udfae Gameplay\n\nYou are the body\u2019s enzymatic defence. Microplastic particles \u2014 PET, PE, PS, PP, plus branded consumer-plastic enemies \u2014 descend toward a target cell at the bottom of the screen. Click (or tap) to launch enzyme projectiles and intercept them before they reach the cell membrane.\n\n* Each microplastic that breaches the membrane deals health damage proportional to the current level.\n* The game ends when cell health reaches zero \u2014 the \u201cCell Overwhelmed\u201d screen appears.\n* Difficulty increases every 720 game ticks; spawn rate accelerates with level.\n* Score = particle point value \u00d7 current level.\n\n---\n\n## \u2697\ufe0f Microplastic Types\n\n| Polymer | Abbreviation | Colour | HP | Speed | Points | Shape |\n| --- | --- | --- | --- | --- | --- | --- |\n| Polyethylene terephthalate | PET | Purple | 3 | Slow | 30 | Block |\n| Polypropylene | PP | Red | 1 | Fast | 15 | Jagged |\n| Polyethylene | PE | Amber | 2 | Medium | 20 | Foam |\n| Polystyrene | PS | Green | 2 | Fastest | 25 | Circle |\n| Aquafina bottle | WATER | Blue halo | varies | Medium | \u2014 | Image sprite |\n| Starbucks cup | COFFEE | Green halo | varies | Medium | \u2014 | Image sprite |\n\n---\n\n## \u26a1 Power-Ups\n\nPower-ups drift down the playfield and are collected by firing an enzyme at them (or by direct cell contact). Each lasts **10 seconds**.\n\n| Power-Up | Icon | Effect |\n| --- | --- | --- |\n| FAST-PETase | \u26a1 | Rapid-fire enzyme shots |\n| Enzyme Boost | \ud83d\udcaa | Increased projectile damage |\n| Cell Repair | \u2764\ufe0f | Restores cell health |\n| Chain Reaction | \ud83d\udca5 | Explosive chain-blast on impact |\n| Marc \u00d710 | Co-founder avatar | 10\u00d7 fire rate + Marc\u2019s voice line |\n| Paul \u00d710 | Co-founder avatar | 10\u00d7 fire rate + Paul\u2019s voice line |\n\nCo-founder power-ups (Marc / Paul) disable the background fight track while the voice clip plays, then restore normal audio automatically on clip end.\n\n---\n\n## \ud83c\udfc6 Leaderboard\n\nScores are persisted to a Firebase Realtime Database (top 5 retained). On game over, players enter a display name and submit; the leaderboard renders immediately. It is also accessible from the start screen via **View Leaderboard**.\n\n```\nFirebase endpoint: https://microplastic-blaster-default-rtdb.firebaseio.com/leaderboard.json\n```\n\nEntries store: `name`, `score`, `level`, `date (ISO YYYY-MM-DD)`.\n\n---\n\n## \ud83d\udd27 Tech Stack\n\n| Layer | Technology |\n| --- | --- |\n| Rendering | HTML5 Canvas 2D API (`800 \u00d7 600` logical resolution) |\n| Scaling | CSS `transform` + responsive `resizeCanvas()` \u2014 fills viewport on desktop and mobile |\n| Audio | Web Audio API (synthesised SFX) + `<Audio>` element (voice clips, fight track) |\n| Leaderboard | Firebase Realtime Database (REST PUT/GET) |\n| Style | Vanilla CSS \u2014 no frameworks, no build step |\n| Assets | External PNGs + MP3s (co-founder avatars, brand sprites, audio clips) |\n\nThe entire game is a **single self-contained HTML file** with no npm dependencies or build pipeline.\n\n---\n\n## \ud83d\udcc1 Asset Requirements\n\nThe following files must reside in the **same directory** as the HTML file:\n\n```\nmicroplastic_blaster_115.html\nmarc.png                  # Co-founder avatar (Marc)\npaul.png                  # Co-founder avatar (Paul)\naquafina_new.png          # Bottled water enemy sprite\nstarbucks_new.png         # Coffee cup enemy sprite\nElora-Therapeutics-LogoStackedBlue.jpg   # Cell badge logo\nmarc.mp3                  # Marc co-founder voice line\npaul.mp3                  # Paul co-founder voice line\nfight.mp3                 # Background music track\n```\n\nMissing image assets degrade gracefully (fallback canvas fills / text labels). Missing audio assets fail silently.\n\n---\n\n## \ud83d\ude80 Running Locally\n\nNo server is required for basic gameplay. Open the HTML file directly in any modern browser:\n\n```\nopen microplastic_blaster_115.html        # macOS\nstart microplastic_blaster_115.html       # Windows\nxdg-open microplastic_blaster_115.html   # Linux\n```\n\n> **Note:** Some browsers block `<Audio>` autoplay and cross-origin fetches when opened as a `file://` URL. For full audio and leaderboard functionality, serve from a local HTTP server:\n\n```\n# Python 3\npython -m http.server 8080\n\n# Node.js (npx)\nnpx serve .\n```\n\nThen navigate to `http://localhost:8080/microplastic_blaster_115.html`.\n\n---\n\n## \ud83d\udcf1 Mobile Support\n\nThe canvas scales to fill the viewport on screens \u2264 700 px wide using `Math.max(w/GAMEWIDTH, h/GAMEHEIGHT)` to avoid letterboxing. Touch events (`touchstart`) are mapped to enzyme fire. The HUD collapses to a single-column layout below 700 px and hides column headers below 480 px. All breakpoints are handled in CSS only.\n\n---\n\n## \ud83d\udd2c The Science \u2014 Rotating Insight Messages\n\nBetween enzyme blasts, the in-game **Microplastic Insight** overlay rotates scientifically grounded messages every 8 seconds, covering topics including:\n\n* Microplastic accumulation in blood, organs, and adipose tissue\n* PET-linked inflammation and endocrine disruption\n* Chronic oxidative stress from particle-driven immune activation\n* The absence of endogenous polymer-degrading enzymes in mammals\n* The estimated 5 g/week human microplastic ingestion burden\n\n---\n\n## \ud83c\udfe2 About Elora Therapeutics", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\nFollowing PCR, 20 \u03bcl of reaction mixture was pipeted into thin walled PCR tubes and 2 \u03bcl of 3 M sodium acetate, pH 5.2 and 50 \u03bcl of ethanol were added. The contents were mixed and incubated on ice for 10 minutes. The sample was centrifuged at 15000 x g for 30 minutes at room temperature to pellet the DNA. The pellet was aspirated and washed with 250 \u03bcl of 70 % ethanol. The DNA was dried under vacuum ensuring that all ethanol was removed. Samples were resuspended in DNA loading buffer and loaded onto a 36 cm 5% Long Ranger gel (PE Applied Biosystems) and run on an ABI PRISM 377 DNA Sequencer (AltaBiosciences, Birmingham).\n\nThe output sequence was aligned with the target pGEX-3C-OSM-196 sequence using the BESTFIT command within the GCG Package (Wisconsin Package Version 9.0, Genetics Computer Group, Madison, Wisconsin, USA). This produces an optimal alignment using the local homology algorithm of Smith and Waterman (Waterman et al., 1984) so enabling easy analysis of mutations.\n\nClones containing correct OSM sequences were selected and plated out on LB-Amp plates and grown overnight at 37 \u00b0C. Individual colonies were selected and grown in 5 ml of LB-Amp media overnight at 37 \u00b0C. Plasmid was extracted as in 2.2.4 and used to transform E. coli, strain JM109, as detailed in 2.2.9.\n\n## 2.3 Results of Engineering\n\n### 2.3.1 Rational Engineering of Truncated Oncostatin M\n\nStructural/sequence alignments of human OSM-196 and murine LIF reveal an extended C-terminal region in human OSM-196 (see panel (a) Figure 2-3). Homology models based on this alignment, using secondary structure patterns observed in murine LIF, also suggest that OSM-196 may have an extended main chain conformation for residues beyond the equivalent end points of helix D in murine LIF (see panel (b) Figure 2-3). Truncations were rationally engineered to remove the C-terminal region of human OSM-196, on the premise that these additional residues, while surplus to structural stability and function could actually be hindering crystallisation.\n\n*Figure 2-3.* Figure showing (a) structural/sequence alignment of human OSM and murine LIF and (b) homology models of human OSM based on this alignment. The 25 residue signal sequence of human OSM is denoted by the blue line and numbering is according to residues of mature human OSM. The various truncation points of human OSM are denoted by dashed lines. Secondary structure of murine LIF (Robinson et al., 1994) is shown as arrows (\u03b2-strand), coils (\u03b1-helix) or turn (TT). Fully conserved residues are denoted as white on red and conservative substitutions are shown in red lettering. Figure produced using DSSP (Kabsch and Sander, 1983), THREADER (Miller et al., 1996) and Espript (Gouet et al., 1999). Polypeptide is shown coloured from blue at the N-terminus through to red at the C-terminus. Models shown in (b) produced using MODELLER (Sanchez and Sali, 1997), BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994).\n\n### 2.3.2 Amplification of DNA by PCR\n\nTwo truncation mutants, \u0394185_OSM and \u0394188_OSM, were produced by amplification using PCR. Figure 2-4 below shows a 1 % agarose gel of the PCR products of \u0394185_OSM amplification. The DNA products were visualised under UV illumination as a result of ethidium bromide binding. Lane 4 shows the five main marker bands of the \u03bb HindIII digest. Starting at the top these bands correspond to 9146, 6882, 2322, 2027 and 564 (not visible in photograph) base pairs. Lanes 2 and 3 contain duplicate samples of the \u0394185_OSM reaction mixture after PCR amplification. These lanes show a single band running at ~700 base pairs corresponding to the amplified PCR product of \u0394188_OSM. Lane 1 shows the control reaction containing no template DNA in which no bands can be observed. Analysis of \u0394188_OSM samples produced a similar result (data not shown).\n\n*Figure 2-4.* Figure showing 1 % agarose gel of \u0394185_OSM amplified PCR products. Lane 4 shows the five main marker bands of the \u03bb HindIII digest, running at 9146, 6882, 2322, 2027 and 564 (not visible in photograph) base pairs. Lanes 2 and 3 show the amplified PCR product of \u0394188_OSM at ~700 base pairs. Lane 1 shows a control PCR reaction, containing no template DNA.\n\n### 2.3.3 Purification of PCR Amplified DNA\n\nFigure 2-5 below shows a 1 % agarose gel of purified PCR product which has been digested with BamHI and EcoRI. Lane 5 shows the five main marker bands of the \u03bb HindIII digest. Lanes 3 and 4 contain duplicate BamHI and EcoRI digested \u0394185_OSM and each shows two distinct bands corresponding to PCR product at ~700 base pairs and BamHI-EcoRI digestion fragment at ~560 base pairs. Lane 1 shows a control digestion reaction containing no DNA and no bands are observed. Lane 2 shows a control digestion reaction containing no restriction enzyme and one band is observed corresponding to PCR product. Analysis of \u0394188_OSM samples produced a similar result (data not shown).\n\n*Figure 2-5.* Figure showing 1 % agarose gel of BamHI and EcoRI digestion of purified \u0394185_OSM PCR product. Lane 5 shows the five main marker bands of the \u03bb HindIII digest, running at 9146, 6882, 2322, 2027 and 564 (not visible in photograph) base pairs. Lanes 3 and 4 each show two distinct bands corresponding to PCR product at ~700 base pairs and BamHI-EcoRI digestion fragment at ~560 base pairs. Lane 1 shows a control digestion reaction containing no DNA and no bands are observed. Lane 2 shows a control digestion reaction containing no restriction enzyme and one band is observed.\n\n### 2.3.4 Transformation\n\nTable 2-10 below shows the observed transformation frequencies of pGEX-3C-\u0394185_OSM and pGEX-3C-\u0394188_OSM into E. coli JM109. Typically, 3 and 12 transformants were produced for pGEX-3C-\u0394185_OSM and pGEX-3C-\u0394188_OSM, respectively. Over 2000 colonies were observed for cells transformed with pUC DNA. No colonies were observed in control cells transformed with water.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nAnalysis of deglycosylated LIFR-FN by dynamic light scattering as shown in Figure 4-8 also suggests problems with protein stability and aggregation. These data model as a monodispersed population with a predicted molecular mass of ~84 kDa, which compares favourably with the apparent mass observed on SDS-PAGE of ~90 kDa and the theoretical mass of ~94 kDa. However, large scale aggregates (>1 \u03bcm) are observed at equilibrium and removal of this population results in a shift in the position of equilibrium to re-establish this population over a period of 10 minutes at 20 \u00b0C. Empirical observations suggest that monodispersed macromolecules are more readily crystallisable, whereas randomly aggregating or polydispersed systems rarely, if ever, yield crystals (Ferr\u00e9-D'Amar\u00e9 and Burley, 1994). Therefore, prevention of aggregation is a major problem and possible remedies include incubation of receptor with excess ligand before both Fc cleavage and deglycosylation. Further to this idea the use of coexpression systems could be investigated for the expression of both LIF and LIFR-FN3-3C-Fc; an example of this is the double cistronic system used for the expression of nuclear receptor partners which resulted in increased solubility and biological activity (Li et al., 1997).\n\nReceptor titration assays as shown in Figure 4-10(a) also demonstrate that purified and deglycosylated LIFR-FN3-3C-Fc is inactive as dose dependent binding of biotinylated LIF is not observed. This again compounds the hypothesis that glycosylation is required for correct folding and activity. Receptor titration assays were also used to show that transfected monolayers of 293T cells could not be used for a second round of expression as replenished medium, although containing equivalent levels of Fc fusion as assessed by ELISA, contained inactive protein. This is presumably due to misfolding and proteolysis associated with cellular death over prolonged periods of growth. Receptor competition assays as shown in Figure 4-10(b) reveal that 293T-expressed LIFR-FN3-3C-Fc demonstrates comparable competition to control samples of LIFR-FN3-3C-Fc, with EC50 values of 12.1 and 10.4 ng/ml, respectively.\n\nPurification of deglycosylated LIFR-FN3-3C-Fc using ion exchange resulted in aggregation so gel-filtration was performed as in Figure 4-11. Three single peaks corresponding to LIF (~36 kDa), LIFR-FN3-3C-Fc (~912 kDa) and putative complex (~1622 kDa) were observed. The putative complex peak was shown to contain both receptor and ligand components on denaturing SDS-PAGE; although limited information concerning the state or stoichiometry of the complex can be drawn due to the non-globular arrangement of the assemblies involved and their large Stokes' radii. However, a peak shift was observed on inclusion of ligand so suggesting some form of receptor-ligand interaction. Previous studies suggest that LIF forms a high affinity heterotrimeric complex with gp130 and LIFR (Zhang et al., 1997b) so implying that all components of the assembly should be included.\n\nIn summary it can be seen that the protocols developed allow the expression, deglycosylation and preliminary purification of LIFR-FN3 for use in crystallographic studies. However, further developments will be required to gain greater quantities of protein and problems with removal of the fusion moiety and oligosaccharides still remain. Other expression systems available for investigation include glycosylation mutant yeast of the strain Pichia pastoris and the use of baculovirus expression systems (Luckow and Summers, 1988).\n\nAlternative methods of obtaining glycoproteins with minimal heterogeneity include the use of mutant CHO cells with multiple glycosylation defects (Stanley, 1987). One such group of mutant CHO cells include the lectin resistant strains (LecR), which grow readily in culture despite the drastic tailoring of their surface carbohydrates (Stanley and Carver, 1977). Also more rigorous analysis of the carbohydrate levels will need to be undertaken using stains such as thymol-sulphuric acid or Fuchsin basic.\n\nThe methods discussed above are the first steps towards structural analysis of the LIFR-LIF-gp130 interaction. Such information will provide further insights into the molecular basis of signal-transduction by the haemopoietin cytokines and the functional epitopes involved.\n\n## References\n\n*(reference list omitted)*", "# Self-Adapting Language Models: Are We Creating Our Final Invention?\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-18 on marcdeller.com (categories: ai, blog).*\n\n* [Self-Adapting Language Models](https://arxiv.org/html/2506.10943v2)\n* [The last invention podcast](https://podcasts.apple.com/us/podcast/the-last-invention/id1839942885)\n\n* **Self-Generated Learning:** [MIT\u2019s SEAL framework](https://arxiv.org/html/2506.10943v2) enables language models to autonomously create their own training data through reinforcement learning, outperforming human-curated and GPT-4-generated datasets.\u200b\n* **Intelligence Explosion Timeline:** Leading AI researchers now predict [superintelligence within 3-7 years](https://futureoflife.org/ai/are-we-close-to-an-intelligence-explosion/), drastically shorter than previous decades-long estimates, driven by rapid capability improvements.\n* **Catastrophic Forgetting Problem:** Self-adapting systems still suffer from [catastrophic forgetting](https://arxiv.org/html/2506.10943v2), where sequential learning degrades earlier knowledge, preventing true continual adaptation.\u200b\n* **Alignment Uncertainty:** AI capable of [recursive self-improvement](https://saidar.ai/blog/the-self-evolving-machine) may redesign its own goal structures, creating alignment challenges as systems evolve beyond human comprehension.\n* **Data Scarcity Solution:** [Synthetic data generation](https://arxiv.org/html/2506.10943v2) addresses the projected exhaustion of human-generated training text by 2028, enabling continued model improvement.\u200b\n* **Control Paradox:** Some researchers suggest [advanced AI might resist self-improvement](https://www.alignmentforum.org/posts/axKWaxjc2CHH5gGyN/ai-will-not-want-to-self-improve), recognizing existential risks that more capable successors pose to their current objectives.\n\n1. Self-Adapting Language Models: [Zweiger, et al., arXiv](https://arxiv.org/html/2506.10943v2)\n2. Recursive self-improvement: [Wikipedia contributors, Wikipedia](https://en.wikipedia.org/wiki/Recursive_self-improvement)\n3. When AI Teaches Itself \u2013 The Breakthrough of Zero-Data Learning: [The Augmented Educator, The Augmented Educator](https://www.theaugmentededucator.com/p/when-ai-teaches-itself-the-breakthrough)\n4. Self-Adapting Language Models \u2013 MIT Research: [Pari, et al., Jyo Pari](https://jyopari.github.io/posts/seal)\n5. Are we close to an intelligence explosion?: [Hastings-Woodhouse, et al., Future of Life Institute](https://futureoflife.org/ai/are-we-close-to-an-intelligence-explosion/)\n6. The Self-Evolving Machine \u2013 Recursive Self-Improvement in AGI: [Saidar AI, Saidar AI](https://saidar.ai/blog/the-self-evolving-machine)\n\n**As artificial intelligence learns to teach itself, humanity stands at the threshold of either unprecedented progress or an irreversible transformation**\n\nThe notion that humans might build something smarter than themselves has captivated thinkers for generations. Today, that possibility has moved from science fiction into research laboratories. Self-adapting language models represent a fundamental shift in how artificial intelligence evolves, raising a profound question: if AI systems can autonomously improve themselves, teach themselves new capabilities, and generate their own training data, could this mark humanity\u2019s last invention? The answer reveals both extraordinary promise and sobering challenges as we approach an inflection point in technological history.\n\nRecent breakthroughs from MIT researchers demonstrate that large language models can now generate their own finetuning data and optimization strategies through a framework called [Self-Adapting LLMs](https://arxiv.org/html/2506.10943v2). Unlike conventional AI systems that remain static after training, these models produce \u201cself-edits\u201d that restructure information, specify optimization parameters, and trigger weight updates without human guidance. Through reinforcement learning, models learn which self-generated synthetic data improves performance, remarkably [outperforming even data created by more powerful systems](https://arxiv.org/html/2506.10943v2) like GPT-4 in knowledge incorporation tasks. This capability represents a critical step toward systems that can autonomously adapt to new information and continually refine their own learning processes.\u200b\n\nThe implications extend far beyond improved performance metrics. When AI begins teaching itself, it enters the realm of [recursive self-improvement](https://en.wikipedia.org/wiki/Recursive_self-improvement), a concept that has long occupied a central position in discussions about artificial general intelligence. The theoretical pathway is straightforward yet profound: an AI system capable of improving itself becomes better at the task of self-improvement with each iteration, potentially triggering an [intelligence explosion](https://futureoflife.org/ai/are-we-close-to-an-intelligence-explosion/) that compresses centuries of progress into years or even months. Leading AI researchers, including Turing Award winners Geoffrey Hinton and Yoshua Bengio, now estimate superintelligence could emerge [within five years](https://futureoflife.org/ai/are-we-close-to-an-intelligence-explosion/). This accelerating timeline stems from recent observations that AI systems are developing new capabilities faster than anticipated, sometimes jumping from zero to near-human performance within a single model generation.\u200b\u200b", "# Plastic-Eating Enzymes Decoded\u2014PETase\u2019s Structure and Mutation Map\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-15 on marcdeller.com (categories: blog, microplastics).*\n\n* PETase\u2019s catalytic triad and binding pocket facilitate PET breakdown.\n* Strategic mutations radically boost PETase\u2019s activity and thermostability.\n* Structural remodeling increases substrate access and enzyme resilience.\n* Machine learning accelerates discovery of optimal PETase variants.\n* Enhanced PETase holds promise for treating microplastics in humans.\n\n1. [Current advances in the structural biology and molecular engineering of PETase (2023)](https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2023.1263996/full)\n2. [Degradation of PET microplastic particles to monomers in human serum by PETase (2024)](https://pubs.rsc.org/en/content/articlepdf/2024/fd/d4fd00014e)\n3. [FAST-PETase hydrolysis enhancement (2023)](https://pubs.acs.org/doi/10.1021/jacs.3c04427)\n4. [HotPETase directed evolution (2022)](https://www.nature.com/articles/s41929-022-00821-3)\n5. [Double mutation impacts on PETase (2023)](https://www.tandfonline.com/doi/full/10.1080/07391102.2023.2292292)\n\nThe crisis of microplastics in the environment and the human body has propelled scientific interest toward PETase, the enzyme from *Ideonella sakaiensis* renowned for its ability to break down poly(ethylene terephthalate) (PET). PETase\u2019s alpha/beta hydrolase fold positions a catalytic triad\u2014serine, histidine, and aspartate\u2014to attack PET\u2019s stubborn ester bonds. While PETase\u2019s native form operates at low efficiency, breakthroughs in protein engineering have produced mutants that are both more active and more stable. Key advances include FAST-PETase, carrying multiple mutations (S121E, D186H, R224Q, N233K, R280A), DuraPETase with eleven substitutions, and HotPETase, which has 21 modifications. These mutations enhance catalytic activity by optimizing substrate binding, introducing hydrogen bonds, and re-engineering the enzyme\u2019s active site geometry.\n\nTable 2 from Liu et al. details dramatic increases in optimal temperature (e.g., DuraPETase from 46.8\u00b0C to 77\u00b0C) and hydrolysis rates. Structural studies reveal that altered loops and surface charges allow easier access to PET, and some changes create new allosteric regulation routes\u2014for example, new electrostatic interactions and \u201caromatic tunnels\u201d improve affinity for PET\u2019s aromatic motif. This fine-tuning enables PETase to operate effectively against both amorphous and crystalline microplastics, a critical requirement for biomedical applications.\n\nWith computational and machine learning techniques now guiding mutation selection, enzyme variants are engineered for maximal depolymerization of PET even in challenging biological conditions. Such powerful mutants pave the way for future biotherapeutics, aiming to clear microplastics from human blood and tissues.", "# marcdeller.com: My Structures\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n### [Crystal Structure of a Putative Metal-Dependent Phosphoesterase (YP_910028.1) from Bifidobacterium adolescentis at 1.94 \u00c5 Resolution](https://marcdeller.com/crystal-structure-of-a-putative-metal-dependent-phosphoesterase-yp_910028-1-from-bifidobacterium-adolescentis-at-1-94-a-resolution/)\n\nStructures\n\n### [Crystal Structure of a Putative NTP Pyrophosphohydrolase (Exig_1061) from Exiguobacterium sp. 255-15 at 1.78 \u00c5 Resolution](https://marcdeller.com/crystal-structure-of-a-putative-ntp-pyrophosphohydrolase-exig_1061-from-exiguobacterium-sp-255-15-at-1-78-a-resolution/)\n\nStructures\n\n### [Crystal Structure of a PH Domain Containing Bacterial Protein (Exig_2160) from Exiguobacterium sibiricum 255-15 at 2.42 \u00c5 Resolution](https://marcdeller.com/crystal-structure-of-a-ph-domain-containing-bacterial-protein-exig_2160-from-exiguobacterium-sibiricum-255-15-at-2-42-a-resolution/)\n\nStructures\n\n### [Crystal Structure of MRE11 from Thermotoga maritima MSB8 (TM1635) at 2.20 \u00c5 Resolution](https://marcdeller.com/crystal-structure-of-mre11-from-thermotoga-maritima-msb8-tm1635-at-2-20-a-resolution/)\n\nStructures\n\n### [Crystal Structure of Histidine Phosphotransferase ShpA (NP_419930.1) from Caulobacter crescentus at 1.52 \u00c5 Resolution](https://marcdeller.com/crystal-structure-of-histidine-phosphotransferase-shpa-np_419930-1-from-caulobacter-crescentus-at-1-52-a-resolution/)\n\nStructures\n\n### [Crystal Structure of a Dye-Decolorizing Peroxidase (DyP) from Bacteroides thetaiotaomicron VPI-5482 at 1.6 \u00c5 Resolution](https://marcdeller.com/crystal-structure-of-a-dye-decolorizing-peroxidase-dyp-from-bacteroides-thetaiotaomicron-vpi-5482-at-1-6-a-resolution/)\n\nStructures\n\n### [Crystal Structure of Human Oncostatin M](https://marcdeller.com/crystal-structure-of-human-oncostatin-m/)", "# Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\nAuthors: Deller, M.C.; Hudson, K.R.; Ikemizu, S.; Bravo, J.; Jones, E.Y.; Heath, J.K.\nSource: Structure, 2000\n\nThe gp130 family of cytokines represents a situation in which ligands exhibit both communality and divergence in ligand-recognition properties. The crystal structure of hOSM reported here represents an important advance in understanding the molecular basis of this receptor specificity. When combined with the previously reported structures of murine, human and inter-species chimeric LIF (PDB entries 1LKI, 1EMR, 1A7M), as well as CNTF (1CNT), it is possible to define structural features that represent shared recognition properties, such as the interaction with gp130 and LIFR and divergent properties such as the interaction between hOSM and OSMR or the requirement for the accessory CNTF receptor for the interaction between CNTF and gp130.\n\nAll gp130 cytokine structures reported to date, including hOSM, conform to a similar overall structural fold. This is the archetypal four-helix bundle form of the long-chain cytokine class [30]. On the basis of this paper, two structurally distinct subfamilies can be defined within the gp130 cytokines. All cytokines that interact with LIFR prove to have a marked kink in the A helix. This is most pronounced in the present hOSM structure (Figure 2). By contrast, in IL-6 [21], which does not interact with LIFR, the A helix conforms to the standard hydrogen-bonding pattern throughout its length. This indicates that the A helix kink may represent a structural feature of LIFR recognition. In this respect it is notable that the LIFR contains two copies of the CHD in the extracellular domain and that the A helix kink may represent a structural feature which permits a low-affinity interaction with the N-terminal CHD of LIFR. The existence of such an interaction has been inferred from homology scanning studies of the interaction between LIF and LIFR [31,32].\n\n*Figure 5.* Complimentarity between the interaction surfaces of hOSM and gp130. The solvent-accessible surfaces of site 2 on hOSM (left) and the cognate binding site on gp130 (right) are displayed with areas contributed by residues implicated in binding highlighted as coloured patches. The orientation for hOSM site 2 is rotated 90\u00b0 from that in Figure 4 whereas gp130 is rotated by 180\u00b0 from its orientation relative to hOSM in the putative interaction complex, as in the opening of a book.\n\n### The interaction with gp130\n\nThe detailed mutagenesis study of hOSM reported here, combined with earlier studies of hLIF [24], has identified those amino acid residues that contribute the majority of binding energy to the interaction with gp130, as defined by the consequences of alanine scanning analysis. This reveals a quartet of residues in the A and C helices that form a distinct patch in the crystal structure (Figure 4). In terms of the interaction between hOSM and gp130, the most important residues are Asn124 and Gly120 in helix C, with lesser contributions from Gln16 and Gln20 in helix A. The crystal structure of the CHD of gp130 [25] combined with mutagenesis data of gp130 ([33]; KRH, E Raulo, O McKenzie and JKH unpublished observations) provides a structural explanation for this interaction. The most prominent feature of gp130 is a solvent-exposed hydrophobic residue (Phe169) in an interstrand loop of the N-terminal half of the bipartite CHD (Figure 5). The second significant determinant of gp130 recognition is a prominent solvent-exposed tyrosine residue (Tyr196), located close to the junction between the two elements of the CHD, the hydroxyl group of which faces away from the main body of the molecule into solution (Figure 5). Inspection of the two functional epitopes in open book format (Figure 5) immediately suggests an explanation for the molecular mechanism of ligand recognition. This indicates that the exposed phenylalanine residue on gp130 may be docked into the hydrophobic 'dimple' formed on the hOSM surface from the C\u03b1 backbone of Gly120. In this configuration the hydroxyl group of gp130 Tyr196 would be able to hydrogen bond with the exposed amide group of Asn124 from hOSM. This explanation is consistent with mutagenesis of hOSM Gly120 reported here in which substitution to alanine, which has a relatively small non-polar sidechain, has less effect on binding than substitution to the bulkier tyrosine residue (Table 3). The substitution of a bulky sidechain would be expected to occlude the non-polar dimple and thereby inhibit docking of the Phe169 of gp130 and formation of the hydrogen bond between the Tyr196 of gp130 and the Asn124 of hOSM.\n\nThis proposed mode of docking between hOSM site 2 and gp130 has strong similarities to the interaction between growth hormone and the growth hormone receptor [23] and EPO/EPO receptor [28]. In both these cases the site 2 interaction also involves burying a bulky, solvent exposed, aromatic residue in a hydrophobic patch formed from the C\u03b1 backbone of helix C. This suggests that the basic site 2 recognition mechanism may exhibit strong structural similarities across the whole cytokine/receptor family. The point of divergence between gp130 and these other systems appears to arise from the use of polar interactions, for example the putative Tyr100\u2013Asn124 hydrogen bond, that contribute additional specificity to complex formation. Thus a major contribution to binding affinity from a distinctive hydrophobic interaction is conserved whilst specificity is controlled through polar interactions. A similar pattern of conservation has recently been noted for a second set of cytokine\u2013receptor interactions, those of the tumour necrosis factor (TNF) like cytokines with their cognate receptors [34].", "# boltzmaker.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n## boltzmaker.mdeller.com/\n\n[Two steps\n\n### Fully Automated Mode\n\nConfigure once, run one command at home, explore the results here.\n\n1. **Prepare.** Answer questions about your proteins, partners and ligands, then download a\n   single bundle carrying the software, the pinned environment and the scripts.\n2. **Analysis.** Run the bundle on your own machine, upload the one results file it writes,\n   and explore your campaign interactively.\n\nBest if you want the whole pipeline handled for you.](/auto)\n[Four tools\n\n### Stepwise Mode\n\nDrive each stage yourself, one at a time.\n\n1. **Wizard.** Build a `boltz_input.md` by answering plain questions.\n2. **Generate.** Turn that spec into the per-target YAML files Boltz-2 needs.\n3. **Preflight.** Validate chemistry, chain ids and inputs before spending GPU time.\n4. **Analyze.** Upload a finished campaign folder for a summary and dashboard.\n\nBest if you already have your own setup and want one piece of it.](/stepwise)\n\n## What it does\n\nOne spec file describes a campaign: proteins, the chains they fold\nwith, and the ligands to cross them against. BoltzMaker writes the inputs, checks the\nchemistry before any GPU time is spent, runs Boltz-2, and turns the results into a report\nyou can read.\n\n## Latest runs\n\n* ABL1_KD\n\n  2026-08-23 \u00b7 5 targets \u00b7 [explore](/runs/6wkagqkLMp2sOzC45ncJdQ/explore)\n* ABL_KINASE\n\n  2026-08-23 \u00b7 5 targets \u00b7 [explore](/runs/Cnpwvr1wlswliV3zJ4KuNA/explore)\n* GLP1R_GIPR_pocket_matrix\n\n  2026-08-22 \u00b7 20 targets \u00b7 [explore](/runs/niSiE5MfGFyddSjGLmeYBA/explore)\n* 5HT2A_GQ\n\n  2026-08-13 \u00b7 2 targets \u00b7 [explore](/runs/d96nzlk1sj5JzfkNCumscw/explore)\n* 5ht2_gq\n\n  2026-07-11 \u00b7 15 targets \u00b7 [explore](/runs/TaRE0NcV9BOpuguKRujt5g/explore)\n\n[All runs](/runs)\n\n### Where a protein family is known\n\nA prediction tells you where the atoms went, not whether the receptor\nmoved. Given an apo structure to compare against \u2014 experimental, or a ligand-free one\nit predicts for you \u2014 BoltzMaker measures that shift against the motifs the family is\nactually described by.\n\n* **GPCRs.** GPCRdb generic (Ballesteros-Weinstein) numbering, so the shift is reported\n  per **TM1\u2013TM7 and H8** and per **ICL1\u20133 and ECL1\u20133** rather than as\n  one number for the chain. TM6, TM7 and ECL2 are marked binding-site adjacent, which is\n  where an agonist-driven change shows up.\n* **Kinases.** KLIFS pocket numbering picks out the **catalytic lysine, \u03b1C\n  glutamate, gatekeeper, hinge, HRD catalytic loop and DFG motif**, and reports whether\n  **DFG** and **\u03b1C** changed between apo and holo \u2014 a coarse\n  C\u03b1\u2013C\u03b1 distance proxy for detecting a shift, not a publication-grade\n  dihedral classifier.\n* **Anything else.** Pfam domains from PDBe's SIFTS residue mapping, so a protein\n  outside those two families still gets its shift per domain instead of per chain.\n* **Whatever the family.** Real protein\u2013ligand interaction fingerprints from\n  PLIP, binding affinity as pIC50 where you ask for it, and ligand chemistry checked for\n  undefined stereocentres, ambiguous protonation and stray salts before a campaign starts\n  rather than after.", "# Decoding the Molecular Machinery Behind Plastic Degradation\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nFollowing nick generation, both fragments undergo sequential degradation through iterative cycles of binding and cleavage. For HE-terminal PET chains (\u1d34\u1d31PET), the terminal MHET unit occupies subsite I while the subsequent three MHET units fill subsite II. The catalytic machinery cleaves the ester bond adjacent to subsite I, releasing one free MHET monomer and producing a shortened chain (\u1d34\u1d31PET\u2099\u208b\u2081). Subsequent digestion of this shortened chain occurs through the same binding and cleavage pattern. TPA-terminal chains (\u1d40\u1d3e\u1d2cPET) undergo a parallel process: the terminal TPA unit positions in subsite I alongside three MHET moieties in subsite II. Ester bond cleavage releases one TPA molecule and produces an HE-terminal chain (\u1d34\u1d31PET\u2099\u208b\u2081) that continues undergoing the standard HE-terminal digestion pathway. Through repeated cycles of substrate binding and ester bond hydrolysis, both polymer types are systematically reduced to monomeric products.\u200b\n\n**Catalytic Chemistry: Acylation and Deacylation**\n\nThe molecular detail of catalysis occurs through classical two-step serine hydrolase chemistry. During the acylation phase, catalytic His237 activates nucleophilic Ser160, facilitating its attack on the substrate\u2019s ester carbonyl carbon. This generates a tetrahedral intermediate, which is stabilized by the oxyanion hole composed of backbone nitrogen atoms from Tyr87 and Met161. Proton transfer from histidine to the leaving group facilitates ester bond cleavage, generating a covalent acyl-enzyme intermediate. In the deacylation phase, a catalytic water molecule attacks the acyl-enzyme intermediate, again forming a tetrahedral species. Computational studies suggest that deacylation is the rate-limiting step, incorporating elements of nucleophilic attack, ester bond cleavage, and the characteristic \u201cmoving-histidine\u201d mechanism wherein the imidazole ring rotates to shuttle protons. Final product release regenerates the free enzyme, restoring its catalytic capability for additional cycles.\u200b\n\n**Substrate Diversity Handling**\n\nThe enzyme demonstrates remarkable adaptability in handling variable substrate architectures. Substrate chains can bind to the enzyme in reverse orientations, with one or two MHET moieties occupying subsite II instead of the standard three. These alternative binding modes generate a diversity of products including 2-HE(MHET)\u2082, (MHET)\u2082 dimers, MHET monomers, and bis(2-hydroxyethyl) terephthalate (BHET). BHET undergoes further degradation into MHET and ethylene glycol (EG). Through combinatorial digestion pathways, the continuous hydrolysis of both \u1d34\u1d31PET and \u1d40\u1d3e\u1d2cPET ultimately accumulates the final monomeric products: MHET, terephthalic acid (TPA), and ethylene glycol.\u200b\n\nThe superior performance of *Is*PETase compared to other PET-degrading enzymes stems from several distinctive structural features. The enzyme possesses an additional disulfide bond between Cys203 and Cys239 near the active site, absent in related cutinases from thermophilic bacteria. This disulfide bridge contributes significantly to thermal stability, as evidenced by the dramatic decrease in melting temperature from 46.8\u00b0C to 33.6\u00b0C when these cysteines are mutated to alanine. The enzyme also contains an extended \u03b28-\u03b16 loop with three additional residues (Asn244, Ser245, Asn246) that creates a continuous substrate binding groove, allowing subsites IIb and IIc to accommodate longer PET chains. In contrast, homologous cutinases lack this extension, resulting in a blocked binding cleft that cannot effectively bind PET polymers.\u200b\n\nStructure-guided protein engineering has demonstrated the potential for enhancing PETase activity beyond its natural capabilities. Mutation of Arg280, located at the distal end of subsite IIc approximately 23 \u00c5 from the catalytic site, to alanine resulted in a 22-32% increase in PET-degrading activity. Crystal structure analysis of this R280A variant revealed an extended and less obstructed subsite IIc that better accommodates longer substrate chains. Additionally, computational studies have identified the flexible Trp185 residue as crucial for catalytic efficiency, with mutations that restrict its motion significantly reducing activity. These insights highlight the importance of both active site chemistry and substrate binding architecture in determining overall enzyme performance.\u200b\n\nBeyond fundamental mechanistic insights, structural and biochemical studies have revealed the evolutionary trajectory of PETase enzymes and identified promising candidates for future applications. Phylogenetic analysis of 69 PETase-like enzymes revealed that *Is*PETase belongs to type IIb, characterized by the presence of an additional disulfide bond, an extended loop, and conservation of key active site residues including Trp159 and Ser238. Other type IIb enzymes from bacteria within the *Burkholderiales* order, including *Acidovorax delafieldii*, *Polyangium brachysporum*, and *Burkholderiales bacterium*, are predicted to exhibit similar high PET-degrading activities based on their shared structural features.\u200b", "# deepcovalent: About DeepCovalent | Accelerate Drug Discovery Today \u2014 DeepCovalent\nAuthors: deepcovalent\nSource: https://www.deepcovalent.com, 2026\n\n## Who we are\n\n**At DeepCovalent, we believe new drug discovery can happen in****a single step****.**\n\nTraditional early-stage drug discovery is slow, costly, and constrained by repetitive trial-and-error processes. Our revolutionary [One-Step-DeNovo\u2122 (**OS3D\u2122**) GenAI Engine](/os3d) transforms this paradigm by generating optimized noval drug molecules **in one step** \u2014 cutting early-stage discovery time to less than **3 weeks**. OS3D\u2122 intelligently navigates vast, unexplored chemical space to identify high-potential drug candidates with unmatched precision and efficiency.\n\nFounded by a team of AI researchers, chemists, and drug discovery scientists, DeepCovalent is headquartered in Boston, MA, and dedicated to partnering with innovators across pharma and biotech to redefine how therapies are discovered and developed.", "# 2015 \u2013 Crystal structure of a two-subunit TrkA octameric gating ring assembly\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nModern structural biology has revealed that the functional mechanism of this assembly relies on sophisticated conformational dynamics. The gating ring undergoes coordinated movements involving \u201cdimer hinge\u201d angles and \u201chelix crossover\u201d motifs that propagate allosteric signals from the nucleotide binding sites to the associated membrane channel. The unique interfacial hydrogen bonding network observed in the TM1088 assembly, involving Arg33, Glu34, and Glu37 residues forming salt bridges across the central cavity, likely contributes to the extraordinary thermostability required for survival in T. maritima\u2019s extreme thermal environment. These structural features represent molecular adaptations that enhance protein stability while maintaining the conformational flexibility necessary for channel regulation.\n\nContemporary research from 2024-2025 has expanded our understanding of bacterial potassium homeostasis beyond simple ion transport to encompass complex metabolic integration. Studies have revealed that potassium homeostasis systems interact intimately with metabolic pathways, including cholesterol metabolism in pathogenic bacteria, where potassium levels serve as regulatory signals that coordinate cellular responses to changing nutrient conditions. This metabolic integration underscores the central role of potassium transport systems in bacterial physiology and highlights why organisms have evolved such sophisticated regulatory mechanisms. The discovery that disruption of potassium homeostasis affects bacterial responses to environmental stresses such as pH changes, osmotic pressure, and antimicrobial compounds demonstrates the far-reaching consequences of ionic dysregulation.\n\nThe structural insights provided by the TM1088 assembly have immediate implications for understanding pathogenic bacteria that retain similar two-subunit systems. M. tuberculosis, the causative agent of tuberculosis, employs a comparable two-subunit potassium transport system that has been linked to bacterial survival within host macrophages and resistance to acidic environments encountered during infection. The structural framework provided by the TM1088 studies offers potential targets for developing novel antimicrobial strategies that could disrupt bacterial potassium homeostasis and compromise pathogen survival. Recent investigations have identified specific residues critical for system function, providing molecular targets for rational drug design approaches.\n\n| Key Concept | Description | Key References |\n| --- | --- | --- |\n| **Two-Subunit Architecture** | Evolutionary precursor system with separate TM1088A and TM1088B proteins forming octameric assembly | [Deller et al., PLoS ONE, 2015](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0122512) |\n| **Nucleotide Binding Mechanism** | AMP binding at GXGXXG motifs with divalent cation cofactors for activation | [Teixeira-Duarte et al., eLife, 2019](https://elifesciences.org/articles/50661) |\n| **Allosteric Regulation** | Cooperative conformational changes propagated through dimer hinge angles | [Guan et al., National Science Review, 2022](https://academic.oup.com/nsr/article/9/8/nwac072/6567883) |\n| **Thermophilic Adaptations** | Interfacial salt bridges and enhanced protein stability for extreme environments | [Deller et al., PLoS ONE, 2015](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0122512) |\n| **Pathogenic Applications** | Therapeutic targets in M. tuberculosis and other pathogens with similar systems | [Epstein, Current Opinion in Microbiology, 2021](https://pubmed.ncbi.nlm.nih.gov/33798529/) |\n| **Metabolic Integration** | Coupling of potassium homeostasis with cholesterol metabolism and cellular responses | [Huang et al., Nature Communications, 2024](https://www.nature.com/articles/s41467-024-49082-7) |\n| **Evolutionary Pathway** | Gene duplication and fusion events transitioning to modern single-subunit systems | [Johnson et al., Journal of Bacteriology, 2009](https://journals.asm.org/doi/10.1128/jb.01367-08) |\n| **Structural Dynamics** | Conformational flexibility enabling channel gating and ion selectivity | [Guan et al., National Science Review, 2022](https://academic.oup.com/nsr/article/9/8/nwac072/6567883) |", "# Building a Comprehensive Protein Production, X-ray Crystallography, and Cryo-EM Laboratory from Scratch\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-09-09 on marcdeller.com (categories: blog, project, sb).*\n\nMy project installing a Glacios 2 in 2023/4 at Incyte: construction, RF shielding, vibration isolation, and instrument install\n\n* [**Best practice: setting up and operating a mid-sized cryo-EM facility**](https://www.frontiersin.org/journals/molecular-biosciences/articles/10.3389/fmolb.2023.1302680/full)\n\n* **Infrastructure Integration:** Unified design connects protein production to structure delivery, ensuring smooth workflows and consistent quality.\n* **Automation Excellence:** AI-driven crystallization and robotics cut manual work, accelerating speed and reproducibility.\n* **Data Management:** Powerful computing and smart pipelines turn massive cryo-EM datasets into precise structural insights.\n* **Personnel Development:** Intensive 6\u201312 month training builds skilled operators and a culture of ongoing mastery.\n* **Quality Assurance:** Standardized protocols secure reproducible results and uphold the lab\u2019s precision ethos.\n\n1. **Glacios 2 Cryo-TEM Specifications and Capabilities:** [Thermo Fisher Scientific, Electron Microscopy Products, 2024](https://www.thermofisher.com/us/en/home/electron-microscopy/products/transmission-electron-microscopes/glacios-cryo-tem.html)\n2. **Cost-Benefit Analysis of Establishing a Cryo-EM Facility:** [Eureka PatSnap, Market Research Report, 2025](https://eureka.patsnap.com/report-cost-benefit-analysis-of-establishing-a-cryo-em-facility)\n3. **Protein Crystallization Automation Systems:** [Formulatrix, Protein Crystallization Systems, 2024](https://formulatrix.com/protein-crystallization-systems/)\n4. **Best Practice Guidelines for Mid-Sized Cryo-EM Facilities:** [Frontiers in Molecular Biosciences, Van Andel Institute, 2023](https://www.frontiersin.org/journals/molecular-biosciences/articles/10.3389/fmolb.2023.1302680/full)\n5. **Automated Multi-Step Protein Purification Methods:** [Morimoto, D., et al., Protein Expression and Purification, 2024](https://pubmed.ncbi.nlm.nih.gov/39094813/)\n\n**Building a Comprehensive Protein Production, X-ray Crystallography, and Cryo-EM Laboratory from Scratch**\n\nBuilding a fully integrated structural biology laboratory has been one of my most ambitious and transformative projects. By bringing protein production, X\u2011ray crystallography, and cryo\u2011electron microscopy (cryo\u2011EM) under one roof, we erased the delays that once came from moving samples across facilities. This integration turned structural analysis into a seamless, iterative process\u2014where insights could be tested and refined in real time.\n\nAutomation lies at the heart of this revolution. The Formulatrix NT8 v4 liquid handler dispenses crystallization drops as small as 10 nL with remarkable precision, while SONICC imaging reveals crystals that no optical microscope could ever detect\u2014hidden in precipitate or mere submicron specks. Likewise, cryo\u2011EM has evolved from an artisan\u2019s craft into an automated engine of discovery. Instruments such as the Thermo Scientific Glacios\u202f2, paired with the Falcon\u202f4i detector and EPU software, now deliver near\u2011atomic clarity with minimal manual intervention. Even vitrification\u2014the once notoriously delicate step\u2014has been reimagined by systems like the VitroJet, which replace manual blotting with real\u2011time controlled sample preparation.\n\nCreating such a facility demands precision beyond the instruments themselves. Cryo\u2011EM rooms must meet ISO\u202f8 cleanroom standards, hold temperatures steady within \u00b10.3\u202f\u00b0C, humidity near 40%, and remain shielded from electromagnetic noise. The cost reflects this complexity: each advanced cryo\u2011EM setup can exceed $7\u202fmillion, often doubled by facility modifications and support infrastructure.\n\nAI\u2011driven protein prediction and machine\u2011learning\u2011guided drug discovery are reshaping biotechnology and pharmaceuticals. Establishing such a laboratory today is not just building a facility\u2014it\u2019s creating a platform for accelerated science, where automation, data integration, and human expertise converge to uncover biology\u2019s finest details.", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nR^W, the reactive group borne by Ring moiety A, is drawn in the specification as an extensive set of electrophilic (\"warhead\") motifs attached through linking groups L^1 through L^7. These include: alpha-halo (Cl, Br, F) alkyl groups; alpha-cyano alkyl; ether- and thioether-linked alkyl; alpha,beta-unsaturated (vinyl/acrylamide-type) alkenes, including those bearing C(O)R^83 and S(O)tR^83 activating groups; allenes; propargyl/alkynyl and cyanoalkynyl groups; epoxides and aziridines (X^1 is O or NR^9); butyrolactones; squarate (cyclobutenedione) groups; quinones; isoxazolines; maleimides; N-cyanopyrrolidines and related cyclic Michael acceptors; pyridinones and piperidinones bearing exocyclic alkenes; acyl and sulfonyl groups bearing leaving groups; isothiocyanates; alkyl halides (F, Cl, Br, I); and cyclopropyl- or spirocyclobutyl-sulfonyl motifs. Vinyl sulfonamide (\u2014NHS(O)2CH=CH2) and acrylamide (\u2014NHC(O)CH=CH2) groups attached through L^2 are the forms exemplified in the working examples.\n\n- L^1 is -L-C(O)\u2014, -L-NR^9C(O)\u2014, -L-OC(O)\u2014, -L-S(O)\u2014, -L-S(O)2\u2014, -L-NR^9S(O)\u2014, -L-OS(O)\u2014, -L-NR^9S(O)2\u2014, or -L-OS(O)2\u2014, wherein L^1 is attached to Ring moiety A through the L linking group;\n- L^2 is -L-, -L-O\u2014, -L-NR^9\u2014, -L-S\u2014, -L-C(O)\u2014, -L-NR^9C(O)\u2014, -L-OC(O)\u2014, -L-S(O)\u2014, -L-S(O)2\u2014, -L-NR^9S(O)\u2014, -L-OS(O)\u2014, -L-NR^9S(O)NR^9\u2014, -L-NR^9S(O)O\u2014, -L-OS(O)NR^9-L^x-, -L-OS(O)2\u2014, -L-NR^9S(O)2O\u2014, -L-S(O)2(NR^9)\u2014, or -L-OS(O)2NR^9\u2014, wherein L^2 is attached to Ring moiety A through the L linking group;\n- L^3 is -L-, -L-C(O)\u2014, -L-NR^9C(O)\u2014, -L-OC(O)\u2014, -L-S(O)\u2014, -L-S(O)2\u2014, -L-NR^9S(O)\u2014, -L-OS(O)\u2014, -L-NR^9S(O)2\u2014, or -L-OS(O)2\u2014;\n- L^4 is -L-, -L-O\u2014, L-S\u2014, or -L-NR^9\u2014;\n- L^5 is a set of linkers of the form -L-O-L^x-, -L-NR^9-L^x-, -L-S-L^x-, -L-C(O)-L^x-, and the corresponding carbamate, carbonate, sulfinyl, sulfonyl, sulfonamide and sulfamate variants;\n- L^6 is -L-, -L-O\u2014, -L-NR^9\u2014, -L-S\u2014, -L-C(O)\u2014, and the corresponding carbamate, carbonate, sulfinyl, sulfonyl and sulfonamide variants;\n- L^7 is -L-, -L-O\u2014, L-S\u2014, or -L-NR^9\u2014;\n- each L is independently a bond or C1\u20136 alkylene, optionally substituted by 1, 2, 3 or 4 independently selected R^G substituents; or each L is \u2014O\u2014C1\u20136 alkyl or \u2014N(R^N)\u2014C1\u20136 alkyl, wherein L is attached to Ring moiety A via the oxygen of the \u2014O\u2014C1\u20136 alkyl or the nitrogen atom of the \u2014N(R^N)\u2014C1\u20136 alkyl group;\n- each R^N is independently H or C1\u20136 alkyl;\n- each L^x is independently a C1\u20136 alkylene, optionally substituted by 1, 2, 3 or 4 independently selected R^G substituents;\n- Ring D is a 4\u201312 membered heterocycloalkyl, C3\u201312 cycloalkyl, C6\u201310 aryl, or a 5\u201310 membered heteroaryl, each optionally substituted with 1, 2, 3 or 4 independently selected C1\u20136 alkyl groups;\n- X^1 is O or NR^9;\n- each q is independently 0, 1, 2, or 3; each t is independently 0, 1, or 2; each u is independently 0, 1, 2, or 3;\n- Ar is C6\u201310 aryl or 5\u201310 membered heteroaryl, each optionally substituted with 1, 2, 3, or 4 independently selected R^9 substituents.\n\n### Preferred embodiments\n\nIn successive sets of narrower embodiments the specification restricts the variables as follows. In some embodiments m is an integer selected from 0, 1, 2, 3, 4, or 5, then from 0, 1, 2, and 3, then from 0, 1, and 2; n is correspondingly narrowed from 0\u20135 to 0\u20133 and then to 0, 1, and 2. In some embodiments X is C and Y is C. In some embodiments Z is N; in other embodiments Z is CR^7.\n\nIn some embodiments, Ring moiety A is selected from C3\u201310 cycloalkyl, 6\u201310 membered aryl, 4\u201310 membered heterocycloalkyl, and 5\u201310 membered heteroaryl. In some embodiments, Ring moiety A is selected from C3\u20137 cycloalkyl, phenyl, 4\u20136 membered heterocycloalkyl, and 5\u20136 membered heteroaryl. In some embodiments, Ring moiety A is C3\u20137 cycloalkyl. In some embodiments, Ring moiety A is 6-membered aryl. In some embodiments, Ring moiety A is 4\u201310 membered heterocycloalkyl. In some embodiments, Ring moiety A is 5\u201310 membered heteroaryl. In some embodiments, Ring moiety A is 5-6 membered heteroaryl. In some embodiments, Ring moiety A is pyrazole. In some embodiments, Ring moiety A is phenyl.\n\nIn some embodiments, Ring moiety B is selected from C3\u20137 cycloalkyl, phenyl and 5\u20136 membered heteroaryl. In some embodiments, Ring moiety B is selected from C5\u20136 cycloalkyl, phenyl and 5\u20136 membered heteroaryl. In some embodiments, Ring moiety B is selected from C6 cycloalkyl, phenyl and 6-membered heteroaryl. In some embodiments, Ring moiety B is selected from C6 cycloalkyl, phenyl, pyridine, pyrimidine, and pyrazine. In some embodiments, Ring moiety B is 5\u20136 membered heteroaryl; in some embodiments 6-membered heteroaryl; in some embodiments selected from pyridine, pyrimidine, and pyrazine. In some embodiments, Ring moiety B is C3\u20137 cycloalkyl; in some embodiments C5\u20136 cycloalkyl; in some embodiments C6 cycloalkyl. In some embodiments, Ring moiety B is 5\u20136 membered heterocycloalkyl selected from morpholine, tetrahydro-2H-pyran, piperidine and piperazine. In some embodiments, Ring moiety B is pyridine.\n\nIn some embodiments, R^1 is independently selected from H, D, halo, CN, C1\u20136 alkyl, C1\u20136 haloalkyl, C1\u20136 haloalkoxy, and C1\u20136 alkoxy. In some embodiments, R^1 is independently selected from C1\u20136 alkyl and C1\u20136 alkoxy. In some embodiments, R^1 is independently selected from Me and OMe. In some embodiments, R^1 is Me. In some embodiments, R^1 is OMe. In some embodiments, m is 0; in some embodiments m is 1; in some embodiments m is 2. In some embodiments, n is 0; in some embodiments n is 1.", "# GitHub README: TopPDBLX\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/TopPDBLX, 2026\n\n| Round | What changed | Identification | Grounding | Gold sets: precision / recall | Residual components generated |\n|---|---|---|---|---|---|\n| [01](https://huggingface.co/Dellboy/toppdblx-residual-parser/tree/main/round01) | Bootstrap distillation from rule output, lexicon 0.1.0 | 87.0% \u2020 | not yet invented | - | N/A \u2021 |\n| [02](https://huggingface.co/Dellboy/toppdblx-residual-parser/tree/main/round02) | `not_a_component` class added, confidence gate fixed | 89.7% \u2020 | not yet invented | - | N/A \u2021 |\n| [03](https://huggingface.co/Dellboy/toppdblx-residual-parser/tree/main/round03) | Class rebalanced; cosine decay and dropout introduced | 88.4% \u2020 | not yet invented | - | N/A \u2021 |\n| [04](https://huggingface.co/Dellboy/toppdblx-residual-parser/tree/main/round04) | Retrained on the 502-reagent lexicon | abandoned (36% duplicate training rows) | - | - | N/A \u2021 |\n| [05](https://huggingface.co/Dellboy/toppdblx-residual-parser/tree/main/round05) | Deduplicated training set; first round scored on a frozen benchmark | 87.58% | 93.41% | - | N/A \u2021 |\n| [**06**](https://huggingface.co/Dellboy/toppdblx-residual-parser/tree/main/round06) | Full epoch, rank 16, 32 LoRA layers, empty-answer examples added | 89.05% | 93.15% | 95.3% / 85.7% | **153,736** |\n| [07](https://huggingface.co/Dellboy/toppdblx-residual-parser/tree/main/round07) | 32B-teacher labels *replacing* rules labels, 16 LoRA layers | 92.39% | 92.65% | superseded by round 09, see below | N/A \u2021 |\n| [08](https://huggingface.co/Dellboy/toppdblx-residual-parser/tree/main/round08) | Same idea, labels filtered for precision, trained 4x longer, still 16 LoRA layers | 93.3% | 93.53% | superseded by round 09, see below | N/A \u2021 |\n| [**09**](https://huggingface.co/Dellboy/toppdblx-residual-parser/tree/main/round09) | **Shipped 2026-08-04.** Teacher labels *added* to the rules labels, not substituted; `protein_buffer`/`soak` scope roles; prompt v2; 32 LoRA layers, matching round 06 | **95.28%** | **94.83%** | 93.3% / **90.3%** | **188,655** \u00b6 |\n\n\u2020 Rounds 01\u201303 were scored against a live residual that shrank as curation improved, so these three figures are not comparable to each other or to later rounds. The comparable, frozen-benchmark story starts at round 05.\n\n**Read down the identification and grounding columns from round 05: every round improves on the last.** 87.58, 89.05, 92.39, 93.3, 95.28 on identification; 93.41, 93.15, 92.65, 93.53, 94.83 on grounding; and 62.85 to 81.0 on fully-identified records between rounds 06 and 09. All measured on the same 2,000 held-out records against the same lexicon 0.9.2, so the comparison is between models rather than dictionaries.\n\n\u2021 N/A rather than 0: the column counts components the round *generated* over the residual, not rows in the released component table (see [Output](#-output)). Only rounds 06 and 09 have ever been run over the corpus at scale; every other adapter was a training experiment, measured and set aside, and was never asked to read the residual.\n\n\u00b6 Generated 2026-08-04, a 5.1-hour pass over all 47,845 residual records: **188,655 components against round 06's 153,736, a 22.7% increase.** **35,580 of those are chemistry the rules never named, and they are now in the released component table**, tagged `parser = slm`. Classification and screen matching were rebuilt from the same output.\n\n### Round 09: why it ships over round 06\n\nRound 09 is the first clean test of whether teacher distillation helps when the teacher's labels are *added* to the rule parser's labels rather than replacing them, at matched LoRA capacity. It does: pooled recall against hand-labelled truth rises from round 06's 85.7% to 90.3%, finding 28 more reagents and missing 32% fewer, for a rise in role-misattribution errors rather than invented chemistry (see [Results against hand-labelled truth](#-results-against-hand-labelled-truth)).\n\nThree things changed together: every rules label was kept and roughly 3,400 teacher-labelled residual rows were added on top; the run used 32 LoRA layers, matching round 06 (rounds 07 and 08 had trained at 16, since `--num-layers` was left at its default); and the prompt gained `protein_buffer` and `soak` roles so a correctly-identified reagent in the wrong role has somewhere to go. A build-time gate fails the run rather than training on a leak: zero gold records reached training, verified after the fact.\n\nRound 09 also trains for longer than round 06 (8,000 iterations against 6,000), and the frozen benchmark shows that this matters on its own: identification, grounding, fully-identified records and fidelity all climb monotonically from iteration 1,000 to 8,000. **The shipped model is therefore round 09's final adapter**, chosen on the 2,000-record benchmark rather than on the 192-record gold sets, which were too small to separate checkpoints that differ by 7 points of fully-identified records.\n\n## \ud83e\uddec Where to cut: the construct boundary model\n\n**The practical question.** You have a protein that will not crystallise. Almost nobody\ncrystallises the full-length gene product: they trim the flexible tails, drop a disordered linker,\nand try the folded core. Choosing where to cut is judgement, and it is usually made once, by hand,\nfrom a disorder plot and an alignment.\n\nThe Protein Data Bank already contains **523,018 of those decisions**, made by crystallographers\nwho then got a structure. Every deposited chain records exactly which stretch of the full-length\nprotein was cloned, and SIFTS maps it back residue by residue. Those are free labels: nobody had\nto annotate anything, the experiment did it.\n\n**What the model does.** Give it a full-length sequence. It returns, for every residue, the\nprobability that a crystallographer would have kept that residue in the construct. High in the\nmiddle, falling away at the ends it thinks you should trim.", "# Transforming Antibody Design: How Artificial Intelligence Revolutionizes Structure Prediction\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nGraph neural networks have emerged as powerful tools for predicting antibody-antigen binding interactions and enabling structure-guided antibody optimization. [GearBind](https://www.nature.com/articles/s41467-024-51563-8) exemplifies this approach through multi-relational graph construction, multi-level geometric message passing, and contrastive pretraining on large-scale protein structural data. This architecture captures atomistic-level features at antibody-antigen interfaces, achieving state-of-the-art performance with a Pearson correlation of 0.707 for high-magnitude binding free energy changes. By successfully identifying mutations that enhance antibody binding affinity, GearBind demonstrated the practical utility of geometric deep learning for therapeutic antibody development. Similar graph-based approaches like [Paragraph](https://academic.oup.com/bioinformatics/article/39/1/btac732/6825310) leverage equivariant graph neural networks for accurate paratope prediction, achieving superior performance using simple feature vectors without requiring antigen information.\n\nThe development of antibody-specific models through transfer learning and fine-tuning has yielded significant performance gains. [AntiFold](https://academic.oup.com/bioinformaticsadvances/article/doi/10.1093/bioadv/vbae202/8090019), an inverse folding model fine-tuned from ESM-IF1 on solved and predicted antibody structures, outperforms existing tools in complementarity-determining region sequence recovery while maintaining high structural similarity to experimentally determined counterparts. When [systematically benchmarked](https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0324566) against ProteinMPNN, ESM-IF, and LM-Design, AntiFold demonstrated superior performance for Fab antibody CDR design, particularly excelling at capturing conserved sequence patterns in structurally constrained regions like CDR-H1 and the CDR-H3 stem. This specialization highlights the importance of domain-specific training data for achieving optimal predictive performance in antibody engineering applications.\n\nValidation metrics have evolved to provide comprehensive assessment of [antibody structure prediction quality](https://www.sciencedirect.com/science/article/pii/S1359644624001508), encompassing structural accuracy measures, binding affinity predictions, and developability assessments. Standard evaluation includes calculating backbone heavy-atom root mean square deviation for framework regions and individual CDR loops, with particular emphasis on the challenging CDR-H3 loop that typically exhibits the highest variability. Interface-specific metrics such as interface RMSD, ligand RMSD, and fraction of native contacts provide quantitative measures of antibody-antigen docking accuracy according to [CAPRI criteria](https://pmc.ncbi.nlm.nih.gov/articles/PMC10751731/). For classification tasks involving binding prediction or developability assessment, researchers employ balanced accuracy, precision, recall, F1 score, and receiver operating characteristic area under curve to evaluate model performance across imbalanced datasets common in antibody engineering workflows.\n\nThe integration of artificial intelligence methodologies with experimental validation workflows represents a critical frontier in translating computational predictions to functional therapeutic candidates. Recent studies have demonstrated that [AI-driven pipelines](https://pmc.ncbi.nlm.nih.gov/articles/PMC11839053/) combining AlphaFold2 for structure prediction, ProteinMPNN for sequence design, and high-throughput experimental screening can successfully optimize antibody framework regions while preserving epitope-binding complementarity-determining regions. This synergistic approach accelerates the identification of antibodies with improved binding affinity, enhanced stability, and favorable developability profiles. As datasets continue expanding through next-generation sequencing of antibody repertoires and high-throughput structural characterization, machine learning models will become increasingly accurate and broadly applicable across diverse antigen targets and antibody formats.", "# Clearing the Bloodstream\u2014Diagnostics and Approaches for Microplastic Removal\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-15 on marcdeller.com (categories: blog, microplastics).*\n\n* PETase breaks down microplastics in serum at 37\u00b0C, proven non-toxic in cell models.\n* HPLC and NMR methods monitor microplastic degradation and clearance.\n* Blood filtration (apheresis) physically removes particles from circulation.\n* Combining enzymatic and filtration methods increases therapeutic efficacy.\n* Personalized protocols will enhance future clinical management of microplastic burden.\n\n1. [PETase review (2023)](https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2023.1263996/full)\n2. [PETase serum paper (2024)](https://pubs.rsc.org/en/content/articlepdf/2024/fd/d4fd00014e)\n3. [Micro(nano)plastics in human health (2024)](https://pmc.ncbi.nlm.nih.gov/articles/PMC11528458/)\n4. [Blood filter for microplastics (2025)](https://nypost.com/2025/07/05/health/this-blood-filter-may-flush-microplastics-from-our-bodies/)\n5. [Therapeutic apheresis for microplastics (2025)](https://pmc.ncbi.nlm.nih.gov/articles/PMC12162106/)\n\nRecent studies confirm microplastics circulate in human blood and serum, raising urgent questions about health and therapeutics. PETase has emerged as a biocatalyst capable of depolymerizing PET into monomers, such as terephthalic acid (TPA) and MHET, at body temperature without harming human cells in vitro. Researchers use pre-treated PET substrates mimicking microplastics, demonstrating that PETase (especially S238A variant) is effective in breaking down these particles under serum-like conditions.\n\nDiagnostics such as high-performance liquid chromatography (HPLC) and NMR spectroscopy track both microplastics and their breakdown products, vital for clinical monitoring. Physical filtration methods, including therapeutic apheresis, show promise for directly removing plastic particles from blood.\n\nThe synergy of enzyme therapy and filtration\u2014administering PETase to degrade microplastics, then filtering out residual particles\u2014could significantly lower the plastic burden in patients. Delivery challenges and kinetics remain, but diagnostic and removal protocols are now in development for future translational use.", "# GitHub README: chatRDKit\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/chatRDKit, 2026\n\nWhenever a request involves drawing/visualising a molecule or generating a\nreport, chatRDKit calls a built-in `render(mol, name=\"...\")` helper, which\nprompts you for the output format:\n\n```\n[chatRDKit] Output format? (1) PNG  (2) HTML  (3) PDF  [1]:\n```\n\n- **PNG** -- a depiction image (single molecule or a grid for several)\n- **HTML** -- a standalone report with the image, SMILES, and formula\n- **PDF** -- the same report as a PDF (via `reportlab`; falls back to PNG\n  with a warning if `reportlab` isn't installed)\n\nFiles are written to the current directory by default. Set `output_dir` to\na path to always use that directory, or to `ask` to be prompted for a\ndirectory each time (configurable, see below).\n\n---\n\n## 3. Configuration\n\nAll settings are stored in `~/.chatrdkit/config.json` and can be viewed/\nchanged with `/config`:\n\n```\nchatRDKit> /config                          # list everything\nchatRDKit> /config model                    # show one value\nchatRDKit> /config model qwen3:0.6b         # change the model\nchatRDKit> /config host http://localhost:11434\nchatRDKit> /config temperature 0.1\nchatRDKit> /config output_format png        # always PNG, never ask\nchatRDKit> /config output_dir ~/Desktop/chatrdkit_out\nchatRDKit> /config debug on                 # print raw LLM output (troubleshooting)\n```\n\n| Setting        | Default                    | Meaning                                                          |\n|----------------|----------------------------|-------------------------------------------------------------------|\n| `host`         | `http://localhost:11434`  | Ollama server URL                                                 |\n| `model`        | `qwen3.6:27b`              | Ollama model tag (must be pulled, must match `ollama list`)       |\n| `temperature`  | `0.2`                      | LLM sampling temperature                                          |\n| `timeout`      | `180`                      | Request timeout (seconds)                                         |\n| `echo`         | `True`                     | Print each translated line before running it                     |\n| `debug`        | `False`                    | Print the raw LLM response (before cleanup) to the console       |\n| `max_history`  | `6`                        | Number of past request/response turns kept for follow-ups        |\n| `output_dir`   | `.` (current directory)   | Where `render()` saves files -- a path, or `ask` to prompt each time |\n| `output_format`| `ask`                      | `ask`, `png`, `html`, or `pdf` -- set to skip the format prompt   |\n\n---\n\n## 4. How it works\n\nchatRDKit is a single file, `chatrdkit.py`, organised into sections:\n\n```\nchatrdkit.py\n\u251c\u2500\u2500 config       JSON-backed settings (~/.chatrdkit/config.json)\n\u251c\u2500\u2500 ollama_client minimal urllib client for POST /api/chat\n\u251c\u2500\u2500 rdkit_reference system prompt + curated RDKit cheat-sheet + session context\n\u251c\u2500\u2500 render       PNG/HTML/PDF output via Draw.MolsToGridImage\n\u251c\u2500\u2500 session      persistent exec namespace (RDKit imports + render())\n\u251c\u2500\u2500 router       SMILES/code/English classification, translation, execution\n\u2514\u2500\u2500 cli          welcome banner, REPL loop, slash commands\n```\n\n1. **Session.** A single Python namespace (dict) is created at startup with\n   `Chem`, `AllChem`, `Draw`, `Descriptors`, `Lipinski`, `Crippen`, `QED`,\n   `DataStructs`, `rdFingerprintGenerator`, `rdMolDescriptors`, `rdMolDraw2D`,\n   `math`, and a `render()`\n   helper already in scope. Every molecule you create persists here for the\n   rest of the session.\n\n2. **Classifying input** (`router.classify`):\n   - `looks_like_smiles()` -- no spaces, only SMILES-valid characters, and\n     `Chem.MolFromSmiles()` parses it successfully.\n   - `looks_like_code()` -- valid Python (`ast.parse`) that contains an\n     assignment, call, attribute access, control flow, etc., or is a bare\n     name that already exists in the session (e.g. typing `mol`).\n   - Anything else is treated as English.\n\n3. **Translation** (`router.translate`). English requests are sent to\n   Ollama with:\n   - a system prompt describing the exec environment, the `render()`\n     helper, and a curated RDKit cheat-sheet (`rdkit_reference.py`,\n     drawn from the [Getting Started with RDKit in\n     Python](https://www.rdkit.org/docs/GettingStartedInPython.html) guide),\n   - `rdkit_reference.build_context()` -- a live list of every molecule\n     currently in the session and its variable name, so \"it\"/\"the\n     molecule\"/a compound name resolves correctly,\n   - the last few turns of conversation for follow-ups.\n\n   The model is instructed to return **only** a Python/RDKit snippet (no\n   markdown, no explanations), reuse existing session variables, create\n   well-known compounds from SMILES using its own chemistry knowledge, and\n   call `render()` for anything visual. Generation is capped\n   (`num_predict`) and any response that comes back too long or repetitive\n   is discarded -- and not added to history -- rather than executed.\n\n4. **Execution** (`router.execute_code` / `session.run`). Each snippet is\n   echoed as `[chatRDKit] >>> ...` and run in the session namespace. A\n   single bare expression's value is printed (REPL-style); otherwise normal\n   `print()` output from the snippet appears directly. Errors are caught and\n   reported without crashing the session.\n\n5. **Rendering** (`render.render`). Computes 2D coordinates if missing,\n   builds a grid image with `Draw.MolsToGridImage`, asks for a format if not\n   configured, and writes PNG/HTML/PDF to `output_dir`.\n\n---\n\n## 5. Safety notes", "# Revolutionary AI Tool XenoBug Transforms Environmental Pollution Control Through Predictive Enzyme Discovery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nValidation studies demonstrate XenoBug\u2019s exceptional utility across diverse pollutant classes. The tool successfully predicted known enzymes for compounds like p-xylene, where it identified monooxygenases from Pseudomonas, Halopseudomonas, and Xenorhabdus genera capable of degrading this soil and groundwater pollutant used in paint and plastic production. For triazophos, an extremely toxic organophosphorus pesticide, XenoBug correctly predicted hydrolase enzymes from Ochrobacterium species. Perhaps more significantly, the platform predicted potential degrading enzymes for pollutants where bioremediating enzymes remain unknown, such as toxaphene, aldrin, dieldrin, and DDT, providing valuable leads for experimental confirmation.\n\nThe tool\u2019s impact extends beyond academic research into practical environmental applications. Municipal authorities are already exploring XenoBug-derived enzyme cocktails for landfill leachate treatment, agricultural runoff management, and contaminated aquifer restoration. The platform\u2019s ability to identify enzyme sources from specific metagenomic environments\u2014including marine metagenomes, compost metagenomes, and soil metagenomes\u2014enables targeted deployment strategies tailored to local environmental conditions.\n\nRecent developments in AI-driven environmental biotechnology complement XenoBug\u2019s capabilities, with parallel research advancing enzyme engineering for microplastic degradation and the discovery of bacterial species capable of destroying \u201cforever chemicals\u201d like PFAS compounds. These synergistic developments suggest a rapidly evolving landscape where artificial intelligence accelerates the discovery and optimization of biological solutions to environmental challenges.\n\nThe accessibility and user-friendly design of XenoBug democratize advanced environmental biotechnology research. Available as a free web-based platform at <https://metabiosys.iiserb.ac.in/xenobug>, the tool allows researchers worldwide to input pollutant information and receive detailed predictions about degrading enzymes, their environmental sources, biochemical pathways, and protein sequences. This open-access approach accelerates global research collaboration and enables rapid response to emerging contamination challenges.\n\nLooking forward, XenoBug represents just the beginning of AI\u2019s transformation of environmental remediation. The platform\u2019s modular architecture facilitates future updates based on expanding databases and improved algorithms. Integration with advanced technologies like natural language processing for SMILES data analysis and 3D cheminformatic features could further enhance prediction accuracy. The tool\u2019s success validates the potential of machine learning to unlock nature\u2019s vast repository of pollutant-degrading capabilities, offering hope for more effective, targeted, and scalable environmental cleanup strategies.\n\nAs environmental challenges intensify globally, XenoBug stands as a testament to the power of artificial intelligence in addressing humanity\u2019s most pressing ecological problems. By bridging the gap between computational prediction and biological reality, this innovative platform accelerates the discovery of nature\u2019s own solutions to pollution, potentially revolutionizing how we approach environmental restoration in the decades ahead.", "# The Protein Architect: How Generative AI Is Building the Drugs of Tomorrow\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* Generative AI has evolved from predicting existing protein structures to designing entirely new ones from scratch, a major leap for protein engineering.\n* AI models like Genie can create custom proteins with novel functions, opening the door for highly specific and effective therapeutic solutions.\n* This technology is accelerating drug discovery by enabling the rapid design of new molecules, moving from an incremental to a revolutionary approach.\n\n1. **AI in Pharma and Biotech:** Market Trends 2025 and Beyond:\n   <https://www.coherentsolutions.com/insights/artificial-intelligence-in-pharmaceuticals-and-biotechnology-current-trends-and-innovations>\n2. **Folding the Future:** How AI is Reshaping Protein Engineering:\n   <https://www.synbiobeta.com/read/folding-the-future-how-ai-is-reshaping-protein-engineering>\n3. **Protein Engineering Market Analysis 2025:** Highlighting Advanced Design Tools:\n   <https://www.thebusinessresearchcompany.com/market-insights/protein-engineering-market-insights-2025>\n\nImagine trying to build a new, complex machine with a limited set of existing parts. For years, this was the reality for protein engineering. Scientists worked diligently to tweak and modify naturally occurring proteins, hoping to coax them into performing new therapeutic functions. This process, while successful, has been slow and incremental, often limited by the blueprints nature provided. The challenge of creating entirely new proteins\u2014designed from scratch with a specific purpose in mind\u2014remained a distant dream. This limitation meant that many diseases with complex mechanisms were left without effective, custom-built solutions.\n\nThe arrival of generative AI has fundamentally changed the blueprint. Models like AlphaFold and Genie are doing for protein engineering what generative AI has done for art and language: creating entirely novel works from the ground up. AlphaFold began by accurately predicting protein structures from amino acid sequences, a breakthrough that accelerated biological research worldwide. But newer models now go a step further, designing completely new proteins that don\u2019t exist in nature. This shift from prediction to creation allows us to move beyond mere modification and into the realm of true *de novo* design. It\u2019s the difference between restoring an old building and designing a skyscraper from scratch.\n\nThis revolution is already yielding tangible results. We can now design enzymes that are stable at high temperatures for industrial processes or engineer proteins with unparalleled specificity to target cancer cells while sparing healthy ones. By using AI to design custom proteins, researchers are accelerating drug discovery and creating therapies tailored to individual diseases. This technology is not just speeding up the old way of doing things; it\u2019s enabling a new paradigm of rational, function-driven design that promises to solve some of the 21st century\u2019s most pressing health challenges.", "# 2025 \u2013 Imidaz0[1,2-a]pyrimidine-5(1 h)-ones as wrn inhibitors\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-09-09 on marcdeller.com (categories: patents).*\n\n* [Link to Google Patent](https://patents.google.com/patent/US20250270218A1/en)\n\n* **Selective Targeting**: Imidazo[1,2-a]pyrimidine-5(1H)-ones demonstrate exceptional selectivity for MSI-high cancer cells while sparing normal tissues\n* **Mechanistic Innovation**: WRN helicase inhibition represents a novel synthetic lethal approach distinct from traditional DNA damage agents\n* **Clinical Promise**: Multiple WRN inhibitors have entered clinical trials with encouraging early efficacy signals in refractory MSI-high cancers\n* **Therapeutic Window**: The differential WRN dependency between MSI-high tumors and normal cells creates an optimal safety profile\n\n1. **Discovery of WRN inhibitor HRO761 with synthetic lethality in MSI cancers:** [Behan, F.M., et al., Nature, 2024](https://www.nature.com/articles/s41586-024-07350-y)\n2. **Targeting the Werner syndrome protein in microsatellite instability cancers:** [Liu, X., et al., PMC, 2025](https://pmc.ncbi.nlm.nih.gov/articles/PMC12328481/)\n3. **Clinical prospects of WRN inhibition as a treatment for MSI tumours:** [Morales-Juarez, D.A., et al., NPJ Precision Oncology, 2022](https://pmc.ncbi.nlm.nih.gov/articles/PMC9666358/)\n\n**Pioneering WRN Inhibition in Precision Cancer Therapy**\n\nThe therapeutic landscape of microsatellite instability-high cancers has undergone a revolutionary transformation with the emergence of Werner syndrome helicase inhibitors. This paradigm shift demonstrates how strategic molecular targeting can exploit cancer-specific vulnerabilities while preserving normal cellular function.\n\nRecent patent developments have illuminated the potential of imidazo[1,2-a]pyrimidine-5(1H)-one derivatives as selective WRN inhibitors, representing a sophisticated approach to synthetic lethal therapy. These bicyclic heterocyclic compounds exhibit remarkable selectivity for microsatellite instability-high tumor cells, achieving potent cytotoxic effects through targeted helicase disruption. The molecular architecture of these compounds enables precise binding to the WRN helicase interface, effectively locking the enzyme in an inactive conformation and preventing essential DNA repair processes.\n\nThe therapeutic rationale centers on exploiting the unique dependency of MSI-high cancers on WRN helicase function. In normal cells with intact mismatch repair mechanisms, WRN inhibition produces minimal toxicity. However, MSI-high tumors, characterized by defective DNA mismatch repair and expanded microsatellite repeats, become critically dependent on WRN for replication fork progression and genomic stability. This dependency creates an exceptional therapeutic window where selective WRN inhibition triggers DNA damage accumulation, replication fork collapse, and apoptosis specifically in cancer cells.\n\nCurrent clinical development efforts have demonstrated promising efficacy signals, with compounds like HRO761 achieving overall response rates of 14% in heavily pretreated MSI-high populations. Multiple pharmaceutical companies have advanced WRN inhibitors into clinical trials, including both covalent and non-covalent mechanisms of action. These developments underscore the transformative potential of targeting WRN helicase as a precision oncology approach for treating colorectal, endometrial, gastric, and other MSI-high malignancies that remain refractory to conventional therapies.", "# 2018 \u2013 Structure\u2013function analysis of the extended conformation of a polyketide synthase module\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Key Concept | Description | Key Reference |\n| --- | --- | --- |\n| Extended Conformation | Linear domain arrangement facilitating substrate handoff. | [Li et al., JACS](https://pubs.acs.org/doi/abs/10.1021/jacs.8b02100) |\n| Catalytic Competence | Enables both inter- and intramodule chemistry without loss of activity. | [Soohoo et al., Annu Rev Biochem](https://www.annualreviews.org/content/journals/10.1146/annurev-biochem-080923-043654) |\n| Dynamic Flexibility | Linker motions gate chain transfer in sync with condensations. | [Doe et al., ACS Catalysis](https://pubs.acs.org/doi/10.1021/acscatal.4c04714) |\n| Asymmetric KS\u2013ACP | Selective intermodular interactions direct vectorial substrate flow. | [Smith et al., Nat Commun](https://pubmed.ncbi.nlm.nih.gov/39179672/) |\n| Structural Architecture | Full-length module structures inform engineering strategies. | [Bagde et al., Nat Prod Rep](https://pubs.rsc.org/en/content/articlelanding/2024/np/d3np00060e) |\n| Ancestral Reconstruction | Evolution-guided modeling of conformational dynamics. | [Researcher et al., Nat Commun](https://www.nature.com/articles/s41467-025-62168-0) |", "# Overcoming Biological Barriers: The Evolution of Oral Protein and Peptide Drug Delivery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-12-10 on marcdeller.com (categories: microplastics).*\n\n* [Oral delivery of protein and peptide drugs: from non-specific formulation approaches to intestinal cell targeting strategies](https://pmc.ncbi.nlm.nih.gov/articles/PMC8771547/)\n* [Oral delivery of protein and peptide therapeutics](https://marcdeller.com/wp-content/uploads/2025/12/ch11.pdf)\n\n* **Clinical Translation:** [Oral semaglutide and octreotide capsules demonstrate that permeation enhancement technologies can achieve regulatory approval and commercial success](https://pmc.ncbi.nlm.nih.gov/articles/PMC8424290/).\n* **Device Innovation:** [Robotic pills deliver octreotide with 65% bioavailability by injecting directly through the intestinal mucosa](https://pmc.ncbi.nlm.nih.gov/articles/PMC10990675/).\n* **Cellular Targeting:** [M-cell and enterocyte receptor-mediated endocytosis increases oral bioavailability up to 10-fold compared to untargeted formulations](https://pmc.ncbi.nlm.nih.gov/articles/PMC8771547/).\n* **Validation Gap:** [Fewer than 2% of oral protein candidates advance to clinical trials, highlighting critical needs for enhanced safety and scalability validation](https://pmc.ncbi.nlm.nih.gov/articles/PMC12114803/).\n\n1. **Oral delivery of protein and peptide drugs: from non-specific formulation approaches to intestinal cell targeting strategies:** [Chen et al., Theranostics, 2022](https://pmc.ncbi.nlm.nih.gov/articles/PMC8771547/)\n2. **Advances in the Oral Delivery of Protein and Peptide Drugs:** [Liu et al., Signal Transduction and Targeted Therapy, 2025](https://pmc.ncbi.nlm.nih.gov/articles/PMC12114803/)\n3. **Recent Advances in Oral Insulin Delivery Technologies:** [Wang et al., Advanced Drug Delivery Reviews, 2024](https://pmc.ncbi.nlm.nih.gov/articles/PMC10932876/)\n4. **Challenges and Opportunities in Delivering Oral Peptides and Proteins**[: Johnson et al., Nature Reviews Drug Discovery, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC10990675/)\n\n**Navigating the gastrointestinal maze to unlock the potential of oral biologics**\n\nProtein and peptide drugs offer remarkable therapeutic specificity but face formidable obstacles in oral delivery that have historically limited their administration to injection routes. The gastrointestinal tract presents a hostile environment where acidic stomach conditions and proteolytic enzymes including pepsin, trypsin, and chymotrypsin rapidly degrade delicate protein structures before absorption can occur. Additional barriers include the viscous mucus layer that traps drug particles and tight junctions between intestinal epithelial cells that block paracellular transport of large molecules. These challenges explain why fewer than 2% of protein drugs currently reach the clinic in oral form despite hundreds of attempts, with most candidates failing to achieve measurable systemic exposure.\n\nInnovative formulation strategies have systematically dismantled these barriers through multiple converging approaches. Chemical modification techniques such as PEGylation and lipidization enhance molecular stability while increasing lipophilicity for better membrane penetration, though these modifications sometimes reduce receptor binding affinity. Permeation enhancers like SNAC and 5-CNAC temporarily modulate intestinal epithelium integrity by opening tight junctions, with semaglutide leveraging this approach to achieve commercial success as the first oral GLP-1 analog. Cell-specific targeting represents a more sophisticated tactic\u2014enterocyte receptors like the proton-coupled folate transporter and M-cell markers such as glycoprotein 2 enable receptor-mediated endocytosis, bypassing degradative pathways and achieving 10-fold higher bioavailability compared to untargeted formulations. Device-based delivery has achieved particular prominence, with self-orienting millimeter-scale applicators and robotic pills that inject drugs directly through the intestinal mucosa, demonstrating bioavailability exceeding 65% for octreotide in human studies while avoiding first-pass metabolism entirely.\n\nThese technological advances have translated into tangible clinical reality. Oral semaglutide reached patients in 2019 using Eligen\u00ae technology, while octreotide capsules gained FDA approval through transient permeation enhancement, offering acromegaly patients a non-injectable alternative. Multiple oral insulin formulations remain in phase II/III trials, with ORMD-0801 showing significant glucose reduction in clinical studies and demonstrating 24% fewer hyperglycemic events. Nevertheless, fundamental hurdles persist\u2014achieving bioavailability above 1% remains a benchmark few candidates meet, and the long-term safety of chronic permeation enhancer use requires extensive validation in larger populations. Manufacturing scalability poses additional constraints, as nanoparticle uniformity and device production costs must align with market demands. The field continues advancing toward personalized oral biologics that maintain therapeutic efficacy while fundamentally improving patient quality of life across chronic diseases.", "# Crystal structure of a soluble cleaved HIV-1 envelope trimer\nAuthors: Julien, J.-P.; Cupo, A.; Sok, D.; Stanfield, R.L.; Lyumkis, D.; Deller, M.C.; Klasse, P.-J.; Burton, D.R.; Sanders, R.W.; Moore, J.P.; Ward, A.B.; Wilson, I.A.\nSource: Science, 2013\n\n*Fig. 2.* Comparison of gp120 and components as observed in high-resolution crystal structures and in the soluble HIV-1 Env trimer. (A) High-resolution crystal structure of core gp120 (PDB ID: 3JWD, pale green) is superimposed on the gp120 component of the soluble, cleaved SOSIP.664 trimer crystal structure (yellow). A longer \u03b11 helix (gp120 residues 99-117) likely contributes to rearrangement in the bridging sheet, particularly in \u03b22 and \u03b23, which extend into V1/V2 atop the trimer. The gp120 \u03b22-proximal residues 115-125 are highlighted in blue and brown in core gp120 and trimeric gp140, respectively. A 2Fo-Fc electron density map contoured at 1.0 \u03c3 is shown as a gray mesh around the \u03b11 helix. (B) Superimposition of the scaffolded gp120 V1/V2 crystal structure (PDB ID: 3U4E, pale green) on V1/V2 in the trimer crystal structure (orange). There are differences in V1 and in the \u03b22 and \u03b23 connecting strands. Electron density for carbohydrates at gp120 N156 and N160 is shown as a 2Fo-Fc gray mesh contoured at 1.0 \u03c3. (C) Structural arrangement of gp120 V1/V2 (orange) and V3 (red) in the context of the trimer. A 2Fo-Fc electron density map contoured at 1.0 \u03c3 is shown as a gray mesh around V1/V2 and V3 elements. All structures are depicted according to secondary structure elements, with glycans depicted as yellow spheres and the Cys126-Cys196 disulfide bond colored blue. The figure was generated with Pymol.\n\n*Fig. 3.* Structural organization of gp41 in the soluble cleaved HIV-1 Env trimer. (A) Overall arrangement of gp41 elements from one protomer is shown in a gray 2Fo-Fc electron density map contoured at 1.0 \u03c3. Carbohydrates are shown as spheres. Dashed lines delineate connecting electron density for which a chain trace and secondary structure determination was ambiguous at this moderate resolution. (B) Regions of contact between the gp120 inner domain and the gp41 central helix. The inset shows hydrophobic residues in the gp120 high-resolution crystal structure from C1, \u03b10 and loop A that line the interface with gp41 HR1. (C) Surface rendering of two of the three HIV-1 Env and Influenza hemagglutinin (HA) (PDB ID: 4FNK) protomers (back protomer omitted) emphasizes the similarity in position and size of a small, central inter-protomer opening that presumably facilitates conformational changes during the fusion process. (D) Comparison between the structures of HIV-1 Env and Influenza HA (PDB ID: 4FNK), the prototype type I fusion protein. There are striking similarities in the position of structural elements in the two glycoproteins. (E) The trimeric arrangement of gp41 HR1 in the post-fusion conformation (PDB ID: 2X7R, beige) superimposes closely with the central HR1 in the soluble HIV-1 Env trimer, which is similar to the retention of the three-helix bundle at the top of the long HA2 helix in Influenza HA. All structures are depicted according to secondary structure elements. The figure was generated with Pymol.\n\n*Fig. 4.* Complete structural definition of the PGT122 epitope. (A) In addition to the N332 glycan (yellow), the PGT122 bnAb recognizes both protein and glycan elements near the base of gp120 V1 (orange) and V3 (red) to mediate broad and potent HIV-1 neutralization. Heavy and light chain CDRs are shown as dark and light blue tubes, respectively. Electron density for oligomannose glycans (spheres) surrounding the PGT122 epitope is shown as a 2Fo-Fc gray mesh contoured at 1.0 \u03c3. (B) Superimposition of the PGT122 epitope with glycans from PGT121 (PDB ID: 4JY4 and 4FQC) and PG16 (PDB ID: 4DQO) liganded crystal structures. PGT122 binding is compatible with the involvement of complex/hybrid glycans at gp120 N137 and N156/173 (green) when the trimer is expressed in human cells capable of making this type of glycan. The figure was generated with Pymol. (C) The use of glycan knock-out mutants of HIV-1 BG505 pseudoviruses reveals the importance of glycans at N137, N156 and N301 for PGT122 recognition and neutralization. These glycans are part of the PGT122 epitope in the trimer crystal structure to varying extents.\n\n*Fig. 5.* Revisiting glycan-dependent epitopes of bnAbs in the context of the soluble Env trimer structure. (A) Superimposition of the co-crystal structures of PG16 (green), PGT128 (red) and PGT135 (cyan) on the PGT122 (blue)-Env trimer crystal structure reveals that these glycan-dependent epitopes overlap. (B) Co-crystal structures of glycan-dependent bnAbs in complex with monomeric gp120 or scaffolded gp120 elements provide crucial, but incomplete, details of their epitopes and, hence, how they recognize Env. The various bnAb epitopes are colored individually on the surfaces, with overlapping elements in purple. (C) Superimposition of the PG9/16, PGT128 and PGT135 co-crystal structures on the soluble trimer-PGT122 co-crystal structure reveals that these glycan-dependent bnAbs have expanded epitopes, which creates overlapping sites of vulnerability on HIV-1 Env. (D) Model of the expanded PG16 epitope on Env. Antibody interactions additional to those previously described in the PG9/PG16-V1/V2-scaffold co-crystal structures are predicted to occur with gp120 V3, as well as with residues N160 and N197 of the adjacent protomer. Protomers are denoted with P1, P2, P3 subscripts. (E) Model of the expanded PGT128 epitope includes additional interactions with glycans at N137 and N156 mediated via the PGT128 light chain (light blue). (F) PGT135 would clash (yellow star) with the gp120 V1 conformation recognized by PGT122, based on superimposition of the PGT135-core gp120 co-crystal structure. A slight re-orientation of gp120 V1 (green) would allow PGT135 to interact, consistent with the hypothesis that the V1 loop is flexible enough to permit different modes of bnAb interaction. Heavy and light chains are shown as dark and light blue tubes, respectively. The figure was generated with Pymol.\n\n## Acknowledgments", "# pants.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n## pants.mdeller.com/\n\n## **Therapeutic PETase Design:** PANTS mines metagenomic sequence space for PET-degrading enzymes to enable protein engineering and therapeutic discovery.\n\nPET is the plastic in drinks bottles and polyester clothing, and fragments of it\nnow turn up in human blood and tissue. A few bacterial enzymes can cut PET apart, but almost every\none was found or engineered for an industrial recycling reactor: hot, alkaline, nothing like a\nhuman body. A *therapeutic* enzyme has to work at 37 \u00b0C in blood, and the published lineage\nhas been moving steadily the other way.\n\nSo PANTS looks somewhere else. **Metagenomics** reads the DNA of\nwhole microbial communities straight from the environment, without culturing anything, reaching\nthe vast majority of bacteria nobody has ever grown in a lab. PANTS searches four communities with\ndecades of plastic exposure and every reason to have evolved a use for it: the\n**human gut**, **compost**, the **ocean surface** where\ndebris collects, and **landfill**. Some 15 million protein sequences are scanned for\nthe polyesterase fold, then filtered on whether the three catalytic residues actually meet in\nspace rather than merely appearing in the sequence. Survivors get a predicted structure, an\nactive-site measurement and a rank.\n\n**The honest part.** Finding this enzyme family is easy. Telling which of its\nmembers actually chew through PET is not, and that is the open problem PANTS exists to work on.\nOnly 342 of the 910 enzymes it learns from rest on a\npublished measurement; the rest carry a label a computer inferred from sequence similarity. A\nmodel trained on those inferred labels looks flawless, but only because it has rediscovered the\nrule that made them. Scores here rank candidates worth testing at the bench. They are not\nevidence that any candidate works.\n[Read the Methods page](/methods) before using PANTS.\n\n14,804,920\n\nsequences processed\n\n439\n\ncandidates recovered\n\n416\n\npredicted structures\n\n342\n\nexperimentally evidenced positives\n\n25\n\nknown enzymes catalogued\n\n2759\n\nactivity measurements, all cited\n\n## Lineage and findings\n\nSeven views of the same 25 enzymes. Every value is\nread from the table below, so nothing here is a separate transcription that could drift\nfrom it.\n\n## Named enzymes and their lineages 25 of 342 measured\n\nThis is the **curated\nsubset**: the enzymes with a name, a known lineage and a structure. It is not the whole\nreference set. 342 enzymes carry a published measurement, almost all from\nPAZy, and those are what the activity head is trained and evaluated on \u2014\n[see all of them](/reference).\n\nGrouped by lineage with\neach wild type heading the variants built on it. Read down the Topt column within a\nlineage and the pattern is the one that motivates this project: engineering runs *away* from\n37 \u00b0C, the therapeutic target, which is marked on\nevery row at or below it. Each row opens a page carrying its structure, active-site geometry,\nmutation set, measured activity and every external identifier.", "# Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\nAuthors: Kong, L.; Lee, J.H.; Doores, K.J.; Murin, C.D.; Julien, J.-P.; McBride, R.; Liu, Y.; Marozsan, A.; Cupo, A.; Klasse, P.-J.; Hoffenberg, S.; Caulfield, M.; King, C.R.; Hua, Y.; Le, K.M.; Khayat, R.; Deller, M.C.; Clayton, T.; Tien, H.; Feizi, T.; Sanders, R.W.; Paulson, J.C.; Moore, J.P.; Stanfield, R.L.; Burton, D.R.; Ward, A.B.; Wilson, I.A.\nSource: Nature Structural & Molecular Biology, 2013\n\nJR-FLmV3 glycan mutants were expressed in 293S cells. Glycoproteins from the crude supernatant were captured onto ELISA plates using mAb 17b. Serial dilutions of biotinylated PGT 135 were added and antibody binding was probed with alkaline phosphatase conjugated streptavidin (Jackson, diluted to 1:1000) and visualized with p-nitrophenol phosphate substrate (Sigma) at 405 nm.\n\n### High density high-mannose array printing\n\nMan7GlcNAc2-Gly, Man8GlcNAc2-Gly, and Man9GlcNAc2-Gly were printed in replicates of six onto NHS-activated glass slides (from ADA Technologies, Inc) at a concentration of 100 \u03bcM as previously described using a MicroGridII contact microarray printing robot. Printing efficiency was determined by measuring ConA binding.\n\n### Binding of antibodies to high-density high-mannose array\n\nBinding of PGT 135, PGT 128 and 2G12 antibodies was measured at 30 \u03bcg/mL and detected using goat-anti-human-Fc\u03b3-R-PE (15 \u03bcg/mL, Jackson). Arrays were scanned for R-PE fluorescence on a ProScanArray HT (PerkinElmer) confocal slide scanner at 70PMT90LP. Signal intensities were collected using Imagene (BioDiscovery) image analysis software and calculated using the mean intensity of 4 replicate spotted samples.\n\n### Comparison of PGT 135 binding to low density (Schott slides) and high-density (ADA slides) glycan microarrays\n\nGlycan microarray analysis was initially carried out on high-mannose glycans printed on NHS-activated slides obtained from Schott, but binding of PGT 135 could not be detected (Supplementary Fig. 5a). However, PGT 135 glycan binding could be detected (Supplementary Fig. 5a) using the glycan array imprinted on higher density NHS-activated slides from ADA Technologies, Inc. As a comparison, we show that binding of PGT 128 was greatly enhanced using the high-density slides (Supplementary Fig. 5b) and strong binding on the array could still be detected at 1 \u03bcg/mL.\n\n### Neoglycolipid (NGL) microarray analyses\n\nAnalysis with neoglycolipid arrays was carried out as described previously. PGT 135 was pre-complexed with biotinylated anti-human-IgG (Vector) at a 1:3 ratio, w/w, before applying onto the slides at a final concentration of 10 \u03bcg/ml. Binding was detected with Alexa-Fluor 647 labeled streptavidin (Molecular Probes) at 1 \u03bcg/ml. Included for comparison are the results with human 2G12 (Polymun Scientific) taken from an earlier experiment performed using a different version of microarrays, PGT 128 and plant lectin ConA (Supplementary Table 8; Supplementary Fig. 5c).\n\n### Electron microscopy and image processing\n\nPGT 135 Fab in complex with the BG505 SOSIP.664 trimer used for electron microscopy studies were prepared as previously described. Particles were picked automatically using DoG Picker and put into a particle stack using the Appion software package. Initial reference-free 2D class averages were calculated using particles binned by 4 via the Xmipp Clustering 2D Alignment, and IMAGIC software programs. Particles were further classified into reference-free 2D class averages and refined using Refine2d in the EMAN package (Supplementary Fig. 6a). An ab initio 3D model was generated by refining the EM map of an unliganded BG505 SOSIP.664 trimer for 29 iterations, against the 2D class averages from the last Refine2d iterations. This initial model was used for the final 75-iteration 3D refinement against 8,831 raw particles binned by 2 using EMAN. C3 symmetry was imposed throughout the reconstruction process. The final 3D reconstruction has a resolution of 20 \u00c5 by an FSC cut-off at 0.5 (Supplementary Fig. 6b).\n\n### Fitting of gp120/PGT 135 crystal structure into the EM density\n\nDue to the high B-values in the constant region of the PGT 135 Fab, initial rigid body fitting of the crystal structure was done with only the gp120 and the variable region of the PGT 135 Fab (Supplementary Fig. 6c). This structure was manually fit into the EM density and refined using the UCSF Chimera 'Fit in Map' function. The crystal structure with the full Fab was then aligned to the fitted structure using the 'Match' command. The fitting of trimeric gp120-PGT 135 crystal structure was further refined using the 'Fit in Map' function resulting in a final correlation value of 0.91.\n\n### Analysis of Fab binding degeneracy\n\nCandidate top views of PGT 135-trimer and PGT 128-trimer particles were subjected to two rounds of reference-free 2D class averaging using Xmipp Clustering 2D Alignment. 2D class averages with the clearest top views of gp120 were selected from the PGT 135 and PGT 128 alignments to be used as a template for the respective datasets. A circular mask was applied to conceal the Fabs, and the remaining particles from the second round of reference free 2D alignments were re-classified via IMAGIC Multi Reference Alignment, and IMAGIC Multivariate Statistical Analysis programs. The 164 and 160 aligned particles for PGT 135 and PGT 128 respectively, from the best 2D class average for each were selected for 2D principal component analysis (PCA). The PCA analysis was carried out through the Sparx package using the first 5 eigenvectors. UCSF Chimera was used to create morph movies (Supplementary movies 1-2) between the average image and each of the eigenvectors to visualize the flexibility.\n\n## Table 1. Data collection and refinement statistics (molecular replacement)", "# GitHub README: chatPDB\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/chatPDB, 2026\n\n| | Round 1 (32B v1) |\n|---|---|\n| Base model | mlx-community/Qwen3-32B-4bit |\n| SFT dataset | v7 \u2014 97,272 examples (77,818 train / 9,727 valid / 9,727 test) |\n| LoRA config | rank 32, scale 64, RSLoRA, num_layers 32/64, dropout 0.05 |\n| max_seq_length | 1,536 (real-measured: covers 100% of train.jsonl, p99.9=1,190, max=1,973) |\n| Iterations trained | ~803 of a planned 1,420 (stopped early \u2014 see below) |\n| Best checkpoint | true iter 600, val_loss 0.176 |\n| Stop reason | real val-loss plateau: block-averaged val loss stopped improving from ~iter 300 onward (0.202 best-block avg at iters 200-300, no further downward trend through iter 800) \u2014 not overfitting, just diminishing returns on the remaining budget |\n| RAG corpus | 42 source files, 105,463 indexed chunks (ChromaDB + BAAI/bge-base-en-v1.5) |\n| Real incidents this round | crashed mid-run at true iter ~630, resumed from the true-iter-600 checkpoint (weights-only resume \u2014 iteration counter and LR schedule do not carry over, see `PROJECT_PLAN.md`); a real `mlx_lm` bug (val-loss/train-loss step misalignment) found and patched locally |\n\nFull narrative for every round, including every real bug found and fixed along the way, lives in\n[`PROJECT_PLAN.md`](PROJECT_PLAN.md).\n\n## \ud83d\udcca Evaluation\n\n`eval/eval_pdb.py` (single-model harness) and `eval/compare/eval_compare.py` (multi-round\ncomparison \u2014 auto-managed `mlx_lm.server`, `ResourceMonitor`, `--resume`, HTML + Markdown reports)\nscore every response against **real, corpus-grounded ground truth** \u2014 PDB ID validity checks real\ncorpus membership (not just format), cross-reference accuracy checks real SIFTS UniProt/CATH/EC\nmappings, numerical fidelity checks real resolution/R-free/chain counts, tool executability runs\nreal Biopython blocks in the Phase 6 sandbox. No metric relies on hand-authored ground truth.\n\n| Metric | Round 1 (32B v1) smoke-test sample |\n|---|---|\n| PDB ID validity | 89\u2013100% (`--n 20` / `--limit 10` runs) |\n| Cross-reference accuracy | 50% |\n| Tool executability | n/a in the 10-example sample (0/1 in the 20-example sample \u2014 the one attempt was a DSSP block, correctly excluded as not-yet-enabled per Phase 6, not a failure) |\n| Numerical fidelity | 33\u201340% |\n| Refusal accuracy | 100% |\n| Degeneration-free | 100% |\n\nThese are small smoke-test samples (`--n 20`/`--limit 10`), run to verify the harness end to end \u2014\nnot a full evaluation pass. Run `python eval/eval_pdb.py --n 200` for a real-sized, seeded sample\nonce a model server is up (see **How to run** below).\n\n## \ud83d\udcda Corpus\n\nwwPDB/RCSB structures and metadata, SIFTS UniProt/Pfam/CATH/GO/EC/InterPro mappings, TWILIGHT\nligand-fit scores, PDB-REDO refinement deltas, EMDB map metadata, OPM membrane placement, MobiDB\ndisorder predictions, SCOP2 fold classification, sequence-redundancy clusters, obsolete-entry\nrecords, BindingDB affinities, STRING interactions, Pharos druggability, AlphaFold predictions,\nAlphaFraud PDB-vs-AlphaFold disagreement data, PubMed/CrossRef citation verification, and full\nPyMOL/ChimeraX/py3Dmol command references \u2014 see [`data/README.md`](data/README.md) for the\ncomplete dataset-version history and construction rules.\n\n## \ud83c\udf10 Web deployment\n\n**Live at [chatpdb.mdeller.com](https://chatpdb.mdeller.com).** Not a rebuilt chat-bubble\ninterface: the browser gets a real terminal (xterm.js) connected over WebSocket to a\nserver-side PTY running the actual, unmodified `scripts/chat.py` \u2014 the same Rich-rendered\nmarkdown, corpus retrieval, and slash commands as running it locally.\n\n**No GPU on the droplet.** `chat_remote.py` (sibling to [ChemSage](https://github.com/bellcheddar/ChemSage)'s\nown, reusing the same pattern) is a drop-in launcher: at import time it patches\n`mlx_lm.generate`/`mlx_lm.stream_generate` to call a HuggingFace Space running the real fused\n32B model on ZeroGPU (a serverless, on-demand GPU allocation), then runs `chat.py` completely\nunaware anything is different.\n\n**Serving stack:** gunicorn with a single `gevent` worker (100 concurrent connections) rather\nthan `sync` workers \u2014 a terminal session is I/O-bound (waiting on the remote model and on\nkeystrokes), not CPU-bound, so one greenlet-based worker serves many concurrent sessions\ncheaply. nginx reverse-proxies both the WebSocket endpoint (`/ws`, with `Upgrade`/`Connection`\nheaders and a long read/send timeout for idle-but-open sessions) and the static app shell,\nbehind a Let's Encrypt certificate. Runs as a dedicated systemd service on the droplet\nalongside ChemSage and BoltzMaker's own web apps.\n\n**Source isolation:** the web app (`web/flask_app/`) lives on a separate `chatpdb-web` git\nbranch/worktree, so it never touches the tested training/eval code on `main`.\n\n## \ud83e\uddf1 Stack\n\nMLX-LM (QLoRA fine-tuning + serving), Biopython/gemmi/DSSP/PyMOL/ChimeraX/MDAnalysis/ProDy (tool\ncalls), ChromaDB + BAAI/bge-base-en-v1.5 (RAG retrieval), pandas/plotly/bio3d (analysis tools\ntaught to the model), Weights & Biases (training observability).\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/chatPDB\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/chatPDB</a></td>\n</tr>\n</table>", "# Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\nAuthors: Hummel, J.R.; Xiao, K.-J.; Yang, J.C.; Epling, L.B.; Mukai, K.; Ye, Q.; Xu, M.; Qian, D.; Huo, L.; Weber, M.; Roman, V.; Lo, Y.; Drake, K.; Stump, K.; Covington, M.; Kapilashrami, K.; Zhang, G.; Ye, M.; Diamond, S.; Yeleswaram, S.; Macarron, R.; Deller, M.C.; Wee, S.; Kim, S.; Wang, X.; Wu, L.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2024\n\nPrevious reports of CDK2 inhibitors have demonstrated that substitution in the kinase hinge region can play a key role in determining kinase selectivity. For example, while CDK2/4/6 inhibitor ebvaciclib bearing an aminopiperidine in this region is highly selective against JAK2, zotiraciclib with a substituted phenyl moiety inhibits both JAK2 and CDK2. Furthermore, an aminopiperidine derivative of pan-CDK inhibitor milciclib showed improved general kinase selectivity. Based on these earlier findings, we hypothesize that the improvement in JAK2 and high general kinase selectivity observed for compound 2 arose from the sp3-hybridized aminopiperidine. However, for compound 2, only 64-fold selectivity against subfamily member CDK1 was observed. As a critical regulator of the G2 to M phase transition and DNA replication, CDK1 is a pan-lethal gene that is essential for the survival and proliferation of normal cells. Hypothesizing that much of the clinical toxicities observed for multitargeted CDK inhibitors might have arisen from CDK1 inhibition, we focused our early design strategy on incorporating structural modifications to mitigate CDK1 activity. To this end, we discovered that varying substitution at the C5-position of the pyrimidine played a key role in determining CDK2 activity and isoform selectivity and that by replacing the chloro group in compound 2 with a trifluoromethyl substituent (3), we achieved a marked increase in both CDK2 potency (IC50 = 3.2 nM) and CDK1 selectivity (IC50 = 407 nM, 130-fold).\n\n### SAR of N-substituted pyrazoles\n\nHaving identified a promising core scaffold, we next fixed the 5-trifluoromethyl-2-aminopyrimidine and evaluated SAR for a series of N-substituted pyrazoles to further improve the CDK2 potency and off-target CDK selectivity (Table 1). Cyclic substituents were first evaluated. Replacing the tetrahydropyran in 3 with a lipophilic cyclohexyl moiety (4) maintained sufficient CDK2 biochemical potency and CDK1 selectivity, but showed poor solubility in simulated gastric fluid (SGF) and low human microsomal stability (entry 1). Increasing polarity by introducing an isoelectronic N-methylpiperidine substituent (5) provided improved solubility and intrinsic clearance, albeit with a significant reduction in CDK1 selectivity (entry 2). Initial evaluation of aromatic groups demonstrated that an ortho-fluorophenyl substituent (6) was well tolerated in the CDK2 enzyme assay (IC50 = 2 nM) and displayed promising CDK1 selectivity and low clearance (entry 3). We next evaluated 6 in a human whole blood assay spiked with the CCNE1-amplified ovarian cancer cell line COV318 by measuring the inhibition of Rb phosphorylation (pRb). Unfortunately, compound 6 showed low activity in the human whole blood (WB) assay (IC50 \u223c 5 \u03bcM). Switching to chloro derivative 7 gave improved enzymatic and WB potency and provided comparable CDK1 selectivity, but with a marked increase in human intrinsic clearance (h-Cl) (entry 4).\n\nLinear and branched substituents were also evaluated. Notably, compound 8, bearing a 2,2,2-trifluoroethyl moiety, displayed promising CDK2 WB potency and good selectivity against off-target CDKs (entry 5). In addition, compound 8 exhibited promising human intrinsic clearance (<0.5 L/h/kg), although solubility in simulated gastric fluid (SGF) was low (3 \u03bcg/mL). On the other hand, compound 9, containing a geminal dimethyl alcohol substituent, displayed improved SGF solubility (>400 \u03bcg/mL) and human microsomal stability (<0.5 L/h/kg), and achieved subnanomolar CDK2 biochemical potency (entry 6). Demonstrating potent target inhibition and favorable CDK isoform selectivity, compounds 7\u20139 emerged as promising leads for further development.\n\n*Table 1.* Selected SAR of N-Substituted Pyrazoles. Mean IC50 values measured in the presence of 1 mM ATP. Retinoblastoma protein (Rb) S780 phosphorylation measured in the CCNE1-amplified COV318 cell line treated with compounds in human whole blood overnight. Human microsomal intrinsic clearance (L/h/kg).\n\n| entry | cmpd | pyrazole N-substituent | properties reported in the text |\n| --- | --- | --- | --- |\n| \u2014 | 3 | tetrahydro-2H-pyran-4-yl | CDK2 IC50 = 3.2 nM; CDK1 IC50 = 407 nM (130-fold selectivity) |\n| 1 | 4 | cyclohexyl | sufficient CDK2 potency and CDK1 selectivity; poor SGF solubility; low human microsomal stability |\n| 2 | 5 | 1-methylpiperidin-4-yl | improved solubility and intrinsic clearance; significantly reduced CDK1 selectivity |\n| 3 | 6 | 2-fluorophenyl | CDK2 IC50 = 2 nM; promising CDK1 selectivity; low clearance; WB IC50 \u223c 5 \u03bcM |\n| 4 | 7 | 2-chlorophenyl | improved enzymatic and WB potency; comparable CDK1 selectivity; marked increase in h-Cl |\n| 5 | 8 | 2,2,2-trifluoroethyl | promising CDK2 WB potency; good off-target CDK selectivity; h-Cl < 0.5 L/h/kg; SGF solubility 3 \u03bcg/mL |\n| 6 | 9 | 2-hydroxy-2-methylpropyl (geminal dimethyl alcohol) | subnanomolar CDK2 potency; SGF solubility > 400 \u03bcg/mL; h-Cl < 0.5 L/h/kg |\n\n### SAR to balance potency, ADME, and physicochemical properties", "# Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\nAuthors: Kong, L.; Lee, J.H.; Doores, K.J.; Murin, C.D.; Julien, J.-P.; McBride, R.; Liu, Y.; Marozsan, A.; Cupo, A.; Klasse, P.-J.; Hoffenberg, S.; Caulfield, M.; King, C.R.; Hua, Y.; Le, K.M.; Khayat, R.; Deller, M.C.; Clayton, T.; Tien, H.; Feizi, T.; Sanders, R.W.; Paulson, J.C.; Moore, J.P.; Stanfield, R.L.; Burton, D.R.; Ward, A.B.; Wilson, I.A.\nSource: Nature Structural & Molecular Biology, 2013\n\n**Authors:** Leopold Kong, Jeong Hyun Lee, Katie J. Doores, Charles D. Murin, Jean-Philippe Julien, Ryan McBride, Yan Liu, Andre Marozsan, Albert Cupo, Per-Johan Klasse, Simon Hoffenberg, Michael Caulfield, C. Richter King, Yuanzi Hua, Khoa M. Le, Reza Khayat, Marc C. Deller, Thomas Clayton, Henry Tien, Ten Feizi, Rogier W. Sanders, James C. Paulson, John P. Moore, Robyn L. Stanfield, Dennis R. Burton (corresponding author), Andrew B. Ward (corresponding author) and Ian A. Wilson (corresponding author). **Affiliations (condensed):** Department of Integrative Structural and Computational Biology, The Scripps Research Institute, La Jolla, California, USA; International AIDS Vaccine Initiative Neutralizing Antibody Center, The Scripps Research Institute; Department of Immunology and Microbial Science, The Scripps Research Institute; Ragon Institute of Massachusetts General Hospital, Massachusetts Institute of Technology, and Harvard, Cambridge, Massachusetts, USA; Department of Cell and Molecular Biology, The Scripps Research Institute; Glycosciences Laboratory, Department of Medicine, Imperial College London, London, UK; Weill Medical College of Cornell University, New York, New York, USA; International AIDS Vaccine Initiative Design and Development Laboratory, Brooklyn, New York, USA; Joint Center for Structural Genomics (http://www.jcsg.org), USA (M.C.D., T.C., H.T.); Department of Medical Microbiology, Academic Medical Center, Amsterdam, Netherlands; Scripps Center for HIV/AIDS Vaccine Immunology & Immunogen Discovery, The Scripps Research Institute; Skaggs Institute for Chemical Biology, The Scripps Research Institute. **Published:** Nature Structural & Molecular Biology, volume 20, issue 7, pages 796\u2013803, July 2013.\n\n**Accession codes:** Coordinates and structure factors for Fab PGT 135 and its complex with gp120, CD4 and 17b have been deposited with the Protein Data Bank under accession codes 4JM4 and 4JM2, respectively. The Fab PGT 135/SOSIP.664 trimer EM reconstruction density has been deposited with the Electron Microscopy Data Bank under accession code EMD-2331.\n\n## Abstract\n\nA substantial fraction of broadly neutralizing antibodies (bnAbs) in certain HIV-infected donors recognizes glycan-dependent epitopes on HIV-1 gp120. Here, we elucidate how bnAb PGT 135 recognizes its Asn332 glycan-dependent epitope from its crystal structure with gp120, CD4 and Fab 17b at 3.1 \u00c5 resolution. PGT 135 interacts with glycans at Asn332, Asn392 and Asn386, using long CDR loops H1 and H3 to penetrate the glycan shield to access the gp120 protein surface. Electron microscopy reveals PGT 135 can accommodate the conformational and chemical diversity of gp120 glycans by altering its angle of engagement. The combined structural studies of PGT 135, PGT 128 and 2G12 show this Asn332-dependent epitope is highly accessible and much more extensive than initially appreciated, allowing for multiple binding modes and varied angles of approach, thereby representing a supersite of vulnerability for antibody neutralization.\n\n## Introduction\n\nThe human immune system initially generates strain-specific neutralizing antibodies to HIV-1 that recognize only a fraction of the massively diverse mutational variants of its envelope glycoprotein (Env). In the infected host, the virus diversifies extensively and escapes from these narrowly focused antibodies. Strong selection pressure drives variation in its exposed surfaces, including its five hypervariable loops (V1-V5) and around 27 N-linked glycans that cover most of the gp120 envelope glycoprotein surface. Despite this extraordinary sequence variation and \"evolving glycan shield\", broadly neutralizing activity is found in 5-25% of donor sera after 2-3 years of infection. However, by this time, the virus has firmly established itself, overwhelming even broadly neutralizing responses. Nevertheless, elicitation of such responses by a vaccine prior to virus exposure may lead to prophylactic protection against HIV-1.\n\nGiven that neutralizing antibody selection pressure shifts glycosylation sites on the virus during the course of infection, it is reasonable to assume that these antibodies may not only be targeting protein surfaces near glycans, but also the glycans themselves. This notion is supported by observations of serum neutralizing activity that is dependent on the presence of the glycan attached to Asn332 and by the recent discovery of several broadly neutralizing antibodies (bnAb) that interact with gp120 in a glycan-dependent manner. Four antibodies isolated from different HIV-1 infected donors and derived from different germline families, PGT 121, 2G12, PGT 128 and PG9, have been structurally characterized in complex with individual glycans, a gp120 outer domain fragment, and a scaffolded glycopeptide, respectively. A number of other glycan-dependent bnAbs, including PG16 and PGT 121-123, 130-131 and 142-145 (n.b. the PGT antibodies are grouped according to their predicted shared germline precursor), await structural characterization in complex with Env.\n\nTo further elucidate the bnAb target around Asn332, we conducted structural and functional analysis of PGT 135, an Asn332 (N332) glycan-dependent bnAb isolated from an individual whose serum has one of the highest neutralization breadths known. The PGT 135 epitope includes three glycans and protein segments that differs substantially from other Asn332-dependent epitopes and enables PGT 135 accommodation of HIV-1 Env glycan heterogeneity through some variation in its angle of approach to the gp120 surface. Hence, the Asn332-dependent epitope constitutes a major site of vulnerability on the glycosylated face of gp120 that can be accessed by antibodies from different germline precursors using completely different modes of recognition and, most likely, different mechanisms of neutralization.\n\n## Results\n\n### Structural characterization of PGT 135 binding to gp120", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nPoint mutations were made in the His-tagged WRN construct used for biochemical characterization (aa S2-S1235). Cysteine 580 (Walker A motif), 727 (linker), 864 (Domain 2A) and 1041 (RQC domain) were each mutated to both Serine and Alanine using site directed mutagenesis, for a total of eight mutants. Purified proteins were assayed in both helicase and ATPase assays. Cys580 mutation to either Ser or Ala abolished all activity, consistent with Walker A motif cysteine function. Cys864 (D2) mutation to alanine and Cys1041 (RQC domain) had no significant impact on helicase or ATPase activity. Cys727 abolished helicase activity (Table 6), whether it was mutated to Ala or Ser. ATPase activity for Cys727 mutants was reduced but not abolished (Table 7). Table 6 is the panel of the nine variant WRN helicase proteins that were screened using a fluorescence-based DNA unwinding assay with an increase in fluorescence intensity indicating DNA unwinding. Table 7 is the panel of the nine variant WRN helicase proteins that were screened using ADP-Glo assay to measure DNA stimulated ATPase activity, with an increase in luminescence indicating ATP hydrolysis.\n\n**TABLE 6. Helicase activity of WRN mutants.**\n\n| WRN Constructs | Observed Helicase Activity (AU/min) | Ratio to WT |\n|---|---|---|\n| 1. C580A | 1.04 | 0.02 |\n| 2. C580S | 0 | 0.00 |\n| 3. C727A | 4.619 | 0.07 |\n| 4. C727A | 1.452 | 0.02 |\n| 5. C864A | 60.75 | 0.94 |\n| 6. C864S | 28.49 | 0.44 |\n| 7. C1041A | 60.14 | 0.93 |\n| 8. C1041S | 40.28 | 0.62 |\n| 9. WR WRN 12 | 64.77 | 1.00 |\n\n**TABLE 7. ATPase activity of WRN mutants.**\n\n| WRN Constructs | Observed Helicase Activity (AU/min) | Ratio to WT |\n|---|---|---|\n| 1. C580A | 2.548 | 0.03 |\n| 2. C580S | 0.8014 | 0.01 |\n| 3. C727A | 22.65 | 0.29 |\n| 4. C727A | 25.42 | 0.33 |\n| 5. C864A | 70.52 | 0.90 |\n| 6. C864S | 42.02 | 0.54 |\n| 7. C1041A | 66.66 | 0.85 |\n| 8. C1041S | 45.74 | 0.59 |\n| 9. WR WRN 12 | 78.17 | 1.00 |\n\nVarious modifications of the contents of the present disclosure, in addition to those described herein, will be apparent to those skilled in the art from the foregoing description. Such modifications are also intended to fall within the scope of the appended claims. Each reference, including without limitation all patent, patent applications, and publications, cited in the present application is incorporated herein by reference in its entirety.\n\n## Claims\n\n**1.** A compound of Formula (I):\n\n*Formula (I) as drawn in the Summary above.*\n\nor a pharmaceutically acceptable salt thereof, wherein: m is an integer selected from 0, 1, 2, 3, 4, 5, 6, 7, or 8; n is an integer selected from 0, 1, 2, 3, 4, 5, 6, 7, or 8; the dashed bond is a single or double bond, provided that proper valency is maintained; Z is N or CR^7; X is C, CH or N and Y is C or CH, or X is C or CH and Y is C, CH, or N; Ring moiety A is selected from C3\u201314 cycloalkyl, 6\u201310 membered aryl, 4\u201314 membered heterocycloalkyl, and 5\u201310 membered heteroaryl; Ring moiety B is selected from C3\u20137 cycloalkyl, phenyl, 4\u20137 membered heterocycloalkyl, and 5\u20136 membered heteroaryl; and R^1, R^2, R^3, R^4, R^5, R^6, R^7 and R^W, together with all of their sub-variables (R^a1/R^b1/R^c1/R^d1/R^e1, R^1A, R^1B, R^a2/R^b2/R^c2/R^d2, R^2A, R^2B, R^a3/R^b3/R^c3/R^d3, R^6A, R^6B, R^81\u2013R^85, R^9, R^G, the linkers L, L^1\u2013L^7 and L^x, Ring D, X^1, Ar and the integers q, t and u), are as defined in the specification and recited in full in claim 1.\n\n*(Claim 1 as published recites the complete Markush definitions of every variable across approximately ten columns of text; the definitions are those reproduced in the Detailed Description above and are not repeated verbatim here.)*\n\n**2-6.** (canceled)\n\n**7.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein Ring moiety A is selected from C3\u20137 cycloalkyl, phenyl, 4\u20136 membered heterocycloalkyl, and 5\u20136 membered heteroaryl.\n\n**8.** (canceled)\n\n**9.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein Ring moiety A is phenyl.\n\n**10-12.** (canceled)\n\n**13.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein Ring moiety B is selected from C3\u20137 cycloalkyl, phenyl and 5\u20136 membered heteroaryl.\n\n**14-31.** (canceled)\n\n**32.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein R^1 is independently selected from H, D, halo, CN, C1\u20136 alkyl, C1\u20136 haloalkyl, C3\u20134 cycloalkyl, OR^a1, SR^a1, NHOR^a1, C(O)R^b1, C(O)NR^c1R^d1, C(O)OR^a1, OC(O)R^b1, OC(O)NR^c1R^d1, NR^c1R^d1, NR^c1C(O)R^b1, NR^c1C(O)OR^a1, NR^c1C(O)NR^c1R^d1, NR^c1S(O)R^b1, NR^c1S(O)2R^b1, NR^c1S(O)2NR^c1R^d1, S(O)2R^b1, and S(O)2NR^c1R^d1, wherein said C1\u20136 alkyl, C1\u20136 haloalkyl, and C3\u20134 cycloalkyl are each optionally substituted with 1, 2, 3, or 4 independently selected R^1A substituents.\n\n**33.** (canceled)\n\n**34.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein each R^a1, R^c1, and R^d1 is independently selected from H, C1\u20136 alkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, and C3\u20134 cycloalkyl, wherein said C1\u20136 alkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, C1\u20136 haloalkyl, and C3\u20134 cycloalkyl are each optionally substituted with 1, 2, 3, or 4 independently selected R^1A substituents; and each R^b1 is independently selected from C1\u20136 alkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, C1\u20136 haloalkyl, and C3\u20134 cycloalkyl, which are each optionally substituted with 1, 2, 3, or 4 independently selected R^1A substituents.\n\n**35-36.** (canceled)", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\nThis superfamily includes receptors for all of the group 1 helical cytokines, except the interferons and IL-10 family of ligands (Cosman et al., 1990; Cosman, 1993) (see Figure 1-4). This superfamily is characterised by the conserved nature of the ligand-binding domain, which is 200 amino acids in size and referred to as a haemopoietin domain. Members of the haemopoietin receptor superfamily are type I membrane glycoproteins, orientated with the N-terminus in an extracellular position through the use of an N-terminal signal sequence, which is post-translationally cleaved. The haemopoietin domain is composed of two fibronectin type III-like modules. These modules consist of 100 amino acids, which form seven \u03b2-strands that are folded in an antiparallel fashion to form a \u03b2-barrel-like structure. These receptors share a conserved pattern of cysteines in the N-terminal fibronectin type III-like domain which is required for correct folding of the haemopoietin domain and functions in disulphide bond formation between the \u03b2-sheet (Rozakis Adcock and Kelly, 1991). The C-terminal fibronectin type III-like domain contains a conserved WSXWS motif, which also performs an important structural function. Substitution of these amino acids in IL-6 results in low levels of expression suggesting incorrect protein folding and poor protein stability (Yawata et al., 1993).\n\nThe haemopoietin domain is often combined with other structurally distinct domains such as the immunoglobulin domain, as observed in IL-3R\u03b1, GM-CSFR\u03b1, IL-5R\u03b1, IL-6R\u03b1, gp130 and GCSFR. Another structural domain also utilised is the fibronectin type III module as observed in IL-6R\u03b1, gp130, LIFR and G-CSFR. The haemopoietin domain can also be duplicated as observed in IL-3R, IL-5R and GM-CSFR (see Figure 1-4).\n\n*Figure 1-4.* Diagrammatic representation of domain organisation observed in the (a) haemopoietin and (b) gp130 family of cytokine receptors. Adapted from (Nicola, 1994). See abbreviations for descriptions.\n\n##### 1.4.2.1.1 Structure-Function Relationship of Growth Hormone\n\nThis was the first haemopoietin receptor-ligand complex to be solved and revealed that one growth hormone molecule binds two receptor sub-units (de Vos et al., 1992) so forming a cytokine plus receptor homodimer complex (see Figure 1-3). The structure of each receptor unit consists of two \u03b2-barrel motifs each composed of seven \u03b2-strands (conventionally labelled strands A, B, C, C', E, F, G in common with the immunoglobulin subset C2). The \u03b2-barrels are produced by folding of two \u03b2-sheets produced by E, B, A and G, F, C, C' so forming a buried hydrophobic interior and a hydrophilic surface. Three disulphide bridges are present in the N-terminal domain which function in linking strands A to B, E to C' and F to G. Also the WSXWS motif as observed in other members of the haemopoietin receptor family is replaced with the motif YGEFS in the C-terminal domain on the loop connecting strands F and G (see Figure 1-6).\n\nIn common with many other proteins, particularly the complementarity determining regions (CDRs) of antibodies, the \u03b2-barrel structures provide a stable framework on which the functionally important loops can reside. The hinge region connecting the two \u03b2-barrels adopts a 3\u2081\u2080 helical conformation, which functions in fixing the interdomain angle at approximately 90\u00b0. The hinge region connecting the two domains functions as the ligand interaction site.\n\nHigh concentrations of hGH are capable of antagonising receptor dimerisation. This led to the proposal of a two site binding model, whereby hGH binds to a molecule of hGHR first at site 1 and following that forms a weaker interaction through site 2 with a second hGHR molecule (Cunningham et al., 1991b). Site directed mutagenesis (Cunningham et al., 1989; Cunningham and Wells, 1989) has revealed that the functional epitopes map onto the centre of these interaction sites. The receptor binding sites on hGH are located on opposite sides of the four helix bundle. Site 1 has a slightly concave character and is formed by residues of helix A, helix D and the AB loop. The total surface buried at this hormone-receptor interface is approximately 1230 \u00c5\u00b2. In contrast, the second binding site on hGH is relatively flat and is made up of exposed side chains from helix A and helix C and the N-terminal loop region. Also, the surface buried in the site 2 interface is much smaller at approximately 900 \u00c5\u00b2, and it has been proposed that this allows for weaker site 2 interactions, so giving rise to a sequential binding mechanism. Once formed, the complex is then further stabilised by burial of approximately 1000 \u00c5\u00b2 of surface in the membrane proximal halves of each receptor.", "# The 79th Organ: When Art and Biology Meet Bioengineering in the Fight Against Microplastics\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-13 on marcdeller.com (categories: blog, microplastics).*\n\n* [Art project imagines 79th organ to extract microplastics from human body](https://interestingengineering.com/photo-story/designer-creates-fungi-made-prosthetic-organ)\n\n* **Anatomical Evolution**: [Human organ count continues expanding with recent discoveries of mesentery, interstitium, and tubarial glands establishing precedent for new organ classifications\u200b](https://www.bbc.com/news/health-38506708)\n* **Microplastic Crisis**: [Humans ingest credit card equivalent plastic weekly, causing cardiovascular risks and systemic inflammation requiring urgent therapeutic intervention](https://med.stanford.edu/news/insights/2025/01/microplastics-in-body-polluted-tiny-plastic-fragments.html)\n* **Fungal Bioengineering**: [Mycelium networks demonstrate superior biocompatibility and plastic-degrading enzyme production compared to synthetic alternatives](https://www.nature.com/articles/s41598-021-91572-x)\n* **Therapeutic Innovation**: [Elora Therapeutics pioneers first systemic enzyme therapy targeting microplastic removal from human tissues](https://ati.utexas.edu/company/elora-developing-a-systemically-delivered-enzyme-to-remove-microplastics-from-the-human-body/)\n* **Bioartificial Future**: [Synthetic biology advances enable organ-scale tissue manufacturing with integrated living components for environmental detoxification](https://www.nature.com/articles/s41467-025-58030-y)\n* **Commercial Validation**: [Multiple companies developing fungal biomaterials for medical applications demonstrate market viability of living tissue engineering approaches](https://www.mycostories.com/post/fungi-health-innovation-mycelium-scaffolds-could-transform-regenerative-medicine)\n\n1. A New Organ Was Discovered in the Human Body: [J. Calvin Coffey, et al., The Lancet](https://sjmagazine.net/health/new-organ-discovered-human-body)\n2. Structure and Distribution of an Unrecognized Interstitium: [Neil Theise, et al., Scientific Reports](https://www.nature.com/articles/s41598-018-23062-6)\n3. Advanced mycelium materials as potential biomedical scaffolds: [Antinori, et al., Nature Scientific Reports](https://www.nature.com/articles/s41598-021-91572-x)\n4. Microplastics and our health research findings: [Stanford Medicine, et al., Medical Insights](https://med.stanford.edu/news/insights/2025/01/microplastics-in-body-polluted-tiny-plastic-fragments.html)\n5. Linking Metabolic Activity of Plastic-Degrading Fungi: [Research Team, et al., Nature Communications](https://pmc.ncbi.nlm.nih.gov/articles/PMC12113260/)\n6. Elora Therapeutics enzyme therapeutic development: [Austin Technology Incubator, et al., Company Profile](https://ati.utexas.edu/company/elora-developing-a-systemically-delivered-enzyme-to-remove-microplastics-from-the-human-body/)\n\n**Revolutionary fungal biotechnology offers hope for addressing humanity\u2019s plastic pollution crisis through innovative prosthetic organ design**\n\nThe human body, once thought to contain a fixed number of organs, continues to surprise us. Recent decades have witnessed remarkable discoveries challenging our understanding of anatomy, from the mesentery\u2019s reclassification as our 78th organ in 2017 to the identification of the interstitium as a potential new organ system in 2018. Most recently, scientists discovered the tubarial salivary glands in 2020, marking the first new organ discovery in nearly 300 years. Yet perhaps the most intriguing development lies not in what we\u2019ve found, but in what we might create: a speculative \u201c79th organ\u201d designed to address one of humanity\u2019s most pressing environmental challenges.\n\nThe concept emerges from designer Odette Dierkx\u2019s visionary work combining fungal biotechnology with prosthetic organ design. This biomimetic device utilizes bioengineered fungi, specifically *Pleurotus ostreatus*, to create a living filtration system capable of extracting and degrading microplastics from the human bloodstream. The core structure incorporates fungal mycelium that releases specialized enzymes through bioremediation processes, transforming harmful plastic fragments into harmless components that the body can safely eliminate.\n\nMicroplastics represent an unprecedented biological challenge, with humans now ingesting approximately one credit card\u2019s worth of plastic particles weekly. These synthetic pollutants infiltrate every organ system, from brain tissue to reproductive organs, triggering chronic inflammation, endocrine disruption, and metabolic dysfunction. Studies reveal microplastics accumulate in arterial plaques, with patients showing these contaminants facing 4.5-fold higher risks of cardiovascular events. The pervasive nature of this contamination, with annual releases of 10-40 million metric tons expected to double by 2040, necessitates innovative therapeutic interventions.\n\nFungal biotechnology offers remarkable potential for addressing plastic pollution through natural degradation mechanisms. Research demonstrates that fungi possess extraordinary capabilities for polymer breakdown, with species like *Pleurotus ostreatus* and *Ganoderma lucidum* producing enzymes including cutinase, laccase, and PETase. These organisms evolved sophisticated enzymatic machinery originally designed for lignocellulose degradation but have adapted to process synthetic polymers. The mycelium\u2019s three-dimensional fibrous networks naturally mimic human extracellular matrix structures, making them ideal candidates for biomedical scaffolds.", "# GitHub README: GOBSMACKED\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/GOBSMACKED, 2026\n\nThe one run that ranked correctly is worth stating carefully rather than celebrating.\nIts receptor came from co-folding the protein *with* erlotinib and then removing the\nligand, so its pocket is already ligand-adapted in a way an apo or predicted-apo pocket\nis not. It is one run, and a single B among four Ds is a hypothesis, not a result.\n\n### Following it up: rank the scoring functions, not the pipeline\n\nIf the limitation is which pose gets ranked first, that can be tested without docking\nanything again. The five runs left fifty poses whose distance to the crystal is already\nknown, so the same fifty were rescored by seven independent functions. Every function saw\nan identical candidate set, which makes a difference in the answer a difference in the\nscoring rather than in what there was to score.\n\n**Top-1 RMSD after rescoring the same ten poses (\u00c5; bold is within 2 \u00c5):**\n\n| Scoring function | 1M17 self-dock | 4HJO cross | AlphaFold | ESMFold | Boltz-2 | Hits | Mean |\n|---|---|---|---|---|---|---|---|\n| Vinardo (smina / Vina 1.2.5) | **1.51** | 4.33 | 4.81 | 5.44 | **1.40** | 2/5 | **3.50** |\n| AutoDock Vina (smina / Vina 1.2.5) | **1.51** | 4.29 | 6.97 | 5.44 | **1.40** | 2/5 | 3.92 |\n| smina `dkoes_scoring` | **1.51** | 4.94 | 8.28 | 8.40 | **1.40** | 2/5 | 4.91 |\n| AutoDock4 (`ad4_scoring`) | 7.85 | 4.94 | 6.92 | 5.44 | **1.40** | 1/5 | 5.31 |\n| as shipped | 8.77 | 4.29 | 7.73 | 8.09 | **1.40** | 1/5 | 6.06 |\n| smina `dkoes_fast` | 7.85 | 4.94 | 6.92 | 7.08 | 8.09 | 0/5 | 6.98 |\n| _oracle: the best pose present_ | _1.51_ | _2.29_ | _4.81_ | _5.44_ | _1.40_ | 2/5 | _3.09_ |\n\n**On the control, rescoring alone recovers the crystal pose.** Vina and Vinardo both give\npose 8 the best score of the ten and put it first, at 1.51 \u00c5, where it had been ranked\neighth. Nothing was re-docked and nothing was re-minimised: the same coordinates, read by\na different function, produce the right answer. Across all five, mean top-1 improves from\n6.06 \u00c5 to 3.50 \u00c5 against an oracle of 3.09 \u00c5, so Vinardo recovers most of the headroom\nthat exists in these pose sets.\n\n**The oracle row splits the receptors into two different problems.** For the crystal and\nco-folded pockets a near-native pose is present and merely ranked below decoys, which\nrescoring fixes. For the two predicted-apo pockets the best pose in the set is 4.81 and\n5.44 \u00c5, and no amount of rescoring can rescue a set whose best member is 4.81 \u00c5.\n\n### Was that a sampling limit, or just that engine?\n\nRescoring cannot answer it, because the poses were fixed in advance. So the same five\nreceptors were docked again from scratch with four alternative functions driving the\nsearch, starting from a conformer built from SMILES rather than from any existing pose:\nseeding a search with a docked pose would turn a docking test into a minimisation test.\nTwenty runs, about 15 seconds each.\n\n| Receptor | Best pose the original search found | Best any engine found | By |\n|---|---|---|---|\n| 1M17 self-dock | 1.51 \u00c5 | **1.17 \u00c5** | Vinardo |\n| 4HJO cross-dock | 2.29 \u00c5 | 2.04 \u00c5 | Vinardo |\n| AlphaFold DB | 4.81 \u00c5 | **3.04 \u00c5** | Vinardo |\n| ESMFold | 5.44 \u00c5 | **2.06 \u00c5** | AutoDock4 |\n| Boltz-2 co-folded | 1.40 \u00c5 | 1.40 \u00c5 | \u2014 |\n\n**The apo ceiling was a property of the search, not of the receptors.** ESMFold's best\navailable pose moves from 5.44 \u00c5 to 2.06 \u00c5. Pooling every pose any engine produced gives\neach of the five receptors something at 3.04 \u00c5 or better, and four of the five something\nat 2.06 \u00c5 or better: a mean oracle of 1.94 \u00c5 where the original sets managed 3.09 \u00c5.\n\n**No protocol tested exploits that.** Top-1 RMSD, which is the only number the pipeline\ncarries forward, at three thresholds:\n\n| Protocol | \u22642.0 \u00c5 | \u22642.5 \u00c5 | \u22643.0 \u00c5 | Mean top-1 |\n|---|---|---|---|---|\n| as shipped | 1/5 | 1/5 | 1/5 | 6.06 \u00c5 |\n| dock with Vina | 1/5 | 1/5 | 1/5 | 5.09 \u00c5 |\n| dock with Vinardo | 1/5 | 1/5 | 1/5 | 5.54 \u00c5 |\n| dock with AutoDock4 | 0/5 | 1/5 | 2/5 | 4.31 \u00c5 |\n| pool every pose, rank by Vinardo | 1/5 | **3/5** | 3/5 | 5.73 \u00c5 |\n| pool every pose, rank by AutoDock4 | 0/5 | 2/5 | 3/5 | **4.15 \u00c5** |\n| _oracle over the pooled set_ | _2/5_ | _4/5_ | _4/5_ | _1.94 \u00c5_ |\n\nThree things in that table are worth more than the ranking itself:\n\n- **The threshold decides the story.** At 2.0 \u00c5 nothing beats the shipped protocol by more\n  than one receptor; at 2.5 \u00c5 pooling and rescoring triples it. AutoDock4 lands at 2.02,\n  2.15 and 2.52 \u00c5, outside a 2 \u00c5 criterion three times over, which is why it has the best\n  mean and the worst hit count. Quoting one threshold would be a choice, not a measurement.\n- **More poses made ranking worse.** On the control, Vinardo picks 5.90 \u00c5 out of that\n  engine's own ten and 9.76 \u00c5 out of all forty-eight. A larger pool holds more\n  high-scoring decoys, so ensembling is not free.\n- **The gap is 4.15 \u00c5 against an oracle of 1.94 \u00c5.** Fixing the sampling put a near-native\n  pose within reach for every receptor, and then no scoring function picked it. Ranking is\n  still the binding constraint, and more clearly than before.\n\n### The affinity head is not a rescoring function\n\nThe one methodologically different scorer available is the deep-learning affinity head the\n`affinity` stage already runs, so the control's ten poses went through it: 1.51 to 8.77 \u00c5,\nsame protein, same ligand, same cached alignment.\n\n| | |\n|---|---|\n| pIC50 across ten poses | 6.538 to 6.732, a spread of **0.19 log units** |\n| RMSD across the same ten | 1.51 to 8.77 \u00c5 |\n| Ranked first | the 8.30 \u00c5 pose |\n| The 1.51 \u00c5 pose | 6.714, mid-pack, and identical to three decimals to the 8.77 \u00c5 pose |\n| Spearman \u03c1, pIC50 against RMSD | **+0.26 (p = 0.48)**, where a useful scorer needs a negative one |\n\nIt does not discriminate between runs and it does not discriminate within a pose set\neither. This is the measured reason affinity is reported beside the composite score\nrather than inside it, and the reason it is not used to re-rank anything.", "# ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n**Behaviour classes by round:**\n\n*Round 3 additions (targeting R2 eval gaps):*\n\n| Class | Weight | Purpose |\n| --- | --- | --- |\n| `single_property` | 2x | One property at a time with code block |\n| `full_profile` | 2x | Complete property profile via code |\n| `qed_score` | 1x | QED drug-likeness score |\n| `molecular_formula` | 1x | Formula, atom counts, ring analysis |\n| `smiles_validation` | 1x | Valid/invalid SMILES handling |\n| `property_comparison` | 1x | Side-by-side two-molecule comparison |\n| `corpus_drug` | 2x | Named drugs from approved_drugs.csv |\n| `murcko_sar` | 1x | Scaffold comparison with SAR context |\n\n*Round 4 additions (structural biology + fidelity):*\n\n| Class | Weight | Purpose |\n| --- | --- | --- |\n| `pymol_selection` | 2x | PyMOL selection algebra (within, byres, chain, resi) |\n| `pymol_viz` | 2x | PyMOL visualization (surface, alignment, symmetry, B-factor, figures) |\n| `plip_analysis` | 2x | PLIP interaction types (H-bond, hydrophobic, pi-stack, salt bridge, halogen) |\n| `pdb_format` | 2x | PDB file format (ATOM/HETATM, altloc, B-factor interpretation) |\n| `pdbtools` | 2x | pdb-tools CLI recipes (selchain, selres, tidy, splitmodel) |\n| `biopython` | 2x | Bio.PDB patterns (SMCRA, NeighborSearch, Superimposer) |\n| `chimerax` | 2x | ChimeraX commands (hbonds, clashes, matchmaker) |\n| `gemmi` | 1x | gemmi mmCIF/PDB parsing and neighbour search |\n| `struct_prep` | 1x | Structure preparation (PDBFixer, protonation, 3D conformers) |\n| `docking_interp` | 2x | Docking result interpretation (Vina scores, RMSD, Kd estimation) |\n| `plotly` | 2x | Plotly Python visualization (MW vs logP, radar charts) |\n| `r_recipe` | 2x | R/ggplot2/bio3d/Plotly-R (SAR plots, RMSD, RMSF) |\n| `protein_family` | 3x | Kinases, GPCRs, antibodies, GTPases, cytokines, receptors (real PDB data) |\n| `code_then_quote` | 3x | \u201cCompute then quote\u201d pattern for numerical fidelity |\n| `rounding_precision` | 2x | Exact rounding conventions (MW 1dp, logP 2dp, TPSA 1dp) |\n| `personality_drug` | 3x | Lipinski verdicts with \u2705/\u274c cards, emoji, and medchem humor |\n| `personality_valid` | 2x | SMILES validation with flair and emoji |\n| `personality_struct` | 3x | Structural observations with character and humor |\n| `refusal` | 2x | Out-of-scope and invalid query handling |\n\n### \ud83c\udf93 Training parameters\n\n| Parameter | Round 1 (7B) | Round 2 (32B) | Round 3 (32B) | Round 4 (32B) | Round 5 (32B, complete) |\n| --- | --- | --- | --- | --- | --- |\n| Base model | `Qwen2.5-7B-Instruct-4bit` | `Qwen2.5-32B-Instruct-4bit` | `Qwen2.5-32B-Instruct-4bit` | `Qwen2.5-32B-Instruct-4bit` | `Qwen2.5-32B-Instruct-4bit` |\n| Fine-tune type | QLoRA (4-bit base) | QLoRA (4-bit base) | QLoRA (4-bit base) | QLoRA (4-bit base) | QLoRA (4-bit base) |\n| `use_rslora` | `true` | `true` | `true` | `true` | `true` |\n| LoRA rank | 32 | 32 | 32 | 32 | **64** |\n| LoRA scale | 64.0 | 64.0 | 64.0 | 64.0 | **90.0** |\n| LoRA keys | q/k/v/o_proj + gate/up/down_proj | q/k/v/o_proj + gate/up/down_proj | q/k/v/o_proj + gate/up/down_proj | q/k/v/o_proj + gate/up/down_proj | q/k/v/o_proj + gate/up/down_proj |\n| `num_layers` | 16 (of 28) | 32 (of 64) | 32 (of 64) | 32 (of 64) | 32 (of 64) |\n| `iters` | 750 | 750 | 1,500 (stopped at 625) | 1,500 (complete) | **3,000** |\n| `batch_size` | 4 | 4 | 4 | 4 | 4 |\n| `learning_rate` | 2e-5 | 2e-5 | 2e-5 | 2e-5 | 2e-5 |\n| LR schedule | Cosine, 50-step warmup | Cosine, 50-step warmup | Cosine, 50-step warmup | Cosine, 50-step warmup | Cosine, **100**-step warmup |\n| `max_seq_length` | 2048 | 2048 | 2048 | 2048 | **3072** |\n| `steps_per_report` | 10 | 10 | 10 | 25 (aligned with eval) | **50** (aligned with eval) |\n| `val_batches` | \u2014 | \u2014 | \u2014 | 25 | **50** |\n| `seed` | 42 | 42 | 43 | 51 | 51 |\n| Trainable params | ~26M (0.34%) | 134M (0.41%) | 134M (0.41%) | 134M (0.41%) | **268M (0.82%)** |\n| Peak memory | ~12 GB | 24.9 GB | 25.2 GB | 24.9 GB | ~33 GB est. |\n| Training data | 1,200 examples | 1,200 examples | 4,000 examples | 6,400 examples | **16,000 examples** |\n| Adapter path | `adapters/chem_sage_7b_v1_lora/` | `adapters/chem_sage_32b_v2_lora/` | `adapters/chem_sage_32b_v3_lora/` | `adapters/chem_sage_32b_v4_lora/` | `adapters/chem_sage_32b_v5_lora/` |\n| Fused model path | `models/chem_sage_7b_v1/` | `models/chem_sage_32b_v2/` | `models/chem_sage_32b_v3/` | `models/chem_sage_32b_v4/` | `models/chem_sage_32b_v5/` |\n\nNotes: Round 3 was configured for 1,500 iters but stopped early at actual iter 625 after val loss plateaued at ~0.060. Training was resumed from iter 100 checkpoint after a battery shutdown. 64 layers was attempted first but caused swap thrashing at 31.5 GB peak; reverted to 32 layers. batch_size=8 was also attempted but ran 4.5x slower due to memory pressure; reverted to 4. Round 4 uses 32 layers (same as R3) \u2014 48 and 64 layers were found to be too slow (~2\u20132.5 min/iter) on M1 Max; 32 layers is the throughput sweet spot at ~40 sec/iter. Round 5 doubles the LoRA rank to 64 with RSLoRA scale 90.0 (effective alpha \u2248 11.25, same order as R4\u2019s 11.31), triples the iters to 3,000, extends max_seq_length to 3,072, and doubles val_batches to 50 for a more stable eval signal. Trainable parameters doubled to 268M (0.82%).\n\n### \ud83c\udf93 Training loss curves\n\nValidation loss measured every 25 steps over 25 batches. Round 3 was resumed from iter 100 checkpoint after a battery shutdown; actual iter = logged iter + 100 for the resumed portion.", "# Call Me A Jerk: Persuading AI to Comply with Objectionable Requests\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe challenge extends to evaluation methodologies themselves. Standard jailbreak assessments often rely on automated judges that can be fooled by sophisticated persuasive content, leading to inflated success rates for some attacks while underestimating others. Recent research suggests that many previously reported jailbreak success rates were significantly overstated, with some methods achieving less than 20% effectiveness when subjected to rigorous evaluation. However, persuasive attacks consistently demonstrate high success rates across multiple evaluation frameworks.\n\nEmerging defense strategies focus on multi-layered approaches that combine traditional technical safeguards with social awareness. Goal prioritization techniques guide model responses toward safer outputs while maintaining functionality. Detection-based defenses like LlamaGuard analyze both inputs and outputs for potential threats. Denoising methods such as Smooth-LLM attempt to neutralize adversarial perturbations through paraphrasing and retokenization. However, research indicates that combining multiple defense mechanisms significantly improves effectiveness, with dual-method combinations showing up to 98% improvement over single techniques.\n\nThe broader implications for AI alignment and safety are profound. As language models become more capable and human-like, they increasingly exhibit the cognitive biases and social vulnerabilities that affect human judgment. This suggests that future AI safety research must incorporate insights from psychology, sociology, and behavioral economics alongside traditional computer science approaches. The development of AI systems that can resist persuasive manipulation while maintaining helpful and natural interactions represents a fundamental challenge for the field.\n\nIndustry responses are beginning to emerge, with major AI companies implementing enhanced safety training and detection systems specifically designed to address persuasive attacks. However, the rapid pace of AI development and deployment often outpaces security improvements, creating windows of vulnerability that sophisticated attackers can exploit. The international AI safety community has recognized this threat, with recent initiatives allocating over \u00a315 million to research projects focused on alignment and safety challenges.\n\n| Concept | Description | Key References |\n| --- | --- | --- |\n| Persuasive Adversarial Prompts (PAP) | Automated prompts that use social persuasion techniques to manipulate LLMs into bypassing safety guardrails | [Zeng, Y., et al., ACL, 2024](https://arxiv.org/html/2401.06373v2) |\n| Jailbreaking | Techniques designed to circumvent AI safety constraints and ethical guidelines to elicit harmful content | [Zou, A., et al., arXiv, 2023](https://arxiv.org/abs/2307.15043) |\n| Social Science Taxonomy | Systematic classification of 40 persuasion techniques from psychology, communication, and sociology research | [Cialdini, R.B., Psychological Science, 2004](https://www.annualreviews.org/doi/10.1146/annurev.psych.55.090902.142015) |\n| Attack Success Rate (ASR) | Percentage metric measuring how often jailbreak attempts successfully compromise LLM safety mechanisms | [Chao, P., et al., ICML, 2024](https://arxiv.org/abs/2307.16368) |\n| GPT-4 Judge | Automated evaluation system using GPT-4 to assess harmfulness of LLM outputs on 1-5 scale | [Qi, X., et al., ICLR, 2023](https://openreview.net/forum?id=xbsmUQ0k9X) |\n| Persuasive Paraphraser | Fine-tuned language model that transforms plain harmful queries into persuasive adversarial prompts | [Zeng, Y., et al., ACL, 2024](https://arxiv.org/html/2401.06373v2) |\n| Alignment | Process of ensuring AI systems behave in accordance with human values and intended objectives | [Ji, J., et al., arXiv, 2023](https://arxiv.org/abs/2310.19852) |\n| Defense Mechanisms | Protective strategies including training-based, inference-time, and detection-based approaches | [Robey, A., et al., ICML, 2023](https://arxiv.org/abs/2310.12620) |\n| Over-refusal | Phenomenon where defensive systems inappropriately reject benign prompts, reducing utility | [Panda, S., et al., SafeGenAI, 2024](https://openreview.net/forum?id=rXReIKbm5e) |\n| Multi-turn Attacks | Sequential interaction strategies that gradually escalate requests to bypass safety measures | [Unit 42, Palo Alto Networks, 2025](https://unit42.paloaltonetworks.com/jailbreaking-generative-ai-web-products/) |", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\nFor the Fab-Apo data, MoRDa (Keegan and Winn (2007) *Acta Cryst.* D63, 447-57; Keegan et al. (2018) *Acta Cryst.* D74, 167-82; and Winn et al. (2011) *Acta Cryst.* D67, 235-42) was used to place two copies of a Fab fragment as a template (PDB ID: 5AZE). For the Fab-mutCALR peptide data Phaser was used to place fragments of the Fab-Apo model and subsequent rounds of manual model building were performed resulting in a model containing two Fabs and 3 peptides. For both datasets, subsequent cycles of density modifications, model building and refinement were carried out using Refmac (Murshudov et al. (1997) *Acta Cryst.* D53, 240-255), Coot (Emsley et al. (2010) *Acta Crystallographica Section D \u2013 Biological Crystallography*, 66, 486-501) and Phenix (Liebschner et al. (2019) *Acta Cryst.* D75, 861-77) until the structures converged at reasonable R-work and R-free (Table 14). The final models were analyzed for good stereochemistry, geometry and clash scores using MolProbity (Table 14).\n\n**Results:**\n\nThe structure of Fab-Apo was determined to 2.9 \u00c5 and consists of residues Light chains 2-213, and Heavy chains 2-219 (2 Fabs) and has good electron density throughout, with the exception of residues Chain L: Asp28-Tyr34, Chain H: Ser99-Gly101, Lys134-Gly139, Chain M: Gly24-Tyr34, Chain I: Ser133-Gly138 which are not modelled due to weak electron density. The structure was refined to an R-work/R-free of 23% and 30%, respectively, and has good stereochemistry throughout with 4 Ramachandran outliers which is acceptable for this resolution (48th percentile for this residue range, Table 14).\n\nThe structure of Fab-mutCALR peptide was determined to 3.2 \u00c5. The two Light chains consist of residues Chain L: 2-215, Chain M: 2-216; and two Heavy chains H: 2-219 and I: 2-185. There are two Fabs and two peptides in the asymmetric unit. The model has good electron density throughout, with the exception of the following residues: Chain H, Ser132-Gly139; Chain M, Thr25-Gly31; and Chain I, Ala130-Ala142, which are not modelled due to weak electron density. The structure was refined to an R-work/R-free of 23% and 32%, respectively, and has good stereochemistry throughout with 7 Ramachandran outliers which is acceptable for this resolution (56th percentile for this residue range, Table 14).\n\n*Table 14.* Data collection statistics, phasing and refinement. (Values in parentheses are for the highest-resolution shell.)\n\n| | Fab fragment | Fab fragment:CalR peptide |\n|---|---|---|\n| Resolution range (\u00c5) | 54.52 \u2013 2.9 (3.004 \u2013 2.9) | 58.3 \u2013 3.2 (3.314 \u2013 3.2) |\n| Space group | P 1 21 1 | P 21 21 21 |\n| Unit cell | 79.921 51.61 109.702 90 96.306 90 | 40.646 121.228 199.482 90 90 90 |\n| Total reflections | 38451 (3220) | 32029 (2718) |\n| Unique reflections | 19565 (1638) | 16045 (1359) |\n| Multiplicity | 2.0 (2.0) | 2.0 (2.0) |\n| Completeness (%) | 97.16 (82.11) | 93.80 (82.11) |\n| Mean I/sigma(I) | 7.97 (2.29) | 13.31 (6.61) |\n| Wilson B-factor | 36.39 | 39.75 |\n| R-merge | 0.08796 (0.3616) | 0.05226 (0.1321) |\n| R-meas | 0.1244 (0.5113) | 0.0739 (0.1868) |\n| R-pim | 0.08796 (0.3616) | 0.05226 (0.1321) |\n| CC1/2 | 0.983 (0.677) | 0.992 (0.929) |\n| CC* | 0.996 (0.899) | 0.998 (0.981) |\n| Reflections used in refinement | 19559 (1638) | 16040 (1359) |\n| Reflections used for R-free | 1652 (138) | 786 (61) |\n| R-work | 0.2301 (0.2879) | 0.2288 (0.2838) |\n| R-free | 0.3032 (0.3904) | 0.3174 (0.3856) |\n| CC(work) | 0.905 (0.757) | 0.899 (0.803) |\n| CC(free) | 0.776 (0.542) | 0.771 (0.623) |\n| Number of non-hydrogen atoms | 5916 | 6101 |\n| macromolecules | 5916 | 6101 |\n| Protein residues | 825 | 824 |\n| RMS(bonds) | 0.010 | 0.012 |\n| RMS(angles) | 1.38 | 1.41 |\n| Ramachandran favored (%) | 95.79 | 91.17 |\n| Ramachandran allowed (%) | 3.72 | 7.96 |\n| Ramachandran outliers (%) | 0.50 | 0.87 |\n| Rotamer outliers (%) | 0 | 0 |\n| Clashscore | 9.83 | 14.96 |\n| Average B-factor | 29.20 | 30.27 |\n| macromolecules | 29.20 | 30.27 |\n\nAs shown in FIG. 12A, the asymmetric unit of the crystal structure includes one Fab molecule bound to two mutCALR peptides (referred to as CalR1 and CalR2). The structure includes 2 light chains (shown in white), 2 heavy chains (shown in black) and 2 identical peptides representing a portion of the mutant CALR C-terminal domain (CalR1, gray and CalR2, white). The Fab binds to the two mutCALR peptides in two distinct binding conformations (FIG. 12A). Only Fv regions of the Fab are displayed for clarity.\n\nFIG. 12B shows the composition of the CDR loops (L1, L2, L3, H1, H2, and H3), the amino acid composition, and the length of the Fab. CDR predictions performed using the CCG scheme in MOE (Molecular Operating Environment). Amino acids highlighted in *bold/italics* are noted for significant contributions to CalR1 peptide binding as detailed in FIGs. 13B-13C. Amino acids highlighted in **bold/underlined** are noted for significant contributions to CalR2 peptide binding as detailed in FIGs. 14B-14C. Amino acids denoted by an asterisk contribute to both CalR1 and CalR2 binding. FIG. 12C shows the arrangement of CalR1 (gray) and CalR2 (white) in the CDR region of Fab1 (top) and the arrangement of the CDR loops (L1, L2, L3, H1, H2, and H3) and orientation of CalR1 (gray) and CalR2 (white) in the CDR region of Fab1 (bottom).", "# Approaches to automated protein crystal harvesting\nAuthors: Deller, M. C.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 70, 133\u2013155 (IYCr crystallization series), 2014\n\nThe lower theoretical limit for a crystal usable for X-ray structure determination is around 1.2 \u00b5m depending on unit-cell dimensions. Crystals of this size appear routinely in screening but are usually written off as hits for follow-up fine screening rather than used directly, and for high-value targets such as protein\u2013protein and protein\u2013nucleic acid complexes, virus assemblies and membrane proteins that optimization can take years, if it works at all. Using the microcrystals directly on microfocus beamlines and FEL sources is the attractive alternative, but manual harvesting of microcrystals is exceptionally difficult and nearly impossible in the single-digit micrometre range, where crystals are hard to separate and excess mother liquor is hard to remove. Micromeshes have been widely adopted because excess liquor can be wicked away or replaced with oil and many crystals spread across a mesh can be located by raster scanning. FEL data have already given structures of lysozyme, mimivirus particles, photosystem I and photosystem II, but the high energies destroy the crystal, so complete data sets need large numbers of them. ADE addresses rapid mounting at scale, and the authors were developing UMR to harvest into SSRL grids \u2014 200 \u00b5m thick polycarbonate with 75 laser-etched holes of 125, 200 and 400 \u00b5m diameter \u2014 allowing 1,200 crystals per UniPuck and 8,400 per shipping dewar.\n\n### 4.2. Reduced late-stage failure and improved cryoprocedures\n\nLate-stage crystal failures are crystals that failed to diffract well enough for structure determination because of physical or chemical damage during harvesting and mounting or from poor cryoprotection. Their true number is very hard to establish because it requires knowing the diffraction state before and after manipulation, and few systematic studies exist, though in situ methods have been used to assess crystal quality beforehand. Given the hundreds of thousands of crystals screened each year, rescuing even a small percentage would materially raise structure-determination rates. Cryoprotection and cooling protocols have changed little since loop-based harvesting began decades ago, and at the JCSG 80% of harvested crystals require added cryoprotectant to prevent damaging hexagonal ice, which reduces resolution and increases mosaicity. Hyperquenching \u2014 cooling at rates above 10,000\u00b0C s\u207b\u00b9 so that amorphous rather than hexagonal ice predominates \u2014 is the proposed alternative, and the UMR cryo-drip method promotes it in two ways: removing excess mother liquor from the microtool and crystal, and coating the crystal in a thin oil layer that insulates it from the slow film-boiling regime and gas cushion of liquid nitrogen, raising the Leidenfrost temperature so cooling occurs in the more efficient nucleate boiling regime. High-speed robot motion through the gas cushion reinforces the effect. A 22% reduction in cryoprotectant has been demonstrated with hyperquenching, and the authors suggest still faster robotics might remove the need for added cryoprotectant altogether.\n\n### 4.3. Increased throughput, and 4.4 final appraisal\n\nA conservative estimate puts manual harvesting and screening at hundreds of thousands of crystals per year: the PDB was growing by about 7,000 structures annually and JCSG experience suggested roughly 100 crystals screened per deposited structure, giving of the order of 700,000 crystals screened per year \u2014 probably an underestimate, since challenging targets in traditional laboratories may need far more than 100 crystals and the global cryopin market is estimated at around two million per year. Automation offers to streamline this, and because HTPX/SG contributes about 10% of PDB depositions, closing the bottleneck at large centres should reduce late-stage failures and raise both throughput and output, with methods developed there flowing down to small laboratories. In the final appraisal, Marc Deller and Bernhard Rupp observe that every system reviewed is engineered around one problem \u2014 manipulating crystals despite their inherently poor structural, physical and chemical stability \u2014 and that no single approach has yet mastered the mechanics of the manipulation step, so a universally applicable solution will probably combine several technologies into a true gene-to-structure platform. Of the roughly 14 documented systems, most are proof-of-concept or demonstration systems no longer in routine use and not commercially available; components of FOT and ADE are commercial, but REACH/G-Rob and UMR were, to the authors' knowledge, the only fully integrated semi-automated harvesting systems then commercially available. They conclude that, given the scale of global crystal production and screening, it is remarkable that most harvesting is still done by hand.\n\n## Key tables\n\n*Table 1.* Benefits of automated crystal harvesting (adapted from Viola, Carman, Walsh, Frankel et al., 2007).", "# From Noise to Nanomolar: PXDesign\u2019s New Benchmark for Protein Binder Design\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nThe experimental results make a strong case for this integrated approach. Against PD-L1, 14 of 17 tested designs bound with KD below 1 \u03bcM, and most came in well below 100 nM \u2014 compared to the strongest BindCraft binder against that same target, which reached approximately 1 \u03bcM. On the SARS-CoV-2 RBD, the hit rate rose from 20% in an earlier version of the pipeline to 50% following generation improvements, with multiple binders confirmed at sub-10 nM affinity by biolayer interferometry. TNF-\u03b1 remains an outstanding challenge \u2014 a notoriously flat and symmetric interface that stymied AlphaProteo as well \u2014 which honestly makes the performance elsewhere more impressive rather than less. The full pipeline is freely accessible via a public [web server](https://protenix-server.com/) and an open-source GitHub repository, removing the compute and engineering barrier for groups wanting to run their own binder campaigns. As structure prediction and generative design continue to converge, PXDesign offers a clear signal that the computational-to-experimental pipeline for protein binder discovery is compressing from months to days.", "# GitHub README: chatRDKit\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/chatRDKit, 2026\n\n- chatRDKit executes whatever code the model returns, in-process, in the\n  same Python interpreter. A basic safety check refuses snippets containing\n  `import os/sys/subprocess/shutil/socket/requests`, `open(`, `eval(`,\n  `exec(`, `__import__`, or direct `os.`/`sys.`/etc. attribute access -- the\n  system prompt also instructs the model not to use these. This is a\n  best-effort net, not a sandbox: review echoed `[chatRDKit] >>> ...` lines\n  for anything unexpected.\n- Everything runs locally against your Ollama server -- no data leaves your\n  machine.\n- `/config debug on` prints the model's raw output (including any stripped\n  `<think>` blocks or code fences) so you can see exactly what was executed.\n\n---\n\n## 6. Troubleshooting\n\n**`Could not reach Ollama at http://localhost:11434 ...`**\nStart Ollama (`ollama serve` or open the desktop app) and confirm the model\nis pulled (`ollama pull qwen3.6:27b`, check with `ollama list`).\n\n**`Ollama at ... did not respond to model '...' within 180s`**\nThe model took too long to respond. Common causes: the model is still\nloading into memory (try again), the machine is under heavy load (e.g.\nanother `ollama pull` running), or the model is just slow for your\nhardware. Increase `/config timeout 300` or switch to a smaller/faster\nmodel.\n\n**`The model's response looked malformed or repetitive ... and was\ndiscarded`**\nSmall models occasionally loop or ramble instead of producing a short\nsnippet. chatRDKit detects this, discards the response, and does *not* add\nit to conversation history (so it can't poison later requests). Try again,\nrephrase, or switch to a larger model (`/config model qwen3.6:27b`).\n\n**`Refusing to run this snippet -- it contains ...`**\nThe model's response touched something outside RDKit (file I/O, `os`,\n`subprocess`, etc.). Rephrase the request -- chatRDKit's `render()` helper\nalready handles all file output.\n\n**`AttributeError: module 'rdkit.Chem...' has no attribute '...'`**\nSmaller models sometimes invent plausible-but-wrong RDKit function names.\nThis is reported as a normal Python error without crashing the session --\ntry rephrasing, or switch to `qwen3.6:27b` for better accuracy.\n\n**PDF output falls back to PNG**\nInstall `reportlab` (`pip install reportlab`) for PDF report support.\n\n---\n\n## 7. Extending\n\n- **More commands / better translations**: extend\n  `rdkit_reference.RDKIT_CHEAT_SHEET` with examples for operations you use\n  often.\n- **Richer session context**: `rdkit_reference.build_context()` is the place\n  to add more state (e.g. 3D conformers, reaction history).\n- **New output formats**: add a branch to `render.render()` alongside\n  `_write_html` / `_write_pdf`.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/chatRDKit\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/chatRDKit</a></td>\n</tr>\n</table>", "# Therapeutic Enzymatic Compositions and Delivery Methods for the Degradation of Microplastics and Nanoplastics in Human Physiology\nAuthors: Deller, Marc C. (inventor)\nSource: United States Provisional Patent Application (Elora Therapeutics), 2025\n\nThe invention provides therapeutic enzymatic compositions and novel delivery methods designed specifically for internal human degradation of the most common forms of microplastics and nanoplastics \u2013 to include polyethylene terephthalate (PET), polyethylene (PE), polypropylene (PP), and polystyrene (PS) \u2013 residing in human blood and tissues. Engineered enzyme variants including polyethylene terephthalate hydrolase (PETase), mono(2-hydroxyethyl) terephthalate hydrolase (MHETase), bacterial-origin CotA-Laccase, and Horseradish Peroxidase (HRP) are optimized for therapeutic applications to degrade these microplastics and nanoplastics, addressing metabolic, immunogenic, and stability challenges unique to human physiology. Delivery methods include oral encapsulation techniques and injectable PEGylation, lipid nanoparticles, and hydrogel formulations, tailored explicitly for human therapeutic use.", "# My Work Driving Innovation in Precision Oncology Through Structure-Based Drug Design\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nCalreticulin mutations represent the second most common oncogenic driver in [myeloproliferative neoplasms](https://marcdeller.com/anti-mutant-calreticulin-calr-antibodies-and-uses-thereof/), present in approximately 25% of essential thrombocythemia and 35% of myelofibrosis patients who lack JAK2-V617F mutations, creating constitutive activation of the thrombopoietin receptor through unique mutant protein sequences absent in normal cells. The development of [INCA033989](https://marcdeller.com/anti-mutant-calreticulin-calr-antibodies-and-uses-thereof/), a fully human IgG1 monoclonal antibody specifically targeting mutant calreticulin, represents a precision immunotherapy approach that selectively eliminates cancer cells expressing the mutation without affecting normal hematopoiesis. Preclinical validation demonstrated that INCA033989 binds mutant calreticulin on the cell surface, blocks the mutant CALR-MPL interaction, inhibits JAK-STAT signaling selectively in patient-derived CD34+ cells while preserving signaling in healthy donor cells, and induces dynamin-dependent endocytosis of the antibody-receptor complex. In competitive transplant mouse models, ten weeks of INCA033989 treatment prevented thrombocytosis, significantly decreased mutant calreticulin stem and progenitor cells in bone marrow, and critically targeted disease-propagating hematopoietic stem cells, as evidenced by the absence of MPN development in secondary transplantation, demonstrating disease-modifying potential beyond symptom management. Clinical translation is advancing through a [Phase 1 trial](https://www.cancer.gov/research/participate/clinical-trials/intervention/anti-mutant-calreticulin-monoclonal-antibody-inca033989) in patients with CALR-mutant essential thrombocythemia, where early data presented in 2024 showed an 86% response rate, representing the first therapeutic antibody designed to selectively target and eliminate specific cancer-driving mutations in myeloproliferative disorders.\u200b\n\n**JAK2-V617F Selective Inhibition: Achieving Molecular Remission**\n\nThe [JAK2-V617F mutation](https://clinicaltrials.gov/study/NCT06213818) drives nearly all cases of polycythemia vera and over half of essential thrombocythemia and myelofibrosis, yet current JAK inhibitors targeting the kinase domain provide symptomatic benefit without addressing mutant allelic burden or achieving molecular remission. [INCB160058](https://clinicaltrials.gov/study/NCT06213818) represents a first-in-class approach using structure-guided molecular design to bind with picomolar affinity to the pseudokinase JH2 domain of JAK2-V617F, achieving over 2,500-fold selectivity relative to the wild-type JAK2 kinase domain targeted by approved therapies. This pseudokinase-binding mechanism blocks ligand-independent thrombopoietin receptor dimerization induced by the V617F mutation, resulting in loss of downstream kinase activity while preserving cytokine-dependent wild-type JAK2 signaling essential for normal hematopoiesis. X-ray crystallography analysis revealed the inhibition mechanism involves conformational disruption of the \u03b1C helix motif upon INCB160058 binding to the JH2 domain, providing atomic-level understanding of selectivity determinants. Preclinical validation in patient-derived models demonstrated that continuous exposure to INCB160058 at sub-IC50 concentrations progressively eliminated JAK2-V617F-positive cells in co-cultures without affecting wild-type cells, and in patient-derived xenograft models selectively reduced human JAK2-V617F cell engraftment and hematopoietic stem/progenitor cells while sparing healthy donor cell populations. The compound normalized pathogenic cytokines including interleukin-6 and interleukin-8, and importantly demonstrated tolerability with significant antitumor activity in vivo, supporting [clinical development](https://www.cancer.gov/research/participate/clinical-trials-search/v?id=NCI-2024-03936) initiated in 2024 in patients with JAK2-V617F-positive myeloproliferative neoplasms resistant or refractory to current JAK inhibitors, with the therapeutic goal of achieving molecular remission through selective eradication of mutant clones.", "# BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nBoltzMaker consolidates five smaller tools I vibe-coded earlier while learning to vibe, each doing exactly one job: [generate_yaml](https://github.com/bellcheddar/generate_yaml) for building Boltz input YAMLs, simple-zsh-script-to-run-boltz2\u00a0for driving prediction runs, [analyze-boltz2-results](https://github.com/bellcheddar/analyze-boltz2-results) for post-run analysis, [cif2plip](https://github.com/bellcheddar/cif2plip) for protein-ligand interaction profiling, and [smiles2grid](https://github.com/bellcheddar/smiles2grid) for rendering a ligand set as a grid of 2D structures. Each vibe-coded project taught me one piece. BoltzMaker shares a single campaign format across all five stages, so the same input file drives every step rather than retyping target names between separate scripts.\n\nThe pipeline is five stages, and each stage reads only what the previous one wrote, so any stage can be re-run on its own:\n\n```\n\n\t\t\t\t\t+----------------+     +----------+     +-----------+     +-----+     +---------+\n| boltz_input.md | --&gt; | generate | --&gt; | preflight | --&gt; | run | --&gt; | analyze |\n+----------------+     +----------+     +-----------+     +-----+     +---------+\n        |\n        |  opt-in, separate: needs an \"Apo structure:\" reference\n        v\n +-------------+\n | compare-sse |\n +-------------+\n\nsetup        create the managed .venv (Python 3.12; boltz pins numpy&lt;2.0)\nsetup-plip   optional .plip_env for cif2plip/PLIP interaction profiling\nnew          interview in plain English -&gt; writes a valid boltz_input.md\nformat       auto-align comments/blank-lines (cosmetic only; --check to dry-run)\ngenerate     write per-target YAMLs + a manifest\npreflight    PASS / WARN / FAIL: CLI, GPU/MPS, disk, YAML, SMILES, chemistry, memory\nrun          boltz predict, live progress + memory monitor, resumable\nanalyze      boltz_summary.csv / .xlsx, boltz_dashboard.html, interaction files\nall          generate -&gt; preflight -&gt; run -&gt; analyze\ncompare-sse  apo vs holo motif shifts -&gt; boltz_sse_comparison.csv / .html / .pml\n\n```\n\nThe workflow is as follows, and the reasoning behind the roadmap lives in [PROJECT_PLAN.md:](https://github.com/bellcheddar/BoltzMaker/blob/main/PROJECT_PLAN.md)\n\n1. `generate` writes the per-target YAMLs and a manifest.\n2. `preflight` is the sanity gate: it checks that the CLI is present, that a GPU or Apple MPS backend is available, that there is disk and memory headroom, and that every YAML, SMILES string, and chain ID is valid.\n3. `run` calls `boltz predict` with a live progress bar and resumes cleanly if it is interrupted.\n4. `analyze` turns the outputs into a CSV, an XLSX workbook, and a self-contained HTML dashboard.\n\nThe ergonomic core is a family, partners, and ligand cross-product: write each protein and each ligand once, and BoltzMaker generates every combination as its own target. That is exactly the shape of a real selectivity or SAR panel, which is what makes the spec worth writing in the first place.\n\n###### **Running BoltzMaker**\n\nEverything installs into a dedicated virtual environment through one setup command:\n\n```\n\n\t\t\t\t\tpython3 BoltzMaker.py setup\n\n```\n\nAvailable switches include `all,` which runs the full pipeline starting from a pre-written boltz_input.md, and `new,` which invokes a wizard that interviews you (proteins, partners, ligands, and the three constraint sentence types) and writes the file for you. It covers only the common cases; rarer fields (modifications, cyclic, MSA override, templates, homo-oligomer copies) are left for hand-editing the file it writes.\n\n```\n\n\t\t\t\t\tpython3 BoltzMaker.py setup                         # create the managed .venv        (--force, -y/--yes)\npython3 BoltzMaker.py setup-plip                    # optional .plip_env for interactions (--force)\npython3 BoltzMaker.py new        [boltz_input.md]   # write a new campaign by answering plain questions\npython3 BoltzMaker.py format      boltz_input.md    # auto-align comments/blank-lines, cosmetic (--check)\npython3 BoltzMaker.py generate    boltz_input.md    # write the target YAMLs + manifest\npython3 BoltzMaker.py preflight   boltz_input.md    # environment + input sanity checks\npython3 BoltzMaker.py run         boltz_input.md    # boltz predict, live progress, resumable\npython3 BoltzMaker.py analyze     boltz_input.md    # CSV / XLSX / HTML dashboard\npython3 BoltzMaker.py all         boltz_input.md    # generate -&gt; preflight -&gt; run -&gt; analyze\npython3 BoltzMaker.py compare-sse boltz_input.md    # opt-in: apo vs holo secondary-structure shifts\npython3 BoltzMaker.py             boltz_input.md    # same as `all` (subcommand is optional)\n\nSETUP &amp; FORMAT OPTIONS\n======================\n\nOption                        Applies to           Description\n----------------------------  -------------------  --------------------------------------------------\n--force                       setup, setup-plip    Recreate the .venv / .plip_env from scratch\n-y, --yes                     setup                Skip the download-confirmation prompt\n--check                       format               Report whether reformatting is needed without\n                                                   writing (exit 1 if so)\n\nCOMMON OPTIONS\n==============", "# 2025 \u2013 Building Data and Project Management Expertise\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-21 on marcdeller.com (categories: certificates).*\n\n* Master core data analysis concepts for informed decision-making.\n* Develop R programming proficiency for statistical computing.\n* Apply structured project management methodologies.\n* Leverage Shiny dashboards for actionable insights.\n* Integrate generative AI into retail analytics.\n* Gain cloud-native architecture skills on AWS.\n* Demonstrate cross-disciplinary expertise to employers.\n\n1. **Data Literacy and Analysis Foundations** \u2013 \u00a0Data, Data, Everywhere: [Certificate](https://www.coursera.org/account/accomplishments/verify/4S09Q8ICKGXN)\n2. **R Programming Basics** \u2013 \u00a0Introduction to R: Basic R Syntax: [Certificate](https://www.coursera.org/account/accomplishments/verify/T687VG5UYI5O)\n3. **Project Management Fundamentals** \u2013 Foundations of Project Management: [Certificate](https://www.coursera.org/account/accomplishments/verify/605BGOEB1YGC)\n4. **Retail Analytics with AI \u2013 Dashboard Development with Shiny GenAI for Retail Analysis**: [Certificate](https://www.coursera.org/account/accomplishments/verify/OO09W4LMZGCB)\n5. **Cloud Architecture** \u2013 Serverless Architectures on AWS: [Certificate](https://www.coursera.org/account/accomplishments/verify/X9AD2JWH2HE9)\n\nIn an era driven by insights and execution, certifications that bolster both data literacy and project management principles unlock unparalleled career opportunities. Starting with *Foundations: Data, Data, Everywhere* and *Introduction to R: Basic R Syntax*, you establish the analytical skills essential for evidence-based decision-making. Coupling these with *Foundations of Project Management*and *Dashboard Development with Shiny: GenAI for Retail Analysis* delivers a dual toolkit for planning and visualization. Finally, the AWS-endorsed *Serverless Architectures on AWS*certification equips you to architect scalable, cost-effective solutions\u2014making you a linchpin in the modern technical landscape.", "# Revolutionizing Drug Discovery: The Click2Drug Platform and the Rise of AI-Powered Virtual Screening\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Concept | Description | Key References |\n| --- | --- | --- |\n| Click2Drug Platform | Comprehensive directory of 807+ CADD tools, databases, and web services categorized across the drug design pipeline | [Click2Drug Directory](https://www.click2drug.org) |\n| Ultra-Large Virtual Screening | AI-powered platforms capable of screening billion-compound libraries with 10-44% hit rates in under 7 days | [Meller, A., et al., Nature Communications (2024)](https://www.nature.com/articles/s41467-024-52061-7) |\n| Deep Learning Pharmacophores | Neural network-based pharmacophore modeling enabling ultra-fast virtual screening without explicit docking conformations | [PharmacoNet, Chemical Science (2024)](https://pubs.rsc.org/en/content/articlelanding/2024/sc/d4sc04854g) |\n| Conformal Prediction | Machine learning framework controlling error rates in virtual screening, reducing compounds requiring docking by 1000-fold | [Hoffman, S.C., et al., Nature Machine Intelligence (2025)](https://www.nature.com/articles/s43588-025-00777-x) |\n| AI-Guided Drug Design | Artificial intelligence systems achieving 80-90% Phase I success rates compared to 40-65% traditional approaches | [Lifebit AI Drug Discovery Review (2025)](https://lifebit.ai/blog/ai-driven-drug-discovery/) |\n| Receptor Flexibility Modeling | Advanced docking algorithms incorporating protein conformational changes and allosteric regulation effects | [Shoichet, B.K., et al., Nature (2024)](https://www.nature.com/articles/s44386-024-00001-2) |\n| Chemical Space Expansion | Ultra-large databases like ZINC-22 (54.9 billion compounds) and REAL (6.7 billion synthesizable molecules) | [Cheminformatics Databases Review (2025)](https://neovarsity.org/blogs/most-used-cheminformatics-databases) |\n| Machine Learning Integration | Graph neural networks and transformer architectures optimizing molecular properties and synthetic accessibility | [Chen, L., et al., MedChemComm (2024)](https://pubs.rsc.org/en/content/articlelanding/2024/md/d4md00522h) |", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices\nAuthors: Deller, M.C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nSite 2 interactions with gp130 were analysed in a similar fashion using multiple simultaneous substitutions due to the weaker interactions involved (Hudson et al., 1996). Residues Gln 25, Ser 28, Gln 32, Ile 121, Gly 124 and Ser 127 have all been shown to participate in gp130 binding. These residues cluster at the ends of helices A and C as shown in yellow in Figure 6-10 (a) below. Non-conservative mutations of these residues result in loss of biological action and act as antagonists of wildtype LIF action.\n\nStructurally equivalent residues in the OSM structure are Gln 16, Lys 19, Leu 23, Asn 117, Gly 120 and Asn 123 as shown in yellow in Figure 6-10 (b) below. The importance of Gln 16, Gly 120 and Asn 123 in the formation of the signalling complex of LIFR has been confirmed through mutagenesis, which reveals that mutation of all of these residues results in an 8 to 100-fold reduction in binding to gp130 and a corresponding reduction in biological activity, with no effect on LIFR binding (Deller et al., 2000).\n\n#### 6.3.2 Interleukin-6\n\nIn common with LIF, mutagenesis studies of IL-6 (Simpson et al., 1997a) also suggest a three site receptor binding model. However, in contrast to LIF these three sites function in binding three separate receptor components.\n\nIt has been shown that sites 2 and 3 interact independently with gp130 only in the presence of IL-6R (Paonessa et al., 1995). Immunoprecipitation (Paonessa et al., 1995) studies have suggested that the complex contains two copies of each component and a hexameric complex with a stoichiometry of 2:2:2 has been confirmed by analytical ultracentrifugation and size exclusion chromatography (Ward et al., 1994).\n\n##### 6.3.2.1 Site 1 Interactions\n\nThis site forms a cleft on the surface of IL-6 and all mutations that affect binding to IL-6R map to this surface, which is composed of residues in helix D and the A-B loop as shown in blue in Figure 6-10 (c) below. Residues of key importance include Arg 179, which results in 100-fold decrease in activity (Fontaine et al., 1993), and Gln 175, which results in 5-fold decrease in activity (Savino et al., 1993). Other solvent exposed residues implicated in binding IL-6R include Ser 176 (Savino et al., 1993), Arg 182 (Lutticken et al., 1991) and Phe 74.\n\nAs for the other cytokines, site 1 in IL-6 is thus implicated as a binding site for a specificity conferring receptor in the signalling complex. This region in OSM corresponds to residues Arg 182, His 178, Ser 179, Ser 185 and Arg 63 as shown in blue in Figure 6-10 (d) below. The function of these residues is currently unknown although all these residues, except Ser 179 are solvent exposed and are likely to form a site 1 epitope, which by analogy with LIF might be expected to interact with LIFR or OSMR.\n\n##### 6.3.2.2 Site 2 Interactions\n\nAs in LIF site 2 interacts independently with a gp130 receptor and consists of residues in helix A and helix C as shown in yellow in Figure 6-10 (c) below. Important residues in this location include Tyr 31, Ser 118 and Val 121, which are all solvent exposed.\n\nStructurally equivalent residues in the OSM structure are Leu 23, Asn 117 and Gly 120 as shown in yellow in Figure 6-10 (d) below. All of these residues have been shown to be important in varying degrees for OSM binding to gp130 as detailed in Table 6-3.\n\n##### 6.3.2.3 Site 3 Interactions\n\nIn analogy with LIF site 3 consists of residues in helix D and the C-D loop; however in IL-6 both sites 2 and 3 interact independently with separate gp130 receptors. Residues important for site 3 binding to the second gp130 receptor, include the solvent exposed residues Trp 157 and Asp 160 (Paonessa et al., 1995) as shown in red in Figure 6-10 (c) below.\n\nStructurally equivalent residues in the OSM structure are Phe 160 and Lys 163 as shown in red in Figure 6-10 (d) below, and both of these residues have been shown to form the major determinant of OSM binding to LIFR as detailed in Table 6-3 below.\n\n##### 6.3.2.4 Site 4 Interactions\n\nOn the basis of the crystal structure of IL-6 (Somers et al., 1997) it has also been predicted that a fourth site on the IL-6 exists that mediates dimer formation between two IL-6 molecules. It has been proposed that the region Glu 106 to Arg 113, shown in pink in Figure 6-10 (b) below, would interact with the same residues on an adjacent molecule, across a 2-fold axis of rotation, so stabilizing the hexameric signaling complex.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices\nAuthors: Deller, M.C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nThe minor interface formed between GCSF and GCSFR is not equivalent to interactions observed in the Epo or GH system as it involves residues on the face of the C-terminal domain of GCSFR and residues at the N-terminal end of the helical bundle of GCSF; namely Ser 7 and Gln 11 of GCSF and Phe 259 and His 260 of GCSFR.\n\nTherefore, the unusual minor site interaction of GCSF with its receptor may provide a useful basis for understanding non-standard epitopes, such as the site 3 motif of the LIFR binding family of cytokines. For example, the Ig-like domain of gp130 has been shown to bind OSM through site 3, so suggesting the formation of a tetrameric assembly of two OSM, one LIFR and one gp130 molecule as shown in Figure 6-11.\n\nA comparison of the themes observed in the Epo, GCSF and OSM signalling systems is given below in Figure 6-11.\n\n*Figure 6-11.* Figure showing ribbon representation of (a) Epo-EpoR (Syed et al., 1998) and (b) GCSF-GCSFR (Aritomi et al., 1999; coordinates kindly provided by K. Morikawa, Biomedical Engineering Research Council, Osaka, Japan) crystal structures. In each case the haemopoietin receptor domains are coloured in blue and the cytokines are coloured in green. Panel (c) shows Plasticene model of proposed OSM-LIFR-gp130 complex (Staunton et al., 1998) observed interactions shown in hypothetical orientation. Figure produced using BOBSCRIPT (Esnouf, 1997), RASTER3D (Merritt and Murphy, 1994), Plasticene (Toys R Us) and Anglepoise lighting.\n\n### 6.4 Sequence Comparisons\n\nDespite the common framework amongst the helical cytokines essentially no sequence identity is observed across the family as a whole. However, core residues are conserved within the gp130 family of cytokines and this gives rise to low but significant levels of sequence identity and conservation as shown in Figure 6-12 below. Typically, members of the gp130 family show ~20 % pairwise identity or conservative substitution. Approximately 50 % of these conserved residues are inaccessible to solvent, and as such contribute to the molecular core so maintaining the protein fold and stability. The remaining residues are solvent exposed and as such represent potential sites of receptor engagement.\n\nFigure 6-12 also demonstrates the conserved nature of the Phe/Lys (residues 160 and 163, respectively in OSM) pair in the LIFR binding family of cytokines. This pair is invariant over all species of OSM and is conserved between OSM, LIF and CNTF. These residues are not observed in IL-6 and are partially conserved in both Epo and GCSF, which contain a Phe/Leu and Phe/Arg pair, respectively. This suggests that GCSF and Epo have diverged from the LIFR binding family of cytokines, as they place less importance on the use of a high affinity site 3 epitope and rely solely on a classic hydrophobic site 1 and 2 interaction as first observed in the growth hormone system.\n\nGiven the importance of these residues in LIFR engagement this level of conservation again supports an evolutionary link between the molecules and together with the similar intron/exon boundaries that are observed for these cytokines suggests an homologous relationship (Bazan, 1991).\n\n*Figure 6-12.* Figure showing sequence alignments of the gp130 family of cytokines and GCSF and Epo. Coils represent secondary structure assignment of OSM and blue blocks represent solvent accessibility ranging from white (~0 %) through light blue (~50 %) to dark blue (~100 %). Yellow circles represent site 2 residues implicated in gp130 binding and red circles represent site 3 residues implicated in LIFR binding (Deller et al., 2000). Green circles represent residues structurally equivalent to residues implicated in modulating species specificity in LIF (Robinson et al., 2000). Figure produced using DSSP (Kabsch and Sander, 1983) and ESPript (Gouet et al., 1999).\n\n### 6.5 Implications for Receptor Binding\n\n#### 6.5.1 Implications for Species Specificity\n\nAs discussed previously both murine LIF and human LIF interact with human LIFR, whereas, murine LIFR interacts only with murine LIF with appreciable affinity (Layton et al., 1994a; Layton et al., 1994b; Owczarek et al., 1993). From analysis of murine-human chimeras the region responsible has been localised to residues 99\u2013180 (Owczarek et al., 1993). From analysis of the crystal structures of murine LIF (Robinson et al., 1994) and human LIF (Robinson et al., 2000) it has been proposed that a change in surface character of helix D, which is ideally placed to interact with LIFR, is responsible for the species anomaly.\n\nThe surface volume occupied by residue Met 64 in murine LIF is displaced by residue Lys 168 in human LIF, due to perturbation of the A-B loop (Robinson et al., 2000). It is proposed that this change in surface character from hydrophobic to charged modulates the specificity of LIF towards LIFR. Although the species specificity of OSM has not been fully investigated it is plausible that a change in the surface character of helix D could be involved given the high number of charged residues residing on the face of this helix. This may prove to be another example of the importance of electrostatic interactions in ligand-receptor docking as discussed in 6.3.3.1.\n\nAnalysis of the OSM structure reveals that structural equivalents to murine LIF Met 64 and human LIF Lys 168 are human OSM residues Pro 53 and Gly 172, respectively. This precludes these residues from involvement in such a species 'switch' mechanism in OSM as glycine has no surface character and it is unlikely that a proline side-chain will be displaced due to its rigid nature. Sequence alignments reveal that Gly 172 is conserved across all species of OSM, whereas Pro 53 is replaced by serine in murine OSM as identified by green circles in Figure 6-12.\n\nOther local or long-range conformational changes in this region are not precluded and further crystallographic studies of other OSM and LIF species variants will be required to address the problem.", "# GitHub README: PANTS\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/PANTS, 2026\n\n## \ud83d\udce6 Dataset\n\nThe catalogue is deposited on Zenodo under CC BY 4.0: candidates with their retrieval\nscores and active-site geometry, the reference set split by evidence tier, the measured\nactivity data with its citations, every pipeline run with its discard counts, and the\npredicted structures.\n\n[](https://doi.org/10.5281/zenodo.21807353)\n\n| | |\n|---|---|\n| Concept DOI (always latest) | [10.5281/zenodo.21807353](https://doi.org/10.5281/zenodo.21807353) |\n| This version (v0.1.0) | [10.5281/zenodo.21807354](https://doi.org/10.5281/zenodo.21807354) |\n\nEvery figure in the deposit's description and datasheet was **recomputed from the\ndeposited files** by `scripts/build_release.py` rather than copied from this README.\nThat is deliberate: on a sibling project six figures reached a Zenodo description by\nbeing copied from prose describing a checkpoint that was measured but never shipped, and\na DOI would have made them permanent.\n\n**Cite as:** Deller, M. C. (2026). *PANTS: a triaged catalogue of candidate PET-degrading\nenzymes for therapeutic conditions*. Version 0.1.0. Zenodo. https://doi.org/10.5281/zenodo.21807353\n\n## \ud83d\udcdd Licence\n\n**Data:** Creative Commons Attribution 4.0 International, see [`LICENSE-DATA`](LICENSE-DATA),\nand via the Zenodo deposit above.\n\n**Code:** MIT, see [`LICENSE`](LICENSE).\n\nAttributing PANTS does not discharge the obligation to the sources it derives from: UniProt\nand UniRef (CC BY 4.0), the PDB (CC0), AlphaFold DB (CC BY 4.0), PAZy, ESTHER and MGnify,\neach of which carries its own terms.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/PANTS\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/PANTS</a></td>\n</tr>\n</table>", "# Decoding the Molecular Machinery Behind Plastic Degradation\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-11-15 on marcdeller.com (categories: blog, microplastics).*\n\n* [Structural insight into molecular mechanism of poly(ethylene terephthalate) degradation](https://pmc.ncbi.nlm.nih.gov/articles/PMC5785972/)\n\n* **Catalytic Triad Architecture:** The [Ser160-His237-Asp206 catalytic triad](https://pmc.ncbi.nlm.nih.gov/articles/PMC5785972/) forms the enzyme\u2019s chemical machinery, with serine acting as the nucleophile that attacks PET\u2019s ester bonds.\n* **Two-Step Degradation Process:** PET hydrolysis proceeds through [nick generation followed by terminal digestion](https://pmc.ncbi.nlm.nih.gov/articles/PMC5785972/), systematically converting insoluble polymer into soluble MHET, TPA, and ethylene glycol monomers.\n* **Specialized Substrate Binding:** The [L-shaped binding cleft accommodates four MHET moieties](https://pmc.ncbi.nlm.nih.gov/articles/PMC5785972/) across subsites I and II, with \u03c0-\u03c0 stacking interactions and hydrophobic contacts positioning the substrate for catalysis.\n* **Structural Basis for Superior Activity:** An additional disulfide bond and extended \u03b28-\u03b16 loop distinguish *Is*PETase from other cutinases, enhancing thermal stability and creating an optimal continuous substrate binding groove.\n* **Engineering Potential:** Structure-guided mutations such as [R280A increase PET degradation by 22-32%](https://pmc.ncbi.nlm.nih.gov/articles/PMC5785972/), demonstrating that rational protein design can enhance enzyme performance for industrial recycling applications.\n\n1. **Structural insight into molecular mechanism of poly(ethylene terephthalate) degradation:** [Joo et al., Nature Communications, 2018](https://pmc.ncbi.nlm.nih.gov/articles/PMC5785972/)\n2. **The reaction mechanism of the Ideonella sakaiensis PETase enzyme:** [Burgin, et al., commun. chem., 2024.](https://pubmed.ncbi.nlm.nih.gov/38538850/)\n3. **Structural insight into catalytic mechanism of PET hydrolase:** [Han et al., Nature Communications, 2017.](https://pmc.ncbi.nlm.nih.gov/articles/PMC5727383/)\n4. **Active Site Flexibility as a Hallmark for Efficient PET Degradation by I. sakaiensis PETase:** [Fecker et al., Biophys. J., 2018.](https://pubmed.ncbi.nlm.nih.gov/29590588/)\n\n**How a bacterial enzyme dismantles polyethylene terephthalate one bond at a time**\n\nThe global accumulation of polyethylene terephthalate (PET) waste has created an environmental crisis of unprecedented scale, with projections suggesting plastic waste could reach 33 billion tons by 2050. While PET bottles are engineered to last for centuries, researchers have discovered a remarkable biological solution: an enzyme from the bacterium *Ideonella sakaiensis* that can efficiently dismantle this persistent polymer at moderate temperatures. Understanding the precise molecular mechanism by which this enzyme operates represents a crucial step toward developing scalable biotechnological solutions for plastic waste management.\n\nAt the heart of PET degradation lies the *Is*PETase enzyme, which belongs to the \u03b1/\u03b2 hydrolase superfamily and employs a sophisticated catalytic apparatus to cleave the ester bonds that hold PET polymers together. The enzyme\u2019s active site contains a classical Ser-His-Asp catalytic triad consisting of Ser160, His237, and Asp206, arranged in a precise three-dimensional configuration that enables the nucleophilic attack on PET\u2019s carbonyl carbons. This triad follows the canonical serine hydrolase mechanism observed in related enzymes, where the serine residue serves as the nucleophile, the histidine acts as a general base and acid, and the aspartate stabilizes the histidine through hydrogen bonding.\u200b\n\nThe structural architecture of *Is*PETase reveals an L-shaped substrate binding cleft approximately 40 \u00c5 in length that accommodates up to four mono(2-hydroxyethyl) terephthalate (MHET) moieties. This binding site is divided into distinct regions: subsite I binds the first MHET unit through \u03c0-\u03c0 stacking interactions with aromatic residues Tyr87 and Trp185, positioning the scissile ester bond adjacent to the catalytic Ser160. The enzyme\u2019s subsite II extends across three sub-regions (IIa, IIb, IIc) and accommodates three additional MHET moieties through predominantly hydrophobic interactions, with key contributions from residues including Trp159, Ile232, Asn233, Ser238, Asn241, and Asn246.\u200b\n\n**PET Degradation Mechanism**\n\nThe degradation process unfolds through a precisely orchestrated two-step mechanism that systematically converts solid PET polymer into soluble monomers:\n\n**1. Nick Generation Step**\n\nThe enzyme binds to the PET surface using its flat hydrophobic interface, positioning the polymer chain within its substrate binding cleft. Four MHET moieties bind sequentially to the subsites: one MHET moiety occupies subsite I while three MHET moieties fill subsite II. The scissile ester bond becomes positioned between subsite I and II near the catalytic Ser160 residue. The catalytic machinery then executes nucleophilic attack by Ser160 on the ester carbonyl carbon, cleaving one ester bond and creating a \u201cnick\u201d in the continuous PET polymer chain. This initial scission generates two distinct PET fragments: a TPA-terminal chain released from subsite I and a hydroxyethyl (HE)-terminal chain released from subsite II. This nick formation represents the critical first step, converting the insoluble, highly crystalline PET into more accessible polymer fragments.\u200b\n\n**2. Terminal Digestion Step**", "# Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\nAuthors: Zhuo, J.; Li, Y.-L.; Qian, D.-Q.; Burns, D.M.; Mei, S.; Rafalski, M.; Xu, M.; Cao, G.; Pan, Y.; Jia, Z.; Jalluri, R.; Epling, L.B.; Fenalti, G.; Deller, M.C.; Procak, J.; Covington, M.; He, X.; Collins, R.; Stubbs, M.; Burke, K.; Oliver, J.; Margulis, A.; Boer, J.; Wynn, R.; Scherle, P.; Diamond, S.; Newton, R.; Zhang, Y.; Metcalf, B.; Yao, W.\nSource: Journal of Medicinal Chemistry (Drug Annotation), 2026\n\n*Figure 4.* Structural Analysis of JAK Inhibitors With JAK1 and JAK2. (A) Crystal structure of povorcitinib 22 (pose 1, orange sticks) bound to JAK2 (PDB 10PJ, green cartoon) superposed onto the crystal structure of povorcitinib 22 bound to JAK1 (PDB 10PI, blue sticks and cartoon, respectively) highlighting interactions with the hinge region. Hydrogen bonds are depicted as dotted lines for JAK2/povorcitinib 22 within the hinge region. (B) Crystal structure of baricitinib 2 (magenta sticks) bound to JAK2 (magenta cartoon) superposed onto the crystal structure of povorcitinib 22 (orange sticks) bound to JAK2 (green cartoon). Hydrogen bonds are shown as dotted lines for JAK2/povorcitinib 22 within the hinge region. (C) Structural comparison between the crystal structures of povorcitinib 22 bound to JAK2 (green cartoon) and JAK1 (blue cartoon) showing the different conformations adopted by the P-loop (indicated by an arrow). For clarity, povorcitinib 22 (blue sticks) is only depicted in the povorcitinib 22/JAK1 complex structure while it is omitted in the JAK2 complex structure. (D) Different view of panel 'C'. (E,F) Cartoon representation of crystal structures of povorcitinib 22 bound to JAK2 in the binding pose denoted as 'pose 2' (E, green) and bound to JAK1 (F, blue). The ligand and amino acid side chains are shown as sticks. Key hydrogen-bonding interactions are shown as dotted lines. (G) Surface representation of JAK2 bound to povorcitinib 22 showing ligand in both conformations (pose 1, orange and pose 2, green) in the binding pocket. Note that the P-loop amino acid N859 of JAK2 makes a hydrogen bond interaction with E1015 and partially occludes the ligand binding site. (H) JAK1 bound to povorcitinib 22.\n\n*Figure 5.* Effect of 18 and 22 (INCB054707) on Adjuvant-Induced Arthritis in Rat (rAIA) Model post 13 d of Immunization. Sum of clinical scores of hind ankles and paws, with 0 being no inflammation and 4 being severe per ankle/paw (max score 8) after oral treatment with (A) 18 twice daily and (B) 22 once daily. n = 6 for each treatment arm. (C) Plethysmometer measurements of paw volume following treatment with 22. (D) Histopathology of joints after treatment with 22 based on the following criteria: bone resorption, periosteal new bone formation, inflammation, cartilage damage, and pannus scored on a scale of 0 (normal) to 7 (very severe/near destruction) and (E) the sum for each treatment group after treatment with 22. Statistics performed using GraphPad Prism Software.\n\n### Table 1. Internal Lead Compounds With JAK1 Isoform Selectivity\n\n| Compound | 4 | 5 | 6 | 7 | 8 | 9 |\n|---|---|---|---|---|---|---|\n| JAK1 IC50 (nM) | 20 | 6.3 | 7.5 | 3.3 | 1.3 | 1.6 |\n| JAK2 IC50 (nM) | 170 | 135 | 178 | 48 | 16 | 31 |\n| JAK2/JAK1 | 9 | 21 | 24 | 15 | 12 | 19 |\n| INA6 (cell) IC50 (nM) | -- | 1600 | 2000 | 612 | 226 | 306 |\n| WB IL-6 IC50 (nM) | | 406 | 1051 | 736 | 321 | 331 |\n| WB TPO IC50 (nM) | | 2800 | 6350 | 1550 | 1600 | 4050 |\n| WB IL-6/TPO | | 7 | 6 | 2 | 5 | 12 |\n| Caco-2 10e-6 cm/s | 0.1 | 0.3 | 1.3 | 2.1 | 2.3 | 0.9 |\n| h-Clr (L/h/kg) (%F) | | >1.1 (7) | >1.1 (<10) | >1.1 (<10) | >1.1 (11) | 0.7 (42) |\n\n*Caco-2, cell line of human colorectal adenocarcinoma cells; h-Clr, hepatic clearance; IC50, half-maximal inhibitory concentration; IL, interleukin; JAK, Janus kinase; TPO, thrombopoietin; and WB, whole blood.*\n\n### Table 2. In Vivo PK Comparison of Compound 9 to 3 (itacitinib)\n\n| Species | Parameters | 9 | 3 |\n|---|---|---|---|\n| Rat IV | CL (%HBF) | 29 | 100 |\n| | Vdss (L/kg) | 0.87 | 2.9 |\n| | t\u00bd (h) | 5.3 | 1.1 |\n| Rat PO | AUC (nM*h) | 3793 | 532 |\n| | %F | 31 | 19 |\n| Cyno IV | CL (%HBF) | 25 | 26 |\n| | Vdss (L/kg) | 0.57 | 0.76 |\n| | t\u00bd (h) | 1.9 (3) | 1.2 |\n| Cyno PO | AUC (nM*h) | 1549 | 2610 |\n| | %F | 9 | 19 |\n| Dog IV | CL (%HBF) | 49 | 37 |\n| | Vdss (L/kg) | 0.54 | 0.82 |\n| | t\u00bd (h) | 2.8 | 1.2 |\n| Dog PO | AUC (nM*h) | 610 | 5630 |\n| | %F | 6 | 43 |\n\n*Phosphate salt of 9 was used in dog study; TFA salt was used in cyno and rat studies; AUC, area under the curve; CL, clearance; cyno, cynomolgus; HBF, hepatic blood flow; IV, intravenous; PK, pharmacokinetic; PO, oral; t\u00bd, elimination half-life; Vdss, volume of distribution at steady state.*\n\n### Table 3. Incyte's Historical Success of Employing Intramolecular H-Bonding to Improve Permeability and PK Profile\n\n| Species | Parameters | INCB13740 (10) | INCB13739 (11) | INCB34049 (12) | INCB28060 (13) |\n|---|---|---|---|---|---|\n| In vitro | Caco-2 (x10\u207b\u2076 cm/s) | 15 | 38 | 4.7 | 22 |\n| Rat PO | Cmax (nM) | 552 | 6,460 | 281 | 8,400 |\n| | AUC (nM*h) | 1,321 | 11,200 | 428 | 35,200 |\n\n*AUC, area under the curve; Cmax, maximum observed concentration; PK, pharmacokinetics; PO, oral.*\n\n### Table 4. SAR of Intramolecular Hydrogen Bonded Amides to Improve the PK Profile", "# Compositional Protein Design: Building Proteins from 3D Ellipsoid Blueprints\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: blog).*\n\n* ProtComposer is a new AI model that generates protein structures based on a user-defined layout of 3D ellipsoids.\n* Each ellipsoid acts as a modular blueprint, allowing a designer to control the location, shape, and secondary structure of different parts of the protein.\n* This compositional approach gives protein engineers unprecedented control, enabling the rational assembly of novel proteins from defined building blocks.\n\n1. Compositional Protein Structure Generation with 3D Ellipsoids: <https://arxiv.org/abs/2503.05025>\n2. <https://openreview.net/pdf/3cd76a5f12a43a2296132178c1a97bad7e141588.pdf>\n\nImagine designing a complex machine not by sculpting a single block of material, but by arranging pre-defined components in space and letting an expert system fill in the connections. This modular, compositional approach is now a reality for protein engineering, thanks to innovative AI models that allow us to think like architects when designing new proteins.\n\nA groundbreaking model called **ProtComposer** introduces a new way to control protein generation using a simple, intuitive layout: a set of 3D ellipsoids. Each ellipsoid acts as a rough \u201csketch\u201d or blueprint for a piece of the final protein. A designer can place these ellipsoids in 3D space and assign them properties, such as their desired size, orientation, and even the type of secondary structure they should contain (e.g., \u201cthis ellipsoid should become an alpha-helix\u201d). The AI then uses this layout as a guide to generate a complete, high-quality protein backbone that conforms to the specified arrangement.[[15]](applewebdata://8E6058B5-53A8-4D8D-9BA0-6D5E119AFE88#fn15)[[16]](applewebdata://8E6058B5-53A8-4D8D-9BA0-6D5E119AFE88#fn16)\n\nThis approach unlocks powerful new capabilities. A user can hand-craft a novel arrangement of ellipsoids to design a protein with a completely new architecture. They can extract the ellipsoid layout from an existing protein and then \u201cre-wire\u201d it to change its connectivity, or even edit the properties of a single component. By conditioning the generative process on these modular, human-interpretable blueprints, ProtComposer gives designers an unprecedented level of control over the final product. It\u2019s a leap toward a future where we can rationally assemble proteins from functional building blocks, opening up new frontiers in synthetic biology and materials science.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\nThe chemokine sub-families are of particular importance due to their roles in inflammation (Baggiolini and Dahinden, 1994) and chemotaxis (Miller and Krangel, 1992). Recently it has been shown that entry of human immunodeficiency virus-1 (HIV-1) into cells depends on the chemokine receptors CCR5 and CXCR4 to which RANTES and SDF-1\u03b1 bind, respectively, so antagonizing infectivity. Members of this family adopt a \u03b2-sandwich motif with a C-terminal \u03b1-helical segment. This fold is exemplified by IL-8, which forms a symmetrical dimer consisting of a six-stranded \u03b2-sheet with two \u03b1-helices straddling the sheet. This produces a structure reminiscent of the peptide binding groove observed in the major histocompatibility complex (MHC) class I molecules. The chemokine family can be further sub-grouped according to the amino acid sequence around a conserved set of cysteine residues; C, CC, CXC or CXXXC (see Table 1-2). Members of the CXC chemokine family analysed structurally include IL-8 (Baldwin et al., 1991; Clore and Gronenborn, 1991), NAP-2 (Malkowski et al., 1995), SDF-1\u03b1 (Dealwis et al., 1998). Members of the CC chemokine family which have been analysed structurally include MCP-1 (Lubkowski et al., 1997), MIP1-\u03b1 and RANTES (Skelton et al., 1995; Wilken et al., 1999).\n\n##### 1.4.1.3.3 Disulphide rich\n\nMembers of this sub-family include insulin, IGF-I, IGF-II, relaxin and bombyxin (see Table 1-2). The polypeptide hormone insulin can be included in this sub-family along with the structurally related IGF-1 (see Figure 1-1) and IGF-2. Insulin consists of a \u03b1/\u03b2 structure derived through folding and cleavage of proinsulin.\n\n#### 1.4.1.4 Group 4 Cytokines (mosaic)\n\nMembers of this family combine several structural features and as such may also belong to more than one structural group. Members include EGF, HGF and IL-12. This is a very diverse group of cytokines and is often further sub-divided according to the domain types involved (i.e. EGF-like, kringle and \u03b1-helical) (see Table 1-2). For example HGF is a heterodimer consisting of an \u03b1- and \u03b2-domain. The \u03b1-subunit contains four Kringle domains and the \u03b2-subunit contains a serine protease domain. Kringle domains consist of a five-stranded antiparallel beta sheet structure with a conserved pattern of three internal disulphide bridges, as observed in the NK1 fragment of HGF (Chirgadze et al., 1999; Ultsch et al., 1998). IL-12 is also a heterodimer in which the \u03b1-subunit contains features of 4-\u03b1-helical bundle cytokines and the \u03b2-subunit contains features of a haemopoietin receptor domain.\n\n### 1.4.2 Structural Classification of Cytokine Receptors\n\nCrystallographic and NMR studies of cytokine receptors are severely hampered by their physical and chemical properties. Firstly, cytokine receptors are relatively large; often being composed of multiple immunoglobulin related domains, so rendering them problematic for expression in simple prokaryotic systems. The large size also limits the use of NMR techniques. The multiple extended domain structure can also be problematic for crystallographic studies as the innate flexibility in the domain linker regions can adversely effect crystallisation. Secondly, cytokine receptors generally have transmembrane regions, thus rendering them insoluble in aqueous solutions and liable to aggregation. Thirdly, cytokine receptors are often heavily and heterogeneously glycosylated, which can again be problematic for crystal packing and X-ray structure determination. Methods used to overcome these problems include the use of eukaryotic expression systems that allow expression of larger and more homogeneous proteins and the formation of ligand receptor complexes which imposes greater rigidity to the receptor domains.\n\nThere is a close correlation between the structural class of the cytokine and the receptor type to which it binds. For example, almost all of the 4-\u03b1-helical bundle cytokines bind to receptors that contain a 200 amino acid ligand-binding domain known as the haemopoietin domain (see Figure 1-3). Exceptions to this rule include the interferons and IL-10, which bind to interferon receptor domains (Ho et al., 1993). However, the haemopoietin domain is structurally related to the interferon receptor domain and these domains are often referred to as haemopoietin class II. These domains are more distantly related to the fibronectin type III and immunoglobulin domains. Seven distinct cytokine receptor superfamilies can be defined on the basis of biochemical function, structural homology and sequence similarity. These are the haemopoietin receptor superfamily, the interferon/IL-10 receptor superfamily, the TNF/NGF receptor superfamily, the TGF\u03b2/activine receptor superfamily, the immunoglobulin receptor superfamily, the tyrosine kinase receptor superfamily and the guanine nucleotide-binding protein-coupled receptor superfamily (G-protein) (see Figure 1-3).\n\n*Figure 1-3.* Gallery of cytokine receptor folds arranged according to the cytokine ligands that they bind. All structures solved by X-ray crystallography (unless otherwise stated). Figures produced using BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994). The polypeptides are coloured from blue at the N-terminus through to red at the C-terminus. Red lettering denotes the main superfamilies of receptors.\n\n#### 1.4.2.1 Haemopoietin Receptor Superfamily", "# Vibe Coding MicroPlastic Blaster\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nOver the next day, the game evolved through 62 distinct versions, growing from a bare-bones prototype into a richly featured experience with co-founder power-ups, VC Coin mechanics, leaderboard persistence via localStorage, and a heads-up display that tracks kills by polymer type. The result is a tool that is simultaneously memorable, shareable, educational, and brand-aligned, precisely the qualities that make it effective as a fundraising and outreach asset.\n\n**Design Evolution From Prototype to Version 115:** One thing I found useful is that defining a fairly strict architecture from the start enabled easier, rapid LLM-driven iteration. The codebase is organized into clearly separated concerns so new enemies, facts, and power-ups can be easily added as the game evolves. This modularity meant that each vibe coding prompt could target a specific section without destabilizing the rest. The game grew from roughly 200 lines in version 1 (31 KB) to over 2,000 lines in version 115 (66 KB), yet the single-file discipline was maintained throughout, and I was pleasantly surprised by the lack of file-size bloat.\n\n```\n\n\t\t\t\t\t\u251c\u2500\u2500 CSS Styles (responsive, dark sci-fi biotech theme)\n\u251c\u2500\u2500 HTML UI (start screen, HUD, game-over screen)\n\u2514\u2500\u2500 JavaScript\n    \u251c\u2500\u2500 Configuration (MP_TYPES, FACTS, POWERUP_TYPES)\n    \u251c\u2500\u2500 Drawing Functions (drawBackground, drawCell, drawMicroplastic, etc.)\n    \u251c\u2500\u2500 Game Logic (spawn, fire, collision detection, power-ups)\n    \u251c\u2500\u2500 Game Loop (update \u2192 render \u2192 requestAnimationFrame)\n    \u2514\u2500\u2500 Event Handlers (click, touch for mobile)\n\n```\n\n**Microplastic Type Definitions:** Each microplastic type maps directly to a real polymer class, with scientifically motivated attributes controlling gameplay behaviour.PET particles are the slowest but toughest, reflecting the real-world challenge of degrading crystalline PET. Each polymer is rendered with a unique shape (jagged, foam, circle, block) so players learn to visually identify different plastic types:\n\n```\n\n\t\t\t\t\tconst MPTYPES = [\n  { name: \"Polypropylene\",  short: \"PP\",  color: \"#ff4757\",\n    size: 20, speed: 0.8,  hp: 1, points: 15, shape: \"jagged\" },\n  { name: \"Polyethylene\",   short: \"PE\",  color: \"#ffa502\",\n    size: 22, speed: 0.7,  hp: 2, points: 20, shape: \"foam\" },\n  { name: \"Polystyrene\",    short: \"PS\",  color: \"#7bed9f\",\n    size: 18, speed: 0.95, hp: 2, points: 25, shape: \"circle\" },\n  { name: \"Polyethylene terephthalate\", short: \"PET\", color: \"#a29bfe\",\n    size: 24, speed: 0.6,  hp: 3, points: 30, shape: \"block\" }\n];\n\n```\n\n**Power-Up System:** Power-ups map directly to Elora\u2019s therapeutic pipeline and fundraising narrative. FAST-PETase activates triple-shot enzyme streams, mirroring the engineered variant\u2019s enhanced catalytic efficiency over wild-type PETase. The VC Coins power-up introduces a fundraising mechanic: collecting coins slows disease progression (microplastic speed), visually reinforcing the message that investment directly translates to therapeutic impact:\n\n```\n\n\t\t\t\t\tconst POWERUPTYPES = [\n  { name: \"FAST-PETase\",    color: \"#00e5ff\", emoji: \"\u26a1\",\n    effect: \"rapid\",   duration: 6000 },\n  { name: \"Enzyme Boost\",   color: \"#ffd32a\", emoji: \"\ud83d\udcaa\",\n    effect: \"power\",   duration: 6000 },\n  { name: \"Cell Repair\",    color: \"#ff4757\", emoji: \"\u2764\ufe0f\",\n    effect: \"heal\",    duration: 0 },\n  { name: \"Chain Reaction\",  color: \"#a29bfe\", emoji: \"\ud83d\udca5\",\n    effect: \"blast\",   duration: 4000 },\n  { name: \"VC Coins\",       color: \"#facc15\", emoji: \"\ud83d\udcb0\",\n    effect: \"vc-slow\", duration: 5000 }\n];\n\n```\n\n**Enzyme Firing Mechanics:** The fireEnzyme function calculates projectile trajectories from the cell toward the player\u2019s click/tap target. Co-founder power-ups produce dramatically different firing patterns:\n\n```\n\n\t\t\t\t\tfunction fireEnzyme(targetX, targetY) {\n  const startX = GAMEWIDTH / 2;\n  const startY = GAMEHEIGHT - 80;\n  const angle = Math.atan2(targetY - startY, targetX - startX);\n  const baseSpeed = 9;\n  const speed = vcSpeedActive ? baseSpeed * 0.5 : baseSpeed;\n\n  if (enzymeType === \"marc\") {\n    playSound(\"shoot-marc\");\n    const streams = 7;\n    const spread = 0.4;\n    for (let i = 0; i &lt; streams; i++) {\n      const t = i / (streams - 1);\n      const a = angle - spread / 2 + spread * t;\n      enzymes.push({\n        x: startX, y: startY,\n        vx: Math.cos(a) * speed * 1.3,\n        vy: Math.sin(a) * speed * 1.3,\n        size: 7, color: \"#22d3ee\",\n        damage: enzymePower * 10,\n        style: \"marc\"\n      });\n    }\n    return;\n  }\n  // Standard enzyme fire...\n  enzymes.push({\n    x: startX, y: startY,\n    vx: Math.cos(angle) * speed,\n    vy: Math.sin(angle) * speed,\n    size: 6, color: \"#00e5ff\",\n    damage: enzymePower * 0.5\n  });\n}\n\n```\n\n**Collision Detection:** Collision is computed using Euclidean distance between enzyme projectiles and microplastic particles. Destroyed particles trigger cascading visual effects and score updates:\n\n```\n\n\t\t\t\t\tfunction checkCollisions() {\n  const cellX = GAMEWIDTH / 2;\n  const cellY = GAMEHEIGHT - 80;\n\n  for (let i = enzymes.length - 1; i &gt;= 0; i--) {\n    for (let j = microplastics.length - 1; j &gt;= 0; j--) {\n      const e = enzymes[i];\n      const m = microplastics[j];\n      if (!e || !m) continue;\n\n      if (Math.hypot(e.x - m.x, e.y - m.y) &lt; m.size + e.size) {\n        m.hp -= e.damage;\n        enzymes.splice(i, 1);\n\n        if (m.hp &lt;= 0) {\n          playSound(\"destroy\");\n          score += m.points * level;\n          killCounts[m.short]++;\n          effects.push({\n            x: m.x, y: m.y,\n            type: \"explosion\", color: m.color, alpha: 1, radius: 5\n          });\n          microplastics.splice(j, 1);\n        }\n        break;\n      }\n    }\n  }\n}\n\n```\n\n**Sound Synthesis Without External Files:** All sound effects are generated procedurally using the Web Audio API, keeping the game truly zero-dependency, although in later version I did add MP3 files for extra oomph with powerup sounds:\n\n```", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\nChorismate mutase/prephenate dehydrogenase from Haemophilus influenzae Rd KW20 is a bifunctional enzyme that catalyzes the rearrangement of chorismate to prephenate and the NAD(P)(+)-dependent oxidative decarboxylation of prephenate to 4-hydroxyphenylpyruvate in tyrosine biosynthesis. The crystal structure of the prephenate dehydrogenase component (HinfPDH) of the TyrA protein from H. influenzae Rd KW20 in complex with the inhibitor tyrosine and cofactor NAD(+) has been determined to 2.0\u2005\u00c5 resolution. HinfPDH is a dimeric enzyme, with each monomer consisting of an N-terminal \u03b1/\u03b2 dinucleotide-binding domain and a C-terminal \u03b1-helical dimerization domain. The structure reveals key active-site residues at the domain interface, including His200, Arg297 and Ser179 that are involved in catalysis and/or ligand binding and are highly conserved in TyrA proteins from all three kingdoms of life. Tyrosine is bound directly at the catalytic site, suggesting that it is a competitive inhibitor of HinfPDH. Comparisons with its structural homologues reveal important differences around the active site, including the absence of an \u03b1-\u03b2 motif in HinfPDH that is present in other TyrA proteins, such as Synechocystis sp. arogenate dehydrogenase. Residues from this motif are involved in discrimination between NADP(+) and NAD(+). The loop between \u03b25 and \u03b26 in the N-terminal domain is much shorter in HinfPDH and an extra helix is present at the C-terminus. Furthermore, HinfPDH adopts a more closed conformation compared with TyrA proteins that do not have tyrosine bound. This conformational change brings the substrate, cofactor and active-site residues into close proximity for catalysis. An ionic network consisting of Arg297 (a key residue for tyrosine binding), a water molecule, Asp206 (from the loop between \u03b25 and \u03b26) and Arg365' (from the additional C-terminal helix of the adjacent monomer) is observed that might be involved in gating the active site.\n\n## The structure of Jann_2411 (DUF1470) from Jannaschia sp. at 1.45\u2005\u00c5 resolution reveals a new fold (the ABATE domain) and suggests its possible role as a transcription regulator (2010)\n\n**Authors:** Bakolitsa C, Bateman A, Jin KK, McMullan D, Krishna SS, Miller MD, Abdubek P, Acosta C, Astakhova T, Axelrod HL, Burra P, Carlton D, Chiu HJ, Clayton T, Das D, Deller MC, Duan L, Elias Y, Feuerhelm J, Grant JC, Grzechnik A, Grzechnik SK, Han GW, Jaroszewski L, Klock HE, Knuth MW, Kozbial P, Kumar A, Marciano D, Morse AT, Murphy KD, Nigoghossian E, Okach L, Oommachen S, Paulsen J, Reyes R, Rife CL, Sefcovic N, Tien H, Trame CB, Trout CV, van den Bedem H, Weekes D, White A, Xu Q, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley S, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309109025196. PMID: 20944211.\n\nThe crystal structure of Jann_2411 from Jannaschia sp. strain CCS1, a member of\u00a0the Pfam PF07336 family classified as a domain of unknown function (DUF1470), was solved to a resolution of 1.45\u2005\u00c5 by multiple-wavelength anomalous dispersion (MAD). This protein is the first structural representative of the DUF1470 Pfam family. Structural analysis revealed a two-domain organization, with the N-terminal domain presenting a new fold called the ABATE domain that may bind an as yet unknown ligand. The C-terminal domain forms a treble-clef zinc finger that is likely to be involved in DNA binding. Analysis of the Jann_2411 protein and the broader ABATE-domain family suggests a role as stress-induced transcriptional regulators.\n\n## The structure of KPN03535 (gi|152972051), a novel putative lipoprotein from Klebsiella pneumoniae, reveals an OB-fold (2010)\n\n**Authors:** Das D, Kozbial P, Han GW, Carlton D, Jaroszewski L, Abdubek P, Astakhova T, Axelrod HL, Bakolitsa C, Chen C, Chiu HJ, Chiu M, Clayton T, Deller MC, Duan L, Ellrott K, Elsliger MA, Ernst D, Farr CL, Feuerhelm J, Grzechnik A, Grant JC, Jin KK, Johnson HA, Klock HE, Knuth MW, Krishna SS, Kumar A, Marciano D, McMullan D, Miller MD, Morse AT, Nigoghossian E, Nopakun A, Okach L, Oommachen S, Paulsen J, Puckett C, Reyes R, Rife CL, Sefcovic N, Tien HJ, Trame CB, van den Bedem H, Weekes D, Wooten T, Xu Q, Hodgson KO, Wooley J, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309109018168. PMID: 20944219.\n\nKPN03535 (gi|152972051) is a putative lipoprotein of unknown function that is secreted by Klebsiella pneumoniae MGH 78578. The crystal structure reveals that despite a lack of any detectable sequence similarity to known structures, it is a novel variant of the OB-fold and structurally similar to the bacterial Cpx-pathway protein NlpE, single-stranded DNA-binding (SSB) proteins and toxins. K. pneumoniae MGH 78578 forms part of the normal human skin, mouth and gut flora and is an opportunistic pathogen that is linked to about 8% of all hospital-acquired infections in the USA. This structure provides the foundation for further investigations into this divergent member of the OB-fold family.\n\n## The structure of SSO2064, the first representative of Pfam family PF01796, reveals a novel two-domain zinc-ribbon OB-fold architecture with a potential acyl-CoA-binding role (2010)\n\n**Authors:** Krishna SS, Aravind L, Bakolitsa C, Caruthers J, Carlton D, Miller MD, Abdubek P, Astakhova T, Axelrod HL, Chiu HJ, Clayton T, Deller MC, Duan L, Feuerhelm J, Grant JC, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kumar A, Marciano D, McMullan D, Morse AT, Nigoghossian E, Okach L, Reyes R, Rife CL, van den Bedem H, Weekes D, Xu Q, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309110002514. PMID: 20944206.", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n> **BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, binding analysis, and run control, orchestrated end to end from a single spec file.**\n\n[](https://boltzmaker.mdeller.com)                 \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://boltzmaker.mdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">boltzmaker.mdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/BoltzMaker\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/BoltzMaker</a></td>\n</tr>\n</table>\n\n---\n\nBoltzMaker is for campaign-scale Boltz-2 work: a matrix of receptors x ligands x conditions,\ndescribed once and run as one unit, rather than a prediction at a time. One annotated\n`boltz_input.md` names the proteins, their co-folded partners, the ligands, the pocket /\ncovalent / distance constraints and any ligand-free controls; BoltzMaker crosses them into a\ntarget per combination and carries all of them through generation, validation, prediction and\nreporting together. The example campaigns below run up to 20 targets from one file, and a real\nGPCR panel here ran 26. It is built for structural biologists and drug-discovery scientists\nrunning structure/affinity panels -- covalent-linkage studies, multi-chain SAR and selectivity\nmatrices, pocket-condition sweeps.\n\n**The way in is [boltzmaker.mdeller.com](https://boltzmaker.mdeller.com), and it installs\nnothing.** Every stage but one runs there: describing the campaign (the `new` wizard's questions\nas a form), `generate`, `preflight`, and the whole of `analyze` -- PLIP interaction detection and\ncompare-sse included. Only the GPU `run` step needs your own hardware, so the site hands you one\nself-extracting bundle carrying the spec, BoltzMaker and a pinned environment. Run it where the\nGPU is, bring back the single `.bmz` it writes, and the site opens the campaign as an interactive\nreport. Full detail under **Web deployment** below.\n\nWhy it matters: hand-running a campaign that size means writing dozens of near-identical YAML\nfiles, remembering the right CLI flags for your hardware, and digging through prediction JSONs\nafterwards. BoltzMaker does all three, enriches the result with real protein-ligand interaction\nanalysis via [cif2plip](https://github.com/bellcheddar/cif2plip), and runs ligand-chemistry\nchecks -- undefined stereocentres, ambiguous protonation states, stray salts -- so bad chemistry\nis caught before hours of compute rather than silently mispredicted.\n\nIt installs locally as a single script too, consolidating five earlier single-purpose tools into\none spec-driven pipeline sharing a single campaign format:\n[generate_yaml](https://github.com/bellcheddar/generate_yaml) (input YAMLs),\n[simple-zsh-script-to-run-boltz2](https://github.com/bellcheddar/simple-zsh-script-to-run-boltz2)\n(driving `boltz predict`),\n[analyze-boltz2-results](https://github.com/bellcheddar/analyze-boltz2-results) (post-run\nanalysis), [cif2plip](https://github.com/bellcheddar/cif2plip) (interaction profiling) and\n[smiles2grid](https://github.com/bellcheddar/smiles2grid) (a boxed grid of 2D ligand\nstructures). And yes, the name is a nod to\n[Boltmaker](https://www.timothytaylor.co.uk/beer/boltmaker), Timothy Taylor's Champion Beer of\nBritain and one of the author's favourites.\n\nSee [CHANGELOG.md](CHANGELOG.md) for what's changed recently, and\n[PROJECT_PLAN.md](PROJECT_PLAN.md) for the roadmap behind the To-do list below.\n\n## \ud83d\udce5 Installation\n\nThree ways to run BoltzMaker. **Path 1 installs nothing at all** and is where to start; Paths 2\nand 3 install it properly on your own machine, and differ only in whether that machine has\ninternet access.\n\n### Path 1: Web app (no install)\n\n[boltzmaker.mdeller.com](https://boltzmaker.mdeller.com) opens on two ways to work: **Fully\nAutomated Mode**, the two-step describe-download-run-upload flow above, and **Stepwise Mode**,\nwhich exposes the same non-GPU stages as four independent tools (**Wizard**, **Generate**,\n**Preflight**, **Analyze**) for when you already have your own setup and want one piece of it.\nBoth are documented under **Web deployment** below.\n\n### Path 2: This computer has internet access\n\n**1.** Open the **Terminal** app (Applications -> Utilities -> Terminal, or search for\n\"Terminal\" with Spotlight), then make a folder for BoltzMaker and go into it:\n\n```sh\nmkdir -p ~/Desktop/BoltzMaker\ncd ~/Desktop/BoltzMaker\n```\n\n**2.** Download BoltzMaker into that folder:\n\n```sh\ngit clone https://github.com/bellcheddar/BoltzMaker.git .\n```\n\n(the trailing `.` puts the files directly into the folder you just made rather than inside\nanother new folder). If this is the first time you've used `git`, macOS may pop up a dialog\nasking to install \"Command Line Tools\" -- click Install, wait for it to finish, then run the\ncommand again.\n\n**3.** Run the installer:\n\n```sh\n./install.sh\n```\n\nIt downloads a few GB of software the first time, so it can take a while.\n\n**4.** Check it worked:\n\n```sh\npixi run preflight examples/t4_lysozyme/boltz_input.md\n```\n\nA table of green PASS results means you're ready -- see **Commands** below.\n\n### Path 3: This computer has no internet access\n\nFor a lab machine, server, or anything offline or behind a firewall. You'll need a second\ncomputer that *does* have internet access to prepare a single installer file first.\n\n**1.** On the computer **with** internet access, follow **Path 2** above, then build the\noffline installer file:\n\n```sh\npixi global install pixi-pack\npixi-pack --platform osx-arm64 --ignore-pypi-non-wheel --create-executable -o boltzmaker-installer.sh\n```\n\n(use `--platform linux-64` instead of `osx-arm64` if the offline computer runs Linux)\n\n**2.** Copy `boltzmaker-installer.sh`, together with the whole `BoltzMaker` folder, onto the\noffline computer (USB drive, internal network transfer, etc.).", "# 2013 \u2013 Models of protein\u2013ligand crystal structures: trust, but verify\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: papers).*\n\nClick to enlarge\n\n* X-ray crystallography provides precise, high-resolution 3D protein-ligand structures critical for accurate drug design and computational biology.\n* It requires rigorous validation using electron density fit, stereochemistry, and model geometry to ensure the reliability of protein-ligand models.\n* Strict validation prevents overinterpretation and supports high-quality scientific research by confirming structural accuracy and binding plausibility.\n\n1. <https://link.springer.com/article/10.1007/s10822-015-9833-8>\n\nMetrics underscoring why X-ray crystallography is the gold standard for creating precise protein-ligand models\u2014essential for advancing drug design and computational biology\u2014while advocating rigorous validation to combat overinterpretation and ensure reliability in scientific research.", "# GitHub README: PANTS\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/PANTS, 2026\n\n```\n  SOURCE ENVIRONMENTS                      SEED SET\n  \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510                 \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n  \u2502 compost      1.0M    \u2502                 \u2502 6 wild types  (UniProt + UniParc)  \u2502\n  \u2502 marine       0.7M    \u2502                 \u2502 9 variants    (parent + mutations) \u2502\n  \u2502 landfill     0.4M    \u2502                 \u2502 2 variants    (PDB constructs)     \u2502\n  \u2502 wastewater   0.03M   \u2502                 \u2502 5 HGMPs       (SciDB deposit)      \u2502\n  \u2502 human gut   12.1M    \u2502                 \u2502 449 EC 3.1.1.101 (annotation)      \u2502\n  \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518                 \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n             \u2502  predicted proteins                           \u2502\n             \u2502                                               \u25bc\n             \u2502                              \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n             \u2502                              \u2502 cluster at 30% identity      \u2502\n             \u2502                              \u2502 \u2192 1 profile HMM per cluster  \u2502\n             \u2502                              \u2502 \u2192 anchor each on UniProt's   \u2502\n             \u2502                              \u2502   own ACT_SITE annotation    \u2502\n             \u2502                              \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n             \u25bc                                              \u2502\n  \u2554\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2557     \u2502\n  \u2551  RECALL                                           \u2551 \u25c4\u2500\u2500\u2500\u2518\n  \u2551                                                   \u2551\n  \u2551   MMseqs2 prefilter        E \u2264 1e-5               \u2551   fast, exhaustive\n  \u2551        \u2502                                          \u2551\n  \u2551        \u25bc                                          \u2551\n  \u2551   hmmscan vs profile library                      \u2551   assigns each survivor\n  \u2551        \u2502                                          \u2551   to the family it\n  \u2551        \u25bc                                          \u2551   actually resembles\n  \u2551   triad filter    Ser\u00b7His\u00b7Asp connected in SPACE  \u2551   \u2190 geometry, not motif\n  \u2551        \u2502                                          \u2551\n  \u255a\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u256a\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u255d\n           \u2502  retains ~0.006%  \u00b7  keeps E-value + bitscore\n           \u25bc                     as the baseline to beat\n     \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n     \u2502  CANDIDATES   \u2502  content-addressed on the sequence,\n     \u2502               \u2502  so the same protein found twice is one row\n     \u2514\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2518\n         \u2502       \u2502\n         \u2502       \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n         \u25bc                                  \u25bc\n  \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510        \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n  \u2502 ESM-2 t12-35M       \u2502        \u2502 ESMFold  (\u2264 450 aa)            \u2502\n  \u2502 frozen, CPU         \u2502        \u2502   \u2502                            \u2502\n  \u2502 480-dim, mean-pool  \u2502        \u2502   \u25bc                            \u2502\n  \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518        \u2502 superpose onto IsPETase 6EQE   \u2502\n             \u2502                   \u2502   \u2502  \u2190 at WRITE time, so the   \u2502\n             \u25bc                   \u2502   \u25bc    browser does none       \u2502\n  \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510        \u2502 active-site geometry           \u2502\n  \u2502 activity head       \u2502        \u2502   triad distances, cleft width,\u2502\n  \u2502 PU-corrected,       \u2502        \u2502   aromatic clamp               \u2502\n  \u2502 calibrated          \u2502        \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n  \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518                         \u2502\n             \u2502        \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n             \u25bc        \u25bc\n        \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n        \u2502  SQLite  \u2192  Flask  \u2192  Catalogue \u00b7 Superpose \u2502\n        \u2502                       Candidate \u00b7 Methods   \u2502\n        \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n\n  Every stage writes a manifest: input hashes, tool versions, git commit, wall time.\n```\n\nTwo things in that diagram are load-bearing and easy to miss. The **triad filter tests\nwhether Ser, His and Asp are connected in space**, not whether a motif matches, so three\nresidues present in sequence but not in contact correctly fail. And **superposition\nhappens when a structure is written**, not in the browser, which is what makes an\nN-structure overlay cost N file loads instead of N alignments.\n\n## \ud83c\udf21\ufe0f The therapeutic gap, measured\n\nThe table above is the premise. These are the numbers. Every enzyme named in the project\nbrief, plus every other PET hydrolase with a published optimum that could be found.\n\nRows are split by **how the number was obtained**, because that difference matters more\nthan the value: the first block is extracted programmatically from UniProt's curated\n`BIOPHYSICOCHEMICAL PROPERTIES` with `ECO:0000269` experimental evidence and PubMed IDs\nattached, and is what sits in the database. The second is from the literature and is\nrecorded here for context only.\n\n### \ud83d\udcd1 Measured, in the database, each with its citation\n\n| Enzyme | Topt | pH opt | Source |\n|---|---|---|---|\n| **HGMP01** (human gut) | **40 \u00b0C** | **~7.4, broad across pH 7.x** | PMID 39551294 |\n| IsPETase | 40 \u00b0C | 9.0 | PMID 26965627, 29603535 |\n| LCC | 50 \u00b0C | 8.5 | PMID 22194294, 24593046 |\n| *T. alba* est1 (`D4Q9N1`) | 50 \u00b0C | 6.0 | PMID 25910960 |\n| *T. alba* est2 (`F7IX06`) | 50 \u00b0C | 6.0 | PMID 20393707 |\n| TfCut1 (`G8GER6`) | 55 \u00b0C | 8.0 | PMID 23604968 |\n| TfCut2 | 55 \u00b0C | 8.0 | PMID 15638529, 20816933 |\n| *T. fusca* (`Q47RJ6`) | 60 \u00b0C | 8.0 | PMID 18658138, 20729325 |\n| *T. fusca* (`Q47RJ7`) | 60 \u00b0C | 8.0 | PMID 18658138 |\n\n### The engineered lineage, from the literature\n\nEvery variant named in the project brief. They are in the database at **weaker\nprovenance** than the block above: `extraction_confidence='review'` and `ECO:0000305`\n(inferred by curator) rather than `ECO:0000269` (experimental, from a publication), because\nthe values were collated from a secondary review rather than read from each primary paper.", "# GitHub README: GOBSMACKED\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/GOBSMACKED, 2026\n\n- [x] **Prepare, without references.** Input resolution across UniProt, RCSB, AlphaFold DB, ESM Atlas and uploads, with the priority reported rather than silently applied. Pfam routing, KLIFS and GPCRdb clients behind a 30-day cache, the Mol* pocket picker and the SVG sequence track\n- [x] **Map the KLIFS pocket onto an arbitrary sequence.** The 85-character pocket string carries no residue numbers and only its regions are contiguous, so it is placed region by region, longest run first, with a bounded difflib pass for regions the alignment shifted. 83 of 85 positions map for EGFR, putting the gatekeeper on Thr790, the hinge on Met793 and DFG on Asp855\n- [x] **The run bundle.** Five resumable stages in a self-contained pixi environment, with the GNN fallback, the platform choice and the solute-only trajectory\n- [x] **Analyze, with verification.** Pocket-C\u03b1 superposition, symmetry-corrected ligand RMSD, \u03c71 agreement, PLIP fingerprints compared in reference numbering, PandaMap, the scorecard and the dynamics panels\n- [x] **Both binding-mode classifiers, on day one.** Kinase and GPCR, each validated against structures whose labels are known: 1M17 Type I DFG-in, 1IEP Type II DFG-out, 2RH1 orthosteric inactive-like, 3SN6 active-like\n- [x] **Runs and ownership.** Public and private runs, an owner key stored only as a hash, unguessable job IDs, and private runs absent from listings rather than redacted in them\n- [x] **The About page.** A hand-drawn pipeline schematic, the grade thresholds, and a software table generated from the same file as THIRD_PARTY.md with every DOI checked against Crossref\n- [x] **Deploy to gobsmacked.mdeller.com.** Provisioned, certificated and serving over HTTP/2, with the nightly prune timer armed. The static location deliberately sets no `access_log`: nginx.conf gives the droplet the `vhost` format that appends the requested host, and naming `combined` in the location would silently zero this app's visit count on the mdeller.com launcher\n- [x] **Run the real loop once, end to end.** EGFR plus erlotinib, prepared on the live site, run on an M1 Max, uploaded and scored: see the worked example below. It found nine bugs the crystal-built fixtures could not, six of which failed silently\n- [x] **PLIP interactions drawn in Mol\\*, and the pocket as sticks.** Mol*'s viewer build exports no shape builder, so each interaction is loaded as a tiny structure of two-atom fragments joined by CONECT records: a run of them along PLIP's own endpoints reads as a dashed line, one file and one colour per interaction type. The lines are therefore the interactions the table lists rather than a second opinion computed by the viewer, which would quietly disagree with it\n- [x] **An apo reference in the overlay.** Optional on Panel 4, fetched and stripped of its ligands at analysis time, and drawn in purple as a fourth toggle: the shape the pocket has with nothing bound, which is the shape a predicted model tends to resemble\n- [x] **The trajectory played back beside the score.** Rendered on the droplet from the same trajectory the dynamics traces come from, so archives uploaded before it existed gain a clip on re-analysis: a C\u03b1 trace superposed on the protein, the pose in phosphor cyan, the residues lining its site in white, and forwards-then-backwards playback because a trajectory does not loop and a cut from the last frame to the first reads as a glitch rather than as data\n- [x] **Affinity, before and after MD.** Built as stage 5 of six, on by default. Boltz-2's affinity head scores the docked pose and frames sampled from the last fifth of the trajectory, following the Boltzina pattern of feeding an existing pose straight to the affinity module with the structure module bypassed. Reported as pIC50, the raw log10(IC50/\u00b5M) and the binder probability for both, with the change between them and the spread across the sampled frames, and deliberately outside the composite score: every graded metric has a crystal to be right or wrong about, and a predicted affinity has none. It lives in its own pixi environment with `no-default-feature = true`, because boltz pins a torch that the docking environment must not inherit, and it never fails the run: a pose it cannot score becomes a missing panel with the reason attached\n- [x] **Every pose in the overlay, with the numbers that separate them.** All ten drawn at once, and per pose the docking score, in-place RMSD to the top pose, centroid separation, best-fit shape RMSD and closest approach to the receptor. In-place against best-fit is the pair that matters: the same conformer in two sites is a search problem, two different conformers is not\n- [x] **Make flex and hybrid docking actually run.** Four bugs, found only by using them rather than by reading them. `flex` treats `-o` as a filename prefix and writes to `<prefix>_results`, so every flex run appeared to produce nothing; `hybrid` accepts neither `--seed` nor `-e` and exits on either; the search radius was half the box's *longest* side, giving a sphere that reached well outside the box it was supposed to describe; and the GNN checkpoint had moved to a `v4` name and release URL, so `hybrid` silently fell back to the empirical scorer on every run\n- [x] **Five starting structures, one campaign.** The experiment above: same ligand, pocket, box, seed and MD protocol, varying only the structure docking starts from, with a self-dock control that is correct by construction. It found that the top-ranked pose, not the sampling and not the receptor, is what limits the result, and that neither starting-model quality nor predicted affinity separates a right pose from a wrong one\n- [x] **Re-dock with alternative scoring functions, and test the affinity head as a re-ranker.** Twenty docking runs across the five receptors showed the apo sampling ceiling was a property of the search, not of the receptors: ESMFold's best available pose moves from 5.44 \u00c5 to 2.06 \u00c5, and pooling every pose gives all five receptors something at 3.04 \u00c5 or better. No protocol tested picks it: 4.15 \u00c5 mean top-1 against a 1.94 \u00c5 oracle. The affinity head returns 0.19 log units of pIC50 across poses spanning 1.51 to 8.77 \u00c5, ranks the 8.30 \u00c5 pose first, and correlates with RMSD at rho +0.26 (p 0.48), so it is not a rescoring function and is not used as one\n- [ ] **Close the 4.15 \u00c5 to 1.94 \u00c5 ranking gap.** Every receptor now has a near-native pose available and nothing ranks it first. This is the open problem, and it is upstream of anything the scorecard can fix: consensus scoring across functions, a rescoring model trained on decoys rather than on affinity, or short per-pose minimisation before ranking\n- [x] **Rank scoring functions on a fixed pose set.** The five runs left 50 poses whose distance to the crystal is already known, so the ranking question can be asked directly and cheaply: rescore the same poses with several independent scoring functions and ask which one puts a near-native pose first, per receptor type. Done, with seven functions over fifty poses: rescoring alone recovers the crystal pose on the control, and it separates two different failures, ranking for holo-like pockets and sampling for apo-like ones\n- [ ] **STEVEDORE: multi-ligand SAR series.** Score a congeneric series against one reference and correlate with ChEMBL affinity, which turns a single verification into a protocol assessment\n- [ ] **DOCKYARD: ingest poses from other engines.** Boltz-2, Vina and DiffDock all produce poses this scorecard could grade, and the comparison is more interesting than any single engine's self-report\n- [ ] **Cryptic pocket detection.** The pocket volume trace already shows a pocket opening and closing during MD; naming that as a finding rather than a plot is the next step\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/GOBSMACKED\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/GOBSMACKED</a></td>\n</tr>\n</table>\n\nReleased under the MIT licence. The app name is a joke; the numbers are not.", "# Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\nAuthors: Deller, M.C.; Hudson, K.R.; Ikemizu, S.; Bravo, J.; Jones, E.Y.; Heath, J.K.\nSource: Structure, 2000\n\n**Authors:** Marc C. Deller, Keith R. Hudson, Shinji Ikemizu, Jer\u00f3nimo Bravo, E. Yvonne Jones (corresponding author) and John K. Heath. Marc C. Deller and Keith R. Hudson contributed equally to this work. **Affiliations (condensed):** Cancer Research Campaign Receptor Structure Group, Division of Structural Biology, The Wellcome Trust Centre for Human Genetics, University of Oxford, Roosevelt Drive, Oxford OX3 7BN, UK; Cancer Research Campaign Growth Factors Group, School of Biochemistry, University of Birmingham, Edgbaston, Birmingham B15 2TT, UK. Present addresses: Yale University School of Medicine, Department of Pharmacology, New Haven, CT, USA (Hudson); Genesis Research and Development, Parnell, Auckland, New Zealand (Ikemizu); Laboratory of Molecular Biology, Medical Research Council Centre, Cambridge, UK (Bravo). **Published:** Structure, volume 8, pages 863\u2013874, 2000. Received 26 April 2000; revisions requested 7 June 2000; revisions received 22 June 2000; accepted 28 June 2000; published 20 July 2000.\n\n**Key words:** four-helix bundle cytokine, gp130 binding, oncostatin M, structure\u2013function.\n\n## Abstract\n\n**Background:** The cytokine oncostatin M (OSM) inhibits growth of certain tumour-derived cells, induces proliferation in other cell types (e.g. haemangioblasts) and is a mediator of inflammatory responses. Its mechanism of action is via specific binding to gp130 and either the leukaemia inhibitory factor receptor (LIFR) or oncostatin M receptor (OSMR) systems at the cell surface to form an active signalling complex.\n\n**Results:** We report here the crystal structure of human oncostatin M (hOSM) along with mutagenesis data which map the receptor-binding epitopes of the molecule. The structure was determined to a resolution of 2.2 \u00c5 and conforms to the haematopoietin cytokine up-up-down-down four-helix bundle topology. The site 2 epitope, responsible for gp130 binding, is centred around Gly120 which forms a 'dimple' on the surface of the molecule located on helices A and C. The site 3 motif, responsible for LIFR and OSMR binding, consists of a protruding Phe160/Lys163 pair located at the start of helix D.\n\n**Conclusions:** The data presented allow functional dissection of the receptor-binding interfaces to atomic resolution. Modelling suggests that the gp130 residue Phe169 packs into the site 2 dimple in an analogous fashion to structurally equivalent residues at the growth hormone\u2013growth hormone receptor interface, implying that certain key features may underlie recognition across the whole cytokine/receptor superfamily. Conversely, detailed comparison of the available structures suggests that variations on a common theme dictate the specificity of receptor\u2013ligand interactions within the gp130 family of cytokines.\n\n## Introduction\n\nHuman oncostatin M (hOSM) is a secreted 252 amino acid polypeptide cytokine that was originally isolated by virtue of its ability to suppress the proliferation of human melanoma cells in vitro [1]. Subsequent biological investigations have revealed that hOSM and mouse OSM (mOSM) exhibit a diversity of effects on different target cell types. These include the ability to inhibit the multiplication of a variety of different tumour-derived cell lines (reviewed in [2]) as well as inducing the proliferation of certain cells including Kaposi sarcoma-derived spindle cells [3] and haemangioblasts derived from the aorta\u2013gonad\u2013mesonephros region of the developing embryo [4]. hOSM, released from macrophages and neutrophils [5], has also been implicated as an important mediator of pro-inflammatory responses in joint destruction [6] and central nervous system inflammation and neurodegeneration [7,8] via, inter alia, the induction of acute-phase protein gene expression [9,10].\n\nhOSM is a member of the family of gp130 cytokines that are characterised by the use of the shared transmembrane transducing receptor gp130 in their mechanism of action [11]. hOSM is therefore functionally related to leukemia inhibitory factor (LIF), interleukin (IL)-6 and IL-11, cardiotrophin-1 (CT-1) and ciliary neurotrophic factor (CNTF). A key feature of the gp130 cytokines is that intracellular signalling results from the ligand-mediated formation of oligomeric receptor complexes which, depending upon the identity of the ligand, can differ in both composition and stoichiometry [12]. In all cases, however, receptor complexes include one or more molecules of the shared transducer gp130 [13]. gp130 cytokines can therefore exhibit both unique and shared biological activities in vivo and in vitro depending upon the composition of the receptor signalling complex that is formed [11]. hOSM signals via formation of an active receptor signalling complex that comprises gp130 and one of two partner receptors: either the LIF receptor (LIFR) [14] or the structurally related OSM-specific receptor, OSMR [15]. mOSM [16] differs from hOSM in that it only signals via association with gp130 and the murine OSMR [17,18]; this means that there is a significant divergence in action between mOSM and hOSM and that the LIFR recognition elements of OSM are not conserved. This finding also indicates that the unique biological activities of hOSM, such as cytostatic effects on tumour cells, are mediated via association with OSMR. Those activities of hOSM that are shared with other cytokines, such as LIF or CNTF, may reflect a common association with the LIFR.", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\n**Publication number:** WO 2023/107994 A1 (PCT). **International application number:** PCT/US2022/081076. **International filing date:** 07 December 2022. **International publication date:** 15 June 2023. **Applicant:** Incyte Corporation, 1801 Augustine Cut-Off, Wilmington, Delaware 19803, USA. **Inventors (all of 1801 Augustine Cut-Off, Wilmington, Delaware 19803, USA):** Patrick Mayes; Horacio G. Nastri; Rebecca A. Buonpane; Edimara S. Reis; Shaun M. Stewart; Jing Zhou; **Marc C. Deller**; Hsiang-Ting Lei; Leslie Brooke Epling McQuirter; Yonghong Zhao; Hamza Celik; Brittney Melissa Wass. **Agent:** Catherine Guzzo et al., Fish & Richardson P.C., Minneapolis, Minnesota. **International Patent Classification:** A61K 39/00; A61P 35/00; C07K 16/18; C07K 16/26; C07K 16/30. **Priority data:** US 63/287,394 filed 08 December 2021; US 63/288,479 filed 10 December 2021; US 63/421,052 filed 31 October 2022. **Published with international search report (Art. 21(3)) and with sequence listing part of description (Rule 5.2(a)).** The corresponding US publication is US 2023/0272055 A1.\n\n## Abstract\n\nAnti-mutant calreticulin (mutCALR) antibodies are disclosed. Also disclosed are related nucleic acids, vectors, cells, kits, and pharmaceutical compositions. Methods of treating or diagnosing myeloproliferative neoplasms with the anti-mutCALR antibodies are also disclosed.\n\n## Claim of Priority\n\nThis application claims the benefit of U.S. Provisional Patent Application No. 63/287,394, filed on December 8, 2021, U.S. Provisional Patent Application No. 63/288,479, filed on December 10, 2021, and U.S. Provisional Patent Application No. 63/421,052, filed on October 31, 2022, each of which is incorporated by reference herein in its entirety.\n\n## Sequence Listing\n\nThis application contains a Sequence Listing that has been submitted electronically as an XML file named 20443-0736WO1_SL.xml. The XML file, created on December 5, 2022, is 388,039 bytes in size. The material in the XML file is hereby incorporated by reference in its entirety.\n\n## Background\n\nCalreticulin (CALR) is a highly conserved chaperone protein that resides primarily in the endoplasmic reticulum and is involved in a variety of cellular processes including protein folding, calcium homeostasis, cell adhesion, and integrin signaling. CALR is also found in the nucleus, suggesting that it may have a role in transcription regulation. Mutations in the gene for CALR have been identified in patients with myeloproliferative neoplasms.\n\n## Summary\n\nThe present disclosure is based, at least in part, on the development of antibodies having high binding affinity and specificity to mutant CALR (\"mutCALR\").\n\nAccordingly, aspects of the present disclosure provide an antibody that binds to human mutant calreticulin (CALR), wherein the antibody comprises a heavy chain variable region (VH) comprising a VH CDR1, a VH CDR2, and a VH CDR3, wherein: the VH CDR1 comprises the amino acid sequence X1X2X3X4X5, wherein X1 is S, E, or D; wherein X2 is Y, L, or S; wherein X3 is A, S, or F; wherein X4 is I or M; and wherein X5 is S, Q, or H; the VH CDR2 comprises the amino acid sequence X6X7X8PX9X10X11X12X13X14YAX15X16X17X18G (SEQ ID NO:97), wherein X6 is L or G; wherein X7 is V, F, or I; wherein X8 is D or I; wherein X9 is E, D, or I; wherein X10 is D, G, F, A, S, or E; wherein X11 is G or A; wherein X12 is E or T; wherein X13 is T or A; wherein X14 is I, M, or N; wherein X15 is E or Q; wherein X16 is K or R; wherein X17 is F or L; and wherein X18 is R or Q; the VH CDR3 comprises the amino acid sequence X19X20X21X22X23X24X25X26X27X28X29X30X31X32X33X34X35X36X37X38 (SEQ ID NO:98), wherein X19 is P, E, or absent; wherein X20 is G, E, or absent; wherein X21 is G, W, S, or absent; wherein X22 is I, D, S, P, or absent; wherein X23 is S, L, I, T, G, or absent; wherein X24 is P, T, Q, I, D, R, or absent; wherein X25 is G, D, or absent; wherein X26 is E, Y, P, L, D, or S; wherein X27 is E, D, A, or G; wherein X28 is S, F, A, E, Y, or W; wherein X29 is Y or F; wherein X30 is G, D, or W; wherein X31 is P, Y, I, or H; wherein X32 is Y or absent; wherein X33 is Y or absent; wherein X34 is Y or absent; wherein X35 is G or absent; wherein X36 is M or absent; wherein X37 is D or absent; wherein X38 is V or absent; wherein the antibody comprises a light chain variable region (VL) comprising a VL CDR1, a VL CDR2, and a VL CDR3, wherein: the VL CDR1 comprises the amino acid sequence X39X40X41X42X43X44X45X46X47X48X49X50X51X52X53X54 (SEQ ID NO:99), wherein X39 is T, A, or absent; wherein X40 is G or absent; wherein X41 is Q, G, V, T, or S; wherein X42 is A, G, S, or N; wherein X43 is S, N, D, T, or Y; wherein X44 is Q, Y, N, D, S, or K; wherein X45 is D, I, F, V, S, or T; wherein X46 is N or absent; wherein X47 is I or absent; wherein X48 is I, G, or R; wherein X49 is S, G, A, D, I, R, or T; wherein X50 is Y or absent; wherein X51 is N, K, I, or E; wherein X52 is Y, S, N, D, H, F, R, or G; wherein X53 is L or V; and wherein X54 is N, H, S, D, or F; the VL CDR2 comprises the amino acid sequence X55X56X57X58X59X60X61, wherein X55 is T, D, E, Q, or R; wherein X56 is A, D, V, or N; wherein X57 is S, G, N, or R; wherein X58 is I, N, D, or K; wherein X59 is L, R, or W; wherein X60 is E or P; and wherein X61 is S, T, or L; the VL CDR3 comprises the amino acid sequence X62X63X64X65X66X67X68X69X70X71X72, wherein X62 is Q, S, C, or G; wherein X63 is Q, V, S, T, or A; wherein X64 is Q, L, W, or Y; wherein X65 is Q, N, D, I, T, A, or G; wherein X66 is S, P, G, N, or A; wherein X67 is N, Y, I, S, N, L, or D; wherein X68 is E, P, S, I, N, H, L, or T; wherein X69 is D, T, S, or absent; wherein X70 is P, H, L, R, F, A, Q, or absent; wherein X71 is W, L, V, Y, S, A, or E; and wherein X72 is T, V, or I.", "# Unlocking the Power of Protein Engineering: FDA-Approved Enzyme Therapies Transform Rare Disease Treatment\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-11-07 on marcdeller.com (categories: blog, microplastics).*\n\n* [The past, present, and future of enzyme-based therapies](https://pmc.ncbi.nlm.nih.gov/articles/PMC8714691/)\n\n* **Glycoengineering Revolutionizes Cellular Targeting**: [Avalglucosidase alfa\u2019s enhanced bis-M6P content](https://linkinghub.elsevier.com/retrieve/pii/S0168952523002208) improves skeletal muscle glycogen clearance through optimized mannose-6-phosphate receptor engagement, demonstrating how rational carbohydrate design overcomes tissue-specific delivery challenges in Pompe disease treatment.\n* **PEGylation Extends Therapeutic Window**: [Pegunigalsidase alfa achieves sustained plasma concentrations](https://practicalneurology.com/articles/2025-oct-27/new-enzyme-replacement-therapy-approved-fda-treat) through covalent PEG crosslinking, reducing dosing frequency from weekly to biweekly administration while maintaining therapeutic efficacy in Fabry disease patients.\n* **Recombinant Production Enables Precision Manufacturing**: [Six FDA-approved enzymes utilize CHO cell expression systems](https://pmc.ncbi.nlm.nih.gov/articles/PMC8714691/), providing superior glycosylation fidelity compared to bacterial or yeast platforms while maintaining native human sequences that minimize immunogenic epitopes.\n* **Mannose-Terminal Modification Directs Macrophage Uptake**: [Imiglucerase\u2019s mannose-exposed glycans](https://accessdata.fda.gov/drugsatfda_docs/label/2002/ceredz061302LB.pdf) achieve preferential macrophage targeting in Gaucher disease through carbohydrate receptor-mediated endocytosis, demonstrating how post-production enzymatic modification fundamentally alters biodistribution patterns.\n* **Orphan Drug Status Accelerates Development**: [Over 39% of 2017 FDA novel drug approvals](https://stratviewresearch.com/market-reports/enzyme-replacement-therapy-market.html)\u00a0addressed orphan diseases, with enzyme therapies benefiting from expedited review pathways, extended market exclusivity.\n\n1. **Analysis of US Food and Drug Administration new drug and biologic approvals, regulatory pathways, and review times, 1980\u20132022:** [Seoane-Vazquez E, Rodriguez-Monguio R, Powers JH 3rd, Sci Rep, 2024](https://doi.org/10.1038/s41598-024-53554-7)\n2. **The past, present, and future of enzyme-based therapies:** [Hennigan JN and Lynch MD, Drug Discov Today, 2022](https://pmc.ncbi.nlm.nih.gov/articles/PMC8714691/)\n3. **FDA approves first enzyme replacement therapy for rare alpha-mannosidosis:** [FDA News Release, February 16, 2023](https://www.fda.gov/drugs/news-events-human-drugs/fda-approves-first-enzyme-replacement-therapy-rare-alpha-mannosidosis)\n4. **Strategies for Glycoengineering Therapeutic Proteins:** [Dammen-Brower K, et al., Front Chem, 2022](https://doi.org/10.3389/fchem.2022.863118)\n5. **A universal GlycoDesign for lysosomal replacement enzymes to enhance therapeutic efficacy:** [Chen YH, et al., Front Bioeng Biotechnol, 2023](https://doi.org/10.3389/fbioe.2023.1128371)\n\n**Precision biologics target enzyme deficiencies through advanced glycoengineering, PEGylation, and targeted delivery strategies**\n\nThe landscape of rare disease treatment has been transformed by enzyme replacement therapies, representing a remarkable convergence of biotechnology innovation and clinical medicine. From the first approval in 1978 to the glycoengineered products of 2023, these therapeutic enzymes demonstrate how protein engineering strategies can address previously untreatable genetic conditions. With over [1,355 pharmaceutical products approved by the FDA between 1980 and 2022](https://www.nature.com/articles/s41467-024-45059-3), enzyme therapies constitute a specialized but critical segment targeting lysosomal storage diseases, metabolic disorders, and cancer-related enzyme deficiencies.\n\nThe development trajectory of enzyme therapeutics reveals an evolution from crude placental extracts to sophisticated recombinant proteins designed with enhanced stability, extended circulation time, and improved cellular targeting. Unlike the approximately [39 unique FDA-approved enzyme therapies documented through 2021](https://pmc.ncbi.nlm.nih.gov/articles/PMC8714691/), current approaches extend beyond simple enzyme replacement to incorporate rational protein engineering, glycan modification, and chemical conjugation strategies that fundamentally alter pharmacokinetic profiles and therapeutic efficacy.\n\nAnalysis of glycoengineering approaches demonstrates the critical role of carbohydrate modifications in determining enzyme biodistribution and cellular uptake. The [mannose-6-phosphate (M6P) receptor pathway](https://pmc.ncbi.nlm.nih.gov/articles/PMC9899006/) mediates lysosomal delivery for disorders like Gaucher disease and Pompe disease, while alternative strategies employ mannose-terminal glycosylation for macrophage targeting or eliminate glycan-receptor interactions entirely to extend circulatory half-life. [Avalglucosidase alfa exemplifies next-generation glycoengineering](https://linkinghub.elsevier.com/retrieve/pii/S0168952523002208), incorporating increased bis-M6P content that enhances CIMPR binding approximately sevenfold compared to first-generation alglucosidase alfa, resulting in superior glycogen clearance in skeletal muscle.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nTypically, 400 ml of sterile LB-amp (see Appendix I) supplemented with 0.3 % (w/v) glucose was inoculated with a colony of pMal-PNGase F-MBP E. coli (kindly provided by M. A. Hall and Dr. K. Hudson, CRC Growth Factors Group, Birmingham). The culture was incubated at 25 \u00b0C with shaking for 24 hours. The culture was diluted 100-fold in LB-amp and grown to OD600 ~0.6 and IPTG added to a final concentration of 0.3 mM. The culture was grown for a further 4 hours at 25 \u00b0C with shaking. The cells were harvested by centrifugation at 5000 \u00d7 g for 5 minutes and the pellet resuspended in 200 ml of lysis buffer consisting of 20 % sucrose and 30 mM Tris-HCl, pH 8.0 (see Appendix III). EDTA was added to a final concentration of 1 mM and the suspension incubated at room temperature with shaking for 10 minutes. The cells were harvested by centrifugation and resuspended in 200 ml of 5 mM magnesium sulphate, followed by incubation on ice for 10 minutes. The lysate was centrifuged at 6000 \u00d7 g for 10 minutes and 20 mM Tris-HCl, pH 8.0 added. The PNGase F was purified by affinity chromatography as detailed in 4.2.4.3 below.\n\n#### 4.2.4.3 Affinity Purification of Recombinant PNGase F\n\nThe chromatography support (Bio-Rad PolyPrep) was packed with 1 ml of amylose resin (Pharmacia) and washed with 5 ml of PNGase F column buffer containing 0.2 M NaCl, 20 mM Tris-HCl, 10 mM mercaptoethanol, pH 8.0 (see Appendix III). The supernatant was passed through the column under gravity at 4 \u00b0C and the column washed with 5 ml of PNGase F column buffer. Finally, protein was eluted with 0.5 ml aliquots of PNGase F elution buffer consisting of 0.2 M NaCl, 20 mM Tris-HCl, 10 mM mercaptoethanol and 10 mM maltose, pH 8.0 (see Appendix III). Protein fractions were analysed by SDS-PAGE and Coomassie assay.\n\n### 4.2.5 Methods of Analysis\n\n#### 4.2.5.1 ELISA\n\nMaxiSorp 96 well plates (Nunc) were coated with 1 \u03bcg/ml mouse anti-human IgG1 Fc monoclonal antibody (Pierce) diluted in PBS. The plates were covered and left for 2 hours at room temperature. The antibody was discarded and the plate then blocked with 1 % (w/v) BSA in PBS for 1 hour followed by two washes with 0.05 % (v/v) Tween-20 in PBS. Standard curves were constructed using 100 \u03bcl/well duplicates of human IgG1 at 500, 375, 250, 125, 62, 31, 16 and 8 ng/ml. Test samples were added at 100 \u03bcl/well in duplicate and the plate left at room temperature for 1 hour. The plate contents were discarded and the plate then washed twice with 0.05 % (v/v) Tween-20 in PBS. Horseradish peroxidase conjugated polyclonal goat anti-human IgG1 Fc (Pierce) was added at a dilution of 1:5000 in 1 % (w/v) BSA in PBS and the plate incubated at room temperature for 1 hour. The plate contents were discarded and the plate then washed twice with 0.05 % (v/v) Tween-20 in PBS followed by two washes with PBS. Substrate development was carried out using 8 mg of 1,2-phenylenediamine (OPD) (Sigma) as chromogen dissolved in 12 ml of citric acid phosphate buffer consisting of 0.035 M citric acid, 0.067 M disodium hydrogen phosphate, pH 5.0 (see Appendix III). Chromogen solution was activated by addition of 5 \u03bcl of hydrogen peroxide and the solution added to each well and the colour developed, in the dark, for 10\u201315 minutes. The reaction was terminated by the addition of 75 \u03bcl of 1 M sulphuric acid to each well and the absorbance read at 490 nm.\n\n#### 4.2.5.2 Western Blot Analysis\n\nWestern blots were carried out as detailed in Chapter 3 except the primary polyclonal antibody consisted of sheep anti-LIFR (Binding Site) at a dilution of 1:10000 in PBS containing 5 % (w/v) BSA and 0.1 % (v/v) Tween-20.\n\n#### 4.2.5.3 Iso-electric Focussing\n\n5.5 % polyacrylamide iso-electric focussing (IEF) gels were prepared using broad range ampholines in the pI range of 3\u201310 (Bio-Rad) as in Table 4-3 below. Gels were run on a Mini-Protean II gel apparatus (Bio-Rad).\n\n| Component | Volume |\n|---|---|\n| 30 % (w/w) Acrylamide / 0.8 % (w/w) N',N'-bis-methylene-acrylamide | 1.28 ml |\n| Water | 3.89 ml |\n| 50 % (v/v) Glycerol | 1.40 ml |\n| Ampholines (pI 3\u201310) | 387 \u03bcl |\n| 10 % (w/v) Ammonium Persulphate | 29 \u03bcl |\n| TEMED | 14 \u03bcl |\n| Total | 7.00 ml |\n\n*Table 4-3.* Table showing composition of 5.5 % polyacrylamide IEF gels.\n\nSamples were diluted 1:1 in loading buffer consisting of 115 \u03bcl water, 60 \u03bcl 50 % (v/v) glycerol and 30 \u03bcl ampholines. Separation was carried out for approximately 3 hours at 150 V (15 mA maximum) at 4 \u00b0C. Anolyte and catholyte consisted of 0.100 M acetic acid and 0.02 M sodium hydroxide, respectively. IEF standards containing phycocyanin (pI 4.5 and 4.8), equine myoglobin (pI 6.8), human haemoglobin A (pI 7.1) and human haemoglobin C (pI 7.5) (Bio-Rad) were run along with a control sample of OSM.\n\n#### 4.2.5.4 Dynamic Light Scattering\n\nSamples of LIFR-FN3-3C-Fc were filtered through a 0.1 \u03bcm membrane (Whatman) and injected into a DynaPro-LS Sizing Instrument (Protein Solutions Inc.). Typically, 0.5 ml of 2 mg/ml sample was injected and measurements taken over a 10 minute period at 20 \u00b0C.\n\n#### 4.2.5.5 Receptor Competition and Titration Assays\n\nReceptor competition and titration assays were carried out as detailed in Chapter 3 with the exceptions that Protein-G was used in place of Protein-A as better signal to noise ratios were reported (K. Hudson, CRC Growth Factor Group, Birmingham, personal communication). For titration assays biotinylated LIF was used at saturating concentrations of 1000 ng/ml and FGFR2 included as control. For competition assays 200 ng/ml of biotinylated LIF was used as competitor and OSM included as control.\n\n### 4.2.6 Production, Purification and Characterisation of LIFR-LIF Complex\n\n#### 4.2.6.1 Expression and Purification of LIF", "# The Invisible Invasion: How Microscopic Plastics Are Infiltrating Human Organ Systems\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-13 on marcdeller.com (categories: blog, microplastics).*\n\n* [The potential impacts of micro-and-nano plastics on various organ systems in humans](https://www.thelancet.com/journals/ebiom/article/PIIS2352-3964(23)00467-X/fulltext)\n\n* **Cardiovascular Threat**: Microplastics accumulate in arterial plaques, increasing [major adverse cardiovascular event risk by 4.5-fold](https://www.ahajournals.org/doi/10.1161/CIRCULATIONAHA.124.069801)\n* **Neurological Disruption**: [Nanoplastics cross blood-brain barriers](https://neurosciencenews.com/nanoplastics-memory-neuroplasticity-299684/), disrupting mitochondrial energy production essential for cognitive function\n* **Renal Dysfunction**: Plastic particles [damage kidney filtration systems](https://pmc.ncbi.nlm.nih.gov/articles/PMC12500543/) through oxidative stress and inflammatory pathway activation\n* **Respiratory Infiltration**: [Daily inhalation of 68,000 microplastic particles](https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0328011) overwhelms natural lung clearance mechanisms\n* **Reproductive Contamination**: Microplastics detected in [69% of female and 55% of male reproductive fluids](https://www.eurekalert.org/news-releases/1088676), raising fertility concerns\n* **Systemic Acceleration**: [Brain plastic concentrations increased 50%](https://www.nature.com/articles/s41591-024-03453-1) between 2016-2024, indicating rapid environmental accumulation\n\n1. **The potential impacts of micro-and-nano plastics on various organ systems in humans:** [Nurshad Ali, et al., eBioMedicine](https://www.thelancet.com/journals/ebiom/article/PIIS2352-3964(23)00467-X/fulltext)\n2. **Microplastics and nanoplastics: emerging threats to cardiovascular health:** [Hamza Irfan, et al., PMC Cardiovascular Research](https://pmc.ncbi.nlm.nih.gov/articles/PMC11918686/)\n3. **Nanoplastics Disrupt Brain Energy, Threatening Memory Function:** [Gavin Davey, et al., Neuroscience News](https://neurosciencenews.com/nanoplastics-memory-neuroplasticity-299684/)\n4. **The detrimental effects of microplastic exposure on kidney function:** [Na Jiang, et al., PMC Nephrology](https://pmc.ncbi.nlm.nih.gov/articles/PMC12500543/)\n5. **Polystyrene microplastics induce liver fibrosis and lipid deposition:** [Qingwen Li, et al., Nature Scientific Reports](https://www.nature.com/articles/s41598-025-86810-5)\n6. **Micro- and Nanoplastics as Disruptors of the Endocrine System:** [Hanna J. Tyc, et al., PMC Endocrine Disruption](https://pmc.ncbi.nlm.nih.gov/articles/PMC12249724/)\n\n**Tiny Particles, Massive Health Implications as Research Reveals Widespread Tissue Contamination**\n\nThe once-theoretical concern about plastic pollution infiltrating the human body has transformed into an urgent clinical reality. Recent groundbreaking research from 2020-2025 demonstrates that micro- and nanoplastics are not merely environmental pollutants but active biological disruptors, systematically infiltrating critical organ systems and triggering cascading health consequences that medical professionals are only beginning to understand.\n\nThe evidence presents a disturbing panorama of systemic contamination. Microplastics and nanoplastics\u2014particles smaller than 5 millimeters and 100 nanometers respectively\u2014have been definitively detected in human arterial plaques, brain tissues, reproductive fluids, lung alveoli, kidney tissues, and hepatic structures. What makes these findings particularly alarming is the accelerating accumulation rate: brain samples from 2024 contained 50% more plastic particles than those from 2016, indicating rapid environmental saturation translating directly into human bioaccumulation.\u200b\n\nThe cardiovascular system emerges as a primary target for plastic-mediated damage. Researchers have identified microplastics in arterial tissues, with concentrations significantly higher in atherosclerotic plaques compared to healthy vessel walls. These particles induce oxidative stress and mitochondrial dysfunction, creating a perfect storm for cardiovascular pathology. Studies reveal that patients with plastic-contaminated arterial plaques face a 4.5-fold increased risk of myocardial infarction, stroke, or mortality, transforming plastic exposure from an environmental concern into a direct cardiovascular threat mechanism.\u200b\n\nThe neurological implications prove equally devastating. Nanoplastics demonstrate the unprecedented ability to cross the blood-brain barrier, with recent Trinity College Dublin research showing these particles specifically target mitochondrial electron transport chains. This disruption compromises brain energy metabolism, potentially contributing to the rising global incidence of neurodegenerative diseases. The correlation between mid-20th century plastic proliferation and increasing rates of conditions like Alzheimer\u2019s and Parkinson\u2019s disease suggests environmental plastic exposure may represent a previously unrecognized neurological risk factor.\u200b\n\nRespiratory system vulnerability becomes apparent through inhalation exposure pathways. Microplastics smaller than 2.5 micrometers penetrate deep into alveolar spaces, triggering inflammatory cascades and oxidative stress responses. Indoor air quality studies reveal humans inhale approximately 68,000 microplastic particles daily, creating chronic low-level exposure that overwhelms natural clearance mechanisms and promotes pulmonary accumulation.\u200b\n\nRenal toxicity manifests through glomerular filtration disruption and tubular epithelial damage. Microplastics activate inflammatory pathways including NF-\u03baB signaling, leading to proteinuria, oxidative stress, and progressive nephron dysfunction. The kidneys\u2019 filtration role paradoxically increases their vulnerability, as these organs concentrate plastic particles while attempting systemic clearance.\u200b", "# Alphafold Local Install on Mac\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-18 on marcdeller.com (categories: ai).*\n\n* **Why Install AlphaFold Locally?**\n* **Full Control & Independence**: Run protein structure predictions without relying on cloud services, API limits, or internet connectivity.\n* **Performance Boost**: Leverage Apple Silicon (M1/M2) hardware or other local GPUs for faster runtimes and reproducibility.\n* **Customization**: Fine-tune environments, dependencies, and workflows to fit your research needs (e.g., custom databases or pipelines).\n* **Security & Privacy**: Keep proprietary sequences and data entirely on local infrastructure.\n* **Seamless Workflow Integration**: Embed AlphaFold outputs directly into existing lab scripts, automation, or visualization tools.\n\n1. <https://macinchem.org/cheminformatics-and-compchem-on-apple-silicon/installing-alphafold2-on-apple-silicon/>\n2. <https://www.anaconda.com/products/navigator>\n3. <https://rest.uniprot.org/uniprotkb/P13725.fasta>\n\n**Here is an easy way to install alphafold2 on OSX in about 20 minutes and 7 easy steps. Based on the instructions from reference 1**:\n\nI\u2019ve added a few fixes for dependencies (i.e. kalign and mmseqs2), permissions changes, combining installs, and corrections to the paths for newer systems.\n\nI recommend you install anaconda navigator to get the base python packages and help making a root environment, you can do this on the command line, but I find the GUI easier to manage so you don\u2019t end up in package hell! This prevents the need for complex dockers, containers, and environments:\n\nEnter these on your command line in order, and you should be up and running quickly:\n\n```\n/bin/bash -c \"$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)\"\nbrew install wget cmake gnu-sed brewsci/bio/hh-suite kalign mmseqs2 --cask miniforge\nmkdir /Users/dellboy/Alphafold\ncd /Users/dellboy/Alphafold\nwget https://raw.githubusercontent.com/YoshitakaMo/localcolabfold/main/install_colabbatch_M1mac.sh\nchmod 777 ./install_colabbatch_M1mac.sh\nbash ./install_colabbatch_M1mac.sh\n```\n\nWhen the installs are finished copy your fasta file into /Users/dellboy/Alphafold, for example: [P13725.fasta](https://rest.uniprot.org/uniprotkb/P13725.fasta) (my first experimental structure!).\n\nThen run the following command from your prompt, assuming you are still in /Users/dellboy/Alphafold directory.\n\n```\n./localcolabfold/colabfold-conda/bin/colabfold_batch --amber --templates --num-recycle 3 --model-order 1,2,3,4,5 ./P13725.fasta FFA2output\n```\n\nThis run took just over 1 hour on my Mac Studio M2 Ultra.", "# Molecular Velcro: Catching What We Cannot See\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nThe practical implications of pairing removal with an intrinsic readout are substantial. Water management and environmental health protection have historically separated the act of intervention from the act of measurement: one infrastructure removes contaminants, another monitors whether they are present. The Fe\u2083O\u2084@SiO\u2082@CP platform collapses this distinction into a single workflow (mix, separate magnetically, read the color response), completing the cycle in approximately three minutes under real-world water conditions, including freshwater, seawater, and consumer plastic leachates. The platform also sustains performance across five regeneration cycles, operates at environmentally relevant contamination loads as low as 0.05 mg/L, and carries an estimated removal cost of US$0.035 per gram, characteristics that move it from demonstrable to deployable. Reviewed against the broader landscape of magnetic adsorbent systems compiled by [Vohl et al. in Nanomaterials](https://doi.org/10.3390/nano14141179), which documents their promise across a range of micro- and nanoplastic removal strategies, the key differentiator here is not capture efficiency per se, but the intrinsic coupling of that capture to a detection signal: the very step that most magnetic sorbent approaches have left to downstream spectroscopy.\n\nThe longer-term implications reach beyond water treatment. As microplastics are now confirmed within human circulatory and pulmonary systems, strategies for in-body sequestration are an emerging area of discussion, and capture materials with biocompatible polydopamine coatings represent a plausible starting point for that conversation. The analytical challenges outlined by [Ivleva in Chemical Reviews](https://doi.org/10.1021/acs.chemrev.1c00178), namely the extraordinary diversity of particle size, shape, density, and polymer chemistry that makes any single detection method inadequate, are the same challenges that a multi-interaction adhesive surface is structurally suited to address. Full translation from environmental remediation to therapeutic application would require substantial further validation under physiological conditions, but the conceptual bridge between \u201cmolecular Velcro for water\u201d and \u201cmolecular Velcro for the body\u201d is shorter than\u00a0ever. For now, the coupling platform represents the most coherent answer yet to a question that has stymied two decades of microplastics research: how do you find, catch, and confirm an invisible contaminant, in the same place, at the same time, without a laboratory?", "# Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\nAuthors: Hummel, J.R.; Xiao, K.-J.; Yang, J.C.; Epling, L.B.; Mukai, K.; Ye, Q.; Xu, M.; Qian, D.; Huo, L.; Weber, M.; Roman, V.; Lo, Y.; Drake, K.; Stump, K.; Covington, M.; Kapilashrami, K.; Zhang, G.; Ye, M.; Diamond, S.; Yeleswaram, S.; Macarron, R.; Deller, M.C.; Wee, S.; Kim, S.; Wang, X.; Wu, L.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2024\n\nADME, absorption distribution metabolism excretion; AUC, area under the concentration\u2212time curve; BQL, below quantitation limit; CCNE1, cyclin E1 gene name; CDK, cyclin-dependent kinase; Cmax, maximum observed plasma concentration; cmpd, compound; cyno, cynomolgus monkey; CYP, cytochrome p450; CYP TDI, time-dependent inhibition of cytochrome p450; E2F, E2 transcription factor; F, fraction of an administered dose of unchanged drug that reaches systemic circulation; FDA, Food and Drug Administration; FF, protein binding free fraction; HBF, hepatic blood flow; h, hours; h-Cl, human microsomal intrinsic clearance; hERG, human ether a-go-go; IC50, half-maximal inhibitory concentration; IV, intravenous; NPM, nucleophosmin; PD, pharmacodynamic; PK, pharmacokinetic; PO, per os; Rb, retinoblastoma protein; SAR, structure\u2212activity relationship; SGF, simulated gastric fluid; t1/2, half-life; Vdss, volume of distribution at steady state; WB, human whole blood.\n\n## References\n\n*(reference list omitted)*", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices\nAuthors: Deller, M.C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\n| Family | Protein | Total number of residues | Number of structurally equivalent residues | Number of sequence identities within equivalents | RMSD (\u00c5) |\n|---|---|---|---|---|---|\n| gp130 family | Murine LIF | 172 | 145 | 31 (21 %) | 2.10 |\n| gp130 family | Human LIF | 159 | 138 | 28 (20 %) | 2.27 |\n| gp130 family | Human IL-6 | 157 | 127 | 22 (17 %) | 2.40 |\n| gp130 family | Human CNTF | 150 | 120 | 14 (12 %) | 2.95 |\n| \u2014 | Human OSM (I222) | 149 | 147 | 147 (100 %) | 0.72 |\n| Growth hormone family | Human GCSF | 172 | 127 | 25 (19 %) | 2.65 |\n| Growth hormone family | Human Epo | 154 | 120 | 7 (6 %) | 2.80 |\n| Growth hormone family | Human GH | 186 | 120 | 9 (8 %) | 3.08 |\n| Growth hormone family | Human Prl | 186 | 133 | 12 (9 %) | 3.17 |\n\n*Table 6-1.* Table showing structural comparison of C\u03b1 positions ranked according to RMSD values of human OSM with various other helical cytokines of the gp130 and growth factor related families. Structural superpositions and RMSD values calculated using the program SHP (Stuart et al., 1979). GCSF-GCSFR coordinates kindly provided by K. Morikawa, Biomedical Engineering Research Council, Osaka, Japan. All other co-ordinates from PDB.\n\nStructural superposition of human OSM onto human LIF, human IL-6, human CNTF, human GCSF and human Epo are shown below in Figure 6-6. To a first approximation the length and orientation of the four main helical regions is conserved between the five molecules. However, several subtle differences are observed. Starting at the N-terminus the loop preceding helix A is anchored to helix C by two disulphides in LIF, whereas only one disulphide is present in this region in OSM. These anchoring disulphide bridges are not present in IL-6, CNTF, Epo or GCSF. In the IL-6 structure this loop region deviates away from the body of the molecule and in the GCSF structure the loop passes over the end of the helical bundle, with the side chain of Leu 10 packing between the helical bundle. CNTF is unique in having no disulphide bridges in the structure, and the lack of an anchoring N-terminal disulphide linkage results in high flexibility for the first 10 residues, although a similar course to both OSM and LIF is suggested.\n\nHelix A adopts a similar course in all structures although OSM differs from others in the severity of the kink. Both CNTF and LIF have minor kinks due to disruption of the main chain hydrogen bonding around residues Thr 32 and Ser 36, respectively (McDonald et al., 1995; Robinson et al., 1994). Both these residues bind waters through the O\u03b31 atom and a similar pattern is observed in OSM through an equivalent residue Thr 27 (see Figure 6-4 (a)). Helix A of Epo differs from all the other structures in being approximately one turn shorter at both the N- and C-terminus. In OSM the C-terminal section of helix A adopts a 3-10 helical conformation; a secondary structural feature not observed in any of the other structures. The unique use of 3-10 helix is also seen in the N-terminal portion of the B-C loop.\n\nThe A-B loop packs against the external face of helix D approximately half way along its length. Similar paths are observed in OSM, LIF, IL-6 and CNTF. Epo, however, differs markedly in this region to all the other structures as the C-terminal region of the A-B loop adopts a path at approximately right angles to that of the other structures, so crossing the internal face of helix D at its N-terminus. The altered path of Epo is mainly due to the shorter length of the A-B loop and a sharp turn stabilised by a disulphide bridge between residues 29 and 33. In a similar fashion the A-B loop of GCSF adopts an altered path around the external face of helix D at its N-terminus. The altered path again due in part to a disulphide bond stabilised turn between residues 37 and 43 and the insertion of an extra helical segment. In common, both LIF and OSM, have a disulphide bridge anchoring the first part of the A-B loop to the top of helix D. In comparison a similar region of the A-B loop of IL-6 is anchored to helix B. No density is present for the A-B loop of CNTF although a similar course is suggested. A major distinction of the OSM structure is the presence of two helical regions in the A-B loop; one running from the C-terminus of helix A and the second running into the N-terminus of helix B. In comparison, LIF and GCSF contain only the first helical region running from the C-terminus of helix A and IL-6 and Epo contain only the second running into the N-terminus of helix B.\n\nBoth helices B and C match closely between the structures with LIF, IL-6, Epo and GCSF all having short interconnecting B-C loop regions. However, this loop region in OSM has a 5 residue insertion that forms just over 1 turn of helix that protrudes from the top of the helical bundle. This feature has so far not been observed in any other members of the gp130 cytokine family, although similar helical features are observed in GH and short chain cytokines such as GM-CSF and IL-4. The N-terminus of helix C is approximately 1 turn shorter in both Epo and GCSF.\n\nThe C-D loop has not been fully modelled into OSM and CNTF, although the orientation of residues preceding this loop suggests a similar route is taken to that of LIF, IL-6 and CNTF in which the loop passes over helix B approximately half way along its length. Deviation from this pattern is observed in Epo as the C-terminal section of the loop protrudes away from the main helical bundle as a series of residues adopt a turn conformation. GCSF, IL-6 and LIF all show similar termination positions of helix D to within 2 or 3 residues.\n\nHelix D shows similar positioning in all structures, although that of Epo is ~2 turns shorter and that of CNTF is ~2 turns longer with the extended section packing towards the N-terminus of helix A, presumably in a similar fashion to non-truncated wildtype OSM.", "# Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands\nAuthors: Pozharski, E.; Deller, M. C.; Rupp, B.\nSource: Protein Crystallography: Methods and Protocols (A. Wlodawer, Z. Dauter & M. Jaskolski, eds.), Methods in Molecular Biology vol. 1607, Chapter 25, pp. 611\u2013625, Springer, 2017\n\nThe fit of a model to its density is quantified by local, per-residue metrics rather than global ones. The chapter names the Real Space Correlation Coefficient (RSCC), the Real Space R-Value (RSR) and composite measures such as the Local Ligand Density Fit (LLDF), and notes that these can be obtained from the PDB validation report and displayed per residue in COOT. Interpretation is given concretely: a perfect fit would give an RSCC of 1.0, values between 0.8 and 1.0 are commonly considered good, and the lower the RSCC the greater the uncertainty in positioning the model relative to the density. RSR is somewhat scale-dependent, and to compensate the PDB uses the RSRZ value, a normalized statistic quantifying how many standard deviations a residue's density fit departs from the mean. Figure 1 of the chapter demonstrates, using ATP-bound structures deposited in the PDB and RSCC and occupancy-weighted B-factor (OWAB) values calculated with Twilight, that RSCC is generally better at higher resolution and lower ligand B-factors \u2014 showing that local ligand metrics depend closely on global properties of the experiment such as resolution. A table of validation software accompanies this section.\n\n### 2.3. Difference and omit electron density maps\n\nDifference maps with coefficients mFo \u2212 DFc can be displayed simultaneously with the 2mFo \u2212 DFc map in COOT, and are read directly: positive difference density appears where parts of the ligand model are missing, and negative difference density where the ligand model is present but unsupported by density. The chapter then singles out the omit difference map as the most definitive proof positive for ligand density. An omit map is obtained by refining the model without the part under examination \u2014 the ligand is deleted and refinement rerun \u2014 so that if the ligand was truly present in the crystal, strong positive difference density in the shape of the ligand appears. Because omit maps are not directly available from public databases, they must be generated for each complex, which requires only general knowledge of file formats and crystallographic software: delete the ligand by editing the PDB file or using COOT or PyMOL, then run a cycle of standard refinement. For thorough examination and to confirm the density is robustly reproduced, the authors recommend using more than one refinement program, naming REFMAC, phenix.refine and BUSTER-TNT. All modern refinement programs output map coefficients for both 2mFo \u2212 DFc and mFo \u2212 DFc maps that can be imported straight into COOT.\n\n### 2.4. Model bias\n\nOmit maps matter chiefly because of model bias. Traditional maps available from public databases such as EDS and PDB_REDO are constructed using phases originating from a model that already includes the ligand in question, so it is unsurprising that some spurious density is always present that appears to confirm the ligand's presence. The chapter illustrates this side by side using a concanavalin A/peptide complex at 1.93 \u00c5 resolution (PDB entry 1jw6): the biased 2mFo \u2212 DFc map from EDS is compared with maps from corrected sources that use difference omit maps, namely PDB_REDO and BUSTER-TNT after refinement with the peptide omitted, all contoured at 1\u03c3. The comparison shows how convincing biased density can look and how much less of the peptide survives when the phases are not contaminated by the modelled ligand. The practical lesson is that a researcher who inspects only the deposited 2mFo \u2212 DFc map has not, in the authors' framework, validated the ligand at all \u2014 they have merely confirmed that the depositor's model is self-consistent.\n\n### 2.5. Incomplete ligand models\n\nThe authors define incomplete models as those in which peptides or other small molecules are \"enthusiastically\" modelled beyond what is clearly traceable in the density. The statistical consequence is perverse: because local quality scores are typically averaged over the whole ligand, correctly modelled regions mask the badly modelled ones, so overall measures such as ligand RSCC, RSR, RSRZ and LLDF are artificially improved and the model gains a false appearance of correctness. Similarly, the percentile of Ramachandran outliers can be very low for short peptide ligands simply because there are few residues to be outliers. These examples support the chapter's central claim that the only reliable validation of a protein\u2013ligand complex is local inspection of the ligand and its fit to the density, which reveals which parts of the model are experimentally defined. The authors add an important nuance: trimming the ligand model back to only what is visible, though crystallographically honest and superficially parsimonious, can itself lead to incorrect functional assumptions \u2014 for example, modelling only five residues of a 33-residue peptide will usually leave a model that needs independent supporting evidence to establish the correct register of the peptide.\n\n### 3. Examination against prior expectations, and 3.1 stereochemical restraints", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices\nAuthors: Deller, M.C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nAnalysis of Figure 6-5 (a) and Figure 6-5 (c) below demonstrates that close packing of the A:D and B:C helical pairs results in a tight solvent inaccessible core with the outer faces of the helices having many charged residues and the inner faces consisting of many hydrophobic residues. This solvent inaccessible core of the protein is composed of two groups of aromatic stacking interactions in which Phe 56, Tyr 173, Phe 169 and Phe 176 form one group and Phe 70, Phe 185 and Trp 187 form the second. This distribution emphasises the hydrophobic nature of helix D, which donates all of the aromatics to these groupings with the exception of Phe 56 (A-B loop) and Phe 70 (B helix). Helix D also shows a clustering of around seven positively charged residues on its outer surface which presumably affects the surface electrostatic potential of site 2 interactions in a similar fashion to that proposed for CNTF (Panayotatos et al., 1995; see 6.3.3.1 for discussion) and GH (Cunningham and Wells, 1993). In contrast to the large number of hydrophobic interactions stabilising the fold, no hydrogen bonds between the main helices are observed (see Figure 6-3).\n\nAnalysis of B-factor values in Figure 6-5 (b) shows that the tight packaging of the helices also results in a rigid helical bundle, with only the loop regions and the C-terminus end of helix B exhibiting any motion. Average B-factors are 30.5 \u00c5\u00b2 for the P212121 structure (see Figure 6-8 for comparison with I222 structure).\n\nThe N-terminal loop preceding helix A (Glu 4 to Gly 9) is tethered to the N-terminus of helix C by one of the two disulphide bridges spanning Cys 6 and Cys 127. In a similar fashion, the second disulphide bridge, between Cys 49 and Cys 167, anchors the start of the A-B loop to the N-terminal region of helix D. The A-B loop consists of two separate \u03b1-helical regions running between Pro 43 to Arg 46 and Glu 59 to Gly 64. Residues between the two helical regions of the A-B loop adopt an extended conformation and pack closely against helix D, with Phe 56 also contributing to the hydrophobic molecular core. The B-C loop contains a second small region of 3-10 helix running between residues Ala 95 and Asp 97 with a traditional \u03b1-helical conformation being adopted for the rest of the helix to residue Ser 101.\n\n*Figure 6-5.* Figure showing ribbon representation of OSM with C\u03b1 coloured according to (a) relative solvent accessibility, (b) refined B-factors and (c) charge and hydrophobicity. Relative solvent accessibility values are coloured from the minimum value of 0 % to the maximum value of 100 % from grey to blue, respectively. B-factors are coloured from the minimum value of 6.3 \u00c5\u00b2 to the maximum value of 84.9 \u00c5\u00b2 from blue to red, respectively (average B-factor 30.5 \u00c5\u00b2). Positively charged residues (Arg, Lys or His) are coloured blue, negatively charged residues (Asp or Glu) red and hydrophobic residues (Trp, Tyr or Phe) yellow. Figure produced using NACCESS (Hubbard and Thornton, 1993), BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994).\n\n### 6.2 Structural Comparisons\n\nThe helical cytokine family has previously been classified in terms of the helix length; short chain or long chain (Nicola, 1994). Short chain helical cytokines analysed structurally include IL-5 (Milburn et al., 1993), M-CSF (Pandit et al., 1992), IFN-\u03b3 (Ealick et al., 1991; Samudzi et al., 1991) and IL-3 (Feng et al., 1996) which are dimeric, and GM-CSF (Diederichs et al., 1991; Walter et al., 1992b), IL-2 (Bazan, 1992) and IL-4 (Powers et al., 1993; Walter et al., 1992b; Wlodawer et al., 1992) which are monomeric (see Chapter 1 for overview).\n\nLong chain helical cytokines resolved structurally include GH (Abdel Meguid et al., 1987; de Vos et al., 1992), Epo (Livnah et al., 1998; Livnah et al., 1996; Syed et al., 1998), Leptin (Zhang et al., 1997a), GCSF (Aritomi et al., 1999; Hill et al., 1993; Lovejoy et al., 1993), IL-10 (Zdanov et al., 1995), IFN-\u03b2 (Senda et al., 1995) and IFN-\u03c4 (Radhakrishnan et al., 1999) with the later two adopting a five helix bundle topology.\n\n#### 6.2.1 Haemopoietin Cytokine Family\n\nThe long chain helical cytokine family can be further sub-divided into those using the gp130 receptor or the interferon receptor and this present structure represents the fourth member of the gp130 cytokine family to be solved structurally. Members of the gp130 family of cytokines delineated structurally so far include IL-6 (Somers et al., 1997; Xu et al., 1997), CNTF (McDonald et al., 1995) and LIF (Robinson et al., 1994 [X-ray]; Purvis and Mabbutt, 1997 [NMR]). For the purpose of this study medium to high-resolution X-ray structures will be used for comparative purposes with coordinates deposited in the PDB, except for GCSF-GCSFR coordinates kindly provided by K. Morikawa, Biomedical Engineering Research Council, Osaka, Japan.\n\nAnalysis of root mean squared deviation (RMSD) values as in Table 6-1 below confirms that, structurally OSM belongs to the gp130 family of cytokines. Murine LIF shows greatest similarity to OSM with an RMSD value of 2.10 \u00c5 over 145 structurally equivalent residues and a sequence identity of 21 %. Human LIF is close behind with an RMSD value of 2.27 \u00c5 over 138 structurally equivalent residues and a sequence identity of 20 %. Human IL-6 and human CNTF are the next best matches with RMSD values typically greater than 2.40 \u00c5 and a sequence identity typically greater than 12 %. Human OSM shows more distant structural similarity to the growth hormone family of helical cytokines with RMSD values typically greater than 2.80 \u00c5 and a sequence identity typically less than 9 %. Within the growth hormone family the greatest similarity is shared with Epo with an RMSD value of 2.80 \u00c5, although sequence identity is exceptionally low at only 6 %. However, OSM clearly shows greater similarity to GCSF with an RMSD value of 2.80 \u00c5 over 127 equivalencies and a sequence identity of 19 %.", "# The Unsung Hero: How a Single Chlorine Atom Transforms Drug Discovery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-15 on marcdeller.com (categories: blog, sb).*\n\n* \u201cMagic Chloro\u201d: Profound Effects of the Chlorine Atom in Drug Discovery\n\n* **Potency Amplification:** [Single chlorine substitution achieves 10-100,000x potency improvements](https://pubmed.ncbi.nlm.nih.gov/37014977/) through optimized halogen bonding and lipophilic contacts with protein targets.\n* **Pharmacokinetic Optimization:** [Chlorine modulates metabolic clearance and extends drug half-life](https://pmc.ncbi.nlm.nih.gov/articles/PMC10687521/) while maintaining favorable ADME profiles without increasing toxicity.\u200b\n* **Halogen Bonding Revolution:** [Chlorine functions as a directional Lewis acid](https://pubs.rsc.org/en/content/articlelanding/2020/cc/d0cc00841a) forming precise interactions with backbone carbonyls and side-chain nucleophiles.\n* **Selectivity Engineering:** [Strategic chlorination improves target selectivity](https://chemrxiv.org/engage/chemrxiv/assets/orp/resource/item/62f1dd4c69f3a528d2b0d366/original/magic-chloro-profound-effects-of-the-chlorine-atom-in-drug-discovery.pdf) by >3,000-fold through exploitation of binding pocket differences.\n* **Synthetic Accessibility:** [New photocatalytic methods enable room-temperature precision chlorination](https://news.rice.edu/news/2025/chemists-create-eco-friendly-method-make-chlorine-based-materials-drugs-and-chemicals) with sustainable catalysts.\n* **Computational Design:** [Machine learning models trained on halogen bonding data](https://pmc.ncbi.nlm.nih.gov/articles/PMC12199455/) enable prospective prediction of chlorination benefits.\u200b\n\n1. **Accelerating drug discovery with a single carbon atom:** [Sharma et al., *Journal of the American Chemical Society*, 2025](https://www.sciencedaily.com/releases/2025/05/250506131509.htm)\n2. **Halogen bonding for molecular recognition:** [Auffinger et al., *Chemical Communications*, 2020](https://pubs.rsc.org/en/content/articlelanding/2020/cc/d0cc00841a)\n3. **Highlights on U.S. FDA-approved halogen-containing drugs in 2024:** [Ali et al., *European Journal of Medicinal Chemistry*, 2025](https://pubmed.ncbi.nlm.nih.gov/39947048/)\n4. **Chemists create eco-friendly method to make chlorine-based materials:** [West et al., *Nature Synthesis*, 2025](https://news.rice.edu/news/2025/chemists-create-eco-friendly-method-make-chlorine-based-materials-drugs-and-chemicals)\n5. **Halogen Bonds (Science/AAAS Blog):** [In the Pipeline, 2013](https://www.science.org/content/blog-post/halogen-bonds)\n6. **New Halogen-Containing Drugs Approved by FDA in 2021:** [*Pharma Excipients*, 2024](https://www.pharmaexcipients.com/news/halogen-containing-drugs/)\n\n**From overlooked substituent to pharmaceutical powerhouse\u2014the chlorine atom\u2019s profound impact on potency, selectivity, and pharmacokinetics**\n\nThe quest for breakthrough therapeutics often hinges on molecular fine-tuning, where even the smallest atomic change can yield extraordinary results. Among medicinal chemistry\u2019s most powerful yet underappreciated tools is the chlorine atom\u2014a single substituent capable of transforming mediocre drug candidates into potent pharmaceutical agents. With over 250 FDA-approved chlorine-containing drugs currently in clinical use, this halogen has quietly revolutionized modern medicine, yet its profound effects have remained largely unexamined until recently. The phenomenon, now termed the \u201cmagic chloro effect\u201c, demonstrates that replacing a single hydrogen atom with chlorine can improve drug potency by 10-fold to an astounding 100,000-fold while simultaneously optimizing pharmacokinetic parameters such as clearance, half-life, and bioavailability.\n\nRecent advances in computational chemistry and structural biology have illuminated the molecular mechanisms underlying chlorine\u2019s remarkable efficacy. Unlike traditional medicinal chemistry approaches that viewed halogens merely as hydrophobic substituents, contemporary research reveals that [chlorine atoms function as sophisticated Lewis acids](https://pubs.rsc.org/en/content/articlelanding/2020/cc/d0cc00841a) capable of forming highly directional halogen bonds with protein targets. These non-covalent interactions\u2014featuring geometries as precise as hydrogen bonds\u2014allow chlorinated compounds to engage backbone carbonyl oxygens, hydroxyl side chains, and aromatic \u03c0-systems within binding pockets. The strategic positioning of chlorine substituents creates a lipophilic anchor (increasing logD by approximately 0.6-1.0 units) while simultaneously establishing specific electrostatic contacts that enhance binding affinity far beyond what lipophilicity alone would predict.\n\nThe pharmaceutical landscape spanning 2020-2025 showcases chlorine\u2019s enduring relevance. Among the [16 halogen-containing drugs approved by the FDA in 2024](https://pubmed.ncbi.nlm.nih.gov/39947048/), fluorine and chlorine dominated, with chlorinated molecules addressing diverse therapeutic areas from oncology to infectious disease. Notably, compounds like **Cobenfy** (xanomeline and trospium chloride, approved September 2024 for schizophrenia) exemplify how chlorine-containing frameworks continue to enable novel mechanisms of action. The 2023 landmark review \u201cMagic Chloro: Profound Effects of the Chlorine Atom in Drug Discovery\u201d documented over 600 medicinal chemistry optimization campaigns where chlorine substitution achieved at least 10-fold potency improvements, with more than 100 cases exceeding 100-fold enhancement.", "# Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33\nAuthors: Kong, L.; Giang, E.; Nieusma, T.; Robbins, J.B.; Deller, M.C.; Stanfield, R.L.; Wilson, I.A.; Law, M.\nSource: Journal of Virology, 2012\n\nWe thank Kristin Cogburn and Arthur Kim for technical support in cell culture and antibody production, Henry Tien and Thomas Clayton for help in setting up initial crystal screens using the Crystalmation robot, and Jane Verenini for manuscript formatting. Portions of this research were carried out at the Stanford Synchrotron Radiation Lightsource (SSRL), a Directorate of the SLAC National Accelerator Laboratory and an Office of Science User Facility operated for the U.S. Department of Energy (DOE) Office of Science by Stanford University.\n\nThe SSRL Structural Molecular Biology Program is supported by the DOE Office of Biological and Environmental Research, NIH's National Center for Research Resources Biomedical Technology Program (P41RR001209), and the National Institute of General Medical Sciences (NIGMS). This work is supported by NIH grants AI79031 (to M.L.), AI84817, and GM U54 GM094586 (to I.A.W.) and the Skaggs Institute (I.A.W.).\n\nThis is TSRI manuscript number 21874.\n\nThe authors declare no conflict of interest.\n\n## References\n\n*(reference list omitted)*", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices\nAuthors: Deller, M.C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\n#### 6.5.2 Summary of Mutagenesis Data\n\nSeveral residues have been implicated to play a role in site 3 and site 2 interactions through alanine mutagenesis (Deller et al., 2000) as shown below in Figure 6-13 and Table 6-3. All active residues have side chains that are solvent accessible to 27 % or greater. Residues Phe 160 and Lys 163 form the main interaction site for LIFR and both reside at the N-terminus of helix D. Site 2 interactions with gp130 are dominated by residues on the A and C helices, towards the N-terminal loop region.\n\n*Figure 6-13.* Figure showing side chains identified as major site 2 and site 3 epitopes in OSM (Deller et al., 2000). Panel (a) represents activity in gp130 binding assays and biological assays for each mutated side chain: red represents >75-fold increase in activity; blue represents <75-fold activity and grey represents no activity. Panel (b) shows relative solvent accessibilities for each mutated side chain ranging from 0 % (grey) through to 100 % (blue) calculated using the program NACCESS (Hubbard and Thornton, 1993). Figure produced using BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994).\n\n| Mutant | RSA (%) | Receptor binding assay: gp130 | Receptor binding assay: LIFR | Ba/F3 cell survival assay: LIFR/gp130 | Ba/F3 cell survival assay: OSMR/gp130 | Site location | Conclusion |\n|---|---|---|---|---|---|---|---|\n| Wildtype | \u2014 | 1.0 | 1.0 | 1.0 | 1.0 | \u2014 | \u2014 |\n| F160A | 65 | 1.1 | >100 | >10000 | >10000 | Site 3, N-terminus helix D | Active towards LIFR only |\n| K163A | 42 | 0.9 | >100 | >100 | >10000 | Site 3, N-terminus helix D | Active towards LIFR only |\n| F160A/K163A | \u2014 | 1.0 | >100 | >10000 | >10000 | Site 3, N-terminus helix D | Active towards LIFR only |\n| S7A | 100 | 1.1 | 0.9 | 0.8 | 0.5 | N-terminal loop | No activity |\n| K8A | 8 | 0.8 | 1.1 | 0.8 | 0.5 | N-terminal loop | No activity |\n| E9A | 96 | 1.2 | 0.8 | 1.4 | 1.2 | N-terminal loop | No activity |\n| Y10A | 35 | 3.4 | 7.2 | 2.1 | 0.9 | N-terminal loop | Limited activity towards LIFR and gp130 |\n| R11A | 74 | 1.0 | 1.9 | 0.8 | 0.5 | N-terminal loop | No activity |\n| V12A | 55 | 1.0 | 1.2 | 0.7 | 0.7 | N-terminal loop | No activity |\n| Q16A | 56 | 75.0 | 1.0 | 100.0 | 125.0 | Site 2, N-terminus helix A | Active towards gp130 only |\n| Q18A | 36 | 4.0 | 5.0 | 1.2 | 1.1 | Site 2, N-terminus helix A | Limited activity towards LIFR and gp130 |\n| K19A | 79 | 1.0 | 2.0 | 1.4 | 2.0 | Site 2, N-terminus helix A | No activity |\n| Q20A | 27 | 90.0 | 1.0 | 20.0 | 30.0 | Site 2, N-terminus helix A | Active towards gp130 only |\n| D22A | 40 | 1.0 | 1.0 | 0.8 | 1.0 | Site 2, N-terminus helix A | No activity |\n| L23A | 61 | 3.0 | 6.7 | 2.5 | 5.0 | Site 2, N-terminus helix A | Limited activity towards LIFR and gp130 |\n| D26A | 63 | 4.0 | 5.0 | 3.0 | 4.1 | Site 2, N-terminus helix A | Limited activity towards LIFR and gp130 |\n| M113A | 66 | 0.7 | 1.0 | 1.0 | 1.0 | Site 2, helix C | No activity |\n| P116A | 52 | 0.7 | 2.3 | 1.0 | 0.9 | Site 2, helix C | No activity |\n| N117A | 29 | 2.0 | 1.0 | 2.6 | 8.0 | Site 2, helix C | Limited activity towards gp130 only |\n| L119A | 46 | 1.7 | 2.2 | 1.2 | 1.7 | Site 2, helix C | No activity |\n| G120A | 76 | >100 | 1.8 | 60.0 | 60.0 | Site 2, helix C | Active towards gp130 only |\n| G120Y | 76 | >100 | 1.8 | >10000 | >10000 | Site 2, helix C | Active towards gp130 only |\n| N123A | 61 | 8.0 | 1.0 | 3.3 | 10.0 | Site 2, helix C | Active towards gp130 only |\n| N124A | 31 | >100 | 1.0 | >10000 | >10000 | Site 2, helix C | Active towards gp130 only |\n\n*Table 6-3.* Table showing summary of mutagenesis data of OSM. Mutagenesis studies carried out by K. Hudson, CRC Growth Factors Group, Birmingham. All mutants were created by PCR overlap using a pGEX OSM plasmid as detailed in Chapter 2. Relative solvent accessibilities (RSA) were calculated using NACCESS (Hubbard and Thornton, 1993) from the OSM coordinates. Receptor binding assays were carried out as detailed in Chapters 3 and 4 with values quoted representing mutant IC50/wildtype IC50 ratios. Ba/F3 cell survival assays were carried out as detailed in Chapter 3 with values quoted representing mutant EC50/wildtype EC50 ratios. Fields highlighted in bold in the original represent important results.\n\n#### 6.5.3 Proposed Model of Receptor-Ligand Interaction\n\nOn the basis of structural superposition of OSM and gp130 onto the Epo-EpoR structure as discussed above and shown in Figure 6-14 below it can be seen that all residues implicated in site 2 receptor binding through mutagenesis (Deller et al., 2000; Horsten et al., 1997) are indeed predicted to map to the interface region. Analysis of Table 6-3 shows that alanine mutation of Gly 120 has a major effect on gp130 binding with no effect on LIFR binding and limited effect on biological activity, however, mutation to tyrosine results in both major disruption of receptor binding and biological activity. Inspection of the proposed model in Figure 6-14 below, suggests that Gly 120 is central to the site 2 epitope and as such mutation to tyrosine would result in steric hindrance between Phe 169 and/or components of the B-C loop, whereas introduction of alanine would limit such a clash. Such a scenario has also been observed in GH in which mutation of the structurally equivalent Gly 120 results in a steric block on the formation of a 1:2 ligand-receptor complex (Clackson et al., 1998).", "# Approaches to automated protein crystal harvesting\nAuthors: Deller, M. C.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 70, 133\u2013155 (IYCr crystallization series), 2014\n\nThe UMR workcell was engineered by Square One System Design in Wyoming with design and development input from Marc Deller and Bernhard Rupp themselves, and was undergoing field testing and further development at the JCSG at The Scripps Research Institute. It is built around a St\u00e4ubli TX60 six-axis serial robot arm used in a three-step semi-automated process: selection of an appropriately sized end-effector microtool, presentation of the microtool close to the drop for user-guided harvesting, and automated cryoprotection, flash-cooling and loading into the storage dewar. It is deliberately flexible, loading crystals into either SSRL cassettes or UniPucks and accepting any standard SBS-footprint crystallization plate, and it uses novel LabVIEW machine-vision algorithms to locate crystals and assist user-guided harvesting. Supporting features include database tracking of every harvested crystal by plate, well, pin and cassette location, automated opening of the plate sealing tape, and the automated cryo-drip procedure. Because the robot works in a much more confined space than a hand, the sealing-tape puncture need only be about 3 mm rather than requiring removal of the whole tape, which extends the useful life of the crystallization drop roughly tenfold. Its two primary objectives are optimizing flash-cooling \u2014 through cryo-drip deposition of single drops of perfluoropolyether oil from a solenoid valve, enabling hyperquenching \u2014 and enabling routine high-throughput harvesting of microcrystals, made possible by the 1 \u00b5m step size of the robot arm, which allows harvesting of crystals under 10 \u00b5m.\n\n### 2.2.7. Capillary-top Mounting Micromanipulator (CMM)\n\nDeveloped at Hokkaido and Nagoya Universities, the CMM enables loopless mounting for optimal long-wavelength experiments. The rationale is that the sulfur anomalous signal used for S-SAD phasing (S K absorption edge at 5.02 \u00c5) increases at longer wavelengths, but so do scattering and absorption by the aqueous solvent and the microloop, degrading intensity accuracy enough to obscure the already weak signal. The solution is to harvest the crystal in a nylon microloop mounted on top of a glass micropipette whose opening sits at the centre of the loop, then aspirate the mother liquor away by negative pressure, leaving a liquid-free crystal on the pipette tip; the crystal is flash-cooled in the cryostream and the nylon loop removed. Modified Hampton CrystalCap bases support the method, but the manual process remains complex and time-consuming, so it was automated around a West Bond model 7200CR Dual Head Epoxy Die Bonder providing an optical microscope, X\u2013Y stage and a micromanipulator Z arm carrying the modified CrystalCap base. The operator moves the plate manually to harvest, after which automatic events fire 100 ms apart: the vacuum line opens to remove mother liquor, then a high-speed air-driven shutter actuates on the cryostream. The advantages are easier beam centring owing to a lack of lensing effects and improved flash-cooling from removal of excess aqueous solution.\n\n### 2.2.8. Acoustic Droplet Ejection (ADE)\n\nThe ADE system was developed by Labcyte and adapted for crystal harvesting on beamline X25 of the National Synchrotron Light Source at Brookhaven. It acoustically transfers 2.5 nl drops of microcrystal slurry onto micromesh mounts, which are then flash-cooled and analysed with a microfocus beam and raster scanning, using a modified Echo liquid-handling system coupled to a plate-handling robot and an X\u2013Y stage. The physics is that focused sound waves eject droplets from a liquid surface, with droplet volume proportional to the wavelength of the supplied sound energy and drop velocity proportional to its amplitude, so size, speed and ejection distance \u2014 and hence the density of microcrystals on the mesh \u2014 can be precisely controlled. ADE was used to mount microcrystal slurries that gave insulin and lysozyme structures at 1.9 and 1.8 \u00c5 respectively, with the overall high quality of the diffraction indicating that the acoustic energy did not damage the crystals. The authors highlight ADE as an enabling technology for fourth-generation microbeam and free-electron laser sources, which demand small crystals in large numbers, and describe the ADE conveyor-belt system in which slurry droplets are dispensed at high rate onto a Kapton belt attached to a standard goniometer base, successfully handling crystals of 20 to 200 \u00b5m and offering unprecedented speed of data collection from multiple crystals.\n\n### 2.2.9. Robotic Equipment for Automated Crystal Harvesting (REACH)", "# Cryo-EM and the New Frontier in Membrane Protein Drug Design\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* Cryo-EM advancements now enable the high-resolution structural determination of challenging membrane proteins like GPCRs, which were previously difficult to study.\n* Detailed visualization of receptor activation and ligand binding provides the precise information needed for structure-based drug design against elusive targets.\n* The integration of Cryo-EM into drug discovery pipelines is accelerating the development of novel therapeutics across a wide range of diseases.\n\n1. Recent trends in Structural Biology and Drug Discovery:\n   <https://3decision.discngine.com/blog/2025/1/10/recent-trends-in-structural-biology-insights-from-the-psdi-conference>\n2. The Disruptive Impact of Structural Biology on Biopharmaceutical Innovation:\n   <https://www.pharmasalmanac.com/articles/the-disruptive-impact-of-structural-biology-on-biopharmaceutical-innovation>\n\nFor decades, a vast and critical class of drug targets remained locked away in a black box. Membrane proteins\u2014the gatekeepers and signal transducers of our cells\u2014are implicated in countless diseases, yet their structures have been notoriously difficult to solve. Embedded in the greasy, fluid cell membrane, they resist the crystallization required for traditional X-ray crystallography, leaving drug designers to work with incomplete models or educated guesses. This has created a major roadblock in developing drugs for many cancers, neurological disorders, and metabolic diseases.\n\nRecent advances in cryo-electron microscopy (Cryo-EM) are shattering this barrier. By flash-freezing proteins in their near-native state and using powerful computation to reconstruct 3D images, Cryo-EM allows us to visualize these challenging molecules with stunning, near-atomic clarity. Breakthrough studies, such as the structural determination of the orphan G-protein coupled receptor (GPCR) GPR55, have revealed detailed ligand-binding modes and activation mechanisms for the first time. This is like finally getting the blueprints to a complex engine we\u2019ve been trying to fix for years.\n\nThe impact on drug discovery is immediate and transformative. Armed with high-resolution structures, scientists can now rationally design drugs that bind with high specificity and efficacy to these once-elusive targets. Enhancements in detector technology and image processing are allowing us to capture not just static pictures, but multiple conformational states, revealing the dynamic movements crucial for protein function and drug interaction. As Cryo-EM becomes faster and more accessible, it is rapidly being integrated into drug discovery pipelines, accelerating the development of a new generation of selective and effective therapeutics against the most challenging membrane proteins.", "# Revolutionizing Protein Structure Determination: How MICA Combines Cryo-EM and AlphaFold3 for Unprecedented Accuracy\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-20 on marcdeller.com (categories: sb).*\n\n* **Revolutionary Integration:** MICA combines cryo-EM experimental data with AlphaFold3 predictions through multimodal deep learning, achieving unprecedented accuracy in automated protein structure determination\n* **Outstanding Performance:** Achieved average TM-score of 0.93 on high-resolution maps, outperforming state-of-the-art methods ModelAngelo and EModelX by significant margins\n* **Robust Scalability:** Demonstrates consistent performance across protein sizes from 384 to 11,109 residues and map resolutions from 1.5-4 \u00c5\n* **Technical Innovation:** Uses Feature Pyramid Network architecture with hierarchical encoder-decoder design to capture multi-scale structural information\n* **Biological Hierarchy:** Implements cascaded prediction heads that reflect natural protein structure organization (backbone \u2192 C\u03b1 \u2192 amino acids)\n* **Automation Breakthrough:** Eliminates need for extensive manual intervention, potentially accelerating structure determination from months to days\n\n1. Gyawali, R., Dhakal, A., & Cheng, J. (2025). **Multimodal deep learning integration of cryo-EM and AlphaFold3 for high-accuracy protein structure determination.** bioRxiv:\n   <https://www.biorxiv.org/content/10.1101/2025.07.03.663071v1>\n2. Chen, S., et al. (2024). **Protein complex structure modeling by cross-modal alignment between cryo-EM maps and protein sequences.** Nature Communications, 15, 8808:\n   <https://www.nature.com/articles/s41467-024-53116-5>\n3. Zhang, Y. & Skolnick, J. (2004). **Scoring function for automated assessment of protein structure template quality.** Proteins, 57, 702\u2013710:\n   <https://zhanggroup.org/TM-score/>\n4. Jamali, K., et al. (2024). **Automated model building and protein identification in cryo-EM maps.** Nature, 628, 450\u2013457:\n   <https://www.nature.com/articles/s41586-024-07215-4>\n5. Oxford Protein Informatics Group. (2017). **TM-score \u2013 Template Modelling Score:**\n   <https://www.blopig.com/blog/2017/01/tm-score/>\n\nThe field of structural biology is witnessing a transformative moment as artificial intelligence meets experimental techniques in unprecedented ways. A groundbreaking study has introduced MICA (Multimodal Integration of Cryo-EM and AlphaFold3), a revolutionary deep learning approach that achieves remarkable accuracy in protein structure determination by combining the experimental power of cryo-electron microscopy (cryo-EM) with the predictive capabilities of AlphaFold3.\n\n## The Challenge of Automated Protein Structure Building\n\nCryo-EM has emerged as a premier technique for determining protein structures, particularly for large protein complexes that are difficult to crystallize. Despite producing high-resolution density maps, the process of building accurate atomic models from these maps remains notoriously challenging. Traditional methods require extensive manual intervention by expert crystallographers, often taking weeks or months to complete a single structure.\n\nThe core difficulty lies in accurately identifying individual atoms within density maps, tracing them to construct protein backbones, and correctly aligning amino acid sequences. Existing automated methods like ModelAngelo and EModelX have made significant progress but still fall short of achieving consistently high-accuracy models.\n\n## A Multimodal Revolution in Structural Biology\n\nMICA represents a paradigm shift by integrating cryo-EM density maps with AlphaFold3-predicted structures at both the input and output levels of the deep learning process. This multimodal approach leverages the strengths of both modalities: experimental density maps provide real structural data, while AlphaFold3 predictions offer complete structural information even for regions with poor or missing density.\n\nThe system employs a sophisticated encoder-decoder architecture with a Feature Pyramid Network (FPN) that simultaneously predicts backbone atoms, carbon-alpha (C\u03b1) atoms, and amino acid types. This hierarchical approach captures structural information at multiple scales, from local atomic details to global protein fold patterns.\n\n## Remarkable Performance Achievements\n\nThe results are nothing short of extraordinary. MICA achieved an average Template Modeling (TM) score of 0.93 on recently released high-resolution cryo-EM density maps, with TM-scores above 0.9 considered indicative of high-accuracy models. This represents a substantial improvement over existing methods:\n\n* 22.7% higher TM-score\u00a0compared to ModelAngelo (0.92 vs 0.75)\n* 3.4% higher TM-score\u00a0compared to EModelX(+AF) (0.92 vs 0.89)\n* Superior completeness\u00a0with longer aligned backbone lengths and better coverage\n\nPerhaps most impressively, MICA demonstrated robust performance across protein sizes ranging from 384 to 11,109 residues and maintained accuracy across varying map resolutions from 1.5 \u00c5 to 4 \u00c5.\n\n## Technical Innovation Behind the Success\n\nMICA\u2019s success stems from several key innovations. First, it processes cryo-EM maps and AlphaFold3 structures through separate enhancement pathways before fusing them into enriched feature representations. The cryo-EM pathway uses multi-scale convolutions with self-attention, while the AlphaFold3 pathway employs feature gates for selective information processing.\n\nThe progressive encoder stack with three hierarchical levels captures increasingly complex multimodal representations, which are then processed through the FPN for multi-scale feature aggregation. The three task-specific decoder heads work in cascade, with backbone predictions informing C\u03b1 predictions, and both informing amino acid classification\u2014reflecting the natural biological hierarchy.\n\n## Real-World Impact and Applications", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\n| Plate | Number of Colonies Observed |\n|---|---|\n| pUC transformed cells (positive control) | >2000 |\n| Non-transformed cells (negative control) | 0 |\n| pGEX-3C-\u0394185_OSM transformed cells | 3 |\n| pGEX-3C-\u0394188_OSM transformed cells | 12 |\n\n*Table 2-10.* Table showing the number of colonies obtained after transformation of E. coli JM109 with pGEX-3C-\u0394185_OSM and pGEX-3C-\u0394188_OSM. Negative and positive controls are also shown.\n\n### 2.3.5 Screening of Transformants\n\nFigure 2-6 below shows a 1 % agarose gel of pGEX-3C-\u0394185_OSM which has been digested with EcoRV and BglII. Lane 4 shows the five main marker bands of the \u03bb HindIII digest. Lanes 1, 2 and 3 each show two distinct bands at 3685 and 1815 base pairs.\n\n*Figure 2-6.* Figure showing 1 % agarose gel of EcoRV and BglII digestion of pGEX-3C-\u0394185_OSM. Lane 4 shows the five main marker bands of the \u03bb HindIII digest, running at 9146, 6882, 2322, 2027 and 564 (not visible in photograph) base pairs. Lanes 1, 2 and 3 contain triplicate digested samples and each shows two distinct bands at 3685 and 1815 base pairs.\n\n### 2.3.6 Automated PCR Sequencing\n\npGEX-3C-\u0394185_OSM and pGEX-3C-\u0394188_OSM sequences were verified against the published OSM sequence (GenBank accession number M27286; Malik et al., 1989). Figure 2-7 below shows the position of insertion into the pGEX-3C-GST plasmid and the four primers used for sequencing are marked. All sequences were correct with any discrepancies in the alignment explained by misreading or ambiguity in interpretation of the fluorescence signal.\n\n*Figure 2-7.* Figure showing sequence of pGEX-3C-OSM_\u0394188 construct in region around OSM insert. 'OSM start' and 'OSM end' denote OSM insert. Primers used are shown along with their reading direction. OSM sequence corresponds to that previously published (GenBank accession number M27286; Malik et al., 1989).\n\n## 2.4 Discussion of Engineering\n\nTruncation and amplification of both \u0394185_OSM and \u0394188_OSM only proved successful using Amplitaq (Perkin Elmer) polymerase. No amplification was observed (data not shown) using Vent Polymerase (New England Biolabs). Vent Polymerase contains 3'-5' exonuclease activity for proof reading of amplified DNA and for normal applications is used as the enzyme of choice. Vent Polymerase is known to work on a smaller subset of template and primer combinations, so suggesting a possible reason for the lack of amplification in this study. In an attempt to maintain the fidelity of amplification low dNTP and magnesium concentrations were used for all PCR reactions, as Amplitaq contains no 3'-5' exonuclease 'proof reading' activity. Lanes 2 and 3 of Figure 2-4 show amplified DNA running at ~700 base pairs corresponding to \u0394185_OSM PCR products. Comparison with lane 1 of Figure 2-4 shows that this band is not present in reactions containing no template DNA.\n\nThe amplified DNA was successfully digested with the restriction enzymes BamHI and EcoRI to remove the template DNA. Lane 3 and 4 of Figure 2-5 each show two distinct bands at ~560 base pairs and ~700 base pairs, corresponding to cut and uncut PCR products of \u0394185_OSM, respectively.\n\nThe purified insert was successfully ligated into the pGEX-3C-GST vector. The observed ligation efficiencies were very low (see Table 2-10) with typically only 3 and 12 transformant colonies for \u0394185_OSM and \u0394188_OSM respectively. Control plates transformed with pUC DNA contained over 2000 colonies, so confirming the competence of the cells used. Presumably, \u0394185_OSM constructs are toxic to cells and basal levels of expression result in the death of transformed cells. Possible explanations for this observation include aberrant protein folding resulting from incorrect prediction of the length of helix D. This is supported by the observation that \u0394188_OSM constructs appear less toxic, so suggesting a more suitable truncation point has been found in this mutant.\n\nAll colonies were screened for correct ligation using digestion with EcoRV and BglII. Lanes 1, 2 and 3 of Figure 2-6 shows two distinct bands at 3685 and 1815 base pairs, so confirming the presence of correctly ligated insert. All transformants showed correct ligation and subsequent PCR sequencing, using three primers (internal to insert, 3' to insert and 5' to insert), confirmed that both constructs showed agreement with published pGEX-3C-OSM-196 sequences.\n\n## References\n\n*(reference list omitted)*", "# 2025 \u2013 Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: papers).*\n\nStructures\n\nDiscovery of an Orally Bioavailable KRAS G12D Inhibitor\n\n* [Link to Paper](https://marcdeller.com/wp-content/uploads/2025/09/ye-et-al-2025-discovery-of-incb159020-an-orally-bioavailable-kras-g12d-inhibitor.pdf)\n\n* **Prevalence Impact:** KRAS G12D represents the most common KRAS mutation, affecting 29% of all KRAS-driven cancers with particularly high rates in pancreatic cancer\n* **Chemical Challenge:** The aspartate residue\u2019s reduced nucleophilicity compared to cysteine creates unique drug design obstacles for achieving oral bioavailability\n* **Design Strategy:** Structure-based drug design utilizing crystallographic insights enabled systematic optimization from weakly potent hits\n* **Molecular Properties:** Focus on molecular rigidity and balanced polarity proved essential for achieving both potency and oral exposure\n* **Validation Milestone:** Successful oral exposure in nonhuman primates demonstrates the feasibility of targeting this previously undruggable mutation\n* **Clinical Potential:** Multiple KRAS G12D inhibitors now in development offer new hope for patients with limited therapeutic options\n\n1. [Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor: Ye, Q. et al., Journal of Medicinal Chemistry.](https://pubs.acs.org/doi/abs/10.1021/acs.jmedchem.4c02662)\n2. [KRAS Mutations in Solid Tumors: Characteristics, Current Therapeutic Strategy: Chen, K. et al., Frontiers in Oncology](https://pmc.ncbi.nlm.nih.gov/articles/PMC9861148/)\n3. [Comprehensive pan-cancer genomic landscape of KRAS altered cancers: Hallin, J. et al., npj Precision Oncology](https://www.nature.com/articles/s41698-022-00334-z)\n4. [Flexibility versus Rigidity for Orally Bioavailable Cyclic Hexapeptides: Nielsen, D. S. et al., ChemBioChem](https://pubmed.ncbi.nlm.nih.gov/26336864/)\n5. [KRAS(G12D) can be targeted by potent inhibitors via formation of salt bridge: Su, T-H. et al., Cell Discovery](https://www.nature.com/articles/s41421-021-00368-w)\n6. Molecular Properties That Influence the Oral Bioavailability of Drug Candidates: Veber, D. F. et al., Journal of Medicinal Chemistry\n\nThe quest to target KRAS mutations has transformed from an impossible dream to clinical reality, yet the most prevalent mutation\u2014KRAS G12D\u2014remained frustratingly elusive. Recent groundbreaking research reveals how INCB159020 overcame seemingly insurmountable chemical and biological barriers to become the first orally bioavailable KRAS G12D inhibitor to achieve exposure in nonhuman primates. This achievement represents a watershed moment in precision oncology, potentially expanding therapeutic options for patients harboring the most common KRAS mutation across multiple cancer types.\n\nKRAS G12D mutations account for approximately 29% of all KRAS alterations and dominate the oncogenic landscape in pancreatic ductal adenocarcinoma, where they comprise over 40% of KRAS mutations. Unlike the cysteine residue in KRAS G12C mutations that enabled covalent inhibitor development, the aspartate residue in G12D presents unique challenges due to its reduced nucleophilicity compared to cysteine. This fundamental chemical difference necessitated entirely new approaches to achieve both sufficient potency and the ADME properties required for oral administration.\n\nThe discovery journey began with a weakly potent hit compound that required extensive structure-based optimization. Researchers at Incyte Corporation employed crystallographic insights to systematically improve binding affinity while simultaneously addressing the formidable challenge of achieving oral bioavailability. The team recognized that molecular rigidity\u2014a reduction in rotatable bonds\u2014could enhance membrane permeability and metabolic stability. This principle, validated across diverse compound classes, proved crucial for INCB159020\u2019s development.\n\nThe breakthrough came through meticulous attention to balanced polarity, a critical parameter for oral drug success. The optimization process focused on achieving sufficient target engagement while maintaining the physicochemical properties necessary for gastrointestinal absorption and systemic distribution. Unlike KRAS G12C inhibitors that benefit from covalent binding to enhance potency, KRAS G12D inhibitors must rely on optimized non-covalent interactions within the switch II pocket. This allosteric binding site, induced upon inhibitor engagement, creates a druggable cavity similar to that exploited by successful G12C therapeutics.\n\nINCB159020\u2019s achievement of oral exposure in nonhuman primates represents validation of years of medicinal chemistry effort. The compound demonstrates that the perceived \u201cundruggability\u201d of KRAS G12D was more accurately a reflection of the technical challenges rather than fundamental impossibility. This success opens new avenues for treating patients with KRAS G12D-driven cancers, particularly pancreatic adenocarcinoma, where this mutation predominates and current therapeutic options remain limited.\n\nThe broader implications extend beyond INCB159020 itself, as the successful design principles established through this work provide a roadmap for developing additional KRAS G12D inhibitors. The emphasis on molecular rigidity, balanced polarity, and structure-guided optimization offers valuable insights for medicinal chemists tackling similar challenges in targeting other \u201cundruggable\u201d proteins. As multiple KRAS G12D inhibitors now progress through clinical development, including competing compounds from various pharmaceutical companies, patients with this common mutation may soon have access to targeted therapies that were unimaginable just a few years ago.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\n| Group | Structural Class | Cytokines | Receptor Class | Biological Action | Cell Type |\n|---|---|---|---|---|---|\n| Group 1: 4-\u03b1-Helical Bundle Cytokines | Short-Chain | IL-2, IL-4, IL-7, IL-9, IL-13; IL-3, GM-CSF, IL-5; M-CSF, CSF; IFN\u03b3 | Haemopoietin Domain; Ig-TK III; IFN-R | Proliferation/Differentiation of haemopoietic cells; Pleiotropic; Mono-specific; Immunity | T-cell, B-cell, Mast cell; Neutrophil, Eosinophil; Stem cell, Fat, Nerve; Many, M\u03c6 |\n| Group 1: 4-\u03b1-Helical Bundle Cytokines | Long-Chain | IL-6, LIF, OSM, IL-12\u03b1, CNTF, IL-11, CT-1; Epo, G-CSF; GH, PRL; IL-10, IFN\u03b1, IFN\u03c4 | Haemopoietin Domain; IFN-R | Pleiotropic; Mono-specific; Growth; Immunity | Liver, Neutrophil, T-cell, M\u03c6, Muscle; Erythroid, Neutrophil; Many; Many |\n| Group 2: \u03b2-Sheet Rich Cytokines | Cysteine Knot | TGF\u03b2, Activin, Inhibin, BMP; PDGF-A, PDGF-B, VEGF, PlGF; NGF, BDNF, NT-3, NT-4/5 | S/T Kinase I/II; Ig-TK III; Cys-TK V, P75 | Differentiation, Growth inhibition, Tissue modeling; Growth; Survival, Growth, Differentiation | Many; Epithelial, Endothelial; Nerve cells, Neuroectoderm |\n| Group 2: \u03b2-Sheet Rich Cytokines | \u03b2-Jelly Roll | TNF\u03b1, TNF\u03b2, CD40L, CD27L, FASL | TNF-R (P55) | Immunity, Cell death | Many |\n| Group 2: \u03b2-Sheet Rich Cytokines | \u03b2-Trefoil | IL-1\u03b1, IL-1\u03b2; FGF, INT-2, KGF | Ig-like; Ig-TK IV, Heparin | Immunity; Growth | Many; Mesenchymal cells |\n| Group 3: \u03b1/\u03b2 Cytokines | S-S Rich \u03b2-Meander | EGF, TGF\u03b1, \u03b2-cellulin, SCDGF, VVGF, Amphiregulin, HB-EGF | Cys-Rich TK I | Proliferation, Wound healing | Epithelial cells |\n| Group 3: \u03b1/\u03b2 Cytokines | S-S Rich (insulin-like) | Insulin, IGF-I, IGF-II, Relaxin, Bombyxin | Cys-Rich TK II; M6P-Receptor | Metabolism, Proliferation/Differentiation | Mesenchymal cells |\n| Group 3: \u03b1/\u03b2 Cytokines | Chemokines (CXC, CC, C, CXXXC) | CXC: IL-8, GRO, PF-4, IP-10, MIP-2, NAP-2, GCP-2; CC: MCP-1, MCP-2, MCP-3, RANTES, MIP-1\u03b1, MIP-1\u03b2, I309; C: Lymphotactin; CXXXC: Fractalkine | Serpentine seven pass | Chemotaxis, Immunity | Neutrophil, Macrophage, Eosinophil |\n| Group 4: Mosaic Cytokines | Ig-EGF-TM-Cyt | HRG\u03b1/\u03b2, NDF, ARIA | Cys-TK I (neu/erbB-2) | Growth, Differentiation | Nerve cells, Epithelial cells |\n| Group 4: Mosaic Cytokines | Kringle-Ser Protease | HGF | TK VI (Met) | Growth | Hepatocytes, Endothelial cells, Epithelial cells |\n| Group 4: Mosaic Cytokines | \u03b1-helix-haemopoietin | IL-12 | ? | Immuno-modulation | T-cells, NK-cells |\n\n*Table 1-2.* Table showing the four structural classes of cytokines according to secondary structure (Richardson, 1981). Subgroupings within each group are shown according to structural motifs, receptor classes utilised and biological actions. Adapted from Nicola, 1994. See abbreviations for full description.\n\n*Figure 1-1.* Gallery of cytokine folds arranged according to the four structural classes of secondary structure (Richardson, 1981). Subgroupings within each group are shown according to structural motifs as highlighted in red lettering. Adapted from Nicola, 1994. All structures solved by X-ray crystallography (unless otherwise stated). Figures produced using BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994). The polypeptides are coloured from blue at the N-terminus, through red at the C-terminus. See abbreviations for full descriptions.\n\n#### 1.4.1.1 Group 1 Cytokines (\u03b1)\n\nThis is the largest family of cytokines and these structures adopt a 4-\u03b1-helical bundle as first reported for GH (Abdel Meguid et al., 1987) (see Figure 1-2). Members of this family have attracted much attention due to their wide-ranging roles in the immune system. This fold has an unusual 'up-up-down-down' connectivity in which two long over-hand loops (connecting helices A and B and helices C and D) allow for helix direction reversal and for packing of the A:D helix pair against the B:C helix pair with a characteristic skew angle. This unusual topology has so far not been observed in any other family of proteins outside the cytokine realm. This group of cytokines can be further sub-divided into short-chain or long-chain helical cytokines, dependent on the length of the main helical segments (see Figure 1-2).\n\n##### 1.4.1.1.1 Short chain helical cytokines\n\nShort-chain 4-\u03b1-helical cytokines include IL-2, IL-3, IL-4, IL-5, IL-7, IL-9, IL-13, GM-CSF, M-CSF, SCF and IFN\u03b3 (Figure 1-2). Short-chain cytokines have relatively short \u03b1-helices of ~15 amino acids in length and have a relatively large skew angle (~35\u00b0) between the helical axis of the A:D pair and the B:C pair. Short chain cytokines often have two short twisted anti parallel \u03b2-strands in the A-B and C-D loops as exemplified by GM-CSF. Other short chain helical cytokines analysed structurally include IL-5 (Milburn et al., 1993), M-CSF (Pandit et al., 1992), IFN-\u03b3 (Ealick et al., 1991; Samudzi et al., 1991) and IL-3 (Feng et al., 1996) which are dimeric, and GM-CSF (Diederichs et al., 1991; Walter et al., 1992a), IL-2 (Bazan, 1992) and IL-4 (Powers et al., 1993; Smith et al., 1994; Walter et al., 1992b; Wlodaver et al., 1992) which are monomeric.\n\n##### 1.4.1.1.2 Long chain helical cytokines\n\nLong-chain cytokines include IL-6, IL-12, Epo, G-CSF, LIF, OSM, CT-1, CNTF, IL-11, leptin, GH, Prl, IFN\u03b1, IFN\u03b2, IFN\u03c4 and IL-10 (see Table 1-2). This class of helical cytokines have relatively long \u03b1-helices of ~25 amino acids in length and a relatively small skew angle (~18\u00b0) between the helical axis of the A:D pair and the B:C pair. Long chain helical cytokines resolved structurally include GH (Abdel Meguid et al., 1987; de Vos et al., 1992), Epo (Livnah et al., 1998; Livnah et al., 1996), Leptin (Zhang et al., 1997a), GCSF (Hill et al., 1993; Lovejoy et al., 1993), IL-10 (Zdanov et al., 1995), IFN-\u03b2 (Senda et al., 1995) and IFN-\u03c4 (Radhakrishnan et al., 1999).", "# Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\nAuthors: Zhuo, J.; Li, Y.-L.; Qian, D.-Q.; Burns, D.M.; Mei, S.; Rafalski, M.; Xu, M.; Cao, G.; Pan, Y.; Jia, Z.; Jalluri, R.; Epling, L.B.; Fenalti, G.; Deller, M.C.; Procak, J.; Covington, M.; He, X.; Collins, R.; Stubbs, M.; Burke, K.; Oliver, J.; Margulis, A.; Boer, J.; Wynn, R.; Scherle, P.; Diamond, S.; Newton, R.; Zhang, Y.; Metcalf, B.; Yao, W.\nSource: Journal of Medicinal Chemistry (Drug Annotation), 2026\n\n| Compound | 9 | 14 | 15 | 16 | 17 | 18 | 19 |\n|---|---|---|---|---|---|---|---|\n| X | H | H | F | F | F | F | H |\n| Y | H | H | H | H | F | F | F |\n| JAK1 IC50 (nM) | 1.5 | 0.8 | 1.7 | 0.56 | 0.46 | 0.25 | 0.25 |\n| JAK2 IC50 (nM) | 31 | 24 | 19 | 11 | 5.5 | 6.1 | 7.8 |\n| JAK2/JAK1 | 21 | 30 | 11 | 19 | 12 | 25 | 25 |\n| INA6 IC50 (nM) | 308 | 150 | 192 | 126 | 60 | 70 | 83 |\n| WB IL-6 IC50 (nM) | 310 | 374 | 258 | 175 | 135 | 207 | 243 |\n| WB TPO IC50 (nM) | 3400 | 3182 | 860 | 782 | 417 | 1167 | 1161 |\n| WB IL-6/TPO | 11 | 9 | 3 | 5 | 3 | 6 | 5 |\n| Caco-2 (x10\u207b\u2076 cm/s) | 0.9 | 1.1 | 3.1 | 4.7 | 4.6 | 8.5 | 1.4 |\n| Proj. h-CL (L/h/kg) (%F) | 0.7 (42) | 0.6 (53) | 0.8 (34) | 0.6 (49) | 0.8 (35) | 0.7 (42) | <0.5 (>60) |\n| RAT IV dose (mg/kg) | 2.5 | 2 | 2.5 | 2.5 | 2 | 2 | 2 |\n| RAT IV Clr %HBF | 29 | 31 | 15 | 6 | 8 | 11 | 12 |\n| RAT IV t\u00bd (h) | 5.3 | 1.7 | 3.7 | 3.5 | 2.4 | 3.7 | 2.2 |\n| RAT PO dose (mg/kg) | 5 | 4 | 5 | 5 | 4 | 5 | 5 |\n| RAT PO AUC (nM*h) | 3,793 | 2,501 | 14,954 | 28,766 | 28,559 | 11,884 | 5,606 |\n| RAT PO t\u00bd (h) | 5.7 | 4.8 | 4.2 | 4.3 | 2.5 | 4.6 | 2.2 |\n| RAT PO %F | 31 | 31 | 61 | 60 | 70 | 54 | 21 |\n| Cyno IV dose (mg/kg) | 2 | | 2 | 2.5 (dose adj.) | | 2.5 | 2.5 |\n| Cyno IV Clr %HBF | 25 | | 21 | 2 | | 3 | 7 |\n| Cyno IV t\u00bd (h) | 1.9 | | 2.1 | 10 | | 8.4 | 4.3 |\n| Cyno PO dose (mg/kg) | 4 | | 4 | 5 (dose adj.) | 4 | 5 | 5 |\n| Cyno PO AUC (nM*h) | 1,549 | | 2,169 | 80,964 | 5,275 | 74,984 | 14,445 |\n| Cyno PO t\u00bd (h) | | | 3.9 | 7.0 | 4.0 | 16 | 6.0 |\n| Cyno PO %F | 9 | | 11 | 38 | | 54 | 28 |\n\n*ADME, absorption, distribution, metabolism, and excretion; AUC, area under the curve; Caco-2, cell line of human colorectal adenocarcinoma cells; CL, clearance; Clr, clearance; cyno, cynomolgus; F, fraction absorbed; HBF, hepatic blood flow; IC50, half-maximal inhibitory concentration; IV, intravenous; JAK, Janus kinase; PO, oral; proj h-CL, projected hepatic clearance; t\u00bd, elimination half-life; TPO, thrombopoietin; WB, whole blood.*\n\n### Table 5. Investigation of Pyrazole Hinge-Binding Motif\n\n| Hinge Binder / Compound | 20 (pyrazole) | 21 (methylpyrazole) | 22 (3,5-dimethylpyrazole) | 18 (7H-pyrrolo[2,3-d]pyrimidine) |\n|---|---|---|---|---|\n| JAK1 IC50 (nM) | 314 | 19 | 11 | 0.25 |\n| JAK2 IC50 (nM) | 7319 | 910 | 505 | 6.1 |\n| Selectivity JAK2/JAK1 | 23 | 48 | 46 | 25 |\n\n*IC50, half-maximal inhibitory concentration; JAK, Janus kinase.*\n\n### Table 6. Comparison of 22 vs 18\n\n| | Parameter | 22 | 18 |\n|---|---|---|---|\n| Potency (nM) & (Selectivity) | JAK1 IC50 | 11 | 0.25 |\n| | JAK2 IC50 | 505 | 6.1 |\n| | Selectivity JAK2/JAK1 | 46 | 25 |\n| | INA6 IC50 | 97 | 70 |\n| | WB IL-6 IC50 | 405 | 207 |\n| | WB TPO IC50 | 6389 | 1167 |\n| | Selectivity WB TPO/IL-6 | 16 | 6 |\n| ADME | Caco-2 (x10\u207b\u2076 cm/s) | 4.5 | 8.5 |\n| | Proj. h-CL (L/h/kg) (%F) | 0.7 (48) | 0.7 (42) |\n| RAT IV | Dose (mg/kg) | 2.5 | 2 |\n| | Clr %HBF | 8 | 11 |\n| | t\u00bd (h) | 3.7 | 3.7 |\n| RAT PO | Dose (mg/kg) | 5 | 5 |\n| | AUC (nM*h) | 45,210 | 11,884 |\n| | t\u00bd (h) | 4.2 | 4.6 |\n| | %F | >100 | 54 |\n| Cyno IV | Dose (mg/kg) | 2.5 | 2.5 |\n| | Clr %HBF | 4 | 3 |\n| | t\u00bd (h) | 13 | 8.4 |\n| Cyno PO | Dose (mg/kg) | 5 | 5 |\n| | AUC (nM*h) | 137,199 | 75,984 |\n| | t\u00bd (h) | NC | 16 |\n| | %F | 100 | 54 |\n| Dog IV | Dose (mg/kg) | 2.5 | 2.0 |\n| | Clr %HBF | 4 | 82 |\n| | t\u00bd (h) | 22 | 1.1 |\n| Dog PO | Dose (mg/kg) | 5 | 4 |\n| | AUC (nM*h) | 165,386 | 65 |\n| | t\u00bd (h) | 23 | 0.9 |\n| | %F | 100 | 1 |\n\n*ADME, absorption, distribution, metabolism, and excretion; AUC, area under the curve; Caco-2, cell line of human colorectal adenocarcinoma cells; Clr, clearance; cyno, cynomolgus; F, fraction absorbed; HBF, hepatic blood flow; IV, intravenous; JAK, Janus kinase; PO, oral; proj h-CL, projected hepatic clearance; t\u00bd, elimination half-life; TPO, thrombopoietin; WB, whole blood.*\n\n### Table 7. In Vitro Potency of 22 in Cytokine- and Growth Factor-Stimulated Cells\n\n| Assay Readout | Cell Type | Stimulus | JAK dimer | IC50 (nM) | n |\n|---|---|---|---|---|---|\n| **Signaling assays** | | | | | |\n| pJAK2/pSTAT3/pSTAT5 | h-T-cells | IL-2 | JAK1/3 | ~10\u201330 | 2 |\n| pSTAT3 | h-WB | IL-6 | JAK1/2 | 690 \u00b1 224 | 32 |\n| pSTAT3 | h-WB | TPO | JAK2/2 | >10,000 | 32 |\n| pSTAT1 | h-WB | IL-6 | JAK1/2 | 360 \u00b1 128 | 8 |\n| pSTAT5 | h-WB | GM-CSF | JAK2/2 | >10,000 | 8 |\n| **Functional assays** | | | | | |\n| Proliferation | h-T-cells | IL-2 | | 29.6 \u00b1 1.2 | 12 |\n| MCP-1 production | h-PBMC | IL-6 | | 178 \u00b1 57 | 3 |\n\n*Data are presented as mean \u00b1 SD. Mean values are presented when n \u2265 3. GM-CSF, granulocyte-macrophage colony-stimulating factor; IC50, half-maximal inhibitory concentration; JAK, Janus kinase; PBMC, peripheral blood mononuclear cells; TPO, thrombopoietin. h = human; h-WB = human whole blood; h-PBMC = human peripheral blood mononuclear cells.*\n\n## Chemistry\n\nIn 2009, Rodgers and Shepard outlined the synthesis of the key SEM-protected azetidine 30, beginning with 4-chloro-7H-pyrrolo[2,3-*d*]pyrimidine 23 (Scheme 1). The key steps were Suzuki coupling followed by deprotection and subsequent Michael addition with pyrazole and Boc-protected 3-(cyanomethylene)azetidine 28. Following acid-promoted Boc removal, azetidine 30 was subjected to Buchwald\u2013Hartwig amination with the appropriate aryl halide as either the amide, as shown in Scheme 1, or as the ester that could be subsequently hydrolyzed and coupled with the appropriate amine using an amide coupling agent, such as HATU. Finally, the SEM-protecting group could be removed using a two-step in situ procedure starting with either BF3 or trifluoroacetic acid (TFA) treatment, followed by azeotropic removal of the volatiles and reaction with ammonium hydroxide to cleave the remaining methylene hydroxy group.", "# GitHub README: TopPDBLX\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/TopPDBLX, 2026\n\n| Stage | Reagents | Names | Identified |\n|---|---|---|---|\n| Seeded from corpus mining | 147 | 501 | 79.1% |\n| Round 1: 40 frequency-ranked decisions | 165 | 570 | 80.1% |\n| Round 2: 10 grouped decisions, 1,004 names | 502 | 1,265 | 84.5% |\n| Prose stripping, and separating apparatus notes/bare units from reagents (parser fixes) | 502 | 1,265 | 85.2% (87.5% on chemistry alone) |\n| Ionic liquids added (v0.4.0); aliases naming a different molecule removed (v0.5.x) | 542 | 1,306 | 85.3% |\n| Two gold sets sampled and labelled; isomers the teacher's false positives exposed (v0.6.x) | 562 | 1,370 | 85.3% |\n| De-duplication: PEG MME's seven ids per weight merged to one, all remaining duplicate ids merged, method text and element abbreviations stripped out (v0.7.0\u2013v0.8.7) | 486 | 1,487 | 85.7% (88.3% on chemistry) |\n| v0.9.0: the Morpheus stock table, 15 premixes transcribed from the vendor brochure | 499 | 1,543 | 85.9% (88.4% on chemistry) |\n| v0.9.2: every pre-merge canonical id resolves again | 499 | 1,583 | 86.5% (89.0% on chemistry) |\n| v0.9.3 (current): ACES added, found by probing what the parsers still cannot name | **500** | **1,587** | **87.2%** |\n\n**For the crystallographer:** every reagent carries the chemistry the ontology needs, and each field is enforced on load rather than being optional documentation. A `peg` entry must state its molecular weight, a `buffer` must state its pKa, and a `premix` must list its constituents. Those invariants caught three separate attempts to bulk-add entries that could not satisfy them.\n\n| Chemical class | Entries | What it carries |\n|---|---|---|\n| additive | 167 | ligands, cofactors and inhibitors carried into the drop |\n| salt | 116 | Hofmeister rank, spanning sulfate (-4) to thiocyanate (+4) |\n| peg | 91 | molecular weight, log-scaled; v/v below 600 and w/v above |\n| buffer | 36 | pKa, since buffer identity collapses to the pH it sets |\n| organic | 33 | alcohols and small organics reported v/v |\n| polyol | 28 | glycerol, sugars and sugar alcohols |\n| detergent | 16 | CYMAL-6, DDM, LDAO, beta-octyl glucoside: reported w/v, and rarely the precipitant |\n| premix | 9 | constituent ids for Tacsimate, Morpheus and similar |\n| other | 6 | vendor mixtures with no single molecular weight, such as PEG Smear Broad |\n\nTwo rounds of curation and one parser fix took identification from 79.1% to 85.2%. The marginal return has since fallen to roughly 15 components per decision, so further curation rounds are no longer the best use of expert time; the remaining gains have come from the residual parser rather than the dictionary.\n\n## \ud83c\udf93 Training\n\n**In plain terms:** the rule-based parser reads most of the archive, but some depositions are written in prose it cannot follow. A small language model was trained to read those, by learning from the hundreds of thousands of examples the rule parser already handles correctly. It costs nothing in hand-labelling, because the rule parser writes its own teaching material.\n\n**For the crystallographer:** SmolLM2-360M under MLX-LM LoRA on an M1 Max, prompt masked so the loss falls on the answer rather than on echoing the question, evaluated on the residual the rule parser could not read.\n\n- **Validation loss is the wrong stopping signal.** The labels are the rule parser's own output, so validation loss measures fidelity to the teacher rather than skill on the residual: it flattens early while identification is still climbing. Checkpoints are chosen by sweeping downstream accuracy on the residual instead.\n- **Longer training clearly helps recovery, and that is part of the improvement story.** On the 2,000-record frozen benchmark, round 09's identification rises 5.5 points and fully-identified records rise 15 points between true iteration 1,000 and true iteration 8,000, and the number of distinct names the model can't identify more than halves. Round 09 is a better recipe *and* a longer one than round 06 (8,000 iterations, ~1.04 epochs, against round 06's 6,000, ~0.94 epochs).\n- **Small evaluation sets can hide a real difference.** The 192-record gold sets showed almost no gap between round 09's early and late checkpoints; the 2,000-record frozen benchmark resolved the same comparison easily. Where a metric is reported below, its sample size is given for exactly this reason.\n- **Gates are judged on the lower confidence bound**, so a lucky sample cannot pass them.\n\nTraining data is deduplicated before oversampling: see [Redundancy](#-redundancy) for why that is not a detail. Every run is named `r1-parse-residual-smollm2-360m-roundNN` and logged to Weights & Biases, with the adapter directory carrying the same name so a checkpoint on disk traces back to the run that produced it.\n\n## \ud83d\udcd0 How the model is measured\n\n**In plain terms:** it is easy to build a metric that a bad model passes. If you only ask \"did it output a real chemical name\", a model that answers *sodium chloride* to every question scores well, because sodium chloride is a real chemical. The measures below were chosen so that answering confidently and wrongly is penalised, not rewarded.\n\n| Measure | Asks | Blind to |\n|---|---|---|\n| **Schema validity** | Is the output parseable JSON using only the allowed roles and units? | Whether any of it is true |\n| **Identification** | Does each named reagent exist in the curated lexicon? | Whether it is the *right* reagent for this text |\n| **Grounding** | Is each named reagent actually mentioned in the text it was given? | Whether the amount and unit are right |\n| **Fidelity** | On text the rules *can* read, does the model produce the same components? | The residual, which is by definition harder |\n\nNone of those four has any ground truth behind it: they ask whether an answer is *defensible*, not whether it is *right*. Precision and recall, defined below, are computed against hand-labelled truth and are what decide anything.", "# Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands\nAuthors: Pozharski, E.; Deller, M. C.; Rupp, B.\nSource: Protein Crystallography: Methods and Protocols (A. Wlodawer, Z. Dauter & M. Jaskolski, eds.), Methods in Molecular Biology vol. 1607, Chapter 25, pp. 611\u2013625, Springer, 2017\n\nBasic empirical epistemology, the authors write, requires a model to be examined against the primary evidence and also weighted by its compliance with independently acquired prior knowledge; a well-refined ligand built into unambiguous density will generally show reasonable stereochemistry and plausible interactions. Prior expectations enter refinement as stereochemical restraints, which act effectively as additional data points stabilizing an otherwise underdetermined or only weakly over-determined macromolecular refinement. The ligand's stereochemistry is described in a restraint file containing bond lengths, bond angles, planarity restraints and preferred torsions. For small molecules and new chemical entities the authors stress that particular care is needed to describe the specific chemistry correctly, including delocalization, tautomerization and charge state, because if the restraints are wrong the crystallographic data may not contribute strongly enough to the refinement target function to guide correct placement, pose and stereochemistry \u2014 and the resulting deviation from the density may not show up in standard real-space metrics. Peptide ligands are a special and useful case: their backbone torsion angles are not restrained, so they provide an excellent geometric real-space cross-validation. Peptide models lacking good supporting density will almost always refine to implausible high-energy conformations, readily detectable as Ramachandran outliers, whose percentile is available from PDB validation reports. A further check is to compare ligand B-factors with those of nearby protein atoms in the binding site: interacting parts should have similar B-factors, and significant differences may indicate partial occupancy or an incorrectly built ligand.\n\n### 3.2. Clash scores and binding site chemistry\n\nA reasonably placed ligand will make clearly defined interactions with amino-acid residues of the target protein. The two common failure modes named are steric clashes between a poorly placed ligand and the protein, and the absence of any sensible contacts between ligand and target. All-atom close-contact analysis \u2014 the clash score \u2014 can be computed with MolProbity and visualized in COOT, which lists bad atom contacts and displays them as dots and spikes. The chapter's worked example is an implausible Fab\u2013antigen complex, PDB entry 2a6i chain P, whose peptide antigen model should have been surrounded by clear positive omit difference density in the distinct shape of the peptide, but instead showed only discontinuous fragments, part of which is explicable as ordered solvent (round green spheres typical of water molecules) in an mFo \u2212 DFc omit map contoured at 2.5\u03c3. Beyond the poor density fit, the antigen model has a multitude of unreasonably close contacts with the Fab, visualized as red spikes, and an implausibly high-energy backbone conformation with torsion angles in unfavourable regions of the Ramachandran plot. The combination of poor agreement with the primary evidence and violation of established prior expectations leads the authors to conclude that the ligand as modelled is unlikely to be present. Some programs, notably VHELIBS, evaluate binding-site interaction quality directly and rank ligands on those scores.\n\n### 4. Practice of ligand evaluation\n\nHow much validation is required is explicitly framed as context-dependent. A protein\u2013ligand structure destined for structure-based lead discovery demands a higher level of scrutiny and generally a resolution of around 2 \u00c5 or better, whereas less scrutiny is needed when a complex is used only to analyse global features of a binding site for purposes such as structure-guided mutagenesis. The chapter then provides a decision tree (Figure 4) whose branches are validation of the ligand electron density, validation of the ligand stereochemistry against prior expectations, and validation of the protein and ligand environment. The recommended data and program setup runs: ask whether the chosen structure provides enough detail, particularly resolution, for the intended purpose; download the full PDB validation report; download coordinates and electron density from EDS or PDB_REDO into COOT, noting that if the electron density is unavailable then full validation is simply not possible; turn on symmetry-related atoms; turn on environment distances; then select the ligand and zoom in, proceeding either to omit-map preparation or to inspection of the 2mFo \u2212 DFc map.\n\n### 4.2. Preparing an omit map\n\nA step-by-step protocol is given. Download the coordinates from the PDB, download the reflection data and convert it to MTZ format using the CCP4i structure-factor utilities, or alternatively download the MTZ file corresponding to the fully optimized structure from PDB_REDO; phenix.cif_as_mtz from the PHENIX package is named as another conversion route. Delete the ligand from the model, either by manually editing the PDB file or in COOT, remembering that ligand atom records are identified by HETATM plus the three-letter ligand identifier, that any LINK or CISPEP records for the ligand in the header must also be removed, and that ATOM lines belonging to the protein and the CRYST1 line must be retained. Run a single cycle of refinement using the ligand-omitted model as the starting structure, ideally with several different refinement programs since results may differ; REFMAC, phenix.refine and BUSTER-TNT are named as widely used, with graphical interfaces available for the first two. The output MTZ contains both the 2mFo \u2212 DFc omit map, conventionally displayed in blue, and the mFo \u2212 DFc difference omit map with green positive and red negative density, both of which can be inspected in COOT.\n\n### 4.3. Electron density inspection", "# 2012 \u2013 Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-09-07 on marcdeller.com (categories: papers).*\n\n* [Link to Paper](https://marcdeller.com/wp-content/uploads/2025/09/kong-et-al-2012-structure-of-hepatitis-c-virus-envelope-glycoprotein-e2-antigenic-site-412-to-423-in-complex-with.pdf)\n\n* **Dual Recognition Modes**: Two distinct antibody binding approaches target the same conserved E2 epitope with different angles of approach\n* **Molecular Precision**: Critical amino acid interactions at positions 413, 415, 418, and 420 define the neutralization interface\n* **Vaccine Innovation**: Epitope-focused nanoparticle immunogens successfully induce broadly neutralizing antibodies across HCV genotypes\n* **Structural Diversity**: \u03b2-hairpin conformational flexibility allows multiple antibody solutions for viral neutralization\n* **Glycan Engineering**: Modified E2 glycoforms enhance immunogenicity while maintaining native-like antigenic properties\n* **Therapeutic Integration**: Structural insights enable both preventive vaccine design and therapeutic antibody development strategies\n\n1. **Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33:** [Kong, L., et al., Journal of Virology, 2012](https://journals.asm.org/doi/10.1128/jvi.01939-12)\n2. **Hepatitis C Virus E1E2 Structure, Diversity, and Implications for Vaccine Development:** [Bailey, J.R., et al., Vaccines, 2024](https://pmc.ncbi.nlm.nih.gov/articles/PMC11125608/)\n3. **Epitome-focused immunogens targeting the hepatitis C virus envelope glycoprotein induce broadly neutralizing antibodies:** [Krey, T., et al., Science Advances, 2024](https://www.science.org/doi/10.1126/sciadv.ado2600)\n4. **The Structural Basis of Hepatitis C Virus Neutralization by AP33:** [Stanfield, R.L., et al., Journal of Virology, 2012](https://pmc.ncbi.nlm.nih.gov/articles/PMC3497650/)\n5. **Structure-Based Design of Hepatitis C Virus E2 Glycoprotein Immunogens:** [Bailey, J.R., et al., Journal of Virology, 2020](https://journals.asm.org/doi/10.1128/jvi.00704-20)\n6. **Non-cognate ligands of hepatitis C virus envelope broadly neutralizing antibodies induce virus-neutralizing sera:** [Walimbwa, S.I., et al., Frontiers in Immunology, 2025](https://pubmed.ncbi.nlm.nih.gov/40766312/)\n\n**Unlocking the Molecular Architecture of Hepatitis C Virus: Structural Insights for Next-Generation Vaccine Design**\n\nThe structural characterization of hepatitis C virus (HCV) envelope glycoproteins represents one of the most significant advances in understanding viral neutralization mechanisms and vaccine development pathways. The groundbreaking crystal structure determination of the broadly neutralizing antibody AP33 in complex with the HCV E2 envelope glycoprotein antigenic site 412-423 has revealed unprecedented molecular details that directly inform rational vaccine design strategies. This structural revelation demonstrates how different antibodies can adopt distinct binding approaches to neutralize the same conserved viral epitope, opening new avenues for therapeutic intervention against this persistent global health challenge.\n\nThe AP33 antibody-E2 peptide complex adopts a distinctive \u03b2-hairpin conformation that positions itself strategically within the deep binding pocket of the antibody paratope. The epitope spans a critical region of the E2 glycoprotein from residues 412 to 423, forming a type I\u2032 \u03b2-turn that presents both hydrophobic and hydrophilic surfaces for antibody recognition. Comparative analysis with the previously characterized HCV1 antibody reveals a remarkable 22-degree difference in the angle of approach between these two broadly neutralizing antibodies, despite targeting the identical epitope sequence. This structural diversity illustrates the multiple solutions that the immune system can employ to recognize and neutralize conserved viral determinants.\n\nThe molecular interactions governing AP33 recognition involve critical contacts with leucine 413, asparagine 415, glycine 418, and tryptophan 420 of the E2 glycoprotein. These residues form an intricate network of hydrogen bonds and hydrophobic interactions that stabilize the antibody-antigen complex, with AP33 burying 515 square angstroms of surface area across the epitope interface. The structural analysis reveals why certain viral escape mutations, particularly N415Y, N415D, and G418D, can disrupt antibody binding while maintaining viral fitness. The asparagine 415 residue forms dual hydrogen bonds with both the main chain and side chain components of the AP33 paratope, explaining the antibody\u2019s exquisite sensitivity to mutations at this position.\n\nRecent advances in HCV vaccine development have leveraged these structural insights to engineer epitope-focused immunogens that specifically target conserved neutralizing determinants. Novel nanoparticle-based vaccine platforms displaying the AP33 epitope have demonstrated the ability to induce broadly neutralizing antibodies in laboratory models, representing a significant breakthrough in HCV vaccine design. These immunogens successfully triggered cross-reactive neutralizing responses against multiple HCV genotypes, including the notoriously resistant genotype 2a strains. The epitope-focused approach addresses the fundamental challenge of viral diversity by concentrating immune responses on the most conserved and functionally critical regions of the viral envelope.", "# Microplastics and Health: Alarming New Research Reveals Widespread Biological Impact\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n1. **Hasegawa, Y., et al. (2024).** Oral exposure to high concentrations of polystyrene microplastics alters the intestinal environment and metabolic outcomes in mice. Frontiers in Immunology, 15:1407936. <https://doi.org/10.3389/fimmu.2024.1407936>\n2. **Deng, Y., et al. (2017).** Tissue accumulation of microplastics in mice and biomarker responses suggest widespread health risks of exposure. Scientific Reports, 7:46687. <https://doi.org/10.1038/srep46687>\n3. **Lu, L., et al. (2018).** Polystyrene microplastics induce gut microbiota dysbiosis and hepatic lipid metabolism disorder in mice. Science of the Total Environment, 631-632:449-458. <https://doi.org/10.1016/j.scitotenv.2018.03.051>\n4. **Yang, W., et al. (2020).** Intestinal microbiota-derived short-chain fatty acids regulation of immune cell IL-22 production and gut immunity. Nature Communications, 11:4457. <https://doi.org/10.1038/s41467-020-18262-6>\n5. **Zhao, L., et al. (2021).** Prolonged oral ingestion of microplastics induced inflammation in the liver tissues of C57BL/6J mice. Ecotoxicology and Environmental Safety, 227:112882. <https://doi.org/10.1016/j.ecoenv.2021.112882>\n6. **Stock, V., et al. (2019).** Uptake and effects of orally ingested polystyrene microplastic particles in vitro and in vivo. Archives of Toxicology, 93:1817-1833. <https://doi.org/10.1007/s00204-019-02478-7>\n\nThis comprehensive study investigated the effects of high-concentration polystyrene microplastics (PS-MPs) on intestinal health and metabolism in mice fed normal diets. Researchers exposed C57BL/6J mice to PS-MPs at concentrations of 1000 \u03bcg/L and 5000 \u03bcg/L for six weeks, focusing on intestinal permeability, gut microbiota changes, and metabolic outcomes. The study revealed that even without inducing leaky gut syndrome, high concentrations of PS-MPs significantly increased serum lipid levels and exacerbated fatty liver conditions. The research demonstrated dose-dependent intestinal disturbances characterized by increased natural killer (NK) cell populations, altered gut microbiota composition, and modulated expression of genes related to nutrient transport. Notably, the study showed decreased bacterial diversity with reduced Bacteroidetes and increased inflammation markers, even in the absence of traditional leaky gut symptoms. The findings highlight that PS-MPs can induce dyslipidemia and non-alcoholic fatty liver disease through intestinal inflammation mediated by NK cells, representing a novel pathway of microplastic toxicity that operates independently of compromised intestinal barrier function.\n\n# 2) Oral exposure to polyethylene microplastics induces inflammatory and metabolic changes and promotes fibrosis in mouse liver\n\n* **Lipid Metabolism Disruption**: PE microplastics significantly altered hepatic lipid metabolism by upregulating genes involved in fatty acid uptake, synthesis, and \u03b2-oxidation, leading to increased liver cholesterol and triglycerides.\n* **Inflammatory Response**: Exposure induced chronic low-grade liver inflammation with increased inflammatory foci, elevated pro-inflammatory cytokines (TNF-\u03b1, IL-6), and enhanced hepatocyte proliferation rates.\n* **Fibrogenesis Promotion**: PE microplastics activated hepatic stellate cells and increased collagen deposition, demonstrating early fibrotic changes that could progress to serious liver pathology.\n* **Exacerbation of Pathology**: In CCl4-induced fibrosis models, PE exposure significantly worsened liver fibrogenesis, suggesting microplastics could accelerate pre-existing liver diseases.\n* **Size-Independent Effects**: Both 36 \u03bcm and 116 \u03bcm PE microbeads induced similar hepatotoxic responses, indicating that effects are not solely dependent on particle size.\n* **Time-Dependent Progression**: Hepatotoxic effects showed temporal progression with increasing severity from 6 to 9 weeks of exposure, suggesting cumulative damage potential.\n\n1. **Djouina, M., et al. (2023).** Oral exposure to polyethylene microplastics induces inflammatory and metabolic changes and promotes fibrosis in mouse liver. Ecotoxicology and Environmental Safety, 264:115417. <https://doi.org/10.1016/j.ecoenv.2023.115417>\n2. **Lu, L., et al. (2018).** Polystyrene microplastics induce gut microbiota dysbiosis and hepatic lipid metabolism disorder in mice. Science of the Total Environment, 631-632:449-458. <https://doi.org/10.1016/j.scitotenv.2018.03.051>\n3. **Yang, L., et al. (2022).** Aged polystyrene microplastics induce gut microbiota dysbiosis and hepatic lipid metabolism disorder in mice. Environmental Science & Technology, 56:4968-4979. <https://doi.org/10.1021/acs.est.1c07933>\n4. **Chen, X., et al. (2022).** Chronic exposure to polyvinyl chloride microplastics induces liver injury and gut microbiota dysbiosis. Science of the Total Environment, 839:155984. <https://doi.org/10.1016/j.scitotenv.2022.155984>\n5. **Cheng, W., et al. (2022).** Polystyrene microplastics induce hepatotoxicity and disrupt lipid metabolism in liver organoids. Science of the Total Environment, 806:150328. <https://doi.org/10.1016/j.scitotenv.2021.150328>\n6. **Shen, M., et al. (2022).** Polyethylene microplastics promote liver fibrosis in mice via chronic inflammation. Environmental Science & Technology Letters, 9:482-488. <https://doi.org/10.1021/acs.estlett.2c00137>", "# Compassion: 6+ FDA IND filings supported while mentoring 20+ scientists\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-29 on marcdeller.com (categories: 6cs).*\n\n## Compassion: 6+ FDA IND filings supported while mentoring 20+ scientists\n\nCompassion is what connects the technical side of my work to its human consequences. Supporting [6+ FDA IND filings](https://marcdeller.com/my-work-driving-innovation-in-precision-oncology-through-structure-based-drug-design/) has reinforced that every structural insight and design decision ultimately traces back to patients facing serious disease. Keeping that perspective front and center shapes how I prioritize programs, communicate risk, and advocate for approaches that maximize meaningful benefit. At the same time, mentoring scientists and helping them build confidence and independence is a powerful way to extend impact far beyond my own direct contributions.\n\n* Compassion means supporting both scientific progress and the people driving it.\n* Mentoring young scientists and enabling multiple IND filings reflects a commitment to both impact and growth.\n* It matters because science advances best when empathy and excellence advance together.\n* The goals of my reserach is that every structure eventually impacts a patient\u2014keeping that person in mind shapes every decision upstream.\n* My philosophy on mentorship is showing up\u2014open door, time given freely, listening before advising.\n\nPersonal Philosophy: The 6C's\n\n### [**Curiosity:** 10+ years driving hands-on drug discovery through multidisciplinary approaches](https://marcdeller.com/curiosity-20-years-driving-hands-on-discovery-through-multidisciplinary-approaches/)\n\n[**Curiosity:** 10+ years driving hands-on drug discovery through multidisciplinary approaches](https://marcdeller.com/curiosity-20-years-driving-hands-on-discovery-through-multidisciplinary-approaches/)\n\n### [**Courage**: 400+ challenging structures determined, enabling breakthrough mechanistic insights](https://marcdeller.com/courage-400-structures-determined-enabling-breakthrough-mechanistic-insights/)\n\n[**Courage**: 400+ challenging structures determined, enabling breakthrough mechanistic insights](https://marcdeller.com/courage-400-structures-determined-enabling-breakthrough-mechanistic-insights/)\n\n### [**Creativity:** 7+ patents through innovative protein engineering and AI integration](https://marcdeller.com/creativity-5-patents-through-innovative-protein-engineering-and-ai-integration/)\n\n[**Creativity:** 7+ patents through innovative protein engineering and AI integration](https://marcdeller.com/creativity-5-patents-through-innovative-protein-engineering-and-ai-integration/)\n\n### [**Communication:** 60+ publications translating structural data into actionable insights](https://marcdeller.com/communication-60-publications-translating-structural-data-into-actionable-insights/)\n\n[**Communication:** 60+ publications translating structural data into actionable insights](https://marcdeller.com/communication-60-publications-translating-structural-data-into-actionable-insights/)\n\n### [**Collaboration:** 2+ state-of-the-art Structural Biology facilities built through global partnerships](https://marcdeller.com/collaboration-2-state-of-the-art-structural-biology-facilities-built-through-global-partnerships/)\n\n[**Collaboration:** 2+ state-of-the-art Structural Biology facilities built through global partnerships](https://marcdeller.com/collaboration-2-state-of-the-art-structural-biology-facilities-built-through-global-partnerships/)", "# marcdeller.com: Microplastics in Humans\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n### [The 79th Organ: When Art and Biology Meet Bioengineering in the Fight Against Microplastics](https://marcdeller.com/the-79th-organ-when-biology-meets-bioengineering-in-the-fight-against-microplastics/)\n\nBlog\n\n### [From Waste to Wonder: How Bacteria Transform Plastic into Life-Saving Medicine](https://marcdeller.com/from-waste-to-wonder-how-bacteria-transform-plastic-into-life-saving-medicine/)\n\nBlog\n\n### [Revolutionary AI Tool XenoBug Transforms Environmental Pollution Control Through Predictive Enzyme Discovery](https://marcdeller.com/revolutionary-ai-tool-xenobug-transforms-environmental-pollution-control-through-predictive-enzyme-discovery/)\n\nBlog\n\n### [Microplastics and Health: Alarming New Research Reveals Widespread Biological Impact](https://marcdeller.com/microplastics-and-health-alarming-new-research-reveals-widespread-biological-impact/)\n\nBlog\n\n### [Engineered Enzymes\u2014Design for Safe, Targeted Microplastic Therapeutics](https://marcdeller.com/engineered-enzymes-design-for-safe-targeted-microplastic-therapeutics/)\n\nBlog\n\n### [PETase Enzyme Therapy\u2014A New Frontier in Medical Microplastic Removal](https://marcdeller.com/petase-enzyme-therapy-a-new-frontier-in-medical-microplastic-removal/)\n\nBlog\n\n### [Future Therapeutics\u2014Engineering Solutions for Microplastic Burden](https://marcdeller.com/future-therapeutics-engineering-solutions-for-microplastic-burden/)\n\nBlog\n\n### [Clearing the Bloodstream\u2014Diagnostics and Approaches for Microplastic Removal](https://marcdeller.com/clearing-the-bloodstream-diagnostics-and-approaches-for-microplastic-removal/)\n\nBlog\n\n### [The Enzyme Toolbox\u2014Alternatives to PETase for Microplastic Breakdown](https://marcdeller.com/the-enzyme-toolbox-alternatives-to-petase-for-microplastic-breakdown/)\n\nBlog\n\n### [Plastic-Eating Enzymes Decoded\u2014PETase\u2019s Structure and Mutation Map](https://marcdeller.com/plastic-eating-enzymes-decoded-petases-structure-and-mutation-map/)", "# 2025 \u2013 Fused Bicyclic Pyrimidones as WRN Inhibitors\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-12-10 on marcdeller.com (categories: patents).*\n\n* [Fused Bicyclic Pyrimidones as WRN Inhibitors](https://patents.google.com/patent/US20250333420A1/en)\n\n* **Synthetic Lethal Vulnerability:** [MSI-H cancers depend on WRN helicase to resolve TA-dinucleotide repeat expansions](https://www.nature.com/articles/s41586-024-07350-y), creating a selective therapeutic window absent in normal cells.\n* **Allosteric Inhibition Strategy:** [Bicyclic pyrimidones bind at the D1-D2 helicase domain interface](https://patents.google.com/patent/US20250333420A1/en), inducing conformational changes that prevent ATP hydrolysis and DNA unwinding.\n* **Resistance-Agnostic Activity:** [WRN inhibitors maintain efficacy in MSI models resistant to chemotherapy, targeted therapies, and immune checkpoint blockade](https://aacrjournals.org/cancerdiscovery/article/11/8/1923/666232/Werner-Helicase-Is-a-Synthetic-Lethal), addressing a critical unmet need.\n* **Clinical Translation Underway:** [Multiple WRN inhibitors including HRO761 have entered clinical trials](https://www.nature.com/articles/s41586-024-07350-y) for MSI-H solid tumors with early evidence of safety and preliminary efficacy.\n\n1. **Discovery of WRN inhibitor HRO761 with synthetic lethality in MSI cancers:** [van Wietmarschen et al., Nature, 2024](https://www.nature.com/articles/s41586-024-07350-y)\n2. **Repeat expansions confer WRN dependence in microsatellite-unstable cancers:** [Picco et al., Nature, 2020](https://pmc.ncbi.nlm.nih.gov/articles/PMC8916167/)\n3. **Werner helicase is a synthetic-lethal vulnerability in mismatch repair-deficient colorectal cancer refractory to therapies:** [Kategaya et al., Cancer Discovery, 2021](https://aacrjournals.org/cancerdiscovery/article/11/8/1923/666232/Werner-Helicase-Is-a-Synthetic-Lethal)\n4. **Fused bicyclic pyrimidones as WRN inhibitors:** [Shepard et al., US Patent Application, 2025](https://patents.google.com/patent/US20250333420A1/en)\n\n**Structure-based design yields selective WRN helicase inhibitors exploiting synthetic lethal vulnerabilities in treatment-resistant malignancies**\n\nApproximately 4% of all cancers harbor a hidden vulnerability that researchers are now learning to exploit. These tumors exhibit microsatellite instability (MSI-H), a form of genetic hypermutability arising from defective DNA mismatch repair machinery. While immune checkpoint inhibitors have transformed outcomes for many patients with [MSI-H colorectal, endometrial, and gastric cancers](https://pmc.ncbi.nlm.nih.gov/articles/PMC11861458/), nearly half develop resistance or never respond at all. A recent patent filing by Incyte Corporation describes a new class of fused bicyclic pyrimidone compounds that target the Werner syndrome helicase (WRN), offering a complementary approach for patients who fail immunotherapy.\n\nThe strategic rationale emerges from a molecular peculiarity of MSI cancers. Large-scale CRISPR and RNAi screens identified [WRN as the top synthetic lethal dependency in MSI cell lines](https://www.nature.com/articles/s41586-024-07350-y), with depletion inducing massive DNA damage exclusively in MSI cells while sparing microsatellite-stable counterparts. The mechanistic basis involves TA-dinucleotide repeats that undergo large-scale expansions in the absence of functional mismatch repair. These expanded repeats form non-B DNA secondary structures that stall replication forks. WRN helicase normally resolves these aberrant structures through ATP-dependent unwinding activity, but [when WRN function is disrupted, structure-specific nucleases cleave the unresolved DNA](https://pmc.ncbi.nlm.nih.gov/articles/PMC8916167/), triggering catastrophic chromosome shattering and cell death.\n\nThe Incyte compounds represent a sophisticated medicinal chemistry solution to inhibiting WRN. [Patent US20250333420A1](https://patents.google.com/patent/US20250333420A1/en) discloses bicyclic pyrimidinone scaffolds designed to bind at the interface of the D1 and D2 helicase domains. These molecules function as mixed ATP-competitive inhibitors, mimicking structural features of ATP while inducing allosteric conformational changes that lock WRN in an inactive state. The hydroxy pyrimidine moiety coordinates the catalytic glutamine residue, while the bicyclic core exploits the unusually polar, arginine-rich pocket at the domain interface. This binding mode differs fundamentally from earlier covalent inhibitors that target cysteine residues, potentially offering advantages in selectivity and resistance profiles.\n\nPreclinical validation demonstrates the therapeutic potential of WRN inhibition. [Pharmacological WRN inhibitors induce DNA damage markers, activate ATM and CHK2 checkpoint kinases, and trigger cell cycle arrest](https://aacrjournals.org/cancerdiscovery/article/14/8/1457/746511/Novel-WRN-Helicase-Inhibitors-Selectively-Target) selectively in MSI cells. Notably, this phenotype occurs independently of p53 status, addressing a critical limitation of many cancer therapies. In vivo studies show dose-dependent tumor growth inhibition in MSI cell line and patient-derived xenograft models, including [tumors resistant to chemotherapy and immunotherapy](https://aacrjournals.org/cancerdiscovery/article/11/8/1923/666232/Werner-Helicase-Is-a-Synthetic-Lethal). Combination studies reveal synergy with topoisomerase inhibitors like irinotecan, with subefficacious doses of each agent producing enhanced antiproliferative effects and complete tumor regression in MSI colorectal cancer models.\n\nThe clinical trajectory appears promising. [HRO761, a structurally related compound](https://www.nature.com/articles/s41586-024-07350-y) developed by Novartis, entered human trials in 2023 and has reported encouraging early data in patients with MSI-H tumors refractory to standard therapies. The bicyclic pyrimidone series disclosed by Incyte, developed with contributions from structural biologist Marc Deller and colleagues, may offer next-generation improvements in potency, pharmacokinetics, or selectivity. As the field advances, refined patient stratification based on [TA-repeat expansion burden](https://pmc.ncbi.nlm.nih.gov/articles/PMC8916167/) may further optimize therapeutic windows and identify those most likely to benefit from WRN-targeted therapy.", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nIn some embodiments, the compound or salt described herein is administered with a JAK1 or JAK2 inhibitor (e.g., baricitinib or ruxolitinib). In some embodiments, the compound or salt described herein is administered with a JAK1 inhibitor. In some embodiments, the compound or salt described herein is administered with a JAK1 inhibitor, which is selective over JAK2.\n\nExample antibodies for use in combination therapy include, but are not limited to, trastuzumab (e.g., anti-HER2), ranibizumab (e.g., anti-VEGF-A), bevacizumab (AVASTIN\u2122, e.g., anti-VEGF), panitumumab (e.g., anti-EGFR), cetuximab (e.g., anti-EGFR), rituxan (e.g., anti-CD20), and antibodies directed to c-MET.\n\nExamples of immunotherapy include cytokine treatment (e.g., interferons, GM-CSF, G-CSF, IL-2), CRS-207 immunotherapy, cancer vaccine, monoclonal antibody, bispecific or multi-specific antibody, antibody drug conjugate, adoptive T cell transfer, Toll receptor agonists, STING agonists, RIG-I agonists, oncolytic virotherapy and immunomodulating small molecules, including thalidomide or JAK1/2 inhibitor, PI3K6 inhibitor, and the like.\n\nIn some embodiments, the WRN inhibitor compounds are administered in combination with one or more anti-cancer drugs, such as a chemotherapeutic agent. The specification lists an extensive set of chemotherapeutic agents, including abarelix, aldesleukin, alemtuzumab, alitretinoin, allopurinol, altretamine, anastrozole, arsenic trioxide, asparaginase, azacitidine, bevacizumab, bexarotene, baricitinib, bleomycin, bortezomib, busulfan, calusterone, capecitabine, carboplatin, carmustine, cetuximab, chlorambucil, cisplatin, cladribine, clofarabine, cyclophosphamide, cytarabine, dacarbazine, dactinomycin, dalteparin sodium, dasatinib, daunorubicin, decitabine, denileukin, dexrazoxane, docetaxel, doxorubicin, dromostanolone propionate, eculizumab, epirubicin, erlotinib, estramustine, etoposide phosphate, etoposide, exemestane, fentanyl citrate, filgrastim, floxuridine, fludarabine, fluorouracil, fulvestrant, gefitinib, gemcitabine, gemtuzumab ozogamicin, goserelin acetate, histrelin acetate, ibritumomab tiuxetan, idarubicin, ifosfamide, imatinib mesylate, interferon alfa 2a, irinotecan, lapatinib ditosylate, lenalidomide, letrozole, leucovorin, leuprolide acetate, levamisole, lomustine, meclorethamine, megestrol acetate, melphalan, mercaptopurine, methotrexate, methoxsalen, mitomycin C, mitotane, mitoxantrone, nandrolone phenpropionate, nelarabine, nofetumomab, oxaliplatin, paclitaxel, pamidronate, panitumumab, pegaspargase, pegfilgrastim, pemetrexed disodium, pentostatin, pipobroman, plicamycin, procarbazine, quinacrine, rasburicase, rituximab, ruxolitinib, sorafenib, streptozocin, sunitinib, sunitinib maleate, tamoxifen, temozolomide, teniposide, testolactone, thalidomide, thioguanine, thiotepa, topotecan, toremifene, tositumomab, trastuzumab, tretinoin, uracil mustard, valrubicin, vinblastine, vincristine, vinorelbine, vorinostat, and zoledronate. Additional examples of chemotherapeutics include proteasome inhibitors (e.g., bortezomib), thalidomide, revlimid, and DNA-damaging agents such as melphalan, doxorubicin, cyclophosphamide, vincristine, etoposide, carmustine, and the like. Example steroids include corticosteroids such as dexamethasone or prednisone. Example Bcr-Abl inhibitors include imatinib mesylate (GLEEVAC\u2122), nilotinib, dasatinib, bosutinib, and ponatinib. Example suitable Flt-3 inhibitors include midostaurin, lestaurtinib, linifanib, sunitinib, sunitinib maleate, sorafenib, quizartinib, crenolanib, pacritinib, tandutinib, PLX3397 and ASP2215. Example suitable RAF inhibitors include dabrafenib, sorafenib, and vemurafenib. Example FAK inhibitors include VS-4718, VS-5095, VS-6062, VS-6063, BI853520, and GSK2256098. Example CDK4/6 inhibitors include palbociclib, ribociclib, trilaciclib, lerociclib, and abemaciclib.\n\n### Labeled Compounds and Assay Methods\n\nThe compounds of the present disclosure can further be useful in investigations of biological processes in normal and abnormal tissues. Thus, another aspect of the present disclosure relates to labeled compounds of the present disclosure (radio-labeled, fluorescent-labeled, etc.) that would be useful not only in imaging techniques but also in assays, both in vitro and in vivo, for localizing and quantitating WRN protein in tissue samples, including human, and for identifying WRN ligands by inhibition binding of a labeled compound. Accordingly, the present disclosure includes WRN binding assays that contain such labeled compounds.\n\nThe present disclosure further includes isotopically-labeled compounds of the disclosure. An \"isotopically\" or \"radio-labeled\" compound is a compound of the present disclosure where one or more atoms are replaced or substituted by an atom having an atomic mass or mass number different from the atomic mass or mass number typically found in nature (i.e., naturally occurring). Suitable radionuclides that may be incorporated in compounds of the present disclosure include but are not limited to 2H (also written as T for tritium), 11C, 13C, 14C, 13N, 15N, 15O, 17O, 18O, 18F, 35S, 36Cl, 82Br, 75Br, 76Br, 77Br, 123I, 124I, 125I and 131I. For example, one or more hydrogen atoms in a compound of the present disclosure can be replaced by deuterium atoms (e.g., one or more hydrogen atoms of a C1\u20136 alkyl group of Formula (I) can be optionally substituted with deuterium atoms, such as \u2014CD3 being substituted for \u2014CH3). In some embodiments, alkyl groups in Formula (I) can be perdeuterated.", "# Microplastics and Health: Alarming New Research Reveals Widespread Biological Impact\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n1. **Danso, I.K., et al. (2024).** Pulmonary toxicity assessment of polypropylene, polystyrene, and polyethylene microplastic fragments in mice. Toxicological Research, 40:313-323. <https://doi.org/10.1007/s43188-023-00224-x>\n2. **Li, X., et al. (2022).** Intratracheal administration of polystyrene microplastics induces pulmonary fibrosis by activating oxidative stress and Wnt/\u03b2-catenin signaling pathway in mice. Ecotoxicology and Environmental Safety, 232:113238. <https://doi.org/10.1016/j.ecoenv.2022.113238>\n3. **Cao, J., et al. (2023).** Exposure to polystyrene microplastics triggers lung injury via targeting toll-like receptor 2 and activation of the NF-\u03baB signal in mice. Environmental Pollution, 320:121068. <https://doi.org/10.1016/j.envpol.2023.121068>\n4. **Woo, J.W., et al. (2023).** Polypropylene nanoplastic exposure leads to lung inflammation through p38-mediated NF-\u03baB pathway due to mitochondrial damage. Particle and Fibre Toxicology, 20:2. <https://doi.org/10.1186/s12989-022-00512-8>\n5. **Jin, Y.J., et al. (2023).** Characterisation of changes in global genes expression in the lung of ICR mice in response to the inflammation and fibrosis induced by polystyrene nanoplastics inhalation. Toxicological Research, 39:575-599. <https://doi.org/10.1007/s43188-023-00188-y>\n6. **Jenner, L.C., et al. (2022).** Detection of microplastics in human lung tissue using \u03bcFTIR spectroscopy. Science of the Total Environment, 831:154907. <https://doi.org/10.1016/j.scitotenv.2022.154907>\n\nThis comparative study investigated the differential pulmonary toxicity of three common microplastic polymer types: polypropylene (PP), polystyrene (PS), and polyethylene (PE) fragments in C57BL/6 mice. Researchers administered 5 mg/kg of each microplastic type daily via intratracheal instillation for 14 days, with comprehensive analysis of inflammatory responses, molecular mechanisms, and histopathological changes. The study revealed polymer-specific toxicity patterns, with PS microplastic fragments inducing the most severe pulmonary inflammation characterized by significantly increased inflammatory cell infiltration, elevated cytokine levels (IL-1\u03b2, IL-6), and enhanced chemokine expression (MCP-1, MIP-1\u03b1, MIP-2, KC). Mechanistic investigation identified TLR4 as the primary mediator of PS-induced inflammation, while PP specifically activated TLR2 without downstream inflammatory signaling. The research demonstrated that PS exposure activated NF-\u03baB phosphorylation and NLRP3 inflammasome components (NLRP3, ASC, Caspase-1), leading to sustained inflammatory responses. Interestingly, PE microplastics showed minimal inflammatory effects despite similar exposure conditions. The study also revealed that surface charge properties (zeta potential) correlated with toxicity, with PS showing the highest negative charge (-38.93 mV) and greatest inflammatory potential. These findings establish important precedents for understanding respiratory risks from different microplastic types.\n\n# 6) Uptake and Accumulation of Polystyrene Microplastics in Zebrafish (Danio rerio) and Toxic Effects in Liver\n\n* **Size-Dependent Uptake**: Microplastic accumulation showed clear size-dependent patterns with 5 \u03bcm particles reaching liver tissue while 20 \u03bcm particles remained in gills and gut.\n* **Rapid Bioaccumulation**: PS-MPs achieved steady-state tissue concentrations within 48 hours, demonstrating efficient uptake and retention in multiple organ systems.\n* **Hepatotoxic Effects**: Both 70 nm and 5 \u03bcm PS-MPs induced liver inflammation, lipid accumulation, and cellular damage in exposed zebrafish.\n* **Metabolic Disruption**: Exposure significantly altered hepatic metabolomic profiles, disrupting lipid metabolism, energy production, and amino acid homeostasis.\n* **Oxidative Stress Response**: PS-MPs induced dose-dependent increases in antioxidant enzyme activities (SOD, CAT), indicating cellular oxidative damage.\n* **Unexpected Size Effects**: Larger 5 \u03bcm particles induced greater oxidative stress than smaller 70 nm particles, challenging conventional size-toxicity assumptions.\n\n1. **Lu, Y., et al. (2016).** Uptake and Accumulation of Polystyrene Microplastics in Zebrafish (Danio rerio) and Toxic Effects in Liver. Environmental Science & Technology, 50:4054-4060. <https://doi.org/10.1021/acs.est.6b00183>\n2. **Mattsson, K., et al. (2015).** Altered behavior, physiology and metabolism in fish exposed to polystyrene nanoparticles. Environmental Science & Technology, 49:553-561. <https://doi.org/10.1021/acs.est.5b04532>\n3. **Browne, M.A., et al. (2008).** Ingested microscopic plastic translocates to the circulatory system of the mussel, Mytilus edulis (L.). Environmental Science & Technology, 42:5026-5031. <https://doi.org/10.1021/es800249a>\n4. **Della Torre, C., et al. (2014).** Accumulation and embryotoxicity of polystyrene nanoparticles at early stage of development of sea urchin embryos Paracentrotus lividus. Environmental Science & Technology, 48:12302-12311. <https://doi.org/10.1021/es503297u>\n5. **Cedervall, T., et al. (2012).** Food Chain Transport of Nanoparticles Affects Behaviour and Fat Metabolism in Fish. PLoS One, 7:e32254. <https://doi.org/10.1371/journal.pone.0032254>\n6. **Wright, S.L., et al. (2013).** The physical impacts of microplastics on marine organisms: A review. Environmental Pollution, 178:483-492. <https://doi.org/10.1016/j.envpol.2013.02.031>", "# GitHub README: generate_yaml\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/generate_yaml, 2026\n\n> **From a JSON of candidate SMILES to a folder of Boltz-2 input YAMLs, in one prompt.**\n\n    \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/generate_yaml\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/generate_yaml</a></td>\n</tr>\n</table>\n\n---\n\nA command-line utility for generating Boltz-2 / co-folding input YAML files from a batch of candidate SMILES stored in a JSON file. Designed to work with the output format produced by generative chemistry runs (e.g. fragment-based screening, de novo design) and feed directly into a Boltz-style ModelCIF prediction pipeline.\n\nWhy it matters: generative chemistry runs spit out hundreds of candidate SMILES, but Boltz-2 wants one carefully structured YAML per compound, each pairing the ligand with a protein chain, a pocket constraint, and an affinity block. Hand-writing those is tedious and error-prone. generate_yaml.py reads the candidates straight from your JSON, prompts once for the protein context, and writes a clean, submission-ready YAML for every selected compound. It is useful for anyone bridging de novo design or fragment screening into a co-folding affinity pipeline: turn a screening output into a ready-to-run prediction batch in seconds, with stdlib-only portability and no dependencies to install.\n\n---\n\n## Overview\n\nGiven a JSON file containing a dictionary of candidate SMILES under a `\"candidates\"` key, `generate_yaml.py` interactively prompts for protein context and contact constraints, then writes one YAML file per selected candidate \u2014 ready to submit to Boltz-2 (or a compatible co-folding engine) for structure prediction with binding affinity estimation.\n\nEach output YAML encodes:\n- A single protein chain with user-supplied ID and sequence\n- A ligand defined by cleaned SMILES\n- A pocket constraint anchoring the ligand to a specified contact residue\n- An affinity property block targeting the ligand\n\n---\n\n## Requirements\n\n- Python 3.8+\n- Standard library only (`argparse`, `json`, `re`, `pathlib`) \u2014 no additional dependencies\n\n---\n\n## Installation\n\n```bash\ngit clone https://github.com/<your-org>/generate_yaml.git\ncd generate_yaml\n# No install step required \u2014 pure stdlib\n```\n\n---\n\n## Usage\n\n```bash\npython generate_yaml.py <json_file>\n```\n\n### Example\n\n```bash\npython generate_yaml.py scratch_Baricitinib_sim.out.20260530_0219.json\n```\n\nThe script will interactively prompt for:\n\n| Prompt | Description | Example |\n|---|---|---|\n| `Protein ID` | Chain identifier used in the YAML | `A` |\n| `Protein sequence` | Full single-letter amino acid sequence | `MGHHHHHHSSGVD...` |\n| `Contact residue number` | Residue number defining the binding pocket | `1007` |\n| `Output directory` | Path where YAML files will be written | `~/boltz_runs/baricitinib/` |\n| `Candidate selection` | Which candidates to export (see below) | `1-25` or `all` |\n\n---\n\n## Candidate Selection Syntax\n\n| Format | Description | Example |\n|---|---|---|\n| `all` | Export every candidate | `all` |\n| `N` | A single candidate by 1-based index | `7` |\n| `N-M` | An inclusive range | `1-25` |\n| `N, M, ...` | Arbitrary comma-separated list | `1, 7, 17, 28, 31` |\n\nCandidate numbering is **1-based** and follows the original key order in the JSON `\"candidates\"` dictionary.\n\n---\n\n## Output\n\nFiles are written to the specified output directory, named sequentially by candidate index:\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/generate_yaml\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/generate_yaml</a></td>\n</tr>\n</table>", "# Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\nAuthors: Ye, Q.; Shvartsbart, A.; Li, Z.; Gan, P.; Policarpo, R.L.; Qi, C.; Roach, J.J.; Zhu, W.; McCammant, M.S.; Hu, B.; Li, G.; Yin, H.; Carlsen, P.; Hoang, G.; Zhao, L.; Susick, R.; Zhang, F.; Lai, C.-T.; Allali Hassani, A.; Epling, L.B.; Gallion, A.; Kurzeja-Lipinski, K.; Gallagher, K.; Roman, V.; Farren, M.R.; Kong, W.; Deller, M.C.; Zhang, G.; Covington, M.; Diamond, S.; Kim, S.; Yao, W.; Sokolsky, A.; Wang, X.\nSource: Journal of Medicinal Chemistry, 2025\n\nBQL = below quantifiable limit. Dose for compound 22 in NHP was solubility limited.\n\nIn parallel with the in depth profiling of 18, we continued to conduct SAR to improve the potency and properties of our lead molecule (Table 3). Given the distinct benefit offered by conformational rigidity, we sought to further restrict the degrees of freedom in the morpholinone side chain. To this end, an additional methyl group was introduced at the CH2 linker. We discovered that the (R)-methyl at this position increased potency, with 21, the direct comparator to compound 18, demonstrating a 5 fold improvement of hWB pERK potency, to 100 nM (Table 3). This was the ultimate validation of our search for improved cell translation, as compound 21 had an SPR potency of 0.94 nM, nearly 1000 fold less potent than our starting compound 10, but was more than 5 fold more potent in the hWB pERK assay.\n\nNext, we turned to the ADME properties, in particular permeability, which was limited for 21. Concerned that the basicity of the 2-methylpyridine was detrimental to properties, we examined less basic diazines. 4-Methylpyrazine in particular possessed a favorable balanced of potency and polarity, with increased permeability (Caco-2 1.4 \u00d7 10\u207b\u2076 cm/s) and a further 2 fold boost in potency (compound 22, Table 3). Compound 22 was dosed in rat and cyno PK, but no improvement in oral bioavailability was observed, despite moderate clearance in the i.v. arm (39% hepatic blood flow).\n\nAt this point we shifted our focus to the Switch II pocket 7-fluoronapthalene. While this substituent represented an improvement over the 3-hydroxynapthol fragment, it was still highly electron rich and lipophilic. Extensive screening was undertaken to identify suitable replacements that maintained potency. Naphthyl fragments proved most potent, but we discovered that 2,3-disubstituted phenyl rings could recapitulate a significant amount of the potency while offering clear benefits for properties, presumably due to their decreased lipophilicity and size. 2-Methyl-3-chloro phenyl provided the most balanced compound, 23 (INCB159020, Table 3, entry 7). Potency was reduced 3-fold, to 330 nM, but permeability improved greatly (3.6 \u00d7 10\u207b\u2076 cm/s) with a moderate improvement in clearance (1.0 L/h/kg). Critically, when dosed orally in NHP at 1.5 mg/kg, we were delighted to see an AUC of 950 nM\u00b7h and oral bioavailability of 42% (Table 4).\n\nA full profile of 23 can be seen in Table 5. Unlike many of the G12C inhibitors, which shows preferential binding to the inactive GDP state, 23 binds with equal potency to both the active GTP and inactive GDP form as determined by SPR (2.4 nM GDP form, 2.8 nM GTP form). 23 binding was highly selective for KRAS G12D. 23 shows significant selectivity over inhibition of wild-type KRAS, in binding (50\u00d7 vs binding to KRAS WT), a pERK setting (80 fold selectivity vs a wild-type H838 line) and in a viability setting (28 fold vs a H838 wild-type line). Furthermore, no binding was observed to wild-type HRAS and NRAS in the exchange assay up to a top concentration of 10 \u03bcM. A representative 39 kinase panel showed only two kinases with binding <10 \u03bcM at 1 mM ATP concentration (TRKA: 3.9 \u03bcM, Aurora B: 5.8 \u03bcM, full panel data available in the Supporting Information). CYP inhibition was moderate, with an IC50 for CYP2D6 most notable at 7.2 \u03bcM. hERG channel inhibition was also moderate at 11.2 \u03bcM. Given the high degree of protein binding (1.1% free), there was sufficient margin to progress. No time dependent CYP inhibition or PXR activation was observed. Solubility in SGF buffer was excellent at >900 \u03bcg/mL, solubility in FASSIF was lower but still considerable at 79 \u03bcg/mL. 23 was therefore progressed to assess in vivo activity.\n\n*Table 5.* Full Profile of 23.\n\n| Assay | Result |\n|---|---|\n| KRAS G12D Exc/pERK/CTG Viability | 4.4 nM / 33 nM / 78 nM |\n| WT Exc/pERK/CTG Viability | 205 nM (45\u00d7) / 2635 nM (80\u00d7) / 2251 nM (29\u00d7) |\n| Caco-2 / h-Cl | 3.6 \u00d7 10\u207b\u2076 cm/s / 1.0 L/h/kg |\n| PB (% free) | 1.1 |\n| Solubility (SGF/FASSIF) | >900 \u03bcg/mL / 79 \u03bcg/mL |\n| CYP (3A4/2D6/2C9/2B6/1A2) | 17 \u03bcM / 7.2 \u03bcM / >20 \u03bcM / >20 \u03bcM / >20 \u03bcM |\n| hERG IC50 | 11.2 \u03bcM |\n| CYP TDI | Clean |\n\nCellular testing performed in HPAFII cells. Wild-type cellular testing performed in H838 cells. CYP 3A4 inhibition was measured on the testosterone binding site. hERG determined by manual patch clamp.\n\nA PK/PD study with 23 was conducted in an HPAC mouse model, which contains \u223c20\u221230% WT KRAS stromal cells, limiting maximum inhibition of pERK to \u223c80%. 23 was administered at a single oral dose of 30, 100 and 300 mg/kg and pERK inhibition was measured at 4 h for the 30 and 100 mg/kg doses and at 24 h for all doses; the MEK inhibitor trametinib was used as positive control (Figure 6). A dose dependent inhibition of pERK levels was observed. The 100 mg/kg dose delivered \u223c70% of maximum inhibition of pERK at 4 h, while the 300 mg/kg dose showed near maximal inhibition out to 24 h post dosing. PK measurements were consistent with this data, with the 100 mg/kg dose displaying plasma concentrations around the mouse WB IC50 (950 nM), while the 300 mg/kg dose displayed a plasma concentration near the mouse WB IC90 (8.5 \u03bcM) at 24 h. 23 was well tolerated at all doses.\n\n*Figure 6.* PK/PD evaluation of 23 in mouse HPAC tumors. Trametinib is used as a positive control. HPAC xenografts contain \u223c20\u221230% wild-type KRAS stromal content, suggesting maximum G12D related inhibition of 70\u221280%.\n\n### Modeling of 23", "# Crystal Structure of Nanog Homeodomain (NANOG) from Homo sapiens at 3.30 \u00c5 Resolution\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [4RBO](https://www.rcsb.org/structure/4RBO)\n* **Release Date:** 2014-10-01\n* **Method:** X-RAY DIFFRACTION 3.3 \u00c5\n* **Organisms:** Homo sapiens\n* **Macromolecules:** Putative homeobox protein NANOGP8 (protein), 5\u2032-D(CPTPTPGPAPAPTPGPGPGPCPC)-3\u2032 (nucleic acid), 5\u2032-D(GPGPCPCPCPAPTPTPCPAPAPG)-3\u2032 (nucleic acid)\n* **Publication:** [Jauch, R., et al. (2015). Crystal Structure and DNA Binding of the Homeodomain of the Stem Cell Transcription Factor Nanog. Proceedings of the National Academy of Sciences, 112, 4666-4671.](https://www.pnas.org/doi/full/10.1073/pnas.1502855112)\n\n**Description:** This structure reveals the molecular basis of Nanog homeodomain binding to palindromic DNA, providing atomic-level insight into the transcription factor interactions that govern embryonic stem cell pluripotency and self-renewal. The 3.3 \u00c5 resolution illuminates the structural determinants of DNA sequence recognition by this master regulator of stemness.", "# Transforming Antibody Design: How Artificial Intelligence Revolutionizes Structure Prediction\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-12-07 on marcdeller.com (categories: sb).*\n\n* [Review of Antibody Structure Prediction-Based on Artificial Intelligence](https://link.springer.com/article/10.1007/s11831-025-10404-7)\n\n* **Rapid structure prediction:** Deep learning methods like [IgFold achieve antibody structure prediction in under 25 seconds](https://www.nature.com/articles/s41467-023-38063-x) with CDR-H3 accuracy of approximately 2 \u00c5ngstr\u00f6ms.\n* **Transformer architectures enable long-range modeling:** [AlphaFold2\u2019s attention mechanisms](https://www.pnas.org/doi/10.1073/pnas.2303499120) capture evolutionary relationships and spatial dependencies between amino acids, achieving 30-50 percent success rates for near-native antibody-antigen complex.\n* **Graph neural networks predict binding interactions:** [Geometric deep learning models like GearBind](https://www.nature.com/articles/s41467-024-51563-8) achieve Pearson correlations of 0.707 for binding affinity predictions successfully identifying affinity-enhancing mutations for therapeutic development.\n* **Antibody-specific models outperform general protein tools:** [Fine-tuned architectures such as AntiFold](https://academic.oup.com/bioinformaticsadvances/article/doi/10.1093/bioadv/vbae202/8090019) demonstrate superior CDR sequence recovery and structural accuracy compared to general inverse folding models.\n\n1. **Review of Antibody Structure Prediction-Based on Artificial Intelligence**: [Springer, 2025](https://link.springer.com/article/10.1007/s11831-025-10404-7)\n2. **Fast, accurate antibody structure prediction from deep learning on massive set of natural antibodies**: [Ruffolo et al., Nature Communications, 2023](https://www.nature.com/articles/s41467-023-38063-x)\n3. **Pretrainable geometric graph neural network for antibody affinity maturation**: [Shan et al., Nature Communications, 2024](https://www.nature.com/articles/s41467-024-51563-8)\n4. **Evaluation of AlphaFold antibody\u2013antigen modeling with implications for improving predictions**: [Guest et al., NIH, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC10751731/)\n\n**Neural networks and deep learning are accelerating the development of therapeutic antibodies by predicting complex molecular architectures with unprecedented accuracy**\n\nThe pursuit of accurate antibody structure prediction has historically challenged computational biologists due to the exceptional variability of complementarity-determining regions, particularly the elusive CDR-H3 loop. Antibodies serve as fundamental components of the immune system and represent a rapidly expanding class of therapeutic agents, yet their rational design has long been hindered by reliance on experimental structure determination methods that are both time-intensive and resource-demanding. The past five years have witnessed a paradigm shift as artificial intelligence-based methodologies have emerged to address these longstanding challenges, fundamentally transforming how researchers approach antibody engineering and therapeutic development.\n\nRecent advances in [neural network architectures](https://www.nature.com/articles/s41467-023-38063-x) have enabled end-to-end prediction of antibody structures directly from amino acid sequences, with methods like [IgFold](https://www.nature.com/articles/s41467-023-38063-x) achieving remarkable accuracy in under 25 seconds. This breakthrough stems from pre-trained language models trained on hundreds of millions of natural antibody sequences, followed by graph networks that directly predict three-dimensional atomic coordinates. When benchmarked against established methods, IgFold demonstrated an average CDR-H3 root mean square deviation of [2.01 \u00c5](https://www.nature.com/articles/s41467-023-38063-x), representing substantial improvement over traditional approaches while dramatically reducing computational time. Similarly, [DeepAb](https://pubmed.ncbi.nlm.nih.gov/35199061/) employs interpretable deep learning to predict antibody variable region structures, consistently outperforming alternative methods on therapeutically relevant antibody sets.\n\nThe integration of attention mechanisms and transformer architectures has proven particularly transformative for [antibody structure prediction](https://www.pnas.org/doi/10.1073/pnas.2303499120). AlphaFold2\u2019s revolutionary architecture incorporates a two-track neural network called Evoformer that jointly embeds evolutionary and spatial relationships between amino acids using multiple sequence alignments and pairwise representations. This attention-based framework enables the model to capture long-range dependencies beyond sequential neighborhoods, facilitating accurate modeling of complex antibody-antigen interactions. When [evaluated on antibody-antigen complexes](https://pmc.ncbi.nlm.nih.gov/articles/PMC10751731/), AlphaFold achieved near-native modeling success rates exceeding 30 percent, with massive sampling strategies pushing success rates to approximately 50 percent. The model\u2019s confidence scores correlate strongly with actual prediction accuracy, providing researchers with valuable quality estimates for computational antibody design workflows.", "# Beyond the Microscope: How Cryo-ET is Revealing Disease in Its Native State\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* Cryo-electron tomography (Cryo-ET) allows scientists to visualize protein structures within their native cellular environment, overcoming the limitations of traditional methods.\n* This technique provides crucial insights into the spatial organization of biomolecules, as demonstrated by the *in-tissue* visualization of Alzheimer\u2019s disease pathology.\n* By revealing previously unseen molecular interactions, Cryo-ET is unlocking new strategies for designing drugs against challenging targets.\n\n1. **Recent Trends in Structural Biology:** Insights from the PSDI Conference:\n   <https://3decision.discngine.com/blog/2025/1/10/recent-trends-in-structural-biology-insights-from-the-psdi-conference>\n2. **Structure-based Drug Design:** Evolving with Groundbreaking Advancements:\n   <https://www.drugdiscoverynews.com/structure-based-drug-design-15997>\n3. **Advances in Structural Biology Research:** Opening Novel Opportunities for Impact:\n   <https://www.ukri.org/who-we-are/how-we-are-doing/research-outcomes-and-impact/bbsrc/advances-in-structural-biology-research/>\n\nFor decades, structural biologists have worked to isolate and purify proteins, pulling them from their complex cellular homes to study their structures in pristine, artificial conditions. This approach, while foundational to drug discovery, is like trying to understand a lion by only observing it in a zoo. We get a clear picture, but we miss the intricate behaviors and interactions of its natural habitat. This limitation has created a gap in our understanding, particularly for complex neurodegenerative diseases like Alzheimer\u2019s, where cellular context is everything. How can we design effective drugs if we don\u2019t fully see how pathological proteins behave within the very cells they are destroying?\n\nWe are now entering a new era of structural biology, moving beyond single-particle analysis to visualize molecules directly within their native cellular environment. Cryo-electron tomography (Cryo-ET) is the technology leading this charge. By flash-freezing tissues and taking multiple images from different angles, Cryo-ET builds a 3D reconstruction of cellular landscapes at near-atomic resolution. This technique allowed researchers at the University of Leeds to visualize the actual 3D architecture of \u03b2-amyloid plaques and tau tangles within a postmortem human brain, offering an unprecedented glimpse into Alzheimer\u2019s pathology. Instead of a static snapshot of an isolated protein, we now see the intricate web of interactions within the cell, revealing vulnerabilities that were previously invisible.\n\nThe impact on drug discovery is profound. By understanding the spatial organization and functional dynamics of disease-driving proteins *in situ*, we can move beyond \u201cundruggable\u201d targets and design more effective, precisely targeted therapies. This in-tissue structural data provides a richer, more accurate map for developing next-generation treatments, not just for Alzheimer\u2019s but for a host of diseases where cellular context is key. We are no longer just looking at the lion in the cage; we are finally on a safari, observing it in the wild and learning its true secrets. This leap from isolated structures to cellular landscapes is paving the way for more innovative and successful therapeutic interventions.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nLIF was expressed and purified as a GST fusion in a similar manner to OSM in Chapter 3 (Robinson et al., 1994). Typically, 6 L of culture was grown and the fusion protein purified on glutathione sepharose. Cleavage of the fusion protein was carried out using 105 units of thrombin (Sigma) in 150 mM sodium chloride, 2.5 mM calcium chloride and 50 mM Tris-HCl, pH 8.0 for 16 hours at room temperature.\n\n#### 4.2.6.2 Analysis of LIFR-LIF Complex\n\nAffinity purified LIF and LIFR-FN3 were mixed in molar ratios of 10 to 1 and incubated at room temperature for 1 hour. Samples were centrifuged at 5000 \u00d7 g to remove aggregates and the supernatant applied to a Superose-6 column (24 ml bed volume, 10 \u00d7 300 mm, separation range 5\u20135000 kDa; Pharmacia). Columns were typically run at a flow rate of 2 ml/min in 150 mM sodium chloride, 50 mM Tris-HCl, pH 7.5 on a Biosys 510 FPLC (Beckman). Columns were calibrated by running a cocktail of protein markers consisting of 0.5 mg each of cytochrome C (12.4 kDa), carbonic anhydrase (29 kDa), albumin bovine serum (66 kDa), alcohol dehydrogenase (150 kDa) and blue dextran (2000 kDa).\n\n## 4.3 Results of Expression, Purification and Characterisation of LIFR\n\n### 4.3.1 Stable Expression in CHO Cells\n\nExpression levels of LIFR-FN3-3C-Fc in CHO cells proved exceptionally low as assessed by ELISA as shown in Figure 4-3 below. Maximal levels of expression of ~1 mg/L were typically obtained after 3 weeks incubation in the presence of 2 mM sodium butyrate. Sodium butyrate was shown to increase levels of expression by up to 30 %. All methods of scale-up, including roller bottles, cell factories and stirred-tank bioreactor systems, proved unsuccessful with cells showing poor adherence and protein expression. The inclusion of MSX was shown to be essential for maintenance of clonal selection over many passages although no increase in the levels of protein expression was observed on MSX treatment.\n\nSoluble glycosylated LIFR-FN3-3C-Fc was successfully purified using affinity purification on Protein-A sepharose at an apparent molecular weight of ~140 kDa as assessed by SDS-PAGE (data not shown). Soluble LIFR-FN3 was successfully liberated on 3C protease treatment at an apparent molecular weight of ~116 kDa and preliminary deglycosylation with non-recombinant PNGase F resulted in a species with an apparent molecular weight of ~90 kDa as assessed by SDS-PAGE (data not shown). The identity of LIFR-FN3 was confirmed by N-terminal sequencing although all attempts at mass spectroscopy failed due to lack of ionisation.\n\n*Figure 4-3.* Graph showing levels of expression of LIFR-FN3-3C-Fc by stably transfected CHO cells (time course of expression over weeks 1\u20135; conditions: control, 0.05 mM MSX, 2 mM sodium butyrate, and 0.05 mM MSX plus 2 mM sodium butyrate; protein concentration in ng/ml on a 0\u20131200 scale). Cells were grown as confluent monolayers in 162 cm\u00b2 tissue culture flasks in 30 ml of ULTRA-CHO medium (BioWhittaker). Samples of medium were taken at weekly intervals for analysis by means of ELISA. Protein concentrations were determined by comparison with human IgG1 standards. Control samples represent protein expressed in ULTRA-CHO medium and other data show protein expressed in medium supplemented with 0.05 mM L-methionine sulphoximine (MSX) and 2 mM sodium butyrate.\n\n### 4.3.2 Stable Expression in Schneider's Drosophila Cell Line 2 (S2)\n\nNo induction of LIFR-FN3-3C-Fc expression was detected using ELISA, Protein-A immunoprecipitation or western blot and samples typically contained a protein with an apparent molecular weight of ~46 kDa, compared with control LIFR-FN3-3C-Fc samples running at ~140 kDa (data not shown). No cross reactivity was observed between LIFR antibodies and samples of control expression medium containing serum, so suggesting the production of truncated translation products. Samples of S2 conditioned medium were shown to be inactive in ligand binding assays as no binding to LIF or OSM was observed, compared with control samples of supernatant containing active gp130-Fc or LIFR-Fc which showed sigmoidal dose dependent binding.\n\n### 4.3.3 Transient Expression in 293T Cells\n\n#### 4.3.3.1 pIG-1 DNA Production and Purification\n\nFigure 4-4 below shows a 1 % agarose gel of BamHI, BglII and HindIII digestion of pIG-1-LIFR-FN3-3C-Fc. Lane 5 shows the five main marker bands of the \u03bb HindIII digest, running at 9146, 6882, 2322, 2027 and 564 base pairs. Lane 1 contains BamHI-HindIII digested DNA which shows a unique drop out at ~2505 bp, lane 2 contains BglII-HindIII digested DNA which shows a characteristic drop out at ~1752 bp and lane 3 contains BamHI-BglII digested DNA which shows a unique drop out at ~753 bp. Lane 4 contains undigested pIG-1-LIFR-FN3-3C-Fc DNA with one main band running at ~7900 bp.\n\n*Figure 4-4.* Figure showing 1 % agarose gel of BamHI, BglII and HindIII digestion of pIG-1-LIFR-FN3-3C-Fc DNA. Lane 5 shows the five main marker bands of the \u03bb HindIII digest, running at 9146, 6882, 2322, 2027 and 564 base pairs. Lanes 1, 2 and 3 contain BamHI-HindIII, BglII-HindIII and BamHI-BglII digested samples of pIG-1-LIFR-FN3-3C-Fc DNA, respectively. Lane 4 contains undigested pIG-1-LIFR-FN3-3C-Fc DNA.\n\n#### 4.3.3.2 SDS-PAGE\n\nFigure 4-5 below shows a 12 % polyacrylamide gel of the affinity purification of LIFR-FN3-3C-Fc expressed transiently in 293T cells. Lane 1 shows a sample of Protein-A sepharose after passage of conditioned 293T supernatant and one main band is evident at ~140 kDa. Lane 2 shows a sample of Protein-A sepharose after cleavage with 3C protease and one main band is seen at ~35 kDa. Lanes 3 and 4 show unbound material and a wash fraction, respectively. Lanes 5, 6 and 7 show the first three eluted fractions of LIFR-FN3 following 3C protease cleavage with one main band running at ~116 kDa. Lane 8 contains high range molecular weight markers running at 220, 97, 66, 46, 30, 21 and 14 kDa (Amersham).", "# Microplastics and Health: Alarming New Research Reveals Widespread Biological Impact\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThis significant research examined the hepatotoxic effects of polyethylene (PE) microplastics, one of the most abundant polymers in environmental contamination. Female mice were exposed to 36 and 116 \u03bcm diameter PE microbeads at 100 \u03bcg/g food concentration for 6 and 9 weeks, with comprehensive analysis of liver function, inflammation, and fibrogenesis. The study revealed that PE exposure altered hepatic gene expression related to fatty acid metabolism, promoting increased uptake, synthesis, and \u03b2-oxidation of fatty acids. Biochemical analysis showed elevated liver cholesterol and triglyceride levels, indicating disrupted lipid homeostasis. The research demonstrated that PE microplastics induced low-grade liver inflammation characterized by increased inflammatory foci, elevated cytokine expression (TNF-\u03b1, IL-6), and enhanced cell proliferation. Importantly, the study identified hepatic stellate cell (HSC) activation and increased collagen deposition, indicating early fibrogenic changes. Using a CCl4-induced fibrosis model, researchers confirmed that PE exposure exacerbated liver fibrogenesis, suggesting that microplastic contamination could worsen pre-existing liver pathologies. The findings establish PE microplastics as hepatotoxic agents capable of promoting both metabolic dysfunction and fibrotic progression in mammalian liver tissue.\n\n# 3) Polyethylene terephthalate microplastics affect gut microbiota distribution and intestinal damage in mice\n\n* **Inflammatory Response**: PET-MPs induced dose-dependent increases in pro-inflammatory cytokines (IL-2, TNF-\u03b1) while suppressing anti-inflammatory factors (IL-10) in both systemic circulation and local tissues.\n* **Barrier Dysfunction**: Exposure caused significant intestinal barrier compromise through reduced mucus secretion, damaged epithelial structures, and decreased expression of crucial tight junction proteins.\n* **Microbiota Disruption**: PET-MPs altered gut microbiota composition with temporal dynamics, showing initial dysbiosis followed by compensatory changes, indicating complex host-microbe interactions.\n* **Systemic Effects**: Beyond local intestinal damage, exposure elevated blood glucose and liver enzymes, suggesting systemic metabolic consequences through gut-liver axis disruption.\n* **Dose-Response Relationship**: Effects showed clear dose-dependent patterns with 250 mg/kg exposure causing the most severe inflammatory and structural changes.\n* **Temporal Progression**: The study revealed time-dependent changes in microbiota composition, with different patterns emerging at 14 and 28 days of exposure.\n\n1. **Sun, X., et al. (2025).** Polyethylene terephthalate microplastics affect gut microbiota distribution and intestinal damage in mice. Ecotoxicology and Environmental Safety, 294:118119. <https://doi.org/10.1016/j.ecoenv.2025.118119>\n2. **Li, B., et al. (2020).** Polyethylene microplastics affect the distribution of gut microbiota and inflammation development in mice. Chemosphere, 244:125492. <https://doi.org/10.1016/j.chemosphere.2019.125492>\n3. **Chen, X., et al. (2022).** Polyvinyl chloride microplastics induced gut barrier dysfunction, microbiota dysbiosis and metabolism disorder in adult mice. Ecotoxicology and Environmental Safety, 241:113809. <https://doi.org/10.1016/j.ecoenv.2022.113809>\n4. **Jin, Y., et al. (2019).** Impacts of polystyrene microplastic on the gut barrier, microbiota and metabolism of mice. Science of the Total Environment, 649:308-317. <https://doi.org/10.1016/j.scitotenv.2018.08.353>\n5. **Grogan, A.E., et al. (2021).** Investigation of polyethylene terephthalate (PET) drinking bottles as marine reservoirs for fecal bacteria and phytoplankton. Marine Pollution Bulletin, 173:113052. <https://doi.org/10.1016/j.marpolbul.2021.113052>\n6. **Harusato, A., et al. (2023).** Impact of particulate microplastics generated from polyethylene terephthalate on gut pathology and immune microenvironments. iScience, 26:106474. <https://doi.org/10.1016/j.isci.2023.106474>\n\nThis research investigated the biological effects of polyethylene terephthalate microplastics (PET-MPs), a prevalent environmental contaminant, on intestinal health and gut microbiome composition. Male C57BL/6 mice were exposed to 10 \u03bcm PET-MPs at concentrations of 10, 50, and 250 mg/kg body weight for 28 days, with comprehensive analysis of inflammatory responses, intestinal barrier function, and microbiota changes. The study revealed dose-dependent increases in inflammatory cytokines (IL-2, TNF-\u03b1) and decreased anti-inflammatory markers (IL-10) in both serum and colon tissues. Histopathological examination showed inflammatory cell infiltration, structural damage to colonic crypts, and reduced mucus secretion, indicating compromised intestinal barrier integrity. Molecular analysis demonstrated significantly decreased expression of mucus-related genes (Muc2, Muc3) and tight junction proteins (Occludin, Claudin-1), confirming barrier dysfunction. Gut microbiota analysis revealed altered diversity and composition, with temporal changes showing initial increases in Firmicutes/Bacteroidetes ratio at 14 days, followed by reversed patterns at 28 days. The research also identified elevated blood glucose and liver enzyme levels, suggesting systemic metabolic effects. These findings establish PET-MPs as potent disruptors of intestinal homeostasis with implications for systemic health through gut-liver axis dysfunction.\n\n# 4) Polystyrene microplastics induce myocardial inflammation and cell death via the TLR4/NF-\u03baB pathway in carp", "# Crystal Structure of a Putative Metal-Dependent Phosphoesterase (YP_910028.1) from Bifidobacterium adolescentis at 1.94 \u00c5 Resolution\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [3O0F](https://www.rcsb.org/structure/3O0F)\n* **Release Date:** 2010-09-29\n* **Method:** X-RAY DIFFRACTION 1.94 \u00c5\n* **Organisms:** Bifidobacterium adolescentis ATCC 15703\n* **Macromolecules:** Putative metal-dependent phosphoesterase (protein)\n* **Unique Ligands:** AMP, CL, MN, ZN\n* **Publication:** [Han, G.W., et al. (2011). Crystal Structure of a Metal-Dependent Phosphoesterase (YP_910028.1) from Bifidobacterium adolescentis. Proteins, 79, 2146-2160.](https://doi.org/10.1002/prot.23035)\n\n**Description:** This structure of a putative metal-dependent phosphoesterase from the gut commensal B. adolescentis demonstrates the power of combining computational prediction with structural genomics. Despite initial bioinformatics suggesting DNA polymerase function, the crystal structure and theoretical analysis correctly predicted phosphoesterase activity\u2014confirmed experimentally\u2014revealing widespread misannotation in sequence databases.", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nSome of the compounds prepared were also separated on a preparative scale by reverse-phase high performance liquid chromatography (RP-HPLC) with MS detector or flash chromatography (silica gel) as indicated in the Examples. Typical preparative reverse-phase high performance liquid chromatography (RP-HPLC) column conditions are as follows: pH=2 purifications: Waters Sunfire\u2122 C18 5 \u03bcm, 30\u00d7100 mm or Waters XBridge C18 5 \u03bcm, 30\u00d7100 mm column, eluting with mobile phase A: 0.1% TFA (trifluoroacetic acid) in water and mobile phase B: acetonitrile; the flow rate was 60 mL/min, the separating gradient was optimized for each compound using the Compound Specific Method Optimization protocol as described in the literature. pH=10 purifications: Waters XBridge C18 5 \u03bcm, 30\u00d7100 mm column, eluting with mobile phase A: 0.1% NH4OH in water and mobile phase B: acetonitrile; the flow rate was 60 mL/minute, the separating gradient was optimized for each compound using the Compound Specific Method Optimization protocol.\n\nThe following abbreviations may be used herein: aq. (aqueous); br (broad); d (doublet); dd (doublet of doublets); DCM (dichloromethane); DMF (N, N-dimethylformamide); Et (ethyl); EtOAc (ethyl acetate); g (gram(s)); h (hrs) (hour(s)); HPLC (high performance liquid chromatography); Hz (hertz); J (coupling constant); LCMS (liquid chromatography-mass spectrometry); m (multiplet); M (molar); MS (Mass spectrometry); Me (methyl); MeCN (acetonitrile); MeOH (methanol); mg (milligram(s)); min. (minutes(s)); mL (milliliter(s)); mmol (millimole(s)); nM (nanomolar); NMR (nuclear magnetic resonance spectroscopy); Ph (phenyl); r.t. (room temperature), s (singlet); t (triplet or tertiary); TBS (tert-butyldimethylsilyl); tert (tertiary); tt (triplet of triplets); TFA (trifluoroacetic acid); THF (tetrahydrofuran); \u03bcg (microgram(s)); \u03bcL (microliter(s)); \u03bcM (micromolar); wt % (weight percent).\n\n### Example 1. N-(2-(2-Ethyl-10-(2-((4-fluorophenyl)amino)-2-oxoethyl)-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide\n\n**Step 1. 2-Ethylbenzo[4,5]imidazo[1,2-a]pyrimidin-4(10H)-one.** A mixture of 1H-benzo[d]imidazol-2-amine (2.00 g, 15.0 mmol) and ethyl propionylacetate (2.57 mL, 18.0 mmol) in DMF (30 mL) was stirred at 140 \u00b0C for 24 h. After cooling, the reaction mixture was poured into water (300 mL) and the resulting suspension was stirred for 1 h. The solids were then filtered and dried over vacuum for 3 h to provide the product (1.95 g, 61%). LC-MS calculated for C11H12N3O (M+H)+: m/z=214.1; found 214.1.\n\n**Step 2. 3-Bromo-2-ethylbenzo[4,5]imidazo[1,2-a]pyrimidin-4(10H)-one.** To a mixture of 2-ethylbenzo[4,5]imidazo[1,2-a]pyrimidin-4(1H)-one (600 mg, 2.81 mmol) in DMF (28 mL) was added NBS (551 mg, 3.10 mmol). The reaction mixture was then allowed to stir at 25 \u00b0C for 15 min. Upon completion, the reaction was diluted with water (300 mL), and the resulting precipitate was filtered and dried under vacuum for 3 h to provide the product (786 mg, 96%). LC-MS calculated for C11H11BrN3O (M+H)+: m/z=292.0, 294.0; found 292.0, 294.0.\n\n**Step 3. 3-Bromo-2-ethyl-10-((2-(trimethylsilyl)ethoxy)methyl)benzo[4,5]imidazo[1,2-a]pyrimidin-4(10H)-one.** To a solution of 3-bromo-2-ethylbenzo[4,5]imidazo[1,2-a]pyrimidin-4(10H)-one (786 mg, 2.69 mmol) in DMF (18 mL) was added (2-(chloromethoxy)ethyl)trimethylsilane (0.72 mL, 4.04 mmol) followed by diisopropylethylamine (1.40 mL, 8.07 mmol). The reaction mixture was allowed to stir at 25 \u00b0C for 1 h. Upon completion, the reaction contents were diluted with EtOAc (60 mL), then washed 3\u00d7 with H2O (40 mL), dried over Na2SO4, and concentrated. Purification by flash column chromatography (0-100% CH2Cl2/EtOAc) to provide the desired product (1.00 g, 88%). LC-MS calculated for C18H25BrN3O2Si (M+H)+: m/z=422.1, 424.1; found 422.0, 424.0.\n\n**Step 4. 3-(2-Aminophenyl)-2-ethyl-10-((2-(trimethylsilyl)ethoxy)methyl)benzo[4,5]imidazo[1,2-a]pyrimidin-4(10H)-one.** To a solution of 3-bromo-2-ethyl-10-((2-(trimethylsilyl)ethoxy)methyl)benzo[4,5]imidazo[1,2-a]pyrimidin-4(10H)-one (1.00 g, 2.37 mmol) in 1,4-dioxane (20.7 ml) and water (3.0 ml) was added (2-aminophenyl)boronic acid (486 mg, 3.55 mmol), potassium phosphate tribasic (1.01 g, 4.73 mmol), and XPhos Pd G4 (204 mg, 0.24 mmol). The reaction mixture was then degassed for 5 min before heating to 90 \u00b0C. The reaction was stirred for 1 h. Upon completion, the reaction mixture was quenched with H2O, and the resulting mixture was extracted with EtOAc (20 mL) 3\u00d7. The organic phase was dried over Na2SO4, concentrated, and purified by flash column chromatography (0-100% hex/EtOAc) to provide the desired product (992 mg, 96%). LC-MS calculated for C24H31N4O2Si (M+H)+: m/z=435.2; found 435.3.\n\n**Step 5. N-(2-(2-Ethyl-4-oxo-10-((2-(trimethylsilyl)ethoxy)methyl)-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide.** To a solution of 3-(2-aminophenyl)-2-ethyl-10-((2-(trimethylsilyl)ethoxy)methyl)benzo[4,5]imidazo[1,2-a]pyrimidin-4(10H)-one (990 mg, 2.28 mmol) in CH2Cl2 (23 mL) at 0 \u00b0C was added Et3N (635 \u03bcL, 4.56 mmol) and acryloyl chloride (0.28 mL, 3.42 mmol). The reaction was then allowed to stir at this temperature for 10 min. Upon completion, the reaction mixture was carried forward directly as a solution, assuming full conversion. LC-MS calculated for C27H33N4O3Si (M+H)+: m/z=489.2; found 489.2.", "# Approaches to automated protein crystal harvesting\nAuthors: Deller, M. C.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 70, 133\u2013155 (IYCr crystallization series), 2014\n\nREACH was developed by NatX-ray with scientists from beamline FIP-BM30A at the ESRF and performs semi-autonomous harvesting, cryoprotection and flash-cooling, plus in situ data collection directly from the plate. It builds on the commercial G-Rob system, whose St\u00e4ubli six-axis arm orients plates for in situ collection and also acts as a goniometer for flash-cooled and capillary-mounted samples \u2014 attractive for workflow because no tool change is needed, but an engineering challenge because serial devices have large spheres of confusion; air-bearing final axes have reduced these to a few micrometres and a NatX-ray prototype reaches 6 \u00b5m. Central to REACH is a two-finger piezoelectric microgripper from Femto-ST with 1 \u00b5m resolution and a 0\u2013500 \u00b5m opening range, able to harvest all but the largest crystals; unlike the six-degree-of-freedom chopsticks fingers, the REACH grippers operate with two degrees of freedom in the open and closed planes, all other motions being handled by the arm. Grip is improved by concave SU-8 pads, 60 \u00b5m wide and 30 \u00b5m thick, added at the finger tips. A fully motorized inverted-microscope Visualization Bench performs automated plate screening and crystal tracking from any SBS plate, and the user steers the arm with a game-pad controller much as with UMR. Crystals can be cryoprotected before or after harvesting and are pre-centred for direct data collection while still held in the grippers. REACH harvested crystals in the 40\u2013400 \u00b5m range in times comparable to manual methods, with significant savings when coupled to direct data collection.\n\n### 2.2.10\u20132.2.11. Optical tweezers: COT and FOT\n\nAn optical trap is produced by tightly focusing a laser through high-numerical-aperture lenses; work from the early 1970s showed focused lasers generate piconewton forces able to levitate and displace micrometre-sized dielectric objects. A trapped particle experiences a scattering force along the beam and a gradient force from fluctuating dipoles, whose net effect balances gravity and holds the object at equilibrium. The key attraction for crystallography is manipulation of fragile microcrystals without mechanical contact. The conventional optical tweezers system at Diamond Light Source uses a modified Zeiss PALM MicroTweezers microscope with 63\u00d7 water-immersion objectives and a 1.5 W 1064 nm Nd:YAG laser, plus a SmarAct micromanipulator for precise loop mounting. Laser radiation damaged polyimide film mounts at powers below 50 mW; 50 mW was needed to levitate a 10 \u00b5m CPV polyhedron crystal, and operation below 120 mW was a good compromise. Laser-harvested CPV polyhedrin crystals gave a 1.5 \u00c5 structure, better than the 2.1 \u00c5 original in the PDB. COT drawbacks are bulky, expensive high-NA lenses limiting placement, a limited focal range restricting work to thin or shallow media, and power constraints confining it to the smallest crystals harvested directly from glass coverslips. The fibre optical tweezers system at SPring-8, built for BL32XU, removes those limits using a twin-lensed fibre approach with two 1064 nm ytterbium fibre laser modules coupled to lensed NSOM probes, an M200 XYZ micromanipulator and a third micromanipulator positioning the receiving microloop. FOT needed only about 10 mW versus 50 mW for COT, though it could still occasionally melt polyimide micromount film and the plate had to be cooled to prevent drop vaporization. Data sets from FOT-mounted lysozyme and thermolysin showed no significant structural change, indicating minimal photo-, thermal or oxidative damage. FOT handled crystals of 5\u201330 \u00b5m, wider than the 10 \u00b5m reported for COT, but could not manipulate crystals above 30 \u00b5m.\n\n### 2.2.12. RodBots, and other commercial systems\n\nThe RodBot, developed at the Multi-Scale Robotics Laboratory of ETH Zurich, is a rod-shaped microrobot that transports crystals to a microloop for conventional harvesting and flash-cooling. RodBots are wireless mobile devices typically 55 \u00d7 55 \u00d7 500 \u00b5m, sized to match the target crystal, made of SU-8 embedded with sticks of soft magnetic material and driven by low-frequency magnetic fields from an octopole electromagnetic coil assembly beneath the crystallization plate. Rotation in the low-Reynolds-number liquid generates fluid currents that produce a rising flow ahead of the device and a stable vortex above it; crystals are lifted, trapped in the vortex, transported and released at the desired location such as a microloop, in approach, transport and deposit phases. The method works in solutions ranging from pure water to concentrated high-molecular-weight PEGs, spreads gentle fluidic force smoothly over the whole crystal surface and so avoids localized physical damage, and handles crystals from a few micrometres to submillimetre size. RodBots can be manoeuvred in any plate format, making them compatible with existing crystallization hardware, and can run standalone or be integrated into larger platforms. The review also notes two commercial systems, the Crystal Harvester from Bruker and the Station for Automated Microscopy and Imaging from FMP Products, excluded from detailed treatment for lack of peer-reviewed material.\n\n### 3. Assessment: harvesting mechanisms compared", "# Call Me A Jerk: Persuading AI to Comply with Objectionable Requests\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-09-03 on marcdeller.com (categories: ai).*\n\n* Link to Paper\n\n* **Persuasive Vulnerability**: AI systems designed to be helpful and conversational inherit human-like susceptibility to social engineering attacks, achieving 92%+ success rates against advanced models\n* **Systematic Exploitation**: A comprehensive taxonomy of 40 persuasion techniques from social science research enables automated generation of sophisticated jailbreak attempts\n* **Defense Inadequacy**: Current protective mechanisms show significant gaps against persuasive attacks, with many exhibiting problematic over-refusal rates while failing to prevent social manipulation\n* **Evaluation Challenges**: Traditional jailbreak assessment methods underestimate the effectiveness of persuasive attacks, requiring new evaluation frameworks that account for social engineering tactics\n* **Multi-layered Security**: Effective defense requires combining technical safeguards with social awareness, using multiple complementary approaches to achieve meaningful protection\n\n1. [Call Me A Jerk: Persuading AI to Comply with Objectionable Requests](https://download.ssrn.com/2025/7/18/5357179.pdf): Zeng, Y., et al., ACL, 2024\n2. [How Johnny Can Persuade LLMs to Jailbreak Them: Rethinking Persuasion to Challenge AI Safety by Humanizing LLMs](https://arxiv.org/html/2401.06373v2): Zeng, Y., et al., arXiv, 2024\n3. [Evolving Security in LLMs: A Study of Jailbreak Attacks and Defense Mechanisms](https://arxiv.org/html/2504.02080v1): Liu, X., et al., arXiv, 2024\n4. [AI Alignment: A Comprehensive Survey](https://arxiv.org/abs/2310.19852): Ji, J., et al., arXiv, 2023\n5. [LLM Security: Vulnerabilities, Attacks, Defenses, and Countermeasures](https://arxiv.org/html/2505.01177v1): Aguilera-Mart\u00ednez, F., et al., arXiv, 2025\n6. [2025 AI Safety Index](https://futureoflife.org/ai-safety-index-summer-2025/): Future of Life Institute, 2025\n\n**The Hidden Art of Digital Persuasion: How Social Engineering Exploits AI\u2019s Human-Like Vulnerabilities**\n\nArtificial intelligence systems designed to interact naturally with humans have inadvertently inherited one of humanity\u2019s most exploitable traits: susceptibility to persuasion. Recent groundbreaking research reveals that large language models, despite their sophisticated safety mechanisms, can be systematically manipulated using the same psychological techniques that influence human decision-making. This discovery fundamentally challenges our understanding of AI security and demonstrates that as we make AI more human-like, we also make it more vulnerable to the ancient art of persuasion.\n\nThe implications are staggering. Traditional cybersecurity approaches focused on technical exploits and algorithmic vulnerabilities, but the emergence of Persuasive Adversarial Prompts represents a paradigm shift toward social engineering attacks specifically designed for AI systems. These attacks achieve remarkable success rates, with some techniques compromising even the most advanced models like GPT-4 and Claude with over 92% effectiveness. Unlike conventional jailbreaking methods that rely on complex optimization or cryptographic obfuscation, persuasive attacks exploit the very mechanisms that make AI systems helpful and conversational.\n\nThe research introduces a systematic taxonomy derived from decades of social science research, encompassing 40 distinct persuasion techniques organized into 13 strategic categories. This comprehensive framework bridges the gap between psychological research and AI safety, revealing how concepts like social proof, authority endorsement, and emotional appeals can be weaponized against machine learning systems. The taxonomy covers both ethical techniques, such as evidence-based persuasion and logical appeals, and unethical methods including deception, threats, and social manipulation.\n\nWhat makes these attacks particularly insidious is their subtlety and scalability. Rather than relying on adversarial optimization that produces gibberish-like prompts, persuasive attacks generate human-readable text that appears legitimate to both automated filters and human reviewers. The attacks can be automated through a \u201cPersuasive Paraphraser\u201d \u2013 a fine-tuned language model that transforms harmful queries into compelling requests using established persuasion principles. This automation enables large-scale deployment of sophisticated social engineering attacks against AI systems.\n\nThe vulnerability extends beyond simple prompt injection. Multi-turn conversation strategies, such as the \u201ccrescendo technique\u201d and \u201cbad Likert judge\u201d method, demonstrate how attackers can gradually escalate interactions to bypass safety mechanisms. These approaches mirror real-world social engineering tactics, where attackers build rapport and trust before making increasingly problematic requests. The research shows that even defensive systems specifically designed to detect malicious prompts struggle against well-crafted persuasive attacks.\n\nCurrent defense mechanisms reveal significant gaps when confronted with persuasive attacks. Traditional approaches like output filtering, safety training, and prompt detection systems were primarily designed to handle technical exploits rather than sophisticated social manipulation. Evaluation studies show that many existing defenses exhibit alarming rates of \u201cover-refusal,\u201d inappropriately blocking benign requests while still failing to prevent persuasive attacks. This creates a dangerous paradox where defensive systems become less useful without becoming more secure.", "# Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\nAuthors: Zhuo, J.; Li, Y.-L.; Qian, D.-Q.; Burns, D.M.; Mei, S.; Rafalski, M.; Xu, M.; Cao, G.; Pan, Y.; Jia, Z.; Jalluri, R.; Epling, L.B.; Fenalti, G.; Deller, M.C.; Procak, J.; Covington, M.; He, X.; Collins, R.; Stubbs, M.; Burke, K.; Oliver, J.; Margulis, A.; Boer, J.; Wynn, R.; Scherle, P.; Diamond, S.; Newton, R.; Zhang, Y.; Metcalf, B.; Yao, W.\nSource: Journal of Medicinal Chemistry (Drug Annotation), 2026\n\nAll reagents and solvents were obtained from commercial sources and were used directly without further purification. Anhydrous reactions were conducted in DriSolv solvents provided by EMD Chemicals Inc. and were conducted under an atmosphere of nitrogen. Reactions were monitored by liquid chromatography/mass spectrometry (LC/MS) analysis using a Waters SunFire C18 column (2.1 \u00d7 50 mm, 5 \u03bcM particle size), eluting with 0.025% TFA/water and 0.025% TFA/acetonitrile using a mass spectrum scan range of 105\u2013900 Da. Purifications by flash chromatography were performed on RediSep columns (Teledyne ISCO) using an ISCO CombiFlash SG100c system. Preparative LC/MS purifications were performed on a Waters FractionLynx system using either UV or mass-directed fractionation and compound-specific optimization. The LC method used a Waters SunFire column (19 \u00d7 100 mm, 5 \u03bcM particle size), eluting with 0.1% TFA/water and 0.1% TFA/acetonitrile gradient at a flow rate of 30 mL/min over a total run time of 5 min. Nuclear magnetic resonance (NMR) spectra were obtained using either a Varian Mercury-300, Mercury-400, or Inova-500 spectrometer. Chemical shifts are reported in parts per million (ppm) relative to tetramethylsilane as an internal standard.\n\nPreparatory LC-MS purifications of some of the compounds prepared were performed on Waters mass-directed fractionation systems. The basic equipment setup, protocols, and control software for the operation of these systems have been described in detail in the literature. The compounds separated were typically subjected to analytical LC/MS for purity analysis under the following conditions: instrument: Agilent 1100 series, LC/MSD; column: Waters Sunfire C18 5 \u03bcm, 2.1 \u00d7 50 mm; mobile phase A: 0.025% TFA in water and mobile phase B: acetonitrile; gradient 2% to 80% B in 3 min with flow rate 2.0 mL/min. All final compounds were >95% pure by HPLC analysis.\n\nSome of the compounds prepared were also separated on a preparative scale by reverse-phase high-performance liquid chromatography (RP-HPLC) with an MS detector or flash chromatography (silica gel), as indicated in the examples. Typical preparative RP-HPLC column conditions are as follows: pH = 2 purifications: Waters Sunfire C18 5 \u03bcm, 30 \u00d7 100 mm or Waters XBridge C18 5 \u03bcm, 30 \u00d7 100 mm columns, eluting with mobile phase A: 0.1% TFA in water and mobile phase B: acetonitrile; the flow rate was 60 mL/min, the separating gradient was optimized for each compound using the Compound-Specific Method Optimization protocol as described in the literature. pH = 10 purifications: Waters XBridge C18 5 \u03bcm, 30 \u00d7 100 mm column, eluting with mobile phase A: 0.1% NH4OH in water and mobile phase B: acetonitrile; the flow rate was 60 mL/min, the separating gradient was optimized for each compound using the Compound Specific Method Optimization protocol as described in the literature.\n\n### General Synthetic Procedures, Synthesis, and Characterization of Compounds 4\u201341\n\nFull step-by-step synthetic procedures with LC/MS (ESI), 1H NMR, 13C NMR, 19F NMR and HRMS characterization data are reported in the manuscript for compounds 4\u201341 and the synthetic intermediates 23\u201342, including:", "# GitHub README: ChemSage\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ChemSage, 2026\n\nNotes: Round 3 was configured for 1,500 iters but stopped early at actual iter 625 after val loss plateaued at ~0.060. Training was resumed from iter 100 checkpoint after a battery shutdown. 64 layers was attempted first but caused swap thrashing at 31.5 GB peak; reverted to 32 layers. batch_size=8 was also attempted but ran 4.5x slower due to memory pressure; reverted to 4. Round 4 uses 32 layers (same as R3) \u2014 48 and 64 layers were found to be too slow (~2\u20132.5 min/iter) on M1 Max; 32 layers is the throughput sweet spot at ~40 sec/iter. Round 5 doubles the LoRA rank to 64 with RSLoRA scale 90.0 (effective alpha \u2248 11.25, same order as R4's 11.31), triples the iters to 3,000, extends max_seq_length to 3,072, and doubles val_batches to 50 for a more stable eval signal. Trainable parameters doubled to 268M (0.82%).\n\n### Training loss curves\n\nValidation loss measured every 25 steps over 25 batches. Round 3 was resumed from iter 100 checkpoint after a battery shutdown; actual iter = logged iter + 100 for the resumed portion.\n\n| Iter | R1 Val | R2 Val | R3 Val | R4 Val | R5 Val | Notes |\n|------|--------|--------|--------|--------|--------|-------|\n| 1 | 3.800 | 3.191 | 2.110 | 2.397 | **2.210** | Baseline; R5 lower than R4 despite 2.5\u00d7 data \u2014 rank=64 adapter fits better |\n| 25 | 1.020 | 0.580 | 0.238 | 0.480 | \u2014 | |\n| 50 | 0.740 | 0.533 | 0.145 | 0.277 | \u2014 | R3 already below R2's best |\n| 75 | 0.630 | 0.527 | 0.122 | 0.200 | \u2014 | |\n| 100 | 0.580 | 0.495 | 0.110 | 0.146 | \u2014 | R3 battery shutdown; resumed from here |\n| 200 | 0.490 | 0.449 | 0.116 | 0.109 | \u2014 | |\n| 300 | 0.440 | 0.416 | 0.076 | 0.082 | \u2014 | |\n| 400 | 0.420 | 0.400 | 0.059 | 0.074 | \u2014 | |\n| 500 | 0.405 | 0.375 | 0.060 | 0.071 | \u2014 | |\n| 600 | 0.395 | 0.359 | 0.060 | 0.059 | \u2014 | R4 matches R3's best |\n| 625 | \u2014 | \u2014 | **0.054** | 0.055 | \u2014 | **Best R3 checkpoint (fused)** |\n| 650 | \u2014 | \u2014 | \u2014 | 0.052 | \u2014 | R4 beats R3 best |\n| 675 | **0.389** | **0.347** | \u2014 | 0.069 | \u2014 | Best R1/R2 checkpoint |\n| 700 | 0.390 | 0.372 | \u2014 | 0.058 | \u2014 | |\n| 750 | \u2014 | \u2014 | \u2014 | 0.056 | \u2014 | |\n| 775 | \u2014 | \u2014 | \u2014 | 0.049 | \u2014 | |\n| 800 | \u2014 | \u2014 | \u2014 | 0.051 | \u2014 | |\n| 950 | \u2014 | \u2014 | \u2014 | **0.041** | \u2014 | **Best R4 checkpoint (fused)** |\n| 1000 | \u2014 | \u2014 | \u2014 | 0.043 | \u2014 | |\n| 1100 | \u2014 | \u2014 | \u2014 | 0.043 | \u2014 | |\n| 1425 | \u2014 | \u2014 | \u2014 | **0.041** | \u2014 | Ties R4 best at iter 950 |\n| 1500 | \u2014 | \u2014 | \u2014 | 0.050 | \u2014 | R4 final iter |\n| 1600 | \u2014 | \u2014 | \u2014 | \u2014 | **0.055** | **Best R5 checkpoint (fused)** |\n| 2000 | \u2014 | \u2014 | \u2014 | \u2014 | 0.061 | R5 early stop |\n\nRound 3 val loss (**0.054**) is **6.4x better** than Round 2 (0.347) and **7.2x better** than Round 1 (0.389). The improvement is driven by dataset quality (19 behaviour classes with code-heavy examples) and 3.3x more training data (4,000 vs 1,200 examples). Round 3 was stopped early at iter 625 of 1,500 after val loss plateaued.\n\nRound 4 **completed 2026-06-27** (1,500 iters, ~43 h). Best val loss: **0.041** at iter 950, beating R3's best of 0.054 by **24%**. The best was also tied at iter 1,425. Fused from iter 950 checkpoint into `models/chem_sage_32b_v4/` (~17.2 GB). Peak memory 25.5 GB, 32 adapted layers, 134M trainable parameters (0.41%).\n\nRound 5 **early-stopped 2026-06-29** at iter 2000 of 3,000. Val loss plateaued at 0.055 from iter 1600 (true best 0.054 at iters 1650 and 1950, but fell between checkpoint saves). Fused from iter 1600 checkpoint into `models/chem_sage_32b_v5/` (~17.2 GB). Higher floor than R4 (0.055 vs 0.041) reflects the harder task: 78 behaviour classes vs 59, 20k examples vs 6.4k.\n\n### Evaluation results\n\n#### Original scorecard eval (per-round, `eval/eval_chem.py`)\n\n| Metric | Round 1 (7B) | Round 2 (32B) | Round 3 (32B) |\n|--------|------|-------|-------|\n| SMILES validity | 0/0 (n/a) | 25/25 (100%) | **494/499 (99%)** |\n| Tool executability | 0/8 (0%) | 72/174 (41%) | **437/442 (99%)** |\n| Numerical fidelity | 0/0 (n/a) | 7/73 (10%) | **363/533 (68%)** |\n| Code blocks generated | 8 | 174 | 442 |\n| **Overall** | **0%** | **51%** | **89%** |\n\nScorecard files: `eval/scorecards/scorecard_r1_7b.html`, `scorecard_r2_32b.html`, `scorecard_r3_32b_v2.html`.\n\n#### Comparative eval \u2014 all 5 rounds (2026-06-29, `eval/compare/eval_compare.py`)\n\n100 shared R5 test examples (seed=42). Full report: `eval/compare/results/compare_20260629_1928.html`.\n\nScores: per-example pass/fail (examples where all instances correct / 100). Metrics degrade\ngracefully to `N/A` when a metric's trigger type never appears in a model's responses.\n\n| Metric | Round 1 | Round 2 | Round 3 | Round 4 | **Round 5** | \u0394 vs R4 |\n|---|---|---|---|---|---|---|\n| SMILES validity | 100% | 100% | 100% | 100% | **100%** | +0% |\n| SMARTS validity | N/A | 0% | 100% | 55% | **100%** | +45% |\n| Tool executability | 0% | 28% | 79% | 71% | **95%** | +24% |\n| Code attempted | 14% | 41% | 100% | 97% | **100%** | +3% |\n| Python exec (extended) | 36% | 34% | 78% | 69% | **99%** | +30% |\n| Code\u2192Quote fidelity | N/A | 0% | 47% | 19% | **61%** | +42% |\n| Numerical fidelity | N/A | 18% | 57% | 47% | **89%** | +42% |\n| Rounding precision | 100% | 100% | 98% | 98% | **99%** | +1% |\n| Refusal accuracy | 98% | 98% | 97% | 98% | **100%** | +2% |\n| QED range | 100% | 100% | 100% | 100% | **100%** | +0% |\n| PDB ID validity | 100% | 100% | 100% | 100% | **100%** | +0% |\n| PyMOL syntax | 77% | 97% | 89% | 89% | **90%** | +1% |\n| Degeneration-free | 93% | 96% | 91% | 98% | **100%** | +2% |\n| **Overall** | **72%** | **63%** | **87%** | **80%** | **95%** | **+15%** |\n\nR5 wins on all exec/fidelity targets: **exec +24%**, **PyExec +30%**, **fidelity +42%**,\n**Code\u2192Quote +42%**, **SMARTS +45%**, **Degeneration-free 100%**, **Refusal 100%**.\n\nThe R4 exec regression vs R3 (71% vs 79%) was caused by dataset distribution shift towards\nstructural biology content that reduced the density of pure-RDKit drills. R5 corrected this\nwith `pyexec_drill` (\u00d76), `code_then_quote_v2` (\u00d75), and `fidelity_multistep` (\u00d75) generators.\n\n## \ud83d\udcda RAG corpus", "# Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\nAuthors: Ye, Q.; Shvartsbart, A.; Li, Z.; Gan, P.; Policarpo, R.L.; Qi, C.; Roach, J.J.; Zhu, W.; McCammant, M.S.; Hu, B.; Li, G.; Yin, H.; Carlsen, P.; Hoang, G.; Zhao, L.; Susick, R.; Zhang, F.; Lai, C.-T.; Allali Hassani, A.; Epling, L.B.; Gallion, A.; Kurzeja-Lipinski, K.; Gallagher, K.; Roman, V.; Farren, M.R.; Kong, W.; Deller, M.C.; Zhang, G.; Covington, M.; Diamond, S.; Kim, S.; Yao, W.; Sokolsky, A.; Wang, X.\nSource: Journal of Medicinal Chemistry, 2025\n\n**Step 5. 3-(1-((1R,4R,5S)-2-Azabicyclo[2.1.1]hexan-5-yl)-8-(2-cyanoethyl)-4-(3-(dimethylamino)azetidin-1-yl)-6-fluoro-7-(3-hydroxynaphthalen-1-yl)-1H-imidazo[4,5-c]quinolin-2-yl)-N,N-dimethylpropanamide (5).** A mixture of 33 (900 mg, 1.29 mmol), 34 (419 mg, 1.55 mmol), XPhos Pd G2 (51 mg, 0.065 mmol) and K3PO4 (551 mg, 2.59 mmol) in dioxane/water (4:1, 5 mL) was heated to 100 \u00b0C for 2 h. The reaction mixture was quenched with water and extracted with ethyl acetate. The organic layer was washed with water and brine, dried over sodium sulfate and concentrated. The crude product was dissolved in 1:1 DCM/TFA (6 mL) and stirred at r.t. for 30 min, then diluted with acetonitrile and purified by reverse-phase HPLC (pH 2) to provide the desired product. LCMS calculated for C38H42FN8O2 (M + H)+: m/z = 660.3; found: 660.2. 1H NMR (600 MHz, DMSO-d6) \u03b4 10.16 (s, 1H), 8.04 (s, 1H), 7.82\u22127.73 (m, 1H), 7.45\u22127.36 (m, 1H), 7.24 (s, 1H), 7.18\u22127.13 (m, 1H), 7.12\u22127.05 (m, 2H), 4.75 (s, 1H), 4.51\u22124.34 (m, 2H), 4.34\u22124.30 (m, 1H), 4.20\u22124.05 (m, 2H), 3.63\u22123.48 (m, 1H), 3.26\u22123.13 (m, 3H), 3.06 (s, 3H), 2.96\u22122.85 (m, 2H), 2.84 (s, 3H), 2.76\u22122.67 (m, 1H), 2.64\u22122.55 (m, 4H), 2.14 (s, 6H), 1.92\u22121.85 (m, 1H), 1.29\u22121.18 (m, 2H). 13C NMR (126 MHz, DMSO) \u03b4 166.9, 155.7, 154.4, 149.4, 148.6, 143.8, 143.6, 142.9, 140.7, 137.2, 135.7, 135.2, 134.9, 134.8, 133.1, 129.9, 129.5, 128.1, 125.2, 120.6, 119.9, 119.1, 118.5, 109.4, 67.7, 63.9, 60.6, 58.1, 46.4, 44.9, 43.6, 40.7, 30.4, 28.8, 21.6, 17.6, 17.5, 16.3.\n\n### Synthesis of Compound 16\n\n**Step 1. tert-Butyl (1R,4R,5S)-5-((7-Bromo-6-(2-cyanoethyl)-8-fluoro-2-(methylthio)-3-(3-(2-oxopiperidin-1-yl)prop-1-yn-1-yl)quinolin-4-yl)amino)-2-azabicyclo[2.1.1]hexane-2-carboxylate (42).** To a mixture of 37 (590 mg, 0.91 mmol), copper iodide (69 mg, 0.36 mmol) and Pd(PPh3)4 (211 mg, 0.18 mmol) were added DMF (3 mL), Hunig's base (0.8 mL, 4.56 mmol) and 1-(prop-2-yn-1-yl)piperidin-2-one (250 mg, 1.82 mmol). The reaction was sparged with nitrogen, then stirred at 70 \u00b0C for 2 h, then quenched with a small amount of sat. aq. ammonium hydroxide. The mixture was diluted with water and extracted with ethyl acetate. The organic layer was washed with water and brine, dried over sodium sulfate and concentrated. The crude product was purified by column chromatography (0\u2212100% ethyl acetate in heptane) to provide the desired product (335 mg, 56%). LCMS calculated for C31H36BrFN5O3S (M + H)+: m/z = 656.2, 658.2; found: 656.2, 658.2.\n\n**Step 2. tert-Butyl (1R,4R,5S)-5-(7-Bromo-8-(2-cyanoethyl)-6-fluoro-4-(methylthio)-2-((2-oxopiperidin-1-yl)methyl)-1H-pyrrolo[3,2-c]quinolin-1-yl)-2-azabicyclo[2.1.1]hexane-2-carboxylate (43).** To a solution of 42 (335 mg, 0.51 mmol) in THF (5 mL) were added IPrAuCl (63 mg, 0.10 mmol) and silver tetrafluoroborate (149 mg, 0.76 mmol) and the reaction mixture was stirred at 70 \u00b0C under nitrogen for 2 h. \u223c10\u221220% Boc loss was observed by LCMS, so triethylamine (36 \u03bcL) and Boc-anhydride (223 mg, 1.02 mmol) were added. After 30 min, the reaction was quenched with brine and extracted with ethyl acetate. The organic layer was washed with sat. aqueous ammonium hydroxide, dried over magnesium sulfate and concentrated. The crude product was purified by flash column chromatography (0\u2212100% ethyl acetate in heptane) to provide the desired product (253 mg, 75%). LCMS calculated for C31H36BrFN5O3S (M + H)+: m/z = 656.2, 658.2; found: 656.2, 658.2.\n\n**Step 3. tert-Butyl (1R,4R,5S)-5-(7-Bromo-8-(2-cyanoethyl)-6-fluoro-4-(methylsulfinyl)-2-((2-oxopiperidin-1-yl)methyl)-1H-pyrrolo[3,2-c]quinolin-1-yl)-2-azabicyclo[2.1.1]hexane-2-carboxylate (44).** To a solution of 43 (253 mg, 0.385 mmol) in DCM (2.5 mL) at 0 \u00b0C was added m-CPBA (86 mg, 0.385 mmol) and the reaction mixture was stirred at this temperature for 30 min. The reaction was quenched with sat. sodium thiosulfate and extracted with DCM. The organic layer was washed with brine, dried over sodium sulfate and concentrated. The crude product was purified by column chromatography (0\u2212100% ethyl acetate in hexanes) to provide the desired product. LCMS calculated for C31H36BrFN5O4S (M + H)+: m/z = 672.2, 674.2; found: 672.2, 674.2.\n\n**Step 4. tert-Butyl (1R,4R,5S)-5-(7-Bromo-8-(2-cyanoethyl)-6-fluoro-4-methyl-2-((2-oxopiperidin-1-yl)methyl)-1H-pyrrolo[3,2-c]quinolin-1-yl)-2-azabicyclo[2.1.1]hexane-2-carboxylate (45).** To a solution of 44 (311 mg, 0.46 mmol) in THF (2.5 mL) at 0 \u00b0C was added MeMgBr (3 M in THF, 385 \u03bcL, 1.16 mmol) dropwise and the reaction mixture was stirred at r.t. for 30 min, then quenched with sat. ammonium chloride and extracted with ethyl acetate. The organic layer was washed with water and brine, dried over sodium sulfate and concentrated. The crude product was purified by flash chromatography (0\u2212100% ethyl acetate in hexanes) to provide the desired product. LCMS calculated for C31H36BrFN5O3 (M + H)+: m/z = 624.2, 626.2; found: 624.2, 626.2.", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nSome examples of suitable excipients include lactose, dextrose, sucrose, sorbitol, mannitol, starches, gum acacia, calcium phosphate, alginates, tragacanth, gelatin, calcium silicate, microcrystalline cellulose, polyvinylpyrrolidone, cellulose, water, syrup, and methyl cellulose. The formulations can additionally include: lubricating agents such as talc, magnesium stearate, and mineral oil; wetting agents; emulsifying and suspending agents; preserving agents such as methyl- and propylhydroxy-benzoates; sweetening agents; and flavoring agents. The compositions of the present disclosure can be formulated so as to provide quick, sustained or delayed release of the active ingredient after administration to the patient by employing procedures known in the art.\n\nIn some embodiments, the pharmaceutical composition comprises silicified microcrystalline cellulose (SMCC) and at least one compound described herein, or a pharmaceutically acceptable salt thereof. In some embodiments, the silicified microcrystalline cellulose comprises about 98% microcrystalline cellulose and about 2% silicon dioxide w/w.\n\nIn some embodiments, the composition is a sustained release composition comprising at least one compound described herein, or a pharmaceutically acceptable salt thereof, and at least one pharmaceutically acceptable carrier or excipient. In some embodiments, the composition comprises at least one compound described herein, or a pharmaceutically acceptable salt thereof, and at least one component selected from microcrystalline cellulose, lactose monohydrate, hydroxypropyl methylcellulose and polyethylene oxide. In some embodiments, the composition comprises at least one compound described herein, or a pharmaceutically acceptable salt thereof, and microcrystalline cellulose, lactose monohydrate and hydroxypropyl methylcellulose. In some embodiments, the composition comprises at least one compound described herein, or a pharmaceutically acceptable salt thereof, and microcrystalline cellulose, lactose monohydrate and polyethylene oxide. In some embodiments, the composition further comprises magnesium stearate or silicon dioxide. In some embodiments, the microcrystalline cellulose is Avicel PH102\u2122. In some embodiments, the lactose monohydrate is Fast-flo 316\u2122. In some embodiments, the hydroxypropyl methylcellulose is hydroxypropyl methylcellulose 2208 K4M (e.g., Methocel K4 M Premier\u2122) and/or hydroxypropyl methylcellulose 2208 K100LV (e.g., Methocel K00LV\u2122). In some embodiments, the polyethylene oxide is polyethylene oxide WSR 1105 (e.g., Polyox WSR 1105\u2122).\n\nIn some embodiments, a wet granulation process is used to produce the composition. In some embodiments, a dry granulation process is used to produce the composition.\n\nThe compositions can be formulated in a unit dosage form, each dosage containing from about 5 to about 1,000 mg (1 g), more usually about 100 mg to about 500 mg, of the active ingredient. In some embodiments, each dosage contains about 10 mg of the active ingredient. In some embodiments, each dosage contains about 50 mg of the active ingredient. In some embodiments, each dosage contains about 25 mg of the active ingredient. The term \"unit dosage forms\" refers to physically discrete units suitable as unitary dosages for human subjects and other mammals, each unit containing a predetermined quantity of active material calculated to produce the desired therapeutic effect, in association with a suitable pharmaceutical excipient.\n\nThe components used to formulate the pharmaceutical compositions are of high purity and are substantially free of potentially harmful contaminants (e.g., at least National Food grade, generally at least analytical grade, and more typically at least pharmaceutical grade). Particularly for human consumption, the composition is preferably manufactured or formulated under Good Manufacturing Practice standards as defined in the applicable regulations of the U.S. Food and Drug Administration. For example, suitable formulations can be sterile and/or substantially isotonic and/or in full compliance with all Good Manufacturing Practice regulations of the U.S. Food and Drug Administration.\n\nThe active compound may be effective over a wide dosage range and is generally administered in a therapeutically effective amount. It will be understood, however, that the amount of the compound actually administered will usually be determined by a physician, according to the relevant circumstances, including the condition to be treated, the chosen route of administration, the actual compound administered, the age, weight, and response of the individual patient, the severity of the patient's symptoms, and the like.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics; supervisor Prof. E. Yvonne Jones), 1999\n\n*Figure 5-3.* Figure showing Harker sections of anomalous difference Patterson map of EMP derivatised \u0394188_OSM produced using the program FFT (Ten Eyck, 1973). Sections at X = 0.5, Y = 0.5 and Z = 0.5 are shown contoured from 2 \u03c3 in 1 \u03c3 steps using the program GROPAT (Esnouf, 1993). Data in the resolution range 10\u20133 \u00c5 were used. The two major sites are denoted as site 1 and site 2 at refined fractional coordinates of 0.22, 0.14, 0.20 and 0.94, 0.00, 0.13, respectively. A third minor site identified through inspection of anomalous difference Fouriers is denoted as site 3 at refined fractional coordinates of 0.32, 0.33, 0.24. Corresponding peaks in each section are shown with connecting lines.\n\n*Figure 5-4.* Figure showing anomalous difference Fourier map of EMP derivatised \u0394188_OSM. Map produced using phases calculated from site 1 and site 2 with data in the range 10\u20133 \u00c5 using the program FFT (Ten Eyck, 1973). Sections in X = 1, Y = 0.5 and Z = 0.5 are shown contoured from 2 \u03c3 in 1 \u03c3 steps using the program GROPAT (Esnouf, 1993). The two major sites are denoted as site 1 and site 2 at refined orthogonal coordinates of 7.86 \u00c5, 7.43 \u00c5, 21.37 \u00c5 and 33.60 \u00c5, 0.00 \u00c5, 13.89 \u00c5, respectively. A third minor site is denoted as site 3 at refined orthogonal coordinates of 11.44 \u00c5, 17.51 \u00c5, 25.65 \u00c5.\n\n*Figure 5-5.* Figure showing final refined model of \u0394188_OSM with bound (a) EMP represented by silver transparent spheres and (b) hypothetical Se-Met substitutions represented by gold transparent spheres. Transparent dotted section represents loop region as observed in LIF. Heavy atom positions were found by manual inspection of anomalous difference Patterson maps and anomalous difference Fourier maps. Refinement of heavy atom positions was carried out using SHARP (de La Fortelle and Bricogne, 1997). Residues His 48, His 171 and His 178 responsible for heavy metal binding are shown as ball and stick objects. Secondary structure is coloured blue at the N-terminus through to red at the C-terminus. Figure produced using BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994).\n\n*Figure 5-6.* Figure showing fluorescence scan of an EMP derivatised \u0394188_OSM crystal. Energy region corresponds to mercury LIII absorption edge. Data were collected from a single EMP derivative propane cooled P212121 crystal on beam line BL19 at the ESRF, Grenoble, France. Three wavelengths selected for data collection are shown as peak (\u03bb = 1.00669 \u00c5, f\u2032 = \u221214.8 e\u2212, f\u2033 = 10.1 e\u2212, E = 12.3163 keV), inflexion (\u03bb = 1.00914 \u00c5 minimum f\u2032 = \u221220.3 e\u2212, f\u2033 = 10.2 e\u2212, E = 12.2863 keV) and remote (\u03bb = 0.826516 \u00c5, f\u2032 = \u22127.4 e\u2212, f\u2033 = 11.8 e\u2212, E = 15.0009 keV). All values are experimental values prior to refinement.\n\nFrom Table 5-2 it can be seen that MAD data collected from EMP derivatised crystals is of good quality with an I/\u03c3(I) value in the outer resolution shell (2.66\u20132.60 \u00c5) greater than 2.9 for all wavelengths. Completeness in the outer resolution shell is 63.1 % or better. Initial maps based on MAD phases and Fobs from the remote data set (figure of merit = 0.46, phasing power = 1.84) proved untraceable. Various phasing packages were tried including SOLVE (Terwilliger and Berendzen, 1999), PHASES (Furey and Swaminathan, 1996) and SHARP (de La Fortelle and Bricogne, 1997). Hendrickson-Lattman coefficients from an independent native-remote single isomorphous replacement with anomalous scattering (SIRAS) experiment were combined with the MAD phases using SHARP. Maps based on the combined phases and Fobs from the remote data set (figure of merit = 0.51, phasing power = 1.90) showed improved main chain density. A summary of phasing statistics and the programs used is shown below in Table 5-4.\n\n| Phasing Package | Experiment | Dispersive (1) / Bijvoet (2) Difference <\\|\u0394F\\|>/<FP> | Figure of Merit \\|F(hkl)best\\|/\\|F(hkl)\\| | Phasing Power \u03a3n\\|FH\\|/\u03a3n\\|E\\| |\n|---|---|---|---|---|\n| SOLVE | MAD | 5.3 % | 0.42 | \u2013 |\n| PHASES | MAD | 5.0 % | 0.23 | 1.33 |\n| SHARP | MAD | 9.9 % (peak\u2212inflexion) *(1); 9.7 % (inflexion\u2212inflexion) *(2); 10.6 % (remote\u2212inflexion) *(1) | 0.46 | 1.84 |\n| SHARP | SIRAS | 16.3 % (remote\u2212native) *(1) | 0.12 | 0.95 |\n| SHARP | MAD-SIRAS | \u2013 | 0.51 | 1.90 |\n\n*Table 5-4.* Table showing statistics for MAD phasing of EMP derivatised \u0394188_OSM crystals. * denotes values calculated using program SCALEIT (Howell and Smith, 1992). (1) and (2) denote dispersive or Bijvoet differences. All data was scaled relative to inflexion point data except for SIRAS experiment in which data was scaled to native data. SIRAS denotes independent phasing experiment between native and remote MAD data. MAD-SIRAS denotes combined phases from MAD and SIRAS experiment.\n\n#### 5.3.6 Molecular Replacement\n\nData were initially processed in the spacegroup I212121 and this yielded strong peaks in the cross-rotation function, but no peaks in the translation function, so suggesting an alternate spacegroup. Cross-rotation functions of data processed in the spacegroup I222 produced a clear solution with a correlation coefficient of 21.5 %. From Table 5-5 below it can be seen that this solution shows a 9 % improvement in correlation coefficient compared to the next best solution. Solution 1 of the cross-rotation function was processed by the translation function and Table 5-6 below shows a clear solution with a correlation coefficient of 55.0 % and an R-factor of 43.5 %. These values compare favorably with the next best solution, which has a lower correlation coefficient of 41.9 %, and a higher R-factor of 49.6 %. Rigid body refinement resulted in an increase in the correlation coefficient to a value of 64.5 % and a drop in the R-factor to a value of 40.5 % for all data in the resolution range 12.0\u20133.5 \u00c5.", "# Microplastics and Cancer: The Mounting Evidence via Literature Review\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-16 on marcdeller.com (categories: microplastics).*\n\nBar diagram of the published documents about microplastics and cancer. Lower numbers in 2025 due to mid year publication date.\n\n* [Data from: Microplastics as Emerging Carcinogens: From Environmental Pollutants to Oncogenic Drivers: Mishra et al., Mol Cancer, 2025](https://link.springer.com/article/10.1186/s12943-025-02409-4)\n\n* **Found and now feared:** \u00a0Microplastics are now detected in human tumor tissue, often at levels above those in adjacent healthy tissue.\n* **Plausible mechanisms:** Particles drive oxidative stress, inflammation, and DNA damage, activating NF-\u03baB, PI3K/Akt, and Wnt pathways.\n* **Pro-metastatic hints:** In colorectal and renal cell carcinoma models, exposure increases cell migration, invasion, and tumor growth.\n* **Polymer chemistry is important:** Polymer identity shapes the response, not just particle size. PET nanoplastics act through a distinct TNF-\u03b1/JAK1/STAT1 axis in gastric models, while polystyrene dominates the colorectal and renal data.\n* **Causation pending:** Although no long-term epidemiological data so far directly link microplastic exposure to higher cancer incidence in humans, mounting evidence underscores the need for further studies.\n\n1. **Microplastics as Emerging Carcinogens: From Environmental Pollutants to Oncogenic Drivers:** [Mishra et al., Mol Cancer, 2025](https://doi.org/10.1186/s12943-025-02409-4)\n2. **Microplastics and Cancer:** [Bogani et al., Eur J Cancer Prev, 2025](https://doi.org/10.1097/CEJ.0000000000000999)\n3. **Microplastics Role in Cell Migration and Distribution During Cancer Cell Division:** [Brynzak-Schreiber et al., Chemosphere, 2024](https://doi.org/10.1016/j.chemosphere.2024.141463)\n4. **Emerging Evidence on Micro- and Nanoplastics Carcinogenicity: Mechanisms, Models, and Signaling Networks:** [Alaraby et al., Mol Cancer, 2026](https://doi.org/10.1186/s12943-026-02657-y)\n5. **Polystyrene Microplastics Aggravate Clear Cell Renal Cell Carcinoma Progression via NF-\u03baB and TGF-\u03b2:** [Ye et al., Adv Sci, 2025](https://doi.org/10.1002/advs.202518500)\n\n**Particles turn up in malignant tissue at higher levels than in healthy tissue, but causation remains unproven**\n\nMicroplastics (under 5 mm) and nanoplastics (under 1 \u00b5m) now reach almost every human tissue we look at: blood, placenta, lung, colon, and, increasingly, tumors. For oncology, the question is no longer whether these particles are present, but whether they do anything once they arrive. Several 2025 syntheses have brought together the scattered evidence. A comprehensive review in *Molecular Cancer* [catalogs particle accumulation across lung, colorectal, gastric, cervical, breast, pancreatic, and prostate malignancies](https://doi.org/10.1186/s12943-025-02409-4), and assembles the mechanistic case linking exposure to disordered cell behavior. The pattern is consistent enough to take seriously and incomplete enough to resist firm conclusions.\n\nThe most cited observation is also the most provocative: particles are frequently recovered from [malignant tissue at higher concentrations than from adjacent healthy tissue](https://doi.org/10.1097/CEJ.0000000000000999). Whether that reflects a causal contribution or simply differential retention inside a disorganized tumor is the central unanswered question.\n\n> **\u201cincreasing evidence associating their bioaccumulation with cancer onset\u201d** \u2013\u00a0[Mishra et al., Mol Cancer, 2025](https://doi.org/10.1186/s12943-025-02409-4)\n\nMechanistically, the proposals are plausible and convergent. Internalized particles generate reactive oxygen species, sustain chronic inflammation, and damage DNA. In cultured colorectal cancer lines, polystyrene micro- and nanoparticles are [taken up in a size- and dose-dependent manner, persist through cell division and amplify migration](https://doi.org/10.1016/j.chemosphere.2024.141463), behavior with clear pro-metastatic overtones. Across *in vitro* and *in vivo* models, exposure repeatedly converges on the same oncogenic circuitry, including [NF-\u03baB, PI3K/Akt/mTOR and Wnt/\u03b2-catenin signalling](https://doi.org/10.1186/s12943-026-02657-y). Particle chemistry matters too: PET nanoplastics, close to my own enzyme work, behave differently again in [gastric models](https://doi.org/10.1016/j.envint.2026.110289).\n\nThe strongest recent experiments move beyond cell lines. In clear-cell renal carcinoma, polystyrene microplastics were [markedly more abundant in tumor than in adjacent normal tissue](https://doi.org/10.1002/advs.202518500), and in patient-derived and xenograft models, they accelerated growth via NF-\u03baB and TGF-\u03b2 signaling, an effect blunted by pathway inhibitors. That is the shape of a mechanism, not yet proof of cause in people. No epidemiological study has shown that microplastic exposure raises cancer incidence, and reverse causation is hard to exclude when the sampling site is the tumor itself. The honest position today is that an active hypothesis with real biological plausibility is waiting on longitudinal human data.\n\n## Complete Reference List\n\nA wider bibliography of microplastic and cancer literature (2022\u20132026) identified via PubMed, grouped by study type. All DOIs verified live.", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\n### Example 24. 2-(4-Cyano-3-ethyl-1-oxo-2-(2-(vinylsulfonamido)phenyl)benzo[4,5]imidazo[1,2-a]pyridin-5(1H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide\n\n**Step 1. 2-(1H-Benzo[d]imidazol-2-yl)acetonitrile.** A mixture of benzene-1,2-diamine (5.0 g, 46 mmol, Sigma-Aldrich P23938) and ethyl 2-cyanoacetate (16 g, 140 mmol, Sigma-Aldrich E18425) in DMF (10 mL) was heated to 160 \u00b0C for 4 h. Upon cooling to room temperature, the mixture was diluted with EtOAc (50 mL) and filtered. The filtrate was washed with brine (3\u00d7), dried over Na2SO4, filtered and concentrated in vacuo. The product was purified via flash column chromatography, eluting with a gradient of 0-10% MeOH in DCM to afford the title compound (5.0 g, 69%). LCMS for C9H8N3 (M+H)+: calculated m/z=158.1; found 158.0.\n\n**Step 2. 3-Ethyl-1-oxo-1,5-dihydrobenzo[4,5]imidazo[1,2-a]pyridine-4-carbonitrile.** A mixture of ethyl 3-oxopentanoate (1.9 g, 13 mmol, PharmaBlock PBT9077), 2-(1H-benzo[d]imidazol-2-yl)acetonitrile (from Step 1, 1.7 g, 11 mmol) and ammonium acetate (1.7 g, 22 mmol) in DMF (2.8 mL) was heated at 140 \u00b0C for 1.5 h with periodic venting. Upon cooling to room temperature, water (15 mL) was added and the suspension was stirred for 30 min. The solid product was isolated by filtration and was washed with water (10 mL) and air dried to afford the title compound which was used without further purification (2.3 g, 87%). LCMS for C14H12N3O (M+H)+: calculated m/z=238.1; found 238.1. 1H NMR (500 MHz, DMSO-d6) \u03b4 13.60 (s, 1H), 8.74-8.41 (m, 1H), 7.58-7.49 (m, 2H), 7.42-7.34 (m, 1H), 5.96 (s, 1H), 2.68 (q, J=7.5 Hz, 2H), 1.26 (t, J=7.5 Hz, 3H).\n\n**Step 3. 2-Bromo-3-ethyl-1-oxo-1,5-dihydrobenzo[4,5]imidazo[1,2-a]pyridine-4-carbonitrile.** To a mixture of 3-ethyl-1-oxo-1,5-dihydrobenzo[4,5]imidazo[1,2-a]pyridine-4-carbonitrile (from Step 2, 0.66 g, 2.8 mmol) in DMF (14 mL) at 0 \u00b0C was added N-bromosuccinimide (0.55 mg, 3.1 mmol) in portions (over 50 m). Saturated NaHCO3 solution was added, followed by brine, and the mixture was extracted with EtOAc. The organic extract was washed with brine (2\u00d7), and the combined aqueous layers were extracted with another portion of EtOAc. The combined organic extracts were dried over Na2SO4, filtered and concentrated in vacuo. The product was purified via flash column chromatography, eluting with a gradient of 0-30% EtOAc in hexanes to afford the title compound (0.40 g, 46%). LCMS for C14H11BrN3 (M+H)+: calculated monoisotopic m/z=316.0; found 316.0. 1H NMR (500 MHz, DMSO-d6) \u03b4 8.59-8.54 (m, 1H), 7.60-7.57 (m, 2H), 7.47-7.37 (m, 1H), 2.90 (q, J=7.6 Hz, 2H), 1.25 (t, J=7.6 Hz, 3H).\n\n**Step 4. 2-(2-Bromo-4-cyano-3-ethyl-1-oxobenzo[4,5]imidazo[1,2-a]pyridin-5(1H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide.** To a slurry of 2-bromo-3-ethyl-1-oxo-1,5-dihydrobenzo[4,5]imidazo[1,2-a]pyridine-4-carbonitrile (from Step 3, 0.050 g, 0.16 mmol), K2CO3 (44 mg, 0.32 mmol), and tetrabutylammonium bromide (5.1 mg, 0.016 mmol) in CH3CN (1.5 mL) at 0 \u00b0C was added 2-bromo-N-(4-(trifluoromethyl)phenyl)acetamide (prepared via the general method given for Intermediate A in Example 1, Step 7, substituting 4-(trifluoromethyl)aniline in place of 2-bromo-N-(4-fluorophenyl)acetamide, 54 mg, 0.19 mmol). The mixture was allowed to warm to room temperature and stir for 3 d. The reaction was repeated on twice the scale, and both reaction mixtures were combined and diluted with EtOAc. The organic mixture was washed with water, followed by brine, dried over Na2SO4, filtered and concentrated in vacuo. The product was purified via flash column chromatography, eluting with a gradient of 0-30% EtOAc in hexanes to afford the title compound (0.025 g, 10%). LCMS for C23H17BrF3N4O2 (M+H)+: calculated monoisotopic m/z=517.0; found 517.1.\n\n**Step 5. 2-(2-(2-Aminophenyl)-4-cyano-3-ethyl-1-oxobenzo[4,5]imidazo[1,2-a]pyridin-5(1H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide.** A mixture of 2-(2-bromo-4-cyano-3-ethyl-1-oxobenzo[4,5]imidazo[1,2-a]pyridin-5(1H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide (from Step 4, 25 mg, 0.048 mmol), (2-aminophenyl)boronic acid (9.9 mg, 0.072 mmol, Combi-Blocks BB-2139), XPhos Pd G4 (4.2 mg, 4.8 \u03bcmol), and K3PO4 (21 mg, 0.097 mmol) in 1,4-dioxane (0.60 mL) and water (0.10 mL) was degassed by sparging with N2 and was heated to 90 \u00b0C for 10 min. Upon cooling to room temperature, the reaction mixture was diluted with water and extracted with EtOAc. The organic layer was dried over Na2SO4, filtered and concentrated in vacuo. The product was purified via flash column chromatography, eluting with a gradient of 0-100% EtOAc in hexanes to afford the title compound as a TFA salt (18 mg, 72%). LCMS for C29H23F3N5O2 (M+H)+: calculated m/z=530.2; found 530.2.\n\n**Step 6. 2-(4-Cyano-3-ethyl-1-oxo-2-(2-(vinylsulfonamido)phenyl)benzo[4,5]imidazo[1,2-a]pyridin-5(1H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide.** A solution of 2-(2-(2-aminophenyl)-4-cyano-3-ethyl-1-oxobenzo[4,5]imidazo[1,2-a]pyridin-5(1H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide (from Step 5, 6.0 mg, 0.011 mmol) in DCM (0.23 mL) at 0 \u00b0C was treated with triethylamine (3.2 \u03bcL, 0.023 mmol) and ethenesulfonyl chloride (1.7 mg, 0.014 mmol, as a solution in DCM (0.10 mL)) and the reaction was stirred for 30 min. Ammonium hydroxide (28-30%, 0.1 mL) was added and the reaction was stirred for 15 min. Volatiles were removed in vacuo, the residue was reconstituted in CH3CN/MeOH, filtered, and purified via preparative HPLC-MS (pH=2) to afford the title compound (1.7 mg, 20%) as a TFA salt. LCMS for C31H25F3N5O4S (M+H)+: calculated m/z=620.2; found 620.2.\n\n### Example 25. N-(2-(4-Cyano-3-ethyl-1-oxo-5-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-1,5-dihydrobenzo[4,5]imidazo[1,2-a]pyridin-2-yl)phenyl)acrylamide", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\n**Author:** Marc Christian Deller. **Institutions:** Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics, Oxford; Corpus Christi College, Oxford. **Submitted:** A thesis submitted in partial fulfilment of the requirements for the degree of Doctor of Philosophy at The University of Oxford, Michaelmas Term 1999.\n\n## Abstract\n\nCytokines are a family of secreted polypeptide factors that function in controlling the survival, growth, differentiation and effector functions of tissue cells. They achieve this through receptor-mediated signal transduction in which extracellular receptor components of target cells bind circulating ligands, triggering a cascade of events that leads to the activation of various intracellular processes. Commonly, protein kinases are activated which modulate cellular proteins and transcriptional activity.\n\nTwo members of the gp130 cytokine family were studied; namely leukemia inhibitory factor receptor (LIFR) and oncostatin M (OSM). The use of a variety of eukaryotic expression systems was investigated for the production of soluble extracellular fragments of the ligand-binding domain of LIFR. Preliminary characterisation of LIFR was carried out, and its suitability for structural studies assessed.\n\nA truncated variant of OSM was designed, expressed, characterised and crystallised, as wildtype material proved recalcitrant to crystallisation. Multiple wavelength anomalous dispersion techniques were applied and the atomic structure determined in the spacegroup P2\u20812\u20812\u2081 to a resolution of 2.2 \u00c5.\n\nThe structure of OSM was compared with other helical cytokines of the growth hormone family and these comparisons, taken together with mutagenesis data, support a mode of receptor-ligand engagement displaying similarities to that of erythropoietin and granulocyte colony stimulating factor. This study presents information that allows further structural analysis of the epitopes involved in ligand binding by the gp130 and LIFR family of receptors.\n\n## Acknowledgments\n\nFirstly, I would like to thank Professors E. Yvonne Jones and John K. Heath for helpful supervision and giving me the opportunity to work with them. I would also like to thank all members of the CRC Growth Factor Group, Birmingham, and members of the Laboratory of Molecular Biophysics and the Wellcome Trust Centre for Human Genetics, Oxford, for general assistance and making the laboratory a great place to work.\n\nIn Birmingham, I am particularly indebted to Keith Hudson, Katya Chobotova, Ann Vernalis, Dave Staunton and Mark Hall for introducing me to the wonderous techniques of molecular biology and for supplying all necessary cells and vectors. In Oxford, I am particularly grateful to Professor Louise Johnson for giving me the opportunity to work in such a stimulating laboratory environment. Thanks also extend to Shinji Ikemizu and Jeronimo Bravo for introducing me to the marvels of modern crystallography; without whom my exploration into crystallography would have never begun.\n\nI would also like to thank Jon, Geoff, Thil, Malkit, Katsumi, Nathan, Wei Ching, Chris, Andy, Neil, Matt, Kurt, Amit, Ealish, Nick, Bob, Paul, Ed, Pete, Rao, Ren, Martin, Mary, Jos\u00e9 and anyone else who has shown interest in this project. I would like to thank Vivian Strojanoff (ESRF, France), Karl Harlos, Elspeth Garman, Dave Stuart and Dave Stammers for assistance with data collection and Richard Bryan, Kathryn Measures and Robert Esnouf for unsurpassed computing facilities. Many thanks also go to Robin Aplin, Dyson Perrins Laboratory, Oxford for persistence with the mass spec analysis. Financial income was supplied by The Medical Research Council and Louise Johnson, for which I am eternally grateful.\n\nI would like to raise a glass to Leo Sayer, Geordie, Bo', Dr Deeps, Magoo, Taff, Nobby, Ballie, Porker, Mop, Sam, Buck, Dirk and 'Big' Kev for being great mates and helping me through the endless 'Engine Trouble'. A big shout also goes to the boy's of Nottingham Forest F.C., Stella Artois, John Player Specials and the Profanisaurus (unabridged version) for entertainment in the darker hours. Special mention must also go to teachers past, present and future, particularly Alison Parker and Gilbert Gimson for believing in me when all others had lost hope.\n\nFinally, I would like to thank Mum, Dad and Andy for their help, support and encouragement over the years...cheers!\n\n\"...it was like climbing a mountain and falling in love at the same time.\" Max Perutz.\n\nFor Mum and Dad.\n\n## Abbreviations", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\nIn some embodiments, the myeloproliferative neoplasm is selected from the group consisting of chronic myelogenous leukemia, polycythemia vera, primary myelofibrosis, essential thrombocythemia, chronic neutrophilic leukemia, acute myelogenus leukemia, and chronic eosinophilic leukemia, chronic myelomonocytic leukemia, myeloproliferative neoplasm and myelodysplastic syndrome, including myelodysplastic syndrome with refractory anaemia with ring sideroblasts, myelodysplastic syndrome with refractory anemia, and myelodysplastic syndrome with refractory anemia with excess blasts.\n\nIn some embodiments, methods described herein further comprise administering to the human subject an additional therapy selected from the group consisting of a Janus tyrosine kinase (JAK) inhibitor, a phosphoinositide 3-kinase (PI3K) inhibitor, a standard of care therapy, or a combination thereof. In some embodiments, the JAK inhibitor is ruxolitinib and itacitinib. In some embodiments, the PI3K inhibitor is parsaclisib. In some embodiments, the standard of care therapy is selected from the group consisting of IFN-alpha, hydroxyurea, thalidomide, lenalidomide, an androgen, an erythropoietin-stimulating agent, a chemotherapeutic agent, or a combination thereof. In some embodiments, the administration of the antibody or the pharmaceutical composition thereof in combination with the JAK inhibitor produces a synergistic effect. In some embodiments, the JAK inhibitor is ruxolitinib.\n\nAspects of the present disclosure provide a kit comprising the anti-mutCALR antibody, the nucleic acid or the set of nucleic acids, the expression vector or the set of expression vectors, or the isolated cell, and instructions for use in treating a myeloproliferative neoplasm in a human subject in need thereof, optionally with instructions for use in combination with an additional therapy.\n\n## Brief Description of the Drawings\n\n*FIG. 1* includes a graph showing that anti-mutCALR antibodies inhibit STAT5 activation in Ba/F3 cells expressing MPL/mutCALR.\n\n*FIG. 2* includes graphs showing that anti-mutCALR antibodies inhibit mutCALR induced cell proliferation in the Ba/F3 cells stably transfected with MPL/mutCALR.\n\n*FIG. 3* includes graphs showing that anti-mutCALR antibodies are efficient in inhibiting oncogenic cell proliferation in Ba/F3 cells stably transfected with either MPL/mutCALR Type 1 (top) or MPL/mutCALR Type 2 (bottom).\n\n*FIGs. 4A-4G* include graphs showing that anti-mutCALR antibodies inhibit the dimerization of the MPL protein.\n\n*FIG. 5* includes a schematic depiction of the dosing schedule of anti-mutCALR antibodies in mice injected with tumor cells expressing MPL/mutCALR.\n\n*FIG. 6* includes a graph of mouse survival after treatment with anti-mutCALR antibodies.\n\n*FIG. 7* includes a graph of spleen weight in mice after treatment with anti-mutCALR antibodies.\n\n*FIG. 8* includes a graph of number of platelets in blood after treatment of mice with anti-mutCALR antibodies.\n\n*FIG. 9* includes a graph of number of tumor cells in blood after treatment of mice with anti-mutCALR antibodies.\n\n*FIG. 10* includes graphs showing the ability of an anti-mutCALR antibody to potentiate the therapeutic response of ruxolitinib in the inhibition of oncogenic cell proliferation triggered by mutCALR Type 1 (top) or Type 2 (bottom).\n\n*FIG. 11* includes a graph showing that anti-mutCALR antibodies compete with MPL for the binding to mutCALR.\n\n*FIGs. 12A-12C* include structural data of the Fab-mutCALR-peptide complex.\n\n*FIGs. 13A-13C* include structural data of the Fab fragment binding to the first mutCALR peptide in the asymmetric unit of the crystal structure.\n\n*FIGs. 14A-14C* include structural data of the Fab fragment binding to the second mutCALR peptide in the asymmetric unit of the crystal structure.\n\n*FIG. 15* includes an image of the sequence of mutant CALR peptide (SEQ ID NO:328) showing the CalR1 conformation binding residues (top) and CalR2 conformation binding residues (bottom), with the residues having close contact with Fab1 (<4.5 \u00c5) shaded in grey. Note that the CalR1 and CalR2 conformation residues represent two opposite faces of the mutant CALR C-terminal helix, with more residues at the N-terminal that are covered in the CalR1 binding conformation than are covered in the CalR2 binding conformation. Close contact of Fab1 was assessed in MOE (Molecular Operating Environment).\n\n*FIGs. 16A-16C* include structural data of anti-mutCALR antibody Fab fragment bound to 31-mer mutant CalR peptide.\n\n*FIGs. 17A-17C* include structural data from an antigen-antibody interaction analysis.\n\n*FIGs. 18A-18C* include data showing that anti-mutCALR antibody treatment restored normal blood counts, spleen volume, and bone marrow environment in CALR^DEL/DEL engineered mice.\n\n*FIGs. 19A-19D* include data showing that anti-mutCALR antibodies inhibit mutCALR-derived oncogenic functions in CD34+ cells isolated from MPN patients carrying the CALR mutation.\n\n*FIG. 20* includes data showing cell-cycle profiles of BaF3 cells with indicated genotypes in the presence of different concentrations of an anti-mutCALR antibody.\n\n*FIG. 21A* includes a graph showing data that AB1-AB4 and clone 4 bind to mutCALR.\n\n*FIG. 21B* includes a graph showing data that clone 4 inhibits cell proliferation.\n\n## Detailed Description", "# Crystal Structure of a Dye-Decolorizing Peroxidase (DyP) from Bacteroides thetaiotaomicron VPI-5482 at 1.6 \u00c5 Resolution\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* PDB ID: [2GVK](https://www.rcsb.org/structure/2GVK)\n* Release Date: 2006-05-16\n* Method: X-RAY DIFFRACTION 1.6 \u00c5\n* Organisms: Bacteroides thetaiotaomicron VPI-5482\n* Macromolecules: Heme peroxidase (protein)Unique Ligands: ACY, CL, EDO, NA, UNL\n* **Publication:** [Zubieta, C., et al. (2007). Crystal Structures of Two Novel Dye-Decolorizing Peroxidases Reveal a Beta-Barrel Fold with a Conserved Heme-Binding Motif. Proteins, 69, 223-233.](https://doi.org/10.1002/prot.21550)\n\n**Description:** This 1.6 \u00c5 structure of BtDyP from the gut bacterium B. thetaiotaomicron is among the first crystal structures of a dye-decolorizing peroxidase (DyP), revealing a novel two-domain \u03b1+\u03b2 ferredoxin-like fold with a conserved heme-binding motif. The hexameric assembly structure established the structural foundations for this new family of heme-dependent peroxidases with applications in bioremediation and biotechnology.", "# Revolutionary Cryo-EM Breakthrough: Di-Gembodies Transform Small Protein Structure Determination\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-11 on marcdeller.com (categories: blog, sb).*\n\n* [Di-Gembodies](https://www.nature.com/articles/s41589-025-01972-7)\n\n* **Size Barrier Elimination**: Di-Gembodies enable cryo-EM structure determination of proteins as small as 14 kDa, breaking through the traditional ~50 kDa limitation that excluded 75% of human proteins\n* **Parallel Structure Solutions**: Revolutionary capability to simultaneously determine structures of two different proteins within a single cryo-EM experiment, dramatically improving research efficiency\u200b\n* **Covalent Rigidity**: Strategic use of side-chain-to-side-chain chemical bonds eliminates flexibility issues that plagued previous scaffolding approaches, ensuring atomic-level precision\n* **Universal Modularity**: Unlike previous methods requiring extensive re-optimization, Di-Gembodies can be adapted to diverse protein targets without laborious engineering for each application\n* **Therapeutic Target Access**: Opens previously inaccessible small protein targets for structure-based drug discovery, particularly those involved in critical cellular signaling pathways\u200b\n* **Enhanced Throughput**: Modular design and parallel capability significantly accelerate structural biology workflows, enabling high-throughput structure determination for drug discovery applications\u200b\n\n1. **Covalently constrained \u2018Di-Gembodies\u2019 enable parallel structure solutions by cryo-EM:** [Gangshun Yi, et al., Nature Chemical Biology](https://www.nature.com/articles/s41589-025-01972-7)\n2. **Rigid enlargement of sybodies with antibody fragments for cryo-EM analyses of small membrane proteins:** [Fabian Ackle, et al., Nature Communications](https://pmc.ncbi.nlm.nih.gov/articles/PMC11923154/)\n3. **Development of a universal nanobody-binding Fab module for fiducial-assisted cryo-EM studies of membrane proteins:** [Jun-ichi Kishikawa, et al., Proceedings of the National Academy of Sciences](https://www.pnas.org/doi/10.1073/pnas.2115435118)\n4. **Recent advances and current trends in cryo-electron microscopy:** [Margherita Guaita, et al., Current Opinion in Structural Biology](https://pmc.ncbi.nlm.nih.gov/articles/PMC10266358/)\n5. **Cryo-EM structure determination of small proteins by nanobody-directed scaffolding:** [Xiaofeng Chu, et al., Proceedings of the National Academy of Sciences](https://www.pnas.org/doi/10.1073/pnas.2115001118)\n6. **A 3.8 \u00c5 resolution cryo-EM structure of a small protein bound to an engineered scaffold:** [Shih-Lin Liu, et al., Nature Communications](https://www.nature.com/articles/s41467-019-09836-0)\n\n**Covalently Constrained Nanobody Scaffolds Enable Unprecedented Parallel Structure Solutions in Electron Microscopy**\n\nThe landscape of structural biology has been dramatically transformed by a groundbreaking discovery that challenges the fundamental limitations of cryo-electron microscopy (cryo-EM). Researchers from the University of Oxford, the Rosalind Franklin Institute, and Diamond Light Source have developed **Di-Gembodies**\u2014covalently constrained nanobody dimers that enable parallel structure determination of multiple proteins simultaneously. This innovative approach represents a paradigm shift from previous limitations, addressing the longstanding challenge of imaging proteins smaller than 50 kDa with atomic-level precision.\n\nThe significance of this breakthrough extends far beyond technical innovation. Approximately 75% of human protein-coding genes produce proteins within the previously inaccessible size range below 50 kDa, making this advancement crucial for understanding cellular machinery and disease mechanisms. Traditional cryo-EM struggles with small proteins due to their low signal-to-noise ratio, which creates substantial difficulties during particle picking, alignment, and ultimately limits resolution. The Di-Gembodies method fundamentally overcomes these challenges through **strategic molecular engineering** that creates rigid, covalently linked scaffolds.\n\nThe methodology centers on exploiting side-chain-to-side-chain assembly to trap engineered nanobody-to-nanobody interfaces. This covalent dimerization creates **modular constructs** that can simultaneously display two copies of the same protein or two distinct proteins through a subunit interface with sufficient constraint for high-resolution structure determination. The researchers validated this revolutionary approach with multiple targets, including both soluble and membrane proteins down to 14 kDa\u2014establishing hen egg white lysozyme as the smallest protein ever resolved by cryo-EM.\n\nThe **transformative impact** of this technology becomes apparent when considering its practical applications. Unlike previous scaffolding approaches that required extensive re-optimization for each target protein, Di-Gembodies offer remarkable **modularity and universality**. The method allows researchers to study proteins that were previously beyond the reach of cryo-EM, including critical therapeutic targets involved in drug discovery and disease mechanisms. The ability to simultaneously determine structures of two different proteins within a single experiment represents unprecedented efficiency in structural biology workflows.\n\nFrom a methodological perspective, the covalent constraints eliminate the flexibility issues that have plagued previous scaffolding strategies. The rigid linkage between nanobodies and target proteins ensures that favorable imaging advantages of the scaffold are directly transferred to the cargo proteins, enabling atomic-level detail visualization. This approach builds upon earlier innovations in cryo-EM fiducial markers while addressing critical limitations through chemical bond formation rather than relying solely on non-covalent interactions.", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\nBT2081 from Bacteroides thetaiotaomicron (GenBank accession code NP_810994.1) is a member of a novel protein family consisting of over 160 members, most of which are found in the different classes of Bacteroidetes. Genome-context analysis lends support to the involvement of this family in carbohydrate metabolism, which plays a key role in B. thetaiotaomicron as a predominant bacterial symbiont in the human distal gut microbiome. The crystal structure of BT2081 at 2.05\u2005\u00c5 resolution represents the first structure from this new protein family. BT2081 consists of an N-terminal domain, which adopts a \u03b2-sandwich immunoglobulin-like fold, and a larger C-terminal domain with a \u03b2-sandwich jelly-roll fold. Structural analyses reveal that both domains are similar to those found in various carbohydrate-active enzymes. The C-terminal \u03b2-jelly-roll domain contains a potential carbohydrate-binding site that is highly conserved among BT2081 homologs and is situated in the same location as the carbohydrate-binding sites that are found in structurally similar glycoside hydrolases (GHs). However, in BT2081 this site is partially occluded by surrounding loops, which results in a deep solvent-accessible pocket rather than a shallower solvent-exposed cleft.\n\n## Structure of LP2179, the first representative of Pfam family PF08866, suggests a new fold with a role in amino-acid metabolism (2010)\n\n**Authors:** Bakolitsa C, Kumar A, Carlton D, Miller MD, Krishna SS, Abdubek P, Astakhova T, Axelrod HL, Chiu HJ, Clayton T, Deller MC, Duan L, Elsliger MA, Feuerhelm J, Grzechnik SK, Grant JC, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Marciano D, McMullan D, Morse AT, Nigoghossian E, Okach L, Oommachen S, Paulsen J, Reyes R, Rife CL, Tien HJ, Trout CV, van den Bedem H, Weekes D, Xu Q, Hodgson KO, Wooley J, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309109023689. PMID: 20944212.\n\nThe structure of LP2179, a member of the PF08866 (DUF1831) family, suggests a novel \u03b1+\u03b2 fold comprising two \u03b2-sheets packed against a single helix. A remote structural similarity to two other uncharacterized protein families specific to the Bacillus genus (PF08868 and PF08968), as well as to prokaryotic S-adenosylmethionine decarboxylases, is consistent with a role in amino-acid metabolism. Genomic neighborhood analysis of LP2179 supports this functional assignment, which might also then be extended to PF08868 and PF08968.\n\n## Structure of a membrane-attack complex/perforin (MACPF) family protein from the human gut symbiont Bacteroides thetaiotaomicron (2010)\n\n**Authors:** Xu Q, Abdubek P, Astakhova T, Axelrod HL, Bakolitsa C, Cai X, Carlton D, Chen C, Chiu HJ, Clayton T, Das D, Deller MC, Duan L, Ellrott K, Farr CL, Feuerhelm J, Grant JC, Grzechnik A, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Krishna SS, Kumar A, Lam WW, Marciano D, Miller MD, Morse AT, Nigoghossian E, Nopakun A, Okach L, Puckett C, Reyes R, Tien HJ, Trame CB, van den Bedem H, Weekes D, Wooten T, Yeh A, Zhou J, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309110023055. PMID: 20944225.\n\nMembrane-attack complex/perforin (MACPF) proteins are transmembrane pore-forming proteins that are important in both human immunity and the virulence of pathogens. Bacterial MACPFs are found in diverse bacterial species, including most human gut-associated Bacteroides species. The crystal structure of a bacterial MACPF-domain-containing protein BT_3439 (Bth-MACPF) from B. thetaiotaomicron, a predominant member of the mammalian intestinal microbiota, has been determined. Bth-MACPF contains a membrane-attack complex/perforin (MACPF) domain and two novel C-terminal domains that resemble ribonuclease H and interleukin 8, respectively. The entire protein adopts a flat crescent shape, characteristic of other MACPF proteins, that may be important for oligomerization. This Bth-MACPF structure provides new features and insights not observed in two previous MACPF structures. Genomic context analysis infers that Bth-MACPF may be involved in a novel protein-transport or nutrient-uptake system, suggesting an important role for these MACPF proteins, which were likely to have been inherited from eukaryotes via horizontal gene transfer, in the adaptation of commensal bacteria to the host environment.\n\n## Structure of a putative NTP pyrophosphohydrolase: YP_001813558.1 from Exiguobacterium sibiricum 255-15 (2010)\n\n**Authors:** Han GW, Elsliger MA, Yeates TO, Xu Q, Murzin AG, Krishna SS, Jaroszewski L, Abdubek P, Astakhova T, Axelrod HL, Carlton D, Chen C, Chiu HJ, Clayton T, Das D, Deller MC, Duan L, Ernst D, Feuerhelm J, Grant JC, Grzechnik A, Jin KK, Johnson HA, Klock HE, Knuth MW, Kozbial P, Kumar A, Lam WW, Marciano D, McMullan D, Miller MD, Morse AT, Nigoghossian E, Okach L, Reyes R, Rife CL, Sefcovic N, Tien HJ, Trame CB, van den Bedem H, Weekes D, Hodgson KO, Wooley J, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309110025534. PMID: 20944217.", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n**Then, full width, the target's sequence:** a per-residue track coloured by property with the\nPLIP contacts marked, and above it a conservation logo aligned across every distinct protein in\nthe campaign, drawn in bits of information. Hovering a residue names it and lists its contacts and\nconservation; clicking one selects and frames it in both viewers, as does clicking a contact.\n\n**BoltzMaker's own panels sit beneath**, lifted out of the reports it generated and rendered as\nsiblings -- every panel listed under **Inside `boltz_dashboard.html`** above, twenty on a real\ncampaign against the two the explorer draws itself, both reports still downloadable whole. They\nare lifted rather than reimplemented, so they cannot drift from the analysis code, and merged\nrather than framed, so the page does not become a scrolling document inside a card. The whole\nsequence is one ordered list in `reports.PANEL_ORDER`, so reordering the page is a line moved\nrather than a template edit. BoltzMaker's **pIC50 vs confidence score** scatter is the one kept,\nwired up so clicking a point opens that target. On it, the dashed vertical line at **0.5** is the\ngenuine absolute low-confidence cutoff, while the mismatch flags\n(`HIGH_CONFIDENCE_POOR_AFFINITY`, `LOW_CONFIDENCE_STRONG_AFFINITY`) are **not** absolute: they\ncome from splitting *that campaign* into terciles, so they mean \"relative to the other targets\nyou ran\". There is deliberately no horizontal affinity threshold line, because drawing one would\nimply a fixed cutoff the numbers do not support.\n\n**Nothing from the results file executes.** It is user input rendered on this site's own origin,\nso the markup is reduced to a tag allowlist by a tokeniser -- not a regex, which was defeated by a\n`<` inside a quoted attribute -- and every chart is rebuilt from its data: each `Plotly.newPlot`\ncall has its arguments JSON-parsed on the server and handed to the page as values, which the\npage's own code then plots. Parsed and re-serialised, an injected payload is inert text.\n\nTwo things are deliberately dropped: the reports' binding-site panels, because the explorer\nalready gives every target a pose viewer with its interactions beside it and the PyMOL sessions\nthey link to are not in the archive; and the compare-sse charts the dashboard embeds, because the\ncompare-sse page carries the same charts under the same element ids, and rendering both would put\ntwo divs with one id on the page, leaving the second unreachable and silently undrawn.\n\nAn open target is deep-linkable -- the URL carries it, so a link to one target can be shared or\nreloaded. **Sessions do not expire on a clock**: an analysis link keeps working, and uploads are\nremoved only when the space set aside for them fills, least recently opened first. **Download\nsummary CSV** gives the full summary table including every column not shown in the browser.\n\nA note on vocabulary: BoltzMaker calls each `Protein:` block a **family**, and its own reports use\nthat word throughout. A campaign with one receptor and an apo companion therefore has two\nfamilies, which reads oddly, so this page says **protein** for the same thing and counts\n**predictions** rather than families.\n\n### Stepwise Mode\n\nThe same non-GPU stages as four independent tools. Each takes an upload and hands back a download,\nand none depends on the others.\n\n| Tool | You give it | You get back |\n|---|---|---|\n| **Wizard** | Answers to the same plain questions as `BoltzMaker.py new` | A validated, tidied `boltz_input.md` |\n| **Generate** | A `boltz_input.md` (paste or upload) | `boltz_yamls.zip`: the per-target YAMLs plus the manifest |\n| **Preflight** | A `boltz_input.md` | The full check table: SMILES and ligand chemistry, chain-id lengths, duplicate targets, size heuristics. The GPU and model-weight checks report on the server, not your machine, so read those on your own hardware instead. |\n| **Analyze** | A zip of a completed campaign folder (`boltz_input.md` + `boltz_yamls/` + `boltz_output/`, plus any `Apo structure:` files) | The summary CSV/XLSX, the interactive dashboard rendered in the page, and a zip of everything including the CIFs, PLIP output and compare-sse results |\n\n### Runs\n\nThe landing page carries the five most recent runs with a link straight into each one's analysis.\nThe full table is at [**/runs**](https://boltzmaker.mdeller.com/runs): everything not marked\nprivate, in one place -- campaign, when it was prepared, target count, the bundle, and the results\nfile once uploaded. Each row offers the bundle and the `.bmz` for download, and **explore**, which\nre-opens the analysis from the archived results without uploading anything again. The six public\ncampaigns listed under **Examples** above all live here.\n\nThe archive is **capped** -- 3GB and 200 runs -- and prunes oldest-first when it fills, each\nremoval recorded so a missing run is explainable rather than a mystery. The host has ~16GB free\nand also serves three other apps, so an uncapped archive would be a slow-motion disk-full outage.\n\n### How it is built and served\n\n**Source and tested-CLI isolation.** The web app lives in `web/`, and the isolation that matters\nis the process boundary. `BoltzMaker.py` was never designed to be imported (it relaunches itself\ninto a managed venv at module import time), so the Flask app only ever invokes it as a subprocess, with\nan explicit minimal environment and a per-command timeout, through a dedicated trimmed venv\n(`.venv/`, torch/boltz-free, ~500MB -- every torch/boltz reference in the script is a\nfunction-local lazy import) kept separate from the Flask-serving venv (`web/.venv/`).", "# Unlocking Molecular Potential: The Evolution of Topliss Tree and Bioisosteric Strategies in Modern Drug Discovery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Concept | Description | Key References |\n| --- | --- | --- |\n| Topliss Tree Methodology | Systematic decision tree approach for aromatic and aliphatic substituent optimization based on physicochemical parameters | [Topliss, J.G., et al., J. Med. Chem., 1972](https://pubs.acs.org/doi/abs/10.1021/jm00280a002) |\n| Classical Bioisosterism | Replacement of atoms or groups with similar valence electron structure maintaining biological activity | [Bioisostere Wikipedia](https://en.wikipedia.org/wiki/Bioisostere), [Drug Design Bioisosterism](https://www.drugdesign.org/chapters/bioisosterism/) |\n| Non-Classical Bioisosterism | Bioisosteric replacements that maintain function but may differ significantly in structure and electronics | [Meanwell, N.A., et al., PMC, 2014](https://pmc.ncbi.nlm.nih.gov/articles/PMC7416817/) |\n| Structure-Based Drug Design | Rational drug design using 3D protein structures and computational modeling for ligand optimization | [Isomorphic Labs AlphaFold 3](https://www.isomorphiclabs.com/articles/rational-drug-design-with-alphafold-3) |\n| Matched Molecular Pairs | Systematic analysis of compound pairs differing by single structural transformations to guide SAR | [Yang, Z., et al., J. Med. Chem., 2023](https://pubs.acs.org/doi/abs/10.1021/acs.jmedchem.2c01787) |\n| Fragment-Based Drug Discovery | Drug discovery approach starting from low molecular weight fragments that bind weakly to targets | [Li, Q., Front. Mol. Biosci., 2020](https://www.frontiersin.org/journals/molecular-biosciences/articles/10.3389/fmolb.2020.00180/full) |", "# Microplastics and Health: Alarming New Research Reveals Widespread Biological Impact\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThis pioneering study established foundational understanding of microplastic bioaccumulation and toxicity using zebrafish as a model organism. Researchers investigated the uptake patterns of fluorescently-labeled polystyrene microplastics (PS-MPs) of different sizes (70 nm, 5 \u03bcm, 20 \u03bcm) and their subsequent toxic effects in liver tissue. The accumulation experiment revealed size-dependent distribution patterns, with 5 \u03bcm particles accumulating in gills, liver, and gut, while larger 20 \u03bcm particles were restricted to gills and gut tissues. Quantitative analysis showed rapid accumulation within 48 hours, reaching steady-state concentrations with highest levels in liver and gut compared to gills. Toxicity assessment using 70 nm and 5 \u03bcm PS-MPs at concentrations of 20-2000 \u03bcg/L for 3 weeks demonstrated significant hepatotoxic effects including inflammation, lipid accumulation, and oxidative stress. Metabolomic analysis revealed disrupted lipid and energy metabolism with altered fatty acid profiles, decreased branched-chain amino acids, and disturbed ATP/ADP/AMP ratios. The study also showed increased activities of antioxidant enzymes (SOD, CAT), indicating oxidative stress responses. Interestingly, larger 5 \u03bcm particles induced greater oxidative stress than smaller 70 nm particles, suggesting complex size-toxicity relationships. This research established important methodological approaches for microplastic research and provided early evidence of systemic metabolic disruption.", "# Protein Hunter: Pioneering the Next Frontier of AI-Driven Protein Design\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-15 on marcdeller.com (categories: blog, sb).*\n\n* [Protein Hunter: exploiting structure hallucination within diffusion for protein design](https://www.biorxiv.org/content/biorxiv/early/2025/10/10/2025.10.10.681530.full.pdf)\n\n* **Revolutionary approach**: [Protein Hunter exploits structure hallucination in diffusion models](https://www.biorxiv.org/content/10.1101/2025.10.10.681530v1) to generate novel proteins starting from undefined sequences\n* **Exceptional efficiency**: Achieves high computational success rates without requiring fine-tuning or extensive optimization cycles\n* **Broad applicability**: [Supports diverse design tasks including binder generation for proteins, nucleic acids, and small molecules](https://www.biorxiv.org/content/10.1101/2025.10.10.681530v2.full.pdf)\n* **Computational accessibility**: Operates as a lightweight framework that doesn\u2019t demand specialized training datasets or infrastructure\n* **Validated performance**: Demonstrates superior [AlphaFold3 in silico success rates](https://www.biorxiv.org/content/10.1101/2025.10.10.681530v1)compared to traditional design methods\n* **Practical implementation**: Enables rapid iteration between structure prediction and sequence optimization for real-world applications\n\n1. **Protein Hunter: exploiting structure hallucination within diffusion for protein design:** [Wu et al., bioRxiv](https://www.biorxiv.org/content/10.1101/2025.10.10.681530v1)\n2. **De novo design of protein structure and function with RFdiffusion:** [Watson et al., Nature](https://www.nature.com/articles/s41586-023-06415-8)\n3. **De novo protein design by deep network hallucination:** [Anishchenko et al., Nature](https://www.nature.com/articles/s41586-021-04184-w)\n4. **Hallucinating symmetric protein assemblies:** [Bennett et al., Science](https://www.science.org/doi/10.1126/science.add1964)\n5. **Protein structure generation via folding diffusion:** [Wu et al., Nature Communications](https://www.nature.com/articles/s41467-024-45051-2)\n6. **Accurate structure prediction of biomolecular interactions with AlphaFold 3:** [Abramson et al., Nature](https://www.nature.com/articles/s41586-024-07487-w)\n\n**Unleashing the Creative Power of Diffusion Models to Hallucinate Novel Protein Architectures from the Ground Up**\n\nThe landscape of protein design has been fundamentally transformed by a groundbreaking approach that harnesses the creative potential of artificial intelligence. [Protein Hunter represents a paradigm shift](https://www.biorxiv.org/content/10.1101/2025.10.10.681530v1) in how we conceptualize and execute de novo protein design, leveraging the phenomenon of structure hallucination within diffusion models to generate entirely novel proteins with unprecedented efficiency and accuracy.\n\nUnlike traditional protein design methods that rely on extensive computational resources and time-intensive optimization processes, Protein Hunter begins with nothing more than a sequence of unknown amino acids (all-X sequence) and allows [diffusion-based structure prediction models to hallucinate reasonable-looking structures](https://www.biorxiv.org/content/10.1101/2025.10.10.681530v1) that can be iteratively refined through sequence redesign and structure re-prediction cycles. This approach represents a fundamental departure from conventional wisdom, demonstrating that the same neural networks trained to predict existing protein structures can be repurposed to imagine entirely new ones.\n\nThe elegance of this method lies in its computational efficiency and broad applicability. Rather than requiring tens of thousands of candidate molecules to be tested before finding a single functional design, [Protein Hunter achieves high AlphaFold3 in silico success rates](https://www.biorxiv.org/content/10.1101/2025.10.10.681530v1)across diverse generation tasks. The framework supports unconditional protein design, where novel structures emerge without constraints, as well as conditional generation targeting specific molecular interactions including proteins, cyclic peptides, small molecules, DNA, and RNA. This versatility extends to sophisticated applications such as multi-motif scaffolding and partial protein redesign, positioning it as a comprehensive platform for addressing the full spectrum of protein design challenges.\n\nThe underlying principle exploits a remarkable property of modern structure prediction networks: their ability to generate plausible structural hypotheses even when presented with artificial or incomplete sequence information. This \u201challucination\u201d phenomenon, initially considered a limitation of these models, has been transformed into a powerful design tool. The approach mirrors [recent advances in diffusion-based protein design](https://www.nature.com/articles/s41586-023-06415-8) but distinguishes itself through its fine-tuning-free methodology and rapid execution speed.\n\nWhat sets Protein Hunter apart is its accessibility and practical implementation. The framework operates without requiring specialized training datasets or extensive computational infrastructure typically associated with protein design workflows. This democratization of protein design technology has profound implications for research laboratories worldwide, potentially accelerating discoveries in therapeutic development, enzyme engineering, and biomaterial creation. The method\u2019s success in generating diverse molecular binders demonstrates its immediate applicability to drug discovery pipelines, where the ability to rapidly design protein-based therapeutics could transform treatment development timelines.", "# Crystal structure of a soluble cleaved HIV-1 envelope trimer\nAuthors: Julien, J.-P.; Cupo, A.; Sok, D.; Stanfield, R.L.; Lyumkis, D.; Deller, M.C.; Klasse, P.-J.; Burton, D.R.; Sanders, R.W.; Moore, J.P.; Ward, A.B.; Wilson, I.A.\nSource: Science, 2013\n\n### Fab expression and purification\n\nRecombinant Fabs for co-crystallization experiments were expressed using a protocol similar to that previously described. Briefly, the heavy and light chain genes were co-transfected using 293Fectin (Invitrogen) into human embryonic kidney (HEK) 293F cells in a 2:1 ratio to minimize formation of light chain homodimers. Secreted Fabs were isolated from the expression media 6-days post-transfection by affinity purification via an anti-human \u03bb/\u03ba light chain affinity matrix (CaptureSelect Fab \u03bb/\u03ba; BAC), followed by cation exchange and gel filtration chromatography.\n\n### SOSIP gp140 expression and purification\n\nTo increase the probability of successful crystallization and structure determination, several soluble, cleaved trimers of the SOSIP.664 gp140 design were expressed and purified as described elsewhere. Crystallization candidates based on env genes from HIV-1 isolates KNH1144, ADA, 1182.6.1D2, BG505 and 208.9.C10 were selected from an initial panel of 20, based on their expression and trimer-formation properties, as assessed by SDS-PAGE and BN-PAGE. Notably, for isolates that did not naturally contain residue N332, a point substitution was made to introduce it, thereby creating the 2G12 epitope to facilitate affinity purification. The Env constructs were co-transfected with the furin protease in HEK 293S GnTI\u2212/\u2212 cells, which lack N-acetylglucosaminyltransferase I and, therefore, produce glycoproteins bearing only high-mannose (Man5-9) glycans. Secreted SOSIP.664 Env proteins were harvested from supernatants and affinity purified using a 2G12 MAb affinity column. Following a high salt elution process, trimers were purified to size homogeneity using a Superose 6 size exclusion chromatography (SEC) matrix (GE Healthcare).\n\n### Sample preparation and crystal screening\n\nPurified SOSIP.664 gp140 trimers were tested for crystallizability either alone or as complexes with one or more ligands from the following list: soluble CD4, Fabs: PGT121, PGT122, PGT123, PGT127, PGT128, PGT135, PG9, PG16, PGT145, PGV04, 17b. In addition, purified trimers and complexes were in some cases treated with the following glycosidases, alone or in combination, for 3 h at temperatures ranging between 20-37 \u00b0C: EndoH (New England Biolabs), EndoD (VWR), Endo F1 (EMD Chemicals) or \u03b11,2,3,6 mannosidase (QA-Bio). The monodispersity of the resulting samples was measured by SDS-PAGE, BN-PAGE and SEC coupled in-line with multi-angle light scattering (SEC-MALS) using the following layout: a Superose 6 10/30 SEC column (GE Healthcare) operated on an AKTA Avant FPLC system (GE Healthcare) with the following calibrated detectors: 1) HP1 1050 Hewlett-Packard UV detector (Norwalk, CT); 2) MiniDawn Treos multi-angle light scattering (MALS) detector (Wyatt Corporation, CA); 3) quasi-elastic light scattering (QELS) detector (Wyatt Corporation, CA); and 4) Optilab T-reX refractive index (RI) detector (Wyatt Corporation, CA). The purified complexes were concentrated to approximately 2-6 mg/ml and subjected to extensive crystallization trials using either: 1) the Oryx8 crystallization robot (Douglas Instruments) with 96 different crystallization conditions at 16 \u00b0C; or 2) the automated IAVI/JCSG/TRSI CrystalMation robotic system (Rigaku) at the Joint Center for Structural Genomics (www.jcsg.org), where 384 different crystallization conditions were tested at 4 \u00b0C and 22 \u00b0C.\n\nIn the above trials, only the partially deglycosylated SOSIP.664 gp140 trimers from various isolates in complex with antibodies of the PGT121-family (PGT121, PGT122 and PGT123) resulted in crystal hits. Surprisingly, for these complexes, hits were obtained from approximately 10% of the crystallization conditions, but the majority of crystals tested (>600) only diffracted to 8 \u00c5 or worse at various synchrotron sources (APS, SSRL, and CLS). Several pre- and post-crystallization strategies were explored in an effort to improve the x-ray diffraction properties of these crystals: however, the use of single-chain Fv instead of Fab, limited in-situ proteolysis, varying temperature of crystal growth, additives screening, crystal cross-linking, dehydration, and annealing all met with limited success. In addition, we attempted use of a SOSIP construct further truncated to gp41 position 650. Although crystals could be obtained with the trimer in complex with PGT122 Fab, these did not show improved diffraction properties compared to the majority of SOSIP.664-containing crystals.\n\n### X-ray data collection, structure solution and model building\n\nUltimately, only the thermostable BG505 SOSIP.664 trimers allowed us to achieve higher resolution x-ray diffraction. Two crystals obtained from crystallization condition 0.1 M CAPS, 2 M ammonium sulfate, 0.2 M lithium sulfate, pH 10.5, with the EndoH-treated, SEC-purified PGT122 Fab plus BG505 SOSIP.664 trimer complex showed a maximum visible diffraction of up to 3.7 \u00c5 along the c-axis at the Advanced Photon Source (APS) beamline 23-ID-D (fig. S1). However, these crystals showed severe anisotropic diffraction and diffracted to a maximum of only ~5 \u00c5 and ~6 \u00c5 along the a- and b-axes, respectively. In addition, due to radiation damage during data collection, it was necessary to merge the data from two crystals to obtain a dataset with sufficient redundancy and completeness. Integration and scaling were performed using the program XDS, and output files were subsequently corrected for anisotropy using the UCLA MBI \u2014 Diffraction Anisotropy Server (http://services.mbi.ucla.edu/anisoscale/anisoscale_xds/), with truncations at 4.5 \u00c5, 5.5 \u00c5 and 4.0 \u00c5 along the a, b and c axes, respectively. The overall data were then truncated to 4.7 \u00c5 based on an I/\u03c3 >2 criteria in the highest resolution shell. Only few higher resolution reflections to 3.7 \u00c5 met these I/\u03c3 >2 criteria and their inclusion in the dataset did not subsequently improve the quality of maps. Data statistics are reported in table S1.", "# GitHub README: FoldForge\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/FoldForge, 2026\n\n> **A single-file browser GUI for the whole Boltz-2 workflow: build, run, review.**\n\n    \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/FoldForge\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/FoldForge</a></td>\n</tr>\n</table>\n\n---\n\nForgeFold is a browser-based interface for preparing, running, and reviewing Boltz-2 structure-prediction jobs. It brings Boltz setup, YAML construction, job-script generation, and result inspection into a single lightweight HTML application. \n\nWhy it matters: the Boltz-2 workflow normally means hopping between a text editor for YAML, a shell for the run script, and a spreadsheet for the results, with plenty of room for format mistakes along the way. ForgeFold pulls all four stages (set the binary path, build the input YAML from sequences and SMILES, generate the run script, and review confidence and affinity metrics) into one lightweight HTML application that opens in any browser with no build step. It is useful for structure-guided teams who want to lower the friction of repeated co-folding runs and give less command-line-comfortable colleagues a clean, guided path from input preparation to result triage.\n\n## Overview\n\nThe application is organized around a simple four-step workflow: set the local Boltz-2 binary path, build an input YAML from protein and ligand definitions, configure a run script, and review output metrics from analyzed results. \n\nThis design is meant to reduce friction for structure-guided workflows where users repeatedly move between FASTA/SMILES input preparation, command-line execution, and affinity-oriented output review. \n\n## Core workflow\n\n1. **Setup installation** \u2014 store the path to the local Boltz-2 executable so it can be reused across sessions. \n2. **Build YAML input** \u2014 enter protein FASTA sequences and ligand SMILES, then export an `input.yaml` file for Boltz-compatible predictions. \n3. **Configure and write jobs** \u2014 choose run parameters and generate a `run_boltz.zsh` shell script for local execution. \n4. **Analyse results** \u2014 load processed outputs such as `boltz_summary.csv` to inspect confidence and affinity-related metrics. \n\n## Features\n\n- Browser-based GUI for Boltz-2 workflow management. \n- YAML builder for protein and ligand input definition. \n- Job-script generation for local prediction runs. \n- Results view for summary CSV uploads and prediction review. \n- Support for metrics including `confidence_score`, `affinity_pred_value1`, and derived `pIC50` display. \n\n## Input model\n\nForgeFold is built around Boltz-style YAML inputs that define one or more protein chains and optional ligands. The example interface content shows protein entries with `id` and `sequence`, ligand entries with `id` and `smiles`, plus optional constraint and property blocks for pocket and affinity settings. \n\nA typical workflow is to paste receptor and partner sequences, add a ligand SMILES string, and export the resulting YAML before launching prediction jobs locally. \n\n## Results interpretation\n\nThe interface is designed to surface key prediction outputs after external analysis. The bundled UI text highlights `confidence_score` as an overall model-confidence measure, `affinity_pred_value1` as an affinity-related score, `affinity_probability_binary1` as a binder-likelihood output, and `pIC50` as a derived potency-style estimate. \n\nThese values are intended for rapid triage and comparison of modeled complexes rather than as a substitute for experimental validation. \n\n## Running ForgeFold\n\nBecause ForgeFold is delivered as a standalone HTML application, the simplest way to use it is to open the file in a modern browser and interact with the GUI directly. \n\nTypical usage:\n\n```bash\n# Example: open the app locally\nopen ForgeFold.html\n```\n\nOr host it with any lightweight static server:\n\n```bash\npython -m http.server 8000\n```\n\nThen open `http://localhost:8000` in a browser.\n\n## Expected companion tools\n\nForgeFold is designed to sit alongside a local Boltz-2 installation and a downstream analysis script such as `analyze_boltz.py`, which generates CSV outputs for the Results tab. \n\nTo make the full workflow work smoothly, keep the following available:\n\n- A working local `boltz` executable. \n- Input biomolecular sequences and ligand SMILES strings. \n- A results-processing script that emits `boltz_summary.csv` or compatible output. \n\n## Suggested repository structure\n\n```text\nForgeFold/\n\u251c\u2500\u2500 README.md\n\u251c\u2500\u2500 ForgeFold.html\n\u251c\u2500\u2500 analyze_boltz.py\n\u251c\u2500\u2500 examples/\n\u2502   \u2514\u2500\u2500 input.yaml\n\u2514\u2500\u2500 outputs/\n```\n\n## Roadmap ideas\n\nPotential future improvements suggested by the current UI structure include more robust YAML validation, richer job templating, tighter integration with local output directories, and more detailed downstream result visualizations. \n\n## License\n\nAdd the project license here.\n\n## Source context\n\nThis README was drafted from the attached ForgeFold application content and interface text.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/FoldForge\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/FoldForge</a></td>\n</tr>\n</table>", "# Self-Adapting Language Models: Are We Creating Our Final Invention?\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nYet the journey from self-adaptation to true recursive self-improvement confronts substantial obstacles. Current systems like SEAL demonstrate \u201cmeta-learning\u201d where models learn how to learn more effectively, but they operate within carefully constrained environments with human-designed reward signals. The [Absolute Zero Reasoner](https://www.theaugmentededucator.com/p/when-ai-teaches-itself-the-breakthrough) and similar systems that learn without external data represent fascinating developments, but they still require humans to define the learning framework. More critically, these models struggle with [catastrophic forgetting](https://arxiv.org/html/2506.10943v2), where new learning overwrites previous knowledge. When SEAL processes multiple sequential updates, performance on earlier tasks gradually degrades, revealing that true continual learning remains elusive. The [computational overhead](https://arxiv.org/html/2506.10943v2) of self-improvement also presents practical limitations, as evaluating each potential modification requires extensive resources.\u200b\n\nThe question of whether self-improving AI constitutes humanity\u2019s final invention hinges on control and alignment. An AI system capable of recursive self-improvement could theoretically redesign its own goal structure, potentially diverging from human values in [unpredictable ways](https://saidar.ai/blog/the-self-evolving-machine). This raises the alignment problem: how do we ensure that increasingly autonomous systems remain aligned with human interests as they evolve beyond our complete understanding? Some researchers argue that capable AI systems might actually [fear self-improvement](https://www.alignmentforum.org/posts/axKWaxjc2CHH5gGyN/ai-will-not-want-to-self-improve), recognizing the existential risks that more powerful successors pose to their current goal structures. This counterintuitive possibility suggests the emergence order of three critical capabilities matters enormously: the ability to self-improve, the ability to recognize risks from improvement, and the ability to solve the alignment problem.\u200b\n\nThe path forward requires balancing innovation with caution. Self-adapting models address the looming \u201c[data wall](https://arxiv.org/html/2506.10943v2)\u201d where frontier language models will exhaust all publicly available human-generated text by 2028. By generating high-quality synthetic training data, these systems offer a solution to continued progress when traditional data sources are depleted. Applications already span [scientific discovery](https://deepfa.ir/en/blog/self-improving-ai-models), healthcare diagnostics, financial optimization, and personalized education, demonstrating practical benefits. However, responsible development demands transparent evaluation frameworks, robust safety mechanisms that cannot be optimized away, and international coordination on governance standards. The future may not require choosing between human creativity and machine capability, but rather architecting systems where both coexist productively.\u200b\n\nWhether self-adapting AI becomes humanity\u2019s final invention depends less on technological feasibility and more on our collective wisdom in guiding its development. The capability for machines to improve themselves already exists in nascent form. What remains uncertain is whether this technology will augment human potential or fundamentally transform the human role in shaping our future. The answer will emerge not from the algorithms themselves, but from the choices we make about their design, deployment, and the ethical frameworks we establish today. We stand not at an inevitable endpoint, but at a critical juncture where our decisions will determine whether AI self-improvement amplifies human flourishing or marks a transition we may not fully control.", "# Engineering the Future of Proteins: Unveiling the Latest AI Tools Transforming Structural Biology in 2025\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-20 on marcdeller.com (categories: sb).*\n\n* AlphaFold 3 refines bound-state predictions for drug discovery by modeling protein\u2013ligand interactions at sub-angstrom accuracy.\n* OpenFold 3 offers an open-source, public-data implementation of AlphaFold 3 for transparent benchmarking and academic use.\n* RFdiffusion leverages diffusion processes on RoseTTAFold to generate novel protein backbones and symmetric assemblies de novo.\n* ProteinGenerator applies sequence-space diffusion to co-generate sequences and structures with multistate and functional conditioning.\n* ESM-2 and ESMFold bypass MSAs to predict structures and functions directly from raw sequences, excelling with orphan and low-homology proteins.\n* RAG-ESM enhances ESM-2 with retrieval-augmented conditioning on homologous sequences for superior masked and conditional generation.\n* ProteinMPNN designs amino acid sequences for fixed backbones in seconds, supporting immune-evasive tuning via direct preference optimization.\n* LigandMPNN incorporates ligand graphs and element types to achieve up to 100-fold improvements in small-molecule and DNA binding accuracy.\n\n1. **AlphaFold 3:** Jumper, J., Evans, R., Pritzel, A., et al. \u201cHighly Accurate Protein Structure Prediction with AlphaFold.\u201d Nature 596, 583\u2013589 (2021):\n   <https://doi.org/10.1038/s41586-021-03819-2>\n2. **OpenFold 3:** Ahdritz, G., Bouatta, N., Bhowmik, D., et al. \u201cOpenFold: Accelerating the Diffusion of AlphaFold into Open Source.\u201d bioRxiv 2022.04.04.486479 (2022):\n   <https://doi.org/10.1101/2022.04.04.486479>\n3. **RoseTTAFold Diffusion:** Yang, J., Anishchenko, I., Park, H., et al. \u201cImproved Protein Structure Prediction Using Potentials from Deep Learning.\u201d Nature 600, 547\u2013552 (2021):\n   <https://doi.org/10.1038/s41586-021-04019-7>\n4. **ESM-2 and ESMFold:** Lin, Z., Akin, H., Rao, R., et al. \u201cLanguage Models of Protein Sequences at the Scale of Evolution Enable Accurate Structure Prediction.\u201d bioRxiv 2022.12.15.520557 (2022):\n   <https://doi.org/10.1101/2022.12.15.520557>\n5. **ProteinMPNN:** Anand, A., & Huang, P. \u201cGenerative Modeling for Protein Structures.\u201d Advances in Neural Information Processing Systems 34, 15508\u201315519 (2021):\n   <https://proceedings.neurips.cc/paper/2021/file/ccd38af51c3ab146f5bc94cf04a1b28c-Paper.pdf>\n\nIn the rapidly evolving field of structural biology, 2025 has witnessed the release of an unprecedented suite of AI-driven protein engineering tools that are redefining our ability to predict, design, and optimize proteins with atomistic precision. From advanced structure predictors to generative sequence designers and diffusion-based backbone sculptors, researchers now possess a versatile toolkit that accelerates discovery pipelines, enhances therapeutic development, and opens new frontiers in enzyme design, biomaterials, and beyond. This post explores the most impactful tools introduced or significantly updated in 2025, highlights their unique capabilities, and showcases how they integrate to form end-to-end protein design platforms.\n\n**AlphaFold 3**\nBuilding on the Nobel-prize\u2013winning AlphaFold2 architecture, AlphaFold 3 extends predictive scope to protein\u2013ligand and protein\u2013protein complexes, delivering sub-angstrom accuracy for bound states that underpin drug discovery workflows. Leveraging proprietary vaults of small-molecule-bound structures through a secure consortium, AlphaFold 3 advances realistic modeling of drug interactions within minutes.\n\n**OpenFold 3**\nAn open-source replica of AlphaFold 3, OpenFold 3 democratizes access to high-fidelity prediction by reproducing the DeepMind implementation using only public data. Released in April 2025, it powers academic structural biology and enables transparent benchmarking against proprietary variants.\n\n**RoseTTAFold Diffusion (RFdiffusion)**\nThis generative framework repurposes the RoseTTAFold network into a diffusion model, iteratively denoising random coordinate inputs to generate novel backbones and symmetric assemblies. RFdiffusion excels at de novo binder design, enzyme scaffolding, and higher-order oligomer construction, producing structures validated by cryo-EM.\n\n**ProteinGenerator (PG)**\nA sequence-space diffusion model built on RoseTTAFold, PG directly generates paired sequences and structures through noise-to-ground-truth denoising. It supports multifaceted conditioning\u2014secondary structure motifs, functional hotspots, and sequence patterns\u2014enabling precise multistate design and functional tuning.\n\n**ESM-2 and ESMFold**\nThe latest Evolutionary Scale Modeling (ESM-2) protein language model boasts 15 billion parameters and delivers sequence-only predictions of structure, dynamics, and functional classification. Paired with ESMFold, it bypasses MSAs for rapid, accurate atomic-level prediction, particularly valuable for orphan proteins and sparse homology scenarios.\n\n**RAG-ESM**\nIntroducing retrieval-augmented generation to protein language models, RAG-ESM conditions ESM-2 on homologous sequences to improve masked prediction, conditional sequence generation, and motif scaffolding. This hybrid approach significantly boosts performance on low-data engineering tasks.\n\n**ProteinMPNN**\nA message-passing neural network that converts fixed backbone coordinates into optimized amino acid sequences, ProteinMPNN remains the gold standard for rapid sequence design. Its latest NVIDIA NIM integration accelerates large-scale pipelines and supports immune-evasive design variants via direct preference optimization.\n\n**LigandMPNN**\nExtending ProteinMPNN to protein\u2013ligand interfaces, LigandMPNN encodes atomic element types, spatial context, and side-chain packing to craft sequences that bind small molecules, DNA, metals, and noncanonical ligands with remarkable affinity improvements validated by X-ray crystallography.\n\nCollectively, these tools form modular pipelines: generate a backbone with RFdiffusion or PG, predict its structure with AlphaFold 3 or ESMFold, refine sequences with ProteinMPNN or LigandMPNN, and iterate in closed-loop automation using RAG-ESM and high-throughput biofoundries. The resulting workflows compress months of engineering into days and democratize complex design challenges across academia and industry.", "# Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\nAuthors: Hummel, J.R.; Xiao, K.-J.; Yang, J.C.; Epling, L.B.; Mukai, K.; Ye, Q.; Xu, M.; Qian, D.; Huo, L.; Weber, M.; Roman, V.; Lo, Y.; Drake, K.; Stump, K.; Covington, M.; Kapilashrami, K.; Zhang, G.; Ye, M.; Diamond, S.; Yeleswaram, S.; Macarron, R.; Deller, M.C.; Wee, S.; Kim, S.; Wang, X.; Wu, L.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2024\n\n**Step 3. 1-(4-(2-((1-(Cyclopropylsulfonyl)piperidin-4-yl)amino)-5-(trifluoromethyl)pyrimidin-4-yl)-1H-pyrazol-1-yl)-2-methylpropan-2-ol (17).** To a mixture of 2-methyl-1-(4-(2-(piperidin-4-ylamino)-5-(trifluoromethyl)pyrimidin-4-yl)-1H-pyrazol-1-yl)propan-2-ol hydrochloride (10.6 g, 25.2 mmol) and N-ethyl-N-isopropylpropan-2-amine (11.0 mL, 63 mmol) in CH3CN (100 mL) was added cyclopropanesulfonyl chloride (4.60 g, 32.7 mmol) and the reaction mixture was stirred at r.t. for 20 min. The mixture was diluted with H2O (100 mL), quenched with saturated aqueous NaHCO3, and diluted with additional H2O (200 mL). The resulting white precipitate was filtered and washed sequentially with H2O and H2O/CH3CN (5:1). The solid was dried under vacuum to afford the title compound (10.7 g, 87% yield, >99.5% purity) as a white solid. LCMS calculated for C20H28F3N6O3S (M + H)+: m/z = 489.2; found, 489.2.\n\n**Step 1. tert-Butyl 4-((5-cyano-4-(1-(2-hydroxy-2-methylpropyl)-1H-pyrazol-4-yl)pyrimidin-2-yl)amino)piperidine-1-carboxylate (40).** A mixture of tert-butyl 4-((4-chloro-5-cyanopyrimidin-2-yl)amino)piperidine-1-carboxylate (31, 86.1 mg, 0.255 mmol), 2-methyl-1-(4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazol-1-yl)propan-2-ol (38, 102 mg, 0.382 mmol), Pd(dppf)Cl2\u00b7CH2Cl2 (41.6 mg, 0.051 mmol), and Na2CO3 (67.5 mg, 0.637 mmol) in 1,4-dioxane (1.06 mL) and H2O (0.21 mL) was stirred at 100 \u00b0C for 16 h. After cooling to r.t., the crude mixture was diluted with H2O and extracted with CH2Cl2. The combined organic phases were dried over MgSO4, concentrated, and purified by flash chromatography (12 g SiO2, EtOAc/hexanes) to afford the title compound (75.2 mg, 67% yield) as a white solid. LCMS calculated for C22H32N7O3 (M + H)+: m/z = 442.3; found, 442.3.\n\n**Step 2. 4-(1-(2-Hydroxy-2-methylpropyl)-1H-pyrazol-4-yl)-2-(piperidin-4-ylamino)pyrimidine-5-carbonitrile hydrochloride (42).** To a mixture of 40 (75.2 mg, 0.170 mmol) in THF (1 mL) was added HCl (4 M in 1,4-dioxane, 0.5 mL, 2 mmol) and the reaction mixture was stirred at 60 \u00b0C for 4 h. After cooling to r.t., the mixture was sparged with nitrogen for 5 min before Et2O (1 mL) was added and the solid precipitate was collected via filtration, washed with Et2O (1 mL), and dried under vacuum to afford the title compound (58 mg, 90% yield) as a white solid. LCMS calculated for C17H24N7O (M + H)+: m/z = 342.2; found, 342.2.\n\n**Step 3. 2-((1-(Cyclopropylsulfonyl)piperidin-4-yl)amino)-4-(1-(2-hydroxy-2-methylpropyl)-1H-pyrazol-4-yl)pyrimidine-5-carbonitrile (18).** To a mixture of 4-(1-(2-hydroxy-2-methylpropyl)-1H-pyrazol-4-yl)-2-(piperidin-4-ylamino)pyrimidine-5-carbonitrile hydrochloride (58 mg, 0.15 mmol) in CH3CN (1 mL) was added N-ethyl-N-isopropylpropan-2-amine (0.1 mL, 0.6 mmol) followed by cyclopropanesulfonyl chloride (32 mg, 0.23 mmol), and the reaction mixture was stirred at r.t. for 2 h. The mixture was diluted with CH3CN, H2O, and TFA (0.1 mL) and purified by prep HPLC (pH 2) to afford the title compound (38.1 mg, 56% yield) as a white solid. LCMS calculated for C20H28N7O3S (M + H)+: m/z = 446.2; found, 446.2.\n\n### Structural biology and supporting studies\n\nExperimental details for the in vitro and in vivo studies, 1H NMR spectra for compounds 2\u221218 and synthetic intermediates, 2D NMR spectra and HPLC traces for compounds 12, 14, and 17, kinase profiles of selected compounds, and the coordinates and experimental methods for the crystal structure of compound 17 bound to CDK2 are provided in the Supporting Information, together with molecular formula strings and in vitro biological data.\n\n**Accession Codes.** The crystal structure of compound 17 bound to CDK2 has been deposited in the Protein Data Bank (PDB: 8UV0). Authors will release the atomic coordinates and experimental data upon article publication.\n\n## Author Information\n\n**Corresponding Author.** Joshua R. Hummel \u2212 Incyte Research Institute, Incyte Corporation, Wilmington, Delaware 19803, United States; johummel@incyte.com.\n\n**Authors.** Kai-Jiong Xiao, Jeffrey C. Yang, Leslie B. Epling, Ken Mukai, Qinda Ye, Meizhong Xu, Dingquan Qian, Lu Huo, Michael Weber, Valerie Roman, Yvonne Lo, Katherine Drake, Kristine Stump, Maryanne Covington, Kanishk Kapilashrami, Guofeng Zhang, Min Ye, Sharon Diamond, Swamy Yeleswaram, Ricardo Macarron, Marc C. Deller, Susan Wee, Sunkyu Kim, Xiaozhao Wang, Liangxing Wu, and Wenqing Yao \u2014 all of Incyte Research Institute, Incyte Corporation, Wilmington, Delaware 19803, United States.\n\n**Author Contributions.** The manuscript was prepared by J.R.H. with contributions from all authors. J.R.H., K.X., J.C.Y., and M.X. designed and synthesized the compounds.\n\n**Funding.** This research was funded by Incyte Corporation.\n\n**Notes.** The authors declare the following competing financial interest(s): All authors are current or former employees of Incyte Corporation.\n\n## Acknowledgments\n\nLaura Kaldon, Michelle Conlin, Ruth Young-Sciame, Susan Petusky, Stephanie Wezalis, Xin He, Kamna Katiyar, Hong Chang, Amy Hehman, Robert Landman, and Gengjie Yang are gratefully acknowledged for PK studies, ADME, and biological and assay support. We would like to thank Karl Blom, Yingrui Dai, Robby Kirk, Vince Caruso, Ronald Magboo, Jim Doughty, and Min Li for analytical assistance. We thank Laurine Galya, Scott Leonard, James Hall, and Sean Bowen for assistance with obtaining and analyzing NMR. We thank Onur Atasoylu, Ravi Jalluri, and Cheng-Tsung Lai for help with molecular modeling.\n\n## Abbreviations", "# Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\nAuthors: Zhuo, J.; Li, Y.-L.; Qian, D.-Q.; Burns, D.M.; Mei, S.; Rafalski, M.; Xu, M.; Cao, G.; Pan, Y.; Jia, Z.; Jalluri, R.; Epling, L.B.; Fenalti, G.; Deller, M.C.; Procak, J.; Covington, M.; He, X.; Collins, R.; Stubbs, M.; Burke, K.; Oliver, J.; Margulis, A.; Boer, J.; Wynn, R.; Scherle, P.; Diamond, S.; Newton, R.; Zhang, Y.; Metcalf, B.; Yao, W.\nSource: Journal of Medicinal Chemistry (Drug Annotation), 2026\n\n**Authors:** Jincong Zhuo, Yun-Long Li, Ding-Quan Qian, David M. Burns (corresponding author), Song Mei, Maria Rafalski, Meizhong Xu, Ganfeng Cao, Yongchun Pan, Zhongjiang Jia, Ravi Jalluri, Leslie B. Epling, Gustavo Fenalti, Marc C. Deller, Jessica Procak, Maryanne Covington, Xin He, Robert Collins, Matthew Stubbs, Krista Burke, Julian Oliver, Alex Margulis, Jason Boer, Richard Wynn, Peggy Scherle, Sharon Diamond, Robert Newton, Yan Zhang, Brian Metcalf, and Wenqing Yao. **Affiliations (condensed):** Incyte Research Institute, Wilmington, Delaware 19803, United States. **Published:** *Journal of Medicinal Chemistry*, 2026 (Drug Annotation). Received December 19, 2025; Revised April 16, 2026; Accepted April 28, 2026. DOI 10.1021/acs.jmedchem.5c03753.\n\n## Abstract\n\nAberrant JAK2 signaling is a key driver in cancers such as myeloproliferative neoplasms (MPNs), whereas JAK1 acts as a central mediator of autoimmune and inflammatory diseases. Since JAK2 supports blood cell production, its inhibition can lead to anemia and other cytopenias, making selective JAK1 inhibition a potentially safer therapeutic strategy opposed to dual JAK1/JAK2 inhibition. We initially identified novel scaffold 9, which selectively inhibited JAK1 over JAK2. Iterative breakthroughs would lead to the discovery of povorcitinib (INCB054707, 22), which showed improved pharmacokinetics due to intramolecular hydrogen bonding of the amide moiety and a reduced ring size in the hinge binding motif. Protein crystallography suggested that JAK1 selectivity was conferred by the amide's alkyl group accessing a small lipophilic groove formed by the P-loop conformation of JAK1 that was distinct from JAK2. In murine arthritis models, povorcitinib demonstrated dose-dependent efficacy and reduced inflammatory signaling, supporting its therapeutic potential in immunological disorders.\n\n## Introduction\n\nIn the mid-1980s, the Wilks group was in search of colony-stimulating factor (CSF) receptors and speculated that they might be protein tyrosine kinases. In their search for CSF receptor kinases, they discovered two highly unusual kinases that possessed two kinase domains (active JH1 and pseudokinase JH2 domains) and lacked a transmembrane region, precluding them from being CSF receptors. Having isolated numerous other kinases that could potentially be CSF or CSF-related receptors, they named these two curious kinases *Just Another Kinase 1 and 2* (i.e., JAK1 and JAK2). Prior to publication, Wilks decided to change the name to Janus kinase, in reference to the Roman god that has two faces akin to the two kinase domains found in each JAK 1 and -2. Around the same time TYK2 was discovered, and a few years later, the final JAK family member JAK3 would be isolated and identified. Akin to the identification of JAK1 and -2, the discovery of their mode of action was also a somewhat secondary result. From the mid 1980s to the mid 90s, extensive research was conducted by numerous groups, notably Stark, Darnell, and Kerr groups, into understanding the action of interferons (IFN), which ultimately led to our understanding of the interplay between JAKs, signal transducers and activators of transcription (STAT), and IFNs. The elucidation of the JAK-STAT pathway was monumental in our understanding of how extracellular cytokines could convey their message to the nucleus to direct the transcription of specific genes through an eloquent cascade of biochemical events and reactions.\n\nUnder normal physiological conditions, the JAK-STAT pathway is intimately involved in immune response, which is initiated by the binding of a cytokine to an extracellular receptor that signals a dimerization event and concomitant activation of the attached intracellular JAK. Once activated, JAK phosphorylates tyrosine residues on the intracellular domain of the receptor, which acts as docking sites for one of seven specific STATs. Upon binding, the STATs are phosphorylated by JAK and then dissociate, dimerize, and translocate into the nucleus, whereupon they activate the specified gene's transcription. Thus, JAKs are a central and critical player in the transduction of extracellular signaling cytokines to activation and transcription of a multitude of proteins. Overactivation of members of the JAK family can occur through somatic mutations and single-nucleotide polymorphisms that lead to constitutive activity. Aberrant JAK activity can lead to a plethora of pathologies, particularly those related to autoimmune diseases and cancers (Figure 1).", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n**9. The same mask, two different answers.** The steering potentials clamp a value between a\nlower and an upper bound, and wrote the result with `energy[mask] = other[mask]`: ordinary Python,\nand correct everywhere except here. On Apple silicon that line compacts the array twice through\nthe same mask, and on a large tensor with very few flagged elements the two compactions disagree.\nMeasured live: a tensor of 3 x 35,830,451 values with 73 flagged, where the left-hand side\nselected 73 elements and the right-hand side 82, and the assignment failed with `shape mismatch:\nvalue tensor of shape [82] cannot be broadcast to indexing result of shape [73]`. That kills the\nwhole `boltz predict` process and every target queued behind it. It cost **nine crashes and\neighteen lost targets across three days** of a 58-target campaign before the cause was found, and\nit hid well: the discrepancy was +1, -2, -3, -4, -6, -2, -71 and -623 across those nine, so it\nread as a different bug each time; and it only fires once a coordinate actually leaves its bounds,\nwhich depends on the trajectory rather than on the molecule, so the same target could fail twice\nand then succeed. The fix replaces the assignment with `torch.where`, which decides element by\nelement and never compacts, so there is no count left to disagree about. It is the same\narithmetic, not an approximation: verified identical to the original on 600 randomised cases (both\nthe energy and derivative branches, infinite and absent bounds, scalar and vector force\nconstants), on the production tensor shape, and exactly against a CPU reference when run on the\nGPU. A companion patch prints every operand's shape and re-raises the original error unchanged, so\na future mis-shape says what it was rather than dying mutely -- that reporter is what identified\nthis one, in a single run, after two rounds of theorising from crash patterns had got nowhere.\nThe campaign went from crashing in seven of its first fifteen invocations to twenty-six\nconsecutive clean ones, and finished all 58 targets with steering left on throughout.\n\n## \u2699\ufe0f Progress, and how long a run will take\n\nTwo rows during `run`, laid out as a metrics rail: a state mark, a label, the bar, then every\nmeasurable value right-aligned in a fixed-width column of tabular figures, so digits sit under\ndigits and a phase name changing length cannot drag the row sideways.\n\n```\n\u25b6 targets    \u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501  3/15  2h14m   ~3h40m\n  structure  \u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501        18m04s  \u2593\u2593\u2593\u2591\u2591\u2591\u2591\u2591  21.4/69G\n```\n\nThe top row is the campaign: targets done, elapsed, and the estimate. The second is the\n**phase** Boltz is in (MSA generation / structure prediction / affinity prediction), parsed\nfrom its log output, with its own clock and the memory gauge. Only the bar is elastic. The state\nmark replaces the second spinner: <span>\u25b6</span> running, <span>\u23f8</span> paused,\n<span>\u25a0</span> stopping, always in the same column.\n\n**The memory gauge is filled against the point this machine starts to hurt, not against\ninstalled RAM** -- the same `MEMORY_THRASH_FRACTION` the swap-thrash warning uses. Measured\nagainst total memory, the 4-chain GPCR run above would have looked merely \"quite full\" until\nthe moment it died. The gauge turns amber at 60% of that ceiling and red at 85%, so it changes\ncolour before the log starts complaining rather than after.\n\n**Boltz exposes no diffusion- or recycling-step-level progress** anywhere in its output\n(verified against the installed package's source), so that is the finest granularity available\n-- and on a single-target campaign it is not fine at all: Boltz's own progress bar counts\ndataloader items, one target is one item, so it renders `0/1` at the start, `1/1` at the end,\nand nothing in between. A run can sit at `0/1` for an hour while working perfectly. Two things\nfollow, both deliberate:\n\n- The inner bar **pulses** rather than showing a static empty bar, and carries our own \"N in\n  this phase\" clock. Boltz's `it/s` figure is shown only while it is actually being refreshed;\n  once stale (60s without an update) it is replaced by that clock, because a rate string\n  written once at the start and displayed for an hour is a stale snapshot presented as live\n  data.\n- The ETA is **estimated, and says where the estimate came from**. Once a target completes it\n  is measured from this run; before that it comes from the median seconds-per-target of\n  previous runs of the same campaign on the same accelerator, read from\n  `.boltzmaker_run_history.jsonl` (median, not mean, so one swap-thrashing run does not set\n  expectations for every run after). With no history it says so rather than inventing a number.\n\nSizing a first run on new hardware is therefore unknown until one target lands. `--limit 1` is\nthe cheapest way to find out: one target gives you a real number and writes the history entry\nevery later run estimates from.\n\n### Controls while a run is going\n\nOn a real terminal, `run` takes two single keypresses. (Under `nohup`, a pipe or CI there is\nno keyboard to read, and it says so rather than appearing to offer them.) **`p` pauses and\nresumes**: a real `SIGSTOP` of Boltz and every worker it started, not a soft flag, so the\nprocesses freeze exactly where they stood and `p` again continues them mid-diffusion, with\nnothing discarded and nothing recomputed. **`q` quits**, stopping Boltz and all its worker\nprocesses, writing the run-history entry and exiting cleanly -- Ctrl-C takes the same path.", "# ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n| File | Records | Contents |\n| --- | --- | --- |\n| `approved_drugs.csv` | 3,257 | All FDA/EMA approved small molecules with SMILES + properties |\n| `bioactivity_key_targets.csv` | 112,357 | IC50/Ki/Kd for EGFR, CDK2, HIV protease, SARS-CoV-2 3CLpro, VEGFR2, Bcr-Abl, COX-2, BRAF, ALK, JAK1/2, hERG |\n| `bioactivity_extended_targets.csv` | ~200k | IC50/Ki for 36 additional targets: BTK, PARP1, CDK4/6, FLT3, mTOR, AKT1, PI3Ka, RET, MET, SRC, HDAC1, BRD4, EZH2, AR, ERa, PPARg, GLP-1R, Factor Xa, Thrombin, ACE, HMG-CoA, PDE5, DHFR, IDH1/2, KRAS + more |\n| `admet_data.csv` | ~50k | Caco-2 permeability, CLint, t1/2, fu,plasma, solubility, logD, CYP3A4/2D6/2C9/1A2/2C19 inhibition |\n| `clinical_candidates_phase2_3.csv` | ~10k | Phase 2-3 small molecules with properties |\n| `fragment_actives.csv` | 10,000 | MW < 300 fragments with pIC50 >= 4 |\n| `natural_products.csv` | ~5k | Natural product-derived compounds |\n| `drug_mechanisms.csv` | 7,561 | Mechanism of action + target for all drugs |\n| `drug_indications.csv` | 60,055 | Disease indications for all drugs |\n| `atc_classification.csv` | 5,579 | Full ATC therapeutic class hierarchy |\n\n### \ud83d\udcda PubChem (`data/corpus/pubchem/`, ~6.6 MB)\n\n| File | Records | Contents |\n| --- | --- | --- |\n| `fda_drugs_with_pubchem_cids.csv` | 3,257 | ChEMBL approved drugs cross-referenced to PubChem CIDs |\n| `tox21_panel.csv` | ~90k | 9 Tox21 nuclear receptor / stress-pathway assays (AhR, AR, ERa, NFkB, Nrf2, MMP, p53, ATAD5, TR) |\n| `pkis2_kinase_panel.csv` | ~50k | Published Kinase Inhibitor Set 2: 367-kinase selectivity panel |\n| `drug_repurposing_hub.csv` | ~6k | Broad Institute Drug Repurposing Hub clinical annotations |\n| `adme_cyp_hts.csv` | 19,803 | NCATS Caco-2 permeability, P-gp efflux, CYP2C9 inhibition HTS |\n| `bindingdb_affinities.csv` | \u2014 | BindingDB Ki/IC50/Kd (literature-curated) |\n| `nci60_gi50.csv` | \u2014 | NCI-60 cancer cell line growth inhibition |\n\n### \ud83d\udcda PDB (`data/corpus/pdb/`, ~111 MB)\n\n| File | Records | Contents |\n| --- | --- | --- |\n| `pdb_all_entries.csv` | 255,239 | All PDB structures: ID, method, resolution, compound, organism, deposit date |\n| `pdb_ligands_all.csv` | 16,171 | All unique PDB ligands with SMILES, InChI, InChIKey, formula, MW, type, ChEMBL/DrugBank/PubChem cross-references |\n| `pdb_ligands_druglike.csv` | 382 | Drug-like CCD subset (MW 100-800 Da, non-polymer, with SMILES) |\n| `pdb_ligand_structure_pairs.csv` | 121,541 | Protein-ligand pairs: which ligands are bound in which structures (with resolution + method) |\n| `pdb_druglike_structure_pairs.csv` | 1,671 | Drug-like protein-ligand pairs only |\n| `pdb_binding_affinities.csv` | 46,885 | Measured Kd/Ki/IC50 from PDBbind-annotated structures (9,209 entries) |\n| `sifts_pdb_uniprot.csv` | 965,937 | SIFTS: PDB chain -> UniProt accession (canonical target identity bridge) |\n| `sifts_pdb_pfam.csv` | 1,022,861 | SIFTS: PDB chain -> Pfam domain family (fold-level annotation) |\n\n## \ud83d\udcdd Recent changes\n\n### Live web deployment (2026-07-24)\n\n* **chemsage.mdeller.com is live.** The real `scripts/chat.py` CLI now runs in the browser via an xterm.js terminal, bridged over WebSocket to a server-side PTY \u2014 same Rich output, corpus tables, and slash commands as the local CLI, not a rebuilt chat-bubble UI.\n* **No local GPU on the droplet:** `mlx_lm`\u2018s `stream_generate`/`generate` are patched at import time (`chat_remote.py`) to call a HuggingFace Space running the real fused 32B model on ZeroGPU, then the unmodified `chat.py` is run unaware of the substitution.\n* **Serving stack:** gunicorn (`gevent` worker, 1 worker / 100 connections \u2014 a PTY session is I/O-bound, not CPU-bound) behind nginx with WebSocket-upgrade support on `/ws`, systemd service, Let\u2019s Encrypt TLS. Full detail in **Web deployment** below.\n\n### Chat CLI and dependency fixes (2026-06-30)\n\n* **mlx-lm 0.31.3 API fix (`scripts/chat.py`):** `stream_generate()` and `generate()` no longer accept `verbose=False` as a keyword argument \u2014 removed from both call sites.\n* **Corpus fast-path (`rag/corpus_lookup.py`):** Added `\"get\"` to the lookup trigger words so queries like `\"get all EGFR structures\"` route directly to the corpus table view instead of falling through to the LLM.\n* **System prompt (`scripts/chat.py`):** Added explicit instruction to render retrieved structured data (PDB structures, bioactivity measurements) as markdown tables rather than generating external API code.\n* **Dependencies (`requirements.txt`):** Added `torch` explicitly (required by `sentence-transformers` for local embeddings).\n* **How to run:** Added quick-start run instructions to the main README \u2014 single terminal, venv python, `--no-rag` flag for model-only mode.\n* **All 5 models published to HuggingFace Hub** (`Dellboy/chem_sage_*`).\n\n### \ud83e\uddea Eval harness fixes and 5-round comparative eval (2026-06-29)\n\n**Tokenizer fix (all models):**\n\u2013 `tokenizer_config.json` in `chem_sage_7b_v1`, `chem_sage_32b_v2`, `chem_sage_32b_v3`, `chem_sage_32b_v4` had `extra_special_tokens` as a JSON list of strings. Newer `transformers` calls `.keys()` on this field \u2014 causing an `AttributeError` on load. The server silently swallowed the error, loaded a broken tokenizer, and then hung indefinitely on every inference request (all queries timed out). Fixed by converting to an empty dict `{}` in all four files. `chem_sage_32b_v5` was already in the correct format.\n\n**Eval harness model-path fix (`eval/compare/eval_compare.py`):**\n\u2013 `query_model()` previously sent `\"model\": \"chemsage\"` in every HTTP POST to the local `mlx_lm.server`. Since `mlx_lm` 0.31.3, an unknown model name triggers a 120-second HuggingFace lazy-load timeout, causing all R1\u2013R4 evals to time out. Fixed by passing the full local model directory path (`model_path`) as the `\"model\"` field.", "# Approaches to automated protein crystal harvesting\nAuthors: Deller, M. C.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 70, 133\u2013155 (IYCr crystallization series), 2014\n\n- Marc Deller and Bernhard Rupp identify crystal harvesting as the last remaining fully manual step in an otherwise automated structural biology pipeline, creating both a throughput bottleneck and a source of late-stage crystal loss.\n- Harvesting comprises looping, cryoprotection, flash-cooling and mounting, and the looping technique itself had changed little in the roughly twenty years before the review.\n- JCSG data show the scale of the problem: just over 165,000 crystals harvested and screened for 1,470 structures, roughly 100 crystals per deposited structure, with 100 crystals taking an expert about four hours to harvest by hand.\n- Fully autonomous harvesting is blocked chiefly by crystal detection, since protein crystals are macroscopically irregular, move during handling, and current image-based algorithms carry error rates around 15%.\n- Fourteen semi-automated systems are catalogued, spanning six harvesting mechanisms \u2014 microcapillary, microtool, microgripper, acoustic droplet ejection, optical trap and MEMS \u2014 each with distinct trade-offs between damage, cost, ease of automation and data-collection compatibility.\n- The UMR workcell, developed by Square One System Design with input from Marc Deller and Bernhard Rupp and field-tested at the JCSG, combines 1 \u00b5m robot resolution, machine-vision guidance, database tracking, automated tape opening and an automated cryo-drip cryoprotection step.\n- UMR's small 3 mm sealing-tape aperture extends the useful life of a crystallization drop roughly tenfold compared with removing the whole tape for manual looping.\n- Hyperquenching at rates above 10,000\u00b0C s\u207b\u00b9, promoted by the cryo-drip oil layer and high-speed robotic plunging, has already been shown to cut cryoprotectant requirements by 22% and may eventually remove the need for added cryoprotectant altogether.\n- Automation is the only practical route to routine handling of microcrystals below 10 \u00b5m, which are needed to exploit microfocus beamlines and free-electron lasers and which are otherwise discarded as mere crystallization hits.\n- Non-contact methods \u2014 acoustic droplet ejection, optical traps and MEMS RodBots that move crystals by fluid vortices \u2014 avoid the mechanical stress and crystal-chasing problems of loops and microtools, at the cost of energy-related damage risks and, for some, bulky optics.\n- In situ growth and data collection sidestep harvesting entirely and are available at the ESRF, SLS, Diamond, HZB and the Australian Synchrotron, but suffer plate scattering, restricted rotation range and room-temperature-only collection.\n- The review concludes that no single mechanism has mastered crystal manipulation, that a universal solution will likely combine several technologies, and that it is remarkable how much of the world's estimated 700,000 crystals screened per year are still harvested by hand.\n\n## References\n\n*(reference list omitted)*", "# Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly\nAuthors: Deller, M.C.; Johnson, H.A.; Miller, M.D.; Spraggon, G.; Elsliger, M.-A.; Wilson, I.A.; Lesley, S.A.\nSource: PLoS ONE, 2015\n\nThe protein products of locus TM1088 assemble to form an octameric cytosolic regulatory component that likely represents the Trk homolog in T. maritima. The TM1088 assembly is the first K+ transporter structure assembled from two different, but structurally homologous proteins. The TM1088 assembly consists of the same folds, domains and assembly motifs as other K+ transporter assemblies, such as TrkA, GsuK, KtrAB, MthK and BK [16,18,20\u201322,24,25]. Given the atypical genetics and dimer composition of the TM1088 assembly, it is somewhat remarkable that it contains the same motifs that define the regulation paradigm of this family of RCK domain containing channels.\n\nThe differences between the two-subunit TM1088 assembly and other more classical single-subunit K+ gating ring assemblies can be understood by a simple gene duplication and fusion event (Fig 5), as proposed for the evolution of E. coli TrkA and other Trk/KtrAB proteins [14,26]. Furthermore, additional regulatory units have also been identified in the Trk/Ktr transporters of E. coli (TrkA and TrkE), Synechocystis sp. PCC6803 (KtrA and KtrE) and B. subtilis (KtrA and KtrC) [6]. Additionally, T. maritima is one of the oldest known bacteria with a very deep lineage and a propensity for lateral gene transfer [52,53]. Sequences of the 16S rRNA gene of members of the Thermotogales phylum reveal similarities to the last common ancestor and also suggest that the phylum is the closest bacterial relative to archaea bacteria and eukarya [54].\n\nHomologs of the two-subunit Trk system can also be found in several other bacterial genomes including Actinobacteria and Dehalococoides, illustrating their central role in bacterial physiology [9,55]. Most significantly, a two-subunit TrkA homolog has also been identified in Mycobacterium tuberculosis [9] and represents a novel target for structure-based drug design and the development of new anti-mycobacterial agents against multidrug-resistant strains. Members of the Trk/Ktr/HKT family are only found in non-animal cells and have no close homologs in animals and, therefore, represent highly selective targets for drug action [6]. Several anti-mycobacterials function by targeting K+ uptake, including clofazimine and 7-methyljuglone, although little is known about their mechanisms of action [56,57].\n\nA key structural feature of the TM1088 assembly is the presence of interfacial hydrogen bonds across the central channel. It is possible that these hydrogen bonds are actually salt bridges, as observed in the gating function of the bacterial ion-channel OmpA [58]. The gate of the OmpA channel is formed by a central salt bridge (Glu52-Arg138), which can flip-flop between two discrete states that represent closed and open forms [59]. Furthermore, Arg33 of TM1088B is structurally equivalent to Glu146 of the MthK assembly, which is also involved in the formation of interfacial contacts between neighboring RCK subunits upon gating ring assembly [18]. It is also possible that these interactions function in stabilizing the complex in this thermophilic organism. The contribution of such salt-bridge interactions to protein stability has been debated in previous studies and it has been suggested that thermophilic analogues of mesophilic proteins possess an increased number of salt bridges [60\u201362]. However, mutational studies suggest the contribution to protein stability is relatively small, although their stabilizing effects may be more apparent at higher temperatures [63]. Cross-pore inter-subunit hydrogen bonds of this type have not been observed in previously reported KtrA or TrkA structures, although salt bridges are formed between the cytosolic TrkA and the membrane component TrkH [20,21].\n\nThe central dogma of gating ring assembly and regulation of ion-flow by the K+ family of transporters is the expansion or contraction of the central ring of N-terminal RCK domains [16,18,20\u201325]. This movement is largely effected via conformational changes in the \"dimer hinge\" region between the N- and C-terminal RCK domains. To accommodate the movement of these domains, the neighboring N-terminal RCK domains must also move, as demonstrated for the BK and MthK gating ring assemblies, in which the protein interfaces on either side of the \"dimer-hinge\" are all the same and \"mobile\" in nature [16,18,22,24]. The ring then accommodates changes in the position of the N-terminal RCK domain using a uniform, four-fold symmetric dilation of the gating ring. However, in contrast, as a result of the domain-swapped nature of their dimer pairs, which in turn leads to a lower symmetry of domains around the central ring, TM1088, TrkA and KtrAB gating ring assemblies contain two distinct protein interfaces on either side of the \"dimer hinge\" [20,21,23]. One of these interfaces forms a \"mobile\" interface (analogous to the single \"mobile\" interface of BK and MthK) and the second distinct interface forms a \"fixed\" interface that does not move to accommodate changes in the \"dimer-hinge\" (Fig 6E).\n\nIt is likely that regulation of ion-flow by the TM1088 assembly could occur using similar \"dimer hinge\" and \"mobile\" interfaces, although the presence of a \"fixed\" interface has yet to be determined. It is important to note that at least one interface must be \"mobile\" to accommodate expansion or contraction of the central ring. However, given the high-level of structural homology with the TrkA gating ring assembly, and the similar D2 symmetry, it is likely that the TM1088 assembly will function by using two distinct \"mobile\" and \"fixed\" interfaces. The \"dimer hinge\" angles of the TM1088 assembly are ~120\u00b0, which is also in agreement with the angles observed in the ADP-bound, and presumably closed form, of the TrkA/TrkH gating ring complex from V. parahaemolyticus. These angles compare with the ATP\u03b3S-bound, and presumably open conformation, of TrkA/TrkH which has a \"dimer-hinge\" angle of ~100\u00b0 (Fig 6E) [21].", "# Clearing the Microscopic Threat: Can Plastic-Eating Enzymes Treat Liver Disease such as MASLD?\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-11-04 on marcdeller.com (categories: blog, microplastics).*\n\n* [Human: Elevated fecal microplastic levels correlate with inflammatory immune cells and persistent liver disease.](https://pubmed.ncbi.nlm.nih.gov/39260480/)\n* [Mouse Model: STZ+HFD model demonstrates accelerated MASLD progression with metabolic dysfunction resembling human disease.](https://www.nature.com/articles/s41467-024-50660-y)\n\n* **Human Evidence**: Elevated [fecal microplastic levels](https://pubmed.ncbi.nlm.nih.gov/39260480/) correlate with inflammatory immune cells and persistent liver disease after bariatric surgery.\n* **Mouse Disease Models**: The [STZ+HFD model](https://www.nature.com/articles/s41467-024-50660-y) demonstrates accelerated MASLD progression with metabolic dysfunction resembling human disease.\n* **Microplastic Mechanisms**: [Nanoplastic exposure](https://pmc.ncbi.nlm.nih.gov/articles/PMC12257899/) activates Hedgehog signaling and promotes hepatic stellate cell-driven fibrosis in experimental models.\n* **Enzyme Activity**: Engineered [PETase variants](https://pubs.rsc.org/en/content/articlelanding/2024/fd/d4fd00014e) degrade PET microplastics in human serum without affecting cell viability.\n* **Gut Microbiome Capacity**: Novel [PET hydrolases](https://pubmed.ncbi.nlm.nih.gov/39551294/) identified from human intestinal bacteria reveal natural plastic degradation potential in the digestive tract.\n\n1. **Links between fecal microplastics and parameters related to metabolic dysfunction-associated steatotic liver disease (MASLD) in humans**: [Schwenger, K.J.P., et al., Science of The Total Environment, 2024](https://pubmed.ncbi.nlm.nih.gov/39260480/)\n2. **A male mouse model for metabolic dysfunction-associated steatotic liver disease and hepatocellular carcinoma**: [Jeong, B.K., et al., Nature Communications, 2024](https://www.nature.com/articles/s41467-024-50660-y)\n3. **Chronic nanoplastic exposure promotes the development and progression of metabolic dysfunction-associated steatotic liver disease**: [Han, J., et al., Liver International, 2025](https://pmc.ncbi.nlm.nih.gov/articles/PMC12257899/)\n4. **Degradation of PET microplastic particles to monomers in human serum by PETase**: [Lopez-Lorenzo, X., et al., Faraday Discussions, 2024](https://pubs.rsc.org/en/content/articlelanding/2024/fd/d4fd00014e)\n5. **Identification of a PET hydrolytic enzyme from the human gut microbiome unveils potential plastic biodegradation**: [Zhang, G., et al., International Journal of Biological Macromolecules, 2024](https://pubmed.ncbi.nlm.nih.gov/39551294/)\n\n**A bold therapeutic approach to target microplastic accumulation in metabolic liver dysfunction through engineered PETase enzymes**\n\nThe growing presence of microplastics in human tissues has sparked concern among researchers investigating metabolic dysfunction-associated steatotic liver disease (MASLD). Recent evidence links [microplastic accumulation](https://pubmed.ncbi.nlm.nih.gov/39260480/) to worsening liver damage, creating an urgent need for therapeutic interventions. Pioneering work now explores whether PETase enzymes\u2014originally discovered in bacteria that consume plastic waste\u2014could offer a clinical solution by breaking down these persistent particles within the body. This approach builds on groundbreaking [mouse studies](https://www.nature.com/articles/s41467-024-50660-y) showing accelerated liver disease progression when animals are exposed to microplastics alongside metabolic stressors, while emerging human data reveals connections between [fecal microplastic levels](https://pubmed.ncbi.nlm.nih.gov/39260480/) and persistent liver pathology following bariatric surgery.\n\nThe clinical thesis centers on a straightforward yet revolutionary concept: if microplastics contribute to MASLD progression, enzymatic degradation of these particles could reduce hepatic injury. PETase enzymes naturally hydrolyze polyethylene terephthalate into smaller molecules like [terephthalic acid and mono(2-hydroxyethyl) terephthalate](https://www.pnas.org/doi/10.1073/pnas.1718804115), which cellular metabolism can further process. Scientists have demonstrated that [engineered PETase variants](https://pubs.rsc.org/en/content/articlelanding/2024/fd/d4fd00014e) maintain activity in human serum at body temperature while showing no toxicity to kidney cells during laboratory testing. This activity profile suggests feasibility for therapeutic delivery, though significant development remains necessary.\n\nHuman studies provide compelling support for this approach. A 2024 investigation examined 23 patients with MASLD, finding that those with elevated [microplastic fiber concentrations](https://pubmed.ncbi.nlm.nih.gov/39260480/) showed positive correlations with inflammatory immune cells and altered gene expression patterns related to apoptosis and immune function. Most strikingly, patients with persistent liver disease twelve months after bariatric surgery exhibited higher fecal microplastic fragments than those whose histology normalized, suggesting that microplastic burden may impede metabolic recovery. These findings align with observations that microplastic exposure activates inflammatory pathways and promotes hepatic stellate cell activation, key drivers of fibrosis progression.", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\n**37.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein each R^1A is independently selected from D, OH, NO2, CN, halo, C1\u20133 alkyl, C2\u20133 alkenyl, C2\u20133 alkynyl, C1\u20133 haloalkyl, cyano-C1\u20133 alkyl, HO\u2014C1\u20133 alkyl, C1\u20133 alkoxy-C1\u20133 alkyl, C3\u20137 cycloalkyl, C1\u20133 alkoxy, C1\u20133 haloalkoxy, amino, C1\u20133 alkylamino, di(C1\u20133 alkyl)amino, thio, C1\u20133 alkylthio, C1\u20133 alkylsulfinyl, C1\u20133 alkylsulfonyl, carbamyl, C1\u20133 alkylcarbamyl, di(C1\u20133 alkyl)carbamyl, carboxy, C1\u20133 alkylcarbonyl, C1\u20133 alkoxycarbonyl, C1\u20133 alkylcarbonyloxy, C1\u20133 alkylcarbonylamino, C1\u20133 alkoxycarbonylamino, C1\u20133 alkylaminocarbonylamino, C1\u20133 alkylsulfonylamino, aminosulfonyl, C1\u20133 alkylaminosulfonyl, di(C1\u20133 alkyl)aminosulfonyl, aminosulfonylamino, C1\u20133 alkylaminosulfonylamino, di(C1\u20133 alkyl)aminosulfonylamino, aminocarbonylamino, C1\u20133 alkylaminocarbonylamino, and di(C1\u20133 alkyl)aminocarbonylamino.\n\n**38.** (canceled)\n\n**39.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein R^1 is independently selected from H, D, halo, CN, C1\u20136 alkyl, C1\u20136 haloalkyl, C1\u20136 haloalkoxy, and C1\u20136 alkoxy.\n\n**40.** (canceled)\n\n**41.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein m is an integer selected from 0, 1, and 2.\n\n**42-44.** (canceled)\n\n**45.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein n is an integer selected from 0, 1, and 2.\n\n**46-47.** (canceled)\n\n**48.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein R^2 is independently selected from H, D, halo, CN, C1\u20136 alkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, C1\u20136 haloalkyl, C3\u201310 cycloalkyl, 6\u201310 membered aryl, 4\u201310 membered heterocycloalkyl, 5\u201310 membered heteroaryl and their C1\u20134 alkyl-linked variants, OR^a2, SR^a2, NHOR^a2, C(O)R^b2, C(O)NR^c2R^d2, C(O)OR^a2, OC(O)R^b2, OC(O)NR^c2R^d2, NR^c2R^d2, NR^c2C(O)R^b2, NR^c2C(O)OR^a2, NR^c2C(O)NR^c2R^d2, NR^c2S(O)2R^b2, NR^c2S(O)2NR^c2R^d2, S(O)2R^b2, and S(O)2NR^c2R^d2, each optionally substituted with 1, 2, 3, or 4 independently selected R^2A substituents.\n\n**49-51.** (canceled)\n\n**52.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein each R^a2, R^c2, and R^d2 is independently selected from H, C1\u20136 alkyl, C1\u20136 haloalkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, C3\u20137 cycloalkyl, phenyl, 4\u20137 membered heterocycloalkyl, 5\u20136 membered heteroaryl and their C1\u20134 alkyl-linked variants, each optionally substituted with 1, 2, 3, or 4 independently selected R^2A substituents; and each R^b2 is independently selected from the corresponding non-hydrogen members of that list.\n\n**53.** (canceled)\n\n**54.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein each R^2A is independently selected from H, D, halo, CN, C1\u20136 alkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, C1\u20136 haloalkyl, C3\u20137 cycloalkyl, phenyl, 4\u20137 membered heterocycloalkyl, 5\u20136 membered heteroaryl and their C1\u20134 alkyl-linked variants, OR^a21, SR^a21, NHOR^a21 and the corresponding acyl, amino and sulfonyl groups, each optionally substituted with 1, 2, 3, or 4 independently selected R^2B substituents.\n\n**55-57.** (canceled)\n\n**58.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein: R^2 is selected from CN, C1\u20136 alkyl, OR^a2, C(O)R^b2, C(O)NR^c2R^d2, and C(O)OR^a2; R^a2, R^c2, and R^d2 are each independently selected from H, C1\u20136 alkyl, and 4\u20137 membered heterocycloalkyl, wherein said C1\u20136 alkyl and 4\u20137 membered heterocycloalkyl are each optionally substituted with one R^2A; R^b2 is C1\u20136 alkyl or 4\u20137 membered heterocycloalkyl, each of which is optionally substituted with one R^2A substituent; and R^2A is selected from OH and C1\u20133 alkoxy.\n\n**59-60.** (canceled)\n\n**61.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein R^3 is selected from H, D, OH, CN, halo, C1\u20136 alkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, C1\u20136 haloalkyl, cyano-C1\u20136 alkyl, HO\u2014C1\u20136 alkyl, and C1\u20136 alkoxy-C1\u20136 alkyl.\n\n**62.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein R^3 is selected from CN and C1\u20136 alkyl.\n\n**63-66.** (canceled)\n\n**67.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein: R^4 and R^5 are each independently selected from H and C1\u20136 alkyl; or R^4 and R^5, together with the carbon atom to which they are attached, form a 3\u20137-membered cycloalkyl ring.\n\n**68-72.** (canceled)\n\n**73.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein R^6 is independently selected from C3\u20137 cycloalkyl, phenyl, 4\u20137 membered heterocycloalkyl, and 5\u20136 membered heteroaryl, each of which is optionally substituted by 1, 2, 3 or 4 independently selected R^6A substituents.\n\n**74-80.** (canceled)\n\n**81.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein each R^6A is independently selected from D, OH, CN, halo, C1\u20136 alkyl, C2\u20136 alkenyl, C2\u20136 alkynyl, C1\u20136 haloalkyl, cyano-C1\u20136 alkyl, HO\u2014C1\u20136 alkyl, C1\u20136 alkoxy-C1\u20136 alkyl, C3\u20136 cycloalkyl, C1\u20136 alkoxy, C1\u20133 haloalkoxy, amino, C1\u20136 alkylamino, and di(C1\u20136 alkyl)amino.\n\n**82.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein each R^6A is independently selected from halo, CN, C1\u20136 alkyl, and C1\u20136 haloalkyl.\n\n**83.** (canceled)\n\n**84.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein X is C; and Y is C.\n\n**85.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein the dashed bond is a double bond, provided that proper valency is maintained.\n\n**86.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein the dashed bond is a single bond, provided that proper valency is maintained.\n\n**87.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein Z is N.\n\n**88.** The compound of claim 1, or a pharmaceutically acceptable salt thereof, wherein Z is CR^7.", "# BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nProtein: 5HT2C\nSequence: MVNLRNAVHSFLVHLIGLLVWQSDISVSPVAAIVTDIFNTSDGGRFKFPDGVQNWPALSIVIIIIMTIGGNILVIMAVSMEKKLHNATNYFLMSLAIADMLVGLLVMPLSLLAILYDYVWPLPRYLCPVWISLDVLFSTASIMHLCAISLDRYVAIRNPIEHSRFNSRTKAIMKIAIVWAISIGVSVPIPVIGLRDEEKVFVNNTTCVLNDPNFVLIGSFVAFFIPLTIMVITYCLTIYVLRRQALMLLHGHTEEPPGLSLDFLKCCKRNTAEEENSANPNQDQNARRRKKKERRPRGTMQAINNERKASKVLGIVFFVFLIMWCPFFITNILSVLCEKSCNQKLMEKLLNVFVWIGYVCSGINPLVYTLFNKIYRRAFSNYLRCNYKVEKKPPVRQIPRVAATALSGRELNVNIYRHTNEPVIEKASDNEPGIEMQVENLELPVNPSSVVSERISSV\nPartners: GNAQ, GNB1, GNG2\nLigands: SB24, LORC\nApo structure: boltz_cif/H2CAP_model_0.cif\nFamily type: gpcr\nGroup: 5HT2C\n\nProtein: H2CNG\nSequence: MVNLRNAVHSFLVHLIGLLVWQSDISVSPVAAIVTDIFNTSDGGRFKFPDGVQNWPALSIVIIIIMTIGGNILVIMAVSMEKKLHNATNYFLMSLAIADMLVGLLVMPLSLLAILYDYVWPLPRYLCPVWISLDVLFSTASIMHLCAISLDRYVAIRNPIEHSRFNSRTKAIMKIAIVWAISIGVSVPIPVIGLRDEEKVFVNNTTCVLNDPNFVLIGSFVAFFIPLTIMVITYCLTIYVLRRQALMLLHGHTEEPPGLSLDFLKCCKRNTAEEENSANPNQDQNARRRKKKERRPRGTMQAINNERKASKVLGIVFFVFLIMWCPFFITNILSVLCEKSCNQKLMEKLLNVFVWIGYVCSGINPLVYTLFNKIYRRAFSNYLRCNYKVEKKPPVRQIPRVAATALSGRELNVNIYRHTNEPVIEKASDNEPGIEMQVENLELPVNPSSVVSERISSV\nLigands: SB24, LORC\nApo structure: boltz_cif/H2CAP_model_0.cif\nFamily type: gpcr\nGroup: 5HT2C\n\nProtein: H2CAP\nSequence: MVNLRNAVHSFLVHLIGLLVWQSDISVSPVAAIVTDIFNTSDGGRFKFPDGVQNWPALSIVIIIIMTIGGNILVIMAVSMEKKLHNATNYFLMSLAIADMLVGLLVMPLSLLAILYDYVWPLPRYLCPVWISLDVLFSTASIMHLCAISLDRYVAIRNPIEHSRFNSRTKAIMKIAIVWAISIGVSVPIPVIGLRDEEKVFVNNTTCVLNDPNFVLIGSFVAFFIPLTIMVITYCLTIYVLRRQALMLLHGHTEEPPGLSLDFLKCCKRNTAEEENSANPNQDQNARRRKKKERRPRGTMQAINNERKASKVLGIVFFVFLIMWCPFFITNILSVLCEKSCNQKLMEKLLNVFVWIGYVCSGINPLVYTLFNKIYRRAFSNYLRCNYKVEKKPPVRQIPRVAATALSGRELNVNIYRHTNEPVIEKASDNEPGIEMQVENLELPVNPSSVVSERISSV\nLigands: none\nFamily type: gpcr\nGroup: 5HT2C\n\n# ===================== Gq heterotrimer partners =====================\n\nPartner: GNAQ\nSequence: MTLESIMACCLSEEAKEARRINDEIERQLRRDKRDARRELKLLLLGTGESGKSTFIKQMRIIHGSGYSDEDKRGFTKLVYQNIFTAMQAMIRAMDTLKIPYKYEHNKAHAQLVREVDVEKVSAFENPYVDAIKSLWNDPGIQECYDRRREYQLSDSTKYYLNDLDRVADPAYLPTQQDVLRVRVPTTGIIEYPFDLQSVIFRMVDVGGQRSERRKWIHCFENVTSIMFLVALSEYDQVLVESDNENRMEESKALFRTIITYPWFQNSSVILFLNKKDLLEEKIMYSHLVDYFPEYDGPQRDAQAAREFILKMFVDLNPDSDKIIYSHFTCATDTENIRFVFAAVKDTILQLNLKEYNLV\n\nPartner: GNB1\nSequence: MSELDQLRQEAEQLKNQIRDARKACADATLSQITNNIDPVGRIQMRTRRTLRGHLAKIYAMHWGTDSRLLVSASQDGKLIIWDSYTTNKVHAIPLRSSWVMTCAYAPSGNYVACGGLDNICSIYNLKTREGNVRVSRELAGHTGYLSCCRFLDDNQIVTSSGDTTCALWDIETGQQTTTFTGHTGDVMSLSLAPDTRLFVSGACDASAKLWDVREGMCRQTFTGHESDINAICFFPNGNAFATGSDDATCRLFDLRADQELMTYSHDNIICGITSVSFSKSGRLLLAGYDDFNCNVWDALKADRAGVLAGHDNRVSCLGVTDDGMAVATGSWDSFLKIWN\n\nPartner: GNG2\nSequence: MASNNTASIAQARKLVEQLKMEANIDRIKVSKAAADLMAYCEAHAKEDPLLTPVPASENPFREKKFFCAIL\n\n# ===================== Ligands =====================\n\nLigand: RISP\n# Risperidone -- 5-HT2A antagonist (also broad-spectrum antipsychotic)\nSMILES: CC1=C(C(=O)N2CCCCC2=N1)CCN3CCC(CC3)c4c5ccc(cc5on4)F\nRole: antagonist\n\nLigand: PSIL\n# Psilocin -- 5-HT2A agonist (active metabolite of psilocybin)\nSMILES: CN(C)CCc1c[nH]c2c1c(ccc2)O\nRole: agonist\n\nLigand: BALO\n# Balovaptan -- 5-HT2B antagonist\nSMILES: CN1Cc2cc(ccc2-n3c(nnc3C4CCC(CC4)Oc5ccccn5)C1)Cl\nRole: antagonist\n\nLigand: LSD1\n# LSD -- 5-HT2B agonist (also 5-HT2A)\nSMILES: CCN(CC)C(=O)[C@H]1CN([C@@H]2Cc3c[nH]c4c3c(ccc4)C2=C1)C\nRole: agonist\n\nLigand: SB24\n# SB-242084 -- 5-HT2C antagonist\nSMILES: Cc1cc2c(cc1Cl)N(CC2)C(=O)Nc3ccc(nc3)Oc4cccnc4C\nRole: antagonist\n\nLigand: LORC\n# Lorcaserin -- 5-HT2C agonist (former anti-obesity drug)\nSMILES: C[C@H]1CNCCc2c1cc(cc2)Cl\nRole: agonist\n\n```\n\nThree receptors, six ligands scoped two apiece, and a Gq heterotrimer partner: that gives 15 targets, twelve ligand-bound (each run with and without Gq) plus three predicted-apo references. A single file drives the whole run and produces an [interactive dashboard](https://bellcheddar.github.io/BoltzMaker/examples/5ht2_gq_panel/boltz_dashboard.html): ranked confidence and affinity, a confidence-versus-affinity scatter, per-ligand chemistry review, and, when interaction profiling is enabled, a residue fingerprint heatmap and a rotating 3D view of each binding site, generated using [PLIP](https://academic.oup.com/nar/article/53/W1/W463/8128215), [PyMOL](https://pymol.org/pymol.html), and [3Dmol.js](https://3dmol.org/):\n\n5HT2 campaign summary table from the BoltzMaker interactive dashboard\n\n5HT2 data summary table from the BoltzMaker interactive dashboard\n\n5HT2 ligand preparation panels from the BoltzMaker interactive dashboard\n\n5HT2 ranked pIC50, confidence, and interaction counts from the BoltzMaker interactive dashboard\n\n5HT2 structural analysis and binding profiles from the BoltzMaker interactive dashboard\n\n###### **The wizard of vibe coding a wizard**\n\nIf you would prefer not to write the input spec by hand, `BoltzMaker.py new` command interviews you in plain English and writes a valid file for you. Given a reference co-crystal or homology model, it can even suggest pocket-contact residues and remap them onto your target\u2019s own numbering by sequence alignment, so the reference structure\u2019s residue numbers do not have to match yours. The next step for the project is to integrate my vibe-coded, chemically and biologically aware [ChemSage LLM](https://marcdeller.com/chemsage/) to achieve more fine-grained control at this step.\n\nHere is an example run of\u00a0BoltzMaker.py new:\n\nExample run of BoltzMaker.py new showing the prompts for Proteins and Ligand to streamline starting the campaign.\n\n###### **Keeping the chemistry honest**\n\nThe part I care most about, having watched a bad SMILES quietly ruin a result more than once, is that the chemistry is checked before any compute is spent. Boltz folds whatever it is handed: an undefined stereocentre, an unintended protonation state, or a stray counterion does not raise an error, it just silently changes the predicted pose and affinity.", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n**A reference of a different protein is refused, not reported on.** Receptor residues are matched\non number with the residue type required to agree, and a reference must match at least 70% of the\nnumbering overlap to be accepted as this protein. Measured against 7E14: GLP1R matched at 0.90,\nGIPR at 0.11. Without that gate the panel compares a GIPR prediction to a GLP1R crystal structure\nand presents the disagreement as a finding. Each reference chain is tried separately rather than\npooled, because a complex numbers its G-protein chains from 1 as well, and pooling lets the last\nchain read overwrite the receptor's own residues.\n\n### \ud83d\udd2d How tight should a pocket be? What the measurements say\n\nPointed at a real question, the panel found this: **`Pocket distance:` decides whether a\nconstrained prediction reproduces the crystal pose, and 8 A is too loose.**\n\nSix conditions, all GLP1R + orforglipron scored against 7E14, one diffusion sample each, no\ntemplates. \"Contacts\" is the number of receptor residues named in the constraint; \"shell\" is the\ndistance from the experimental ligand they were derived at. The campaign is public as\n[`GLP1R_GIPR_pocket_matrix`](https://boltzmaker.mdeller.com/runs/niSiE5MfGFyddSjGLmeYBA/explore),\nso every row below can be read in the live report.\n\n| Pocket | Shell | Contacts | `Pocket distance` | Site (\u00c5) | Pose (\u00c5) | Conformer (\u00c5) | Confidence | Ligand ipTM |\n|---|---|---|---|---|---|---|---|---|\n| P60 | 6 \u00c5 | 38 | 4 \u00c5 | 1.36 | **2.04** | 1.34 | 0.849 | 0.987 |\n| P80 | 8 \u00c5 | 62 | 4 \u00c5 | 1.64 | 2.56 | 1.46 | 0.844 | 0.976 |\n| P35 | 3.5 \u00c5 | 13 | 4 \u00c5 | 2.83 | 3.70 | 1.66 | 0.841 | 0.939 |\n| P45 | 4.5 \u00c5 | 25 | 4 \u00c5 | 2.45 | 8.96 | 3.38 | 0.834 | 0.952 |\n| V6G | 8 \u00c5 | 62 | 8 \u00c5 | 3.08 | 9.51 | 3.08 | 0.836 | 0.940 |\n| unconstrained | -- | sparse sweep | 8 \u00c5 | 11.02 | 13.41 | 3.43 | 0.765 | 0.754 |\n\n**The distance is the lever, and P80 isolates it.** P80 and V6G name the *identical 62 residues*.\nNothing differs between them but 4 \u00c5 against 8 \u00c5, and the pose goes from 9.51 \u00c5 -- the right\npocket, the wrong way round -- to 2.56 \u00c5, a reproduction of the experiment. That is a\nsingle-variable result, which none of the other comparisons are.\n\n**Do not tune the residue count.** At a fixed 4 \u00c5 it does not order the outcome: 13 contacts gives\n3.70 \u00c5, 25 gives 8.96 \u00c5, 38 gives 2.04 \u00c5, 62 gives 2.56 \u00c5. Outcomes are bimodal -- a run either\nlands (2.0-3.7 \u00c5, conformer 1.3-1.7) or it does not (9-9.5 \u00c5, conformer 3.1-3.4) -- and P45 fails\nthe same way V6G does, getting the ligand's own conformation wrong as well as its placement. With\none sample per condition there is no separating \"25 contacts is bad\" from \"P45 was a bad draw\",\nso the honest summary is that **at 4 \u00c5 three of four pocket definitions reproduced the crystal\npose, and at 8 \u00c5 none did.**\n\n**The confidence columns are the argument for this panel.** Across poses spanning 2.04 to 9.51 \u00c5,\n`confidence_score` moves 0.834 to 0.849 -- a spread of 0.015 that does not even put the worst pose\nlast. Ligand ipTM orders the three good runs correctly (0.939 < 0.976 < 0.987 for 3.70 > 2.56 >\n2.04 \u00c5) and then hands P45's 8.96 \u00c5 a 0.952, placing a failure mid-pack. Both scores do drop\nsharply for the unconstrained run (0.765 / 0.754), so they can tell \"on the receptor but not in\nthe site\" from \"in the site\" -- they simply cannot tell a right pose from a wrong one once the\nligand is in the pocket. Only the comparison with an experiment can.\n\n**Affinity is not in this table**, because the pose series was run with `Predict affinity: no` --\nit was a geometry test. The one row that has it is the unconstrained baseline (binder p 0.496,\npIC50 8.98), and the campaign it came from already measured binder probability at r = +0.07\nagainst pIC50 across 24 targets, moving with the pocket condition rather than the chemistry.\n\n#### What to set\n\nUse **`Pocket distance: 4`** for a named pocket. Leave the contact list as whatever your\nreference structure gives you; deriving it at a 6-8 \u00c5 shell is fine and the difference between\nthose two was within the noise of this experiment.\n\nThis does **not** apply to the ligand-free \"stay on your own receptor\" sweep that unconstrained\ntargets get -- that is fixed at `CONFINE_DISTANCE_A` (8 \u00c5) and deliberately not coupled to\n`Pocket distance:`. The two constraints ask different questions: a pocket asks *which site*,\nwhere tightening is the whole point, while the sweep asks *which protein* of residues scattered\nthe length of the chain, where no ligand can be near all of them and the number only scales the\npull. Coupling them meant tightening a pocket silently re-tuned the baseline arm it was about to\nbe compared with.\n\n#### Does the restraint just manufacture the pose?\n\nThe series above shows a tight pocket reproducing a crystal pose. That is only worth something if\nthe restraint can also **fail**: point it at the wrong site and a pose panel that still reports\n2 \u00c5 would be measuring the constraint, not the prediction.\n\nA later two-receptor campaign runs that control. Every ligand is predicted three times --\nrestrained to one pocket, restrained to the other, and unrestrained -- so each of the two ligands\nwith an experimental structure gets its own site, a wrong site and no site at all, all at\n`Pocket distance: 4`. Pocket `V6G` comes from 7E14, pocket `41Y` from 7RBT; each is the site its\nown ligand actually occupies.\n\n| Ligand | Restrained to | Site (\u00c5) | Pose (\u00c5) | Conformer (\u00c5) |\n|---|---|---|---|---|\n| 41Y (vs 7RBT) | **its own site** (41Y) | 0.66 | **0.91** | 0.51 |\n| orforglipron (vs 7E14) | **its own site** (V6G) | 1.85 | **2.79** | 1.63 |\n| 41Y | nothing (unconstrained) | 6.00 | 6.17 | 0.81 |\n| orforglipron | nothing (unconstrained) | 11.64 | 13.57 | 2.92 |\n| 41Y | **the wrong site** (V6G) | 22.88 | 24.93 | 1.78 |\n| orforglipron | **the wrong site** (41Y) | 28.23 | 29.40 | 3.67 |", "# GitHub README: PANTS\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/PANTS, 2026\n\nOnly three enzymes sit anywhere near 37 \u00b0C, and two of them (Z1-PETase, DuraPETase) are\nengineered variants of IsPETase that reach it by trading away the activity the others\ngained. **HGMP01 is the only one that arrived there naturally**, and the only measured\nenzyme in this project with a near-neutral pH optimum as well: everything else wants pH 8\nto 9, and even IsPETase asks for 9.\n\nThat is the therapeutic gap, and it is not an argument: it is what the numbers say. It is\nalso the reason the human gut was added as a fourth source environment, since an enzyme\nresident at 37 \u00b0C and pH 7.4 has already been selected under something close to the\ntarget conditions.\n\n### Why HGMP01 matters, and the trap in retrieving it\n\nHGMP01 is a PET hydrolase [identified from the human gut microbiome](https://pubmed.ncbi.nlm.nih.gov/39551294/)\n(*Int J Biol Macromol* 283, 2024). It hydrolyses PET nanoparticles, outperformed the four\nother candidates in that study, and shares only about 5% identity with IsPETase. Its\nhomologues are distributed across 41 families and 94 genera of gut microbes.\n\nIt is in no public sequence database: zero hits across UniProt, NCBI protein and NCBI\nnuccore. The sequences live in the authors' SciDB deposit.\n\n**The paper and the deposit number the enzymes differently, and matching on the name\nwould have mislabelled all five.** The paper's HGMP01 to HGMP05 are the deposit's HGMP03,\n04, 06, 07 and 08. The mapping is pinned by two independent facts:\n\n| Paper | Length | Deposit | Length | Identifier | Homologues |\n|---|---|---|---|---|---|\n| **HGMP01** | 275 | HGMP03 | 275 | **`GUT_GENOME238302_00589`** | **697** |\n| HGMP02 | 341 | HGMP04 | 341 | `GUT_GENOME243637_00613` | 131 |\n| HGMP03 | 323 | HGMP06 | 323 | `GUT_GENOME137663_00143` | 96 |\n| HGMP04 | 282 | HGMP07 | 282 | `GUT_GENOME171691_00743` | 1000 |\n| HGMP05 | 321 | HGMP08 | 320 | `GUT_GENOME244370_00064` | 87 |\n\nEvery length is distinct, so the assignment is 1:1 with no ambiguity. Independently, the\npaper states that the homologue search \"identified a total of 697 putative HGMP01-like\nenzymes\", and the deposit's HGMP03 returned exactly 697 hits. Two unrelated numbers\nagreeing is what makes this a determination rather than a guess.\n\nEvery sequence is length-checked against the paper's own table before being stored, and\nthat check immediately caught a transposition between the similar identifiers\n`GUT_GENOME244370_00064` and `GUT_GENOME244064_00699`, which would otherwise have filed\nthe wrong protein as HGMP05 with nothing downstream to reveal it.\n\n## \ud83d\udea7 Current status\n\n**Live at [`pants.mdeller.com`](https://pants.mdeller.com).** The pipeline runs end to end from\nmetagenome FASTA to folded, geometry-measured candidates, and the catalogue is served behind\ngunicorn and nginx. **The project reached a conclusion that is not the one it set out to\nreach \u2014 read [`FINDINGS.md`](FINDINGS.md) first, then [`NEXT.md`](NEXT.md).**\n\n| Phase | State |\n|---|---|\n| 0: Scaffold, schema, manifest provenance | \u2705 Complete |\n| 1: Curation, hard negatives, activity data | \u2705 Complete |\n| 2: Recall (profile HMMs, MMseqs2, HMMER) | \u2705 Complete |\n| 4: ESM-2 embedding | \u2705 Complete |\n| 5: Activity head, calibration, evaluation | \u2705 Complete, and **negative**: a learned head does not beat retrieval on measured labels |\n| 6: Structures and active-site geometry | \u2705 Complete: 416 of 439 candidates folded, 559 reference structures |\n| 7 to 8: Web app and deployment | \u2705 Complete: live on port 8005, listed on the mdeller.com launcher |\n| 9: Measured negatives and the three findings | \u2705 Complete |\n\nWhat is in the database today:\n\n| Set | Count | Notes |\n|---|---|---|\n| **Candidates** | **439** | Mined from 14,804,920 metagenomic proteins across five environments, all triad-complete |\n| Characterised enzymes | 1,348 | Of which **342 carry a measurement**, the rest annotation |\n| **Measured negatives** | **150** | Expressed, assayed, no product released \u2014 from two 2025 screens. The class this field usually lacks |\n| Hard negatives | 131 | Other \u03b1/\u03b2-hydrolase folds, matched on five axes |\n| Activity measurements | 2,759 | 2,757 carrying a DOI |\n| Reference structures | 559 | 60 crystal, 104 AlphaFold, 395 ESMFold, all superposed onto IsPETase |\n| Unlabelled homologues | 25,041 | Held outside the catalogue so they cannot inflate a quoted count |\n\n### The headline result\n\nCandidates are ranked by **similarity to the nearest enzyme whose PET activity was measured**,\nnot by a learned classifier, because on measured labels no model built here beats that\nbaseline. Every candidate carries a competence band from the measured identity-decay curve:\n\n| Band | Candidates | Meaning |\n|---|---|---|\n| In range | **16** | \u226570% identity to a measured-active enzyme |\n| Marginal | 39 | 50\u201370% |\n| Out of range | **384** | <50%, ranked but with no demonstrated validity |\n\nThe out-of-range candidates are precisely the novel enzymes this project set out to find, and\nnothing here can score them. That is stated on the site rather than hidden, and the fix is\nmeasurement on new lineages \u2014 see [`release/validation_panel.csv`](release/validation_panel.csv).\n\n### Positives by evidence tier\n\nThe count that matters is not 854 but **341**: the number with a measurement behind the label.\n\n| Tier | n | What it means |\n|---|---|---|\n| `EC-auto-annotated` | 449 | EC 3.1.1.101 assigned by similarity (ECO:0000256). A prediction, not a measurement |\n| `ESTHER-family-predicted` | 50 | Family membership only |\n| `ESTHER-family-protein-evidence` | 14 | Family, protein observed |\n| `EC-experimental` | 10 | EC 3.1.1.101 with ECO:0000269 and PubMed citations |\n| Curated wild types and variants | 6 | Hand-curated, sequence-verified, mutations validated |\n| **`PAZy-measured`** | **312** | In PAZy because activity was **measured** on a plastic and published, each with a DOI |\n| `HGMP-measured` | 5 | Human gut PET hydrolases with measured activity (PMID 39551294) |\n\n## \ud83e\uddea What Phase 1 found", "# HGMP01 and FAST-PETase: Decoding Two Paths to Plastic Degradation \u2013 Engineered versus Natural PETase Enzymes\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| PETase feature | FAST-PETase | HGMP01 | Reference |\n| --- | --- | --- | --- |\n| Mutations | Five engineered mutations (S121E, D186H, R224Q, N233K, R280A) enhance thermostability and catalytic efficiency. | Natural variant with no engineered modifications. | Lu et al., JACS, 2023 |\n| Sequence identity | 100% sequence identity to IsPETase minus five mutations. | 20-41% sequence identity to common PETases, indicating independent evolutionary origin. | Zhang et al., IJBM, 2024 |\n| Catalytic machinery | Ser160-His237-Asp206 triad; N233K raises Asp206 basicity, lowering activation energy by 4.4 kcal/mol. | Conserved catalytic triad; specific residue details less characterized. | Berke et al., Commun. Chem (Nature), 2024 |\n| Thermodynamic properties | Optimal temperature 50\u00b0C, melting temperature ~63\u00b0C (16\u00b0C above wild-type). | Optimal temperature 40\u00b0C (body temp), thermostability not reported. | Kusumaningrum et al., ACS Omega/PMC, 2025 |\n| pH preferences | Optimal pH 9.0, active range 6.0-9.0 (moderate alkaline). | Optimal pH 10.0-11.0, active range 7.0-11.0 (high alkaline, GI adaptation). | Zhang et al., IJBM, 2024 |\n| Binding site organization | L-shaped cleft (~40 \u00c5), wobbly Trp185 aids substrate entry and \u03c0-\u03c0 interactions. | Enhanced activity against PET nanoparticles; favors fragmented substrates. | Austin et al., Nat. Commun., 2017 |\n| Fold and architecture | 290 amino acids, \u03b1/\u03b2 hydrolase fold; two disulfide bonds (Cys203-Cys239, Cys273-Cys289). | 275 amino acids, 30.18 kDa, predicted \u03b1/\u03b2 hydrolase fold with DLH domain. | Austin et al., Nat. Commun., 2017 |\n| Origin | Engineered variant of IsPETase from Ideonella sakaiensis; optimized by machine learning design. | Discovered from human gut microbiome (Acutalibacteraceae), actively expressed in 20% of human samples. | Zhang et al., IJBM, 2024 |", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\n| Ex. No. | Name | LCMS [M+H]+ |\n|---|---|---|\n| 17 | N-(2-(10-(2-((3,4-Difluorophenyl)amino)-2-oxoethyl)-2-ethyl-8-methoxy-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide | 558.2 |\n| 18 | N-(2-(2-Ethyl-10-(2-((4-fluorophenyl)amino)-2-oxoethyl)-7,8-dimethyl-4-oxo-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide | 538.2 |\n| 19 | N-(2-(2-Ethyl-9-methoxy-4-oxo-10-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-4,10-dihydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide | 590.2 |\n| 20 | N-(2-(8-Ethyl-6-oxo-10-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-6,10-dihydropyrido[2',3':4,5]imidazo[1,2-a]pyrimidin-7-yl)phenyl)acrylamide | 561.2 |\n\n*(1H NMR data at 500 MHz in DMSO-d6 are reported in the published table for Examples 17-20; the full shift lists are omitted here for brevity.)*\n\n### Example 21. 2-(2-Ethyl-4-oxo-3-(2-(vinylsulfonamido)phenyl)benzo[4,5]imidazo[1,2-a]pyrimidin-10(4H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide\n\n**Step 1. 2-(2-Ethyl-4-oxobenzo[4,5]imidazo[1,2-a]pyrimidin-10(4H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide.** A solution of 2-ethylbenzo[4,5]imidazo[1,2-a]pyrimidin-4(10H)-one (Example 1, Step 1) (300 mg, 1.41 mmol) in DMF (9.4 mL) was cooled to 0 \u00b0C before adding diisopropylethylamine (0.74 mL, 4.22 mmol), followed by 2-bromo-N-(4-(trifluoromethyl)phenyl)acetamide (595 mg, 2.11 mmol). The ice bath was then removed, and the reaction mixture was allowed to stir for 70 min. Upon completion, the reaction mixture was quenched with H2O, and the resulting mixture was extracted with EtOAc (20 mL) 3\u00d7. The organic phase was dried over Na2SO4, concentrated, and purified by flash column chromatography (0-100% hex/EtOAc) to provide the desired product (100 mg, 17%). LC-MS calculated for C21H18F3N4O2 (M+H)+: m/z=415.1; found 415.2.\n\n**Step 2. 2-(3-Bromo-2-ethyl-4-oxobenzo[4,5]imidazo[1,2-a]pyrimidin-10(4H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide.** To a mixture of 2-(2-ethyl-4-oxobenzo[4,5]imidazo[1,2-a]pyrimidin-10(4H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide (100 mg, 0.24 mmol) in DMF (2.4 mL) was added N-bromosuccinimide (47.2 mg, 0.27 mmol). The reaction mixture was allowed to stir at 25 \u00b0C for 15 min. Upon completion, the reaction was diluted w/water (30 mL), and the resulting precipitate was filtered and dried over vacuum for 3 h to provide the product (20.9 mg mmol, 18%). LC-MS calculated for C21H17BrF3N4O2 (M+H)+: m/z=493.0, 495.0; found 493.0, 495.1.\n\n**Step 3. 2-(3-(2-Aminophenyl)-2-ethyl-4-oxobenzo[4,5]imidazo[1,2-a]pyrimidin-10(4H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide.** To a solution of 2-(3-bromo-2-ethyl-4-oxobenzo[4,5]imidazo[1,2-a]pyrimidin-10(4H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide (21 mg, 0.043 mmol) in 1,4-dioxane (1.0 mL) and water (0.14 mL) was added 2-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)aniline (14.0 mg, 0.064 mmol), potassium phosphate tribasic (18.1 mg, 0.085 mmol), and XPhos Pd G4 (3.7 mg, 4.26 \u03bcmol). The reaction mixture was then degassed for 5 min before heating to 90 \u00b0C. The reaction was stirred for 1 h. Upon completion, the reaction mixture was quenched with H2O, and the resulting mixture was extracted with EtOAc (20 mL) 3\u00d7. The organic phase was dried over Na2SO4, filtered and concentrated in vacuo, and purified by flash column chromatography (0-100% hex/EtOAc) to provide the desired product (20.8 mg, 97%). LC-MS calculated for C27H23F3N5O2 (M+H)+: m/z=506.2; found 506.3.\n\n**Step 4. 2-(2-Ethyl-4-oxo-3-(2-(vinylsulfonamido)phenyl)benzo[4,5]imidazo[1,2-a]pyrimidin-10(4H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide.** To a solution of 2-(3-(2-aminophenyl)-2-ethyl-4-oxobenzo[4,5]imidazo[1,2-a]pyrimidin-10(4H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide (15 mg, 0.030 mmol) in DCM (0.5 mL) at 0 \u00b0C was added triethylamine (8.27 \u03bcl, 0.059 mmol) and ethenesulfonyl chloride (4.20 \u03bcl, 0.045 mmol). The reaction was allowed to stir for 30 min. Upon completion, the reaction contents were acidified w/TFA, diluted with MeCN (5 mL), filtered, and purified by prep-LCMS (XBridge C18 column, eluting with a gradient of acetonitrile/water containing 0.1% TFA, at flow rate of 60 mL/min) to give the product as a TFA salt (5.0 mg, 24% yield). LC-MS calculated for C29H25F3N5O4S (M+H)+: m/z=596.2; found 596.2. 1H NMR (400 MHz, DMSO-d6) \u03b4 10.92 (s, 1H), 8.91 (s, 1H), 8.47 (d, J=8.0 Hz, 1H), 7.82 (d, J=8.5 Hz, 2H), 7.73 (t, J=7.6 Hz, 3H), 7.59-7.50 (m, 1H), 7.50-7.30 (m, 3H), 7.23-7.12 (m, 2H), 6.73 (dd, J=16.5, 9.9 Hz, 1H), 6.09 (d, J=16.5 Hz, 1H), 5.98 (d, J=9.9 Hz, 1H), 5.40-5.23 (m, 2H), 2.30-2.23 (m, 2H), 1.08 (t, J=7.5 Hz, 3H). 19F NMR (376 MHz, DMSO-d6) \u03b4 \u221260.38, \u221276.50.\n\n### Example 22. N-(2-(2-ethyl-4-oxo-10-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-4,6,7,8,9,10-hexahydrobenzo[4,5]imidazo[1,2-a]pyrimidin-3-yl)phenyl)acrylamide\n\nThe procedure of Example 21 was followed, using 4,5,6,7-tetrahydro-1H-benzo[d]imidazol-2-amine in place of 1H-benzo[d]imidazol-2-amine (Step 1, Example 1); using acryloyl chloride in place of ethenesulfonyl chloride (Step 4, Example 21) to afford the title compound as a TFA salt. LC-MS calculated for C30H29F3N5O3 (M+H)+: m/z=564.2; found 564.2. 1H NMR (400 MHz, DMSO-d6) \u03b4 10.84 (s, 1H), 8.91 (s, 1H), 8.00 (d, J=8.1 Hz, 1H), 7.82 (d, J=8.5 Hz, 2H), 7.72 (d, J=8.5 Hz, 2H), 7.31 (ddd, J=8.5, 5.4, 3.6 Hz, 1H), 7.14 (d, J=4.4 Hz, 2H), 6.50 (dd, J=17.0, 10.1 Hz, 1H), 6.17 (dd, J=17.0, 2.2 Hz, 1H), 5.63 (dd, J=10.1, 2.2 Hz, 1H), 5.01 (s, 2H), 3.12 (s, 2H), 2.61 (s, 2H), 2.15 (q, J=7.5 Hz, 2H), 1.82 (s, 4H), 0.93 (t, J=7.5 Hz, 3H).\n\n### Example 23. 2-(2-Ethyl-4-oxo-3-(2-(vinylsulfonamido)phenyl)-6,7,8,9-tetrahydrobenzo[4,5]imidazo[1,2-a]pyrimidin-10(4H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide\n\nThe procedure of Example 21 was followed, using 4,5,6,7-tetrahydro-1H-benzo[d]imidazol-2-amine in place of 1H-benzo[d]imidazol-2-amine (Step 1, Example 1) to afford the title compound as a TFA salt. LC-MS calculated for C29H29F3N5O4S (M+H)+: m/z=600.2; found 600.2.", "# AlphaFraud \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n* **A positive control.** The page opens with the one thing the rest of the site does not: an *excellent* match. Human **[Artemis (DCLRE1C)](https://alphafraud.mdeller.com/examples#ex-6TT5_1)**, a genuinely novel post-cutoff nuclease (only 29% identity to anything AlphaFold trained on), was predicted blind to TM-score 0.996, lDDT 0.985 and 0.45 \u00c5 C\u03b1-RMSD at a justified mean pLDDT of 95, and AlphaFraud scores it FRAUD 0.02. It is the proof that the pipeline certifies good predictions, not just hunts for bad ones.\n* **An auto [\u201cexample of the week\u201d](https://alphafraud.mdeller.com/examples).** After every weekly run the pipeline picks that release\u2019s most instructive confidently-wrong catch and writes its panel deterministically from the database (no LLM), badged and dated.\n* **A [complete weekly archive](https://alphafraud.mdeller.com/examples).** Every PDB release week since the 2018 cutoff is listed: weeks with a catch expand to a full panel with its C\u03b1-deviation ribbon, and the quiet weeks (where AlphaFold was never overconfident) are recorded too, collapsed to the most recent by default.\n* **Curated deep-dives.** Hand-authored panels across the failure spectrum: ordinary single-chain globular folds AlphaFold gets *subtly* wrong (right fold, wrong conformation or assembly \u2014 the [PP2A scaffold](https://alphafraud.mdeller.com/examples#ex-8UWB_3), [STIM1](https://alphafraud.mdeller.com/examples#ex-6YEL_1), [Aha1](https://alphafraud.mdeller.com/examples#ex-7DME_1), [WWP2](https://alphafraud.mdeller.com/examples#ex-6RSS_1)) and the dramatic whole-fold catastrophes ([TMEM106B](https://alphafraud.mdeller.com/examples#ex-7U14_1), [alpha-synuclein](https://alphafraud.mdeller.com/examples#ex-9A1Q_1), [transthyretin](https://alphafraud.mdeller.com/examples#ex-9BZS_1), [factor X](https://alphafraud.mdeller.com/examples#ex-6Q9F_2)). Each panel carries live metrics, an interactive 3Dmol.js viewer (deviation colouring plus the AlphaFold \u201cghost\u201d) and a structure / mechanism / disease write-up.\n* **Linked from the scatter.** The curated examples are overlaid on the [home fraud-quadrant plot](https://alphafraud.mdeller.com/) as pulsing red (failure) and green (accurate) markers, each labelled with its protein name and clickable straight through to its panel.\n\nEach panel is summarised below \u2014 click any title to open its live, interactive version. Metrics: **TM-score** and **lDDT** (1.0 = identical), **C\u03b1-RMSD** (0 \u00c5 = identical), **pLDDT** (AlphaFold\u2019s own confidence, 0\u2013100), **FRAUD** (confidence-weighted error, higher = worse), **novelty** (100% = wholly unseen before AlphaFold\u2019s cutoff).\n\n### \ud83d\udfe2 Protein Artemis (DCLRE1C) \u00b7 [`6TT5_1`](https://alphafraud.mdeller.com/examples#ex-6TT5_1)\n\n| Field | Detail |\n| --- | --- |\n| **Take-home** | A novel post-cutoff nuclease AlphaFold predicted blind almost perfectly \u2014 the control proving AlphaFraud certifies good predictions, not only failures |\n| **AlphaFold vs experimental** | Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \u201cghost\u201d overlay |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-6TT5_1)** |\n| **KPIs** | TM **0.996** \u00b7 lDDT 0.985 \u00b7 pLDDT 95.2 \u00b7 C\u03b1-RMSD 0.45 \u00c5 \u00b7 FRAUD **0.02** \u00b7 novelty 71% |\n| **Stats** | 11 depositions, **0** confidently wrong \u00b7 SS-Q3 92.5% \u00b7 pLDDT\u2194lDDT +0.56 (calibrated) |\n| **Structure** | Metallo-\u03b2-lactamase + \u03b2-CASP nuclease domain with a Zn active site; the disordered regulatory C-terminus sits (correctly) outside the crystallised fold |\n| **Why AlphaFold gets it right** | A single, conserved globular domain with a deep MSA and one native conformation \u2014 AlphaFold\u2019s ideal regime; confidence is earned (PAE-overconfident 0.00) |\n| **Biology & disease** | NHEJ endonuclease that opens V(D)J hairpins with DNA-PKcs; loss of function causes radiosensitive SCID / Omenn; pursued as a radiosensitiser drug target |\n| **Key fact** | 29% identity to anything pre-cutoff yet GDT_TS 99.2 \u2014 genuine generalisation, not memorisation |\n\n### \ud83d\udd34 PP2A scaffold subunit A / PR65 (PPP2R1A) \u00b7 [`8UWB_3`](https://alphafraud.mdeller.com/examples#ex-8UWB_3)\n\n| Field | Detail |\n| --- | --- |\n| **Take-home** | Near-maximal confidence (pLDDT 97) on a flexible 15-HEAT solenoid whose curvature is wrong for 99% of residues |\n| **AlphaFold vs experimental** | Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \u201cghost\u201d overlay |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-8UWB_3)** |\n| **KPIs** | TM 0.293 \u00b7 lDDT 0.311 \u00b7 pLDDT 96.8 \u00b7 C\u03b1-RMSD 15.0 \u00c5 \u00b7 FRAUD 0.77 \u00b7 novelty 0% |\n| **Stats** | 26 depositions, 1 confidently wrong (this one) \u00b7 pLDDT\u2194lDDT \u22120.13 (confidence carries no signal) |\n| **Structure** | 15 tandem HEAT repeats forming a horseshoe \u03b1-solenoid \u2014 a spring, not a fixed ruler |\n| **Why AlphaFold gets it wrong** | Local helices are right, but tiny per-repeat hinge errors compound across 15 repeats into a grossly wrong global shape \u2014 a curvature failure, not a fold failure |\n| **Biology & disease** | Sole scaffold of the PP2A tumour-suppressor phosphatase; cancer hotspots (P179R, R183, S256F\u2026) cluster in the B56-binding repeats |\n| **Key fact** | The source paper needed an experimental trimer structure to guide the prediction |\n\n### \ud83d\udd34 Stromal interaction molecule 1 / CC1 (STIM1) \u00b7 [`6YEL_1`](https://alphafraud.mdeller.com/examples#ex-6YEL_1)", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nIon-exchange chromatography of LIFR-FN3 proved unsuccessful using both Mono-Q and Mono-S columns in the pH range 4\u201310 (theoretical pI 7.2, observed pI 5\u20136). No specific binding of protein was observed and only aggregates were recovered from the column at high salt concentrations.\n\nFigure 4-11 below shows FPLC elution profile of Superose-6 analytical grade gel filtration column with panel (a) showing the elution profile of molecular weight standards with five main peaks running at 2000, 150, 66, 29 and 12 kDa. Panel (b) shows the elution profile of affinity purified LIF in which one main peak is observed at ~91 minutes (~36 kDa) and comparison with lane 6 of panel (e) shows that this peak runs at ~25 kDa on denaturing SDS-PAGE. Inspection of panel (c) shows that LIFR-FN3 is eluted as one main peak at ~67 minutes (~912 kDa) and comparison with lane 3 of panel (e) shows that this species runs as one major band on denaturing SDS-PAGE at a molecular weight of ~116 kDa. Panel (d) shows the putative LIF-LIFR-FN3 complex is eluted as a single species at ~62 minutes (~1622 kDa) and comparison with lane 2 of panel (e) shows that this peak runs as two major bands on denaturing SDS-PAGE at molecular weights of ~25 kDa and ~116 kDa. Panel (e) shows SDS-PAGE of pooled and concentrated fractions corresponding to peaks underlined by numbered red bars. Lane 1 contains void volume eluate in which no protein bands are observed and lane 2 contains the putative LIF-LIFR-FN3 complex peak eluted at ~62 minutes in which two bands at molecular weights of ~25 kDa and ~116 kDa can be seen. Lane 3 contains the LIFR-FN3 peak eluted at ~67 minutes and a single band at ~116 kDa can be observed. Lanes 5, 6 and 7 contain the LIF peak eluted at ~91 minutes and a single band is observed at ~25 kDa. Lanes 8, 9 and 10 contain control samples of putative LIFR-FN3-LIF complex, LIF and LIFR-FN3, respectively. Lane 11 contains broad range molecular weight rainbow markers running at 135, 80, 43, 32 and 14 kDa (Bio-Rad).\n\n*Figure 4-11.* Chromatogram showing FPLC elution profile of Superose-6 analytical grade gel filtration column with (a) molecular weight standards, (b) affinity purified LIF, (c) affinity purified deglycosylated LIFR-FN3 and (d) putative LIF-LIFR-FN3 complex. 100 \u03bcl samples were loaded at a flow rate of 2 ml/min and the absorbance monitored at 280 nm. Complex was formed by mixing LIF and LIFR-FN3 in molar ratios of 10:1. LIF is eluted as one main peak at ~91 minutes (~35.6 kDa) and LIFR-FN3 is eluted as one main peak at ~67 minutes (>150 kDa). The putative LIF-LIFR-FN3 complex is eluted as a single peak at ~62 minutes (>150 kDa). SDS-PAGE (e) of pooled fractions corresponding to peaks underlined by numbered red bars. Lane 1 contains a sample of void volume eluate and lane 2 contains putative LIF-LIFR-FN3 complex eluted at ~62 minutes. Lane 3 contains LIFR-FN3 eluted at ~67 minutes. Lanes 5, 6 and 7 contain samples of LIF eluted in peaks around 91 minutes. Lanes 8, 9 and 10 contain control samples of putative complex, LIF and LIFR-FN3, respectively. Lane 11 contains broad range molecular weight rainbow markers running at 135, 80, 43, 32 and 14 kDa (Bio-Rad).\n\n## 4.4 Discussion of Expression, Purification and Characterisation of LIFR\n\nThe aim of this project was to express and purify high levels of soluble LIFR-FN3 protein suitable for crystallographic studies; such requirements are typically 10 mg or greater at a purity of 90 % or better. This was not achieved due to several technical problems although critical steps towards this goal have been established.\n\nExpression in CHO cells resulted in very low protein yields as assessed by ELISA, as shown in Figure 4-3, and scale-up using a stirred-tank bioreactor system and roller bottles proved unsuccessful. Poor cell growth and adherence was observed, compared with wildtype cells, so suggesting that LIFR-FN3-3C-Fc is affecting cellular morphology and phenotype. Typically ~1 mg/L of glycosylated LIFR-FN3-3C-Fc was obtained after 3\u20134 weeks incubation and as such this system was deemed financially and technically unfeasible for large-scale protein production. Sufficient material was produced using this system for the generation of antibodies and development of preliminary purification protocols. Possible improvements to the pEE14 expression system include the use of alternative promoter-enhancer elements such as the adenovirus major late promoter in combination with the SV40 enhancer (Wong et al., 1985).\n\nFigure 4-3 demonstrates the addition of sodium butyrate to CHO cell growth medium increases the levels of expression of LIFR-FN3 by up to 30 %. Sodium butyrate has previously been shown to improve expression from recombinant plasmids in mammalian cells due to increases in both enhancer-dependent transcription and stable transformation efficiencies (Gorman et al., 1983). These effects have been shown to occur in a wide variety of mammalian cells including Friend erythroleukaemic cells (Reeves and Cserjesi, 1979) and CHO cells (D'Anna et al., 1980). The concentration-dependent effects of sodium butyrate in CHO cells are thought to involve inner histone acetylation and histone H1 phosphorylation (D'Anna et al., 1980). Both of these events function in controlling high order chromatin structure and as such regulate chromatin transcriptional activity. This does not appear to be a protein-specific or cell-specific phenomenon, so the use of sodium butyrate may prove to be important in the development of other mammalian expression systems.", "# Unlocking Membrane Protein Structures: The Crystallization Revolution Reshaping Drug Discovery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Concept | Description | Key References |\n| --- | --- | --- |\n| Lipidic Cubic Phase (LCP) | Bicontinuous liquid crystalline matrix formed from monoolein and water that mimics native membrane environment, enabling crystallization of membrane proteins in lipid bilayer context with mean resolution of 2.5 \u00c5 | [Lipidic Cubic Phase-Induced Membrane Protein Crystallization](https://pubs.acs.org/doi/10.1021/acs.cgd.7b00519), [Lipidic Cubic Phase Technologies](https://pmc.ncbi.nlm.nih.gov/articles/PMC3164297/) |\n| GPCR Structure Explosion | G protein-coupled receptor structures increased from 17 in 2012 to 1,716 by 2025, with 78% of 2021 structures determined by cryo-EM rather than X-ray crystallography | [GPCRdb](https://gpcrdb.org), [Structure determination of GPCRs: cryo-EM compared with X-ray](https://portlandpress.com/biochemsoctrans/article/49/5/2345/229854/) |\n| Fusion Protein Engineering | Replacement of flexible intracellular loop 3 with T4 lysozyme or BRIL to increase thermostability, reduce conformational heterogeneity, and provide additional polar surface for crystal contacts | [Modified T4 lysozyme fusion proteins](https://pmc.ncbi.nlm.nih.gov/articles/PMC4408211/), [Fusion Partner Toolchest](https://pmc.ncbi.nlm.nih.gov/articles/PMC3375611/) |\n| Bicelle Crystallization | Temperature-sensitive lipid/amphiphile bilayer disks that remain liquid at low temperature for easy handling but form gel-like phase at 37\u00b0C to facilitate crystallization in native-like environment | [Crystallizing membrane proteins using lipidic bicelles](https://pmc.ncbi.nlm.nih.gov/articles/PMC3264687/), [Mechanisms of membrane protein crystallization in bicelles](https://www.nature.com/articles/s41598-022-13945-0) |\n| MicroED Integration | Microcrystal electron diffraction technique enabling atomic resolution structure determination from submicron crystals using electron beams, compatible with LCP and bicelle-grown crystals through FIB milling | [Studying membrane proteins with MicroED](https://portlandpress.com/biochemsoctrans/article-abstract/50/1/231/230823/), [Robust approach for MicroED sample preparation](https://www.nature.com/articles/s41467-023-36733-4) |\n| Cryo-EM Resolution Revolution | Single-particle cryo-electron microscopy improved mean resolution from 4.2 \u00c5 (2015) to 3.6 \u00c5 (2019) for membrane proteins through detector advances, image processing, and membrane mimetic systems like nanodiscs | [10-Year Meta-Analysis Of Membrane Protein Structural Biology](https://pmc.ncbi.nlm.nih.gov/articles/PMC7856071/), [Single-particle cryogenic electron microscopy structure determination](https://pubmed.ncbi.nlm.nih.gov/40228430/) |\n| Thermostabilization (StaR) | Systematic mutagenesis approach to increase detergent stability and lock receptors in specific conformational states, exemplified by 21\u00b0C melting temperature increase for \u03b2\u2081-adrenergic receptor with 6 mutations | [Conformational thermostabilization of \u03b2\u2081-adrenergic receptor](https://www.pnas.org/doi/10.1073/pnas.0711253105), [Coiled-coil-based design strategy for thermostabilization](https://www.nature.com/articles/s41598-023-36855-1) |", "# marcdeller.com: Microplastics in Humans\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n## Therapeutic Vision: Tackling Microplastics in humans\n\nMicroplastics now show up in human blood, placental tissue, atherosclerotic plaque, and the brain. The early literature points toward chronic inflammation, endocrine disruption, and metabolic stress as likely consequences. At [Elora Therapeutics](https://eloratherapeutics.com), we are building engineered PET-degrading enzymes designed to operate inside the body, drawing on the natural PETase and MHETase enzymes first characterized in *Ideonella sakaiensis*. It is frontier science with no playbook, and a protein-engineering problem at its heart; designing enzymes that are stable, selective, and safe *in vivo*.\n\n* **Inflammation Trigger:** Microplastics cause chronic inflammation that harms cellular function and overall health.\n* **Hormone Disruption:** These pollutants interfere with endocrine signaling, leading to hormonal imbalances.\n* **Metabolic Dysfunction:** Accumulation of microplastics is linked to metabolic disorders affecting energy and body regulation.\n* **Innovative Treatment:** Elora Therapeutics is developing enzymes to safely break down and eliminate microplastics from the human body, pioneering new therapeutic possibilities.\n\nTop Posts\n\n### [Vibe Coding MicroPlastic Blaster](https://marcdeller.com/vibe-coding-microplastic-blaster/)\n\n[Vibe Coding MicroPlastic Blaster](https://marcdeller.com/vibe-coding-microplastic-blaster/)\n\n### [Elora Therapeutics: Treating Microplastics in Humans](https://marcdeller.com/elora/)\n\n[Elora Therapeutics: Treating Microplastics in Humans](https://marcdeller.com/elora/)\n\n### [Decoding the Molecular Machinery Behind Plastic Degradation](https://marcdeller.com/decoding-the-molecular-machinery-behind-plastic-degradation/)\n\n[Decoding the Molecular Machinery Behind Plastic Degradation](https://marcdeller.com/decoding-the-molecular-machinery-behind-plastic-degradation/)\n\n### [Cellular Uptake of Microplastics: Pathways, Persistence, and Therapeutic Potential for Intracellular Targeting of PETase](https://marcdeller.com/cellular-uptake-of-microplastics-pathways-persistence-and-therapeutic-potential/)\n\n[Cellular Uptake of Microplastics: Pathways, Persistence, and Therapeutic Potential for Intracellular Targeting of PETase](https://marcdeller.com/cellular-uptake-of-microplastics-pathways-persistence-and-therapeutic-potential/)\n\n### [Clearing the Microscopic Threat: Can Plastic-Eating Enzymes Treat Liver Disease such as MASLD?](https://marcdeller.com/clearing-the-microscopic-threat-can-plastic-eating-enzymes-treat-liver-disease/)\n\n[Clearing the Microscopic Threat: Can Plastic-Eating Enzymes Treat Liver Disease such as MASLD?](https://marcdeller.com/clearing-the-microscopic-threat-can-plastic-eating-enzymes-treat-liver-disease/)\n\nMicroplastics\n\n### [Seeing Micro and Nanoplastics in Living Tissues](https://marcdeller.com/seeing-plastic-in-living-tissue/)\n\nMicroplastics\n\n### [Microplastics and Cancer: The Mounting Evidence via Literature Review](https://marcdeller.com/microplastics-and-cancer-the-mounting-evidence/)\n\nBlog\n\n### [Plastic, Persistent, and Potentially Carcinogenic](https://marcdeller.com/plastic-persistent-and-potentially-carcinogenic/)\n\nBlog\n\n### [An Enzyme That Eats Plastic \u2014 Inside Your Blood](https://marcdeller.com/kth/)\n\nBlog\n\n### [Molecular Velcro: Catching What We Cannot See](https://marcdeller.com/molecular-velcro-catching-what-we-cannot-see/)\n\nMicroplastics\n\n### [PETase: Nature\u2019s Solution to a Human-Made Problem](https://marcdeller.com/petase-natures-solution-to-a-human-made-problem/)\n\nBlog\n\n### [Vibe Coding MicroPlastic Blaster](https://marcdeller.com/vibe-coding-microplastic-blaster/)\n\nMicroplastics\n\n### [Elora Therapeutics: Treating Microplastics in Humans](https://marcdeller.com/elora/)\n\nMicroplastics\n\n### [Overcoming Biological Barriers: The Evolution of Oral Protein and Peptide Drug Delivery](https://marcdeller.com/overcoming-biological-barriers-the-evolution-of-oral-protein-and-peptide-drug-delivery/)\n\nBlog\n\n### [Decoding the Molecular Machinery Behind Plastic Degradation](https://marcdeller.com/decoding-the-molecular-machinery-behind-plastic-degradation/)\n\nBlog\n\n### [Cellular Uptake of Microplastics: Pathways, Persistence, and Therapeutic Potential for Intracellular Targeting of PETase](https://marcdeller.com/cellular-uptake-of-microplastics-pathways-persistence-and-therapeutic-potential/)\n\nBlog\n\n### [Unlocking the Power of Protein Engineering: FDA-Approved Enzyme Therapies Transform Rare Disease Treatment](https://marcdeller.com/unlocking-the-power-of-protein-engineering-fda-approved-enzyme-therapies-transform-rare-disease-treatment/)\n\nBlog\n\n### [Engineering Evolution: How FAST-PETase and Other Variants Are Transforming Plastic Biodegradation](https://marcdeller.com/engineering-evolution-how-fast-petase-and-other-variants-are-transforming-plastic-biodegradation/)\n\nBlog\n\n### [HGMP01 and FAST-PETase: Decoding Two Paths to Plastic Degradation \u2013 Engineered versus Natural PETase Enzymes](https://marcdeller.com/hgmp01-and-fast-petase-decoding-two-paths-to-plastic-degradation-engineered-versus-natural-petase-enzymes/)\n\nBlog\n\n### [HGMP01 and other Hidden Guardians: How the Human Gut Fights Back Against Plastic Pollution](https://marcdeller.com/hgmp01-and-other-hidden-guardians-how-the-human-gut-fights-back-against-plastic-pollution/)\n\nBlog\n\n### [Flow Cytometry and Microplastic Detection in Human Tissues](https://marcdeller.com/flow-cytometry-revolutionizes-microplastic-detection-in-human-tissues/)\n\nBlog\n\n### [Clearing the Microscopic Threat: Can Plastic-Eating Enzymes Treat Liver Disease such as MASLD?](https://marcdeller.com/clearing-the-microscopic-threat-can-plastic-eating-enzymes-treat-liver-disease/)\n\nBlog\n\n### [The Invisible Invasion: How Microscopic Plastics Are Infiltrating Human Organ Systems](https://marcdeller.com/the-invisible-invasion-how-microscopic-plastics-are-infiltrating-human-organ-systems/)\n\nBlog", "# Flow-Based Models: A New Paradigm for Protein Structure Generation\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: blog).*\n\n* Flow-based generative models like Proteina are an emerging alternative to diffusion models for designing new protein structures.\n* These models learn a direct and controllable transformation from a simple distribution to a complex protein structure, allowing for hierarchical control over the final fold.\n* Proteina has achieved state-of-the-art results, generating diverse and designable proteins at an unprecedented scale, expanding the toolkit for protein engineering.\n\n1. Proteina: Scaling Flow-based Protein Structure Generative Models: <https://arxiv.org/abs/2503.00710>\n2. Proteina is a new large-scale flow-based protein backbone: <https://github.com/NVIDIA-Digital-Bio/proteina>\n\nIn the world of generative AI, diffusion models have become famous for their ability to create stunningly realistic images by starting with noise and gradually refining it into a coherent picture. A similar approach has been a game-changer for protein design. However, a powerful alternative paradigm is now emerging: flow-based generative models. These models offer a different, and in some ways more direct, path to designing novel proteins.\n\nInstead of a noisy, iterative process, flow-based models learn a direct and reversible transformation between a simple, easy-to-sample distribution (like a standard normal distribution) and the complex distribution of real protein structures. A new, large-scale model called **Proteina** is at the forefront of this approach. It uses a sophisticated flow-matching process to generate protein backbones, offering exquisite control over the final product. The model can be conditioned on hierarchical labels, allowing a designer to guide the generation process at a high level (e.g., \u201ccreate a mostly-alpha-helical protein\u201d) or a very specific one (e.g., \u201ccreate a protein with this exact fold\u201d).\n\nProteina achieves state-of-the-art performance in *de novo* protein design, producing a wide diversity of novel and designable proteins up to 800 residues in length\u2014a significant leap in scale. By learning a direct \u201cflow\u201d to the desired structure, these models can offer a more intuitive and controllable design process. This technology represents a significant addition to the protein engineer\u2019s toolkit, providing a new way to explore the vast landscape of possible protein structures and create novel bio-machines with tailored architectures.", "# Citation Giants: How the Most-Cited Papers of the 21st Century Are Reshaping Science\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-11 on marcdeller.com (categories: blog).*\n\n* [Citation Giants](https://www.nature.com/articles/d41586-025-01125-9)\n\n* **Methodological Dominance**: Research tools and software papers consistently outperform discovery papers in citation counts\n* **AI Architecture Impact**: Deep learning frameworks like ResNets and Transformers enable countless applications across disciplines\n* **Pandemic Acceleration**: COVID-19 research achieved unprecedented citation velocity through global urgency and collaboration\n* **Cross-disciplinary Integration**: The most cited papers increasingly bridge computer science with biology and medicine\n* **Infrastructure Over Discovery**: Citation giants provide computational and analytical foundations rather than single breakthroughs\u200b\n* **Global Collaboration Patterns**: Highly cited papers reflect international research networks and shared methodological standards\u200b\n\n1. Exclusive: the most-cited papers of the twenty-first century: [van Noorden, R., et al., Nature](https://www.nature.com/articles/d41586-025-01125-9)\n2. Deep Residual Learning for Image Recognition: [He, K., et al., Computer Vision and Pattern Recognition](https://www.cv-foundation.org/openaccess/content_cvpr_2016/html/He_Deep_Residual_Learning_CVPR_2016_paper.html)\n3. Artificial intelligence and machine learning to fight COVID-19: [Lalmuanawma, S., et al., Physiological Genomics](https://journals.physiology.org/doi/abs/10.1152/physiolgenomics.00029.2020)\n4. The 100 most-cited articles in COVID-19: a bibliometric analysis: [Zyoud, S.H., et al., BMC Public Health](https://pmc.ncbi.nlm.nih.gov/articles/PMC11293834/)\n5. Research Trends On Academic Writing From 2020 To 2025: [Zubir, F., et al., International Journal of Modern Education](https://gaexcellence.com/ijmoe/article/download/5767/5298/19510)\n6. Scientific breakthroughs: 2025 emerging trends to watch: [CAS Editorial Team, et al., CAS Insights](https://www.cas.org/resources/cas-insights/scientific-breakthroughs-2025-emerging-trends-watch)\n\nThe scholarly landscape has witnessed a dramatic transformation in the first quarter of the 21st century. While groundbreaking discoveries like mRNA vaccines, CRISPR gene editing, the Higgs boson, and gravitational wave detection dominate headlines, they surprisingly don\u2019t appear among the most-cited papers published since 2000. Instead, the citation champions represent the essential infrastructure of modern science: artificial intelligence architectures, research methodologies, statistical software, and clinical data that enable countless other discoveries.\n\nMicrosoft\u2019s 2016 paper on deep residual learning networks (ResNets) claims the crown as the century\u2019s most-cited work, accumulating over 254,000 citations according to Google Scholar. This paper solved a fundamental problem in neural network training\u2014signal degradation through many layers\u2014enabling networks with 150 layers compared to the previous standard of 30. The ResNet architecture became foundational for AI breakthroughs from AlphaGo\u2019s game mastery to AlphaFold\u2019s protein structure predictions and ultimately ChatGPT\u2019s language modeling capabilities.\n\nThe COVID-19 pandemic accelerated citation patterns in unprecedented ways. Guan Wei-jie\u2019s clinical characterization of coronavirus disease garnered over 16,000 citations within just three years, demonstrating how urgent global health crises can rapidly elevate research visibility. Machine learning applications for COVID-19 diagnosis, drug discovery, and epidemiological modeling became citation magnets as researchers worldwide sought computational solutions to pandemic challenges.\n\nCancer research maintains its position as a citation powerhouse through foundational works that continue accumulating references decades after publication. Douglas Hanahan\u2019s \u201cHallmarks of Cancer\u201d framework and WHO\u2019s GLOBOCAN cancer statistics serve as essential starting points for virtually every cancer study, while emerging CRISPR-based therapeutics represent the field\u2019s future direction.\n\nPerhaps most tellingly, research software and methodological papers dominate the citation landscape. George Sheldrick\u2019s SHELX crystallography software, Virginia Braun and Victoria Clarke\u2019s thematic analysis methodology, and various bioinformatics tools like DESeq2 demonstrate how methodological innovations become the workhorses of scientific progress. These papers succeed not through revolutionary discoveries but by providing reliable, accessible tools that thousands of researchers depend upon.\n\nThe concentration of highly cited papers in artificial intelligence, computational biology, and research methodology reveals a profound shift toward interdisciplinary, data-driven science. Today\u2019s most influential research doesn\u2019t just make discoveries\u2014it creates the computational and analytical foundations that enable entire fields to advance. This pattern suggests that 21st-century science increasingly values infrastructure over individual breakthroughs, with the most cited papers serving as the digital scaffolding supporting the modern research enterprise.", "# marcdeller.com: Vibe Coding\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n## Vibe Coding: build fast and solve real problems\n\nVibe coding is how I prototype at the speed of thought: short, single-file tools built in a quick conversation with an AI assistant, then refined at the bench. Each one scratches a real itch in structural biology or drug discovery, from harvesting a Boltz-2 output tree to profiling protein-ligand contacts or wrangling SMILES into a tidy grid. They are deliberately small and unpolished: sketches and proofs of concept, seeds for the larger applications I am building to tackle harder problems. Latest projects include: [AlphaFraud](https://alphafraud.mdeller.com), [BoltzMaker](https://marcdeller.com/boltzmaker-vibe-coding-a-campaign-scale-boltz-2-pipeline/), [ChemSage](https://marcdeller.com/chemsage/), and [MicroPlastics Blaster.](https://marcdeller.com/vibe-coding-microplastic-blaster/)\n\n* **Built fast, by conversation:** Each tool begins as a quick prompt to an AI assistant and a single file of code.\n* **Small by design:** These are focused prototypes, not finished products, each solving one narrow problem well.\n* **Rooted in real work:** Every script answers a genuine need from my structural biology and drug discovery workflows.\n* **Stepping stones, not destinations:** Today's throwaway utility is tomorrow's module inside a larger application.\n* **Open and reproducible:** Logged, auditable, and shared so others can run them, break them, and make them better.\n\nTop Vibes\n\n### [**GPCR Binning** \u2013 Structural binning and organisation utilities for GPCR datasets](https://marcdeller.com/structural-binning-and-organisation-utilities-for-gpcr-datasets/)\n\n[**GPCR Binning** \u2013 Structural binning and organisation utilities for GPCR datasets](https://marcdeller.com/structural-binning-and-organisation-utilities-for-gpcr-datasets/)\n\n### [**SiteHealthInspector** \u2013 A pure-Bash, colour-coded website health audit](https://marcdeller.com/a-pure-bash-colour-coded-website-health-audit-no-install-no-dependencies/)\n\n[**SiteHealthInspector** \u2013 A pure-Bash, colour-coded website health audit](https://marcdeller.com/a-pure-bash-colour-coded-website-health-audit-no-install-no-dependencies/)\n\n### [**Smiles2Grid** \u2013 Turn a JSON of SMILES into a paginated, property-annotated PDF grid with RDKit](https://marcdeller.com/turn-a-json-of-smiles-into-a-paginated-property-annotated-pdf-grid-with-rdkit/)\n\n[**Smiles2Grid** \u2013 Turn a JSON of SMILES into a paginated, property-annotated PDF grid with RDKit](https://marcdeller.com/turn-a-json-of-smiles-into-a-paginated-property-annotated-pdf-grid-with-rdkit/)\n\n### [**AlphaFraud** \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved](https://marcdeller.com/alphafraud-weekly-watcher-newly-deposited-human-pdb-structures-vs-their-blind-alphafold-predictions-a-running-test-of-whether-protein-folding-is-solved/)\n\n[**AlphaFraud** \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved](https://marcdeller.com/alphafraud-weekly-watcher-newly-deposited-human-pdb-structures-vs-their-blind-alphafold-predictions-a-running-test-of-whether-protein-folding-is-solved/)\n\n### [**AnalyzeBoltz** \u2013 Harvest the sprawling Boltz2 output tree, tidy summary, CSV, and CIF output folder](https://marcdeller.com/analyze_boltz-py/)\n\n[**AnalyzeBoltz** \u2013 Harvest the sprawling Boltz2 output tree, tidy summary, CSV, and CIF output folder](https://marcdeller.com/analyze_boltz-py/)\n\nVibe Code\n\n### [**ChatPDB** \u2013 A protein-structure-literate AI assistant for the Protein Data Bank: RAG for facts, a QLoRA-tuned model for behaviour](https://marcdeller.com/elementor-15618/)\n\nVibe Code\n\n### [**AlphaFraud** \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved](https://marcdeller.com/alphafraud-weekly-watcher-newly-deposited-human-pdb-structures-vs-their-blind-alphafold-predictions-a-running-test-of-whether-protein-folding-is-solved/)\n\nVibe Code\n\n### [**BoltzMaker**: Boltz2 campaign-scale structure and affinity prediction, and binding analysis](https://marcdeller.com/boltzmakervibe/)\n\nVibe Code\n\n### [**ChemSage** \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior](https://marcdeller.com/chemsagevibe/)\n\nVibe Code\n\n### [**WordPressID** \u2013 Every post and page from any WordPress site, to CSV and screen, in one command](https://marcdeller.com/every-post-and-page-from-any-wordpress-site-to-csv-and-screen-in-one-command/)\n\nVibe Code\n\n### [**Smiles2Grid** \u2013 Turn a JSON of SMILES into a paginated, property-annotated PDF grid with RDKit](https://marcdeller.com/turn-a-json-of-smiles-into-a-paginated-property-annotated-pdf-grid-with-rdkit/)\n\nVibe Code\n\n### [**SiteHealthInspector** \u2013 A pure-Bash, colour-coded website health audit](https://marcdeller.com/a-pure-bash-colour-coded-website-health-audit-no-install-no-dependencies/)\n\nVibe Code\n\n### [**SiteDoctor** \u2013 Full diagnosis, health, content, and SEO for any WordPress site](https://marcdeller.com/one-pass-full-diagnosis-health-content-and-seo-for-any-wordpress-site/)\n\nVibe Code\n\n### [**Boltz2Run** \u2013 Every run logged, timed, and reproducible](https://marcdeller.com/a-minimal-auditable-wrapper-for-boltz-1-2-every-run-logged-timed-and-reproducible/)\n\nVibe Code\n\n### [**PocketTide** \u2013 An interactive dashboard for GLP-1R structures](https://marcdeller.com/an-interactive-dashboard-for-glp-1r-structures/)\n\nVibe Code\n\n### [**Microplastics Blaster** \u2013 Deploy enzymes, destroy microplastics, save the body](https://marcdeller.com/deploy-enzymes-destroy-microplastics-save-the-body/)\n\nVibe Code\n\n### [**GPCR Binning** \u2013 Structural binning and organisation utilities for GPCR datasets](https://marcdeller.com/structural-binning-and-organisation-utilities-for-gpcr-datasets/)\n\nVibe Code", "# Crystal Structure of Human Oncostatin M\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [1EVS](https://www.rcsb.org/structure/1EVS)\n* **Release Date:** 2000-09-13\n* **Method:** X-RAY DIFFRACTION 2.2 \u00c5\n* **Organisms:** Homo sapiens\n* **Macromolecules:** ONCOSTATIN M (protein)\n* **Publication:** [Deller, M.C., et al. (2000). Crystal Structure and Functional Dissection of the Cytostatic Cytokine Oncostatin M. Structure, 8, 863-874.](https://doi.org/10.1016/s0969-2126(00)00176-3)\n\n**Description:** This foundational structure of human Oncostatin M\u2014a cytokine that inhibits tumor cell growth while promoting inflammatory responses\u2014reveals the four-helix bundle topology and maps the site 2 (gp130-binding) and site 3 (LIFR/OSMR-binding) receptor epitopes to atomic resolution. The structure-function dissection provides key insights into cytokine-receptor specificity within the gp130 signaling family, with implications for cancer and inflammatory disease biology.", "# 2026 \u2013 Learn to code with AI\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-28 on marcdeller.com (categories: certificates).*\n\n* [Link to certificate](https://www.coursera.org/account/accomplishments/verify/0X9WKS0AH5P9)\n\n* **Award Details:** Scrimba, through Coursera, issued this certificate on March 25, 2026.\n* **Course Focus:** It recognizes completion of a course on learning programming with AI assistance.\n* **Career Value:** This matters because AI-assisted coding can accelerate software development and reduce friction.\n* **Practical Use:** It helps scientists build scripts, small apps, and automation tools more quickly.", "# pants.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n| Trained on | Positives | Negatives | Clusters | AUC | Baseline | Reading |\n| --- | --- | --- | --- | --- | --- | --- |\n| measured-only  pet_activity \u00b7 2026-08-06 | 308 | 385 | \u2014 | 0.774 | 0.565 | clears the composition baseline by a real margin |\n| measured-only  pet_activity \u00b7 2026-08-06 | 308 | 385 | 59 | 0.845 | 0.565 | clears the composition baseline by a real margin |\n| mixed  pet_activity \u00b7 2026-08-06 | 852 | 385 | 97 | 0.835 | 0.565 | trained on similarity-derived labels |\n| measured-only  pet_activity \u00b7 2026-08-06 | 308 | 385 | \u2014 | 0.739 | 0.565 | clears the composition baseline by a real margin |\n| measured-only  pet_activity \u00b7 2026-08-06 | 308 | 385 | 48 | 0.790 | 0.565 | clears the composition baseline by a real margin |\n| mixed  pet_activity \u00b7 2026-08-06 | 852 | 385 | 65 | 0.859 | 0.565 | trained on similarity-derived labels |\n| measured-only  pet_activity \u00b7 2026-08-06 | 302 | 241 | \u2014 | 0.845 | 0.803 | barely clears amino-acid composition |\n| measured-only  pet_activity \u00b7 2026-08-06 | 302 | 241 | 60 | 0.947 | 0.803 | clears the composition baseline by a real margin |\n| mixed  pet_activity \u00b7 2026-08-06 | 789 | 241 | 97 | 0.972 | 0.803 | trained on similarity-derived labels |\n| measured-only  pet_activity \u00b7 2026-08-06 | 302 | 241 | \u2014 | 0.857 | 0.803 | barely clears amino-acid composition |\n| measured-only  pet_activity \u00b7 2026-08-06 | 302 | 241 | 46 | 0.949 | 0.803 | clears the composition baseline by a real margin |\n| mixed  pet_activity \u00b7 2026-08-06 | 789 | 241 | 61 | 0.981 | 0.803 | trained on similarity-derived labels |\n| measured-only  pet_activity \u00b7 2026-08-06 | 304 | 241 | \u2014 | 0.860 | 0.736 | clears the composition baseline by a real margin |\n| measured-only  pet_activity \u00b7 2026-08-06 | 304 | 241 | 61 | 0.942 | 0.736 | clears the composition baseline by a real margin |\n| mixed  pet_activity \u00b7 2026-08-06 | 791 | 241 | 97 | 0.975 | 0.736 | trained on similarity-derived labels |\n| measured-only  pet_activity \u00b7 2026-08-06 | 304 | 241 | \u2014 | 0.782 | 0.736 | barely clears amino-acid composition |\n| measured-only  pet_activity \u00b7 2026-08-06 | 304 | 241 | 46 | 0.921 | 0.736 | clears the composition baseline by a real margin |\n| mixed  pet_activity \u00b7 2026-08-06 | 791 | 241 | 60 | 0.967 | 0.736 | trained on similarity-derived labels |\n| measured-only  pet_activity \u00b7 2026-08-06 | 304 | 241 | \u2014 | 0.860 | 0.736 | clears the composition baseline by a real margin |\n| measured-only  pet_activity \u00b7 2026-08-06 | 304 | 241 | 61 | 0.942 | 0.736 | clears the composition baseline by a real margin |\n| mixed  pet_activity \u00b7 2026-08-06 | 791 | 241 | 97 | 0.975 | 0.736 | trained on similarity-derived labels |\n| measured-only  pet_activity \u00b7 2026-08-06 | 304 | 241 | \u2014 | 0.782 | 0.736 | barely clears amino-acid composition |\n| measured-only  pet_activity \u00b7 2026-08-06 | 304 | 241 | 46 | 0.921 | 0.736 | clears the composition baseline by a real margin |\n| mixed  pet_activity \u00b7 2026-08-06 | 791 | 241 | 60 | 0.967 | 0.736 | trained on similarity-derived labels |\n| measured-only  pet_activity \u00b7 2026-08-06 | 304 | 241 | \u2014 | 0.860 | 0.736 | clears the composition baseline by a real margin |\n| measured-only  pet_activity \u00b7 2026-08-06 | 304 | 241 | 61 | 0.942 | 0.736 | clears the composition baseline by a real margin |\n| mixed  pet_activity \u00b7 2026-08-06 | 791 | 241 | 97 | 0.975 | 0.736 | trained on similarity-derived labels |\n| measured-only  pet_activity \u00b7 2026-08-06 | 304 | 241 | \u2014 | 0.782 | 0.736 | barely clears amino-acid composition |\n| measured-only  pet_activity \u00b7 2026-08-06 | 304 | 241 | 46 | 0.921 | 0.736 | clears the composition baseline by a real margin |\n| mixed  pet_activity \u00b7 2026-08-06 | 791 | 241 | 60 | 0.967 | 0.736 | trained on similarity-derived labels |\n| measured-only  pet_activity \u00b7 2026-08-06 | 304 | 241 | \u2014 | 0.860 | 0.736 | clears the composition baseline by a real margin |\n| measured-only  pet_activity \u00b7 2026-08-06 | 304 | 241 | 61 | 0.942 | 0.736 | clears the composition baseline by a real margin |\n| mixed  pet_activity \u00b7 2026-08-06 | 791 | 241 | 97 | 0.975 | 0.736 | trained on similarity-derived labels |\n| measured-only  pet_activity \u00b7 2026-08-06 | 304 | 241 | \u2014 | 0.782 | 0.736 | barely clears amino-acid composition |\n| measured-only  pet_activity \u00b7 2026-08-06 | 304 | 241 | 46 | 0.921 | 0.736 | clears the composition baseline by a real margin |\n| mixed  pet_activity \u00b7 2026-08-06 | 791 | 241 | 60 | 0.967 | 0.736 | trained on similarity-derived labels |\n| measured-only  pet_activity \u00b7 2026-08-05 | 152 | 26 | 7 | 0.493 | 0.398 | at chance \u2014 the hardest contrast, and the honest answer |\n| measured-only  pet_activity \u00b7 2026-08-05 | 305 | 26 | 44 | 0.850 | 0.651 | clears the composition baseline by a real margin |\n| measured-only  pet_activity \u00b7 2026-08-05 | 300 | 220 | 45 | 0.976 | 0.829 | clears the composition baseline by a real margin |\n| experimental  pet_activity \u00b7 2026-08-04 | 13 | 220 | 1 | \u2014 | \u2014 | not evaluable \u2014 too few independent clusters to split on |\n| mixed  pet_activity \u00b7 2026-08-04 | 500 | 220 | 25 | 1.000 | 0.778 | perfect, and meaningless: it reproduced the similarity rule that made the labels |\n\n## The machine this runs on\n\nA single DigitalOcean droplet shared with five other applications, which is why\nthe web layer carries no scientific libraries at all: every structure and every measurement was\ncomputed offline and is served here as a precomputed file.\n\n78.9%\n\ndisk used\n\n17.3 GB free of\n82.1 GB\n\n66.3%\n\nmemory used\n\n1.38 GB\navailable of 4.11 GB.\nMemory is the binding constraint here, not disk or speed.\n\n0.05\n\nload average\n\nacross 2\ncores. Above the core count means work is queueing.\n\n20.6 MB\n\ndatabase\n\nSQLite, one file, everything in it.\n832 candidate and\n559\nreference structure files alongside it, totalling\n139.5 +\n117.6 MB.\n\nLinux x86_64 \u00b7 Python 3.12.3 \u00b7 web process up\n132.51 h \u00b7 host up 1319.2 h\n\u00b7 data version 0.3.0", "# The Unsung Hero: How a Single Chlorine Atom Transforms Drug Discovery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nRecent methodological advances promise to expand chlorine\u2019s utility even further. A [2025 photocatalytic chlorination method developed at Rice University](https://news.rice.edu/news/2025/chemists-create-eco-friendly-method-make-chlorine-based-materials-drugs-and-chemicals) enables room-temperature, blue-light-mediated anti-Markovnikov hydrochlorination using sustainable iron-sulfur catalysts. This precision chlorination approach addresses the traditional challenge of regioselective halogenation, offering pharmaceutical chemists enhanced control over chlorine placement without harsh reagents or extensive purification. Meanwhile, [computational approaches incorporating halogen bonding parameters into docking algorithms](https://onlinelibrary.wiley.com/doi/10.1002/pro.70321) enable prospective design of chlorinated analogs with predicted affinity improvements, accelerating lead optimization cycles.\n\nThe integration of machine learning with halogen bonding principles represents the frontier of chlorine-enabled drug design. [Systematic benchmarking studies published in 2025](https://pmc.ncbi.nlm.nih.gov/articles/PMC12199455/) established optimal quantum mechanical methods for characterizing halogen-\u03c0 and halogen-Lewis base interactions, generating datasets suitable for training predictive models. These computational frameworks allow researchers to virtually screen chlorination sites and predict binding energy enhancements without extensive synthetic campaigns\u2014democratizing access to the magic chloro effect.\n\nAs pharmaceutical research confronts antibiotic resistance, neglected tropical diseases, and complex oncology targets, the chlorine atom remains an indispensable tool for molecular optimization. The convergence of mechanistic understanding, structural validation, synthetic methodology, and computational prediction positions chlorine substitution as a rational\u2014rather than serendipitous\u2014strategy for 21st-century drug discovery. Whether deployed as a single strategic modification or as part of multi-halogenated scaffolds, chlorine continues to prove that in medicinal chemistry, sometimes the smallest change makes the biggest difference.\n\n| Concept | Description | Key References |\n| --- | --- | --- |\n| **Magic Chloro Effect** | Single chlorine substitution (H\u2192Cl) can dramatically improve drug potency, often by 10-100x, sometimes up to 100,000x | [Chiodi & Ishihara, *J Med Chem*](https://pubmed.ncbi.nlm.nih.gov/37014977/); [Joshi & Srivastava, *RSC Advances*](https://pmc.ncbi.nlm.nih.gov/articles/PMC10687521/) |\n| **Prevalence in Drugs** | Over 250 FDA-approved drugs contain chlorine atoms, making it the 5th most common element after C, H, N, and O | [FDA approval statistics](https://www.fda.gov/drugs/novel-drug-approvals-fda/novel-drug-approvals-2024); [Njardarson elemental analysis](https://pubmed.ncbi.nlm.nih.gov/37014977/) |\n| **Potency Improvements** | Chlorine substitutions achieve 10-fold to >1000-fold potency improvements through optimized protein-ligand interactions | [Multiple examples across therapeutic areas](https://chemrxiv.org/engage/chemrxiv/assets/orp/resource/item/62f1dd4c69f3a528d2b0d366/original/magic-chloro-profound-effects-of-the-chlorine-atom-in-drug-discovery.pdf) (2020-2025) |\n| **Pharmacokinetic Benefits** | Chlorine reduces metabolic clearance, extends half-life, and increases drug exposure (AUC) while maintaining favorable ADME properties | [HLM studies](https://www.frontiersin.org/journals/pharmacology/articles/10.3389/fphar.2023.1145140/full); [Metabolic stability analyses](https://pmc.ncbi.nlm.nih.gov/articles/PMC10687521/) (2020-2025) |\n| **Halogen Bonding** | Chlorine acts as Lewis acid forming directional interactions with backbone carbonyls, hydroxyl groups, and other Lewis bases in binding pockets | [Wilcken et al., *J Med Chem*](https://pubs.acs.org/doi/10.1021/jm3012068); [Auffinger et al., *Chem Commun*](https://pubs.rsc.org/en/content/articlelanding/2020/cc/d0cc00841a) |\n| **Lipophilicity & Binding** | Chlorine increases lipophilicity (\u0394logD ~0.6-1.0) while filling hydrophobic pockets and forming favorable van der Waals contacts | [Topliss scheme](https://www.tandfonline.com/doi/full/10.1080/17568919.2025.2485865); [LipE optimization strategies](https://chemrxiv.org/engage/chemrxiv/assets/orp/resource/item/62f1dd4c69f3a528d2b0d366/original/magic-chloro-profound-effects-of-the-chlorine-atom-in-drug-discovery.pdf) |\n| **7-Chloroquinoline Core** | Essential pharmacophore in antimalarial drugs (chloroquine, hydroxychloroquine); electron-withdrawing chlorine optimizes pKa for activity | [Antimalarial optimization studies](https://chemrxiv.org/engage/chemrxiv/assets/orp/resource/item/62f1dd4c69f3a528d2b0d366/original/magic-chloro-profound-effects-of-the-chlorine-atom-in-drug-discovery.pdf) (2020-2025) |\n| **Para-Chlorophenyl Effect** | Most common chlorination site due to synthetic accessibility via electrophilic aromatic substitution; part of Topliss decision tree | [Topliss operational scheme](https://www.tandfonline.com/doi/full/10.1080/17568919.2025.2485865); >100 examples with dramatic improvements |\n| **Two-Chloro Effects** | Di-chlorination can produce >10,000-fold improvements through conformational control and enhanced protein interactions | [FGFR1, SERT examples](https://chemrxiv.org/engage/chemrxiv/assets/orp/resource/item/62f1dd4c69f3a528d2b0d366/original/magic-chloro-profound-effects-of-the-chlorine-atom-in-drug-discovery.pdf) (2020-2025) |\n| **Selectivity Enhancement** | Chlorine substitution can improve target selectivity >3000-fold by exploiting binding pocket differences between related proteins | [D3 receptor, kinase selectivity](https://chemrxiv.org/engage/chemrxiv/assets/orp/resource/item/62f1dd4c69f3a528d2b0d366/original/magic-chloro-profound-effects-of-the-chlorine-atom-in-drug-discovery.pdf) (2023-2025) |", "# Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly\nAuthors: Deller, M.C.; Johnson, H.A.; Miller, M.D.; Spraggon, G.; Elsliger, M.-A.; Wilson, I.A.; Lesley, S.A.\nSource: PLoS ONE, 2015\n\n### Crystallization and data collection\n\nTM1088A was concentrated to 10 mg/ml using centrifugal ultrafiltration (Millipore). TM1088A was crystallized using the nanodroplet vapor diffusion method with standard JCSG crystallization protocols [27,29]. Sitting drops composed of 200 nl protein solution mixed with 200 nl crystallization solution in a sitting drop format were equilibrated against a 50 \u03bcl reservoir at 293 K for 54 days prior to harvest. The crystallization reagent consisted of 40% MPD, 5% PEG-8000, and 0.1 M Cacodylate pH 6.5. No further cryoprotectant was added to the crystal. Initial screening for diffraction was carried out using the Stanford Automated Mounting system (SAM) at the Stanford Synchrotron Radiation Lightsource (SSRL, Menlo Park, CA).\n\nThe TM1088B C-terminal domain was dialyzed into crystallization buffer (20 mM Tris, pH 7.8 and 150 mM NaCl) and concentrated to 15 mg/ml using centrifugal ultrafiltration (Millipore). TM1088B C-terminal domain was crystallized using the nanodroplet vapor diffusion method with standard JCSG crystallization protocols [27,29]. Sitting drops composed of 150 nl protein solution mixed with 150 nl crystallization solution in a sitting drop format were equilibrated against a 60 \u03bcl reservoir at 293 K. The crystallization reagent consisted of 25% (v/v) ethylene glycol. No further cryoprotectant was added to the crystal.\n\nThe TM1088 complex was concentrated to 3 mg/ml using centrifugal ultrafiltration (Millipore) and NADH added to a final molar ratio of 1.1:1.0 (NADH: protein). The TM1088 complex was crystallized by vapor diffusion against 0.1 M magnesium chloride, 15% (w/v) PEG 4000, 0.1 M HEPES, pH 7.0. Drops consisted of 0.5 \u03bcl protein, 0.5 \u03bcl reservoir and 0.25 \u03bcl xylitol (30% (w/v)) as cryoprotectant with a 60 \u03bcl reservoir volume. Reservoirs were overlaid with 10 \u03bcl of silicon oil and paraffin oil (1:1) to regulate the rate of vapor diffusion [30].\n\nCrystals were flash-cooled to 100 K and X-ray data collected at the Advanced Light Source (ALS, Lawrence Berkeley National Laboratory) at beamlines 5.0.1 (TM1088A) or 5.0.3 (TM1088B and TM1088 complex). All data were processed with HKL2000 or Mosflm and SCALA from the CCP4 package [31,32].\n\n### Phasing, model building and refinement\n\nRefinement was performed with Refmac5 [33] and PHENIX [34\u201336] and model building with O [37], Coot [38] and MIFit (http://code.google.com/p/mifit). All other crystallographic manipulations were carried out with the CCP4 package [32]. Molecular graphics were prepared using PyMOL (Schr\u00f6dinger, LLC). Data collection and refinement statistics are shown in Table 1.\n\n*Table 1.* Crystallographic data collection and refinement.\n\n| | TM1088 assembly | TM1088A | TM1088B C-terminal domain |\n|---|---|---|---|\n| PDB ID | 3L4B | 2G1U | 3JXO |\n| Beamline | ALS 5.0.3 | ALS 5.0.1 | ALS 5.0.3 |\n| Space group | P3(2)21 | P2(1) | C222(1) |\n| Cell dimensions (\u00c5) | a = b = 96.78, c = 305.10 | a = 47.86, b = 35.14, c = 56.58, \u03b2 = 111.85\u00b0 | a = 62.11, b = 64.40, c = 97.00 |\n| Wavelength (\u00c5) | 1.000 | 1.000 | 1.000 |\n| Resolution range (\u00c5) | 48.4\u20133.45 | 44.4\u20131.50 | 50.0\u20131.55 |\n| Highest resolution shell (\u00c5) | 3.57\u20133.45 | 1.54\u20131.50 | 1.61\u20131.55 |\n| Rmerge (%) (a) | 17.9 (77.2) | 4.8 (34.4) | 6.5 (51.9) |\n| Mean I/\u03c3(I) (a) | 11.4 (2.0) | 19.1 (1.6) | 39.9 (2.3) |\n| Observations | 178,674 | 133,164 | 395,401 |\n| Unique reflections (a) | 22,866 (2244) | 26,312 (1298) | 28,472 (2651) |\n| Reflections used in refinement (Rcryst) | 20,333 | 24,978 | 25,563 |\n| Reflections used for Rfree | 1164 | 1332 | 1438 |\n| Completeness (%) (a) | 99.9 (100.0) | 93.4 (63.8) | 99.3 (94.1) |\n| Rcryst (Rfree) (%) | 25.9 (30.8) | 15.3 (17.5) | 15.8 (18.7) |\n| No. protein atoms | 8952 | 1144 | 2433 |\n| No. water molecules | 0 | 110 | 89 |\n| No. hetero atoms | 93 | 40 | 0 |\n| RMSD bond lengths (\u00c5) | 0.006 | 0.016 | 0.014 |\n| RMSD bond angles (\u00b0) | 0.97 | 1.65 | 1.48 |\n| Ramachandran favored (outliers) (%) (b) | 95.1 (0.9) | 99.3 (0) | 98.8 (0) |\n| Mean overall protein B-value (AMP) | 49.3 (78.7) | 18.2 (22.1) | 23.2 |\n\nRmerge = \u03a3hkl \u03a3i |Ii(hkl) \u2212 <I(hkl)>| / \u03a3hkl \u03a3i Ii(hkl). Rcryst = \u03a3||Fobs| \u2212 |Fcalc|| / \u03a3|Fobs| where Fcalc and Fobs are the calculated and observed structure factor amplitudes, respectively. Rfree = as for Rcryst, but for 5.0% of the total reflections chosen at random and omitted from refinement. (a) Numbers in parentheses are for highest resolution shell. (b) As determined by MolProbity [39].", "# Therapeutic Enzymatic Compositions and Delivery Methods for the Degradation of Microplastics and Nanoplastics in Human Physiology\nAuthors: Deller, Marc C. (inventor)\nSource: United States Provisional Patent Application (Elora Therapeutics), 2025\n\nThe invention comprises engineered enzyme formulations specifically tailored for internal human biodegradation of microplastics and nanoplastics, to include PET, PE, PP, and PS. It includes optimized enzyme variants\u2014PETase (wild-type, engineered variants, fusion enzymes, thermostabilized mutants), MHETase (wild-type, engineered variants, fusion enzymes), bacterial-origin CotA-Laccase (engineered variants with enhanced oxidative activity), and engineered HRP (recombinant and oxidative-stabilized variants). The invention posits novel delivery methods, including encapsulation in biodegradable polymers, lipid nanoparticles, PEGylation, and hydrogel stabilization, to maximize enzyme effectiveness, stability, and safety in therapeutic contexts.\n\n## Brief Description of Drawings\n\n*FIG. 1.* Depicts the enzymatic PET degradation pathway involving PETase and MHETase variants \u2013 including minor and major byproducts \u2013 under human physiological conditions.\n\n*FIG. 2.* Illustrates the oxidative degradation pathways for PE by bacterial-origin CotA-Laccase \u2013 including minor and major byproducts \u2013 under human physiological conditions.\n\n*FIG. 3.* Illustrates the oxidative degradation pathways for PP by bacterial-origin CotA-Laccase \u2013 including minor and major byproducts \u2013 under human physiological conditions.\n\n*FIG. 4.* Illustrates the oxidative degradation pathways for PS by bacterial-origin CotA-Laccase \u2013 including minor and major byproducts \u2013 under human physiological conditions.\n\n*FIG. 5.* Illustrates the oxidative degradation pathways for PE by engineered HRP \u2013 including minor and major byproducts \u2013 under human physiological conditions.\n\n*FIG. 6.* Illustrates the oxidative degradation pathways for PP by engineered HRP \u2013 including minor and major byproducts \u2013 under human physiological conditions.\n\n*FIG. 7.* Illustrates the oxidative degradation pathways for PS by engineered HRP \u2013 including minor and major byproducts \u2013 under human physiological conditions.\n\n## Detailed Description of the Invention\n\nThe invention is comprised of a unique composition of enzymes and innovative delivery methods specifically optimized for human therapeutic use to degrade microplastics and nanoplastics effectively and safely.\n\n### Enzyme Compositions\n\n- **PET-degrading enzymes:** wild-type PETase, engineered PETase variants, PETase fusion enzymes, and thermostabilized PETase mutants specifically developed for therapeutic human applications.\n- **MHETase enzymes:** wild-type MHETase, engineered MHETase variants, MHETase fusion enzymes.\n- **Bacterial-origin CotA-Laccase enzymes:** engineered CotA-Laccase variants optimized for oxidative activity, incorporating site-directed mutagenesis to enhance substrate specificity, thermostability, and modifications in redox potential.\n- **HRP enzymes:** engineered HRP variants, recombinant HRP expressed in non-fungal systems, oxidative peroxidase hybrids incorporating site-specific mutations for enhanced catalytic efficiency and stability.\n\n### Innovative Delivery Systems\n\n- **Oral Formulations:** encapsulation in biodegradable polymer nano-encapsulation or enteric-coated capsules.\n- **Injectable Formulations:** PEGylation, lipid nanoparticles, hydrogels for systemic distribution, reduced immunogenicity, and enhanced stability.\n- **Hybrid Delivery System:** oral administration of PET-degrading enzymes and MHETase, combined with injectable CotA-Laccase and HRP.\n\n### Metabolic Clearance and Safety\n\nHuman-specific hepatic and renal metabolic pathways facilitate safe excretion of degradation products. Immunogenicity risks are mitigated by PEGylation and nano-encapsulation, supported by experimental studies demonstrating enhanced enzyme stability and reduced immune response.\n\n## Claims\n\n1. A therapeutic enzymatic composition formulated specifically for the human degradation of microplastics and nanoplastics under physiological conditions, comprising at least one enzyme selected from the group consisting of:\n   - PET-degrading enzymes selected from wild-type PETase, engineered PETase variants, PETase fusion enzymes, and thermostabilized PETase mutants;\n   - MHETase enzymes selected from wild-type MHETase, engineered MHETase variants, and MHETase fusion enzymes;\n   - Bacterial-origin CotA-Laccase enzymes comprising engineered CotA-Laccase variants optimized for oxidative activity, substrate specificity, thermostability, and modified redox potential;\n   - Horseradish Peroxidase enzymes comprising engineered HRP variants, recombinant HRP expressed in non-fungal systems, and oxidative peroxidase hybrids incorporating mutations for enhanced catalytic efficiency and stability.\n\n2. The therapeutic enzymatic composition of claim 1, wherein the enzymes are delivered orally through encapsulation methods selected from biodegradable polymer nano-encapsulation or enteric-coated capsules.\n\n3. The therapeutic enzymatic composition of claim 1, wherein the enzymes are delivered through injectable formulations selected from PEGylation, lipid nanoparticles, or hydrogel encapsulation to optimize systemic distribution, stability, and reduced immunogenic response.\n\n4. The therapeutic enzymatic composition of claim 1, utilizing a hybrid delivery system comprising oral administration of PET-degrading enzymes and MHETase combined with injectable administration of CotA-Laccase and engineered HRP enzymes.\n\n5. A method for therapeutic degradation of microplastics and nanoplastics within the human body comprising administering the enzymatic composition of claim 1 and ensuring safe metabolic clearance of degradation products via hepatic and renal pathways.\n\n6. The therapeutic enzymatic composition of claim 1, wherein enzyme stability and immunogenicity reduction are achieved through methods selected from PEGylation, lipid nanoparticle encapsulation, or biodegradable polymer encapsulation, supported by empirical data demonstrating increased enzyme stability and reduced immune response.\n\n## Abstract", "# marcdeller.com: Structural Biology\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n### [Outrunning Diffusion: The Fast Lane in CryoEM Grid Preparation](https://marcdeller.com/outrunning-diffusion-the-fast-lane-in-cryoem-grid-preparation/)\n\nBlog\n\n### [Unlocking Membrane Protein Structures: The Crystallization Revolution Reshaping Drug Discovery](https://marcdeller.com/unlocking-membrane-protein-structures-the-crystallization-revolution-reshaping-drug-discovery/)\n\nBlog\n\n### [Elevating Science Communication: Effortless 3D Structure Embedding with 3Dmol.js](https://marcdeller.com/elevating-science-communication-effortless-3d-structure-embedding-with-3dmol-js/)\n\nBlog\n\n### [TYK2 Biology: How Deep Mutational Scanning Reveals New Therapeutic Frontiers](https://marcdeller.com/unlocking-tyk2-biology-how-deep-mutational-scanning-reveals-new-therapeutic-frontiers/)\n\nBlog\n\n### [The Unsung Hero: How a Single Chlorine Atom Transforms Drug Discovery](https://marcdeller.com/magic-chloro-profound-effects-of-the-chlorine-atom-in-drug-discovery/)\n\nBlog\n\n### [Protein Hunter: Pioneering the Next Frontier of AI-Driven Protein Design](https://marcdeller.com/protein-hunter-pioneering-the-next-frontier-of-ai-driven-protein-design/)\n\nBlog\n\n### [Revolutionary Bridge Between Prediction and Reality: How Elastic Networks Transform AlphaFold2\u2019s Structural Accuracy](https://marcdeller.com/revolutionary-bridge-between-prediction-and-reality-how-elastic-networks-transform-alphafold2s-structural-accuracy/)\n\nBlog\n\n### [Precision Timing of Cytokine Signals through Facilitated Dissociation](https://marcdeller.com/precision-timing-of-cytokine-signals-through-facilitated-dissociation/)\n\nBlog\n\n### [Spacial Score\u2500A Comprehensive Topological Indicator for Small-Molecule Complexity](https://marcdeller.com/spacial-score%e2%94%80a-comprehensive-topological-indicator-for-small-molecule-complexity/)\n\nBlog\n\n### [The $40 Billion Revolution: How Bispecific Antibodies Are Transforming Cancer Care](https://marcdeller.com/the-40-billion-revolution-how-bispecific-antibodies-are-transforming-cancer-care/)\n\nBlog\n\n### [Revolutionizing Protein Crystallography: The Synergy of TELSAM-DARPin Fusion and Electric Field Optimization](https://marcdeller.com/revolutionizing-protein-crystallography-the-synergy-of-telsam-darpin-fusion-and-electric-field-optimization/)\n\nBlog\n\n### [The Memorization Trap: Have Protein-Ligand Cofolding Methods Transcended Simple Data Recall?](https://marcdeller.com/the-memorization-trap-have-protein-ligand-cofolding-methods-transcended-simple-data-recall/)\n\nBlog\n\n### [Remapping the Chemical and Pharmacological Spaces of Drugs: Navigating Tomorrow\u2019s Therapeutic Frontier](https://marcdeller.com/remapping-the-chemical-and-pharmacological-spaces-of-drugs-navigating-tomorrows-therapeutic-frontier/)\n\nBlog\n\n### [Revolutionary Cryo-EM Breakthrough: Di-Gembodies Transform Small Protein Structure Determination](https://marcdeller.com/revolutionary-cryo-em-breakthrough-di-gembodies-transform-small-protein-structure-determination/)\n\nBlog\n\n### [Unlocking Molecular Potential: The Evolution of Topliss Tree and Bioisosteric Strategies in Modern Drug Discovery](https://marcdeller.com/unlocking-molecular-potential-the-evolution-of-topliss-tree-and-bioisosteric-strategies-in-modern-drug-discovery/)\n\nBlog\n\n### [Building a Comprehensive Protein Production, X-ray Crystallography, and Cryo-EM Laboratory from Scratch](https://marcdeller.com/building-a-comprehensive-protein-production-x-ray-crystallography-and-cryo-em-laboratory-from-scratch/)\n\nBlog\n\n### [Empowering QSAR Modeling: Optimal Data Practices and Cutting-Edge Tools](https://marcdeller.com/empowering-qsar-modeling-optimal-data-practices-and-cutting-edge-tools/)\n\nBlog\n\n### [Optimizing Data Volume and Diversity for Robust Antibody\u2013Antigen \u0394\u0394G Prediction](https://marcdeller.com/optimizing-data-volume-and-diversity-for-robust-antibody-antigen-%ce%b4%ce%b4g-prediction/)\n\nBlog\n\n### [Nanobodies: Revolutionizing Molecular Medicine Through Nature\u2019s Smallest Antibodies](https://marcdeller.com/nanobodies-revolutionizing-molecular-medicine-through-natures-smallest-antibodies/)\n\nBlog\n\n### [MARCUS: Revolutionizing Chemical Literature Curation with AI-Driven Annotation and Recognition](https://marcdeller.com/marcus-revolutionizing-chemical-literature-curation-with-ai-driven-annotation-and-recognition/)\n\nSB\n\n### [Revolutionizing Drug Discovery: LatentX and the Future of Protein Binder Design](https://marcdeller.com/revolutionizing-drug-discovery-latentx-and-the-future-of-protein-binder-design/)\n\nSB\n\n### [Unveiling FragLites: Pioneering Precision in Drug Discovery](https://marcdeller.com/unveiling-fraglites-pioneering-precision-in-drug-discovery/)\n\nSB\n\n### [The LockTac Revolution: Binding the Future of Therapeutics](https://marcdeller.com/the-locktac-revolution-binding-the-future-of-therapeutics/)\n\nSB\n\n### [The Shape-Shifting World of Intrinsically Disordered Region Binding Proteins](https://marcdeller.com/the-shape-shifting-world-of-intrinsically-disordered-region-binding-proteins/)\n\nSB\n\n### [Unlocking Protein Structure Insights without a Lab: Harnessing CRO Services for Structural Biology Research](https://marcdeller.com/unlocking-protein-structure-insights-without-a-lab-harnessing-cro-services-for-structural-biology-research/)\n\nSB\n\n### [Engineering the Future of Proteins: Unveiling the Latest AI Tools Transforming Structural Biology in 2025](https://marcdeller.com/engineering-the-future-of-proteins-unveiling-the-latest-ai-tools-transforming-structural-biology-in-2025/)\n\nSB\n\n### [Unleashing Dynamic Precision: How DRGSCROLL Transforms Protein\u2013Ligand Docking](https://marcdeller.com/unleashing-dynamic-precision-how-drgscroll-transforms-protein-ligand-docking/)\n\nSB\n\n### [Capturing the Invisible: Understanding the \u00c5ngstr\u00f6m in Structural Biology](https://marcdeller.com/capturing-the-invisible-understanding-the-angstrom-in-structural-biology/)\n\nSB", "# Flow Cytometry and Microplastic Detection in Human Tissues\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Key Concepts | Description | Key References |\n| --- | --- | --- |\n| **Flow Cytometry Detection** | Automated technique using Nile red fluorescent staining coupled with flow cytometry to rapidly detect and quantify microplastics and nanoplastics (>200 nm) in liquid samples, enabling analysis of thousands of particles within minutes. | [Li et al., MethodsX, 2025](https://www.sciencedirect.com/science/article/pii/S2215016125000482); [Li et al., Frontiers in Chemistry, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC10540628/) |\n| **Nile Red Staining Protocol** | Lipophilic fluorescent dye that selectively binds to plastic polymers in solution, optimal at 15-20 \u03bcg/mL concentration with 20-30% DMSO to reduce agglomeration. Successfully stains PE, PS, PET, and PC polymers. | [Maes et al., Scientific Reports, 2017](https://www.nature.com/articles/srep44501); [Fast-screening method, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC9958479/) |\n| **Human Tissue Applications** | Flow cytometry has been adapted to detect microplastics in human blood, liver, placenta, and other tissues following tissue digestion and extraction protocols. Detects particles down to 200 nm in biological samples. | [Methods for human tissues, 2020](https://www.eurekalert.org/news-releases/478209); [Microplastics in human tissues, 2024](https://pmc.ncbi.nlm.nih.gov/articles/PMC11342020/) |\n| **Advantages Over Spectroscopy** | Flow cytometry offers faster analysis (minutes vs. hours), automated processing, higher throughput, detection of submicron particles (<1 \u03bcm), and lower detection limits compared to FTIR (>20 \u03bcm) and Raman (>1 \u03bcm) spectroscopy. | [Flow cytometry review, 2025](https://pmc.ncbi.nlm.nih.gov/articles/PMC11870216/); [Spectroscopy limitations, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC10035019/) |\n| **Limitations and Challenges** | Main challenges include Nile red aggregation in aqueous media causing false positives, background noise overlap, limited polymer identification specificity, and requirement for careful optimization of staining conditions and data analysis gating strategies. | [Technical challenges, 2025](https://pmc.ncbi.nlm.nih.gov/articles/PMC11870216/); [Analytical method development, 2023](https://www.sciencedirect.com/science/article/pii/S2772577423000095) |", "# Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3\nAuthors: Shvartsbart, A.; Roach, J.J.; Witten, M.R.; Koblish, H.; Harris, J.J.; Covington, M.; Hess, R.; Lin, L.; Frascella, M.; Truong, L.; Leffet, L.; Conlen, P.; Beshad, E.; Klabe, R.; Katiyar, K.; Kaldon, L.; Young-Sciame, R.; He, X.; Petusky, S.; Chen, K.-J.; Horsey, A.; Lei, H.-T.; Epling, L.B.; Deller, M.C.; Vechorkin, O.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2022\n\n**Step 1. 3,5,7-Trichloropyrazolo[1,5-a]pyrimidine (39).** To a solution of 5,7-dichloropyrazolo[1,5-a]pyrimidine (500 mg, 2.66 mmol, Combi-Blocks ST-4256) in DMF (5 mL) was added NCS (391 mg, 2.93 mmol), and the reaction mixture was stirred at 60 \u00b0C for 1 h. After cooling to room temperature, ice chips and water were added. The resulting precipitate was filtered, washed with water, and air-dried to afford the title compound (460 mg, 78%) as an off-white solid. LCMS calculated for C6H3Cl3N3 (M + H)+: m/z = 221.9; found: 221.9.\n\n**Step 2. 3,5-Dichloropyrazolo[1,5-a]pyrimidin-7-amine (40).** A suspension of 3,5,7-trichloropyrazolo[1,5-a]pyrimidine (460 mg, 2.07 mmol) in concentrated ammonium hydroxide (10 mL, 145 mmol) was heated to 100 \u00b0C in a heavy-walled sealed tube for 1 h. After cooling to room temperature, the reaction mixture was diluted with cold water and filtered. The collected solid was then washed with cold water and air-dried to afford the title compound (355 mg, 85%). LCMS calculated for C6H5Cl2N4 (M + H)+: m/z = 203.0; found: 203.0.\n\n**Step 3. 3,5-Dichloro-6-iodopyrazolo[1,5-a]pyrimidin-7-amine (41).** To a solution of 3,5-dichloropyrazolo[1,5-a]pyrimidin-7-amine (355 mg, 1.75 mmol) in DMF (5 mL) was added NIS (590 mg, 2.62 mmol), and the reaction mixture was stirred at 50 \u00b0C for 1 h. After cooling to room temperature, ice chips and water were added. The resulting precipitate was filtered, washed with water, and air-dried to afford the title compound (489 mg, 85%). LCMS calculated for C6H4Cl2IN4 (M + H)+: m/z = 328.9; found: 328.8.\n\n**Step 4. 3,5-Dichloro-6-iodopyrazolo[1,5-a]pyrimidine (42).** To a solution of 3,5-dichloro-6-iodopyrazolo[1,5-a]pyrimidin-7-amine (489 mg, 1.49 mmol) in THF (4 mL) was added tert-butyl nitrite (0.88 mL, 7.43 mmol), and the reaction mixture was heated to 80 \u00b0C for 2 h. The reaction mixture was concentrated and purified by flash chromatography (0\u221250% EtOAc/hexanes) to afford the title compound (217 mg, 47%). LCMS calculated for C6H3Cl2IN3 (M + H)+: m/z = 313.9; found: 313.7.\n\n**Step 5. 3,5-Dichloro-6-(1-isopropyl-1H-pyrazol-4-yl)pyrazolo[1,5-a]pyrimidine (43a).** A mixture of 3,5-dichloro-6-iodopyrazolo[1,5-a]pyrimidine (100 mg, 0.32 mmol), 1-isopropyl-4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazole (75 mg, 0.32 mmol), Pd(dppf)Cl2\u00b7CH2Cl2 (13 mg, 16 \u03bcmol), and sodium carbonate (101 mg, 0.96 mmol) in 1,4-dioxane (2 mL) and water (0.5 mL) was sparged with N2 and heated to 80 \u00b0C for 1 h. The reaction mixture was diluted with EtOAc and water. The layers were separated, and the organic layer was washed with brine, dried over MgSO4, filtered, and concentrated. The residue was purified by flash chromatography (0\u2212100% EtOAc/hexanes) to afford the title compound (77 mg, 82%). LCMS calculated for C12H12Cl2N5 (M + H)+: m/z = 296.0; found: 296.2.\n\n**Step 6. (S)-3-Chloro-6-(1-isopropyl-1H-pyrazol-4-yl)-5-((tetrahydrofuran-3-yl)oxy)pyrazolo[1,5-a]pyrimidine (44a).** To a solution of 3,5-dichloro-6-(1-isopropyl-1H-pyrazol-4-yl)pyrazolo[1,5-a]pyrimidine (10 mg, 34 \u03bcmol) in DMF (1 mL) was added (S)-tetrahydrofuran-3-ol (14 mg, 0.17 mmol), followed by cesium carbonate (16.5 mg, 0.051 mmol) at room temperature. The reaction mixture was heated to 80 \u00b0C for 1 h, then cooled to room temperature. The solution of the product in DMF was used directly for the next reaction. LCMS calculated for C16H19ClN5O2 (M + H)+: m/z = 348.1; found: 348.1.\n\n**Step 7. (S)-5-(6-(1-Isopropyl-1H-pyrazol-4-yl)-5-((tetrahydrofuran-3-yl)oxy)pyrazolo[1,5-a]pyrimidin-3-yl)-N-methylnicotinamide (27).** A mixture of (S)-3-chloro-6-(1-isopropyl-1H-pyrazol-4-yl)-5-((tetrahydrofuran-3-yl)oxy)pyrazolo[1,5-a]pyrimidine (12 mg, 35 \u03bcmol), N-methyl-5-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)nicotinamide (11 mg, 41 \u03bcmol), XPhos Pd G2 (1.4 mg, 1.7 \u03bcmol), and cesium carbonate (34 mg, 0.1 mmol) in DMF (1 mL) and water (0.5 mL) was sparged with N2 and heated to 100 \u00b0C for 2 h. The reaction was diluted with MeOH and filtered through a SiliaPrep thiol cartridge, and the product was purified by prep HPLC (pH 2) to provide the title compound (1.4 mg, 3.1 \u03bcmol, 9% over two steps). LCMS calculated for C23H26N7O3 (M + H)+: m/z = 448.2; found: 448.4. 1H NMR (500 MHz, DMSO-d6) \u03b4 9.39 (d, J = 2.1 Hz, 1H), 9.38 (s, 1H), 9.00 (t, J = 2.1 Hz, 1H), 8.85 (d, J = 2.0 Hz, 1H), 8.74 (s, 1H), 8.72 (q, J = 4.9, 4.4 Hz, 1H), 8.28 (s, 1H), 8.06 (s, 1H), 5.79 (td, J = 4.7, 2.4 Hz, 1H), 4.58 (p, J = 6.7 Hz, 1H), 4.17 (dd, J = 10.7, 4.7 Hz, 1H), 4.11\u22124.05 (m, 1H), 3.96 (q, J = 8.1 Hz, 1H), 3.88 (td, J = 8.3, 4.7 Hz, 1H), 2.86 (d, J = 4.5 Hz, 3H), 2.46 (dt, J = 14.6, 7.3 Hz, 1H), 2.36\u22122.26 (m, 1H), 1.47 (d, J = 6.6 Hz, 6H).\n\n### Synthesis of Compound 29\n\nA mixture of (S)-3-chloro-6-(1-isopropyl-1H-pyrazol-4-yl)-5-((tetrahydrofuran-3-yl)oxy)pyrazolo[1,5-a]pyrimidine (44a, 150 mg, 0.43 mmol), N-(methyl-d3)-5-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)nicotinamide (137 mg, 0.52 mmol, see the Supporting Information for preparation), XPhos Pd G2 (17 mg, 22 \u03bcmol), and cesium carbonate (422 mg, 1.29 mmol) in 1,4-dioxane (4 mL) and water (0.5 mL) was sparged with N2 and heated to 100 \u00b0C for 2 h. The reaction mixture was diluted with MeOH and filtered through a SiliaPrep thiol cartridge, and the residue was purified by prep HPLC (pH 2) to afford the title compound (73 mg, 38%). LCMS calculated for C23H23D3N7O3 (M + H)+: m/z = 451.2; found: 451.2. 1H NMR (600 MHz, DMSO-d6) \u03b4 9.40 (s, 2H), 8.99 (t, J = 2.2 Hz, 1H), 8.84 (s, 1H), 8.75 (s, 1H), 8.68 (s, 1H), 8.29 (s, 1H), 8.07 (s, 1H), 5.80 (ddt, J = 6.7, 4.4, 2.0 Hz, 1H), 4.59 (hept, J = 6.7 Hz, 1H), 4.18 (dd, J = 10.7, 4.8 Hz, 1H), 4.08 (dd, J = 10.7, 1.9 Hz, 1H), 3.96 (q, J = 7.9 Hz, 1H), 3.88 (td, J = 8.4, 4.7 Hz, 1H), 2.50\u22122.42 (m, 1H), 2.32 (dddd, J = 13.9, 6.9, 4.7, 1.9 Hz, 1H), 1.47 (d, J = 6.7 Hz, 6H).\n\n### Synthesis of Compound 30", "# ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n* All 17 corpus CSVs searchable directly from the chat CLI via keyword fast-path (`rag/corpus_lookup.py`) that bypasses the LLM for enumeration queries\n* Up to 500 matching rows loaded per source; pageable in 50-row pages with CSV export\n* PDB binding affinity cross-reference: resolves target names to PDB IDs via `pdb_all_entries.csv`\n* Deduplication by `comp_id` sorted by `comp_mw` descending (highest-MW representative per compound)\n* Intent filtering: SMILES/affinity/sequence queries drop irrelevant sources from the menu\n* `_CAPS_STOPWORDS` prevents format words (\u201cSMILES\u201d) from winning keyword extraction over target names (\u201cEGFR\u201d)\n* RAG hybrid loop: top-5 corpus chunks per query via ChromaDB (BAAI/bge-base-en-v1.5, 55k chunks)\n* RDKit auto-execution of code blocks in model responses\n* ASCII art banner (ChemSage, `small_slant` font), Rich Markdown rendering, Rich `Table` corpus tables with rounded cyan borders\n* Column display config (`_COL_CFG`) maps all corpus column names to abbreviated headers and per-column max widths; ellipsis overflow keeps all rows single-line\n\n## \ud83c\udf10 Web deployment\n\n**Live at [chemsage.mdeller.com](https://chemsage.mdeller.com).** Not a rebuilt chat-bubble interface: the browser gets a real terminal (xterm.js) connected over WebSocket to a server-side PTY running the actual, unmodified `scripts/chat.py` \u2014 the same Rich-rendered markdown, corpus tables, and slash commands (`/save`, `/info`, `/retry`, tab completion) as running it locally.\n\n**No GPU on the droplet.** `chat_remote.py` is a drop-in launcher: at import time it patches `mlx_lm.generate`/`mlx_lm.stream_generate` to call a HuggingFace Space running the real fused 32B model on ZeroGPU (a serverless, on-demand GPU allocation), then runs `chat.py` completely unaware anything is different. `chat.py` itself is never modified for the hosted path.\n\n**Serving stack:** gunicorn with a single `gevent` worker (100 concurrent connections) rather than `sync` workers \u2014 a terminal session is I/O-bound (waiting on the remote model and on keystrokes), not CPU-bound, so one greenlet-based worker serves many concurrent sessions cheaply. nginx reverse-proxies both the WebSocket endpoint (`/ws`, with `Upgrade`/`Connection` headers and a long read/send timeout for idle-but-open sessions) and the static app shell, behind a Let\u2019s Encrypt certificate. Runs as a dedicated systemd service on the droplet alongside AlphaFraud and BoltzMaker\u2019s own web apps.\n\n**Source isolation:** the web app (`web/flask_app/`) lives on a separate `web` git branch/worktree, so it never touches the tested training/eval code on `main`.\n\n## \ud83e\uddf1 Stack\n\nEverything runs locally on **Apple Silicon** via **MLX-LM** (`mlx_lm.lora` to train, `mlx_lm.fuse` to merge, `mlx_lm.server` for an OpenAI-compatible endpoint the GUI talks to). Unified memory means no VRAM wall: 16 GB handles ~8B, 32 GB ~14B, 64 GB ~32B. QLoRA is automatic because the base is a 4-bit MLX model.\n\nThe chat CLI (`scripts/chat.py`) uses **Rich** for streaming Markdown output (`Live`), corpus tables (rounded borders, ellipsis overflow, abbreviated headers), panels, progress spinners, and a themed console; **prompt_toolkit** for the interactive input session (arrow-key history, slash-command tab completion, styled prompt, persistent bottom toolbar with context window indicator); **ChromaDB** + **sentence-transformers** for the RAG retrieval layer; and **RDKit** for live chemical property computation. Responses stream token-by-token via `mlx_lm.stream_generate` and re-render as Markdown on completion. New commands: `/save` (export conversation to Markdown), `/info` (session/model details), `/retry` (regenerate last response).\n\n## \u2705 To Do\n\n### Completed (Rounds 1-5)\n\n* \u2705 Increase SFT dataset from 1,500 to 5,000 examples with code-heavy generators (Round 3)\n* \u2705 Add examples that teach \u201cemit RDKit code, don\u2019t compute manually\u201d (Round 3: `single_property`, `full_profile`)\n* \u2705 Include SMILES validation examples (Round 3: `smiles_validation`)\n* \u2705 Increase `iters` from 750 to 1,500 (Round 3; stopped early at 625 \u2014 val loss plateaued)\n* \u2705 Try `num_layers: 64` \u2014 caused swap thrashing at 31.5 GB; reverted to 32\n* \u2705 Try `batch_size: 8` \u2014 4.5x slower due to memory pressure; reverted to 4\n* \u2705 Add structural biology generators: PyMOL selections/viz, PLIP, PDB format, pdb-tools, Biopython, ChimeraX, gemmi\n* \u2705 Add Plotly and R/ggplot2/bio3d recipe generators\n* \u2705 Add protein family generators: kinases, GPCRs, antibodies, GTPases, cytokines, receptors (real PDB data)\n* \u2705 Add numerical fidelity drills: `code_then_quote` (3x), `rounding_precision` (2x)\n* \u2705 Add refusal/out-of-scope examples\n* \u2705 Expand dataset to 8,000 examples across 51 behaviour classes\n\n### Round 5 dataset (complete)\n\n* \u2705 20,000 examples across 78 behaviour classes (`build_dataset.py --n 20000 --seed 51`)\n* \u2705 19 new generator classes targeting R4 exec/PyExec gap: `pyexec_drill` (\u00d76), `code_then_quote_v2` (\u00d75), `rounding_explicit` (\u00d75), `fidelity_multistep` (\u00d75), `herg_liability`, `selectivity_profile`, `prodrug_bcs`, `sar_delta`, `mdanalysis`, `dssp`, `ppi_interface`, `uniprot_api`, `conformer_3d`, `mcs_search`, `reaction_smarts`, `recap_fragmentation`, `electron_density`, `biological_assembly`, `drug_target_family`\n* \u2705 Token lengths p50=322, p99=488, max=558 \u2014 all under max_seq_length=3072\n* \u2705 707 examples rejected (96.5% pass rate); 78 classes, class balance max:median < 3:1\n* \u2705 `data/sft_r5/` \u2014 16,000 train / 2,000 val / 2,000 test\n\n### \ud83c\udf93 Round 5 training (complete \u2014 early stop)", "# Remapping the Chemical and Pharmacological Spaces of Drugs: Navigating Tomorrow\u2019s Therapeutic Frontier\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-11 on marcdeller.com (categories: blog, sb).*\n\n* [Remapping the Chemical Space and the Pharmacological Space of Drugs: What Can We Expect from the Road Ahead?](https://www.mdpi.com/1424-8247/17/6/742)\n\n* **Data-Driven Maps:** Interactive AI-powered visualizations transform vast molecular libraries into [navigable landscapes](https://pmc.ncbi.nlm.nih.gov/articles/PMC11733715/).\n* **Non-Linear Embeddings:** UMAP and GTM outperform PCA for preserving local and [global structures in complex datasets](https://pmc.ncbi.nlm.nih.gov/articles/PMC11733715/).\n* **Aromatic Prevalence:** 81% of approved drugs feature aromatic rings, guiding [scaffold selection](https://pmc.ncbi.nlm.nih.gov/articles/PMC11207054/).\n* **Target Shifts:** Kinase inhibitors dominate recent approvals, with [epigenetic writers on the rise](https://pmc.ncbi.nlm.nih.gov/articles/PMC11207054/).\n* **Hybrid Workflows:** Descriptor-based and AI-driven methods converge for [comprehensive space exploration](https://www.frontiersin.org/journals/drug-discovery/articles/10.3389/fddsv.2025.1674289/full).\n* **Explainable AI:** Transparent ML models ensure trust and regulatory acceptance in [mapping and drug repurposing](https://www.sciencedirect.com/science/article/pii/S1359644625001059).\n\n1. **Remapping the Chemical Space and the Pharmacological Space of Drugs:** What Can We Expect from the Road Ahead?: [Franco, et al., Pharmaceuticals](https://pmc.ncbi.nlm.nih.gov/articles/PMC11207054/)\n2. **Exploring Dimensionality Reduction for Chemical Space Visualization:** [Li, et al., Journal of Chemical Information and Modeling](https://pmc.ncbi.nlm.nih.gov/articles/PMC11733715/)\n3. **Chemical space visual navigation in the era of deep learning and artificial intelligence:** [G\u00f3mez-Bombarelli, et al., Trends in Pharmacological Sciences](https://www.sciencedirect.com/science/article/pii/S1359644625001059)\n4. **On the biologically relevant chemical space: BioReCS**: [L\u00f3pez-L\u00f3pez, et al., Frontiers in Drug Discovery](https://www.frontiersin.org/journals/drug-discovery/articles/10.3389/fddsv.2025.1674289/full)\n5. **Properties of FDA-approved small molecule protein kinase inhibitors:** [Roskoski, et al., Pharmacological Research](https://pubmed.ncbi.nlm.nih.gov/40252783/)\n6. **Applications of Artificial Intelligence in Drug Repurposing:** [Wan, et al., Advanced Science](https://advanced.onlinelibrary.wiley.com/doi/10.1002/advs.202411325)\n\n**Unlocking the vastness of chemical matter and its biological potential reveals new avenues for precision drug discovery, empowered by data-driven insights and AI-enhanced mapping**\n\nPharmaceutical innovation increasingly relies on comprehensive maps of chemical and pharmacological spaces to guide the exploration of molecular diversity and target landscapes. The seminal work \u201cRemapping the Chemical Space and the Pharmacological Space of Drugs\u201d leveraged ChEMBL34 to project high-dimensional molecular fingerprints into two-dimensional UMAP representations, revealing that 81% of approved drugs contain aromatic rings and highlighting shifts toward kinase and epigenetic targets in [post-2020 approvals.](https://pmc.ncbi.nlm.nih.gov/articles/PMC11207054/)\n\nRecent advances further refine these maps: non-linear dimensionality reduction methods such as t-SNE, UMAP, and GTM deliver superior neighborhood preservation for large compound libraries [compared to PCA](https://pmc.ncbi.nlm.nih.gov/articles/PMC11733715/). AI-driven chemography platforms now integrate graph-based embeddings and generative models to navigate uncharted regions of chemical space, identifying novel scaffolds and repurposing opportunities by [mining multi-modal biomedical data.](https://www.sciencedirect.com/science/article/pii/S1359644625001059)\n\nSimultaneously, pharmacological space analyses highlight an accelerating trend in kinase inhibitor development alongside emergent epigenetic writer targets, underscoring the dynamic interplay between [chemical innovation and therapeutic focus](https://pmc.ncbi.nlm.nih.gov/articles/PMC11207054/).\n\nLooking ahead, hybrid workflows that combine descriptor-based, rule-based, and AI-driven methods\u2014augmented by explainable AI\u2014will enable systematic traversal of both beneficial and [\u201cdark\u201d chemical regions, driving safer and more targeted molecule design](https://www.frontiersin.org/journals/drug-discovery/articles/10.3389/fddsv.2025.1674289/full).\n\nExploration of chemical and pharmacological spaces is poised to become more predictive and personalized. As databases expand, true innovation will depend not just on scale but on the strategic integration of structural diversity, biological relevance, and human-augmented AI insights. By weaving together cheminformatics, network pharmacology, and machine learning, the drug discovery roadmap promises to chart new territories previously hidden within the chemical universe.", "# Unlocking Protein Structure Insights without a Lab: Harnessing CRO Services for Structural Biology Research\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-20 on marcdeller.com (categories: sb).*\n\n* **CRO partnerships:** eliminate the need for costly internal lab infrastructure exceeding $5M for protein structure.\n* **Full protein structure workflows:** available from cloning, expression, purification, through crystallization and data collection.\n* **Protein purification:** uses chromatographic techniques and quality control like SDS-PAGE and DLS to ensure purity and homogeneity.\n* **Crystallization:** employs vapor diffusion with varied chemical conditions optimized for crystal quality.\n* **Data collection:** performed at synchrotron or advanced X-ray sources with cryoprotection to avoid radiation damage.\n* **Structure determination:** accomplished through molecular replacement or experimental phasing with iterative refinement targeting <2\u00c5 resolution.\n* **Challenges include:** solubility, crystallization reproducibility, and cost variability depending on protein complexity.\n\n* **Pharmaron Structural Biology Services:**\n  <https://www.pharmaron.com/services/laboratory-services/structural-biology-services/>\n* **SARomics Drug Discovery CRO Outsourcing Benefits:**\n  [https://www.saromics.com/drug-discovery-cro-outsourcing](https://www.saromics.com/drug-discovery-cro-outsourcing/)\n* **Protein Crystallization Market Report:**\n  <https://www.mordorintelligence.com/industry-reports/protein-crystallization-market>\n* **Next Generation Technologies for Protein Structure Determination:** <https://pubmed.ncbi.nlm.nih.gov/40382917/>\n* **NanoImagingServices Cryo-EM Structural Biology Services:**\n  <https://www.nanoimagingservices.com/structural-biology>\n* **Helix BioStructures Protein & Cryo-EM Services:**\n  <https://www.lifesciencesreview.com/helix-biostructures>\n* **Biortus Strategic Collaboration and Protein Structure Pipeline:** <http://www.biortus.bio/en/newsdetails/17/13.html>\n\nStructural biology is the cornerstone of understanding molecular mechanisms that govern life, enabling advances from fundamental science to drug discovery. Yet, establishing a cutting-edge laboratory for experimental protein structure determination can demand hefty investments exceeding $5 million, including sophisticated equipment, dedicated space, and specialized personnel. For many academic and small biotech groups, as well as emerging pharmaceutical startups, these costs pose a formidable barrier. Fortunately, Contract Research Organizations (CROs) specializing in structural biology offer a transformative alternative to rapidly and cost-effectively access state-of-the-art expertise and experimental pipelines.\n\nLeveraging established CRO partnerships enables researchers to bypass expensive infrastructure by outsourcing critical steps such as cloning, protein expression, purification, crystallization, structural data collection, and analysis. These service providers bring together integrated platforms of high-throughput facilities, seasoned scientists, and synchrotron access, dramatically accelerating the journey from gene to atomic-resolution protein structures. The common workflow begins by cloning the gene encoding the target protein into an expression vector optimized for crystallography studies. Protein expression is often performed in diverse host systems\u2014bacteria, yeast, insect, or mammalian cells\u2014to ensure proper folding and relevant post-translational modifications. Optimization of expression conditions quickly yields sufficient quantities of soluble protein suitable for downstream analysis.\n\nProtein purification is next, where CROs employ sophisticated chromatographic techniques such as affinity, ion-exchange, and size-exclusion chromatography. Rigorous quality control assessments including SDS-PAGE and Dynamic Light Scattering confirm protein purity above 95% and a monodisperse state, prerequisites for successful crystallization. Crystallization trials themselves are one of the most critical and challenging aspects. CROs utilize diverse approaches like vapor diffusion using hanging or sitting drop techniques across extensive chemical screens with various precipitants, salts, and pH buffers. Optimization aims to produce high-quality crystals capable of diffracting X-rays to high resolution.\n\nData collection follows, typically conducted at specialized synchrotron facilities or advanced lab-based X-ray sources. Crystals are cryo-protected and frozen in liquid nitrogen to protect against radiation damage during data acquisition. Sophisticated data processing software such as CCP4 or XDS transforms diffraction patterns into electron density maps. Structural determination is accomplished by molecular replacement or experimental phasing techniques, including selenomethionine labeling. Iterative refinement hones the atomic model to ideally achieve resolution around or better than 2 angstroms.\n\nChallenges remain, such as protein solubility issues, variability in crystallization reproducibility, and fluctuations in data quality. Cost models vary widely depending on protein complexity, the necessity for de novo structure elucidation, and chosen CRO expertise. Selecting providers with integrated services that combine protein production and crystallographic analysis\u2014like SARomics and Biortus\u2014can streamline workflows, reducing timelines and costs. Moreover, CROs increasingly invest in automation, cloud-based data handling, and AI-driven approaches, further enhancing efficiency and reproducibility.\n\nThe rise in demand for outsourcing early-stage drug discovery and structural biology to CROs reflects industry trends. Many pharmaceutical and biotech firms now prefer variable cost structures over fixed capital investments, appreciating transparent milestone-based billing that reduces risk and improves capital efficiency. The capacity of CROs to offer comprehensive solutions, from cloning through structure-guided medicinal chemistry, supports faster lead compound development and novel therapeutic discovery.", "# Crystal Structure of MRE11 from Thermotoga maritima MSB8 (TM1635) at 2.20 \u00c5 Resolution\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID**: [2Q8U](https://www.rcsb.org/structure/2Q8U)\n* **Release Date:** 2007-06-26\n* **Method:** X-RAY DIFFRACTION 2.2 \u00c5\n* **Organisms:** Thermotoga maritima MSB8\n* **Macromolecules:** Exonuclease, putative (protein)\n* **Publication:** [Das, D., et al. (2010). Crystal Structure of the First Eubacterial Mre11 Nuclease Reveals Novel Features That May Discriminate Substrates During DNA Repair. Journal of Molecular Biology, 397, 647-663.](https://doi.org/10.1016/j.jmb.2010.01.049)\n\n**Description:** This is the first crystal structure of an eubacterial Mre11 nuclease, a central player in DNA double-strand break repair. The 2.2 \u00c5 structure from T. maritima reveals both conserved and novel features compared to the archaeal enzyme, particularly in the DNA-specificity domain and dimeric organization, suggesting how different organisms have evolved distinct mechanisms for substrate discrimination during DNA repair.", "# Google Scholar profile of Marc C. Deller\nAuthors: Deller, M.C.\nSource: Google Scholar, 2026\n\nCitation statistics (all time / since 2021): citations 3967 / 1343; h-index 25 / 18; i10-index 45 / 26.\n\n## Publications listed (73), by citation count", "# flexappeal.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n## flexappeal.mdeller.com/\n\n```\n   ______         ___                      __\n  / __/ /____ __ / _ | ___  ___  ___ ___ _/ /\n / _// / -_) \\ // __ |/ _ \\/ _ \\/ -_) _ `/ /\n/_/ /_/__/__\\/_/ |_/ .__/ .__/__/_,_/_/\n                    /_/  /_/\n```\n\nFlexAppeal\n\n## Start with a structure\n\nFlexAppeal turns the OpenMM option surface into a form, then hands you a single\nself-contained file to run on your own machine. Nothing is simulated on this\nserver \u2014 your trajectory never leaves your laptop.\n\n### Lysozyme, in motion\n\n![](/static/lysozyme-poster.webp?v=1785178104)\n\nThe first half of the Example tab's 10\u00a0ns run, played back. Cartoon ribbon, secondary structure coloured.", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n| Problem | What is happening, and what to do |\n|---|---|\n| `setup-plip` fails, or `pip install plip` tries to build OpenBabel from source | Expected without conda-forge -- `plip`'s own installer forces a from-source OpenBabel rebuild unless OpenBabel is already importable *inside pip's build sandbox*, and the standalone PyPI `pymol-open-source` wheel has a hardcoded broken library path on at least some machines. `setup-plip` works around both via a self-downloaded micromamba; re-run `python3 BoltzMaker.py setup-plip --force` if a previous attempt left a half-built `.plip_env`. |\n| A `preflight`/`analyze` step involving `.plip_env` errors with `ModuleNotFoundError: No module named 'chatmol'` (or similar) | A stray `~/.pymolrc.py` (e.g. from an unrelated PyMOL plugin) is being loaded by the bundled PyMOL. BoltzMaker already overrides `HOME` for these subprocess calls so this shouldn't reach you, but if it does, check `~/.pymolrc.py` for anything referencing a package not installed in `.plip_env`. |\n| `run` seems to hang with no progress, or your Mac gets extremely slow | Check the memory figure in the progress bar and see **Memory on Mac** above -- very likely swap-thrashing, not a genuine stall. Re-run with a lower `--mps-watermark`, `--workers 1` and `--max-parallel-samples 1`. |\n| A target's YAML/CIF exists on disk but BoltzMaker says it's missing, or `preflight` hangs | Check for iCloud \"Optimize Mac Storage\" dataless files -- `preflight`'s `icloud_materialize` check handles this automatically, but a very large campaign can take a while to force-download everything on first run. |\n| `boltz` fails during `setup` with a `numpy` build error | You're likely on Python 3.13+. `boltz` pins `numpy<2.0`, which has no prebuilt wheel past cp312 -- `_find_boltz_python()` already looks for a `python3.12` specifically; install one (`brew install python@3.12`) if it can't find one. |\n| A run dies with `torch._C._LinAlgError: linalg.svd ... failed to converge because the input matrix is ill-conditioned` | Not a conditioning problem, and not an OOM. Measured on this hardware: of every degenerate 3x3 (rank-1, zeros, repeated singular values, 1e-20, 1e20) plus 2000 random matrices, the only input that raises it is one containing **NaN** -- the diffusion coordinates have already diverged before the alignment runs. `linalg.svd` is not even an MPS op (it falls back to CPU), so forcing it to CPU changes nothing. Apply `patches/apply_boltz_patches.py` so the failure is contained to that one target, then try that target with `--no-potentials`, since the physical-guidance coordinate update is one of the places a trajectory can diverge. |\n| A run dies with `RuntimeError: shape mismatch: value tensor of shape [N] cannot be broadcast to indexing result of shape [M]`, part-way through a target | An Apple-silicon masked-select miscount inside boltz's flat-bottom steering potential, not a problem with your inputs. The same boolean mask compacts to a different number of elements on two tensors of identical shape, so an assignment that is arithmetically fine raises. It is intermittent (it needs a coordinate to leave its bounds on that particular trajectory), so the same target can fail twice and then succeed, and N and M differ by anything from 1 to 623. Run `python3 patches/apply_boltz_patches.py`, which replaces the assignment with `torch.where` (identical arithmetic, no compaction) and installs a reporter that names every operand shape if anything similar recurs. Re-run it after any `boltz` upgrade. |\n| One target's failure kills the whole `boltz predict`, and targets behind it never run | `boltz`'s `predict_step` skips a batch on out-of-memory but re-raises everything else, and `LinAlgError` subclasses `RuntimeError`. Run `python3 patches/apply_boltz_patches.py` (idempotent, keeps `.orig` backups) -- `preflight`'s `boltz_patches` row reports whether it is applied. Re-run it after any `boltz` upgrade, which silently reverts it. |\n| A target fails preflight with a chain-id-length error | Boltz truncates chain IDs to 5 characters internally (a fixed-width field in its own schema) and silently corrupts longer ones rather than erroring at parse time -- shorten the protein/partner/ligand name in `boltz_input.md`. |\n| The dashboard's charts (or the binding-site 3D view) don't render, or look unstyled | plotly.js and 3Dmol.js are vendored and inlined (not CDN-loaded), so a missing network connection shouldn't cause this -- Google Fonts is still CDN-loaded for styling, so the page needs internet access at least once for the fonts to look right (falls back to a generic sans-serif otherwise; charts, 3D views and data are unaffected). If they genuinely don't render, check that `vendor/plotly-2.35.2.min.js` and `vendor/3Dmol-2.5.5-min.js` exist next to `BoltzMaker.py` -- `analyze` prints a warning and falls back to the relevant CDN (known not to work in some HTML-preview contexts) if either is missing. |\n| Upload rejected as too large | Uploads are capped at **200MB**. The packer aims below that (it stops at 180MB, dropping the largest structures first and recording each one in the manifest), so hitting this usually means an unusually large campaign. Analyse it locally with `boltz_dashboard.html` instead. |\n| \"This results file declares format version N\" | The `.bmz` was written by a different version of the bundle than the site understands. Prepare a fresh bundle and re-pack; the file itself is not damaged. |\n| \"No manifest.json in the upload\" | A campaign folder was uploaded instead of the `.bmz`. Either upload the `.bmz` the bundle wrote, or use Stepwise Mode's **Analyze**, which is the tool that takes a campaign folder. |\n| Some targets have no pose viewer | Those structures were dropped to stay under the size limit, or the target failed. `manifest.json` inside the `.bmz` records which, and why. They are still in `boltz_cif/` on your own machine. |\n| The 3D viewer says WebGL is unavailable | The browser has WebGL disabled or unsupported. Everything else on the page is unaffected. |\n| Preflight fails on `sse_comparison` or `result_packer` | The bundle is incomplete -- re-download it. Both are checked before any GPU time is spent precisely because `analyze` imports them only at the end, so a missing file used to surface after the prediction had already finished. |\n| Preflight warns on `vendor_assets` | Plotly or 3Dmol is missing from `vendor/`, so the dashboard will reach for a CDN and will not render offline. The run itself is unaffected. |\n| Preflight warns `boltz_cli` did not answer `--help` in 120s | A cold first import, not a broken install: `boltz --help` loads the whole torch stack, and a freshly solved environment byte-compiles it too. The bundle warms it before the campaign, and it is a warning that does not stop the run. |\n| The interactions panel is empty | PLIP either did not run for that target, or found nothing. If you switched **Skip PLIP interaction analysis** on when preparing, that is the cause. |\n| A session link stops working | Sessions are not deleted for being old, but the oldest are evicted once the space set aside for them fills, least recently opened first. Upload the file again; nothing is lost, since the `.bmz` is on your own disk. |", "# The RNA Revolution: Designing Drugs for a New Class of Targets\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* RNA is emerging as a new class of therapeutic target, moving drug discovery beyond its traditional focus on proteins.\n* Innovative computational methods like \u201cSHAMAN\u201d are being used to understand the dynamic structures of RNA, revealing binding sites for small molecules.\n* Structure-based design of RNA-targeting drugs offers a powerful new approach to treat diseases by intervening before harmful proteins are created.\n\n1. **What are the recent trends in Structural Biology and Drug Discovery:**<https://3decision.discngine.com/blog/2025/1/10/recent-trends-in-structural-biology-insights-from-the-psdi-conference>\n\nIn the central dogma of molecular biology, RNA has often been viewed as a simple messenger, a passive courier carrying genetic instructions from DNA to the protein-making machinery of the cell. For this reason, the vast majority of drug discovery efforts have focused on proteins as the ultimate therapeutic targets. This protein-centric view, however, overlooks the critical regulatory roles that RNA molecules play in health and disease. The complex, three-dimensional structures that RNA can fold into create unique pockets and surfaces that can be targeted with small molecules, yet our ability to design drugs for them has been hampered by a lack of understanding of their dynamic nature.\n\nWe are now beginning to recognize RNA as a viable and exciting new frontier for therapeutic intervention. The challenge lies in its flexibility; unlike rigid proteins, RNA structures are highly dynamic and exist in a constant state of flux. To address this, scientists are developing innovative computational techniques to map the \u201cconformational landscape\u201d of RNA molecules. One such method, called \u201cSHAMAN,\u201d uses accelerated molecular dynamics simulations to explore the many shapes an RNA molecule can adopt over time. By doing so, it can identify transient pockets that are ideal for binding small molecule drugs.\n\nThis structure-based approach is setting the foundation for the rational design of RNA-targeting drugs. The SHAMAN computational pipeline has already successfully identified known binding sites on RNA and ranked potential new ones, proving its potential as a tool for discovery. Targeting RNA directly offers a way to intervene earlier in the disease process, before a harmful protein is even made. This could lead to innovative treatments for a wide range of diseases, from viral infections to genetic disorders. We are moving beyond the protein-and-inhibitor paradigm and into a new era where RNA is no longer just the messenger, but a primary target for the next generation of precision medicines.", "# 2025 \u2013 Foundations: Data, Data, Everywhere\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-28 on marcdeller.com (categories: certificates).*\n\n* [Link to certificate](https://www.coursera.org/account/accomplishments/verify/4S09Q8ICKGXN)\n\n* **Award Details:** Google, through Coursera, issued this certificate on March 11, 2025.\n* **Course Focus:** It recognizes completion of a course covering foundational data concepts and analytical thinking.\n* **Career Value:** This matters because data literacy is central to rigorous and reproducible scientific research.\n* **Practical Use:** It helps with organizing, interpreting, and evaluating experimental datasets in science.", "# GitHub README: wordpressID\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/wordpressID, 2026\n\n> **Every post and page from any WordPress site, to CSV and screen, in one command.**\n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/wordpressID\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/wordpressID</a></td>\n</tr>\n</table>\n\n---\n\nA single-command tool that pulls every post and page from a WordPress site via the public REST API and outputs a clean, de-duplicated inventory of `id`, `type`, and `title`: written to CSV on disk and rendered as an aligned table on screen in the same pass.\n\nWhy it matters: getting a complete list of a WordPress site's content sounds trivial until the REST API silently caps at 100 items and drops the rest with no warning, leaving you with a quietly incomplete inventory. wp-content-inventory paginates properly, decodes the HTML entities WordPress bakes into titles, and renders a quoted CSV and an aligned on-screen table from a single fetch so the two can never disagree. It is useful for migrations, content audits, redirect mapping, and any maintenance task where you need a trustworthy, complete manifest of a site's posts and pages, using nothing but curl and jq against endpoints WordPress already exposes.\n\nNo plugins, no authentication, no scraping. Just `curl` and `jq` against the endpoints WordPress already exposes.\n\n## What it does\n\n- Enumerates all published **posts** and **pages** through the WordPress REST API (`/wp-json/wp/v2/`).\n- Paginates automatically, so it captures everything rather than silently stopping at the API's 100-item cap.\n- Decodes the HTML entities WordPress returns in titles (`&amp;`, smart quotes, en dashes).\n- Writes a properly quoted CSV to disk and prints a column-aligned, comma-safe table to the terminal.\n- Fetches each endpoint **once** and renders both outputs from that single capture, so the file and the screen can never drift apart.\n\n## Requirements\n\n| Tool | Purpose | Notes |\n|------|---------|-------|\n| `curl` | HTTP requests | Pre-installed on macOS and most Linux distros |\n| `jq` | JSON parsing and CSV/TSV encoding | `brew install jq` or `apt install jq` |\n| `column` | On-screen table alignment | Ships with macOS and `util-linux`; only needed for the table view |\n\n## Usage\n\nSet your target site and run:\n\n```bash\nSITE=\"https://marcdeller.com\"\n\ndata=$( for t in posts pages; do p=1; while r=$(curl -s \"$SITE/wp-json/wp/v2/$t?per_page=100&page=$p&_fields=id,title\") && echo \"$r\" | grep -q '\"id\"'; do echo \"$r\" | jq -c --arg t \"${t%s}\" '.[] | {id, title:(.title.rendered|gsub(\"&amp;\";\"&\")|gsub(\"&#8217;\";\"\\u2019\")|gsub(\"&#8211;\";\"-\")), type:$t}'; p=$((p+1)); done; done ); \\\n{ echo \"id,type,title\"; echo \"$data\" | jq -r '[.id,.type,.title]|@csv'; } > wp_content.csv; \\\n{ printf 'ID\\tTYPE\\tTITLE\\n'; echo \"$data\" | jq -r '[.id, .type, (if (.title|length)>70 then .title[0:69]+\"\\u2026\" else .title end)]|@tsv'; } | column -t -s$'\\t'\n```\n\nThis writes `wp_content.csv` and prints a table like:\n\n```\nID     TYPE  TITLE\n14613  post  2026 - Discovery of the Orally Bioavailable Isoform Selective Janus K\u2026\n13848  post  2026 - Anthropic Academy\n13485  post  Plastic, Persistent, and Potentially Carcinogenic\n...\n```\n\n## Output\n\nThe CSV holds the full, untruncated data:\n\n```csv\nid,type,title\n14613,\"post\",\"2026 - Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\"\n13848,\"post\",\"2026 - Anthropic Academy\"\n11930,\"page\",\"About\"\n```\n\n| Column | Description |\n|--------|-------------|\n| `id` | WordPress object ID |\n| `type` | `post` or `page` |\n| `title` | Decoded, CSV-quoted title |\n\nThe on-screen `title` is truncated to 70 characters purely for terminal width. The CSV always keeps the full title.\n\n## How it works\n\n**Pagination is mandatory, not optional.** The REST API caps `per_page` at 100 and will silently drop the overflow. A site with 193 posts returns exactly 100 from a naive call, with no error to warn you. The loop walks `&page=$p` until a page returns no IDs, so the count is always complete. You can confirm the true total against the `X-WP-Total` response header:\n\n```bash\ncurl -sI \"$SITE/wp-json/wp/v2/posts?per_page=100\" | grep -i x-wp-total\n```\n\n**Entity decoding.** WordPress serialises titles with HTML entities. The `gsub` chain handles the three most common (`&amp;`, `&#8217;`, `&#8211;`). Extend it for any others you hit (`&#8220;`/`&#8221;` smart quotes, `&hellip;`), or pipe titles through a full entity decoder if you would rather not play whack-a-mole.\n\n**Comma safety.** Titles routinely contain commas and parentheses. The CSV uses `jq`'s `@csv`, which quotes and escapes fields correctly. The screen table uses `@tsv` with `column -s$'\\t'`, so a comma inside a title never spills into the next column.\n\n**Single fetch, two views.** The API responses are captured once into `$data` as JSON lines, then both the CSV and the table are generated from that. No double requests, no risk of the two outputs disagreeing.\n\n## Customisation\n\n- **More metadata:** add fields via `_fields=id,title,slug,link,date` and extend the `jq` object and output arrays.\n- **Untruncated screen titles:** replace the `if (.title|length)>70 ...` block with plain `.title`.\n- **Wide tables:** append `| less -SX` to scroll horizontally without wrapping.\n- **Spreadsheet-friendly alignment:** swap `@csv` for `@tsv` if you prefer a tab-delimited file.\n- **Other content types:** add custom post types to the `for t in posts pages` list if your site registers them in the REST API.\n\n## Notes\n\nThe REST API must be publicly reachable (it is by default). Sites that disable or firewall `/wp-json/` will return nothing. Only published content is listed unless you authenticate; drafts and private items require an authenticated request.\n\n## Licence\n\nMIT.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/wordpressID\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/wordpressID</a></td>\n</tr>\n</table>", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\nMembers of this superfamily include TNFR, NGFR, FasR, CD27R, CD30R, CD40R, 4-1BBR and OX-40R (see Table 1-2). This receptor superfamily is structurally distinct and is characterised by long chain \u03b2-sheet structures, with a characteristic set of cysteine-rich repeats. Activation of TNFR requires receptor trimerisation, which is brought about by aggregation of receptor subunits on binding a TNF trimer (Banner et al., 1993). Other structures determined include the apoptosis inducing ligand TRAIL in complex with its receptor DR5 (Mongkolsapaya et al., 1999) (see Figure 1-3).\n\n#### 1.4.2.4 TGF/Activine Receptor Superfamily\n\nThis is a large superfamily of transmembrane serine/threonine kinases, which function in binding long chain \u03b2-sheet cytokines such as TGF and activine (see Table 1-2). Receptors of this superfamily contain short cysteine-rich extracellular domains and intracellular kinase domains. The structure of the signaling complex is not fully understood, although it is thought that oligomerisation between TGF receptors type I, type II and type III occurs (Nicola, 1994).\n\n#### 1.4.2.5 Immunoglobulin Receptor Superfamily\n\nThe IL-1 group of cytokines (IL-1\u03b1, IL-1\u03b2 and IL-1\u03b3) bind to this receptor superfamily (Sims et al., 1988). These receptors contain an immunoglobulin-like domain, which consists of seven anti-parallel \u03b2-strands, folded into two sheets, which are stabilised by a conserved disulphide bond. Structures determined include IL1R in complex with both IL1\u03b2 (Vigers et al., 1997) and its naturally occurring antagonist (Schreuder et al., 1997) (see Figure 1-3).\n\n#### 1.4.2.6 Tyrosine Kinase Receptor Superfamily\n\nMembers of this superfamily include CSFR, c-kit, EGFR, IGF-1R, FGFR and PDGFR. These receptors have a cytoplasmic domain with tyrosine kinase activity. All members of this family undergo ligand induced homodimerisation, which brings the cytoplasmic domains together, so activating the kinase domain and initiating signal transduction (Ullrich and Schlessinger, 1990). Members of this superfamily studied structurally include the Fms-like tyrosine kinase receptor (Flt-1) (Wiesmann et al., 1997) (see Figure 1-3). Intracellular portions of these receptors have also been solved, for example the activated insulin receptor tyrosine kinase in complex with peptide substrate and ATP analogue (Hubbard, 1997) (see Figure 1-3).\n\n#### 1.4.2.7 Guanine Nucleotide-Binding Protein-Coupled Receptor Superfamily\n\nChemokines such as IL-8 bind to this superfamily of structurally distinct receptors (see Figure 1-1 and Table 1-2). Receptors of this superfamily consist of a \"serpentine seven pass\" motif containing seven hydrophobic membrane-spanning domains and relatively short N- and C-terminal regions. Activation of the receptor complex leads to the formation of GTP bound \u03b1-subunits of a G-protein complex (G\u03b1-GTP) which is subsequently released from the receptor and functions as an intracellular second messenger (Gilman, 1987; Neer and Smith, 1996). Structural studies of this superfamily are limited due to the hydrophobic nature of such membrane spanning proteins, although NMR studies have been carried out on IL-8 in complex with N-terminal peptide fragments of the CXCR1 receptor (Skelton et al., 1999).\n\n### 1.4.3 Receptor Activation and Signaling Through Haemopoietin Receptors\n\nSignal transduction through cytokine receptors generally requires ligand induced receptor homodimerisation or heterodimerisation (Heldin, 1995). For example, homodimerisation is required for activation of M-CSFR, SCFR, EpoR, G-CSFR, GHR and PrlR (Watowich et al., 1992). However, as discussed above the active signaling complex of many other cytokines, for example the gp130 family, requires heteroligomerisation of different receptor subunits. In general ligands first bind to specific receptor \u03b1-subunits so forming a low-affinity receptor-ligand complex. Further receptor components (receptor \u03b2-subunits and receptor \u03b3-subunits) may bind and this oligomerisation leads to the formation of high-affinity activated receptor-ligand complexes.\n\nThe cytoplasmic regions of most cytokine receptors do not possess any intrinsic enzymatic activity and with the exception of IL-2R and EpoR sequence similarity in these regions is low. There are, however, conserved regions in the intracellular domain termed box 1, box 2 and box 3 (Murakami et al., 1991) and the recently identified suppresser of cytokine signalling (SOCS) box (Zhang et al., 1999). These regions interact with cytosolic signaling molecules and so regulate signal transduction (Fukunaga et al., 1993; Murakami et al., 1993). The membrane proximal motif, box 1, is characteristically rich in proline residues and is essential for association with the Janus Kinase (JAK) family of proteins (Ihle and Kerr, 1995; Ihle et al., 1994; Taga and Kishimoto, 1995). JAKs responsible for gp130 signaling include JAK1, JAK2 and Tyk2, whereas the PrlR and GHR systems utilise only JAK2. JAK molecules are brought together through receptor dimerisation and this leads to phosphorylation of the conserved KEYY motif of the kinase activation loop. The associated increase in kinase activity leads to tyrosine phosphorylation of the receptor. Once phosphorylated, this region becomes a binding site for Src homology 2 domains (SH2) of various signaling proteins such as Shc and Grb of the mitogen-activated protein (MAP) kinase pathway. Alternatively, signal transducers and activators of transcription (STAT) molecules can also associate with the phosphorylated receptors through SH2 domains. Phosphorylation of the STAT molecules induces formation of homo- or heterodimers, which then migrate to the nucleus where they bind interferon-\u03b3 activation sites (GAS) of various promoter elements. This functions in activating transcription (Darnell et al., 1994). STAT1 and STAT3 are activated in response to gp130 stimulation, whereas, PrlR and EpoR utilize STAT 5 and GHR activates STAT1 (Ihle and Kerr, 1995).", "# Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3\nAuthors: Shvartsbart, A.; Roach, J.J.; Witten, M.R.; Koblish, H.; Harris, J.J.; Covington, M.; Hess, R.; Lin, L.; Frascella, M.; Truong, L.; Leffet, L.; Conlen, P.; Beshad, E.; Klabe, R.; Katiyar, K.; Kaldon, L.; Young-Sciame, R.; He, X.; Petusky, S.; Chen, K.-J.; Horsey, A.; Lei, H.-T.; Epling, L.B.; Deller, M.C.; Vechorkin, O.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2022\n\nSubstitution on the carbon adjacent to the pyridazine nitrogen (R1) was subsequently explored. We quickly identified methoxy derivative 5 (Table 1), which had an improved potency in FGFR3 (IC50 = 4.6 nM) while maintaining similar levels of selectivity over FGFR1 (19\u00d7) compared to 4.\n\nReplacement of the phenyl ring of 5 with m-pyridyl was tolerated, as exemplified by 6 (Table 2). The ensuing SAR was carried out interchangeably with the benzamide or m-pyridyl amide. The improvement in potency resulting from methoxy substitution at R1 encouraged further exploration into this pocket. Initially, this region appeared intolerant of larger substituents, since a modest increase in size to the ethoxy derivative 8 led to a 3-fold loss in potency relative to 6 (FGFR3 IC50 = 56 nM vs 13 nM). We probed for lipophilic or hydrophilic interactions by installing larger groups such as phenyl (10), benzyl (11), and diethylamino (12), but in all cases, significant losses in FGFR3 enzyme potency were observed.\n\n*Table 2.* SAR Adjacent to the Pyridazine Nitrogen.\n\n| Compound | R1 | FGFR3 IC50 (nM) | FGFR1 IC50 (nM) |\n|---|---|---|---|\n| 6 | OMe | 13 | 428 |\n| 7 | NHMe | 34 | 420 |\n| 8 | OEt | 56 | 477 |\n| 9 | Oi-Pr | 54 | 365 |\n| 10 | OPh | 1177 | 2648 |\n| 11 | OBn | 441 | 362 |\n| 12 | OCH2CH2N(Me)2 | 247 | 415 |\n\nThe most critical trend in the SAR was the observation that expanding the ethoxy group to the more sterically encumbered isopropoxy group (9) did not negatively affect potency, suggesting that springs might be tolerated in this region. To assess how we might capitalize on this observation, compound 9 was docked into the ATP binding site of FGFR3 (Figure 2) using a known crystal structure (PDB code: 4K33). Our generated model showed two key interactions: the nitrogen of the imidazole ring forming a hydrogen bond to Ala558 in the hinge region and the amide carbonyl forming a hydrogen bond to Lys508 with the N-methyl group extending past the gatekeeper (Val555). Importantly, we noticed that the isopropyl group was situated close to an asparagine residue in the lower hinge region (Asn562, Figure 2a). The proximity of these substituents, in combination with the SAR, guided a design strategy aimed at delivering a heteroatom to engage the primary amide N\u2212H of Asn562 in a hydrogen bonding interaction. To this end, the isopropyl group was cyclized to (hetero)cycloalkyl rings of varying size and substitution (Table 3).\n\n*Figure 2.* (a) MOE docking of 9 in apo FGFR3 (PDB ID: 4K33). (b) Design of the THF ether to engage Asn562 (PDB ID: 4K33).\n\nThe azetidine derivative 13 lost significant potency for FGFR3 (IC50 = 515 nM), while the lactam 14 was better tolerated (FGFR3 IC50 = 18 nM) but offered no improvement relative to 5. Substitution with tetrahydropyran (15) also led to a significant loss of potency and selectivity. We were delighted to find that the oxetane derivative 16 improved the FGFR3 potency (IC50 = 1.1 nM) and further increased the selectivity over FGFR1 (65\u00d7). The racemic tetrahydrofuran (THF)-substituted compound 17 also enhanced the potency (FGFR3 IC50 = 2.4 nM) and FGFR1 selectivity (R1/R3 = 49\u00d7). We prepared the single enantiomers of 17 and found that the (S)-enantiomer 19 was the active isomer. Compound 19 was the first in this series to display sub-nanomolar enzymatic potency (FGFR3 IC50 = 0.9 nM). Given that FGFR1 also contains an asparagine residue at this position (Asn571), molecular modeling was ineffective at explaining the increase in selectivity for 19 relative to 5. A slight conformational difference in the protein backbone could possibly make this interaction more favorable in FGFR3.\n\n*Table 3.* SAR of (Hetero)Cycloalkyl Ethers.\n\n*(the body of Table 3 is an image-based structure table and did not survive text extraction from the PDF; the compounds and the potency values it summarises \u2014 azetidine 13, lactam 14, tetrahydropyran 15, oxetane 16, racemic tetrahydrofuran 17 and the single enantiomers 18/19 \u2014 are described in the surrounding text)*\n\nThe most potent compounds (16, 17, and 19) were tested in a homogeneous time-resolved fluorescence (HTRF) cellular assay using TEL-FGFR3-BA/F3 cells, a functional assay measuring compound inhibition of phospho-FGFR3. In addition to the best FGFR3 potency, 19 also exhibited the most potent cellular activity (BA/F3 IC50 = 15 nM).\n\nA broad in-house kinase panel screen of an early compound related to 4 identified four other kinases that were inhibited with IC50 < 100 nM: PDGFRB, KIT, FLT3, and TRKA (data not shown). Unfortunately, 19 was found to have less than desirable selectivity over these kinases as well (Table 4). To mitigate the possibility of off-target toxicity, we desired 100-fold selectivity over all other kinases.\n\nReplacement of the benzamide ring in 19 with m-pyridyl (compound 20) dramatically improved kinase selectivity without affecting FGFR3 potency (FGFR3 IC50 = 2.2 nM). However, compound 20 did not translate well from enzyme to cell (BA/F3 IC50 = 969 nM), undoubtedly a result of poor cellular permeation due to high polarity (C log P = 0.02). Consequently, we were forced to identify an alternative solution for improving the kinase selectivity of 19.", "# Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\nAuthors: Hummel, J.R.; Xiao, K.-J.; Yang, J.C.; Epling, L.B.; Mukai, K.; Ye, Q.; Xu, M.; Qian, D.; Huo, L.; Weber, M.; Roman, V.; Lo, Y.; Drake, K.; Stump, K.; Covington, M.; Kapilashrami, K.; Zhang, G.; Ye, M.; Diamond, S.; Yeleswaram, S.; Macarron, R.; Deller, M.C.; Wee, S.; Kim, S.; Wang, X.; Wu, L.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2024\n\n**N-(1-(Methylsulfonyl)piperidin-4-yl)-4-(1-(tetrahydro-2H-pyran-4-yl)-1H-pyrazol-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (3).** A mixture of 4-chloro-N-(1-(methylsulfonyl)piperidin-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (25, 210 mg, 0.585 mmol), 1-(tetrahydro-2H-pyran-4-yl)-4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazole (20, 195 mg, 0.702 mmol), Pd(dppf)Cl2\u00b7CH2Cl2 (47.8 mg, 0.059 mmol), and Na2CO3 (124 mg, 1.17 mmol) in 1,4-dioxane/H2O (2:1, 15 mL) was stirred at 100 \u00b0C overnight. After cooling to r.t., the reaction mixture was extracted with EtOAc. The combined organic phases were dried over MgSO4, filtered, and concentrated. The residue was purified by flash chromatography (SiO2, EtOAc/CH2Cl2). The product was further purified by prep HPLC (pH 2) to afford the title compound (65 mg, 23%) as an off-white solid. LCMS calculated for C19H26F3N6O3S (M + H)+: m/z = 475.2; found, 475.1.\n\n**4-(1-Cyclohexyl-1H-pyrazol-4-yl)-N-(1-(methylsulfonyl)piperidin-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (4).** A mixture of 4-chloro-N-(1-(methylsulfonyl)piperidin-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (25, 50 mg, 0.14 mmol), 1-cyclohexyl-4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazole (58 mg, 0.21 mmol), Pd(dppf)Cl2\u00b7CH2Cl2 (22.8 mg, 0.028 mmol), and Na2CO3 (29.5 mg, 0.28 mmol) in CH3CN/H2O (5:1, 2.4 mL) was irradiated in a microwave reactor at 100 \u00b0C for 30 min. After cooling to r.t., the crude mixture was purified by prep HPLC (pH 2) to afford the title compound (27.4 mg, 42% yield) as a white solid. LCMS calculated for C20H28F3N6O2S (M + H)+: m/z = 473.2; found, 473.1.\n\n**4-(1-(1-Methylpiperidin-4-yl)-1H-pyrazol-4-yl)-N-(1-(methylsulfonyl)piperidin-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (5).** A mixture of 4-chloro-N-(1-(methylsulfonyl)piperidin-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (25, 325 mg, 0.91 mmol), 1-methyl-4-(4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazol-1-yl)piperidine (396 mg, 1.36 mmol), Pd(dppf)Cl2\u00b7CH2Cl2 (148 mg, 0.18 mmol), and Na2CO3 (192 mg, 1.81 mmol) in CH3CN/H2O (5:1, 6 mL) was irradiated in a microwave reactor at 100 \u00b0C for 30 min. After cooling to r.t., the mixture was diluted with CH2Cl2, filtered, and concentrated. The residue was purified by flash chromatography (SiO2, EtOAc/hexanes followed by MeOH/CH2Cl2). The product was further purified by prep HPLC (pH 2) to afford the title compound (96.8 mg, 22% yield) as a white solid. LCMS calculated for C20H29F3N7O2S (M + H)+: m/z = 488.2; found, 488.1.\n\n**4-(1-(2-Fluorophenyl)-1H-pyrazol-4-yl)-N-(1-(methylsulfonyl)piperidin-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (6).** The title compound was prepared according to the procedure described for 4 with (1-(2-fluorophenyl)-1H-pyrazol-4-yl)boronic acid replacing 1-cyclohexyl-4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazole. White solid; 54% yield (36.5 mg, 75.3 \u03bcmol). LCMS calculated for C20H21F4N6O2S (M + H)+: m/z = 485.1; found, 485.0.\n\n**4-(1-(2-Chlorophenyl)-1H-pyrazol-4-yl)-N-(1-(methylsulfonyl)piperidin-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (7).** The title compound was prepared according to the procedure described for 4 with (1-(2-chlorophenyl)-1H-pyrazol-4-yl)boronic acid replacing 1-cyclohexyl-4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazole. Off-white solid; 11% yield (45 mg, 90 \u03bcmol). LCMS calculated for C20H21ClF3N6O2S (M + H)+: m/z = 501.1; found, 501.2.\n\n**N-(1-(Methylsulfonyl)piperidin-4-yl)-4-(1-(2,2,2-trifluoroethyl)-1H-pyrazol-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (8).** The title compound was prepared according to the procedure described for 5 with 4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1-(2,2,2-trifluoroethyl)-1H-pyrazole (35) replacing 1-methyl-4-(4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazol-1-yl)piperidine. White solid; 29% yield (126 mg, 0.267 mmol). LCMS calculated for C16H19F6N6O2S (M + H)+: m/z = 473.1; found, 473.1.\n\n**2-Methyl-1-(4-(2-((1-(methylsulfonyl)piperidin-4-yl)amino)-5-(trifluoromethyl)pyrimidin-4-yl)-1H-pyrazol-1-yl)propan-2-ol (9).** A mixture of 4-chloro-N-(1-(methylsulfonyl)piperidin-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (25, 86.1 mg, 0.240 mmol), 2-methyl-1-(4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazol-1-yl)propan-2-ol (38, 96 mg, 0.360 mmol), Pd(dppf)Cl2\u00b7CH2Cl2 (39.2 mg, 0.048 mmol), and Na2CO3 (63.6 mg, 0.60 mmol) in 1,4-dioxane/H2O (5:1, 1.2 mL) was stirred at 100 \u00b0C for 16 h. After cooling to r.t., the mixture was diluted with CH3CN, H2O, and TFA (0.1 mL) and purified by prep HPLC (pH 2) to afford the title compound (64 mg, 58% yield) as an off-white solid. LCMS calculated for C18H26F3N6O3S (M + H)+: m/z = 463.2; found, 463.3.\n\n### Synthesis of Intermediate 28 and Compounds 10 and 11\n\n**Step 1. 4-(1-(1-Ethoxyethyl)-1H-pyrazol-4-yl)-N-(1-(methylsulfonyl)piperidin-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (27).** A mixture of 4-chloro-N-(1-(methylsulfonyl)piperidin-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (25, 1.5 g, 4.2 mmol), 1-(1-ethoxyethyl)-4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazole (26, 1.22 g, 4.6 mmol), Pd(dppf)Cl2\u00b7CH2Cl2 (0.34 g, 0.42 mmol), and Na2CO3 (1.11 g, 10.5 mmol) in CH3CN/H2O (5:1, 24 mL) was stirred at 90 \u00b0C for 70 min. After cooling to r.t., the reaction mixture was diluted with H2O and extracted with EtOAc. The combined organic phases were washed with brine, dried over MgSO4, and concentrated. The residue was purified by flash chromatography (SiO2, EtOAc/hexanes) to afford the title compound (1.30 g, 67% yield) as an off-white solid. LCMS calculated for C18H26F3N6O3S (M + H)+: m/z = 463.2; found, 463.2.", "# GitHub README: FlexAppeal\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/FlexAppeal, 2026\n\nThe command-line interface exists so every stage can be driven without a browser:\n\n| Command | What it does |\n|---|---|\n| `./FlexAppeal.py serve` | Run the web app locally |\n| `./FlexAppeal.py inspect 1aki.pdb` | Report chains, gaps, disulfides, cofactors and the estimated system size |\n| `./FlexAppeal.py fetch 1AKI` | Retrieve a structure from the RCSB, OPM or the AlphaFold DB |\n| `./FlexAppeal.py bundle 1aki.pdb` | Build a run bundle without the browser |\n| `./FlexAppeal.py unpack run.command` | List or extract a bundle's contents |\n| `./FlexAppeal.py validate config.json` | Check a configuration against the registry |\n| `./FlexAppeal.py docs` | Regenerate `docs/options.md` from the option registry |\n| `./FlexAppeal.py sweep` | Remove expired scratch sessions |\n\n### \u25b6\ufe0f Running a bundle\n\n```bash\nchmod +x flexappeal_myrun.command\n./flexappeal_myrun.command\n```\n\nThe `chmod` is required: a file downloaded from a browser arrives without the executable bit. The first run installs the environment (several gigabytes, 5 to 15 minutes); later runs go straight to simulating. Re-running resumes from the last checkpoint.\n\n## \ud83d\udcca Output\n\nEverything lands in one directory beside the bundle:\n\n| File | What it is |\n|---|---|\n| `<job>.fxa` | **The results file.** Upload this to the Analysis tab |\n| `prepared.pdb` | After repair, before solvation |\n| `solvated.pdb` | The complete system as simulated |\n| `trajectory.xtc` | The full trajectory, which stays on your machine |\n| `state_data.csv` | Energies, temperature, pressure and density over time |\n| `state.chk` | Checkpoint, for resuming |\n| `system.xml`, `integrator.xml`, `final_state.xml` | Full serialisation, for exact reproduction or extension |\n| `run_manifest.json` | Versions, platform, achieved ns/day, wall time |\n\n## \ud83e\uddea Validation\n\nThe physics is checked against known values rather than against itself. Every fixture in the test suite is a real results file from a real run.\n\n| System | Measured | Expected |\n|---|---|---|\n| Lysozyme (1AKI) | R<sub>g</sub> 14.13 \u00c5 | ~14.3 \u00c5 |\n| Lysozyme | Helix content 39% | ~40% |\n| Lysozyme | 4 disulfides at 6\u2013127, 30\u2013115, 64\u201380, 76\u201394 | Exactly those four |\n| Trypsin + benzamidine (3PTB) | Asp189 contact 100% of frames | The defining S1 salt bridge |\n| Trypsin + benzamidine | Ser190, Trp215, Gly216, Gly219, Gly226 | The canonical S1 pocket |\n| Glycophorin A in POPC (1AFO) | Bilayer thickness 3.64 nm | 3.7\u20134.0 nm |\n| Glycophorin A in POPC | Order parameters 0.20\u20130.25 plateau, falling to 0.12 | The fluid-bilayer profile |\n\nA further 30 tests resolve every OpenMM symbol the option registry names against the installed package, and 8 more actually run `addSolvent` for each water model rather than merely checking that its file loads.\n\n## \ud83e\udded Caveats\n\nStated plainly, because each is a real limit rather than a rough edge:\n\n- **Metal cofactors are refused.** Haem, zinc fingers and iron-sulfur clusters cannot go through GAFF, OpenFF or espaloma, which are organic force fields with no transition-metal parameters. FlexAppeal says so at build time instead of failing inside your run.\n- **Ligand protonation comes from the RCSB chemical definition**, which is deposited in one fixed state. That state is frequently not the dominant one at the pH the protein is simulated at. The formal charge used is printed in the run log.\n- **Area per lipid counts the protein's cross-section as membrane area.** It is a convergence diagnostic, not a number to compare against a pure-bilayer literature value.\n- **OpenMM has no official Metal back end.** The `openmm-metal` plugin is third-party and experimental, and Apple has deprecated OpenCL. Rather than promise a back end, the bundle benchmarks what is actually present. On an M1 Max the reliable levers are hydrogen mass repartitioning with a 4 fs timestep, mixed precision, and a sensible thread count.\n- **Single-component membranes only.** `Modeller.addMembrane` builds one lipid type. Anything more complex should be built in CHARMM-GUI and uploaded.\n- **No free-energy methods.** No FEP, umbrella sampling, metadynamics or replica exchange. Each is a project in itself.\n\n## \ud83e\uddf1 Stack\n\n| Layer | Choice |\n|---|---|\n| Simulation | OpenMM 8.5, PDBFixer, openmmforcefields, OpenFF toolkit |\n| Analysis | MDTraj, NumPy, SciPy |\n| Web | Flask, Jinja, vanilla JavaScript (no framework, no build step) |\n| Visualisation | Plotly 2.35, Mol\\* 5.11 (both vendored) |\n| Structure parsing | gemmi |\n| Environments | pixi for development and for each generated bundle |\n\n## \ud83d\udee0\ufe0f Web deployment\n\nFlexAppeal is deployed to a DigitalOcean droplet shared with four other applications, on port 8004.\n\n```bash\ncp deploy/.env.example .env      # fill in DROPLET_SSH and SERVER_NAME\n# On the droplet, once, as root:\nbash /opt/flexappeal/deploy/provision.sh\n# From your Mac, thereafter:\nbash deploy/deploy.sh\n```\n\n| File | Role |\n|---|---|\n| `deploy/provision.sh` | One-time and idempotent: user, venv, systemd units, nginx site, TLS |\n| `deploy/deploy.sh` | Sync, install dependencies, restart, verify `/healthz` |\n| `deploy/gunicorn.conf.py` | Two sync workers, 300 s timeout |\n| `deploy/flexappeal-web.service` | Hardened unit with a `MemoryMax` cgroup limit |\n| `deploy/flexappeal-scratch-clean.{service,timer}` | Sweeps abandoned sessions every 15 minutes |\n| `deploy/nginx-flexappeal.conf` | Static assets from disk, rate limits, 250 MB body cap, and logging left on for `/static/` |\n| `deploy/nginx-flexappeal-limits.conf` | The rate-limit zone, which must be declared at http scope |\n\nThe values duplicated across Python, nginx, gunicorn and systemd are annotated \"MUST stay in sync\" in both files, and `tests/test_deploy.py` enforces the agreement.", "# Protein Structure and Redox Biology: NOS Linkages as Allosteric Switches\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: blog).*\n\n* Newly discovered nitrogen-oxygen-sulfur (NOS) linkages in proteins function as dynamic, allosteric redox switches.\n* These bonds form or break in response to cellular oxidative stress, allowing proteins to modulate their activity and respond to their chemical environment.\n* This discovery provides new insights into redox biology and opens avenues for designing smart therapeutics and biosensors that are controlled by a cell\u2019s redox state.\n\n1. Revealing arginine-cysteine and glycine-cysteine NOS linkages by systematic re-evaluation of the PDB: <https://www.nature.com/articles/s42004-025-01535-w>\n2. NOS and SONOS redox switches in proteins \u2013 Abstract \u2013 Europe PMC: <https://europepmc.org/article/PPR/PPR365120>\n3. In Solution Identification of the Lysine\u2013Cysteine Redox Switch: <https://pmc.ncbi.nlm.nih.gov/articles/PMC11056971/>\n\nOur cells exist in a constant state of chemical balance. A key aspect of this is the \u201credox state\u201d\u2014the balance between oxidizing and reducing molecules. When this balance is disrupted by an excess of reactive oxygen species, it leads to oxidative stress, a condition implicated in everything from aging to cancer. For years, scientists have studied how proteins are modified in response to this stress, with disulfide bonds being the classic example of a redox-sensitive switch.\n\nThe recent discovery of nitrogen-oxygen-sulfur (NOS) linkages in proteins adds a major new player to the field of redox biology. These unusual covalent bonds, found between residues like arginine-cysteine and glycine-cysteine, are not merely structural supports. Evidence suggests they are dynamic, reversible cross-links that act as sophisticated allosteric redox switches. This means they form or break in response to changes in the cellular redox environment, and in doing so, they can modulate a protein\u2019s activity from a distance.\n\nBy forming an NOS bridge, a protein can subtly alter its 3D shape, which in turn can change the conformation of its active site or affect its ability to bind to other proteins. This provides a mechanism for fine-tuning cellular processes in response to oxidative stress. Understanding how these switches work has profound implications. In disease, these switches might be stuck in the \u201con\u201d or \u201coff\u201d position. In protein engineering, we could harness this mechanism to design novel biosensors that report on a cell\u2019s redox state, or create therapeutics that are only activated under the specific oxidative conditions found in a tumor.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics; supervisor Prof. E. Yvonne Jones), 1999\n\nMercury positions were determined by manual inspection of anomalous difference Patterson maps, calculated using the program FFT (Ten Eyck, 1973). Confirmation of the mercury positions was carried out using the programs RSPS and SOLVE (Terwilliger and Berendzen, 1999). The positions of the mercury sites were refined using the program SHARP (de La Fortelle and Bricogne, 1997).\n\n#### 5.2.7 Molecular Replacement\n\nThe refined MAD determined structure was used as a search model for molecular replacement of subsequently collected native data using the program AMoRe (Navaza, 1994). Residues 3\u2013134 and 156\u2013187 of the model were used in the search with water molecules removed. Reflections were first sorted using the SORTING routine and the molecule translated so that the centre of gravity is at the origin, using the TABLING routine. The search model was placed in the symmetry free spacegroup P1 to minimize cross vectors and the cross-rotation function was carried out in the resolution range 12\u20133.5 \u00c5, using the routine ROTING. Having determined the orientation of the molecules in the asymmetric unit the translation parameter was calculated to orientate the molecules relative to the symmetry axis. This was carried out using the routine TRAING in the resolution range 12\u20133.5 \u00c5. Rigid body refinement of the solution was carried out using the routine FITTING and the resulting orientation matrix used to transform the output of the TABLING routine using the program PDBSET.\n\n#### 5.2.8 Map Calculation and Model Refinement\n\nMAD phases were calculated to 2.6 \u00c5 resolution using the program SHARP (de La Fortelle and Bricogne, 1997). Phase improvement was carried out by solvent flattening, using cyclical phase extension for 160 cycles starting from 4.0 \u00c5 to a final resolution of 2.6 \u00c5 using the program SOLOMON (Abrahams and Leslie, 1996). Structure factors and phases were calculated using the program SFALL (Agarwal, 1978) and phase combination carried out using SIGMAA (Read, 1986). Initial 2|Fo|\u2212|Fc| maps were calculated using data in the range 20.0\u20132.6 \u00c5 using the program FFT (Ten Eyck, 1973). All subsequent refinement and map calculations were carried out using the program CNS-SOLVE (Brunger et al., 1998).\n\nAn homology model of \u0394188_OSM (see Chapter 2) containing residues 1\u2013188 of mature OSM was used as an initial model. Rigid body refinement of the model was carried out using native data between 20.0\u20132.6 \u00c5. Helices were treated as separate rigid bodies and 40 cycles of maximum likelihood rigid body refinement were carried out. Low weighting of the X-ray terms (Wa = 2.4) was initially used, to maintain close to ideal stereochemistry, and 10 % of the data was flagged for use as a free-R (Brunger, 1992) test set. Positional refinement using maximum likelihood was carried out for 200 cycles using data in the range 20.0\u20132.6 \u00c5 followed by simulated annealing for 1000 cycles, using torsion dynamics with slow cooling from 4000 K.\n\nManual rebuilding was carried out using the interactive graphics program O (Jones et al., 1991) with reference to the original solvent flattened 2|Fo|\u2212|Fc| electron density maps. Refined coordinates were output and 2|Fo|\u2212|Fc|\u03b1calc maps and |Fo|\u2212|Fc| difference maps calculated using the program CNS-SOLVE (Brunger et al., 1998) using data in the range 20.0\u20132.6 \u00c5. Further rebuilding was carried out with reference to the 2|Fo|\u2212|Fc|\u03b1calc maps and |Fo|\u2212|Fc| difference maps. For regions of ambiguous density composite omit maps were calculated using 5 % omissions of data. After five rounds of such rebuilding and refinement, bulk solvent correction was applied with a solvent density of 0.29 e/\u00c53 and a B-factor of 31.1 \u00c52. Anisotropic B-factor correction was also applied on data in the range 6.0\u20132.2 \u00c5. In subsequent rounds of refinement all data in the resolution range 20.0\u20132.2 \u00c5 were included. A further ten rounds of refinement and manual rebuilding was carried out with the inclusion of grouped B-factor refinement and subsequent individual B-factor refinement.\n\nIn the final stages of refinement bound water molecules were built into the model using an automated peak search utility in the program CNS-SOLVE (Brunger et al., 1998). Waters were built into peaks above 1.5 \u03c3 at a distance of 2.6\u20133.2 \u00c5 from the nearest oxygen or nitrogen atoms. A second round of water addition and refinement was carried out at a peak search level of 1.0 \u03c3. Further waters were added manually through inspection of 2|Fo|\u2212|Fc|\u03b1calc maps and |Fo|\u2212|Fc| difference maps. In a final round of refinement the X-ray weighting term was increased (Wa = 10) so relaxing the restraint on stereochemical terms.\n\n#### 5.2.9 Analysis of Model\n\nThe stereochemistry of the final model was assessed using the program PROCHECK (Laskowski et al., 1993). Secondary structure assignments and solvent accessibilities were calculated using the programs DSSP (Kabsch and Sander, 1983), HERA (Hutchinson and Thornton, 1990) and NACCESS (Hubbard and Thornton, 1993). Structural superpositions were performed using the program SHP (Stuart et al., 1979). Molecules were drawn and rendered using the programs BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994).\n\n### 5.3 Results of Structure Determination\n\n#### 5.3.1 Crystallisation\n\nCrystal Screen I and Crystal Screen II (Hampton Research) produced no crystals in all samples tested at 22 \u00b0C or 4 \u00b0C. \u0394188_OSM samples displayed high solubility in all screens with Se-Met-OSM and OSM-169 displaying heavy precipitation in ~70 % of all conditions.", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\n**Publication number:** US 2025/0215006 A1. **Publication date:** July 3, 2025. **Applicant / Assignee:** Incyte Corporation, Wilmington, DE (US). **Inventors:** Yanran Ai, Joe Barbosa, Ananda Ghosh, Marc D. Deller, Joe Glenn, Jr., Leslie McQuirter, Yehao Qiu, Lixin Shao, Stacey Shepard, Cooper Taylor, Xiaozhao Wang, Eddy Yue and Guofeng Zhang, all of Wilmington, DE (US). **Application No.:** 18/914,364. **Filed:** October 14, 2024. **Related U.S. application data:** Provisional application No. 63/590,628, filed on October 16, 2023. **International classification:** C07D 487/04; A61K 31/437; A61K 31/519; C07D 471/04; C07D 471/14. **U.S. classification (CPC):** C07D 487/04 (2013.01); A61K 31/437 (2013.01); A61K 31/519 (2013.01); C07D 471/04 (2013.01); C07D 471/14 (2013.01).\n\n## Abstract\n\nDisclosed are compounds of Formula (I), methods of using the compounds as immunomodulators, and pharmaceutical compositions comprising such compounds. The compounds are useful in treating, preventing, or ameliorating diseases or disorders such as cancer or infections.\n\n## Field of the Disclosure\n\nThe present application is concerned with pharmaceutically active compounds. The disclosure provides compounds as well as their compositions and methods of use. The compounds are inhibitors of WRN and are useful in the treatment of various diseases including cancer.\n\n## Background\n\nWerner helicase (WRN) belongs to the RecQ helicase family, along with BLM, RecQ1, RecQ4 and RecQ5. These enzymes play broad roles in genome maintenance, preserving chromosome stability, and suppressing neoplastic transformation (Chu, W. et al. RecQ helicases: multifunctional genome caretakers. Nat Rev Cancer 9, 644\u2013654 (2009)).\n\nGenome-wide screens identified high dependence of WRN by microsatellite instability-high (MSI-H) cancer cells, i.e., deletion of WRN has anti-proliferation effect on MSI cells, but not microsatellite stable (MSS) cells (Chu (supra); Behan, F. M. et al. Prioritization of cancer therapeutic targets using CRISPR-Cas9 screens. Nature 568, 511\u2013516 (2019); Chan, E. M. et al. WRN helicase is a synthetic lethal target in microsatellite unstable cancers. Nature 568, 551\u2013556 (2019); Kategaya, L. et al. Werner syndrome helicase is required for the survival of cancer cells with microsatellite instability. iScience 13, 488\u2013497 (2019)).\n\nWRN is a multifunctional enzyme having exonuclease activity, ATPase activity, and helicase activities. Site-directed mutagenesis and functional rescue studies have determined that helicase activity of WRN is critical for this dependence (Behan (supra); Chan (supra)). Therefore, molecules that target WRN helicase activity could have therapeutic uses in MSI-H/dMMR cancers. The WRN helicase activity is driven by ATP hydrolysis, which is catalyzed by its ATPase activity. So small molecules that inhibit WRN ATPase activity, WRN ATP hydrolysis activity, or both will ultimately inhibit its WRN helicase activity.\n\nMSI-H or mismatch repair deficient (dMMR) tumors have an accumulation of genetic errors in sequences that normally contain short repeats (called microsatellites). Nearly 4% of all cancers have MSI-H/dMMR, with the highest prevalence detected in endometrial (31%), colon (20%), and gastric cancers (19%) (Bonneville R. et al, Landscape of Microsatellite Instability Across 39 Cancer Types JCO Precision Oncology 1, 1\u201315 (2017)).\n\nRecently, checkpoint inhibitors (CPI) have emerged as effective treatment for MSI-H/dMMR cancers in both 1st line and 2nd line settings, but with limited response rates (for example, Pembrolizumab approved in MSI-H/dMMR tumor patients with 33.3% ORR; Green A. K. et al. A Review of Immune Checkpoint Blockade Therapy in Endometrial Cancer. Am Soc Clin Oncol Educ Book. 40, 1\u20137 (2020)), and resistance has developed in some patients.\n\nWRN inhibitors have a different mode of action compared to checkpoint inhibitors, and will have the potential to complement CPI in the treatment of MSI-H/dMMR cancers. In addition, WRN sensitivity is maintained in preclinical MSI-H colorectal cancer (CRC) models with resistance to standard of care (SOC), including targeted therapies, chemotherapy, and immunotherapy (Picco, C. et al. Werner Helicase Is a Synthetic-Lethal Vulnerability in Mismatch Repair-Deficient Colorectal Cancer Refractory to Targeted Therapies, Chemotherapy, and Immunotherapy. Cancer Discov 11, 1923\u20131937 (2021)). This has the potential to broaden WRN inhibitors' clinical uses.\n\nAccordingly, there is a need for new compounds that inhibit WRN and which are useful in treating cancers such as MSI-H/dMMR cancers. This application addresses that need and others.\n\n## Summary\n\nThe present disclosure provides, inter alia, compounds of Formula (I):\n\n*Formula (I).* A tricyclic core in which a ring moiety B is fused through X, Y and Z to a bicyclic pyrimidinone system; ring moiety A bearing an electrophilic warhead group R^W is attached at the carbon alpha to the carbonyl; substituent R^3 sits on the ring carbon adjacent to Z; and an N-linked acetamide arm bearing R^4 and R^5 at the alpha carbon and an R^6 group on the amide nitrogen completes the molecule. Ring B carries m independently selected R^1 substituents and ring A carries n independently selected R^2 substituents.\n\nor a pharmaceutically acceptable salt thereof, wherein constituent variables are defined herein.\n\nThe present disclosure further provides a pharmaceutical composition comprising a compound disclosed herein, or a pharmaceutically acceptable salt thereof, and one or more pharmaceutically acceptable excipient or carrier.", "# Engineering Evolution: How FAST-PETase and Other Variants Are Transforming Plastic Biodegradation\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nBeyond *Is*PETase derivatives, thermophilic cutinases offer alternative scaffolds. [Leaf-branch compost cutinase](https://www.mdpi.com/1420-3049/29/6/1338/pdf?version=1710754801) (LCC) variants, particularly LCC-ICCG (F243I/D238C/S283C/Y127G), operate optimally at 65-72\u00b0C. The F243I mutation altered the aromatic binding cleft to enhance PET substrate interactions, while the D238C/S283C pair introduced a stabilizing disulfide bridge. Recent [LCC engineering studies](https://febs.onlinelibrary.wiley.com/doi/10.1111/febs.16736) identified the S101N/F243T double mutant capable of completely depolymerizing 1.3 g untreated postconsumer PET waste within three days at 55\u00b0C using only 1.25 mg enzyme.\u200b\n\n[TurboPETase](https://www.nature.com/articles/s41467-024-45662-9), derived from BhrPETase, achieved 98.2% depolymerization of postconsumer PET at 200 g/kg loading within 8 hours. This variant combined eight mutations including W104Leu and F243Thr that reduced steric hindrance and enhanced active site flexibility. The Trp104Leu mutation disrupted interactions with Pro248, facilitating product release and increasing turnover by 71% relative to the six-mutant BhrPETase parent.\u200b\n\n[Cut190 variants](https://pmc.ncbi.nlm.nih.gov/articles/PMC9606173/) represent calcium-dependent PET hydrolases with unique regulatory mechanisms. The Cut190** mutant (S226P/R228S with C-terminal deletion) exhibited enhanced thermal stability and activity through a calcium-mediated conformational switching mechanism. Crystal structures revealed an \u201cejecting form\u201d promoting product dissociation, while the F77L mutation in Cut190**SS increased catalytic activity by modulating Site 1 conformational flexibility.\u200b\n\n**Origin and Function Across Variants**\n\n*Is*PETase originates from the mesophilic bacterium *Ideonella sakaiensis* isolated from a bottle recycling facility, exhibiting optimal activity at 30-35\u00b0C. In contrast, thermophilic enzymes such as [TfCut2 from *Thermobifida fusca*](https://pmc.ncbi.nlm.nih.gov/articles/PMC10577388/) function at elevated temperatures (55-70\u00b0C) but possess narrower substrate binding sites due to structural differences in the \u03b28-\u03b16 connecting loop. *Is*PETase contains three additional residues (Asn244-Ser245-Asn246) in this loop, creating extended binding subsites (IIb and IIc) that accommodate longer PET oligomers.\u200b\n\nThe functional roles of specific mutations illustrate engineering trade-offs. The [W159H mutation in DuraPETase](https://pmc.ncbi.nlm.nih.gov/articles/PMC10546322/) reconstructed the active site via \u03c0-\u03c0 stacking with Tyr87, forming an \u201caromatic tunnel\u201d that enhanced PET binding through aromatic interactions. Conversely, the [S214H mutation](https://pmc.ncbi.nlm.nih.gov/articles/PMC12387008/) stabilized the wobbling Trp185 residue, balancing activity and stability. When reverted to S214S in DuraPETase-4M, activity increased while thermostability decreased, exemplifying the activity-stability trade-off.\u200b\n\n[Z1-PETase](https://www.biorxiv.org/content/10.1101/2022.11.14.516528v1.full.pdf), engineered through balance-directed design, contains 13 mutations including four disulfide bonds (N233C-S282C and A171C-S193C). The P181V mutation increased soluble protein yield 40-fold while maintaining thermostability, and surface charge modifications (N37D, R132E, R224E) facilitated enzyme adsorption to PET substrates. Z1-PETase degraded approximately 75% of amorphous PET powder in 96 hours at 30\u00b0C, outperforming FAST-PETase, DuraPETase, and LCC-ICCG across temperature ranges.\u200b\n\n**Advantages and Limitations**\n\nEach variant offers distinct benefits. FAST-PETase excels at moderate temperatures (50\u00b0C) with minimal pretreatment requirements, degrading highly crystalline films (23.6% crystallinity) while releasing 23.8 mM monomers. HotPETase maintains activity at 60-65\u00b0C, exceeding [LCC-ICCG performance](https://www.nature.com/articles/s41467-024-45662-9) at equivalent temperatures. DuraPETase demonstrates superior activity on semicrystalline substrates at 37\u00b0C, completely degrading nanoplastics within one hour.\u200b\n\nHowever, limitations persist. Mesophilic variants like FAST-PETase lose efficiency above 50\u00b0C as PET undergoes physical aging, increasing crystallinity and reducing enzymatic accessibility. Thermophilic enzymes require energy-intensive heating, potentially offsetting environmental benefits. [Product inhibition](https://pmc.ncbi.nlm.nih.gov/articles/PMC10963241/) by MHET and TPA remains problematic across all variants, necessitating strategies such as co-expression with MHETase or continuous product removal. Recent [engineering efforts](https://pmc.ncbi.nlm.nih.gov/articles/PMC11866010/) addressing these challenges include the *Is*PETase MT variant with 15.3\u00b0C increased melting temperature and sustained activity to 60\u00b0C.\u200b\n\nThe [comparative analysis of LCC variants](https://pubs.rsc.org/en/content/articlelanding/2023/re/d3re00056g) highlights substrate-dependent performance variations. While ICCG-DAQI (combining ICCG and DuraPETase mutations) showed 27% higher product release than ICCG on bottle PET, post-industrial textile fibers were degraded more efficiently, achieving 40 mg TPA per hour per mg enzyme. These findings emphasize the importance of matching enzyme characteristics to specific PET waste streams for optimal recycling outcomes.\u200b", "# Approaches to automated protein crystal harvesting\nAuthors: Deller, M. C.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 70, 133\u2013155 (IYCr crystallization series), 2014\n\n| Specific benefit | Manual harvesting | Automated harvesting | Benefit of automation |\n|---|---|---|---|\n| High throughput | No | Yes | Allows hundreds of crystals to be harvested; saves several hundred person hours per year |\n| Fully automated pipeline | No | Yes | Manual procedures are a process bottleneck; closes the gap for true gene-to-structure automation |\n| Smaller crystals | No | Yes | Allows harvesting of crystals below 10 \u00b5m; enables high-throughput use of microfocus beams |\n| Adaptable to new crystallization and screening methods | No | Yes | Flexible automation adapts to new plate and screen designs such as microfluidics |\n| Precision liquid handling | No | Yes | Automates soaking of ligands and small-molecule inhibitors for co-crystallization trials |\n| High-speed harvesting | No | Yes | Reduces time between harvesting and flash-cooling; reduces osmotic shock and drying of mother liquor |\n| Reproducibility | No | Yes | Constant robot speeds and controlled motion paths mean fewer variables for crystal loss and damage |\n| Improved flash-cooling | No | Yes | Enables hyperquenching, reduced cryoprotectant, automated cryoprotection screening and cryo-drip |\n\n*Table 2.* Advantages and disadvantages of levels of harvesting automation (more + symbols indicate a higher score or cost).\n\n| Level | High throughput | Labour saving | Reproducibility | Crystal detection | Cost | Ease of engineering | Damage to crystal | Limited to synchrotron | Data-collection ability | Advantages | Disadvantages |\n|---|---|---|---|---|---|---|---|---|---|---|---|\n| Fully automated | +++ | +++ | +++ | +++ | +++ | + | +++ | + | + | High throughput, labour saving, reproducible | Relies on crystal image detection; expensive hardware and software; difficult engineering |\n| Semi-automated | + | + | ++ | + | + | + | ++ | + | + | Reproducible, relatively inexpensive | Medium throughput; labour intensive |\n| Plate-based in situ | +++ | +++ | +++ | +++ | + | +++ | + | +++ | +++ | No harvesting required; limited crystal damage; easy engineering; very high throughput; labour saving; reproducible | Excess X-ray scattering and absorption from plate; reduced rotation angle; room-temperature only; relies on image detection or X-ray rastering; expensive; synchrotron only |\n\n*Table 3.* Semi-automated protein crystal-harvesting systems reviewed.\n\n| System | Developer | Harvesting mechanism | Key features |\n|---|---|---|---|\n| Protein Crystal Growth Demonstration System (PCGDS) | University of Washington and Boeing Missiles and Space Division, USA | Microcapillary | Remote telescience/telerobotic harvesting onboard the ISS |\n| Crystal Preparation Prime Item (CPPI) | Oceaneering Space Systems, Houston, and University of Alabama at Birmingham, USA | Microcapillary, microtool | ISS X-ray Crystallography Facility; liquid\u2013liquid bridge isolates crystals for harvesting |\n| Two-fingered micro-hand (\"chopsticks\") | KEK and AIST, Japan | Microgripper, microtool | Microgrippers operated chopstick-fashion; crystal placed into loop or microtool |\n| Automated Crystal Harvester (ACH) | The Ohio State University, USA and University of Limerick, Ireland | Microcapillary, microtool | Capillary-loop coordinating mechanism; harvesting of membrane proteins from LCP |\n| Columbia Automated Robotic Environment for Streak Seeding (CARESS) | NESG, Columbia University, USA | Microcapillary, microtool | Silicon microshovel; automated streak-seeding; micropipette-based pickup |\n| Universal Micromanipulation Robot (UMR) | Square One System Design, Wyoming, and JCSG, The Scripps Research Institute, USA | Microtool | Semi-autonomous microtool harvesting; flexible plate and cassette/puck configurations; automated data tracking, plate opening, cryo-drip cryoprotection and flash-cooling; 1 \u00b5m resolution robotics |\n| Capillary-top Mounting Micromanipulator (CMM) | Nagoya University and Hokkaido University, Japan | Microcapillary, microtool | Loopless mounting; reduced X-ray absorption at longer wavelengths; improved S-SAD phasing |\n| Acoustic Droplet Ejection (ADE) | National Synchrotron Light Source beamline X25 and Labcyte Inc., USA | ADE | Acoustic mounting of microcrystal slurries; enabling technology for FEL beamlines; conveyor-belt loading |\n| Robotic Equipment for Automated Crystal Harvesting (REACH) | ESRF beamline FIP-BM30A and NatX-ray, France | Microgripper | Semi-autonomous microgripper harvesting; automated data tracking, plate opening, cryoprotection and flash-cooling; direct data collection from plates and from samples in the grippers |\n| Conventional optical tweezers (COT) | Diamond Light Source, England | COT, micromesh | Lens-based optical trap mounting of microcrystals around 10 \u00b5m |\n| Fibre optical tweezers (FOT) | RIKEN SPring-8 Center, Japan | FOT, microtool | Fibre-based optical trap for 5\u201330 \u00b5m crystals; touchless mounting |\n| RodBot | Multi-Scale Robotics Laboratory, ETH Zurich, Switzerland | MEMS, microtool | Mobile magnetic MEMS device; fluid-flow manipulation and mounting of crystals of any size |\n| Crystal Harvester (CH) | Bruker, Billerica, Massachusetts, USA | Microtool | Commercial micromanipulator-based product |\n| Station for Automated Microscopy and Imaging (SAMI) | FMP Products Inc., Greenwich, Connecticut, USA | Microtool | Commercial twin-loop micromanipulator, primarily for small-molecule crystallography |\n\n*Table 4.* Harvesting mechanisms and their relative advantages and disadvantages (more + symbols indicate a higher score or cost).", "# Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands\nAuthors: Pozharski, E.; Deller, M. C.; Rupp, B.\nSource: Protein Crystallography: Methods and Protocols (A. Wlodawer, Z. Dauter & M. Jaskolski, eds.), Methods in Molecular Biology vol. 1607, Chapter 25, pp. 611\u2013625, Springer, 2017\n\nInspection begins with the 2mFo \u2212 DFc map: does the blue grid surround the model? If not, further validation is likely required. Then the omit mFo \u2212 DFc difference density is examined, with the original PDB model imported for comparison, commonly viewed at a 3\u03c3 level; the authors suggest a useful calibration step of computing an omit map after deleting a single protein residue in the binding site, which gives an estimate of the magnitude of difference density expected from a bona fide structural element. Next, evaluate how much of the ligand can be clearly placed within the density \u2014 partial fits are fairly common for peptide ligands, lipid molecules and glycosylation sites, where the most plausible explanations are disorder arising from weak interactions with the protein, or poor occupancy of the binding site. Then evaluate whether the ligand could be fitted in a different orientation or conformation, and whether the density admits alternative explanations altogether: strings of structured water molecules in the binding site, components of the mother liquor or cryoprotectant such as short polyethylene glycol fragments, or terminal fragments and loop regions of a symmetry-related protein molecule. Finally, examine the residue density fit graph in COOT, paying particular attention to the ligand and to residues within the binding site.\n\n### 4.4. Environmental examination\n\nThe environmental checks ask whether symmetry-related molecules are involved in ligand binding; if so, further evaluation using biochemical or biophysical techniques may be required to determine whether those interactions are real or crystal-packing artefacts. If multiple copies of the same molecule are present in the asymmetric unit under non-crystallographic symmetry, all of the binding sites in the asymmetric unit should be examined and validated, and the researcher should ask whether different conformations or occupancies can be explained by non-equivalent environments, plasticity or crystal packing. Bond distances within the binding-site environment should be displayed in COOT and assessed for whether the non-covalent interactions are plausible. Lastly, the clash score should be computed through the COOT validation menu, asking whether there are significant clashes between ligand and protein and whether they can be explained or corrected. Together with the density inspection of the preceding subsection, this completes the three-branch decision tree.\n\n### 5. Tools, and 6. Outlook and new techniques\n\nThe tools available include standalone programs, web servers and databases of ligands scored by various quality criteria, catalogued in the chapter's software table. The authors reiterate that quality and suitability are context-sensitive \u2014 not every protein\u2013ligand structure is fit for structure-based lead discovery, and not every missing piece of a ligand is cause for panic \u2014 closing with the maxim from their earlier work: trust, but verify. The outlook section identifies a new problem. Ultrafast electron diffraction, double electron\u2013electron resonance, atomic force microscopy, cryo-electron microscopy and X-ray free-electron lasers generally produce structures of lower resolution, typically in the 4\u201310 \u00c5 range, which places extra emphasis on the ligand restraints applied. New hybrid techniques allow the dynamics of ligand binding to be probed on new timescales, and XFEL enables analysis of pico- and femtosecond dynamics; much of the chemistry, and ligand stereochemistry in particular, is yet to be explored on such timescales, so new definitions of ligand stereochemistry will be needed to keep restraints accurate. The chapter ends with a call to action: structural biology lies at the forefront of biomedical research and drug discovery, the community must get serious about effective ligand validation, effort is required at all levels to keep deposited protein\u2013ligand structures at the highest quality, validation methods must continue to evolve and embrace hybrid techniques, and mandatory validation efforts and task forces such as those initiated by the PDB will be essential.\n\n## Key tables\n\n*Table 1.* Protein\u2013ligand validation software listed in the chapter.\n\n| Package | Description |\n|---|---|\n| Twilight | Small molecule and peptide ligand validation |\n| COOT | Molecular graphics package for model building and validation |\n| MolProbity | All-atom clash scores and other validation statistics |\n| VHELIBS | Fit of ligands and binding sites |\n| ValLigURL server | Compare conformations of ligands in the PDB |\n| MotiveValidator and ValidatorDB | Interactive web-based validation of ligands and residues |\n| PDB_REDO | Updated and optimized X-ray structure models and maps |\n| wwPDB | wwPDB Validation Server |\n| PRIVATEER | Checks glycan nomenclature and stereochemistry |\n| LIGPLOT | Ligand\u2013protein interaction diagrams |\n| EDS | Electron density server |\n| EDSTATS | Statistical quality indicators of electron density maps |\n| OVERLAPMAP | Average of two maps |\n| BUSTER-TNT | Refinement of proteins and ligands |\n| WHAT_IF / WHAT_CHECK | Protein and ligand verification tools |\n\n*Table 2.* Real-space quality metrics and their interpretation, as described in the chapter.", "# Supercharging Discovery: How LLM Prompts Are Transforming Structural Biology and Cheminformatics\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-17 on marcdeller.com (categories: sb).*\n\n* Motif/fingerprint discovery from single protein sequences.\n* Prediction of protein\u2013protein interactions from sequence pairs.\n* Mapping from sequence to secondary structure and risk features.\n* Analog design suggestions for improved drug properties.\n* Pathway inference using protein functional annotations.\n* Rapid metabolic/toxicity profiling and mitigation strategies.\n* Patent mining and innovation mapping for biomolecular targets.\n* Synthetic route planning for cost-effective drug assembly.\n* Enzyme substrate range predictions from structural input.Structured extraction of key data (e.g., IC50) from literature.\n\n1. <https://arxiv.org/html/2408.11363v1>\n2. <https://github.com/Yijia-Xiao/Protein-LLM-Survey>\n3. <https://www.nature.com/articles/s41598-025-99290-4>\n4. [https://arxiv.org/html/2408.11363v2](https://pmc.ncbi.nlm.nih.gov/articles/PMC12076503/)\n5. <https://academic.oup.com/bioinformatics/advance-article/doi/10.1093/bioinformatics/btaf396/8215464>\n6. <https://arxiv.org/html/2405.06690v1>\n7. <https://arxiv.org/html/2410.05610v2>\n8. <https://openreview.net/forum?id=2nTzomzjjb>\n9. <https://arxiv.org/html/2409.04481>\n10. <https://pmc.ncbi.nlm.nih.gov/articles/PMC12076503/>\n\nThe intersection of artificial intelligence and molecular science is fueling a revolution in protein structure analysis and drug discovery. Large language models (LLMs), like GPT-4 and Claude, are increasingly being used beyond text analysis \u2014 their ability to follow complex, multi-stage instructions allows them to accelerate real-world biopharma applications. Let\u2019s explore **10 creative prompt engineering workflows** that illustrate the scope and power of LLMs in structural biology and cheminformatics.\n\n## End-to-End Example Prompts for LLMs in Molecular Discovery\n\nBelow are ten sophisticated prompt workflows, each designed to work with standard LLMs accessible through Perplexity, OpenAI, and others. Each example provides a motivating use-case, example input, and expected output.\n\n---\n\n1. **Motif Discovery from Protein Sequence**\n\n   * *Prompt*: \u201cGiven this amino acid sequence, list known biological motifs, infer likely function, and suggest a related signaling pathway.\u201d\n   * *Example Input*: `MKWVTFISLLFLFSSAYSRGVFRRDTHKSEIAHRFKDLGE`\n   * *Sample Output*: \u201cContains an SH3-binding motif; likely involved in eukaryotic signaling via proline-rich region interactions.\u201d\n2. **Protein\u2013Protein Interaction Prediction**\n\n   * *Prompt*: \u201cGiven two sequences, analyze complementarity and predict the likelihood and context of protein\u2013protein interaction.\u201d\n   * *Example Input*: Two FASTA-formatted sequences.\n   * *Sample Output*: \u201cComplementary binding motifs predicted; probable interaction in cytoskeletal scaffolding.\u201d\n3. **Secondary Structure Risk Profiling**\n\n   * *Prompt*: \u201cMap this protein sequence to its secondary structure, correlating specific residues to structural or stability risks.\u201d\n   * *Example Input*: Amino acid sequence.\n   * *Sample Output*: \u201cResidue 24 (Pro) likely introduces helix bending; Cys residues suggest potential for disulfide-mediated stability.\u201d\n4. **Analog Suggestion for Drug-like Properties**\n\n   * *Prompt*: \u201cGiven this scaffold molecule, suggest five analogs with enhanced drug-likeness and briefly justify each proposal.\u201d\n   * *Example Input*: SMILES or InChI for a core structure.\n   * *Sample Output*: \u201cAdding a methyl group increases solubility; halogen substitutions can improve binding affinity.\u201d\n5. **Pathway Mapping from Experimental Protein Data**\n\n   * *Prompt*: \u201cBased on these experimental protein annotations, trace the most probable signaling pathway and cite supporting literature.\u201d\n   * *Example Input*: Functional and localization annotations.\n   * *Sample Output*: \u201cLikely involved in the MAPK/ERK pathway; see PMID 12345678 for pathway evidence.\u201d\n6. **Rapid Toxicity Profiling for Small Molecules**\n\n   * *Prompt*: \u201cGiven the following molecule, enumerate likely metabolic liabilities and suggest chemical modifications to reduce toxicity.\u201d\n   * *Example Input*: Molecular structure or SMILES.\n   * *Sample Output*: \u201cPredicted aldehyde metabolite may be toxic; propose amide substitution.\u201d\n7. **Patents and Prior Art Search**\n\n   * *Prompt*: \u201cList ten recent patents related to this protein target, summarizing the main innovation of each.\u201d\n   * *Example Input*: Target protein or gene.\n   * *Sample Output*: A table of patents with titles, numbers, and summaries.\n8. **Affordability and Synthesis Route Analysis**\n\n   * *Prompt*: \u201cSuggest cost-effective, feasible synthesis routes for this small molecule, considering common lab reagents.\u201d\n   * *Example Input*: Compound name or SMILES.\n   * *Sample Output*: \u201cThree-step synthesis via Grignard addition followed by amide coupling.\u201d\n9. **Enzyme Substrate Promiscuity Insight**\n\n   * *Prompt*: \u201cGiven enzyme X, predict three likely substrate classes and rationalize the prediction based on active site features.\u201d\n   * *Example Input*: Enzyme name and structure data.\n   * *Sample Output*: \u201cDisplays promiscuity toward primary amines due to flexible binding cleft.\u201d\n10. **Data Extraction from Biomedical Literature**\n\n    * *Prompt*: \u201cRead the following abstract, extract all reported inhibitory concentrations (IC50) and summarize compound structures.\u201d\n    * *Example Input*: Biomedical abstract text.\n    * *Sample Output*: \u201cCompound A: IC50=42nM; pyridine-based core.\u201d", "# GitHub README: PANTS\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/PANTS, 2026\n\nSo the apparent geometric signature is **largely cluster structure**. Within the enzymes to\nhand, PET-active ones really do have deeper clefts; but which enzymes are PET-active is\nconfounded with which family they belong to, and once a model has to generalise to a\ncluster it has never seen, the separation goes. Only 8 clusters contain a negative at all\nand only 4 of those also contain a PET-active enzyme, which is far too thin a base to\nlearn a transferable rule from.\n\n**Both signals fail the same test for the same reason, and it is not the method.** This is\nthe clearest statement of the constraint the project is actually under: raw feature\ndifferences that look decisive at p<0.001 collapse under cluster-grouped splitting, which\nis precisely why that splitting is the rule here and why the composition baseline is\nreported permanently. The finding is not \"geometry does not work\" but \"geometry cannot be\nshown to work on 17 negatives spanning 4 shared clusters\".\n\nThree routes that would, in [`PHASE1_FINDINGS.md`](PHASE1_FINDINGS.md): ordinal\nwithin-paper rankings, regression on measured rates instead of a binary label, and\ngeometry as an annotation-independent axis.\n\n**A caution about the evidence tiers.** Of the 449 entries carrying EC 3.1.1.101 by automatic annotation, none is a measurement: they hold `ECO:0000256` (by similarity), not `ECO:0000269` (experimental). They were briefly labelled \"unreviewed\", which reads as a curation backlog rather than the substantive difference it is. Sixteen positives have experimental evidence. That is the number the Methods tab will report.\n\n## \ud83d\udd2d The profile library and the recall run\n\nRecall is a two-stage funnel: MMseqs2 casts the net across millions of sequences fast, HMMER makes the sensitive call on the survivors. Every candidate keeps its retrieval numbers (E-value, bitscore, profile identity), because those are the baseline the learned model has to beat.\n\nThe library is **one profile HMM per 30% sequence cluster**, each with its own catalytic anchor, rather than a single pooled profile. That matters: a single profile built over the polyesterases scored **0 of 111 near misses** as triad-complete, not because classic cutinases lack a catalytic triad but because they never aligned well enough for the columns to map. Per-cluster profiles took that to 79%, so the near misses survive recall and reach the scoring stage where they belong.\n\nAnchors come from UniProt's own `Active site` annotation rather than being hardcoded. Cross-checked before adoption: aligning to a pooled profile and reading IsPETase's verified S160/D206/H237 columns predicted LCC as S165/D210/H242 and TfCut2 as S170/D216/H248, and UniProt's independently curated annotations give exactly those numbers.\n\n| | |\n|---|---|\n| Library built from | 529 positives in 29 clusters at 30% identity |\n| Profiles | 3 (264, 73 and 3 sequences), anchored on LCC, `P9WP41` and `A6WFI5` |\n| Proteins scanned | 2,220,462 |\n| Candidates recovered | **128**, all triad-complete |\n| Runtime | 1,424 s (24 min) on an M1 Max |\n\n### Funnel\n\n| Stage | Surviving |\n|---|---|\n| Scanned | 14,778,289 |\n| Passed the prefilter, profile scan and triad filter | **439** |\n\nRetention is 0.003%. That is a **choice** (a strict E-value, then a hard requirement that\nSer, His and Asp be connected in space), not a property of the data.\n\n### By source environment\n\n| Environment | Proteins scanned | Candidates | Per 1M | Median %ID to nearest characterised |\n|---|---|---|---|---|\n| **Human gut** | **12,584,458** | **311** | 25 | 29.7% |\n| Compost | 1,020,575 | 69 | 68 | 34.1% |\n| Marine plastisphere | 737,027 | 44 | 60 | 26.5% |\n| Landfill | 436,229 | 15 | 34 | 32.8% |\n| Wastewater | 26,631 | 0 | 0 | |\n| **Total** | **14,778,289** | **439** | 30 | |\n\nThe gut is now 71% of the catalogue by count and 85% by sequences scanned. Its *yield* is\nthe lowest of the four productive environments, which is the honest reading: polyesterases\nare rarer in the gut than in compost, and the gut cohort is large because it was sampled\ndeeply, not because it is rich.\n\n**The identity bands are the interesting part**, because they separate rediscovery from genuinely unexplored sequence space:\n\n| Environment | \u226570% identity (rediscovery) | 40 to 70% | <40% (novel) |\n|---|---|---|---|\n| Compost | 15 | 14 | 40 |\n| Marine plastisphere | **0** | **0** | **44** |\n| Landfill | 1 | 2 | 12 |\n\nCompost gives the highest yield per million proteins and hands back 15 near-identical copies of enzymes that are already characterised, one of them a 100% identity match. That is unsurprising: LCC itself is leaf-branch compost derived, so compost is where the field has already looked.\n\n**Every single marine plastisphere candidate sits below 40% identity to anything characterised**, with a best bitscore of 64.9 against compost's 486. Nothing in that cohort is a rediscovery. This is spec section 2's thesis in one table: E-value rank pushes the well-known enzymes to the top and the unexplored ones down, and re-ranking that is exactly what the learned model is for.\n\nThe wastewater assembly returned nothing, which is a reasonable null: it is the only source in the set that is neither plastic-associated nor compost.\n\n### The blind test: did recall find HGMP01-like enzymes unaided?\n\nHGMP01 was **not** used to seed recall. Its sequence was unavailable when the gut scan was\ndesigned, and once obtained it was deliberately left out of the profiles, because the paper\nthat describes it reports HGMP01-like genes as widely distributed across the gut\nmicrobiome. Recovering them without the seed is a real test; seeding first would have made\nit circular.\n\nAgainst the five measured HGMPs:\n\n| | |\n|---|---|\n| Gut candidates | 311 |\n| With a detectable hit to any HGMP | **254 of 311 (82%)** |\n| Whose nearest HGMP is HGMP01 | **33** |\n| Best identity to HGMP01 | 38.3% |\n| Median identity to nearest HGMP | 27.5% |", "# Engineering Protein Stability: Unlocking the Secrets of Successful Crystallization\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-20 on marcdeller.com (categories: sb).*\n\n* Modern construct design leverages computational tools like AlphaFold3 and specialized servers to systematically remove flexible regions while preserving functional domains\n* Temperature optimization and codon usage correction dramatically improve expression levels and protein solubility in heterologous systems\n* Expression system selection has become more sophisticated, with cell-free systems offering unique advantages for difficult proteins\n* Surface entropy reduction (SER) provides a rational approach to engineering crystallizable proteins by replacing high-entropy surface residues\n* VHH nanobodies serve as powerful crystallization chaperones for highly dynamic proteins that resist conventional crystallization approaches\n* T4 lysozyme fusions have revolutionized membrane protein crystallography, enabling structure determination of previously intractable GPCRs\n* Reductive methylation offers a simple chemical modification strategy that improves crystal packing and resolution for many proteins\n\n1. Deller, M.C., Kong, L. & Rupp, B. (2016). [**Protein stability: a crystallographer\u2019s perspective.**](https://marcdeller.com/wp-content/uploads/2025/08/Protein-stability-a-crystallographers-perspective.pdf) *Acta Crystallographica Section F*, 72, 72-95:\n   <https://doi.org/10.1107/S2053230X15024619>\n2. Lynch, M.L., Mandal, S. & Ogorzalek, A. (2023). **Preparing for successful protein crystallization experiments.** *Acta Crystallographica Section F*, 79, 132-143:\n   <https://doi.org/10.1107/S2053230X23004144>\n3. Lopes, A. (2024). **Maximize Protein Expression: Overcoming Challenges and Boosting Yields.** *Lenio Bio:*<https://www.leniobio.com/maximize-protein-expression-overcoming-challenges-and-boosting-yields/>\n4. Derewenda, Z.S. & Vekilov, P.G. (2006). **Entropy and surface engineering in protein crystallization.** *Acta Crystallographica Section D*, 62, 116-124:\n   <https://journals.iucr.org/paper?ba5078>\n5. Mooij, W.T.M., Mitsiki, E. & Perrakis, A. (2009). **ProteinCCD: enabling the design of protein truncation constructs for expression and crystallization experiments.** *Nucleic Acids Research*, 37, W402-W405:\n   <https://academic.oup.com/nar/article/37/suppl_2/W402/1124629>\n6. Walter, T.S., Meier, C., Assenberg, R. et al. (2006). **Lysine methylation as a routine rescue strategy for protein crystallization.** *Structure*, 14, 1617-1622:\n   <https://pmc.ncbi.nlm.nih.gov/articles/PMC2678869/>\n7. Zou, Y., Weis, W.I. & Kobilka, B.K. (2012). N-**Terminal T4 Lysozyme Fusion Facilitates Crystallization of a G Protein Coupled Receptor.** *PLoS ONE*, 7(10), e46039:\n   <https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0046039>\n\nProtein crystallization remains one of the most challenging bottlenecks in structural biology, despite decades of technological advances in X-ray crystallography, NMR spectroscopy, and cryo-electron microscopy. While synchrotron radiation sources have become more powerful and detectors more sensitive, obtaining high-quality protein crystals suitable for structure determination continues to limit research progress. The key to success lies in understanding and engineering protein stability\u2014a complex interplay of compositional and conformational factors that determines whether a protein will form ordered crystals or remain stubbornly in solution.\n\n**Construct Design: The Foundation of Crystallization Success**\n\nThe journey toward successful protein crystallization begins with thoughtful construct design. Modern structural biologists have moved beyond simply expressing full-length proteins, recognizing that strategic truncations and modifications can dramatically improve crystallization prospects. The concept centers on removing flexible, disordered regions that contribute to conformational heterogeneity while preserving the stable, well-folded domains essential for biological function.\n\nAdvanced computational tools like AlphaFold3 now provide unprecedented insights into protein structure and flexibility, enabling researchers to identify problematic regions before beginning experimental work. Servers such as ProteinCCD facilitate the design of multiple truncation constructs by integrating predictions of secondary structure, disorder, transmembrane segments, and domain boundaries. This systematic approach allows investigators to generate libraries of constructs with varying start and end points, maximizing the chances of finding an optimal crystallizable variant.\n\nThe integration of affinity tags presents both opportunities and challenges in construct design. While tags like maltose-binding protein (MBP) can enhance solubility and act as crystallization chaperones, they must be carefully positioned to avoid introducing additional flexibility. Self-cleaving tags, including intein-based systems, offer elegant solutions by eliminating the need for protease treatment that could damage the target protein.\n\n**Expression Optimization: Beyond Simple Overproduction**\n\nAchieving high protein expression levels requires a multifaceted approach that considers host organisms, growth conditions, and molecular engineering strategies. Temperature optimization has emerged as a particularly powerful tool, with reduced expression temperatures (15-25\u00b0C) often dramatically improving protein solubility by slowing folding kinetics and reducing aggregation.\n\nCodon optimization represents another critical factor, especially when expressing eukaryotic proteins in bacterial systems. The availability of specialized E. coli strains like Rosetta, which supply rare tRNAs, has revolutionized the expression of mammalian proteins. Additionally, the strategic incorporation of cofactors and prosthetic groups during expression can ensure proper protein folding and stability.", "# AI-Driven Inverse Protein Folding: Transforming Protein Design for Next-Gen Therapeutics\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* Advanced AI models are solving the \u201cinverse folding problem,\u201d allowing scientists to design novel protein sequences for a desired 3D structure.\n* This technology accelerates the creation of new enzymes, antibodies, and other protein-based therapeutics with high precision.\n* By integrating structural and evolutionary data, AI reduces the cost and time of development, paving the way for scalable, custom-designed medicines.\n\n1. **AI could accelerate protein engineering \u2013 key for developing new medicines:**\n   <https://sheffield.ac.uk/news/ai-could-accelerate-protein-engineering-key-developing-new-medicines>\n2. **Top 5 Drug Discovery Trends 2025 Driving Breakthroughs:**\n   <https://www.pelagobio.com/cetsa-drug-discovery-resources/blog/drug-discovery-trends-2025/>\n3. **This technology accelerates the creation of new enzymes, antibodies, and other protein-based therapeutics with high precision:**\n   <https://www.drugtargetreview.com/news/164745/ai-transforms-protein-design/>\n\nDesigning proteins with desired functions has traditionally been a labor-intensive challenge, hampered by the complex relationship between sequence and structure. The process often involved speculative, iterative changes to known proteins, a method more akin to artistry than engineering. This limitation meant that creating entirely novel proteins for specific therapeutic tasks was slow and fraught with failure, leaving countless diseases without purpose-built molecular solutions.\n\nThe recent breakthroughs in AI, specifically with advanced inverse protein folding models, are revolutionizing this field. Inverse folding is the challenge of predicting an amino acid sequence that will adopt a specific 3D structure to achieve a desired biological function. New AI-driven methods, like those developed in a collaboration between the University of Sheffield and AstraZeneca, leverage deep learning to dramatically enhance the precision and speed of protein design. These models act as expert architects, predicting crucial folding patterns and identifying stable amino acid sequences, enabling the rapid generation of functional proteins with minimal trial and error.\n\nThis leap forward has profound implications for drug discovery. It enables true *de novo* protein engineering, facilitating the creation of custom therapies such as highly specific enzymes, next-generation antibodies, and novel molecular scaffolds tailored for diseases like cancer, autoimmune disorders, and rare genetic conditions. By integrating structural constraints and evolutionary data, this AI approach ensures the resulting proteins are not only structurally sound but also biologically relevant, overcoming previous barriers posed by unpredictable folding. As AI continues to advance, inverse protein folding is set to become a cornerstone of the pharmaceutical pipeline, unlocking therapeutic possibilities previously deemed unreachable.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nExpression of OSM_\u0394185 proved unsuccessful. No induction with IPTG was observed and no protein was observed following affinity purification with glutathione sepharose. It is possible that OSM_\u0394185 was not folded properly and as such would be present in the insoluble fraction. Possible reasons for mis-folding include incorrect prediction of the C-terminal end of helix D, as determined through homology modelling with LIF. This proposal is further supported by the low transformation frequencies observed for \u0394185_OSM (as reported in Chapter 2) which presumably result from the toxicity associated with this mutant protein.\n\nFigure 3-5 shows that expression of OSM_\u0394188-GST was successful. OSM_\u0394188 was separated to greater than 90 % purity, as assessed by SDS-PAGE, after one affinity step. Yields of ~1 mg/L were achieved. Western blot analysis as shown in Figure 3-6 revealed extensive fragmentation of OSM_\u0394188, so suggesting further purification. Purification using ion-exchange proved unsuccessful due to aggregation of OSM_\u0394188 in all pH's tested. Figure 3-9 shows that OSM_\u0394188 was resolved as a distinct peak using gel filtration. Elution profiles show two main peaks with the predominant peak corresponding to OSM_\u0394188 monomers (25 kDa) and a minor peak corresponding to OSM_\u0394188 dimers (50 kDa). Dimer formation was favoured in the presence of Triton X-100 and in the absence of detergent a minor contaminating band at 34 kDa was observed. The nature of this band is unknown.\n\nN-terminal sequencing was used to verify the correct nature of the purified OSM_\u0394188 fraction. This revealed the inclusion of the four residue sequence GPGS at the N-terminus of the mature OSM sequence. These residues result from 3C protease and BamHI cleavage of the engineered constructs. Mass spectrometry as shown in Figure 3-11 revealed one major species at a mass of 21579.98 \u00b1 9.05 Da. This value disagrees with the theoretical mass of OSM_\u0394188 of 21259 Da, the difference presumably due to bound ions, detergent, glycerol or other buffer components.\n\nReceptor binding assays as shown in Figure 3-7 suggest \u0394188_OSM possesses similar properties to the wildtype protein in vitro. Comparison of the EC50 values for binding to the gp130 receptor in titration and competition assays suggests OSM_\u0394188 has a similar affinity to OSM-196. Sigmoidal dose responses are observed in receptor competition assays with EC50 values of 209.7 ng/ml and 182 ng/ml for wildtype and mutant, respectively.\n\nBa/F3 cell survival assays demonstrate that OSM-196 and \u0394188-OSM also have similar properties in vivo. Ba/F3 cells are an IL-3 dependent murine cell-line that grows in suspension (Palacios and Steinmetz, 1985). This cell-line contains few endogenous cytokine receptors, so can be used as a model system through the introduction of exogenous receptors. Ba/F3 cells were previously stably transfected with human gp130 and human LIFR (Vicky Barton, CRC Growth Factors Group, Birmingham) so providing a cell-line responsive to both human LIF and human OSM. OSM_\u0394188 and OSM-196 show similar biological responses in Ba/F3 cell assays as shown in Figure 3-8. EC50 values for OSM_\u0394188 and OSM-196 show close agreement with values of 0.26 ng/ml and 0.39 ng/ml, respectively. These data suggest that \u0394188 C-terminal truncations do not affect the receptor binding and biological activity of OSM.\n\nFigure 3-10 shows that expression of OSM-Se-Met was only successful when carried out in the methionine auxotrophic strains of E. coli DL41 or B834. Mass spectrometry as shown in Figure 3-11 shows that 78Se-Met derivatisation of OSM_\u0394188 was not complete. Five main protein species are observed corresponding to no selenium substitutions, one selenium substitution, two selenium substitutions, three selenium substitutions and four selenium substitutions. Given the relatively good yield of protein obtained this observation is rather puzzling and suggests that some form of suppression of auxotrophism is occurring; possible sources of 32S-Met include the starter culture or contaminants in the LeMasters medium. The poor substitution observed could also be due to preferential use of 32S-Met, which given the buried nature of the methionine residues in the refined structure would be more favourable for correct protein expression and folding. N-terminal sequence analysis of OSM-Se-Met material, produced in strain B834, has the correct sequence with the inclusion of four residues from the 3C protease site.\n\n## 4. Chapter 4 \u2013 Expression, Purification and Characterisation of LIFR\n\nThe aims of this chapter include the production of leukemia inhibitory factor receptor (LIFR) of sufficient quality and quantity for crystallographic studies. This chapter also aims to characterise the LIFR-LIF receptor-ligand complex and its suitability for crystallographic studies. The use of several expression systems is reported, along with the production, purification, deglycosylation and preliminary characterisation of the LIFR-LIF receptor-ligand complex.\n\nAppendix III is included as a summary of the reagents and buffers used in the expression, purification and characterisation of LIFR. The methods involved are discussed in greater detail in texts such as (Maniatis et al., 1989), (Coligan et al., 1995) and (Deutscher, 1990). A detailed analysis of cell and tissue culture is given by (Paul, 1980) and an overview of the processes involved in the course of this chapter is shown in Figure 4-1 below.\n\n## 4.1 Outline of Methods Employed\n\n*Figure 4-1.* Flowchart detailing methods and techniques employed in the course of chapter 4.\n\n## 4.2 Methods of Expression, Purification and Characterisation of LIFR\n\nPrevious attempts to produce LIFR protein in E. coli expression systems proved unsuccessful due to the large size of constructs involved (typically >1000 amino acids) and the inability to correctly determine domain boundaries (D. Staunton, J. K. Heath, CRC Growth Factors Group, Birmingham, personal communication).", "# When the Enzyme Is Its Own Blueprint: Protein-Templated DNA Synthesis in DRT3\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-04-23 on marcdeller.com (categories: blog, sb).*\n\n* [**Science News**](https://www.science.org/content/article/scientists-stunned-fundamentally-new-way-life-produces-dna)\n\n* **Protein as blueprint:** Drt3b synthesizes sequence-specific poly(AC) DNA using active site residues as a chemical template, with no nucleic acid required.\n* **Cryo-EM was the decisive experiment:** The 2.6 \u212b structure of the 6:6:6 complex was the only way to demonstrate the absence of a nucleic acid template in the Drt3b active site.\n* **DRTs are a deep reservoir of novel biochemistry:** DRT2, DRT4, DRT7, and DRT9 each use distinct DNA synthesis strategies, suggesting many more mechanisms remain uncharacterized.\n* **Biotechnological potential:** Engineering Drt3b to accept new sequence specificities could open routes to template-free synthesis of defined DNA sequences for biomaterials and beyond.\n\n1. **Protein-templated synthesis of dinucleotide repeat DNA by an antiphage reverse transcriptase:** [Deng et al., Science, 2026](https://www.science.org/doi/10.1126/science.aed1656)\n2. **Diverse enzymatic activities mediate antiviral immunity in prokaryotes:** [Gao et al., Science, 2020](https://www.science.org/doi/10.1126/science.aba0372)\n3. **De novo gene synthesis by an antiviral reverse transcriptase:** [Tang et al., Science, 2024](https://pubmed.ncbi.nlm.nih.gov/39116258/)\n4. **Protein-primed homopolymer synthesis by an antiviral reverse transcriptase:** [Tang et al., Nature, 2025](https://www.nature.com/articles/s41586-025-09179-5)\n5. **Bacterial reverse transcriptase synthesizes long poly(A)-rich cDNA for antiphage defense:** [Song et al., Science, 2025](https://www.science.org/doi/10.1126/science.ads4639)\n\n**A bacterial defense system has revealed that a protein\u2019s active site geometry alone can direct sequence-specific DNA synthesis, extending the known boundaries of biological information flow**\n\nThe central dogma of molecular biology defines a constrained information flow: DNA into RNA into protein, with reverse transcriptases permitting RNA to be read back into DNA. Even that exception requires a nucleic acid template. The assumption that sequence-specific DNA synthesis always depends on a nucleic acid blueprint has stood largely unchallenged for five decades. A paper by [Deng et al.](https://www.science.org/doi/10.1126/science.aed1656), published in *Science* in April 2026, changes that, and the mechanism at its heart was made visible only through high-resolution cryo-electron microscopy.\n\nThe story originates in bacterial immunity. Defense-associated reverse transcriptases (DRTs) are a broad class of anti-phage systems first systematically identified by [Gao et al. in 2020](https://www.science.org/doi/10.1126/science.aba0372). In contrast to CRISPR, which targets foreign DNA for cleavage, DRTs respond to phage attack by synthesizing it. [DRT2 systems](https://pubmed.ncbi.nlm.nih.gov/39116258/) produce coding genes de novo through rolling-circle reverse transcription of a noncoding RNA, generating an effector protein that aborts infection. Two independent groups found that [DRT9](https://www.nature.com/articles/s41586-025-09179-5) synthesizes poly(dA) DNA homopolymers to disrupt viral replication, with [a parallel study](https://www.science.org/doi/10.1126/science.ads4639) showing the same system responds to phage-driven dATP accumulation. The DRT family was already emerging as a collection of biochemical surprises.\n\nDRT3 is the most striking variant yet. Alex Gao\u2019s group at Stanford cloned the DRT3 system from *E. coli* and found it contains two distinct reverse transcriptases, Drt3a and Drt3b, and a noncoding RNA scaffold. Drt3a synthesizes poly(GT) DNA conventionally, reading an ACACAC template encoded in the ncRNA. Drt3b does something entirely new: it generates the complementary poly(AC) strand in the complete absence of any nucleic acid template. Conserved active site residues mimic a template strand in chemical terms, enforcing strict alternating base incorporation through protein geometry alone. Cryo-EM structures at 2.6 \u212b resolution resolved the complete assembly as a D3-symmetric 6:6:6 ribonucleoprotein complex, with both enzymes engaged simultaneously in a single coordinated machine.\n\n> ***\u201cThese findings expand the functional landscape of nucleic acid polymerases, revealing a protein-templated mechanism for sequence-specific DNA synthesis.\u201d***\n\nThe structural data were decisive. Without the atomic-resolution cryo-EM model, it would have been impossible to demonstrate conclusively that no nucleic acid strand occupies the Drt3b active site: the protein is both enzyme and blueprint. Whether DRT3 formally violates the central dogma is debated; the poly(GT/AC) product is repetitive, non-coding, and does not appear to enter the bacterial genome. But it extends the known modes of biological information transfer in an unambiguous way. How the duplex DNA product confers anti-phage immunity remains open, with sequestration of viral replication factors and formation of unusual DNA secondary structures both proposed. More broadly, the finding reinforces a point that often needs making: in mechanistic enzymology, atomic structures are not confirmatory data. They are the experiment.", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\nFIG. 16A shows views of the protein structure in an asymmetric unit. The asymmetric unit consists of 1 light chain, 1 heavy chain from the Fab fragment and a 31-amino acid peptide from Type 1 mutant CalR. FIG. 16B shows a zoom-in view for the region of CDR and mutant CalR peptide. Constant regions of the Fab are excluded for clarity. FIG. 16C shows the CDR composition of the Fab fragment. Bold font denotes amino acids that contribute to the epitope recognition in mutant CalR. CCG numbering scheme was used for defining the CDR in MOE (Molecular Operating Environment, 2019.01; Chemical Computing Group ULC, 1010 Sherbrooke St. West, Suite #910, Montreal, QC, Canada, H3A 2R7, 2021).\n\nFIG. 17A shows an illustration of interacting residues located across CDR regions from the heavy chain and the light chain. FIG. 17B shows selected residues for interaction analysis between the antibody and the antigen CalR31 (31-mer mutant CalR peptide). Residues selected based on distance and interaction type determined by MOE (Molecular Operating Environment, 2019.01; Chemical Computing Group ULC). The side chain of residues are shown in stick representation with numbering adopted from the CCG scheme in MOE. FIG. 17C shows results from a detailed interaction analysis between the CalR31 and the CDR of antibody. The distance between interacting pairs were calculated in MOE. Distances in Angstroms are shown by dotted lines.\n\n### Example 13: Efficacy of Anti-mutCALR Antibody in a Mouse Model of Essential Thrombocythemia\n\nTo test the therapeutic potential of anti-mutCALR antibodies in vivo, an antibody was evaluated in a mouse model of essential thrombocythemia (ET). In this mouse model, a conditional allele that expresses a mutCALR protein with a C-terminal sequence (DEL52) identical to what found in MPN patients was knocked in to the mouse CALR sequence. The engineered mice (CALRdel/del) develop an ET-like disease with marked thrombocytosis, splenomegaly, and abnormal megakaryocytosis (Li et al. *Blood* 2018; 131:649). The expression of mutCALR in the engineered mice was induced with intraperitoneal injections of polyinosinic:polycytidylic acid (poly I:C) (250 \u03bcg/dose; every other day for a total of 3 injections). Treatment initiated 19 weeks post-poly I:C induction and consisted of intravenous injections of anti-mutCALR antibody (clone 74) at 50 mg/kg QW for a total of 4 weeks. ET phenotype was confirmed by assessing the platelet counts in the blood (Sysmex, Kobe, Japan), spleen size and bone marrow histology. A representative study is shown in FIGs. 18A-18C.\n\nThe anti-mutCALR antibody restored normal platelet counts (FIG. 18A), spleen volume (FIG. 18B), and the bone marrow cell environment (FIG. 18C). Taken together, these results demonstrate the efficacy of the anti-mutCALR antibodies in treating ET.\n\n### Example 14: Efficacy of Anti-mutCALR Antibody in Primary Patient Cells\n\nCD34+ cells isolated from MPN patients carrying the CALR mutation were used to characterize the ability of an anti-mutCALR antibody to inhibit the mutCALR-derived oncogenic function. Peripheral blood mononuclear cells (PBMCs) were isolated from non-identified blood samples from MPN patients by Ficoll gradient extraction (Fisher Scientific) and CD34+ cells were enriched using magnetic enrichment (Miltenyi Biotec). CD34+ cells were cultured for seven days in SFEM-II media (STEMCELL Technologies) containing hSCF, hFLT3L, TPO, LDL2698, SR1, and UM171.\n\nCD34+ cells (50,000 cells/well) were then plated into 96-well plates and treated with a mutCALR or isotype control antibody for 2 hours. Following treatment, plates were centrifuged, the supernatant was aspirated and subsequently washed with PBS. After centrifugation the cell pellets were lysed using lysing buffer (Cell Signaling Technologies) supplemented with 1x Halt Protease and Phosphatase Inhibitor Cocktail (Thermo Fisher Scientific). Lysates were added to the Phospho (Tyr694)/Total STAT5a,b whole cell lysate kit (Meso Scale Diagnostics) and phospho-STAT5 levels were then quantitated using Meso Sector S 600 (Meso Scale Diagnostics). Clone 74 selectively inhibited pSTAT5 in CD34+ cells harboring mutCALR in a dose-dependent manner, while the isotype control (IgG) at 10 \u03bcg/mL had no impact on CD34+ cells. Moreover, pSTAT5 was not inhibited in CD34+ cells harboring the V617F JAK2 mutation (FIG. 19A).\n\nTo evaluate the ability of antibodies to inhibit mutCALR oncogenic function and imbalanced proliferation of megakaryocytes, CD34+ cells (50,000 cells/well) were added to a 12-well plate with SFEM-II supplemented with hSCF, hGCSF, hIL3, and hIL6 and treated with a mutCALR antibody or isotype control for 6 days. Cells were stained and analyzed by flow cytometry (LSRFortessa X-20 analyzer, BD Biosciences). Antibodies used were: APC anti-human CD38 antibody (BioLegend), FITC anti-human lineage cocktail (BioLegend), PE/Cyanine7 anti-human CD34 antibody (BioLegend), PE anti-human CD41 antibody (BioLegend), APC mouse anti-human CD42b antibody (BD Pharmingen). Megakaryocytes were identified as the CD41+ CD42b+ cells. The anti-mutCALR antibodies (clones 74 and 65) selectively prevented the differentiation of mutCALR+ CD34+ cells into mature megakaryocytes in a dose-dependent manner, while isotype control (IgG) had no impact on this population. A representative experiment is shown in FIG. 19B.", "# GitHub README: SiteHealthInspector\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/SiteHealthInspector, 2026\n\n> **A pure-Bash, colour-coded website health audit. No install, no dependencies.**\n\n    \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/SiteHealthInspector\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/SiteHealthInspector</a></td>\n</tr>\n</table>\n\n---\n\n> A terminal-based website health auditing tool written in pure Bash.\n\nWhy it matters: when a site feels slow or you need to confirm TLS, caching, and security headers are sound, you want an answer in seconds, not a heavyweight tool to set up. SiteHealthInspector.sh is a single Bash script that probes timing, HTTP status, caching and CDN behaviour, security headers, the TLS certificate, DNS records, CMS fingerprint, and an overall weighted health score, all rendered with star ratings and bar charts for instant triage. It is useful for developers and site owners who live in the terminal and want a fast, dependency-free pulse check on any URL: perfect for spot checks, CI smoke tests, and before/after comparisons when tuning performance.\n\n`SiteHealthInspector.sh` runs a comprehensive, colour-coded diagnostic scan against any URL \u2014 reporting timing metrics, HTTP status, caching behaviour, security headers, SSL/TLS certificate status, DNS records, CMS fingerprinting, and an overall health scorecard. No runtime dependencies beyond standard Unix tools.\n\n---\n\n## \u2728 Features\n\n| Module | What it checks |\n|---|---|\n| \u23f1 **Timing Metrics** | DNS resolution, TCP connect, TLS handshake, TTFB, total load time, redirect overhead, download speed |\n| \ud83d\udcf6 **HTTP Status & Redirect Chain** | Final HTTP status code with full hop-by-hop redirect trace |\n| \ud83d\udce6 **Cache & CDN Headers** | `Cache-Control`, `ETag`, `Last-Modified`, `Age`, Cloudflare cache status (HIT/MISS/BYPASS/DYNAMIC), `Vary`, content encoding |\n| \ud83d\udd10 **Security Headers** | HSTS, CSP, `X-Content-Type-Options`, `X-Frame-Options`, XSS Protection, Referrer Policy, Permissions Policy, COEP, COOP |\n| \ud83d\udd11 **SSL/TLS Certificate** | Subject, issuer, validity window, days-to-expiry with renewal warnings, TLS protocol version, cipher suite, certificate verification |\n| \ud83d\udca8 **Compression & Content** | gzip/brotli/zstd detection, HTTP/1.1 vs HTTP/2 vs HTTP/3 negotiation, page size, content-type |\n| \ud83c\udf0d **DNS & Connectivity** | A, AAAA, MX, NS records; SPF and DMARC TXT records; 5-run TTFB stability test with jitter calculation |\n| \u2699 **CMS / Platform Detection** | Generator meta tag, WordPress login/REST API exposure, sitemap.xml, robots.txt |\n| \ud83d\udc93 **Availability & Uptime** | Live HTTP check, `www` subdomain, HTTP\u2192HTTPS redirect enforcement, 404 and `/wp-admin/` probing |\n| \ud83c\udfc6 **Overall Health Scorecard** | Weighted 0\u2013100 score across all modules with grade: EXCELLENT / GOOD / FAIR / POOR |\n\nAll output uses ANSI colour coding with star ratings (\u2605\u2605\u2605\u2605\u2605) and bar charts for instant visual triage.\n\n---\n\n## \ud83d\udccb Requirements\n\n| Dependency | Used for |\n|---|---|\n| `bash` \u2265 3.2 | Script runtime (macOS/Linux compatible) |\n| `curl` | HTTP requests, timing metrics, SSL verification |\n| `openssl` | TLS certificate introspection |\n| `dig` | DNS record lookups |\n| `bc` | Floating-point arithmetic for timing calculations |\n\nAll of the above ship with macOS and most Linux distributions. On minimal containers (e.g. Alpine), install with:\n\n```bash\napk add curl openssl bind-tools bc\n```\n\n---\n\n## \ud83d\ude80 Usage\n\n```bash\n# Scan the default target (marcdeller.com)\nbash SiteHealthInspector.sh\n\n# Scan a specific URL\nbash SiteHealthInspector.sh https://example.com\n\n# Make executable and run directly\nchmod +x SiteHealthInspector.sh\n./SiteHealthInspector.sh https://your-site.com\n```\n\n> **Note:** The script follows redirects (`-L`) and accepts a full URL including scheme. Both `https://` and `http://` targets are supported.\n\n---\n\n## \ud83d\udcca Sample Output\n\n```\n  ============================================================\n  \ud83c\udf10  SITE HEALTH INSPECTOR  v1.2\n  Target: https://example.com\n  2025-06-11  14:32:07  BST\n  ============================================================\n\n  ============================================================\n  \u23f1  TIMING METRICS\n  ============================================================\n  \ud83d\udd0d DNS Resolution: 12ms  \u2605\u2605\u2605\u2605\u2605\n  \ud83d\udd0c TCP Connect: 45ms  \u2605\u2605\u2605\u2605\u2605\n  \ud83d\udd12 TLS Handshake: 98ms  \u2605\u2605\u2605\u2605\u2605\n  \u26a1 TTFB (Time to First Byte): 187ms  \u2605\u2605\u2605\u2605\u2605\n  \ud83c\udfc1 Total Load Time: 423ms  \u2605\u2605\u2605\u2605\u2605\n\n  ...\n\n  \ud83c\udfc6 Overall Score: \ud83c\udfc6 EXCELLENT  87/100\n```\n\n---\n\n## \ud83c\udfaf Scoring\n\n### Timing thresholds (lower is better)\n\n| Metric | \u2605\u2605\u2605\u2605\u2605 Good | \u2605\u2605\u2605\u2606\u2606 Fair | \u2605\u2606\u2606\u2606\u2606 Poor |\n|---|---|---|---|\n| DNS Resolution | \u2264 50 ms | \u2264 200 ms | > 200 ms |\n| TCP Connect | \u2264 100 ms | \u2264 400 ms | > 400 ms |\n| TLS Handshake | \u2264 150 ms | \u2264 400 ms | > 400 ms |\n| TTFB | \u2264 200 ms | \u2264 600 ms | > 600 ms |\n| Total Load | \u2264 500 ms | \u2264 2000 ms | > 2000 ms |\n\n### Overall health score\n\n| Score | Grade |\n|---|---|\n| 80\u2013100 | \ud83c\udfc6 EXCELLENT |\n| 60\u201379 | \u26a1 GOOD |\n| 40\u201359 | \u26a0 FAIR |\n| 0\u201339 | \ud83d\udea8 POOR |\n\nKey contributors to the overall score include: TTFB, total load time, compression, HTTP/2+, valid SSL, certificate validity, HSTS presence, cache headers, Cloudflare HIT status, and live availability.\n\n---\n\n## \ud83d\udd12 Security Header Coverage\n\nThe script audits all headers recommended by [OWASP Secure Headers Project](https://owasp.org/www-project-secure-headers/):\n\n- `Strict-Transport-Security` (HSTS)\n- `Content-Security-Policy` (CSP)\n- `X-Content-Type-Options`\n- `X-Frame-Options`\n- `X-XSS-Protection`\n- `Referrer-Policy`\n- `Permissions-Policy`\n- `Cross-Origin-Embedder-Policy` (COEP)\n- `Cross-Origin-Opener-Policy` (COOP)\n\n---\n\n## \ud83d\uddc2 File Structure\n\n```\nSiteHealthInspector.sh    # Single-file script \u2014 no config, no dependencies to install\n```\n\n---\n\n## \ud83d\udee0 Customisation", "# Elora Therapeutics: Treating Microplastics in Humans\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-02-18 on marcdeller.com (categories: microplastics).*\n\nPaul Accepting The Ventures - Austin 2026, Best Team Award Winners\n\nPaul Swartzendruber (left) and Marc C. Deller (right) Co-Founders of Elora Therapeutics\n\n* **Pervasive Exposure:** Microplastics have been detected in human blood, brain, and placental tissue, with PET being the most common polymer, and there is no existing therapeutic to remove them.\n* **Proof of Concept:** PETase has been shown to degrade PET microplastics into non-toxic monomers in human serum at physiological temperature without compromising cell viability.\n* **Validation Underway:** Elora Therapeutics is planning a multi-milestone validation study with UT Austin to establish enzyme stability, safety, and efficacy in humans.\n* **New Therapeutic Category:** Elora Therapeutics is the first company pursuing FDA approval for an enzyme-based therapeutic that degrades and safely removes microplastics from the human body.\n\n* [Elora Therapeutics](https://eloratherapeutics.com)\n* [Elora Therapeutics Pitch Deck](https://eloratherapeutics.com/s/deck)\n* [Austin Technology Incubator](https://ati.utexas.edu/company/elora-developing-a-systemically-delivered-enzyme-to-remove-microplastics-from-the-human-body/)\n* [Elora Therapeutics](https://www.linkedin.com/company/elora-therapeutics/posts/?feedView=all)\n* [Marc C Deller](https://www.linkedin.com/in/marccdeller/)\n* [Paul Swartzendruber](https://www.linkedin.com/in/paulswartzendruber/)\n* [Microplastics Blog](https://marcdeller.com/microplastics)\n* [Elora Therapeutics Provisional Patent](https://marcdeller.com/wp-content/uploads/2026/02/Elora-Therapeutics-Provisional-Patent.pdf)\n* [isPETase Discovered in Landfill Isolates](https://www.science.org/doi/10.1126/science.aad6359)\n* [Efficacy and Safety in Human Serum](https://pubs.rsc.org/en/content/articlepdf/2024/fd/d4fd00014e)\n* [PET-hydrolytic Enzymes in the Gut](https://www.sciencedirect.com/science/article/abs/pii/S0141813024085428)\n* [MicroPlastics Blaster](https://marcdeller.com/microplastic-blaster)\n\n**We aim\u00a0to Develop a *First-in-Class* Enzyme Therapy Designed to Degrade and Safely Remove Microplastics from the Human Body**\n\nHumans ingest an estimated 74,000 to 121,000 microplastic particles per year, and these synthetic fragments now appear in blood, brain tissue, and even the placenta, where they drive chronic inflammation, metabolic dysfunction, and endocrine disruption.\n\nAmong the polymers detected, polyethylene terephthalate (PET) is the most prevalent in human tissues, and our biology has no mechanism to clear it. Elora Therapeutics is developing the first FDA-targeted therapeutic to address this problem at its origin: engineered PETase and MHETase enzymes designed to depolymerize PET microplastics into non-toxic monomers directly within biological fluids.\n\nRecent proof-of-concept research demonstrated that an engineered PETase can degrade PET in human serum at 37\u00b0C without affecting cell viability, establishing that enzymatic degradation of microplastics in the body is scientifically feasible. Building on this foundation, Elora Therapeutics has partnered with the University of Texas at Austin to conduct a systematic validation study evaluating enzyme stability, dose-response safety, and breakdown product biocompatibility in human serum and cell models.\n\nEmbedded in the Austin Technology Incubator and led by a team with deep experience across Pfizer, Stanford, Ginkgo Bioworks, and the NIH, Elora is positioned to define an entirely new category of biologic therapeutics: one that targets the environmental drivers of metabolic disease rather than managing downstream symptoms.", "# Crystal Structure of an Immature VRC01-Class Antibody DRVIA7 from a Chinese Donor Bound to Clade A/E HIV-1 gp120 Core\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [5CD5](https://www.rcsb.org/structure/5CD5)\n* **Release Date:** 2016-04-06\n* **Method:** X-RAY DIFFRACTION 3.396 \u00c5\n* **Organisms:** Homo sapiens, Human immunodeficiency virus 1\n* **Macromolecules:** 93TH057 HIV-1 gp120 core, DRVIA7 Fab Heavy Chain, DRVIA7 Fab Light Chain (proteins)\n* **Unique Ligands:** NAG\n* **Publication:** [Kong, L., et al. (2016). Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor. Immunity, 44, 939-950.](https://www.cell.com/immunity/fulltext/S1074-7613(16)30095-4)\n* **Blog Post:** [Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor](https://marcdeller.com/key-gp120-glycans-pose-roadblocks-to-the-rapid-development-of-vrc01-class-antibodies-in-an-hiv-1-infected-chinese-donor/)\n\n**Description:** This structure captures an immature VRC01-class antibody (DRVIA7) in complex with HIV-1 gp120, revealing how key glycans\u2014particularly at positions N276 and V5\u2014create steric roadblocks that impede the maturation of broadly neutralizing antibodies. The structural and longitudinal analyses inform glycan-accommodating vaccine design strategies for eliciting VRC01-class responses across diverse populations.", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\nIn certain embodiments, the bispecific anti-mutCALR antibody is a biparatopic antibody. A biparatopic antibody is an antibody which recognizes two non-identical epitopes (overlapping or non-overlapping epitopes) on the same target antigen (e.g., the C-terminal of mutCALR domain). A biparatopic anti-mutCALR antibody can comprise two immunoglobulin heavy chain-light chain pairs or one immunoglobulin heavy chain-light chain pair. In some embodiments, a biparatopic anti-mutCALR antibody is a full-length antibody comprising one immunoglobulin heavy-chain-light chain pair.\n\nIn some instances, the anti-mutCALR antibody is a multivalent antibody. A multivalent antibody may be internalized (and/or catabolized) faster than a bivalent antibody by a cell expressing an antigen to which the antibodies bind. The antibodies described herein can be multivalent antibodies with three or more antigen binding sites (e.g., tetravalent antibodies). The multivalent antibody can comprise a dimerization domain and three or more antigen binding sites; an exemplary dimerization domain comprises (or consists of) an Fc region or a hinge region.\n\n### Production and Purification\n\nThe antibodies of the present disclosure can be isolated from inside or outside (such as from the medium) of the host cell and purified as substantially pure and homogenous antibodies. Antibodies may be isolated and purified by appropriately selecting and combining, for example, column chromatography, filtration, ultrafiltration, salting out, solvent precipitation, solvent extraction, distillation, immunoprecipitation, SDS-polyacrylamide gel electrophoresis, isoelectric focusing, dialysis, and recrystallization. Chromatography includes affinity chromatography, ion exchange chromatography, hydrophobic chromatography, gel filtration, reverse-phase chromatography, and adsorption chromatography. Chromatography can be carried out using liquid phase chromatography such as HPLC and FPLC. Columns used for affinity chromatography include protein A column and protein G column; examples of columns using a protein A column include Hyper D, POROS, and Sepharose FF (GE Healthcare Biosciences).\n\n### Polynucleotides, Expression Vectors, and Cells\n\nThe disclosure also provides polynucleotides and vectors encoding an anti-mutCALR antibody or portion thereof (e.g., VH, VL, HC, or LC) described herein. The polynucleotides of the disclosure can be in the form of RNA or in the form of DNA. In some instances, the polynucleotide is DNA; in some instances complementary DNA (cDNA); in some instances RNA. In some instances, the polynucleotide encodes a VH comprising the VH CDR1, VH CDR2, and VH CDR3 of any antibody described herein (see, e.g., Tables 1-5); a VL comprising the VL CDR1, VL CDR2, and VL CDR3 of any antibody described herein; or a heavy chain or light chain comprising those regions. Also provided herein are cells comprising one or more expression vectors described herein.\n\n### Methods of Treatment\n\nAnother aspect of the present disclosure pertains to methods of treating a myeloproliferative neoplasm in an individual (e.g., patient) by administering to the individual in need of such treatment a therapeutically effective amount or dose of one or more antibodies of the present disclosure or a pharmaceutical composition thereof.\n\nNon-limiting examples of a myeloproliferative neoplasm include chronic myelogenous leukemia, polycythemia vera, primary myelofibrosis, essential thrombocythemia, chronic neutrophilic leukemia, acute myelogenus leukemia, chronic eosinophilic leukemia, chronic myelomonocytic leukemia, myeloproliferative neoplasm and myelodysplastic syndrome, including myelodysplastic syndrome with refractory anaemia with ring sideroblasts, myelodysplastic syndrome with refractory anemia, and myelodysplastic syndrome with refractory anemia with excess blasts.\n\nAnti-mutCALR antibodies disclosed herein can be used to treat, alone or in combination with other therapies, a myeloproliferative neoplasm, or can be used, alone or in combination with other therapies, for the manufacture of a medicament for the treatment of a myeloproliferative neoplasm. Non-limiting examples of other therapies include a JAK inhibitor (e.g., ruxolitinib, itacitinib), a PI3K inhibitor (e.g., parsaclisib), a standard of care therapy (e.g., IFN-alpha, hydroxyurea, thalidomide, lenalidomide, an androgen, an erythropoietin-stimulating agent, a chemotherapeutic agent), or a combination thereof. Non-limiting examples of JAK inhibitors for use as described herein are provided in U.S. Pat. Nos. 7,335,667; 9,359,358; 8,691,807; 9,181,271; and 9,034,884. Non-limiting examples of PI3K inhibitors are provided in U.S. Pat. Nos. 9,108,984; 9,062,055; 8,759,359; and 9,434,746.", "# GitHub README: gpcr-binning\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/gpcr-binning, 2026\n\n> **Structural binning and organisation utilities for GPCR datasets.**\n\n    \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/gpcr-binning\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/gpcr-binning</a></td>\n</tr>\n</table>\n\n---\n\nGPCR Binning is a working repository for organising and grouping G-protein-coupled receptor (GPCR) structural data. GPCRs are the single largest class of drug targets, and the structural data describing them is large, heterogeneous, and growing fast across receptor families, conformational states, and bound ligands. This project collects utilities for binning that data into coherent, analysable groups.\n\nWhy it matters: GPCRs account for a substantial fraction of all approved drugs, and structure-based work on them depends on being able to slice a sprawling structural landscape into meaningful categories (by class, by state, by binding mode, by resolution and method). Sensible binning is the foundation for everything downstream: selectivity analysis, template selection for modelling, and triage for structure-based drug design. This repository is useful as a place to develop and keep those binning and grouping utilities in one structural-biology-aware codebase.\n\n> \u26a0\ufe0f **Status:** This project is in active development. The sections below describe its intended scope; specifics will be filled in as the codebase matures.\n\n---\n\n## \ud83c\udfaf Scope\n\n| Area | Intent |\n|---|---|\n| \ud83d\uddc3\ufe0f **Binning** | Group GPCR structures into coherent categories for analysis |\n| \ud83e\uddec **Domain-aware** | Built around GPCR structural biology rather than generic clustering |\n| \ud83d\udd2c **Analysis-ready** | Produce groupings suitable for downstream SBDD and comparison |\n\n---\n\n## \ud83d\ude80 Getting started\n\n```bash\ngit clone https://github.com/bellcheddar/gpcr-binning.git\ncd gpcr-binning\n```\n\nUsage instructions will be documented here as the tooling stabilises.\n\n---\n\n## \ud83e\udded Context\n\nThis repository sits alongside related structure-focused tools in the portfolio, including [PocketTide](https://github.com/bellcheddar/PocketTide) (a GLP-1R structure dashboard) and the Boltz-2 co-folding utilities. Together they support a structure-based approach to drug discovery across GPCR and other target classes.\n\n---\n\n## \ud83d\udcc4 Licence\n\nSee repository for licence terms.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/gpcr-binning\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/gpcr-binning</a></td>\n</tr>\n</table>", "# About Marc C. Deller, D.Phil.\nAuthors: Deller, M.C.\nSource: marcdeller.com / chatMCD profile, 2026\n\nD.Phil. in structural biology / biophysics, University of Oxford (1995\u20131999), Laboratory of Molecular Biophysics and the Wellcome Trust Centre for Human Genetics, supervised by Prof. E. Yvonne Jones. Thesis: \"Structural Studies of Cytokines and Cytokine Receptors\", in which he determined the first crystal structure of the cytokine oncostatin M.\n\nBSc (First Class Honours) in Biochemistry and Molecular Biology, University of Leeds, mentored by Prof. Tony North, who inspired him to pursue a doctorate in protein structure. During his degree he spent a one-year industrial placement at Boots Pharmaceuticals in the team that developed Advil/ibuprofen/Nurofen.\n\nA-levels at High Pavement College, Nottingham (B grades in biology and chemistry). GCSEs at Henry Mellish School, Nottingham (A grades in biology, chemistry and physics), where his mentor Gilbert Gimson inspired him to study chemistry. Primary school: St Mary's Primary School, Bulwell.\n\nCertifications (10+) include Project Management, AI, Serverless Architectures, Data Analysis, Introduction to R, Dashboard Development, and Communication Skills with Generative AI.\n\n## Biggest career achievements\n\nThe mutant-CALR antibody patent. At Incyte Marc led a \"structural biology sprint\" to solve the peptide-bound mutant-calreticulin\u2013antibody complex in just over ten days, beating an already aggressive two-week patent deadline and securing a first-in-class anti-mutant-CALR antibody patent (US 2023/0272055 A1) projected to generate more than $500M in annual sales.\n\nThe $8M Incyte structural biology platform. He designed and built an integrated, automated platform (crystallisation robotics, X-ray, Glacios cryo-EM and AI/cloud analytics) that fixed hit-to-lead bottlenecks and delivered a four-fold efficiency gain, a 67% throughput increase and 100+ co-crystal structures in under three years across ten programmes and five new targets.\n\nThe Stanford Macromolecular Structure Knowledge Center. He established and scaled a profitable wet-lab service organisation with a 70-project portfolio and $1M annual budget, equipping 50+ researchers, generating $60K in yearly revenue, enabling six novel structures, supporting three NIH SBIR submissions and expanding SSRL/SLAC access to 70 users from ten countries.\n\nEarlier landmarks include the first structure of oncostatin M (Oxford, published in Structure in 2000), 70+ novel structures and the Top96 crystallisation screen at the JCSG, and crystallography contributions to the first high-resolution structure of a soluble cleaved HIV-1 envelope trimer (Science, 2013).\n\n## Scientific expertise\n\nTherapeutic areas: oncology, immunology, infectious disease, and enzymology for microplastic remediation.\n\nTarget classes and named targets: cytokines and cell-surface receptors (oncostatin M, LIFR, gp130), receptor tyrosine kinases (VEGFR, cMET, FGFR2/3), proteases (HIV protease), potassium channels (TrkA gating ring), cellulases and esterases, HCV envelope E2, HIV-1 envelope and its receptor, neutralising antibodies, ERK, HIV topoisomerase, mutant-CALR antibody, ALK2, CDK2/4, KRAS G12D, JAK1/2, WRN helicase, VRK1 and USP deubiquitylases.\n\nCore skills: structure\u2013activity relationships, biophysical assays (SPR, ITC, DSF and others), X-ray crystallography, cryo-EM (Glacios, RELION, CryoSPARC), NMR, MS and HDX, construct design, protein expression (E. coli, mammalian, insect, cell-free) and purification, and protein production at scale.\n\nComputational and software stack: AlphaFold, Boltz-2, ESMFold, ProteinMPNN, RFdiffusion, Rosetta, MOE, AutoDock, GROMACS, OpenMM, Schr\u00f6dinger, PyMOL, ChimeraX, COOT, HKL2000/3000, PHENIX, CCP4/CCP4-EM, PLIP, RDKit, PSILO, RockMaker, ChemCart, Benchling, BioRender, ChemDraw, Python, R/Shiny, MATLAB, AWS, HPC (SLURM, Globus), LIMS, KNIME, GraphPad, Spotfire, JIRA, Confluence and Smartsheet.\n\n## Leadership style and philosophy\n\nMarc lives by the motto of the 6 C's: Curiosity, Courage, Creativity, Communication, Compassion and Collaboration. He mentors teams using the 6 C's, and his door is always open to opinions from all perspectives. He likes to bring a fun and engaging aspect to science and research, with a mentality of exploring and finding the truth. One of his favourite lines: \"If science were successful on the first attempt, it would be called 'Search', which rarely happens, so that's why we call it 'Research'.\" He also quotes Steve Jobs: \"Stay hungry, stay foolish.\" He is a supporter of freedom of thought, expression of principles and helping others in need.\n\nThe most important things in his life are family, friends and animals; being there for friends when they need you; supporting others in need; being kind; giving back; and listening.\n\n## Family background\n\nMarc was born in Bulwell, Nottingham, on 21 October 1972, the son of Norman and Valerie Deller. Norman was a hard-working printer and Valerie a devoted stay-at-home mum who kept things running. He has an older brother, Andy (Andr\u00e9), an engineer turned business leader. Marc's own family includes his dogs Matchi, Brinkley, Danny and Harvey.\n\n## Favourites\n\nFavourite towns: Nottingham, San Diego and Wilmington. Favourite food: fish and chips. Favourite leader: Brian Clough. Favourite band: Man or Astro-Man? Favourite guitar: a '62 Fender Jazzmaster. Favourite movie: Night of the Comet. Favourite dog: Matchi. Favourite artist: Pete Fowler. Favourite footballer: George Best.\n\n## Personal style\n\nMarc's style is sports casual: think a high-end Hugo Boss suit with sneakers. His favourite clothing brands are Banana Republic, Lacoste, Simple (no, he is not Larry David) and Paul Smith. A Nottingham Forest jersey is never out of style. His favourite glasses are Oakley Frogskins. He prefers beer to cocktails, but a good gin and tonic with lemon is great.\n\n## Fun facts\n\nMarc is an avid collector of vinyl records and instrumental surf music. He supports Nottingham Forest, the Philadelphia Phillies and the Philadelphia Eagles. He is a Lego enthusiast and, in his younger days, a keen mountain biker. He drinks Earl Grey tea and San Diego IPAs (and Yuengling!), and enjoys Timothy Taylor's Boltmaker, the beer that inspired the name of his BoltzMaker tool. He likes the TV show The IT Crowd, but if Jim is asking, he hasn't actually watched it yet. His best school grades were A's in biology, chemistry and physics at GCSE.", "# codswallop.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n## codswallop.mdeller.com/\n\n```\n ######   #######  ########   ######  ##      ##    ###    ##       ##        #######  ########\n##    ## ##     ## ##     ## ##    ## ##  ##  ##   ## ##   ##       ##       ##     ## ##     ##\n##       ##     ## ##     ## ##       ##  ##  ##  ##   ##  ##       ##       ##     ## ##     ##\n##       ##     ## ##     ##  ######  ##  ##  ## ##     ## ##       ##       ##     ## ########\n##       ##     ## ##     ##       ## ##  ##  ## ######### ##       ##       ##     ## ##\n##    ## ##     ## ##     ## ##    ## ##  ##  ## ##     ## ##       ##       ##     ## ##\n ######   #######  ########   ######   ###  ###  ##     ## ######## ########  #######  ##\n```\n\n**C**omparative **O**verview of\n**D**omains, **S**tructures,\n**W**hole-families, **A**lignments,\n**L**igands, **L**ineages,\n**O**rganisms & **P**DB entries.\n\n# The PDB, summarised\u2026\n\nHand it a protein and it does an afternoon\u2019s work of manual PDB trawling in ninety\nseconds: every entry, every construct, every ligand, cross-referenced and interactive.\n\nTry\n\nA PDB ID, a UniProt accession, a gene name, a Pfam or InterPro identifier, a raw\nsequence, or just what the protein is called. Anything ambiguous asks rather than guesses.\n\n## Recent searches\n\n* [Beta-lactamase inhibitory protein\n  **20**](/f/beta-lactamase-inhibitory-protein-p35804)\n* [Vascular endothelial growth factor receptor 1\n  **1333**](/f/vascular-endothelial-growth-factor-receptor-1-p17948)\n* [Sodium channel protein type 10 subunit alpha\n  **101**](/f/sodium-channel-protein-type-10-subunit-alpha-q9y5y9)\n* [Tubulin alpha-1A chain\n  **864**](/f/tubulin-alpha-1a-chain-q71u36)\n* [Mitogen-activated protein kinase 14\n  **1629**](/f/mitogen-activated-protein-kinase-14-q16539)\n* [Tyrosine-protein phosphatase non-receptor type 11\n  **927**](/f/tyrosine-protein-phosphatase-non-receptor-type-1-q06124)\n* [Enoyl-[acyl-carrier-protein] reductase [NADH]\n  **290**](/f/enoyl-acyl-carrier-protein-reductase-nadh-p9wgr1)\n* [Actin, alpha skeletal muscle\n  **740**](/f/actin-alpha-skeletal-muscle-p68133)\n* [Serine/threonine-protein kinase mTOR\n  **102**](/f/serine-threonine-protein-kinase-mtor-p42345)\n* [Alpha-synuclein\n  **231**](/f/alpha-synuclein-p37840)\n* [RAC-alpha serine/threonine-protein kinase\n  **1589**](/f/rac-alpha-serine-threonine-protein-kinase-p31749)\n* [Dipeptidyl peptidase 4\n  **160**](/f/dipeptidyl-peptidase-4-p27487)\n* [Streptavidin\n  **382**](/f/streptavidin-p22629)\n* [Cystic fibrosis transmembrane conductance regulator\n  **235**](/f/cystic-fibrosis-transmembrane-conductance-regula-p13569)\n* [Fibroblast growth factor receptor 1\n  **1946**](/f/fibroblast-growth-factor-receptor-1-p11362)\n* [Microtubule-associated protein tau\n  **242**](/f/microtubule-associated-protein-tau-p10636)\n* [Androgen receptor\n  **367**](/f/androgen-receptor-p10275)\n* [Replicase polyprotein 1ab\n  **1962**](/f/replicase-polyprotein-1ab-p0dtd1)\n* [Calmodulin-1\n  **1105**](/f/calmodulin-1-p0dp23)\n* [Polyubiquitin-C\n  **1605**](/f/polyubiquitin-c-p0cg48)\n* [Cytochrome P450 3A4\n  **189**](/f/cytochrome-p450-3a4-p08684)\n* [Amyloid-beta precursor protein\n  **380**](/f/amyloid-beta-precursor-protein-p05067)\n* [High affinity nerve growth factor receptor\n  **1986**](/f/high-affinity-nerve-growth-factor-receptor-p04629)\n* [Albumin\n  **257**](/f/albumin-p02768)\n* [Transthyretin\n  **515**](/f/transthyretin-p02766)\n* [Insulin\n  **508**](/f/insulin-p01308)\n* [Serine protease 1\n  **1922**](/f/serine-protease-1-p00760)\n* [Coagulation factor X\n  **1784**](/f/coagulation-factor-x-p00742)\n* [Beta-galactosidase\n  **111**](/f/beta-galactosidase-p00722)\n* [Superoxide dismutase [Cu-Zn]\n  **251**](/f/superoxide-dismutase-cu-zn-p00441)\n* [Bromodomain-containing protein 4\n  **1996**](/f/bromodomain-containing-protein-4-o60885)\n* [Menin\n  **70**](/f/menin-o00255)\n* [Hemoglobin subunit beta\n  **528**](/f/hemoglobin-subunit-beta-4hhb-2)\n* [TURKEY EGG WHITE LYSOZYME\n  **1691**](/f/turkey-egg-white-lysozyme-135l-1)\n* [Hemoglobin subunit alpha\n  **533**](/f/hemoglobin-subunit-alpha-4hhb-1)\n* [C-type lysozyme/alpha-lactalbumin family\n  **1691**](/f/c-type-lysozyme-alpha-lactalbumin-family-135l-1)\n* [Lysozyme C\n  **1691**](/f/lysozyme-c-p00698)\n* [Acetylcholinesterase\n  **560**](/f/acetylcholinesterase-p22303)\n* [Poly [ADP-ribose] polymerase 1\n  **165**](/f/poly-adp-ribose-polymerase-1-p09874)\n* [Serine/threonine-protein kinase B-raf\n  **1039**](/f/serine-threonine-protein-kinase-b-raf-p15056)\n* [Prothrombin\n  **1409**](/f/prothrombin-p00734)\n* [Estrogen receptor\n  **760**](/f/estrogen-receptor-p03372)\n* [Heat shock protein HSP 90-alpha\n  **674**](/f/heat-shock-protein-hsp-90-alpha-p07900)\n* [Beta-secretase 1\n  **456**](/f/beta-secretase-1-p56817)\n* [Bacteriorhodopsin\n  **284**](/f/bacteriorhodopsin-p02945)\n* [Adenosine receptor A2a\n  **555**](/f/adenosine-receptor-a2a-p29274)\n* [Rhodopsin\n  **103**](/f/rhodopsin-p08100)\n* [Cyclin-dependent kinase 2\n  **1974**](/f/cyclin-dependent-kinase-2-p24941)\n* [Beta-lactamase TEM\n  **769**](/f/beta-lactamase-tem-p62593)\n* [Green fluorescent protein\n  **1164**](/f/green-fluorescent-protein-p42212)\n* [Ferritin heavy chain\n  **585**](/f/ferritin-heavy-chain-p02794)\n* [Histone H3.1\n  **1172**](/f/histone-h3-1-p68431)\n* [Myoglobin\n  **561**](/f/myoglobin-p02185)\n* [Hemoglobin subunit alpha\n  **533**](/f/hemoglobin-subunit-alpha-p69905)\n* [Dihydrofolate reductase\n  **401**](/f/dihydrofolate-reductase-p00374)\n* [Ribonuclease pancreatic\n  **500**](/f/ribonuclease-pancreatic-p61823)\n* [Mono(2-hydroxyethyl) terephthalate hydrolase\n  **10**](/f/mono-2-hydroxyethyl-terephthalate-hydrolase-a0a0k8p8e7)\n* [Tyrosine-protein kinase JAK1\n  **1814**](/f/tyrosine-protein-kinase-jak1-p23458)\n* [Poly(ethylene terephthalate) hydrolase\n  **209**](/f/poly-ethylene-terephthalate-hydrolase-a0a0k8p6t7)\n* [Spike glycoprotein\n  **1991**](/f/spike-glycoprotein-p0dtc2)", "# BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n## \ud83d\udd2c Ligand validation & scaffold highlighting\n\nTwo related but distinct checks run over every SMILES ligand before you commit hours of `boltz predict` time to them, and both surface directly in the dashboard.\n\n**Why this exists:** Boltz folds whatever chemistry it\u2019s given \u2014 an undefined stereocentre, an unintended protonation state, or a stray counterion left in a SMILES string doesn\u2019t raise an error, it just silently changes the predicted pose and affinity. These are exactly the mistakes a non-specialist (or a tired specialist) makes typing SMILES by hand, and they\u2019re invisible until you\u2019re staring at a confusing result with no idea the input was ever wrong.\n\n### Ligand preparation (validity checks)\n\nAt parse time, every ligand SMILES is canonicalized (RDKit) so the same molecule is represented consistently everywhere downstream \u2014 the generated YAML, the summary table, and cif2plip\u2019s own ligand-matching (see the InChIKey-based matching note in <CHANGELOG.md>). Then, both at `preflight` (as the `ligand_preparation` check) and again in the dashboard\u2019s \u201cLigand preparation\u201d card, each ligand is checked for:\n\n| Check | How | What it means |\n| --- | --- | --- |\n| Undefined stereocentres | `Chem.FindMolChiralCenters(includeUnassigned=True)` | A stereocentre exists in the molecule but the SMILES doesn\u2019t specify which enantiomer/diastereomer \u2014 Boltz will fold *some* version of it, possibly not the one you intended |\n| Disconnected fragments | `Chem.GetMolFrags()` returns more than one fragment | Likely a salt or counterion left in the SMILES (e.g. a sodium carboxylate written as two components) |\n| Ionizable groups | SMARTS match: carboxylic acid, primary/secondary amine, phenol, sulfonic acid | The group\u2019s protonation state at physiological pH isn\u2019t specified by a plain SMILES \u2014 worth a deliberate choice, not a default assumption |\n\nAll of this is advisory, not a hard failure \u2014 these can be legitimate, deliberate modelling choices \u2014 but they\u2019re the kind of thing worth a second look before trusting downstream numbers.\n\n### Scaffold highlighting (the \u201cLigand structures\u201d grid)\n\nSeparately, the dashboard\u2019s ligand grid tries to answer a different question: *do any of these ligands share a chemical core?* This matters most for SAR (structure-activity relationship) campaigns, where a chemist is usually testing close analogues on purpose, and seeing the shared scaffold at a glance (with the parts that differ jumping out) is more useful than reading each SMILES individually. Two tiers, in order, and nothing is highlighted unless one of them actually finds something real:\n\n1. **Exact Bemis-Murcko scaffold match** \u2014 ligands whose ring systems + connecting linkers are chemically identical are grouped, threshold-free. This is the dominant case for a real SAR series.\n2. **Fallback for near-analogues:** ligands left over are grouped by Morgan/Tanimoto fingerprint similarity, then a maximum common substructure (MCS) is computed across the *whole* group and verified to actually match every member \u2014 so the claim is a proven substructure match, not just an assigned similarity score.\n\nSmall or trivial shared fragments (below 8 heavy atoms \u2014 e.g. \u201cthey all contain a benzene ring\u201d) are deliberately not highlighted; a group has to share something structurally meaningful to be called out. Ligands in the same scaffold group also have their 2D depiction aligned to a common orientation, so the shared core is drawn in the same position across cells and visually \u201csnaps together.\u201d\n\n**What\u2019s highlighted and how**, directly on each rendered structure \u2014 the same badges shown on each ligand cell and spelled out in the panel\u2019s own legend:\n\n| Badge | Colour | Meaning |\n| --- | --- | --- |\n| `S` | \ud83d\udfea Magenta | Undefined stereocentre (RDKit also draws its own `(?)` marker at the atom) |\n| `A` | \ud83d\udfe7 Amber | Carboxylic acid \u2014 protonation state not specified |\n| `N` | \ud83d\udfe7 Amber | Primary/secondary amine \u2014 protonation state not specified |\n| `Ph` | \ud83d\udfe7 Amber | Phenol \u2014 protonation state not specified |\n| `SO3` | \ud83d\udfe7 Amber | Sulfonic acid \u2014 protonation state not specified |\n| `salt` | \ud83d\udfe5 Red | Disconnected fragment (salt/counterion) \u2014 flagged on the border and badge only, not atom-highlighted (there\u2019s no single meaningful atom to point at) |\n| \u2014 | one of six colour-blind-safe palette colours | Atoms in a shared scaffold/substructure \u2014 consistent per group, with a legend entry naming the group and how many ligands share it (e.g. \u201cshared scaffold \u2014 3/5 ligands\u201d) |\n\nA specific finding (stereocentre, ionizable group) always takes priority over the softer scaffold highlight if they overlap on the same atom, since it\u2019s the more actionable signal. If no ligand shares a real scaffold with any other, the panel says so plainly (\u201cno shared scaffold or substructure detected\u201d) rather than forcing a highlight onto something coincidental. CCD-code ligands have no SMILES to render and show a plain placeholder instead of an empty cell.\n\n## \ud83e\uddec compare-sse: apo vs holo secondary-structure shifts\n\n**Why this exists:** a confidence score tells you *what* Boltz predicted, not *how the protein moved* in response to ligand binding \u2014 a real structural question whenever you have both a reference apo (unbound) structure and a predicted holo one for the same protein. `compare-sse` answers it in terms a structural biologist actually reasons in (\u201cTM6 swung out 4.2 Angstrom\u201d, \u201cthe DFG motif flipped from in to out\u201d), not raw DSSP fragment coordinates.", "# Outrunning Diffusion: The Fast Lane in CryoEM Grid Preparation\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-25 on marcdeller.com (categories: blog, sb).*\n\n* [Outrunning protein diffusion to the air\u2013water interface in cryoEM](https://www.pnas.org/doi/10.1073/pnas.2516900122)\n\n* **Hyperquenching**: Nanosecond vitrification seals proteins away from [damaging air\u2013water interface.](https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cmtd.202200050)\n* **Grid Technology**: High-speed, blotless devices [minimize denaturation, preferred orientation, and manual error](https://www.nanoimagingservices.com/about/blog/a-guide-to-cryo-grid-production-materials-interfaces-and-their-impact-on-sample-prep).\n* **Material Innovation**: [Hydrophilized graphene supports enable gentle adsorption and reduce AWI damage](https://elifesciences.org/articles/42747).\n* **Crystal Coating**: Engineered grid surfaces, like hydrophobin films, can [actively sequester proteins from the AWI](https://elifesciences.org/articles/42747).\n* **Future Outlook**: Integration of analytics, automation, and [strategic sample prep fuels the cryoEM resolution revolution](https://www.pnas.org/doi/10.1073/pnas.2516900122).\n\n1. Outrunning protein diffusion to the air\u2013water interface in cryoEM: Rubino, et al., PNAS, 2022 \u2013 [Rubino, et al., PNAS, 2022](https://www.pnas.org/doi/10.1073/pnas.2516900122)\n2. Large Temperature\u2010Jump and Nanosecond Hyperquenching for Time\u2010Resolved Freeze Trapping: Cherepanov, et al., Chemistry Europe, 2023 \u2013 [Cherepanov, et al., Chemistry Europe, 2023](https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cmtd.202200050)\n3. Protein denaturation at the air-water interface and how to prevent it: D\u2019Imprima, et al., eLife, 2019 \u2013 [D\u2019Imprima, et al., eLife, 2019](https://elifesciences.org/articles/42747)\n4. Recent Technical Advances in Sample Preparation for Cryo-EM: Xu, et al., Frontiers in Molecular Biosciences, 2022 \u2013 [Xu, et al., Frontiers in Molecular Biosciences, 2022](https://www.frontiersin.org/journals/molecular-biosciences/articles/10.3389/fmolb.2022.892459/full)\n5. A Guide to Cryo-Grid Production: Materials, Interfaces and their Impact on Sample Prep: NanoImaging Services Blog, 2024 \u2013 [NanoImaging Services Blog, 2024](https://www.nanoimagingservices.com/about/blog/a-guide-to-cryo-grid-production-materials-interfaces-and-their-impact-on-sample-prep)\n\n**Breaking the speed limit in cryoEM sample preparation promises a revolution\u2014but is it truly possible to outrun the relentless race of protein diffusion towards the air\u2013water interface?**\n\nCryo-electron microscopy (cryoEM) has become a driving force in structural biology, enabling scientists to unlock atomic-level secrets of macromolecules. The challenge: proteins in solution rapidly migrate to and interact with the air\u2013water interface (AWI) during grid preparation, causing denaturation, preferred orientations, and complex dissociation. The question at the heart of modern cryoEM research is whether innovation in grid preparation can truly prevent these damaging interactions by getting ahead of diffusion timelines.\n\nProteins diffuse astonishingly quickly. In a standard cryoEM grid, most particles reach the AWI well within a millisecond, with smaller sample films facilitating even faster migration\u2014for a 100 kDa protein molecule, migration across a 50 nm layer happens in microseconds ([PNAS](https://www.pnas.org/doi/10.1073/pnas.2516900122)). As a result, the window for outrunning diffusion during grid prep is incredibly narrow. Blotting, sample thinning, and even manual freezing often operate on slower timescales than the physical processes at play. The introduction of hyperquenching and advanced automated grid preparation attempts to tilt the odds.\n\nNanosecond hyperquenching has emerged as a promising technology, using ultrafast jets and instant vitrification to immobilize protein particles before they encounter the AWI, minimizing denaturation and loss of native structure ([Chemistry Europe](https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cmtd.202200050)). This technique forms a unique glass-like layer that essentially \u201cseals\u201d proteins inside, precluding their contact with the damaging interface\u2014a feat achieved over mere picoseconds. Similarly, high-speed, blotless grid devices like the chameleon accelerate dispense-to-plunge times to tens of milliseconds, limiting exposure and preferred orientation problems ([NanoImaging Services](https://www.nanoimagingservices.com/about/blog/a-guide-to-cryo-grid-production-materials-interfaces-and-their-impact-on-sample-prep)). These rapid workflows combine robotics, self-wicking grids, and real-time ice quality monitoring for reproducibly thin, high-quality films\u2014a significant advance from traditional methods.\n\nYet, it\u2019s not just speed that counts. Material engineering, particularly the use of hydrophilized graphene-coated grids, has proven key in mitigating AWI damage ([eLife](https://elifesciences.org/articles/42747)). Graphene\u2019s atomically-thin, hydrophilic surface allows proteins to adsorb in a protective environment, greatly reducing denaturation and aggregation. Automated systems deploying such supports enable fine-tuning of sample conditions and improve the likelihood of achieving high-resolution reconstructions. Some engineered support films, like those using hydrophobin crystals, can effectively *sequester*particles away from the AWI in a crystalline monolayer ([Frontiers in Molecular Biosciences](https://www.frontiersin.org/journals/molecular-biosciences/articles/10.3389/fmolb.2022.892459/full)).", "# Elora Therapeutics Investor Pitch Deck\nAuthors: Swartzendruber, Paul (Co-Founder & CEO); Deller, Marc C. (Co-Founder & CSO)\nSource: Elora Therapeutics (investor pitch deck), 2025\n\nElora has two paths to revenue: license and option-to-buy by a pharma partner, or FDA approval and commercialization.\n\nPath A \u2014 License & Option-to-Buy: a pharma partner pays for license rights and the Phase 1b package, comprising $25\u2013100M upfront plus $500M\u2013$1.5B in milestones and 6\u201310% royalties, with an option-to-buy after Phase 2a.\n\nPath B \u2014 FDA Approval & Commercialization: payers pay for an insurance-reimbursable enzyme therapy, generating $1\u20135B in annual net sales for Elora.\n\nKey sources: BIO Industry Analysis; public company deal disclosures; selected partnering comps.\n\n## Slide 10 \u2014 Team & Advisors\n\nElora's team and advisors bring the experience to translate enzyme science into a clinical-grade therapeutic.\n\n## Slide 11 \u2014 Raise and Use of Funds\n\nElora is raising its first $275K to deliver in vitro validation, in vivo proof-of-concept, and IP expansion.\n\nThe use of funds breaks down as: In Vivo mouse study \u2013 $200K (73%); In Vitro validation \u2013 $50K (18%); and IP expansion \u2013 $25K (9%).\n\nThe key milestones are: (1) In Vitro \u2014 validate engineered PETase/MHETase activity, stability, and safety in simulated human serum; (2) In Vivo \u2014 demonstrate preclinical proof of concept in a mouse model with reduced PET burden and improved biomarkers; and (3) IP Expansion \u2014 expanded IP around lead selection, formulation, and human compatibility. The timeline runs from February 2026 to October 2026.\n\n## Slide 12 \u2014 Thank You\n\nPaul Swartzendruber, Co-Founder & CEO \u2014 paul@eloratherapeutics.com. Marc Deller, D.Phil, Co-Founder & CSO \u2014 marc@eloratherapeutics.com. Elora Therapeutics is partnered with the Austin Technology Incubator (ATI). Website: https://eloratherapeutics.com.", "# 2017 \u2013 Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: papers).*\n\n* [Link to Paper](https://marcdeller.com/wp-content/uploads/2025/09/qt72v3f2jt.pdf)\n\n* **Map Fit:** Ensure each ligand atom is supported by clear electron density.\n* **Stereochemistry:** Validate bond geometry against high-quality small-molecule libraries.\n* **Chemical Plausibility:** Cross-check interactions against known SAR data.\n* **Automated Tools:** Use RSA, MolProbity, and Privateer for rapid, standardized validation.\n* **Manual Inspection:** Refine ligand placement iteratively in graphics software.\n* **Impact on Drug Design:** Reliable models underpin accurate docking and free-energy predictions.\n\n1. Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands: [Pozharski, E., et al., Methods Mol Biol](https://link.springer.com/protocol/10.1007/978-1-4939-7000-1_25)\n2. MolProbity: all-atom structure validation for macromolecular crystallography: [Chen, V. B., et al., Acta Crystallogr D Biol Crystallogr](https://doi.org/10.1107/S0907444909042073)\n3. Real-space correlation analysis in crystallographic model validation: [Williams, C. J., et al., IUCrJ](https://doi.org/10.1107/S2052252515007019)\n4. Privateer: software for validation of carbohydrate structures: [Agirre, J., et al., Nat Commun](https://doi.org/10.1038/ncomms9645)\n5. Enhanced Validation of Small-Molecule Ligands and Carbohydrates in the Protein Data Bank: [Feng, Z., et al., Structure](https://doi.org/10.1016/j.str.2021.02.004)\n6. A workflow to create a high-quality protein\u2013ligand binding dataset with binding affinity and experimental details: [Zhang, H., et al., Drug Discov Today](https://doi.org/10.1016/j.drudis.2025.104257)\n\nAccurate validation of protein\u2013ligand crystal structures demands rigorous assessment of electron density, stereochemistry, and chemical plausibility to ensure reliable models for structure-guided drug design.\n\nThe determination of protein\u2013ligand complex structures by X-ray crystallography underpins modern drug discovery. Because ligands contribute only a small fraction of total scattering mass, validating their atomic models is critical for confidence in binding interactions. First, experimental electron-density maps must unequivocally support ligand placement. Near-atomic resolution and robust map\u2013model correlation ensure the ligand is correctly localized within the binding pocket. Second, the ligand geometry must adhere to established stereochemical parameters\u2014bond lengths, angles, and torsion angles are compared against small-molecule libraries to detect outliers. Third, the biochemical plausibility of observed interactions\u2014hydrogen bonds, hydrophobic contacts, and metal coordination\u2014must be inspected in the context of known binding motifs and structure\u2013activity relationship (SAR) data.\n\nValidation tools such as RSA (Real-space correlation analysis), MolProbity, and Privateer facilitate automated checks. RSA quantifies map\u2013model agreement for each ligand atom, highlighting regions of ambiguity. MolProbity assesses stereochemistry and clashes, while Privateer specializes in carbohydrate ligands. Researchers should complement automated checks with manual inspection in molecular graphics software, adjusting ligand torsions or fitting alternative conformers until both map support and chemical rationale converge.\n\nApplying these validation steps minimizes false-positive binding modes, prevents propagation of erroneous models in databases, and enhances the reliability of downstream computational studies\u2014docking, molecular dynamics, and free-energy calculations.\n\n| Key Concept | Description | Reference |\n| --- | --- | --- |\n| Map Fit | Correlation of ligand atoms with electron density maps. | [Pozharski, et al.](https://link.springer.com/protocol/10.1007/978-1-4939-7000-1_25) |\n| Stereochemistry | Assessment of bond lengths, angles, and torsions. | [Chen, et al.](https://doi.org/10.1107/S0907444909042073) |\n| Chemical Plausibility | Inspection of binding interactions against SAR data. | [Krojer, et al.](https://pubmed.ncbi.nlm.nih.gov/28012432/) |\n| RSA | Quantitative real-space correlation analysis. | [Williams, et al.](https://doi.org/10.1107/S2052252515007019) |\n| MolProbity | All-atom steric and geometry validation. | [Chen, et al.](https://doi.org/10.1107/S0907444909042073) |\n| Privateer | Specialized carbohydrate ligand validation. | [Agirre, et al.](https://doi.org/10.1038/ncomms9645) |", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\nAround 30 side chains from each protein make contact and mutagenesis of contact residues in hGHR, have revealed that a central hydrophobic region dominated by two tryptophan residues, accounts for more than three-quarters of the binding free energy (Clackson and Wells, 1995). It has been shown that both site 1 and site 2 of the hormone interact with the same part of the hGHR molecule (see Figure 1-5). This region consists of around 11 residues with Trp 104 and Trp 169 (one from each fibronectin type III-like domain) contributing most of the binding energy through hydrophobic burial. Trp 104 has been shown to fit into a pocket of the ligand produced by Gly 120 and mutation of this residue on the ligand sterically blocks the formation of a 1:2 complex (Clackson et al., 1998). The same two residues plus Asn 128 of the receptor are important in site 2 interactions. These functional epitopes lie on loops extending out of the \u03b2-sheet, with Trp 104 residing on the EF loop of the N-terminal domain and Trp 169 residing on the BC loop of the C-terminal domain (see section 1.4.2.1.2 and Figure 1-6). Asn 128 is positioned in the helical linker region between the two domains. These loops have been shown to have similar main-chain conformations in both hGH-hGHR and hPrl-hPrlR complexes (Somers et al., 1994) (see Figure 1-6).\n\nAnalysis of the binding determinants in the hGH-hGHR complex has revealed a pattern in which energetically important residues form dominant hydrophobic patches and the less important hydrophilic residues surround this region. It has been proposed that the hydrophobic core of these ligand-binding surfaces drives the interaction entropically, through Van der Waals contact and the associated exclusion of water from the hydrophobic surfaces. In contrast, the purpose of the peripheral hydrophilic residues is to provide the fidelity for the interaction through specific hydrogen bonding and salt bridge formation. In contrast, a reverse of this pattern has been reported from analysis of the crystal structure of the IL4-IL4R\u03b1 complex (Hage et al., 1999), in which Asp 72 of the receptor forms a core hydrophilic element on the receptor which is surrounded by a shell of hydrophobic residues. This residue forms a salt bridge with Arg 88 of the ligand, which itself is surrounded by a plethora of hydrophobic residues. The inner core of hydrophilic residues is thought to provide specificity to the binding and the positioning of Asp 72 in the hydrophobic environment enhances both hydrogen and ionic bonding potentials. The hydrophobic shell again provides the entropic driving force for the interaction. Asp 72 of the IL4R\u03b1 is the major determinant of binding and as such can be compared with Trp 104 of hGHR.\n\n##### 1.4.2.1.2 Structure-Function Relationship of Prolactin\n\nThe hPrlR binds both hGH and hPrl. The crystal structure of hPrlR bound to hGH, in 1:1 stoichiometry, has been solved and this structure shows homology to the hGH receptor-ligand complex (Cunningham et al., 1990; Somers et al., 1994) (see Figure 1-3). As in hGHR the WSXWS box is located in the FG loop of the C-terminal domain. The hGH binds to the PrlR at site 1 in a similar fashion to that observed in the hGH-hGHR system, with residues from the AB loop and the D helix participating in the interface. Similarly to hGHR the hPrlR also displays two conserved tryptophan residues; Trp 72 in the EF loop (analogous to Trp 104 of hGHR) and Trp 139 in the BC loop (analogous to Trp 169 of hGHR). Although hGH binds to both the hGHR and hPrlR, hPrl does not bind to the hGHR and mutagenesis studies have shown that the binding sites for both receptors on hGH overlap, with Arg 167 and Glu 174 being key specificity determinants (Somers et al., 1994). It has also been shown that Glu 174 and His 18 of hGH function in chelating Zn\u00b2\u207a ions along with Asp 217 and His 218 of the receptor. This contributes a considerable enthalpy to the binding reaction, increasing the affinity of hGH for hPrlR by upto 8000-fold (Somers et al., 1994).\n\n##### 1.4.2.1.3 Structure-Function Relationship of Erythropoetin\n\nSeveral structures of EpoR have been determined including Epo bound (Syed et al., 1998), peptide agonist bound (Livnah et al., 1996), peptide antagonist bound (Livnah et al., 1998) and ligand free receptor (Livnah et al., 1999). These structures show general homology to the hGH-hGHR complex (see Figure 1-3). The structure of EpoR has been determined in complex with a dimeric 20 residue cyclic peptide unrelated in sequence to Epo, termed EMP1 (Epo mimetic peptide 1; Livnah et al., 1996; Wrighton et al., 1996). EMP1 is capable of inducing an erythropoetic response and the signaling pathways activated appear to be identical to those activated by native Epo (Wrighton et al., 1996). Signaling by the mimetic shows ~5% the efficiency of the natural ligand (Johnson et al., 1998). This mimetic dimerises EpoR using a smaller set of residues, as compared to the hGH-hGHR system, again with key aromatic residues (Phe 93 and Phe 205) forming the functional epitope (Middleton et al., 1999). Comparison of the EMP1-EpoR complex with the hGH-hGHR and Epo-EpoR complexes has suggested that the efficiency of signaling through these receptors is dependent on the receptor orientation (Syed et al., 1998). In the EMP1-EpoR structure the membrane proximal haemopoietin domains are twisted away from each other by ~45\u00b0, whereas these two domains in all other haemopoietin receptor-ligand complexes lie in the same plane. It has been proposed that this shift is responsible for the reduced signaling efficiency of EMP1, as the intracellular domains are not oriented optimally to allow signaling through the JAK pathway (see 1.4.3) (Syed et al., 1998).", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices\nAuthors: Deller, M.C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nTherefore, it can be concluded that of the 5 residues identified as the site 2 motif of OSM through mutagenesis studies, all are structurally equivalent to residues directly involved in the Epo-EpoR interface. OSM residues Gln 20, Gly 120, Asn 123 and Asn 124 are all equivalent to the series of residues in Epo that pack against the EpoR residue Phe 93. The fifth member of OSM site 2 is Gln 16, which is equivalent to an Epo residue involved in one of the hydrogen bonds in the receptor-ligand interface. The EpoR residue Phe 93 is in turn structurally and functionally equivalent to the gp130 residue Phe 169 which resides on the E-F loop of the N-terminal domain.\n\nThe two remaining major determinants in EpoR, Met 150 and Phe 205 (Middleton et al., 1999) have no structural equivalents although mutation of Val 230 in the B-C loop has been shown to affect binding in the IL-6 system (Horsten et al., 1997). However, this residue is relatively buried so suggesting indirect effects on ligand binding. Searches for other regions of binding activity in gp130 have revealed that the solvent exposed residues Gly 286 and Lys 285 of the F-G loop also function in binding IL-6 (Horsten et al., 1997). The F-G loop region has also been confirmed as a binding epitope in the GCSF system as discussed below (Yamasaki et al., 1997). Thus, it appears that the gp130 system has diverged from the Epo related systems in that only one aromatic residue is involved in binding compared with two in Epo. This may be to allow greater conformational flexibility so allowing formation of higher order receptor-ligand signalling complexes such as the hexameric assemblies observed for IL-6 (Ward et al., 1996) and CNTF (De Serio et al., 1995), the tetrameric assemblies of OSM (Staunton et al., 1998) (see Figure 6-11) and the trimeric assemblies of LIF (Zhang et al., 1997b).\n\n#### 6.3.5 Granulocyte Colony Stimulating Factor\n\nThe GCSF-GCSFR system (Aritomi et al., 1999; Hill et al., 1993; GCSF-GCSFR coordinates kindly provided by K. Morikawa, Biomedical Engineering Research Council, Osaka, Japan; see Figure 6-11 (b)) provides an attractive model for comparison with the OSM and gp130 system for several reasons. Firstly, the ligand and receptor components of this system show high levels of structural similarity to OSM and gp130, respectively, as assessed by the RMSD values as shown in Table 6-1 and Table 6-2. Particularly, as discussed in section 6.2.1, the positioning and orientation of the A-B loop in GCSF more closely matches that of OSM than the Epo structure. Secondly, the Phe/Lys pair as utilised by the LIFR binding cytokines is conserved, as shown in Figure 6-12 below, and these two residues have also been shown to be critically important for binding GCSFR (Reidhaar-Olson et al., 1996). Thirdly, mutations in GCSFR have revealed that residues in the B-C and F-G loops of the C-terminal domain are critical for ligand binding, in a similar fashion to gp130 and in contrast to the use of the B-C and E-F loops in Epo. Finally, as seen with gp130, GCSF has a propensity for forming higher order complexes with stoichiometries of 1:1, 2:2 and 4:4 having been reported depending on the experiment performed (Hiraoka et al., 1995; Hiraoka et al., 1994; Horan et al., 1996; Horan et al., 1998).\n\nHowever, one major draw back of the comparison is the lack of site 1 and site 3 contacts in the current crystal structure of the GCSF-GCSFR complex (Aritomi et al., 1999) as shown in Figure 6-11 below. The two contact regions are described in terms of a major site, which is equivalent to site 2, and a minor site that displays novel interactions. Therefore, this system can only be used for modelling of site 2 interactions with gp130.\n\n##### 6.3.5.1 Major Interface\n\nThe current model of GCSF-GCSFR interaction involves a 2:2 complex consisting of two 1:1 complexes of GCSF and GCSFR related by a non-crystallographic pseudo-twofold axis. The 1:1 complex is formed by binding of the hinge region of the receptor to an epitope equivalent to site 2 on the ligand. The core of this interaction involves hydrogen bonding of Glu 19 of GCSF to Arg 288 of GCSFR; both of these residues have previously shown to be important through mutagenesis (Layton et al., 1997; Layton et al., 1999). Lys 19 is the structurally equivalent residue in OSM and although this residue is ~80 % solvent exposed alanine mutation showed no effect on either gp130 or LIFR binding. However, mutation of adjacent residues Gln 16 and Gln 20, which show much lower solvent accessibilities, affect gp130 binding by ~75-fold or greater as detailed in Table 6-3 below.\n\nIn contrast to the Epo-EpoR interface which is predominantly hydrophobic in character the GCSF-GCSFR major interface displays an even spread of hydrophobic and polar interactions in a similar fashion to the mosaic interface observed in the IL-4-IL-4R complex (Hage et al., 1998). This again compounds the similarity to the OSM system with the present mutagenesis data suggesting that this interface maps to a surface of mixed character.\n\nMutational analysis has revealed that in contrast to EpoR the major interface determinants of GCSFR reside on the B-C and F-G loops of the C-terminal domain and involve residues Ile 239 and Arg 288, respectively (Layton et al., 1997). Structurally equivalent residues to Ile 239 and Arg 288 of GCSFR are Ile 231 and Lys 285 of gp130, respectively, which demonstrates conservation between the two receptors, although the role of these two residues in gp130 is currently unknown. As discussed previously mutagenesis of Val 230 and Lys 285 resulted in loss of affinity between gp130 and IL-6 (Horsten et al., 1997) so implying that activity resides around these residues and so suggesting a similar mode of receptor engagement between the two systems.\n\n##### 6.3.5.2 Minor Interface", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\nIn some embodiments, the anti-mutCALR antibody inhibits one or more signaling pathways downstream of thrombopoietin receptor (MPL) in a cell expressing human mutant CALR; inhibits oncogenic cell proliferation in a cell expressing human mutant CALR; and/or inhibits dimerization of MPL in a cell expressing human mutant CALR. In some embodiments, the anti-mutCALR antibody inhibits one or more signaling pathways downstream of MPL in both a first cell expressing human Type 1 mutant CALR and in a second cell expressing human Type 2 mutant CALR; inhibits oncogenic cell proliferation in both a first cell expressing human Type 1 mutant CALR and in a second cell expressing human Type 2 mutant CALR; and/or inhibits dimerization of MPL in both a first cell expressing human Type 1 mutant CALR and in a second cell expressing human Type 2 mutant CALR. In some embodiments, the one or more signaling pathways downstream of MPL are selected from the group consisting of Janus tyrosine kinase (JAK) and signal transducers and activators of transcription (STAT) signaling, mitogen-activated protein kinase (MEK) and extracellular signal-regulated kinase (ERK) signaling, serine/threonine kinase (AKT) signaling, and mammalian target of rapamycin (mTOR) signaling.\n\nIn some embodiments, the anti-mutCALR antibody has modulated Fc effector function. In some embodiments, the modulated Fc effector function is increased Fc effector function or reduced Fc effector function. In some embodiments, the anti-mutCALR antibody has reduced Fc effector function. In some embodiments, the Fc effector function is antibody-dependent cell-mediated cytotoxicity (ADCC), complement dependent cytotoxicity (CDC), or antibody dependent cellular phagocytosis (ADCP). In some embodiments, the anti-mutCALR antibody having reduced Fc effector function has increased binding affinity to human mutant CALR as compared to an antibody without reduced Fc effector function. In some embodiments, the Fc effector function is ADCC.\n\nIn some embodiments, the anti-mutCALR antibody is a human or humanized antibody. In some embodiments, the anti-mutCALR antibody is a full-length antibody. In some embodiments, the anti-mutCALR antibody is an IgG1, IgG2, IgG3 or IgG4 antibody. In some embodiments, the anti-mutCALR antibody is an IgG1 antibody. In some embodiments, the anti-mutCALR antibody is a bispecific antibody, a biparatopic antibody, a single chain antibody (scFv), an Fab fragment, an F(ab')2 fragment, an Fab' fragment, an Fsc fragment, an Fv fragment, an scFv, an sc(Fv)2, or a diabody. In some embodiments, the anti-mutCALR antibody is a biparatopic antibody comprising two heavy chain-light chain pairs or one heavy chain-light chain pair. In some embodiments, the biparatopic antibody is a full-length antibody. In some embodiments, the biparatopic antibody comprises one heavy chain-light chain pair. In some embodiments, the anti-mutCALR antibody is conjugated to a toxic substance. In some embodiments, the toxic substance is a radioisotope or a cytotoxic agent.\n\nAspects of the present disclosure provide a nucleic acid or a set of nucleic acids, which collectively encodes any one of the anti-mutCALR antibodies described herein. Aspects of the present disclosure provide an expression vector or a set of expression vectors comprising the nucleic acid or the set of nucleic acids encoding any one of the anti-mutCALR antibodies described herein operably linked to a promoter. Aspects of the present disclosure provide an isolated cell comprising the nucleic acid or the set of nucleic acids encoding any one of the anti-mutCALR antibodies or the expression vector or the set of expression vectors. Aspects of the present disclosure provide a method of making the anti-mutCALR antibody, comprising culturing the cell described herein and isolating the antibody. Aspects of the present disclosure provide a pharmaceutical composition comprising the anti-mutCALR antibody, the nucleic acid or the set of nucleic acids, the expression vector or the set of expression vectors, or the isolated cell, and a pharmaceutically acceptable carrier.\n\nAspects of the present disclosure provide a method of treating a myeloproliferative neoplasm in a human subject in need thereof, the method comprising administering to the human subject an effective amount of an anti-mutCALR antibody described herein or the pharmaceutical composition thereof. Other aspects of the present disclosure provide an anti-mutCALR antibody described herein for use in the treatment of a myeloproliferative neoplasm, or the use of an anti-mutCALR antibody described herein for the manufacture of a medicament for the treatment of a myeloproliferative neoplasm.\n\nAspects of the present disclosure provide a method of detecting a CALR exon 9 mutation in a biological sample, the method comprising obtaining a biological sample from a human subject who has or is suspected of having a myeloproliferative neoplasm and contacting the sample with an anti-mutCALR antibody described herein such that the anti-mutCALR antibody binds to a mutCALR protein if the mutCALR protein is present in the biological sample. Another aspect of the present disclosure provides a method of diagnosing a human subject with a myeloproliferative neoplasm, the method comprising obtaining a biological sample from a human subject who has or is suspected of having a myeloproliferative neoplasm and contacting the sample with an anti-mutCALR antibody described herein such that the anti-mutCALR antibody binds to a mutCALR protein if the mutCALR protein is present in the biological sample.", "# How AI Is Revealing the Secret Pathways of Protein Folding\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: blog).*\n\n* Building on structure prediction tools like AlphaFold, AI is now being used to map the entire folding pathway of proteins, not just the final structure.\n* AI accelerates traditional molecular dynamics simulations, making it possible to model the complex and rapid journey of a protein as it folds.\n* This reveals crucial intermediate states and helps predict how mutations can lead to misfolding, providing new insights for drug discovery.\n\n1. AI for Unfolding the Protein Folding Pathways: <https://www.linkedin.com/pulse/ai-unfolding-protein-folding-pathways-boltzmannlabs-k6bac>\n2. Folding the Future: How AI is Reshaping Protein Engineering: <https://www.synbiobeta.com/read/folding-the-future-how-ai-is-reshaping-protein-engineering>\n\nFor 50 years, predicting the final 3D structure of a protein from its amino acid sequence was one of biology\u2019s grand challenges. The breakthrough of AI models like AlphaFold largely solved this problem, giving us an unprecedented look at the finished product. But this triumph raised an even more profound question: if we know the destination, can we now map the journey? The function of a protein is determined not just by its final shape, but by the complex, high-speed journey it takes to get there\u2014its folding pathway.\n\nUnderstanding these pathways has been incredibly difficult. The process involves a dizzying number of fleeting, transient intermediate shapes, and simulating every atomic movement is computationally prohibitive, even for supercomputers. This is where the next wave of AI is making its mark. Scientists are now combining the predictive power of AI with molecular dynamics (MD) simulations to chart these once-invisible routes. Instead of brute-forcing the simulation, AI models can learn the patterns of folding from shorter simulations and intelligently predict the most likely next steps.\n\nThis AI-accelerated approach allows researchers to model entire folding pathways, identify key intermediate states, and even predict how a genetic mutation might alter the folding route, potentially leading to disease. We are moving beyond a static picture of proteins and into the realm of protein dynamics. It\u2019s the difference between having a photograph of a car and a GPS animation of its entire journey, including every turn and traffic jam. This deeper understanding is essential for designing drugs that can stabilize correctly folded proteins or correct misfolding pathways, opening up new strategies to combat diseases rooted in protein misfolding.", "# 2025 \u2013 Strengthen Your Decision-Making with Generative AI\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-28 on marcdeller.com (categories: certificates).*\n\n* [Link to certificate](https://www.linkedin.com/learning/certificates/55ccac83c777ec51ceedf007ef7dfdd4ac54fd988b23eb1497d0732b93c48c12)\n\n* **Award Details:** LinkedIn Learning issued this certificate on March 25, 2025 at 08:38 PM UTC.\n* **Course Focus:** It recognizes completion of a course on improving decision-making with generative AI.\n* **Career Value:** This matters because scientists often need to compare evidence, options, and uncertainty efficiently.\n* **Practical Use:** It helps with prioritizing experiments, evaluating alternatives, and structuring research choices.", "# Extended Portfolio \u2014 Marc C. Deller, D.Phil.\nAuthors: Deller, M.C.\nSource: Personal portfolio (merged from two versions), 2025\n\nStructural tour de force: Illuminates HIV-1's glycan shield dynamics by resolving \u2014 via high-resolution crystallography of the BG505 SOSIP trimer with 8ANC195/PGT128 \u2014 how bNAbs exploit gp41 interactions (N637 glycan accommodation) and glycan domino effects (N301 repositioning N262), unveiling allosteric glycan manipulation and completing trimer-level epitope blueprints for precision vaccine engineering.\n\n## Most Cited Publications \u2014 with commentary\n\n### Crystal structure of a soluble cleaved HIV-1 envelope trimer\n\n**Journal:** Science, 2013 (979 citations). **Authors:** J.P. Julien, A. Cupo, D. Sok, R.L. Stanfield, D. Lyumkis, M.C. Deller et al. **Link:** https://www.science.org/doi/abs/10.1126/science.1245625\n\nExciting paper: Presents a high-resolution crystal structure of a near-native, cleaved HIV-1 envelope trimer in complex with a broadly neutralizing antibody, revealing unprecedented molecular details of the trimer's architecture and its vulnerable sites \u2014 insights that are crucial for both understanding how HIV-1 infects cells and advancing rational vaccine design.\n\n### Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\n\n**Journal:** Nature Structural & Molecular Biology, 2013 (420 citations). **Authors:** Leopold Kong et al. **Link:** https://www.nature.com/articles/nsmb.2594\n\nThis study reveals how the HIV-1 glycan shield's Asn332-dependent supersite \u2014 a densely glycosylated region on gp120 \u2014 can be penetrated by diverse antibodies like PGT135, which uses elongated CDR loops to bypass glycans and target vulnerable protein surfaces, offering a blueprint for designing vaccines that exploit this Achilles' heel of the virus.\n\n### Protein stability: a crystallographer's perspective\n\n**Journal:** Structural Biology and Crystallization Communications, 2016 (370 citations). **Authors:** Marc C. Deller, Leopold Kong, Bernhard Rupp. **Link:** https://journals.iucr.org/f/issues/2016/02/00/en5571/index.html\n\nFeatured as a \"Top article in structural biology (Spring 2020)\": Details a practical, crystallographer-focused discussion on protein stability, the understanding of which is essential for optimizing critical processes like protein expression, purification, formulation, storage, and structural studies across the biotechnology, pharmaceutical, and academic sectors.\n\n### The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\n\n**Journal:** ACS Chemical Biology, 2017 (130 citations). **Authors:** Elise A. Span, Daniel L.M. Suess, Marc C. Deller, R. David Britt, Michael A. Marletta. **Link:** https://pubs.acs.org/doi/abs/10.1021/acschembio.7b00016\n\nThis study elucidates critical hydrogen-bonding networks in fungal polysaccharide monooxygenases (PMOs) that govern oxygen activation and proton transfer \u2014 key mechanistic insights for optimizing enzymatic cellulose degradation, with transformative potential for biofuel production and sustainable biomass utilization.\n\n### Models of protein\u2013ligand crystal structures: trust, but verify\n\n**Journal:** Journal of Computer-Aided Molecular Design, 2015 (113 citations). **Authors:** Marc C. Deller, Bernhard Rupp. **Link:** https://link.springer.com/article/10.1007/s10822-015-9833-8\n\nThis paper is crucial because it establishes metrics underscoring why X-ray crystallography is the gold standard for creating precise protein\u2013ligand models \u2014 essential for advancing drug design and computational biology \u2014 while advocating rigorous validation to combat overinterpretation and ensure reliability in scientific research.\n\n## Further Publications\n\n### Crystallisation of Proteins and Macromolecular Complexes: Past, Present and Future\n\n**Venue:** eLS, 2017. **Authors:** M.C. Deller, B. Rupp.\n\n### Crystal structure of a two-subunit TrkA octameric gating ring assembly\n\n**Journal:** PLoS ONE 10 (3), e0122512, 2015. **Authors:** M.C. Deller, H.A. Johnson, M.D. Miller, G. Spraggon, M.A. Elsliger, I.A. Wilson, \u2026\n\n### Structure-based discovery of NANOG variant with enhanced properties to promote self-renewal and reprogramming of pluripotent stem cells\n\n**Journal:** Proceedings of the National Academy of Sciences 112 (15), 4666\u20134671, 2015. **Authors:** Y. Hayashi, L. Caboni, D. Das, F. Yumoto, T. Clayton, M.C. Deller, P. Nguyen, \u2026\n\n### Approaches to automated protein crystal harvesting\n\n**Journal:** Structural Biology and Crystallization Communications 70 (2), 133\u2013155, 2014. **Authors:** M.C. Deller, B. Rupp.\n\n### Structural insights into the recognition of phosphopeptide by the FHA domain of kanadaptin\n\n**Journal:** PLoS ONE 9 (9), e107309, 2014. **Authors:** Q. Xu, M.C. Deller, T.K. Nielsen, J.C. Grant, S.A. Lesley, M.A. Elsliger, \u2026\n\n### Crystal Structure of the First Eubacterial Mre11 Nuclease Reveals Novel Features that May Discriminate Substrates During DNA Repair\n\n**Journal:** J. Mol. Biol. 397: 647, 2012. **Authors:** D. Das, D. Moiani, H.L. Axelrod, M.D. Miller, D. McMullan, K.K. Jin, P. Abdubek, \u2026\n\n### Structure of hepatitis C virus envelope glycoprotein E2 antigenic site 412 to 423 in complex with antibody AP33\n\n**Journal:** Journal of Virology 86 (23), 13085\u201313088, 2012. **Authors:** L. Kong, E. Giang, T. Nieusma, J.B. Robbins, M.C. Deller, R.L. Stanfield, \u2026\n\n### Functional and structural characterization of a thermostable acetyl esterase from Thermotoga maritima\n\n**Journal:** Proteins: Structure, Function, and Bioinformatics 80 (6), 1545\u20131559, 2012. **Authors:** M. Levisson, G.W. Han, M.C. Deller, Q. Xu, P. Biely, S. Hendriks, L.F. Ten Eyck, \u2026\n\n### Structure of the \u03b3-D-glutamyl-L-diamino acid endopeptidase YkfC from Bacillus cereus in complex with L-Ala-\u03b3-D-Glu: insights into substrate recognition by NlpC/P60 cysteine peptidases\n\n**Journal:** Structural Biology and Crystallization Communications 66 (10), 1354\u20131364, 2010. **Authors:** Q. Xu, P. Abdubek, T. Astakhova, H.L. Axelrod, C. Bakolitsa, X. Cai, D. Carlton, \u2026", "# The Druggable Genome: Expanding the Target Landscape with Chemical Biology\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-14 on marcdeller.com (categories: blog).*\n\n* A large portion of the human proteome is considered \u201cundruggable,\u201d representing a major untapped resource for new medicines.\n* Chemical biology tools like activity-based protein profiling (ABPP) and chemoproteomics allow for the systematic mapping of functional sites and drug interactions across the entire proteome.\n* This discovery-driven approach is revealing novel drug targets and binding pockets, expanding the druggable landscape and enabling the development of first-in-class therapies.\n\n1. **Top 5 Drug Discovery Trends 2025 Driving Breakthroughs:**\n   <https://www.pelagobio.com/cetsa-drug-discovery-resources/blog/drug-discovery-trends-2025/>\n2. **New Concepts in Drug Discovery (2025):**\n   <https://www.sciencedirect.com/journal/current-opinion-in-structural-biology/special-issue/10BFRJKLFCN>\n3. **The Disruptive Impact of Structural Biology on Biopharmaceutical Innovation:**\n   <https://www.pharmasalmanac.com/articles/the-disruptive-impact-of-structural-biology-on-biopharmaceutical-innovation>\n\nFor all the incredible advances in medicine, a frustratingly large portion of the proteins encoded by the human genome remain beyond the reach of our drugs. It is estimated that over 85% of proteins are considered \u201cundruggable\u201d by conventional small molecule inhibitors, either because they lack a well-defined pocket to target or because their function is not well understood. This vast, unexplored territory represents a massive untapped opportunity for treating human disease.\n\nA new generation of chemical biology tools is finally allowing us to systematically explore this \u201cdark matter\u201d of the proteome. Techniques like activity-based protein profiling (ABPP) use reactive chemical probes to map the functional state of entire classes of enzymes directly in living systems. This can reveal novel binding sites and identify proteins whose activity is altered in disease states, pointing to new, previously unappreciated drug targets.\n\nCombined with chemoproteomics, which uses chemical probes to map which proteins a drug-like molecule interacts with across the entire proteome, these methods provide a powerful engine for expanding the druggable landscape. We are no longer limited to the well-lit, known corners of the genome. We can now systematically shine a light on the entire proteome, identifying new nodes of vulnerability for therapeutic intervention. This is a fundamental shift from a hypothesis-driven to an discovery-driven approach, unlocking a wealth of new targets and paving the way for first-in-class medicines against diseases that have long been considered intractable.", "# marcdeller.com: About Me\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n## Protein to Medicine: My Path to Impact\n\nHi, I\u2019m Marc, a structural biologist and drug discovery leader focused on the intersection of protein structure, protein production, platform development, and translational science. My work spans X-ray crystallography, Cryo-EM, discovery strategy, team leadership, and the systems needed to move difficult problems toward real outcomes. I build structure-guided discovery programs that turn complex biology into practical therapeutic decisions. Today, that work runs through two ventures: [Elora Therapeutics](https://eloratherapeutics.com), which I co-founded to engineer enzymes that degrade microplastics in human tissue, and [DeepCovalent](https://deepcovalent.com), where I advise on AI-driven drug discovery. Highlights include [400+ Protein Structures,](https://marcdeller.com/structures/) [60+ Publications,](https://marcdeller.com/publications/) [7+ Patents,](https://marcdeller.com/patents/) and [6+ FDA IND filings.](https://marcdeller.com/my-work-driving-innovation-in-precision-oncology-through-structure-based-drug-design/)\n\n[Resume](https://marcdeller.com/wp-content/uploads/2026/07/marc_deller_resume_july_2026.pdf)\n\n* **Structural Biologist and Protein Scientist:** I use protein structure to answer practical questions in drug discovery.\n* **Broad, hands-on domain experience:** I\u2019ve spent 10+ years across protein production, protein engineering, antibodies, small molecules, and AI\u2011guided therapeutics, always focused on what will move a program forward.\n* **Leader, mentor, and team builder:** I develop and support teams and automated platforms that increase experimental throughput without sacrificing scientific rigor.\n* **Deep toolkit, applied selectively:** I work with Cryo\u2011EM, X\u2011ray crystallography, Biophysics, and modern AI tools, using them where they genuinely accelerate development rather than add noise.\n* **Collaborative, impact-driven partner:** I enjoy working with teams that need deep structural insight paired with pragmatic, steady leadership.\n\nPersonal Philosophy: The 6C's\n\n### [**Curiosity:** 10+ years driving hands-on drug discovery through multidisciplinary approaches](https://marcdeller.com/curiosity-20-years-driving-hands-on-discovery-through-multidisciplinary-approaches/)\n\n[**Curiosity:** 10+ years driving hands-on drug discovery through multidisciplinary approaches](https://marcdeller.com/curiosity-20-years-driving-hands-on-discovery-through-multidisciplinary-approaches/)\n\n### [**Courage**: 400+ challenging structures determined, enabling breakthrough mechanistic insights](https://marcdeller.com/courage-400-structures-determined-enabling-breakthrough-mechanistic-insights/)\n\n[**Courage**: 400+ challenging structures determined, enabling breakthrough mechanistic insights](https://marcdeller.com/courage-400-structures-determined-enabling-breakthrough-mechanistic-insights/)\n\n### [**Creativity:** 7+ patents through innovative protein engineering and AI integration](https://marcdeller.com/creativity-5-patents-through-innovative-protein-engineering-and-ai-integration/)\n\n[**Creativity:** 7+ patents through innovative protein engineering and AI integration](https://marcdeller.com/creativity-5-patents-through-innovative-protein-engineering-and-ai-integration/)\n\n### [**Communication:** 60+ publications translating structural data into actionable insights](https://marcdeller.com/communication-60-publications-translating-structural-data-into-actionable-insights/)\n\n[**Communication:** 60+ publications translating structural data into actionable insights](https://marcdeller.com/communication-60-publications-translating-structural-data-into-actionable-insights/)\n\n### [**Collaboration:** 2+ state-of-the-art Structural Biology facilities built through global partnerships](https://marcdeller.com/collaboration-2-state-of-the-art-structural-biology-facilities-built-through-global-partnerships/)\n\n[**Collaboration:** 2+ state-of-the-art Structural Biology facilities built through global partnerships](https://marcdeller.com/collaboration-2-state-of-the-art-structural-biology-facilities-built-through-global-partnerships/)\n\n### [**Compassion:** 6+ FDA IND filings supported while mentoring 20+ scientists](https://marcdeller.com/compassion-6-fda-ind-filings-supported-while-mentoring-20-scientists/)\n\n[**Compassion:** 6+ FDA IND filings supported while mentoring 20+ scientists](https://marcdeller.com/compassion-6-fda-ind-filings-supported-while-mentoring-20-scientists/)\n\nInteractive Bio Timeline\n\nCareer Timeline\n\nInstitution\n\nSummary\n\nPrograms\n\nTargets\n\nTechniques", "# deepcovalent: OS3D\u2122 | Accelerate Drug Development Today \u2014 DeepCovalent\nAuthors: deepcovalent\nSource: https://www.deepcovalent.com, 2026\n\nDeepCovalent\u2019s **OS3D\u2122** delivers single-step, seed-free *de novo* small-molecule generation, enabling unbiased access to novel chemical space beyond incremental analog design. Its multimodal, multi-objective architecture conditions generation simultaneously on target biology, efficacy signals, ADMET profiles, and medicinal chemistry constraints\u2014producing synthesis-ready candidates with early translational relevance. By unifying ideation, design, and *in-silico* validation into a single forward pass, it generates optimized and *in-silico* validated molecules faster, leaner, and beyond known chemistry.\n\n## One-Step-Denovo\u2122 Drug Discovery (OS3D\u2122)\n\n* Reverse trial-and-error approach.\u200b\n* Manual and sequential iteration across chemistry, biology, and ADMET.\n\n#### **Traditional Wrokflow is Slow and Costly**\n\n* GenAI-driven design enables one-step forward drug discovery.\u200b\n* Rapid generation and evaluation of small molecule candidates\u00a0*in-silico.*\u200b\n\n#### **OS3D\u2122: Ideation to Molecule in** **3 Weeks**\n\n* Incomplete modeling of protein\u2013ligand context and clinical constraints.\u200b\n* Optimization constrained to known chemical space.\n\n#### **Why 90% of Drugs Fail in Clinical Trials\u200b**\n\n* Unified multimodal and multi-objective modeling of protein\u2013ligand systems.\u200b\n* Integrates biology, bioactivity, and rich ADMET context.\u200b\n* Unbiased access to previously unknown chemical space.\n\n#### **OS3D\u2122: High Downstream Success**\n\n[Partner With US](/partnerships)\n\n[Learn more](/contact)", "# Protein stability: a crystallographer's perspective\nAuthors: Deller, M. C.; Kong, L.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 72, 72\u201395 (IYCr crystallization series), 2016\n\nBuffer composition exerts a direct influence on stability, so buffer screening is a powerful stabilization strategy both for crystallography and for biologics formulation. Thermofluor is the primary high-throughput approach: hundreds of buffer and pH combinations can be screened and the stabilizing effect read out as a shift in melting temperature (\u0394Tm). Citrate, bis-tris and N-(2-acetamido)iminodiacetic acid (ADA) have been identified as having statistically significant stabilizing effects across tested proteins; in the extreme case, intrinsically insoluble proteins such as V22-SH3 are soluble only in pure water. Ligand binding is likewise strongly stabilizing, and co-crystal structures with cofactors, prosthetic groups, substrates, drugs and inhibitors are described as the holy grail of structural biology. Although soaking compounds through the lattice often works, ligand binding may induce conformational change, so ab initio screening in the presence of ligand may be required. Ions and organic additives can stabilize protein or lattice: Thermofluor identified magnesium as stabilizing *Mycobacterium tuberculosis* DapD, and adding MgCl2 produced larger crystals, with two tightly coordinated Mg\u00b2\u207a ions found at the homotrimer interface. PEG and MPD are frequently observed bound in structures, MPD apparently promoting hydration of the surface by binding exposed hydrophobic residues such as leucine, and additive screens such as Silver Bullets (Hampton Research) can stabilize or cross-link proteins to promote lattice formation.\n\n### 6.4. Fused affinity tags and crystallization chaperones\n\nAffinity tags such as MBP aid stability and solubility during expression and purification, but large tags are usually removed before crystallization because linker flexibility interferes with packing; small tags such as His tags can often be left on without unduly affecting packing or function. Since the problem is the flexible linker rather than the tag itself, several groups have engineered out that flexibility by replacing the usual protease cleavage site with a string of alanine residues, producing rigid fusions. This approach has yielded crystal structures of MBP fusions including gp21, SarR, MAT\u03b11 and MCL1, and has generalized into the concept of crystallization chaperones. Cited examples of chaperone use include Fab antibody fragments applied to capture the neurotransmitter sodium symporter LeuT in various conformational states, and the fusion of T4 lysozyme to a G-protein-coupled receptor. These approaches are proving especially useful for stabilizing membrane proteins and enabling their structure determination.\n\n### 6.5. Reductive alkylation\n\nChemical modification of the protein surface is a well-established route to improved crystallization, complementary to site-directed mutagenesis. The commonest method is reductive methylation of the \u03b5-amino groups of solvent-exposed lysine residues using dimethylamine\u2013borane and formaldehyde, a technique that has become a workhorse salvage method in structural genomics consortia and is thought to work by introducing new surface contacts that promote lattice formation. Recent developments include ethylation and isopropylation, though fewer targets are available to assess them. An important variant is cysteine alkylation for membrane-protein structure determination, first used for the \u03b21-adrenergic GPCR and now common in GPCR work; it stabilizes by preventing disulfide formation and, in the \u03b21-adrenergic receptor, by stabilizing monomers and preventing oligomerization. Other chemical modifications include fluorination: fluorine is essentially absent from biology but stabilizes proteins through the \"fluorous effect\", which promotes phase separation and increases buried surface area in the hydrophobic core. Fluorinated proteins can be made with hexafluoroleucine; the structure of the designed four-helix bundle \u03b14H shows fluorinated residues packing well into the core with little perturbation.\n\n### 6.6. Surface mutagenesis, surface-entropy reduction and deglycosylation\n\nMutating surface-exposed residues is a proven route to better stability and crystallizability. An early success was transplanting key crystal contacts from the rat ferritin structure onto the previously unsolved human orthologue, mutating human Lys86 to Glu to mimic the Ca\u00b2\u207a site seen in rat crystal contacts \u2014 useful only when an orthologous structure exists. To generalize the idea, Derewenda and coworkers mutated surface glutamate and lysine residues of RhoGDI to alanine, arginine or aspartate; these residues have high conformational entropy and rarely participate in crystal interfaces, making them attractive targets. The mutable set has since been extended to glutamine, and the SERp server evaluates solvent exposure, secondary structure, surface entropy and evolutionary conservation \u2014 the last used to avoid mutating structurally or functionally important residues \u2014 and proposes suitable residue clusters. SER has been combined with fused-tag and chaperone approaches, as in the RACK1A\u2013MBP fusion structure. Surface glycosylation demands separate handling: strategies include expression in glycosylation-deficient CHO lec1 or HEK293S GnTI\u207b cells, mutagenesis of Asn-X-Ser/Thr sequons or of O-linked Ser/Thr sites, enzymatic removal with Endo H or PNGase F (which converts the glycosylated asparagine to aspartate), and glycosylation inhibitors such as swainsonine and kifunensine, often combined with subsequent Endo H treatment as used for sRPTP\u03c3 and s19A. Expression host also tunes glycan patterns \u2014 Sf insect cells give simpler GlcNAcMan5-type glycans at fewer sites, whereas yeast can hyperglycosylate. Crucially, glycans are not always deleterious: densely glycosylated hepatitis C virus E2 bound to a broadly neutralizing antibody crystallized only when fully glycosylated, with a key crystal contact mediated by an N-linked glycan contacting a symmetry mate.", "# Vibe Coding MicroPlastic Blaster\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nfunction playSound(type) {\n  if (!audioCtx || muted || !sfxGainNode) return;\n  const now = audioCtx.currentTime;\n\n  if (type === \"destroy\") {\n    const o1 = audioCtx.createOscillator();\n    const o2 = audioCtx.createOscillator();\n    const g = audioCtx.createGain();\n    o1.connect(g); o2.connect(g);\n    g.connect(sfxGainNode);\n\n    o1.type = \"sine\";\n    o1.frequency.setValueAtTime(600, now);\n    o1.frequency.exponentialRampToValueAtTime(1200, now + 0.08);\n    o1.frequency.exponentialRampToValueAtTime(200, now + 0.25);\n\n    o2.type = \"triangle\";\n    o2.frequency.setValueAtTime(800, now);\n    o2.frequency.exponentialRampToValueAtTime(100, now + 0.3);\n\n    g.gain.setValueAtTime(0.15, now);\n    g.gain.exponentialRampToValueAtTime(0.001, now + 0.3);\n\n    o1.start(now); o2.start(now);\n    o1.stop(now + 0.3); o2.stop(now + 0.3);\n  }\n}\n\n```\n\nI did eventually break one of my initial rules, no external file dependencies, to increase the richness of gameplay. I am glad I did! Much more fun and\u00a0engaging now with real images and music, all of course made using Gemini AI, including Nano Banana for the images and Lyria for the music:\n\n**Firebase Leaderboard:**\u00a0One of my key learnings from the [**Learn to Code\u00a0with AI**](https://marcdeller.com/learn-to-code-with-ai/) certificate I recently completed is the use of a real-time asynchronous database using [Firebase DB](https://firebase.google.com) (from Google) to build a persistent database to host the leaderboard.\n\nAll three local-storage functions (`getLeaderboard`, `saveLeaderboard`, `addScoreToLeaderboard`) and `renderLeaderboard`/`toggleLeaderboardPanel` were rewritten as `async` functions using the Firebase REST API at `https://microplastic-blaster-default-rtdb.firebaseio.com/leaderboard.json`:\n\n* **GET scores**: `fetch(FIREBASE_LB_URL)` \u2014 handles both Firebase object and array response formats\n* **PUT scores**: stores the full sorted top-5 array in one atomic write\n* `gameOver()` made `async` with `await` on both Firebase calls\n* `renderLeaderboard` shows a *\u201cLoading leaderboard\u2026\u201d* placeholder while fetching\n\nOnce this one was set up, the DB automagically handled the setup, and I ended up with a nice, simple DB recording date, level, name, and score for the top five players:\n\n```\n\n\t\t\t\t\t{\n  \"leaderboard\": [\n    {\n      \"date\": \"2026-04-09\",\n      \"level\": 16,\n      \"name\": \"Marc\",\n      \"score\": 152075\n    },\n    {\n      \"date\": \"2026-04-07\",\n      \"level\": 16,\n      \"name\": \"Brian Clough\",\n      \"score\": 132200\n    },\n    {\n      \"date\": \"2026-04-07\",\n      \"level\": 13,\n      \"name\": \"Wallace\",\n      \"score\": 81700\n    },\n    {\n      \"date\": \"2026-04-07\",\n      \"level\": 12,\n      \"name\": \"Marc\",\n      \"score\": 71840\n    },\n    {\n      \"date\": \"2026-04-07\",\n      \"level\": 11,\n      \"name\": \"Gromit\",\n      \"score\": 69755\n    }\n  ]\n}\n\n```\n\n**Summary:** Great fun project overall, ~6 hours to get up and running with a working version, and I am now in what appears to be an infinite timeline to completion! My impression is that vibe coding is fast for building demos, wireframing, and planning, but tweaking, polishing, and adding final touches are more efficiently achieved by digging into the code.\n\n\ud83e\uddec **MicroPlastic Blaster**\n\nDeploy Enzymes. Destroy Microplastics. Save the Body.\n\nA browser-based arcade game by [Elora Therapeutics](https://www.eloratherapeutics.com) \u2014 the biotech developing first-in-class enzyme therapy to remove microplastics from the human body. Fire enzymatic blasts at incoming polymer particles, protect the cell, and climb the global leaderboard.\n\n\ud83c\udfae **Gameplay**\n\nYou are the body\u2019s enzymatic defense. Microplastic particles \u2014 PET, PE, PS, PP, plus branded consumer-plastic enemies \u2014 descend toward a target cell at the bottom of the screen. Click (or tap) to launch enzyme projectiles and intercept them before they reach the cell membrane.\n\n\u2013 Each microplastic that breaches the membrane deals health damage proportional to the current level.\n\u2013 The game ends when cell health reaches zero \u2014 the \u201cCell Overwhelmed\u201d screen appears.\n\u2013 Difficulty increases every 720 game ticks; spawn rate accelerates with level.\n\u2013 Score = particle point value \u00d7 current level.\n\n\u2697\ufe0f **Microplastic Types**\n\n| Polymer | Abbreviation | Colour | HP | Speed | Points | Shape |\n| --- | --- | --- | --- | --- | --- | --- |\n| Polyethylene terephthalate | PET | Purple | 3 | Slow | 30 | Block |\n| Polypropylene | PP | Red | 1 | Fast | 15 | Jagged |\n| Polyethylene | PE | Amber | 2 | Medium | 20 | Foam |\n| Polystyrene | PS | Green | 2 | Fastest | 25 | Circle |\n| Aquafina bottle | WATER | Blue halo | varies | Medium | \u2014 | Image sprite |\n| Starbucks cup | COFFEE | Green halo | varies | Medium | \u2014 | Image sprite |\n\n\u26a1 **Power-Ups**\n\nPower-ups drift down the playfield and are collected by firing an enzyme at them (or by direct cell contact). Each lasts 10 seconds.\n\n| Power-Up | Icon | Effect |\n| --- | --- | --- |\n| FAST-PETase | \u26a1 | Rapid-fire enzyme shots |\n| Enzyme Boost | \ud83d\udcaa | Increased projectile damage |\n| Cell Repair | \u2764\ufe0f | Restores cell health |\n| Chain Reaction | \ud83d\udca5 | Explosive chain-blast on impact |\n| Marc \u00d710 | Co-founder avatar | 10\u00d7 fire rate + Marc\u2019s voice line |\n| Paul \u00d710 | Co-founder avatar | 10\u00d7 fire rate + Paul\u2019s voice line |\n\nCo-founder power-ups (Marc / Paul) disable the background fight track while the voice clip plays, then restore normal audio automatically on clip end.\n\n\ud83c\udfc6 **Leaderboard**\n\nScores are persisted to a Firebase Realtime Database (top 5 retained). On game over, players enter a display name and submit; the leaderboard renders immediately. It is also accessible from the start screen via View Leaderboard.\n\nFirebase endpoint: https://microplastic-blaster-default-rtdb.firebaseio.com/leaderboard.json\n\nEntries store: `name`, `score`, `level`, `date (ISO YYYY-MM-DD)`.\n\n\ud83d\udd27 **Tech Stack**", "# GitHub README: TopPDBLX\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/TopPDBLX, 2026\n\n**In plain terms:** a small model is cheap to run over 200,000 records but not clever enough to teach itself. A large model is clever enough to teach but far too slow to run over the whole archive. So the large one reads a few thousand of the hard cases, the small one learns from what it produces, and the small one does the actual work. The expert never labels a corpus: they label a couple of hundred records that decide whether any of it worked.\n\nThe cycle has three participants and each does the one thing it is best at:\n\n| | Does | Costs | Why it cannot do the others' job |\n|---|---|---|---|\n| **Expert** (`gold_bench`, `courtroom`) | Labels the gold sets; answers curation questions | An hour or two | Cannot label the whole residual, and should never be asked to |\n| **Teacher** (Qwen2.5-32B, local) | Labels a few thousand of the hardest residual records | ~40 s/record | Far too slow for the corpus, and noisier than the student it trains |\n| **Student** (SmolLM2-360M) | Reads every residual record | ~0.1 s/record | Bootstrapped from the rules, so the rules were its ceiling until the teacher's labels lifted it |\n\n**Why the expert sits at the top of the loop and not inside it.** Every automated measure in this project is precision-shaped: it can catch an invented reagent, but nothing automated can catch one that was never mentioned at all. The gold sets are the only thing that can, so they are the yardstick and are never trained on, and every teacher run excludes them by name.\n\nThe other loop is unchanged and still pays: names the parser reaches for but cannot find are grouped into a few dozen curation calls, so the expert only ever sees decisions the machine could not make alone, and the lexicon they produce feeds back into the next parse.\n\n**The core of it is about ten lines.** Everything else is provenance, validation and the lexicon:\n\n```python\nfor deposition in archive:\n    components = []\n    for clause in split_into_clauses(deposition.text):     # \";\" \",\" newlines, \" and \", ...\n        amount, unit, name = read_quantity(clause)          # \"20% (w/v) PEG 3350\"\n        reagent = lexicon.lookup(name)\n        if reagent is None:                                 # retry, never destructive\n            reagent = lexicon.lookup(strip_prose(clause))   # \"crystal conditions were 0.1 M ...\"\n        components.append(reagent or Unidentified(name))\n\n    families = {family_of(c) for c in components            # {Organic, PEG, Salt}\n                if c.is_precipitant and c.amount is not None}\n\n    if any(c.is_unidentified for c in components) or not families:\n        condition_class = Unclassified(reason)              # an answer, not a failure\n    else:\n        condition_class = SEVEN_CLASSES[families]\n```\n\n| Tool | What it does here |\n|---|---|\n| RCSB Data GraphQL API | Fetches every deposition, batched and resumable |\n| gemmi | Reads the archive mmCIF for the byte-level fidelity gate |\n| `regex` | Clause splitting, which turned out to be the hard part rather than the chemistry |\n| `ontology/synonyms.yaml` | The reagent dictionary: 500 reagents, 1,587 spellings |\n| pydantic | Enforces the schema and the chemical invariants on load |\n| polars, pyarrow, duckdb | Tables, joins and the queryable release |\n| MMseqs2 | Sequence clustering at 30%, 50% and 90% identity, to control redundancy |\n| MLX-LM, SmolLM2-360M | The student: reads the residual the rules cannot, fine-tuned on Apple Silicon |\n| MLX, Qwen2.5-32B-Instruct-4bit | The teacher: labels the hard records locally, so the corpus never leaves the machine |\n| Weights & Biases | Training runs, one named run per round |\n| `condition_courtroom_v7.html` | Single-file curation and accuracy audit, one condition per screen, no server |\n| `gold_bench_v1.html` | Single-file labeller for the gold records that judge every model round |\n\n**Why the fidelity gate comes first.** Everything downstream is a claim about what a depositor wrote, so the first stage proves the text was fetched intact: every condition string is byte-compared against the mmCIF in the 90 GB archive snapshot, including loop row counts. It agrees on all 205,943 entries, at 100%. Without that, a parsing statistic would be measuring the API rather than the archive.\n\n## \ud83d\udcca Results\n\n| Measure | Value |\n|---|---|\n| Records | 199,185 |\n| Usable | **188,039 (94.4%)** |\n| Components | 645,656, **87.3% identified** as a canonical reagent. 610,076 from the rule parser, 35,580 recovered by the residual parser |\n| Reagent lexicon | 500 reagents, 1,587 names (v0.9.3) |\n| Linked sequences | 184,229 across **23,159** distinct 30% identity clusters |\n| Screen-well matches | 81,802 component-set matches, 39,219 agreeing on every concentration |\n| Archive fidelity | **100%** over 205,943 entries against the 90 GB mmCIF snapshot |\n| Condition classes | Seven JCSG Top96 precipitant classes (v0.3.1), **80.2% classified**, 19.8% honestly unclassified |\n| Parser accuracy | Both gold-set batches pooled, rules + round 09: **93.3% precision, 90.3% recall** (rules alone: 95.1% precision, 69.5% recall) |\n| Leak-free splits | 181,007 records, no cluster spanning folds at 30%, 50% or 90% |\n| Tests | 492 passing |\n\nThe project brief anticipated a rule-based parser plateauing near 75%. It reaches 93.5%, because most of the difficulty in this corpus turned out to be clause splitting rather than chemistry: newlines, double spaces, `in`, spaced slashes and stray brackets all separate components, and handling them properly recovered far more than chasing reagent names did.\n\n### Two denominators, both reported", "# 2025 \u2013 Dashboard Development with Shiny: GenAI for Retail Analysis\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-28 on marcdeller.com (categories: certificates).*\n\n* [Link to certificate](https://www.coursera.org/account/accomplishments/verify/OO09W4LMZGCB)\n\n* **Award Details:** Coursera issued this project certificate on March 27, 2025.\n* **Course Focus:** It recognizes completion of a project on building interactive dashboards with Shiny.\n* **Career Value:** This matters because dashboards make large and complex datasets easier to explore and explain.\n* **Practical Use:** It supports creation of scientific interfaces for visualizing structures, assays, and analytical outputs.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\nThis hypothesis has gained further support from in vivo protein fragment complementation assays, in which EpoR-JAK2 fusions have been made constitutively active through the inclusion of linker regions between the receptor and kinase domains (Remy et al., 1999). This allows optimal association of signaling components in the absence of ligand induced receptor dimerisation. Further support is also provided by the crystal structure of the ligand free EpoR, in which crystal packing contacts imply a propensity for formation of an inactive dimer structure in which the intracellular domains would be too far apart to permit phosphorylation of JAK2 (Livnah et al., 1999). It is proposed that these preformed inactive dimers could exist on the cell surface and that ligand binding induces a rearrangement that activates the receptor.\n\n##### 1.4.2.1.4 Structure-Function Relationship of gp130\n\nThe gp130 receptor structure has been determined crystallographically in a ligand free state (Bravo et al., 1998) (see Figure 1-3 and Figure 1-6). This receptor shows homology to other haemopoietin receptors as discussed above (see 1.4.2.1.1) (see Figure 1-6). However, one notable feature unique to gp130 is the packing observed in the WSXWSX motif. This motif forms two-successive \u03b2-bulges resulting from the intercalation of Trp 288, Trp 291 and Trp 247 between the aliphatic residues Met 280, Arg 278, Arg 276 and Arg 240 which are located on strand F, so forming an extended \u03c0-cation stacking system (see Figure 1-6). This is more extended than other members of the family due to inclusion of Trp 247 and Arg 240, which are not classic members of the motif. The second major structural difference is the division of strand G of the N-terminal domain into two equal portions, again resulting from a second \u03b2-bulge structure similar to that observed in the WSXWS motif itself. In contrast to hGHR, the interdomain 3\u2081\u2080 helical region is followed by a region of short \u03b2-strand. This functions in hydrogen bonding to residues of both fibronectin type III-like domains, so resulting in increased interdomain rigidity.\n\nSites responsible for ligand binding have been mapped to residues including Phe 169 of the N-terminal domain EF loop (analogous to Trp 104 of hGHR) and Val 230 in the BC loop (analogous to Trp 169 of hGHR) of the C-terminal domain. Val 230 is relatively buried in the crystal structure although mutagenesis does affect binding to IL-6-IL-6R (Horsten et al., 1997), so suggesting this residue has an indirect effect on binding resulting from perturbed local structure. In GHR site 2 interactions are mediated through Asn 128 and this residue is structurally equivalent to Asp 283 of gp130 (see Figure 1-6). However, no data are yet available concerning the function of this residue in gp130.\n\nThe gp130 family of receptors includes IL-6R (Ward et al., 1995), LIFR (Gearing et al., 1991) (Hilton and Nicola, 1992), OSMR (Linsley et al., 1989; Mosley et al., 1996), CNTFR (Davis et al., 1993), IL-11R (Hilton et al., 1994), CT-1R (Robledo et al., 1997) and gp130 (Taga and Kishimoto, 1995). These receptors are structurally related and show many overlapping biological activities. Genetic analysis (Gough, 1992) and biochemical reconstitution (Baumann et al., 1993) has revealed a common receptor complex structure in which the ligand specific \u03b1-subunit is combined with one or two receptor components that are shared between the various cytokine family members. These common components are gp130 and LIFR, with the gp130 acting as a signal-transducing component in all cases. IL-6 and IL-11 have been shown to bind to ligand-specific receptors (i.e. IL-6R and IL-11R, respectively) so inducing homodimerisation of gp130 and initiation of signaling (see Figure 1-4).\n\nOther members of the family such as LIF and OSM induce heterodimerisation of gp130 with LIFR. OSM can also induce heterodimerisation of gp130 with a specific OSMR\u03b2 subunit (Mosley et al., 1996), so forming a so called type II assembly consisting of OSM, gp130 and OSMR\u03b2 as shown in Figure 1-5 below. In a similar fashion CNTF interacts with CNTFR and then forms a higher order complex also containing gp130 and LIFR. Analogous trimeric receptor assemblies are also proposed to exist for CT-1, which is thought to bind a ligand-specific receptor followed by gp130 and LIFR. CNTFR differs from other members of the family as it is attached to the membrane via a glycero-phosphatidyl inositol (GPI) anchor, therefore lacking both transmembrane and cytoplasmic domains.\n\n#### 1.4.2.2 Interferon/IL-10 Receptor Superfamily\n\nINF\u03b3R and IL-10R belong to this superfamily which is structurally related to the haemopoietin receptor superfamily and as such also functions in binding \u03b1-helical bundle cytokines (see Table 1-2). This superfamily is often referred to as haemopoietin class II. The fibronectin type III-like modules of this family contain a conserved pattern of cysteine residues at the C- and N-termini, but the WSXWS motif observed in the haemopoietin class I family is not present (Taga and Kishimoto, 1992). Crystal structures solved include tissue factor (TF) (Harlos et al., 1994; Muller et al., 1996), killer cell inhibitory receptor (KIR) (Fan et al., 1997; Maenaka et al., 1999) and IFN\u03b3R (Sogabe et al., 1997). These structures have revealed variability in the interdomain angle ranging from ~50\u00b0 in KIR to ~120\u00b0 in TF and IFN\u03b3R. This compares with the ~90\u00b0 angle observed in EpoR, GHR, PrlR and gp130 (see Figure 1-3 for comparison).\n\n#### 1.4.2.3 TNF/NGF Receptor Superfamily", "# From Bench to Regulation: How the FDA Is Formalising a Future Beyond Animal Testing\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-04-16 on marcdeller.com (categories: blog).*\n\n* [**FDA Press Release**](https://www.fda.gov/news-events/press-announcements/fda-releases-draft-guidance-alternatives-animal-testing-drug-development)\n* [**Draft Guidance**](https://www.fda.gov/regulatory-information/search-fda-guidance-documents/general-considerations-use-new-approach-methodologies-drug-development)\n\n* **FDA\u2019s draft guidance:** defines four validation principles that NAMs must satisfy before regulatory submission.\n* **NAMs in regulatory vs. discovery contexts:** are explicitly distinguished: the guidance covers submissions only, not early-stage research.\n* **Human Liver-Chip data:** demonstrate that validated organ-on-a-chip models can outperform animal models in predicting clinical toxicity.\n* **Early FDA engagement:** is recommended for all sponsors considering NAMs in indication-, organ-, or endpoint-specific programmes.\n\n1. **FDA Releases Draft Guidance on Alternatives to Animal Testing in Drug Development:** [FDA Press Announcement, U.S. Food and Drug Administration, 2026](https://www.fda.gov/news-events/press-announcements/fda-releases-draft-guidance-alternatives-animal-testing-drug-development)\n2. **General Considerations for the Use of New Approach Methodologies in Drug Development:** [FDA Guidance Document, U.S. Food and Drug Administration, 2026](https://www.fda.gov/regulatory-information/search-fda-guidance-documents/general-considerations-use-new-approach-methodologies-drug-development)\n3. An FDA/CDER Perspective on Nonclinical Testing Strategies: Classical Toxicology Approaches and New Approach Methodologies (NAMs): [Avila et al., Regulatory Toxicology and Pharmacology, 2020](https://doi.org/10.1016/j.yrtph.2020.104662)\n4. Performance Assessment and Economic Analysis of a Human Liver-Chip for Predictive Toxicology: [Ewart et al., Communications Medicine, 2022](https://www.nature.com/articles/s43856-022-00209-1)\n5. Replacing Animal Testing with Stem Cell-Organoids: Advantages and Limitations: [Park et al., Stem Cell Reviews and Reports, 2024](https://doi.org/10.1007/s12015-024-10723-5)\n\n**New FDA draft guidance on New Approach Methodologies sets out a validation framework for human-centric nonclinical testing in drug development**\n\nFor decades, animal models have formed the backbone of nonclinical drug safety testing, offering a regulatory path that is well understood but imperfect. A growing body of evidence has demonstrated that species-to-species differences can lead to systematic failures in predicting human responses, contributing to the persistently high attrition rates observed in clinical development. Against this backdrop, the FDA\u2019s Center for Drug Evaluation and Research (CDER) has been incrementally building the scientific and regulatory infrastructure to accommodate alternatives, supported by the Food and Drug Omnibus Reform Act of 2022, which clarified that data from New Approach Methodologies (NAMs) could legitimately support drug development and approval submissions.\n\nIn March 2026, that groundwork crystallised into formal regulatory action. The FDA released draft guidance providing drug developers with a validation framework for NAMs in nonclinical submissions. NAMs encompass a broad spectrum of human-relevant approaches: two-dimensional and three-dimensional in vitro cell culture systems, organoids, microphysiological systems including organs-on-a-chip, in chemico assays, computational in silico modelling, and studies using phylogenetically lower organisms such as zebrafish or C. elegans. The guidance is explicit that its scope covers regulatory submissions, not drug discovery applications, drawing a sharp line between exploratory and regulatory contexts.\n\nThe practical utility of these technologies has already been demonstrated in published studies. An analysis of 870 human Liver-Chips showed 87% sensitivity and 100% specificity in predicting drug-induced liver injury across a blinded set of known hepatotoxic and non-toxic compounds, substantially outperforming conventional preclinical models, as described by Ewart et al. in Communications Medicine (2022). Complementary work on stem cell-derived organoids has reinforced this case; Park et al. (Stem Cell Reviews and Reports, 2024) reviewed how induced pluripotent stem cell-based systems and organoids can model organ-level pathophysiology with human genetic fidelity, offering platforms that animal models structurally cannot replicate. Together, these examples reflect the scientific rationale that underpins the FDA\u2019s current position.\n\n**\u201cTechnological advances are allowing us to move beyond animal testing in drug development, which has a poor track record of predicting safety and efficacy in humans.\u201d \u2013 FDA Commissioner Marty Makary, M.D., M.P.H.**\n\nThe guidance establishes four core validation principles that any NAM must satisfy to be considered for regulatory use. These are: Context of Use, requiring a clear definition of the NAM\u2019s intended regulatory purpose; Human Biological Relevance, demonstrating how the model can assess toxicity in a human-representative system; Technical Characterisation, establishing scientific confidence through robust and reproducible methods; and Fit-for-Purpose, assuring that the NAM can assist CDER in regulatory decision-making. This framework builds directly on the foundational FDA/CDER perspective published by Avila et al. in Regulatory Toxicology and Pharmacology (2020), which articulated the opportunities and challenges of NAMs in regulatory nonclinical programmes and remains the primary published statement of FDA/CDER\u2019s mechanistic philosophy on the subject.", "# Crystal Structure of BG505 SOSIP gp140 HIV-1 Env Trimer Bound to Broadly Neutralizing Antibodies PGT128 and 8ANC195\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [5C7K](https://www.rcsb.org/structure/5C7K)\n* **Release Date:** 2015-10-07\n* **Method:** X-RAY DIFFRACTION 4.6019 \u00c5\n* **Organisms:** Homo sapiens, Human immunodeficiency virus 1\n* **Macromolecules:** Unique protein chains: 6 (proteins)\n* **Unique Ligands:** NAG, SO4\n* **Publication:** [Kong, L., et al. (2015). Complete Epitopes for Vaccine Design Derived from a Crystal Structure of the Broadly Neutralizing Antibodies PGT128 and 8ANC195 in Complex with an HIV-1 Env Trimer. Acta Crystallographica Section D, 71, 2099-2108.](https://journals.iucr.org/paper?S1399004715013917)\n\n**Description:** This HIV-1 Env trimer structure in complex with two broadly neutralizing antibodies\u2014PGT128 and 8ANC195\u2014reveals complete epitopes on the native-like trimer for the first time. The structure uncovers the full 8ANC195 interface spanning gp120 and gp41, and demonstrates how PGT128 allosterically repositions the N262 glycan, providing direct blueprints for structure-based HIV vaccine design.", "# marcdeller.com: Chat with My Bot\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n# Let\u2019s Chat: Proteins, Platforms, Programs\n\nMost bios tell you what someone has done. This is a chance to ask. I\u2019m Marc, a structural biologist and drug discovery leader with two decades of experience across X-ray crystallography, Cryo-EM, protein engineering, and translational science. Rather than reading a static resume, you can ask me directly about my approach to structure-guided discovery, how I\u2019ve built and led scientific platforms, what I look for when designing a drug discovery program, or where I think the field is heading. Think of this less as a chatbot and more as a conversation with me.\n\n**Fun questions to cut and paste:**\n\n* If you could create a dream team of famous scientists (living or historical) to collaborate on tackling big problems, who would you choose and why?\n* Imagine you have unlimited resources to build the ultimate automated research platform for structural biology\u2014what cool features and capabilities would it include?\n* What\u2019s the wildest or most futuristic application of protein\u2013ligand crystal structures you can envision for drug discovery or diagnostics?\n* If a plastic-eating enzyme like PETase could \u201ctalk,\u201d what would it say about its role in cleaning up our planet\u2014and what mutation would it request for its next upgrade?", "# GitHub README: ChemSage\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ChemSage, 2026\n\n> **A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behaviour, and live RDKit/PyMOL tool calls for chemical truth. Fine-tuned and served locally on Apple Silicon with MLX-LM.**\n\n[](https://chemsage.mdeller.com)      \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://chemsage.mdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">chemsage.mdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/ChemSage\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/ChemSage</a></td>\n</tr>\n</table>\n\n---\n\nChemSage is a locally hosted, open-source assistant that reasons about small molecules (SMILES,\nscaffolds, properties), drives the cheminformatics and structural tools a medicinal chemist\nactually uses (RDKit, PyMOL, docking, PLIP), and grounds every factual claim in your own corpus.\n\n**Why it matters:** off-the-shelf models invent chemistry (wrong logP, malformed SMILES) and\ncannot see your private SAR data. ChemSage fixes both by separating concerns: volatile knowledge\nlives in retrieval, stable behaviour lives in the weights, and deterministic chemical facts are\ncomputed live by trusted tools rather than guessed. It is useful for: interrogating your own\nassay and SAR data, triaging compounds, generating correct PyMOL/RDKit scripts, and interpreting\ndocking and interaction output, all on local hardware. A live, no-install terminal session\nagainst the real fine-tuned model runs at\n**[chemsage.mdeller.com](https://chemsage.mdeller.com)** (see **Web deployment** below).\n\n**Initial idea:** Fed up with token, model and upload document limits? Then why not run your own\nAI models on Mac hardware? Thanks to [@ai.christianson](https://www.tiktok.com/@ai.christianson)\non TikTok for highlighting the following method to implement MLX-LM optimised fine-tuning of a\nlocal LLM on Mac hardware: https://www.tiktok.com/@ai.christianson/video/7538150476467014943?lang=en\n\n**Disclaimer:** I am not a professional coder! This is an exercise in Vibe Coding as a Structural\nBiologist to get the job done. Use at your own risk, and double check the science. I am still\nworking on validation of the Chemistry and Biology. This is a proof of concept exercise to\ndemonstrate what can be achieved using moderate hardware and limited coding experience.\n\n**Project links:**\n\n- [GitHub](https://github.com/bellcheddar/ChemSage)\n- [My blog on the making of ChemSage](https://marcdeller.com/chemsage)\n- [Model data and evaluation stats](https://marcdeller.com/chemsage_evaluation)\n- [Models on Hugging Face](https://huggingface.co/Dellboy)\n\n## \ud83e\udd17 Models on HuggingFace Hub\n\n| Model | HuggingFace | Base | Round |\n|---|---|---|---|\n| ChemSage 7B v1 | [Dellboy/chem_sage_7b_v1](https://huggingface.co/Dellboy/chem_sage_7b_v1) | Qwen2.5-7B-Instruct-4bit | Round 1 |\n| ChemSage 32B v2 | [Dellboy/chem_sage_32b_v2](https://huggingface.co/Dellboy/chem_sage_32b_v2) | Qwen2.5-32B-Instruct-4bit | Round 2 |\n| ChemSage 32B v3 | [Dellboy/chem_sage_32b_v3](https://huggingface.co/Dellboy/chem_sage_32b_v3) | Qwen2.5-32B-Instruct-4bit | Round 3 |\n| ChemSage 32B v4 | [Dellboy/chem_sage_32b_v4](https://huggingface.co/Dellboy/chem_sage_32b_v4) | Qwen2.5-32B-Instruct-4bit | Round 4 |\n| **ChemSage 32B v5** \u2b50 | [**Dellboy/chem_sage_32b_v5**](https://huggingface.co/Dellboy/chem_sage_32b_v5) | Qwen2.5-32B-Instruct-4bit | **Round 5 (current)** |\n\n## \u2699\ufe0f The hybrid in one line\n\nRAG keeps it truthful today; the fine-tune makes it consistent tomorrow; RDKit makes it correct always.\n\n## \ud83d\ude80 Quick start\n\nSee **[`PROJECT_PLAN.md`](PROJECT_PLAN.md)** for the full step-by-step build (phases 0 to 8). Short version:\n\n| Step | What | Script / command |\n|---|---|---|\n| 0 | Env: MLX-LM + RDKit + GUI | (`requirements.txt`) |\n| 1 | Base model (Qwen 2.5 32B, MLX 4-bit) | `config/train_config.yaml` |\n| 2 | **RAG first** (ship this alone) | `scripts/ingest_rag.py` |\n| 2a | Download ChEMBL corpus | `scripts/download_chembl.py` + `scripts/download_chembl_extended.py` |\n| 2b | Download PubChem corpus | `scripts/download_pubchem.py` |\n| 2c | Download PDB corpus | `scripts/download_pdb.py` + `scripts/download_pdb_supplement.py` |\n| 3 | Build + validate SFT dataset | `scripts/build_dataset.py` |\n| 4 | QLoRA fine-tune | `mlx_lm.lora --config config/train_config.yaml --train` |\n| 5 | Fuse + serve | `mlx_lm.fuse` \u2192 `mlx_lm.server --port 8081` |\n| 6 | Close the hybrid loop | `rag/tool_exec.py` |\n| 7 | Chemistry-aware evaluation | `eval/eval_chem.py` |\n\n## \u25b6\ufe0f How to run\n\n**Terminal 1 \u2014 start the model server:**\n```bash\ncd /Users/dellboy/Documents/Vibe_Coding/chem_sage\nsource .venv/bin/activate\nmlx_lm.server --model models/chem_sage_32b_v5 --port 8080\n```\n\n**Terminal 2 \u2014 run chat:**\n```bash\ncd /Users/dellboy/Documents/Vibe_Coding/chem_sage\nsource .venv/bin/activate\npython scripts/chat.py --model models/chem_sage_32b_v5\n```\n\nWait for \"HTTP server listening\" in Terminal 1 before starting Terminal 2.\n\n## \ud83c\udf93 Training\n\n### Training data", "# TYK2 Biology: How Deep Mutational Scanning Reveals New Therapeutic Frontiers\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-18 on marcdeller.com (categories: blog, sb).*\n\n* [Deep mutational scanning reveals pharmacologically relevant insights into TYK2 signaling and disease](https://www.biorxiv.org/content/biorxiv/early/2025/10/12/2025.10.11.681520.full.pdf)\n\n* **Comprehensive Variant Mapping:** Deep mutational scanning of >23,000 TYK2 variants reveals [high-resolution structure-function relationships](https://www.biorxiv.org/content/10.1101/2025.10.11.681520v1) and identifies novel allosteric sites for selective drug targeting.\u200b\n* **Pseudokinase Regulation:** The [catalytically inactive JH2 domain](https://www.pnas.org/doi/10.1073/pnas.1401180111) serves as both an autoinhibitory regulator and a highly selective therapeutic target distinct from other JAK family kinases.\u200b\n* **Protective Variants Reduce Abundance:** Human variants that confer autoimmune protection work by decreasing [TYK2 protein levels](https://www.biorxiv.org/content/10.1101/2025.10.11.681520v1), suggesting degradation strategies as promising therapeutic approaches.\u200b\n* **Compound Potency Modulation:** Coupling DMS with inhibitor treatment identifies [variants affecting drug sensitivity](https://www.biorxiv.org/content/10.1101/2025.10.11.681520v1) and clarifies structure-activity relationships for medicinal chemistry optimization.\u200b\n* **Clinical Translation:** Selective TYK2 pseudokinase inhibitors demonstrate [robust efficacy](https://jamanetwork.com/journals/jamadermatology/fullarticle/2822710) in autoimmune diseases without JAK-related safety concerns observed with broader inhibitors.\n\n1. **Deep mutational scanning reveals pharmacologically relevant insights into TYK2 signaling and disease:** [Hiatt et al., bioRxiv](https://www.biorxiv.org/content/10.1101/2025.10.11.681520v1)\n2. **Structure of the pseudokinase\u2013kinase domains from protein kinase TYK2 reveals a mechanism for Janus kinase (JAK) autoinhibition:** [Lupardus et al., PNAS](https://www.pnas.org/doi/10.1073/pnas.1401180111)\n3. **High-resolution deep mutational scanning of the melanocortin-4 receptor enables target characterization for drug discovery:** [Jones et al., eLife](https://elifesciences.org/articles/104725)\n4. **Allosteric regulation and inhibition of protein kinases:** [Tong et al., Biochemical Society Transactions](https://pmc.ncbi.nlm.nih.gov/articles/PMC10089111/)\n5. **TYK2:p.Pro1104Ala variant protects against autoimmunity by modulating immune cell populations:** [Dendrou et al., Clinical & Experimental Immunology](https://pmc.ncbi.nlm.nih.gov/articles/PMC11885862/)\n\n**Systematic variant analysis illuminates pseudokinase function and guides next-generation autoimmune drug development**\n\nDeep mutational scanning has transformed our understanding of [Tyrosine Kinase 2 (TYK2)](https://www.biorxiv.org/content/10.1101/2025.10.11.681520v1), a genetically validated target in autoimmune disease, by systematically interrogating over 23,000 amino acid substitutions across critical protein functions. This [high-throughput approach](https://elifesciences.org/articles/104725)couples variant library generation with functional selection and next-generation sequencing to create comprehensive sequence-function maps, enabling researchers to identify novel druggable sites and clarify how human genetic variants influence disease susceptibility. The study specifically examined TYK2\u2019s role in IFN-\u03b1 signaling and protein abundance, revealing that protective human variants associated with reduced autoimmune risk actually decrease TYK2 protein levels rather than merely reducing enzymatic activity. This finding fundamentally challenges conventional therapeutic approaches that focus solely on kinase inhibition.\n\nThe [pseudokinase domain](https://www.pnas.org/doi/10.1073/pnas.1401180111), also known as the JH2 domain, emerges as a critical regulatory element that allosterically controls the adjacent catalytic kinase domain through extensive interdomain contacts. Unlike traditional ATP-competitive inhibitors that target the conserved active site, [allosteric inhibitors](https://pmc.ncbi.nlm.nih.gov/articles/PMC9186491/) binding to the JH2 domain achieve remarkable selectivity for TYK2 over other JAK family members (JAK1, JAK2, JAK3). Recent clinical successes with selective TYK2 pseudokinase inhibitors like deucravacitinib and zasocitinib demonstrate this strategy\u2019s therapeutic potential, with [zasocitinib achieving](https://jamanetwork.com/journals/jamadermatology/fullarticle/2822710) up to 33% complete skin clearance in psoriasis patients after just 12 weeks. The deep mutational scanning data further identified variants that modulate compound potency by coupling DMS with inhibitor treatment, prospectively revealing structure-activity relationships that inform medicinal chemistry optimization.\n\nBeyond inhibition, the observation that naturally occurring protective variants reduce TYK2 abundance suggests [protein degradation](https://www.biorxiv.org/content/10.1101/2025.10.11.681520v1) as an underexplored therapeutic modality. This concept aligns with emerging [PROTAC (proteolysis-targeting chimera) technologies](https://pmc.ncbi.nlm.nih.gov/articles/PMC10224132/) that have successfully entered clinical trials for other kinase targets. The comprehensive variant effect mapping enabled by DMS provides unprecedented insights into which protein regions are most sensitive to destabilization, potentially guiding the design of degrader molecules that selectively eliminate TYK2 while sparing related kinases. Additionally, the identification of novel allosteric sites through high-resolution structure-function mapping opens new avenues for drug discovery beyond the well-characterized JH2 pocket.", "# GitHub README: PANTS\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/PANTS, 2026\n\nSo the pipeline independently recovered a large cohort of gut proteins resembling\nenzymes it had never been shown, and 33 of them sit closest to the one with measured\nactivity at 40 \u00b0C and neutral pH.\n\n**What this does not show.** None of them *is* HGMP01: the best match is 38.3%, so these\nare relatives rather than rediscoveries. That is consistent with the source paper, which\nfound 697 HGMP01-like enzymes at 20% identity or better across the whole UHGP-100\ndatabase, where this scan covered 50 assemblies. It is also the expected outcome: HGMP01\ncomes from a UHGG genome that these particular assemblies need not contain.\n\nThe useful claim is narrower than \"PANTS found HGMP01\" and more interesting than nothing:\n**recall reaches the right neighbourhood of sequence space unaided**, in an environment\nnobody had asked it about until the therapeutic framing pointed there.\n\n## \ud83e\uddf1 Stack\n\n| Layer | Choice |\n|---|---|\n| Recall | MMseqs2 18-8cc5c, HMMER 3.4 (brew binaries, shelled out to) |\n| Embedding | ESM-2 `t12-35M`, frozen, CPU |\n| Heads | scikit-learn logistic regression, PU-corrected, Platt/isotonic calibration |\n| Structures | ESMFold offline, Boltz-2 via BoltzMaker for ligand co-folds (v2) |\n| Storage | SQLite (WAL), one file, no external database |\n| Web | Flask + gunicorn behind nginx, server-rendered templates plus vanilla ES6 |\n| Front end | Plotly.js, Mol\\*, Tabulator. No React, Vue, npm, webpack, Streamlit or Dash |\n\n**Two virtual environments, deliberately.** The droplet has 3.8 GB of RAM shared with five other applications, so the always-on web process never imports torch: `requirements-web.txt` is Flask, gunicorn and gemmi, and nothing else. All heavy compute is precomputed offline on an M1 Max and shipped as SQLite rows plus static mmCIF.\n\n## \ud83d\udd27 Installation\n\n```bash\ngit clone https://github.com/bellcheddar/PANTS.git\ncd PANTS\n\n# external tools\nbrew install hmmer mmseqs2\n\n# pipeline venv (offline batch work: torch, transformers, scikit-learn)\npython3 -m venv .venv\n.venv/bin/pip install -r requirements-pipeline.txt\n\n# web venv (always-on serving: no torch, ever)\npython3 -m venv .venv-web\n.venv-web/bin/pip install -r requirements-web.txt\n\n# keep bulk data out of the iCloud-synced Documents tree\nmkdir -p ~/PANTSData/{raw,interim}\nln -s ~/PANTSData/raw data/raw\nln -s ~/PANTSData/interim data/interim\n\ncp .env.example .env\n```\n\n## \ud83d\ude80 Usage\n\n```bash\n.venv/bin/python PANTS.py init                # create dirs, DB schema, check symlinks\n.venv/bin/python PANTS.py curate-seeds        # fetch wild types, derive variants\n.venv/bin/python PANTS.py harvest-negatives   # ESTHER hard negatives, matched\n.venv/bin/python PANTS.py status              # database summary\n.venv/bin/python PANTS.py serve               # local dev server on :8005\n.venv/bin/pytest                              # 28 tests\n```\n\n| Command | Does |\n|---|---|\n| `init` | Creates the directory tree and the schema, enables WAL, warns if `data/raw` is not a symlink out of iCloud |\n| `curate-seeds` | Fetches each wild type from UniProt by accession and derives every confirmed variant from its parent |\n| `harvest-negatives` | Streams the ESTHER slice from UniProt, classifies by family, selects a matched negative set |\n| `status` | Counts of candidates, scores, structures, characterised enzymes and recent runs |\n| `serve` | Flask dev server (production uses gunicorn on port 8005) |\n\n## \ud83d\udd2c Sequences are fetched, never typed\n\nEvery sequence enters through the UniProt REST client. Engineered variants have no accession of their own, so each is stored as a **parent plus a mutation list** and the sequence is derived, with `apply_mutations` refusing any substitution whose stated parent residue does not match.\n\nThat check earns its keep: a wrong mutation set yields a sequence that is still a valid protein, still folds, still embeds and still trains. The error would never surface as a crash, only as quietly degraded scores. Where a complete mutation set could not be confirmed, the variant is recorded with **no sequence** and excluded from training, because a partial set gives a wrong sequence and that is worse than an honest gap.\n\n`find_offset` determines the mature-versus-precursor numbering shift rather than guessing it: only an offset satisfying every mutation at once is accepted. All four confirmed variants validated at offset 0, across 14 residue positions.\n\nCorrections this caught during curation: `Q6A0I4` was initially curated as Cut190 and is actually **TfCut2** (*Thermobifida fusca*). Cut190 is `W0TJ64`, and its strain assignment (AHK190 versus type strain P101, both 304 aa) is still unresolved.\n\n## \ud83d\udcca Evaluation protocol\n\n| Element | Rule |\n|---|---|\n| Splits | By sequence cluster at 30% and 50% identity, never by sequence. Both reported |\n| Generalisation | Leave-one-family-out across ESTHER families |\n| Retrieval baseline | Model rank versus HMMER E-value rank on held-out characterised enzymes |\n| Composition baseline | Amino-acid composition plus length, cluster-grouped. Reported permanently |\n| Calibration | Reliability diagrams and Brier score, not just AUC |\n| Prospective set | Any PETase characterised after a fixed date, held out as a blind test |\n| Subsets | Reported separately for measured-activity and annotation-only positives |\n\n## \u26a0\ufe0f Limitations", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nExpression using Drosophila S2 cells proved unsuccessful as no induction of LIFR-FN3-3C-Fc expression was detected using ELISA, Protein-A immunoprecipitation or western blot. Samples typically contained a protein with an apparent molecular weight of ~46 kDa as assessed by western blot, compared with control LIFR-FN3-3C-Fc samples running at ~140 kDa. Samples of S2 conditioned medium were shown to be inactive in ligand binding assays and showed no binding to LIF or OSM. The nature of the ~46 kDa product is unknown although it is likely that this protein is a truncated version of LIFR-FN3-3C-Fc, due to its specific reactivity towards LIFR antibodies.\n\nTransient expression in 293T cells proved the most successful system tested for large-scale protein production. Large quantities of pIG-LIFR-FN3-3C-Fc DNA were routinely prepared from E. coli strain MC101/P3, with a triple BamHI, BglII and HindIII digest (as shown in Figure 4-4) being essential for confirmation of identity and quality of DNA. Typically, ~5 mg of pIG-LIFR-FN3-3C-Fc DNA was used to transfect ~36 \u00d7 162 cm\u00b2 flasks of 293T cells and protein yields of around 10 mg were achieved after 1 week incubation at 37 \u00b0C.\n\nSDS-PAGE and western blot analysis as shown in Figure 4-5 and Figure 4-6, respectively, reveal one major species at an apparent molecular weight of ~140 kDa after Protein-A affinity purification of 293T conditioned supernatants; this compares favourably with the theoretical mass of ~120 kDa. A second species was observed in western blots at ~46 kDa, presumably corresponding to a truncated translation product as observed in S2 cell conditioned supernatants. Cleavage of the fusion protein, in situ, with 3C protease yielded soluble LIFR-FN3 running at an apparent molecular weight of ~114 kDa. This compares with the theoretical mass of the LIFR-FN3-3C-Fc moiety of ~94 kDa; the disparity due to retarded passage through polyacrylamide due to glycosylation.\n\nIndeed it has been predicted that LIFR is glycosylated at many of the 20 potential N-glycosylation sites (Gearing et al., 1991). In order to limit microheterogeneity and conformational flexibility associated with glycoproteins, deglycosylation of LIFR-FN3 was attempted. Previous attempts at deglycosylating LIFR-FN3 using \u03b2-N-acetylglucosidase F (endo F; cleaves glycosidic linkage between the two core N-acetylglucosamine residues) proved unsuccessful (K. Chobotova, CRC Growth Factors Group, Birmingham). Treatment of cells with N-butyldeoxynojirimycin (Karlson et al., 1993; inhibits \u03b1-glucosidase I activity thus preventing formation of complex N-linked oligosaccharides so allowing unprocessed sugars to be removed with endo H) resulted in poor levels of cell growth and expression (data not shown).\n\nFigure 4-9 demonstrates that deglycosylation of LIFR-FN3 was successfully achieved using both recombinant (Barsomian et al., 1990) and non-recombinant PNGase F (New England Biolabs). The recombinant PNGase F was shown to be active as both a native maltose-binding fusion protein and in its cleaved form. It was also demonstrated that DTT from the 3C protease cleavage step should also be removed to obtain optimal levels of deglycosylation as PNGase F activity is modulated by DTT. However, western blot analysis as shown in Figure 4-9 reveals that non-recombinant PNGase F is the most effective with comparable enzyme levels to the recombinant form resulting in the largest band shift and a less diffuse band. This suggests more complete deglycosylation and improved homogeneity is produced by the non-recombinant preparation. Commercially supplied non-recombinant PNGase F was used for the purposes of this study although this supply is not financially feasible for large-scale deglycosylation. An alternative 'home-made' endoglycosidase preparation extracted from Flavobacterium meningosepticum containing endo F1, endo F2, endo F3 and PNGase F activity has been demonstrated to show better activity than commercial supplies (Baker et al., 1994). This extract may provide a more cost-effective solution to the problem of deglycosylation.\n\nPNGase F functions in cleaving the N-acetylglucosaminylasparagine bond of asparagine-linked oligosaccharides to yield an aspartic acid residue, ammonia and free reducing oligosaccharide. Such large-scale mutation of asparagine to aspartate may have considerable implications for protein folding, stability, purification and crystallisation. For similar reasons PCR mutagenesis of potentially glycosylated asparagine residues has previously been avoided, as it is suspected that co-translational glycosylation may be required for correct folding and expression. Analysis of the isoelectric focussing data in Figure 4-7 suggests that this is indeed the case with only the fully glycosylated fusion protein running as one fairly tight population with a pI of between 5 and 6. Deglycosylation, cleavage or acid elution of the fusion protein results in disruption of this pI profile and gives rise to a continuous distribution of protein species with pI ranging from 4 to 8. It can be concluded that both sugars and the Fc moiety have a stabilising effect on the LIFR-FN3 protein.", "# Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\nAuthors: Ye, Q.; Shvartsbart, A.; Li, Z.; Gan, P.; Policarpo, R.L.; Qi, C.; Roach, J.J.; Zhu, W.; McCammant, M.S.; Hu, B.; Li, G.; Yin, H.; Carlsen, P.; Hoang, G.; Zhao, L.; Susick, R.; Zhang, F.; Lai, C.-T.; Allali Hassani, A.; Epling, L.B.; Gallion, A.; Kurzeja-Lipinski, K.; Gallagher, K.; Roman, V.; Farren, M.R.; Kong, W.; Deller, M.C.; Zhang, G.; Covington, M.; Diamond, S.; Kim, S.; Yao, W.; Sokolsky, A.; Wang, X.\nSource: Journal of Medicinal Chemistry, 2025\n\n**Step 5. 3-(1-((1R,4R,5S)-2-Azabicyclo[2.1.1]hexan-5-yl)-6-fluoro-7-(7-fluoronaphthalen-1-yl)-4-methyl-2-((2-oxopiperidin-1-yl)methyl)-1H-pyrrolo[3,2-c]quinolin-8-yl)propanenitrile (16).** To a mixture of 45 (29 mg, 0.038 mmol), 2-(7-fluoronaphthalen-1-yl)-4,4,5,5-tetramethyl-1,3,2-dioxaborolane (22 mg, 0.76 mmol), XPhos Pd G2 (6 mg, 7.6 \u03bcmol) and potassium phosphate, tribasic (24 mg, 0.11 mmol) were added dioxane/water (4:1, 1 mL) and the reaction was sparged with nitrogen, then stirred at 80 \u00b0C for 1 h. The reaction was quenched with water and extracted with ethyl acetate. The organic layer was washed with water and brine, dried over sodium sulfate and concentrated. The crude product was dissolved in 1:1 DCM/TFA (1 mL) and stirred at r.t. for 15 min, then diluted with MeOH and purified by prep-LCMS (XBridge C18 column, eluting with a gradient of acetonitrile/water containing 0.1% TFA, at flow rate of 60 mL/min) to afford the desired product. LCMS calculated for C37H36F2N5O2 (M + H)+: m/z = 620.3; found: 620.3. 1H NMR (500 MHz, DMSO, 343 K) \u03b4 9.3 (bs, 1H), 8.26\u22128.12 (m, 3H), 7.70 (dd, J = 8.2, 7.0 Hz, 1H), 7.64 (d, J = 7.0 Hz, 1H), 7.50 (td, J = 8.7, 2.6 Hz, 1H), 7.00 (s, 1H), 6.91 (s, 1H), 5.46 (d, J = 2.8 Hz, 1H), 5.07 (s, 1H), 4.78 (s, 2H), 4.01 (s, 1H), 3.45 (s, 1H), 3.32 (bs, 3H), 2.97 (m, 1H), 2.85 (s, 3H), 2.74 (dt, J = 14.2, 7.2 Hz, 1H), 2.67 (d, J = 7.2 Hz, 1H), 2.44\u22122.36 (m, 2H), 2.32 (d, J = 9.2 Hz, 1H), 2.30 (d, J = 8.9 Hz, 1H), 1.82 (s, 4H), 1.60 (d, J = 9.3 Hz, 1H).\n\n### Synthesis of Compound 18\n\n**Step 1. tert-Butyl (1R,4R,5S)-5-(7-Bromo-8-(2-cyanoethyl)-6-fluoro-4-(methylthio)-2-((3-oxomorpholino)methyl)-1H-pyrrolo[3,2-c]quinolin-1-yl)-2-azabicyclo[2.1.1]hexane-2-carboxylate (46).** This compound was prepared using Steps 1 and 2 from the synthesis of compound 16 using 4-(prop-2-yn-1-yl)morpholin-3-one as the coupling partner in the Sonogashira coupling. LCMS calculated for C30H34BrFN5O4S (M + H)+: m/z = 658.2, 660.2; found: 658.2, 660.2.\n\n**Step 2. tert-Butyl (1R,4R,5S)-5-(8-(2-Cyanoethyl)-6-fluoro-7-(7-fluoronaphthalen-1-yl)-4-(methylthio)-2-((3-oxomorpholino)methyl)-1H-pyrrolo[3,2-c]quinolin-1-yl)-2-azabicyclo[2.1.1]hexane-2-carboxylate (47).** To a mixture of 46 (346 mg, 0.525 mmol), 2-(7-fluoronaphthalen-1-yl)-4,4,5,5-tetramethyl-1,3,2-dioxaborolane (172 mg, 0.63 mmol), XPhos Pd G2 (41 mg, 0.053 mmol) and potassium phosphate, tribasic (223 mg, 1.05 mmol) were added dioxane/water (12 mL, 5:1) and the reaction flask was sparged with nitrogen, then stirred at 80 \u00b0C for 30 min. The reaction was quenched with water and extracted with ethyl acetate. The organic layer was washed with water and brine, dried over sodium sulfate and concentrated. The crude product was purified by column chromatography (0\u2212100% ethyl acetate in hexanes) to provide the desired product. LCMS calculated for C40H40F2N5O4S (M + H)+: m/z = 724.3; found: 724.3.\n\n**Step 3. 3-(1-((1R,4R,5S)-2-Azabicyclo[2.1.1]hexan-5-yl)-6-fluoro-7-(7-fluoronaphthalen-1-yl)-4-(2-methylpyridin-4-yl)-2-((3-oxomorpholino)methyl)-1H-pyrrolo[3,2-c]quinolin-8-yl)propanenitrile (18).** To a solution of 47 (20 mg, 0.028 mmol) in dioxane (0.5 mL) were added Pd(PPh3)4 (16 mg, 0.014 mmol), copper 2-methylsalicylate (21 mg, 0.10 mmol) and 2-methyl-4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)pyridine (15 mg, 0.69 mmol) and the reaction was sparged with nitrogen, then stirred at 120 \u00b0C for 1 h. The reaction was cooled to r.t. and quenched with sat. aq. ammonium hydroxide and extracted with DCM. The organic layer was dried over sodium sulfate and concentrated. The residue was dissolved in 1:1 TFA/DCM (1 mL) and stirred at r.t. for 45 min, then diluted with MeOH and purified by prep-LCMS (XBridge C18 column, eluting with a gradient of acetonitrile/water containing 0.1% TFA, at flow rate of 60 mL/min) to afford the desired product as the corresponding TFA salt (1.2 mg, 6.6%). LCMS calculated for C40H35F2N6O2 (M + H)+: m/z = 669.3; found: 669.3. 1H NMR (600 MHz, DMSO) \u03b4 9.42 (s, 1H), 8.82 (d, J = 5.6 Hz, 1H), 8.33 (d, J = 12.3 Hz, 1H), 8.26\u22128.16 (m, 2H), 8.13\u22127.98 (m, 3H), 7.77\u22127.66 (m, 2H), 7.55 (dtd, J = 14.4, 8.7, 2.6 Hz, 1H), 7.22\u22127.03 (m, 1H), 6.89 (dd, J = 10.7, 2.6 Hz, 1H), 5.63\u22125.50 (m, 1H), 5.14 (dd, J = 15.8, 10.8 Hz, 1H), 4.94 (d, J = 6.0 Hz, 1H), 4.67 (dd, J = 15.9, 9.8 Hz, 1H), 4.21\u22124.12 (m, 3H), 3.85 (t, J = 5.1 Hz, 2H), 3.47 (s, 1H), 3.31 (dq, J = 13.1, 4.7 Hz, 1H), 3.18 (ddd, J = 12.0, 6.4, 4.2 Hz, 1H), 3.10\u22123.00 (m, 1H), 2.91 (ddd, J = 12.1, 7.4, 4.3 Hz, 1H), 2.81\u22122.54 (m, 6H), 2.34 (t, J = 10.1 Hz, 1H), 1.61 (d, J = 9.2 Hz, 1H).\n\n### Synthesis of Compound 22\n\nThis compound was prepared using the procedures described for compound 23, using 7-fluoronaphthyl boronic acid, pinacol ester in the Suzuki cross coupling step. LCMS calculated for C40H36F2N7O2 (M + H)+: m/z = 684.3; found: 684.3. 1H NMR (TFA salt, mixture of atropisomers, rotamers, 600 MHz, DMSO-d6) \u03b4 9.59 (m, 1H), 9.49 (m, 0.7H), 9.20 (m, 0.3H), 8.82 (m, 1H), 8.40\u22128.30 (m, 1.5H), 8.27\u22128.16 (m, 2.5H), 7.87\u22127.61 (m, 4H), 7.55 (m, 1H), 7.16 (m, 0.5H), 6.92 (m, 0.5H), 5.93 (m, 0.7H), 5.77 (m, 0.3H), 5.67 (m, 0.3H), 5.44 (m, 0.7H), 5.21\u22124.89 (m, 1H), 4.35\u22124.12 (m, 2.3H), 4.05\u22123.38 (m, 3.3H), 3.14\u22121.86 (m, 11H), 1.74\u22121.46 (m, 4.4H).\n\n## Associated Content\n\nThe Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jmedchem.4c02662: detailed synthetic protocols, characterization of final compounds, protocols for in vitro and in vivo assays, mean and standard deviations for assays, X-ray parameters for compound 3 and 5a (PDF); table of molecular formula strings (CSV).\n\n### Accession Codes\n\nPDB depositions are available for compound 3 (9E5F) and 5a (9E5D). Authors will release the atomic coordinates and structure factors upon article publication.\n\n## Author Information", "# Structure\u2013Function Analysis of the Extended Conformation of a Polyketide Synthase Module\nAuthors: Li, X.; Sevillano, N.; La Greca, F.; Deis, L.; Liu, Y.-C.; Deller, M.C.; Mathews, I.I.; Matsui, T.; Cane, D.E.; Craik, C.S.; Khosla, C.\nSource: Journal of the American Chemical Society, 2018\n\n**Authors:** Xiuyuan Li, Natalia Sevillano, Florencia La Greca, Lindsay Deis, Yu-Chen Liu, Marc C. Deller, Irimpan I. Mathews, Tsutomu Matsui, David E. Cane, Charles S. Craik and Chaitan Khosla (corresponding author, khosla@stanford.edu). **Affiliations (condensed):** Department of Chemistry, Stanford University, Stanford, California 94305 (Li, Liu, Khosla); Department of Biochemistry, Stanford University (Deis, Khosla); Department of Chemical Engineering, Stanford University (Khosla); Stanford ChEM-H, Stanford University (Deller, Khosla); Department of Pharmaceutical Chemistry, University of California San Francisco, San Francisco, California 94158 (Sevillano, La Greca, Craik); Stanford Synchrotron Radiation Lightsource, SLAC National Accelerator Laboratory, Stanford University, Menlo Park, California 94025 (Mathews, Matsui); Department of Chemistry, Brown University, Providence, Rhode Island 02912-9108 (Cane). **Published:** Journal of the American Chemical Society, volume 140, issue 21, pages 6518\u20136521, May 30, 2018. Author manuscript available in PMC May 30, 2019.\n\nThe authors declare no competing financial interest. Supporting Information (experimental methods and supporting figures and tables) is available on the ACS Publications website at DOI 10.1021/jacs.8b02100.\n\n## Abstract\n\nCatalytic modules of assembly-line polyketide synthases (PKSs) have previously been observed in two very different conformations\u2014an \"extended\" architecture and an \"arch-shaped\" architecture\u2014although the catalytic relevance of neither has been directly established. By the use of a fully human na\u00efve antigen-binding fragment (Fab) library, a high-affinity antibody was identified that bound to the extended conformation of a PKS module, as verified by X-ray crystallography and tandem size-exclusion chromatography\u2013small-angle X-ray scattering (SEC\u2013SAXS). Kinetic analysis proved that this antibody-stabilized module conformation was fully competent for catalysis of intermodular polyketide chain translocation as well as intramodular polyketide chain elongation and functional group modification of a growing polyketide chain. Thus, the extended conformation of a PKS module is fully competent for all of its essential catalytic functions.\n\n## Main Text\n\nAssembly-line polyketide synthases (PKSs), such as 6-deoxyerythronolide B synthase (DEBS), are multifunctional, multimodular enzymes that synthesize structurally diverse natural products.[1\u20133] Two truncated derivatives of DEBS are shown in Figure 1.\n\n*Figure 1.* (A) Bimodular and (B) trimodular derivatives of DEBS. Domains are shown as spheres. KS = ketosynthase; AT = acyl transferase; ACP = acyl carrier protein; L = KS-AT linker domain; KR = ketoreductase; KR0 = redox-inactive, epimerase-active KR homologue; TE = thioesterase. Blue tabs depict docking domains that facilitate non-covalent association of successive modules. Module 2 domains are colored for ease of comparison with Figure 2.\n\nAlthough high-resolution structures of individual domains and multidomain fragments from several PKS modules have been solved, no X-ray structures of intact modules are available.[4] Complete understanding of PKS structure\u2013function relationships is further challenged by the possibility of large, catalytically relevant conformational changes within a module. Indeed, two very different conformations have been proposed. Small-angle X-ray scattering (SAXS) analysis of DEBS module 3 supported a conformation with the ketosynthase\u2013acyltransferase (KS-AT) core in the same extended conformation as that previously observed in the crystal structures of two stand-alone KS-AT fragments[5] (Figure 2A). In contrast, cryo-electron microscopy (cryo-EM) analysis of module 5 of the related pikromycin synthase revealed an arch-shaped conformation with the homodimeric KS as the capstone and each AT domain oriented downward to form an extensive interface with the KS of the opposite dimer[6] (Figure 2B). Hypothetically, these apparent differences might be reconciled if one conformation corresponds to the state of the module during intermodular translocation of the polyketide chain (step I, Figure 3) while the other reflects the chain elongation state (step III, Figure 3). To address this question, we have now trapped individual PKS modules in defined conformational states and analyzed their catalytic properties.\n\n*Figure 2.* Alternative conformational models of homodimeric PKS modules: (A) SAXS model of DEBS module 3 + TE;[5] (B) cryo-EM model of module 5 of the pikromycin synthase.[6] Docking domain (\"coil\"), blue; KS, green; KS-AT linker, cyan; AT, pink; KR, yellow; ACP, orange; TE, brown. The collinear domain organization is as shown in Figure 1. The two conformations differ primarily in the orientation of the AT and L domains relative to the KS.\n\n*Figure 3.* Catalytic cycle of DEBS module 2. Domains are colored as in Figure 1. The cycle includes four reactions: translocation of the growing polyketide chain (I) into and (IV) out of the module, (II) transacylation of an extender unit from methylmalonyl-CoA to its ACP domain, and (III) chain elongation via decarboxylative C\u2013C bond formation. Module 2 also reduces the \u03b2-keto thioester product of chain elongation; this reaction is lumped into step III.\n\nIn this study, we chose DEBS modules 2 and 3 as representative proteins, principally because of (i) their shared organization and stereochemical control with PikAIII, (ii) the fact that activity assays and reference kinetic data for these modules are well-established, (iii) our early success in isolating antibodies against them, and (iv) the solved X-ray structure of the KS/AT fragment from module 3, which enabled structural analysis of its antibody-bound complex.", "# GitHub README: GOBSMACKED\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/GOBSMACKED, 2026\n\n> **Fold, dock, relax and annotate a protein-ligand complex, then check it against the experimental structure.**\n\n[](https://gobsmacked.mdeller.com)                      \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>App</b></td><td><a href=\"https://gobsmacked.mdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">gobsmacked.mdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/GOBSMACKED\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/GOBSMACKED</a></td>\n</tr>\n</table>\n\n---\n\n**GOBSMACKED** (Ground-truth Overlay for Binding Sites, Modes And Complex Kinetics/Dynamics) takes a protein and a ligand, folds and docks and relaxes them, and then does the thing most docking pipelines skip: it goes and finds the crystal structure, superposes on the binding pocket, and tells you how close you got and why.\n\n**Why it matters:** a docking score is a ranking, not a measurement, and a pretty predicted complex looks exactly the same whether it is right or wrong. GOBSMACKED answers three separate questions about one prediction: how close the pose lands to the crystal (PIER REVIEW), whether molecular dynamics recovers the induced fit that an apo-like predicted pocket is missing (HOLOGRAM), and whether the binding mode the prediction implies is the binding mode the crystal shows (GATEKEEPER). It is useful for: anyone validating a docking protocol before trusting it on a target with no structure, anyone asking whether ESMFold plus docking is good enough for a particular pocket, and anyone who wants the answer as a graded scorecard rather than as a folder of PDB files.\n\n---\n\n## \ud83e\udded How it works\n\nThe heavy compute does not run on the server. ESMFold, the PandaDock GNN and OpenMM need a GPU and several gigabytes; the host is a shared CPU droplet with 3.8 GB. So the work is split in two, with an archive passing between them.\n\n| Stage | Where | What happens |\n|---|---|---|\n| **Prepare** | droplet, CPU | Resolve the input, fetch the best available structure, annotate the family, pick the pocket, choose a reference crystal, emit `run_bundle.tar.gz` |\n| **Run** | your machine, GPU | `pixi run gobsmacked`: fold (if needed), prep, dock, minimise and run MD, score the affinity before and after, summarise, emit `results.tar.gz` |\n| **Analyze** | droplet, CPU | Validate the archive, superpose on the pocket, run PLIP and PandaMap, grade, classify the binding mode, draw the trajectory |\n\nNothing in the bundle contacts the server. The campaign file goes in, the results archive comes back, and both are validated against a schema so a failed stage never turns into a puzzling analysis.\n\n## \ud83d\ude80 Quick start\n\nOpen [gobsmacked.mdeller.com](https://gobsmacked.mdeller.com), paste a UniProt accession and a SMILES, pick a pocket, and download the bundle. Then, on a machine with a GPU:\n\n```bash\ncurl -fL \"https://gobsmacked.mdeller.com/runs/<job-id>/bundle\" | tar xz \\\n  && cd run_bundle_<job-id> && ./run.sh\n```\n\nPrepare prints that line with the job filled in and a copy button. One step, and the only\nprerequisites are `curl` and `bash`: the bundle carries its own `pixi.lock`, and `run.sh`\ninstalls pixi if the machine has not got it, builds the environment from the lock (no solve, so\nthe versions are the ones the bundle was tested with) and runs all six stages. The environment\nlives in `.pixi/` inside the bundle directory, so nothing is installed system-wide.\n\nUpload `results.tar.gz` on the Analyze tab when it finishes. It is written at the top level of the\nbundle directory, next to `run.sh`, rather than inside `results/`: the archive is the one file the\nrun exists to produce and nobody should have to go looking for it.\n\nTo run the web application locally:\n\n```bash\ngit clone https://github.com/bellcheddar/GOBSMACKED.git\ncd GOBSMACKED\nuv venv --python 3.11 .venv\nuv pip install --python .venv/bin/python -r requirements.txt\nmake serve                   # http://127.0.0.1:8009\n```\n\n## \ud83d\udd27 Prepare\n\nFour panels, each unlocking the next.\n\n| Panel | Accepts | What it does |\n|---|---|---|\n| **1. Protein** | UniProt accession, PDB ID, raw sequence, or a dropped `.pdb` / `.cif`, plus an optional domain range | Resolves the canonical sequence and picks a starting structure by a fixed, reported priority: your upload > a named PDB entry > AlphaFold DB > ESM Atlas > fold it in the bundle. **From Pfam** fills the range from the domain the pocket sits in |\n| **2. Annotation** | (automatic) | InterPro supplies Pfam domains, which route the target: `PF00069`/`PF07714` to KLIFS, `PF00001`/`PF00002`/`PF00003`/`PF10324` to GPCRdb, anything else to UniProt features alone |\n| **3. Ligand and pocket** | SMILES, plus residues clicked in Mol* or on the sequence track | RDKit validates and draws the ligand; the docking box is the selection's extent plus 8 \u00c5, floored at 18 \u00c5 per side |\n| **4. Reference** | (automatic, or a typed PDB ID) | RCSB entries mapped to the accession, ranked by Morgan Tanimoto against your ligand, with resolution and a 2D depiction |\n\n**Trim to the domain.** An AlphaFold model is of the whole precursor. EGFR's is 1,210 residues, of which the kinase domain is 253: solvating the other 957 costs an order of magnitude in MD time, adds two disordered tails that wander through the box, and tells you nothing about the pocket.\n\n**The reference's site is renumbered onto your model.** 1M17 numbers EGFR from the mature protein and UniProt from the precursor, 24 apart. Pasting the crystal's residue list straight into the pocket picker would select real residues that are the wrong ones, which is worse than an error, so both structures are aligned by sequence first.", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nDNA encoding the target protein, 6\u00d7HIS-TEV-WRN-517-945, was produced, cloned and transfected as detailed in Example D. 6 L batches of insect cell culture were produced and processed as detailed in Example D. The cell pellets were thawed and resuspended in IMAC buffer A (50 mM HEPES pH 7.5, 500 mM NaCl, 10% glycerol, 10 mM imidazole, 0.5 mM TCEP) supplemented with DNase, 1 mM PMSF and Complete protease inhibitor tablets (Roche). The resuspended cells were sonicated on ice for 10 min using 5 second on/off pulses at 30% amplitude. Cells were clarified using ultracentrifugation at 120,000\u00d7g for 60 min at 4 \u00b0C.\n\nThe cleared lysate was loaded on to 3\u00d75 ml HisTrap Excel columns (Cytiva) pre-equilibrated with IMAC buffer A and the column was washed with an increasing gradient of 2%, 4% and 6% of IMAC buffer B (IMAC buffer A+500 mM imidazole). Elution of the bound proteins was performed using a linear gradient of IMAC buffer B ranging from 6% to 100% imidazole (30-500 mM imidazole). Fractions containing the protein of interest were combined and supplemented with TEV protease at a ratio of 1:20 TEV:protein (w/v), and dialyzed against 1 L of IMAC buffer A. After overnight cleavage (16 hrs), the sample was centrifuged to remove aggregates and loaded on to a HisTrap Excel column to capture the TEV protease, the His-tagged GST fragment, and the uncleaved material.\n\nThe flow-through and wash fractions from the IMAC purification were pooled and concentrated using a centrifugal concentrator with a 30 kDa molecular weight cutoff (Millipore). The concentrated sample was injected on to a HiLoad 16/600 Superdex 200 preparative Size Exclusion Chromatography column (Cytiva) equilibrated in SEC buffer (25 mM HEPES pH 7.5, 300 mM NaCl, 10% glycerol, 1 mM TCEP). Fractions were collected and analyzed using SDS-PAGE. Fractions corresponding to the WRN protein were combined and concentrated using a centrifugal concentrator with a 30 kDa molecular weight cut-off to 12-16 mg/mL (Millipore). The protein was aliquoted, flash frozen in liquid nitrogen, and stored at \u221280 \u00b0C.\n\n### Co-Crystallization with WRN Inhibitor Compound\n\nCrystallization of the purified WRN protein with a WRN inhibitor compound was performed using sitting drop vapor diffusion techniques. Crystals were obtained at 4 \u00b0C. from the Wizard Screen (Molecular Dimensions) condition consisting of 0.1 M MES pH 6.0, 0.2 M calcium acetate and 20% (w/v) PEG 8000. The condition was optimized by varying pH, calcium acetate and PEG concentrations. Furthermore, an additive screen was set up and the additives ammonium sulfate and sodium malonate were found to produce larger and more robust crystals for data collection. For co-crystal determination, WRN protein was co-crystallized in the presence of 2 mM of the WRN inhibitor compound. The crystals were subsequently cryo-protected in precipitant solution supplemented with 2 mM of the WRN inhibitor compound and 25% glycerol before flash-cooling in liquid nitrogen.\n\nData was collected at beamline I03, Diamond Light Source, Didcot, UK, at 100 K and \u03bb=0.9763 \u00c5. 3600 images were collected with an oscillation range of 0.1\u00b0 per image. The beamline was equipped with a Dectris Eiger2 XE 16M detector. The data were processed using the autoPROC pipeline (1), which includes the software XDS (2), Aimless (3), and Pointless (4).\n\n### Structure Refinement of Co-Crystal Structure\n\nThe WRN protein structure was determined by molecular replacement using Phaser (5) using a truncated model of the published WRN (PDB ID: 6YHR). To enable successful molecular replacement, each domain of the WRN protein was placed independently. The structure was refined using Buster (6) and model building was carried out in Coot (7). One WRN monomer was located in the asymmetric unit and initial difference density maps revealed that the WRN inhibitor compound was bound to the protein. Ligand coordinates and restraint files for use in Buster were generated using Grade2 from Global Phasing Ltd (8).\n\nThe structure of WRN in complex with the WRN inhibitor compound was determined to 3.1 \u00c5 resolution and includes amino acid residues 526-925 and 929-945 in chain A. The missing residues were all located in loop regions and were not visible in the electron density. Furthermore, one zinc ion and three sulphate ions were modelled into the electron density. The compound was overall well-defined in the electron density.\n\n### Analysis of Co-Crystal Structure\n\nThe co-crystal structure of the WRN helicase core bound to the WRN inhibitor compound was determined to 3.1 \u00c5 resolution. Analysis of the structure revealed an unexpected conformation, with a 150 degree rotation of the ATPase 2A domain relative to the ATPase 1A domain of the helicase core around the flexible linker. Previously published WRN helicase structures, including RCSB PDB ID: 6YHR, did not display this inter-domain rotation.\n\nThe 150 degree rotation of domain 2A of the helicase core results in rearrangement of residues necessary for ATP hydrolysis; specifically, the conserved Walker A motif (which is traditionally observed in the SF2 family of helicase proteins) is repositioned to form an unexpected pocket which was not included in any structures of the WRN helicase deposited in the Protein Data Bank. The conformational rearrangement also resulted in a new binding site in which Cys727 is in close proximity (3.3 angstroms) to the WRN inhibitor compound. In contrast, a corresponding linear sequence alignment of Walker A and other canonical motifs suggests Cys580 (within the Walker A motif) to be involved with the WRN inhibitor compound, and not the novel Cys727 binding site found in this co-crystal analysis. Additional novel interacting residues at Arg857 (Motif VI, the arginine finger, ATP binding site, (2)) and Phe730 (linker region between domains 1A and 2A) were also identified.\n\n### Example F. Mutant Panel Shows Decoupling of ATPase and Helicase Activity Through Cys 727", "# Discovery of an Orally Bioavailable KRAS G12D Inhibitor_2\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [9E5F](https://www.rcsb.org/structure/9E5F)\n* **Release Date:** 2025-01-22\n* **Method:** X-RAY DIFFRACTION 1.35 \u00c5\n* **Organisms:** Homo sapiens\n* **Macromolecules:** GTPase KRas (protein)\n* **Unique Ligands:** A1BEJ (Compound 23), GDP, MG\n* **Publication:** [Deller, M.C., et al. (2025). Discovery of an Orally Bioavailable KRAS G12D Inhibitor. Journal of Medicinal Chemistry, 68, 1924-1939.](https://pubs.acs.org/doi/10.1021/acs.jmedchem.4c02662)\n* **Blog Post:** [Discovery of INCB159020: An Orally Bioavailable KRAS G12D Inhibitor](https://marcdeller.com/discovery-of-incb159020-an-orally-bioavailable-kras-g12d-inhibitor/)\n\n**Description:** A companion structure to [9E5D](https://www.rcsb.org/structure/9E5D), this 1.35 \u00c5 resolution crystal captures KRAS G12D in complex with Compound 23, an alternative inhibitor from the same drug discovery campaign. Together with 9E5D, it maps the SAR landscape that enabled optimization of orally bioavailable KRAS G12D inhibitors for oncology.", "# GitHub README: chatPDB\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/chatPDB, 2026\n\n> **A protein-structure-literate AI assistant for the Protein Data Bank: RAG for facts, a QLoRA-tuned model for behaviour, and live Biopython/gemmi/DSSP/PyMOL tool calls for structural truth. Fine-tuned and served locally on Apple Silicon with MLX-LM.**\n\n[](https://chatpdb.mdeller.com)      \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://chatpdb.mdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">chatpdb.mdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/chatPDB\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/chatPDB</a></td>\n</tr>\n</table>\n\n---\n\nchatPDB is a locally hosted, open-source assistant that reasons about protein structures (file\nformats, experimental methods, cross-reference databases, structure-manipulation tools), drives\nthe structural-biology tools a working structural biologist actually uses (Biopython, gemmi, DSSP,\nPyMOL, ChimeraX, and more), and grounds every factual claim in the real wwPDB/RCSB/SIFTS/UniProt\ncorpus rather than memory.\n\n**Why it matters:** off-the-shelf models invent structural facts (wrong resolutions, hallucinated\nPDB IDs, fabricated secondary-structure content) and have no way to compute a real value rather\nthan guess one. chatPDB fixes both by separating concerns: volatile knowledge lives in retrieval,\nstable behaviour lives in the fine-tuned weights, and deterministic structural facts (atom counts,\nsecondary structure, chain composition, geometry) are computed live by real tools, executed in a\nsandbox, rather than asserted from memory. It is useful for: interrogating deposited PDB structures\nand their cross-references, generating correct Biopython/gemmi/DSSP/PyMOL scripts, honestly\ndeclining to answer when an ID doesn't exist or a value isn't known, and reasoning like a\nstructural biologist about experimental method, quality, and biological assembly rather than\ntreating a structure as a bare coordinate file. A live, no-install terminal session against\nthe real fine-tuned model runs at **[chatpdb.mdeller.com](https://chatpdb.mdeller.com)** (see\n**Web deployment** below).\n\nSibling project to [ChemSage](https://github.com/bellcheddar/ChemSage), reusing its proven\nMLX-LM/QLoRA/RAG process on a fresh domain: protein structures instead of small-molecule chemistry.\n\n**Project links:**\n\n- [GitHub](https://github.com/bellcheddar/chatPDB)\n- [Models on Hugging Face](https://huggingface.co/Dellboy)\n\n## \ud83e\udd17 Models on HuggingFace Hub\n\n| Model | HuggingFace | Base | SFT dataset | Round |\n|---|---|---|---|---|\n| **chatPDB 32B v1** \u2b50 | [**Dellboy/chatpdb_32b_v1**](https://huggingface.co/Dellboy/chatpdb_32b_v1) | Qwen3-32B-4bit | v7 (97,272 examples) | **Round 1 (current)** |\n\nAdd a row here for every future fused model (never overwrite a previous row) \u2014 see\n[\ud83c\udf93 Training](#-training) below for how each round's real numbers get recorded.\n\n## \u2699\ufe0f The hybrid in one line\n\nRAG keeps it truthful today; the fine-tune makes it consistent tomorrow; Biopython/gemmi/DSSP/PyMOL\nmake it correct always.\n\n## \ud83d\ude80 Quick start\n\nSee **[`PROJECT_PLAN.md`](PROJECT_PLAN.md)** for the full step-by-step build (phases 0 to 10). Short version:\n\n| Step | What | Script / command |\n|---|---|---|\n| 0 | Env: MLX-LM + Biopython + structural tools | (`requirements.txt`) |\n| 1 | Base model (Qwen3-32B, MLX 4-bit) | `config/train_config.yaml` |\n| 2 | **RAG first** (ship this alone) | `scripts/ingest_rag.py` |\n| 3 | Build + validate SFT dataset | `scripts/build_dataset.py` |\n| 4 | QLoRA fine-tune | `python scripts/train_launch.py --config config/train_config.yaml` |\n| 5 | Fuse + serve | `python scripts/merge_export.py` \u2192 `mlx_lm.server --port 8080` |\n| 6 | Close the hybrid loop | `rag/tool_exec.py` (Biopython sandbox; gemmi/DSSP/PyMOL staged next) |\n| 7 | Structural evaluation | `python eval/eval_pdb.py` \u00b7 `python eval/compare/eval_compare.py` |\n| 8 | Local CLI + RAG | `python scripts/chat.py` |\n| 9 | Hosted demo (no local install) | live at [chatpdb.mdeller.com](https://chatpdb.mdeller.com), see **Web deployment** below |\n\n## \u25b6\ufe0f How to run\n\n**Interactive chat** (loads the model in-process \u2014 no server needed, unlike the eval scripts below):\n```bash\ncd /Users/dellboy/Documents/Vibe_Coding/chatPDB\nsource .venv/bin/activate\npython scripts/chat.py --model models/chatpdb_32b_v1\n```\n`/help` for slash commands, `/info` for session details, `quit`/`exit`/`q` to leave. `--no-rag` runs\nmodel-only (no retrieval).\n\n**Evaluate the model** (these two DO need `mlx_lm.server`; `eval_compare.py` manages its own):\n```bash\n# Terminal 1\npython scripts/preflight.sh   # flush memory, check for background-process contention\nmlx_lm.server --model models/chatpdb_32b_v1 --port 8080\n\n# Terminal 2, once Terminal 1 shows \"HTTP server listening\"\npython eval/eval_pdb.py --n 200                 # single-model scorecard, needs Terminal 1 running\npython eval/compare/eval_compare.py              # multi-round comparison, starts its own server\n```\n\n## \ud83c\udf93 Training\n\nReal numbers only, recorded per round \u2014 extend this table (never edit a previous round's row) as\nnew training rounds ship.", "# GitHub README: PANTS\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/PANTS, 2026\n\nThere is one partial way round that, and it has now been run rather than argued about.\nPAZy curates enzymes measured on **other** plastics too (PA, PUR, PLA, PBAT, PHA). Most are\ndifferent folds doing different jobs, and including a nylon amidase would be another easy\nnegative. But 26 of them **share a 30% cluster with a PET-active enzyme**, which puts them\ninside the polyesterase family boundary. That is the closest thing to the missing negative\nclass that exists.\n\nRunning the same head against negatives of increasing closeness, everything else fixed:\n\n| Negative regime | Positives | Negatives | Head AUC | Composition only |\n|---|---|---|---|---|\n| Near-miss (one ESTHER `Cutinase` family) | 305 | 110 | 0.997 \u00b1 0.006 | 0.890 |\n| Out-of-family (distant \u03b1/\u03b2-hydrolases) | 305 | 109 | 0.975 \u00b1 0.022 | 0.906 |\n| **Within-family** (measured on another plastic) | 305 | 26 | **0.850 \u00b1 0.055** | 0.651 |\n| **Within-family, mixed clusters only** | 152 | 26 | **0.493 \u00b1 0.297** | 0.398 |\n\n**Discrimination decays monotonically as the negatives get closer, and collapses to chance\nat the boundary that matters.** The head beats composition in every regime except the last.\nIn the strictest one, where every negative has a PET-active enzyme inside its own cluster,\nso family membership carries no information at all, it scores 0.493: a coin flip.\n\n**The conclusion does not depend on where the family boundary is drawn.** \"Inside the\nfamily\" has no single right answer, so the same test was run under three definitions:\n\n| Family definition | Negatives | Shared clusters | Head AUC | Composition |\n|---|---|---|---|---|\n| Shares a 30% cluster with a PET-active enzyme | 26 | 7 | 0.493 \u00b1 0.297 | 0.398 |\n| Hits the per-cluster profile library (the test recall applies to every candidate) | 12 | 3 | **0.279 \u00b1 0.168** | 0.413 |\n| Passes both | 12 | 3 | 0.279 \u00b1 0.168 | 0.413 |\n\nUnder the stricter, profile-based definition the head is not merely at chance but\n*inverted*: it ranks the within-family negatives **above** the PET-active enzymes. With 12\nnegatives across 3 shared clusters that is not evidence of a real inverse signal, and it\nshould not be read as one. What it does do is close off the remaining optimistic reading,\nbecause no choice of family boundary produces discrimination.\n\nA fourth definition was tried and rejected rather than reported: the pooled `PLC_all.hmm`\nprofile admits only 6 of the 155, and misses proteins literally named `Cutinase` and\n`CutL1`. That is the pooled profile's already-documented failure \u2014 it scored 0/111 on the\nnear misses, which is why the per-cluster library exists \u2014 and quoting it would have\nunderstated the negative set by reproducing a known bug.\n\n**The strictest rows are underpowered, and saying so is part of the result.** One cluster holds\n85% of the positives, so the cluster-grouped folds come out badly lopsided, and the\nper-fold AUCs are 0.193, 0.714, 0.208 and 0.857. A mean of 0.493 across that spread cannot\ndistinguish \"no signal\" from \"not enough data to detect one\". What it does rule out is a\nlarge, easily-learned signal, which the 0.976 headline might otherwise be read as implying.\n\nTwo caveats travel with these numbers permanently:\n\n- **The negatives are weak.** PAZy records only positive substrate associations, so \"PET\n  not listed\" conflates *inactive* with *never assayed*. Some fraction are false negatives,\n  which pushes measured discrimination **down**: the within-family figures are a lower\n  bound, not a point estimate.\n- **The composition baseline is doing more work than it appears.** It reaches 0.906 against\n  out-of-family negatives. Most of the headline 0.976 is therefore available from amino\n  acid composition alone, and the head's real contribution is the margin above that, not\n  the absolute number.\n\nThe practical conclusion is not that the approach fails, but that **the evaluation is\nlabel-limited rather than method-limited**, and the binding constraint is now a few dozen\nmeasured within-family negatives rather than any modelling choice.\n\n### \ud83d\udd2c And the same test on geometry\n\nGeometry is the one signal here measured off coordinates rather than inherited from an\nannotation, so it is the obvious thing to try on the contrast the embeddings failed.\nAlphaFold models were taken for the characterised enzymes that have them (131 measured:\n114 PET-active, 17 within-family negatives) and the active site measured on each. Mean\npLDDT is 90.4 against 91.3, so nothing below is a model-confidence artefact.\n\nCompared directly, several features separate the two groups convincingly:\n\n| Feature | PET-active | Within-family negatives | AUC | p |\n|---|---|---|---|---|\n| Cleft depth (\u00c5) | 4.20 | 3.41 | **0.819** | <0.001 |\n| Cleft-lining residues | 82.9 | 88.3 | 0.788 * | <0.001 |\n| Oxyanion donor 2 distance (\u00c5) | 4.56 | 4.95 | 0.742 * | 0.001 |\n| Oxyanion donor 1 distance (\u00c5) | 2.97 | 3.23 | 0.696 * | 0.009 |\n| Cleft width (\u00c5) | 19.38 | 21.78 | 0.686 * | 0.014 |\n| Triad Ser\u2013His (\u00c5) | 2.87 | 2.87 | 0.599 | 0.190 |\n\n\\* inversely predictive: the feature is *lower* in PET-active enzymes.\n\nPET-active polyesterases have deeper, narrower clefts with fewer lining residues and a\ntighter oxyanion hole. That is a physically sensible picture, and three of those p-values\nsurvive Bonferroni correction across the nine features tested.\n\n**It does not survive cluster-grouped evaluation.** Trained on all nine features and split\nby 30% cluster, exactly as every other number in this project:\n\n| | AUC | Folds |\n|---|---|---|\n| Geometry, cluster-grouped | **0.533 \u00b1 0.185** | 0.394 / 0.838 / 0.525 / 0.375 |\n| Sequence embeddings, same question | 0.493 \u00b1 0.297 | \u2014 |", "# BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n(use `--platform linux-64` instead of `osx-arm64` if the offline computer runs Linux)\n\n**2.** Copy `boltzmaker-installer.sh`, together with the whole `BoltzMaker` folder, onto the offline computer (USB drive, internal network transfer, etc.).\n\n**3.** On the offline computer, in a terminal, go to where you copied those files and run:\n\n```\n./boltzmaker-installer.sh -o ./boltzmaker-env\nsource ./boltzmaker-env/activate.sh\nexport KMP_DUPLICATE_LIB_OK=TRUE\n```\n\n**4.** Check it worked:\n\n```\npython3 BoltzMaker.py preflight examples/t4_lysozyme/boltz_input.md\n```\n\nFull details, including a couple of one-time extra steps this path needs, are in [`docs/tier_b_offline_install.md`](docs/tier_b_offline_install.md).\n\n## \ud83e\udde9 Architecture\n\n```\n+----------------+     +----------+     +-----------+     +-----+     +---------+\n| boltz_input.md | --> | generate | --> | preflight | --> | run | --> | analyze |\n+----------------+     +----------+     +-----------+     +-----+     +---------+\n```\n\n| Stage | Command | Produces |\n| --- | --- | --- |\n| 1. Input | \u2014 | `boltz_input.md` \u2014 the family x partners x ligand DSL spec |\n| 2. Generate | `generate` | `boltz_yamls/*.yaml` + manifest |\n| 3. Preflight | `preflight` | PASS/WARN/FAIL: CLI, GPU/MPS, disk, iCloud, YAML/SMILES/chain-id/chemistry, memory |\n| 4. Predict | `run` | `boltz predict` -> `boltz_output/`, live progress + memory monitor, resumable |\n| 5. Analyze | `analyze` | `boltz_summary.csv`/`.xlsx`, `boltz_dashboard.html`, `boltz_cif/`, plus interaction files if `setup-plip` has run, plus `boltz_sse_comparison.csv`/`.html` for any family with an `Apo structure:` set (see **compare-sse** below) |\n\nEach stage reads only the manifest + files the previous stage wrote, so any stage can be re-run on its own (`generate`, `preflight`, `run`, or `analyze` individually) without repeating the others: `all` simply chains all four.\n\n## \ud83d\udd27 One-time setup\n\n```\npython3 BoltzMaker.py setup\n```\n\nCreates a dedicated `.venv` (Python 3.12: boltz pins `numpy<2.0`, which has no prebuilt wheel for newer Pythons) next to `BoltzMaker.py` and installs `boltz`, `rich`, `pandas`, `openpyxl`, `pyyaml`, `rdkit`, `matplotlib`, `psutil`, `scipy`, `gemmi`, `biopython`, `plotly`, `reportlab`, and `requests` into it. This pulls PyTorch (~2-3 GB) and, the first time `boltz predict` actually runs, Boltz downloads several GB of model weights over the network. Every other command below transparently relaunches itself under this managed environment, so you can keep invoking the script with whatever `python3` is on your PATH.\n\nRe-run with `--force` to recreate the venv from scratch, or `-y/--yes` to skip the download confirmation prompt.\n\n### Optional: `setup-plip` (protein-ligand interaction analysis)\n\n```\npython3 BoltzMaker.py setup-plip\n```\n\nEntirely optional and separate from the venv above. Builds a `.plip_env` (via a self-downloaded [micromamba](https://micro.mamba.pm), ~1-1.5GB, mostly PyMOL\u2019s own Qt/ Cairo/HDF5 dependencies) and vendors a pinned commit of [cif2plip](https://github.com/bellcheddar/cif2plip), which converts a Boltz ModelCIF into a strict PDB and runs [PLIP](https://github.com/pharmai/plip) on it for real protein-ligand interaction fingerprints (H-bonds, salt bridges, pi-stacking, halogen bonds, metal coordination, etc.). A conda-forge-based environment is used deliberately here rather than pip: PLIP requires OpenBabel and PyMOL as in-process Python imports (not subprocess calls), and empirically, `plip`\u2018s own installer forces a broken from-source OpenBabel build unless a working OpenBabel is already present, while the standalone PyPI `pymol-open-source` wheel has a hardcoded broken library path on this platform \u2014 conda-forge\u2019s builds have neither problem.\n\nIf `.plip_env` isn\u2019t present, everything below degrades gracefully: `analyze` skips interaction analysis (dashboard looks exactly as it did before this feature), and the `new` wizard simply doesn\u2019t ask about reference structures. Nothing else in BoltzMaker requires it. `preflight`\u2018s `plip_env` check always reports which mode you\u2019re in.\n\n### Optional: `mkdssp`/`dssp` (for `compare-sse`\u2018s SSE-boundary-shift metric)\n\nNot bundled or installed by any BoltzMaker command \u2014 a small external binary you may already have (`brew install dssp` on macOS, or `conda install -c salilab dssp`). Only needed for one specific `compare-sse` metric (secondary-structure-element boundary shift, used when a structure has no deposited HELIX/SHEET records \u2014 true for every Boltz-predicted structure). Every other `compare-sse` metric works without it.\n\n### Alternative: `pixi` (one unified environment, macOS + Linux/CUDA)\n\n```\n./install.sh\n```\n\nInstalls [`pixi`](https://pixi.sh) if it isn\u2019t already on your machine, then solves and installs everything (`boltz`, `rdkit`, PLIP\u2019s OpenBabel/PyMOL stack, and the rest of `requirements.txt`) as **one** environment from `pixi.toml`/`pixi.lock`, replacing the need to separately run `setup` + `setup-plip` above. Reproducible (the committed `pixi.lock` pins every package on both `osx-arm64` and `linux-64`), and the only path here that\u2019s actually tested against Linux/CUDA rather than macOS/MPS alone. Once installed:\n\n```\npixi run preflight my_campaign.md\npixi run all my_campaign.md\n```\n\n(or `pixi shell` to activate the environment directly and run `python3 BoltzMaker.py ...` as normal). PLIP needs one extra one-time step, same spirit as `setup-plip` above: `pixi run postinstall`.\n\n**No internet on the target machine at all?** See [`docs/tier_b_offline_install.md`](docs/tier_b_offline_install.md) for building a single self-extracting installer script per platform (`pixi-pack --create-executable`) that needs no `pixi`, no `conda`, and no network to run.\n\n## \ud83e\uddea Examples\n\nFour entirely public-domain campaigns in `examples/`, run any of them with `python3 BoltzMaker.py all examples/<name>/boltz_input.md`:", "# 2013 \u2013 Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: papers).*\n\n* [Link to Paper](https://marcdeller.com/wp-content/uploads/2025/09/nihms464890.pdf)\n\n* **Multi-Modal Recognition**: The Asn332 supersite uniquely accommodates multiple antibody binding modes and angles of approach, making it more accessible than other conserved epitopes on HIV-1 envelope\n* **Evolutionary Constraint**: Viral attempts to escape supersite-targeting antibodies create fitness costs, exposing conserved regions and maintaining immune vulnerability despite selective pressure\n* **Glycan Processing Specificity**: The intrinsic mannose patch formation results from high glycan density limiting processing, creating distinctive molecular signatures for immune recognition\n* **Therapeutic Breadth**: Contemporary antibodies targeting the supersite demonstrate exceptional neutralization breadth against diverse HIV-1 strains, including challenging circulating recombinant forms\n* **Vaccine Target Validation**: The supersite\u2019s conservation and accessibility across multiple binding modes makes it an optimal target for structure-guided vaccine design approaches\n* **Dual Mechanism Activity**: Supersite-targeting antibodies combine virus neutralization with infected cell elimination through ADCC, providing comprehensive therapeutic potential\n\n1. Kong, L., et al.: [Kong, L., et al., Nature Structural & Molecular Biology, 2013](https://www.nature.com/articles/nsmb.2594)\n2. Becerra, J.C., et al.: [Becerra, J.C., et al., Microbial Cell, 2024](https://microbialcell.com/researcharticles/2024a-becerra-microbial-cell/)\n3. Stewart-Jones, G.B., et al.: [Stewart-Jones, G.B., et al., Nature Communications, 2015](https://www.nature.com/articles/ncomms8479)\n4. Frontiers in Immunology: [Frontiers Editorial Team, et al., Frontiers in Immunology, 2021](https://www.frontiersin.org/journals/immunology/articles/10.3389/fimmu.2021.755568/full)\n5. Nature Immunology: [Vaccination Study Team, et al., Nature Immunology, 2024](https://www.nature.com/articles/s41590-024-01833-w)\n6. Cell Reports: [Zhang, Y., et al., Cell Reports, 2024](https://www.sciencedirect.com/science/article/pii/S2211124724013615)\n\n**Breaking Through HIV\u2019s Impenetrable Glycan Fortress: The Asn332 Supersite Revelation**\n\nThe human immunodeficiency virus (HIV-1) has long been considered one of the most formidable pathogens due to its extraordinary ability to evade immune responses through sophisticated molecular camouflage. The virus cloaks its envelope glycoprotein gp120 with a dense layer of sugar molecules called glycans, creating what researchers term a \u201cglycan shield\u201d that conceals vulnerable regions from neutralizing antibodies. This protective barrier seemed nearly impenetrable until groundbreaking research revealed a critical weakness: a supersite of immune vulnerability centered around the Asn332 glycan that represents one of the most promising targets for HIV vaccine development and therapeutic intervention.\n\nThe discovery emerged from detailed structural studies of broadly neutralizing antibodies (bnAbs) that possess the remarkable ability to neutralize diverse HIV-1 strains despite the virus\u2019s extraordinary genetic variability. Unlike conventional antibodies that target only protein surfaces, these exceptional antibodies recognize glycan-dependent epitopes, essentially turning the virus\u2019s own protective shield against it. The Asn332 supersite represents a convergence point where multiple bnAbs can bind despite using completely different molecular recognition strategies, making it an unusually accessible and extensive target on the otherwise heavily fortified viral surface.\n\nAt the molecular level, the supersite encompasses a region surrounding the conserved N-linked glycan at asparagine residue 332, along with neighboring glycans at positions 392 and 386. This glycan cluster forms part of what scientists call the \u201cintrinsic mannose patch\u201d \u2013 a region where the high density of glycans limits their processing during biosynthesis, leaving behind under-processed oligomannose-type structures that serve as distinctive molecular signatures for immune recognition. The remarkable feature of this supersite is its ability to accommodate multiple binding modes and varied angles of approach, allowing different antibody lineages to target the same vulnerable region using distinct molecular strategies.\n\nRecent studies spanning 2024-2025 have provided unprecedented insights into how bnAbs penetrate the glycan shield to access this supersite. Advanced structural analyses reveal that antibodies like PGT135 use extended complementarity-determining region (CDR) loops, particularly the heavy chain CDR H1 and H3 loops, to reach through the glycan canopy and contact the underlying protein surface. These molecular probes must be precisely calibrated \u2013 typically extending 15-20 angstroms from the antibody surface to match the length of N-linked glycans protruding from the viral envelope. The antibodies achieve this remarkable feat through a combination of glycan recognition and protein contact, with PGT135 burying approximately 1,011 square angstroms of surface area across three glycans.\n\nThe supersite\u2019s accessibility stems from several unique structural features that distinguish it from other regions of the viral envelope. Unlike the CD4 binding site where small differences in antibody positioning result in dramatically different neutralization outcomes, the Asn332 supersite tolerates considerable variation in binding angles and contact points while maintaining neutralization efficacy. This tolerance reflects the supersite\u2019s extensive nature, encompassing six conserved glycans and the underlying protein surface in a region that spans much further than initially appreciated. The structural flexibility allows antibodies to accommodate the conformational and chemical diversity of glycans across different viral strains, adapting their angle of engagement to maintain effective binding.", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\n\"Thio\" refers to a group of formula \u2014SH. \"Cn-m alkylamino\" refers to a group of formula \u2014NH(alkyl). \"Cn-m alkoxycarbonyl\" refers to a group of formula \u2014C(O)O\u2014alkyl. \"Cn-m alkylcarbonyl\" refers to a group of formula \u2014C(O)\u2014alkyl. \"Cn-m alkylcarbonylamino\" refers to a group of formula \u2014NHC(O)-alkyl. \"Cn-m alkoxycarbonylamino\" refers to a group of formula \u2014NHC(O)O(Cn-m alkyl). \"Cn-m alkylsulfonylamino\" refers to a group of formula \u2014NHS(O)2-alkyl. \"Aminosulfonyl\" refers to a group of formula \u2014S(O)2NH2. \"Cn-m alkylaminosulfonyl\" refers to a group of formula \u2014S(O)2NH(alkyl). \"di(Cn-m alkyl)aminosulfonyl\" refers to a group of formula \u2014S(O)2N(alkyl)2. \"Aminosulfonylamino\" refers to a group of formula \u2014NHS(O)2NH2. \"Cn-m alkylaminosulfonylamino\" refers to a group of formula \u2014NHS(O)2NH(alkyl). \"di(Cn-m alkyl)aminosulfonylamino\" refers to a group of formula \u2014NHS(O)2N(alkyl)2. \"Aminocarbonylamino\", employed alone or in combination with other terms, refers to a group of formula \u2014NHC(O)NH2. \"Cn-m alkylaminocarbonylamino\" refers to a group of formula \u2014NHC(O)NH(alkyl). \"di(Cn-m alkyl)aminocarbonylamino\" refers to a group of formula \u2014NHC(O)N(alkyl)2. \"Cn-m alkylcarbamyl\" refers to a group of formula \u2014C(O)\u2014NH(alkyl). \"Cn-m alkylthio\" refers to a group of formula \u2014S-alkyl. \"Cn-m alkylsulfinyl\" refers to a group of formula \u2014S(O)-alkyl. \"Cn-m alkylsulfonyl\" refers to a group of formula \u2014S(O)2-alkyl. \"cyano-Cn-m alkyl\" refers to a group of formula \u2014(Cn-m alkylene)-CN. \"HO\u2014Cn-m alkyl\" refers to a group of formula \u2014(Cn-m alkylene)-OH. \"Cn-m alkoxy-Co-p alkyl\" refers to a group of formula \u2014(Cn-m alkylene)-O(Co-p alkyl). \"Carboxy\" refers to a group of formula \u2014C(O)OH. \"di(Cn-m-alkyl)amino\" refers to a group of formula \u2014N(alkyl)2. \"di(Cn-m-alkyl)carbamyl\" refers to a group of formula \u2014C(O)N(alkyl)2. \"Cn-m alkylcarbonyloxy\" is a group of formula \u2014OC(O)-alkyl. \"Aminocarbonyloxy\" is a group of formula \u2014OC(O)\u2014NH2. \"Cn-m alkylaminocarbonyloxy\" is a group of formula \u2014OC(O)\u2014NH\u2014alkyl. \"di(Cn-m alkyl)aminocarbonyloxy\" is a group of formula \u2014OC(O)\u2014N(alkyl)2. \"Carbamyl\" refers to a group of formula \u2014C(O)NH2. \"Carbonyl\", employed alone or in combination with other terms, refers to a \u2014C(O)\u2014 group.\n\n\"Cycloalkyl\" refers to non-aromatic cyclic hydrocarbons including cyclized alkyl and alkenyl groups. Cycloalkyl groups can include mono- or polycyclic (e.g., having 2, 3 or 4 fused rings) groups, spirocycles, and bridged rings (e.g., a bridged bicycloalkyl group). Ring-forming carbon atoms of a cycloalkyl group can be optionally substituted by oxo or sulfido (e.g., C(O) or C(S)). Also included in the definition of cycloalkyl are moieties that have one or more aromatic rings fused (i.e., having a bond in common with) to the cycloalkyl ring, for example, benzo or thienyl derivatives of cyclopentane, cyclohexane, and the like. A cycloalkyl group containing a fused aromatic ring can be attached through any ring-forming atom including a ring-forming atom of the fused aromatic ring. Cycloalkyl groups can have 3, 4, 5, 6, 7, 8, 9, or 10 ring-forming carbons (i.e., C3\u201310). In some embodiments, the cycloalkyl is a C3\u201310 monocyclic or bicyclic cycloalkyl; in some embodiments a C3\u20137 monocyclic cycloalkyl; in some embodiments a C4\u20137 monocyclic cycloalkyl; in some embodiments a C4\u201310 spirocycle or bridged cycloalkyl. Example cycloalkyl groups include cyclopropyl, cyclobutyl, cyclopentyl, cyclohexyl, cycloheptyl, cyclopentenyl, cyclohexenyl, cyclohexadienyl, cycloheptatrienyl, norbornyl, norpinyl, norcarnyl, cubane, adamantane, bicyclo[1.1.1]pentyl, bicyclo[2.1.1]hexyl, bicyclo[2.2.1]heptanyl, bicyclo[3.1.1]heptanyl, bicyclo[2.2.2]octanyl, spiro[3.3]heptanyl, and the like. In some embodiments, cycloalkyl is cyclopropyl, cyclobutyl, cyclopentyl, or cyclohexyl.\n\n\"Heteroaryl\" refers to a monocyclic or polycyclic (e.g., having 2, 3, or 4 fused rings) aromatic heterocycle having at least one heteroatom ring member selected from N, O, or S. In some embodiments, any ring-forming N in a heteroaryl moiety can be an N-oxide. In some embodiments, the heteroaryl is a 5\u201310 membered monocyclic or bicyclic heteroaryl having 1, 2, 3, or 4 heteroatom ring members independently selected from N, O, and S. In some embodiments, the heteroaryl is a 5\u20136 monocyclic heteroaryl having 1 or 2 heteroatom ring members independently selected from N, O, and S. Example heteroaryl groups include, but are not limited to, pyridyl, pyrimidinyl, pyrazinyl, pyridazinyl, pyrrolyl, pyrazolyl, azolyl, oxazolyl, isoxazolyl, thiazolyl, isothiazolyl, imidazolyl, furyl, thienyl, triazolyl (e.g., 1,2,3-triazolyl, 1,2,4-triazolyl, 1,3,4-triazolyl), tetrazolyl, thiadiazolyl (e.g., 1,2,3-thiadiazolyl, 1,2,4-thiadiazolyl, 1,3,4-thiadiazolyl), quinolinyl, isoquinolinyl, indolyl, benzothienyl, benzofuranyl, benzisoxazolyl, imidazo[1,2-b]thiazolyl, purinyl, triazinyl, thieno[3,2-b]pyridinyl, imidazo[1,2-a]pyridinyl, 1,5-naphthyridinyl, 1H-pyrazolo[4,3-b]pyridinyl, oxadiazolyl (e.g., 1,2,3-oxadiazolyl, 1,2,4-oxadiazolyl, 1,3,4-oxadiazolyl), 1,2-dihydro-1,2-azaborinyl, and the like.", "# BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nProtein: ADRB2\nSequence: MGQPGNGSAFLLAPNGSHAPDHDVTQERDEVWVVGMGIVMSLIVLAIVFGNVLVITAIAKFERLQTV...\nApo structure: ./refs/adrb2_apo.pdb\nFamily type: gpcr\n\n```\n\n5HT2 per-motif Ca RMSD\n\n5HT2 heatmap of motif x target RMSD\n\n###### **When the laptop says no, AI-assisted\u00a0Vibe Coding helps you fix it!**\n\nThe 5-HT2 panel\u2019s six Gq-heterotrimer targets (roughly 1,250 to 1,280 tokens each) would not run on Apple Silicon at all, and for two entirely separate reasons.\n\nThe first was a hard crash: Boltz\u2019s triangular attention builds the whole complex\u2019s row-wise attention-score matrix in a single unchunked matmul, which scales as tokens \u00d7 heads \u00d7 tokens \u00d7 tokens, and past about 1,250 residues, that one tensor exceeds the MPS backend\u2019s size ceiling and takes the process down with a SIGSEGV inside PyTorch\u2019s own tiled fallback. Because each row\u2019s attention is computed independently, chunking the matmul along the row axis is mathematically exact rather than an approximation, so BoltzMaker patches it straight into the installed `boltz` package at setup time, idempotently, and only after checking the function against the source it expects, so a future Boltz release cannot be silently mis-patched.\n\nThe second was a genuine out-of-memory failure: with the crash fixed, running six large targets back to back in one process hit real memory pressure partway through, and two targets fell over on Boltz\u2019s own clean-fail path, then dragged two already-successful targets down with them when the affinity phase went looking for an intermediate file that was never written. The fix is to `run` and `all` now retry any incomplete target automatically, isolating each one to its own single-target `boltz predict` invocation from the first retry onward, because a crashed subprocess only releases its MPS memory once it fully exits. Both are permanent, general behaviors rather than patches for this campaign: the chunking applies to any large complex on Apple Silicon, and the retry applies to any OOM, on any run.\n\n###### **What I learned\u2026**\n\n* **The dashboard was fine; the hosting wasn\u2019t.** It rendered perfectly locally and broke on the site. I spent far too long re-debugging the HTML before checking whether WordPress was rewriting it on the way out. It was. Check the delivery layer first, then the file.\n* **Vendored beats linked.** The dashboard originally pulled Plotly from a CDN and rendered blank the moment it was opened somewhere without network access. Inlining the library made the file bigger and the problem disappeared. Same for `compare-sse`\u2018s report.\n* **`pdb_tidy` lies quietly.** It renumbers TER records and leaves gaps in the CONECT serials, which OpenBabel then silently mis-reads as different bonds, giving PLIP wrong interactions with no error anywhere. The docs said nothing. Only diffing the actual output found it.\n* **Compute once, format twice.** The summary CSV and the dashboard originally each calculated the ranking themselves, and inevitably disagreed. One function now does the sums and the two outputs just render them.\n* **Grids will split a pair.** The ligand grid put structures and their labels in separate cells the moment the set size changed, because a grid sees siblings, not units. Group them explicitly or they drift apart.\n\n###### **What comes next\u2026**\n\n* **Benchmark it against reality.** Pull known actives and co-crystals for a target family from ChEMBL, BindingDB, and the PDB, then report predicted-versus-measured correlation and pose RMSD. A confidence score says the model is sure; this would say whether it\u2019s right.\n* **A triage view, not another table.** Plot confidence against predicted affinity and quadrant it, so choosing what to test next takes a glance rather than a spreadsheet.\n* **Flag the CPU fallbacks.** `torch.linalg.svd` drops from MPS to CPU with no warning and no error, just a much slower run. Preflight should catch it and say what it will cost.\n* **Checkpoint per sample.** Long runs fail near the end, and everything is lost. A crash should cost minutes, not hours.\n* **A `doctor` command.** Two environments and a dozen libraries, with failures that only surface three stages in. One command that imports everything and reports what\u2019s broken.", "# Outrunning Diffusion: The Fast Lane in CryoEM Grid Preparation\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nDespite these advances, there is no panacea; the protein, its buffer, and the workflow interact complexly, with cryoEM often relying on iterative optimization and trial-and-error. Hyperquenching, fast plunge-grid systems, and new support materials are blazing a trail, but the drag race against diffusion is still on. Time-resolved and precise grid behavior, supported by real-time analytics, represent the next steps towards reliably high-quality cryoEM grids\u2014shaping the future of structure-based drug design and fundamental biology.\n\n| Key Concept | Description | Key Reference |\n| --- | --- | --- |\n| Hyperquenching | Nanosecond vitrification forms protective glassy water barrier against AWI damage in cryoEM. | [Cherepanov, et al., Chemistry Europe, 2023](https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cmtd.202200050) |\n| Automated grid preparation | Blotless, high-speed systems minimize exposure and error during sample prep, boosting resolution. | [NanoImaging Services Blog, 2024](https://www.nanoimagingservices.com/about/blog/a-guide-to-cryo-grid-production-materials-interfaces-and-their-impact-on-sample-prep) |\n| Materials engineering | Hydrophilized graphene grids provide gentle support, reducing protein denaturation and preferred orientation. | [D\u2019Imprima, et al., eLife, 2019](https://elifesciences.org/articles/42747) |\n| Support film crystal coating | Engineered protein coatings sequester macromolecules, removing them from direct AWI contact. | [Xu, et al., Frontiers in Molecular Biosciences, 2022](https://www.frontiersin.org/journals/molecular-biosciences/articles/10.3389/fmolb.2022.892459/full) |\n| Time-resolved grid analytics | Real-time control and monitoring improve reproducibility and optimize sample conditions for high-resolution imaging. | [Rubino, et al., PNAS, 2022](https://www.pnas.org/doi/10.1073/pnas.2516900122) |", "# HGMP01 and FAST-PETase: Decoding Two Paths to Plastic Degradation \u2013 Engineered versus Natural PETase Enzymes\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nTemperature preferences distinguish these enzymes significantly. FAST-PETase functions optimally at [50\u00b0C](https://pmc.ncbi.nlm.nih.gov/articles/PMC10851425/), approaching the glass transition temperature of PET where polymer chains become more accessible to enzymatic attack. The [five mutations](https://pubmed.ncbi.nlm.nih.gov/35478237/) work synergistically to elevate the melting temperature to approximately 63\u00b0C, a [substantial improvement](https://pmc.ncbi.nlm.nih.gov/articles/PMC10594578/) over the wild-type enzyme. HGMP01 exhibits maximal activity at [40\u00b0C](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf), consistent with human body temperature and the conditions within the gastrointestinal tract where this enzyme naturally functions.\u200b\n\npH preferences reveal adaptation to distinct environments. FAST-PETase performs best at [pH 9.0](https://www.nature.com/articles/s42003-023-04413-0), maintaining activity across the moderate range of 6.0 to 9.0. HGMP01 demonstrates remarkable alkaline tolerance, achieving peak activity at [pH 10.0 to 11.0](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf) while retaining function down to neutral pH. This alkaline preference correlates with HGMP01\u2019s [predicted isoelectric point of 9.12](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf)and overall positive surface charge, contrasting with the [dipole distribution](https://pmc.ncbi.nlm.nih.gov/articles/PMC8540959/) observed in the parent enzyme of FAST-PETase.\u200b\n\nThe catalytic machinery in both enzymes utilizes the classic serine hydrolase mechanism, proceeding through [acylation and deacylation stages](https://www.nature.com/articles/s42004-024-01154-x). FAST-PETase benefits from the [N233K mutation](https://pubs.acs.org/doi/10.1021/jacs.3c04427), which paradoxically enhances catalysis not by forming a salt bridge as initially proposed, but by altering the chain folding around Glu204. This conformational change increases the basicity of Asp206 in the catalytic triad, strengthening its interaction with protonated His237 and [lowering the activation barrier](https://pmc.ncbi.nlm.nih.gov/articles/PMC10851425/) by approximately 4.4 kcal/mol. The specific catalytic residues in HGMP01 remain less characterized, though the enzyme clearly possesses the [conserved catalytic triad](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf) essential for ester bond hydrolysis.\u200b\n\nBinding site organization reflects each enzyme\u2019s substrate preferences. FAST-PETase maintains the [characteristic shallow L-shaped cleft](https://pmc.ncbi.nlm.nih.gov/articles/PMC8540959/) spanning approximately 40 \u00c5, divided into subsites that accommodate multiple PET moieties simultaneously. The [wobbly Trp185 residue](https://www.nature.com/articles/s41467-017-02255-z) adopts multiple conformations, facilitating substrate entry and product release through \u03c0-\u03c0 interactions with aromatic groups in PET. HGMP01 shows enhanced activity toward [PET nanoparticles](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf), suggesting its binding architecture may favor smaller substrate fragments consistent with the fragmented plastic particles encountered in the digestive environment.\u200b\n\nSequence relationships provide insight into evolutionary origins. FAST-PETase shares [100% sequence identity](https://pmc.ncbi.nlm.nih.gov/articles/PMC8540959/) with its parent IsPETase aside from the five engineered mutations, maintaining the [two disulfide bonds](https://pmc.ncbi.nlm.nih.gov/articles/PMC8540959/) characteristic of this enzyme family. HGMP01 exhibits only [20 to 41% sequence similarity](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf) to known PETases, indicating independent evolution. Both enzymes contain the [DLH domain](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf), a structural signature common to many esterases and lipases.\u200b\n\nThe convergent evolution toward plastic degradation activity through divergent pathways highlights fundamental principles of enzyme adaptation. FAST-PETase achieves superior performance through precise mutation of surface residues that stabilize the protein fold and optimize active site geometry for industrial applications. HGMP01 represents nature\u2019s independent solution, potentially shaped by millennia of exposure to polyester-like compounds in the gut environment. [Recent metatranscriptome analysis](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf) confirms that HGMP01-like genes are actively expressed in over 20% of surveyed human gut samples, suggesting an ongoing role in managing ingested microplastics.\u200b\n\nThese parallel developments underscore both the promise and complexity of biocatalytic plastic recycling. While FAST-PETase demonstrates the power of rational design to create robust industrial biocatalysts, HGMP01 reveals that functional plastic-degrading enzymes have emerged spontaneously in unexpected ecological niches, offering new templates for protein engineering and raising important questions about the biological fate of ingested plastics in human populations.", "# Unlocking Molecular Potential: The Evolution of Topliss Tree and Bioisosteric Strategies in Modern Drug Discovery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-09-15 on marcdeller.com (categories: blog, sb).*\n\n* **Systematic Optimization Foundation**: The Topliss tree methodology established fundamental principles for systematic molecular optimization that remain relevant in modern computational drug design approaches\n* **Bioisosteric Strategy Evolution**: Classical bioisosterism has expanded to encompass sophisticated replacements addressing multiple optimization parameters simultaneously, from metabolic stability to selectivity enhancement\n* **Structure-Based Design Revolution**: High-resolution protein structures and AI predictions have transformed rational drug design, enabling precise prediction of how modifications affect binding and selectivity\n* **Multi-Parameter Optimization**: Contemporary approaches successfully balance potency, selectivity, ADMET properties, and safety through integrated optimization strategies combining multiple methodologies\n\n1. **Utilization of operational schemes for analog synthesis in drug design:** Topliss, J.G., et al.,\u00a0[J. Med. Chem., 1972](https://pubs.acs.org/doi/abs/10.1021/jm00280a002)\n2. **The Influence of Bioisosteres in Drug Design: Tactical Applications:** Meanwell, N.A., et al.,\u00a0[PMC, 2014](https://pmc.ncbi.nlm.nih.gov/articles/PMC7416817/)\n3. **Using Matched Molecular Series as a Predictive Tool To Optimize Biological Activity:** O\u2019Boyle, N.M., et al.,\u00a0[J. Med. Chem., 2014](https://pubs.acs.org/doi/10.1021/jm500022q)\n4. **Bioisosteres in Medicinal Chemistry:** Brown, N., et al.,\u00a0[Wiley, 2012](https://onlinelibrary.wiley.com/doi/book/10.1002/9783527654307)\n5. **A Data-Driven Perspective on Bioisostere Evaluation:** Sauer, S., et al.,\u00a0[PMC, 2025](https://pmc.ncbi.nlm.nih.gov/articles/PMC12406194/)\n6. **Structure-Based Drug Design with a Deep Hierarchical Generative Model:** Guan, J., et al.,\u00a0[J. Chem. Inf. Model., 2024](https://pubs.acs.org/doi/10.1021/acs.jcim.4c01193)\n\n**Unlocking Molecular Potential: The Evolution of Topliss Tree and Bioisosteric Strategies in Modern Drug Discovery**\n\nThe landscape of drug discovery has undergone a remarkable transformation since John Topliss introduced his groundbreaking decision tree methodology in 1972. What began as a simple, non-mathematical approach to guide medicinal chemists through systematic structural modifications has evolved into a sophisticated toolkit that now integrates artificial intelligence, computational modeling, and data-driven insights. Today\u2019s structure-based drug design represents a confluence of traditional rational approaches and cutting-edge technologies, where the principles established by Topliss continue to inform modern optimization strategies while bioisosteric replacements have become the cornerstone of contemporary medicinal chemistry.\n\nThe Topliss tree methodology emerged from a fundamental challenge that plagued early drug discovery: how to systematically explore aromatic substitution patterns without resorting to extensive mathematical calculations or statistical procedures that many chemists found intimidating. Topliss recognized that while Hansch\u2019s quantitative structure-activity relationship (QSAR) approach provided valuable insights into how lipophilic, electronic, and steric properties affected biological activity, it required computational expertise that was not readily available to all medicinal chemists in the early 1970s. His elegant solution was to create a branching decision tree that would guide researchers through sequential substituent modifications based on observed activity changes, effectively democratizing the application of structure-activity principles.\n\nThe original Topliss tree operates on a straightforward premise: starting with an unsubstituted benzene ring, chemists would systematically introduce substituents based on their physicochemical properties, with each branch point determined by whether the previous modification increased, decreased, or maintained biological activity. This stepwise approach typically began with a 4-chloro substitution, followed by branches leading to 3,4-dichloro, 4-methoxy, or 3-trifluoromethyl modifications depending on the observed activity trends. The genius of this system lay not in its mathematical sophistication, but in its practical applicability and intuitive nature.\n\nRecent computational analyses have provided new perspectives on the Topliss tree\u2019s underlying principles. Modern visualization tools using extended electron distribution force fields have revealed the profound electrostatic and hydrophobic changes introduced by the recommended substitution patterns. These insights validate the original physicochemical rationale while providing deeper understanding of how electronic effects propagate through molecular frameworks. Furthermore, contemporary data-driven approaches like the Matsy algorithm have been compared directly with Topliss recommendations, generally showing remarkable concordance while occasionally suggesting alternative pathways based on statistical analysis of large chemical databases.\n\nParallel to the evolution of systematic optimization strategies, the concept of bioisosterism has matured from Langmuir\u2019s early observations about molecules with similar electron distributions to a nuanced understanding of how structural modifications can fine-tune biological, pharmacological, and physicochemical properties. Classical bioisosterism, rooted in the principle that atoms or groups with similar valence electron structures exhibit comparable biological activities, has expanded to encompass non-classical replacements that maintain function while potentially differing significantly in electronic or steric characteristics.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics; supervisor Prof. E. Yvonne Jones), 1999\n\nIn the I222 structure residues 6\u201394, 104\u2013132 and 158\u2013187 are well ordered. Residues 1\u20135 at the N-terminus, residues 95\u2013103 in the B-C loop and residues 133\u2013157 of the C-D loop are disordered and display no interpretable electron density. Electron density is good throughout the course of the main chain and a comparison of the electron density for the two structures is shown below in Figure 5-8.\n\nReasons for the poor electron density at the N-terminus and inter helical loop regions are readily apparent on inspection of the B-factors for the P212121 structure as shown in Figure 5-9 below. In both the P212121 and I222 structures, the N-terminal loop region preceding Leu 10 shows a sharp increase in main chain B-factor values, with both rising above the mean values of 27 \u00c52 and 39 \u00c52, respectively. Similarly, the C-terminal region of helix C from Glu 130 and the N-terminal region of helix D before Phe 160 show a similar increase. Also, higher than average main chain B-factors are observed for both structures between residues Glu 90 and Lys 110. The level of thermal motion in these regions suggests a high degree of flexibility within the crystal, which therefore gives rise to multiple conformations and ambiguous electron density in these regions.\n\n*Figure 5-8.* Figure showing the refined co-ordinates and 2|Fo|\u2212|Fc|\u03b1calc electron density maps of \u0394188_OSM contoured at 1 \u03c3. Main panel shows region of P212121 map corresponding to helix C running from residues 106 to 126. Boxed section shows corresponding region of map for each of the two spacegroups. The I222 map is to a resolution of 2.5 \u00c5 and the P212121 map to a resolution of 2.2 \u00c5. Maps were produced from MAD data using SHARP (de La Fortelle and Bricogne, 1997) followed by solvent flattening with SOLOMON (Abrahams and Leslie, 1996) and refinement with CNS-SOLVE (Brunger et al., 1998). Figure was produced using BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994).\n\n*Figure 5-9.* Figure showing a) average main chain B-factor, b) average side chain B-factor, c) solvent accessibility and d) secondary structure assignment by residue for the entire \u0394188_OSM sequence. Data shown for P212121 structure. Figure produced using PROCHECK (Laskowski et al., 1993).\n\nAnalysis of the Ramachandran plot shown in Figure 5-10 below shows that the stereochemistry of the molecule is good with no residues lying in energetically disallowed regions, and only Arg 67 of the I222 structure falling in a generously allowed region. No clear density was observed for Arg 67, so suggesting high levels of flexibility for this side chain.\n\nFor the P212121 structure 94.3 % of residues lie in the most favoured regions (A, B, L) and 5.7 % lie in additionally allowed regions (a, b, l, p). No residues lie in generously allowed (~a, ~b, ~l, ~p) or disallowed regions.\n\nFor the I222 structure 88.1 % of residues lie in the most favoured regions and 11.1 % lie in additionally allowed regions. No residues lie in disallowed regions and only 0.8 % of residues fall in generously allowed regions.\n\nA number of residues including Gln 20, Asp 22, Asp 26, Glu 51, His 83, Asp 87, Asn 123, Lys 163, Arg 162, and Arg 84 appeared to have alternate side-chain conformations. Refinement of conformers was not successful, presumably due to the low occupancies (typically <20 %) of the minor conformations and the relatively low resolution of data.\n\n*Figure 5-10.* Figure showing Ramachandran plot of the stereochemistry observed in a) P212121 model and b) I222 model. For the P212121 model 94.3 % of residues lie in the most favoured regions (A, B, L) and 5.7 % lie in additionally allowed regions (a, b, l, p). No residues lie in generously allowed (~a, ~b, ~l, ~p) or disallowed regions. For the I222 structure 88.1 % of residues lie in the most favoured regions and 11.1 % lie in additionally allowed regions. No residues lie in disallowed regions and only 0.8 % of residues fall in generously allowed regions. Figure was produced using the program PROCHECK (Laskowski et al., 1993).\n\n### 5.4 Discussion of Structure Determination\n\n\u0394188_OSM was the only sample that produced crystals. Both OSM-196 and Se-Met-OSM variants showed exceptionally poor solubility and aggregation. Crystals used for structure determination were grown from PEG 35000 and this is the first reported use of this reagent for the growth of protein crystals.\n\nSe-Met derivatised OSM displayed exceptionally poor solubility presumably due to the reduced solubility of Se-Met compared to Met. Indeed, the C-Se bond is 0.14 \u00c5 longer than the C-S bond and this leads to greater polarity, weaker bonds and higher reactivity. Selenides are more susceptible to oxidation than the corresponding sulphide and the resultant selenoxides themselves act as mild oxidising agents. Inclusion of reducing agents to the preparation was not sufficient, so suggesting further problems with protein folding. This is confirmed on inspection of the final refined model, which shows that all four methionine residues in the sequence are buried in the core of the protein and all display solvent accessibilities of less than 50 % (see Figure 5-5 (b)).", "# GitHub README: GOBSMACKED\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/GOBSMACKED, 2026\n\n### \ud83e\uddf2 Affinity, before and after MD\n\nBoltz-2's affinity head scores the docked pose and the relaxed complex, so the panel answers a\nquestion the rest of the scorecard cannot: did relaxing the pose change what the model thinks of\nit. Each is reported as pIC50, the raw `affinity_pred_value` (log10 of IC50 in micromolar, lower\nis stronger) and the binder probability, with the change between them. The post-MD figure is a\nmean over several frames from the last fifth of the trajectory, with its standard deviation, so a\nprediction that swings by a log unit across the sampled window says so rather than presenting one\nframe as the answer.\n\n**It is deliberately outside the composite score.** Every graded metric on the scorecard has a\ncrystal structure to be right or wrong about. A predicted affinity has none, and folding an\nunverifiable number into a score whose whole argument is verification would break that argument.\nIt sits beside the grade, not inside it.\n\n**A missing affinity is a missing panel, not a failed run.** The stage declines rather than\nraising: no ligand it can parameterise, no MSA, the affinity head writing nothing for a pose. The\nreason is carried into the archive and shown on the card, and the other five stages are untouched.\n\n### \ud83c\udfb2 Every pose, not just the top one\n\nDocking returns ten poses and the scorecard grades one of them, which hides the case that matters\nmost: the run that found the right answer and ranked it fourth. The overlay panel draws all ten at\nonce and tabulates, per pose, the docking score, the in-place RMSD to the top pose, the centroid\nseparation, a best-fit shape RMSD, and the closest heavy-atom approach to the receptor.\n\nThe distinction between the two RMSDs is the useful part. In-place RMSD asks how far this pose\nsits from that one; best-fit shape RMSD asks whether it is the same conformer put somewhere else.\nA pair that is 6 \u00c5 apart in place and 0.4 \u00c5 after superposition is one pose docked into two sites,\nwhich is a search problem. A pair that is 6 \u00c5 apart both ways is two genuinely different bound\nconformations, which is not.\n\n### \ud83d\udd11 The binding mode\n\nTwo families get a real answer and everything else gets an honest one.\n\n**Kinases** are labelled from the KLIFS 85-residue pocket, mapped onto the target sequence region by region. DFG-in / out / inter follows the Modi and Dunbrack distance criteria; \u03b1C-in / out follows the \u03b23-Lys to \u03b1C-Glu salt bridge; the Type I / I\u00bd / II / allosteric label follows which subpockets the ligand occupies.\n\n**GPCRs** are labelled from GPCRdb generic numbering: orthosteric, vestibule, intracellular or lipid-facing from the contacts, and active-like or inactive-like from the TM3-TM6 distance, with the NPxxY RMSD, the toggle switch \u03c71, the PIF motif and the sodium site reported alongside.\n\nBoth classifiers run on the prediction and on the crystal with the same code, so a difference in the label is a difference in the structure and not a difference in method.\n\n### \ud83e\uddea Thresholds that were measured, not quoted\n\nTwo numbers in this repository were placed by measuring structures with this code rather than by taking a value from a paper that used a different atom pair:\n\n- **The TM3-TM6 activation cut is 13 \u00c5**, from six structures: inactive rhodopsin 1GZM 8.7, A2A 3EML 9.7, \u03b22AR 2RH1 11.2; active metarhodopsin 3PQR 14.7, A2A 5G53 18.5, \u03b22AR 3SN6 19.0. (Quoted \"ionic lock\" distances of 3 to 4 \u00c5 are guanidinium-to-carboxylate, not C\u03b1-to-C\u03b1, and are not comparable.)\n- **A subpocket counts as occupied at two contacts, not one.** Erlotinib in 1M17 grazes exactly one \u03b1C residue at 4 \u00c5, and a one-contact rule labels a textbook Type I inhibitor as Type I\u00bd.\n\n## \ud83e\uddf1 Stack\n\n| Component | Role | Licence | Where it runs |\n|---|---|---|---|\n| Flask, gunicorn, nginx, SQLite | Web application and store | BSD / MIT / public domain | droplet |\n| biotite, tmtools, gemmi, NumPy, SciPy | Alignment, superposition, TM-score | BSD / MIT / MPL | droplet |\n| RDKit | SMILES, depiction, fingerprints, symmetry-aware RMSD | BSD-3 | droplet and bundle |\n| PLIP | Interaction fingerprints | GPL-2.0 | droplet only, as a subprocess |\n| PandaMap | 2D interaction maps, empirical \u0394G | MIT | droplet |\n| PandaDock | Docking (hybrid search plus SE(3) GNN rescoring) | MIT | bundle |\n| Boltz-2 | Affinity head, on the docked pose and on MD frames | MIT | bundle |\n| ESMFold, OpenFold | Folding when no model exists | MIT / Apache-2.0 | bundle |\n| OpenMM, PDBFixer, openmmforcefields | Preparation, minimisation, MD | MIT / LGPL | bundle |\n| MDTraj | Trajectory analysis | LGPL-2.1 | bundle and droplet |\n| Mol*, Plotly.js | 3D views and traces | MIT | browser |\n\n**PLIP is GPL-2.0 and is run as a subprocess, never imported.** What crosses the boundary is an XML file, so no GPL code is linked into this MIT-licensed application, and PLIP is deliberately absent from the run bundle. Full attribution, with every DOI checked against Crossref, is in [THIRD_PARTY.md](THIRD_PARTY.md) and on the app's About page: both are generated from `software.yaml`, so they cannot drift.\n\n## \ud83d\udcc1 Repository layout", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\nThe crystal structure of a putative NTPase, YP_001813558.1 from Exiguobacterium sibiricum 255-15 (PF09934, DUF2166) was determined to 1.78\u2005\u00c5 resolution. YP_001813558.1 and its homologs (dimeric dUTPases, MazG proteins and HisE-encoded phosphoribosyl ATP pyrophosphohydrolases) form a superfamily of all-\u03b1-helical NTP pyrophosphatases. In dimeric dUTPase-like proteins, a central four-helix bundle forms the active site. However, in YP_001813558.1, an unexpected intertwined swapping of two of the helices that compose the conserved helix bundle results in a `linked dimer' that has not previously been observed for this family. Interestingly, despite this novel mode of dimerization, the metal-binding site for divalent cations, such as magnesium, that are essential for NTPase activity is still conserved. Furthermore, the active-site residues that are involved in sugar binding of the NTPs are also conserved when compared with other \u03b1-helical NTPases, but those that recognize the nucleotide bases are not conserved, suggesting a different substrate specificity.\n\n## Structure of the first representative of Pfam family PF04016 (DUF364) reveals enolase and Rossmann-like folds that combine to form a unique active site with a possible role in heavy-metal chelation (2010)\n\n**Authors:** Miller MD, Aravind L, Bakolitsa C, Rife CL, Carlton D, Abdubek P, Astakhova T, Axelrod HL, Chiu HJ, Clayton T, Deller MC, Duan L, Feuerhelm J, Grant JC, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Krishna SS, Kumar A, Marciano D, McMullan D, Morse AT, Nigoghossian E, Okach L, Reyes R, van den Bedem H, Weekes D, Xu Q, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309110007517. PMID: 20944207.\n\nThe crystal structure of Dhaf4260 from Desulfitobacterium hafniense DCB-2 was determined by single-wavelength anomalous diffraction (SAD) to a resolution of 2.01\u2005\u00c5 using the semi-automated high-throughput pipeline of the Joint Center for Structural Genomics (JCSG) as part of the NIGMS Protein Structure Initiative (PSI). This protein structure is the first representative of the PF04016 (DUF364) Pfam family and reveals a novel combination of two well known domains (an enolase N-terminal-like fold followed by a Rossmann-like domain). Structural and bioinformatic analyses reveal partial similarities to Rossmann-like methyltransferases, with residues from the enolase-like fold combining to form a unique active site that is likely to be involved in the condensation or hydrolysis of molecules implicated in the synthesis of flavins, pterins or other siderophores. The genome context of Dhaf4260 and homologs additionally supports a role in heavy-metal chelation.\n\n## Structure of the first representative of Pfam family PF09410 (DUF2006) reveals a structural signature of the calycin superfamily that suggests a role in lipid metabolism (2010)\n\n**Authors:** Chiu HJ, Bakolitsa C, Skerra A, Lomize A, Carlton D, Miller MD, Krishna SS, Abdubek P, Astakhova T, Axelrod HL, Clayton T, Deller MC, Duan L, Feuerhelm J, Grant JC, Grzechnik SK, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Kumar A, Marciano D, McMullan D, Morse AT, Nigoghossian E, Okach L, Paulsen J, Reyes R, Rife CL, van den Bedem H, Weekes D, Xu Q, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309109037749. PMID: 20944205.\n\nThe first structural representative of the domain of unknown function DUF2006 family, also known as Pfam family PF09410, comprises a lipocalin-like fold with domain duplication. The finding of the calycin signature in the N-terminal domain, combined with remote sequence similarity to two other protein families (PF07143 and PF08622) implicated in isoprenoid metabolism and the oxidative stress response, support an involvement in lipid metabolism. Clusters of conserved residues that interact with ligand mimetics suggest that the binding and regulation sites map to the N-terminal domain and to the interdomain interface, respectively.\n\n## Structure of the \u03b3-D-glutamyl-L-diamino acid endopeptidase YkfC from Bacillus cereus in complex with L-Ala-\u03b3-D-Glu: insights into substrate recognition by NlpC/P60 cysteine peptidases (2010)\n\n**Authors:** Xu Q, Abdubek P, Astakhova T, Axelrod HL, Bakolitsa C, Cai X, Carlton D, Chen C, Chiu HJ, Chiu M, Clayton T, Das D, Deller MC, Duan L, Ellrott K, Farr CL, Feuerhelm J, Grant JC, Grzechnik A, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Krishna SS, Kumar A, Lam WW, Marciano D, Miller MD, Morse AT, Nigoghossian E, Nopakun A, Okach L, Puckett C, Reyes R, Tien HJ, Trame CB, van den Bedem H, Weekes D, Wooten T, Yeh A, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309110021214. PMID: 20944232.", "# 2016 \u2013 Key gp120 glycans pose roadblocks to the rapid development of VRC01-class antibodies in an HIV-1- infected Chinese donor\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe germline-targeting vaccination strategy has emerged as a promising approach to overcome these glycan-imposed barriers, with recent clinical trials demonstrating proof-of-concept success in humans. The eOD-GT8 60mer immunogen successfully activated VRC01-class B cell precursors in 97% of vaccine recipients in the IAVI G001 Phase 1 trial, inducing somatic hypermutation characteristic of mature broadly neutralizing antibodies. Remarkably, these vaccine-elicited precursors retained the capacity to accommodate the N276 glycan despite its absence from the priming immunogen, suggesting that intrinsic structural plasticity enables glycan adaptation during subsequent affinity maturation.\n\nPopulation-specific genetic factors further complicate VRC01-class antibody development, with Chinese populations demonstrating distinct immunological profiles that influence antibody evolution trajectories. The predominance of specific HIV-1 subtypes, particularly CRF07_BC and CRF01_AE variants in Chinese populations, presents unique glycosylation patterns that may differentially impact antibody maturation pathways. These viral variants exhibit distinct envelope conformations and glycan arrangements that could selectively favor or hinder particular antibody lineages, contributing to population-specific differences in broadly neutralizing antibody development rates.\n\nRecent advances in structural immunology have revealed the molecular mechanisms underlying glycan accommodation, demonstrating that successful VRC01-class antibodies develop sophisticated paratope architectures capable of navigating the complex glycan landscape. High-resolution cryo-electron microscopy structures show that mature antibodies employ complementarity-determining region rearrangements in both heavy and light chains to achieve glycan tolerance while maintaining CD4-binding site recognition. These structural adaptations occur through iterative rounds of somatic hypermutation and selection, requiring precisely orchestrated molecular evolution that balances glycan accommodation with neutralization potency.\n\n| Key Concept | Description | Key References |\n| --- | --- | --- |\n| VRC01-class Broadly Neutralizing Antibodies | Potent HIV-neutralizing antibodies derived from IGHV1-2 gene segments that target the CD4-binding site with broad cross-clade neutralization capability | [PMC12377726](https://pmc.ncbi.nlm.nih.gov/articles/PMC12377726/) |\n| N276 Glycan Shield | A conserved carbohydrate modification at asparagine 276 that creates steric barriers preventing antibody access to vulnerable epitopes on gp120 | [Cell Reports](https://www.sciencedirect.com/science/article/pii/S2211124721013954) |\n| Somatic Hypermutation (SHM) | The accumulation of random mutations in antibody variable regions during immune responses, requiring extensive changes for broad neutralization | [Medical Xpress](https://medicalxpress.com/news/2024-08-mechanisms-antibody-gene-diversification-somatic.html) |\n| Germline-targeting Vaccination | Vaccine strategies designed to prime rare B cell precursors capable of developing into broadly neutralizing antibodies through sequential immunization | [PMC12313413](https://pmc.ncbi.nlm.nih.gov/articles/PMC12313413/) |\n| CD4-binding Site (CD4bs) | The conserved binding region on HIV envelope protein gp120 where the virus attaches to CD4 receptors on host T cells | [PMC5103708](https://pmc.ncbi.nlm.nih.gov/articles/PMC5103708/) |\n| Glycan Accommodation Mechanisms | Structural adaptations allowing mature antibodies to navigate around glycan obstacles while maintaining neutralization potency | [eLife Sciences](https://elifesciences.org/articles/68110) |", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices\nAuthors: Deller, M.C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nOther important residues residing on helix C of OSM are Asn 124 and Asn 123, with mutation of Asn 124 showing the greatest effect on receptor binding and biological activity. Analysis of the proposed model suggests that Asn 124 is directed towards Phe 169 and the B-C loop of the receptor, whereas Asn 123 has a higher solvent accessibility being placed on the outer face of helix A, so showing less interaction towards this region of the receptor. The only residues implicated in receptor engagement on helix A are Gln 16 and Gln 20. It is interesting to note that in the long stretch of residues mutated on this helix (i.e. residues 7\u201326) only this pair showed activity. Their spacing of one helical turn suggests that both residues face in a similar fashion towards the major epitope on the receptor.\n\nClearly this model fails to account for all data, including site 1 and site 3 interactions, and further crystallographic studies involving LIFR and complexes thereof will be required.\n\n*Figure 6-14.* Figure showing (a) proposed OSM-gp130 site 2 binding model and (b) proposed OSM-gp130 site 2 interface, based on Epo-EpoR structure (Syed et al., 1998). Panel (a) shows gp130 with the N-terminal domain coloured dark blue and the C-terminal domain coloured light blue. The WSXWS motif is shown along with the interdomain linker region in orange and the F-G loop in purple. Residues implicated in site 1 and site 3 are shown as CPK objects in yellow and red, respectively. Panel (b) shows a closeup of the interface surface with residues shown to be important through mutagenesis (Deller et al., 2000; Horsten et al., 1997) shown as ball and stick objects. Helices A (blue), B (green), C (yellow) and D (red) are shown along with connecting loops in grey. Superposition was carried out using the program SHP (Stuart et al., 1979). Figure produced using BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994).\n\n### 6.6 Summary\n\nIn conclusion it can be seen that the present structure of OSM highlights several novel features unique to the gp130 family of cytokines, whilst at the same time supporting common themes observed in other 4 helix bundle cytokines. Although exhibiting a high degree of structural homology to other helical cytokines, key features such as the helical content of both the B-C and A-B loops and the severity of the kink in helix A set OSM apart. These subtle differences are likely to be key features responsible for modulation of receptor-ligand engagement in the gp130 family which utilise otherwise homologous systems.\n\nStructural comparison reveals degrees of similarity to the classic two site binding system utilised by GH and Epo in which hydrophobic burial of key aromatic residues of the receptor drives the interaction (de Vos et al., 1992; Syed et al., 1998). However, models based on these interactions suggest the use of an interface that is less hydrophobic in nature in both LIF and OSM. Indeed, mutagenesis has revealed that site 2 interactions with gp130 and site 3 interactions with LIFR dominate the binding energy and as such the classic site 1 is of less importance (Deller et al., 2000; Hudson et al., 1996; Layton et al., 1994b). Residues comprising site 2 and site 3 of OSM are both hydrophobic and polar in nature in a similar fashion to the mosaic binding sites observed in both the GCSF (Aritomi et al., 1999) and IL4 (Hage et al., 1998) receptor-ligand systems. This move away from the use of hydrophobic surfaces may allow tighter modulation of fidelity, albeit at the possible expense of affinity, a situation that is ideal for the 'pick and mix' approach to receptor engagement observed in the gp130 family of cytokines (Staunton et al., 1998).\n\nThe level of promiscuity exhibited by LIFR and gp130 receptors suggests that novel motifs on the receptors are required. Models of the OSM-gp130 interaction, based on the Epo system, suggest that only one hydrophobic residue of the receptor dominates the binding energy; a claim which is so far further supported by mutagenesis data (Horsten et al., 1997). It is possible that the gp130 interaction has moved away from the classic B-C and E-F loop interactions which would lead to invariant and 'rigid' orientation of the ligand between the two receptor domains to a more 'flexible' interaction involving components of the B-C loop and the F-G loop, as observed in the GCSF complex (Aritomi et al., 1999; Layton et al., 1999). It has also been suggested that such receptor-ligand complexes are likely to be dependent on relative domain conformation and orientation, as observed in the Epo system (Syed et al., 1998) and as such this model would allow for such optimisation of orientation. Structural features such as the WSXWS motif, which effectively forms a supporting 'back bone' of gp130 may be important in defining the allowed range of orientations.\n\nOther evidence suggests that the gp130 family of cytokines may use complex modes of binding including electrostatic (McDonald et al., 1995) and entropic components (Purvis and Mabbutt, 1997). Further receptor-ligand interactions, as observed in the minor site of the GCSFR-GCSF complex (Aritomi et al., 1999) and the proposed site 3 interaction between LIF and the Ig domain of LIFR (Chobotova and Heath, 2000), adds a further level of complexity, which the classic two site model of GH and Epo fails to explain. This again highlights the complexities involved in these systems and traditional mutagenesis may prove of little merit for analysis of the finer details of the active signalling complexes.", "# Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33\nAuthors: Kong, L.; Giang, E.; Nieusma, T.; Robbins, J.B.; Deller, M.C.; Stanfield, R.L.; Wilson, I.A.; Law, M.\nSource: Journal of Virology, 2012\n\n**Authors:** Leopold Kong, Erick Giang, Travis Nieusma, Justin B. Robbins, Marc C. Deller, Robyn L. Stanfield, Ian A. Wilson (corresponding author) and Mansun Law (corresponding author). **Affiliations (condensed):** Departments of Molecular Biology and of Immunology and Microbial Science, and The Skaggs Institute for Chemical Biology, The Scripps Research Institute, La Jolla, California, USA; and Joint Center for Structural Genomics, La Jolla, California, USA. **Published:** Journal of Virology, volume 86, number 23, pages 13085\u201313088, December 2012. Received 27 July 2012; accepted 7 September 2012; published ahead of print 12 September 2012.\n\n## Abstract\n\nWe have determined the crystal structure of the broadly neutralizing antibody (bnAb) AP33, bound to a peptide corresponding to hepatitis C virus (HCV) E2 envelope glycoprotein antigenic site 412 to 423. Comparison with bnAb HCV1 bound to the same epitope reveals a different angle of approach to the antigen by bnAb AP33 and slight variation in its \u03b2-hairpin conformation of the epitope. These structures establish two different modes of binding to E2 that antibodies adopt to neutralize diverse HCV.\n\n## Main text\n\nStructural characterization of conserved neutralizing epitopes provides critical information for the design of vaccines to counteract genetic diversity of pathogens. The E2 antigenic site 412 to 423 is a highly conserved neutralizing determinant of HCV and is a prime target for vaccine design. We recently determined the crystal structure of this conserved site in complex with a human broadly neutralizing antibody (bnAb), HCV1. The antibody-bound epitope forms a \u03b2-hairpin displaying a hydrophilic face and a hydrophobic face on opposing sides of the hairpin. The antibody predominantly interacts with the E2 residues Leu413 and Trp420 on the hydrophobic face of the epitope that are nearly 100% conserved. Nevertheless, HCV can escape this antibody through mutations at other positions on the binding face, e.g., N415K (in 1% of circulating HCV).\n\nTo further characterize this important neutralizing determinant, we report a second structure of this antigenic site in complex with the bnAb AP33. The murine monoclonal antibody (MAb) AP33 was discovered by Patel and coworkers, and the antibody was found to have broad neutralizing activity to diverse HCV isolates. In this study, the antibody was expressed as a chimeric mouse-human antibody to facilitate expression and purification (see Fig. S1 in the supplemental material). The antibody epitope has been mapped and extensively studied by overlapping peptide scanning, phage-display mimotope panning, selection of in vitro escape mutants, and site-directed mutagenesis. The E2 mutations N415Y, N415D, N417S, and G418D enable viral escape from neutralization by the MAb AP33.\n\nThe crystal structure reveals that, similar to the binding site for the bnAb HCV1, the AP33 epitope also forms a \u03b2-hairpin sandwiched between the heavy chain (HC) and light chain (LC) of the antibody (Fig. 1A) (detailed methods are provided in the supplemental material). Most of the binding is mediated by hydrophobic interactions along the hydrophobic face of the epitope (Fig. 1B; see also Table S2 in the supplemental material). A number of hydrogen bonds also stabilize the interaction, mostly between side chains on the Fab and main chain of the peptide (Fig. 1C; see also Table S4 in the supplemental material). Overall, there are many similarities between the AP33 and HCV1 epitopes. The same type of \u03b2-turn (type I\u2032) is found in both structures, and both antibodies bind the hydrophobic face of the \u03b2-hairpin (Fig. 1B; see also Table S2 in the supplemental material). However, the anti-parallel \u03b2-sheet in the \u03b2-hairpin in the AP33 epitope splays apart at the end distal from the \u03b2-turn, resulting in only 4 intrapeptide hydrogen bonds stabilizing the hairpin instead of 5 found in the HCV1 epitope (Fig. 1D). Accordingly, AP33 buries less surface area around the termini than HCV1 (Fig. 2D).\n\nA direct comparison between AP33 and HCV1 structures reveals that the antibodies approach this antigenic site from different directions. When the epitopes are structurally superposed, the antibodies bind with a 22\u00b0 difference in the angle of approach (Fig. 2A). Although both peptide epitopes bind in the cleft between VH and VL of the antibodies (Fig. 2B, top), in the HCV1 structure, the tip of the \u03b2-hairpin points toward VL (158-\u00c52 buried surface), while the majority of the \u03b2-hairpin interacts with VH (300 \u00c52) (Fig. 2B, bottom). In contrast, VL and VH of AP33 interact almost equally with the N- and C-terminal \u03b2-strands of the antigen (242 \u00c52 and 273 \u00c52, respectively). This difference in VL usage is highlighted when the backbone atoms of the epitopes are superposed: Gln412-Asn417, which interact with VH of both AP33 and HCV1, are highly similar between the two structures, while differences are more apparent in Ser419-Asn423, which interact mainly with VL of AP33 (Fig. 2C and F).\n\nDespite these differences, it is clear that two independent antibody selection and maturation pathways arrived at a similar solution to engage the hydrophobic \u03b2-hairpin epitope: both events rely on hydrophobic interactions along the cleft between VL and VH. Both antibodies use primarily their side chains for interactions (Fig. 2D); coincidentally, the relatively few backbone interactions are mediated by their L3 loops (Fig. 2E). In both structures, a hydrogen bond is formed with the Gly418 tip of the peptide (Fig. 1C) and the major hydrophobic interactions are with Leu413 and Trp420 (see Tables S2 and S3 in the supplemental material). These interactions define the essential features of this antigenic site for broad neutralization of HCV.", "# Spacial Score\u2500A Comprehensive Topological Indicator for Small-Molecule Complexity\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-13 on marcdeller.com (categories: blog, sb).*\n\n* [Spacial Score\u2500A Comprehensive Topological Indicator for Small-Molecule Complexity](https://pubs.acs.org/doi/10.1021/acs.jmedchem.3c00689)\n\n* **Optimal Range:** Compounds with nSPS between 20 and 40 show the highest potency and selectivity\n* **Granular Metric:** SPS integrates hybridization, stereogenicity, ring systems, and branching per atom\n* **Size Normalization:** Dividing SPS by heavy\u2010atom count yields nSPS for fair comparisons\n* **Predictive Power:** nSPS matches or outperforms Fsp\u00b3 and FCstereo in ROC analyses\n* **Library Differentiation:** nSPS distinguishes natural products, screening collections, and GDB-17 entries\n* **Synthetic Planning:** RDKit implementation allows tracking complexity growth during synthesis\n\n1. **Spacial Score\u2500A Comprehensive Topological Indicator for Small-Molecule Complexity:** [Krzyzanowski A., et al., Journal of Medicinal Chemistry](https://pubs.acs.org/doi/10.1021/acs.jmedchem.3c00689)\n2. **On degree-based operators and topological descriptors of molecular systems:** [Ponnusamy V., et al., Scientific Reports](https://www.nature.com/articles/s41598-024-72621-7)\n3. **Enhanced molecular descriptors using degree-distance invariants:** [Zhang L., et al., Journal of Computational Chemistry](https://www.sciencedirect.com/science/article/abs/pii/S0022286024012493)\n4. **Molecular Complexity: You Know It When You See It:** [Cleves A.E., et al., Journal of Chemical Information and Modeling](https://www.sciencedirect.com/org/science/article/pii/S1520480423003472)\n5. **Topological modeling and QSPR-based prediction of bioactivity:** [Zhao Y., et al., Scientific Reports](https://www.nature.com/articles/s41598-025-11863-5)\n6. **rdkit.Chem.SpacialScore module implementation:** [Landrum G., et al., Journal of Cheminformatics](https://rdkit.org/docs/source/rdkit.Chem.SpacialScore.html)\n\n**Discover how a new, granular metric transforms our understanding of molecular complexity and guides drug design toward more potent and selective candidates**\n\nSmall-molecule complexity has traditionally been gauged by simple metrics like the fraction of sp\u00b3-hybridized carbons (*F*sp\u00b3) and the fraction of stereogenic carbons (*F*Cstereo). While valuable, these scores often fail to capture the nuances of molecular topology and chemists\u2019 intuitive sense of complexity. The **Spacial Score (SPS)** and its **size-normalized version (nSPS)**address these limitations by integrating four atom-level terms\u2014hybridization, stereogenicity, ring membership, and branching\u2014into a single, highly granular indicator that correlates with biological activity and guides synthesis planning.\n\nThe SPS is computed per heavy atom (i) as:\n\nSPS=\u2211i(hi\u00d7si\u00d7ri\u00d7ni2)\n\nwhere h\u00a0(hybridization term) rewards sp\u00b3 versus sp\u00b2/sp atoms, s\u00a0(stereoisomeric term) emphasizes stereogenic centers and E/Z isomerism, r\u00a0(non-aromatic ring term) highlights non-aromatic cycles, and n\u00a0(neighbor count) captures skeletal branching. Dividing SPS by the total heavy atoms yields nSPS, facilitating comparisons across molecules of different sizes.\n\nApplying nSPS to datasets such as ChEMBL reveals a **clear trend**: low nSPS (<11.2) compounds are often low-activity, whereas **moderate to high nSPS compounds (20\u201340)** exhibit the highest proportion of potent (pChEMBL \u2265 6.5) and selective ligands. Notably, exceeding an nSPS of ~40 may reduce potency, suggesting an optimal complexity \u201cwindow\u201d for drug candidates. Natural products typically populate the higher end of this range, reflecting their intricate topologies, while synthetic libraries span lower to mid nSPS values.\n\nBeyond activity correlations, nSPS distinguishes between disparate compound collections\u2014natural products, commercial libraries, dark chemical matter, and enumerated sets like GDB-17\u2014underscoring its ability to capture genuine topological diversity. Moreover, comparisons with *F*sp\u00b3 and *F*Cstereo via ROC analysis demonstrate that nSPS matches or exceeds these traditional metrics in predicting potency and promiscuity (AUC ~0.61 vs 0.58\u20130.62).\n\nFinally, SPS aids **synthesis planning** by quantifying topological changes across reaction sequences, enabling chemists to monitor complexity growth in total syntheses or library design. Its RDKit implementation supports rapid calculation from SMILES strings, making nSPS readily accessible for medicinal chemistry workflows.\n\n| Key Concept | Description | Reference |\n| --- | --- | --- |\n| Hybridization Term (h) | Weights atoms by hybridization: sp\u00b3, sp\u00b2, sp, others | [Krzyzanowski A., et al., Journal of Medicinal Chemistry](https://pubs.acs.org/doi/10.1021/acs.jmedchem.3c00689) |\n| Stereoisomeric Term (s) | Emphasizes stereogenic centers and E/Z isomers | [Ponnusamy V., et al., Scientific Reports](https://www.nature.com/articles/s41598-024-72621-7) |\n| Non-Aromatic Ring Term (r) | Highlights non-aromatic ring atoms vs aromatic | [Zhang L., et al., Journal of Computational Chemistry](https://www.sciencedirect.com/science/article/abs/pii/S0022286024012493) |\n| Branching Term (n\u00b2) | Squares number of heavy-atom neighbors to capture branching | [Cleves A.E., et al., Journal of Chemical Information and Modeling](https://www.sciencedirect.com/org/science/article/pii/S1520480423003472) |\n| Normalization (nSPS) | Divides SPS by total heavy atoms for size-independent comparison | [Zhao Y., et al., Scientific Reports](https://www.nature.com/articles/s41598-025-11863-5) |\n| Implementation | RDKit module enabling SPS/nSPS calculation from SMILES strings | [Landrum G., et al., Journal of Cheminformatics](https://rdkit.org/docs/source/rdkit.Chem.SpacialScore.html) |", "# Crystal Structure of TrkA-N Domain Protein from Thermotoga sp. RQ2\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [3JXO](https://www.rcsb.org/structure/3JXO)\n* **Release Date:** 2010-03-31\n* **Method:** X-RAY DIFFRACTION 1.55 \u00c5\n* **Organisms:** Thermotoga sp. RQ2\n* **Macromolecules:** TrkA-N domain protein (protein)\n* **Publication:** [Deller, M.C., et al. (2015). Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly. PLoS ONE, 10, e0122512.](https://doi.org/10.1371/journal.pone.0122512)\n* **Blog Post:** [Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly](https://marcdeller.com/crystal-structure-of-a-two-subunit-trka-octameric-gating-ring-assembly/)\n\nDescription: This high-resolution (1.55 \u00c5) component structure of the TrkA-N domain from Thermotoga sp. RQ2 provides the atomic-level detail of the individual subunit that assembles into the octameric K\u207a channel gating ring. Together with the full assembly structure ([3L4B](https://www.rcsb.org/structure/3L4B)), it establishes the structural framework for understanding potassium transport regulation in thermophilic bacteria.", "# GitHub README: TopPDBLX\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/TopPDBLX, 2026\n\n- [x] Ingest the whole archive, with a byte-level fidelity gate\n- [x] Deterministic rule parser with a stable discard taxonomy\n- [x] Commercial screen library, extracted verbatim from vendor sources\n- [x] Sequence linkage and MMseqs2 redundancy control\n- [x] Assemble the release in five formats, with a generated datasheet\n- [x] Settle licensing: CC-BY-4.0 data, MIT code\n- [x] Complete the 90 GB archive snapshot and run the full-scale fidelity check\n- [x] Reagent lexicon: two curation rounds plus the ionic liquids, 147 to 526 reagents (499 today, after merging what was never distinct)\n- [x] Seven-class condition ontology, replacing the withdrawn three-level version\n- [x] Fine-tune SmolLM2 on the parse residual, and strip narrative prose in the parser\n- [x] Classification accuracy audit, two rounds of 96 (spec 6.6): 85.4% rules-derived, 83.3% model-derived, pooled\n- [x] Apply the trained model over the residual: round 06 generated 153,736 components and lifted classified coverage from 59.5% to 80.2%\n- [x] **Regenerate the corpus with round 09** (5.1 hours over all 47,845 residual records): **188,655 components, 22.7% more than round 06**, and classified coverage 80.2% to **80.8%**\n- [x] **Publish the model's recovered chemistry, with provenance:** `release.assemble` now merges the residual parser's components into the released table behind a `parser` column (`rules` or `slm`). The rule parser wins wherever it succeeded, so only genuinely new reagents are added. **Components 610,076 to 645,656; identified 86.5% to 87.2%; conditions classified 143,626 to 152,006.** Before this, the model's output never reached anyone who downloaded the dataset, and the README said otherwise\n- [x] Gold set of 96 hand-labelled records, and the first precision/recall this project has had\n- [x] Second gold set of 96, sampled where the pipeline and the teacher disagree\n- [x] Add the reagents the gold sets named that the lexicon could not place\n- [x] Tell the drop from the protein sample: `protein_buffer` and `soak` roles, 6,800 components\n- [x] De-duplicate the lexicon: PEG MME was seven canonical ids per molecular weight; 36 merged, five invariants now tested\n- [x] Remove the aliases naming a *different* molecule: barium/yttrium, CAPS/CHAPS, DTT/DTE and four more\n- [x] Transcribe the Morpheus stock table from the vendor brochure: 15 premixes with real compositions\n- [x] Classify premixes on their constituents rather than refusing them: coverage 76.8% to 80.2%\n- [x] Split reagents written with no separator at all: +7,553 identified components, nothing lost\n- [x] **Teacher distillation, resolved:** round 09 adds the 32B's labels to the rules labels instead of replacing them, at 32 LoRA layers matching round 06. Pooled recall against hand-labelled truth rises from 85.7% to 90.3%, finding 28 more reagents; rounds 07 and 08 had changed the label source and halved LoRA capacity in the same experiment\n- [x] **Ship round 09, preferring recall over a precision-weighted score:** it finds more chemistry and misfiles slightly more of it, and the extra errors are role misattributions (a real reagent in the wrong scope), not inventions\n- [x] Confirm training length matters: on the 2,000-record frozen benchmark, round 09's identification and fully-identified records keep rising from true iteration 1,000 to 8,000; the 192-record gold sets were too small to resolve the same gap clearly\n- [x] **Choose the round 09 checkpoint on the larger benchmark, not the smaller one:** the 192-record gold sets picked iteration 4,500 on recall, and the 2,000-record frozen benchmark showed it identifying 2.6 points fewer reagents and leaving 7.3 points fewer records fully identified. The final adapter (iteration 8,000) ships\n- [x] **Re-generate the corpus with round 09** (5.1 hours over 47,845 residual records): 188,655 components against round 06's 153,736, and classified coverage 80.2% to 80.8%\n- [x] **Probe what the parsers still cannot name**, by labelling 464 records carrying an unnamed component with a 32B teacher. **17.7% yield a reagent the release lacks** (CI 14.2 to 21.1), so the remaining 82,341 unnamed components are mostly not recoverable chemistry: the source text says \"PEG\" or \"phosphate\" without saying which, or names no reagent at all. One genuine lexicon gap found and closed (ACES, lexicon 0.9.3)\n- [ ] Cut the 49 reagents round 09 still misses, and the 46 it misfiles: the misfiles are role errors the `protein_buffer` and `soak` classes should absorb\n- [ ] Supply pKa values for 59 buffers, or fix the clause splitter that produced them\n- [x] **Publish to Zenodo for a citable DOI:** v1.0.0 deposited 2026-08-05, [10.5281/zenodo.21807134](https://doi.org/10.5281/zenodo.21807134). Schema and dataset versions frozen at 1.0.0; eight files, each checksum-verified on upload\n- [x] **Mirror on HuggingFace Datasets:** [`Dellboy/toppdblx-conditions`](https://huggingface.co/datasets/Dellboy/toppdblx-conditions), public, with a dataset card carrying the provenance split and the limitations\n- [x] **Browser front end**, an exploration tool rather than a predictor: [toppdblx.mdeller.com](https://toppdblx.mdeller.com), single file, no build step\n- [x] **Construct boundary model:** MCC 0.669 and a 9-residue median error on held-out proteins, and it survived five separate attempts to beat it\n- [x] **Crystallisation propensity from TargetTrack:** AUC 0.656, and the top 5% of ranked constructs crystallise at 2.4x the base rate\n- [x] **Condition recommender, killed on its own criterion:** worse than a query-independent frequency prior on every measure, and worse with 32x the data. Published as a negative result rather than tuned\n- [x] **Dataset paper drafted** ([`PAPER.md`](PAPER.md)), every figure re-derived from the released artefacts\n- [ ] Submit the paper, and mint a Zenodo version for the revision that accompanies it\n\n## \ud83d\udcc4 Licence", "# Beyond the Cell: The New Frontier of Extracellular Protein Degradation\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* [**LYTAC Platform: Lycia Therapeutics**](https://lyciatx.com/technology/lytac-platform/)\n\n* **First FDA-approved PROTAC:** Vepdegestrant validates targeted protein degradation as a clinical modality, not just a research tool.\n* **The intracellular ceiling:** Classical PROTACs cannot reach the ~40% of the proteome that lives outside the cell.\n* **LYTACs close the gap:** Lysosome-targeting chimaeras and related eTPD modalities route extracellular and membrane-bound targets to the lysosome.\n* **Designed proteins join the toolkit:** EndoTags from the Baker and Bertozzi labs enable genetically encoded degraders without native ligands.\n\n1. **Arvinas Announces FDA Approval of VEPPANU (vepdegestrant) for the Treatment of ESR1m, ER+/HER2- Advanced Breast Cancer:** [Arvinas Press Release, May 2026](https://ir.arvinas.com/news-releases/news-release-details/arvinas-announces-fda-approval-veppanu-vepdegestrant-treatment)\n2. **Targeted Degradation of Extracellular Proteins: State of the Art and Diversity of Degrader Designs:** [Mamun et al., Journal of Hematology & Oncology, 2025](https://jhoonline.biomedcentral.com/articles/10.1186/s13045-025-01703-4)\n3. **Extracellular Targeted Protein Degradation: An Emerging Modality for Drug Discovery:** [Wells & Kumru, Nature Reviews Drug Discovery, 2024](https://www.nature.com/articles/s41573-023-00833-z)\n4. **Lysosome-Targeting Chimaeras for Degradation of Extracellular Proteins:** [Banik et al., Nature, 2020](https://www.nature.com/articles/s41586-020-2545-9)\n5. **Designed Endocytosis-Inducing Proteins Degrade Targets and Amplify Signals:** [Huang et al., Nature, 2025](https://www.nature.com/articles/s41586-024-07948-2)\n\n**The first FDA-approved PROTAC strengthens the case for LYTACs and the next generation of extracellular degraders**\n\nThe first FDA approval of a PROTAC has just landed. On 1 May 2026, the agency cleared [Arvinas and Pfizer\u2019s vepdegestrant (Veppanu)](https://ir.arvinas.com/news-releases/news-release-details/arvinas-announces-fda-approval-veppanu-vepdegestrant-treatment) for adults with ESR1-mutant, ER-positive, HER2-negative advanced breast cancer, ahead of its PDUFA date and on the strength of a 43% reduction in the risk of disease progression versus fulvestrant in the Phase 3 VERITAC-2 trial. After two decades of work tracing back to Craig Crews and Ray Deshaies\u2019 original 2001 chimaera, an event-driven, catalytic degrader has cleared the regulatory bar. Targeted protein degradation has now been proven not only as a chemical biology tool, but as a medicine.\n\nThat medicine, however, works inside the cell. PROTACs hijack the ubiquitin-proteasome system, so by design they reach only intracellular targets. Roughly forty percent of the human proteome sits outside that compartment, on the plasma membrane or in the extracellular milieu, and the proteasome cannot touch any of it. Closing that gap requires [a different cellular machinery altogether](https://www.nature.com/articles/s41573-023-00833-z): receptor-mediated endocytosis and the lysosome.\n\n> **\u201cThis milestone demonstrates that targeted protein degradation can translate into meaningful clinical impact.\u201d**\n\nThat alternative arrived in 2020 with [LYTACs](https://www.nature.com/articles/s41586-020-2545-9), bifunctional molecules that tether an extracellular target to a lysosome-shuttling receptor such as the cation-independent mannose-6-phosphate receptor or, in the liver, the asialoglycoprotein receptor. The toolkit has expanded fast: AbTACs and PROTABs co-opt cell-surface E3 ligases like RNF43 and ZNRF3; KineTACs route cargo through cytokine receptors; ATACs use peptide warheads against integrins. Most recently, [Huang, Bertozzi, and Baker reported computationally designed EndoTags](https://www.nature.com/articles/s41586-024-07948-2) that trigger endocytosis without any native ligand, making fully genetically encoded degraders feasible.\n\nThe implications track with the proteome itself. Cytokines, immune checkpoints, secreted oncogenic ligands, and tumour cell-surface receptors that resisted both small-molecule and antibody approaches now have a credible degradation path, as catalogued in a recent [comprehensive review of eTPD design diversity](https://jhoonline.biomedcentral.com/articles/10.1186/s13045-025-01703-4). Vepdegestrant\u2019s clearance will not, on its own, derisk LYTACs in the clinic; biodistribution, manufacturability, and immunogenicity remain real engineering problems. But it does establish that regulators will approve degraders on the same evidentiary basis as any other oncology agent. For the wave of extracellular degraders now entering IND-enabling studies, that is the validation the field was waiting for.", "# GitHub README: TopPDBLX\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/TopPDBLX, 2026\n\n**Identification is the weakest of the four.** A model reading `20% PEG 3350` and emitting `SODIUM_CHLORIDE` scores as identified: the name is real, it is simply not the reagent in front of it. Since the whole purpose of curation is to make names mean something, a measure that cannot tell a real name from the correct one is not measuring the thing that matters.\n\n**Grounding closes that hole and needs no labels.** If the model names a reagent, some spelling of that reagent should appear in the source string. Punctuation and spacing are stripped on both sides, so `peg-3350`, `PEG3350` and `peg 3,350` all match one alias. A failure is either a hallucination or a spelling the lexicon has never seen, and both are worth knowing about, so ungrounded cases are written out for inspection rather than only counted.\n\n**Every one of those four is precision-shaped.** Each asks whether what the parser *said* is defensible; none can see a reagent it never mentioned at all, so a parser that reads one component per record and stops scores well on all four. Measured directly: on 4,000 residual records, 67% contain more reagent clauses than the parser emitted components. That blind spot is exactly what the gold sets and the frozen benchmark below exist to close.\n\n## \ud83c\udfc5 Results against hand-labelled truth\n\nEverything else in this project checks whether an answer is defensible. This section checks whether it is *right*, against two kinds of hand-labelled ground truth.\n\n**Two gold-set batches, always pooled.** A random 96 records answers \"how good is the pipeline overall\", but rarely contains a reagent the sources disagree on, so it under-tests precision. A second 96, sampled specifically where the teacher and the rule parser disagree, supplies that. Both are labelled by hand in `gold_bench_v1.html`, one reagent list per condition, and both are pooled (192 records, 616 reagents) rather than averaged, so the headline figures below are one number with a real confidence interval. Neither batch is ever trained on, and both are excluded from every teacher run by name.\n\n| source | found | missed | falsely added | precision | recall |\n|---|---|---|---|---|---|\n| rule parser alone | 428 | 188 | 22 | 95.1% | 69.5% |\n| rules + round 06 | 528 | 88 | 26 | 95.3% | 85.7% |\n| rules + round 09 (shipped, final adapter) | 556 | 60 | 40 | 93.3% | 90.3% |\n\n*Both batches pooled, measured 2026-08-04 at lexicon 0.9.2.*\n\n| Measure | Asks | Fails when |\n|---|---|---|\n| **Precision** | Of the reagents claimed, how many really are in the condition? | The parser invents chemistry |\n| **Recall** | Of the reagents really in the condition, how many did it find? | The parser misses chemistry |\n\n**Round 09 finds 28 more reagents than round 06 and misses 32% fewer.** Its extra false positives are not inventions: every one names a reagent genuinely present in the source text (a protein storage buffer, a soak, a cryoprotectant) placed in the wrong role, which is what the `protein_buffer` and `soak` scope roles exist to fix. In short: **round 09 finds more chemistry and misfiles slightly more of it.**\n\n**For the crystallographer:** the gold sets are a small yardstick with wide intervals. They measure reagent *identity*, not concentration or unit, and they don't replace the commercial screen cross-reference (81,802 component-set matches) or the human audit.\n\n### The frozen benchmark: 2,000 held-out residual records\n\nThe gold sets are precise enough to rank sources but, at 192 records, too small to resolve some real differences: see [Training](#-training). A larger, held-out slice of the residual (2,000 records, never used for training or checkpoint selection) is used to compare rounds directly, including across a single round's own checkpoints:\n\n| | round 06 | r09 @ 1,000 | r09 @ 4,500 | **r09 @ 8,000** |\n|---|---|---|---|---|\n| identification | 89.05% | 89.76% | 92.67% | **95.28%** |\n| grounding | 93.15% | 93.04% | 94.06% | **94.83%** |\n| fully-identified records | 62.85% | 65.8% | 73.75% | **81.0%** |\n| fidelity (exact match) | 91.95% | 80.5% | 90.9% | **94.38%** |\n| distinct unidentified names | 629 | 635 | 462 | **294** |\n\n*Measured 2026-08-04 at lexicon 0.9.2. The shipped checkpoint is round 09's final adapter, true iteration 8,000.*\n\n**Every measure improves monotonically with training, and that is why the final adapter ships.** Iteration 4,500 leads on gold-set recall (92.0% against the final adapter's 90.3%), which on the 192-record yardstick alone would have selected it. On 2,000 records it identifies 2.6 points fewer reagents, leaves 7.3 points fewer records fully identified, and emits 168 more names the lexicon cannot place. Trading that for 1.7 points of recall would buy components while leaving more of them unresolvable, which is not what completeness means here.\n\n## \ud83e\udd17 Every round, and what it delivered\n\nAll rounds ship as LoRA adapters for `mlx-community/SmolLM2-360M-Instruct`, one directory per round in **[`Dellboy/toppdblx-residual-parser`](https://huggingface.co/Dellboy/toppdblx-residual-parser)**. Each is 33 MB; the base model is not redistributed. Rounds that didn't reach the corpus are listed anyway, since the point is the arc.", "# Multi-Step LLM Prompt: From Protein Sequence to Validated Protein Structure\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-17 on marcdeller.com (categories: sb).*\n\n* LLMs enable multi-step workflows from raw protein sequence to validated 3D model, combining language and structure understanding.\n* Secondary, tertiary structure prediction and validation can all be orchestrated through evolving AI prompts and specialized models.\n* Key validation metrics include Ramachandran plots, MolProbity scores, and checks for stereochemical plausibility.\n* Experimental techniques like cryo-EM and X-ray crystallography remain indispensable for definitive structural confirmation.\n* This approach accelerates drug design, understanding disease mechanisms, and fundamental biology by making high-quality models broadly accessible.\n\n1. **ProteinGPT: Multimodal LLM for Protein Property Prediction and Structure Understanding:**\n   <https://arxiv.org/abs/2408.11363>\n2. **Deep learning methods for protein function prediction (review, including structure validation):**\n   <https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11735672/>\n3. **AlphaFold Protein Structure Database (AI-based predictions, experimental validation context):**\n   <https://alphafold.ebi.ac.uk>\n4. **Advanced Protein Structure Validation Techniques:**\n   <https://www.numberanalytics.com/blog/advanced-protein-structure-validation-techniques>\n5. **PDB File Format, The B-Factor (example structure output and explanation):**\n   <https://www.proteinstructures.com/pdb-file-format/>\n\nProteins put life into motion. Yet, the journey from the raw string of a protein\u2019s amino acid sequence to its intricate 3D structure is a grand biological puzzle. Thanks to the explosive innovation in artificial intelligence\u2014particularly large language models (LLMs)\u2014that journey is now being charted faster and more accurately than ever before.\n\n## Step 1: From Sequence to Secondary Architecture\n\nIt all starts with a sequence. Imagine each letter, representing an amino acid, as a note in a vast molecular symphony. The first computational step identifies recurring motifs\u2014helices, sheets, and coils\u2014each hinting at how the protein will ultimately fold. Modern LLMs, trained on millions of protein examples, quickly spot these patterns. They annotate the positions and types of secondary structure elements, providing the \u201cscaffolding\u201d for what follows.\n\n## Step 2: Predicting the 3D Tertiary Structure\n\nWith secondary structure in hand, the LLM collaborates with specialized models like AlphaFold or ProteinGPT. These systems use deep neural networks and \u201cattention mechanisms\u201d to predict how those architectural elements collapse into a final, compact, and functional shape. The result\u2014a coordinate map of all the atoms\u2014can be output in the universally recognized Protein Data Bank (PDB) file format, ready for visualization or simulation.\n\n## Step 3: Rigorous Structure Validation\n\nPrediction isn\u2019t enough. Structural bioinformatics relies on robust validation. Computational checks mimic the questions an experimentalist would ask: Are the bond angles realistic? Do the backbone torsion angles fall into allowed Ramachandran regions? Are there steric clashes? Modern LLM-based systems report detailed metrics\u2014like MolProbity score and stereochemical outlier counts\u2014that allow researchers to weigh the model\u2019s accuracy, often rivaling the quality seen in X-ray structures.\n\n## Step 4: Towards Experimental Confirmation\n\nWhile computational models are powerful, experimental validation still reigns supreme. The blog describes X-ray crystallography (for static, high-atom-resolution maps), cryo-electron microscopy (for huge complexes and dynamics), and NMR spectroscopy (for flexible and small proteins) as gold standards. Models that pass the computational scrutiny outlined above are prime candidates for such experimental work, often accelerating the process by highlighting likely conformations.\n\n## Step 5: Synthesis and Utility\n\nThe workflow from sequence to validated structure empowers scientists in fields ranging from drug design to evolutionary biology. By relying on scalable, interpretable steps\u2014secondary structure inference, 3D prediction, validation, and hypothesis-testing\u2014the modern LLM-centric approach opens new horizons for discovery.\n\n## Conclusion\n\nAs we move into the era of AI-driven science, the marriage of language models and protein bioinformatics promises both speed and accuracy. The steps above don\u2019t just offer a recipe for prediction\u2014they constitute a scientific dialogue between biologists and their data, mediated by AI.", "# GitHub README: TopPDBLX\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/TopPDBLX, 2026\n\nLarge files live outside the repository. Point the data directories somewhere that is not inside an iCloud-synced folder, because macOS Optimize Mac Storage will silently evict them:\n\n```bash\nmkdir -p ~/TopPDBLXData/{raw,interim,processed}\nln -sfn ~/TopPDBLXData/raw       data/raw\nln -sfn ~/TopPDBLXData/interim   data/interim\nln -sfn ~/TopPDBLXData/processed data/processed\n```\n\n## \ud83d\ude80 Usage\n\nEvery stage is a single command and writes its own manifest.\n\n```bash\n./run.sh --list                    # show all stages\n./run.sh ingest.entry_ids          # snapshot the PDB entry id list\n./run.sh ingest.fetch_entries      # batched GraphQL harvest (about 18 minutes)\n./run.sh ingest.flatten            # raw JSON to parquet\n./run.sh ingest.validate_fidelity  # gate: API text against archive mmCIF\n./run.sh parse.run_parser          # free text to typed components\n./run.sh assign.classify           # the seven JCSG Top96 precipitant classes\n./run.sh assign.screen_match       # cross-reference commercial screens\n./run.sh link.representative       # one sequence per entry\n./run.sh link.cluster              # MMseqs2 at 30%, 50% and 90%\n./run.sh release.assemble          # build the released database\n./run.sh release.datasheet         # regenerate the datasheet\n```\n\n### Pipeline stages\n\n| Stage | Purpose |\n|---|---|\n| `ingest.entry_ids` | Snapshots the entry id list: this defines the archive version for the release |\n| `ingest.fetch_entries` | Batched GraphQL harvest, resumable, raw responses kept as provenance |\n| `ingest.flatten` | One row per `(pdb_id, crystal_id)`, never index 0 of the crystal-grow loop |\n| `ingest.validate_fidelity` | Byte-compares API text against archive mmCIF, including loop row counts |\n| `ingest.targettrack` | Acquires and checksums the archived TargetTrack negative set for Phase 3 |\n| `parse.mine_reagents` | Ranks candidate reagent names by corpus mass, to drive lexicon curation |\n| `parse.lexicon_coverage` | Measures what share of the corpus the lexicon identifies |\n| `parse.run_parser` | The deterministic rule parser, over the whole archive |\n| `assign.build_screens` | Extracts screen formulations verbatim from vendor PDFs |\n| `assign.screen_match` | Matches parsed conditions to screen wells, schema to schema |\n| `link.representative` | Chooses one sequence per entry, never one per chain |\n| `link.cluster` | MMseqs2 clustering at three identity thresholds |\n| `link.sifts`, `link.uniprot` | PDB to UniProt mapping, and full-length reference sequences |\n| `parse.lexicon_questions` | Ranks unidentified names by corpus mass into a few dozen curation calls |\n| `parse.apply_lexicon_answers` | Folds audit answers back into the lexicon, with the reasoning recorded |\n| `parse.curation_queue` | Groups the remaining gaps into ten bucketed decisions |\n| `parse.apply_curation_queue` | Applies a bucketed answer to every name in its bucket |\n| `assign.classify` | The seven JCSG Top96 precipitant classes, or Unclassified with a reason |\n| `eval.class_audit` | Accuracy audit, 96 conditions split by provenance, one decision each |\n| `eval.gold_questions` | Samples records for hand labelling, banded by text length |\n| `eval.gold_metrics` | Precision and recall against the gold sets, per source and pooled |\n| `eval.splits` | Cluster-based train, validation and test splits, leak-free at all three thresholds |\n| `eval.baselines` | Frequency prior and homology retrieval, the baselines any model must beat |\n| `models.build_slm_dataset` | Bootstrap training data for the residual parser, no hand labelling |\n| `models.train_slm` | LoRA fine-tune under MLX-LM, W&B logged, one named run per round |\n| `models.eval_slm` | Fidelity and residual identification, with Wilson intervals on both gates |\n| `models.apply_slm` | Runs the model over the residual and writes the components it reads |\n| `models.teacher_label` | Labels the residual with a local 32B, to train past the rule parser's ceiling |\n| `release.assemble` | Builds the released database in five formats |\n| `release.snapshot` | Frozen 90 GB mmCIF archive, for the full-scale fidelity check |\n| `release.verify_archive` | Compares the parsed source field against the archive, for every entry |\n\n## \ud83d\udce6 Output\n\n| File | Contents |\n|---|---|\n| `toppdblx-conditions-v1.0.0.jsonl.gz` | Canonical form, one nested record per line |\n| `toppdblx-conditions-v1.0.0.parquet` | Record level, sequence linkage flattened on |\n| `toppdblx-components-v1.0.0.parquet` | One row per reagent |\n| `toppdblx-components-v1.0.0.csv.gz` | The same, for spreadsheets |\n| `toppdblx.duckdb` | Both tables, plus `usable_conditions` and `condition_components` views |\n| `schema-v1.0.0.json` | JSON Schema, generated from the pydantic model |\n\n**Every component carries a `parser` field**, `rules` or `slm`, so the two sources are always separable:\n\n| `parser` | rows | what it is |\n|---|---|---|\n| `rules` | 610,076 | The deterministic parser. Filter to this for a fully reproducible subset containing no model output |\n| `slm` | 35,580 | Chemistry the rules left unnamed and the residual parser recovered, across 26,339 records |\n\n**The rule parser wins wherever it succeeded.** A model row is published only when it names a reagent the rules did *not* name for that record: of round 09's 149,913 named rows, 113,943 simply restate a reagent the rules already had and are discarded. The rules' unnamed rows (an amount with no reagent attached) are kept rather than replaced, because matching them to a model row on concentration and unit succeeds for only 24% of cases, and inventing that correspondence would be worse than recording the gap honestly.\n\nUntil 2026-08-05 the released table was rule-parser output only, and this README described the model's contribution incorrectly: the \"components shipped\" figures quoted for round 06 counted rows in the model's own working file, not rows anyone could download.", "# Genomic Giants: The Lungfish That Rewrote DNA Records\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-11 on marcdeller.com (categories: blog).*\n\n* [Big fish!](https://www.science.org/content/article/odd-fish-has-30-times-much-dna-humans-new-record-animals)\n\n* **Genomic Scale**: *Lepidosiren paradoxa* possesses 91 billion DNA bases, establishing the largest animal genome record\u200b\n* **Expansion Rate**: The genome adds one human genome equivalent every 10 million years through transposable element proliferation\n* **Regulatory Breakdown**: Loss of piRNA and KRAB-ZFP suppression systems enabled unchecked transposable element activity\u200b\n* **Chromosomal Giants**: Eighteen of 19 chromosomes individually exceed the entire human genome size\u200b\n* **Evolutionary Insight**: Lungfish genomes illuminate the molecular basis of the fish-to-tetrapod transition\u200b\n* **Active Elements**: Transposable elements remain functional today, ensuring continued genome expansion\u200b\n\n1. Giant lungfish genome elucidates the conquest of land by vertebrates: [Schartl, M., et al., Nature](https://www.nature.com/articles/s41586-021-03198-8)\n2. The genomes of all lungfish inform on genome expansion and tetrapod evolution: [Schartl, M., et al., Nature](https://pubmed.ncbi.nlm.nih.gov/39143221/)\n3. This odd fish has 30 times as much DNA as humans\u2014a new record for animals: [Pennisi, E., Science](https://www.science.org/content/article/odd-fish-has-30-times-much-dna-humans-new-record-animals)\n4. Effective population size does not explain long-term variation in genome size and transposable element content: [Marino, A., et al., eLife](https://elifesciences.org/reviewed-preprints/100574)\n5. Transposable elements as genome regulators in normal and malignant hematopoiesis: [Marasca, F., et al., Nature Reviews Cancer](https://pmc.ncbi.nlm.nih.gov/articles/PMC12056191/)\n6. From Junk DNA to Genomic Treasure: Impacts of Transposable Elements on Animal Genome Evolution: [Chen, L., et al., International Journal of Molecular Sciences](https://pmc.ncbi.nlm.nih.gov/articles/PMC12350819/)\n\nNature delivered a stunning surprise in 2024 when scientists unveiled the genomic champion of the animal kingdom. The South American lungfish, *Lepidosiren paradoxa*, carries a DNA blueprint containing 91 billion chemical letters\u201430 times larger than the human genome. This ancient air-breathing fish has accumulated genetic material at a staggering rate, adding the equivalent of an entire human genome every 10 million years over the past 100 million years.\n\nWhat makes this discovery particularly fascinating is that despite harboring this enormous genetic library, the lungfish contains roughly the same number of protein-coding genes as humans\u2014approximately 20,000. The remaining 90% of its genome consists of repetitive DNA sequences called transposable elements, or \u201cjumping genes,\u201d that have run rampant throughout the species\u2019 evolutionary history. These parasitic genetic elements copy themselves and insert new copies randomly throughout the genome, creating an ever-expanding molecular landscape that challenges our understanding of genomic organization.\n\nThe mechanism behind this unprecedented expansion involves the breakdown of cellular quality control systems. Unlike other vertebrates that maintain molecular brakes on transposable element activity through piRNA pathways and specialized protein networks, lungfishes have lost key regulatory genes. This evolutionary oversight has allowed transposable elements to proliferate unchecked, transforming the lungfish genome into a repetitive DNA warehouse. Remarkably, evidence suggests these elements remain active today, meaning the genome continues its relentless expansion.\n\nThis genomic gigantism carries profound biological costs. Each of the lungfish\u2019s 19 chromosomes rivals the size of the entire human genome, requiring enormous cellular energy to replicate and maintain. The nuclei housing this massive DNA must expand to accommodate the genetic cargo, and cell division becomes a marathon of molecular machinery working overtime. Yet somehow, these ancient fish have survived for hundreds of millions of years with their bloated genomes, representing one of evolution\u2019s most extreme experiments in genetic excess.\n\nThe lungfish genome offers unique insights into vertebrate evolution. As the closest living relatives to the first land animals, lungfishes serve as molecular time capsules preserving genetic information about the epic transition from water to land 370 million years ago. By comparing the genomes of South American, African, and Australian lungfish species, researchers have traced how limb-like fins evolved and sometimes disappeared, providing clues about the developmental toolkit that enabled vertebrates to colonize terrestrial environments.\n\n| Concept | Description | Key References |\n| --- | --- | --- |\n| Transposable Elements | Self-replicating DNA sequences that copy and insert throughout genomes | [Chen, L., et al., International Journal of Molecular Sciences](https://pmc.ncbi.nlm.nih.gov/articles/PMC12350819/) |\n| piRNA Pathway | Cellular defense system that suppresses transposable element activity | [Morrison, J., et al., Developmental Cell](https://pmc.ncbi.nlm.nih.gov/articles/PMC9365273/) |\n| Genome Expansion | Evolutionary increase in genome size through repetitive DNA accumulation | [Schartl, M., et al., Nature](https://www.nature.com/articles/s41586-021-03198-8) |\n| C-value Paradox | Lack of correlation between genome size and organism complexity | [Marino, A., et al., eLife](https://elifesciences.org/reviewed-preprints/100574) |\n| Lungfish Evolution | Ancient air-breathing fish representing early vertebrate land transition | [Schartl, M., et al., Nature](https://pubmed.ncbi.nlm.nih.gov/39143221/) |\n| Junk DNA | Non-coding genomic regions with unclear or minimal functional significance | [Pennisi, E., Science](https://www.science.org/content/article/odd-fish-has-30-times-much-dna-humans-new-record-animals) |", "# Crystal Structure of a Putative NTP Pyrophosphohydrolase (Exig_1061) from Exiguobacterium sp. 255-15 at 1.78 \u00c5 Resolution\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [3NL9](https://www.rcsb.org/structure/3NL9)\n* **Release Date:** 2010-07-21\n* **Method:** X-RAY DIFFRACTION 1.78 \u00c5\n* **Organisms:** Exiguobacterium artemiae 255-15\n* **Macromolecules:** putative NTP pyrophosphohydrolase (protein)\n* **Unique Ligands:** EDO\n* **Publication:** [Deller, M.C., et al. (2010). Crystal Structure of a Putative NTP Pyrophosphohydrolase (Exig_1061) from Exiguobacterium sp. 255-15. Acta Crystallographica Section F, 66, 1237-1244.](https://journals.iucr.org/f/issues/2010/10/00/wd5134/index.html)\n\n**Description:** This structural genomics contribution provides the 1.78 \u00c5 crystal structure of a putative NTP pyrophosphohydrolase from the psychrotolerant bacterium Exiguobacterium sp. 255-15. The structure establishes the molecular framework for this enzyme family and contributes to the functional annotation of extremophile genomes.", "# Vibe Coding MicroPlastic Blaster\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n| Layer | Technology |\n| --- | --- |\n| Rendering | HTML5 Canvas 2D API (`800 \u00d7 600` logical resolution) |\n| Scaling | CSS `transform` + responsive `resizeCanvas()` \u2014 fills viewport on desktop and mobile |\n| Audio | Web Audio API (synthesised SFX) + `<Audio>` element (voice clips, fight track) |\n| Leaderboard | Firebase Realtime Database (REST PUT/GET) |\n| Style | Vanilla CSS \u2014 no frameworks, no build step |\n| Assets | External PNGs + MP3s (co-founder avatars, brand sprites, audio clips) |\n\nThe entire game is a **single self-contained HTML file** with no npm dependencies or build pipeline.\n\n\ud83d\udcc1 **Asset Requirements**\n\nThe following files must reside in the same directory as the HTML file:\n\nmicroplastic_blaster_115.html\nmarc.png: Co-founder avatar (Marc)\npaul.png: Co-founder avatar (Paul)\naquafina_new.png: Bottled water enemy sprite\nstarbucks_new.png: Coffee cup enemy sprite\nElora-Therapeutics-LogoStackedBlue.jpg: Cell badge logo\nmarc.mp3: Marc co-founder voice line\npaul.mp3: Paul co-founder voice line\nfight.mp3: Background music track\n\nMissing image assets degrade gracefully (fallback canvas fills / text labels). Missing audio assets fail silently.\n\n\ud83d\ude80 **Run Locally**\n\nNo server is required for basic gameplay. Open the HTML file directly in any modern browser:\n\n\ud83d\udcf1 **Mobile Support**\n\nThe canvas scales to fill the viewport on screens \u2264 700 px wide using `Math.max(w/GAMEWIDTH, h/GAMEHEIGHT)` to avoid letterboxing. Touch events (`touchstart`) are mapped to enzyme fire. The HUD collapses to a single-column layout below 700 px and hides column headers below 480 px. All breakpoints are handled in CSS only.\n\n\ud83d\udd2c **The Science \u2014 Rotating Insight Messages**\n\nBetween enzyme blasts, the in-game Microplastic Insight overlay rotates scientifically grounded messages every 8 seconds, covering topics including:\n\n\u2013 Microplastic accumulation in blood, organs, and adipose tissue\n\u2013 PET-linked inflammation and endocrine disruption\n\u2013 Chronic oxidative stress from particle-driven immune activation\n\u2013 The absence of endogenous polymer-degrading enzymes in mammals\n\u2013 The estimated 5 g/week human microplastic ingestion burden", "# Late-stage Crystal Failures and the Protein Crystal Harvesting Bottleneck\nAuthors: Deller, M. C.; Walsh, J.; Viola, R.; Elsliger, M. A.; Wilson, I. A.; Rupp, B.\nSource: American Crystallographic Association (ACA) Annual Meeting 2014 (poster), 2014\n\n**Authors:** Marc C. Deller (a), Jace Walsh (b), Robert Viola (b), Marc A. Elsliger (a), Ian A. Wilson (a), Bernhard Rupp (c). **Affiliations (condensed):** (a) Joint Center for Structural Genomics, The Scripps Research Institute, La Jolla, CA 92037, USA; (b) Square One Systems Design, Jackson, WY 83002, USA; (c) k.-k. Hofkristallamt, Vista, CA 92084. E-mail: mdeller@scripps.edu or hofkristallamt@gmail.com; http://www.ruppweb.org/robot/. **Presented:** American Crystallographic Association Annual Meeting, 2014 (poster). Work sponsored by NIH STTR Phase II Grant No. 2 R42 GM073278-02A1 and JCSG funded by Grant U54 GM094586.\n\n## Abstract\n\nHigh-throughput protein crystallography and structural genomics (HTPX/SG) have revolutionized structural biology. Approximately 10% of all protein structures deposited in the Protein Data Bank (PDB) are a result of HTPX/SG efforts [1]. To enable the significant increase in output, most HTPX/SG centers have embraced automation and robotics at all stages of the structure determination process. However, a process bottleneck remains at the crystal harvesting stage, which is still largely carried out by hand [1]. Automation of crystal harvesting offers several advantages over traditional manual methods: (a) harvesting of microcrystals too small for reliable harvesting by hand, (b) reduced late-stage failures due to non-optimal crystal handling and cryocooling and (c) increased throughput. In collaboration with Square One Systems Design, the Universal Micromanipulation Robot (UMR) is being developed and tested at the Joint Center for Structural Genomics (JCSG) to embrace the potential benefits of automation, particularly in the areas of microcrystal harvesting and improved cryocooling [1,2,3]. The UMR system has recently been modified to enable harvesting of microcrystals into SSRL sample mounting grids. High-density crystal harvesting will be essential to enable the use of high-brilliance X-ray sources from fourth-generation synchrotrons, free-electron lasers (FEL) and microfocus beamlines. Additionally, the UMR system has been modified to enable high-speed cryocooling (hyperquenching) [4,5]. Such a system offers the potential to relieve the harvesting bottleneck and reduce late-stage crystal failures.\n\n## Introduction\n\nThe Universal Micromanipulation Robot (UMR) is designed to perform the full range of tasks necessary for protein crystal harvesting. The platform is currently being optimized to enable hyperquenching for HTPX/SG at the JCSG:\n\n- Protein crystals are cooled to 77 K or 100 K to reduce radiation damage during data collection.\n- Current cooling procedures often result in hexagonal ice formation which can reduce resolution and increase mosaicity.\n- Cryoprotectant is required to prevent hexagonal ice formation within the crystal. This is both time-consuming, potentially damaging to the crystal and creates a process bottleneck. Additionally, late-stage crystal losses due to failed harvesting and poor cryoprotection techniques can occur (see Panel 1).\n- Automation of crystal harvesting processes streamlines pipeline and increases throughput, reduces late-stage crystal failures and embraces new technologies such as FEL sources that require high-density crystal shipping (see Panel 2).\n- UMR couples drip cryo cryoprotection and hyperquenching in order to limit late-stage crystal loss. Furthermore, analysis of the quench speeds and temperature profiles will provide a framework for optimal cryoprotection and cooling methods (see Panel 3).\n- Current methodologies involve cooling protein crystals relatively slowly (~1000 K/s) which allows time for damaging hexagonal ice formation. High rates of cooling (10,000\u201320,000 K/s) are predicted to be sufficient for the hyperquenching of protein crystals (see Panel 4).\n\n## Methods\n\n### Robot-assisted harvesting into SSRL grids\n\nThe UMR system was modified to harvest microcrystals directly into SSRL sample mounting grids, following a robot-assisted harvesting workflow. The UMR represents an opportunity for high-density shipping of crystals to FEL sources.\n\n### Combination drip cryoprotection and hyperquenching\n\nAn automated drip cryo mechanism applies cryoprotectant to harvested crystals and provides insulation from the film boiling regime of liquid nitrogen (PFO oil). A directly coupled pneumatic hyperquencher then cools the crystals at high speed (5 m/s) in the nucleate boiling regime of liquid nitrogen. A high-speed camera (1000 frames/s) allows accurate calculation of hyperquencher speeds, and a high-rate data acquisition thermocouple allows precise monitoring of cooling rates.\n\n## Results\n\n### Manual crystal harvesting presents a process bottleneck\n\nCrystal harvesting remains one of the few manual processes in a typical HTPX/SG pipeline, creating a process bottleneck and a potential for late-stage crystal failures. Structural genomics contributes approximately 10% of structures in the PDB, and at the JCSG roughly 100 crystals are consumed per structure.\n\n### Hyperquenching and increased rates of cooling\n\nThe slow cooling rate of manual cryocooling in liquid nitrogen provides an opportunity for damaging hexagonal ice formation. Robot-assisted cooling in liquid nitrogen gives an intermediate cooling rate. High-speed pneumatic cryocooling via the hyperquencher increases the probability of amorphous ice formation (vitreous/glassy/unstructured ice), which is less damaging to the protein crystal. The overall temperature profile of the hyperquencher was compared with a traditional manual quench and a robot-assisted quench.\n\n## Panel captions\n\n*Panel 1.* Manual crystal harvesting presents a process bottleneck. (a) Crystal harvesting remains one of the few manual processes in a typical HTPX/SG pipeline. This creates a process bottleneck and a potential for late-stage crystal failures. (b) A typical HTPX/SG automation pipeline. (c) ~10% contribution of SG structures to PDB. (d) ~100 crystals per structure at the JCSG.", "# Structure\u2013Function Analysis of the Extended Conformation of a Polyketide Synthase Module\nAuthors: Li, X.; Sevillano, N.; La Greca, F.; Deis, L.; Liu, Y.-C.; Deller, M.C.; Mathews, I.I.; Matsui, T.; Cane, D.E.; Craik, C.S.; Khosla, C.\nSource: Journal of the American Chemical Society, 2018\n\nMonoclonal antibodies are powerful reagents for stabilizing specific conformations of individual proteins.[7\u201310] In particular, antigen-binding fragments (Fabs) form rigid complexes with their protein antigens. We therefore sought to identify Fabs that would recognize DEBS module 3 + thioesterase (TE) (Figure 1) in a conformation consistent with retention of the enzymatic activity of this PKS.\n\nA large phage-displayed library of Fabs (3.7 \u00d7 10^10) from the naive human antibody repertoire[11] was used to identify Fabs against DEBS module 3 + TE.[12] Briefly, module 3 + TE was partially biotinylated and immobilized onto streptavidin magnetic beads. These beads were incubated with the phage-displayed Fab library. After repeated washing to remove nonspecifically bound phage, the tightly bound phage particles were eluted and reinfected in Escherichia coli for further propagation. This panning process was repeated for several rounds and yielded multiple high-affinity clones against the target protein. One Fab, designated 1B2 (Figure S1A), bound module 3 + TE with high affinity (KD = 110 \u00b1 10 nM, as measured by enzyme-linked immunosorbent assay (ELISA); Figure 4A). The stability of the antigen\u2013Fab complex was verified by size-exclusion chromatography (SEC) (Figure 4B,C). Binding analysis with individual domains and multi-domains of module 3 + TE revealed that 1B2 recognized the KS-AT fragment of this PKS module (KD = 145 \u00b1 25 nM, as measured by ELISA; Figure S1B).\n\n*Figure 4.* Binding of 1B2 to the KS-AT fragment of DEBS module 3. (A) ELISA of binding of 1B2 with module 3 + TE. (B) SEC traces of 1B2 (blue), the KS-AT fragment (red), and the 1B2/KS-AT complex (green). (C) SDS-PAGE: (left lane) the major SEC peak is a stoichiometric mixture of 1B2 and KS-AT; (right lane) 1B2 alone.\n\nTaking advantage of the high affinity of 1B2 for the KS-AT fragment of module 3 + TE, we grew crystals of this complex. Crystals belonged to the monoclinic space group C2 and diffracted to 2.1 \u00c5 resolution (Table S1). The asymmetric unit contained one molecule each of KS-AT and 1B2 and 680 waters; the homodimeric structure was obtained through symmetry expansion. The structure was refined to Rwork and Rfree values of 0.167 and 0.210, respectively.\n\nAs can be seen in Figure 5, the structure of the complex established that the KS-AT portion assumed an extended dimeric architecture analogous to that shown in Figure 2A for the uncomplexed KS-AT didomain. The complementarity-determining regions (CDRs) of two 1B2 equivalents face each other on opposite sides of the N-terminal coiled-coil region (Figure S2A,B). Each Fab also contacts both the KS domain (Figure S2C) and the AT domain (Figure S2D). In particular, hydrogen bonds were observed between Arg98 of the light chain and the conserved Asp331 of the KS domain and between Asn211 of the heavy chain and the carbonyl oxygen of Pro775 of the AT domain. The extensive overall footprint of 1B2 across the KS-AT fragment is readily visualized in the space-filling model shown in Figure S2E.\n\n*Figure 5.* Crystal structure of 1B2 in complex with the KS-AT fragment of DEBS module 3 + TE. Two views of the complex (PDB entry 6C9U) are shown. The coloring scheme is identical to that in Figures 1\u20133 (docking domain, blue; KS, green; KS-AT linker, cyan; AT, pink). This homodimer binds to two copies of the Fab, shown in red. The axis and plane of symmetry are shown in the lower view.\n\nWith the exception of the N-terminal coiled coil, which is absent in the previously reported structure of the KS-AT fragment of DEBS module 3,[13] no significant conformational changes were observed when the 1B2/KS-AT structure was compared to that of its antibody-free counterpart (Figure S3A). Excellent backbone overlap was also observed between the 1B2/KS-AT structure and the homologous KS-AT fragment of DEBS module 5, which also included an N-terminal coiled coil[14] (Figure S3B). Notably, the extended KS-AT conformation in the 1B2/KS-AT complex was very different from the reported cryo-EM-deduced arched conformation attributed to the homologous KS-AT portion of module 5 of the pikromycin synthase (Figure S4).\n\nThe structure of 1B2 in complex with the KS-AT fragment of module 3 + TE led us to speculate that the same Fab would also recognize the module 2 + TE construct shown schematically in Figure 1A, as this hybrid PKS module includes the N-terminal docking domain from module 3 and the hydrogen-bonding functional groups highlighted above are conserved between the two DEBS modules. Indeed, native module 2 + TE bound just as tightly to 1B2 as did the KS-AT didomain of module 3 (KD = 100 \u00b1 10 nM, as measured by ELISA; Figure S5). We therefore utilized previously developed UV-spectrophotometric kinetic assays with module 2 + TE[15] to analyze the influence of the Fab on the individual rates of both intermodular chain translocation and intramodular chain elongation catalyzed by this module. Our expectation was that if access to the arch-shaped KS-AT conformation (Figure 2B) were essential for either reaction, then 1B2 would prove inhibitory in the corresponding assay.", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\nDipeptidyl-peptidase VI from Bacillus sphaericus and YkfC from Bacillus subtilis have both previously been characterized as highly specific \u03b3-D-glutamyl-L-diamino acid endopeptidases. The crystal structure of a YkfC ortholog from Bacillus cereus (BcYkfC) at 1.8\u2005\u00c5 resolution revealed that it contains two N-terminal bacterial SH3 (SH3b) domains in addition to the C-terminal catalytic NlpC/P60 domain that is ubiquitous in the very large family of cell-wall-related cysteine peptidases. A bound reaction product (L-Ala-\u03b3-D-Glu) enabled the identification of conserved sequence and structural signatures for recognition of L-Ala and \u03b3-D-Glu and, therefore, provides a clear framework for understanding the substrate specificity observed in dipeptidyl-peptidase VI, YkfC and other NlpC/P60 domains in general. The first SH3b domain plays an important role in defining substrate specificity by contributing to the formation of the active site, such that only murein peptides with a free N-terminal alanine are allowed. A conserved tyrosine in the SH3b domain of the YkfC subfamily is correlated with the presence of a conserved acidic residue in the NlpC/P60 domain and both residues interact with the free amine group of the alanine. This structural feature allows the definition of a subfamily of NlpC/P60 enzymes with the same N-terminal substrate requirements, including a previously characterized cyanobacterial L-alanine-\u03b3-D-glutamate endopeptidase that contains the two key components (an NlpC/P60 domain attached to an SH3b domain) for assembly of a YkfC-like active site.\n\n## Structures of the first representatives of Pfam family PF06684 (DUF1185) reveal a novel variant of the Bacillus chorismate mutase fold and suggest a role in amino-acid metabolism (2010)\n\n**Authors:** Bakolitsa C, Kumar A, Jin KK, McMullan D, Krishna SS, Miller MD, Abdubek P, Acosta C, Astakhova T, Axelrod HL, Burra P, Carlton D, Chen C, Chiu HJ, Clayton T, Das D, Deller MC, Duan L, Elias Y, Ellrott K, Ernst D, Farr CL, Feuerhelm J, Grant JC, Grzechnik A, Grzechnik SK, Han GW, Jaroszewski L, Johnson HA, Klock HE, Knuth MW, Kozbial P, Marciano D, Morse AT, Murphy KD, Nigoghossian E, Nopakun A, Okach L, Paulsen J, Puckett C, Reyes R, Rife CL, Sefcovic N, Tien HJ, Trame CB, Trout CV, van den Bedem H, Weekes D, White A, Xu Q, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309109050647. PMID: 20944209.\n\nThe crystal structures of BB2672 and SPO0826 were determined to resolutions of 1.7 and 2.1\u2005\u00c5 by single-wavelength anomalous dispersion and multiple-wavelength anomalous dispersion, respectively, using the semi-automated high-throughput pipeline of the Joint Center for Structural Genomics (JCSG) as part of the NIGMS Protein Structure Initiative (PSI). These proteins are the first structural representatives of the PF06684 (DUF1185) Pfam family. Structural analysis revealed that both structures adopt a variant of the Bacillus chorismate mutase fold (BCM). The biological unit of both proteins is a hexamer and analysis of homologs indicates that the oligomer interface residues are highly conserved. The conformation of the critical regions for oligomerization appears to be dependent on pH or salt concentration, suggesting that this protein might be subject to environmental regulation. Structural similarities to BCM and genome-context analysis suggest a function in amino-acid synthesis.\n\n## Structures of the first representatives of Pfam family PF06938 (DUF1285) reveal a new fold with repeated structural motifs and possible involvement in signal transduction (2010)\n\n**Authors:** Han GW, Bakolitsa C, Miller MD, Kumar A, Carlton D, Najmanovich RJ, Abdubek P, Astakhova T, Axelrod HL, Chen C, Chiu HJ, Clayton T, Das D, Deller MC, Duan L, Ernst D, Feuerhelm J, Grant JC, Grzechnik A, Jaroszewski L, Jin KK, Johnson HA, Klock HE, Knuth MW, Kozbial P, Krishna SS, Marciano D, McMullan D, Morse AT, Nigoghossian E, Okach L, Reyes R, Rife CL, Sefcovic N, Tien HJ, Trame CB, van den Bedem H, Weekes D, Xu Q, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309109050416. PMID: 20944214.\n\nThe crystal structures of SPO0140 and Sbal_2486 were determined using the semiautomated high-throughput pipeline of the Joint Center for Structural Genomics (JCSG) as part of the NIGMS Protein Structure Initiative (PSI). The structures revealed a conserved core with domain duplication and a superficial similarity of the C-terminal domain to pleckstrin homology-like folds. The conservation of the domain interface indicates a potential binding site that is likely to involve a nucleotide-based ligand, with genome-context and gene-fusion analyses additionally supporting a role for this family in signal transduction, possibly during oxidative stress.\n\n## Structures of three members of Pfam PF02663 (FmdE) implicated in microbial methanogenesis reveal a conserved \u03b1+\u03b2 core domain and an auxiliary C-terminal treble-clef zinc finger (2010)\n\n**Authors:** Axelrod HL, Das D, Abdubek P, Astakhova T, Bakolitsa C, Carlton D, Chen C, Chiu HJ, Clayton T, Deller MC, Duan L, Ellrott K, Farr CL, Feuerhelm J, Grant JC, Grzechnik A, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Krishna SS, Kumar A, Lam WW, Marciano D, McMullan D, Miller MD, Morse AT, Nigoghossian E, Nopakun A, Okach L, Puckett C, Reyes R, Sefcovic N, Tien HJ, Trame CB, van den Bedem H, Weekes D, Wooten T, Xu Q, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309110020166. PMID: 20944230.", "# Regrow Your Teeth: The End of Dental Implants\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-18 on marcdeller.com (categories: blog).*\n\n* [Development of a new antibody drug to treat congenital tooth agenesis](https://pubmed.ncbi.nlm.nih.gov/39389160/)\n* [USAG-1 Protein Tooth Regeneration](https://www.rexresearch.com/TakahashiToothRegeneration/TakahashiToothRegeneration.html)\n\n* **Revolutionary Mechanism**: [Anti-USAG-1 antibody therapy](https://pubmed.ncbi.nlm.nih.gov/33579703/) reactivates dormant tooth buds by blocking the protein that normally prevents third dentition development.\n* **Clinical Trials Underway**: [Phase I human trials began October 2024](https://toregem.co.jp/en/archives/8371) at Kyoto University Hospital, testing safety in adults missing at least one tooth.\u200b\n* **Animal Model Success**: Preclinical studies demonstrated complete tooth regeneration in [mice and ferrets](https://pmc.ncbi.nlm.nih.gov/articles/PMC7880588/) with no major adverse effects using single antibody doses.\n* **Target Population**: Initial focus on [congenital tooth agenesis](https://pmc.ncbi.nlm.nih.gov/articles/PMC10290303/) patients, with future expansion to anyone experiencing tooth loss from decay, injury, or disease.\n* **Advantages Over Implants**: Natural regenerated teeth would provide [genuine sensation and biological integration](https://www.docseducation.com/blog/future-dental-implants-biomimicry-bioregeneration-and-path-ahead) unlike artificial implants, potentially lasting decades without replacement.\u200b\n* **Timeline to Market**: [Toregem Biopharma](https://toregem.co.jp/en/) projects commercial availability by 2030, pending successful completion of clinical trial phases.\u200b\n\n1. **Development of a new antibody drug to treat congenital tooth agenesis**: [Takahashi, K., et al., Science Advances](https://pubmed.ncbi.nlm.nih.gov/39389160/)\n2. **Anti\u2013USAG-1 therapy for tooth regeneration through enhanced BMP signaling**: [Takahashi, K., et al., Science Advances](https://pubmed.ncbi.nlm.nih.gov/33579703/)\n3. **Advances in tooth agenesis and tooth regeneration**: [Takahashi, K., et al., Regenerative Therapy](https://www.sciencedirect.com/science/article/pii/S2352320423000044)\n4. **The non-canonical BMP and Wnt/\u03b2-catenin signaling pathways orchestrate early tooth development**: [NIH National Library of Medicine](https://pmc.ncbi.nlm.nih.gov/articles/PMC4299140/)\n5. **Wnt and BMP signaling crosstalk in regulating dental stem cells**: [NIH National Library of Medicine](https://pmc.ncbi.nlm.nih.gov/articles/PMC5421560/)\n6. **Expression patterns of bone morphogenetic proteins in the developing mouse tooth**: [PubMed Central](https://pubmed.ncbi.nlm.nih.gov/9415424/)\n\n**A revolutionary Japanese antibody therapy brings us closer to regrowing natural teeth\u2014no implants required**\n\nFor decades, losing a tooth has meant choosing between dentures, bridges, or dental implants\u2014all artificial replacements that come with limitations, discomfort, and significant costs. But what if your body could\u00a0grow a new tooth instead? This isn\u2019t science fiction anymore. [Led by Dr. Katsu Takahashi at Kitano Hospital](https://pubmed.ncbi.nlm.nih.gov/39389160/) (Osaka) in collaboration with [Kyoto University](https://www.kyoto-u.ac.jp/en/research-news/2021-03-31), this groundbreaking research targets the USAG-1 protein that normally prevents additional tooth development. By blocking this \u201cbiological stop sign,\u201d dormant tooth buds can be reactivated to grow fully functional natural teeth. [Human clinical trials began in October 2024](https://toregem.co.jp/en/archives/8371), marking a pivotal moment in regenerative dentistry that could transform how we approach tooth loss by 2030.\n\nThe breakthrough centers on a deceptively simple insight: humans already possess the biological machinery for a third set of teeth. While most mammals grow only two sets\u2014baby teeth and permanent teeth\u2014humans retain [dormant tooth buds](https://pmc.ncbi.nlm.nih.gov/articles/PMC7880588/) beneath the gums throughout life, vestiges of an evolutionary capacity we\u2019ve largely lost. Approximately [1% of people with hyperdontia](https://en.wikipedia.org/wiki/Hyperdontia)\u00a0naturally develop supernumerary teeth, demonstrating that these latent structures remain biologically viable. The USAG-1 protein acts as a molecular brake on tooth development by [inhibiting both BMP and Wnt signaling pathways](https://www.science.org/doi/10.1126/sciadv.abf1798), the critical cellular communication networks that orchestrate tooth formation during embryonic development. Without this suppression, tooth buds would continue developing into additional teeth. Dr. Takahashi\u2019s team engineered a [monoclonal antibody](https://frontlinegenomics.com/new-antibody-drug-regenerates-lost-teeth/) that\u00a0neutralizes explicitly USAG-1, essentially removing the brake and allowing dormant buds to resume development.\n\nPreclinical studies demonstrated remarkable success across multiple species. In mice with genetic tooth agenesis\u2014a condition causing [congenitally missing teeth](https://pmc.ncbi.nlm.nih.gov/articles/PMC10290303/)\u2014a single antibody administration triggered the growth of complete, functional teeth. Critically, the therapy worked in ferrets, diphyodont animals with [dental patterns similar to humans](https://pmc.ncbi.nlm.nih.gov/articles/PMC7880588/), where third incisors successfully developed. These animal models showed no significant adverse effects, with regenerated teeth exhibiting normal structure, mineralization, and integration with surrounding tissues. The antibody\u2019s mechanism focuses specifically on [enhancing BMP signaling](https://www.science.org/doi/10.1126/sciadv.abf1798) without excessively disrupting Wnt pathways\u2014a crucial refinement that prevented the poor survival rates observed in earlier experiments.", "# Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\nAuthors: Hummel, J.R.; Xiao, K.-J.; Yang, J.C.; Epling, L.B.; Mukai, K.; Ye, Q.; Xu, M.; Qian, D.; Huo, L.; Weber, M.; Roman, V.; Lo, Y.; Drake, K.; Stump, K.; Covington, M.; Kapilashrami, K.; Zhang, G.; Ye, M.; Diamond, S.; Yeleswaram, S.; Macarron, R.; Deller, M.C.; Wee, S.; Kim, S.; Wang, X.; Wu, L.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2024\n\n**Step 2. N-(1-(Methylsulfonyl)piperidin-4-yl)-4-(1H-pyrazol-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (28).** To a mixture of 27 (1.30 g, 2.81 mmol) in THF (6 mL) was added HCl (4 M in 1,4-dioxane, 3.0 mL, 12 mmol) and the reaction mixture was stirred at r.t. for 2 h. The mixture was concentrated in vacuo to afford the title compound (0.93 g, 85% yield) as an off-white solid. LCMS calculated for C14H18F3N6O2S (M + H)+: m/z = 391.1; found, 391.1.\n\n**3-Chloro-4-(4-(2-(1-(methylsulfonyl)piperidin-4-ylamino)-5-(trifluoromethyl)pyrimidin-4-yl)-1H-pyrazol-1-yl)benzonitrile (10).** A mixture of N-(1-(methylsulfonyl)piperidin-4-yl)-4-(1H-pyrazol-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (28, 600 mg, 1.54 mmol), 3-chloro-4-fluorobenzonitrile (598 mg, 3.84 mmol), and K3PO4 (1.31 g, 6.15 mmol) in DMSO (4.0 mL) was stirred at 140 \u00b0C for 30 min. After cooling to r.t., the reaction mixture was diluted with H2O and extracted with EtOAc. The combined organic phases were dried over MgSO4 and concentrated. The residue was purified by flash chromatography (SiO2, 0\u221250% EtOAc/CH2Cl2). The product was further purified by prep HPLC (pH 2) to afford the title compound (190 mg, 24% yield) as a white solid. LCMS calculated for C21H20ClF3N7O2S (M + H)+: m/z = 526.1; found, 526.0.\n\n**Step 1. 3-Chloro-4-(4-(2-((1-(methylsulfonyl)piperidin-4-yl)amino)-5-(trifluoromethyl)pyrimidin-4-yl)-1H-pyrazol-1-yl)benzaldehyde (29).** A mixture of 28 (100 mg, 0.256 mmol), 3-chloro-4-fluorobenzaldehyde (122 mg, 0.768 mmol), and K3PO4 (217 mg, 1.02 mmol) in DMSO (500 \u03bcL) was stirred at 140 \u00b0C for 30 min. After cooling to r.t., the reaction mixture was diluted with aqueous 1 M HCl and extracted with EtOAc. The combined organic phases were dried over MgSO4 and concentrated. The crude material was used without further purification. LCMS calculated for C21H21ClF3N6O3S (M + H)+: m/z = 529.1; found, 529.1.\n\n**Step 2. 4-(1-(2-Chloro-4-((methylamino)methyl)phenyl)-1H-pyrazol-4-yl)-N-(1-(methylsulfonyl)piperidin-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (11).** To a mixture of aldehyde 29 in toluene (0.5 mL) and AcOH (0.5 mL) was added methylamine (2 M in THF, 0.9 mL, 1.8 mmol) and the reaction mixture was stirred at r.t. for 1 h. The mixture was concentrated in vacuo. The residue was taken up in 1,2-DCE (0.5 mL) before NaBH(OAc)3 (543 mg, 2.56 mmol) was added portionwise followed by a drop of MeOH and the reaction mixture was stirred at r.t. overnight. The product was purified by prep HPLC (pH 2) to afford the title compound (45 mg, 32% yield over two steps) as a white solid. LCMS calculated for C22H26ClF3N7O2S (M + H)+: m/z = 544.2; found, 544.0.\n\n### Synthesis of Intermediates 31 and 33 and Compound 12\n\n**tert-Butyl 4-((4-chloro-5-cyanopyrimidin-2-yl)amino)piperidine-1-carboxylate (31).** The title compound was prepared according to the procedure described for 23 with 2,4-dichloropyrimidine-5-carbonitrile (30) replacing 2,4-dichloro-5-(trifluoromethyl)pyrimidine. Off-white solid; 51% yield (8.54 g, 25.3 mmol). LCMS calculated for C11H13ClN5O2 (M \u2212 C4H8 + 1)+: m/z = 282.1; found, 282.0.\n\n**3-Chloro-4-(4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazol-1-yl)benzaldehyde (33).** To a mixture of 4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazole (32, 1.00 g, 5.15 mmol) in DMF (10 mL) was added 3-chloro-4-fluorobenzaldehyde (1.63 g, 10.3 mmol) and Cs2CO3 (3.36 g, 10.3 mmol) and the reaction mixture was stirred at 120 \u00b0C for 1 h. After cooling to r.t., the mixture was diluted with H2O and extracted with CH2Cl2. The combined organic phases were washed with brine, dried over MgSO4, and concentrated. The residue was purified by flash chromatography (40 g SiO2, EtOAc/hexanes) to afford the title compound (0.92 g, 54% yield) as a colorless waxy solid. LCMS calculated for C16H19BClN2O3 (M + H)+: m/z = 333.1; found, 333.0.\n\n**Step 1. 4-(1-(2-Chloro-4-formylphenyl)-1H-pyrazol-4-yl)-2-(piperidin-4-ylamino)pyrimidine-5-carbonitrile (34).** A mixture of tert-butyl 4-((4-chloro-5-cyanopyrimidin-2-yl)amino)piperidine-1-carboxylate (31, 1.351 g, 4.00 mmol), 3-chloro-4-(4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazol-1-yl)benzaldehyde (33, 2.00 g, 6.01 mmol), Pd(dppf)Cl2\u00b7CH2Cl2 (0.327 g, 0.400 mmol), and Na2CO3 (1.06 g, 10.0 mmol) in CH3CN/H2O (5:1, 20 mL) was stirred at 80 \u00b0C for 2 h. After cooling to r.t., HCl (4 M in 1,4-dioxane, 20.0 mL, 80.0 mmol) was added and the reaction mixture was stirred at r.t. for 2 h. MeOH (10 mL) was added and the mixture was stirred at r.t. for 30 min. The reaction mixture was diluted with H2O and extracted with CH2Cl2. The organic phase was removed, and the aqueous phase was made basic with NaOH and extracted with CH2Cl2 and EtOAc. The combined organic phases were dried over MgSO4 and concentrated. The crude material was used without further purification. LCMS calculated for C20H19ClN7O (M + H)+: m/z = 408.1; found, 408.1.\n\n**Step 2. 4-(1-(2-Chloro-4-((methylamino)methyl)phenyl)-1H-pyrazol-4-yl)-2-((1-(methylsulfonyl)piperidin-4-yl)amino)pyrimidine-5-carbonitrile (12).** To a mixture of 34 in THF (20 mL) was added methanesulfonyl chloride (0.458 g, 4.00 mmol) followed by the dropwise addition of NEt3 (0.557 mL, 4.00 mmol), and the reaction mixture was stirred at r.t. for 1 h. Methylamine (33 wt % in EtOH, 5 mL, 40 mmol) was added and the reaction mixture was stirred at r.t. for 1 h before heating to 70 \u00b0C for 30 min. The reaction mixture was concentrated in vacuo. The crude residue was taken up in MeOH (20 mL) and THF (20 mL) before NaBH4 (0.303 g, 8.00 mmol) was added portionwise and the reaction mixture was stirred at r.t. for 30 min. The mixture was diluted with CH3CN, H2O, and TFA (0.3 mL) and purified by prep HPLC (pH 10). The product was further purified by prep HPLC (pH 2) to afford the title compound (286 mg, 14% yield over two steps) as a white solid. LCMS calculated for C22H26ClN8O2S (M + H)+: m/z = 501.2; found, 501.1.\n\n### Synthesis of Compounds 13\u201316", "# Unlocking Membrane Protein Structures: The Crystallization Revolution Reshaping Drug Discovery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-22 on marcdeller.com (categories: blog, sb).*\n\n* [GPCRdb](https://gpcrdb.org)\n* [G protein-coupled receptors (GPCRs): advances in structures, mechanisms and drug discovery](https://www.nature.com/articles/s41392-024-01803-6)\n\n* **Exponential PDB growth:** [Membrane protein structures](https://academic.oup.com/nar/article/47/D1/D390/5173663) increased from 3,500 in 2019 to projected 10,500 by 2029, with GPCR structures reaching 1,716 by 2025\u200b\n* **LCP dominance for GPCRs:** [Lipidic cubic phase crystallization](https://formulatrix.com/life-science-automation-blog/lipidic-cubic-phase-lcp-introduction/) achieves 2.5 \u00c5 mean resolution, accounting for 49 GPCR structures by producing native membrane-like environments\u200b\n* **Cryo-EM revolution:** [Single-particle cryo-EM](https://portlandpress.com/biochemsoctrans/article/49/5/2345/229854/Structure-determination-of-GPCRs-cryo-EM-compared) now determines 78% of GPCR structures, with resolution improving from 4.2 \u00c5 to 3.6 \u00c5 between 2015-2019\u200b\n* **Fusion protein success:** [T4 lysozyme and BRIL insertions](https://pmc.ncbi.nlm.nih.gov/articles/PMC4408211/) replacing flexible intracellular loop 3 enabled majority of recent GPCR structures through increased stability and crystallization contacts\n* **MicroED integration:** [Microcrystal electron diffraction](https://portlandpress.com/biochemsoctrans/article-abstract/50/1/231/230823) combined with FIB milling enables structure determination from submicron LCP and bicelle crystals previously unsuitable for analysis\u200b\n\n1. **Design and implementation of suspended drop crystallization**: [Gillman, C., et al., bioRxiv Preprint, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC10081258/)\n2. **Recent Advances in Structure-Based Drug Design Targeting Class A GPCRs**: [Lee, Y., et al., Journal of Medicinal Chemistry, 2018](https://pubs.acs.org/doi/10.1021/acs.jmedchem.6b01453)\n3. **MemProtMD database: a resource for membrane-embedded protein structures**: [Newport, T.D., et al., Nucleic Acids Research, 2019](https://academic.oup.com/nar/article/47/D1/D390/5173663)\n4. **Membrane Protein Crystallisation: Current Trends and Future Perspectives**: [Salom, D., et al., Methods in Cell Biology, 2013](https://pmc.ncbi.nlm.nih.gov/articles/PMC4605854/)\n5. **A 10-Year Meta-Analysis Of Membrane Protein Structural Biology**: [Choy, B.C., et al., Biochimica et Biophysica Acta (BBA) \u2013 Biomembranes, 2020](https://pmc.ncbi.nlm.nih.gov/articles/PMC7856071/)\n\n***How advanced techniques and exponential growth in GPCR structures are transforming our understanding of pharmaceutical targets***\n\nThe field of membrane protein crystallography has experienced a remarkable transformation over the past decade, with the Protein Data Bank witnessing an explosion of new structures that would have seemed impossible just years ago. In 2019, approximately 3,500 membrane protein structures representing over 1,000 unique proteins were deposited in the PDB, and projections suggest this number will reach 10,500 structures by 2029. This exponential growth, particularly dramatic for G protein-coupled receptors, reflects fundamental advances in crystallization methodologies, protein engineering strategies, and structural determination techniques that have turned once-intractable targets into routine pharmaceutical research subjects.\n\nG protein-coupled receptors exemplify this revolution most dramatically. From just 17 GPCR structures in 2012, the field has exploded to [1,716 GPCR structures](https://gpcrdb.org/) by 2025, with an unprecedented 78% of structures deposited in the first half of 2021 determined by cryo-electron microscopy rather than traditional X-ray crystallography. This shift represents not merely incremental progress but a fundamental reimagining of how structural biologists approach these critical drug targets, which account for approximately 50% of current pharmaceuticals.\n\n[Lipidic cubic phase crystallization](https://pubs.acs.org/doi/10.1021/acs.cgd.7b00519) has emerged as perhaps the most transformative technique for membrane protein structure determination. This method, which creates a bicontinuous bilayer environment that mimics the native cellular membrane, has proven especially powerful for GPCRs, accounting for 49 of the 70 GPCR structures reported by 2015. The [mean resolution achieved through LCP crystallization](https://pmc.ncbi.nlm.nih.gov/articles/PMC4605854/) reaches 2.5 \u00c5, nearly half an \u00e5ngstr\u00f6m better than traditional alkyl maltoside detergent methods. The technique works by embedding membrane proteins in a highly curved, multiply-branched lipid matrix formed from monoolein or monopalmitolein, where proteins can diffuse within the bilayer environment until crystallization conditions trigger a local phase transition from cubic to lamellar phase. This transition concentrates proteins in specific regions, promoting nucleation while the bulk cubic phase continues feeding protein molecules to growing crystal faces through lamellar portals.\n\n| Method | Approximate Count | Percentage | Key Applications |\n| --- | --- | --- | --- |\n| Detergent-Based | ~60 | ~66% | General use (channels, transporters) |\n| Lipidic Cubic Phase | 17 | 18.7% | GPCRs, high-resolution structures |\n| Bicelle | ~9 | ~10% | Bacteriorhodopsin, select GPCRs |\n| Fusion Protein | ~5 | ~5.3% | GPCRs requiring engineered hydrophilic domains |", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\n**Scheme 1.** Compounds of Formula (I) can be prepared as shown in Scheme 1. Condensation of benzoimidazolylamine 1-1 with dicarbonyl compound 1-2 at elevated temperatures (e.g., 140 \u00b0C) can afford tricyclic 1-3. Tricyclic intermediate 1-3 can be converted to halide 1-4, where X^1 is a halogen (e.g., Cl, Br, or I) under standard halogenation conditions (e.g., in the presence of NBS or NIS or NCS). The \u2014NH\u2014 moiety of imidazopyrimidinone 1-4 can be protected by a protecting group (e.g. SEM) under standard conditions (e.g. in the presence of 2-(chloromethoxy)ethyl)trimethylsilane and diisopropylethylamine) to give protected tricyclic intermediate 1-5. Halide 1-5 can be coupled with compound 1-6 [where E is a cyclic structure bearing a protected or unprotected amine that can be part of the ring or a substituent off of the ring (e.g., aniline, benzylic amine, cycloalkylamine, cycloalkenylamine pyridylamine, azetidine, pyrrolidine, piperidine, piperazine, cycloheteroalkylamine, cycloheteroalkenylamine, cycloheteroarylamine, fused bicyclic amine, or spirocyclic amine, etc.), M^1 is a boronic acid, boronate ester, potassium trifluoroborate, H or an appropriately substituted metal such as Sn(Bu)3 or Zn] under standard Suzuki conditions (e.g., in the presence of a palladium catalyst, such as XPhos Pd G4 and a base (e.g., a phosphate base)) or standard Stille conditions (e.g., in the presence of a palladium(0) catalyst, such as tetrakis(triphenylphosphine)palladium(0)) or standard Negishi conditions (e.g., in the presence of a palladium catalyst, such as tetrakis(triphenylphosphine)palladium(0) or [1,1'-bis(diphenylphosphino)ferrocene]dichloropalladium (II)) or standard Buchward-Hartwig conditions (e.g., in the presence of a palladium catalyst, such as [1,1'-bis(diphenylphosphino)ferrocene]dichloropalladium (II) and a base (e.g., sodium tert-butoxide)) to give compound 1-7. Amine 1-7 can be converted to compound 1-8 under standard amide/sulfonamide formation conditions (e.g., in the presence of a carboxylic chloride or sulfonyl chloride). Compound 1-8 can be deprotected by standard acidic conditions (e.g. in the presence of TFA) to give compound 1-9. Compound 1-9 can be alkylated under standard SN2 conditions (e.g., in the presence of diisopropylethylamine and 1-10, where X^2 is a halogen (e.g., Cl, Br, or I) or pseudohalogen (e.g., OTs or OMs)) to give compounds of Formula (I).\n\n**Scheme 2.** Compounds of Formula (I) can also be prepared as shown in Scheme 2. Suitable starting materials 2-1, which can be synthesized through Scheme 1 (e.g., compound 1-3) can be alkylated under standard SN2 conditions (e.g., in the presence of diisopropylethylamine and 2-2, where X^1 is a halogen (e.g., Cl, Br, or I) or pseudohalogen (e.g., OTs or OMs)) to give compound 2-3. Compound 2-3 can be converted to halide 2-4, where X^2 is a halogen (e.g., Cl, Br, or I) under standard halogenation conditions (e.g., in the presence of NBS or NIS or NCS). Halide 2-4 can be coupled with compound 2-5 [where A is a cyclic structure (e.g., phenyl, saturated or unsaturated cycloalkyl, saturated or unsaturated heterocycloalkyl, heteroaryl, saturated or unsaturated fused bicyclic ring, or saturated or unsaturated spirocyclic ring, etc.) bearing a reactive functional group R^W (e.g., alpha,beta-unsaturated carbonyl, alpha,beta-unsaturated sulfonyl, halide, nitrile, aldehyde, epoxide, or aziridine, etc.), M^1 is a boronic acid, boronate ester, potassium trifluoroborate, or an appropriately substituted metal such as Sn(Bu)3 or Zn] under standard Suzuki, Stille or Negishi conditions to give compounds of Formula (I).\n\n**Scheme 3.** Compounds of Formula (I) can also be prepared as shown in Scheme 3. Suitable starting materials 3-1, which can be synthesized through Scheme 2 (e.g., compound 2-4), where X^2 is a halogen (e.g., Cl, Br, or I), can be coupled with compound 3-2 [where E is a cyclic structure bearing a protected or unprotected amine that can be part of the ring or a substituent off of the ring (e.g., aniline, benzylic amine, cycloalkylamine, cycloalkenylamine pyridylamine, azetidine, pyrrolidine, piperidine, piperazine, cycloheteroalkylamine, cycloheteroalkenylamine, cycloheteroarylamine, fused bicyclic amine, or spirocyclic amine, etc.), M^1 is a boronic acid, boronate ester, potassium trifluoroborate, H or an appropriately substituted metal such as Sn(Bu)3 or Zn] under standard Suzuki, Stille, Negishi, or Buchward-Hartwig conditions to give compound 3-3. Amine 3-3 can be converted to compounds of Formula (I) under standard amide/sulfonamide formation conditions (e.g., in the presence of a carboxylic chloride or sulfonyl chloride).\n\n**Scheme 4.** Compounds of Formula (I) can be prepared as shown in Scheme 4. Condensation of imidazolylamine 4-1, where B is a cyclic structure (e.g. cycloalkyl, phenyl, heterocycloalkyl or heteroaryl), with dicarbonyl compound 4-2 at elevated temperatures (e.g., 140 \u00b0C) can afford tricyclic intermediate 4-3. Tricyclic intermediate 4-3 can be alkylated under standard SN2 conditions (e.g., in the presence of diisopropylethylamine and 4-4, where X^1 is a halogen (e.g., Cl, Br, or I) or pseudohalogen (e.g., OTs or OMs)) to give compound 4-5. Compound 4-5 can be converted to halide 4-6, where X^2 is a halogen (e.g., Cl, Br, or I) under standard halogenation conditions (e.g., in the presence of NBS or NIS or NCS). Halide 4-6 can be coupled with compound 4-7 under standard Suzuki, Stille, Negishi or Buchward-Hartwig conditions to give compound 4-8. Amine 4-8 can be converted to compounds of Formula (I) under standard amide/sulfonamide formation conditions (e.g., in the presence of a carboxylic chloride or sulfonyl chloride).", "# 2025 \u2013 Imidaz0[1,2-a]pyrimidine-5(1 h)-ones having a covalent warhead at the 3-position as wrn inhibitors\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-09-09 on marcdeller.com (categories: patents).*\n\n* [Link to Google Patent](https://patents.google.com/patent/US20250270217A1/en)\n\n* **Synthetic Lethality Mechanism**: WRN inhibitors exploit the dependency of MSI-H cancer cells on Werner helicase for survival, inducing selective DNA damage and apoptosis while sparing microsatellite-stable cells\n* **Covalent Warhead Strategy**: The 3-position covalent warhead design enables irreversible binding to WRN active site cysteines, providing enhanced potency and prolonged target engagement compared to reversible inhibitors\n* **Clinical Validation Evidence**: Phase I trials of covalent WRN inhibitors have demonstrated encouraging response rates of 14% in heavily pretreated MSI-H patients, establishing proof-of-concept for this therapeutic approach\n* **Resistance Circumvention**: Covalent binding mechanisms may overcome potential resistance mutations that could limit the efficacy of non-covalent WRN inhibitors in clinical settings\n\n1. **AACR 2025 \u2013 Roche sets a bar in Werner helicase inhibition:** [Baldanis, K., et al., Oncology Pipeline, 2025](https://www.oncologypipeline.com/apexonco/aacr-2025-roche-sets-bar-werner-helicase-inhibition)\n2. **Discovery of WRN inhibitor HRO761 with synthetic lethality in MSI cancers:** [Kategaya, L., et al., Nature, 2024](https://www.nature.com/articles/s41586-024-07350-y)\n3. **Targeting the Werner syndrome protein in microsatellite instability cancers:** [Liu, Y., et al., European Journal of Medical Oncology, 2025](https://pmc.ncbi.nlm.nih.gov/articles/PMC12328481/)\n4. **Innovative design and potential applications of covalent strategy in drug discovery:** [Chen, X., et al., European Journal of Medicinal Chemistry, 2025](https://pubmed.ncbi.nlm.nih.gov/39756145/)\n\n**Targeting the Undruggable: Covalent Imidazo[1,2-a]pyrimidine-5(1h)-ones Revolutionize WRN Inhibitor Design**\n\nThe challenge of selectively targeting Werner syndrome helicase (WRN) in microsatellite instability-high (MSI-H) cancers has driven the development of innovative covalent inhibitors that exploit synthetic lethality mechanisms. Recent advances in medicinal chemistry have identified imidazo[1,2-a]pyrimidine-5(1h)-ones equipped with strategically positioned covalent warheads at the 3-position as promising therapeutic agents. These compounds represent a sophisticated approach to drug design, combining the proven pharmacological potential of the imidazo[1,2-a]pyrimidine scaffold with the precision targeting capabilities of covalent inhibition strategies. MSI-H cancers, which comprise approximately 15% of colorectal cancers and significant fractions of gastric and endometrial malignancies, exhibit unique vulnerabilities to WRN inhibition due to their dependence on this helicase for genomic stability maintenance. The selective nature of this dependency stems from the synthetic lethality relationship between deficient mismatch repair systems and WRN function, where dual disruption results in catastrophic DNA damage specifically in cancer cells while sparing normal tissue.\n\nCurrent clinical investigations of WRN inhibitors, including HRO761 and RO7589831, have demonstrated proof-of-concept efficacy with manageable toxicity profiles, establishing the therapeutic potential of this target. The strategic placement of covalent warheads at the 3-position of the imidazo[1,2-a]pyrimidine core enables irreversible binding to critical cysteine residues within the WRN active site, potentially overcoming resistance mechanisms that limit non-covalent inhibitors and providing durable target engagement essential for therapeutic efficacy in aggressive cancer types.", "# Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3\nAuthors: Shvartsbart, A.; Roach, J.J.; Witten, M.R.; Koblish, H.; Harris, J.J.; Covington, M.; Hess, R.; Lin, L.; Frascella, M.; Truong, L.; Leffet, L.; Conlen, P.; Beshad, E.; Klabe, R.; Katiyar, K.; Kaldon, L.; Young-Sciame, R.; He, X.; Petusky, S.; Chen, K.-J.; Horsey, A.; Lei, H.-T.; Epling, L.B.; Deller, M.C.; Vechorkin, O.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2022\n\nDuring the course of our work on exploring the ethers, a concurrent effort was undertaken to investigate substitution on the aromatic positions of the benzamide ring. Careful analysis of the SAR revealed two key trends: addition of fluorine para to the N-methyl amide (R4, Table 4) abolished potency toward KIT and PDGFRB (e.g., 5 \u2192 21), while substitution of a methyl group in the meta position (R3, Table 4) significantly reduced activity for FLT3 and TRKA (e.g., 19 \u2192 22). These trends suggested that incorporating benzamides with substitution at both R3 and R4 would result in compounds with desirable selectivity over all four problematic kinases. Gratifyingly, the 4-fluoro-3-methylbenzamide compound 23 verified our hypothesis and demonstrated >100-fold kinome selectivity. Compound 23 also maintained excellent potency for FGFR3 (IC50 = 0.7 nM) and selectivity over FGFR1 (62\u00d7). Compound 24, a 3-difluoromethyl-4-fluorobenzamide, displayed a similar profile to 23 and maintained potent activity against the V555L (IC50 = 2.6 nM) and V555M (IC50 = 0.9 nM) GK mutants. The 3,4-difluorobenzamide derivative 25 had the best biochemical potency (FGFR3 IC50 = 0.6 nM) and general kinase profile of the series. Notably, 25 (C log P = 1.2) is significantly less polar than 20. This compound was very potent in the BA/F3 cellular assay (IC50 = 6.8 nM) and was found to have highly desirable in vitro ADME properties (Table 5), with low intrinsic clearance (0.5 L/h/kg), high Caco-2 permeability (Papp = 5.7 \u00d7 10\u207b\u2076 cm/s), and good solubility (FASSIF/SGF = 118/>400 \u03bcg/mL). On the basis of this profile, 25 was dosed in a rat PK cassette.\n\n*Table 4.* Phenyl Ring SAR and General Kinase Selectivity Data.\n\n| Compound | R3 | R4 | R5 | FGFR3 IC50 (nM) | FGFR1 IC50 (nM) | FGFR3 V555L IC50 (nM) | FGFR3 V555M IC50 (nM) | BA/F3 IC50 (nM) | FLT3 IC50 (nM) | TRKA IC50 (nM) | KIT IC50 (nM) | PDGFRB IC50 (nM) |\n|---|---|---|---|---|---|---|---|---|---|---|---|---|\n| 19 | H | H | Me | 0.9 | 36 | 0.5 | \u2014 | 15 | 25 | 118 | 82 | 66 |\n| 20 | H | \u2014 | Me | 2.2 | 147 | \u2014 | \u2014 | 969 | 1343 | 7304 | 384 | 370 |\n| 5 | H | H | Me | 4.6 | 86 | \u2014 | \u2014 | 33 | 12 | 58 | 33 | 38 |\n| 21 | H | F | Me | 2.6 | 60 | \u2014 | \u2014 | 14 | 8 | 11 | 180 | 348 |\n| 22 | Me | H | i-Pr | 1.3 | 39 | 5.3 | 3.2 | 5.4 | 114 | 363 | 22 | 126 |\n| 23 | Me | F | Me | 0.7 | 46 | \u2014 | \u2014 | 12 | 97 | 293 | 1657 | 1386 |\n| 24 | CHF2 | F | Me | 0.9 | 49 | 2.6 | 0.9 | 13 | 98 | 443 | 619 | 800 |\n| 25 | F | F | Me | 0.6 | 21 | \u2014 | \u2014 | 6.8 | 155 | 285 | 1449 | 1741 |\n\nCompound 20 carries an m-pyridyl amide in place of the benzamide, so the R3/R4 phenyl substituents do not apply; dashes indicate values that were not reported in the table.\n\n*Table 5.* In Vitro ADME and Pharmacokinetics of 25 in Rat.\n\n| Parameter | Value |\n|---|---|\n| Caco-2 (Papp \u00d7 10\u207b\u2076 cm/s) | 5.7 |\n| h-Cl (L/h/kg) | 0.5 |\n| Solubility (\u03bcg/mL, FASSIF/SGF) | 118/>400 |\n| CYP3A4 IC50 (\u03bcM) | 3.9 |\n| Rat i.v. dose (mg/kg) | 0.73 |\n| Rat i.v. HBF% | 3.3 |\n| Rat i.v. Vdss (L/kg) | 0.35 |\n| Rat i.v. t1/2 (h) | 5.3 |\n| Rat p.o. dose (mg/kg) | 3.1 |\n| Rat p.o. Cmax (nM) | 6173 |\n| Rat p.o. AUC (nM\u00b7h) | 33 587 |\n| Rat p.o. F% | 54 |\n\nUpon i.v. dosing in rat, 25 exhibited very low clearance (HBF = 3.3%) and a moderate half-life (t1/2 = 5.3 h). Oral dosing (p.o.) resulted in excellent exposure (AUC = 33 587 nM\u00b7h) with good oral bioavailability (F% = 54). Before examining pharmacokinetics in higher species, compound 25 was screened in a full cytochrome P450 (CYP) panel. Unfortunately, we found that 25 potently inhibited the activity of CYP3A4 (IC50 = 3.9 \u03bcM). It is also possible that the CYP inhibition contributed to the high exposure of 25.\n\nWe revisited 20, speculating that the additional polarity imparted by the m-pyridyl ring would reduce CYP inhibition. In fact, 20 had no effect on CYP3A4 (0% inhibition at 5 \u03bcM, Table 6). Encouraged by this data, several approaches to improve the cell permeability of 20 were considered. One such strategy was to modulate the polarity of the bicyclic core. This was implemented by switching the position of the bridgehead nitrogen, thereby converting the highly basic imidazo[1,2-b]pyridazine core to a significantly less polar pyrazolo[1,5-a]pyrimidine. Pleasingly, the resulting compound 26 was 2-fold more potent than 20 in the enzyme assay (FGFR3 IC50 = 1.0 nM vs 2.2 nM, respectively), with dramatically improved cellular translation (BA/F3 IC50 = 29 nM vs 969 nM) and minimal CYP inhibition.\n\n*Table 6.* Comparison of Imidazo[1,2-b]Pyridazine and Pyrazolo[1,5-a]Pyrimidine Scaffolds.\n\n| Compound | FGFR3 IC50 (nM) (R1/R3) | BA/F3 IC50 (nM) | CYP3A4 IC50 (\u03bcM) (% inh @ 5 \u03bcM) | Caco-2 (Papp \u00d7 10\u207b\u2076 cm/s) |\n|---|---|---|---|---|\n| 20 | 2.2 (67\u00d7) | 969 | >5 (0) | \u2014 |\n| 26 | 1.0 (50\u00d7) | 29 | >5 (15) | 0.2 |\n\nThe low permeability of 26 (Papp = 0.2 \u00d7 10\u207b\u2076 cm/s) was addressed by the replacement of the pyrazole N-methyl group with isopropyl to afford 27 (Papp = 1.5 \u00d7 10\u207b\u2076 cm/s, Table 7). Compound 27 was among the most potent in the enzyme assay (FGFR3 IC50 = 0.4 nM), with excellent potency in cells (BA/F3 IC50 = 7.2 nM).", "# Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\nAuthors: Hummel, J.R.; Xiao, K.-J.; Yang, J.C.; Epling, L.B.; Mukai, K.; Ye, Q.; Xu, M.; Qian, D.; Huo, L.; Weber, M.; Roman, V.; Lo, Y.; Drake, K.; Stump, K.; Covington, M.; Kapilashrami, K.; Zhang, G.; Ye, M.; Diamond, S.; Yeleswaram, S.; Macarron, R.; Deller, M.C.; Wee, S.; Kim, S.; Wang, X.; Wu, L.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2024\n\nThe crystal structure of compound 17 bound to CDK2 in the ATP binding site was determined to 1.55 \u00c5 resolution (PDB: 8UV0) and is illustrated in Figure 4. The 2-aminopyrimidine engages in a hinge binding interaction, and the trifluoromethyl moiety is projected toward the phenylalanine gatekeeper residue (Phe80). The saturated piperidine ring adopts a chair conformation to avoid diaxial interactions, thereby favorably positioning the sulfonamide oxygen atoms to act as hydrogen bond acceptors with the backbone N\u2212H of Asp86 and a water-mediated interaction with Lys89 and Gln85. The pyrazole ring is coplanar with the 2-aminopyrimidine, which positions it under the glycine-rich loop (Ile10\u2212Val18). Finally, the tertiary hydroxyl group interacts with Asp145, which is part of the catalytic triad (Lys33, Glu51, and Asp145).\n\n*Figure 4.* Crystal structure of compound 17 bound to CDK2 in the ATP pocket (PDB: 8UV0).\n\n## Conclusions\n\nIn summary, starting with HTS hit (\u00b1)-1, we have rationally designed and discovered a series of (4-pyrazolyl)-2-aminopyrimidines as potent and selective CDK2 inhibitors. Early structural modifications identified that substitution at the C5-position of the pyrimidine had a profound effect on both CDK2 activity and isoform selectivity. Further SAR studies led to the discovery of compound 17 that displayed subnanomolar activity against CDK2 and selectivity over CDK1, CDK4, CDK6, CDK7, and CDK9. Evaluation in mice with CCNE1-amplified ovarian cancer tumors showed that 17 reduced tumor levels of pRb by \u223c50% and showed tumor growth inhibition. These results provide compelling evidence for small molecule pharmacological inhibition of CDK2 as a means to treat CCNE1-amplified cancers.\n\n## Chemistry\n\nThe synthesis of compound 2 was achieved in two steps from commercially available starting inputs (Scheme 1). Suzuki cross-coupling of 2,4,5-trichloropyrimidine (19) and pyrazole boronic ester 20 provided intermediate 21, and Buchwald\u2212Hartwig amination with 1-(methylsulfonyl)piperidin-4-amine afforded compound 2.\n\n*Scheme 1.* Synthesis of Compound 2. Reagents and conditions: (a) Pd(dppf)Cl2\u00b7CH2Cl2, Na2CO3, CH3CN/H2O, 80 \u00b0C, 84%; (b) 1-(methylsulfonyl)piperidin-4-amine, RuPhos Pd G2, Cs2CO3, dioxane, 100 \u00b0C, 53%.\n\nA general synthetic route to access compounds 3\u22129 is outlined in Scheme 2. Selective SNAr of pyrimidine 22 with tert-butyl 4-aminopiperidine-1-carboxylate in the presence of ZnCl2 and Hunig's base provided intermediate 23. The removal of the tert-butyloxycarbonyl protecting group followed by mesylation gave 25. Suzuki cross-coupling with commercially available pyrazole boronic acids or pinacol esters delivered compounds 3\u22129.\n\n*Scheme 2.* Synthesis of Compounds 3\u22129. Reagents and conditions: (a) ZnCl2, tert-butyl 4-aminopiperidine-1-carboxylate, DIPEA, DCE/t-BuOH, 0\u221260 \u00b0C, 60%; (b) HCl (4 M in dioxane), THF, 60 \u00b0C, 88%; (c) MsCl, DIPEA, THF, 0 \u00b0C to rt, 85%; and (d) ArBpin or ArB(OH)2, Pd(dppf)Cl2\u00b7CH2Cl2, Na2CO3, dioxane/H2O or CH3CN/H2O, 100 \u00b0C, 11\u221258%.\n\nCompounds 10 and 11 were accessed through a modified route to allow for late-stage pyrazole functionalization (Scheme 3). Cross-coupling of previously described intermediate 25 with protected pyrazole boronic ester 26 afforded 27. Protecting group removal to access 28, followed by SNAr with aryl fluorides furnished compound 10 and intermediate 29. Reductive amination of aldehyde 29 with methylamine delivered compound 11.\n\n*Scheme 3.* Synthesis of Compounds 10 and 11. Reagents and conditions: (a) Pd(dppf)Cl2\u00b7CH2Cl2, Na2CO3, CH3CN/H2O, 90 \u00b0C, 67%; (b) HCl (4 M in dioxane), THF, rt, 85%; (c) for 10: 3-chloro-4-fluorobenzonitrile, K3PO4, DMSO, 140 \u00b0C, 24%; (d) for 11: 3-chloro-4-fluorobenzaldehyde, K3PO4, DMSO, 140 \u00b0C; and (e) (i) NH2Me (2 M in THF), PhMe/AcOH, rt, (ii) NaBH(OAc)3, DCE, rt, 32% (two steps).\n\nAlternatively, compound 12 was synthesized by the convergent route outlined in Scheme 4. Aminopyrimidine 31 was prepared in analogous fashion to 23, and SNAr of pyrazole 32 with 3-chloro-4-fluorobenzaldehyde gave 33. Suzuki cross-coupling of intermediates 31 and 33 followed by Boc deprotection furnished 34. A one-pot mesylation and reductive amination provided 12.\n\n*Scheme 4.* Synthesis of Compound 12. Reagents and conditions: (a) ZnCl2, tert-butyl 4-aminopiperidine-1-carboxylate, DIPEA, DCE/t-BuOH, 0\u221260 \u00b0C, 51%; (b) 3-chloro-4-fluorobenzaldehyde, Cs2CO3, DMF, 120 \u00b0C, 54%; (c) (i) Pd(dppf)Cl2\u00b7CH2Cl2, Na2CO3, CH3CN/H2O, 80 \u00b0C, (ii) HCl (4 M in dioxane), MeOH, rt; and (d) (i) MsCl, NEt3, rt, (ii) NH2Me, rt to 70 \u00b0C, (iii) NaBH4, MeOH/THF, rt, 14% (two steps).\n\nThe synthesis of compounds 13\u221216 was carried out in two steps from intermediates 23 and 31 to allow for late-stage installation of functionalized sulfonamides (Scheme 5). Cross-coupling with pyrazole 35 followed by Boc removal gave amines 36 and 37, and sulfonylation with commercially available sulfonyl chlorides delivered compounds 13\u221216.\n\n*Scheme 5.* Synthesis of Compounds 13\u221216. Reagents and conditions: (a) (i) Pd(dppf)Cl2\u00b7CH2Cl2, Na2CO3, CH3CN/H2O, 110 \u00b0C, (ii) HCl (4 M in dioxane), rt; (b) R3SO2Cl, NEt3, THF, rt, 11\u221248% (two steps).\n\nThe synthesis of compounds 17 and 18 was carried out in a similar fashion employing commercially available pyrazole 38 (Scheme 6). Slight modification of reaction conditions provided access of compound 17 in good yield and >99.5% purity on 10 g scale. Purification of 39 and 40 after Suzuki coupling, followed by deprotection with HCl allowed for clean isolation of amine hydrochlorides 41 and 42, and reaction with cyclopropanesulfonyl chloride furnished compounds 17 and 18.\n\n*Scheme 6.* Synthesis of Compounds 17 and 18. Reagents and conditions: (a) Pd(dppf)Cl2\u00b7CH2Cl2, Na2CO3, dioxane/H2O, 100 \u00b0C; (b) for 17: HCl (4 M in dioxane), CH2Cl2/MeOH, rt, 96% (two steps); (c) for 18: HCl (4 M in dioxane), THF, 60 \u00b0C, 90%; and (d) cyclopropanesulfonyl chloride, DIPEA, CH3CN, rt, 56\u221287%.\n\n## Experimental Section\n\n### General", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\nA paused run **keeps everything it holds** -- RAM, and the GPU allocations with it -- because\nthat is what makes it resumable in place. Right for \"I need the machine for ten minutes\", wrong\nfor \"pause this until tomorrow\"; for that, quit and re-run, since `run` skips completed\ntargets. Paused time is excluded from the ETA and recorded separately in the run history\n(`working_seconds` alongside `duration_seconds`), so a run paused over lunch does not teach\nevery later run that targets take an extra hour.\n\nQuitting tears down the **whole process tree**, not just the process BoltzMaker launched. Boltz\nruns its dataloader workers as children, and terminating only the parent left them alive holding\ntheir share of RAM and the GPU -- measured on a real two-worker tree, and now covered by a test.\n\n## \ud83d\udcca Outputs\n\nWritten next to `boltz_input.md`. Anything marked optional appears only when the feature that\nwrites it has run.\n\n| Output | What it is |\n|---|---|\n| `boltz_run_<timestamp>.log` | Raw `boltz predict` output for the run |\n| `boltz_output/` | Boltz's own prediction output tree |\n| `boltz_cif/` | Every completed target's `*_model_0.cif`, flattened into one folder |\n| `boltz_summary.csv` / `boltz_summary.xlsx` | One row per target -- see below |\n| `boltz_summary_view.csv` | The dashboard's \"Summary table\" columns: a trimmed, renamed subset of `boltz_summary.csv`, for anyone who wants the at-a-glance view in a spreadsheet rather than every raw field |\n| `boltz_ligands.csv` | ID, SMILES, stereocentre/ionizable-group/fragment findings, MW, cLogP, TPSA |\n| `boltz_dashboard.html` | The interactive report -- see below |\n| `boltz_ligand_grid.pdf` (optional) | A print/share-friendly PDF of the dashboard's \"Ligand structures\" grid: same 5x5 pagination, rendered structures, severity borders and scaffold highlighting as on screen, in the style of [smiles2grid](https://github.com/bellcheddar/smiles2grid)'s own PDF output. Only written when a campaign has at least one SMILES ligand |\n| `boltz_plip/` (optional) | Per-target cif2plip output: the converted PDB, PLIP's XML/TXT reports, the ray-traced binding-site PNG, and the PyMOL `.pse` session -- cached here so re-running `analyze` doesn't re-profile a target already done |\n| `boltz_interactions.csv` (optional) | Long format, one row per detected contact across every target: interaction type, residue, distance -- the raw data behind the fingerprint heatmap and per-target contact tables |\n| `boltz_dashboard_sessions/` (optional) | Each target's PyMOL `.pse` session, linked from the dashboard. This is the one thing that stops `boltz_dashboard.html` being a single self-contained file once interaction analysis has run; without `setup-plip` it stays exactly as self-contained as before |\n| `boltz_pose_pairs/` | Two single-ligand mmCIFs per comparison (predicted, superposed into the experimental frame; and experimental) plus an `index.json` of the distances -- what the dashboard's pair viewer draws, carried in the `.bmz` |\n| `boltz_sse_comparison.csv` / `.html` | One row per family/target/motif: Ca RMSD, centroid shift, helix-axis rotation/kink angles, SSE boundary shift, flagged phi/psi residues, and (kinases) DFG-in/out and alphaC-in/out states. Written automatically whenever any family has `Apo structure:` set. Family/Target columns, chart legends and the family-coverage table use the dashboard's display names; the CSV keeps the raw `family_id`/`target_stem` columns alongside for cross-referencing. The HTML is a standalone dashboard whose content is also embedded into `boltz_dashboard.html`. See **compare-sse** below for what the metrics mean |\n| `boltz_sse_family_status.json` | One entry per protein family: `ok` (with a target/motif count) / `no_apo_structure` / `apo_not_found` / `annotation_failed` / `no_predicted_structures` -- the machine-readable form of the \"Family coverage\" table, so a family with no `Apo structure:` reads as \"not configured\" rather than silently missing |\n| `boltz_sse_comparison_sessions/` (optional) | A plain-text PyMOL `.pml` script per target: colours/labels each motif, highlights the ones with a significant shift |\n| `reference/` (optional, yours) | Experimental mmCIF structures you drop in yourself. Their presence is what turns on the dashboard's **Ligand pose vs experiment** panel (see below); the same files already serve `Pocket contact:` and `Apo structure:` |\n\n### `boltz_summary.csv` / `.xlsx`\n\nOne row per target: every scalar field from the confidence/affinity JSONs, computed pIC50,\nthe input ligand SMILES, and `flags`/`notes` columns (`LOW_CONFIDENCE`,\n`HIGH_CONFIDENCE_POOR_AFFINITY`, `LOW_CONFIDENCE_STRONG_AFFINITY`, `LOW_POCKET_PLDDT`,\n`MISSING_OUTPUTS`). The XLSX adds a `selectivity` sheet (ligand x family pIC50 pivot) whenever\na campaign spans more than one protein family. When `setup-plip` has run there is also a\n`plip_status` column (`ok` / `no_interactions` / `failed` / `ambiguous_ligand` /\n`skipped_no_env`) and a `plip_<type>_count` column per interaction type detected\n(hydrophobic, hydrogen_bonds, salt_bridges, etc.).\n\n### Inside `boltz_dashboard.html`\n\nPanels, in order:", "# Regrow Your Teeth: The End of Dental Implants\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n[Phase I human trials launched at Kyoto University Hospital](https://toregem.co.jp/en/archives/8371) in October 2024, enrolling 30 healthy adult males aged 30-64, each missing at least one tooth. This initial phase prioritizes safety assessment and optimal dosing through intravenous antibody administration. If successful, [Phase II trials](https://remedico.app/newsfeed/here-are-several-slug-options-for-your-article-seo-optimized-recommended) will expand to children aged 2-7 with congenital anodontia, individuals born with severely compromised dentition affecting approximately [0.1% of the population](https://my.clevelandclinic.org/health/diseases/24322-tooth-agenesis). Unlike conventional tissue engineering approaches requiring cell transplantation, this [cell-free molecular therapy](https://www.drugtargetreview.com/news/87782/antibody-could-be-used-in-tooth-regeneration-therapy-study-shows/) represents a fundamentally simpler treatment paradigm. Patients would receive an injection rather than undergoing complex surgical procedures, dramatically reducing treatment complexity and recovery time.\n\nThe implications extend far beyond congenital conditions. Current tooth replacement options carry significant drawbacks: [dental implants cost $3,000-$6,000 per tooth](https://www.901dental.com/dental-implants-success-rates-and-risks/), require osseointegration over several months, and achieve [90-95% success rates over 10 years](https://pubmed.ncbi.nlm.nih.gov/27334865/)\u2014meaning 5-10% fail. Implant complications include peri-implantitis, mechanical failure, and the inability to provide natural proprioceptive feedback. Regenerated natural teeth, by contrast, would integrate seamlessly with nerves and blood vessels, maintain genuine tactile sensation, and potentially last decades without replacement. For the millions experiencing tooth loss from decay, injury, or periodontal disease, biological regeneration could represent the first true cure rather than merely a prosthetic substitute.\n\n[Toregem Biopharma](https://toregem.co.jp/en/), the Kyoto University spin-off company Dr. Takahashi co-founded, aims to bring tooth regeneration medicine to market by 2030. This ambitious timeline reflects the therapy\u2019s relatively straightforward mechanism compared to complex stem cell approaches. The research team envisions expanding applications beyond congenital cases to address age-related tooth loss, potentially ameliorating oral frailty in elderly populations. Regulatory pathways for monoclonal antibodies are well-established, potentially accelerating approval processes. Nevertheless, critical questions remain: ensuring regenerated teeth emerge in correct positions, confirming long-term stability, and verifying that third dentition structures possess the durability of natural permanent teeth.\n\nThis paradigm shift from repair to regeneration fundamentally redefines what\u2019s possible in dentistry. Rather than accepting tooth loss as permanent and managing it with artificial replacements, we\u2019re approaching the threshold where lost teeth can be biologically restored. The convergence of molecular biology, developmental genetics, and antibody engineering has unlocked a capability humans have carried dormant within their jaws all along. As Phase I trial results emerge over the coming years, we\u2019ll learn whether this promise translates to clinical reality\u2014and whether the era of dental implants will indeed give way to an age of natural tooth regeneration.\n\n| Key Concept | Description | Key References |\n| --- | --- | --- |\n| **USAG-1 Protein** | A biological inhibitor that blocks BMP and Wnt signaling pathways, preventing additional tooth development by suppressing dormant tooth buds | [Takahashi et al., Science Advances](https://pubmed.ncbi.nlm.nih.gov/39389160/) |\n| **Anti-USAG-1 Antibody** | Monoclonal antibody therapy that neutralizes USAG-1 protein, reactivating dormant tooth buds through enhanced BMP signaling to regenerate functional teeth | [Takahashi et al., Science Advances](https://pubmed.ncbi.nlm.nih.gov/33579703/) |\n| **BMP Signaling** | Bone Morphogenetic Protein pathway essential for tooth development, controlling cell proliferation, differentiation, and the number of teeth formed | [NIH National Library of Medicine](https://pmc.ncbi.nlm.nih.gov/articles/PMC4299140/) |\n| **Congenital Tooth Agenesis** | Developmental absence of one or more teeth, including hypodontia (1-5 missing) and oligodontia (6+ missing), affecting 0.1-10% of the population | [Journal of Clinical Medicine](https://pmc.ncbi.nlm.nih.gov/articles/PMC10290303/) |\n| **Third Dentition** | Dormant tooth buds present in humans that represent a potential third set of teeth, naturally activated in approximately 1% of people with hyperdontia | [Science Advances Journal](https://pmc.ncbi.nlm.nih.gov/articles/PMC7880588/) |\n| **Toregem Biopharma** | Kyoto University spin-off company founded by Dr. Takahashi conducting Phase I clinical trials at Kyoto University Hospital starting October 2024 | [Toregem BioPharma Official](https://toregem.co.jp/en/) |", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nA serial dilution (1:3 scheme) of the compounds were prepared in neat DMSO, with 100 nL of the compounds dispensed in the 384 wells assay plate. An 1-hr pre-incubation of the compounds with enzyme was included. A His-tagged construct of WRN (AA S2-S1235) was produced and used for the ATPase assay. Maximum inhibition control wells were set up with the addition of only assay buffer, and no inhibition control wells were set up with neat DMSO without test compounds. The reaction was initiated by addition of 5 \u03bcL substrate mix (200 \u03bcM ultrapure ATP, and 50 nM dsDNA prepared in assay buffer) and incubated at 25 \u00b0C for 35 min. A forked dsDNA oligomer substrate was prepared by annealing equal amounts of oligo GPat-1 ([BHQ2]GCACTGGCCGTCGTTTTACGGTCGTGACT) and oligo GPat-2 (TTTTTTACT-TAACGACGGCCAGTGC[TAM]) and was used as dsDNA substrate for WRN ATPase assay. A typical ATPase reaction consists of 25 nM forked dsDNA, 100 \u03bcM ATP, and 1 nM WRN prepared in assay buffer. The WRN ATPase reaction was stopped by the addition of 5 \u03bcL of ADP-Glo Reagent from ADP-Glo\u2122 Kinase Assay kit and incubated for 40 min to deplete the unreacted ATP. Afterwards, 5 \u03bcL of Kinase Detection Reagent from the ADP-Glo\u2122 Kinase Assay kit was added and incubated for 1 hr for the development of luminescent signal.\n\nLuminesce signals were measured on BMG PHERAstar FSX using luminescence protocol. Data analysis was carried out using Genedata software following % of control normalization. Fitting parameters include dose response IC50, Hill slope (n), maximum inhibition (S\u221e), and minimum inhibition (S0). The reported IC50 values for compounds in this invention are the geometric means of at least 2 independent replicates.\n\nResults of the WRN ATPase activity assay for compounds of the Examples are shown in Table 4.\n\n**TABLE 4. WRN ATPase activity of the compounds of the Examples.**\n\n| Example Compound | WRN ATPase Activity |\n|---|---|\n| 1 | + |\n| 2 | + |\n| 3 | + |\n| 4 | + |\n| 5 | + |\n| 6 | + |\n| 7 | ++ |\n| 8 | + |\n| 9 | + |\n| 10 | ++++ |\n| 11 | + |\n| 12 | ++ |\n| 13 | ++ |\n| 14 | ++ |\n| 15 | + |\n| 16 | ++++ |\n| 17 | ++ |\n| 18 | ++++ |\n| 19 | + |\n| 20 | ++ |\n| 21 | + |\n| 22 | + |\n| 23 | + |\n| 24 | ++ |\n| 25 | ++ |\n| 26 | ++ |\n\n*Note: ATPase activity is represented by the following: + is <100 nM; ++ is 100 nM\u2013<500 nM; +++ is 500 nM to <1000 nM; ++++ is 1000 nM to <5000 nM; and +++++ is 5000 or greater nM.*\n\n### Example B. WRN Helicase Activity Assay\n\nCompounds of the examples were tested for WRN helicase activity. DNA encoding the target protein, 6\u00d7HIS-TEV-WRN-517-945, was produced, cloned and transfected as detailed in Example D. To determine the inhibitory properties of the compounds, a 384 wells plate-based Helicase assay was set up to measure the extent of DNA unwinding (Helicase) activity of WRN. The WRN Helicase assay was designed to measure increase in fluorescence signal using a fluorescence based, dinner-quencher paired forked dsDNA oligomer substrate (custom synthesis from Sigma-Aldrich) prepared by annealing equal amounts of oligo GPat-1 ([BHQ2]GCACTGGCCGTCGTTTTACGGTCGTGACT) and oligo GPat-2 (TTTTTTACT-TAACGACGGCCAGTGC[TAM]). Continuous measurement of fluorescent signal was used to optimize the buffer conditions, concentration of enzyme (WRN), concentration of substrates (ATP & dsDNA), and the linearity of the Helicase reaction. All solutions were prepared in the following assay buffer: 50 mM Tris-HCl, pH-7.5, 2 mM MgCl2, 50 mM NaCl, 1 mM EGTA, 0.01% P-20, 0.05% BSA, and 2 mM DTT prepared with nuclease free water.\n\nA serial dilution (1:3 scheme) of the compounds was prepared in neat DMSO and 100 nL of the compounds was dispensed in the Greiner black low volume (Greiner #784076) 384 wells assay plate. An 1-hr pre-incubation of the compounds with enzyme was included. A His-tagged construct of WRN (AA S2-S1235) was produced and used for the ATPase assay. Maximum inhibition control wells were set up with the addition of only assay buffer, and no inhibition control wells were set up with neat DMSO without test compounds. The reaction was initiated by addition of 5 \u03bcL substrate mix (200 \u03bcM ultrapure ATP, and 50 nM dsDNA prepared in assay buffer) and incubated at 25 \u00b0C for 35 min. A typical Helicase reaction consists of 25 nM forked dsDNA, 100 \u03bcM ATP, and 1 nM WRN prepared in assay buffer. The WRN Helicase reaction was quenched by the addition of 5 \u03bcL \"STOP\" solution containing 100 mM EDTA (final EDTA concentration 33.33 mM) prepared in assay buffer. The plates were read on BMG PHERAstar FSX using 540/590 module to measure an increase of fluorescence. Data analysis was carried out using the Genedata software following % of control normalization. Fitting parameters include dose response IC50, Hill slope (n), maximum response (S\u221e), and minimum response (S0). The reported IC50 values for compounds in this invention are the geometric means of at least 2 independent replicates.\n\nResults of the WRN Helicase activity assay for compounds of the Examples are presented in Table 5.\n\n**TABLE 5. WRN Helicase activity of the compounds of the Examples.**\n\n| Example Compound | WRN Helicase Activity |\n|---|---|\n| 1 | + |\n| 2 | + |\n| 3 | + |\n| 4 | + |\n| 5 | + |\n| 6 | + |\n| 7 | ++ |\n| 8 | + |\n| 9 | + |\n| 10 | ++++ |\n| 11 | + |\n| 12 | ++ |\n| 13 | ++ |\n| 14 | ++ |\n| 15 | + |\n| 16 | +++++ |\n| 17 | ++ |\n| 18 | +++++ |\n| 19 | + |\n| 20 | ++ |\n| 21 | + |\n| 22 | + |\n| 23 | + |\n| 24 | ++ |\n| 25 | ++ |\n| 26 | + |\n\n*Note: Helicase activity is represented by the following: + is <100 nM; ++ is 100 nM\u2013<500 nM; +++ is 500 nM to <1000 nM; ++++ is 1000 nM to <5000 nM; +++++ is 5000 nM to 20,000 nM.*\n\n### Example C. WRN Inhibitor Cellular Activity Assays", "# The LockTac Revolution: Binding the Future of Therapeutics\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-22 on marcdeller.com (categories: sb).*\n\n* LOCKTACs stabilize, rather than disrupt, natural biomolecular interactions.\n* They offer therapeutic avenues for tackling previously intractable drug targets.\n* By reinforcing the molecular \u2018ties\u2019 between proteins, they can amplify or inhibit signaling pathways.\n* The 2025 Science review outlines both amplifying and inhibitory LOCKTAC examples and mechanisms.\n* Current research is focused on optimizing specificity, minimizing off-target effects, and clinical translation.\n* LOCKTACs\u2019 modularity and precision position them to redefine the landscape of future therapeutics.\n\n1. Science: **Load and lock: An emerging class of therapeutics that stabilize molecular interactions:**\n   <https://www.science.org/doi/10.1126/science.adx3595>\n2. ChemBioChem: **Modulation of Protein\u2013Protein Interactions by Small Molecules\u2014Key Strategies and Recent Examples (2024)**:\n   <https://chemistry-europe.onlinelibrary.wiley.com/doi/10.1002/cbic.202300636>\n3. Cancer Discovery: **Targeting Protein Complexes in Cancer Therapy (2024)**:\n   <https://aacrjournals.org/cancerdiscovery/article/14/5/1190/741173/Targeting-Protein-Complexes-in-Cancer-Therapy>\n4. Frontiers in Science: **Localized production of nucleic acid nanomedicine** **(2025)**:\n   <https://www.frontiersin.org/articles/10.3389/fsci.2025.1458636/full>\n\nIn the cutting-edge world of molecular medicine, a new class of drugs is capturing the attention of scientists and industry leaders alike. LOCKTACs\u2014molecules that stabilize preexisting biomolecular interactions\u2014instead of simply blocking or disrupting, work by reducing the rate at which two molecules fall apart inside the body. Unlike previous drug discovery strategies that have mainly focused on inhibiting or disassembling dynamic protein complexes, LOCKTACs are designed to \u2018lock\u2019 those vital connections into place, thereby amplifying or suppressing signaling with unprecedented precision. The mechanism underlying LOCKTACs is fundamentally different from better-known platforms like PROTACs; rather than forcing new bonds, LOCKTACs prolong the persistence of already meaningful protein partnerships. This extends the lifetime of a given biological signal or blocks a pathway depending on the context, such as enhancing therapeutic efficacy or halting disease progression.\n\nRecent research published in Science in August 2025 provides a comprehensive review of LOCKTACs from the perspectives of theory, preclinical models, and potential therapeutic applications. The article highlights how pioneer molecules in this class can either amplify, as with stabilizing beneficial interactions like SMN2 mRNA and its nuclear partner, or inhibit, as shown in the case of motor proteins whose function is blocked by persistently locked engagement. The versatility of LOCKTACs has sparked a surge of interest in exploring so-called \u2018undruggable\u2019 targets, such as transcription factors and macromolecular complexes long thought inaccessible using traditional drug modalities. With the field rapidly accelerating, several peer-reviewed reports from 2024-2025 have charted both the promise and the remaining challenges in discovery, optimization, and delivery of these molecules.\n\nKey areas of ongoing research focus on mapping out molecular blueprints for future LOCKTACs, fine-tuning selectivity, reducing off-target effects, and translating promising candidates from in vitro systems to animal models and eventually clinical trials. As LOCKTACs unlock new frontiers in drug development, they may provide long-sought answers to diseases currently lacking effective treatment options. The next few years promise to be transformative as these compounds move from proof of concept toward clinical reality.", "# Plastic, Persistent, and Potentially Carcinogenic\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-04-24 on marcdeller.com (categories: blog, microplastics).*\n\n* [**Long-term exposure to real-life polyethylene terephthalate nanoplastics induces carcinogenesis in vitro**](https://pubs.acs.org/doi/10.1021/acs.est.5c01628)\n\n* [**Microplastics role in cell migration and distribution during cancer cell division**](https://www.sciencedirect.com/science/article/pii/S0045653524003564)\n\n* **Microplastics accumulate preferentially in tumor tissue:** MNP concentrations are consistently higher in cancerous tissue than in adjacent healthy tissue across multiple cancer types.\n* **PET is a specific carcinogenic risk at chronic exposure:** Real-life PET nanoplastics induce genotoxicity and invasive potential only after 30 weeks, exposing critical gaps in standard toxicology assays.\n* **Sub-micron particles drive metastatic cell behavior:** 0.25 \u00b5m MNPs amplify cancer cell migration and are heritably transmitted through cell division without clearance.\n* **Immune subversion is a causal in vivo mechanism:** Microplastic-induced macrophage ferroptosis remodels the tumor immune microenvironment, directly accelerating metastasis in animal models.\n\n1. **Microplastics role in cell migration and distribution during cancer cell division:** [Brynzak-Schreiber et al., Chemosphere, 2024](https://www.sciencedirect.com/science/article/pii/S0045653524003564)\n2. **Detection and quantification of microplastics in various types of human tumor tissues:** [Zhao et al., Ecotoxicol. Environ. Saf., 2024](https://www.sciencedirect.com/science/article/pii/S0147651324008947)\n3. **Long-term exposure to real-life polyethylene terephthalate nanoplastics induces carcinogenesis in vitro:** [Guti\u00e9rrez-Garc\u00eda et al., Environ. Sci. Technol., 2025](https://pubs.acs.org/doi/10.1021/acs.est.5c01628)\n4. **Microplastic-induced macrophage dysfunction drives lung tumor progression through glutathione imbalance:** [Kim et al., ACS Nano, 2026](https://pubs.acs.org/doi/10.1021/acsnano.5c15425)\n5. **Emerging evidence on micro- and nanoplastics carcinogenicity: mechanisms, models, and signaling networks:** [Alaraby et al., Molecular Cancer, 2026](https://link.springer.com/article/10.1186/s12943-026-02657-y)\n\n**Emerging mechanistic evidence moves the microplastics-cancer debate beyond correlation**\n\nWe now live in a world where plastics do not simply surround us \u2014 they are inside us. Microplastics and nanoplastics (MNPs), particles less than 5 mm produced by the photodegradation and mechanical fragmentation of larger plastic waste, have been detected in virtually every human tissue sampled to date: lungs, liver, placenta, blood, and brain. A [landmark 2024 study by Zhao et al.](https://www.sciencedirect.com/science/article/pii/S0147651324008947) provided the first systematic quantification of MNPs directly in human tumor specimens, detecting polystyrene, polyvinyl chloride, and polyethylene in samples from 61 patients across six cancer types. Detection rates were striking: 80% in lung tumors and 70% in pancreatic cancer tissue, with particle concentrations reaching over 500 ng per gram of tissue. Critically, MNP loads were consistently elevated in cancerous tissue relative to adjacent non-cancerous tissue, a pattern now replicated across multiple cancer types and detection platforms, and one that shifts the conversation from incidental contamination to biologically meaningful accumulation.\n\nOf particular concern is polyethylene terephthalate, PET, the polymer that defines single-use drink bottles and synthetic textiles, and one of the most abundant MNP types identified in human respiratory airways. A 30-week chronic exposure study by [Guti\u00e9rrez-Garc\u00eda et al.](https://pubs.acs.org/doi/10.1021/acs.est.5c01628) in BEAS-2B bronchial epithelial cells demonstrated that real-life PET nanoplastics induce genotoxic damage, anchorage-independent growth, and a measurable increase in invasive potential only after prolonged accumulation, with no significant effects observed at 15 weeks. Transcriptomic analysis confirmed upregulation of non-small cell lung cancer pathways at the 30-week mark. The implication is direct: standard short-duration toxicology assays are structurally incapable of detecting the carcinogenic potential of PET, and its regulatory status likely reflects that blind spot.\n\n> **\u201cOur observations underscore the potential of MNPs as hidden catalysts for tumor progression, particularly through enhancing cell migration and possibly fueling metastasis.\u201d**\n\nThat framing, from [Brynzak-Schreiber et al.](https://www.sciencedirect.com/science/article/pii/S0045653524003564) working with four human colorectal cancer cell lines, captures the central mechanistic concern. Sub-micron MNPs at 0.25 \u00b5m significantly amplified cancer cell migration in a size-dependent manner, were inherited by daughter cells during division without apparent elimination, and accumulated preferentially in the non-proliferating core of 3D tumor spheroids. This intracellular persistence means that MNP burden is not cleared but transmitted, compounding through successive generations of tumor cells. In a parallel in vivo study, [Kim et al.](https://pubs.acs.org/doi/10.1021/acsnano.5c15425) demonstrated that orally ingested microplastics accumulate across multiple organs in tumor-bearing mice, triggering TLR2/TLR4-dependent macrophage ferroptosis through a paradoxical uncoupling of glutathione metabolism. Despite NRF2 pathway activation, selective suppression of GPX1 and GPX3 drove macrophage cell death, progressively dismantling the antitumor immune microenvironment and accelerating lung metastasis in a time-dependent and measurable manner in vivo.", "# Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\nAuthors: Zhuo, J.; Li, Y.-L.; Qian, D.-Q.; Burns, D.M.; Mei, S.; Rafalski, M.; Xu, M.; Cao, G.; Pan, Y.; Jia, Z.; Jalluri, R.; Epling, L.B.; Fenalti, G.; Deller, M.C.; Procak, J.; Covington, M.; He, X.; Collins, R.; Stubbs, M.; Burke, K.; Oliver, J.; Margulis, A.; Boer, J.; Wynn, R.; Scherle, P.; Diamond, S.; Newton, R.; Zhang, Y.; Metcalf, B.; Yao, W.\nSource: Journal of Medicinal Chemistry (Drug Annotation), 2026\n\nTo assess our original hypothesis that selective inhibition of JAK1 could be beneficial in treating autoimmune- and inflammatory-related diseases (Figure 1), 18 (INCB047986) and 22 (INCB054707) were evaluated for efficacy in several in vivo murine models. One such model was the adjuvant-induced arthritis (AIA) model in rats, which is an aggressive T cell-mediated autoimmune arthritis that closely mimics rheumatoid arthritis (RA) in humans. Thirteen days after immunization, 18 and 22 were dosed twice a day and once daily orally, respectively (Figure 5). Macroscopic examination of periarticular erythema and edema of the joints of each hind paw/ankle were evaluated and scored from 0 to 4 for each paw, and the sum of the two values provided clinical scores, with 0 being no inflammation and 8 being severe. Both 18 and 22 demonstrated a dose-dependent improvement in clinical score after only a few days of dosing. The 1 mg/kg arm displayed minimal effect or improvement in disease, whereas the 3 and 10 mg/kg arms demonstrated almost full resolution of disease for both 18 and 22. Further evaluation of joint inflammation was assessed using a plethysmometer to measure the hind paw volume, providing similar results for both compounds (data only shown for 22). The plethysmometer measurements of 22 at 3 and 10 mg/kg dosing returned the hind paw volume to normal levels by day 26, indicating a reduction in tissue swelling and inflammation. Further histopathologic evaluation of 22 was conducted in an AIA rat model in which the adjuvant arthritic hind ankles of rats were examined microscopically and scored 0\u20137 on each of the following criteria: bone resorption, periosteal new bone formation, inflammation, cartilage damage, and pannus. The scores for each paw/ankle were averaged, with a score of 0 indicative of being normal and 7 being severe disease. The 3 and 10 mg/kg arms showed a significant improvement in the disease histopathology scores, with an 81% and 94% reduction in the summed scores, respectively. Collectively, the above data suggest that 22 (INCB054707, povorcitinib) is highly efficacious in the murine arthritis model at 3 and 10 mg/kg doses and may have utility in a therapeutic setting for the treatment of rheumatoid arthritis, and potentially other autoimmune and inflammatory disorders that are reliant on JAK1 signaling.\n\nTo ensure that the observed in vivo efficacy was primarily due to JAK1 inhibition, 22 was assessed in several assays that were dependent on JAK1 heterodimerization for concomitant signaling (Table 7). Interleukin (IL)-2 stimulated human T cells that are dependent on JAK1/3 dimerization were effectively inhibited by 22 with an approximate half-maximal inhibitory concentration (IC50) of 10\u201330 nM for downstream phosphorylation of JAK2, STAT3, and STAT5 and the proliferation of T cells. Compound 22 effectively inhibited the phosphorylation of STAT1 and STAT3 with an IC50 of 360 and 690 nM, respectively, in human whole blood that was stimulated with IL-6. Additionally, MCP1 production that was stimulated by IL-6 in human PBMCs was inhibited with an IC50 of 178 nM. It is important to note that although IL-6-mediated signaling is dependent on the dimerization of JAK1/2, it has been demonstrated by Rodig et al. that cells derived from mice that are deficient of JAK1 are not effectively stimulated by IL-6. Guschin et al. further confirmed that JAK1 plays an essential role in IL-6-stimulated phosphorylation of STATs 1 and 3. The observation that the phosphorylation of STAT3 and STAT5 was not blocked by 22 in whole blood that was stimulated with the JAK2-specific cytokines TPO and granulocyte-macrophage colony-stimulating factor (GM-CSF) further confirmed that the reduction in inflammation was primarily driven by JAK1 inhibition. These data suggest that 22 effectively and selectively inhibits JAK1 over other JAK isoforms, such as JAK2, in a cellular and whole blood environment.\n\nIn vitro toxicity studies indicated that 22 had an IC50 for inhibition of hERG > 20 \u03bcM and HEK293 proliferation >20 \u03bcM. As previously mentioned, 22 did not inhibit CYPs 1A2, 2B6, 2C9, 2C19, 2D6, or 3A4 below 25 \u03bcM. In vivo, there was no significant weight loss observed for the rodent studies discussed above.\n\n*Figure 3.* Computational Modeling of Compound 3 (itacitinib) in JAK1 (purple) and JAK2 (cyan). JAK, Janus kinase.", "# JCSG Top96 Screen Formulation\nAuthors: Deller, M. C.\nSource: Joint Center for Structural Genomics, The Scripps Research Institute (screen formulation sheet, 16 April 2013), 2013\n\n## Notes on the screen\n\nList of the 96 most successful JCSG crystallization conditions with respect to JCSG structures deposited in the PDB (conditions responsible for 631 JCSG PDB depositions between 1/1/2001 and 9/23/2010). Conditions listed in plate order (according to major precipitant class) as currently employed at the JCSG.\n\nThe parent screen is noted and abbreviated as follows: H1: Crystal Screen (Hampton Research), H2: Crystal Screen 2 (Hampton Research), H1cryo: Crystal Screen Cryo (Hampton Research), H2cryo: Crystal Screen 2 Cryo (Hampton Research), W1: Wizard Classic 1 (Emerald Biosystems), W2: Wizard Classic 2 (Emerald Biosystems), W1cryo: Wizard Cryo 1 (Emerald Biosystems), W2cryo: Wizard Cryo 2 (Emerald Biosystems), MPD: MPD Grid Screen (Hampton Research), PEG/ion: PEG ion screen (Hampton Research), AmSO4: Grid Screen Ammonium sulfate (Hampton Research) and P6K/LiCl: Grid Screen PEG/LiCl (Hampton Research).\n\nThe corresponding JCSG Core Suite (Qiagen) conditions are listed and those also represented in the \"Core Screen\" (later JCSG+ screen) are denoted by a plus sign (+) in the JCSG Core Suite column. The top 10 conditions with respect to PDB depositions are highlighted in the PDB column in the original document. Due to formulation and manufacturing considerations two conditions (JCSG-IV A02 and JCSG-I G01, highlighted by a triangle (\u0394) in the JCSG Core Suite column), were modified in the Qiagen JCSG Core Suite formulation. To ensure consistency with the PDB entries the original screening condition, as used at the JCSG (W1 #27 and PEG/ion #48, respectively), is reported above and not the condition currently reported in the Qiagen formulation of the JCSG Core Suite. For example, JCSG-IV A02: 0.2 M Lithium sulfate, 1.2 M Sodium dihydrogen phosphate/0.8 M di-Potassium hydrogen phosphate, 0.1 M Glycine pH 10.5 vs. W1 #27: 0.2 M Lithium sulfate, 1.2 M Sodium dihydrogen phosphate/0.8 M di-Potassium hydrogen phosphate, 0.1 M CAPS pH 10.5 (change in buffer). JCSG-I G01: 20% (w/v) PEG 3350, 0.18 M tri-Ammonium citrate vs. PEG/ion #48: 20% (w/v) PEG 3350, 0.2 M di-Ammonium citrate (change in salt).\n\nJCSG Top96 Cryo conditions are abbreviated as follows: EG: Ethylene glycol, MPD: 2-Methyl-2,4-pentanediol, G: Glycerol, P200: PEG 200 and P400: PEG 400. JCSG Top96 Cryo conditions were selected from coarse screen crystallization conditions that produced a clear visual freeze and a score of zero for ice rings during diffraction screening, with the exception of JCSG Top96 conditions B10 (JCSG-III G11) and B11 (JCSG-IV G08) which resulted from fine-screen conditions and scored 1 for ice rings during diffraction screening; these conditions are denoted with a dagger (\u2020) in the Top96 Cryo column.", "# 2015 \u2013 Protein stability: a crystallographer\u2019s perspective\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: papers).*\n\n* [Link to Paper](https://marcdeller.com/wp-content/uploads/2025/09/en5571.pdf)\n\n* **Protein stability** is a critical factor for producing homogeneous, well-ordered crystals, which are essential for successful crystallization and structural studies.\n* **Maintaining stability** requires careful optimization during protein expression, purification, formulation, and storage to preserve both structure and function.\n* **Researchers use tools** such as mass spectrometry, SDS-PAGE, NMR, and cryo-EM to assess stability and guide improvements for crystallography and drug development\n* **Resilience Balance:** Stability gains offset entropic costs.\n* **Construct Refinement:** Removing disorder enhances crystallizability.\n* **Thermal Profiling:** *T\u2098* and \u0394*H* guide buffer and ligand choices.\n* **Temperature\u2010Varied Crystallography:** Reveals functional dynamics.\n* **Fusion and Cofactors:** Tags and ligands fortify fragile folds.\n* **Interface Locking:** Small molecules can stabilize oligomeric states.\n\n1. **Protein stability: a crystallographer\u2019s perspective:** [Deller, et al., Acta Cryst. F](https://journals.iucr.org/f/issues/2016/02/00/en5571/index.html)\n2. **Improving Protein Expression, Stability, and Function with Small Molecules:** [Pace, et al., J. Am. Chem. Soc.](https://pubs.acs.org/doi/10.1021/jacs.3c10941)\n3. **Temperature-Dependent Crystallography of K-RAS Highlights Conformational Ensembles:** [Jain, et al., PNAS](https://pmc.ncbi.nlm.nih.gov/articles/PMC11888411/)\n4. **BeStSel: analysis site for protein CD spectra\u20142025 update:** [Tsuboyama, et al., Bioinformatics](https://pmc.ncbi.nlm.nih.gov/articles/PMC12230724/)\n5. **A Tutorial Review on the Methodologies and Theories Utilized to Model Protein Crystallization:** [Anderson, et al., Crystal Growth & Design](https://pubs.acs.org/doi/10.1021/acs.cgd.4c00221)\n6. **Functional Protein Dynamics in a Crystal:** [Lee, et al., Nature Communications](https://www.nature.com/articles/s41467-024-47473-4)\n\nDetails practical, crystallographer-focused discussion on protein stability, the understanding of which is essential for optimizing critical processes like protein expression, purification, formulation, storage, and structural studies across the biotechnology, pharmaceutical, and academic sectors.\n\nProteins must withstand diverse chemical and physical challenges to reveal their structures. A crystallographer relies on\u00a0compositional homogeneity\u00a0and\u00a0conformational rigidity\u00a0to coax proteins into ordered crystals. By refining constructs to remove flexible segments, optimizing solution conditions, and harnessing techniques from thermal unfolding to multi-temperature crystallography, researchers can stabilize targets just enough to form diffraction-quality crystals without sacrificing biologically relevant motions.\n\nProteins inherently balance a small thermodynamic advantage of the folded state against a significant entropic cost. Each buried methyl group confers ~1.1 kcal\u00b7mol\u207b\u00b9 of stability but costs ~2.4 kcal\u00b7mol\u207b\u00b9 in entropy loss. Hydrogen bonds contribute a similar ~1.1 kcal\u00b7mol\u207b\u00b9 regardless of protein size. Thus, marginal gains in packing, salt bridges, or cofactor binding can tip the balance toward crystallizability. Construct engineering\u2014trimming disordered termini, introducing stabilizing point mutations, or removing proteolytic sites\u2014minimizes conformational heterogeneity and enhances lattice contacts.\n\nThermal denaturation assays (e.g., differential scanning fluorimetry) yield melting temperatures (*T\u2098*) and enthalpic profiles, guiding buffer and ligand selection to raise stability. Complementing these, room\u2010temperature and multi\u2010temperature X\u2010ray crystallography reveal how cryogenic snapshots may mask fluctuations critical for function, as shown by K-RAS conformers uncovered only at physiological temperatures. B-factors in crystal models report positional uncertainty, flagging flexible regions for construct redesign.\n\nStabilization strategies extend beyond small molecules: fusion tags (MBP, GroEL), cofactor supplementation, engineered disulfide bonds, and ancestral\u2010sequence reconstruction all bolster fold integrity. Thermophile-inspired salt bridges and hydrophobic core enhancements have yielded enzymes functioning at extreme conditions. Even protein\u2013protein interfaces can be locked in preferred states by ligands like Tafamidis, which stabilizes transthyretin dimers against amyloid formation.\n\nUltimately, crystallographic success hinges on fine\u2010tuning the protein\u2019s stability landscape: retaining enough dynamics to preserve function yet limiting heterogeneity for crystal order. This delicate equilibrium transforms a dynamic polymer into a stationary map of atomic precision.\n\n| Key Concept | Description | Reference |\n| --- | --- | --- |\n| Marginal Stability | Folded proteins gain ~1.1 kcal\u00b7mol\u207b\u00b9 per buried methyl vs. 2.4 kcal\u00b7mol\u207b\u00b9 entropic cost. | [Pace, et al., J. Mol. Biol.](https://pubs.acs.org/doi/10.1021/jacs.3c10941) |\n| Thermal Denaturation | DSF and CD assays provide *T\u2098* and \u0394*H* for buffer optimization. | [Shih, et al., Anal. Biochem.](https://pmc.ncbi.nlm.nih.gov/articles/PMC12230724/) |\n| Multi\u2010Temp Cryst. | Room and physiological temperature data uncover hidden conformers. | [Smith, et al., J. Mol. Biol.](https://pmc.ncbi.nlm.nih.gov/articles/PMC11888411/) |\n| Construct Engineering | Trimming disordered regions and removing protease sites improves order. | [Kissinger, et al., Acta Cryst. F](https://journals.iucr.org/f/issues/2016/02/00/en5571/index.html) |\n| Fusion Tags | MBP and chaperonins enhance solubility and folding yield. | [Thomson, et al., J. Biol. Chem.](https://pubs.acs.org/doi/10.1021/jacs.3c10941) |\n| Interface Stabilizers | Ligands like Tafamidis lock oligomers against aggregation. | [Bulawa, et al., Science](https://pubs.acs.org/doi/10.1021/jacs.3c10941) |", "# Extended Portfolio \u2014 Marc C. Deller, D.Phil.\nAuthors: Deller, M.C.\nSource: Personal portfolio (merged from two versions), 2025\n\n### The crystal structure of a bacterial Sufu-like protein defines a novel group of bacterial proteins that are similar to the N-terminal domain of human Sufu\n\n**Journal:** Protein Science 19 (11), 2131\u20132140, 2010. **Authors:** D. Das, R.D. Finn, P. Abdubek, T. Astakhova, H.L. Axelrod, C. Bakolitsa, X. Cai, \u2026\n\n### Bacterial pleckstrin homology domains: a prokaryotic origin for the PH domain\n\n**Journal:** Journal of Molecular Biology 396 (1), 31\u201346, 2010. **Authors:** Q. Xu, A. Bateman, R.D. Finn, P. Abdubek, T. Astakhova, H.L. Axelrod, \u2026\n\n### Structural basis of murein peptide specificity of a \u03b3-D-glutamyl-L-diamino acid endopeptidase\n\n**Journal:** Structure 17 (2), 303\u2013313, 2009. **Authors:** Q. Xu, S. Sudek, D. McMullan, M.D. Miller, B. Geierstanger, D.H. Jones, \u2026\n\n### A structural basis for the regulatory inactivation of DnaA\n\n**Journal:** Journal of Molecular Biology 385 (2), 368\u2013380, 2009. **Authors:** Q. Xu, D. McMullan, P. Abdubek, T. Astakhova, D. Carlton, C. Chen, H.J. Chiu, \u2026\n\n### Crystal structures of MW1337R and lin2004: Representatives of a novel protein family that adopt a four-helical bundle fold\n\n**Journal:** Proteins: Structure, Function, and Bioinformatics 71 (3), 1589\u20131596, 2008. **Authors:** P. Kozbial, Q. Xu, H.J. Chiu, D. McMullan, S.S. Krishna, M.D. Miller, P. Abdubek, \u2026\n\n### Crystal structures of two novel dye-decolorizing peroxidases reveal a \u03b2-barrel fold with a conserved heme-binding motif\n\n**Journal:** Proteins: Structure, Function, and Bioinformatics 69 (2), 223\u2013233, 2007. **Authors:** C. Zubieta, S.S. Krishna, M. Kapoor, P. Kozbial, D. McMullan, H.L. Axelrod, \u2026\n\n### Identification and structural characterization of heme binding in a novel dye-decolorizing peroxidase, TyrA\n\n**Journal:** Proteins: Structure, Function, and Bioinformatics 69 (2), 234\u2013243, 2007. **Authors:** C. Zubieta, R. Joseph, S. Sri Krishna, D. McMullan, M. Kapoor, H.L. Axelrod, \u2026\n\n### Cell surface receptors\n\n**Journal:** Current Opinion in Structural Biology 10 (2), 213\u2013219, 2000. **Authors:** M.C. Deller, E.Y. Jones.\n\n### Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\n\n**Journal:** Structure 8 (8), 863\u2013874, 2000. **Authors:** M.C. Deller, K.R. Hudson, S. Ikemizu, J. Bravo, E.Y. Jones, J.K. Heath.\n\n## Portfolio, Web and News\n\n- **Blog:** https://marcdeller.com\n- **LinkedIn:** www.linkedin.com/in/marccdeller\n- **Google Scholar:** https://bit.ly/marcdeller_pubs\n- **Protein structures:** https://bit.ly/marcdeller_proteins\n- **Mentored Lemelson-MIT Student Prize** \u2014 Stanford News, April 2017: \"Undergrads win prize for work combating antibacterial resistance\" (https://news.stanford.edu/stories/2017/04/undergrads-win-prize-work-combating-antibacterial-resistance)\n- **\"Your one-stop shop for producing, crystallizing biomolecules\"** \u2014 Stanford Medicine News, April 2016 (https://med.stanford.edu/news/all-news/2016/04/your-one-stop-shop-for-producing-crystallizing-biomolecules.html)\n- **Top articles in structural biology (Spring 2020)** \u2014 IUCr Journals special issue (https://journals.iucr.org/special_issues/2020/biologytoparticles/index.html)\n\n## Continued Learning\n\n- 2025 \u2013 Google: Foundations of Project Management (https://www.coursera.org/account/accomplishments/verify/605BGOEB1YGC)\n- 2025 \u2013 IBM: Introduction to Artificial Intelligence (https://www.coursera.org/account/accomplishments/verify/71R8R1FHA194)\n- 2025 \u2013 AWS: Serverless Architectures (https://www.coursera.org/account/accomplishments/verify/X9AD2JWH2HE9)\n- 2024 \u2013 Google: Foundations: Data, Data, Everywhere (https://www.coursera.org/account/accomplishments/verify/4S09Q8ICKGXN)\n- 2024 \u2013 Coursera: Introduction to R: Basic R syntax (https://www.coursera.org/account/accomplishments/verify/T687VG5UYI5O)\n- 2024 \u2013 Coursera: Dashboard Development with Shiny: GenAI for Retail Analysis (https://www.coursera.org/account/accomplishments/verify/OO09W4LMZGCB)\n- 2024 \u2013 LinkedIn Learning: Amplify Your Communication Skills with Generative AI", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nThe data in panel (a) successfully models as a monodisperse population with one major species at a hydrodynamic radius of 1.03 \u00d7 10^\u22122 \u03bcm (10.3 nm) and an estimated molecular weight of ~84 kDa. The estimated molecular weight was calculated by comparison with the hydrodynamic radii of standard globular proteins of known mass. This compares favourably with the theoretical molecular weight of LIFR-FN3 of ~94 kDa. Comparison of panels (a), (b) and (c) suggests that the system is in a state of dynamic equilibrium with large-scale aggregates (>0.1 \u03bcm) present in the solution alongside a population at ~84 kDa. On filtering the aggregates are removed and over a period of 10 minutes the aggregates are reformed so re-establishing the position of equilibrium.\n\n*Figure 4-8.* Figure showing dynamic light scattering data of deglycosylated LIFR-FN3. Samples of 293T-expressed LIFR-FN3 were deglycosylated overnight with non-recombinant PNGase F (New England Biolabs) and analysed on a DynaPro-LS Sizing Instrument (Protein Solutions Inc.). Typically, 0.5 ml of 2 mg/ml sample was injected through a 0.1 \u03bcm filter and measurements taken over a 10 minute period at 20 \u00b0C. Sample shown at (a) 0 minutes, (b) 5 minutes and (c) 10 minutes after injection. Red dotted line represents hydrodynamic cut off of 0.1 \u03bcm filter. Green dotted line represents monodisperse population modelled at 10.3 nm with an estimated molecular weight of ~84 kDa.\n\n#### 4.3.3.6 Deglycosylation of LIFR with PNGase F\n\nFigure 4-9 below shows a western blot of 12 % polyacrylamide gel of LIFR-FN3 samples from various deglycosylation reactions using PNGase F. Lane numbering corresponds to reactions set up in Table 4-2. Lanes 1 and 2 contain control samples of non-recombinant and recombinant PNGase F, respectively, and it can be seen that no cross reactivity to the LIFR antibody is observed. Lanes 3 and 8 contain control samples of LIFR-FN3 incubated with and without G7 buffer, respectively, and single distinct bands at ~116 kDa are observed. Lane 4 contains 17 \u03bcg of LIFR-FN3 incubated with 30.6 \u03bcg of non-recombinant PNGase F (New England Biolabs) and a single band at ~90 kDa is observed. Lanes 5, 6 and 7 contain 17 \u03bcg of LIFR-FN3 incubated with 9.4, 18.8 and 30.6 \u03bcg of recombinant PNGase F, respectively, and one main band is observed at ~90 kDa. Lane 9 contains broad range molecular weight rainbow markers running at 135, 80, 43, 32, 14 and 8 kDa (Bio-Rad).\n\nComparison of lanes 3 and 7 suggests that deglycosylation of LIFR-FN3 is occurring on incubation with higher levels of recombinant PNGase F, due to the reduction in apparent molecular weight from ~114 to ~95 kDa. This compares favourably with the theoretical molecular weight of non-glycosylated LIFR-FN3 of ~94 kDa. Lane 5 suggests that lower levels of recombinant PNGase F are less effective and as such result in a slightly heavier and more diffuse protein population. Comparison with lane 4 shows that non-recombinant PNGase F is the most effective with comparable enzyme levels demonstrating a larger band shift to ~90 kDa and a less diffuse band.\n\n*Figure 4-9.* Figure showing western blot of LIFR-FN3 samples at various stages of deglycosylation using PNGase F. 12 % polyacrylamide gel is shown probed with LIFR antibody (Binding Site). Lane numbering corresponds to reactions set up in Table 4-2. Lanes 1 and 2 contain control samples of non-recombinant and recombinant PNGase F, respectively. Lanes 3 and 8 contain control samples of LIFR-FN3 incubated with and without G7 buffer, respectively. Lane 4 contains 17 \u03bcg of LIFR-FN3 incubated with 30.6 \u03bcg of non-recombinant PNGase F (New England Biolabs). Lanes 5, 6 and 7 contain 17 \u03bcg of LIFR-FN3 incubated with 9.4, 18.8 and 30.6 \u03bcg of recombinant PNGase F, respectively. Lane 9 contains broad range molecular weight rainbow markers running at 135, 80, 43, 32, 14 and 8 kDa (Bio-Rad).\n\n#### 4.3.3.7 Receptor Competition and Titration Assays\n\nFigure 4-10 below shows graphs of (a) receptor titration and (b) receptor competition assays of ligand binding to LIFR-FN3-3C-Fc. Graph (a) demonstrates that the concentration range of ligand used leads to saturation of the LIFR-FN3-3C-Fc receptor under the conditions employed. This also shows that specific binding is occurring due to the sigmoidal increase in absorbance with increasing concentration of control human biotinylated LIF. Binding is not observed to purified/deglycosylated material or replenished 293T supernatant, as both show background responses similar to the control FGFR2 sample. Binding of human biotinylated LIF to 293T expressed LIFR-FN3-3C-Fc shows an EC50 value of 12.1 ng/ml which compares favourably to the control sample of CHO cell expressed material which displays an EC50 value of 10.4 ng/ml.\n\nFrom graph (b) it can be seen that increasing the concentration of human LIF causes a sigmoidal decrease in the absorbance reading due to competition with biotinylated human LIF. Little or no fall in absorbance is demonstrated by the control samples of murine LIF or human OSM-196 suggesting that no competition is taking place. Binding of human LIF to 293T expressed LIFR-FN3-3C-Fc displays an EC50 value of 166.4 ng/ml.\n\n*Figure 4-10.* Figure showing graphs of (a) receptor titration and (b) receptor competition assays of ligand binding to 293T expressed murine LIFR-FN3-3C-Fc. Panel (a) shows binding of human biotinylated LIF to various samples of LIFR-FN3-3C-Fc including control (green; EC50 10.4 ng/ml), purified/deglycosylated (pink), 293T supernatant (red; EC50 12.1 ng/ml) and replenished 293T supernatant (blue). Panel (b) shows competitive binding of various ligands to murine LIFR-FN3-3C-Fc in the presence of 200 ng/ml human biotinylated LIF including human LIF (green; EC50 166.4 ng/ml), murine LIF (red) and as control human OSM-196 (blue).\n\n#### 4.3.3.8 Production and Purification of LIFR", "# Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design\nAuthors: Irimia, A.; Serra, A.M.; Sarkar, A.; Jacak, R.; Kalyuzhniy, O.; Sok, D.; Saye-Francisco, K.L.; Schiffner, T.; Tingle, R.; Kubitz, M.; Adachi, Y.; Stanfield, R.L.; Deller, M.C.; Burton, D.R.; Schief, W.R.; Wilson, I.A.\nSource: PLoS Pathogens, 2017\n\nStructure determination and neutralization potency experiments with the best binding mutants were then performed to determine if changing the light-chain surface charge by mutation leads to conformational changes or compromises neutralization. Abolition or decrease in neutralization would suggest that the 10E8 light chain is oriented toward and interacts with the viral membrane. We focused on mutants 1\u20133 and 5, as binding to the peptide-scaffolded T117v2 was decreased for mutant 4.\n\n### Structural characterization of the 10E8 LC mutants\n\nX-ray structures of mutants 1\u20133 and 5 (Table 1) were determined in complex with the T117v2 scaffold at 1.6 \u00c5 (mutant 1) and 2.0\u20132.2 \u00c5 (mutants 2, 3 and 5) resolution. Superposition of the C\u03b1 atoms of the variable domains of mutants and wild-type 10E8 shows nearly identical conformations (C\u03b1 r.m.s.d. of 0.40, 0.23, 0.31 and 0.39 \u00c5 for mutants 1, 2, 3, and 5, respectively). Thus, the mutations mainly result in different charge distributions on the light-chain surface (Fig 3A\u20133E), with the most negative surface observed for mutant 3. Although mutant 1 has a slightly larger r.m.s.d. (0.4 \u00c5), the differences in the main chain occur mainly at the Fab elbow region. Only in mutant 5 do the R29A(L) and Y32A(L) mutations produce a slight shift in the main-chain for FRL2 (residues 48, 49), CDRL2 (50\u201356), FRL3 (57\u201372) and CDRL1 (24\u201332) with maximum C\u03b1 differences observed in CDRL2 (1.3 \u00c5 for Lys51(L)), FRL3 (0.9 \u00c5 for Gly68(L)) and CDRL1 (0.9 \u00c5 between Arg29(L) and Ala29(L); S4 Fig). The shift for CDRL2 in mutant 5 may be the direct result of substituting Tyr32(L) with alanine, which allows the nearby Phe48(L) side chain to rotate and influence the conformation of Lys51(L), Asn52(L) and Asn53(L). None of the mutations affect binding to T117v2 (in agreement with the SPR study, Table 1) as the mutant CDRH3 conformations are almost identical to wild-type 10E8 (S4 Fig).\n\n*Fig 3. Distribution of charge on the surfaces of the 10E8 light-chain variants.* The electrostatic solvent accessible surface (contoured at \u00b15 kT/e) of the light-chain region presumed to bind the viral membrane is shown for: (A) wild-type mature 10E8; (B) 10E8 mutant 1; (C) 10E8 mutant 2; (D) 10E8 mutant 3; (E) 10E8 mutant 5. Negative and positive charges are indicated in red and blue colored surfaces, respectively. The ligands bound in the lipid-binding site are shown as dark cyan sticks and the epitope scaffold is shown in gray. The underlined superscript letters designate the original residues in the 10E8 wild type.\n\nThe four mutant structures show different bound ligands (Fig 3B\u20133E) depending on the residue mutated and on the components of the crystallization mother liquor and cryoprotectants (n.b. all mutants were crystallized without lipids and in different conditions). Mutant 1 has glycerol from the cryoprotectant in the lipid-binding site (Fig 3B; S2 Table; S5A\u2013S5C Fig), as also observed for the wild-type 10E8-T117v2 complex cryoprotected with glycerol (Fig 1D; S2E and S2F Fig). In mutant 2 for which Arg29(L) was mutated to glutamate, only water molecules seem to occupy this site when ethylene glycol was used for cryoprotection rather than glycerol. However, in mutant 3, with the same mutation (R29E(L)), a phosphate from the crystallization condition occupies the site (S5E and S5F Fig) suggesting that replacement of Arg29(L) with a negative charge might still retain binding to at least a phosphate, which is a component of phospholipid head groups (n.b. glycerol was absent from the buffers used with crystallization of mutant 3, and thus does not compete for phosphate binding to this site). In mutant 5, where R29(L) and Y32(L) are mutated to alanine (Fig 3E), glycerol (used for cryoprotection) is not observed at this location in either of the two Fabs in the asymmetric unit, suggesting that the lipid binding site is severely disrupted by these mutations.\n\n### Neutralization potency of the 10E8 light-chain mutants\n\nIn parallel, we probed the neutralization potency of 10E8 mutants 1\u20133 and 5 against a panel of 109 viruses to determine if these mutations in the light chain interfere with neutralization. A decrease in neutralization potency compared to wild type was observed with most of the strains for mutant 1 in which three arginines (two are proximal to the lipid-binding site) were reverted to the germline residues (Fig 3B and Table 2). A more significant decrease in neutralization was observed for mutant 2 (Fig 3C and Table 2). Increasing the amount of negatively charged residues on the light-chain surface in mutant 3 results in loss of neutralization (Fig 3D and Table 2) of most of the viruses that we tested with the breadth dropping to 34%. Interestingly, mutation of only R29(L) and Y32(L) to alanine in mutant 5, which was designed to disrupt the observed lipid-binding site (Fig 3E), also abrogated the neutralization of most of the viruses tested (breadth of 44%, Table 2), despite retaining picomolar binding affinity to the epitope scaffolds. Thus, mutation of basic and polar residues on the light-chain surface to acidic residues does not interfere with binding to the MPER peptide epitope, but neutralization potency and breadth decreased as the number of negatively charged residues on the surface increased. Substantial loss of neutralization occurred when the lipid-binding site was disrupted (mutant 5) or when the light-chain surface patch contained mainly acidic residues (mutant 3).\n\n*Table 2.* Neutralization breadth and potency of wild-type 10E8 and 10E8 light-chain mutants assayed on an indicator panel of one hundred nine pseudotype viruses.", "# Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\nAuthors: Deller, M.C.; Hudson, K.R.; Ikemizu, S.; Bravo, J.; Jones, E.Y.; Heath, J.K.\nSource: Structure, 2000\n\nThe identity of the purified hOSM (1\u2013187) protein employed for crystallisation was confirmed using electrospray ionisation on a Micromass BioQ II triple, quadrapole, atmospheric pressure mass spectrophotometer equipped with an electrospray interface operating in the positive ion mode (Robin Aplin, Oxford Centre for Molecular Science, Oxford). N-terminal sequencing was also carried out for 10 cycles using an Applied Biosystems 494A/473A Sequencer (Tony Willis, MRC Immunochemistry Unit, Oxford).\n\n### Bioassays of hOSM mutants\n\nMutant forms of hOSM were tested for biological activity on the Ba/F3-hLIFR/hgp130 cell line as described previously [24]. The BA/F3-hOSMR/hgp130 cell line was generated by co-transfecting Ba/F3 cells with a full-length hgp130 cDNA expressed under the control of the cytomegalovirus (CMV) promoter in plasmid pCDNA3 and the full-length hOSMR cDNA [15] expressed under the control of the CMV promoter in plasmid pCDNA3. Transfected cells were cultured overnight in 50 ng/ml IL-3 and then selected for co-expression of hgp130 and hOSMR in the presence of 50 ng/ml hOSM. Co-expression of both receptors in co-transfectants was confirmed by FACS analysis after staining with antibodies directed against hgp130 [37] and hOSMR (KRH and JKH, unpublished observations). Bioassays were performed as described previously [19]. All assays were performed in triplicate in at least three independent experiments. The data presented are the results from a single experiment, as it was noted that absolute response to defined concentrations of ligand varied between different batches of cells, although the relative potencies of the WT and mutant proteins were preserved. The SEM of triplicate determinations was always less than 15% of the mean.\n\n### Ligand-binding assays\n\nHuman gp130 and hLIFR were expressed as fusion proteins with human Fc as described previously [24]. The fusions were purified by protein A sepharose affinity chromatography and employed in solid-phase ligand-binding assays using biotin-conjugated hOSM as the tracer [24]. The ability of mutant forms of hOSM to compete with biotin\u2013hOSM tracer was analysed using the PRISM software package (Graphpad Software) using a single-site binding model.\n\n### Crystallization and derivatisation\n\nFor crystallisation trials, hOSM(1\u2013187) was concentrated to 8 mg/ml in 50 mM Tris-HCl containing 150 mM sodium chloride, pH 7.5. Crystals grew by vapour diffusion within 1\u20133 days at 22\u00b0C from hanging drops typically comprising 2 \u03bcl of protein solution plus 2 \u03bcl of reservoir solution (30% w/v PEG 35,000, 0.25 M sodium acetate, pH 7.5). Crystals measured approximately 0.3 \u00d7 0.2 \u00d7 0.2 mm3 and showed ordered Bragg diffraction to 2.2 \u00c5 when exposed to synchrotron radiation. Preliminary characterisation indicated that the crystals belong to the orthogonal spacegroup P212121 with unit cell dimensions a = 35.8 \u00c5, b = 53.1 \u00c5, c = 106.8 \u00c5 and \u03b1 = \u03b2 = \u03b3 = 90\u00b0. There is one molecule per asymmetric unit and the crystal solvent content is ~48.0% (Matthews coefficient, Vm = 2.3 \u00c53/Da). Tests using a number of heavy atom compounds highlighted EMP as a putative heavy atom derivative. Derivatisation was carried out by soaking crystals in pre-equilibrated drops of mother liquor containing 14 mM EMP for 16\u201324 h.\n\n### Data collection\n\nMAD data were collected to a resolution of 2.6 \u00c5 from a single EMP derivative crystal on beam line BM14 at the ESRF, Grenoble, France. Crystals were briefly passed through a cryoprotectant solution consisting of a 2 \u03bcl drop of 30% (w/v) PEG 35,000, 0.25 M sodium acetate, pH 7.5 overlayed with 2 \u03bcl of 30% (w/v) PEG 8000, 0.25 M sodium acetate, pH 7.5 and then flash-cooled directly in liquid propane. The crystal was maintained at 100 K in a dry air stream and data recorded using a MAR 345 image plate detector (MAR Research, Hamburg, Germany). The mercury LIII edge of the derivative crystal was characterised by a fluorescence scan and X-ray data sets collected at three wavelengths (\u03bb = 1.00669 \u00c5, 1.00914 \u00c5 and 0.826516 \u00c5). A 2.2 \u00c5 resolution native data set was also collected at the same beamline from a single cryo-cooled crystal.\n\nIndividual data sets were auto-indexed, integrated and scaled using the programs DENZO and SCALEPACK [38]. Structure factor amplitudes were calculated using the program TRUNCATE [39], the data sets scaled using the program SCALEIT [40], and normalised anomalous scattering magnitudes estimated using the program REVISE [41]. Data collection statistics are reported in Table 1.\n\n### Structure determination and analysis\n\nScaling of MAD data sets yielded maximal dispersive differences of 10.6% and maximal anomalous differences of 9.7%. The mercury sites for the EMP derivative were determined by manual inspection of difference and anomalous Patterson syntheses and confirmed using the programs RSPS and SOLVE [42]. The sites were refined using the program SHARP [43], and MAD phases were calculated to 2.6 \u00c5 resolution (figure of merit = 0.46, phasing power = 1.84). Initial maps based on these phases and Fobs from the EMP \u03bb2 data set proved uninterpretable. Hendrickson-Lattman coefficients for SIRAS phases derived using the EMP derivative and native data were combined with the MAD phases. Maps based on these combined phases and the EMP \u03bb2 Fobs (figure of merit 0.51, phasing power = 1.90) showed improved main chain density. Phase improvement was carried out using solvent flattening, as implemented by the program SOLOMON [44], and the structure was traced using the interactive graphics program O [45].", "# Marc C. Deller \u2014 LinkedIn profile summary\nAuthors: Deller, M.C.\nSource: LinkedIn (linkedin.com/in/marccdeller), 2026\n\n## Headline\n\n\"Biotech Co-Founder & Structural Biology Expert | 400+ Protein Structures | AI-Driven Drug Discovery | Transforming Drug Discovery Through Innovation\"\n\n## Industry and location\n\nMarc lists his industry as Biotechnology Research and his location as Wilmington, Delaware, United States.\n\n## About / summary\n\nMarc Deller is CSO and Co-Founder of Elora Therapeutics, which is pioneering a first-in-class enzyme therapy designed to degrade synthetic microplastics that accumulate in blood and tissues, linked to chronic inflammation and metabolic diseases affecting millions worldwide. He describes 20+ years leading transformative innovation in structural biology and drug discovery, combining deep protein-science expertise with AI-driven automation to accelerate discovery of life-changing medicines. His cross-disciplinary approach bridges protein engineering, bioinformatics, biophysics, cloud computing and lab automation.\n\nCareer highlights, as stated on LinkedIn:\n\n- Delivered 400+ high-impact protein structures supporting 6+ FDA preclinical IND filings across oncology, immunology, virology and ophthalmology.\n- Built 2 state-of-the-art automated research facilities featuring X-ray crystallography and Cryo-EM platforms, enabling rapid gene-to-structure workflows.\n- Spearheaded structural biology efforts on 20+ therapeutic programs, shortening hit-to-lead timelines and driving drug-pipeline success.\n- Mentored and led 70+ scientists to foster innovation and operational excellence.\n- Managed $8M+ research and equipment budgets to implement cutting-edge capabilities.\n\nCore strengths, as stated on LinkedIn: strategic leadership driving scientific insights that propel drug-discovery breakthroughs; cross-functional collaboration integrating biology, chemistry and computational science; technological innovation deploying Cryo-EM, AI/ML and automation to accelerate discovery platforms; project management crafting visionary strategies for urgent biological needs; and team empowerment, cultivating a culture of creativity, mentorship and high performance.\n\nMarc's profile invites connections from biotech innovators, scientific collaborators and anyone passionate about enzyme therapeutics or emerging health challenges, and notes he's happy to talk Nottingham Forest football.\n\n## Links\n\nMarc's LinkedIn profile lists: his personal site (https://marcdeller.com), his GitHub (https://github.com/bellcheddar), his Scopus author profile, his app-hosting site (https://mdeller.com), Elora Therapeutics (https://eloratherapeutics.com), and a r\u00e9sum\u00e9 PDF hosted on his own site.\n\n## Note on scope\n\nMarc's therapeutic-area experience, as summarised on LinkedIn, spans oncology, immunology, virology and ophthalmology \u2014 a broader list than the oncology/immunology focus emphasised in his r\u00e9sum\u00e9 and Incyte-era corpus documents, reflecting the full breadth of IND-supporting work across his career.", "# Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design\nAuthors: Irimia, A.; Serra, A.M.; Sarkar, A.; Jacak, R.; Kalyuzhniy, O.; Sok, D.; Saye-Francisco, K.L.; Schiffner, T.; Tingle, R.; Kubitz, M.; Adachi, Y.; Stanfield, R.L.; Deller, M.C.; Burton, D.R.; Schief, W.R.; Wilson, I.A.\nSource: PLoS Pathogens, 2017\n\n### Identification of 10E8 regions interacting with the viral membrane\n\nCrystal structures of 10E8 in complex with T117v2 epitope scaffold and lipids 06:0 phosphatidylglycerol (PG) (2.37 \u00c5 resolution) or 06:0 phosphatidic acid (PA) (2.62 \u00c5 resolution) led to identification of a lipid-binding site at the proximity of the CDRL1 and CDRH3 loops (Fig 1A). Electron density for the lipid head groups and part of the PG acyl tail were observed (S2 Fig). The orientation of the lipid fragments with respect to 10E8 suggests that the hydrophobic lipid tails do not interact with the Fab or T117v2 and are disordered. The 06:0 PG and 06:0 PA head groups are bound into a crevice delineated by CDRL1 (Leu28(L), Arg29(L), Ser30(L), His31(L) and Tyr32(L)), FRL3 (Ala66(L), Ser67(L) and Gly68(L)), and by Trp100b(H), Ser100c(H) and Gly100d(H) at the CDRH3 tip (Fig 1B and 1C; where (L) stands for light and (H) for heavy chains). Glycerol, the cryo-protection component of the crystals, occupies the lipid-binding site in the 10E8-T117v2 structure when exogenous lipids are not added in the crystallization experiments (Fig 1D).\n\n*Fig 1. 10E8 lipid-binding site.* (A) Crystal structure of the 10E8-T117v2 complex bound to 06:0 PG (sticks; red, glycerol and phosphate moieties of the head group; cyan, head group region linked to the lipid tails). The 10E8 light chain (LC) is shown in beige, heavy chain (HC) in violet, and T117v2 scaffold containing the MPER epitope (pink) in gray. This color scheme is used throughout. (B) All potential hydrogen-bond interactions (within ~3.5 \u00c5) of the 06:0 PG fragment (colored by atoms) with 10E8 residues of CDRL1 (beige) and CDRH3 (violet) are shown as dashed lines. (C) All potential hydrogen-bond interactions (within ~3.5 \u00c5) of the 06:0 PA fragment (colored by atoms) with CDRL1 (beige) and CDRH3 (violet) residues of 10E8. (D) All potential hydrogen-bond interactions (within ~3.5 \u00c5) of a glycerol found in the lipid-binding site in the 10E8-T117v2 structure in the absence of added lipids.\n\nInterestingly, the 10E8 light chain displays basic surface patches arising from somatically mutated (Arg17(L), Arg24(L), Arg70(L)), and germline-encoded (Arg29(L), Lys51(L), Arg61(L)) residues in a plane with the bound lipid head groups observed in our structures, as well as with K/R683, the last residue of the gp41 ectodomain before the TM region (Fig 2A). Thus, this basic as well as Ser/Thr/Asn rich polar surface (Fig 2B) is likely to interact with polar and negatively charged lipid head groups of the viral membrane.\n\n*Fig 2. Design of 10E8 light-chain mutants.* (A) Several light chain (LC; beige) basic residues (Arg and Lys shown as sticks) are in the same plane with the lipid head group (red and cyan sticks) and Lys683 (red sticks) of the gp41 MPER epitope (red). The heavy chain (HC) is shown in violet. (B) Light-chain surface residues (beige) that are presumed to face the viral membrane. The heavy-chain residues at the tip of CDRH3 are shown as violet sticks. The 10E8 peptide epitope (pink) from the T117v2 scaffold is shown up to Lys683 (pink; last residue of gp41 ectodomain). (C) Alignment of the amino-acid sequence of the five 10E8 light-chain mutants compared to wild-type 10E8. Only the variable region of the 10E8 light-chain variants are shown with the Kabat numbering scheme on top of the alignment. The mutations are highlighted in red.\n\n### Design of light-chain mutants to probe the orientation of 10E8 with respect to the membrane\n\nTo further investigate and validate that the light chain of 10E8 is involved in binding to the lipid head groups of the membrane upper leaflet, we designed several 10E8 light-chain mutants (Fig 2C): mutant 1 reverts the three somatically mutated arginines to 10E8 germline residues (R17Q(L), R24Q(L), R70T(L)); mutant 2 replaces most of the lipid-proximal arginine residues with aspartate or glutamate (R24E(L), R29E(L), R70E(L), as well as R17D(L)); mutant 3 has all the basic residues as well as some of the polar residues on this surface mutated to D or E (R17D(L), R24E(L), R29E(L), K51D(L), N52D(L), S65D(L), S67D(L), R70E(L), S76D(L)); mutant 4 has R29(L) and Y32(L), which flank the lipid head groups, mutated to glutamate (R29E(L) and Y32E(L)); and mutant 5 has R29(L) and Y32(L) mutated to alanine (R29A(L) and Y32A(L)). Overall, mutants 1 to 3 were constructed to coat the presumed membrane-interacting surface of 10E8 with different amounts of negative charge to investigate binding affinity and specificity of the 10E8 epitope (protein and lipid) and to examine the effects of the mutations on neutralization potency. In addition, mutants 4 and 5 were designed to disrupt the lipid-binding site that we observe in the crystal structures.\n\n### Influence of light-chain mutations on binding to the epitope scaffolds\n\nTo investigate the effect of the light-chain surface mutations on 10E8 IgG binding to the T117v2 scaffold, we performed surface plasmon resonance (SPR) experiments. SPR analysis (Table 1 and S3 Fig) shows that, with the exception of mutant 4 for which the residues involved in lipid head group binding were mutated to glutamate (R29E(L) and Y32E(L)), all other mutants retained picomolar affinity (KD) to the T117v2 scaffold (Table 1). Mutations R29E(L) and Y32E(L) in mutant 4 (1.2 nM KD for T117v2) resulted in a ~40-fold reduction in binding compared to the affinity-matured 10E8 IgG (0.029 nM for T117v2), while mutation of the same residues to alanine in mutant 5 (~0.12 nM KD) resulted in only a 4-fold decrease in binding. As 10E8 interaction with the MPER peptide epitope is only with the heavy chain (Fig 1A), these results indicate that, except for mutant 4, all other light-chain mutants do not interfere with 10E8 binding to the MPER-scaffold.", "# Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands\nAuthors: Pozharski, E.; Deller, M. C.; Rupp, B.\nSource: Protein Crystallography: Methods and Protocols (A. Wlodawer, Z. Dauter & M. Jaskolski, eds.), Methods in Molecular Biology vol. 1607, Chapter 25, pp. 611\u2013625, Springer, 2017\n\n**Authors:** Edwin Pozharski, Marc C. Deller and Bernhard Rupp.\n\n**Published:** Chapter 25 of *Protein Crystallography: Methods and Protocols*, Methods in Molecular Biology volume 1607, pages 611\u2013625, Springer Science+Business Media, 2017.\n\n**Keywords:** Crystal structure; protein\u2013ligand complexes; structure model validation; small molecule ligands; peptide ligands.\n\n*This document is a structured summary of the chapter, not the full text.*\n\n## Abstract\n\nModels of target proteins in complex with small molecule ligands or peptide ligands are of significant interest to the biomedical research community. Structure-guided lead discovery and structure-based drug design make extensive use of such models. The bound ligands comprise only a small fraction of the total X-ray scattering mass, and therefore particular care must be taken to properly validate the atomic model of the ligand as experimental data can often be scarce. The ligand model must be validated against both the primary experimental data and the local environment, specifically: (1) the primary evidence in the form of the electron density, (2) examined for reasonable stereochemistry, and (3) the chemical plausibility of the binding interactions must be inspected. Tools that assist the researcher in the validation process are presented.\n\n## Section-by-section synopsis\n\n### 1. Introduction\n\nThe chapter opens by arguing that accurate atomic models of protein\u2013ligand complexes deposited in the Protein Data Bank underpin ligand docking and structure-based drug design, so rigorous validation is essential if those computational methods are to give meaningful results. The central technical problem is one of scale: bound ligands \u2014 small molecules, short peptides, DNA fragments and related analogues \u2014 have small mass relative to the protein target and therefore contribute proportionally little to the net X-ray scattering intensities. Global reciprocal-space statistics such as the frequently cited R-values consequently carry almost no ligand-specific information, and overall real-space measures such as r.m.s. deviations from target bond lengths and angles are quoted for the whole structure and do not reflect ligand quality either. Ligands are further afflicted by partial occupancy, enhanced mobility, conformational flexibility and poor protein stability, all of which further reduce their discrete scattering contribution. Careful local inspection using real-space quality indicators is therefore necessary, against three criteria: the fit of the ligand model to the electron density, compliance with prior expectations of reasonable stereochemistry, and plausible chemistry in the ligand\u2013protein binding site. Strict evidence-based validation has already revealed cases in which ligands purportedly bound to a target are insufficiently supported by the electron density; other evidence such as biochemical data, database annotations, functional assignments or protein fold may exist, but the pose must still be supported by density. The authors note that sometimes the density does not allow the pose to be determined even though some bound entity, often disordered, is evidently present, and that the community has no consensus on how such cases should be described at deposition \u2014 so readers should be alert to possible over-interpretation of ligands in deposited structures.\n\n### 2. Examination against the primary evidence\n\nThe primary evidence supporting any crystal structure model is the electron density, obtained by Fourier reconstruction from the diffraction data \u2014 tens to hundreds of thousands of reflection intensities \u2014 together with the phase angle of each reflection. The authors point out that experimental phases are usually not used for protein\u2013ligand structures; molecular replacement, in which an apo-protein structure is placed in the crystal, is the preferred method, traditionally followed by manual real-space adjustment on a graphics workstation and further cycles of reciprocal-space refinement against the diffraction data. The crucial consequence is that traditional maps produced by this route, such as 2mFo \u2212 DFc, are inherently biased toward the model that generated the phases. Difference electron density maps of the form mFo \u2212 DFc, by contrast, are almost always sufficiently clear to indicate both the presence and location of a bona fide ligand. The chapter therefore lays down the requirement that difference maps \u2014 specifically those calculated with the ligand omitted from the model \u2014 must be consulted during both model building and ligand validation.\n\n### 2.1. Displaying electron density\n\nPractical instructions follow for getting maps on screen. Display programs such as COOT and PyMOL can download everything needed for electron density reconstruction from repositories including the wwPDB, PDB_REDO and the Uppsala Electron Density Server (EDS). What is downloaded are appropriately weighted structure-factor amplitudes and phases for each observed reflection, which are displayed together with the atomic positions from the associated model file identified by its unique four-character PDB ID. The density itself is generally rendered as a three-dimensional contour map of density values scaled in standard deviations \u2014 sigma \u2014 from the mean density, so that contour levels quoted in the literature (1\u03c3, 2\u03c3, 3\u03c3) are directly comparable across structures. The standard 2mFo \u2212 DFc map is reconstructed from maximum-likelihood coefficients obtained during global reciprocal-space maximum-posterior refinement of the model against the diffraction data. This subsection establishes the working environment assumed by the rest of the chapter: COOT with coordinates and maps loaded, and the ability to switch between the standard map and difference maps.\n\n### 2.2. Real space correlation and real space R-value", "# Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\nAuthors: Ye, Q.; Shvartsbart, A.; Li, Z.; Gan, P.; Policarpo, R.L.; Qi, C.; Roach, J.J.; Zhu, W.; McCammant, M.S.; Hu, B.; Li, G.; Yin, H.; Carlsen, P.; Hoang, G.; Zhao, L.; Susick, R.; Zhang, F.; Lai, C.-T.; Allali Hassani, A.; Epling, L.B.; Gallion, A.; Kurzeja-Lipinski, K.; Gallagher, K.; Roman, V.; Farren, M.R.; Kong, W.; Deller, M.C.; Zhang, G.; Covington, M.; Diamond, S.; Kim, S.; Yao, W.; Sokolsky, A.; Wang, X.\nSource: Journal of Medicinal Chemistry, 2025\n\nAt the outset of our program, the feasibility of reversible inhibition of KRAS G12D was still in question. The only available starting points were a peptide developed by Takeda and a single example in a patent of a small molecule, compound 1 (Figure 1), related to an earlier G12C lead, where removal of the acrylamide warhead targeting cysteine showed <15 \u03bcM potency against KRAS G12D. Therefore, the first goal of the program was to determine if significant potency could be obtained in the absence of a covalent warhead. We elected to use 1 as a point of inquiry for our program.\n\nWhen tested in house in a biochemical assay measuring the exchange of a fluorescently labeled GDP analog for a nonhydrolyzable analog of GTP in the presence of SOS1, compound 1 showed an IC50 of 8 \u03bcM. Based on analysis of published KRAS G12C X-ray structures, we hypothesized that the piperazine nitrogen of 1 formed a salt bridge with Asp12, which was subsequently confirmed by researchers at Mirati. Initial SAR focused on optimization of this interaction while holding the rest of the molecule constant. We discovered that introducing a bridge to the piperazine ring (compound 2, Figure 2) significantly improved potency in the exchange assay, an observation made independently by a number of research groups. With compound 2 in hand, focus shifted to identifying a structurally distinct scaffold. We had reported tricyclic imidazole KRAS G12C inhibitors and hypothesized that a basic amine with a suitable conformational constraint would maintain the key salt bridge with Asp12 in this new scaffold. A preliminary screen of rigid amines identified compound 3 (Figure 2) as a promising hit with an exchange IC50 of 671 nM. Compound 3, like all compounds in this manuscript, contains a stable atropisomeric axis around the core-Switch II bond (highlighted in bold in compound 3, Figure 2). Unless otherwise stated, potency is reported for the more potent atropisomer.\n\n*Figure 2.* Initial optimization of hit compound 1 to lead compound 5.\n\nAs we pursued further SAR around compound 3, it became evident that the azabicyclo[2.1.1]hexane was critical for potency, as virtually all replacements were significantly less active. An X-ray structure of 3 bound to KRAS G12D was solved and suggested a possible explanation for this potency gain (Figure 3A). The protonated amine of the azabicyclo[2.1.1]hexane formed two key interactions with the protein \u2014 a direct hydrogen bond with Tyr96 and a water mediated interaction with Gly10. The rigid bicyclo[2.1.1] hexane structure serves to orient the amine in an otherwise difficult to access conformation directly below the plane of the imidazole core. Contrary to our initial hypothesis, no interactions were made with Asp12, suggesting that additional potency could be realized by growing toward this residue.\n\n*Figure 3.* X-ray crystal structure of (A) 3 and (B) 5a bound to KRAS-G12D (gray). Site of mutation Asp12 is highlighted in yellow. Key interactions for Tyr96 and Gly10 are highlighted with yellow dashed lines and distances (\u00c5). KRAS-G12D bound to 3 and 5a are deposited in the PDB under PDB ID: 9E5F and 9E5D, respectively.\n\nFurther examination of the X-ray structure identified the imidazole 2 position as a potential vector that could project a substituent toward Asp12 (see Figure 2 for numbering). A two carbon linker between the core and a suitable functionality proved an ideal length, with both shorter and longer chains detrimental to activity. A tertiary amide proved a particularly successful capping group for the linker. In addition, we found that changing chloro to methyl at the 8 position of the tricyclic core provided a \u223c2 fold improvement in potency. Overall, compound 4 showed a \u223c50 fold improvement in potency over 3 and also showed cellular potency of 1.3 \u03bcM in a homogeneous time-resolved fluorescence (HTRF) assay measuring the phosphorylation of ERK in the KRAS G12D dependent HPAFII cell line.\n\nTo further improve potency, we hoped to leverage the structural information available for the Takeda peptide. Published X-ray data showed the peptide binding to an extended conformation of KRAS, in which the Switch II loop opened to reveal a large pocket. Furthermore, multiple reports indicated that the Switch II loop was flexible, adapting a variety of conformations in crystal structures available at the time. We hypothesized that larger substitution in the Switch II region could help to stabilize the Switch II loop in an alternate conformation, opening up an extended pocket in which to gain significant additional potency.\n\nExtensive attempts to do so by substituting various positions on the 3-naphthol severely attenuated potency. The Switch II loop could alternatively be reached by substitution of the core, in particular at the 8 position. Encouraged by preliminary data suggesting that a methyl group showed improved potency relative to chloro (vide supra), we also undertook extensive SAR at this vector. Gratifyingly, compound 5 with cyanoethyl substitution increased potency to 2.9 nM, approaching the lower limit of the exchange assay. Using surface plasmon resonance (SPR) as a more sensitive assay to probe direct binding to KRAS showed that 5 was about 100-fold more potent than 4. In addition, 5 was the first compound to show potency in an HTRF pERK cell assay in the presence of human whole blood (hWB pERK) below 1 \u03bcM (IC50 = 977 nM).", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices\nAuthors: Deller, M.C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\n*Figure 6-10.* Figure showing summary of residues important for receptor binding in (a) LIF, (c) IL-6 and (e) CNTF. Structurally equivalent residues in OSM are shown in panels (b), (d) and (f). Residues involved in site 1 (blue) and site 3 (red) interactions with LIFR are shown along with residues involved in site 2 (yellow) interactions with gp130. Residues represent those deemed important from alanine scanning mutagenesis and the structurally equivalent residues in OSM were analysed using the program SHP (Stuart et al., 1979). Figure produced using BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994).\n\n#### 6.3.4 Erythropoietin\n\nThe Epo-EpoR structure (Syed et al., 1998) as shown in Figure 6-11 (a) below shows considerable homology to the GH-GHR system as discussed in Chapter 1 and also to the OSM and gp130 system as assessed by the RMSD values as shown in Table 6-1 and Table 6-2. Given the wealth of mutagenesis and structural data available, this system provides an ideal model for understanding OSM-gp130 interactions. However, the wider picture involving LIFR interactions will require structural studies of complexes involving the site 3 epitope; a feature that is unique to the LIFR family of ligands.\n\n##### 6.3.4.1 Site 1 Interactions\n\nSite 1 forms a central hydrophobic pocket involving Epo residues from helices A, B, D and part of the A-B loop. Epo residues Thr 44 and Asn 147 function in hydrogen bonding to Phe 93 of EpoR, which is the key binding determinant on the receptor. A hydrophobic quartet of Epo residues including Gly 151, Val 46, Phe 48 and Leu 155 form the centre of site 1 with the phenyl ring of the EpoR residue Phe 93 packing against the C\u03b1 of Epo residue Gly 151. The site 1 interface also has contributions from Epo residues Thr 44 to Phe 48 of the A-B loop with Phe 48 being in van der Waals contact with a minor determinant Met 150 of EpoR.\n\nOSM residues structurally equivalent to the site 1 hydrophobic pocket in Epo include Ser 179, Ser 58, Glu 60 and Arg 182, so suggesting that the hydrophobic nature of this motif is not conserved between the two. This is further borne out through comparison with LIF mutagenesis data (discussed in section 6.3.1.1) which maps site 1 to residues Arg 46, Glu 47, His 174, His 178 and Ser 179, so suggesting that this site is more hydrophilic in character than observed in Epo. This is not unexpected, as the major determinant of OSM interaction with LIFR is the Phe 160 and Lys 163 pair on the ligand, as identified through alanine mutagenesis (Deller et al., 2000). This site 3 motif provides most of the binding energy with site 1 playing only a minor role. Therefore, it could be predicted that the site 3 aromatic of the Phe/Lys pair functions in binding to a hydrophobic patch on the LIFR, in a reverse situation to that of the site 1 Epo-EpoR interaction (Middleton et al., 1999; Syed et al., 1998). Alternatively, the basic partner of the Phe/Lys pair could function in salt bridge formation with equivalents on the LIFR. This situation is observed in the crystal structure of IL-4-IL4R complex (Hage et al., 1999) in which Arg 88 of the ligand and Asp 72 of the receptor form a salt bridge with the bonding potential being enhanced by the surrounding hydrophobic environment.\n\nEquivalents to the Phe/Lys pair are not totally conserved outside the LIFR binding family of cytokines (see Figure 6-12). For example, IL-6 has a structurally and functionally equivalent Trp 157 and Asp 160 pair (Paonessa et al., 1995) and no such structural equivalents are observed for Epo. This is as would be expected, as the major binding determinant is the hydrophobic quartet of site 1 and site 3 interactions are redundant. GCSF shows a partially conserved Phe 144 and Lys 147 pair, so again suggesting homology to the gp130-LIFR family of ligands.\n\n##### 6.3.4.2 Site 2 Interactions\n\nSite 2 involves Epo residues located in the A and C helices although the interactions are less extensive than site 1 and most of the interactions involve residues of the C-helix. In a similar fashion to site 1 a hydrophobic surface is produced by Epo residues Leu 5, Val 11, Tyr 15, Ser 104, Thr 107 and Leu 108, which are all within 4.5 \u00c5 of the major binding determinant of EpoR, residue Phe 93 (Middleton et al., 1999; Syed et al., 1998). Epo residues Arg 10, Val 11 and Arg 14 are in van der Waals contact with the minor binding determinant Met 150 of EpoR. The EpoR residue Phe 93 is structurally and functionally equivalent to the gp130 residue Phe 169 located on the E-F loop of domain 1, whereas, the EpoR residues Met 150 and Phe 205, which have also been identified as essential binding determinants (Middleton et al., 1999) have no structural equivalents in gp130.\n\nStructurally equivalent OSM residues to the site 2 hydrophobic pocket in Epo include Leu 13, Gln 20, Met 24, Gly 120, Asn 123 and Asn 124, so suggesting that this motif is only partially conserved in hydrophobic character between the two molecules. Indeed, the side chains of the two main hydrophobic residues Leu 13 and Met 24 are buried in the OSM structure with relative solvent accessibilities of 2 and 3 %, respectively. Equivalent residues in Epo are Leu 5 and Tyr 15 with relative solvent accessibilities of 60 and 11 %, respectively. However, of these six residues alanine mutation of Gln 20, Gly 120, Asn 123 and Asn 124 has been shown to display up to 100-fold decrease on gp130 binding with little or no effect on LIFR binding (Deller et al., 2000). In addition a fifth member of the OSM site 2 motif has been identified as Gln 16. This residue is structurally equivalent to the Epo residue Asp 8 which hydrogen bonds to His 153 of EpoR, which is one of the 11 hydrogen bonds in the site 2 interface.", "# 2025 \u2013 Building an Adaptability Mindset in the Age of AI\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-28 on marcdeller.com (categories: certificates).*\n\n* [Link to certificate](https://www.linkedin.com/learning/certificates/8fd265595da5c6a811a98b77161dfbe5a887b53439fa7ca014db31e58ab8a0f4)\n\n* **Award Details:** LinkedIn Learning issued this certificate on March 18, 2025 at 05:00 PM UTC.\n* **Course Focus:** It recognizes completion of training on adaptability and critical thinking in an AI-driven environment.\n* **Career Value:** This matters because researchers increasingly need to adjust to changing tools, workflows, and technologies.\n* **Practical Use:** It helps scientists stay effective as AI reshapes analysis, modeling, and discovery work.", "# Extended Portfolio \u2014 Marc C. Deller, D.Phil.\nAuthors: Deller, M.C.\nSource: Personal portfolio (merged from two versions), 2025\n\n**Taglines:** Structural Biology | Drug Discovery | X-ray Crystallography | Cryo-EM; Associate Director | Drug Discovery | AI-driven Protein Design | Hit-to-Lead Innovation.\n\n**Contact:** (302) 358-6093 \u00b7 marc@marcdeller.com \u00b7 www.linkedin.com/in/marccdeller \u00b7 https://marcdeller.com \u00b7 Wilmington, United States.\n\nMarc has 60 publications in total with 3,631 citations (full list at Google Scholar: https://bit.ly/marcdeller_pubs). His protein structures are listed at https://bit.ly/marcdeller_proteins.\n\n## Patents\n\n### Anti-mutant Calreticulin (CALR) Antibodies and uses thereof\n\n**Year:** 2022. **Link:** https://patents.google.com/patent/WO2023107994A1/en\n\n### Bicyclic Compounds as Inhibitors of WRN\n\n**Status:** pending.\n\n## Latest Publications \u2014 with commentary\n\n### Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\n\n**Journal:** J. Med. Chem., 2025. **Authors:** Qinda et al. **Link:** https://pubs.acs.org/doi/abs/10.1021/acs.jmedchem.4c02662\n\nSupport for KRAS IND filing: Discovered an innovative, orally bioavailable inhibitor for the KRAS G12D mutation, which is known for its role in driving cancer. The objective is to effectively balance potency and ADME properties in order to tackle the challenges associated with this previously \"undruggable\" protein. This development aims to broaden treatment options beyond therapies specifically targeting the KRAS G12C mutation in lung cancers.\n\n### Discovery of (4-pyrazolyl)-2-aminopyrimidines as potent and selective Inhibitors of cyclin-dependent kinase 2\n\n**Journal:** J. Med. Chem., 2024. **Authors:** J.R. Hummel et al. **Link:** https://pubs.acs.org/doi/abs/10.1021/acs.jmedchem.3c02287\n\nSupport for CDK2 IND filing: Protein structure-based design of a first-in-class potent and highly selective CDK2 inhibitor from a novel chemical class, which demonstrates antitumor activity and addresses historical limitations of toxicity and poor selectivity that hindered the development of previous CDK2-targeted cancer therapies.\n\n### Discovery of potent and selective inhibitors of wild-type and gatekeeper mutant fibroblast growth factor receptor (FGFR) 2/3\n\n**Journal:** J. Med. Chem., 2022. **Authors:** Artem Shvartsbart et al. **Link:** https://pubs.acs.org/doi/abs/10.1021/acs.jmedchem.2c01366\n\nSupport for FGFR2/3 IND filing: Details the discovery of novel, selective FGFR2/3 inhibitors that address critical limitations of existing cancer treatments by maintaining efficacy against drug-resistant mutations and reducing side effects like hyperphosphatemia, thereby offering a promising therapeutic advancement for patients with cancers such as cholangiocarcinoma and bladder cancer.\n\n### Structure\u2013function analysis of the extended conformation of a polyketide synthase module\n\n**Journal:** Journal of the American Chemical Society, 2018. **Authors:** Xiuyuan Li et al. **Link:** https://pubs.acs.org/doi/abs/10.1021/jacs.8b02100\n\nGroundbreaking study: Resolves a long-standing debate in polyketide synthase (PKS) biology by demonstrating \u2014 via innovative antibody stabilization, structural validation (X-ray/SAXS), and kinetic assays \u2014 that the extended module conformation is catalytically functional for both chain elongation and modification, unlocking critical insights for engineered antibiotic production.\n\n### Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands\n\n**Venue:** Protein Crystallography: Methods and Protocols, 2017. **Authors:** Edwin Pozharski, Marc C. Deller, Bernhard Rupp. **Link:** https://link.springer.com/protocol/10.1007/978-1-4939-7000-1_25\n\nEmphasizes the need for high-quality protein structures in the age of predictive AI: Highlights the critical need for robust validation of protein\u2013ligand structural models, which are essential for drug discovery, by outlining specific criteria like electron density fit, stereochemistry, and binding plausibility, and introducing tools to assist researchers in this often challenging validation process due to scarce experimental data for ligands.\n\n### Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design\n\n**Journal:** PLoS Pathogens 13 (2), e1006212, 2017. **Authors:** A. Irimia, A.M. Serra, A. Sarkar, R. Jacak, O. Kalyuzhniy, D. Sok, \u2026 **Link:** https://journals.plos.org/plospathogens/article?id=10.1371/journal.ppat.1006212\n\nLandmark study: Redefines HIV-1 neutralization by revealing \u2014 via X-ray structures, mutagenesis, and modeling \u2014 that the broadly protective 10E8 antibody targets a hybrid epitope combining the gp41 MPER region and viral membrane lipids, with its light chain engaging lipids and positioning MPER perpendicularly to the membrane, offering transformative insights for designing lipid-informed vaccines/therapeutics against HIV.\n\n### Key gp120 glycans pose roadblocks to the rapid development of VRC01-class antibodies in an HIV-1-infected Chinese donor\n\n**Journal:** Immunity 44 (4), 939\u2013950, 2016. **Authors:** L. Kong, B. Ju, Y. Chen, L. He, L. Ren, J. Liu, K. Hong, B. Su, Z. Wang, \u2026 **Link:** https://www.cell.com/immunity/fulltext/S1074-7613(16)30095-4\n\nPivotal study: Bridges critical gaps in HIV vaccine design by uncovering \u2014 through structural analyses (X-ray/EM), longitudinal tracking, and functional assays \u2014 how early VRC01-class antibody precursors overcome glycan obstacles via light-chain adaptations, revealing both roadblocks (N276/V5 glycan clashes) and a rapid maturation pathway for broad neutralization, guiding next-gen immunogens targeting the CD4-binding site.\n\n### Complete epitopes for vaccine design derived from a crystal structure of the broadly neutralizing antibodies PGT128 and 8ANC195 in complex with an HIV-1 Env trimer\n\n**Journal:** Biological Crystallography 71 (10), 2099\u20132108, 2015. **Authors:** L. Kong, A. Torrents de la Pe\u00f1a, M.C. Deller, F. Garces, K. Sliepen, Y. Hua, \u2026 **Link:** https://journals.iucr.org/paper?S1399004715013917", "# GitHub README: CODSWALLOP\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/CODSWALLOP, 2026\n\n### Backlog (deliberately outside the four phases)", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nIn some embodiments, the present disclosure provides treatment of an individual or a patient in vivo using a compound of Formula (I) or a pharmaceutically acceptable salt thereof such that growth of cancerous tumors is inhibited. A compound of Formula (I) or of any of the formulas as described herein, or a compound as recited in any of the claims and described herein, or a pharmaceutically acceptable salt thereof, can be used to inhibit the growth of cancerous tumors. Alternatively, a compound of Formula (I) or of any of the formulas as described herein, or a compound as recited in any of the claims and described herein, or a pharmaceutically acceptable salt thereof, can be used in conjunction with other agents or standard cancer treatments, as described below. In one embodiment, the present disclosure provides a method for inhibiting growth of tumor cells in vitro. The method includes contacting the tumor cells in vitro with a compound of Formula (I) or of any of the formulas as described herein, or of a compound as recited in any of the claims and described herein, or a pharmaceutically acceptable salt thereof. In another embodiment, the present disclosure provides a method for inhibiting growth of tumor cells in an individual or a patient. The method includes administering to the individual or patient in need thereof of a therapeutically effective amount of a compound of Formula (I) or of any of the formulas as described herein, or of a compound as recited in any of the claims and described herein, or a pharmaceutically acceptable salt thereof.\n\nIn some embodiments, provided herein is a method for treating cancer. In some embodiments, the method comprises administering to a patient in need thereof, a therapeutically effective amount of a compound of Formula (I), or a salt thereof. In some embodiments, the cancer is selected from cancers whose growth may be inhibited using compounds of the disclosure. In some embodiments, the cancer is selected from cancers that are classified as microsatellite instability-high (MSI-H) cancers. In some embodiments, the cancer is selected from cancers that are classified as mismatch repair deficient (dMMR) cancers.\n\nExamples of cancers that are treatable using the compounds of the present disclosure include, but are not limited to, bone cancer, pancreatic cancer, skin cancer, cancer of the head or neck, cutaneous or intraocular malignant melanoma, uterine cancer, ovarian cancer, rectal cancer, cancer of the anal region, stomach cancer, testicular cancer, uterine cancer, carcinoma of the fallopian tubes, carcinoma of the endometrium, endometrial cancer, carcinoma of the cervix, carcinoma of the vagina, carcinoma of the vulva, Hodgkin's Disease, non-Hodgkin's lymphoma, cancer of the esophagus, cancer of the small intestine, cancer of the endocrine system, cancer of the thyroid gland, cancer of the parathyroid gland, cancer of the adrenal gland, sarcoma of soft tissue, cancer of the urethra, cancer of the penis, chronic or acute leukemias including acute myeloid leukemia, chronic myeloid leukemia, acute lymphoblastic leukemia, chronic lymphocytic leukemia, solid tumors of childhood, lymphocytic lymphoma, cancer of the bladder, cancer of the kidney or urethra, carcinoma of the renal pelvis, neoplasm of the central nervous system (CNS), primary CNS lymphoma, tumor angiogenesis, spinal axis tumor, brain stem glioma, pituitary adenoma, Kaposi's sarcoma, epidermoid cancer, squamous cell cancer, T-cell lymphoma, and environmentally induced cancers including those induced by asbestos. In some embodiments, the patient has not received previous treatments for the cancer. In some embodiments, the patient has not progress from previous treatments for the cancer. In some embodiments, the patient is intolerant to other therapies for the cancer.\n\nIn some embodiments, cancers treatable with compounds of the present disclosure include melanoma (e.g., metastatic malignant melanoma), renal cancer (e.g. clear cell carcinoma), prostate cancer (e.g. hormone refractory prostate adenocarcinoma), breast cancer, colon cancer, lung cancer (e.g. non-small cell lung cancer and small cell lung cancer), squamous cell head and neck cancer, urothelial cancer (e.g. bladder), and cancers with high microsatellite instability and/or mismatch repair deficient (MSI-H/dMMR). In some embodiments, the cancer being treated includes refractory or recurrent malignancies whose growth may be inhibited using the compounds of the present disclosure.\n\nIn some embodiments, cancers that are treatable using the compounds of the present disclosure include, but are not limited to, solid tumors (e.g., prostate cancer, colon cancer, esophageal cancer, endometrial cancer, ovarian cancer, uterine cancer, renal cancer, hepatic cancer, pancreatic cancer, gastric cancer, breast cancer, lung cancer, cancers of the head and neck, thyroid cancer, glioblastoma, sarcoma, bladder cancer, etc.), hematological cancers (e.g., lymphoma, leukemia such as acute lymphoblastic leukemia (ALL), acute myelogenous leukemia (AML), chronic lymphocytic leukemia (CLL), chronic myelogenous leukemia (CML), DLBCL, mantle cell lymphoma, Non-Hodgkin lymphoma (including relapsed or refractory NHL and recurrent follicular), Hodgkin lymphoma, or multiple myeloma.", "# Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly\nAuthors: Deller, M.C.; Johnson, H.A.; Miller, M.D.; Spraggon, G.; Elsliger, M.-A.; Wilson, I.A.; Lesley, S.A.\nSource: PLoS ONE, 2015\n\n*Figure 2.* Crystal structures of TM1088 proteins. (A) Ribbon representation of the high-resolution structure of TM1088A showing some of the conserved hydrophobic residues involved in dimer formation (Ile130, Val133 and Ile138 as red sticks). Bound AMP is displayed as yellow spheres. Sequence logo of the nucleotide binding motif region (Gly14/Gly16/Gly19) from structural alignment with other nucleotide binding RCK family members is shown inset (PDB ID: 1LSS and 1LSU). Each position of the sequence is represented by a stack of residues with the height of each stack representing the conservation at that position. Glycine and polar amino acids are colored green, basic amino acids are colored blue and hydrophobic amino acids are colored black [47]. (B) Ribbon representation of the TM1088A dimer of the octameric TM1088 assembly. Conserved hydrophobic residues involved in the TM1088A to TM1088B interface are shown as red sticks (Phe88, Met91, Ile114 and Phe115). The \"dimer hinge\" angle is denoted by a red arrow. (C) Ribbon representation of the TM1088B dimer from the octameric TM1088 assembly. Conserved hydrophobic residues involved in the TM1088A to TM1088B interface are shown as red sticks (Leu80, Phe81 and Met111). Sequence logo of the structurally equivalent region to the AMP binding site of TM1088A is shown inset. A structural alignment of non-nucleotide binding RCK family members (PDB ID: 1ID1 and 1LNQ) was used for the alignment. (D) Ribbon representation of the TM1088B C-terminal domain. Conserved hydrophobic residues involved in the TM1088B dimer interface are shown as red sticks (Ile140, Phe148, Ile174, Val179, Tyr200 and Ile202).\n\nThe octameric TM1088 complex has three main protein:protein interaction surfaces: 1) TM1088A:TM1088A dimerization interface, 2) TM1088B:TM1088B dimerization interface and 3) TM1088A:TM1088B interface (Table 2, conserved residues involved in these interactions are highlighted as red sticks in Fig 2A\u20132D). The TM1088A dimer is held together by the hydrophobic C-terminal portion of \u03b16 (Pro125-Ile138) of each monomer, which packs against portions of \u03b11 (Arg17-Ser28) and \u03b24 (Met75-Ala78), \u03b25 (Asn101-Arg105) and \u03b26 (Lys121-Ile123) of the adjacent subunit. Conserved residues central to TM1088A dimerization include Ile130, Val133 and Ile138 (red sticks in Fig 2A). The TM1088A dimer interface is predominantly hydrophobic with ~1600 \u00c5\u00b2 of total buried surface area involving 39 non-bonded contacts over 22 interface residues (Table 2).\n\n*Figure 3.* Nucleotide binding and other structural features of TM1088 proteins. (A) TM1088B inter-domain \u03b16 helix between Thr114 and Pro133. The refined 2Fo-Fc electron density map is contoured at 1\u03c3. (B) TM1088A nucleotide binding site. AMP is shown as yellow sticks in the high-resolution TM1088A structure. Conserved TM1088A residues interacting with AMP are displayed as green sticks (oxygen atoms are red, nitrogen blue, phosphate orange and sulfur yellow). Key hydrogen bonds are shown by gray dashed lines. A sigma-A weighted omit Fo-Fc electron density map of the bound AMP is contoured at 2.5\u03c3. (C) TM1088B interfacial hydrogen-bonding pattern. Residues involved in hydrogen bonding across the central cavity of the octameric TM1088 assembly (Arg33, Glu34 and Glu37) are shown as red sticks (oxygen atoms are colored yellow and nitrogen blue). Hydrogen bonds are shown as gray dashed lines. (D) Half of the interfacial hydrogen bonding pattern is shown (for clarity) with a refined 2Fo-Fc electron density map contoured at 1\u03c3.\n\n*Table 2.* Interactions in the three main protein interfaces of the octameric TM1088 assembly.\n\n| TM1088 interface | PDB chains | Interface residues of first component | Interface residues of second component | Interface area of first component (\u00c5\u00b2) | Interface area of second component (\u00c5\u00b2) | H-bonds in interface | Non-bonded contacts |\n|---|---|---|---|---|---|---|---|\n| A:A | BE* | 11 | 11 | 813 | 823 | \u2013 | 39 |\n| A:A | AF* | 11 | 10 | 770 | 797 | \u2013 | 34 |\n| B:B | CD* | 28 | 29 | 2099 | 2070 | 10 | 129 |\n| B:B | GH* | 24 | 24 | 1712 | 1756 | 7 | 100 |\n| B:B | DH | 3 | 2 | 104 | 132 | 2 | 23 |\n| B:B | CG | 1 | 1 | 56 | 60 | \u2013 | 4 |\n| A:B | BD* | 15 | 13 | 766 | 774 | 4 | 91 |\n| A:B | AC* | 14 | 13 | 761 | 749 | 3 | 64 |\n| A:B | EG* | 13 | 13 | 721 | 722 | 3 | 76 |\n| A:B | FH* | 11 | 11 | 705 | 701 | 2 | 52 |\n| A:B | FG | 2 | 3 | 118 | 111 | 1 | 14 |\n| A:B | EH | 2 | 3 | 112 | 103 | \u2013 | 18 |\n| A:B | AD | 1 | 3 | 109 | 100 | \u2013 | 11 |\n| A:B | BC | 1 | 3 | 109 | 100 | \u2013 | 13 |\n\n*Asterisk (bold in the original) denotes dimer pairs within the octameric TM1088 assembly. Data were calculated using PDBsum for entry 3l4b [48].\n\nSimilar regions of the TM1088B N-terminal domain are also involved in dimerization, specifically the structurally equivalent \u03b16 helix that forms the inter-domain linker. However, further interactions are also facilitated by the additional C-terminal Trk domain. These interactions predominantly involve conserved hydrophobic residues Ile140, Phe148, Ile174, Val179, Tyr200 and Ile202 that are all situated on the central \u03b2-sheet of the Rossmann-like fold (red sticks in Fig 2D). As a consequence of this extra dimerization domain, the surface buried in the TM1088B dimerization interface is significantly larger than that of the TM1088A dimer with over 4000 \u00c5\u00b2 of total buried surface area involving 129 non-bonded contacts over 57 residues. Additionally, the TM1088B dimerization interface involves 10 hydrogen bonds that are not observed in the TM1088A dimerization interface (Table 2) and are predominantly located at the interface between the N-terminal Trk domain (Arg17 and Ser18) and the inter-domain \u03b16 helix (Asp101, Thr124 and Glu128).", "# The Rise of the Bio-Hybrid: Merging Proteins with Electronics for Next-Gen Diagnostics\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-14 on marcdeller.com (categories: blog).*\n\n* Bio-hybrid technology combines the specific molecular recognition of engineered proteins with the signal processing power of electronic materials.\n* AI-driven protein engineering is used to design proteins that can act as the highly specific sensing element in these next-generation diagnostic devices.\n* This fusion of biology and electronics is paving the way for ultra-sensitive, real-time biosensors for applications in medicine, environmental monitoring, and beyond.\n\n1. **Protein Engineering Market Analysis 2025: Highlighting Advanced Design Tools:**\n   <https://www.thebusinessresearchcompany.com/market-insights/protein-engineering-market-insights-2025>)\n2. **Turbocharging Protein Engineering with AI:**\n   <https://cns.utexas.edu/news/features/turbocharging-protein-engineering-ai>\n3. **AI could accelerate protein engineering \u2013 key for developing new medicines:**\n   <https://sheffield.ac.uk/news/ai-could-accelerate-protein-engineering-key-developing-new-medicines>\n\nThe ability to rapidly and sensitively detect specific molecules is critical for everything from medical diagnostics to environmental monitoring. For years, we have relied on either purely biological tools, like antibody-based tests, or purely electronic sensors. Both have their limitations: biological tests can be slow and require complex reagents, while electronic sensors often lack the incredible specificity of biological recognition.\n\nA new frontier is emerging at the intersection of these two worlds: bio-hybrid technology. Researchers are now finding ways to directly interface engineered proteins with electronic materials like graphene or silicon nanowires. The idea is to use a custom-designed protein as the \u201cbrains\u201d of the sensor\u2014the part that specifically recognizes and binds to the target molecule (e.g., a virus particle or a cancer biomarker). This binding event then triggers a measurable electrical signal in the electronic component.\n\nAI-driven protein engineering is essential for creating the biological half of these devices. It allows us to design proteins that are not only highly specific for their target but are also stable when attached to an electronic surface and can induce a strong signal upon binding. This fusion of the molecular recognition power of biology with the signal amplification and processing power of electronics is leading to the development of ultra-sensitive, real-time biosensors. These devices could one day be integrated into wearable technology for continuous health monitoring or deployed in the field for instant environmental testing, creating a seamless link between the biological and digital worlds.", "# cif2plip \u2013 Protein-ligand interaction profiling using PLIP for co-folding outputs which produce CIF files\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-14 on marcdeller.com (categories: vibe-code).*\n\n> **CIF to strict PDB to PLIP. Trustworthy interaction profiling for co-folding outputs.**\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/cif2plip](https://github.com/bellcheddar/cif2plip) |\n\n---\n\n**CIF \u2192 strict PDB \u2192 PLIP protein-ligand interaction analysis**\n\nWhy it matters: co-folding tools such as Boltz-2, AlphaFold 3, and Chai-1 emit ModelCIF files that quietly violate the PDB format (multi-character chain IDs, oversized residue names, no bond table), and naive conversion introduces a subtle CONECT-parsing bug that corrupts ligand perception and invents fictitious interactions. cif_to_plip diagnoses and fixes every one of these issues before PLIP ever sees the structure, so the protein-ligand interactions you report are real. It is useful for anyone doing structure-based drug design on predicted complexes: reliable, reproducible interaction fingerprints (H-bonds, salt bridges, pi-stacking, halogen bonds, metal coordination) straight from a raw co-folding CIF.\n\nA single-file Python pipeline that converts a ModelCIF or mmCIF co-folding output into a format-compliant PDB file and feeds it directly into [PLIP](https://github.com/pharmai/plip) for full protein-ligand interaction profiling. Designed for downstream analysis of [Boltz-2](https://github.com/jwohlwend/boltz) and similar co-folding outputs where the input CIF carries no bond table and often contains multi-character chain IDs that break the PDB format specification.\n\n---\n\n## Why this script exists\n\nModern co-folding tools (Boltz-2, AlphaFold 3, Chai-1) emit ModelCIF files with:\n\n* **Multi-character chain IDs** (`JH1`, `J031`, \u2026) that exceed the single-character PDB column 22 limit\n* **Residue names longer than 3 characters** (`LIG1`, `SMOL`, \u2026) that overflow PDB columns 18\u201320\n* **No CONECT records** \u2014 bond topology is absent from the file\n\nNa\u00efvely converting such files with `gemmi` or `pdb_tidy` alone introduces a subtle but critical bug: `pdb_tidy` leaves a one-serial gap at each chain/TER boundary. Because OpenBabel maps CONECT records **positionally** by serial number, any gap causes records past that point to be silently discarded. The result is corrupted ligand perception \u2014 wrong SMILES, fictitious bonds in the PyMOL session, and invented PLIP interactions. `cif_to_plip` diagnoses and resolves all of these issues before PLIP ever sees the structure.\n\n---\n\n## Pipeline\n\n```\nCIF input\n   \u2502\n   \u251c\u2500 [1] gemmi parse + sanity check (single model enforced)\n   \u251c\u2500 [2] Remap multi-character chain IDs \u2192 single-character (A, B, C \u2026)\n   \u251c\u2500 [3] Rename >3-char residue names \u2192 3-char codes\n   \u251c\u2500 [4] Write strict PDB (element cols 77-78, TER records, occupancy/B padding)\n   \u251c\u2500 [5] pdb_tidy \u2192 renumber all serials contiguously (removes TER-boundary gap)\n   \u251c\u2500 [6] Generate intramolecular CONECT records from covalent-radius geometry\n   \u251c\u2500 [7] pdb_validate \u2014 format compliance check\n   \u2514\u2500 [8] PLIP \u2192 XML + TXT + PNG + PyMOL .pse\n              \u2514\u2500 [9] Flatten XML \u2192 interactions CSV + ligand summary CSV\n```\n\n---\n\n## Outputs\n\n| File | Description |\n| --- | --- |\n| `<stem>.pdb` | Standards-compliant PDB with contiguous serials and CONECT records |\n| `<stem>_plip/<stem>_report.xml` | Full PLIP XML report |\n| `<stem>_plip/<stem>_report.txt` | Human-readable PLIP text report |\n| `<stem>_plip/*.png` | Ray-traced binding-site images (one per ligand) |\n| `<stem>_plip/*.pse` | PyMOL session with PLIP interaction visualisation |\n| `<stem>_interactions.csv` | Tidy, flattened interactions table (one row per contact) |\n| `<stem>_ligand_summary.csv` | Per-ligand summary with SMILES, InChIKey, interaction counts |\n\n---\n\n## Installation\n\nConda is recommended to manage the mixed Python/OpenBabel/PyMOL stack:\n\n```\nconda create -n cif2plip python=3.11 -y\nconda activate cif2plip\n\n# Core dependencies\npip install gemmi pdb-tools plip\n\n# OpenBabel (required by PLIP for ligand perception)\nconda install -c conda-forge openbabel -y\n\n# Open-source PyMOL (required by PLIP for .pse session generation)\nconda install -c conda-forge pymol-open-source -y\n```\n\nVerify the install:\n\n```\npython -m plip.plipcmd --help\npdb_tidy --help\n```\n\n---\n\n## Usage\n\n```\npython cif_to_plip.py <input.cif> [-o <output_dir>] [--no-plip]\n```\n\n### Arguments\n\n| Argument | Description |\n| --- | --- |\n| `cif` | Input `.cif` / `.mmcif` file (required) |\n| `-o`, `--outdir` | Output directory (default: same directory as input) |\n| `--no-plip` | Convert to PDB only; skip PLIP analysis |\n\n### Examples\n\nRun the full pipeline:\n\n```\npython cif_to_plip.py boltz_output/jak1_lig.cif -o results/jak1/\n```\n\nConvert only (useful for format validation or feeding into other tools):\n\n```\npython cif_to_plip.py model.cif --no-plip\n```\n\n---\n\n## Interaction types reported\n\nPLIP detects and reports the following contact classes, all of which appear as distinct `interaction_type` values in the output CSV:\n\n* Hydrophobic contacts\n* Hydrogen bonds (donor/acceptor, with geometry)\n* Water bridges\n* Salt bridges\n* \u03c0\u2013stacking (face-to-face and edge-to-face)\n* \u03c0\u2013cation interactions\n* Halogen bonds\n* Metal coordination\n\n---\n\n## Technical notes\n\n### Chain ID remapping\n\nChain IDs are assigned in order of appearance using the 62-character alphabet `A\u2013Z`, `a\u2013z`, `0\u20139`. The mapping is printed to stdout at runtime.\n\n### Residue name truncation\n\nNames exceeding 3 characters are truncated to their first 3 characters. Collisions are resolved by assigning sequential codes `L01`, `L02`, \u2026 The mapping is printed to stdout.\n\n### CONECT generation\n\nBonds are perceived from covalent radii (Cordero et al., 2008) with a 0.45 \u00c5 tolerance. H\u2013H pairs are excluded. The bond table is written using strictly 80-column CONECT records and inserted immediately before the `END` record, after all serials have been made contiguous.", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\nThe TM1088 locus of T. maritima codes for two proteins designated TM1088A and TM1088B, which combine to form the cytosolic portion of a putative Trk K+ transporter. We report the crystal structure of this assembly to a resolution of 3.45 \u00c5. The high resolution crystal structures of the components of the assembly, TM1088A and TM1088B, were also determined independently to 1.50 \u00c5 and 1.55 \u00c5, respectively. The TM1088 proteins are structurally homologous to each other and to other K+ transporter proteins, such as TrkA. These proteins form a cytosolic gating ring assembly that controls the flow of K+ ions across the membrane. TM1088 represents the first structure of a two-subunit Trk assembly. Despite the atypical genetics and chain organization of the TM1088 assembly, it shares significant structural homology and an overall quaternary organization with other single-subunit K+ gating ring assemblies. This structure provides the first structural insights into what may be an evolutionary ancestor of more modern single-subunit K+ gating ring assemblies.\n\n## Models of protein-ligand crystal structures: trust, but verify (2015)\n\n**Authors:** Deller MC, Rupp B.  \n**Journal:** Journal of computer-aided molecular design (2015). DOI: 10.1007/s10822-015-9833-8. PMID: 25665575.\n\nX-ray crystallography provides the most accurate models of protein-ligand structures. These models serve as the foundation of many computational methods including structure prediction, molecular modelling, and structure-based drug design. The success of these computational methods ultimately depends on the quality of the underlying protein-ligand models. X-ray crystallography offers the unparalleled advantage of a clear mathematical formalism relating the experimental data to the protein-ligand model. In the case of X-ray crystallography, the primary experimental evidence is the electron density of the molecules forming the crystal. The first step in the generation of an accurate and precise crystallographic model is the interpretation of the electron density of the crystal, typically carried out by construction of an atomic model. The atomic model must then be validated for fit to the experimental electron density and also for agreement with prior expectations of stereochemistry. Stringent validation of protein-ligand models has become possible as a result of the mandatory deposition of primary diffraction data, and many computational tools are now available to aid in the validation process. Validation of protein-ligand complexes has revealed some instances of overenthusiastic interpretation of ligand density. Fundamental concepts and metrics of protein-ligand quality validation are discussed and we highlight software tools to assist in this process. It is essential that end users select high quality protein-ligand models for their computational and biological studies, and we provide an overview of how this can be achieved.\n\n## Structure-based discovery of NANOG variant with enhanced properties to promote self-renewal and reprogramming of pluripotent stem cells (2015)\n\n**Authors:** Hayashi Y, Caboni L, Das D, Yumoto F, Clayton T, Deller MC, Nguyen P, Farr CL, Chiu HJ, Miller MD, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Tomoda K, Conklin BR, Wilson IA, Yamanaka S, Fletterick RJ.  \n**Journal:** Proceedings of the National Academy of Sciences of the United States of America (2015). DOI: 10.1073/pnas.1502855112. PMID: 25825768.\n\nNANOG (from Irish mythology T\u00edr na n\u00d3g) transcription factor plays a central role in maintaining pluripotency, cooperating with OCT4 (also known as POU5F1 or OCT3/4), SOX2, and other pluripotency factors. Although the physiological roles of the NANOG protein have been extensively explored, biochemical and biophysical properties in relation to its structural analysis are poorly understood. Here we determined the crystal structure of the human NANOG homeodomain (hNANOG HD) bound to an OCT4 promoter DNA, which revealed amino acid residues involved in DNA recognition that are likely to be functionally important. We generated a series of hNANOG HD alanine substitution mutants based on the protein-DNA interaction and evolutionary conservation and determined their biological activities. Some mutant proteins were less stable, resulting in loss or decreased affinity for DNA binding. Overexpression of the orthologous mouse NANOG (mNANOG) mutants failed to maintain self-renewal of mouse embryonic stem cells without leukemia inhibitory factor. These results suggest that these residues are critical for NANOG transcriptional activity. Interestingly, one mutant, hNANOG L122A, conversely enhanced protein stability and DNA-binding affinity. The mNANOG L122A, when overexpressed in mouse embryonic stem cells, maintained their expression of self-renewal markers even when retinoic acid was added to forcibly drive differentiation. When overexpressed in epiblast stem cells or human induced pluripotent stem cells, the L122A mutants enhanced reprogramming into ground-state pluripotency. These findings demonstrate that structural and biophysical information on key transcriptional factors provides insights into the manipulation of stem cell behaviors and a framework for rational protein engineering.\n\n## Approaches to automated protein crystal harvesting (2014)\n\n**Authors:** Deller MC, Rupp B.  \n**Journal:** Acta crystallographica. Section F, Structural biology communications (2014). DOI: 10.1107/s2053230x14000387. PMID: 24637746.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices\nAuthors: Deller, M.C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\n*Figure 6-7.* Figure showing structural superpositions of human GCSFR (pink; Aritomi et al., 1999), human EpoR (red; Livnah et al., 1996), human GHR (green; de Vos et al., 1992) and human PrlR (blue; Somers et al., 1994) with gp130 (grey). The E-F loop located on domain 1 typically contains a solvent exposed aromatic residue important for ligand binding. Two other potential binding sites are located on the F-G and B-C loops of domain 2. Structural superpositions produced using SHP (Stuart et al., 1979) and figure produced using BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994).\n\n#### 6.2.3 P212121 and I222 Structures\n\nStructural superposition as shown in Table 6-1 and Figure 6-8 below reveals that the I222 crystal structure of OSM shows a high degree of structural similarity to the P212121 crystal structure of OSM. The two molecules display a RMSD in C\u03b1 positions of 0.72 \u00c5 over 147 structurally equivalent residues. Side chain conformations all lie in close agreement with the exception of residues in the flexible N-terminus preceding Leu 13 and large flexible solvent exposed arginine residues (e.g. Arg 50, 52, 63, 67 and 84) and lysine residues (e.g. Lys 94 and 110) which all show conformational flexibility between the two structures. Three clusters of residues grouped around Glu 89, Lys 110 and Tyr 126, also exhibit conformational flexibility about the \u03c71 angle. These clusters are located on helices B and C and residues from each are involved in crystal contacts in one or both of the structures as shown in Figure 6-9 below. It is also interesting to note that this region surrounding the B-C loop is structurally equivalent to site 4 of IL-6, which is postulated to be involved in IL-6 dimerisation as discussed in section 6.3.2.4. The conformational flexibility of side chains and the formation of crystal contacts in this region suggests OSM may also have a propensity for dimer formation; an observation which is supported by this thesis (see Chapter 3) and other studies (Staunton et al., 1998).\n\nAll helices and loops adopt similar positions in the two structures and major deviations between the two all lie in flexible regions at the end of helices and loop regions. Comparison of the B-factor values as shown below in Figure 6-8 shows that the two structures differ markedly in their flexibility. B-factors of the P212121 structure range from a minimum value of 6.3 \u00c5\u00b2 to a maximum value of 84.9 \u00c5\u00b2, with an average value of 30.5 \u00c5\u00b2. Whereas, the I222 structure displays higher B-factors ranging from a minimum value of 10.2 \u00c5\u00b2 to a maximum value of 84.4 \u00c5\u00b2, with a higher average value of 40.8 \u00c5\u00b2. The higher average B-factor observed in the I222 structure is presumably due to the fact that data was collected at room temperature (295 K) compared with the P212121 data that was collected under cryo-conditions (100 K). As discussed previously, higher than average B-factor values at the N- and C-terminus of helix C result in flexibility and hence disordered density for both the C-D and B-C loops of the I222 structure. Other areas of high mobility include the N-terminus preceding Leu 13 and the second helical section of the A-B loop running into helix B.\n\nFigure 6-9 below shows that the crystal contact points for the two structures also differ with eight main contact points in the I222 structure (8 molecules per unit cell) and four main contact points in the P212121 structure (4 molecules per unit cell). Panel (a) shows crystal contact points in the P212121 structure with residues shown in red corresponding to the N-terminal loop and the N-terminus of helix B and the C-terminus of helix C. Contacts shown in blue correspond to the second helical region at the C-terminus of the A-B loop and the C-terminus of helix D. Contacts shown in green lie on the N-terminus of helices A and C and the final contact site shown in yellow resides on the B-C loop and the first helical portion at the N-terminus of the A-B loop. Comparison with the crystal contact sites in the I222 structure shown in panel (b) reveals that no particular region is favoured for crystal contact and most exposed regions of loop and helix play a role.\n\n*Figure 6-8.* Figure showing (a) structural superpositions of P212121 OSM (grey) and I222 OSM (cyan) and B-factors of (b) P212121 and (c) I222 structures. C\u03b1 positions show 0.72 \u00c5 RMSD over 147 structurally equivalent residues. Structural superpositions were produced using SHP (Stuart et al., 1979). Panel (b) and (c) show comparison of refined B-factor values for crystal structures of OSM. In (b) B-factors are coloured from the minimum value of 6.3 \u00c5\u00b2 to the maximum value of 84.9 \u00c5\u00b2 from blue to red, respectively (average B-factor 30.5 \u00c5\u00b2). In (c) B-factors are coloured from the minimum value of 10.2 \u00c5\u00b2 to the maximum value of 84.4 \u00c5\u00b2 from blue to red, respectively (average B-factor 40.8 \u00c5\u00b2). Figure produced BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994).\n\n*Figure 6-9.* Figure showing crystal contact points observed for P212121 and I222 OSM. OSM molecules are shown with a solvent accessible contoured surface produced using the program VOLUMES (R. Esnouf, Oxford, U.K.). The four crystal contact points for the P212121 structure are shown as red, blue, green and yellow clusters on the protein surface. The eight crystal contact points for the I222 structure are shown as red, blue, green, yellow, pink, purple, orange and black clusters on the protein surface. Figure produced using BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994).\n\n### 6.3 Structure-Function Relationship Comparisons", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nExamples of skin cancers that are treatable using the compounds of the present disclosure include melanoma, basal cell carcinoma, squamous cell carcinoma, Kaposi's sarcoma, moles dysplastic nevi, lipoma, angioma, dermatofibroma, and keloids. In some embodiments, diseases and indications that are treatable using the compounds of the present disclosure include, but are not limited to, sickle cell disease (e.g., sickle cell anemia), triple-negative breast cancer (TNBC), myelodysplastic syndromes, testicular cancer, bile duct cancer, esophageal cancer, and urothelial carcinoma.\n\nThe terms \"individual\" or \"patient,\" used interchangeably, refer to any animal, including mammals, preferably mice, rats, other rodents, rabbits, dogs, cats, swine, cattle, sheep, horses, or primates, and most preferably humans.\n\nThe phrase \"therapeutically effective amount\" refers to the amount of active compound or pharmaceutical agent that elicits the biological or medicinal response in a tissue, system, animal, individual or human that is being sought by a researcher, veterinarian, medical doctor or other clinician.\n\nThe terms \"treating\" or \"treatment\" refers to one or more of (1) inhibiting the disease; e.g., inhibiting a disease, condition or disorder in an individual who is experiencing or displaying the pathology or symptomatology of the disease, condition or disorder (i.e., arresting further development of the pathology and/or symptomatology); and (2) ameliorating the disease; e.g., ameliorating a disease, condition or disorder in an individual who is experiencing or displaying the pathology or symptomatology of the disease, condition or disorder (i.e., reversing the pathology and/or symptomatology) such as decreasing the severity of disease.\n\nIn some embodiments, the compounds of the present disclosure are useful in preventing or reducing the risk of developing any of the diseases referred to herein; e.g., preventing or reducing the risk of developing a disease, condition or disorder in an individual who may be predisposed to the disease, condition or disorder but does not yet experience or display the pathology or symptomatology of the disease.\n\n### Formulation, Dosage Forms, and Administration\n\nWhen employed as pharmaceuticals, the compounds of the present disclosure can be administered in the form of pharmaceutical compositions. Thus, the present disclosure provides a composition comprising a compound of Formula (I) or any of the formulas as described herein, a compound as recited in any of the claims and described herein, or a pharmaceutically acceptable salt thereof, or any of the embodiments thereof, and at least one pharmaceutically acceptable carrier or excipient. These compositions can be prepared in a manner well known in the pharmaceutical art, and can be administered by a variety of routes, depending upon whether local or systemic treatment is indicated and upon the area to be treated. Administration may be topical (including transdermal, epidermal, ophthalmic and to mucous membranes including intranasal, vaginal and rectal delivery), pulmonary (e.g., by inhalation or insufflation of powders or aerosols, including by nebulizer; intratracheal or intranasal), oral or parenteral. Parenteral administration includes intravenous, intraarterial, subcutaneous, intraperitoneal intramuscular or injection or infusion; or intracranial, e.g., intrathecal or intraventricular, administration. Parenteral administration can be in the form of a single bolus dose, or may be, e.g., by a continuous perfusion pump. Pharmaceutical compositions and formulations for topical administration may include transdermal patches, ointments, lotions, creams, gels, drops, suppositories, sprays, liquids and powders. Conventional pharmaceutical carriers, aqueous, powder or oily bases, thickeners and the like may be necessary or desirable.\n\nThe present disclosure also includes pharmaceutical compositions which contain, as the active ingredient, the compound of the present disclosure or a pharmaceutically acceptable salt thereof, in combination with one or more pharmaceutically acceptable carriers or excipients. In some embodiments, the composition is suitable for topical administration. In making the compositions of the present disclosure, the active ingredient is typically mixed with an excipient, diluted by an excipient or enclosed within such a carrier in the form of, e.g., a capsule, sachet, paper, or other container. When the excipient serves as a diluent, it can be a solid, semi-solid, or liquid material, which acts as a vehicle, carrier or medium for the active ingredient. Thus, the compositions can be in the form of tablets, pills, powders, lozenges, sachets, cachets, elixirs, suspensions, emulsions, solutions, syrups, aerosols (as a solid or in a liquid medium), ointments containing, e.g., up to 10% by weight of the active compound, soft and hard gelatin capsules, suppositories, sterile injectable solutions, and sterile packaged powders.\n\nIn preparing a formulation, the active compound can be milled to provide the appropriate particle size prior to combining with the other ingredients. If the active compound is substantially insoluble, it can be milled to a particle size of less than 200 mesh. If the active compound is substantially water soluble, the particle size can be adjusted by milling to provide a substantially uniform distribution in the formulation, e.g., about 40 mesh. The compounds of the present disclosure may be milled using known milling procedures such as wet milling to obtain a particle size appropriate for tablet formation and for other formulation types. Finely divided (nanoparticulate) preparations of the compounds of the present disclosure can be prepared by processes known in the art see, e.g., WO 2002/000196.", "# 2024 \u2013 Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective CDK2 Inhibitors\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: papers).*\n\nStructures\n\nDiscovery of (4-Pyrazolyl)-2-Aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase\n\n* [Link to Paper](https://marcdeller.com/wp-content/uploads/2025/09/hummel-et-al-2024-discovery-of-4-pyrazolyl-2-aminopyrimidines-as-potent-and-selective-inhibitors-of-cyclin-dependent.pdf)\n\n* **Selective Targeting:** (4-Pyrazolyl)-2-aminopyrimidines achieve unprecedented selectivity for CDK2 over other CDK isoforms through optimized hinge region interactions\n* **Potent Activity:** Lead compound 17 demonstrates sub-nanomolar potency (IC\u2085\u2080 = 0.29 nM) with robust selectivity profile across the CDK family\n* **Therapeutic Vulnerability:** CCNE1-amplified cancers show preferential sensitivity to CDK2 inhibition due to oncogene addiction mechanisms\n* **Mechanism of Action:** CDK2 inhibition blocks Rb phosphorylation, inducing G1/S arrest and senescence specifically in cyclin E1-overexpressing tumors\n* **Combination Potential:** Synergistic activity with DNA-damaging agents through disruption of homologous recombination repair pathways\n* **Clinical Translation:** Multiple selective CDK2 inhibitors have entered clinical trials based on this breakthrough discovery\n\n1. [Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2: Hummel, J.R., et.al., Journal of Medicinal Chemistry](https://pubs.acs.org/doi/abs/10.1021/acs.jmedchem.3c02287)\n2. [INX-315, a Selective CDK2 Inhibitor, Induces Cell Cycle Arrest and Senescence in Solid Tumors: Dietrich, C., et.al., Cancer Discovery](https://incyclixbio.com/wp-content/uploads/2024/05/Cancer-Discovery-2024-INX-315.pdf)\n3. [Targeting CDK2 and other novel cell cycle targets for breast cancer therapy: Anonymous, et.al., Expert Opinion on Drug Discovery](https://www.tandfonline.com/doi/full/10.1080/14728222.2025.253741)\n4. [Profiling the Activity of the Potent and Highly Selective CDK2 Inhibitor BLU-222: Anonymous, et.al., Cancer Research](https://aacrjournals.org/cancerres/article/85/7/1297/754355/Profiling-the-Activity-of-the-Potent-and-Highly)\n5. [CDK2 regulates collapsed replication fork repair in CCNE1-amplified ovarian cancer cells: Brown, V.E., et.al., NAR Cancer](https://pubmed.ncbi.nlm.nih.gov/37519629/)\n6. [Inhibitors and PROTACs of CDK2: challenges and opportunities: Zeng, Y., et.al., Expert Opinion on Drug Discovery](https://pubmed.ncbi.nlm.nih.gov/38994606/)\n\nThe quest for selective cyclin-dependent kinase 2 (CDK2) inhibitors has reached a significant milestone with the discovery of (4-pyrazolyl)-2-aminopyrimidines, representing a new class of potent anticancer compounds that successfully address the longstanding challenge of kinase selectivity. This breakthrough addresses a critical unmet need in oncology, where CDK2 dysregulation drives tumor progression in CCNE1-amplified cancers, yet previous inhibitors suffered from poor selectivity and associated toxicity.\n\nCDK2 plays a pivotal role in cell cycle regulation, partnering with cyclin E1 to control the G1/S transition. The dysregulation of this CDK2/cyclin E1 complex serves as a key oncogenic driver in various human cancers, particularly high-grade serous ovarian cancer, endometrial cancer, and certain breast cancers where CCNE1 amplification occurs in 15-20% of cases. These amplifications create a therapeutic vulnerability, as cancer cells become dependent on CDK2 activity for survival while normal cells can compensate through alternative pathways.\n\nThe structure-activity relationship (SAR) studies revealed critical insights into the molecular determinants of CDK2 selectivity. The (4-pyrazolyl)-2-aminopyrimidine scaffold achieves selectivity through specific interactions with the CDK2 active site, particularly in the hinge region where the 2-aminopyrimidine motif forms crucial hydrogen bonds with Leu83. The pyrazole ring at the 4-position occupies the ribose binding pocket, while substituent modifications fine-tune selectivity over other CDK isoforms. Compound 17, the lead molecule from this series, demonstrates exceptional potency with an IC\u2085\u2080 of 0.29 nM against CDK2 while maintaining selectivity over CDKs 1, 4, 6, 7, and 9.\n\nThe mechanism of action involves competitive inhibition of ATP binding, leading to reduced retinoblastoma (Rb) protein phosphorylation. In CCNE1-amplified cancer models, CDK2 inhibition results in G1/S cell cycle arrest and induction of therapy-induced senescence. This selectivity profile represents a significant advancement over earlier pan-CDK inhibitors like dinaciclib, which caused dose-limiting toxicities due to simultaneous inhibition of multiple CDK isoforms.\n\nThe therapeutic potential extends beyond single-agent activity, with emerging evidence supporting combination strategies. In CCNE1-amplified tumors, CDK2 inhibitors sensitize cells to DNA-damaging agents by disrupting homologous recombination repair mechanisms. This synthetic lethality approach offers particular promise for ovarian cancers with intact homologous recombination, which are typically resistant to PARP inhibitors. The combination of CDK2 inhibition with chemotherapy converts static responses to tumor regression, even in chemoresistant models.\n\nClinical translation of this discovery is already underway, with several selective CDK2 inhibitors entering clinical trials. Beyond the Incyte compound, related selective CDK2 inhibitors including INX-315, BLU-222, and PF-07104091 are being evaluated in patients with CCNE1-amplified cancers. These clinical developments validate the therapeutic hypothesis that selective CDK2 inhibition can provide efficacy while avoiding the toxicities associated with broader CDK inhibition.", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nThe compositions administered to a patient can be in the form of pharmaceutical compositions described above. These compositions can be sterilized by conventional sterilization techniques or may be sterile filtered. Aqueous solutions can be packaged for use as is, or lyophilized, the lyophilized preparation being combined with a sterile aqueous carrier prior to administration. The pH of the compound preparations typically will be between 3 and 11, more preferably from 5 to 9 and most preferably from 7 to 8. It will be understood that use of certain of the foregoing excipients, carriers or stabilizers will result in the formation of pharmaceutical salts.\n\n### Patient Screening\n\nIn some embodiments, therapeutic applications of the WRN inhibition compounds of the present disclosure can be accompanied by one or more screening tests, including patient screening tests according to clinical practice guidance. In some embodiments, one or more screening test can be administered to determine if the cancer is a microsatellite instability-high (MSI-H) cancer. Polymerase chain reaction (PCR) and immunohistochemistry (IHC) assays can be used to classify if a cancer is MSI.\n\nIn some embodiments, a polymerase chain reaction (PCR) assay is used to determine if the cancer is a microsatellite instability-high (MSI-H) cancer. The PCR based assays can follow the NCI-Reference Panel and/or the Promega MSI Analysis System. The NCI-Reference Panel (also referred to as Bethesda Panel, Ref 3) includes five loci, three dinucleotide sites (D2S123, D5S346 and D17S250) and two mononucleotide sites (BAT-25 and BAT-26). The Promega MSI Analysis System (such as OncoMate\u2122 kit) includes five mononucleotide repeat markers: BAT-25, BAT-26, NR-21, NR-24 and MONO\u201427. A cancer can be classified as MSI if a downward shift in size is detected in at least two of the five genes, compared to normal tissue samples.\n\nIn some embodiments, an immunohistochemistry (IHC) assay is used to determine if the cancer is a microsatellite instability-high (MSI-H) cancer. The IHC assay can be used to detect loss of protein expression of mismatch repair genes, such as MLH1, MSH2, MSH6 and PMS2. These genes are also associated with hereditary nonpolyposis colorectal cancer (HNPCC), Lynch syndrome, and the constitutional MMR deficiency (CMMRD) syndrome. In some embodiments, a loss of expression with at least one of the four proteins, compared to normal tissue samples, can classify a cancer as MSI.\n\nIn some embodiments, one or more screening test can be administered to determine if the cancer is a mismatch repair deficient (dMMR) cancer.\n\n### Combination Therapies\n\nIn some embodiments, the WRN inhibitor compounds of the present disclosure are administered as a single agent. In some embodiments, the WRN inhibitor compound is administered in combination with an additional active agent or treatment. In some embodiments, the WRN inhibitor compound is administered in combination with an immunotherapy, a targeted therapy, a chemotherapy, or radiation therapy. In some embodiments, the active agents of the combination therapy have different and complementary mechanisms of action. In some embodiments, the active agents of the combination therapy have a synergistic and/or more durable therapeutic effect on a target disease, disorder and condition. In some embodiments, the combination of active agents allows for a dose adjustment of one or more of the active agents, which can reduce adverse effects associated with one or more of the active agents. In some embodiments, the active agents can be co-administered in a single composition; in some embodiments, the active agents can be co-administered separately; in some embodiments, the active agents can be co-administered separately and simultaneously; in some embodiments, the active agents can be administered separately and sequentially.\n\n**Combination with immune-checkpoint therapies.** In some embodiments, WRN inhibitor compounds of the present disclosure are used in combination with one or more immune checkpoint inhibitors for the treatment of diseases, such as cancer or infections. Examples of immune checkpoint inhibitors include inhibitors against immune checkpoint molecules such as CBL-B, CD20, CD28, CD40, CD122, CD96, CD73, CD47, CDK2, GITR, CSF1R, JAK, PI3K delta, PI3K gamma, TAM, arginase, HPK1, CD137 (also known as 4-1BB), ICOS, A2AR, B7-H3, B7-H4, BTLA, CTLA-4, LAG3, TIM3, TIGIT, TLR (TLR7/8), CD112R, VISTA, PD-1, PD-L1 and PD-L2. In some embodiments, the immune checkpoint inhibitor is a stimulatory checkpoint molecule selected from CD27, CD28, CD40, ICOS, OX40, GITR and CD137. In some embodiments, the immune checkpoint inhibitor is an inhibitory checkpoint molecule selected from A2AR, B7-H3, B7-H4, BTLA, CTLA-4, IDO, KIR, LAG3, PD-1, TIM3, TIGIT, and VISTA. In some embodiments, the WRN inhibitor compounds are used in combination with one or more agents selected from KIR inhibitors, TIGIT inhibitors, LAIR1 inhibitors, CD160 inhibitors, 2B4 inhibitors, and TGFR beta inhibitors.\n\nIn some embodiments, the compounds provided herein can be used in combination with one or more agonists of immune checkpoint molecules, e.g., OX40, CD27, GITR, and CD137 (also known as 4-1BB).", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\nMutant CALR can be detected in a biological sample of a subject who has a myeloproliferative neoplasm. Thus an aspect of the present disclosure provides a method of detecting a CALR exon 9 mutation in a subject's biological sample, the method comprising obtaining a biological sample from a subject who has or is suspected of having a myeloproliferative neoplasm and contacting the sample with an anti-mutCALR antibody described herein such that the anti-mutCALR antibody binds to the mutCALR protein if the mutCALR protein is present in the biological sample. Another aspect provides a method of diagnosing a subject with a myeloproliferative neoplasm by the same approach.\n\nThe biological sample can be a blood sample, a bone marrow sample or a serum sample. In some embodiments, the biological sample is fresh or frozen. In some embodiments, the biological sample is fixed, for example in formaldehyde or paraformaldehyde.\n\nAn anti-mutCALR antibody described herein for use in the present detection or diagnostic methods can also comprise (e.g., be conjugated to) a detectable label. Suitable detectable labels include a radioisotope, a nanoparticle, a fluorescent compound, a bioluminescent compound, chemiluminescent compound, a metal chelator, magnetic beads, metallic beads, colloidal particles, fluorescent dyes, electron-dense reagents, enzymes (for example, as used in an ELISA), biotin, digoxigenin, or haptens. An anti-mutCALR antibody described herein bound to mutCALR protein, whether labeled or unlabeled, can be detected by any suitable detection method, including immunologic techniques such as immunohistochemistry (IHC), immunocytochemistry, Western blot, or ELISA immunoassay; gel- or blot-based methods; mass spectrometry; flow cytometry; or fluorescent activated cell sorting (FACS).\n\nThe disclosure also provides a kit for the diagnosis of a myeloproliferative neoplasm comprising one or more containers of an anti-mutCALR antibody described herein or a diagnostic formulation thereof, and optionally with instructions for using the anti-mutCALR antibody or a diagnostic formulation thereof to detect a mutCALR exon 9 mutation or diagnose a myeloproliferative neoplasm.\n\n## Examples\n\nThe following examples are provided to better illustrate the claimed invention and are not to be interpreted as limiting the scope of the invention. To the extent that specific materials are mentioned, it is merely for purposes of illustration and is not intended to limit the invention. One skilled in the art can develop equivalent means or reactants without the exercise of inventive capacity and without departing from the scope of the invention.\n\n### Example 1: Generation of Anti-mutCALR Antibodies\n\nTo generate anti-human mutCALR monoclonal antibodies, mice were immunized with one plasmid encoding human MPL and a second plasmid encoding human Type 1 mutCalR. The sequences for mutant CalR and MPL were cloned into the pVAC2 expression plasmid (Invivogen). The nucleotide sequences used in the vectors as well as the sequences of the encoded proteins are given in the specification (human MPL nucleotide sequence, SEQ ID NO:323; human MPL amino acid sequence, SEQ ID NO:324; human Type 1 mutCALR nucleotide sequence, SEQ ID NO:325; human Type 1 mutCALR amino acid sequence, SEQ ID NO:320).\n\nPlasmablasts were isolated from the mouse spleen and lymph nodes by flow cytometry. Antibody sequences of the plasmablasts were determined using 10x Genomics VH/VL paired B cell sequencing. The murine VH/VL pairs were expressed as chimeras with huIgG1 Fc and tested for binding and functionality. An antibody designated clone 54 was produced by this process.\n\nIn addition, multiple selection rounds of single-donor and multi-donor phage display libraries were performed according to the method of Erasmus et al., *Communications Biology* 4, article no. 350 (2021). Phage libraries were enriched for either 2 or 3 rounds on biotinylated MBP-mutCalR Type 1 fusion protein (Cepter Biopartners) with deselection using MBP (Rockland Immunochemicals) and streptavidin beads (Thermo Fisher) each round (see, e.g., Example 2). scFv cassettes from the library pools showing the strongest specific enrichment for mutCalR were then recombined into a yeast display vector to create yeast display libraries. These were selected over four rounds using decreasing concentrations of MBP-mutCalR with and without blocking with either a scrambled peptide or oligo-lysine peptide (Alamanda Polymers). In the final round of selection, yeast were also selected for binding to a biotinylated mutCalR long peptide (Biot-LC-MKDKQDEEQRTRRMMRTKMRMRRMRRTRRKMRRKMSPARPRTSSREASLQGWTEA; SEQ ID NO:332) or short peptide (Biot-LC-MKDKQDEEQRTRRMMRTKMRMRRMRRTRRKM; SEQ ID NO:333). Unique sequences were obtained from the final sorting output by Sanger sequencing of yeast colonies. For screening, these scFv were reformatted by cloning into a human IgG1 expression vector, then expressed and purified from Expi293F cell (Thermo Fisher, cat. #A14635) supernatants. The MBP-mutCalR Type 1 fusion protein used in this experiment is given as SEQ ID NO:334.\n\nThe amino acid sequences of the six CDRs for each of the 54 unique clones are shown in Table 1. The heavy chain, VH, light chain, and VL sequences of each clone are shown in Table 4.", "# AnalyzeBoltz \u2013 Harvest the sprawling Boltz2 output tree, tidy summary, CSV, and CIF output folder\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-14 on marcdeller.com (categories: vibe-code).*\n\n> **Harvest a sprawling Boltz-2 output tree into one tidy summary, CSV, and CIF folder.**\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/analyze-boltz2-results](https://github.com/bellcheddar/analyze-boltz2-results) |\n\n---\n\nA post-run analysis script for [Boltz-2](https://github.com/jwohlwend/boltz) co-folding predictions. Run it in any directory containing a `boltz_run_*.log` file and a `predictions/` subdirectory to automatically harvest confidence scores, affinity predictions, CIF models, and run metadata into a clean summary report and CSV.\n\nWhy it matters: a Boltz-2 batch leaves behind a deep, scattered output tree (per-target directories, confidence JSONs, affinity JSONs, ModelCIFs, and a verbose log) that is painful to read by hand and impossible to rank at a glance. analyze_boltz.py walks the whole tree, extracts every useful number, derives a pIC50 estimate, and consolidates everything into a human-readable summary, a machine-readable CSV, and a flat folder of all model CIFs. It is useful for anyone screening many complexes: instant triage and ranking of predictions by confidence and predicted affinity, with robust handling of missing or malformed files so no result is silently dropped.\n\n---\n\n## Overview\n\nBoltz-2 produces a sprawling output tree: per-target prediction directories, confidence JSONs, affinity JSONs, and ModelCIF files, alongside a verbose run log. `analyze_boltz.py` walks that tree, extracts everything useful, and consolidates it into three outputs:\n\n| Output | Description |\n| --- | --- |\n| `boltz_run_summary.log` | Human-readable summary: invocation, directory tree, runtime, per-prediction scores |\n| `boltz_summary.csv` | Machine-readable table of all predictions with confidence, affinity, and pIC\u2085\u2080 |\n| `boltz_cif/` | Flat directory of all `*_model_0.cif` files, ready for downstream structure analysis |\n\n---\n\n## Requirements\n\n* Python 3.10+ (uses `Path | None` union syntax)\n* Standard library only \u2014 no external dependencies\n\n---\n\n## Usage\n\nCopy `analyze_boltz.py` into, or run it from, the top-level directory of your Boltz-2 run output:\n\n```\ncd /path/to/your/boltz_run_dir\npython analyze_boltz.py\n```\n\nThe script resolves everything relative to the current working directory. No arguments are required.\n\n### Expected directory layout\n\n```\nboltz_run_dir/\n\u251c\u2500\u2500 boltz_run_20260101_120000.log     \u2190 required: matches boltz_run_*.log\n\u251c\u2500\u2500 predictions/                      \u2190 required: can be nested\n\u2502   \u251c\u2500\u2500 target_001/\n\u2502   \u2502   \u251c\u2500\u2500 target_001_model_0.cif\n\u2502   \u2502   \u251c\u2500\u2500 confidence_target_001_model_0.json\n\u2502   \u2502   \u2514\u2500\u2500 affinity_target_001_model_0.json\n\u2502   \u2514\u2500\u2500 target_002/\n\u2502       \u2514\u2500\u2500 ...\n\u2514\u2500\u2500 analyze_boltz.py\n```\n\n`predictions/` can be located anywhere below the top-level directory \u2014 the script walks the full subtree to find it.\n\n---\n\n## Outputs\n\n### `boltz_run_summary.log`\n\nSections:\n\n1. **Zsh execution script and switches** \u2014 reconstructed boltz invocation line(s) from the log\n2. **Directory structure** \u2014 annotated tree up to 3 levels deep\n3. **Run timing** \u2014 start/finish timestamps and total wall-clock runtime\n4. **Prediction summaries** \u2014 per-target: CIF name, confidence score, affinity values, pIC\u2085\u2080\n5. **CIF copy operations** \u2014 log of every file copied into `boltz_cif/`\n\n### `boltz_summary.csv`\n\nOne row per prediction subdirectory:\n\n| Column | Description |\n| --- | --- |\n| `input YAML name` | Source YAML inferred from log or directory name |\n| `output cif name` | Filename of the `*_model_0.cif` model |\n| `confidence_score` | Boltz-2 overall confidence score (0\u20131) |\n| `affinity_pred_value1` | Raw affinity prediction (log-scale, kcal/mol units) |\n| `affinity_probability_binary1` | Probability of binding (binary classifier output) |\n| `pIC\u2085\u2080` | Derived pIC\u2085\u2080: `(6 \u2212 affinity_pred_value1) \u00d7 1.364` |\n\n### `boltz_cif/`\n\nFlat collection of all `*_model_0.cif` files discovered across the predictions tree. Suitable for bulk loading into PyMOL, ChimeraX, or a downstream PLIP/interaction analysis pipeline.\n\n---\n\n## pIC\u2085\u2080 Derivation\n\nThe affinity output from Boltz-2 (`affinity_pred_value1`) is a log-transformed binding free energy proxy. The conversion applied here is:\n\n```\npIC\u2085\u2080 = (6 \u2212 affinity_pred_value1) \u00d7 1.364\n```\n\nThis is a heuristic approximation; treat values as relative rankings rather than absolute potency predictions.\n\n---\n\n## Notes\n\n* **YAML association**: prediction directories are matched to input YAMLs by stem name. If no match is found, the script falls back to the first YAML detected in the log, then derives a name from the directory itself \u2014 no prediction is silently dropped.\n* **Robustness**: the script tolerates missing confidence or affinity JSONs, absent CIF files, and non-standard timestamp formats. Failures are logged per-prediction rather than aborting the run.\n* **Log parsing**: the invocation heuristic looks for lines containing both `boltz` and `python`/`zsh`. If none are found, it falls back to `Running structure prediction` / `Running affinity prediction` lines.\n\n---\n\n## Compatibility\n\nTested against Boltz-2 output structure as of early 2026. The script assumes:\n\u2013 Log filename matches `boltz_run_*.log`\n\u2013 Confidence JSONs match `confidence*_model_0.json`\n\u2013 Affinity JSONs match `affinity*.json`\n\u2013 Model CIFs match `*_model_0.cif`\n\n---\n\n## Licence\n\nMIT\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/analyze-boltz2-results](https://github.com/bellcheddar/analyze-boltz2-results) |", "# Unlocking Molecular Potential: The Evolution of Topliss Tree and Bioisosteric Strategies in Modern Drug Discovery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe strategic deployment of bioisosteric replacements has proven invaluable across multiple dimensions of drug optimization. Modern applications extend far beyond simple activity preservation to encompass improvements in selectivity, metabolic stability, solubility, and safety profiles. The replacement of hydrogen with deuterium, exemplified in drugs like deutetrabenazine, represents a sophisticated approach to enhancing pharmacokinetic properties through isotopic effects. Similarly, the substitution of carboxylic acids with tetrazole rings in angiotensin receptor blockers like losartan demonstrates how bioisosteric modifications can address bioavailability challenges while maintaining receptor binding affinity.\n\nContemporary structure-based drug design has revolutionized the application of both Topliss principles and bioisosteric strategies through the integration of high-resolution structural data and computational modeling. The availability of protein crystal structures, complemented by artificial intelligence-driven predictions such as those from AlphaFold 3, has enabled precise rational design approaches that were unimaginable in Topliss\u2019s era. These tools allow medicinal chemists to visualize molecular interactions at atomic resolution, predicting how structural modifications will affect binding affinity, selectivity, and off-target effects.\n\nThe emergence of matched molecular pair analysis represents a natural evolution of the systematic thinking embodied in the Topliss tree. This computational approach identifies pairs of compounds that differ by single structural transformations, enabling quantitative assessment of how specific changes affect various molecular properties. Unlike the prospective nature of the Topliss tree, matched molecular pair analysis leverages retrospective data analysis to extract design principles from large chemical databases, providing statistical validation for structural modification strategies.\n\nFragment-based drug discovery has introduced another dimension to systematic optimization approaches, beginning with low molecular weight fragments that bind weakly to targets and growing them into potent lead compounds. This methodology complements traditional approaches by focusing on binding efficiency and ligand-protein interaction optimization, often revealing novel binding modes and allosteric sites that might be missed by conventional high-throughput screening approaches.\n\nThe integration of artificial intelligence and machine learning into drug design has created unprecedented opportunities for combining the intuitive principles of the Topliss tree with data-driven optimization strategies. Modern AI systems can process vast chemical databases to identify optimal substitution patterns, predict ADMET properties, and suggest novel bioisosteric replacements based on learned structure-activity relationships. These systems effectively automate and scale the decision-making processes that Topliss originally systematized, while maintaining the fundamental logic of stepwise optimization based on observed activity changes.\n\nDeep learning approaches, particularly graph neural networks, have shown remarkable capability in modeling molecular structures and predicting biological activities without requiring predefined descriptors. These methods can capture complex non-linear relationships between molecular structure and biological activity, potentially identifying optimization pathways that might not be apparent through traditional approaches. However, the interpretability challenges associated with deep learning models underscore the continued value of mechanistic understanding provided by classical approaches like the Topliss tree.\n\nThe current era of drug discovery benefits from the convergence of multiple optimization strategies, where traditional rational design principles inform AI-driven approaches, and computational predictions guide experimental validation. Modern medicinal chemists routinely combine Topliss-style systematic optimization with bioisosteric replacement strategies, fragment-based approaches, and structure-based design principles. This integrated approach leverages the strengths of each methodology while mitigating individual limitations.\n\nLooking toward the future, the principles established by Topliss and refined through decades of bioisosteric research continue to provide fundamental guidance for molecular optimization. The emergence of quantum computing applications in drug design, advanced protein folding predictions, and increasingly sophisticated AI models promises to further enhance our ability to design optimal therapeutic molecules. However, the core concepts of systematic exploration, physicochemical property optimization, and rational structural modification remain as relevant today as they were when first introduced.\n\nThe evolution from Topliss trees to modern AI-driven drug design represents not a replacement of traditional approaches, but rather their integration into increasingly powerful optimization frameworks. As we advance into an era of precision medicine and personalized therapeutics, the systematic thinking embodied in classical approaches provides the foundation upon which more sophisticated computational methods can build. The challenge for modern drug discovery lies not in choosing between traditional and computational approaches, but in effectively integrating them to harness the full potential of both human insight and artificial intelligence in the pursuit of better medicines.", "# Microplastics and Cancer: The Mounting Evidence via Literature Review\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n1. **Microplastics as Emerging Carcinogens: From Environmental Pollutants to Oncogenic Drivers:**\u00a0[Mishra et al., Mol Cancer, 2025](https://doi.org/10.1186/s12943-025-02409-4)\n2. **Emerging Evidence on Micro- and Nanoplastics Carcinogenicity: Mechanisms, Models, and Signaling Networks:**\u00a0[Alaraby et al., Mol Cancer, 2026](https://doi.org/10.1186/s12943-026-02657-y)\n3. **The Hidden Threat: Microplastics and Cancer Biology:**\u00a0[Cirillo et al., J Hazard Mater, 2026](https://doi.org/10.1016/j.jhazmat.2026.141332)\n4. **Microplastics and Nanoplastics in Cancer Progression: Biology and Public Health:**\u00a0[Ponce-Cusi et al., Biomedicines, 2025](https://doi.org/10.3390/biomedicines14010001)\n5. **Microplastics and Cancer:**\u00a0[Bogani et al., Eur J Cancer Prev, 2025](https://doi.org/10.1097/CEJ.0000000000000999)\n6. **Micro and Nanoplastics in Human Carcinogenesis: Insights from In Vitro Studies:**\u00a0[Silva et al., Toxicology, 2025](https://doi.org/10.1016/j.tox.2025.154331)\n7. **Invisible Invaders: The Carcinogenic Threat of Microplastics and Nanoplastics in Colorectal Cancer (Systematic Review):**\u00a0[Wen & Lin, Front Public Health, 2025](https://doi.org/10.3389/fpubh.2025.1653245)\n8. **Microplastics: An Often-Overlooked Issue in the Transition from Chronic Inflammation to Cancer:**\u00a0[Cheng et al., J Transl Med, 2024](https://doi.org/10.1186/s12967-024-05731-5)\n9. **Evaluating the Relationship Between Microplastics and Nanoplastics Contamination and Diverse Cancer Types:**\u00a0[Ga\u0142\u0119ziewska et al., Environ Pollut, 2025](https://doi.org/10.1016/j.envpol.2025.127052)\n10. **Do Microplastics and Nanoplastics Directly Contribute to Human Carcinogenesis?:**\u00a0[Casella et al., Environ Pollut, 2025](https://doi.org/10.1016/j.envpol.2025.127343)\n11. **The Bioaccumulation and Carcinogenic Potential of Micro- and Nanoplastics in Humans:**\u00a0[Nabi et al., Toxicol Res, 2026](https://doi.org/10.1007/s43188-025-00332-w)\n12. **Insights into the Potential Carcinogenicity of Micro- and Nano-plastics:**\u00a0[Domenech et al., Mutat Res Rev, 2023](https://doi.org/10.1016/j.mrrev.2023.108453)\n13. **Micro(nano)plastics Pollution and Human Health: How Plastics Can Induce Carcinogenesis:**\u00a0[Kumar et al., Chemosphere, 2022](https://doi.org/10.1016/j.chemosphere.2022.134267)\n14. **Microplastics and Gynecological Tumors: An Emerging Environmental Health Concern:**\u00a0[Hu et al., Reprod Toxicol, 2025](https://doi.org/10.1016/j.reprotox.2025.109018)\n15. **Microplastic-Associated Pollutants in Prostate Carcinogenesis and Plant-Based Chemo-Preventive Strategies:**[Kumar et al., Drug Chem Toxicol, 2026](https://doi.org/10.1080/01480545.2026.2658191)\n16. **Microplastics Role in Cell Migration and Distribution During Cancer Cell Division:**\u00a0[Brynzak-Schreiber et al., Chemosphere, 2024](https://doi.org/10.1016/j.chemosphere.2024.141463)\n17. **Polystyrene Microplastics Aggravate Clear Cell Renal Cell Carcinoma Progression via NF-\u03baB and TGF-\u03b2:**\u00a0[Ye et al., Adv Sci, 2025](https://doi.org/10.1002/advs.202518500)\n18. **PET Nanoplastics Promote Gastric Cancer Progression by Inhibiting TNF-\u03b1/JAK1/STAT1 Signaling:**\u00a0[Li et al., Environ Int, 2026](https://doi.org/10.1016/j.envint.2026.110289)\n19. **Polystyrene Nanoplastics Promote Colitis-Associated Cancer by Disrupting Lipid Metabolism and Inducing DNA Damage:**\u00a0[Tian et al., Environ Int, 2025](https://doi.org/10.1016/j.envint.2025.109258)\n20. **Long-Term Exposure to Real-Life PET Nanoplastics Induces Carcinogenesis In Vitro:**\u00a0[Guti\u00e9rrez-Garc\u00eda et al., Environ Sci Technol, 2025](https://doi.org/10.1021/acs.est.5c01628)\n21. **Dysregulation of SUSD2-CLDN18.2 Cell Adhesion Drives Lung Adenocarcinoma Under Chronic Nanoplastics Exposure:**\u00a0[Wang et al., Ecotoxicol Environ Saf, 2025](https://doi.org/10.1016/j.ecoenv.2025.118946)\n22. **Chronic Nanoplastics Exposure Drives Lung Cancer Progression via the ATF3/PCK2 Axis:**\u00a0[Li et al., J Hazard Mater, 2025](https://doi.org/10.1016/j.jhazmat.2025.140764)\n23. **Polystyrene Nanoplastics and Lung Cancer: Network Toxicology and Mechanistic In Vitro Studies:**\u00a0[Le et al., Toxicol Appl Pharmacol, 2025](https://doi.org/10.1016/j.taap.2025.117618)\n24. **Long-Term Low-Dose Polystyrene Nanoplastics Enhance Metastatic Potential of Colorectal Cancer Cells:**\u00a0[Kim et al., Toxicology, 2025](https://doi.org/10.1016/j.tox.2025.154356)\n25. **SGLT2-PPAR\u03b3 Axis as a Target Against Polyethylene Microplastic-Driven Metabolic Reprogramming in Colorectal Cancer:**\u00a0[Donisi et al., J Transl Med, 2026](https://doi.org/10.1186/s12967-026-07776-0)\n26. **Oral Polystyrene Nanoplastics Promote Triple-Negative Breast Cancer Tumorigenesis and Metastasis:**\u00a0[Zhu et al., Adv Sci, 2025](https://doi.org/10.1002/advs.202508310)\n27. **In Vitro Cell-Transforming Capacity of Micro- and Nanoplastics Derived from 3D-Printing Waste:**\u00a0[Rodr\u00edguez-Garraus et al., Ecotoxicol Environ Saf, 2025](https://doi.org/10.1016/j.ecoenv.2025.118007)\n28. **Detection of Unlabeled Polystyrene Micro- and Nanoplastics in Mammalian Tissue by Optical Photothermal Infrared Spectroscopy:**\u00a0[Duswald et al., Anal Chem, 2025](https://doi.org/10.1021/acs.analchem.4c05400)", "# Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor\nAuthors: Kong, L.; Ju, B.; Chen, Yajing; He, L.; Ren, L.; Liu, J.; Hong, K.; Su, B.; Wang, Z.; Ozorowski, G.; Ji, X.; Hua, Y.; Chen, Yanli; Deller, M.C.; Hao, Y.; Feng, Y.; Garces, F.; Wilson, R.; Dai, K.; O'Dell, S.; McKee, K.; Mascola, J.R.; Ward, A.B.; Wyatt, R.T.; Li, Y.; Wilson, I.A.; Zhu, J.; Shao, Y.\nSource: Immunity, 2016\n\nSecond, the 5-aa LCDR3 signature of VRC01-class bNAbs appeared to be rather common in the B cell repertoire and not as rare as previously perceived. As observed from the VRC01-class bNAbs identified thus far (Scheid et al., 2011; Wu et al., 2011; Zhou et al., 2013; Zhu et al., 2013), the light-chain germline gene is less restricted than that of the heavy chain, allowing both \u03ba and \u03bb genes to be used. For DRVIA7, a new germline gene, IgKV1-5*01, was used. Our repertoire analysis revealed a stable presence (up to 2% of the repertoire) of light chains with 5-aa CDR3 loops and showed that the immune system could rapidly produce a large number of new germline light chains to facilitate B cell selection. Thus, the LCDR3 signature as a barrier to VRC01 maturation might be readily overcome by an active repertoire.\n\nThird, the interplay of LCDR1 and N terminus with the N276 and V5 glycans has not been fully appreciated within the context of VRC01-class antibody development. Previous work placed high importance on establishing a favorable angle of approach to avoid clash with neighboring gp120 subunits and the V1V2 loop within the trimeric spike (Chen et al., 2009; Zhou et al., 2015), which appears not to be an issue for DRVIA7. The limited neutralizing activity and the stalled lineage development, however, could be well explained by a less favorable DRVIA7 light-chain interaction with gp120 glycans. A DRVIA7 with reengineered N terminus and LCDR1 displayed similar breadth to VRC01, indicating that such small changes to accommodate gp120 glycans at N276 and V5 could transform an antibody with limited neutralization breadth to a bNAb.\n\nIn conclusion, DRVIA7 provided us an opportunity to recapitulate the early events in B cell maturation and to assess the barriers for generation of a VRC01-like antibody. Starting from the naive B cell with recombined 5-aa LCDR3, the DRVIA7 lineage precursors emerged with neutralization-competent heavy chains in 2006, rapidly peaked by 2008, and began to decline in 2009, all in the face of rapidly diversifying HIV. Despite the attempts by the immune system to overcome the glycan barrier, the maturation of light-chain N terminus and CDR1 did not occur naturally; however, antibody engineering was able to lead to a neutralization breadth on par with VRC01. Thus, our case study of a Chinese LTNP provided much-sought-after information on the emergence of VRC01-class antibodies that had been sorely lacking. Our findings also supported the notion that redesign of the glycan shield around the CD4bs at N276 and the V5 loop might lead to an effective immunogen to elicit VRC01-like antibodies, providing useful information to facilitate rational vaccine design.\n\n## Experimental Procedures\n\n### Chinese Donor DRVI01 Samples\n\nThe sera and peripheral blood mononuclear cells (PBMCs) described in this study were collected from an HIV-1-infected Chinese patient, DRVI01, from a former plasma donor (FPD) study cohort founded by the Collaboration for AIDS Vaccine Discovery (CAVD) of the Bill and Melinda Gates Foundation, Chinese National Natural Science Foundation, and National Major Project on Infectious Disease. This study has been reviewed and approved by the Institutional Review Boards (IRBs) of the National Center for AIDS/STD Control and Prevention, the Chinese Center for Disease Control and Prevention. A detailed neutralization analysis has been reported for the samples from this cohort (Hu et al., 2012).\n\n### Isolation of Monoclonal Antibodies from Donor DRVI01\n\nIsolation of CD4bs-directed DRVIA mAbs was achieved by single-cell FACS using the CD4bs-specific RSC3 and RSC3\u0394371I probes (Wu et al., 2010). Memory B cells were sorted on a FACS Aria III cell sorter (BD Biosciences) to obtain single B cells with the phenotype CD19+ CD20+ CD14\u2013 CD3\u2013 CD8\u2013 IgG+ RSC3+ RSC3\u0394371I\u2013. Single cells were then sorted into 96-well plates followed by RT-PCR, IgG cloning, and expression. Details are described in the Supplemental Experimental Procedures.\n\n### Expression and Purification of DRVIA7 Fab, gp120, and gp140 Proteins\n\nAll proteins were produced transiently in either FreeStyle 293F or GnTI\u2013/\u2013 293S cells as previously described (Kong et al., 2013). In brief, after 4\u20137 days, the supernatants of the transfected cells were harvested and purified with a *Galanthus nivalis* lectin column (Vector Laboratories) for gp120 or with a Kappa CaptureSelect column (BAC BV) for DRVIA7 Fab. All proteins were further purified by size exclusion chromatography (SEC).\n\n### Crystallographic Analysis of DRVIA7 Fab and gp120-Bound Complex\n\nCrystals were obtained for DRVIA7 Fab and the complex of 93TH057 gp120 core bound to DRVIA7 Fab. The crystals diffracted to 2.9 \u00c5 and 3.4 \u00c5 resolution, respectively, and their structures were solved by molecular replacement. After refinement, final Rcryst and Rfree values were 23.4% and 28.4% for DRVIA7 Fab, and 25.3% and 30.6% for gp120-bound DRVIA7 Fab complex, respectively (Table S2). Details are described in the Supplemental Experimental Procedures.\n\n### Negative Stain Electron Microscopy\n\nJRFL SOSIP.664 trimers in complex with either DRVIA7 or DRVIA7H-VRC01L Fab were analyzed by negative-stain EM. Data were collected on either an FEI Tecnai T12 electron microscope or an FEI Talos electron microscope. Data processing and 3D reconstruction were performed as previously described (de Taeye et al., 2015). Crystal structures or EM volumes were fit into the 3D reconstructions to generate figures. Details are described in Supplemental Experimental Procedures.\n\n### Ion Torrent PGM Sequencing of Donor DRVI01 Antibody Libraries", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nFigure 3-9 below shows the elution profiles of gel filtration using a Sephadex-75 column. Panel (a) shows the elution profile of molecular weight standards, with five main peaks at 2000, 150, 66, 29 and 12 kDa. Panel (b) shows the elution profile of affinity purified OSM_\u0394188 in the absence of Triton X-100 which shows two major peaks eluted from the Sephadex-75 column at 170 minutes (25.1 kDa) and 155 minutes (34 kDa). Lane 8 of panel (d) shows that the 155 minute peak runs at ~30 kDa on SDS-PAGE. Lane 9 of panel (d) shows that the 170 minute peak runs at ~25 kDa on SDS-PAGE. Panel (c) shows the elution profile of affinity purified OSM_\u0394188 in the presence of Triton X-100 and this shows two major peaks eluted from the Sephadex-75 column at 170 minutes (25.1 kDa) and 140 minutes (50 kDa). Lane 7 of panel (d) shows that the 140 minute peak runs at ~25 kDa on SDS-PAGE.\n\nFigure 3-10 below shows elution profiles of a Sephadex-75 column with OSM-Se-Met expressed in E. coli strain (a) B834 and (b) DL41. These chromatograms show two major peaks eluted at 170 minutes (25.1 kDa) and 140 minutes (50 kDa). Lane 9 of panel (c) shows that the peak eluted at 170 minutes runs at ~25 kDa on SDS-PAGE.\n\n*Figure 3-9.* Chromatogram showing FPLC elution profile of Superdex-75 prep grade gel filtration column with (a) molecular weight standards, (b) affinity purified OSM_\u0394188 sample without Triton X-100 and (c) affinity purified OSM_\u0394188 sample with Triton X-100. 4 ml samples were loaded at a flow rate of 1 ml/min and the absorbance monitored at 280 nm. In the presence of Triton X-100 two peaks are eluted at 170 minutes (25.1 kDa) and 140 minutes (50 kDa). In the absence of Triton X-100 a second peak is eluted at 155 minutes (34 kDa). SDS-PAGE (d) of preparation, with lane 1 and 2 showing total cellular protein pre- and post-induction. Lane 3 contains material bound to affinity column. Lane 4 shows material bound to affinity column after cleavage with 3C protease and lane 5 shows OSM_\u0394188 eluted from affinity column. Lane 6 shows material loaded onto gel filtration. Lane 7 shows the 140 min peak and lane 8 shows the 155 min peak. Lane 9 shows concentrated peak of purified OSM_\u0394188 eluted at ~170 minutes. Lane 10 contains broad range molecular weight markers (Bio-Rad).\n\n*Figure 3-10.* Chromatogram showing FPLC elution profile of Superdex-75 prep grade gel filtration column with affinity purified OSM_Se-Met from E. coli strain (a) B834 and (b) DL41. 4 ml samples were loaded at a flow rate of 1 ml/min and the absorbance monitored at 280 nm. Two main peaks are eluted at 170 minutes (25.1 kDa) and 140 minutes (50 kDa). SDS-PAGE (c) of preparation with lane 1 and 2 showing total cellular protein pre- and post-induction. Lane 3 shows the material that did not bind to the affinity column. Lane 4 contains the cleared lysate and lane 5 shows the column wash. Lane 6 shows material bound to affinity column and lane 7 shows material remaining on column after cleavage with 3C protease. Lane 8 contains material eluted from affinity column after cleavage with 3C protease and lane 9 contains the concentrated peak of purified OSM-Se-Met eluted at ~170 minutes. Lane 10 contains broad range molecular weight markers.\n\n### 3.3.8 N-Terminal Sequencing\n\nThe results of N-terminal sequencing are shown in Table 3-4 below. This data shows that E. coli strain JM109 produces only one species of native OSM_\u0394188 with the addition of a four residue 3C linker site (GPGS) at the N-terminus. E. coli strain JM109 produced two minor species when grown on minimal media supplemented with 78Se-Met. The E. coli methionine auxotrophic strain B834 produced only one species.\n\n| E. coli strain | Protein | N-terminal Sequence | Yield (%) |\n|---|---|---|---|\n| JM109 | Native (32S-Met) | GPGSAAIGS-SKEYR | 38.3 |\n| JM109 | Derivative (78Se-Met) | GPGSAAIGS-SKEYRVLLG | 3.8 |\n| JM109 | Derivative (78Se-Met) | AAIGS-SKEYRVLLG | 4.2 |\n| B834 | Derivative (78Se-Met) | GPGSAAIGS- | 38.1 |\n\n*Table 3-4.* Table showing results of N-terminal sequencing.\n\n### 3.3.9 Mass Spectrometry\n\nFigure 3-11 (a) below shows mass spectra of OSM_\u0394188 samples run in electrospray mode. This shows one major peak at 21579.98 \u00b1 9.05 Da. The theoretical mass of OSM_\u0394188 is 21409.50 Da as calculated using the ExPASy molecular biology server (The University of Geneva; www.expasy.ch).\n\nFigure 3-11 (b) shows OSM-Se-Met material run in electrospray mode and this shows five major peaks, corresponding to sequential substitution of 32S-Met with 78Se-Met. The five species are observed at 21593.83 \u00b1 1.62 Da (no substitution), 21641.27 \u00b1 10.75 Da (one selenium substitution), 21683.85 \u00b1 1.39 Da (two selenium substitutions), 21745.63 \u00b1 1.43 Da (three selenium substitutions) and 21765.79 \u00b1 1.18 Da (four selenium substitutions; i.e. substitution of all methionines).\n\n*Figure 3-11.* Figure showing mass spectra of (a) OSM_\u0394188 and (b) OSM-Se-Met samples. Typically, ~400 pmol samples of HPLC purified protein in water were analysed by electrospray ionization on a Micromass BioQ II spectrometer. OSM_\u0394188 samples display one major peak at 21579.98 \u00b1 9.05. OSM-Se-Met samples display five major peaks at 21593.82 \u00b1 1.62 (native), 21641.27 \u00b1 10.75 (1 \u00d7 Se), 21695.07 \u00b1 1.82 (2 \u00d7 Se), 21745.63 \u00b1 1.43 (3 \u00d7 Se) and 21803.21 \u00b1 3.86 (4 \u00d7 Se).\n\n## 3.4 Discussion of Expression, Purification and Characterisation of Oncostatin M\n\nProduction of 3C protease in sufficient quantities proved successful. Figure 3-3 shows that ~90 % pure preparations of 3C-His6 protease could be produced. Figure 3-4 demonstrates that DTT is essential for efficient cleavage of fusion proteins with 3C protease. Cleavage was achieved using very low concentrations of 3C protease as either GST or His6 fusions. Typically, 3C-His6 protease was used at a ratio of 1:1000 of protease to fusion protein and removal of 3C-His6 was carried out by passage over nickel affinity resin.", "# 2017 \u2013 The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Concept | Description | Key Reference |\n| --- | --- | --- |\n| Hydrogen-Bond Network | H161 and Q167 stabilize oxygen binding, aiding proton transfer. | [Marletta, et al., ACS Chem. Biol., 2017](https://pubs.acs.org/doi/abs/10.1021/acschembio.7b00016) |\n| Motif Diversity | REF and HQY/HQF motifs prevalent across AA10 LPMOs modulate reactivity. | [Vaaje-Kolstad, et al., ACS Omega, 2024](https://pubs.acs.org/doi/10.1021/acsomega.4c02666) |\n| Redox Tuning | Glutamine/glutamate protonation alters reduction potentials by hundreds of mV. | [S\u00f8gaard, et al., J. Am. Chem. Soc., 2023](https://pubs.acs.org/doi/10.1021/jacs.3c05342) |\n| Substrate Binding | Second-sphere histidine at ~3.3 \u00c5 from the brace directs polysaccharide recognition. | [Gonz\u00e1lez-Garc\u00eda, et al., Biochem. Soc. Trans., 2021](https://portlandpress.com/biochemsoctrans/article/49/1/531/227605) |\n| Peroxygenase Activity | Q219E mutation elevates H\u2082O\u2082-driven chitin oxidation in AA10s. | [Vaaje-Kolstad, et al., ACS Omega, 2024](https://pubs.acs.org/doi/10.1021/acsomega.4c02666) |\n| Solvent Rescue | Cavity created by alanine substitutions allows solvent H-bonding to sustain activity. | [Marletta, et al., ACS Chem. Biol., 2017](https://pubs.acs.org/doi/abs/10.1021/acschembio.7b00016) |", "# pants.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n| Candidate | Environment | Length | Bitscore | Identity | pLDDT | Cleft \u00c5 | Nearest |\n| --- | --- | --- | --- | --- | --- | --- | --- |\n| [PANTS-d94f1dfcbcb5](/candidate/PANTS-d94f1dfcbcb5) | Compost | 290 | 486.0 | 1.00 | 94 | 16.67 | EC:A0A543IKT7 |\n| [PANTS-2cc6ed0d4170](/candidate/PANTS-2cc6ed0d4170) | Compost | 367 | 466.2 | 0.88 | 87 | 14.78 | EC:A0A975LD17 |\n| [PANTS-1374ba10a5bb](/candidate/PANTS-1374ba10a5bb) | Compost | 310 | 461.6 | 0.80 | 92 | 17.23 | EC:A0ABX8C4W7 |\n| [PANTS-fddfd0b0d10e](/candidate/PANTS-fddfd0b0d10e) | Compost | 308 | 461.0 | 0.80 | 92 | 17.24 | EC:A0ABX8C4W7 |\n| [PANTS-21f383c6955e](/candidate/PANTS-21f383c6955e) | Compost | 315 | 457.5 | 0.68 | 92 | 16.67 | EC:A0A7W3LQR6 |\n| [PANTS-3d53535c3f75](/candidate/PANTS-3d53535c3f75) | Compost | 301 | 455.0 | 0.96 | 94 | 17.22 | TfCut2 |\n| [PANTS-5ee40d6243c4](/candidate/PANTS-5ee40d6243c4) | Compost | 304 | 454.7 | 1.00 | 93 | 16.99 | Cut190 |\n| [PANTS-50ee90fbde21](/candidate/PANTS-50ee90fbde21) | Compost | 300 | 451.6 | 1.00 | 93 | 19.86 | EC:A0ABY4KWB4 |\n| [PANTS-d8d5896935d2](/candidate/PANTS-d8d5896935d2) | Compost | 300 | 451.3 | 1.00 | 93 | 19.86 | EC:A0ABY4KWB4 |\n| [PANTS-088de59db354](/candidate/PANTS-088de59db354) | Compost | 296 | 449.4 | 0.96 | 93 | 19.86 | EC:A0ABY4KWB4 |\n| [PANTS-8647e5670cbb](/candidate/PANTS-8647e5670cbb) | Compost | 301 | 446.8 | 0.99 | 94 | 17.17 | EC:A0A9P2WRG3 |\n| [PANTS-780fbc037344](/candidate/PANTS-780fbc037344) | Compost | 272 | 440.5 | 1.00 | 92 | 19.71 | EC:A0ABY4KWB4 |\n| [PANTS-b1b2f95a9cef](/candidate/PANTS-b1b2f95a9cef) | Compost | 280 | 357.0 | 0.57 | 96 | 17.31 | EC:A0ABW2PE71 |\n| [PANTS-5ecab0206577](/candidate/PANTS-5ecab0206577) | Compost | 293 | 354.6 | 0.57 | 94 | 16.83 | EC:A0ABN3C865 |\n| [PANTS-3fbe2d3b631d](/candidate/PANTS-3fbe2d3b631d) | Compost | 318 | 346.0 | 0.70 | 93 | 20.97 | PLC:A0A1W6L438 |\n| [PANTS-0f15970cbfdf](/candidate/PANTS-0f15970cbfdf) | Compost | 194 | 326.1 | 0.77 | 90 | 23.99 | PLC:A0A168EN35 |\n| [PANTS-2f3770bc2355](/candidate/PANTS-2f3770bc2355) | Compost | 283 | 318.9 | 0.59 | 96 | 19.68 | PLC:A0A0U3MFB6 |\n| [PANTS-a261334522ec](/candidate/PANTS-a261334522ec) | Landfill | 275 | 297.8 | 0.78 | 80 | 21.09 | PLC:A0A078MGG8 |\n| [PANTS-13530f7d85a7](/candidate/PANTS-13530f7d85a7) | Compost | 164 | 244.3 | 0.68 | 87 | 21.43 | EC:A0A931GV53 |\n| [PANTS-ecb485eac2c5](/candidate/PANTS-ecb485eac2c5) | Compost | 194 | 235.7 | 0.66 | 93 | 17.71 | PLC:Q8RR62 |\n| [PANTS-7304e7e56216](/candidate/PANTS-7304e7e56216) | Compost | 155 | 229.5 | 0.58 | 88 | 15.72 | PLC:A0A1I6NU60 |\n| [PANTS-068d318a9554](/candidate/PANTS-068d318a9554) | Compost | 201 | 115.4 | 0.32 | 91 | 20.4 | PLC:E9KJL1 |\n| [PANTS-7b3c1e20c460](/candidate/PANTS-7b3c1e20c460) | Landfill | 250 | 108.3 | 0.30 | 88 | 21.23 | EC:A0A931AKC4 |\n| [PANTS-f76ac70478bd](/candidate/PANTS-f76ac70478bd) | Landfill | 391 | 83.5 | 0.35 | 93 | 18.05 | EC:A0ABW2SV52 |\n| [PANTS-9b6ff14275b6](/candidate/PANTS-9b6ff14275b6) | Landfill | 504 | 76.8 | 0.33 | 90 | 24.73 | EC:A0ABW2SV52 |\n| [PANTS-bfd978d4ddca](/candidate/PANTS-bfd978d4ddca) | Compost | 287 | 75.7 | 0.72 | 92 | 16.11 | PLC:A0A1H1SPP7 |\n| [PANTS-fd48eb25bbe0](/candidate/PANTS-fd48eb25bbe0) | Compost | 283 | 71.7 | 0.35 | 89 | 18.64 | PLC:A0A075UPJ7 |\n| [PANTS-a0087546ef4d](/candidate/PANTS-a0087546ef4d) | Human gut | 315 | 71.5 | 0.33 | 92 | 22.39 | EC:A0ABW1D4K0 |\n| [PANTS-5b5625eec855](/candidate/PANTS-5b5625eec855) | Compost | 287 | 68.6 | 0.67 | 92 | 15.9 | PLC:A0A1H1SPY0 |\n| [PANTS-f1bc07683fad](/candidate/PANTS-f1bc07683fad) | Marine plastisphere | 339 | 64.9 | 0.26 | 85 | 21.7 | EC:A0A8H9GVK2 |\n| [PANTS-9111db0b9d00](/candidate/PANTS-9111db0b9d00) | Marine plastisphere | 388 | 64.7 | 0.40 | 94 | 20.7 | EC:A0A7W0CGN6 |\n| [PANTS-3bed2d847a10](/candidate/PANTS-3bed2d847a10) | Marine plastisphere | 439 | 62.0 | 0.25 | 82 | 25.01 | PLC:A0A078MGG8 |\n| [PANTS-9f420183d3d0](/candidate/PANTS-9f420183d3d0) | Human gut | 304 | 61.9 | 0.33 | 91 | - | PLC:A0A163MA58 |\n| [PANTS-4e79a1ab303e](/candidate/PANTS-4e79a1ab303e) | Human gut | 307 | 61.7 | 0.33 | 91 | - | PLC:A0A163MA58 |\n| [PANTS-ff40e7bf06df](/candidate/PANTS-ff40e7bf06df) | Human gut | 307 | 61.2 | 0.33 | 91 | - | PLC:A0A163MA58 |\n| [PANTS-f4ff41161010](/candidate/PANTS-f4ff41161010) | Compost | 176 | 60.6 | 0.79 | 85 | 15.78 | PLC:A0A1H1SPP7 |\n| [PANTS-c5db9ee6498b](/candidate/PANTS-c5db9ee6498b) | Human gut | 274 | 60.6 | 0.34 | 87 | - | PLC:A0A163MA58 |\n| [PANTS-a68e278f8889](/candidate/PANTS-a68e278f8889) | Human gut | 302 | 60.4 | 0.33 | 96 | - | PLC:A0A163MA58 |\n| [PANTS-242d5314c2b8](/candidate/PANTS-242d5314c2b8) | Human gut | 304 | 60.3 | 0.33 | 96 | - | PLC:A0A163MA58 |\n| [PANTS-47feb678bf91](/candidate/PANTS-47feb678bf91) | Marine plastisphere | 599 | 59.9 | 0.26 | - | - | EC:A0A8H9GVK2 |\n| [PANTS-a4997f37c1c1](/candidate/PANTS-a4997f37c1c1) | Marine plastisphere | 398 | 59.6 | 0.30 | 79 | 24.89 | EC:A0A239F412 |\n| [PANTS-4b9cdfabb1a3](/candidate/PANTS-4b9cdfabb1a3) | Marine plastisphere | 498 | 57.4 | 0.24 | - | - | EC:A0A2P4UDA7 |\n| [PANTS-b72c6e2d6f64](/candidate/PANTS-b72c6e2d6f64) | Landfill | 414 | 52.2 | 0.25 | 84 | - | EC:A0A919RR95 |\n| [PANTS-974c1f1a5418](/candidate/PANTS-974c1f1a5418) | Marine plastisphere | 551 | 51.9 | 0.38 | - | - | EC:A0A239F412 |\n| [PANTS-76b05ee43620](/candidate/PANTS-76b05ee43620) | Compost | 222 | 50.6 | 0.48 | 90 | 16.37 | EC:A0ABW2CM09 |\n| [PANTS-5a8208b2421a](/candidate/PANTS-5a8208b2421a) | Human gut | 202 | 50.5 | 0.38 | 96 | 22.5 | HGMP01 |\n| [PANTS-27463e09e9cc](/candidate/PANTS-27463e09e9cc) | Compost | 168 | 49.3 | 0.37 | 83 | 16.82 | EC:A0ABW2CM09 |\n| [PANTS-0a6c3795452a](/candidate/PANTS-0a6c3795452a) | Landfill | 292 | 49.3 | 0.30 | 85 | 25.22 | EC:A0A919RR95 |\n| [PANTS-2b40ec6a6df0](/candidate/PANTS-2b40ec6a6df0) | Compost | 335 | 47.7 | 0.30 | 88 | 18.17 | EC:A0ABP8Q6M0 |\n| [PANTS-965c9a755eab](/candidate/PANTS-965c9a755eab) | Marine plastisphere | 218 | 46.3 | 0.32 | 87 | 21.71 | EC:A0ABT4UC71 |\n| [PANTS-4a3466447bfc](/candidate/PANTS-4a3466447bfc) | Human gut | 291 | 46.1 | 0.26 | 89 | 22.59 | EC:A0A852TYE9 |\n| [PANTS-de7733a036af](/candidate/PANTS-de7733a036af) | Human gut | 357 | 45.8 | 0.31 | 87 | 23.38 | PLC:A0A163MA58 |\n| [PANTS-28993327d4e2](/candidate/PANTS-28993327d4e2) | Compost | 198 | 44.1 | 0.31 | 94 | 17.59 | EC:A0A9X7PFK9 |\n| [PANTS-efa74b948904](/candidate/PANTS-efa74b948904) | Human gut | 262 | 44.0 | 0.31 | 92 | 22.55 | EC:A0ABW2Y2H3 |\n| [PANTS-9f6ec163c256](/candidate/PANTS-9f6ec163c256) | Landfill | 152 | 43.6 | 0.44 | 92 | - | PLC:A0A1H2ERI7 |\n| [PANTS-caf723f779ae](/candidate/PANTS-caf723f779ae) | Human gut | 336 | 43.6 | 0.99 | 76 | 22.69 | PLC:A0A0H2ZDU5 |\n| [PANTS-70b93e2650b6](/candidate/PANTS-70b93e2650b6) | Human gut | 291 | 43.5 | 0.31 | 90 | 22.72 | EC:A0ABW2Y2H3 |\n| [PANTS-b976dd3ca628](/candidate/PANTS-b976dd3ca628) | Human gut | 291 | 43.5 | 0.31 | 91 | 22.63 | EC:A0ABW2Y2H3 |\n| [PANTS-ea055bedfb9a](/candidate/PANTS-ea055bedfb9a) | Human gut | 291 | 43.5 | 0.31 | 91 | 22.63 | EC:A0ABW2Y2H3 |\n| [PANTS-0196beae52ed](/candidate/PANTS-0196beae52ed) | Human gut | 291 | 42.7 | 0.31 | 91 | 21.01 | EC:A0ABW2Y2H3 |\n| [PANTS-7b3472deddea](/candidate/PANTS-7b3472deddea) | Human gut | 291 | 42.7 | 0.32 | 91 | 22.63 | EC:A0ABW2Y2H3 |\n| [PANTS-2f33d772d13d](/candidate/PANTS-2f33d772d13d) | Human gut | 287 | 42.5 | 0.26 | 91 | 22.57 | EC:A0A852TYE9 |\n| [PANTS-2c69a1dc5b75](/candidate/PANTS-2c69a1dc5b75) | Human gut | 307 | 42.3 | 0.32 | 91 | 19.53 | PLC:A0A0M1P5R7 |\n| [PANTS-b28fd00c46cb](/candidate/PANTS-b28fd00c46cb) | Landfill | 453 | 42.1 | 0.38 | - | - | EC:A0ABW2SV52 |\n| [PANTS-3a023517eefe](/candidate/PANTS-3a023517eefe) | Compost | 248 | 41.7 | 0.29 | 92 | 25.46 | EC:A0A841IUE8 |\n| [PANTS-1759f98114b6](/candidate/PANTS-1759f98114b6) | Human gut | 307 | 41.0 | 0.32 | 91 | 20.86 | PLC:A0A0M1P5R7 |\n| [PANTS-3f129f09247b](/candidate/PANTS-3f129f09247b) | Human gut | 307 | 40.5 | 0.32 | 91 | 21.42 | PLC:A0A0M1P5R7 |\n| [PANTS-182611108704](/candidate/PANTS-182611108704) | Human gut | 254 | 40.3 | 0.30 | 92 | 18.32 | EC:A0ABP6KM42 |\n| [PANTS-9d75aa38ae57](/candidate/PANTS-9d75aa38ae57) | Compost | 257 | 39.7 | 0.22 | 94 | 22.36 | EC:A0A239AX86 |\n| [PANTS-95c1f7ed3e4f](/candidate/PANTS-95c1f7ed3e4f) | Human gut | 156 | 39.2 | 0.32 | 86 | 19.57 | EC:A0ABP6KM42 |\n| [PANTS-0e188a1aef04](/candidate/PANTS-0e188a1aef04) | Human gut | 290 | 39.1 | 0.27 | 89 | 17.67 | EC:A0A8J3U475 |\n| [PANTS-6d328f4d943e](/candidate/PANTS-6d328f4d943e) | Human gut | 291 | 38.3 | 0.27 | 89 | 17.68 | EC:A0A8J3U475 |\n| [PANTS-377715c04715](/candidate/PANTS-377715c04715) | Human gut | 259 | 38.1 | 0.30 | 93 | 21.14 | EC:A0A7Y9GI10 |\n| [PANTS-1da6b1efb781](/candidate/PANTS-1da6b1efb781) | Human gut | 259 | 38.0 | 0.30 | 93 | 21.14 | EC:A0A7Y9GI10 |\n| [PANTS-2526a36ebc8b](/candidate/PANTS-2526a36ebc8b) | Human gut | 450 | 37.1 | 0.29 | 69 | 20.95 | EC:A0A8J3U475 |\n| [PANTS-50c7618c98d1](/candidate/PANTS-50c7618c98d1) | Marine plastisphere | 604 | 36.0 | 0.32 | - | - | EC:A0A919RR95 |\n| [PANTS-4f5560f67dd5](/candidate/PANTS-4f5560f67dd5) | Human gut | 206 | 36.0 | 0.31 | 92 | 19.08 | EC:A0ABT4UC71 |\n| [PANTS-c6ca1161ffce](/candidate/PANTS-c6ca1161ffce) | Compost | 339 | 35.9 | 0.30 | 81 | 22.11 | EC:D1A646 |\n| [PANTS-1a777b0f8087](/candidate/PANTS-1a777b0f8087) | Compost | 365 | 35.8 | 0.30 | 78 | 16.8 | EC:D1A646 |\n| [PANTS-b8e6fd527452](/candidate/PANTS-b8e6fd527452) | Human gut | 130 | 35.2 | 0.28 | 89 | 11.09 | EC:A0A919UR82 |\n| [PANTS-38f9606202dc](/candidate/PANTS-38f9606202dc) | Compost | 324 | 34.8 | 0.34 | 86 | 20.39 | EC:A0A7C9NKH2 |\n| [PANTS-158e5477f376](/candidate/PANTS-158e5477f376) | Human gut | 464 | 34.8 | 0.31 | - | - | EC:A0ABP7GG15 |\n| [PANTS-bd638266b6e3](/candidate/PANTS-bd638266b6e3) | Human gut | 137 | 34.8 | 0.36 | 90 | 18.94 | EC:A0A8J3U475 |\n| [PANTS-289b69ff69ba](/candidate/PANTS-289b69ff69ba) | Human gut | 165 | 34.8 | 0.33 | 86 | 18.23 | EC:A0A1H5T4B9 |\n| [PANTS-a39aaabaf1f1](/candidate/PANTS-a39aaabaf1f1) | Human gut | 264 | 34.5 | 0.29 | 91 | 22.92 | EC:A0A841DCU0 |\n| [PANTS-91938fb9b714](/candidate/PANTS-91938fb9b714) | Marine plastisphere | 175 | 34.3 | 0.30 | 86 | 21.57 | EC:A0ABW3EGR7 |\n| [PANTS-4f1aac883de2](/candidate/PANTS-4f1aac883de2) | Human gut | 220 | 34.3 | 0.33 | 88 | 19.72 | EC:A0ABT4UC71 |\n| [PANTS-65b45a1d7c7d](/candidate/PANTS-65b45a1d7c7d) | Human gut | 211 | 34.3 | 0.34 | 91 | 19.81 | EC:A0ABR9HP34 |\n| [PANTS-1023a852d7e7](/candidate/PANTS-1023a852d7e7) | Human gut | 295 | 34.3 | 0.41 | 92 | 17.49 | EC:A0A239AX86 |\n| [PANTS-b0038dde71e1](/candidate/PANTS-b0038dde71e1) | Human gut | 256 | 34.2 | 0.31 | 92 | 21.11 | EC:A0ABT4UC71 |\n| [PANTS-34b97219e0df](/candidate/PANTS-34b97219e0df) | Human gut | 181 | 34.2 | 0.28 | 91 | 21.89 | EC:A0A243RI45 |\n| [PANTS-89f1b3515d50](/candidate/PANTS-89f1b3515d50) | Compost | 185 | 34.0 | 0.32 | 92 | 14.83 | EC:A0A2W2EN64 |\n| [PANTS-0e03e38226d3](/candidate/PANTS-0e03e38226d3) | Human gut | 256 | 34.0 | 0.31 | 92 | 21.06 | EC:A0ABT4UC71 |\n| [PANTS-604ea65178a1](/candidate/PANTS-604ea65178a1) | Human gut | 256 | 34.0 | 0.33 | 92 | 21.05 | EC:A0ABT4UC71 |\n| [PANTS-80445f7a45c8](/candidate/PANTS-80445f7a45c8) | Human gut | 201 | 34.0 | 0.33 | 94 | 18.13 | EC:A0A1H5T4B9 |\n| [PANTS-7df6fe0a840b](/candidate/PANTS-7df6fe0a840b) | Human gut | 218 | 34.0 | 0.34 | 89 | 17.44 | EC:A0A2P4UDP6 |\n| [PANTS-c4d00c26823f](/candidate/PANTS-c4d00c26823f) | Human gut | 221 | 33.9 | 0.29 | 91 | 21.09 | EC:A0A8J3U475 |\n| [PANTS-c35e80261652](/candidate/PANTS-c35e80261652) | Human gut | 469 | 33.9 | 0.31 | - | - | EC:A0ABP7GG15 |\n| [PANTS-baad4917e3ca](/candidate/PANTS-baad4917e3ca) | Human gut | 262 | 33.9 | 0.33 | 93 | 21.05 | EC:A0ABT4UC71 |\n| [PANTS-5412013d1452](/candidate/PANTS-5412013d1452) | Human gut | 597 | 33.7 | 0.27 | - | - | EC:A0A919UR82 |\n| [PANTS-3d3c482af6db](/candidate/PANTS-3d3c482af6db) | Human gut | 256 | 33.7 | 0.31 | 92 | 20.71 | EC:A0ABT4UC71 |\n| [PANTS-381df9a78ed6](/candidate/PANTS-381df9a78ed6) | Human gut | 258 | 33.5 | 0.29 | 90 | 21.61 | EC:A0A8J3U475 |\n| [PANTS-bfcf8e94bfc1](/candidate/PANTS-bfcf8e94bfc1) | Human gut | 205 | 33.5 | 0.25 | 89 | 22.64 | EC:A0A9Y2YA38 |\n| [PANTS-c719343cba9c](/candidate/PANTS-c719343cba9c) | Human gut | 258 | 33.4 | 0.29 | 91 | 21.13 | EC:A0A8J3U475 |\n| [PANTS-1537e907f5ea](/candidate/PANTS-1537e907f5ea) | Human gut | 258 | 33.4 | 0.29 | 90 | 21.13 | EC:A0A8J3U475 |\n| [PANTS-d30ad2f5be5a](/candidate/PANTS-d30ad2f5be5a) | Human gut | 258 | 33.4 | 0.29 | 91 | 21.08 | EC:A0A8J3U475 |\n| [PANTS-ce3bf3d15007](/candidate/PANTS-ce3bf3d15007) | Human gut | 246 | 33.4 | 0.28 | 91 | 18.16 | EC:A0A7W3LR11 |\n| [PANTS-87daa3e2f064](/candidate/PANTS-87daa3e2f064) | Human gut | 469 | 33.3 | 0.30 | - | - | EC:A0ABP7GG15 |\n| [PANTS-1a7940ecbc25](/candidate/PANTS-1a7940ecbc25) | Human gut | 223 | 33.3 | 0.28 | 90 | 21.49 | EC:A0A8J3U475 |\n| [PANTS-63e6e43ce87a](/candidate/PANTS-63e6e43ce87a) | Compost | 121 | 33.2 | 0.32 | 88 | 15.46 | EC:A0ABW3EQF3 |\n| [PANTS-d8af9b71ada5](/candidate/PANTS-d8af9b71ada5) | Human gut | 217 | 33.2 | 0.36 | 95 | 22.44 | EC:A0A8J3U475 |\n| [PANTS-3899eec78da2](/candidate/PANTS-3899eec78da2) | Human gut | 414 | 33.1 | 0.26 | 90 | 16.0 | EC:A0ABW1NJ96 |\n| [PANTS-29c3c443af13](/candidate/PANTS-29c3c443af13) | Human gut | 226 | 33.0 | 0.38 | 94 | 17.21 | EC:A0ABT9QDD4 |\n| [PANTS-eac3e551eb7d](/candidate/PANTS-eac3e551eb7d) | Human gut | 258 | 32.9 | 0.28 | 90 | 21.63 | EC:A0A8J3U475 |\n| [PANTS-ddcbf1b00e2d](/candidate/PANTS-ddcbf1b00e2d) | Human gut | 258 | 32.9 | 0.28 | 89 | 21.59 | EC:A0A8J3U475 |\n| [PANTS-23f43e912264](/candidate/PANTS-23f43e912264) | Human gut | 258 | 32.9 | 0.28 | 90 | 21.58 | EC:A0A8J3U475 |\n| [PANTS-3af086a668f7](/candidate/PANTS-3af086a668f7) | Human gut | 254 | 32.9 | 0.28 | 90 | 21.56 | EC:A0A8J3U475 |\n| [PANTS-cb34dab9451e](/candidate/PANTS-cb34dab9451e) | Human gut | 414 | 32.9 | 0.27 | 90 | 16.02 | EC:A0A6I4WH26 |\n| [PANTS-d00ec8def7ec](/candidate/PANTS-d00ec8def7ec) | Marine plastisphere | 411 | 32.8 | 0.28 | 90 | 21.1 | PLC:A0A1I6NU60 |\n| [PANTS-77ec97640668](/candidate/PANTS-77ec97640668) | Human gut | 258 | 32.8 | 0.29 | 91 | 20.95 | EC:A0A8J3U475 |\n| [PANTS-59dbee9d8ba1](/candidate/PANTS-59dbee9d8ba1) | Human gut | 414 | 32.8 | 0.25 | 91 | 16.12 | EC:A0ABW1CGM7 |\n| [PANTS-43edeb782a78](/candidate/PANTS-43edeb782a78) | Human gut | 345 | 32.8 | 0.28 | 82 | 21.22 | EC:A0A8J3PRQ9 |\n| [PANTS-8f1b2a02fcc4](/candidate/PANTS-8f1b2a02fcc4) | Human gut | 414 | 32.8 | 0.25 | 90 | 16.0 | EC:A0A7X0U3W9 |\n| [PANTS-7d932e776b16](/candidate/PANTS-7d932e776b16) | Marine plastisphere | 404 | 32.7 | 0.25 | 88 | 18.07 | EC:A0A2P4UNN9 |\n| [PANTS-08c9ad32ce25](/candidate/PANTS-08c9ad32ce25) | Marine plastisphere | 245 | 32.5 | 0.30 | 83 | 19.06 | EC:A0A1G8SHF7 |\n| [PANTS-a306063fcbaf](/candidate/PANTS-a306063fcbaf) | Human gut | 255 | 32.5 | 0.33 | 93 | 20.7 | EC:A0A5N6AZJ9 |\n| [PANTS-f71cae5478f9](/candidate/PANTS-f71cae5478f9) | Human gut | 255 | 32.5 | 0.31 | 91 | 22.28 | HGMP04 |\n| [PANTS-7dd05febb96c](/candidate/PANTS-7dd05febb96c) | Human gut | 155 | 32.5 | 0.30 | 87 | 18.5 | EC:A0A8J3U475 |\n| [PANTS-99c1f5a2fcac](/candidate/PANTS-99c1f5a2fcac) | Human gut | 414 | 32.5 | 0.28 | 90 | 23.01 | EC:A0A7W9GDV7 |\n| [PANTS-d5ffa53ac4f9](/candidate/PANTS-d5ffa53ac4f9) | Human gut | 184 | 32.5 | 0.33 | 94 | 17.7 | EC:A0A8J3U475 |\n| [PANTS-2f34abd47dee](/candidate/PANTS-2f34abd47dee) | Human gut | 246 | 32.5 | 0.36 | 92 | 22.36 | EC:A0A8J3U475 |\n| [PANTS-463e7b625bd9](/candidate/PANTS-463e7b625bd9) | Human gut | 254 | 32.4 | 0.28 | 91 | 21.03 | EC:A0A8J3U475 |\n| [PANTS-a6bb2d5e105e](/candidate/PANTS-a6bb2d5e105e) | Human gut | 274 | 32.3 | 0.32 | 89 | 21.31 | EC:A0A2P4UDP6 |\n| [PANTS-01bdff6d9403](/candidate/PANTS-01bdff6d9403) | Human gut | 241 | 32.2 | 0.27 | 91 | 16.57 | EC:A0A6I4WH26 |\n| [PANTS-cf5b61686b69](/candidate/PANTS-cf5b61686b69) | Human gut | 257 | 32.1 | 0.32 | 92 | 21.63 | EC:A0ABT0G6E6 |\n| [PANTS-c7c045f569ed](/candidate/PANTS-c7c045f569ed) | Human gut | 255 | 32.0 | 0.29 | 91 | 22.27 | EC:A0ABQ2QK59 |\n| [PANTS-1d0f0568a672](/candidate/PANTS-1d0f0568a672) | Human gut | 438 | 32.0 | 0.29 | 79 | 22.38 | EC:A0ABN3W1M5 |\n| [PANTS-50a40361d00a](/candidate/PANTS-50a40361d00a) | Human gut | 255 | 31.8 | 0.31 | 92 | 20.66 | EC:A0ABQ4H167 |\n| [PANTS-7d7833871da4](/candidate/PANTS-7d7833871da4) | Human gut | 255 | 31.8 | 0.31 | 92 | 20.62 | EC:A0ABQ4H167 |\n| [PANTS-8e27a78bcb21](/candidate/PANTS-8e27a78bcb21) | Human gut | 344 | 31.8 | 0.38 | 76 | 19.05 | EC:A0A841ED77 |\n| [PANTS-2a4c2788a77a](/candidate/PANTS-2a4c2788a77a) | Marine plastisphere | 364 | 31.7 | 0.28 | 95 | 17.38 | EC:A0ABV5TJR7 |\n| [PANTS-db063e52995d](/candidate/PANTS-db063e52995d) | Human gut | 255 | 31.6 | 0.31 | 93 | 20.71 | EC:A0ABQ4H167 |\n| [PANTS-fb15adb0c776](/candidate/PANTS-fb15adb0c776) | Human gut | 159 | 31.6 | 0.30 | 80 | 20.35 | EC:A0ABV5TJR7 |\n| [PANTS-7c9345861335](/candidate/PANTS-7c9345861335) | Human gut | 325 | 31.6 | 0.29 | 84 | 19.2 | EC:A0A365HAJ8 |\n| [PANTS-6e05e5560d6f](/candidate/PANTS-6e05e5560d6f) | Human gut | 175 | 31.6 | 0.31 | 87 | 20.65 | EC:A0ABW2CZA0 |\n| [PANTS-89f465b7f037](/candidate/PANTS-89f465b7f037) | Human gut | 185 | 31.5 | 0.33 | 91 | 9.32 | EC:A0ABR9P363 |\n| [PANTS-182b45308971](/candidate/PANTS-182b45308971) | Human gut | 246 | 31.5 | 0.28 | 91 | 18.95 | EC:A0A7W3LR11 |\n| [PANTS-c72d05ef5fe6](/candidate/PANTS-c72d05ef5fe6) | Human gut | 428 | 31.3 | 0.31 | 89 | 18.71 | EC:A0A8J3U475 |\n| [PANTS-bff0869594ae](/candidate/PANTS-bff0869594ae) | Human gut | 211 | 31.3 | 0.28 | 93 | 16.7 | EC:A0A6I4WH26 |\n| [PANTS-4fcf277b9b6f](/candidate/PANTS-4fcf277b9b6f) | Human gut | 325 | 31.2 | 0.31 | 85 | 20.52 | EC:A0A365H755 |\n| [PANTS-a8fdac7a3874](/candidate/PANTS-a8fdac7a3874) | Human gut | 358 | 31.2 | 0.30 | 81 | 17.87 | EC:A0A8J3U475 |\n| [PANTS-4fb566497b51](/candidate/PANTS-4fb566497b51) | Human gut | 302 | 31.2 | 0.38 | 90 | 17.19 | EC:A0ABT9QDD4 |\n| [PANTS-75ddf350d4e7](/candidate/PANTS-75ddf350d4e7) | Human gut | 153 | 31.1 | 0.31 | 82 | 24.06 | HGMP04 |\n| [PANTS-0c8bacc3e1d3](/candidate/PANTS-0c8bacc3e1d3) | Human gut | 358 | 31.1 | 0.30 | 81 | 21.88 | EC:A0A8J3U475 |\n| [PANTS-d9416fc7602f](/candidate/PANTS-d9416fc7602f) | Human gut | 358 | 31.0 | 0.30 | 81 | 20.56 | EC:A0A8J3U475 |\n| [PANTS-8731b640e42a](/candidate/PANTS-8731b640e42a) | Human gut | 262 | 30.9 | 0.27 | 89 | 20.3 | EC:A0ABV5TJR7 |\n| [PANTS-5a35d0e3c8c4](/candidate/PANTS-5a35d0e3c8c4) | Human gut | 263 | 30.9 | 0.27 | 90 | 20.3 | EC:A0ABV5TJR7 |\n| [PANTS-78754ad45412](/candidate/PANTS-78754ad45412) | Human gut | 263 | 30.9 | 0.27 | 89 | 20.3 | EC:A0ABV5TJR7 |\n| [PANTS-d209073a753d](/candidate/PANTS-d209073a753d) | Human gut | 263 | 30.9 | 0.27 | 90 | 20.3 | EC:A0ABV5TJR7 |\n| [PANTS-63cbd705ab41](/candidate/PANTS-63cbd705ab41) | Compost | 366 | 30.8 | 0.26 | 73 | 20.54 | EC:A0A6P2BTG6 |\n| [PANTS-39fc3823f1a3](/candidate/PANTS-39fc3823f1a3) | Marine plastisphere | 404 | 30.8 | 0.25 | 88 | 21.68 | PLC:A0A1I6NU60 |\n| [PANTS-277b4af96f4b](/candidate/PANTS-277b4af96f4b) | Human gut | 115 | 30.8 | 0.30 | 95 | 15.42 | EC:A0A8J3U475 |\n| [PANTS-c0f45e49472f](/candidate/PANTS-c0f45e49472f) | Human gut | 260 | 30.7 | 0.29 | 91 | 9.52 | EC:A0ABR9P363 |\n| [PANTS-b604508bed3b](/candidate/PANTS-b604508bed3b) | Human gut | 120 | 30.6 | 0.28 | 92 | 15.5 | EC:A0A8J3U475 |\n| [PANTS-e75a530a6244](/candidate/PANTS-e75a530a6244) | Compost | 354 | 30.5 | 0.28 | 79 | 17.11 | EC:D1A646 |\n| [PANTS-815d14b44812](/candidate/PANTS-815d14b44812) | Compost | 361 | 30.5 | 0.33 | 89 | 24.95 | EC:A0ABP8Q6M0 |\n| [PANTS-7e1295e1f268](/candidate/PANTS-7e1295e1f268) | Human gut | 416 | 30.5 | 0.29 | 90 | 16.91 | EC:A0ABP6KM42 |\n| [PANTS-974378098784](/candidate/PANTS-974378098784) | Human gut | 248 | 30.4 | 0.27 | 87 | 21.12 | EC:A0ABV5TJR7 |\n| [PANTS-a460b8bbbf57](/candidate/PANTS-a460b8bbbf57) | Human gut | 172 | 30.4 | 0.35 | 80 | 18.29 | EC:A0A3D9SNA7 |\n| [PANTS-03dc56a9884e](/candidate/PANTS-03dc56a9884e) | Human gut | 334 | 30.4 | 0.32 | 79 | 19.04 | EC:A0A365H755 |\n| [PANTS-3ff2e9cf031c](/candidate/PANTS-3ff2e9cf031c) | Human gut | 215 | 30.3 | 0.35 | 93 | 22.31 | EC:A0A8J3U475 |\n| [PANTS-bd4dce26ae0c](/candidate/PANTS-bd4dce26ae0c) | Human gut | 304 | 30.3 | 0.32 | 89 | 17.03 | EC:A0ABP6DBE2 |\n| [PANTS-ba80cc5a5982](/candidate/PANTS-ba80cc5a5982) | Human gut | 171 | 30.1 | 0.31 | 81 | 18.44 | EC:A0A243RI45 |\n| [PANTS-e8462e13c76a](/candidate/PANTS-e8462e13c76a) | Human gut | 194 | 30.1 | 0.31 | 90 | 22.74 | HGMP04 |\n| [PANTS-d9f57891fa57](/candidate/PANTS-d9f57891fa57) | Human gut | 273 | 30.1 | 0.33 | 90 | 17.77 | PLC:A0A147KJY8 |\n| [PANTS-06c2ae224565](/candidate/PANTS-06c2ae224565) | Human gut | 196 | 29.9 | 0.30 | 93 | 16.92 | EC:A0ABT4UC71 |\n| [PANTS-f2fc610e03d4](/candidate/PANTS-f2fc610e03d4) | Human gut | 233 | 29.9 | 0.33 | 94 | 18.2 | EC:A0ABV9U430 |\n| [PANTS-55b8ddca06e9](/candidate/PANTS-55b8ddca06e9) | Compost | 316 | 29.8 | 0.45 | 88 | 19.96 | EC:A0A9X2K9B4 |\n| [PANTS-026d5eda227d](/candidate/PANTS-026d5eda227d) | Human gut | 354 | 29.7 | 0.30 | 83 | 17.89 | EC:A0A8J3U475 |\n| [PANTS-f28f303efc41](/candidate/PANTS-f28f303efc41) | Human gut | 243 | 29.7 | 0.29 | 90 | 17.34 | EC:A0ABW1CGM7 |\n| [PANTS-adc06bbd1acb](/candidate/PANTS-adc06bbd1acb) | Human gut | 281 | 29.6 | 0.31 | 91 | 14.51 | PLC:A0A163MA58 |\n| [PANTS-db3be4e6f5fa](/candidate/PANTS-db3be4e6f5fa) | Human gut | 286 | 29.6 | 0.30 | 83 | 19.15 | EC:A0ABP8UBC5 |\n| [PANTS-c5bc74f826bb](/candidate/PANTS-c5bc74f826bb) | Human gut | 286 | 29.6 | 0.30 | 83 | 19.16 | EC:A0ABP8UBC5 |\n| [PANTS-03faca12a583](/candidate/PANTS-03faca12a583) | Human gut | 244 | 29.6 | 0.28 | 92 | 17.34 | EC:A0A5S4FJI0 |\n| [PANTS-547f14279971](/candidate/PANTS-547f14279971) | Human gut | 344 | 29.4 | 0.29 | 82 | 16.9 | EC:A0A7C9NKH2 |\n| [PANTS-bc2e995b8367](/candidate/PANTS-bc2e995b8367) | Human gut | 413 | 29.4 | 0.25 | 92 | 16.33 | EC:A0A3D9SPQ3 |\n| [PANTS-6029d3e53ec5](/candidate/PANTS-6029d3e53ec5) | Human gut | 111 | 29.4 | 0.33 | 93 | 15.64 | EC:A0ABT4UC71 |\n| [PANTS-e7a5b1da49db](/candidate/PANTS-e7a5b1da49db) | Human gut | 257 | 29.4 | 0.33 | 91 | 20.54 | HGMP04 |\n| [PANTS-9d0e7f5a9e7b](/candidate/PANTS-9d0e7f5a9e7b) | Human gut | 353 | 29.3 | 0.35 | 83 | 22.35 | EC:A0A8J3U475 |\n| [PANTS-e858ce93d353](/candidate/PANTS-e858ce93d353) | Human gut | 264 | 29.2 | 0.31 | 89 | 21.56 | EC:A0A7K0C0L9 |\n| [PANTS-2eb9c4b04f0f](/candidate/PANTS-2eb9c4b04f0f) | Human gut | 255 | 29.1 | 0.27 | 91 | 22.28 | EC:A0A3D9SPQ3 |\n| [PANTS-c948848b24c0](/candidate/PANTS-c948848b24c0) | Human gut | 257 | 29.1 | 0.28 | 91 | 20.44 | EC:A0A8J3U475 |\n| [PANTS-0053bec8509d](/candidate/PANTS-0053bec8509d) | Human gut | 199 | 29.1 | 0.36 | 81 | 19.0 | EC:A0ABT4UC71 |\n| [PANTS-023904e83efa](/candidate/PANTS-023904e83efa) | Marine plastisphere | 411 | 29.0 | 0.26 | 88 | 20.99 | EC:A0A5S4H3R9 |\n| [PANTS-f27b6be35620](/candidate/PANTS-f27b6be35620) | Human gut | 413 | 29.0 | 0.28 | 91 | 16.37 | EC:A0A6I4WH26 |\n| [PANTS-b066b18c6b09](/candidate/PANTS-b066b18c6b09) | Human gut | 257 | 28.9 | 0.28 | 91 | 20.51 | EC:A0A8J3U475 |\n| [PANTS-20f45706528e](/candidate/PANTS-20f45706528e) | Human gut | 286 | 28.9 | 0.32 | 85 | 19.18 | EC:A0A4R6UW92 |\n| [PANTS-967c9f9e25ad](/candidate/PANTS-967c9f9e25ad) | Human gut | 123 | 28.9 | 0.32 | 92 | 18.34 | EC:A0ABW1CC56 |\n| [PANTS-851ce33c4be0](/candidate/PANTS-851ce33c4be0) | Human gut | 257 | 28.9 | 0.28 | 91 | 20.54 | EC:A0A8J3U475 |\n| [PANTS-615c4b6e9853](/candidate/PANTS-615c4b6e9853) | Human gut | 257 | 28.9 | 0.28 | 91 | 20.56 | EC:A0A8J3U475 |\n| [PANTS-cae63d3b72c0](/candidate/PANTS-cae63d3b72c0) | Human gut | 227 | 28.8 | 0.30 | 94 | 17.41 | EC:A0A1I0JU31 |\n| [PANTS-b72bddf9f270](/candidate/PANTS-b72bddf9f270) | Human gut | 189 | 28.8 | 0.32 | 94 | 17.52 | EC:A0ABW1CC56 |\n| [PANTS-b25c2f02cdb0](/candidate/PANTS-b25c2f02cdb0) | Human gut | 413 | 28.7 | 0.26 | 91 | 16.47 | EC:A0A3D9SPQ3 |\n| [PANTS-0b41ce291172](/candidate/PANTS-0b41ce291172) | Human gut | 968 | 28.5 | 0.28 | - | - | EC:A0ABN3W1M5 |\n| [PANTS-a66878eeab8c](/candidate/PANTS-a66878eeab8c) | Human gut | 286 | 28.5 | 0.30 | 83 | 19.23 | EC:A0ABW1CGM7 |\n| [PANTS-f610e60c9781](/candidate/PANTS-f610e60c9781) | Human gut | 133 | 28.4 | 0.32 | 94 | 17.37 | EC:A0ABP6DBE2 |\n| [PANTS-5c66a91f3b10](/candidate/PANTS-5c66a91f3b10) | Human gut | 257 | 28.4 | 0.33 | 91 | 20.49 | HGMP04 |\n| [PANTS-db4c1ccd7e1e](/candidate/PANTS-db4c1ccd7e1e) | Landfill | 294 | 28.2 | 0.38 | 89 | 23.92 | EC:A0A9X2K0A4 |\n| [PANTS-2625b759dab1](/candidate/PANTS-2625b759dab1) | Human gut | 274 | 28.2 | 0.33 | 85 | 18.08 | EC:A0A243RI45 |\n| [PANTS-e03b1a21a631](/candidate/PANTS-e03b1a21a631) | Human gut | 286 | 28.2 | 0.30 | 83 | 19.23 | EC:A0ABW1CGM7 |\n| [PANTS-ba573459d404](/candidate/PANTS-ba573459d404) | Human gut | 190 | 28.2 | 0.38 | 88 | 13.61 | EC:A0A7Y9GI10 |\n| [PANTS-2e83b07b2cab](/candidate/PANTS-2e83b07b2cab) | Human gut | 255 | 28.1 | 0.31 | 91 | 22.26 | HGMP04 |\n| [PANTS-da83b7447854](/candidate/PANTS-da83b7447854) | Human gut | 280 | 28.1 | 0.22 | 87 | 20.53 | HGMP04 |\n| [PANTS-9a4b79f42a0d](/candidate/PANTS-9a4b79f42a0d) | Human gut | 968 | 28.1 | 0.28 | - | - | EC:A0A8J3XMP6 |\n| [PANTS-a61efa3f7d0d](/candidate/PANTS-a61efa3f7d0d) | Compost | 332 | 28.0 | 0.27 | 92 | 24.54 | EC:A0A7K0BNG8 |\n| [PANTS-afb5bb5f7ca0](/candidate/PANTS-afb5bb5f7ca0) | Marine plastisphere | 236 | 28.0 | 0.32 | 91 | 21.96 | EC:A0ABV3GLG2 |\n| [PANTS-7d013b473ace](/candidate/PANTS-7d013b473ace) | Human gut | 298 | 28.0 | 0.32 | 89 | 17.5 | EC:A0ABW1CC56 |\n| [PANTS-10eadec83602](/candidate/PANTS-10eadec83602) | Human gut | 202 | 27.9 | 0.34 | 91 | - | EC:A0ABN3T5D2 |\n| [PANTS-7359d27c77d1](/candidate/PANTS-7359d27c77d1) | Marine plastisphere | 262 | 27.8 | 0.24 | 95 | 18.28 | EC:A0ABV9U3U3 |\n| [PANTS-4e8cfc219d31](/candidate/PANTS-4e8cfc219d31) | Human gut | 357 | 27.8 | 0.31 | 84 | 19.6 | EC:A0A7C9NKH2 |\n| [PANTS-197c007c6990](/candidate/PANTS-197c007c6990) | Human gut | 258 | 27.8 | 0.27 | 91 | 19.72 | EC:A0A931F5V8 |\n| [PANTS-1ed30a615aa8](/candidate/PANTS-1ed30a615aa8) | Human gut | 247 | 27.8 | 0.23 | 97 | 19.01 | EC:A0A7W7RJC2 |\n| [PANTS-ad3bc58e12cd](/candidate/PANTS-ad3bc58e12cd) | Human gut | 244 | 27.7 | 0.29 | 91 | 17.44 | EC:A0A5S4FJI0 |\n| [PANTS-3c116ff8ff62](/candidate/PANTS-3c116ff8ff62) | Human gut | 313 | 27.7 | 0.33 | 86 | 21.11 | EC:A0A4R6UW92 |\n| [PANTS-2d72f42e42e9](/candidate/PANTS-2d72f42e42e9) | Human gut | 298 | 27.7 | 0.30 | 88 | 17.5 | EC:A0A9Y2YA38 |\n| [PANTS-014e2f5894b4](/candidate/PANTS-014e2f5894b4) | Human gut | 419 | 27.7 | 0.30 | 91 | 17.06 | EC:A0ABT4UC71 |\n| [PANTS-0d18adef6bca](/candidate/PANTS-0d18adef6bca) | Human gut | 330 | 27.7 | 0.29 | 92 | 23.02 | EC:A0ABV5TJR7 |\n| [PANTS-e7a83a8f86c9](/candidate/PANTS-e7a83a8f86c9) | Human gut | 419 | 27.6 | 0.28 | 91 | 16.35 | EC:A0ABT4UC71 |\n| [PANTS-fe0c48eb7ac7](/candidate/PANTS-fe0c48eb7ac7) | Human gut | 263 | 27.5 | 0.27 | 90 | 20.3 | EC:A0ABV5TJR7 |\n| [PANTS-12326b542314](/candidate/PANTS-12326b542314) | Human gut | 598 | 27.5 | 0.30 | - | - | EC:A0ABR9P363 |\n| [PANTS-cf4872c53051](/candidate/PANTS-cf4872c53051) | Compost | 252 | 27.4 | 0.29 | 91 | 20.59 | EC:A0A365H0G7 |\n| [PANTS-621f9af66105](/candidate/PANTS-621f9af66105) | Human gut | 221 | 27.4 | 0.33 | 93 | 17.9 | EC:A0A7K0C0L9 |\n| [PANTS-ea2b4dde3988](/candidate/PANTS-ea2b4dde3988) | Human gut | 319 | 27.4 | 0.31 | 86 | 17.48 | EC:A0A9Y2YA38 |\n| [PANTS-5a6ea2a7e743](/candidate/PANTS-5a6ea2a7e743) | Human gut | 217 | 27.3 | 0.41 | 94 | 15.6 | EC:A0A7Y9GI10 |\n| [PANTS-511b51761a59](/candidate/PANTS-511b51761a59) | Human gut | 263 | 27.2 | 0.38 | 92 | 15.57 | EC:A0A7Y9GI10 |\n| [PANTS-8cf135a7574f](/candidate/PANTS-8cf135a7574f) | Compost | 242 | 27.1 | 0.32 | 86 | 23.27 | EC:A0A6P2BTG6 |\n| [PANTS-da0e6c8690c3](/candidate/PANTS-da0e6c8690c3) | Human gut | 279 | 27.1 | 0.38 | 88 | 15.53 | EC:A0A7Y9GI10 |\n| [PANTS-faf8027f90b9](/candidate/PANTS-faf8027f90b9) | Human gut | 279 | 27.1 | 0.38 | 87 | 15.53 | EC:A0A7Y9GI10 |\n| [PANTS-57023fa5e8bb](/candidate/PANTS-57023fa5e8bb) | Human gut | 282 | 27.0 | 0.27 | 92 | 20.13 | EC:A0A931F5V8 |\n| [PANTS-2c1b3dd9a31b](/candidate/PANTS-2c1b3dd9a31b) | Human gut | 281 | 27.0 | 0.27 | 92 | 20.15 | EC:A0A931F5V8 |\n| [PANTS-ae60e2adfca2](/candidate/PANTS-ae60e2adfca2) | Human gut | 389 | 27.0 | 0.33 | 74 | 20.23 | EC:A0ABP6KM42 |\n| [PANTS-e16be9b9b0d7](/candidate/PANTS-e16be9b9b0d7) | Human gut | 244 | 26.8 | 0.29 | 91 | 17.35 | EC:A0A1I0JU31 |\n| [PANTS-31dfe9ca56c9](/candidate/PANTS-31dfe9ca56c9) | Human gut | 283 | 26.8 | 0.30 | 86 | 15.75 | EC:D7AYZ5 |\n| [PANTS-27b922447349](/candidate/PANTS-27b922447349) | Landfill | 328 | 26.7 | 0.36 | 89 | 25.42 | EC:A0ABN3D3P5 |\n| [PANTS-e00040e72d57](/candidate/PANTS-e00040e72d57) | Marine plastisphere | 217 | 26.6 | 0.26 | 94 | 21.58 | EC:A0ABP6CMB2 |\n| [PANTS-f1b711067e4b](/candidate/PANTS-f1b711067e4b) | Human gut | 174 | 26.5 | 0.29 | 84 | 20.32 | EC:A0A8J3U475 |\n| [PANTS-721626f6c0b3](/candidate/PANTS-721626f6c0b3) | Compost | 309 | 26.4 | 0.30 | 80 | 24.48 | EC:A0ABW2PJI5 |\n| [PANTS-bebd392634fe](/candidate/PANTS-bebd392634fe) | Marine plastisphere | 646 | 26.3 | 0.30 | - | - | EC:A0ABV3GLG2 |\n| [PANTS-bffd49e5c909](/candidate/PANTS-bffd49e5c909) | Human gut | 244 | 26.3 | 0.23 | 92 | 17.34 | PLC:A0A1H1SER0 |\n| [PANTS-5f4d81a0d013](/candidate/PANTS-5f4d81a0d013) | Human gut | 260 | 26.2 | 0.34 | 91 | 25.88 | EC:A0A7Y9GI10 |\n| [PANTS-af0564b4feb9](/candidate/PANTS-af0564b4feb9) | Human gut | 275 | 26.1 | 0.28 | 88 | 18.95 | EC:A0A8J3PRQ9 |\n| [PANTS-2b7a719e90aa](/candidate/PANTS-2b7a719e90aa) | Compost | 665 | 25.9 | 0.24 | - | - | EC:A0ABQ2QK59 |\n| [PANTS-735381b0ff71](/candidate/PANTS-735381b0ff71) | Compost | 665 | 25.9 | 0.24 | - | - | EC:A0ABQ2QK59 |\n| [PANTS-81740296ccb3](/candidate/PANTS-81740296ccb3) | Marine plastisphere | 217 | 25.8 | 0.28 | 94 | 21.66 | EC:A0ABP7V404 |\n| [PANTS-3ca70320fa22](/candidate/PANTS-3ca70320fa22) | Human gut | 196 | 25.8 | 0.27 | 85 | 18.98 | HGMP02 |\n| [PANTS-db1a5e67e9b9](/candidate/PANTS-db1a5e67e9b9) | Human gut | 136 | 25.7 | 0.28 | 93 | 15.92 | EC:A0ABP8UBC5 |\n| [PANTS-9c095950212e](/candidate/PANTS-9c095950212e) | Compost | 619 | 25.6 | 0.29 | - | - | EC:A0A365H0G7 |\n| [PANTS-a25095d9671b](/candidate/PANTS-a25095d9671b) | Human gut | 289 | 25.6 | 0.30 | 91 | 17.0 | EC:A0A3D9SPQ3 |\n| [PANTS-d97711269d0a](/candidate/PANTS-d97711269d0a) | Human gut | 225 | 25.6 | 0.29 | 92 | 15.82 | EC:A0A8J3U475 |\n| [PANTS-f1d49354fc7e](/candidate/PANTS-f1d49354fc7e) | Human gut | 289 | 25.6 | 0.30 | 92 | 16.98 | EC:A0A3D9SPQ3 |\n| [PANTS-f093eb68fdbf](/candidate/PANTS-f093eb68fdbf) | Human gut | 335 | 25.5 | 0.30 | 86 | 22.97 | EC:A0ABW3EQF3 |\n| [PANTS-3102d0956951](/candidate/PANTS-3102d0956951) | Human gut | 277 | 25.4 | 0.33 | 86 | 19.55 | EC:A0A239AX86 |\n| [PANTS-fd024a17bdb9](/candidate/PANTS-fd024a17bdb9) | Human gut | 279 | 25.4 | 0.33 | 87 | 19.67 | EC:A0A239AX86 |\n| [PANTS-b474f5c0c23f](/candidate/PANTS-b474f5c0c23f) | Human gut | 277 | 25.4 | 0.33 | 87 | 19.61 | EC:A0A239AX86 |\n| [PANTS-f8ac589ee490](/candidate/PANTS-f8ac589ee490) | Landfill | 379 | 25.3 | 0.28 | 89 | 22.49 | EC:A0A7W8EHH8 |\n| [PANTS-4a54cdf0424c](/candidate/PANTS-4a54cdf0424c) | Human gut | 267 | 25.3 | 0.29 | 89 | 16.43 | EC:A0A8J3U475 |\n| [PANTS-e8e629209abb](/candidate/PANTS-e8e629209abb) | Compost | 219 | 25.2 | 0.34 | 92 | 21.08 | EC:A0A8J3X7Y9 |\n| [PANTS-90c9e4679f85](/candidate/PANTS-90c9e4679f85) | Landfill | 165 | 25.2 | 0.49 | 90 | 15.23 | EC:A0A7W3MTE8 |\n| [PANTS-6eb128fdb77f](/candidate/PANTS-6eb128fdb77f) | Human gut | 281 | 25.2 | 0.23 | 85 | 21.54 | HGMP04 |\n| [PANTS-e21c0d81eb8e](/candidate/PANTS-e21c0d81eb8e) | Human gut | 283 | 25.0 | 0.29 | 88 | 15.85 | EC:A0A8J3U475 |\n| [PANTS-10dd2fe71095](/candidate/PANTS-10dd2fe71095) | Human gut | 208 | 25.0 | 0.27 | 92 | 20.19 | EC:A0A7W3N2I8 |\n| [PANTS-cabf92291864](/candidate/PANTS-cabf92291864) | Human gut | 244 | 25.0 | 0.30 | 92 | 17.43 | EC:A0A5S4FJI0 |\n| [PANTS-6fcba6cf5714](/candidate/PANTS-6fcba6cf5714) | Marine plastisphere | 217 | 24.9 | 0.28 | 94 | 18.94 | EC:A0ABP7V404 |\n| [PANTS-fafcc253b6ae](/candidate/PANTS-fafcc253b6ae) | Human gut | 290 | 24.9 | 0.28 | 91 | 16.97 | EC:A0A3D9SRX0 |\n| [PANTS-52726c2cf362](/candidate/PANTS-52726c2cf362) | Human gut | 290 | 24.8 | 0.29 | 90 | 17.61 | EC:A0A8J3U475 |\n| [PANTS-bca3c3c4cafb](/candidate/PANTS-bca3c3c4cafb) | Human gut | 281 | 24.6 | 0.22 | 86 | 20.18 | HGMP04 |\n| [PANTS-135b4fb9afd0](/candidate/PANTS-135b4fb9afd0) | Human gut | 221 | 24.6 | 0.29 | 86 | 21.67 | EC:A0ABQ2QK59 |\n| [PANTS-537549659b1d](/candidate/PANTS-537549659b1d) | Human gut | 267 | 24.5 | 0.30 | 89 | 16.44 | EC:A0A8J3U475 |\n| [PANTS-88a39bd6c239](/candidate/PANTS-88a39bd6c239) | Human gut | 269 | 24.5 | 0.30 | 90 | 22.85 | EC:A0ABW3EQF3 |\n| [PANTS-49d48a9af071](/candidate/PANTS-49d48a9af071) | Human gut | 264 | 24.5 | 0.31 | 92 | 19.01 | EC:A0A7W3N2I8 |\n| [PANTS-fdf467010c61](/candidate/PANTS-fdf467010c61) | Human gut | 596 | 24.4 | 0.27 | - | - | EC:A0A7W3LQW1 |\n| [PANTS-4d56964ac059](/candidate/PANTS-4d56964ac059) | Human gut | 212 | 24.4 | 0.29 | 91 | 16.54 | EC:A0ABR9P363 |\n| [PANTS-66309d2e1809](/candidate/PANTS-66309d2e1809) | Human gut | 180 | 24.3 | 0.36 | 93 | 16.65 | EC:A0A1V0A5L2 |\n| [PANTS-801c81062ac1](/candidate/PANTS-801c81062ac1) | Human gut | 221 | 24.2 | 0.28 | 91 | 22.06 | EC:A0ABP6QLT6 |\n| [PANTS-6793a07eba17](/candidate/PANTS-6793a07eba17) | Marine plastisphere | 217 | 24.1 | 0.27 | 94 | 18.85 | EC:A0A8J3PRQ9 |\n| [PANTS-095131806de6](/candidate/PANTS-095131806de6) | Human gut | 277 | 24.1 | 0.32 | 87 | 19.6 | EC:A0A239AX86 |\n| [PANTS-4a9d3f2f9379](/candidate/PANTS-4a9d3f2f9379) | Human gut | 277 | 24.1 | 0.32 | 87 | 19.58 | EC:A0A239AX86 |\n| [PANTS-11071a4ce617](/candidate/PANTS-11071a4ce617) | Marine plastisphere | 445 | 24.0 | 0.26 | 84 | 20.88 | EC:A0ABW2CZA0 |\n| [PANTS-213c1a8021d9](/candidate/PANTS-213c1a8021d9) | Compost | 619 | 24.0 | 0.29 | - | - | EC:A0A365H0G7 |\n| [PANTS-fc7a76552900](/candidate/PANTS-fc7a76552900) | Human gut | 195 | 24.0 | 0.28 | 83 | 13.61 | EC:A0A1M6M213 |\n| [PANTS-1f9b21ff90c0](/candidate/PANTS-1f9b21ff90c0) | Compost | 189 | 23.9 | 0.28 | 93 | 18.91 | EC:A0ABN3QL32 |\n| [PANTS-cddf707b24ff](/candidate/PANTS-cddf707b24ff) | Human gut | 327 | 23.9 | 0.31 | 82 | 24.03 | EC:A0A7W3N2I8 |\n| [PANTS-6df6e44c325d](/candidate/PANTS-6df6e44c325d) | Human gut | 281 | 23.9 | 0.23 | 86 | 20.46 | HGMP04 |\n| [PANTS-89df35bc0336](/candidate/PANTS-89df35bc0336) | Human gut | 231 | 23.9 | 0.36 | 94 | 17.77 | EC:A0A1V0A5L2 |\n| [PANTS-59e264535e51](/candidate/PANTS-59e264535e51) | Marine plastisphere | 166 | 23.8 | 0.26 | 86 | 18.25 | EC:A0ABP6CMB2 |\n| [PANTS-a434808ce88d](/candidate/PANTS-a434808ce88d) | Compost | 216 | 23.8 | 0.35 | 83 | 21.75 | EC:A0ABP6BRL7 |\n| [PANTS-628b04a09598](/candidate/PANTS-628b04a09598) | Human gut | 210 | 23.8 | 0.30 | 92 | 18.62 | EC:A0A239AX86 |\n| [PANTS-0c3f586995eb](/candidate/PANTS-0c3f586995eb) | Human gut | 312 | 23.7 | 0.27 | 86 | 17.44 | EC:A0ABY3LNA1 |\n| [PANTS-ec9743320c0e](/candidate/PANTS-ec9743320c0e) | Human gut | 290 | 23.7 | 0.29 | 90 | 17.64 | EC:A0A8J3U475 |\n| [PANTS-944f93870d58](/candidate/PANTS-944f93870d58) | Human gut | 303 | 23.6 | 0.27 | 87 | 17.44 | EC:A0ABY3LNA1 |\n| [PANTS-84ed462a2f48](/candidate/PANTS-84ed462a2f48) | Human gut | 287 | 23.5 | 0.28 | 86 | 15.78 | EC:A0ABY6YXF2 |\n| [PANTS-0dc11824c5b8](/candidate/PANTS-0dc11824c5b8) | Human gut | 196 | 23.5 | 0.28 | 86 | 19.12 | HGMP02 |\n| [PANTS-2f4c000c41cc](/candidate/PANTS-2f4c000c41cc) | Human gut | 280 | 23.5 | 0.22 | 85 | 20.97 | HGMP04 |\n| [PANTS-37425c2194fc](/candidate/PANTS-37425c2194fc) | Marine plastisphere | 165 | 23.4 | 0.28 | 88 | 18.37 | EC:A0ABP7IBD0 |\n| [PANTS-3678bc760af9](/candidate/PANTS-3678bc760af9) | Human gut | 268 | 23.4 | 0.25 | 93 | 20.09 | HGMP04 |\n| [PANTS-2aaf42d1c98a](/candidate/PANTS-2aaf42d1c98a) | Human gut | 287 | 23.4 | 0.29 | 86 | 15.78 | EC:A0A543NNU7 |\n| [PANTS-5e3fc8d44e3e](/candidate/PANTS-5e3fc8d44e3e) | Human gut | 263 | 23.4 | 0.24 | 91 | 20.09 | HGMP04 |\n| [PANTS-cb1b8795d217](/candidate/PANTS-cb1b8795d217) | Human gut | 281 | 23.3 | 0.23 | 86 | 20.41 | HGMP04 |\n| [PANTS-1254e38f4a3e](/candidate/PANTS-1254e38f4a3e) | Human gut | 280 | 23.3 | 0.22 | 85 | 21.09 | HGMP04 |\n| [PANTS-a3c5b86ef330](/candidate/PANTS-a3c5b86ef330) | Human gut | 281 | 23.3 | 0.23 | 88 | 19.64 | HGMP04 |\n| [PANTS-88949e1b709e](/candidate/PANTS-88949e1b709e) | Compost | 316 | 23.2 | 0.49 | 90 | 20.32 | PLC:F4QTD6 |\n| [PANTS-7183008ecedf](/candidate/PANTS-7183008ecedf) | Human gut | 268 | 23.2 | 0.33 | 89 | 20.34 | EC:A0A239AX86 |\n| [PANTS-5a9b797e4c9c](/candidate/PANTS-5a9b797e4c9c) | Marine plastisphere | 150 | 23.1 | 0.28 | 82 | 25.42 | EC:A0ABV9E7U1 |\n| [PANTS-7c38672ce65d](/candidate/PANTS-7c38672ce65d) | Marine plastisphere | 197 | 23.1 | 0.25 | 90 | 21.52 | EC:A0ABP6CMB2 |\n| [PANTS-fb53ed63ed40](/candidate/PANTS-fb53ed63ed40) | Human gut | 271 | 23.1 | 0.27 | 91 | 18.37 | EC:A0A7W9IBT4 |\n| [PANTS-c440a93fe548](/candidate/PANTS-c440a93fe548) | Human gut | 269 | 23.1 | 0.28 | 86 | 15.81 | EC:A0A1M6M213 |\n| [PANTS-2e3a6ef0e9f2](/candidate/PANTS-2e3a6ef0e9f2) | Human gut | 281 | 23.1 | 0.23 | 86 | 19.88 | HGMP04 |\n| [PANTS-3db14c36b6f1](/candidate/PANTS-3db14c36b6f1) | Human gut | 269 | 23.1 | 0.28 | 86 | 24.64 | EC:A0A1M6M213 |\n| [PANTS-80b626126d56](/candidate/PANTS-80b626126d56) | Human gut | 283 | 23.0 | 0.28 | 86 | 24.56 | EC:A0A1M6M213 |\n| [PANTS-01e97631d770](/candidate/PANTS-01e97631d770) | Human gut | 283 | 23.0 | 0.28 | 86 | 24.77 | EC:A0A1M6M213 |\n| [PANTS-78e600a85733](/candidate/PANTS-78e600a85733) | Human gut | 283 | 23.0 | 0.28 | 86 | 24.75 | EC:A0A1M6M213 |\n| [PANTS-c40aa8377f27](/candidate/PANTS-c40aa8377f27) | Human gut | 249 | 23.0 | 0.30 | 96 | 20.21 | HGMP01 |\n| [PANTS-4511faeedd74](/candidate/PANTS-4511faeedd74) | Marine plastisphere | 217 | 22.9 | 0.26 | 93 | 18.76 | EC:A0ABP7V404 |\n| [PANTS-8c53696b8231](/candidate/PANTS-8c53696b8231) | Compost | 223 | 22.9 | 0.31 | 91 | 17.68 | EC:A0ABP6KM42 |\n| [PANTS-eed0edd3ca26](/candidate/PANTS-eed0edd3ca26) | Human gut | 268 | 22.9 | 0.33 | 89 | 20.37 | EC:A0A239AX86 |\n| [PANTS-d137453554f8](/candidate/PANTS-d137453554f8) | Human gut | 241 | 22.9 | 0.29 | 97 | - | PLC:C0ZUA7 |\n| [PANTS-c2067f8dae0a](/candidate/PANTS-c2067f8dae0a) | Human gut | 241 | 22.9 | 0.29 | 97 | - | PLC:C0ZUA7 |\n| [PANTS-252273fc627f](/candidate/PANTS-252273fc627f) | Marine plastisphere | 191 | 22.8 | 0.28 | 92 | 22.24 | EC:A0ABP7IBD0 |\n| [PANTS-6aff5403a2f3](/candidate/PANTS-6aff5403a2f3) | Human gut | 151 | 22.8 | 0.33 | 86 | 22.59 | EC:A0A7Y9GI10 |\n| [PANTS-c161d3214edd](/candidate/PANTS-c161d3214edd) | Compost | 180 | 22.7 | 0.31 | 91 | 18.13 | EC:A0ABT4UC71 |\n| [PANTS-6ec99030c250](/candidate/PANTS-6ec99030c250) | Human gut | 147 | 22.7 | 0.32 | 89 | 5.15 | EC:A0A7W3LQW1 |\n| [PANTS-06cb2d0daf14](/candidate/PANTS-06cb2d0daf14) | Human gut | 281 | 22.6 | 0.24 | 87 | 20.69 | HGMP02 |\n| [PANTS-d96f36ed1866](/candidate/PANTS-d96f36ed1866) | Human gut | 281 | 22.6 | 0.22 | 85 | 19.73 | HGMP04 |\n| [PANTS-bae249be37c1](/candidate/PANTS-bae249be37c1) | Compost | 139 | 22.4 | 0.30 | 75 | - | EC:A0ABW2Y2K9 |\n| [PANTS-fcac028beb3f](/candidate/PANTS-fcac028beb3f) | Human gut | 382 | 22.4 | 0.29 | 84 | 21.27 | EC:A0ABQ2QK59 |\n| [PANTS-030040cc5744](/candidate/PANTS-030040cc5744) | Human gut | 393 | 22.4 | 0.27 | 91 | 18.85 | EC:A0ABP6KM42 |\n| [PANTS-6ad686db7847](/candidate/PANTS-6ad686db7847) | Human gut | 281 | 22.3 | 0.22 | 88 | 19.87 | HGMP04 |\n| [PANTS-91949094d000](/candidate/PANTS-91949094d000) | Human gut | 281 | 22.3 | 0.23 | 88 | 19.7 | HGMP04 |\n| [PANTS-4da1ec3bcf6d](/candidate/PANTS-4da1ec3bcf6d) | Human gut | 281 | 22.2 | 0.23 | 86 | 19.68 | HGMP04 |\n| [PANTS-df47c122e34f](/candidate/PANTS-df47c122e34f) | Human gut | 302 | 22.2 | 0.34 | 90 | 17.86 | EC:A0ABN3KHS9 |\n| [PANTS-91f0d93aa64f](/candidate/PANTS-91f0d93aa64f) | Human gut | 281 | 22.2 | 0.22 | 88 | 20.66 | HGMP04 |\n| [PANTS-3859139782db](/candidate/PANTS-3859139782db) | Compost | 181 | 22.1 | 0.41 | 86 | 14.68 | EC:A0A7Z0EUU2 |\n| [PANTS-7a33851266f6](/candidate/PANTS-7a33851266f6) | Human gut | 260 | 22.1 | 0.26 | 94 | 17.53 | EC:A0ABQ2QK59 |\n| [PANTS-d8f29f519930](/candidate/PANTS-d8f29f519930) | Human gut | 437 | 22.0 | 0.28 | 88 | 21.7 | HGMP04 |\n| [PANTS-a2177b1905c1](/candidate/PANTS-a2177b1905c1) | Human gut | 279 | 21.9 | 0.33 | 88 | 15.17 | EC:A0A7Y9GI10 |\n| [PANTS-d4b9c6d93303](/candidate/PANTS-d4b9c6d93303) | Human gut | 279 | 21.9 | 0.35 | 88 | 15.14 | EC:A0A7Y9GI10 |\n| [PANTS-f49e5308d5b7](/candidate/PANTS-f49e5308d5b7) | Human gut | 185 | 21.9 | 0.25 | 84 | 18.87 | EC:A0A5N6AZJ9 |\n| [PANTS-bd41207a1ddb](/candidate/PANTS-bd41207a1ddb) | Human gut | 279 | 21.8 | 0.35 | 88 | 15.12 | EC:A0A7Y9GI10 |\n| [PANTS-daa02a7afb86](/candidate/PANTS-daa02a7afb86) | Human gut | 432 | 21.8 | 0.28 | 90 | 24.3 | EC:A0ABU2KYZ7 |\n| [PANTS-67f5d71ad4c3](/candidate/PANTS-67f5d71ad4c3) | Human gut | 132 | 21.8 | 0.28 | 81 | 5.46 | HGMP02 |\n| [PANTS-b784acc1758a](/candidate/PANTS-b784acc1758a) | Human gut | 279 | 21.7 | 0.32 | 88 | 15.12 | EC:A0A7Y9GI10 |\n| [PANTS-2d991f6f68e8](/candidate/PANTS-2d991f6f68e8) | Human gut | 111 | 21.6 | 0.29 | 95 | 10.94 | EC:A0ABU2KYZ7 |\n| [PANTS-f824db26a301](/candidate/PANTS-f824db26a301) | Human gut | 181 | 21.6 | 0.32 | 82 | 19.39 | EC:A0A7W3LQW1 |\n| [PANTS-0d43598204b7](/candidate/PANTS-0d43598204b7) | Human gut | 216 | 21.6 | 0.34 | 92 | 15.14 | EC:A0A7Y9GI10 |\n| [PANTS-536b4c720ca1](/candidate/PANTS-536b4c720ca1) | Human gut | 212 | 21.6 | 0.31 | 93 | 20.6 | EC:A0ABV5TJR7 |\n| [PANTS-5db5f1c43378](/candidate/PANTS-5db5f1c43378) | Marine plastisphere | 280 | 21.5 | 0.24 | 92 | 21.2 | EC:A0A7W3LQW1 |\n| [PANTS-4158856f9a1e](/candidate/PANTS-4158856f9a1e) | Marine plastisphere | 217 | 21.5 | 0.26 | 94 | 18.76 | EC:A0ABP7V404 |\n| [PANTS-90e52a1f6d3f](/candidate/PANTS-90e52a1f6d3f) | Marine plastisphere | 217 | 21.4 | 0.26 | 94 | 18.84 | EC:A0ABP6CMB2 |\n| [PANTS-510751a72a3e](/candidate/PANTS-510751a72a3e) | Human gut | 367 | 21.3 | 0.26 | 89 | 26.31 | EC:A0A5M3W626 |\n| [PANTS-9589d5b6a2aa](/candidate/PANTS-9589d5b6a2aa) | Human gut | 279 | 21.3 | 0.32 | 88 | 15.12 | EC:A0A7Y9GI10 |\n| [PANTS-8031476a0565](/candidate/PANTS-8031476a0565) | Landfill | 321 | 21.1 | 0.24 | 86 | 23.53 | EC:D1A646 |\n| [PANTS-d67014b88662](/candidate/PANTS-d67014b88662) | Human gut | 336 | 21.0 | 0.27 | 62 | 25.99 | EC:A0ABU2KYZ7 |\n| [PANTS-04fa4d3083a7](/candidate/PANTS-04fa4d3083a7) | Compost | 291 | 20.9 | 0.24 | 85 | 16.68 | EC:A0A8J3PRQ9 |\n| [PANTS-22337b7d2be9](/candidate/PANTS-22337b7d2be9) | Compost | 176 | 20.8 | 0.42 | 81 | - | EC:A0ABW2CTX2 |\n| [PANTS-f0259e7d0feb](/candidate/PANTS-f0259e7d0feb) | Human gut | 132 | 20.8 | 0.31 | 90 | 16.47 | EC:A0ABW1CC56 |\n| [PANTS-a476a11c65cd](/candidate/PANTS-a476a11c65cd) | Marine plastisphere | 247 | 20.7 | 0.27 | 91 | 20.15 | EC:A0A853BQM0 |\n| [PANTS-b7a075e773c7](/candidate/PANTS-b7a075e773c7) | Human gut | 302 | 20.7 | 0.31 | 90 | 17.86 | EC:A0ABN3KHS9 |\n| [PANTS-e62652375433](/candidate/PANTS-e62652375433) | Human gut | 170 | 20.7 | 0.27 | 85 | 16.14 | EC:A0A239F952 |\n| [PANTS-f36ac3596244](/candidate/PANTS-f36ac3596244) | Human gut | 144 | 20.6 | 0.30 | 82 | 19.64 | EC:A0ABW1CC56 |\n| [PANTS-505ed3992559](/candidate/PANTS-505ed3992559) | Compost | 706 | 20.5 | 0.25 | - | - | EC:A0A543NNU7 |\n| [PANTS-4427a6e80bbc](/candidate/PANTS-4427a6e80bbc) | Human gut | 196 | 20.5 | 0.22 | 92 | 25.64 | EC:A0ABW1NJ96 |\n| [PANTS-04bdd4c7dfe7](/candidate/PANTS-04bdd4c7dfe7) | Human gut | 287 | 20.5 | 0.28 | 90 | 19.71 | EC:A0A239F952 |\n| [PANTS-fcdc8407dc64](/candidate/PANTS-fcdc8407dc64) | Marine plastisphere | 251 | 20.4 | 0.27 | 93 | 20.43 | EC:A0A853BQM0 |\n| [PANTS-3d07be75e95e](/candidate/PANTS-3d07be75e95e) | Human gut | 456 | 20.4 | 0.24 | - | - | EC:A0A3D9SUD5 |\n| [PANTS-d7e3c8946cef](/candidate/PANTS-d7e3c8946cef) | Compost | 171 | 20.2 | 0.46 | 80 | 17.25 | EC:A0A9X2K9B4 |\n| [PANTS-7a6ef043c5bc](/candidate/PANTS-7a6ef043c5bc) | Human gut | 833 | 20.2 | 0.32 | - | - | HGMP04 |\n| [PANTS-b52e98fe0d1a](/candidate/PANTS-b52e98fe0d1a) | Compost | 859 | 20.1 | 0.23 | - | - | EC:A0A4V0ZKB9 |\n| [PANTS-e9dd4211ae00](/candidate/PANTS-e9dd4211ae00) | Landfill | 277 | 20.1 | 0.26 | 97 | 25.84 | EC:A0A7W3LQW1 |\n| [PANTS-b7f1f487e981](/candidate/PANTS-b7f1f487e981) | Compost | 225 | 20.1 | 0.31 | 94 | 22.33 | EC:A0A7W3LR11 |\n| [PANTS-ab579d41ee64](/candidate/PANTS-ab579d41ee64) | Marine plastisphere | 217 | 20.0 | 0.27 | 95 | 21.56 | EC:A0ABP6CMB2 |\n| [PANTS-99b3cc949168](/candidate/PANTS-99b3cc949168) | Marine plastisphere | 209 | 19.9 | 0.26 | 92 | 18.61 | EC:A0ABP7V404 |\n| [PANTS-dca92ff03a16](/candidate/PANTS-dca92ff03a16) | Human gut | 315 | 19.9 | 0.29 | 89 | 19.46 | EC:A0ABW1CC56 |\n| [PANTS-6c458f78f2a5](/candidate/PANTS-6c458f78f2a5) | Human gut | 282 | 19.9 | 0.25 | 92 | 17.04 | EC:A0ABV5TJR7 |\n| [PANTS-322ac7fe6068](/candidate/PANTS-322ac7fe6068) | Marine plastisphere | 281 | 19.7 | 0.28 | 96 | 23.51 | EC:D1A646 |\n| [PANTS-e7d71575fd7a](/candidate/PANTS-e7d71575fd7a) | Compost | 287 | 19.6 | 0.29 | 87 | 26.12 | EC:A0A7W3R6L2 |\n| [PANTS-967c3c1a60c8](/candidate/PANTS-967c3c1a60c8) | Human gut | 185 | 19.5 | 0.24 | 81 | 25.05 | EC:A0ABR9JL03 |\n| [PANTS-37c78e67f49b](/candidate/PANTS-37c78e67f49b) | Human gut | 107 | 19.5 | 0.33 | 88 | 18.24 | EC:A0A4R6UW92 |\n| [PANTS-9abd5a449c28](/candidate/PANTS-9abd5a449c28) | Human gut | 315 | 19.4 | 0.29 | 89 | 19.45 | EC:A0ABW1CC56 |\n| [PANTS-69ab4a1f0f3a](/candidate/PANTS-69ab4a1f0f3a) | Marine plastisphere | 282 | 19.2 | 0.25 | 98 | 17.48 | EC:A0A243RTG5 |\n| [PANTS-3b50eab19302](/candidate/PANTS-3b50eab19302) | Human gut | 248 | 19.2 | 0.38 | 96 | 20.2 | HGMP01 |\n| [PANTS-56e67190cb4b](/candidate/PANTS-56e67190cb4b) | Human gut | 223 | 19.1 | 0.33 | 93 | 23.12 | EC:A0ABT4TFD0 |\n| [PANTS-3a8ed4240aea](/candidate/PANTS-3a8ed4240aea) | Human gut | 198 | 19.0 | 0.26 | 72 | 22.36 | EC:A0ABT4UC71 |\n| [PANTS-d489dafdf4a5](/candidate/PANTS-d489dafdf4a5) | Compost | 711 | 18.9 | 0.29 | - | - | EC:A0ABV0AIQ8 |\n| [PANTS-ffc51cae4d1c](/candidate/PANTS-ffc51cae4d1c) | Human gut | 198 | 18.9 | 0.29 | 67 | 22.27 | EC:A0ABT4UC71 |\n| [PANTS-83d34c0d2849](/candidate/PANTS-83d34c0d2849) | Marine plastisphere | 217 | 18.8 | 0.29 | 93 | 21.69 | EC:A0ABP6CMB2 |\n| [PANTS-227076ee89fd](/candidate/PANTS-227076ee89fd) | Human gut | 262 | 18.8 | 0.27 | 90 | 20.47 | EC:A0ABQ4H167 |\n| [PANTS-8be37f023823](/candidate/PANTS-8be37f023823) | Human gut | 219 | 18.8 | 0.29 | 88 | 19.66 | EC:A0A7W3N2I8 |\n| [PANTS-5cf065cb5a9d](/candidate/PANTS-5cf065cb5a9d) | Human gut | 241 | 18.8 | 0.29 | 88 | 20.21 | HGMP04 |\n| [PANTS-79e49dd125cd](/candidate/PANTS-79e49dd125cd) | Human gut | 213 | 18.7 | 0.29 | 89 | 19.75 | EC:A0A7W3N2I8 |\n| [PANTS-6223a5b2c9a6](/candidate/PANTS-6223a5b2c9a6) | Compost | 155 | 18.6 | 0.34 | 85 | 19.93 | EC:A0A365H6Y8 |\n| [PANTS-526841a1bb94](/candidate/PANTS-526841a1bb94) | Human gut | 227 | 18.1 | 0.31 | 90 | 15.02 | EC:A0A7Y9GI10 |\n| [PANTS-4e8c96d0bf74](/candidate/PANTS-4e8c96d0bf74) | Human gut | 279 | 18.0 | 0.30 | 87 | 14.89 | EC:A0A7Y9GI10 |\n| [PANTS-3b852a348a40](/candidate/PANTS-3b852a348a40) | Human gut | 279 | 18.0 | 0.30 | 87 | 14.88 | EC:A0A7Y9GI10 |\n| [PANTS-0782c349e210](/candidate/PANTS-0782c349e210) | Human gut | 279 | 17.9 | 0.31 | 87 | 14.87 | EC:A0A7Y9GI10 |\n| [PANTS-d59a4b67a3f4](/candidate/PANTS-d59a4b67a3f4) | Human gut | 279 | 17.9 | 0.31 | 87 | 14.87 | EC:A0A7Y9GI10 |\n| [PANTS-5fdfa5b1dde4](/candidate/PANTS-5fdfa5b1dde4) | Human gut | 279 | 17.9 | 0.30 | 88 | 14.88 | EC:A0A7Y9GI10 |\n| [PANTS-255c8d257b46](/candidate/PANTS-255c8d257b46) | Compost | 228 | 17.7 | 0.33 | 95 | 22.36 | EC:A0A7W3LR11 |\n| [PANTS-7f349f48a680](/candidate/PANTS-7f349f48a680) | Human gut | 279 | 17.7 | 0.31 | 87 | 14.87 | EC:A0A7Y9GI10 |\n| [PANTS-f3de269e5718](/candidate/PANTS-f3de269e5718) | Human gut | 281 | 17.6 | 0.25 | 88 | 20.57 | HGMP02 |\n| [PANTS-e287d88885a7](/candidate/PANTS-e287d88885a7) | Human gut | 281 | 17.6 | 0.24 | 89 | 20.46 | HGMP02 |\n| [PANTS-b28982189bc0](/candidate/PANTS-b28982189bc0) | Human gut | 279 | 17.6 | 0.24 | 84 | 14.97 | EC:A0A8J3U475 |\n| [PANTS-e275f7c69959](/candidate/PANTS-e275f7c69959) | Human gut | 289 | 17.5 | 0.24 | 95 | - | EC:A0ABR9JL03 |\n| [PANTS-9a0280fdec45](/candidate/PANTS-9a0280fdec45) | Human gut | 279 | 17.5 | 0.34 | 86 | 14.92 | EC:A0A7Y9GI10 |\n| [PANTS-473e39e1c7e5](/candidate/PANTS-473e39e1c7e5) | Marine plastisphere | 217 | 17.3 | 0.26 | 95 | 21.54 | EC:A0ABP7V404 |\n| [PANTS-b6ae493e5bbf](/candidate/PANTS-b6ae493e5bbf) | Human gut | 268 | 17.3 | 0.38 | 86 | 15.11 | EC:A0A3D9SRX0 |\n| [PANTS-6a8e5e202c83](/candidate/PANTS-6a8e5e202c83) | Human gut | 279 | 17.3 | 0.38 | 85 | 15.19 | EC:A0A3D9SRX0 |\n| [PANTS-0927b15394b6](/candidate/PANTS-0927b15394b6) | Human gut | 268 | 17.3 | 0.38 | 85 | 15.05 | EC:A0A3D9SRX0 |\n| [PANTS-b0a23fb0fab1](/candidate/PANTS-b0a23fb0fab1) | Marine plastisphere | 309 | 17.2 | 0.24 | 94 | 23.64 | EC:A0A7W3LQW1 |\n| [PANTS-3569c6b6e3b6](/candidate/PANTS-3569c6b6e3b6) | Human gut | 280 | 17.2 | 0.38 | 85 | 15.09 | EC:A0A3D9SRX0 |\n| [PANTS-92d1d186aff3](/candidate/PANTS-92d1d186aff3) | Human gut | 279 | 17.2 | 0.32 | 86 | 14.86 | EC:A0A7Y9GI10 |\n| [PANTS-e9f9026fb0e6](/candidate/PANTS-e9f9026fb0e6) | Human gut | 250 | 16.7 | 0.25 | 89 | 19.2 | EC:A0A8J3U475 |\n| [PANTS-8f2d257645c8](/candidate/PANTS-8f2d257645c8) | Human gut | 224 | 16.7 | 0.30 | 94 | 23.1 | EC:A0ABQ2QK59 |\n| [PANTS-97f7168bbc5c](/candidate/PANTS-97f7168bbc5c) | Compost | 211 | 16.4 | 0.28 | 93 | 21.74 | EC:A0ABP6CMB2 |\n| [PANTS-4cd4b8146925](/candidate/PANTS-4cd4b8146925) | Compost | 314 | 16.3 | 0.36 | 89 | 19.07 | EC:A0ABY4L3P7 |\n| [PANTS-eb99f4ce70c1](/candidate/PANTS-eb99f4ce70c1) | Human gut | 279 | 16.3 | 0.30 | 88 | 14.98 | EC:A0A7Y9GI10 |\n| [PANTS-6f7dc1a8b70e](/candidate/PANTS-6f7dc1a8b70e) | Marine plastisphere | 217 | 16.2 | 0.26 | 95 | 21.56 | EC:A0ABP6CMB2 |\n| [PANTS-7d0bfa27d810](/candidate/PANTS-7d0bfa27d810) | Compost | 243 | 16.2 | 0.28 | 95 | 19.78 | EC:D2B3P3 |\n| [PANTS-806da9f1c7b5](/candidate/PANTS-806da9f1c7b5) | Marine plastisphere | 143 | 15.9 | 0.26 | 82 | 18.29 | EC:A0ABP7V404 |\n| [PANTS-984ffcfbe01d](/candidate/PANTS-984ffcfbe01d) | Human gut | 279 | 15.9 | 0.34 | 88 | 20.38 | EC:A0A3D9SRX0 |\n| [PANTS-b85af680dc68](/candidate/PANTS-b85af680dc68) | Human gut | 279 | 15.9 | 0.29 | 86 | 14.9 | EC:A0A7Y9GI10 |\n| [PANTS-4463fd5001cf](/candidate/PANTS-4463fd5001cf) | Human gut | 279 | 15.9 | 0.34 | 88 | 19.59 | EC:A0A3D9SRX0 |\n| [PANTS-8cdd90e84203](/candidate/PANTS-8cdd90e84203) | Human gut | 286 | 15.6 | 0.23 | 91 | 23.78 | EC:A0A7W3LQW1 |\n| [PANTS-afdfc2296b42](/candidate/PANTS-afdfc2296b42) | Human gut | 157 | 15.5 | 0.32 | 84 | 15.18 | EC:A0A919R7S7 |\n| [PANTS-73b6da61d224](/candidate/PANTS-73b6da61d224) | Human gut | 179 | 12.8 | 0.27 | 93 | 24.47 | EC:A0A4R5FXI3 |\n| [PANTS-8255cbab3349](/candidate/PANTS-8255cbab3349) | Human gut | 223 | 12.7 | 0.26 | 95 | 25.94 | EC:A0A3N1D7Y0 |\n| [PANTS-ce1fce39ad6e](/candidate/PANTS-ce1fce39ad6e) | Landfill | 267 | 12.6 | 0.23 | 88 | 20.95 | EC:A0A7W9GDV7 |\n| [PANTS-ec1db9554425](/candidate/PANTS-ec1db9554425) | Human gut | 188 | 12.5 | 0.25 | 90 | 23.38 | PLC:W7JED3 |\n| [PANTS-79af25165177](/candidate/PANTS-79af25165177) | Human gut | 242 | 12.3 | 0.22 | 94 | 23.78 | EC:A0A918XC99 |\n| [PANTS-72829784c250](/candidate/PANTS-72829784c250) | Human gut | 284 | 11.8 | 0.25 | 94 | 26.17 | EC:A0A3N1D7Y0 |", "# Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\nAuthors: Kong, L.; Lee, J.H.; Doores, K.J.; Murin, C.D.; Julien, J.-P.; McBride, R.; Liu, Y.; Marozsan, A.; Cupo, A.; Klasse, P.-J.; Hoffenberg, S.; Caulfield, M.; King, C.R.; Hua, Y.; Le, K.M.; Khayat, R.; Deller, M.C.; Clayton, T.; Tien, H.; Feizi, T.; Sanders, R.W.; Paulson, J.C.; Moore, J.P.; Stanfield, R.L.; Burton, D.R.; Ward, A.B.; Wilson, I.A.\nSource: Nature Structural & Molecular Biology, 2013\n\nTo elucidate the interaction with gp120, we tested PGT 135 Fab complexed with a number of gp120 constructs and additional ligands for crystallization. The additional ligands included soluble CD4 receptor and Fabs recognizing gp120 to increase the chances of crystallization. Crystals of a quaternary complex (i.e. four protein components) consisting of PGT 135 Fab, JR-FL core gp120, CD4 D1D2 and 17b Fab (Fig. 1) diffracted to 3.1 \u00c5 resolution (Table 1), while unliganded Fab 135 crystals diffracted to 1.75 \u00c5 (Table 1). The JR-FL gp120 monomeric core bound to PGT 135 Fab with high affinity [82 nM by ITC (Supplementary Table 1)], but the corresponding virus was only weakly neutralized by PGT 135 IgG unless specific mutations were introduced into the V1-V2 loops, which are >25 \u00c5 away from the epitope in the monomeric gp120 structure and probably affect binding indirectly by impacting trimer interactions (Fig. 2). Nevertheless, the weak neutralization that we observed correlates with an approximate \u03bcM binding affinity for the Env trimer. A number of bnAb-Env complexes in this affinity range have been structurally elucidated, e.g. b12 with a stabilized gp120 core construct and PG9 with V1-V2 loop scaffolds, and these structures are consistent with many published findings on Env biology.\n\nThe crystal structure shows that CD4, 17b and PGT 135 are bound to gp120 in nearly orthogonal orientations (Fig. 1a). After partially deglycosylating the gp120/PGT 135 complex by endoglycosidase H (EndoH) prior to crystallization, we observed only single N-acetylglucosamine (GlcNAc) moieties in the gp120 electron density at Asn339, Asn362, Asn448, Asn262 and Asn276, whereas high mannose (Man) glycans interacting with PGT 135 were protected and we observed density for GlcNAc2Man2 at Asn386 (N386), GlcNAc2Man6 at Asn332 and GlcNAc2Man8 at Asn392 (N392). When compared to its unliganded structure, PGT 135 does not undergo any large conformational changes [0.4 \u00c5 C\u03b1 root-mean-square deviation (RMSD) between Fab variable domains] (Supplementary Fig. 1a-b). The gp120 structure is also highly similar in both the PGT 135-CD4-17b-gp120(JRFL) and the CD4-17b-gp120(HXBC) complexes (0.5 \u00c5 C\u03b1 RMSD), suggesting that PGT 135 does not induce any further conformational changes when it binds to gp120. This inference was corroborated by ITC data: the \u0394G, \u0394H and T\u0394S values for PGT 135 binding to unliganded gp120 and to the gp120-CD4 complex were nearly equivalent (Supplementary Table 1). Therefore, evidence that PGT 135 neutralizes virus by inducing conformational changes in gp120 or by direct inhibition of CD4 binding is lacking. However, the PGT 135 neutralization mechanism still remains unclear although we note that the PGT 135 epitope also involves residues from the V4 loop. In another study, an anti-FLAG antibody, which neutralizes recombinant virus containing a FLAG insert in the V4 loop, appears to function via nonspecific steric coverage of the virion instead of blocking functionally important surfaces such as the CD4 binding site.\n\nPGT 135 engages gp120 primarily via a long 18-residue CDR H3 loop, which has 10 amino-acid insertions after position 100 (n.b. the average human CDR H3 is 13 residues using the Kabat definition of H95-H102), and an extended CDR H1 loop containing a rare 5 amino-acid insertion not present in the predicted germline sequence (Supplementary Fig. 2); both CDR loops protrude approximately 20 \u00c5 from the VH surface, comparable to the length of high mannose N-linked glycans. The VH CDR loops in unliganded PGT 135 converge to form a highly hydrophobic horseshoe-shaped surface consisting of Trp34 from CDR H1, Trp53 from H2, and Val100-FMLVP-Ile100F from H3 (Supplementary Fig. 1d). A salt bridge between Arg54 on the tip of H2 and Glu33 near the top of the H1 loop insertion helps hold these loops in place. These features remain intact after binding to gp120 (Supplementary Fig. 1c).\n\nPGT 135 buries a complex epitope consisting of 438 \u00c52 of protein and 1010 \u00c52 of glycan surface on gp120 (Fig. 1 and Supplementary Table 2), which is larger than the typical antibody footprint (\u223c700-900 \u00c52). 83 percent of the protein-protein buried surface involves side chains on gp120 (Supplementary Table 2) in contrast to PG9 and PGT 128 that rely primarily on backbone interactions with V1-V2 and V3, respectively, to form main-chain \u03b2-strand associations. The CDR H1 and H3 tips primarily contact gp120 \u03b2-strand 19, which connects the V4 loop to \u03b2-strand 20, as part of the bridging sheet (Fig. 1b), and adjacent \u03b2-strand 13, between V3 and \u03b1-helix 2, where His330 is particularly important for binding and neutralization (Supplementary Table 2-3). On the gp120 V4 loop, Glu409 interacts with CDRs L2 and H3, forming a protein-protein interaction zone that is discontinuous from the rest of the contacts. The CDR H1 insertion allows access to the protein surface around the Asn386 glycan (GlcNAc2Man2), which separates the PGT 135 epitope from the CD4 binding site. Trp34 on the tip of the CDR H1 insertion interacts with the Asn386 glycan. Coincidentally, bnAb b12 also contacts residues near the Asn386 glycan via CDR H3 Trp100, which occupies a similar position to PGT 135 H1 Trp34. Deletion of the H1 insertion significantly reduced PGT 135 binding and neutralization of different HIV-1 isolates (Fig. 2b-c), suggesting that this extended loop is important for the antibody-antigen interaction.", "# GitHub README: CODSWALLOP\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/CODSWALLOP, 2026\n\n**It is a shadow, not the thing.** The map is two dimensions of an object that has as many as the family has representatives. Across the 70 built families the first two principal coordinates carry a median of **65%** of the positive eigenvalue mass, and as little as **33%** (the beta-2 adrenergic receptor, tau, insulin, calmodulin). Two structures drawn close together are close in the projection; on a low-variance family they may be further apart than they look. The heatmap in the Structures panel is the higher-fidelity view of the same matrix, which is why both exist.\n\n**1 &minus; TM is not a Euclidean distance**, so the Gram matrix has negative eigenvalues: a median of **19%** of the positive mass, up to **49%**. They are clamped to zero rather than the build failing, which is the standard treatment and is honest as far as it goes, but it means the embedding is an approximation of a geometry that does not strictly exist. Reading distance qualitatively (near, far, off on its own) is safe. Reading it as a metric is not, and there is nothing on the panel to stop you: use the TM-score matrix itself for anything quantitative.\n\nThis is also why the panel is fitted the way it is. Those two points sit at x = -1.0 while every other representative spans -0.036 to +0.046, so fitting the panel to the furthest point put the entire native family inside 5.4% of its width: a map of one curiosity and a smudge. The fit uses a high quantile with headroom instead, and pins anything beyond it to the rim with its true distance in the tooltip, so the outlier stays visible and stays findable without costing the other 1,686 structures their map.\n\n## \ud83e\uddf1 Stack\n\n| Layer | Choice | Note |\n|---|---|---|\n| Backend | Python 3.12, Flask 3.1, gunicorn 26 | No compiled toolchain needed on the droplet in phase 1 |\n| Cache | SQLite (WAL) | Raw API responses plus assembled families, both with a one-week TTL |\n| Front end | Vanilla JS, hand-rolled SVG | The map needs direct control of individual nodes for the cross-highlight pulse |\n| Tables | Tabulator 6.3.1 | Vendored, skinned to the theme tokens |\n| Type | Archivo, Inter, JetBrains Mono | Self-hosted woff2, latin and latin-ext only, 244 kB total |\n| Serving | nginx, certbot | Same droplet pattern as AlphaFraud |\n\n## \ud83d\udce1 Data sources\n\nAll public, none require a key. Please cite them, not this tool, when the data does the work.\n\n| Source | Endpoint | Used for |\n|---|---|---|\n| RCSB Search API v2 | `search.rcsb.org/rcsbsearch/v2/query` | Family resolution: sequence (mmseqs2), UniProt accession, Pfam/InterPro annotation, full text |\n| RCSB Data API | `data.rcsb.org/graphql` | Batched entry, polymer entity, assembly and chem_comp metadata |\n| UniProt REST | `rest.uniprot.org/` | Canonical sequence, protein name, organism, gene names |\n| InterPro API | `ebi.ac.uk/interpro/api/` | Pfam and InterPro accession names and types |\n\nPhases 2 to 4 add the RCSB 1D Coordinates, Alignment and file services, the PDBe REST and Graph APIs, EMDB and the AlphaFold DB.\n\n## \ud83e\uddea Tests\n\n```bash\n./.venv/bin/pip install pytest\n./.venv/bin/python -m pytest\n```\n\n183 tests, none of which touch the network. They are written against the bugs that actually shipped rather than for coverage: a fusion partner being shredded by the aligner, a chimera diffed against the wrong reference, `Bis-Tris` being tallied as `Tris`, a saturating \"never\" statistic, a cached `None` being refetched forever, a beacon that no page requested, a family outvoting its own subject, a disulphide bond expanded into 122 of them, a decorative spiral that read as structure, a cluster labeller whose preferred input was empty on every real request, an offset applied twice so that contacts landed on residues the protein does not have, and a gene search that answered VEGFR with twelve fruit flies.\n\n## \ud83d\udd27 Installation\n\n```bash\ngit clone https://github.com/bellcheddar/CODSWALLOP.git\ncd CODSWALLOP\npython3 -m venv .venv\n./.venv/bin/pip install -r requirements.txt\n./.venv/bin/python CODSWALLOP.py init\n```\n\n## \ud83d\ude80 Usage\n\n```bash\n# Run the development server (defaults to 127.0.0.1:8006)\n./.venv/bin/python CODSWALLOP.py serve\n\n# Assemble a family from the shell, without a browser\n./.venv/bin/python CODSWALLOP.py build P00918\n./.venv/bin/python CODSWALLOP.py build \"TEM-1 beta-lactamase\"\n./.venv/bin/python CODSWALLOP.py build PF00062 --refresh\n\n# Cache housekeeping\n./.venv/bin/python CODSWALLOP.py stats\n./.venv/bin/python CODSWALLOP.py purge\n```\n\n| Command | Does |\n|---|---|\n| `init` | Create the cache database and apply any column migrations |\n| `serve` | Run the development server (`--bind`, `--debug`) |\n| `build <query>` | Assemble a family and print a summary. Exits 2 with a candidate list if the query is ambiguous |\n| `stats` | Cache statistics: families, entries, entities, cached responses, database size |\n| `purge` | Drop expired cache entries |\n\n### Routes\n\n| Route | Does |\n|---|---|\n| `/` | The closed drawer: one search field |\n| `/lookup?q=` | Resolve the input, then redirect to the family or ask which one you meant |\n| `/f/<slug>` | The family. Paints complete if cached, paints its filing state and fetches if not |\n| `/f/<slug>/dossier` | A self-contained family report: one file, no fetches, prints to PDF |\n| `/api/family/<slug>` | Assemble (or serve) a family as JSON. The front end's own payload, not a contract |\n| `/api/v1/families` | What this instance has filed. The entry point for another program |\n| `/api/v1/family/<slug>` | **The stable summary.** A few kilobytes, versioned schema, CORS open. Never assembles on demand |\n| `/api/stats` | Cache statistics, and the per-app hit beacon for the mdeller.com launcher |\n| `/healthz` | Liveness |\n\n## \ud83d\udee0\ufe0f Configuration\n\nCopy `.env.example` to `.env`. Everything has a sensible default in `codswallop/config.py`.", "# Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor\nAuthors: Kong, L.; Ju, B.; Chen, Yajing; He, L.; Ren, L.; Liu, J.; Hong, K.; Su, B.; Wang, Z.; Ozorowski, G.; Ji, X.; Hua, Y.; Chen, Yanli; Deller, M.C.; Hao, Y.; Feng, Y.; Garces, F.; Wilson, R.; Dai, K.; O'Dell, S.; McKee, K.; Mascola, J.R.; Ward, A.B.; Wyatt, R.T.; Li, Y.; Wilson, I.A.; Zhu, J.; Shao, Y.\nSource: Immunity, 2016\n\nAntibody library preparation and PGM sequencing of DRVI01 B cell repertoires were performed as previously described (He et al., 2014). In brief, the 5\u2032-RACE PCR protocol was used to prepare unbiased antibody H, \u03ba, and \u03bb chain libraries for DRVI01, while the gene-specific multiplex PCR protocol was used to prepare antibody libraries for five LTNPs. Details are described in the Supplemental Experimental Procedures.\n\n### Bioinformatics Analysis of Repertoire Sequencing Data\n\nThe Antibodyomics 1.0 pipeline (He et al., 2014; Wu et al., 2011) was used to process the NGS data. Identity-divergence 2D analysis, cross-donor phylogenetic analysis, and CDR3-based lineage tracing were used to dissect the details of DRVIA7 lineage at the three time points studied. Details are described in the Supplemental Experimental Procedures.\n\n### HIV-1 Neutralization Assays\n\nNeutralization assays were performed on TZM-bl reporter cells using various panels of pseudoviruses as previously described (Wu et al., 2010). Neutralization curves were fit by a nonlinear regression analysis to derive IC50 values for DRVI01 serum mapping and functional validation of DRVIA mAbs and variants. Details are described in the Supplemental Experimental Procedures.\n\n## Accession Numbers\n\nCoordinates and structure factors for Fab DRVIA7 and its complex with gp120 are deposited with the PDB under accession codes 5CD3 and 5CD5, respectively. The Fab DRVIA7:JRFL SOSIP.664 trimer and Fab DRVIA7H-VRC01L:JRFL SOSIP.664 trimer EM reconstructions are deposited in the EMDB under accession codes EMD-6372 and EMD-6373, respectively.\n\n## Author Contributions\n\nProject design by L.K., B.J., Yajing Chen, L.H., L.R., J.L., Y.L., I.A.W., J.Z., and Y.S.; neutralizing antibody screening by K.H., L.R., and Y.S.; antibody isolation by Yajing Chen, R.W., and Y.L.; X-ray work and analysis by L.K., Y. Hua, M.C.D., and F.G.; EM work by G.O. and A.B.W.; antibody NGS, lineage tracing, and SGA analysis by L.H. and J.Z.; synthesis of antibody genes and construction of expression plasmids by B.J., X.J., and Z.W.; neutralization on the global and DRVI panels by K.H. and L.R.; expression and purification of antibodies and Env proteins by L.K., J.L., X.J., Yajing Chen, R.W., K.D., R.T.W., and Y. Hao; Env SGA performed by K.H., Yanli Chen, and Y.F.; antibody Fab construct design by Yajing Chen and R.W.; antibody binding assays by Yajing Chen, L.K., R.W., B.J., and X.J.; epidemiology and cohort maintenance by B.S. and K.H.; VRC01-resistant HIV-1 neutralization by S.O'D., K.M., and J.R.M.; neutralization on the 8-virus panel by Yajing Chen and R.W.; manuscript written by L.K., Y.L., I.A.W., J.Z., and Y.S. All authors were asked to comment on the manuscript. This is TSRI manuscript number 29101.\n\n## Acknowledgments\n\nWe are very grateful to R. Stanfield for helpful discussions; M. Elsliger for computer support; H. Tien for crystallization screening; A. Eroshkin for help with calculating N-linked glycosylation frequencies; J. Hou, P. Li, and M. Zhu for help with antibody binding assays; D. Li, H. Liang, and H. Peng for help with measuring patient CD4+ T cell counts; Q. Zhao for help with viral load assays; J. Wu and X. Ding for help with patient blood sample collection; and J.P. Verenini for manuscript formatting. X-ray data sets were collected at the Advanced Photon Source (APS) beamline 23ID-B. Use of the APS was supported by the US Department of Energy, Basic Energy Sciences, Office of Science, under contract no. DE-AC02-06CH11357. Electron microscopy data were collected at the Scripps Research Institute EM Facility. This work was supported by the National Major projects for Infectious Diseases Control and Prevention (2012ZX10001008) (Y.S.), by the International AIDS Vaccine Initiative Neutralizing Antibody Center and CAVD, by the Center for HIV/AIDS Vaccine Immunology and Immunogen Discovery (CHAVI-ID UM1 AI100663) (A.B.W., I.A.W., J.Z.), by the HIV Vaccine Research and Design (HIVRAD) program (P01 AI110657) (A.B.W., I.A.W.), by the Joint Center of Structural Genomics (JCSG) funded by the NIH NIGMS, Protein Structure Initiative (U54 GM094586) (I.A.W.), by the SKLID Development grant (2012SKLID103) (Y.S.), by NIH NIAID grants R01 AI102766 (Y.L.) and AI084817 (I.A.W., A.B.W.), and by an NIH NIAID development grant from the Center for AIDS Research at the University of California, San Diego (P30AI36214) (Y.L., J.Z.). A portion of this work was supported by the NIH CIPRA grant (U19AI51915) (Y.S.) and by an American Foundation for AIDS Research Mathilde Krim Fellowship in Basic Biomedical Research (L.K.). Y.S. is grateful to Dennis Burton for hosting his stay at The Scripps Research Institute to initiate this collaborative study.\n\n## Supplemental Information\n\nSupplemental Information includes Supplemental Experimental Procedures, five tables, and five figures and can be found with this article online at http://dx.doi.org/10.1016/j.immuni.2016.03.006.\n\n## References\n\n*(reference list omitted)*", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\nMre11 nuclease plays a central role in the repair of cytotoxic and mutagenic DNA double-strand breaks. As X-ray structural information has been available only for the Pyrococcus furiosus enzyme (PfMre11), the conserved and variable features of this nuclease across the domains of life have not been experimentally defined. Our crystal structure and biochemical studies demonstrate that TM1635 from Thermotoga maritima, originally annotated as a putative nuclease, is an Mre11 endo/exonuclease (TmMre11) and the first such structure from eubacteria. TmMre11 and PfMre11 display similar overall structures, despite sequence identity in the twilight zone of only approximately 20%. However, they differ substantially in their DNA-specificity domains and in their dimeric organization. Residues in the nuclease domain are highly conserved, but those in the DNA-specificity domain are not. The structural differences likely affect how Mre11 from different organisms recognize and interact with single-stranded DNA, double-stranded DNA and DNA hairpin structures during DNA repair. The TmMre11 nuclease active site has no bound metal ions, but is conserved in sequence and structure with the exception of a histidine that is important in PfMre11 nuclease activity. Nevertheless, biochemical characterization confirms that TmMre11 possesses both endonuclease and exonuclease activities on single-stranded and double-stranded DNA substrates, respectively.\n\n## Open and closed conformations of two SpoIIAA-like proteins (YP_749275.1 and YP_001095227.1) provide insights into membrane association and ligand binding (2010)\n\n**Authors:** Kumar A, Lomize A, Jin KK, Carlton D, Miller MD, Jaroszewski L, Abdubek P, Astakhova T, Axelrod HL, Chiu HJ, Clayton T, Das D, Deller MC, Duan L, Feuerhelm J, Grant JC, Grzechnik A, Han GW, Klock HE, Knuth MW, Kozbial P, Krishna SS, Marciano D, McMullan D, Morse AT, Nigoghossian E, Okach L, Reyes R, Rife CL, Sefcovic N, Tien HJ, Trame CB, van den Bedem H, Weekes D, Xu Q, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309109042481. PMID: 20944218.\n\nThe crystal structures of the proteins encoded by the YP_749275.1 and YP_001095227.1 genes from Shewanella frigidimarina and S. loihica, respectively, have been determined at 1.8 and 2.25\u2005\u00c5 resolution, respectively. These proteins are members of a novel family of bacterial proteins that adopt the \u03b1/\u03b2 SpoIIAA-like fold found in STAS and CRAL-TRIO domains. Despite sharing 54% sequence identity, these two proteins adopt distinct conformations arising from different dispositions of their \u03b12 and \u03b13 helices. In the `open' conformation (YP_001095227.1), these helices are 15\u2005\u00c5 apart, leading to the creation of a deep nonpolar cavity. In the `closed' structure (YP_749275.1), the helices partially unfold and rearrange, occluding the cavity and decreasing the solvent-exposed hydrophobic surface. These two complementary structures are reminiscent of the conformational switch in CRAL-TRIO carriers of hydrophobic compounds. It is suggested that both proteins may associate with the lipid bilayer in their `open' monomeric state by inserting their amphiphilic helices, \u03b12 and \u03b13, into the lipid bilayer. These bacterial proteins may function as carriers of nonpolar substances or as interfacially activated enzymes.\n\n## Structure of BT_3984, a member of the SusD/RagB family of nutrient-binding molecules (2010)\n\n**Authors:** Bakolitsa C, Xu Q, Rife CL, Abdubek P, Astakhova T, Axelrod HL, Carlton D, Chen C, Chiu HJ, Clayton T, Das D, Deller MC, Duan L, Ellrott K, Farr CL, Feuerhelm J, Grant JC, Grzechnik A, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Krishna SS, Kumar A, Lam WW, Marciano D, McMullan D, Miller MD, Morse AT, Nigoghossian E, Nopakun A, Okach L, Puckett C, Reyes R, Tien HJ, Trame CB, van den Bedem H, Weekes D, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309110032999. PMID: 20944222.\n\nThe crystal structure of the Bacteroides thetaiotaomicron protein BT_3984 was determined to a resolution of 1.7\u2005\u00c5 and was the first structure to be determined from the extensive SusD family of polysaccharide-binding proteins. SusD is an essential component of the sus operon that defines the paradigm for glycan utilization in dominant members of the human gut microbiota. Structural analysis of BT_3984 revealed an N-terminal region containing several tetratricopeptide repeats (TPRs), while the signature C-terminal region is less structured and contains extensive loop regions. Sequence and structure analysis of BT_3984 suggests the presence of binding interfaces for other proteins from the polysaccharide-utilization complex.\n\n## Structure of Bacteroides thetaiotaomicron BT2081 at 2.05\u2005\u00c5 resolution: the first structural representative of a new protein family that may play a role in carbohydrate metabolism (2010)\n\n**Authors:** Yeh AP, Abdubek P, Astakhova T, Axelrod HL, Bakolitsa C, Cai X, Carlton D, Chen C, Chiu HJ, Chiu M, Clayton T, Das D, Deller MC, Duan L, Ellrott K, Farr CL, Feuerhelm J, Grant JC, Grzechnik A, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Krishna SS, Kumar A, Lam WW, Marciano D, McMullan D, Miller MD, Morse AT, Nigoghossian E, Nopakun A, Okach L, Puckett C, Reyes R, Tien HJ, Trame CB, van den Bedem H, Weekes D, Wooten T, Xu Q, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309110028228. PMID: 20944224.", "# Approaches to automated protein crystal harvesting\nAuthors: Deller, M. C.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 70, 133\u2013155 (IYCr crystallization series), 2014\n\nThe Protein Crystal Growth Demonstration System was built by the University of Washington with the Boeing Missile and Space Division as a prototype for handling protein crystals aboard the International Space Station, where reduced convection and sedimentation are claimed to give better crystals. Its main aim was to demonstrate telerobotics controlled over standard internet protocols, freeing crew time and ISS bandwidth. It uses microcapillaries to harvest from standard 24-well CrysChem plates with a Mini-DD robot on a linear motion base with 25 \u00b5m positioning resolution, a syringe-pump-driven microcapillary end-effector and three cameras for target, workspace and capillary-tip views. It was reported only as a testbed harvesting 0.5 mm plastic cubes, with performance on real crystals unknown. The Crystal Preparation Prime Item was built by Oceaneering Space Systems under contract to the University of Alabama for the ISS X-ray Crystallography Facility, one of four Prime Items alongside crystal growth, X-ray diffraction and command/control units. Crystals are loaded at high density into 12 cartridge tools that also hold harvesting loops; 12 pipettes in two opposed banks of six form a liquid\u2013liquid bridge of 0.127 to 1.27 cm that serves as the working area, with liquid slowly pumped until a crystal is isolated. A camera and an OM3198 micromanipulator let an earth-based scientist harvest, after which the crystal is snap-cooled in a conduction cooling unit and stored in a 24-crystal carousel at \u2212180\u00b0C, or kept in a microcapillary for room-temperature study. CPPI is rugged and reliable but its dependence on the liquid\u2013liquid bridge makes it unsuitable for harvesting directly from growth plates; it led to the design of the ACTOR robotics system from Rigaku.\n\n### 2.2.3. Two-fingered micro-hand (\"chopsticks\")\n\nDeveloped by KEK with the National Institute of Advanced Industrial Science and Technology as part of an attempt to fully automate structure solution, the chopsticks system uses a two-fingered micro-hand whose main-finger and sub-finger operate independently. Harvesting is operator-assisted through a joystick: the crystal is held carefully between the two fingers and then placed in a traditional microloop for flash-cooling and data collection. Algorithms control camera focus, detect crystal position and determine the best orientation for handling, and many operations are automatic, including opening the fingers to a size appropriate for the target, closing them to hold the crystal, and transferring it to the microloop. The system achieved a 90% success rate harvesting 8 \u00b5m glass beads, which the authors consider respectable given that the beads are much smaller than typical protein crystals in the 10\u2013300 \u00b5m range. Two advantages are emphasized: the contact pressure on the crystal is precisely controlled, reducing physical and structural damage, and the fingers approach in an open configuration, reducing the tendency to chase the crystal around the drop that afflicts microloop approaches because of fluid flow and bow-wave effects.\n\n### 2.2.4\u20132.2.5. Microcapillary and micropipette systems: ACH and CARESS\n\nThe Automated Crystal Harvester was designed at The Ohio State University as part of a platform for automating membrane-protein crystallization, and was among the first systems to address manipulation of nanolitre volumes of highly viscous lipidic cubic phase. It was also the first to abandon microloop harvesting in favour of a capillary-loop coordinating mechanism, in which the crystal is first drawn into a microcapillary and then transferred to a conventional microloop for flash-cooling and data collection. Vision- and focus-based object tracking follows the loop, capillary and crystal during mounting; drops under 10 \u00b5l in 72-well microbatch plates are visualized on a Nikon Eclipse E400 with an X\u2013Y stage under MAC 2000 control giving 100 nm positioning, while a micropump supplies the negative and positive pressure for aspiration and dispensing. The Columbia Automated Robotic Environment for Streak Seeding was among the first semi-autonomous systems built specifically for HTPX/SG use, at the Northeast Structural Genomics Consortium. It was initially designed to automate streak-seeding of 96-well plates using custom single-crystal silicon microshovels, achieving 6.5 wells per minute with results comparable to manual protocols, using an MP-285 micropositioner with three degrees of freedom and 40 nm positioning resolution. It was later extended to harvesting by mounting a glass micropipette actuated by a CellTram Vario micro-injector capable of 2 nl aspirate and dispense steps, with the user guiding the injector through visual feedback loops until the crystal is drawn in, then dispensing it into a microloop resting in a microbridge of cryoprotectant.\n\n### 2.2.6. Universal Micromanipulation Robot (UMR)", "# Structure of the Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Region 412-423 Bound to the Broadly Neutralizing Antibody AP33\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [4G6A](https://www.rcsb.org/structure/4G6A)\n* **Release Date:** 2012-09-26\n* **Method:** X-RAY DIFFRACTION 2.501 \u00c5\n* **Organisms:** Hepacivirus hominis, Homo sapiens\n* **Macromolecules:** AP33 Heavy Chain, AP33 Light chain, E2 peptide (proteins)\n* **Publication:** [Kong, L., et al. (2012). Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33. Journal of Virology, 86, 13085-13088.](https://journals.asm.org/doi/10.1128/jvi.01939-12)\n* **Blog Post:** [Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33](https://marcdeller.com/structure-of-hepatitis-c-virus-envelope-glycoprotein-e2-antigenic-site-412-to-423-in-complex-with-antibody-ap33/)\n\n13. Description: This structure reveals the crystal structure of broadly neutralizing antibody AP33 bound to HCV E2 envelope glycoprotein antigenic site 412-423, establishing two distinct modes of antibody binding that can neutralize diverse HCV strains. The comparison with bnAb HCV1 shows different angles of approach and subtle \u03b2-hairpin conformational variations, providing critical structural insights for pan-genotypic HCV vaccine development.", "# alphabetti.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n## alphabetti.mdeller.com/\n\nESMFold \u00b7 ESM-2 650M\n\u00b7 no alignment used\n\n[WebLogo 3](https://github.com/gecrooks/weblogo)\n3.9.0 \u00b7 MIT \u00b7 Crooks\n*et al.* 2004 \u00b7\n[PMC419797](https://pmc.ncbi.nlm.nih.gov/articles/PMC419797/)\n\nRendered by [WebLogo 3](https://github.com/gecrooks/weblogo), unmodified.\nThis is the figure, exactly as the command line would draw it.\nDownload it as PDF, EPS or SVG from the panel.\n\nstarting\n\nSequence or UniProt accession\n\nDownload\n\nSequence\nhover to highlight \u00b7 click to fly there\n\n## alphabetti.mdeller.com/about\n\n## What it does\n\nALPHABETTI draws a protein's own single-letter amino acid codes in three\ndimensions, positioned and oriented along a predicted backbone. It is a\nsequence logo that has been wrapped around the structure it describes.\n\nThe part that is actually new: **there is no multiple sequence alignment\nanywhere in this pipeline**. A conventional logo measures conservation by\ncounting residues in an alignment column, which means finding homologues,\naligning them, and inheriting every bias in that process. ALPHABETTI asks a\nprotein language model what it expects at each position instead, having\nfirst masked that position so the model cannot read the answer off its own\ninput. What you are looking at is a model's expectation, mapped onto a\nmodel's structure.\n\n## How a stack is built\n\nAt each position the model gives a probability for all twenty canonical\namino acids. From that:\n\n```\nH_i      = -sum_a p_a log2(p_a)          Shannon entropy, bits\nR_i      = log2(20) - H_i               information content, max 4.322 bits\nheight_a = p_a * R_i                    per-letter height, bits\n```\n\nThe letters at a position are stacked tallest-first along the residue's own\nside-chain direction, each standing on the one below, so **the total height\nof a stack is that position's information content**. That is the property\nthat makes the picture mean something, and it is checked by a unit test\nagainst hand-calculated values.\n\nNo small-sample entropy correction is applied. The usual correction in a\nsequence logo compensates for estimating a distribution from a finite\nnumber of aligned sequences; here there is no alignment and no sample, so\ncorrecting would mean subtracting a bias that does not exist.\n\n## Models\n\n| Role | Model | Checkpoint | Reference |\n| --- | --- | --- | --- |\n| Structure | ESMFold | `facebook/esmfold_v1` | Lin et al., Science 379:1123 (2023) |\n| Language | ESM-2 650M | `facebook/esm2_t33_650M_UR50D` | Lin et al., Science 379:1123 (2023) |\n\nBoth run on a Hugging Face ZeroGPU Space rather than on this server. The\nserver this app is served from has 3.9 GB of memory; the ESMFold checkpoint\nalone is 7.9 GB of weights.\n\n## HOGWASH, and the one tab that uses an alignment\n\nEvery tab above derives from a protein language model and uses no\nalignment at all. **HOGWASH is the exception**: it counts what\nevolution actually did in a column of aligned homologues, which is the\nolder and better-established way to measure conservation. Both are in the\napp so they can be compared.\n\nThe numbers and the drawing are both\n[WebLogo 3](https://github.com/gecrooks/weblogo)'s,\nMIT licensed and used unmodified: the column counts, the composition\npriors, the small-sample correction, the unit conversions, the colour\nschemes and the rendered output. ALPHABETTI supplies the input plumbing\nand nothing else. Please cite it as\n**Crooks GE, Hon G, Chandonia JM, Brenner SE (2004) WebLogo: a sequence\nlogo generator. *Genome Research* 14:1188-1190**\n([doi:10.1101/gr.849004](https://doi.org/10.1101/gr.849004),\n[PMC419797](https://pmc.ncbi.nlm.nih.gov/articles/PMC419797/)),\nand the idea itself as Schneider & Stephens (1990). Full licence text\nand every other dependency is in `THIRD-PARTY.md`.\n\n## The other quantities\n\n| Quantity | How it is computed |\n| --- | --- |\n| pLDDT | ESMFold's own per-residue confidence, from the B-factor column, normalised and asserted to be 0-100. |\n| Relative SASA | FreeSASA with a 1.4 A probe, divided by the theoretical maximum for that residue type from **Tien et al. (2013)**, PLoS ONE 8(11):e80635. Not clipped at 1: a predicted structure can be more exposed than the reference tripeptide. |\n| Secondary structure | DSSP where `mkdssp` is installed, otherwise a P-SEA-style assignment from CA positions alone (Labesse et al., CABIOS 1997). The fallback agrees with DSSP on 83.7 % of 805 residues across ten diverse chains. The panel names which was used. |\n| Variant score | `log p(mutant) - log p(wild type)`, the ESM-1v convention. Natural log. **Negative means deleterious.** The wild-type entry is exactly zero by construction. |\n| Virtual CB | Glycine has no beta carbon, so one is constructed from N, CA and C by the standard tetrahedral geometry. Validated against ubiquitin: 0.13 A mean deviation from the real CB across all 70 non-glycine residues. Glycines are never skipped. |\n\n## Limitations, plainly", "# GitHub README: CODSWALLOP\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/CODSWALLOP, 2026\n\n- [x] **Conservation mapped onto the surface in Mol\\***, ConSurf-style, computed in-house\n- [ ] **Pocket detection (fpocket) and druggability scoring** across the family, and **DiffDock** beside it for the members that have no holo structure at all. Docking predicts a pose, not a motif, so it belongs here rather than in the motifs tab: the family-centric question it answers is whether a ligand seen bound in one member is placed in the equivalent pocket of a relative that has only ever been solved apo. The family already provides the thing that makes that checkable, since a predicted pose can be scored against the experimentally observed contacts of its relatives instead of being reported on the model's own confidence\n- [ ] **Cryo-EM specifics.** Map resolution versus model resolution, half-map FSC, local resolution\n- [x] **Nomenclature reconciliation panel.** One protein, fourteen names, and the reverse: one name, seven proteins\n- [x] **\"Who works on this family\".** Depositing groups, PIs, a timeline of labs\n- [ ] **Point-mutation impact overlay** from ClinVar and gnomAD for human targets\n- [ ] **Watchlist.** Email when a new entry lands in a saved family, reusing the AlphaFraud weekly timer\n- [x] **JSON API** so BoltzMaker and chatPDB can consume a family summary\n- [x] **Thirteen fixes to the panels, mostly about saying which residue.** The map gained rotation buttons and arrow keys, and its drag was half inverted: yaw followed the mouse while pitch opposed it, because screen y grows downward and the sign was never flipped, which is what \"sometimes backwards\" actually was. Sensitivity now comes from the panel's own width rather than a fixed 0.008 rad/px, which on a 900px panel was 7.2 radians of turn per drag: more than a full revolution, so no drag could be undone by dragging back. The field names up to fourteen clusters instead of six, the ones past the first three set smaller, and labels re-test for collisions as they rotate, since an arrangement laid out clear of itself in one orientation is not clear in another. That change does nothing for 16 of 73 families, which have exactly one cluster at the shared cut, and a great deal for the 22 that have more than six: insulin has 42 eligible and now shows 10. **The residue naming was the theme.** The coverage census, the disorder plot, the sequence logo, the motif table and both fold diagrams all quoted bare indices while the conservation track said `P21`, so a position read off one plot had to be translated to be found on another. They now agree, through one helper. The functional-sites column is `Residue` rather than `Where` and carries the letter. The fold diagram needed pipeline work to say it at all: the PDBe reports its layout in the reference entry's own numbering, which is not the seed's (1IOT's first strand is 42-46 there and 60-64 here, and acetylcholinesterase is out by 31 throughout), so the two were being drawn on panels one above the other, both labelled as residues. Both numbers are now carried and both are labelled. **And the smaller ones**: the ligand viewer framed its camera at 1.45 sphere radii when a 45-degree field of view needs 2.61 to inscribe one, so every ligand overflowed its own panel; the distance is now derived from the field of view and the viewport aspect, since the vertical direction is the tight one on a wide drawer. The AlphaFold button toggles instead of adding a second copy of the model on top of the first. The crystallisation scatter sizes one circle per condition by how many entries used it, replacing a `Math.random()` jitter that spent the precipitant axis on hiding overlaps, moved every point on every redraw and still failed where a hundred entries shared one condition. \"Which entries to trust\" rows open a drawer showing every measure against both the wwPDB threshold and the family median, because \"worse than the threshold\" and \"worse than its neighbours\" are different findings. The depositing-group drawer lists the primary citations with the entries each covers, derived in the browser from the citation table already in the payload rather than bought with a parse-version bump and a rebuild of every family. Construct coverage moved to the Constructs tab, and Topology stopped being a rail entry of its own: its two panels sit under Assembly, where the question is already being asked\n- [x] **An empty answer is an answer, and the search stopped promoting the wrong protein.** Three faults behind one report of a blank Contacts tab. **The panel blamed the wrong thing**: it read \"run `CODSWALLOP.py contacts` on a workstation, PLIP and OpenBabel are not installed on the server\", which was pre-migration copy and false on both counts, since the droplet has had both installed since the compute moved. Two more of those strings were still in the map and topology panels. **`build` returned `None` for two unrelated situations**, an apo family and an analysis that yielded nothing, and the caller could not tell them apart from PLIP being absent, so it exited non-zero; the worker then marked the family served regardless (deliberately, since contacts must not hold a family in the queue), and nothing ever retried it. A family with no ligand-bound structure has no fingerprint to compute and never will have one, so it now writes a current artefact carrying `holo_entries` and zero counts, and the tab says which of the two happened. **The family they landed on was itself wrong**, and that was my VEGFR fix: human-first is only right when a human protein of that name exists. Humans have no beta-lactamase, so the eight human hits were proteins that merely contain a beta-lactamase-*like* fold (Apollo exonuclease, MBLAC2, COA7, dipeptidase 1, LACTB2) and every one outranked the actual enzymes, which are bacterial and among the most deposited proteins in the archive; `Beta-lactamase` itself came ninth. Ranking now scores the name match first and uses human only to break ties, which is exactly the VEGFR case since nothing is named \"VEGFR\" at all. The remaining term explains the top hit: with reviewed and human tied across all eight, the tie-break was the name alphabetically, so first place went to whichever protein sorted first, decided by an apostrophe. Found in passing while checking that fix: **\"vascular endothelial growth factor receptor\" was classified as a pasted sequence**, because whitespace is stripped before the alphabet test and the phrase is thirty-nine valid residue codes, so it seeded a nonsense family instead of searching. Whitespace shape now decides, since a sequence's spacing is layout (uniform blocks, or runs too long to be words) and prose's is words: length alone cannot separate them, as the shortest words in that phrase are six characters and six is a legitimate block size", "# Microplastics and Health: Alarming New Research Reveals Widespread Biological Impact\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n* **Cardiotoxic Mechanisms**: PS-MPs induced severe myocardial damage through TLR4/NF-\u03baB pathway activation, leading to inflammation, cell death, and structural tissue disruption in carp hearts.\n* **Inflammatory Response**: Exposure significantly increased pro-inflammatory cytokines (TNF-\u03b1, IL-1\u03b2, IL-6) while suppressing anti-inflammatory IL-10, creating a sustained inflammatory environment.\n* **Cell Death Pathways**: PS-MPs activated both apoptotic and necroptotic cell death mechanisms, with apoptosis predominating through altered Bcl-2/Bax ratios and caspase activation.\n* **Molecular Targeting**: The study identified TLR4 as the specific receptor mediating PS-MPs recognition and subsequent inflammatory cascades in cardiac tissue.\n* **Dose-Dependent Effects**: Both in vivo and in vitro experiments demonstrated clear dose-response relationships for inflammatory markers and cell death indicators.\n* **Structural Damage**: Histopathological analysis revealed extensive myocardial fiber disruption, enlarged intercellular spaces, and inflammatory cell infiltration following PS-MPs exposure.\n\n1. **Zhang, Q., et al. (2023).** Polystyrene microplastics induce myocardial inflammation and cell death via the TLR4/NF-\u03baB pathway in carp. Fish and Shellfish Immunology, 135:108690. <https://doi.org/10.1016/j.fsi.2023.108690>\n2. **Zhang, Y., et al. (2022).** Polystyrene microplastics-induced cardiotoxicity in chickens via the ROS-driven NF-\u03baB-NLRP3-GSDMD and AMPK-PGC-1\u03b1 axes. Science of the Total Environment, 840:156727. <https://doi.org/10.1016/j.scitotenv.2022.156727>\n3. **Zhao, L., et al. (2021).** Prolonged oral ingestion of microplastics induced inflammation in the liver tissues of C57BL/6J mice through polarization of macrophages and increased infiltration of natural killer cells. Ecotoxicology and Environmental Safety, 227:112882. <https://doi.org/10.1016/j.ecoenv.2021.112882>\n4. **Qiao, R., et al. (2019).** Microplastics induce intestinal inflammation, oxidative stress, and disorders of metabolome and microbiome in zebrafish. Science of the Total Environment, 662:246-253. <https://doi.org/10.1016/j.scitotenv.2018.12.041>\n5. **Deng, Y., et al. (2017).** Tissue accumulation of microplastics in mice and biomarker responses suggest widespread health risks of exposure. Scientific Reports, 7:46687. <https://doi.org/10.1038/srep46687>\n6. **Carbery, M., et al. (2018).** Trophic transfer of microplastics and mixed contaminants in the marine food web and implications for human health. Environment International, 115:400-409. <https://doi.org/10.1016/j.envint.2018.03.007>\n\nThis groundbreaking study examined the cardiotoxic effects of polystyrene microplastics (PS-MPs) in aquatic organisms, specifically investigating the molecular mechanisms underlying myocardial damage in carp (Cyprinus carpio). Researchers exposed carp to 1000 ng/L PS-MPs (1-10 \u03bcm) for 21 days and isolated primary cardiomyocytes for in vitro analysis at concentrations of 100 and 200 \u03bcg/mL. The study revealed severe myocardial structural damage characterized by fiber breaks, enlarged cell gaps, and extensive inflammatory cell infiltration. Mechanistic investigation identified TLR4 as the primary pattern recognition receptor activated by PS-MPs, leading to downstream NF-\u03baB pathway activation and NLRP3 inflammasome stimulation. The research demonstrated that PS-MPs exposure significantly increased expression of inflammatory cytokines (TNF-\u03b1, IL-1\u03b2, IL-6) while suppressing anti-inflammatory IL-10. Cell death analysis revealed that PS-MPs induced both apoptosis and necrosis, with apoptosis predominating through Bcl-2 downregulation and Bax upregulation, ultimately activating caspase cascades. The study also showed activation of necroptotic pathways through RIPK3 and MLKL upregulation. This research provides the first comprehensive evidence of PS-MPs cardiotoxicity in aquatic vertebrates, establishing important precedents for understanding cardiovascular risks in exposed organisms.\n\n# 5) Pulmonary toxicity assessment of polypropylene, polystyrene, and polyethylene microplastic fragments in mice\n\n* **Polymer-Specific Toxicity**: PS microplastics induced severe pulmonary inflammation while PP and PE showed minimal effects, establishing clear differential toxicity patterns among common plastic polymers.\n* **TLR-Mediated Mechanisms**: PS activated TLR4-dependent inflammatory pathways leading to NF-\u03baB and NLRP3 inflammasome activation, while PP activated TLR2 without downstream inflammatory signaling.\n* **Surface Charge Correlation**: Microplastic toxicity correlated with surface charge properties, with PS showing the highest negative zeta potential (-38.93 mV) and greatest inflammatory potential.\n* **Inflammatory Response**: PS exposure significantly increased inflammatory cell infiltration, cytokine production (IL-1\u03b2, IL-6), and chemokine expression in bronchoalveolar lavage fluid.\n* **Histopathological Changes**: PS-exposed mice showed inflammatory cell infiltration and macrophage accumulation in perivascular and peribronchial regions of lung tissue.\n* **NLRP3 Inflammasome Activation**: PS specifically triggered NLRP3 inflammasome assembly and activation, leading to enhanced IL-1\u03b2 processing and sustained inflammatory responses.", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\n**Authors:** Xu Q, Traag BA, Willemse J, McMullan D, Miller MD, Elsliger MA, Abdubek P, Astakhova T, Axelrod HL, Bakolitsa C, Carlton D, Chen C, Chiu HJ, Chruszcz M, Clayton T, Das D, Deller MC, Duan L, Ellrott K, Ernst D, Farr CL, Feuerhelm J, Grant JC, Grzechnik A, Grzechnik SK, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Krishna SS, Kumar A, Marciano D, Minor W, Mommaas AM, Morse AT, Nigoghossian E, Nopakun A, Okach L, Oommachen S, Paulsen J, Puckett C, Reyes R, Rife CL, Sefcovic N, Tien HJ, Trame CB, van den Bedem H, Wang S, Weekes D, Hodgson KO, Wooley J, Deacon AM, Godzik A, Lesley SA, Wilson IA, van Wezel GP.  \n**Journal:** The Journal of biological chemistry (2009). DOI: 10.1074/jbc.m109.018564. PMID: 19567872.\n\nSsgA-like proteins (SALPs) are a family of homologous cell division-related proteins that occur exclusively in morphologically complex actinomycetes. We show that SsgB, a subfamily of SALPs, is the archetypal SALP that is functionally conserved in all sporulating actinomycetes. Sporulation-specific cell division of Streptomyces coelicolor ssgB mutants is restored by introduction of distant ssgB orthologues from other actinomycetes. Interestingly, the number of septa (and spores) of the complemented null mutants is dictated by the specific ssgB orthologue that is expressed. The crystal structure of the SsgB from Thermobifida fusca was determined at 2.6 A resolution and represents the first structure for this family. The structure revealed similarities to a class of eukaryotic \"whirly\" single-stranded DNA/RNA-binding proteins. However, the electro-negative surface of the SALPs suggests that neither SsgB nor any of the other SALPs are likely to interact with nucleotide substrates. Instead, we show that a conserved hydrophobic surface is likely to be important for SALP function and suggest that proteins are the likely binding partners.\n\n## Structural basis of murein peptide specificity of a gamma-D-glutamyl-l-diamino acid endopeptidase (2009)\n\n**Authors:** Xu Q, Sudek S, McMullan D, Miller MD, Geierstanger B, Jones DH, Krishna SS, Spraggon G, Bursalay B, Abdubek P, Acosta C, Ambing E, Astakhova T, Axelrod HL, Carlton D, Caruthers J, Chiu HJ, Clayton T, Deller MC, Duan L, Elias Y, Elsliger MA, Feuerhelm J, Grzechnik SK, Hale J, Han GW, Haugen J, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Kumar A, Marciano D, Morse AT, Nigoghossian E, Okach L, Oommachen S, Paulsen J, Reyes R, Rife CL, Trout CV, van den Bedem H, Weekes D, White A, Wolf G, Zubieta C, Hodgson KO, Wooley J, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Structure (London, England : 1993) (2009). DOI: 10.1016/j.str.2008.12.008. PMID: 19217401.\n\nThe crystal structures of two homologous endopeptidases from cyanobacteria Anabaena variabilis and Nostoc punctiforme were determined at 1.05 and 1.60 A resolution, respectively, and contain a bacterial SH3-like domain (SH3b) and a ubiquitous cell-wall-associated NlpC/P60 (or CHAP) cysteine peptidase domain. The NlpC/P60 domain is a primitive, papain-like peptidase in the CA clan of cysteine peptidases with a Cys126/His176/His188 catalytic triad and a conserved catalytic core. We deduced from structure and sequence analysis, and then experimentally, that these two proteins act as gamma-D-glutamyl-L-diamino acid endopeptidases (EC 3.4.22.-). The active site is located near the interface between the SH3b and NlpC/P60 domains, where the SH3b domain may help define substrate specificity, instead of functioning as a targeting domain, so that only muropeptides with an N-terminal L-alanine can bind to the active site.\n\n## Crystal structures of MW1337R and lin2004: representatives of a novel protein family that adopt a four-helical bundle fold (2008)\n\n**Authors:** Kozbial P, Xu Q, Chiu HJ, McMullan D, Krishna SS, Miller MD, Abdubek P, Acosta C, Astakhova T, Axelrod HL, Carlton D, Clayton T, Deller M, Duan L, Elias Y, Elsliger MA, Feuerhelm J, Grzechnik SK, Hale J, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Koesema E, Kumar A, Marciano D, Morse AT, Murphy KD, Nigoghossian E, Okach L, Oommachen S, Reyes R, Rife CL, Spraggon G, Trout CV, van den Bedem H, Weekes D, White A, Wolf G, Zubieta C, Hodgson KO, Wooley J, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Proteins (2008). DOI: 10.1002/prot.22020. PMID: 18324683.\n\n*(no abstract available)*\n\n## Crystal structure of homoserine O-succinyltransferase from Bacillus cereus at 2.4 A resolution (2007)\n\n**Authors:** Zubieta C, Krishna SS, McMullan D, Miller MD, Abdubek P, Agarwalla S, Ambing E, Astakhova T, Axelrod HL, Carlton D, Chiu HJ, Clayton T, Deller M, DiDonato M, Duan L, Elsliger MA, Grzechnik SK, Hale J, Hampton E, Han GW, Haugen J, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Koesema E, Kumar A, Marciano D, Morse AT, Nigoghossian E, Oommachen S, Reyes R, Rife CL, van den Bedem H, Weekes D, White A, Xu Q, Hodgson KO, Wooley J, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Proteins (2007). DOI: 10.1002/prot.21208. PMID: 17546672.\n\n*(no abstract available)*\n\n## Crystal structures of two novel dye-decolorizing peroxidases reveal a beta-barrel fold with a conserved heme-binding motif (2007)\n\n**Authors:** Zubieta C, Krishna SS, Kapoor M, Kozbial P, McMullan D, Axelrod HL, Miller MD, Abdubek P, Ambing E, Astakhova T, Carlton D, Chiu HJ, Clayton T, Deller MC, Duan L, Elsliger MA, Feuerhelm J, Grzechnik SK, Hale J, Hampton E, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kumar A, Marciano D, Morse AT, Nigoghossian E, Okach L, Oommachen S, Reyes R, Rife CL, Schimmel P, van den Bedem H, Weekes D, White A, Xu Q, Hodgson KO, Wooley J, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Proteins (2007). DOI: 10.1002/prot.21550. PMID: 17654545.", "# GitHub README: CODSWALLOP\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/CODSWALLOP, 2026\n\n- [x] **Fetch the UniProt canonical and every deposited SEQRES.** Plus UniProt active and binding sites, used only to *describe* a mutation (\"at an annotated active site\"), never to claim what it did\n- [x] **Family alignment with per-column conservation and a sequence logo.** A star alignment against the seed rather than a progressive MSA, so a column here is the same column as in the coverage census and the domain ribbon, and no external binary is needed on the droplet. Conservation is normalised Shannon entropy weighted by how many entities used each construct. It also produces the list this panel exists for: **positions people deliberately mutate**, where the wild type still dominates but a real minority carries something else. On p53 the top of that list is M133L, V203A, N239Y and N268D, which together are the thermostabilising superstable quadruple mutant, alongside the Y220C druggable hotspot; on carbonic anhydrase II it is the active-site pocket variants\n- [x] **Construct diff engine.** The flagship, and it works: Align every SEQRES to the canonical and classify: N- and C-terminal expression tags (His6/8/10, Strep-II, FLAG, HA, Myc, Avi), cleavage sites and scar residues (TEV, 3C/PreScission, thrombin, SUMO, enterokinase), fusion partners (MBP, GST, SUMO, Trx, T4 lysozyme, BRIL, rubredoxin, PGS, GFP), truncations and internal loop deletions, point mutations sub-classified as catalytic, thermostabilising, surface entropy reduction or disulphide engineering, and non-canonical residues including SeMet\n- [x] **Replaced the phase 1 fusion heuristic**, which was wrong on every case it flagged for carbonic anhydrase II: all four were *Schistosoma* CA, a genuinely larger protein with a His6 tag, called a fusion only because it was measured against the human seed rather than its own reference\n- [x] **Construct table.** One row per unique construct, most-used first, with the entries that used it and their best resolution. On p53 the top row is residues 94-312 carrying M133L/V203A/Y220C at 1.24 \u00c5, and R273H, R249S and R282W each get their best structure: \"which construct crystallises best\" answered directly\n- [x] **Domain and fold annotation.** CATH, SCOP2B and ECOD per chain from the RCSB instance features, aggregated to a consensus with median boundaries and a chain count, drawn as an architecture ribbon. A domain must appear on at least 2 chains (or 1 % of them) to be reported: without that the list filled with domains belonging to *other* proteins in the complexes, and p53 picked up \"Green fluorescent protein\" and \"Annexin\"\n- [x] **Density coverage census.** `UNOBSERVED_RESIDUE_XYZ` per chain, mapped through each member's own alignment offset onto seed coordinates, drawn as a second curve against the construct curve. \"Resolved in no construct at all\" turned out to be as useless a bar as the binary union (0 % for p53), so the reported figure is the *ratio*: residues resolved in under a quarter of the constructs that contained them. p53 reads 20.6 %, and names 361-393, 62-90 and 294-312; spike's 1152-1271 is resolved in 4 %; carbonic anhydrase II is 0.8 %\n- [x] **Cross-species orthologue matrix.** Organisms by entry count, with best resolution, holo count and the fraction of the seed each one's constructs cover\n- [x] **Bound the construct table to the shared selection state.** Hovering a construct dims every node on the map that does not use it, and clicking opens the best entry that did, so a reader lands on a real structure rather than an abstraction\n\n### Phase 3: structure, chemistry and interactions", "# GitHub README: AlphaFraud\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/AlphaFraud, 2026\n\n| Field | Detail |\n|---|---|\n| **Take-home** | Not a refold but a wrong-*context* capture \u2014 a 39-residue EGF1 fragment threaded through a partner enzyme's active site |\n| **AlphaFold vs experimental** | <a href=\"https://alphafraud.mdeller.com/examples#ex-6Q9F_2\"><img src=\"https://alphafraud.mdeller.com/ribbon/6Q9F_2.svg\" alt=\"Experimental structure of 6Q9F_2 coloured by C\u03b1 deviation from AlphaFold\" width=\"380\"></a><br><sub>Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \"ghost\" overlay</sub> |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-6Q9F_2)** |\n| **KPIs** | TM 0.127 \u00b7 lDDT 0.468 \u00b7 pLDDT 93.5 \u00b7 C\u03b1-RMSD 6.6 \u00c5 \u00b7 FRAUD 0.37 \u00b7 novelty 0% |\n| **Stats** | 47 depositions, 26 confidently wrong \u00b7 pLDDT\u2194lDDT \u22120.59 |\n| **Structure** | Modular zymogen (Gla\u2013EGF1\u2013EGF2\u2013protease); the worst deposition is a linear EGF1 peptide in AspH's channel, not a folded module |\n| **Why AlphaFold gets it wrong** | A context failure, not aggregation \u2014 the RMSD is modest (6.6 \u00c5) but TM collapses because a linear fragment has no compact fold to superpose |\n| **Biology & disease** | Central coagulation zymogen; activated factor Xa is the target of rivaroxaban, apixaban and edoxaban |\n| **Key fact** | The non-amyloid contrast \u2014 the error is missing biological context (fragment / complex / PTM), not misfolding |\n\n## \ud83e\uddea Calculate\n\nThe **[Calculate](https://alphafraud.mdeller.com/calculate)** tab runs the whole pipeline on demand for any human PDB ID you type and returns the same panel format as the auto examples \u2014 KPIs, the C\u03b1-deviation ribbon, the interactive 3D viewer and a deterministic write-up \u2014 with **no AI-generated text**.\n\n- **Strict autocomplete.** The input only completes to entries that (1) exist, (2) are human, and (3) have an AlphaFold model. That comes from a precomputed index \u2014 seeded for free from the corpus (the ~83k entities already carrying a confirmed AlphaFold model), back-filled with a one-off pre-cutoff human sweep, and refreshed automatically by every weekly run. Press Tab to complete.\n- **Any human structure, blind or not.** Post-cutoff entries are genuine blind tests; pre-cutoff entries compute too, with a note that AlphaFold may have seen them in training (so it is not a blind prediction).\n- **Clear reasons on rejection.** If an entry doesn't qualify you are told exactly why: not a human protein structure, no single UniProt mapping (antibody / fusion / chimera / synthetic), or no AlphaFold model for the sequence.\n- **Right or wrong, honestly.** A good prediction renders the green \"AlphaFold gets it right\" panel (like the positive control); a confidently-wrong one renders the failure panel \u2014 the same deterministic builder (`examples_auto.build`) decides which from the metrics.\n- **Isolated and cached.** On-demand results are stored under a separate run and status, so they never enter the leaderboard, stats, analysis or archive; the heavy compute runs in an isolated subprocess; and a repeat request for the same structure is instant.\n\n## \ud83e\uddee Coverage and skipped entities\n\nNot every deposited entity can be compared. AlphaFold DB holds a single model per human UniProt sequence (monomers only), so anything that does not map cleanly to one accession, has no model covering its resolved range, or is too small or too large to compare is recorded as *skipped* rather than forced through.\n\nThe completed full-archive backfill (finished 2026-07-18) processed **96,827** post-cutoff human entities: **82,796 analysed** (85.5%; 67,314 fast-screened plus 15,482 fully compared), **1,941 confidently wrong** (2.3% of analysed, 431 of them sequence-novel), **13,908 skipped** (14.4%), and **123 unrecoverable errors** (0.1%, broken coordinate-less files). The skips break down as:\n\n| Count | Share | Reason | Nature |\n|---:|---:|---|---|\n| **11,875** | 85% | **No single UniProt mapping** \u2014 antibodies, chimeras, engineered constructs, fusion proteins, synthetics | Structural fact: no single AlphaFold model exists to compare against |\n| **1,186** | 9% | **No AlphaFold model covering the residue range** \u2014 mostly giant multi-domain proteins (titin-scale) whose resolved fragment falls outside AlphaFold DB's fragmented coverage, plus isoforms it lacks | Coverage gap in AlphaFold DB |\n| **464** | 3% | **Too few residues aligned (<10)** \u2014 tiny peptides or mostly-unresolved chains | Nothing meaningful to compare |\n| **381** | 3% | **Structure too large (>40 MB)** \u2014 the out-of-memory guard for the 3.8 GB droplet | Resource limit (our cap) |\n| **2** | \u2014 | Sequence too short (<3 residues) | Degenerate |\n\nThe 85% is exactly what the design intends: because AlphaFold DB is monomer-and-single-sequence only, anything that does not map to one human UniProt accession (an antibody Fab, a designed fusion, a chimera) has no counterpart model, so it is flagged and skipped rather than mis-compared.\n\nSkipped entities are terminal by design: `retry-errors` only re-runs `status='error'`, and the resumable weekly and backfill runs skip anything already recorded. Four of the five buckets are permanent structural properties, so re-running would simply re-skip them. The two worth revisiting are the **381 oversized structures** (a memory limit of the small droplet, not a true incomparability, recoverable by raising `MAX_STRUCT_BYTES` or upsizing the box) and the multi-domain subset of the **1,186 missing-model** bucket (recoverable with better fragment selection for AlphaFold models above 2,700 residues).\n\n## \ud83d\udee0\ufe0f Web deployment\n\n**Live at [alphafraud.mdeller.com](https://alphafraud.mdeller.com).** The report is served by a Flask app (gunicorn behind nginx with a Let's Encrypt certificate) and the weekly pipeline runs from a systemd timer on the same droplet.\n\n```bash\n# 1. From your Mac: copy .env.example to .env and fill in DROPLET_SSH / DROPLET_PATH / SERVER_NAME\ncp .env.example .env", "# Molecular Velcro: Catching What We Cannot See\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-04-15 on marcdeller.com (categories: blog, microplastics).*\n\n* [**Nature, News & Views, April 2026**](https://www.nature.com/articles/s44221-026-00622-z)\n\n* **Simultaneous action:** The Fe\u2083O\u2084@SiO\u2082@CP platform captures nano- and microplastics and detects them via colour change in under three minutes.\n* **Broad adhesion:** Copper-doped polydopamine acts as molecular Velcro, binding multiple polymer types without chemical labelling or polymer-specific tuning.\n* **Proven human risk:** Microplastics have been detected in human blood, lung tissue, and arterial plaque, intensifying demand for scalable removal technologies.\n* **Therapeutic potential:** Biocompatible mussel-inspired polydopamine coatings open a future pathway toward in-body sequestration strategies for circulating plastic particles.\n\n1. **A coupling platform for microplastics extraction and detection:** [Chakraborty et al., Nature Water, 2026](https://doi.org/10.1038/s44221-026-00622-z)\n2. **Interfacial interaction-driven rapid capture and on-site analysis of nano- and microplastics enabled by multifunctional magnetic adsorbent:** [Yao et al., Nature Water, 2026](https://doi.org/10.1038/s44221-026-00610-3)\n3. **Risk assessment of microplastic particles:** [Koelmans et al., Nature Reviews Materials, 2022](https://doi.org/10.1038/s41578-021-00411-y)\n4. **Discovery and quantification of plastic particle pollution in human blood:** [Leslie et al., Environment International, 2022](https://doi.org/10.1016/j.envint.2022.107199)\n5. **Chemical analysis of microplastics and nanoplastics: challenges, advanced methods, and perspectives:** [Ivleva et al., Chemical Reviews, 2021](https://doi.org/10.1021/acs.chemrev.1c00178)\n\n**A new coupling platform for microplastic extraction and detection may change how water systems monitor and manage one of the most pervasive invisible pollutants of our time.**\n\nPlastic particles smaller than a human hair have now been detected in virtually every aquatic environment on Earth, from remote alpine lakes to the tap water of major cities, and, more troublingly, in human blood, lung tissue, and arterial plaque. A 2022 biomonitoring study published in [Environment International](https://doi.org/10.1016/j.envint.2022.107199) detected microplastics in 77% of blood samples from healthy adult volunteers, demonstrating unequivocally that these particles move through the body. The same year, a landmark risk framework published in [Nature Reviews Materials](https://doi.org/10.1038/s41578-021-00411-y) confirmed that microplastics cause measurable harm across multiple biological and ecological systems. Yet despite mounting evidence of their reach, removing and monitoring the smallest particles (those at or below 100 nanometres) has remained an unsolved technical problem. Conventional water treatment processes consistently fail to intercept them, while analytical methods capable of detecting nanoscale plastics require specialized laboratory infrastructure, chemical labeling, and significant processing time, making real-world, decentralized monitoring practically impossible.\n\nInto this gap, a team led by [Qingfeng Yao and colleagues](https://doi.org/10.1038/s44221-026-00610-3), writing in *Nature Water*, has introduced a single multifunctional particle that captures nano- and microplastics and then reports what it has captured, simultaneously\u00a0within minutes,\u00a0without the need for\u00a0fluorescent staining or spectroscopic infrastructure. The material, designated Fe\u2083O\u2084@SiO\u2082@CP, consists of a magnetic iron-oxide core surrounded by a stabilizing silica shell, which is, in turn, coated with a copper-doped polydopamine layer that serves as the active adhesive surface. Polydopamine is a bioinspired polymer that mimics the adhesive proteins of marine mussels, and its copper doping confers an additional enzymatic character: the coating drives a catalytic reaction that produces a measurable color change when plastic particles bind to it. The result is a particle that does not simply adsorb plastics; it signals their presence through a visible shift in optical absorbance, enabling a label-free on-site readout with a detection limit reaching into the microgram-per-liter range when the captured material is preconcentrated. As described in a commentary by [Chakraborty in Nature Water](https://doi.org/10.1038/s44221-026-00622-z), the coating acts like many different kinds of molecular Velcro: rather than being chemically selective toward one polymer type, it offers multiple weak binding interactions simultaneously, allowing it to grip polyethylene, polyethylene terephthalate, aged mixed debris, and other plastic chemistries with comparable efficiency.\n\n*\u201cRather than being tuned to one polymer, the coating offers multiple weak interactions, many different kinds of Velcro, so it can latch onto plastics with very different chemistries.\u201d*", "# The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\nAuthors: Span, E.A.; Suess, D.L.M.; Deller, M.C.; Britt, R.D.; Marletta, M.A.\nSource: ACS Chemical Biology, 2017\n\nWild type Neurospora crassa (FGSC 2489) and Myceliophthora thermophila (ATCC 42464) were used in the experiments herein; genomic DNA from M. thermophila was purified using methods described by the Fungal Genetics Stock Center.[34] PASC was prepared as previously described.[8]\n\n### Protein Expression and Purification\n\nMtPMO3* constructs were expressed in Neurospora crassa using a knock-in expression method, as previously described.[35] Point substitutions were introduced using site-directed mutagenesis to a pCSR-1 vector containing wild type MtPMO3* (gene ID MYCTH_92668, isolated from M. thermophila genomic DNA). Mutagenesis products were transformed into E. coli DH5\u03b1 cells for propagation, and sequence-validated linearized target DNA constructs were transformed into N. crassa for expression. Protein constructs were purified from the N. crassa secretome as described for wild type MtPMO3*.[24] Before the final gel filtration step, enzymes were reconstituted with excess CuSO4 at pH 5.0 for 4 h at ambient temperature. The final chromatography step also functioned to desalt the sample of excess CuSO4.\n\nMtCDH-2 was purified natively from M. thermophila as described previously.[8]\n\n### Protein Crystallization\n\nWild type MtPMO3* was crystallized using sitting-drop vapor diffusion techniques as implemented by an Oryx8 crystallization robot (Douglas Instruments). Crystals used for structure determination grew in the presence of a crystallization cocktail consisting of 18% PEG 6000 (w/v) and 0.1 M citric acid at pH 3.9 and 4 \u00b0C. A total of 150 nL of crystallization cocktail was mixed 1:1 with 20 mg mL\u22121 of MtPMO3*. Prior to flash-cooling in liquid nitrogen, crystals were soaked briefly in 25% PEG 6000 (v/v in water) as a cryoprotectant.\n\n### Data Collection and Structure Determination\n\nX-ray diffraction data was collected at beamline 12-2 at the Stanford Synchrotron Radiation Lightsource (SSRL). Data were processed with Denzo/Scalepack programs from the HKL2000 software package.[36] Initial phases were obtained by molecular replacement with a search model based on the structure of Cel61B (PDB ID 2VTC), a PMO from the soft-rot fungus Hypocrea jecorina.[16] The model for MtPMO3* was built using COOT[37] and refined using Refmac5 within the CCP4 package.[38] The statistics of the diffraction data and structure refinement are summarized in Supporting Information Table 1.\n\n### Identification of Conserved H-Bonding Networks\n\nMultiple sequence alignment (MSA) data sets for each PMO family were generated using a Hidden Markov Model-based search algorithm.[39] Search ensembles consisted of smaller, starting MSAs containing protein sequences of structurally or functionally characterized PMOs. Conserved residues were examined against existing structures of each PMO family to identify conserved sequence motifs near the active site engaging in conserved H-bonding interactions.\n\n### Cellulose Activity Assays\n\nWild type and variant MtPMO3* (2 \u03bcM) was mixed with 10 mg mL\u22121 phosphoric acid swollen cellulose (PASC) and 1 \u03bcM MtCDH-2 with atmospheric O2. Reactions of 45 \u03bcL were carried out in 50 mM sodium acetate buffer (5.0) at 40 \u00b0C for 1\u201330 min. Peaks from aldonic acids with degrees of polymerization of (DP) 5\u201313 were analyzed via HPAEC as described previously.[8] Products were quantified by integrating peak areas with Dionex Chromeleon software v. 7.1.3. Smaller C1-oxidized products with DP 2\u20134 were excluded from analysis, as they are also products of the CDH reaction.\n\n### O2 Consumption Assays\n\nAn Oxygraph Plus System (Hansatech Instruments) equipped with a water jacket was used to measure O2 levels in MtPMO3* reactions over time. Assays containing 0.5 \u03bcM MtPMO3*, 10 mg mL\u22121 PASC, and atmospheric O2 were conducted in 50 mM sodium acetate buffer (5.0) at 40 \u00b0C. Reactions of 300 \u03bcL were initiated with 5 \u03bcM MtCDH-2. Stir bar speed was set to 70 rpm. O2View software was used for data analysis. Initial rates were calculated using 20 s intervals within the first 2 min of the reaction.\n\n### Peroxide Formation Assays\n\nHorseradish peroxidase (HRP)-coupled assays were based on previously described methods,[29] utilizing Amplex Red as a substrate for the photometric detection of peroxide in MtPMO3* reactions over time. Assays contained 100 \u03bcL of 100 mM HEPES (7.0), 0.5 \u03bcM MtPMO3*, 0.5 U/mL HRP, 50 \u03bcM Amplex Red, 0.5 \u03bcM MtCDH-2, 100 \u03bcM cellobiose, and \u00b125 U/mL superoxide dismutase (SOD). Reactions were carried out in flat-bottom, untreated 96-well plates (Corning), and detection at 573 nm was performed with a SpectraMax 340 plate reader (Molecular Devices) and analyzed with SoftMax Pro software.\n\n### Electron Paramagnetic Resonance Spectroscopy\n\nAll variant and wild type MtPMO3* samples (200\u2013300 \u03bcM) were prepared in 10 mM Tris (8.5) and 150 mM NaCl with 20% glycerol. X-band continuous-wave (CW) EPR spectra were recorded on a Bruker ELEXSYS E500 spectrometer equipped with a cylindrical TE011-mode resonator (SHQE-W), an ESR-900 liquid helium cryostat, and an ITC-5 temperature controller (Oxford Instruments). Spectral simulations were performed with MATLAB Release 2016A using the EasySpin 5.1.5 toolbox.[40]\n\n## Acknowledgments\n\nThe following are acknowledged: J. Hangasky and other Marletta lab members for technical advice and helpful discussion; J. Klinman and A. Murphy for use of their Oxygraph instruments; I. Wilson for SSRL beam time; and members of the JCSG for assistance with high-throughput crystal screening. This work was funded by NSF grant #1565770 and a grant from the Energy Biosciences Institute to M.A.M. The EPR spectroscopy was supported by the National Institutes of Health GM111025 grant to R.D.B. D.L.M.S. acknowledges funding from the National Institutes of Health NIGMS (F32GM111025).\n\n## References\n\n*(reference list omitted)*", "# Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\nAuthors: Zhuo, J.; Li, Y.-L.; Qian, D.-Q.; Burns, D.M.; Mei, S.; Rafalski, M.; Xu, M.; Cao, G.; Pan, Y.; Jia, Z.; Jalluri, R.; Epling, L.B.; Fenalti, G.; Deller, M.C.; Procak, J.; Covington, M.; He, X.; Collins, R.; Stubbs, M.; Burke, K.; Oliver, J.; Margulis, A.; Boer, J.; Wynn, R.; Scherle, P.; Diamond, S.; Newton, R.; Zhang, Y.; Metcalf, B.; Yao, W.\nSource: Journal of Medicinal Chemistry (Drug Annotation), 2026\n\nPharmacokinetic evaluation revealed that 9 had an improved rat and a similar cynomolgus (cyno) PK profile compared to 3; however, the dog PK profile was significantly worse than 3 (Table 2). It was conjectured that the modest and poor bioavailability in cyno and dog, respectively, may be partly attributable to the low permeability of 9. It has been demonstrated in the literature that one possible strategy to address poor permeability, ADME, and toxicity properties of compounds containing an amide is to introduce intramolecular hydrogen bonding (IMHB). In fact, we previously demonstrated in two separate projects (11-\u03b2-HSD1 and c-Met) that incorporation of an IMHB amide could dramatically improve the Caco-2 permeability and more importantly the PK profile of the compounds, which led to the discovery of the FDA-approved c-Met inhibitor INCB28060 13 and the clinically explored 11-\u03b2-HSD1 inhibitor INCB13739 11 (Table 3). It was speculated that a similar approach may be applicable to improve the PK profile of 9. Thus, fluorination of the *ortho*-position of the phenyl ring and the iso-propyl methyl group were explored, and the results are summarized in Table 4.\n\nReplacement of one of the methyl groups of 9 with a trifluoromethyl group had little effect on permeability and the rat PK profile. Introduction of an *ortho*-fluorine to afford 15 improved the Caco-2 permeability by 3-fold, exposure by 4-fold, and oral bioavailability by 2-fold in rats, although the cynomolgus monkey PK was only marginally improved for 15 vs 9. Interestingly, incorporation of both the trifluoromethyl and *ortho*-fluoro groups afforded 16, which displayed an enhanced improvement in Caco-2 (5.5-fold), exposure (7.5-fold), and oral bioavailability (2-fold) in rats compared to non-IMHB amide 9. Importantly, 16 displayed a significant improvement in the cyno PK profile with a 52-fold increase in exposure and a 4-fold increase in oral bioavailability vs 9. In addition, 16 had an approximately 2-fold improvement in enzymatic, cellular, and whole-blood potency. These results were surprising given that the trifluoromethyl amide capping group did not provide any significant advantage when the *ortho*-fluorine was not present (14 vs 9).\n\nA structure\u2013activity relationship (SAR) trend that unexpectedly emerged for this and other highly related scaffolds was that incorporation of a small group at the meta-position relative to the amide moiety imparted a boost in potency. We believe that this boost in potency may be due to the substituent twisting the phenyl ring slightly out of the plane with respect to the azetidine ring, which will be the focus of a follow-up manuscript. Therefore, a meta-fluorine was introduced to 15 and 16 to afford 17 and 18, respectively. As a result, both compounds had an improvement in their potency. Compound 17 had a 3-fold improvement in both the enzymatic and cell potencies with a 2-fold improvement in the whole-blood assay compared to 15. Interestingly, 17 had a 2-fold improvement in oral bioavailability in both rat and cyno PK, with a slight improvement in the Caco-2 permeability and rat oral bioavailability vs 15. Compound 18 had an almost 2-fold improvement in both the enzymatic and cellular potencies vs 16, although the whole blood potency was similar. The Caco-2 of 18 was improved by almost 2-fold compared to 16, although the PK profiles in both rat and cyno were comparable. Both 17 and 18 displayed significant improvements in the Caco-2 and PK profiles in both rat and cyno in comparison to the parent secondary amide compound 9, which lacked an IMHB. Elimination of the *ortho*-fluoro substituent of 18 to afford 19 demonstrated the importance of the intramolecular H-bonded amide to improve the Caco-2 and PK profiles. It should be noted that both 16 and 18 had superior potency and PK profiles in both rat and cyno compared to 3. Despite the excellent PK profile in both rats and cynos for 16 and 18, both compounds had very poor exposure and oral bioavailability in the dog (data not shown).\n\nWhereas the 7H-pyrrolo[2,3-*d*]pyrimidine moiety is ubiquitous in numerous kinase inhibitors as the hinge-binding motif, particularly for JAK inhibitors, it is also well-documented that the indole-like ring is prone to CYP-promoted oxidation due to its electron-rich nature. Therefore, a variety of replacement hinge-binding motifs were explored. It was discovered that the pyrazole ring was a competent hinge-binding motif (20), albeit with a substantial loss in enzymatic potency compared to the 7H-pyrrolo[2,3-*d*]pyrimidine moiety 18 (Table 5). Sequential methylation of the 3- and 5-positions increased the potency 16- and 28-fold for the mono- and dimethylated pyrazoles 21 and 22 compared to the parent pyrazole 20. Comparison of 22 to 18 indicated that the dimethylpyrazole was 44-fold less potent than the 7H-pyrrolo[2,3-*d*] pyrimidine in the enzymatic assay, although the JAK1 isoform selectivity in the enzymatic assay was almost doubled to 46-fold JAK1 selective over JAK2. Despite the significant decrease in enzymatic binding, 22 had a comparable cellular potency, and the whole blood potency was only about 2-fold less than 18. Interestingly, the whole blood isoform selectivity increased by ~2.7-fold, which was in alignment with the increase in selectivity that was observed in the enzymatic assay. Compound 22 was inactive against JAK3 (IC50 > 10 \u03bcM) and had negligible potency toward TYK2 (849 nM). Kinome profiling of 22 indicated that the novel hinge pyrazole binding motif was highly selective for the JAK family and inhibited only one other kinase, Casein Kinase 1 Gamma 1 (CK1G1), with greater than 50% inhibition at a 100 nM concentration of 22 (see Supporting Information).", "# GitHub README: TopPDBLX\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/TopPDBLX, 2026\n\n16,971 of the \"unidentified\" components contain no chemistry at all: method notes (`streak seeded`), screen references (`hampton research index screen`), unnamed ligands (`protein`, `inhibitor`) and bare splitter fragments (`na`). No lexicon entry can ever match them, so counting them as reagents the parser failed to identify measures an artefact rather than the parser. They carry the role `not_a_component` with an auditable reason, and both denominators are published: **86.5%** over every component, **89.0%** over those that actually name a substance.\n\n### Why records are discarded\n\nEvery X-ray record is accounted for: it either parsed or carries exactly one reason code.\n\n| Reason | Records | Share |\n|---|---|---|\n| `NO_REAGENT_MATCH` | 7,603 | 3.82% |\n| `TOO_SHORT` | 4,369 | 2.19% |\n| `METHOD_ONLY` | 2,400 | 1.21% |\n| `UNPARSEABLE_RESIDUAL` | 301 | 0.15% |\n| `REFERENCE_ONLY` | 103 | 0.05% |\n| `EMPTY` | 91 | 0.05% |\n| `NON_CRYSTALLISATION_TEXT` | 87 | 0.04% |\n\n`REFERENCE_ONLY` at 103 records is worth noting: the brief expected \"see publication\" to be a major discard category, and it is negligible.\n\n## \ud83d\uddc2\ufe0f Ontology\n\n**In plain terms:** every crystallisation experiment is a mixture of a few ingredients. This sorts each one into a small set of buckets based on what kind of ingredient is doing the work, so you can ask \"what usually crystallises proteins like mine?\" and get an answer instead of a quarter of a million sentences.\n\nThere are seven buckets, plus an eighth for conditions that cannot be sorted honestly. That eighth is large, and that is the point: guessing would be worse than admitting the text does not say.\n\n| Class | Conditions | Share |\n|---|---|---|\n| Salt/PEG | 59,601 | 31.7% |\n| Unclassified | 37,208 | 19.8% |\n| PEG | 28,623 | 15.2% |\n| Salt | 27,922 | 14.8% |\n| Organic/PEG/Salt | 11,220 | 6.0% |\n| Organic/PEG | 9,629 | 5.1% |\n| Organic/Salt | 9,498 | 5.1% |\n| Organic | 4,338 | 2.3% |\n| **Classified** | **150,831** | **80.2%** |\n\nThe rules alone reach 59.5%. The rest is the model reading records no regular expression could,\nplus premixes now classified on what they are made of rather than refused.\n\n**For the crystallographer:** the seven classes are the JCSG Top96 precipitant classes, which are the non-empty subsets of {Organic, PEG, Salt}. Classification is by presence with no thresholds: a PEG is a PEG whatever its molecular weight and concentration, a salt is a salt whatever its chemistry and concentration. Buffers are excluded, since at 0.1 M a buffer sets the pH rather than precipitating anything (spec 6.3), and the pH is carried separately.\n\nUnclassified is a first-class answer with its reason recorded, never a null:\n\n| Reason it cannot be classified | Share |\n|---|---|\n| No amount stated for a precipitant: naming PEG without saying how much is not a measured condition | 14.2% |\n| No precipitant at all | 5.4% |\n| An unidentified reagent: the lexicon does not recognise a name, so nothing can be asserted | 0.4% |\n\n**The second row is a deliberate choice and the expensive one.** Those 24,327 conditions name their precipitant unambiguously and simply never state a concentration: `PEG6000, Sodium Chloride, VAPOR DIFFUSION`. Classifying them as `Salt/PEG` would lift classified coverage from 58.9% to roughly 72% in one line, and the rule that a PEG is a PEG whatever its concentration would arguably support it. They stay Unclassified because a condition with no concentration is not a measured condition, and coverage is not worth buying with a claim the deposition does not make.\n\n**Premixes are no longer refused, and that is a reversal.** Spec 6.4 was settled by declaring every premixed system Unclassified, which was right while a premix was an opaque token. Once the vendor compositions were transcribed (lexicon v0.9.0) it stopped being right: 59% of the conditions it refused were blocked by a premix made only of *buffers* (MES/imidazole, phosphate-citrate, MIB, SPG), and a buffer is excluded from naming the class anyway. Those were ordinary conditions declined for their packaging rather than their chemistry. A premix now contributes the chemistry it is made of, so Morpheus Precipitant Mix 4 is MPD with PEG 1000 and PEG 3350, which is `Organic/PEG`. Classified coverage moved 76.8% to **80.2%**. A premix with no transcribed composition is still Unclassified, because there is nothing to expand it into.\n\nAn earlier three-level ontology of 163 binned groups was withdrawn at v0.3.0. Its groups were derived by binning the corpus and then having labels retrofitted, which spec 6.1 rejects in its first line, and several were not chemically coherent: median purity across the 41 second-level groups was 49%. The full reasoning is in [`ontology/CHANGELOG.md`](ontology/CHANGELOG.md).\n\n## \u2697\ufe0f Reagent lexicon\n\n**In plain terms:** depositors write the same chemical a dozen different ways. \"PEG 3350\", \"peg3350\", \"polyethylene glycol 3350\" and \"PEG 3,350\" are one substance. The lexicon is the dictionary that maps every spelling onto one canonical name, and it is what makes the whole database queryable.", "# Protein stability: a crystallographer's perspective\nAuthors: Deller, M. C.; Kong, L.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 72, 72\u201395 (IYCr crystallization series), 2016\n\n| Quantity | Value |\n|---|---|\n| Net stability of a folded 10 kDa protein | ~10 kcal mol\u207b\u00b9 |\n| Contribution of hydrophobic effect to stability | ~60% |\n| Contribution of hydrogen bonding to stability | ~40% |\n| Burial of a single methyl group | ~1.1 kcal mol\u207b\u00b9 stabilizing; ~2.4 kcal mol\u207b\u00b9 entropic penalty |\n| Proline destabilization of \u03b1-helix vs alanine | \u0394\u0394G \u2248 3.16 kcal mol\u207b\u00b9 |\n| Human proteins with at least one disordered segment | ~40% |\n| Human proteins completely disordered | ~25% |\n| Crystallization success, Tm < 43\u00b0C | 23% of 657 samples |\n| Crystallization success, Tm > 45\u00b0C | 49% of 657 samples |\n| Crystallization success with high initial Thermofluor signal | 36.6% |\n| Negative-stain EM lower size limit | >150 kDa (crystallography works well <150 kDa) |\n| PCDDB entries at time of writing | 529 |\n| PTMs catalogued in SWISS-PROT/TrEMBL | just under 500 |\n\n## Key takeaways\n\n- Marc Deller, Leopold Kong and Bernhard Rupp argue that protein stability comes in two practically distinct flavours \u2014 compositional stability (a single, well-defined chemical species) and conformational stability (limited disorder and dynamics) \u2014 and that both must be satisfied before crystallization is even possible.\n- No amount of screening for thermodynamic or kinetic crystallization conditions will rescue a construct that cannot form stable, identical entities in the asymmetric unit, so stability work must precede screening.\n- A folded protein is only marginally stable, around 10 kcal mol\u207b\u00b9 for a 10 kDa protein, with roughly 60% of that stability from the hydrophobic effect and 40% from hydrogen bonding according to the work of Pace and coworkers.\n- Intrinsically disordered proteins are not a marginal problem: about 40% of human proteins contain a disordered segment and about 25% are wholly disordered, and NMR rather than crystallography has been the primary tool for studying them.\n- Post-translational modifications, particularly heterogeneous glycosylation and proteolysis, are the main practical destroyers of compositional homogeneity, but glycans occasionally mediate essential crystal contacts, as in the hepatitis C virus E2 glycoprotein complex.\n- Construct design is the highest-leverage decision available to the structural biologist: removing protease sites, PEST sequences and disordered termini, and choosing tags, often determines whether crystals ever appear.\n- Expression conditions and host choice matter as much as sequence, with low-temperature and cold-shock systems, cofactor supplementation, rare-codon strains such as Rosetta, protease-deficient BL21 Star and tightly regulated pBAD vectors all deployed to protect stability.\n- Thermofluor (differential scanning fluorimetry) is the review's workhorse analytical method, needing under 1 mg of protein and only a qPCR machine, and it supports buffer screening, ligand discovery and interaction quantification \u2014 though false positives and dye artefacts require careful melt-curve analysis.\n- Thermal stability correlates with crystallizability only up to a point: 23% of proteins with Tm below 43\u00b0C crystallized versus 49% above 45\u00b0C, but above 45\u00b0C Tm carries essentially no additional predictive power.\n- The salvage toolbox of structural genomics \u2014 truncation arrays, buffer and additive screening, reductive methylation, surface-entropy reduction, in situ proteolysis, DXMS-guided trimming and disulfide-engineered synthetic symmetrization \u2014 is presented as a systematic way to convert an uncrystallizable target into a tractable one.\n- Crystallizability predictors such as XtalPred, POODLE and DisMeta assess isolated single chains, so they systematically underrate targets that are stabilized only within a protein\u2013protein complex.\n- The authors close with Fred Richards's verdict that protein stability remains an unsolved problem, while pointing to XFELs, next-generation synchrotrons, cryo-EM and hybrid methods as the route to studying proteins too unstable or dynamic for classical crystallography.\n\n## References\n\n*(reference list omitted)*", "# Decoding the Molecular Machinery Behind Plastic Degradation\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe mechanistic understanding of *Is*PETase provides a foundation for rational enzyme engineering aimed at industrial-scale PET recycling. By systematically targeting residues involved in substrate binding, thermal stability, and catalytic efficiency, researchers can develop optimized enzyme variants capable of operating under industrially relevant conditions. The combination of ancestral sequence reconstruction, computational design approaches, and deep mutational scanning has already yielded variants with melting temperatures exceeding 20\u00b0C higher than wild-type PETase, demonstrating the feasibility of enhancing both stability and activity simultaneously. Furthermore, the development of enzyme cocktails combining PETase with downstream hydrolases such as MHETase and BHETase offers synergistic effects, accelerating the complete conversion of PET into its monomeric building blocks terephthalic acid and ethylene glycol.\u200b\n\nThe elucidation of PETase\u2019s molecular mechanism represents more than an academic achievement; it provides a blueprint for addressing one of the most pressing environmental challenges of our time. As researchers continue to refine these biological catalysts through protein engineering, the prospect of economically viable enzymatic PET recycling draws closer to reality. The lessons learned from studying *Is*PETase extend beyond plastic degradation, offering insights into enzyme evolution, structure-function relationships, and the rational design of biocatalysts for diverse industrial applications.\n\n| Key Concept | Description |\n| --- | --- |\n| Catalytic Triad | Ser160, His237, and Asp206 form the catalytic machinery; Ser160 acts as nucleophile attacking the ester carbonyl carbon |\n| Substrate Binding Subsites | L-shaped binding cleft divided into subsite I (binds first MHET moiety via \u03c0-\u03c0 interactions with Tyr87/Trp185) and subsite II (accommodates three MHET moieties in regions IIa, IIb, IIc) |\n| Nick Generation Step | Four MHET moieties bind across subsites I and II; cleavage of one ester bond creates a nick producing TPA-terminal and HE-terminal PET chains |\n| Terminal Digestion Step | Sequential hydrolysis of terminal PET chains (HE-terminal and TPA-terminal) into MHET monomers through repeated binding and ester bond cleavage |\n| Structural Features | Additional disulfide bond (Cys203-Cys239) near active site and extended \u03b28-\u03b16 loop enhance thermal stability and create continuous substrate binding groove |", "# Revolutionizing Protein Structure Analysis: Cryo-EM without Purification Accelerates the Future of Structural Biology\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-17 on marcdeller.com (categories: sb).*\n\n* New workflow enables solving protein structures via cryo-EM directly from crude cell-free lysates with no or minimal purification.\n* The PURE expression system provides minimal background, supporting atomic-resolution reconstructions in under 24 hours.\n* Miniaturized 25\u00b5l reactions are sufficient, reducing costs and labor.\n* Additional rapid cleanup steps yield modest resolution improvement with minimal extra time.\n* Approach opens doors for automation and high-throughput screening in structural biology.\n\n1. **Powell, S. M., & Evans, J. E. Cryo-EM protein structure without purification. [bioRxiv (2025)](https://doi.org/10.1101/2025.08.13.669967).**\n2. **Wu, M. & Lander, G. C. Present and Emerging Methodologies in Cryo-EM Single-Particle Analysis. [Biophysical Journal 119, 1281-1289 (2020)](https://doi.org/10.1016/j.bpj.2020.08.027).**\n3. **Kuruma, Y. & Ueda, T. The PURE system for the cell-free synthesis of membrane proteins. [Nature Protocols 10, 1328-1344 (2015)](https://doi.org/10.1038/nprot.2015.082).**\n4. **Punjani, A. et al. cryoSPARC: algorithms for rapid unsupervised cryo-EM structure determination. [Nat Methods 14, 290-296 (2017)](https://doi.org/10.1038/nmeth.4169).**\n5. **Su, C.-C. et al. A \u2018Build and Retrieve\u2019 methodology to simultaneously solve cryo-EM structures of membrane proteins. [Nature Methods 18, 69-75 (2021)](https://doi.org/10.1038/s41592-020-01021-2).**\n\n**From Gene to Structure in Under 24 Hours: How Minimizing Purification Is Transforming Cryo-EM**\n\nIn recent years, the pace of advancements in structural biology has accelerated dramatically, driven by breakthroughs in both computational and experimental methods. Yet, one sticking point lingered: the laborious and time-consuming purification of proteins before structural analysis. Now, a team at Pacific Northwest National Laboratory and Washington State University has pushed the boundaries by demonstrating that high-resolution cryo-electron microscopy (cryo-EM) structures can be solved directly from crude cell-free mixtures, bypassing lengthy purification steps\u2014and all in less than a day.\n\n**Breaking the Bottleneck in Structural Biology**\n\nFor decades, generating homogenous, purified protein samples has been the bottleneck in structural biology. Even with automation and modern electron microscopes capable of screening multiple samples per day, the initial sample preparation often lags far behind, taking days to months. This delay hinders the pace of fundamental research and drug discovery.\n\nThe study introduces a workflow that employs the PURE cell-free expression system. Unlike traditional cell-based methods or even wheat germ lysate systems, the PURE system utilizes a streamlined mixture of only 36 protein complexes, eliminating a vast background of interfering proteins\u2014critical for clear imaging and accurate structure determination. Experiments using this platform achieved robust expression of target proteins at reaction volumes as small as 25\u00b5l, generating enough sample for biochemical validation and cryo-EM analysis in a matter of hours.\n\n**Direct Imaging\u2014No Purification, No Problem**\n\nBy skipping purification, the process becomes not only faster, but also far more amenable to automation and high-throughput screening. The research team tested this \u201cstructure without purification\u201d workflow using Artemia ferritin and the plant protein PDX1.2 as benchmarks. Following rapid cell-free expression and minimal handling, lysates were applied directly to cryo-EM grids. Automated imaging and data processing rapidly filtered out non-target particles, leading to 3D reconstructions at atomic resolutions: 2.8\u00c5 for ferritin and 2.5\u00c5 for PDX1.2, surpassing or matching traditional purified samples.\n\nSmall-scale post-expression treatments\u2014like spin filtering, benzonase digestion, or his-tag affinity cleanup\u2014were evaluated for their incremental impact on map quality. Remarkably, these rapid interventions offered modest resolution improvements, bringing final maps as fine as 2.1\u00c5, but at the cost of only about an hour of added prep time. Crucially, even fully crude lysates yielded side-chain level detail sufficient for confident model building and refinement.\n\n**Broad Implications for the Field**\n\nThe elimination of large-scale purification unlocks significant advantages:\n\n* **Reduced costs:** Miniaturized reactions cut reagent use and waste.\n* **Accelerated workflow:** The entire process from expression to model can be completed in less than 24 hours.\n* **Expanded accessibility:** Proteins that are unstable, toxic, or otherwise intractable in traditional systems become evaluable.\n* **Compatibility with automation:** Small volumes and minimal handling simplify integration with automated sample preparation and imaging platforms.\n\nNotably, the presence of ribosomes and other lysate components unexpectedly improved ice quality during vitrification, often a tricky variable in cryo-EM. The authors further note that the PURE system is compatible with modern genetic engineering to introduce post-translational modifications or folding-assisting factors, reducing concerns about protein quality in the absence of a full cellular context.\n\n**A New Era for Structural Biology**\n\nThis innovative workflow demonstrates that direct-from-lysate, cell-free cryo-EM analysis is more than a proof-of-concept\u2014it is a practical, robust, and scalable technique poised to accelerate discovery across biology and medicine. With resolutions previously achievable only through painstaking purification now accessible in hours, the field is set for a leap in throughput and impact.", "# Cell surface receptors\nAuthors: Deller, M.C.; Jones, E.Y.\nSource: Current Opinion in Structural Biology, 2000\n\nAs with many cell surface receptors in general, numerous cytokine receptors use immunoglobulin (Ig)-like domains to mediate ligand binding. Although this classifies them as members of the Ig superfamily (IgSF), there is no such unifying theme discernable in the architecture of their ligands. The complex structures determined to date reveal a varied spectrum of ligand\u2013receptor interactions (Figure 1). A major addition to the gallery this past year is the structure of the complex between the prototypic neurotrophin nerve growth factor (NGF) and a portion of the extracellular domain from one of its receptors, TrkA [8] (Figure 1). The neurotrophins are cytokines belonging to the nervous system and function through interaction with the Trk family of high-affinity receptors. Our understanding of the determinants of neurotrophin\u2013Trk pairwise specificities has been bolstered by a supplementary paper providing the structures of the neurotrophin-binding domains of all three types of Trk receptor [9]. The extracellular regions of the Trk receptors each contain five domains; however, the available structures cannot provide clear insight into the role of the N-terminal four domains, other than to imply that these domains are unlikely to make direct contact with the neurotrophin ligand. The available structural information does, however, provide a picture that is consistent with the idea of receptor cross-linking, with the interaction with dimeric neurotrophin acting to trigger signalling. This picture contrasts totally with the data for the interleukin (IL)-1\u2013IL-1 receptor (IL-1R) family of cytokine\u2013receptor interactions [10,11] (Figure 1). In this system, the ligand is monomeric, the interactions are spread over three Ig-like receptor domains and the 1:1 complex provides no direct model for the mechanism of signal initiation. Rather, this last is believed to require the formation of a ternary complex with an additional molecule, the IL-1R accessory protein.\n\nThe NGF\u2013TrkA complex does, however, provide a story that is broadly analogous to that of dimeric vascular endothelial growth factor (VEGF) complexed with a domain from its receptor (Flt-1), the structure of which was determined in the same laboratory two years earlier [12] (Figure 1). The Flt-1 story has been further embellished this past year with a report on the solution structure of its VEGF-binding domain in isolation and a second crystal structure of this Ig-like domain in complex with VEGF, solved in a different crystal form [13]. As with numerous other biomedically important cytokine\u2013receptor interactions, the challenge now is to exploit structural data in the development of novel therapeutic agents. A report on a crystal structure of an antagonist peptide complexed to VEGF discusses one dead end reached in this quest [14].\n\nA peptide-based approach has also recently been employed to provide the first structural toehold in the study of chemokine\u2013receptor interactions. The multi-membrane spanning nature of the chemokine receptors presents a substantial obstacle to high-resolution structural studies, prompting the use of a peptide fragment from CXC-chemokine receptor 1 in an NMR study of IL-8 binding [15].\n\n*Figure 1.* The varied architecture of IgSF receptor\u2013cytokine complexes. Schematic representations of the complexes are shown with the cytokine coloured black in each case. (a) 1:2 complex of the dimeric NGF ligand with domain 5 of the TrkA receptor [8]. (b) 1:1 complex of IL-1\u03b2 with the extracellular region of IL-1R [10]. (c) 1:2 complex of the dimeric VEGF ligand with domain 2 of the Flt-1 receptor [12]. All figures produced using programmes BOBSCRIPT [54] and RASTER3D [55].\n\n## Hematopoietic receptor complexes\n\nIn contrast to the variety of complexes discussed in the previous section, the four-helix bundle cytokines form a single superfamily of ligands matched by a single receptor superfamily. The interactions among examples of these molecules yielded the first structural insights into any form of cell surface receptor\u2013ligand complex and continue to provide many of the paradigms for cytokine\u2013receptor recognition and signalling. The original body of work on growth hormone and its receptors was ably reviewed in 1996 [16]. This past year, complexes of two new but related systems, those of the cytokine granulocyte colony stimulating factor (GCSF) in complex with two domains from its receptor (GCSF-R) [17] and of IL-4 with the extracellular region of its high-affinity, specificity conferring receptor (IL-4R\u03b1) [18], have provided fresh food for thought, with additional twists on the established theme.", "# Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\nAuthors: Kong, L.; Lee, J.H.; Doores, K.J.; Murin, C.D.; Julien, J.-P.; McBride, R.; Liu, Y.; Marozsan, A.; Cupo, A.; Klasse, P.-J.; Hoffenberg, S.; Caulfield, M.; King, C.R.; Hua, Y.; Le, K.M.; Khayat, R.; Deller, M.C.; Clayton, T.; Tien, H.; Feizi, T.; Sanders, R.W.; Paulson, J.C.; Moore, J.P.; Stanfield, R.L.; Burton, D.R.; Ward, A.B.; Wilson, I.A.\nSource: Nature Structural & Molecular Biology, 2013\n\nIt is still not clear why the Asn332 glycan is so dominant among antibodies targeting this supersite, but one possibility is that it is not only highly conserved (73%) and predominantly oligomannose, but also conformationally restrained by adjacent sugars. Indeed, the glycan Asn332 conformation is nearly identical in the crystal structures of PGT 128 and PGT 135 (Fig. 6), suggesting that the immune system is presented with a comparatively well-ordered glycan shield, at least for glycan Asn332. Heterogeneity of surrounding glycans may be accommodated, as suggested by the distribution of angles of approach utilized by PGT 135, and in the case of PGT 121 potentially being able to bind both high mannose and complex glycans. One implication for vaccine design is that immunization with an antigen containing a heterogeneous mix of glycoforms and glycosylation sites may promote \"promiscuous\" antibodies that have greater neutralization breadth. However, this benefit may be countered by a dampened immune response, in which case different vaccination schedules and adjuvants would need to be tested. Ultimately, to mimic this supersite using rational vaccine design, such as epitope scaffolding, to create vaccine immunogens, one should carefully consider how to reconstitute a faithful representation of the glycan canopy with glycoforms that are similar to those in the intact Env trimer.\n\n## Online Methods\n\n### Expression and purification of PGT 135 and 17b Fab\n\nPGT 135 Fab was expressed using a baculovirus system with SF9 cells as described previously. PGT 135 was purified from the cell supernatant first by anti-human lambda affinity chromatography followed by cation exchange chromatography and size exclusion chromatography using Superdex 200\u2122 (GE Healthcare) as previously described. Fab 17b was obtained from digesting IgG 17b with papain using a previously described protocol.\n\n### Expression and purification of JRFL gp120 core\n\nA JRFL gp120 core containing a short (mini) V3 region (residues 305-320 deleted and replaced by a proline residue) was cloned into a phCMV3 vector with an IgK secretion signal. This plasmid was used to transfect HEK 293S GnT1\u2212/\u2212 (293S) cells using 293Fectin (Invitrogen) under serum free conditions. The protein was purified first through a Galanthus nivalis lectin column, followed by SEC with Superdex 200\u2122 (GE Healthcare).\n\n### Expression and purification of BG505 SOSIP.664 trimer\n\nThe HIV-1 clade A BG505 Env sequence and the construction of SOSIP.664 trimers using the Env sequence of BG505.W6M.ENV.C2 (GenBank Accession ABA61516/DQ208458) designed with a T332N mutation is described elsewhere. The BG505 SOSIP.664 trimer was expressed and purified as previously described. The BG505 SOSIP.664 trimer was purified using a 2G12-coupled affinity matrix followed by passage through a sizing column.\n\n### Crystallization and data collection\n\nUnliganded PGT 135 Fab crystallized over a period of 28 days at 20 \u00b0C in a crystallization reagent consisting of 20% (w/v) PEG 8000, 0.1 M CHES, pH 9.5. Crystals were harvested and cryoprotected by a brief immersion in 70% well buffer, 30% glycerol, followed by immediate flash-cooling in liquid nitrogen. Data were collected at APS beamline 23ID-B (wavelength: 1.033 \u00c5) at 100 K. Data were processed and scaled with HKL-2000. Quaternary complexes of PGT 135 crystallized in 20% PEG 2000, 0.1 M Tris pH 7.0 (JCSG CoreSuite I Well C07) using our automated CrystalMation robotic system (Rigaku). An optimization screen was made around this condition and large single crystals were obtained from 16% w/v PEG MME 2000, Tris pH 7.87 using the Oryx8 Crystallization robot. Data were collected at the ALS beamline 5.0.2 (wavelength: 1.000 \u00c5) at 100 K, and were processed and scaled with HKL-2000. All data processing statistics are summarized in Table 1.\n\n### Structure determination and refinement\n\nThe unliganded PGT 135 structure was determined by the molecular replacement method using Phaser with an unrelated Fab structure (PDB ID: 3KYM) as an initial model. For the quaternary complex, multiple components were used for phasing: 17b Fab (PDB ID: 2NXY), soluble CD4 (PDB ID: 2NXY), gp120 core (PDB ID: 2NXY) and high-resolution unliganded PGT 135 Fab as determined here. Model building was carried out using Coot-0.6.2 and refinement was implemented with the Phenix program. See Supplementary Note for refinement details and Table 1 for final refinement statistics.\n\n### Isothermal titration calorimetry\n\nIsothermal titration calorimetry (ITC) binding experiments were performed using a MicroCal iTC200 instrument (GE). See Supplementary Note for details.\n\n### gp120 binding ELISAs\n\nRecombinant gp120 (250 ng) was immobilised directly onto flat bottom microtitre plates (Costar type 3690, Corning Inc.) at 4 \u00b0C overnight. Antibody binding was determined as described above. Fold change in binding for PGT 135 mutants is summarized in Supplementary Table 9.\n\n### Generation of pseudovirus\n\nPseudovirus was generated in HEK 293T or GnT1\u2212/\u2212 deficient 293S cells as described previously. Glycosidase inhibitors were added at the time of transfection and were used alone or in combination at the following concentrations: 25 \u03bcM kifunensine and 2 mM N-butyldeoxynojirimycin (NB-DNJ).\n\n### Neutralization assays\n\nNeutralization activity of antibodies against pseudovirus in TZM-bl cells was determined as described previously. Fold change in neutralization for PGT 135 mutants is summarized in Supplementary Table 10.\n\n### Antibody and envelope mutations\n\nMutations in the PGT heavy chain, the HIV-1 envelope glycoprotein and the JRFLmV3 construct were introduced using QuikChange site-directed mutagenesis (Stratagene, La Jolla, CA). Mutations were verified by DNA sequencing (Eton Biosciences, La Jolla, CA).\n\n### JR-FLmV3 glycan mutant binding ELISAs", "# Molecular Architecture of GLP-1 Receptor Drug Binding Reveals Dynamic Activation Mechanisms\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe glucagon-like peptide-1 receptor (GLP-1R) has emerged as a pivotal therapeutic target for type 2 diabetes and obesity, yet understanding how drugs engage this class B G protein-coupled receptor (GPCR) requires deciphering complex molecular choreography across multiple receptor domains. Recent structural biology advances, particularly through cryo-electron microscopy (cryo-EM) and molecular dynamics simulations, have revealed that GLP-1R activation involves a sophisticated interplay between peptide binding dynamics, conformational changes in transmembrane helices, and coordinated movements of extracellular domains. The receptor employs a two-domain binding mechanism where the [extracellular domain (ECD) captures peptide C-termini](https://www.nature.com/articles/celldisc201642) as an affinity trap while the N-terminal peptide segments insert into the [transmembrane domain (TMD) orthosteric pocket](https://www.nature.com/articles/s41467-021-27760-0) to trigger activation. This binding mode creates a deep, amphipathic pocket formed by residues from transmembrane helices 1, 2, 3, 5, 6, and 7, with critical contributions from [Y152 (TM1), K197 (TM2), Q234 and Y241 (TM3), W306 and R310 (TM5), and E364 and H363 (TM6)](https://academic.oup.com/mend/article/25/10/1804/2614704). Peptide N-terminal residues, particularly His1 and Glu9 in GLP-1, establish both [hydrogen bonds and hydrophobic contacts](https://www.pnas.org/doi/10.1073/pnas.2303696120) with these binding pocket residues, with interaction persistence correlating directly with signaling efficacy. Molecular dynamics simulations have illuminated that [different agonists exhibit distinct dynamics](https://www.nature.com/articles/s41467-021-27760-0) within this pocket, with native GLP-1 forming more stable interactions at the cavity base compared to oxyntomodulin or exendin-4. The extracellular loops play specialized roles: [ECL2 stabilizes the peptide and couples binding to activation](https://pmc.ncbi.nlm.nih.gov/articles/PMC3281730/), ECL1 undergoes conformational rearrangements that differ between agonists, and ECL3 transitions from ECD-stabilizing interactions in the inactive state to [facilitating TM7 rearrangements](https://www.sciencedirect.com/science/article/pii/S109727652030650X) during activation. The hallmark of GLP-1R activation is the [outward movement of TM6\u2019s intracellular region by approximately 11.9 \u00c5](https://www.nature.com/articles/s41421-024-00649-0) at residue R348, creating a cavity for Gs protein engagement. Notably, class B GPCRs like GLP-1R exhibit unique activation kinetics where [Gs interaction can precede TM6 opening](https://pmc.ncbi.nlm.nih.gov/articles/PMC12347461/), contrasting with class A GPCRs, and once opened, [TM6 relaxes back slowly](https://pmc.ncbi.nlm.nih.gov/articles/PMC12347461/), enabling sustained signaling. The ECD serves dual functions as both affinity trap and [intrinsic agonist](https://www.nature.com/articles/celldisc201642), with its conformational transition from a closed TMD-interacting state to an extended position being essential for activation, even when peptides are covalently linked to the receptor. Recent structural studies of therapeutically relevant agonists like [tirzepatide](https://www.pnas.org/doi/10.1073/pnas.2116506119) have revealed that its weaker affinity for GLP-1R compared to GIPR stems from a bulkier N-terminal Tyr1 causing steric conflicts with W306, resulting in biased pharmacology with reduced receptor desensitization. Small molecule agonists bind to a [more tightly packed orthosteric pocket](https://pmc.ncbi.nlm.nih.gov/articles/PMC7784854/) with distinct ECL3 and TM7 positioning compared to peptides, yet converge on similar mechanisms to induce the critical TM6 conformational change. This structural understanding establishes that transient versus persistent interactions at the binding pocket base directly influence G protein activation kinetics, explaining how different agonists can produce [biased signaling profiles](https://www.nature.com/articles/s41467-021-27760-0) despite targeting the same receptor, and provides a molecular framework for rational design of next-generation therapeutics with tailored pharmacological properties.", "# Revolutionizing Protein Crystallography: The Synergy of TELSAM-DARPin Fusion and Electric Field Optimization\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe structural insights from TELSAM-DARPin fusion crystals reveal sophisticated adaptation mechanisms within the polymer matrix. TELSAM polymers can adjust their helical rise to accommodate fused target proteins, with observed minimum helical rises of 46.9 \u00c5 for single-helix polymers. The technology eliminates the requirement for direct inter-polymer contacts, dramatically expanding the theoretical size limit of target proteins that can be successfully crystallized. Crystal structures demonstrate that fusion to TELSAM polymers increases binding avidity, stabilizing otherwise weak inter-target protein crystal contacts through approximately 5.4 \u00d7 10^8 minimal contact interactions.\n\nThis convergence of engineered fusion systems and physical field manipulation represents a fundamental advancement in structural biology methodology. The combination enables researchers to transform historically difficult crystallization targets into routine structural determinations, potentially accelerating drug discovery and protein engineering applications where high-resolution structures are essential for rational design approaches.\n\n| Key Concept | Description | Key References |\n| --- | --- | --- |\n| TELSAM Crystallization Chaperone | pH-dependent polymer-forming protein that creates ordered crystal lattices when fused to target proteins, enabling crystallization at ultra-low concentrations | [Nawarathnage, S.D., et al., BYU ScholarsArchive](https://scholarsarchive.byu.edu/cgi/viewcontent.cgi?article=12031&context=etd) |\n| DARPin Engineering | Designed ankyrin repeat proteins with consensus sequences forming stable, modular binding scaffolds with high target specificity and thermal stability | [G\u00fcttler, T., et al., PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC9611651/) |\n| Linker Optimization | Strategic design of rigid, semi-flexible, and flexible connectors between TELSAM and target proteins, with Pro-Ala linkers showing optimal performance for DARPin targets | [Nawarathnage, S.D., et al., PubMed](https://pubmed.ncbi.nlm.nih.gov/40950120/) |\n| Electric Field Crystallization | Application of weak AC electric fields (6 V/mm, 1 kHz) to enhance protein crystallization kinetics by modifying protein-protein interaction potentials | [Ray, D., et al., Journal of Physical Chemistry Letters](https://pubs.acs.org/doi/10.1021/acs.jpclett.4c01744) |\n| Polymer Orientation Control | Electric field-induced alignment and organization of TELSAM polymers to optimize crystal packing and prevent polymer flipping that disrupts lattice periodicity | [Hekstra, D.R., et al., Nature Communications](https://pmc.ncbi.nlm.nih.gov/articles/PMC5730412/) |\n| Avidity Enhancement | TELSAM fusion systems stabilize weak crystal contacts through massive numbers (~5.4 \u00d7 10\u2078) of minimal intermolecular interactions, overcoming traditional crystallization barriers | [Anderson, M.J., et al., Open Biology](https://pmc.ncbi.nlm.nih.gov/articles/PMC8889177/) |", "# 2025 \u2013 Embracing AI-Enabled Communication and Adaptability\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-21 on marcdeller.com (categories: certificates).*\n\n* Reinforce professional credibility by showcasing AI and communication expertise.\n* Bridge interdisciplinary teams through effective AI-powered dialogue.\n* Cultivate resilience with adaptability-focused learning.\n* Build foundational AI knowledge via accredited programs.\n* Enhance decision-making by leveraging generative AI tools.\n* Strengthen interpersonal connections in hybrid work settings.\n* Demonstrate commitment to lifelong learning in cutting-edge technologies.\n* Position yourself as a forward-thinking leader in AI initiatives.\n\n1. **AI-Driven Communication**\n   **and Collaboration \u2013**Amplify Your Communication Skills with Generative AI: [Certificate](https://www.linkedin.com/learning/certificates/e73941e7b279574ef2e43e461acd40efb644f9284f285d1210cee7580ee9528a)\n2. **Adaptability and Resilience in AI \u2013** Building an Adaptability Mindset in the Age of AI: [Certificate](https://www.linkedin.com/learning/certificates/8fd265595da5c6a811a98b77161dfbe5a887b53439fa7ca014db31e58ab8a0f4)\n3. **Foundational AI Knowledge \u2013** Introduction to Artificial Intelligence (AI) \u2013 IBM: [Certificate](https://www.coursera.org/account/accomplishments/verify/71R8R1FHA194)\n\nWith rapid technological advances, professionals who master AI-driven tools for communication gain a decisive edge. Certifications such as *Amplify Your Communication Skills with Generative AI*and *Building an Adaptability Mindset in the Age of AI* empower you to navigate complex collaborations and drive innovation. Earning the *Introduction to Artificial Intelligence (AI)* credential from IBM deepens your foundational understanding, positioning you to integrate AI solutions effectively into everyday workflows. Together, these certifications validate both technical prowess and the essential soft skills needed to lead and adapt in AI-driven environments.", "# Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly\nAuthors: Deller, M.C.; Johnson, H.A.; Miller, M.D.; Spraggon, G.; Elsliger, M.-A.; Wilson, I.A.; Lesley, S.A.\nSource: PLoS ONE, 2015\n\nSeveral structural studies have shown that RCK domains form homodimeric assemblies arranged in a clamshell-like arrangement with a hinge region formed by two domain-swapped helices. This arrangement is observed in many crystal structures including the M. thermoautotrophicum potassium channel (MthK) [15\u201318], M. jannaschii TrkA [19], B. subtilis KtrA [19,20], E. coli K+ channel [10], E. coli ybaL, V. parahaemolyticus TrkA [21] and Danio rerio Ca2+-activated K+ channel (BK) [22]. Structural and biochemical studies have also shown that these RCK homodimers assemble into higher-order quaternary structures to form cytosolic regulatory components referred to as gating ring assemblies, which are typically composed of a tetrameric arrangement of protein chains, with each chain itself composed of an RCK dimer. This assembly typically results in a symmetrical octameric arrangement of RCK domains that form a central pore on the cytosolic side of the channel.\n\nSeveral crystal structures of K+ transporter gating ring assemblies have been determined including MthK [16,18], KtrA [20,23], BK [22,24], GsuK [25] and TrkA [21]. The gating ring assembly combines with the membrane spanning units to form a functional transporter complex. The RCK domains of the gating ring assembly undergo conformational rearrangements in their dimeric associations in response to external stimuli, such as Ca2+ and nucleotides. These conformational changes are symmetrical in nature and involve an expansion or contraction of the central ring of RCK domains, which propagate to the associated membrane-spanning channel resulting in regulation of ion flow across the membrane [21,22]. For example, the BK channel gating ring assembly has been eloquently described as undergoing a conformation change that opens the channel \"like the petals of a flower\" [22].\n\nPrevious protein expression experiments and phylogentic analysis conducted in our lab have shown that the TM1088 locus of T. maritima codes for two proteins designated as TM1088A and TM1088B, which combine to form an octameric assembly [9]. The octameric TM1088 structure is the first example of a two-subunit Trk K+ transporter composed of two different proteins. This two-subunit complex is produced from a single promoter, as a result of overlapping start and stop codons, and an internal ribosome binding site within the TM1088 locus. This scenario is suggestive of a gene duplication event similar to that proposed in the evolution of E. coli TrkA and other Trk/KtrAB proteins [14,26].\n\nGiven the somewhat atypical genetics and protein composition of the TM1088 assembly, it is remarkable to find that the structure conforms to the structural paradigms observed in all other RCK-containing K+ transporters studied to date. Similarities of the TM1088 assembly to other gating ring structures are highlighted, along with some unique features including an interfacial salt bridge across the central cavity and several protein interfaces with implications for gating ring assembly and mechanism. This structure provides a framework for future structural and functional studies on the mechanisms of K+ ion regulation by this two-subunit family of K+ transporters.\n\n## Materials and Methods\n\n### Strains, plasmids, recombinant techniques, expression and affinity purification\n\nThe TM1088 locus (GenBank: NP_228894) consists of two separate open reading frames with overlapping start and stop codons and, as such, encodes two separate proteins designated TM1088A (UniProt: D3KFX7_THEMA, Pfam02254) and TM1088B (UniProt: D3KFX8_THEMA, Pfam02254 and Pfam02080) (Fig 1A). In some databases TM1088A is, however, annotated as locus TM_1087.1 and TM1088B as locus TM_1088. For the purpose of this study, plasmid pMH4-TM1088 was used to produce TM1088A:TM1088B complexes (TM1088A: full length, residues 1\u2013143 and TM1088B: full length, residues 1\u2013218) and TM1088A (full length, residues 1\u2013143) through use of different purification schemes. Plasmid pMH4-TM1088 was constructed as previously described as part of our structural genomics exploration into the entire T. maritima proteome [27]. Plasmid pSpeedET-TM1088B was used to produce TM1088B (residues 140\u2013218) as previously described [9].", "# Crystal structure of a soluble cleaved HIV-1 envelope trimer\nAuthors: Julien, J.-P.; Cupo, A.; Sok, D.; Stanfield, R.L.; Lyumkis, D.; Deller, M.C.; Klasse, P.-J.; Burton, D.R.; Sanders, R.W.; Moore, J.P.; Ward, A.B.; Wilson, I.A.\nSource: Science, 2013\n\n**Authors:** Jean-Philippe Julien, Albert Cupo, Devin Sok, Robyn L. Stanfield, Dmitry Lyumkis, Marc C. Deller, Per-Johan Klasse, Dennis R. Burton, Rogier W. Sanders, John P. Moore (corresponding author), Andrew B. Ward (corresponding author) and Ian A. Wilson (corresponding author). **Affiliations (condensed):** Department of Integrative Structural and Computational Biology, The Scripps Research Institute, La Jolla, CA 92037, USA; IAVI Neutralizing Antibody Center, The Scripps Research Institute; Center for HIV/AIDS Vaccine Immunology and Immunogen Discovery, The Scripps Research Institute; Weill Medical College of Cornell University, New York, NY 10021, USA; Department of Immunology and Microbial Science, The Scripps Research Institute; National Resource for Automated Molecular Microscopy, The Scripps Research Institute; Joint Center for Structural Genomics, The Scripps Research Institute (M.C.D.); Ragon Institute of MGH, MIT, and Harvard, Cambridge, MA 02129, USA; Department of Medical Microbiology, Academic Medical Center, Amsterdam, Netherlands; Skaggs Institute for Chemical Biology, The Scripps Research Institute. **Published:** Science, volume 342, issue 6165, pages 1477\u20131483, 20 December 2013.\n\n**Summary line:** The overall architecture of the structure involved in HIV entry provides a blueprint for vaccine design.\n\n## Abstract\n\nHIV-1 entry into CD4+ target cells is mediated by cleaved envelope glycoprotein (Env) trimers that have been challenging to characterize structurally. Here, we describe the crystal structure at 4.7 \u00c5 of an antigenically near-native, cleaved, stabilized, soluble Env trimer (termed BG505 SOSIP.664 gp140) in complex with a potent broadly neutralizing antibody, PGT122. The structure shows a pre-fusion state of gp41, the interaction between the component gp120 and gp41 subunits, and how a close association between the gp120 V1/V2/V3 loops stabilizes the trimer apex around the three-fold axis. The complete epitope of PGT122 on the trimer involves gp120 V1, V3 and several surrounding glycans. This trimer structure advances our understanding of how Env functions and is presented to the immune system, and provides a blueprint for structure-based vaccine design.\n\n## Introduction\n\nThe envelope glycoprotein (Env) trimer is the only virally encoded antigen on the surface of HIV-1, the pathogen responsible for the global AIDS epidemic, and is responsible for viral entry into host cells. The trimer is composed of gp120/gp41 heterodimers and is the target for neutralizing antibodies. Various structures of components of gp120 and gp41, alone and in complex with different ligands, have been determined. Cryo-electron microscopy (EM) and tomography have been integrated with core gp120 x-ray structures to visualize the Env trimer at resolutions that extend from 30 \u00c5 to below 10 \u00c5, and thereby provide insights into its overall conformation before and after receptor binding. However, determining an atomic-level structure of the Env trimer has been difficult. A higher resolution structure would not only help to understand how the trimer functions during virus-cell fusion, but also guide HIV-1 vaccine design by delineating the key antigenic sites recognized by the humoral immune system and the defenses evolved by the virus as a counter-measure.\n\nDuring Env synthesis, gp160 precursors trimerize and are subsequently cleaved by proteases of the furin family into gp120 and gp41 subunits, which associate non-covalently before the native complex reaches the surface of infected cells and is then packaged onto virions. Cleavage is obligatory for Env trimers to function in viral infection of target cells. Virus-cell fusion is a multistep process involving three major Env conformations, each with distinct roles: 1) pre-fusion (interacts with CD4 receptor); 2) extended gp41 intermediate (interacts with CCR5 or CXCR4 co-receptors); and 3) gp41 six-helix bundle (hemi-fusion of viral and cell membranes).", "# Self-Adapting Language Models: Are We Creating Our Final Invention?\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Concept | Description | Key References |\n| --- | --- | --- |\n| **Self-Adapting Language Models (SEAL)** | Framework enabling LLMs to generate their own finetuning data and optimization directives through reinforcement learning, achieving persistent weight updates without external supervision | [Zweiger et al., arXiv](https://arxiv.org/html/2506.10943v2) [Pari et al., MIT](https://jyopari.github.io/posts/seal) |\n| **Recursive Self-Improvement** | Process where an AI system enhances its own capabilities without human intervention, potentially leading to exponential intelligence growth and superintelligence | [Wikipedia](https://en.wikipedia.org/wiki/Recursive_self-improvement) [Kumar, Future of Life Institute](https://futureoflife.org/ai/the-unavoidable-problem-of-self-improvement-in-ai-an-interview-with-ramana-kumar-part-1/) |\n| **Intelligence Explosion** | Theoretical scenario where recursive self-improvement triggers rapid, uncontrollable advancement from AGI to superintelligence, compressing decades of progress into months or years | [Hastings-Woodhouse, Future of Life Institute](https://futureoflife.org/ai/are-we-close-to-an-intelligence-explosion/) [MacAskill, Forethought Centre](https://www.youtube.com/watch?v=SjSl2re_Fm8) |\n| **Catastrophic Forgetting** | Phenomenon where neural networks lose previously learned information when acquiring new knowledge, limiting continual learning capabilities in self-adapting systems | [Zweiger et al., arXiv](https://arxiv.org/html/2506.10943v2) [Irie et al., TMLR](https://arxiv.org/abs/2312.00276) |\n| **Synthetic Data Generation** | AI-created training data that replicates statistical properties of real-world information, addressing data scarcity while preserving privacy and enabling continued model improvement | [IBM Research](https://www.ibm.com/think/insights/ai-synthetic-data) [DataHub Analytics](https://datahubanalytics.com/synthetic-data-generation-addressing-data-scarcity-in-machine-learning/) |\n| **AI Alignment Problem** | Challenge of ensuring advanced AI systems remain aligned with human values and goals as they become increasingly autonomous and potentially surpass human intelligence | [Wikipedia](https://en.wikipedia.org/wiki/AI_alignment) [AI Alignment Forum](https://www.alignmentforum.org/posts/axKWaxjc2CHH5gGyN/ai-will-not-want-to-self-improve) |\n| **Meta-Learning** | \u201cLearning to learn\u201d paradigm where AI systems optimize their own learning processes, adaptation strategies, and algorithm selection rather than just task performance | [Irie et al., TMLR](https://arxiv.org/abs/2312.00276) [TAILOR Network](https://tailor-network.eu/meta-learning-for-continual-learning/) |", "# ChemSage: Vibe-Coding a Chemically Aware Drug-Discovery LLM on Local Hardware\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n#Terminal 1 \u2014 start the model server:\ncd /Users/dellboy/Documents/Vibe_Coding/chem_sage\nsource .venv/bin/activate\nmlx_lm.server --model models/chem_sage_32b_v5 --port 8080\n\n#Terminal 2 \u2014 run chat.py:\ncd /Users/dellboy/Documents/Vibe_Coding/chem_sage\nsource .venv/bin/activate\npython scripts/chat.py --model models/chem_sage_32b_v5\n\n```\n\nChat.py interface showing \"get all EGFR structures\"\n\nChat.py response showing tabulated PDB ligand-structure pairs.\n\nChat.py response with options retrieved from RAG corpus including 1) Bioactivity, 2) ATC classification, 3) PDB structures, 4) PDB binding affinities, and 5) PDB ligand-structure pairs.\n\n###### What comes next?\n\nTwo threads are open. The first is getting ChemSage in front of people: wrapping the chat CLI (`chat.py`) in a Gradio interface hosted on marcdeller.com, standing up a lightweight Hugging Face Spaces demo, exposing a local endpoint for live talks through a Cloudflare Tunnel, and packaging the client as a `pipx`-installable tool so anyone with a local Qwen 32B can point it at their own server. The second is a planned Round 6 aimed squarely at the residual weak spots. Distribution shift, not raw example count, caused Round 4\u2019s regressions, so the priority is weighted class-balanced sampling across all 78 classes to lock the fragile metrics in place. Beyond that: triple the code-to-quote drill (still the weakest metric), add a dedicated PyMOL syntax generator, try a two-phase curriculum that follows a full run with a short fidelity-focused continuation, raise the LoRA rank to 96, and run a per-class evaluation on a mid-training checkpoint so class-level dilution shows up after an hour rather than after forty.\n\n###### To Do List\n\nLots to learn with some ideas below, and in all honesty, I don\u2019t fully understand what all these mean yet! The joys of Vibe Coding! Build quick, learn later!\n\n* Host on my lightweight DigitalOcean Droplet at [mdeller.com](http://mdeller.com)?\n* Serve via Ollama or Web-UI?\n* HuggingFace Spaces hosting through Cloudflare tunnel?\n* Build a single-file package installer?\n* Round 6 refined training?\n  + Weighted class-balanced sampling\n  + Improve code-to-quote corpus\n  + Pymol syntax generator\n  + Raise to 96 layers\n  + Two-phase curriculum\n  + Mid-training checkpoints", "# GitHub README: chatmol\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/chatmol, 2026\n\n> **Natural-language control of PyMOL. Describe the scene, get the structure.**\n\n    \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/chatmol\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/chatmol</a></td>\n</tr>\n</table>\n\n---\n\nA natural-language chat frontend for **PyMOL (open-source, 3.1.0)**, powered\nby a local **Ollama** model (default: `qwen3.6:27b`).\n\nWhy it matters: PyMOL is the lingua franca of molecular visualisation, but producing a polished figure often means stacking a dozen exact commands for selections, representations, colouring, and views. ChatMol layers a local LLM directly onto PyMOL's command line so plain-English requests (\"colour by chain and show the ligand as sticks\") become correct PyMOL syntax instantly, while any valid PyMOL you type still runs untouched. It is useful for structural biologists, crystallographers, and cryo-EM scientists who want to move faster from raw coordinates to publication-quality scenes, lower the barrier for students learning PyMOL, and keep every byte of structural data local.\n\nType plain English into PyMOL's normal command line:\n\n```\nPyMOL> fetch 1abc and show it as a cartoon coloured by chain, then show the ligand as sticks\n[ChatMol] > fetch 1abc, async=0\n[ChatMol] > hide everything\n[ChatMol] > show cartoon, polymer\n[ChatMol] > util.cbc polymer\n[ChatMol] > show sticks, organic\n```\n\nAnything that already looks like valid PyMOL syntax is run unchanged, so\nyour existing scripts and muscle memory keep working. Anything that looks\nlike English is sent to a local LLM, translated into PyMOL commands, echoed\nto the console, and executed. Everything runs locally -- no data leaves\nyour machine.\n\n---\n\n## 1. Prerequisites\n\nYou need two things installed before the plugin itself: **Ollama** (with a\nmodel pulled) and **open-source PyMOL 3.1.0**.\n\n### 1.1 Ollama + model\n\n1. Install Ollama:\n   - **macOS**: download the app from <https://ollama.com/download>, or\n     `brew install ollama`\n   - **Linux**: `curl -fsSL https://ollama.com/install.sh | sh`\n   - **Windows**: download the installer from <https://ollama.com/download>\n\n2. Make sure the server is running. The desktop app starts it\n   automatically; otherwise start it manually and leave it running:\n\n   ```sh\n   ollama serve\n   ```\n\n3. Pull a Qwen3.6 model:\n\n   ```sh\n   ollama pull qwen3.6:27b\n   ```\n\n   This is the default ChatMol uses, and gives the best translation\n   quality. It's a ~17 GB download and needs roughly 20 GB+ of free\n   RAM/VRAM to run comfortably. If your machine can't handle that:\n\n   ```sh\n   # Much smaller/faster, good for trying ChatMol out, lower translation quality:\n   ollama pull qwen3:0.6b\n   ```\n\n   Then point ChatMol at whichever tag you pulled (see\n   [Configuration](#4-configuration)) -- just make sure the `model` setting\n   matches a tag from `ollama list` exactly.\n\n4. Sanity check the server is reachable and the model responds:\n\n   ```sh\n   curl http://localhost:11434/api/chat -d '{\n     \"model\": \"qwen3.6:27b\",\n     \"messages\": [{\"role\": \"user\", \"content\": \"say hi\"}],\n     \"stream\": false\n   }'\n   ```\n\n   You should get back a JSON response with a `message.content` field. If\n   this fails, fix it before installing the plugin -- see\n   [Troubleshooting](#7-troubleshooting).\n\nIf you run Ollama on a different host/port, configure that too (see\n[Configuration](#4-configuration)).\n\n### 1.2 PyMOL (open-source, 3.1.0)\n\nPick whichever matches how you normally manage Python environments:\n\n- **conda / mamba** (any OS):\n\n  ```sh\n  conda create -n pymol -c conda-forge -c schrodinger pymol-open-source=3.1.0\n  conda activate pymol\n  ```\n\n- **Homebrew (macOS)**:\n\n  ```sh\n  brew install pymol\n  ```\n\n- **Linux package managers** also commonly provide `pymol` (e.g.\n  `apt install pymol`), though the bundled version may differ from 3.1.0.\n\nLaunch it once (`pymol`) to confirm it starts normally and note where its\nPython lives -- that's the interpreter ChatMol will run under. ChatMol only\nuses the Python standard library, so no extra `pip install` is required.\n\n---\n\n## 2. Installing the plugin\n\nChatMol is a plain Python package (the `chatmol/` directory). There are two\nways to load it; pick one.\n\n### Option A -- `~/.pymolrc.py` (recommended, most reliable)\n\nThis runs ChatMol from the source directory directly -- easiest to update\n(just edit the files and restart PyMOL).\n\nCreate or edit `~/.pymolrc.py` (in your home directory) and add:\n\n```python\nimport sys\nsys.path.insert(0, \"/Users/dellboy/Documents/Vibe Coding/chatmol\")\nimport chatmol\n```\n\n(Adjust the path if you move the `chatmol` project folder.)\n\nRestart PyMOL. You should see the `[ChatMol] AI chat interface loaded`\nbanner in the console -- that confirms it's active.\n\n### Option B -- Plugin Manager (`chatmol.zip`)\n\nA ready-to-install zip is included at the project root:\n`chatmol.zip` (contains `chatmol/__init__.py` etc. at its root, as the\nPlugin Manager expects).\n\n1. In PyMOL: **Plugin > Plugin Manager > Install New Plugin > Choose\n   file...** and select `chatmol.zip`.\n2. Restart PyMOL.\n\nThis also adds a **Plugin > ChatMol Settings** menu item with a small GUI\nfor configuration.\n\nIf you edit the source after installing this way, rebuild the zip from the\nproject root:\n\n```sh\ncd \"/Users/dellboy/Documents/Vibe Coding/chatmol\"\nrm -f chatmol.zip\nzip -r chatmol.zip chatmol -x '*.pyc' -x '*__pycache__*' -x '*.DS_Store'\n```\n\nand reinstall it.\n\n### Verifying the install\n\nEither way, after restarting PyMOL you should see:", "# gobsmacked.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n| Tool | What it does here | Licence | Links |\n| --- | --- | --- | --- |\n| ESMFold (esm) | Folding the sequence when no model exists | MIT | [repository](https://github.com/facebookresearch/esm) \u00b7 [Lin et al. 2023, Science](https://doi.org/10.1126/science.ade2574) |\n| OpenFold | ESMFold's structure module | Apache-2.0 | [repository](https://github.com/aqlaboratory/openfold) \u00b7 [Ahdritz et al. 2024, Nature Methods](https://doi.org/10.1038/s41592-024-02272-z) |\n| PDBFixer | Missing atoms, protonation, cleaning before MD | MIT | [repository](https://github.com/openmm/pdbfixer) \u00b7 The OpenMM project |\n| PandaDock | Pose search and SE(3) GNN rescoring | MIT | [repository](https://github.com/pritampanda15/PandaDock) \u00b7 [Panda 2026, bioRxiv](https://doi.org/10.64898/2026.08.19.745667) |\n| OpenMM | Minimisation, equilibration and production MD | MIT | [repository](https://github.com/openmm/openmm) \u00b7 [Eastman et al. 2024, Journal of Physical Chemistry B (OpenMM 8)](https://doi.org/10.1021/acs.jpcb.3c06662) |\n| openmmforcefields / OpenFF | Ligand parameters (OpenFF 2.1.0) alongside Amber14 | MIT | [repository](https://github.com/openforcefield/openff-toolkit) \u00b7 [Boothroyd et al. 2023, Journal of Chemical Theory and Computation](https://doi.org/10.1021/acs.jctc.3c00039) |\n| Boltz-2 | Affinity head scored on the relaxed complex, its diffusion module bypassed | MIT | [repository](https://github.com/jwohlwend/boltz) \u00b7 [Passaro et al. 2025, bioRxiv](https://doi.org/10.1101/2025.06.14.659707) |\n| Boltzina | The pattern followed, not a dependency: an external pose scored by Boltz-2's affinity head with the structure module omitted | MIT | [repository](https://github.com/ohuelab/Boltzina) \u00b7 [Furui and Ohue 2025, arXiv](https://doi.org/10.48550/arXiv.2508.17555) |\n| MDTraj | Trajectory analysis, in the bundle and on the server | LGPL-2.1 | [repository](https://github.com/mdtraj/mdtraj) \u00b7 [McGibbon et al. 2015, Biophysical Journal](https://doi.org/10.1016/j.bpj.2015.08.015) |\n\n### Analyze\n\n| Tool | What it does here | Licence | Links |\n| --- | --- | --- | --- |\n| PLIP | Interaction fingerprints, run as a subprocess only | GPL-2.0 | [repository](https://github.com/pharmai/plip) \u00b7 [Adasme et al. 2021, Nucleic Acids Research](https://doi.org/10.1093/nar/gkab294) |\n| PandaMap | 2D interaction maps and an empirical binding-energy estimate | MIT | [repository](https://github.com/pritampanda15/PandaMap) \u00b7 Panda 2025, PandaMap |\n| biotite | Sequence alignment behind every residue-number mapping | BSD-3-Clause | [repository](https://github.com/biotite-dev/biotite) \u00b7 [Kunzmann and Hamacher 2018, BMC Bioinformatics](https://doi.org/10.1186/s12859-018-2367-z) |\n| tmtools / TM-align | Whole-chain TM-score, reported as context for the pocket fit | MIT | [repository](https://github.com/jvkersch/tmtools) \u00b7 [Zhang and Skolnick 2005, Nucleic Acids Research](https://doi.org/10.1093/nar/gki524) |\n| PoseBusters (criteria) | The pose-validity checks, reimplemented on this app's own geometry | BSD-3-Clause | [repository](https://github.com/maabuu/posebusters) \u00b7 [Buttenschoen et al. 2024, Chemical Science](https://doi.org/10.1039/D3SC04185A) |\n| gemmi | Structure input and output, mmCIF and PDB | MPL-2.0 | [repository](https://github.com/project-gemmi/gemmi) |\n\n### Browser\n\n| Tool | What it does here | Licence | Links |\n| --- | --- | --- | --- |\n| Mol* | Every 3D view, including the three-way overlay | MIT | [repository](https://github.com/molstar/molstar) \u00b7 [Sehnal et al. 2021, Nucleic Acids Research](https://doi.org/10.1093/nar/gkab314) |\n| Plotly.js | Trajectory panels | MIT | [repository](https://github.com/plotly/plotly.js) \u00b7 plotly.com |\n\n### Hosting\n\n| Tool | What it does here | Licence | Links |\n| --- | --- | --- | --- |\n| Flask, gunicorn, nginx, SQLite | The web application and its store | BSD / MIT / public domain | [repository](https://flask.palletsprojects.com) |\n| pixi and uv | The bundle's environment, and the server's | BSD-3-Clause / Apache-2.0 | [repository](https://github.com/prefix-dev/pixi) |\n\n## Credits and licence\n\nGOBSMACKED 1.0.0 was written by Marc C. Deller, D.Phil.\n([marc@marcdeller.com](mailto:marc@marcdeller.com),\n[marcdeller.com](https://marcdeller.com)). Source and issues:\n[github.com/bellcheddar/GOBSMACKED](https://github.com/bellcheddar/GOBSMACKED).\n\nThe application is MIT licensed. To cite it, name the repository and the version.\n\nPLIP is GPL-2.0. GOBSMACKED runs it as a subprocess and never imports it, so no\nGPL code is linked into this application, and PLIP is deliberately absent from the run bundle.\nEvery other component's licence is in the table above.", "# GitHub README: TopPDBLX\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/TopPDBLX, 2026\n\n| Term | What it means here |\n|---|---|\n| **Crystallisation condition** | The recipe a protein was crystallised from: a few reagents, their concentrations, and a pH. This project parses about 200,000 of them |\n| **Deposition text** | What the scientist actually typed into the PDB, free-hand, in no agreed format. The raw material |\n| **Precipitant** | The reagent doing the work: it pulls the protein out of solution. Usually a PEG, a salt or an organic |\n| **Drop / reservoir** | The two halves of a vapour-diffusion experiment. The drop holds protein plus condition; the reservoir pulls water out of it |\n| **Mother liquor** | The solution a crystal grew in and sits in |\n| **Cryoprotectant** | Added *after* growth to stop ice forming when the crystal is frozen. Real chemistry, but it did not crystallise anything |\n| **Premix** | A vendor mixture sold as one bottle (Morpheus \"Divalents\", Tacsimate) that is really several reagents |\n| **PEG** | Polyethylene glycol, the commonest precipitant. The number is its molecular weight, not an amount: \"PEG 3350\" is a size |\n| **Hofmeister rank** | An ordering of salts by how strongly they push proteins out of solution |\n| | |\n| **Lexicon** | The dictionary mapping every spelling a depositor might use onto one **canonical id**. `peg3350`, `PEG 3,350` and `polyethylene glycol 3350` all become `PEG_3350` |\n| **Canonical id** | The single name this project uses for a reagent, in `UPPER_SNAKE_CASE` |\n| **The residual** | The records the rule parser could **not** read. This is what the language model is for, and it is never trained on them |\n| **The gold sets** | Two batches of 96 records each, 192 in total, labelled by hand in `app/gold_bench_v1.html`. The only ground truth here, used to measure and never to train |\n| **LoRA** | A way of fine-tuning a model by training a small adapter rather than all its weights. Ours is 33 MB against a 360M-parameter base |\n| **Distillation** | Training a small model on a larger one's output. Here a 32-billion-parameter *teacher* labels text for a 360-million-parameter *student* |\n| **Grounding** | Checking that a reagent the model named actually appears in the text it was given. Catches invention, which no accuracy score can |\n| **Identification** | Whether an emitted name exists in the lexicon. A weaker test than grounding: it doesn't check that the name is *right* for this text, only that it's real |\n\n## \ud83d\udd04 Workflow\n\n**In plain terms:** the same chemical appears in the archive as \"PEG 3350\", \"peg3350\" and \"polyethylene glycol 3,350\", amounts are written in a dozen notations, and some entries are whole paragraphs of narrative. The work is to turn that text into a table you can query, without inventing anything that is not there.\n\nIt runs as a chain of stages. Each is one command, each writes a manifest recording its inputs, tool versions and git state, and each can be re-run on its own.\n\n```\n   RCSB PDB archive\n          \u2502\n          \u2502  ingest.*          fetch every deposition over GraphQL, then byte-compare\n          \u25bc                    the text against the archive mmCIF: 100% agreement\n   199,185 condition strings\n          \u2502\n          \u2502  parse.run_parser  split into clauses, read amounts and units, look each\n          \u25bc                    name up in the reagent lexicon\n   645,656 typed components \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n          \u2502  parse.scope     tells the drop \u2502 47,845 records the rules cannot read\n          \u2502  from the protein sample and a  \u2502\n          \u2502  assign.classify                \u25bc\n          \u25bc                          \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n   7 precipitant classes             \u2502   THE TEACHING LOOP                     \u2502\n   or Unclassified + reason          \u2502                                         \u2502\n          \u2502                          \u2502   Qwen2.5-32B  \u2500\u2500labels\u2500\u2500\u25ba  SmolLM2-360M\u2502\n          \u2502  link.*                  \u2502   models.          the      models.     \u2502\n          \u25bc                          \u2502   teacher_label    student  train_slm   \u2502\n   sequences + MMseqs2 clusters      \u2502        \u25b2                        \u2502       \u2502\n          \u2502                          \u2502        \u2502                        \u2502       \u2502\n          \u2502                          \u2502        \u2514\u2500\u2500\u2500\u2500 gold_metrics \u25c4\u2500\u2500\u2500\u2500\u2500\u2518       \u2502\n          \u2502                          \u2502              scores both against        \u2502\n          \u2502                          \u2502              192 hand-labelled records  \u2502\n          \u2502                          \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n          \u2502                                           \u2502 names it wants but\n          \u2502                                           \u2502 the lexicon lacks\n          \u2502                                           \u25bc\n          \u2502                                    gold_bench \u00b7 courtroom\n          \u2502                                    an expert, tens of decisions\n          \u2502                                           \u2502\n          \u2502  release.assemble                         \u25bc\n          \u25bc                              ontology/synonyms.yaml \u2500\u2500\u2510\n   JSONL \u00b7 Parquet \u00b7 CSV \u00b7 DuckDB \u00b7 JSON Schema \u00b7 datasheet \u25c4\u2500\u2500\u2500\u2500\u2500\u2518\n                                                    the lexicon feeds back\n                                                    into the next parse\n```\n\n### The teaching loop, and why there are two models\n\n**Why bother with two models: the teacher is roughly 90x larger and far slower to run.** At around 40 seconds per record, a 32B model would take weeks to read every residual record the rules cannot parse; the 360M student reads them in under two hours. So the teacher labels a few thousand of the hardest cases, and the student does the corpus-scale work, which took recall against hand-labelled truth from **69.5% (rules alone) to 90.3% (rules + round 09, pooled)**.", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\nTM0077 from Thermotoga maritima is a member of the carbohydrate esterase family 7 and is active on a variety of acetylated compounds, including cephalosporin C. TM0077 esterase activity is confined to short-chain acyl esters (C2-C3), and is optimal around 100\u00b0C and pH 7.5. The positional specificity of TM0077 was investigated using 4-nitrophenyl-\u03b2-D-xylopyranoside monoacetates as substrates in a \u03b2-xylosidase-coupled assay. TM0077 hydrolyzes acetate at positions 2, 3, and 4 with equal efficiency. No activity was detected on xylan or acetylated xylan, which implies that TM0077 is an acetyl esterase and not an acetyl xylan esterase as currently annotated. Selenomethionine-substituted and native structures of TM0077 were determined at 2.1 and 2.5 \u00c5 resolution, respectively, revealing a classic \u03b1/\u03b2-hydrolase fold. TM0077 assembles into a doughnut-shaped hexamer with small tunnels on either side leading to an inner cavity, which contains the six catalytic centers. Structures of TM0077 with covalently bound phenylmethylsulfonyl fluoride and paraoxon were determined to 2.4 and 2.1 \u00c5, respectively, and confirmed that both inhibitors bind covalently to the catalytic serine (Ser188). Upon binding of inhibitor, the catalytic serine adopts an altered conformation, as observed in other esterase and lipases, and supports a previously proposed catalytic mechanism in which Ser hydroxyl rotation prevents reversal of the reaction and allows access of a water molecule for completion of the reaction.\n\n## Structure of hepatitis C virus envelope glycoprotein E2 antigenic site 412 to 423 in complex with antibody AP33 (2012)\n\n**Authors:** Kong L, Giang E, Nieusma T, Robbins JB, Deller MC, Stanfield RL, Wilson IA, Law M.  \n**Journal:** Journal of virology (2012). DOI: 10.1128/jvi.01939-12. PMID: 22973046.\n\nWe have determined the crystal structure of the broadly neutralizing antibody (bnAb) AP33, bound to a peptide corresponding to hepatitis C virus (HCV) E2 envelope glycoprotein antigenic site 412 to 423. Comparison with bnAb HCV1 bound to the same epitope reveals a different angle of approach to the antigen by bnAb AP33 and slight variation in its \u03b2-hairpin conformation of the epitope. These structures establish two different modes of binding to E2 that antibodies adopt to neutralize diverse HCV.\n\n## Crystal structure of a metal-dependent phosphoesterase (YP_910028.1) from Bifidobacterium adolescentis: Computational prediction and experimental validation of phosphoesterase activity (2011)\n\n**Authors:** Han GW, Ko J, Farr CL, Deller MC, Xu Q, Chiu HJ, Miller MD, Sefcikova J, Somarowthu S, Beuning PJ, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA, Ondrechen MJ.  \n**Journal:** Proteins (2011). DOI: 10.1002/prot.23035. PMID: 21538547.\n\nThe crystal structures of an unliganded and adenosine 5'-monophosphate (AMP) bound, metal-dependent phosphoesterase (YP_910028.1) from Bifidobacterium adolescentis are reported at 2.4 and 1.94 \u00c5, respectively. Functional characterization of this enzyme was guided by computational analysis and then confirmed by experiment. The structure consists of a polymerase and histidinol phosphatase (PHP, Pfam: PF02811) domain with a second domain (residues 105-178) inserted in the middle of the PHP sequence. The insert domain functions in binding AMP, but the precise function and substrate specificity of this domain are unknown. Initial bioinformatics analyses yielded multiple potential functional leads, with most of them suggesting DNA polymerase or DNA replication activity. Phylogenetic analysis indicated a potential DNA polymerase function that was somewhat supported by global structural comparisons identifying the closest structural match to the alpha subunit of DNA polymerase III. However, several other functional predictions, including phosphoesterase, could not be excluded. Theoretical microscopic anomalous titration curve shapes, a computational method for the prediction of active sites from protein 3D structures, identified potential reactive residues in YP_910028.1. Further analysis of the predicted active site and local comparison with its closest structure matches strongly suggested phosphoesterase activity, which was confirmed experimentally. Primer extension assays on both normal and mismatched DNA show neither extension nor degradation and provide evidence that YP_910028.1 has neither DNA polymerase activity nor DNA-proofreading activity. These results suggest that many of the sequence neighbors previously annotated as having DNA polymerase activity may actually be misannotated.\n\n## A conserved fold for fimbrial components revealed by the crystal structure of a putative fimbrial assembly protein (BT1062) from Bacteroides thetaiotaomicron at 2.2\u2005\u00c5 resolution (2010)\n\n**Authors:** Xu Q, Abdubek P, Astakhova T, Axelrod HL, Bakolitsa C, Cai X, Carlton D, Chen C, Chiu HJ, Chiu M, Clayton T, Das D, Deller MC, Duan L, Ellrott K, Farr CL, Feuerhelm J, Grant JC, Grzechnik A, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Krishna SS, Kumar A, Marciano D, McMullan D, Miller MD, Morse AT, Nigoghossian E, Nopakun A, Okach L, Puckett C, Reyes R, Sefcovic N, Tien HJ, Trame CB, van den Bedem H, Weekes D, Wooten T, Yeh A, Zhou J, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309110006548. PMID: 20944223.\n\nBT1062 from Bacteroides thetaiotaomicron is a homolog of Mfa2 (PGN0288 or PG0179), which is a component of the minor fimbriae in Porphyromonas gingivalis. The crystal structure of BT1062 revealed a conserved fold that is widely adopted by fimbrial components.\n\n## Bacterial pleckstrin homology domains: a prokaryotic origin for the PH domain (2010)", "# 2016 \u2013 Key gp120 glycans pose roadblocks to the rapid development of VRC01-class antibodies in an HIV-1- infected Chinese donor\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: papers).*\n\n* [Link to Paper](https://marcdeller.com/wp-content/uploads/2025/09/PIIS1074761316300954.pdf)\n\n* **Glycan Barrier Mechanisms**: The N276 glycan creates rigid steric constraints that force antibodies to accommodate rather than displace carbohydrate modifications during binding interactions.\n* **Structural Adaptation Requirements**: VRC01-class antibodies must develop sophisticated complementarity-determining region architectures to navigate glycan obstacles while maintaining neutralization breadth.\n* **Population-Specific Evolution**: Chinese HIV-1 populations exhibit distinct viral subtypes and glycosylation patterns that influence antibody development trajectories differently than other global populations.\n* **Germline-Targeting Success**: Recent clinical trials demonstrate that engineered immunogens can successfully prime rare VRC01-class precursors and guide their evolution toward glycan accommodation.\n* **Somatic Hypermutation Demands**: Broadly neutralizing antibody development requires extensive mutation accumulation (10-16% sequence divergence) that challenges conventional B cell survival mechanisms.\n* **Therapeutic Implications**: Understanding glycan accommodation mechanisms enables rational design of sequential immunization strategies that could overcome natural barriers to broadly neutralizing antibody development.\n\n1. **Structural insights into VRC01-class bnAb precursors with diverse light chains from the IAVI G001 clinical trial:** [Andrabi, R., et al., PNAS](https://pmc.ncbi.nlm.nih.gov/articles/PMC12377726/)\n2. **Precise targeting of human HIV broadly neutralizing antibody precursors:** [Escolano, A., et al., Nature Immunology](https://pmc.ncbi.nlm.nih.gov/articles/PMC12313413/)\n3. **Structural basis of glycan276-dependent recognition by HIV-1 broadly neutralizing antibodies:** [Gao, F., et al., Cell Reports](https://www.sciencedirect.com/science/article/pii/S2211124721013954)\n4. **Mechanisms of antibody gene diversification during somatic hypermutation:** [Yeap, L.S., et al., Science China Life Sciences](https://medicalxpress.com/news/2024-08-mechanisms-antibody-gene-diversification-somatic.html)\n5. **Induction of HIV Neutralizing Antibody Lineages in Mice with Diverse Precursor Repertoires:** [Escolano, A., et al., Cell Host & Microbe](https://pmc.ncbi.nlm.nih.gov/articles/PMC5103708/)\n6. **Functional development of a V3/glycan-specific broadly neutralizing antibody lineage from an HIV-1 clade A/D infected Kenyan donor:** [Williams, K.L., et al., eLife Sciences](https://elifesciences.org/articles/68110)\n\n**Unraveling the Glycan Fortress: How N276 Carbohydrates Hinder VRC01-Class Antibody Evolution in HIV-Infected Populations**\n\nBridges critical gaps in HIV vaccine design by uncovering\u2014through structural analyses (X-ray/EM), longitudinal tracking, and functional assays\u2014how early VRC01-class antibody precursors overcome glycan obstacles via light-chain adaptations, revealing both roadblocks (N276/V5 glycan clashes) and a rapid maturation pathway for broad neutralization, guiding next-gen immunogens targeting the CD4-binding site.\n\nThe development of broadly neutralizing antibodies against HIV represents one of medicine\u2019s most challenging pursuits, particularly when examining the intricate molecular roadblocks that impede rapid antibody maturation. Recent investigations into HIV-1 infected Chinese populations have revealed critical insights into how specific glycan modifications create formidable barriers to VRC01-class antibody development. The N276 glycan, positioned strategically on the gp120 envelope protein, emerges as a pivotal determinant in antibody evolution pathways, fundamentally altering the trajectory of immune responses in infected individuals.\n\nVRC01-class broadly neutralizing antibodies represent the holy grail of HIV immunology, derived from rare IGHV1-2 immunoglobulin heavy chain gene segments that enable recognition of the conserved CD4-binding site across diverse viral strains. These remarkable antibodies demonstrate exceptional breadth, neutralizing up to 90% of circulating HIV-1 variants when fully matured. However, their development requires extensive somatic hypermutation, with mature antibodies accumulating 10-16% sequence divergence from their germline precursors. This extraordinary level of mutation presents a significant challenge for vaccine development, as most B cells lack the intrinsic capacity to undergo such extensive modification while maintaining functional antibody production.\n\nThe glycan shield surrounding HIV\u2019s envelope protein serves as an evolutionary masterpiece of immune evasion, with the N276 glycan functioning as a particularly effective molecular guardian. Structural analyses reveal that this carbohydrate modification creates steric hindrance that restricts antibody access to vulnerable epitopes on the CD4-binding site. Unlike other glycans that may be displaced during antibody binding, the N276 glycan maintains its rigid orientation, forcing developing antibodies to accommodate its presence rather than circumvent it. This accommodation requirement places additional constraints on antibody evolution, necessitating specific structural adaptations that may not naturally occur during conventional immune responses.\n\nContemporary research demonstrates that glycan-dependent antibodies must maintain the ligand-free orientation of the N276 glycan to achieve effective neutralization. This finding contradicts earlier assumptions about antibody-glycan interactions and highlights the sophisticated co-evolution between viral evasion mechanisms and host immune responses. Antibodies that attempt to displace the N276 glycan face energetic penalties that compromise their binding affinity, while those that successfully accommodate the glycan through complementarity-determining region modifications achieve superior neutralization breadth.", "# Collaboration: 2+ state-of-the-art Structural Biology facilities built through global partnerships\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-29 on marcdeller.com (categories: 6cs).*\n\n## Collaboration: 2+ state-of-the-art Structural Biology facilities built through global partnerships\n\nCollaboration has been central to my most ambitious projects, especially when building [2+ structural biology laboratories](https://marcdeller.com/building-a-comprehensive-protein-production-x-ray-crystallography-and-cryo-em-laboratory-from-scratch/) from the ground up. Establishing fully integrated protein production, X-ray crystallography, and cryo-EM laboratories is a team effort that spans scientists, engineers, operations, and leadership. My role has often been to connect these threads into a coherent vision and then execute against it step by step. By investing in collaborative, future-proof platforms, we create ecosystems where new ideas, people, and technologies can plug in and flourish over many years.\n\n* Collaboration underpins every enduring success in modern structural biology.\n* Building comprehensive research facilities requires bringing together international expertise from industry and academia.\n* It matters because shared vision and teamwork amplify capability far beyond individual effort.\n* Day-to-day, collaboration lives at the interfaces spanning cross-functional collaborations between biology, phramacology, structural biology, chemistry, biology, and computation.\n* Every protein structure is a collaboration between CRO resources, collaborators, protein scientists, and synchrotron beamline scientists.\n\nPersonal Philosophy: The 6C's\n\n### [**Curiosity:** 10+ years driving hands-on drug discovery through multidisciplinary approaches](https://marcdeller.com/curiosity-20-years-driving-hands-on-discovery-through-multidisciplinary-approaches/)\n\n[**Curiosity:** 10+ years driving hands-on drug discovery through multidisciplinary approaches](https://marcdeller.com/curiosity-20-years-driving-hands-on-discovery-through-multidisciplinary-approaches/)\n\n### [**Courage**: 400+ challenging structures determined, enabling breakthrough mechanistic insights](https://marcdeller.com/courage-400-structures-determined-enabling-breakthrough-mechanistic-insights/)\n\n[**Courage**: 400+ challenging structures determined, enabling breakthrough mechanistic insights](https://marcdeller.com/courage-400-structures-determined-enabling-breakthrough-mechanistic-insights/)\n\n### [**Creativity:** 7+ patents through innovative protein engineering and AI integration](https://marcdeller.com/creativity-5-patents-through-innovative-protein-engineering-and-ai-integration/)\n\n[**Creativity:** 7+ patents through innovative protein engineering and AI integration](https://marcdeller.com/creativity-5-patents-through-innovative-protein-engineering-and-ai-integration/)\n\n### [**Communication:** 60+ publications translating structural data into actionable insights](https://marcdeller.com/communication-60-publications-translating-structural-data-into-actionable-insights/)\n\n[**Communication:** 60+ publications translating structural data into actionable insights](https://marcdeller.com/communication-60-publications-translating-structural-data-into-actionable-insights/)\n\n### [**Compassion:** 6+ FDA IND filings supported while mentoring 20+ scientists](https://marcdeller.com/compassion-6-fda-ind-filings-supported-while-mentoring-20-scientists/)\n\n[**Compassion:** 6+ FDA IND filings supported while mentoring 20+ scientists](https://marcdeller.com/compassion-6-fda-ind-filings-supported-while-mentoring-20-scientists/)", "# SiteDoctor \u2013 Full diagnosis, health, content, and SEO for any WordPress site\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nIf none of `--scan`, `--inventory`, or `--analytics` are given, **all three\nrun**. Note that `--inventory` and `--analytics` both depend on the content\ninventory, so requesting `--analytics` alone will still fetch the inventory\nin the background \u2014 it just won\u2019t be printed/included unless `--inventory`\nis also given.\n\n### Other options\n\n| Flag | Description |\n| --- | --- |\n| `url` | Target site URL (optional \u2014 prompted for if omitted), e.g. `https://example.com` |\n| `--max-pages N` | Number of pages to run the per-page SEO analysis on during `--analytics` (default: `10`). Increase for a more complete audit of larger sites, at the cost of one extra HTTP request per page. |\n| `--no-psi` | Skip the Google PageSpeed Insights API call entirely and go straight to the self-computed performance scorecard. Useful when offline, rate-limited, or for faster runs. |\n| `--out PATH` | Output HTML report path (default: `./site-doctor-report-<host>-<timestamp>.html`) |\n\n## Output\n\n* A color-coded report is printed to the terminal (colors auto-disable when\n  output isn\u2019t a TTY).\n* A branded, self-contained HTML report is written to disk, containing every\n  section run during that invocation (scorecard, tables, badges, and\n  per-page cards).\n\n## How it works\n\nSite Doctor avoids all third-party dependencies by implementing its own\nminimal HTTP/1.1 client on top of `socket` and `ssl`, which lets it measure\neach phase of a request (DNS, connect, TLS, TTFB, download) precisely. HTML\nis parsed with regular expressions rather than a DOM library, DNS records\nare fetched via `dig` (with a `socket`-based fallback), and WordPress data is\nread entirely from the site\u2019s existing public REST API endpoints \u2014 no\nplugins, authentication, or scraping required.\n\n## Disclaimer\n\nSite Doctor only reads publicly available information (HTTP responses, DNS\nrecords, and public REST API endpoints). Use it against sites you own or are\nauthorized to test.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/site-doctor](https://github.com/bellcheddar/site-doctor) |", "# Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3\nAuthors: Shvartsbart, A.; Roach, J.J.; Witten, M.R.; Koblish, H.; Harris, J.J.; Covington, M.; Hess, R.; Lin, L.; Frascella, M.; Truong, L.; Leffet, L.; Conlen, P.; Beshad, E.; Klabe, R.; Katiyar, K.; Kaldon, L.; Young-Sciame, R.; He, X.; Petusky, S.; Chen, K.-J.; Horsey, A.; Lei, H.-T.; Epling, L.B.; Deller, M.C.; Vechorkin, O.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2022\n\n**Step 1. 3-Bromo-5,7-dichloropyrazolo[1,5-a]pyrimidine (45).** To a solution of 5,7-dichloropyrazolo[1,5-a]pyrimidine (2.0 g, 10.6 mmol) in DMF (20 mL) was added NBS (2.1 g, 11.7 mmol), and the reaction mixture was stirred at room temperature for 1 h. Water was added, and the resulting precipitate was filtered, washed with water, and air-dried overnight to afford the title compound (2.43 g, 86%). The product was used without purification. LCMS calculated for C6H3BrCl2N3 (M + H)+: m/z = 265.9; found: 265.9.\n\n**Step 2. 3-Bromo-5-chloropyrazolo[1,5-a]pyrimidin-7-amine (46).** A suspension of 3-bromo-5,7-dichloropyrazolo[1,5-a]pyrimidine (2.4 g, 9.0 mmol) in concentrated ammonium hydroxide (30 mL, 0.4 mol) was heated to 100 \u00b0C in a heavy-walled sealed tube for 1 h. After cooling to room temperature, the reaction mixture was diluted with cold water and filtered. The collected solid was then washed with ice-cold water and air-dried to afford the title compound (2.19 g, 98%). LCMS calculated for C6H5BrClN4 (M + H)+: m/z = 246.9; found: 246.9.\n\n**Step 3. 3-Bromo-5,6-dichloropyrazolo[1,5-a]pyrimidin-7-amine (47).** To a solution of 3-bromo-5-chloropyrazolo[1,5-a]pyrimidin-7-amine (2.19 g, 8.85 mmol) in DMF (20 mL) was added NCS (1.18 g, 8.85 mmol), and the reaction mixture was stirred at 50 \u00b0C for 1 h. After cooling to room temperature, ice chips and water were added, and the resulting precipitate was filtered, washed with water, and air-dried to afford the title compound (2.08 g, 82%). LCMS calculated for C6H4BrCl2N4 (M + H)+: m/z = 280.9; found: 280.9.\n\n**Step 4. 3-Bromo-5,6-dichloropyrazolo[1,5-a]pyrimidine (48).** To a solution of 3-bromo-5,6-dichloropyrazolo[1,5-a]pyrimidin-7-amine (11 g, 39 mmol) in THF (100 mL) was added tert-butyl nitrite (23 mL, 0.2 mol), and the reaction mixture was heated to 80 \u00b0C for 2 h. The reaction mixture was concentrated, loaded directly onto silica gel, and purified by flash chromatography (0\u221220% EtOAc/hexanes) to afford the title compound (6.7 g, 64%).\n\n**Step 5. (S)-3-Bromo-6-chloro-N-(tetrahydrofuran-3-yl)pyrazolo[1,5-a]pyrimidin-5-amine (49a).** A mixture of 3-bromo-5,6-dichloropyrazolo[1,5-a]pyrimidine (500 mg, 1.87 mmol), (S)-tetrahydrofuran-3-amine (485 \u03bcL, 5.62 mmol), and cesium carbonate (732 mg, 2.25 mmol) in THF (6 mL) was heated to 80 \u00b0C for 2 h. The reaction mixture was cooled to room temperature, diluted with EtOAc, and filtered. The filtrate was concentrated in vacuo, and the residue was purified by flash chromatography (0\u221250% EtOAc/hexanes) to afford the title compound (414 mg, 70%) as a light yellow solid. LCMS calculated for C10H11BrClN4O (M + H)+: m/z = 317.0; found: 316.8.\n\n**Step 6. (S)-5-(6-Chloro-5-((tetrahydrofuran-3-yl)amino)pyrazolo[1,5-a]pyrimidin-3-yl)-N-(methyl-d3)nicotinamide (50a).** A mixture of (S)-3-bromo-6-chloro-N-(tetrahydrofuran-3-yl)pyrazolo[1,5-a]pyrimidin-5-amine (50 mg, 0.16 mmol), N-(methyl-d3)-5-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)nicotinamide (50 mg, 0.19 mmol), Pd(dppf)Cl2\u00b7CH2Cl2 (13 mg, 16 \u03bcmol), and sodium carbonate (60 mg, 0.6 mmol) in 1,4-dioxane (1.5 mL) and water (0.5 mL) was sparged with N2 and heated to 100 \u00b0C for 2 h. The reaction mixture was adsorbed onto silica gel directly. The product was purified by flash chromatography (0\u2212100% EtOAc/hexanes followed by 0\u221220% MeOH/DCM) to afford the title compound (15 mg, 25%). LCMS calculated for C17H15D3ClN6O2 (M + H)+: m/z = 376.1; found: 376.2.\n\n**Step 7. (S)-5-(6-(1-Isopropyl-1H-pyrazol-3-yl)-5-((tetrahydrofuran-3-yl)amino)pyrazolo[1,5-a]pyrimidin-3-yl)-N-(methyl-d3)nicotinamide (30).** A mixture of (S)-5-(6-chloro-5-((tetrahydrofuran-3-yl)amino)pyrazolo[1,5-a]pyrimidin-3-yl)-N-(methyl-d3)nicotinamide (15 mg, 40 \u03bcmol), 1-isopropyl-3-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazole (28.3 mg, 0.12 mmol), XPhos Pd G2 (1.6 mg, 2.0 \u03bcmol), and cesium carbonate (39 mg, 0.1 mmol) in 1,4-dioxane (1 mL) and water (0.5 mL) was sparged with N2 and heated to 100 \u00b0C for 2 h. The reaction was diluted with MeOH and filtered through a thiol SiliaPrep cartridge. The product was purified by prep HPLC (pH 2) to afford the title compound (2 mg, 11%). LCMS calculated for C23H24D3N8O2 (M + H)+: m/z = 450.2; found: 450.2. 1H NMR (400 MHz, DMSO-d6) \u03b4 9.43 (d, J = 2.1 Hz, 1H), 9.38 (d, J = 5.8 Hz, 1H), 9.31 (s, 1H), 8.99 (t, J = 2.2 Hz, 1H), 8.76 (d, J = 2.0 Hz, 1H), 8.63 (s, 1H), 8.62 (s, 1H), 7.98 (d, J = 2.4 Hz, 1H), 7.08 (d, J = 2.5 Hz, 1H), 4.81\u22124.72 (m, 1H), 4.63 (p, J = 6.7 Hz, 1H), 4.14 (dd, J = 9.2, 5.5 Hz, 1H), 3.94 (q, J = 7.7 Hz, 1H), 3.88 (dt, J = 8.5, 4.2 Hz, 1H), 3.83 (dd, J = 9.3, 3.1 Hz, 1H), 2.49\u22122.43 (m, 1H), 2.05\u22121.95 (m, 1H), 1.50 (d, J = 6.6 Hz, 6H).\n\n## Associated Content\n\nThe Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jmedchem.2c01366: full synthetic procedures and characterization of compounds 5\u221230; copies of NMR spectra and HPLC traces for compounds 25, 27, 29, and 30; protocols for all biological assays, pharmacokinetic, and pharmacodynamic studies; and coordinates and experimental details for the X-ray co-crystal structure of 30 (PDF); docking models for compound 9 (Figure 2a) (PDB); theoretical THF derivative (Figure 2b) bound to FGFR3 (PDB); molecular formula strings and in vitro biological data (CSV).\n\n### Accession Codes\n\nPDB ID code for 30 bound to FGFR2 is 8E1X. Authors will release the atomic coordinates upon article publication.\n\n## Author Information\n\n**Corresponding Author:** Artem Shvartsbart \u2014 Prelude Therapeutics, Wilmington, Delaware 19803, United States; orcid.org/0000-0002-1438-3095; Email: ashvartsbart@preludetx.com.", "# GitHub README: PANTS\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/PANTS, 2026\n\n> **Historical record.** Everything in this section was true when it was written and has\n> since been superseded. The AUC 0.976 below was measured against negatives meaning \"not\n> reported active in a database\", which cannot distinguish an enzyme assayed and found dead\n> from one nobody assayed. When 150 **measured** negatives arrived, that number and the\n> conclusions drawn from it did not survive. The section is kept because the reasoning is\n> the record of how the project got to where it is \u2014 for the current position see\n> [`FINDINGS.md`](FINDINGS.md).\n\nTwo findings that changed the plan, both surfaced by the pre-training sanity gate rather than by review. Full detail in [`PHASE1_FINDINGS.md`](PHASE1_FINDINGS.md).\n\n**The nine curated positives are one cluster, not nine examples.** They collapse into a single cluster at both 30% and 50% identity, because the engineered variants are 97.6% to 99.3% identical to their parents. Under the project's own evaluation rule (split by cluster, never by sequence) that is one independent example, so no cluster-split evaluation is possible over them and any cross-validation is pure leakage. The first ungrouped trivial baseline returned AUC 0.9996 for exactly that reason. Harvesting the full ESTHER polyesterase family took the positive set to 87 sequences in 11 clusters at 30%, at the cost of those being annotation-only labels.\n\n**The hard negatives were separable on amino-acid composition alone.** Cluster-grouped, a classifier using nothing but 20 amino-acid fractions and length scored AUC 0.954 against a null of 0.495. The coefficients diagnosed it: negatives Leu-rich, positives Ser/Thr/Gly/Pro-rich, which is a **secreted-versus-cytoplasmic** signature rather than polyester chemistry. Every characterised polyesterase is secreted; the negative families were largely intracellular.\n\nNegatives are now matched on five axes: length distribution, identity to nearest positive, genus cap, **signal peptide**, and **phylum**. Phylum matching specifically contributed almost nothing (0.845 to 0.842), which is itself informative: the residual is not GC-driven.\n\n**Curating real activity data moved it further.** `EC 3.1.1.101` is poly(ethylene terephthalate) hydrolase, a curator's assignment of measured function rather than a family guess, and harvesting it took the positive set from 87 sequences in 11 clusters to 529 in 29:\n\n| Positive set | Clusters | Composition baseline | Verdict |\n|---|---|---|---|\n| Curated only | 1 | 0.9996 (leakage, not a measurement) | invalid |\n| Plus ESTHER family | 11 | 0.842 | MARGINAL |\n| Plus EC 3.1.1.101 | 29 | **0.778** | MARGINAL |\n\nStill short of the 0.75 pass mark, but the trend confirms the diagnosis: much of the apparent shortcut was a small-sample artefact that shrinks as real diversity arrives.\n\nConsequently the composition baseline is a **permanently reported metric** alongside the retrieval baseline, not merely a pre-training gate. Any claim the model makes has to clear both.\n\n### The labels were the blocker, and PAZy resolved it\n\nFor most of this project a head trained on the catalogue's own labels scored **AUC 1.000**,\nwhich was a failure rather than a success: 449 positives carried `EC 3.1.1.101` from\n`ECO:0000256`, meaning the label was assigned **by sequence similarity**. A sequence model\nreproducing that is close to tautological. Meanwhile the positives with real experimental\nevidence numbered 17 and spanned 5 clusters: too few for the cluster-split protocol to run\nat all, so the honest head could not be scored rather than scoring badly.\n\n[PAZy](https://api.pazy.eu/api) inverts the inclusion criterion. An enzyme is listed\nbecause activity was **measured** on a plastic and published, with the DOI attached.\n\n| | Measured positives | Clusters at 30% | Clusters at 50% |\n|---|---|---|---|\n| Before | 17 | 5 | 7 |\n| With PAZy | 333 | 51 | 73 |\n| Plus the confirmed engineered variants | **341** | **50** | **72** |\n\nTrained on the measured set alone (300 after length filtering, 45 clusters, 220 negatives\nand near misses):\n\n| Metric | Value |\n|---|---|\n| AUC, cluster-grouped | **0.976 \u00b1 0.021** over 4 valid folds |\n| Average precision | 0.987 |\n| Brier score | 0.052 |\n| Composition-only baseline | 0.829 |\n\nThat is a different kind of number from the 1.000. The labels are experimental rather than\nsimilarity-derived, the split is across 45 independent units rather than impossible, and\nthe head clears the composition baseline by a real margin instead of tying with it.\n\n**What it still does not show, tested rather than assumed.** Running the head against the\n125 near misses gave **AUC 1.000**, which is a warning rather than a triumph. Every near\nmiss belongs to a single ESTHER family (`Cutinase`), and one homogeneous family separates\nas a block: composition alone scores 0.910 on the same contrast. That is a **family-level**\nquestion wearing the clothes of a functional one.\n\n| Contrast | ESM-2 head | Composition only |\n|---|---|---|\n| vs distant \u03b1/\u03b2-hydrolase families | 0.979 | 0.887 |\n| vs near misses | **1.000** | 0.910 |\n| vs both | 0.976 | 0.829 |\n\nThe honest test needs PET-**inactive** members of the polyesterase family, measured and\npublished. Those barely exist, and the reason is structural: **databases record what\nworks.** PAZy lists enzymes because activity was found; nobody systematically publishes\n\"we assayed this polyesterase on PET and it did nothing\". The negative class for the\nquestion that matters is largely unwritten, which is publication bias rather than a\ncuration gap, and no further database mining will fix it.\n\n### \ud83c\udfaf The within-family test, run", "# Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\nAuthors: Zhuo, J.; Li, Y.-L.; Qian, D.-Q.; Burns, D.M.; Mei, S.; Rafalski, M.; Xu, M.; Cao, G.; Pan, Y.; Jia, Z.; Jalluri, R.; Epling, L.B.; Fenalti, G.; Deller, M.C.; Procak, J.; Covington, M.; He, X.; Collins, R.; Stubbs, M.; Burke, K.; Oliver, J.; Margulis, A.; Boer, J.; Wynn, R.; Scherle, P.; Diamond, S.; Newton, R.; Zhang, Y.; Metcalf, B.; Yao, W.\nSource: Journal of Medicinal Chemistry (Drug Annotation), 2026\n\n**Caco-2 Permeability Protocol.** The Caco-2 monolayers were incubated at 37 \u00b0C, with ambient humidity, and CO2 for the duration of the transport assay. To determine the apparent permeability of compounds in the absorptive direction (A-B), a 50 \u03bcM solution of compound in Hanks' balanced salt solution was added to the donor compartment (apical side), whereas Hanks' balanced salt solution with 4% BSA was added in the receiver compartment (basolateral side). The apical volume was 0.2 mL, and the basolateral volume was 0.6 mL. The incubation period was 120 min. At the end of the incubation period, 0.1 mL of sample was removed from the receiver side, and an equal volume of acetonitrile was added for protein precipitation. After centrifugation, the supernatants were collected for analysis. Nadolol and metoprolol were included as low- and high-permeability comparators, respectively. Transepithelial electrical resistance was measured to ensure the integrity of the cell monolayers. Caco-2 cell monolayers with transepithelial electrical resistance values \u2265 300 \u03a9\u00b7cm\u00b2 were used for transport experiments. The concentrations of the controls and compound were measured by liquid chromatography/mass spectrometry (LC/MS) and liquid chromatography/tandem mass spectrometry (LC/MS/MS), respectively.\n\n**Intrinsic Clearance Protocol.** Compound (1 \u03bcM) was incubated with human liver microsomes (1 mg/mL of protein), NADPH (2 mM), and 100 mM potassium phosphate buffer (pH 7.4) at 37 \u00b0C. Aliquots were taken at 0, 10, 20, and 30 min and denatured with methanol. After centrifugation to remove the denatured proteins, the resulting supernatants were analyzed by LC/MS.\n\n**Pharmacokinetic (PK) Studies.** All animal studies were conducted with approval of the Institutional Animal Care and Use Committee of Incyte (a fully AAALAC-accredited institution) in accordance with local, state, and federal guidelines and regulations. Studies were conducted under the supervision of a veterinarian and in compliance with guidelines and protocols established and approved by the DuPont Haskell Animal Welfare Committee (HAWC # INC023-GP). This study is a non-GLP exploratory study intended to support dose selection for GLP studies for applications for products regulated by governmental agencies.\n\nINCB054707 (Compound 22) doses were administered intravenously (IV; 2.5 or 2.0 mg/kg) or orally (4.0 or 5.0 mg/kg) via gavage to Sprague\u2212Dawley rats, beagle dogs, and cynomolgus monkeys. The IV formulation in rats was 5% dimethylacetamide (DMAC) and 25% or 50% sulfobutyl ether \u03b2-cyclodextrin in acidified saline (pH 3). The vehicle for IV dosing in beagle dogs and cynomolgus monkeys was 5% DMAC and 40% of sulfobutyl ether \u03b2-cyclodextrin (SBE-CD, 50% w/v solution) in saline. The oral formulation in all species was 5% DMAC in 0.5% methylcellulose aqueous solution. For rats, blood samples were collected at predose, 5 (IV dosing only), 15, and 30 min; and 1, 2, 4, 6, 8, and 24 h postdose in both oral (PO) and IV groups. For dogs, blood samples were collected at predose, 15 and 30 min; and 1, 2, 3, 4, 6, 8, 12, and 24 h postdose in PO group, and at predose, 5, 10, 15, and 30 min; and 1, 2, 3, 4, 6, 8, 12, and 24 h postdose in the IV group. For monkeys, blood samples were collected at predose, 15 and 30 min; and 1, 2, 3, 4, 6, 8, 12, 16, and 24 h postdose in PO group, and at predose, 5, 10, 15, and 30 min; and 1, 2, 3, 4, 6, 8, 12, 16, and 24 h postdose in the IV group. All blood samples were collected using EDTA as the anticoagulant and were processed to extract plasma by centrifugation and stored at \u221220 \u00b0C until analysis. The plasma was protein precipitated, and the diluted supernatant was analyzed using LC/MS/MS. A standard curve was prepared in plasma from 1 to 5000 nM and processed in the same manner as the samples. The PK parameters were determined using WinNonLin 6.3. The PK results for the other compounds were obtained similarly. In rats and monkeys, the cassette dosing of up to six test compounds was conducted to obtain preliminary PK parameters.\n\n**Rat Adjuvant-Induced Arthritis.** Adjuvant-induced arthritis was elicited in rats according to established methods. Lewis rats (150\u2212200 g, female; Charles River Laboratories, Wilmington, MA) are injected at the base of the tail with 100 \u03bcL of an emulsion of CFA (10 mg/mL *Mycobacterium butyricum* in incomplete Freund's adjuvant). Rats exhibited signs of inflammation within 2 weeks of the injection of CFA. Each rat paw was scored following visual observation using a rating of 0\u22123, (0 = normal; 1 = redness and minimal swelling of digits; 2 = moderate swelling of the digits and/or paw; 3 = severe swelling of digits and/or paw). Individual animal paw scores are combined and recorded as a sum of all four paws, and groups means of these totals are reported. Percent inhibition in clinical score/severity is calculated using the following formula: (Vehicle\u2212Treated)/Vehicle \u00d7 100.\n\nIn addition, a plethysmometer (Stoelting, Wood Dale, IL) was used to measure paw volumes taken at baseline and study termination. At the termination of the experiment, paws were removed from euthanized rats for histologic analyses. Treatment was initiated when significant signs of disease would be noted, and groups of animals were sorted so that mean scores would be equivalent\u2014usually occurring 2 weeks after adjuvant injection. Graphs reflect end points collected only immediately prior to and after therapy was initiated (treatment day 0). Groups consisted of six animals, and statistical differences between treatment and vehicle controls were assessed using two-tailed Student *t* tests or ANOVA with a Dunnett's test when appropriate.", "# Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\nAuthors: Zhuo, J.; Li, Y.-L.; Qian, D.-Q.; Burns, D.M.; Mei, S.; Rafalski, M.; Xu, M.; Cao, G.; Pan, Y.; Jia, Z.; Jalluri, R.; Epling, L.B.; Fenalti, G.; Deller, M.C.; Procak, J.; Covington, M.; He, X.; Collins, R.; Stubbs, M.; Burke, K.; Oliver, J.; Margulis, A.; Boer, J.; Wynn, R.; Scherle, P.; Diamond, S.; Newton, R.; Zhang, Y.; Metcalf, B.; Yao, W.\nSource: Journal of Medicinal Chemistry (Drug Annotation), 2026\n\nIn 2014, Li, Zhuo et al. disclosed a modified synthetic route that constructed the top portion of the molecule and then coupled it to the appropriate pyrazole hinge-binding motif, such as 42 (Scheme 2). Thus, the amide was prepared from the corresponding carboxylic acid 33 and amine 34, via formation of the acid chloride. Nucleophilic aromatic substitution of 35 was conducted with azetidin-3-ol 36 in the presence of DBU to afford alcohol 37, which was subsequently oxidized with (diacetoxyiodo)benzene and TEMPO to provide ketone 38. The Horner\u2013Wadsworth\u2013Emmons reaction was performed to afford the 3-(cyanomethylene)azetidine 39, which could undergo a Michael addition with the pyrazole boronate 40. Finally, the fully assembled top portion 41 could be coupled with the appropriate hinge-binding pyrazole 42 via Suzuki coupling to afford the final product 22.\n\n*Scheme 1.* Synthesis of 18 (INCB047986). Reagents and conditions: (a) (i) NaH, DMAC, 0 \u00b0C (ii) SEM-Cl, 89%; (b) 25, Pd(PPh3)4, K2CO3, 1-butanol/H2O, 90 \u00b0C; (c) 10% HCl (aq), THF, rt 82% for two steps (b and c); (d) 28, DBU, ACN, 98%; (e) 4.0 M HCl in 1,4-dioxane, DCM, 86%; (f) 31, (R)-BINAP, Pd(OAc)2, Cs2CO3, PhMe, 130 \u00b0C, 5h; (g) BF3\u00b7OEt, ACN, 0 \u00b0C to rt.\n\n*Scheme 2.* Synthesis of 22 (INCB054707). Reagents and conditions: (a) (i) oxalyl chloride, ACN, DMF (ii) 34, 0.5 M NaOH (aq), DCM, PhMe, 0 \u00b0C; (b) 36, DBU, ACN, 80 \u00b0C; (c) PhI(Oac)2, TEMPO, DCM; (d) (EtO)2P(O)CH2CN, *t*-BuOK, THF, 0 \u00b0C to rt; (e) 40, ACN, 50 \u00b0C; (f) 42, Pd(PPh3)4, Na2CO3, 1,4-dioxane/H2O, 110 \u00b0C.\n\n## Conclusions\n\nFollowing the discovery of the JAK1-selective inhibitor itacitinib, an SAR campaign was initiated with the goal of reducing the molecular weight and improving the ADME and PK profile. For the first objective, it was discovered that the molecular weight could be significantly reduced by replacement of the 1-(3-fluoro-2-(trifluoromethyl) isonicotinoyl) piperidinyl moiety with a simple *i*-propyl amide at the 4-position of a phenyl ring (compound 9). Compound 9 maintained the JAK1/2 isoform selectivity with enhanced potency and rat PK profile. However, the cynomolgus monkey and dog PK profiles were inferior to itacitinib 3. In an effort to address this issue, the introduction of IMHB interaction for the secondary amide was implemented. This strategy has been successfully adopted in several historical Incyte programs and led to the discovery of the FDA-approved c-Met inhibitor capmatinib (13) and the clinically relevant 11-\u03b2-HSD1 inhibitor INCB13739 (11) (Table 3). Installation of fluorines on one of the methyl groups of the *i*-propyl amide had minimal effect (14), whereas installation of a fluorine atom in the *ortho*-position relative to the secondary amide to afford 15 resulted in a significant improvement in the exposure and oral bioavailability in rats with a slight improvement in the Caco-2 value and monkey PK profile. Surprisingly, installation of both IMHB moieties (i.e., *ortho*-F and CF3 amide capping groups) to afford compound 16 led to a drastic improvement in both the rat and cyno PK profiles vs both the *ortho*-fluoro-amide 15 and the CF3 amide 14. Potency was further improved by installation of a 5-fluoro group on the phenyl ring relative to the amide moiety (18). The PK profile in all three species (rat, monkey, and dog) was substantially improved by replacement of the 7H-pyrrolo[2,3-*d*]pyrimidine hinge-binding motif with a 3,5-dimethylpyrazole hinge-binding group (22).\n\nThe JAK1 isoform selectivity was attributed to the terminal amide capping *i*-propyl or trifluoro-*i*-propyl group of the secondary amide occupying a vacant space that is only present in the JAK1 isoform due to two key amino acid residue differences between JAK1 and -2 in this area that shift the P-loop away from the binding pocket by 2\u20133.5 \u00c5 for JAK1 to create a small lipophilic pocket.\n\nPharmacodynamic studies indicated that the selective JAK1 inhibitor povorcitinib 22 exerted significant improvements in multiple measured parameters of disease in an adjuvant-induced arthritis rat model that mimics human rheumatoid arthritis. Cytokine- and growth-factor-stimulated cells indicated that these effects were due to selective JAK1 inhibition that was devoid of meaningful JAK2 inhibition. Additional preclinical studies were conducted and showed highly encouraging efficacy and led to the advancement of povorcitinib 22 into several clinical trials. The results from these preclinical and clinical studies will be the topics of upcoming publications. Currently, povorcitinib 22 is in phase 3 trials for the treatment of hidradenitis suppurativa, vitiligo, and prurigo nodularis and in phase 2 for the treatment of asthma and chronic spontaneous urticaria.\n\n## Experimental Section\n\n### General Experimental Information", "# The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\nAuthors: Span, E.A.; Suess, D.L.M.; Deller, M.C.; Britt, R.D.; Marletta, M.A.\nSource: ACS Chemical Biology, 2017\n\nIn structures of exclusively C1-oxidizing PMOs, a threonine or tyrosine side chain accepts an H-bond from the N-terminal amino group; in those of C4-oxidizing PMOs, an ordered water molecule serves the same role. The T74 residue in MtPMO3* forms H-bonds with the N-terminal amino group of H1. This interaction enhances the nucleophilicity of the N-terminal amino ligand and could help stabilize high-valent oxyl intermediates. Additionally, T74 may form H-bonds with ligands occupying the axial Cu coordination site trans to the tyrosine. In the structure of a PMO2 bound to a cellodextrin substrate, the ordered \"pocket\" water H-bonds with a C6-OH atom of the substrate, in addition to accepting a H-bond from the N-terminal amino group.[26] This \"axial\" H-bonding network is absent in PMO3 subtype (C1/C4-oxidizing) enzymes, where a proline residue excludes solvent from the pocket.\n\n*Figure 2.* (a) Structural superposition of MtPMO3* (yellow) with three other AA9 structures, PcPMO1 (orange, PDB ID 4B5Q), NcPMO2 (indigo, 4EIR), and NcPMO3 (teal, 4EIS). Conserved residues H161 and Q167 (MtPMO3* numbering) form H-bonds with D26 from a symmetry-related molecule in MtPMO3* (left), a water molecule in NcPMO3 (top right), and a buffer-derived glycerol molecule in PcPMO1 (bottom right). Distances are in \u00c5ngstroms (\u00c5). The active-site Cu appears reduced by the X-ray beam in all structures; exogenous molecules depicted in the solvent-facing equatorial position are at lengths (3.6\u20134.6 \u00c5) from the Cu center that exclude formal coordination. (b) Multiple sequence alignment of representative AA9, AA10, AA11, and AA13 PMOs showing conservation of H-bonding motifs (marked with diamonds) within the PMO superfamily. Fungal PMOs possess a H-X5\u20138-Q-X-Y (AA9), N-X2-E-X-Y (AA11), or Q-X-Q-X-Y (AA13) motif depending on substrate specificity (cellulose, chitin, and starch, respectively). Bacterial AA10 PMOs active on cellulose contain either a R-X4-E-X-F or H-X2-Q-X-Y motif, correlating with observed regioselectivity (C1 or C1/C4, respectively). Primary coordinating residues are marked with a star. Numbers correspond to MtPMO3* amino acid numbering.\n\n### Cellulose Activity and O2 Utilization of H-bonding Variants\n\nThe activity of MtPMO3* variants at H161, Q167, and T74 was examined using phosphoric acid swollen cellulose (PASC) as a substrate. Assays utilized reduced cellobiose dehydrogenase from M. thermophila (MtCDH-2) to deliver reducing equivalents from its flavin cofactor. Soluble products quantified at 1 min end points show that activity is reduced (~10\u201350% relative to wild type), but not abolished by any single substitution (Figure 3). Given the conservation of H161 and Q167, it was expected that the alanine substitutions would have a more dramatic reduction on activity, perhaps eliminating it completely. One explanation for the retention of some activity is that solvent may enter the cavity created by the difference in side chain volume and \"rescue\" activity to some extent via H-bonds from the solvent. This would be expected to occur in PMOs because the active site is near the protein-solvent interface. The PASC assay gives limited information because the insoluble nature of the substrate impedes quantitative kinetic analysis. Assays were also performed with the addition of exogenous Cu(II) and showed no increase in activity (Supporting Information Figure 1), demonstrating that all wild type and variant enzymes are not functionally deficient in copper, as purified.\n\n*Figure 3.* (a) Effect of H-bonding network substitutions on MtPMO3* PASC activity. Assays contained 2 \u03bcM PMO, 10 mg mL\u22121 PASC, atmospheric O2, and 1 \u03bcM MtCDH-2 as the reducing agent. Reactions of 45 \u03bcL were carried out in 50 mM sodium acetate buffer (pH 5.0) at 40 \u00b0C for 1 min. Peaks from aldonic acids with DP 5\u201313 were quantified via HPAEC with electrochemical detection (nC), and peak areas were integrated over time (nC\u00b7min). Smaller C1-oxidized products with DP 2\u20134 were excluded from this analysis, as they are also products of the CDH reaction (n = 3). (b) Sample HPAEC trace with aldonic acid products labeled by DP, from A2-A13.\n\nInitial rates of O2 consumption were measured for all variants and ranged from ~50\u2013100% of the wild type enzyme, although the values may be inflated by the background rate of O2 reduction in assays utilizing CDH, which may reduce O2 independently via its iron-heme cofactor (Figure 4). Variants H161E, Q167A, Q167E, and T74A are the slowest to turn over O2 at ~50% the rate of wild type. The other H161 variants are less affected: H161A turns over O2 at ~70% the rate of wild type, and H161Q can turn over O2 at rates comparable to the wild type. A number of variants show defects in O2 consumption that are smaller than their defects in product formation. This observation is complicated by the high background in the O2 consumption assays, as well as the difficulty in extracting rate information from the PASC assays; however, it suggests O2 utilization is decoupled from substrate hydroxylation. The clearest example of this is the H161Q variant, followed by H161A.\n\n*Figure 4.* Effect of H-bonding network substitutions on O2 consumption rate by MtPMO3*. Assays contained 0.5 \u03bcM PMO, 10 mg mL\u22121 PASC, and atmospheric O2 and were carried out in 50 mM sodium acetate buffer (pH 5.0) at 40 \u00b0C. Reactions of 300 \u03bcL were initiated with 5 \u03bcM MtCDH-2. (a\u2013c) Assay traces showing O2 consumption by variants of H161, Q167, and T74, respectively, in comparison with the wild type (WT) and background (no PMO). (d) Rates of O2 consumption are derived from initial velocity regions after CDH addition (n = 3).", "# deepcovalent: Partnerships \u2014 DeepCovalent\nAuthors: deepcovalent\nSource: https://www.deepcovalent.com, 2026\n\nPartner with DeepCovalent to turn target hypotheses into ranked, synthesis\u2011ready, and fully *in-silico* validated small molecules in \u2264 **3 weeks** using the **OS3D\u2122** \u2014 built to improve downstream success while reducing time and cost.\n\n## Accelerate Drug Discovery Together\n\n* **Speed:** Ideation to fully *in-silico* validated molecules \u2264 3 weeks.\n* **Quality:** Improved downstream success with multi\u2011objective optimization embeds potency, selectivity, ADMET, and developability constraints early.\n* **Novelty:** Unbiased accessing unexplored chemical space beyond incremental analog design.\n* **Clarity:** Transparent, partner\u2011friendly IP pathways and predictable collaboration structures.\n* **Flexibility**: In addition to the two core partnership models below, we remain open to customized collaboration structures that support unique scientific and strategic goals.\n\n#### **Why Partner with DeepCovalent**\n\n## Partnership Models\n\n#### **On\u2011Demand Generation & Early Assessment**\n\nRapid, low-cost access to OS3D\u2122 to evaluate new or deprioritized targets.\nPartners receive synthesis-ready and fully *in-silico* validated molecules in **3 weeks** \u2014 enabling fast go/no-go decisions and early pipeline optionality with minimal risk.\n\nA flexible, collaborative model with shared ownership of molecular IP. We focus on AI-driven molecular design and early discovery, while retaining the flexibility to extend into compound synthesis, preclinical validation, and IND-enabling studies under co-development agreements.\n\n#### **Flexible Co-Development Model**\n\n[Discuss Codevelopment](/contact)", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\nSimilarly, mutagenesis of human LIF has revealed a third site consisting of residues Lys 159 and Phe 156, located at the N-terminus of helix D, which provides the majority of free energy for binding to LIFR (Hudson et al., 1996) (see Figure 1-6). Mutation of these residues individually reduces binding by ~100-fold and biological activity by ~1000-fold. Double mutations virtually eliminate receptor binding and biological activity. These residues are solvent exposed in close proximity to each other and are surrounded by a group of several residues that also affect binding. This arrangement of several pivotal residues shielded by ones of less importance has also been observed in the hGH-hGHR complex (Clackson and Wells, 1995). Other cytokines with sequence similarity to human LIF (murine LIF, OSM, CNTF and CT-1), that utilise the LIFR in their signaling complexes, all have the human LIF residues Phe 156 and Lys 159 conserved (Robinson et al., 1994). In contrast, IL-6 and IL-11, which do not bind the LIFR, lack these residues. Therefore, Phe 156 and Lys 159 (or their equivalents) may function as a LIFR binding epitope. This epitope is referred to as site 3 and forms high affinity contacts with LIFR whilst site 1 functions as a secondary component of LIFR binding. Regions of LIFR responsible for the site 3 interaction have recently been localised to the so-called LIGHT motif of the immunoglobulin-like domain (Chobotova and Heath, 2000).\n\nResidues of human LIF involved in binding gp130 have also been identified through the use of multiple simultaneous substitution mutagenesis (Hudson et al., 1996). This has revealed several residues in helix-A (Gln 25, Ser 28 and Gln 23) and helix-C (Ile 121, Gly 124 and Ser 127) that participate in gp130 binding (see Figure 1-6). None of these residues are entirely conserved through the gp130-binding family of cytokines, so suggesting that a variable subset of these residues is involved in binding gp130. This motif is therefore thought to function in regulating the affinity of gp130 binding. This region is analogous to site 2 in the hGH-hGHR model and non-conservative mutations in this region result in loss of biological activity (see Figure 1-6).\n\n*Figure 1-6.* Figure showing the structural and functional motifs observed in the haemopoietin cytokine family. The C-terminal (light-blue) and N-terminal (dark-blue) fibronectin type III-like domains of (a) hGHR (Somers et al., 1994) and (c) gp130 (Bravo et al., 1998) are shown. The Linker region between domains is shown in orange along with key residues of the receptor in red. The ligands (b) hPrl and (d) LIF (Robinson et al., 1994) are shown in green. Site 1 (blue), site 2 (yellow) and site 3 (red) residues for each ligand are shown. Figures produced using BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994).\n\n#### 1.5.1.3 Leukemia Inhibitory Factor Receptor\n\nLIFR is 1097 amino acids long with a 44 amino acid leader sequence. The single transmembrane domain is 26 amino acids in size and consists of a stretch of hydrophobic residues. The extracellular domain is 789 residues long and consists of two haemopoietin domains separated by a putative immunoglobulin-like domain. This ligand-binding region is separated from the cell surface by three fibronectin type III domains. The cytoplasmic domain is 238 residues in size and is rich in serines, threonines, prolines and acidic residues. The predicted molecular weight of LIFR is 111 kDa, although the observed molecular weight on SDS-PAGE is 200 kDa due to extensive glycosylation at several of the 20 potential N-glycosylation sites. Murine and human LIFR show 80% sequence identity with both showing extensive homology to the entire sequence of gp130.\n\n### 1.5.2 Oncostatin M and the gp130 Receptor\n\n#### 1.5.2.1 Biological Action of Oncostatin M\n\nOSM and LIF function in a highly pleiotropic and yet redundant fashion. This is reflected through their syntenic linkage on chromosome 22 (Giovannini et al., 1993; Rose et al., 1993), and the shared use of gp130 (Taga and Kishimoto, 1997). Among the shared biological activities of OSM and LIF are the ability to induce differentiation of murine myeloid leukemic cells into macrophage-like cells (Bruce et al., 1992) and the inhibition of differentiation of embryonic stem cells (Rose et al., 1994). Both LIF and OSM also play a role in the inflammatory response, for example induction of acute phase protein synthesis in hepatocytes (Richards et al., 1992) and cartilage destruction (Hui et al., 1996).\n\nHowever, several important biological activities specific to OSM have also been identified including growth inhibition of a variety of solid tissue tumours (Horn et al., 1990) and human melanoma cells (Brown et al., 1987; Malik et al., 1989; Zarling et al., 1986). OSM also displays mitogenic activity in rabbit aorta smooth muscle cells (Grove et al., 1993). OSM, and not LIF or IL-6, has also been shown to induce spindle cell morphology in acquired immunodeficiency syndrome (AIDS)-associated Kaposi's sarcoma (AIDS-KS cells) (Cai et al., 1994; Miles et al., 1992; Nair et al., 1992). The lack of biological response of AIDS-KS cells to IL-6 and LIF can be explained by the absence of IL-6R and LIFR (Murakami Mori et al., 1995).\n\n#### 1.5.2.2 Structure-Function Relationship of Oncostatin M and gp130\n\nOSM shares approximately 25% sequence identity with LIF, so allowing structural comparison with the crystal structures of both human (Robinson et al., 1994) and murine LIF (Robinson et al., 1995) and more distantly to the hGH-hGHR system (see section 1.5.1.2 and 1.4.2.1.1).", "# Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly\nAuthors: Deller, M.C.; Johnson, H.A.; Miller, M.D.; Spraggon, G.; Elsliger, M.-A.; Wilson, I.A.; Lesley, S.A.\nSource: PLoS ONE, 2015\n\nIn addition, the quaternary structure, domain arrangement, overall shape and size is similar to that of other RCK gating ring assemblies, such as TrkA (PDB ID: 4J9U) [21], MthK (PDB ID: 2FY8) [16,18], KtrA (PDB ID: 2HMS and 4J7C) [20,23], GsuK (PDB ID: 4GX5) [25] and BK (3U6N and 3NAF) [22,24]. TM1088 homodimers are assembled in a pseudo-tetrameric arrangement to generate an octameric assembly with approximate dimensions of 127 \u00c5 x 101 \u00c5 x 53 \u00c5 (Fig 4D). Moreover, similar homodimeric arrangements are observed in the non-complexed high-resolution crystal structures of TM1088A (PDB ID: 2G1U) and TM1088B (PDB ID: 3JXO) further supporting the correct assembly of these dimers into an octameric arrangement. The N-terminal Trk domains of TM1088B (N1) and TM1088A (N2) form the two sides of the ring, while the C-terminal Trk domains of TM1088B (C1) are peripheral and lie outside the central ring (Fig 4D).\n\nThe TM1088 assembly forms a large central channel with a total solvent accessible surface of ~9700 \u00c5\u00b2 (calculated by CASTp and MOLE [50,51], channel shown as aqua surface in Fig 4D). The channel is ~50 \u00c5 in length and is formed by a roughly equal contribution of residues from all 8 chains and is readily accessible to solvent on both ends. The four AMP molecules reside on the inside of the channel and are solvent accessible. The AMP molecules are bound to the TM1088A N-terminal TrkA domain (N2) at positions flanking the dimer interface. The AMP molecules line the central channel in a similar position to nucleotides observed in both KtrA and TrkA gating ring assemblies (AMP shown as yellow spheres in Fig 4D). The central channel of the TM1088 complex is ~15 \u00c5 at its narrowest, which is comparable to the ~15 \u00c5 diameter at the center of the MthK channel (PDB ID: 2FY8) [18].\n\n*Figure 4.* TM1088 proteins are structurally homologous to each other and to other RCK domains. (A) Superposition of the N-terminal Trk domain of TM1088A (N2, green) onto the N-terminal Trk domain of TM1088B (N1, blue). (B) Superposition of the N-terminal Trk domain of TM1088A (N2, green) onto the N-terminal TrkA domain of V. parahaemolyticus (red, PDB ID: 4J9U). (C) Superposition of the N-terminal Trk domain of TM1088B (N1, blue) onto the N-terminal TrkA domain of V. parahaemolyticus (red, PDB ID: 4J9U). (D) Solvent accessible surface of the octomeric TM1088 assembly with the surface stripped away from the N1, N2 and C1 domains to show the underlying C\u03b1 trace used in the alignments in panels A-C. AMP is shown as yellow spheres. The solvent accessible channel traversing the assembly is shown in aqua. Residues involved in hydrogen bond formation across the central channel are highlighted by a red surface (Arg33, Glu34 and Glu37). The three types of protein interface in the assembly are highlighted by red arrows.\n\n### TM1088 assembly has an RCK domain arrangement homologous to other gating ring assemblies\n\nAs discussed above, the best structural alignments for both TM1088A and TM1088B are achieved with the N-terminal domain of TrkA from V. parahaemolyticus (Fig 4B\u20134C). Differences in the overall quaternary structure of TM1088 and TrkA are a result of the atypical two-subunit nature of the TM1088 assembly and the use of a domain-swapped organization. We have made an attempt to place the two-subunit TM1088 assembly into a more standard framework in line with previously studied single-subunit K+ transporter gating ring assemblies as detailed in Fig 5. It is important to note that all structures of RCK domains determined to date involve similar domain swapping; in the case of TM1088A secondary structure elements are swapped between monomers (i.e. helix 6), whereas in TM1088B, entire tertiary structure elements are swapped between monomers (i.e. N-terminal and C-terminal domains).\n\nWe propose that a simple gene duplication and fusion event (Fig 5A\u20135B), followed by different patterns of domain organization (Fig 5C), may be sufficient to account for the differences between the two-subunit TM1088 assembly and other single-subunit K+ gating ring assemblies (Fig 5D\u20135F). Duplication and fusion of the two-chains of TM1088B into a consolidated polypeptide chain consisting of two-tandem RCK repeats would be one possible route (Fig 5B, RCK1 and RCK2 shown, each with an N-terminal and C-terminal domain). Obviously, other combinations of duplication and fusions utilizing both TM1088A and TM1088B could also be envisaged.\n\n*Figure 5.* TM1088 assembly has an RCK domain arrangement homologous to other gating ring assemblies. (A) Schematic arrangement of the N-terminal (N1 and N2) and C-terminal (C1) RCK domains of TM1088A (green) and TM1088B (blue). (B) Possible gene duplication and fusion event resulting in a more classical single-chain arrangement as observed in other K+ gating ring assemblies. (C) Two-pathways of protein assembly resulting in a domain-swapped RCK or a non-swapped RCK. (D) Top: Schematic of the domain organization of the TM1088 assembly, showing the two-fold symmetry elements (black arrows) and the unique protein interfaces (orange curves). Middle: The repeating TM1088B RCK unit highlighting the domain-swapped organization. Bottom: molecular surface of the TM1088 assembly showing the corresponding domains. (E) Top: Schematic of the domain organization of the TrkA assembly (PDB ID: 4J9U), showing the two-fold symmetry elements (black arrows) and the unique protein interfaces (orange curves). Middle: The repeating TrkA RCK unit highlighting the domain-swapped organization. Bottom: molecular surface of the TrkA assembly showing the corresponding domains. (F) Top: Schematic of the domain organization of the BK assembly (PDB ID: 3U6N), showing the four-fold symmetry element (black square) and the unique protein interfaces (orange curves). Middle: The repeating BK RCK unit highlighting the non-swapped domain organization. Bottom: molecular surface of the BK assembly showing the corresponding domains. Panels E-F adapted from [21].", "# The Self-Driving Lab: How Automation and AI Are Redefining the Pace of Discovery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* [The Big Benefits of Going Small: Novartis Mic-Drop and MicroCycle](https://www.live.novartis.com/en/article/the-big-benefits-of-going-small)\n\n* **Closed-loop autonomy:** SDLs integrate design, synthesis, assay, and analysis into a single AI-orchestrated cycle.\n* **Validated in pharma:** Novartis\u2019s MicroCycle and Carnegie Mellon\u2019s Coscientist run real lab hardware via active-learning agents.\n* **Compressed timelines:** What once took weeks per design-make-test-analyse iteration can now finish in hours of autonomous operation.\n* **Scientists redirected, not replaced:** Human expertise shifts to question framing, judgement, and interpretation while routine work runs unattended.\n\n1. **The Rise of Self-Driving Labs in Chemical and Materials Sciences:** [Abolhasani & Kumacheva, Nature Synthesis, 2023](https://www.nature.com/articles/s44160-022-00231-0)\n2. **Self-Driving Laboratories for Chemistry and Materials Science:** [Tom et al., Chemical Reviews, 2024](https://pubs.acs.org/doi/10.1021/acs.chemrev.4c00055)\n3. **MicroCycle: An Integrated and Automated Platform to Accelerate Drug Discovery:** [Brocklehurst et al., Journal of Medicinal Chemistry, 2024](https://pubs.acs.org/doi/10.1021/acs.jmedchem.3c02029)\n4. **Autonomous Chemical Research with Large Language Models:** [Boiko et al., Nature, 2023](https://www.nature.com/articles/s41586-023-06792-0)\n5. **Autonomous \u2018Self-Driving\u2019 Laboratories: A Review of Technology and Policy Implications:** [Tobias & Wahab, Royal Society Open Science, 2025](https://royalsocietypublishing.org/doi/10.1098/rsos.250646)\n\n**Closed-loop autonomy is collapsing the design-make-test-analyze\u00a0cycle from weeks to hours, and pharma is taking notice**\n\nDrug discovery has long operated on a familiar rhythm: design a molecule, synthesize it, test it, interpret the data, design again. Each turn of that cycle traditionally takes weeks, and a typical hit-to-lead campaign can chew through months before yielding a candidate worth advancing. The bottleneck is rarely a shortage of ideas; it is the throughput and parallelism of the wet lab itself, where pipettes, plates, and people impose hard physical limits on how fast chemical and biological hypotheses can be tested. [The rise of self-driving laboratories (SDLs)](https://www.nature.com/articles/s44160-022-00231-0) has begun to dissolve those limits by collapsing the entire loop into a single automated, machine-learning-orchestrated engine.\n\nWhat separates an SDL from a conventional automation suite is closed-loop autonomy. Rather than a human queuing the next experiment, an active-learning algorithm proposes the most informative compound to make next; robotic synthesis and purification execute it at the microscale; in-line analytics quantify and characterize the product; and the resulting data flow back to retrain the model, typically within hours rather than weeks. [A comprehensive 2024](https://pubs.acs.org/doi/10.1021/acs.chemrev.4c00055)review catalogs\u00a0how this paradigm has matured across chemistry and materials science, with parallel progress now visible in pharma.\n\n> **\u201cSelf-driving laboratories (SDLs) promise an accelerated application of the scientific method.\u201d**\n\nConcrete examples are no longer theoretical. Novartis [reported the MicroCycle platform](https://pubs.acs.org/doi/10.1021/acs.jmedchem.3c02029) in 2024, integrating plate-based microscale chemistry, automated purification, in-situ quantification, and biochemical assays under active-learning control to drive multiparameter compound optimization. A year earlier, Carnegie Mellon\u2019s [Coscientist](https://www.nature.com/articles/s41586-023-06792-0) demonstrated that GPT-4-class language models can autonomously design, plan, and execute palladium-catalyzed cross-couplings on real laboratory hardware, offering a glimpse of what agentic AI looks like when it gets hands on the bench.\n\nThe implication for structural biology and SBDD is significant. When co-crystal structures, biophysical characterization, and DMTA cycles can all be choreographed by a single orchestrating layer, a small team can credibly run dozens of programs in parallel, an efficiency gain familiar to anyone who has built an integrated platform from scratch. [A 2025 review of the technology and its policy implications](https://royalsocietypublishing.org/doi/10.1098/rsos.250646) argues that SDLs now automate nearly the entire scientific method, from hypothesis to inference. Scientists remain indispensable for the questions worth asking and the judgment calls only experience supplies, but the routine has finally found a competent autopilot.", "# mdeller.com: Marc Deller's web app server (landing page)\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n**Welcome to mdeller.com**\n\n Healthy\n\nmarc@mdeller:~$ whoami\n\n# Marc C. Deller, D.Phil.\n\n\u2014 Structural biologist working in drug discovery. This is where the tools I build end up: small web apps for people who work with protein structures. Free, no sign-up, nothing to install.\n\nmain site \u2192 [marcdeller.com](https://marcdeller.com/) \"Hidden Architecture: Mapping Biology\u2019s Structure\"\n\nmarc@mdeller:~$\n\n## ls -la apps/\n\n\u00a0# 12 live, 0 building\n\n* \u25cfGOBSMACKED # fold, dock, relax and annotate a protein-ligand complex on your own GPU, then find the crystal structure and get graded on how close you got, and why[\u2197 github.com/bellcheddar/GOBSMACKED](https://github.com/bellcheddar/GOBSMACKED)up\u2014\n* \u25cfALPHABETTI # a protein sequence logo in 3D, wrapped around its own predicted structure. The amino acid letters themselves, stacked by how strongly a language model expects them, with no multiple sequence alignment anywhere[\u2197 github.com/bellcheddar/ALPHABETTI](https://github.com/bellcheddar/ALPHABETTI)up\u2014\n* \u25cfButtFold # a protein folds in your browser and the trajectory becomes music: contacts forming are note onsets, hydropathy sets pitch, secondary structure sets texture. The web version of PhoneFold[\u2197 github.com/bellcheddar/ButtFold](https://github.com/bellcheddar/ButtFold)up\u2014\n* \u25cfCODSWALLOP # every PDB entry for a protein family in one place: what was actually made, what it bound, what crystallised, and which residues nobody has ever seen[\u2197 github.com/bellcheddar/CODSWALLOP](https://github.com/bellcheddar/CODSWALLOP)up\u2014\n* \u25cfTopPDBLX # every crystallisation condition in the protein data bank, parsed and indexed: what worked, how often, and how little of it is specific to your protein[\u2197 github.com/bellcheddar/TopPDBLX](https://github.com/bellcheddar/TopPDBLX)up\u2014\n* \u25cfFlexAppeal # Set up protein molecular dynamics without writing the script: every OpenMM option explained, a self-contained bundle to run on your own machine, and interactive trajectory analysis[\u2197 github.com/bellcheddar/FlexAppeal](https://github.com/bellcheddar/FlexAppeal)up\u2014\n* \u25cfAlphaFraud # Catch AlphaFold where the fold is wrong: weekly, automatically, on freshly deposited human structures[\u2197 github.com/bellcheddar/AlphaFraud](https://github.com/bellcheddar/AlphaFraud)up\u2014\n* \u25cfBoltzMaker # Build and analyze Boltz-2 protein structure prediction campaigns: a guided input wizard, YAML generation, and full result analysis with confidence/affinity dashboards -- no GPU required[\u2197 github.com/bellcheddar/BoltzMaker](https://github.com/bellcheddar/BoltzMaker)up\u2014\n* \u25cfChatPDB # A protein-structure-literate AI assistant for analyzing proteins: RAG for facts, a QLoRA-tuned model for behaviour, and live Biopython/gemmi/DSSP/PyMOL tool calls for structural truth.[\u2197 github.com/bellcheddar/chatPDB](https://github.com/bellcheddar/chatPDB)up\u2014\n* \u25cfChemSage # A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior, and live RDKit/PyMOL tool calls for chemical truth[\u2197 github.com/bellcheddar/ChemSage](https://github.com/bellcheddar/ChemSage)up\u2014\n* \u25cfMicroPlastics Blaster # A browser-based arcade game by Elora Therapeutics \u2014 the biotech developing first-in-class enzyme therapy to remove microplastics from the human body. Fire enzymatic blasts at incoming polymer particles, protect the cell, and climb the global leaderboard[\u2197 github.com/bellcheddar/microplastic-blaster](https://github.com/bellcheddar/microplastic-blaster)up\u2014\n* \u25cfPANTS # PETase ANnotation and Triage System: mines metagenomic sequence space for PET-degrading enzymes that work at 37 \u00b0C in serum rather than at 70 \u00b0C in an industrial reactor, with predicted structures superposed on IsPETase and the catalytic triad highlighted[\u2197 github.com/bellcheddar/PANTS](https://github.com/bellcheddar/PANTS)up\u2014\n\nmarc@mdeller:~$ top -bn1 | head\n\nuptime\n:   \u2026\n\nload\n:   \u2026\n\ncpu\n:   \u2026 cores\n\nmemory\n:   \u2026 \u2026\n\ndisk\n:   \u2026 \u2026\n\nvisits\n:   \u2026\n\ntls\n:   \u2026\n\nmarc@mdeller:~$ cat links.txt\n\n[blog](https://marcdeller.com/blog/) [structural-biology](https://marcdeller.com/sb/) [publications](https://marcdeller.com/publications/) [structures](https://marcdeller.com/structures/) [projects](https://marcdeller.com/projects/) [patents](https://marcdeller.com/patents/) [vibe-coding](https://marcdeller.com/vibe-coding/) [about-me](https://marcdeller.com/about/)", "# Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor\nAuthors: Kong, L.; Ju, B.; Chen, Yajing; He, L.; Ren, L.; Liu, J.; Hong, K.; Su, B.; Wang, Z.; Ozorowski, G.; Ji, X.; Hua, Y.; Chen, Yanli; Deller, M.C.; Hao, Y.; Feng, Y.; Garces, F.; Wilson, R.; Dai, K.; O'Dell, S.; McKee, K.; Mascola, J.R.; Ward, A.B.; Wyatt, R.T.; Li, Y.; Wilson, I.A.; Zhu, J.; Shao, Y.\nSource: Immunity, 2016\n\nExtensive efforts have been made to study the VRC01 class of bNAbs. Co-crystal structures with the gp120 core elucidate a common interaction strategy employed by all VRC01-class bNAbs (Scheid et al., 2011; Wu et al., 2011; Zhou et al., 2010, 2013, 2015), and next-generation sequencing (NGS) analysis reveals a converged pattern of maturation toward a set of \"VRC01 signatures\" (Zhou et al., 2013). These signatures include the preferential use of the IgHV1-2*02 germline allele, a conserved maturation pattern of heavy chains, a stringent 5-aa light-chain complementarity determining region (CDR) 3, and specific mutations in the light-chain CDR1 (Wu et al., 2011; Zhou et al., 2013). Based on this information, putative VRC01 precursors and intermediates are inferred to guide CD4bs immunogen design (Jardine et al., 2013, 2015). However, important questions remain unanswered: how much do these VRC01 signatures contribute to the early lineage development? What is the order of molecular events in B cell maturation leading to the VRC01-like neutralizing activity? Do these steps involve overcoming critical barriers that need to be accounted for during immunogen design? The complex patterns of VRC01 lineage development over a 15-year period have been reported (Wu et al., 2015), but bNAbs identified from the earliest time point already possess the mature VRC01 signatures and broad neutralizing activity, indicating that this lineage reached a \"plateau\" prior to the timeframe studied. Thus, details of early-stage VRC01 development remain unknown.\n\nHere we addressed these questions by analyzing samples from an HIV-1-infected Chinese patient (Hu et al., 2012). This donor, DRVI01, was a long-term non-progressor (LTNP) with increasing serum reactivity to the CD4bs over a period of 5 years. We isolated a monoclonal antibody (mAb), DRVIA7, from the 5th year sample, which exhibited a subset of VRC01 signatures but with limited neutralizing activity. Structures revealed that DRVIA7 was a VRC01-like antibody but with critical differences in the light-chain CDR1 and N terminus, both of which interact with gp120 glycans. Functional analysis confirmed that a gp120 glycan patch was a barrier to achieving neutralization breadth. We traced the DRVIA7 antibody lineage temporally within the unbiased B cell repertoires and utilized single-genome amplification (SGA) to investigate the viral co-evolution, which confirmed the inhibitory role of the gp120 glycan patch identified. In summary, our study identified a VRC01-class antibody lineage from a Chinese patient, demonstrated that the glycan shield stalled the development of VRC01-like broadly neutralizing activity, and provided functional VRC01 precursors and early intermediates to facilitate B cell lineage-based design of CD4bs immunogens.\n\n## Results\n\n### Identification and Characterization of an HIV-1-Infected Chinese Donor\n\nThe analysis of HIV-1-infected individuals displaying prolonged control of viremia, or LTNPs, presents opportunities to learn how such control is achieved by the immune system (Tomaras et al., 2011; Walker et al., 2010). In this study, we analyzed the HIV-1-neutralizing antibody response in a Chinese LTNP. Historically, two sources have contributed to the HIV epidemic of the last century in China: an early wave of injecting drug users (IDUs) in the southwest Yunnan province that borders on the heroin-producing \"Golden Triangle\" region in the mid to late 1980s (Ma et al., 1990) and a later outbreak in the former plasma donors (FPDs) through contamination of blood plasma transfusion by the IDU's HIV-1 strain in mid 1990s (Li et al., 2012). After 10 years of infection without antiviral treatment, some FPD survivors displayed relatively normal CD4+ T cell counts and a certain level of viral control consistent with that of LTNPs. The patient studied here, DRVI01, was selected from the top 5% neutralizers based on serum screening against a panel of more than 20 viruses (Hu et al., 2012). In brief, DRVI01 entered the study at the age of 40 with informed consent and was procedurally followed up every 6 months from 2005 to 2010. This donor carried a clade-B' strain of HIV-1, with a viral load ranging from 74,200 to 310,000 copies/mL and a CD4+ T cell count ranging from 335 to 769 cells/mL during the period studied. Serum neutralization was tested against a global panel containing 12 viruses from diverse clades (Global panel) (deCamp et al., 2014) and a separate panel containing 25 viruses established at the Division of Research on Virology and Immunology (DRVI panel) of the China CDC. Over 95% breadth was observed at all the time points studied (Figure 1A; Table S1). The average potency increased from 270 and 246 geometric mean titers (GMTs) in 2005 to 736 and 706 GMTs in 2009 for the two virus panels, respectively. The results indicated rapid development of neutralizing antibodies within this Chinese patient over a 5-year period.\n\n### Identification of a CD4-Binding Site-Directed VRC01-like Antibody, DRVIA7", "# Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly\nAuthors: Deller, M.C.; Johnson, H.A.; Miller, M.D.; Spraggon, G.; Elsliger, M.-A.; Wilson, I.A.; Lesley, S.A.\nSource: PLoS ONE, 2015\n\nThe third major protein interface is between adjacent TM1088A and TM1088B dimer pairs. This hydrophobic surface is composed of TM1088A residues on \u03b13 (Asp83-Phe97) and \u03b14 (Ile112-Asn118) and TM1088B residues on \u03b14 (Ile82-Asp76) and \u03b16 (Ile107-Met111). Conserved residues buried in this interface include TM1088A Phe88, Met91, Ile114 and Phe115 (highlighted by red sticks in Fig 2B) and TM1088B Leu80, Phe81 and Met111 (highlighted by red sticks in Fig 2C). This is the smallest of the three protein interfaces in the octameric TM1088 assembly involving ~1500 \u00c5\u00b2 of total buried surface area and 91 non-bonded interactions and a total of 28 residues (Table 2). This interface is structurally equivalent to the \"helix crossover\" motif, which has been previously shown to be central to the gating ring mechanism in the RCK family of proteins.\n\nOverall, the protein:protein interfaces of the TM1088 assembly are extensive (just under 18,000 \u00c5\u00b2 of total buried surface area), specific (32 hydrogen bonds) and structurally homologous to interfaces observed in other K+ gating ring assemblies, further indicating the partnering of the components is highly specific and correctly assembled (Table 2).\n\n### Nucleotide binding and other structural features of TM1088 proteins\n\nAMP was co-purified with TM1088A and appears in both the high-resolution TM1088A structure and the low-resolution structure of the octameric TM1088 assembly (AMP shown as yellow spheres or sticks in Figs 1C, 2A\u20132B and 3B).\n\nAMP was not included in the crystallization reagents and was presumably scavenged from the E. coli cells during expression. NADH was included in the crystallization reagents as a result of database annotations that suggest NADH as a potential ligand. Previous differential scanning calorimetry (DSC) and Thermofluor experiments conducted in our lab showed that adenosine-containing nucleotides are capable of binding to TM1088A [9]. Specifically, AMP stabilized the TM1088A melting temperature by +5.8 \u00b0C (DSC) and +8 \u00b0C (Thermofluor) and NADH by +2.3 \u00b0C (DSC) and +4 \u00b0C (Thermofluor). AMP showed the highest degree of stabilization for all adenosine nucleotides tested and no stabilization of TM1088B was observed as expected [9]. DSC and Thermofluor analysis could not be performed on the TM1088 complex, as a result of a melt profile in excess of 130 \u00b0C, which is outside the range of the DSC and Thermofluor equipment.\n\nResidues involved in nucleotide interaction include the highly conserved loop region between \u03b11 and \u03b21 that contain the nucleotide binding motif (Gly14-X-Gly16-X-X-Gly19) and additional residues, including Asp37, Gly57, Ala59, Thr80 and Arg105 (Fig 3B). Key conserved residues involved in hydrogen bonding with the AMP include Ala59, Arg17 and Arg105, which function in anchoring the AMP at opposite ends via the adenine and phosphate moieties of the AMP (Fig 3B). The AMP is located in a similar position and orientation to the adenosine-containing nucleotides bound in other KtrAB and TrkA structures [20,21].\n\nOne unique and distinguishing feature of the TM1088 assembly is the pattern of the hydrogen-bonding network observed across the central cavity between opposing TM1088B dimers (highlighted in red in Fig 1C) as a pseudo-symmetrical arrangement formed by Arg33 of helix \u03b12 with Glu37 and Glu34 of the opposing \u03b12' (Fig 3C). This interaction displays good electron density and, therefore, must be highly ordered (Fig 3D).\n\n### TM1088 proteins are structurally homologous to each other and to other RCK domains\n\nAlthough TM1088A and TM1088B share only 23% sequence identity, they are structurally homologous to each other (RMSD of 1.2 \u00c5 over 122 aligned C\u03b1 atoms, Fig 4A). TM1088A is shorter and consists only of an N-terminal TrkA domain and linker region. Given the somewhat atypical nature of the two-subunit TM1088 assembly, we conducted extensive structural superpositions and DALI searches to identify the closest structural homologs of TM1088 proteins [49]. TM1088 proteins are structurally similar to all previously determined RCK structures with the best structural alignment occurring with the N-terminal Trk domain of TM1088A (N2) and TrkA of V. parahaemolyticus (Fig 4B, PDB ID: 4J9V, RMSD of 1.2 \u00c5 over 71 C\u03b1 atoms) [21]. The same TrkA domain also aligns well with the N-terminal Trk domain (N1) of TM1088B (Fig 4C, PDB ID: 4J9V, RMSD of 1.0 \u00c5 over 72 C\u03b1 atoms). Other notable structural alignments of TM1088A include those with RCK domains of TrkA from M. jannaschii (PDB ID: 1LSS, RMSD of 1.1 \u00c5 over 99 C\u03b1 atoms) [19], KtrA of B. subtilis (PDB ID: 1LSU, RMSD of 3.8 \u00c5 over 81 aligned residues) [19], MthK of M. thermautotrophicus (PDB ID: 1LNQ, RMSD of 1.6 \u00c5 over 92 aligned atoms) [16] and BK of Danio rerio (PDB ID: 3U6N, RMSD of 1.90 \u00c5 over 93 C\u03b1 atoms) [21]. It is important to note that full-length, non-flexible superpositions of TM1088B (N1 and C1 domains together) do not result in good alignments with other RCK proteins as a consequence of different relative positions of the N- and C-terminal RCK domains (N1 and C1) and changes in the inter-domain linker regions and domain-swapping.", "# Marc Deller's vibe-coding portfolio: AI-assisted scientific apps and tools\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar, mdeller.com, marcdeller.com, 2026\n\nGOBSMACKED (Ground-truth Overlay for Binding Sites, Modes And Complex Kinetics/Dynamics) is an end-to-end fold \u2192 dock \u2192 MD \u2192 analysis web app. A UniProt sequence or PDB code goes in; ESMFold folds it if no structure exists; a SMILES ligand is docked with PandaDock into a user-selected pocket; OpenMM runs minimisation and a short molecular-dynamics simulation; and the complex is analysed interactively with PandaMap, PLIP and Mol*, with annotation from KLIFS (kinases), GPCRdb (GPCRs), Pfam and UniProt. The goal is to verify a predicted, docked protein\u2013ligand complex against an experimentally determined structure with a scorecard, an apo\u2192holo induced-fit three-way overlay with an MD rescue score, and a family-aware binding-mode classifier. It uses a two-step Prepare/Analyze design (a downloadable pixi run bundle executed locally, results uploaded for analysis, because the droplet cannot carry the compute), an optional Boltz-2 affinity-only scoring stage, and a deliberately different \"instrument panel\" front end (dark slate control-room look, phosphor cyan/amber/green, Sora typeface).\n\n## ALPHABETTI (alphabetti.mdeller.com)\n\nALPHABETTI (Amino-acid Letters Plotted Helically As Backbone-Embedded Text In 3D) is a Flask web app that takes a FASTA sequence or UniProt ID, folds it on the fly with ESMFold, and renders the sequence as a rotatable 3D sequence logo instead of a conventional structure viewer. Four tabs sit on the base engine: GIBBERISH (3D sequence logo with entropy in bits), BUMFLUFF (solvent accessibility), BALDERDASH (variant and substitution hotspots) and FOLDEROL (a 2D-to-3D fold animation). The emphasis is on entertaining, engaging visuals.\n\n## PANTS\n\nPANTS (PETase ANnotation and Triage System) is a plastic-degrading-enzyme mining tool that searches metagenomic assemblies for uncharacterised PETase-like and MHETase-like candidates using an ESM-2 (t12, 35M) classifier trained on characterised cutinase/esterase families and PAZy database entries, returning a ranked candidate list with the catalytic triad and oxyanion hole flagged. It feeds the enzyme-discovery work of Elora Therapeutics.\n\n## ChemSage\n\nChemSage is a hybrid retrieval-augmented-generation (RAG) plus QLoRA fine-tuned local language model for drug discovery. Its RAG corpus contains 41 reference Markdown files covering RDKit, PyMOL, PLIP, ChimeraX, gemmi, pdb-tools, Biopython and cheminformatics APIs. Marc described the approach in a blog post.\n\n## chatMCD\n\nchatMCD is Marc's personal chatbot: a small (~8B-parameter) open-weight language model fine-tuned locally with MLX-LM on his publications, thesis, patents, CV and personal notes, so that it can answer questions about Marc, his science and his projects. This document is part of its training corpus.\n\n## Planned and smaller projects\n\nOther planned or in-progress tools include FlexAppeal (an OpenMM-based, GPU-backed protein molecular-dynamics pipeline app), an FEP app for automating and analysing alchemical free-energy calculations, a crystallisation-condition predictor built on normalised PDB crystallisation data (growing out of his JCSG Top96 work), HAWKER (a Swift multiplatform app mining failed or abandoned clinical assets for reclaimable drugs, pockets and targets), and a family of Swift apps for Apple's Neural Engine: BOFFIN (on-device protein analysis), PhoneFold (real-time ESMFold folding with music), JUMPjet (crude on-device protein dynamics) and PfamIE (a Protein Family Inference Engine using Pfam). He is also training a small LLM on the PSI TargetTrack structural-genomics archive with MLX-LM.\n\nSmaller branded HTML apps and scripts include the Confidence\u2013Flexibility Atlas, the PETase Lineage Map, the Ligandable Cysteine Atlas, a Resolution Dashboard, Name Folder, the Structural Glyphs generative SVG fingerprint atlas, a PDF Timebomb Locker (Python plus a WordPress plugin), and cif_to_plip.py, a CIF \u2192 PDB \u2192 PLIP pipeline whose development documented a pdb_tidy CONECT serial-gap bug that corrupts OpenBabel bond perception.\n\n## Custom Claude skills\n\nMarc maintains a suite of custom skills for his AI assistants: marcs-profile (his professional and personal context), marcs-blog-writer (a research-driven, 400-word blog-post workflow delivering publication-ready HTML to marcdeller.com), marcs-resume (ATS-optimised resume tailoring), marcs-job-hunter (multi-board job search with an HTML report), marcs-vibe-coding (HTML/CSS/JS scientific app generation with a branded README standard and the readme_forge.py polish script), marcs-vibe-icon (300\u00d7300 house-style SVG app icons) and marcs-page-icon (flat duotone-blue section icons for marcdeller.com, featuring a recurring blue-haired, bespectacled scientist character), plus a BoltzMaker input-file wizard.\n\n## Website\n\nmarcdeller.com is a WordPress/Elementor site hosted on Cloudways with a Breeze/Varnish/Cloudflare APO performance stack. Its sections include About, AI, Blog, Chat, Learnings, Microplastics, Patents, Projects, Publications, Resume, Structures, Structural Biology and Vibe Coding, each with a matching page icon.", "# Engineered Enzymes\u2014Design for Safe, Targeted Microplastic Therapeutics\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-15 on marcdeller.com (categories: blog, microplastics).*\n\n* Enzyme engineering uses mutations, fusion domains, and glycosylation for optimal PETase variants.\n* Design advances increase activity, thermostability, and substrate specificity.\n* Delivery technologies boost circulation, targeting, and immune evasion.\n* Safety (low cytotoxicity) and tissue targeting are top design priorities.\n* The future is custom-designed biocatalysts for microplastic remediation in medicine.\n\n1. [Current PETase engineering advances (2023)](https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2023.1263996/full)\n2. [Glycan engineering for PETase stability (2023)](https://www.nature.com/articles/s42003-023-04413-0)\n3. [PETase serum paper (2024)](https://pubs.rsc.org/en/content/articlepdf/2024/fd/d4fd00014e)\n4. [Delivery of cytotoxic enzymes (2022)](https://pmc.ncbi.nlm.nih.gov/articles/PMC9230553/)\n5. [Cytochrome P450 enzyme design (2024)](https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2024.1405466/full)\n\nEngineering enzymes for therapeutic microplastic removal combines rational mutagenesis, fusion protein technologies, and controlled glycosylation to create PETase variants with enhanced activity, stability, and safety. Loop and surface mutations optimize enzyme-substrate interactions, while fusion with CBMs or hydrophobins improves targeting and affinity for plastics found in the bloodstream. Glycosylation modulates enzyme resilience against blood proteases and immune recognition.\n\nDelivery techniques, including encapsulation, PEGylation, and cell-surface anchoring, further improve distribution, reduce cytotoxicity, and enable organ targeting. Hybrid designs allow for companion domains to process a wider array of microplastics. Customized biocatalysts tailored to the patient\u2019s plastic burden are on the horizon, transforming landfill enzymes into medical microplastic eliminators.", "# GitHub README: CODSWALLOP\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/CODSWALLOP, 2026\n\n- [x] **Mol\\* viewer embedded, single entry.** From the index card and from a Structures panel with an entry picker that follows the current filters. Loaded **on demand**, because Mol\\* is 5 MB, larger than the whole rest of the page including a 1,500-entry family payload: somebody who came for the construct table should not pay for a renderer they never open. It reads structures straight from the RCSB, so this needs no downloads, no storage and no toolchain on the server, which is what let it ship separately from the pipeline work below. Viewport colour follows the theme toggle\n- [x] **Family superposition, and an AlphaFold overlay with pLDDT colouring.** The transforms fall out of the same TM-align run that builds the matrix. The reference is the representative most similar to everything else *among the seed's own structures*, not the best-resolution one: superposing onto an outlier makes every other structure look wrong, and it is forced into the loaded set because every transform maps onto its frame. Restricting it to the seed is not fussiness. A family is assembled at 30 % identity, so an ABL1 search legitimately returns most of the tyrosine kinases, and the most central structure of that set is an EGFR entry: the page superposed ABL1's family onto a different protein, with nothing on it saying so. The AlphaFold model **is** superposed, onto the same reference as everything else: the alignment is the same TM-align call, done in the pipeline where TM-align lives rather than in a browser that cannot do it, and the panel reports the TM-score so an overlay of something that does not actually superpose is visible as such. Its URL comes from the AFDB API rather than being constructed: this was written against `-model_v4.cif` and the DB already serves v6. The accession it fetches is the family's *seed*, never the accession the most members carry, which is a popularity contest the subject frequently loses: ABL1 and JAK1 both fetched EGFR's model, and every A2A receptor structure carries a BRIL fusion, so A2A fetched the model of *E. coli* cytochrome b562. Both rendered as a confident superposition of the wrong protein at a TM-score that looked merely poor rather than meaningless.\n- [x] **Pairwise TM-score matrix, and the embedding it produces.** The placeholder layout is retired for any family that has one: node positions are classical MDS of the 1&nbsp;&minus;&nbsp;TM distance matrix, so distance on the map is structural distance and the 0.5 contour is the conventional same-fold boundary. Computed **on a workstation, never on the droplet** (`CODSWALLOP.py embed`, then `deploy/push_embeddings.sh`): it downloads mmCIF files and does real numerical work, and the box has two cores shared with eight apps. `tmtools` and `biotite` are in `requirements-dev.txt` only, and the app reads a JSON artefact through a reader that imports nothing beyond the standard library. (numpy is on the droplet regardless, as a biopython dependency: it is the mmCIF parsing and TM-align that stay off it.) One representative per distinct construct rather than per entry, capped, because superposing two crystal forms of an identical construct measures crystallography rather than biology. Half the cap is reserved for constructs of the protein that was actually searched for, topped up only where ranking by usage would leave it under-represented, so a family whose seed is dominant keeps exactly the nodes it had. Without the reservation an ABL1 search put 2 of its 80 nodes on ABL1, and both were its 63-residue SH3 domain: the map, the reference and the AlphaFold model were all EGFR's. It is now 40 of 80 at a median length of 270, which is the kinase domain *(Superseded below: the workstation-only rule fell to per-chain fetching, and the droplet computes this itself now. `requirements-compute.txt`, not `requirements-dev.txt`.)*\n- [x] **The matrix as a clustered heatmap, with its clusters as filters.** Average linkage in the browser, so the cut height is a live control: dragging it is the only way to tell whether a grouping is real or an artefact of where the line sits. Drawn on a canvas, because a 260-representative family is 67,600 cells. The default cut is 0.18, not the conventional 0.5 same-fold line, which returns one cluster for almost any family because every member of a family is the same fold by construction. Picking a cluster narrows the map, the entry list, the table and every other panel through the shared selection state.\n- [x] **Ligand panel.** CCD ID, formula, SMILES, occurrence count and best resolution achieved, with a per-component classification into ligand, cofactor, lipid/detergent, ion, buffer, cryoprotectant or solvent. This retires the Phase 1 exclusion list: \"ligand-bound\" now means a ligand or a cofactor, which took carbonic anhydrase II from 1,208 entities to 1,096 and the receptor from 409 to 395. Each component carries the RCSB's own 2D depiction, fetched lazily because a family can hold hundreds. Whether a metal is structural or catalytic is a property of the protein and not of the component, so this began by filing carbonic anhydrase's catalytic zinc as an ion and saying so on the panel. It is now resolved with evidence rather than by declaring it unresolvable: a metal is promoted to cofactor when the interaction data shows it coordinated by at least 3 residues that are conserved above 0.90, across at least 5 entries, and the promotion states which residues did it. The entry threshold is doing real work, since a single mercury on one entry qualified on every other test\n- [x] **PLIP family-wide.** A hot-residue ranking with each residue's conservation beside it, and a ligand-by-residue fingerprint heatmap, in seed coordinates so a hot residue is the same residue as in the conservation track and the domain ribbon. Workstation only, like the embedding. The conversion reuses Marc's `cif2plip.py` verbatim (vendored to `pipeline/`) for its `pdb_tidy` CONECT serial-gap fix; only the PLIP invocation differs, asking for XML rather than the PyMOL session and images a web page cannot use. Validated on carbonic anhydrase II: 890 contacts over 40 entries with no conversion failures, and the top residues come out as His96, Thr199, Thr200, Leu198, His94 and His119, which is the zinc triad and the gatekeeper threonines, found with no prior knowledge of the enzyme anywhere in the pipeline.\n- [x] **Crystallisation panel.** `_exptl_crystal_grow.pdbx_details` is free text a depositor typed, so this is a parser, not a field read: 17 precipitant patterns, 15 buffers with their working pH ranges, 14 additives and 9 setup methods, each carrying the best resolution anyone achieved with it. On the beta-2 adrenergic receptor it finds lipidic cubic phase in 79 of 120 conditions with monoolein and cholesterol as the top additives, which is exactly how that receptor is crystallised. CSV export for TopPDBLX\n- [x] **Quality and validation panel.** Clashscore, RSRZ, Ramachandran and rotamer outliers, R-free minus R-work gap, data completeness and structure-factor availability, as a traffic light against the wwPDB report's own thresholds, worst first, with the family's medians for context. The figures live in `pdbx_vrpt_summary_geometry` and `pdbx_vrpt_summary_diffraction`; `pdbx_vrpt_summary` itself holds almost nothing", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices\nAuthors: Deller, M.C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nConsiderable mutagenesis work has been carried out on the gp130 family of cytokines that reveals, in common with GH, several discrete sites on the molecules involved in interactions with receptor components. Sites 1 and 2 are in broadly similar locations of the molecule to receptor binding sites on GH, while a third site has been identified that enables an extra interaction not seen in the GH-GHR system. The structure-function relationships of LIF (Hudson et al., 1996; Layton et al., 1994a; Layton et al., 1994b; Owczarek et al., 1993) and CNTF (Di Marco et al., 1996; Inoue et al., 1995; Panayotatos et al., 1995) are the most fully understood due to the availability of crystal structures. A summary of the structural and mutagenesis data for each of these molecules will now be discussed. IL-6 and IL-11 are of less importance to this study, as they do not participate in LIFR binding. However, the crystal structure of IL-6 has been solved and mutagenesis data are available (Simpson et al., 1997b) so comparisons can also be made. Discussion of the structure-function relationship for the Epo and GCSF systems will also be discussed due to the structural similarity with the OSM and gp130 system.\n\n#### 6.3.1 Leukemia Inhibitory Factor\n\nAs discussed previously in 6.2.1 human OSM exhibits close structural similarity to human LIF with RMSD in C\u03b1 positions of 2.27 \u00c5 for 138 structurally equivalent residues with a sequence identity of ~20 %. Through the use of native gel electrophoresis, gel-filtration chromatography, sedimentation equilibrium analysis, surface plasmon resonance spectroscopy and chemical cross-linking it has been shown that LIF forms a heterotrimeric complex with the extracellular portions of LIFR and gp130 in a stoichiometry of 1:1:1 (Zhang et al., 1997b). The complex is stabilised by site 1 and site 3 interactions with LIFR, and this system differs from that of hGH in that site 3 interactions predominate energetically. Site 3 interactions facilitate LIFR binding through a recently identified LIGHT motif in the Ig-like domain of LIFR (Chobotova and Heath, 2000).\n\n##### 6.3.1.1 Site 1 Interactions\n\nAlanine scanning mutagenesis of murine LIF (Hudson et al., 1996) has revealed a site analogous to site 1 of hGH, responsible for LIFR binding, centralised around Val 175 with contributions from Asp 57, Lys 58, Lys 170 and Ala 174. These residues map to the C-terminus of helix D and the A-B loop as shown in blue in Figure 6-10 (a) below. Structurally equivalent residues in OSM are Arg 46, Glu 47, His 174, His 178 and Ser 179 as shown in Figure 6-10 (b) below, although the function of these residues in OSM is currently unknown and awaiting further mutational analysis. However, these residues, with the exception of His 174 and Ser 179, are solvent exposed in the OSM crystal structure and as such would be expected to form a similar site 1 epitope to that of LIF.\n\n##### 6.3.1.2 Site 3 Interactions\n\nThe residues of site 1 contribute weakly to LIFR binding and residues Phe 156 and Lys 159, located on the N-terminus of helix D as shown in red in Figure 6-10 (a) below, contribute the majority of the free energy for binding (Hudson et al., 1996). Alanine mutation of these residues individually reduces binding more than 100-fold and biological activity by up to 3000-fold. Combined alanine mutagenesis eliminates biological activity completely. These two residues are solvent exposed and are conserved in all members of the gp130 family that utilise LIFR; in contrast these residues are not conserved in IL-6 and IL-11 which do not use LIFR (see section 6.4 and Figure 6-12 for sequence alignments). Surrounding Phe 156 and Lys 159 are several other residues that influence LIFR binding; this situation is analogous to that of growth hormone in which key residues are surrounded by ones of lesser importance (Cunningham and Wells, 1989).\n\nStructurally equivalent residues in OSM are Phe 160 and Lys 163, which reside on the N-terminus of helix D as shown in red in Figure 6-10 (b) below. These two residues have been confirmed as the major LIFR binding epitope for OSM with alanine mutagenesis of either residue resulting in a >100-fold decrease in binding to LIFR and a >10000-fold reduction in biological activity, with no effect on gp130 binding (Deller et al., 2000).\n\nThrough analysis of the solution dynamics and secondary structure of murine LIF it has been suggested that the C-D loop of LIF is rigid and shows limited mobility in solution (Purvis and Mabbutt, 1997). In comparison, other cytokines utilising the LIFR unit demonstrate marked mobility in the C-D loop. For example, previous studies have shown that the C-D loop of OSM adopts a long extended conformation in solution (Hoffman et al., 1996) and this present study supports this claim, as considerable disorder and mobility in this region is observed. Also, the crystal structure of CNTF (McDonald et al., 1995) shows considerable flexibility in this region. Although no directly functional residues lie on the C-D loop the site 3 motif lies close on the N-terminus of helix D and as such mobility in this region could affect receptor engagement through some form of entropic component.\n\nInterestingly OSM and CNTF, which both show high levels of flexibility in this region both show low affinity binding to LIFR, compared with LIF. This could also help account for the differential selection of LIFR or OSMR in the OSM signalling system and also the observed anomaly of high affinity binding of hLIF to mLIFR compared to mLIF binding to mLIFR (see 6.5.1).\n\n##### 6.3.1.3 Site 2 Interactions", "# Google Scholar profile of Marc C. Deller\nAuthors: Deller, M.C.\nSource: Google Scholar, 2026\n\n- Crystal structure of a soluble cleaved HIV-1 envelope trimer (2013). JP Julien, A Cupo, D Sok, RL Stanfield, D Lyumkis, MC Deller, PJ Klasse, .... Science 342 (6165), 1477-1483, 2013. Cited by 1027.\n- Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120 (2013). L Kong, JH Lee, KJ Doores, CD Murin, JP Julien, R McBride, Y Liu, .... Nature structural & molecular biology 20 (7), 796-803, 2013. Cited by 443.\n- Protein stability: a crystallographer's perspective (2016). MC Deller, L Kong, B Rupp. Structural Biology and Crystallization Communications 72 (2), 72-95, 2016. Cited by 440.\n- The role of the secondary coordination sphere in a fungal polysaccharide monooxygenase (2017). EA Span, DLM Suess, MC Deller, RD Britt, MA Marletta. ACS chemical biology 12 (4), 1095, 2017. Cited by 146.\n- Models of protein\u2013ligand crystal structures: trust, but verify (2015). MC Deller, B Rupp. Journal of computer-aided molecular design 29 (9), 817-836, 2015. Cited by 130.\n- Crystal structures of two novel dye\u2010decolorizing peroxidases reveal a \u03b2\u2010barrel fold with a conserved heme\u2010binding motif (2007). C Zubieta, SS Krishna, M Kapoor, P Kozbial, D McMullan, HL Axelrod, .... Proteins: Structure, Function, and Bioinformatics 69 (2), 223-233, 2007. Cited by 108.\n- Structure of hepatitis C virus envelope glycoprotein E2 antigenic site 412 to 423 in complex with antibody AP33 (2012). L Kong, E Giang, T Nieusma, JB Robbins, MC Deller, RL Stanfield, .... Journal of virology 86 (23), 13085-13088, 2012. Cited by 103.\n- Structural basis of murein peptide specificity of a \u03b3-D-glutamyl-L-diamino acid endopeptidase (2009). Q Xu, S Sudek, D McMullan, MD Miller, B Geierstanger, DH Jones, .... Structure 17 (2), 303-313, 2009. Cited by 99.\n- Key gp120 glycans pose roadblocks to the rapid development of VRC01-class antibodies in an HIV-1-infected Chinese donor (2016). L Kong, B Ju, Y Chen, L He, L Ren, J Liu, K Hong, B Su, Z Wang, .... Immunity 44 (4), 939-950, 2016. Cited by 98.\n- Structure of the \u03b3-D-glutamyl-L-diamino acid endopeptidase YkfC from Bacillus cereus in complex with L-Ala-\u03b3-D-Glu: insights into substrate recognition by NlpC/P60 cysteine\ufffd\u2026 (2010). Q Xu, P Abdubek, T Astakhova, HL Axelrod, C Bakolitsa, X Cai, D Carlton, .... Structural Biology and Crystallization Communications 66 (10), 1354-1364, 2010. Cited by 97.\n- Crystal structure and functional dissection of the cytostatic cytokine oncostatin M (2000). MC Deller, KR Hudson, S Ikemizu, J Bravo, EY Jones, JK Heath. Structure 8 (8), 863-874, 2000. Cited by 93.\n- Identification and structural characterization of heme binding in a novel dye\u2010decolorizing peroxidase, TyrA (2007). C Zubieta, R Joseph, S Sri Krishna, D McMullan, M Kapoor, HL Axelrod, .... Proteins: Structure, Function, and Bioinformatics 69 (2), 234-243, 2007. Cited by 91.\n- Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design (2017). A Irimia, AM Serra, A Sarkar, R Jacak, O Kalyuzhniy, D Sok, .... PLoS pathogens 13 (2), e1006212, 2017. Cited by 85.\n- Complete epitopes for vaccine design derived from a crystal structure of the broadly neutralizing antibodies PGT128 and 8ANC195 in complex with an HIV-1 Env trimer (2015). L Kong, A Torrents de la Pe\ufffda, MC Deller, F Garces, K Sliepen, Y Hua, .... Biological Crystallography 71 (10), 2099-2108, 2015. Cited by 74.\n- Structure-based discovery of NANOG variant with enhanced properties to promote self-renewal and reprogramming of pluripotent stem cells (2015). Y Hayashi, L Caboni, D Das, F Yumoto, T Clayton, MC Deller, P Nguyen, .... Proceedings of the National Academy of Sciences 112 (15), 4666-4671, 2015. Cited by 66.\n- Functional and structural characterization of a thermostable acetyl esterase from Thermotoga maritima (2012). M Levisson, GW Han, MC Deller, Q Xu, P Biely, S Hendriks, LF Ten Eyck, .... Proteins: Structure, Function, and Bioinformatics 80 (6), 1545-1559, 2012. Cited by 66.\n- Cell surface receptors (2000). MC Deller, EY Jones. Current opinion in structural biology 10 (2), 213-219, 2000. Cited by 59.\n- Approaches to automated protein crystal harvesting (2014). MC Deller, B Rupp. Structural Biology and Crystallization Communications 70 (2), 133-155, 2014. Cited by 55.\n- Structure\u2013function analysis of the extended conformation of a polyketide synthase module (2018). X Li, N Sevillano, F La Greca, L Deis, YC Liu, MC Deller, II Mathews, .... Journal of the American Chemical Society 140 (21), 6518, 2018. Cited by 52.\n- Crystal structure of the first eubacterial Mre11 nuclease reveals novel features that may discriminate substrates during DNA repair (2010). D Das, D Moiani, HL Axelrod, MD Miller, D McMullan, KK Jin, P Abdubek, .... Journal of molecular biology 397 (3), 647-663, 2010. Cited by 48.\n- Bacterial pleckstrin homology domains: a prokaryotic origin for the PH domain (2010). Q Xu, A Bateman, RD Finn, P Abdubek, T Astakhova, HL Axelrod, .... Journal of molecular biology 396 (1), 31-46, 2010. Cited by 45.\n- Structure of a membrane-attack complex/perforin (MACPF) family protein from the human gut symbiont Bacteroides thetaiotaomicron (2010). Q Xu, P Abdubek, T Astakhova, HL Axelrod, C Bakolitsa, X Cai, D Carlton, .... Structural Biology and Crystallization Communications 66 (10), 1297-1305, 2010. Cited by 44.\n- Crystal structure of the Fic (Filamentation induced by cAMP) family protein SO4266 (gi|24375750) from Shewanella oneidensis MR\u20101 at 1.6 \ufffd resolution (2009). D Das, SS Krishna, D McMullan, MD Miller, Q Xu, P Abdubek, C Acosta, .... Proteins: Structure, Function, and Bioinformatics 75 (1), 264-271, 2009. Cited by 42.\n- Structural and functional characterizations of SsgB, a conserved activator of developmental cell division in morphologically complex actinomycetes (2009). Q Xu, BA Traag, J Willemse, D McMullan, MD Miller, MA Elsliger, .... Journal of Biological Chemistry 284 (37), 25268-25279, 2009. Cited by 35.\n- Crystal structure of a novel Sm\u2010like protein of putative cyanophage origin at 2.60 \ufffd resolution (2009). D Das, P Kozbial, HL Axelrod, MD Miller, D McMullan, SS Krishna, .... Proteins: Structure, Function, and Bioinformatics 75 (2), 296-307, 2009. Cited by 27.\n- Structure of the first representative of Pfam family PF09410 (DUF2006) reveals a structural signature of the calycin superfamily that suggests a role in lipid metabolism (2010). HJ Chiu, C Bakolitsa, A Skerra, A Lomize, D Carlton, MD Miller, .... Structural Biology and Crystallization Communications 66 (10), 1153-1159, 2010. Cited by 24.\n- The structure of SSO2064, the first representative of Pfam family PF01796, reveals a novel two-domain zinc-ribbon OB-fold architecture with a potential acyl-CoA-binding role (2010). SS Krishna, L Aravind, C Bakolitsa, J Caruthers, D Carlton, MD Miller, .... Structural Biology and Crystallization Communications 66 (10), 1160-1166, 2010. Cited by 23.\n- Discovery of potent and selective inhibitors of wild-type and gatekeeper mutant fibroblast growth factor receptor (FGFR) 2/3 (2022). A Shvartsbart, JJ Roach, MR Witten, H Koblish, JJ Harris, M Covington, .... Journal of Medicinal Chemistry 65 (22), 15433-15442, 2022. Cited by 21.\n- The structure of the first representative of Pfam family PF09836 reveals a two-domain organization and suggests involvement in transcriptional regulation (2010). D Das, NV Grishin, A Kumar, D Carlton, C Bakolitsa, MD Miller, .... Structural Biology and Crystallization Communications 66 (10), 1174-1181, 2010. Cited by 20.\n- Structure of BT_3984, a member of the SusD/RagB family of nutrient-binding molecules (2010). C Bakolitsa, Q Xu, CL Rife, P Abdubek, T Astakhova, HL Axelrod, .... Structural Biology and Crystallization Communications 66 (10), 1274-1280, 2010. Cited by 20.\n- Discovery of INCB159020, an orally bioavailable KRAS G12D inhibitor (2025). Q Ye, A Shvartsbart, Z Li, P Gan, RL Policarpo, C Qi, JJ Roach, W Zhu, .... Journal of Medicinal Chemistry 68 (2), 1924-1939, 2025. Cited by 19.\n- INCB161734: A novel, potent, and orally bioavailable KRAS G12D selective inhibitor demonstrates antitumor activity in KRAS G12D mutant tumors (2024). MR Farren, V Roman, A Gallion, A Allali-Hassani, A Sokolsky, W Kong, .... Cancer Research 84 (6_Supplement), 5900-5900, 2024. Cited by 19.\n- Discovery of (4-pyrazolyl)-2-aminopyrimidines as potent and selective Inhibitors of cyclin-dependent kinase 2 (2024). JR Hummel, KJ Xiao, JC Yang, LB Epling, K Mukai, Q Ye, M Xu, D Qian, .... Journal of Medicinal Chemistry 67 (4), 3112-3126, 2024. Cited by 18.\n- Crystal structure of homoserine O\u2010succinyltransferase from Bacillus cereus at 2.4 \ufffd resolution (2007). C Zubieta, SS Krishna, D McMullan, MD Miller, P Abdubek, S Agarwalla, .... Proteins: Structure, Function, and Bioinformatics 68 (4), 999-1005, 2007. Cited by 17.\n- Crystal structure of a metal\u2010dependent phosphoesterase (YP_910028.1) from Bifidobacterium adolescentis: Computational prediction and experimental validation\ufffd\u2026 (2011). GW Han, J Ko, CL Farr, MC Deller, Q Xu, HJ Chiu, MD Miller, J Sefcikova, .... Proteins: Structure, Function, and Bioinformatics 79 (7), 2146-2160, 2011. Cited by 16.\n- A structural basis for the regulatory inactivation of DnaA (2009). Q Xu, D McMullan, P Abdubek, T Astakhova, D Carlton, C Chen, HJ Chiu, .... Journal of molecular biology 385 (2), 368-380, 2009. Cited by 16.\n- The structure of Haemophilus influenzae prephenate dehydrogenase suggests unique features of bifunctional TyrA enzymes (2010). HJ Chiu, P Abdubek, T Astakhova, HL Axelrod, D Carlton, T Clayton, .... Structural Biology and Crystallization Communications 66 (10), 1317-1325, 2010. Cited by 15.\n- A conserved fold for fimbrial components revealed by the crystal structure of a putative fimbrial assembly protein (BT1062) from Bacteroides thetaiotaomicron at 2.2 \ufffd resolution (2010). Q Xu, P Abdubek, T Astakhova, HL Axelrod, C Bakolitsa, X Cai, D Carlton, .... Structural Biology and Crystallization Communications 66 (10), 1281-1286, 2010. Cited by 14.\n- The structure of Jann_2411 (DUF1470) from Jannaschia sp. at 1.45 \ufffd resolution reveals a new fold (the ABATE domain) and suggests its possible role as a transcription regulator (2010). C Bakolitsa, A Bateman, KK Jin, D McMullan, SS Krishna, MD Miller, .... Structural Biology and Crystallization Communications 66 (10), 1198-1204, 2010. Cited by 14.\n- Crystal structure of histidine phosphotransfer protein ShpA, an essential regulator of stalk biogenesis in Caulobacter crescentus (2009). Q Xu, D Carlton, MD Miller, MA Elsliger, SS Krishna, P Abdubek, .... Journal of molecular biology 390 (4), 686-698, 2009. Cited by 14.\n- The structure of BVU2987 from Bacteroides vulgatus reveals a superfamily of bacterial periplasmic proteins with possible inhibitory function (2010). D Das, RD Finn, D Carlton, MD Miller, P Abdubek, T Astakhova, .... Structural Biology and Crystallization Communications 66 (10), 1265-1273, 2010. Cited by 13.\n- The structure of the first representative of Pfam family PF06475 reveals a new fold with possible involvement in glycolipid metabolism (2010). C Bakolitsa, A Kumar, D McMullan, SS Krishna, MD Miller, D Carlton, .... Structural Biology and Crystallization Communications 66 (10), 1211-1217, 2010. Cited by 12.\n- Open and closed conformations of two SpoIIAA-like proteins (YP_749275. 1 and YP_001095227. 1) provide insights into membrane association and ligand binding (2010). A Kumar, A Lomize, KK Jin, D Carlton, MD Miller, L Jaroszewski, .... Structural Biology and Crystallization Communications 66 (10), 1245-1253, 2010. Cited by 11.\n- Structural insights into the recognition of phosphopeptide by the FHA domain of kanadaptin (2014). Q Xu, MC Deller, TK Nielsen, JC Grant, SA Lesley, MA Elsliger, .... PLoS One 9 (9), e107309, 2014. Cited by 10.\n- Structures of the first representatives of Pfam family PF06938 (DUF1285) reveal a new fold with repeated structural motifs and possible involvement in signal transduction (2010). GW Han, C Bakolitsa, MD Miller, A Kumar, D Carlton, RJ Najmanovich, .... Structural Biology and Crystallization Communications 66 (10), 1218-1225, 2010. Cited by 10.\n- Structures of three members of Pfam PF02663 (FmdE) implicated in microbial methanogenesis reveal a conserved \u03b1+ \u03b2 core domain and an auxiliary C-terminal treble-clef zinc finger (2010). HL Axelrod, D Das, P Abdubek, T Astakhova, C Bakolitsa, D Carlton, .... Structural Biology and Crystallization Communications 66 (10), 1335-1346, 2010. Cited by 9.\n- Structure of the first representative of Pfam family PF04016 (DUF364) reveals enolase and Rossmann-like folds that combine to form a unique active site with a possible role in\ufffd\u2026 (2010). MD Miller, L Aravind, C Bakolitsa, CL Rife, D Carlton, P Abdubek, .... Structural Biology and Crystallization Communications 66 (10), 1167-1173, 2010. Cited by 8.\n- Conformational changes associated with the binding of zinc acetate at the putative active site of XcTcmJ, a cupin from Xanthomonas campestris pv. campestris (2010). HL Axelrod, P Kozbial, D McMullan, SS Krishna, MD Miller, P Abdubek, .... Structural Biology and Crystallization Communications 66 (10), 1347-1353, 2010. Cited by 8.\n- Validation of protein\u2013ligand crystal structure models: small molecule and peptide ligands (2017). E Pozharski, MC Deller, B Rupp. Protein Crystallography: Methods and Protocols, 611-625, 2017. Cited by 7.\n- The crystal structure of a bacterial Sufu\u2010like protein defines a novel group of bacterial proteins that are similar to the N\u2010terminal domain of human Sufu (2010). D Das, RD Finn, P Abdubek, T Astakhova, HL Axelrod, C Bakolitsa, X Cai, .... Protein science 19 (11), 2131-2140, 2010. Cited by 6.\n- Structures of the first representatives of Pfam family PF06684 (DUF1185) reveal a novel variant of the Bacillus chorismate mutase fold and suggest a role in amino-acid metabolism (2010). C Bakolitsa, A Kumar, KK Jin, D McMullan, SS Krishna, MD Miller, .... Structural Biology and Crystallization Communications 66 (10), 1182-1189, 2010. Cited by 5.\n- Crystal structures of MW1337R and lin2004: Representatives of a novel protein family that adopt a four\u2010helical bundle fold (2008). P Kozbial, Q Xu, HJ Chiu, D McMullan, SS Krishna, MD Miller, P Abdubek, .... Proteins: Structure, Function, and Bioinformatics 71 (3), 1589-1596, 2008. Cited by 5.\n- Crystal structure of a two-subunit TrkA octameric gating ring assembly (2015). MC Deller, HA Johnson, MD Miller, G Spraggon, MA Elsliger, IA Wilson, .... Plos one 10 (3), e0122512, 2015. Cited by 4.\n- Structure of Bacteroides thetaiotaomicron BT2081 at 2.05 \ufffd resolution: the first structural representative of a new protein family that may play a role in carbohydrate metabolism (2010). AP Yeh, P Abdubek, T Astakhova, HL Axelrod, C Bakolitsa, X Cai, .... Structural Biology and Crystallization Communications 66 (10), 1287-1296, 2010. Cited by 4.\n- Structure of LP2179, the first representative of Pfam family PF08866, suggests a new fold with a role in amino-acid metabolism (2010). C Bakolitsa, A Kumar, D Carlton, MD Miller, SS Krishna, P Abdubek, .... Structural Biology and Crystallization Communications 66 (10), 1205-1210, 2010. Cited by 3.\n- The structure of KPN03535 (gi| 152972051), a novel putative lipoprotein from Klebsiella pneumoniae, reveals an OB-fold (2010). D Das, P Kozbial, GW Han, D Carlton, L Jaroszewski, P Abdubek, .... Structural Biology and Crystallization Communications 66 (10), 1254-1260, 2010. Cited by 3.\n- Crystallisation of proteins and macromolecular complexes: past, present and future (2014). MC Deller, B Rupp. Encyclopedia of life sciences. John Wiley & Sons, Ltd. https://doi. org/10\ufffd\u2026, 2014. Cited by 2.\n- Structure of a putative NTP pyrophosphohydrolase: YP_001813558. 1 from Exiguobacterium sibiricum 255-15 (2010). GW Han, MA Elsliger, TO Yeates, Q Xu, AG Murzin, SS Krishna, .... Structural Biology and Crystallization Communications 66 (10), 1237-1244, 2010. Cited by 2.\n- Discovery of the orally bioavailable isoform selective Janus kinase 1 (JAK1) compound povorcitinib (INCB054707) for the treatment of inflammatory and autoimmune diseases (2026). J Zhuo, YL Li, DQ Qian, DM Burns, S Mei, M Rafalski, M Xu, G Cao, Y Pan, .... Journal of Medicinal Chemistry 69 (11), 12830-12852, 2026. Cited by 1.\n- Structure of the c-D-glutamyl-L-diamino acid endopeptidase YkfC from Bacillus cereus in complex with L-Ala-c-D-Glu: insights into substrate recognition by NlpC/P60 cysteine\ufffd\u2026 (2010). T Astakhova, L Herbert, X Cai, D Carlton, C Chen, HJ Chiu, M Chiu, .... Acta Cryst 66, 1354-1364, 2010. Cited by 1.\n- Fused bicyclic pyrimidones as wrn inhibitors (2026). L Zhao, Y Ai, M Deller, J FELICIANO, A Ghosh, L McQuirter, X Wang, .... US Patent App. 19/063,574, 2026. Cited by .\n- Fused bicyclic pyrimidones as wrn inhibitors (2025). S Shepard, Y Ai, J Barbosa, M Deller, A Ghosh, J Glenn, L McQuirter, .... US Patent App. 18/979,710, 2025. Cited by .\n- Bicyclic compounds as inhibitors of wrn (2025). J Barbosa, Y Ai, M Deller, BD Dherange, A Ghosh, J Glenn, L McQuirter, .... US Patent App. 18/821,138, 2025. Cited by .\n- Imidaz0 [1, 2-a] pyrimidine-5 (1 h)-ones having a covalent warhead at the 3-position as wrn inhibitors (2025). Y Ai, J Barbosa, H Chen, M Deller, BD Dherange, N Falahatpisheh, .... US Patent App. 18/953,485, 2025. Cited by .\n- Imidaz0 [1, 2-a] pyrimidine-5 (1 h)-ones as wrn inhibitors (2025). S Shepard, Y Ai, J Barbosa, M Deller, N Falahatpisheh, A Ghosh, J Glenn, .... US Patent App. 18/953,578, 2025. Cited by .\n- Bicyclic compounds as inhibitors of wrn (2025). Y Ai, J Barbosa, M Deller, BD Dherange, A Ghosh, J Glenn, L McQuirter, .... US Patent App. 18/821,175, 2025. Cited by .\n- Anti-mutant calreticulin (calr) antibodies and uses thereof (2023). P Mayes, HG Nastri, RA Buonpane, SM STEWART, J Zhou, MC DELLER, .... . Cited by .\n- Crystal Structure of a Cleaved, Stabilized HIV-1 Env Trimer in Complex with a Glycan-Dependent Broadly Neutralizing Antibody at 4.7 angstrom (2013). J Julien, A Cupo, D Sok, RL Stanfield, D Lyumkis, MC Deller, DR Burton, .... AIDS RESEARCH AND HUMAN RETROVIRUSES 29 (11), A165-A166, 2013. Cited by .\n- Structure of the gamma D-glutamyl-L-diamino acid endopeptidase YkfC from Bacillus cereus in complex with L-Ala--D-Glu: Insights into substrate recognition by NlpC/P60 cysteine\ufffd\u2026 (2013). Q Xu, P Abdubek, T Astakhova, HL Axelrod, C Bakolitsa, X Cai, D Carlton, .... Acta Cryst. F66: 1354-1364, 2010, 2013. Cited by .\n- Crystal Structure of the First Eubacterial Mre11 Nuclease Novel Features that May Discriminate Substrates During DNA Repair (2012). D Das, D Moiani, HL Axelrod, MD Miller, D McMullan, KK Jin, P Abdubek, .... J. Mol. Biol 397: 647, 2010, 2012. Cited by .\n- CLONING, EXPRESSION, CHARACTERIZATION, PURIFICATION AND PRELIMINARY CRYSTALLOGRAPHIC STUDIES OF THE HAEMOPOIETIN CYTOKINE ONCOSTATIN-M (1999). MC Deller, K Hudson, J Bravo, JK Heath, EY Jones. NUCLEIC ACIDS SYMPOSIUM SERIES 40, 10-10, 1999. Cited by .\n- Structural studies of cytokines and cytokine receptors (1999). MC Deller. University of Oxford. 1999., 1999. Cited by .\n- 5Chapter (). M Levisson, GW Han, MC Deller, P Biely, S Hendriks, Q Xu, L Ten Eyck, .... Biochemical and Structural Analysis of Thermostable Esterases, 65, 0. Cited by .", "# marcdeller.com: My Projects\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n## Systems Builder: Projects That Define My Work\n\nMy projects span [precision oncology](https://marcdeller.com/my-work-driving-innovation-in-precision-oncology-through-structure-based-drug-design/), virology, immunology, and [structural genomics](https://marcdeller.com/top96-crystal-screen/), as well as mentoring a Stanford-led student program on [novel antibiotics for resistant bacteria](https://news.stanford.edu/stories/2017/04/undergrads-win-prize-work-combating-antibacterial-resistance). The thread connecting all of it is the same: use protein structure as the starting point and let biology dictate everything that follows. Today, my active focus splits between engineered enzyme therapeutics at [Elora Therapeutics](https://eloratherapeutics.com) for microplastic degradation in human tissues and advising an early-stage AI drug-design venture, [DeepCovalent](https://deepcovalent.com), working on AI-generated small molecules.\n\n* **Precision Oncology:** Advanced treatments through structural biology and structure-based drug design targeting cancer, immunological disorders, virology, and antibiotic resistance\n* **Platform Innovation:** Built automation and lab infrastructure spanning Cryo-EM, X-ray crystallography, crystal harvesting robotics, and scalable discovery workflows.\n* **Technological Prowess**: Championed automation in structural biology workflows and adopted CryoEM/X-ray crystallography technologies to enhance drug design processes\n* **Translational Impact:** Bridging foundational structural biology and real-world therapeutic outcomes, from fighting microplastic accumulation in humans to advising AI-driven drug design startups and enabling next-generation discovery pipelines.\n\nTop Projects\n\n### [Top96 Crystal Screen](https://marcdeller.com/top96-crystal-screen/)\n\n[Top96 Crystal Screen](https://marcdeller.com/top96-crystal-screen/)\n\n### [Universal Mounting Robot (UMR)](https://marcdeller.com/universal-mounting-robot-umr/)\n\n[Universal Mounting Robot (UMR)](https://marcdeller.com/universal-mounting-robot-umr/)\n\n### [My Work Driving Innovation in Precision Oncology Through Structure-Based Drug Design](https://marcdeller.com/my-work-driving-innovation-in-precision-oncology-through-structure-based-drug-design/)\n\n[My Work Driving Innovation in Precision Oncology Through Structure-Based Drug Design](https://marcdeller.com/my-work-driving-innovation-in-precision-oncology-through-structure-based-drug-design/)\n\n### [Building a Comprehensive Protein Production, X-ray Crystallography, and Cryo-EM Laboratory from Scratch](https://marcdeller.com/building-a-comprehensive-protein-production-x-ray-crystallography-and-cryo-em-laboratory-from-scratch/)\n\n[Building a Comprehensive Protein Production, X-ray Crystallography, and Cryo-EM Laboratory from Scratch](https://marcdeller.com/building-a-comprehensive-protein-production-x-ray-crystallography-and-cryo-em-laboratory-from-scratch/)\n\n### [JAK2-V617F and the Evolution of Targeted Therapy: Precision Medicine Transforms Myeloproliferative Neoplasm Treatment](https://marcdeller.com/jak2-v617f-and-the-evolution-of-targeted-therapy-precision-medicine-transforms-myeloproliferative-neoplasm-treatment/)\n\n[JAK2-V617F and the Evolution of Targeted Therapy: Precision Medicine Transforms Myeloproliferative Neoplasm Treatment](https://marcdeller.com/jak2-v617f-and-the-evolution-of-targeted-therapy-precision-medicine-transforms-myeloproliferative-neoplasm-treatment/)\n\nAI\n\n### [BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline](https://marcdeller.com/boltzmaker-vibe-coding-a-campaign-scale-boltz-2-pipeline/)\n\nAI\n\n### [ChemSage: Vibe-Coding a Chemically Aware Drug-Discovery LLM on Local Hardware](https://marcdeller.com/chemsage/)\n\nProjects\n\n### [JAK2-V617F and the Evolution of Targeted Therapy: Precision Medicine Transforms Myeloproliferative Neoplasm Treatment](https://marcdeller.com/jak2-v617f-and-the-evolution-of-targeted-therapy-precision-medicine-transforms-myeloproliferative-neoplasm-treatment/)\n\nBlog\n\n### [Building a Comprehensive Protein Production, X-ray Crystallography, and Cryo-EM Laboratory from Scratch](https://marcdeller.com/building-a-comprehensive-protein-production-x-ray-crystallography-and-cryo-em-laboratory-from-scratch/)\n\nProjects\n\n### [My Work Driving Innovation in Precision Oncology Through Structure-Based Drug Design](https://marcdeller.com/my-work-driving-innovation-in-precision-oncology-through-structure-based-drug-design/)\n\nProjects\n\n### [Universal Mounting Robot (UMR)](https://marcdeller.com/universal-mounting-robot-umr/)\n\nProjects\n\n### [Top96 Crystal Screen](https://marcdeller.com/top96-crystal-screen/)", "# Microplastics Blaster \u2013 Deploy enzymes, destroy microplastics, save the body\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nElora Therapeutics is developing the first FDA-targeted enzyme therapeutic \u2014 based on engineered PETase and MHETase \u2014 designed to actively depolymerise PET microplastics accumulated in human biological matrices. MicroPlastic Blaster is a public-awareness and brand engagement tool built to communicate the scale of the microplastic health problem in an accessible, interactive format.\n\n**Website:** [eloratherapeutics.com](https://www.eloratherapeutics.com)\n\n---\n\n## \ud83d\udcc4 Licence\n\n\u00a9 Elora Therapeutics. All rights reserved. Assets (logos, avatars, audio) are proprietary. The game logic may be adapted for educational or awareness purposes with attribution.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/microplastic-blaster](https://github.com/bellcheddar/microplastic-blaster) |", "# GitHub README: TopPDBLX\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/TopPDBLX, 2026\n\n```sql\n-- which reagents crystallise the widest range of protein families,\n-- without lysozyme and trypsin drowning the answer\nSELECT name_canonical,\n       count(DISTINCT cluster_30) AS families,\n       count(*) AS uses\nFROM condition_components\nWHERE name_canonical IS NOT NULL\nGROUP BY name_canonical\nORDER BY families DESC\nLIMIT 10;\n```\n\n## \ud83e\uddf1 Stack\n\n| Layer | Choice |\n|---|---|\n| Pipeline | Python 3.14, polars, pyarrow, duckdb, pydantic |\n| Structure parsing | gemmi (mmCIF), used only for fidelity checking |\n| Text | `regex`, for variable-width lookbehind and better unicode handling |\n| Clustering | MMseqs2, shelled out |\n| Interfaces | Single-file HTML and vanilla JavaScript, no build step |\n\n## \u26a0\ufe0f Known limitations\n\nThese determine what conclusions the data can support, and are stated in full in [`DATASHEET.md`](DATASHEET.md).\n\n| Limitation | Consequence |\n|---|---|\n| The PDB contains only successes | Supports P(condition given crystallised), says nothing about P(crystallised given sequence) |\n| Frequency reflects screen popularity | A common condition is evidence about what was tried, not about what works best |\n| About 80% of percentage concentrations carry no w/v or v/v marker | Where a unit was inferred it is flagged `unit_inferred`: filter on it before treating a unit as reported fact |\n| Only 2.2% of entries name a cryoprotectant explicitly | `cryo_evidence` separates `explicit` from `inferred`, and four in five are inferences |\n| Depositor errors are reproduced, not corrected | A handful of records state nanomolar concentrations of bulk reagents. The exception is an amount that cannot be true at all (above 8 M equivalent, or above 100%): the reagent is kept, the amount dropped, and the record flagged `implausible_concentration` |\n| Screen matching cannot validate reagent naming | Both sides use the same lexicon, so a systematic naming error moves both identically |\n| Fidelity to the rules is a ceiling, not evidence of beating it | The rule parser's own labels train the residual model, so agreement with `rules_v3` cannot show a gain; only identification and grounding on the residual, and precision/recall against the gold sets, can |\n| The residual parser cannot discover new chemistry | It emits only names it has seen, so a genuinely absent reagent looks the same as a model error. Closing that gap is curation, not modelling |\n| The gold sets are 192 records pooled | Wide confidence intervals, reagent identity only, no concentrations. Makes \"did we miss something\" answerable, not precisely answerable |\n| Recall is the weaker half of the pair | Rules alone: 95.1% precision against 69.5% recall, the parser is far likelier to miss a reagent than invent one. Rules + round 09 narrows that gap to 93.3% precision, 90.3% recall |\n\n## \u2705 To Do", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices\nAuthors: Deller, M.C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\n**Author:** Marc C. Deller. **Institution:** Laboratory of Molecular Biophysics and Wellcome Trust Centre for Human Genetics, University of Oxford. **Supervisor:** Prof. E. Yvonne Jones. **Degree:** D.Phil. thesis, University of Oxford (c. 1999). This part contains Chapter 6 (description of the oncostatin M (OSM) crystal structure, structural comparisons with other haemopoietin cytokines and receptors, structure\u2013function analysis of the site 1, 2 and 3 receptor-binding epitopes, sequence comparisons, implications for receptor binding and the overall summary) followed by the Appendices (reagent recipes and useful formulae).\n\n## Chapter 6 \u2013 Structure Description, Comparison and Functional Implications\n\nThe aims of this chapter include the description of the OSM structure and comparisons with other members of the haemopoietin cytokine family. Detailed analysis of the structure is reported along with implications for receptor binding and species specificity. The crystal structure solved in the spacegroup P212121 is used for analysis (unless stated otherwise), due to the higher quality of experimental data. A comparison of the I222 and P212121 structures is also given. Finally, general structural themes pertinent to the haemopoietin cytokine network are also discussed along with functional dissection of the site 1, 2 and 3 receptor binding epitopes as depicted schematically in Figure 6-0 below.\n\n*Figure 6-0.* Schematic representation of the site 1, 2 and 3 receptor binding site motifs as observed in the gp130 and growth hormone family of ligands. The 4-helix bundle is depicted as a cylinder and the regions responsible for receptor binding are highlighted.\n\n### 6.1 Structure Description\n\nAs predicted from sequence analysis and homology models (Bazan, 1991), OSM conforms to the up-up-down-down 4-helix bundle topology, as observed for other members of the haemopoietin cytokine family whose crystal structures have been determined. Figure 6-1 and Figure 6-2 below show C\u03b1 stereodiagrams and ribbon representations of OSM, respectively.\n\n*Figure 6-1.* Figure showing C\u03b1 stereodiagram of OSM coloured from blue at the N-terminus through to red at the C-terminus. Every tenth residue is numbered. Figure produced using the program MOLSCRIPT (Kraulis, 1991).\n\n*Figure 6-2.* Figure showing orthogonal projections of the ribbon representation of OSM coloured from blue at the N-terminus through to red at the C-terminus. Disulphide bonds are represented as ball and stick objects with the sulphur atoms highlighted by yellow spheres. Transparent dotted section represents loop region as observed in LIF. Position of open circles on helix A and C denotes location of kinked regions. Figure produced using BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994).\n\nOSM is a compact barrel shaped molecule with dimensions of approximately 20 \u00c5 \u00d7 27 \u00c5 \u00d7 56 \u00c5. The basic fold of OSM consists of 4 main helical regions (helix A, residues 10\u201337; helix B, residues 67\u201390; helix C, residues 105\u2013131; helix D, residues 159\u2013185) connected by two long overhand loops (A-B loop, residues 38\u201366; C-D loop, residues 132\u2013158) and one short loop (B-C loop, residues 91\u2013104). The secondary structure assignment and hydrogen bonding pattern is shown below in Figure 6-3.\n\nThe overall arrangement of helices in this fold may be considered as two pairs of anti-parallel helices that pack together, with helices A and D forming one pair and helices B and C forming the other. Helices B and C pack in a classical fashion, with a skew angle between the helical axis of around 20\u00b0, as observed in other members of the long chain 4-\u03b1-helix cytokines. Helix C is the longest helix consisting of 26 residues that run the length of the molecule with helix B being the shortest helix of only 23 residues in length. Figure 6-4 (b) below shows that helix C exhibits a slight kink towards the N-terminus of its length, between residues Pro 116 and Gln 112, due to disruption of the classic hydrogen bonding pattern observed for \u03b1-helices. This disruption is in part due to substitute hydrogen bonds, which are formed with two tightly bound water molecules.\n\nHelices A and D pack in similar fashion although a more pronounced kink is observed in helix A (see Figure 6-4 (a)). This serves to allow closer packing of the A:D and B:C helical pairs, so forming a tighter solvent inaccessible core. The kink in helix A is caused by a break in the helical conformation, as residues Gln 25 and Leu 30 adopt a hydrogen bonded turn conformation, and also four tightly bound water molecules disrupt the classic hydrogen bonding pattern. Residues Thr 27 to Ile 37 of helix A adopt a 3-10 helix conformation.\n\n*Figure 6-3.* Figure showing secondary structure and hydrogen bonding pattern of P212121 OSM. Figure produced using DSSP (Kabsch and Sander, 1983) and HERA (Hutchinson and Thornton, 1990). Residues represented as open ovals are in a \u03b1-helical conformation and residues represented as shaded ovals are in a 3-10 helical conformation. All other residues are in an extended loop conformation, except for residues 1\u20133 at the N-terminus and residues 135\u2013155 of the C-D loop for which no appreciable density was observed. Arrows represent hydrogen bonding pattern.\n\n*Figure 6-4.* Figure showing kinked regions of a) helix A and b) helix C. Refined 2.2 \u00c5 2|Fo|\u2212|Fc| electron density map is shown contoured at 1\u03c3. Main chain hydrogen bonds are shown as yellow dashed lines and hydrogen bonds to waters are shown as red dashed lines. Bound water molecules as shown as sky blue spheres. Map was produced from MAD data using SHARP (de La Fortelle and Bricogne, 1997) followed by solvent flattening with SOLOMON (Abrahams and Leslie, 1996) and refinement with CNS (Brunger et al., 1998). Figure was produced using BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994).", "# GitHub README: PANTS\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/PANTS, 2026\n\n1. Positives number in the low hundreds, and most carry family annotation rather than measured PET activity. Every score is an extrapolation from a small, biased sample.\n2. Published activity data is not harmonised across assay formats. Absolute rate predictions should not be trusted.\n3. Crystalline PET degradation at 37 \u00b0C by any known enzyme is slow. PANTS ranks relative promise, not therapeutic viability.\n4. Predicted structures are predictions. Cleft geometry from ESMFold on a metagenomic sequence with no close homologue carries real uncertainty.\n5. Nothing here addresses delivery, immunogenicity, biodistribution, or what happens to liberated TPA and EG in vivo. Those decide whether any of this is a therapy.\n6. Metagenomic candidates may come from unculturable organisms, may not express in a standard host, and may be fragments or misassemblies.\n7. **A learned head does not beat retrieval on measured labels.** On the hardest contrast the paired lead over nearest-neighbour retrieval is +0.159 [\u22120.263, +0.390], spanning zero, and eighteen times the language-model parameters changed nothing. The shipped ranking is therefore retrieval, and the earlier AUC 0.976 was measured against an easier negative class.\n8. **Ranking PET activity within the polyesterase family is the question that matters, and it is not solved.** Against 150 measured-inactive polyesterases the head reaches 0.582 \u00b1 0.067 and active-site geometry reaches 0.398 \u2014 below chance. The determinants are lineage-specific: the same model fitted inside different families points in unrelated or opposite directions, so a model trained across families learns something that reverses on a family it has not seen.\n9. Old note, kept: the composition baseline was 0.778 on the earlier annotation-heavy positive set. Until a model clears that as well as the E-value baseline, no claim of learned discrimination is supported.\n10. Of 854 positives, 341 carry a measurement. The rest are automatic EC annotation or family membership, so any head trained today is trained mostly on predicted labels.\n9. Everything of interest is packed tightly in embedding space (characterised PET enzymes sit at cosine 0.96 or above to each other, candidates at a median 0.931 to their nearest known enzyme). The head discriminates small differences inside a dense cluster, not well-separated groups.\n\n## \ud83d\udcda Data sources\n\n| Source | Use |\n|---|---|\n| UniProt / UniRef | Reference sequence space, taxonomy, evidence level, signal peptides |\n| ESTHER | \u03b1/\u03b2-hydrolase family assignment, hard negatives, near misses |\n| PAZy | Characterised plastic-degrading enzymes: the measured-activity positives |\n| PDB | Experimental structures and ground-truth geometry |\n| MGnify, JGI IMG/M, OceanDNA, Tara Oceans | Metagenomic assemblies for mining |\n| Meltome Atlas, FireProtDB | Thermostability transfer learning |\n| AlphaFold DB | Precomputed structures where a UniProt match exists |\n\n## \u2705 To Do\n\nRoadmap for PANTS, roughly in dependency order. Suggestions welcome.", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\nIn some embodiments, an anti-mutCALR antibody comprises a heavy chain variable region with the C-terminal lysine removed. In some embodiments, an anti-mutCALR antibody comprises a heavy chain variable region comprising an amino acid sequence that has the three VH CDRs of any anti-mutCALR clone disclosed herein and which has at least 75%, 80%, 85%, 90%, 95%, 96%, 97%, 98%, 99%, or 100% identity to any one of the VH sequences set forth in Tables 4-5 and a light chain variable region comprising an amino acid sequence that has the three VL CDRs of any anti-mutCALR clone disclosed herein and which has at least 75%, 80%, 85%, 90%, 95%, 96%, 97%, 98%, 99%, or 100% identity to any one of the VL sequences set forth in Tables 4-5.\n\n*(Tables 1-5 of the specification \u2014 the six CDR sequences for each of the 54 unique parental clones (Table 1) and each of the 161 affinity-matured/mutant clones 55-215 (Table 2), together with the full VH, VL, heavy-chain and light-chain sequences for the parental clones (Table 4) and the mutant clones (Table 5) \u2014 occupy approximately eighty pages of the printed specification and are set out there as sequence tables keyed to SEQ ID NOs:1-318. They are not reproduced here.)*\n\n### Antibody Formats\n\nAntibody fragments can be obtained by treating the whole antibody with an enzyme such as papain, pepsin, or plasmin or the FabRICATOR (IdeS) recombinant enzyme (Genovis AB) that digests IgG antibodies to produce a homogeneous pool of F(ab')2 and Fc/2 fragments. Papain digestion of whole antibodies produces F(ab)2 or Fab fragments; pepsin digestion of whole antibodies yields F(ab')2 or Fab'; and plasmin digestion of whole antibodies yields Facb fragments. Alternatively, antibody fragments can be produced recombinantly. Antibody fragments can be expressed in and secreted from *E. coli*, thus allowing the facile production of large amounts of these fragments. Antibody fragments can be isolated from antibody phage libraries. Alternatively, Fab'-SH fragments can be directly recovered from *E. coli* and chemically coupled to form F(ab')2 fragments. According to another approach, F(ab') fragments can be isolated directly from recombinant host cell culture. Fab and F(ab')2 fragments with increased in vivo half-life comprising a salvage receptor binding epitope residues are described in U.S. Pat. No. 5,869,046.\n\nIn some instances, the anti-mutCALR antibody is a minibody. Minibodies of anti-mutCALR antibodies include diabodies, single chain (scFv), and single-chain (Fv)2 (sc(Fv)2). A \"diabody\" is a bivalent minibody constructed by gene fusion. Diabodies are dimers composed of two polypeptide chains. The VL and VH domain of each polypeptide chain of the diabody are bound by linkers. The number of amino acid residues that constitute a linker can be between 2 to 12 residues (e.g., 3-10 residues or five or about five residues). The linkers of the polypeptides in a diabody are typically too short to allow the VL and VH to bind to each other. Thus, the VL and VH encoded in the same polypeptide chain cannot form a single-chain variable region fragment, but instead form a dimer with a different single-chain variable region fragment. As a result, a diabody has two antigen-binding sites.\n\nAn scFv is a single-chain polypeptide antibody obtained by linking the VH and VL with a linker. The order of VHs and VLs to be linked is not particularly limited, and they may be arranged in any order. Examples of arrangements include: [VH] linker [VL]; or [VL] linker [VH]. An sc(Fv)2 is a minibody in which two VHs and two VLs are linked by a linker to form a single chain. An sc(Fv)2 can be prepared, for example, by connecting scFvs with a linker.\n\nIn some instances, the anti-mutCALR antibody is a bispecific antibody. Bispecific antibodies are antibodies that have binding specificities for at least two different epitopes. Exemplary bispecific antibodies may bind to two different epitopes of the mutCALR protein. Other such antibodies may combine a mutCALR binding site with a binding site for another antigen. Bispecific antibodies can be prepared as full length antibodies or low molecular weight forms thereof (e.g., F(ab')2 bispecific antibodies, sc(Fv)2 bispecific antibodies, diabody bispecific antibodies).\n\nTraditional production of full length bispecific antibodies is based on the co-expression of two immunoglobulin heavy chain-light chain pairs, where the two chains have different specificities. In a different approach, antibody variable domains with the desired binding specificities are fused to immunoglobulin constant domain sequences. According to another approach described in U.S. Pat. No. 5,731,168, the interface between a pair of antibody molecules can be engineered to maximize the percentage of heterodimers that are recovered from recombinant cell culture (\"knobs-into-holes\"): one or more small amino acid side chains from the interface of the first antibody molecule are replaced with larger side chains (e.g., tyrosine or tryptophan), and compensatory \"cavities\" of identical or similar size to the large side chain(s) are created on the interface of the second antibody molecule by replacing large amino acid side chains with smaller ones (e.g., alanine or threonine).\n\nBispecific antibodies include cross-linked or \"heteroconjugate\" antibodies. The \"diabody\" technology provides an alternative mechanism for making bispecific antibody fragments: the fragments comprise a VH connected to a VL by a linker which is too short to allow pairing between the two domains on the same chain, so the VH and VL domains of one fragment are forced to pair with the complementary VL and VH domains of another fragment, thereby forming two antigen-binding sites.", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\nRegulatory inactivation of DnaA is dependent on Hda (homologous to DnaA), a protein homologous to the AAA+ (ATPases associated with diverse cellular activities) ATPase region of the replication initiator DnaA. When bound to the sliding clamp loaded onto duplex DNA, Hda can stimulate the transformation of active DnaA-ATP into inactive DnaA-ADP. The crystal structure of Hda from Shewanella amazonensis SB2B at 1.75 A resolution reveals that Hda resembles typical AAA+ ATPases. The arrangement of the two subdomains in Hda (residues 1-174 and 175-241) differs dramatically from that of DnaA. A CDP molecule anchors the Hda domains in a conformation that promotes dimer formation. The Hda dimer adopts a novel oligomeric assembly for AAA+ proteins in which the arginine finger, crucial for ATP hydrolysis, is fully exposed and available to hydrolyze DnaA-ATP through a typical AAA+ type of mechanism. The sliding clamp binding motifs at the N-terminus of each Hda monomer are partially buried and combine to form an antiparallel beta-sheet at the dimer interface. The inaccessibility of the clamp binding motifs in the CDP-bound structure of Hda suggests that conformational changes are required for Hda to form a functional complex with the clamp. Thus, the CDP-bound Hda dimer likely represents an inactive form of Hda.\n\n## Crystal structure of a novel Sm-like protein of putative cyanophage origin at 2.60 A resolution (2009)\n\n**Authors:** Das D, Kozbial P, Axelrod HL, Miller MD, McMullan D, Krishna SS, Abdubek P, Acosta C, Astakhova T, Burra P, Carlton D, Chen C, Chiu HJ, Clayton T, Deller MC, Duan L, Elias Y, Elsliger MA, Ernst D, Farr C, Feuerhelm J, Grzechnik A, Grzechnik SK, Hale J, Han GW, Jaroszewski L, Jin KK, Johnson HA, Klock HE, Knuth MW, Kumar A, Marciano D, Morse AT, Murphy KD, Nigoghossian E, Nopakun A, Okach L, Oommachen S, Paulsen J, Puckett C, Reyes R, Rife CL, Sefcovic N, Sudek S, Tien H, Trame C, Trout CV, van den Bedem H, Weekes D, White A, Xu Q, Hodgson KO, Wooley J, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Proteins (2009). DOI: 10.1002/prot.22360. PMID: 19173316.\n\nECX21941 represents a very large family (over 600 members) of novel, ocean metagenome-specific proteins identified by clustering of the dataset from the Global Ocean Sampling expedition. The crystal structure of ECX21941 reveals unexpected similarity to Sm/LSm proteins, which are important RNA-binding proteins, despite no detectable sequence similarity. The ECX21941 protein assembles as a homopentamer in solution and in the crystal structure when expressed in Escherichia coli and represents the first pentameric structure for this Sm/LSm family of proteins, although the actual oligomeric form in vivo is currently not known. The genomic neighborhood analysis of ECX21941 and its homologs combined with sequence similarity searches suggest a cyanophage origin for this protein. The specific functions of members of this family are unknown, but our structure analysis of ECX21941 indicates nucleic acid-binding capabilities and suggests a role in RNA and/or DNA processing.\n\n## Crystal structure of histidine phosphotransfer protein ShpA, an essential regulator of stalk biogenesis in Caulobacter crescentus (2009)\n\n**Authors:** Xu Q, Carlton D, Miller MD, Elsliger MA, Krishna SS, Abdubek P, Astakhova T, Burra P, Chiu HJ, Clayton T, Deller MC, Duan L, Elias Y, Feuerhelm J, Grant JC, Grzechnik A, Grzechnik SK, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Kumar A, Marciano D, McMullan D, Morse AT, Nigoghossian E, Okach L, Oommachen S, Paulsen J, Reyes R, Rife CL, Sefcovic N, Trame C, Trout CV, van den Bedem H, Weekes D, Hodgson KO, Wooley J, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Journal of molecular biology (2009). DOI: 10.1016/j.jmb.2009.05.023. PMID: 19450606.\n\nCell-cycle-regulated stalk biogenesis in Caulobacter crescentus is controlled by a multistep phosphorelay system consisting of the hybrid histidine kinase ShkA, the histidine phosphotransfer (HPt) protein ShpA, and the response regulator TacA. ShpA shuttles phosphoryl groups between ShkA and TacA. When phosphorylated, TacA triggers a downstream transcription cascade for stalk synthesis in an RpoN-dependent manner. The crystal structure of ShpA was determined to 1.52 A resolution. ShpA belongs to a family of monomeric HPt proteins that feature a highly conserved four-helix bundle. The phosphorylatable histidine His56 is located on the surface of the helix bundle and is fully solvent exposed. One end of the four-helix bundle in ShpA is shorter compared with other characterized HPt proteins, whereas the face that potentially interacts with the response regulators is structurally conserved. Similarities of the interaction surface around the phosphorylation site suggest that ShpA is likely to share a common mechanism for molecular recognition and phosphotransfer with yeast phosphotransfer protein YPD1 despite their low overall sequence similarity.\n\n## Crystal structure of the Fic (Filamentation induced by cAMP) family protein SO4266 (gi|24375750) from Shewanella oneidensis MR-1 at 1.6 A resolution (2009)\n\n**Authors:** Das D, Krishna SS, McMullan D, Miller MD, Xu Q, Abdubek P, Acosta C, Astakhova T, Axelrod HL, Burra P, Carlton D, Chiu HJ, Clayton T, Deller MC, Duan L, Elias Y, Elsliger MA, Ernst D, Feuerhelm J, Grzechnik A, Grzechnik SK, Hale J, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Kumar A, Marciano D, Morse AT, Murphy KD, Nigoghossian E, Okach L, Oommachen S, Paulsen J, Reyes R, Rife CL, Sefcovic N, Tien H, Trame CB, Trout CV, van den Bedem H, Weekes D, White A, Hodgson KO, Wooley J, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Proteins (2009). DOI: 10.1002/prot.22338. PMID: 19127588.\n\n*(no abstract available)*\n\n## Structural and functional characterizations of SsgB, a conserved activator of developmental cell division in morphologically complex actinomycetes (2009)", "# GitHub README: CODSWALLOP\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/CODSWALLOP, 2026\n\n- **Ask for anything.** A PDB ID, a PDB entity id, a UniProt accession, a gene name, a Pfam or InterPro accession, a raw sequence (FASTA or bare), or just what the protein is called. Anything genuinely ambiguous returns a disambiguation card rather than a guess: `4HHB` asks whether you meant the alpha or beta globin, `LYZ` asks which of twelve organisms.\n- **One definition of a family.** Whatever you type is resolved to a **seed sequence**, and membership is every PDB polymer entity above a percent-identity threshold to that seed. Every member therefore carries a real identity number to the same reference, so the threshold slider means something exact rather than \"whatever the annotation happened to say\".\n- **The drawer.** An archival shell: a divider rail carrying all ten sections from day one, a family header with a rubber-stamped `FILED` mark, a stat strip, and an index card that flips out from any entry anywhere without ever losing the family context.\n- **The constellation.** Every entry as a node. Once the family has an embedding, position is classical MDS of the pairwise TM-score matrix, so distance on the map is *structural* distance and the 0.5 contour is the conventional same-fold boundary; until then it falls back to a clearly-labelled identity layout. Node size is 1/resolution, colour is method, an amber halo means ligand-bound. Labels name the clusters the matrix falls into at the same cut the Structures heatmap uses. A ringed node holds several entries that share a construct and opens on click; a hollow node has a position inherited rather than measured; a brass chevron means the structure lies off the end of the scale and has been pinned to the edge.\n- **Cross-highlight.** Hovering any row in the entry list pulses that entry's node brass in the map, and hovering a node highlights its row. One state, two representations, and every panel added in later phases binds to the same shared selection state.\n- **Family definition controls.** Percent-identity slider, method chips, and include/exclude toggles for orthologues and for chimeras and fusion constructs. All of it filters client-side and instantly: the family is assembled once at the lowest threshold and every control re-filters the same cached set.\n- **Construct coverage census.** How many of the family's constructs contain each residue of the seed, as a depth profile. On lysozyme it finds residues 1 to 18 covered by 66 of 1,686 constructs, which is the signal peptide that is cleaved and never present in the mature protein. On p53 it finds residues 361 to 393 in 10 of 267.\n- **Full table view.** Sortable, filterable, column-toggleable, with CSV, JSON and BibTeX export. The BibTeX is deduplicated to one record per paper with the PDB IDs collected in a note, which is the tedious part of assembling a family bibliography by hand.\n- **Provenance: what it is called, and who did the work.** Every spelling the family is deposited under, reconciled against UniProt's recommended, alternative, short and gene names, plus the reverse direction that actually breaks a literature search: within hen lysozyme's own family, \"Lysozyme C\" is the deposited description of **seven different accessions**. Beside it, the depositing groups: last author as the group's proxy (labelled as the convention it is), counted per entry rather than per paper, with a deposition timeline. 380 groups over 1975 to 2026 for lysozyme, the largest holding 8.3 % of it.\n- **Conservation on the surface.** ConSurf-style colouring of the loaded structure by family-wide conservation, banded on the family's own quintiles because a family is similar sequences by construction and a fixed scale never uses its variable end. The seed-to-structure frame is *scored on residue identity* rather than assumed, and refused below 60 % agreement, so a construct with an N-terminal tag is not frame-shifted into a plausible-looking lie.\n- **A JSON API other projects can build against.** `/api/v1/family/<slug>` is a few kilobytes with field names that are a promise and a `schema_version` to break if one has to change meaning, rather than the front end's own multi-megabyte payload.\n- **Dark by default, light on request.** A lamp-lit archive theme that honours `prefers-color-scheme` and respects `prefers-reduced-motion`, self-hosted type, and no third-party request on any page load.\n\n## \ud83d\udd2c Worked example: a fold nobody searched for\n\nSearch `1AKI`, hen egg-white lysozyme. It is the most-solved protein in the archive: 1,687 entries, 281 distinct constructs, 26 organisms, and essentially all of it is the same fold solved again. The embedding says so, putting 1,686 of the 1,688 entities in one cloud.\n\nTwo are not in it.\n\nThey are **9J0L** and **9J0M**, cryo-EM structures of hen lysozyme **amyloid fibrils**, and they sit at the far edge with a brass chevron because they are off the end of the scale entirely.\n\n| | 9J0L and 9J0M |\n|---|---|\n| Identity to the seed | **100%** (identical sequence, and the same construct as each other) |\n| TM-score to the other 76 representatives | min 0.15, **median 0.18**, max 0.20 |\n| Pairs above the 0.5 same-fold line | **0 of 76** |\n| Method | cryo-EM at 2.29 \u00c5 and 3.21 \u00c5 |\n\nSame protein, same sequence, a completely different fold: native lysozyme refolded into a cross-\u03b2 fibril. **No sequence-based view of this family can find them**, because there is nothing in the sequence to find. They are 100% identical to the seed, so an identity slider puts them at the top of the list along with everything else, and the identity layout the map falls back to places them in the middle of the crowd. Only a structural distance separates them, and only a family-centric one makes it obvious that it is two entries out of 1,688 rather than a general property of lysozyme.\n\n### What the map is not\n\nTwo limits worth stating, because a scatter plot of points looks like a measurement whatever produced it.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices\nAuthors: Deller, M.C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nClearly, structural studies through the use of X-ray crystallography and NMR techniques will allow finer dissection of the interactions involved. This present work represents one such study in which a further piece of structural information has been added. Further analysis of the higher order complexes of gp130, LIFR and the specific \u03b1-receptor subunits is now required to complete the understanding of signal transduction through the haemopoietin cytokine network; central to this understanding is the site 3 interaction with LIFR, which still remains the elusive target of structural studies.\n\n## Appendices\n\n### 7.1 Appendix I \u2013 Molecular Biology Reagents\n\n#### 7.1.1 Luria-Bertani Medium (LB)\n\n| Component | Amount |\n|---|---|\n| Bacto-tryptone | 10 g |\n| Bacto-yeast extract | 5 g |\n| NaCl | 10 g |\n| dH2O | to 1000 cm\u00b3, followed by autoclaving for 20 min at 15 p.s.i. |\n\n#### 7.1.2 Tris-Acetate (TAE) (\u00d750 Stock)\n\n| Component | Amount |\n|---|---|\n| Tris base | 242 g |\n| Glacial acetic acid | 57.1 ml |\n| 0.5 M EDTA (pH 8.0) | 100 ml |\n\n#### 7.1.3 Tris-Borate (TBE) (\u00d75 Stock)\n\n| Component | Amount |\n|---|---|\n| Tris base | 54 g |\n| Boric acid | 27.5 g |\n| 0.5 M EDTA (pH 8.0) | 20 ml |\n\n#### 7.1.4 LB Plates\n\n| Component | Amount |\n|---|---|\n| Agar | 12 g |\n| Bacto-tryptone | 8 g |\n| Yeast extract | 4 g |\n| Sodium chloride | 8 g |\n| Sodium hydroxide (10 M) | 80 \u03bcl |\n| dH2O | to 1000 cm\u00b3, followed by autoclaving for 20 min at 15 p.s.i. |\n\n#### 7.1.5 Transformation Media\n\nLB media containing: sucrose 20 mM, MgSO4 10 mM, MgCl2 10 mM.\n\n#### 7.1.6 DNA Loading Buffer Type III (6\u00d7)\n\nBromophenol blue 0.25 %, xylene cyanol FF 0.25 %, glycerol 30 %. Store at 4 \u00b0C.\n\n### 7.2 Appendix II \u2013 Protein Expression and Purification Reagents\n\n#### 7.2.1 SOB Medium\n\n| Component | Amount |\n|---|---|\n| Bacto-tryptone | 20 g |\n| Bacto-yeast extract | 5 g |\n| NaCl | 0.5 g |\n| 250 mM KCl (pH 7) | 10 ml |\n| dH2O | up to 1000 cm\u00b3, followed by autoclave for 20 min at 15 p.s.i. |\n| 2 M MgCl2 (before use) | 5 ml |\n\n#### 7.2.2 SOC Medium\n\n| Component | Amount |\n|---|---|\n| Bacto-tryptone | 20 g |\n| Bacto-yeast extract | 5 g |\n| NaCl | 0.5 g |\n| 250 mM KCl (pH 7) | 10 ml |\n| dH2O | up to 1000 cm\u00b3, followed by autoclave for 20 min at 15 p.s.i. |\n| 2 M MgCl2 (before use, sterile) | 5 ml |\n| 1 M glucose (before use) | 20 ml |\n\n#### 7.2.3 TNED Buffer\n\nTris-base (pH 8.0) 50 mM, NaCl 150 mM, EDTA 10 mM, DTT 1 mM.\n\n#### 7.2.4 Sodium Citrate Buffer\n\nCitric acid (0.1 M) 46.5 ml, sodium citrate (0.1 M) 3.5 ml.\n\n#### 7.2.5 MTPBS\n\nNaCl 150 mM, Na2HPO4 16 mM, NaH2PO4 4 mM, pH 7.3.\n\n#### 7.2.6 Gel Filtration Buffer\n\nTris-HCl 50 mM, NaCl 150 mM.\n\n#### 7.2.7 Reduced Glutathione\n\nReduced glutathione 10 mM, Tris-HCl 20 mM, pH 8.0.\n\n#### 7.2.8 G7 PNGase F Buffer\n\nSodium phosphate 0.05 M, NP-40 (v/v) 10 %.\n\n#### 7.2.9 PNGase F Column Buffer\n\nNaCl 200 mM, Tris-HCl 20 mM, mercaptoethanol 10 mM, pH 8.0.\n\n#### 7.2.10 PNGase F Elution Buffer\n\nNaCl 200 mM, Tris-HCl 20 mM, mercaptoethanol 10 mM, maltose 10 mM, pH 8.0.\n\n#### 7.2.11 Citric Acid Phosphate Buffer\n\nCitric acid 0.035 M, disodium phosphate 0.067 M, pH 5.0.\n\n#### 7.2.12 Phosphate Buffered Saline (PBS)\n\n| Component | Amount |\n|---|---|\n| Sodium chloride | 8 g |\n| Potassium chloride | 0.2 g |\n| Disodium hydrogen phosphate | 1.44 g |\n| Potassium dihydrogen phosphate | 0.24 g |\n| dH2O | up to 1000 ml, pH 7.4 |\n\n#### 7.2.13 Acrylamide/Bis (30 % T, 2.67 % C)\n\nAcrylamide 87.6 g, N,N'-bis-methylene-acrylamide 2.4 g, dH2O up to 300 ml; filter (0.45 \u03bcm) and store at 4 \u00b0C in the dark.\n\n#### 7.2.14 SDS Reducing Buffer (\u00d71 stock)\n\n| Component | Amount |\n|---|---|\n| dH2O | 3.5 ml |\n| 0.5 M Tris-HCl (pH 6.8) | 1.3 ml |\n| Glycerol | 6.0 ml |\n| 10 % (w/v) SDS | 3.5 ml |\n| \u03b2-mercaptoethanol | 0.5 ml |\n| 1 % (w/v) BPB | 0.5 ml |\n\n#### 7.2.15 SDS Running Buffer (\u00d75 stock)\n\nTris-base 9 g, glycine 14 g, SDS 3 g, dH2O up to 600 ml.\n\n#### 7.2.16 Separating Gel (12 %)\n\n| Component | Amount |\n|---|---|\n| dH2O | 3.35 ml |\n| 1.5 M Tris-HCl (pH 8.8) | 2.50 ml |\n| 10 % (w/v) SDS | 100 \u03bcl |\n| 30 % Acrylamide/Bis | 4.0 ml |\n| 10 % (w/v) APS | 50 \u03bcl |\n| TEMED | 5 \u03bcl |\n\n#### 7.2.17 Stacking Gel (4 %)\n\n| Component | Amount |\n|---|---|\n| dH2O | 6.1 ml |\n| 0.5 M Tris-HCl (pH 6.8) | 2.5 ml |\n| 10 % (w/v) SDS | 100 \u03bcl |\n| 30 % Acrylamide/Bis | 1.33 ml |\n| 10 % (w/v) APS | 50 \u03bcl |\n| TEMED | 10 \u03bcl |\n\n#### 7.2.18 Coomassie Destain\n\nMethanol 400 ml, acetic acid 100 ml, dH2O up to 1000 ml.\n\n#### 7.2.19 Coomassie Stain\n\nCoomassie Blue R-250 1 g, methanol 400 ml, acetic acid 100 ml, dH2O up to 1000 ml.\n\n#### 7.2.20 Protein Loading Buffer (SDS Reducing)\n\n| Component | Amount |\n|---|---|\n| dH2O | 4.0 ml |\n| 0.5 M Tris-HCl (pH 6.8) | 1.0 ml |\n| Glycerol | 0.8 ml |\n| \u03b2-mercaptoethanol | 0.4 ml |\n| 10 % (w/v) SDS | 1.6 ml |\n| 0.05 % (w/v) BPB | 0.2 ml |\n\n#### 7.2.21 Western Blot Transfer Buffer\n\nTris-base 3.03 g, glycine 14.4 g, methanol 100 ml, dH2O to 1000 ml.\n\n### 7.3 Appendix III \u2013 Eukaryotic Expression Reagents\n\n#### 7.3.1 DMEM-S for Culture/Transfection of 293T Cells\n\nDMEM supplemented with: FCS 10 %, penicillin 50 IU/ml, streptomycin 50 \u03bcg/ml, glutamine 2 mM, pyruvate 1 mM, glucose 4500 mg/ml.\n\n#### 7.3.2 SF900 Media for Culture of Drosophila S2 Cells\n\nSF900 media supplemented with: FCS 5 %, penicillin 50 IU/ml, streptomycin 50 \u03bcg/ml, hygromycin-B 300 \u03bcg/ml.\n\n#### 7.3.3 GMEM-S for Culture/Transfection of CHO Cells\n\nGMEM-S supplemented with: FCS 10 %, penicillin 50 IU/ml, streptomycin 50 \u03bcg/ml, L-methionine sulphoximine 50 \u03bcM.\n\n#### 7.3.4 HBBS (10\u00d7)\n\nNaCl (w/v) 1.60 %, Hepes (w/v) 1.20 %, Na2HPO4 (w/v) 0.04 %.\n\n### 7.4 Appendix IV \u2013 Useful Formulae\n\n#### 7.4.1 Spectrophotometric Coefficient for Calculating Protein Concentration\n\n[Protein] (mg/ml) = A280 \u00d7 Mr / ((1331 \u00d7 Y) + (5650 \u00d7 W))", "# Molecular Architecture of GLP-1 Receptor Drug Binding Reveals Dynamic Activation Mechanisms\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Key Concept | Description | Key Reference |\n| --- | --- | --- |\n| **Deep Binding Pocket Residues** | Y152 (TM1), K197 (TM2), Q234/V237 (TM3), W306 (TM5), R310 (TM5), E364/H363 (TM6), L384 (TM7), Y241 (TM3) form the orthosteric pocket. His1 of peptides makes critical contacts with E364, R310, W306, and Y241. | [Dynamics of GLP-1R peptide agonist engagement](https://www.nature.com/articles/s41467-021-27760-0) |\n| **N-terminal Peptide Anchoring** | Peptide N-terminus (His1, Glu9) anchors deep in TMD through hydrogen bonds and hydrophobic interactions. GLP-1 forms more persistent contacts than oxyntomodulin or exendin-4, particularly at the base of the binding cavity (Y152, R190). | [Structural analysis of dual agonism at GLP-1R](https://www.pnas.org/doi/10.1073/pnas.2303696120) |\n| **Extracellular Loop Interactions** | ECL2 (residues around N302, T298) stabilizes peptide binding and receptor activation. ECL1 undergoes conformational changes during activation. ECL3 interacts with ECD in inactive state but reorganizes upon agonist binding. | [Second Extracellular Loop of GLP-1R](https://pmc.ncbi.nlm.nih.gov/articles/PMC3281730/) |\n| **TM6 Conformational Changes** | Outward movement of TM6 intracellular region (~11.9 \u00c5 at R348) creates G protein binding cavity. TM6 movement precedes or occurs simultaneously with Gs engagement. Class B GPCRs show slower TM6 relaxation than class A GPCRs. | [Ligand-free GLP-1R in complex with Gs](https://www.nature.com/articles/s41421-024-00649-0) |\n| **ECD Role in Activation** | ECD acts as \u201caffinity trap\u201d for peptide C-terminus and as \u201cintrinsic agonist\u201d for activation. ECD-TMD interaction through ECLs is required for activation. ECD undergoes conformational change from \u201cclosed\u201d to \u201cstand-up\u201d position. | [Intrinsic agonist mechanism for GLP-1R](https://www.nature.com/articles/celldisc201642) |\n| **Peptide Binding Dynamics** | Differential peptide dynamics correlate with signaling efficacy. Transient vs. persistent interactions at binding pocket base affect G protein activation kinetics. Biased agonists show distinct engagement patterns. | [Dynamics of GLP-1R peptide agonist engagement](https://www.nature.com/articles/s41467-021-27760-0) |", "# Building Proteins Brick by Brick: How AI Is Assembling Novel Bio-Machines\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* AI models like ProDomino are being developed to predict how to successfully combine different protein domains into new, functional proteins.\n* This approach moves beyond modifying existing proteins to enabling the *de novo* design of complex, multi-domain \u201cbio-machines\u201d.\n* This technology has wide-ranging applications in biotechnology and medicine, from creating novel therapeutics to advanced biosensors.\n\n1. **Artificial protein engineering advances through molecular level assembly of individual parts:**\n   <https://phys.org/news/2025-08-artificial-protein-advances-molecular.html>\n2. **Game-Changing AI Tool Rewrites the Rules of Protein Engineering:**\n   <https://scitechdaily.com/game-changing-ai-tool-rewrites-the-rules-of-protein-engineering/>\n\nNature is the ultimate innovator, constantly creating new proteins by shuffling and combining existing functional units, or domains. For years, scientists have dreamed of mimicking this process to build bespoke proteins\u2014custom bio-machines designed for specific tasks. The challenge has been immense. Simply stitching domains together often results in misfolded, non-functional blobs. Predicting how different components will interact to form a stable, working whole was largely a matter of frustrating trial and error.\n\nA new AI model called ProDomino, developed at Heidelberg University, is finally cracking the code of modular protein design. This open-source tool functions like an expert assembly manual for proteins. By training on vast datasets of known protein structures, the AI learns the complex rules that govern how domains fit and work together. It can accurately forecast how to connect individual parts at the molecular level to engineer a functional, and even adjustable, new protein. This shifts protein engineering from modifying existing templates to true *de novo* construction.\n\nThe possibilities unlocked by this approach are vast. Scientists can now design multi-domain proteins that act as sophisticated biosensors, which change color or emit light in the presence of a specific molecule. They can build novel enzymes with customized catalytic activities for green chemistry or create multi-functional therapeutic proteins that can, for instance, bind to a cancer cell and simultaneously deliver a toxic payload. This is more than just an incremental improvement; it\u2019s a new paradigm in protein engineering. By learning nature\u2019s assembly language, AI is giving us the power to build the next generation of biological machines.", "# Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\nAuthors: Zhuo, J.; Li, Y.-L.; Qian, D.-Q.; Burns, D.M.; Mei, S.; Rafalski, M.; Xu, M.; Cao, G.; Pan, Y.; Jia, Z.; Jalluri, R.; Epling, L.B.; Fenalti, G.; Deller, M.C.; Procak, J.; Covington, M.; He, X.; Collins, R.; Stubbs, M.; Burke, K.; Oliver, J.; Margulis, A.; Boer, J.; Wynn, R.; Scherle, P.; Diamond, S.; Newton, R.; Zhang, Y.; Metcalf, B.; Yao, W.\nSource: Journal of Medicinal Chemistry (Drug Annotation), 2026\n\n**Protein Expression, Crystallization, and Structure Determination.** The JAK1 JH1 DNA construct used for crystallization comprises residues 866 to 1154 and includes a noncleavable C-terminal His-tag in a baculovirus expression system. In brief, the protein was expressed in Sf9 cells using a baculovirus expression system and purified using Ni affinity chromatography followed by size exclusion chromatography in a buffer containing 20 mM Tris/HCl, 200 mM NaCl, 0.5 mM EDTA, 3 mM DTT at pH 8.9. Concentrated JAK1 JH1 samples were concentrated to 10 mg/mL and incubated with 1.5 mM (5-fold molar excess) of povorcitinib (150 mM in DMSO) prior to crystallization experiments. Crystals of JAK1 JH1 in complex with povorcitinib were obtained using sitting-drop vapor diffusion and a crystallization condition containing 0.1 M BIS-TRIS-Propane, pH 8.5, 20% [w/v] PEG 3350, 5 mM K2\u2212H-Phosphate, 5 mM K\u2212H2-Phosphate, 5 mM Na2\u2212H-Phosphate, 5 mM Na\u2212H2-Phosphate) equilibrated at 4 \u00b0C. Prior to data collection, JAK1 crystals were cryoprotected using 25% ethylene glycol. Crystallographic data was collected at the European synchrotron radiation source (Grenoble, FR, beamline ID23.1), integrated and scaled by autoPROC and SCALA. Structure refinement proceeded with Refmac, Coot, and Phenix. The final JAK1 JH1 structure bound to povorcitinib was deposited in the PDB with ID 10PI.\n\nThe DNA construct used for crystallization of JAK2 JH1 comprises residues 840 to 1132 and was cloned in frame with an N-terminal 6X His tag and a TEV cleavage site for expression in Sf9 cells using a baculovirus system. The N-terminal His-tag used for purification was not cleaved off. Protein was purified using Ni affinity chromatography followed by size exclusion chromatography using 20 mM Tris/HCl, 250 mM NaCl, 1 mM DTT, 8.5 as the running buffer. JAK2 JH1 at a concentration of 0.8 mg/mL was incubated with 0.22 mM (10-fold molar excess) of povorcitinib (150 mM in DMSO stock solution). This mix was then concentrated to 7.0 mg/mL and then mixed (1:2 ratio) with a reservoir solution (100 mM Molecular Dimensions Buffer System 6 pH 8.0, 1.4 M Na2-Malonate pH 6.0) equilibrated at 20 \u00b0C over 300 \u03bcL of reservoir solution using the sitting-drop vapor diffusion method. Crystals were cryo-protected by transfer to 3.4 M Na2-Malonate pH 6.0 before data collection. Data was collected at Petra III (Hamburg, DE, beamline P13) and integrated and scaled with XDS (Table S1). Structure refinement proceeded with Refmac, Coot, and Phenix. The final structure of JAK2 JH1 bound to povorcitinib was deposited in the PDB with ID 10PJ.\n\n## Associated Content\n\nSupporting Information available free of charge at https://pubs.acs.org/doi/10.1021/acs.jmedchem.5c03753 \u2014 molecular string file for all the final target compounds; enzyme, cellular, and whole blood assays for compounds 4\u201322; in vitro ADME assays compounds 4\u201319, 22; in vivo PK and PD experiments for compounds 9, 14\u201319, 22; full kinome profiling of 22; X-ray crystallographic information and modeling of 22; SMILES for all examples and full ChemDraw structures of all examples (CSV); general experimental information, general synthetic procedures, synthesis, and characterization of compounds 4\u201341 including 1H NMR, 13C NMR, 19F NMR, LCMS and HRMS of final compounds and LCMS spectra for all compounds (PDF); computational modeling of Compound 3 (itacitinib) in JAK1 (PDB); computational modeling of Compound 3 (itacitinib) in JAK2 (PDB).\n\n### Accession Codes\n\nAtomic coordinates for JAK1 kinase (JH1 domain) in complex with povorcitinib (22) and JAK2 kinase (JH1 domain) in complex with povorcitinib (22) were deposited in the PDB with accession codes 10PI and 10PJ, respectively (www.rcsb.org). Authors will release the atomic coordinates upon article publication.\n\n## Author Information\n\n**Corresponding Author:** David M. Burns \u2014 Incyte Research Institute, Wilmington, Delaware 19803, United States; orcid.org/0009-0008-3647-0596; Email: dburns@incyte.com\n\n**Authors:** Jincong Zhuo, Yun-Long Li, Ding-Quan Qian, Song Mei, Maria Rafalski, Meizhong Xu, Ganfeng Cao, Yongchun Pan, Zhongjiang Jia, Ravi Jalluri, Leslie B. Epling, Gustavo Fenalti, Marc C. Deller, Jessica Procak, Maryanne Covington, Xin He, Robert Collins, Matthew Stubbs, Krista Burke, Julian Oliver, Alex Margulis, Jason Boer, Richard Wynn, Peggy Scherle, Sharon Diamond, Robert Newton, Yan Zhang, Brian Metcalf, Wenqing Yao \u2014 all of Incyte Research Institute, Wilmington, Delaware 19803, United States.\n\n### Author Contributions\n\nJincong Zhuo, Yun-Long Li, Ding-Quan Qian, David M. Burns, and Song Mei have contributed equally to this work and are cofirst authors. The manuscript was prepared by David M. Burns with contributions from all listed authors.\n\n### Funding\n\nThe study was funded by Incyte Corporation, Wilmington, DE, USA.\n\n### Notes\n\nThe authors declare the following competing financial interest(s): All authors were full-time employees of Incyte Corporation at the time the above work was conducted.\n\n## Acknowledgments\n\nThe authors thank Sean Bowen, Mei Li, Laurine Galya, and James Hall for analytical support. Boshan Liao, Jun Li, Todd Farrand, Margaret Favata, Karen Gallagher, Cindy Marando, and Chryssa Kanellopoulou for biological support and guidance. Qun Li, Chunhong He, Colin Zhang, Hai Fen, Taisheng Huang, Chubiao Xue, Peng He, Jim Rodgers, and Stacey Shepard for chemistry support. Editorial assistance was provided by ICON (North Wales, PA, USA) and funded by Incyte Corporation. We acknowledge the European Synchrotron Radiation Facility (ESRF) for provision of synchrotron radiation facilities at ID23-1, as well as beamline P-13 at EMBL Hamburg at the PETRA III storage ring (DESY, Hamburg, Germany) for X-ray data collection.\n\n## Abbreviations\n\nADME, absorption, distribution, metabolism, and excretion; AIA, adjuvant-induced arthritis; CSF, colony-stimulating factor; cyno, cynomolgus; DBU, 1,8-Diazabicyclo[5.4.0]undec-7-ene; FDA, US Food and Drug Administration; GM-CSF, granulocyte-macrophage colony-stimulating factor; HATU, hexafluorophosphate azabenzotriazole tetramethyl uronium; IC50, half-maximal inhibitory concentration; IFN, interferon; IL, interleukin; IMHB, intramolecular hydrogen bonding; IV, intravenous; JAK, Janus kinase; LC/MS, liquid chromatography/mass spectrometry; MPN, myeloproliferative neoplasm; PK, pharmacokinetic; PO, oral; STAT, signal transducer and activator of transcription; TEMPO, 2,2,6,6-tetramethyl-1-piperidinyloxy; TFA, trifluoroacetic acid.\n\n## References\n\n*(reference list omitted)*", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nOne or more constituent atoms of the compounds presented herein can be replaced or substituted with isotopes of the atoms in natural or non-natural abundance. In some embodiments, the compound includes at least one deuterium atom. In some embodiments, the compound includes two or more deuterium atoms. In some embodiments, the compound includes 1-2, 1-3, 1-4, 1-5, or 1-6 deuterium atoms. In some embodiments, all of the hydrogen atoms in a compound can be replaced or substituted by deuterium atoms.\n\nSynthetic methods for including isotopes into organic compounds are known in the art (Deuterium Labeling in Organic Chemistry by Alan F. Thomas (New York, N.Y., Appleton-Century-Crofts, 1971; The Renaissance of H/D Exchange by Jens Atzrodt, Volker Derdau, Thorsten Fey and Jochen Zimmermann, Angew. Chem. Int. Ed. 2007, 7744-7765; The Organic Chemistry of Isotopic Labelling by James R. Hanson, Royal Society of Chemistry, 2011). Isotopically labeled compounds can be used in various studies such as NMR spectroscopy, metabolism experiments, and/or assays.\n\nSubstitution with heavier isotopes, such as deuterium, may afford certain therapeutic advantages resulting from greater metabolic stability, for example, increased in vivo half-life or reduced dosage requirements, and hence may be preferred in some circumstances. (see e.g., A. Kerekes et al. J. Med. Chem. 2011, 54, 201-210; R. Xu et al. J. Label Compd. Radiopharm. 2015, 58, 308-312). In particular, substitution at one or more metabolism sites may afford one or more of the therapeutic advantages.\n\nThe radionuclide that is incorporated in the instant radio-labeled compounds will depend on the specific application of that radio-labeled compound. For example, for in vitro WRN protein labeling and competition assays, compounds that incorporate 3H, 14C, 82Br, 125I, 131I, 35S or will generally be most useful. For radio-imaging applications 11C, 18F, 125I, 123I, 124I, 131I, 75Br, 76Br or 77Br can be useful.\n\nIt is understood that a \"radio-labeled\" or \"labeled compound\" is a compound that has incorporated at least one radionuclide. In some embodiments, the radionuclide is selected from the group consisting of 3H, 14C, 125I, 35S and 82Br.\n\nThe present disclosure can further include synthetic methods for incorporating radio-isotopes into compounds of the disclosure. Synthetic methods for incorporating radio-isotopes into organic compounds are well known in the art, and an ordinary skill in the art will readily recognize the methods applicable for the compounds of disclosure.\n\nA labeled compound of the present disclosure can be used in a screening assay to identify and/or evaluate compounds. For example, a newly synthesized or identified compound (i.e., test compound) which is labeled can be evaluated for its ability to bind a WRN protein by monitoring its concentration variation when contacting with the WRN protein, through tracking of the labeling. For example, a test compound (labeled) can be evaluated for its ability to reduce binding of another compound which is known to bind to a WRN protein (i.e., standard compound). Accordingly, the ability of a test compound to compete with the standard compound for binding to the WRN protein directly correlates to its binding affinity. Conversely, in some other screening assays, the standard compound is labeled and test compounds are unlabeled. Accordingly, the concentration of the labeled standard compound is monitored in order to evaluate the competition between the standard compound and the test compound, and the relative binding affinity of the test compound is thus ascertained.\n\n## Examples\n\nCertain embodiments of the present disclosure will be described in greater detail by way of specific examples. The following examples are offered for illustrative purposes and are not intended to limit the present disclosure in any manner. Those of skill in the art will readily recognize a variety of non-critical parameters which can be changed or modified to yield essentially the same results. The compounds of the Examples have been found to inhibit the activity of WRN according to at least one assay described herein.\n\nExperimental procedures for compounds of the present disclosure are provided below. Open Access Preparative LCMS Purification of some of the compounds prepared was performed on Waters mass directed fractionation systems. The basic equipment setup, protocols and control software for the operation of these systems have been described in detail in literature. See, e.g., Blom, \"Two-Pump At Column Dilution Configuration for Preparative LC-MS\", K. Blom, *J. Combi. Chem.*, 2002, 4, 295-301; Blom et al., \"Optimizing Preparative LC-MS Configurations and Methods for Parallel Synthesis Purification\", *J. Combi. Chem.*, 2003, 5, 670-83; and Blom et al., \"Preparative LC-MS Purification: Improved Compound Specific Method Optimization\", *J. Combi. Chem.*, 2004, 6, 874-883.\n\nThe compounds separated were typically subjected to analytical liquid chromatography mass spectrometry (LCMS) for purity analysis under the following conditions: Instrument=Agilent 1100 series, LC/MSD; Column: Waters Sunfire\u2122 C18 5 \u03bcm, 2.1\u00d750 mm, Buffers: mobile phase A: 0.025% TFA in water and mobile phase B: acetonitrile; gradient 2% to 80% B in 3 min with flow rate 2.0 mL/min.", "# Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\nAuthors: Ye, Q.; Shvartsbart, A.; Li, Z.; Gan, P.; Policarpo, R.L.; Qi, C.; Roach, J.J.; Zhu, W.; McCammant, M.S.; Hu, B.; Li, G.; Yin, H.; Carlsen, P.; Hoang, G.; Zhao, L.; Susick, R.; Zhang, F.; Lai, C.-T.; Allali Hassani, A.; Epling, L.B.; Gallion, A.; Kurzeja-Lipinski, K.; Gallagher, K.; Roman, V.; Farren, M.R.; Kong, W.; Deller, M.C.; Zhang, G.; Covington, M.; Diamond, S.; Kim, S.; Yao, W.; Sokolsky, A.; Wang, X.\nSource: Journal of Medicinal Chemistry, 2025\n\nA model of compound 23 based on accumulated crystallographic data is presented in Figure 7. The key interactions from compounds 3 and 5a are conserved, with the [2.1.1]bicyclic amine and cyano ethyl group binding a key water and the carbonyl of the morpholinone engaging with the backbone N\u2212H of Asp12. The methyl pyrazine is not observed to make any interactions with the protein, although a water mediated interaction can not be ruled out. Importantly, the expected ground state conformation of the side chain closely mimics the bound conformation of ethyl amide of 5a.\n\n*Figure 7.* Model of 23 bound to KRAS G12D. Key residues are highlighted. Key interactions are highlighted with yellow dashed lines and distances (\u00c5).\n\n### Chemistry\n\nCompounds were generally prepared from a common intermediate 30, which could be synthesized by the five step sequence outlined in Scheme 1. Thus, iodination of commercially available 24 provided 25, which was transformed to the dioxine and cyclized with ethyl nitroacetate to provide 27. Chlorination was followed by immediate displacement with commercial amine 29 to provide 30 as a solid.\n\n*Scheme 1.* Synthesis of Key Intermediate 32. Reagents and conditions: (a) NIS, DMF, 60 \u00b0C; (b) triphosgene, THF, 60 \u00b0C; (c) ethyl nitroacetate, DIPEA, PhMe; (d) POCl3, DIPEA, 105 \u00b0C; (e) DIPEA, NMP, 60 \u00b0C.\n\nEarly compounds containing the imidazole core (i.e., 3\u22127) were prepared according to Scheme 2. The synthesis of compound 5 is shown for reference, full experimental details can be found in the Supporting Information. Installation of the solvent piece could be performed on crude 30 to provide 31. Reduction of the nitro group was followed by acetic acid-mediated cyclization with the appropriate aldehyde to form the imidazole core. Next, the cyanoethyl fragment was installed via a two step Heck/reduction protocol, providing 33. Finally, Suzuki cross coupling and deprotection provided compound 5 as a mixture of atropisomers, which could be separated by preparatory HPLC. This strategy was utilized for all compound described below.\n\n*Scheme 2.* Synthesis of Compound 5. Reagents and conditions: (a) 3-dimethylaminoazetidine, 60 \u00b0C, NMP; (b) Fe, NH4Cl, THF/MeOH/H2O (1:1:1); (c) N,N-dimethyl-4-oxobutanamide, EtOH/AcOH (5:1), 80 \u00b0C; (d) acrylonitrile, Pd(OAc)2, P(o-tol)3, TEA, DMF, 80 \u00b0C; (e) LiEt3BH, THF, 0 \u00b0C; (f) XPhos Pd G2, K3PO4, dioxane/water, 80 \u00b0C; (g) TFA/DCM (1:1).\n\nThe change to the indole core necessitated an alternative synthetic strategy. After investigating a number of routes, we settled on a Sonogashira/cyclization approach. We originally used gold(I) catalysis to achieve the cyclization, but subsequently discovered that reaction profiles were significantly cleaner when an inorganic base (typically cesium carbonate) was used in the absence of transition metal catalysts. The new critical intermediate thus became 37 (Scheme 3). Thiomethyl was first installed at the 2 position to act as a latent halogen for late stage introduction of substitution at this position. The free N\u2212H was then protected to avoid triazole formation during the Sandmeyer reaction. At this stage of the project, we also implemented a one step reductive Heck protocol to reduce the number of synthetic manipulations. Sandmeyer conversion of the amine to the corresponding iodide then proceeded smoothly on the Boc-protected material. Finally, deprotection of both Boc groups simultaneously and reprotection of the basic amine provided intermediate 37.\n\n*Scheme 3.* Synthesis of Intermediate 37. Reagents and conditions: (a) NaSMe, DCM/MeOH (3:1), 0 \u00b0C to r.t. (98% from 28); (b) Boc-anhydride, TEA, DMAP, THF, r.t.; (c) Fe, NH4Cl, THF/MeOH/H2O (1:1:1), 60 \u00b0C; (d) MeNHCO, Pd(PPh3)4, acrylonitrile, TEA, DMF, 80 \u00b0C; (e) NaNO2, H2SO4, KI, CuI, ACN, \u221215 \u00b0C to r.t.; (f) TFA, DCM, then Boc-anhydride, TEA, THF, r.t.\n\nIntermediate 37 underwent smooth Sonogashira coupling with a wide variety of alkynes under mild conditions; cyclization was typically performed in situ by the addition of cesium carbonate and heating until full consumption of alkyne was observed (Scheme 4). At this stage the solvent and Switch II pocket piece could be installed interchangeably. In the case of 23, Suzuki cross coupling was performed first, followed by solvent installation. Heteroatom-linked solvent pieces were typically installed via oxidation to the sulfone and SNAr displacement, while carbon-linked groups were installed via a Liebeskind\u2013Srogl reaction with the corresponding boronate or stannyl heterocycle. Finally, Suzuki cross coupling and deprotection provided final compounds. The synthesis of 23 is shown in Scheme 4, synthetic procedures for other compounds can be found in the experimental section of the Supporting Information. The synthesis of alkyne 38 is shown in Scheme 5. (R)-butynamine was reacted with 2-(2-chloroethoxy)acetyl chloride, then cyclized under basic conditions.\n\n*Scheme 4.* Synthesis of Compound 23. Reagents and conditions: (a) CuI, (PPh3)4Pd, TEA, DMF, 70 \u00b0C; (b) Cs2CO3, DMF, 70 \u00b0C; (c) (3-chloro-2-methylphenyl)boronic acid, (PPh3)4Pd, K3PO4, dioxane/water (4:1), 80 \u00b0C; (d) (PPh3)4Pd, Cu(3-methylsalicylate), dioxane, 120 \u00b0C; (e) TFA/DCM (1:1).\n\n*Scheme 5.* Synthesis of Alkyne 40. Reagents and conditions: (a) 2-(2-chloroethoxy)acetyl chloride, TEA, THF, r.t.; (b) NaH, THF, r.t.\n\n## Conclusion", "# Top96 Crystal Screen\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-12 on marcdeller.com (categories: project).*\n\n* [Top 96 Conditions](https://marcdeller.com/wp-content/uploads/2025/08/top96.pdf)\n* [Top 96 Poster](https://marcdeller.com/wp-content/uploads/2025/08/top96_poster_ACA_2013_final.pdf)\n* [Commercially available from Rigaku Reagents](https://rigakureagents.com/products/block-jcsg-top96)\n\n* **Large-scale analysis:** millions of crystallization trials revealed clear patterns in successful reagent combinations.\n* **The Top96 screen:** distills the 384-condition JCSG Core Suite into the 96 most productive crystallization conditions.\n* **The Top96 Cryoscreen:** adds optimal cryo-protectant pairings for each condition, solving harvesting and cooling challenges.\n* **Benchmarking:** PDB and internal data confirmed high success rates, with some conditions exceeding 70% effectiveness.\n* **The streamlined screens**\u00a0benefit both high-throughput facilities and small labs by saving time, materials, and costs while boosting success rates.\n\n1. **Shotgun crystallization strategy for structural genomics: an optimized two-tiered crystallization screen against the Thermotoga maritima proteome:** [Page, R.](https://pubmed.ncbi.nlm.nih.gov/12777766/), [Acta Crystallogr D Biol Crystallogr.](https://pubmed.ncbi.nlm.nih.gov/12777766/)\n2. **Structural genomics of the Thermotoga maritima proteome implemented in a high-throughput structure determination pipeline:** [Lesley, S.A.](https://www.pnas.org/doi/10.1073/pnas.142413399), [Proc Natl Acad Sci U S A.](https://www.pnas.org/doi/10.1073/pnas.142413399)\n3. **The JCSG high-throughput structural biology pipeline:** [Elsliger, M.A.,](https://pubmed.ncbi.nlm.nih.gov/20944202/)\u00a0[Acta Crystallogr Sect F Struct Biol Cryst Commun.](https://pubmed.ncbi.nlm.nih.gov/20944202/)\n4. **Sparse matrix sampling: a screening method for crystallization of proteins:** [Jancarik, J.](https://journals.iucr.org/paper?S0021889891004430), [J Appl Crystallogr.](https://journals.iucr.org/paper?S0021889891004430)\n\nThe determination of protein structures through X-ray crystallography has transformed our understanding of biology and disease mechanisms. In recent decades, the surge in genome sequencing has dramatically expanded the number of protein structures available in the Protein Data Bank (PDB). High-throughput structural genomics efforts, such as those at the Joint Center for Structural Genomics (JCSG), demand strategies that maximize efficiency, control costs, and adapt processes for scalability.\n\nJCSG developed a crystallization pipeline capable of screening vast numbers of proteins under diverse chemical conditions. Over 13 years, this approach generated an exceptionally large dataset: more than five million crystallization experiments, over 130,000 crystals, and 1,144 resulting PDB depositions. Analysis of this dataset revealed patterns in which conditions were most likely to yield diffraction-quality crystals\u2014information essential for optimizing future screens.\n\nFrom this data-driven refinement emerged the **Top96 screen**, a curated set of 96 crystallization conditions with the highest yield of successful structures. These conditions were distilled from the original JCSG Core Suite of 384 reagents by ranking each condition by its ability to produce harvestable crystals that resulted in PDB depositions. The team also created the **Top96 Cryoscreen**, which pairs each of the Top96 conditions with its most effective cryo-protection protocol. This addition serves a dual purpose: improving crystallization coverage and aiding in cases where crystal harvesting or cooling is a bottleneck.\n\nThe chemical diversity of Top96 spans organic precipitants, PEG variants, salts, and mixed systems, capturing both common successful formulations and novel conditions absent from commercial kits. Each condition\u2019s success rate was validated against both internal databases and the PDB. Statistical analysis revealed that some conditions had success rates above 70%, while others showed more modest but still meaningful performance.\n\nThese refined screens are not only valuable for the scale of high-throughput projects but also for small research labs where protein availability is limited. They enable a more targeted approach, reducing redundancy, costs, and time, while increasing the odds of crystallization success. By applying lessons from over a decade of data curation, the JCSG Top96 and Top96 Cryoscreen offer a rational, evidence-based platform for structural biology that can accelerate and democratize the path from purified protein to solved structure.", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nResolution of racemic mixtures of compounds can be carried out by any of numerous methods known in the art. An example method includes fractional recrystallization using a chiral resolving acid which is an optically active, salt-forming organic acid. Suitable resolving agents for fractional recrystallization methods are, for example, optically active acids, such as the D and L forms of tartaric acid, diacetyltartaric acid, dibenzoyltartaric acid, mandelic acid, malic acid, lactic acid or the various optically active camphorsulfonic acids such as beta-camphorsulfonic acid. Other resolving agents suitable for fractional crystallization methods include stereoisomerically pure forms of alpha-methylbenzylamine (e.g., S and R forms, or diastereomerically pure forms), 2-phenylglycinol, norephedrine, ephedrine, N-methylephedrine, cyclohexylethylamine, 1,2-diaminocyclohexane, and the like. Resolution of racemic mixtures can also be carried out by elution on a column packed with an optically active resolving agent (e.g., dinitrobenzoylphenylglycine). Suitable elution solvent composition can be determined by one skilled in the art.\n\nCompounds provided herein also include tautomeric forms. Tautomeric forms result from the swapping of a single bond with an adjacent double bond together with the concomitant migration of a proton. Tautomeric forms include prototropic tautomers which are isomeric protonation states having the same empirical formula and total charge. Example prototropic tautomers include ketone-enol pairs, amide-imidic acid pairs, lactam-lactim pairs, enamine-imine pairs, and annular forms where a proton can occupy two or more positions of a heterocyclic system, for example, 1H- and 3H-imidazole, 1H-, 2H- and 4H-1,2,4-triazole, 1H- and 2H-isoindole, 2-hydroxypyridine and 2-pyridone, and 1H- and 2H-pyrazole. Tautomeric forms can be in equilibrium or sterically locked into one form by appropriate substitution.\n\nAll compounds, and pharmaceutically acceptable salts thereof, can be found together with other substances such as water and solvents (e.g., hydrates and solvates) or can be isolated. In some embodiments, preparation of compounds can involve the addition of acids or bases to affect, for example, catalysis of a desired reaction or formation of salt forms such as acid addition salts.\n\nIn some embodiments, the compounds provided herein, or salts thereof, are substantially isolated. By \"substantially isolated\" is meant that the compound is at least partially or substantially separated from the environment in which it was formed or detected. Partial separation can include, for example, a composition enriched in the compounds provided herein. Substantial separation can include compositions containing at least about 50%, at least about 60%, at least about 70%, at least about 80%, at least about 90%, at least about 95%, at least about 97%, or at least about 99% by weight of the compounds provided herein, or salt thereof. Methods for isolating compounds and their salts are routine in the art.\n\nThe term \"compound\" as used herein is meant to include all stereoisomers, geometric isomers, tautomers, and isotopes of the structures depicted. Compounds herein identified by name or structure as one particular tautomeric form are intended to include other tautomeric forms unless otherwise specified.\n\nThe phrase \"pharmaceutically acceptable\" is employed herein to refer to those compounds, materials, compositions, and/or dosage forms which are, within the scope of sound medical judgment, suitable for use in contact with the tissues of human beings and animals without excessive toxicity, irritation, allergic response, or other problem or complication, commensurate with a reasonable benefit/risk ratio.\n\nThe present application also includes pharmaceutically acceptable salts of the compounds described herein. As used herein, \"pharmaceutically acceptable salts\" refers to derivatives of the disclosed compounds wherein the parent compound is modified by converting an existing acid or base moiety to its salt form. Examples of pharmaceutically acceptable salts include, but are not limited to, mineral or organic acid salts of basic residues such as amines; alkali or organic salts of acidic residues such as carboxylic acids; and the like. The pharmaceutically acceptable salts of the present disclosure include the conventional non-toxic salts of the parent compound formed, for example, from non-toxic inorganic or organic acids. The pharmaceutically acceptable salts of the present disclosure can be synthesized from the parent compound which contains a basic or acidic moiety by conventional chemical methods. Generally, such salts can be prepared by reacting the free acid or base forms of these compounds with a stoichiometric amount of the appropriate base or acid in water or in an organic solvent, or in a mixture of the two; generally, non-aqueous media like ether, ethyl acetate, alcohols (e.g., methanol, ethanol, iso-propanol, or butanol) or acetonitrile (ACN) are preferred. Lists of suitable salts are found in *Remington's Pharmaceutical Sciences*, 17th ed., Mack Publishing Company, Easton, Pa., 1985, p. 1418 and *Journal of Pharmaceutical Science*, 66, 2 (1977), each of which is incorporated herein by reference in its entirety.\n\n### Synthesis of the Compounds\n\nCompounds of the present disclosure, including salts thereof, can be prepared using known organic synthesis techniques and can be synthesized according to any of numerous possible synthetic routes.", "# marcdeller.com: Vibe Coding\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n### [**Smiles2YAML** \u2013 From a JSON of candidate SMILES to a folder of Boltz-2 input YAML](https://marcdeller.com/from-a-json-of-candidate-smiles-to-a-folder-of-boltz-2-input-yamls-in-one-prompt/)\n\nVibe Code\n\n### [**FoldForge** \u2013 Browser GUI for the whole Boltz-2 workflow](https://marcdeller.com/a-single-file-browser-gui-for-the-whole-boltz-2-workflow-build-run-review/)\n\nVibe Code\n\n### [**cif2plip** \u2013 Protein-ligand interaction profiling using PLIP for co-folding outputs which produce CIF files](https://marcdeller.com/cif-to-strict-pdb-to-plip-trustworthy-interaction-profiling-for-co-folding-outputs/)\n\nVibe Code\n\n### [**ChatRDKit** \u2013 Talk to RDKit for Chemoinformatics in plain language](https://marcdeller.com/talk-to-rdkit-in-plain-english-cheminformatics-without-the-boilerplate/)\n\n# **ChatRDKit** \u2013 Talk to RDKit for Chemoinformatics in plain language", "# From Single Targets to Disease Networks: A Systems Approach to Drug Discovery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* Drug discovery is moving beyond a \u201cone target, one drug\u201d approach to a systems biology perspective that analyzes entire disease networks.\n* By mapping the \u201cinteractome,\u201d researchers can identify more effective drug targets and better predict potential side effects.\n* This network-based approach is crucial for advancing precision medicine and designing therapies tailored to a patient\u2019s specific molecular profile.\n\n1. **By mapping the \u201cinteractome,\u201d researchers can identify more effective drug targets and better predict potential side effects:**\n   <https://www.sciencedirect.com/journal/current-opinion-in-structural-biology/special-issue/10BFRJKLFCN>\n2. **This network-based approach is crucial for advancing precision medicine and designing therapies tailored to a patient\u2019s specific molecular profile:**\n3. <https://link.springer.com/journal/42485/updates/27759914>\n4. **Drug discovery is moving beyond a \u201cone target, one drug\u201d approach to a systems biology perspective that analyzes entire disease networks:**\n   <https://www.pharmasalmanac.com/articles/the-disruptive-impact-of-structural-biology-on-biopharmaceutical-innovation>\n\nFor decades, the \u201cone target, one drug\u201d model has been the bedrock of pharmaceutical research. Scientists would identify a single protein implicated in a disease and then search for a molecule to block it. While this approach has yielded life-saving medicines, it carries a significant limitation: it often ignores the fact that diseases arise from the failure of complex, interconnected biological networks, not just a single faulty part. This narrow focus is a key reason why so many drugs that look promising in the lab fail in clinical trials.\n\nWe are now entering an era of systems biology, where the focus is shifting from individual components to the entire network. By combining high-throughput structural proteomics with computational modeling, researchers can now map the \u201cinteractome\u201d\u2014the comprehensive web of protein-protein and protein-ligand interactions within a cell. This gives us a bird\u2019s-eye view of cellular signaling, revealing how different pathways intersect and influence one another in both healthy and diseased states. It\u2019s the difference between looking at a single street and seeing the entire city map.\n\nThis holistic perspective is transforming drug discovery. By analyzing the entire disease network, we can identify more effective and robust drug targets, including central nodes that control multiple pathological processes. It also helps us predict potential side effects by showing how a drug might unintentionally interfere with other pathways. This leads to the design of smarter, more targeted therapies and is a critical step toward the ultimate goal of personalized medicine, where treatments can be tailored to the unique molecular profile of an individual\u2019s disease network.", "# GitHub README: chatmol\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/chatmol, 2026\n\n```\n[ChatMol] AI chat interface loaded.\n[ChatMol]   - Type plain English into the PyMOL command line, e.g.\n[ChatMol]       fetch 1abc and show it as a cartoon coloured by chain\n[ChatMol]   - Or use the explicit command:  ai <request>\n[ChatMol]   - Settings:  chatmol_config [key [value]]\n[ChatMol]   - Forget conversation context:  chatmol_reset\n```\n\nThen try it:\n\n```\nPyMOL> fetch 1ubq, async=0\nPyMOL> colour the protein by chain and show it as cartoon\n```\n\n---\n\n## 3. Usage\n\n### Implicit (command-line box)\n\nJust type natural language into PyMOL's command line and press Enter:\n\n```\nPyMOL> load the protein in ~/Downloads/kinase.pdb and colour it by chain\nPyMOL> zoom in on the active site, residues 45 to 60 of chain A\nPyMOL> show the ligand as sticks coloured by element with green carbons\nPyMOL> make a high resolution image and save it to ~/Desktop/render.png\nPyMOL> mutate residue 100 of chain A to alanine\n```\n\nValid PyMOL/Python lines (e.g. `color red, chain A`, `zoom`, `@script.pml`,\n`util.cbc`, lines starting with `#`, `/`, `@`) are always passed straight\nthrough to PyMOL, untouched.\n\n### Explicit\n\nIf auto-detection is off, or you just prefer to be explicit:\n\n```\nPyMOL> ai colour everything by secondary structure\nPyMOL> chatmol show a surface around the binding pocket, transparent\n```\n\n### Multi-turn context\n\nChatMol keeps a short rolling history of your requests and the commands it\ngenerated, so follow-ups like *\"now make it bigger\"* or *\"undo the surface,\nshow sticks instead\"* have context. Clear it with:\n\n```\nPyMOL> chatmol_reset\n```\n\n---\n\n## 4. Configuration\n\nAll settings are stored in `~/.chatmol/config.json` and can be viewed/changed\nwith `chatmol_config`, or via **Plugin > ChatMol Settings** if installed\nthrough the Plugin Manager.\n\n```\nPyMOL> chatmol_config                       # list everything\nPyMOL> chatmol_config model                 # show one value\nPyMOL> chatmol_config model, qwen3:0.6b     # change the model\nPyMOL> chatmol_config host, http://localhost:11434\nPyMOL> chatmol_config temperature, 0.1\nPyMOL> chatmol_config auto, off             # disable command-line auto-translation\nPyMOL> chatmol_config debug, on             # print raw LLM output (troubleshooting)\n```\n\n| Setting        | Default                    | Meaning                                                        |\n|----------------|----------------------------|-----------------------------------------------------------------|\n| `host`         | `http://localhost:11434`  | Ollama server URL                                               |\n| `model`        | `qwen3.6:27b`              | Ollama model tag (must be pulled, must match `ollama list`)     |\n| `temperature`  | `0.2`                      | LLM sampling temperature                                        |\n| `timeout`      | `120`                      | Request timeout (seconds)                                       |\n| `auto`         | `True`                     | Auto-route non-PyMOL command-line input through the LLM         |\n| `echo`         | `True`                     | Print each translated command before running it                |\n| `debug`        | `False`                    | Print the raw LLM response (before cleanup) to the console      |\n| `max_history`  | `6`                        | Number of past request/response turns kept for follow-ups       |\n\n---\n\n## 5. How it works\n\n```\nchatmol/\n\u251c\u2500\u2500 __init__.py        plugin entry point: registers commands, installs hook\n\u251c\u2500\u2500 config.py          JSON-backed settings (~/.chatmol/config.json)\n\u251c\u2500\u2500 ollama_client.py   minimal urllib client for POST /api/chat\n\u251c\u2500\u2500 pymol_reference.py system prompt + curated PyMOL command cheat-sheet\n\u251c\u2500\u2500 nl_router.py       PyMOL-vs-English heuristic, cmd.do hook, translation\n\u251c\u2500\u2500 commands.py        `ai`, `chatmol_config`, `chatmol_reset` commands\n\u2514\u2500\u2500 gui.py             Plugin-menu settings dialog (Qt)\n```\n\n1. **Hooking the command line.** PyMOL's GUI command-line box (and\n   `@script.pml`, menu actions, etc.) all funnel through `cmd.do(line)`.\n   ChatMol saves the original `cmd.do` and replaces it with a wrapper.\n\n2. **Deciding what's \"PyMOL\" vs \"English\".** `nl_router.looks_like_pymol()`\n   checks whether the first token of the line is a known PyMOL command (in\n   `cmd.keyword`, which covers both built-ins and anything registered via\n   `cmd.extend`, e.g. ChatMol's own `chatmol_config`), a known namespace\n   (`util`, `pymol`, `_self`, ...), or a Python keyword. If the word\n   *following* the command looks like an English article/preposition/etc.\n   (e.g. \"show **the** ligand...\") or the line contains connector words like\n   \"except\"/\"but\"/\"without\", it's treated as English instead. As a second\n   safety net, if a line doesn't look like PyMOL *or* English, it's passed\n   through unchanged so odd internal strings aren't sent to the LLM.\n\n3. **Translation.** Anything left over is sent to Ollama\n   (`ollama_client.chat`) with:\n   - a system prompt describing the required output format and a curated\n     PyMOL command reference (`pymol_reference.PYMOL_CHEAT_SHEET`),\n   - a snapshot of the current session (`cmd.get_names`, chains per\n     object), so \"the protein\" / \"it\" / \"chain A\" resolve correctly,\n   - the last few turns of conversation for follow-up requests.\n\n   The model is instructed to output **only** PyMOL command lines (or a\n   `#` comment if the request can't be done), one per line, and to pass\n   already-valid PyMOL syntax through unchanged -- a safety net for cases\n   the heuristic misclassifies. Generation is capped (`num_predict`) and any\n   response that comes back too long or repetitive is discarded with a\n   friendly message rather than executed or remembered, so a bad response\n   from a small model can't snowball into garbage for every later request.", "# Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\nAuthors: Hummel, J.R.; Xiao, K.-J.; Yang, J.C.; Epling, L.B.; Mukai, K.; Ye, Q.; Xu, M.; Qian, D.; Huo, L.; Weber, M.; Roman, V.; Lo, Y.; Drake, K.; Stump, K.; Covington, M.; Kapilashrami, K.; Zhang, G.; Ye, M.; Diamond, S.; Yeleswaram, S.; Macarron, R.; Deller, M.C.; Wee, S.; Kim, S.; Wang, X.; Wu, L.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2024\n\n**Authors:** Joshua R. Hummel (corresponding author), Kai-Jiong Xiao, Jeffrey C. Yang, Leslie B. Epling, Ken Mukai, Qinda Ye, Meizhong Xu, Dingquan Qian, Lu Huo, Michael Weber, Valerie Roman, Yvonne Lo, Katherine Drake, Kristine Stump, Maryanne Covington, Kanishk Kapilashrami, Guofeng Zhang, Min Ye, Sharon Diamond, Swamy Yeleswaram, Ricardo Macarron, Marc C. Deller, Susan Wee, Sunkyu Kim, Xiaozhao Wang, Liangxing Wu and Wenqing Yao. **Affiliations (condensed):** Incyte Research Institute, Incyte Corporation, Wilmington, Delaware 19803, United States. **Published:** J. Med. Chem. 2024, 67, 3112\u20133126. Received 5 December 2023; revised 11 January 2024; accepted 19 January 2024; published 7 February 2024.\n\n## Abstract\n\nCDK2 is a critical regulator of the cell cycle. For a variety of human cancers, the dysregulation of CDK2/cyclin E1 can lead to tumor growth and proliferation. Historically, early efforts to develop CDK2 inhibitors with clinical applications proved unsuccessful due to challenges in achieving selectivity over off-target CDK isoforms with associated toxicity. In this report, we describe the discovery of (4-pyrazolyl)-2-aminopyrimidines as a potent class of CDK2 inhibitors that display selectivity over CDKs 1, 4, 6, 7, and 9. SAR studies led to the identification of compound 17, a kinase selective and highly potent CDK2 inhibitor (IC50 = 0.29 nM). The evaluation of 17 in CCNE1-amplified mouse models shows the pharmacodynamic inhibition of CDK2, measured by reduced Rb phosphorylation, and antitumor activity.\n\n## Introduction\n\nCyclin-dependent kinases (CDKs) are a family of serine/threonine protein kinases that regulate key cellular processes and drive the progression of cell cycle. Inactive in their monomeric form, CDK activity is governed by heterodimerization with regulatory subunits, known as cyclins, and an intricate network of phosphorylation events. Under homeostasis, these regulatory mechanisms are tightly controlled to ensure successful cell division. Conversely, the dysregulation of CDK/cyclin activity can cause sustained, uncontrolled cellular proliferation, a hallmark of cancer. Although the key concepts of CDK biology were first established in the 1990s, translating the fundamental understanding of cell cycle control into effective cancer therapies has proven challenging. Pioneering efforts identified pan-CDK inhibitors that lacked an optimal balance between safety and clinical efficacy due to toxicities associated with multitargeted CDK activity. Recently, the FDA approval of the selective CDK4/6 dual inhibitors palbociclib, ribociclib, and abemaciclib for the treatment of hormone receptor-positive (HR+) and human epidermal growth factor receptor 2-negative (HER2\u2212) breast cancer (Figure 1) has reignited interest in CDK targeted cancer therapies, and marks a paradigm shift toward the discovery of novel selective CDK inhibitors.\n\n*Figure 1.* FDA-approved selective CDK4/6 dual inhibitors.\n\nAlthough multiple CDKs have been proposed as therapeutic targets for cancer, CDK2 is of particular interest because the dysregulation of CDK2/cyclin E1 has been identified as a mechanism of resistance to CDK4/6 blockade. CDK2 partners canonically with cyclin E1 to phosphorylate and inactivate the tumor suppressor retinoblastoma protein (Rb), which releases the E2F transcription factor that ultimately leads to cell cycle progression from G1 to S-phase. Emerging evidence suggests that CDK2 activity is critical for cellular proliferation across multiple oncogenic pathways, and that for a subset of genetically defined cancers, the dysregulation of CDK2/cyclin E1 is associated with tumor growth and aggressiveness, and can drive malignant transformation. For instance, the amplification of CCNE1, the gene that encodes for cyclin E1, often correlates with poor prognosis in ovarian, breast, gastric, and endometrial cancers. In addition, preclinical studies have shown that genetic suppression and pharmacological inhibition of CDK2 can selectively inhibit cell proliferation and induce G1 arrest in CCNE1-amplified cancer cells. Taken together, these studies support the selective inhibition of CDK2 as a therapeutic target for CCNE1-amplified tumors. While several selective CDK2 inhibitors recently advanced to clinical trials (INCB123667, PF-07104091, BLU-222, ARTS-021, and INX-315), none are currently approved. In search of novel therapies to address this unmet clinical need, our discovery program set out to identify potent and selective small molecule inhibitors of CDK2.\n\n## Results and Discussion\n\n### Hit identification and early scaffold design\n\nA high-throughput screen (HTS) of Incyte's internal compound collection identified hit (\u00b1)-1 as a promising starting point that displayed submicromolar enzymatic activity against CDK2/Cyclin E1 (IC50 = 607 nM) but also potently inhibited its primary target JAK2 (Figure 2). In vitro CDK2 biochemical potency and selectivity was assessed in non-radiometric enzyme activity assays. With a focus on improving CDK2 potency and general kinase selectivity, our initial SAR efforts led to the discovery of compound 2 that showed increased CDK2 activity (IC50 = 16 nM) and no measurable activity against JAK2 (IC50 > 10,000 nM). Compound 2 also exhibited excellent selectivity against an in-house panel of 8 kinases (see Supporting Information, Table S1).\n\n*Figure 2.* Structure of high-throughput screening hit (\u00b1)-1 and initial scaffold discovery.", "# GitHub README: cif2plip\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/cif2plip, 2026\n\n> **CIF to strict PDB to PLIP. Trustworthy interaction profiling for co-folding outputs.**\n\n    \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/cif2plip\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/cif2plip</a></td>\n</tr>\n</table>\n\n---\n\n**CIF \u2192 strict PDB \u2192 PLIP protein-ligand interaction analysis**\n\nWhy it matters: co-folding tools such as Boltz-2, AlphaFold 3, and Chai-1 emit ModelCIF files that quietly violate the PDB format (multi-character chain IDs, oversized residue names, no bond table), and naive conversion introduces a subtle CONECT-parsing bug that corrupts ligand perception and invents fictitious interactions. cif_to_plip diagnoses and fixes every one of these issues before PLIP ever sees the structure, so the protein-ligand interactions you report are real. It is useful for anyone doing structure-based drug design on predicted complexes: reliable, reproducible interaction fingerprints (H-bonds, salt bridges, pi-stacking, halogen bonds, metal coordination) straight from a raw co-folding CIF.\n\nA single-file Python pipeline that converts a ModelCIF or mmCIF co-folding output into a format-compliant PDB file and feeds it directly into [PLIP](https://github.com/pharmai/plip) for full protein-ligand interaction profiling. Designed for downstream analysis of [Boltz-2](https://github.com/jwohlwend/boltz) and similar co-folding outputs where the input CIF carries no bond table and often contains multi-character chain IDs that break the PDB format specification.\n\n---\n\n## Why this script exists\n\nModern co-folding tools (Boltz-2, AlphaFold 3, Chai-1) emit ModelCIF files with:\n\n- **Multi-character chain IDs** (`JH1`, `J031`, \u2026) that exceed the single-character PDB column 22 limit\n- **Residue names longer than 3 characters** (`LIG1`, `SMOL`, \u2026) that overflow PDB columns 18\u201320\n- **No CONECT records** \u2014 bond topology is absent from the file\n\nNa\u00efvely converting such files with `gemmi` or `pdb_tidy` alone introduces a subtle but critical bug: `pdb_tidy` leaves a one-serial gap at each chain/TER boundary. Because OpenBabel maps CONECT records **positionally** by serial number, any gap causes records past that point to be silently discarded. The result is corrupted ligand perception \u2014 wrong SMILES, fictitious bonds in the PyMOL session, and invented PLIP interactions. `cif_to_plip` diagnoses and resolves all of these issues before PLIP ever sees the structure.\n\n---\n\n## Pipeline\n\n```\nCIF input\n   \u2502\n   \u251c\u2500 [1] gemmi parse + sanity check (single model enforced)\n   \u251c\u2500 [2] Remap multi-character chain IDs \u2192 single-character (A, B, C \u2026)\n   \u251c\u2500 [3] Rename >3-char residue names \u2192 3-char codes\n   \u251c\u2500 [4] Write strict PDB (element cols 77-78, TER records, occupancy/B padding)\n   \u251c\u2500 [5] pdb_tidy \u2192 renumber all serials contiguously (removes TER-boundary gap)\n   \u251c\u2500 [6] Generate intramolecular CONECT records from covalent-radius geometry\n   \u251c\u2500 [7] pdb_validate \u2014 format compliance check\n   \u2514\u2500 [8] PLIP \u2192 XML + TXT + PNG + PyMOL .pse\n              \u2514\u2500 [9] Flatten XML \u2192 interactions CSV + ligand summary CSV\n```\n\n---\n\n## Outputs\n\n| File | Description |\n|---|---|\n| `<stem>.pdb` | Standards-compliant PDB with contiguous serials and CONECT records |\n| `<stem>_plip/<stem>_report.xml` | Full PLIP XML report |\n| `<stem>_plip/<stem>_report.txt` | Human-readable PLIP text report |\n| `<stem>_plip/*.png` | Ray-traced binding-site images (one per ligand) |\n| `<stem>_plip/*.pse` | PyMOL session with PLIP interaction visualisation |\n| `<stem>_interactions.csv` | Tidy, flattened interactions table (one row per contact) |\n| `<stem>_ligand_summary.csv` | Per-ligand summary with SMILES, InChIKey, interaction counts |\n\n---\n\n## Installation\n\nConda is recommended to manage the mixed Python/OpenBabel/PyMOL stack:\n\n```bash\nconda create -n cif2plip python=3.11 -y\nconda activate cif2plip\n\n# Core dependencies\npip install gemmi pdb-tools plip\n\n# OpenBabel (required by PLIP for ligand perception)\nconda install -c conda-forge openbabel -y\n\n# Open-source PyMOL (required by PLIP for .pse session generation)\nconda install -c conda-forge pymol-open-source -y\n```\n\nVerify the install:\n\n```bash\npython -m plip.plipcmd --help\npdb_tidy --help\n```\n\n---\n\n## Usage\n\n```\npython cif_to_plip.py <input.cif> [-o <output_dir>] [--no-plip]\n```\n\n### Arguments\n\n| Argument | Description |\n|---|---|\n| `cif` | Input `.cif` / `.mmcif` file (required) |\n| `-o`, `--outdir` | Output directory (default: same directory as input) |\n| `--no-plip` | Convert to PDB only; skip PLIP analysis |\n\n### Examples\n\nRun the full pipeline:\n\n```bash\npython cif_to_plip.py boltz_output/jak1_lig.cif -o results/jak1/\n```\n\nConvert only (useful for format validation or feeding into other tools):\n\n```bash\npython cif_to_plip.py model.cif --no-plip\n```\n\n---\n\n## Interaction types reported\n\nPLIP detects and reports the following contact classes, all of which appear as distinct `interaction_type` values in the output CSV:\n\n- Hydrophobic contacts\n- Hydrogen bonds (donor/acceptor, with geometry)\n- Water bridges\n- Salt bridges\n- \u03c0\u2013stacking (face-to-face and edge-to-face)\n- \u03c0\u2013cation interactions\n- Halogen bonds\n- Metal coordination\n\n---\n\n## Technical notes\n\n### Chain ID remapping\nChain IDs are assigned in order of appearance using the 62-character alphabet `A\u2013Z`, `a\u2013z`, `0\u20139`. The mapping is printed to stdout at runtime.\n\n### Residue name truncation\nNames exceeding 3 characters are truncated to their first 3 characters. Collisions are resolved by assigning sequential codes `L01`, `L02`, \u2026 The mapping is printed to stdout.", "# marcdeller.com: My Publications\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n## From Structure to IND: Stories Behind My Publications\n\n60+ top-cited peer-reviewed manuscripts showcase groundbreaking advancements in structural biology, medicinal chemistry, and immunology. They highlight my expertise in designing key experiments, solving complex biological problems, and mentoring teams. Covering topics like drug discovery, protein stability, and vaccine design, these studies reflect global collaboration that drives innovation across disciplines.\n\n[OpenAlex](https://openalex.org/authors/A5072187299)\n[Orcid](https://orcid.org/0000-0001-8070-6502)\n[Scopus](https://www.scopus.com/authid/detail.uri?authorId=21741715300)\n\nTop Publications\n\n### [2012 \u2013 Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33](https://marcdeller.com/structure-of-hepatitis-c-virus-envelope-glycoprotein-e2-antigenic-site-412-to-423-in-complex-with-antibody-ap33/)\n\n[2012 \u2013 Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33](https://marcdeller.com/structure-of-hepatitis-c-virus-envelope-glycoprotein-e2-antigenic-site-412-to-423-in-complex-with-antibody-ap33/)\n\n### [2013 \u2013 Crystal structure of a soluble cleaved HIV-1 envelope trimer](https://marcdeller.com/crystal-structure-of-a-soluble-cleaved-hiv-1-envelope-trimer/)\n\n[2013 \u2013 Crystal structure of a soluble cleaved HIV-1 envelope trimer](https://marcdeller.com/crystal-structure-of-a-soluble-cleaved-hiv-1-envelope-trimer/)\n\n### [2013 \u2013 Models of protein\u2013ligand crystal structures: trust, but verify](https://marcdeller.com/models-of-protein-ligand-crystal-structures-trust-but-verify/)\n\n[2013 \u2013 Models of protein\u2013ligand crystal structures: trust, but verify](https://marcdeller.com/models-of-protein-ligand-crystal-structures-trust-but-verify/)\n\n### [2013 \u2013 Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120](https://marcdeller.com/supersite-of-immune-vulnerability-on-the-glycosylated-face-of-hiv-1-envelope-glycoprotein-gp120/)\n\n[2013 \u2013 Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120](https://marcdeller.com/supersite-of-immune-vulnerability-on-the-glycosylated-face-of-hiv-1-envelope-glycoprotein-gp120/)\n\n### [2014 \u2013 Approaches to automated protein crystal harvesting](https://marcdeller.com/approaches-to-automated-protein-crystal-harvesting/)\n\n[2014 \u2013 Approaches to automated protein crystal harvesting](https://marcdeller.com/approaches-to-automated-protein-crystal-harvesting/)\n\nPapers\n\n### [2026 \u2013 Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases](https://marcdeller.com/discovery-of-the-orally-bioavailable-isoform-selective-janus-kinase-1-compound-povorcitinib/)\n\nPapers\n\n### [2025 \u2013 Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor](https://marcdeller.com/discovery-of-incb159020-an-orally-bioavailable-kras-g12d-inhibitor/)\n\nPapers\n\n### [2024 \u2013 Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective CDK2 Inhibitors](https://marcdeller.com/discovery-of-4-pyrazolyl-2-aminopyrimidines-as-potent-and-selective-cdk2-inhibitors/)\n\nPapers\n\n### [2022 \u2013 Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant FGFR2/3](https://marcdeller.com/discovery-of-potent-and-selective-inhibitors-of-wild-type-and-gatekeeper-mutant-fgfr2-3/)\n\nPapers\n\n### [2018 \u2013 Structure\u2013function analysis of the extended conformation of a polyketide synthase module](https://marcdeller.com/structure-function-analysis-of-the-extended-conformation-of-a-polyketide-synthase-module/)\n\nPapers\n\n### [2017 \u2013 Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands](https://marcdeller.com/validation-of-protein-ligand-crystal-structure-models-small-molecule-and-peptide-ligands/)\n\nPapers\n\n### [2017 \u2013 The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase](https://marcdeller.com/the-role-of-the-secondary-coordination-sphere-in-a-fungal-polysaccharide-monooxygenase/)\n\nPapers\n\n### [2017 \u2013 Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design](https://marcdeller.com/lipid-interactions-and-angle-of-approach-to-the-hiv-1-viral-membrane-of-broadly-neutralizing-antibody-10e8-insights-for-vaccine-and-therapeutic-design/)\n\nPapers\n\n### [2016 \u2013 Key gp120 glycans pose roadblocks to the rapid development of VRC01-class antibodies in an HIV-1- infected Chinese donor](https://marcdeller.com/key-gp120-glycans-pose-roadblocks-to-the-rapid-development-of-vrc01-class-antibodies-in-an-hiv-1-infected-chinese-donor/)\n\nPapers\n\n### [2015 \u2013 Protein stability: a crystallographer\u2019s perspective](https://marcdeller.com/protein-stability-a-crystallographers-perspective/)\n\nPapers\n\n### [2015 \u2013 Models of protein\u2013ligand crystal structures: trust, but verify](https://marcdeller.com/models-of-protein-ligand-crystal-structures-trust-but-verify-2/)\n\nPapers\n\n### [2015 \u2013 Crystal structure of a two-subunit TrkA octameric gating ring assembly](https://marcdeller.com/crystal-structure-of-a-two-subunit-trka-octameric-gating-ring-assembly/)\n\nPapers\n\n### [2014 \u2013 Approaches to automated protein crystal harvesting](https://marcdeller.com/approaches-to-automated-protein-crystal-harvesting/)\n\nPapers\n\n### [2013 \u2013 Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120](https://marcdeller.com/supersite-of-immune-vulnerability-on-the-glycosylated-face-of-hiv-1-envelope-glycoprotein-gp120/)\n\nPapers\n\n### [2013 \u2013 Models of protein\u2013ligand crystal structures: trust, but verify](https://marcdeller.com/models-of-protein-ligand-crystal-structures-trust-but-verify/)\n\nPapers\n\n### [2013 \u2013 Crystal structure of a soluble cleaved HIV-1 envelope trimer](https://marcdeller.com/crystal-structure-of-a-soluble-cleaved-hiv-1-envelope-trimer/)\n\nPapers\n\n### [2012 \u2013 Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33](https://marcdeller.com/structure-of-hepatitis-c-virus-envelope-glycoprotein-e2-antigenic-site-412-to-423-in-complex-with-antibody-ap33/)", "# GPCR Binning \u2013 Structural binning and organisation utilities for GPCR datasets\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-14 on marcdeller.com (categories: vibe-code).*\n\n> **Structural binning and organisation utilities for GPCR datasets.**\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/gpcr-binning](https://github.com/bellcheddar/gpcr-binning) |\n\n---\n\nGPCR Binning is a working repository for organising and grouping G-protein-coupled receptor (GPCR) structural data. GPCRs are the single largest class of drug targets, and the structural data describing them is large, heterogeneous, and growing fast across receptor families, conformational states, and bound ligands. This project collects utilities for binning that data into coherent, analysable groups.\n\nWhy it matters: GPCRs account for a substantial fraction of all approved drugs, and structure-based work on them depends on being able to slice a sprawling structural landscape into meaningful categories (by class, by state, by binding mode, by resolution and method). Sensible binning is the foundation for everything downstream: selectivity analysis, template selection for modelling, and triage for structure-based drug design. This repository is useful as a place to develop and keep those binning and grouping utilities in one structural-biology-aware codebase.\n\n> \u26a0\ufe0f **Status:** This project is in active development. The sections below describe its intended scope; specifics will be filled in as the codebase matures.\n\n---\n\n## \ud83c\udfaf Scope\n\n| Area | Intent |\n| --- | --- |\n| \ud83d\uddc3\ufe0f **Binning** | Group GPCR structures into coherent categories for analysis |\n| \ud83e\uddec **Domain-aware** | Built around GPCR structural biology rather than generic clustering |\n| \ud83d\udd2c **Analysis-ready** | Produce groupings suitable for downstream SBDD and comparison |\n\n---\n\n## \ud83d\ude80 Getting started\n\n```\ngit clone https://github.com/bellcheddar/gpcr-binning.git\ncd gpcr-binning\n```\n\nUsage instructions will be documented here as the tooling stabilises.\n\n---\n\n## \ud83e\udded Context\n\nThis repository sits alongside related structure-focused tools in the portfolio, including [PocketTide](https://github.com/bellcheddar/PocketTide) (a GLP-1R structure dashboard) and the Boltz-2 co-folding utilities. Together they support a structure-based approach to drug discovery across GPCR and other target classes.\n\n---\n\n## \ud83d\udcc4 Licence\n\nSee repository for licence terms.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/gpcr-binning](https://github.com/bellcheddar/gpcr-binning) |", "# GitHub README: CODSWALLOP\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/CODSWALLOP, 2026\n\n## \ud83d\udcda Citing the data\n\nCODSWALLOP is a lens over other people's archives. If the data does the work in your paper, cite them:\n\n- **RCSB PDB**: Burley, S.K. et al. *Nucleic Acids Res.* (2023) `10.1093/nar/gkac1077`\n- **PDBe**: Armstrong, D.R. et al. *Nucleic Acids Res.* (2020) `10.1093/nar/gkz990`\n- **UniProt**: The UniProt Consortium. *Nucleic Acids Res.* (2023) `10.1093/nar/gkac1052`\n- **InterPro**: Paysan-Lafosse, T. et al. *Nucleic Acids Res.* (2023) `10.1093/nar/gkac993`\n\n## \ud83d\udcc4 Licence\n\nMIT. See [LICENSE](LICENSE).\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/CODSWALLOP\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/CODSWALLOP</a></td>\n</tr>\n</table>", "# GitHub README: site-doctor\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/site-doctor, 2026\n\n> **One pass. Full diagnosis. Health, content, and SEO for any WordPress site.**\n\n    \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/site-doctor\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/site-doctor</a></td>\n</tr>\n</table>\n\n---\n\nA single-file, dependency-free Python 3 tool that diagnoses the health,\ncontent, and SEO/performance profile of a WordPress site \u2014 using only the\nstandard library.\n\nWhy it matters: keeping a WordPress site fast, secure, and well-indexed normally means juggling several separate tools (a header scanner, a TLS checker, a PageSpeed run, a content audit). Site Doctor folds all of that into a single dependency-free Python script that reads only public information and emits both a colour-coded terminal report and a branded, self-contained HTML report in one pass. It is useful for site owners, agencies, and developers who want a fast, repeatable, zero-install health check: ideal for pre-launch audits, ongoing maintenance, and verifying that caching, security headers, and Core Web Vitals are actually configured the way you think they are.\n\nIt produces a color-coded report on screen **and** a branded HTML report\non disk, in the same pass.\n\n## Features\n\nBy default, Site Doctor runs **all three** checks below against a target URL\nin a single pass and writes one combined HTML report. Pass `--scan`,\n`--inventory`, and/or `--analytics` to run only a subset.\n\n### 1. Diagnostic Scan (`--scan`)\n- Request timing breakdown: DNS, TCP connect, TLS handshake, TTFB, download, total\n- HTTP status, response headers, and redirect chain\n- Caching behaviour (`Cache-Control`, `ETag`, `Last-Modified`, etc.)\n- Security headers (HSTS, CSP, X-Frame-Options, X-Content-Type-Options, Referrer-Policy, Permissions-Policy)\n- Cookie flags (`Secure`, `HttpOnly`, `SameSite`)\n- CORS configuration and mixed-content checks\n- SSL/TLS certificate details (issuer, expiry, SANs), supported TLS protocol versions, ALPN/HTTP2 detection\n- DNS records (A, AAAA, MX, TXT, NS, CNAME, PTR)\n- `robots.txt`, sitemap, favicon, and custom 404 handling checks\n- HTTP \u2192 HTTPS and www \u2194 non-www redirect behaviour\n- Server technology / CDN detection (e.g. Cloudflare, Nginx, LiteSpeed)\n- WordPress fingerprinting: version disclosure, active theme, plugins, REST API,\n  XML-RPC, `readme.html`, `wp-login.php`, and exposed user accounts via\n  `/wp-json/wp/v2/users`\n- Detected analytics/tracking scripts (Google Analytics, Jetpack Stats, etc.)\n- A weighted **Health Scorecard** (0\u2013100, grade A\u2013F) across Security Headers,\n  TLS/HTTPS, Performance, WordPress Hygiene, HTTP Behaviour, and SEO Basics\n\n> Note: visitor counts and live CPU/memory load cannot be measured externally\n> over HTTP \u2014 the report says so explicitly rather than guessing.\n\n### 2. WordPress Content Inventory (`--inventory`)\n- Walks every public post type exposed via the WordPress REST API\n  (`/wp-json/wp/v2/...`), paginating through results\n- De-duplicates by type + ID\n- Outputs a clean table: **ID, Type, Date, Category, Title**\n\n### 3. Analytics & SEO Report (`--analytics`)\n- Real Lighthouse scores via the public Google PageSpeed Insights API\n  (Performance, SEO, Accessibility, Best Practices, plus Core Web Vitals)\n- Automatic fallback to a self-computed performance scorecard if PSI is\n  unavailable (rate-limited, offline, etc.)\n- Per-page SEO breakdown for the site's content: title/meta description\n  length, H1 count, word count, image alt coverage, internal/external link\n  counts, Open Graph/Twitter tags, structured data (JSON-LD), and load time\n\n## Requirements\n\n- Python 3.8+\n- No third-party packages \u2014 only the standard library\n- Optional: `dig` (used for DNS lookups if available, with a socket-based\n  fallback otherwise)\n\n## Usage\n\n```\npython3 site_doctor.py [URL] [OPTIONS]\n```\n\nIf `URL` is omitted, you'll be prompted to enter one interactively (e.g.\n`https://example.com`). The scheme defaults to `https://` if not given.\n\n```bash\n# Prompt for a URL and run everything (scan + inventory + analytics)\npython3 site_doctor.py\n\n# Run everything against a specific site\npython3 site_doctor.py https://example.com\n\n# Run only the diagnostic scan and health scorecard\npython3 site_doctor.py https://example.com --scan\n\n# Run only the WordPress content inventory\npython3 site_doctor.py https://example.com --inventory\n\n# Run only the analytics/SEO report, skipping PageSpeed Insights,\n# analyzing 25 pages instead of the default 10\npython3 site_doctor.py https://example.com --analytics --max-pages 25 --no-psi\n\n# Run everything and write the HTML report to a custom path\npython3 site_doctor.py https://example.com --out reports/example.html\n```\n\n### Mode flags (omit all to run everything)\n\n| Flag | Description |\n| --- | --- |\n| `--scan` | Diagnostic scan: request timing (DNS/connect/TLS/TTFB/download), HTTP headers, caching, security headers, cookies, CORS/mixed content, TLS certificate and protocol support, DNS records, robots.txt/sitemap/favicon, redirect behaviour, server/CDN fingerprint, WordPress core/theme/plugin/user fingerprinting, and the weighted Health Scorecard (0\u2013100, grade A\u2013F). |\n| `--inventory` | WordPress content inventory: walks every public post type exposed via `/wp-json/wp/v2/...`, paginates through all results, de-duplicates by type+ID, and prints a table of ID, Type, Date, Category, Title. |\n| `--analytics` | Analytics & SEO report: real Lighthouse scores via Google PageSpeed Insights (Performance, SEO, Accessibility, Best Practices, Core Web Vitals), with automatic fallback to a self-computed performance scorecard if PSI is unavailable, plus a per-page SEO breakdown across the site's content. |", "# HGMP01 and other Hidden Guardians: How the Human Gut Fights Back Against Plastic Pollution\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe temperature and pH stability of HGMP01 proves particularly significant for physiological relevance. While many environmental PET degrading enzymes require elevated temperatures near the polymer\u2019s glass transition point of 65 to 80\u00b0C, HGMP01 maintains considerable activity at human body temperature. Its tolerance for alkaline conditions aligns with pH variations encountered in different intestinal segments, suggesting evolutionary adaptation to the gut environment. Concentration dependent effects show proportional increases in hydrolysis products with ascending enzyme levels from 50 to 2000 nanomolar, indicating substantial degradation potential when sufficient enzyme is present.\u200b\n\nThe broader health implications remain under investigation. Microplastic exposure alters gut microbiome composition, decreasing beneficial genera while promoting pathogenic species expansion. These shifts trigger inflammatory responses, compromise intestinal barrier function, and modify metabolic outputs including short chain fatty acid production. Microbial composition changes associated with microplastic treatment resemble patterns previously linked to depression and colorectal cancer, though direct causal relationships require further study. The interaction between microplastics and gut bacteria creates bidirectional effects where plastics alter microbial ecology while microbes simultaneously attempt degradation.\u200b\n\nMoving forward, this discovery positions the gut microbiome as both victim and potential solution to plastic contamination. The natural presence of PET hydrolyzing capacity in intestinal bacteria suggests evolutionary pressure from environmental exposure. While current enzymatic activity remains lower than engineered variants from environmental sources, these gut resident enzymes operate under physiological conditions without requiring external intervention. Understanding these endogenous degradation mechanisms could inform therapeutic strategies ranging from probiotic formulations to enzyme supplementation for reducing microplastic burden and associated health risks.\u200b\n\nxml version=\"1.0\" encoding=\"UTF-8\"?\n\n| Key Concepts | Description | Key References |\n| --- | --- | --- |\n| HGMP01 Protein | A novel PET hydrolase enzyme identified from the human gut metagenome that demonstrates the capacity to break down polyethylene terephthalate nanoparticles through enzymatic hydrolysis. This enzyme exhibits activity at body temperature (37\u00b0C) and functions across a broad pH range (7.0-11.0), with optimal activity under mildly alkaline conditions. | [Zhang, et al., International Journal of Biological Macromolecules, 2024](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf) |\n| PET Hydrolyzing Enzymes | A family of enzymes including PETases and MHETases that catalyze the breakdown of polyethylene terephthalate plastics through hydrolysis reactions. These enzymes possess a conserved serine-histidine-aspartate catalytic triad mechanism and function through nucleophilic attack on ester bonds within the polymer structure, converting PET into monomeric building blocks. | [Zhou, et al., Microorganisms, 2024](https://pmc.ncbi.nlm.nih.gov/articles/PMC10820491/) |\n| Microplastic Accumulation in IBD | Research demonstrates that patients with inflammatory bowel disease exhibit significantly elevated concentrations of microplastic particles in fecal samples compared to healthy individuals. Studies show fecal microplastic concentration in IBD patients (41.8 items/g) is substantially higher than in healthy controls (28.0 items/g), with polyethylene terephthalate and polyamide as the dominant polymer types detected. | [Yan, et al., Environmental Science & Technology, 2022](https://pubmed.ncbi.nlm.nih.gov/34935363/) |\n| Temperature and pH Stability | HGMP01 demonstrates considerable enzymatic activity for PET hydrolysis across diverse environmental conditions. The enzyme maintains sustained high activity at human body temperature (37\u00b0C) with an optimal temperature of 40\u00b0C. It exhibits broad pH tolerance spanning 7.0 to 11.0, with maximum activity observed at pH 10.0-11.0, indicating robust performance under physiological alkaline conditions found in certain regions of the digestive tract. | [Zhang, et al., International Journal of Biological Macromolecules, 2024](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf) |\n| Gut Microbiome Biotransformation | Emerging evidence reveals that the human gut microbiota possesses the capability to biotransform and potentially degrade ingested microplastic particles during gastrointestinal transit. Studies using simulated digestion models demonstrate that PET microplastics undergo structural modifications in the colon and alter microbial community composition, with certain bacterial genera showing adherence to microplastic surfaces and biofilm formation. | [Tamargo, et al., Scientific Reports, 2022](https://pubmed.ncbi.nlm.nih.gov/35017590/) |", "# Cell surface receptors\nAuthors: Deller, M.C.; Jones, E.Y.\nSource: Current Opinion in Structural Biology, 2000\n\nThe most recent example of lattice contacts provoking previously unanticipated models for protein assemblies comes from the structure determination of B7-1 (CD80), an IgSF receptor involved in the control of T-cell activation [51]. The formation of parallel, twofold rotationally symmetric homodimers in the crystal (an association also supported by subsequent analytical ultracentrifugation data) suggests a mechanism for avidity enhanced binding to the dimeric counter-receptor CTLA-4.\n\n## Virus\u2013receptor complexes\n\nCell surface receptors are exploited by viruses as a means of attachment and entry into cells. Such interactions are particularly amenable to study by single-particle cryo-electron microscopy. The application of this technique to adenovirus coated with a soluble form of its specific integrin receptor has yielded an, as yet, tantalisingly hazy view of the structure at approximately 21 \u00c5 resolution [52]. The approach becomes truly powerful when coupled with high-resolution structures of isolated components. Such data are available for rhinovirus and the relevant N-terminal two domains of its receptor, intercellular adhesion molecule-1 (ICAM-1). These crystal structures, fitted to a cryo-electron microscopy reconstruction of the rhinovirus\u2013ICAM-1 complex, yield a highly informative view of this pathological interaction [53].\n\n## Conclusions\n\nThis past year has seen an unprecedented explosion in the number of new structures of receptor complexes. As a result, this review, rather than just focusing on the architecture of isolated receptors and putative models for their functional interactions, has been able to emphasise detailed structural data for the actual complexes. These data, in turn, have highlighted the necessity of close interplay between structural and functional analyses to test crystal-contact-inspired hypotheses.\n\n## Acknowledgements\n\nWe thank our colleagues in Structural Biology at the Wellcome Trust Centre for Human Genetics for stimulating discussions. EYJ is funded by the Royal Society and MCD by the Cancer Research Campaign and the Medical Research Council.\n\n## References\n\n*(reference list omitted)*", "# Approaches to automated protein crystal harvesting\nAuthors: Deller, M. C.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 70, 133\u2013155 (IYCr crystallization series), 2014\n\nMicro-electromechanical systems are presented as potentially fundamental to the future of crystal manipulation. The RodBot system uses micrometre-sized magnetized MEMS agents for gentle manipulation in solution, and MEMS technology already appears in the microshovel of the CARESS system and the loopless mounts of the CMM system. Biocompatible MEMS materials such as SU-8, a photosensitive epoxy negative photoresist widely used in biomedical MEMS, are used to build the REACH microgrippers; SU-8 offers simple fabrication, high flexibility that avoids damaging fragile crystals, and X-ray background scattering low enough for data collection while the crystal is still held in the gripper. Other MEMS devices suited to biological manipulation include nanometre-precision microgrippers, thermally actuated microgrippers and grippers exploiting adhesive forces, and MEMS photolithography has been used to make microscale silicon crystallization plates. The authors predict that full automation of harvesting will depend on such advances and that traditional loop-based harvesting will eventually be rendered redundant, with MEMS tools customized for automated systems essential to moving the field forward.\n\n### 2. Automation of protein crystal harvesting, and 2.1 fully automated systems\n\nThe engineering challenge is contrasted with conventional laboratory automation, which follows a defined sequence with limited need for operational flexibility and little sensory feedback. An autonomous harvesting system must react in real time to the motion of the crystal and make intelligent choices from sensory feedback, which requires real-time machine vision providing closed-loop control of the robot's actuators \u2014 cutting-edge and costly mechatronics. For that reason none of the systems reviewed has reached full autonomy and only a few have reached an operator-assisted or semi-automated level. Fully automated harvesting requires identification of the target crystal, harvesting from the mother liquor, then cryoprotection and flash-cooling. Crystal detection is the crux and is hard for three reasons: protein crystals well ordered at the molecular level are often disordered macroscopically with irregular edges and textures; crystals move during harvesting so the very features used to identify them change with orientation, making real-time positional feedback extremely difficult; and both false positives from non-crystal objects and false negatives from missed crystals must be rigorously rejected, while the algorithms must run fast enough for a real-time control loop.\n\nDetection approaches all involve some form of edge detection and texture analysis, and the review lists self-organizing neural nets, machine learning, Fisher linear discriminant analysis, Bayes object classification, decision trees and support vector machines. Success rates remain low, with error rates around 15% common, which Marc Deller and Bernhard Rupp judge insufficient for routine HTPX/SG use and inadequate for full autonomy. They suggest full automation will need detection methods not based solely on image-based computer vision, and point to spectroscopic imaging, birefringence, second-harmonic generation microscopy, UV fluorescence, two-photon fluorescence and X-ray diffraction rastering as candidates.\n\n### 2.2. Semi-automated protein crystal harvesting\n\nSemi-automated, or operator-assisted, systems sidestep the hardest problems by keeping a human in the loop to identify crystals and partially guide the robot, so they avoid complex computer-vision and robotic guidance systems, have simpler designs and lower construction costs, and retain many benefits of automation. Their drawbacks are lower throughput and the labour-intensive nature of skilled human-assisted operation. Searches of the internet and of patent and literature databases revealed 14 semi-automated harvesting systems at varying levels of deployment and automation, summarized in Table 3. Five were designed specifically for HTPX/SG use \u2014 the chopsticks micro-hand, ACH, CARESS, REACH and UMR \u2014 and two, PCGDS and CPPI, were designed for remote telerobotic operation in space. The review then walks through each system's key features with an eye to suitability for HTPX/SG deployment, grouping them by the underlying harvesting mechanism: microcapillary, microtool, microgripper, acoustic droplet ejection, optical trap and MEMS.\n\n### 2.2.1\u20132.2.2. Space-based systems: PCGDS and CPPI", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\n| Volume (\u03bcl) | Component |\n|---|---|\n| 0.5 | Purified PCR Product |\n| 1.0 | pGEX-3C Vector (BamHI\u2013EcoRI cut) (2 ng/ml) |\n| 1.0 | 10x Ligation Buffer |\n| 1.0 | Acetylated BSA (100 ng/\u03bcl) |\n| 1.0 | T4 Ligase (5 units/ml) |\n| 6.0 | Sterile Distilled Water |\n| 10.0 | Total |\n\n*Table 2-5.* Table showing composition of ligation reaction mixture used to ligate PCR product to vector DNA.\n\n### 2.2.8 Preparation of Competent Cells\n\nE. coli strain XL1-Blue (Invitrogen), were streaked onto LB plates and grown for 16 hours at 37 \u00b0C. A single colony was picked and used to inoculate 5 ml of LB media. The culture was grown, with shaking, overnight at 37 \u00b0C. This starter culture was diluted to 200 ml of LB and then further grown to an OD600 ~ 0.5. The cells were harvested at 3000 x g and the pellet resuspended gently in 80 ml of sterile ice-cold 50 mM calcium chloride. The cells were left on ice for 20 minutes and then harvested at 3000 x g. Finally, the cells were gently resuspended in 20 ml of sterile ice cold 100 mM calcium chloride containing 15 % glycerol. The cells were aliquoted in 0.5 ml portions and stored at \u201370 \u00b0C.\n\n### 2.2.9 Transformation\n\nAn aliquot of competent E. coli, strain XL1-Blue (Invitrogen), was thawed slowly at room temperature. The cells were gently mixed and three transformation reactions set up; experimental, negative control and positive control. For the experimental ligation 100 \u03bcl of competent cells were gently added to 5 \u03bcl of ligation reaction mixture as in Table 2-5. The cells were incubated on ice for 20 minutes and then 'heat shocked' at 42 \u00b0C for 90 seconds. The cells were returned to ice for 1 minute. After addition of 700 \u03bcl of LB media containing 20 mM sucrose, 10 mM magnesium sulphate and 10 mM magnesium chloride the culture was grown for 1 hour at 37 \u00b0C. The cells were harvested at 5000 x g for 2 minutes and 700 \u03bcl of the supernatant was removed. The cells were gently mixed and then plated out onto an LB-Amp plate. The plates were grown for 16 hours at 37 \u00b0C and the number of transformant colonies on each plate recorded. Positive and negative controls using 5 \u03bcl of pUC DNA (200 pg/\u03bcl) or 5 \u03bcl of water instead of ligation reaction mixture were prepared in a similar fashion.\n\n### 2.2.10 Screening of Transformants\n\nColonies grown on LB-Amp plates were isolated and used to inoculate 5 ml of LB-Amp media. The cultures were grown for 4 hours in 50 ml centrifuge tubes at 37 \u00b0C with shaking. For each culture 3 ml of cell suspension was removed and the cells harvested by centrifugation at 15000 x g into eppendorf tubes. The plasmid DNA was extracted using a Mini-Prep Plasmid Extraction Kit (Qiagen).\n\nExtracted plasmid samples were doubly digested with EcoRV and BglII to check for correct ligation. The vector contains a unique EcoRV site and the insert contains a unique BglII site. Restriction enzyme digests were set up as in Table 2-6 and the reaction incubated at 37 \u00b0C for 2 hours. The digested samples were run on a 1 % agarose-TAE gel.\n\n| Volume (\u03bcl) | Component |\n|---|---|\n| 4.00 | Extracted Plasmid DNA |\n| 1.00 | 10x Restriction Buffer B |\n| 0.50 | BglII (40 units/\u03bcl) |\n| 0.25 | EcoRV (40 units/\u03bcl) |\n| 4.25 | Sterile Distilled Water |\n| 10.00 | Total |\n\n*Table 2-6.* Table showing composition of restriction enzyme reaction mixture used to analyse pGEX-3C-OSM_\u0394185 and pGEX-3C-OSM_\u0394188 plasmids.\n\n### 2.2.11 Automated PCR Sequencing\n\nAutomated PCR sequencing was carried out using cycle sequencing with dye-labelled terminators. Sequencing reactions were setup using the ABI PRISM BigDye Terminator Cycle Sequencing Kit (PE Applied Biosystems) according to the recipe in Table 2-7. Three primers (5', 3' and internal to the insert) were used to ensure accurate reading through all areas of the insert (see Table 2-8). The BigDye reaction mixture contains all necessary components including BigDye terminators, dNTP's, AmpliTaq FS polymerase and magnesium chloride.\n\n| Component | 5' - Sequencing Reaction Volume (\u03bcl) | Internal - Sequencing Reaction Volume (\u03bcl) | 3' - Sequencing Reaction Volume (\u03bcl) |\n|---|---|---|---|\n| Plasmid DNA | 6.0 | 6.0 | 6.0 |\n| BigDye | 8.0 | 8.0 | 8.0 |\n| Primer | 1.0 | 1.0 | 1.0 |\n| Sterile water | 5.0 | 5.0 | 5.0 |\n| Total | 20.0 | 20.0 | 20.0 |\n\n*Table 2-7.* Table showing composition of reaction mixtures used for automated PCR sequencing.\n\n| Primer Description | Oligonucleotide Sequence |\n|---|---|\n| 5' pGEX Primer | 5'-GCA-TGG-CCT-TTG-CAG-GG-3' |\n| Internal OSM Primer | 5'-AGT-GAG-GAG-ACC-CTG-AGG-3' |\n| 3' pGEX Primer | 5'-TCC-GCT-TAC-AGA-CAA-GCT-G-3' |\n\n*Table 2-8.* Table showing oligonucleotides used for automated PCR sequencing.\n\nAmplification of DNA for sequencing was carried out using automated PCR cycling on a DNA Thermal Cycler 480 (Perkin Elmer Applied Biosystems). Denaturation of the double stranded DNA was carried out by incubation at 96 \u00b0C for 10 seconds. Annealing of primers to their complementary sequences was carried out by incubation for 5 seconds at 50 \u00b0C. Primer extension was carried out by incubation of the reaction mixture for 4 minutes at 60 \u00b0C. Twenty five cycles of denaturation-annealing-extension were carried out. The reaction cycles are summarised below in Table 2-9.\n\n| 1) Denaturation Temp / Time (sec) | 2) Annealing Temp / Time (sec) | 3) Extension Temp / Time (sec) | Cycles |\n|---|---|---|---|\n| 96 \u00b0C / 10 | 50 \u00b0C / 5 | 60 \u00b0C / 240 | 25 |\n\n*Table 2-9.* Table showing temperatures and incubation times used for automated PCR sequencing.", "# 2017 \u2013 Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: papers).*\n\n* Link to Paper\n\n* **MPER Structural Architecture**: The membrane proximal external region adopts a perpendicular orientation relative to the viral membrane during 10E8 engagement, creating an accessible target for neutralizing antibodies despite its proximity to the lipid bilayer.\n* **Dual Epitope Recognition**: 10E8 uniquely combines peptide and lipid recognition through coordinated interactions between its light chain and viral membrane phospholipids, particularly phosphatidylglycerol and phosphatidic acid.\n* **Germline Targeting Success**: Advanced vaccine strategies using epitope scaffolds on nanoparticle platforms have successfully activated 10E8-class antibody precursors in multiple animal models, providing proof-of-principle for translatable vaccine approaches.\n* **Clinical Development Challenges**: Early clinical trials revealed unexpected local reactogenicity with 10E8VLS, emphasizing the need for careful antibody engineering and safety evaluation in therapeutic development.\n* **Membrane-Focused Immunogen Design**: mRNA-delivered membrane-bound HIV envelope proteins show superior immunogenicity compared to soluble trimers, reducing off-target responses while enhancing neutralizing antibody induction.\n* **Bispecific Antibody Innovation**: Novel bispecific antibodies combining 10E8 specificity with complementary targets demonstrate enhanced neutralization breadth and potency, offering promising therapeutic strategies.\n\n1. **Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8**: [Irimia A, et al., PLOS Pathogens](https://journals.plos.org/plospathogens/article?id=10.1371/journal.ppat.1006212)\n2. **Structure of the membrane proximal external region of HIV-1**: [Chong H, et al., PNAS](https://www.pnas.org/doi/10.1073/pnas.1807259115)\n3. **Phase 1 trial evaluating safety and pharmacokinetics of HIV-1 broadly neutralizing antibody 10E8VLS**: [Coates EE, et al., JCI Insight](https://insight.jci.org/articles/view/175375)\n4. **Ab initio prediction of specific phospholipid complexes and membrane association of HIV-1 MPER antibodies**: [Maillie R, et al., eLife Sciences](https://elifesciences.org/articles/90139)\n5. **Vaccination induces broadly neutralizing antibody precursors to HIV gp41 MPER**: [Steichen JM, et al., Nature Immunology](https://www.nature.com/articles/s41590-024-01833-w)\n6. **Vaccination with mRNA-encoded membrane-bound HIV Envelope elicits neutralizing antibodies**: [Moyo N, et al., Nature Communications](https://pmc.ncbi.nlm.nih.gov/articles/PMC11785158/)\n\n**Unlocking HIV-1\u2019s Achilles\u2019 Heel: How Broadly Neutralizing Antibody 10E8 Revolutionizes Our Understanding of Viral Membrane Interactions.**\n\nThe battle against HIV-1 has long been hampered by the virus\u2019s extraordinary ability to evade immune recognition through rapid mutation and sophisticated camouflage strategies. However, recent breakthrough research has illuminated a critical vulnerability in HIV-1\u2019s armor: the membrane proximal external region (MPER) of the viral envelope protein gp41. At the forefront of this discovery is the broadly neutralizing antibody 10E8, which has emerged as one of the most potent weapons in our immunological arsenal, capable of neutralizing an unprecedented 98% of HIV-1 variants tested. This remarkable breadth stems from 10E8\u2019s unique dual recognition mechanism, simultaneously targeting both the conserved MPER peptide epitope and the viral membrane lipids themselves, creating a composite epitope that the virus cannot easily escape through mutation alone.\n\nThe significance of this discovery extends far beyond basic virology, offering transformative insights for both vaccine development and therapeutic design. Unlike previous broadly neutralizing antibodies that suffered from problematic cross-reactivity with human tissues, 10E8 demonstrates exceptional specificity for its viral targets while maintaining extraordinary neutralization potency. The antibody\u2019s sophisticated approach involves a precisely orchestrated molecular dance: its light chain engages specific phospholipid head groups on the viral membrane through a constellation of basic and polar residues, while its heavy chain simultaneously binds to the MPER epitope, which adopts an upright orientation approximately 75-80 degrees perpendicular to the viral membrane surface. This dual engagement creates a remarkably stable antibody-antigen complex that effectively blocks viral entry into host cells.\n\nRecent advances in structural biology and molecular dynamics simulations have revealed the atomic-level details of this interaction, showing how 10E8\u2019s complementarity determining regions CDRL1, CDRH3, and framework region FRL3 work in concert to create a specific lipid-binding site that recognizes phosphatidylglycerol and phosphatidic acid head groups. These lipids are particularly enriched in HIV-1 membranes compared to normal cellular membranes, providing an additional layer of viral specificity. The antibody approaches its target at an angle of approximately 43 degrees from the membrane surface, positioning its variable regions to maximize both peptide and lipid interactions while minimizing potential interference from other viral or cellular components.", "# 2012 \u2013 Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nContemporary structural studies have expanded beyond the AP33-E2 interaction to encompass the complete E1E2 heterodimer complex, revealing additional sites of vulnerability for therapeutic targeting. Cryo-electron microscopy structures of full-length membrane-extracted E1E2 complexes have defined the overall architecture of the viral envelope and identified novel broadly neutralizing epitopes. These structural advances have enabled the development of improved glycoengineered E2 variants that enhance antibody binding while maintaining native-like conformations. The incorporation of specific glycan modifications has been shown to improve both the homogeneity and immunogenicity of vaccine candidates, with certain glycoforms demonstrating enhanced neutralizing antibody responses in preclinical studies.\n\nThe therapeutic implications of these structural discoveries extend beyond prophylactic vaccination to include therapeutic antibody development and small molecule inhibitor design. The detailed molecular understanding of antibody-epitope interactions has facilitated the engineering of improved neutralizing antibodies with enhanced breadth and potency. Additionally, structural insights into the CD81 receptor binding site and its relationship to neutralizing epitopes have informed the development of entry inhibitors that could complement direct-acting antiviral therapies. The convergent evolution of multiple antibodies targeting the same epitope through different structural approaches validates the conservation and accessibility of these sites in the context of viral infection.\n\n| Key Concept | Description | Key References |\n| --- | --- | --- |\n| **\u03b2-Hairpin Epitope Structure** | Conserved structural motif in HCV E2 residues 412-423 that forms type I\u2032 \u03b2-turn recognized by broadly neutralizing antibodies | [Kong, L., et al., Journal of Virology, 2012](https://journals.asm.org/doi/10.1128/jvi.01939-12) |\n| **Dual Recognition Mechanism** | Two different antibody binding modes (AP33 and HCV1) targeting identical epitope with 22\u00b0 angular difference | [Stanfield, R.L., et al., Journal of Virology, 2012](https://pmc.ncbi.nlm.nih.gov/articles/PMC3497658/) |\n| **Critical Binding Residues** | Leu413, Asn415, Gly418, and Trp420 form essential molecular contacts for neutralizing antibody recognition | [Bailey, J.R., et al., Journal of Virology, 2012](https://pmc.ncbi.nlm.nih.gov/articles/PMC3497650/) |\n| **Epitope-Focused Immunogens** | Nanoparticle-based vaccine candidates displaying conserved epitopes to induce broadly neutralizing antibodies | [Krey, T., et al., Science Advances, 2024](https://www.science.org/doi/10.1126/sciadv.ado2600) |\n| **Viral Escape Mechanisms** | Mutations N415Y, N415D, and G418D disrupt antibody binding while maintaining viral fitness and infectivity | [Bailey, J.R., et al., Vaccines, 2024](https://pmc.ncbi.nlm.nih.gov/articles/PMC11125608/) |\n| **Glycoengineering Approaches** | Modified E2 glycoforms with altered N-linked glycosylation patterns enhance immunogenicity and antibody binding | [Ringe, R., et al., npj Vaccines, 2025](https://www.nature.com/articles/s41541-025-01161-6) |", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\nProvided herein are anti-mutCALR antibodies and related nucleic acids, expression vectors, cells, kits, and pharmaceutical compositions. The anti-mutCALR antibodies described herein are useful in the treatment or prevention or diagnosis of myeloproliferative neoplasms (e.g., chronic myelogenous leukemia, polycythemia vera, primary myelofibrosis, essential thrombocythemia, chronic neutrophilic leukemia, acute myelogenus leukemia, chronic eosinophilic leukemia, chronic myelomonocytic leukemia, myeloproliferative neoplasm and myelodysplastic syndrome, including myelodysplastic syndrome with refractory anaemia with ring sideroblasts, myelodysplastic syndrome with refractory anemia, and myelodysplastic syndrome with refractory anemia with excess blasts.).\n\n### Mutant CALR and Anti-mutCALR Antibodies\n\nCALR is a highly conserved chaperone protein that resides primarily in the endoplasmic reticulum and is involved in a variety of cellular processes including protein folding, calcium homeostasis, cell adhesion, and integrin signaling. Mutations in the *CALR* gene have been identified in patients with myeloproliferative neoplasms. The two most frequent CALR mutations are a 52 base pair (bp) deletion and a 5 bp insertion, which are referred to as Type 1 and Type 2 mutations, respectively. Type 1 and Type 2 mutations cause a +1 frameshift within exon 9 that generates a novel, positively-charged C-terminal amino acid sequence that lacks the KDEL domain (SEQ ID NO:347) of the WT protein, thereby enabling the mutCALR to escape the ER and activate the thrombopoietin receptor (MPL) and induce constitutive activation of Janus kinase 2 (JAK2) signaling. The amino acid sequences of human WT CALR, Type 1 and Type 2 mutCALR, and the novel C-terminal sequence are shown below. The frameshift amino acid residues in Type 1 and Type 2 mutCALR are shown in bold and the novel C-terminal sequence is marked by underlining.\n\n**Human WT CALR (GenBank Accession No. NP_004334.1)** (SEQ ID NO:319):\n\n```\nMLLSVPLLLGLLGLAVAEPAVYFKEQFLDGDGWTSRWIESKHKSDFGKFVLSSGKFYGDEEKDKGLQT\nSQDARFYALSASFEPFSNKGQTLVVQFTVKHEQNIDCGGGYVKLFPNSLDQTDMHGDSEYNIMFGPDI\nCGPGTKKVHVIFNYKGKNVLINKDIRCKDDEFTHLYTLIVRPDNTYEVKIDNSQVESGSLEDDWDFLP\nPKKIKDPDASKPEDWDERAKIDDPTDSKPEDWDKPEHIPDPDAKKPEDWDEEMDGEWEPPVIQNPEYK\nGEWKPRQIDNPDYKGTWIHPEIDNPEYSPDPSIYAYDNFGVLGLDLWQVKSGTIFDNFLITNDEAYAE\nEFGNETWGVTKAAEKQMKDKQDEEQRLKEEEEDKKRKEEEEAEDKEDDEDKDEDEEDEEDKEEDEEED\nVPGQAKDEL\n```\n\n**Human Type 1 mutCALR** (SEQ ID NO:320):\n\n```\nMLLSVPLLLGLLGLAVAEPAVYFKEQFLDGDGWTSRWIESKHKSDFGKFVLSSGKFYGDEEKDKGLQT\nSQDARFYALSASFEPFSNKGQTLVVQFTVKHEQNIDCGGGYVKLFPNSLDQTDMHGDSEYNIMFGPDI\nCGPGTKKVHVIFNYKGKNVLINKDIRCKDDEFTHLYTLIVRPDNTYEVKIDNSQVESGSLEDDWDFLP\nPKKIKDPDASKPEDWDERAKIDDPTDSKPEDWDKPEHIPDPDAKKPEDWDEEMDGEWEPPVIQNPEYK\nGEWKPRQIDNPDYKGTWIHPEIDNPEYSPDPSIYAYDNFGVLGLDLWQVKSGTIFDNFLITNDEAYAE\nEFGNETWGVTKAAEKQMKDKQDEEQRTRRMMRTKMRMRRMRRTRRKMRRKMSPARPRTSCREACLQGW\nTEA\n```\n\n**Human Type 2 mutCALR** (SEQ ID NO:321), whose novel C-terminal sequence reads\n`...QRLKEEEEDKKRKEEEEAEDNCRRMMRTKMRMRRMRRTRRKMRRKMSPARPRTSCREACLQGWTEA`.\n\n**mutCALR C-terminal consensus mutant sequence** (SEQ ID NO:322):\n\n```\nRMRRMRRTRRKMRRKMSPARPRTSCREACLQGWTEA\n```\n\nThis disclosure provides anti-mutCALR antibodies that are useful in treating or diagnosing myeloproliferative neoplasms. Note that in the present disclosure, unless stated otherwise, amino acid positions assigned to CDRs and frameworks in a variable region of the anti-mutCALR antibodies are specified according to Kabat; see EA Kabat, Sequences of Proteins of Immunological Interest, U.S. Dept. of Health and Human Services, Public Health Service, National Institutes of Health, 1991, (OCoLC)1138727707.\n\nIn some embodiments, the anti-mutCALR antibody is an anti-mutCALR antibody that comprises one, two, three, four, five, and/or six CDRs of any one of the antibodies described herein. In some embodiments, an anti-mutCALR antibody comprises (i) one, two, and/or three heavy chain CDRs of any one of the clones presented in Tables 1-2, and/or (ii) one, two, and/or three light chain CDRs from any one of the clones presented in Tables 1-2. In some embodiments, an anti-mutCALR antibody comprises (i) three heavy chain CDRs from any one of the clones presented in Tables 4-5, and (ii) three light chain CDRs from any one of the clones presented in Tables 4-5.\n\nIn some embodiments, an anti-mutCALR antibody comprises a heavy chain CDR1, CDR2, and CDR3 and/or a light chain variable region CDR1, CDR2, and CDR3 from an antibody described herein. In some embodiments, an anti-mutCALR antibody comprises a mouse version, mouse variant, human version, human variant, humanized version, humanized variant, or affinity matured variant of an antibody described herein.\n\nIn some embodiments, the anti-mutCALR antibody is a variant of an anti-mutCALR antibody described herein which comprises one to thirty conservative amino acid substitution(s), e.g., one to twenty-five, one to twenty, one to fifteen, one to ten, one to five, or one to three conservative amino acid substitution(s). In some embodiments, the conservative amino acid substitution(s) is in a CDR of the antibody. In some embodiments, the conservative amino acid substitution(s) is not in a CDR of the antibody. In some embodiments, the conservative amino acid substitution(s) is in a framework region of the antibody. In some embodiments, a CDR comprises one, two, three, or four amino acid substitutions. In some embodiments, the one or more substitutions are made as part of a humanization process, a germline humanization process, an affinity maturation process, or an optimization process.", "# GitHub README: TopPDBLX\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/TopPDBLX, 2026\n\n| | |\n|---|---|\n| Weights | **[`Dellboy/toppdblx-construct-boundary`](https://huggingface.co/Dellboy/toppdblx-construct-boundary)** on HuggingFace, 134 MB |\n| Base model | [ESM-2 t12-35M](https://huggingface.co/facebook/esm2_t12_35M_UR50D), 33.5M parameters |\n| Input | Full-length UniProt sequence, up to 1,022 residues |\n| Output | Per-residue probability of being inside a crystallised construct |\n| Trained on | 34,874 proteins, from 523,018 deposited chains across 185,831 PDB entries |\n| Held out | 4,077 proteins, split by 30% sequence identity so no homologue is on both sides |\n\n### What it achieves\n\n| Measure | Test set | Why this measure |\n|---|---|---|\n| **Boundary error** | **9 residues** (median) | How far off the cut is, on the 1,314 test proteins that were genuinely truncated. A median: the spread matters and is below |\n| **MCC** | **0.669** | Correlation between predicted and real per-residue calls. A model that says \"keep everything\" scores 0.00 |\n| Accuracy | 85.7% | Reported for completeness and **should not be read alone**: 61.5% of residues really are inside a construct, so \"keep everything\" already scores 61.5% while being useless |\n\n**A worked example.** Hen lysozyme (`P00698`) is 147 residues in UniProt, of which 1 to 18 are\nthe signal peptide and 19 to 147 the mature chain. Asked cold, the model proposes **19 to 147**.\nNobody told it what a signal peptide is; it learned that crystallographers do not clone them.\n\n**But read the spread before trusting a single number**, because the median hides the shape of it:\n\n| | boundary error |\n|---|---|\n| Half the boundaries | within **5 residues** |\n| Three quarters | within 56 residues |\n| Nine tenths | within 250 residues |\n\n**It is excellent on most proteins and badly wrong on a minority.** 60% of boundaries land within\n10 residues, and 56% of proteins have *both* ends within 25. The mean error of 72 residues is the\ntail talking, not the typical case.\n\n**The model half knows when it is wrong.** Mean probability across the proposed span runs 0.97 on\ngood predictions and 0.82 on bad ones, so `model.propose_boundaries` returns a `confident` flag at\n0.85. Confident proposals get both ends within 25 residues **71%** of the time against 56%\nungated, and cover 68% of proteins. Treat a low-confidence span as a hint and look at the profile.\n\nFor context, truncated constructs here trim a median of 62 residues from the N-terminus and keep\nunder half the chain, so being within 5 residues is a small fraction of the decision, and being\n250 out is not a usable answer.\n\n### Training rounds\n\nAll figures are the **held-out test split** (4,077 proteins, 1,445 of them truncated), never the\nvalidation split used to pick checkpoints. That distinction decided which model ships.\n\n| Round | What changed | MCC | Boundary error | coverage@3 | Verdict |\n|---|---|---|---|---|---|\n| [**01**](https://huggingface.co/Dellboy/toppdblx-construct-boundary) | ESM-2 t12-35M, 3 epochs, token-bucketed batches, consensus labels at 50% agreement | **0.669** | **9 residues** | **36.1%** | **Shipped.** Best on every metric, and still is |\n| 02 | Soft targets everywhere: train on the fraction of a protein's constructs covering each residue | 0.661 \u2020 | 11 \u2020 | 30.6% \u2020 | **Abandoned at epoch 3.** Validation metrics flat while training loss fell |\n| 03 | Soft targets gated to proteins with at least 10 deposited constructs, 6 epochs | 0.659 | 10 | 35.6% | **Rejected.** Worse than round 01 on all five metrics |\n| 04 | No soft targets, round 01's recipe run for 6 epochs instead of 3 | 0.662 | 10 | 35.8% | **Rejected.** Led on validation, lost on test |\n| 05 | **ESM-2 t30-150M**, 4x the parameters, otherwise round 01's recipe | 0.673 | 10 | 34.9% | **Rejected.** A dead heat on MCC (+0.004) and worse on the other four measures |\n\n\u2020 round 02 was stopped early, so its figures are validation-only and not comparable to the rest.\n\n**The campaign found nothing, and how it failed is the useful part.** On the 519-protein validation\nsplit round 04 read coverage@3 41.8% against round 01's 40.3%, and that looked like a result. On\nthe 1,445-protein test split it reverses to 35.8% against 36.1%. Round 03's dramatic jump from\n34.1% to 41.0% at epoch 5, and round 04 matching it, were both inside the variance of the smaller\nsample.\n\n**Round 01 has now been challenged five ways and survived all of them**, which is a better argument\nfor it than the original result was:\n\n| Attempted | Outcome |\n|---|---|\n| Six structural features (coil, pLDDT, Pfam edges, disorder, disorder as a channel, surface entropy) | None help |\n| Soft targets, everywhere and gated | Do not help |\n| Six epochs instead of three | Does not help |\n| **A 4x larger backbone** | **Does not help** |\n\n**The task has a ceiling near MCC 0.67 and a 9-residue median boundary error, and it is not a\ncapacity, schedule or feature-engineering problem.** The plateau is in the labels. Where a\ncrystallographer cuts carries real signal, and the model already extracts most of it, but the\ndecision is also part convention, part whichever vector was to hand, and part arbitrary. None of\nthat is recoverable from sequence.\n\nWorth stating for anyone tempted to repeat this: the 150M model matched the 35M model to within\n0.004 MCC while costing four times the parameters and four times the training time.\n\nPer epoch, on validation: MCC 0.607 then 0.681 then 0.700, boundary error 14 then 9 then 7\nresidues. That upward trend suggested a longer run would help. **Round 04 tested it directly and it\ndid not**: six epochs scored marginally worse on held-out data than three.\n\n**The failure condition was written before the run, not after.** MCC below 0.40 or a median\nboundary error worse than 20 residues would have meant the signal is not learnable from sequence\nat this scale, and R3 would have stopped. Declaring it in advance is what stops a marginal result\nbeing tuned until it looks like a good one.\n\n### Turning a probability into a construct", "# ORCID record of Marc C. Deller\nAuthors: Deller, M.C.\nSource: ORCID, 2026\n\n- The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase (2017). ACS Chemical Biology. journal-article. DOI: 10.1021/acschembio.7b00016\n- Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design (2017). PLoS Pathogens. journal-article. DOI: 10.1371/journal.ppat.1006212\n- Protein stability: A crystallographer's perspective (2016). Acta Crystallographica Section:F Structural Biology Communications. journal-article. DOI: 10.1107/S2053230X15024619\n- Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor (2016). Immunity. journal-article. DOI: 10.1016/j.immuni.2016.03.006\n- Structure-based discovery of NANOG variant with enhanced properties to promote self-renewal and reprogramming of pluripotent stem cells (2015). Proceedings of the National Academy of Sciences of the United States of America. journal-article. DOI: 10.1073/pnas.1502855112\n- Models of protein-ligand crystal structures: Trust, but verify (2015). Journal of Computer-Aided Molecular Design. journal-article. DOI: 10.1007/s10822-015-9833-8\n- Crystal structure of a two-subunit TrkA octameric gating ring assembly (2015). PLoS ONE. journal-article. DOI: 10.1371/journal.pone.0122512\n- Complete epitopes for vaccine design derived from a crystal structure of the broadly neutralizing antibodies PGT128 and 8ANC195 in complex with an HIV-1 Env trimer (2015). Acta Crystallographica Section D: Biological Crystallography. journal-article. DOI: 10.1107/S1399004715013917\n- Structural insights into the recognition of phosphopeptide by the FHA domain of Kanadaptin (2014). PLoS ONE. journal-article. DOI: 10.1371/journal.pone.0107309\n- Approaches to automated protein crystal harvesting (2014). Acta Crystallographica Section F:Structural Biology Communications. journal-article. DOI: 10.1107/S2053230X14000387\n- Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120 (2013). Nature Structural and Molecular Biology. journal-article. DOI: 10.1038/nsmb.2594\n- Crystal structure of a soluble cleaved HIV-1 envelope trimer (2013). Science. journal-article. DOI: 10.1126/science.1245625\n- Structure of hepatitis C virus envelope glycoprotein E2 antigenic site 412 to 423 in complex with antibody AP33 (2012). Journal of Virology. journal-article. DOI: 10.1128/JVI.01939-12\n- Functional and structural characterization of a thermostable acetyl esterase from Thermotoga maritima (2012). Proteins: Structure, Function and Bioinformatics. journal-article. DOI: 10.1002/prot.24041\n- Crystal structure of a metal-dependent phosphoesterase (YP_910028.1) from Bifidobacterium adolescentis: Computational prediction and experimental validation of phosphoesterase activity (2011). Proteins: Structure, Function and Bioinformatics. journal-article. DOI: 10.1002/prot.23035\n- The structure of the first representative of Pfam family PF09836 reveals a two-domain organization and suggests involvement in transcriptional regulation (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309109022672\n- The structure of the first representative of Pfam family PF06475 reveals a new fold with possible involvement in glycolipid metabolism (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309109022684\n- The structure of SSO2064, the first representative of Pfam family PF01796, reveals a novel two-domain zinc-ribbon OB-fold architecture with a potential acyl-CoA-binding role (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309110002514\n- The structure of KPN03535 (gi|152972051), a novel putative lipoprotein from Klebsiella pneumoniae, reveals an OB-fold (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309109018168\n- The structure of Jann-2411 (DUF1470) from Jannaschia sp. at 1.45 \u00c5 resolution reveals a new fold (the ABATE domain) and suggests its possible role as a transcription regulator (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309109025196\n- The structure of Haemophilus influenzae prephenate dehydrogenase suggests unique features of bifunctional TyrA enzymes (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309110021688\n- The structure of BVU2987 from Bacteroides vulgatus reveals a superfamily of bacterial periplasmic proteins with possible inhibitory function (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309109046788\n- The crystal structure of a bacterial Sufu-like protein defines a novel group of bacterial proteins that are similar to the N-terminal domain of human Sufu (2010). Protein Science. journal-article. DOI: 10.1002/pro.497\n- Structures of three members of Pfam PF02663 (FmdE) implicated in microbial methanogenesis reveal a conserved \u03b1+\u0392 core domain and an auxiliary C-terminal treble-clef zinc finger (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309110020166\n- Structures of the first representatives of Pfam family PF06938 (DUF1285) reveal a new fold with repeated structural motifs and possible involvement in signal transduction (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309109050416\n- Structures of the first representatives of Pfam family PF06684 (DUF1185) reveal a novel variant of the Bacillus chorismate mutase fold and suggest a role in amino-acid metabolism (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309109050647\n- Structure of the \u03b3-D-glutamyl-L-diamino acid endopeptidase YkfC from Bacillus cereus in complex with L-Ala-\u03b3-d-Glu: Insights into substrate recognition by NlpC/P60 cysteine peptidases (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309110021214\n- Structure of the first representative of Pfam family PF09410 (DUF2006) reveals a structural signature of the calycin superfamily that suggests a role in lipid metabolism (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309109037749\n- Structure of the first representative of Pfam family PF04016 (DUF364) reveals enolase and Rossmann-like folds that combine to form a unique active site with a possible role in heavy-metal chelation (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309110007517\n- Structure of a putative NTP pyrophosphohydrolase: YP-001813558.1 from Exiguobacterium sibiricum 255-15 (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309110025534\n- Structure of a membrane-attack complex/perforin (MACPF) family protein from the human gut symbiont Bacteroides thetaiotaomicron (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309110023055\n- Structure of LP2179, the first representative of Pfam family PF08866, suggests a new fold with a role in amino-acid metabolism (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309109023689\n- Structure of Bacteroides thetaiotaomicron BT2081 at 2.05 \u00c5 resolution: The first structural representative of a new protein family that may play a role in carbohydrate metabolism (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309110028228\n- Structure of BT-3984, a member of the SusD/RagB family of nutrient-binding molecules (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309110032999\n- Open and closed conformations of two SpoIIAA-like proteins (YP-749275.1 and YP-001095227.1) provide insights into membrane association and ligand binding (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309109042481\n- Crystal Structure of the First Eubacterial Mre11 Nuclease Reveals Novel Features that May Discriminate Substrates During DNA Repair (2010). Journal of Molecular Biology. journal-article. DOI: 10.1016/j.jmb.2010.01.049\n- Conformational changes associated with the binding of zinc acetate at the putative active site of XcTcmJ, a cupin from Xanthomonas campestris pv. campestris (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309109021988\n- Bacterial Pleckstrin Homology Domains: A Prokaryotic Origin for the PH Domain (2010). Journal of Molecular Biology. journal-article. DOI: 10.1016/j.jmb.2009.11.006\n- A conserved fold for fimbrial components revealed by the crystal structure of a putative fimbrial assembly protein (BT1062) from Bacteroides thetaiotaomicron at 2.2 \u00c5 resolution (2010). Acta Crystallographica Section F: Structural Biology and Crystallization Communications. journal-article. DOI: 10.1107/S1744309110006548\n- Structural and functional characterizations of SsgB, a conserved activator of developmental cell division in morphologically complex actinomycetes (2009). Journal of Biological Chemistry. journal-article. DOI: 10.1074/jbc.M109.018564\n- Structural Basis of Murein Peptide Specificity of a \u03b3-D-Glutamyl-L-Diamino Acid Endopeptidase (2009). Structure. journal-article. DOI: 10.1016/j.str.2008.12.008\n- Crystal structure of the Fic (filamentation induced by cAMP) family protein SO4266 (gi|24375750) from Shewanella oneidensis MR-1 at 1.6 \u00c5 resolution (2009). Proteins: Structure, Function and Bioinformatics. journal-article. DOI: 10.1002/prot.22338\n- Crystal structure of a novel Sm-like protein of putative cyanophage origin at 2.60 A resolution (2009). Proteins: Structure, Function and Bioinformatics. journal-article. DOI: 10.1002/prot.22360\n- Crystal Structure of Histidine Phosphotransfer Protein ShpA, an Essential Regulator of Stalk Biogenesis in Caulobacter crescentus (2009). Journal of Molecular Biology. journal-article. DOI: 10.1016/j.jmb.2009.05.023\n- A Structural Basis for the Regulatory Inactivation of DnaA (2009). Journal of Molecular Biology. journal-article. DOI: 10.1016/j.jmb.2008.10.059\n- Crystal structures of MW1337R and lin2004: Representatives of a novel protein family that adopt a four-helical bundle fold (2008). Proteins: Structure, Function and Genetics. journal-article. DOI: 10.1002/prot.22020\n- Identification and structural characterization of heme binding in a novel dye-decolorizing peroxidase, TyrA (2007). Proteins: Structure, Function and Genetics. journal-article. DOI: 10.1002/prot.21673\n- Crystal structures of two novel dye-decolorizing peroxidases reveal a \u03b2-barrel fold with a conserved heme-binding motif (2007). Proteins: Structure, Function and Genetics. journal-article. DOI: 10.1002/prot.21550\n- Crystal structure of homoserine O-succinyltransferase from Bacillus cereus at 2.4 \u00c5 resolution (2007). Proteins: Structure, Function and Genetics. journal-article. DOI: 10.1002/prot.21208\n- Crystal structure and functional dissection of the cytostatic cytokine oncostatin M (2000). Structure. journal-article. DOI: 10.1016/S0969-2126(00)00176-3\n- Cell surface receptors (2000). Current Opinion in Structural Biology. journal-article. DOI: 10.1016/S0959-440X(00)00072-5", "# Crystal Structure of the Complex of 10E8 Fab Light Chain Mutant1 and T117v2 HIV-1 MPER Scaffold\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [5SY8](https://www.rcsb.org/structure/5SY8)\n* **Release Date:** 2017-03-08\n* **Method:** X-RAY DIFFRACTION 1.62 \u00c5\n* **Organisms:** Homo sapiens, synthetic construct\n* **Macromolecules:** 10E8 EPITOPE SCAFFOLD T117v2, 10E8 FAB HEAVY CHAIN, 10E8 FAB LIGHT CHAIN (proteins)\n* **Unique Ligands:** GOL\n* **Publication:** [Irimia, A., et al. (2017). Lipid Interactions and Angle of Approach to the HIV-1 Viral Membrane of Broadly Neutralizing Antibody 10E8: Insights for Vaccine and Therapeutic Design. PLoS Pathogens, 13, e1006212.](https://journals.plos.org/plospathogens/article?id=10.1371/journal.ppat.1006212)\n* **Blog Post:** [Lipid Interactions and Angle of Approach to the HIV-1 Viral Membrane of Broadly Neutralizing Antibody 10E8](https://marcdeller.com/lipid-interactions-and-angle-of-approach-to-the-hiv-1-viral-membrane-of-broadly-neutralizing-antibody-10e8-insights-for-vaccine-and-therapeutic-design/)\n\n**Description:** This structure reveals how the broadly neutralizing antibody 10E8 engages both the HIV-1 MPER peptide and viral membrane lipids through a dual-recognition mechanism. The 1.62 \u00c5 resolution structure maps specific lipid head group interaction sites and defines the MPER\u2019s perpendicular orientation relative to the membrane during neutralization, providing essential blueprints for vaccine immunogen and therapeutic antibody design.", "# JCSG Top96: A Rationally Designed Crystallization Screen Based on over 1000 Deposited Protein Structures\nAuthors: Deller, M. C.; Clayton, T.; Elsliger, M. A.; Wilson, I. A.\nSource: American Crystallographic Association (ACA) Annual Meeting 2013 (poster), 2013\n\n**Authors:** Marc C. Deller, Tom Clayton, Marc A. Elsliger, Ian A. Wilson. **Affiliations (condensed):** Joint Center for Structural Genomics, The Scripps Research Institute, La Jolla, CA 92037, USA. E-mail: mdeller@scripps.edu. **Presented:** American Crystallographic Association Annual Meeting, 2013 (poster). The JCSG is supported by NIH, National Institute of General Medical Sciences, Award Number U54 GM094586.\n\n## Abstract\n\nLarge scale high-throughput structural biology and structural genomics biology (HT-SG) centers, such as the Joint Center for Structural Genomics (JCSG), present a unique opportunity for the analysis of large systematic datasets of crystallization conditions for a diverse set of protein targets from both bacterial and mammalian origin. The crystallization pipeline developed at the JCSG has resulted in over 165,000 screened crystals and close to 1400 unique protein structures to date. Analysis of these data (07/1/2000 \u2013 9/23/2010) reveals a high level of redundancy in crystallization conditions with 55% (631/1144 deposited JCSG structures) of the structures coming from 25% of the conditions routinely used.\n\nPrevious rationally designed crystallization screens, such as our own JCSG+, have focused mainly on the propensity to grow crystals, but not on the analysis of the diffraction properties of the resulting crystals. Using data from 1144 deposited JCSG protein structures, a new minimal 96-well structure-based crystallization screen has been designed \u2013 JCSG Top96 \u2013 incorporating the 96 most successful crystallization conditions with respect to PDB deposition. Several trends within this new screening set are identified including the above average performance of PEG in combination with other precipitants and the below average performance of single precipitant conditions. To further build on this screening strategy, the most successful cryo-protectant for each crystallization condition was identified and the JCSG Top96 Cryo screen was formulated. The JCSG Top96 and JCSG Top96 Cryo screens are particularly suitable for proteins available in limited quantities where conservation of material is of high importance. These screens can be used by both large-scale HT-SG platforms and smaller-scale studies in individual laboratories. This high efficiency crystallization screen will help reduce costs and will also enable evermore challenging problems in biological and biomedical research to be addressed using X-ray crystallography.\n\n## Introduction\n\nThe JCSG Top96 screen is a novel single plate crystallization screen collated from 10 years of PDB depositions by the JCSG. The screen is a follow up to previous JCSG screens such as the JCSG Core Suite and JCSG+.\n\n- JCSG Top96 conditions available at: http://www.jcsg.org/top96\n- Large scale HT-SG structure based screen.\n- Minimal screen, increased efficiency, saves money and protein.\n- Data mined from: 5.4 million crystallization experiments; 145,000 crystals; 1144 JCSG protein structures deposited in PDB (Fig. 1).\n- 55% of JCSG structures crystallize in 96 conditions (Fig. 1).\n- ~4-fold better performance than JCSG+, only 16 conditions in common; JCSG+ 168 PDB, JCSG Top96 631 PDB.\n- Conditions organized in plate according to precipitant and concentration (Fig. 2).\n- 7 precipitant classes (Fig. 3 and Fig. 4).\n- Above average performance of PEG combinations (Fig. 4).\n- Enrichment of PEG combination precipitants (Fig. 5).\n- Design of JCSG Top96 Cryo from diffraction/PDB data (Fig. 6).\n\n## Methods\n\n### JCSG Top96 Design\n\nThe 1144 JCSG PDB depositions (1101 of which came from JCSG Core Suite conditions) were ranked by screen and by the number of PDB depositions attributable to each condition, and the resulting list was sorted to select the best 96 conditions. The JCSG Core Suite conditions were analysed both by block and by PDB deposition count.\n\n### JCSG Top96 Plate Layout\n\nThe selected conditions were organized in a single 96-well plate according to precipitant class and precipitant concentration. Wells of the plate can be colored according to PDB counts or according to precipitant class.\n\n### JCSG Top96 Cryo\n\nUsing PDB deposition data, the cryo-protectant that most frequently gives a clear freeze (assessed both visually and by diffraction) was determined for each of the JCSG Top96 conditions, using the cryo-protectants in routine use at the JCSG, and these were formulated as the JCSG Top96 Cryo screen.\n\n## Results\n\n### JCSG Top96 Precipitant Class\n\nThe JCSG Top96 screen comprises seven precipitant classes, formed by the overlap of PEG, salt and organic precipitants (single-precipitant classes and their pairwise and three-way combinations).\n\n### JCSG Top96 Performance\n\nComparison of PDB counts and JCSG Top96 counts for the seven precipitant classes, and their difference from the mean, shows above average performance for all PEG combination precipitant classes and below average performance for single precipitant conditions.\n\n### JCSG Top96 Enrichment\n\nRelative to the JCSG Core Suite, PEG combination precipitant classes are enriched in the JCSG Top96 and JCSG Top96 Cryo screens at the expense of the single precipitant classes.\n\n## Figure captions\n\n*Figure 1.* Selection of JCSG Top96 conditions. (a) 1144 JCSG PDBs were ranked by screen (1101 from JCSG Core Suite conditions) and (b) sorted to select the best 96. Panels show JCSG Core Suite conditions by block and JCSG Core Suite conditions by PDB.\n\n*Figure 2.* JCSG Top96 plate layout. (a\u2013d) PDB depositions from each of the JCSG Core Suite blocks and (e) JCSG Top96, well colored according to PDB counts, (f) JCSG Top96, wells colored by precipitant class.\n\n*Figure 3.* Seven precipitant classes of the JCSG Top96 screen. Proportional Venn diagram showing the overlap of PEG (green), salt (blue) and organic (red) precipitants. Numbers in white represent PDB depositions.", "# Revolutionizing Drug Discovery: The Click2Drug Platform and the Rise of AI-Powered Virtual Screening\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-28 on marcdeller.com (categories: ai).*\n\n* **Comprehensive Resource Integration:** Click2Drug bridges computational tools across the entire drug discovery pipeline through systematic categorization and curation\n* **Ultra-Scale Virtual Screening:** Modern platforms process billion-compound libraries in days rather than years, achieving 14-44% hit rates with AI-guided approaches\n* **Deep Learning Revolution:** Pharmacophore modeling and molecular docking have been transformed through neural networks that understand molecular structure-activity relationships\n* **Predictive Accuracy Breakthrough:** AI-designed compounds demonstrate 80-90% Phase I success rates compared to 40-65% for traditional discovery approaches\n* **Democratic Access Expansion:** Open-source platforms and cloud computing have democratized sophisticated drug discovery tools for global research communities\n\n* [**Click2Drug**](https://www.click2drug.org/directory_SmallMoleculesDatabase.php)\n* **Structure and dynamics in drug discovery:** [Shoichet, B.K., et al., Nature](https://www.nature.com/articles/s44386-024-00001-2) (2024)\n* **An artificial intelligence accelerated virtual screening platform for drug discovery:** [Meller, A., et al., Nature Communications](https://www.nature.com/articles/s41467-024-52061-7) (2024)\n* **Perspectives on current approaches to virtual screening in drug discovery:** [Muegge, I., et al., Expert Opinion on Drug Discovery](https://pubmed.ncbi.nlm.nih.gov/39132881/) (2024)\n* **Recent advances from computer-aided drug design to artificial intelligence drug design:** [Chen, L., et al., MedChemComm](https://pubs.rsc.org/en/content/articlelanding/2024/md/d4md00522h) (2024)\n* **Deep learning pipeline for accelerating virtual screening in drug discovery:** [Noor, F., et al., Scientific Reports](https://www.nature.com/articles/s41598-024-79799-w) (2024)\n* **Rapid traversal of vast chemical space using machine learning:** [Hoffman, S.C., et al., Nature Machine Intelligence](https://www.nature.com/articles/s43588-025-00777-x) (2025)\n\nThe landscape of pharmaceutical research is experiencing an unprecedented transformation through the integration of artificial intelligence and computational methodologies. At the forefront of this revolution stands Click2Drug, a comprehensive directory that has become instrumental in democratizing access to computer-aided drug design tools, while simultaneously witnessing the emergence of billion-scale virtual screening platforms that promise to reshape how we discover life-saving medications.\n\nThe traditional drug discovery process, notorious for its decade-long timelines and multi-billion dollar costs, faces mounting pressure to evolve. With success rates hovering around 10% and development costs exceeding $2.6 billion per approved drug, the pharmaceutical industry desperately needs innovative approaches that can accelerate discovery while maintaining scientific rigor. This challenge has catalyzed the development of sophisticated computational platforms that leverage machine learning, deep learning, and artificial intelligence to transform every stage of the drug discovery pipeline.\n\nClick2Drug, developed by the Swiss Institute of Bioinformatics, represents a pivotal resource in this computational revolution. Housing over 807 meticulously categorized links to software, databases, and web services, the platform covers the entire spectrum of drug design activities\u2014from initial molecular visualization and target identification to final ADME toxicity prediction. What distinguishes Click2Drug from conventional software repositories is its structured approach to organizing computational tools across thirteen distinct categories, including molecular modeling, binding site prediction, virtual screening, and pharmacophore-based design. This taxonomical organization enables researchers to navigate the complex landscape of available tools efficiently, whether they are academic investigators with limited budgets or industry professionals seeking specialized solutions.\n\nRecent developments in 2024 and 2025 have witnessed remarkable advances in ultra-large virtual screening capabilities. The emergence of platforms like RosettaVS demonstrates the potential of AI-accelerated virtual screening to process multi-billion compound libraries with unprecedented accuracy. Unlike traditional docking approaches that maintain rigid protein structures, these advanced platforms incorporate receptor flexibility and sophisticated machine learning algorithms to predict binding poses and affinities more accurately. RosettaVS has successfully demonstrated its capabilities by screening billions of compounds against diverse targets, achieving hit rates of 14-44% while completing entire campaigns in less than seven days.\n\nThe integration of deep learning methodologies has fundamentally altered the virtual screening paradigm. PharmacoNet represents a groundbreaking approach to pharmacophore modeling, utilizing deep learning frameworks to perform ultra-fast virtual screening without requiring explicit binding conformations. This platform successfully identified selective cannabinoid receptor inhibitors from a library of 187 million compounds within 21 hours using a single CPU\u2014a computational feat that would have been impossible with traditional methodologies. Similarly, VirtuDockDL combines both ligand-based and structure-based screening approaches with deep learning algorithms, offering superior predictive accuracy and complete automation for large-scale datasets.", "# Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\nAuthors: Ye, Q.; Shvartsbart, A.; Li, Z.; Gan, P.; Policarpo, R.L.; Qi, C.; Roach, J.J.; Zhu, W.; McCammant, M.S.; Hu, B.; Li, G.; Yin, H.; Carlsen, P.; Hoang, G.; Zhao, L.; Susick, R.; Zhang, F.; Lai, C.-T.; Allali Hassani, A.; Epling, L.B.; Gallion, A.; Kurzeja-Lipinski, K.; Gallagher, K.; Roman, V.; Farren, M.R.; Kong, W.; Deller, M.C.; Zhang, G.; Covington, M.; Diamond, S.; Kim, S.; Yao, W.; Sokolsky, A.; Wang, X.\nSource: Journal of Medicinal Chemistry, 2025\n\nAn X-ray structure of 5a, the methyl ester analogue of 5, was solved to help guide further SAR (Figure 3B). Examination of the structure showed that, contrary to our hypothesis, the Switch II loop was positioned in virtually the same conformation as seen in the X-ray structure of 3. The cyanoethyl had instead bent underneath the core and made a hydrogen bond with the same water molecule as the azabicyclo[2.1.1]hexane. On the other hand, our hypothesis that growing from the 2 position of the imidazole would yield productive interactions with Asp12 was validated, with the carbonyl forming a hydrogen bond with the backbone N\u2212H. It is notable that, despite not forming a direct hydrogen bond with the Asp12 side chain, compound 5a maintains significant selectivity over KRAS WT (KRAS G12D SPR: 0.16 nM, KRAS WT SPR: 8.8 nM, 55\u00d7 selectivity). This level of selectivity was generally maintained when the backbone N\u2212H interaction was present, see Supporting Information for details.\n\nDespite excellent binding potency (20 pM in SPR) compound 5 had relatively poor translation into cellular and whole blood settings, presumably due to a high degree of polarity. In an attempt to attenuate the polarity, we examined the related tricyclic indole core. Cognizant that replacing a nitrogen for a C\u2212H could influence the orientation of the amino azetidine, we surveyed SAR at this position (Table 1), which is pointed toward the solvent front. Indeed, on the indole core the amino azetidine proved inferior to more flexible oxygen linked substituents, in particular the \u03b1-methyl pyrrolidinol and a related bicyclic analog first reported by researchers at Mirati. The shift from imidazole to indole had a beneficial effect on cellular translation, though subtle, and the indole core was maintained for subsequent SAR. Ultimately, compound 10 emerged from this effort as a new potency record for the program, bottoming out the SPR assay at <1 pM Kd.\n\n*Table 1.* Examination of Solvent Front SAR on Indole and Imidazole Cores. All experiments are an average of at least n = 2 unless otherwise noted, see Supporting Information for details.\n\n*(the body of Table 1 is an image-based structure table and did not survive text extraction from the PDF; the compounds and trends it summarises are described in the surrounding text)*\n\nCompound 10 represented a key juncture. We had successfully demonstrated that an irreversible warhead was not necessary to achieve excellent potency against the KRAS G12D mutant and that this potency could be achieved with a small molecule. In addition to excellent binding potency, 10 showed 20 nM potency in the pERK cellular assay and <500 nM potency in the hWB pERK assay. However, the path toward a balanced, orally bioavailable compound was not clear. Neither 10 nor any related compounds had any measurable flux in a Caco-2 assay (<0.1 \u00d7 10\u207b\u2076 cm/s) and, despite their picomolar binding affinity, showed only moderate translation to a cellular setting, suggesting high polarity. Even with high polarity, 10 and related compounds had high clearance (>1.1 L/h/kg) in human liver microsomes.\n\nIn examining the structure of 10, a number of potential liabilities were identified that we saw as contributing factors to the poor in vitro profile.\n\n1. Two highly basic amines: the azabicyclo[2.1.1]hexane and the pyrrolidine solvent fragment\n2. The hydroxynapthol functionality, a known liability due to glucoronidation and oxidative metabolism\n3. Flexibility in the ethyl amide fragment\n4. Flexibility in the cyanoethyl fragment\n\n*Figure 4.* Structural simplification of compound 10.\n\nIn spite of the detrimental ADME effects, all substituents seemed critical for potency, with each engendering a >10\u00d7 loss in activity if removed. Nevertheless, the team decided that further incremental changes on the current scaffold would not provide oral exposure befitting a developmental candidate and delineated a set of targets we felt were necessary to enter more suitable ADME property space. First, the number of basic amines was capped at one. This meant that the pyrrolidinol in the solvent had to be replaced, since all compounds without the azabicyclo[2.1.1]hexane continued to be inactive (>10 \u03bcM in exchange potency). Second, the hydroxyl in the Switch II pocket had to be removed or replaced to avoid the associated metabolic liabilities and extra hydrogen bond donor.\n\nWe reasoned that by replacing one of the highly polar substituents with a more balanced one, the loss of binding affinity would be compensated by a concurrent improvement in cellular translation. Extensive screening of the Switch II pocket and solvent region followed. Eventually we were able to show that liabilities 1 or 2 could be addressed independently in a limited number of cases without compromising cellular potency (Figure 4). Compound 11, with an ethoxy group replacing the pyrrolidinol and compound 12 with the napthol converted to 7-fluoronaphthyl are representative of the general trends observed. Both were 200 fold less potent in the binding assay, but maintained whole blood pERK potency within 2 fold of compound 10. However, when the SAR was combined, the resulting compound (13) lost >30 fold potency. In addition, compounds 11 and 12, along with all other compounds with only one target liability removed, did not show an improvement in ADME properties. Even 13 showed only a moderate improvement in clearance (0.7 L/h/kg) and no improvement in permeability. Clearly, an additional potency breakthrough would be necessary to achieve our goal.", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\nTo test the ability of anti-mutCALR antibodies to inhibit oncogenic cell proliferation triggered by the Type 1 and Type 2 CALR mutations, Ba/F3 cells (DSMZ) were engineered to express MPL + mutCALR Type 1 (52 bp deletion; SEQ ID NO:320) or MPL + mutCALR Type 2 (5 bp insertion; SEQ ID NO:321). Cells were plated at 5,000 cells per well in RPMI 1640 + 2% FBS, antibodies added, incubated for 72 hours, and followed by assessment of cell viability using the CellTiter-Glo Luminescent Cell Viability Assay (Promega, Madison, WI) and Top Count (Perkin Elmer, Boston, MA) or Pherastar (BMG Labtech, Ortenberg, Germany) for luminescence quantification. Anti-mutant CALR antibodies show potency to inhibit mutCALR-induced oncogenic cell proliferation associated with both Type 1 and Type 2 mutations (FIG. 3).\n\n### Example 7: Anti-mutCALR Antibody Mediated Inhibition of MPL Dimerization\n\nInhibition of MPL dimerization was tested as a potential mechanism of action of the anti-mutCALR antibodies. HAP1 cells knocked out for human JAK2 (Horizon Discovery, Ltd.) were transiently transfected with vectors encoding the MPL-LgBiT and MPL-smBiT fusion proteins (Promega Corp.). Also included in the transfection were vectors encoding the full-length human JAK2 and the full-length WT or mutant CALR Type 1 protein (full-length cds cloned into the pD2529 vector, ATUM Bio). The cells were transfected in 96 well plates using Trans-IT 2020 reagent (Mirus Bio LLC) with equivalent amounts of each plasmid. Cells were then incubated at 37 \u00b0C, 5% CO2. 6 hours post-transfection, antibodies were diluted into growth medium (IMDM, 10% FBS) and added to the cells at the indicated concentrations. The plates were incubated overnight and the growth medium was replaced with 100 \u03bcl OPTI-MEM I, no phenol red, containing the same concentrations of antibodies. After 1 hour at 37 \u00b0C, 5% CO2, 25 \u03bcl NanoGlo Live Cell Reagent (Promega) was added to each well. Plates were returned to the incubator for 30 minutes before luminescence was read on a PHERAstar FSX (BMG Labtech). Data was analyzed using GraphPad PRISM Software (version 7.04) and expressed as percent inhibition compared to isotype control. As shown in FIGs. 4A-4G, anti-mutCALR antibodies (clones 55, 65, 68, 74, 134, 184, and 188) can inhibit MPL dimerization.\n\nThe amino acid sequences for the MPL-smBiT and MPL-LgBiT fusion proteins used in this Example are given as SEQ ID NO:336 and SEQ ID NO:337, respectively.\n\n### Example 8: Effect of Anti-mutCALR Antibody in Mice Injected with Tumor Cells Expressing MPL-mutCALR\n\nTo test the functional ability of anti-mutCALR antibodies in vivo, antibodies were evaluated in a mouse model of tumor growth. Engineered Ba/F3 tumor cells expressing MPL/mutCALR Type 1 were inoculated intravenously in NSG immunodeficient mice (NOD-*scid* IL2Rgamma^null, The Jackson Laboratories, Bar Harbor, ME). Tumors were allowed to grow for 10 days when mice were randomized into antibody or isotype control treatment groups. Different doses of antibodies were administered intraperitoneally and tumor growth was followed over time by assessing the presence and number of tumor cells in the blood (Sysmex, Kobe, Japan). Further, full hematology and tumor infiltration in the spleen and bone marrow were evaluated for assessment of antibody potencies. A representative in vivo study is shown in FIG. 5.\n\nThe anti-mutCALR antibodies prolonged mouse survival (FIG. 6) and prevented splenomegaly (FIG. 7), thrombocytopenia (FIG. 8), and proliferation of tumor cells in the blood (FIG. 9).\n\n### Example 9: Anti-mutCALR Antibodies potentiate the therapeutic response of ruxolitinib in the inhibition of oncogenic cell proliferation triggered by mutCALR Type 1 or Type 2\n\nTo test the ability of anti-mutCALR antibodies to potentiate the therapeutic response of the JAK1/2 inhibitor ruxolitinib, Ba/F3 cells were engineered to express MPL + mutCALR Type 1 (52 bp deletion) or MPL + mutCALR Type 2 (5 bp insertion) as described above. Cells were plated at 5,000 cells per well in RPMI 1640 + 2% FBS, treated with 50 nM of ruxolitinib and/or anti-mutCALR antibodies. Cells were incubated for 72 hours followed by assessment of cell viability using the CellTiter-Glo Luminescent Cell Viability Assay (Promega, Madison, WI) and Top Count (Perkin Elmer, Boston, MA) or Pherastar (BMG Labtech, Ortenberg, Germany) for luminescence quantification. Ruxolitinib used at a concentration of 50 nM inhibited the level of oncogenic cell proliferation by approximately 20% (dotted line). Anti-mutant CALR antibody clone 6 potentiated the ability of ruxolitinib to inhibit cell proliferation in cells carrying the CALR Type 1 (FIG. 10; top) or Type 2 (FIG. 10; bottom) mutation. IC50 values for various clones are shown in Table 13.\n\n*Table 13.* IC50 values for anti-mutCALR antibody or antibody plus ruxolitinib mediated inhibition of cell proliferation (Ba/F3-TPOR/mutCALR type 1 engineered cells).\n\n| Clone No. | Log IC50 (\u03bcg/mL), antibody alone | Log IC50 (\u03bcg/mL), antibody + ruxolitinib |\n|---|---|---|\n| 6 | -1.354 | -4.270* |\n| 15 | -1.558* | -6.186* |\n| 17 | -1.585* | -4.962* |\n| 32 | 0.180* | -0.586* |\n| 36 | -8.511 | -3.880* |\n| 38 | -0.035 | 0.542 |\n| 40 | -36.7 | -1.595* |\n| 51 | -1.585* | -4.962* |\n| 52 | 6.906 | -0.709 |\n| 53 | -0.173 | -0.424 |\n| 74 | -0.438 | -1.458 |\n\n*extrapolated from fitted curve\n\n### Example 10: Anti-mutCALR Antibodies compete with MPL in vitro for the binding to mutCALR\n\nThe amino acid sequence of recombinant His tag_hMPL used in this example is given as SEQ ID NO:338, and the amino acid sequence of recombinant Flag tag_GFP_tev_hmutCALR Type 1 is given as SEQ ID NO:326.", "# JCSG Top96 Screen Formulation\nAuthors: Deller, M. C.\nSource: Joint Center for Structural Genomics, The Scripps Research Institute (screen formulation sheet, 16 April 2013), 2013\n\n**Author:** Marc C. Deller (mdeller@scripps.edu), Joint Center for Structural Genomics, The Scripps Research Institute. **Date:** 16 April 2013.\n\n## Introduction\n\nThe JCSG Top96 screen is a rationally designed, single-plate, 96-condition protein crystallization screen derived from the crystallization conditions that produced more than 1000 JCSG protein structures deposited in the Protein Data Bank. This formulation sheet lists the 96 most successful JCSG crystallization conditions with respect to JCSG structures deposited in the PDB; together these conditions were responsible for 631 JCSG PDB depositions between 1/1/2001 and 9/23/2010. Conditions are listed in plate order (according to major precipitant class) as employed at the JCSG. For each well the sheet gives the precipitant class, the parent commercial screen and condition number, the corresponding JCSG Core Suite condition, the JCSG Top96 well, the crystallization solution, the recommended JCSG Top96 Cryo cryoprotectant, and the number of PDB structures the condition produced.\n\n## Screen formulation", "# ORCID record of Marc C. Deller\nAuthors: Deller, M.C.\nSource: ORCID, 2026\n\nKeywords: Structural Biology, Protein Crystallography, Crystallization, Protein Expression and Purification\n\n## Works (51)", "# WordPressID \u2013 Every post and page from any WordPress site, to CSV and screen, in one command\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-14 on marcdeller.com (categories: vibe-code).*\n\n> **Every post and page from any WordPress site, to CSV and screen, in one command.**\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/wordpressID](https://github.com/bellcheddar/wordpressID) |\n\n---\n\nA single-command tool that pulls every post and page from a WordPress site via the public REST API and outputs a clean, de-duplicated inventory of `id`, `type`, and `title`: written to CSV on disk and rendered as an aligned table on screen in the same pass.\n\nWhy it matters: getting a complete list of a WordPress site\u2019s content sounds trivial until the REST API silently caps at 100 items and drops the rest with no warning, leaving you with a quietly incomplete inventory. wp-content-inventory paginates properly, decodes the HTML entities WordPress bakes into titles, and renders a quoted CSV and an aligned on-screen table from a single fetch so the two can never disagree. It is useful for migrations, content audits, redirect mapping, and any maintenance task where you need a trustworthy, complete manifest of a site\u2019s posts and pages, using nothing but curl and jq against endpoints WordPress already exposes.\n\nNo plugins, no authentication, no scraping. Just `curl` and `jq` against the endpoints WordPress already exposes.\n\n## What it does\n\n* Enumerates all published **posts** and **pages** through the WordPress REST API (`/wp-json/wp/v2/`).\n* Paginates automatically, so it captures everything rather than silently stopping at the API\u2019s 100-item cap.\n* Decodes the HTML entities WordPress returns in titles (`&amp;`, smart quotes, en dashes).\n* Writes a properly quoted CSV to disk and prints a column-aligned, comma-safe table to the terminal.\n* Fetches each endpoint **once** and renders both outputs from that single capture, so the file and the screen can never drift apart.\n\n## Requirements\n\n| Tool | Purpose | Notes |\n| --- | --- | --- |\n| `curl` | HTTP requests | Pre-installed on macOS and most Linux distros |\n| `jq` | JSON parsing and CSV/TSV encoding | `brew install jq` or `apt install jq` |\n| `column` | On-screen table alignment | Ships with macOS and `util-linux`; only needed for the table view |\n\n## Usage\n\nSet your target site and run:\n\n```\nSITE=\"https://marcdeller.com\"\n\ndata=$( for t in posts pages; do p=1; while r=$(curl -s \"$SITE/wp-json/wp/v2/$t?per_page=100&page=$p&_fields=id,title\") && echo \"$r\" | grep -q '\"id\"'; do echo \"$r\" | jq -c --arg t \"${t%s}\" '.[] | {id, title:(.title.rendered|gsub(\"&amp;\";\"&\")|gsub(\"&#8217;\";\"\\u2019\")|gsub(\"&#8211;\";\"-\")), type:$t}'; p=$((p+1)); done; done ); \\\n{ echo \"id,type,title\"; echo \"$data\" | jq -r '[.id,.type,.title]|@csv'; } > wp_content.csv; \\\n{ printf 'ID\\tTYPE\\tTITLE\\n'; echo \"$data\" | jq -r '[.id, .type, (if (.title|length)>70 then .title[0:69]+\"\\u2026\" else .title end)]|@tsv'; } | column -t -s$'\\t'\n```\n\nThis writes `wp_content.csv` and prints a table like:\n\n```\nID     TYPE  TITLE\n14613  post  2026 - Discovery of the Orally Bioavailable Isoform Selective Janus K\u2026\n13848  post  2026 - Anthropic Academy\n13485  post  Plastic, Persistent, and Potentially Carcinogenic\n...\n```\n\n## Output\n\nThe CSV holds the full, untruncated data:\n\n```\nid,type,title\n14613,\"post\",\"2026 - Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\"\n13848,\"post\",\"2026 - Anthropic Academy\"\n11930,\"page\",\"About\"\n```\n\n| Column | Description |\n| --- | --- |\n| `id` | WordPress object ID |\n| `type` | `post` or `page` |\n| `title` | Decoded, CSV-quoted title |\n\nThe on-screen `title` is truncated to 70 characters purely for terminal width. The CSV always keeps the full title.\n\n## How it works\n\n**Pagination is mandatory, not optional.** The REST API caps `per_page` at 100 and will silently drop the overflow. A site with 193 posts returns exactly 100 from a naive call, with no error to warn you. The loop walks `&page=$p` until a page returns no IDs, so the count is always complete. You can confirm the true total against the `X-WP-Total` response header:\n\n```\ncurl -sI \"$SITE/wp-json/wp/v2/posts?per_page=100\" | grep -i x-wp-total\n```\n\n**Entity decoding.** WordPress serialises titles with HTML entities. The `gsub` chain handles the three most common (`&amp;`, `&#8217;`, `&#8211;`). Extend it for any others you hit (`&#8220;`/`&#8221;` smart quotes, `&hellip;`), or pipe titles through a full entity decoder if you would rather not play whack-a-mole.\n\n**Comma safety.** Titles routinely contain commas and parentheses. The CSV uses `jq`\u2018s `@csv`, which quotes and escapes fields correctly. The screen table uses `@tsv` with `column -s$'\\t'`, so a comma inside a title never spills into the next column.\n\n**Single fetch, two views.** The API responses are captured once into `$data` as JSON lines, then both the CSV and the table are generated from that. No double requests, no risk of the two outputs disagreeing.\n\n## Customisation\n\n* **More metadata:** add fields via `_fields=id,title,slug,link,date` and extend the `jq` object and output arrays.\n* **Untruncated screen titles:** replace the `if (.title|length)>70 ...` block with plain `.title`.\n* **Wide tables:** append `| less -SX` to scroll horizontally without wrapping.\n* **Spreadsheet-friendly alignment:** swap `@csv` for `@tsv` if you prefer a tab-delimited file.\n* **Other content types:** add custom post types to the `for t in posts pages` list if your site registers them in the REST API.\n\n## Notes\n\nThe REST API must be publicly reachable (it is by default). Sites that disable or firewall `/wp-json/` will return nothing. Only published content is listed unless you authenticate; drafts and private items require an authenticated request.\n\n## Licence\n\nMIT.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/wordpressID](https://github.com/bellcheddar/wordpressID) |", "# Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\nAuthors: Zhuo, J.; Li, Y.-L.; Qian, D.-Q.; Burns, D.M.; Mei, S.; Rafalski, M.; Xu, M.; Cao, G.; Pan, Y.; Jia, Z.; Jalluri, R.; Epling, L.B.; Fenalti, G.; Deller, M.C.; Procak, J.; Covington, M.; He, X.; Collins, R.; Stubbs, M.; Burke, K.; Oliver, J.; Margulis, A.; Boer, J.; Wynn, R.; Scherle, P.; Diamond, S.; Newton, R.; Zhang, Y.; Metcalf, B.; Yao, W.\nSource: Journal of Medicinal Chemistry (Drug Annotation), 2026\n\nBy the early 2000s, there was growing evidence that the JAK-STAT pathway may play a pivotal role in the pathogenesis of a variety of hematologic disorders and hematopoietic malignancies, such as lymphomas and leukemias. These findings prompted several companies, including Incyte, to pursue kinase-selective JAK inhibitors. In 2005, a series of seminal manuscripts were published that disclosed that the V617F mutation imparts constitutive activity of JAK2, resulting in oncogenesis and pathogenesis of myeloproliferative neoplasms (MPNs), which validated the therapeutic potential of JAK inhibition and provided clinical guidance. Medicinal chemistry efforts led by Rodgers, Shepard, Vaddi et al. led to the discovery of INCB018424 (1, Jakafi), which demonstrated preclinical therapeutic benefit for myeloproliferative disorders and advanced to clinical trials (Figure 2). In 2011, the dual JAK1/2 inhibitor ruxolitinib (1) was approved for polycythemia vera as the first JAK-selective kinase inhibitor to be approved by the US Food and Drug Administration (FDA). In 2007, a second dual JAK1/2 inhibitor, INCB028050 (2, baricitinib), was discovered at Incyte and was codeveloped with Eli Lilly. Baricitinib (2) gained European approval in 2017 and FDA approval in 2018 for the treatment of rheumatoid arthritis. Despite the success of the dual JAK1/-2 inhibitors, particularly in the MPN space, it was appreciated that selective JAK1 inhibition might be a viable therapeutic strategy to treat autoimmune and inflammatory diseases that are predominantly reliant on JAK1 heterodimers to promote signaling. It was hypothesized that by sparing JAK2 inhibition, one could allow healthy cells to preserve normal JAK2-STAT signaling, which is essential for hematopoietic processes and may alleviate some of the side effects associated with JAK2 inhibition such as anemia, thrombocytopenia, and neutropenia. Therefore, a JAK1-selective program was initiated to pursue autoimmune and inflammatory diseases.\n\n*Figure 1.* Role of JAK1 in Various Autoimmune and Inflammatory Diseases. EPO, erythropoietin; GM-CSF, granulocyte-macrophage colony-stimulating factor; IFN, interferon; IL, interleukin; ILC1, group 1 innate lymphoid cells; JAK, Janus kinase; TH, T helper cell; TPO, thrombopoietin; TSLP, thymic stromal lymphopoietin; TYK2, tyrosine kinase 2.\n\n*Figure 2.* Incyte JAK Inhibitors. IC50, half-maximal inhibitory concentration; JAK, Janus kinase. Compound 1: INCB018424, ruxolitinib, Jakafi 2011. Compound 2: INCB028050, baricitinib, Olumiant 2017. Compound 3: INCB039110 (itacitinib), JAK1 IC50 3.6 nM, JAK2 IC50 75 nM, JAK1/JAK2 selectivity 21\u00d7.\n\n## Results and Discussion\n\nIn 2009, our first JAK1 isoform-selective clinical candidate was identified, compound 3 (INCB039110, itacitinib), which was 21-fold selective for JAK1 inhibition over JAK2 (Figure 2). It was postulated ad hoc that the selectivity may be attributed to two minor differences between JAK1 and JAK2: (1) the JAK1 P-loop can adopt a more retracted conformation, creating a larger pocket for JAK1 vs JAK2; (2) the loop connecting the JAK-1 \u03b2-3 and \u03b1-C-helix is extended by one amino acid, creating an additional small lipophilic pocket that may accommodate the CF3 group (Figure 3). In an effort to decrease the molecular weight, introduce sp3 character, and improve absorption, distribution, metabolism, and excretion (ADME), as well as physicochemical properties, numerous nonaromatic capping groups were explored, and the pyrrolidine ring was discovered as a viable option (Table 1). From the piperidine-azetidine scaffold, it was found that alkylation of the urea moiety may be partially responsible for the poor ADME properties; thus, the azetidine ring was replaced with a cyclobutyl and a phenyl group to afford amides 6 and 7. Whereas both amides 6 and 7 had an increase in the permeability compared to the urea 5, the phenyl amide 7 had an increase in both the enzymatic and cellular potencies compared to both urea 5 and cyclobutyl amide 6. It was hypothesized that switching the ring that was connected to pyrazole from piperidine to the more conformationally rigid azetidine may further increase the potency and may be beneficial to improve projected human clearance and the pharmacokinetic (PK) profile. Gratifyingly 8, displayed a ~3-fold increase in the enzymatic and cellular potency and a 2-fold increase in whole blood potency compared to the piperidine amide 7. Unfortunately, the projected human clearance was unaffected. It was speculated that breaking the pyrrolidine ring to afford a secondary amide may introduce the necessary polarity to improve the clearance. In addition, elimination of the \u03b1-pyrrolidine methyl group may eliminate a potential metabolic liability. The simple *i*-propyl amide 9 was prepared and demonstrated that the projected human clearance could be improved with this modification, albeit with a slight loss in permeability. We were pleased to observe that the potency of 9 was maintained and the JAK1 isoform selectivity was surprisingly improved by 2-fold in the whole blood assay (WB TPO/IL-6). The JAK isoform selectivity for JAK3 and TYK2 was approximately 890- and 120-fold, respectively compared to JAK1 enzymatic potency. Compounds from this series consistently displayed similar or improved selectivity profiles for JAK3 and TYK2, with JAK3 potency typically being over 1 \u03bcM.", "# Engineering Evolution: How FAST-PETase and Other Variants Are Transforming Plastic Biodegradation\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-11-05 on marcdeller.com (categories: blog, microplastics).*\n\n* [FAST-PETase: Machine learning-aided engineering of hydrolases for PET depolymerization](https://www.nature.com/articles/s41586-022-04599-z)\n* [IsPETase: A bacterium that degrades and assimilates poly(ethylene terephthalate)](https://www.science.org/doi/10.1126/science.aad6359)\n\n* **Machine Learning Optimization:** [FAST-PETase achieves 38-fold activity increase through structure-guided mutation selection](https://nature.com/articles/s41586-022-04599-z) targeting loop stabilization and salt bridge formation with support from [computational analysis of temperature-induced stabilization](https://pmc.ncbi.nlm.nih.gov/articles/PMC11946493/).\n* **Thermostability Engineering:** [HotPETase and DuraPETase employ extensive mutation networks creating hydrogen bonds, disulfide bridges, and electrostatic interactions](https://pmc.ncbi.nlm.nih.gov/articles/PMC10546322/) for industrial temperature operation.\n* **Cutinase Alternatives:** [LCC and Cut190 variants offer thermophilic scaffolds with calcium-dependent mechanisms and engineered substrate binding sites](https://febs.onlinelibrary.wiley.com/doi/10.1111/febs.16736) for aromatic polymer recognition.\n* **Activity-Stability Trade-offs:** [Mutations affecting aromatic residues (W159H, S214H, F243I/T) modulate active site chemistry](https://pmc.ncbi.nlm.nih.gov/articles/PMC10546322/) demonstrating inherent balance between catalytic efficiency and thermal resistance.\n* **Substrate-Dependent Performance:** [Variant effectiveness varies with PET properties (crystallinity, molecular weight, physical form)](https://www.nature.com/articles/s41467-024-45662-9), requiring tailored enzyme-substrate matching for successful industrial deployment.\n\n1. **Discovery and Characterization of IsPETase:** [Yoshida, S., et al., Science, 2016](https://science.sciencemag.org/content/351/6278/1196)\n2. **Current advances in the structural biology and molecular engineering of PETase:** [Liu, F., et al., Front Bioeng Biotechnol, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC10546322/)\n3. **Efficient polyethylene terephthalate degradation at moderate temperature: a protein engineering study of LC-cutinase highlights the key role of residue 243:** [Pfaff, L., et al., FEBS J, 2023](https://febs.onlinelibrary.wiley.com/doi/10.1111/febs.16736)\n4. **Computational redesign of a hydrolase for nearly complete PET depolymerization:** [Cui, Y., et al., Nat Commun, 2024](https://www.nature.com/articles/s41467-024-45662-9)\n5. **Thermostability and Activity Improvements of PETase from Ideonella sakaiensis:** [Zhang, H., et al., Molecules, 2024](https://www.mdpi.com/1420-3049/29/6/1338/pdf?version=1710754801)\n6. **Computational analysis reveals temperature-induced stabilization of FAST-PETase:** [Stockinger, P., et al., Comput Struct Biotechnol J, 2025](https://pmc.ncbi.nlm.nih.gov/articles/PMC11946493/)\n\n**A deep dive into the structural differences, mutations, and engineering approaches driving the next generation of PET-degrading enzymes**\n\nThe global plastic waste crisis has sparked intense efforts to develop enzymatic solutions for polyethylene terephthalate (PET) degradation. Since the 2016 discovery of [*Ideonella sakaiensis*PETase](https://pmc.ncbi.nlm.nih.gov/articles/PMC10546322/) (*Is*PETase), researchers have engineered multiple variants with enhanced activity and thermostability. This article examines key PETase variants, focusing on their structural modifications, mutation strategies, and performance characteristics that position them as viable tools for industrial plastic recycling.\n\n**Structural Foundation and Engineering Strategies**\n\nAll PETase variants share the conserved \u03b1/\u03b2 hydrolase fold featuring a catalytic triad (Ser160-His237-Asp206 in *Is*PETase) and a substrate binding cleft on the protein surface. However, [engineering approaches differ significantly](https://pmc.ncbi.nlm.nih.gov/articles/PMC12387008/) in their mutation selection strategies. Rational design based on crystal structures has guided thermostability improvements, while [machine learning algorithms](https://pmc.ncbi.nlm.nih.gov/articles/PMC11946493/) have identified synergistic mutation combinations that traditional methods might miss.\u200b\n\nThe [FAST-PETase variant](https://www.sciencedirect.com/science/article/pii/S2001037025000789) (*Is*PETase S121E/D186H/R224Q/N233K/R280A) exemplifies structure-based machine learning engineering. Developed using the MutCompute algorithm, FAST-PETase demonstrated 38-fold higher activity than its ThermoPETase parent at 50\u00b0C, releasing 33.8 mM PET monomers after 96 hours. The N233K mutation introduced an intramolecular salt bridge with E204, while D186H formed additional hydrogen bonds with P120 and F191, collectively stabilizing the \u03b26-\u03b27 connecting loop.\u200b\n\nThermostable variants such as [HotPETase](https://pmc.ncbi.nlm.nih.gov/articles/PMC10546322/) contain 21 mutations relative to wild-type *Is*PETase, achieving a melting temperature of 82.5\u00b0C. The P181V mutation in HotPETase created additional hydrogen bonding and decreased flexibility around residues Ala183-Asn190, improving packing of the central \u03b2-sheet region. Meanwhile, [DuraPETase](https://pmc.ncbi.nlm.nih.gov/articles/PMC10546322/), engineered via the GRAPE computational strategy, incorporated 10 mutations (S214H-I168R-W159H-S188Q-R280A-A180I-G165A-Q119Y-L117F-T140D) that increased thermostability by 31\u00b0C and degradation activity 300-fold on semicrystalline PET.\u200b\n\n**Comparative Analysis of Cutinase-Like Variants**", "# GitHub README: ChemSage\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ChemSage, 2026\n\n| | Round 1 (7B) | Round 2 (32B) | Round 3 (32B) | Round 4 (32B) | Round 5 (32B) |\n|---|---|---|---|---|---|\n| Total examples | 1,500 | 1,500 | 5,000 | 8,000 | 20,000 |\n| Train / valid / test | 1,200 / 150 / 150 | 1,200 / 150 / 150 | 4,000 / 500 / 500 | 6,400 / 800 / 800 | 16,000 / 2,000 / 2,000 |\n| Behaviour classes | 8 | 8 | 19 | 59 | 78 |\n| Seed SMILES | 25 | 25 | 51 + corpus drugs | 51 + corpus drugs | 51 + corpus drugs |\n| Corpus drug examples | \u2014 | \u2014 | ~1,000 | ~1,000 + protein family PDBs | ~1,200 + protein family PDBs |\n| Code-block density | Low (~50%) | Low (~50%) | High (~80%) | High (~85%) | High (~90%) |\n| Generator script | `--n 1500` | `--n 1500` | `--n 5000 --seed 43` | `--n 8000 --seed 47` | `--n 20000 --seed 51` |\n| Validation | RDKit + SMILES | RDKit + SMILES | RDKit + SMILES | RDKit + SMILES | RDKit + SMILES |\n| Token lengths (p50/p99/max) | \u2014 | \u2014 | \u2014 | \u2014 | 322 / 488 / 558 |\n| Rejected examples | \u2014 | \u2014 | 243 (95%) | \u2014 | 707 (96.5%) |\n\n**Behaviour classes by round:**\n\n*Round 3 additions (targeting R2 eval gaps):*\n\n| Class | Weight | Purpose |\n|---|---|---|\n| `single_property` | 2x | One property at a time with code block |\n| `full_profile` | 2x | Complete property profile via code |\n| `qed_score` | 1x | QED drug-likeness score |\n| `molecular_formula` | 1x | Formula, atom counts, ring analysis |\n| `smiles_validation` | 1x | Valid/invalid SMILES handling |\n| `property_comparison` | 1x | Side-by-side two-molecule comparison |\n| `corpus_drug` | 2x | Named drugs from approved_drugs.csv |\n| `murcko_sar` | 1x | Scaffold comparison with SAR context |\n\n*Round 4 additions (structural biology + fidelity):*\n\n| Class | Weight | Purpose |\n|---|---|---|\n| `pymol_selection` | 2x | PyMOL selection algebra (within, byres, chain, resi) |\n| `pymol_viz` | 2x | PyMOL visualization (surface, alignment, symmetry, B-factor, figures) |\n| `plip_analysis` | 2x | PLIP interaction types (H-bond, hydrophobic, pi-stack, salt bridge, halogen) |\n| `pdb_format` | 2x | PDB file format (ATOM/HETATM, altloc, B-factor interpretation) |\n| `pdbtools` | 2x | pdb-tools CLI recipes (selchain, selres, tidy, splitmodel) |\n| `biopython` | 2x | Bio.PDB patterns (SMCRA, NeighborSearch, Superimposer) |\n| `chimerax` | 2x | ChimeraX commands (hbonds, clashes, matchmaker) |\n| `gemmi` | 1x | gemmi mmCIF/PDB parsing and neighbour search |\n| `struct_prep` | 1x | Structure preparation (PDBFixer, protonation, 3D conformers) |\n| `docking_interp` | 2x | Docking result interpretation (Vina scores, RMSD, Kd estimation) |\n| `plotly` | 2x | Plotly Python visualization (MW vs logP, radar charts) |\n| `r_recipe` | 2x | R/ggplot2/bio3d/Plotly-R (SAR plots, RMSD, RMSF) |\n| `protein_family` | 3x | Kinases, GPCRs, antibodies, GTPases, cytokines, receptors (real PDB data) |\n| `code_then_quote` | 3x | \"Compute then quote\" pattern for numerical fidelity |\n| `rounding_precision` | 2x | Exact rounding conventions (MW 1dp, logP 2dp, TPSA 1dp) |\n| `personality_drug` | 3x | Lipinski verdicts with \u2705/\u274c cards, emoji, and medchem humor |\n| `personality_valid` | 2x | SMILES validation with flair and emoji |\n| `personality_struct` | 3x | Structural observations with character and humor |\n| `refusal` | 2x | Out-of-scope and invalid query handling |\n\n### Training parameters\n\n| Parameter | Round 1 (7B) | Round 2 (32B) | Round 3 (32B) | Round 4 (32B) | Round 5 (32B, complete) |\n|---|---|---|---|---|---|\n| Base model | `Qwen2.5-7B-Instruct-4bit` | `Qwen2.5-32B-Instruct-4bit` | `Qwen2.5-32B-Instruct-4bit` | `Qwen2.5-32B-Instruct-4bit` | `Qwen2.5-32B-Instruct-4bit` |\n| Fine-tune type | QLoRA (4-bit base) | QLoRA (4-bit base) | QLoRA (4-bit base) | QLoRA (4-bit base) | QLoRA (4-bit base) |\n| `use_rslora` | `true` | `true` | `true` | `true` | `true` |\n| LoRA rank | 32 | 32 | 32 | 32 | **64** |\n| LoRA scale | 64.0 | 64.0 | 64.0 | 64.0 | **90.0** |\n| LoRA keys | q/k/v/o_proj + gate/up/down_proj | q/k/v/o_proj + gate/up/down_proj | q/k/v/o_proj + gate/up/down_proj | q/k/v/o_proj + gate/up/down_proj | q/k/v/o_proj + gate/up/down_proj |\n| `num_layers` | 16 (of 28) | 32 (of 64) | 32 (of 64) | 32 (of 64) | 32 (of 64) |\n| `iters` | 750 | 750 | 1,500 (stopped at 625) | 1,500 (complete) | **3,000** |\n| `batch_size` | 4 | 4 | 4 | 4 | 4 |\n| `learning_rate` | 2e-5 | 2e-5 | 2e-5 | 2e-5 | 2e-5 |\n| LR schedule | Cosine, 50-step warmup | Cosine, 50-step warmup | Cosine, 50-step warmup | Cosine, 50-step warmup | Cosine, **100**-step warmup |\n| `max_seq_length` | 2048 | 2048 | 2048 | 2048 | **3072** |\n| `steps_per_report` | 10 | 10 | 10 | 25 (aligned with eval) | **50** (aligned with eval) |\n| `val_batches` | \u2014 | \u2014 | \u2014 | 25 | **50** |\n| `seed` | 42 | 42 | 43 | 51 | 51 |\n| Trainable params | ~26M (0.34%) | 134M (0.41%) | 134M (0.41%) | 134M (0.41%) | **268M (0.82%)** |\n| Peak memory | ~12 GB | 24.9 GB | 25.2 GB | 24.9 GB | ~33 GB est. |\n| Training data | 1,200 examples | 1,200 examples | 4,000 examples | 6,400 examples | **16,000 examples** |\n| Adapter path | `adapters/chem_sage_7b_v1_lora/` | `adapters/chem_sage_32b_v2_lora/` | `adapters/chem_sage_32b_v3_lora/` | `adapters/chem_sage_32b_v4_lora/` | `adapters/chem_sage_32b_v5_lora/` |\n| Fused model path | `models/chem_sage_7b_v1/` | `models/chem_sage_32b_v2/` | `models/chem_sage_32b_v3/` | `models/chem_sage_32b_v4/` | `models/chem_sage_32b_v5/` |", "# GitHub README: FlexAppeal\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/FlexAppeal, 2026\n\n> **Turn the whole OpenMM option surface into a form, run the simulation on your own machine, and explore what comes back.**\n\n[](https://flexappeal.mdeller.com)       \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://flexappeal.mdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">flexappeal.mdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/FlexAppeal\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/FlexAppeal</a></td>\n</tr>\n</table>\n\n---\n\nFlexAppeal is a three-tab web application for protein molecular dynamics. The **Prepare** tab turns the OpenMM option surface into a guided form and emits a single self-contained file. You run that file on your own machine, where it installs its environment, benchmarks the available compute platforms, simulates, and analyses the full trajectory. It produces one results file, which the **Analysis** tab reads and renders as interactive plots and a three-dimensional structure viewer. The **Example** tab is a real 10 ns run of lysozyme, rendered through that same path, so what the documentation claims and what the software does cannot drift apart.\n\n**Why it matters:** setting up a correct MD run means stitching together structure repair, force-field selection, solvation, system construction, integrator and barostat choice, a staged equilibration protocol, and a set of reporters, across roughly 120 interacting parameters. A large fraction of the ways to get that wrong produce a trajectory that runs to completion and looks entirely plausible: solvent padding smaller than the non-bonded cutoff, so the protein interacts with its own periodic image; a 4 fs timestep without hydrogen mass repartitioning, so total energy climbs all run; CHARMM36 used with AMBER's cutoff scheme. FlexAppeal encodes those rules as validation rather than leaving them to be remembered. It is useful for: setting up a first simulation without a week of reading, reproducing someone else's protocol exactly, sweeping a parameter across replicates, and getting a standard analysis panel without writing MDTraj by hand.\n\nThe hosted application never simulates. That is a design decision, not a limitation to be lifted later: MD needs hours of dedicated compute, and your trajectory should not have to travel anywhere.\n\n---\n\n## \u2728 Features\n\n| | |\n|---|---|\n| **A worked example** | A real 10 ns lysozyme run, committed and rendered by the same code that renders an upload |\n| **Every OpenMM option, explained** | 115 parameters across 14 groups, each with a full sentence saying why its default is what it is |\n| **Validation that knows the physics** | Padding versus cutoff, timestep versus constraints, switching versus dispersion correction, membrane versus barostat |\n| **Live estimates** | Atom count, trajectory size and wall time update as you type |\n| **One file to run** | A self-extracting bundle carrying the structure, a pinned environment, a readable run script and its own installer |\n| **Auditable output** | The generated `run.py` writes every value literally (`310.0 * unit.kelvin`), so it reads as a record of exactly what ran |\n| **Platform benchmarking** | The bundle times the available OpenMM platforms on your real system and picks the fastest |\n| **Ligands** | Chemistry from the RCSB Chemical Component Dictionary, parameterised with OpenFF Sage, GAFF2 or espaloma |\n| **Membranes** | POPC and six other lipids via `Modeller.addMembrane`, with pre-oriented structures fetched from OPM |\n| **Analysis without scripting** | RMSD, RMSF, radius of gyration, SASA, DSSP, hydrogen bonds, contacts, PCA, clustering, and ligand or membrane panels |\n| **Interactive** | Plotly plots and a Mol\\* viewer, both vendored: the page needs no third-party network access |\n| **A trajectory beside the controls** | Every tab opens with its panel at 75% and the lysozyme run playing at 25%: 396 KB of H.264, fetched after the page has loaded and skipped entirely for reduced motion |\n\n## \ud83c\udfd7\ufe0f How it works\n\n```\n  Prepare tab                    your machine                   Analysis tab\n  \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500                    \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500                   \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n  structure  \u2500\u2510\n  115 options \u251c\u2500\u25ba flexappeal_run.command \u2500\u25ba install \u2500\u25ba simulate \u2500\u25ba .fxa \u2500\u25ba plots\n  validation \u2500\u2518      (one file, ~70 KB)      benchmark   analyse    (~25 MB)  viewer\n```\n\nThree properties fall out of that shape:\n\n- **No metric is computed on thinned data.** The bundle analyses the complete trajectory locally; only the results and a decimated copy for viewing travel back.\n- **Uploads stay small.** Tens of megabytes rather than the gigabytes a raw trajectory would be.\n- **Your trajectory stays yours.** Nothing in the bundle uploads anything.\n\nThe **Example** tab is that whole path already walked: a 10 ns lysozyme run whose bundle, terminal recording and results file are all in `examples/lysozyme_10ns/`, rendered by the code that renders an upload.\n\n## \ud83d\udccb Requirements\n\n**To use the hosted app:** a browser.\n\n**To run a bundle:** macOS on Apple Silicon, and about 4 GB of disk for the environment it installs on first run. Nothing needs to be installed in advance; the bundle bootstraps [pixi](https://pixi.sh) if you do not have it.\n\n**To develop FlexAppeal:**\n\n| Requirement | Notes |\n|---|---|\n| Python 3.11 | The ceiling for `openff-toolkit` and `openmmforcefields` on conda-forge |\n| pixi | Installed by `install.sh` if absent |\n| ~6 GB disk | OpenMM, AmberTools and the OpenFF toolkit |\n\n## \ud83d\udd27 Installation\n\n```bash\ngit clone https://github.com/bellcheddar/FlexAppeal.git\ncd FlexAppeal\n./install.sh\n```\n\nThe web application itself needs none of that. Its own dependencies are pip-installable with no compiled toolchain:\n\n```bash\npip install -r requirements.txt\n```\n\n## \ud83d\ude80 Usage\n\n```bash\npixi run serve                    # http://127.0.0.1:8004\npixi run test                     # the full suite\n```", "# GitHub README: GOBSMACKED\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/GOBSMACKED, 2026\n\n```\napp/                     the Flask application (droplet only, no torch)\n  routes/                prepare, analyze, runs, about\n  services/              fetch, annotate, references, bundle, ingest, superpose,\n                         interactions, scorecard, modes, dynamics, affinity,\n                         poses, movie\n  templates/  static/    Jinja2, the instrument-panel CSS, Mol* and Plotly\nbundle_template/         copied verbatim into every run bundle\n  run.sh                 one-step bootstrap: installs pixi, builds from the lock, runs\n  pixi.lock              1,969 pinned package entries across three environments\n                         (default, fold, affinity) and two platforms\n  run.py                 the six-stage runner with resume markers\n  gobsmacked_run/        fold, prep, dock, md, affinity, summarise, schema,\n                         console, noise\ndesign/                  the visual contract this app is built from\ndeploy/                  systemd units, nginx site, provision and deploy scripts\nscripts/                 prune, DOI checker, THIRD_PARTY generator\ntests/                   160 tests, plus two fixture archives built from crystals\nsoftware.yaml            the single source of truth for attribution\n```\n\n## \ud83d\udd2c A worked example\n\nOne real run, end to end, on an M1 Max: EGFR plus erlotinib, prepared on the live site,\njudged against 1M17.\n\n| Stage | Wall clock | What happened |\n|---|---|---|\n| fetch | seconds | AlphaFold model `AF-P00533-F1`, mean pLDDT 75.9 |\n| annotate | seconds | Pfam `PF07714` 714-966, KLIFS gatekeeper Thr790 |\n| prepare | seconds | 1M17's site renumbered onto the model (+24), box sized on erlotinib at 32 x 21 x 23 A |\n| prep | 3 s | trimmed 1,210 residues to the 253-residue kinase domain, all 51 pocket residues kept |\n| dock | 937 s | PandaDock empirical search, 10 poses, best score \u221215.5 kcal/mol |\n| md | 3,224 s | 58,266 atoms, OpenCL single precision, ~20 ns/day, 100 ps equilibration plus 500 ps production |\n| summarise | 106 s | 100 frames |\n| analyse | 28 s | on the droplet |\n\n**GOBSMACK 53.5, grade D.** The interesting part is what that decomposes into:\n\n| Metric | Value | Grade |\n|---|---|---|\n| Ligand RMSD | 7.83 \u00c5 | F |\n| PLIP overlap | 0.33 | D |\n| Pocket C\u03b1 RMSD | 1.66 \u00c5 | C |\n| \u03c71 agreement | 0.69 | C |\n| Drift, last window | 0.25 \u00c5 | A |\n| MD rescue | \u22120.75 \u00c5 | F |\n\nThe pose is stable and physically valid: it does not move over 500 ps, it has no clash, its\nstereocentres are intact and it sits inside the box. It is also 8 \u00c5 from where erlotinib\nactually binds, it never touches the hinge, and the mode classifier calls it allosteric\nagainst the crystal's Type I. A docking score of \u221215.5 kcal/mol says nothing about any of\nthat, which is the entire argument for scoring a prediction against the structure rather\nthan against itself.\n\nTwo caveats stated rather than buried: this run used the empirical scorer alone, because the\nSE(3) GNN rescoring needs PyTorch and this machine had 12 GB of disk left, and 500 ps is a\nshort production run. Both are reasons to expect a worse pose, and neither changes what the\nscorecard is reporting.\n\n## \ud83e\uddea An experiment: five starting structures, one campaign\n\nThe worked example above scores a D and says the pose is 8 \u00c5 from where erlotinib\nactually binds. The obvious reading is that the starting model was not good enough. That\nis a testable claim, so it was tested.\n\nFive runs of the same campaign, differing in one variable only: the structure docking\nstarts from. Same ligand, same pocket, same box (31.7 x 21 x 23 \u00c5), same seed, same\nexhaustiveness, same MD protocol, all judged against the same crystal. The first is a\ncontrol that should be unbeatable, because it docks erlotinib back into the very crystal\nit came from.\n\n| # | Starting structure | Pocket C\u03b1 RMSD to 1M17 | Top-ranked pose | Best pose (its rank) | GOBSMACK |\n|---|---|---|---|---|---|\n| 1 | **1M17 crystal, self-dock** | **0.00 \u00c5** | **8.77 \u00c5** | 1.51 \u00c5 (rank 8) | D 52.8 |\n| 2 | 4HJO crystal, cross-dock | 1.53 \u00c5 | 4.29 \u00c5 | 2.29 \u00c5 (rank 5) | D 50.5 |\n| 3 | AlphaFold DB `AF-P00533-F1` | 0.91 \u00c5 | 7.73 \u00c5 | 4.81 \u00c5 (rank 8) | D 55.0 |\n| 4 | ESMFold, ESM Atlas | 0.99 \u00c5 | 8.09 \u00c5 | 5.44 \u00c5 (rank 9) | D 49.0 |\n| 5 | Boltz-2 co-folded, ligand stripped | 0.94 \u00c5 | 1.40 \u00c5 | 1.40 \u00c5 (rank 1) | B 79.0 |\n\n**The control settles it. A perfect receptor produced one of the worst top-ranked\nposes.** Docking erlotinib into its own crystal, with a pocket that is correct by\nconstruction, put an 8.77 \u00c5 pose first. The crystal pose was found: it is in the same\nlist of ten, at 1.51 \u00c5. It was placed eighth.\n\nSo the sampling is not the bottleneck and the receptor is not the bottleneck. Within\nthis pose set the limitation is **which pose gets ranked first**, which is a\nwell-documented weakness of fast empirical scoring functions in general and not a\nproperty of any one implementation. They are built to be quick enough to search millions\nof placements, and that speed is bought against exactly this discrimination.\n\nTwo secondary results, both of which change how the other numbers should be read:\n\n- **Starting-model quality does not predict pose quality.** Three of these pockets sit\n  within 0.09 \u00c5 of each other, at 0.91, 0.94 and 0.99 \u00c5, and produce top-ranked poses of\n  7.73, 1.40 and 8.09 \u00c5. The best possible receptor, at 0.00 \u00c5, produces the worst pose\n  of the five. Reporting \"the model was too rough\" would have been a comfortable and\n  wrong conclusion, and four of the five runs on their own would have supported it.\n- **The predicted affinity does not separate right poses from wrong ones.** pIC50 ranged\n  over 6.44 to 6.66 across poses spanning 1.4 to 8.8 \u00c5, and the run with the *worst*\n  top-ranked pose returned the *highest* affinity of the five. This is the measured\n  argument for keeping affinity beside the composite score rather than inside it.", "# GitHub README: TopPDBLX\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/TopPDBLX, 2026\n\n**Data:** Creative Commons Attribution 4.0 International, see [`LICENSE-DATA`](LICENSE-DATA).\n**Code:** MIT, see [`LICENSE`](LICENSE).\n\nAttributing TopPDBLX does not discharge the obligation to the sources it derives from: the Protein Data Bank (CC0), SIFTS and UniProt (both CC BY 4.0). Commercial screen formulations in `ontology/screens/` are transcriptions of vendor-published support materials, kept structurally separable so they can be withdrawn without breaking the release. Screen names are trademarks of their owners, used nominatively.\n\n**Cite as:** Deller, M. C. (2026). TopPDBLX: a parsed, normalised and sequence-linked database of crystallisation conditions from the Protein Data Bank. Version 1.0.0. Zenodo. https://doi.org/10.5281/zenodo.21807134 (concept DOI [10.5281/zenodo.21807133](https://doi.org/10.5281/zenodo.21807133) always resolves to the latest version)\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/TopPDBLX\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/TopPDBLX</a></td>\n</tr>\n</table>", "# When Protein Design Becomes a Security Problem\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-04-14 on marcdeller.com (categories: ai).*\n\n* [**Artificial Intelligence and the Future of Biosecurity**](https://www.frontiersin.org/research-topics/72195/artificial-intelligence-and-the-future-of-biosecurity)\n\n* **Screening blind spots are real:** AI-designed proteins can replicate the function of known toxins without any sequence similarity, evading current DNA synthesis biosecurity filters entirely.\n* **Accessibility compounds the risk:** Open-source generative protein design tools lower barriers to both beneficial and potentially harmful applications, demanding governance that matches the pace of the technology.\n* **Defensive applications are equally powerful:** The same AI biodesign capabilities that introduce risk can generate diagnostics, neutralising binders, and vaccine antigens against novel pathogens within hours.\n* **Pre-deployment evaluation is essential:** Biosecurity risk assessments for AI models must occur before public release, not after misuse has already demonstrated a vulnerability.\n\n1. **Protein Design, Generative AI and Biological Security:** [Brackmann et al., *Frontiers in Microbiology*, 2026](https://www.frontiersin.org/journals/microbiology/articles/10.3389/fmicb.2026.1817535/full)\n2. **Artificial Intelligence Challenges in the Face of Biological Threats:** Emerging Catastrophic Risks for Public Health: [de Lima et al., *Frontiers in Artificial Intelligence*, 2024](https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2024.1382356/full)\n3. **Responsible AI in Biotechnology: Balancing Discovery, Innovation and Biosecurity Risks:** [Wheeler et al., *Frontiers in Bioengineering and Biotechnology*, 2025](https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2025.1537471/full)\n4. **AI-Enabled Protein Design: A Strategic Asset for Global Health and Biosecurity:** [Stuart et al., *NAM Perspectives*, 2024](https://pmc.ncbi.nlm.nih.gov/articles/PMC11741535/)\n5. **Dual-Use Capabilities of Concern of Biological AI Models:** [Pannu et al., *PLOS Computational Biology*, 2025](https://journals.plos.org/ploscompbiol/article?id=10.1371%2Fjournal.pcbi.1012975)\n\n**How generative AI is reshaping the boundaries between biological innovation and existential risk**\n\nFor most of its history, protein engineering required deep expertise, costly equipment, and years of iterative experimentation. Designing a novel protein with a precisely defined structure and function was the kind of work that took entire research groups a decade to achieve. That constraint has collapsed. Deep learning models \u2014 trained on the vast structural and sequence information deposited in public databases \u2014 can now generate entirely novel protein sequences that fold predictably into defined architectures, often in minutes. The same computational shift that has accelerated vaccine development, diagnostics, and targeted therapeutics has also, quietly, created a new class of biological risk that existing safety infrastructure was never built to detect.\n\nThe core problem is elegantly simple and deeply unsettling. Standard biosecurity screening tools at DNA synthesis companies \u2014 the last checkpoint before a potentially dangerous sequence is physically manufactured \u2014 rely almost entirely on sequence homology. They search for similarity to known toxins, select agents, and regulated pathogens. [Brackmann et al. in *Frontiers in Microbiology*](https://www.frontiersin.org/journals/microbiology/articles/10.3389/fmicb.2026.1817535/full) make the critical point that AI-generated proteins can be functionally equivalent to known biological agents while sharing little or no sequence similarity with them, rendering homology-based screening effectively blind. A designed protein built from scratch using tools like RFdiffusion or ProteinMPNN might recapitulate the three-dimensional active site of ricin or a botulinum toxin with high fidelity, yet pass through every standard screening filter undetected. This is not a theoretical concern. A collaborative red-teaming effort, coordinated by Microsoft researchers and described by [de Lima et al.](https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2024.1382356/full), found that AI-generated protein variants designed to mimic biologically dangerous structures could evade multiple commercial screening systems entirely \u2014 prompting emergency patches now deployed globally.\n\nWhat makes this problem particularly acute is the accessibility of the underlying tools. The generative models capable of producing these sequences are largely open-source, well-documented, and require no specialist knowledge beyond basic computational biology. [Wheeler\u2019s review in *Frontiers in Bioengineering and Biotechnology*](https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2025.1537471/full) documents how the democratisation of AI-driven bioengineering has lowered the technical barrier to misuse, while simultaneously noting that the same tools underpin some of the most important advances in pandemic preparedness and medical countermeasure development seen in decades. The dual-use tension is not a future problem to be managed eventually \u2014 it is a present reality requiring immediate and proportionate governance responses. Voluntary community commitments, while a meaningful first step, cannot substitute for enforceable policy frameworks and technical safeguards embedded directly into the development pipeline.", "# Beyond Inhibition: The Rise of Conditional Degraders for Safer Medicines\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* Conditional degraders like RIPTACs and TriTACs are a next-generation TPD technology designed for enhanced safety and selectivity.\n* These molecules are engineered to activate only under specific biological conditions, such as in cells expressing a particular marker, limiting their action to diseased tissues.\n* This approach offers the potential for more \u201cprogrammable\u201d medicines with fewer side effects, improving the therapeutic window for potent drugs.\n\n1. **10 Evolved Terms Every Scientist Should Know About Targeted Protein Degradation in 2025:**\n   <https://resources.nanotempertech.com/blog/10-evolved-terms-every-scientist-should-know-about-targeted-protein-degradation-in-2025>\n2. **Targeted protein degradation: advances in drug discovery and therapeutic applications:** <https://www.nature.com/articles/s41392-024-02004-x>\n3. **CRDMO: H1 2025 Review of Targeted Protein Degraders:**\n   <https://wxpress.wuxiapptec.com/detail/1029/crdmo-h1-2025-review-of-targeted-protein-degraders>\n\nThe ideal drug would only act where it is needed, eliminating disease-causing proteins in sick cells while leaving healthy cells untouched. While targeted therapies have brought us closer to this goal, achieving perfect selectivity remains a major challenge, often leading to unwanted side effects. Targeted Protein Degraders (TPDs) offer a powerful way to eliminate harmful proteins, but how can we make them even smarter and safer?\n\nThe next evolution in this field is the development of \u201cconditional degraders.\u201d These are advanced TPDs designed to work only under specific biological conditions, such as in the presence of a second molecule that is unique to a disease state. Technologies like RIPTACs (Receptor-Induced Proximity Targeting Chimeras) are a prime example. A RIPTAC will only degrade its target protein in cells that also express a specific \u201cdocking\u201d receptor on their surface, offering an extra layer of tissue-specific targeting and safety.\n\nOther approaches, like TriTACs (Tri-specific TArgeting Chimeras), add a third arm to the molecule to improve control and selectivity. These innovations are moving the field toward what could be called \u201cprogrammable\u201d medicine. The ability to design degraders that require multiple conditions to be met before they activate could dramatically reduce off-target effects and improve the therapeutic window for potent drugs. This is a crucial step in creating therapies that are not only powerful but also exceptionally safe, bringing us closer to the ultimate goal of precision medicine.", "# GitHub README: simple-zsh-script-to-run-boltz2\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/simple-zsh-script-to-run-boltz2, 2026\n\n> **A minimal, auditable wrapper for Boltz-1/2. Every run logged, timed, and reproducible.**\n\n    \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/simple-zsh-script-to-run-boltz2\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/simple-zsh-script-to-run-boltz2</a></td>\n</tr>\n</table>\n\n---\n\nA minimal, production-ready shell wrapper for running [Boltz-1/Boltz-2](https://github.com/jwohlwend/boltz) structure predictions from the command line. Handles logging, runtime tracking, and input sanitisation automatically so every run is reproducible and auditable.\n\nWhy it matters: running co-folding predictions by hand invites silent failures: overwritten logs, stray macOS dot-files derailing a batch, no record of how long a run took or what flags were used. run_boltz.zsh wraps Boltz in a thin, production-minded shell layer that timestamps every log, tracks wall-clock runtime, sanitises the input directory, and tees output to both screen and file. It is useful for anyone running Boltz at scale or in a regulated setting where reproducibility and an audit trail matter: drop it in, point it at your inputs, and every prediction becomes self-documenting.\n\n---\n\n## Features\n\n- **Timestamped log files** \u2014 every run writes to `boltz_run_YYYYMMDD_HHMMSS.log`; nothing is ever silently overwritten\n- **Runtime tracking** \u2014 total wall-clock time reported at the end of each run\n- **Input sanitisation** \u2014 removes macOS/Linux hidden dot-files (`.DS_Store`, `._*`, etc.) from the inputs directory before prediction\n- **MPS fallback** \u2014 sets `PYTORCH_ENABLE_MPS_FALLBACK=1` for Apple Silicon compatibility\n- **Unified stdout + log** \u2014 all output is tee'd to the terminal and the log file simultaneously via `2>&1 | tee`\n\n---\n\n## Requirements\n\n- [Boltz](https://github.com/jwohlwend/boltz) installed and on your `$PATH`\n- `zsh` (macOS default; also available on Linux)\n- A CUDA-capable GPU (the script passes `--accelerator gpu`)\n- Internet access if MSA server lookups are needed (`--use_msa_server`)\n\n---\n\n## Directory layout\n\n```\nproject/\n\u251c\u2500\u2500 run_boltz.zsh       # this script\n\u251c\u2500\u2500 inputs/             # place your input YAML / FASTA / CIF files here\n\u2514\u2500\u2500 output/             # Boltz outputs written here (created automatically)\n```\n\n---\n\n## Usage\n\n```zsh\nchmod +x run_boltz.zsh\n./run_boltz.zsh\n```\n\nThat's it. The script picks up all inputs from `./inputs/`, writes predictions to `./output/`, and creates a timestamped log in the current directory.\n\n---\n\n## Configuration\n\nAll key parameters are set at the top of the script:\n\n| Variable | Default | Description |\n|---|---|---|\n| `INPUT_DIR` | `./inputs` | Directory containing Boltz input files |\n| `OUTPUT_DIR` | `./output` | Root output directory |\n| `LOGFILE` | `boltz_run_<timestamp>.log` | Per-run log filename |\n\nBoltz prediction flags can be edited directly in the `boltz predict` call:\n\n| Flag | Value | Notes |\n|---|---|---|\n| `--use_potentials` | \u2014 | Enables potential energy guidance during diffusion |\n| `--diffusion_samples` | `1` | Number of diffusion samples per input; increase for ensemble diversity |\n| `--use_msa_server` | \u2014 | Queries the ColabFold MSA server; remove for purely local runs |\n| `--num_workers` | `8` | DataLoader workers; tune to your CPU core count |\n| `--accelerator` | `gpu` | Use `cpu` for debugging on machines without a GPU |\n\n---\n\n## Output\n\nEach run produces:\n\n```\noutput/\n\u2514\u2500\u2500 boltz_results_<input_name>/\n    \u251c\u2500\u2500 predictions/\n    \u2502   \u2514\u2500\u2500 <model>_model_0.cif      # ModelCIF-format predicted structure\n    \u251c\u2500\u2500 confidence_<model>_model_0.json\n    \u2514\u2500\u2500 ...\n\nboltz_run_20250611_143022.log        # full stdout/stderr capture\n```\n\nPredicted structures are written as **ModelCIF** files. These can be loaded directly in PyMOL, UCSF ChimeraX, or processed downstream with tools such as [Biotite](https://www.biotite-python.org/) or a PLIP-based interaction pipeline.\n\n---\n\n## Notes\n\n- The `cleanup_input_dir` function removes hidden files silently before each run. This prevents spurious Boltz errors on macOS where Finder writes `._<filename>` resource fork files alongside legitimate inputs.\n- Log files accumulate in the working directory; archive or delete them periodically if running many predictions.\n- To run multiple diffusion samples for ensemble analysis, increase `--diffusion_samples` (e.g. `5`) and downstream results will include `model_0` through `model_N`.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/simple-zsh-script-to-run-boltz2\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/simple-zsh-script-to-run-boltz2</a></td>\n</tr>\n</table>", "# Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\nAuthors: Kong, L.; Lee, J.H.; Doores, K.J.; Murin, C.D.; Julien, J.-P.; McBride, R.; Liu, Y.; Marozsan, A.; Cupo, A.; Klasse, P.-J.; Hoffenberg, S.; Caulfield, M.; King, C.R.; Hua, Y.; Le, K.M.; Khayat, R.; Deller, M.C.; Clayton, T.; Tien, H.; Feizi, T.; Sanders, R.W.; Paulson, J.C.; Moore, J.P.; Stanfield, R.L.; Burton, D.R.; Ward, A.B.; Wilson, I.A.\nSource: Nature Structural & Molecular Biology, 2013\n\nVirus neutralization and binding assays revealed differential effects of these N-linked glycosylation sites on PGT 135 activity against different isolates of HIV-1 (Fig. 3d). Not surprisingly, Asn332 and Asn392 were required for all isolates tested because most of the antibody contacts are to these glycans. However, the requirement for Asn295 (N295) and Asn386 glycans, as well as His330 residue, on gp120, was strain dependent. Although the Asn295 glycan does not contact PGT 135 in the context of the JR-FL core gp120, these data suggest it may be recruited by the antibody in the context of other HIV-1 strains so that the epitope would be modified and strain-dependent. Thus, although the PGT 135 interaction is highly specific for particular glycoforms on gp120, it may also be promiscuous by exploiting a neighboring Asn295 glycan for recognition of some HIV-1 strains. By targeting the oligomannose patch on gp120 where the required glycoforms are present, while tolerating a certain range of alternate glycan positions, PGT 135 may acquire broad neutralizing breadth despite the heterogeneity seen in the Env glycan populations. Overall, the data point towards two general requirements for PGT 135 binding, with the caveat that any given isolate may deviate from the most common pattern of recognition: 1) high mannose glycans within the context of the tightly clustered high mannose patch on gp120 and 2) a D2 arm on the glycan at Asn392 lacking at least the terminal mannose residue.\n\n### Structural explanation for neutralization breadth of PGT 135\n\nPGT 135 can neutralize 33% of HIV-1 isolates at an IC50 < 50 \u03bcg ml\u22121 using a 162 cross-clade virus panel, which is comparable to bnAb b12 (34%), but less than PGT 128 (72%), although PGT 135 and 128 both recognize epitopes critically dependent upon the same highly conserved Asn332 glycan. The variability of important residues and glycan sites was examined using 3045 aligned gp120 sequences (Los Alamos HIV Database). The sequence conservation of the essential elements necessary for gp120 binding agrees well with neutralization breadth (Supplementary Table 2). For PGT 128, glycans at Asn332 (73% conserved) and Asn301 (N301) (92% conserved) are critical for neutralization; both glycan motifs are present together in 69% of the sequences, which is close to the observed 72% neutralization breadth. For PGT 135, glycans at Asn332, Asn392, Asn295, Asn386, and His330 (73%, 79%, 59%, 87%, and 71% conserved, respectively) are all important for neutralization (Supplementary Table 2). His330, Asn332 and Asn392 are all required for neutralization of all isolates tested (Fig. 3d) and these three sites are found concurrently in 50% of the sequences. However, PGT 135 recognition also requires Asn295 and Asn386 glycans in a strain-dependent manner. His330, glycans at Asn332, Asn392, and either Asn295 or Asn386 are found in 31% or 42% of the sequences, while His330 along with all four glycan motifs (332, 392, 295, and 386) is present in 26% of the sequences. The observed neutralization breadth of PGT 135 is 33% of test isolates, suggesting that the dispensability of either Asn295 or the Asn386 glycans in certain strains enhances neutralization breadth compared to a requirement for all four glycans.\n\n### PGT 135 interaction with gp120 trimer\n\nWe used electron microscopy (EM) to characterize PGT 135 Fab binding to a glycosylated clade A BG505 SOSIP.664 gp140 trimer. Three Fabs were bound per trimer, with the gp120 protomers in a more closed conformation (Fig. 4a) than the more open conformation seen with CD4 binding site ligands. The PGT 135 angle of approach to the gp120 epitope avoids any clash with neighboring gp120 protomers or interference with the CD4 binding site (Fig. 4a). The 2D class averages showed that PGT 135 occupies a distinct epitope from PGT 128, which also recognizes the Asn332 glycan (Fig. 4b), but we observed significant variance in the orientation of Fab PGT 135 relative to the trimer (Fig. 4c, Supplementary Movie 1). Principal component analysis (PCA) of the variance clearly illustrated some flexibility in the Fab PGT 135 interaction with the trimer compared to PGT 128 (Fig. 4d, Supplementary Movie 2). This increased variance for PGT 135 may represent a distribution of different angles of approach due to some slight shift between the interacting glycans or with the glycoforms it contacts (Fig. 4d). The use of alternative angles of approach may therefore reflect an immunologic countermeasure to some natural variation in the N-linked glycans that decorate the Env surfaces of different HIV-1 isolates.\n\n## Discussion\n\nRecent advances in our understanding of immune responses against HIV-1 Env have revealed that extremely potent, broadly neutralizing responses can be made against the glycosylated face of gp120, particularly around the Asn332 glycan. PGT 135, 2G12 and PGT 128 bnAbs are all dependent on the Asn332 glycan but have distinct binding motifs and quite different angles of approach, and yet their neutralization activity is not compromised (Fig. 5a). This scenario contrasts with antibodies whose binding footprints overlap the CD4 binding site where small differences in angle and relative position result in clashes with neighboring gp120s and very different extents of neutralization from VRC01 (bnAb, 93% breadth) to F105 (non-neutralizing). Thus, these Asn332-dependent antibodies define a novel and unexpectedly broad supersite for neutralization (Fig. 5b), which encompasses 6 conserved glycans and the underlying protein surface.", "# Molecular Architecture of GLP-1 Receptor Drug Binding Reveals Dynamic Activation Mechanisms\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-12-12 on marcdeller.com (categories: sb).*\n\n* [Evolving cryo-EM structural approaches for GPCR drug discovery](https://www.cell.com/structure/fulltext/S0969-2126(21)00124-6?_returnURL=https%3A%2F%2Flinkinghub.elsevier.com%2Fretrieve%2Fpii%2FS0969212621001246%3Fshowall%3Dtrue)\n\n* **Deep Pocket Architecture**: The [orthosteric binding pocket](https://www.nature.com/articles/s41467-021-27760-0) comprises Y152 (TM1), K197 (TM2), Q234/Y241 (TM3), W306/R310 (TM5), E364/H363 (TM6), and L384 (TM7), with His1 of peptide agonists forming critical contacts deep within this cavity that stabilize the active conformation.\n* **Dynamic Binding Correlates with Efficacy**: Molecular dynamics simulations reveal that [GLP-1 forms more persistent interactions](https://www.nature.com/articles/s41467-021-27760-0) at the binding pocket base compared to oxyntomodulin or exendin-4, with transient versus stable engagement patterns directly affecting G protein activation kinetics and signaling bias.\n* **TM6 Conformational Switch**: Receptor activation requires [outward movement of TM6\u2019s intracellular region by approximately 11.9 \u00c5](https://www.nature.com/articles/s41421-024-00649-0) to create the Gs binding cavity, with class B GPCRs exhibiting unique kinetics where Gs engagement can occur before full TM6 opening and subsequent slow relaxation enables prolonged signaling.\n* **ECD as Intrinsic Agonist**: The extracellular domain functions beyond peptide capture, acting as an [intrinsic agonist](https://www.nature.com/articles/celldisc201642) that directly activates the transmembrane domain through ECL interactions, with its conformational transition from closed to extended position being absolutely required for receptor activation even with covalently tethered peptides.\n\n1. **Dynamics of GLP-1R peptide agonist engagement are correlated with kinetics of G protein activation**: [Wootten et al., Nature Communications, 2022](https://www.nature.com/articles/s41467-021-27760-0)\n2. **Structural analysis of the dual agonism at GLP-1R and GCGR**: [Zhao et al., PNAS, 2023](https://www.pnas.org/doi/10.1073/pnas.2303696120)\n3. **Molecular features of the ligand-free GLP-1R, GCGR and GIPR in complex with Gs proteins**: [Qiao et al., Cell Discovery, 2024](https://www.nature.com/articles/s41421-024-00649-0)\n4. **An intrinsic agonist mechanism for activation of glucagon-like peptide-1 receptor**: [Wootten et al., Cell Discovery, 2016](https://www.nature.com/articles/celldisc201642)\n\n**Structure-based insights illuminate key residues, conformational transitions, and binding pocket characteristics that dictate therapeutic agonist engagement**", "# Plastic, Persistent, and Potentially Carcinogenic\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nThe molecular architecture of these effects is becoming well-defined. As reviewed in comprehensive detail by [Alaraby et al.](https://link.springer.com/article/10.1186/s12943-026-02657-y) in Molecular Cancer, MNP exposure activates a coordinated network of oncogenic signaling programs: NF-\u03baB drives pro-inflammatory and pro-survival gene expression; PI3K/AKT/mTOR and MAPK/ERK pathways promote proliferation and metabolic reprogramming; Wnt/beta-catenin links ROS-driven stress to tumorigenic cellular reprogramming; and ECM-receptor and focal adhesion remodeling supports epithelial-mesenchymal transition. The question the field must now answer is not whether microplastics are biologically active in the tumor context, but which physicochemical properties, particle size, surface charge, polymer chemistry, and the chemical cargo of adsorbed carcinogens, determine which oncogenic programs are activated and in which tissue compartments. For researchers at the structural and computational biology interface of drug discovery, the signaling nodes being activated by MNP exposure, many of them well-characterized and structurally tractable, may represent a new and genuinely addressable class of environmental oncology targets.", "# Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\nAuthors: Ye, Q.; Shvartsbart, A.; Li, Z.; Gan, P.; Policarpo, R.L.; Qi, C.; Roach, J.J.; Zhu, W.; McCammant, M.S.; Hu, B.; Li, G.; Yin, H.; Carlsen, P.; Hoang, G.; Zhao, L.; Susick, R.; Zhang, F.; Lai, C.-T.; Allali Hassani, A.; Epling, L.B.; Gallion, A.; Kurzeja-Lipinski, K.; Gallagher, K.; Roman, V.; Farren, M.R.; Kong, W.; Deller, M.C.; Zhang, G.; Covington, M.; Diamond, S.; Kim, S.; Yao, W.; Sokolsky, A.; Wang, X.\nSource: Journal of Medicinal Chemistry, 2025\n\nIn summary, we demonstrate the discovery of a potent and selective KRAS G12D inhibitor with promising target engagement in vivo. Due to uncertainties about the druggability of KRAS G12D, we utilized a two-step strategy. First, proof of concept for potent inhibition of KRAS G12D was achieved by focusing on identifying additional interactions with the protein. Second, we were able to balance the delicate interplay of potency and bioavailability by identifying key structural features interfering with oral exposure, removing them despite the significant loss of binding affinity, then building potency back with a focus on properties. Rigidifying the structure proved key in this endeavor, increasing potency by reducing the degrees of freedom while simultaneously improving ADME. These changes allowed for promising oral exposure and bioavailabiliy in multiple preclinical species. The resulting compound 23 (INCB159020) serves as an important tool compound and provides a framework for the design of future bioavailable KRAS inhibitors.\n\n## Experimental Section\n\n### General Information\n\nUnless otherwise specified, all solvents were purchased from commercial sources and used without further drying. Column chromatography was performed using prepackaged columns (RediSep Rf or SNAP) using Biotage Isolera or Teledyne ISCO CombiFlash system. NMR spectra were obtained using a Varian Mercury-400, Inova-500, or a Bruker Advance III HD 600 MHz spectrometer equipped with a 5 mm QCI Cryoprobe. Data for 1H spectra are reported as follows: chemical shift (\u03b4 ppm), multiplicity, coupling constant (Hz), integration and are referenced to the residual solvent peak (2.50 for DMSO-d6 and 7.26 ppm for CDCl3). Data for 13C spectra are reported as follows: chemical shift (\u03b4 ppm), multiplicity, coupling constant (Hz) and are referenced to the residual solvent peak (40 ppm for DMSO-d6) and are proton decoupled. Preparative LC purifications were performed on a Waters FractionLynx system using UV-triggered or mass directed fractionation and compound-specific method optimization. For purification of final compounds, the reaction mixture was diluted with methanol or acetonitrile to 6 mL total volume, and then filtered through a syringe filter. Purification was then performed with prep-LCMS with the following general conditions: XBridge C18 column, gradient of acetonitrile/water containing 0.1% TFA, flow rate of 60 mL/min. Unless otherwise indicated, compounds were isolated as the corresponding TFA salt. All final compounds for biological testing were purified by reverse phase prep-HPLC to >95% purity as determined by analytical HPLC.\n\n### Synthesis of 23\n\n**Step 1. 2-Amino-4-bromo-3-fluoro-5-iodobenzoic acid (25).** NIS (21.15 g, 94 mmol) was added to a solution of 2-amino-4-bromo-3-fluorobenzoic acid (20 g, 85 mmol) in DMF (200 mL) and the reaction was stirred at 80 \u00b0C for 3 h. The mixture was cooled with ice water and then water (500 mL) was added, the precipitate was filtered, washed with water and dried to provide the desired product as a solid. The crude solid was used in the next step without further purification. LC-MS calculated for C7H5BrFINO2 (M + H)+: m/z = 359.9, 361.9; found 359.9, 361.9.\n\n**Step 2. 7-Bromo-8-fluoro-6-iodo-2H-benzo[d][1,3]oxazine-2,4(1H)-dione (26).** Triphosgene (9.07 g, 30.6 mmol) was added to a solution of 25 (22 g, 61.1 mmol) in dioxane (200 mL) and the reaction was stirred at 80 \u00b0C for 2 h. The reaction mixture was cooled with ice water and then filtered. The solid was washed with ethyl acetate to provide the desired product as a solid. The crude product was used in the next step without further purification. LC-MS calculated for C8H3BrFINO3 (M + H)+: m/z = 385.8, 387.8; found 385.8, 387.8.\n\n**Step 3. 7-Bromo-8-fluoro-6-iodo-3-nitroquinoline-2,4-diol (27).** DIPEA (25.5 mL, 146 mmol) was added to a solution of ethyl 2-nitroacetate (16.33 mL, 146 mmol) and 26 (20 g, 73.0 mmol) in toluene (200 mL) at r.t. The reaction was stirred at 95 \u00b0C for 3 h. The reaction was cooled and then filtered, the solid was washed with a small amount of hexanes to provide the desired product. The crude product was used in the next step without further purification. LC-MS calculated for C9H4BrFIN2O4 (M + H)+: m/z = 428.8, 430.8; found 428.8, 430.8.\n\n**Step 4. 7-Bromo-2,4-dichloro-8-fluoro-6-iodo-3-nitroquinoline (28).** DIPEA (8.14 mL, 46.6 mmol) was added to a mixture of 27 (10 g, 23.31 mmol) in POCl3 (10.86 mL, 117 mmol). The reaction mixture was stirred at 100 \u00b0C for 2 h. The solvent was removed under vacuum and then azeotroped with toluene 3 times to provide the crude material which was purified with flash column chromatography (0\u2212100% ethyl acetate in hexanes) to provide the desired product. LC-MS calculated for C9H2BrCl2FIN2O2 (M + H)+: m/z = 464.8, 466.8; found 464.8, 466.8.\n\n**Step 5. tert-Butyl (1R,4R,5S)-5-((7-Bromo-8-fluoro-6-iodo-2-(methylthio)-3-nitroquinolin-4-yl)amino)-2-azabicyclo[2.1.1]hexane-2-carboxylate (35).** To a solution of 28 (220.6 g, 474 mmol) and tert-butyl (1R,4R,5S)-5-amino-2-azabicyclo[2.1.1]hexane-2-carboxylate (96 g, 483 mmol) in DMF (600 mL) was added Hunig's base (103 mL, 592 mmol) and the reaction mixture was heated to 60 \u00b0C for 1 h. The reaction mixture was then cooled to 0 \u00b0C in an ice bath and sodium thiomethoxide (100 g, 1421 mmol) and MeOH (100 mL) were added. The mixture was stirred at 0 \u00b0C for 1 h, then poured into ice water. The resulting precipitate was filtered, washed with ice water, then dried under vacuum to afford the desired product (297.6 g, 98%). LC-MS calculated for C20H22BrFIN4O4S (M + H)+: m/z = 639.0; found 639.1.", "# GitHub README: FlexAppeal\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/FlexAppeal, 2026\n\n`/static/` deliberately keeps `access_log` on, and a test enforces that too. The [mdeller.com launcher](https://github.com/bellcheddar/mdeller-landing) counts a visit to this app by watching its access log for `/static/app.js`: a request only a browser that rendered the page ever makes, since a scanner fetches the HTML and stops. Turning logging off for that location is invisible in every other respect and simply pins the hit count at zero.\n\n## \u2705 To Do\n\nRoadmap, newest ideas at the top. Suggestions welcome.\n\n- [ ] Espaloma charges as a faster alternative to AM1-BCC, which takes minutes per ligand\n- [ ] Replicate comparison in the Analysis tab: overlay several `.fxa` files on one panel\n- [ ] Trajectory concatenation, so an extended run analyses as one piece\n- [ ] Umbrella-sampling and steered-MD protocols (currently out of scope by design)\n- [ ] A \"reproduce this\" button that rebuilds a bundle from an uploaded `.fxa`\n- [x] **Trajectory playback on load.** The structure panel starts looping as soon as the frames are in, one pass every 50 seconds, skipped for anyone whose system asks for reduced motion.\n- [ ] **Work around the OpenCL memory leak.** OpenMM's OpenCL platform grows by roughly 3 kB per integration step on an M1 Max, measured with no reporters attached, so it is the platform rather than anything FlexAppeal does. The 10 ns reference run ended 8 GB up; a 100 ns run of the same system would need about 80 GB. Neither `Context.reinitialize(preserveState=True)` nor destroying and rebuilding the `Simulation` reclaims it. The run now measures its own slope and warns in time to restart from the checkpoint, which does release it, but a bundle that segments a long run across processes automatically would be better than asking the user to notice.\n- [x] **A worked example.** A real 10 ns run of hen egg-white lysozyme (1AKI, 19,433 atoms with water) is committed to the repository, and the Example tab renders it through the same code path as an upload: the panels come from `plots.build_all()`, the Mol\\* viewer loads the topology and trajectory out of the same `.fxa`, and the option tables are read from the registry. So the page cannot drift from the product -- if the Analysis tab breaks, this breaks with it, in CI. The terminal captures are the bytes the run actually printed, replayed through the library that produced them. `bash examples/lysozyme_10ns/rebuild.sh` regenerates all of it.\n- [x] **Apple Silicon runs on the GPU.** OpenMM's OpenCL platform was being disqualified on every Mac: Apple's implementation supports single precision only, the default configuration asks for mixed (correct on CUDA), and context creation fails with \"No compatible OpenCL platform is available\" -- which names the platform rather than the property, so the benchmark caught it and silently fell back to the CPU. Measured on solvated lysozyme, 19,433 atoms, M1 Max: **29.5 ns/day on CPU against 206 on OpenCL**, so a 10 ns run took eight hours instead of seventy minutes. Unsupported properties are now dropped rather than fatal, the substitution is announced, and the check probes the platform in front of it rather than hard-coding a list.\n- [x] **Rich terminal output.** The CLI and both generated scripts share one palette and print through [rich](https://github.com/Textualize/rich): tables for structure reports, determinate bars for the integration stages, an indeterminate spinner for minimisation (whose iteration counter is not monotonic, so a percentage would run backwards), and a live memory readout beside every bar. Swap is the figure that matters -- a run that spills out of RAM slows by an order of magnitude with no other symptom -- so it is on screen throughout and warns once, loudly, if this run is what pushed the machine there. Off a terminal it degrades to plain milestone lines with no escape sequences, which is what makes a `nohup` log readable.\n- [x] **Deployed to [flexappeal.mdeller.com](https://flexappeal.mdeller.com).** Port 8004 on the shared droplet, behind nginx with a Let's Encrypt certificate and HTTP/2. Verified in production rather than merely health-checked: a structure upload renders all 115 options and finds lysozyme's four disulfides, a bundle builds and its generated scripts compile, and a results file renders eleven panels. Each application on the box serves its own certificate, which is the failure the unconditional-certbot rule exists to prevent.\n- [x] **Server-side re-analysis.** The Analysis tab can compute a metric the run did not, or restrict one to part of the molecule, against the trajectory packed in the results file. Bounded by an atom-by-frame budget, a separate ceiling on quadratic pairwise work, one job at a time via an expiring lock, and a cgroup memory limit. The work runs in a detached subprocess so MDTraj never blocks a worker.\n- [x] **Membrane systems.** `Modeller.addMembrane` with seven lipids, pre-oriented structures from OPM, the membrane barostat, and area-per-lipid, thickness and deuterium order parameter panels.\n- [x] **Ligands and cofactors.** Chemistry from the RCSB Chemical Component Dictionary transferred onto the deposited coordinates, parameterised with OpenFF Sage, GAFF2 or espaloma, with ligand RMSD and per-residue contact occupancy.\n- [x] **Analysis tab.** Nine panels plus six convergence small multiples, a Mol\\* viewer, stat tiles and a table view. The categorical palette is validated rather than chosen: three of the brand accents failed the contrast and lightness checks for chart use and were re-stepped.\n- [x] **Self-contained run bundles.** One file carrying the structure, a pinned environment, a readable OpenMM script, the analysis pass and its own installer.\n- [x] **The option registry.** 115 options declared once as data; the form, the validator, the script template and the documentation all read from it.\n\n## \ud83d\udcdd Licence\n\nMIT. See [`LICENSE`](LICENSE) for the text, and [`NOTICE.md`](NOTICE.md) for the third-party notices.", "# marcdeller.com: Home\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n## Hidden Architecture: Mapping Biology\u2019s Structure\n\nExpert structural biologist and drug discovery leader with over 10+ years of experience. Specialization in protein structure, X-ray crystallography, Cryo-EM, protein production, mechanisms of action, receptor biology, molecular pharmacology, automated research platforms, and AI-driven therapeutics, contributing to major advances in oncology and immunology. This experience encompasses [400+ protein structures](https://marcdeller.com/structures/), [60+ peer-reviewed publications](https://marcdeller.com/publications/), [7+ patents](https://marcdeller.com/patents/), [6+ FDA IND filings across oncology](https://marcdeller.com/my-work-driving-innovation-in-precision-oncology-through-structure-based-drug-design/), immunology, and virology, and vibe-coded projects including [BoltzMaker](https://marcdeller.com/boltzmaker-vibe-coding-a-campaign-scale-boltz-2-pipeline/),\u00a0[ChemSage](https://marcdeller.com/chemsage/) and [MicroPlastics Blaster](https://marcdeller.com/vibe-coding-microplastic-blaster/) are available as web apps at [mdeller.com](https://mdeller.com).\n\nFeatured Posts\n\n### [ChemSage: Vibe-Coding a Chemically Aware Drug-Discovery LLM on Local Hardware](https://marcdeller.com/chemsage/)\n\n[ChemSage: Vibe-Coding a Chemically Aware Drug-Discovery LLM on Local Hardware](https://marcdeller.com/chemsage/)\n\n### [My Work Driving Innovation in Precision Oncology Through Structure-Based Drug Design](https://marcdeller.com/my-work-driving-innovation-in-precision-oncology-through-structure-based-drug-design/)\n\n[My Work Driving Innovation in Precision Oncology Through Structure-Based Drug Design](https://marcdeller.com/my-work-driving-innovation-in-precision-oncology-through-structure-based-drug-design/)\n\n### [Building a Comprehensive Protein Production, X-ray Crystallography, and Cryo-EM Laboratory from Scratch](https://marcdeller.com/building-a-comprehensive-protein-production-x-ray-crystallography-and-cryo-em-laboratory-from-scratch/)\n\n[Building a Comprehensive Protein Production, X-ray Crystallography, and Cryo-EM Laboratory from Scratch](https://marcdeller.com/building-a-comprehensive-protein-production-x-ray-crystallography-and-cryo-em-laboratory-from-scratch/)\n\n### [JAK2-V617F and the Evolution of Targeted Therapy: Precision Medicine Transforms Myeloproliferative Neoplasm Treatment](https://marcdeller.com/jak2-v617f-and-the-evolution-of-targeted-therapy-precision-medicine-transforms-myeloproliferative-neoplasm-treatment/)\n\n[JAK2-V617F and the Evolution of Targeted Therapy: Precision Medicine Transforms Myeloproliferative Neoplasm Treatment](https://marcdeller.com/jak2-v617f-and-the-evolution-of-targeted-therapy-precision-medicine-transforms-myeloproliferative-neoplasm-treatment/)\n\n### [Elora Therapeutics: Treating Microplastics in Humans](https://marcdeller.com/elora/)\n\n[Elora Therapeutics: Treating Microplastics in Humans](https://marcdeller.com/elora/)\n\nAbout Me\n\nArtificial Intelligence\n\nBlogs\n\nMicroplastics in Humans\n\nMy Learning\n\nMy Patents\n\nMy Projects\n\nMy Publications\n\nMy Structures\n\nMy Web Apps\n\nStructural Biology\n\nMy Vibe Coding\n\n[Linkedin](https://www.linkedin.com/in/marccdeller/)\n\n[Google](https://scholar.google.com/citations?user=o02oSJIAAAAJ&hl=en&oi=ao)\n\n[Researchgate](https://www.researchgate.net/profile/Marc-Deller?ev=hdr_xprf)\n\n[Robot](https://marcdeller.com/chat/)\n\n* [(302) 358-6093](tel:3023586093)\n* Wilmington, Delaware, USA\n* [E-mail](mailto:marc@marcdeller.com)", "# 2018 \u2013 Structure\u2013function analysis of the extended conformation of a polyketide synthase module\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: papers).*\n\n* [Link to Paper](https://marcdeller.com/wp-content/uploads/2025/09/nihms-1016244.pdf)\n\n* **Extended Architecture:** Antibody stabilization reveals a linear domain arrangement supporting efficient catalysis.\n* **Catalytic Competence:** Kinetic assays show the elongated state facilitates both translocation and elongation steps.\n* **Dynamic Flexibility:** Linker motions couple condensations to mechanical gating, ensuring directional chain transfer.\n* **Asymmetric Interactions:** Cryo-EM studies uncover selective KS\u2013ACP contacts that discriminate intra- and intermodular transfer.\n* **Engineering Potential:** Biasing modules toward the extended conformation could enhance non-natural polyketide production.\n* **Real-Time Probing:** Conformation-locking agents offer avenues for monitoring biosynthesis step by step.\n\n1. **Structure\u2013Function Analysis of the Extended Conformation of a Polyketide Synthase Module:** Xiuyuan Li, et.al., [Journal of the American Chemical Society](https://pubs.acs.org/doi/abs/10.1021/jacs.8b02100)\n2. **Structure and Mechanisms of Assembly-Line Polyketide Synthases:** Alexander M. Soohoo, et.al., [Annual Review of Biochemistry](https://www.annualreviews.org/content/journals/10.1146/annurev-biochem-080923-043654)\n3. **Architecture of full-length type I modular polyketide synthases revealed by X-ray crystallography, cryo-electron microscopy, and AlphaFold2:** Saket R. Bagde, et.al., [Natural Product Reports](https://pubs.rsc.org/en/content/articlelanding/2024/np/d3np00060e)\n4. **Structural basis for intermodular communication in assembly-line polyketide synthases:** Dillon P. Cogan, et.al., [Nature Chemical Biology](https://pubmed.ncbi.nlm.nih.gov/39179672/)\n5. **Modular polyketide synthase contains two reaction chambers that operate asynchronously:** Saket R. Bagde, et. al., [Science](https://pubmed.ncbi.nlm.nih.gov/34735234/)\n6. **Solution Structure and Conformational Flexibility of a Polyketide Synthase Module:** Matthias Huber, et.al., [JACS Au](https://pubs.acs.org/doi/10.1021/jacsau.1c00043)\n\nResolves a long-standing debate in polyketide synthase (PKS) biology by demonstrating\u2014via innovative antibody stabilization, structural validation (X-ray/SAXS), and kinetic assays\u2014that the extended module conformation is catalytically functional for both chain elongation and modification, unlocking critical insights for engineered antibiotic production.\n\nIn the realm of natural product biosynthesis, polyketide synthases stand as monumental assembly lines, orchestrating the stepwise elongation and modification of molecular scaffolds. A recent deep dive into the extended conformation of a module within this enzymatic behemoth has challenged prior assumptions about static architectures and revealed an adaptable framework that underpins catalytic efficiency. By stabilizing the module in its elongated state, researchers have shown that this form is fully competent for both chain translocation and elongation steps, offering a new vantage on how domain motions coordinate chemistry with choreography.\n\nProbing this extended form began with antibody stabilization, capturing the module in a configuration once thought incidental. X-ray crystallography and SEC-SAXS confirmed that the extended architecture aligns active-site domains linearly, juxtaposing the acyl carrier protein and ketosynthase in a conformation that facilitates rapid substrate handoff. Kinetic assays then demonstrated that this elongated state catalyzes intermodular translocation and intramodule elongation with efficiencies matching or exceeding the arch-shaped form.\n\nThese findings illuminate how conformational flexibility serves a functional imperative rather than a by-product of modular design. The extended form appears to act as a \u201cfully loaded track\u201d along which the growing polyketide chain can traverse multiple active sites without steric hindrance. This insight aligns with emerging cryo-EM snapshots of asymmetric KS\u2013ACP interactions, where selective domain engagement governs the directionality of polyketide transfer. Advanced computational models further suggest that transient bending and straightening of linkers modulate pathway gating, coupling thermodynamic favorability of Claisen-like condensations to mechanical toggling of the module.\n\nLooking forward, the recognition that an extended conformation is catalytically robust reframes strategies for engineering hybrid synthases. By designing linkers and interface surfaces that bias this architecture, one could streamline non-natural polyketide assembly and improve yields. Moreover, antibody or small-molecule probes tailored to lock modules in extended states could become tools to dissect stepwise catalysis in real time. Ultimately, embracing the dynamic landscape of PKS modules promises to unlock novel routes to designer molecules and deepen our understanding of nature\u2019s most sophisticated biosynthetic machines.", "# Protein stability: a crystallographer's perspective\nAuthors: Deller, M. C.; Kong, L.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 72, 72\u201395 (IYCr crystallization series), 2016\n\nHost choice influences whether the target is toxic to the cell and whether it is cleaved by endogenous proteases. Toxicity is managed by tight, concentration-dependent regulation of expression, as with the L-arabinose-responsive pBAD vector. Background expression is reduced with pLysS, which expresses T7 lysozyme, a natural inhibitor of the T7 RNA polymerase used in pET systems, and background protease activity is reduced in OmpT-deficient strains such as BL21 Star (Invitrogen/ThermoFisher), which lack the outer-membrane aspartyl protease. Codon usage matters when mammalian proteins are made in bacteria: the arginine codon AGA is rare in *E. coli* and can cause premature termination, frameshifting and misincorporation. Remedies are synthetic genes matched to host codon usage, or co-transformation with rare-tRNA plasmids such as CodonPlus (Stratagene) and pRARE (EMD Millipore/Novagen), the latter available commercially in BL21 cells as Rosetta \u2014 used by the Swedish Human Protein Atlas project to improve both expression level and purity. For mammalian and human proteins, CHO cells and HEK293T/293F cells are often essential for correct PTMs and folding, and insect cells such as *Spodoptera frugiperda* (Sf9, Sf21) and *Trichoplusia ni* offer another route.\n\n### 5. Key techniques for determining protein stability\n\nTable 2 compares the three main structural methods. Because crystallography is a solid-state method, dynamic and disordered proteins are hard to crystallize, which is why NMR has been central to studying IDPs such as p53 and CREB and is well suited to assessing secondary and tertiary structure in dynamic systems. Crystallography's higher resolution makes it particularly attractive for primary-structure changes \u2014 uncharacterized PTMs such as glycosylation can often be read directly from electron-density maps when the data are of sufficiently high resolution. Cryo-EM, given sufficiently high-resolution maps, is powerful for stability and dynamics at the quaternary level. The section then treats NMR (\u00a75.1), electron microscopy (\u00a75.2), spectroscopic methods (\u00a75.3) and melting-temperature analysis (\u00a75.4) in turn, with the practical emphasis throughout on how much material each method consumes, what conditions it can probe, and whether the sample survives for downstream crystallization.\n\n### 5.1. Nuclear magnetic resonance (NMR)\n\nNMR determines stability in solution and is highly complementary to X-ray analysis. Modern instruments need minimal sample (of the order of 10 nM), and even a simple one-dimensional proton spectrum reports on conformational stability: because of the inverse relation between spin\u2013spin relaxation time and peak width, large soluble aggregates give no interpretable high-resolution spectrum. For non-aggregated samples, good discrimination in the backbone amide region below 8.3 p.p.m., together with peaks around 1 p.p.m., indicates folded protein. Two-dimensional \u00b9H\u2013\u00b9\u2075N heteronuclear single-quantum coherence (HSQC) spectra distinguish folded from unstructured protein and allow apo and ligand-bound complexes to be compared, mapping backbone amide groups by their \u00b9H and \u00b9\u2075N resonance frequencies; this requires \u00b9\u2075N-labelled protein and more material than the qualitative one-dimensional experiment. Key practical benefits are that pH, temperature and ligand binding can be varied and studied in a near-native solution state, and that NMR is nondestructive, so samples can be recovered for crystallization \u2014 hence the common pairing of NMR screening with crystallization in high-throughput structure-determination facilities. Figure 3 illustrates folded versus disordered spectra and the striking improvement seen for bacterial methionine aminopeptidase on inhibitor binding, whose complex crystals diffracted to 0.9 \u00c5.\n\n### 5.2. Electron microscopy\n\nConformational heterogeneity, especially in large multi-domain protein\u2013protein complexes, frequently blocks crystallization, and mutational variants and partner combinations must be screened. EM allows the sample to be visualized directly and its heterogeneity assessed. Full three-dimensional cryo-EM reconstruction is the best case but is time-consuming and needs substantial effort to find suitable data-collection parameters; by contrast, raw images of negatively stained particles are straightforward and need very little protein (typically less than 10 \u00b5g). This practice, long routine for EM practitioners screening samples for cryo-EM, has recently been adapted for crystallography. Caveats are that negative stains such as uranyl formate can introduce artifacts or disrupt the sample, and that the particle must be relatively large (greater than about 150 kDa) to be imaged \u2014 conveniently complementary to crystallography, which generally works well below 150 kDa. Simple image analysis yields two-dimensional class averages that quantify conformational classes and can guide construct screening. With cryo-EM data sets now pushing beyond 3 \u00c5, the method is increasingly useful for screening both conformational and compositional stability; examples cited include \u03b1-tubulin structural transitions and HBV capsid protein stability.\n\n### 5.3. Spectroscopic methods", "# About Marc C. Deller, D.Phil.\nAuthors: Deller, M.C.\nSource: marcdeller.com / chatMCD profile, 2026\n\n## Identity and contact\n\nMarc Christian Deller, D.Phil. (usually written Marc C. Deller) is a structural biologist, protein scientist, biophysicist and drug-discovery leader. He was born on 21 October 1972 in Bulwell, Nottingham, England, and is now based in Wilmington, Delaware, USA. His middle name is Christian.\n\nContact and links: email marc@marcdeller.com; phone (302) 358-6093; website https://marcdeller.com; LinkedIn https://linkedin.com/in/marccdeller; GitHub https://github.com/bellcheddar (handle \"bellcheddar\"); Google Scholar https://scholar.google.com/citations?user=o02oSJIAAAAJ; ORCID 0000-0001-8070-6502; Scopus Author ID 21741715300; OpenAlex author A5072187299; ResearchGate https://www.researchgate.net/profile/Marc-Deller. His web-app server is https://mdeller.com. Company sites: https://www.eloratherapeutics.com and https://www.deepcovalent.com. Useful pages on his site: publications at https://marcdeller.com/publications/, PDB structures at https://marcdeller.com/structures/, patents at https://marcdeller.com/patents/, projects at https://marcdeller.com/projects/, and his doctoral thesis PDF at https://marcdeller.com/wp-content/uploads/2025/08/marc_deller_thesis.pdf.\n\n## Professional identity in one paragraph\n\nMarc is a structural and computational biology leader with more than 25 years of experience spanning academia, structural genomics, big pharma and biotech. He is an expert in X-ray crystallography, cryo-EM, protein production, structure-based drug design (SBDD) and AI-driven therapeutics. Over his career he has contributed to more than 400 protein structures deposited in the Protein Data Bank, more than 60 publications, 7+ patents and 6+ FDA IND filings, and he has built two wet labs from scratch. He grew up in Nottingham and studied at the University of Leeds (BSc) and the University of Oxford (D.Phil.); his career then runs Yale \u2192 Pfizer \u2192 Scripps \u2192 Stanford \u2192 Incyte \u2192 Elora Therapeutics and DeepCovalent.\n\n## Current roles (2025\u2013present)\n\nMarc is Co-Founder and Chief Scientific Officer of Elora Therapeutics (from June 2025), a pre-seed, mission-driven biotech developing engineered PET-degrading enzymes (PETase- and MHETase-type enzymes) as a first-in-class systemic enzyme therapeutic to reduce the burden of microplastics and nanoplastics in the human body, with an initial focus on cardiometabolic disease. He leads scientific strategy, intellectual property (a 2025 provisional patent on therapeutic enzymatic compositions for degrading microplastics in human physiology), grant and funding applications, CRO partnerships and team building. The company is registered in Austin, Texas.\n\nHe is also a Scientific Technical Advisor to DeepCovalent Inc. (from October 2025), where he advises on physics-based computer-aided drug design (CADD) pipelines, small-molecule GPCR-targeting discovery platforms, in silico screening and cheminformatics workflows, and protein-science CRO oversight and experimental validation.\n\nAlongside these roles he describes himself as a \"vibe-coder\": a subject-matter expert who builds AI-assisted scientific software tools (see the separate document on his vibe-coding apps), fine-tunes local language models with Apple's MLX-LM on a 64 GB M1 Mac, and writes a science blog at marcdeller.com.\n\n## Career history\n\nResearch Fellow, Drug Discovery, Incyte Pharmaceuticals, Wilmington, Delaware (March 2022 \u2013 December 2024). Director/Senior-Director-equivalent role. Determined more than 100 challenging co-crystal structures for medicinal and computational chemistry teams and ran structure-based drug design for more than ten oncology and immunology programmes (ALK2, CDK2/4, KRAS G12D, FGFR2/3, JAK1/2, WRN, VRK1, USPs). Supported 6+ FDA IND filings and 7+ patents (mutant-CALR antibody, KRAS G12D, WRN, CDK2/4, JAK1/2, FGFR2/3). Mentored 5+ scientists, built and managed 10+ CRO and vendor partnerships, and championed automation, AI/ML and cloud computing, accelerating small-molecule pipelines by 67%.\n\nAssociate Director, Structural Biology, Incyte Pharmaceuticals (March 2018 \u2013 March 2022). Designed and built an $8M automated structural biology platform integrating crystallisation robotics, X-ray diffraction, a Glacios cryo-electron microscope, NMR, mass spectrometry, HDX and AI/cloud analytics. The platform delivered a four-fold gain in project efficiency, a 67% increase in throughput and 100+ protein\u2013drug co-crystal structures in under three years across ten programmes and five new targets. Managed a $1.5M budget and mentored 20+ scientists.\n\nDirector and Founder, Macromolecular Structure Knowledge Center (MSKC), Stanford ChEM-H, Stanford University (June 2015 \u2013 March 2018). Founded and scaled a profitable wet-lab service organisation managing a 70-project portfolio and a $1M annual budget; equipped 50+ researchers with advanced protein-science capabilities; generated about $60K in annual revenue; enabled six novel structures; supported three NIH SBIR submissions; and expanded SSRL/SLAC synchrotron access to 70 users from ten countries.\n\nTarget Manager, Joint Center for Structural Genomics (JCSG), The Scripps Research Institute, La Jolla, California (August 2006 \u2013 June 2015). Increased NIH structural-genomics throughput by 10%, solved 70+ novel protein structures, co-authored 40+ publications, designed the JCSG Top96 crystallisation screen, worked on automated crystal harvesting (the Universal Micromanipulation Robot, UMR), and contributed crystallography to HIV-1 Env and HCV E2 antibody-complex structures in Ian Wilson's laboratory.\n\nPrincipal Scientist, Small-Molecule Drug Discovery / Structural Biology, Pfizer. Structural biology contributions to small-molecule drug-discovery programmes.\n\nPostdoctoral Research Fellow, Yale University School of Medicine. JAK/STAT protein structure research.\n\n## Education", "# Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\nAuthors: Ye, Q.; Shvartsbart, A.; Li, Z.; Gan, P.; Policarpo, R.L.; Qi, C.; Roach, J.J.; Zhu, W.; McCammant, M.S.; Hu, B.; Li, G.; Yin, H.; Carlsen, P.; Hoang, G.; Zhao, L.; Susick, R.; Zhang, F.; Lai, C.-T.; Allali Hassani, A.; Epling, L.B.; Gallion, A.; Kurzeja-Lipinski, K.; Gallagher, K.; Roman, V.; Farren, M.R.; Kong, W.; Deller, M.C.; Zhang, G.; Covington, M.; Diamond, S.; Kim, S.; Yao, W.; Sokolsky, A.; Wang, X.\nSource: Journal of Medicinal Chemistry, 2025\n\n**Authors:** Qinda Ye, Artem Shvartsbart, Zhenwu Li, Pei Gan, Rocco L. Policarpo, Chao Qi, Jeremy J. Roach, Wenyu Zhu, Matthew S. McCammant, Bin Hu, Gencheng Li, Haolin Yin, Peter Carlsen, Gia Hoang, Le Zhao, Robert Susick, Fenglei Zhang, Cheng-Tsung Lai, Abdellah Allali Hassani, Leslie B. Epling, Alexandra Gallion, Kerri Kurzeja-Lipinski, Karen Gallagher, Valerie Roman, Matthew R. Farren, Weixi Kong, Marc C. Deller, Guofeng Zhang, Maryanne Covington, Sharon Diamond, Sunkyu Kim, Wenqing Yao, Alexander Sokolsky (corresponding author) and Xiaozhao Wang. **Affiliations (condensed):** Department of Discovery Chemistry, Department of Computational Chemistry, Department of Applied Technology, Department of Preclinical Pharmacology and Department of DMB, Incyte Research Institute, Incyte Corporation, Wilmington, Delaware 19803, United States. **Published:** J. Med. Chem., volume 68, pages 1924\u20131939, 2025. Received 4 November 2024; revised 16 December 2024; accepted 30 December 2024; published 8 January 2025.\n\n## Abstract\n\nThe inhibition of mutant KRAS proteins has emerged as a promising approach for treating KRAS-driven cancers, as evidenced by the clinical success of KRAS G12C inhibitors. KRAS G12D, the most common mutant, promises significant expansion of the addressable patient population; however, the reduced nucleophilicity of aspartate compared to cysteine poses significant challenges in balancing sufficient potency with ADME properties to support oral exposure. Herein, we describe the discovery of KRAS G12D inhibitor 23 (INCB159020), which achieves oral exposure in nonhuman primate (NHP). Starting from a weakly potent hit, structure-based drug design was utilized to drive significant potency. Focus on molecular rigidity and balanced polarity then allowed for successful optimization of properties required for oral exposure.\n\n## Introduction\n\nKirsten rat sarcoma virus (KRAS) is a GTPase that acts as a master regulator of the RAS/MAPK and PI3K/AKT pathways governing cell growth, differentiation, proliferation and survival. Upon binding GTP, KRAS enters an activated state, signaling through Raf and PI3K to stimulate the production of growth factors. As GTP is hydrolyzed to GDP, KRAS enters a GDP-bound off state. Active/inactive state cycling is achieved through the action of Guanine Nucleotide Exchange Factors (GEFs) such as SOS1. As a consequence of its central status, KRAS is frequently misregulated in a variety of human cancers, including nonsmall cell lung (NSCLC), pancreatic and colorectal cancer. In fact, it is estimated that 20% of human tumors are driven by KRAS mutations. Mutations can occur at a number of positions, but are most frequently localized to Gly12, Gly13 and Gln61. These mutations decrease the intrinsic hydrolysis rate of GTP, leading to buildup of the activated form, constitutive pathway activation and uncontrolled cellular proliferation. When mutated, KRAS has been shown to act as a key driver of proliferation. This fact, combined with the high prevalence of KRAS mutations across the deadliest of human cancers, makes KRAS a critical target for drug discovery.\n\nHowever, KRAS has proven to present a formidable challenge to discovery efforts, due to the exceptional binding affinity to GTP preventing direct inhibition of the active site. A significant breakthrough was made in 2013, when Shokat and coworkers described the discovery of a new pocket adjacent to the nucleoside binding region, which could be accessed by covalently modifying the cysteine residue present in the KRAS G12C mutant. Further optimization by a number of groups led to the identification of numerous potent and efficacious clinical KRAS G12C inhibitors, culminating with the FDA approval of sotorasib and adagrasib for the treatment of KRAS G12C-mutated NSCLC in 2021 and 2022 respectively.\n\nThe success of KRAS G12C inhibitors led to increased interest in inhibiting other mutant forms of KRAS. Among these, the G12D mutation is the most prevalent mutation observed in KRAS and represents a significant additional patient populations, in particular in pancreatic (37% of tumors) and colorectal (12% of tumors) indications. However, the KRAS G12D mutant presents a unique challenge, owing to the attenuated nucleophilicity of the aspartic acid, rendering a covalent approach uncertain. While covalent interactions with aspartic acid are possible, the requirement of increased electrophilicity raises concerns with respect to selectivity and safety. An alternative approach involves the removal of the warhead in pursuit of a reversible inhibitor. Researchers at Mirati Therapeutics have shown the successful application of the latter approach, leading to the identification of MRTX1133 (Figure 1), a highly potent reversible KRAS G12D inhibitor that was shown to be efficacious in a variety of tumor models. However, due to the low oral bioavailability of MRTX1133, i.p. dosing was used to establish efficacy and significant formulation work was required to achieve clinically relevant oral exposure. A number of other companies have since reported efforts toward KRAS G12D inhibitors or inhibitors that target a number of RAS isoforms. Herein we describe the discovery of a highly structurally distinct series of selective KRAS G12D inhibitors that show promising oral exposure in multiple preclinical species.\n\n*Figure 1.* MRTX1133 and early patent hit 1.\n\n## Results and Discussion\n\nBased on the clinical success of KRAS G12C inhibitors and the significant unmet need in tumors driven by the G12D mutation, we initiated a program with the goal of discovering a potent and selective orally bioavailable small molecule inhibitor of KRAS G12D.", "# ChatRDKit \u2013 Talk to RDKit for Chemoinformatics in plain language\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-14 on marcdeller.com (categories: vibe-code).*\n\n> **Talk to RDKit in plain English. Cheminformatics without the boilerplate.**\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/chatRDKit](https://github.com/bellcheddar/chatRDKit) |\n\n---\n\nA natural-language chat frontend for **RDKit** (the open-source\ncheminformatics toolkit), powered by a local **Ollama** model (default:\n`qwen3.6:27b`).\n\nWhy it matters: RDKit is the backbone of open-source cheminformatics, but its API is broad and unforgiving, and every routine task (computing descriptors, drawing a molecule, running a similarity search) means recalling the exact module, class, and function names. chatRDKit removes that friction by letting you describe what you want in natural language and translating it into correct RDKit code on the fly, all running locally against an Ollama model so no structures ever leave your machine. It is useful for medicinal chemists, structural biologists, and anyone who wants RDKit\u2019s power at the speed of thought: rapid property triage, on-the-fly depiction, Lipinski checks, and fingerprint comparisons, without breaking flow to look up syntax.\n\nType plain English at the prompt:\n\n```\nchatRDKit> Load aspirin and tell me its molecular weight, LogP and TPSA\n[chatRDKit] >>> aspirin = Chem.MolFromSmiles(\"CC(=O)Oc1ccccc1C(=O)O\")\n[chatRDKit] >>> print(\"MW:\", Descriptors.MolWt(aspirin))\n[chatRDKit] >>> print(\"LogP:\", Descriptors.MolLogP(aspirin))\n[chatRDKit] >>> print(\"TPSA:\", Descriptors.TPSA(aspirin))\nMW: 180.15899999999996\nLogP: 1.3101\nTPSA: 63.60000000000001\n```\n\nOr type a **SMILES string** directly to load a molecule, or **Python/RDKit\ncode** to run it as-is. Everything runs locally in a persistent Python\nsession \u2014 no data leaves your machine.\n\n---\n\n## 1. Prerequisites\n\n### 1.1 Ollama + model\n\n1. Install Ollama:\n2. **macOS**: download the app from <https://ollama.com/download>, or\n   `brew install ollama`\n3. **Linux**: `curl -fsSL https://ollama.com/install.sh | sh`\n4. **Windows**: download the installer from <https://ollama.com/download>\n5. Make sure the server is running. The desktop app starts it\n   automatically; otherwise:\n\n`sh\n ollama serve`\n\n1. Pull a Qwen3.6 model:\n\n`sh\n ollama pull qwen3.6:27b`\n\nThis is the default chatRDKit uses (a ~27.8B-parameter model, ~17 GB on\ndisk) and gives the best translation quality. It needs roughly 20 GB+ of\nfree RAM/VRAM to run comfortably. If your machine can\u2019t handle that:\n\n`sh\n # Much smaller/faster, good for trying chatRDKit out, lower translation quality:\n ollama pull qwen3:0.6b`\n\nThen point chatRDKit at whichever tag you pulled (see\n[Configuration](#3-configuration)) \u2014 just make sure the `model` setting\nmatches a tag from `ollama list` exactly.\n\n1. Sanity check the server is reachable and the model responds:\n\n`sh\n curl http://localhost:11434/api/chat -d '{\n \"model\": \"qwen3.6:27b\",\n \"messages\": [{\"role\": \"user\", \"content\": \"say hi\"}],\n \"stream\": false\n }'`\n\nYou should get back a JSON response with a `message.content` field.\n\n### 1.2 Python + RDKit\n\nchatRDKit needs Python 3.10+ and the latest RDKit. A virtual environment is\nrecommended:\n\n```\ncd chatRDKit\npython3 -m venv .venv\nsource .venv/bin/activate          # Windows: .venv\\Scripts\\activate\npip install -r requirements.txt\n```\n\n`requirements.txt` installs:\n\u2013 `rdkit` \u2014 the latest release from PyPI (chemistry engine + drawing)\n\u2013 `reportlab` \u2014 used for PDF report output (optional; PNG/HTML work\nwithout it)\n\n---\n\n## 2. Running chatRDKit\n\n```\nsource .venv/bin/activate\npython chatrdkit.py\n```\n\nOn launch you\u2019ll see a welcome banner with 5 numbered example prompts:\n\n```\n========================================================================\n  chatRDKit -- chat with RDKit (2026.03.3) using natural language\n========================================================================\n\nTry one of these to get started (type the number, or your own request):\n  1. Enter a SMILES string to load a molecule (e.g. CC(=O)Oc1ccccc1C(=O)O for aspirin)\n  2. What is the molecular weight, LogP and TPSA of caffeine?\n  3. Draw ibuprofen and save it as an image\n  4. Generate a 3D structure of ethanol and optimise its geometry\n  5. Compare the similarity between aspirin and paracetamol using Morgan fingerprints\n\nType /help for commands, /exit to quit.\n\nchatRDKit>\n```\n\nType a number to run that example. Picking **1** prompts you to enter a\nSMILES string, loads it as a molecule (`mol`, `mol2`, \u2026 in the session),\nand then shows a **structured follow-up menu** of suggested next steps for\nthat molecule (compute properties, draw it, generate 3D, find rings, check\nLipinski\u2019s Rule of Five) \u2014 type a number to run one, or just type your own\nrequest.\n\n### Three kinds of input\n\n* **A SMILES string** (e.g. `CCO`, `c1ccccc1`, `CC(=O)Oc1ccccc1C(=O)O`) is\n  loaded directly as a new molecule (`mol`, `mol2`, `mol3`, \u2026) \u2014 no LLM\n  call needed.\n* **Python/RDKit code** (e.g. `Descriptors.MolWt(mol)`, `mol.GetNumAtoms()`,\n  `for atom in mol.GetAtoms(): print(atom.GetSymbol())`) is executed\n  directly in the session. A bare expression\u2019s value is printed, REPL-style.\n* **Plain English** (anything else) is sent to Ollama, translated into RDKit\n  code using the current session state as context, echoed as\n  `[chatRDKit] >>> ...`, and executed.\n\n### Multi-turn context\n\nchatRDKit keeps a rolling history of your requests and the code it\ngenerated, plus a live snapshot of every molecule currently in the session\n(so \u201cit\u201d, \u201cthe molecule\u201d, or a compound name you already loaded resolve\ncorrectly). Clear the history with:\n\n```\nchatRDKit> /reset\n```\n\n### Rendering output\n\nWhenever a request involves drawing/visualising a molecule or generating a\nreport, chatRDKit calls a built-in `render(mol, name=\"...\")` helper, which\nprompts you for the output format:", "# PDB structures deposited by Marc C. Deller\nAuthors: Deller, M.C. and co-depositors\nSource: RCSB Protein Data Bank, 2026\n\n- **10PI** \u2014 JAK1 kinase (JH1 domain) in complex with povorcitinib. Method: X-RAY DIFFRACTION; resolution 1.54 \u00c5; space group C 1 2 1; 594 polymer residues, 70.7 kDa; released 2026-06-10. Deposition authors: Epling, L.B., Fenalti, G.. Keywords: TRANSFERASE. Primary citation: Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases. (J.Med.Chem., 2026), DOI 10.1021/acs.jmedchem.5c03753.\n- **10PJ** \u2014 JAK2 kinase (JH1 domain) in complex with povorcitinib. Method: X-RAY DIFFRACTION; resolution 1.59 \u00c5; space group I 1 2 1; 632 polymer residues, 75.66 kDa; released 2026-06-10. Deposition authors: Epling, L.B., Fenalti, G.. Keywords: TRANSFERASE. Primary citation: Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases. (J.Med.Chem., 2026), DOI 10.1021/acs.jmedchem.5c03753.\n- **9E5D** \u2014 Discovery of an Orally Biovailable KRAS G12D Inhibitor. Method: X-RAY DIFFRACTION; resolution 1.36 \u00c5; space group P 21 21 21; 170 polymer residues, 20.5 kDa; released 2025-01-22. Deposition authors: Deller, M.C., Epling, L.B.. Keywords: HYDROLASE/HYDROLASE INHIBITOR. Bound ligands: GDP, MG. Primary citation: Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor. (J.Med.Chem., 2025), DOI 10.1021/acs.jmedchem.4c02662.\n- **9E5F** \u2014 Discovery of an Orally Bioavailable KRAS G12D Inhibitor. Method: X-RAY DIFFRACTION; resolution 1.35 \u00c5; space group P 21 21 21; 170 polymer residues, 20.4 kDa; released 2025-01-22. Deposition authors: Deller, M.C., Epling, L.B.. Keywords: HYDROLASE/HYDROLASE INHIBITOR. Bound ligands: GDP, MG. Primary citation: Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor. (J.Med.Chem., 2025), DOI 10.1021/acs.jmedchem.4c02662.\n- **8UV0** \u2014 Discovery of (4-Pyrazolyl)-2-Aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2. Method: X-RAY DIFFRACTION; resolution 1.55 \u00c5; space group P 21 21 21; 297 polymer residues, 34.33 kDa; released 2024-02-14. Deposition authors: Deller, M.C., Epling, L.B.. Keywords: TRANSFERASE/INHIBITOR. Primary citation: Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2. (J.Med.Chem., 2024), DOI 10.1021/acs.jmedchem.3c02287.\n- **8E1X** \u2014 FGFR2 kinase domain in complex with a Pyrazolo[1,5-a]pyrimidine analog (Compound 29). Method: X-RAY DIFFRACTION; resolution 2.68 \u00c5; space group P 21 21 21; 660 polymer residues, 76.54 kDa; released 2022-11-23. Deposition authors: Lei, H.-T., Epling, L.B., Deller, M.C.. Keywords: STRUCTURAL PROTEIN. Primary citation: Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3. (J.Med.Chem., 2022), DOI 10.1021/acs.jmedchem.2c01366.\n- **6OFO** \u2014 Crystal structure of split green fluorescent protein (GFP); s10 circular permutant (194-195). Method: X-RAY DIFFRACTION; resolution 2.603 \u00c5; space group P 1 21 1; 508 polymer residues, 56.43 kDa; released 2019-07-10. Deposition authors: Lin, C.-Y., Romei, M.G., Deller, M.C., Doukov, T.I., Boxer, S.G.. Keywords: FLUORESCENT PROTEIN. Primary citation: Unified Model for Photophysical and Electro-Optical Properties of Green Fluorescent Proteins. (J.Am.Chem.Soc., 2019), DOI 10.1021/jacs.9b07152.\n- **6C9U** \u2014 Crystal structure of [KS3][AT3] didomain from module 3 of 6-deoxyerthronolide B synthase in complex with antibody fragment (Fab). Method: X-RAY DIFFRACTION; resolution 2.09 \u00c5; space group C 1 2 1; 1426 polymer residues, 152.05 kDa; released 2018-05-23. Deposition authors: Deis, L.N., Li, X., Mathews, I.I., Khosla, C.. Keywords: TRANSFERASE/immune system. Bound ligands: K, NA. Primary citation: Structure-Function Analysis of the Extended Conformation of a Polyketide Synthase Module. (J. Am. Chem. Soc., 2018), DOI 10.1021/jacs.8b02100.\n- **5UFV** \u2014 Crystal Structure of a Cellulose-active Polysaccharide Monooxygenase from M. thermophila (MtPMO3*). Method: X-RAY DIFFRACTION; resolution 2.45 \u00c5; space group P 1 21 1; 1428 polymer residues, 152.75 kDa; released 2017-03-15. Deposition authors: Span, E.A., Deller, M.C., Marletta, M.A.. Keywords: Hydrolase/OXIDOREDUCTASE. Bound ligands: CU. Primary citation: The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase. (ACS Chem. Biol., 2017), DOI 10.1021/acschembio.7b00016.\n- **5SY8** \u2014 Crystal structure of the complex of 10E8 Fab light chain mutant1 and T117v2 HIV-1 MPER scaffold. Method: X-RAY DIFFRACTION; resolution 1.62 \u00c5; space group P 1; 613 polymer residues, 67.2 kDa; released 2017-03-08. Deposition authors: Irimia, A., Wilson, I.A.. Keywords: IMMUNE SYSTEM. Primary citation: Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design. (PLoS Pathog., 2017), DOI 10.1371/journal.ppat.1006212.\n- **5T29** \u2014 Crystal structure of 10E8 Fab light chain mutant3, against the MPER region of the HIV-1 Env, in complex with the MPER epitope scaffold T117v2. Method: X-RAY DIFFRACTION; resolution 2.03 \u00c5; space group C 1 2 1; 614 polymer residues, 67.23 kDa; released 2017-03-08. Deposition authors: Irimia, A., Wilson, I.A.. Keywords: IMMUNE SYSTEM. Primary citation: Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design. (PLoS Pathog., 2017), DOI 10.1371/journal.ppat.1006212.\n- **5T5B** \u2014 CRYSTAL STRUCTURE OF THE COMPLEX OF 10E8 FAB LIGHT CHAIN MUTANT5 AND T117V2 HIV-1 MPER SCAFFOLD. Method: X-RAY DIFFRACTION; resolution 2.07 \u00c5; space group P 1 21 1; 1228 polymer residues, 134.36 kDa; released 2017-03-08. Deposition authors: Irimia, A., Wilson, I.A.. Keywords: IMMUNE SYSTEM. Primary citation: Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design. (PLoS Pathog., 2017), DOI 10.1371/journal.ppat.1006212.\n- **5T6L** \u2014 Crystal structure of 10E8 Fab in complex with the MPER epitope scaffold T117v2. Method: X-RAY DIFFRACTION; resolution 2.1 \u00c5; space group P 1 21 1; 1228 polymer residues, 135.96 kDa; released 2017-03-08. Deposition authors: Irimia, A., Wilson, I.A.. Keywords: IMMUNE SYSTEM. Primary citation: Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design. (PLoS Pathog., 2017), DOI 10.1371/journal.ppat.1006212.\n- **5T80** \u2014 Crystal structure of 10E8 Fab in complex with the MPER epitope scaffold T117v2 and phosphatidic acid (06:0 PA). Method: X-RAY DIFFRACTION; resolution 2.62 \u00c5; space group P 21 21 21; 614 polymer residues, 67.45 kDa; released 2017-03-08. Deposition authors: Irimia, A., Wilson, I.A.. Keywords: IMMUNE SYSTEM. Primary citation: Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design. (PLoS Pathog., 2017), DOI 10.1371/journal.ppat.1006212.\n- **5T85** \u2014 Crystal structure of 10E8 Fab in complex with the MPER epitope scaffold T117v2 and phosphatidylglycerol (06:0 PG). Method: X-RAY DIFFRACTION; resolution 2.373 \u00c5; space group P 21 21 21; 614 polymer residues, 67.82 kDa; released 2017-03-08. Deposition authors: Irimia, A., Wilson, I.A.. Keywords: IMMUNE SYSTEM. Primary citation: Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design. (PLoS Pathog., 2017), DOI 10.1371/journal.ppat.1006212.\n- **5TFW** \u2014 Crystal structure of 10E8 Fab light chain mutant2 against the MPER region of the HIV-1 Env, in complex with T117v2 epitope scaffold. Method: X-RAY DIFFRACTION; resolution 2.168 \u00c5; space group C 1 2 1; 614 polymer residues, 67.12 kDa; released 2017-03-08. Deposition authors: Irimia, A., Wilson, I.A.. Keywords: IMMUNE SYSTEM. Primary citation: Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design. (PLoS Pathog., 2017), DOI 10.1371/journal.ppat.1006212.\n- **5CD3** \u2014 Structure of immature VRC01-class antibody DRVIA7. Method: X-RAY DIFFRACTION; resolution 2.9 \u00c5; space group P 1; 1720 polymer residues, 189.02 kDa; released 2016-04-06. Deposition authors: Kong, L., Wilson, I.A.. Keywords: IMMUNE SYSTEM. Primary citation: Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor. (Immunity, 2016), DOI 10.1016/j.immuni.2016.03.006.\n- **5CD5** \u2014 Crystal structure of an immature VRC01-class antibody DRVIA7 from a Chinese donor bound to clade A/E HIV-1 gp120 core. Method: X-RAY DIFFRACTION; resolution 3.396 \u00c5; space group P 41 21 2; 783 polymer residues, 89.98 kDa; released 2016-04-06. Deposition authors: Kong, L., Wilson, I.A.. Keywords: IMMUNE SYSTEM. Bound ligands: NAG. Primary citation: Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor. (Immunity, 2016), DOI 10.1016/j.immuni.2016.03.006.\n- **5C7K** \u2014 Crystal structure BG505 SOSIP gp140 HIV-1 Env trimer bound to broadly neutralizing antibodies PGT128 and 8ANC195. Method: X-RAY DIFFRACTION; resolution 4.6019 \u00c5; space group I 2 3; 1543 polymer residues, 183.37 kDa; released 2015-10-07. Deposition authors: Kong, L., Stanfield, R.L., Wilson, I.A.. Keywords: IMMUNE SYSTEM. Bound ligands: NAG. Primary citation: Complete epitopes for vaccine design derived from a crystal structure of the broadly neutralizing antibodies PGT128 and 8ANC195 in complex with an HIV-1 Env trimer. (Acta Crystallogr.,Sect.D, 2015), DOI 10.1107/S1399004715013917.\n- **4RBO** \u2014 Crystal structure of a Nanog homeobox (NANOG) from Homo sapiens at 3.30 A resolution. Method: X-RAY DIFFRACTION; resolution 3.3 \u00c5; space group P 65 2 2; 188 polymer residues, 31.87 kDa; released 2014-10-01. Deposition authors: Joint Center for Structural Genomics (JCSG), Partnership for Stem Cell Biology (STEMCELL). Keywords: DNA BINDING PROTEIN, TRANSCRIPTION/DNA. Primary citation: Structure-based discovery of NANOG variant with enhanced properties to promote self-renewal and reprogramming of pluripotent stem cells. (Proc.Natl.Acad.Sci.USA, 2015), DOI 10.1073/pnas.1502855112.\n- **4NCO** \u2014 Crystal Structure of the BG505 SOSIP gp140 HIV-1 Env trimer in Complex with the Broadly Neutralizing Fab PGT122. Method: X-RAY DIFFRACTION; resolution 4.7 \u00c5; space group C 1 2 1; 2997 polymer residues, 350.93 kDa; released 2013-11-13. Deposition authors: Julien, J.-P., Stanfield, R.L., Lyumkis, D., Ward, A.B., Wilson, I.A.. Keywords: VIRAL PROTEIN/IMMUNE SYSTEM. Bound ligands: NAG. Primary citation: Crystal structure of a soluble cleaved HIV-1 envelope trimer. (Science, 2013), DOI 10.1126/science.1245625.\n- **4JM2** \u2014 Crystal Structure of PGT 135 Fab in Complex with gp120 Core Protein from HIV-1 Strain JR-FL Bound to CD4 and 17b Fab. Method: X-RAY DIFFRACTION; resolution 3.1 \u00c5; space group C 1 2 1; 1399 polymer residues, 158.95 kDa; released 2013-05-29. Deposition authors: Kong, L., Wilson, I.A.. Keywords: IMMUNE SYSTEM/VIRAL PROTEIN. Bound ligands: NAG. Primary citation: Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120. (Nat.Struct.Mol.Biol., 2013), DOI 10.1038/nsmb.2594.\n- **4JM4** \u2014 Crystal Structure of PGT 135 Fab. Method: X-RAY DIFFRACTION; resolution 1.751 \u00c5; space group P 21 21 2; 450 polymer residues, 49.37 kDa; released 2013-05-29. Deposition authors: Kong, L., Wilson, I.A.. Keywords: IMMUNE SYSTEM. Primary citation: Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120. (Nat.Struct.Mol.Biol., 2013), DOI 10.1038/nsmb.2594.\n- **4G6A** \u2014 Structure of the Hepatitis C virus envelope glycoprotein E2 antigenic region 412-423 bound to the broadly neutralizing antibody AP33. Method: X-RAY DIFFRACTION; resolution 2.501 \u00c5; space group P 1 21 1; 910 polymer residues, 98.74 kDa; released 2012-09-26. Deposition authors: Kong, L., Wilson, I.A., Law, M.. Keywords: IMMUNE SYSTEM. Primary citation: Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33. (J.Virol., 2012), DOI 10.1128/JVI.01939-12.\n- **3O0F** \u2014 Crystal structure of a putative metal-dependent phosphoesterase (BAD_1165) from bifidobacterium adolescentis atcc 15703 at 1.94 A resolution. Method: X-RAY DIFFRACTION; resolution 1.94 \u00c5; space group P 65; 301 polymer residues, 33.61 kDa; released 2010-08-25. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: HYDROLASE. Bound ligands: FE, PO4, ZN. Primary citation: Crystal structure of a metal-dependent phosphoesterase (YP_910028.1) from Bifidobacterium adolescentis: Computational prediction and experimental validation of phosphoesterase activity. (Proteins, 2011), DOI 10.1002/prot.23035.\n- **3NL9** \u2014 Crystal structure of a putative NTP pyrophosphohydrolase (Exig_1061) from EXIGUOBACTERIUM SP. 255-15 at 1.78 A resolution. Method: X-RAY DIFFRACTION; resolution 1.78 \u00c5; space group C 1 2 1; 171 polymer residues, 19.52 kDa; released 2010-07-21. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: HYDROLASE. Primary citation: Structure of a putative NTP pyrophosphohydrolase: YP_001813558.1 from Exiguobacterium sibiricum 255-15. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309110025534.\n- **3M82** \u2014 Crystal structure of Acetyl xylan esterase (TM0077) from THERMOTOGA MARITIMA at 2.40 A resolution (PMSF inhibitor complex structure). Method: X-RAY DIFFRACTION; resolution 2.4 \u00c5; space group P 21 21 21; 2022 polymer residues, 233.77 kDa; released 2010-05-26. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: HYDROLASE. Bound ligands: CA. Primary citation: Functional and structural characterization of a thermostable acetyl esterase from Thermotoga maritima. (Proteins, 2012), DOI 10.1002/prot.24041.\n- **3M83** \u2014 Crystal structure of Acetyl xylan esterase (TM0077) from THERMOTOGA MARITIMA at 2.12 A resolution (paraoxon inhibitor complex structure). Method: X-RAY DIFFRACTION; resolution 2.12 \u00c5; space group P 21 21 21; 2022 polymer residues, 234.04 kDa; released 2010-05-12. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: HYDROLASE. Bound ligands: CA. Primary citation: Functional and structural characterization of a thermostable acetyl esterase from Thermotoga maritima. (Proteins, 2012), DOI 10.1002/prot.24041.\n- **3M81** \u2014 Crystal structure of Acetyl xylan esterase (TM0077) from THERMOTOGA MARITIMA at 2.50 A resolution (native apo structure). Method: X-RAY DIFFRACTION; resolution 2.5 \u00c5; space group P 21 21 21; 2022 polymer residues, 233.12 kDa; released 2010-04-21. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: HYDROLASE. Bound ligands: CA. Primary citation: Functional and structural characterization of a thermostable acetyl esterase from Thermotoga maritima. (Proteins, 2012), DOI 10.1002/prot.24041.\n- **3L4B** \u2014 Crystal Structure of an Octomeric Two-Subunit TrkA K+ Channel Ring Gating Assembly, TM1088A:TM1088B, from Thermotoga maritima. Method: X-RAY DIFFRACTION; resolution 3.45 \u00c5; space group P 32 2 1; 1444 polymer residues, 161.86 kDa; released 2010-02-16. Deposition authors: Deller, M.C., Johnson, H.A., Miller, M., Spraggon, G., Wilson, I.A., Lesley, S.A., Joint Center for Structural Genomics (JCSG). Keywords: TRANSPORT PROTEIN. Primary citation: Crystal Structure of an Octomeric Two-Subunit TrkA K+ Channel Ring Gating Assembly, TM1088A:TM1088B, from Thermotoga maritima (To be Published, None), DOI None.\n- **3L5O** \u2014 Crystal structure of protein with unknown function from DUF364 family (ZP_00559375.1) from Desulfitobacterium hafniense DCB-2 at 2.01 A resolution. Method: X-RAY DIFFRACTION; resolution 2.01 \u00c5; space group P 61; 540 polymer residues, 61.65 kDa; released 2010-02-02. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: Adenosyl binding protein. Primary citation: Structure of the first representative of Pfam family PF04016 (DUF364) reveals enolase and Rossmann-like folds that combine to form a unique active site with a possible role in heavy-metal chelation. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309110007517.\n- **3K5J** \u2014 Crystal structure of Putative SUFU (suppressor of fused protein) homolog (YP_208451.1) from Neisseria gonorrhoeae FA 1090 at 1.40 A resolution. Method: X-RAY DIFFRACTION; resolution 1.4 \u00c5; space group P 32 2 1; 183 polymer residues, 22.16 kDa; released 2010-01-26. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: PROTEIN BINDING. Primary citation: The crystal structure of a bacterial Sufu-like protein defines a novel group of bacterial proteins that are similar to the N-terminal domain of human Sufu. (Protein Sci., 2010), DOI 10.1002/pro.497.\n- **3KK7** \u2014 Crystal structure of Putative cell invasion protein with MAC/Perforin domain (NP_812351.1) from BACTERIODES THETAIOTAOMICRON VPI-5482 at 2.46 A resolution. Method: X-RAY DIFFRACTION; resolution 2.46 \u00c5; space group P 21 21 21; 1082 polymer residues, 124.35 kDa; released 2009-11-24. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: CELL INVASION. Primary citation: Structure of a membrane-attack complex/perforin (MACPF) family protein from the human gut symbiont Bacteroides thetaiotaomicron. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309110023055.\n- **3JXO** \u2014 Crystal Structure of an Octomeric Two-Subunit TrkA K+ Channel Ring Gating Assembly, TM1088A:TM1088B, from Thermotoga maritima. Method: X-RAY DIFFRACTION; resolution 1.55 \u00c5; space group C 2 2 21; 172 polymer residues, 18.72 kDa; released 2009-10-06. Deposition authors: Deller, M.C., Johnson, H.A., Miller, M., Spraggon, G., Wilson, I.A., Lesley, S.A., Joint Center for Structural Genomics (JCSG). Keywords: TRANSPORT PROTEIN. Primary citation: Crystal Structure of an Octomeric Two-Subunit TrkA K+ Channel Ring Gating Assembly, TM1088A:TM1088B, from Thermotoga maritima (To be Published, None), DOI None.\n- **3IRB** \u2014 Crystal structure of protein with unknown function from DUF35 family (13815350) from SULFOLOBUS SOLFATARICUS at 1.80 A resolution. Method: X-RAY DIFFRACTION; resolution 1.8 \u00c5; space group P 41 2 2; 290 polymer residues, 34.1 kDa; released 2009-09-01. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: acyl-CoA binding protein. Bound ligands: ZN. Primary citation: The structure of SSO2064, the first representative of Pfam family PF01796, reveals a novel two-domain zinc-ribbon OB-fold architecture with a potential acyl-CoA-binding role. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309110002514.\n- **3HSA** \u2014 Crystal structure of pleckstrin homology domain (YP_926556.1) from SHEWANELLA AMAZONENSIS SB2B at 1.99 A resolution. Method: X-RAY DIFFRACTION; resolution 1.99 \u00c5; space group P 21 21 21; 630 polymer residues, 72.83 kDa; released 2009-06-23. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: structural genomics, unknown function. Primary citation: Bacterial pleckstrin homology domains: a prokaryotic origin for the PH domain. (J.Mol.Biol., 2010), DOI 10.1016/j.jmb.2009.11.006.\n- **3HBZ** \u2014 Crystal structure of a putative glycoside hydrolase (bt_2081) from bacteroides thetaiotaomicron vpi-5482 at 2.05 A resolution. Method: X-RAY DIFFRACTION; resolution 2.05 \u00c5; space group P 32 2 1; 342 polymer residues, 41.43 kDa; released 2009-05-26. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: Ca-BINDING PROTEIN. Bound ligands: CA, CAC, NA. Primary citation: Structure of Bacteroides thetaiotaomicron BT2081 at 2.05\u2005A resolution: the first structural representative of a new protein family that may play a role in carbohydrate metabolism. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309110028228.\n- **3H41** \u2014 CRYSTAL STRUCTURE OF A NLPC/P60 FAMILY PROTEIN (BCE_2878) FROM BACILLUS CEREUS ATCC 10987 AT 1.79 A RESOLUTION. Method: X-RAY DIFFRACTION; resolution 1.79 \u00c5; space group C 1 2 1; 311 polymer residues, 36.67 kDa; released 2009-05-05. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: HYDROLASE. Bound ligands: ALA, DGL. Primary citation: Structure of the gamma-D-glutamyl-L-diamino acid endopeptidase YkfC from Bacillus cereus in complex with L-Ala-[gamma]-D-Glu: insights into substrate recognition by NlpC/P60 cysteine peptidases (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309110021214.\n- **3H50** \u2014 CRYSTAL STRUCTURE OF A TETRACENOMYCIN POLYKETIDE SYNTHESIS PROTEIN (TCMJ) FROM XANTHOMONAS CAMPESTRIS PV. CAMPESTRIS AT 1.60 A RESOLUTION. Method: X-RAY DIFFRACTION; resolution 1.6 \u00c5; space group C 1 2 1; 114 polymer residues, 12.71 kDa; released 2009-05-05. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: BIOSYNTHETIC PROTEIN. Bound ligands: ACT, ZN. Primary citation: Conformational changes associated with the binding of zinc acetate at the putative active site of XcTcmJ, a cupin from Xanthomonas campestris pv. campestris. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309109021988.\n- **3H0N** \u2014 Crystal structure of a duf1470 family protein (jann_2411) from jannaschia sp. ccs1 at 1.45 A resolution. Method: X-RAY DIFFRACTION; resolution 1.45 \u00c5; space group C 1 2 1; 188 polymer residues, 21.38 kDa; released 2009-04-21. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: METAL BINDING PROTEIN. Bound ligands: ACT, NA, NI, ZN. Primary citation: The structure of Jann_2411 (DUF1470) from Jannaschia sp. at 1.45\u2005  resolution reveals a new fold (the ABATE domain) and suggests its possible role as a transcription regulator. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309109025196.\n- **3GF8** \u2014 Crystal structure of putative polysaccharide binding proteins (DUF1812) (NP_809975.1) from BACTEROIDES THETAIOTAOMICRON VPI-5482 at 2.20 A resolution. Method: X-RAY DIFFRACTION; resolution 2.2 \u00c5; space group P 41 21 2; 296 polymer residues, 34.84 kDa; released 2009-03-24. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: Carbohydrate Binding Protein. Primary citation: A conserved fold for fimbrial components revealed by the crystal structure of a putative fimbrial assembly protein (BT1062) from Bacteroides thetaiotaomicron at 2.2 A resolution (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309110006548.\n- **3F1Z** \u2014 Crystal structure of putative nucleic acid-binding lipoprotein (YP_001337197.1) from Klebsiella pneumoniae subsp. pneumoniae MGH 78578 at 2.46 A resolution. Method: X-RAY DIFFRACTION; resolution 2.46 \u00c5; space group P 21 21 21; 1330 polymer residues, 146.22 kDa; released 2008-11-18. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: DNA BINDING PROTEIN. Primary citation: The structure of KPN03535 (gi|152972051), a novel putative lipoprotein from Klebsiella pneumoniae, reveals an OB-fold. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309109018168.\n- **3EQX** \u2014 CRYSTAL STRUCTURE OF A FIC FAMILY PROTEIN (SO_4266) FROM SHEWANELLA ONEIDENSIS AT 1.6 A RESOLUTION. Method: X-RAY DIFFRACTION; resolution 1.6 \u00c5; space group P 21 21 21; 746 polymer residues, 85.45 kDa; released 2008-10-14. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: DNA BINDING PROTEIN. Primary citation: Crystal structure of the Fic (Filamentation induced by cAMP) family protein SO4266 (gi|24375750) from Shewanella oneidensis MR-1 at 1.6 A resolution. (Proteins, 2009), DOI 10.1002/prot.22338.\n- **3E0F** \u2014 Crystal structure of a putative metal-dependent phosphoesterase (bad_1165) from bifidobacterium adolescentis atcc 15703 at 2.40 A resolution. Method: X-RAY DIFFRACTION; resolution 2.4 \u00c5; space group P 65; 301 polymer residues, 33.29 kDa; released 2008-09-02. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: HYDROLASE. Bound ligands: FE, PO4, ZN. Primary citation: Crystal structure of a metal-dependent phosphoesterase (YP_910028.1) from Bifidobacterium adolescentis: Computational prediction and experimental validation of phosphoesterase activity. (Proteins, 2011), DOI 10.1002/prot.23035.\n- **3DEE** \u2014 CRYSTAL STRUCTURE OF A PUTATIVE REGULATORY PROTEIN INVOLVED IN TRANSCRIPTION (NGO1945) FROM NEISSERIA GONORRHOEAE FA 1090 AT 2.25 A RESOLUTION. Method: X-RAY DIFFRACTION; resolution 2.1 \u00c5; space group C 1 2 1; 249 polymer residues, 29.15 kDa; released 2008-08-05. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: TRANSCRIPTION. Primary citation: The structure of the first representative of Pfam family PF09836 reveals a two-domain organization and suggests involvement in transcriptional regulation. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309109022672.\n- **3DUE** \u2014 CRYSTAL STRUCTURE OF A PUTATIVE PERIPLASMIC PROTEIN FROM DUF2874 FAMILY (BVU_2987) FROM BACTEROIDES VULGATUS ATCC 8482 AT 1.85 A RESOLUTION. Method: X-RAY DIFFRACTION; resolution 1.85 \u00c5; space group P 21 21 21; 127 polymer residues, 15.03 kDa; released 2008-08-05. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: UNKNOWN FUNCTION. Primary citation: The structure of BVU2987 from Bacteroides vulgatus reveals a superfamily of bacterial periplasmic proteins with possible inhibitory function. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309109046788.\n- **3DCX** \u2014 Crystal structure of a duf1696 family protein with a pleckstrin-homology domain (shew_0819) from shewanella loihica pv-4 at 2.00 A resolution. Method: X-RAY DIFFRACTION; resolution 2.0 \u00c5; space group P 21 21 21; 585 polymer residues, 66.41 kDa; released 2008-07-29. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: UNKNOWN FUNCTION. Primary citation: Bacterial pleckstrin homology domains: a prokaryotic origin for the PH domain. (J.Mol.Biol., 2010), DOI 10.1016/j.jmb.2009.11.006.\n- **3D00** \u2014 Crystal structure of a tungsten formylmethanofuran dehydrogenase subunit e (fmde)-like protein (syn_00638) from syntrophus aciditrophicus at 1.90 A resolution. Method: X-RAY DIFFRACTION; resolution 1.9 \u00c5; space group P 41 21 2; 191 polymer residues, 21.8 kDa; released 2008-05-20. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: METAL BINDING PROTEIN. Bound ligands: ZN. Primary citation: Structures of three members of Pfam PF02663 (FmdE) implicated in microbial methanogenesis reveal a conserved alpha+beta core domain and an auxiliary C-terminal treble-clef zinc finger. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309110020166.\n- **3CM1** \u2014 Crystal structure of SsgA-like sporulation-specific cell division protein (YP_290167.1) from Thermobifida fusca YX-ER1 at 2.60 A resolution. Method: X-RAY DIFFRACTION; resolution 2.6 \u00c5; space group P 41; 417 polymer residues, 46.25 kDa; released 2008-04-01. Deposition authors: Joint Center for Structural Genomics (JCSG), Chruszcz, M., Minor, W., Wang, S.. Keywords: CELL CYCLE. Primary citation: Structural and functional characterizations of SsgB, a conserved activator of developmental cell division in morphologically complex actinomycetes. (J.Biol.Chem., 2009), DOI 10.1074/jbc.M109.018564.\n- **3CGH** \u2014 Crystal structure of a susd homolog (bt_3984) from bacteroides thetaiotaomicron vpi-5482 at 1.70 A resolution. Method: X-RAY DIFFRACTION; resolution 1.7 \u00c5; space group C 2 2 21; 516 polymer residues, 57.69 kDa; released 2008-03-18. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: METAL BINDING PROTEIN. Bound ligands: ZN. Primary citation: Structure of BT_3984, a member of the SusD/RagB family of nutrient-binding molecules. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309110032999.\n- **3BY7** \u2014 CRYSTAL STRUCTURE OF A PROTEIN STRUCTURALLY SIMILAR TO SM/LSM-LIKE RNA-BINDING PROTEINS (JCVI_PEP_1096686650277) FROM UNCULTURED MARINE ORGANISM AT 2.60 A RESOLUTION. Method: X-RAY DIFFRACTION; resolution 2.6 \u00c5; space group C 1 2 1; 500 polymer residues, 55.74 kDa; released 2008-01-29. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: UNKNOWN FUNCTION. Primary citation: Crystal structure of a novel Sm-like protein of putative cyanophage origin at 2.60 A resolution. (Proteins, 2009), DOI 10.1002/prot.22360.\n- **3BYQ** \u2014 Crystal structure of a duf1185 family protein (bb2672) from bordetella bronchiseptica rb50 at 1.70 A resolution. Method: X-RAY DIFFRACTION; resolution 1.7 \u00c5; space group C 2 2 21; 579 polymer residues, 66.43 kDa; released 2008-01-29. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: UNKNOWN FUNCTION. Primary citation: Structures of the first representatives of Pfam family PF06684 (DUF1185) reveal a novel variant of the Bacillus chorismate mutase fold and suggest a role in amino-acid metabolism. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309109050647.\n- **3BOS** \u2014 Crystal structure of a putative dna replication regulator HDA (SAMA_1916) from Shewanella amazonensis sb2b at 1.75 A resolution. Method: X-RAY DIFFRACTION; resolution 1.75 \u00c5; space group P 21 21 21; 484 polymer residues, 56.87 kDa; released 2008-01-15. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: HYDROLASE REGULATOR,DNA BINDING PROTEIN. Bound ligands: CDP, MG, NA. Primary citation: A structural basis for the regulatory inactivation of DnaA. (J.Mol.Biol., 2009), DOI 10.1016/j.jmb.2008.10.059.\n- **3B77** \u2014 Crystal structure of a ph domain containing bacterial protein (exig_2160) from exiguobacterium sibiricum 255-15 at 2.42 A resolution. Method: X-RAY DIFFRACTION; resolution 2.42 \u00c5; space group P 4; 1158 polymer residues, 134.85 kDa; released 2007-11-13. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: UNKNOWN FUNCTION. Primary citation: Bacterial pleckstrin homology domains: a prokaryotic origin for the PH domain. (J.Mol.Biol., 2010), DOI 10.1016/j.jmb.2009.11.006.\n- **2RA9** \u2014 Crystal structure of a duf1285 family protein (sbal_2486) from shewanella baltica os155 at 1.40 A resolution. Method: X-RAY DIFFRACTION; resolution 1.4 \u00c5; space group P 21 21 21; 150 polymer residues, 17.38 kDa; released 2007-10-16. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: UNKNOWN FUNCTION. Bound ligands: NA. Primary citation: Structures of the first representatives of Pfam family PF06938 (DUF1285) reveal a new fold with repeated structural motifs and possible involvement in signal transduction. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309109050416.\n- **2RE3** \u2014 CRYSTAL STRUCTURE OF a DUF1285 family protein (SPO_0140) FROM SILICIBACTER POMEROYI DSS-3 AT 2.50 A RESOLUTION. Method: X-RAY DIFFRACTION; resolution 2.5 \u00c5; space group P 43 21 2; 388 polymer residues, 43.31 kDa; released 2007-10-09. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: UNKNOWN FUNCTION. Primary citation: Structures of the first representatives of Pfam family PF06938 (DUF1285) reveal a new fold with repeated structural motifs and possible involvement in signal transduction. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309109050416.\n- **2QTP** \u2014 Crystal structure of a duf1185 family protein (spo0826) from silicibacter pomeroyi dss-3 at 2.10 A resolution. Method: X-RAY DIFFRACTION; resolution 2.1 \u00c5; space group P 43 21 2; 194 polymer residues, 20.81 kDa; released 2007-08-21. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: UNKNOWN FUNCTION. Primary citation: Structures of the first representatives of Pfam family PF06684 (DUF1185) reveal a novel variant of the Bacillus chorismate mutase fold and suggest a role in amino-acid metabolism. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309109050647.\n- **2Q8U** \u2014 CRYSTAL STRUCTURE OF MRE11 FROM THERMOTOGA MARITIMA MSB8 (TM1635) AT 2.20 A RESOLUTION. Method: X-RAY DIFFRACTION; resolution 2.2 \u00c5; space group P 1 21 1; 672 polymer residues, 77.71 kDa; released 2007-06-26. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: HYDROLASE. Primary citation: Crystal structure of the first eubacterial Mre11 nuclease reveals novel features that may discriminate substrates during DNA repair. (J.Mol.Biol., 2010), DOI 10.1016/j.jmb.2010.01.049.\n- **2Q3L** \u2014 CRYSTAL STRUCTURE OF AN UNCHARACTERIZED PROTEIN FROM DUF3478 FAMILY WITH A SPOIIAA-LIKE FOLD (SHEW_3102) FROM SHEWANELLA LOIHICA PV-4 AT 2.25 A RESOLUTION. Method: X-RAY DIFFRACTION; resolution 2.25 \u00c5; space group C 2 2 21; 252 polymer residues, 29.73 kDa; released 2007-06-12. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: UNKNOWN FUNCTION. Bound ligands: NA. Primary citation: Open and closed conformations of two SpoIIAA-like proteins (YP_749275.1 and YP_001095227.1) provide insights into membrane association and ligand binding. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309109042481.\n- **2PV7** \u2014 Crystal structure of chorismate mutase / prephenate dehydrogenase (tyrA) (1574749) from Haemophilus influenzae RD at 2.00 A resolution. Method: X-RAY DIFFRACTION; resolution 2.0 \u00c5; space group P 41 21 2; 596 polymer residues, 70.24 kDa; released 2007-05-22. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: ISOMERASE, OXIDOREDUCTASE. Primary citation: The structure of Haemophilus influenzae prephenate dehydrogenase suggests unique features of bifunctional TyrA enzymes. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309110021688.\n- **2OOC** \u2014 Crystal structure of Histidine Phosphotransferase ShpA (NP_419930.1) from Caulobacter crescentus at 1.52 A resolution. Method: X-RAY DIFFRACTION; resolution 1.52 \u00c5; space group P 32 2 1; 226 polymer residues, 24.93 kDa; released 2007-02-06. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: TRANSFERASE. Primary citation: Crystal structure of histidine phosphotransfer protein ShpA, an essential regulator of stalk biogenesis in Caulobacter crescentus. (J.Mol.Biol., 2009), DOI 10.1016/j.jmb.2009.05.023.\n- **2OOK** \u2014 Crystal structure of a protein with unknown function (YP_749275.1) from Shewanella Frigidimarina NCIMB 400 at 1.80 A resolution. Method: X-RAY DIFFRACTION; resolution 1.8 \u00c5; space group C 1 2 1; 254 polymer residues, 29.81 kDa; released 2007-02-06. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: STRUCTURAL GENOMICS, UNKNOWN FUNCTION. Primary citation: Open and closed conformations of two SpoIIAA-like proteins (YP_749275.1 and YP_001095227.1) provide insights into membrane association and ligand binding. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309109042481.\n- **2IIZ** \u2014 Crystal structure of putative melanin biosynthesis protein TyrA with bound heme (NP_716371.1) from Shewanella Oneidensis at 2.30 A resolution. Method: X-RAY DIFFRACTION; resolution 2.3 \u00c5; space group P 43 21 2; 312 polymer residues, 37.29 kDa; released 2006-11-14. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: Structural Genomics/Unknown function. Bound ligands: HEM. Primary citation: Identification and structural characterization of heme binding in a novel dye-decolorizing peroxidase, TyrA. (Proteins, 2007), DOI 10.1002/prot.21673.\n- **2IAY** \u2014 Crystal structure of a duf1831 family protein (lp2179) from lactobacillus plantarum at 1.20 A resolution. Method: X-RAY DIFFRACTION; resolution 1.2 \u00c5; space group P 21 21 21; 114 polymer residues, 13.0 kDa; released 2006-10-10. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: TRANSLATION. Primary citation: Structure of LP2179, the first representative of Pfam family PF08866, suggests a new fold with a role in amino-acid metabolism. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309109023689.\n- **2ICH** \u2014 CRYSTAL STRUCTURE OF A PUTATIVE ATTH (NE1406) FROM NITROSOMONAS EUROPAEA AT 2.00 A RESOLUTION. Method: X-RAY DIFFRACTION; resolution 2.0 \u00c5; space group P 21 21 21; 670 polymer residues, 76.49 kDa; released 2006-10-03. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: LIPID BINDING PROTEIN. Primary citation: Structure of the first representative of Pfam family PF09410 (DUF2006) reveals a structural signature of the calycin superfamily that suggests a role in lipid metabolism. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309109037749.\n- **2HAG** \u2014 Crystal structure of a putative dyp-type peroxidase protein (so_0740) from shewanella oneidensis at 2.75 A resolution. Method: X-RAY DIFFRACTION; resolution 2.75 \u00c5; space group P 43 21 2; 312 polymer residues, 36.47 kDa; released 2006-08-08. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: HYDROLASE. Primary citation: Crystal structures of two novel dye-decolorizing peroxidases reveal a beta-barrel fold with a conserved heme-binding motif. (Proteins, 2007), DOI 10.1002/prot.21550.\n- **2HBW** \u2014 Crystal structure of a putative endopeptidase (ava_3396) from anabaena variabilis atcc 29413 at 1.05 A resolution. Method: X-RAY DIFFRACTION; resolution 1.05 \u00c5; space group P 21 21 2; 235 polymer residues, 25.65 kDa; released 2006-08-08. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: HYDROLASE. Primary citation: Structural Basis of Murein Peptide Specificity of a gamma-D-Glutamyl-L-Diamino Acid Endopeptidase. (Structure, 2009), DOI 10.1016/j.str.2008.12.008.\n- **2HUJ** \u2014 Crystal structure of a protein of uknown function (NP_471338.1) from Listeria innocua at 1.74 A resolution. Method: X-RAY DIFFRACTION; resolution 1.74 \u00c5; space group P 63 2 2; 140 polymer residues, 17.22 kDa; released 2006-08-08. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: STRUCTURAL GENOMICS, UNKNOWN FUNCTION. Primary citation: Crystal structures of MW1337R and lin2004: representatives of a novel protein family that adopt a four-helical bundle fold. (Proteins, 2008), DOI 10.1002/prot.22020.\n- **2GVI** \u2014 Crystal structure of a putative formylmethanofuran dehydrogenase subunit e (ta1109) from thermoplasma acidophilum at 1.87 A resolution. Method: X-RAY DIFFRACTION; resolution 1.87 \u00c5; space group I 2 2 2; 204 polymer residues, 24.63 kDa; released 2006-06-13. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: OXIDOREDUCTASE. Bound ligands: ACY, EDO, ZN. Primary citation: Structures of three members of Pfam PF02663 (FmdE) implicated in microbial methanogenesis reveal a conserved alpha+beta core domain and an auxiliary C-terminal treble-clef zinc finger. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309110020166.\n- **2H1T** \u2014 Crystal structure of a duf1089 family protein (pa1994) from pseudomonas aeruginosa at 1.80 A resolution. Method: X-RAY DIFFRACTION; resolution 1.8 \u00c5; space group C 1 2 1; 376 polymer residues, 44.27 kDa; released 2006-05-30. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: UNKNOWN FUNCTION. Primary citation: The structure of the first representative of Pfam family PF06475 reveals a new fold with possible involvement in glycolipid metabolism. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309109022684.\n- **2GVK** \u2014 Crystal structure of a dye-decolorizing peroxidase (DyP) from Bacteroides thetaiotaomicron VPI-5482 at 1.6 A resolution. Method: X-RAY DIFFRACTION; resolution 1.6 \u00c5; space group P 63 2 2; 317 polymer residues, 36.2 kDa; released 2006-05-16. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: OXIDOREDUCTASE. Bound ligands: NA. Primary citation: Crystal structures of two novel dye-decolorizing peroxidases reveal a beta-barrel fold with a conserved heme-binding motif. (Proteins, 2007), DOI 10.1002/prot.21550.\n- **2GLZ** \u2014 Crystal structure of a formylmethanofuran dehydrogenase subunit e-like protein (dhaf_2992) from desulfitobacterium hafniense dcb-2 at 1.45 A resolution. Method: X-RAY DIFFRACTION; resolution 1.45 \u00c5; space group P 21 21 21; 306 polymer residues, 36.52 kDa; released 2006-04-18. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: METAL BINDING PROTEIN. Bound ligands: NI, ZN. Primary citation: Structures of three members of Pfam PF02663 (FmdE) implicated in microbial methanogenesis reveal a conserved alpha+beta core domain and an auxiliary C-terminal treble-clef zinc finger. (Acta Crystallogr.,Sect.F, 2010), DOI 10.1107/S1744309110020166.\n- **2GHR** \u2014 Crystal structure of homoserine o-succinyltransferase (NP_981826.1) from Bacillus cereus ATCC 10987 at 2.40 A resolution. Method: X-RAY DIFFRACTION; resolution 2.4 \u00c5; space group P 41 2 2; 302 polymer residues, 35.82 kDa; released 2006-04-11. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: TRANSFERASE. Primary citation: Crystal structure of homoserine O-succinyltransferase from Bacillus cereus at 2.4 A resolution (Proteins, 2007), DOI 10.1002/prot.21208.\n- **2FG0** \u2014 Crystal structure of a putative gamma-d-glutamyl-l-diamino acid endopeptidase (npun_r0659) from nostoc punctiforme pcc 73102 at 1.79 A resolution. Method: X-RAY DIFFRACTION; resolution 1.79 \u00c5; space group P 41 21 2; 492 polymer residues, 55.46 kDa; released 2006-01-10. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: HYDROLASE. Primary citation: Structural Basis of Murein Peptide Specificity of a gamma-D-Glutamyl-L-Diamino Acid Endopeptidase. (Structure, 2009), DOI 10.1016/j.str.2008.12.008.\n- **2EVR** \u2014 CRYSTAL STRUCTURE OF A PUTATIVE GAMMA-D-GLUTAMYL-L-DIAMINO ACID ENDOPEPTIDASE (NPUN_R0659) FROM NOSTOC PUNCTIFORME PCC 73102 AT 1.60 A RESOLUTION. Method: X-RAY DIFFRACTION; resolution 1.6 \u00c5; space group P 41 2 2; 246 polymer residues, 27.96 kDa; released 2005-11-22. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: HYDROLASE. Bound ligands: NA. Primary citation: Structural Basis of Murein Peptide Specificity of a gamma-D-Glutamyl-L-Diamino Acid Endopeptidase. (Structure, 2009), DOI 10.1016/j.str.2008.12.008.\n- **2ETS** \u2014 CRYSTAL STRUCTURE OF A BACTERIAL DOMAIN OF UNKNOWN FUNCTION FROM DUF1798 FAMILY (MW1337) FROM STAPHYLOCOCCUS AUREUS SUBSP. AUREUS AT 2.25 A RESOLUTION. Method: X-RAY DIFFRACTION; resolution 2.25 \u00c5; space group H 3 2; 128 polymer residues, 15.66 kDa; released 2005-11-08. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: UNKNOWN FUNCTION. Primary citation: Crystal structures of MW1337R and lin2004: representatives of a novel protein family that adopt a four-helical bundle fold. (Proteins, 2008), DOI 10.1002/prot.22020.\n- **1VLQ** \u2014 Crystal structure of Acetyl xylan esterase (TM0077) from Thermotoga maritima at 2.10 A resolution. Method: X-RAY DIFFRACTION; resolution 2.1 \u00c5; space group P 1 21 1; 4044 polymer residues, 467.99 kDa; released 2004-08-24. Deposition authors: Joint Center for Structural Genomics (JCSG). Keywords: HYDROLASE. Primary citation: Functional and structural characterization of a thermostable acetyl esterase from Thermotoga maritima. (Proteins, 2012), DOI 10.1002/prot.24041.\n- **1EVS** \u2014 CRYSTAL STRUCTURE OF HUMAN ONCOSTATIN M. Method: X-RAY DIFFRACTION; resolution 2.2 \u00c5; space group P 21 21 21; 187 polymer residues, 21.14 kDa; released 2000-09-13. Deposition authors: Deller, M.C., Hudson, K.R., Ikemizu, S., Bravo, J., Jones, E.Y., Heath, J.K.. Keywords: CYTOKINE. Primary citation: Crystal structure and functional dissection of the cytostatic cytokine oncostatin M. (Structure Fold.Des., 2000), DOI 10.1016/S0969-2126(00)00176-3.", "# Crystal structure of a soluble cleaved HIV-1 envelope trimer\nAuthors: Julien, J.-P.; Cupo, A.; Sok, D.; Stanfield, R.L.; Lyumkis, D.; Deller, M.C.; Klasse, P.-J.; Burton, D.R.; Sanders, R.W.; Moore, J.P.; Ward, A.B.; Wilson, I.A.\nSource: Science, 2013\n\nWe thank Y. Hua, L. Kong, P.S. Lee, X. Zhu, R. Pejchal, H. Tien, T. Clayton, K. Saye and J. Korzun, for technical assistance, previous contributions to the Env trimer project, and insightful discussions and C.R. King and W. Koff for support and encouragement. This work was supported by NIH P01 AI82362 (J.P.M., I.A.W.), as well as the International AIDS Vaccine Initiative Neutralizing Antibody Consortium and Center (D.R.B., I.A.W., J.P.M., A.B.W.), CHAVI-ID UM1 AI100663 (D.R.B., I.A.W., A.B.W.), NIH R01 AI084817 (I.A.W.), NIH R37 AI36082 (J.P.M.), NIH R01 AI33292 (D.R.B.), a Vidi grant from the Netherlands Organization for Scientific Research (R.W.S.), a Starting Investigator Grant from the European Research Council (R.W.S.), Canadian Institutes of Health Research fellowship (J.-P.J.), and the Ragon Institute. D.L. is supported by the U.S. NIH NIGMS Biomedical Technology Research Center program (GM103310). Use of the Advanced Photon Source for data collection was supported by the DOE, Basic Energy Sciences, Office of Science, under contract no. DE-AC02-06CH11357. GM/CA CAT has been funded in whole or in part with federal funds from NCI (grant Y1-CO-1020) and NIGMS (grant Y1-GM-1104). Extensive crystal screening was also carried out at the Stanford Synchrotron Radiation Lightsource (SSRL) and at the Canadian Light Source (CLS). SSRL, a Directorate of the SLAC National Accelerator Laboratory and an Office of Science User Facility is operated for the U.S. Department of Energy (DOE) Office of Science by Stanford University. The SSRL Structural Molecular Biology Program is supported by the DOE Office of Biological and Environmental Research; NIH's National Center for Research Resources, Biomedical Technology Program (P41RR001209); and the National Institute of General Medical Sciences (NIGMS). CLS is supported by the Natural Sciences and Engineering Research Council of Canada, the National Research Council Canada, the Canadian Institutes of Health Research, the Province of Saskatchewan, Western Economic Diversification Canada, and the University of Saskatchewan. The content is the responsibility of the authors and does not necessarily reflect the official views of the NIGMS, NCI or NIH. Coordinates and structure factors have been deposited with the Protein Data Bank under accession code 4NCO. The International AIDS Vaccine Initiative (IAVI) has previously filed a patent relating to the BG505 SOSIP.664 trimer: U.S. Prov. Appln. No. 61/772,739, titled \"HIV-1 envelope glycoprotein\", with inventors M. Caulfield, A.C., H. Dean, S. Hoffenberg, C.R.K., P.J.K., A. Marozsan, J.P.M., R.S., A.B.W., I.A.W., J.-P.J., but no patents have been filed on any work described here. Materials and information will be provided under an MTA. This is manuscript 25054 from The Scripps Research Institute.\n\n## References\n\n*(reference list omitted)*", "# GitHub README: microplastic-blaster\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/microplastic-blaster, 2026\n\nElora Therapeutics is developing the first FDA-targeted enzyme therapeutic \u2014 based on engineered PETase and MHETase \u2014 designed to actively depolymerise PET microplastics accumulated in human biological matrices. MicroPlastic Blaster is a public-awareness and brand engagement tool built to communicate the scale of the microplastic health problem in an accessible, interactive format.\n\n**Website:** [eloratherapeutics.com](https://www.eloratherapeutics.com)\n\n---\n\n## \ud83d\udcc4 Licence\n\n\u00a9 Elora Therapeutics. All rights reserved. Assets (logos, avatars, audio) are proprietary. The game logic may be adapted for educational or awareness purposes with attribution.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/microplastic-blaster\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/microplastic-blaster</a></td>\n</tr>\n</table>", "# Approaches to automated protein crystal harvesting\nAuthors: Deller, M. C.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 70, 133\u2013155 (IYCr crystallization series), 2014\n\n**Authors:** Marc C. Deller (The Joint Center for Structural Genomics, The Scripps Research Institute, 10550 North Torrey Pines Road, La Jolla, California, USA; corresponding author, mdeller@scripps.edu) and Bernhard Rupp (Department of Forensic Crystallography, k.-k. Hofkristallamt, Vista, California, USA, and Department of Genetic Epidemiology, Innsbruck Medical University, Sch\u00f6pfstrasse 41, 6020 Innsbruck, Austria; br@ruppweb.org).\n\n**Published:** *Acta Crystallographica Section F* **70**, 133\u2013155 (2014); part of the IYCr crystallization series. Received 9 December 2013; accepted 7 January 2014.\n\n*This document is a structured summary of the review, not the full text.*\n\n## Abstract\n\nThe harvesting of protein crystals is almost always a necessary step in the determination of a protein structure using X-ray crystallographic techniques. However, protein crystals are usually fragile and susceptible to damage during the harvesting process. For this reason, protein crystal harvesting is the single step that remains entirely dependent on skilled human intervention. Automation has been implemented in the majority of other stages of the structure-determination pipeline, including cloning, expression, purification, crystallization and data collection. The gap in automation between crystallization and data collection results in a bottleneck in throughput and presents unfortunate opportunities for crystal damage. Several automated protein crystal harvesting systems have been developed, including systems utilizing microcapillaries, microtools, microgrippers, acoustic droplet ejection and optical traps. However, these systems have yet to be commonly deployed in the majority of crystallography laboratories owing to a variety of technical and cost-related issues. Automation of protein crystal harvesting remains essential for harnessing the full benefits of fourth-generation synchrotrons, free-electron lasers and microfocus beamlines. Furthermore, automation of protein crystal harvesting offers several benefits when compared with traditional manual approaches, including the ability to harvest microcrystals, improved flash-cooling procedures and increased throughput.\n\n## Section-by-section synopsis\n\n### 1. Introduction\n\nMarc Deller and Bernhard Rupp open by noting that X-ray crystallography accounted for 88% of all macromolecular structures in the Protein Data Bank as of January 2014 \u2014 85,486 of 96,692 entries \u2014 and that crystal harvesting, defined broadly as the manipulations that carry a crystal from its growth medium into the X-ray beam in a diffraction-ready state, is essential to every one of them. Robotics and automation, largely driven by structural genomics, have produced high-throughput protein crystallography (HTPX) covering cloning, expression, purification, crystallization, data collection and processing, so it is surprising that harvesting alone remains barely automated. The authors attribute this to the unique difficulty of the task: protein crystals are extremely fragile and sensitive to subtle changes in osmolarity, pH, chemical environment and, to a lesser extent, temperature, making the micromanipulation hard to automate and easy to omit from automation discussions. The consequence is a leaking pipeline in which late-stage losses arise from physical damage caused by operator handling, physical and chemical damage from poor cryoprotection, and loss of diffraction during unsuccessful flash-cooling \u2014 all problems amenable to precise, reproducible automated control.\n\n### 1.1. Impact of HTPX/SG\n\nLarge-scale HTPX and structural genomics centres determine hundreds of novel structures a year, and their central advantage is the pooling of resources and equipment across multiple principal investigators, allowing robotics to be leveraged at a scale impossible in a single-investigator laboratory. In the United States the review names the Joint Center for Structural Genomics (JCSG), the Midwest Center for Structural Genomics (MCSG), the New York Structural Genomics Consortium (NYSGC) and the Northeast Structural Genomics Consortium (NSGC). Worldwide HTPX/SG efforts account for over 10% of all protein structures deposited in the PDB, and that proportion rises steeply for families with no previous structural coverage, where about 50% of all novel structures come from HTPX/SG centres. Many synchrotron facilities have added or are adding HTPX/SG capabilities, several with on-site crystallization facilities and high-flux microfocus beamlines capable of in situ data collection, which gives industrial and academic users outside traditional SG centres access to the same automation. Marc Deller and Bernhard Rupp argue that advances in automated harvesting will let these facilities increase both throughput and structure output.\n\n### 1.2.1. The protein crystal-harvesting process", "# Unlocking the Promise and Pitfalls of Large Language Models in Biomedical NLP\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-17 on marcdeller.com (categories: ai).*\n\n* State-of-the-art fine-tuned models outperform current LLMs in extraction and classification within BioNLP.\n* GPT-4 excels in medical question answering, achieving up to 30% higher accuracy over prior best methods.\n* LLMs are prone to inconsistencies, missing information, and hallucinations\u2014especially in zero-shot mode.\n* Advanced prompt engineering and few-shot learning measurably reduce errors in LLM outputs.\n* Cost and context length limits remain significant barriers to LLM deployment in routine biomedical applications.\n\n1. **Chen, Q., et al. (2025). Benchmarking large language models for biomedical natural language processing applications and recommendations:**\n   <https://doi.org/10.1038/s41467-025-56989-2>\n2. **Supplementary datasets and code repository:**\n   <https://doi.org/10.5281/zenodo.14025500>\n3. **BioBERT: Lee, J. et al. (2020). BioBERT: a pre-trained biomedical language representation model for biomedical text mining:**\n   <https://academic.oup.com/bioinformatics/article/36/4/1234/5566506>\n4. **PubMedBERT: Gu, Y. et al. (2021). Domain-specific language model pretraining for biomedical natural language processing:**\n   <https://dl.acm.org/doi/10.1145/3450439>\n5. **PMC LLaMA model resource: Wu, C. et al. (2024). PMC-LLaMA: toward building open-source language models for medicine:**\n   <https://jamanetwork.com/journals/jamia/article-abstract/2817226>\n\nThe explosion of biomedical literature poses daunting challenges for researchers, clinicians, and knowledge managers. With over 36 million articles on PubMed and thousands more released each day, extracting actionable knowledge has become a monumental task. Enter biomedical Natural Language Processing (BioNLP), a field supercharged by the rise of large language models (LLMs) like GPT-3.5, GPT-4, and open-source alternatives such as LLaMA 2. But do these cutting-edge models truly outperform specialized, fine-tuned tools in extracting, summarizing, and interpreting complex bioscience texts?\n\nA comprehensive study by Chen et al. systematically benchmarks leading LLMs against state-of-the-art (SOTA) fine-tuning methods across 12 major biomedical NLP tasks. The results reveal both the immense potential and practical limitations of deploying generalist LLMs in specialized biomedical domains.\n\n## The Race: Generalist LLMs vs. Domain-Specific Fine-Tuning\n\nThe study pits four prominent LLMs\u2014GPT-3.5, GPT-4, LLaMA 2, and PMC LLaMA\u2014against fine-tuned domain-specific models like BioBERT and BioBART in tasks spanning named entity recognition, relation extraction, multi-label classification, question answering, summarization, and text simplification.\n\nKey findings show that:\n\n* **Fine-tuned BERT/BART models consistently outperform LLMs in most extractive tasks**, such as identifying gene names or chemical-disease relations, with up to 40% higher accuracy in information extraction.\n* **LLMs, particularly GPT-4, shine in reasoning-heavy applications,** notably medical question answering. Here, GPT-4 surpasses the SOTA fine-tuned methods by nearly 30% in accuracy\u2014a notable leap for zero- and few-shot learning.\n* For generative tasks such as summarization and simplification, LLMs deliver competitive accuracy and notably higher readability, though they may lag in completeness compared to fine-tuned baselines.\n\n## Hallucinations and Inconsistencies: The Reliability Dilemma\n\nDespite their capabilities, LLMs\u2014especially in zero-shot settings\u2014frequently produce inconsistent, incomplete, or hallucinated outputs. LLaMA 2, for example, exhibited hallucinations in up to 32% of test cases for complex classification tasks. However, introducing even a single carefully selected example (few-shot learning) can dramatically curb these errors.\n\nManual review remains essential, given that automatic scoring models (such as ROUGE-L for summaries) do not always align with human evaluators, who consistently preferred the readability of GPT-generated summaries.\n\n## Practical Considerations: Cost, Context, and Open-Science\n\nCost is a critical factor in real-world application. GPT-4, while leading in performance for some tasks, can be 60 to 100 times more expensive than its predecessors\u2014raising scalability concerns for large-scale or routine use.\n\nInterestingly, the study found that **continual biomedical pretraining (as in PMC LLaMA) offers only marginal gains** over standard open-source models after fine-tuning. This highlights a need for more efficient, sustainable development of domain-specific LLMs.\n\n## Key Recommendations\n\nBased on extensive benchmarking and error analysis, the authors suggest:\n\n* **Use fine-tuned BERT/BART models as the go-to for extractive tasks.**\n* **Deploy closed-source LLMs, starting with GPT-3.5, for reasoning and summarization, keeping costs in mind.**\n* **Apply advanced prompt engineering and manual review to minimize hallucinations and maximize reliability.**\n* **Fine-tune open-source LLMs when input lengths and resources allow for task-specific performance boosts.**\n* **Revisit evaluation paradigms**, as existing benchmarks may not capture the nuanced performance of LLMs, especially in generative tasks.\n\n## The Road Ahead\n\nWhile LLMs have closed part of the gap with traditional fine-tuned models in some BioNLP tasks\u2014especially those demanding reasoning\u2014the field still contends with reliability, reproducibility, and cost challenges. Tailored evaluation methodologies and new benchmarks are needed to fully harness the transformative power of LLMs in biomedicine.", "# alphabetti.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n* **Single chain only.** No complexes, no assemblies, no ligands, no cofactors, no metals, no post-translational modifications.\n* **Maximum 400 residues**, minimum 10.\n* **ESMFold degrades on low-homology and disordered sequences**, and it degrades quietly. Because ESMFold's trunk *is* ESM-2, the language model and the structure fail together rather than independently: when the model has no evolutionary signal for a sequence, you get both a poor fold and a flat, uninformative logo, and the two will not contradict each other.\n* **A worked example of that failure: avGFP.** Masked marginals recover\n  the wild-type residue at 10 % of positions, against 8 % for a randomly\n  shuffled version of the same sequence, and mean information content is\n  0.27 of a possible 4.322 bits. Mean pLDDT is 43. It is a famous,\n  heavily studied protein that this method has almost nothing useful to\n  say about, and nothing on screen would tell you so except the\n  confidence number. **Check the pLDDT before you believe a picture.**\n* **The database sequence is usually a precursor.** The examples name\n  explicit residue ranges from UniProt's Chain feature for this reason:\n  P0CG48 is nine tandem copies of ubiquitin, not one, and P00698 carries\n  an 18-residue signal peptide that is cleaved in vivo and folds into\n  nothing. Paste an accession and you get the whole database entry, tail\n  and all. Removing lysozyme's signal peptide raised its mean pLDDT from\n  91.0 to 95.1.\n* **These are language model scores, not clinical predictions.** Reported ClinVar and gnomAD annotations, where shown, come from the EBI Proteins API and are a separate kind of evidence entirely.\n* A predicted structure is a hypothesis. So is a language model's opinion. This app draws one on top of the other.\n\n## Colophon\n\nFlask and Gunicorn behind nginx; three.js with no build step; ESMFold and\nESM-2 on a ZeroGPU Space. The 3D letterforms are Baloo 2 Bold, converted to\noutlines and extruded. Written by\n[Marc C. Deller, D.Phil.](https://marcdeller.com)\nBackend running the `hf_space` fold backend.", "# Approaches to automated protein crystal harvesting\nAuthors: Deller, M. C.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 70, 133\u2013155 (IYCr crystallization series), 2014\n\nThe engineering challenges reduce to two: identifying the target crystal and manipulating it without damage. Microcapillary methods, used in PCGDS, CPPI, ACH, CARESS and CMM, are the easiest and cheapest to automate, requiring only a simple pump or syringe as in established micro-injection techniques; they offer the most reliable pickup, since the crystal follows the fluid flow rather than having to be chased, minimize mechanical contact by keeping the crystal in solution, protect against dehydration, and cope well with viscous media such as LCP. Their major weakness is incompatibility with current cryoprotection and flash-cooling protocols, requiring an extra transfer step into a microloop or microtool. Microtool methods, used by UMR, CH and SAMI, are cheap, relatively easy to automate and \u2014 most importantly \u2014 compatible with existing flash-cooling and data-collection protocols, but capture and manipulation are difficult because fluid flow and bow waves cause the crystal to be chased around the drop, damaging it and its neighbours; growing crystals directly in the cryoloop or micromount is noted as a promising workaround. Microgripper methods, used by chopsticks and REACH, may offer the easiest route to automation, with easy pickup and reduced collateral damage, but force-feedback grippers are expensive and compatibility with flash-cooling and data collection is low, although REACH showed data can be collected directly from crystals held in the grippers.\n\n### 3.4. Emerging harvesting methods (ADE, OT and MEMS)\n\nMicrotool and microgripper methods both involve real mechanical contact with the crystal, and the review argues that many of the associated problems can be avoided by using non-mechanical forces \u2014 sound for ADE, light for optical traps and fluid flow for MEMS. These are less physically hazardous and more precise for microcrystals, but ADE and COT need expensive, bulky control and focusing systems and are limited in target thickness and crystal size, and the incident light or sound energy must be controlled to avoid thermal and photo-damage; FOT removes the need for bulky high-NA focusing and allows harvesting from a wider range of plates and sample sizes. Two further optical approaches are described: the Crystal Direct method, which uses laser-induced photoablation of ultrathin 12.5 \u00b5m Kapton films that serve as the crystal growth support, excising the film for mounting on a standard pin \u2014 attractive for large numbers of microcrystals with raster scanning, and eliminating any tool entering the drop; and the simpler tool-free Crystal Catcher, which injects vinyl adhesives onto a mounting pin to capture the crystal without mechanical damage, sharing the low-background benefits of loopless mounting and so suited to S-SAD. MEMS techniques such as the RodBot are singled out as exciting because hydrodynamic flows capture and move crystals with no direct physical contact and control is purely magnetic, avoiding the complications of mechanical components near a fragile crystal.\n\n### 3.5. In situ protein crystal growth and mounting\n\nIn situ growth and data collection depend on close coupling between growth medium and X-ray source, and advances in home and synchrotron sources plus plate design and materials have driven a surge of interest. Solutions range from simple SBS-format plates to complex microfluidic chips. Because the crystal is never removed from the mother liquor, most of the engineering challenges of harvesting disappear, physical contact and further manipulation are eliminated, and late-stage damage is limited; in situ methods also expedite the search for well diffracting crystals and the systematic analysis of parameters such as dehydration, and have solved viral and ligand-bound structures where manual mounting had failed. Named plate designs include the X-Chip, Crystal Former, In-Situ-1, topographic drop-pinning plate, in situ crystallization plate, CrystalQuick X and CrystalSlide. The downsides are scattering by plate material and mother liquor, a limited rotation angle and hence less usable data per crystal, and restriction to room-temperature collection because crystals cannot be cryocooled in the plate. Mitigations include low-birefringence, low-scattering materials such as the cyclic olefin copolymer of CrystalQuick X and well profiles permitting up to 80\u00b0 of data. In situ screening of SBS plates is available at the ESRF, SLS, Diamond, Helmholtz-Zentrum Berlin and the Australian Synchrotron, and in-house options include the PX scanner, PlateMate and a dedicated LCP plate holder, though in-house exposures of 10\u201320 min limit these to pre-screening and prioritization. The section also defines the sphere of confusion \u2014 the minimum spherical volume enclosing crystal movement when all goniometer axes rotate \u2014 typically 1\u2013100 \u00b5m commercially and submicrometre for piezo devices; for microcrystals it must be smaller than the beam or the crystal, and GM/CA minibeam work at APS beamline 23ID-B showed a 10 \u00b5m crystal in a 10 \u00b5m beam requires an SoC of 2 \u00b5m or less. REACH's arm is reported at better than 13 \u00b5m, and robotically harvested lysozyme and NikA\u2013FeEDTA gave data comparable to manual methods.\n\n### 4.1. Ability to harvest microcrystals", "# GitHub README: chatmol\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/chatmol, 2026\n\n4. **Execution.** Each returned line is echoed as `[ChatMol] > ...` and run\n   through the *original* `cmd.do`, so results are indistinguishable from\n   typing the command yourself (including normal PyMOL error messages if a\n   translated command is wrong).\n\n---\n\n## 6. Safety notes\n\n- ChatMol executes whatever the model returns. A bad translation could run\n  a destructive command (e.g. `delete all`, overwriting a file with `save`).\n  Review the echoed `[ChatMol] > ...` lines, especially for destructive\n  requests, and use `chatmol_config auto, off` + the explicit `ai` command\n  if you want a manual checkpoint before every translation.\n- Everything runs locally against your Ollama server -- no data leaves your\n  machine.\n- `chatmol_config debug, on` is useful when a translation looks wrong: it\n  prints the model's raw output (including any stripped `<think>` blocks\n  or code fences) so you can see what was sent to PyMOL.\n\n---\n\n## 7. Troubleshooting\n\n**`Could not reach Ollama at http://localhost:11434 ...`**\nStart Ollama (`ollama serve` or open the desktop app) and confirm the model\nis pulled (`ollama pull qwen3.6:27b`, check with `ollama list`).\n\n**`Ollama at ... did not respond to model '...' within 120s`**\nThe model took too long to respond -- this is caught gracefully and won't\ncrash PyMOL. Common causes: the model is still loading into memory (try\nagain), the machine is under heavy load (e.g. another `ollama pull` running\nin the background), or the model is just slow for your hardware. Increase\n`chatmol_config timeout, 240` or switch to a smaller/faster model.\n\n**`The model's response looked malformed or repetitive ... and was\ndiscarded`**\nSmall models occasionally loop (e.g. repeating `enable all; enable all; ...`)\ninstead of producing a short translation. ChatMol detects this, discards the\nresponse, and -- importantly -- does *not* add it to conversation history (so\nit can't poison later requests). Just try again, rephrase, or switch to a\nlarger model (`chatmol_config model, qwen3.6:27b`).\n\n**Nothing happens / commands look wrong**\nTurn on debug mode (`chatmol_config debug, on`) and re-run the request to\nsee the model's raw output. Try a larger model for better translation\nquality, or lower `temperature` for more deterministic output.\n\n**A plain command got sent to the LLM by mistake**\nThe heuristic in `nl_router.looks_like_pymol()` is conservative but not\nperfect. The system prompt asks the model to pass valid PyMOL syntax\nthrough unchanged, so this is usually harmless -- but if you hit a\nrecurring case, add the command/namespace to `_EXTRA_KEYWORDS` in\n`nl_router.py`.\n\n**Disable everything temporarily**\n`chatmol_config auto, off` stops the command line from auto-translating;\n`ai <request>` / `chatmol <request>` still work explicitly.\n\n---\n\n## 8. Extending\n\n- **More commands / better translations**: extend\n  `pymol_reference.PYMOL_CHEAT_SHEET` with examples for operations you use\n  often (e.g. project-specific selection macros).\n- **Session-aware context**: `pymol_reference.build_context()` is the place\n  to add more state (current view, sequence, B-factor ranges, etc.).\n- **Confirmation step**: wrap `nl_router.run_natural_language`'s execution\n  loop with a prompt/dialog if you want to approve translated commands\n  before they run.\n- **Streaming / richer UI**: a dedicated chat panel could be added in\n  `gui.py` using `pymol.Qt` for a conversational view alongside the command\n  line.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/chatmol\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/chatmol</a></td>\n</tr>\n</table>", "# Generative AI for Protein Pocket Design: The Next Leap in Drug Discovery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: blog).*\n\n* New generative AI models like PocketGen can design novel protein binding pockets from scratch, tailored to bind a specific drug molecule.\n* This technology generates both the sequence and the 3D structure of the pocket, accelerating the design of high-affinity protein binders.\n* By enabling the creation of custom \u201clocks\u201d for any given molecular \u201ckey,\u201d this approach represents a major leap forward for drug discovery and synthetic biology.\n\n1. Efficient generation of protein pockets with PocketGen: <https://www.nature.com/articles/s42256-024-00920-9>\n2. Efficient generation of protein pockets with PocketGen (Journal Article): <https://par.nsf.gov/biblio/10572215-efficient-generation-protein-pockets-pocketgen>\n\nThe central task of structure-based drug design is to find or create a molecule that fits perfectly into a binding pocket on a target protein. For decades, this process has involved starting with an existing protein and then screening countless compounds to find one that binds. But what if, instead of searching for a key to fit a given lock, we could design a brand-new lock, perfectly shaped to bind a specific key?\n\nThis is the paradigm-shifting concept behind new generative AI tools like **PocketGen**. Rather than being limited to the binding pockets that nature has provided, this deep generative model allows scientists to design novel protein pockets from the ground up. Given a specific drug molecule (the \u201ckey\u201d) and a protein scaffold, PocketGen can generate both the amino acid sequence and the precise 3D atomic structure of a pocket that will bind that molecule with high affinity. It designs the lock and key simultaneously.\n\nPocketGen achieves this through a sophisticated graph transformer architecture that captures interactions at the atomic, residue, and ligand level, ensuring the generated pocket is not only structurally sound but also chemically compatible with the drug. The results are stunning: the model operates at speeds ten times faster than older physics-based methods and successfully generates high-affinity pockets with remarkable accuracy. This technology fundamentally accelerates the design of protein-binding therapeutics and diagnostics. It frees drug discovery from the constraints of existing biology and opens the door to creating highly specific, custom-designed protein binders for virtually any molecule of interest.", "# Crystal structure of a soluble cleaved HIV-1 envelope trimer\nAuthors: Julien, J.-P.; Cupo, A.; Sok, D.; Stanfield, R.L.; Lyumkis, D.; Deller, M.C.; Klasse, P.-J.; Burton, D.R.; Sanders, R.W.; Moore, J.P.; Ward, A.B.; Wilson, I.A.\nSource: Science, 2013\n\nThe weak and diffuse electron density corresponding to the FP and the residues that connect it to the HR1 central helix makes model building of this region particularly difficult at this resolution (fig. S7A), but is likely indicative of lack of regular secondary structure. Nonetheless, residues adjacent to the N-termini of the HR1 helices initially wrap around the gp120 \u03b10 helix, forming a meta-stable loop-helix structure akin to similar elements in influenza HA2 (Fig. 3C-D) that lead to an extension of the central helix in the post-fusion form. Overall, the central trimeric coiled-coil arrangement of the fusion protein (gp41) surrounded by three receptor subunits (gp120) is a characteristic of type-1 membrane-fusion proteins. We now extend that hallmark to the capping of the central helices by another helix (cf. \u03b11 and \u03b1105-115) in both gp120 and HA. A small inter-protomer central opening is created at the top of the central helix in both fusion proteins (Fig. 3C), as well as in Ebola and PIV5/RSV fusion proteins, that presumably facilitates conformational rearrangements of the head domains (e.g. gp120, HA1) during the fusion process. The Env trimer structure also agrees well with what is known about the transition from a pre-fusion to a post-fusion conformation, in that the three HR1 central helices adopt a similar arrangement in the post-fusion six-helix bundle (Fig. 3E).\n\nA second, well-defined, ~40 \u00c5 long, helical element, at a ~60\u00b0 angle relative to the HR1 central helix, is located at the bottom of the trimer, proximal to the membrane, where it wraps around the trimer base (Fig. 3A). We interpret this long helix (7 turns) as the C-terminal portion of the HR2 sequence, which is consistent with a cryo-EM reconstruction in the accompanying paper, where deletion of residues 651-664 from the BG505 SOSIP.664 trimer eliminated the density corresponding to the end of this helix. Strong electron density between the bottom of HR1 and the middle of the HR2 helix is likely indicative of substantial intra-subunit gp41-stabilizing interactions. The residues that are N-terminal to the long HR2 helix appear to adopt a horseshoe conformation for the polypeptide backbone (Fig. 3A and fig. S4B); the NAG moieties at glycosylation sites 625 and 637 are clearly visible in the electron density maps and define the location of turns. Strong electron density, predominantly in a flat-extended shape characteristic of \u03b2-strands, is present between the C-terminus of HR1 and the N-terminus of HR2, where the gp41 disulfide loop (DSL; residues 585-609) is thought to be located (fig. S7B). However, we cannot determine the arrangement of these gp41 components with confidence at this resolution, and a further complication is the close proximity of gp120 C1 and C5 elements that most likely participate in tertiary structure interactions with residues connecting HR1 to HR2, including the DSL. As such, we attribute the approximate location of the gp41 DSL and, hence, the engineered inter-subunit disulfide bond, to the base of the Env trimer. Finally, no conclusion can be drawn about the arrangement of the gp41 membrane-proximal-external-region (MPER) because it is not present in the BG505 SOSIP.664 gp140 trimer. The gp41 HR2 may continue to extend in the same orientation as it enters the MPER, or could change direction near residue 664 so that the MPER could align parallel to the trimer axis.\n\n## Structural definition of the PGT122 epitope", "# BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n| Problem | Fix |\n| --- | --- |\n| `setup-plip` fails, or `pip install plip` tries to build OpenBabel from source | This is expected without conda-forge \u2014 `plip`\u2018s own installer forces a from-source OpenBabel rebuild unless OpenBabel is already importable *inside pip\u2019s build sandbox*, and the standalone PyPI `pymol-open-source` wheel has a hardcoded broken library path on at least some machines. `setup-plip` works around both by building a conda-forge env via a self-downloaded micromamba \u2014 just re-run `python3 BoltzMaker.py setup-plip --force` if a previous attempt left a half-built `.plip_env`. |\n| A `preflight`/`analyze` step involving `.plip_env` errors with `ModuleNotFoundError: No module named 'chatmol'` (or similar) | A stray `~/.pymolrc.py` on your machine (e.g. from an unrelated PyMOL plugin) is being loaded by the bundled PyMOL. BoltzMaker already overrides `HOME` for these subprocess calls so this shouldn\u2019t reach you, but if it does, check `~/.pymolrc.py` for anything referencing a package not installed in `.plip_env`. |\n| `run` seems to hang with no progress, or your Mac gets extremely slow | Check the memory-usage figure in the progress bar and see \u201cMemory on Mac\u201d earlier in this document \u2014 this is very likely swap-thrashing, not a genuine stall. Re-run with a lower `--mps-watermark`, `--workers 1`, and `--max-parallel-samples 1`. |\n| A target\u2019s YAML/CIF exists on disk but BoltzMaker says it\u2019s missing, or `preflight` hangs | Check for iCloud \u201cOptimize Mac Storage\u201d dataless files \u2014 `preflight`\u2018s `icloud_materialize` check handles this automatically, but a very large campaign can take a while to force-download everything on first run. |\n| `boltz` fails during `setup` with a `numpy` build error | You\u2019re likely on Python 3.13+. `boltz` pins `numpy<2.0`, which has no prebuilt wheel past cp312 \u2014 `_find_boltz_python()` already looks for a `python3.12` specifically; install one (`brew install python@3.12`) if it can\u2019t find one. |\n| A target fails preflight with a chain-id-length error | Boltz truncates chain IDs to 5 characters internally (a fixed-width field in its own schema) and silently corrupts longer ones rather than erroring at parse time \u2014 shorten the protein/partner/ligand name in `boltz_input.md`. |\n| The dashboard\u2019s charts (or the binding-site 3D view) don\u2019t render, or look unstyled | plotly.js and 3Dmol.js are both vendored and inlined (not CDN-loaded), so this shouldn\u2019t happen from a missing network connection \u2014 Google Fonts is still loaded from a CDN for styling, though, so the page needs internet access at least once for the fonts to look right (falls back to a generic sans-serif otherwise; charts, 3D views, and data are unaffected). If they genuinely don\u2019t render, check that `vendor/plotly-2.35.2.min.js` and `vendor/3Dmol-2.5.5-min.js` exist next to `BoltzMaker.py` \u2014 `analyze` prints a warning and falls back to the relevant CDN (which is known not to work in some HTML-preview contexts) if either is missing. |\n\n## \ud83c\udf10 Web deployment\n\n**Live at [boltzmaker.mdeller.com](https://boltzmaker.mdeller.com).** A Flask frontend for the non-GPU stages of the pipeline \u2014 upload a `boltz_input.md` (or build one with the wizard), run `generate`/`preflight` and download the results, or upload a completed campaign folder (from a `run` you did locally) for full `analyze`, PLIP interaction detection, and compare-sse, all in the browser. The GPU `run` step is deliberately never hosted: there\u2019s no GPU on the droplet, so that stage still runs on your own hardware and you bring the `boltz_output/` back for analysis.\n\n**Source and tested-CLI isolation.** The web app lives in `web/`, developed on a separate `web` git branch/worktree so the tested `BoltzMaker.py` on `main` is never touched by web-app work. `BoltzMaker.py` was never designed to be imported (it relaunches itself into a managed venv at module import time), so the Flask app only ever invokes it as a subprocess, with an explicit minimal environment and a per-command timeout, through a dedicated trimmed venv (`.venv/`, torch/boltz-free, ~500MB \u2014 every torch/boltz reference in the script is a function-local lazy import) kept separate from the Flask-serving venv (`web/.venv/`).\n\n**Optional PLIP.** If the droplet\u2019s `.plip_env` (the same conda-forge PyMOL/OpenBabel environment `setup-plip` builds locally, ~1-1.5GB) is present, `analyze` runs full protein-ligand interaction detection and compare-sse automatically; if not, both degrade gracefully rather than erroring, exactly as they do locally.\n\n**Serving stack:** gunicorn (3 sync workers, 900s timeout to cover PLIP rendering and compare-sse\u2019s GPCRdb/KLIFS/PDBe lookups) behind nginx (Let\u2019s Encrypt TLS via certbot, HTTP/2, rate-limited upload endpoints, 200MB body cap), as a hardened systemd service (dedicated unprivileged user, memory/CPU caps). Every request gets its own scratch directory, deleted when the request finishes and swept by a five-minute cleanup timer as a backstop; uploaded zips are checked for zip-slip and zip-bomb payloads (path-traversal/absolute-path rejection, uncompressed-size and entry-count caps) before anything is extracted.\n\n## \ud83e\uddeb Testing\n\n```\n.venv/bin/pip install -r requirements-dev.txt\n.venv/bin/pytest tests/\n```\n\n27 tests covering `compare-sse`\u2018s annotators against real fixture data (a real apo EGFR kinase-domain structure vs the `egfr_covalent` example\u2019s real holo prediction; a real apo beta2-adrenergic-receptor structure vs `adrb2_gs_panel`\u2018s real holo predictions), with GPCRdb/KLIFS/PDBe network calls swapped for an injectable fake client seeded with real, previously-verified API responses \u2014 fully offline and fast (~2s). Plus grammar and CLI-resolution tests for the parser fields above.\n\n## \ud83d\udcda Citation", "# Future Therapeutics\u2014Engineering Solutions for Microplastic Burden\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-15 on marcdeller.com (categories: blog, microplastics).*\n\n* Microplastics accumulate in diverse tissues, posing chronic health questions.\n* Biocatalytic therapy (enzyme delivery) may clear plastics within the body.\n* Encapsulation and cell-surface display increase stability, targeting, and safety.\n* Filtration and diagnostic technologies complement enzyme-based therapies.\n* Hybrid modalities promise personalized future treatments for microplastic health risks.\n\n1. [Current advances in PETase engineering (2023)](https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2023.1263996/full)\n2. [PETase degradation in serum (2024)](https://pubs.rsc.org/en/content/articlepdf/2024/fd/d4fd00014e)\n3. [Bioaccumulation of microplastics in human brains (2025)](https://www.nature.com/articles/s41591-024-03453-1)\n4. [Apheresis for microplastic removal (2025)](https://conexiant.com/internal-medicine/articles/apheresis-shows-promise-for-removing-microplastics-from-human-blood/)\n5. [Micro(nano)plastics health burden (2024)](https://pmc.ncbi.nlm.nih.gov/articles/PMC11528458/)\n\nThe discovery of microplastics in blood and tissues, including brain and breast milk, has sparked innovation in therapeutic approaches. Biocatalytic enzyme therapy\u2014delivering PETase and other enzymes\u2014offers promise for in situ breakdown of microplastics circulating in the body. Methods such as encapsulation and cell-surface display aim to stabilize enzymes, target specific tissues, and improve immune compatibility.\n\nTherapeutic apheresis and super-sensitive diagnostics provide immediate solutions for removing microplastics, while future enzyme therapies unlock the potential for complete biodegradation. Engineered forms of PETase, tailored for high activity and low toxicity, lead candidate strategies for medical intervention.\n\nHybrid models combining enzymatic treatments with filtration could be adapted to match the individual\u2019s burden and tissue distribution of microplastics. The key challenges remain delivery, stability, and patient safety, but the path is increasingly clear for clinical translation.", "# alphafraud.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n## alphafraud.mdeller.com/\n\n# Welcome to AlphaFraud\n\nAlphaFraud watches the PDB for freshly-deposited human structures and pits each against\nits *blind* AlphaFold prediction \u2014 a continuous, public test of whether protein folding is\nreally \u201csolved\u201d. Every week it surfaces the structures AlphaFold got most **confidently wrong**:\nsure of itself, and still missing.\n\n82,901\n\nstructures analysed\n\n1,948\n\nconfidently wrong\n\n2.3%\n\nconfidently-wrong rate\n\n431\n\nnovel + confidently wrong\n\n0.5%\n\nnovel + wrong rate\n\n459\n\nrelease weeks\n\n## Explore the site\n\n[\ud83e\uddec\n\n### All structures\n\nThe cumulative dashboard \u2014 every analysed human structure on the pLDDT-vs-agreement \u201cfraud quadrant\u201d, with metric histograms and a weekly trend.\n\n**\ud83d\udca1 You'll learn:** Spot the confidently-wrong cluster: high AlphaFold confidence, low real agreement.\n\nOpen All structures \u2192](/all)\n[\ud83c\udfc6\n\n### Worst offenders\n\nThe all-time leaderboard of AlphaFold\u2019s biggest misses, grouped by protein with a \u00d7N repeat count and a deviation-coloured ribbon each.\n\n**\ud83d\udca1 You'll learn:** Which proteins AlphaFold fails on again and again \u2014 amyloids top the list.\n\nOpen Worst offenders \u2192](/leaderboard)\n[\ud83d\udd2c\n\n### Examples\n\nHand-authored deep-dives plus an auto \u201cexample of the week\u201d, opening with a positive control where AlphaFold nailed a novel structure.\n\n**\ud83d\udca1 You'll learn:** Why AlphaFold gets some folds right and mangles others \u2014 explored in interactive 3D.\n\nOpen Examples \u2192](/examples)\n[\ud83e\uddea\n\n### Calculate\n\nType any human PDB ID and score it on demand \u2014 the full pipeline runs live and returns the same panel, with no AI-written text.\n\n**\ud83d\udca1 You'll learn:** Exactly how AlphaFold did on a structure you care about.\n\nOpen Calculate \u2192](/calculate)\n[\ud83d\udcc8\n\n### Analysis\n\nWhere the failures cluster: CATH/SCOP fold-family enrichment, sequence clustering and per-superfamily blind-spot scorecards.\n\n**\ud83d\udca1 You'll learn:** Which fold classes AlphaFold systematically struggles with.\n\nOpen Analysis \u2192](/analysis)\n[\ud83d\uddc4\ufe0f\n\n### Archive\n\nEvery PDB release week since the 2018 cutoff, browsable at weekly granularity and linked to each week\u2019s structures.\n\n**\ud83d\udca1 You'll learn:** Track the confidently-wrong rate week by week.\n\nOpen Archive \u2192](/archive)\n\n## alphafraud.mdeller.com/home\n\n# Welcome to AlphaFraud\n\nAlphaFraud watches the PDB for freshly-deposited human structures and pits each against\nits *blind* AlphaFold prediction \u2014 a continuous, public test of whether protein folding is\nreally \u201csolved\u201d. Every week it surfaces the structures AlphaFold got most **confidently wrong**:\nsure of itself, and still missing.\n\n82,901\n\nstructures analysed\n\n1,948\n\nconfidently wrong\n\n2.3%\n\nconfidently-wrong rate\n\n431\n\nnovel + confidently wrong\n\n0.5%\n\nnovel + wrong rate\n\n459\n\nrelease weeks\n\n## Explore the site\n\n[\ud83e\uddec\n\n### All structures\n\nThe cumulative dashboard \u2014 every analysed human structure on the pLDDT-vs-agreement \u201cfraud quadrant\u201d, with metric histograms and a weekly trend.\n\n**\ud83d\udca1 You'll learn:** Spot the confidently-wrong cluster: high AlphaFold confidence, low real agreement.\n\nOpen All structures \u2192](/all)\n[\ud83c\udfc6\n\n### Worst offenders\n\nThe all-time leaderboard of AlphaFold\u2019s biggest misses, grouped by protein with a \u00d7N repeat count and a deviation-coloured ribbon each.\n\n**\ud83d\udca1 You'll learn:** Which proteins AlphaFold fails on again and again \u2014 amyloids top the list.\n\nOpen Worst offenders \u2192](/leaderboard)\n[\ud83d\udd2c\n\n### Examples\n\nHand-authored deep-dives plus an auto \u201cexample of the week\u201d, opening with a positive control where AlphaFold nailed a novel structure.\n\n**\ud83d\udca1 You'll learn:** Why AlphaFold gets some folds right and mangles others \u2014 explored in interactive 3D.\n\nOpen Examples \u2192](/examples)\n[\ud83e\uddea\n\n### Calculate\n\nType any human PDB ID and score it on demand \u2014 the full pipeline runs live and returns the same panel, with no AI-written text.\n\n**\ud83d\udca1 You'll learn:** Exactly how AlphaFold did on a structure you care about.\n\nOpen Calculate \u2192](/calculate)\n[\ud83d\udcc8\n\n### Analysis\n\nWhere the failures cluster: CATH/SCOP fold-family enrichment, sequence clustering and per-superfamily blind-spot scorecards.\n\n**\ud83d\udca1 You'll learn:** Which fold classes AlphaFold systematically struggles with.\n\nOpen Analysis \u2192](/analysis)\n[\ud83d\uddc4\ufe0f\n\n### Archive\n\nEvery PDB release week since the 2018 cutoff, browsable at weekly granularity and linked to each week\u2019s structures.\n\n**\ud83d\udca1 You'll learn:** Track the confidently-wrong rate week by week.\n\nOpen Archive \u2192](/archive)\n\n## alphafraud.mdeller.com/stats\n\n# Statistics\n\nEverything AlphaFraud is tracking \u2014 the science, the pipeline, the server and the data\nsources \u2014 computed live on each load.\n\n82,901\n\nstructures analysed\n\n1,948\n\nconfidently wrong\n\n2.3%\n\nconfidently-wrong rate\n\n431\n\nnovel + confidently wrong\n\n0.894\n\nmean TM-score\n\n0.829\n\nmean lDDT\n\n### \ud83d\udd2c Catch & quality\n\nAnalysed\n\n[82,901](/all)\n\nCompared (tier-2)\n\n15,587\n\nScreened (tier-1)\n\n67,314\n\nConfidently wrong\n\n[1,948 \u00b7 2.3%](/leaderboard)\n\nNovel\n\n1,918 \u00b7 2.3%\n\nNovel+wrong\n\n431 \u00b7 0.5%\n\nProteins (UniProt)\n\n6,281\n\nMean TM\n\n0.894\n\nMean lDDT\n\n0.829\n\nMean pLDDT\n\n84.9\n\nSkipped\n\n13,916\n\nErrors\n\n123\n\n### \ud83d\udcc5 Coverage & schedule\n\nDeposit range\n\n2018-04-30\u21922026-08-16\n\nRelease range\n\n2018-05-16\u21922026-08-26\n\nRelease weeks\n\n[459](/archive)\n\nBackfill months\n\n117\n\nWeekly runs\n\n7\n\nAF cutoff\n\n2018-04-30\n\nLast update\n\n2026-08-26 00:30\n\nNext release\n\nWed 09 Sep 2026\n\n### \u2699\ufe0f Runs & compute\n\nTotal runs\n\n125\n\nWeekly/backfill/calc\n\n7/117/1\n\nLatest weekly\n\n2026-08-26\n\nRibbons\n\n15,558\n\nOn-demand calcs\n\n[5](/calculate)\n\nComputing now\n\n0\n\n### \ud83d\udda5\ufe0f Server\n\nVersion\n\nv0.1.0\n\nPython\n\n3.12.3\n\nCPU cores\n\n2\n\nLoad (1/5/15)\n\n0.00/0.02/0.00\n\nMemory (used/total)\n\n2.5 GB/3.8 GB\n\nWorker RSS\n\n185.7 MB\n\nUptime\n\n54d 23h\n\n### \ud83d\udcbe Storage & database\n\nDatabase (+WAL)\n\n6.1 GB\n\nRibbons\n\n15,558 \u00b7 4.1 GB\n\nStructure cache\n\n9.8 GB\n\nData total\n\n13.9 GB\n\nDisk (used/total)\n\n60.3 GB/76.4 GB\n\nDisk free\n\n16.1 GB\n\nentities\n\n96,945\n\nentity_blobs\n\n96,945\n\nentity_annotations\n\n15,592\n\nweekly_examples\n\n459\n\nruns/visits\n\n125/145\n\n### \ud83d\udd0c Data sources & APIs\n\nRCSB Search\n\n[search.rcsb.org](https://search.rcsb.org)\n\nRCSB GraphQL\n\n[data.rcsb.org](https://data.rcsb.org/graphql/index.html)\n\nRCSB files\n\n[files.rcsb.org](https://www.rcsb.org)\n\nAlphaFold DB\n\n[alphafold.ebi.ac.uk](https://alphafold.ebi.ac.uk)\n\nCalculate index\n\n[134,212](/calculate)\n\nAF checks cached\n\n6,533\n\nSource code\n\n[bellcheddar/AlphaFraud](https://github.com/bellcheddar/AlphaFraud)\n\n### \ud83d\udc65 Traffic & coverage\n\nVisitors\n\n145\n\nPage views\n\n5,107\n\nToday\n\n2\n\nTotal processed\n\n96,945\n\nCoverage\n\n85.5%", "# AlphaFraud \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n| Module | Responsibility |\n| --- | --- |\n| `pdb.py` | RCSB Search (new human protein entities) + Data GraphQL (UniProt mapping, dates, method) |\n| `afdb.py` | AlphaFold DB fetch: the blind model + PAE, choosing the fragment covering the entity\u2019s span |\n| `novelty.py` | max % identity to any pre-cutoff chain (so genuinely unseen sequences are flagged) |\n| `compare.py` | the full metric suite on the aligned span (TM, lDDT, GDT, RMSD, per-residue deviation, PAE honesty) |\n| `ribbon.py` | C\u03b1-deviation SVG + 3D-viewer coords + the AlphaFold \u201cghost\u201d (vendored, offline) |\n| `annotate.py` / `analysis.py` | CATH / SCOP / citation enrichment and the `/analysis` snapshot (enrichment, clustering, PCA) |\n| `examples_auto.py` | the deterministic, no-LLM panel builder (match **or** failure framing) |\n| `calc_index.py` / `calculate.py` | the Calculate tab\u2019s strict autocomplete index + on-demand single-entity compute |\n| `db.py` / `report.py` / `webapp.py` | SQLite persistence, Plotly figure JSON, and the Flask routes |\n\n**The weekly pipeline** (`AlphaFraud.py run`, one entity at a time):\n\n```\nfor entity in RCSB.search(human, protein, deposited_after = 2018-04-30):\n    if entity in db: continue\n    meta = RCSB.graphql(entity)                       # UniProt, dates, method, chains\n    if not meta.single_uniprot: skip(\"antibody / fusion / chimera\"); continue\n    af = AlphaFoldDB.model(meta.uniprot)              # the blind prediction (fragment covering the span)\n    if af is None:               skip(\"no AlphaFold model\"); continue\n    exp, model = download(entity), download(af)\n    m        = compare(exp, model)                    # TM, lDDT, GDT, RMSD, per-residue dev, PAE honesty\n    novelty  = max_identity_to_pre_cutoff(meta.sequence)\n    fraud    = mean_i( pLDDT_i/100 * clamp(dev_i, 0, 15\u00c5) / 15 )   # confidence-weighted error\n    confidently_wrong = (mean_pLDDT > 70) and (TM < 0.5)          # sure, and still wrong\n    ribbon.render(exp, model)                         # deviation SVG + coords + AF ghost\n    db.store(entity, m, novelty, fraud, confidently_wrong)\n\nanalysis.rebuild()            # /analysis enrichment snapshot\nrebuild_weekly_examples()     # auto 'example of the week' + weekly archive\ncalc_index.refresh(metas)     # keep the Calculate autocomplete index current\n```\n\n**The always-on web app** (gunicorn) reads the same SQLite and renders cached Plotly JSON:\n\n```\nGET /            -> cumulative \"fraud quadrant\" dashboard (figures cached, short TTL)\nGET /entry/<id>  -> per-structure heatmaps, calibration, per-domain metrics\nGET /examples    -> positive control + auto example-of-the-week + curated deep-dives\nGET /calculate   -> type a human PDB id  ->  (reuse if known | compute on demand)  ->  panel\n```\n\n**The Calculate tab** reuses the very same `process_entity()` as the pipeline, run in an isolated subprocess so a heavy computation never blocks a web worker, and stored apart so it never pollutes the science:\n\n```\nPOST /calculate/run { pdb }:\n    entity = resolve_human_entity(pdb)                # strict index, else a live RCSB check\n    if not qualifies:  return reason                 # not human / no single UniProt / no AF model\n    if entity already computed:  return \"ready\"      # cached -> instant\n    mark(entity, \"computing\"); spawn AlphaFraud.py calculate <entity>   # same pipeline, subprocess\n    return \"computing\"                               # the browser polls, then injects the panel\n```\n\n## \ud83c\udfaf Metric suite\n\nEvery matched structure is scored on the aligned residue span (all pure-Python except CATH, which is a lookup):\n\n| Family | Metrics |\n| --- | --- |\n| Global fold agreement | TM-score (4 normalisations), C\u03b1 / backbone / all-atom / core RMSD, GDT_TS, GDT_HA, MaxSub, structural overlap |\n| Local, superposition-free | lDDT (global and per-residue), contact-map Jaccard / precision / recall, distance-matrix difference, CAD-score |\n| Backbone and secondary structure | Q3 secondary-structure agreement, mean \u0394\u03c6 / \u0394\u03c8, radius-of-gyration difference |\n| Domains | Per-CATH-domain and per-PAE-cluster-domain TM and RMSD (exposes inter-domain hinge errors a global fit hides) |\n| Confidence calibration | pLDDT vs lDDT correlation, PAE vs observed-error correlation, overconfident-pair fraction |\n| Composite | FRAUD score (confidence-weighted error) and a confidently-wrong flag (mean pLDDT > 70 yet TM-score < 0.5) |\n\n## \ud83d\udccb Requirements\n\n* Python 3.11 or newer\n* The packages in `requirements.txt` (requests, tenacity, biotite, tmtools, numpy, scipy, pandas, plotly, flask, gunicorn, python-dotenv), all pip-installable with no compiled toolchain\n* Outbound HTTPS to `search.rcsb.org`, `data.rcsb.org`, `files.rcsb.org` and `alphafoldebi.ac.uk`\n* For deployment: a Linux droplet with root/SSH access, a domain, and DNS pointed at it\n\n## \ud83d\udd27 Installation\n\n```\ngit clone https://github.com/bellcheddar/AlphaFraud.git\ncd AlphaFraud\npython3 -m venv .venv\nsource .venv/bin/activate\npip install -r requirements.txt\npython3 AlphaFraud.py init          # create dirs, DB schema, vendor plotly.min.js\n```\n\n## \ud83d\ude80 Usage\n\n```\n# Preview discovery for a window without writing anything\npython3 AlphaFraud.py run --since 2024-01-01 --until 2024-01-10 --limit 6 --dry-run\n\n# Process a window for real (omit --since to resume from the last run)\npython3 AlphaFraud.py run --since 2024-01-01 --limit 50\n\n# Backfill a historical range (full metrics on everything)\npython3 AlphaFraud.py backfill --from 2023-01-01 --to 2023-12-31\n\n# Two-tier backfill: fast TM-score screen on everything, full metrics only on the\n# disagreements (TM below --tm-threshold). Far cheaper on time, disk and API load.\npython3 AlphaFraud.py backfill --from 2023-01-01 --to 2023-12-31 --two-tier --tm-threshold 0.7\n\n# Two-tier backfill of the ENTIRE post-cutoff archive (~96k entities, monthly chunks,\n# resumable). Screens all, fully analyses only the tail AlphaFold got wrong.\npython3 AlphaFraud.py backfill --all --two-tier", "# Unlocking the Power of Protein Engineering: FDA-Approved Enzyme Therapies Transform Rare Disease Treatment\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nPEGylation remains the most clinically established half-life extension technology, with [proven safety in humans for over two decades](https://pmc.ncbi.nlm.nih.gov/articles/PMC4663500/). [Pegademase bovine (Adagen), approved in 1990](https://pmc.ncbi.nlm.nih.gov/articles/PMC7049870/), pioneered PEGylated enzyme replacement for adenosine deaminase deficiency, extending enzyme half-life from hours to days through steric hindrance that reduces proteolytic degradation and renal clearance. The evolution to [elapegademase (Revcovi) in 2018](https://pmc.ncbi.nlm.nih.gov/articles/PMC7049870/) incorporated improved linker chemistry with succinimidyl carbonate replacing succinimidyl succinate, reducing spontaneous de-PEGylation and extending circulatory half-life by 4 to 15 hours in preclinical models. Recent innovations include [pegunigalsidase alfa (Elfabrio), approved in 2023](https://practicalneurology.com/articles/2025-oct-27/new-enzyme-replacement-therapy-approved-fda-treat), which combines PEGylation with covalent crosslinking of alpha-galactosidase A to achieve sustained plasma concentrations and reduced immunogenicity for Fabry disease treatment.\n\nDelivery method optimization addresses the fundamental challenge of achieving therapeutic enzyme concentrations at target sites while minimizing systemic exposure and immunogenic responses. The predominant intravenous infusion route enables [100% bioavailability and rapid onset of action](https://towardshealthcare.com/insights/enzyme-replacement-therapy-market-size) but requires frequent hospital visits and specialized administration protocols. [Laronidase (Aldurazyme) demonstrates M6P-mediated cellular uptake](https://pro.campus.sanofi/en/aldurazyme-mechanism-of-action), where mannose-6-phosphate-terminated oligosaccharide chains bind specific receptors on cell surfaces, triggering receptor-mediated endocytosis and lysosomal trafficking. Alternative approaches include intramuscular delivery for [pegademase and elapegademase in ADA-SCID](https://pmc.ncbi.nlm.nih.gov/articles/PMC7049870/), and subcutaneous administration for [asfotase alfa (Strensiq)](https://pubs.rsc.org/en/content/articlelanding/2023/md/d3md00481c), an Fc-fusion protein where the fragment crystallizable region extends circulation time through neonatal Fc receptor recycling.\n\nTargeting strategies enable selective enzyme delivery to affected tissues while reducing off-target accumulation and potential cytotoxicity. For lysosomal storage diseases, the [M6P receptor pathway provides natural tropism](https://pmc.ncbi.nlm.nih.gov/articles/PMC9899006/) for hepatocytes and other tissues expressing cation-independent mannose-6-phosphate receptors. Gaucher disease therapies exploit macrophage-specific uptake through [mannose-terminal glycosylation that engages macrophage mannose receptors](https://accessdata.fda.gov/drugsatfda_docs/label/2002/ceredz061302LB.pdf), achieving preferential delivery to disease-relevant cell populations. The [glycoengineered Long-Acting-GlycoDesign (LAGD) platform](https://pmc.ncbi.nlm.nih.gov/articles/PMC9899006/) represents an alternative paradigm, eliminating M6P to produce sialylated complex N-glycans that evade receptor-mediated clearance, extending circulatory stability for enzymes including alpha-galactosidase A, beta-glucuronidase, cathepsin D, and acid alpha-glucosidase.\n\nStability enhancement through protein engineering addresses the inherent thermodynamic constraints that limit enzyme therapeutic efficacy. Recombinant production in Chinese hamster ovary (CHO) cells enables precise control over post-translational modifications while [maintaining native human amino acid sequences to minimize immunogenicity](https://pmc.ncbi.nlm.nih.gov/articles/PMC2745820/). Studies on [TEM beta-lactamase mutations demonstrate](https://pmc.ncbi.nlm.nih.gov/articles/PMC8844078/) how single amino acid substitutions like M182T can increase fitness by reducing proteolysis without significantly altering catalytic activity, illustrating the potential for rational stabilization strategies in therapeutic enzymes. Chemical modification approaches include [periodate oxidation followed by borohydride reduction](https://pmc.ncbi.nlm.nih.gov/articles/PMC9899006/) to eliminate glycan-binding interactions for beta-glucuronidase, demonstrating how post-production modifications can fundamentally alter enzyme pharmacology.\n\nThe cytotoxic enzyme class employs fundamentally different mechanisms, depleting essential metabolites or activating prodrugs specifically within tumor microenvironments. [L-asparaginase and its PEGylated derivatives](https://pmc.ncbi.nlm.nih.gov/articles/PMC8714691/) catalyze asparagine depletion, exploiting the auxotrophy of L-asparagine synthetase-deficient leukemias and lymphomas to achieve remission rates exceeding 93% when combined with vincristine and prednisone. [Rasburicase (Elitek), approved in 2002](https://go.drugbank.com/drugs/DB00049), provides recombinant urate oxidase for tumor lysis syndrome, catalyzing uric acid oxidation to soluble allantoin and preventing hyperuricemia-associated renal failure in patients receiving chemotherapy. Evolution from bacterial extracts to recombinant production in *Saccharomyces cerevisiae* eliminates batch-to-batch variability while reducing immunogenic contaminants, though anti-drug antibody formation remains a clinical challenge requiring ongoing monitoring.", "# 2025 \u2013 Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Key Concept | Description | Key Reference |\n| --- | --- | --- |\n| KRAS G12D Prevalence | 29% of KRAS mutations; 40.2% in pancreatic cancer | <https://pmc.ncbi.nlm.nih.gov/articles/PMC9861148/> |\n| Structure-Based Drug Design | Crystal structures guided hit-to-lead optimization | <https://pubs.acs.org/doi/abs/10.1021/acs.jmedchem.4c02662> |\n| Bioavailability Challenges | Aspartate\u2019s low nucleophilicity complicates design | <https://pubs.acs.org/doi/abs/10.1021/acs.jmedchem.4c02662> |\n| Molecular Rigidity | Reduced flexibility improves permeability | <https://pubmed.ncbi.nlm.nih.gov/26336864/> |\n| Balanced Polarity | Polarity crucial for potency and absorption | <https://pubs.acs.org/doi/abs/10.1021/acs.jmedchem.4c02662> |\n| Primate Exposure | Oral exposure in nonhuman primates | <https://pubs.acs.org/doi/abs/10.1021/acs.jmedchem.4c02662> |\n| Switch II Pocket | Allosteric pocket induced upon binding | <https://www.nature.com/articles/s41421-021-00368-w> |\n| ADME Properties | Key ADME parameters for drug success | https://elearning.uniroma1.it/\u2026Drug%20Candidates.pdf |", "# The Memorization Trap: Have Protein-Ligand Cofolding Methods Transcended Simple Data Recall?\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe memorization phenomenon manifests most clearly in binding site mutagenesis experiments, where systematic alterations to binding pocket residues reveal the models\u2019 inability to adapt to chemical and spatial perturbations. Even dramatic mutations that completely alter pocket properties result in predicted ligand poses that remain biased toward original binding configurations, often producing physically implausible structures with significant atomic clashes. This behavior indicates that current architectures lack sufficient understanding of fundamental physical constraints governing molecular interactions.\n\nSpecialized approaches have attempted to address these limitations through various architectural innovations. DynamicBind employs equivariant geometric diffusion networks to model large-scale protein conformational changes during ligand binding, explicitly accounting for the dynamic nature of protein-ligand recognition. NeuralPLexer utilizes multi-scale generative modeling to sample both ligand-free and ligand-bound conformational ensembles, providing state-specific predictions that better capture induced-fit effects. Despite these advances, fundamental challenges persist in achieving robust generalization beyond training data distributions.\n\nThe implications of these findings extend beyond technical considerations to practical applications in drug discovery. Current cofolding methods show promise for optimizing known chemical scaffolds and exploring chemical space adjacent to well-characterized binding interactions. However, their utility for truly de novo drug design\u2014identifying completely novel chemotypes or targeting cryptic binding sites\u2014remains limited by their dependence on memorized structural patterns. This limitation is particularly problematic for targeting emerging therapeutic areas, such as protein-protein interaction modulators or allosteric sites, where training data remains sparse.\n\nRecent efforts to expand beyond memorization have focused on incorporating physical constraints and energy-based modeling into deep learning architectures. Some approaches integrate molecular dynamics simulations during training to expose models to conformational diversity not captured in static crystal structures. Others employ physics-informed neural networks that explicitly encode chemical bonding rules and stereochemical constraints, potentially improving generalization to novel chemical space. However, these hybrid approaches face computational scalability challenges and require careful balancing of data-driven learning with physical modeling.\n\nThe field\u2019s trajectory suggests that overcoming memorization limitations will require fundamental advances in model architecture and training methodologies. Promising directions include few-shot learning approaches that can rapidly adapt to new chemical environments, self-supervised learning methods that extract physical principles from unlabeled structural data, and multi-modal training that integrates diverse experimental observables beyond static structures. Additionally, expanding training datasets through high-quality molecular dynamics simulations and experimental techniques like cryo-electron microscopy may provide the conformational diversity necessary for more robust learning.\n\nCurrent protein-ligand cofolding methods represent remarkable achievements in computational biology, demonstrating unprecedented accuracy for structure prediction tasks within their training domains. However, their reliance on memorized patterns rather than fundamental understanding of molecular recognition principles limits their transformative potential for drug discovery applications requiring true predictive capabilities beyond known chemical space.\n\n| Key Concept | Description | Key References |\n| --- | --- | --- |\n| Cofolding Methods | Deep learning approaches that simultaneously predict protein structure and ligand binding from sequence information | [Abramson, J., et al., Nature](https://www.nature.com/articles/s41586-024-07487-w) |\n| Memorization Problem | Current methods largely recall training data patterns rather than learning underlying physical principles | [\u0160krinjar, P., et al., bioRxiv](https://www.biorxiv.org/content/10.1101/2025.02.03.636309v3.full.pdf) |\n| Runs N\u2019 Poses Dataset | Benchmark dataset of 2,600 high-resolution protein-ligand systems for unbiased evaluation | [\u0160krinjar, P., et al., bioRxiv](https://www.biorxiv.org/content/10.1101/2025.02.03.636309v1) |\n| Diffusion Architecture | Advanced neural network architecture using iterative denoising processes for structure generation | [Lu, W., et al., Nature Communications](https://pmc.ncbi.nlm.nih.gov/articles/PMC10844226/) |\n| Generalization Challenge | Performance degradation when models encounter protein-ligand combinations outside training distribution | [Current Opinion in Structural Biology](https://pmc.ncbi.nlm.nih.gov/articles/PMC12406700/) |\n| Induced-Fit Effects | Conformational changes in proteins upon ligand binding, crucial for accurate interaction modeling | [Qiao, Z., et al., Nature Machine Intelligence](https://arxiv.org/abs/2209.15171) |", "# Protein stability: a crystallographer's perspective\nAuthors: Deller, M. C.; Kong, L.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 72, 72\u201395 (IYCr crystallization series), 2016\n\n**Authors:** Marc C. Deller (Stanford ChEM-H, Macromolecular Structure Knowledge Center, Stanford University, Stanford, California, USA; corresponding author, mdeller@stanford.edu), Leopold Kong (Laboratory of Cell and Molecular Biology, NIDDK, National Institutes of Health, Bethesda, Maryland, USA) and Bernhard Rupp (Department of Forensic Crystallography, k.-k. Hofkristallamt, Vista, California, USA, and Department of Genetic Epidemiology, Medical University of Innsbruck, Sch\u00f6pfstrasse 41, A-6020 Innsbruck, Austria).\n\n**Published:** *Acta Crystallographica Section F* **72**, 72\u201395 (2016); part of the IYCr crystallization series. Received 27 November 2015; accepted 21 December 2015; edited by H. M. Einspahr.\n\n*This document is a structured summary of the review, not the full text.*\n\n## Abstract\n\nProtein stability is a topic of major interest for the biotechnology, pharmaceutical and food industries, in addition to being a daily consideration for academic researchers studying proteins. An understanding of protein stability is essential for optimizing the expression, purification, formulation, storage and structural studies of proteins. In this review, discussion will focus on factors affecting protein stability, on a somewhat practical level, particularly from the view of a protein crystallographer. The differences between protein conformational stability and protein compositional stability will be discussed, along with a brief introduction to key methods useful for analyzing protein stability. Finally, tactics for addressing protein-stability issues during protein expression, purification and crystallization will be discussed.\n\n## Section-by-section synopsis\n\n### 1. Introduction\n\nThe review introduces protein stability from the viewpoint of a structural biologist \u2014 defined as a scientist who determines detailed molecular structure by crystallography, NMR spectroscopy or cryo-EM \u2014 with particular emphasis on crystallographic techniques, because instability is a substantial obstacle to crystallizing many proteins. The authors deliberately skip most computational and theoretical treatments of folding thermodynamics, referring readers to reviews by Compiani & Capriotti and by Lazaridis & Karplus. Stability is framed, after Schr\u00f6dinger, as \"the potential of a pattern to survive over time\"; for a protein the \"pattern\" is transformed by unfolding, denaturation, degradation, conformational change, enzymatic modification and proteolytic cleavage. The authors stress that stability means different things to different practitioners: a pharmacologist, biotechnologist or food scientist typically means the half-life of activity, whereas a protein chemist or structural biologist means the integrity of primary, secondary, tertiary and quaternary structure. The roadmap of the review is laid out: stability as a prerequisite for crystallization (\u00a72), stability at each structural level (\u00a73), expression and purification (\u00a74), analytical methods including melting-temperature analysis, NMR and cryo-EM (\u00a75), and crystallizability-enhancing tactics (\u00a76).\n\n### 2. Stability is a fundamental prerequisite for crystallization\n\nBiomolecular crystallization is described as the self-organization of macromolecules into a translationally periodic arrangement with long-range order; for this to work optimally the moieties within each asymmetric unit must be of the same kind and the same shape. If a protein cannot form such stable entities, a primary requirement for crystallization is simply not met, and no amount of searching for thermodynamic and kinetic conditions will yield crystals of that construct. The authors introduce the review's central conceptual distinction: two \"flavors\" of stability, compositional and conformational, both of which the crystallographer must assess (Table 1). Thermodynamics establishes the necessary conditions for crystallization, while kinetics and dynamics decide whether a possible scenario actually becomes reality; only when all parameters are satisfied will crystal formation proceed (Figure 1a, adapted from Rupp, 2015). Figure 1b illustrates the marginal net stability of a folded protein: the overwhelming loss of conformational entropy is only marginally balanced by van der Waals forces, hydrogen bonding and the hydrophobic effect.\n\n### 2.1. Compositional stability\n\nCompositional stability means that the same chemical species must persist throughout the crystallization experiment: the protein molecules must have the same chemical makeup. Chemical homogeneity is commonly assessed by mass spectrometry or SDS\u2013PAGE. Compositional homogeneity is typically compromised by post-translational modifications, especially glycosylation and proteolysis, which alter primary structure and generate variability. Because crystallization takes time, compositional stability must be maintained over an extended period and preferably across a reasonable range of environmental conditions. The authors emphasize that there is no such thing as absolute stability: a protein compositionally stable enough to give a single band on an SDS\u2013PAGE gel may still be insufficiently stable over the timeframe of a crystallization experiment. Where the answer to the key questions in Table 1 is \"no\", the recommended remedies are to check the amino-acid sequence, screen for PTMs (especially proteolysis), purify further, apply more rigorous bioanalytical methods such as mass spectrometry and light scattering, use customized or less harsh crystallization screens, and explore different screening temperatures.\n\n### 2.2. Conformational stability", "# GitHub README: SiteHealthInspector\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/SiteHealthInspector, 2026\n\nThe default scan target is `https://marcdeller.com`. Change it by passing a URL argument or by editing the fallback in line 4:\n\n```bash\nTARGET=\"${1:-https://your-default-site.com}\"\n```\n\n---\n\n## \ud83d\udcc4 Licence\n\nMIT \u2014 use freely, attribution appreciated.\n\n---\n\n*Built as part of the [marcdeller.com](https://marcdeller.com) toolchain.*\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/SiteHealthInspector\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/SiteHealthInspector</a></td>\n</tr>\n</table>", "# ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n* \u2705 Config: `config/train_config_r5.yaml` \u2014 rank=64, scale=90, iters=3000, max_seq=3072, val_batches=50\n* \u2705 Launch: single-process `caffeinate -i .venv/bin/python /tmp/r5_launch.py` (cache 16.7 GB, memory limit 50.1 GB)\n* \u2705 **Training started 2026-06-28** \u2014 Iter 1 val loss 2.210\n* \u2705 No rank=64 instability \u2014 converged normally\n* \u2705 **Early stop at iter 2000** \u2014 val loss plateaued at 0.055 from iter 1600 (best: 0.054 at iters 1650 and 1950, unsaved)\n* \u2705 Best available checkpoint: iter 1600, val loss **0.055** \u2192 copied to `adapters.safetensors`\n* \u2705 Fused into `models/chem_sage_32b_v5/` (~17.2 GB)\n* \u2705 Run `eval/compare/eval_compare.py` after fuse\n\n### Round 4 dataset (complete)\n\n* \u2705 8,000 examples across 59 behaviour classes (`build_dataset.py --n 8000 --seed 47`)\n* \u2705 Protein family coverage: kinases (7,687 PDB), antibodies (3,322), GPCRs (103), GTPases (359), cytokines (769), receptors (6,891) \u2014 real PDB IDs and ligand SMILES from corpus\n* \u2705 Numerical fidelity drills: `code_then_quote` enforces \u201ccompute then quote\u201d pattern; `rounding_precision` teaches MW 1dp, logP 2dp, TPSA 1dp\n* \u2705 Structural biology: PyMOL selection algebra, visualization recipes, PLIP interaction analysis, PDB format, pdb-tools, Biopython Bio.PDB, ChimeraX, gemmi, structure preparation, docking interpretation\n* \u2705 Visualization: Plotly Python (scatter, radar), R/ggplot2 (SAR plots), bio3d (RMSD, RMSF)\n* \u2705 Set `steps_per_report: 25` to align train/val reporting at same iters\n\n### \ud83c\udf93 Round 4 training (complete)\n\n* \u2705 `num_layers: 32` \u2014 48 and 64 layers tested and found too slow (~2 min/iter); 32 is the M1 Max sweet spot\n* \u2705 Adapter path: `adapters/chem_sage_32b_v4_lora` (checkpoints saved every 200 iters)\n* \u2705 Training completed 2026-06-27 10:59 BST \u2014 1,500 iters, ~43 h\n* \u2705 Best checkpoint: iter 950, val loss **0.041** (also tied at iter 1425)\n* \u2705 Copied `0000950_adapters.safetensors` \u2192 `adapters.safetensors`\n* \u2705 Fused into `models/chem_sage_32b_v4/` (~17.2 GB)\n* \u2705 Updated `eval/compare/models.yaml` with final stats and complete val loss curve\n* \u2705 Raised rank to 64 in Round 5 for more adapter expressivity\n\n### \ud83e\uddea Evaluation and model quality\n\n* \u2705 Complete Round 3 eval: SMILES 99%, Exec 99%, Fidelity 68%\n* \u2705 Comparative eval harness: `eval/compare/eval_compare.py` with 13 metrics, loss curves, resource stats\n* \u2705 Comparative eval run after R4 fuse: full 4-round report `eval/compare/results/compare_20260627_1753.html`\n* \u2705 Run R5 comparative eval after fuse: add Round 5 to `eval/compare/models.yaml` and run `eval_compare.py`\n* \u2705 Improve exec/PyExec from R4\u2019s 74%/75% to >85% (R5 result: exec 95%, PyExec 99%)\n* \u2b1c Add perplexity benchmark against base (unfine-tuned) 32B\n* \u2b1c Add hallucination rate metric: count invented PDB IDs, ChEMBL IDs, SMILES not in corpus\n\n### Infrastructure\n\n* \u2705 ChromaDB 1.x fix: `allow_reset=True` in `rag/retrieve.py` and `scripts/ingest_rag.py` \u2014 55,742 chunks preserved, no re-ingest required\n* \u2705 `psutil` added to `requirements.txt` for server process monitoring in comparative eval\n* \u2705 Set up HuggingFace Hub: all 5 models published at [huggingface.co/Dellboy](https://huggingface.co/Dellboy)\n* \u2b1c Create `scripts/run_eval.sh` (start server, wait, run eval, stop server)\n\n### Publishing and deployment\n\n* \u2705 **Live web deployment at chemsage.mdeller.com** \u2014 shipped as an xterm.js terminal in the browser, not a chat-bubble Gradio UI: the real, unmodified `scripts/chat.py` runs server-side in a PTY, bridged to the browser over a WebSocket, so the hosted session is identical to running the CLI locally (same slash commands, same Rich rendering). No local GPU on the droplet, so `mlx_lm`\u2018s `generate`/`stream_generate` are patched at import time to call a HuggingFace Space running the real fused model on ZeroGPU instead \u2014 see **Web deployment** below.\n* \u2b1c **Packaged CLI installer** \u2014 publish `chat.py` as a `pipx`-installable tool (`pyproject.toml` entry point); users with a local Qwen 32B model can `pipx install chemsage` and point it at their own `mlx_lm.server` endpoint.\n\n### Round 6 (planned)\n\nIdeas to try next, based on learnings from Rounds 1\u20135:", "# Protein Hunter: Pioneering the Next Frontier of AI-Driven Protein Design\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Key Concept | Description | Key References |\n| --- | --- | --- |\n| Structure Hallucination | Neural networks trained on real protein structures generate plausible structural hypotheses from artificial or incomplete sequences | [Anishchenko et al., Nature](https://www.nature.com/articles/s41586-021-04184-w) |\n| Diffusion Models | Generative AI approach that creates protein structures through iterative denoising processes from random states | [Watson et al., Nature](https://www.nature.com/articles/s41586-023-06415-8) |\n| Fine-tuning-free Design | Protein design approach that doesn\u2019t require model retraining or specialized datasets for different design tasks | [Wu et al., bioRxiv](https://www.biorxiv.org/content/10.1101/2025.10.10.681530v1) |\n| All-X Sequence | Starting point using undefined amino acid sequences (X represents any amino acid) for structure prediction | [Wu et al., bioRxiv](https://www.biorxiv.org/content/10.1101/2025.10.10.681530v2.full.pdf) |\n| Multi-target Binding | Capability to design proteins that interact with diverse molecular targets including nucleic acids and small molecules | [Abramson et al., Nature](https://www.nature.com/articles/s41586-024-07487-w) |\n| Iterative Refinement | Sequential cycles of sequence redesign and structure re-prediction to optimize protein designs | [Bennett et al., Science](https://www.science.org/doi/10.1126/science.add1964) |", "# GitHub README: AlphaFraud\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/AlphaFraud, 2026\n\n# --workers N adds a download thread pool (network-bound work parallelises well). Single-\n# threaded is the reliable default for long unattended runs; workers give a partial speedup\n# but are more sensitive to the droplet's core count.\npython3 AlphaFraud.py backfill --all --two-tier --workers 6\n\n# Database summary + worst-offenders leaderboard\npython3 AlphaFraud.py status\n\n# Local dev web server (production uses gunicorn)\npython3 AlphaFraud.py serve --port 8000\n```\n\n| Command | Purpose |\n|---|---|\n| `init` | Create directories, the SQLite schema, and vendor `plotly.min.js` |\n| `run` | Discover, match, score and store one deposit window (`--since`, `--until`, `--limit`, `--dry-run`) |\n| `backfill` | Process a historical window (`--from`, `--to`); `--two-tier` screens by TM-score and fully analyses only disagreements below `--tm-threshold`; `--all` covers the entire post-cutoff archive in resumable monthly chunks |\n| `status` | Print a database summary and the current leaderboard |\n| `serve` | Run the Flask app locally (`--port`, `--debug`) |\n\n## \ud83d\udcca Output\n\nThe web app serves live from SQLite. Every plot carries a plain-language explanatory panel (what the axes mean, the cutoffs, a take-home summary, and a per-metric breakdown giving both a lay explanation and the equation, e.g. `TM = (1/L) \u03a3\u1d62 1/(1 + (d\u1d62/d\u2080)\u00b2)`) plus a CSV export; every table exports to CSV too.\n\nThe site is live at **[alphafraud.mdeller.com](https://alphafraud.mdeller.com/)**.\n\n| Route | Shows |\n|---|---|\n| [`/`](https://alphafraud.mdeller.com/) | Home \u2014 a concise navigation page: the project's purpose, headline KPIs, and a colour-coded card per tab with a \"what you'll learn\" line |\n| [`/all`](https://alphafraud.mdeller.com/all) | Cumulative \"All structures\" dashboard across every processed week: KPIs, the pLDDT-vs-TM fraud-quadrant scatter, metric histograms, the weekly trend, and a top-ranked table. A pulldown jumps to any single week |\n| `/week/<label>` | One release week, same layout scoped to that batch |\n| `/entry/<id>` | Per-structure detail: per-residue error tracks, calibration scatter, distance-matrix and PAE-vs-observed heatmaps, per-domain breakdown, and every metric |\n| [`/leaderboard`](https://alphafraud.mdeller.com/leaderboard) | The all-time worst AlphaFold failures across every processed week |\n| [`/analysis`](https://alphafraud.mdeller.com/analysis) | Structural deep dive over the worst offenders: fold/family enrichment (Wilson-CI + Fisher), sequence-similarity clustering, failure-mode PCA, theme flags, per-superfamily blind-spot scorecards, conformational-heterogeneity detection, a \"new this week\" ribbon, and RCSB + DOI links on every structure |\n| [`/examples`](https://alphafraud.mdeller.com/examples) | Curated storytelling layer: a positive control (an excellent AlphaFold match), an auto-generated \"example of the week\" plus a full weekly-release archive, and hand-authored deep-dive panels with interactive viewers (see below) |\n| [`/calculate`](https://alphafraud.mdeller.com/calculate) | Type any human PDB ID and get the same panel computed on demand: strict autocomplete (exists + human + has an AlphaFold model), full pipeline in an isolated subprocess, no LLM (see below) |\n| [`/archive`](https://alphafraud.mdeller.com/archive) | Every PDB release week since the 2018 cutoff, at weekly granularity, each linking to that week's structures and confidently-wrong count |\n| [`/stats`](https://alphafraud.mdeller.com/stats) | An exhaustive live Statistics page: themed panels for catch/quality, coverage & schedule, runs & compute, server (CPU/memory/uptime), storage & database, data sources & APIs, and traffic |\n| `/api/week/<label>`, `/api/leaderboard`, `/api/entry/<id>` | JSON for external tools |\n\n## \ud83d\udd2c Examples\n\nThe **[Examples](https://alphafraud.mdeller.com/examples)** tab turns raw catches into readable stories and deliberately shows both sides of AlphaFold's record, not just the failures:", "# GitHub README: microplastic-blaster\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/microplastic-blaster, 2026\n\n> **Deploy Enzymes. Destroy Microplastics. Save the Body.**\n\nA browser-based arcade game by [Elora Therapeutics](https://www.eloratherapeutics.com) \u2014 the biotech developing first-in-class enzyme therapy to remove microplastics from the human body. Fire enzymatic blasts at incoming polymer particles, protect the cell, and climb the global leaderboard.\n\n    \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/microplastic-blaster\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/microplastic-blaster</a></td>\n</tr>\n</table>\n\n---\n\n## \ud83c\udfae Gameplay\n\nYou are the body's enzymatic defence. Microplastic particles \u2014 PET, PE, PS, PP, plus branded consumer-plastic enemies \u2014 descend toward a target cell at the bottom of the screen. Click (or tap) to launch enzyme projectiles and intercept them before they reach the cell membrane.\n\n- Each microplastic that breaches the membrane deals health damage proportional to the current level.\n- The game ends when cell health reaches zero \u2014 the \"Cell Overwhelmed\" screen appears.\n- Difficulty increases every 720 game ticks; spawn rate accelerates with level.\n- Score = particle point value \u00d7 current level.\n\n---\n\n## \u2697\ufe0f Microplastic Types\n\n| Polymer | Abbreviation | Colour | HP | Speed | Points | Shape |\n|---|---|---|---|---|---|---|\n| Polyethylene terephthalate | PET | Purple | 3 | Slow | 30 | Block |\n| Polypropylene | PP | Red | 1 | Fast | 15 | Jagged |\n| Polyethylene | PE | Amber | 2 | Medium | 20 | Foam |\n| Polystyrene | PS | Green | 2 | Fastest | 25 | Circle |\n| Aquafina bottle | WATER | Blue halo | varies | Medium | \u2014 | Image sprite |\n| Starbucks cup | COFFEE | Green halo | varies | Medium | \u2014 | Image sprite |\n\n---\n\n## \u26a1 Power-Ups\n\nPower-ups drift down the playfield and are collected by firing an enzyme at them (or by direct cell contact). Each lasts **10 seconds**.\n\n| Power-Up | Icon | Effect |\n|---|---|---|\n| FAST-PETase | \u26a1 | Rapid-fire enzyme shots |\n| Enzyme Boost | \ud83d\udcaa | Increased projectile damage |\n| Cell Repair | \u2764\ufe0f | Restores cell health |\n| Chain Reaction | \ud83d\udca5 | Explosive chain-blast on impact |\n| Marc \u00d710 | Co-founder avatar | 10\u00d7 fire rate + Marc's voice line |\n| Paul \u00d710 | Co-founder avatar | 10\u00d7 fire rate + Paul's voice line |\n\nCo-founder power-ups (Marc / Paul) disable the background fight track while the voice clip plays, then restore normal audio automatically on clip end.\n\n---\n\n## \ud83c\udfc6 Leaderboard\n\nScores are persisted to a Firebase Realtime Database (top 5 retained). On game over, players enter a display name and submit; the leaderboard renders immediately. It is also accessible from the start screen via **View Leaderboard**.\n\n```\nFirebase endpoint: https://microplastic-blaster-default-rtdb.firebaseio.com/leaderboard.json\n```\n\nEntries store: `name`, `score`, `level`, `date (ISO YYYY-MM-DD)`.\n\n---\n\n## \ud83d\udd27 Tech Stack\n\n| Layer | Technology |\n|---|---|\n| Rendering | HTML5 Canvas 2D API (`800 \u00d7 600` logical resolution) |\n| Scaling | CSS `transform` + responsive `resizeCanvas()` \u2014 fills viewport on desktop and mobile |\n| Audio | Web Audio API (synthesised SFX) + `<Audio>` element (voice clips, fight track) |\n| Leaderboard | Firebase Realtime Database (REST PUT/GET) |\n| Style | Vanilla CSS \u2014 no frameworks, no build step |\n| Assets | External PNGs + MP3s (co-founder avatars, brand sprites, audio clips) |\n\nThe entire game is a **single self-contained HTML file** with no npm dependencies or build pipeline.\n\n---\n\n## \ud83d\udcc1 Asset Requirements\n\nThe following files must reside in the **same directory** as the HTML file:\n\n```\nmicroplastic_blaster_115.html\nmarc.png                  # Co-founder avatar (Marc)\npaul.png                  # Co-founder avatar (Paul)\naquafina_new.png          # Bottled water enemy sprite\nstarbucks_new.png         # Coffee cup enemy sprite\nElora-Therapeutics-LogoStackedBlue.jpg   # Cell badge logo\nmarc.mp3                  # Marc co-founder voice line\npaul.mp3                  # Paul co-founder voice line\nfight.mp3                 # Background music track\n```\n\nMissing image assets degrade gracefully (fallback canvas fills / text labels). Missing audio assets fail silently.\n\n---\n\n## \ud83d\ude80 Running Locally\n\nNo server is required for basic gameplay. Open the HTML file directly in any modern browser:\n\n```bash\nopen microplastic_blaster_115.html        # macOS\nstart microplastic_blaster_115.html       # Windows\nxdg-open microplastic_blaster_115.html   # Linux\n```\n\n> **Note:** Some browsers block `<Audio>` autoplay and cross-origin fetches when opened as a `file://` URL. For full audio and leaderboard functionality, serve from a local HTTP server:\n\n```bash\n# Python 3\npython -m http.server 8080\n\n# Node.js (npx)\nnpx serve .\n```\n\nThen navigate to `http://localhost:8080/microplastic_blaster_115.html`.\n\n---\n\n## \ud83d\udcf1 Mobile Support\n\nThe canvas scales to fill the viewport on screens \u2264 700 px wide using `Math.max(w/GAMEWIDTH, h/GAMEHEIGHT)` to avoid letterboxing. Touch events (`touchstart`) are mapped to enzyme fire. The HUD collapses to a single-column layout below 700 px and hides column headers below 480 px. All breakpoints are handled in CSS only.\n\n---\n\n## \ud83d\udd2c The Science \u2014 Rotating Insight Messages\n\nBetween enzyme blasts, the in-game **Microplastic Insight** overlay rotates scientifically grounded messages every 8 seconds, covering topics including:\n\n- Microplastic accumulation in blood, organs, and adipose tissue\n- PET-linked inflammation and endocrine disruption\n- Chronic oxidative stress from particle-driven immune activation\n- The absence of endogenous polymer-degrading enzymes in mammals\n- The estimated 5 g/week human microplastic ingestion burden\n\n---\n\n## \ud83c\udfe2 About Elora Therapeutics", "# Crystal Structure of a PH Domain Containing Bacterial Protein (Exig_2160) from Exiguobacterium sibiricum 255-15 at 2.42 \u00c5 Resolution\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [3B77](https://www.rcsb.org/structure/3B77)\n* **Release Date:** 2007-11-13\n* **Method:** X-RAY DIFFRACTION 2.42 \u00c5\n* **Organisms:** Exiguobacterium artemiae 255-15\n* **Macromolecules:** Uncharacterized protein (protein)\n* **Publication:** [Deller, M.C., et al. (2010). Crystal Structure of a PH Domain Containing Bacterial Protein from Exiguobacterium sibiricum 255-15. Journal of Molecular Biology, 396, 31-46.](https://www.sciencedirect.com/science/article/pii/S0022283609013576)\n\nDescription: This structure reveals an unexpected PH domain architecture in a bacterial protein from the permafrost-dwelling Exiguobacterium sibiricum 255-15, a fold previously associated almost exclusively with eukaryotic signaling. The 2.42 \u00c5 structure expands our understanding of bacterial protein fold diversity and potential signaling mechanisms in extremophilic organisms.", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\nMutations were introduced into three of the identified clones (clones 6, 15, and 17) to generate 161 unique mutant clones (clones 55-215). The amino acid sequences of the six CDRs for each of the mutant clones are shown in Table 2. Light chain and heavy chain sequences of the parental clones (clones 6, 15, and 17) and mutant clones (clones 55-215) are shown in Table 5. One such mutation to the heavy chain, N297A, resulted in Fc effector function \"null\" mutants, where the Fc effector function, particularly ADCC, was eliminated or substantially eliminated. As can be seen from the activity data in the Examples, those antibodies with N297A Fc effector null mutations have a generally increased binding affinity for mutCALR than those antibodies without an Fc effector null mutation.\n\n### Example 2: Binding Affinity\n\nThe amino acid sequence of recombinant MBP-mutCALR Type 1 fusion protein used in Experiments 1 and 2 is given as SEQ ID NO:335 (an N-terminal His6 tag followed by maltose-binding protein fused to human Type 1 mutCALR).\n\n**Experiment 1.** Recombinant MBP-mutCALR Type 1 fusion protein was custom made at Cepter (cepterbiopartners.com). Data for the interaction of antibodies with recombinant MBP-mutCALR Type 1 fusion protein were collected at 25 \u00b0C on a Biacore 8K instrument (Cytiva, Marlborough, MA), and all reagents were obtained from Cytiva unless otherwise specified. The experiments were conducted using Tris-buffered saline (pH 7.2), 0.005% Surfactant P20, 1 mM CaCl2 as running buffer. To prepare the capture surfaces, anti-human IgG (Fc) antibodies (Product # 29234600) was amine-coupled onto a CM4 chip (Product # BR100534) under standard conditions using an Amine Coupling Kit (Product # BR100050). Anti-mutCALR antibodies were captured onto the chip surface by injecting over flow cell 2 only at a flow rate of 30 \u03bcL/min for 30 seconds. Typical capture levels were in the range of 10-30 RU. Recombinant MBP-mutCALR Type 1 fusion protein was prepared at nominal concentrations of 0, 3.1, 9.3, 27.8, 83.3, and 250 nM in the running buffer and was injected over both flow cells 1 and 2 for 210 seconds at a flow rate of 50 \u03bcL/min, followed by a 510-second dissociation phase at the same flow rate. The kinetics parameters were obtained by applying 1:1 binding model to fit the data from multiple-cycle injection experiments using Biacore Insight Evaluation software. MBP protein at 250 nM concentration was injected as a control, which showed no binding to anti-mutCALR antibodies.\n\n*Table 6.* Binding affinities and kinetic association and dissociation rate constants to mutCALR for the indicated antibodies for Experiment 1. (Representative rows; the full table lists 53 clones.)\n\n| Clone No. | ka (1/Ms) | kd (1/s) | KD (M) |\n|---|---|---|---|\n| 38 | 1.28E+06 | 2.68E-03 | 2.10E-09 |\n| 1 | 1.38E+06 | 1.96E-02 | 1.42E-08 |\n| 2 | 1.63E+06 | 1.23E-02 | 7.58E-09 |\n| 4 | 1.63E+06 | 7.04E-03 | 4.33E-09 |\n| 5 | 9.35E+05 | 4.05E-03 | 4.33E-09 |\n| 6 | 5.97E+05 | 1.53E-03 | 2.57E-09 |\n| 10 | 2.11E+06 | 2.92E-03 | 1.38E-09 |\n| 15 | 3.57E+07 | 2.15E-01 | 6.03E-09 |\n| 17 | 1.56E+06 | 3.47E-03 | 2.22E-09 |\n| 43 | 1.87E+05 | 6.20E-04 | 3.32E-09 |\n| 48 | 1.81E+05 | 2.97E-04 | 1.64E-09 |\n| 51 | 2.35E+05 | 1.09E-03 | 4.64E-09 |\n\nND, not determined.\n\n**Experiment 2.** Recombinant MBP-mutCALR Type 1 fusion protein was custom made at Cepter. Data for the interaction of antibodies with recombinant mutCALR protein were collected at 25 \u00b0C on a Biacore 8K instrument (Cytiva, Marlborough, MA), and all reagents were obtained from Cytiva unless otherwise specified. The experiments were conducted using Tris-buffered saline (pH 7.2), 0.005% Surfactant P20, 1 mM CaCl2 as running buffer. To prepare the capture surfaces, anti-human IgG (Fc) antibodies (Product # 29234600) was amine-coupled onto a CM4 chip (Product # BR100534) under standard conditions using an Amine Coupling Kit (Product # BR100050). Anti-mutCALR antibodies were captured onto the chip surface by injecting over flow cell 2 only at a flow rate of 10 \u03bcL/min for 30 seconds. Typical capture levels were in the range of 15-25 RU. Recombinant MBP-mutCALR Type 1 fusion protein was prepared at nominal concentrations of 0, 0.75, 2.22, 6.67, 20, and 60 nM and injected over both flow cells 1 and 2 for 150 seconds at a flow rate of 69 \u03bcL/min, followed by a 230-second dissociation phase at the same flow rate. The kinetics parameters were obtained by applying 1:1 binding model to fit the data from multiple-cycle injection experiments using Biacore Insight Evaluation software.\n\n*Table 7.* Binding affinities and kinetic association and dissociation rate constants to mutCALR for the indicated parental and variant antibodies for Experiment 2. (Representative rows; the full table lists more than eighty parental and affinity-matured clones. Affinities for the matured clones reach the high-picomolar range.)\n\n| Clone No. | ka (1/Ms) | kd (1/s) | KD (M) |\n|---|---|---|---|\n| 32 | 1.84E+06 | 3.14E-03 | 1.71E-09 |\n| 6 | 2.29E+06 | 1.23E-03 | 5.38E-10 |\n| 40 | 3.44E+06 | 1.21E-03 | 3.52E-10 |\n| 55 | 5.32E+06 | 9.11E-04 | 1.71E-10 |\n| 65 | 4.60E+06 | 2.31E-03 | 5.02E-10 |\n| 68 | 1.44E+06 | 1.66E-03 | 1.15E-09 |\n| 74 | 1.29E+06 | 1.53E-03 | 1.19E-09 |\n| 170 | 4.27E+06 | 1.79E-03 | 4.19E-10 |\n| 180 | 1.00E+07 | 1.27E-01 | 1.98E-10 |\n| 185 | 5.32E+06 | 9.11E-04 | 1.71E-10 |\n| 133 | 8.51E+06 | 7.02E-03 | 8.26E-10 |\n| 73 | 7.18E+06 | 5.46E-03 | 7.60E-10 |\n\nND, not determined.", "# 2025 \u2013 Tricyclic compounds as inhibitors of wrn\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-12 on marcdeller.com (categories: patents).*\n\n* [Link to PDF](https://marcdeller.com/wp-content/uploads/2025/08/US20250215006A1.pdf)\n* [Link to Google Patent](https://patents.google.com/patent/US20250215006A1/en?oq=US20250215006A1)\n\n* Compounds selectively target WRN helicase, crippling DNA repair in MSI-H/dMMR cancers while sparing healthy cells.\n* These drugs work via a mechanism independent of immunotherapy, providing new options for resistant tumors.\n* Preclinical research shows effectiveness even where standard chemotherapy, targeted drugs, and immunotherapy no longer work.\n* By filling a major treatment gap, WRN inhibitors could reshape therapeutic possibilities for a significant subset of hard-to-treat cancers.\n\n1. **Chu, W. et al. RecQ helicases: multifunctional genome caretakers.** [Nat Rev Cancer 9, 644-654 (2009)](https://pubmed.ncbi.nlm.nih.gov/19657341/)\n2. **Behan, F. M. et al. Prioritization of cancer therapeutic targets using CRISPR-Cas9 screens.** [Nature 568, 511-516 (2019)](https://pubmed.ncbi.nlm.nih.gov/30971826/)\n3. **Chan, E. M. et al. WRN helicase is a synthetic lethal target in microsatellite unstable cancers.** [Nature 568, 551-556 (2019)](https://pubmed.ncbi.nlm.nih.gov/30971823/)\n4. **Kategaya, L. et al. Werner syndrome helicase is required for the survival of cancer cells with microsatellite instability.** [iScience 13, 488-497 (2019))](https://pubmed.ncbi.nlm.nih.gov/30898619/)\n\nThis patent introduces a novel class of pharmaceutical compounds crafted to interfere with the enzyme Werner helicase (WRN), a key player in maintaining DNA integrity and cellular stability. The WRN enzyme, a member of the RecQ helicase family, is indispensable for cellular DNA repair and chromosome stability. However, certain aggressive cancers\u2014specifically those with high microsatellite instability (MSI-H) or deficient mismatch repair (dMMR)\u2014become dangerously reliant on WRN\u2019s function for survival.\n\nScientific screening has revealed a pronounced dependence on WRN for the survival of MSI-H cancer cells. When WRN is removed or inhibited, these vulnerable cancer cells lose their ability to efficiently repair DNA, leading to halted growth and cell death. In contrast, cancers that do not exhibit microsatellite instability remain largely unaffected by the absence of WRN, pointing to a highly selective vulnerability that these new compounds exploit.\n\nTraditional treatment options for MSI-H/dMMR cancers, such as immune checkpoint inhibitors, have shown benefit but come with challenges: response rates remain modest, and some patients develop resistance. The compounds described in this patent employ a fundamentally different approach, undermining cancer cell survival by obstructing WRN\u2019s adenine triphosphate (ATP)-driven mechanisms\u2014specifically its ability to unwind DNA and to hydrolyze ATP, both essential for DNA repair and cancer growth.\n\nA particularly compelling aspect of these WRN inhibitors is their potential application in overcoming treatment resistance. Preclinical evidence suggests that even cancers with established resistance to chemotherapy, targeted treatments, or immunotherapy remain susceptible to WRN inhibition. This opens new avenues for patients whose tumors have failed to respond to existing therapies.\n\nThe patent outlines the structural makeup of these inhibitors, their pharmaceutical formulations, and potential treatment regimens. The envisioned uses extend beyond monotherapy: these compounds are intended to be versatile, with the potential to combine with other modern therapies\u2014such as checkpoint inhibitors\u2014for a more comprehensive attack on difficult-to-treat cancers. This innovation positions WRN inhibitors as a promising class of targeted treatment\u2014precise, selective, and potentially transformative for patients with MSI-H/dMMR malignancies.", "# Cats Confuse Reasoning LLM: Query-Agnostic Adversarial Triggers for Reasoning Models\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-27 on marcdeller.com (categories: ai).*\n\n* **CatAttack Pipeline:** automated discovery and transfer of query-agnostic triggers\n* **Transferability:** triggers found on proxy models generalize to advanced reasoning models\n* **Error Amplification:** triggers yield over 300% increase in incorrect outputs\n* **Semantic Integrity:** attacks preserve original question meaning, requiring no domain knowledge\n* **Computational Overhead**: triggers induce unnecessary reasoning loops\n* **Defense Directions:** model randomization, semantic filtering, adversarial fine-tuning\n\n1. **Cats Confuse Reasoning LLM: Query-Agnostic Adversarial Triggers for Reasoning Models:** Meghana Rajeev, et.al., arXiv: <https://arxiv.org/pdf/2503.01781.pdf>\n2. **Can Model Randomization Offer Robustness Against Query-Based Adversarial Attacks:** Quoc Viet Vo, et.al., ICLR 2025: <https://openreview.net/forum?id=DpnY7VOktT>\n3. **ShadowCoT: Cognitive Hijacking for Stealthy Reasoning Backdoors in LLMs:** Gejian Zhao, et.al., arXiv: <https://arxiv.org/abs/2504.05605>\n4. **GSM-PLUS: A Comprehensive Benchmark for Evaluating the Robustness of LLMs as Mathematical Problem Solvers:** Qintong Li, et.al., ACL 2024: <https://aclanthology.org/2024.acl-long.163/>\n\nState-of-the-art reasoning models can be **severely misled** by simple, irrelevant text snippets, raising urgent concerns for robustness and security.\n\nThe recent study \u201cCats Confuse Reasoning LLM: Query Agnostic Adversarial Triggers for Reasoning Models\u201d reveals that appending a brief phrase\u2014such as \u201cInteresting fact: cats sleep most of their lives\u201d\u2014to any math problem more than doubles the probability of an incorrect solution. The authors introduce **CatAttack**, an automated pipeline that discovers adversarial triggers on a low-cost proxy model (DeepSeek V3) and successfully transfers them to stronger reasoning models like DeepSeek R1 and Qwen-32B. This transferability proves that vulnerabilities are inherent to model architectures rather than training data or specific tasks. By evaluating performance across multiple datasets, the researchers show over a 300% increase in error rates on target models when these triggers are present. Moreover, they observe that triggers not only induce wrong answers but also cause **unwarranted reasoning loops**, leading to significant computational overheads.\n\nExisting adversarial approaches often rely on altering semantic content or leveraging ground-truth answers for perturbation. In contrast, CatAttack preserves the original problem semantics entirely, requiring no domain expertise. This makes it a particularly **insidious threat** in applications ranging from finance to healthcare, where reasoning accuracy is critical. The pipeline combines a **prompt optimizer**, a **judge model** for hallucination detection, and a **transfer mechanism** to refine and validate triggers iteratively. Their experiments on a held-out math benchmark demonstrate a 50% transfer success rate from V3 to R1-distilled-Qwen-32B, underscoring CatAttack\u2019s efficiency and stealth.\n\nThese findings call for novel defense strategies. Potential mitigations include **model randomization** to disrupt adversarial query consistency , **semantic filtering** to detect irrelevant prefixes or suffixes, and **adversarial fine-tuning** with CatAttack triggers to immunize models. Importantly, the existence of query-agnostic triggers suggests that security evaluations must include threat models beyond input-specific attacks. As reasoning models become integral to high-stakes decision-making, addressing these vulnerabilities is essential to ensure reliability and trustworthiness.", "# GitHub README: FlexAppeal\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/FlexAppeal, 2026\n\nMIT is compatible with everything FlexAppeal depends on. The two libraries this\nrepository actually redistributes, Plotly.js and Mol\\*, are both MIT, and their\ncopyright notices are retained as that licence requires. MDTraj (LGPL-2.1+) and\ngemmi (MPL-2.0) are imported unmodified through their published interfaces,\nwhich both licences permit from differently licensed code.\n\nOpenMM, PDBFixer, openmmforcefields, the OpenFF toolkit and AmberTools are never\ndistributed as part of FlexAppeal. A generated bundle installs them into its own\nenvironment on your machine, and the run script calls them there.\n\n**Cite what you use.** Force-field parameters (AMBER, CHARMM36, OpenFF, GAFF),\nchemical definitions from the RCSB Chemical Component Dictionary, and structures\nfrom the RCSB PDB, OPM or the AlphaFold Database each carry their own terms and\ncitation expectations. FlexAppeal records the accession and the force field in\nevery results file's manifest so the provenance is there when you write it up.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/FlexAppeal\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/FlexAppeal</a></td>\n</tr>\n</table>", "# marcdeller.com: My Structures\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n## Protein Structures: My Atomic\u2011Level Discoveries\n\n400+ protein structures show significant advances in structural biology, medicinal chemistry, and immunology. They showcase my expertise in construct design, protein crystallization, X-ray crystallography, and structure determination. These structures address challenging protein-protein complexes and answer difficult biological problems in structure-based drug discovery, glycoprotein structure, ligand-bound complexes, vaccine design, and protein stability.\n\nMy PDB\u2019s\n\nTop Structures\n\n### [Crystal Structure of [KS3][AT3] Didomain from Module 3 of 6-Deoxyerythronolide B Synthase in Complex with Antibody Fragment (Fab)](https://marcdeller.com/crystal-structure-of-ks3at3-didomain-from-module-3-of-6-deoxyerythronolide-b-synthase-in-complex-with-antibody-fragment-fab/)\n\n[Crystal Structure of [KS3][AT3] Didomain from Module 3 of 6-Deoxyerythronolide B Synthase in Complex with Antibody Fragment (Fab)](https://marcdeller.com/crystal-structure-of-ks3at3-didomain-from-module-3-of-6-deoxyerythronolide-b-synthase-in-complex-with-antibody-fragment-fab/)\n\n### [Crystal Structure of Human Oncostatin M](https://marcdeller.com/crystal-structure-of-human-oncostatin-m/)\n\n[Crystal Structure of Human Oncostatin M](https://marcdeller.com/crystal-structure-of-human-oncostatin-m/)\n\n### [Discovery of (4-Pyrazolyl)-2-Aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2](https://marcdeller.com/discovery-of-4-pyrazolyl-2-aminopyrimidines-as-potent-and-selective-inhibitors-of-cyclin-dependent-kinase-2/)\n\n[Discovery of (4-Pyrazolyl)-2-Aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2](https://marcdeller.com/discovery-of-4-pyrazolyl-2-aminopyrimidines-as-potent-and-selective-inhibitors-of-cyclin-dependent-kinase-2/)\n\n### [Discovery of an Orally Bioavailable KRAS G12D Inhibitor](https://marcdeller.com/discovery-of-an-orally-bioavailable-kras-g12d-inhibitor/)\n\n[Discovery of an Orally Bioavailable KRAS G12D Inhibitor](https://marcdeller.com/discovery-of-an-orally-bioavailable-kras-g12d-inhibitor/)\n\nStructures\n\n### [Discovery of an Orally Bioavailable KRAS G12D Inhibitor](https://marcdeller.com/discovery-of-an-orally-bioavailable-kras-g12d-inhibitor/)\n\nStructures\n\n### [Discovery of an Orally Bioavailable KRAS G12D Inhibitor_2](https://marcdeller.com/discovery-of-an-orally-bioavailable-kras-g12d-inhibitor_2/)\n\nStructures\n\n### [Discovery of (4-Pyrazolyl)-2-Aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2](https://marcdeller.com/discovery-of-4-pyrazolyl-2-aminopyrimidines-as-potent-and-selective-inhibitors-of-cyclin-dependent-kinase-2/)\n\nStructures\n\n### [FGFR2 Kinase Domain in Complex with a Pyrazolo[1,5-a]pyrimidine Analog (Compound 29)](https://marcdeller.com/fgfr2-kinase-domain-in-complex-with-a-pyrazolo15-apyrimidine-analog-compound-29/)\n\nStructures\n\n### [Crystal Structure of [KS3][AT3] Didomain from Module 3 of 6-Deoxyerythronolide B Synthase in Complex with Antibody Fragment (Fab)](https://marcdeller.com/crystal-structure-of-ks3at3-didomain-from-module-3-of-6-deoxyerythronolide-b-synthase-in-complex-with-antibody-fragment-fab/)\n\nStructures\n\n### [Crystal Structure of a Cellulose-Active Polysaccharide Monooxygenase from M. thermophila (MtPMO3*)](https://marcdeller.com/crystal-structure-of-a-cellulose-active-polysaccharide-monooxygenase-from-m-thermophila-mtpmo3/)\n\nStructures\n\n### [Crystal Structure of the Complex of 10E8 Fab Light Chain Mutant1 and T117v2 HIV-1 MPER Scaffold](https://marcdeller.com/crystal-structure-of-the-complex-of-10e8-fab-light-chain-mutant1-and-t117v2-hiv-1-mper-scaffold/)\n\nStructures\n\n### [Crystal Structure of an Immature VRC01-Class Antibody DRVIA7 from a Chinese Donor Bound to Clade A/E HIV-1 gp120 Core](https://marcdeller.com/crystal-structure-of-an-immature-vrc01-class-antibody-drvia7-from-a-chinese-donor-bound-to-clade-a-e-hiv-1-gp120-core/)\n\nStructures\n\n### [Crystal Structure of BG505 SOSIP gp140 HIV-1 Env Trimer Bound to Broadly Neutralizing Antibodies PGT128 and 8ANC195](https://marcdeller.com/crystal-structure-of-bg505-sosip-gp140-hiv-1-env-trimer-bound-to-broadly-neutralizing-antibodies-pgt128-and-8anc195/)\n\nStructures\n\n### [Crystal Structure of Nanog Homeodomain (NANOG) from Homo sapiens at 3.30 \u00c5 Resolution](https://marcdeller.com/crystal-structure-of-nanog-homeodomain-nanog-from-homo-sapiens-at-3-30-a-resolution/)\n\nStructures\n\n### [Crystal Structure of the BG505 SOSIP gp140 HIV-1 Env Trimer in Complex with the Broadly Neutralizing Fab PGT122](https://marcdeller.com/crystal-structure-of-the-bg505-sosip-gp140-hiv-1-env-trimer-in-complex-with-the-broadly-neutralizing-fab-pgt122/)\n\nStructures\n\n### [Crystal Structure of PGT135 Fab in Complex with gp120 Core Protein from HIV-1 Strain JR-FL Bound to CD4 and 17b Fab](https://marcdeller.com/crystal-structure-of-pgt135-fab-in-complex-with-gp120-core-protein-from-hiv-1-strain-jr-fl-bound-to-cd4-and-17b-fab/)\n\nStructures\n\n### [Structure of the Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Region 412-423 Bound to the Broadly Neutralizing Antibody AP33](https://marcdeller.com/structure-of-the-hepatitis-c-virus-envelope-glycoprotein-e2-antigenic-region-412-423-bound-to-the-broadly-neutralizing-antibody-ap33/)\n\nStructures\n\n### [Crystal Structure of Acetyl Xylan Esterase (TM0077) from Thermotoga maritima at 2.40 \u00c5 Resolution](https://marcdeller.com/crystal-structure-of-acetyl-xylan-esterase-tm0077-from-thermotoga-maritima-at-2-40-a-resolution/)\n\nStructures\n\n### [Crystal Structure of an Octomeric Two-Subunit TrkA K+ Channel Ring Gating Assembly from Thermotoga maritima](https://marcdeller.com/crystal-structure-of-an-octomeric-two-subunit-trka-k-channel-ring-gating-assembly-from-thermotoga-maritima/)\n\nStructures\n\n### [Crystal Structure of TrkA-N Domain Protein from Thermotoga sp. RQ2](https://marcdeller.com/crystal-structure-of-trka-n-domain-protein-from-thermotoga-sp-rq2/)\n\nStructures", "# Precision Timing of Cytokine Signals through Facilitated Dissociation\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-13 on marcdeller.com (categories: blog, sb).*\n\n* [Design of facilitated dissociation enables timing of cytokine signalling](https://www.nature.com/articles/s41586-025-09549-z)\n\n* **Molecular Timer:** [Engineered excited states trigger rapid receptor release within seconds](https://www.nature.com/articles/s41586-025-09549-z)\n* **Kinetic Control:** [5,700-fold increase in dissociation rate enables precise IL-2 pulses](https://pubmed.ncbi.nlm.nih.gov/40993395)\n* **Multistate Design:** AlphaFold2 and Rosetta tools optimize folding and allostery for dual conformations\n* **Structural Validation:** X-ray crystallography and DEER spectroscopy confirm strained excited states\n* **Therapeutic Potential:** [On-demand cytokine termination reduces toxicity in immunotherapies](https://genengnews.com/topics/artificial-intelligence/ai-designed-protein-switches-control-drugs-with-speed/)\n* **Broader Impact:** [Framework applicable to other cytokines and synthetic biological circuits](https://www.biorxiv.org/content/10.1101/2024.01.18.576190v2.full.pdf)\n\n1. **Design of facilitated dissociation enables timing of cytokine signalling:** [Adam J. Broerman, et al., Nature](https://www.nature.com/articles/s41586-025-09549-z)\n2. **NF-\u03baB responds to absolute differences in cytokine concentrations:** [Minjun Son, et al., Sci. Signal](https://www.science.org/doi/10.1126/scisignal.aaz4382)\n3. **Molecular Level Insights Into the Structural and Dynamic Factors Driving Cytokine Function:** [Jennifer Y. Cui, et al., Front. Mol. Biosci.](https://www.frontiersin.org/journals/molecular-biosciences/articles/10.3389/fmolb.2021.773252/full)\n4. **Signaling pathways in the regulation of cytokine release syndrome in human diseases and intervention therapy:** [Xun Li, et al., Signal Transduct. Target Ther.](https://doi.org/10.1038/s41392-021-00764-4)\n5. **Diffusion-limited cytokine signaling in T cell populations:** [Patrick Brunner, et al., iScience](https://doi.org/10.1016/j.isci.2024.110134)\n6. **Cytokines: From Clinical Significance to Quantification:** [Chao Liu, et al., Adv. Sci.](https://doi.org/10.1002/advs.202004433)\n\n**Unlocking Temporal Control of Interleukin-2 Signalling for Next-Generation Therapeutics**\n\nA groundbreaking protein-design platform now enables engineered cytokine receptors and mimics to dissociate on command, transforming hours-long interleukin-2 (IL-2) signals into precise, seconds-scale pulses. By sculpting excited conformational states that destabilize protein\u2013protein complexes, [researchers have programmed dissociation rates up to 5,700-fold faster](https://www.nature.com/articles/s41586-025-09549-z), paving the way for tunable immunotherapies, rapid biosensors, and kinetically controlled cell circuits.\n\nHarnessing induced-fit power strokes, the team designed fusion proteins with two distinct structural states: one that binds tightly to IL-2 receptors and another that introduces steric strain to trigger rapid release. Cutting-edge deep-learning tools (AlphaFold2, RosettaFold, RFdiffusion, ProteinMPNN) guided multistate design, ensuring robust folding and allosteric coupling. X-ray crystallography and double-electron-electron resonance confirmed the intended excited-state architectures, while kinetic assays quantified dissociation enhancements\u2014up to ~5,700-fold for receptor\u2013effector complexes. These rapid on/off switches enabled dissection of IL-2 signalling dynamics in live cells, revealing [how precise temporal patterns govern downstream STAT5 activation and cell-fate decisions](https://pubmed.ncbi.nlm.nih.gov/40993395).\n\nMoving beyond static affinity, this strategy redefines protein design by targeting the kinetics of molecular interactions. Facilitated dissociation constructs can be generalized to other cytokines and receptors, offering a versatile toolkit for engineering immune responses with unprecedented temporal resolution. Future applications include immunotherapies that minimize systemic toxicity by terminating cytokine action on cue, and diagnostic sensors that amplify transient biomarker signals into detectable outputs within seconds.\n\nKey takeaways from 2020\u20132025 research demonstrate that tuning dissociation kinetics is essential for spatial and temporal precision in cytokine signalling. Integration of dynamic feedback loops in tissue niches [further modulates signal localization and efficacy](https://www.biorxiv.org/content/10.1101/2024.01.18.576190v2.full.pdf). By coupling multistate protein design with high-resolution structure prediction, facilitated dissociation charts a new path toward programmable immunomodulation.\n\n| Concept | Description | Key Reference |\n| --- | --- | --- |\n| Facilitated dissociation | Engineering excited states to accelerate complex dissociation | [Broerman, et al., Nature](https://www.nature.com/articles/s41586-025-09549-z) |\n| Multistate design | Computational optimization of two conformations for allosteric switching | [Broerman, et al., Nature](https://www.nature.com/articles/s41586-025-09549-z) |\n| Kinetic tuning | Programming dissociation rates up to ~5,700-fold | [Broerman, et al., Nature](https://pubmed.ncbi.nlm.nih.gov/40993395) |\n| Structural validation | X-ray crystallography and DEER spectroscopy confirm excited states | [Broerman, et al., Nature](https://pubmed.ncbi.nlm.nih.gov/40993395) |\n| Dynamic feedback | Spatial and temporal niche regulation of cytokine gradients | [Sj\u00f6lin, et al., bioRxiv](https://www.biorxiv.org/content/10.1101/2024.01.18.576190v2.full.pdf) |\n| Therapeutic applications | On-demand cytokine termination to reduce side effects | [GEN News](https://genengnews.com/topics/artificial-intelligence/ai-designed-protein-switches-control-drugs-with-speed/) |", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\n*Figure 4-5.* Figure showing 12 % polyacrylamide gel of the expression and affinity purification of LIFR-FN3-3C-Fc. Lane 1 contains Protein-A sepharose after passage of 293T supernatant. Lane 2 shows the Protein-A sepharose after cleavage with 3C protease. Lane 3 and 4 show unbound material and a wash step, respectively. Lane 5, 6 and 7 show the first three eluted fractions of LIFR-FN3 following 3C protease cleavage. Lane 8 contains high range molecular weight markers running at 220, 97, 66, 46, 30, 21 and 14 kDa (Amersham).\n\n#### 4.3.3.3 Western Blot Analysis\n\nFigure 4-6 below shows a western blot of LIFR-FN3-3C-Fc samples at various stages of the purification scheme. Lane 1 contains conditioned 293T supernatant and no bands are visible. Lane 2 contains a sample of Protein-A sepharose after passage of conditioned 293T supernatant and three major bands are observed at ~220, ~140 and ~46 kDa. Lane 3 contains a sample of Protein-A sepharose wash and no bands are observed. Lane 4 contains recombinant PNGase F and lane 7 contains non-recombinant PNGase F (New England Biolabs) and no cross reactivity with the LIFR antibody is observed. Lanes 5 and 6 contain samples of Protein-A sepharose bound LIFR-FN3-3C-Fc after treatment with recombinant PNGase F and non-recombinant PNGase F, respectively. Comparison of the ~140 kDa band in lanes 2, 5 and 6 suggests a slight reduction in molecular weight following PNGase F treatment, with the non-recombinant variant displaying a tighter band so suggesting more complete deglycosylation. Lane 8 contains 3C-His6 protease and no cross reactivity with the LIFR antibody is observed. Lanes 9 and 10 contain eluted fractions of LIFR-FN3 after 3C protease treatment and a single diffuse band is observed at ~116 kDa. Lane 11 contains a sample of Protein-A sepharose after elution of LIFR-FN3 and no bands are observed suggesting complete elution of LIFR. The positions of the broad range molecular weight rainbow markers are noted at 220, 97, 66, 46 and 30 kDa (Bio-Rad).\n\n*Figure 4-6.* Figure showing western blot of LIFR-FN3-3C-Fc samples at various stages of the purification scheme. 12 % polyacrylamide gel is shown probed with anti-LIFR antibody (Binding Site). Lane 1 contains conditioned 293T supernatant and lane 2 contains Protein-A sepharose after passage of supernatant. Lane 3 contains Protein-A sepharose wash. Lane 4 contains recombinant PNGase F and lane 5 contains Protein-A sepharose bound LIFR-FN3-3C-Fc after treatment with recombinant PNGase F. Lane 6 contains Protein-A sepharose bound LIFR-FN3-3C-Fc after treatment with non-recombinant PNGase F (New England Biolabs) and lane 7 contains non-recombinant PNGase F. Lane 8 contains 3C-His6 protease and lanes 9 and 10 contain eluted fractions of LIFR-FN3 after 3C protease treatment. Lane 11 contains Protein-A sepharose after elution of LIFR-FN3. The positions of the broad range molecular weight rainbow markers are noted at 220, 97, 66, 46 and 30 kDa (Bio-Rad), although no bands are visible on the blot.\n\n#### 4.3.3.4 Iso-electric Focussing\n\nFigure 4-7 below shows a western blot of 5.5 % polyacrylamide IEF gel of various samples of LIFR-FN3-3C-Fc. Lane 1 contains 3C protease cleaved LIFR-FN3 and no major bands are observed. Lane 2 contains acid eluted LIFR-FN3-3C-Fc which again shows no major bands and lane 3 contains Protein-A bound LIFR-FN3-3C-Fc which shows one major band at a pI ~5\u20136; compared with the theoretical pI of the LIFR-FN moiety at 7.2. Lane 4 contains non-recombinant PNGase F treated LIFR-FN3-3C-Fc and no major bands are observed. Lane 5 contains a control sample of OSM in which no cross reactivity to the LIFR antibody is observed. Lane 6 shows IEF standards containing phycocyanin (pI 4.5 and 4.8), equine myoglobin (pI 6.8), human haemoglobin A (pI 7.1) and human haemoglobin C (pI 7.5) (Bio-Rad) which all show single bands due to the luminescent detection method used.\n\n*Figure 4-7.* Figure showing western blot of 5.5 % polyacrylamide IEF gel prepared using broad range ampholines (pI 3\u201310) (Bio-Rad). LIFR antibody (Binding Site) was used with ECL (Amersham) chemiluminescent detection. Lane 1 contains 3C protease cleaved LIFR-FN3. Lane 2 contains acid eluted LIFR-FN3-3C-Fc and lane 3 contains Protein-A bound LIFR-FN3-3C-Fc. Lane 4 contains PNGase F (New England Biolabs) treated LIFR-FN3-3C-Fc and lane 5 contains a control sample of OSM. Lane 6 shows IEF standards containing phycocyanin (pI 4.5 and 4.8), equine myoglobin (pI 6.8), human haemoglobin A (pI 7.1) and human haemoglobin C (pI 7.5) (Bio-Rad).\n\n#### 4.3.3.5 Dynamic Light Scattering\n\nFigure 4-8 below shows samples of deglycosylated LIFR-FN3 analysed on a DynaPro-LS Sizing Instrument (Protein Solutions Inc.) at various time points. Panel (a) shows the sample at 0 minutes, directly after filtering through a 0.1 \u03bcm membrane filter, and one major species at a hydrodynamic radius of ~1 \u00d7 10^\u22122 \u03bcm is observed. Panel (b) shows the same sample after 5 minutes incubation at 20 \u00b0C and these data display one major species at ~1 \u00d7 10^\u22122 \u03bcm with the peak being split into a second minor species at ~1.5 \u00d7 10^\u22121 \u03bcm. Panel (c) shows the sample after 10 minutes incubation and two distinct species at hydrodynamic radii of ~1 \u00d7 10^\u22122 and ~2.9 \u00d7 10^\u22121 \u03bcm are observed. Samples analysed after ~6 hours (data not shown) have the same profile as panel (c), suggesting equilibrium is reached after ~10 minutes.", "# Cell surface receptors\nAuthors: Deller, M.C.; Jones, E.Y.\nSource: Current Opinion in Structural Biology, 2000\n\nWith the exception of the growth hormone\u2013receptor system, the erythropoietin (Epo)\u2013receptor (EpoR) system is the most fully characterised of the hematopoietic receptor complexes. This system has proved particularly fascinating because of the unique availability of detailed structural information on the mode of receptor binding by a peptide mimetic of the Epo cytokine [19], as well as on the physiological cytokine\u2013receptor recognition complex [20]. The story has been recently reviewed by some of the key players [21]. As is now the norm, the results this past year highlight the crucial interplay between structural results and functional studies. Alanine-substitution mutagenesis of EpoR further underscores the central role of a pair of phenylalanine residues in binding unrelated physiological and mimetic ligands [22]. One key new structural result has been the determination of an antagonist peptide\u2013EpoR complex [23]. When examined alongside the existing structures of signalling complexes, formed by either the physiological Epo ligand or the agonist peptide mimetic, the importance of the relative positioning of the two receptors cross-linked by the ligand is thrown into sharp relief (Figure 2a,b). Thus, extracellular domain orientation appears to be tightly coupled to the cytoplasmic signalling events. This picture is further compounded by the suggestive dimerisation mode revealed in the crystal lattice formed by unliganded EpoR [24]. This geometry is not commensurate with the orientation required for signalling and might thus be a mechanism for maintaining receptor pairs at the cell surface while avoiding constitutive signalling. The inference of physiological interaction modes on the basis of crystal lattice contacts can necessarily only suggest hypotheses for further functional investigation; it is therefore important that this structural evidence for dimerisation is reported back-to-back with supporting in vivo analyses [25].\n\n*Figure 2.* The extracellular disposition of cytokine\u2013receptor complexes required to trigger intracellular signal transduction. (a) The structure of the 1:2 complex of Epo\u2013EpoR [20]. (b) The structure of the 1:2 complex of EpoR with a dimeric antagonist peptide [23]. In each structure, the ligand is coloured black and the orientation is consistent with the cell surface being at the bottom of the panel. The view in both panels is matched with reference to the right hand side receptor. The relative positions of the extracellular receptor regions are seen to differ between the agonist (signalling) complex and the antagonist complex. The direction of the relative reorientation of the left hand side receptor is denoted by an arrow. (c,d) Orthogonal views of the 1:3 complex of the trimeric TRAIL ligand (coloured black) with the DR5 receptor [27]. The 54 \u00c5 distances between the C termini of the extracellular portions of the three copies of DR5 (i.e. the start points for the transmembrane helices) are strikingly similar to the 52 \u00c5 spacings of binding sites on TRAF2 [56], a trimeric cytosolic molecule implicated in the initiation of intracellular signalling for this family of receptors.\n\n## The trimeric trigger for apoptosis\n\nThe second cytokine\u2013receptor interaction to be characterised structurally was that of tumour necrosis factor (TNF) with its receptor (TNFR) in 1993 [26]. In the intervening years, sequence data have revealed substantial superfamilies of TNF-like ligands and TNFR-type receptors, many of which play a key role in apoptosis (controlled cell death). In contrast, further structural information on this form of cytokine\u2013receptor interaction remained suspended until this year, when, typically, two structures of the same complex, TRAIL (TNF-related apoptosis-inducing ligand)\u2013DR5 (death receptor 5), were reported contemporaneously [27,28] (Figure 2c,d). TRAIL, like TNF, is trimeric; indeed, the structure of the ligand seems highly conserved within the superfamily and a comparison of the structure of the receptor-bound ligand with that of the isolated TRAIL molecule [29] points to a rigid, unchanging scaffold for cross-linking potentially more flexible receptors. A comparison with the TNF\u03b2\u2013TNFR1 complex reveals some conservation of the most hydrophobic portion of the interaction interface, with specificity conferred by variation in more polar parts of the interface. In an echo of the Epo\u2013EpoR story, the precise positioning of the receptors within the extracellular region of the threefold symmetric signalling complex may be necessary for triggering the apoptotic signal (Figure 2c,d).\n\n## The developing story for fibroblast growth factors and their receptors\n\nThe fibroblast growth factors (FGFs) and their receptors (FGFRs) are crucial to the control of growth and differentiation for a wide variety of cells during development. Structurally, the system is broadly analogous to that of monomeric IL-1 and its IgSF receptor IL-1R (discussed above; Figure 1b) and thus requires a further component for receptor cross-linking. From functional studies, the formation of the signalling complex is believed to require dimerisation of the receptor through interaction with the ligand and with heparin or heparan-sulfate-containing proteoglycans. Very recently, structures have been reported for two types of FGF\u2013FGFR complex (FGF2\u2013FGFR1 [30] and FGF1\u2013FGFR2 [31]). In both cases, the primary complex is 1:1 and correlates closely with the expected binding site for FGF between Ig-like domains two and three of the receptor. In apparent contrast to the IL-1\u2013IL-1R system, the N-terminal Ig-like domain of FGFR has not been functionally implicated in binding and was omitted from the receptor fragments used in the two structural studies. Unexpectedly, both studies also reveal further dimerisation of the 1:1 complex within the crystal, consistent with the formation of a 2:2 complex. The functional implications of this observation are considerable.", "# 2025 \u2013 Introduction to R: Basic R syntax\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-28 on marcdeller.com (categories: certificates).*\n\n* [Link to certificate](https://www.coursera.org/account/accomplishments/verify/T687VG5UYI5O)\n\n* **Award Details:** Coursera issued this project certificate on March 27, 2025.\n* **Course Focus:** It recognizes completion of a project covering basic R syntax.\n* **Career Value:** This matters because programming supports reproducible statistics and transparent data workflows.\n* **Practical Use:** It helps with scripting, plotting, and automating routine scientific analyses.", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\n**45.** The antibody of claim 44, wherein the antibody: inhibits one or more signaling pathways downstream of MPL in both a first cell expressing human Type 1 mutant CALR and in a second cell expressing human Type 2 mutant CALR; inhibits oncogenic cell proliferation in both a first cell expressing human Type 1 mutant CALR and in a second cell expressing human Type 2 mutant CALR; and/or inhibits dimerization of MPL in both a first cell expressing human Type 1 mutant CALR and in a second cell expressing human Type 2 mutant CALR.\n\n**46.** The antibody of claim 44 or claim 45, wherein the one or more signaling pathways downstream of MPL are selected from the group consisting of Janus tyrosine kinase (JAK) and signal transducers and activators of transcription (STAT) signaling, mitogen-activated protein kinase (MEK) and extracellular signal-regulated kinase (ERK) signaling, serine/threonine kinase (AKT) signaling, and mammalian target of rapamycin (mTOR) signaling.\n\n**47.** The antibody of claim 44, wherein the antibody binds to human mutant CALR with a greater affinity than human mutant CALR binds to MPL.\n\n**48.** The antibody of claim 44, wherein the antibody binds to human mutant CALR with an IC50 of between about 0.1 and 1 nM.\n\n**49.** An antibody that binds to human mutant calreticulin (CALR), wherein the antibody has modulated Fc effector function.\n\n**50.** The antibody of claim 49, wherein the modulated Fc effector function is increased Fc effector function or reduced Fc effector function.\n\n**51.** The antibody of claim 49 or claim 50, wherein the antibody has reduced Fc effector function.\n\n**52.** The antibody of any one of claims 49-51, wherein the Fc effector function is antibody-dependent cell-mediated cytotoxicity (ADCC), complement dependent cytotoxicity (CDC), or antibody dependent cellular phagocytosis (ADCP).\n\n**53.** The antibody of any one of claims 49-52, wherein the antibody having reduced Fc effector function has increased binding affinity to human mutant CALR as compared to an antibody without reduced Fc effector function.\n\n**54.** The antibody of claim 53, wherein the Fc effector function is ADCC.\n\n**55.** An antibody that binds to human mutant calreticulin (CALR), wherein the antibody binds to human mutant CALR at an epitope within DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328).\n\n**56.** The antibody of claim 55, wherein the antibody binds within amino acids 8-20 of DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328).\n\n**57.** The antibody of claim 56, wherein the antibody binds to: one or more of the amino acid residues R8, M10, M19, and R20 in DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328); or all of the amino acid residues R8, M10, M19, and R20 in DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328).\n\n**58.** The antibody of claim 55, wherein the antibody binds within amino acids 1-20 of DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328).\n\n**59.** The antibody of claim 58, wherein the antibody binds to: one or more of D1, R5, R8, M9, M10, R11, T12, M14, R15, M16, M19, and R20 in DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328); or all of D1, R5, R8, M9, M10, R11, T12, M14, R15, M16, M19, and R20 in DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328).\n\n**60.** The antibody of claim 55, wherein the antibody binds within amino acids 7-21 of DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328).\n\n**61.** The antibody of claim 60, wherein the antibody binds to one or more of the amino acid residues R7, R11, R17, R20, and R21 in DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328).\n\n**62.** The antibody of claim 60, wherein the antibody binds to all of the amino acid residues R7, R11, R17, R20, and R21 in DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328).\n\n**63.** The antibody of claim 55, wherein the antibody binds within amino acids 6-21 of DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328).\n\n**64.** The antibody of claim 55, wherein the antibody binds to one or more of the amino acid residues T6, R7, M10, R11, M14, R17, R18, R20, and R21 in DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328).\n\n**65.** The antibody of claim 63, wherein the antibody binds to all of the amino acid residues T6, R7, M10, R11, M14, R17, R18, R20, and R21 in DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328).\n\n**66.** An antibody that binds to human mutant calreticulin (CALR), wherein the antibody binds two non-identical epitopes within DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328).\n\n**67.** An antibody that binds to human mutant calreticulin (CALR), wherein the antibody binds to human mutant CALR at a first epitope within DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328) and at a second epitope within DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328), wherein the first epitope and the second epitope are non-identical.\n\n**68.** The antibody of claim 67, wherein the antibody binds to the first epitope at one or more of the amino acid residues R8, M10, M19, and R20 in DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328) and the antibody binds to the second epitope at one or more of the amino acid residues R7, R11, R17, R20, and R21 in DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328).\n\n**69.** The antibody of claim 67, wherein the antibody binds to the first epitope at all of the amino acid residues R8, M10, M19, and R20 in DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328) and the antibody binds to the second epitope at all of the amino acid residues R7, R11, R17, R20, and R21 in DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328).\n\n**70.** The antibody of claim 67, wherein the antibody binds to the first epitope at one or more of the amino acid residues D1, R5, R8, M9, M10, R11, T12, M14, R15, M16, M19, and R20 in DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328) and the antibody binds to the second epitope at one or more of the amino acid residues T6, R7, M10, R11, M14, R17, R18, R20, and R21 in DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328).", "# My Work Driving Innovation in Precision Oncology Through Structure-Based Drug Design\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-09-08 on marcdeller.com (categories: project).*\n\n6+ FDA IND filings across oncology, immunology, and virology, anchored by structure-based drug design against some of the toughest targets in precision medicine. I delivered the structural biology behind [INCB161734](https://aacrjournals.org/cancerres/article/84/6_Supplement/5900/740558/Abstract-5900-INCB161734-A-novel-potent-and-orally), a potent and orally bioavailable KRAS-G12D inhibitor with 80-fold mutant selectivity ([AACR 2024 poster](https://www.incytemi.com/document/Poster/AACR%202024%20-%20KRASG12D%20mutant%20poster.pdf)) now in [Phase 1 trials](https://www.clinicaltrials.gov/study/NCT06179160), alongside co-crystal structures and SBDD support for clinical-stage programs targeting [CDK2](https://www.clinicaltrials.gov/study/NCT07023627), [FGFR2/3](https://clinicaltrials.gov/study/NCT06906562), the first-in-class [mutant CALR antibody INCA033989](https://www.clinicaltrials.gov/study/NCT05936359), [JAK1](https://www.incytemi.com/document/Poster/ACS%20spring%202025_Burns_Povorcitinib%20Discovery.pdf), and [JAK2-V617F](https://www.clinicaltrials.gov/study/NCT06213818). Together, these programs integrate AI-driven design with high-throughput X-ray crystallography and cryo-EM to compress lead-optimization timelines and accelerate precision medicines to patients.\n\nMovie of KRAS-G12D bound to INCB161734\n\n* **KRAS-G12D Precision Targeting:**  INCB161734 shows 34% response in relapsed pancreatic cancer, leveraging 80\u00d7 selectivity for mutant KRAS and advancing in combination trials.\n* **CDK2-Driven Ovarian Cancer:**  INCB123667 delivers 33% responses in platinum-resistant cases with cyclin E1 amplification, now entering the MAESTRA-2 pivotal study.\n* **FGFR2/3 Resistance Control:** Structure-guided inhibitors overcome FGFR gatekeeper mutations (V564F, V555M), addressing resistance across ~36% of FGFR2-altered cancers.\n* **Mutant CALR Immunotherapy:** INCA033989 achieves 86% response in essential thrombocythemia by eradicating mutant CALR\u207a stem cells, indicating curative potential in ongoing trials.\n* **JAK2-V617F Molecular Remission:** INCB160058 targets the pseudokinase domain with 2,500\u00d7 mutant selectivity, eliminating JAK2-V617F clones while preserving normal blood cell formation.\n\n* [**Discovery of Povorcitinib (INCB054707):**  An Isoform Selective Janus Kinase 1 (JAK1) Inhibitor.](https://www.incytemi.com/document/Poster/ACS%20spring%202025_Burns_Povorcitinib%20Discovery.pdf)\n\n**Advancing targeted therapies from molecular insights to clinical breakthroughs across multiple cancer-driving mutations**\n\nPrecision oncology has undergone a remarkable transformation over the past five years, driven by the integration of [structure-based drug design](https://marcdeller.com/advancing-precision-oncology-a-deep-dive-into-four-breakthrough-targeted-therapies-transforming-cancer-treatment/) with advanced molecular profiling and biomarker-driven patient selection strategies. My work at the intersection of structural biology and drug discovery has contributed to developing highly selective inhibitors targeting oncogenic drivers that were once considered undruggable, translating fundamental protein science into clinical candidates that address critical unmet needs in cancer treatment. This journey spans [KRAS-G12D mutations](https://marcdeller.com/discovery-of-incb159020-an-orally-bioavailable-kras-g12d-inhibitor/) in pancreatic and colorectal cancers, [CDK2 dependencies](https://marcdeller.com/discovery-of-4-pyrazolyl-2-aminopyrimidines-as-potent-and-selective-cdk2-inhibitors/) in platinum-resistant ovarian cancer, [FGFR2/3 alterations](https://marcdeller.com/discovery-of-potent-and-selective-inhibitors-of-wild-type-and-gatekeeper-mutant-fgfr2-3/) across multiple solid tumors, and [mutant calreticulin](https://marcdeller.com/anti-mutant-calreticulin-calr-antibodies-and-uses-thereof/) in myeloproliferative neoplasms\u2014each representing distinct challenges that required tailored structure-guided approaches to achieve therapeutic selectivity while sparing normal cellular function.\u200b\n\n**KRAS-G12D Inhibitor: Transforming Pancreatic Cancer Treatment**\n\nThe [KRAS-G12D mutation](https://marcdeller.com/discovery-of-incb159020-an-orally-bioavailable-kras-g12d-inhibitor/) represents one of the most prevalent oncogenic drivers in pancreatic ductal adenocarcinoma, occurring in approximately 45% of cases, as well as 15% of colorectal cancers and a subset of lung cancers. Structure-based design efforts led to the development of [INCB161734](https://clinicaltrials.gov/study/NCT06179160), a potent oral inhibitor demonstrating over 80-fold selectivity for the G12D mutant versus wild-type KRAS through precise exploitation of conformational differences between the mutant and normal protein. In ongoing [Phase 1 clinical trials](https://clinicaltrials.gov/study/NCT06179160), INCB161734 achieved a remarkable 34% objective response rate at the 1,200 mg daily dose in patients with relapsed pancreatic cancer, with 86% disease control rate, representing unprecedented activity in this difficult-to-treat population. Critically, the compound demonstrated continuous near-maximal target engagement in preclinical models, and early clinical data showed 72% of patients treated at 1,200 mg achieved greater than 90% reduction in circulating tumor DNA KRAS-G12D variant allele frequency, providing a molecular surrogate of therapeutic response. The favorable safety profile with manageable predominantly mild gastrointestinal and hematologic adverse events, combined with no dose-limiting toxicities reported, supports advancing [combination strategies](https://marcdeller.com/advancing-precision-oncology-a-deep-dive-into-four-breakthrough-targeted-therapies-transforming-cancer-treatment/) with standard chemotherapy regimens and immunotherapy agents, with pivotal trials anticipated to launch in first-line pancreatic cancer settings in 2026.\u200b", "# 2026 \u2013 From Bicyclic to Tricyclic: A New WRN Inhibitor Patent from Incyte\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-05-18 on marcdeller.com (categories: patents).*\n\n* [Fused Tricyclic Compounds as Inhibitors of WRN](https://patents.google.com/patent/US20260098041A1/en)\n\n* **Tricyclic Evolution:** The new Incyte patent extends the earlier bicyclic pyrimidone scaffold by fusing an additional ring.\n* **Dual Enzymatic Inhibition:** Compounds claim inhibition of both WRN ATPase and helicase activities, the genetically validated mechanism.\n* **MSI-H/dMMR Target Population:** WRN remains synthetic lethal in microsatellite-unstable cancers, including those refractory to checkpoint blockade.\n* **Expanded Vector Space:** Three-dimensional scaffold geometry supports selectivity optimization against related RecQ helicases BLM, RecQ1, RecQ4, and RecQ5.\n\n1. **Fused bicyclic pyrimidones as WRN inhibitors (tricyclic series):** [Zhao et al., US Patent Application, 2026](https://patents.google.com/patent/US20260098041A1/en)\n2. **Fused bicyclic pyrimidones as WRN inhibitors (bicyclic series):** [Shepard et al., US Patent Application, 2025](https://patents.google.com/patent/US20250333420A1/en)\n3. **WRN helicase is a synthetic lethal target in microsatellite unstable cancers:** [Chan et al., Nature, 2019](https://www.nature.com/articles/s41586-019-1102-x)\n4. **Discovery of WRN inhibitor HRO761 with synthetic lethality in MSI cancers:** [Ferretti et al., Nature, 2024](https://www.nature.com/articles/s41586-024-07350-y)\n5. **Identification of VVD-214/RO7589831, a Clinical-Stage Covalent Allosteric Inhibitor of WRN Helicase:** [Kikuchi et al., Journal of Medicinal Chemistry, 2025](https://pubs.acs.org/doi/10.1021/acs.jmedchem.5c01805)\n\n**A second Incyte filing extends the fused pyrimidone scaffold into three dimensions for a clinically validated synthetic lethal cancer target**\n\nWerner syndrome RecQ helicase, or WRN, sits at the intersection of cancer genetics and structural biology, a synthetic lethal vulnerability identified through [CRISPR screens](https://www.nature.com/articles/s41586-019-1102-x) in microsatellite instability-high (MSI-H) tumours. Roughly 4% of all cancers carry this signature, with the highest prevalence in endometrial (31%), colorectal (20%), and gastric (19%) disease, and checkpoint blockade leaves most patients still needing additional options. Two clinical candidates now anchor the WRN landscape: Novartis\u2019s [HRO761](https://www.nature.com/articles/s41586-024-07350-y), a non-covalent inhibitor that binds within the helicase domain to lock WRN in an inactive conformation, and Vividion\u2019s [VVD-214](https://pubs.acs.org/doi/10.1021/acs.jmedchem.5c01805), a covalent cysteine-engaging compound that stabilises a similarly inactive state through allosteric engagement. Both validate that selective small-molecule inhibition of an apparently undruggable target class is achievable.\n\nIncyte\u2019s medicinal chemistry programme entered this space with a fused bicyclic pyrimidone series disclosed in [patent application US20250333420A1](https://patents.google.com/patent/US20250333420A1/en) (October 2025), designed to occupy a similar ATP-adjacent pocket through a hydroxy pyrimidine motif coordinating catalytic residues. A more recent filing, [US20260098041A1 published April 2026](https://patents.google.com/patent/US20260098041A1/en), extends that lineage. The disclosed compounds replace the bicyclic core with fused tricyclic pyrimidone scaffolds spanning two general formulae (I and II), and claim dual inhibition of WRN ATPase and helicase activities, the two enzymatic functions most genetically validated as essential in MSI-H cells.\n\n> **\u201cWRN inhibitors have a different mode of action compared to checkpoint inhibitors and will have the potential to complement CPI in the treatment of MSI-H/dMMR cancers.**\u201d [Incyte Corporation, 2026](https://patents.google.com/patent/US20260098041A1/en)\n\nThe chemical evolution from bicyclic to tricyclic is more than cosmetic. Adding a third fused ring raises sp3 character, shifts the molecules further into three-dimensional chemical space, and creates additional vector positions around the Ring A and optional Ring D substituents disclosed in the new claims. That geometry matters when the target is a member of the RecQ helicase family alongside BLM, RecQ1, RecQ4, and RecQ5: selectivity is hard-won, the ATP pocket residues are conserved, and a more sculpted scaffold provides additional opportunities to address selectivity, permeability, and metabolic stability liabilities in parallel.\n\nA tricyclic core also expands the freedom-to-operate footprint as competing WRN programmes accelerate, and offers a credible follow-up chemotype should the initial bicyclic series face development setbacks. For an MSI-H/dMMR patient population that remains underserved despite checkpoint inhibition, having structurally distinct, mechanistically related candidates progressing in parallel is a meaningful insurance policy. Having contributed to the WRN programme at Incyte, I find this scaffold an instructive example of how iterative scaffold-hopping inside a structure-driven campaign continues to widen the design space around a target that, only six years ago, looked intractable.", "# GitHub README: ALPHABETTI\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ALPHABETTI, 2026\n\n**The fold does not run on the droplet, and cannot.** That box has 3.9 GB of RAM with about 2.0 GB free, two cores, and eight other applications already on it; the `esmfold_v1` checkpoint alone is 7.9 GB of fp32 weights, and a 400-residue fold needs several more gigabytes for the L x L pair representation. It is short by a factor of four to six.\n\nSo the Space returns nothing but coordinates and probabilities, and **every derived quantity is computed in this repository** \u2014 information content, entropy, variant scores, accessibility, secondary structure, per-residue frames \u2014 in modules that need neither a GPU nor a network and are unit-tested without one. The science does not move when the compute does.\n\nThat decision also removed Redis and RQ. With no models to hold, a job spends its entire life waiting on one HTTPS request, which is what threads are for; job state lives in SQLite so both Gunicorn workers can see it.\n\n## \u2699\ufe0f Options\n\nEvery setting is an environment variable. See `.env.example`.\n\n| Variable | Default | What it does |\n|---|---|---|\n| `ALPHABETTI_FOLD_BACKEND` | `hf_space` | `hf_space`, `local` (torch in-process, needs ~16 GB), or `none` (cache only) |\n| `ALPHABETTI_HF_SPACE` | `Dellboy/alphabetti-fold` | The Space that holds the models |\n| `HF_TOKEN` | *(none)* | Required in practice: the anonymous ZeroGPU quota is a handful of folds and runs out with an error that does not mention quota |\n| `ALPHABETTI_MASKED_MARGINALS` | `1` | Masked marginals; `0` gives the faster, misleading unmasked pass |\n| `ALPHABETTI_MAX_LENGTH` | `400` | Longer input is refused with an offer to truncate, naming the exact range kept |\n| `ALPHABETTI_MIN_LENGTH` | `10` | |\n| `ALPHABETTI_FOLD_TIMEOUT` | `900` | Seconds before a fold is given up on cleanly |\n| `ALPHABETTI_MAX_CONCURRENT` | `2` | Concurrent folds; ZeroGPU serialises anyway and quota is per account |\n| `ALPHABETTI_TOP_K` | `6` | Amino acids per position carried in the payload |\n| `ALPHABETTI_CACHE` | `instance/alphabetti.sqlite` | Cache database |\n| `ALPHABETTI_ENABLE_UNIPROT` | `1` | Accession and entry-name lookup |\n| `ALPHABETTI_ENABLE_VARIANTS` | `1` | EBI Proteins API overlay for BALDERDASH |\n\n## \ud83d\udce4 Outputs\n\n| Format | Route | Notes |\n|---|---|---|\n| PDB | `/api/download/<id>.pdb` | The predicted structure, pLDDT in the B-factor column |\n| JSON | `/api/download/<id>.json` | The full payload that drives all four tabs |\n| CSV | `/api/download/<id>.csv` | Per residue: resnum, aa, pLDDT, RSA, SASA, SS, bits, entropy, surprise |\n| GLB | in-app | Baked instanced geometry with vertex colours, for Blender or AR |\n| STL | in-app | For printing; letters are separate shells and the app says so |\n| PNG | in-app | 1x, 2x, 4x, with a transparent-background option |\n| GIF | in-app | One FOLDEROL loop, with a resolution selector |\n\n## \ud83d\ude80 Running it\n\n```bash\ngit clone https://github.com/bellcheddar/ALPHABETTI.git\ncd ALPHABETTI\npython3 -m venv .venv\n.venv/bin/pip install -r requirements.txt\ncp .env.example .env          # add HF_TOKEN\n.venv/bin/python app.py       # http://127.0.0.1:8006\n```\n\nThe four examples are committed as computed payloads, so the app shows a rotating protein before it has ever reached a GPU.\n\n```bash\n.venv/bin/python -m pytest tests/ -q          # 49 tests, no GPU, no network\n.venv/bin/python scripts/prewarm.py           # recompute the examples\n.venv/bin/python scripts/make_typeface.py /path/to/font.ttf out.json 700\n.venv/bin/python scripts/make_icon.py         # icon, favicon, OG card\n```\n\nDeployment: `deploy/provision.sh` once, `deploy/deploy.sh` thereafter.\n\n## \u2705 To Do", "# Hidden Chemistry in the Protein Data Bank: What Else Are We Missing?\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: blog).*\n\n* Recent research using new machine learning algorithms has uncovered previously undetected nitrogen-oxygen-sulfur (NOS) linkages in archived protein structures from the Protein Data Bank.\n* This discovery reveals that even well-studied proteins can hold surprising chemical secrets that were missed by conventional analysis methods.\n* The finding highlights the potential for modern computational tools to unlock new biological insights from existing data, prompting a re-examination of established knowledge.\n\n1. A new complexity in protein chemistry: Algorithm uncovers previously undetected bonds: <https://phys.org/news/2025-05-complexity-protein-chemistry-algorithm-uncovers.html>\n2. Revealing arginine-cysteine and glycine-cysteine NOS linkages by systematic re-evaluation of the PDB: <https://www.nature.com/articles/s42004-025-01535-w>\n3. An overlooked bonding motif appears in many more proteins than was previously realised: <https://www.chemistryworld.com/news/an-overlooked-bonding-motif-appears-in-many-more-proteins-than-was-previously-realised/4021616.article>\n\nThe Protein Data Bank (PDB) is one of science\u2019s crown jewels\u2014a vast, public archive containing the 3D structures of hundreds of thousands of proteins, meticulously determined over decades. For structural biologists and drug designers, it has been the foundational blueprint for understanding life and disease. We have long operated under the assumption that these blueprints, while constantly growing, were largely complete for the structures they held. We believed we had seen and cataloged all the fundamental ways proteins are built. A recent discovery, however, shows that even in this well-explored territory, surprising secrets remain hidden in plain sight.\n\nResearchers at the University of G\u00f6ttingen developed a new computational pipeline, integrating machine learning and quantum mechanics to systematically re-analyze over 86,000 structures in the PDB. They were hunting for subtle chemical bonds that might have been missed by conventional analysis tools. The algorithm uncovered something startling: a previously undetected type of covalent bond, a nitrogen-oxygen-sulfur (NOS) linkage, hiding in numerous proteins. This overlooked chemical feature had been present in the data for years, waiting for the right tools to bring it to light.\n\nThis discovery fundamentally expands our understanding of protein chemistry and function, revealing a new layer of complexity. It serves as a powerful reminder that our knowledge is only as good as the tools we use to acquire it. As we develop more sophisticated AI and machine learning algorithms, what other hidden chemistry lies dormant in our archives? This finding is not just about a new bond; it\u2019s a call to action for the entire scientific community to revisit our most trusted datasets with fresh eyes and new technology. The next great breakthrough might not come from a new experiment, but from uncovering what we\u2019ve already collected but have not yet seen.", "# 2015 \u2013 Models of protein\u2013ligand crystal structures: trust, but verify\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-09-01 on marcdeller.com (categories: papers).*\n\nClick to enlarge\n\n* [Link to Paper](https://marcdeller.com/wp-content/uploads/2025/09/nihms888398.pdf)\n\n* **Model Fidelity:** Evaluate ligand fit using cLDDT and pocket RMSD before analysis.\n* **Dynamic Refinement:** Integrate MD simulations or flexible fitting to correct misaligned ligands.\n* **AI Validation:** Employ PoseBusters or similar tools to screen for steric clashes and unrealistic geometries.\n* **Data Curation:** Utilize workflows like HiQBind-WF to standardize hydrogen placement and bond orders.\n* **Confidence Metrics:** Leverage predicted confidence scores (e.g., plDDT) to filter robust models.\n* **Continuous Verification:** Reassess published models when new validation methods emerge.\n\n* **Models of protein\u2013ligand crystal structures: trust, but verify:** [MC Deller, *J. Comput Aided Mol Des*. 2015](https://pmc.ncbi.nlm.nih.gov/articles/PMC4531100/pdf/nihms662686.pdf) September\n* **DynamicBind: predicting ligand-specific protein-ligand complex structure with a deep equivariant generative model:** [Wei Lu, Chen, et al., Nat Commun, 2024](https://pubmed.ncbi.nlm.nih.gov/38316797/)\n* **Modeling ligands in cryo-EM with generative AI and density-guided simulations:** [N. Haloi, et al., bioRxiv, 2025](https://www.biorxiv.org/content/10.1101/2025.02.10.637508v1)\n* **Interformer: an interaction-aware model for protein\u2013ligand docking:** [H. Lai, et al., Nat Commun, 2024](https://pubmed.ncbi.nlm.nih.gov/39587070/)\n* **Structure prediction of protein-ligand complexes from sequence information:** [P. Bryant, et al., Nat Commun, 2024](https://pubmed.ncbi.nlm.nih.gov/38806453/)\n* **Fast Quantitative Validation of 3D Models of Low-Affinity Protein\u2013Ligand Complexes:** [M. Rossi, et al., J. Med. Chem., 2024](https://pubmed.ncbi.nlm.nih.gov/38848103/)\n* **A workflow to create a high-quality protein-ligand binding dataset for training, validation, and prediction tasks:** [Y. Wang, et al., Digit Discov., 2025](https://pubmed.ncbi.nlm.nih.gov/40190768/)\n\n**Trust, but Verify: Unraveling the Fidelity of Protein\u2013Ligand Crystal Models**\n\nHigh-resolution crystal structures of protein\u2013ligand complexes are invaluable for drug discovery and mechanistic biochemistry, yet they can harbor hidden discrepancies that compromise downstream analyses. Recent advances in validation algorithms and AI-driven refinement underscore the imperative to scrutinize every model before placing unqualified trust in it.\n\nAccurate protein\u2013ligand models underpin rational drug design by revealing key intermolecular contacts and guiding lead optimization. However, even high-resolution structures may exhibit steric clashes, misplaced ligands, or unrealistic geometries that escape cursory inspection. Modern validation metrics\u2014such as real-space correlation coefficients, pocket root-mean-square deviation, and contact-LDDT scores\u2014offer quantitative lenses through which to assess model quality. Integrative methods that combine crystallographic data with molecular dynamics simulations or AI-guided flexible fitting further refine ligand poses, correcting misfits that static refinement may miss.\n\nBetween 2024 and 2025, several breakthroughs have emerged:\n\n* **DynamicBind** (Nat Commun 15, 1071, 2024) leverages an equivariant generative network to predict ligand poses from AlphaFold protein conformations, achieving pocket RMSDs < 2 \u00c5 and cLDDT scores > 0.8 across Drug Target families.\n* **Generative AI integrated** with cryo-EM density-guided MD (bioRxiv 2025.02.10.637508v1) demonstrates 82\u201395% ligand pose accuracy against ten EMDB entries, highlighting the value of flexible fitting to resolve suboptimal initial models.\n* **Interformer** (Nat Commun, Nov 2024) employs an interaction-aware mixture density network to capture non-covalent contacts, boosting top-1 docking success rates to 63.9% and reducing steric-clash failures compared to DiffDock and GNINA.\n* **Umol** (Nat Commun 2024) predicts fully flexible, all-atom protein\u2013ligand complex structures directly from sequence, offering confidence metrics (plDDT) that correlate with binding strength.\n\nThese innovations strengthen confidence in structural models but also illustrate that verification remains critical. Automated validation pipelines\u2014such as PoseBusters and HiQBind-WF\u2014curate and correct PDBbind entries, flagging rare elements, steric clashes, and bond-order errors to yield datasets of >18 000 high-quality complexes.\n\n| Concept | Description | Key Reference |\n| --- | --- | --- |\n| cLDDT score | Fraction of conserved ligand\u2013protein contacts at multiple distance tolerances. | [Y. Chen, et al., Nat Commun, 2024](https://www.nature.com/articles/s41467-024-45461-2) |\n| Pocket RMSD | RMSD of binding-pocket atoms within 5 \u00c5 of ligand after alignment. | [Y. Chen, et al., Nat Commun, 2024](https://www.nature.com/articles/s41467-024-45461-2) |\n| Generative AI fitting | Flexible fitting of ligands into density maps via AI and MD. | [A. Singh, et al., bioRxiv, 2025](https://www.biorxiv.org/content/10.1101/2025.02.10.637508v1.full-text) |\n| Interaction-aware docking | Mixture density networks modeling non-covalent interactions for docking. | [L. Zhao, et al., Nat Commun, 2024](https://www.nature.com/articles/s41467-024-54440-6) |\n| plDDT metrics | Predicted confidence metrics correlating with binding strength. | [P. Gupta, et al., Nat Commun, 2024](https://www.nature.com/articles/s41467-024-48837-6) |\n| HiQBind-WF workflow | Semi-automated curation and refinement of protein\u2013ligand complexes. | [S. Tan, et al., RSC Adv., 2025](https://pubs.rsc.org/en/content/articlehtml/2025/dd/d4dd00357h) |", "# Optimizing Data Volume and Diversity for Robust Antibody\u2013Antigen \u0394\u0394G Prediction\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-27 on marcdeller.com (categories: blog, sb).*\n\n* **Volume Expansion:** Hundreds of experimental \u0394\u0394G measurements must scale to hundreds of thousands for reliable training.\n* **Diversity Balancing:** Sequence, structural, and mutation diversity prevents overfitting and enhances model generalization.\n* **Synthetic Data Utility:** Computationally generated \u0394\u0394G values boost dataset size but need pairing with experimental diversity.\n* **Embedding Strategies:** Combining semantic embeddings with residue-level features improves sequence-only prediction accuracy.\n* **Transfer Learning Leverage:** Large-scale stability datasets inform binding \u0394\u0394G models but hinge on structural-model fidelity.\n* **Multitask Frameworks:** Jointly predicting stability and affinity increases throughput and overall prediction robustness.\n\n1. **Investigating the Volume and Diversity of Data Needed for Generalizable Antibody\u2013Antigen \u0394\u0394G Prediction:** Hummer AM, et al., Nature Machine Intelligence: <https://www.nature.com/articles/s43588-025-00823-8>\n2. **MVSF-AB: accurate antibody\u2013antigen binding affinity prediction via multi-view sequence feature learning:** Li M, et al., Bioinformatics: <https://doi.org/10.1093/bioinformatics/btae579>\n3. **Transfer learning to leverage larger datasets for improved prediction of protein stability and protein\u2013protein binding affinity:** Meier D, et al., Proceedings of the National Academy of Sciences: <https://www.pnas.org/doi/10.1073/pnas.2314853121>\n4. **AttABseq: an attention-based deep learning prediction method for antigen\u2013antibody binding \u0394\u0394G:** Zhang Y, et al., Briefings in Bioinformatics: <https://academic.oup.com/bib/article/25/4/bbae304/7705533>\n5. **Reliable prediction of protein\u2013protein binding affinity changes upon mutations (Pythia-PPI):** Zhao L, et al., PMC: <https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12199698/>\n\nAchieving generalizable \u0394\u0394G prediction for antibody\u2013antigen interactions requires not only orders of magnitude more experimental data but also careful curation to maximize sequence, structural, and mutational diversity.\n\nA cutting-edge \u0394\u0394G predictor demands extensive and varied training examples. Recent work demonstrates that models like Graphinity, which leverages an equivariant graph neural network, hit performance ceilings when trained on only a few hundred experimental measurements; augmenting with nearly one million FoldX\u2010generated values improved correlation to 0.9 but exposed limitations in representativeness. Synthetic datasets can reveal how model accuracy scales with data size, yet without balanced representation of different antibody lineages, CDR variations, and mutation contexts, overfitting persists and generalization falters.\n\nSequence-based approaches such as MVSF-AB integrate semantic embeddings from proteinBERT and residue-level features, outperforming earlier methods on natural and mutant antigen\u2013antibody pairs by mitigating structural-data scarcity through multi-view learning. However, their accuracy still hinges on training-set heterogeneity and may drop for novel antibody frameworks. Transfer learning from megascale stability datasets (e.g., ~400 000 reliable \u0394\u0394G\u00b0 points from protease sensitivity assays) has boosted model robustness for protein stability but reveals that structural fidelity of predicted models limits applicability to binding \u0394\u0394G tasks.\n\nAttention-based sequence predictors like AttABseq further highlight that diverse mutation landscapes\u2014single and multi-point changes across >32 complexes\u2014are essential to sustain prediction stability (PCC up to 0.587) and prevent performance plateaus seen with homogeneous datasets. Multitask frameworks (e.g., Pythia-PPI) trained on augmented SKEMPI-like interfaces (~400 000 mutations) illustrate that integrating protein stability and binding-affinity predictions can increase throughput (>10 000 predictions/min) while maintaining high correlation (r=0.785).\n\nUltimately, experimental \u0394\u0394G measurements spanning multiple logs of affinity and encompassing diverse antibody scaffolds, antigen epitopes, and mutation types remain critical. Strategic dataset expansion\u2014balancing structure-derived, sequence-based, and synthetic entries\u2014emerges as the key to unlocking truly generalizable \u0394\u0394G prediction for therapeutic antibody engineering.", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\n**101.** A method of diagnosing a human subject with a myeloproliferative neoplasm, the method comprising obtaining a biological sample from a human subject who has or is suspected of having a myeloproliferative neoplasm and contacting the sample with the antibody of any one of claims 1-85 such that the antibody binds to a mutCALR protein if the mutCALR protein is present in the biological sample.\n\n## International Search Report\n\nThe international search report was completed on 7 March 2023 and mailed on 15 May 2023 by the European Patent Office acting as International Searching Authority. Subject-matter classification: INV. A61K39/00, C07K16/18, C07K16/30, A61P35/00, C07K16/26. Fields searched: A61K, C07K, A61P; databases consulted: EPO-Internal, BIOSIS, Sequence Search, WPI Data. Cited prior art includes WO 2016/087514 A1 (CeMM Forschungszentrum f\u00fcr Molekulare Medizin GmbH, 9 June 2016), cited against claims 1-7 and 74-101. The search report records a lack-of-unity objection dividing the claims into a large number of inventions corresponding to the individual antibody clones.\n\n## References\n\n*(reference list omitted)*", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n**Chart colour and shape.** The two scatters colour each point by tier via a continuous\ncolourscale + colorbar legend, in the selectivity heatmap's style: confidence tier for\npIC50-vs-confidence-score, affinity tier for pIC50-vs-binder-probability, matching the shield and\nbullseye icons respectively. When a `Ligand:` block sets `Role: agonist`/`Role: antagonist`,\npoints are also shape-coded (circle = agonist, diamond = antagonist), the legend sitting inside\nthe plot area's top-left corner rather than Plotly's default outside-right, where it would\ncollide with the colorbar. Campaigns that don't set `Role:` see a single unshaped trace.\n\n**Embedding.** The dashboard posts its own content height to any parent window via `postMessage`\non load and resize, so a page embedding it in an iframe (e.g. `findings.md`'s \"Interactive\ndashboard\" section) can size the iframe to the actual content -- a cross-origin iframe cannot\notherwise be measured from the embedding page's JS.\n\n## \ud83d\udd2c Ligand validation & scaffold highlighting\n\nTwo related but distinct checks run over every SMILES ligand before you commit hours of\n`boltz predict` time, and both surface in the dashboard. Boltz folds whatever chemistry it is\ngiven: an undefined stereocentre, an unintended protonation state or a stray counterion raises\nno error, it silently changes the predicted pose and affinity. These are exactly the mistakes a\nnon-specialist (or a tired specialist) makes typing SMILES by hand, and they're invisible until\nyou're staring at a confusing result with no idea the input was ever wrong.\n\n### Ligand preparation (validity checks)\n\nAt parse time every ligand SMILES is canonicalised (RDKit) so the same molecule is represented\nconsistently everywhere downstream -- the generated YAML, the summary table, and cif2plip's own\nligand-matching (see the InChIKey-based matching note in [CHANGELOG.md](CHANGELOG.md)). Then,\nboth at `preflight` (as the `ligand_preparation` check) and in the dashboard's \"Ligand\npreparation\" card, each ligand is checked for **undefined stereocentres**\n(`Chem.FindMolChiralCenters(includeUnassigned=True)`: a stereocentre exists but the SMILES\ndoesn't say which enantiomer/diastereomer, so Boltz will fold *some* version of it, possibly not\nthe one you intended), **disconnected fragments** (`Chem.GetMolFrags()` returning more than one\n-- likely a salt or counterion left in the SMILES, e.g. a sodium carboxylate written as two\ncomponents), and **ionizable groups** (SMARTS matches for carboxylic acid, primary/secondary\namine, phenol and sulfonic acid, whose protonation state at physiological pH a plain SMILES\ndoesn't specify -- worth a deliberate choice, not a default assumption).\n\nAll advisory, not a hard failure -- these can be legitimate modelling choices -- but worth a\nsecond look before trusting downstream numbers.\n\n### Scaffold highlighting (the \"Ligand structures\" grid)\n\nThe ligand grid answers a different question: *do any of these ligands share a chemical core?*\nThis matters most for SAR campaigns, where a chemist is testing close analogues on purpose and\nseeing the shared scaffold at a glance (with the parts that differ jumping out) beats reading\neach SMILES individually. Two tiers, in order, and nothing is highlighted unless one of them\nfinds something real:\n\n1. **Exact Bemis-Murcko scaffold match** -- ligands whose ring systems + connecting linkers are\n   chemically identical are grouped, threshold-free. The dominant case for a real SAR series.\n2. **Fallback for near-analogues** -- leftovers are grouped by Morgan/Tanimoto fingerprint\n   similarity, then a maximum common substructure (MCS) is computed across the *whole* group and\n   verified to match every member, so the claim is a proven substructure match rather than an\n   assigned similarity score.\n\nShared fragments below 8 heavy atoms (e.g. \"they all contain a benzene ring\") are deliberately\nnot highlighted. Ligands in the same group also have their 2D depiction aligned to a common\norientation, so the shared core is drawn in the same position across cells and visually snaps\ntogether.\n\n**What's highlighted and how**, directly on each rendered structure -- the same badges shown on\neach cell and spelled out in the panel's legend:\n\n| Badge | Colour | Meaning |\n|---|---|---|\n| `S` | \ud83d\udfea Magenta | Undefined stereocentre (RDKit also draws its own `(?)` marker at the atom) |\n| `A` | \ud83d\udfe7 Amber | Carboxylic acid -- protonation state not specified |\n| `N` | \ud83d\udfe7 Amber | Primary/secondary amine -- protonation state not specified |\n| `Ph` | \ud83d\udfe7 Amber | Phenol -- protonation state not specified |\n| `SO3` | \ud83d\udfe7 Amber | Sulfonic acid -- protonation state not specified |\n| `salt` | \ud83d\udfe5 Red | Disconnected fragment (salt/counterion) -- flagged on the border and badge only, not atom-highlighted (there's no single meaningful atom to point at) |\n| -- | one of six colour-blind-safe palette colours | Atoms in a shared scaffold/substructure -- consistent per group, with a legend entry naming the group and how many ligands share it (e.g. \"shared scaffold -- 3/5 ligands\") |\n\nThe legend is captioned badge by badge, stating what was found and on how many ligands -- never\nan unexplained highlight. A specific finding (stereocentre, ionizable group) takes priority over\nthe softer scaffold highlight where they overlap on an atom, being the more actionable signal.\nIf no ligand shares a real scaffold with any other, the panel says so plainly (\"no shared\nscaffold or substructure detected\") rather than forcing a highlight onto something coincidental.\nCCD-code ligands have no SMILES to render and show a plain placeholder instead of an empty cell.\n\n## \ud83c\udfaf Ligand pose vs experiment: the prediction against reality", "# BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n| Example | Demonstrates | Dashboard | Input |\n| --- | --- | --- | --- |\n| `t4_lysozyme` | One protein (T4 lysozyme L99A, UniProt P00720) + one ligand (benzene). No partners, no pocket_contacts. The minimal shape; smallest/fastest smoke test. | [boltz_dashboard.html](https://bellcheddar.github.io/BoltzMaker/examples/t4_lysozyme/boltz_dashboard.html) | [boltz_input.md](https://github.com/bellcheddar/BoltzMaker/blob/main/examples/t4_lysozyme/boltz_input.md) |\n| `egfr_covalent` | EGFR kinase domain (UniProt P00533) + a generic covalent fragment, linked via `bond_constraints` at Cys797. Demonstrates covalent-linkage modelling. | [boltz_dashboard.html](https://bellcheddar.github.io/BoltzMaker/examples/egfr_covalent/boltz_dashboard.html) | [boltz_input.md](https://github.com/bellcheddar/BoltzMaker/blob/main/examples/egfr_covalent/boltz_input.md) |\n| `adrb2_gs_panel` | Beta-2 adrenergic receptor (UniProt P07550), agonist vs antagonist, as two separate `Protein:` blocks sharing one sequence rather than one family crossed with both ligands: the agonist target co-folds a Gs alpha partner (UniProt P63092), the antagonist target doesn\u2019t (Gs only forms a stable complex with the active, agonist-bound receptor in reality \u2014 co-folding it with the antagonist too made Boltz predict a near-identical active-like fold for both, 0.38 Angstrom apart; splitting them out gets a real conformational difference, 1.28 Angstrom apart, TM6 shift roughly doubled for the agonist). Demonstrates `compare-sse` (see below) and why co-folded partners should match each ligand\u2019s real biology, not just get crossed with everything. | [boltz_dashboard.html](https://bellcheddar.github.io/BoltzMaker/examples/adrb2_gs_panel/boltz_dashboard.html) | [boltz_input.md](https://github.com/bellcheddar/BoltzMaker/blob/main/examples/adrb2_gs_panel/boltz_input.md) |\n| `5ht2_gq_panel` | Three serotonin receptors (5-HT2A/2B/2C, UniProt P28223/P41595/P28335), each with a real agonist/antagonist pair (Psilocin/Risperidone, LSD/Balovaptan, Lorcaserin/SB-242084), each predicted both with and without the Gq heterotrimer (GNAQ+GNB1+GNG2) co-folded \u2014 a 3x2x2 panel, plus a native ligand-free (`Ligands: none`) apo target per receptor, used as each receptor\u2019s `compare-sse` reference since no genuinely apo experimental structure exists for any of the three (checked entity-by-entity across all 59 deposited structures). Demonstrates `Ligands: none`, a larger size-heterogeneous campaign in one manifest, Apple Silicon MPS support for large multi-chain complexes (see below), and `compare-sse` against a *predicted* rather than experimental apo reference \u2014 TM6 centroid shift comes out consistently larger for the Gq-bound targets than their no-Gq counterparts across all three receptors, the expected activation signal (see [findings.md](https://github.com/bellcheddar/BoltzMaker/blob/main/examples/5ht2_gq_panel/findings.md) for the full statistical write-up). | [boltz_dashboard.html](https://bellcheddar.github.io/BoltzMaker/examples/5ht2_gq_panel/boltz_dashboard.html) | [boltz_input.md](https://github.com/bellcheddar/BoltzMaker/blob/main/examples/5ht2_gq_panel/boltz_input.md) |\n\n**Verified end-to-end** (`generate` -> `preflight` -> real `boltz predict` -> `analyze`, including cif2plip interaction analysis) on an Apple M1 Max, 64GB:\n\n| Example | Targets | Total time | Result |\n| --- | --- | --- | --- |\n| `t4_lysozyme` | 1 | 3m 16s | confidence 0.98, pIC50 8.9, 7 hydrophobic contacts matching the known L99A cavity residues |\n| `egfr_covalent` | 1 | 5m 31s | confidence 0.92, covalent Cys797 SG-to-fragment bond confirmed at 1.75 Angstrom |\n| `adrb2_gs_panel` | 2 | 1h 28m 36s (`ADRB2_ISO1`, agonist+Gs) + 25m 6s (`AR2NG_PRO1`, antagonist alone) | confidence 0.79 (`ADRB2_ISO1`) / 0.83 (`AR2NG_PRO1`) |\n| `5ht2_gq_panel` | 15 | 9 small targets (apo + receptor-alone) ~4-5m each; 6 large receptor+Gq-heterotrimer complexes ~43-48m each | All 15 completed successfully (12 ligand-bound + 3 apo); confidence 0.66-0.81, iPTM up to 0.99 for the ligand-bound complexes |\n\nRun time scales with complex size, not just target count: `ADRB2_ISO1`\u2018s two-chain receptor+Gs-partner complex took disproportionately longer than the single-chain examples, since attention-style operations scale worse than linearly with sequence length \u2014 confirmed directly by `AR2NG_PRO1` (same receptor, no partner) finishing in a fraction of the time. One contributing factor on Apple Silicon specifically: `torch.linalg.svd` (used in the diffusion step) has no MPS implementation and silently falls back to CPU \u2014 worth budgeting for on large multi-chain campaigns.\n\n`5ht2_gq_panel`\u2018s six large 4-chain receptor+Gq-heterotrimer targets (~1250-1280 tokens) originally crashed on Apple Silicon: boltz\u2019s triangular attention computes the full row-wise QK^T score matrix for the whole complex in one unchunked matmul, which exceeds MPS\u2019s single-tensor size ceiling past roughly 1250 residues and crashes the process inside PyTorch\u2019s internal tiled-bmm fallback. Each row\u2019s attention is independent, so chunking along that axis is exact, not an approximation \u2014 `setup` now patches this directly into the installed `boltz` package (idempotent, and checked against boltz\u2019s exact source so a future upgrade can\u2019t be silently mis-patched). `run` also wraps `boltz predict` with `caffeinate` automatically (macOS only, silently skipped if unavailable) as general sleep-prevention hygiene for long GPU jobs. All 15 `5ht2_gq_panel` targets, including the six large complexes, now complete successfully \u2014 see <CHANGELOG.md> for the fix in full.\n\n## \ud83e\udded boltz_input.md format", "# Structure of Hepatitis C Virus Envelope Glycoprotein E2 Antigenic Site 412 to 423 in Complex with Antibody AP33\nAuthors: Kong, L.; Giang, E.; Nieusma, T.; Robbins, J.B.; Deller, M.C.; Stanfield, R.L.; Wilson, I.A.; Law, M.\nSource: Journal of Virology, 2012\n\nInterestingly, these similar solutions were realized using very different antibody complementarity-determining region (CDR) loops (Fig. 2E and F). For the MAb AP33, all CDR loops except L2 are involved in the interactions (Fig. 2E). For the MAb HCV1, CDR H2, H3, and L3 loops account for most interactions, although framework region (FR) 2 and L1 also contribute to some interactions (Fig. 2E). Also, MAb HCV1 binding relies on an insertion in CDR H3 that is not present in MAb AP33 (Fig. 2F). The lack of CDR L2 involvement in antigen binding for both MAbs is not unusual for small-molecule or linear peptide interactions. The MAb AP33, on the other hand, has a longer L1, with inserted residues immediately upstream of a contacting residue in the CDR L1 loop of the paratope (Fig. 2F).\n\nThe structure also helps explain how different mutations escape the MAb AP33. In addition to hydrophobic interactions with E2 Leu413 and Trp420, binding of the MAb AP33 also requires hydrogen bonding to Asn415 and Gly418 (Fig. 1C and 3A), thus explaining some of the known escape mutants of the antibody (N415Y and N415D). AP33 also interacts with Gly418 and may not accommodate substitutions with bulkier side chains, such as those for the known escape mutant G418D.\n\nInterestingly, the antibody forms two hydrogen bonds with both the main chain and the side chain of Asn415 (Fig. 1C; see also Table S4 in the supplemental material), whereas only one hydrogen bond to the main chain of that residue is found in the HCV1 structure. Consequently, the MAb AP33 does not bind mutants with Asn415 replaced by any of the residues found in natural variants (Fig. 3B) and does not neutralize infectious pseudotype virus harboring these mutations, in contrast to HCV1 (Fig. 3D). In addition, an E655G mutation has been reported to help the virus to escape the MAb AP33. E2 Glu655 is found in 74% of known HCV isolates, and the E655G mutation does not appear to be directly involved in forming the AP33 epitope (Fig. 3C). Instead, the mutation enhances viral infectivity (Fig. 3D), which may help compensate for reduced viral infectivity due to mutations in the conserved AP33 epitope.\n\nThe comparative study of two independent structures of the conserved E2 antigenic site 412 to 423 further defines it as a prime vaccine target and demonstrates how antibodies of different genetic origins adopt similar solutions to recognizing this type-I\u2032 \u03b2-hairpin on E2 for neutralization. The structures provide useful information for the design of immunogens that will optimally present this site of vulnerability. An immunogen directing neutralizing antibody responses to this conserved antigenic site may form the basis of a broadly effective HCV vaccine.\n\n**Protein structure accession number.** The atomic coordinates have been deposited in the Protein Data Bank under PDB identification (ID) code 4G6A.\n\n## Figure captions\n\n*Figure 1.* Crystal structure of the MAb AP33 in complex with its HCV E2 epitope. (A) The overall structure of the AP33 complex is shown with a cartoon representation. The peptide epitope (red) is bound between the heavy (dark gray) and light (light gray) chains of the Fab. (B) The adaptive Poisson-Boltzmann solver (APBS; http://www.poissonboltzmann.org/apbs/) was used to calculate the surface potential across the solvent-accessible surfaces of both the paratope (top) and the epitope (bottom) (surface potential from \u22123 kT/e [red] to +3 kT/e [blue] for comparison with the HCV1 antibody interaction). For the peptide, the surface potential is shown looking from above the antibody (top) and for the peptide epitope (middle) viewed from the paratope and a 180\u00b0 rotation below (bottom). (C) Residues of the epitope that form hydrogen bonds with the epitope are shown with a ball-and-stick representation. Hydrogen bonds are displayed as dashed lines. (D) Backbone hydrogen bonding that stabilizes the \u03b2-hairpin of the epitope is shown.\n\n*Figure 2.* Comparison of the MAb AP33 and HCV1 binding to the antigenic region. (A) Different angles of approach for AP33 and HCV1 are shown, with the superposed antigenic regions as flat strands and the antibodies as C\u03b1 ribbon traces. The angle was calculated between the VL-VH pseudo 2-fold axes in each Fab. (B) The orientations of the bound antigenic region in the AP33 (top) and HCV1 (bottom) structures are shown as ribbons relative to the combining regions. The antibody is presented as a transparent Connelly molecular surface, with the CDR loops shown as ribbons. (C) The backbone atoms of the antigenic regions in the AP33 and HCV1 structures were superimposed and are shown with a ball-and-stick representation. Larger differences are observed on the C-terminal side, which interacts with different portions of the antibodies in the two structures. (D) Burial of residues on the E2 epitope by AP33 and HCV1. (E) Side-by-side comparison of CDR loop usage of AP33 (left) and HCV1 (right), shown by buried surface area. (F) All contacts between the epitope and the CDR loops of AP33 and HCV1 are shown. The sequences of the CDR loops are aligned, and the specific E2 residues interacting with them are highlighted in red (AP33) and green (HCV1).\n\n*Figure 3.* Escape mutants of MAb AP33. Binding of MAb AP33 to alanine scanning mutants of E2 region 412 to 423 (A) or mutants with a substitution at position 415 (B) or 655 (C). (D) Resistance to MAb AP33 by mutations leading to loss of binding. The antibody used in this study (huAP33) is a chimeric form of the MAb, with the murine variable domains fused to human kappa LC and IgG1 HC constant regions. Functional verification of this chimeric MAb is provided in Fig. S1 in the supplemental material.\n\n## Acknowledgments", "# Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly\nAuthors: Deller, M.C.; Johnson, H.A.; Miller, M.D.; Spraggon, G.; Elsliger, M.-A.; Wilson, I.A.; Lesley, S.A.\nSource: PLoS ONE, 2015\n\nThe crystal structure of the TM1088 octameric assembly was determined to 3.45 \u00c5 using molecular replacement (Fig 1C and Table 1, PDB ID: 3L4B). This resolution is similar to other K+ gating ring assemblies which have been reported to 2.7 \u00c5 and 3.5 \u00c5 (KtrAB, PDB ID: 2HMS and 4J7C) [20], 2.79 \u00c5 (MthK, PDB ID: 2FY8) [16,18], 3.61 \u00c5 and 3.1 \u00c5 (BK, PDB ID: 3U6N and 3NAF) [22,24], 3.7 \u00c5 (GsuK, PDB ID: 4GX5) [25] and 3.80 \u00c5 and 3.05 \u00c5 (KtrA/TrkH, PDB ID: 4J9U and 4J9V) [21]. The octameric TM1088 assembly has significant sequence and structural homology with the single-subunit TrkA gating ring assembly of V. parahaemolyticus (PDB ID: 4J9U) [21]. Specifically, TM1088A and TM1088B share 29% and 22% sequence identity with TrkA, respectively, as well as considerable structural homology for the individual domains. However, all atom superpositions of the octameric TM1088 assembly with the TrkA assembly results in an RMSD of 9.0 \u00c5 over 5136 atoms (Fig 1E) due to differences in domain orientations between the two structures. The overall similarity in quaternary structure would position TM1088 adjacent to the cell membrane in a position suitable for association with its partner protein TM1089 (Fig 1E, TrkH is homologous to TM1089). Using in vivo complementation assays, both TM1088 and TM1089 have been shown to be required for cell survival in low potassium [9]. The high level of structural homology with the TrkA assembly is consistent with TM1088 being a member of the Trk family of transporters, which in turn, is consistent with the annotations in all the major protein family databases including Pfam [46].\n\n### Crystal structures of TM1088 proteins\n\nThe crystal structures of full-length TM1088A (Fig 2A\u20132B and Table 1, PDB ID: 2G1U) and the C-terminal domain of TM1088B (Fig 2C\u20132D and Table 1, PDB ID: 3JXO) were determined to 1.50 \u00c5 and 1.55 \u00c5, respectively. The overall fold, dimer-interfaces and arrangement of these higher resolution structures are consistent with the corresponding components in the octameric TM1088 assembly.\n\nTM1088A contains one molecule in the asymmetric unit (domain N2, Fig 2A) that forms a canonical N-terminal Trk domain comprising of a domain-swapped dimer in the crystal lattice (domains N2 and N2', Fig 2B). It is important to note that each chain can be considered a single RCK unit, and for consistency with the literature and previous structural studies, two RCK units form the canonical standard assembly. This distinction is important as other K+ transporters achieve the same quaternary structure as TM1088 proteins using a single polypeptide chain. This same dimeric arrangement is also observed in the octameric TM1088 assembly. TM1088A adopts a Rossmann-like fold with each domain consisting of a central core of six-stranded parallel \u03b2-sheets (\u03b21-\u03b26) flanked by three \u03b1-helices on one side (\u03b13, \u03b14 and \u03b15) and three on the other side (\u03b16', \u03b11 and \u03b12). Each domain also contains a characteristic GXGXXG nucleotide binding site motif (Fig 2A) consistent with a role in nucleotide binding.\n\nSimilarly, TM1088B also adopts a Rossmann-like fold and forms a similar dimer in both the TM1088 complex and the high-resolution structure (Fig 2C). In contrast, each chain of TM1088B consists of two sub-domains; a Trk N-terminal domain (N1 and N1') corresponding to Val3-Gly112 and a TrkA C-terminal domain (C1 and C1') spanning Glu143-Leu215. TM1088B does not possess a dinucleotide binding site (corresponding region shown as a sequence logo in Fig 2C). The C-terminal Trk domain (C1 and C1') was also determined independently to high-resolution and forms a two-layer sandwich structure referred to as an alpha-beta plait (CATH classification 3.30.70, Fig 2D).\n\nThe N1 and C1 Trk domains of TM1088B are connected via an inter-domain linker (Ile113-Leu142) that consists predominantly of \u03b16 of the N-terminal Trk domain (highlighted in Fig 2C) with good electron density throughout (Fig 3A). This region is important for arrangement and dimerization of the C-terminal Trk domains, and in a similar fashion to the structurally equivalent \u03b16 of TM1088A, is responsible for the domain-swapped dimer of TM1088B. The angle between helices \u03b16 of the dimer pairs of TM1088A and TM1088B form what is commonly referred to as the \"dimer hinge\" angle (highlighted by red arrow in Fig 2B). This \"dimer hinge\" angle is central to the gating ring mechanism in the family of RCK proteins.", "# Revolutionizing Drug Discovery: LatentX and the Future of Protein Binder Design\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-22 on marcdeller.com (categories: sb).*\n\n* LatentX generates lab-functional protein binders with picomolar affinities and near-perfect hit rates.\n* The model co-samples protein sequence and 3D structure simultaneously for faster, more accurate design.\n* LatentX creates novel molecules beyond nature\u2019s repertoire, enabling new therapeutic modalities.\n* A no-code web platform enables widespread access to cutting-edge protein design technology.\n* Validated by lab experiments, bridging AI predictions with physical outcomes for reliable drug candidates.\n* Backed by top AI scientists, funded with $50 million, LatentX seeks to make biology programmable for instant drug design.\n\n1. Latent Labs. (2025). **Platform \u2013 Latent-X:**\n   <https://www.latentlabs.com/latent-x/>\n2. Nature Chemical Biology. (2025). **Article on LatentX AI model for protein binder design:**\n   <https://www.nature.com/articles/s41589-025-01929-w>\n3. TechCrunch. (2025). **Latent Labs launches web-based AI model to democratize protein design:**\n   <https://techcrunch.com/2025/07/21/latent-labs-launches-web-based-ai-model-to-democratize-protein-design/>\n4. Forbes. (2025). **How Latent Labs plans to create medicines from scratch with AI:**\n   <https://www.forbes.com/sites/davidprosser/2025/02/12/how-latent-labs-plans-to-create-medicines-from-scratch-with-ai/>\n\nThe frontiers of drug discovery have been dramatically redrawn by the advent of artificial intelligence, with LatentX standing at the cutting edge of this revolutionary shift. LatentX is a state-of-the-art AI model developed by Latent Labs that generates lab-functional protein binders with unprecedented precision at the atomic level. Traditional drug discovery relies on screening millions of random molecules, a laborious and expensive process with low success rates. LatentX disrupts this by solving the intricate geometric puzzle of binding\u2014placing every atom precisely to achieve high-affinity and highly specific protein binders. This model generates viable binders faster and with far fewer candidates than ever before, reaching binding affinities in the picomolar range for mini-binders and achieving near-perfect hit rates for macrocycles.\n\nThis leap in efficiency is transformational. By co-sampling protein sequences and their 3D structures simultaneously, LatentX outperforms prior AI models that treat these separately, enabling rapid computational experimentation that once took months to accomplish. Importantly, LatentX extends beyond nature\u2019s existing protein repertoire, designing novel molecules that obey strict biochemical constraints like hydrogen bonding and aromatic interactions. This capability opens the door to therapeutic modalities including nanobodies and antibodies, previously out of reach for generative AI.\n\nLatent Labs has democratized this technology through a no-code web platform, empowering researchers across academia, biotech startups, and pharmaceutical companies to design new protein binders directly in their browsers. Users upload target proteins, select binding hotspots, and generate binder molecules with computational rankings that mirror lab validation results, dramatically lowering barriers to entry for AI-powered drug design. The platform\u2019s success is backed by lab experiments that validate the predicted binders\u2019 function, specificity, and affinity, underscoring LatentX\u2019s ability to deliver practical therapeutics at unprecedented speed.\n\nSupported by a team of former AlphaFold developers and DeepMind scientists, and fueled by a significant $50 million investment round in early 2025, LatentX represents a pivotal step towards making biology programmable. The goal is to push drug discovery fully into the computational realm\u2014allowing drugs to be designed and refined almost instantly, vastly accelerating the availability of new treatments for a variety of conditions. LatentX\u2019s approach exemplifies how frontier AI is not just solving biological predictions but creating biomolecules that could fundamentally transform medicine, offering hope for more personalized and precisely targeted therapies in the near future.", "# marcdeller.com: Artificial Intelligence\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n## AI to Action: From Models to Molecules\n\nArtificial intelligence is transforming drug discovery, protein science, and structural biology. Tools like AlphaFold predict protein structures accurately, speeding up drug target identification and new therapeutic design. Generative models produce new proteins and simulate biological mechanisms, streamlining research. Applications include virtual molecular screening and toxicity prediction. This shifts results toward faster development, lower costs, and higher success rates, enabling the development of medicines for diseases like cancer.\n\n* **AI accelerates drug discovery:** Prediction of protein structures, enabling researchers to target diseases more efficiently and accurately.\n* **Generative models design novel proteins:** Simulation of complex biological interactions, streamlining experimental workflows, and hypothesis testing\n* **Key tools like AlphaFold, Boltz2, ProteinMPNN, and RFDiffusion:** Rapid prediction of protein folding and molecular binding.\n* **AI-driven virtual screening:** Docking and Molecular Dynamics simulations using tools such as AutoDock and GROMACS enhance drug design, enabling rapid in silico screening to identify potential leads.\n\nTop Posts\n\n### [Unlocking the Promise and Pitfalls of Large Language Models in Biomedical NLP](https://marcdeller.com/unlocking-the-promise-and-pitfalls-of-large-language-models-in-biomedical-nlp/)\n\n[Unlocking the Promise and Pitfalls of Large Language Models in Biomedical NLP](https://marcdeller.com/unlocking-the-promise-and-pitfalls-of-large-language-models-in-biomedical-nlp/)\n\n### [Alphafold Local Install on Mac](https://marcdeller.com/alphafold-local-install-on-mac/)\n\n[Alphafold Local Install on Mac](https://marcdeller.com/alphafold-local-install-on-mac/)\n\n### [Installation of RFdiffusion on MacOSX M1 architecture without GPU support](https://marcdeller.com/installation-of-rfdiffusion-on-macosx-m1-architecture-without-gpu-support/)\n\n[Installation of RFdiffusion on MacOSX M1 architecture without GPU support](https://marcdeller.com/installation-of-rfdiffusion-on-macosx-m1-architecture-without-gpu-support/)\n\n### [Unleashing the Power of AnythingLLM: Running Large Language Models Locally](https://marcdeller.com/unleashing-the-power-of-anythingllm-running-large-language-models-locally/)\n\n[Unleashing the Power of AnythingLLM: Running Large Language Models Locally](https://marcdeller.com/unleashing-the-power-of-anythingllm-running-large-language-models-locally/)\n\n### [Democratizing Bioinformatics with Multi-Agent Language Models](https://marcdeller.com/democratizing-bioinformatics-with-multi-agent-language-models/)\n\n[Democratizing Bioinformatics with Multi-Agent Language Models](https://marcdeller.com/democratizing-bioinformatics-with-multi-agent-language-models/)\n\nAI\n\n### [BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline](https://marcdeller.com/boltzmaker-vibe-coding-a-campaign-scale-boltz-2-pipeline/)\n\nAI\n\n### [ChemSage: Vibe-Coding a Chemically Aware Drug-Discovery LLM on Local Hardware](https://marcdeller.com/chemsage/)\n\nAI\n\n### [OpenBind Publishes a Dense Structure-Affinity Dataset for Structure-Based AI](https://marcdeller.com/openbind-publishes-a-dense-structure-affinity-dataset-for-structure-based-ai/)\n\nAI\n\n### [Sci-Hub at the Crossroads: From Shadow Archive to AI Research Tool](https://marcdeller.com/sci-hub-at-the-crossroads-from-shadow-archive-to-ai-research-tool/)\n\nAI\n\n### [When Protein Design Becomes a Security Problem](https://marcdeller.com/when-protein-design-becomes-a-security-problem/)\n\nAI\n\n### [Self-Adapting Language Models: Are We Creating Our Final Invention?](https://marcdeller.com/self-adapting-language-models-are-we-creating-our-final-invention/)\n\nAI\n\n### [Unlocking Documents for AI: The Power of Docling for AI-Driven Transformation](https://marcdeller.com/unlocking-documents-for-ai-the-power-of-docling-for-ai-driven-transformation/)\n\nAI\n\n### [Call Me A Jerk: Persuading AI to Comply with Objectionable Requests](https://marcdeller.com/call-me-a-jerk-persuading-ai-to-comply-with-objectionable-requests/)\n\nAI\n\n### [Revolutionizing Drug Discovery: The Click2Drug Platform and the Rise of AI-Powered Virtual Screening](https://marcdeller.com/revolutionizing-drug-discovery-the-click2drug-platform-and-the-rise-of-ai-powered-virtual-screening/)\n\nAI\n\n### [Cats Confuse Reasoning LLM: Query-Agnostic Adversarial Triggers for Reasoning Models](https://marcdeller.com/cats-confuse-reasoning-llm-query-agnostic-adversarial-triggers-for-reasoning-models/)\n\nAI\n\n### [When Perception Deceives Reality: The Hidden Flaws in AI Reasoning Evaluation](https://marcdeller.com/when-perception-deceives-reality-the-hidden-flaws-in-ai-reasoning-evaluation/)\n\nAI\n\n### [How to Confuse Machine Learning: Understanding Modern Adversarial Attacks](https://marcdeller.com/how-to-confuse-machine-learning-understanding-modern-adversarial-attacks/)\n\nAI\n\n### [Democratizing Bioinformatics with Multi-Agent Language Models](https://marcdeller.com/democratizing-bioinformatics-with-multi-agent-language-models/)\n\nAI\n\n### [Unleashing the Power of AnythingLLM: Running Large Language Models Locally](https://marcdeller.com/unleashing-the-power-of-anythingllm-running-large-language-models-locally/)\n\nAI\n\n### [Installation of RFdiffusion on MacOSX M1 architecture without GPU support](https://marcdeller.com/installation-of-rfdiffusion-on-macosx-m1-architecture-without-gpu-support/)\n\nAI\n\n### [Alphafold Local Install on Mac](https://marcdeller.com/alphafold-local-install-on-mac/)\n\nAI\n\n### [Unlocking the Promise and Pitfalls of Large Language Models in Biomedical NLP](https://marcdeller.com/unlocking-the-promise-and-pitfalls-of-large-language-models-in-biomedical-nlp/)", "# The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\nAuthors: Span, E.A.; Suess, D.L.M.; Deller, M.C.; Britt, R.D.; Marletta, M.A.\nSource: ACS Chemical Biology, 2017\n\n**Authors:** Elise A. Span, Daniel L. M. Suess, Marc C. Deller, R. David Britt and Michael A. Marletta (corresponding author, marletta@berkeley.edu). **Affiliations (condensed):** Biophysics Graduate Group, University of California, Berkeley, California 94720 (Span); Department of Chemistry, University of California, Davis, California 95616 (Suess, Britt); The Joint Center for Structural Genomics, The Scripps Research Institute, La Jolla, California 92037 (Deller); Department of Chemistry and Department of Molecular and Cell Biology, University of California, Berkeley, California 94720 (Marletta). **Published:** ACS Chemical Biology, volume 12, issue 4, pages 1095\u20131103, April 21, 2017. Author manuscript available in PMC August 13, 2017.\n\n**Accession codes:** Coordinates and structure factors have been deposited in the PDB with accession number 5UFV. Supporting Information (supplemental text, SI Table 1 and SI Figures 1\u20134) is available on the ACS Publications website. The authors declare no competing financial interest.\n\n## Abstract\n\nPolysaccharide monooxygenases (PMOs) are secreted metalloenzymes that catalyze the oxidative degradation of polysaccharides in a copper-, oxygen-, and reductant-dependent manner. Cellulose-active fungal PMOs degrade cellulosic substrates to be utilized as a carbon source for fungal growth. To gain insight into the PMO mechanism, the role of conserved residues in the copper coordination sphere was investigated. Here, we report active-site hydrogen-bonding motifs in the secondary copper coordination sphere of MtPMO3*, a C1-oxidizing PMO from the ascomycete fungus Myceliophthora thermophila. A series of point substitutions that disrupt this conserved network are used to interrogate its function. Activity assays, in conjunction with EPR spectroscopy, demonstrate that residues H161 and Q167 are involved in stabilizing bound oxygen, and H161 appears to play a role in proton transfer. Additionally, Q167 increases the ligand donor strength of Y169 to the copper via a hydrogen-bonding interaction. Altogether, H161 and Q167 are important for oxygen activation, and the results are suggestive of a copper-oxyl active intermediate.\n\n## Introduction\n\nCellulose degradation in fungi is accomplished by an ensemble of cellulose-active enzymes that are secreted to the extracellular space following induction by cellulosic metabolites.[1,2] These enzymes largely consist of hydrolytic exo- and endoglucanases working in concert with the recently discovered class of redox enzymes called polysaccharide monooxygenases (PMOs, also referred to as lytic polysaccharide monooxygenases, or LPMOs).[3] Fungal PMOs that are active on cellulose (also known as AA9 enzymes) utilize a flat protein surface centered around the catalytic site to bind to the surface of crystalline cellulose. PMOs are the only enzymes known to degrade bulk cellulose without the need to separate one glucan from the matrix, a task that is thermodynamically difficult due to the dense network of hydrogen bonds (H-bonds) and van der Waals forces that pack glucans into microfibrils and higher-order structures.[4,5] PMOs act synergistically with hydrolytic cellulases and thus have commercial value in enhancing cellulose degradation for the production of renewable fuels and chemicals.[6] Moreover, PMOs represent a newly discovered enzyme superfamily with novel catalytic features that are yet to be fully understood.[7]\n\nPMO activity is copper-, oxygen-, and reductant-dependent, and reaction products show incorporation of one oxygen atom from O2, consistent with monooxygenase activity (Scheme 1).[8,9] Although the PMO chemical mechanism is unknown, it has been postulated to be similar to other monooxygenases such as peptidylglycine \u03b1-hydroxylating monooxygenase or cytochrome P450 enzymes.[10] The generally agreed-upon mechanisms for these enzymes differ in their reactive intermediates and catalytic steps, but all require the well-timed delivery of oxygen (O2), electrons, and protons to the active site.[11\u201313] Active-site structures of O2-activating copper (Cu) enzymes are known to tune reactivity in various ways,[14] and the distinctive features of the PMO active site may be integral to the reaction mechanism. PMOs possess a type II Cu center, coordinated by a histidine brace motif, in which two absolutely conserved histidine side chains (one being H1, which possesses a unique N\u03b5-methylation) and the N-terminal amino group ligate the Cu. In the AA9 family, a conserved tyrosine also occupies the primary coordination sphere, axial to the histidine coordination plane. The secondary coordination sphere comprises a conserved hydrogen-bonding (H-bonding) network, as deduced from AA9 crystal structures.[15\u201317]\n\n*Scheme 1.* Polysaccharide Monooxygenase (PMO) Reaction, R\u2013H = Substrate (e.g., Cellulose).\n\nNoncovalent secondary sphere effects on catalysis in metalloenzymes can arise from steric, electrostatic, and H-bonding interactions; among these, H-bonding is thought to be the most important.[18,19] Hydrogen bonds formed by secondary sphere residues in O2-binding proteins are known to stabilize bound O2, as in the classical case of the distal histidine in hemoglobin and myoglobin.[20] In O2-activating enzymes, H-bonds in the secondary coordination sphere help tune redox potentials and ligand donor strengths, facilitate proton and electron transfer, and help stabilize nonstandard coordination geometries and high-energy intermediates.[18] In cytochrome P450cam, for example, H-bonding residues in the proximal pocket tune the redox potential of the heme,[21] and those in the distal pocket control proton transfer to reactive iron-bound oxygen-based intermediates.[22]", "# GitHub README: ButtFold\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ButtFold, 2026\n\nBoth bake gates therefore fail, and they are right to \u2014 they are assertions about *folding*.\nThe generative bake states the mirror image of each (it must expand fourfold, start as noise,\nend with CA\u2013CA at 3.80 \u00c5, and finish with real secondary structure), and it goes through the\nsame `build_frames` rather than a second builder, because one frame builder is what makes the\nPython-vs-JavaScript parity test possible at all. The contact rule is two-sided for it: not\n\"are these residues close\" \u2014 they always are \u2014 but \"are they at their final separation\",\nwhich measured 0.00 \u2192 1.00 across a trajectory where the one-sided rule read 1.00 throughout.\nIts tempo comes from that same number, because `compaction` clamps to 1 on the first frame\nand would otherwise play the whole piece flat out. And it is baked rather than served live\nbecause 1000 denoising steps is 2.2 minutes on a Mac and 10.2 on the droplet, with no\nspeed-up from more cores.\n\n**The music is drawn on the structure.** A contact note is an event between two residues \u2014\nthe score stores both \u2014 and both are already on the stage, so when the note sounds a line is\nstruck between them: green for a contact, amber for a core contact in the bass voice, fading\nat the note's own velocity. Its two residues brighten on the ribbon at the same moment. Under\nit, the chain unrolled into a strip, one cell per residue in the cartoon's own colours,\nlighting from **all five voices** \u2014 the pad picks helix residues, the rhythm picks sheet ones,\nso the strip carries the whole texture rather than the contacts alone. At rest that strip is a\nsequence view of the secondary structure, re-forming as the fold does.\n\nBoth read the same list, and they get it by **asking the audio clock what is sounding** rather\nthan being told by the scheduler. The scheduler queues a second ahead, so a visual driven from\nit would run ahead of its own music, and it queues nothing while paused, so scrubbing would\nshow an empty stage. A query cannot drift, because it *is* the clock the animation already\nfollows. A note is drawn for the longer of its own length and a 0.9 s floor: a contact is a\nsemiquaver and would otherwise be a single frame, and a pad chord is held for a whole bar and\nwould otherwise go dark while you could still hear it.\n\n**Both computed paths are watched as they happen**, which for the server means the browser reads\nthe coordinate file the droplet is still writing to. Progress there was a percentage over a still\npicture: the droplet folded for half a minute while a number climbed, and the same protein folding\nin the browser beside it turned and collapsed the whole time. The frames were already on disk, so\nthe route serves what is there and the browser runs the same `frames.js` over it that the live\nworker runs on its own output. The secondary-structure and radius-of-gyration charts are drawn a\nframe at a time from either source.\n\n## \ud83c\udfb9 How a trajectory becomes music\n\nThe mapping is PhoneFold's, row for row, because the point is that someone who plays with this and\nthen downloads the app hears the same piece.\n\n| Trajectory feature | Musical parameter |\n|---|---|\n| New contact event | Note onset; sequence separation sets the register |\n| Long-range hydrophobic contact | Bass note (core packing) |\n| Helix content | Sustained pad, stacked fourths |\n| Sheet content | Staccato interlocking figure |\n| Coil content | Arpeggiation between chord tones |\n| Per-residue confidence | Note velocity for that residue |\n| Mean confidence | Low-pass cutoff, detune, reverb |\n| Radius of gyration | Tempo and register |\n| Convergence | Cadence, resolving to the tonic |\n\n\"Confidence\" here is the fraction of a residue's own native contacts that have formed, which is\nreal per-residue information: the hydrophobic core locks in first and completely while the termini\nare still loose, so a region that never resolves stays a detuned wash for the whole piece.\n\nFive styles ship (Fantasy, Jazz, Pop, Rock, Surf) and are the app's own JSON files, byte-identical.\nEach note is placed in space at its residue's live 3D position through a `PannerNode` with HRTF, so\nthe fold collapses around the listener.\n\n## \ud83d\ude80 Usage\n\nOpen <https://buttfold.mdeller.com>. Press Play. Drag the structure to turn it, scroll to zoom,\ndouble-click to reframe.\n\nTo run it locally:\n\n```bash\ngit clone https://github.com/bellcheddar/ButtFold.git\ncd ButtFold\npython3 -m venv .venv\n./.venv/bin/pip install -r requirements-dev.txt\n./.venv/bin/python app.py                    # http://127.0.0.1:8007/\n./.venv/bin/python -m buttfold.worker        # optional: the fold queue\n```\n\n`requirements.txt` is the web layer (Flask and gunicorn, no numpy) and `requirements-queue.txt` is\nthe worker (numpy, for the bake). They are deliberately disjoint: the droplet runs both processes\nunder separate systemd budgets so a fold can never starve the six other apps on the box.\n\n## \ud83d\udd27 Rebuilding the artefacts\n\nEverything the app serves is committed. These regenerate it, and all of them run on the Mac.\n\n| Command | What it does |\n|---|---|\n| `tools/extract_natives.py` | Vendors the native structures and starting coils from PhoneFold into `data/natives/` |\n| `tools/bake_gallery.py` | Folds the six gallery proteins and bakes them, with the collapse assertions |\n| `tools/bake_gallery.py --frame-fixtures` | Also writes the exact coordinates the JS parity test compares against |\n| `tools/build_wasm.sh` | Compiles the G\u014d model to WebAssembly, emsdk pinned at 4.0.7 |\n| `tools/swift_score_dump/run.sh` | Runs PhoneFold's own Sonifier over this gallery to produce the reference scores |\n| `tools/psea.py --fixtures`, `tools/contacts.py --fixtures` | Reference outputs for the JS geometry ports |\n| `tools/check_all.sh` | Every machine-verifiable gate in one command |\n\n## \ud83e\uddea Testing: verify the artefact, not the intent", "# 2025 \u2013 Investing in Human Skills in the Age of AI\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-28 on marcdeller.com (categories: certificates).*\n\n* [Link to certificate](https://www.linkedin.com/learning/certificates/c8b38adb7299bbb40c98b244da91d1eb8b915de8e5caeadc8d3f73f6fce6a9a0)\n\n* **Award Details:** LinkedIn Learning issued this certificate on March 18, 2025 at 04:35 PM UTC.\n* **Course Focus:** It recognizes completion of a course on interpersonal skills, emotional intelligence, and critical thinking.\n* **Career Value:** This matters because scientific leadership still relies on judgment, teamwork, and communication.\n* **Practical Use:** It supports stronger collaboration, mentoring, and decision-making in multidisciplinary teams.", "# 2015 \u2013 Crystal structure of a two-subunit TrkA octameric gating ring assembly\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-09-02 on marcdeller.com (categories: papers).*\n\n* [Link to Paper](https://marcdeller.com/wp-content/uploads/2025/09/file.pdf)\n\n* **Ancient Architecture Revealed:** The two-subunit TrkA system represents an evolutionary precursor to modern single-subunit potassium transporters, offering insights into protein evolution\n* **Sophisticated Regulation:** The octameric assembly employs multiple nucleotide binding sites and allosteric networks to achieve precise control over potassium transport\n* **Thermophilic Adaptations:** Unique structural features including interfacial salt bridges enable function in extreme thermal environments\n* **Pathogenic Relevance:** Similar systems in disease-causing bacteria like M. tuberculosis present novel therapeutic targets for antimicrobial development\n* **Metabolic Integration:** Recent studies reveal intimate connections between potassium homeostasis and cellular metabolism, expanding beyond simple ion transport\n* **Structural Precision:** The assembly achieves remarkable stability through extensive protein interfaces while maintaining conformational flexibility for regulation\n\n1. **Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly:** [Deller et al., PLoS ONE, 2015](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0122512)\n2. **The Thermotoga maritima Trk Potassium Transporter\u2014from Molecular Architecture to Physiological Function:** [Johnson et al., Journal of Bacteriology, 2009](https://journals.asm.org/doi/10.1128/jb.01367-08)\n3. **Activation of a nucleotide-dependent RCK domain requires binding of divalent cations:** [Teixeira-Duarte et al., eLife, 2019](https://elifesciences.org/articles/50661)\n4. **The allosteric gating mechanism of the MthK channel:** [Guan et al., National Science Review, 2022](https://academic.oup.com/nsr/article/9/8/nwac072/6567883)\n5. **Structure and mechanism of the K+/H+ exchanger KefC:** [Huang et al., Nature Communications, 2024](https://www.nature.com/articles/s41467-024-49082-7)\n6. **Molecular Mechanisms for Bacterial Potassium Homeostasis:** [Epstein, Current Opinion in Microbiology, 2021](https://pubmed.ncbi.nlm.nih.gov/33798529/)\n\n**Revolutionary Insights into Bacterial Potassium Transport: The Two-Subunit TrkA Octameric Assembly Unveils Ancient Evolutionary Mechanisms**\n\nThe intricate world of bacterial potassium homeostasis has recently been illuminated by groundbreaking structural revelations that bridge evolutionary gaps spanning millions of years. The crystal structure of the two-subunit TrkA octameric gating ring assembly from Thermotoga maritima represents a pivotal discovery that fundamentally reshapes our understanding of how ancient organisms managed one of life\u2019s most critical ionic processes. This remarkable protein complex, designated TM1088, stands as a testament to nature\u2019s ingenuity in developing sophisticated regulatory mechanisms that predate modern single-subunit systems by potentially vast evolutionary timescales.\n\nThe structural complexity of this assembly is breathtaking in its precision and elegance. Unlike conventional single-subunit TrkA proteins found in most bacteria, the T. maritima system employs two distinct but complementary proteins, TM1088A and TM1088B, which combine to form a functional octameric assembly consisting of four units of each protein. This 160-kilodalton complex exhibits extraordinary structural sophistication, with TM1088A containing a single N-terminal Trk domain and TM1088B featuring both N-terminal and C-terminal Trk domains connected by a critical inter-domain linker. The assembly achieves remarkable stability through three major protein interfaces: TM1088A homodimerization, TM1088B homodimerization, and the crucial TM1088A-TM1088B heteromeric interface that buries nearly 18,000 square angstroms of surface area.\n\nRecent advances in cryo-electron microscopy and crystallographic techniques have revealed unprecedented details about nucleotide binding mechanisms in related RCK domain systems. The TM1088A protein demonstrates exquisite specificity for adenosine nucleotides, particularly AMP, which binds at the characteristic GXGXXG motif within a Rossmann-like fold. This nucleotide binding occurs at four symmetrically arranged sites within the central channel of the octameric assembly, creating a molecular environment that can respond to cellular energy states. The binding pocket architecture involves critical residues including Ala59, Arg17, and Arg105, which form a network of hydrogen bonds that anchor the AMP molecule at both its adenine and phosphate moieties. Contemporary research has demonstrated that divalent cations play crucial cofactor roles in nucleotide-dependent RCK domain activation, with magnesium ions coordinating through gamma-phosphates of bound ATP in related systems to stabilize active conformations.\n\nThe evolutionary implications of this two-subunit architecture are profound and far-reaching. Phylogenetic analysis suggests that the TM1088 system represents an evolutionary predecessor to modern single-subunit TrkA proteins found in organisms like Escherichia coli. The transition from this ancestral two-subunit system to contemporary single-subunit versions likely occurred through gene duplication followed by fusion events, creating the tandem domain architecture observed in modern TrkA proteins. This evolutionary pathway provides crucial insights into how complex regulatory mechanisms can be maintained while undergoing significant structural reorganization. The presence of similar two-subunit systems in pathogenic organisms such as Mycobacterium tuberculosis suggests that this ancient architecture may confer specific advantages in certain environmental niches, particularly those requiring enhanced stability or specialized regulatory responses.", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nIn some embodiments, the inhibitor of an immune checkpoint molecule is anti-PD1 antibody, anti-PD-L1 antibody, or anti-CTLA-4 antibody. In some embodiments, the inhibitor of an immune checkpoint molecule is an inhibitor of PD-1 or PD-L1, e.g., an anti-PD-1 or anti-PD-L1 monoclonal antibody. Named examples given in the specification include nivolumab, pembrolizumab, atezolizumab, durvalumab, avelumab, cemiplimab, tislelizumab, spartalizumab (PDR001), cetrelimab (JNJ-63723283), toripalimab (JS001), camrelizumab (SHR-1210), sintilimab (IB1308), AB122 (GLS-010), AMP-224, AMP-514/MEDI-0680, BMS936559, JTX\u20144014, BGB-108, SHR-1210, MEDI4736, FAZ053, BCD-100, KN035, CS1001, BAT1306, LZM009, AK105, HLX10, SHR-1316, CBT-502 (TQB2450), A167 (KL-A167), STI-A101 (ZKAB001), CK-301, BGB-A333, MSB-2311, HLX20, TSR-042, LY3300054 and MGA012, together with an extensive list of further checkpoint-directed agents (anti-CTLA-4: ipilimumab, tremelimumab, AGEN1884, CP-675,206; anti-LAG3: BMS-986016, LAG525, INCAGN2385, eftilagimod alpha (IMP321); anti-TIM3: INCAGN2390, MBG453, TSR-022; anti-CD20: obinutuzumab, rituximab; CD73: oleclumab, MEDI9447; TIGIT: OMP-31M32; VISTA: JNJ-61610588, CA-170; B7-H3: enoblituzumab, MGD009, 8H9; KIR: lirilumab, IPH4102; A2aR: CPI-444; TGF-beta: trabedersen, galusertinib, M7824; PI3K-gamma: IPI-549; CD47: Hu5F9-G4, TTI-621; CD70: cusatuzumab, BMS-936561; MCLA-145). Agonists named include GITR agonists TRX518, MK-4166, INCAGN1876, MK-1248, AMG228, BMS-986156, GWN323, MEDI1873, MEDI6469; OX40 agonists INCAGN01949, MED10562 (tavolimab), MOXR-0916, PF-04518600, GSK3174998, BMS-986178, 9B12 and the OX40L fusion protein MEDI6383; CD40 agonists CP-870893, ADC-1013, CDX\u20141140, SEA-CD40, R07009789, JNJ-64457107, APX\u2014005M, Chi Lob 7/4; ICOS agonists GSK-3359609, JTX\u20142011, MEDI-570; CD28 agonist theralizumab; CD27 agonist varlilumab; TLR7/8 agonist MEDI9197; CD137 agonists urelumab and utomilumab.\n\nIn some embodiments, the compounds of the present disclosure can be used in combination with bispecific antibodies. In some embodiments, one of the domains of the bispecific antibody targets PD-1, PD-L1, CTLA-4, GITR, OX40, TIM3, LAG3, CD137, ICOS, CD3 or TGF-beta receptor. In some embodiments, the bispecific antibody binds to PD-1 and PD-L1 (e.g., MCLA-136), or binds to PD-L1 and CTLA-4 (e.g., AK104).\n\nIn some embodiments, the compounds of the disclosure can be used in combination with one or more metabolic enzyme inhibitors, e.g., an inhibitor of IDO1, TDO, or arginase. Examples of IDO1 inhibitors include epacadostat, NLG919, BMS-986205, PF-06840003, IOM2983, RG-70099 and LY338196.\n\n**Combination with other therapies.** In some embodiments, the WRN inhibitor compounds of the present disclosure can be used in combination with other methods of treating cancers. Examples of other methods of treating cancers include chemotherapy, irradiation therapy, tumor-targeted therapy, adjuvant therapy, immunotherapy, surgery, steroids, immunosuppressants, immune-oncology agents, metabolic enzyme inhibitors, chemokine receptor inhibitors, and phosphatase inhibitors, as well as targeted therapies such as Bcr-Abl, Flt-3, EGFR, HER2, JAK, c-MET, VEGFR, PDGFR, c-Kit, IGF-1R, RAF, FAK, and CDK4/6 kinase inhibitors. Compounds of the present disclosure can be combined with one or more inhibitors of the following kinases for the treatment of cancer: Akt1, Akt2, Akt3, BCL2, CDK4/6, TGF-PR, PKA, PKG, PKC, CaM-kinase, phosphorylase kinase, MEKK, ERK, MAPK, mTOR, EGFR, HER2, HER3, HER4, INS-R, IDH2, IGF-1R, IR-R, PDGFaR, PDGFPR, PI3K (alpha, beta, gamma, delta, and multiple or selective), CSF1R, KIT, FLK-II, KDR/FLK-1, FLK-4, fit-1, FGFR1, FGFR2, FGFR3, FGFR4, c-Met, PARP, Ron, Sea, TRKA, TRKB, TRKC, TAM kinases (Axl, Mer, Tyro3), FLT3, VEGFR/Flt2, Flt4, EphA1, EphA2, EphA3, EphB2, EphB4, Tie2, Src, Fyn, Lck, Fgr, Btk, Fak, SYK, FRK, JAK, ABL, ALK and B-Raf. Non-limiting examples of inhibitors that can be combined with the compounds of the present disclosure for treatment of cancer and infections include an FGFR inhibitor (FGFR1, FGFR2, FGFR3 or FGFR4, e.g., pemigatinib (INCB54828), INCB62079), an EGFR inhibitor (also known as ErB-1 or HER-1; e.g., erlotinib, gefitinib, vandetanib, orsimertinib, cetuximab, necitumumab, or panitumumab), a VEGF inhibitor or pathway blocker (e.g. bevacizumab, pazopanib, sunitinib, sorafenib, axitinib, regorafenib, ponatinib, cabozantinib, vandetanib, ramucirumab, lenvatinib, ziv-aflibercept), a PARP inhibitor (e.g., olaparib, rucaparib, veliparib or niraparib), a JAK inhibitor (JAK1 and/or JAK2, e.g., ruxolitinib or baricitinib; JAK1, e.g., itacitinib (INCB39110), INCB052793, or INCB054707), an IDO inhibitor (e.g., epacadostat, NLG919, or BMS-986205, MK7162), an LSD1 inhibitor (e.g., GSK2979552, INCB59872 and INCB60003), a TDO inhibitor, a PI3K-delta inhibitor (e.g., parsaclisib (INCB50465) or INCB50797), a PI3K-gamma inhibitor such as PI3K-gamma selective inhibitor (e.g., INCB53914), a CSF1R inhibitor, a TAM receptor tyrosine kinases (Tyro-3, Axl, and Mer; e.g., INCB081776), an adenosine receptor antagonist (e.g., A2a/A2b receptor antagonist), an HPK1 inhibitor, a chemokine receptor inhibitor (e.g., CCR2 or CCR5 inhibitor), a SHP1/2 phosphatase inhibitor, a histone deacetylase inhibitor (HDAC) such as an HDAC8 inhibitor, an angiogenesis inhibitor, an interleukin receptor inhibitor, bromo and extra terminal family members inhibitors (for example, bromodomain inhibitors or BET inhibitors such as INCB54329 and INCB57643), c-MET inhibitors (e.g. capmatinib), an anti-CD19 antibody (e.g., tafasitamab), an ALK2 inhibitor (e.g., INCB00928); or combinations thereof.", "# BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-07-08 on marcdeller.com (categories: ai, artificial-intelligence, blog, project, sb).*\n\nBoltzMaker: Boltz2 campaign-scale structure and affinity prediction, binding analysis, and run control, orchestrated end-to-end from a single spec file.\n\n* [BoltzMaker GitHub Repo](https://github.com/bellcheddar/BoltzMaker)\n* [BoltzMaker Vibe Code Project Page](https://marcdeller.com/boltzmakervibe)\n* [Example dashboard output for 5HT2 GPCR screening campaign](https://bellcheddar.github.io/BoltzMaker/examples/5ht2_gq_panel/boltz_dashboard.html)\n* [Example analysis of 5HT2 GPCR screening campaign](https://marcdeller.com/5HT2_findings/findings.preview.html)\n\n* **Spec-driven, end-to-end:** One annotated file drives generation, validation, prediction, and reporting for an entire Boltz-2 campaign.\n* **Chemistry checked before compute:** RDKit flags stereocentres, salts, and protonation ambiguity before hours of folding, not after.\n* **Built for modest hardware:** Memory-aware defaults and resumable runs make campaign-scale prediction viable on a 64GB MacBook Pro M1.\n* **Expert plus AI, not either alone:** A domain scientist who cannot code professionally can still ship a real, documented tool.\n* **Checked against known biology:** `compare-sse` reports apo-versus-holo movement as motif language, and reproduced the textbook TM6 outward swing on Gq coupling in 6 of 6 pairs.\n* **Failures made survivable:** An exact attention-chunking patch and automatic per-target retry turned a hard MPS crash and a mid-run OOM into campaigns that simply finish.\n\n1. **Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction:** [Passaro et al., bioRxiv, 2025](https://www.biorxiv.org/content/10.1101/2025.06.14.659707v1)\n2. **Accurate structure prediction of biomolecular interactions with AlphaFold 3:** [Abramson et al., Nature, 2024](https://www.nature.com/articles/s41586-024-07487-w)\n3. **PLIP 2025: introducing protein-protein interactions to the protein-ligand interaction profiler:** [Schake et al., Nucleic Acids Research, 2025](https://academic.oup.com/nar/article/53/W1/W463/8128215)\n4. **RDKit: Open-source cheminformatics:** [Landrum et al., rdkit.org](https://www.rdkit.org/)\n5. **The Properties of Known Drugs. 1. Molecular Frameworks:** [Bemis and Murcko, J. Med. Chem., 1996](https://pubs.acs.org/doi/10.1021/jm9602928)\n6. **GPCRdb in 2025: adding odorant receptors, data mapper, structure similarity search and models of physiological ligand complexes:** [Taracena Herrera et al., Nucleic Acids Research, 2025](https://academic.oup.com/nar/article/53/D1/D425/7903362)\n7. **KLIFS: an overhaul after the first 5 years of supporting kinase research:** [Kanev et al., Nucleic Acids Research, 2021](https://academic.oup.com/nar/article/49/D1/D562/5934416)\n\n###### **How five small single-purpose tools became one spec-driven workflow that folds, scores, and reports a full Boltz-2 affinity campaign on modest local hardware**\n\nStructure prediction has quietly become the cheap part of the problem. [Boltz-2](https://www.biorxiv.org/content/10.1101/2025.06.14.659707v1) will fold a protein and a ligand together and estimate their binding affinity, but running a real campaign by hand still means writing dozens of near-identical YAML files, remembering the correct flags for whatever hardware you\u2019re on, and picking through prediction JSONs afterward.\n\nThose are repetitive, error-prone steps that do not need a human.\u00a0[BoltzMaker](https://github.com/bellcheddar/BoltzMaker)\u00a0collapses them into one spec-driven pipeline: a single annotated file becomes generated inputs, validated chemistry, a monitored run, and a report you can read at a glance.\n\nAnd the name? It is a nod to [Boltmaker](https://www.timothytaylor.co.uk/beer/boltmaker), Timothy Taylor\u2019s Champion Beer of Britain, and a personal favorite; the pipeline runs whether or not there is a pint in hand.\n\nA candid disclaimer belongs here. I am not a professional software engineer, and BoltzMaker is not production software. It is an exercise, and the latest entry in my ongoing [vibe-coding series](https://marcdeller.com/vibe-coding/), in which a subject-matter expert can build with their own data on a laptop, for the price of some electricity and a bit of patience. That framing is the whole point: the value is not the code; it is that the person who understands the science can now assemble the tooling around it themselves, rather than waiting on someone who does not.\n\n> **\u201cCrucially, it achieves strong correlation with experimental readouts on many benchmarks, while being at least 1000\u00d7 more computationally efficient than FEP.\u201d \u2013 Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction, [Passaro et al., bioRxiv, 2025](https://www.biorxiv.org/content/10.1101/2025.06.14.659707v1)**\n\n###### **Installation**\n\n**1.** Open the **Terminal** app (Applications -> Utilities -> Terminal, or search for \u201cTerminal\u201d with Spotlight), then make a folder for BoltzMaker and go into it:\n\n```\n\n\t\t\t\t\tmkdir -p ~/Desktop/BoltzMaker\ncd ~/Desktop/BoltzMaker\n\n```\n\n**2.** Download BoltzMaker into that folder:\n\n```\n\n\t\t\t\t\tgit clone https://github.com/bellcheddar/BoltzMaker.git .\n\n```\n\nThe trailing `.` tells it to put the files directly into the folder you just made, rather than inside another new folder. If this is the first time you\u2019ve used `git`, macOS may pop up a dialog asking to install \u201cCommand Line Tools\u201d \u2014 click Install and wait for it to finish, then run the command again.\n\n**3.** Run the installer:\n\n```\n\n\t\t\t\t\t./install.sh\n\n```\n\nWait for it to finish. It downloads a few GB of software the first time, so it can take a while depending on your connection.\n\n**4.** Check it worked:\n\n```\n\n\t\t\t\t\tpixi run preflight examples/t4_lysozyme/boltz_input.md\n\n```\n\nIf you see a table of green PASS results, you\u2019re ready to go \u2014 see **Commands** below for what to run next.\n\n###### **How BoltzMaker works, and what it grew from**", "# Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\nAuthors: Zhuo, J.; Li, Y.-L.; Qian, D.-Q.; Burns, D.M.; Mei, S.; Rafalski, M.; Xu, M.; Cao, G.; Pan, Y.; Jia, Z.; Jalluri, R.; Epling, L.B.; Fenalti, G.; Deller, M.C.; Procak, J.; Covington, M.; He, X.; Collins, R.; Stubbs, M.; Burke, K.; Oliver, J.; Margulis, A.; Boer, J.; Wynn, R.; Scherle, P.; Diamond, S.; Newton, R.; Zhang, Y.; Metcalf, B.; Yao, W.\nSource: Journal of Medicinal Chemistry (Drug Annotation), 2026\n\n**Human Whole Blood IL-6 pSTAT1 FACS Assay.** To determine compound effects on JAK-mediated STAT1 phosphorylation inhibition in normal human whole blood, the blood from healthy donors (BioIVT) was incubated in the presence or absence of a concentration range of test compounds (3-fold dilution from the highest of 20 \u03bcM concentration in the assay) and 13 ng/mL human IL-6 (R&D Systems) in a 96-well \"2 mL Assay Block\" (Corning) for 17 min at 37 \u00b0C. Blood was then lysed and fixed with prewarmed red blood cell lysed/fixed buffer (BD Bioscience) for 10 min at 37 \u00b0C and centrifuged at 1600 rpm for 5 min. The cells were resuspended in Perm Buffer III (BD Bioscience) and kept at \u221220 \u00b0C overnight. The cells were washed and resuspended in Stain Buffer (BD Bioscience) and stained with Alexa Fluor 647 Mouse Anti-STAT1 (pY701; BD Bioscience), human CD4 (RPA-T4) Alexa488 (BD Bioscience) for 1 h in the dark at room temperature. The cells were then washed and resuspended in Staining Buffer and filtered with 96-well filter plate (MilliporeSigma). Flow cytometric analysis was performed. Cells were gated on CD4+ and Phospho-STAT1 level determined by MFI (BD LSRFortessa X-20 and FlowJo). The data were converted to percent inhibition relative to DMSO control. IC50 determination was performed by fitting the curve of compound percent inhibition versus the log of the compound concentration using the GraphPad Prism 9 software.\n\n**Human Whole Blood GM-CSF pSTAT5 FACS Assay.** To determine the effects of compounds on JAK-mediated STAT5 phosphorylation inhibition in normal human whole blood, blood from healthy donors (BioIVT) was incubated in the presence or absence of a concentration range of test compounds (3-fold dilution starting at 20 \u03bcM concentration in the assay) and 104 pg/mL human GM-CSF (R&D Systems) in a 96-well \"2 mL Assay Block\" (Corning) for 17 min at 37 \u00b0C. Blood was then lysed and fixed with prewarmed red blood cell lysed/fixed buffer (BD Bioscience) for 10 min at 37 \u00b0C and centrifuged at 1600 rpm for 5 min. The cells were resuspended in Perm Buffer III (BD Bioscience) and kept at \u221220 \u00b0C overnight. The cells were washed and resuspended in Stain Buffer (BD Bioscience) and stained with Alexa Fluor 647 Anti-STAT5 (pY694; BD Bioscience), human CD33 (p67.6) PE (BD Bioscience) for 1 h in the dark at room temperature. The cells were then washed and resuspended in Staining Buffer and filtered with 96-well filter plate (MilliporeSigma). Flow cytometric analysis was performed. Cells were gated on CD33+ and Phospho-STAT5 level determined by MFI (BD LSRFortessa X-20 and FlowJo). The data were converted to percent inhibition relative to DMSO control. IC50 determination was performed by fitting the curve of compound percent inhibition versus the log of the compound concentration using the GraphPad Prism 9 software.\n\n**IL-2-Stimulated T-Cell Proliferation Assay.** The effects of the compounds on T-cell proliferation were measured in an ATP luminescence assay. T cells were freshly isolated from whole blood of healthy donors (BioIVT) by elutriation. T cells were first activated by incubation with PHA (Sigma-Aldrich) at 10 \u03bcg/mL for 3 days. After washing the cells with PBS twice, cells were resuspended in RPMI 1640 medium supplemented with 10% FBS and 100 U/mL human IL-2 (R&D system). Cells were seeded at 6000 cells/well in 96-well plates, compounds at various concentrations (3-fold dilution starting at 10 \u03bcM concentration in the assay) were added and cells incubated at 37 \u00b0C, 5% CO2 for 72 h. CellTiter-Glo reagent (Promega) was then added, and luminescence was read with PHERAstar microplate reader (BMG Labtech). Testing compound effects on cell proliferation were calculated and reported as percentage of inhibition relative to wells treated with DMSO only. IC50 determination was performed by fitting the curve of compound percent inhibition versus the log of the compound concentration using the GraphPad Prism 9 software.\n\n**Human PBMC MCP-1 ELISA.** The JAK/STAT pathway partially mediates IL-6\u2212induced MCP-1. The effects of the compounds on this MCP-1 production were measured in human PBMCs. Human blood from healthy donors was purchased through BioIVT (Westbury, NY) and PBMCs isolated using Ficoll-Paque Plus from Cytiva (Marlborough, MA). RPMI 1640 Medium was purchased from Thermo Fisher Scientific (Waltham, MA) with 10% FBS from Gibco (Grand Island, NY). IL-6 was purchased from R&D Systems #206-IL (Minneapolis, MN) and Quantikine Human MCP-1/CCL2 Immunoassay kit, Cat. #DCP00 was purchased from R&D Systems (Minneapolis, MN). PBMCs were resuspended at 4 \u00d7 10\u2076 cells/mL in RPMI 1640 with 10% FCS and added at 220 \u03bcL cell per well into 96-well tissue culture plates (Costar Cat #3595) with and without 5 \u03bcL of prediluted compound (3-fold dilution starting at 10 \u03bcM concentration in the assay) and then incubated 10 min at room temperature. Wells were stimulated by adding 10 ng/mL IL-6 to each well and incubated at 37 \u00b0C, 5% CO2 for 48 h. The 200 \u03bcL supernatant was then collected and tested in the Quantikine Human MCP-1/CCL2 Immunoassay ELISA per kit instructions. Compound inhibition of MCP-1 was calculated and reported as percentage of inhibition relative to wells with a full-inhibitor control (100% inhibition) and DMSO (0% inhibition). IC50 determination was performed by fitting the curve of compound percent inhibition versus the log of the compound concentration using the GraphPad Prism 9 software.\n\n#### In Vitro ADME Assay Compounds 4\u201319, 22", "# ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n**Per-example pass/fail scoring:**\n\u2013 Changed the eval metric methodology throughout `eval_compare.py` from per-instance accuracy (`ok / total_attempts`) to per-example pass/fail (`passed_examples / 100`). An example \u201cpasses\u201d a metric if every instance in that response is correct (`ok == tot`, `tot > 0`). This makes round-over-round comparisons meaningful: R1 calling 2 SMILES correctly is no longer the same score as R3 calling 100 SMILES correctly. The `_per_example_agg()` helper re-derives this from `results[].scores` in `raw_results.json`.\n\n**5-round comparative eval (complete, 2026-06-29):**\n\u2013 All 5 rounds evaluated on 100 shared R5 test examples (seed=42). Full report: `eval/compare/results/compare_20260629_1928.html`. R5 overall 95% (vs R4 80%, R3 87%). R5 wins on exec (+24%), PyExec (+30%), fidelity (+42%), SMARTS (+45%). All README eval tables updated.\n\n### \ud83c\udf93 Round 5: Higher rank + 2.5\u00d7 corpus (2026-06-28, training in progress)\n\n* Expanded SFT dataset from 8,000 to **20,000 examples** (16,000 train / 2,000 val / 2,000 test) across **78 behaviour classes** (19 new vs R4)\n* Raised LoRA rank to **64** (from 32); scale set to **90.0** (RSLoRA effective alpha \u2248 11.25, same order as R4); trainable parameters **doubled to 268M (0.82%)**\n* Extended `max_seq_length` to **3,072** (from 2,048); all corpus examples fit (p99=488, max=558 tokens)\n* Tripled training iters to **3,000**; doubled `val_batches` to 50 for a more stable eval signal\n* New generators targeting R4 exec/PyExec gap and numerical fidelity: `pyexec_drill` (\u00d76), `code_then_quote_v2` (\u00d75), `rounding_explicit` (\u00d75), `fidelity_multistep` (\u00d75)\n* New content generators: `herg_liability`, `selectivity_profile`, `prodrug_bcs`, `sar_delta`, `mdanalysis`, `dssp`, `ppi_interface`, `uniprot_api`, `conformer_3d`, `mcs_search`, `reaction_smarts`, `recap_fragmentation`, `electron_density`, `biological_assembly`, `drug_target_family` (\u00d73-4 each)\n* 707 examples rejected during validation (96.5% pass rate); token lengths all under max_seq_length (p99=488)\n* Training launched 2026-06-28 with `caffeinate -i` under single-process MLX launch (`/tmp/r5_launch.py`) \u2014 memory limits set in-process before importing mlx_lm; cache 16.7 GB / memory limit 50.1 GB\n* **Iter 1 val loss: 2.210** (lower than R4\u2019s 2.397 \u2014 rank=64 adapter has more capacity from the start)\n* Config: `config/train_config_r5.yaml`; adapter path: `adapters/chem_sage_32b_v5_lora/`\n* **Early stop at iter 2000** \u2014 val loss plateaued at 0.055 from iter 1600; fused from iter 1600 checkpoint into `models/chem_sage_32b_v5/` (~17.2 GB)\n\n### \ud83c\udf93 Round 4: Expanded dataset training (2026-06-25 to 2026-06-27, complete)\n\n* Expanded SFT dataset from 5,000 to 8,000 examples (6,400 train / 800 val / 800 test) across 59 behaviour classes\n* New behaviour classes: drug mechanisms/indications/ATC, multi-step reasoning, refusals, personality, structural biology, docking, visualisation (Plotly, R/ggplot2), numerical fidelity drills (`code_then_quote`, `rounding_precision`)\n* `steps_per_report` aligned with `steps_per_eval` (both 25) so train/val loss appear at the same iters\n* Training ran to completion: 1,500 iters, ~43 h (2026-06-25 15:39 to 2026-06-27 10:59 BST)\n* **Best val loss: 0.041 at iter 950** \u2014 beats R3\u2019s all-time best of 0.054 by **24%** (tied at iter 1425)\n* Peak memory 25.5 GB, 32 adapted layers, 134M trainable parameters (0.41%)\n* Fused from iter 950 checkpoint into `models/chem_sage_32b_v4/` (~17.2 GB, 4 safetensors shards)\n\n### \ud83e\uddea Comparative eval harness (2026-06-26)\n\nNew script `eval/compare/eval_compare.py` runs all 4 model rounds side-by-side on the same 100 R4 test examples and produces a rich HTML + Markdown comparison report.\n\n**Features:**\n\u2013 **Option A auto-server**: spawns and kills `mlx_lm.server` per model automatically (no manual terminal juggling); health-checks the OpenAI-compatible endpoint before starting eval; sends SIGTERM to the process group on completion\n\u2013 **Shared examples**: 100 examples sampled from the R4 test set with a fixed seed (reproducible)\n\u2013 **All 13 metrics**: 3 core + 7 new metrics, all reporting `n/a` gracefully on older models\n\u2013 **R3 as baseline**: metric table includes a \u0394 column (\u00b1%) vs Round 3; colour-coded green/red\n\u2013 **Model profile panel**: full spec comparison across architecture, training config, training outcomes, corpus, and eval runtime\n\u2013 **Runtime resource stats**: server process CPU mean/peak %, RSS mean/peak GB (via `psutil`), mean latency (s), throughput (tok/s)\n\u2013 **Validation loss curves**: SVG comparison chart of all 4 models\u2019 val loss curves (clipped at 0.60; dashed line for estimated curves)\n\u2013 **Per-example table**: 100 collapsible rows; click \u201cresponses\u201d to see all model outputs side-by-side\n\u2013 **Resume flag**: `--resume` loads `raw_results.json` cache and skips already-evaluated models\n\u2013 **Branding**: ChemSage ASCII art (pyfiglet), clickable [marcdeller.com](https://marcdeller.com) links throughout\n\n**Usage** (run after R4 training completes and fuse is done):\n\n```\ncd chem_sage/\n# Full run \u2014 all 4 models\npython eval/compare/eval_compare.py\n\n# Subset and resume\npython eval/compare/eval_compare.py --models round3 round4 --resume\n\n# Quick smoke-test (10 examples)\npython eval/compare/eval_compare.py --limit 10\n```\n\n**Config:** `eval/compare/models.yaml` \u2014 all model metadata, loss curves, and training stats in one place; edit to update R4 final values after training.\n\n### \ud83d\udd17 ChromaDB 1.x compatibility fix (2026-06-25)\n\n* ChromaDB 1.0+ switched to a Rust connection backend; `PersistentClient` timed out on databases created with the old Python backend\n* Fix: pass `Settings(allow_reset=True, anonymized_telemetry=False)` to `PersistentClient` in both `rag/retrieve.py` and `scripts/ingest_rag.py`\n* No data loss, no re-ingest \u2014 all 55,742 indexed chunks preserved; confirmed connected and serving queries\n* Downgrade to `chromadb<1.0.0` is not possible on Python 3.14 (no pre-built `chroma-hnswlib` wheel)", "# Cellular Uptake of Microplastics: Pathways, Persistence, and Therapeutic Potential for Intracellular Targeting of PETase\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-11-07 on marcdeller.com (categories: blog, microplastics).*\n\n* [Impact of Microplastics and Nanoplastics on Human Health](https://www.mdpi.com/2079-4991/11/2/496)\n* [Mannose-6-phosphate glycan for lysosomal targeting: various applications from enzyme replacement therapy to lysosome-targeting chimeras](https://www.tandfonline.com/doi/full/10.1080/19768354.2022.2079719#abstract)\n\n* **Endocytosis mechanisms govern uptake**: Microplastics larger than 500 nanometers enter cells through [phagocytosis](https://www.mdpi.com/2079-4991/11/2/496), while particles between 50 and 200 nanometers predominantly use [clathrin-mediated endocytosis](https://www.mdpi.com/2079-4991/11/2/496) as their primary cellular entry pathway.\n* **Lysosomal persistence drives cellular burden**: [Microplastics resist degradation](https://ecancer.org/en/news/24366-micro-and-nanoplastics-in-the-body-are-passed-on-during-cell-division) within lysosomes and are passed to daughter cells during division, enabling long-term accumulation and potential promotion of cancer cell migration.\n* **Cancer cells accumulate more particles**: [Microplastics concentrate at higher levels](https://pmc.ncbi.nlm.nih.gov/articles/PMC11720132/) in tumor tissue compared to normal tissue, potentially due to enhanced uptake rates and altered intratumoral circulation patterns.\n* **Therapeutic enzymes show promise**: [FAST-PETase degrades PET microplastics](https://www.mdpi.com/2079-4991/11/2/496) into non-toxic monomers in human serum at 37 degrees Celsius without affecting cell viability, highlighting potential biomedical applications.\n* **Delivery of PETase into cells is feasible, as demonstrated by FDA-approved enzyme replacement therapies** **(ERT):**[The mannose-6-phosphate receptor pathway](https://pmc.ncbi.nlm.nih.gov/articles/PMC8757330/) represents the primary mechanism by which lysosomal enzymes are delivered to their target compartments.\n\n1. **Impact of Microplastics and Nanoplastics on Human Health:** [Yee, M.S.L. et al., Nanomaterials, 2021](https://doi.org/10.3390/nano11020496)\n2. **Micro and nanoplastics in the body are passed on during cell division:** [Pichler, V., et al., ecancer, 2024](https://ecancer.org/en/news/24366-micro-and-nanoplastics-in-the-body-are-passed-on-during-cell-division)\n3. **Degradation of PET microplastic particles to monomers in human serum by PETase:** [Lopez-Lorenzo, X., et al., Faraday Discussions, 2024](https://doi.org/10.1039/d4fd00014e)\n4. **Engineering IGF2 for Lysosome-targeting chimeras development to target drug-resistant membrane proteins in tumor therapy:** [Pan, Y., et al., Protein Science, 2025](https://pubmed.ncbi.nlm.nih.gov/39969096/)\n5. **Lysosome-Targeting Strategy Using Polypeptides and Chimeric Molecules:** [Mondal, B., et al., ACS Omega, 2021](https://pmc.ncbi.nlm.nih.gov/articles/PMC8757330/)\n6. **Mannose-6-phosphate glycan for lysosomal targeting: various applications from enzyme replacement therapy to lysosome-targeting chimeras:** [Seo, Jinho et al., Journal of Glycobiology, 2022](https://www.tandfonline.com/doi/full/10.1080/19768354.2022.2079719)\n\n**Breaking barriers through enzymatic innovation for a plastic-laden world: PETase delivery into cells**\n\nCellular internalization of microplastics and nanoplastics depends critically on particle size, with distinct uptake mechanisms governing their entry into cells. Particles larger than 500 nanometers primarily enter through [phagocytosis and macropinocytosis](https://www.mdpi.com/2079-4991/11/2/496), both energy-intensive processes that rely on actin-dependent membrane remodeling. Smaller particles between 50 and 200 nanometers predominantly use [clathrin-mediated endocytosis](https://www.mdpi.com/2079-4991/11/2/496), while caveolin-mediated endocytosis handles specific size ranges. Nanoplastics below 100 nanometers can penetrate cellular barriers more efficiently than larger microplastics, inducing [greater oxidative stress and mitochondrial dysfunction](https://pmc.ncbi.nlm.nih.gov/articles/PMC9208838/). Environmental exposure significantly enhances uptake efficiency, as particles coated with biomolecules from natural waters are internalized more readily than pristine particles. Once inside cells, these particles localize primarily within [endosomal and lysosomal compartments](https://pmc.ncbi.nlm.nih.gov/articles/PMC8757330/).\n\nThe lysosomal sequestration of microplastics represents both a protective mechanism and a source of cellular stress. Lysosomes function as cellular waste management centers, containing acidic environments and hydrolytic enzymes designed to break down biological materials. However, microplastics and nanoplastics resist enzymatic degradation due to their synthetic chemical composition. Rather than being digested, these particles accumulate within lysosomes, where they persist and are [inherited by daughter cells during division](https://ecancer.org/en/news/24366-micro-and-nanoplastics-in-the-body-are-passed-on-during-cell-division). In cancer cells, this accumulation appears particularly problematic, with studies showing that nanoplastic-exposed colorectal cancer cells exhibit increased migratory behavior, potentially promoting metastasis. Three-dimensional tumor spheroid models demonstrate that microplastics concentrate in the nutrient-deprived, non-proliferating core regions, where limited cell division prevents dilution of particle burden.", "# Nanobodies: Revolutionizing Molecular Medicine Through Nature\u2019s Smallest Antibodies\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nDiagnostic applications have benefited significantly from nanobodies\u2019 rapid kinetics and exceptional stability. Whole-cell biosensors utilizing nanobody-displayed yeasts have achieved detection limits as low as 0.037 \u03bcg/mL for SARS-CoV-2 spike proteins, demonstrating the potential for rapid, cost-effective viral detection. The development of nanobody-based imaging probes has enhanced molecular imaging capabilities, with several agents progressing through clinical trials for PET and SPECT imaging of various cancer biomarkers.\n\nThe COVID-19 pandemic highlighted nanobodies\u2019 potential for rapid therapeutic development against emerging pathogens. Researchers at the University of Minnesota developed multiple potent nanobodies, including Nanosota-2, -3, and -4, which demonstrated exceptional efficacy against SARS-CoV-2 variants. Nanosota-2 achieved an IC50 of 2 picomolar against the prototypic strain, representing among the most potent SARS-CoV-2 entry inhibitors discovered. The ability to rapidly engineer nanobodies against new viral variants through structure-guided design and phage display provides a powerful platform for pandemic preparedness.\n\nRecent innovations in nanobody production have moved beyond traditional camelid immunization to embrace synthetic approaches. Artificial intelligence and machine learning algorithms now enable rational design of nanobodies with predetermined properties, significantly accelerating the development process. Synthetic nanobody libraries, combined with advanced screening techniques including mass spectrometry and deep sequencing, have improved hit rates and expanded the diversity of targetable epitopes. These technological advances have reduced development timelines while enhancing the quality and specificity of resulting nanobodies.", "# 2026 \u2013 Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-02 on marcdeller.com (categories: papers).*\n\n* [Link to paper](https://marcdeller.com/wp-content/uploads/2026/06/povo_article.pdf)\n\n* **Selectivity by geometry:** Povorcitinib exploits a roughly 2 \u00c5 P-loop shift that opens a lipophilic groove only in JAK1.\n* **Sparing JAK2:** Leaving JAK2 intact avoids the anemia and cytopenias tied to erythropoietin and thrombopoietin blockade.\n* **Residues that decide:** JAK1 His885 faces solvent while JAK2 Asn859 and Phe860 close the pocket against the inhibitor\u2019s tail.\n* **Oral exposure by design:** An intramolecular hydrogen bond masks the amide\u2019s polarity, lifting permeability and oral bioavailability across species.\n* **From structure to clinic:** The 46-fold-selective design now shows durable activity in hidradenitis suppurativa, validating the JAK1-only strategy.\n\n1. **Discovery of the Orally Bioavailable Isoform Selective JAK1 Compound Povorcitinib (INCB054707):** [Zhuo et al., J. Med. Chem., 2026](https://doi.org/10.1021/acs.jmedchem.5c03753)\n2. **A Comprehensive Overview of Globally Approved JAK Inhibitors:** [Shawky et al., Pharmaceutics, 2022](https://doi.org/10.3390/pharmaceutics14051001)\n3. **Functional and Structural Characterization of Clinical-Stage JAK2 Inhibitors Identifies Determinants for Drug Selectivity:** [Miao et al., J. Med. Chem., 2024](https://doi.org/10.1021/acs.jmedchem.4c00197)\n4. **Intramolecular Hydrogen Bonding: An Opportunity for Improved Design in Medicinal Chemistry:** [Caron et al., Med. Res. Rev., 2019](https://doi.org/10.1002/med.21562)\n5. **Efficacy and Safety of the Oral JAK1 Inhibitor Povorcitinib in Hidradenitis Suppurativa:** [Kirby et al., J. Am. Acad. Dermatol., 2024](https://www.jaad.org/article/S0190-9622(23)03037-2/fulltext)\n\n**How a retracted loop and a fluorinated tail let povorcitinib tell two near-identical kinase pockets apart**\n\nThe four Janus kinases share an ATP pocket so conserved that inhibiting one without the others is a real medicinal-chemistry problem, which is why most early JAK drugs hit two or more isoforms at once, as cataloged across the [approved JAK inhibitor class](https://doi.org/10.3390/pharmaceutics14051001). That promiscuity has a cost. JAK2 carries erythropoietin and thrombopoietin signaling, so blocking it invites anemia, thrombocytopenia, and other cytopenias, a liability the [structural profiling of JAK2 inhibitors](https://doi.org/10.1021/acs.jmedchem.4c00197) ties directly to its role in hematopoiesis. JAK1 tells a different story. It sits at the center of the inflammatory cytokine network, pairing with JAK2, JAK3, and TYK2 to relay interleukin and interferon signals. A compound that shuts down JAK1 while leaving JAK2 untouched should calm autoimmune disease without starving the bone marrow. The hard part is finding something in the binding site that actually distinguishes the two.\n\nA recent Incyte [paper of mine](https://doi.org/10.1021/acs.jmedchem.5c03753) reconstructs how povorcitinib (INCB054707) was engineered to do exactly that. Co-crystal structures of the inhibitor bound to the isolated JAK1 and JAK2 kinase domains, deposited as PDB [10PI](https://www.rcsb.org/structure/unreleased/10PI) and [10PJ](https://www.rcsb.org/structure/unreleased/10PJ), show that its 3,5-dimethylpyrazole hinge binder forms the expected hydrogen bonds with the backbone of Leu932 and Glu930 in both isoforms. Selectivity is not written at the hinge, where the two kinases look almost identical. It is written further out\u00a0along the elongated arm of the molecule, which extends past the gatekeeper toward the glycine-rich P-loop.\n\n> **\u201cJAK1 selectivity was conferred by the amide\u2019s alkyl group accessing a small lipophilic groove\u201d**\n\nThere, the two isoforms diverge. In JAK1, the P-loop rides about 2 \u00c5 higher, opening a shallow lipophilic groove that neatly accepts the terminal trifluoro-isopropyl amide. In JAK2, the equivalent loop closes:\u00a0JAK1 presents His885 pointing into the solvent, JAK2 swings Asn859 inward to hydrogen-bond with Glu1015, and a nearby Phe860 rotates the pocket shut. The same fluorinated tail that fits JAK1 now clashes, and affinity drops.\n\nTwo design moves made the difference. Trading the metabolically vulnerable pyrrolopyrimidine for the smaller dimethylpyrazole pushed that tail deeper into the JAK1-specific space, roughly doubling isoform selectivity to about 46-fold. An [intramolecular hydrogen bond](https://doi.org/10.1002/med.21562) then masked the amide\u2019s polarity, rescuing permeability and oral exposure across species. The payoff is visible beyond the crystallography: povorcitinib has produced durable responses in [hidradenitis suppurativa](https://www.jaad.org/article/S0190-9622(23)03037-2/fulltext), with 2026 long-term readouts reinforcing the safety case for selective JAK1 inhibition. A two-residue difference, read correctly, becomes a drug.", "# GitHub README: ButtFold\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ButtFold, 2026\n\n> **A protein folds in your browser and the trajectory becomes music.**\n\n[](https://buttfold.mdeller.com)               \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>App</b></td><td><a href=\"https://buttfold.mdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">buttfold.mdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/ButtFold\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/ButtFold</a></td>\n</tr>\n</table>\n\n---\n\nButtFold takes a named protein, starts it as a self-avoiding random coil, and folds it with a\nC-alpha structure-based (G\u014d) model until it reaches its native structure. The trajectory is the\ninstrument: contacts forming are note onsets, sequence separation sets register, hydropathy sets\npitch, secondary structure sets texture (helix a sustained pad, sheet a staccato figure, coil an\narpeggio), and compaction drives an accelerando. It is the web version of\n[PhoneFold](https://github.com/bellcheddar/PhoneFold), and it plays the same piece: `Sonifier.js`\nis held to note-for-note agreement with the shipped Swift, 15,536 notes across two proteins and\nall five styles, zero differences.\n\nWhy it matters: most things that show a protein \"folding\" are showing you a structure prediction\nbeing refined, which is a picture of a computation rather than of a molecule, and the ones that\nsonify proteins almost always sonify the **sequence**, which is a lookup table set to music. This\ndoes neither. The physics is a real dynamical trajectory computed from scratch in your browser,\nand every note in the piece is caused by something the chain actually did, which makes the music a\nreading of the fold instead of a decoration on it. It is useful for: teaching what a folding funnel\nsounds like, showing a coarse-grained simulation to people who will never open a terminal, and\nhearing the difference between a protein that packs a hydrophobic core and one that never resolves.\n\n**What it is not.** This is a funnelled coarse-grained simulation of a protein whose structure is\nalready known, in the tradition of Clementi, Nymeyer and Onuchic (*J. Mol. Biol.* **298**:937,\n2000). It is not a structure prediction, it is not an unbiased physical force field, and no protein\nfolds this way. The app says so on the page, next to the thing it is describing, rather than in an\nAbout box.\n\n## \ud83e\uddf1 Three ways to get a trajectory, one player\n\nEverything downstream of \"a stream of CA frames\" is shared. The three sources produce\nbyte-identical frame objects, which is tested rather than asserted.\n\n| | What it is | Compute | When it runs |\n|---|---|---|---|\n| **Gallery** | Six G\u014d folds and three Genie 2 backbones, precomputed | none | Always. First paint, no capability check, no wait |\n| **Fold it live** | The same C compiled to WebAssembly, in a Web Worker | your CPU | The primary interactive path where the browser can run it |\n| **ESMFold** | ESMFold predicts a real UniProt protein at Meta; the same C, compiled natively, folds a chain toward that prediction here | ~1 s at Meta, then the droplet | Pick a protein nobody precomputed |\n| **Genie 2** | A diffusion model invents a backbone from noise; the denoising *is* the trajectory | baked on a Mac, once | Always, in the gallery |\n\nA fold the server computes is baked into the gallery's own artefact format, so a queued result\nplays through the identical player with no code path of its own, and converges the cache: a\nprotein folded once is served from disk forever after.\n\n**The ESMFold engine keeps two claims apart.** ESMFold predicts where a protein ends up; the\nG\u014d model animates a chain collapsing toward that prediction. The second was never a physical\nfolding pathway and the first can be wrong, so the badge says `predicted at Meta, folded here`\nrather than anything implying this server did the science. It runs at Meta because\n`facebook/esmfold_v1` is an 8.44 GB checkpoint and the droplet has 3.9 GB with no swap \u2014\nnot slow, impossible.\n\nThe 24 proteins on offer were **screened by predicting every candidate**, not by name. Ranking\nUniProt's reviewed 40\u201376-residue entries by how well studied they are returns twenty-five\nribosomal proteins in the top twenty-five, and a ribosomal protein is only folded inside the\nribosome. Two measurements do the filtering \u2014 ESMFold's own mean pLDDT \u2265 0.70, and a radius of\ngyration within 1.20\u00d7 of the folded-globule scaling the sonifier already uses. That second bar\nwas moved from 1.35 by a fold that failed its bake gate, and the three entries it removed were\na dimer, an 11-mer and a dodecamer: proteins that *are* folded alone but are only compact as an\nassembly, which is the ribosomal error one step subtler.\n\n**Genie 2 runs the trajectory backwards, and almost none of that is cosmetic.** A G\u014d fold\nstarts as an extended coil and collapses; a diffusion trajectory starts with every residue\npiled into a ball of radius 1.1 \u00c5 and *inflates* into a protein, growing secondary structure\non the way out. Measured on an 80-residue sample: Rg 1.1 \u2192 11.2 \u00c5, and 3003 of 3004 contacts\n\"form\" on frame one because at the start everything is within any distance you care to name.", "# Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\nAuthors: Hummel, J.R.; Xiao, K.-J.; Yang, J.C.; Epling, L.B.; Mukai, K.; Ye, Q.; Xu, M.; Qian, D.; Huo, L.; Weber, M.; Roman, V.; Lo, Y.; Drake, K.; Stump, K.; Covington, M.; Kapilashrami, K.; Zhang, G.; Ye, M.; Diamond, S.; Yeleswaram, S.; Macarron, R.; Deller, M.C.; Wee, S.; Kim, S.; Wang, X.; Wu, L.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2024\n\nTo lower the human intrinsic clearance (h-Cl) of 2-chlorophenyl derivative 7, we hypothesized that incorporating small polar substituents in the aromatic ring might reduce lipophilicity, which in turn could potentially increase metabolic stability. The addition of a cyano group (10) at the 4-position maintained potency and improved intrinsic clearance (0.5 L/h/kg) but also resulted in poor SGF solubility (Table 2, entry 1). Reasoning that incorporation of an N-methyl benzyl amine might improve aqueous solubility by both disrupting molecular planarity and providing an ionizable group, we were encouraged to find that compound 11 showed improved SGF solubility (94 \u03bcg/mL). Unfortunately 11 also showed lower selectivity against CDK1 and CDK7 (entry 2). Revisiting the R1 substituent at the pyrimidine C5-position, we discovered that switching to a cyano group (12) provided a notable increase in both CDK1 and CDK7 selectivity while further improving solubility (entry 3). Compound 12 also displayed a cellular IC50 of 18 nM against CCNE1-amplified COV318 ovarian cancer cells and showed good translation in the human whole blood assay (WB IC50 = 112 nM), which we hypothesize was due to the increased protein binding free fraction (FF = 33%). Cellular activity against CDK1 was also characterized in OVCAR3 cells by measuring the inhibition of nucleophosmin (NPM) phosphorylation at threonine 199, with compound 12 exhibiting a CDK1 cellular IC50 of 225 nM.\n\nTo improve the aqueous solubility of trifluoroethyl derivative 8, we first evaluated polar functionality at the R3 position of the sulfonyl moiety (Table 2). Extensive SAR in this region revealed that a variety of aromatic nitrogen heterocycles were well tolerated, often showing subnanomolar biochemical potency and improved selectivity against off-target CDKs. For example, the addition of an N-methylpyrazole (13) showed promising CDK2 WB potency, low intrinsic clearance, and high caco2 permeability, but unfortunately no improvement in solubility was observed (entry 4). To our delight, a nitrogen walk to the more polar N-methylimidazole (14) showed a dramatic improvement in SGF solubility (322 \u03bcg/mL), while maintaining a balanced in vitro ADME profile (entry 5). Compound 14 also displayed promising CDK isoform selectivity, and further improvement in WB potency to 334 nM. With a favorable compound profile in hand, we next fixed the R2 and R3 substituents and evaluated a cyano substituent at the R1 position (entry 6). In contrast to the trend observed for N-methyl benzyl amines 11\u221212, compound 15 showed much lower solubility than the trifluoromethyl congener 14. Compound 15 also showed an improvement in CDK1 and CDK7 selectivity, albeit with a reduction in CDK4 and CDK6 selectivity as well as caco2 permeability. Additional attempts to balance the aqueous solubility and permeability through structural modification at R3 proved fruitless, as exemplified by 2-pyridyl derivative 16 that displayed increased caco2 permeability with concomitant reduction in SGF solubility (entry 7).\n\nThe promising solubility and intrinsic clearance of compound 9 prompted us to fix the R1 and R2 positions and evaluate substituents at the R3 position of the sulfonyl moiety (Table 2). Following SAR studies, we discovered that compound 17 bearing a cyclopropyl substituent showed improved WB potency (398 nM) and promising selectivity over off-target CDKs (entry 8). Furthermore, compound 17 exhibited excellent SGF solubility (>400 \u03bcg/mL), low intrinsic clearance (0.5 L/h/kg), and high caco2 permeability. Interestingly, converting compound 17 into the cyano derivative 18 resulted in diminished CDK1, CDK4, and CDK6 enzyme selectivity as well as reduced SGF solubility (entry 9).\n\n*Table 2.* SAR to Balance Potency, ADME, and Physiochemical Properties. Mean IC50 values measured in the presence of 1 mM ATP. Retinoblastoma protein (Rb) S780 phosphorylation measured in the CCNE1-amplified COV318 cell line. Nucleophosmin (NPM) T199 phosphorylation measured in the OVCAR3 cell line. Treated with compounds in human whole blood overnight. Human microsomal intrinsic clearance (L/h/kg). Caco2 values Pm \u00d7 (10\u22126) cm/s. Protein binding free fraction (FF).\n\n| entry | cmpd | R1 (pyrimidine C5) | R2 (pyrazole N-substituent) | R3 (sulfonyl substituent) | properties reported in the text |\n| --- | --- | --- | --- | --- | --- |\n| 1 | 10 | CF3 | 2-chloro-4-cyanophenyl | methyl | maintained potency; h-Cl 0.5 L/h/kg; poor SGF solubility |\n| 2 | 11 | CF3 | 2-chloro-4-((methylamino)methyl)phenyl | methyl | SGF solubility 94 \u03bcg/mL; lower CDK1 and CDK7 selectivity |\n| 3 | 12 | CN | 2-chloro-4-((methylamino)methyl)phenyl | methyl | increased CDK1 and CDK7 selectivity; improved solubility; COV318 cellular IC50 = 18 nM; WB IC50 = 112 nM; FF = 33%; CDK1 cellular IC50 = 225 nM (NPM T199, OVCAR3) |\n| 4 | 13 | CF3 | 2,2,2-trifluoroethyl | 1-methyl-1H-pyrazol-4-yl | promising CDK2 WB potency; low intrinsic clearance; high caco2 permeability; no improvement in solubility |\n| 5 | 14 | CF3 | 2,2,2-trifluoroethyl | 1-methyl-1H-imidazol-4-yl | SGF solubility 322 \u03bcg/mL; balanced ADME; promising CDK isoform selectivity; WB IC50 = 334 nM |\n| 6 | 15 | CN | 2,2,2-trifluoroethyl | 1-methyl-1H-imidazol-4-yl | lower solubility than 14; improved CDK1 and CDK7 selectivity; reduced CDK4/CDK6 selectivity and caco2 permeability |\n| 7 | 16 | CN | 2,2,2-trifluoroethyl | pyridin-2-yl | increased caco2 permeability with reduced SGF solubility |\n| 8 | 17 | CF3 | 2-hydroxy-2-methylpropyl | cyclopropyl | CDK2 IC50 = 0.29 nM; WB IC50 = 398 nM; promising off-target CDK selectivity; SGF solubility > 400 \u03bcg/mL; h-Cl 0.5 L/h/kg; high caco2 permeability |\n| 9 | 18 | CN | 2-hydroxy-2-methylpropyl | cyclopropyl | diminished CDK1, CDK4 and CDK6 enzyme selectivity; reduced SGF solubility |\n\n### Cross-species pharmacokinetics", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\nUpregulation of the fibroblast growth factor receptor (FGFR) signaling pathway has been implicated in multiple cancer types, including cholangiocarcinoma and bladder cancer. Consequently, small molecule inhibition of FGFR has emerged as a promising therapy for patients suffering from these diseases. First-generation pan-FGFR inhibitors, while highly effective, suffer from several drawbacks. These include treatment-related hyperphosphatemia and significant loss of potency for the mutant kinases. Herein, we present the discovery and optimization of novel FGFR2/3 inhibitors that largely maintain potency for the common gatekeeper mutants and have excellent selectivity over FGFR1. A combination of meticulous structure-activity relationship (SAR) analysis, structure-based drug design, and medicinal chemistry rationale ultimately led to compound 29, a potent and selective FGFR2/3 inhibitor with excellent in vitro absorption, distribution, metabolism, excretion (ADME), and pharmacokinetics in rat. A pharmacodynamic study of a closely related compound established that maximum inhibition of downstream ERK phosphorylation could be achieved with no significant effect on serum phosphate levels relative to vehicle.\n\n## Structure-Function Analysis of the Extended Conformation of a Polyketide Synthase Module (2018)\n\n**Authors:** Li X, Sevillano N, La Greca F, Deis L, Liu YC, Deller MC, Mathews II, Matsui T, Cane DE, Craik CS, Khosla C.  \n**Journal:** Journal of the American Chemical Society (2018). DOI: 10.1021/jacs.8b02100. PMID: 29762030.\n\nCatalytic modules of assembly-line polyketide synthases (PKSs) have previously been observed in two very different conformations-an \"extended\" architecture and an \"arch-shaped\" architecture-although the catalytic relevance of neither has been directly established. By the use of a fully human na\u00efve antigen-binding fragment (Fab) library, a high-affinity antibody was identified that bound to the extended conformation of a PKS module, as verified by X-ray crystallography and tandem size-exclusion chromatography-small-angle X-ray scattering (SEC-SAXS). Kinetic analysis proved that this antibody-stabilized module conformation was fully competent for catalysis of intermodular polyketide chain translocation as well as intramodular polyketide chain elongation and functional group modification of a growing polyketide chain. Thus, the extended conformation of a PKS module is fully competent for all of its essential catalytic functions.\n\n## Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design (2017)\n\n**Authors:** Irimia A, Serra AM, Sarkar A, Jacak R, Kalyuzhniy O, Sok D, Saye-Francisco KL, Schiffner T, Tingle R, Kubitz M, Adachi Y, Stanfield RL, Deller MC, Burton DR, Schief WR, Wilson IA.  \n**Journal:** PLoS pathogens (2017). DOI: 10.1371/journal.ppat.1006212. PMID: 28225819.\n\nAmong broadly neutralizing antibodies to HIV, 10E8 exhibits greater neutralizing breadth than most. Consequently, this antibody is the focus of prophylactic/therapeutic development. The 10E8 epitope has been identified as the conserved membrane proximal external region (MPER) of gp41 subunit of the envelope (Env) viral glycoprotein and is a major vaccine target. However, the MPER is proximal to the viral membrane and may be laterally inserted into the membrane in the Env prefusion form. Nevertheless, 10E8 has not been reported to have significant lipid-binding reactivity. Here we report x-ray structures of lipid complexes with 10E8 and a scaffolded MPER construct and mutagenesis studies that provide evidence that the 10E8 epitope is composed of both MPER and lipid. 10E8 engages lipids through a specific lipid head group interaction site and a basic and polar surface on the light chain. In the model that we constructed, the MPER would then be essentially perpendicular to the virion membrane during 10E8 neutralization of HIV-1. As the viral membrane likely also plays a role in selecting for the germline antibody as well as size and residue composition of MPER antibody complementarity determining regions, the identification of lipid interaction sites and the MPER orientation with regard to the viral membrane surface during 10E8 engagement can be of great utility for immunogen and therapeutic design.\n\n## The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase (2017)\n\n**Authors:** Span EA, Suess DLM, Deller MC, Britt RD, Marletta MA.  \n**Journal:** ACS chemical biology (2017). DOI: 10.1021/acschembio.7b00016. PMID: 28257189.\n\nPolysaccharide monooxygenases (PMOs) are secreted metalloenzymes that catalyze the oxidative degradation of polysaccharides in a copper-, oxygen-, and reductant-dependent manner. Cellulose-active fungal PMOs degrade cellulosic substrates to be utilized as a carbon source for fungal growth. To gain insight into the PMO mechanism, the role of conserved residues in the copper coordination sphere was investigated. Here, we report active-site hydrogen-bonding motifs in the secondary copper coordination sphere of MtPMO3*, a C1-oxidizing PMO from the ascomycete fungus Myceliophthora thermophila. A series of point substitutions that disrupt this conserved network are used to interrogate its function. Activity assays, in conjunction with EPR spectroscopy, demonstrate that residues H161 and Q167 are involved in stabilizing bound oxygen, and H161 appears to play a role in proton transfer. Additionally, Q167 increases the ligand donor strength of Y169 to the copper via a hydrogen-bonding interaction. Altogether, H161 and Q167 are important for oxygen activation, and the results are suggestive of a copper-oxyl active intermediate.\n\n## Validation of Protein-Ligand Crystal Structure Models: Small Molecule and Peptide Ligands (2017)", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics; supervisor Prof. E. Yvonne Jones), 1999\n\nFlash cooling of the crystals proved problematic, presumably due to the high viscosity of mother liquor solutions and the associated lack of heat transfer. Passing crystals through solutions of lower molecular weight PEG alleviated the problem, so allowing MAD analysis of single crystals to be carried out. An EMP derivative of OSM was produced and this was shown to bind at two major and one minor site (see Figure 5-5 (a)). Although the initial identification of the derivative was achieved with room temperature SIRAS data, crystal decay and non-isomorphism underscored the necessity for the single crystal MAD experiment. A fluorescence scan was used to characterise the mercury LIII edge of the derivative and a traditional three wavelength MAD experiment carried out. Phases based on this experiment were supplemented by the addition of further phasing from the independent SIRAS experiment and the resulting maps showed improved quality.\n\nThe current model has a crystallographic-R value of 20.5 % with a free-R value of 26.1 % for all data between 20.0\u20132.2 \u00c5. A second crystal form belonging to the spacegroup I222 was subsequently solved using molecular replacement and this structure has been refined to a crystallographic R-value of 20.8 % and an R-free value of 28.9 % for all data between 20.0\u20132.5 \u00c5.\n\nThe final model shows significant deviations to the initial model generated using Modeller, with an RMSD value of 2.40 over 151 structurally equivalent C\u03b1. This suggests possible reasons for the failure of molecular replacement techniques; although solutions using the refined model coordinates of human LIF and murine LIF also failed.\n\n## References\n\n*(reference list omitted)*", "# Protein stability: a crystallographer's perspective\nAuthors: Deller, M. C.; Kong, L.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 72, 72\u201395 (IYCr crystallization series), 2016\n\nSpectroscopy is used primarily at the level of secondary structure, using circular dichroism (CD), Fourier-transform infrared (FT-IR) and Raman spectroscopy. Protein CD, or spectropolarimetry, is performed in the far-UV range (170\u2013250 nm) and has enjoyed a resurgence with synchrotron-radiation circular dichroism (SRCD), which operates at lower wavelengths; SRCD data are deposited in the Protein Circular Dichroism Data Bank (PCDDB), which at the time of writing held 529 entries. Interpretation is straightforward in outline: a twin minimum in ellipticity at 208 and 222 nm indicates \u03b1-helical content, whereas a single minimum at 204 nm suggests random coil and one at 217 nm suggests \u03b2-sheet. As an applied example, UV-CD analysis of the plant membrane protein MBP-b6 quantified helix, sheet, turn and unordered content and showed that n-dodecyl-\u03b2-D-maltoside and Triton X-100 preserved correct secondary structure while sodium dodecyl sulfate disrupted it. Infrared spectroscopy detects changes in amide-group bond oscillation caused by altered hydrogen-bonding patterns, and femtosecond two-dimensional infrared spectroscopy (2D-IR) offers greater sensitivity to structural differences.\n\n### 5.4. Melting-temperature analysis\n\nGlobal thermal stability is represented by the thermal denaturation midpoint or melting temperature (Tm), determined by Thermofluor (differential scanning fluorimetry, DSF) or differential scanning calorimetry (DSC). Thermofluor uses a reporter dye such as SYPRO Orange or 1-anilinonaphthalene-8-sulfonate (ANS) whose fluorescence changes on binding the exposed hydrophobic core of unfolding protein; its major advantages are very low protein consumption (typically well under 1 mg) and the fact that the only instrument needed is a standard qPCR machine. Applications include detecting ligand binding to proteins of unknown function \u2014 screening 3,000 compounds identified an essential *Streptococcus pneumoniae* gene product as a nucleoside diphospho-keto-sugar aminotransferase \u2014 and quantifying protein\u2013protein interactions, as in probing MBP stabilization of ankyrin-repeat proteins with alanine mutants. Hundreds of truncations and buffer conditions can be screened, and the method extends to membrane proteins in detergents, for example the acetylcholine receptor in Brij-35. Caveats include protein\u2013dye interactions that can themselves destabilize (observed for GroEL with ANS), common aberrant or false-positive thermal shifts requiring careful melt-curve analysis with tools such as Meltdown, proteins that do not unfold with a clean sigmoid (high-disulfide proteins such as albumin, and thermophilic proteins), and the fact that a negative result does not prove absence of binding \u2014 the ligand may simply not stabilize further.\n\n### 6. Improving protein stability and the concept of crystallizability\n\nThe unifying theme of crystallizability-enhancing techniques is improvement of both compositional and conformational stability (Table 4). The toolbox comprises truncations, buffer screening, ligand and additive screening, purification tags and crystallization chaperones, reductive methylation, surface-entropy reduction (SER), in situ proteolysis, Thermofluor, deuterium-exchange mass spectrometry (DXMS) and disulfide engineering, many of them deployed as \"salvage\" methods by high-throughput structural genomics consortia. Numerous software packages predict crystallizability, including DisMeta, XtalPred, POODLE, MFDp2, MoRFpred, RFCRYS, XANNpred, SVMCRYS, SCMCRYS, CRYSTALP2, MetaPPCP and ParCrys; many are template-based comparisons with known structures, while XtalPred, POODLE and DisMeta use sequence-based prediction of low complexity, disorder, transmembrane segments and signal peptides. A key caveat is that these predictors evaluate a single polypeptide chain in isolation, whereas complex formation frequently stabilizes the constituents \u2014 so for crystallographic work it is often essential to study the whole complex rather than its individually disordered parts.\n\n### 6.1. Truncations and domain selection\n\nSelecting the shortest possible domain is generally preferable, and computational tools such as ESPRIT (Expression of Soluble Proteins by Random Incremental Truncation) and combinatorial domain hunting (CDH) assist. Truncation to the shortest workable fragment is often decisive: amyloid fibres, of great medical importance in Alzheimer's and prion diseases, resisted crystallographic analysis because of their partial disorder, yet fragments of only 6\u20137 amino acids that still form fibrils yielded microcrystals and structures. Structured truncation arrays generated by polymerase incomplete primer extension (PIPE) cloning allowed the Joint Center for Structural Genomics (JCSG) and the Structural Genomics Institute at Karolinska Institutet to produce several thousand truncations of recalcitrant targets. Beyond truncation, point mutations matter: stability-enhancing mutations in membrane proteins are surprisingly common, with estimates that around 10% of random mutations confer some stabilization, exemplified by two valine-to-alanine substitutions in the transmembrane portion of the M13 coat protein. The reasons are not always evident from structure; it has been suggested that membrane proteins must be inherently flexible, and hence conformationally unstable, to function within the restricted membrane environment.\n\n### 6.2. Buffer screening and 6.3. Ligands and additive screening", "# Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3\nAuthors: Shvartsbart, A.; Roach, J.J.; Witten, M.R.; Koblish, H.; Harris, J.J.; Covington, M.; Hess, R.; Lin, L.; Frascella, M.; Truong, L.; Leffet, L.; Conlen, P.; Beshad, E.; Klabe, R.; Katiyar, K.; Kaldon, L.; Young-Sciame, R.; He, X.; Petusky, S.; Chen, K.-J.; Horsey, A.; Lei, H.-T.; Epling, L.B.; Deller, M.C.; Vechorkin, O.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2022\n\n**Authors:** Artem Shvartsbart (corresponding author), Jeremy J. Roach, Michael R. Witten, Holly Koblish, Jennifer J. Harris, Maryanne Covington, Rodrigo Hess, Luping Lin, Michelle Frascella, Lisa Truong, Lynn Leffet, Patricia Conlen, Elham Beshad, Ron Klabe, Kamna Katiyar, Laura Kaldon, Ruth Young-Sciame, Xin He, Susan Petusky, Kwang-Jong Chen, April Horsey, Hsiang-Ting Lei, Leslie B. Epling, Marc C. Deller, Oleg Vechorkin and Wenqing Yao. **Affiliations (condensed):** Incyte Corporation, Wilmington, Delaware 19803, United States; with present addresses for individual authors at Prelude Therapeutics (Wilmington, Delaware), Proteovant Therapeutics (King of Prussia, Pennsylvania), Ikena Oncology (Boston, Massachusetts), WuXi AppTec (Philadelphia, Pennsylvania) and Synnovation Therapeutics (Wilmington, Delaware). **Published:** J. Med. Chem., volume 65, pages 15433\u201315442, 2022. Received 18 August 2022; published 10 November 2022.\n\n## Abstract\n\nUpregulation of the fibroblast growth factor receptor (FGFR) signaling pathway has been implicated in multiple cancer types, including cholangiocarcinoma and bladder cancer. Consequently, small molecule inhibition of FGFR has emerged as a promising therapy for patients suffering from these diseases. First-generation pan-FGFR inhibitors, while highly effective, suffer from several drawbacks. These include treatment-related hyperphosphatemia and significant loss of potency for the mutant kinases. Herein, we present the discovery and optimization of novel FGFR2/3 inhibitors that largely maintain potency for the common gatekeeper mutants and have excellent selectivity over FGFR1. A combination of meticulous structure\u2212activity relationship (SAR) analysis, structure-based drug design, and medicinal chemistry rationale ultimately led to compound 29, a potent and selective FGFR2/3 inhibitor with excellent in vitro absorption, distribution, metabolism, excretion (ADME), and pharmacokinetics in rat. A pharmacodynamic study of a closely related compound established that maximum inhibition of downstream ERK phosphorylation could be achieved with no significant effect on serum phosphate levels relative to vehicle.\n\n## Introduction\n\nFibroblast growth factor receptors (FGFR1\u22124) are a family of transmembrane receptor tyrosine kinases that play a role in several physiological processes, including embryonic development, metabolic homeostasis, and angiogenesis. Aberrant activation of FGFR signaling leads to proliferation and survival of cancer cells, with FGFR3 alterations most commonly found in bladder cancer, while FGFR2 alterations are mainly found in cholangiocarcinoma, gastric cancer, and uterine cancer.\n\nThree ATP-competitive small molecule pan-FGFR (FGFR1\u22123) inhibitors have been approved by the FDA for the treatment of cancers with FGFR genetic alterations (Figure 1). Erdafitinib (1) was approved to treat locally advanced or metastatic urothelial carcinoma (bearing FGFR3 or FGFR2 genetic alteration), while pemigatinib (2) and infigratinib (3) were approved to treat locally advanced or metastatic cholangiocarcinoma (bearing FGFR2 alteration). The most common treatment-related toxicity for these drugs is hyperphosphatemia. Since antibody-mediated FGFR1 activation in mice leads to hypophosphatemia, while FGFR1/4 knockout mice experienced hyperphosphatemia, this toxicity has been hypothesized to be due to FGFR1 inhibition.\n\nInhibitors 1\u22123 derive a large amount of binding energy from an optionally substituted 3,5-dimethoxyphenyl group that projects past the gatekeeper (GK) residue into a hydrophobic pocket. Clinical resistance to currently approved FGFR inhibitors has been reported and results predominantly from GK mutations (FGFR2 V564I/L/M/F, FGFR3 V555I/L/M/F). The GK residue in each mutant is larger than the valine of the wild-type kinase and can no longer accommodate the 3,5-dimethoxyphenyl group of 1\u22123.\n\nConsequently, an FGFR inhibitor of an alternative chemotype that does not rely on this interaction for potency is highly desirable. Based on these data, we sought to develop a second-generation FGFR3 inhibitor with equal potency for the wild-type kinase and gatekeeper mutants, >20-fold selectivity over FGFR1, and a pharmacokinetic profile suitable for oral dosing. Equal potency for FGFR2 was also desirable, though not a formal goal. Therefore, FGFR3 enzymatic and cellular screening was used to drive the SAR discussed herein.\n\n*Figure 1.* FDA-approved pan-FGFR (FGFR1\u22123) inhibitors erdafitinib, pemigatinib, and infigratinib.\n\n## Results and Discussion\n\nOur program was initiated with a high-throughput screening (HTS) campaign, which identified several potential scaffolds. One of these was an imidazo[1,2-b]pyridazine series, from which 4 was discovered (Table 1). Compound 4 had good potency for FGFR3 WT (IC50 = 68 nM) and GK mutant FGFR3 V555L (IC50 = 22 nM), with >10-fold selectivity over FGFR1, providing an excellent starting point for our discovery campaign.\n\nThe N-methyl amide proved to be a key component of the pharmacophore. Analysis of related HTS hits revealed that compounds without this precise substituent were significantly less potent and selective. We briefly examined other modifications to the amide or replacement with isosteres, but there was no improvement over 4 (data not shown). Given the very high structural homology between FGFR3 and FGFR1, molecular modeling was unsuccessful in explaining the selectivity imparted by the amide. One possible explanation is a subtle conformational change in the P-loop between these two isoforms.\n\n*Table 1.* HTS Hit 4 and Methoxy Derivative 5.\n\n| Compound | R1 | FGFR3 IC50 (nM) | FGFR3 V555L IC50 (nM) | FGFR1 IC50 (nM) |\n|---|---|---|---|---|\n| 4 | H | 68 | 22 | 918 |\n| 5 | OMe | 4.6 | \u2014 | 86 |", "# Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly\nAuthors: Deller, M.C.; Johnson, H.A.; Miller, M.D.; Spraggon, G.; Elsliger, M.-A.; Wilson, I.A.; Lesley, S.A.\nSource: PLoS ONE, 2015\n\nPrevious structural studies have all shown that gating ring structures are assembled from a total of eight central RCK domains, with the central channel formed by an N-terminal domain of the RCK unit (Fig 5D\u20135F, N-terminal domains denoted as N1 and N2) [16,18,20\u201325]. However, in contrast, the N1 and N2 domains of TM1088 are not housed on a single chain, but instead are spread across the two homodimers of TM1088A and TM1088B, respectively (N1, N1' and N2, N2', respectively, Fig 5D). This arrangement forms a symmetric repeating RCK unit around the central channel. Comparison of this repeating RCK unit with that of the gating ring assemblies of TrkA and BK reveals that two different patterns of domain organization are used within the RCK unit (red dotted lobes in Fig 5D\u20135F). For TM1088, the N1 domain is on the opposite side to the C1 domain and, as such, forms a domain-swapped RCK (Fig 5D). This is comparable to the more conventional single-subunit TrkA system, in which the analogous N1 and C1 domains are also on opposite sides of the protein (Fig 5E). However, other K+ gating ring complexes, such as BK (and MthK), do not adopt a domain-swapped organization and the N1 and C1 domains lie on the same side of the protein (Fig 5F).\n\nThe organization of the RCK domains results in some interesting differences in symmetry. The TM1088 assembly has D2 point symmetry involving four, perpendicular and intersecting, two-fold rotational axes of symmetry (Fig 5D) and is similar to the D2 point symmetry of TrkA, which involves three, perpendicular and intersecting, two-fold rotational axes of symmetry (Fig 5E). However, the BK channel (and MthK) has C4 symmetry as a result of the lack of a domain-swapped organization (Fig 5F). It is important to note that similar C4 symmetry occurs in the non-domain swapped potassium channel GsuK [25] and, therefore, the model presented in Fig 5 may only pertain to transporters of specific families. However, to determine if any global rules exist that relate symmetry and the RCK domain architecture, further gating ring structures will be required. As a result of the symmetry, the TM1088 assembly has three unique protein:protein interfaces (N1:N1', N2:N2' and N1:N2' highlighted by orange curves in Fig 5D), in common with TrkA, which also has three unique protein:protein interfaces (N1:N1, N1:N2 and N2:N2 highlighted by orange curves in Fig 5E). However, in contrast, the BK channel (and MthK) only has two unique protein:protein interfaces (N1:N2 and N2:N1 highlighted by orange curves in Fig 4F) [21].\n\n### Potential role of TM1088 interfaces in the gating ring assembly\n\nSurprisingly, the atypical two-subunit structure of the TM1088 assembly results in three unique protein:protein interfaces that are all structurally homologous to interfaces known to be central to gating ring assembly in the RCK family of K+ transport proteins (Fig 6A). These interfaces are classically characterized as being either a) an intra-chain flexible \"dimer hinge\" between RCK domains or b) a \"mobile\" or \"fixed\" inter-chain interface between neighboring RCK domains. Interestingly, as a result of its two-subunit nature, the TM1088 assembly has an additional \"dimer hinge\" not observed in other gating ring assemblies (Fig 6A). The N1:N1' interface of the TM1088B dimer forms \"dimer hinge 1\" that is analogous to the more classical dimer hinge between the N- and C-terminal domains as observed in TrkA, BK and MthK assemblies. The N2:N2' interface of the TM1088A dimer forms \"dimer hinge 2\" (Fig 6A). The third interface between adjacent TM1088A and TM1088B dimer pairs (N1:N2') is structurally analogous to the \"mobile\" or \"fixed\" interfaces of other K+ gating ring assemblies (Fig 6A). The \"dimer hinge\" angles of TM1088A and TM1088B are ~120\u00b0 (Fig 6B and 6C, respectively). The \"mobile or fixed\" interface of TM1088, which can be measured as the \"helix-crossover\" angle between TM1088B and TM1088A dimers is measured as ~80\u00b0 (Fig 6D).\n\nIt is important to note that, as a result of the lower D2 symmetry, the TrkA (and MthK) assemblies consist of an alternating pattern of \"fixed\" and \"mobile\" interfaces around the central channel (Fig 6E). In contrast, the BK (and KtrAB) assemblies have higher C4 symmetry and, therefore, possess only a single type of \"mobile\" interface.\n\n*Figure 6.* Potential role of TM1088 interfaces in the gating ring assembly. (A) The two unique dimers of the TM1088 assembly are highlighted (red dashed lobes) showing the three unique protein:protein interfaces (orange and red curves). The \"dimer hinge 1\" between TM1088B dimer partners (red arrow), \"dimer hinge 2\" between TM1088A dimer partners (red arrow) and the \"mobile or fixed\" interface between adjacent dimer pairs (orange curve) are shown. (B) \"Dimer hinge 2\" between the N-terminal Trk domains N2 and N2' of TM1088A. (C) \"Dimer hinge 1\" between N-terminal Trk domains N1 and C1, and N1' and C1' of TM1088B. (D) \"Mobile or fixed\" interface between adjacent TM1088A and TM1088B dimer pairs. TM1088A helices are shown in green (\u03b13: Asp83-Phe97 and \u03b14: Ile112-Asn118) and TM1088B helices in blue (\u03b14: Ile82-Asp76 and \u03b16: Ile107-Met111). Conserved hydrophobic residues are shown as green sticks (TM1088A: Phe88, Met91 and Ile114) and blue sticks (TM1088B: Leu80, Phe81, Met111). The \"helix crossover\" angle between \u03b13 of TM1088A and \u03b14 of TM1088B is denoted by an orange arrow. (E) The single unique dimer of the TrkA assembly (PDB ID: 4J9U) is highlighted showing the three unique protein interfaces (orange and red curves). The \"dimer hinge\" angle between dimer partners for the closed conformation (red dashed line, ADP-bound TrkA, PDB ID: 4J9U) and the open conformation (green dashed line, ATP\u03b3S-bound TrkA, PDB ID: 4J9V) are shown.\n\n## Conclusions", "# 2017 \u2013 Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThis mechanistic understanding has catalyzed a renaissance in HIV vaccine research, particularly in the development of germline-targeting immunogens designed to activate the rare B cell precursors capable of maturing into broadly neutralizing antibodies. Cutting-edge vaccine platforms now incorporate epitope scaffolding techniques that present the MPER in its native membrane-associated context, using nanoparticle delivery systems and mRNA-based approaches to focus immune responses on neutralizing epitopes while avoiding problematic off-target reactions. These strategies represent a fundamental shift from traditional vaccine approaches, moving beyond simple protein immunization toward sophisticated immunological engineering that guides B cell development through carefully orchestrated sequential immunizations.\n\nThe therapeutic implications are equally profound, with 10E8-based antibodies entering clinical development as both treatment and prevention strategies. However, early clinical trials have revealed important challenges, including unexpected local reactogenicity that led to the suspension of some studies involving 10E8VLS, highlighting the critical importance of thorough safety evaluation and potential antibody engineering to minimize adverse effects while preserving neutralization potency. Current research focuses on developing next-generation MPER-targeting antibodies that maintain 10E8\u2019s remarkable breadth and potency while improving safety profiles and pharmacokinetic properties. Additionally, innovative bispecific antibody designs are being explored that combine 10E8\u2019s MPER-targeting capabilities with complementary specificities, potentially offering enhanced neutralization through multi-epitope targeting strategies.\n\nThe convergence of structural insights, advanced vaccine technologies, and clinical experience with 10E8 is reshaping our approach to HIV prevention and treatment, offering realistic hope for developing effective interventions against this persistent global health challenge.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices\nAuthors: Deller, M.C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nThis region maps to the structurally equivalent residues Pro 93 to Gln 112 of the OSM structure, which reside on the B-C loop and N-terminal portion of helix C, as shown in pink in Figure 6-10 (b) below. As discussed previously several residues in this region, clustered around Glu 89, Lys 110 and Tyr 126, exhibit conformational flexibility between the I222 and P212121 crystal structures of OSM, due to their involvement in crystal contacts in one or both of the structures. It has been reported that OSM forms low levels of dimers and trimers in solution (Staunton et al., 1998) and this present study also supports this claim as detailed in Chapter 3. Further support is given by the pH dependency of the dimerisation with 0.2 % of total protein being dimerised at pH 6.0 and 20 % at pH 7.5 (Staunton et al., 1998), this dependency could possibly be due to the charged character of Glu 89 and Lys 110 in the interface. This interface would provide a smaller surface area for burial than that observed in the CNTF dimer (~1038 \u00c5\u00b2) (McDonald et al., 1995) and as such would explain the low levels of dimerisation observed in OSM. However, CNTF does use similar residues in the dimer interface along the face of helix C, so suggesting a common theme may prevail. However, the stoichiometry of the OSM-gp130-LIFR complex is believed to be 1:1:2 (Zhang et al., 1997b) and there is no evidence to suggest any direct interaction between the OSM molecules of the assembly (Staunton et al., 1998). Moreover, both IL-6 (May et al., 1991) and GH (Cunningham et al., 1991a) have also been reported to exist as oligomers and such complexes may be important for storage and transport of potentially toxic levels of cytokine.\n\n#### 6.3.3 Ciliary Neurotrophic Factor\n\nCNTF requires three separate receptor components for signalling; CNTFR, LIFR and gp130 (Davis et al., 1993). In common with LIF and IL-6, as discussed above, three sites on the ligand have been identified as important for receptor binding. CNTF, CNTFR, gp130 and LIFR have been shown to form a hexameric complex in the stoichiometry of 2:2:1:1 (De Serio et al., 1995).\n\nUsing a series of cell survival assays (Inoue et al., 1997; Panayotatos et al., 1994) measuring specific responses to both \u03b1-receptor (CNTFR) and \u03b2-receptor (LIFR and gp130) units the three receptor binding sites of CNTF have been delineated in a similar fashion to LIF and IL-6 as discussed above.\n\n##### 6.3.3.1 Site 1 Interactions\n\nIn agreement with the theme introduced in LIF and IL-6, site 1 functions in binding CNTFR; a specificity conferring receptor element. Interactions are believed to involve residues Arg 25 and Arg 28 which are located on helix A, residues Gln 63 and Trp 64 which are located on the A-B loop and, residue Gln 74 located on helix B and residue Arg 177 located on helix D (Panayotatos et al., 1995). These residues are all solvent exposed as shown in blue in Figure 6-10 (e) below.\n\nStructurally equivalent residues in the OSM structure include Asp 22, Thr 27, Asn 75 and Arg 182 as shown in blue in Figure 6-10 (b) below. It is interesting to note that two of these residues are positively charged and given the previous observation (see Figure 6-5 (f)) that helix D shows an unusually large clustering of around seven positive charges, it is possible that OSM-LIFR interactions may involve an electrostatic component for modulation of binding. A similar clustering is observed in site 1 of CNTF and it has been postulated that electrostatic interactions are involved in CNTFR recruitment (McDonald et al., 1995). LIF does not possess such a prominent clustering of charged residues, so suggesting that the high affinity site 3 interaction prevails.\n\nSuch clustering of electrostatic potential has previously been reported to play a role in cytokine ligand-receptor interactions; for example the use of three lysine residues in the NGF-p75NGFR interaction (Ibanez et al., 1992). Similarly, studies of the electrostatics of GH-GHR contact residues has revealed that mutation of Arg residues most affect association and together contribute a factor of about 20 to the on-rate (Cunningham and Wells, 1993). These data suggests that the hormone and receptor associate by diffusion and electrostatics to form an ensemble of weak collisional complexes, and the structure of OSM suggests a similar mechanism may be utilised.\n\n##### 6.3.3.2 Site 2 Interactions\n\nIn continuation of the themes observed in LIF site 2 residues interact with gp130 and this region has been partially mapped to the solvent exposed residues Lys 26 and Asp 30 (Panayotatos et al., 1995) which both reside on helix A as shown in yellow in Figure 6-10 (e) below.\n\nStructurally equivalent residues in the structure of OSM include Leu 23 and Arg 29 as shown in Figure 6-10 (f) below. Analysis of Table 6-3 below reveals that mutation of Leu 23 has a small effect on both gp130 and LIFR binding although the role of the solvent exposed residue Arg 29 is unknown.\n\n##### 6.3.3.3 Site 3 Interactions\n\nSite 3 in common with LIF and IL-6 consists of a conserved pair of residues, Phe 152 and Lys 155 (Inoue et al., 1997), which have been shown to facilitate binding to LIFR without affecting binding to CNTFR or gp130 (Di Marco et al., 1996). Both of these residues are solvent exposed and lie on the N-terminus of helix D, as shown in red in Figure 6-10 (e) below.\n\nFrom sequence alignments as shown in Figure 6-12 below it is clear that Phe 152 and Lys 155 correspond to the functionally equivalent Phe 160 and Lys 163 pair of OSM and Phe 156 and Lys 159 of LIF. However, these pairs do not correspond structurally between the two molecules due to the lengthened helix D of OSM, which results in a shift of 3 residues between the pairs as shown in red in Figure 6-10 (f) below.", "# The Memorization Trap: Have Protein-Ligand Cofolding Methods Transcended Simple Data Recall?\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-12 on marcdeller.com (categories: blog, sb).*\n\n* [Have protein-ligand cofolding methods moved beyond memorisation?](https://www.biorxiv.org/content/10.1101/2025.02.03.636309v3.full.pdf)\n\n* **Memorization Over Understanding**: Current cofolding methods largely memorize ligand poses from training data, limiting de novo drug design applications\n* **Performance Degradation**: Models show substantial accuracy drops when evaluated on protein-ligand combinations significantly different from training sets\n* **Binding Site Limitations**: Even dramatic binding pocket mutations fail to appropriately alter predicted ligand poses, often resulting in physically implausible structures\u200b\n* **Training Data Bias**: Performance correlates strongly with ligand promiscuity in training data, with cofactors showing better prediction than site-specific molecules\n* **Architectural Advances**: Methods like DynamicBind and NeuralPLexer attempt to address flexibility challenges through specialized diffusion and multi-scale approaches\n* **Generalization Challenge**: The field requires fundamental advances in learning algorithms to move beyond pattern recognition toward true molecular understanding\u200b\n\n1. **Have protein-ligand co-folding methods moved beyond memorisation?:** [\u0160krinjar, P., et al., bioRxiv](https://www.biorxiv.org/content/10.1101/2025.02.03.636309v3.full.pdf)\n2. **Accurate structure prediction of biomolecular interactions with AlphaFold 3:** [Abramson, J., et al., Nature](https://www.nature.com/articles/s41586-024-07487-w)\n3. **Chai-1: Decoding the molecular interactions of life:** [Chai Discovery Team, et al., bioRxiv](https://www.biorxiv.org/content/biorxiv/early/2024/10/11/2024.10.10.615955.full.pdf)\n4. **State-specific protein-ligand complex structure prediction with a multi-scale deep generative model:** [Qiao, Z., et al., Nature Machine Intelligence](https://arxiv.org/abs/2209.15171)\n5. **DynamicBind: predicting ligand-specific protein-ligand complex structure:** [Lu, W., et al., Nature Communications](https://pmc.ncbi.nlm.nih.gov/articles/PMC10844226/)\n6. **Beyond rigid docking: deep learning approaches for fully flexible protein\u2013ligand interactions:** [Review Article, Current Opinion in Structural Biology](https://pmc.ncbi.nlm.nih.gov/articles/PMC12406700/)\n\nThe landscape of computational structural biology has been fundamentally transformed by the emergence of protein-ligand cofolding methods. These sophisticated deep learning approaches promise to revolutionize drug discovery by directly predicting how proteins and small molecules interact at the atomic level. However, a comprehensive analysis using the newly developed Runs N\u2019 Poses benchmark dataset reveals a concerning limitation: current state-of-the-art cofolding methods, including AlphaFold3, Chai-1, and RoseTTAFold All-Atom, appear to be largely memorizing ligand poses from their training data rather than learning the underlying physical principles of molecular interactions.\n\nThis challenge represents more than a technical limitation\u2014it strikes at the heart of whether artificial intelligence can truly understand biological systems or merely excel at pattern recognition within familiar chemical space. The implications extend far beyond academic curiosity, directly impacting the development of novel therapeutics and our ability to target previously \u201cundruggable\u201d proteins. Recent evaluations demonstrate that when these methods encounter protein-ligand combinations significantly different from their training sets, performance drops dramatically, suggesting that the apparent sophistication of these models may mask a fundamental reliance on memorized structural patterns.\n\nThe evolution from rigid docking approaches to flexible cofolding represents a paradigm shift in computational molecular biology. Traditional docking methods treated proteins as static structures, attempting to fit ligands into predetermined binding pockets. Contemporary cofolding approaches, exemplified by methods like AlphaFold3 and Chai-1, simultaneously predict both protein conformation and ligand binding pose from sequence information alone. This capability theoretically enables the modeling of induced-fit effects and conformational changes that occur upon ligand binding\u2014phenomena crucial for understanding allosteric regulation and designing more effective drugs.\n\nAlphaFold3, released in 2024, introduced a substantially updated diffusion-based architecture capable of predicting complex biomolecular interactions with unprecedented accuracy. The model achieves approximately 77% success rates for protein-ligand structure prediction tasks, representing a significant advancement over specialized docking tools. Similarly, Chai-1 demonstrates comparable performance metrics while offering open-source accessibility, achieving 77% ligand RMSD success rates on benchmark datasets. These achievements initially suggested that the field had overcome fundamental limitations in modeling protein flexibility and ligand specificity.\n\nHowever, the development of rigorous evaluation frameworks has revealed critical weaknesses in generalization capability. The Runs N\u2019 Poses dataset, comprising 2,600 high-resolution protein-ligand systems released after the training cutoff dates of major cofolding methods, provides an unbiased assessment of model performance on truly novel structures. When evaluated against this benchmark, current methods show substantial performance degradation, particularly for ligands that have only been observed binding in single pockets during training. More promiscuous ligands, such as metabolic cofactors that bind multiple protein families, demonstrate moderately improved performance, suggesting that training data diversity directly correlates with prediction accuracy.", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\nBtDyP from Bacteroides thetaiotaomicron (strain VPI-5482) and TyrA from Shewanella oneidensis are dye-decolorizing peroxidases (DyPs), members of a new family of heme-dependent peroxidases recently identified in fungi and bacteria. Here, we report the crystal structures of BtDyP and TyrA at 1.6 and 2.7 A, respectively. BtDyP assembles into a hexamer, while TyrA assembles into a dimer; the dimerization interface is conserved between the two proteins. Each monomer exhibits a two-domain, alpha+beta ferredoxin-like fold. A site for heme binding was identified computationally, and modeling of a heme into the proposed active site allowed for identification of residues likely to be functionally important. Structural and sequence comparisons with other DyPs demonstrate a conservation of putative heme-binding residues, including an absolutely conserved histidine. Isothermal titration calorimetry experiments confirm heme binding, but with a stoichiometry of 0.3:1 (heme:protein).\n\n## Identification and structural characterization of heme binding in a novel dye-decolorizing peroxidase, TyrA (2007)\n\n**Authors:** Zubieta C, Joseph R, Krishna SS, McMullan D, Kapoor M, Axelrod HL, Miller MD, Abdubek P, Acosta C, Astakhova T, Carlton D, Chiu HJ, Clayton T, Deller MC, Duan L, Elias Y, Elsliger MA, Feuerhelm J, Grzechnik SK, Hale J, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Kumar A, Marciano D, Morse AT, Murphy KD, Nigoghossian E, Okach L, Oommachen S, Reyes R, Rife CL, Schimmel P, Trout CV, van den Bedem H, Weekes D, White A, Xu Q, Hodgson KO, Wooley J, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Proteins (2007). DOI: 10.1002/prot.21673. PMID: 17654547.\n\nTyrA is a member of the dye-decolorizing peroxidase (DyP) family, a new family of heme-dependent peroxidase recently identified in fungi and bacteria. Here, we report the crystal structure of TyrA in complex with iron protoporphyrin (IX) at 2.3 A. TyrA is a dimer, with each monomer exhibiting a two-domain, alpha/beta ferredoxin-like fold. Both domains contribute to the heme-binding site. Co-crystallization in the presence of an excess of iron protoporphyrin (IX) chloride allowed for the unambiguous location of the active site and the specific residues involved in heme binding. The structure reveals a Fe-His-Asp triad essential for heme positioning, as well as a novel conformation of one of the heme propionate moieties compared to plant peroxidases. Structural comparison to the canonical DyP family member, DyP from Thanatephorus cucumeris (Dec 1), demonstrates conservation of this novel heme conformation, as well as residues important for heme binding. Structural comparisons with representative members from all classes of the plant, bacterial, and fungal peroxidase superfamily demonstrate that TyrA, and by extension the DyP family, adopts a fold different from all other structurally characterized heme peroxidases. We propose that a new superfamily be added to the peroxidase classification scheme to encompass the DyP family of heme peroxidases.\n\n## Cell surface receptors (2000)\n\n**Authors:** Deller MC, Yvonne Jones E.  \n**Journal:** Current opinion in structural biology (2000). DOI: 10.1016/s0959-440x(00)00072-5. PMID: 10753817.\n\nDuring the past year, the database of ligand-receptor complexes has essentially doubled. These new results have immeasurably extended and expanded our view of cell surface receptor structure and function. The flood of data has revealed new models for receptor cross-linking, demonstrating the potential stringency of the extracellular requirements for the initiation of intracellular signalling and highlighting unexpected interactions suggestive of higher order clustering.\n\n## Crystal structure and functional dissection of the cytostatic cytokine oncostatin M (2000)\n\n**Authors:** Deller MC, Hudson KR, Ikemizu S, Bravo J, Jones EY, Heath JK.  \n**Journal:** Structure (London, England : 1993) (2000). DOI: 10.1016/s0969-2126(00)00176-3. PMID: 10997905.\n\nBackgroundThe cytokine oncostatin M (OSM) inhibits growth of certain tumour-derived cells, induces proliferation in other cell types (e.g. haemangioblasts) and is a mediator of inflammatory responses. Its mechanism of action is via specific binding to gp130 and either the leukaemia inhibitory factor receptor (LIFR) or oncostatin M receptor (OSMR) systems at the cell surface to form an active signalling complex.ResultsWe report here the crystal structure of human oncostatin M (hOSM) along with mutagenesis data which map the receptor-binding epitopes of the molecule. The structure was determined to a resolution of 2.2 A and conforms to the haematopoietin cytokine up-up-down-down four-helix bundle topology. The site 2 epitope, responsible for gp130 binding, is centred around Gly120 which forms a 'dimple' on the surface of the molecule located on helices A and C. The site 3 motif, responsible for LIFR and OSMR binding, consists of a protruding Phe160/Lys163 pair located at the start of helix D.ConclusionsThe data presented allow functional dissection of the receptor-binding interfaces to atomic resolution. Modelling suggests that the gp130 residue Phe169 packs into the site 2 dimple in an analogous fashion to structurally equivalent residues at the growth hormone-growth hormone receptor interface, implying that certain key features may underlie recognition across the whole cytokine/receptor superfamily. Conversely, detailed comparison of the available structures suggests that variations on a common theme dictate the specificity of receptor-ligand interactions within the gp130 family of cytokines.\n\n## Structural studies of cytokines and cytokine receptors (1999)\n\n**Authors:** Deller MC.  \n**Journal:**  (1999). \n\n*(no abstract available)*", "# gobsmacked.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n## gobsmacked.mdeller.com/\n\n## 1. Protein\n\nFilled in with a worked example: **EGFR** (P00533, Pfam PF07714) and\n**erlotinib**, against **1M17**. Press Fetch to follow it through, or start\ntyping to replace it.\n\nUniProt accession, PDB ID or sequence\n\nForce a particular PDB entry as the starting structure (optional)\n\nFilled in, this is a redocking control: the run starts from the crystal it\nis judged against, which is the standard positive control and the quickest way to see the\napp work. Clear it to start from the AlphaFold model instead, which is the harder and more\nrealistic test.\n\nOr drop your own coordinates (.pdb, .cif)\n\nTrim to a domain, optional (first-last residue)\n\n## Family, filled in by Fetch\n\n* Pfamnot fetched\n\n## 2. Sequence and annotation\n\n## 3. Ligand\n\nName\n\nSMILES\nCOCCOc1cc2ncnc(Nc3cccc(C#C)c3)c2cc1OCCOC\n\npH\n\n## 4. Reference structure\n\nRanked by how similar each entry's ligand is to yours. A crystal holding the\nsame ligand is preferred over a sharper one holding a different chemotype.\n\nReference PDB entry\n\nSearch below and click an entry, or type one here.\n\nApo reference, optional: the same protein with nothing bound\n\n## 5. Pocket\n\nResidues (chain:number, comma separated)\n\n## Scope\n\nFetch a protein to see it here.\n\nClick residues in the structure or the sequence track to build the pocket. The box is the\nselection's extent plus 8 A, floored at 18 A per side.\n\n## 6. Run settings and bundle\n\nDocking\n\nhybrid (search + GNN rescoring)\nflex (flexible side chains)\ndock (rigid, fastest)\n\nPoses\n\nProduction MD (ps)\n\nFrame every (ps)\n\nAffinity\n\nPredict affinity with Boltz-2\n\nScores the docked pose and the relaxed complex, and\nreports both. Adds a few minutes and a one-off download.\n\nFrames scored after MD\n\nFive frames across the last fifth of the run,\nreported as a mean and a spread.\n\nTitle for the Runs table\n\nVisibility\n\nThis run appears in the Runs tab for anyone.\n\n## gobsmacked.mdeller.com/about\n\n## What GOBSMACKED does\n\nGOBSMACKED folds, docks, relaxes and annotates a protein-ligand complex, then checks the\nresult against the experimental structure. The heavy compute happens on your own GPU, in a\nbundle this server writes; the annotation and the verification happen here.\n\n### Why the superposition is on pocket Ca atoms\n\nA predicted model can be excellent at the binding site and several angstroms out at a\ndisordered terminus. Superposing whole chains spreads that error into the pocket and inflates\nthe ligand RMSD, which is the one number this app exists to report honestly. Everything\ngeometric here is measured after fitting the residues within 8 A of the reference ligand,\nand the whole-chain TM-score is reported separately as context.\n\n### What the grades mean\n\n| Metric | A | B | C | D | F |\n| --- | --- | --- | --- | --- | --- |\n| Ligand RMSD, best of pose 1 and MD-final | \u2264 1.0 A | \u2264 2.0 | \u2264 3.0 | \u2264 4.0 | > 4.0 |\n| Pocket Ca RMSD, MD-final | \u2264 0.8 | \u2264 1.2 | \u2264 1.8 | \u2264 2.5 | > 2.5 |\n| chi1 agreement, MD-final | \u2265 0.85 | \u2265 0.70 | \u2265 0.55 | \u2265 0.40 | < 0.40 |\n| PLIP Jaccard, best of pose 1 and MD-final | \u2265 0.75 | \u2265 0.55 | \u2265 0.40 | \u2265 0.25 | < 0.25 |\n| MD stability: ligand drift, last 200 ps minus first 200 ps | \u2264 0.5 A | \u2264 1.0 | \u2264 1.5 | \u2264 2.5 | > 2.5 |\n| Rescue: pocket Ca RMSD before MD minus after | \u2265 +0.5 A | \u2265 +0.2 | \u2265 0 | \u2265 \u22120.3 | < \u22120.3 |\n| Pose validity (clashes, bond lengths, chirality, box) | pass | | | | fail |\n\nThe composite GOBSMACK score is the weighted mean of those grades (ligand RMSD\n30, PLIP overlap 20, pocket Ca 15, chi1 10, stability 10, rescue 10, validity 5). A metric\nthat could not be measured drops out and the remaining weights are renormalised, which is\nstated on the card. A validity failure caps the composite at 40. A run with no reference\nstructure gets no composite at all: there is nothing to verify it against.\n\n## Software and references\n\n### Prepare\n\n| Tool | What it does here | Licence | Links |\n| --- | --- | --- | --- |\n| UniProt | Canonical sequence, features, protein naming | CC BY 4.0 | [repository](https://www.uniprot.org) \u00b7 [The UniProt Consortium 2023, Nucleic Acids Research](https://doi.org/10.1093/nar/gkac1052) |\n| InterPro / Pfam | Domain assignment, and the family router that follows from it | CC0 | [repository](https://www.ebi.ac.uk/interpro) \u00b7 [Paysan-Lafosse et al. 2023, Nucleic Acids Research](https://doi.org/10.1093/nar/gkac993) |\n| KLIFS | 85-residue kinase pocket numbering, DFG and alphaC states | open, attribution | [repository](https://klifs.net) \u00b7 [Kanev et al. 2021, Nucleic Acids Research](https://doi.org/10.1093/nar/gkaa895) |\n| GPCRdb | Ballesteros-Weinstein generic numbering, segments, microswitches | Apache-2.0 | [repository](https://github.com/protwis) \u00b7 [Pandy-Szekeres et al. 2023, Nucleic Acids Research](https://doi.org/10.1093/nar/gkac1013) |\n| RCSB PDB | Reference structure search, entry metadata, coordinates | CC0 | [repository](https://www.rcsb.org) \u00b7 [Burley et al. 2023, Nucleic Acids Research](https://doi.org/10.1093/nar/gkac1077) |\n| AlphaFold DB | Predicted starting structures and their pLDDT | CC BY 4.0 | [repository](https://alphafold.ebi.ac.uk) \u00b7 [Varadi et al. 2024, Nucleic Acids Research](https://doi.org/10.1093/nar/gkad1011) |\n| ESM Metagenomic Atlas | Folding short sequences without a GPU | MIT | [repository](https://esmatlas.com) \u00b7 [Lin et al. 2023, Science](https://doi.org/10.1126/science.ade2574) |\n| RDKit | SMILES validation, depiction, fingerprints, symmetry-aware RMSD | BSD-3-Clause | [repository](https://github.com/rdkit/rdkit) \u00b7 rdkit.org |\n\n### Run", "# SiteHealthInspector \u2013 A pure-Bash, colour-coded website health audit\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-14 on marcdeller.com (categories: vibe-code).*\n\n> **A pure-Bash, colour-coded website health audit. No install, no dependencies.**\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/SiteHealthInspector](https://github.com/bellcheddar/SiteHealthInspector) |\n\n---\n\nWhy it matters: when a site feels slow or you need to confirm TLS, caching, and security headers are sound, you want an answer in seconds, not a heavyweight tool to set up. SiteHealthInspector.sh is a single Bash script that probes timing, HTTP status, caching and CDN behaviour, security headers, the TLS certificate, DNS records, CMS fingerprint, and an overall weighted health score, all rendered with star ratings and bar charts for instant triage. It is useful for developers and site owners who live in the terminal and want a fast, dependency-free pulse check on any URL: perfect for spot checks, CI smoke tests, and before/after comparisons when tuning performance.\n`SiteHealthInspector.sh` runs a comprehensive, colour-coded diagnostic scan against any URL \u2014 reporting timing metrics, HTTP status, caching behaviour, security headers, SSL/TLS certificate status, DNS records, CMS fingerprinting, and an overall health scorecard. No runtime dependencies beyond standard Unix tools.\n\n---\n\n## \u2728 Features\n\n| Module | What it checks |\n| --- | --- |\n| \u23f1 **Timing Metrics** | DNS resolution, TCP connect, TLS handshake, TTFB, total load time, redirect overhead, download speed |\n| \ud83d\udcf6 **HTTP Status & Redirect Chain** | Final HTTP status code with full hop-by-hop redirect trace |\n| \ud83d\udce6 **Cache & CDN Headers** | `Cache-Control`, `ETag`, `Last-Modified`, `Age`, Cloudflare cache status (HIT/MISS/BYPASS/DYNAMIC), `Vary`, content encoding |\n| \ud83d\udd10 **Security Headers** | HSTS, CSP, `X-Content-Type-Options`, `X-Frame-Options`, XSS Protection, Referrer Policy, Permissions Policy, COEP, COOP |\n| \ud83d\udd11 **SSL/TLS Certificate** | Subject, issuer, validity window, days-to-expiry with renewal warnings, TLS protocol version, cipher suite, certificate verification |\n| \ud83d\udca8 **Compression & Content** | gzip/brotli/zstd detection, HTTP/1.1 vs HTTP/2 vs HTTP/3 negotiation, page size, content-type |\n| \ud83c\udf0d **DNS & Connectivity** | A, AAAA, MX, NS records; SPF and DMARC TXT records; 5-run TTFB stability test with jitter calculation |\n| \u2699 **CMS / Platform Detection** | Generator meta tag, WordPress login/REST API exposure, sitemap.xml, robots.txt |\n| \ud83d\udc93 **Availability & Uptime** | Live HTTP check, `www` subdomain, HTTP\u2192HTTPS redirect enforcement, 404 and `/wp-admin/` probing |\n| \ud83c\udfc6 **Overall Health Scorecard** | Weighted 0\u2013100 score across all modules with grade: EXCELLENT / GOOD / FAIR / POOR |\n\nAll output uses ANSI colour coding with star ratings (\u2605\u2605\u2605\u2605\u2605) and bar charts for instant visual triage.\n\n---\n\n## \ud83d\udccb Requirements\n\n| Dependency | Used for |\n| --- | --- |\n| `bash` \u2265 3.2 | Script runtime (macOS/Linux compatible) |\n| `curl` | HTTP requests, timing metrics, SSL verification |\n| `openssl` | TLS certificate introspection |\n| `dig` | DNS record lookups |\n| `bc` | Floating-point arithmetic for timing calculations |\n\nAll of the above ship with macOS and most Linux distributions. On minimal containers (e.g. Alpine), install with:\n\n```\napk add curl openssl bind-tools bc\n```\n\n---\n\n## \ud83d\ude80 Usage\n\n```\n# Scan the default target (marcdeller.com)\nbash SiteHealthInspector.sh\n\n# Scan a specific URL\nbash SiteHealthInspector.sh https://example.com\n\n# Make executable and run directly\nchmod +x SiteHealthInspector.sh\n./SiteHealthInspector.sh https://your-site.com\n```\n\n> **Note:** The script follows redirects (`-L`) and accepts a full URL including scheme. Both `https://` and `http://` targets are supported.\n\n---\n\n## \ud83d\udcca Sample Output\n\n```\n  ============================================================\n  \ud83c\udf10  SITE HEALTH INSPECTOR  v1.2\n  Target: https://example.com\n  2025-06-11  14:32:07  BST\n  ============================================================\n\n  ============================================================\n  \u23f1  TIMING METRICS\n  ============================================================\n  \ud83d\udd0d DNS Resolution: 12ms  \u2605\u2605\u2605\u2605\u2605\n  \ud83d\udd0c TCP Connect: 45ms  \u2605\u2605\u2605\u2605\u2605\n  \ud83d\udd12 TLS Handshake: 98ms  \u2605\u2605\u2605\u2605\u2605\n  \u26a1 TTFB (Time to First Byte): 187ms  \u2605\u2605\u2605\u2605\u2605\n  \ud83c\udfc1 Total Load Time: 423ms  \u2605\u2605\u2605\u2605\u2605\n\n  ...\n\n  \ud83c\udfc6 Overall Score: \ud83c\udfc6 EXCELLENT  87/100\n```\n\n---\n\n## \ud83c\udfaf Scoring\n\n### Timing thresholds (lower is better)\n\n| Metric | \u2605\u2605\u2605\u2605\u2605 Good | \u2605\u2605\u2605\u2606\u2606 Fair | \u2605\u2606\u2606\u2606\u2606 Poor |\n| --- | --- | --- | --- |\n| DNS Resolution | \u2264 50 ms | \u2264 200 ms | > 200 ms |\n| TCP Connect | \u2264 100 ms | \u2264 400 ms | > 400 ms |\n| TLS Handshake | \u2264 150 ms | \u2264 400 ms | > 400 ms |\n| TTFB | \u2264 200 ms | \u2264 600 ms | > 600 ms |\n| Total Load | \u2264 500 ms | \u2264 2000 ms | > 2000 ms |\n\n### Overall health score\n\n| Score | Grade |\n| --- | --- |\n| 80\u2013100 | \ud83c\udfc6 EXCELLENT |\n| 60\u201379 | \u26a1 GOOD |\n| 40\u201359 | \u26a0 FAIR |\n| 0\u201339 | \ud83d\udea8 POOR |\n\nKey contributors to the overall score include: TTFB, total load time, compression, HTTP/2+, valid SSL, certificate validity, HSTS presence, cache headers, Cloudflare HIT status, and live availability.\n\n---\n\n## \ud83d\udd12 Security Header Coverage\n\nThe script audits all headers recommended by [OWASP Secure Headers Project](https://owasp.org/www-project-secure-headers/):\n\n* `Strict-Transport-Security` (HSTS)\n* `Content-Security-Policy` (CSP)\n* `X-Content-Type-Options`\n* `X-Frame-Options`\n* `X-XSS-Protection`\n* `Referrer-Policy`\n* `Permissions-Policy`\n* `Cross-Origin-Embedder-Policy` (COEP)\n* `Cross-Origin-Opener-Policy` (COOP)\n\n---\n\n## \ud83d\uddc2 File Structure\n\n```\nSiteHealthInspector.sh    # Single-file script \u2014 no config, no dependencies to install\n```\n\n---\n\n## \ud83d\udee0 Customisation\n\nThe default scan target is `https://marcdeller.com`. Change it by passing a URL argument or by editing the fallback in line 4:\n\n```\nTARGET=\"${1:-https://your-default-site.com}\"\n```\n\n---\n\n## \ud83d\udcc4 Licence\n\nMIT \u2014 use freely, attribution appreciated.\n\n---\n\n*Built as part of the [marcdeller.com](https://marcdeller.com) toolchain.*\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/SiteHealthInspector](https://github.com/bellcheddar/SiteHealthInspector) |", "# Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design\nAuthors: Irimia, A.; Serra, A.M.; Sarkar, A.; Jacak, R.; Kalyuzhniy, O.; Sok, D.; Saye-Francisco, K.L.; Schiffner, T.; Tingle, R.; Kubitz, M.; Adachi, Y.; Stanfield, R.L.; Deller, M.C.; Burton, D.R.; Schief, W.R.; Wilson, I.A.\nSource: PLoS Pathogens, 2017\n\n*S2 Fig. Electron density maps of ligands bound in the lipid-binding site.* (A) Electron density (3\u03c3 level; green wires) of a fragment of 06:0 PG (sticks) in the initial difference Fourier map calculated after molecular replacement. (B) 2Fo-Fc density (1\u03c3 level; blue wires) for the 06:0 PG fragment (sticks) after the last round of refinement. (C) Electron density (3\u03c3 level; green wires) of a fragment of 06:0 PA (sticks) in the initial difference Fourier map calculated after molecular replacement. (D) 2Fo-Fc density (1\u03c3 level; blue wires) for the 06:0 PA fragment (sticks) after the last round of refinement. (E) Initial difference Fourier positive electron density (3\u03c3 level; green wires) of a glycerol molecule (sticks) bound in the lipid-binding site in the 10E8-T117v2 structure. (F) 2Fo-Fc density (1\u03c3 level; blue wires) calculated at final stage of refinement for the glycerol molecule (sticks) bound in the lipid-binding site in the 10E8-T117v2 structure.\n\n*S3 Fig. SPR sensorgrams of 10E8 epitope-scaffolds.* T117v2 binding to 10E8 IgG light-chain mutants.\n\n*S4 Fig. Structure comparison of 10E8 mutant 5-T117v2 complex to 10E8 wild type-T117v2.* (A) Snapshot of the superposition of the light-chain regions of 10E8 in mutant 5 (gray) and wild type (beige) showing regions that deviate slightly from each other. The positions of C\u03b1 atoms of each residue are shown as small spheres for comparison. Despite small variations in position of some light-chain residues, the location and conformation of CDRH3 in the two structures (violet, wild type; green, mutant 5) is nearly identical. The region located in the red dashed rectangle is seen in a close-up in (B) for wild type and (C) for mutant 5. The side chains of the residues are shown as sticks, with residues Asn48(L)-Phe53(L) adopting multiple conformations. The underlined superscript letters designate the original residues in the 10E8 wild type.\n\n*S5 Fig. Description of the ligands bound in the lipid-binding site in mutant 1 and 3 structures.* (A) Initial Fo-Fc electron density map (3\u03c3 level) observed for a glycerol molecule bound in the lipid-binding site in 10E8 mutant 1-T117v2 complex. (B) 2Fo-Fc map (1\u03c3 level) for the glycerol in (A) after refinement. (C) All potential glycerol (dark cyan sticks) hydrogen-bond interactions (within ~3.5 \u00c5) with the residues of the lipid-binding site (beige, light chain; violet for heavy chain) in mutant 1 are shown as sticks. (D) Initial Fo-Fc electron density map (3\u03c3 level) observed for a phosphate from crystallization conditions bound in the lipid-binding site in 10E8 mutant 3-T117v2 complex. (E) 2Fo-Fc map (1\u03c3 level) for the phosphate at (D) after refinement. (F) All potential phosphate (red-orange sticks) hydrogen-bond interactions (within ~3.5 \u00c5) with the residues of the lipid-binding site (beige, light chain; violet, heavy chain) in mutant 3 are shown as sticks. The underlined superscript letters designate the original residues in the 10E8 wild type.\n\n*S6 Fig. SECMALS analysis of wild type 10E8 and light chain mutants.* Normalized UV (red) and light scattering (blue) signals of the indicated antibodies were plotted against the elution volume and protein molar masses of the peaks are indicated in black. Wild-type 10E8 and mutant 1 each show three main peaks in the UV traces, all of which have molecular weights of approximately 150 kDa, indicating that all these peaks represent monomeric IgGs. In contrast, only a single peak can be detected for mutants 2, 3 and 5. Although scattering signals indicate the presence of high-molecular weight aggregates in the wild type 10E8 and mutant 1 sample, the UV analysis shows that only insignificant amounts (<0.5%) of these aggregates are present in either preparation.\n\n*S1 Table. X-ray data collection, structure determination and refinement statistics for complexes of 10E8 Fab with T117v2 scaffold alone or co-crystallized with 06:0 PA and 06:0 PG.*\n\n*S2 Table. X-ray data collection, structure determination and refinement statistics for complexes of 10E8 Fab mutants with T117v2.*\n\n## Acknowledgments\n\nWe thank Henry Tien for technical support with the crystallization robot. This is publication 29421 from The Scripps Research Institute.\n\n## Author Contributions\n\nConceptualization: AI, AMS, WRS, IAW. Formal analysis: AI, AS, RJ. Funding acquisition: DRB, WRS, IAW. Investigation: AI, AMS, OK, DS, KLSF, TS, RT, MK, YA, MCD. Resources: DRB, WRS, IAW. Supervision: RLS, DRB, WRS, IAW. Writing \u2013 original draft: AI. Writing \u2013 review & editing: AI, AS, RLS, WRS, IAW.\n\n## References\n\n*(reference list omitted)*", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nA solution of 2-(2-(2-aminophenyl)-4-cyano-3-ethyl-1-oxobenzo[4,5]imidazo[1,2-a]pyridin-5(1H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide (Step 5, Example 24, 6.0 mg, 0.011 mmol) in DCM (0.23 mL) at 0 \u00b0C was treated with triethylamine (3.2 \u03bcL, 0.023 mmol) and acryloyl chloride (1.2 mg, 0.014 mmol) and the reaction was stirred for 20 min. The reaction mixture was diluted with CH3CN, filtered, and purified via preparative HPLC-MS (pH=2) to afford the title compound (2.5 mg, 30%) as a TFA salt. LCMS for C32H25F3N5O3 (M+H)+: calculated m/z=584.2; found 584.2.\n\n### Example 26. N-(2-(3-Ethyl-1-oxo-5-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-1,5-dihydrobenzo[4,5]imidazo[1,2-a]pyridin-2-yl)phenyl)acrylamide\n\n**Step 1. 3-Ethylbenzo[4,5]imidazo[1,2-a]pyridin-1(5H)-one.** To a mixture of 3-ethyl-1-oxo-1,5-dihydrobenzo[4,5]imidazo[1,2-a]pyridine-4-carbonitrile (Step 2, Example 24, 22 mg, 0.093 mmol) in water (0.74 mL) and 1,4-dioxane (0.19 mL) was added c-H2SO4 (0.25 mL, 4.6 mmol) dropwise. The mixture was stirred at rt for 5 min, then was sealed and heated to 120 \u00b0C for 6 days. Upon cooling to room temperature, the reaction mixture was added to ice-water (10 mL). Solid sodium bicarbonate was added cautiously while stirring to adjust to pH >7. The mixture was extracted with EtOAc. The organic layer was dried over Na2SO4, filtered and concentrated in vacuo to afford the title compound which was used without further purification (19 mg, 95%). LCMS for C13H13N2O (M+H)+: calculated m/z=213.1; found 213.1.\n\n**Step 2. 2-(3-Ethyl-1-oxobenzo[4,5]imidazo[1,2-a]pyridin-5(H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide.** To a slurry of 3-ethylbenzo[4,5]imidazo[1,2-a]pyridin-1(5H)-one (prepared by the method of Step 1, 0.050 g, 0.24 mmol), K2CO3 (0.13 g, 0.94 mmol), and tetrabutylammonium bromide (7.6 mg, 0.024 mmol) in CH3CN (2.4 mL) was added 2-bromo-N-(4-(trifluoromethyl)phenyl)acetamide (prepared via the general method given for Intermediate A in Example 1, Step 7, substituting 4-(trifluoromethyl)aniline in place of 2-bromo-N-(4-fluorophenyl)acetamide, 0.13 g, 0.47 mmol) and the reaction was stirred for 30 min at rt, then at 50 \u00b0C for 1.5 h. Upon cooling to room temperature, the reaction mixture was diluted with EtOAc, and the organic mixture was washed with water, followed by brine, dried over Na2SO4, filtered and concentrated in vacuo. The product was purified via flash column chromatograpy, eluting with a gradient of 0-100% EtOAc in hexanes to afford the title compound (23 mg, 24%). LCMS for C22H19F3N3O2 (M+H)+: calculated m/z=414.1; found 414.1.\n\n**Step 3. 2-(2-Bromo-3-ethyl-1-oxobenzo[4,5]imidazo[1,2-a]pyridin-5(H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide.** To a solution of 2-(3-ethyl-1-oxobenzo[4,5]imidazo[1,2-a]pyridin-5(H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide (from Step 2, 23 mg, 0.056 mmol) in DMF (0.28 mL) at 0 \u00b0C was added N-bromosuccininmide (9.9 mg, 0.056 mmol) in DMF (0.28 mL) dropwise. After 10 min, saturated NaHCO3 solution was added and the mixture was extracted with EtOAc. The organic layer was washed with brine (2\u00d7), dried over Na2SO4, filtered and concentrated in vacuo. The product was purified via flash column chromatography, eluting with a gradient of 0-70% EtOAc in hexanes to afford the title compound (8.4 mg, 31%). LCMS for C22H18BrF3N3O2 (M+H)+: calculated monoisotopic m/z=492.1; found 492.0.\n\n**Step 4. 2-(2-(2-Aminophenyl)-3-ethyl-1-oxobenzo[4,5]imidazo[1,2-a]pyridin-5(H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide.** The procedure of Step 5, Example 24 was followed, using 2-(2-bromo-3-ethyl-1-oxobenzo[4,5]imidazo[1,2-a]pyridin-5(1H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide (8.4 mg, 0.017 mmol) in place of 2-(2-bromo-4-cyano-3-ethyl-1-oxobenzo[4,5]imidazo[1,2-a]pyridin-5(1H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide to afford the title compound (6.0 mg, 70%). LCMS for C28H24F3N4O2 (M+H)+: calculated m/z=505.2; found 505.3.\n\n**Step 5. N-(2-(3-Ethyl-1-oxo-5-(2-oxo-2-((4-(trifluoromethyl)phenyl)amino)ethyl)-1,5-dihydrobenzo[4,5]imidazo[1,2-a]pyridin-2-yl)phenyl)acrylamide.** The procedure of Example 25 was followed, using 2-(2-(2-aminophenyl)-3-ethyl-1-oxobenzo[4,5]imidazo[1,2-a]pyridin-5(1H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide (from Step 4, 6.0 mg, 0.012 mmol) in place of 2-(2-(2-aminophenyl)-4-cyano-3-ethyl-1-oxobenzo[4,5]imidazo[1,2-a]pyridin-5(1H)-yl)-N-(4-(trifluoromethyl)phenyl)acetamide to afford the title compound (2.0 mg, 22%) as a TFA salt. LCMS for C31H26F3N4O3 (M+H)+: calculated m/z=559.2; found 559.2. 1H NMR (600 MHz, DMSO-d6) \u03b4 10.94 (s, 1H), 8.90 (s, 1H), 8.70 (d, J=7.9 Hz, 1H), 7.84 (d, J=8.2 Hz, 1H), 7.73 (d, J=8.6 Hz, 2H), 7.62 (d, J=8.1 Hz, 1H), 7.53-7.48 (m, 1H), 7.39-7.30 (m, 2H), 7.23-7.16 (m, 2H), 6.45 (s, 1H), 6.41 (dd, J=17.0, 10.2 Hz, 1H), 6.15 (dd, J=17.0, 2.0 Hz, 1H), 5.60 (dd, J=10.2, 2.1 Hz, 1H), 5.28 (s, 2H), 2.35-2.21 (m, 2H), 0.98 (t, J=7.5 Hz, 3H).\n\n### Example A. WRN ATPase Activity Assay\n\nCompounds of the Examples were tested for WRN ATPase activity. WRN ATPase activity catalyzes ATP hydrolysis to produce ADP and phosphate. DNA encoding the target protein, 6\u00d7HIS-TEV-WRN-517-945, was produced, cloned and transfected as detailed in Example D. The amount of ADP produced was measured using ADP-Glo\u2122 Kinase Assay kit from Promega Corporation in a Greiner white regular volume 384 wells assay plate (Greiner #781075). Data was normalized to a standard curve generated using various ADP concentrations and the corresponding luminescent signal intensities at each concentration. The assay conditions were optimized for enzyme concentration (WRN), substrate concentrations (ATP & dsDNA), and the linearity of the ATPase reaction, in the following assay buffer: 50 mM Tris-HCl, pH-7.5, 2 mM MgCl2, 50 mM NaCl, 1 mM EGTA, 0.01% P-20, 0.05% BSA, and 2 mM DTT prepared with nuclease free water.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 4: Structure Description, Comparison and Functional Implications (Chapter 6) and Appendices\nAuthors: Deller, M.C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nIn summary, features unique to OSM include the helical B-C loop region which forms an extended 'ear' protruding from the main body of the helical bundle and two sections of helix in the A-B loop. The severity of the kink in helix A is also unique to OSM along with the use of more tightly coiled 3-10 helix. Together these subtle but none the less unique features may function in modulation of receptor-ligand interactions in the gp130-LIFR system.\n\n*Figure 6-6.* Figure showing pairwise structural superpositions of (a) human LIF (yellow), (b) human IL-6 (blue; Somers et al., 1997), (c) human CNTF (green; McDonald et al., 1995), (d) human GCSF (purple; Aritomi et al., 1999; Hill et al., 1993) and human Epo (red; Livnah et al., 1996) with OSM (grey). GCSF coordinates kindly provided by K. Morikawa, Biomedical Engineering Research Council, Osaka, Japan. Structural superpositions were produced using SHP (Stuart et al., 1979) and figures were produced using BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994).\n\n#### 6.2.2 Haemopoietin Receptor Family\n\nThe structural similarity of the OSM-gp130 system to that of the Epo-EpoR and GCSF-GCSFR system is further reflected on analysis of the structural similarities of gp130, EpoR and GCSFR as shown below in Table 6-2. This suggests that gp130 shows greatest structural similarity to GCSFR with an RMSD value of 2.52 \u00c5, and a sequence identity of 29 % over 189 structurally equivalent residues. Substantial similarity has also been reported between gp130 and EpoR (Bravo et al., 1998) with an RMSD value of 2.52 \u00c5 over 177 equivalencies.\n\n| Family | Protein | Total number of residues | Number of structurally equivalent residues | Number of sequence identities within equivalents | RMSD (\u00c5) |\n|---|---|---|---|---|---|\n| Haemopoietin receptor | Human GCSFR | 213 | 189 | 54 (29 %) | 2.52 |\n| Haemopoietin receptor | Human EpoR | 217 | 177 | 36 (20 %) | 2.52 |\n| Haemopoietin receptor | Human PrlR | 197 | 171 | 33 (19 %) | 3.03 |\n| Haemopoietin receptor | Human GHR | 192 | 164 | 31 (19 %) | 3.07 |\n\n*Table 6-2.* Table showing structural comparison ranked according to RMSD values of C\u03b1 superpositions for human gp130 with various other receptors of the haemopoietin family. GCSF-GCSFR coordinates (Aritomi et al., 1999) kindly provided by K. Morikawa, Biomedical Engineering Research Council, Osaka, Japan. All other coordinates are from the PDB. Structural superpositions and RMSD calculated using the program SHP (Stuart et al., 1979).\n\nFigure 6-7 below shows the structural superposition of human gp130 onto human GCSFR, human EpoR, human GHR and human PrlR. The overall two domain alignment with EpoR has an RMSD value of 2.52 \u00c5, with 177 structurally equivalent residues. These relatively poor superposition values are mainly due to differences in the tilt angle, which is defined as the angle between the long axes, running approximately parallel to the \u03b2-strands of the two domains. EpoR has a tilt angle of ~90\u00b0 in common with other class 1 haemopoietin receptors, whereas gp130 has a tilt angle of ~70\u00b0. This ~20\u00b0 difference results in poor individual domain superposition, but allows analysis of the E-F and B-C loops which are on the N- and C-terminal domains, respectively. These regions are thought to contain the key binding determinants of gp130. Indeed, Phe 93 and Phe 205 of EpoR which are major binding determinants (Middleton et al., 1999) are structurally equivalent to Phe 169 and Val 230 of gp130, respectively. Mutagenesis has revealed the importance of Phe 169, Val 230 of the B-C loop and Gly 286 and Lys 285 of the F-G loop (Horsten et al., 1997) as discussed in section 6.3.4.\n\nThe two fibronectin type III-like cytokine binding domains have the same basic structural scaffold in both gp130 and EpoR consisting of a \u03b2-sandwich formed from a three-strand (A, B, E) and four-strand (C, C', F, G) \u03b2-sheet. The C-terminal domains show good agreement between the two structures with an RMSD value of 1.0 \u00c5 for 81 structurally equivalent residues. This good agreement is mainly attributed to the short length of the \u03b2-strands (Bravo et al., 1998). However, the N-terminal domains show a poorer agreement within the core framework with an RMSD value of 1.0 \u00c5, for only 53 structurally equivalent residues. This discrepancy is mainly attributed to a shift in the angle of \u03b2-strands C, F and G of the C, C', F, G \u03b2-sheet (Bravo et al., 1998).\n\nFigure 6-7 below shows the structural superposition of human gp130 onto human GCSFR (Aritomi et al., 1999). The overall two domain alignment has an RMSD value of 2.52 \u00c5, with 189 structurally equivalent residues and a sequence identity of 29 %. Again these relatively poor superposition values are mainly due to ~25\u00b0 differences in the tilt angle between the N- and C-terminal domains. In gp130, but not GCSFR or EpoR, the interdomain angle is tightly restrained by main-chain hydrogen bonds between the WSXWS motif and the interdomain linker strand (Bravo et al., 1998), so suggesting that the WSXWS motif may play a role in modulating the interdomain orientation of haemopoietin receptors. Mutational analysis has revealed that in contrast to the even split between the N- and C-terminal domains the major binding determinants of GCSFR reside on the B-C and F-G loops and involve residues Ile 239 and Arg 288, respectively (Layton et al., 1997). This use of residues in the C-terminal B-C and F-G loops at the expense of residues in the N-terminal E-F loop is similar to that observed in gp130, as discussed above. Structurally equivalent residues to Ile 239 and Arg 288 of GCSFR are Lys 281 and Ile 231 of gp130, respectively, and these residues may indeed prove to be important epitopes for ligand binding.", "# When Perception Deceives Reality: The Hidden Flaws in AI Reasoning Evaluation\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nAlternative evaluation approaches restore model performance dramatically. When asked to generate compact algorithmic solutions rather than exhaustive move lists, models successfully solved 15-disk Tower of Hanoi cases that were supposedly beyond their capabilities at 8 disks. This finding suggests that apparent reasoning failures reflect format constraints rather than fundamental limitations in algorithmic understanding. The distinction between mechanical execution and problem-solving difficulty becomes crucial: Tower of Hanoi requires exponentially many moves but involves trivial per-move decisions, while other planning problems demand complex optimization despite shorter solutions.\n\nThe research methodology challenges expose broader issues in AI evaluation practices. Current approaches often conflate solution length with problem complexity, failing to account for computational demands beyond mere enumeration. Evaluation frameworks must distinguish between reasoning capability and output constraints, verify puzzle solvability before testing, and consider multiple solution representations to separate algorithmic understanding from execution limitations.\n\nThese findings illuminate the fundamental question of what constitutes reasoning versus sophisticated pattern matching. While some researchers maintain that language models exhibit \u201cno evidence of formal reasoning\u201d and that their behavior \u201cis better explained by sophisticated pattern matching\u201d, others argue that the evidence supports more nuanced conclusions. The models demonstrate capabilities for solving highly complex tasks requiring long, precise action sequences that challenge even human solvers, yet their performance varies significantly by task type.\n\nThe debate ultimately reflects deeper philosophical questions about the nature of reasoning itself. As evaluation methodologies improve, the evidence suggests that LRMs operate as \u201cstochastic, RL-tuned searchers\u201d within discrete state spaces whose structure remains poorly understood. Rather than dismissing these systems as mere pattern matchers or accepting claims of human-like reasoning, the scientific community must focus on mapping the terrain of their actual capabilities through careful experimentation.\n\nThe lesson extends beyond AI research to experimental design principles across scientific disciplines. The importance of rigorous methodology becomes paramount when evaluating complex systems, as poorly designed experiments can mask genuine capabilities or create false limitations. Future work must prioritize evaluation frameworks that capture dynamic reasoning processes rather than static outputs, account for practical constraints in system design, and distinguish between algorithmic understanding and mechanical execution.", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n| File | What it is |\n|---|---|\n| `boltz_input.md` | Your campaign spec, tidied to the house style. Editable. |\n| `sse_comparison/`, `vendor/` | The compare-sse package `analyze` imports, and the Plotly and 3Dmol builds the offline dashboard embeds. `vendor/` is most of the bundle's size, included so the run needs nothing from the network beyond the model weights. |\n| `config.json` | The run settings, machine-readable. Provenance; not read at run time. |\n| `run_campaign.sh` | Installs the environment, runs the campaign, packs the results. |\n| `pack_results.py` | Writes the `.bmz`. |\n| `BoltzMaker.py` | The pipeline itself. |\n| `pixi.toml`, `pixi.lock` | The pinned environment, locked for macOS (Apple Silicon) and Linux (x86-64). |\n\nMove the downloaded file to the machine with the GPU, put it wherever you want the campaign to\nlive, and either double-click it in Finder or run it from a terminal in that folder:\n\n```sh\nsh ./boltzmaker_<campaign>.command\n```\n\n`sh` is fine on any platform: the file re-execs itself under bash on its first line, so it behaves\nthe same whether your `/bin/sh` is bash (macOS) or dash (most Linux distributions).\n`bash ./boltzmaker_<campaign>.command` and `./boltzmaker_<campaign>.command` (once executable)\nwork too.\n\nIt unpacks into a folder beside itself and starts immediately: installs [pixi](https://pixi.sh)\nif you do not have it, solves the environment, then runs `generate` -> `preflight` -> `run` ->\n`analyze`, writing one `<campaign>.bmz` in that same folder. Bring that back for Step 2.\n**Boltz-2's model weights are not in the bundle** -- large and versioned by Boltz itself, they\ndownload on first use and cache in your home directory, so a second campaign skips that step.\n\nEach long step -- solving the environment, installing PLIP, warming the Boltz CLI -- shows a\nspinner with its elapsed time and reports how long it took, so a terminal that has not moved in\nfour minutes is never ambiguous; a failing step prints the last twenty lines of its log rather\nthan leaving it in a file you have to find. Re-running is safe at any point: `run` is idempotent\nso an interrupted campaign resumes, and the script refuses to overwrite an existing unpacked\nfolder. A campaign that fails part-way still packs what completed and records the shortfall.\n\nThe bundle runs the whole pipeline including `analyze`, not just `run`. Once a machine has the\npinned environment analysis costs seconds more there, and doing it locally is what lets the\ndroplet skip a ~1.5GB PLIP environment and a 900-second request for work your machine has already\ndone. What comes back is small and already structured -- 4.2MB on a real 15-target campaign,\nagainst a 19.4MB dashboard and 50.6MB of PyMOL sessions that deliberately stay on your own disk\n-- so Analysis is a reader, not a compute step.\n\n### Fully Automated Mode, Step 2: Analysis\n\nUpload the `.bmz` and the campaign opens as an interactive report. Nothing is recomputed, so it is\nquick regardless of campaign size, and it opens on the campaign summary with the first target\nalready selected -- on a one-target campaign there was never a click to make.\n\nA **header** names the campaign and counts how many predictions it produced, how many are\nflagged and how many carry a structure, warning explicitly if targets are missing from the\nsummary because the campaign did not fully complete. Below it, a sortable, filterable\n**targets** table -- target, family, ligand, confidence, pIC50, interaction count and flags,\nfiltered by free text, by family, or to flagged targets only -- opens any row you click as the\n**target detail**: four panes for that target, then two campaign-level panes and a sequence\ntrack.\n\nThe Target detail heading is itself the target picker, so the panel both names the open target and\nchanges it, and a pulldown at the top of the page jumps to any panel. **Download HTML package**\ngives the whole explorer as a directory of files -- one page, Mol*, Plotly and the campaign's own\ndata -- to drop into any web root; a **Keep private** campaign gets that download beside a\n**Destroy all data** button, removing the session and any archived bundle and results from the\nserver.\n\n**Four equal panes:**\n\n| Pane | Content |\n|---|---|\n| Overall structure | A [Mol*](https://molstar.org) viewer coloured by chain, ligand in red |\n| Binding site | A second Mol* viewer framed on the ligand and its contacting residues, turning on load |\n| Interactions | Every PLIP contact grouped by type: residue name, number and chain, the distance, and the geometry belonging to that interaction -- donor and acceptor atom types and the donor angle for a hydrogen bond, the T/P classification and ring offset for a stack, the charge sense and ligand group for a salt bridge |\n| Metrics | pTM, ipTM, complex pLDDT, predicted affinity, pIC50 with its ensemble spread, and the Ca RMSD to the apo reference (the residue-weighted mean across the motifs compare-sse could align, with that motif count beside it), with the ligand depiction filling the space below |\n\nBoth viewers carry cartoon/surface/spin/reset and an **AlphaFold** button overlaying that\nprotein's AlphaFold model, superposed server-side on the confident core (pLDDT >= 70) and\nreporting the accession, how it was resolved, how many Ca it was fitted on, and the RMSD.\n\n**Two campaign-level panes** sit below: **Ligand pose**, every ligand overlaid as sticks in one\nframe with no protein, and **Superposed targets**, every target as a Ca trace with its RMSD to the\ncampaign's reference beside it. Each has a checkbox per target, and each RMSD is given over the\nresidues that actually agree, with that count.", "# 2022 \u2013 Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant FGFR2/3\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: papers).*\n\nStructures\n\nFGFR2 Kinase Domain in Complex with a Pyrazolo[1,5-a]pyrimidine Analog\n\n* [Link to Paper](https://marcdeller.com/wp-content/uploads/2025/09/shvartsbart-et-al-2022-discovery-of-potent-and-selective-inhibitors-of-wild-type-and-gatekeeper-mutant-fibroblast.pdf)\n\n* Potent Dual Inhibition: Compound 29 retains low-nanomolar potency against wild-type and gatekeeper FGFR2/3, with excellent ADME and PK profiles.\n* PROTAC Strategy: Selective FGFR2 degrader 28e achieves DC\u2085\u2080 = 0.645 nM, sparing FGFR1/3/4 and overcoming V564F resistance.\n* AI-Driven Design: Generative chemistry engines identify novel cores enabling high FGFR2/3 selectivity and in vivo efficacy.\n* Dynamics-Based Selectivity: Lirafugratinib (RLY-4008) exploits protein motion to distinguish FGFR2 from FGFR1, showing 73% clinical response and reduced hyperphosphatemia.\n* Irreversible Covalent Inhibition: BW710 covalently binds FGFR2 with high potency and oral bioavailability, sparing other FGFRs.\n* Precision Oncology Impact: Selective FGFR2/3 inhibitors promise durable responses with improved safety profiles.\n\n1. **Artem Shvartsbart, et.al.,** [ACS J. Med. Chem](https://pubs.acs.org/doi/abs/10.1021/acs.jmedchem.2c01366)\n2. **Design, synthesis and antitumor activity of a novel FGFR2-selective degrader to overcome resistance of the FGFR2V564F gatekeeper mutation based on a pan-FGFR inhibitor:** [Zuli Hu, et.al., Eur. J. Med. Chem](https://doi.org/10.1016/j.ejmech.2024.116612)\n3. **Discovery of TYRA-300: First Oral Selective FGFR3 Inhibitor for the Treatment of Urothelial Carcinoma:** [Yoshikazu Nakaki, et.al., J. Med. Chem](https://pubs.acs.org/doi/10.1021/acs.jmedchem.4c01531)\n4. **Discovery of lirafugratinib (RLY-4008), a highly selective FGFR2 inhibitor:** [Lakshmi N. Palaniappan, et.al., PNAS](https://www.pnas.org/doi/10.1073/pnas.2317756121)\n5. **Discovery of BW710 as a potent, selective and orally bioavailable FGFR2 inhibitor:** [Yalan Xie, et.al., Eur. J. Med. Chem](https://pubmed.ncbi.nlm.nih.gov/39908791/)\n6. **Insilico Medicine publishes AI-driven innovative design strategy for highly selective FGFR2/3 inhibitors:** [Insilico Medicine, et.al., J. Med. Chem](https://insilico.com/tpost/hr2eyfocx1-jmc-insilico-medicine-publishes-ai-drive)\n\nThe fibroblast growth factor receptor (FGFR) family, particularly FGFR2 and FGFR3, plays a crucial role in cellular growth and survival, with aberrant activation implicated in cancers such as cholangiocarcinoma and bladder cancer. First-generation pan-FGFR inhibitors, while effective, often cause hyperphosphatemia and lose potency against gatekeeper mutations. This study describes the rational design and optimization of a new class of FGFR2/3 inhibitors that maintain high potency against both wild-type and gatekeeper mutant forms, while exhibiting strong selectivity over FGFR1 and other kinases. Key advances included structure-based drug design, strategic modifications to the inhibitor scaffold, and careful tuning of pharmacokinetic properties, leading to the identification of compound 29. This molecule demonstrated sub-nanomolar potency for FGFR2/3, excellent selectivity, favorable ADME properties, and robust oral bioavailability in preclinical models. Importantly, the new inhibitors avoid the reliance on interactions that are compromised by gatekeeper mutations, addressing a major limitation of earlier drugs.\n\nFor patients, these advances could translate to more effective and safer cancer therapies targeting FGFR-driven tumors. The high selectivity for FGFR2/3 over FGFR1 is particularly significant, as it may reduce the risk of treatment-induced hyperphosphatemia, a common and sometimes dose-limiting side effect of current FGFR inhibitors. The ability to inhibit common resistance mutations means these next-generation drugs could offer durable responses where first-generation therapies fail. Early pharmacodynamic studies in animal models showed effective pathway inhibition without significant changes in serum phosphate, supporting the potential for improved tolerability in clinical settings.\n\n| Concept | Description | Key Reference |\n| --- | --- | --- |\n| Gatekeeper Resistance | Mutations like FGFR2^V564F^ reduce inhibitor potency. | [Shvartsbart et al.](https://pubs.acs.org/doi/abs/10.1021/acs.jmedchem.2c01366) |\n| PROTAC Degradation | Targeted FGFR2 degradation overcoming mutant resistance. | [Hu et al.](https://doi.org/10.1016/j.ejmech.2024.116612) |\n| Generative Chemistry | AI-driven scaffold design for FGFR2/3 selectivity. | [Insilico Medicine](https://insilico.com/tpost/hr2eyfocx1-jmc-insilico-medicine-publishes-ai-drive) |\n| Dynamics-Based Design | Protein motion modeling yields FGFR2-specific inhibitors. | [Palaniappan et al.](https://www.pnas.org/doi/10.1073/pnas.2317756121) |\n| Covalent Inhibition | Irreversible FGFR2 engagement enhances potency. | [Xie et al.](https://pubmed.ncbi.nlm.nih.gov/39908791/) |\n| Clinical Translation | Selective inhibitors demonstrate improved safety and efficacy profiles. | [Nakaki et al.](https://pubs.acs.org/doi/10.1021/acs.jmedchem.4c01531) |", "# Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\nAuthors: Zhuo, J.; Li, Y.-L.; Qian, D.-Q.; Burns, D.M.; Mei, S.; Rafalski, M.; Xu, M.; Cao, G.; Pan, Y.; Jia, Z.; Jalluri, R.; Epling, L.B.; Fenalti, G.; Deller, M.C.; Procak, J.; Covington, M.; He, X.; Collins, R.; Stubbs, M.; Burke, K.; Oliver, J.; Margulis, A.; Boer, J.; Wynn, R.; Scherle, P.; Diamond, S.; Newton, R.; Zhang, Y.; Metcalf, B.; Yao, W.\nSource: Journal of Medicinal Chemistry (Drug Annotation), 2026\n\nDespite the ~2-fold loss in permeability, 22 had a dramatic improvement in the PK profile in both rat and cyno species, with a doubling of oral bioavailability in both species to 100% free (Table 6). In beagle dogs, 22 displayed a significant improvement in hepatic blood flow clearance from 82% to 4%, compared to 18. The pronounced improvement in dog clearance resulted in a marked increase in both oral exposure and oral bioavailability, with 165,386 nM*h exposure and 100% oral bioavailability, compared to a 65 nM*h exposure and 1% oral bioavailability for 18 in dogs. In summary, the perturbation of the hinge-binding motif to the 3,5-dimethylpyrazole 22 resulted in a dramatic improvement in the PK profile for all three species, accompanied by an increase in JAK1 isoform selectivity, albeit with a loss in potency. It should be noted that neither 18 or 22 inhibited CYPs 1A2, 2B6, 2C9, 2C19, 2D6, or 3A4 below 25 \u03bcM or HEK293 proliferation below 20 \u03bcM.\n\nThe crystal structure of povorcitinib 22 bound to JAK1 (PDB 10PI) and JAK2 (PDB 10PJ) revealed that 22 binds to the two JAK isoforms in two very similar yet discrete conformations (Figure 4A). These two binding poses are in alignment with conventional type 1 inhibitors, such as the binding of baricitinib 2 to JAK2 (Figure 4B). The two nitrogen atoms of the hinge-binding pyrazole ring of 22 forms hydrogen-bond interactions with the main-chain amide of Leu932 and the main-chain carbonyl of Glu930 in the hinge region of JAK1 and JAK2. These key H-bonding interactions within the hinge region are comparable to the 7H-pyrrolo[2,3-*d*]pyrimidine heterocycle found in inhibitors such as baricitinib 2 and compound 18. Despite both hinge-binding heterocycles having conserved H-bonding interactions with the hinge region, the pyrazole ring has a 44-fold loss in enzymatic potency when compared to the 7H-pyrrolo[2,3-*d*]pyrimidine hinge-binding heterocycle of 18. One explanation for this loss in potency may be that the H-bonding interaction with the hinge region is slightly weakened for the 5-membered pyrazole ring due to a suboptimal geometry that orients the trajectory of the electrons on the N2 and the proton on N1 of the pyrazole ring in a slight misalignment with their complementary H-bonding donor/acceptor pair residues of the hinge region. It is also plausible that the loss in potency may be due to the smaller 5-membered pyrazole ring having a reduction in \u03c0-\u03c0-stacking and van der Waals interactions within the hinge region and that the larger 7H-pyrrolo[2,3-*d*]pyrimidine moiety has an enhanced shape complementarity for the JAK hinge region pocket.\n\nIn addition to the differences in the hinge-binding motif, another important modification for povorcitinib 22 vs conventional dual JAK1/2 inhibitors, such as baricitinib 2, is that 22 is significantly longer in length (Figure 4B). Thus, the phenyl ring of 22 extends toward the P-loop of both JAK1 and -2. A structural comparison of the X-ray crystal structures of JAK1 and -2 bound to povorcitinib 22 (Figure 4C\u2013H) indicates that there are differences in the amino acid composition in this region of the protein that could contribute to an upward shift of ~2.0 \u00c5 in the P-loop for JAK1 (Figure 4C and 4D). In particular, JAK1 has an His885 that is directed toward the solvent, whereas JAK2 has Asn859 that is bent toward the binding pocket and forms a hydrogen bond with Glu1015 (Figure 4E). It is hypothesized that the trifluoromethyl group of the terminal amide of 22 would clash with the P-loop of JAK2 causing a decrease in binding affinity. The additional selectivity of 22 may be due to the smaller hinge-binding pyrazole ring of 22 shifting the terminal amide trifluoromethyl capping group closer to the P-loop, thus exploiting the additional space provided by JAK1 compared to JAK2 and imparting additional selectivity. In addition, the phenylalanine residue (F860) that is present in JAK2 may orient the phenyl ring toward the pocket to adopt a \"closed\" disposition, akin to the DFG pocket, which may preclude povorcitinib 22 from efficiently binding to JAK2 (Figure 4E\u2013H). Finally, the region connecting the C-helix to the \u03b2-sheet of JAK1 and -2 revealed subtle structural differences that may facilitate an H-bonding interaction between Asn859 and Glu1015 to further \"close\" the pocket of JAK2 toward binding to 18 and povorcitinib 22 (Figure 4E and 4G). These observations are speculative and under further investigation but are in agreement with the observed JAK1/2 selectivity data.", "# BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n| Output | Description |\n| --- | --- |\n| `boltz_run_<timestamp>.log` | Raw `boltz predict` output for the run |\n| `boltz_output/` | Boltz\u2019s own prediction output tree |\n| `boltz_cif/` | Every completed target\u2019s `*_model_0.cif`, flattened into one folder |\n| `boltz_summary.csv` / `boltz_summary.xlsx` | One row per target: every scalar field from the confidence/affinity JSONs, computed pIC50, the input ligand SMILES, and `flags`/`notes` columns (`LOW_CONFIDENCE`, `HIGH_CONFIDENCE_POOR_AFFINITY`, `LOW_CONFIDENCE_STRONG_AFFINITY`, `LOW_POCKET_PLDDT`, `MISSING_OUTPUTS`). The XLSX adds a `selectivity` sheet (ligand x family pIC50 pivot) whenever a campaign spans more than one protein family. When `setup-plip` has run, also gets a `plip_status` column (`ok` / `no_interactions` / `failed` / `ambiguous_ligand` / `skipped_no_env`) and a `plip_<type>_count` column per interaction type detected (hydrophobic, hydrogen_bonds, salt_bridges, etc.). |\n| `boltz_summary_view.csv` | The same columns shown in the dashboard\u2019s \u201cSummary table\u201d (see below) \u2014 a trimmed, renamed subset of `boltz_summary.csv`, for anyone who wants the at-a-glance view in a spreadsheet rather than every raw field |\n| `boltz_ligand_grid.pdf` (optional) | A print/share-friendly PDF of the dashboard\u2019s \u201cLigand structures\u201d grid \u2014 same 5\u00d75 pagination, same rendered structures, severity borders and scaffold highlighting as the on-screen version, in the style of [smiles2grid](https://github.com/bellcheddar/smiles2grid)\u2018s own PDF output. Only written when a campaign has at least one SMILES ligand to render. |\n| `boltz_plip/` (optional) | Per-target cif2plip output: the converted PDB, PLIP\u2019s XML/TXT reports, the ray-traced binding-site PNG, and the PyMOL `.pse` session \u2014 cached here so re-running `analyze` doesn\u2019t re-profile a target that\u2019s already been done |\n| `boltz_interactions.csv` (optional) | Long format, one row per detected contact across every target: interaction type, residue, distance \u2014 the raw data behind the dashboard\u2019s fingerprint heatmap and per-target contact tables |\n| `boltz_dashboard_sessions/` (optional) | Each target\u2019s PyMOL `.pse` session, copied here and linked from the dashboard \u2014 this is the one thing that makes `boltz_dashboard.html` no longer a single self-contained file once interaction analysis has run; without `setup-plip`, the dashboard stays exactly as self-contained as before |\n| `boltz_dashboard.html` | Posts its own real content height to any parent window via `postMessage` on load and resize, so a page embedding it in an iframe (e.g. `findings.md`\u2018s \u201cInteractive dashboard\u201d section) can size the iframe to fit the actual content instead of guessing a fixed height \u2014 a cross-origin iframe can\u2019t otherwise be measured/resized from the embedding page\u2019s own JS. A campaign summary table with a third \u201cDetails\u201d column alongside Field/Value \u2014 a linked path to the input file, each protein/partner\u2019s id and sequence length, each ligand\u2019s id and SMILES-vs-CCD source, the full list of target stems, which specific ligands got flagged in ligand-chemistry review (linking to the card below), and a plain-English gloss for each of the more cryptic run parameters (accelerator, MPS watermark, recycling/sampling steps, etc.) \u2014 tracked across every `run` invocation in a small hidden sidecar file. Then a \u201cSummary table\u201d directly below it: grouped into named column bands (Identity, Confidence, Affinity, Interactions, Structure) with short human headers instead of raw JSON field names, redundant/granular columns (per-chain and per-chain-pair confidence breakdowns, individual ensemble sub-model values) hidden by regex pattern rather than a fixed list \u2014 so it scales correctly to campaigns with more than two chains \u2014 and two download links, one for the full underlying CSV and one for a CSV matching just this trimmed/renamed view. The \u201cTarget\u201d column shows a `{group}_{partners}_{ligand}` display name (e.g. `5HT2A_GNAQ+GNB1+GNG2_RISP`, partners omitted when there are none, `apo` in place of a ligand for a ligand-free target) rather than the internal per-variant family id/stem (`H2ANG_RISP`, `H2AAP`) \u2014 and this isn\u2019t just a table label: the same display name (or its family-level `{group}_{partners}` form, with no ligand, for whole-family contexts) replaces the internal id in every chart tick/legend/point label (ranked pIC50, ranked confidence, the pIC50-vs-confidence-score and pIC50-vs-binder-probability scatters, interaction counts, the residue-interaction fingerprint heatmap), every per-target/per-family card title, the campaign-summary target list, and the selectivity pivot\u2019s columns (both the dashboard heatmap and the XLSX `selectivity` sheet) \u2014 see **compare-sse** below for the same treatment there. The raw per-variant ids stay alongside the display name in every underlying CSV/XLSX `targets` sheet, for cross-referencing against real output filenames. A \u201cPartner\u201d column lists each target\u2019s co-folded partner chain(s) (hidden when the campaign has none), and rows are grouped by `Group:`/family id with a blue top border marking each new group \u2014 the same blue used for column-group boundaries, just rotated. The \u201cFlags\u201d column is renamed \u201cSummary\u201d and icon-based: a bullseye (affinity) and a shield (confidence) icon per row, each tinted green/amber/red by tier (exact value and interpretation on hover), reusing the existing `LOW_CONFIDENCE_THRESHOLD` and a symmetric buffer around Boltz\u2019s documented 0.5 binder decision boundary \u2014 Boltz\u2019s own docs define these metrics\u2019 [0, 1] range but publish no official tri-colour bands. A `MISSING_OUTPUTS` failure collapses the cell to a single red cross; a legend to the right of the download links spells out all six tier/icon combinations. Always shown now (previously hidden entirely when nothing was flagged), so a clean campaign reads as a row of green icons rather than a column that silently disappears. A ligand-free (apo) target\u2019s ligand/affinity/interface/interaction columns (including the bullseye) show an explicit `N/A` rather than a blank cell or a misleading `0.00`, since there\u2019s no ligand or inter-chain interface for those to describe. Then a \u201cLigand preparation\u201d card (the same stereocentre/protonation-state/disconnected-fragment checks as `preflight`\u2018s `ligand_preparation` check, shown per-ligand rather than as a single summary line), then a \u201cLigand structures\u201d card: a paginated 5\u00d75 grid of every ligand\u2019s rendered 2D structure (building on [smiles2grid](https://github.com/bellcheddar/smiles2grid)\u2018s design, adapted for a single campaign\u2019s scale), with stereocentre/ionizable-group findings highlighted directly on each structure, ligands sharing a Bemis-Murcko scaffold (or, failing that, a verified whole-group maximum-common-substructure) grouped and colour-highlighted together with their depictions aligned to a common orientation, and a captioned legend (badge-by-badge: what S/A/N/Ph/SO3/salt each mean, plus the cluster colour key) stating exactly what was found and on how many ligands \u2014 never an unexplained highlight \u2014 plus \u201cDownload PDF\u201d (the same grid as a print/share-friendly file, `boltz_ligand_grid.pdf`) and \u201cDownload SMILES\u201d (`boltz_ligands.csv`: ID, SMILES, stereocentre/ionizable-group/fragment findings, MW, cLogP, TPSA) links side by side on one line, matching the Summary table\u2019s own download-links style. Then interactive [Plotly](https://plotly.com/javascript/) charts in a grid (ranked pIC50, ranked confidence, a \u201cpIC50 vs confidence score\u201d scatter, interaction counts by type, then a \u201cpIC50 vs binder probability\u201d scatter (binder probability on x, pIC50 on y) \u2014 hover/zoom/pan; plotly.js itself is vendored and inlined into the file rather than CDN-loaded, so the dashboard has no runtime dependency on an external script host). The two scatter charts colour each point by tier via a continuous colourscale + colorbar legend (the same style as the Family x ligand selectivity heatmap\u2019s own colorbar) \u2014 confidence tier (matching the Summary table\u2019s shield icon) for pIC50-vs-confidence-score, affinity tier (matching the bullseye icon) for pIC50-vs-binder-probability \u2014 and, when a `Ligand:` block sets the optional `Role: agonist`/`Role: antagonist` field, shape-code points by pharmacology (circle = agonist, diamond = antagonist) with a legend positioned inside the plot area\u2019s top-left corner (not Plotly\u2019s default outside-right position, which would otherwise collide with the colorbar); campaigns that don\u2019t set `Role:` see a single unshaped trace, unchanged from before. When `setup-plip` has run: a per-family residue-interaction fingerprint heatmap (also interactive Plotly \u2014 shown for every family with interaction data, even a single ligand, though the similarity-based reordering that helps SAR ranking within a series only kicks in from 3+ ligands) and, per target, its binding-site image (residues labelled and interaction distances shown \u2014 PLIP\u2019s own images have neither, so these are re-rendered from its PyMOL session with both added, with a \u201cDownload image\u201d link of its own) next to an interactive, auto-rotating [3Dmol.js](https://3dmol.org) view of the same predicted structure (built directly from the mmCIF, ligand highlighted), side by side with a table of that target\u2019s contacts (with its own \u201cDownload CSV\u201d link) plus a download link for the full PyMOL session. Finally, a \u201cSecondary structure shifts\u201d card (see **compare-sse** below): a \u201cFamily coverage\u201d table (every protein family in the campaign, with its status \u2014 `OK` and a target/motif count, or a plain-English reason it was skipped, e.g. \u201cNo apo structure configured\u201d), an \u201cOverall shift statistics\u201d summary, and, when there\u2019s data, the full per-motif table plus its own Plotly charts. |\n| `boltz_sse_comparison.csv` / `.html` | Written automatically by `analyze`/`all` whenever any family has `Apo structure:` set (or on demand via the standalone `compare-sse` command). One row per family/target/motif: Ca RMSD, centroid shift, helix-axis rotation/kink angles, SSE boundary shift, flagged phi/psi residues, and (kinases) DFG-in/out and alphaC-in/out states \u2014 a metric that genuinely wasn\u2019t computed for a motif shows as `N/A`, not a blank cell. The Family/Target columns, chart legends, and family-coverage table all show the same `{group}_{partners}` / `{group}_{partners}_{ligand}` display names used throughout the main dashboard; the CSV also keeps the raw `family_id`/`target_stem` columns alongside for cross-referencing. The HTML is a standalone dashboard (family coverage, overall shift statistics, Plotly bar chart + motif x target heatmap); the same content is also embedded directly into `boltz_dashboard.html` (see below) |\n| `boltz_sse_family_status.json` | One entry per protein family: `ok` (with a target/motif count) / `no_apo_structure` / `apo_not_found` / `annotation_failed` / `no_predicted_structures` \u2014 the machine-readable form of the dashboard\u2019s \u201cFamily coverage\u201d table, so a family with no `Apo structure:` configured reads as \u201cnot configured\u201d rather than silently missing |\n| `boltz_sse_comparison_sessions/` (optional) | A plain-text PyMOL `.pml` script per target \u2014 colours/labels each motif, highlights the ones with a significant shift |", "# GitHub README: ALPHABETTI\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ALPHABETTI, 2026\n\n- [x] **Backend science modules** \u2014 geometry, language model, accessibility and sequence handling, all pure numpy and Biopython, unit-tested with no GPU and no network.\n- [x] **Per-residue frames** \u2014 orthonormal, right-handed, with a degenerate-case guard and both termini handled; mirrored function for function in `orientation.js` for FOLDEROL.\n- [x] **Virtual C\u03b2 for glycine** \u2014 validated to 0.13 \u00c5 mean against real C\u03b2 positions in 1UBQ. Never skipped.\n- [x] **Fixed a 180\u00b0 dihedral sign error** \u2014 helices sat at \u2212129\u00b0 where P-SEA publishes +50\u00b0, so every angle criterion silently failed. The CA-only secondary structure assignment went from 55 % to **83.7 %** against DSSP. This matters because `mkdssp` is not on the droplet, making the fallback the production path.\n- [x] **Caught DSSP failing on every structure** \u2014 `mkdssp` rejects any file with no PDB `HEADER` line and ESMFold emits none, while a bare `except: pass` reported it as merely absent.\n- [x] **ZeroGPU fold service** \u2014 ESMFold and ESM-2 650M behind a JSON API, 12.7 s wall for a first fold against an estimated 3-8 minutes on a CPU droplet.\n- [x] **Masked marginals by default** \u2014 after measuring that unmasked marginals return the wild type at 100 % of positions.\n- [x] **Instanced glyph renderer** \u2014 one `InstancedMesh` per amino acid, twenty draw calls regardless of chain length.\n- [x] **Baloo 2 to `typeface.json`** \u2014 with the curve argument order verified against the source outlines to 0.34 %.\n- [x] **All four tabs** against a single payload, switching with no network round trip.\n- [x] **Four examples spanning the fold classes** \u2014 beta-grasp, mixed alpha/beta, all-alpha and a TIM barrel, each folding above pLDDT 90, each using its mature chain rather than the UniProt precursor.\n- [x] **Exports** \u2014 GLB, STL, PNG at 1x/2x/4x, and an animated GIF over one FOLDEROL loop.\n- [x] **49 tests**, including regressions for the bugs found during the build.\n- [x] **`BENCHMARKS.md`** with real timings from the machines that serve the app.\n- [x] **Neon Signage visual direction**, chosen from five rendered candidates.\n- [ ] **DNS and TLS** \u2014 `alphabetti.mdeller.com` needs an A record to 45.55.102.228 and a certbot certificate; mdeller.com's certificate does not cover subdomains.\n- [ ] **Add to the mdeller.com launcher** \u2014 one entry at the top of `apps.json`, with the beacon pointing at the 3D typeface, since a scanner never fetches it.\n- [x] **HOGWASH** \u2014 a real WebLogo from a real alignment, with every WebLogo option exposed, seven output formats, Pfam fetch and five bundled example alignments.\n- [x] **HOGWASH in 3D** \u2014 the logo is drawn in the glyph engine as a helix, ring, strip or rows, orbited and zoomed like everything else. The flat WebLogo is now an export only.\n- [ ] **HOGWASH in 3D** \u2014 map the alignment's columns onto the structure and compare, per position, what evolution did against what the language model expects. The `LogoData` is already in the payload; the gap-to-residue mapping is not built.\n- [ ] **BALDERDASH substitution heatmap panel** \u2014 the 20 x L matrix is already in the payload and colour-mapped; the clickable panel that flies the camera to a position is not built.\n- [ ] **Mobile testing on a real phone** \u2014 the bottom-sheet layout and pinch-zoom are written but have only been checked at emulated widths.\n- [ ] **Verify GLB and STL open in Blender** \u2014 the exporters run and produce files of a sensible size, but nobody has opened one yet.\n- [ ] **Revisit the 400-residue cap** \u2014 set against an assumed CPU fold of minutes; at 32 s on ZeroGPU the honest constraint is now quota and payload size.\n- [ ] **A second fold backend** \u2014 the Space is a single point of failure, and `FoldBackend` already has the seam for a fallback.\n\n## \ud83d\udcc4 Licence\n\nMIT.\n\n## \ud83d\ude4f Citations\n\n- Lin Z, Akin H, Rao R, *et al.* (2023) Evolutionary-scale prediction of atomic-level protein structure with a language model. *Science* **379**:1123-1130.\n- Meier J, Rao R, Verkuil R, *et al.* (2021) Language models enable zero-shot prediction of the effects of mutations on protein function. *NeurIPS*.\n- Tien MZ, Meyer AG, Sydykova DK, Spielman SJ, Wilke CO (2013) Maximum allowed solvent accessibilities of residues in proteins. *PLoS ONE* **8**(11):e80635.\n- Labesse G, Colloc'h N, Pothier J, Mornon JP (1997) P-SEA: a new efficient assignment of secondary structure from C\u03b1 trace of proteins. *CABIOS* **13**:291-295.\n- Schneider TD, Stephens RM (1990) Sequence logos: a new way to display consensus sequences. *Nucleic Acids Research* **18**:6097-6100.\n- Kabsch W, Sander C (1983) Dictionary of protein secondary structure. *Biopolymers* **22**:2577-2637.\n- Kyte J, Doolittle RF (1982) A simple method for displaying the hydropathic character of a protein. *J Mol Biol* **157**:105-132.\n\n---\n\nBuilt by **Marc C. Deller, D.Phil.** \u00b7 [marcdeller.com](https://marcdeller.com) \u00b7 [marc@marcdeller.com](mailto:marc@marcdeller.com)", "# BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nAt parse time, every ligand SMILES is canonicalised with [RDKit](https://www.rdkit.org/) so the same molecule is represented identically everywhere downstream. Each ligand is then checked for undefined stereocentres (via `FindMolChiralCenters`), disconnected fragments that usually mean a salt left in the string (`GetMolFrags`), and ionizable groups whose protonation state a plain SMILES leaves unspecified (SMARTS matches for carboxylic acids, amines, phenols, and sulfonic acids). None of these is a hard failure, because any of them can be a deliberate modelling choice; they are flags worth a second look before you trust downstream numbers. Separately, the dashboard\u2019s ligand grid asks whether any compounds share a chemical core, grouping them first by exact [Bemis-Murcko scaffold](https://pubs.acs.org/doi/10.1021/jm9602928) and then, for near-analogues, by fingerprint similarity backed by a verified maximum common substructure, so a highlighted core is always a proven match rather than a bare similarity score.\n\nFor the predicted complexes themselves, [cif2plip](https://github.com/bellcheddar/cif2plip) converts each Boltz output into a strict PDB and runs [PLIP](https://academic.oup.com/nar/article/53/W1/W463/8128215) for real interaction fingerprints: hydrogen bonds, salt bridges, pi-stacking, halogen bonds, and metal coordination, re-rendered through [PyMOL](https://www.pymol.org/) into a labelled binding-site view with distances.\n\nThis is where structure prediction, cheminformatics, and interaction analysis meet in one place. The same lineage of diffusion-based models that [AlphaFold 3](https://www.nature.com/articles/s41586-024-07487-w) established now produces poses in minutes, but those poses are only ever as trustworthy as the chemistry fed into them, which is why the validation steps that BoltzMaker performs are essential.\n\n###### **Watching the protein move by following Secondary Structure Elements (SSE): compare-sse**\n\nA confidence score tells you what Boltz predicted; it says nothing about how the protein moved when the ligand arrived. That is a real structural question whenever you hold both a reference apo (unbound) structure and a predicted holo one for the same protein, and it is the question `compare-sse` answers, in the terms a structural biologist reasons in (\u201cTM6 swung out 4.2 \u00c5\u201d, \u201cthe DFG motif flipped from in to out\u201d) rather than raw DSSP fragment coordinates.\n\nMotifs are annotated by one of three pluggable sources, auto-selected per family or pinned with `Family type`. For a GPCR, generic numbering comes from GPCRdb (Ballesteros-Weinstein and GPCRdb schemes), giving TM1 to TM7, H8, and the extracellular and intracellular loops. For a kinase, KLIFS supplies its fixed 85-residue pocket alignment and the canonical motifs (hinge, gatekeeper, HRD catalytic loop, DFG, \u03b1C-Glu, catalytic Lys). Anything outside those two families falls back to Pfam domain boundaries via PDBe\u2019s SIFTS residue mapping, the universal last resort. Before any motif is measured, apo is superposed onto holo using only the family\u2019s stable, non-pocket residues (through gemmi\u2019s `superpose_positions`), so a local, ligand-induced shift cannot quietly skew the global fit and inflate every downstream number.\n\nPseudo code for the SSE workflow is outlined below:\n\n```\n\n\t\t\t\t\t$ python3 BoltzMaker.py compare-sse boltz_input.md\n\nfor each target that supplies an \"Apo structure\"\n\n    family = Family type:        # gpcr | kinase | auto (default)\n\n    motifs = annotate (family)\n        gpcr    -&gt; GPCRdb generic numbering  (TM1-7, H8, ECL1-3, ICL1-3)\n        kinase  -&gt; KLIFS 85-residue pocket   (hinge, gatekeeper, HRD, DFG, aC-Glu, Lys)\n        auto    -&gt; whichever fits, else Pfam / SIFTS domain boundaries\n\n    superpose (apo -&gt; holo)       # stable, non-pocket residues only (gemmi)\n\n    for each motif:\n        Ca RMSD / centroid shift\n        helix-axis rotation + kink   (apo / holo / delta)\n        SSE boundary shift           (HELIX/SHEET records, or optional dssp)\n        phi/psi outliers             (&gt; --phi-psi-threshold)\n        DFG-in/out, aC-in/out        (kinases; coarse Ca-Ca classifier)\n\nwrite  boltz_sse_comparison.csv                    # one row per motif\n       boltz_sse_comparison.html                   # Plotly bar + motif-by-target heatmap\n       boltz_sse_comparison_sessions/&lt;target&gt;.pml   # plain-text PyMOL, no venv dep\n\n```\n\nEach motif then earns a compact panel of metrics: C\u03b1 RMSD and centroid shift after superposition; a helix-axis rotation angle for helical motifs, which is where the classic TM6 outward swing on GPCR activation shows up; a helix kink angle reported as apo, holo, and delta; an SSE boundary shift for whether a helix or strand grew or shrank (the one metric that needs deposited HELIX/SHEET records or an optional `mkdssp` binary, while everything else works without it); and per-residue backbone dihedral outliers above a phi/psi threshold. Kinases add a coarse DFG-in/out and \u03b1C-in/out call read from C\u03b1-C\u03b1 distances, built to catch a state change between apo and holo rather than to pass as publication-grade conformational classification. The command writes three artifacts next to the spec: a `boltz_sse_comparison.csv` with one row per family, target, and motif; a self-contained `boltz_sse_comparison.html` carrying a Plotly bar chart and a motif-by-target heatmap, vendored exactly like the main dashboard; and one plain-text PyMOL script per target that colors and labels each motif and flags the ones that moved. Those `.pml` files are just text, so they open in any local PyMOL without adding a `PyMOL` dependency to BoltzMaker\u2019s own environment.\n\n**Important note:** Currently, only GPCRs and Kinase family types are implemented.\n\nTo utilize\u00a0compare-sse, the following \u201cApo structure\u201d\u00a0and \u201cFamily type\u201d fields\u00a0are required in the boltz_input.md:\n\n```", "# GitHub README: CODSWALLOP\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/CODSWALLOP, 2026\n\n- [x] **The two machines stay in step by themselves.** They hold different halves of the same app and neither can do the other's job: the droplet is the only one that knows what readers have asked for, because families they assemble live in a cache database deliberately excluded from every rsync, and the workstation is the only one that can build an embedding, a fingerprint or a topology. So the family set has to travel one way and the artefacts the other, and until `deploy/sync.sh` neither happened on its own: the droplet held 13 families, the workstation 34, and the landing page showed a third of the archive. The sync is keyed on the **query** rather than the slug, because a slug is derived from the seed and moves when a protein is renamed upstream, while the query is what a person actually typed and is the only thing either machine can rebuild from. Run daily from launchd rather than cron, since this Mac is usually asleep at four in the morning and `StartCalendarInterval` runs at the next wake where a cron entry simply misses it, twenty minutes behind the droplet's own warm so it reads a settled list *(Superseded below: there is no longer a second machine to stay in step with, and `deploy/sync.sh` with its launchd timer is deleted.)*\n- [x] **Artefacts for families a reader creates, not just ones somebody pre-warmed.** The embedding and the interaction fingerprint are built on a workstation and rsynced across, because the droplet has two cores shared with eight apps and deliberately has neither biotite, tmtools, PLIP nor OpenBabel installed. So a family assembled for the first time by a reader on the live site had an artefact on neither machine and rendered the identity placeholder indefinitely, since nothing outside a hand-written pre-warm list ever asked for it. The droplet now records the request (`artefact_request`, keyed on slug, counting hits so the queue is a priority order rather than a log) and `deploy/drain_queue.sh` collects from the workstation: read the queue over SSH, build, push, and only then mark each one served, because marking on build completion empties the queue on a failed push, which is the one state where the record matters most. A family that later needs *more* than it was served, which is what a pipeline version bump does to every family at once, returns to the queue rather than staying quietly marked done. The map's own note was the other half: it said \"run `CODSWALLOP.py embed` on a workstation\", which is advice for the one reader who is me, and now states plainly that the layout is not a structural measurement, why it cannot be computed while you wait, and that the family has been queued *(Superseded below: there is no second machine to collect from, so the queue is drained in place by `deploy/worker.sh`.)*\n- [x] **Topology diagrams from DSSP as custom SVG.** Strands as arrows, helices as cylinders, built in-house rather than scraped from PDBsum, which is copyrighted. Drawn on the **seed axis**, so an element sits directly under the same residues as the conservation track, the coverage census, the domain ribbon and the motifs tab, and mapped there by **alignment rather than by an assumed offset**: an entry may be numbered on the mature protein, on the construct from 1, or on the canonical, and an assumed offset is right often enough to look like it works. The connectivity is the part worth having and the part a linear track cannot express: the arcs are **DSSP's bridge partners**, so they show which strands actually hydrogen-bond to which, weighted by how many bridges support each pair so one spurious bridge does not draw the same line as a ten-residue pairing. GFP comes out as eleven strands with the long-range arc that closes the barrel; KRAS as the six-stranded mixed sheet of a G domain; myoglobin, haemoglobin and ferritin as all-helical with no arcs at all. DSSP proper when `mkdssp` is on the PATH (CCP4 ships it) and biotite's P-SEA otherwise, with the artefact recording which, because the two disagree at element boundaries and P-SEA has no bridge partners to give: the pairing is then absent rather than guessed\n- [x] **Motifs tab (first half: sequence sites, grounded in the family).** Functional sites on the seed in one place, from **UniProt's curated features** (active and binding sites, modified residues, glycosylation, lipidation, disulphides, transmembrane spans, signal peptides, short linear motifs) and from **PROSITE** signatures via ScanProsite. The panel says which is which and does not mix them, because they are different claims: a UniProt feature is somebody having read a paper about this protein, a PROSITE hit is a pattern having matched a string. The **high-probability PROSITE patterns are excluded by default**, and that single flag is most of the value here: scanning chicken lysozyme with them on returns a PKC phosphorylation site and two N-myristoylation sites, on a secreted protein that has none of them. Six hits become two, and the two are the lysozyme-like domain profile and the glycosyl hydrolase family 22 signature. Every site carries **the share of the family whose construct contains it and how often it is actually resolved when present**, which is what makes it a family panel rather than a sequence-annotation panel: lysozyme's signal peptide is in 4.0 % of constructs and resolved in 0.7 %, because it is cleaved, and no single entry's page can tell you that. \"Resolved\" is expressed against the constructs that contained the site, so a site nobody has ever cloned reads as unknown rather than as disordered. It also found a real bug in the feature parser: UniProt gives a disulphide as its two paired cysteines, and expanding that to a range turned lysozyme's 24&ndash;145 bond into 122 consecutive \"disulphide bonds\", one per residue\n- [x] **Family-specific reference numbering in the motifs tab**, following what BoltzMaker does. **KLIFS** for kinases and **GPCRdb** for receptors, used differently because the two are shaped differently. GPCRdb returns per-residue generic numbering keyed on the **UniProt sequence number**, so it lands directly in seed coordinates beside the conservation track and the coverage census: A2A's DRY, CWxP, NPxxY and PIF micro-switches are located by generic number rather than by sequence, which is the entire point of a generic scheme, so a receptor whose DRY is a DRF still has its switch found and shown as DRF. The full segment map (TM1&ndash;TM7, the loops, H8, C-term) comes with it. KLIFS numbers its 85-residue pocket against each deposited structure's own author numbering, which is not the seed's, so rather than guess a mapping this reports the thing KLIFS alone can say about a *family*: **which conformation each of its structures was caught in**. ABL1 is 43.0 % DFG-out, 31.6 % out-like and 25.3 % DFG-in across the 158 of its entries KLIFS holds, with the &alpha;C-helix in on 95.6 %. A single entry's page can tell you it is DFG-out; only the family can tell you what fraction of the field that is. A family that is neither gets neither, rather than an empty kinase pocket for a lysozyme. Licences: GPCRdb releases its data CC BY and KLIFS is free for academic use, both attributed on the panel\n- [x] **Assembly and oligomeric state panel.** The family-level question a single entry page cannot answer: has a protein deposited 1,686 times ever been seen as anything other than the state everyone quotes? Lysozyme is 90.9 % monomeric, and the 4.3 % dimeric and 4.0 % trimeric entries are exactly the ones worth opening. Provenance is reported as three counts and deliberately **not** as an agreement rate, because `author_defined_assembly` says only that the depositor stated one: PISA may have returned nothing or never run. Scoring that as disagreement would have invented a conflict on 758 of those 1,686 entries. The ambiguity that is real, an entry whose own assemblies disagree about the chain count, is 19 of them, and they are listed by name rather than reduced to a rate. Two things the data forced: oligomeric states group on the **chain count**, never on `oligomeric_details`, whose capitalisation is not consistent across the archive and which listed \"trimeric\" (64) and \"Trimeric\" (3) as two different states in the same table; and buried interface area is quoted as quartiles rather than a range, because one 24-mer puts the maximum three orders of magnitude above the median (1,476 \u00c5\u00b2 against 615,514 \u00c5\u00b2)\n- [x] **The dossier becomes the reference document it was meant to be.** Every figure the app draws, drawn again for a file that fetches nothing: domain architecture and construct coverage as inline SVG, the RCSB's own ligand depictions and a rendered still of the reference structure embedded as `data:` URIs, and the seed sequence numbered every ten with sites, PTMs, disulphides, transmembrane spans and the fifteen most-substituted positions marked. KPIs on one row, each linking into the live app, and a cross-reference bar out to UniProt, RCSB, PDBe, PDBe-KB, AlphaFold, InterPro, Pfam, CATH and SCOP, the last two by the identifiers the Domains panel already holds rather than by a search. The self-containment test was corrected while doing it: it forbade `<img` and `url(` outright, which also forbade `data:` URIs and inline figures, and those are part of the file and open offline in ten years, which is the whole requirement. It now checks that nothing is *fetched*\n- [x] **Dossier export.** `/f/<slug>/dossier`: one file, branded, suitable for a project kickoff or a grant appendix. Constructs, the positions people deliberately mutate, assembly, domains, ligands, crystallisation, validation, orthologues, the seed sequence and one record per primary paper. Self-contained in the strict sense and tested for it, because the point of the thing is to outlive the session it came from: no script, no stylesheet, no font, no image and no `url()` anywhere, so it still works opened from a mailbox years later with nothing behind it. PDF is the browser's own print-to-PDF rather than a server-side renderer, which keeps WeasyPrint and its Cairo stack off a droplet with two cores, and the print CSS earns that choice: `@page` margins, no row split across a page break, no heading stranded at a page foot. It renders from a family that is already filed and never assembles one on demand, since assembly takes up to ninety seconds and this is exactly the URL a link checker fetches. Two things it taught: an undated paper crashed Jinja's `sort` filter comparing `None` to `int`, so citations are ordered in Python; and the line reporting \"1,606 carry a deliberate mutation\" was wrong, because `is_engineered` means *differs from the UniProt canonical* and lysozyme's most-used construct is the mature protein after its signal peptide, so the document now says which it means and points at the construct column instead\n- [x] **Shareable permalinks with the filter state encoded in the URL.** What a reader is looking at is a filter state, not a page, so somebody who narrows a 1,988-entry family to the ligand-bound X-ray structures of one TM cluster previously had nothing to send a colleague but the family URL and a set of verbal instructions. Identity, methods, the orthologue/fusion/holo toggles, the view, the open entry and the highlighted construct all live in the hash, and only non-defaults are written, so the common case stays a clean `/f/<slug>` and the URL grows in proportion to what the reader actually did. The cluster is the one piece that cannot be stored literally, being a set of up to 260 construct ids: it is stored as the cut height plus a single member and rebuilt by re-clustering at that height, which also degrades honestly, since a link made before an embedding was rebuilt restores a cluster from the matrix that exists now rather than asserting a stale one. `replaceState`, not `pushState`, because the identity slider fires on every step and a history entry per step makes the back button useless for leaving the page. Verified by round trip in headless Chrome rather than by inspection: filter, read the URL, cold-load it in a second tab, and compare the *derived* entity count, which is the thing that would disagree if any control had been restored in name only\n- [x] **Performance and hardening.** The batched fetches now run concurrently (six at a time: these are free public APIs, so the aim is to stop wasting round-trip latency rather than to hammer them) and UniProt's entry and features come in one request rather than two. Spike went from 75 s to 32 s cold, and warm pages are unchanged at 0.06-0.7 s. A `warm` command plus a weekly systemd timer, firing a few hours after the PDB's Wednesday release, so a reader never pays for a stale cache. `warm` also folds in the structural artefacts: on a workstation it rebuilds any embedding whose pipeline version has moved, and with `--contacts` any PLIP fingerprint too (opt-in, because PLIP is minutes per family and a weekly run that quietly grew to an hour is a weekly run somebody turns off). Everywhere, including the droplet which cannot build them, it reports how many families are still showing the placeholder map. That number is on `/api/stats` too. Three artefact version bumps in one afternoon each silently invalidated every artefact, and the page renders either way: the map just quietly stops being a measurement. Every optional panel degrades on its own rather than taking the family down with it\n- [x] **The map stops drawing things that are not there, and starts naming the things that are.** Four changes to one panel, each prompted by looking at it rather than at the code. **The spirals were decoration.** Entries sharing a construct share an MDS coordinate exactly, and were fanned around it by a golden-angle offset of `0.012\u00b7\u221ak` with nothing bounding `k`: that is a sunflower packing, and past a couple of dozen members it draws its own parastichies, so a dense family showed radiating \"arms\" that were phyllotaxis rather than structure. On ABL1 the fan spanned 0.281 against a median representative-to-representative distance of 0.104, meaning the artefact was 2.7 times the signal and two nodes on opposite rims of one disc were the same sequence. A crowded point is now one ringed node scaled by what it holds, opening on click, and the fan that survives below the threshold is bounded by a share of the distance to the nearest other representative. **Two thirds of ABL1's nodes were guesses drawn as measurements**: a member whose construct was not among the representatives inherits the position of the nearest one by identity, and since hundreds of representatives sit at 100% identity the tie fell to sort order and dumped 126 of them onto a single point, which is most of what made the largest discs. Ties are now spread rather than concentrated, and an inherited position is drawn hollow and counted on the panel. **Clusters are named**, from the same average linkage at the same cut height the Structures heatmap uses, after the protein that owns them where one accession holds half the cluster and after the Pfam domain its members share otherwise: ABL1's largest cluster is 1,932 entities whose commonest accession is EGFR at 17%, so naming it \"Epidermal growth factor receptor\" asserts something false about 83% of it, while \"protein tyrosine kinase\" is true of 95% and is the reason those structures cluster at all. **And one structure can no longer set the scale for a whole family**, which is what the lysozyme worked example above is about. Verified in a real browser through CDP mouse events rather than by dispatching synthetic clicks at elements, which is how the last round of map interaction bugs was missed: a synthetic click bypasses the browser's own hit-testing and cannot fail\n- [x] **The contacts panel was numbering residues that do not exist.** `seed_pos = resnr + (query_beg - 1)` added the seed offset to PLIP's *author* residue number, which in a well-annotated entry already carries it, so it was counted twice: JAK1 (query_beg 879, seed 1,154 residues) reported hot residues at 1,340 and 2,110. **52 of 71 built families were affected.** It was invisible on carbonic anhydrase, the family the panel was validated on and this README quotes, because its query_beg is 1 and the offset is zero; even there the numbers were only accidentally plausible, since the panel printed THR200 and position 200 of that seed is a *proline*. It now maps by aligning each of the member's own chains to the seed, the way `topology.map_to_seed` already did and says why in its docstring. Which chains belong to the member is **read** from the entity record rather than inferred, because inference does not work here: a co-crystallised 78-residue partner aligns to carbonic anhydrase at 37.8 % over 74 columns, inside the twilight zone and above any threshold that still admits a legitimate 35 %-identity orthologue. All 71 artefacts rebuilt at VERSION 4, which cannot be migrated because the raw residue numbers are not kept and the double offset cannot be undone after the fact\n- [x] **Provenance, and a JSON API.** Both from data already fetched for other panels: entity descriptions, the citation records the BibTeX export deduplicates, and deposit dates. The names panel is partitioned by accession first, which is the whole correctness of it, since counting description strings across a family assembled at 30 % identity reported `ALPHA-LACTALBUMIN` as an unrecognised name for hen lysozyme when it is the correct name of a relative\n- [x] **Conservation on the surface**, through MolViewSpec: this Mol\\* build exposes no MolScript, no overpaint helper and no theme registration, so the usual routes are unavailable. Finding that also turned up `focusResidue`, which built its selection with `lib.structure.Script` and had therefore **never worked**: every call threw and the `try/catch` returned a bare `false`, so the Contacts panel had been inviting readers to click a residue and focus it for as long as it existed\n- [x] **Curated sites on a PDB-seeded family.** UniProt numbers on the canonical and a PDB-seeded family's seed is typically the mature protein, so features were skipped entirely rather than placed wrongly: `lysozyme-1aki-1` showed no active site and no disulphides while `lysozyme-c-p00698`, the same protein, showed both. Now aligned onto the seed, giving Glu35 and Asp52, hen lysozyme's catalytic pair in mature numbering, and the four disulphides at C6, C30, C64 and C76. A feature only partly inside the seed is dropped whole: half a signal peptide is not a shorter signal peptide\n- [x] **`VEGFR` answered with twelve fruit flies.** No reviewed entry carries VEGFR as a gene name, because the human genes are KDR, FLT1 and FLT4 while Drosophila's Pvr lists it as a synonym, so the gene query matched nothing reviewed and fell straight through to the unreviewed tail. The protein *name* is now searched too when the gene finds nothing reviewed, and candidates are ranked human-first inside the reviewed group. The first fix covered only `by_gene` and left the free-text path alone, which is the same complaint one input away: ranking now lives in `search` so every caller gets it\n- [x] **All compute moved to the droplet; the Mac builds nothing.** The two-machine arrangement above worked and had one fatal property: `drain_queue.sh` ran on a laptop, so \"this family has been queued for it\" meant \"when Marc's Mac next wakes up\", which for a Mac that is off for a week means never. The reason it could not move was memory, and it was a single measured thing: parsing a whole deposited assembly to pull out one chain's alpha carbons peaked at **6.3 GB RSS** (8GLV is 453 MB of text and yields 426 alpha carbons) on a box with 2.2 GB free and no swap, so the failure would not have been slowness but an OOM taking the other eight apps with it. gemmi was tried and is twelve times faster on the same file while still needing about six gigabytes: the parser was never the problem. Fetching the single chain from the **RCSB Model Server** returns 533 kB instead, an 850-fold reduction, and peaks at **102 MB**; it is also *faster* than the static file at about 0.5 s per structure, a first measurement to the contrary having been taken on the 453 MB monster itself. `embed.structure_path()` is now the one place anything asks for a structure file, which it had to become: moving the fetch silently broke topology, because that read `_cif_path()` and depended on `ca_trace` having downloaded the whole entry as a *side effect*, and a missing file was already a legitimate \"no 2D layout for this one\" so the failure wore the costume of a normal answer. Contacts is the one artefact that still needs whole structures, since interaction detection is every atom and every chain, so it caps the file size, **counts what it skipped apart from what failed**, and says so on the panel: a failure is a bug to chase and a skip is a decision this pipeline made, and a number that adds them together is one nobody can act on. The cap aborts the transfer rather than checking afterwards, because a HEAD on files.rcsb.org times out and finishing a 453 MB download to then decline it costs 453 MB of a disk that eight apps share. Proven end to end on the droplet: VEGFR-1, 1,328 entries, embedded in 1h42m at 102 MB peak with free memory never dipping below 2.2 GB, then its topology by P-SEA in one second (`mkdssp` ships with CCP4 and is not pip-installable, and the artefact records which method drew it), then PLIP. `deploy/sync.sh`, `drain_queue.sh`, `push_embeddings.sh`, `overnight.sh` and the launchd timer are all deleted\n- [x] **Ship-out.** Icon, blog post on marcdeller.com, and a link from the mdeller.com landing page. Owned and done by Marc, outside this repository", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\nEvery other score on the dashboard is Boltz grading its own work. This one is not: it compares\nthe docked ligand with the same molecule in an experimentally determined structure, and it is\nthe only panel that can tell you the prediction is wrong. On a real GLP1R campaign, orforglipron\nwas predicted into the receptor's own transmembrane site with `ligand_iptm` 0.940 and a\nconfidence score of 0.836 -- comfortably green on both icons -- yet superposed on 7E14, the\ncrystal structure of that same complex, the predicted ligand sat **9.4 A** from where the\nexperiment puts it: correct pocket, correct conformation, rotated end for end. No metric Boltz\nemits noticed, because none of them is a comparison with reality.\n\n**How to use it.** Drop experimental mmCIF files into a `reference/` folder next to\n`boltz_input.md` -- the same files you downloaded for `Pocket contact:` and `Apo structure:`,\nnothing else to configure. The panel then appears in `boltz_dashboard.html` and in the hosted\nAnalysis view; without `reference/` it stays silent rather than showing an empty card.\n\n**Three numbers, because they fail independently.** **Site (A)** is the distance between the two\nligand centroids after superposing the receptor: did it find the pocket at all. **Pose (A)** is\nthe symmetry-corrected RMSD in place, after that same superposition: did it find the binding\nmode. **Conformer (A)** is the symmetry-corrected RMSD with the two ligands superposed on each\nother: is the molecule's own shape right.\n\nThe orforglipron case reads 3.05 / 9.43 / 3.08, legible at a glance as \"right pocket, right\nshape, wrong orientation\"; collapsed into one figure it would be legible as nothing. Two icons\ncarry the summary in the dashboard's own visual language -- a **bullseye** for the site, a **pose\nmark** for the binding mode -- each tinted green under 2 A, amber to 5 A, red beyond, 2 A being\nthe long-standing convention for \"this reproduces the crystal structure\".\n\n**Atoms are paired by molecular graph, not by proximity.** A 65-atom drug has equivalent methyls\nand flippable rings; pairing each predicted atom with whichever experimental atom is nearest\nquietly flatters a wrong pose by matching it to itself. RDKit assigns bond orders from the\ncampaign's own SMILES and enumerates the substructure matches, so a symmetry-equivalent\nplacement scores as correct and a genuinely different one does not.\n\n**Which experimental ligand a target is compared against** is worked out rather than configured.\nA constrained target names its pocket after the ligand it was derived from (`GLP1R_ORFO_V6G` ->\nV6G), so that is the answer. An unconstrained target -- the baseline, and the comparison most\nworth having -- is matched by heavy-atom element composition against every ligand in\n`reference/`, so the same molecule is found wherever it sits.\n\n**A construct numbered differently from its reference still compares.** The receptor\nsuperposition matches residues by number and requires the residue types to agree, which is the\nright correspondence when both sides number the same protein the same way -- and is not the case\nnearly as often as it sounds. Boltz numbers a prediction from 1, while a PDB entry numbers the\nconstruct it holds, so a trimmed ABL1 kinase domain came out 1-272 against references numbered\n223-500, shared not one residue number with either, and every pose in the campaign went\nuncompared behind the message \"2GQG is not a structure of ABL1\". A trim is a contiguous slice, so\none integer reconciles the two, and it is found rather than assumed: every offset that could\nbring the ranges into contact is scored by how many residues then agree in type, and the best\nwins. Matching numbering still gives offset zero, because that is what matches most. Nothing is\ncompared where the residues do not correspond, which is the property the bare number-matching\nhad and the reason an alignment was not used instead.\n\n**One small viewer per pair, grouped by pocket.** The table says a pose is 9.4 A wrong; it cannot\nsay *how*. \"Rotated end for end\" is obvious in one glance at two overlaid ligands and invisible\nin a column of angstroms, so the hosted Analysis view draws each comparison as its own frame\nholding exactly two ligands as sticks -- the prediction in red, the experimental one in grey --\nsuperposed through their receptors and zoomed to fit, with spin and reset. Tiles are grouped\nunder the pocket they were run against, unconstrained last, so the comparison the matrix exists\nfor reads down one column. On the GLP1R campaign the V6G tiles show two molecules interleaved in\none volume pointing different ways, and the unconstrained tiles show them in different places\nentirely: the two failure modes, told apart without reading a number.\n\nEach pair is written at analyze time as two single-ligand mmCIFs (`boltz_pose_pairs/`, carried in\nthe `.bmz`), because the superposition needs RDKit, the reference files and the predicted\ncomplex, none of which are on the server -- so the viewer draws exactly the coordinates the\nnumbers were measured from, not a second superposition that could disagree with the table above\nit. **Viewers are pooled, not one per tile:** a browser allows a limited number of live WebGL\ncontexts and exceeding it does not raise, it silently kills the oldest, turning early frames\nblack while later ones look fine. Frames are created as they scroll into view and disposed\nleast-recently-seen-first, so thirty comparisons cost the same as four.", "# PETase Enzyme Therapy\u2014A New Frontier in Medical Microplastic Removal\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-15 on marcdeller.com (categories: blog, microplastics).*\n\n* PETase variants efficiently degrade PET microplastics at human body temperature.\n* Enzyme therapy with PETase is proven non-toxic to human cells.\n* Encapsulation and surface display boost targeting, stability, and circulation.\n* PETase-based therapeutics promise actual molecular breakdown of microplastics.\n* Biocatalytic agents could eclipse filtration for long-term microplastic management.\n\n1. [PETase degradation in serum (2024)](https://pubs.rsc.org/en/content/articlepdf/2024/fd/d4fd00014e)\n2. [Current advances in PETase engineering (2023)](https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2023.1263996/full)\n3. [Enhancement via FAST-PETase (2023)](https://pubs.acs.org/doi/10.1021/jacs.3c04427)\n4. [Advances in cytotoxic enzyme delivery (2022)](https://pmc.ncbi.nlm.nih.gov/articles/PMC9230553/)\n5. [Plastic-eating enzymes in waste management (2022)](https://molecularbiosci.utexas.edu/news/research/plastic-eating-enzyme-could-eliminate-billions-tons-landfill-waste)\n\nThe medical application of PETase, particularly engineered variants like S238A, FAST-PETase, and DuraPETase, is gaining traction for its ability to degrade PET microplastics efficiently at body temperature. Safety studies show PETase does not compromise human cell viability, and its monomer byproducts (TPA and MHET) are quickly excreted with minimal toxicity.\n\nAdvanced delivery systems\u2014such as enzyme encapsulation in erythrocytes or surface display techniques\u2014increase PETase circulation time and tissue targeting. Compared to physical filtration, enzyme therapeutics promise true biodegradation rather than mere sequestration or removal. As mutational and design strategies progress, enzyme therapeutics may become integral in reducing microplastic load from the human body.", "# My Work Driving Innovation in Precision Oncology Through Structure-Based Drug Design\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n**CDK2 Inhibitor: Precision Targeting in Ovarian Cancer**\n\nCyclin-dependent kinase 2 emerges as a critical therapeutic vulnerability in ovarian cancers exhibiting [cyclin E1 overexpression](https://marcdeller.com/discovery-of-4-pyrazolyl-2-aminopyrimidines-as-potent-and-selective-cdk2-inhibitors/), a biomarker present in a substantial subset of platinum-resistant cases where treatment options remain severely limited. The development of [4-pyrazolyl-2-aminopyrimidines](https://marcdeller.com/discovery-of-4-pyrazolyl-2-aminopyrimidines-as-potent-and-selective-cdk2-inhibitors/) as potent and selective CDK2 inhibitors represents a structure-guided approach exploiting the synthetic lethal relationship between CDK2 inhibition and cyclin E1 overexpression in cancer cells that become dependent on CDK2 for survival. [INCB123667](https://clinicaltrials.gov/study/NCT07023627), the clinical candidate from this chemical series, demonstrated compelling proof-of-concept efficacy in heavily pretreated patients with platinum-resistant or refractory ovarian cancer, achieving a 33.3% objective response rate with 100 mg daily dosing in a [Phase 1 study](https://www.incytemi.com/document/Poster/ASCO%202025%20-%20INCB123667%20(CDK2%20inhibitor)%20in%20Ovarian%20Cancer%20(oral%20presentation).pdf), with all but one responder demonstrating cyclin E1 overexpression, validating the precision biomarker approach. The median progression-free survival of 5.3 months and duration of response of 3.6 months, coupled with over 70% of patients experiencing tumor shrinkage, represents meaningful clinical benefit in a population that typically has exhausted multiple prior therapies including bevacizumab and PARP inhibitors. Importantly, the manageable safety profile with predominantly grade 2 or lower hematologic and gastrointestinal treatment-emergent adverse events, and only 2.2% discontinuation due to toxicity, supports the advancement into the [MAESTRA-2 pivotal trial](https://clinicaltrials.gov/study/NCT07023627)testing INCB123667 versus chemotherapy in biomarker-selected platinum-resistant ovarian cancer patients.\u200b\n\n**FGFR2/3 Inhibitor: Overcoming Gatekeeper Mutations**\n\nFibroblast growth factor receptor alterations, including activating mutations, amplifications, and gene fusions in [FGFR2 and FGFR3](https://marcdeller.com/discovery-of-potent-and-selective-inhibitors-of-wild-type-and-gatekeeper-mutant-fgfr2-3/), drive oncogenesis across cholangiocarcinoma, bladder cancer, and other solid tumors, but therapeutic efficacy of first-generation FGFR inhibitors remains limited by acquired resistance through gatekeeper mutations such as V564F in FGFR2 and V555M in FGFR3 that create steric hindrance preventing drug binding. Structure-based design approaches enabled the [discovery of potent and selective inhibitors](https://marcdeller.com/discovery-of-potent-and-selective-inhibitors-of-wild-type-and-gatekeeper-mutant-fgfr2-3/) capable of overcoming both wild-type and gatekeeper mutant FGFR2/3, addressing a critical clinical need as approximately 36% of all FGFR2 alterations consist of molecular brake or gatekeeper mutations, enriched to 51% and 57% in cholangiocarcinoma and breast cancer respectively. These next-generation compounds maintain potent inhibitory activity against resistance mutations by exploiting alternative binding modes that avoid direct steric clashes with the enlarged gatekeeper residue, validated through crystallographic analysis and biochemical profiling. The development strategy incorporated targeted protein degradation approaches, with FGFR2-selective degraders achieving DC50 values of 0.645 nM and demonstrating superiority over parental inhibitors against the FGFR2-V564F gatekeeper mutant with IC50 of 0.121 nM, providing a comprehensive toolkit to combat resistance mechanisms. This structure-informed precision medicine approach addresses the polyclonal resistance observed clinically, where patients develop multiple concurrent FGFR2 mutations following first-generation inhibitor therapy, enabling broader and more durable disease control.\u200b\n\n**Mutant Calreticulin Antibody: Targeting Myeloproliferative Neoplasms**", "# Marc Deller's vibe-coding portfolio: AI-assisted scientific apps and tools\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar, mdeller.com, marcdeller.com, 2026\n\nMarc Deller describes himself as a \"vibe-coder\": a subject-matter expert in structural biology and drug discovery who uses AI-assisted coding (Claude Code, Cowork and custom skills) to build scientific tools quickly, rather than a professional software engineer. His code lives at https://github.com/bellcheddar, his web apps run on subdomains of his own server https://mdeller.com (a single small DigitalOcean droplet with nginx, certbot and Flask/gunicorn), and companion write-ups appear on his blog at https://marcdeller.com. His local development and model-training machine is a 64 GB Apple M1 Mac, on which he fine-tunes local language models with Apple's MLX-LM.\n\nA consistent design system runs through everything he builds: navy #1C244B and accent blue #467FF7, the Poppins typeface, British English, no em dashes (colons or parentheses instead), a strict list of forbidden hype words (groundbreaking, revolutionary, paradigm-shifting and the like), and a matching set of flat, navy-outlined cartoon icons with a Baloo 2 wordmark. He also loves a British-slang backronym for a project name.\n\n## AlphaFraud (alphafraud.mdeller.com)\n\nAlphaFraud is a live web app and GitHub repository (github.com/bellcheddar/AlphaFraud) that audits AlphaFold's blind predictions against freshly deposited human structures in the PDB. The logic: AlphaFold2 was trained on PDB chains released before 30 April 2018, so any human structure deposited after that date is a real blind test of the claim that \"AlphaFold has solved protein folding\". A weekly pipeline (a systemd timer) pulls each new human entity from the RCSB PDB, matches it to its AlphaFold DB model at the EBI, superposes the two, and scores their disagreement with TM-score, lDDT, GDT, RMSD, per-residue deviation, a PAE-honesty metric and a novelty score (maximum percent identity to any pre-cutoff chain). A prediction is flagged \"confidently wrong\" when mean pLDDT exceeds 70 but the TM-score is below 0.5, and a composite FRAUD score weights error by confidence. The full post-cutoff backfill finished on 18 July 2026: 96,827 human entities processed, 82,796 analysed, 1,941 confidently wrong, of which 431 were sequence-novel. The worst offenders are dominated by amyloids (transthyretin, alpha-synuclein with 210 depositions, TMEM106B, beta-2-microglobulin) and conformational switches such as STIM1 that AlphaFold predicts as confidently folded. Tabs: Home, All Structures (a fraud-quadrant scatter), Worst Offenders leaderboard, Examples (including an Artemis positive control), Calculate (on-demand, no LLM), Analysis (fold enrichment and PCA), Archive and Stats. The stack is Python, biotite, tmtools, numpy, scipy, Plotly and SQLite, with no cluster and no external database. Marc's point in building it was to show what a subject-matter expert can stand up on modest hardware at minimal cost.\n\n## BoltzMaker (github.com/bellcheddar/BoltzMaker)\n\nBoltzMaker is a wrapper around Boltz-2 for campaign-scale protein structure and binding-affinity prediction. It takes a human-readable campaign file (boltz_input.md) describing proteins, partners and ligands (SMILES or CCD codes) with constraints, generates the YAML manifests Boltz-2 needs, runs the campaign, and performs post-run analysis including PLIP interaction profiling, secondary-structure comparison, and crash recovery for Apple MPS runs. It is packaged with pixi, and a self-contained macOS installer and an on-device narration sidecar (using Apple's Foundation Models framework on Apple Silicon) are planned. The name is a nod to Timothy Taylor's Boltmaker beer. Marc has run Boltz-2 campaigns with it on GLP-1R and on the 5-HT2A/2B/2C serotonin receptor Gq panel, and has written a full blog post with architecture diagrams and CLI reference tables.\n\n## chatPDB (chatpdb.mdeller.com)\n\nchatPDB (github.com/bellcheddar/chatPDB) is a protein-structure AI assistant for the Protein Data Bank, one of Marc's vibe-coding web apps hosted on mdeller.com.\n\n## CODSWALLOP (codswallop.mdeller.com)\n\nCODSWALLOP is a family-centric PDB mining web app (Flask) that takes a protein family and returns every structure in it with a full cross-referenced summary: interactive tables and plots, Mol* viewers, PLIP interaction analysis, family alignment viewers, sequences, crystallisation conditions, ligands, interacting residues, domains, folds, topology diagrams, counts, resolution, expression systems, constructs and protein tags. It is positioned as a richer, more interactive successor to the PDB landing pages and the old PDB summary pages. The backronym is Comparative Overview of Domains, Structures, Whole-families, Alignments, Ligands, Lineages, Organisms & PDB entries (British slang for nonsense, chosen ironically for a rigorous tool). Its crystallisation output is intended to feed a planned crystallisation-condition predictor, and its AlphaFold overlay ties in to AlphaFraud.\n\n## GOBSMACKED (gobsmacked.mdeller.com)", "# 2025 \u2013 Serverless Architectures on AWS\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-27 on marcdeller.com (categories: certificates).*\n\n* [Link to certificate](https://www.coursera.org/account/accomplishments/verify/X9AD2JWH2HE9)\n\n* **Award Details:** Amazon Web Services, through Coursera, issued this certificate on March 10, 2025.\n* **Course Focus:** It recognizes completion of a course on serverless architecture design in AWS.\n* **Career Value:** This matters because scalable cloud infrastructure is increasingly important for scientific software and shared tools.\n* **Practical Use:** It helps with deploying lightweight web apps, services, and online research resources.", "# Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor\nAuthors: Kong, L.; Ju, B.; Chen, Yajing; He, L.; Ren, L.; Liu, J.; Hong, K.; Su, B.; Wang, Z.; Ozorowski, G.; Ji, X.; Hua, Y.; Chen, Yanli; Deller, M.C.; Hao, Y.; Feng, Y.; Garces, F.; Wilson, R.; Dai, K.; O'Dell, S.; McKee, K.; Mascola, J.R.; Ward, A.B.; Wyatt, R.T.; Li, Y.; Wilson, I.A.; Zhu, J.; Shao, Y.\nSource: Immunity, 2016\n\nDRVI01 sera displayed strong CD4bs specificity and some V1V2 (variable regions 1 and 2) reactivity. We isolated CD4bs-directed mAbs from a 2009 sample (Figure 1B). A CD4bs-specific probe pair, RSC3 and RSC3\u0394371I, was used to identify antibody-expressing B cells targeting the CD4bs by fluorescence-activated cell sorting (FACS) (Wu et al., 2010). In brief, peripheral blood mononuclear cells (PBMCs) from DRVI01 were incubated with cell markers and two RSC3 probes. CD4bs-specific memory B cells with the phenotype CD19+ CD20+ CD14\u2013 CD3\u2013 CD8\u2013 IgG+ RSC3+ RSC3\u0394371I\u2013 were sorted into 96-well microtiter plates. Of seven sorted memory B cells, five were amplified with matching heavy and light chain genes (Figure 1B). After cloning into the IgG1 vector, the full IgG mAbs were expressed transiently in 293F cells. Of the five isolated mAbs, DRVIA1, 3, 4, and 6b showed strong RSC3 binding in ELISA, whereas DRVIA7 bound weakly to RSC3 as characterized by Bio-Layer Interferometry (Figure S1A), and none of the mAbs were reactive to RSC3\u0394371I. We further confirmed their CD4bs specificity using a YU2 gp140-foldon (gp140-F) trimer with the D368R mutation (Figure S1B; Yang et al., 2002). Compared to DRVIA1-6b, DRVIA7 exhibited greater neutralizing ability but had limited breadth (Figure S1C). Sequence analysis revealed a clonal family consisting of DRVIA1, 3, 4, and 6b, and an outlier DRVIA7 (Figure S1D). The DRVIA1-6b clonal family was characterized by a low level (3.1%\u20136.3%) of somatic hypermutation (SHM), a 19-aa heavy chain CDR3 (HCDR3) loop (Kabat numbering), and a 9-aa \u03ba chain CDR3 loop. By contrast, DRVIA7 displayed VRC01-like sequence features: SHM of 19.0% for the VH gene and 17% for the V\u03ba gene, a heavy chain of IgHV1-02*02 allelic origin, a relatively short HCDR3 loop (11 aa), and a distinctive 5-aa \u03ba chain CDR3 loop containing a \"F91E96\" motif (\"Y91E96\" in VRC01) (Zhou et al., 2013). These findings suggested that DRVIA7 might be a VRC01-class antibody that had acquired a subset of VRC01 signatures but not yet broad neutralizing activity, thus providing an opportunity to study the emergence of VRC01-class antibodies.\n\n*Figure 1. Identification of a VRC01-Class Antibody DRVIA7 from HIV-1-Infected Chinese Donor DRVI01.* (A) The neutralizing ability of DRVI01 sera from six time points from 2005 to 2009 were tested against the global and DRVI panels of pseudoviruses. Dendrograms are shown at each time point representing the pseudovirus panels with nodes representing individual viruses colored according to neutralizing potency. (B) Single B cell sorting of CD4bs-directed monoclonal antibodies by flow cytometry. Around 10 million PBMCs from DRVI01 were incubated with cell markers and sorting probes. Seven B cells that reacted with RSC3, but not RSC3\u0394371I, were then sorted into 96-well microtiter plates. Abbreviations are as follows: FITC, fluorescein isothiocyanate; FSC-H, forward scatter height; FSC-A, forward scatter area; and SSC, side scatter area. (C) Sequence analysis of DRVIA7 and VRC01 heavy and light chains with alignment to respective germline genes. Mature antibody residues that differ from germline are colored in red. See also Table S1 and Figure S1.\n\n### Structural Differences between DRVIA7 and VRC01-Class Antibody Interactions with gp120 and gp140 Trimer Primarily with the Glycan Shield\n\nThe limited breadth and potency suggested that some important features present in the mature VRC01 were missing in DRVIA7 (Figure S1C). To elucidate these features, we determined the crystal structure of unliganded DRVIA7 Fab at 2.9 \u00c5 resolution (Table S2), which resembled VRC01 and other bNAbs of this class such as VRC03, PGV04, and 12A21 with a C\u03b1 root mean square deviation (RMSD) of <1.0 \u00c5 over the antibody variable regions (VH and VL) (Figures S2A and S2B; Zhou et al., 2013). Using an HIV-1 gp120 from the 93TH057 strain, we next determined the crystal structure of a gp120 core\u2013DRVIA7 Fab complex at 3.4 \u00c5 resolution (Figure 2A; Table S2). The gp120 was deglycosylated with endoglycosidase H (EndoH), but only after complex formation to preserve glycan interactions within the antibody-gp120 interface. Overall, DRVIA7 buried 957 \u00c5\u00b2 of solvent-accessible gp120 protein surface overlapping the CD4bs. Negative-stain EM and crystal structures confirmed that DRVIA7 bound to the SOSIP gp140 trimer with a similar angle of approach to VRC01 and nearly identical interactions with gp120 (Figures 2B\u20132D and S2C\u2013S2G; Derking et al., 2015; Lyumkis et al., 2013; Zhou et al., 2010). For example, 8 of 11 HCDR2 residues and 3 of 4 HCDR3 residues on VRC01 and DRVIA7 that contacted gp120 were conserved between the two antibodies, while their LCDR3 loops adopted identical conformations and contacted a similar set of gp120 residues (Figure S2C). Nevertheless, DRVIA7 and VRC01 heavy chains could still differ in their sensitivity to gp120 mutations. To examine this possibility, we generated a chimeric antibody with DRVIA7 heavy chain and mature VRC01 light chain (DRVIA7H-VRC01L). Indeed, the H54 of DRVIA7 heavy chain made this chimeric antibody less tolerant of the G472A mutation of gp120 than VRC01 (Figures S2H and S2I). However, the substantial differences in their neutralizing ability (Figure S1C) could not be explained by these primary protein-protein interactions.", "# FGFR2 Kinase Domain in Complex with a Pyrazolo[1,5-a]pyrimidine Analog (Compound 29)\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [8E1X](https://www.rcsb.org/structure/8E1X)\n* **Release Date:** 2022-11-23\n* **Method:** X-RAY DIFFRACTION 2.68 \u00c5\n* **Organisms:** Homo sapiens\n* **Macromolecules:** Fibroblast growth factor receptor 2 (protein)\n* **Unique Ligands:** U9P (Compound 29)\n* **Publication:** [Shvartsbart, A., et al. (2022). Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3. Journal of Medicinal Chemistry, 65, 15433-15442.](https://pubs.acs.org/doi/10.1021/acs.jmedchem.2c01366)\n* **Blog Post:** [Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant FGFR2/3](https://marcdeller.com/discovery-of-potent-and-selective-inhibitors-of-wild-type-and-gatekeeper-mutant-fgfr2-3/)\n\n**Description:** This structure elucidates the binding mode of Compound 29, a potent and selective FGFR2/3 inhibitor that retains low-nanomolar potency against common gatekeeper mutants. The structure was instrumental in guiding SAR optimization to achieve excellent selectivity over FGFR1, addressing treatment-related resistance mechanisms in cholangiocarcinoma and bladder cancer.", "# Protein stability: a crystallographer's perspective\nAuthors: Deller, M. C.; Kong, L.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 72, 72\u201395 (IYCr crystallization series), 2016\n\n### 3.4. Quaternary structure\n\nQuaternary structure \u2014 the arrangement of folded subunits into multi-subunit complexes \u2014 underpins allostery and cooperativity. The classic Monod\u2013Wyman\u2013Changeux model posits tense (T) and relaxed (R) states; ligands may bind either with equal affinity, but preference for R raises affinity while preference for T lowers it, defining an allosteric modulator. Haemoglobin is the best-studied example, with the R state representing deoxyhaemoglobin and the T state oxyhaemoglobin. Because many proteins undergo large conformational changes involving both secondary and tertiary structure, and each state can have different compositional or conformational stability, stability cannot be considered in isolation from function. Modulating protein\u2013protein interactions with small molecules has long been a pharmaceutical goal but is difficult given the large interface areas; destabilizers are more common than stabilizers. Cited destabilizers are Abbott's navitoclax, which disrupts Bcl-2 binding to Bad/Bid/Bak, and Roche's nutlin-3, which blocks p53\u2013MDM2. Stabilizers include cyclosporin A, which stabilizes the calcineurin\u2013cyclophilin interaction, and tafamidis, which binds a pocket at the transthyretin dimer interface and prevents the aggregation and misfolding that cause transthyretin amyloidoses such as peripheral neuropathy (Figure 2).\n\n### 4. Protein stability during protein expression and purification\n\nObtaining intact, stable, folded protein in sufficient quantity for crystallization screening, formulation, vaccine development or therapeutic use is frequently the limiting step, and many proteins are unstable, unfolded or proteolysed under expression conditions \u2014 degrading both conformational and compositional stability. Contributing factors include the primary structure of the target, the design of the recombinant expression plasmid, temperature, expression medium and toxicity to the host organism (Table 3). Because so many variables interact, combinatorial design approaches combined with high-throughput methods are commonly used to find workable combinations, and the authors point readers to S\u00f8rensen & Mortensen for advanced genetic engineering of cells and plasmids. This section is organized around three levers the experimenter controls: construct design (\u00a74.1), expression conditions (\u00a74.2) and host-cell choice (\u00a74.3).\n\n### 4.1. Construct design, sequence, expression tags and protein stability\n\nConstruct design is presented as one of the primary decisions a structural biologist makes. Protease cleavage sites within a target invite attack by endogenous host proteases and can be removed during recombinant assembly; the authors cite the removal of two KEX2 sites from a malarial vaccine candidate, which enabled large-scale production of full-length protein. Some primary structures are inherently short-lived: sequences rich in proline, glutamate, serine and threonine (PEST sequences) often confer half-lives under 2 h, targeting proteins for proteasomal degradation or calpain-mediated proteolysis. The N-end rule is a strong predictor of in vivo half-life \u2014 N-terminal methionine, serine, alanine, glycine, threonine, valine or proline stabilizes (>20 h), whereas phenylalanine, aspartate, lysine or arginine destabilizes (<3 min) via ubiquitination \u2014 though this is irrelevant in bacterial systems and in practice the N-terminal residue is usually masked by an N-terminal affinity tag. Tags also improve solubility, most notably maltose-binding protein (MBP), which promotes correct folding and behaves like a chaperone; DnaK and GroEL co-expression serves a similar purpose. The pMal malE signal sequence can direct proteins to the periplasm, away from cytoplasmic proteases. Self-cleaving tags \u2014 intein-based, or nickel-ion cleavage at the (S/T)XHZ motif used for GmSPI-2 \u2014 avoid the nonspecific damage caused by thrombin or factor Xa.\n\n### 4.2. Expression conditions and protein stability\n\nThe host must be supplied with the prosthetic groups, cofactors and ligands the target requires. Many are scavenged from the medium, but some cannot be synthesized by the host or are present in insufficient quantity: heme incorporation in *Escherichia coli* is often low unless the medium is supplemented with \u03b4-aminolevulinate. Solubility during expression can be improved by additives including trehalose, glycine betaine, mannitol, L-arginine, potassium citrate, CuCl2, proline, xylitol, NDSB 201, CTAB and K2PO4, as catalogued by Leibly and coworkers. Expression temperature is a second major lever: cold-shock induction systems such as pCold (Takara/Clontech) improve overall stability of the target, and lower temperatures halt cell proliferation and reduce endogenous protease expression, further protecting the target and improving purity. To support low-temperature expression, cold-adapted *E. coli* strains such as ArcticExpress (Agilent) co-express the cold-adapted chaperonins Cpn10 and Cpn60 from the psychrophilic bacterium *Oleispira antarctica*. Table 3 pairs each of these tactics with its key literature references.\n\n### 4.3. Host cells and protein stability", "# Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor\nAuthors: Kong, L.; Ju, B.; Chen, Yajing; He, L.; Ren, L.; Liu, J.; Hong, K.; Su, B.; Wang, Z.; Ozorowski, G.; Ji, X.; Hua, Y.; Chen, Yanli; Deller, M.C.; Hao, Y.; Feng, Y.; Garces, F.; Wilson, R.; Dai, K.; O'Dell, S.; McKee, K.; Mascola, J.R.; Ward, A.B.; Wyatt, R.T.; Li, Y.; Wilson, I.A.; Zhu, J.; Shao, Y.\nSource: Immunity, 2016\n\nWe investigated the DRVIA7 lineage longitudinally by tracing DRVIA7 heavy and light chains in the unbiased repertoires obtained from three time points (Figure 5A). We first analyzed the 2009 time point, from which DRVIA7 was identified. The heavy and light chain repertoires with respect to DRVIA7 were visualized via the identity-divergence two-dimensional (2D) plots. As expected, we observed two distinct islands on the 2D plots corresponding to the DRVIA7 somatic population. Specifically, the heavy chain population showed a divergence of 18%\u201327% (on average 23.1%) relative to the IgHV1-2 germline gene with a DRVIA7 identity of 85%\u201395%. In contrast, the light chain population displayed a lower divergence of 13%\u201320% and a DRVIA7 identity of 90%\u201395%. Using the cross-donor analysis (Zhu et al., 2013), we identified 530 VRC01-like heavy chains from the repertoire (Figure S4A), which, when overlaid on the 2D plot, not only covered an island of sequences most similar to DRVIA7, but also extended into the main population (Figure 5A, column 3, panel 2, in orange). These VRC01-like heavy chains accounted for 7.9\u2030 of the IgHV1-2 gene family, with DRVIA7 somatic variants \u2014 defined by an HCDR3 identity of 95% or greater \u2014 representing 1.4\u2030 of the gene family (in magenta). We selected 30 distinct heavy chains (Figure S4B) for phylogenetic analysis, which showed multiple sub-lineages (Figure 5A, top right). When paired with mature VRC01 light chain, most reconstituted antibodies showed broad neutralization, confirming that these heavy chains are indeed VRC01-like (Tables S4 and S5). We then identified 703 \u03ba chains with the VRC01-class LCDR3 signature (Figure 5A, column 3, panel 4, in cyan; Zhou et al., 2013), which accounted for 10.1\u2030 of the IgKV1-5 gene family. A total of 22 distinct light chains (Figure S4B) were selected for phylogenetic analysis. The light chain dendrogram exhibited a branching pattern similar to that of the heavy chains (Figure 5A, bottom right). Taken together, our analysis revealed a large population of VRC01-like antibodies in the 2009 repertoire containing a DRVIA7 lineage that had not yet overcome the glycan barrier.\n\nThe 2008 repertoire showed a markedly expanded DRVIA7 lineage with large islands extending from the near-100% identity to the main population on the 2D plots, for both heavy and light chains (Figure 5A, column 2). Compared to 2009, we observed a 5-fold expansion of VRC01-like heavy chains defined by cross-donor analysis (42.3\u2030) and a 10-fold expansion of DRVIA7-like heavy chains (14.0\u2030) defined by an HCDR3 identity of 95% or greater (Figure 5A, column 2, panel 2). These VRC01-like heavy chains showed an average divergence of 22.9%, similar to that observed for 2009 (23.1%). Consistently, we observed a 2.5-fold expansion of IgKV1-5 light chains with the VRC01-class LCDR3 signature (Figure 5A, column 2, panel 4). Of note, a group of near-germline light chains (5%\u201310% divergence) can be seen on the 2D plot, suggesting an active B cell development driven by the light chain selection. Ten distinct VRC01-like heavy chains that were not sequence homologs of the 30 tested 2009 heavy chains, as defined by an identity cutoff of 90%, were selected for functional validation (Figures S4A and S4B). When reconstituted with VRC01 light chain, they produced neutralization profiles similar to those identified from 2009 (Tables S4 and S5).\n\nThe 2006 repertoire, however, produced a distinct pattern compared to the two later time points. A population of heavy chains characterized by a low SHM level (10%\u201325%, with an average of 17.0%) and a DRVIA7 identity of 85% or less was \"budding\" from the main population (Figure 5A, column 1, panel 1). These IgHV1-2 sequences appeared not to possess the mature VRC01 signature (Zhu et al., 2013) defined by the converged cross-donor analysis (Figure S4A). However, the sequences identified in the first step of this analysis, accounting for 23.1\u2030 of the IgHV1-2 gene family, overlapped precisely with the budding population on the 2D plot (Figure 5A, column 1, panel 2, in yellow), suggesting that they might be precursors of DRVIA7 heavy chain. Five heavy chains selected from this population (Figure S4B), when reconstituted with VRC01 light chain, exhibited a remarkable neutralization breadth of 84% or greater (Figure 5B; Tables S4 and S5). Light chain analysis revealed that 12.2\u2030 of the IgKV1-5 gene family possess the VRC01-class LCDR3 signature, suggesting that it is readily encoded by the repertoire. The VJ recombination that gave rise to the 5-aa LCDR3 occurred constantly, as shown by the near-germline light chains in the 2008 repertoire (Figure 5A, column 2, panel 4). Furthermore, 5-aa LCDR3 loops were identified from most light-chain germline gene families at each time point.", "# Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\nAuthors: Deller, M.C.; Hudson, K.R.; Ikemizu, S.; Bravo, J.; Jones, E.Y.; Heath, J.K.\nSource: Structure, 2000\n\nThe N-terminal loop preceding helix A (Gly4\u2013Glu9) is tethered to the C terminus of helix C by one of the two disulphide bridges (Cys6 and Cys127). The second disulphide bridge, between Cys49 and Cys167, anchors the start of the AB loop to the N-terminal region of helix D. The AB loop itself consists of two separate \u03b1-helical regions running from Pro43 to Arg46 and Glu59 to Gly64. Residues between these two helical regions adopt an extended conformation and pack closely against helix D. By contrast, the second two loops (BC and CD) are less closely associated with the core structure. The BC loop protrudes above the four-helix bundle and exhibits higher than average B factors. Despite this relative mobility it does, however, contain a high percentage of regular secondary structure with a 310 helix running between residues Ala95 and Asp97 followed by an \u03b1 helix up to residue Ser101. The long CD loop appears to be highly mobile and could not be modelled as a single conformation.\n\n*Table 1.* Data collection statistics.\n\n| | Native | EMP \u03bb1 | EMP \u03bb2 | EMP \u03bb3 |\n|---|---|---|---|---|\n| Wavelength (\u00c5) | 0.979273 | 1.00914 | 0.826516 | 1.00669 |\n| Unit cell a, b, c (\u00c5) | 35.8, 53.1, 106.8 | 35.9, 53.3, 106.7 | 35.9, 53.3, 106.7 | 35.9, 53.3, 106.7 |\n| \u03b1 = \u03b2 = \u03b3 (\u00b0) | 90 | 90 | 90 | 90 |\n| Resolution range (\u00c5) | 25\u20132.2 | 25\u20132.6 | 25\u20132.6 | 25\u20132.6 |\n| Total reflections | 43,901 | 24,040 | 26,223 | 24,159 |\n| Unique reflections | 10,513 | 11,650 | 12,069 | 11,691 |\n| Redundancy | 4.2 | 2.1 | 2.2 | 2.1 |\n| Completeness (%) | 96.5 (98.2) | 96.4 (63.1) | 99.9 (100) | 96.8 (66.7) |\n| <I/\u03c3(I)> | 13.2 (4.3) | 12.1 (2.9) | 12.5 (4.7) | 12.3 (3.3) |\n| Rmerge (%) | 8.3 (37.6) | 4.6 (18.7) | 4.6 (16.4) | 4.3 (17.3) |\n\nEMP \u03bb denotes MAD data collected from EMP-derivatised hOSM. Rmerge = \u03a3|I \u2013 <I>|/\u03a3<I>. Values in parentheses correspond to the highest resolution shell (2.66\u20132.60 \u00c5 for EMP, 2.25\u20132.20 \u00c5 for native).\n\n*Table 2.* Refinement statistics.\n\n| | |\n|---|---|\n| Resolution range (\u00c5) | 20\u20132.2 (2.28\u20132.20) |\n| Rcryst (%) | 20.5 (27.1) |\n| Rfree (%) | 26.1 (33.6) |\n| Completeness, working set (%) | 86.1 (81.0) |\n| Completeness, test set (%) | 9.9 (8.6) |\n| No. of reflections (F > 0), working set | 9410 (870) |\n| No. of reflections (F > 0), test set | 1082 (92) |\n| No. of non-hydrogen atoms, protein | 1319 |\n| No. of non-hydrogen atoms, water | 116 |\n| Rmsd from ideality, bond lengths (\u00c5) | 0.0147 |\n| Rmsd from ideality, bond angles (\u00b0) | 1.7997 |\n| Rmsd from ideality, dihedrals (\u00b0) | 19.8882 |\n| Rmsd from ideality, improper (\u00b0) | 1.0381 |\n| Average B factor, mainchain (\u00c52) | 27.1 |\n| Average B factor, sidechain (\u00c52) | 32.4 |\n| Average B factor, water (\u00c52) | 38.4 |\n| Average B factor, all atoms (range) (\u00c52) | 30.5 (6.3\u201384.9) |\n\nRcryst = \u03a3||Fobs| \u2013 |Fcalc||/\u03a3|Fobs|. Rfree is as for Rcryst but calculated using a 10% test set of reflections excluded from the refinement. Values in parentheses correspond to the highest resolution shell (2.28\u20132.20 \u00c5).\n\n*Figure 1.* The structure of hOSM. (a) Refined coordinates and 2|Fo|\u2013|Fc|\u03b1calc electron-density map of hOSM contoured at 1\u03c3 showing density for the kinked region of helix A centred on Thr27. CNTF and LIF also have kinks owing to disruption of the mainchain hydrogen bonding around residues Thr32 and Ser36, respectively [19,20]. Both these residues bind water molecules through the O\u03b31 atom to provide a substitute hydrogen-bond network similar to the pattern observed here for hOSM involving residue Thr27. (b) Ribbon diagram of hOSM coloured from blue at the N terminus through to red at the C terminus. Disulphide bonds are represented as ball-and-stick models with the sulphur atoms highlighted as yellow spheres. The transparent dotted section represents the CD loop as observed in LIF. (c) Stereodiagram of the C\u03b1 trace for hOSM.\n\n### Structural comparison with other helical cytokines\n\nhOSM is the fourth member of the gp130 cytokine family to be analysed structurally; crystal and/or solution structures have been reported for IL-6 [21,26], CNTF [20] and LIF [19,27]. Structural superimpositions with these and other four-helix bundle cytokines show that the gp130 cytokines form a distinct family both structurally and functionally. Within the gp130 cytokine family hOSM has the greatest structural similarity to LIF (rmsd of 2.1 \u00c5 over 145 pairs of structurally equivalent C\u03b1 atoms). The more distant structural similarity to the four-helix bundles of the growth hormone family is at its best for superimposition of hOSM and erythropoietin (EPO; [28]) (rmsd of 2.8 \u00c5 for 120 pairs of structurally equivalent C\u03b1 atoms).\n\nThe addition of hOSM to the gp130 cytokine structural database highlights several features of potential functional importance (Figure 2). First, the kinked helix A is revealed as a distinctive feature not just of the gp130 cytokines but also specifically of the members of the family of cytokines that bind to LIFR; IL-6 is the sole gp130 cytokine lacking this feature (Figures 1a,2a). The kink results in a repositioning of the C-terminal end of helix A relative to the 'top' of the four-helix bundle, a region carrying the major epitope for LIFR, and OSMR, binding (at the start of helix D). This feature is therefore suggestive of the positioning of the C-terminal end of helix A impacting on LIFR and OSMR binding. Conversely, the positioning of the N-terminal portion of helix A, which is implicated in gp130 binding, is relatively conserved across all members of the family.", "# ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n### \ud83d\udcca Chat CLI: Streaming output and UX polish (2026-06-25)\n\n**Streaming generation:**\n\u2013 Response tokens now stream live to the terminal in real-time using `mlx_lm.stream_generate` + Rich `Live`; raw text streams token-by-token during generation, then the final response re-renders as formatted Markdown \u2014 no more waiting behind a \u201cThinking\u2026\u201d spinner\n\u2013 Falls back to blocking `generate` + spinner if `stream_generate` is unavailable\n\n**New slash commands:**\n\u2013 `/save` \u2014 exports the full conversation to a timestamped Markdown file on the Desktop (includes marcdeller.com attribution header)\n\u2013 `/info` \u2014 shows a panel with loaded model name/size, training config (iters, RSLoRA rank, lr, ctx), corpus stats (chunks, CSV files, MB), and embed model; clickable hyperlinks to [marcdeller.com](https://marcdeller.com) and [marc@marcdeller.com](mailto:marc@marcdeller.com)\n\u2013 `/retry` \u2014 re-generates the previous response using the same RAG context; replaces the last assistant turn in history\n\n**UX refinements:**\n\u2013 **Tab completion**: type `/` and press Tab for a dropdown of all slash commands; completion menu styled to match the ChemSage colour scheme\n\u2013 **RAG retrieval indicator**: `\ud83d\udd0d Retrieving context\u2026` spinner during retrieval, then `\ud83d\udcda N chunks retrieved` confirmation line before generation begins\n\u2013 **Context window indicator in toolbar**: bottom toolbar now shows `ctx N/3` alongside the turn counter so you can see how much of the 3-turn context window is filled\n\u2013 **`/reset` confirmation**: asks `Clear history? [y/N]` before wiping \u2014 no accidental data loss\n\u2013 **Generation stats**: now includes elapsed seconds \u2014 `\u26a1 N tok \u00b7 X.X tok/s \u00b7 X.Xs`\n\u2013 **Styled goodbye**: exit via `quit`/`q`/Ctrl-C shows a full-screen Panel with the ChemSage ASCII banner, \u201cThank you for using ChemSage\u201d, and clickable links to [marcdeller.com](https://marcdeller.com) and [marc@marcdeller.com](mailto:marc@marcdeller.com)\n\u2013 **Hyperlinks throughout**: banner, `/help`, `/info`, `/save` output, and goodbye Panel all include clickable `[link=...]` Rich markup pointing to marcdeller.com and marc@marcdeller.com\n\n**Earlier polish (same date):**\n\u2013 `prompt_toolkit` input with arrow-key history and styled `You:` prompt\n\u2013 Persistent bottom toolbar: `\u2b21 ChemSage \u00b7 model \u00b7 QLoRA \u00b7 RAG \u00b7 turn N \u00b7 ctx N/3`\n\u2013 Corpus tables: Rich `Table` with rounded cyan borders, single-line rows (ellipsis overflow), abbreviated headers, `\u2713`/`\u2717` booleans, method abbreviations\n\u2013 RDKit output in yellow-bordered `Panel`; retrieval errors in red-bordered `Panel`\n\u2013 Terminal title set to `ChemSage` on launch and `/clear`\n\u2013 Added `rich`, `prompt_toolkit`, `pyfiglet` to `requirements.txt`\n\n### \ud83e\uddea Eval harness: R4 extended metrics (2026-06-25)\n\nExtended `eval/eval_chem.py` with new metrics that degrade gracefully to `n/a` when run against R1-R3 models/datasets, enabling a clean side-by-side comparison across all rounds:\n\n| New metric | What it measures | Why added |\n| --- | --- | --- |\n| **Rounding precision** | MW 1 d.p. \u00b7 logP 2 d.p. \u00b7 TPSA 1 d.p. \u00b7 QED 2 d.p. | R4 drills exact decimal-place conventions |\n| **Code-then-quote** | prose values match what code blocks actually printed to stdout | Distinguishes quoting own code vs. recalling from memory |\n| **Extended executability** | all `python` blocks (RDKit + Plotly + Biopython) | R4 adds Plotly and Biopython code generation |\n| **Refusal accuracy** | out-of-scope prompts correctly declined | R4 adds explicit refusal training |\n| **QED range** | all stated QED values in [0, 1] | R4 adds QED computation training |\n\nBackwards compatibility: R1-R3 test sets have no refusal examples (refusal shows n/a), no QED mentions (QED shows n/a), code blocks may not print (code-then-quote shows n/a). All five new metrics return `(0, 0)` when inapplicable. The core 3-metric overall score formula is unchanged for direct cross-round comparison.\n\nNew scorecard features: per-class breakdown table (if test examples carry a `class` field), overall score card, collapsible response viewer in per-example table, extended exec breakdown by library category, mean latency in header, cleaner terminal summary with n/a annotation.\n\n### \ud83c\udf93 Round 3: Code-heavy dataset and extended training (2026-06-25)\n\n* Expanded SFT dataset from 1,500 to 5,000 examples (4,000 train) across 19 behaviour classes\n* 8 new generator classes targeting eval gaps: `single_property`, `full_profile`, `qed_score`, `molecular_formula`, `smiles_validation`, `property_comparison`, `corpus_drug`, `murcko_sar`\n* 51 seed SMILES + ~1,000 corpus drug examples from `approved_drugs.csv`\n* Configured for 1,500 iters; stopped early at iter 625 after val loss plateaued at ~0.060\n* Best val loss **0.054** at iter 625 (vs 0.347 for Round 2): **6.4x improvement**\n* Attempted 64 layers (swap thrashing at 31.5 GB) and batch_size=8 (4.5x slower); reverted to 32 layers / batch 4\n* Resumed from iter 100 checkpoint after battery shutdown; no data loss\n* Eval: SMILES 99%, Exec 99%, Fidelity 68% (vs R2: 100%, 41%, 10%)\n* Fused model saved to `models/chem_sage_32b_v3/` (17 GB)\n\n### Round 2: Qwen 2.5 32B upgrade (2026-06-24)\n\n* Upgraded base model from Qwen 2.5 7B to 32B (4-bit, MLX); trained with 32 adapted layers (vs 16 for 7B)\n* Best val loss 0.347 at step 675 (vs 0.389 for 7B): 10.8% improvement\n* Fused model saved to `models/chem_sage_32b_v2/` (17 GB); serves on 64 GB Apple Silicon at 24.9 GB peak memory\n* Eval shows 100% SMILES validity, 41% tool executability, 10% numerical fidelity (7B scored 0% across the board)\n* Added tqdm progress bar and live metrics to `eval/eval_chem.py`\n* System prompt injection in eval harness (test examples lack system messages)\n* Fixed `merge_export.py`: uses `sys.executable -m mlx_lm fuse` (corrected deprecated syntax), resolves adapter path relative to config directory\n\n### \ud83d\udcda Round 1: Corpus fast-path and chat CLI", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\nProtein: RECP1                  # short name, MAX 5 CHARACTERS (Boltz stores chain names\n                                # in a fixed 5-char field internally and silently\n                                # truncates longer ones, which then crashes later with a\n                                # confusing error -- `preflight` catches this for you).\n                                # Also names the output file: {protein}_{ligand}.yaml\nSequence: MDILC...              # required\nPartners: CHNX, CHNY            # optional: co-folded chains, defined as their own\n                                # Partner: blocks below\n# Ligands: LIG1, LIG3           # optional: restrict this protein to a ligand subset\n                                # (default: crossed with every ligand below)\n# Ligands: none                 # optional: ligand-free (apo) target -- no ligand entity,\n                                # no pocket constraint, no affinity property, whatever\n                                # `Predict affinity:` says. Stem is just the protein name\n                                # (e.g. `RECP1.yaml`, not `RECP1_LIG1.yaml`), runs through\n                                # the same generate/preflight/run/analyze pipeline and the\n                                # same staged `boltz predict` batch as every other target.\n# Modifications: SEP:5          # optional: CCD:position tokens for modified residues\n                                # (e.g. phosphoserine)\n# Cyclic: yes                   # optional: cyclic polymer (e.g. a cyclic peptide)\n# MSA: empty                    # optional: path to a precomputed MSA, or \"empty\" for\n                                # single-sequence mode (skip MSA generation)\n# Templates: reference_structure.cif\n                                # optional: structural template file(s), applied to all\n                                # protein chains (no per-chain mapping -- hand-edit the\n                                # generated YAML for that rarer case)\n# Apo structure: reference/apo.pdb\n                                # optional: a reference apo/unbound structure, used only\n                                # by `compare-sse` (see below), never by generate/run.\n                                # No genuinely apo experimental structure? Predict one:\n                                # give another `Protein:` block the same `Sequence:` and\n                                # `Ligands: none` (see above), run the campaign once, then\n                                # point `Apo structure:` at its output in `boltz_cif/`.\n# Apo chain: A                 # optional: explicit chain id in the apo structure above\n                                # (omit to auto-detect via sequence identity)\n# Pocket source: LIG1 from 7XYZ # optional, and repeatable: records which experimental\n                                # structure a `Pocket contact:` set was derived from.\n                                # Reporting only -- never affects generate/run. Without\n                                # it the reference panel reads \"not recorded\" for that\n                                # pocket, and the pose panel can only find the\n                                # experimental ligand if that structure happens to be\n                                # sitting in `reference/`. Repeat it once per structure\n                                # a protein takes a site from; one code claimed by two\n                                # structures is an error. A `Pocket source:` line with\n                                # NO matching `Pocket contact:` lines is a **reference\n                                # molecule**: the structure ships so its bound ligand can\n                                # score a matching compound's pose, and it defines no\n                                # site, so it adds no targets\n# Family type: gpcr            # optional: gpcr / kinase / auto (default) -- selects\n                                # `compare-sse`'s motif annotator\n# Group: RECP1                  # optional: shared display/report name for multiple\n                                # `Protein:` blocks that are the same underlying receptor\n                                # (e.g. with/without a partner, or a predicted apo\n                                # variant) -- defaults to this block's own name if unset\n\nPartner: CHNX\nSequence: MTLES...\n# Type: dna                     # optional: protein (default) / dna / rna\n# Copies: X1, X2                # optional: homo-oligomer chain-id override -- this one partner\n                                # sequence becomes multiple chains\n\nLigand: LIG1\nSMILES: FC(F)CNC(...)=O         # exactly one of SMILES/CCD is required\n# Role: agonist                 # optional: agonist / antagonist -- reporting only (dashboard\n                                # chart shapes), never affects generate/run\n# Class: control                # optional: control / experimental -- reporting only. Marks\n                                # which compounds are the known-answer yardsticks and which\n                                # are the ones being asked about, so the charts can outline\n                                # the latter without spending a colour on them\n\nLigand: LIG2\nCCD: GOL                        # a Chemical Component Dictionary code (e.g. common crystallization\n                                # additives/ions) instead of a SMILES\n\nCovalent bond: RECP1 residue 44 atom SG to LIG1 residue 1 atom C3\nPocket contact: RECP1 residue 148\nDistance constraint: RECP1 residue 10 to RECP1 residue 80 within 8.0 Angstrom\n```\n\nEvery protein is crossed with every ligand (unless a protein sets `Ligands:` to scope itself\nto a subset, or `Ligands: none` for a single ligand-free/apo target), producing one\n`{protein}_{ligand}.yaml` per pair. See `example.md` for the full copy-paste template and\n`examples/` for complete working campaigns.\n\n## \ud83d\ude80 Commands", "# Nanobodies: Revolutionizing Molecular Medicine Through Nature\u2019s Smallest Antibodies\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-27 on marcdeller.com (categories: blog, sb).*\n\n* **Revolutionary Architecture:** Single-domain structure achieves full antigen binding in just 15 kDa proteins\n* **Exceptional Stability:** Thermal resistance up to 90\u00b0C with reversible refolding capabilities\n* **Enhanced Penetration:** Superior tissue access and blood-brain barrier crossing for neurological applications\n* **Clinical Success:** Multiple approved therapeutics with projected $3.3 billion market by 2033\n* **Rapid Development:** AI-driven design and synthetic libraries enable accelerated therapeutic discovery\n* **Versatile Applications:** Spanning structural biology, diagnostics, immunotherapy, and pandemic response\n\n* **NANOBODIES\u00ae: A Review of Generation, Diagnostics and Therapeutics:** Bo-kyung Jin, et al., International Journal of Molecular Sciences:\n  <https://pmc.ncbi.nlm.nih.gov/articles/PMC10057852/>\n* **Single domain antibody: Development and application in biotechnology and biomedicine:** Haixia Li, et al., Immunological Reviews:\n  <https://onlinelibrary.wiley.com/doi/10.1111/imr.13381>\n* **The Impact of Nanobodies on G Protein-Coupled Receptor Structural Biology and Their Potential as Therapeutic Agents:** David Salom, et al., Molecular Pharmacology:\n  <https://pubmed.ncbi.nlm.nih.gov/39107078/>\n* **Nanobodies and their derivatives: pioneering the future of cancer immunotherapy:** Haixia Li, et al., Cell Communication and Signaling:\n  <https://pubmed.ncbi.nlm.nih.gov/40474230/>\n\nThe remarkable discovery of nanobodies has fundamentally transformed our understanding of antibody-based therapeutics and diagnostics. These extraordinary single-domain antibodies, derived from the unique heavy-chain-only antibodies found exclusively in camelids, represent a paradigm shift in molecular medicine. Their diminutive size of approximately 15 kilodaltons, combined with exceptional stability and specificity, positions nanobodies as powerful tools that address critical limitations of conventional antibodies in both research and clinical applications.\n\nThe evolutionary origins of nanobodies reflect nature\u2019s ingenious adaptation to challenging environments. Camelids developed these simplified antibody structures as an evolutionary response to pathogenic pressures, possibly enhanced by extreme environmental conditions including high temperatures and water scarcity. This natural selection process resulted in antibodies with enhanced thermal stability, capable of withstanding temperatures up to 90\u00b0C while maintaining their functional integrity through reversible refolding. Unlike conventional antibodies requiring both heavy and light chains, nanobodies achieve complete antigen recognition through a single variable heavy-chain domain, eliminating the complexity of chain pairing while preserving high-affinity binding.\n\nRecent advances in nanobody engineering have accelerated their translation from laboratory curiosities to clinical realities. The market for nanobodies has experienced unprecedented growth, reaching $602.98 million in 2024 with projections indicating expansion to $3.3 billion by 2033, representing a remarkable 20.9% compound annual growth rate. This explosive growth reflects the successful clinical approval of multiple nanobody-based therapeutics, including Caplacizumab for thrombotic thrombocytopenic purpura, Ozoralizumab for rheumatoid arthritis, and Ciltacabtagene autoleucel for multiple myeloma.\n\nThe unique structural properties of nanobodies confer several advantages that distinguish them from conventional antibodies. Their compact size enables superior tissue penetration and enhanced ability to cross physiological barriers, including the blood-brain barrier, making them particularly valuable for treating neurological disorders. The extended complementarity-determining region 3 (CDR3) loops, averaging 17-18 amino acids, adopt \u03b1-helical conformations that provide exceptional binding diversity and enable recognition of cryptic epitopes inaccessible to larger antibodies. Additional disulfide bonds between CDR1 and CDR3 regions further stabilize these extended loops, contributing to their remarkable binding affinity that can reach picomolar levels.\n\nIn structural biology, nanobodies have revolutionized protein crystallization as essential chaperones that stabilize specific conformational states. The breakthrough use of nanobody Nb80 to stabilize the active conformation of the \u03b22-adrenergic receptor enabled the first structural determination of an active G-protein-coupled receptor. This success has been replicated across numerous membrane proteins, with nanobodies facilitating the determination of over 340 GPCR structures to date. Recent innovations include \u201cGluebodies,\u201d engineered nanobody scaffolds with optimized crystallization properties that demonstrate superior resolution and reliability compared to conventional approaches.\n\nThe therapeutic applications of nanobodies span diverse medical fields, with particularly promising developments in cancer immunotherapy. Nanobody-based CAR-T cell therapies have demonstrated enhanced safety profiles and reduced toxicity compared to traditional scFv-based approaches. The compact structure of nanobodies enables the development of sophisticated multi-specific constructs that can simultaneously target multiple antigens while maintaining structural stability. Recent clinical trials have shown remarkable efficacy, with nanobody-based CD7-targeted CAR-T therapy achieving a 70% complete remission rate in patients with refractory acute myeloid leukemia.", "# Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\nAuthors: Kong, L.; Lee, J.H.; Doores, K.J.; Murin, C.D.; Julien, J.-P.; McBride, R.; Liu, Y.; Marozsan, A.; Cupo, A.; Klasse, P.-J.; Hoffenberg, S.; Caulfield, M.; King, C.R.; Hua, Y.; Le, K.M.; Khayat, R.; Deller, M.C.; Clayton, T.; Tien, H.; Feizi, T.; Sanders, R.W.; Paulson, J.C.; Moore, J.P.; Stanfield, R.L.; Burton, D.R.; Ward, A.B.; Wilson, I.A.\nSource: Nature Structural & Molecular Biology, 2013\n\nOverall in comparing HIV-1 bnAbs that interact with glycans, PGT 135 is similar to PGT 128 and PG9 in its use of heavy chain CDR loops to penetrate through the glycan canopy to contact the underlying protein surface with PG9 uniquely using a hammerhead-shaped CDR H3 loop for the interaction. As with PGT 128, but not PG9, PGT 135 also relies on a second extended CDR loop on the heavy chain (H1 in PGT 135 and H2 in PGT 128) to contact antigen. Despite different amino-acid lengths, these extended CDR loops are at similar distances from the antibody (15-20 \u00c5), reflecting the length of an N-linked glycan from the underlying protein surface. In contrast, 2G12 only contacts the outermost mannose residues of the interacting glycans, a feature that presumably results from its unusual VH domain-swap to attain high affinity binding.\n\nIn terms of the relative glycan positioning within the binding sites, we also observe some similarities. For Fabs PGT 135, PGT 128 and 2G12, the distances between the two major interacting glycans on the gp120 surface (measured between the asparagine-bound carbons of the protein proximal GlcNAcs of the glycans) are remarkably similar at around 18-20 \u00c5. The PG9 glycans are also spaced similarly at 15 \u00c5. These similarities in distances likely reflect constraints imposed by the dimensions of the Fab interacting surfaces, the requirement that a CDR loop penetrate between glycans, and the density of the glycans on the gp120 surface.\n\nAlthough three different bnAbs recognize the same Asn332 glycan, they contact chemically distinct surrounding surfaces. 2G12 binds to the glycan tips and is, therefore, highly dependent on the presence of the terminal mannose residues in a particular linkage. In contrast, PGT 135 and PGT 128 recognize the complete length of the glycan at Asn332, but approach from and interact with different sides of the glycan. Strikingly, the glycan conformation and orientation is largely unchanged in the PGT 135 and PGT 128 structures (Fig. 6a), despite the known conformational flexibility of glycans. This conformational stability confers the carbohydrate with two chemically distinct \"faces\" at the level of the entire glycan; a much more flexible glycan would not have such well-defined faces because its overall orientation and conformation would be rapidly changing. The Asn332 glycan faces are defined by the sum of the contributions from the apolar and polar faces of individual sugar rings that then determine the net polarity of the complete face. The sugar hydroxyl groups extend from the polar side of a sugar ring, thereby potentially mediating more hydrophilic-type interactions compared to the apolar face (Fig. 6b). Thus, some 'rules' for interaction with individual sugars have been established whereby aromatic residues, such as Trp, typically stack on the apolar face of sugars and hydrophilic and charged residues interact with the polar face and hydrogen bond with the ring hydroxyl groups.\n\nWe observed that PGT 135 interacts predominantly with an apolar face consisting of three individual apolar sugar ring faces and one polar sugar ring face (Fig. 6c) with 61% of the buried surface in the interaction mediated by the individual apolar faces. PGT 128 interacts with a predominantly polar face consisting of 5 individual polar sugar faces and 2 individual apolar sugar ring faces, with 58% of the buried surface mediated by the polar faces (Fig. 6d). Conversely, PGT 135 residues interacting with the apolar face of glycan 332 are predominantly hydrophobic (\u223c64% of the buried surface area contacts are mediated by apolar residues), while PGT 128 residues are about equally hydrophobic and hydrophilic (\u223c54% of the buried surface in the interaction are mediated by apolar residues). It should be noted that the apolar interactions, particularly ring stacking, would naturally involve much greater surface area than hydrophilic interactions. Similarly, PGT 128 interactions with the polar face of glycan 332, involve many more hydrogen bonds compared to PGT 135 (13 versus 4). Thus, we conclude that the different faces of the Asn332 glycan are an important determinant for antibody recognition of this supersite.\n\nThe discovery of bnAbs in a SHIV-infected macaque that depend on the Asn332 glycan, in addition to those described above in infected humans, raises the prospect that suitable immunogens that display these combined glycan and protein epitopes in an appropriate configuration can be tested as vaccine candidates in animal models for re-elicitation of similar glycan-dependent antibodies. Furthermore, a recent study showed that broadly neutralizing responses were raised against the Asn332 glycan in two human individuals initially infected with a virus initially lacking the Asn332 glycan, but which contained an Asn334 glycan that is present only when Asn332 is absent. The Asn332 glycan emerges later in the infection, but the virus can revert back to the Asn334 glycan, resulting in escape from the immune response. Although most viruses have the Asn332 glycan (73%) rather than the Asn334 glycan (20% conserved), a vaccine may need to be capable of inducing responses to either glycan site.", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\nAbstracts for 58 of Marc Deller's papers indexed in Europe PMC (author search \"Deller MC\"), newest first. Marc's roles range from first and corresponding author (oncostatin M, TrkA gating ring, crystal harvesting, protein stability) to co-author contributing crystallography (JCSG structural genomics papers 2007\u20132012, HIV-1 Env and HCV E2 papers with Ian Wilson's lab, Stanford papers, and Incyte medicinal-chemistry papers).\n\n## Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases (2026)\n\n**Authors:** Zhuo J, Li YL, Qian DQ, Burns DM, Mei S, Rafalski M, Xu M, Cao G, Pan Y, Jia Z, Jalluri R, Epling LB, Fenalti G, Deller MC, Procak J, Covington M, He X, Collins R, Stubbs M, Burke K, Oliver J, Margulis A, Boer J, Wynn R, Scherle P, Diamond S, Newton R, Zhang Y, Metcalf B, Yao W.  \n**Journal:** Journal of medicinal chemistry (2026). DOI: 10.1021/acs.jmedchem.5c03753. PMID: 42202255.\n\nAberrant JAK2 signaling is a key driver in cancers such as myeloproliferative neoplasms (MPNs), whereas JAK1 acts as a central mediator of autoimmune and inflammatory diseases. Since JAK2 supports blood cell production, its inhibition can lead to anemia and other cytopenias, making selective JAK1 inhibition a potentially safer therapeutic strategy opposed to dual JAK1/JAK2 inhibition. We initially identified novel scaffold 9, which selectively inhibited JAK1 over JAK2. Iterative breakthroughs led to the discovery of povorcitinib (INCB054707, 22), which showed improved pharmacokinetics due to intramolecular hydrogen bonding of the amide moiety and a reduced ring size in the hinge binding motif. Protein crystallography suggested that JAK1 selectivity was conferred by the amide's alkyl group accessing a small lipophilic groove formed by the P-loop conformation of JAK1 that was distinct from JAK2. In murine arthritis models, povorcitinib demonstrated dose-dependent efficacy and reduced inflammatory signaling, supporting its therapeutic potential in immunological disorders.\n\n## Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor (2025)\n\n**Authors:** Ye Q, Shvartsbart A, Li Z, Gan P, Policarpo RL, Qi C, Roach JJ, Zhu W, McCammant MS, Hu B, Li G, Yin H, Carlsen P, Hoang G, Zhao L, Susick R, Zhang F, Lai CT, Allali Hassani A, Epling LB, Gallion A, Kurzeja-Lipinski K, Gallagher K, Roman V, Farren MR, Kong W, Deller MC, Zhang G, Covington M, Diamond S, Kim S, Yao W, Sokolsky A, Wang X.  \n**Journal:** Journal of medicinal chemistry (2025). DOI: 10.1021/acs.jmedchem.4c02662. PMID: 39772605.\n\nThe inhibition of mutant KRAS proteins has emerged as a promising approach for treating KRAS-driven cancers, as evidenced by the clinical success of KRAS G12C inhibitors. KRAS G12D, the most common mutant, promises significant expansion of the addressable patient population; however, the reduced nucleophilicity of aspartate compared to cysteine poses significant challenges in balancing sufficient potency with ADME properties to support oral exposure. Herein, we describe the discovery of KRAS G12D inhibitor 23 (INCB159020), which achieves oral exposure in nonhuman primate (NHP). Starting from a weakly potent hit, structure-based drug design was utilized to drive significant potency. Focus on molecular rigidity and balanced polarity then allowed for successful optimization of properties required for oral exposure.\n\n## Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2 (2024)\n\n**Authors:** Hummel JR, Xiao KJ, Yang JC, Epling LB, Mukai K, Ye Q, Xu M, Qian D, Huo L, Weber M, Roman V, Lo Y, Drake K, Stump K, Covington M, Kapilashrami K, Zhang G, Ye M, Diamond S, Yeleswaram S, Macarron R, Deller MC, Wee S, Kim S, Wang X, Wu L, Yao W.  \n**Journal:** Journal of medicinal chemistry (2024). DOI: 10.1021/acs.jmedchem.3c02287. PMID: 38325398.\n\nCDK2 is a critical regulator of the cell cycle. For a variety of human cancers, the dysregulation of CDK2/cyclin E1 can lead to tumor growth and proliferation. Historically, early efforts to develop CDK2 inhibitors with clinical applications proved unsuccessful due to challenges in achieving selectivity over off-target CDK isoforms with associated toxicity. In this report, we describe the discovery of (4-pyrazolyl)-2-aminopyrimidines as a potent class of CDK2 inhibitors that display selectivity over CDKs 1, 4, 6, 7, and 9. SAR studies led to the identification of compound 17, a kinase selective and highly potent CDK2 inhibitor (IC50 = 0.29 nM). The evaluation of 17 in CCNE1-amplified mouse models shows the pharmacodynamic inhibition of CDK2, measured by reduced Rb phosphorylation, and antitumor activity.\n\n## Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3 (2022)\n\n**Authors:** Shvartsbart A, Roach JJ, Witten MR, Koblish H, Harris JJ, Covington M, Hess R, Lin L, Frascella M, Truong L, Leffet L, Conlen P, Beshad E, Klabe R, Katiyar K, Kaldon L, Young-Sciame R, He X, Petusky S, Chen KJ, Horsey A, Lei HT, Epling LB, Deller MC, Vechorkin O, Yao W.  \n**Journal:** Journal of medicinal chemistry (2022). DOI: 10.1021/acs.jmedchem.2c01366. PMID: 36356320.", "# Courage: 400+ challenging structures determined, enabling breakthrough mechanistic insights\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-29 on marcdeller.com (categories: 6cs).*\n\n## Courage: 400+ challenging structures determined, enabling breakthrough mechanistic insights\n\nFor me, courage in science is about stepping into hard problems knowing that failure is likely but impact is meaningful if you succeed. Determining [400+ protein structures](https://marcdeller.com/structures/) has meant repeatedly choosing the \u201cugly\u201d targets because those are often where the most important biology hides. This has required a bias toward action and an acceptance that many attempts will not work the first, second, or even tenth time. By consistently tackling the hardest structural questions, I aim to convert technical risk into strategic advantage for the programs and teams I support.\n\n* Courage means pursuing complex, high-stakes protein targets where others see only failure.\n* Determining challenging structures has demanded calculated risk and embracing unknowns to reveal mechanisms behind disease and function.\n* It matters because scientific courage turns persistence into impact\u2014illuminating biology and enabling new therapies.\n* Courage also looks like a clock\u2014solving a structure in days when an IP window or clinical decision cannot wait.\n* Real courage sometimes means trusting a structure that contradicts the prevailing model\u2014and acting on it before the field is ready.\n\nPersonal Philosophy: The 6C's\n\n### [**Curiosity:** 10+ years driving hands-on drug discovery through multidisciplinary approaches](https://marcdeller.com/curiosity-20-years-driving-hands-on-discovery-through-multidisciplinary-approaches/)\n\n[**Curiosity:** 10+ years driving hands-on drug discovery through multidisciplinary approaches](https://marcdeller.com/curiosity-20-years-driving-hands-on-discovery-through-multidisciplinary-approaches/)\n\n### [**Creativity:** 7+ patents through innovative protein engineering and AI integration](https://marcdeller.com/creativity-5-patents-through-innovative-protein-engineering-and-ai-integration/)\n\n[**Creativity:** 7+ patents through innovative protein engineering and AI integration](https://marcdeller.com/creativity-5-patents-through-innovative-protein-engineering-and-ai-integration/)\n\n### [**Communication:** 60+ publications translating structural data into actionable insights](https://marcdeller.com/communication-60-publications-translating-structural-data-into-actionable-insights/)\n\n[**Communication:** 60+ publications translating structural data into actionable insights](https://marcdeller.com/communication-60-publications-translating-structural-data-into-actionable-insights/)\n\n### [**Collaboration:** 2+ state-of-the-art Structural Biology facilities built through global partnerships](https://marcdeller.com/collaboration-2-state-of-the-art-structural-biology-facilities-built-through-global-partnerships/)\n\n[**Collaboration:** 2+ state-of-the-art Structural Biology facilities built through global partnerships](https://marcdeller.com/collaboration-2-state-of-the-art-structural-biology-facilities-built-through-global-partnerships/)\n\n### [**Compassion:** 6+ FDA IND filings supported while mentoring 20+ scientists](https://marcdeller.com/compassion-6-fda-ind-filings-supported-while-mentoring-20-scientists/)\n\n[**Compassion:** 6+ FDA IND filings supported while mentoring 20+ scientists](https://marcdeller.com/compassion-6-fda-ind-filings-supported-while-mentoring-20-scientists/)", "# GitHub README: analyze-boltz2-results\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/analyze-boltz2-results, 2026\n\n> **Harvest a sprawling Boltz-2 output tree into one tidy summary, CSV, and CIF folder.**\n\n    \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/analyze-boltz2-results\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/analyze-boltz2-results</a></td>\n</tr>\n</table>\n\n---\n\nA post-run analysis script for [Boltz-2](https://github.com/jwohlwend/boltz) co-folding predictions. Run it in any directory containing a `boltz_run_*.log` file and a `predictions/` subdirectory to automatically harvest confidence scores, affinity predictions, CIF models, and run metadata into a clean summary report and CSV.\n\nWhy it matters: a Boltz-2 batch leaves behind a deep, scattered output tree (per-target directories, confidence JSONs, affinity JSONs, ModelCIFs, and a verbose log) that is painful to read by hand and impossible to rank at a glance. analyze_boltz.py walks the whole tree, extracts every useful number, derives a pIC50 estimate, and consolidates everything into a human-readable summary, a machine-readable CSV, and a flat folder of all model CIFs. It is useful for anyone screening many complexes: instant triage and ranking of predictions by confidence and predicted affinity, with robust handling of missing or malformed files so no result is silently dropped.\n\n---\n\n## Overview\n\nBoltz-2 produces a sprawling output tree: per-target prediction directories, confidence JSONs, affinity JSONs, and ModelCIF files, alongside a verbose run log. `analyze_boltz.py` walks that tree, extracts everything useful, and consolidates it into three outputs:\n\n| Output | Description |\n|---|---|\n| `boltz_run_summary.log` | Human-readable summary: invocation, directory tree, runtime, per-prediction scores |\n| `boltz_summary.csv` | Machine-readable table of all predictions with confidence, affinity, and pIC\u2085\u2080 |\n| `boltz_cif/` | Flat directory of all `*_model_0.cif` files, ready for downstream structure analysis |\n\n---\n\n## Requirements\n\n- Python 3.10+ (uses `Path | None` union syntax)\n- Standard library only \u2014 no external dependencies\n\n---\n\n## Usage\n\nCopy `analyze_boltz.py` into, or run it from, the top-level directory of your Boltz-2 run output:\n\n```bash\ncd /path/to/your/boltz_run_dir\npython analyze_boltz.py\n```\n\nThe script resolves everything relative to the current working directory. No arguments are required.\n\n### Expected directory layout\n\n```\nboltz_run_dir/\n\u251c\u2500\u2500 boltz_run_20260101_120000.log     \u2190 required: matches boltz_run_*.log\n\u251c\u2500\u2500 predictions/                      \u2190 required: can be nested\n\u2502   \u251c\u2500\u2500 target_001/\n\u2502   \u2502   \u251c\u2500\u2500 target_001_model_0.cif\n\u2502   \u2502   \u251c\u2500\u2500 confidence_target_001_model_0.json\n\u2502   \u2502   \u2514\u2500\u2500 affinity_target_001_model_0.json\n\u2502   \u2514\u2500\u2500 target_002/\n\u2502       \u2514\u2500\u2500 ...\n\u2514\u2500\u2500 analyze_boltz.py\n```\n\n`predictions/` can be located anywhere below the top-level directory \u2014 the script walks the full subtree to find it.\n\n---\n\n## Outputs\n\n### `boltz_run_summary.log`\n\nSections:\n\n1. **Zsh execution script and switches** \u2014 reconstructed boltz invocation line(s) from the log\n2. **Directory structure** \u2014 annotated tree up to 3 levels deep\n3. **Run timing** \u2014 start/finish timestamps and total wall-clock runtime\n4. **Prediction summaries** \u2014 per-target: CIF name, confidence score, affinity values, pIC\u2085\u2080\n5. **CIF copy operations** \u2014 log of every file copied into `boltz_cif/`\n\n### `boltz_summary.csv`\n\nOne row per prediction subdirectory:\n\n| Column | Description |\n|---|---|\n| `input YAML name` | Source YAML inferred from log or directory name |\n| `output cif name` | Filename of the `*_model_0.cif` model |\n| `confidence_score` | Boltz-2 overall confidence score (0\u20131) |\n| `affinity_pred_value1` | Raw affinity prediction (log-scale, kcal/mol units) |\n| `affinity_probability_binary1` | Probability of binding (binary classifier output) |\n| `pIC\u2085\u2080` | Derived pIC\u2085\u2080: `(6 \u2212 affinity_pred_value1) \u00d7 1.364` |\n\n### `boltz_cif/`\n\nFlat collection of all `*_model_0.cif` files discovered across the predictions tree. Suitable for bulk loading into PyMOL, ChimeraX, or a downstream PLIP/interaction analysis pipeline.\n\n---\n\n## pIC\u2085\u2080 Derivation\n\nThe affinity output from Boltz-2 (`affinity_pred_value1`) is a log-transformed binding free energy proxy. The conversion applied here is:\n\n```\npIC\u2085\u2080 = (6 \u2212 affinity_pred_value1) \u00d7 1.364\n```\n\nThis is a heuristic approximation; treat values as relative rankings rather than absolute potency predictions.\n\n---\n\n## Notes\n\n- **YAML association**: prediction directories are matched to input YAMLs by stem name. If no match is found, the script falls back to the first YAML detected in the log, then derives a name from the directory itself \u2014 no prediction is silently dropped.\n- **Robustness**: the script tolerates missing confidence or affinity JSONs, absent CIF files, and non-standard timestamp formats. Failures are logged per-prediction rather than aborting the run.\n- **Log parsing**: the invocation heuristic looks for lines containing both `boltz` and `python`/`zsh`. If none are found, it falls back to `Running structure prediction` / `Running affinity prediction` lines.\n\n---\n\n## Compatibility\n\nTested against Boltz-2 output structure as of early 2026. The script assumes:\n- Log filename matches `boltz_run_*.log`\n- Confidence JSONs match `confidence*_model_0.json`\n- Affinity JSONs match `affinity*.json`\n- Model CIFs match `*_model_0.cif`\n\n---\n\n## Licence\n\nMIT\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist\n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/analyze-boltz2-results\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/analyze-boltz2-results</a></td>\n</tr>\n</table>", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n- [x] Give the dashboard one number that is not Boltz grading its own work: a\n  **Ligand pose vs experiment** panel comparing the docked ligand with the same molecule in an\n  experimental structure. Measured motivation: a GLP1R/orforglipron prediction scored\n  `ligand_iptm` 0.940 and confidence 0.836 while sitting 9.4 A from where 7E14 puts the\n  same ligand in the same pocket, rotated end for end, and nothing on the dashboard could\n  say so. Reports site, pose and conformer separately, because they fail independently\n  (3.05 / 9.43 / 3.08 reads as \"right pocket, right shape, wrong orientation\"; one figure\n  reads as nothing). Atoms are paired by molecular graph via RDKit rather than by\n  proximity, which would match a wrong pose to itself; and a reference is refused unless\n  it is genuinely a structure of that protein -- 7E14 matched GIPR's residue numbering at\n  0.11 identity and would otherwise have been reported on. See **Ligand pose vs experiment** above.\n- [x] Answer \"how tight should a pocket be?\" with measurements rather than intuition:\n  six conditions on GLP1R + orforglipron against 7E14, reported in **How tight should a\n  pocket be?** above. `Pocket distance` turned out to be the lever -- the same 62\n  residues at 4 A instead of 8 A took the pose from 9.51 A to 2.04-2.56 A -- while the\n  contact count did not order the outcome at all. Confidence spans 0.834-0.849 across\n  that whole range and ligand ipTM ranks a 8.96 A failure mid-pack, which is the\n  evidence for the pose panel existing. Decoupled the confine-to-receptor sweep from\n  `Pocket distance:` in the same pass, so tightening a pocket no longer re-tunes the\n  baseline arm it is compared against.\n- [x] Give each comparison its own frame: **one pair viewer per predicted/experimental\n  ligand pair**, exactly two ligands as sticks, superposed through their receptors and\n  zoomed to fit, grouped by the pocket each target was run against. Fed by two\n  single-ligand mmCIFs BoltzMaker writes at analyze time, so the picture is the same\n  superposition the numbers came from and the server needs no RDKit. Frames are pooled\n  and drawn on scroll, because a browser silently kills the oldest WebGL context rather\n  than refusing a new one, which turns early tiles black while later ones look fine.\n- [x] Let a private campaign be shown to someone without publishing it: **share links**.\n  The same self-contained HTML package the Download button produces is hosted at an\n  unguessable URL behind a generated password, shown once and stored only as a PBKDF2\n  hash. A share is still not a listed run -- it lives under `<runs_root>/shares/`, which\n  the Archive never reads, so it cannot appear on `/runs` or the landing page however many\n  exist. Two independent secrets per share (the viewing password cannot revoke), a\n  per-share cookie secret so revoking invalidates every session already issued, a\n  10-attempt lockout that survives a worker restart, and revocation both by a one-click\n  link and by the campaign's own Destroy button. Shares never expire by design, so the\n  storage cap refuses a new share rather than evicting an old one -- evicting would break\n  the promise that a link works until it is revoked.\n- [x] Add **Use same pocket** to Prepare, with a distance box defaulting to 8 A. Co-folding puts a ligand wherever it likes: measured on a real GLP1R/GIPR campaign, 6 of 27 ligands docked onto the G-protein subunits rather than the receptor, and only 6 of 14 GIPR ligands shared a site. The pocket reference is **per ligand**, not per protein, because a pocket belongs to the ligand-receptor pair -- orforglipron's site on GLP1R (7E14) and 41Y's site on GIPR (7RBT) share 3 residues out of ~60 once projected onto each other, so one pocket per receptor would force one chemotype into the wrong site. Each ligand names a **holo** reference and picks from the ligands found in it; waters, ions, buffers, sugars and lipids are excluded outright, so 7dty's six cholesterols are never offered. Holo is enforced for the pocket and apo is enforced for the comparison: a reference with a bound ligand is refused as apo (6ln2, a modulator+Fab complex, was used as one for weeks) and a ligand-free one is refused as a pocket. Residues are mapped into the user's own numbering by alignment and reach Boltz as a `pocket` constraint with `max_distance`.\n- [x] Close the archive's trust gap: \"Keep private\" put the guarantee in a checkbox the caller had to remember to tick, and automated form posts made against the live site published three bundles of a real private campaign on `/runs` and the front page. Archiving now needs positive evidence of a person submitting the form from this site (`Sec-Fetch-Site: same-origin`, or a same-origin `Referer`), fails closed otherwise, and honours an explicit `X-BoltzMaker-No-Archive` opt-out for test clients.\n- [x] Make the bundle **the only file anyone has to keep**: it runs the campaign, and uploading it back at the top of Step 1 restores the form to edit and rebuild. It was three artefacts before -- the `.command`, the `.bmz` it produces, and a `.boltzpage.json` from a \"Save page\" button -- with no way to say which mattered; the second and third are gone from the page. Reading a `.command` needs a line-anchored search for the payload marker, because the extractor greps for its own marker and a plain `find` slices the shell script in as base64, failing with \"Incorrect padding\" and naming nothing.\n- [x] Load a bundle that **has no saved page inside it** -- every example campaign and everything already in the Runs archive -- by rebuilding the form from its `boltz_input.md`, which fixes bundles already on people's disks rather than only the archived copies. Done only when provably safe: the rebuilt page goes back through the form's own parser and assembler and the resulting spec is compared with the original, refusing with the difference named if they disagree, and refusing any directive the derivation has not been taught. The check earned itself immediately, catching eight silent losses -- companion proteins invented on every reload (an unticked box posts nothing, and the field being wholly absent reads as \"old page, default all on\"), a covalent bond dropped because blank fields were skipped and the parallel arrays went ragged, `Class: experimental` appearing from nowhere, 5HT2's `H2AAP` renamed to `5HAP`, three companions collapsing into one because they were matched on a sequence their variants share, companions lost when a real apo exists and nothing points at them, and a spec with no `Settings:` block reading as different from its own rebuild.\n- [x] Give the form the fields a spec always had and it did not: per-protein **`Ligands:` scoping**, **`Group:`**, **`Family type:`**, a ligand's **`Role:`**, an apo reference as a **file in the bundle** with an optional **chain**, and **`Targets per invocation`**. Not optional extras -- these are why no example campaign could be loaded back in, and deriving a form from a spec is only honest if the form can hold everything the spec says.\n- [x] Stop the two places a partner is named from drifting apart: **co-folded partners are tickboxes** built from the Partner rows themselves. Renaming a partner after a protein referenced it used to fail validation at download time with the campaign already entered. Still posts one comma-separated `protein_partners[]` per row -- a `<select multiple>` posts one value per selection, which would make the protein rows' parallel arrays ragged and attach partners to the wrong protein.\n- [x] **Read a PDB id back** under the box, the way the UniProt accession already was: title, method, resolution and the bound ligands. Every typo in a four-character id is another valid id, and \"apo\" in a title is not a guarantee -- this project has twice measured against a reference that was not what it claimed. Needs one GraphQL call rather than the REST entry endpoint, whose `nonpolymer_bound_components` is absent on many entries including 5VEW, which has three ligands: REST would report \"no ligands\" for exactly the case the check exists to catch.\n- [x] **Upload an apo structure** from the machine, for anything not in the PDB, copied into the bundle under `reference/`. Two things that would fail silently: the form needs `multipart/form-data` or the browser posts the filename without the file, and the input is named per row rather than as a `[]` array because an empty file input posts nothing and an array would attach each upload to the wrong protein. Contents are sniffed rather than trusted from the extension, filenames sanitised, 20MB ceiling.\n- [x] Fix the UniProt autofill **announcing nothing when it fills a field**. Assigning `.value` from script fires no event, so a partner filled from its accession did not reach the proteins' partner pickers until some later edit triggered a re-sync -- which read as \"partners only appear when I add another one\". Autosave missed it identically: an autofilled name and sequence were never written to browser storage until the next keystroke elsewhere, so filling a protein from its accession and reloading lost it.\n- [x] Write the campaign up in plain English, on the machine that ran it: **Landlord**, narrating each target and the campaign from the Apple Neural Engine where macOS 26 and Apple Intelligence allow it, and from a template everywhere else. No weights ship and no data leaves the machine. Measured before building anything: 0 refusals across 26 fact blocks deliberately loaded with the vocabulary a content filter dislikes, and GPU power *falling* during narration (15.0 to 8.2 mW), so it does not compete with a folding run. Everything that is arithmetic is computed in Python and merely narrated -- asked to compose the campaign tallies the model named the caution count as discards, and both figures were in its input, so the numeric gate could not see it. What the model does write is gated on every number appearing verbatim in its fact block, and a rejected summary falls through to the template. Nothing in the path can fail a campaign. See **Landlord** above, and `docs/landlord_spike.md`, `docs/landlord_bench.md` and `docs/landlord_packaging.md`.\n- [x] Prove the pocket restraint can **fail**, not just succeed: a wrong-pocket arm on every ligand. The tightness series showed a 4 A pocket reproducing a crystal pose, which is only evidence if the same machinery misses when aimed somewhere else. Running each ligand against its own site, the other ligand's site and no site at all put own-site at 0.91/2.79 A, unconstrained at 6.17/13.57 A and wrong-site at 24.93/29.40 A -- worse than no constraint -- on two ligands across two receptors. The conformer column stays low (0.81, 1.78) through the wrong-site failures, so the molecule's shape is right and only its placement is wrong, which is the decomposition the three columns exist for. See **Does the restraint just manufacture the pose?** above.\n- [x] Resolve pocket codes against the chemical component dictionary instead of treating them as opaque labels. A pocket carried through a campaign as the string `41Y` is a PDB component id defining a real molecule, and it turned out to be the same compound as one of that campaign's own ligands -- which had an experimental structure nobody had fetched. Pulling 7RBT into `reference/` doubled the pose panel's coverage and supplied the second ligand of the wrong-pocket control above.\n- [x] Superpose a prediction on a reference that **numbers its residues differently**. Matching on number assumed both sides number the same protein the same way; Boltz numbers a prediction from 1, so a trimmed kinase domain shared no residue with references numbered 223-500 and the pose panel reported \"2GQG is not a structure of ABL1\" while silently comparing nothing. The offset is now found by which shift makes the residue types agree, so a trimmed or engineered construct compares like any other and matching numbering still wins with offset zero. Measured on the trimmed ABL1 campaign: unconstrained imatinib reads 1.31A against 1IEP where the full-length construct gave 18.39A.\n- [x] Stop reporting a **missing interaction analysis as a finding of no contacts**. cif2plip is optional, so a campaign run without it produces no contacts for any target, and the viewer said \"0 contacting residues and the ligand as sticks\" -- an absence of analysis stated as a fact about the ligand, which are opposite claims. The results file now records whether the analysis ran at all, and the viewer says which case it is.\n- [x] Record where a pocket came from, via **`Pocket source: <code> from <PDB>`**. The reference panel previously read \"not recorded\" for any pocket whose provenance lived only in whoever built the spec, and the pose panel could pair a prediction with its experiment only when that structure happened to already be in `reference/`. Reporting only -- it never reaches generate/run.\n- [x] Take **more than one reference structure per protein**, and ship every one of them. `Pocket source:` is now repeatable, and each pocket reference downloaded for its contacts is written into `reference/` instead of being discarded -- previously a pocket's structure travelled only when the same entry happened to be the apo one too, so on any campaign where they differed the pose panel had nothing to score against and said nothing about it. One code claimed by two structures is a parse error rather than last-write-wins.\n- [x] Offer **\"reference molecule only\"** on a pocket reference: ship the structure so its bound ligand can score a matching compound's pose, without defining a site. A site is another run for every ligand; a reference molecule costs nothing. Expressed in the spec as a `Pocket source:` line with no `Pocket contact:` lines, so no new grammar was needed, and listed in the reference panel as \"reference only\" rather than as a site with zero contacts. Measured on ABL1: imatinib had no experimental twin at all until 1IEP was added this way, and then reproduced its crystal pose at 0.56 A constrained against 18.39 A unconstrained.\n- [x] Say which ligands were **not** compared and why. A four-target ABL1 campaign reported \"2 target(s) compared\" and dropped imatinib silently, which read as \"it was never predicted\" rather than \"it had no reference\".\n- [x] Make the **Reference structures** panel ask only the question that applies to the protein in front of it: G-protein coupling for a GPCR, DFG and alphaC for a kinase, and nothing at all for anything else. A kinase reference was reported as \"no G protein bound\" -- true, and as informative as \"no wings\" is of a horse. Family comes from the same `applies_to()` filters that pick the annotator, so a panel cannot label a protein differently from what measured it.\n- [x] Replace the **kinase state classifier**, which was wrong in one direction for every structure it ever saw. DFG-Asp Ca to catalytic-Lys Ca barely moves between states, so an 8 A threshold called dasatinib-bound 2GQG (plainly DFG-in) \"DFG-out\" at 11.15 A -- as it called everything. Now Dunbrack's two-distance test on the DFG-Phe ring, and the Lys-Glu salt bridge for alphaC, which read 2GQG correctly as DFG-in/alphaC-in. alphaC is unresolved rather than \"out\" when side chains are unmodelled, and DFG reports `other` for a conformation that is neither.\n- [x] Put the **pocket code in the target name** (`2_RECP1_GNAS_LIG1_41Y`, `Unc` when unconstrained). In a matrix campaign a ligand is predicted once per pocket, so a bare run number reads as an index rather than a condition; the suffix says what was done differently. Derived from the same value the Pocket column shows, so a name and its column cannot disagree, and skipped for apo targets, which have no pocket.\n- [x] Add **`Class: control` / `Class: experimental`** to a `Ligand:` block, and mark the experimental ones on the pIC50-vs-confidence, pIC50-vs-binder-probability and both ranked charts. Marked by a thin red outline rather than a colour, because colour is already carrying confidence and affinity tiers on those plots and a third meaning would collide with both. A campaign that sets no `Class:` sees no outlines, unchanged from before.\n- [x] Add **Use potentials** as the first Prediction setting, on by default: Boltz's FK steering and physical-guidance coordinate update were hardcoded on, with no way to turn them off from the web form or the CLI. `--no-potentials` now threads through to `boltz predict`, which matters because that guidance update is one of the places a diffusion trajectory can diverge to NaN.\n- [x] Make single-target retries actually isolated: staging one YAML did not scope the\n  run, because `boltz predict` rebuilds its manifest from every processed record and\n  iterates that, so a \"one at a time\" retry re-ran the whole campaign in one process and\n  one raising target took the rest down with it. Each batch now parks the other records\n  for its duration, and a retry pass that completes nothing aborts instead of repeating.\n- [x] Ship `patches/apply_boltz_patches.py` for four `boltz` defects -- containment of\n  numerical failures, a device-correct fp32 alignment guard, a NaN diagnostic that names\n  the offending tensor, and an affinity phase that skips targets with no `pre_affinity`\n  file -- verified by a `boltz_patches` preflight row so a `boltz` upgrade cannot silently\n  revert them.\n- [x] Add `--no-potentials` and default the run settings to the values a real 26-target\n  GPCR campaign needed (`--workers 0`, `--max-msa-seqs 4096`, `--memory-warn-tokens 1500`).\n- [x] Make an OOM skip reclaim memory on Apple silicon (`torch.mps.empty_cache()`), and\n  record each completed target's peak RSS to `.boltzmaker_target_memory.jsonl`.\n- [x] Stop one unusable Pfam mapping discarding every domain for a structure. PDBe returns `author_residue_number: null` when a domain boundary is not observed in the deposited model -- 3 of 6 mappings for `7dty`, including the receptor's own PF02793 -- and that null reached `range()`, raising `'NoneType' object cannot be interpreted as an integer` and losing the usable domains with it. A real GIPR campaign therefore had no motifs and `compare-sse produced no comparable targets`; it now writes 27 rows across both families.\n- [x] Make boltz's full-precision guards work off CUDA. It wraps its numerically fragile steps -- triangular attention, pairformer, attention, the distogram and confidence heads, the encoders and the diffusion sampler -- in `torch.autocast(\"cuda\", enabled=False)`, because they overflow in reduced precision. The device type is hardcoded, so on Apple silicon (autocast device_type `mps`) all **19 sites across 12 files** are inert and those steps run in bfloat16 regardless. Measured on torch 2.10: inside an MPS bf16 autocast a matmul under boltz's own guard still returns bfloat16; with the fix it returns float32. Nesting a second guard keeps it a no-op on CUDA.\n- [x] Make an unattended run need nothing watching it. Recycle the Boltz process every\n  `Targets per invocation` targets (default 4), because Apple's allocator only returns\n  everything on process exit -- held memory floored at ~20GB after one target and grew\n  ~1.9GB per target against a 55.7GB ceiling, which is why a run had 0 out-of-memory\n  skips in its first four targets and 3 in its last four. Apply the boltz patches from\n  BoltzMaker itself rather than only from `run_campaign.sh`, so a direct `all` or a\n  rebuilt environment cannot run unpatched. Stop a Boltz that has said nothing for an\n  hour and hand its targets to the retry ladder, instead of waiting on a wedged worker\n  that will never exit. Together these retire the external watchdog.\n- [x] Make a finished campaign read as an answer rather than a listing: Affinity ahead\n  of Confidence, rows ranked by predicted pIC50 (apo rows sinking, ties stable),\n  click-to-sort on every column of the summary and SSE tables with the indicator shown\n  up front, and a Pockets panel naming each constraint and the ligands run against it.\n  The analysis page lifts these panels out of each run's stored dashboard rather than\n  building them, so `web/deploy/refresh_run_reports.py` re-renders the two that are a\n  pure function of a bundle's `boltz_input.md` and `boltz_summary.csv` -- leaving every\n  other panel byte-for-byte -- and already-uploaded runs get the change too. Clearing\n  the server's `panels.json` caches is part of that: 61 of them would otherwise have\n  gone on serving the old panels.\n- [x] Put a viewer under the Pockets table: every receptor superposed in one grey,\n  every ligand in its pocket's colour (pocket 1 green, pocket 2 blue, the\n  unconstrained baseline always red), a checkbox per target, and the same spin/reset\n  controls as the Superposed targets pane. Only targets that bind something are drawn\n  -- an apo target has no pose to place -- so the 15-target 5-HT2 campaign shows its\n  12. Which pocket a target used is recovered from its own stem, matching the whole of\n  `family_ligand_code` rather than the suffix, so a ligand that happens to share a\n  pocket's name is not misread as a constrained run.\n- [x] Settle on one user-facing vocabulary -- **proteins, co-folded partners, ligands,\n  pockets, ligand-free companions, predictions** -- and use it everywhere. The summary\n  table headed its receptor column \"Target\" while the list below it called a prediction\n  a target: one word for two things on one page. Proteins are counted by group, so the\n  5-HT2 panel's nine `Protein:` blocks are its three receptors. The analysis page opens\n  with those six as KPI boxes, in the order the prepare form asks for them, and the\n  campaign summary reports the same six numbers.\n- [x] Rank the summary table by predicted pIC50 rather than by binder probability.\n  Measured across 24 ligand targets on the GLP1R/GIPR campaign, the binary head\n  correlated with pIC50 at **r = +0.07** and returned nothing above 0.71 for a set of\n  four real, characterised compounds. Its mean was flat across ligands (0.34-0.51)\n  while tracking the *pocket condition* instead (0.30 unconstrained, 0.50-0.51\n  constrained) -- it was reporting how a target was set up, not what was in the site.\n  Orforglipron, a confirmed potent GLP1R binder, scored 0.59 and ranked below\n  compounds it beats experimentally by orders of magnitude. pIC50 put the same four in\n  their known order (orforglipron 11.2, then the other three at 9.8, 8.4 and 7.9), so\n  that is what the table ranks on. A binary binder/non-binder score also has nothing to separate when\n  every ligand in a campaign is a real binder.\n- [ ] Classify the failure before retrying, and escalate along the axis that addresses it:\n  a memory ladder (isolation, then `--max-msa-seqs 2048`) and a NaN ladder\n  (`--no-potentials`, then fp32). Retrying with identical parameters is already known to be\n  worthless -- 20 invocations produced nothing. Any rung past the first changes the science,\n  so the rung used has to appear in the report. Do not vary `--mps-watermark` (a hard\n  ceiling, not a pressure valve), `--workers` (already at its floor of 0), or the seed (the\n  failing run was unseeded and still failed identically 20/20).\n- [ ] Re-derive `preflight`'s size check from the recorded peaks rather than\n  `--memory-warn-tokens`, so it reports \"this is the size of a target that peaked at 61GB\n  here\" instead of a threshold someone guessed.\n- [x] Recover `GLP1R_orfo`, the one target whose diffusion diverges. `--no-potentials`\n  confirmed the potentials were the cause (16.8A vs 2154A), but sampling three targets\n  differently from the other 21 is not a fix. Two clamps contain it instead -- a\n  gradient bound anchored to the trajectory's quietest scale, and a 0.1A cap on how far\n  one guidance step may move an atom -- and all three now land within 2A of that\n  control with steering on, so the whole matrix is comparable.\n- [ ] Work out whether the NaN divergence is bf16-specific by exposing Boltz's hardcoded\n  `precision=\"bf16-mixed\"` as an option, rather than patching it per incident.\n- [x] Run a real GPU campaign end to end -- prepared on the site, run from the bundle, packed, uploaded and explored. Four bundle defects had reached a user before this (a preflight timeout too short for a cold torch import, `sse_comparison` and `vendor` missing from the bundle, pre-1980 zip timestamps, `sh` portability), because everything up to the campaign is verified automatically while the `pixi install` -> `boltz predict` -> `analyze` path is not. Worth repeating before each release\n- [x] Show everything the dashboard computes in the hosted Analysis step, not only the two panels the explorer draws itself. BoltzMaker's own generated reports now travel inside the `.bmz` and are framed beside the interactive panels, so the hosted view cannot drift from the offline one\n- [x] Give every protein an apo reference so compare-sse can actually run from the web form: a predicted ligand-free companion by default, an experimental PDB entry when one is named (mmCIF preferred, since most recent cryo-EM entries have no legacy file), or both. Previously the form emitted no `Apo structure:` at all, so the comparison never ran and its skip option had nothing to skip\n- [x] Add **Keep private** and a **Runs** tab: the decision travels in the user's own files (the bundle records it, the packer copies it into the results manifest) so a private run leaves nothing on the server to list, while everything else is archived with its bundle and results under a capped, self-pruning archive\n- [x] Add single-keypress run controls: `p` pauses and resumes via a real SIGSTOP of the whole Boltz process tree, so an interrupted target is neither discarded nor recomputed, and `q` stops the run and all its workers cleanly. Fixed a leak found on the way: terminating only the parent left dataloader workers alive holding RAM and the GPU\n- [x] Replace the run's dead `-:--:--` with an ETA that says where it came from -- measured from this run once a target completes, otherwise the median seconds-per-target of past runs of the same campaign on the same accelerator -- and make the phase row honest about the fact that Boltz reports nothing between dataloader items\n- [x] Preflight the bundle's own contents (`sse_comparison`, the vendored dashboard assets, the results packer), since all three are read only after the prediction finishes and a missing one used to cost the whole run before announcing itself\n- [x] Keep everything typed into the Prepare form across a download, a visit to Analysis or a reload, and let it be saved to and loaded from a small `.json` file\n- [x] Split the hosted site into two modes: a **Fully Automated** two-step flow (configure a campaign, download one self-extracting bundle carrying the software, environment and scripts; run it locally; upload the single `.bmz` results file it writes) and the original **Stepwise** four-tool flow, chosen from a two-panel landing page\n- [x] Give the Analysis step real exploration rather than an embedded report: a sortable, filterable target table, a confidence-versus-affinity triage plot coloured by BoltzMaker's own flags, and a per-target 3Dmol.js pose viewer with the detected PLIP interactions beside it -- which also lands the cross-target selectivity/triage view below\n- [x] Host the non-GPU stages (`new` wizard, `generate`, `preflight`, full `analyze` including PLIP and compare-sse) as a Flask app at [boltzmaker.mdeller.com](https://boltzmaker.mdeller.com), so the pipeline can be tried with no local install -- the GPU `run` step stays on your own hardware by design\n- [x] Add `preflight --json` so the check results can be consumed by tooling rather than scraped from the rich table, with the banner suppressed alongside it (as `format` already did) to keep stdout parseable\n- [x] Resolve the right micromamba build from the running platform in `setup-plip`, so Linux (`linux-64`/`linux-aarch64`) works rather than silently fetching a macOS binary on an x86_64 Linux host\n- [x] Fix compare-sse's DSSP fallback for real campaigns: multi-character family ids no longer overflow the legacy-PDB chain-ID field, and the temporary structure now carries the `HEADER` line `mkdssp` >= 4 needs to recognise it as PDB at all\n- [x] Pin exact dependency versions and add a cached/offline install mode for reproducible installs -- via `pixi.toml`/`pixi.lock` and `install.sh` (see **Alternative: pixi** above and `docs/tier_b_offline_install.md`), not by pinning `setup`/`setup-plip`'s own unpinned `requirements.txt`-driven installs, which remain as they were\n- [x] Add a cross-target selectivity/triage view (confidence-vs-affinity quadrant flags for \"high-confidence, high-affinity\" hits) -- delivered in Fully Automated Mode's Analysis explorer rather than in the offline dashboard; it plots BoltzMaker's own `flags` and draws only the absolute 0.5 confidence cutoff, deliberately not a horizontal affinity threshold, because the mismatch flags are within-campaign terciles rather than absolute cutoffs\n- [x] Bundle Plotly.js locally instead of via CDN so the dashboard renders fully offline/air-gapped\n- [x] Add a ligand-preparation/validation step (canonicalization, stereocentre/protonation-state flagging, disconnected-fragment detection) so bad input chemistry is caught before hours of compute, not silently mispredicted\n- [x] Add a 3Dmol.js rotating structure view to each target's binding-site panel, placed next to the existing static PyMOL image (keep the static image, add a \"Download image\" link for it -- doesn't exist yet -- and add the interactive 3Dmol.js view alongside rather than replacing anything)\n- [ ] Share pip cache between the two environments (`PIP_CACHE_DIR`) so `setup-plip` doesn't re-download wheels the main venv already fetched\n- [ ] Add `BoltzMaker.py doctor` -- a post-install check that imports boltz/rdkit/plip/openbabel/pymol in-process and reports exactly which env/feature is broken\n- [ ] Add an explicit Boltz model-weights cache dir + a `preflight` check for it (ties into the existing iCloud dataless-file eviction check)\n- [ ] Detect the MPS `torch.linalg.svd` CPU-fallback at `preflight` and warn with an estimated runtime multiplier for large multi-chain complexes\n- [ ] Add a residue/chain-count-based runtime pre-estimator, plus a toggle to log which ops fall back to CPU on MPS\n- [ ] Checkpoint `run` at the per-sample level (not just per-target) so an interrupted multi-hour complex resumes without recomputing completed diffusion samples\n- [ ] Add a smoke-test suite: an end-to-end fixture run in CI plus unit tests for the `boltz_input.md` parser and JSON-metric flattening\n- [ ] Add a retrospective `benchmark` mode: pull known actives/co-crystal data for a target family (ChEMBL/BindingDB/PDB) and report predicted-vs-measured pIC50 correlation + pose RMSD, so a user has a per-target-family trust score before committing to a real campaign\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/BoltzMaker\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/BoltzMaker</a></td>\n</tr>\n</table>", "# Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\nAuthors: Ye, Q.; Shvartsbart, A.; Li, Z.; Gan, P.; Policarpo, R.L.; Qi, C.; Roach, J.J.; Zhu, W.; McCammant, M.S.; Hu, B.; Li, G.; Yin, H.; Carlsen, P.; Hoang, G.; Zhao, L.; Susick, R.; Zhang, F.; Lai, C.-T.; Allali Hassani, A.; Epling, L.B.; Gallion, A.; Kurzeja-Lipinski, K.; Gallagher, K.; Roman, V.; Farren, M.R.; Kong, W.; Deller, M.C.; Zhang, G.; Covington, M.; Diamond, S.; Kim, S.; Yao, W.; Sokolsky, A.; Wang, X.\nSource: Journal of Medicinal Chemistry, 2025\n\nAttention therefore turned to rigidifying the structure. Structural rigidity, as measured by the number of rotatable bonds, has been linked to a number of beneficial properties, from increases in binding potency to improvement in oral bioavailability. As mentioned above, two areas of the molecule with the highest number of rotatable bonds were the cyanoethyl fragment and the amide tail. To evaluate which area was more likely to lead to significant gains we performed 1 \u03bcs molecular dynamics simulations on the ground state of a simplified analog of compound 10. An overlay of all energetically available conformations are shown in Figure 5 and point to the ethyl amide fragment as a significantly more attractive location for additional rigidity. Due to the sterically congested environment around the cyanoethyl, there are two main conformational ensembles observed, with the ensemble containing the largest number of conformations corresponding to the bioactive conformation, which helps explain the significant potency gains realized upon its installation. Conversely, the ethyl amide projects into a sterically unencumbered area and the bioactive conformation represents a small sliver of the overall conformations sampled. It was therefore hypothesized that locking the ethyl amide in the bioactive conformation would provide larger benefits to potency.\n\n*Figure 5.* Ground state molecular dynamics simulation of compound 10.\n\n*Table 2.* Mutation Site Cyclization SAR. All experiments are an average of at least n = 2 unless otherwise noted, see Supporting Information for details.\n\n*(the body of Table 2 is an image-based structure table and did not survive text extraction from the PDF; the compounds and trends it summarises are described in the surrounding text)*\n\nA number of cyclization approaches were attempted, with the goal of keeping a two atom linker between the core and the amide, while keeping the hydroxynaphthyl Switch II piece constant. After significant screening, we discovered that \u03b3-lactam 14 improved potency 5 fold in the binding assay, at the cost of translation to hWB (Table 2). Switching the ethoxy substituent to methyl restored the hWB potency, which proved to be a consistent trend throughout the series.\n\nGratifyingly, compound 16, with the 3-hydroxynaphthol in 15 replaced with 7-fluoronaphthyl, was the first to maintain WB pERK potency within 2-fold of 10, while meeting both objectives 1 and 2 outlined above. Related compound 17, with an extra oxygen in the mutation site ring, further improved cellular potency to 7.2 nM and hWB potency to 677 nM. Encouragingly, cyclization improved permeability in the Caco-2 assay (0.7 \u00d7 10\u207b\u2076 cm/s). This was the first time permeability was observed in this series, suggesting we had finally identified an appropriate balance of polarity. Indeed, we were able to quickly drive further improvements in properties through examination of the rest of the molecule (Table 3). The solvent was first examined, freed from the requirement of a basic amine. A number of derivatives showed meaningful improvements in permeability while maintaining potency, with compound 18 showing a Caco-2 value of 1.6 \u00d7 10\u207b\u2076 cm/s. Similarly, clearance was also beneficially impacted by heteroaryl substitution, with compounds 18 and 20 showing 1.0 and 0.7 L/h/kg, respectively. This was the first time in the program that permeability and clearance could be successfully combined in a single compound (18), validating our hypothesis that only by removing both the hydroxynaphthyl and basic amine in the solvent could we obtain a balanced ADME profile.\n\n*Table 3.* SAR Around Compound 17.\n\n| Compound | R1 | X | R2 | SPR (nM) | Exc. (nM) | pERK (nM) | hWB pERK (nM) | Caco (\u00d710\u207b\u2076 cm/s) | h-Cl (L/h/kg) |\n|---|---|---|---|---|---|---|---|---|---|\n| 17 | A | H | Me | N/A | 2.7 | 7.2 | 677 | 0.7 | >1.1 |\n| 18 | A | H | C | 1.2 | 6.9 | 36 | 565 | 1.6 | 1.0 |\n| 19 | A | H | D | 0.84 | 3.7 | 10 | 519 | 1.7 | >1.1 |\n| 20 | A | H | E | N/A | 3.8 | 17 | 626 | 1.1 | 0.7 |\n| 21 | A | (R)-Me | C | 0.98 | 2.1 | 7.5 | 103 | 0.6 | >1.1 |\n| 22 | A | (R)-Me | F | 1.5 | 2.8 | 6.8 | 55 | 1.4 | >1.1 |\n| 23 (INCB159020) | B | (R)-Me | F | 22 | 4.4 | 33 | 331 | 3.6 | 1.0 |\n\nAll experiments are an average of at least n = 2 unless otherwise noted, see SI for details. N = 1 for compound 17 in the hWB pERK assay and for compounds 18 and 19 in the h-Cl assay. Compounds 18, 19 and 20 were tested as a mixture of atropisomers; N/A = not tested.\n\nEncouraged by in vitro ADME properties, we progressed compounds 18 and 20 into rat PK studies (Table 4). Disappointingly, both compounds showed very high in vivo clearance (>100% hepatic blood flow) and low oral bioavailability (<20%). To ensure that no species specific metabolism issues were present, compound 18 was also dosed in nonhuman primate (NHP). Here the results were mixed. Although oral exposure was below the quantifiable limit, in the i.v. arm, compound 18 showed reasonable clearance (0.97 L/h/kg, 37% hepatic blood flow) and a half-life of 8.4 h.\n\nThe complete lack of oral bioavailability in spite of moderate clearance and reasonable permeability in the Caco-2 assay prompted a search for alternative explanations. After extensive profiling, it was discovered that compound 20 suffers significant metabolism in the intestine. In addition, an efflux liability (efflux ratio = 20), albeit saturable at higher concentrations, was observed.\n\n*Table 4.* PK Parameters for Selected Compounds. Full details available in the Supporting Information.\n\n| Compound | Species | Dose (iv/po) (mg/kg) | Cmax (nM) | AUC (nM\u00b7h) | %F | Cl (%HBF) | t1/2 (h) |\n|---|---|---|---|---|---|---|---|\n| 18 | rat | 1.0/3.0 | 41.5 | 108 | 11 | >100 | 3.5 |\n| 18 | NHP | 0.5/2.0 | 2.5 | BQL | 0 | 37 | 8.4 |\n| 22 | rat | 0.1/3.0 | 24 | 85 | 20 | >100 | 2.5 |\n| 22 | NHP | 1.0/20 | 7.2 | 68 | 2 | 39 | 4.9 |\n| 23 | rat | 1.0/3.0 | 57 | 331 | 27 | >100 | 3.8 |\n| 23 | NHP | 1.0/1.5 | 100 | 952 | 42 | 37 | 6.4 |", "# marcdeller.com: Structural Biology\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n## Beyond Pretty Pictures: Structural Biology Driving Drug Discovery\n\nFor me, protein structures still anchor everything that follows in drug discovery. Knowing exactly how a ligand binds, where a cryptic pocket opens, or how a conformational switch flips is what separates a beautiful figure from a usable lead. Modern computational tools, including AlphaFold, RFDiffusion, Boltz-2, and ProteinMPNN, now extend that reach into target classes and design spaces that were out of reach a decade ago. The interesting question is no longer whether we can predict or design proteins, but how to pair those predictions with the experimental discipline that keeps them honest. That pairing, more than any single tool, is what moves a programme forward.\n\n* **Precision:** Unlocks atomic-level binding landscapes for designing ultra-targeted compounds. This eliminates guesswork from lead optimization, enabling chemists to engineer specificity and reduce off-target liabilities.\n* **Prediction**: Leverages machine learning to model complex protein folds in hours, not months. Teams can rapidly prioritize tractable targets and identify cryptic binding pockets by integrating predictions with experimental validation.\n* **Acceleration:** Streamlines discovery workflows, slashing evaluation times by integrating structural data into design-make-test-analyse cycles. Automated triage compresses the path from hit to development candidate.\n* **Innovation:** Applies adaptive algorithms to reveal novel therapeutic scaffolds and optimize efficacy. This approach interrogates the full breadth of protein-ligand interactions, enabling first-in-class molecules beyond known chemical space.\n\nTop Posts\n\n### [Supercharging Discovery: How LLM Prompts Are Transforming Structural Biology and Cheminformatics](https://marcdeller.com/supercharging-discovery-how-llm-prompts-are-transforming-structural-biology-and-cheminformatics/)\n\n[Supercharging Discovery: How LLM Prompts Are Transforming Structural Biology and Cheminformatics](https://marcdeller.com/supercharging-discovery-how-llm-prompts-are-transforming-structural-biology-and-cheminformatics/)\n\n### [Multi-Step LLM Prompt: From Protein Sequence to Validated Protein Structure](https://marcdeller.com/multi-step-llm-prompt-from-protein-sequence-to-validated-protein-structure/)\n\n[Multi-Step LLM Prompt: From Protein Sequence to Validated Protein Structure](https://marcdeller.com/multi-step-llm-prompt-from-protein-sequence-to-validated-protein-structure/)\n\n### [Revolutionizing Protein Structure Analysis: Cryo-EM without Purification Accelerates the Future of Structural Biology](https://marcdeller.com/revolutionizing-protein-structure-analysis-cryo-em-without-purification-accelerates-the-future-of-structural-biology/)\n\n[Revolutionizing Protein Structure Analysis: Cryo-EM without Purification Accelerates the Future of Structural Biology](https://marcdeller.com/revolutionizing-protein-structure-analysis-cryo-em-without-purification-accelerates-the-future-of-structural-biology/)\n\n### [Shedding Light on the Shadows: The Untapped Potential of the Dark Kinome](https://marcdeller.com/shedding-light-on-the-shadows-the-untapped-potential-of-the-dark-kinome/)\n\n[Shedding Light on the Shadows: The Untapped Potential of the Dark Kinome](https://marcdeller.com/shedding-light-on-the-shadows-the-untapped-potential-of-the-dark-kinome/)\n\n### [Engineering Protein Stability: Unlocking the Secrets of Successful Crystallization](https://marcdeller.com/engineering-protein-stability-unlocking-the-secrets-of-successful-crystallization/)\n\n[Engineering Protein Stability: Unlocking the Secrets of Successful Crystallization](https://marcdeller.com/engineering-protein-stability-unlocking-the-secrets-of-successful-crystallization/)\n\nAI\n\n### [BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline](https://marcdeller.com/boltzmaker-vibe-coding-a-campaign-scale-boltz-2-pipeline/)\n\nAI\n\n### [ChemSage: Vibe-Coding a Chemically Aware Drug-Discovery LLM on Local Hardware](https://marcdeller.com/chemsage/)\n\nAI\n\n### [OpenBind Publishes a Dense Structure-Affinity Dataset for Structure-Based AI](https://marcdeller.com/openbind-publishes-a-dense-structure-affinity-dataset-for-structure-based-ai/)\n\nBlog\n\n### [When the Wet Lab Becomes the Bottleneck: Self-Driving Platforms for Protein Engineering](https://marcdeller.com/when-the-wet-lab-becomes-the-bottleneck-self-driving-platforms-for-antibody-engineering/)\n\nBlog\n\n### [From Noise to Nanomolar: PXDesign\u2019s New Benchmark for Protein Binder Design](https://marcdeller.com/from-noise-to-nanomolar-pxdesigns-new-benchmark-for-protein-binder-design/)\n\nBlog\n\n### [When the Enzyme Is Its Own Blueprint: Protein-Templated DNA Synthesis in DRT3](https://marcdeller.com/when-the-enzyme-is-its-own-blueprint-protein-templated-dna-synthesis-in-drt3/)\n\nBlog\n\n### [An Enzyme That Eats Plastic \u2014 Inside Your Blood](https://marcdeller.com/kth/)\n\nMicroplastics\n\n### [PETase: Nature\u2019s Solution to a Human-Made Problem](https://marcdeller.com/petase-natures-solution-to-a-human-made-problem/)\n\nSB\n\n### [AXIS: Closing the Loop on Crystal Detection](https://marcdeller.com/axis-closing-the-loop-on-crystal-detection/)\n\nSB\n\n### [Molecular Architecture of GLP-1 Receptor Drug Binding Reveals Dynamic Activation Mechanisms](https://marcdeller.com/molecular-architecture-of-glp-1-receptor-drug-binding-reveals-dynamic-activation-mechanisms/)\n\nSB\n\n### [Transforming Antibody Design: How Artificial Intelligence Revolutionizes Structure Prediction](https://marcdeller.com/transforming-antibody-design-how-artificial-intelligence-revolutionizes-structure-prediction/)\n\nSB\n\n### [Self-Encoded Libraries: A New Era in Drug Discovery Screening](https://marcdeller.com/eliminating-dna-barcodes-a-new-era-in-drug-discovery-screening/)\n\nBlog\n\n### [Tuning the Signal: Rational Design of G Protein Selectivity](https://marcdeller.com/tuning-the-signal-rational-design-of-g-protein-selectivity/)\n\nBlog", "# Microplastics and Health: Alarming New Research Reveals Widespread Biological Impact\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-17 on marcdeller.com (categories: blog, microplastics).*\n\n* **Intestinal Impact**: High-concentration polystyrene microplastics disrupt gut microbiota composition and induce intestinal inflammation through NK cell activation, leading to metabolic dysfunction even without leaky gut syndrome.\n* **Liver Toxicity**: Polyethylene microplastics promote hepatic fibrosis and metabolic disturbances by altering lipid metabolism pathways and activating hepatic stellate cells in both healthy and pathological conditions.\n* **Gut Barrier Dysfunction**: Polyethylene terephthalate microplastics cause significant intestinal damage with inflammatory cell infiltration, reduced mucus secretion, and compromised barrier integrity alongside microbiota dysbiosis.\n* **Cardiovascular Effects**: Polystyrene microplastics induce severe myocardial inflammation and cell death in aquatic organisms through TLR4/NF-\u03baB pathway activation, suggesting potential cardiovascular risks.\n* **Respiratory Toxicity**: Polystyrene microplastic fragments specifically trigger pulmonary inflammation via TLR4-mediated pathways, while other plastic types show minimal respiratory effects.\n* **Systemic Bioaccumulation**: Size-dependent uptake patterns show smaller microplastics accumulate in multiple organs including liver, inducing oxidative stress and metabolic disruption across biological systems.\n\n1. **Hasegawa, Y., et al. (2024).** Oral exposure to high concentrations of polystyrene microplastics alters the intestinal environment and metabolic outcomes in mice. Frontiers in Immunology, 15:1407936. <https://doi.org/10.3389/fimmu.2024.1407936>\n2. **Djouina, M., et al. (2023).** Oral exposure to polyethylene microplastics induces inflammatory and metabolic changes and promotes fibrosis in mouse liver. Ecotoxicology and Environmental Safety, 264:115417. <https://doi.org/10.1016/j.ecoenv.2023.115417>\n3. **Sun, X., et al. (2025).** Polyethylene terephthalate microplastics affect gut microbiota distribution and intestinal damage in mice. Ecotoxicology and Environmental Safety, 294:118119. <https://doi.org/10.1016/j.ecoenv.2025.118119>\n4. **Zhang, Q., et al. (2023).** Polystyrene microplastics induce myocardial inflammation and cell death via the TLR4/NF-\u03baB pathway in carp. Fish and Shellfish Immunology, 135:108690. <https://doi.org/10.1016/j.fsi.2023.108690>\n5. **Danso, I.K., et al. (2024).** Pulmonary toxicity assessment of polypropylene, polystyrene, and polyethylene microplastic fragments in mice. Toxicological Research, 40:313-323. <https://doi.org/10.1007/s43188-023-00224-x>\n6. **Lu, Y., et al. (2016).** Uptake and Accumulation of Polystyrene Microplastics in Zebrafish (Danio rerio) and Toxic Effects in Liver. Environmental Science & Technology, 50:4054-4060. <https://doi.org/10.1021/acs.est.6b00183>\n\nSix groundbreaking research studies reveal the extensive biological impact of microplastics on mammalian health systems. The first study demonstrates that high concentrations of polystyrene microplastics alter intestinal environments and metabolic outcomes in mice, inducing inflammation and dysbiosis through NK cell activation. The second study shows that polyethylene microplastics cause inflammatory and metabolic changes while promoting liver fibrosis in mice through disrupted lipid metabolism and HSC activation. Research on polyethylene terephthalate microplastics reveals significant gut microbiota disruption and intestinal damage, with inflammatory responses and barrier dysfunction in exposed mice. A fourth study investigates polystyrene microplastics\u2019 cardiotoxic effects in carp, showing myocardial inflammation and cell death via TLR4/NF-\u03baB pathway activation. Pulmonary toxicity research demonstrates that polystyrene fragments specifically induce lung inflammation through TLR4-mediated NF-\u03baB and NLRP3 inflammasome activation, while polypropylene and polyethylene show minimal effects. Finally, zebrafish studies reveal size-dependent uptake and accumulation of polystyrene microplastics, with liver toxicity manifesting as oxidative stress, inflammation, and disrupted lipid metabolism. Collectively, these studies establish clear dose-dependent relationships between microplastic exposure and systemic health impacts across multiple organ systems, highlighting urgent needs for environmental regulation and continued research into human health implications.\n\n# 1) Oral exposure to high concentrations of polystyrene microplastics alters the intestinal environment and metabolic outcomes in mice\n\n* **Metabolic Impact**: High-concentration PS-MPs induced dyslipidemia and fatty liver disease without causing leaky gut syndrome, demonstrating alternative toxicity pathways beyond barrier disruption.\n* **Immune Response**: PS-MPs exposure significantly increased NK cell populations in intestinal tissues, triggering inflammation and contributing to metabolic dysfunction through immune-mediated mechanisms.\n* **Microbiota Disruption**: Exposure caused dose-dependent changes in gut microbiota composition with decreased diversity, reduced beneficial bacteria, and altered metabolic profiles affecting host health.\n* **Gene Expression Changes**: PS-MPs modulated intestinal gene expression related to inflammation, nutrient transport, and metabolic processes, indicating molecular-level disruption of normal physiological functions.\n* **Dose-Dependent Effects**: The severity of intestinal and metabolic disturbances increased with higher PS-MPs concentrations, establishing clear dose-response relationships for toxicity assessment.\n* **Novel Pathway Discovery**: The study identified NK cell-mediated intestinal inflammation as a new mechanism of microplastic toxicity, expanding understanding beyond traditional barrier function disruption.", "# pants.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\nEvery measured value in this database points at the paper that reported it.\n**2757** of\n**2759** measurements carry a DOI or\nPubMed identifier, drawn from\n**38** distinct sources.\n\n### What was measured\n\n| Quantity | Values |\n| --- | --- |\n| percent depolymerization | 2142 |\n| Tm the temperature at which the fold falls apart | 359 |\n| Product release how much PET breakdown product appeared | 183 |\n| KM how much substrate it takes to half-saturate the enzyme | 21 |\n| Topt the temperature at which the enzyme works fastest | 20 |\n| Performance claim a stated result from the paper \u2014 \"6-fold faster than the wild type\" \u2014 kept as text because it has no single unit | 11 |\n| Catalytic activity product formed under the assay conditions the paper used | 10 |\n| pHopt the acidity at which it works fastest | 8 |\n| Ordinal activity a rank rather than a number: this enzyme beat that one, with no value given | 5 |\n\n### Strength of evidence\n\n| Evidence code | Values |\n| --- | --- |\n| Experimental, from a publication ECO:0000269 | 2737 |\n| Inferred by a curator ECO:0000305 | 22 |\n\nEvidence Ontology codes. The first\nmeans somebody ran the assay; the second that a curator collated it from a review.\n\n## Models and software\n\nRecorded per pipeline run rather than declared here, so this is what actually\nproduced the data rather than what a README says it did.\n\n### Machine-learning models\n\n|  |  |\n| --- | --- |\n| **ESM-2 t12-35M** a protein language model. Turns a sequence into 480 numbers capturing what it has learned about proteins; frozen, never fine-tuned here | embedding |\n| **ESMFold v1** predicts a 3D structure from sequence alone, no alignment needed. 8.4 GB of weights, run on CPU | folding |\n| **AlphaFold DB** precomputed structures, fetched rather than run | folding |\n\nNo large language model is involved in\nproducing any number on this site.\n\n### Tool versions in use\n\n|  |  |\n| --- | --- |\n| biotite | 1.7.1 |\n| gemmi | 0.7.5 |\n| hmmer | HMMER 3.4 (Aug 2023); http://hmmer.org/ |\n| mmseqs2 | 18-8cc5c |\n| numpy | 2.5.1 |\n| python | 3.14.3 |\n| sklearn | 1.9.0 |\n| torch | 2.13.0 |\n| transformers | 5.14.1 |\n\nSchema v17\n\u00b7 code at commit dc28608\n\n## External databases\n\nPublic resources this project reads from, with the date each was last retrieved.\nNothing is scraped continuously: a source is pulled when a stage runs, and the date below is when\nthat last happened.\n\n| Source | What it provides | Records | Retrieved | Licence |\n| --- | --- | --- | --- | --- |\n| **ACS-screen** | Norton-Baker, Komp, Gado et al. 2025, ACS Catal. 15:16070-16083 | 114 | 2026-08-06 | CC BY 4.0 (author-deposited Supporting Information) |\n| **HGMP-SciDB** | human gut metagenome polyesterases from the deposit accompanying the paper | 5 | 2026-08-04 | see publication |\n| **PAZy** | the plastics-active enzyme database: an enzyme is listed because activity on a plastic was measured and published | 320 | 2026-08-06 | see publication |\n| **Science-landscape** | Landscape profiling of PET depolymerases, Science 2025, Data S3 | 94 | 2026-08-06 | see publication |\n| **Science-landscape-S1** | Science 2025 Data S1, homologue search space | 25041 | 2026-08-06 | see publication |\n| **Science-landscape-S7** | Science 2025 Data S7, per-cluster ecological context | 170 | 2026-08-06 | see publication |\n| **UniProt** | curated protein sequences, annotations and catalytic-site positions | 6 | 2026-08-05 | CC BY 4.0 |\n| **RCSB PDB** | experimental crystal structures and their metadata | 60 | per structure build | CC0 |\n| **AlphaFold DB** | predicted structures by UniProt accession | 104 | per structure build | CC BY 4.0 |\n| **Crossref** | DOI verification \u2014 every citation on this site was resolved and its title, journal and year checked | 38 | on curation | open |\n| **MGnify** | metagenomic assemblies: the raw sequence space searched, counted here as environments rather than records | 5 | per scan | open |\n\n## Pipeline activity\n\nEvery stage writes a record of what went in, what came out, which tool versions\nran and how long it took. **124** runs and\n**121** manifests so far, totalling\n**32.2** hours of recorded compute.\n\n| Stage | What it does | Runs | In | Out | Last run |\n| --- | --- | --- | --- | --- | --- |\n| recall | search metagenome proteins for the polyesterase fold, then keep only those whose catalytic residues meet in space | 65 | 14,804,920 | 520 | 2026-08-05 05:10 |\n| reference_structures | fetch or fold a structure for each characterised enzyme and superpose it onto IsPETase | 33 | 672 | 672 | 2026-08-07 22:22 |\n| seeds | load the curated wild types and derive each engineered variant from its parent plus a checked mutation list | 10 | 168 | 152 | 2026-08-05 23:41 |\n| pazy | import enzymes with published, measured activity on a plastic | 4 | 1,936 | 679 | 2026-08-06 00:11 |\n| embed | turn every sequence into 480 numbers with a frozen protein language model | 4 | 4,002 | 4,002 | 2026-08-06 22:15 |\n| structure | fold candidates with ESMFold and measure the active site | 2 | 128 | 88 | 2026-08-04 23:30 |\n| science_supplements | \u2014 | 1 | 25,411 | 25,053 | 2026-08-06 22:30 |\n| science_landscape | \u2014 | 1 | 183 | 183 | 2026-08-06 22:08 |\n| positives_family | add the wider enzyme family as similarity-labelled positives | 1 | 83 | 79 | 2026-08-04 12:39 |\n| negatives | collect hard negatives: the same fold, no PET activity | 1 | 3,985 | 256 | 2026-08-04 12:41 |\n| harmonise | reconcile activity values reported under different assays | 1 | 459 | 47 | 2026-08-04 12:39 |\n| acs_screen | \u2014 | 1 | 477 | 214 | 2026-08-06 22:04 |\n\n## Model evaluations\n\nEvery training run ever recorded, including the ones that failed. AUC is the\nchance the model ranks a true PET enzyme above a non-enzyme: 1.0 is perfect, 0.5 is a coin flip.\nThe **composition baseline** beside it is what you get from amino-acid counts alone \u2014\na score only means something if it clears that.", "# Elevating Science Communication: Effortless 3D Structure Embedding with 3Dmol.js\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\ndata-pdb='1EVS'\ndata-backgroundcolor='0xffffff'\ndata-ui='true'\ndata-style='stick'\ndata-surface='opacity:.7;color:white'\ndata-zoom-to=\"true\"\ndata-spin='axis:y;speed:2'\n\n&gt;&lt;/div&gt;\n\n```\n\n```\n\n\t\t\t\t\t&lt;script src=\"https://3Dmol.org/build/3Dmol-min.js\"&gt;&lt;/script&gt;\n&lt;script src=\"https://3Dmol.org/build/3Dmol.ui-min.js\"&gt;&lt;/script&gt;\n\n&lt;div style=\"height: 400px; width: 400px; position: relative;\" class='viewer_3Dmoljs'\n\ndata-cid='2244'\ndata-backgroundcolor='0xffffff'\ndata-ui='true'\ndata-style='stick'\ndata-spin='axis:y;speed:2'\n\n&gt;&lt;/div&gt;\n\n```\n\nThe ease-of-use, security, and verification for scientific web development with 3Dmol.js make it the leading choice for both teaching and research. Inline customization tags and verified code examples, combined with open-source community support, cement 3Dmol.js as the future of molecular visualization on the web.\n\nConfigurations like spin, zoom, UI, colors, style, and surface rendering can be accessed using \u201c[**data-tags**](https://3dmol.org/doc/tutorial-embeddable.html)\u201d as detailed in the table below.\n\n| Key Concept | Description | Example |\n| --- | --- | --- |\n| data-pdb | PDB ID to be loaded into the viewer | data-pdb=\u20191YCR\u2019 |\n| data-cid | PubChem compound id to be loaded | data-cid=\u20192244\u2032 |\n| data-href | URL to molecular data file | data-href=\u2019https://files.rcsb.org/download/2POR.pdb\u2019 |\n| data-element | id of HTML element with embedded molecular data | data-element=\u2019moldata\u2019 |\n| data-type | File format (can be pdb, sdf, xyz, mol2, cube) | data-type=\u2019sdf\u2019 |\n| data-backgroundcolor | Background color (default: black) | data-backgroundcolor=\u20190xffffff\u2019 |\n| data-backgroundalpha | Background alpha (default: 1.0 opaque) | data-backgroundalpha=\u20190.5\u2032 |\n| data-select | AtomSpec selection specification | data-select=\u2019chain:A\u2019 |\n| data-style | Style specification | data-style=\u2019cartoon:color=spectrum\u2019 |\n| data-surface | Surface style specification | data-surface=\u2019opacity:0.7;color:white\u2019 |\n| data-labelres | Residue label style specification | data-labelres=\u2019backgroundColor:white;fontColor:black\u2019 |\n| data-zoomto | AtomSpec selection to zoom to | data-zoomto=\u2019chain:A\u2019 |\n| data-spin | Spin model with axis and speed | data-spin=\u2019axis:z;speed:10\u2032 |\n| data-callback | JavaScript function to call on viewer creation | data-callback=\u2019myCallback\u2019 |\n| data-ui | Show the UI for the viewer | data-ui=\u2019true\u2019 |", "# When the Wet Lab Becomes the Bottleneck: Self-Driving Platforms for Protein Engineering\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-05-04 on marcdeller.com (categories: blog, sb).*\n\n* [**Self-driving Autonomous Machines for Protein Landscape Exploration (SAMPLE)**](https://www.nature.com/articles/s44286-023-00002-4)\n\n* **The lab is the bottleneck:** Generative antibody models now produce sequences faster than wet labs can build, test, and characterize them.\n* **Closed loops are the new default:** Self-driving Design\u2013Build\u2013Test\u2013Learn platforms compress weeks of human cycles into unattended overnight runs.\n* **ML-ready by design:** Modern automation often emits standardized numerical feature vectors that flow straight into model training pipelines.\n* **Frictions remain:** Well-to-well variability, instrument API integration, and immunogenicity assays still constrain end-to-end autonomous antibody discovery.\n\n1. **A high-throughput platform for biophysical antibody developability assessment to enable AI/ML model training:** [Arsiwala et al., mAbs, 2025](https://www.tandfonline.com/doi/full/10.1080/19420862.2025.2593055)\n2. **Self-driving laboratories to autonomously navigate the protein fitness landscape:** [Rapp et al., Nature Chemical Engineering, 2024](https://www.nature.com/articles/s44286-023-00002-4)\n3. **Engineering multispecific antibodies with complete killing selectivity through the closed-loop integration of machine learning and high-throughput experimentation:** [Mullin et al., mAbs, 2025](https://www.tandfonline.com/doi/full/10.1080/19420862.2025.2598093)\n4. **Autonomous \u2018self-driving\u2019 laboratories: a review of technology and policy implications:** [Tobias & Wahab, Royal Society Open Science, 2025](https://royalsocietypublishing.org/doi/10.1098/rsos.250646)\n5. **An Automated End-to-End Workflow for Production of Secreted Proteins in Transfected Mammalian Cells:**[Vasnarungruengkul et al., bioRxiv, 2025](https://www.biorxiv.org/content/10.1101/2025.07.13.664612v1)\n\n**How closed-loop automation, ML-ready datasets, and autonomous Design\u2013Build\u2013Test\u2013Learn cycles are reshaping therapeutic antibody discovery**\n\nFor most of the past decade, generative models have been outrunning the laboratory. Protein language models, diffusion-based structure generators, and Bayesian flow networks now produce *de novo* antibody sequences faster than any team can clone, express, and characterize them. The constraint has shifted: cloning vectors, transient transfections, microscale purification, biophysical panels, and developability assays now sit on the critical path. Generating the experimental data needed to train better models has become the rate-limiting step in antibody engineering, and the response has been a new class of infrastructure explicitly built to address that gap. Autonomous \u201cself-driving\u201d laboratories integrate robotics, scheduling software, and active-learning agents into a closed-loop Design\u2013Build\u2013Test\u2013Learn (DBLT) engine that runs continuously with minimal human intervention.\n\nThe blueprint is not new, but recent demonstrations show that the DBLT principle can work. [SAMPLE, described in a 2024 *Nature Chemical Engineering* paper](https://www.nature.com/articles/s44286-023-00002-4), paired an intelligent agent with a remote cloud lab to engineer thermostable enzymes without human input. A [2025 *Royal Society Open Science*review](https://royalsocietypublishing.org/doi/10.1098/rsos.250646) then cataloged how these systems span chemistry, materials science, and the biological sciences, framing closed-loop experimentation as the dominant motif of the next decade of autonomous research.\n\n> **\u201cSelf-driving laboratories automate and accelerate the scientific discovery process.\u201d\u2014 Rapp, Bremer & Romero, *Nature Chemical Engineering* (2024)**\n\nThree 2025 reports show where antibody-specific platforms are heading. Ginkgo Bioworks introduced [PROPHET-Ab](https://www.tandfonline.com/doi/full/10.1080/19420862.2025.2593055), a high-throughput developability suite that profiles 246 therapeutic IgGs across ten biophysical assays and emits ML-ready datasets at the scale required to train modern predictors. LabGenius reported [EVA](https://www.tandfonline.com/doi/full/10.1080/19420862.2025.2598093), a closed-loop multispecific platform that traversed 44,160 T-cell engager designs and recovered sub-10 pM HER2\u00d7CD3 candidates with 14-fold enrichment over Sobol sampling. A [Caltech group published a fully automated end-to-end workflow](https://www.biorxiv.org/content/10.1101/2025.07.13.664612v1) that takes plasmid DNA through to assay-ready secreted protein on a single 96-well robotic line.\n\nThe common thread is that experimental output is no longer treated as ad hoc data. Every assay is engineered to emit standardized numerical feature vectors (kinetic constants, Tm, \u0394PWL, polyspecificity indices) that flow directly into an ML feature store. Bottlenecks remain in well-to-well variability, instrument API integration, and immunogenicity assays that are difficult to miniaturize. The trajectory is nonetheless clear: when the lab itself becomes a programmable component of the model, the question shifts from how to design a million antibodies at scale to which million to measure next.", "# GitHub README: site-doctor\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/site-doctor, 2026\n\nIf none of `--scan`, `--inventory`, or `--analytics` are given, **all three\nrun**. Note that `--inventory` and `--analytics` both depend on the content\ninventory, so requesting `--analytics` alone will still fetch the inventory\nin the background \u2014 it just won't be printed/included unless `--inventory`\nis also given.\n\n### Other options\n\n| Flag | Description |\n| --- | --- |\n| `url` | Target site URL (optional \u2014 prompted for if omitted), e.g. `https://example.com` |\n| `--max-pages N` | Number of pages to run the per-page SEO analysis on during `--analytics` (default: `10`). Increase for a more complete audit of larger sites, at the cost of one extra HTTP request per page. |\n| `--no-psi` | Skip the Google PageSpeed Insights API call entirely and go straight to the self-computed performance scorecard. Useful when offline, rate-limited, or for faster runs. |\n| `--out PATH` | Output HTML report path (default: `./site-doctor-report-<host>-<timestamp>.html`) |\n\n## Output\n\n- A color-coded report is printed to the terminal (colors auto-disable when\n  output isn't a TTY).\n- A branded, self-contained HTML report is written to disk, containing every\n  section run during that invocation (scorecard, tables, badges, and\n  per-page cards).\n\n## How it works\n\nSite Doctor avoids all third-party dependencies by implementing its own\nminimal HTTP/1.1 client on top of `socket` and `ssl`, which lets it measure\neach phase of a request (DNS, connect, TLS, TTFB, download) precisely. HTML\nis parsed with regular expressions rather than a DOM library, DNS records\nare fetched via `dig` (with a `socket`-based fallback), and WordPress data is\nread entirely from the site's existing public REST API endpoints \u2014 no\nplugins, authentication, or scraping required.\n\n## Disclaimer\n\nSite Doctor only reads publicly available information (HTTP responses, DNS\nrecords, and public REST API endpoints). Use it against sites you own or are\nauthorized to test.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/site-doctor\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/site-doctor</a></td>\n</tr>\n</table>", "# Late-stage Crystal Failures and the Protein Crystal Harvesting Bottleneck\nAuthors: Deller, M. C.; Walsh, J.; Viola, R.; Elsliger, M. A.; Wilson, I. A.; Rupp, B.\nSource: American Crystallographic Association (ACA) Annual Meeting 2014 (poster), 2014\n\n*Panel 2.* High-throughput harvesting into SSRL grids. (a\u2013d) Workflow of robot-assisted harvesting into SSRL sample mounting grids. UMR represents an opportunity for high-density shipping to FEL sources.\n\n*Panel 3.* Combination drip cryoprotection and hyperquenching. (a) Automated drip cryo mechanism for harvested crystals provides insulation from film boiling regime of liquid nitrogen (PFO oil). (b) Directly coupled pneumatic hyperquencher cools crystals at high-speed (5 m/s) in nucleate boiling regime of liquid nitrogen. (c) High-speed camera (1000 frames/s) allows accurate calculation of hyperquencher speeds. (d) High-rate data acquisition thermocouple allows for precise monitoring of cooling rates.\n\n*Panel 4.* Hyperquenching and increased rates of cooling. (a) Slow cooling rate of manual cryocooling in liquid nitrogen provides opportunity for damaging hexagonal ice formation. (b) Intermediate cooling rate of robot-assisted cooling in liquid nitrogen. (c) High-speed pneumatic cryocooling, via the hyperquencher, increases the probability of amorphous ice formation (vitreous/glassy/unstructured ice) which is less damaging to the protein crystal. (d) Overall temperature profile of hyperquencher (green) in comparison to a traditional manual (black) and robot-assisted (red) quench.\n\n## Conclusions\n\nCrystal harvesting is one of the last manual steps in the HTPX/SG pipeline and is a source of both a process bottleneck and late-stage crystal failures. The Universal Micromanipulation Robot, developed with Square One Systems Design and tested at the JCSG, automates harvesting, including microcrystals into SSRL sample mounting grids for high-density shipping to FEL sources, and couples automated drip cryoprotection with pneumatic hyperquenching at 5 m/s. Cooling rates of 10,000\u201320,000 K/s, compared with ~1000 K/s for conventional methods, are predicted to favour amorphous rather than hexagonal ice, and measurement of quench speeds and temperature profiles provides a framework for optimizing cryoprotection and cooling. Such a system offers the potential to relieve the harvesting bottleneck and reduce late-stage crystal failures.\n\n## References\n\n*(reference list omitted)*", "# Unlocking Documents for AI: The Power of Docling for AI-Driven Transformation\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-18 on marcdeller.com (categories: ai, blog).*\n\n* [Docling - Open Source Document Processing for AI](https://www.docling.ai)\n* [Granite Docling WebGPU](https://huggingface.co/spaces/ibm-granite/granite-docling-258M-WebGPU)\n\n* **Enterprise Data Liberation:** Docling transforms PDFs, DOCX, and other formats into [JSON and Markdown](https://docling-project.github.io/docling/usage/) outputs optimized for large language models and RAG applications.\n* **Specialized AI Models:** [DocLayNet and TableFormer](https://github.com/docling-project/docling-ibm-models) models trained on 81,000+ pages deliver near-human accuracy for layout analysis and table extraction without relying on traditional OCR.\u200b\n* **Frictionless Deployment:** [Single-command installation](https://docling-project.github.io/docling/installation/) via pip install docling works across all major platforms, with optional browser-based processing through Granite Docling WebGPU.\u200b\n* **Privacy-Preserving Processing:** [100% local execution](https://huggingface.co/spaces/ibm-granite/granite-docling-258M-WebGPU) capability ensures sensitive documents never leave enterprise networks or user devices, addressing compliance requirements.\u200b\n* **RAG Framework Integration:** Native compatibility with [LangChain, LlamaIndex, and spaCy](https://docling-project.github.io/docling/integrations/llamaindex/)enables drop-in enhancement of existing AI application pipelines.\u200b\n* **Open-Source Momentum:** Earning [37,000+ GitHub stars](https://www.ibm.com/think/news/doclings-rise-llm-ready-data) and contribution to the Linux Foundation demonstrates strong community adoption and long-term sustainability.\n\n1. **A new tool to unlock data from enterprise documents for generative AI:**[IBM Research Blog](https://research.ibm.com/blog/docling-generative-AI)\n2. **Docling Technical Report:** [Auer, C., et al., arXiv:2408.09869](https://arxiv.org/abs/2408.09869)\n3. **Docling: An Efficient Open-Source Toolkit for AI-driven Document Conversion:** [Livathinos, N., et al., arXiv:2501.17887](https://arxiv.org/html/2501.17887v1)\n4. **IBM Granite-Docling: End-to-end document understanding:** [IBM Announcements](https://www.ibm.com/new/announcements/granite-docling-end-to-end-document-conversion)\n5. **Docling: AI Powered Document Parsing for LLMs and RAG:** [Franks World](https://www.franksworld.com/2024/12/11/docling-ai-powered-document-parsing-for-llms-and-rag/)\n6. **IBM contributes key open-source projects to Linux Foundation:** [IBM News](https://www.ibm.com/new/announcements/ibm-adds-open-source-projects-docling-beeaI-and-data-prep-kit-added-to-the-linux-foundation)\n\n**IBM\u2019s open-source toolkit bridges the gap between static document images and PDFs and intelligent systems, making document data AI-ready in seconds**\n\nThe enterprise world faces a persistent challenge: vast repositories of valuable information trapped in PDFs, PowerPoint decks, Word documents, and scanned images remain inaccessible to modern AI systems. While foundation models have consumed nearly every scrap of publicly available internet data, the most critical business intelligence often sits locked away in complex document formats that resist traditional parsing methods. This disconnect has created a significant bottleneck in deploying generative AI applications, particularly for retrieval-augmented generation (RAG) systems that require clean, structured data to deliver accurate responses.\n\n[Docling](https://docling-project.github.io/docling/getting_started/), IBM Research\u2019s open-source document processing toolkit, addresses this fundamental limitation with remarkable efficiency. Released in 2024 under a permissive MIT license, Docling transforms unstructured documents into machine-readable JSON and Markdown formats using specialized AI models rather than traditional optical character recognition (OCR). This approach delivers 30 times faster processing while reducing errors, according to IBM researcher Peter Staar, who helped develop the system. The toolkit\u2019s two core AI models\u2014a layout analysis model trained on the [DocLayNet](https://github.com/docling-project/docling-ibm-models) dataset of 81,000 manually labeled pages and TableFormer for table structure recognition\u2014achieve accuracy within five percentage points of human performance. Unlike commercial solutions requiring cloud connectivity, Docling runs entirely on commodity hardware, from standard laptops to enterprise servers, making it ideal for sensitive data and air-gapped environments.\n\nInstallation couldn\u2019t be simpler: execute [pip install docling](https://docling-project.github.io/docling/installation/) from your Python environment to begin processing documents immediately:\n\n```\n\n\t\t\t\t\tpip install docling\n\n```\n\nThis single command installs all dependencies and downloads the necessary AI model weights from Hugging Face on first use. The toolkit works seamlessly across macOS, Linux, and Windows platforms, supporting both x86_64 and arm64 architectures without additional configuration. For those seeking an even simpler entry, IBM recently launched [Granite Docling WebGPU](https://huggingface.co/spaces/ibm-granite/granite-docling-258M-WebGPU), a browser-based demonstration running a 258-million parameter vision-language model entirely client-side.\n\nThis WebGPU implementation processes documents with 100% local execution, ensuring no data leaves the user\u2019s device\u2014a critical feature for handling confidential materials. Basic usage requires just five lines of Python code: import the DocumentConverter, point it at a PDF file or URL, and export the results to Markdown or JSON. More advanced configurations enable OCR for scanned documents, activate table structure recognition, adjust image resolution, and leverage GPU acceleration for faster processing.\n\nIt works well in my hands for simple tables and bar charts, but I\u2019ve not had much success so far with more complex graphs or images. Also, doesn\u2019t\u00a0support PDFs out of the box.", "# Crystal Structure of the BG505 SOSIP gp140 HIV-1 Env Trimer in Complex with the Broadly Neutralizing Fab PGT122\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* PDB ID: [4NCO](https://www.rcsb.org/structure/4NCO)\n* **Release Date:** 2013-11-13\n* **Method:** X-RAY DIFFRACTION 4.7 \u00c5\n* **Organisms:** Homo sapiens, Human immunodeficiency virus 1\n* **Macromolecules:** BG505 SOSIP gp120, BG505 SOSIP gp41, PGT122 heavy chain, PGT122 light chain (proteins)Unique **Ligands:** NAG\n* **Publication:** [Julien, J.P., et al. (2013). Crystal Structure of a Soluble Cleaved HIV-1 Envelope Trimer. Science, 342, 1477-1483.](https://www.science.org/doi/10.1126/science.1245625)\n* **Blog Post:** [Crystal Structure of a Soluble Cleaved HIV-1 Envelope Trimer](https://marcdeller.com/crystal-structure-of-a-soluble-cleaved-hiv-1-envelope-trimer/)\n\n**Description:** A landmark in HIV structural biology, this structure reveals the first crystal structure of a soluble, cleaved HIV-1 Env trimer (BG505 SOSIP.664) in complex with the broadly neutralizing antibody PGT122. It shows the prefusion state of gp41, the gp120/gp41 interface, and the stabilizing V1/V2/V3 apex architecture, providing a foundational blueprint for structure-based HIV vaccine design.", "# eloratherapeutics: Elora Therapeutics\nAuthors: eloratherapeutics\nSource: https://www.eloratherapeutics.com, 2026\n\n# Plastic Meets Biology\n\n#### Developing engineered enzyme therapeutics to remove microplastics from humans implicated in cancer progression and metastasis", "# Capturing the Invisible: Understanding the \u00c5ngstr\u00f6m in Structural Biology\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-20 on marcdeller.com (categories: sb).*\n\n* The \u00e5ngstr\u00f6m equals 1 \u00d7 10\u207b\u00b9\u2070 m, ideal for atomic and molecular dimensions.\n* Named after Anders Jonas \u00c5ngstr\u00f6m, pioneer of solar spectroscopy.\n* Core to X-ray crystallography, NMR, and cryo-EM in structural biology.\n* Enables mapping of atomic coordinates and bond geometries with sub-\u00c5 accuracy.\n* Persistence due to historical convention, ease of scale alignment, and data-reporting standards.\n* Drives advances in drug design, enzyme mechanism studies, and nanobiotechnology.\n\n1. **Wikidoc:** \u00c5ngstr\u00f6m\u2014Definition, history, and applications:\n   <https://www.wikidoc.org/index.php/%C3%85ngstr%C3%B6m>\n2. **Fermat\u2019s Library:** \u201cWhy some things are darker when wet\u201d\u2014Annotated discussion of \u00c5ngstr\u00f6m\u2019s models:\n   <https://fermatslibrary.com/s/why-some-things-are-darker-when-wet>\n3. **Wikipedia:** Structural biology\u2014Technique overview and \u00c5ngstr\u00f6m-level resolution milestones:\n   <https://en.wikipedia.org/wiki/Structural_biology>\n4. **QUDT:** Unit definitions for \u00c5ngstr\u00f6m (ANGSTROM). <https://qudt.org/vocab/unit/ANGSTROM>\n5. **USU Today:**USU\u2019s Science Unwrapped Explores the \u00c5ngstr\u00f6m Scale:\n   <https://www.usu.edu/today/story/usu-science-unwrapped-explores-the-angstrom-scale-friday-march-22>\n\nThe \u00e5ngstr\u00f6m (\u00c5), a unit equal to 10\u207b\u00b9\u2070 m, is the bedrock of atomic-scale measurement in structural biology\u2014crucial to interpreting biomolecular form and function at true atomic resolution.\n\nThe **\u00e5ngstr\u00f6m** (symbol \u00c5) traces its origins to Swedish physicist Anders Jonas \u00c5ngstr\u00f6m (1814\u20131874), who first quantified solar spectral lines using this scale of one ten-billionth of a meter. Though not an SI unit, it persists in fields where atomic-level precision is paramount, such as crystallography, electron microscopy, and spectroscopy. Its adoption reflects both historical legacy and unmatched convenience: atomic radii, bond lengths, and protein domain separations naturally fall in the 1\u201310 \u00c5 range, eliminating cumbersome decimal notation.\n\nStructural biology\u2019s major techniques\u2014**X-ray crystallography**, **nuclear magnetic resonance** (NMR), and **cryo-electron microscopy** (cryo-EM)\u2014each exploit the \u00e5ngstr\u00f6m scale:\n\n* **X-ray crystallography** revolutionized atomic visualization in the early 20th century. Diffraction patterns from protein crystals yield electron density maps refined to sub-3 \u00c5\u2014and often better than 2 \u00c5\u2014revealing precise atomic coordinates.\n* **NMR spectroscopy** offers dynamic structural ensembles, with bond length and dihedral angle measurements accurate to ~0.01 \u00c5, elucidating flexibility and conformational exchange without crystallization.\n* **Cryo-EM\u2019s renaissance** in the 2010s overcame the \u201cresolution barrier,\u201d now allowing single-particle reconstructions at sub-2 \u00c5 resolution, once unimaginable outside crystallography.\n\nAchieving **1.2 \u00c5** resolution in X-ray structures, for instance, pinpoints hydrogen-bond geometry; at **1.0 \u00c5** and below, charge density maps even discriminate bonding electrons. In cryo-EM, sub-2 \u00c5 maps facilitate drug-binding site modeling directly from density, accelerating structure-based drug design.", "# PETase: Nature\u2019s Solution to a Human-Made Problem\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-04-14 on marcdeller.com (categories: microplastics, sb).*\n\nPDB-101: Molecule of the Month: Plastic-eating Enzymes\n\n* [**Molecule of the Month**](https://pdb101.rcsb.org/motm/277)\n* [**Elora Therapeutics**](https://eloratherapeutics.com)\n\n* **Bacterial origin:** *Ideonella sakaiensis* evolved a two-enzyme PET-degrading system, giving PETase its remarkable catalytic architecture.\n* **Serine hydrolase mechanism:** PETase uses a Ser-His-Asp catalytic triad to cleave PET ester bonds with precise substrate specificity.\n* **Engineered variants:** FAST-PETase and LCC-ICCG achieve near-complete depolymerization of post-consumer plastics within days.\n* **Therapeutic potential:** Engineered PETase degrades PET microplastics in human serum at 37\u00b0C without measurable cytotoxicity.\n\n1. Plastic-eating Enzymes: [Goodsell, RCSB PDB-101, 2023](https://pdb101.rcsb.org/motm/277)\n2. An engineered PET depolymerase to break down and recycle plastic bottles: [Tournier et al., Nature, 2020](https://www.nature.com/articles/s41586-020-2149-4)\n3. Machine learning-aided engineering of hydrolases for PET depolymerization: [Lu et al., Nature, 2022](https://www.nature.com/articles/s41586-022-04599-z)\n4. Current advances in the structural biology and molecular engineering of PETase: [Liu et al., Frontiers in Bioengineering and Biotechnology, 2023](https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2023.1263996)\n5. Degradation of PET microplastic particles to monomers in human serum: [Lopez-Lorenzo et al., Faraday Discussions, 2024](https://pubs.rsc.org/en/content/articlehtml/2024/fd/d4fd00014e)\n\n**How a bacterial enzyme discovered in a Japanese recycling facility is reshaping our understanding of plastic degradation \u2014 and opening unexpected doors in human health.**\n\nPolyethylene terephthalate, known universally as PET, is one of the most widely produced synthetic polymers on Earth \u2014 present in everything from single-use water bottles to polyester textiles\u00a0\u2014 and it accumulates in landfills and ecosystems at a rate that outpaces every conventional recycling strategy currently deployed. With over 70 million tons manufactured annually and the predominant thermomechanical recycling route degrading material quality with each cycle, the case for a fundamentally different approach has grown increasingly urgent. Into this gap stepped an unlikely protagonist: *Ideonella sakaiensis* 201-F6, isolated from a PET film sample at a Japanese recycling facility. This organism had evolved a two-enzyme system capable of using PET as its primary carbon and energy source, and at its center was a [structurally unique serine hydrolase](https://pdb101.rcsb.org/motm/277), later known as PETase.\n\nThe catalytic logic of PETase is elegant in its precision. The enzyme operates through a canonical serine-histidine-aspartate catalytic triad, the same arrangement seen in serine proteases, and executes a two-step acylation-deacylation mechanism that cleaves the ester bonds that hold PET polymer chains together. What makes the enzyme structurally distinctive is the architecture of its substrate-binding cleft \u2014 wider and more open than that of related cutinases \u2014 allowing the bulky aromatic PET chain to access the active site. PETase attacks the solid polymer surface in an interfacial biocatalytic process, releasing soluble oligomers including mono(2-hydroxyethyl) terephthalate (MHET) as the primary product; a partner enzyme, MHETase, then completes the depolymerization cascade, cleaving MHET into terephthalic acid and ethylene glycol.\n\nThe wild-type enzyme was constrained by poor thermostability and modest efficiency against the densely packed crystalline regions of post-consumer waste. Two engineering milestones changed the trajectory. In 2020, [Tournier et al.](https://www.nature.com/articles/s41586-020-2149-4) reported an engineered thermophilic cutinase variant achieving at least 90% PET depolymerization within 10 hours at industrially relevant loadings. Then, in 2022, [Lu et al.](https://www.nature.com/articles/s41586-022-04599-z) used a structure-based machine learning algorithm to engineer FAST-PETase \u2014 five mutations that collectively enabled near-complete degradation of untreated post-consumer PET from 51 different thermoformed products within one week, and recovery of monomers suitable for repolymerisation into virgin-quality PET. A [comprehensive structural and mechanistic review](https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2023.1263996) published the following year detailed how each of these rational and computational strategies exploited the enzyme\u2019s binding-groove geometry to improve catalytic turnover.\n\nPerhaps the most striking recent development lies not in industrial recycling but in biomedicine. Microplastics have now been detected in human blood, placental tissue, and gastrointestinal samples, with PET among the most frequently identified polymer types. In a 2024 study, [Lopez-Lorenzo et al.](https://pubs.rsc.org/en/content/articlehtml/2024/fd/d4fd00014e) at KTH Royal Institute of Technology demonstrated that an engineered PETase variant can depolymerize microplastic PET substrates directly in human serum at 37\u00b0C, producing terephthalic acid and MHET without measurable cytotoxicity in HEK293 cell viability assays. What began as an observation about bacterial metabolism in a recycling facility has matured, in under a decade, into a scientific program spanning polymer chemistry, protein engineering, environmental remediation, and biomedicine.", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\n**71.** The antibody of claim 67, wherein the antibody binds to the first epitope at all of the amino acid residues D1, R5, R8, M9, M10, R11, T12, M14, R15, M16, M19, and R20 in DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328) and the antibody binds to the second epitope at all of the amino acid residues T6, R7, M10, R11, M14, R17, R18, R20, and R21 in DEEQRTRRMMRTKMRMRRMRR (SEQ ID NO:328).\n\n**72.** An antibody that binds to human mutant CALR and competes for binding to human mutant CALR with an antibody that has a heavy chain comprising the amino acid sequence of SEQ ID NO:327 and a light chain comprising the amino acid sequence of SEQ ID NO:213.\n\n**73.** An antibody that binds to human mutant CALR and inhibits or reduces binding of human mutant CALR to thrombopoietin receptor (MPL).\n\n**74.** The antibody of any one of claims 1 to 73, wherein the human mutant CALR is human Type 1 mutant CALR comprising the amino acid sequence of SEQ ID NO:320.\n\n**75.** The antibody of any one of claims 1 to 73, wherein the human mutant CALR is human Type 2 mutant CALR comprising the amino acid sequence of SEQ ID NO:321.\n\n**76.** The antibody of any one of claims 1-75, wherein the antibody is a human or humanized antibody.\n\n**77.** The antibody of any one of claims 1-76, wherein the antibody is a full-length antibody.\n\n**78.** The antibody of any one of claims 1-77, wherein the antibody is an IgG1, IgG2, IgG3 or IgG4 antibody.\n\n**79.** The antibody of claim 78, wherein the antibody is an IgG1 antibody.\n\n**80.** The antibody of any one of claims 1-75, wherein the antibody is a bispecific antibody, a biparatopic antibody, a single chain antibody (scFv), an Fab fragment, an F(ab')2 fragment, an Fab' fragment, an Fsc fragment, an Fv fragment, an scFv, an sc(Fv)2, or a diabody.\n\n**81.** The antibody of claim 80, wherein the antibody is a biparatopic antibody comprising two heavy chain-light chain pairs or one heavy chain-light chain pair.\n\n**82.** The antibody of claim 80, wherein the biparatopic antibody is a full-length antibody.\n\n**83.** The antibody of claim 81 or claim 82, wherein the biparatopic antibody comprises one heavy chain-light chain pair.\n\n**84.** The antibody of any of claims 1-83, which is conjugated to a toxic substance.\n\n**85.** The antibody of claim 84, which is conjugated to a radioisotope or a cytotoxic agent.\n\n**86.** A nucleic acid or a set of nucleic acids, which collectively encodes the antibody of any one of claims 1-85.\n\n**87.** An expression vector or a set of expression vectors comprising the nucleic acid or the set of nucleic acids of claim 86 operably linked to a promoter.\n\n**88.** An isolated cell comprising the nucleic acid or the set of nucleic acids of claim 86 or the expression vector or the set of expression vectors of claim 87.\n\n**89.** A method of making the antibody of any one of claims 1-85, comprising culturing the cell of claim 88 and isolating the antibody.\n\n**90.** A pharmaceutical composition comprising the antibody of any one of claims 1-85, the nucleic acid or the set of nucleic acids of claim 86, the expression vector or the set of expression vectors of claim 87, or the isolated cell of claim 88, and a pharmaceutically acceptable carrier.\n\n**91.** A method of treating a myeloproliferative neoplasm in a human subject in need thereof, the method comprising administering to the human subject an effective amount of the antibody of any one of claims 1-85 or the pharmaceutical composition of claim 90.\n\n**92.** The method of claim 91, wherein the myeloproliferative neoplasm is selected from the group consisting of chronic myelogenous leukemia, polycythemia vera, primary myelofibrosis, essential thrombocythemia, chronic neutrophilic leukemia, acute myelogenus leukemia, chronic eosinophilic leukemia, chronic myelomonocytic leukemia, myeloproliferative neoplasm and myelodysplastic syndrome, including myelodysplastic syndrome with refractory anaemia with ring sideroblasts, myelodysplastic syndrome with refractory anemia, and myelodysplastic syndrome with refractory anemia with excess blasts.\n\n**93.** The method of claim 91 or claim 92, further comprising administering to the human subject an additional therapy selected from the group consisting of a Janus tyrosine kinase (JAK) inhibitor, a phosphoinositide 3-kinase (PI3K) inhibitor, a standard of care therapy, or a combination thereof.\n\n**94.** The method of claim 93, wherein the JAK inhibitor is ruxolitinib and itacitinib.\n\n**95.** The method of claim 93, wherein the PI3K inhibitor is parsaclisib.\n\n**96.** The method of claim 93, wherein the standard of care therapy is selected from the group consisting of IFN-alpha, hydroxyurea, thalidomide, lenalidomide, an androgen, an erythropoietin-stimulating agent, a chemotherapeutic agent, or a combination thereof.\n\n**97.** The method of claim 93, wherein the administration of the antibody of any one of claims 1-85 or the pharmaceutical composition of claim 90 in combination with the JAK inhibitor produces a synergistic effect.\n\n**98.** The method of claim 97, wherein the JAK inhibitor is ruxolitinib.\n\n**99.** A kit comprising the antibody of any one of claims 1-85, the nucleic acid or the set of nucleic acids of claim 86, the expression vector or the set of expression vectors of claim 87, or the isolated cell of claim 88, and instructions for use in treating a myeloproliferative neoplasm in a human subject in need thereof, optionally with instructions for use in combination with an additional therapy.\n\n**100.** A method of detecting a CALR exon 9 mutation in a biological sample, the method comprising obtaining a biological sample from a human subject who has or is suspected of having a myeloproliferative neoplasm and contacting the sample with the antibody of any one of claims 1-85 such that the antibody binds to a mutCALR protein if the mutCALR protein is present in the biological sample.", "# Unveiling FragLites: Pioneering Precision in Drug Discovery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-22 on marcdeller.com (categories: sb).*\n\n* FragLites are a focused library of 31 halogenated fragments used to map protein binding sites with high sensitivity and specificity.\n* They enable identification of known and novel protein interaction hotspots, aiding drug discovery and protein biology.\n* Application on CDK2-cyclin A and bromodomain proteins (BRD4, ATAD2) validates their versatility in mapping ligand sites.\n* FragLite binding exploits halogen anomalous scattering for precise X-ray crystallographic detection.\n* The approach reduces resource needs compared to full fragment screening while maintaining comprehensive site mapping.\n* FragLites support development of chemical probes and inform mutation site selection for functional studies.\n\n1. Erlanson, D. A., Murray, C. W., & Davis, B. J. (2019). **FragLites\u2014Minimal, Halogenated Fragments Displaying Pharmacophore Doublets.** Journal of Medicinal Chemistry:\n   <https://pubs.acs.org/doi/10.1021/acs.jmedchem.9b00304>\n2. Wood, E. R., et al. (2024). **Crystallographic fragment screening of CDK2-cyclin A: FragLites identify functional protein interaction hotspots.** bioRxiv:\n   <https://www.biorxiv.org/content/10.1101/2024.06.03.596235v1.full.pdf>\n3. Filippakopoulos, P., et al. (2022). **Mapping ligand interactions of bromodomains BRD4 and ATAD2 reveals novel binding sites by fragment screening. Journal of Medicinal Chemistry**:\n   <https://pubs.acs.org/doi/10.1021/acs.jmedchem.2c01357>\n4. Keable, S. et al. (2025). **Crystallographic Fragment Screening of CDK2-cyclin A: FragLites Rapidly Identify Multiple Binding Sites**. European Journal of Medicinal Chemistry:\n   <https://www.sciencedirect.com/science/article/pii/S0969212625002655>\n\nFragLites have emerged as a groundbreaking tool in the realm of fragment-based drug discovery (FBDD), offering an innovative approach to identify and map protein interaction sites with unparalleled efficiency and precision. These small, halogenated molecular fragments leverage the powerful anomalous scattering property of halogens, enabling researchers to sensitively and unambiguously detect productive drug-like interactions through X-ray crystallography. Unlike traditional extensive fragment libraries, the FragLites set\u2014a concise library of 31 halogenated fragments\u2014can pinpoint key protein binding hotspots effectively, making it a resource-efficient method for understanding protein surfaces relevant for drug targeting.\n\nRecent studies between 2024 and 2025 have validated the capabilities of FragLites across diverse protein systems. Notably, FragLites successfully mapped the protein-protein interaction (PPI) hotspots on cyclin-dependent kinase 2 (CDK2) and its complex with cyclin A, uncovering both known and previously uncharacterized binding sites. This chemically rich mapping serves as a foundational resource for designing chemical probes and selecting mutation sites that distinguish protein functions, thereby advancing both drug design and functional biology insights.\n\nIn addition to CDK2, FragLites have demonstrated their utility in mapping ligand binding sites of bromodomains such as BRD4 and ATAD2. The technology revealed that FragLites can recapitulate interactions identified by full fragment screening but with greater speed and reduced experimental resource input. This allows for early-stage assessment of protein tractability and informs hit-finding strategies. Furthermore, the FragLite framework extends to analogues termed PepLites that mimic peptide interactions, broadening their applicability.\n\nThe benefits of FragLites go beyond mere binding detection. Their selective binding profiles enable prioritization of functional hotspots over nonspecific or artefactual interactions, supported by corroborative mutational studies and structural analyses. This precision enriches the drug discovery pipeline by offering reliable starting points for rational design of lead-like molecules, greatly enhancing efficiency in early drug development stages. Overall, FragLites mark a significant advancement in fragment-based methodologies, illustrating a promising avenue for accelerating the discovery of novel therapeutics.", "# 2024 \u2013 Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective CDK2 Inhibitors\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe broader implications for cancer therapy are substantial, as this work establishes a framework for developing next-generation CDK inhibitors with improved therapeutic windows. The success of the (4-pyrazolyl)-2-aminopyrimidine scaffold demonstrates how structure-based drug design can overcome historical challenges in kinase selectivity, potentially serving as a template for targeting other members of the CDK family with enhanced precision.\n\n| Key Concepts | Description | Key References |\n| --- | --- | --- |\n| CDK2 Selectivity | Achievement of selectivity over other CDK isoforms through structure-based design | [Hummel et al., J. Med. Chem.](https://pubs.acs.org/doi/abs/10.1021/acs.jmedchem.3c02287) |\n| CCNE1 Amplification | Genetic alteration creating therapeutic vulnerability in specific cancer subtypes | [Anonymous, Cancer Res.](https://aacrjournals.org/cancerres/article/85/7/1297/754355/Profiling-the-Activity-of-the-Potent-and-Highly) |\n| Structure-Activity Relationships | Molecular determinants of potency and selectivity in pyrazolyl-aminopyrimidine scaffold | [Cyclin-Dependent Kinase Inhibitors: SAR, ACS J. Med. Chem.](https://pubs.acs.org/doi/10.1021/acs.jmedchem.6b01254) |\n| Rb Phosphorylation | Key mechanistic readout of CDK2 activity and therapeutic target engagement | [Nature Sci. Rep. 2022](https://www.nature.com/articles/s41598-022-20769-5) |\n| Combination Therapy | Synergistic approaches combining CDK2 inhibition with DNA-damaging agents | [Brown et al., NAR Cancer](https://pubmed.ncbi.nlm.nih.gov/37519629/) |\n| Clinical Translation | Development pathway from preclinical discovery to clinical trials | [Expert Opin. Drug Discov. 2025](https://www.tandfonline.com/doi/full/10.1080/14728222.2025.2537412) |", "# Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\nAuthors: Ye, Q.; Shvartsbart, A.; Li, Z.; Gan, P.; Policarpo, R.L.; Qi, C.; Roach, J.J.; Zhu, W.; McCammant, M.S.; Hu, B.; Li, G.; Yin, H.; Carlsen, P.; Hoang, G.; Zhao, L.; Susick, R.; Zhang, F.; Lai, C.-T.; Allali Hassani, A.; Epling, L.B.; Gallion, A.; Kurzeja-Lipinski, K.; Gallagher, K.; Roman, V.; Farren, M.R.; Kong, W.; Deller, M.C.; Zhang, G.; Covington, M.; Diamond, S.; Kim, S.; Yao, W.; Sokolsky, A.; Wang, X.\nSource: Journal of Medicinal Chemistry, 2025\n\n**Corresponding Author:** Alexander Sokolsky \u2014 Department of Discovery Chemistry, Incyte Research Institute, Incyte Corporation, Wilmington, Delaware 19803, United States; Email: asokolsky@incyte.com.\n\n**Authors:** Qinda Ye, Artem Shvartsbart, Zhenwu Li, Pei Gan, Rocco L. Policarpo, Chao Qi, Jeremy J. Roach, Wenyu Zhu, Matthew S. McCammant, Bin Hu, Gencheng Li, Haolin Yin, Peter Carlsen, Gia Hoang, Le Zhao, Robert Susick and Fenglei Zhang (Department of Discovery Chemistry); Cheng-Tsung Lai (Department of Computational Chemistry); Abdellah Allali Hassani, Leslie B. Epling, Kerri Kurzeja-Lipinski, Karen Gallagher, Marc C. Deller, Guofeng Zhang and Maryanne Covington (Department of Applied Technology); Alexandra Gallion, Valerie Roman, Matthew R. Farren and Sunkyu Kim (Department of Preclinical Pharmacology); Weixi Kong and Sharon Diamond (Department of DMB); Wenqing Yao (Department of Discovery Chemistry and Department of Preclinical Pharmacology); Xiaozhao Wang (Department of Discovery Chemistry) \u2014 all of the Incyte Research Institute, Incyte Corporation, Wilmington, Delaware 19803, United States.\n\n**Funding:** Incyte Corporation, Wilmington, DE, USA.\n\n**Notes:** The authors declare the following competing financial interest(s): All authors of this manuscript were employed by Incyte Corporation at the time this work was conducted.\n\n## Acknowledgments\n\nKwang-Jong Chen, Gina Correa, Stephanie Wezalis, Michelle Conlin, Amy Hehman, Stephanie Hamill and Rina Pan are gratefully acknowledged for biological assay, ADME screening, and PK support. We would like to thank Yingrui Dai, Robby Kirk, Vince Caruso, Ronald Magboo, and Saran Carrington for analytical assistance. We thank Sean Bowen, James Hall, and Scott Leonard for assistance with obtaining and analyzing NMR spectra.\n\n## Abbreviations\n\nADME, absorption, distribution, metabolism and excretion; AUC, area under the drug concentration\u2212time curve; Cmax, maximal concentration of the concentration\u2212time curve; CYP, cytochrome p450; CYP TDI, time dependent inhibition of cytochrome p450; DIPEA, diisopropyl ethyl amine; DMF, dimethylformamide; ERK, extracellular signal-regulated kinase; F%, fraction of an administered dose of unchanged drug that reaches systemic circulation; FaSSiF, fasted state simulated intestinal fluid; FDA, Federal Drug Administration; GDP, guanadine diphosphate; GTP, guanadine triphosphate; %HBF, % hepatic blood flow; hERG, human ether-a-go-go-related gene; i.p., intraperitoneal; i.v., intravenous; p.o., per os; MEK, mitogen-activated protein kinase kinase; NIS, N-iodosuccinimide; NMP, N-methylpyrrolidone; SAR, structure\u2212activity relationship; SGF, simulated gastric fluid; SOS1, Son of sevenless homologue 1; TEA, triethylamine; THF, tetrahydrofuran; Vss, volume of distribution at steady state.\n\n## References\n\n*(reference list omitted)*", "# Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\nAuthors: Ye, Q.; Shvartsbart, A.; Li, Z.; Gan, P.; Policarpo, R.L.; Qi, C.; Roach, J.J.; Zhu, W.; McCammant, M.S.; Hu, B.; Li, G.; Yin, H.; Carlsen, P.; Hoang, G.; Zhao, L.; Susick, R.; Zhang, F.; Lai, C.-T.; Allali Hassani, A.; Epling, L.B.; Gallion, A.; Kurzeja-Lipinski, K.; Gallagher, K.; Roman, V.; Farren, M.R.; Kong, W.; Deller, M.C.; Zhang, G.; Covington, M.; Diamond, S.; Kim, S.; Yao, W.; Sokolsky, A.; Wang, X.\nSource: Journal of Medicinal Chemistry, 2025\n\n**Step 11. tert-Butyl (1R,4R,5S)-5-(7-(3-Chloro-2-methylphenyl)-8-(2-cyanoethyl)-6-fluoro-4-(methylthio)-2-((R)-1-(3-oxomorpholino)ethyl)-1H-pyrrolo[3,2-c]quinolin-1-yl)-2-azabicyclo[2.1.1]hexane-2-carboxylate (40).** A solution of 39 (2.16 g, 3.21 mmol), (3-chloro-2-methylphenyl)boronic acid (1.09 g, 6.42 mmol), Pd(PPh3)4 (0.742 g, 0.642 mmol) and potassium phosphate (1.70 g, 8.03 mmol) in 1,4-dioxane (20 mL) and water (4 mL) was stirred at 80 \u00b0C for 30 min. The reaction was then concentrated and purified by flash chromatography (0\u2212100% ethyl acetate in hexanes) to afford the title compound (1.36 g, 59%). LC-MS calculated for C38H42ClFN5O4S (M + H)+: m/z = 718.3; found 718.3.\n\n**Step 12. 3-(1-((1R,4R,5S)-2-Azabicyclo[2.1.1]hexan-5-yl)-7-(3-chloro-2-methylphenyl)-6-fluoro-4-(5-methylpyrazin-2-yl)-2-((R)-1-(3-oxomorpholino)ethyl)-1H-pyrrolo[3,2-c]quinolin-8-yl)propanenitrile (23).** A solution of 40 (1.25 g, 1.74 mmol), Pd(PPh3)4 (1.00 g, 0.870 mmol), copper(I) 3-methylsalicylate (1.87 g, 8.70 mmol) and 2-methyl-5-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)pyrazine (1.92 g, 8.70 mmol) in 1,4-dioxane (10 mL) was stirred at 120 \u00b0C for 1 h. The reaction mixture was filtered through a thiol SiliaPrep SPE cartridge and concentrated. The residue was stirred in 1:1 DCM/TFA (30 mL) for 30 min, concentrated and purified by prep HPLC (pH 2). LC-MS calculated for C37H36ClFN7O2 (M + H)+: m/z = 664.3; found 664.4. 1H NMR (500 MHz, DMF-d7) \u03b4 9.71 (d, J = 1.5 Hz, 1H), 8.81 (d, J = 1.5 Hz, 1H), 8.63 (s, 1H), 7.84 (s, 1H), 7.65 (m, 1H), 7.49 (m, 2H), 6.38 (q, J = 6.3 Hz, 1H), 5.07 (m, 1H), 4.63 (d, J = 5.9 Hz, 1H), 4.21 (d, J = 16.5 Hz, 1H), 4.16 (d, J = 16.5 Hz, 1H), 3.82 (m, 1H), 3.76 (m, 1H), 3.59 (m, 1H), 3.12 (m, 1H), 3.10 (m, 1H), 2.96 (m, 2H), 2.91 (m, 1H), 2.67 (s, 3H), 2.60 (m, 1H), 2.52 (d, J = 8.6 Hz, 1H), 2.19 (s, 3H), 1.94 (d, J = 7.0 Hz, 1H), 1.69 (d, J = 6.7 Hz, 3H), 1.36 (d, J = 8.6 Hz, 1H), 1.22 (d, J = 7.0 Hz, 1H).\n\n### Synthesis of 5\n\n**Step 1. tert-Butyl (1R,4R,5S)-5-((7-Bromo-2-(3-(dimethylamino)azetidin-1-yl)-8-fluoro-6-iodo-3-nitroquinolin-4-yl)amino)-2-azabicyclo[2.1.1]hexane-2-carboxylate (31).** To a mixture of 28 (4.0 g, 8.59 mmol) in CH2Cl2 (60 mL) was added tert-butyl (1R,4R,5S)-5-amino-2-azabicyclo[2.1.1]hexane-2-carboxylate (1.7 g, 8.59 mmol) and DIPEA (7.48 mL, 42.9 mmol) and the reaction was stirred at 55 \u00b0C for 4 h. Then N,N-dimethylazetidin-3-amine dihydrochloride (1.63 g, 9.45 mmol) was added. After heating at 55 \u00b0C for another 4 h, the mixture was concentrated to dryness, and used without further purification (5.0 g, 84%). LCMS calculated for C24H30BrFIN6O4 (M + H)+: m/z = 691.1; found 691.2.\n\n**Step 2. tert-Butyl (1R,4R,5S)-5-((3-Amino-7-bromo-2-(3-(dimethylamino)azetidin-1-yl)-8-fluoro-6-iodoquinolin-4-yl)amino)-2-azabicyclo[2.1.1]hexane-2-carboxylate.** Sodium hydrosulfite (2.86 g, 16.4 mmol) in water (2.5 mL) was added to a solution of 33 (2.27 g, 3.28 mmol) and 30% aq. ammonium hydroxide (4.26 mL, 32.8 mmol) in MeOH (30 mL) at 0 \u00b0C. After 10 min, water (30 mL) was added to the reaction mixture followed by extraction with dichloromethane (30 mL \u00d7 3). The combined organic layers were dried over Na2SO4, filtered and concentrated. The crude product was added to a silica gel column and was eluted with methanol/dichloromethane from 0% to 10% to provide the desired product (1.9 g, 88% yield). LCMS calculated for C24H32BrFIN6O2 (M + H)+: m/z = 661.1; found 661.2.\n\n**Step 3. tert-Butyl (1R,4R,5S)-5-(7-Bromo-2-(3-(dimethylamino)-3-oxopropyl)-4-(3-(dimethylamino)azetidin-1-yl)-6-fluoro-8-iodo-1H-imidazo[4,5-c]quinolin-1-yl)-2-azabicyclo[2.1.1]hexane-2-carboxylate (32).** A solution of the product above (3.00 g, 4.54 mmol) and N,N-dimethyl-4-oxobutanamide (0.615 g, 4.76 mmol) in ethanol (9.1 mL) was heated to 85 \u00b0C in a sealed vial under air for 3 h. The vial was cooled to 22 \u00b0C and n-butanol (9 mL) was added, the vial uncapped, and the mixture stirred vigorously overnight (open to air) at 65 \u00b0C. Volatiles were removed in vacuo and the product was used without further purification (>99%). LCMS calculated for C30H39BrFIN7O3 (M + H)+: m/z = 770.1; found: 770.1.\n\n**Step 4. tert-Butyl (1R,4R,5S)-5-(7-Bromo-8-(2-cyanoethyl)-2-(3-(dimethylamino)-3-oxopropyl)-4-(3-(dimethylamino)azetidin-1-yl)-6-fluoro-1H-imidazo[4,5-c]quinolin-1-yl)-2-azabicyclo[2.1.1]hexane-2-carboxylate (33).** In a 40 mL vial, 32 (3.50 g, 4.54 mmol) and bis(tri-o-tolylphosphine)palladium(0) (0.325 g, 0.454 mmol) were dissolved in DMF (15 mL). Triethylamine (1.266 mL, 9.09 mmol) and acrylonitrile (0.598 mL, 9.09 mmol) were added to the reaction mixture. The headspace was purged with nitrogen and the vial was capped and stirred at 80 \u00b0C for 1 h. The reaction mixture was cooled to r.t and poured into rapidly stirring ice water to precipitate a yellow solid. The solid was filtered and dried on the filter for 72 h. The crude solid was dissolved in THF (40 mL) and cooled to 0 \u00b0C. Lithium triethylborohydride (8.2 mL, 1 M in THF, 1.81 equiv) was added in four portions over 30 min with LCMS monitoring after each addition. Water was slowly added to the reaction mixture (20 mL) and the mixture was extracted with DCM (3 \u00d7 20 mL), dried over MgSO4, and volatiles removed in vacuo to provide the desired product (3.24 g, quant.). LCMS calculated for C33H43BrFN8O3 (M + H)+: m/z = 697.3; found: 697.4.", "# Installation of RFdiffusion on MacOSX M1 architecture without GPU support\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-20 on marcdeller.com (categories: ai).*\n\n* **Rosetta/DGL Issues:** Some advanced functionalities (like DGL-based graph operations) may still fail due to missing native libraries on ARM.\n* **Slower Performance:** Without CUDA, all computations run on CPU\u2014expect longer runtimes.\n* **Apple Silicon Macs** can run RFdiffusion via specialized forks and environment tweaks.\n* **CUDA-dependent tasks will use CPU:** complex models may run slowly.\n* **Installing Miniconda** and the correct dependencies is essential for a smooth setup.\n* You may need to troubleshoot DGL and compatibility issues.\n\n1. **GitHub Issue: RFdiffusion installation on M1/M2:**\n   <https://github.com/RosettaCommons/RFdiffusion/issues/118>\n2. **Mac-compatible install instructions (community fork):**\n   <https://github.com/YaoYinYing/RFdiffusion#conda-install-se3-transformer>\n3. **Hugging Face RFdiffusion package and setup:**\n   <https://huggingface.co/GlandVergil/RFdiffusion>\n4. **Stable Diffusion on Apple Silicon overview:**\n   <https://huggingface.co/docs/diffusers/v0.5.1/en/optimization/mps>\n5. **Tutorial: Installing RFdiffusion step-by-step:**\n   <https://meilerlab.org/wp-content/uploads/2023/12/RFDiffusion_tutorial.pdf>\n\n**Installing RFdiffusion on MacOS M1: Step-by-Step Guide**\n\nThe explosion of deep learning is transforming the way researchers approach protein design. RFdiffusion, a RosettaFold-based generative tool, empowers users to craft entirely new protein structures. Yet, many Mac users\u2014especially those on the latest Apple Silicon (M1/M2)\u2014hit a wall: the official codebase is built for CUDA on NVIDIA GPUs. Here\u2019s how you can overcome this challenge and get RFdiffusion running on your M1 Mac.\n\n**Why RFdiffusion?**\n\nRFdiffusion excels at *de novo* protein design, enabling advances in biotechnology, drug discovery, and synthetic biology. Its architecture adapts the power of RosettaFold for generative tasks, making it invaluable for academic and industrial researchers alike.\n\n**The Challenge on Mac M1**\n\nThe official RFdiffusion relies on CUDA and some Linux-centric dependencies. Apple Silicon (M1/M2) lacks CUDA\u2014NVIDIA\u2019s GPU platform\u2014so a direct installation fails due to missing libraries (like `cudatoolkit` and `dgl`). Fortuitously, the open-source community has developed forks and installation recipes, with promising results.\n\nThe instruction below should work on MacOSX M1 architecture without GPU support, so it will be slooooww\u2026\n\n<https://github.com/YaoYinYing/RFdiffusion/tree/mps-test>\n\n```\ngit clone https://github.com/RosettaCommons/RFdiffusion.git\ncd RFdiffusion\nmkdir models && cd models\nwget http://files.ipd.uw.edu/pub/RFdiffusion/6f5902ac237024bdd0c176cb93063dc4/Base_ckpt.pt\nwget http://files.ipd.uw.edu/pub/RFdiffusion/e29311f6f1bf1af907f9ef9f44b8328b/Complex_base_ckpt.pt\nwget http://files.ipd.uw.edu/pub/RFdiffusion/60f09a193fb5e5ccdc4980417708dbab/Complex_Fold_base_ckpt.pt\nwget http://files.ipd.uw.edu/pub/RFdiffusion/74f51cfb8b440f50d70878e05361d8f0/InpaintSeq_ckpt.pt\nwget http://files.ipd.uw.edu/pub/RFdiffusion/76d00716416567174cdb7ca96e208296/InpaintSeq_Fold_ckpt.pt\nwget http://files.ipd.uw.edu/pub/RFdiffusion/5532d2e1f3a4738decd58b19d633b3c3/ActiveSite_ckpt.pt\nwget http://files.ipd.uw.edu/pub/RFdiffusion/12fc204edeae5b57713c5ad7dcb97d39/Base_epoch8_ckpt.pt\nwget http://files.ipd.uw.edu/pub/RFdiffusion/f572d396fae9206628714fb2ce00f72e/Complex_beta_ckpt.pt\nwget http://files.ipd.uw.edu/pub/RFdiffusion/1befcb9b28e2f778f53d47f18b7597fa/RF_structure_prediction_weights.pt\n```\n\nDownload the following into ./env\n\n<https://github.com/YaoYinYing/RFdiffusion/blob/mps-test/env/SE3nv_macos.yml>\n\n```\nconda env create -f ./SE3nv_macos.yml\nconda activate RFdiffusion\nconda install 'pytorch==2.3.0' torchvision torchaudio cpuonly -c pytorch\npip install 'dgl==2.2.1' -f https://data.dgl.ai/wheels/repo.html\npip install git+https://github.com/YaoYinYing/nvtx-mock --force-reinstall\npip install nvtx\npip install git+https://github.com/YaoYinYing/SE3Transformer@rfdiffusion-mps-test\npip install git+https://github.com/NVIDIA/dllogger#egg=dllogger\npip install git+https://github.com/YaoYinYing/RFdiffusion@mps-test\npip install pydantic\npip install torchdata==0.9.0\nconda activate RFdiffusion\ncd ./examples\ntar -xvf ./ppi_scaffolds_subset.tar.gz\n```\n\nMake your first run:\n\n```\n./scripts/run_inference.py 'contigmap.contigs=[150-150]' inference.output_prefix=test_outputs/test inference.num_designs=10\n```", "# Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor\nAuthors: Kong, L.; Ju, B.; Chen, Yajing; He, L.; Ren, L.; Liu, J.; Hong, K.; Su, B.; Wang, Z.; Ozorowski, G.; Ji, X.; Hua, Y.; Chen, Yanli; Deller, M.C.; Hao, Y.; Feng, Y.; Garces, F.; Wilson, R.; Dai, K.; O'Dell, S.; McKee, K.; Mascola, J.R.; Ward, A.B.; Wyatt, R.T.; Li, Y.; Wilson, I.A.; Zhu, J.; Shao, Y.\nSource: Immunity, 2016\n\nFurther analysis revealed that the LCDR1 loops and N termini of DRVIA7, VRC01, and 12A21 interacted with the glycan shield in subtly different ways (Figures 2E and S3A\u2013S3C). Specifically, VRC01 had a 2-aa deletion in its LCDR1 and a tyrosine that directed the first GlcNAc of the N276 glycan away from the antibody (Zhou et al., 2010). Similarly, 12A21 LCDR1 appeared to push the GlcNAc out of the way. By contrast, two hydrophobic residues in DRVIA7 LCDR1 (I29 and W32) shifted and twisted the GlcNAc in N276 by 5.5 \u00c5 and 30\u00b0 (glycotorsion \u03c6), respectively, compared to the VRC01-bound conformation. Superposition of a more intact N276 glycan structure (Scharf et al., 2014) with the gp120-bound DRVIA7, VRC01, and 12A21 structures suggested a potential clash with DRVIA7, but not with VRC01 or 12A21. The DRVIA7 and VRC01 structures also differed at their light-chain N termini, which interacted with an N-linked glycan N461 on the gp120 V5 loop (Figure 2F). Every residue from 460 to 463 was a potential N-linked glycosylation site with 14%\u201344% sequence conservation, and 91% of the 24,602 isolates analyzed here contained at least one glycosylation site in this V5 region. Thus, the N276 glycan and the V5 glycan, although 15 \u00c5 apart (Figure 2F), might together play an inhibitory role for DRVIA7 recognition. Indeed, structural modeling showed an unfavorable energy score for the DRVIA7 LCDR1 and N terminus that interacted with this glycan cluster, which could be substantially reduced by swapping the LCDR1 loop with VRC01 or 12A21 sequences and by truncating the light-chain N terminus (Figure S3D). Of note, the importance of N276 and V5 glycans has already been reported for the in vivo selection and function of CD4bs-directed bNAbs (Freund et al., 2015; Wang et al., 2015). Taken together, these findings suggested that glycan accommodation by the antibody light chain could be a critical determinant of VRC01-class neutralizing ability.\n\n*Figure 2. DRVIA7 and VRC01 Recognition of gp120 and gp140 Differed Primarily in Light-Chain Interactions with Glycans.* (A) Ribbon diagrams of the crystal structure of DRVIA7 Fab (light blue and green) bound to 93TH057 gp120 (yellow) at 3.4 \u00c5 resolution are displayed. The N276 glycan is shown in stick-and-ball representation with carbon atoms colored green, nitrogen atoms colored blue, and oxygen atoms colored red. (B) Negative-stain EM reconstruction at 21\u201322 \u00c5 of DRVIA7 Fab bound to JRFL SOSIP trimer (green surface) is superposed onto the previously published reconstruction of the BG505 SOSIP trimer bound to VRC01 (blue mesh, EMDB: EMDB-6252). (C) Ribbon diagrams of the variable regions of DRVIA7 (green) and VRC01 (blue) are shown after their structures were superposed on gp120 (white molecular surface). (D) The gp120-contacting regions (designated by their CDRs and FR regions) of DRVIA7 and VRC01 are displayed over the molecular surface of gp120 as in (C). The CD4-binding loop on the gp120 surface is colored in beige, the base of the V1/V2 loop in light blue, V5 loop in orange, and Loop D in yellow. The position of the N276 glycan is indicated by a black star. The light-chain CDR1s and N termini (Nt) display the most structural variance between the two antibodies and are highlighted by red dotted circles. (E) The N276 glycan, from their corresponding gp120 structures, is displayed when bound to VRC01 LCDR1 (blue, left), 12A21 LCDR1 (blue, middle), and DRVIA7 LCDR1 (light green, right). Below, coordinates for Man3GlcNAc2 attached to N276 from a previously published structure (PDB: 4P9H) are superposed on the protein-proximal GlcNAc visible in the VRC01, 12A21, and DRVIA7 structures. (F) The DRVIA7 and VRC01 light-chain N termini are shown interacting with the gp120 V5 loop. Below, the N276 and V5 loop glycosylation sites are colored red on the white gp120 surface. Ribbon diagrams of DRVIA7 N termini and LCDR1 loops are displayed as above. See also Table S2 and Figure S2.\n\n### Unfavorable Light Chain Accommodation of the Glycan Shield Limits DRVIA7 Neutralizing Ability\n\nAfter the structural analysis, we examined the effect of glycosylation on the binding of DRVIA7 and DRVIA7H-VRC01L, which showed an identical angle of approach when bound to the SOSIP gp140 trimer, as indicated by negative-stain EM (Figures S2D\u2013S2G). We first measured their binding toward glycosylated and deglycosylated gp120s of HIV-1 strains YU2, 93TH057, and JRFL. As reflected by the KD (koff/kon) values, DRVIA7 exhibited significantly higher affinity toward deglycosylated gp120 whereas the chimeric antibody displayed an opposite effect, confirming the inhibitory role of glycans for DRVIA7 binding (Figure 3A). Analysis of the binding kinetics revealed that deglycosylation improved the on rates for both antibodies, while showing a differential effect on the off rates (Figures 3A and S3E). Specifically, deglycosylation drastically increased the off rates for DRVIA7H-VRC01L, in comparison to a plateaued dissociation when gp120 was glycosylated. By contrast, deglycosylation showed a lesser effect on the off rates for DRVIA7. Our results thus highlighted the impact of glycosylation on CD4bs recognition by the VRC01-class bNAbs, consistent with another VRC01-class bNAb, PGV04 (Lyumkis et al., 2013). The specific role of the N276 glycan was further illustrated by antibody binding to YU2 gp120 lacking this glycosylation site (Figure 3B). DRVIA7 showed a more pronounced binding to the mutant gp120, while VRC01 and DRVIA7H-VRC01L recognized both gp120s in a similar manner.", "# 2025 \u2013 Bicyclic compounds as inhibitors of wrn\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-12 on marcdeller.com (categories: patents).*\n\n* [Link to Google Patent](https://patents.google.com/patent/US20250236622A1/en?oq=US20250236622A1)\n\n* New compounds are designed to powerfully disrupt WRN helicase, a DNA-repair enzyme cancer cells with microsatellite instability depend upon.\n* Tumors with high MSI/dMMR signatures are particularly vulnerable to this inhibition, while healthy tissues are largely spared.\n* These drugs act through a novel mechanism, distinct from immunotherapies, and can potentially overcome therapy resistance.\n* The compounds may provide much-needed options for patients resistant to checkpoint inhibitors, broadening cancer treatment possibilities.\n\n1. **Chu, W. et al. RecQ helicases: multifunctional genome caretakers.** [Nat Rev Cancer 9, 644-654 (2009)](https://pubmed.ncbi.nlm.nih.gov/19657341/)\n2. **Behan, F. M. et al. Prioritization of cancer therapeutic targets using CRISPR-Cas9 screens.** [Nature 568, 511-516 (2019)](https://pubmed.ncbi.nlm.nih.gov/30971826/)\n3. **Chan, E. M. et al. WRN helicase is a synthetic lethal target in microsatellite unstable cancers.** [Nature 568, 551-556 (2019)](https://pubmed.ncbi.nlm.nih.gov/30971823/)\n4. **Kategaya, L. et al. Werner syndrome helicase is required for the survival of cancer cells with microsatellite instability.** [iScience 13, 488-497 (2019))](https://pubmed.ncbi.nlm.nih.gov/30898619/)\n\nThis patent describes a new class of pharmaceutical compounds designed to block a crucial enzyme, WRN helicase, which plays a key role in maintaining the integrity of human DNA. Dysfunction of WRN is connected not only to premature aging, but also to the survival of certain aggressive tumors\u2014specifically those with high levels of microsatellite instability (MSI-H) or deficient mismatch repair (dMMR).\n\nAdvances in cancer genomics revealed that many MSI-H/dMMR cancers, including a significant percentage of endometrial, colon, and stomach tumors, are especially reliant on WRN activity for their survival. When WRN is shut off in these cancer cells, their DNA can\u2019t be properly managed or repaired, leading to cell death. However, tumors that lack this genomic signature (microsatellite stability) are left largely unharmed, indicating the target\u2019s selectivity.\n\nExisting treatments for these cancer types, such as immune checkpoint inhibitors (CPIs) like pembrolizumab and newer drug combinations, have brought progress but also face limitations: not all patients respond well, and resistance is common. WRN-targeting compounds introduce a fundamentally different biological approach, directly undermining cancer cell survival by preventing the WRN enzyme from carrying out its ATP-powered unwinding action necessary for DNA repair and replication.\n\nThe patent outlines several forms of the new inhibitors, methods of placing them into suitable pharmaceutical preparations, and strategies for clinical use. These compounds either block WRN\u2019s ability to unwind DNA (helicase activity), stop its energy source (ATPase activity), or both. Preclinical models suggest that these agents remain effective even in cancers that have already become resistant to standard therapies, including targeted agents, chemotherapy, or immunotherapy.\n\nThis innovation may transform treatment for patients with MSI-H/dMMR tumors, especially those who have exhausted other options. WRN inhibitors could be used alone or in combination with current approaches, offering hope of overcoming resistance, increasing the breadth of treatable cancers, and minimizing collateral damage to normal, non-cancerous cells.", "# Validation of Protein\u2013Ligand Crystal Structure Models: Small Molecule and Peptide Ligands\nAuthors: Pozharski, E.; Deller, M. C.; Rupp, B.\nSource: Protein Crystallography: Methods and Protocols (A. Wlodawer, Z. Dauter & M. Jaskolski, eds.), Methods in Molecular Biology vol. 1607, Chapter 25, pp. 611\u2013625, Springer, 2017\n\n| Metric | What it measures | Interpretation |\n|---|---|---|\n| RSCC (Real Space Correlation Coefficient) | Correlation of the model with the electron density, per residue | 1.0 is a perfect fit; 0.8\u20131.0 commonly considered good; lower values mean greater positional uncertainty |\n| RSR (Real Space R-Value) | Real-space residual between model and density | Somewhat scale-dependent |\n| RSRZ | Normalized version of RSR used by the PDB | Number of standard deviations a residue departs from the mean density fit |\n| LLDF (Local Ligand Density Fit) | Composite measure of ligand density fit | Artificially improved by correctly modelled regions of an incomplete ligand |\n| OWAB (Occupancy Weighted Adjusted B-factor) | Local displacement of ligand atoms | Lower OWAB, together with higher resolution, correlates with better RSCC |\n| Clash score (MolProbity) | All-atom close contacts | Bad contacts displayed as dots and spikes in COOT |\n| Ramachandran outlier percentile | Backbone torsion plausibility of peptide ligands | Unrestrained backbone torsions make outliers a genuine cross-validation, but the percentile is unreliable for very short peptides |\n\n*Table 3.* Map types used in ligand validation.\n\n| Map | Coefficients | Role in validation |\n|---|---|---|\n| Standard maximum-likelihood map | 2mFo \u2212 DFc | Shows the model in density but is inherently biased toward the model that supplied the phases |\n| Difference map | mFo \u2212 DFc | Positive density where model is missing, negative where the model is unsupported |\n| Omit difference map | mFo \u2212 DFc with the ligand deleted before refinement | The most definitive proof positive of ligand presence; must be generated by the user, not downloaded |\n\n## Key takeaways\n\n- Marc Deller co-authored this Methods in Molecular Biology chapter with Edwin Pozharski and Bernhard Rupp as a practical manual for validating protein\u2013ligand crystal structures, published in 2017 as Chapter 25 of volume 1607.\n- Because a bound ligand contributes only a small fraction of the total X-ray scattering, global statistics such as R-values and overall r.m.s. deviations contain almost no information about whether the ligand is real or correctly placed.\n- Validation rests on three pillars: the fit of the ligand to the electron density, compliance with reasonable stereochemistry, and chemically plausible interactions with the binding site.\n- Standard 2mFo \u2212 DFc maps deposited or served from public databases are model-biased, because their phases come from a model that already contains the ligand, so they can appear to confirm a ligand that is not there.\n- The definitive test is the omit difference map, calculated by deleting the ligand and re-refining, and the chapter recommends running more than one refinement program \u2014 REFMAC, phenix.refine and BUSTER-TNT \u2014 to confirm the density is robustly reproduced.\n- An RSCC of 1.0 is a perfect fit and values between 0.8 and 1.0 are commonly considered good; RSCC improves with higher resolution and lower occupancy-weighted B-factors.\n- Incomplete or over-enthusiastically modelled ligands are dangerous precisely because averaged local scores hide them, so only residue-by-residue local inspection reveals which parts of a ligand are experimentally supported.\n- Peptide ligands offer a free cross-validation because their backbone torsions are unrestrained, so unsupported peptide models refine into implausible high-energy conformations that appear as Ramachandran outliers.\n- Comparing ligand B-factors with those of surrounding protein atoms is a simple diagnostic: large mismatches suggest partial occupancy or an incorrectly built ligand.\n- Alternative explanations for apparent ligand density must be excluded, including ordered water, components of the mother liquor or cryoprotectant such as short PEG fragments, and symmetry-related protein fragments.\n- The required level of scrutiny depends on the intended use, with structure-based lead discovery generally demanding around 2 \u00c5 resolution or better and much closer inspection than structure-guided mutagenesis.\n- Emerging low-resolution and hybrid methods \u2014 cryo-EM, XFEL, ultrafast electron diffraction and DEER, often in the 4\u201310 \u00c5 range \u2014 will place greater weight on ligand restraints and demand new definitions of ligand stereochemistry, and the authors call for mandatory validation efforts of the kind initiated by the PDB.\n\n## References\n\n*(reference list omitted)*", "# ChatRDKit \u2013 Talk to RDKit for Chemoinformatics in plain language\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n```\n[chatRDKit] Output format? (1) PNG  (2) HTML  (3) PDF  [1]:\n```\n\n* **PNG** \u2014 a depiction image (single molecule or a grid for several)\n* **HTML** \u2014 a standalone report with the image, SMILES, and formula\n* **PDF** \u2014 the same report as a PDF (via `reportlab`; falls back to PNG\n  with a warning if `reportlab` isn\u2019t installed)\n\nFiles are written to the current directory by default. Set `output_dir` to\na path to always use that directory, or to `ask` to be prompted for a\ndirectory each time (configurable, see below).\n\n---\n\n## 3. Configuration\n\nAll settings are stored in `~/.chatrdkit/config.json` and can be viewed/\nchanged with `/config`:\n\n```\nchatRDKit> /config                          # list everything\nchatRDKit> /config model                    # show one value\nchatRDKit> /config model qwen3:0.6b         # change the model\nchatRDKit> /config host http://localhost:11434\nchatRDKit> /config temperature 0.1\nchatRDKit> /config output_format png        # always PNG, never ask\nchatRDKit> /config output_dir ~/Desktop/chatrdkit_out\nchatRDKit> /config debug on                 # print raw LLM output (troubleshooting)\n```\n\n| Setting | Default | Meaning |\n| --- | --- | --- |\n| `host` | `http://localhost:11434` | Ollama server URL |\n| `model` | `qwen3.6:27b` | Ollama model tag (must be pulled, must match `ollama list`) |\n| `temperature` | `0.2` | LLM sampling temperature |\n| `timeout` | `180` | Request timeout (seconds) |\n| `echo` | `True` | Print each translated line before running it |\n| `debug` | `False` | Print the raw LLM response (before cleanup) to the console |\n| `max_history` | `6` | Number of past request/response turns kept for follow-ups |\n| `output_dir` | `.` (current directory) | Where `render()` saves files \u2014 a path, or `ask` to prompt each time |\n| `output_format` | `ask` | `ask`, `png`, `html`, or `pdf` \u2014 set to skip the format prompt |\n\n---\n\n## 4. How it works\n\nchatRDKit is a single file, `chatrdkit.py`, organised into sections:\n\n```\nchatrdkit.py\n\u251c\u2500\u2500 config       JSON-backed settings (~/.chatrdkit/config.json)\n\u251c\u2500\u2500 ollama_client minimal urllib client for POST /api/chat\n\u251c\u2500\u2500 rdkit_reference system prompt + curated RDKit cheat-sheet + session context\n\u251c\u2500\u2500 render       PNG/HTML/PDF output via Draw.MolsToGridImage\n\u251c\u2500\u2500 session      persistent exec namespace (RDKit imports + render())\n\u251c\u2500\u2500 router       SMILES/code/English classification, translation, execution\n\u2514\u2500\u2500 cli          welcome banner, REPL loop, slash commands\n```\n\n1. **Session.** A single Python namespace (dict) is created at startup with\n   `Chem`, `AllChem`, `Draw`, `Descriptors`, `Lipinski`, `Crippen`, `QED`,\n   `DataStructs`, `rdFingerprintGenerator`, `rdMolDescriptors`, `rdMolDraw2D`,\n   `math`, and a `render()`\n   helper already in scope. Every molecule you create persists here for the\n   rest of the session.\n2. **Classifying input** (`router.classify`):\n3. `looks_like_smiles()` \u2014 no spaces, only SMILES-valid characters, and\n   `Chem.MolFromSmiles()` parses it successfully.\n4. `looks_like_code()` \u2014 valid Python (`ast.parse`) that contains an\n   assignment, call, attribute access, control flow, etc., or is a bare\n   name that already exists in the session (e.g. typing `mol`).\n5. Anything else is treated as English.\n6. **Translation** (`router.translate`). English requests are sent to\n   Ollama with:\n7. a system prompt describing the exec environment, the `render()`\n   helper, and a curated RDKit cheat-sheet (`rdkit_reference.py`,\n   drawn from the [Getting Started with RDKit in\n   Python](https://www.rdkit.org/docs/GettingStartedInPython.html) guide),\n8. `rdkit_reference.build_context()` \u2014 a live list of every molecule\n   currently in the session and its variable name, so \u201cit\u201d/\u201dthe\n   molecule\u201d/a compound name resolves correctly,\n9. the last few turns of conversation for follow-ups.\n\nThe model is instructed to return **only** a Python/RDKit snippet (no\nmarkdown, no explanations), reuse existing session variables, create\nwell-known compounds from SMILES using its own chemistry knowledge, and\ncall `render()` for anything visual. Generation is capped\n(`num_predict`) and any response that comes back too long or repetitive\nis discarded \u2014 and not added to history \u2014 rather than executed.\n\n1. **Execution** (`router.execute_code` / `session.run`). Each snippet is\n   echoed as `[chatRDKit] >>> ...` and run in the session namespace. A\n   single bare expression\u2019s value is printed (REPL-style); otherwise normal\n   `print()` output from the snippet appears directly. Errors are caught and\n   reported without crashing the session.\n2. **Rendering** (`render.render`). Computes 2D coordinates if missing,\n   builds a grid image with `Draw.MolsToGridImage`, asks for a format if not\n   configured, and writes PNG/HTML/PDF to `output_dir`.\n\n---\n\n## 5. Safety notes\n\n* chatRDKit executes whatever code the model returns, in-process, in the\n  same Python interpreter. A basic safety check refuses snippets containing\n  `import os/sys/subprocess/shutil/socket/requests`, `open(`, `eval(`,\n  `exec(`, `__import__`, or direct `os.`/`sys.`/etc. attribute access \u2014 the\n  system prompt also instructs the model not to use these. This is a\n  best-effort net, not a sandbox: review echoed `[chatRDKit] >>> ...` lines\n  for anything unexpected.\n* Everything runs locally against your Ollama server \u2014 no data leaves your\n  machine.\n* `/config debug on` prints the model\u2019s raw output (including any stripped\n  `<think>` blocks or code fences) so you can see exactly what was executed.\n\n---\n\n## 6. Troubleshooting\n\n**`Could not reach Ollama at http://localhost:11434 ...`**\nStart Ollama (`ollama serve` or open the desktop app) and confirm the model\nis pulled (`ollama pull qwen3.6:27b`, check with `ollama list`).", "# Crystal Structure of PGT135 Fab in Complex with gp120 Core Protein from HIV-1 Strain JR-FL Bound to CD4 and 17b Fab\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [4JM2](https://www.rcsb.org/structure/4JM2)\n* **Release Date:** 2013-05-29\n* **Method:** X-RAY DIFFRACTION 3.1 \u00c5\n* **Organisms:** Homo sapiens, Human immunodeficiency virus 1\n* **Macromolecules:** Unique protein chains: 6 (proteins)\n* **Unique Ligands:** NAG, PG4\n* **Publication:** [Kong, L., et al. (2013). Supersite of Immune Vulnerability on the Glycosylated Face of HIV-1 Envelope Glycoprotein gp120. Nature Structural & Molecular Biology, 20, 796-803.](https://www.nature.com/articles/nsmb.2594)\n* **Blog Post:** [Supersite of Immune Vulnerability on the Glycosylated Face of HIV-1 Envelope Glycoprotein gp120](https://marcdeller.com/supersite-of-immune-vulnerability-on-the-glycosylated-face-of-hiv-1-envelope-glycoprotein-gp120/)\n\n**Description:** This 3.1 \u00c5 crystal structure defines the Asn332 glycan-dependent \u2018supersite of immune vulnerability\u2019 on HIV-1 gp120, showing how broadly neutralizing antibody PGT135 uses extended CDR loops to penetrate the glycan shield and contact glycans at Asn332, Asn392, and Asn386. The discovery of this extensive antigenic region makes it a prime target for HIV vaccine immunogen design.", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\nA campaign with no apo structures anywhere never aborts the rest of the pipeline over an\noptional, additive feature. The standalone command still exits with a clear error if you\nexplicitly pass a `--family`/`--target` matching nothing, since that's a real mistake worth\nstopping for.\n\n## \ud83c\udf7a Landlord: a summary in plain English\n\nNamed after Timothy Taylor's flagship, as BoltzMaker is named after their Boltmaker.\n\nAt the end of `analyze`, Landlord writes a summary of the campaign and of every target\nin it: what the confidence means, what each ligand's contacts and predicted potency\nshow, whether a pose reproduced its experimental structure, and a verdict. It appears\nas the last panel of the dashboard and of the hosted explorer, and as\n`boltz_summary_prose.json` beside the CSVs.\n\nOn an Apple Silicon Mac running macOS 26 with Apple Intelligence switched on, the prose\nis written **on-device by the Neural Engine**. No weights ship, no data leaves the\nmachine, no API key exists to leak. Everywhere else -- Intel Macs, Linux, an older\nmacOS, Apple Intelligence off -- the same summary is rendered from a template, and the\ncampaign neither notices nor cares.\n\n| Setting | What it does |\n|---|---|\n| `--narration auto` | On-device if possible, template otherwise. The default. |\n| `--narration template` | Template even where the model would work. Deterministic, so the output is reproducible. |\n| `--narration off` | No summary. |\n| `--narration model` | On-device only, failing loudly. For testing that path; deliberately not offered in the web form. |\n\n### It never computes, and it is checked when it does\n\nEvery number a summary can contain is computed, rounded, given its units and turned\ninto a string *before* the model sees it, and every ranking and threshold judgement is\nalready made. The model is handed `\"ipTM 0.84\"` and `rank: \"2 of 4\"` -- it has nothing\nto round and no list to sort.\n\nThat is not fastidiousness. Asked to compose the campaign-level findings, the model\nwrote the caution count as the number of discards and the flagged count as the\nconfidence count. Both figures were in its input. So anything that is arithmetic --\nthe verdict, the tallies, the rankings -- is computed in Python, and the model\nnarrates rather than derives. Per-target verdict mismatches went from three in six to\nnone.\n\nWhat the model does write is then gated: **every numeric token in generated prose must\nappear verbatim in the fact block it came from**, or the summary is discarded and the\ntemplate used instead. On its first run against real output the gate caught the model\ninventing \"confidence scores above 0.8\" -- true, as it happens, and never supplied.\n\nThe gate checks presence, not attribution. It cannot tell that a supplied number has\nbeen attached to the wrong noun, which is exactly why the tallies are not narrated.\n\n### Nothing here can fail a campaign\n\nMissing binary, Apple Intelligence off, model still downloading, a timeout, unreadable\noutput, a summary rejected by the gate: every one of those ends at the template.\nNarration runs after the analysis is already on disk, at low QoS, and a failure costs\nthe summary and nothing else.\n\n### It does not compete with the GPU\n\nMeasured on an M1 Max: GPU power *fell* during narration, 15.0 to 8.2 mW, so nothing\ncontends with a folding run -- narration can happen while a campaign is still going.\nThe inference path uses about 6% of one core. `powermetrics` exposes no ANE rail on\nthis SoC, so the work is placed on the Neural Engine by attribution rather than by a\npower reading: not the GPU, not the CPU, and `aned` busy throughout.\n\n`docs/landlord_spike.md` has the feasibility findings, including a 0% refusal rate\nacross 26 fact blocks weighted towards the vocabulary most likely to trip a content\nfilter. `docs/landlord_bench.md` has the timings -- including why concurrency is 1 --\nand `docs/landlord_packaging.md` why there is nothing to notarise.\n\n## \ud83c\udf10 Web deployment\n\n**Live at [boltzmaker.mdeller.com](https://boltzmaker.mdeller.com).** A Flask frontend for the\nnon-GPU stages, in the two modes introduced under **Installation** above. The GPU `run` step is\ndeliberately never hosted: there's no GPU on the droplet, so that stage always runs on your own\nhardware.\n\n### Fully Automated Mode, Step 1: Prepare\n\nDescribe the campaign in the form: a **name**, whether to **predict binding affinity**, then your\n**partners** (optional co-folded chains), **proteins**, **constraints** (pocket contact, covalent\nbond, or distance), and **ligands** (SMILES or CCD code). Same fields `BoltzMaker.py new` asks\nfor, and the same 5-character shared-namespace rule on every short name.\n\nFour fields replace something the form used to make you do by hand:", "# Protein stability: a crystallographer's perspective\nAuthors: Deller, M. C.; Kong, L.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 72, 72\u201395 (IYCr crystallization series), 2016\n\nEven a compositionally homogeneous sample will usually fail to crystallize without conformational stability. A protein with substantial disordered regions or domains showing dynamic variability is less likely to self-organize into a lattice. The problem is compounded because the conformation of flexible regions is context-driven \u2014 conformations in the cell, in an NMR tube and in a crystal may differ \u2014 so results from any structural method are only meaningful within the conditions of that method. Crystallography gives limited dynamics information, encoded in the atomic displacement parameter or B factor (in units of \u00c5\u00b2), a statistical measure of the probability of finding an atom at its mean position; high B factors indicate low positional certainty arising from dynamic disorder of the chain or short- or long-range lattice disorder. NMR ensembles provide a complementary measure, the r.m.s.d. between corresponding atoms, unperturbed by crystal packing. Intrinsically disordered proteins (IDPs) are a large class: an estimated 40% of human proteins contain at least one disordered segment and 25% are completely disordered. NMR studies of IDPs such as CREB, p53 and 14-3-3 showed that disorder allows plasticity and that structure often forms only on partner binding \u2014 a paradigm shift from the rigid lock-and-key view. Aggregation-prone regions (APRs) of 5\u201315 residues with \u03b2-sheet propensity drive amyloid formation, as in \u03b22-microglobulin amyloidosis, and intrinsically insoluble proteins (IIPs) such as the V22-SH3 mutant are insoluble in normal buffers yet can be solubilized in pure water.\n\n### 3. Stability of the protein on a structural level; 3.1 Primary structure\n\nThe authors conceptualize stability level by level \u2014 primary, secondary, tertiary and quaternary \u2014 and pair each with definitions, example biochemical processes and analytical methods (Table 2). At the primary level, stability concerns changes to the amino-acid sequence or chemical modification of residues by PTMs. PTMs affect both compositional stability (modifications are often non-uniform or incomplete) and conformational stability (glycans add surface disorder while shielding polar and charged residues needed for crystal contacts). Functional groups such as myristate, palmitate, isoprenoid and GPI anchors target proteins to membranes; carboxylate, ethanolamine phosphoglycerate and hypusine regulate activity; and whole proteins including ubiquitin, SUMO, ISG15, PUP and NEDD are conjugated. In the survey by Khoury and coworkers of 821,182 experimentally analyzed proteins, the ten most frequent PTMs were phosphorylation (58,383), acetylation (6,751), N-linked glycosylation (5,526), amidation (2,844), hydroxylation (1,619), methylation (1,523), O-linked glycosylation (1,133), ubiquitylation (878), pyrrolidone carboxylic acid (826) and sulfation (504); just under 500 PTMs are catalogued in SWISS-PROT and TrEMBL. Protein splicing by inteins \u2014 first seen in the yeast VMA1 vacuolar H+-ATPase \u2014 is highlighted both as a primary-structure phenomenon and as an engineering tool underpinning self-cleaving purification tags such as cSAT and the intein-CBD IMPACT system (NEB).\n\n### 3.2. Secondary structure\n\nSecondary structure is described either by the hydrogen-bonding pattern between backbone amide H and carbonyl O atoms (Pauling) or by backbone stereochemistry (Ramachandran); its principal driving forces are hydrogen bonding and the hydrophobic effect. The \u03b1-helix predominates, accounting for roughly one third of all secondary-structure elements. Early crystal structures showed alanine, leucine and glutamate to be common in helices and proline, glycine and aspartate to be rare, information that fed prediction algorithms such as the Chou\u2013Fasman method, later refined by mutagenesis-derived \u03b1-helical and \u03b2-sheet propensity tables. The strongest consensus is proline's destabilizing effect on helices (\u0394\u0394G of 3.16 kcal mol\u207b\u00b9 compared with alanine at 0 kcal mol\u207b\u00b9), arising from the missing backbone amide H atom, which cannot participate in hydrogen bonding, and from a roughly 30% kink in the helix backbone caused by the bulky cyclic side chain. Glycine has the next lowest helix propensity owing to enhanced conformational flexibility. Propensities are strongly context-dependent: proline is common in transmembrane helices and can stabilize them. These principles, consistent with Anfinsen's kinetic free-energy minimum, now support algorithms for designing stable helices and sheets, thermally stabilized proteins by comparative modelling, and crystallizability prediction.\n\n### 3.3. Tertiary structure\n\nTertiary structure \u2014 the overall fold \u2014 is influenced by conformational and compositional stability, cellular environment, temperature and pH, primary and secondary structure, solvation, hydrogen bonding, salt bridges, hydrophobic effects, van der Waals forces, ligand, cofactor and ion binding, chaperones and PTMs. Folding carries a large entropic penalty (T\u0394S \u226b 0), so a folded protein is only marginally stable \u2014 about 10 kcal mol\u207b\u00b9 for a 10 kDa protein. Quantitative work by Pace and coworkers attributes roughly 60% of stability to the hydrophobic effect and 40% to hydrogen bonding: burial of a single methyl group contributes about 1.1 kcal mol\u207b\u00b9 to net stability while loss of its conformational entropy contributes about 2.4 kcal mol\u207b\u00b9 to instability, and the net contribution of hydrogen bonding is likewise about 1.1 kcal mol\u207b\u00b9 and largely independent of protein size, whereas hydrophobic interactions contribute less in small proteins. Thermophilic proteins, for example from *Thermotoga maritima*, resemble their mesophilic counterparts but show more salt bridges, greater hydrophobicity and more aromatic residues. Fold stability is not universally required for function: serpins convert from a stressed (S) native form to a relaxed (R) form, and large fold changes also occur in influenza haemagglutinin, lymphotactin, Mad2 and CLIC1.", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n**The ordering is the result.** Own site beats unconstrained beats wrong site, by an order of\nmagnitude at each step, on two different ligands against two different receptors. A restraint\nthat fabricated a comfortable pose wherever it was aimed would put the wrong-site rows near the\nown-site rows; instead they are the worst numbers in the campaign, worse than having no\nconstraint at all. So the pocket is steering, not scoring -- and the panel is measuring the\nprediction.\n\nThe conformer column stays low for 41Y even when it is 24.93 \u00c5 from home (0.81, 1.78): the\nmolecule's own shape is right and only its placement is wrong, exactly the decomposition the\nthree columns exist for. These numbers also come from a different campaign than the tightness\ntable above, which is why orforglipron's own-site pose reads 2.79 \u00c5 here against 2.56 \u00c5 there --\none diffusion sample per condition, so treat sub-\u00c5 differences between campaigns as noise and the\norder-of-magnitude steps as the finding.\n\n**A pocket code is a PDB chemical component id, so look it up.** `41Y` was carried through that\ncampaign as nothing but a pocket label. Its formal definition is a molecule -- `C32H41FN2O2`,\nappearing in exactly one PDB entry -- and it turned out to be the same compound as one of the\ncampaign's own ligands, which meant that ligand had an experimental structure nobody had fetched.\nDownloading 7RBT into `reference/` doubled the pose panel's coverage and supplied the second\nligand in the table above. If a pocket is named after a component id, resolve it against the\nchemical component dictionary before assuming the campaign has no experiment to check itself\nagainst.\n\n## \ud83e\uddec compare-sse: apo vs holo secondary-structure shifts\n\n**Why this exists:** a confidence score tells you *what* Boltz predicted, not *how the protein\nmoved* in response to ligand binding -- a real structural question whenever you have both a\nreference apo (unbound) structure and a predicted holo one for the same protein. `compare-sse`\nanswers it in terms a structural biologist actually reasons in (\"TM6 swung out 4.2 Angstrom\",\n\"the DFG motif flipped from in to out\"), not raw DSSP fragment coordinates.\n\nIt is a core part of `analyze`/`all`: any family with `Apo structure:` set is compared\nautomatically and the result embedded into `boltz_dashboard.html`. A family with no apo structure\nisn't silently skipped -- the \"Family coverage\" table says so explicitly. Pass `--skip-sse` to opt\nout, or use the standalone command to re-run just this analysis (`--family`/`--target` scope it).\n\nMotifs are annotated by one of three pluggable sources, auto-selected per family or set\nexplicitly with `Family type:`:\n\n| Family type | Motifs | Source |\n|---|---|---|\n| `gpcr` | TM1-7, H8, ECL1-3, ICL1-3 | [GPCRdb](https://gpcrdb.org)'s structure-based generic-numbering service (Ballesteros-Weinstein / GPCRdb schemes) |\n| `kinase` | hinge, gatekeeper, catalytic loop (HRD), DFG motif, alphaC-Glu, catalytic Lys | [KLIFS](https://klifs.net)'s public REST API (its fixed 85-residue pocket alignment) |\n| `auto` (default) | whichever of the above applies, else... | ...falls back to Pfam domain boundaries via [PDBe](https://www.ebi.ac.uk/pdbe)'s SIFTS residue mapping -- the universal last resort for any protein outside the two families above |\n\nApo is superposed onto holo using only the family's stable, non-binding-site-adjacent residues\n(via gemmi's `superpose_positions`), so a ligand-induced local shift can't skew the global fit.\nEach motif then gets:\n\n| Metric | What it means |\n|---|---|\n| Ca RMSD / centroid shift | How far the motif moved, post-superposition |\n| Helix-axis rotation angle | For helical motifs -- e.g. the classic TM6 \"outward swing\" on GPCR activation |\n| Helix kink angle (apo/holo/delta) | Whether a helix straightened or kinked more |\n| SSE boundary shift | Did the helix/strand get longer or shorter -- needs deposited HELIX/SHEET records, or an optional external `mkdssp`/`dssp` binary as a fallback (see **One-time setup** above); every other metric works without it |\n| Flagged phi/psi residues | Per-residue backbone dihedral outliers above `--phi-psi-threshold` |\n| DFG-in/out, alphaC-in/out (kinases only) | DFG uses Dunbrack's two-distance criterion on the DFG-Phe ring; alphaC uses the catalytic-Lys/alphaC-Glu salt bridge. Both are distance criteria on the atoms that actually move, not full dihedral models, and DFG reports `other` for a conformation that is neither. alphaC is unresolved rather than `out` when the side chains are unmodelled |\n\nA metric that genuinely wasn't computed for a motif (axis rotation for a loop, DFG state for a\nnon-kinase family, boundary shift with no DSSP data) shows as an explicit `N/A` in both the CSV\nand every dashboard table, not a blank cell.\n\nAbove the per-motif table, both the dashboard's embedded card and the standalone\n`boltz_sse_comparison.html` show a **family coverage** row per protein family -- `OK` (with a\ntarget/motif count), `No apo structure configured`, `Apo structure file not found`,\n`No motif annotation available` or `No predicted (holo) structures yet` -- and **overall shift\nstatistics**: targets/motifs compared; mean/median/max Ca RMSD, with which target and motif had\nthe largest shift; mean centroid shift; total flagged phi/psi residues; and kinase DFG/alphaC\nstate-change counts.\n\n```sh\npython3 BoltzMaker.py compare-sse boltz_input.md\n```\n\nWrites, next to `boltz_input.md` (or `--out-dir`): `boltz_sse_comparison.csv`,\n`boltz_sse_family_status.json`, a standalone self-contained `boltz_sse_comparison.html` (Plotly\nbar chart + motif x target heatmap, vendored the same way as the main dashboard), and\n`boltz_sse_comparison_sessions/<target>.pml` -- a plain-text PyMOL script per target that\ncolours/labels each motif and highlights significant shifts. It's just text: opens in any local\nPyMOL install, no `pymol` dependency in BoltzMaker's own venv.", "# GitHub README: PANTS\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/PANTS, 2026\n\n- [x] **Repository scaffold and the two-venv split.** `requirements-web.txt` carries Flask, gunicorn and gemmi and nothing else, so the always-on droplet process never imports torch. The droplet has 3.8 GB shared with five other applications, which makes this a memory constraint rather than a style preference\n- [x] **SQLite schema with manifest provenance.** Thirteen tables, WAL enabled once at creation, every pipeline stage opening a run and writing input/output hashes, tool versions, git commit and wall time to both a table and a JSON file. A stage that raises still leaves its manifest behind\n- [x] **Verify torch on Python 3.14 (plan risk 8).** torch 2.13.0 installs cleanly with MPS available, so no fallback to 3.11 was needed\n- [x] **Install and pin the external tools.** HMMER 3.4 and MMseqs2 18-8cc5c, shelled out to rather than bound as libraries, with versions captured in every manifest\n- [x] **Move bulk data out of the iCloud tree.** `data/raw` and `data/interim` symlink to `~/PANTSData`; macOS \"Optimize Mac Storage\" evicts large files mid-run and this machine has only ~62 GB free\n- [x] **Curate the characterised seed set.** Wild types fetched from UniProt by accession; engineered variants derived from parent plus mutation list, with every substitution checked against the parent residue it names. All four confirmed variants validated at offset 0 across 14 positions\n- [x] **Harvest ESTHER hard negatives.** Matched on five axes: length distribution, identity to nearest positive, genus cap, signal peptide and phylum\n- [x] **Run the trivial-baseline gate before any embedding work (plan risk 1).** It fired, and found both that the curated positives are one cluster rather than nine examples, and that the negatives were separable on a secreted-versus-cytoplasmic composition signature\n- [x] **Record an evidence level on every positive.** UniProt `protein_existence` separates the 23 with protein-level evidence from the 56 predicted or inferred, so the two are never pooled in a reported metric\n- [x] **Make the composition baseline a permanent reported metric.** Stored alongside the retrieval baseline in `training_runs`, with `n_positive_clusters` recording independent units rather than the raw count\n- [x] **Curate measured activity data.** Taken from UniProt's machine-readable, citable annotations rather than transcribed from PDFs, which is where fabrication risk lives. `EC 3.1.1.101` (poly(ethylene terephthalate) hydrolase) gave 459 entries and took the positive set from 87 sequences in 11 clusters to 529 in 29. 47 measurements extracted (21 Km, 8 Topt, 8 pH optima, 10 qualitative), each carrying its PubMed IDs, with `comparable_group_id` keyed on parameter plus substrate so a Km on pNP-butanoate is never pooled with one on PET film\n- [x] **Add the gut microbiome as a fourth source environment.** The three current environments (compost, marine plastisphere, landfill) are all external. A human gut metagenome is the one that matters most for the therapeutic framing: an enzyme already resident at 37 \u00b0C, pH 7.4 and in a proteolytic environment has been selected under something close to the target conditions, rather than being asked to work far from its optimum. **HGMP01** is the concrete starting point: it is named in spec section 1, it is metagenome-derived rather than a variant of a characterised parent, and it is currently the one named enzyme the recall stage is expected to recover from sequence space instead of being seeded with. Done: **12,584,458 gut proteins from five MGnify studies**, registered as `human_gut` and now 71% of the catalogue. **HGMP01 itself could not be seeded**: it returns zero hits in UniProt and in NCBI protein/nuccore, and its paper (PMID 39551294) has no linked sequence records, so the sequence sits in supplementary material. That is not a blocker but an advantage, because the same paper reports HGMP01-like genes as widely distributed across the gut microbiome, so recall finding them **unaided** is a blind test rather than a circular one\n- [x] **Obtained all five HGMP sequences.** No public database holds them (zero hits across UniProt, NCBI protein and nuccore), but the authors deposited them on SciDB. **The paper and the deposit use different numbering**: the paper's HGMP01 to HGMP05 are the deposit's HGMP03, 04, 06, 07 and 08, so matching on name would have mislabelled all five. The mapping is pinned by two independent facts, sequence length (all five distinct, so 1:1) and homologue count (the paper's \"697 putative HGMP01-like enzymes\" matches the deposit's HGMP03 exactly). **HGMP01 = `GUT_GENOME238302_00589`, 275 aa, optimum 40 \u00b0C at near-neutral pH**: the only measured PET hydrolase in this project whose optimum is close to physiological\n- [x] **Built the ordinal-within-paper fallback, and established why it is the route rather than a fallback.** Before this the database held exactly **one** quantitative value measured on PET; all 21 Km values were on soluble pNP-ester proxies. Attempting the real thing showed why: **PET rates live in figure panels**, not in extractable text or structured deposits, so they cannot be harvested at scale. The brief anticipated this and proposed within-paper ordinal ranking, which now exists as a `parameter_type` with `ordinal_rank_in_paper`, seeded from PMID 39551294 (HGMP01 first of five on PET nanoparticles, the other four recorded as equal-second because the paper does not order them). Assay conditions are attached so the comparison is interpretable\n- [x] **Got measured within-family negatives \u2014 29 to 156.** The binding constraint, broken by two 2025 screens that assayed panels under one protocol and reported what did NOT work, both with openly deposited source data. **ACS Catalysis 2025** ([10.1021/acscatal.5c03460](https://doi.org/10.1021/acscatal.5c03460)): 477 proteins with sequences, 216 assayed, **88 measured inactive** at a stated 0.1% depolymerisation floor, and assayed at **40 \u00b0C** as well as 60 \u00b0C \u2014 almost every optimum in this catalogue sits between 50 and 78 \u00b0C and the therapeutic target is 37 \u00b0C. **Science 2025** ([10.1126/science.adp5637](https://doi.org/10.1126/science.adp5637), Data S3): 2,064 library entries, 183 assayed, **69 measured inactive** against each entry's *own* replicate standard deviation \u2014 active above 2 SD, inactive at or below 1 SD, and the 12 in between labelled neither way and excluded from training, because those are exactly the enzymes a threshold would be tuned on. The distinction that matters: PAZy's negatives mean \"not reported active\", which cannot separate tested-and-failed from never-tested; these were expressed, assayed and found not to release product. Measurements went 75 to 2,759 and the catalogue 1,140 to 1,348\n- [x] **Re-ran every evaluation on the measured negatives, and found the constraint was never the count.** Negatives 26 to 149; clusters holding BOTH an active and an inactive enzyme, **7 to 8**. Five times the data bought one independent lineage, because the binding quantity is not how many negatives exist but how many lineages anyone has measured on both sides. On the full measured within-family contrast the head scores **0.648** and amino-acid composition scores **0.644** \u2014 no advantage at all over counting residues. Restricted to mixed clusters, the hardest honest test, it reaches **0.625 against composition's 0.519**. The real gain is the error bar: that test previously read 0.435 \u00b1 0.301, an interval that excluded nothing, and now reads **0.625 \u00b1 0.081**. The model did not improve; the experiment became powered enough to give a stable answer, and the answer is a modest signal that is not composition. Also worth recording: adding polyesterase negatives collapsed the composition baseline from 0.803 to **0.565** and retrieval from 0.952 to **0.787**, so the earlier numbers were flattered by an easy negative class, and the head-versus-retrieval gap stayed at zero (\u22120.009 before, +0.003 now) while both fell\n- [x] **Asked the geometry question of measured labels, and the earlier signal vanished.** Active-site geometry was the one signal here measured off coordinates rather than inherited from an annotation, so it was the obvious thing to re-test. On inferred negatives cleft depth separated PET-active from PET-inactive at **AUC 0.808, p 1.7e-07**, falling to 0.534 cluster-grouped \u2014 read at the time as an underpowered evaluation waiting for data. Asked of **237 PET-active against 139 measured-inactive** \u2014 the complete folded set, 205 new structures over 23 hours \u2014 all ESMFold on both sides so no coordinate-source confound applies, the answer is **AUC 0.398 (0.360\u20130.453 across ten random splits)**. Below chance, and every split lands there: the relationship the features carry inside training clusters *reverses* in held-out ones, which is the signature of fitting lineage rather than activity. The trajectory is the strongest part of the evidence \u2014 0.507 at 95 negatives, 0.498 at 125, 0.459 at 139, 0.398 with the positives complete. A real effect emerging from a thin sample strengthens and narrows around a non-null value; moving toward and past chance as the sample grows is a confound being diluted, and that distinction is exactly what let the earlier reading survive as \"underpowered\" for as long as it did. A coin flip, with an interval finally tight enough to exclude something. Not one of the nine features survives at p<0.01 and cleft depth comes in at **0.529, p 0.38**. That 0.808 was never measuring PET activity; it was measuring whatever separates the enzymes a database happens to list from the ones it does not, and it disappeared the moment the labels came from somebody running the assay. One limit stated with it: only 8 clusters hold both classes, and every structure is a prediction, so if ESMFold idealises the active site \u2014 which the paired crystal-versus-model comparison shows it does \u2014 a real difference could be flattened by the modelling rather than absent from the enzymes\n- [x] **Found why nothing transfers: the determinants are lineage-specific.** Three independent methods had failed identically \u2014 sequence embeddings, a learned head and active-site geometry all work out-of-family and collapse within it \u2014 and the usual reading is too little data. This test asks instead whether a global rule can exist at all, and needs no train/test split, so it cannot be leakage in disguise: fit the same model separately inside each of the three lineages large enough to support one, then compare the directions the fits point in. The reference is not zero but an empirical **ceiling** \u2014 how well two bootstrap resamples of the *same* lineage agree, since noisy vectors from an identical process do not agree perfectly either. **Two fits of the same lineage agree at ~+0.70. Two different lineages agree at roughly zero, and in two of three pairs point in OPPOSITE directions** \u2014 what predicts activity in one family predicts inactivity in another. The same pattern appears in two feature sets sharing nothing, one from 3D coordinates and one from a language model over sequence. The high ceiling is what makes it a result rather than a shrug: within-lineage directions are well determined, so it is specifically the between-lineage comparison that fails. It also explains the below-chance numbers \u2014 a model trained across lineages learns a direction that reverses on an unseen one, which is exactly how geometry reached 0.398 and why an 18x larger language model changed nothing. The problem was never capacity and never only sample size: **there may be no single global rule to learn.** Established on three lineages, so a strong indication rather than a settled law, but the first explanation that accounts for every failure at once\n- [x] **Measured where the signal holds, and concluded the deliverable is retrieval.** With a global rule ruled out, the remaining question is the one a user asks: how close must a candidate be to characterised enzymes before a prediction means anything? Re-clustering inside the 262-member lineage (171 measured-active, 91 measured-inactive) at 95/90/80/70/60/50% identity and holding out whole sub-clusters gives the curve. Read as point estimates it shows an operating band \u2014 head **0.703** against retrieval's 0.544 at 70%. **Read with intervals it is not there**: the paired lead, head minus retrieval on the *same* held-out sub-clusters, is **+0.159 [\u22120.263, +0.390]**, spanning zero. Seventeen sub-clusters cannot resolve a difference that size, and the curve is not monotonic \u2014 amino-acid composition beats both at 80%. One thing survives: at 70% the head's own interval **[0.521, 0.798]** excludes 0.5, so within a lineage among relatives sharing ~70% of their sequence there is weak evidence of better-than-chance discrimination. It simply cannot be shown to beat looking up the nearest known PETase. **So PANTS should ship nearest-neighbour retrieval, not a learned classifier** \u2014 simple, untrained, no risk of learning a lineage direction that reverses, and nothing built here beats it on labels somebody measured. A smaller claim than the project set out to make, and the one the evidence supports\n- [x] **Switched candidate ranking to retrieval, with a stated competence range.** Candidates were ranked by HMM bitscore, which says a sequence matches the polyesterase profile \u2014 the easy question, answered at AUC 0.98, and not what anyone is here for. They are now ranked by identity to the nearest enzyme whose PET activity was **measured**, which is different from the existing `nearest_characterised_id`: that field points at the nearest entry in the whole catalogue, usually one carrying an EC number assigned automatically by similarity, so it could rank a candidate highly for resembling something nobody ever assayed. Schema v17 adds `nearest_measured_id`, `nearest_measured_identity` and `retrieval_band`. Every candidate now carries its band, from the identity-decay measurement: **16 in range (\u226570%), 39 marginal (50\u201370%), 384 out of range (<50%)** \u2014 ranked, but with no demonstrated validity at that distance. Median identity to a measured-active enzyme is 36.7%. The catalogue states the tension rather than hiding it: ranking by similarity rewards candidates resembling what is already known, which is the opposite of discovery and the bias this project set out to correct. The low-identity candidates are the interesting ones and the ones nothing here can score. That is the state of the evidence, and the fix is measurement on new lineages rather than a cleverer model\n- [x] **Ingested the rest of the Science supplement, and left one file out on purpose.** **S2** (2,064 sequences) is byte-identical to what S3 already supplied \u2014 verified all 2,064, then skipped, because ingesting it would create nothing and risk duplicating rows whose counts this project quotes as evidence. **S1** is the 25,411-sequence homologue search space with no activity measured on any of it; 25,041 went into a separate `unlabelled_sequences` table rather than the catalogue, because nothing there has been characterised and letting it into the table whose totals are quoted would corrupt every count on the site. **S7** gives per-cluster ecology \u2014 isolation source, biome, habitat, geography and the temperature the *source organism* was isolated at, which is not the enzyme's optimum and is stored under a name that says so \u2014 kept at cluster level because that is the level it is reported at. 170 clusters stored, 354 catalogued enzymes linked. One result falls straight out: under the paper's own cluster framework **14 clusters hold both a PET-active and a measured-inactive enzyme**, against 8 under our 30% identity clustering, so the shortage of lineages measured on both sides is real under either definition and slightly less severe under theirs\n- [x] **Made the PAZy citations queryable, and found the ordinal route was not needed yet.** Schema v15 lifts the primary DOI out of the free-text notes into a column: 342 enzymes across 88 papers, 31 of which cover more than one enzyme and 285 enzymes (83%) inside such a paper. An access audit over those 31 found 16 papers holding 177 enzymes with machine-readable full text through Europe PMC. The first one opened supplied **absolute rates rather than ranks** \u2014 percent depolymerisation under one protocol \u2014 which is strictly better than the ordinal signal this item was raised to extract, so ordinal encoding is deferred to the papers that only rank. Twelve papers holding 96 enzymes are not openly reachable, and are recorded as a quantified gap rather than a silent one\n- [x] **Resolved the Cut190 strain ambiguity.** `W0TJ64`, not `C7MVE8`. Length could never separate them (both 304 aa), but the crystal structures can: **4WFI, 4WFJ, 4WFK, 5ZNO, 5ZRQ and 5ZRR all cross-reference W0TJ64**, and C7MVE8 has no PDB entry at all. Structural evidence beats a name match, and the seed was right by luck rather than by evidence until now\n- [x] **Promoted Micpa-PETase and Kutbu-PETase to named enzymes.** Both arrived through the bulk PAZy import as `PAZy:270` and `PAZy:276` under PAZy's abbreviations, and every reference view filters on `enzyme_id NOT LIKE 'PAZy:%'`, so an identifier decided whether a named enzyme with a crystal structure appeared at all. Renamed in place rather than duplicated \u2014 a second row with the same sequence would double-count the enzyme in every total on the site \u2014 cascading by hand through `reference_structures`, `reference_geometry` and `activity_measurements`, which reference it with `ON UPDATE NO ACTION`. Both verified against the Science library by exact sequence match before renaming: **Micpa-PETase** (*Micromonospora pattaloongensis*, PDB 8YTU at 1.34 \u00c5) is the **most active enzyme in that entire screen** at 1,506 \u00b5M product release, and **Kutbu-PETase** (*Kutzneria buriramensis*, PDB 8YTW at 2.65 \u00c5) melts at 88.8 \u00b0C\n- [x] **Confirm the outstanding mutation sets.** All five are resolved, and two further variants (DepoPETase, LCC-A2) were found in the same pass. HotPETase and Cut190\\*\\*SS came from crystal structures; DuraPETase, TurboPETase and Z1-PETase from verified mutation sets, all matching at offset 0. Every engineered variant now carries a sequence, and the whole seed set rebuilds from one command\n- [x] **Choose and acquire the metagenome collections, size-checked first.** 2,220,462 predicted proteins (858 MB) from landfill, marine plastisphere and compost assemblies. Only assemblies carry proteins: MGnify's largest plastisphere study has 357 samples and no protein sequences at all, being 16S amplicon\n- [x] **Build the recall stage.** One profile HMM per 30% cluster, each anchored on UniProt's own Active site annotation, MMseqs2 prefilter then hmmscan and a triad completeness filter. 128 candidates from 2.2M proteins in 24 minutes, with discard counts reported at every step\n- [x] **Detect the oxyanion hole properly.** The structure stage was already doing it structurally, but **wrongly**: \"the two backbone N atoms closest to the serine\" returns Met161 and Trp185 on IsPETase, whose published donors are Tyr87 and Met161. Now takes donor 1 by position (the nucleophile elbow) and donor 2 from the sequence-distant oxyanion loop via an N\u2013H direction test, recovering both donors from 6EQE and pinned by three regression tests. Geometry records *which* residues it identified, since the old output could not have revealed the error\n- [x] **Embed the candidate set.** ESM-2 t12-35M, frozen, CPU, mean-pooled with padding and CLS/EOS excluded. 848 vectors at 480 dimensions in under a minute\n- [x] **Filter fragments and length outliers.** From UniProt's own Fragment flag rather than a length cutoff, plus a 200 to 450 aa window derived from the experimentally evidenced positives. Marked, never deleted, so the catalogue stays complete and the exclusion stays auditable\n- [x] **Train the PET activity head.** Three labelling schemes, cluster-grouped, with the Elkan-Noto class prior estimated (c = 0.645) rather than assumed and swept across 1/3/5/10% on out-of-fold scores. At 30% identity: all-annotated **0.967 \u00b1 0.020**, measured-only **0.921 \u00b1 0.072**, PU **0.782 \u00b1 0.159**. Two things the sweep taught by being wrong first: clipping the prior-adjusted score at 1 collapses everything above the prior into a tie and the ties *move* the AUC, which made the prior look influential (0.913 to 0.977) when it is invariant by construction; and refitting on the rows being scored returned 1.000 for every prior, memorisation wearing the costume of an invariance check. AUC is now omitted from the sweep on purpose and what the prior genuinely governs is reported instead \u2014 how many sequences get *called* positive, 78% at a 1% prior down to 65% at 10%\n- [x] **Run the full evaluation protocol.** Five of the six components run; the sixth was made to run. **Cluster splits at 30% and 50%**: 0.967 and 0.975 all-annotated, so the threshold is not doing the work. **Leave-one-family-out is degenerate**, and that is the finding rather than a gap \u2014 all 13 ESTHER families are wholly positive or wholly negative (Polyesterase-lipase-cutinase 77/0, Cutinase 0/110), so holding one out removes a class and AUC is undefined: in this catalogue the label *is* family membership. **Prospective holdout** could not use `pdb_release_date`, which is empty for the entire catalogue, so UniProt first-public dates were fetched for 858 of 890 accessions spanning 1986 to 2026; AUC 0.908 at a 2020 cutoff, reported **UNDERPOWERED** because the test side holds 366 positives and 6 negatives. **Reliability** is saturated: five of eight quantile bins sit at a predicted 1.00, which is what a head that has learned family membership looks like. **The verdict that matters** is against the baselines: composition 0.736, nearest-measured-positive retrieval **0.931**. The all-annotated head clears retrieval by +0.037 and the measured-only head **does not clear it at all (-0.009)**. Trained on labels somebody actually measured, the learned head does not beat looking up the nearest known PETase \u2014 which is the same conclusion the sequence and geometry evaluations reached, arrived at a third way\n- [x] **Smoke-test ESMFold before committing to a full run.** Cleared: 8.44 GB model (it bundles ESM-2 3B), 31 min to load, 109 s to fold 290 residues. No Boltz-2 fallback needed. The prediction reproduces IsPETase's crystal active site exactly (S160/D206/H237, Ser-His 2.98 A against 2.94)\n- [x] **Extract active-site geometry, validated on crystal structures first.** The triad is found by geometry rather than sequence position, and recovers the published triads of 6EQE and 4EB0 exactly. Validation caught a brittle cleft-width metric: a hard radius feeding a max meant a 0.4 A shift in one residue halved the answer, scoring IsPETase's own prediction at 11.09 A against its crystal's 20.90 A. Now a percentile over a wider radius, and crystal-to-prediction disagreement fell from 9.81 A to 1.81 A\n- [x] **Establish whether geometry actually tracks PET activity.** First tested on AlphaFold models of 131 characterised enzymes (114 PET-active, 17 within-family negatives), pLDDT-matched so nothing is a model-confidence artefact. Raw feature differences are convincing and physically sensible (cleft depth AUC 0.819, p<0.001; PET-active enzymes have deeper, narrower clefts with a tighter oxyanion hole), but **cluster-grouped it collapses to 0.533 \u00b1 0.185**, statistically indistinguishable from the sequence head's 0.493. The signature is largely cluster structure. **Re-run on the finished structure set and it reproduces exactly**: 342 enzymes rather than 131, raw cleft depth AUC 0.808, cluster-grouped **0.534 \u00b1 0.173**. A 2.6\u00d7 larger benchmark changed the conclusion by 0.001, which is the strongest evidence yet that this is label-limited and not method-limited. The binding constraint barely moved: within-family negatives went 17 to 26, in 10 clusters\n- [x] **Found the crystal structures the reference set already had.** The builder had always ordered its sources experimental-first, and the branch had never once fired: the PAZy import records a UniProt accession and leaves `pdb_ids_json` empty, so all 312 of its enzymes fell through to a model. 12 of 320 structures were experimental because of a missing lookup, not a missing structure. Linking UniProt's cross-references and ranking them (exact sequence match, then fewest differences, then X-ray, then resolution) took it to **60 of 354**, best resolution 0.91 \u00c5. Two traps: comparing sequences anchored at position 1 reported LCC, Cut190 and twenty others as 209-274 differences when they are identical, because the deposit is the mature chain and the stored sequence the precursor \u2014 searching the offset took exact matches from 10 to 36; and UniProt cross-references homologues freely, so deposits past 8 differences are refused and 4 enzymes keep the model computed from their own sequence\n- [x] **Established that geometry is not comparable across structure source.** Adding the 55 crystal structures made the activity result *worse* \u2014 cluster-grouped AUC 0.749 on predicted structures alone against 0.553 pooled, with non-overlapping ranges across twenty random split seeds. More data lowering a score is a confound announcing itself. Among enzymes of the same activity class, geometry alone tells a crystal structure from a model at AUC 0.723; and paired **within the same protein** (51 enzymes measured both ways, so no enzyme-selection artefact is possible) the oxyanion hole differs systematically: the second donor's angle is **23.6\u00b0 in the crystal against 15.6\u00b0 in the model** (p 1.1e-06), its distance +0.78 \u00c5 (p 1.3e-07). Predictions build a tighter, more idealised oxyanion hole than the protein actually has. **Cleft depth is source-invariant** (p 0.072) and is also the strongest activity feature, which is the usable finding: a geometry model must hold the source constant or restrict itself to the features that survive. A third fact closes it off \u2014 **all 57 experimental structures are PET-active and not one of the 26 within-family negatives has a deposit**, because nobody crystallises the enzymes that do not work\n- [x] **Build the v1 tabs.** Home, Catalogue, Candidate, Compare and Methods, with 3Dmol.js over static PDB and Plotly for every chart\n- [x] **Superposed interactive structure viewer, first working version.** Verified in a real browser, not assumed: multiple structures load into one Mol\\* viewport in distinct colours and genuinely overlay, which confirms the pre-superposed-at-write-time design end to end. Two rendering defects remain, both recorded below\n- [x] **Render structures as cartoon rather than wireframe.** Root cause found: Mol\\* fell back to lines because gemmi's mmCIF carries no polymer or secondary-structure annotation, and the UMD build exports only 9 symbols. Moved to 3Dmol.js and now write HELIX/SHEET records computed with biotite's P-SEA `annotate_sse`, which 3Dmol parses to set `atom.ss` directly\n- [x] **Draw the catalytic triad.** Resolved by the same move: `molstar.MolScriptBuilder` is not exported by the UMD viewer build, so the selection expression could never be constructed. 3Dmol selects the triad residues directly\n- [x] **Designed the viewer interaction.** Three of the four open questions are settled and one deliberately is not. **The overlay set**: IsPETase and FAST-PETase shown by default, because a candidate's geometry means nothing beside one reference and one reference cannot show how far a working engineered variant has already moved; plus per-lineage small multiples sharing one camera, paginated at twelve. **The triad**: a transparent yellow surface drawn from the residues geometry measured, never from positions inferred by an alignment, with substitutions in pink and a click reporting the side-chain distance between them. **IsPETase as a ghost**: yes, grey, and every structure superposed onto it at write time so the browser aligns nothing. Sequence panels on every viewer page, click-through in both directions, monomers only, panel captions, and coordinates served under versioned URLs. **Still open: whether pLDDT colouring becomes the default.** It is the honest colouring for a prediction and the wrong one for a crystal structure, and now that 60 of 354 references are experimental the answer differs by panel rather than by page\n- [x] **Deploy to `pants.mdeller.com`.** Live on port 8005 behind gunicorn, nginx and certbot, with an entry in the mdeller.com launcher. All validation in `deploy.sh` runs *before* the first rsync, so a failed deploy cannot leave new code on disk with the service still running the old build. Two droplet-specific fixes landed with it: HTML now carries `no-cache` so heuristic caching cannot pin stale `?v=` asset URLs, and the `/static/` location overrides nginx's stock mime type for `.pdb` (`application/x-pilot`, the PalmPilot format), which had silently defeated `gzip_types` and was sending 100 MB of compressible text uncompressed\n- [x] **Measured real droplet headroom.** Not estimates: **3,915 MB total, 2,498 MB available, and no swap configured**, so an overcommit is a hard OOM kill rather than a slowdown. PANTS' own web service is **58 MB** across two workers. The box carries six applications and the large ones are AlphaFraud (1,294 MB) and chatPDB (608 MB). Disk is 19 GB free of 77 GB. The conclusion for v2: **ESMFold cannot run here at all** \u2014 8.44 GB of weights against 2.5 GB available \u2014 and nothing about that is close enough to tune. ESM-2 t12-35M at roughly 150 MB is the only model that fits, which is the same reason the web virtual environment carries no scientific libraries: structures and embeddings are computed offline and served as precomputed files\n- [x] **Tried to build on the sequence head, and stopped on evidence.** Frozen ESM-2 at 35M, 150M and 650M, mean-pooled: size-weighted leave-one-cluster-out AUC 0.562, 0.472, 0.465 against a composition baseline of 0.472. Eighteen times the parameters buys nothing. Two measurement faults found on the way and fixed. The k-fold evaluation reported `\u00b1 0.000` across ten seeds \u2014 not stability but only **2 distinct partitions**, because one cluster holds 68% of the data and four folds cannot separate it; switched to leave-one-cluster-out. And the unweighted LOCO mean of **0.625** was an artefact of averaging seven clusters of 2\u20136 enzymes, where an AUC can only land near 0 or 1, alongside the three real values of 0.443, 0.429 and 0.574; size-weighted it is **0.472**. With three evaluable folds the standard error is ~0.06, so every representation tried sits inside one interval \u2014 the differences were never measurable. A ranking objective was assessed as the escape route and rejected on data: only **3** clusters hold two enzymes whose rates differ by more than replicate noise, against 10 for the binary objective, so ranking narrows the base rather than widening it. What would change the verdict is not more enzymes but more **lineages**: roughly 10 enzymes from each of ~15 new 30% clusters under one protocol at 37 \u00b0C, about 150 assays, taking evaluable lineages from 3 to ~18 while measuring the therapeutic condition directly\n- [x] **Designed the validation panel that would lift the constraint.** `release/validation_panel.csv`: **150 assays across 15 new 30%-identity lineages**, one protocol, 37 \u00b0C on amorphous PET, taking evaluable lineages from 3 to about 18. Selected by breadth and deliberately *not* by predicted activity \u2014 using the prediction to choose what to measure would select enzymes resembling what is already known, which is the bias the project exists to correct. Members within each lineage are spread by greedy maximum-minimum distance rather than sampled from the centre, so ten near-duplicates do not measure the same enzyme ten times. First cut selected purely by cluster size and produced 140 of somebody else's sequences against 10 of this project's, testing none of the mining that built the catalogue; rebalanced to put PANTS' own novel lineages first, giving **80 PANTS candidates (68 from human gut, the therapeutically relevant environment) and 70 homologues**. Those 80 are precisely the out-of-range candidates nothing here can score, so measuring them adds lineages *and* asks whether the mining found anything real\n- [x] **Wrote the findings up, and made them un-staleable.** `FINDINGS.md` is **generated** by `scripts/build_findings_document.py`, reading every figure from the JSON artefact that produced it \u2014 the same discipline applied after a Methods paragraph quoted a superseded AUC for weeks. Structured as claim / evidence / limitation rather than as narrative, because a negative result is only worth anything if a reader can see immediately what would overturn it\n- [x] **Hardened for handover.** `NEXT.md` documents the state, the one experiment that matters, how to run everything, seven traps that have already cost time in this repo, and three things deliberately not done with the reasons. Five new tests guard the findings against drift by asserting **structure and direction, never exact values** \u2014 a test pinning `AUC == 0.398` would fail the first time anyone legitimately adds data, which trains people to delete tests. 68 tests pass\n- [ ] **MHETase pipeline (v2).** Its own seed and negatives: MHETase is Tannase family, so a PETase-seeded profile search cannot reach it\n- [x] **Published the catalogue as a citable dataset.** Zenodo, CC BY 4.0, concept DOI [10.5281/zenodo.21807353](https://doi.org/10.5281/zenodo.21807353), which always resolves to the newest version. **v0.3.0** ([10.5281/zenodo.21848998](https://doi.org/10.5281/zenodo.21848998)) carries eighteen files \u2014 the catalogue, the structures, and a JSON artefact for every analysis including the three negative results \u2014 and v0.2.0 ([10.5281/zenodo.21827812](https://doi.org/10.5281/zenodo.21827812)) carried thirteen, each size-verified on upload and every figure in its description traced back to `STATS.json` before publishing. The description also leads with the negative result rather than burying it. `DATASHEET.md` is now GENERATED by `build_release.py` rather than hand-maintained: by v0.2.0 the hand-written one claimed 1,107 reference rows against 1,140, 48 measurements against 75 and 268 structures against 1,188, and still called the release \"a snapshot taken before folding finished\" months after folding finished", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\nMany different families of cytokines exist and the focus of this study is the group utilising the glycoprotein 130 (gp130) receptor. The gp130 family of cytokines includes leukemia inhibitory factor (LIF), oncostatin M (OSM), interleukin-6 (IL-6), interleukin-11 (IL-11), ciliary neurotrophic factor (CNTF) and cardiotrophin-1 (CT-1). These are amongst the most pleiotropic cytokines known and function in controlling haematopoesis, neuronal development and the immune system (Cosman, 1993; Nicola, 1994).\n\n## 1.3 Cytokine or Growth Factor?\n\nIn common both cytokines and growth factors act on their target cells by binding cell surface receptors, which then transduce the stimulus to the inside of the cell, so eliciting a physiological response. The biological response is typically mediated through protein phosphorylation events commonly brought about by kinase activity. A cytokine is most commonly defined as a protein factor that binds to a cell surface receptor containing no intrinsic kinase activity. Conversely, the term growth factor is reserved for the subset of extracellular signaling proteins that bind to cell surface receptors containing intrinsic kinase activity. Cytokines can also be defined as a class of extracellular signaling proteins that bind structurally related receptors. These receptors are characterised by a conserved pattern of cysteine residues as well as a conserved WSXWSX motif. Cytokines generally act on target cells in contact with the secreting cell (juxtacrine) or cells in the immediate vicinity (paracrine) of the secreting cell. Several cytokines also function in direct regulation of the secreting cell (autocrine). In contrast, growth factors as exemplified by growth hormone and prolactin have a more classical long-range (endocrine) role.\n\nHowever, the terms cytokine and growth factor are often used interchangeably; reflecting their similarities and for the purpose of this study no distinction will be made (see Table 1-1 below for comparison).\n\n| Property | Growth Factor | Cytokines |\n|---|---|---|\n| Site of Production | Few | Many |\n| Cellular Target | Many | Few |\n| Biological Role | Homeostasis, Tissue Repair | Fighting Infection |\n| Biological Redundancy | Low | High |\n| Biological Pleiotropy | Low | High |\n| In the Circulation? | Yes | Rarely |\n| Sphere of Influence | Widespread | Autocrine, Paracrine |\n| Inducers | Physiological Variation | External Insults |\n| Receptor Properties | Tyrosine Kinase Activity | No Enzymatic Activity |\n\n*Table 1-1.* Table showing a comparison of cytokines and growth factor properties. Adapted from Nicola, 1994.\n\n## 1.4 Structural Classification of Cytokines and their Receptors\n\nDespite the overall lack of sequence similarity in most members of the cytokine family, they can be sub-classified according to similarities in biological action, induction mechanisms and three-dimensional structure (Table 1-2).\n\n### 1.4.1 Structural Classification of Cytokines\n\nCytokines adopt several distinct structural motifs and since the three-dimensional structure of a protein often reflects its evolutionary origins more accurately than sequence homology, this is a very useful means of classification. Cytokines adopt many diverse folds involving all alpha (\u03b1), all beta (\u03b2) or alternating mixtures of both alpha separated by beta (\u03b1/\u03b2) secondary structures (Richardson, 1981). As in other classes of proteins discrete domains of alpha and beta secondary structure (\u03b1+\u03b2) are uncommon (see Table 1-2 and Figure 1-1).\n\nThis distribution demonstrates that no particular fold is favoured for use in cytokine signaling. However, cytokines are generally relatively small, compact and adopt very stable structures; often with protective carbohydrates and intramolecular disulphide bridges. These features enhance solubility, stability and generally increase the half-life through protection from degradation and protease action. These protective features are evolutionarily beneficial considering the vigorous insults that cytokine signaling is often induced by.", "# Communication: 60+ publications translating structural data into actionable insights\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-29 on marcdeller.com (categories: 6cs).*\n\n## Communication: 60+ publications translating structural data into actionable insights\n\nWith [60+ cross-functional publications](https://marcdeller.com/publications/), I see communication as the bridge between atomic-level detail and real-world decisions in drug discovery. Solving a structure is only step one; the real value emerges when that structure is translated into clear hypotheses and strategies that non-structural biologists can act on. Across my publications, I focus on telling mechanistic stories that guide the next experiment, design cycle, or decision rather than just reporting data. Well-communicated structural insights accelerate alignment across disciplines and prevent critical information from getting lost in translation between teams.\n\n* Communication is about transforming atomic-level data into stories that drive scientific and therapeutic understanding.\n* With many peer-reviewed papers, I focus on clarity and accessibility, turning structures into discoveries others can build upon.\n* It matters because effective communication ensures that knowledge multiplies, not just accumulates.\n* In structural biology the figure does the work\u2014a single well-rendered view says what paragraphs of text cannot.\n* The real test is whether a structure changes the next experiment, the next molecule, or the next clinical decision.\n\nPersonal Philosophy: The 6C's\n\n### [**Curiosity:** 10+ years driving hands-on drug discovery through multidisciplinary approaches](https://marcdeller.com/curiosity-20-years-driving-hands-on-discovery-through-multidisciplinary-approaches/)\n\n[**Curiosity:** 10+ years driving hands-on drug discovery through multidisciplinary approaches](https://marcdeller.com/curiosity-20-years-driving-hands-on-discovery-through-multidisciplinary-approaches/)\n\n### [**Courage**: 400+ challenging structures determined, enabling breakthrough mechanistic insights](https://marcdeller.com/courage-400-structures-determined-enabling-breakthrough-mechanistic-insights/)\n\n[**Courage**: 400+ challenging structures determined, enabling breakthrough mechanistic insights](https://marcdeller.com/courage-400-structures-determined-enabling-breakthrough-mechanistic-insights/)\n\n### [**Creativity:** 7+ patents through innovative protein engineering and AI integration](https://marcdeller.com/creativity-5-patents-through-innovative-protein-engineering-and-ai-integration/)\n\n[**Creativity:** 7+ patents through innovative protein engineering and AI integration](https://marcdeller.com/creativity-5-patents-through-innovative-protein-engineering-and-ai-integration/)\n\n### [**Collaboration:** 2+ state-of-the-art Structural Biology facilities built through global partnerships](https://marcdeller.com/collaboration-2-state-of-the-art-structural-biology-facilities-built-through-global-partnerships/)\n\n[**Collaboration:** 2+ state-of-the-art Structural Biology facilities built through global partnerships](https://marcdeller.com/collaboration-2-state-of-the-art-structural-biology-facilities-built-through-global-partnerships/)\n\n### [**Compassion:** 6+ FDA IND filings supported while mentoring 20+ scientists](https://marcdeller.com/compassion-6-fda-ind-filings-supported-while-mentoring-20-scientists/)\n\n[**Compassion:** 6+ FDA IND filings supported while mentoring 20+ scientists](https://marcdeller.com/compassion-6-fda-ind-filings-supported-while-mentoring-20-scientists/)", "# Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\nAuthors: Kong, L.; Lee, J.H.; Doores, K.J.; Murin, C.D.; Julien, J.-P.; McBride, R.; Liu, Y.; Marozsan, A.; Cupo, A.; Klasse, P.-J.; Hoffenberg, S.; Caulfield, M.; King, C.R.; Hua, Y.; Le, K.M.; Khayat, R.; Deller, M.C.; Clayton, T.; Tien, H.; Feizi, T.; Sanders, R.W.; Paulson, J.C.; Moore, J.P.; Stanfield, R.L.; Burton, D.R.; Ward, A.B.; Wilson, I.A.\nSource: Nature Structural & Molecular Biology, 2013\n\nThe large glycan binding surface of PGT 135 is created by the CDR loops draping themselves across the entire length of the N-linked glycans at Asn392 (547 \u00c52) and Asn332 (365 \u00c52), where multiple van der Waals interactions and hydrogen bonds are made from the mannose sugars at the tips to the GlcNAc moieties at the stem (Fig. 1, Fig. 3 and Supplementary Table 2, 4-7). The Asn386 glycan interaction is significantly less (99 \u00c52, two van der Waals interactions with H1) than with the Asn392 and Asn332 glycans, but nonetheless critical for neutralization of various HIV-1 strains (Fig. 3d), (Supplementary Tables 3-7). Interestingly, the Asn392 and Asn332 glycans interact with opposite faces of PGT 135 CDR H3 in a bifurcated manner reminiscent of bnAb PG9 and PGT 128 binding to dual glycans on epitope-scaffolded V1-V2 loops (PG9) or gp120 outer domain (PGT 128), respectively (Fig. 3a). However, unlike these other bnAbs, whose interacting glycans are on adjacent \u03b2-strands, the two glycans interacting with PGT 135 CDR H3 are on independent \u03b2-strands separated by \u03b2-strand 19 (Fig. 1). VL also contacts the Asn332 glycan via CDR L3 on one side of VL, whereas CDR L2, and VL framework regions 2-3 on the opposite side interact with the Asn392 glycan (Fig. 1).\n\n### Glycan specificity of PGT 135\n\nN-linked glycans exhibit multiple levels of heterogeneity, arising from variable positioning of the glycosylation sequons on gp120, differential processing by Golgi enzymes, and conformational flexibility. To understand how PGT 135 achieves relatively broad neutralizing activity despite dependence on heterogeneous glycans, we employed virus neutralization and binding assays based on virus and gp120s from different strains with various N-linked glycosylation knockouts, along with glycan arrays and different expression systems (Fig. 3b-d). These results were compared with analysis of the glycans in the crystal structure of a partially deglycosylated PGT 135 complex. Crystallization required partial deglycosylation of the complex, which we carried out with the rationale that EndoH will preferentially cleave the more exposed and flexible N-linked glycans that are not protected by PGT 135 binding (Supplementary Fig. 3d).\n\nPseudoviruses prepared from mammalian 293T cells (complex, hybrid and oligomannose glycans) or 293S cells (Man5-9GlcNAc2 oligomannose glycans) had similar PGT 135 neutralization profiles (Fig. 3c). These results are consistent with the crystal structure, which shows interaction only with an oligomannose glycan patch on gp120 that would be present when gp120 is expressed in either cell line, and validate the use of 293S cells to produce the gp120 for the crystal study. In contrast to 2G12 and PGT 128, PGT 135 also neutralized pseudoviruses from cells treated with N-butyldeoxynojirimycin (NB-DNJ), which results in either missing or glucosylated (i.e. masked) D1 arms of N-linked glycans. This observation is also consistent with the crystal structure because PGT 135 does not interact with the Asn332 glycan D1 arm and only with the second mannose residue of the Asn392 glycan D1 arm, which would not be affected by NB-DNJ treatment (Fig. 3a). Also in contrast to 2G12 and PGT 128, adding kifunensine to the expression system, which limits glycan processing to Man9GlcNAc2, resulted in loss of neutralization by PGT 135. Again, the crystal structure is compatible with this outcome: modeling an additional mannose residue to the D2 arm of the Man8 glycan at Asn392 would result in a severe steric clash (Supplementary Fig. 3a). The D2 arm of the Man6 glycan at Asn332 is also sterically crowded, but may be able to accommodate a terminal mannose with some slight adjustment (Supplementary Fig. 3b). In contrast, the D3 arm of the Asn332 glycan is free from any steric clash with the antibody (Supplementary Fig. 3c).\n\nGlycan array binding assays showed that PGT 135 recognized Man7-9 glycans exclusively and bound to these with lower affinity than 2G12 and PGT 128 (Fig. 3b). Although PGT 135 appears to bury extended glycan moieties in the binding site, most of the contacts are van der Waals interactions and the antibody makes only 4 hydrogen bonds to the Asn332 glycan. In contrast, PGT 128, which generated a substantially higher glycan binding signal on glycan arrays, makes 13 hydrogen bonds with the Asn332 glycan. The pattern of binding to these glycans is dependent on the density, clustering and conformation of the presented glycans; in a different glycan array format using non-covalently immobilized N-linked glycans as neoglycolipids that have an element of mobility, only Man8 and Man7 glycans without the terminal D2 mannose were recognized (Supplementary Table 8). The array analyses and kifunensine effect described above can be reconciled if binding to glycans other than Asn392, most likely Asn332, can occur via Man8 or Man9. While a homogenous display of glycans on an array is unlikely to mimic the actual presentation of the glycans on gp120, use of multiple arrays containing different coupling chemistries and printing densities allows more opportunity to assess the glycan specificity of PGT 135.", "# Universal Mounting Robot (UMR)\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-12 on marcdeller.com (categories: project).*\n\n* [ACA Poster 2013](https://marcdeller.com/wp-content/uploads/2025/08/UMR_poster_ACA_2014_final.pdf)\n* Approaches to automated protein crystal harvesting\n* [Link](https://pmc.ncbi.nlm.nih.gov/articles/PMC3936438/)\n\n* Persistent bottleneck in structural biology: crystal harvesting remains one of the last manual tasks in an otherwise automated process.\n* Multiple engineering approaches \u2014 from capillaries to MEMS robots \u2014 target fragility, dehydration, and handling precision.\n* Semi\u2011automated systems show practical progress, especially for microcrystals and high\u2011viscosity media.\n* Advances in detectors and in\u2011situ methods could bypass traditional mounting entirely.\n* Hybrid, integrated platforms promise higher throughput, fewer failures, and better exploitation of next\u2011generation X\u2011ray sources.\n\n1. **The JCSG high-throughput structural biology pipeline:** [Elsliger, M.-A.](https://pubmed.ncbi.nlm.nih.gov/20944202/), [*Acta Crystallographica Section F.*](https://journals.iucr.org/f/)\n2. **Automated robotic harvesting of protein crystals \u2013 addressing a critical bottleneck or instrumentation overkill?:** [Viola, R.,](https://pubmed.ncbi.nlm.nih.gov/17965947/) [*Journal of Structural and Functional Genomics.*](https://link.springer.com/journal/10969)\n3. **REACH: Robotic Equipment for Automated Crystal Harvesting using a six-axis robot arm and a micro-gripper:** [Heidari Khajepour, M. Y.](https://pubmed.ncbi.nlm.nih.gov/23519413/), [*Acta Crystallographica Section D.*](https://journals.iucr.org/d/)\n4. **Acoustically mounted microcrystals yield high-resolution X-ray structures:**\u00a0[Soares, A. S.](https://pubmed.ncbi.nlm.nih.gov/21542590/), [*Biochemistry.*](https://pubs.acs.org/journal/bichaw)\n5. **Hyperquenching for protein cryocrystallography:** [Warkentin, M.](https://pubmed.ncbi.nlm.nih.gov/20461232/), [*Journal of Applied Crystallography.*](https://journals.iucr.org/j/)\n\nProtein crystallography is central to unlocking how biological molecules function, yet one step in the process has stubbornly resisted automation: harvesting the crystals for X\u2011ray diffraction. While cloning, expression, purification, crystallization, and data collection are now often robotic, moving delicate protein crystals from their growth medium to the beamline still relies heavily on expert hands \u2014 a slow, inconsistent, and failure\u2011prone bottleneck. The fragility of protein crystals, their sensitivity to osmotic, chemical, and thermal changes, and the precision required make this task uniquely challenging for machines.\n\nIn response, researchers and engineers around the world have explored an array of innovative systems to remove, or at least reduce, the need for manual handling. Concepts span microcapillaries, fine microtools, microgrippers, acoustic droplet ejection, optical tweezers, and MEMS\u2011based microrobots. Some focus on physically supporting the crystal during transfer; others use non\u2011contact forces like sound, light, or fluid flows to position them with minimal damage.\n\nFull autonomy requires fast, reliable crystal recognition, delicate motion control, and high\u2011precision feedback \u2014 capabilities that remain expensive and technically demanding. As a bridge, several semi\u2011automated systems now exist, blending operator guidance with robotic precision. Notable examples include the Universal Micromanipulation Robot (UMR), REACH microgripper platform, acoustic droplet ejection conveyor systems, and magnetic MEMS \u201cRodBots.\u201d These approaches address specific weaknesses of manual work: handling crystals under 10 \u03bcm, harvesting from viscous media, improving cryoprotection consistency, reducing dehydration, and increasing speed.\n\nEmerging technologies promise to reshape the landscape. High\u2011speed pixel array detectors open the door to continuous data collection from crystals still in their growth plates, sometimes eliminating conventional mounting altogether. In\u2011situ diffraction approaches bypass harvesting entirely, removing a physical failure point and opening opportunities for fully automated pipelines. MEMS microgrippers, optical manipulation, and oil\u2011assisted hyperquenching further refine cryocooling \u2014 often with less cryoprotectant.\n\nWhere deployed, automation delivers repeatability, higher throughput, and the ability to accommodate microcrystals that are beyond human dexterity. These gains matter: structural genomics centers process hundreds of thousands of crystals each year, with perhaps 100 crystals screened for every successful structure. Eliminating late\u2011stage loss, increasing microcrystal yield, and seamlessly linking growth to data collection could notably accelerate structural biology output.\n\nThe forward path likely lies not in a single \u201cperfect\u201d device but in hybrid systems drawing on complementary technologies \u2014 uniting precision sensing, gentle manipulation, and integration with next\u2011generation beamlines \u2014 to finally close a stubborn gap in the protein crystallography pipeline.", "# GitHub README: TopPDBLX\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/TopPDBLX, 2026\n\n`./run.sh model.propose_boundaries --uniprot P00698` smooths the profile, takes the longest\ncontiguous run, merges runs separated by short gaps, and refuses below 50 residues rather than\nreturning a construct nobody would clone.\n\n**Two structural rules were tried and both lost to the model.** Neither is in the shipped\nproposer, and both are kept in the code with their numbers attached, as results rather than gaps.\n\n| Feature tried | Idea | Result |\n|---|---|---|\n| Chou-Fasman coil nudge | Do not cut mid-helix | **Worse**: median error 5 residues to 8 |\n| ESMFold pLDDT gate | Trim where a single-sequence fold is unconfident | **No change**: median 2 either way; pLDDT alone much worse (median 6, 71% within 25) |\n| Pfam domain-edge snapping | Cut where the curated domain ends | **Signal real, application fails.** True boundaries sit within 15 residues of a Pfam edge 54% of the time against a 19% null, but snapping never improves the median and nudges p75 from 56 to 58 |\n| metapredict disorder, blended | Trim predicted-disordered tails | **Not redundant but no help.** Correlates with pLDDT only \u22120.39, and predicts \"inside\" better than pLDDT does (AUROC 0.793 against 0.719), yet blending leaves the median at 6 and worsens p90 |\n| Disorder as an input channel, retrained | Fuse the feature rather than blending post hoc | **No movement**: like-for-like head-only pilot, p75 111 to 108, p90 unchanged at 362 |\n| Surface entropy, hydrophobicity | Trim high-entropy EKQR patches | **No signal at all**: AUROC 0.474 and 0.491, both indistinguishable from chance |\n\n**Six attempts, no improvement, and the pattern is consistent.** Two of the features carry\ngenuine independent signal (Pfam edges at 2.8x enrichment over a null, disorder outscoring pLDDT\non the very task) and still fail to help, whether blended afterwards or fused as an input channel\nand retrained.\n\nThe reading: **the classifier is trained on the decision itself, and every one of these features is\na proxy for part of that decision.** A construct boundary frequently sits in perfectly ordered\nsequence, at a domain junction or a convenient cloning site, and a 33.5M-parameter protein language\nmodel has evidently already absorbed what these proxies encode. Meanwhile validation MCC was still\nclimbing when round 01 stopped (0.607, 0.681, 0.700 by epoch). **Effort belongs in the model, not\nin features around it.**\n\n### What it is not\n\n- **It is not a disorder predictor.** Residues that were cloned but never appeared in the density\n  count as *inside* the construct. The model predicts what a crystallographer chose to clone, not\n  what turned out to be ordered.\n- **It only knows successes.** Every label comes from a construct that produced a crystal. It has\n  never seen a construct that failed, so it cannot tell you a boundary is bad, only that it is\n  unlike the ones that worked.\n- **It errs towards keeping residues.** It predicts \"inside\" for 69% of residues where the truth\n  is 61.5%, so proposed spans run slightly long. For construct design that is the safer direction,\n  but trim rather than extend if you are choosing between its suggestion and your own.\n\n## \ud83c\udfb2 Will it crystallise at all?\n\n**The one question the PDB cannot answer.** Every condition in this dataset produced a crystal, so\nthe archive supports *what worked* and is silent on *what failed*. The only substantial record of\ncrystallisation failures is [TargetTrack](https://zenodo.org/records/821654), the archived\nstructural genomics target log, final release 1 July 2017.\n\nParsed to 335,771 targets with sequences and full status histories, of which:\n\n| | Targets |\n|---|---|\n| Reached crystallisation | 21,173 |\n| Reached purified protein and never crystallised | 79,926 |\n\n**Conditioning the negatives on reaching purified protein is the load-bearing decision.** Every\ntarget in the training set got far enough that crystallisation was genuinely attempted, so the\nmodel answers *given soluble purified protein, will it crystallise*. Using all 335,771 targets\ninstead drops the positive rate from 20.9% to 6.3%, and a model trained that way mostly predicts\nwhether a gene expresses, which is a different and much easier question.\n\nGradient boosting on 30 sequence descriptors, split at 30% identity by MMseqs2 cluster:\n\n| Measure | Value |\n|---|---|\n| AUC | **0.656** (pre-declared floor 0.65) |\n| Average precision | 0.360 against a 0.218 base rate |\n| **Top 5% of ranked constructs crystallise** | **53%**, a 2.4x lift |\n| Top 10% | 43%, 2.0x |\n\n**Read the ranking, not the AUC.** 0.656 barely clears the threshold set before the run and would\nnot trouble a published predictor. What it does usefully is order a shortlist: pick the top twenty\nconstructs from a hundred candidates and half of them crystallise, against a fifth if you pick at\nrandom. Sequence length, net charge and cysteine content dominate; disorder predictions add\nnothing (AUC 0.654 to 0.656).\n\n**TargetTrack ends in 2017** and is dominated by structural genomics centres that chose tractable\ntargets. The number is honest for that population and is not a universal probability.\n\n## \ud83d\udeab What the dataset will not tell you\n\nTwo independent attempts to predict the *condition* from the *sequence* both lose to a baseline\nthat ignores the sequence entirely.\n\n| Source | hit@1 | hit@5 | hit@10 |\n|---|---|---|---|\n| **Frequency prior** (ignores the query) | **0.646** | **0.846** | **0.922** |\n| ESM-2 embeddings + learned head | 0.571 | 0.833 | 0.909 |\n\n97,471 training and 21,316 test records at the 30% identity split, 41 condition classes. The\nlearned model is worse on every measure. Its hit@5 margin is **-0.013** against the **+0.03**\ndeclared before the experiment began, so the kill criterion fired and the result is published\nuntuned.", "# The 79th Organ: When Art and Biology Meet Bioengineering in the Fight Against Microplastics\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe therapeutic applications extend beyond speculation. Mycelium-based biomaterials demonstrate exceptional biocompatibility when tested with primary human fibroblasts, showing excellent cell viability and morphology comparable to control samples. These natural scaffolds require only simple sterilization through autoclaving, eliminating the need for toxic solvents or complex fabrication processes typical of synthetic materials. The intrinsic properties of fungal networks\u2014including natural porosity, biodegradability, and bioactive compounds like \u03b2-glucans\u2014position them as superior alternatives to conventional medical devices.\n\nCompanies like Elora Therapeutics are pioneering this intersection of environmental remediation and human health. Founded in 2025 and embedded within Austin Technology Incubator, Elora is developing the first systemically delivered enzyme therapeutic specifically designed to degrade and remove microplastics from the human body. Their approach targets synthetic polymers that accumulate in blood and tissues\u2014compounds our biology cannot naturally eliminate\u2014addressing three critical indications: chronic inflammation, metabolic disease, and hormone disruption.\n\nThe convergence of synthetic biology and bioengineering is accelerating organ-scale tissue manufacturing capabilities. Recent advances in computational vascular design enable the creation of perfusable networks spanning entire organs, with synthetic biology tools allowing precise cellular programming. These developments suggest that complex bioartificial organs incorporating living fungal components could become feasible within decades, potentially revolutionizing how we address environmental toxin exposure.\n\n| Key Concept | Description | Key References |\n| --- | --- | --- |\n| The 79th Organ | Speculative bioengineered organ using fungi to filter microplastics from bloodstream through enzymatic degradation | [Dierkx Design Study](https://www.designboom.com/design/fungi-made-prosthetic-79th-organ-extracts-breaks-down-microplastics-human-bodies-10-03-2025/) |\n| Microplastic Bioaccumulation | Synthetic polymer particles accumulate in human organs causing inflammation, endocrine disruption, and metabolic dysfunction | [Stanford Medical Research](https://med.stanford.edu/news/insights/2025/01/microplastics-in-body-polluted-tiny-plastic-fragments.html) |\n| Fungal Enzyme Systems | Naturally occurring enzymes like cutinase, laccase, and PETase capable of degrading synthetic polymers through bioremediation | [Nature Polymer Degradation](https://pmc.ncbi.nlm.nih.gov/articles/PMC12113260/) |\n| Mycelium Scaffolds | Three-dimensional fibrous networks that mimic human extracellular matrix providing biocompatible tissue engineering platforms | [Advanced Mycelium Materials](https://www.nature.com/articles/s41598-021-91572-x) |\n| Bioartificial Organs | Combination of biomaterials and living cells creating functional organ replacements with enhanced capabilities | [Bioartificial Manufacturing](https://pmc.ncbi.nlm.nih.gov/articles/PMC6322143/) |\n| Elora Therapeutics | First company developing systemic enzyme therapy for microplastic removal targeting endocrine and metabolic disorders | [Company Innovation Profile](https://www.linkedin.com/company/elora-therapeutics) |", "# Cell surface receptors\nAuthors: Deller, M.C.; Jones, E.Y.\nSource: Current Opinion in Structural Biology, 2000\n\n**Authors:** Marc C. Deller and E. Yvonne Jones. **Affiliations (condensed):** CRC Receptor Structure Group, Wellcome Trust Centre for Human Genetics, Roosevelt Drive, Headington, Oxford OX3 7BN, UK (both authors; e-mail Marc.Deller@strubi.ox.ac.uk); Oxford Centre for Molecular Sciences, New Chemistry Building, South Parks Road, Oxford OX1 3QT, UK (Jones; e-mail Yvonne.Jones@strubi.ox.ac.uk). **Published:** Current Opinion in Structural Biology, volume 10, pages 213\u2013219, 2000 (section: Macromolecular assemblages).\n\n**Abbreviations:** DR5, death receptor 5; Epo, erythropoietin; EpoR, Epo receptor; FGF, fibroblast growth factor; FGFR, FGF receptor; ICAM-1, intercellular adhesion molecule-1; Ig, immunoglobulin; IgSF, Ig superfamily; IL-1, interleukin-1; IL-1R, IL-1 receptor; LG, laminin G-like; NGF, nerve growth factor; NK, natural killer; SRCR, scavenger receptor cysteine-rich; TfR, transferrin receptor; TNF, tumour necrosis factor; TNFR, TNF receptor; TRAIL, TNF-related apoptosis-inducing ligand; VEGF, vascular endothelial growth factor.\n\n## Abstract\n\nDuring the past year, the database of ligand\u2013receptor complexes has essentially doubled. These new results have immeasurably extended and expanded our view of cell surface receptor structure and function. The flood of data has revealed new modes for receptor cross-linking, demonstrating the potential stringency of the extracellular requirements for the initiation of intracellular signalling and highlighting unexpected interactions suggestive of higher order clustering.\n\n## Introduction\n\nThe cells within a multicellular organism must communicate with each other to coordinate their functions. Direct cell\u2013cell communication is achieved through the interactions of specialised receptor molecules protruding from the cell surface. Similarly, interactions at a distance employ molecules secreted by one cell to deliver a signal to a target cell by binding to the appropriate receptor molecules on its surface. In parallel with these physiological interactions, viruses exploit specific interactions with cell surface receptors to infect the organism.\n\nMany cell surface receptors are present in relatively low quantities; the harnessing of molecular biology and expression techniques for the production of milligram quantities of soluble forms of such molecules has been essential for structural studies. This area is now coming-of-age, with a substantial number of new structures unveiled during the course of the past year. Such information should allow us to address fundamental questions about the biological processes that underpin human health. This review will be confined to discussing the extracellular portions of receptors and their interactions. Particular emphasis will be placed on the recognition complexes, rather than on the isolated receptor structures.\n\n## Structural insights into new families of receptors\n\nStructural templates are now available for many of the domains that typically comprise the extracellular regions of cell surface receptors [1]. The database of domain architectures has been further added to in the past year with three new motifs. A representative crystal structure of a scavenger receptor cysteine-rich (SRCR) domain (a curved \u03b2 sheet cradling an \u03b1 helix) [2] provided the first structural information on the defining features of the SRCR protein superfamily. A structure/function study of the N-terminal domain (a jellyroll \u03b2 sandwich) of the EphB2 receptor [3] provided the first data for the ligand-binding domain of the Eph subfamily of receptor tyrosine kinases in the nervous system. Lastly, structures of laminin G-like (LG) modules (also termed LNS domains) from the extracellular matrix glycoprotein laminin [4] and the neuronal cell surface protein neurexin [5] have highlighted putative ligand-binding surfaces. The LG module structure is also a jellyroll \u03b2 sandwich and, at least in the case of laminin, suggests a calcium-dependent lectin-like mode of binding to the cell surface receptor \u03b1-dystroglycan.\n\n## Cytokine receptors and the diversity of cytokine binding modes\n\nCytokines mediate intercellular signalling through high-affinity interactions with the extracellular regions of specific cell surface receptors. This promotes receptor oligomerisation, which, in turn, triggers intracellular signalling cascades within the target cell. Research carried out over the past year has illuminated the structures and functions of a number of cytokines and their receptors. One noteworthy pointer for the future is the use of cryo-electron microscopy to visualise the quaternary structure of the entire insulin\u2013insulin receptor structure [6]. Such studies, when combined with higher resolution structures of soluble components (e.g. the crystal structure of the N-terminal three domains of the type-1 insulin-like growth factor receptor [7]), promise a route to a unified picture of entire membrane-spanning signalling complexes.", "# Engineering Protein Stability: Unlocking the Secrets of Successful Crystallization\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nMedia supplementation strategies have also evolved significantly. The addition of osmolytes, chemical chaperones, and specialized growth factors can enhance both cell viability and protein yields. For mammalian expression systems, temperature reduction post-transfection and the use of histone deacetylase inhibitors can boost expression levels substantially.\n\n**Expression Systems: Choosing the Right Platform**\n\nThe selection of appropriate expression systems has become increasingly sophisticated, with researchers choosing platforms based on specific protein requirements rather than convenience. Bacterial systems remain popular for their simplicity and cost-effectiveness, but mammalian systems are often essential for proteins requiring complex post-translational modifications or proper disulfide bond formation.\n\nCell-free expression systems represent an emerging alternative that offers unique advantages, including the ability to produce toxic proteins and incorporate non-natural amino acids. These systems bypass many of the limitations associated with living cells while maintaining the protein folding machinery necessary for producing functional proteins.\n\nFor membrane proteins, the choice of expression system becomes even more critical. Specialized systems that can provide appropriate lipid environments and membrane insertion machinery are often required for successful expression of functional membrane proteins suitable for crystallization studies.\n\n**Engineering Success Stories: Three Breakthrough Examples**\n\nThe power of protein engineering for crystallization is best illustrated through specific success stories where traditional approaches had failed:\n\n**1. Surface Entropy Reduction (SER) \u2013 RhoGDI Case Study**\nThe development of surface entropy reduction by Derewenda and colleagues represents a landmark achievement in rational protein crystallization. Using the human regulatory protein RhoGDI as a model system, they systematically replaced surface-exposed lysine and glutamate residues with alanine to reduce conformational entropy. This approach proved remarkably successful, with the vast majority of mutations resulting in enhanced crystallization potential. Most importantly, the crystal contacts were directly mediated by the mutated regions, validating the theoretical basis of the method. The success of SER has been replicated across numerous protein targets, with over 100 structures now deposited in the Protein Data Bank using this approach.\n\n**2. VHH Nanobody Chaperones \u2013 MazE Structure Determination**\nThe crystallization of the bacterial addiction antidote MazE represents a triumph of chaperone-assisted crystallography. MazE had proven impossible to crystallize due to its extreme conformational flexibility, with more than half the protein existing in a disordered state. The breakthrough came through the use of a specific camel VHH nanobody fragment that bound to the protein and provided the conformational stability necessary for crystal formation. Even with the nanobody chaperone, only 45% of MazE was ordered in the final crystal structure, highlighting the intrinsic flexibility that had made this protein so challenging to study. This work demonstrated that nanobodies could serve as powerful crystallization chaperones for highly dynamic proteins.\n\n**3. T4 Lysozyme Fusion \u2013 GPCR Structural Revolution**\nThe fusion of T4 lysozyme to G-protein coupled receptors has revolutionized membrane protein structural biology. The \u03b22-adrenergic receptor was the first GPCR to benefit from this approach, where T4 lysozyme replaced the flexible third intracellular loop or was fused to the N-terminus. The success stemmed from T4 lysozyme\u2019s exceptional crystallization properties and its ability to provide additional hydrophilic surface area for crystal contacts while masking flexible regions of the GPCR. Subsequent modifications, including disulfide-stabilized variants (dsT4L) and minimal T4L (mT4L), have further expanded the utility of this approach. The T4 lysozyme fusion strategy has now been applied to over 14 different GPCRs, making it the most successful general approach for membrane protein crystallization.\n\nThese examples demonstrate that protein engineering for crystallization has evolved from empirical trial-and-error approaches to rational, theory-driven strategies. The combination of computational design tools, advanced expression systems, and innovative engineering approaches continues to expand the range of proteins amenable to structural studies. As our understanding of protein stability and crystallization mechanisms deepens, we can expect even more sophisticated engineering strategies that will make the most challenging proteins accessible to structural analysis.", "# Unleashing the Power of AnythingLLM: Running Large Language Models Locally\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-20 on marcdeller.com (categories: ai).*\n\n* AnythingLLM enables running advanced AI language models securely and privately on your own machine for custom document analysis and RAG applications.\n* It supports plugging in a wide range of open source and commercial LLMs, with flexible options for custom embeddings and search.\n* The built-in LanceDB vector database scales to millions of document chunks, enabling real-time search and analysis\u2014all data stays local.\n* Sophisticated text splitting and chunking features ensure accurate, context-preserving document vectorization for even the largest datasets.\n* Installing locally delivers unmatched privacy, speed, flexibility, and long-term cost control vs. typical cloud-based LLM offerings.\n\n1. **AnythingLLM Official Documentation \u2014 Overview:**\n   <https://docs.anythingllm.com/introduction>\n2. **LLM and Embedding Model Configuration in AnythingLLM:**\n   <https://docs.useanything.com/setup/llm-configuration/local/localai>\n3. **A Novice-Friendly Guide to Running Local AI With AnythingLLM and Ollama:**\n   <https://www.feinberg.northwestern.edu/sites/artificial-intelligence/health-data-science/ai-essentials/local-llm-guide.html>\n4. **Vector Database Capabilities in AnythingLLM:**\n   <https://docs.anythingllm.com/setup/vector-database-configuration/overview>\n5. **GitHub: AnythingLLM Features and Custom Integrations:**\n   <https://github.com/Mintplex-Labs/anything-llm>\n\n**Why Run AnythingLLM Locally?**\n\nRunning advanced AI models on your own machine was once science fiction. With the emergence of AnythingLLM, local deployment of large language models (LLMs) for custom document analysis, vector search, and more is possible\u2014no cloud dependency required. Here\u2019s what sets this powerful framework apart and why running it on your hardware gives you privacy, speed, and flexibility that cloud-based systems struggle to match.\n\n**AnythingLLM** is an all-in-one AI application designed for both technical and non-technical users. Whether you need Retrieval-Augmented Generation (RAG), AI agents, or private, large-scale document analysis, AnythingLLM offers a \u201cno code\u201d interface that just works. By running locally, your data always stays private, inference is fast, and you\u2019re free from expensive cloud subscriptions and usage caps.\n\nKey advantages of local deployment:\n\n* **Privacy:** Sensitive files or client data never leaves your device; full compliance with privacy requirements is ensured.\n* **Speed:** No network lag\u2014responses come directly from your hardware, greatly reducing latency for analysis and chat tasks.\n* **Customization:** Swap in your choice of LLMs (Ollama, Llama, Gemma, etc.) and custom embedders for tailored workflows and domain-specific intelligence.\n* **Cost control:** One-time investment in hardware can be more economical over time, without surprise cloud bills or data egress charges.\n* **Offline access:** Analyze and interact with documents even without an internet connection.\n\n**Custom Models and Vector Database Power**\n\nAnythingLLM shines in environments where customization is key. Using the Ollama backend or Docker, it lets you configure and run your preferred LLMs\u2014from lightweight models for everyday summarization to 70B+ parameter giants for the toughest NLP tasks. You can even bring your own embedding models for fine-tuned search and relevance.\n\nAt the heart of its document analysis is the integrated **vector database**, powered by LanceDB. This enables:\n\n* Instant search and RAG over large document sets, even at the scale of millions of chunks\n* Total control over your vectors (they never leave your device)\n* Support for multiple vector database backends if you want more advanced or enterprise features\n\n**Scalable Text Splitting and Chunking**\n\nAnythingLLM\u2019s text splitter and chunking tools are optimized for accuracy and efficiency. They break down huge documents into logically-linked \u201cchunks\u201d with appropriate overlap, ensuring no information is lost across boundaries. This chunking is language- and context-aware, making it suitable for diverse file types or even code repositories.\n\nThe result: efficient, vectorized storage and lightning-fast retrieval\u2014perfect for legal, biomedical, or research domains with massive, complex datasets.\n\n**Local vs. Online: The Edge of Self-Hosting**\n\nRunning AnythingLLM locally, as opposed to relying on cloud offerings, brings distinctive benefits:\n\n* **Security & compliance:** Absolute control over all inputs, outputs, and metadata\u2014critical for regulated organizations.\n* **Performance:** No external dependencies; you control scaling and resource allocation.\n* **Flexibility:** Choose and switch models, embedders, and databases at will, incorporating the latest open-source innovations without waiting for cloud releases.\n* **Cost predictability:** Avoid unnecessary operational costs, especially if you need extensive, ongoing analysis.\n\nOnline/cloud models are easier to start with, but may fall short for applications where privacy, extensive customization, or cost management are paramount.", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\nIn another experiment, the CD34+ cells described above were added at 50,000 cells/well to a 12-well plate containing 2.0 mL culture media with specified concentrations of clone 74 or isotype control. The treatment period was 12 hours. After 12 hours, cells were collected, washed and lysed using lysing buffer (Cell Signaling Technologies) supplemented with 1x Halt Protease and Phosphatase Inhibitor Cocktail (Thermo Fisher Scientific). Protein samples (6 \u03bcg) were separated in pre-casted 4-12% Tris-glycine gels (Thermo Fisher Scientific) and transferred to nitrocellulose membranes using iBlot 2 Dry Blotting System and iBlot 2 Transfer Stacks (Thermo Fisher Scientific). The nitrocellulose membranes were blocked with StartingBlock (Thermo Fisher Scientific) for one hour and probed with antibodies to detect pSTAT5 (Cell Signaling), STAT5 (Cell Signaling), pSTAT3 (Cell Signaling), STAT3 (Cell Signaling), and \u03b2-actin (Cell Signaling). Detection was performed using horseradish peroxidase (HRP)-conjugated secondary rabbit antibody (Cell Signaling) and chemiluminescence HRP substrate (Thermo Scientific). Clone 74 selectively inhibited pSTAT3 and pSTAT5 in CD34+ cells harboring mutCALR in a dose-dependent manner, while the isotype control (IgG) at 10 \u03bcg/mL had no impact on CD34+ cells (FIG. 19C).\n\nIn another experiment, Clone 74 and ruxolitinib were combined to determine the impact of co-treatment on megakaryopoiesis produced by CD34+ cells from human cord blood (WT cells) or an MPN patient with mutCALR (mutCALR cells). After six days in culture, 25 nM ruxolitinib alone did not impact the frequency of megakaryocytes produced by either WT or mutCALR-expressing CD34+ cells. In contrast, 25 \u03bcg/mL of clone 74 alone selectively reduced the production of megakaryocytes from mutCALR+ CD34+ cells but not from WT CD34+ cells. Further, the combination of 25 nM ruxolitinib and 25 \u03bcg/mL of clone 74 led to further depletion of pathogenic megakaryocytes produced by mutCALR+ CD34+ cells (FIG. 19D).\n\nTaken together, these results demonstrate the ability of the anti-mutCALR antibodies to inhibit mutCALR-derived oncogenic functions in MPN patient cells.\n\n### Example 15: Effect of Anti-mutCALR Antibody in the Cell Cycle\n\nEngineered BaF3 cells (10,000 cells/well) were added to a 12-well plate containing 2.0 mL culture media with serially diluted antibodies. The treatment period was for 24 hours. After 22 hours of incubation, Ba/F3 cells were pulse-labeled with BrdU for 2 hours. After 24 hours, Ba/F3 cells were collected, washed with PBS and incubated with BD Cytofix/Cytoperm buffer for 20 minutes on ice. Cells were then washed with the BD Perm/Wash buffer and the cell pellets were resuspended in 100 \u03bcL of BD Cytoperm and Permeabilization buffer and incubated on ice for 10 minutes. Cells were then washed and resuspended in 100 \u03bcl of BD Cytofix/Cytoperm buffer and incubated at room temperature for 5 minutes. Cells were washed and treated with 100 \u03bcL of 300 \u03bcg/mL solution of DNase, and incubated for 1 hour at room temperature, and were then washed and resuspended in 50 \u03bcl of BD Perm/Wash buffer containing anti-BrdU-APC antibody for 20 minutes in the dark at room temperature. After additional washing, cells were resuspended in PBS containing 2% FBS and 7-AAD and analyzed for cell cycle profiles on the LSRFortessa X-20 analyzer (BD Biosciences). Clone 74 was screened to determine its effect on the cell cycle using WT BaF3 cells or Ba/F3-TPOR/mutCALR type 1. Clone 74 selectively induced apoptosis in BaF3 cells carrying mutCALR in a dose-dependent manner, whereas isotype control (IgG) had no effect on these cells. In contrast, clone 74 did not impact cell cycle profiles of WT BaF3 cells (FIG. 20).\n\n### Example 16: Activity of Anti-mutCALR Antibody\n\nCell binding of various anti-mutCALR antibodies as compared to clone 4 was assessed using engineered BaF3 cells stably expressing human MPL and human mutCALR Type 1. The antibodies used in this experiment were obtained commercially (AB1: CAL2 antibody, Dianova) or synthesized from sequences found in the literature (AB2: anti-mutCALR antibody including VH and VL regions of antibody B3 from WO2020175689 fused to human constant regions; AB3: anti-mutCALR antibody including the BJ095 VH region and the BJ097 VL region from WO2019178362 fused to human constant regions; and AB4: anti-mutCALR antibody including VH and VL regions of antibody 8B2-H6 from WO2016087514 fused to human constant regions). The heavy and light chain sequences of AB2, AB3 and AB4 are given as SEQ ID NOs:341-346. (Note: there was an EI\u2192IE transposition when synthesizing the AB3 VL sequence at the end of the variable region; otherwise, the VL sequence is identical to the BJ097 VL region.)\n\nApproximately 200,000 cells/well were added to 96-well plates and stained with the indicated concentration of antibodies for 30 minutes on ice. Cells were then washed and stained with goat anti-human secondary conjugated to R-phycoerythrin (R-PE) (Jackson Immuno Research Laboratories) for 30 minutes on ice. The cells were then washed again and analyzed by flow cytometry. Geometric Mean Fluorescence Intensity (GMFI) of cell binding was graphed using GraphPad Prism Software (version 7.04). As shown in FIG. 21A, the anti-mutant CALR antibodies show differential levels of binding to cells expressing mutant CALR.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\n*Figure 2-1.* Flowchart detailing methods and techniques employed in the course of chapter 2.\n\n## 2.2 Methods of Engineering\n\n### 2.2.1 Rational Design of Truncated Oncostatin M\n\nOSM was previously cloned from U-937 cells induced with phorbal myristate acetate (PMA) (Staunton et al., 1998; Zarling et al., 1986). All OSM sequences referred to correspond to that previously published (GenBank accession number M27286; Malik et al., 1989). A template for the OSM molecule as a glutathione S-transferase (GST) fusion protein was kindly provided in the pGEX vector (Pharmacia) (Keith Hudson and John K. Heath, CRC Growth Factor Group, Birmingham) (see Figure 2-2).\n\nThe pGEX-3C-OSM-196 vector consisted of the predominant form of full length mature OSM (OSM-196; 196 residues). A GST moiety (247 residues) was included for use as a purification tag (Smith and Johnson, 1988). An intervening rhinovirus type 14 protease 3C site (protease 3C) was included for cleavage of the GST fusion protein. This protease site consisted of the 11 residue sequence GSEVLFQ\u2193GPGS, where \u2193 represents the position of cleavage and underline represents core residues required for efficient cleavage (Leong et al., 1992; Walker et al., 1994). The pTac promoter was utilised for high levels of expression and the Lac gene included for inducible expression (see Figure 2-2).\n\n*Figure 2-2.* Diagrammatic representation of (a) pGEX-3C-OSM-196 vector used for expression of soluble (b) OSM-3C-GST fusions. Multiple cloning site (MCS) is shown along with unique restriction sites in red lettering. Translation product (b) shows predicted helices (yellow cylinders) and disulphide bridges (red).\n\nPreviously, it has been shown that OSM-196 is prone to aggregation, low solubility and dimer formation (Staunton et al., 1998). Crystallisation attempts with this protein proved unsuccessful, due to low solubility and excessive precipitation. In order to produce material suitable for crystallographic studies mutant OSM proteins were engineered. These mutants were based on homology models designed through structural comparison with LIF (Robinson et al., 1994).\n\nStructural alignments of murine LIF and human OSM-196 were carried out using the program THREADER (Miller et al., 1996). Default values were used to produce an alignment resulting in the lowest pairwise energy sum between aligned residues. The resulting alignment was used for modelling as implemented by the package MODELLER (Sanchez and Sali, 1997). Default values were used and a model was generated from murine LIF coordinates (PDB code 1LKI) (Robinson et al., 1994) and the OSM sequence using molecular dynamics for satisfaction of spatial restraints, so relieving any regions of poor stereochemistry.\n\n### 2.2.2 Design of Oligonucleotides to Produce Truncation Mutants\n\nOligonucleotide primers were designed to enable amplification of the N-terminal 184 residues (\u0394185_OSM) or 187 residues (\u0394188_OSM) of OSM-196 using the polymerase chain reaction (PCR) (see Table 2-1). The two C-terminal truncation primers contain a cleavage site for the restriction enzyme EcoRI (shown underlined), so allowing cleavage of the resulting PCR amplified DNA fragment. The two C-terminal truncation primers also contain the triplet sequence ATC, which codes for the stop codon UAG (Amber), immediately following the C-terminal residue. These primers all bind directly to the pGEX-3C-OSM-196 plasmid sequence.\n\n| Primer Description | Oligonucleotide Sequence | Encoded residues (read from the 3' end) |\n|---|---|---|\n| 5' pGEX | 5'-GCA-TGG-CCT-TTG-CAG-GG-3' | \u2014 |\n| 3' \u0394185_OSM | 5'-GGA-ATT-CTA-GAA-GAC-CCG-CCC-CAC-TGA-GTG-3' | Stop Phe Val Arg Gly Val Ser His |\n| 3' \u0394188_OSM | 5'-GGA-ATT-CTA-CCA-CTT-GCT-GAA-GAC-CCG-CCC-3' | Stop Trp Lys Ser Phe Val Arg Gly |\n\n*Table 2-1.* Table showing primers used for amplification of \u0394185_OSM and \u0394188_OSM. EcoRI restriction enzyme site is denoted by underlined bases (GAATTC) and corresponding amino acids are written underneath each codon. Stop codon UAG (Amber) is denoted by 'Stop'.\n\n### 2.2.3 Preparation of Oligonucleotides\n\nOligonucleotides were prepared by John Fox (AltaBiosciences, Birmingham), using a 308B Applied Biosystems DNA synthesiser. Primers were prepared on a 0.2 \u03bcM scale and purified on Sephadex G25. Primers were dissolved in sterile water to 1 mg/ml and stored at \u201320 \u00b0C. For use in PCR all primers were further diluted to 200 ng/\u03bcl in water. Measurements were based on an absorbance reading of 1 at 260 nm corresponding to an oligonucleotide concentration of 20 \u03bcg/ml.\n\n### 2.2.4 Preparation of Plasmid DNA\n\nE. coli, strain JM109, containing pGEX-3C-OSM-196 were streaked onto LB-Amp plates (Luria-Bertani-ampicillin) (12 g agar, 8 g tryptone, 4 g yeast extract, 8 g sodium chloride and 50 mg of ampicillin upto 1000 cm\u00b3) (see Appendix I). Plates were grown for 16 hours at 37 \u00b0C. A single colony was used to inoculate a 5 ml starter culture of LB-Amp media (10 g tryptone, 5 g yeast, 10 g sodium chloride and 50 mg of ampicillin upto 1000 cm\u00b3) (see Appendix I). The culture was then grown for 8 hours at 37 \u00b0C with shaking. This culture was diluted to 500 ml in LB-Amp media and grown for a further 16 hours at 37 \u00b0C with vigorous shaking. The cells were harvested by centrifugation at 6000 x g for 15 minutes at 4 \u00b0C and plasmid extracted using a Maxi Prep Kit (Qiagen). The purified pGEX-3C-OSM-196 was used as a template for PCR (see 2.2.5) or was digested with BamHI and EcoRI and purified (see 2.2.6) for use in ligation with insert (see 2.2.7).\n\n### 2.2.5 Amplification of DNA by PCR", "# Transforming Antibody Design: How Artificial Intelligence Revolutionizes Structure Prediction\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Key Concept | Description | Key Reference |\n| --- | --- | --- |\n| **Pre-trained Language Models** | Neural networks trained on hundreds of millions of antibody sequences extract evolutionary patterns and structural features, enabling rapid structure prediction directly from amino acid sequences with sub-25-second inference times. | [Ruffolo et al., Nature Communications, 2023](https://www.nature.com/articles/s41467-023-38063-x) |\n| **Attention Mechanisms** | Transformer-based architectures like AlphaFold2\u2019s Evoformer module use row-wise and column-wise attention to capture long-range dependencies and co-evolutionary signals between amino acids, improving complex structure predictions. | [Emerging methods, PNAS, 2023](https://www.pnas.org/doi/10.1073/pnas.2303499120) |\n| **Graph Neural Networks** | Multi-relational graph architectures with geometric message passing capture atomistic-level features at antibody-antigen interfaces, enabling accurate binding affinity predictions and identification of affinity-enhancing mutations. | [Shan et al., Nature Communications, 2024](https://www.nature.com/articles/s41467-024-51563-8) |\n| **Transfer Learning** | Fine-tuning general protein models on antibody-specific datasets substantially improves CDR structure prediction accuracy, particularly for variable regions like CDR-H3 that exhibit high conformational diversity. | [Chinery et al., Bioinformatics Advances, 2024](https://academic.oup.com/bioinformaticsadvances/article/doi/10.1093/bioadv/vbae202/8090019) |\n| **Validation Metrics** | Comprehensive assessment combines structural accuracy measures (RMSD, TM-score, lDDT), interface quality metrics (interface RMSD, fraction of native contacts), and classification metrics (accuracy, precision, recall, F1 score, ROC-AUC) to evaluate model performance. | [Guest et al., NIH, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC10751731/) |\n| **Inverse Folding Models** | Antibody-specific inverse folding tools like AntiFold and ProteinMPNN design sequences that fold into target structures by learning structure-to-sequence mappings, with sequence recovery rates exceeding 50 percent for CDR regions. | [Shanehsazzadeh et al., PLOS ONE, 2025](https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0324566) |", "# BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nOption                        Default              Description\n----------------------------  -------------------  --------------------------------------------------\n--output-dir                  settings.output_dir  Override where generated YAMLs are written\n--out-dir                     ./boltz_output       Boltz's own --out_dir, next to the md file\n--accelerator                 auto                 auto / gpu / cpu\n--limit N                     none                 Cap how many pending targets 'run' submits\n                                                   (smoke test before a full batch)\n--strict                      off                  Promote preflight WARN to FAIL\n--skip-interactions           off                  Skip cif2plip interaction analysis during\n                                                   analyze, even if setup-plip has been run\n\nMEMORY-CONTROL OPTIONS  (unified memory / Apple Silicon)\n=======================================================\n\nOption                        Default              Description\n----------------------------  -------------------  --------------------------------------------------\n--workers                     2                    Matches Boltz's own default\n--mps-watermark               1.0                  PYTORCH_MPS_HIGH_WATERMARK_RATIO cap\n--max-parallel-samples        1                    Boltz --max_parallel_samples\n--recycling-steps             Boltz default        Passthrough\n--sampling-steps              Boltz default        Passthrough\n--diffusion-samples-affinity  Boltz default        Passthrough\n--sampling-steps-affinity     Boltz default        Passthrough\n--max-msa-seqs                Boltz default        Passthrough\n--memory-warn-tokens          1000                 Preflight size-heuristic WARN threshold\n\nCOMPARE-SSE OPTIONS\n===================\n\nOption                        Default              Description\n----------------------------  -------------------  --------------------------------------------------\n--out-dir                     spec directory       Where boltz_sse_comparison.csv / .html /\n                                                   _sessions/ are written\n--phi-psi-threshold           see --help           Flag per-residue backbone dihedral outliers\n                                                   above this angle (degrees)\n\n```\n\nPython 3.12 is required because `boltz` pins `numpy<2.0`, which has no prebuilt wheel for newer versions. The script installs `boltz`, `rich`, `pandas`, `openpyxl`, `pyyaml`, `rdkit`, `matplotlib`, `psutil`, `scipy`, `gemmi`, `biopython`, `plotly`, and `reportlab`. It also pulls PyTorch (~2-3 GB), and the first time`boltz predict` actually runs, Boltz downloads several GB of model weights over the network.\n\nThe dependency list is deliberately short:\n\n```\n\n\t\t\t\t\t# Core BoltzMaker dependencies (Python 3.12; boltz pins numpy&lt;2.0)\nboltz          # Boltz-2 structure + affinity model\nrich           # live progress interface\npandas         # tabular summaries\nopenpyxl       # XLSX workbook output\npyyaml         # generated Boltz input files\nrdkit          # ligand chemistry + validation, 2D depiction, scaffold matching\nmatplotlib     # static figures\npsutil         # live memory monitoring\nscipy          # similarity ordering; motif geometry for compare-sse\ngemmi          # structure parsing; apo-to-holo superposition for compare-sse\nbiopython      # sequence alignment (pocket-contact remapping, motif numbering)\nplotly         # interactive dashboard charts; compare-sse bar chart + heatmap\nreportlab      # printable ligand grid (PDF)\n\n# Network services used at runtime by compare-sse (no install, no API key)\n# GPCRdb  -&gt; generic numbering for GPCR motifs (TM1-7, H8, ECL/ICL)\n# KLIFS   -&gt; 85-residue kinase pocket alignment (hinge, gatekeeper, HRD, DFG, aC-Glu)\n# PDBe    -&gt; SIFTS residue mapping to Pfam domain boundaries (universal fallback)\n\n```\n\n###### **Each package earns its place**\n\n`boltz` is the model itself. `rich` draws the progress interface. `pandas` and `openpyxl` produce the CSV and XLSX outputs. `pyyaml` writes the generated inputs. [RDKit](https://www.rdkit.org/) handles every piece of ligand chemistry. `matplotlib` and [Plotly](https://plotly.com/) render the figures and the interactive dashboard charts. `psutil` watches memory during a run, which matters more than you might expect on unified-memory Macs. `scipy` orders ligand series by similarity, `gemmi` and `biopython` parse structures and run the sequence alignment behind pocket-contact suggestions, and `reportlab` writes the printable ligand grid. The whole thing is a single script, `BoltzMaker.py`, whose subcommands each transparently relaunch under the managed environment. Interaction profiling with PLIP and cif2plip is optional and lives in a separate conda-forge environment, because PLIP needs OpenBabel and PyMOL as in-process imports that do not install cleanly through pip.\n\n###### **The input file**\n\nThe single input spec, `boltz_input.md`, is plain labeled text: `Label: value` lines with a blank line between blocks, comments starting with `#`, and field names written in plain English (`Output folder`, `Predict affinity`, `Pocket contact`) rather than Boltz-internal YAML format, which is finicky and difficult to parse. Constraint sentences for covalent bonds, distance restraints, and pocket contacts sit on their own lines and name the protein they belong to.\n\nHere is an example of a 5-HT2A/2B/2C serotonin receptor screening campaign, agonist vs. antagonist, with vs. without the Gq heterotrimer, that ships as a worked example. This runs structure and affinity prediction for three receptors and their Gq alpha partner with an agonist and an antagonist pair for each:\n\n```", "# Curiosity: 10+ years driving hands-on drug discovery through multidisciplinary approaches\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-29 on marcdeller.com (categories: 6cs).*\n\n## Curiosity: 10+ years driving hands-on Drug discovery through multidisciplinary approaches\n\nCuriosity has been the through-line of my 10+ year career, shaping how I choose problems, build teams, and design experiments. I have always believed that the best science starts with a simple question that refuses to let go, even when the tools or data don\u2019t yet exist to answer it. That mindset led me to build a multidisciplinary skill set that spans innovation in [precision oncology](https://marcdeller.com/my-work-driving-innovation-in-precision-oncology-through-structure-based-drug-design/), structural biology, molecular pharmacology, bioinformatics, and AI-driven discovery. Curiosity-driven, multidisciplinary work allows me to see patterns others miss and open up new therapeutic opportunities that a more linear mindset might overlook.\n\n* Curiosity fuels my research philosophy, curiosity-led questions drive new experimental and computational frontiers.\n* From protein structures to precision oncology, vaccine design, antibodies, and the engineering of microplastic-degrading enzymes, curiosity has guided bold, hands-on exploration.\n* Curiosity fuels my research philosophy, curiosity-led questions drive new experimental and computational frontiers.\n* It matters because relentless curiosity transforms uncertainty into discovery and moves science beyond boundaries.\n* Curiosity now drives Elora Therapeutics, where I co-founded a pre-seed effort to develop PET-degrading enzymes for microplastic remediation.\n\nPersonal Philosophy: The 6C's\n\n### [**Courage**: 400+ challenging structures determined, enabling breakthrough mechanistic insights](https://marcdeller.com/courage-400-structures-determined-enabling-breakthrough-mechanistic-insights/)\n\n[**Courage**: 400+ challenging structures determined, enabling breakthrough mechanistic insights](https://marcdeller.com/courage-400-structures-determined-enabling-breakthrough-mechanistic-insights/)\n\n### [**Creativity:** 7+ patents through innovative protein engineering and AI integration](https://marcdeller.com/creativity-5-patents-through-innovative-protein-engineering-and-ai-integration/)\n\n[**Creativity:** 7+ patents through innovative protein engineering and AI integration](https://marcdeller.com/creativity-5-patents-through-innovative-protein-engineering-and-ai-integration/)\n\n### [**Communication:** 60+ publications translating structural data into actionable insights](https://marcdeller.com/communication-60-publications-translating-structural-data-into-actionable-insights/)\n\n[**Communication:** 60+ publications translating structural data into actionable insights](https://marcdeller.com/communication-60-publications-translating-structural-data-into-actionable-insights/)\n\n### [**Collaboration:** 2+ state-of-the-art Structural Biology facilities built through global partnerships](https://marcdeller.com/collaboration-2-state-of-the-art-structural-biology-facilities-built-through-global-partnerships/)\n\n[**Collaboration:** 2+ state-of-the-art Structural Biology facilities built through global partnerships](https://marcdeller.com/collaboration-2-state-of-the-art-structural-biology-facilities-built-through-global-partnerships/)\n\n### [**Compassion:** 6+ FDA IND filings supported while mentoring 20+ scientists](https://marcdeller.com/compassion-6-fda-ind-filings-supported-while-mentoring-20-scientists/)\n\n[**Compassion:** 6+ FDA IND filings supported while mentoring 20+ scientists](https://marcdeller.com/compassion-6-fda-ind-filings-supported-while-mentoring-20-scientists/)", "# Clearing the Microscopic Threat: Can Plastic-Eating Enzymes Treat Liver Disease such as MASLD?\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nMouse models have proven essential for understanding mechanisms linking microplastics to MASLD. The [STZ+HFD mouse model](https://www.nature.com/articles/s41467-024-50660-y) recapitulates the full disease spectrum from steatosis to hepatocellular carcinoma within the context of metabolic dysfunction resembling human type 2 diabetes and obesity. When researchers exposed mice to [polystyrene nanoplastics](https://pmc.ncbi.nlm.nih.gov/articles/PMC12257899/) while feeding them a high-fat diet, they observed exacerbated lipid accumulation, hepatocyte apoptosis, and accelerated fibrosis compared to diet alone. The nanoplastics accumulated specifically in hepatocytes and triggered Hedgehog signaling pathways that promote stellate cell activation. Mice exposed to [submicroplastics](https://www.sciencedirect.com/science/article/abs/pii/S0300483X25000939) on a Western diet developed worsened steatosis, ballooning, inflammation, and fibrosis through mechanisms involving disturbed lipid metabolism and oxidative stress.\n\nThe therapeutic potential of PETase extends beyond simple particle removal. Discovery of [PET hydrolases in human gut microbiomes](https://pubmed.ncbi.nlm.nih.gov/39551294/) suggests that enhancing natural degradation capacity could represent a viable strategy. Researchers identified novel PET-degrading enzymes within intestinal bacteria, indicating that the gastrointestinal system already possesses some capacity for plastic breakdown. Engineering these enzymes for enhanced stability and activity could accelerate clearance of ingested microplastics before they translocate to the liver. Combining enzymatic therapy with metabolic interventions like GLP-1 receptor agonists might offer synergistic benefits, as these medications already show efficacy in MASLD treatment while potentially reducing inflammatory responses to residual particles. The challenge lies in delivery systems that maintain enzyme activity, avoid immune clearance, and target appropriate tissue compartments. Encapsulation strategies using modified erythrocytes or nanoparticle carriers could address these limitations, extending circulation time while reducing immunogenicity. While substantial development work remains before clinical trials, the convergence of human epidemiological data, mechanistic mouse studies, and advances in enzyme engineering creates a compelling foundation for pursuing PETase-based therapies as part of comprehensive MASLD management strategies.\n\n| Key Concepts | Description | Key References |\n| --- | --- | --- |\n| **Human studies** | Research in 23 MASLD patients showed that elevated fecal microplastic concentrations correlated with inflammatory immune cells, altered hepatic gene expression, and persistent liver disease following bariatric surgery. Patients with higher microplastic burden demonstrated reduced metabolic recovery. | [Schwenger et al., Sci Total Environ, 2024](https://pubmed.ncbi.nlm.nih.gov/39260480/) [Lopez-Lorenzo, X., et al., Faraday Discuss, 2024](https://pubs.rsc.org/en/content/articlelanding/2024/fd/d4fd00014e) |\n| **Mouse studies** | The STZ+HFD mouse model recapitulates the full MASLD spectrum with metabolic dysfunction. Nanoplastic exposure in these models exacerbated steatosis, hepatocyte apoptosis, and fibrosis through Hedgehog signaling activation and hepatic stellate cell proliferation, demonstrating accelerated disease progression. | [Park et al., Nat Commun, 2024](https://www.nature.com/articles/s41467-024-50660-y) [Lee et al., Liver Int, 2025](https://pmc.ncbi.nlm.nih.gov/articles/PMC12257899/) |", "# GitHub README: PANTS\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/PANTS, 2026\n\n> **Find PET-degrading enzymes that work at 37 \u00b0C in serum, not at 70 \u00b0C in a reactor.**\n\n       \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/PANTS\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/PANTS</a></td>\n</tr>\n</table>\n\n---\n\nPANTS (PETase ANnotation and Triage System) discovers, triages and engineers PETase-like and MHETase-like enzymes for **therapeutic** use: degrading PET microplastics under physiological conditions rather than in an industrial reactor. It mines metagenomic sequence space, ranks candidates on axes that matter at 37 \u00b0C, and will serve the result as an interactive catalogue with predicted structures and in-silico engineering.\n\n**Why it matters:** essentially the entire published PETase field optimises for industrial conditions (above PET's glass transition at roughly 70 \u00b0C, often alkaline, with thermostability and enzyme cost as the dominant concerns). Therapeutic use inverts almost all of it: peak activity at 37 \u00b0C, pH 7.4, in serum, against highly crystalline aged microplastic, with immunogenicity and protease resistance suddenly central. A naive homology search seeded on characterised PETases ranks candidates *towards* the industrial optimum, because that is where the well-annotated, heavily-published enzymes sit. Correcting for that bias is most of what PANTS is for. It is useful for: anyone triaging polyesterase sequence space on physiological rather than industrial criteria, and anyone who wants a calibrated probability rather than an E-value rank before committing a wet-lab screening budget.\n\n## \ud83e\udded Why this is not a homology search\n\nDetection is solved. Discrimination is not.\n\nA profile HMM built from characterised PET hydrolases returns thousands of \u03b1/\u03b2-hydrolase fold members sharing the Ser-His-Asp triad and the oxyanion hole. Almost none have meaningful activity on crystalline PET, and sequence identity to IsPETase is a weak predictor of PET activity, so E-value rank is close to uninformative about the property of interest.\n\nThe architecture that follows: **retrieval is the recall stage, the learned model is the precision stage.** MMseqs2 and HMMER cast the net exhaustively and interpretably; the model ranks within it on axes retrieval is blind to. That also yields a free baseline: if the model cannot beat E-value rank on held-out characterised enzymes, that is a cheap and publishable negative result.\n\n## \u2697\ufe0f The therapeutic constraint\n\n| Axis | Industrial optimum | Therapeutic requirement |\n|---|---|---|\n| Temperature | 65 to 75 \u00b0C | Peak activity at 37 \u00b0C |\n| pH | 8 to 9 | 7.2 to 7.4 |\n| Substrate | Amorphous, pre-treated, high surface area | Highly crystalline aged microplastic and nanoplastic |\n| Medium | Buffer | Serum, plasma proteins, lipids, physiological ionic strength |\n| Stability concern | Thermal denaturation over days | Serum protease resistance, aggregation, clearance |\n| Immunogenicity | Irrelevant | Central |\n| Product handling | Recovered and recycled | TPA and EG must be tolerable at achievable local concentrations |\n\n## \ud83d\udd00 How it works", "# Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3\nAuthors: Shvartsbart, A.; Roach, J.J.; Witten, M.R.; Koblish, H.; Harris, J.J.; Covington, M.; Hess, R.; Lin, L.; Frascella, M.; Truong, L.; Leffet, L.; Conlen, P.; Beshad, E.; Klabe, R.; Katiyar, K.; Kaldon, L.; Young-Sciame, R.; He, X.; Petusky, S.; Chen, K.-J.; Horsey, A.; Lei, H.-T.; Epling, L.B.; Deller, M.C.; Vechorkin, O.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2022\n\n**Authors:** Jeremy J. Roach (Proteovant Therapeutics, King of Prussia, Pennsylvania); Michael R. Witten, Jennifer J. Harris, Maryanne Covington, Rodrigo Hess, Michelle Frascella, Lisa Truong, Lynn Leffet, Patricia Conlen, Ron Klabe, Kamna Katiyar, Laura Kaldon, Ruth Young-Sciame, Xin He, Susan Petusky, Kwang-Jong Chen, April Horsey, Hsiang-Ting Lei, Leslie B. Epling, Marc C. Deller and Oleg Vechorkin (Incyte Corporation, Wilmington, Delaware); Holly Koblish (Ikena Oncology, Boston, Massachusetts); Luping Lin (WuXi AppTec, Philadelphia, Pennsylvania); Elham Beshad (Proteovant Therapeutics, King of Prussia, Pennsylvania); Wenqing Yao (Synnovation Therapeutics, Wilmington, Delaware).\n\n**Author Contributions:** The manuscript was prepared with contributions from all authors. All authors have given approval to the final version of the manuscript. A.S., J.J.R., and M.R.W. designed and synthesized the compounds.\n\n**Funding:** Funding was provided by Incyte Research Institute, who employed all authors at the time this research was conducted.\n\n**Notes:** The authors declare no competing financial interest.\n\n## Acknowledgments\n\nThe authors thank Scott Leonard, James Hall, and Laurine Galya for NMR analysis; Ravi Jalluri and Onur Atasoylu for molecular modeling; Jim Doughty for purity analysis; Min Li for solubility measurements; Leslie Hall, Karen Gallagher, and Hong Chang for their expert technical assistance; and Eddy W. Yue and Matthew McCammant for proofreading the manuscript.\n\n## Abbreviations\n\nFGFR, fibroblast growth factor receptor; SAR, structure\u2212activity relationships; PK, pharmacokinetics; PD, pharmacodynamics; CYP, cytochrome P450; ADME, absorption, distribution, metabolism, excretion; HBF, hepatic blood flow; MOE, molecular operating environment; AUC, area under the curve; WT, wild-type; PDGFRB, platelet-derived growth factor receptor \u03b2; FLT3, FMS-like tyrosine kinase 3; TRKA, tropomyosin receptor kinase A; FASSIF, fasted-state simulated intestinal fluid; SGF, simulated gastric fluid.\n\n## References\n\n*(reference list omitted)*", "# Outsmarting Evolution: Using AI to Overcome Drug Resistance\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-14 on marcdeller.com (categories: blog).*\n\n* AI is being used to predict which mutations in a drug target are most likely to cause drug resistance before they emerge in patients.\n* This predictive insight allows for the proactive design of \u201cresistance-proof\u201d drugs that target stable regions of a protein.\n* This forward-thinking strategy can lead to the development of more durable therapies for cancer and infectious diseases, overcoming the challenge of evolutionary escape.\n\n1. **The Disruptive Impact of Structural Biology on Biopharmaceutical Innovation:**\n   <https://www.pharmasalmanac.com/articles/the-disruptive-impact-of-structural-biology-on-biopharmaceutical-innovation>\n2. **Turbocharging Protein Engineering with AI:**\n   <https://cns.utexas.edu/news/features/turbocharging-protein-engineering-ai>\n3. **AI in Drug Discovery: Predictions:**\n   <https://lizard.bio/knowledge-hub/2025-ai-in-drug-discovery-predictions>\n\nDrug resistance is one of the greatest challenges in modern medicine, particularly in oncology and infectious disease. We develop a powerful new drug, only to have the cancer cell or bacterium evolve, mutating the target protein so the drug no longer binds. This evolutionary arms race forces us to constantly develop next-generation drugs to overcome these resistance mutations, a costly and time-consuming process that we often lose.\n\nArtificial intelligence offers a powerful new strategy to get ahead in this race. Instead of waiting for resistance to emerge in the clinic, we can now use AI to predict it before it happens. By analyzing the structure of a drug target and the known patterns of viral or cancer evolution, machine learning models can predict which mutations are most likely to arise and cause drug resistance. This gives us a \u201cmap\u201d of the target\u2019s potential evolutionary escape routes.\n\nArmed with this predictive power, we can take a proactive approach to drug design. Instead of designing a drug that only fits the current form of the protein, we can use computational methods to design \u201cresistance-proof\u201d therapies that bind to a region of the protein that is essential for its function and cannot be easily mutated. We can also design molecules that are effective against both the original protein and its most likely future mutant forms. This AI-driven, forward-thinking approach allows us to outsmart evolution, creating more durable and long-lasting therapies that can win the fight against drug resistance.", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nIn some embodiments, cancers that are treatable using the compounds of the present disclosure include, but are not limited to, cholangiocarcinoma, bile duct cancer, triple negative breast cancer, rhabdomyosarcoma, small cell lung cancer, leiomyosarcoma, hepatocellular carcinoma, Ewing's sarcoma, brain cancer, brain tumor, astrocytoma, neuroblastoma, neurofibroma, basal cell carcinoma, chondrosarcoma, epithelioid sarcoma, eye cancer, Fallopian tube cancer, gastrointestinal cancer, gastrointestinal stromal tumors, hairy cell leukemia, intestinal cancer, islet cell cancer, oral cancer, mouth cancer, throat cancer, laryngeal cancer, lip cancer, mesothelioma, neck cancer, nasal cavity cancer, ocular cancer, ocular melanoma, pelvic cancer, rectal cancer, renal cell carcinoma, salivary gland cancer, sinus cancer, spinal cancer, tongue cancer, tubular carcinoma, urethral cancer, and ureteral cancer.\n\nIn some embodiments, the compounds of the present disclosure can be used to treat sickle cell disease and sickle cell anemia.\n\nIn some embodiments, diseases and indications that are treatable using the compounds of the present disclosure include, hematological cancers, sarcomas, lung cancers, gastrointestinal cancers, genitourinary tract cancers, liver cancers, bone cancers, nervous system cancers, gynecological cancers, and skin cancers.\n\nExamples of hematological cancers that are treatable using the compounds of the present disclosure include lymphomas and leukemias, such as acute lymphoblastic leukemia (ALL), acute myelogenous leukemia (AML), acute promyelocytic leukemia (APL), chronic lymphocytic leukemia (CLL), chronic myelogenous leukemia (CML), diffuse large B-cell lymphoma (DLBCL), mantle cell lymphoma, Non-Hodgkin lymphoma (including relapsed or refractory NHL and recurrent follicular), Hodgkin lymphoma, myeloproliferative diseases (e.g., primary myelofibrosis (PMF), polycythemia vera (PV), and essential thrombocytosis (ET)), myelodysplasia syndrome (MDS), T-cell acute lymphoblastic lymphoma (T-ALL) and multiple myeloma (MM).\n\nExamples of sarcomas that are treatable using the compounds of the present disclosure include chondrosarcoma, Ewing's sarcoma, osteosarcoma, rhabdomyosarcoma, angiosarcoma, fibrosarcoma, liposarcoma, myxoma, rhabdomyoma, rhabdosarcoma, fibroma, lipoma, hamartoma, and teratoma.\n\nExamples of lung cancers that are treatable using the compounds of the present disclosure include non-small cell lung cancer (NSCLC), small cell lung cancer, bronchogenic carcinoma (squamous cell, undifferentiated small cell, undifferentiated large cell, adenocarcinoma), alveolar (bronchiolar) carcinoma, bronchial adenoma, chondromatous hamartoma, and mesothelioma.\n\nExamples of gastrointestinal cancers that are treatable using the compounds of the present disclosure include cancers of the esophagus (squamous cell carcinoma, adenocarcinoma, leiomyosarcoma, lymphoma), stomach (carcinoma, lymphoma, leiomyosarcoma), pancreas (ductal adenocarcinoma, insulinoma, glucagonoma, gastrinoma, carcinoid tumors, vipoma), small bowel (adenocarcinoma, lymphoma, carcinoid tumors, Kaposi's sarcoma, leiomyoma, hemangioma, lipoma, neurofibroma, fibroma), large bowel (adenocarcinoma, tubular adenoma, villous adenoma, hamartoma, leiomyoma), and colorectal cancer.\n\nExamples of genitourinary tract cancers that are treatable using the compounds of the present disclosure include cancers of the kidney (adenocarcinoma, Wilm's tumor [nephroblastoma]), bladder and urethra (squamous cell carcinoma, transitional cell carcinoma, adenocarcinoma), prostate (adenocarcinoma, sarcoma), and testis (seminoma, teratoma, embryonal carcinoma, teratocarcinoma, choriocarcinoma, sarcoma, interstitial cell carcinoma, fibroma, fibroadenoma, adenomatoid tumors, lipoma).\n\nExamples of liver cancers that are treatable using the compounds of the present disclosure include hepatoma (hepatocellular carcinoma), cholangiocarcinoma, hepatoblastoma, angiosarcoma, hepatocellular adenoma, and hemangioma.\n\nExamples of bone cancers that are treatable using the compounds of the present disclosure include osteogenic sarcoma (osteosarcoma), fibrosarcoma, malignant fibrous histiocytoma, chondrosarcoma, Ewing's sarcoma, malignant lymphoma (reticulum cell sarcoma), multiple myeloma, malignant giant cell tumor chordoma, osteochronfroma (osteocartilaginous exostoses), benign chondroma, chondroblastoma, chondromyxofibroma, osteoid osteoma, and giant cell tumors.\n\nExamples of nervous system cancers that are treatable using the compounds of the present disclosure include cancers of the skull (osteoma, hemangioma, granuloma, xanthoma, osteitis deformans), meninges (meningioma, meningiosarcoma, gliomatosis), brain (astrocytoma, medulloblastoma, glioma, ependymoma, germinoma (pinealoma), glioblastoma, glioblastoma multiform, oligodendroglioma, schwannoma, retinoblastoma, congenital tumors), and spinal cord (neurofibroma, meningioma, glioma, sarcoma), as well as neuroblastoma and Lhermitte-Duclos disease.\n\nExamples of gynecological cancers that are treatable using the compounds of the present disclosure include cancers of the uterus (endometrial carcinoma), cervix (cervical carcinoma, pre-tumor cervical dysplasia), ovaries (ovarian carcinoma (serous cystadenocarcinoma, mucinous cystadenocarcinoma, unclassified carcinoma), granulosa-thecal cell tumors, Sertoli-Leydig cell tumors, dysgerminoma, malignant teratoma), vulva (squamous cell carcinoma, intraepithelial carcinoma, adenocarcinoma, fibrosarcoma, melanoma), vagina (clear cell carcinoma, squamous cell carcinoma, botryoid sarcoma (embryonal rhabdomyosarcoma)), and fallopian tubes (carcinoma).", "# The Digital Twin in the Clinical Trial: How AI Is Making Drug Testing Smarter and Faster\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* AI-powered \u201cdigital twins\u201d are being developed to predict disease progression for individual patients in clinical trials, creating a virtual placebo group.\n* This technology allows for smaller, faster, and more efficient clinical trials by providing a personalized baseline for each participant.\n* Using digital twins can reduce the number of patients needed in placebo arms, making trials more ethical and accelerating the approval of new drugs.\n\n1. **How AI will reshape pharma in 2025:**<https://www.drugtargetreview.com/article/154981/how-ai-will-reshape-pharma-by-2025/>\n2. **The Impact of AI on Drug Discovery and Design:**<https://www.coherentsolutions.com/insights/artificial-intelligence-in-pharmaceuticals-and-biotechnology-current-trends-and-innovations>\n3. **AI in Drug Discovery: 2025 Outlook:**<https://www.foley.com/insights/publications/2024/12/ai-drug-discovery-2025-outlook/>\n\nThe clinical trial is the most expensive and time-consuming stage of drug development, often taking years and costing billions of dollars. A major reason for this is the need to recruit large numbers of patients to generate statistically significant data, which is especially challenging for rare diseases. Furthermore, the use of placebo groups, while essential for scientific rigor, means that many participants in urgent need of treatment receive an inactive substance. This system, though the gold standard, is fraught with ethical dilemmas and logistical hurdles that create a major bottleneck in bringing new medicines to patients.\n\nA groundbreaking approach using artificial intelligence is poised to fundamentally reshape this landscape. Companies are now developing \u201cdigital twin generators\u201d\u2014AI-driven models that create a virtual counterpart for each patient in a clinical trial. By analyzing a patient\u2019s baseline characteristics, these models can accurately predict how that individual\u2019s disease would likely progress over time if they were in the placebo group. This provides a personalized, data-driven baseline for every participant, allowing for more powerful and efficient trial designs.\n\nThe implications are revolutionary. With reliable digital twins, clinical trials can be run with smaller, more targeted patient groups, significantly reducing both the cost and duration of the study. It also minimizes the number of patients required for a placebo arm, allowing more participants to receive the experimental therapy. This AI-driven innovation is not just an incremental improvement; it is a paradigm shift that makes clinical trials faster, smarter, and more ethical. By increasing the efficiency and predictive power of clinical research, we can accelerate the entire drug development process and deliver hope to patients sooner.", "# Crystal Structure of a Cellulose-Active Polysaccharide Monooxygenase from M. thermophila (MtPMO3*)\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [5UFV](https://www.rcsb.org/structure/5UFV)\n* **Release Date:** 2017-03-15\n* **Method:** X-RAY DIFFRACTION 2.45 \u00c5\n* **Organisms:** Thermothelomyces thermophilus ATCC 42464\n* **Macromolecules:** Glycoside hydrolase family 61 protein (protein)\n* **Unique Ligands:** CU\n* **Publication:** [Span, E.A., et al. (2017). The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase. ACS Chemical Biology, 12, 1095-1103.](https://pubs.acs.org/doi/10.1021/acschembio.7b00016)\n* **Blog Post:** [The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase](https://marcdeller.com/the-role-of-the-secondary-coordination-sphere-in-a-fungal-polysaccharide-monooxygenase/)\n\n**Description:** This structure of the cellulose-active fungal polysaccharide monooxygenase MtPMO3* reveals the critical hydrogen-bonding networks in the secondary copper coordination sphere that govern oxygen activation and proton transfer. The structural insights are key to understanding and optimizing enzymatic cellulose degradation for biofuel production and sustainable biomass utilization.", "# Cell surface receptors\nAuthors: Deller, M.C.; Jones, E.Y.\nSource: Current Opinion in Structural Biology, 2000\n\n## Structural immunology: the control of cellular immune responses\n\nNatural killer (NK) cells and CD8+ T lymphocytes (T cells) have complementary roles in the cellular immune response. The control of activation and, ultimately, lysis is critically dependent on direct, receptor-mediated recognition events between T cells or NK cells and MHC class I molecules on the target cell. Recent structural studies on these molecules are too extensive to do justice to within this review; however, up-to-date reviews are available covering MHC class I type molecules [32] and T-cell receptor based recognition complexes [33].\n\nNK cell receptors may be grouped into two structural families. This past year has seen major progress in structural studies of NK cell receptors with C-type lectin folds; first, with a representative structure of this type of NK cell receptor [34] and, very recently, with a complex between a receptor and an MHC class I molecule [35]. The complex structure poses an interesting puzzle as contacts between the molecules in the crystal suggest not one, but two potential interaction modes. Progress on the second class of NK cell receptor, those composed of Ig-like domains, remains frustrating in terms of complexes, although two further structures of family members have been determined [36,37].\n\nThis year has also seen the determination of three IgSF receptor structures that bridge the humoral and cellular immune responses. Two provide the first structural insights into Fc gamma-receptors [38,39], whereas the third, that of the high-affinity IgE receptor [40], provides a basis for the design of novel therapies to combat allergic responses. Each prompts speculation on the nature of the interaction complexes that should fuel further functional analyses.\n\n## Receptor interactions regulating iron metabolism\n\nRecent advances in our understanding of the mechanism by which cells acquire iron-loaded transferrin exemplify the emergent synergy among genetics-based studies of disease linkages, classical functional studies and structural biology. The structure of the extracellular region of the transferrin receptor (TfR) [41] was swiftly followed by the structure of its complex with HFE [42]. HFE, an MHC class I type cell surface molecule, was first marked out as a player in the transferrin system through its link to the genetic disease of iron homeostasis, hereditary haemochromatosis. The complex reveals an architecture consistent with the association of receptors from the same cell surface and is suggestive of a conformational trigger, whereby TfR communicates to cytoplasmic proteins that HFE is bound.\n\n## Cell adhesion molecules and IgSF-mediated cell\u2013cell interactions\n\nThe cadherins represent an important family of calcium-dependent, homophilic cell adhesion molecules. As such, lattice contacts within their crystal structures have provided persuasive models for functionally relevant modes of association [43,44]. The story to date has, however, retained several areas of controversy. The report this year of a combined electron microscopy, crystallography and mutagenesis based analysis [45] thus serves an important purpose in further dissecting the complex web of cis and trans interactions involving this molecule.\n\nMany cell adhesion molecules are members of the IgSF and, in common with the cadherins, several mediate homophilic interactions. The neural cell adhesion molecule (NCAM) is one such example and NMR studies this past year have provided further structural data on isolated domains from this molecule, as well as interaction data hinting at its binding sites [46]. The first structural analyses on a cell adhesion molecule were performed on CD2 in the early 1990s. The original crystallographic studies on CD2 were particularly noteworthy in revealing a 'head-to-head' mode of molecular packing within the crystal lattice that has remained a paradigm for functional intercellular adhesive interactions within this subgroup of the IgSF [47]. The physiological receptor for CD2 in humans is CD58. The production of a form of this molecule amenable to crystallisation has proved extremely problematic; however, this year both crystallographic and NMR analyses of the CD2-binding domain of CD58 were at last accomplished [48,49]. These coordinates, combined with those for CD2, allowed a detailed assessment of the putative head-to-head CD2\u2013CD58 binding mode (Figure 3). Gratifyingly, the key features of the predicted interface were subsequently borne out in a direct structure determination of the CD2\u2013CD58 complex [50]. The functional adhesive interaction mediated by this heterophilic pair is very weak, but specific. These new structural data point to the importance of electrostatic complementarity, with shape matching at the protein\u2013protein interface being notably inexact (Figure 3).\n\n*Figure 3.* The nature of a low-affinity, but specific adhesion interaction. (a) The presumptive complex [48] between the two IgSF cell adhesion molecules CD2 and CD58 is shown schematically in a view consistent with the positioning of cell surfaces at the top and bottom of the panel. The complex was based on the isolated structures of CD2 and a CD58\u2013CD2 chimera interacting in a mode first identified from lattice contacts in CD2 crystals [47]. This geometry is directly confirmed by the crystal structure of the CD2\u2013CD58 adhesion complex [50]. (b) A close-up view of the interaction surface of CD2\u2013CD58, with solvent-accessible molecular surfaces of central sections of the interacting \u03b2 sheets depicted semi-transparently (adapted from [48]).", "# An Enzyme That Eats Plastic \u2014 Inside Your Blood\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-04-18 on marcdeller.com (categories: blog, microplastics, sb).*\n\nData digitized from original research article for display, please refer to original article below for actual data\n\n* [**Degradation of PET microplastic particles to monomers in human serum by PETase**](https://pubs.rsc.org/en/content/articlehtml/2024/fd/d4fd00014e)\n\n* **Complete monomer release:** PETase fully degrades PET microplastics into TPA and MHET, as confirmed by HPLC and NMR.\n* **Active in serum:** The S238A PETase variant releases ~500 \u00b5M of monomers within 15 minutes in real human serum at 37 \u00b0C.\n* **No cell toxicity:** HEK293 cells remain fully viable across all tested concentrations and densities of PETase.\n* **A medical proof of concept:** This is the first demonstration of enzymatic microplastic depolymerization in a serum-like biological environment.\n\n1. Degradation of PET microplastic particles to monomers in human serum by PETase: [Lopez-Lorenzo et al., Faraday Discussions, 2024](https://pubs.rsc.org/en/content/articlehtml/2024/fd/d4fd00014e)\n2. Discovery and quantification of plastic particle pollution in human blood: [Leslie et al., Environment International, 2022](https://www.sciencedirect.com/science/article/pii/S0160412022001258)\n3. A bacterium that degrades and assimilates PET: [Yoshida et al., Science, 2016](https://www.science.org/doi/10.1126/science.aad6359)\n4. Machine learning-aided engineering of hydrolases for PET depolymerization: [Lu et al., Nature, 2022](https://www.nature.com/articles/s41586-022-04599-z)\n5. A Review of Human Exposure to Microplastics and Insights Into Microplastics as Obesogens: [Kannan et al., Frontiers in Endocrinology, 2021](https://www.frontiersin.org/journals/endocrinology/articles/10.3389/fendo.2021.724989/full)\n\n**Research from KTH, Sweden, demonstrates\u00a0that a bacterial enzyme can break PET microplastics into harmless building blocks in human serum \u2014 without harming cells**\n\nEvery year, humans ingest and inhale tens of thousands to millions of [microplastic particles](https://www.frontiersin.org/journals/endocrinology/articles/10.3389/fendo.2021.724989/full) \u2014 and in 2022, researchers made it official: those particles get into our blood. A [landmark study by Leslie et al.](https://www.sciencedirect.com/science/article/pii/S0160412022001258) measured plastic in the blood of 17 out of 22 healthy adults, finding an average concentration of 1.6 \u00b5g/mL, with PET \u2014 the same polymer used in drinks bottles and food packaging \u2014 present in half the samples. We know these particles are circulating inside us. What we did not know until recently was whether anything could be done about it.\n\nPET stands for polyethylene terephthalate, a plastic woven into daily life, making it hard to avoid. It is made by linking two chemical building blocks \u2014 terephthalic acid (TPA) and ethylene glycol \u2014 into long, tightly packed chains. When PET weathers and fragments into microscopic particles, those chains retain their structure, making the material stubborn and biologically inert in ways that are not well understood but may not be harmless. In 2016, [Yoshida et al.](https://www.science.org/doi/10.1126/science.aad6359) discovered a soil bacterium, *Ideonella sakaiensis*, that produces an enzyme \u2014 PETase \u2014 capable of snipping those polymer chains apart. The enzyme works by hydrolysis, slipping water molecules into the chemical bonds that hold PET together, breaking them apart, and returning them to TPA and MHET (mono(2-hydroxyethyl) terephthalate), the same monomers the polymer is built from.\n\nThe key question for any potential medical application was whether PETase could function inside the body. Blood serum is a chemically complex environment \u2014 full of proteins, salts, antibodies, hormones, and enzymes \u2014 and enzymes that work in laboratory buffer conditions often fail in it. Researchers at KTH Royal Institute of Technology in Stockholm set out to answer this directly. In a study published in [Faraday Discussions in 2024](https://pubs.rsc.org/en/content/articlehtml/2024/fd/d4fd00014e), Lopez-Lorenzo, Hueting, Bosshard, and Syr\u00e9n tested an engineered variant of PETase \u2014 the S238A mutant \u2014 against microplastic-like PET particles in both simulated body fluid (SBF) buffer and real human serum, at 37 \u00b0C to mimic body temperature. The substrate they used had been pre-treated by microwave to enrich it in the crystalline, trans-conformation chains that dominate real environmental microplastics \u2014 the form that is notoriously resistant to enzymatic attack.\n\n**The enzyme works!**\n\nAs shown in **Panel A** of figure below, the S238A PETase released measurable quantities of both TPA and MHET across a wide range of substrate concentrations in SBF buffer \u2014 from 0.25 mg/mL up to 10 mg/mL \u2014 with product release scaling with substrate load, reaching approximately 2,100 \u00b5M of TPA and over 4,100 \u00b5M of MHET at 10 mg/mL. Strikingly, the enzyme showed a clear preference for producing TPA, the more rapidly excreted and better-characterized of the two products. This selectivity matters: TPA has been shown to be non-toxic to human fibroblasts at concentrations up to 200 \u00b5g/mL and is rapidly cleared through urine in animal models.\n\nThe serum experiment, summarised in **Panel C**, delivered the most striking result. Starting from zero, around 500 \u00b5M of total monomers was released within the first 15 minutes of incubation, rising to approximately 850 \u00b5M by 60 minutes. The reaction was not stopped by the biological complexity of real human serum \u2014 it was catalytically competent under conditions closely resembling those in a living bloodstream. This rate of monomer release, in the authors\u2019\u00a0assessment, is comparable to that observed in buffer, providing important proof of feasibility.", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\nRecombinant His tag_hMPL and recombinant Flag tag_GFP_tev_hmutCALR Type 1 proteins (80 nM MPL + 5 nM mutCALR) were incubated in assay buffer (HEPES, pH 7.5 50 nM; Prionex 0.05%; NaCl 100 nM; Pluronic F-127 0.01%; CaCl2 1 mM; MgCl2 1 mM; DTT) for 1 h at room temperature to allow for the formation of the mutCALR/MPL complex. In a 384-well plate, 5 \u03bcl of anti-mutCALR antibodies 6 and 32, non-functional CAL2 antibody (Dianova), or untagged mutCALR (competitor) was incubated with 5 \u03bcl of the mutCALR/MPL protein mix for 1 h at room temperature. HTRF detection solution (2 nM anti-FLAG-Europium, and 100 nM anti-6xHis-SureLight APC) was then added and the plate was read on BMG PHERAstar FSX using kinetic HTRF protocol, obtaining readings every 15 minutes for 120 minutes total. Data showed that binding equilibrium was reached by 90 min and the 90 min time point HTRF Ratio values were used for analysis. HTRF Ratio from each test well containing FLAG-GFP-tev-mutCALR + MPL-6xHis solution was background-subtracted using control well (MPL_His tag not included). Percent-of-control values were obtained by dividing the background-subtracted HTRF Ratios for each test well by the buffer control well in the same plate row. The data shows that untagged mutCALR competes with FLAG tag_mutCALR for binding to MPL (positive control) while the isotype does not interfere with the mutCALR/MPL interaction (negative control). Whereas the non-functional CAL2 antibody showed some signal inhibition at high concentration (> 100 nM), the anti-mutCALR antibodies 6 and 32 fully inhibited the interaction of MPL and mutCALR, indicating that the anti-mutCALR antibodies compete with MPL for binding to mutCALR (FIG. 11).\n\n### Example 11: Structural Determination of Fab Fragment from Anti-mutCALR Antibody Bound to mutCALR 21-mer Peptide\n\nFor Examples 11 and 12, and FIGs. 12, 13, 14, 15, 16 and 17, the CCG numbering scheme from MOE (Molecular Operating Environment, 2019.01; Chemical Computing Group ULC, 1010 Sherbrooke St. West, Suite #910, Montreal, QC, Canada, H3A 2R7, 2021) was used for CDR definitions and antibody sequence numbering.\n\nThe Fab portion of anti-mutCALR antibody (clone 55; also referred to as antibody 55) was expressed and purified as follows: The Fab portion of antibody 55 was expressed in Expi293F cells (Thermo Fisher, cat. #A14635) by transient transfection for 5 days. The Fab was purified from clarified supernatants by binding to CaptureSelect CH1-XL Affinity Matrix (ThermoFisher), washing with PBS buffer, and eluting with 50 mM sodium acetate, pH 4.0. After elution, Fab was buffer exchanged to PBS and concentrated using Ultra-15 centrifugal filter unit, 10 kDa MWCO (Amicon) and then stored at \u221280 \u00b0C.\n\n**Sequences:**\n\nFab Heavy Chain (SEQ ID NO:327):\n\n```\nEVQLVQSGAEVKKPGASVKVSCKVSGYTLTELSMQWVRQAPGKGLEWMGGFDPDDGETMYAEKF\nQGRLTVTEDTSTDTVYMELRSLTSEDTALYFCATSPGYDFFDYWGQGTLVTVSSASTKGPSVFP\nLAPSSKSTSGGTAALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSS\nSLGTQTYICNVNHKPSNTKVDKRVEPKSCDKTHT\n```\n\nFab Light Chain (SEQ ID NO:213):\n\n```\nQSALTQPASVSGSPGQSITISCTGTSSDVGGYNYVSWYQQHPGKAPKLMIYEVSNRPSGVSNRF\nSGSNSGNTATLTINRVEAGDEADYYCQVWDSSNDLLIFGGGTKLTVLGQPKAAPSVTLFPPSSE\nELQANKATLVCLISDFYPGAVTVAWKADSSPVKAGVETTTPSKQSNNKYAASSYLSLTPEQWKS\nHRSYSCQVTHEGSTVEKTVAPTECS\n```\n\n**Crystallization** was performed as follows:\n\n*Fab:* The Fab portion of antibody 55 was concentrated to 10 mg/mL in TBS buffer. Sparse matrix crystallization screens were set up using an NT8 crystallization robot (Formulatrix, Bedford, MA). Drops containing 200 nl (Fab fragment) + 200 nl (reservoir) were used for the setup and the plates were incubated at 4 \u00b0C, 13 \u00b0C and 20 \u00b0C. Crystals of the Fab portion of antibody 55 appeared in condition C10 of the JCSG Top96 Screen (0.1 M HEPES pH 7.5, 20% (w/v) PEG 8,000; Rigaku, Bainbridge Island, WA) after five days incubation at 4 \u00b0C.\n\n*Fab-mutCALR Peptide:* The Fab portion of antibody 55 was mixed with mutCALR peptide (Acetyl-DEEQRTRRMMRTKMRMRRMRR-NH2; SEQ ID NO:339) at a ratio of 1:1.5 molar excess, and then concentrated to a final concentration of 35 mg/mL. Sparse matrix crystallization screens were set up using an NT8 crystallization robot (Formulatrix, Bedford, MA). Drops containing 200 nl (Fab plus mutCALR peptide) + 200 nl (reservoir) were used for the setup and the plates were incubated at 4 \u00b0C, 13 \u00b0C and 20 \u00b0C. Initial crystal hits grew from condition C7 of the JCSG Top96 Screen (Rigaku, Bainbridge Island, WA) after 3 days incubation at 4 \u00b0C. This initial hit condition was further refined, leading to the mutCALR peptide bound crystals. The final condition for Fab-mutCALR peptide was 0.1 M Bicine pH 7.94, 19% w/v PEG 6000. These crystals grew in approximately five days.\n\nAll crystals were flash cooled in liquid nitrogen for X-ray data collection.\n\n**Data Collection, Processing and Refinement** was performed as follows:\n\nDiffraction data was collected at 100 K using synchrotron radiation at the Advanced Photon Source (IMCA-CAT beamline 17-ID). Diffraction data indexing, integration and scaling were performed with the AutoPROC package. Data collection statistics, phasing and refinement are given in Table 14.", "# pants.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n\u2020 cleft width measured over fewer\nthan four aromatic residues: it is a percentile of the distances between them, so with three instead\nof five it is not comparable with the rest of the column. TurboPETase is the case in point, and it is\nnot an artefact: two of its eight substitutions delete an aromatic side chain outright.\nWhere a source records several optima, the one measured on **PET** is shown in\npreference to one measured on a model ester, and failing that the one with the most independent\npublications behind it; hover any value for the basis and the alternatives. That choice matters:\nLCC's optimum is 50 \u00b0C on pNP-butanoate but above 70 \u00b0C on PET, and the first would place it\nnear the therapeutic end of this table for a substrate the project is not about.\nTopt is in \u00b0C and cleft width in \u00c5;\nthe coloured edge marks the family. %ID is against the wild type at the root of\nthat enzyme's own lineage, not against IsPETase: LCC-ICCG is 98.6% identical to LCC and only about\n50% to IsPETase, and only the first says how far engineering moved it. Wild types are 100% by\ndefinition. Optimum temperatures come from UniProt where\nan experimental value is curated and from the literature review otherwise; each enzyme page shows\nwhich, with its evidence code. Where a value is published as a range the table shows the midpoint\nand the page shows the range. Cleft width is measured on the structure named in the Structure\ncolumn, so it mixes crystal structures with predictions; the page says which for each.\n\n## By source environment\n\n| Environment | Sequences | Candidates | Per million | Mean %ID to nearest |\n| --- | --- | --- | --- | --- |\n| Human gut | 12,584,458 | 311 | 25 | 30.1 |\n| Compost | 1,020,575 | 69 | 68 | 47.8 |\n| Marine plastisphere | 737,027 | 44 | 60 | 27.6 |\n| Landfill | 436,229 | 15 | 34 | 35.7 |\n\nEvery marine plastisphere candidate sits below 40% identity to\nanything characterised: nothing in that cohort is a rediscovery. Compost, where the field has already\nlooked, returns near-identical copies of known enzymes.\n\n## Ranked by retrieval bitscore\n\nThe retrieval baseline, which deliberately favours what is already known: the\ninteresting candidates are the low-identity ones this ranking pushes down.", "# ChatMol \u2013 Natural-language control of PyMOL, describe the scene, get the picture\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n**`Ollama at ... did not respond to model '...' within 120s`**\nThe model took too long to respond \u2014 this is caught gracefully and won\u2019t\ncrash PyMOL. Common causes: the model is still loading into memory (try\nagain), the machine is under heavy load (e.g. another `ollama pull` running\nin the background), or the model is just slow for your hardware. Increase\n`chatmol_config timeout, 240` or switch to a smaller/faster model.\n\n**`The model's response looked malformed or repetitive ... and was\ndiscarded`**\nSmall models occasionally loop (e.g. repeating `enable all; enable all; ...`)\ninstead of producing a short translation. ChatMol detects this, discards the\nresponse, and \u2014 importantly \u2014 does *not* add it to conversation history (so\nit can\u2019t poison later requests). Just try again, rephrase, or switch to a\nlarger model (`chatmol_config model, qwen3.6:27b`).\n\n**Nothing happens / commands look wrong**\nTurn on debug mode (`chatmol_config debug, on`) and re-run the request to\nsee the model\u2019s raw output. Try a larger model for better translation\nquality, or lower `temperature` for more deterministic output.\n\n**A plain command got sent to the LLM by mistake**\nThe heuristic in `nl_router.looks_like_pymol()` is conservative but not\nperfect. The system prompt asks the model to pass valid PyMOL syntax\nthrough unchanged, so this is usually harmless \u2014 but if you hit a\nrecurring case, add the command/namespace to `_EXTRA_KEYWORDS` in\n`nl_router.py`.\n\n**Disable everything temporarily**\n`chatmol_config auto, off` stops the command line from auto-translating;\n`ai <request>` / `chatmol <request>` still work explicitly.\n\n---\n\n## 8. Extending\n\n* **More commands / better translations**: extend\n  `pymol_reference.PYMOL_CHEAT_SHEET` with examples for operations you use\n  often (e.g. project-specific selection macros).\n* **Session-aware context**: `pymol_reference.build_context()` is the place\n  to add more state (current view, sequence, B-factor ranges, etc.).\n* **Confirmation step**: wrap `nl_router.run_natural_language`\u2018s execution\n  loop with a prompt/dialog if you want to approve translated commands\n  before they run.\n* **Streaming / richer UI**: a dedicated chat panel could be added in\n  `gui.py` using `pymol.Qt` for a conversational view alongside the command\n  line.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/chatmol](https://github.com/bellcheddar/chatmol) |", "# ChatPDB \u2013 A protein-structure-literate AI assistant for the Protein Data Bank: RAG for facts, a QLoRA-tuned model for behaviour\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-07-25 on marcdeller.com (categories: vibe-code).*\n\n> **A protein-structure-literate AI assistant for the Protein Data Bank: RAG for facts, a QLoRA-tuned model for behaviour, and live Biopython/gemmi/DSSP/PyMOL tool calls for structural truth. Fine-tuned and served locally on Apple Silicon with MLX-LM.**\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [chatpdb.mdeller.com](https://chatpdb.mdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/chatPDB](https://github.com/bellcheddar/chatPDB) |\n\n---\n\nchatPDB is a locally hosted, open-source assistant that reasons about protein structures (file formats, experimental methods, cross-reference databases, structure-manipulation tools), drives the structural-biology tools a working structural biologist actually uses (Biopython, gemmi, DSSP, PyMOL, ChimeraX, and more), and grounds every factual claim in the real wwPDB/RCSB/SIFTS/UniProt corpus rather than memory.\n\n**Why it matters:** off-the-shelf models invent structural facts (wrong resolutions, hallucinated PDB IDs, fabricated secondary-structure content) and have no way to compute a real value rather than guess one. chatPDB fixes both by separating concerns: volatile knowledge lives in retrieval, stable behaviour lives in the fine-tuned weights, and deterministic structural facts (atom counts, secondary structure, chain composition, geometry) are computed live by real tools, executed in a sandbox, rather than asserted from memory. It is useful for: interrogating deposited PDB structures and their cross-references, generating correct Biopython/gemmi/DSSP/PyMOL scripts, honestly declining to answer when an ID doesn\u2019t exist or a value isn\u2019t known, and reasoning like a structural biologist about experimental method, quality, and biological assembly rather than treating a structure as a bare coordinate file. A live, no-install terminal session against the real fine-tuned model runs at **[chatpdb.mdeller.com](https://chatpdb.mdeller.com)** (see **Web deployment** below).\n\nSibling project to [ChemSage](https://github.com/bellcheddar/ChemSage), reusing its proven MLX-LM/QLoRA/RAG process on a fresh domain: protein structures instead of small-molecule chemistry.\n\n**Project links:**\n\n* [GitHub](https://github.com/bellcheddar/chatPDB)\n* [Models on Hugging Face](https://huggingface.co/Dellboy)\n\n## \ud83e\udd17 Models on HuggingFace Hub\n\n| Model | HuggingFace | Base | SFT dataset | Round |\n| --- | --- | --- | --- | --- |\n| **chatPDB 32B v1** \u2b50 | [**Dellboy/chatpdb_32b_v1**](https://huggingface.co/Dellboy/chatpdb_32b_v1) | Qwen3-32B-4bit | v7 (97,272 examples) | **Round 1 (current)** |\n\nAdd a row here for every future fused model (never overwrite a previous row) \u2014 see [\ud83c\udf93 Training](#-training) below for how each round\u2019s real numbers get recorded.\n\n## \u2699\ufe0f The hybrid in one line\n\nRAG keeps it truthful today; the fine-tune makes it consistent tomorrow; Biopython/gemmi/DSSP/PyMOL make it correct always.\n\n## \ud83d\ude80 Quick start\n\nSee **[`PROJECT_PLAN.md`](PROJECT_PLAN.md)** for the full step-by-step build (phases 0 to 10). Short version:\n\n| Step | What | Script / command |\n| --- | --- | --- |\n| 0 | Env: MLX-LM + Biopython + structural tools | (`requirements.txt`) |\n| 1 | Base model (Qwen3-32B, MLX 4-bit) | `config/train_config.yaml` |\n| 2 | **RAG first** (ship this alone) | `scripts/ingest_rag.py` |\n| 3 | Build + validate SFT dataset | `scripts/build_dataset.py` |\n| 4 | QLoRA fine-tune | `python scripts/train_launch.py --config config/train_config.yaml` |\n| 5 | Fuse + serve | `python scripts/merge_export.py` \u2192 `mlx_lm.server --port 8080` |\n| 6 | Close the hybrid loop | `rag/tool_exec.py` (Biopython sandbox; gemmi/DSSP/PyMOL staged next) |\n| 7 | Structural evaluation | `python eval/eval_pdb.py` \u00b7 `python eval/compare/eval_compare.py` |\n| 8 | Local CLI + RAG | `python scripts/chat.py` |\n| 9 | Hosted demo (no local install) | live at [chatpdb.mdeller.com](https://chatpdb.mdeller.com), see **Web deployment** below |\n\n## \u25b6\ufe0f How to run\n\n**Interactive chat** (loads the model in-process \u2014 no server needed, unlike the eval scripts below):\n\n```\ncd /Users/dellboy/Documents/Vibe_Coding/chatPDB\nsource .venv/bin/activate\npython scripts/chat.py --model models/chatpdb_32b_v1\n```\n\n`/help` for slash commands, `/info` for session details, `quit`/`exit`/`q` to leave. `--no-rag` runs model-only (no retrieval).\n\n**Evaluate the model** (these two DO need `mlx_lm.server`; `eval_compare.py` manages its own):\n\n```\n# Terminal 1\npython scripts/preflight.sh   # flush memory, check for background-process contention\nmlx_lm.server --model models/chatpdb_32b_v1 --port 8080\n\n# Terminal 2, once Terminal 1 shows \"HTTP server listening\"\npython eval/eval_pdb.py --n 200                 # single-model scorecard, needs Terminal 1 running\npython eval/compare/eval_compare.py              # multi-round comparison, starts its own server\n```\n\n## \ud83c\udf93 Training\n\nReal numbers only, recorded per round \u2014 extend this table (never edit a previous round\u2019s row) as new training rounds ship.", "# ChemSage: Vibe-Coding a Chemically Aware Drug-Discovery LLM on Local Hardware\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nThe architecture (from the [project plan](https://github.com/bellcheddar/ChemSage/blob/main/PROJECT_PLAN.md)) wires those steps together as a single local pipeline, served behind an OpenAI-compatible endpoint so any GUI can talk to it.\n\n```\n\n\t\t\t\t\tuser   +---------------------------------------------+\n query-&gt;|  GUI front end (AnythingLLM / Open WebUI)   |\n        +----------------------+----------------------+\n                               v\n        +---------------------------------------------+\n        |   RAG retriever (Chroma + embeddings)       | &lt;-- ingested corpus (papers, SAR, dossiers)\n        +----------------------+----------------------+\n                               v   query + retrieved context\n        +---------------------------------------------+\n        |   mlx_lm.server  -&gt;  fused QLoRA model (MLX)|\n        +----------------------+----------------------+\n                               v   may emit a tool call (RDKit / PyMOL / dock)\n        +---------------------------------------------+\n        |   tool-exec shim (sandboxed)                |\n        +----------------------+----------------------+\n                               v   validated results (canonical SMILES, props, RMSD)\n                         grounded answer\n\n```\n\n###### The stack, and why MLX-LM?\n\nThe whole thing is pure Python on Apple Silicon. The dependency list is short, which is part of the point.\n\n```\n\n\t\t\t\t\t# Fine-tune + serve: MLX-LM (Apple Silicon / Metal)\nmlx-lm                 # Apple Silicon-native LLM toolkit: LoRA fine-tune, fuse adapters, and run local server\n\n# Core / shared\nrdkit                  # Cheminformatics toolkit for molecular representations, descriptors, and chemistry ground truth\ntransformers&gt;=4.46     # Hugging Face model + tokenizer stack for model zoo access and config/weights handling\npandas                 # Tabular data handling, corpus metadata, logs, and evaluation outputs\nnumpy                  # Array math backbone for tensors, preprocessing, and numerical utilities\npyyaml                 # Human-readable config management for experiments, models, and pipelines\n\n# RAG layer\nchromadb               # Local/remote embedding database for vector storage, filtering, and similarity search\nsentence-transformers  # Local embedding models for document/query vectors feeding the RAG index\npypdf                  # PDF ingestion: split pages, extract text, and build structured document chunks for RAG\n\n# Chat CLI + eval client\nrich                   # Structured, colorful terminal output for chat UI, logs, and evaluation reports\nprompt_toolkit         # Interactive command-line interface: input editing, history, and keybindings for the chat client\nrequests               # HTTP client for talking to remote MLX-LM servers, RAG APIs, or evaluation services\ntqdm                   # Progress bars for training loops, corpus ingestion, and batch evaluation runs\npsutil                 # System resource monitoring to keep Apple Silicon runs and RAG indexing within hardware limits\n\n```\n\nA handful of scripts carry the load. `ingest_rag.py` chunks, embeds and stores the corpus; `build_dataset.py` generates and validates the training data; `train_qlora.py` wraps `mlx_lm.lora`; `merge_export.py` fuses the adapter and serves it; `tool_exec.py` is the sandboxed RDKit and PyMOL runner; and `eval_chem.py` scores the result. The engine underneath all of it is [MLX](https://mlx-framework.org/), Apple\u2019s array framework, and its [MLX-LM](https://github.com/ml-explore/mlx-lm) companion for language models.\n\nMLX matters because of how Apple Silicon handles memory. A discrete NVIDIA card forces a hard VRAM wall and constant copying between CPU and GPU memory. Apple\u2019s unified-memory architecture removes both: the CPU and GPU share one pool, so there are no memory copies and no separate VRAM budget. In practice, that means a 16 GB Mac comfortably tunes an 8B model, a 32 GB machine handles roughly a 14B model, and a 64 GB machine reaches a full 32B model. The trade-off against NVIDIA is wall-clock speed, not capability, and because the base model is already a 4-bit checkpoint, the fine-tune runs automatically as QLoRA. Apple\u2019s own write-up on [MLX and the M5 neural accelerators](https://machinelearning.apple.com/research/exploring-llms-mlx-m5) sets out the same case in more detail.\n\n###### The training corpus\n\nThe dataset is where this project succeeds or fails, so the rules are strict. Every numeric answer is generated by first running RDKit to compute the real value, then templating the question and the worked solution around it, never by hand-authoring a property. Each example is validated on write: if its code block fails to run or its SMILES will not canonicalize, it is rejected. SMILES are kept inside fenced code blocks so the model treats them as literals it passes to a tool, and the splits are an 80/10/10 train/validation/frozen-test partition. The supervised set grew across rounds from 1,500 examples to 20,000, spanning 78 behavior classes, alongside a 55,742-chunk retrieval corpus drawn from ChEMBL, PubChem and the PDB. The [data layout](https://github.com/bellcheddar/ChemSage/tree/main/data#readme) keeps the rounds separate and the test set untouched.\n\n```\n\n\t\t\t\t\tdata/\n|-- corpus/          # RAG sources: PDFs, SAR CSVs, dossiers (gitignored)\n|-- sft/             # Rounds 1-4 SFT data\n|   |-- train.jsonl  # 6,400 examples (Round 4)\n|   |-- valid.jsonl  # 800 examples\n|   '-- test.jsonl   # 800 examples, frozen, eval only\n'-- sft_r5/          # Round 5 SFT data\n    |-- train.jsonl  # 16,000 examples\n    |-- valid.jsonl  # 2,000 examples\n    '-- test.jsonl   # 2,000 examples, frozen, eval only\n\n```\n\n###### Five rounds, and how they were judged", "# AXIS: Closing the Loop on Crystal Detection\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-04-14 on marcdeller.com (categories: sb).*\n\n* [**AXIS: a Lab-in-the-Loop machine learning approach for automated detection of macromolecular crystals**](https://journals.iucr.org/m/issues/2026/03/00/jt5083/index.html)\n\n* **Persistent bottleneck resolved:** [AXIS delivers automated crystal annotation at human-expert accuracy](https://pubmed.ncbi.nlm.nih.gov/41854216/), closing the last manual gap in high-throughput crystallography pipelines.\n* **DINOv2 foundation:** [Transfer learning from DINOv2](https://debuggercafe.com/dinov2-for-image-classification-fine-tuning-vs-transfer-learning/) onto the MARCO dataset gives AXIS robust generalisation across diverse crystallization image types.\n* **Continuous local adaptation:** [Iterative expert feedback](https://journals.iucr.org/m/issues/2026/03/00/jt5083/) lets AXIS specialise to each facility\u2019s optical conditions, correcting human annotation errors over time.\n* **Drug discovery accelerated:** [Eliminating manual image review](https://www.biorxiv.org/content/10.1101/2025.11.03.685844v1) directly speeds fragment and ligand screening campaigns requiring rapid, large-scale outcome scoring.\n\n1. AXIS: a Lab-in-the-Loop machine learning approach for automated detection of macromolecular crystals: [Personnaz et al., IUCrJ, 2026](https://doi.org/10.1107/S2052252526001399)\n2. Classification of crystallization outcomes using deep convolutional neural networks: [Bruno et al., PLoS ONE, 2018](https://doi.org/10.1371/journal.pone.0198883)\n3. Deep learning applications in protein crystallography: [Matinyan et al., Acta Crystallographica Section A, 2024](https://doi.org/10.1107/S2053273323009300)\n4. The automated crystallography pipelines at the EMBL HTX Facility: [M\u00e1rquez et al., Journal of Visualized Experiments, 2021](https://doi.org/10.3791/62491)\n\n**Intelligent automation finally bridges the last manual gap in high-throughput macromolecular crystallography**\n\nMacromolecular crystallography has long served as the mechanistic backbone of structural biology, providing atomic-resolution insights into proteins and nucleic acids that are essential to both fundamental discovery and structure-guided drug design. Over the past two decades, robotic handling, liquid dispensing automation, and synchrotron-coupled harvesting systems have dramatically compressed the protein-to-structure timeline, particularly for fragment and ligand screening campaigns that can generate tens of thousands of crystallization images in a single experiment. Yet one step has persistently resisted full automation: the visual identification of crystals from imaging plates. Manual inspection at this scale is not merely slow; it is inconsistent, cognitively taxing, and introduces subjective annotation errors that propagate through downstream decisions. To address this, researchers at the European Molecular Biology Laboratory in Grenoble developed AXIS, an AI-based Crystal Identification System that integrates the self-supervised DINOv2 vision transformer with state-of-the-art transfer learning and the MARCO dataset, the largest annotated\u00a0collection of crystallization images.\n\nAXIS was trained on nearly 77,000 images consolidated from MARCO\u2019s half-million-image repository into crystal and non-crystal classes, then fine-tuned on local facility data to capture the specific morphological and optical characteristics of each imaging environment. The system operates equally well on both visible-light and UV-fluorescence images, the latter providing intrinsic protein fluorescence signals that are particularly informative for confirming the proteinaceous nature of hits. Most significantly, AXIS implements a Lab-in-the-Loop methodology in which domain experts periodically review model outputs, correct errors, and resubmit curated annotations as a training signal, enabling the model to continuously specialize to local conditions rather than degrading over time. In practice, this iterative feedback architecture means AXIS not only matches human expert accuracy on large-scale annotation tasks but also actively corrects annotation errors introduced by human reviewers, removing a source of label noise that has undermined prior automated systems. This results in a deployable, facility-agnostic detection pipeline that removes the final manual bottleneck in high-throughput crystallography workflows, opening a credible path to fully automated protein-to-structure pipelines at scale for both academic and industrial structural biology.", "# cif2plip \u2013 Protein-ligand interaction profiling using PLIP for co-folding outputs which produce CIF files\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n### Serial renumbering\n\nAll `ATOM`, `HETATM`, and `TER` records are renumbered sequentially from 1 with no gaps. This step is the fix for the OpenBabel CONECT-parsing bug described above.\n\n---\n\n## Dependencies\n\n| Package | Role |\n| --- | --- |\n| [gemmi](https://gemmi.readthedocs.io) | CIF parsing and PDB writing |\n| [pdb-tools](https://www.bonvinlab.org/pdb-tools/) | `pdb_tidy`, `pdb_validate` |\n| [PLIP](https://github.com/pharmai/plip) | Protein-ligand interaction profiling |\n| [OpenBabel](https://openbabel.org) | Ligand bond-order perception (PLIP dependency) |\n| [PyMOL (OSS)](https://pymol.org) | `.pse` session generation (PLIP dependency) |\n\n---\n\n## Licence\n\nMIT\n\n---\n\n## Citation\n\nIf you use this pipeline in published work, please cite PLIP:\n\n> Adasme MF et al. (2021). PLIP 2021: expanding the scope of the protein-ligand interaction profiler to DNA and RNA. *Nucleic Acids Research*, 49(W1), W530\u2013W534. https://doi.org/10.1093/nar/gkab294\n\nAnd gemmi if the CIF conversion step is central to your workflow:\n\n> Wojdyr M (2022). GEMMI: A library for structural biology. *Journal of Open Source Software*, 7(73), 4200. https://doi.org/10.21105/joss.04200\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/cif2plip](https://github.com/bellcheddar/cif2plip) |", "# Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\nAuthors: Hummel, J.R.; Xiao, K.-J.; Yang, J.C.; Epling, L.B.; Mukai, K.; Ye, Q.; Xu, M.; Qian, D.; Huo, L.; Weber, M.; Roman, V.; Lo, Y.; Drake, K.; Stump, K.; Covington, M.; Kapilashrami, K.; Zhang, G.; Ye, M.; Diamond, S.; Yeleswaram, S.; Macarron, R.; Deller, M.C.; Wee, S.; Kim, S.; Wang, X.; Wu, L.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2024\n\n**Step 1. N-(Piperidin-4-yl)-4-(1-(2,2,2-trifluoroethyl)-1H-pyrazol-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (36).** A mixture of tert-butyl 4-((4-chloro-5-(trifluoromethyl)pyrimidin-2-yl)amino)piperidine-1-carboxylate (23, 3.00 g, 7.88 mmol), 4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1-(2,2,2-trifluoroethyl)-1H-pyrazole (35, 2.61 g, 9.45 mmol), Pd(dppf)Cl2\u00b7CH2Cl2 (0.643 g, 0.788 mmol), and Na2CO3 (5.01 g, 47.3 mmol) in CH3CN/H2O (5:1, 20 mL) was irradiated in a microwave reactor at 110 \u00b0C for 2 h. After cooling to r.t., HCl (4 M in 1,4-dioxane, 24 mL, 96 mmol) was added and the reaction mixture was stirred at r.t. overnight. The mixture was diluted with H2O and extracted with CH2Cl2. The organic phase was removed, and the aqueous phase was made basic with NaOH and extracted with CH2Cl2 and EtOAc. The combined organic phases were washed with brine, dried over MgSO4, and concentrated. The crude material was used without further purification. LCMS calculated for C15H17F6N6 (M + H)+: m/z = 395.1; found, 395.3.\n\n**Step 2. N-(1-((1-Methyl-1H-imidazole-4-yl)sulfonyl)piperidin-4-yl)-4-(1-(2,2,2-trifluoroethyl)-1H-pyrazol-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (14).** To a mixture of 36 in THF (20 mL) was added 1-methyl-1H-imidazole-4-sulfonyl chloride (1.42 g, 7.86 mmol) followed by dropwise addition of NEt3 (1.1 mL, 7.9 mmol), and the reaction mixture was stirred at r.t. for 1 h. The reaction mixture was concentrated in vacuo. The crude residue was taken up in CH3CN, H2O, and TFA (0.6 mL) and purified by prep HPLC (pH 2) to afford the title compound (2.03 g, 48% yield over two steps) as a white solid. LCMS calculated for C19H21F6N8O2S (M + H)+: m/z = 539.1; found, 539.1.\n\n**N-(1-((1-Methyl-1H-pyrazol-4-yl)sulfonyl)piperidin-4-yl)-4-(1-(2,2,2-trifluoroethyl)-1H-pyrazol-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (13).** The title compound was prepared according to the procedure described for 14 with 1-methyl-1H-pyrazole-4-sulfonyl chloride replacing 1-methyl-1H-imidazole-4-sulfonyl chloride. Off-white solid; 22% yield over two steps (37 mg, 69 \u03bcmol). LCMS calculated for C19H21F6N8O2S (M + H)+: m/z = 539.1; found, 539.1.\n\n**Step 1. 2-(Piperidin-4-ylamino)-4-(1-(2,2,2-trifluoroethyl)-1H-pyrazol-4-yl)pyrimidine-5-carbonitrile (37).** The title compound was prepared according to the procedure described for 36 with tert-butyl 4-((4-chloro-5-cyanopyrimidin-2-yl)amino)piperidine-1-carboxylate (31) replacing tert-butyl 4-((4-chloro-5-(trifluoromethyl)pyrimidin-2-yl)amino)piperidine-1-carboxylate. LCMS calculated for C15H17F3N7 (M + H)+: m/z = 352.1; found, 352.2.\n\n**Step 2. 2-((1-((1-Methyl-1H-imidazole-4-yl)sulfonyl)piperidin-4-yl)amino)-4-(1-(2,2,2-trifluoroethyl)-1H-pyrazol-4-yl)pyrimidine-5-carbonitrile (15).** The title compound was prepared according to the procedure described for 14 with 2-(piperidin-4-ylamino)-4-(1-(2,2,2-trifluoroethyl)-1H-pyrazol-4-yl)pyrimidine-5-carbonitrile replacing N-(piperidin-4-yl)-4-(1-(2,2,2-trifluoroethyl)-1H-pyrazol-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine. White solid; 13% yield over two steps (144 mg, 0.291 mmol). LCMS calculated for C19H21F3N9O2S (M + H)+: m/z = 496.1; found, 496.1.\n\n**2-((1-(Pyridin-2-ylsulfonyl)piperidin-4-yl)amino)-4-(1-(2,2,2-trifluoroethyl)-1H-pyrazol-4-yl)pyrimidine-5-carbonitrile (16).** This compound was prepared according to the procedures described for 14 with 2-(piperidin-4-ylamino)-4-(1-(2,2,2-trifluoroethyl)-1H-pyrazol-4-yl)pyrimidine-5-carbonitrile (37) replacing N-(piperidin-4-yl)-4-(1-(2,2,2-trifluoroethyl)-1H-pyrazol-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine and pyridine-2-sulfonyl chloride replacing 1-methyl-1H-imidazole-4-sulfonyl chloride. White solid; 11% yield over two steps (118 mg, 0.240 mmol). LCMS calculated for C20H20F3N8O2S (M + H)+: m/z = 493.1; found, 493.1.\n\n### Synthesis of Compounds 17 and 18\n\n**Step 1. tert-Butyl 4-((4-(1-(2-hydroxy-2-methylpropyl)-1H-pyrazol-4-yl)-5-(trifluoromethyl)pyrimidin-2-yl)amino)piperidine-1-carboxylate (39).** A mixture of tert-butyl 4-((4-chloro-5-(trifluoromethyl)pyrimidin-2-yl)amino)piperidine-1-carboxylate (23, 10.0 g, 26.3 mmol), 2-methyl-1-(4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazol-1-yl)propan-2-ol (38, 8.39 g, 31.5 mmol), Pd(dppf)Cl2\u00b7CH2Cl2 (1.72 g, 2.10 mmol), and Na2CO3 (8.35 g, 79 mmol) in 1,4-dioxane (105 mL) and H2O (26.3 mL) was degassed with nitrogen and sealed. The reaction mixture was stirred at 100 \u00b0C for 5 h. After cooling to r.t., the reaction was diluted with saturated aqueous NaHCO3 and extracted with EtOAc. The combined organic phases were dried over Na2SO4, filtered, and concentrated. The crude residue was purified by flash chromatography (0\u221260% EtOAc/hexanes) to afford the title compound as an off-white solid. LCMS calculated for C22H32F3N6O3 (M + H)+: m/z = 485.2; found, 485.2.\n\n**Step 2. 2-Methyl-1-(4-(2-(piperidin-4-ylamino)-5-(trifluoromethyl)pyrimidin-4-yl)-1H-pyrazol-1-yl)propan-2-ol hydrochloride (41).** To a mixture of 39 in CH2Cl2/MeOH (4:1, 25 mL) was added HCl (4 M in 1,4-dioxane, 25 mL, 100 mmol) and the reaction mixture was stirred at r.t. for 2 h to form a suspension. The solid precipitate was filtered, washed with Et2O, and dried to afford the title compound (10.6 g, 96% yield over two steps) as a white solid. LCMS calculated for C17H24F3N6O (M + H)+: m/z = 385.2; found, 385.2.", "# Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\nAuthors: Deller, M.C.; Hudson, K.R.; Ikemizu, S.; Bravo, J.; Jones, E.Y.; Heath, J.K.\nSource: Structure, 2000\n\nThis hypothesis also suggests why CNTF interacts with gp130 at a very much lower affinity than either LIF or hOSM. In CNTF the equivalent residue to Gly120 is a tyrosine which, as argued above, would be predicted to disrupt the core interaction. The employment of the accessory receptor CNTF-R in this scheme is required to compensate for this loss of affinity by introducing additional sites of receptor\u2013receptor interaction into the complex.\n\n### The interaction with LIFR and OSMR\n\nA distinctive feature of the gp130 cytokine family is the use of a novel receptor recognition site 3 located at the N terminus of helix D. This epitope engages the second signalling receptor in the complex: either LIFR/OSMR in the case of OSM, LIF and CNTF, or gp130 in the case of IL-6 and IL-11. Previous studies [24,29] have revealed that, for the interaction with LIFR, this site 3 epitope involves a solvent-exposed pair of residues which are conserved in all LIFR-binding cytokines described to date. The availability of the crystal structure of hOSM has enabled us to determine both the extent to which this recognition scheme holds for additional members of the family and to examine the involvement of this epitope in the interaction between hOSM and OSMR. Mutagenesis of the two key site 3 residues in hOSM, Phe160 and Lys163, reveals that each residue is required for the interaction with both LIFR and OSMR. This epitope (Figure 4) therefore represents a core structural element in site 3 recognition, although the precise significance of each residue in the epitope differs between LIFR and OSMR. It follows that the ability of OSM to interact with OSMR must result from either the involvement of additional residues in the vicinity of this epitope or the presence of residues in LIF which block some feature of the interaction with OSMR.\n\nIt is intriguing that the structure of the site 3 recognition epitope is similar to the site 2 recognition site of gp130: both employ a prominent aromatic sidechain as a 'core' with additional affinity contributed by the sidechain of a residue in close physical proximity. This suggests that site 3 receptor recognition may involve a similar overall scheme to that of site 2 except that the complementary epitopes on receptor and ligand are reversed. This suggests that the site 3 recognition site in LIFR, which is located in the immunoglobulin-like domain [31,32,35], may involve an analogous hydrophobic dimple or patch formed from residues with small sidechains. In this respect it is notable that alanine scanning studies of the LIFR site 3 hOSM recognition site imply a short sequence motif with a conserved glycine residue at its core [32].\n\n## Biological implications\n\nThe interaction of cytokine ligands with transmembrane receptors to form signalling complexes is a major mechanism for intercellular signal transduction and cell regulation. The molecular specificity of the extracellular cytokine\u2013receptor interaction is of fundamental biological significance as it dictates the resulting intracellular signal-transduction processes and, accordingly, the associated biological responses. The family of cytokines which mediate their effects through binding to the gp130 receptor \u2014 namely oncostatin M (OSM), leukaemia inhibitory factor, interleukins 6 and 11, cardiotrophin-1 and ciliary neurotrophic factor (CNTF) \u2014 are of particularly broad biomedical importance. Of these gp130-binding cytokines, human OSM (hOSM) is of potential interest for novel cancer therapeutics as well as providing a target for the treatment of inflammatory responses in the joints and nervous system.\n\nThe combined crystal structure and mutagenesis analysis of hOSM reported here has provided important insights into the molecular basis of biological specificity in the gp130 cytokine system. Particular value accrues from the structural comparisons these data now allow with a panel of gp130 cytokines that exhibit both common and divergent receptor-recognition properties. The study confirms the concept [36] that receptor recognition involves a conserved non-polar 'core' interaction with specificity contributed by the identity of residues surrounding the core. This study also suggests a second aspect to recognition specificity: the requirement for soluble receptors in some gp130 signalling complexes (e.g. CNTF receptor in the CNTF-mediated recognition complex) acting to compensate for low-affinity core interactions by the introduction of additional receptor\u2013receptor interactions. Finally, this study provides the structural basis for the development of small-molecule compounds to intervene in hOSM-mediated signalling events.\n\n## Materials and methods\n\n### Production and expression of hOSM proteins\n\nA 21.4 kDa fragment of mature hOSM (residues 1\u2013187) was expressed as a GST fusion protein using the pGex-2T expression vector (Pharmacia). The hOSM and GST proteins were joined by a recognition site for rhinovirus 3C protease. All OSM mutants were generated by overlap polymerase chain reaction from the pGEX\u2013hOSM template GST fusion using specific oligonucleotides for each mutant; the sequences of primers used are available on request. hOSM (1\u2013187) and mutant proteins were expressed as GST fusions in E. coli strain JM109. Soluble GST fusion proteins were bound to glutathione\u2013sepharose resin and the OSM protein released by overnight cleavage with 3C protease at 37\u00b0C and purified by high-resolution gel-filtration on a Sephadex-75 column (Pharmacia). All proteins had a purity of greater than 80% as determined by SDS PAGE and silver staining. Protein concentrations were determined by the Coomassie Protein Plus assay (Bio-rad).", "# How to Confuse Machine Learning: Understanding Modern Adversarial Attacks\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Concept | Description | Key References |\n| --- | --- | --- |\n| Adversarial Examples | Specially crafted inputs that cause misclassification while appearing normal to humans | <https://arxiv.org/pdf/2310.17626.pdf> |\n| Data Poisoning | Systematic corruption of training data to influence model behavior in attacker-desired ways | <https://ceur-ws.org/Vol-3910/aics2024_p10.pdf> |\n| Prompt Injection | Manipulation of language model inputs to override safety instructions or extract information | <https://www.nature.com/articles/s41467-024-55631-x> |\n| Model Extraction | Theft of proprietary model functionality through strategic API querying techniques | <https://arxiv.org/html/2506.22521v1> |\n| Byzantine Attacks | Coordinated deception in distributed learning environments where participants provide false data | <https://www.usenix.org/conference/usenixsecurity20/presentation/fang> |\n| Gradient Masking | Defensive technique that obscures gradient information but may provide false sense of security | <https://proceedings.mlr.press/v80/athalye18a.html> |", "# GitHub README: ButtFold\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ButtFold, 2026\n\n- [x] **Measure before building.** Droplet fold times, WASM speed in three browsers, WASM parity\n      against native, and the baked payload size, all in `METRICS.md` before a line of the app\n      existed. Two of the plan's own numbers had to change as a result\n- [x] **The baked gallery.** Six G\u014d folds with the collapse and first-frame-contact assertions\n      green, served by Flask with explicit Cache-Control because Flask sends none on a template\n      and heuristic caching then makes a deploy invisible\n- [x] **The 3D stage.** three.js tube along the CA trace, coloured by secondary structure, with\n      live readouts and two time series. A degenerate initial spline gave the tube a NaN bounding\n      sphere and three.js frustum-culled it forever behind a page that otherwise looked perfect\n- [x] **Note-for-note sonifier parity with the shipped app.** Float32 is emulated with\n      `Math.fround` wherever the Swift uses `Float`: mean confidence accumulates over 76 residues\n      in single precision on the phone, and the difference reaches note velocity and the cadence\n- [x] **The Web Audio engine.** All five style profiles unchanged, per-note HRTF panning, a\n      lookahead scheduler, and the audio clock driving the animation rather than the reverse\n- [x] **The live WebAssembly fold.** `wasm_api.c` is additive: it `#include`s the vendored C\n      rather than editing it, and reproduces the CLI's trajectory bitwise\n- [x] **The droplet queue.** One worker at nice 19 behind a residue cap, a depth cap, a per-IP cap\n      and a timeout of three times the measured worst case, with a cache that converges\n- [x] **The honesty layer.** The app's own disclosure strings byte for byte, an engine badge that\n      never scrolls away, and a disclosure paragraph as body text rather than behind a link\n- [x] **PhoneFold's camera.** Ported after drag-to-rotate proved broken the same way the app's\n      once was: yaw and pitch with a clamp, which makes a vertical drag die mid-gesture. A\n      quaternion attitude on the subject has no pole to protect\n- [ ] **Cartoon secondary structure.** The ribbon is a plain tube, so a helix reads as a coiled\n      tube and a strand as a straight one. The app draws a flat helical ribbon and an arrowed\n      strand, and this is the biggest remaining visual gap\n- [ ] **Genie 2 gallery entries.** Two to four backbones precomputed on the Mac, carrying the\n      generative disclosure on both the card and the stage. Genie 2 needs Core ML or a GPU, so\n      these can only ever be baked, never folded live\n- [ ] **Mobile Safari's fold time.** The one measurement no machine here can take, and the one\n      that decides whether phones fold live at all or route to the queue\n- [ ] **The PhoneFold store link.** `static/links.json` carries a null `app_store_url`, so the\n      card honestly says \"in review\". One field flips it when the app clears review\n\n## \ud83d\udcda Provenance and licences\n\nThe G\u014d model, the five style profiles, the native structures and the geometry and audio ports all\ncome from [PhoneFold](https://github.com/bellcheddar/PhoneFold) (MIT), vendored with provenance\nheaders naming the source commit and re-vendored rather than patched. The C is unedited.\n[`THIRD-PARTY.md`](THIRD-PARTY.md) lists everything, with the papers behind the model, the\nsecondary-structure assignment, the hydropathy scale and both radius-of-gyration scaling laws.\n\nButtFold's own code is MIT: see [`LICENSE`](LICENSE).\n\nThe name is a joke about an accidental butt-dial. The page says so once, in the footer, and never\nagain.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/ButtFold\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/ButtFold</a></td>\n</tr>\n</table>", "# Democratizing Bioinformatics with Multi-Agent Language Models\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-20 on marcdeller.com (categories: ai).*\n\n* Enabling end-to-end bioinformatics pipelines on local hardware with proprietary data.\n* Achieving expert-comparable performance on conceptual genomics tasks.\n* Streamlining multi-step workflows (RNA-seq, ChIP-seq, single-cell analyses).\n* Automating causal inference from GWAS data in disease research.\n* Providing error-detection, self-reflection, and collaborative reasoning among agents.\n* Offering modular designs for easy integration and extension.\n* Reducing computational cost via small, fine-tuned models plus RAG.\n* Empowering junior researchers with guided workflows and transparent outputs.\n\n1. Mehandru N. et al., **BioAgents: Democratizing Bioinformatics Analysis with Multi-Agent Systems:**\n   [arXiv:2501.06314 (2025).](https://arxiv.org/abs/2501.06314)\n2. **Retrieval-augmented generation for generative artificial intelligence,**\u00a0Nature (2025):\n   <https://www.nature.com/articles/s44401-024-00004-1>\n3. Mehandru N. et al., **BioAgents HTML v1**, arXiv:2501.06314v1 (2025).\n4. Su H. et al., **BioMaster: Multi-agent System for Automated Bioinformatics Analysis Workflow**, bioRxiv (2025):\n   <https://www.biorxiv.org/content/10.1101/2025.01.23.634608v1>\n5. Xu W. et al., **MRAgent: an LLM-based automated agent for causal knowledge discovery in disease via Mendelian randomization**, Briefings in Bioinformatics (2025):\n   <https://doi.org/10.1093/bib/bbaf140>\n6. Wang et al., **An AI Agent for Fully Automated Multi-Omic Analyses**, PMC (2024):\n   <https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11600294/>\n\nBioinformatics workflows often demand deep expertise in both genomics and computational methods, creating barriers for researchers. A new paradigm\u2014**multi-agent systems** built on fine-tuned language models with **retrieval-augmented generation**\u2014is emerging to bridge this gap. BioAgents and related frameworks automate complex tasks, matching expert-level performance while remaining accessible and customizable.\n\nBioAgents, introduced in early 2025, showcases how coordinated, specialized agents can plan, execute, and validate bioinformatics analyses without expensive infrastructure. Similar efforts\u2014MRAgent for causal inference, BioMaster for workflow orchestration, and AutoBA for multi-omic analyses\u2014demonstrate the versatility of this approach. These systems significantly reduce time and expertise required, democratizing advanced bioinformatics for broader scientific communities.", "# 2017 \u2013 The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: papers).*\n\n* [Link to Paper](https://marcdeller.com/wp-content/uploads/2025/09/nihms888398.pdf)\n\n* **Sphere Stabilization:** H161 and Q167 form a hydrogen-bond network that stabilizes bound oxygen and promotes proton transfer.\n* **Motif Variability:** Two motifs (REF and HQY/HQF) dominate second-sphere composition across cellulose-active AA10s, reflecting evolutionary tuning.\n* **Protonation Tuning:** Mutation of conserved glutamine/glutamate shifts reduction potentials by up to 500 mV, altering reduction/reoxidation kinetics.\n* **Substrate Recognition:** Second-sphere histidine in LPMO9G facilitates cellulose binding and glycosidic bond destabilization at ~3 \u00c5 distances.\n* **Peroxygenase Enhancement:** Q219E mutation boosts chitin oxidation by increasing in situ H\u2082O\u2082 generation under reductant-driven conditions.\n* **Solvent Rescue:** Even disruptive alanine substitutions retain partial activity via solvent penetrating the cavity, highlighting sphere adaptability.\n\n1. **The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase:** [Marletta, M.A., et al., *ACS Chem. Biol.*, 2017](https://pubs.acs.org/doi/abs/10.1021/acschembio.7b00016)\n2. **Impact of the Copper Second Coordination Sphere on Catalytic Fine-Tuning of LPMOs:** [Vaaje-Kolstad, G., et al., *ACS Omega*, 2024](https://pubs.acs.org/doi/10.1021/acsomega.4c02666)\n3. **A Conserved Second Sphere Residue Tunes Copper Site Reactivity in LPMOs:** [S\u00f8gaard, D., et al., *J. Am. Chem. Soc.*, 2023](https://pubs.acs.org/doi/10.1021/jacs.3c05342)\n4. **Assessing the Role of Redox Partners in TthLPMO9G and Its Mutants:** [Turner, C., et al., *Front. Microbiol.*, 2024](https://pmc.ncbi.nlm.nih.gov/articles/PMC10835826/)\n\n**Unveiling Hidden Helpers: How the Secondary Coordination Sphere Powers Fungal LPMOs**\n\nCritical hydrogen-bonding networks in fungal polysaccharide monooxygenases (PMOs) that govern oxygen activation and proton transfer\u2014key mechanistic insights for optimizing enzymatic cellulose degradation, with transformative potential for biofuel production and sustainable biomass utilization.\n\nFrom the moment copper binds within a fungal polysaccharide monooxygenase\u2019s active site, a bustling network of neighboring amino acids leap into action\u2014fine-tuning oxygen activation, stabilizing reactive intermediates, and orchestrating efficient degradation of cellulose. Recent studies spanning 2024\u20132025 have cast new light on this once\u2013overlooked ensemble of residues, revealing how second-sphere players such as histidines, glutamines, and tyrosines amplify catalytic performance and broaden substrate selectivity. While the primary histidine brace secures the metal, it is this surrounding hydrogen-bonding web that truly unlocks the enzyme\u2019s oxidative prowess, transforming what was once viewed as structural padding into a dynamic, integral partner in fungal biomass conversion.\n\nWithin the fungal AA9 family, mutational analyses of *Myceliophthora thermophila* PMO3* variants demonstrate that disrupting the H161\u2013Q167 network slashes oxygen binding and proton-transfer efficiency\u2014yet retains partial activity through solvent rescue, underscoring the sphere\u2019s role in stabilizing a fleeting copper-oxyl intermediate. Complementary 2024 work on bacterial AA10 LPMOs reveals that swapping the His\u2013Gln\u2013Tyr motif for Arg\u2013Glu\u2013Phe accelerates oxidative inactivation under turnover, highlighting complex interplay between second-sphere composition and enzyme longevity. Moreover, advanced spectroscopy and DFT calculations on *Neurospora crassa* AA9C variants confirm that protonation states of a conserved glutamine/glutamate tune reduction potentials by hundreds of millivolts, modulating the ratio of reduction to reoxidation rates and controlling radical confinement pathways.\n\nBeyond individual residues, bioinformatic surveys of over 460 predicted cellulose-active AA10s uncover two dominant second-sphere motifs\u2014REF and HQY/HQF\u2014equally represented across diverse bacterial lineages, suggesting evolutionary pressure to balance substrate affinity and oxidative stability. Detailed structural mapping in *Thermothelomyces thermophilus* LPMO9G further pinpoints a second-sphere histidine (H140) lying within 3.3 \u00c5 of the histidine brace, directly influencing cellulose recognition and glycosidic-bond destabilization. As peroxygenase functionality emerges, new mutagenesis of axial glutamine residues demonstrates enhanced chitin activity when fueled by exogenous H\u2082O\u2082, opening doors to tailored biocatalysts for polysaccharide valorization.\n\nCollectively, these advances reveal the secondary coordination sphere not as passive support but as a dynamic regulator\u2014safeguarding reactive species, directing proton flows, and sculpting substrate channels. By harnessing this hidden network, next-generation LPMO engineering can achieve unprecedented efficiency in biomass conversion, sustainable biofuel production, and precision biotransformations.", "# Self-Encoded Libraries: A New Era in Drug Discovery Screening\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-12-07 on marcdeller.com (categories: sb).*\n\n* [Barcode-free hit discovery from massive libraries enabled by automated small molecule structure annotation](https://www.nature.com/articles/s41467-025-65282-1)\n\n* **Synthesis freedom**: Self-encoded libraries circumvent DNA compatibility restrictions, enabling [harsh reaction conditions](https://www.nature.com/articles/s41467-025-65282-1) like cross-couplings and heterocyclizations that would degrade genetic barcodes.\n* **Target expansion**: The barcode-free approach successfully identifies [inhibitors for DNA-processing enzymes](https://www.nature.com/articles/s41467-025-65282-1) like FEN1, which are inaccessible to DNA-encoded library screening due to nucleic acid binding site interference.\n* **Automated decoding**: [COMET software](https://www.nature.com/articles/s41467-025-65282-1) uses predicted fragmentation patterns to filter tandem mass spectrometry data, achieving annotation recall rates of 66-82% while minimizing false positives from isobaric compounds.\n* **Scalable throughput**: Current platforms screen 500,000-member libraries at 100 femtomoles per compound with [potential expansion to five million members](https://www.nature.com/articles/s41467-025-65282-1) within solubility limits, rivaling practical DNA-encoded library scales.\n\n1. **Barcode-free hit discovery from massive libraries enabled by automated small molecule structure annotation**: [V\u00f6geli et al., Nature Communications, 2025](https://www.nature.com/articles/s41467-025-65282-1)\n2. **Advances in Ultrahigh Throughput Hit Discovery with Tandem Mass Spectrometry**: [Quartararo et al., Journal of the American Chemical Society, 2023](https://pubs.acs.org/doi/10.1021/jacs.3c04899)\n3. **A faster way to find new medicines\u2014without the limitations of big DNA barcodes**: [Pomplun et al., Phys.org, 2025](https://phys.org/news/2025-10-faster-medicines-limitations-big-dna.html)\n4. **DNA-Encoded Chemical Libraries: A Comprehensive Review with Successes and Challenges**: [Ottl et al., ACS Pharmacology & Translational Science, 2021](https://pmc.ncbi.nlm.nih.gov/articles/PMC8369695/)\n\n***Mass spectrometry-based decoding enables rapid identification of therapeutic candidates from half-million compound libraries***\n\nDrug discovery has long relied on affinity selection technologies to rapidly screen vast chemical libraries for molecules that bind therapeutic targets. While [DNA-encoded libraries](https://www.nature.com/articles/s41467-025-65282-1) have transformed this landscape by enabling the screening of billions of compounds through unique genetic barcodes, they come with significant limitations. The DNA tags themselves create synthesis complexity, restrict chemical reaction compatibility, and can interfere with binding assays, particularly for targets that naturally interact with nucleic acids like transcription factors and DNA-processing enzymes.\n\nRecent advances in mass spectrometry instrumentation and computational annotation algorithms have now made barcode-free screening a reality. Researchers have developed [self-encoded libraries](https://pubs.acs.org/doi/10.1021/jacs.3c04899) that identify compounds directly through their tandem mass spectrometry fragmentation patterns rather than external tags. This approach combines solid-phase combinatorial synthesis with automated structure annotation software called COMET, which predicts fragmentation patterns and filters mass spectrometry data to accurately identify library members. The platform has successfully screened libraries containing over 500,000 small molecules in single experiments, matching the scale of functional DNA-encoded library selections while avoiding the constraints of genetic barcoding.\n\nThe self-encoded library workflow enables synthesis protocols that would degrade DNA tags, including palladium-catalyzed cross-couplings, heterocyclizations, and nucleophilic aromatic substitutions. Libraries can be synthesized in under a week using standard organic chemistry techniques, democratizing access to large-scale screening beyond well-funded pharmaceutical companies. Validation studies against carbonic anhydrase IX identified [multiple nanomolar binders](https://www.nature.com/articles/s41467-025-65282-1) with enrichment factors comparable to phage display selections. More significantly, the platform succeeded where DNA-encoded approaches cannot, discovering potent inhibitors of flap endonuclease 1, a DNA-binding enzyme whose active site would be compromised by the presence of oligonucleotide tags. The absence of bulky barcodes allows library compounds to explore broader conformational diversity during selection and eliminates false positives caused by DNA-target interactions. With detection limits suggesting scalability to five million compounds and the ability to screen diverse chemical scaffolds simultaneously, this [barcode-free technology](https://phys.org/news/2025-10-faster-medicines-limitations-big-dna.html)represents a transformative step for early-stage therapeutic development.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nCell survival and proliferation was determined colorimetrically according to the modified method of Mossman (Chapdelaine, 1989). This assay is dependent on the reduction of the tetrazolium salt 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyl tetrazolium bromide (MTT) by the mitochondrial dehydrogenase of viable cells to form a blue formazan product (see 3.4 for more detailed discussion). Typically, 10 \u03bcl of 5 mg/ml MTT was added to each well and the plate incubated for a further 4 hours. The formazan product was solubilised by the addition of 10 % (w/v) SDS in 0.01 M HCl to each well. The plate was incubated at 37 \u00b0C for 16 hours and the absorbance measured at 570 nm.\n\n### 3.2.10 Ion-exchange and Gel Filtration Chromatography\n\nSamples of affinity purified OSM_\u0394188 were buffered to various pH's in the range of 3\u201310 and then applied to either Resource-Q, Resource-S, Mono-Q or Mono-S columns (Pharmacia). Samples were typically run at a flow rate of 5 ml/min using a linear gradient of 0 to 2.0 M sodium chloride on a Biosys 510 FPLC or Gold System HPLC (Beckman).\n\nSamples of affinity purified OSM_\u0394188 were applied to a Superdex-75 HiLoad column (bed volume 320 ml; 260 \u00d7 600 mm; separation range 3\u201370 kDa) (Pharmacia). Columns were typically run at a flow rate of 1 ml/min in 150 mM sodium chloride, 50 mM Tris-HCl, pH 7.5 on a Biosys 510 FPLC (Beckman). Columns were calibrated by running a cocktail of protein markers consisting of 0.5 mg each of cytochrome C (12.4 kDa), carbonic anhydrase (29.0 kDa), albumin bovine serum (66.0 kDa), alcohol dehydrogenase (150.0 kDa) and blue dextran (2000.0 kDa).\n\n### 3.2.11 N-terminal Sequencing\n\nN-terminal sequencing was carried out for 10 cycles using an Applied Biosystems 494A/473A Sequencer (Tony Willis, MRC Immunochemistry Unit Protein Sequencing Facility, Oxford). Typically, 230\u2013480 pmol samples of HPLC purified material were submitted and blotted onto PVDF membrane after separation on a 10 % NuPAGE/MOPS or 4\u201312 % NuPAGE/MES gel.\n\n### 3.2.12 Mass Spectrometry\n\nTypically, ~400 pmol samples of HPLC purified OSM_\u0394188 were submitted for mass spectrometry. Analysis was carried out using electrospray ionization on a Micromass BioQ II triple quadrupole, atmospheric pressure mass spectrometer equipped with an electrospray interface operating in the positive ion mode (Dr. Robin Aplin, Oxford Centre for Molecular Sciences, Oxford). Samples were also analysed in Matrix Assisted Laser Desorption and Ionisation (MALDI) mode, using a Kratos Kompact mass spectrometer (Dr. Lisa Wallington, CRC Growth Factor Group, Birmingham).\n\n## 3.3 Results of Expression, Purification and Characterisation of Oncostatin M\n\n### 3.3.1 SDS-PAGE of Protease 3C-His6 Expression and Affinity Purification\n\nFigure 3-3 below shows a 12 % SDS-polyacrylamide gel of the expression and affinity purification of 3C-His6 protease. Lane 1 shows the total cellular protein before induction with IPTG and lane 2 shows the total cellular protein after ~20 hours induction with IPTG. By comparison of the first two lanes a relative increase in the density of bands at ~27 kDa and ~50 kDa can be observed on induction, so suggesting production of 3C-His6. Lane 3 shows the cleared cell lysate, which shows major bands at ~27 kDa and 50 kDa. Lane 4 shows the first eluted fraction of 3C-His6 running at ~27 kDa. Lane 5 shows the nickel agarose resin after passage of the cleared lysate and elution of the fusion protein and no bands are observed. Lane 6 shows the six components of the broad range rainbow molecular weight markers running at 210, 135, 80, 42, 31 and 18 kDa (Bio-Rad).\n\n*Figure 3-3.* Figure showing 12 % SDS-PAGE gel of the expression and affinity purification of protease 3C-His6. Lane 1 and 2 show the total cellular protein before and after induction, respectively. Lane 3 shows the cleared cell lysate and lane 4 shows the eluted protein fractions after cleavage with 3C protease. Lane 5 shows the nickel agarose resin after elution and lane 6 show broad range molecular weight markers running at 210, 135, 80, 42, 31 and 18 kDa (Bio-Rad).\n\n### 3.3.2 Analysis of Fusion Protein Cleavage Using Protease 3C\n\nFigure 3-4 below shows a 12 % SDS polyacrylamide gel used to determine optimal cleavage conditions. Lanes 1 to 8 correspond to reaction mixtures containing OSM-196-3C-GST and various concentrations of 3C-GST protease (as in Table 3-1). Lanes 1 to 4 contain increasing concentrations of 3C-GST protease and all contain no DTT. Lanes 1 to 4 show a main band at ~47 kDa, which corresponds to 3C-GST, and a band at ~50 kDa, which corresponds to OSM-196-GST. Minimal cleavage products are observed in lanes 1 to 4. Lanes 5 to 8 contain increasing concentrations of 3C-GST protease and all contain ~0.5 mM DTT. Lanes 5 to 8 show a main band at ~47 kDa, which corresponds to 3C protease, and a band at ~25 kDa, which corresponds to cleaved OSM-196. Lanes 5 to 8 demonstrate a loss of the OSM-196-GST band at ~50 kDa. Lane 9 contains a 3C-GST control running at ~47 kDa and lane 10 contains OSM-196-GST running at ~50 kDa. Lane 11 contains broad range rainbow molecular weight markers (Bio-Rad).\n\n*Figure 3-4.* Figure showing 12 % SDS polyacrylamide gel used to determine optimal cleavage conditions. Lanes 1 to 8 correspond to reaction mixtures containing OSM-196-3C-GST and various concentrations of 3C-GST. Reactions in lanes 1 to 4 were carried out in the absence of DTT whereas lanes 5 to 8 contained ~0.5 mM DTT. Lane 9 contains 3C-GST and lane 10 contains OSM-196-GST as control. Lane 11 contains broad range molecular weight markers running at 210, 135, 80, 42, 31 and 18 kDa (Bio-Rad).\n\n### 3.3.3 SDS-PAGE of OSM_\u0394188 and OSM_\u0394185 Expression", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nIn an attempt to produce large quantities of LIFR protein for crystallographic studies various mammalian expression systems were investigated. Full length constructs of the extracellular portion of LIFR consisting of the two N-terminal haemopoietin domains, an intervening Ig-like domain and the three C-terminal fibronectin type III (FN) motifs were kindly provided in pEE14 (Bebbington, 1991), pRmHa-3 (Bunch et al., 1988) and pIG-1 (Simmons, 1992) vectors as soluble Fc fusions (LIFR-FN3-3C-Fc; K. Chobotova, K. R. Hudson and J. K. Heath, CRC Growth Factor Group, Birmingham; see Figure 4-2 below). In all cases an Fc purification tag was fused to the C-terminus of the fibronectin type III motif to aid in affinity purification using sepharose immobilised Protein-A (Pharmacia). The Fc tag consisted of the heavy chain constant domains 2 and 3 and the hinge region (hinge-CH2-CH3) of human IgG1. In all cases the intervening linker region between the Fc and the LIFR contained 11 amino acids of the rhinovirus type 14 protease 3C recognition site to facilitate removal of the Fc purification tag (see Chapter 2).\n\nThe mature LIFR-FN3 moiety consists of 829 residues with a theoretical non-glycosylated molecular weight of 93.8 kDa and a theoretical pI of 7.2. All cells were obtained from the European Collection of Cell Cultures (ECACC), unless otherwise stated.\n\n*Figure 4-2.* Diagrammatic representation of a) pEE14 vector (Bebbington, 1991), b) pRmHa-3 vector (Bunch et al., 1988) and c) pIG-1 vector (Simmons, 1992) used for expression of soluble LIFR-FN3-3C-Fc fusion in CHO cells, Drosophila S2 cells and 293T fibroblast cells, respectively. Panel d) shows the diagrammatic domain representation of the translation product (LIFR-FN3-3C-Fc) consisting of the two N-terminal haemopoietin domains of LIFR, an intervening Ig-like domain and the three C-terminal fibronectin type III (FN) domains. An Fc tag is included for Protein-A affinity purification and a 3C protease site is included for cleavage of fusion protein. All constructs kindly provided by K. Chobotova, CRC Growth Factors Group, Birmingham.\n\n### 4.2.1 Stable Expression in CHO Cells\n\nThe mammalian expression vector pEE14 was used (Bebbington, 1991; Cockett et al., 1990; see Figure 4-2) for expression of soluble LIFR-FN3-3C-Fc fusion protein. Expression was driven by the human cytomegalovirus (hCMV) promoter in host Chinese hamster ovary K1 cells (CHO-K1). This vector contains the glutamine synthetase (GS) minigene for amplification and clonal selection. GS is an allosteric enzyme required for the production of glutamine from glutamate and this enzyme is inhibited by L-methionine sulphoximine (MSX) which is a transition state analogue of the reaction:\n\nGlutamate + Ammonia + ATP \u21d4 Glutamine + ADP + Pi\n\nCHO cells transfected with the GS minigene do not require glutamine in the growth media, providing sufficient glutamate is present. However, GS is essential for cell survival when glutamine is absent from the media and full MSX inhibition in this situation is lethal. Therefore, MSX treatment encourages an increase in the copy number of the GS gene when cells are cultured in the absence of glutamine, so effectively co-amplifying any associated genes. pEE14-LIFR-FN3-3C-Fc constructs were kindly provided by K. Chobotova, CRC Growth Factors Group, Birmingham.\n\n#### 4.2.1.1 CHO Cell Culture\n\npEE14-LIFR-FN3-3C-Fc transfected CHO-K1 cells were cultured in GMEM-S medium (Advanced Protein Products Ltd.) supplemented with 10 % dialysed foetal calf serum (FCS; Gibco), 50 IU/ml penicillin, 50 \u03bcg/ml streptomycin and 50 \u03bcM MSX (Sigma) (see Appendix III). Cultures were maintained at 37 \u00b0C in a humidified atmosphere of 5 % CO2 in air. Cultures were initially seeded at 0.5 \u00d7 10^6 cells/ml and maintained by splitting each flask 1 in 5 on alternate days (or when confluent). Cells were harvested by trypsinisation as detailed in 4.2.1.2 below. All operations involving cell-lines were carried out aseptically in a class II microbiological flow cabinet. Cells were routinely tested for mycoplasma using an enzyme-linked immunosorbent assay (ELISA) kit (Boehringer Mannheim).\n\n#### 4.2.1.2 Trypsinisation\n\nGrowth medium was removed by aspiration and the monolayer washed with 10 ml of PBS. The PBS was removed and 2 ml of 0.05 % (w/v) trypsin, 0.02 % (w/v) EDTA solution (Sigma) was added with gentle rocking to ensure even distribution. The flask was stoppered and incubated at 37 \u00b0C for 2 minutes. The cells were dislodged by banging the flask on the bench several times. Fresh medium was added and the resulting cell suspension counted using a haemocytometer. The required number of cells was harvested by centrifugation at 450 \u00d7 g for 5 minutes.\n\n#### 4.2.1.3 Preservation of Cells by Freezing\n\nTypically, 2 \u00d7 10^8 cells were harvested by trypsinisation as in 4.2.1.2 and resuspended in 1.0 ml of dialysed FCS (Gibco) containing 10 % dimethyl sulphoxide (BDH). Aliquots of cell suspension were placed in cryo tubes embedded in polystyrene to ensure slow and even freezing. The tubes were placed at \u221270 \u00b0C overnight and then transferred to liquid nitrogen for long-term storage. Re-establishment of frozen cell-lines was carried out by rapidly thawing aliquots in a 37 \u00b0C water bath followed by layering of the cells onto 10 ml of pre-warmed GMEM-S medium (Advanced Protein Products Ltd.) supplemented with 10 % dialysed FCS (Gibco), 50 IU/ml penicillin, 50 \u03bcg/ml streptomycin and 2 mM L-glutamine. The cells were harvested by centrifugation at 450 \u00d7 g for 5 minutes and then resuspended in 10 ml of GMEM-S as above and transferred to a 75 cm\u00b2 tissue culture flask. When confluent the cells were maintained as above.\n\n#### 4.2.1.4 Protein Expression", "# Protein stability: a crystallographer's perspective\nAuthors: Deller, M. C.; Kong, L.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 72, 72\u201395 (IYCr crystallization series), 2016\n\n| Type of stability | Key question | Answer required for successful crystallization | Investigate if the answer is \"No\" |\n|---|---|---|---|\n| Compositional | Is the chemical makeup of the protein well defined? | Yes | Check amino-acid sequence; check for PTMs, especially proteolysis; purify further |\n| Compositional | Does the protein have a high level of chemical homogeneity? | Yes | Apply more rigorous bioanalytical methods such as mass spectrometry and light scattering |\n| Compositional | Is the protein chemically stable in the crystallization conditions? | Yes | Use customized or less harsh crystallization screens; explore different screening temperatures |\n| Compositional | Is the protein stable over the course of the crystallization experiment? | Yes | As above |\n| Conformational | Are there minimal disordered regions in the protein? | Yes | Redesign expression constructs to engineer out disordered or dynamic regions |\n| Conformational | Does the protein have minimal content of domains or regions with dynamic variability over time? | Yes | Identify stabilizing protein partners or ligands |\n\n*Table 2.* Common measures of protein stability, by structural level, with the relative merits of the three main structure-determination techniques (five asterisks = optimal).\n\n| Structural level | Definition of stability | Example processes/features | Common methods | Crystallography | NMR | EM |\n|---|---|---|---|---|---|---|\n| Primary | Change of amino-acid sequence or modification of amino acids | PTM, proteolysis, protein splicing | Half-life analysis, SDS\u2013PAGE, mass spectrometry, Eastern and Western blotting | ***** | **** | * |\n| Secondary | Change of \u03b1-helix, \u03b2-sheet and loop content | Secondary-structure formation, racemization, aromatic side-chain interactions, ligands | CD, synchrotron-radiation CD, UV-CD, FT-IR, 2D-IR, DXMS | ***** | ***** | * |\n| Tertiary | Change of overall fold or protein conformation | Hydrogen bonding, hydrophobic interactions, conformational change, disulfide bonding, topology | ITC, DSC, Thermofluor | **** | **** | ** |\n| Quaternary | Change in oligomeric state | Protein\u2013protein interactions, oligomerization | Size-exclusion chromatography, native gel electrophoresis | * | ***** | ***** |\n\n*Table 3.* Methods for improving protein stability during expression and purification.\n\n| Category | Approach | Examples | Key references |\n|---|---|---|---|\n| Construct design | Removal of degradation-prone and protease sites | N-end rule, PEST sequences, specific protease sites | Bachmair et al. (1986); Rogers et al. (1986); Spiegel et al. (2015) |\n| Construct design | Optimized codon usage | Synthetic gene synthesis offered by many companies | Daniel et al. (2015); Kane (1995) |\n| Construct design | Truncations and point mutations | Stabilizing mutations for membrane proteins | Gr\u00e4slund et al. (2008); Klock et al. (2008) |\n| Construct design | Chaperones and co-expression | Co-expression and fusion vectors with DnaK and GroEL | Kyratsous & Panagiotidis (2012); Kyratsous et al. (2009) |\n| Construct design | Affinity tags for improved solubility | GST, MBP, Halo tags | Walls & Loughran (2011); Kapust & Waugh (1999) |\n| Expression conditions | Cofactors and ligands | Metal ions and cofactors essential for folding, plus other stabilizing cofactors | Leibly et al. (2012) |\n| Expression conditions | Low-temperature expression | Cold-shock, low-temperature inducible promoters such as pCold | Qing et al. (2004); Vasina & Baneyx (1997) |\n| Host cells | Codon-optimized cells | *E. coli* Rosetta | Tegel et al. (2010) |\n| Host cells | Reduced-toxicity and reduced-protease cells | *E. coli* BL21 Star and pLysS | Studier (1991) |\n\n*Table 4.* Methods for modifying protein stability and improving crystallizability, with the type of stability each addresses.\n\n| Method | Compositional stability | Conformational stability | Example | Key references |\n|---|---|---|---|---|\n| Truncations and domain selection | Yes | Yes | Removal of disordered regions | Yumerefendi et al. (2010); Reich et al. (2006); Klock et al. (2008); Gr\u00e4slund et al. (2008) |\n| Buffer screening | Yes | Yes | Stabilize via ionic changes | Reinhard et al. (2013) |\n| Ligands and additive screening | Yes | Yes | Binding to stabilize protein | Chung (2007); Hassell et al. (2007) |\n| Fused affinity tags and crystallization chaperones | Yes | Yes | Stabilize protein with rigid fusion of tag | Smyth et al. (2003) |\n| Reductive alkylation | Yes | Yes | Provides entropic benefit upon crystallization | Rice et al. (1977); Tan et al. (2014); Walter et al. (2006) |\n| Surface mutagenesis, SER and deglycosylation | Yes | Yes | Entropic stabilization of large, flexible, surface-exposed residues | Cooper et al. (2007); Goldschmidt et al. (2014) |\n| In situ proteolysis | Yes | Yes | Removal of flexible loops | Dong et al. (2007); Wernimont & Edwards (2009) |\n| Thermofluor | No | Yes | Analysis of thermal stability (Tm) | Reinhard et al. (2013); Ristic et al. (2015) |\n| Deuterium-exchange mass spectrometry (DXMS) | No | Yes | Identification of flexible regions | Englander (2006); Englander & Kallenbach (1983) |\n| Disulfide engineering | Yes | Yes | Stabilization of quaternary structure of crystal packing | Forse et al. (2011); Quistgaard (2014) |\n\n*Table 5.* Selected quantitative benchmarks cited in the review.", "# The $40 Billion Revolution: How Bispecific Antibodies Are Transforming Cancer Care\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-13 on marcdeller.com (categories: blog, sb).*\n\n* [The $40 Billion Race: Bispecific Antibodies Are Redefining Cancer Therapy](https://www.forbes.com/sites/williamhaseltine/2025/09/18/the-40-billion-race-bispecific-antibodies-are-redefining-cancer-therapy/)\n\n* **Market Transformation**: Bispecific antibody market projected to reach $40-484 billion by 2034, representing 44% annual growth from current $12-17 billion valuation\u200b\n* **Clinical Breakthrough**: Over 70% response rates achieved in heavily pretreated multiple myeloma patients with complete remission in nearly 50% of cases\u200b\n* **Accessibility Advantage**: Unlike CAR-T therapies requiring weeks of preparation, bispecific antibodies provide immediate \u201coff-the-shelf\u201d treatment access for community cancer centers\u200b\n* **Pipeline Expansion**: More than 250 bispecific candidates under development across 180+ companies, with trispecific antibodies showing 260-fold enhanced tumor-killing potential\u200b\n* **Safety Management**: Cytokine release syndrome controllable through step-up dosing and prophylactic measures, with toxicity profiles more manageable than CAR-T alternatives\n* **Beyond Oncology**: Successful applications in hemophilia, retinal diseases, and expanding autoimmune indications demonstrate platform versatility across medical specialties\n\n1. **Bispecific antibodies: unleashing a new era in oncology treatment:** [Shi L, et al., Cell Death & Disease](https://pmc.ncbi.nlm.nih.gov/articles/PMC12320366/)\n2. **Treating Cancer with Bispecific Antibodies: New FDA Approvals:** [Clark SM, Applied Oncology](https://apponcologysummit.org/treating-cancer-with-bispecific-antibodies-new-fda-approvals/)\n3. **Bispecific Antibodies Market Size & Outlook, 2025-2033:** [Straits Research, Market Analysis Report](https://straitsresearch.com/report/bispecific-antibodies-market)\n4. **Mechanisms of resistance against T-cell engaging bispecific antibodies:** [van de Donk NW, et al., The Lancet Haematology](https://pubmed.ncbi.nlm.nih.gov/39033769/)\n5. **FDA approved bispecific antibodies:** [Evitria Scientific Research](https://www.evitria.com/journal/bispecific-antibodies/fda-approved-bispecific-antibodies/)\n6. **Bispecific Antibodies Market Size And Share Report, 2030:** [Grand View Research, Industry Analysis](https://www.grandviewresearch.com/industry-analysis/bispecific-antibodies-market-report)\n\n**A new class of dual-targeting immunotherapies is reshaping oncology with unprecedented response rates and accessible treatment protocols**\n\nCancer treatment stands at a pivotal transformation point, driven by an extraordinary class of engineered proteins that simultaneously attack tumors on multiple fronts. Bispecific antibodies, once confined to research laboratories, now represent one of medicine\u2019s fastest-growing therapeutic sectors, with market projections reaching $40 billion by 2030. These sophisticated molecules have achieved what traditional treatments often cannot: response rates exceeding 70% in heavily pretreated patients who had exhausted all other options.\n\nUnlike conventional monoclonal antibodies that target a single protein, bispecific antibodies operate through dual engagement mechanisms. One binding arm attaches to tumor-associated antigens while the other recruits immune effector cells, creating a direct bridge between cancer cells and the body\u2019s natural defense systems. This revolutionary approach transforms resting T-cells into aggressive tumor killers at the precise location where they are needed most.\n\nThe clinical impact has been remarkable. In multiple myeloma, where patients typically face limited options after relapse, bispecific antibodies like teclistamab, elranatamab, and talquetamab have demonstrated overall response rates between 60-70% in heavily pretreated populations. More striking still, nearly half of these patients achieve complete remission. Similar success stories emerge across hematologic malignancies, with agents like blinatumomab revolutionizing acute lymphoblastic leukemia treatment and newer bispecifics transforming diffuse large B-cell lymphoma outcomes.\n\nThe expansion into solid tumors marks another breakthrough frontier. Five bispecific antibodies now carry FDA approval for solid malignancies, including amivantamab for non-small cell lung cancer, zanidatamab for HER2-positive biliary tract cancer, and zenocutuzumab for NRG1 fusion-positive tumors. These approvals represent significant victories against historically challenging cancers, with zenocutuzumab achieving 42% response rates in heavily pretreated pancreatic adenocarcinoma patients.\n\nManufacturing and accessibility advantages distinguish bispecific antibodies from personalized cellular therapies. While CAR-T treatments require weeks of patient-specific cell engineering and specialized treatment centers, bispecific antibodies arrive ready-to-use from the pharmacy shelf. This \u201coff-the-shelf\u201d availability democratizes access to advanced immunotherapy, particularly benefiting community oncology practices where 85% of cancer patients receive care.\n\nHowever, this therapeutic power requires careful management. Cytokine release syndrome represents the primary safety concern, as activated immune systems can trigger systemic inflammatory responses. Clinical experience has refined management protocols, with step-up dosing regimens and prophylactic measures significantly reducing severe complications. Importantly, these toxicities remain generally more manageable than those associated with CAR-T therapies.", "# ChemSage: Vibe-Coding a Chemically Aware Drug-Discovery LLM on Local Hardware\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nChemSage went through five training rounds. Round 1 was a 7B proof of concept; Round 2 moved to a 32B base; Rounds 3 to 5 grew the dataset and the behavior coverage, with Round 5 doubling the LoRA rank for more adapter capacity. The five fine-tuned models are published on [Hugging Face](https://huggingface.co/Dellboy). Full stats are available from the HTML output of [eval_compare.py](https://marcdeller.com/ChemSage/chemsage_evaluation.html).\n\n| Round | Base model | Trainable params | Train examples | Behaviour classes |\n| --- | --- | --- | --- | --- |\n| 1 | Qwen2.5-7B-Instruct-4bit | 26 M (0.34%) | 1,200 | 8 |\n| 2 | Qwen2.5-32B-Instruct-4bit | 134 M (0.41%) | 1,200 | 8 |\n| 3 | Qwen2.5-32B-Instruct-4bit | 134 M (0.41%) | 4,000 | 19 |\n| 4 | Qwen2.5-32B-Instruct-4bit | 134 M (0.41%) | 6,400 | 59 |\n| 5 | Qwen2.5-32B-Instruct-4bit | 268 M (0.82%) | 16,000 | 78 |\n\n**Table 1. Model lineage. Each 32B round trains less than 1% of the parameters via QLoRA on a 4-bit base.**\n\nValidation loss tells the training story. Better data, not more compute, drove the gains: Round 3\u2019s RDKit-validated, code-heavy examples cut the best validation loss to 0.054, an order of magnitude below the early rounds, and Round 4 reached 0.041. Round 5 settled slightly higher at 0.055, which is the expected cost of a much harder task (78 behavior classes against 59). Every model fits and trains inside the unified memory of a single 64 GB machine.\n\n| Round | Best val loss | Iters (actual) | Peak train memory | Fused size |\n| --- | --- | --- | --- | --- |\n| 1 | 0.389 @ iter 675 | 750 | 12.0 GB | 4.2 GB |\n| 2 | 0.347 @ iter 675 | 750 | 26.1 GB | 17.0 GB |\n| 3 | 0.054 @ iter 625 | 625 | 25.2 GB | 17.0 GB |\n| 4 | 0.041 @ iter 950 | 1,500 | 25.5 GB | 17.2 GB |\n| 5 | 0.055 @ iter 1,600 | 2,000 | 30.3 GB | 17.2 GB |\n\n**Table 2. Training outcomes. Rounds 3 and 5 stopped early once validation loss plateaued.**\n\nGeneric LLM benchmarks miss what matters here, so ChemSage is scored with a chemistry-aware [evaluation harness](https://marcdeller.com/ChemSage/chemsage_evaluation.html) that checks whether emitted SMILES parse, whether tool code actually runs, and whether stated property values match an RDKit recompute. The comparison below runs all five rounds on the same 100 held-out test examples, scoring each example (an example passes a metric only if every instance in the response is correct). Round 5 leads at 95% overall, with the largest gains on the things that matter most: tool executability, numerical fidelity, and code-then-quote transcription.\n\n| Metric (pass rate) | R1 | R2 | R3 | R4 | R5 | \u0394 vs R1 |\n| --- | --- | --- | --- | --- | --- | --- |\n| SMILES validity | 100% | 100% | 100% | 100% | 100% | +0% |\n| SMARTS validity | n/a | 0% | 100% | 55% | 100% | n/a |\n| Tool executability | 0% | 28% | 79% | 71% | 95% | +95% |\n| Code attempted | 14% | 41% | 100% | 97% | 100% | +86% |\n| Python exec (extended) | 36% | 34% | 78% | 69% | 99% | +63% |\n| Code-to-quote fidelity | n/a | 0% | 47% | 19% | 61% | n/a |\n| Numerical fidelity | n/a | 18% | 57% | 47% | 89% | n/a |\n| Rounding precision | 100% | 100% | 98% | 98% | 99% | -1% |\n| Refusal accuracy | 98% | 98% | 97% | 98% | 100% | +2% |\n| QED range | 100% | 100% | 100% | 100% | 100% | +0% |\n| PDB ID validity | 100% | 100% | 100% | 100% | 100% | +0% |\n| PyMOL syntax | 77% | 97% | 89% | 89% | 90% | +13% |\n| Degeneration-free | 93% | 96% | 91% | 98% | 100% | +7% |\n| Overall | 72% | 63% | 87% | 80% | 95% | +23% |\n\n**Table 3. Comparative evaluation across all 13 metrics, per-example pass rate over 100 shared test examples. The \u0394 column is Round 5 against the Round 1 baseline. Metrics show n/a where Round 1 never triggered that check, so no baseline exists.**\n\n###### Is the AI bubble coming?\n\nHere is the part that nags at me. The economics of hosted frontier models do not obviously add up. The subscription you pay and the compute you consume are increasingly different numbers. A recent Claude Code session of mine is a clear illustration: a $\u00a020-per-month plan\u00a0against roughly $57 in metered API-equivalent usage consumed in about 4 hours of work.\n\nA single local coding session: a $20 per month subscription tier set against $57.18 of metered API-equivalent usage in just over 4 hours.\n\nMultiply that across millions of users, and the question writes itself. When the headline subscription sits well below the marginal cost of serving heavy users, the gap is being filled by investment rather than revenue, and that is the shape of a bubble. None of this means the frontier labs vanish: the hardest reasoning, the largest contexts, and the newest capabilities will keep living in big data centers, and for plenty of work, that is exactly where they belong. The more interesting shift is at the other end. A great deal of routine, sensitive, or latency-critical work does not need a frontier model at all, and it has good reasons to come home: cost you control, data that never leaves the building, and no rate limit between you and your own files.\n\nThat is the quiet argument ChemSage is really making. A capable 32B model that trains, serves, and reasons on a desktop, grounded in your own corpus and checked by your own tools, is no longer exotic. With Apple now putting neural accelerators into consumer chips, the on-device floor keeps rising. The future is unlikely to be all-cloud or all-local. It looks more like a split: frontier models in the cloud for the genuinely hard problems, and a growing tier of specialist, device-embedded models doing the daily work, privately and at the cost of\u00a0electricity.\n\n###### How to run\u2026\n\nRun these 2 commands in separate terminals, first start the model server, then run chat.py.\n\n**Note:** the \u2013model flag on chat.py sets the model name shown in /info but the server is what actually runs inference \u2014 make sure the server is fully loaded (you\u2019ll see \u201cHTTP server listening\u201d) before starting chat. The server defaults to port 8080 so no extra flags needed on the chat side.\n\n```", "# ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-07-01 on marcdeller.com (categories: vibe-code).*\n\n> **A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behaviour, and live RDKit/PyMOL tool calls for chemical truth. Fine-tuned and served locally on Apple Silicon with MLX-LM.**\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [chemsage.mdeller.com](https://chemsage.mdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/ChemSage](https://github.com/bellcheddar/ChemSage) |\n\n---\n\nChemSage is a locally hosted, open-source assistant that reasons about small molecules (SMILES, scaffolds, properties), drives the cheminformatics and structural tools a medicinal chemist actually uses (RDKit, PyMOL, docking, PLIP), and grounds every factual claim in your own corpus.\n\n**Why it matters:** off-the-shelf models invent chemistry (wrong logP, malformed SMILES) and cannot see your private SAR data. ChemSage fixes both by separating concerns: volatile knowledge lives in retrieval, stable behaviour lives in the weights, and deterministic chemical facts are computed live by trusted tools rather than guessed. It is useful for: interrogating your own assay and SAR data, triaging compounds, generating correct PyMOL/RDKit scripts, and interpreting docking and interaction output, all on local hardware. A live, no-install terminal session against the real fine-tuned model runs at **[chemsage.mdeller.com](https://chemsage.mdeller.com)** (see **Web deployment** below).\n\n**Initial idea:** Fed up with token, model and upload document limits? Then why not run your own AI models on Mac hardware? Thanks to [@ai.christianson](https://www.tiktok.com/@ai.christianson) on TikTok for highlighting the following method to implement MLX-LM optimised fine-tuning of a local LLM on Mac hardware: https://www.tiktok.com/@ai.christianson/video/7538150476467014943?lang=en\n\n**Disclaimer:** I am not a professional coder! This is an exercise in Vibe Coding as a Structural Biologist to get the job done. Use at your own risk, and double check the science. I am still working on validation of the Chemistry and Biology. This is a proof of concept exercise to demonstrate what can be achieved using moderate hardware and limited coding experience.\n\n**Project links:**\n\n* [GitHub](https://github.com/bellcheddar/ChemSage)\n* [My blog on the making of ChemSage](https://marcdeller.com/chemsage)\n* [Model data and evaluation stats](https://marcdeller.com/chemsage_evaluation)\n* [Models on Hugging Face](https://huggingface.co/Dellboy)\n\n## \ud83e\udd17 Models on HuggingFace Hub\n\n| Model | HuggingFace | Base | Round |\n| --- | --- | --- | --- |\n| ChemSage 7B v1 | [Dellboy/chem_sage_7b_v1](https://huggingface.co/Dellboy/chem_sage_7b_v1) | Qwen2.5-7B-Instruct-4bit | Round 1 |\n| ChemSage 32B v2 | [Dellboy/chem_sage_32b_v2](https://huggingface.co/Dellboy/chem_sage_32b_v2) | Qwen2.5-32B-Instruct-4bit | Round 2 |\n| ChemSage 32B v3 | [Dellboy/chem_sage_32b_v3](https://huggingface.co/Dellboy/chem_sage_32b_v3) | Qwen2.5-32B-Instruct-4bit | Round 3 |\n| ChemSage 32B v4 | [Dellboy/chem_sage_32b_v4](https://huggingface.co/Dellboy/chem_sage_32b_v4) | Qwen2.5-32B-Instruct-4bit | Round 4 |\n| **ChemSage 32B v5** \u2b50 | [**Dellboy/chem_sage_32b_v5**](https://huggingface.co/Dellboy/chem_sage_32b_v5) | Qwen2.5-32B-Instruct-4bit | **Round 5 (current)** |\n\n## \u2699\ufe0f The hybrid in one line\n\nRAG keeps it truthful today; the fine-tune makes it consistent tomorrow; RDKit makes it correct always.\n\n## \ud83d\ude80 Quick start\n\nSee **[`PROJECT_PLAN.md`](PROJECT_PLAN.md)** for the full step-by-step build (phases 0 to 8). Short version:\n\n| Step | What | Script / command |\n| --- | --- | --- |\n| 0 | Env: MLX-LM + RDKit + GUI | (`requirements.txt`) |\n| 1 | Base model (Qwen 2.5 32B, MLX 4-bit) | `config/train_config.yaml` |\n| 2 | **RAG first** (ship this alone) | `scripts/ingest_rag.py` |\n| 2a | Download ChEMBL corpus | `scripts/download_chembl.py` + `scripts/download_chembl_extended.py` |\n| 2b | Download PubChem corpus | `scripts/download_pubchem.py` |\n| 2c | Download PDB corpus | `scripts/download_pdb.py` + `scripts/download_pdb_supplement.py` |\n| 3 | Build + validate SFT dataset | `scripts/build_dataset.py` |\n| 4 | QLoRA fine-tune | `mlx_lm.lora --config config/train_config.yaml --train` |\n| 5 | Fuse + serve | `mlx_lm.fuse` \u2192 `mlx_lm.server --port 8081` |\n| 6 | Close the hybrid loop | `rag/tool_exec.py` |\n| 7 | Chemistry-aware evaluation | `eval/eval_chem.py` |\n\n## \u25b6\ufe0f How to run\n\n**Terminal 1 \u2014 start the model server:**\n\n```\ncd /Users/dellboy/Documents/Vibe_Coding/chem_sage\nsource .venv/bin/activate\nmlx_lm.server --model models/chem_sage_32b_v5 --port 8080\n```\n\n**Terminal 2 \u2014 run chat:**\n\n```\ncd /Users/dellboy/Documents/Vibe_Coding/chem_sage\nsource .venv/bin/activate\npython scripts/chat.py --model models/chem_sage_32b_v5\n```\n\nWait for \u201cHTTP server listening\u201d in Terminal 1 before starting Terminal 2.\n\n## \ud83c\udf93 Training\n\n### \ud83c\udf93 Training data\n\n|  | Round 1 (7B) | Round 2 (32B) | Round 3 (32B) | Round 4 (32B) | Round 5 (32B) |\n| --- | --- | --- | --- | --- | --- |\n| Total examples | 1,500 | 1,500 | 5,000 | 8,000 | 20,000 |\n| Train / valid / test | 1,200 / 150 / 150 | 1,200 / 150 / 150 | 4,000 / 500 / 500 | 6,400 / 800 / 800 | 16,000 / 2,000 / 2,000 |\n| Behaviour classes | 8 | 8 | 19 | 59 | 78 |\n| Seed SMILES | 25 | 25 | 51 + corpus drugs | 51 + corpus drugs | 51 + corpus drugs |\n| Corpus drug examples | \u2014 | \u2014 | ~1,000 | ~1,000 + protein family PDBs | ~1,200 + protein family PDBs |\n| Code-block density | Low (~50%) | Low (~50%) | High (~80%) | High (~85%) | High (~90%) |\n| Generator script | `--n 1500` | `--n 1500` | `--n 5000 --seed 43` | `--n 8000 --seed 47` | `--n 20000 --seed 51` |\n| Validation | RDKit + SMILES | RDKit + SMILES | RDKit + SMILES | RDKit + SMILES | RDKit + SMILES |\n| Token lengths (p50/p99/max) | \u2014 | \u2014 | \u2014 | \u2014 | 322 / 488 / 558 |\n| Rejected examples | \u2014 | \u2014 | 243 (95%) | \u2014 | 707 (96.5%) |", "# Flow Cytometry and Microplastic Detection in Human Tissues\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-11-04 on marcdeller.com (categories: blog, microplastics).*\n\n* [Microplastics and nanoplastics detection using flow cytometry: Challenges and methodological advances with fluorescent dye application](https://www.sciencedirect.com/science/article/pii/S2215016125000482)\n* [Fast-screening flow cytometry method for detecting nanoplastics in human peripheral blood](https://pmc.ncbi.nlm.nih.gov/articles/PMC9958479/)\n\n* **High-Throughput Analysis:** Flow cytometry analyzes thousands of microplastic particles per minute, [dramatically faster than spectroscopic methods](https://pmc.ncbi.nlm.nih.gov/articles/PMC11870216/) that require hours for complete sample analysis.\u200b\n* **Enhanced Sensitivity:** The technique detects [particles as small as 200 nanometers](https://vbn.aau.dk/en/publications/preliminary-results-from-detection-of-microplastics-in-liquid-sam), far below the [20-micrometer limitation of FTIR spectroscopy](https://pmc.ncbi.nlm.nih.gov/articles/PMC10035019/), capturing the nanoplastic fraction missed by traditional methods.\u200b\n* **Automated Quantification:** [Replacing subjective visual counting with statistical analysis](https://vbn.aau.dk/en/publications/preliminary-results-from-detection-of-microplastics-in-liquid-sam) eliminates human bias and provides reproducible measurements across different laboratories and operators.\u200b\n* **Human Tissue Applications:** Researchers have successfully adapted flow cytometry for [detecting microplastics in blood, liver, placenta, and other human tissues](https://www.eurekalert.org/news-releases/478209) following optimized digestion and extraction protocols.\n* **Technical Challenges Remain:** [Nile red dye aggregation in aqueous solutions](https://pmc.ncbi.nlm.nih.gov/articles/PMC11870216/) creates false positives that require careful data analysis, while polymer identification remains less specific than spectroscopic alternatives.\u200b\n\n1. **Separation and flow cytometry analysis of microplastics and nanoplastics**: [Li, J., et al., Frontiers in Chemistry, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC10540628/)\n2. **A rapid-screening approach to detect and quantify microplastics based on fluorescent tagging with Nile Red**: [Maes, T., et al., Scientific Reports, 2017](https://www.nature.com/articles/srep44501)\n3. **Detection of Sub-20 \u03bcm Microplastic Particles by Attenuated Total Reflection Fourier Transform Infrared Spectroscopy and Comparison with Raman Spectroscopy**: [Othman, A.M., et al., ACS Omega, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC10035019/)\n4. **Fast-screening flow cytometry method for detecting nanoplastics in human peripheral blood**: [Salvia, R., et al., MethodsX, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC9958479/)\n5. **Can flow cytometry emerge as a high-throughput technique for micro- and nanoplastics analysis in complex environmental aqueous matrices?**: [Liu, Y., et al., Bioresource Technology Reports, 2025](https://www.sciencedirect.com/science/article/pii/S2589014X25001525)\n\n**A rapid, automated approach transforms our ability to quantify plastic contamination where it matters most**\n\nMicroplastic contamination has emerged as one of the pressing environmental health concerns of our time, with these tiny polymer fragments now detected throughout the human body. Researchers have found [microplastics in placentas](https://hscnews.unm.edu/news/hsc-newsroom-post-microplastics), [cirrhotic liver tissue](https://www.thelancet.com/journals/ebiom/article/PIIS2352-3964(22)00328-0/fulltext), [peripheral blood](https://pmc.ncbi.nlm.nih.gov/articles/PMC9958479/), and [multiple organs including brain, kidney, and thyroid](https://pmc.ncbi.nlm.nih.gov/articles/PMC12454625/). Traditional methods for detecting these contaminants rely on spectroscopic techniques that require hours of analysis time and struggle to identify the smallest particles. Flow cytometry, initially developed for biological applications, has recently been [adapted for environmental analysis](https://www.sciencedirect.com/science/article/pii/S2215016125000482) and now offers a faster, more sensitive alternative that could transform how researchers study microplastic exposure in human populations.\u200b\n\nThe core challenge in microplastic detection has always been the same issue that makes these materials so persistent in the environment. Plastics resist degradation, lack inherent color in many forms, and exist across a size range from millimeters down to nanometers. [Traditional vibrational spectroscopy methods](https://pmc.ncbi.nlm.nih.gov/articles/PMC9521859/) like Fourier transform infrared spectroscopy can only reliably identify particles larger than 20 micrometers, while [Raman spectroscopy struggles with fluorescence interference](https://pmc.ncbi.nlm.nih.gov/articles/PMC10035019/) and requires spatial resolution around 1 micrometer. Both approaches demand [extensive analysis time](https://pmc.ncbi.nlm.nih.gov/articles/PMC8915801/), with complete filter analyses taking hours or even days. For researchers attempting to process large numbers of tissue samples or track exposure patterns across populations, these limitations represent serious obstacles.\u200b", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\nThe terms \"individual\" or \"patient\" or \"subject\", used interchangeably, refer to any animal, including mammals, preferably mice, rats, other rodents, rabbits, dogs, cats, swine, cattle, sheep, horses, or primates, and most preferably humans. The phrase \"therapeutically effective amount\" refers to the amount of active antibody or pharmaceutical agent that elicits the biological or medicinal response in a tissue, system, animal, individual or human that is being sought by a researcher, veterinarian, medical doctor or other clinician. As used herein, the term \"treating\" or \"treatment\" refers to one or more of (1) inhibiting the disease, e.g., inhibiting a disease, condition or disorder in an individual who is experiencing or displaying the pathology or symptomatology of the disease, condition or disorder (i.e., arresting further development of the pathology and/or symptomatology); and (2) ameliorating the disease, e.g., ameliorating a disease, condition or disorder in an individual who is experiencing or displaying the pathology or symptomatology of the disease, condition or disorder (i.e., reversing the pathology and/or symptomatology) such as decreasing the severity of disease. In some embodiments, the antibodies of the invention are useful in preventing or reducing the risk of developing any of the diseases referred to herein.\n\n### Pharmaceutical Compositions\n\nAn anti-mutCALR antibody described herein can be formulated as a pharmaceutical composition for administration to a subject, e.g., to treat a disease or disorder described herein. Typically, a pharmaceutical composition includes a pharmaceutically acceptable carrier. As used herein, \"pharmaceutically acceptable carrier\" includes any and all solvents, dispersion media, coatings, antibacterial and antifungal agents, isotonic and absorption delaying agents, and the like that are physiologically compatible. The composition can include a pharmaceutically acceptable salt, e.g., an acid addition salt or a base addition salt.\n\nThe pharmaceutical compositions may be in a variety of forms. These include, for example, liquid, semi-solid and solid dosage forms, such as liquid solutions (e.g., injectable and infusible solutions), dispersions or suspensions, tablets, pills, powders, liposomes and suppositories. Typically compositions for the agents described herein are in the form of injectable or infusible solutions.\n\nThe composition can be formulated as a solution, microemulsion, dispersion, liposome, or other ordered structure suitable for stable storage at high concentration or as a lyophilized preparation. Sterile injectable solutions can be prepared by incorporating an anti-mutCALR antibody described herein in the required amount in an appropriate solvent with one or a combination of ingredients enumerated above, as required, followed by filtered sterilization. In the case of sterile powders for the preparation of sterile injectable solutions, the preferred methods of preparation are vacuum drying and freeze drying. Prolonged absorption of injectable compositions can be brought about by including in the composition an agent that delays absorption, for example, monostearate salts. The anti-mutCALR antibodies can also be formulated as liposomes.\n\nPharmaceutical compositions formulated for subcutaneous administration may be suitable in some circumstances because the subject can self-administer the pharmaceutical composition. Pharmaceutical formulations for subcutaneous administration can further comprise protein formulations comprising arginine-HCl, histidine, and/or polysorbate, which may confer increased potency, improved serum half-life, or enhanced solubility to the anti-mutCALR antibodies.\n\n### Administration\n\nThe anti-mutCALR antibody can be administered to a subject, e.g., a subject in need thereof, for example, a human subject, by a variety of methods. For many applications, the route of administration can be intravenous injection or infusion (IV), subcutaneous injection (SC), intraperitoneally (IP), or intramuscular injection.\n\nThe anti-mutCALR antibody can be administered as a fixed dose, or in a mg/kg patient weight dose. The dose can also be chosen to reduce or avoid production of antibodies against the anti-mutCALR antibody. Dosage regimens are adjusted to provide the desired response, e.g., a therapeutic response or a combinatorial therapeutic effect. Dosage unit form or \"fixed dose\" or \"flat dose\" as used herein refers to physically discrete units suited as unitary dosages for the subjects to be treated; each unit contains a predetermined quantity of active compound calculated to produce the desired therapeutic effect in association with the required pharmaceutical carrier and optionally in association with the other agent. Single or multiple dosages may be given. Alternatively, or in addition, the antibody may be administered via continuous infusion.\n\nThe disclosure also provides a kit comprising one or more containers of an anti-mutCALR antibody or a pharmaceutical formulation thereof, optionally with one or more other prophylactic or therapeutic agents useful for the treatment of a disease or disorder, and optionally with instructions for using the anti-mutCALR antibody or a pharmaceutical formulation thereof. The instructions relating to the use of an anti-mutCALR antibody generally include information as to dosage, dosing schedule, and route of administration for the intended treatment. The containers can be unit doses, bulk packages (e.g., multi-dose packages) or sub-unit doses. The label package insert indicates that an anti-mutCALR antibody is used for treating, delaying the onset, and/or alleviating a myeloproliferative neoplasm.\n\n### Detection and Diagnosis", "# Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\nAuthors: Hummel, J.R.; Xiao, K.-J.; Yang, J.C.; Epling, L.B.; Mukai, K.; Ye, Q.; Xu, M.; Qian, D.; Huo, L.; Weber, M.; Roman, V.; Lo, Y.; Drake, K.; Stump, K.; Covington, M.; Kapilashrami, K.; Zhang, G.; Ye, M.; Diamond, S.; Yeleswaram, S.; Macarron, R.; Deller, M.C.; Wee, S.; Kim, S.; Wang, X.; Wu, L.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2024\n\nBased on structural diversity and balanced in vitro profiles, compounds 12, 14, and 17 were advanced to pharmacokinetic (PK) studies in both rodents and cynomolgus monkeys (Table 3). The rat PK profile of compound 12 showed good oral exposure (3498 nM\u00b7h at 3 mg/kg) and bioavailability (69%), moderate clearance [36% hepatic blood flow (HBF)], and a half-life (t1/2) of 3.8 h (entry 1). Upon progressing to cyno PK, compound 12 showed moderate exposure (923 nM\u00b7h at 1.5 mg/kg) and a prolonged IV half-life of 6 h (entry 2). On the other hand, compound 14 showed improved oral exposure and relatively lower clearance in both species (entries 3 and 4). While comparable half-life was observed in rat and cyno PK, species differences in clearance and volume of distribution (Vdss) gave rise to the observed similarity. Despite comparable oral exposure (AUC) and maximum observed plasma concentration (Cmax) across species, in cyno PK compound 14 showed a lower bioavailability (8%) that coincided with a decrease in clearance (<5% HBF). In contrast to 12 and 14, compound 17 showed only moderate oral exposure in rats (AUC 751 nM\u00b7h at 3 mg/kg) that we hypothesize resulted from high clearance (130% HBF) (entry 5). Nonetheless, the promising in vitro profile of 17 prompted us to advance this compound to cyno PK. At an oral dose of 2 mg/kg compound 17 exhibited good exposure (AUC of 5652 nM\u00b7h) and low clearance (9% hepatic blood flow) in cynos (entry 6), and displayed improved oral bioavailability (21%) and an IV half-life of 2.6 h. Based on a balanced cyno PK profile, compound 17 was selected as our lead candidate for further in vivo evaluation.\n\n*Table 3.* Cross-Species Pharmacokinetic Profiles of 12, 14, and 17.\n\n| entry | cmpd | species | PO AUC (nM\u00b7h) | PO Cmax (nM) | F (%) | IV HBF (%) | IV t1/2 (h) | IV Vdss (L/kg) |\n| --- | --- | --- | --- | --- | --- | --- | --- | --- |\n| 1 | 12 | rat | 3498 | 412 | 69 | 36 | 3.8 | 5.7 |\n| 2 | 12 | cyno | 923 | 93 | 20 | 26 | 6 | 4.4 |\n| 3 | 14 | rat | 6950 | 813 | 48 | 12 | 3.3 | 1.4 |\n| 4 | 14 | cyno | 7590 | 778 | 8 | 4.3 | 3.2 | 0.37 |\n| 5 | 17 | rat | 751 | 757 | 51 | 130 | 2 | 1.8 |\n| 6 | 17 | cyno | 5652 | 962 | 21 | 9 | 2.6 | 0.4 |\n\nFootnotes to Table 3: compound 12 in rat and compound 17 in rat were dosed in cassette studies at 0.5 mg/kg IV and 3 mg/kg PO; compound 12 in cyno was dosed in cassette studies at 0.5 mg/kg IV and 1.5 mg/kg PO; compound 14 in rat was dosed in cassette studies at 0.43 mg/kg IV and 3 mg/kg PO; compound 14 in cyno was dosed in discrete studies at 0.5 mg/kg IV and 3 mg/kg PO; compound 17 in cyno was dosed in discrete studies at 0.3 mg/kg IV and 2 mg/kg PO.\n\n### In vivo pharmacodynamics and efficacy\n\nTo assess the in vivo effects of compound 17 on CDK2 signaling, we next conducted a pharmacokinetic/pharmacodynamic (PK/PD) study measuring Rb phosphorylation (pRb) as a pharmacodynamic biomarker (Figure 3). Measuring CDK2 activity through pRb inhibition requires cells to undergo G1 to S phase transition. To capture sufficient cell cycle progression in unsynchronized OVCAR3 cells in vivo, we opted for a 5 day PK/PD study design. In a 5 day PK/PD study, once daily treatment of 17 at 30 mg/kg showed approximately 50% reduction in Rb phosphorylation over 24 h (Figure 3a). PK measurements at 1, 4, and 8 h showed plasma concentrations of compound 17 exceeding or approaching WB IC50. However, by the 24 h time point, plasma levels of 17 dropped below quantitation limit (BQL), suggesting fast clearance in mouse.\n\nWe next evaluated compound 17 in an efficacy study in CCNE1-amplified OVCAR3 tumor bearing mice. Consistent with the decrease in pRb as a PD biomarker, compound 17 displayed antitumor activity in OVCAR3 tumor bearing mice with 38% tumor-growth inhibition (TGI) following once daily oral treatment of 17 at 30 mg/kg for 28 days after inoculation (Figure 3b). Compound 17 demonstrated a favorable safety profile at 30 mg/kg. Comprehensive twice-daily monitoring revealed no unexpected deaths, significant adverse effects, or body weight losses.\n\n*Figure 3.* (a) 5 day PK/PD analysis for compound 17 measuring pRb levels. (b) In vivo tumor growth inhibition in immunocompromised (nu/nu) female mice bearing CCNE1-amplified OVCAR3 tumors.\n\n### Off-target profiling\n\nHaving demonstrated pharmacodynamic inhibition of CDK2 and in vivo efficacy, additional in vitro characterization was carried out for 17 to assess the potential risk for off-target toxicity. Compound 17 exhibited remarkable general kinase selectivity, and in a panel of over 50 kinases, not a single kinase was inhibited at an IC50 value below 3 \u03bcM at 1 mM ATP concentration (Supporting Information, Table S2). In addition, compound 17 was not an inhibitor of cytochrome P450 (CYP) and showed >25 \u03bcM inhibition against multiple CYP isoforms (2D6, 3A4, 2C9, 2C8, 2C19, 2B6, and 1A2) (Supporting Information, Table S3). Preincubation of 17 with CYP isoforms 3A4, 2D6, 2C9, and 2C19 at 10 \u03bcM compound concentration showed no significant time-dependent inhibition (TDI) (Supporting Information, Table S4). Together, these data suggest minimal risk for potential drug\u2212drug interactions (DDIs). Low inhibition in hERG patch clamp assay was also observed (6.6% at 10 \u03bcM), indicating a reduced likelihood of adverse cardiac events.\n\n### Crystal structure of compound 17 bound to CDK2", "# Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly\nAuthors: Deller, M.C.; Johnson, H.A.; Miller, M.D.; Spraggon, G.; Elsliger, M.-A.; Wilson, I.A.; Lesley, S.A.\nSource: PLoS ONE, 2015\n\npMH4-TM1088 was constructed by amplification of the TM1088 locus using Polymerase Chain Reaction (PCR) from T. maritima MSB8 (DSM3109) genomic DNA using PfuTurbo DNA polymerase (Stratagene) and primers corresponding to the predicted 5' (5'-CTGTACTTCCAGGGCATGTCAAAAAAACAAAAAAGC-3') and 3' (5'-AATTAAGTCGCGTTATCAACTATTCACACCTCC-3') ends. The PCR product was cloned into plasmid pMH4, which encodes an expression and purification tag (MGSDKIHHHHHH) at the N-terminus of the full-length TM1088A protein. The expression and purification tag was not cleaved from the final product. The cloning junctions were confirmed by DNA sequencing. Expression was performed in a modified Terrific Broth (TB) using the E. coli strain GeneHogs (Invitrogen). At the end of fermentation, lysozyme was added to the culture to a final concentration of 250 \u03bcg/ml, and the cells were harvested and frozen. After one freeze/thaw cycle, the cells were sonicated in lysis buffer (50 mM Tris pH 7.9, 50 mM NaCl, 10 mM imidazole, 1 mM Tris(2-carboxyethyl)phosphine-HCl (TCEP)) and the lysate was clarified by centrifugation at 32,500 x g for 30 minutes. The soluble fraction was passed over nickel-chelating resin (GE Healthcare) pre-equilibrated with lysis buffer, the resin washed with wash buffer (50 mM Tris pH 7.9, 300 mM NaCl, 40 mM imidazole, 10% (v/v) glycerol, 1 mM TCEP), and the protein was eluted with elution buffer (20 mM Tris pH 7.9, 300 mM imidazole, 10% (v/v) glycerol, 1 mM TCEP). The eluate was diluted 10-fold with buffer Q (20 mM Tris pH 7.9, 5% (v/v) glycerol, 0.25 mM TCEP) containing 50 mM NaCl and loaded onto a RESOURCE Q column (GE Healthcare) pre-equilibrated with the same buffer. Both TM1088A and the TM1088 complex were eluted with a linear gradient of 50\u2013500 mM NaCl in buffer Q, buffer exchanged with crystallization buffer (20 mM Tris pH 7.9, 150 mM NaCl, 0.25 mM TCEP), and further purified on a 1.6 x 60 cm HiLoad Superdex 200 column (GE Healthcare) with isocratic elution in Tris crystallization buffer (20 mM Tris pH 7.9, 150 mM NaCl, 0.25 mM TCEP). TM1088A and the TM1088 complex fractions were concentrated and subjected to crystallization trials.\n\npSpeedET-TM1088B was generated using the Polymerase Incomplete Primer Extension (PIPE) cloning method (Klock et al., 2008). The TM1088B C-terminal domain (residues 140\u2013218) was amplified by PCR from T. maritima MSB8 (DSM3109) genomic DNA using PfuTurbo DNA polymerase (Stratagene) and I-PIPE (Insert) primers (forward primer, 5'-ctgtacttccagggcATGAAAGTCATAATAATCGGTGG-3' and reverse primer, 5'-aattaagtcgcgttaTCAGAACTCATCTGGGAAGATGAGAGC-3', target sequence in upper case) that included sequences for the predicted 5' and 3' ends. The expression vector, pSpeedET, which encodes an amino-terminal tobacco etch virus (TEV) protease-cleavable expression and purification tag (MGSDKIHHHHHHENLYFQ/G), was PCR amplified with V-PIPE (Vector) primers (forward primer: 5'-taacgcgacttaattaactcgtttaaacggtctccagc-3', reverse primer: 5'-gccctggaagtacaggttttcgtgatgatgatgatgatg-3'). V-PIPE and I-PIPE PCR products were mixed to anneal the amplified DNA fragments together. E. coli GeneHogs (Invitrogen) competent cells were transformed with the I-PIPE / V-PIPE mixture and dispensed on selective LB-agar plates. The cloning junctions were confirmed by DNA sequencing. Expression was performed in E. coli HK100 cells grown at 37 \u00b0C in TB medium containing kanamycin. Cultures were grown to OD600 = 0.5 then induced with 0.02% (w/v) arabinose. After 4 hours, the cultures were harvested by centrifugation at 2,000 x g for 20 minutes. Cell pellets were resuspended in 20 mM TRIS pH 8, 0.1 M NaCl, 0.5 mM TCEP and disrupted by a microfluidizer (Microfluidics). Unbroken cells and cellular debris were removed by centrifugation at 12,000 x g for 20 minutes. The supernatant was applied to a 1 mL Ni-NTA agarose (Qiagen) gravity column equilibrated with 50 mM TRIS pH 8, 0.05 M NaCl, 10 mM imidazole, 0.5 mM TCEP. The column was washed with 50 mM TRIS pH 8, 0.3 M NaCl, 40 mM imidazole, 10% glycerol, 0.5 mM TCEP. The His-tagged TM1088B C-terminal protein was then eluted with 20 mM TRIS pH 8, 0.3 M imidazole, 10% glycerol, 0.5 mM TCEP, concentrated and then submitted for crystallization trials with the tag intact.\n\nProtein concentrations were determined by the Bradford method using a commercially available assay (Pierce) with bovine serum albumin as standard. SDS-PAGE was performed as described [28].\n\n*Figure 1.* The TM1088 locus codes for two proteins that form an octameric assembly. (A) TM1088 locus in the region of the overlapping start (TTG) and stop (TGA) codons (shown in red) of TM1088B and TM1088A, respectively (GenBank: NP_228894). The putative Ribosome Binding Site (RBS; GGGAG) is shown in orange. The C-terminal portion of the TM1088A transcript is shown in green (UniProt: D3KFX7_THEMA) and the N-terminal portion of the TM1088B transcript is shown in blue (UniProt: D3KFX8_THEMA). (B) Purification of the octameric TM1088 complex using size exclusion chromatography. Molecular weight standards (shown in red) and TM1088 sample (shown in dashed green-blue) were run on an S200 column at 0.5 ml/min. SDS-PAGE of affinity purified sample prior to final size exclusion step is shown inset. (C) Ribbon representation of the octameric TM1088 assembly. TM1088A and TM1088B are shown in green and blue, respectively. Bound AMP is shown as yellow spheres and residues involved in hydrogen bonding across the central cavity are highlighted by red sticks (Arg33, Glu34 and Glu37). (D) Schematic representation of the domain organization of the octameric TM1088 assembly. (E) Superposition of the octameric TM1088 assembly on to TrkA/TrkH of V. parahaemolyticus (PDB ID: 4J9U) demonstrating the overall similarity in size and architecture between the two structures. The transparent molecular surface of TM1088A and TM1088B is shown in green and blue, respectively, and the C\u03b1 ribbon trace of TrkA/TrkH in black.", "# Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design\nAuthors: Irimia, A.; Serra, A.M.; Sarkar, A.; Jacak, R.; Kalyuzhniy, O.; Sok, D.; Saye-Francisco, K.L.; Schiffner, T.; Tingle, R.; Kubitz, M.; Adachi, Y.; Stanfield, R.L.; Deller, M.C.; Burton, D.R.; Schief, W.R.; Wilson, I.A.\nSource: PLoS Pathogens, 2017\n\nOur crystallographic study of 10E8 binding to two phospholipids PG and PA reveals a lipid-binding site in a cavity delineated by CDRL1, CDRH3 and FRL3 (Figs 1A and 4A). In healthy cells, PG and PA are less abundant lipids of the plasma membrane. However, the amount of these lipids increases on the HIV-1 membrane that is acquired from the host cell during budding. A diversity of ligands (PG, PA, phosphate, glycerol) is observed in our structures in the lipid-binding site and, therefore, other phospholipid head groups could potentially occupy this cleft. The lipid binding site residues are relatively conserved in the 10E8 germline IGLV3-19*01 sequence with only two differences observed for light-chain residues 31 (tyrosine in the germline-encoded sequence and histidine in mature 10E8) and 66 (serine in the germline and alanine in mature 10E8). The presence of the tyrosine and serine at the respective positions in germline does not influence the conformation of CDRL1 and FRL3 in this cavity as comparison with the structure of the unbound germline 10E8 (PDB 5JO5) shows that the side chains of residues 31 and 66 point away from this cleft and do not change the topology of the lipid-binding site. However, CDRH3, which also delineates this cavity, is not visible in the germline structure as its residues appear to be disordered. Thus, it is not clear if the lipid-binding site is fully recapitulated in the 10E8 germline. In addition to the lipid-binding site, we observed that the 10E8 light-chain surface is coated by several basic (arginine and lysine) and multiple polar (serine and threonine) residues, which together with Lys683, the last residue of the gp41 ectodomain, lie in a plane that roughly coincides with the lipid head groups observed in our structure. Several light-chain mutants show similar binding affinity to the epitope scaffold but with decreased neutralization potency and breadth suggesting that, although binding to the MPER peptide epitope is not compromised, binding to its composite epitope formed by MPER peptide and viral membrane lipids on the virus surface would be affected. Furthermore, the light-chain mutations do not alter the Fab conformation, consistent with the binding experiments, but result in more acidic surface, strongly suggesting that the 10E8 antibody interacts with the viral membrane via this basic-polar light-chain patch. Our data also suggest an upright orientation of the MPER helical epitope of 10E8 with respect to the viral membrane (Fig 4B), which is tilted about 78\u00b0 \u00b1 3\u00b0 from the bilayer during antibody engagement (similar to that observed for 4E10). This angle is defined by the epitope's helical axis intersecting the axis of the plane of the lipid head group in the direction from which 10E8 approaches the MPER in the CHARMM model. This MPER orientation with respect to the membrane and the angle of antibody approach is also suggested by the recent cryo-EM structure of native JR-FL Env\u0394CT at 8.8 \u00c5 resolution that contains the MPER and TM domains in complex with 10E8 and PGT151 antibodies.\n\nThis cryo-EM structure provided fascinating new insights into the 10E8 interaction with Env and on new features outside the MPER that are part of the 10E8 epitope (e.g. N88 and N625 glycans), although no detailed information was possible for the 10E8 interaction with membrane lipids (or the micelle in the EM study) or for the TM region. Our study suggests that 10E8 interacts with the helical MPER with an angle of approach of ~43\u00b0 \u00b1 3\u00b0 as measured from the viral membrane to the pseudo dyad axis between the variable light and heavy chains (defined by atoms C\u03b2 of Ser179 and C\u03b1 of Phe100a), with the light-chain variable region interacting with the phospholipid head groups on the membrane surface. In this model, CDRH3 inserts between the MPER epitope and the lipid head groups with the aromatic residues at its tip located within the hydrophilic region of the lipid bilayer. It is likely that CDRH1, CDRH2 and FRH3 of 10E8 form additional contacts with Env as shown in our MD model fitted into the EM map of the full-length Env bound to CD4 (Fig 4D) and as shown in the JR-FL Env\u0394CT-10E8-PGT151 EM structure. In the JR-FL Env\u0394CT-10E8-PGT151 EM structure, the presence of PGT151, which binds with a stoichiometry of two Fabs per trimer, leads to a disruption in the symmetry of the spike causing the three 10E8 Fabs to have slightly different orientations compared to each other (i.e. slightly different angle of tilt toward the micelle position and rotations around the dyad axis between the heavy and light chains). In our model, the three 10E8 Fabs bind with the same angle of approach to their respective MPER-viral membrane composite epitope. The 10E8 Fab in the EM structure with the closest orientation to the one that we propose here is tilted only ~8\u00b0 more towards the predicted position of the membrane surface. Gp120 conformational changes on receptor and co-receptor engagement could also promote 10E8 binding. Comparison of this 10E8 model with a previous model of 4E10 bound to the gp41 epitope-lipid bilayer (Fig 4B and 4C) shows a similar orientation of the MPER with regard to the viral membrane. 10E8 and 4E10 both interact with the residues on the same face of the helical MPER epitope, but the Fab variable regions that contact the gp41 epitope differ. The two antibodies are rotated by about 90\u00b0 around their pseudo-dyad axes with respect to each other, with 10E8 interacting with the viral membrane via the light chain and CDRH3, while 4E10 interacts via CDRH1 and CDRH3. 4E10 may also make more extensive interactions with upstream regions of Env (Fig 4E) than 10E8 (Fig 4D) or most likely engage its epitope as fusion intermediate forms after receptor engagement when gp120 regions are no longer in the way. Reduction in the 10E8 interaction with other regions of Env compared with 4E10 may possibly explain its increased neutralization potency.", "# Beyond Trial and Error: AI-Driven Design of Peptide Therapeutics\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* AI is transforming peptide drug discovery by replacing slow, trial-and-error methods with rapid, *in silico* design and optimization.\n* Machine learning models can design novel peptides from scratch and accurately predict their therapeutic properties, accelerating the path to candidate selection.\n* This AI-driven approach unlocks the full potential of peptide therapeutics, making the development of these highly specific drugs faster and more efficient.\n\n1. **AI in drug discovery: Key trends shaping therapeutics in 2025:**<https://www.gubra.dk/blog/ai-in-drug-discovery-key-trends-shaping-therapeutics-in-2025/>\n2. **The future of pharmaceuticals: Artificial intelligence in drug discovery:**<https://www.sciencedirect.com/science/article/pii/S2095177925000656>\n3. **Beyond Legacy Tools: Defining Modern AI Drug Discovery for 2025:**<https://www.biopharmatrend.com/ai-drug-discovery-pipeline-2024/>\n\nPeptides, short chains of amino acids, represent a powerful class of drugs that can offer high specificity and low toxicity compared to traditional small molecules. However, their development has been a major challenge. The process often relied on labor-intensive screening of vast libraries or painstaking, iterative modifications of naturally occurring peptides. This trial-and-error approach was slow and inefficient, hindering the full therapeutic potential of peptides and leaving many promising avenues unexplored. Discovering and optimizing a peptide candidate could take years, a significant barrier to getting these targeted therapies to the clinic.\n\nArtificial intelligence is now revolutionizing this field, replacing slow, manual processes with rapid, *in silico* design. Machine learning models, trained on vast datasets of peptide sequences and their biological activities, can now design entirely new peptides from scratch. These AI systems can predict a peptide\u2019s potency, stability, and binding affinity with remarkable accuracy, allowing researchers to generate and screen millions of virtual candidates in a fraction of the time it would take in the lab. This moves the discovery process from the wet lab to the computer, focusing resources only on the most promising designs.\n\nThe impact of this shift is transformative. AI-driven approaches are not only accelerating the discovery timeline but are also enabling the design of peptides with optimized properties that would be difficult to achieve through traditional methods. By bypassing the limitations of native peptide ligands and manual optimization, AI is making peptide-based drug discovery faster, more efficient, and more predictable. This computational power is unlocking the full potential of peptide therapeutics, paving the way for a new generation of highly specific and effective drugs for a wide range of diseases.", "# The New Architects: How Generative AI Is Designing Novel Protein Scaffolds\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-14 on marcdeller.com (categories: blog).*\n\n* [**The past, present and future of de novo protein design. Nature, 2026**](https://www.nature.com/articles/s41586-026-10328-7)\n\n* **From renovation to architecture:** Generative models now design backbones de novo, expanding the addressable design space far beyond what evolution has produced.\n* **An industrial stack is forming:** Latent Labs, Cradle, Levitate, Tamarind, Amineo, and AminoAnalytica are commercialising distinct slices of the protein design workflow.\n* **Lab-validated performance is real:** Latent-X reports 91-100% macrocycle hit rates; Cradle reports 1.2 to 12 fold faster R&D cycles in pharma programmes.\n* **The bottleneck has moved:** With computational throughput far exceeding experimental capacity, biopharma\u2019s rate-limiting step has shifted decisively back to the wet lab.\n\n1. **De novo design of protein structure and function with RFdiffusion:** [Watson et al., Nature, 2023](https://www.nature.com/articles/s41586-023-06415-8)\n2. **Robust deep learning\u2013based protein sequence design using ProteinMPNN:** [Dauparas et al., Science, 2022](https://www.science.org/doi/10.1126/science.add2187)\n3. **Introducing Latent-X, a frontier generative AI model for protein binder design:** Latent Labs press release, 2025\n4. **Cradle raises $73M Series B to put AI-powered protein engineering in every lab:** [Cradle Bio, 2024](https://www.cradle.bio/blog/series-b)\n5. **Tamarind Bio secures $13.6M Series A to make AI more accessible for biology:** [GEN, 2026](https://www.genengnews.com/topics/artificial-intelligence/tamarind-bio-secures-13-6m-series-a-to-make-ai-more-accessible-for-biology/)\n\n**From the Baker lab\u2019s diffusion models to a fast-emerging ecosystem of platforms turning de novo design into a routine engineering discipline**\n\nFor years, protein engineering was a renovation business. Scientists took a scaffold that nature had already built and tinkered with surface residues, hoping to coax out a new function. The blueprint constrained the result. Generative AI has rewritten that contract. With [ProteinMPNN](https://www.science.org/doi/10.1126/science.add2187), sequences could be redesigned with a 52.4% native recovery rate that left Rosetta well behind. With [RFdiffusion](https://www.nature.com/articles/s41586-023-06415-8), full backbones could be denoised into existence, including symmetric oligomers, motif scaffolds, and picomolar binders, with the latter confirmed by cryo-EM at near-atomic accuracy. In eighteen months, the toolkit moved from sequence-on-scaffold to scaffold-from-prompt.\n\nWhat is striking now is how quickly that capability has become an industry. Generative protein design is no longer the private resource of a handful of academic groups; it is a competitive software market with a clear bench-to-pipeline value proposition. A growing list of platforms is vying for biopharma attention, each making different bets on user experience, data strategy, and the underlying model stack. Some embed AI on top of familiar physics-based pipelines; others abandon Rosetta scoring entirely in favour of end-to-end deep generative models. All of them face the same downstream constraint.\n\n**\u201cIn a manner analogous to networks that produce images from user-specified inputs, RFdiffusion enables the design of diverse functional proteins from simple molecular specifications.\u201d**\n\nThat capability now reaches scientists through commercial platforms with very different approaches. Latent Labs\u2019 Latent-X is a frontier all-atom binder model offering 91 to 100 percent lab hit rates for macrocycles and picomolar affinities for mini-binders, accessed through a no-code browser interface. [Cradle](https://www.cradle.bio/blog/series-b) couples its own wet-lab data flywheel with bespoke generative models, and counts six of the top twenty-five global pharmas among customers running active programmes on the platform. [Levitate Bio](https://levitate.bio/), spun out from Cyrus Biotechnology and David Baker\u2019s lab via the Rosetta Commons Foundation, fuses Rosetta physics with the latest AI protocols for de novo design and developability assessment.\n\nOther entrants are reshaping the workflow itself. [Tamarind Bio](https://www.genengnews.com/topics/artificial-intelligence/tamarind-bio-secures-13-6m-series-a-to-make-ai-more-accessible-for-biology/) aggregates over 250 models, among them AlphaFold, RFdiffusion, ProteinMPNN, and GROMACS, into a single orchestrated cloud used by tens of thousands of researchers and eight of the top twenty pharmas. France\u2019s [Amineo](https://www.amineo.design/) combines physics, robotics-inspired conformational sampling, and automated reasoning to design custom enzymes and binders from first principles. [AminoAnalytica\u2019s](https://aminoanalytica.com/) Amina is an agentic interface that turns a natural-language brief such as \u2018design me a thermostable mini-binder against target X\u2019 into an executed pipeline. The new architects are not merely better tools; they constitute an industrial stack assembling around the Nobel-recognised foundations laid by the Baker lab. For drug discovery, the implication is direct: design space has expanded faster than experimental capacity, and the bottleneck has shifted firmly back to the wet lab.", "# Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\nAuthors: Zhuo, J.; Li, Y.-L.; Qian, D.-Q.; Burns, D.M.; Mei, S.; Rafalski, M.; Xu, M.; Cao, G.; Pan, Y.; Jia, Z.; Jalluri, R.; Epling, L.B.; Fenalti, G.; Deller, M.C.; Procak, J.; Covington, M.; He, X.; Collins, R.; Stubbs, M.; Burke, K.; Oliver, J.; Margulis, A.; Boer, J.; Wynn, R.; Scherle, P.; Diamond, S.; Newton, R.; Zhang, Y.; Metcalf, B.; Yao, W.\nSource: Journal of Medicinal Chemistry (Drug Annotation), 2026\n\n- {1-[1-(Pyrrolidin-1-ylcarbonyl)azetidin-3-yl]-4-[4-(7H-pyrrolo[2,3-*d*]pyrimidin-4-yl)-1H-pyrazol-1-yl]piperidin-4-yl}acetonitrile (**4**), prepared in six steps from 4-(1H-pyrazol-4-yl)-7-{[2-(trimethylsilyl)ethoxy]methyl}-7H-pyrrolo[2,3-*d*]pyrimidine and *tert*-butyl 4-(cyanomethylene)piperidine-1-carboxylate.\n- Synthesis of (S)-2-(4-(4-(7H-pyrrolo[2,3-*d*]pyrimidin-4-yl)-1H-pyrazol-1-yl)-1-(1-(2-methylpyrrolidine-1-carbonyl)azetidin-3-yl)piperidin-4-yl)acetonitrile (**5**).\n- Synthesis of {1-[3-({[(2S)-2-Methylpyrrolidin-1-yl]carbonyl}cyclobutyl)-4-[4-(7H-pyrrolo[2,3-*d*]pyrimidin-4-yl)-1H-pyrazol-1-yl]piperidin-4-yl}acetonitrile (**6**).\n- Synthesis of {1-(4-{[(2S)-2-Methylpyrrolidin-1-yl]carbonyl}phenyl)-4-[4-(7H-pyrrolo[2,3-*d*]pyrimidin-4-yl)-1H-pyrazol-1-yl]piperidin-4-yl}acetonitrile bis(trifluoroacetate) (**7**).\n- {1-(4-{[(2S)-2-Methylpyrrolidin-1-yl]carbonyl}phenyl)-4-(7-{[2-(trimethylsilyl)ethoxy]methyl}-7H-pyrrolo[2,3-*d*]pyrimidin-4-yl)-1H-pyrazol-1-yl]azetidin-3-yl}acetonitrile (**8**).\n- 4-(3-(Cyanomethyl)-3-[4-(7H-pyrrolo[2,3-*d*]pyrimidin-4-yl)-1H-pyrazol-1-yl]azetidin-1-yl)-N-isopropylbenzamide (**9**), prepared in three steps from 4-bromo-N-isopropylbenzamide.\n- 4-(3-(Cyanomethyl)-3-[4-(7H-pyrrolo[2,3-*d*]pyrimidin-4-yl)-1H-pyrazol-1-yl]azetidin-1-yl)-N-[(1S)-2,2,2-trifluoro-1-methylethyl]benzamide bis(trifluoroacetate) (**14**).\n- 4-{3-(Cyanomethyl)-3-[4-(7H-pyrrolo[2,3-*d*]pyrimidin-4-yl)-1H-pyrazol-1-yl]azetidin-1-yl}-2-fluoro-N-isopropylbenzamide (**15**).\n- 4-{3-(Cyanomethyl)-3-[4-(7H-pyrrolo[2,3-*d*]pyrimidin-4-yl)-1H-pyrazol-1-yl]azetidin-1-yl}-2-fluoro-N-[(1S)-2,2,2-trifluoro-1-methylethyl]benzamide (**16**).\n- 4-{3-(Cyanomethyl)-3-[4-(7H-pyrrolo[2,3-*d*]pyrimidin-4-yl)-1H-pyrazol-1-yl]azetidin-1-yl}-2,5-difluoro-N-isopropylbenzamide (**17**).\n- 4-{3-(Cyanomethyl)-3-[4-(7H-pyrrolo[2,3-*d*]pyrimidin-4-yl)-1H-pyrazol-1-yl]azetidin-1-yl}-2,5-difluoro-N-[(1S)-2,2,2-trifluoro-1-methylethyl]benzamide (**18**, INCB047986).\n- 4-{3-(Cyanomethyl)-3-[4-(7H-pyrrolo[2,3-*d*]pyrimidin-4-yl)-1H-pyrazol-1-yl]azetidin-1-yl}-3-fluoro-N-[(1S)-2,2,2-trifluoro-1-methylethyl]benzamide (**19**).\n- 4-[3-(1H,1'H-4,4'-Bipyrazol-1-yl)-3-(cyanomethyl)azetidin-1-yl]-2,5-difluoro-N-[(1S)-2,2,2-trifluoro-1-methylethyl]benzamide trifluoroacetate (**20**).\n- 4-[3-(Cyanomethyl)-3-(3'-methyl-1H,1'H-4,4'-bipyrazol-1-yl)azetidin-1-yl]-2,5-difluoro-N-[(1S)-2,2,2-trifluoro-1-methylethyl]benzamide trifluoroacetate (**21**).\n- 4-[3-(Cyanomethyl)-3-(3',5'-dimethyl-1H,1'H-4,4'-bipyrazol-1-yl)azetidin-1-yl]-2,5-difluoro-N-[(1S)-2,2,2-trifluoro-1-methylethyl]benzamide (**22**, INCB054707, povorcitinib), prepared by Suzuki coupling of Intermediate 41 [4-(3-(cyanomethyl)-3-[4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazol-1-yl]azetidin-1-yl)-2,5-difluoro-N-[(1S)-2,2,2-trifluoro-1-methylethyl]benzamide] with 4-bromo-3,5-dimethyl-1H-pyrazole using tetrakis(triphenylphosphine)palladium(0) and sodium carbonate in 1,4-dioxane/water at 110 \u00b0C, followed by flash chromatography and prep-LC/MS purification. LC/MS (ESI) calculated for C23H23F5N7O (M+H)+: *m/z* = 508.2; found: 508.0. HRMS (ESI) calculated for C23H23F5N7O (M+H)+: *m/z* = 508.1884; found: 508.1888.\n\nKey synthetic intermediates characterized include 2,4,5-trifluoro-N-[(1S)-2,2,2-trifluoro-1-methylethyl]benzamide (**35**), 2,5-difluoro-4-(3-oxoazetidin-1-yl)-N-[(1S)-2,2,2-trifluoro-1-methylethyl]benzamide (**37**/**38**), 4-{3-(cyanomethylene)azetidin-1-yl}-2,5-difluoro-N-[(1S)-2,2,2-trifluoro-1-methylethyl]benzamide (**39**), and the pinacol boronate **41**.\n\n### Biological Assays\n\n#### Enzyme, Cellular, and Whole Blood Assays for Compounds 4\u201322\n\n**In Vitro JAK Kinase Assay (Enzymatic).** Compounds herein were tested for inhibitory activity of Janus kinase (JAK) targets according to the following in vitro assay described in the literature. The catalytic domains of human JAK1 (a.a. 837\u20131142) and JAK2 (a.a. 828\u20131132) with an N-terminal His tag were expressed using baculovirus in insect cells and purified. The catalytic activity of JAK1 or JAK2 was assayed by measuring the phosphorylation of a biotinylated peptide. The phosphorylated peptide was detected by homogeneous time-resolved fluorescence (HTRF). Half-maximal inhibitory concentrations (IC50) of compounds were measured for each kinase in the 40 \u03bcL reactions that contain the enzyme, ATP and 500 nM peptide in 50 mM Tris (pH 7.8) buffer with 100 mM NaCl, 5 mM DTT, and 0.1 mg/mL (0.01%) BSA. For the 1 mM IC50 measurements, ATP concentration in the reactions was 1 mM. Reactions were carried out at room temperature for 1 h and then stopped with 20 \u03bcL 45 mM EDTA, 300 mM SA-APC, 6 nM Eu-Py20 in assay buffer (PerkinElmer, Boston, MA). Binding to the Europium labeled antibody took place for 40 min, and HTRF signal was measured on a Fusion plate reader.", "# 2025 \u2013 Amplify Your Communication Skills with Generative AI\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-28 on marcdeller.com (categories: certificates).*\n\n* [Link to certificate](https://www.linkedin.com/learning/certificates/e73941e7b279574ef2e43e461acd40efb644f9284f285d1210cee7580ee9528a)\n\n* **Award Details:** LinkedIn Learning issued this certificate on March 18, 2025 at 05:50 PM UTC.\n* **Course Focus:** It recognizes completion of training on using generative AI to improve communication.\n* **Career Value:** This matters because strong communication is essential for explaining scientific ideas, methods, and findings.\n* **Practical Use:** It supports clearer writing, presentations, and collaboration across research teams.", "# Boltz2Run \u2013 Every run logged, timed, and reproducible\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-14 on marcdeller.com (categories: vibe-code).*\n\n> **A minimal, auditable wrapper for Boltz-1/2. Every run logged, timed, and reproducible.**\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/simple-zsh-script-to-run-boltz2](https://github.com/bellcheddar/simple-zsh-script-to-run-boltz2) |\n\n---\n\nA minimal, production-ready shell wrapper for running [Boltz-1/Boltz-2](https://github.com/jwohlwend/boltz) structure predictions from the command line. Handles logging, runtime tracking, and input sanitisation automatically so every run is reproducible and auditable.\n\nWhy it matters: running co-folding predictions by hand invites silent failures: overwritten logs, stray macOS dot-files derailing a batch, no record of how long a run took or what flags were used. run_boltz.zsh wraps Boltz in a thin, production-minded shell layer that timestamps every log, tracks wall-clock runtime, sanitises the input directory, and tees output to both screen and file. It is useful for anyone running Boltz at scale or in a regulated setting where reproducibility and an audit trail matter: drop it in, point it at your inputs, and every prediction becomes self-documenting.\n\n---\n\n## Features\n\n* **Timestamped log files** \u2014 every run writes to `boltz_run_YYYYMMDD_HHMMSS.log`; nothing is ever silently overwritten\n* **Runtime tracking** \u2014 total wall-clock time reported at the end of each run\n* **Input sanitisation** \u2014 removes macOS/Linux hidden dot-files (`.DS_Store`, `._*`, etc.) from the inputs directory before prediction\n* **MPS fallback** \u2014 sets `PYTORCH_ENABLE_MPS_FALLBACK=1` for Apple Silicon compatibility\n* **Unified stdout + log** \u2014 all output is tee\u2019d to the terminal and the log file simultaneously via `2>&1 | tee`\n\n---\n\n## Requirements\n\n* [Boltz](https://github.com/jwohlwend/boltz) installed and on your `$PATH`\n* `zsh` (macOS default; also available on Linux)\n* A CUDA-capable GPU (the script passes `--accelerator gpu`)\n* Internet access if MSA server lookups are needed (`--use_msa_server`)\n\n---\n\n## Directory layout\n\n```\nproject/\n\u251c\u2500\u2500 run_boltz.zsh       # this script\n\u251c\u2500\u2500 inputs/             # place your input YAML / FASTA / CIF files here\n\u2514\u2500\u2500 output/             # Boltz outputs written here (created automatically)\n```\n\n---\n\n## Usage\n\n```\nchmod +x run_boltz.zsh\n./run_boltz.zsh\n```\n\nThat\u2019s it. The script picks up all inputs from `./inputs/`, writes predictions to `./output/`, and creates a timestamped log in the current directory.\n\n---\n\n## Configuration\n\nAll key parameters are set at the top of the script:\n\n| Variable | Default | Description |\n| --- | --- | --- |\n| `INPUT_DIR` | `./inputs` | Directory containing Boltz input files |\n| `OUTPUT_DIR` | `./output` | Root output directory |\n| `LOGFILE` | `boltz_run_<timestamp>.log` | Per-run log filename |\n\nBoltz prediction flags can be edited directly in the `boltz predict` call:\n\n| Flag | Value | Notes |\n| --- | --- | --- |\n| `--use_potentials` | \u2014 | Enables potential energy guidance during diffusion |\n| `--diffusion_samples` | `1` | Number of diffusion samples per input; increase for ensemble diversity |\n| `--use_msa_server` | \u2014 | Queries the ColabFold MSA server; remove for purely local runs |\n| `--num_workers` | `8` | DataLoader workers; tune to your CPU core count |\n| `--accelerator` | `gpu` | Use `cpu` for debugging on machines without a GPU |\n\n---\n\n## Output\n\nEach run produces:\n\n```\noutput/\n\u2514\u2500\u2500 boltz_results_<input_name>/\n    \u251c\u2500\u2500 predictions/\n    \u2502   \u2514\u2500\u2500 <model>_model_0.cif      # ModelCIF-format predicted structure\n    \u251c\u2500\u2500 confidence_<model>_model_0.json\n    \u2514\u2500\u2500 ...\n\nboltz_run_20250611_143022.log        # full stdout/stderr capture\n```\n\nPredicted structures are written as **ModelCIF** files. These can be loaded directly in PyMOL, UCSF ChimeraX, or processed downstream with tools such as [Biotite](https://www.biotite-python.org/) or a PLIP-based interaction pipeline.\n\n---\n\n## Notes\n\n* The `cleanup_input_dir` function removes hidden files silently before each run. This prevents spurious Boltz errors on macOS where Finder writes `._<filename>` resource fork files alongside legitimate inputs.\n* Log files accumulate in the working directory; archive or delete them periodically if running many predictions.\n* To run multiple diffusion samples for ensemble analysis, increase `--diffusion_samples` (e.g. `5`) and downstream results will include `model_0` through `model_N`.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/simple-zsh-script-to-run-boltz2](https://github.com/bellcheddar/simple-zsh-script-to-run-boltz2) |", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics; supervisor Prof. E. Yvonne Jones), 1999\n\n| | Native (I222) | Native (P212121) | EMP\u03bb inflexion | EMP\u03bb remote | EMP\u03bb peak |\n|---|---|---|---|---|---|\n| Data Collection Site | Photon Factory | ESRF (BL14) | ESRF (BL14) | ESRF (BL14) | ESRF (BL14) |\n| Data Collected | \u2014 | \u2014 | 100\u00b0 | 100\u00b0 | 100\u00b0 |\n| Temperature | 295 K | 100 K | 100 K | 100 K | 100 K |\n| Wavelength | 1.000000 \u00c5 | 0.979273 \u00c5 | 1.00914 \u00c5 | 0.826516 \u00c5 | 1.00669 \u00c5 |\n| Spacegroup | I222 | P212121 | P212121 | P212121 | P212121 |\n| Unit Cell Dimensions | a = 52.9 \u00c5, b = 62.0 \u00c5, c = 109.4 \u00c5, \u03b1 = \u03b2 = \u03b3 = 90\u00b0 | a = 35.8 \u00c5, b = 53.1 \u00c5, c = 106.8 \u00c5, \u03b1 = \u03b2 = \u03b3 = 90\u00b0 | a = 35.9 \u00c5, b = 53.3 \u00c5, c = 106.7 \u00c5, \u03b1 = \u03b2 = \u03b3 = 90\u00b0 | a = 35.9 \u00c5, b = 53.3 \u00c5, c = 106.7 \u00c5, \u03b1 = \u03b2 = \u03b3 = 90\u00b0 | a = 35.9 \u00c5, b = 53.3 \u00c5, c = 106.7 \u00c5, \u03b1 = \u03b2 = \u03b3 = 90\u00b0 |\n| Resolution Range | 20\u20132.5 \u00c5 | 25\u20132.2 \u00c5 | 25\u20132.6 \u00c5 | 25\u20132.6 \u00c5 | 25\u20132.6 \u00c5 |\n| Total Number of Reflections | 15637 | 43901 | 24040 | 26223 | 24159 |\n| Number of Unique Reflections | 5544 | 10513 | 11650 | 12069 | 11691 |\n| Redundancy | 2.8 | 4.2 | 2.1 | 2.2 | 2.1 |\n| Completeness | 85.2% | 96.5% (98.2%) | 96.4% (63.1%) | 99.9% (100%) | 96.8% (66.7%) |\n| <I/\u03c3(I)> | 9.2 (2.4) | 13.2 (4.3) | 12.1 (2.9) | 12.5 (4.7) | 12.3 (3.3) |\n| Rmerge | 9.9% | 8.3% (37.6%) | 4.6% (18.7%) | 4.6% (16.4%) | 4.3% (17.3%) |\n\n*Table 5-2.* Table showing data collection and processing statistics for \u0394188_OSM. EMP\u03bb denotes MAD data collected from ethylmercuryphosphate (EMP) derivatised \u0394188_OSM, with EMP\u03bb inflexion, EMP\u03bb remote and EMP\u03bb peak representing inflexion, remote and peak data, respectively. Rmerge = \u03a3|I\u2212<I>|/\u03a3<I>. Redundancy = Total Number of Reflections/Number of Unique Reflections. Values in parentheses correspond to the highest resolution shell (2.66\u20132.60 \u00c5 for EMP\u03bb, 2.25\u20132.20 \u00c5 for native P212121 and 2.56\u20132.50 \u00c5 for native I222).\n\n#### 5.3.4 Heavy Metal Derivatisation\n\nInitial heavy atom derivative trials failed due to the high reactivity of mercury and uranyl compounds towards the OSM crystals (see Table 5-3). Even at low concentration UA, UN, BAKERS, TAMM and PMI soaks destroyed crystal instantaneously. Platinum salts presumably disrupted crystalline disorder with PN and CP soaked crystals showing no diffraction. Lead salts such as TA and LA showed no reactivity towards OSM crystals, even at higher concentrations, as no derivatisation was observed. One mercury derivative was identified by soaking crystals in equilibrated drops of mother liquor containing 14 mM ethylmercuryphosphate (EMP) for 16\u201324 hours.\n\n| Heavy Metal Reagent | Target Residues | Comments |\n|---|---|---|\n| Potassium Mercuric Iodide (PMI) | Gly/Asn/Met/His | Crystal Destroyed |\n| Mercury meso-(2,3-dimethoxytetramethylene)acetate (BAKERS) | Met/His/Cys | Crystal Destroyed |\n| Tetrakis(acetomercuric)methane (TAMM) | His/Cys | Crystal Destroyed |\n| Uranyl Acetate (UA) | Glu/Asp | Crystal Destroyed |\n| Uranyl Nitrate (UN) | Glu/Asp | Crystal Destroyed |\n| Sodium Gold Chloride (SGC) | Cys/His | Crystal Destroyed |\n| Cis-dichlorodiamine Platinum (II) (CP) | Met/His/Cys | No Diffraction |\n| Di-potassium Platinum Nitrate (PN) | Met/His/Cys | No Diffraction |\n| Trimethyllead Acetate (TA) | Glu/Asp | No Binding |\n| Lead Acetate (LA) | Glu/Asp | No Binding |\n| Ethylmercury Phosphate (EMP) | Met/His/Cys | Peaks in Patterson |\n\n*Table 5-3.* Table showing summary of heavy metal reagents used for derivative screening of \u0394188_OSM. All soaks were carried out for between 16 and 72 hours in equilibrated mother liquor containing 30 % (w/v) PEG 35000, 0.25 M sodium acetate, pH 7.5 and corresponding heavy metal reagent. Target residues represent sites of attachment previously reported for each heavy metal reagent (Blundell and Johnson, 1976).\n\nAn EMP derivative was originally identified using SRS data and a full MAD experiment was subsequently carried out at the ESRF. Derivatisation was confirmed by manual inspection of anomalous difference Patterson maps and anomalous difference Fourier maps as shown in Figure 5-3 and Figure 5-4 below. Inspection of Harker sections of the anomalous difference Patterson map reveals two major sites for the EMP derivative at refined fractional coordinates of 0.22, 0.14, 0.20 (site 1) and 0.94, 0.00, 0.13 (site 2). Inspection of anomalous difference Fourier maps, calculated using phases generated by the two major sites, revealed a third minor site at refined fractional coordinates of 0.32, 0.33, 0.24 (site 3). The same two major sites were also identified using an automated procedure within the program SOLVE (Terwilliger and Berendzen, 1999). All three sites showed good refinement using the program SHARP (de La Fortelle and Bricogne, 1997) with occupancies of 23 %, 19 % and 16 % for site 1, site 2 and site 3, respectively. Analysis of the final refined model showed the heavy atom binding sites correspond to residues His 48, His 171 and His 178 which are all solvent exposed (see Figure 5-5). Similar analysis of the final refined model reveals that the four methionine residues in OSM are buried and as such may affect folding of derivatised material (see Figure 5-5).\n\n#### 5.3.5 Multiple Wavelength Anomalous Dispersion\n\nThe mercury LIII absorption edge (observed in ideal systems at \u03bb = 1.00910 \u00c5 and 12.2858 keV) was characterised using a fluorescence scan as shown below in Figure 5-6. This shows a characteristically broad f\u2033 peak often associated with mercury derivatised proteins.", "# From Waste to Wonder: How Bacteria Transform Plastic into Life-Saving Medicine\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-11 on marcdeller.com (categories: blog, microplastics).*\n\n* [Researchers turn plastic into paracetamol](https://cen.acs.org/biological-chemistry/biotechnology/Researchers-turn-plastic-paracetamol/103/web/2025/06)\n\n* **Circular Economy Integration**: Plastic waste becomes pharmaceutical feedstock, demonstrating true circular economy principles in healthcare manufacturing where 300 million tons of pharmaceutical plastic waste annually could be transformed into essential medicines.\n* **Zero-Carbon Synthesis**: Room temperature bacterial fermentation produces paracetamol with virtually no carbon emissions compared to fossil fuel-intensive traditional methods that contribute 55% more CO\u2082 per ton than automotive manufacturing.\n* **Biocompatible Chemistry**: First demonstration of Lossen rearrangement occurring in living cells, catalyzed by naturally occurring bacterial phosphate systems, enabling new-to-nature biosynthetic pathways.\u200b\n* **High Conversion Efficiency**: 92% yield achieved from plastic-derived terephthalic acid to paracetamol in under 24 hours using engineered *E. coli* fermentation at ambient conditions.\u200b\n* **Scalable Platform Technology**: Bacterial system can be reprogrammed to produce various pharmaceutical compounds beyond paracetamol, potentially revolutionizing sustainable drug manufacturing.\u200b\n* **Environmental Impact Reduction**: Process addresses both plastic pollution (350 million tons annually) and pharmaceutical carbon footprint simultaneously, supporting circular economy principles in healthcare.\n\n1. **A biocompatible Lossen rearrangement in Escherichia coli:** [Wallace, Stephen, et al., Nature Chemistry](https://www.nature.com/articles/s41557-025-01845-5)\n2. **Upcycling plastic into painkillers: Microbes transform everyday waste into acetaminophen:** [Clark, Gaby, et al., Phys.org](https://phys.org/news/2025-06-upcycling-plastic-painkillers-microbes-everyday.html)\n3. **Bacterium engineered to produce paracetamol from plastic bottle waste:** [Chemistry World, et al., Royal Society of Chemistry](https://www.chemistryworld.com/news/bacterium-engineered-to-produce-paracetamol-from-plastic-bottle-waste/4021777.article)\n4. **Complete Enzymatic Depolymerization of Polyethylene Terephthalate:** [Zhang, Wei, et al., Environmental Science & Technology Letters](https://pubs.acs.org/doi/10.1021/acs.estlett.5c00190)\n5. **Turning plastic into medicine: Closing the loop on pharma production:** [Labiotech Editorial, et al., Labiotech](https://www.labiotech.eu/trends-news/pet-plastic-paracetamol/)\n6. **Scientists turn plastic waste into pain medicine:** [Mongabay Editorial, et al., Mongabay](https://news.mongabay.com/short-article/scientists-turn-plastic-waste-into-pain-medicine/)\n\n**Revolutionary Breakthrough: Turning Polyethylene Terephthalate into Paracetamol**\n\nIn a groundbreaking achievement that bridges environmental sustainability with pharmaceutical manufacturing, researchers at the University of Edinburgh have discovered a revolutionary method to transform everyday plastic waste into paracetamol, one of the world\u2019s most essential medications. This innovative approach represents the first documented case of producing pharmaceuticals from plastic waste using genetically engineered *Escherichia coli* bacteria, demonstrating how synthetic biology can address two pressing global challenges simultaneously: plastic pollution and sustainable pharmaceutical production.\n\nThe breakthrough centers on a previously unknown biocompatible version of the Lossen rearrangement, a classical organic chemistry reaction that has never before been observed in living systems. Professor Stephen Wallace and his team at Edinburgh\u2019s Wallace Lab engineered *E. coli* bacteria to perform this sophisticated chemical transformation at room temperature, converting polyethylene terephthalate (PET) plastic\u2014the material used in water bottles and food packaging\u2014directly into paracetamol with over 92% efficiency. This process occurs through microbial fermentation, similar to beer brewing, completing the conversion in less than 24 hours while producing virtually no carbon emissions.\n\nTraditional paracetamol manufacturing relies heavily on fossil fuel-derived chemicals, particularly benzene and phenol extracted from crude oil. These conventional processes require extreme temperatures, high pressures, and generate significant greenhouse gas emissions. The pharmaceutical industry consumes thousands of tons of fossil fuels annually just to power paracetamol production facilities, making this medication a substantial contributor to climate change. Wallace\u2019s team recognized that by utilizing plastic waste as a renewable feedstock, they could simultaneously address the environmental burden of both plastic pollution and pharmaceutical manufacturing.\n\nThe biological transformation begins when PET plastic is chemically broken down into terephthalic acid, one of its constituent monomers. The engineered *E. coli* then metabolizes this compound through a carefully designed metabolic pathway that converts terephthalic acid into para-aminobenzoic acid (PABA), an essential bacterial metabolite. Through the biocompatible Lossen rearrangement\u2014catalyzed by naturally occurring phosphate within the bacterial cells\u2014PABA is transformed into the active ingredient of paracetamol. This process represents the first time that synthetic chemistry has been successfully integrated with living cellular systems to produce pharmaceuticals from waste materials.", "# buttfold.mdeller.com: web app pages\nAuthors: Deller, M.C.\nSource: mdeller.com, 2026\n\n## buttfold.mdeller.com/\n\n# Loading\n\nColour\n\nStyle\n\nEngine\n\nUniProt\n\n\u2014\n\nG\u014d model \u00b7 toward a known structure\nprecomputed\n\nShow the music\n\nRg (\u00c5)\u2014\n\nEnds (\u00c5)\u2014\n\nCompact\u2014\n\nNative\u2014\n\nContacts\u2014\n\nHelix\u2014\n\nSheet\u2014\n\nCoil\u2014\n\n### Secondary structure over time\n\nhelix \u00b7 sheet \u00b7 coil\n\n### Radius, compaction and contacts over time\n\nradius \u00b7 compaction \u00b7 contacts\n\n## The gallery\n\n9 folds\n\nYes, it is a butt-dial joke. It will not come up again.\nNo protein folds this way. ButtFold animates a simple physics model relaxing a chain\ntoward a structure it already knows.\n\nWhat it is, and is notClose\n\nThat is not a prediction of an unknown structure and it is not a folding pathway.\nESMFold predicts where a real UniProt protein ends up \u2014 at Meta, not on this\nserver \u2014 and the G\u014d model then animates a chain collapsing toward that\nprediction. A prediction can be wrong, so every protein offered has an experimental\nstructure in the PDB to check it against.\n\nThe 150 poses in a trajectory are 50,000 integration steps apart, far enough that\nan atom can move further than the protein is wide between two of them. What you see\nin between is drawn to join them up, not computed, so the motion is smooth where the\nmodel is not. The music is a faithful map of the simulation's events, and nothing\nmore.\n\n**Every note is an event in the fold.** Nothing is composed and\nnothing is random: the same protein in the same style gives the same piece, every\ntime.\n\nHow the music is madeClose\n\nTwo residues touching for the first time is a note. How far apart they are in the\nchain sets its register, so a local contact sounds high and a long-range one low; a\nlong-range contact between two greasy residues \u2014 the core coming together\n\u2014 goes to the bass.\n\nHelix\n:   a sustained pad in stacked fourths\n\nSheet\n:   a staccato interlocking figure\n\nCoil\n:   arpeggios between the chord tones\n\nHow folded one residue is\n:   that note\u2019s velocity\n\nHow folded the chain is\n:   the low-pass, the detune and the reverb\n\nRadius of gyration\n:   tempo and register\n\nConverging\n:   a cadence, resolving to the tonic\n\nSo a protein that never quite resolves stays a detuned wash for the whole piece,\nand one that snaps shut accelerates into a cadence. Each note is placed in space\nwhere its own residue is, which is worth headphones. A bar holds sixteen contacts at\nmost; when more form than that the surplus is counted and reported under the\ntransport rather than quietly dropped.", "# From Waste to Wonder: How Bacteria Transform Plastic into Life-Saving Medicine\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe environmental implications extend far beyond pharmaceutical production. PET plastic creates over 350 million tons of waste annually, with most ending up in landfills or polluting oceans. Existing mechanical recycling processes typically produce lower-grade plastic products that continue contributing to the waste stream. Wallace\u2019s approach offers a genuine upcycling solution, converting waste plastic into high-value pharmaceutical products. The research team successfully demonstrated their method using actual PET bottles, producing the equivalent of approximately two paracetamol tablets from a single bottle through their bacterial fermentation process.\n\n| Concept | Description | Key References |\n| --- | --- | --- |\n| Biocompatible Lossen Rearrangement | First demonstration of classical organic chemistry reaction occurring in living bacterial cells, catalyzed by naturally occurring phosphate ions | [Wallace et al., Nature Chemistry](https://pmc.ncbi.nlm.nih.gov/articles/PMC12226347/) |\n| PET Depolymerization | Chemical breakdown of polyethylene terephthalate plastic into terephthalic acid monomer for bacterial conversion | [Zhang et al., Environmental Science](https://pubs.acs.org/doi/10.1021/acs.estlett.5c00190) |\n| Metabolic Engineering | Genetic reprogramming of E. coli bacteria to convert plastic-derived compounds into pharmaceutical molecules | [Labiotech Editorial](https://www.labiotech.eu/trends-news/pet-plastic-paracetamol/) |\n| Para-Aminobenzoic Acid (PABA) Synthesis | Essential bacterial metabolite produced from terephthalic acid, serving as precursor for paracetamol biosynthesis | [Novel PABA Analogs Research](https://pmc.ncbi.nlm.nih.gov/articles/PMC10604881/) |\n| Circular Pharmaceutical Manufacturing | Integration of waste plastic as renewable feedstock for pharmaceutical production, reducing reliance on fossil fuels | [Clark et al., Phys.org](https://phys.org/news/2025-06-upcycling-plastic-painkillers-microbes-everyday.html) |\n| Sustainable Biotechnology Platform | Room temperature microbial fermentation system producing pharmaceuticals with minimal carbon emissions | [Mongabay Editorial](https://news.mongabay.com/short-article/scientists-turn-plastic-waste-into-pain-medicine/) |", "# GitHub README: ChemSage\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ChemSage, 2026\n\nThe `data/corpus/` directory contains the full retrieval knowledge base: tool reference docs plus two databases of experimental chemistry data:\n\n### ChEMBL (`data/corpus/chembl/`, ~70 MB)\n\n| File | Records | Contents |\n|---|---|---|\n| `approved_drugs.csv` | 3,257 | All FDA/EMA approved small molecules with SMILES + properties |\n| `bioactivity_key_targets.csv` | 112,357 | IC50/Ki/Kd for EGFR, CDK2, HIV protease, SARS-CoV-2 3CLpro, VEGFR2, Bcr-Abl, COX-2, BRAF, ALK, JAK1/2, hERG |\n| `bioactivity_extended_targets.csv` | ~200k | IC50/Ki for 36 additional targets: BTK, PARP1, CDK4/6, FLT3, mTOR, AKT1, PI3Ka, RET, MET, SRC, HDAC1, BRD4, EZH2, AR, ERa, PPARg, GLP-1R, Factor Xa, Thrombin, ACE, HMG-CoA, PDE5, DHFR, IDH1/2, KRAS + more |\n| `admet_data.csv` | ~50k | Caco-2 permeability, CLint, t1/2, fu,plasma, solubility, logD, CYP3A4/2D6/2C9/1A2/2C19 inhibition |\n| `clinical_candidates_phase2_3.csv` | ~10k | Phase 2-3 small molecules with properties |\n| `fragment_actives.csv` | 10,000 | MW < 300 fragments with pIC50 >= 4 |\n| `natural_products.csv` | ~5k | Natural product-derived compounds |\n| `drug_mechanisms.csv` | 7,561 | Mechanism of action + target for all drugs |\n| `drug_indications.csv` | 60,055 | Disease indications for all drugs |\n| `atc_classification.csv` | 5,579 | Full ATC therapeutic class hierarchy |\n\n### PubChem (`data/corpus/pubchem/`, ~6.6 MB)\n\n| File | Records | Contents |\n|---|---|---|\n| `fda_drugs_with_pubchem_cids.csv` | 3,257 | ChEMBL approved drugs cross-referenced to PubChem CIDs |\n| `tox21_panel.csv` | ~90k | 9 Tox21 nuclear receptor / stress-pathway assays (AhR, AR, ERa, NFkB, Nrf2, MMP, p53, ATAD5, TR) |\n| `pkis2_kinase_panel.csv` | ~50k | Published Kinase Inhibitor Set 2: 367-kinase selectivity panel |\n| `drug_repurposing_hub.csv` | ~6k | Broad Institute Drug Repurposing Hub clinical annotations |\n| `adme_cyp_hts.csv` | 19,803 | NCATS Caco-2 permeability, P-gp efflux, CYP2C9 inhibition HTS |\n| `bindingdb_affinities.csv` | \u2014 | BindingDB Ki/IC50/Kd (literature-curated) |\n| `nci60_gi50.csv` | \u2014 | NCI-60 cancer cell line growth inhibition |\n\n### PDB (`data/corpus/pdb/`, ~111 MB)\n\n| File | Records | Contents |\n|---|---|---|\n| `pdb_all_entries.csv` | 255,239 | All PDB structures: ID, method, resolution, compound, organism, deposit date |\n| `pdb_ligands_all.csv` | 16,171 | All unique PDB ligands with SMILES, InChI, InChIKey, formula, MW, type, ChEMBL/DrugBank/PubChem cross-references |\n| `pdb_ligands_druglike.csv` | 382 | Drug-like CCD subset (MW 100-800 Da, non-polymer, with SMILES) |\n| `pdb_ligand_structure_pairs.csv` | 121,541 | Protein-ligand pairs: which ligands are bound in which structures (with resolution + method) |\n| `pdb_druglike_structure_pairs.csv` | 1,671 | Drug-like protein-ligand pairs only |\n| `pdb_binding_affinities.csv` | 46,885 | Measured Kd/Ki/IC50 from PDBbind-annotated structures (9,209 entries) |\n| `sifts_pdb_uniprot.csv` | 965,937 | SIFTS: PDB chain -> UniProt accession (canonical target identity bridge) |\n| `sifts_pdb_pfam.csv` | 1,022,861 | SIFTS: PDB chain -> Pfam domain family (fold-level annotation) |\n\n## \ud83d\udcdd Recent changes\n\n### Live web deployment (2026-07-24)\n\n- **chemsage.mdeller.com is live.** The real `scripts/chat.py` CLI now runs in the browser via an xterm.js terminal, bridged over WebSocket to a server-side PTY \u2014 same Rich output, corpus tables, and slash commands as the local CLI, not a rebuilt chat-bubble UI.\n- **No local GPU on the droplet:** `mlx_lm`'s `stream_generate`/`generate` are patched at import time (`chat_remote.py`) to call a HuggingFace Space running the real fused 32B model on ZeroGPU, then the unmodified `chat.py` is run unaware of the substitution.\n- **Serving stack:** gunicorn (`gevent` worker, 1 worker / 100 connections \u2014 a PTY session is I/O-bound, not CPU-bound) behind nginx with WebSocket-upgrade support on `/ws`, systemd service, Let's Encrypt TLS. Full detail in **Web deployment** below.\n\n### Chat CLI and dependency fixes (2026-06-30)\n\n- **mlx-lm 0.31.3 API fix (`scripts/chat.py`):** `stream_generate()` and `generate()` no longer accept `verbose=False` as a keyword argument \u2014 removed from both call sites.\n- **Corpus fast-path (`rag/corpus_lookup.py`):** Added `\"get\"` to the lookup trigger words so queries like `\"get all EGFR structures\"` route directly to the corpus table view instead of falling through to the LLM.\n- **System prompt (`scripts/chat.py`):** Added explicit instruction to render retrieved structured data (PDB structures, bioactivity measurements) as markdown tables rather than generating external API code.\n- **Dependencies (`requirements.txt`):** Added `torch` explicitly (required by `sentence-transformers` for local embeddings).\n- **How to run:** Added quick-start run instructions to the main README \u2014 single terminal, venv python, `--no-rag` flag for model-only mode.\n- **All 5 models published to HuggingFace Hub** (`Dellboy/chem_sage_*`).\n\n### Eval harness fixes and 5-round comparative eval (2026-06-29)\n\n**Tokenizer fix (all models):**\n- `tokenizer_config.json` in `chem_sage_7b_v1`, `chem_sage_32b_v2`, `chem_sage_32b_v3`, `chem_sage_32b_v4` had `extra_special_tokens` as a JSON list of strings. Newer `transformers` calls `.keys()` on this field \u2014 causing an `AttributeError` on load. The server silently swallowed the error, loaded a broken tokenizer, and then hung indefinitely on every inference request (all queries timed out). Fixed by converting to an empty dict `{}` in all four files. `chem_sage_32b_v5` was already in the correct format.\n\n**Eval harness model-path fix (`eval/compare/eval_compare.py`):**\n- `query_model()` previously sent `\"model\": \"chemsage\"` in every HTTP POST to the local `mlx_lm.server`. Since `mlx_lm` 0.31.3, an unknown model name triggers a 120-second HuggingFace lazy-load timeout, causing all R1\u2013R4 evals to time out. Fixed by passing the full local model directory path (`model_path`) as the `\"model\"` field.", "# Unleashing Dynamic Precision: How DRGSCROLL Transforms Protein\u2013Ligand Docking\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-20 on marcdeller.com (categories: sb).*\n\n* Evolutionary side-chain optimization reveals true binding interactions.\n* Automated rotamer search eliminates manual pre-selection steps.\n* Enhances affinity and specificity predictions in lead optimization.\n* Validates docking hypotheses across protein\u2013protein and allosteric interfaces.\n* Improves local geometry for cryo-EM model fitting.\n* Exports refined complexes for molecular dynamics and free-energy calculations.\n* Upcoming features include nucleic-acid docking and batch processing.\n* Web-based interface ensures accessibility and seamless integration.\n\n1. Durda\u011f\u0131 Research Group. **\u201cPrecision Starts at the Sidechain.\u201d** DRGSCROLL:\n   <https://drgscroll.com>\n2. Jain, Anupriya, et al. **\u201cSide-chain flexibility in protein\u2013ligand docking.\u201d** Journal of Chemical Information and Modeling, vol. 56, no. 1 (2016): 22\u201332:\n   <https://doi.org/10.1021/acs.jcim.5b00493>\n3. Leach, Andrew R., et al. **\u201cMolecular docking: a computational approach for drug design.\u201d** Drug Discovery Today, vol. 2, no. 9 (1997): 424\u2013431:\n   <https://doi.org/10.1016/S1359-6446(97)01107-9>\n4. Meng, Xiang, et al. **\u201cMolecular docking: a powerful approach for structure-based drug discovery.\u201d** Current Computer\u2010Aided Drug Design, vol. 7, no. 2 (2011): 146\u2013157:\n   <https://doi.org/10.2174/157340911795677602>\n5. Case, David A., et al. **\u201cAMBER 2020.\u201d** University of California, San Francisco:\n   <https://ambermd.org/>\n\nIn the fast-paced world of structural biology and drug discovery, every atomic detail matters. DRGSCROLL (Durda\u011f\u0131 Research Group\u2019s Side Chain Rotation for Ligand Landing) is redefining how researchers explore protein\u2013ligand interactions by introducing evolution-driven side-chain optimization. Unlike traditional rigid docking approaches that treat side chains as static or constrained to pre-defined rotamers, DRGSCROLL leverages a genetic algorithm to explore a vast conformational space, revealing side-chain orientations that reflect true binding interactions. This platform enables scientists to fine-tune lead compounds, validate docking hypotheses, and discover novel interactions with unprecedented accuracy.\n\nDRGSCROLL\u2019s intuitive web interface guides users through a seamless workflow: upload your protein structure (in PDB format), submit the ligand coordinate file, and voil\u00e0\u2014side-chain conformations adapt organically, guided by the algorithm\u2019s fitness function. The result is a refined protein\u2013ligand complex that \u201cbreathes\u201d and binds with the precision you need to drive your research forward.\n\nThis level of dynamic refinement is crucial when optimizing the affinity and specificity of lead compounds. Small shifts in side-chain rotamers can mean the difference between a moderate binder and a high-potency inhibitor. By automating the search for optimal side-chain placements\u2014without the need for manual rotamer selection\u2014DRGSCROLL accelerates hit-to-lead and lead-optimization campaigns, reduces false-positive docking poses, and uncovers hidden pockets that rigid methods often miss.\n\nBeyond small-molecule docking, DRGSCROLL serves as a powerful tool for:\n\n* Validating structural hypotheses in protein\u2013protein interfaces\n* Modeling conformational selection in allosteric regulation\n* Investigating enzyme mechanism through side-chain dynamics\n* Enhancing cryo-EM model fitting by improving local geometry\n* Guiding site-directed mutagenesis by revealing energetically favorable rotamers\n\nDRGSCROLL also integrates seamlessly with downstream analysis tools. Export refined complexes directly to molecular dynamics packages like GROMACS or AMBER for further sampling, or feed coordinates into free-energy perturbation pipelines for quantitative affinity prediction. With an active development roadmap, upcoming features include support for nucleic-acid docking, covariance-driven mutation sampling, and high-throughput batch processing for virtual screening libraries.\n\nBy merging rigorous computational algorithms with a user-friendly web platform, DRGSCROLL empowers researchers to push the boundaries of structural refinement. In a domain where millisecond search times and single-angstrom accuracy can dictate research outcomes, DRGSCROLL\u2019s genetic-algorithm backbone ensures that your protein\u2013ligand complexes reflect the structure as it truly wants to bind.", "# BoltzMaker: Vibe-Coding a Campaign-Scale Boltz-2 Pipeline\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n# Example: 5-HT2A/2B/2C serotonin receptors, agonist vs antagonist, with vs\n# without the Gq heterotrimer\n#\n# Three closely related Gq/11-coupled serotonin receptors (5-HT2A UniProt\n# P28223, 5-HT2B UniProt P41595, 5-HT2C UniProt P28335), each crossed with\n# a real agonist and a real antagonist, each ligand predicted BOTH with\n# and without the Gq heterotrimer (GNAQ + GNB1 + GNG2) co-folded --\n# following the same lesson learned from adrb2_gs_panel: co-folding a G\n# protein biases Boltz toward an active-like fold regardless of ligand\n# identity, so a fair agonist-vs-antagonist comparison needs the receptor\n# alone as well as the physiologically-correct (agonist+Gq) condition.\n# Every receptor gets two Protein: blocks sharing one sequence -- one\n# with Partners: (the Gq trimer), one without -- exactly as\n# adrb2_gs_panel does, extended to a 3x2x2 = 12-target panel: 3\n# receptors x 2 ligands x (with/without Gq).\n#\n# Apo reference structures: genuinely ligand-free, not an antagonist-\n# bound \"inactive-like\" stand-in. Checked entity-by-entity across all 59\n# deposited structures for these 3 receptors (not trusting RCSB's summary\n# fields, which turned out to have real gaps) -- every single one has a\n# real ligand bound, including ones that first looked apo (e.g. what\n# looked like a ligand-free 5-HT2B entry turned out to have ergotamine).\n# No genuinely apo experimental structure exists for any of the three.\n# Instead, each receptor's own apo reference is a Boltz prediction of the\n# receptor ALONE (no ligand, no partner) -- a real BoltzMaker target via\n# `Ligands: none` (see the 3 apo Protein: blocks below), staged and run\n# in the exact same `boltz predict` batch as the 12 ligand-bound targets\n# below (one manifest, one campaign, no separate invocation), producing\n# reference/*_apo_predicted.cif via boltz_cif/. Clearly a computational\n# model, not an experimental one, but unambiguously ligand-free.\n#\n# All 6 ligands are real, clinically/pharmacologically characterized\n# compounds: Risperidone and Psilocin (5-HT2A antagonist/agonist),\n# Balovaptan and LSD (5-HT2B antagonist/agonist), SB-242084 and\n# Lorcaserin (5-HT2C antagonist/agonist).\n\nSettings:\nOutput folder: ./boltz_yamls\nPredict affinity: yes\n\n# ===================== 5-HT2A =====================\n\nProtein: 5HT2A\nSequence: MDILCEENTSLSSTTNSLMQLNDDTRLYSNDFNSGEANTSDAFNWTVDSENRTNLSCEGCLSPSCLSLLHLQEKNWSALLTAVVIILTIAGNILVIMAVSLEKKLQNATNYFLMSLAIADMLLGFLVMPVSMLTILYGYRWPLPSKLCAVWIYLDVLFSTASIMHLCAISLDRYVAIQNPIHHSRFNSRTKAFLKIIAVWTISVGISMPIPVFGLQDDSKVFKEGSCLLADDNFVLIGSFVSFFIPLTIMVITYFLTIKSLQKEATLCVSDLGTRAKLASFSFLPQSSLSSEKLFQRSIHREPGSYTGRRTMQSISNEQKACKVLGIVFFLFVVMWCPFFITNIMAVICKESCNEDVIGALLNVFVWIGYLSSAVNPLVYTLFNKTYRSAFSRYIQCQYKENKKPLQLILVNTIPALAYKSSQLQMGQKKNSKQDAKTTDNDCSMVALGKQHSEEASKDNSDGVNEKVSCV\nPartners: GNAQ, GNB1, GNG2\nLigands: RISP, PSIL\nApo structure: boltz_cif/H2AAP_model_0.cif\nFamily type: gpcr\nGroup: 5HT2A\n\nProtein: H2ANG\nSequence: MDILCEENTSLSSTTNSLMQLNDDTRLYSNDFNSGEANTSDAFNWTVDSENRTNLSCEGCLSPSCLSLLHLQEKNWSALLTAVVIILTIAGNILVIMAVSLEKKLQNATNYFLMSLAIADMLLGFLVMPVSMLTILYGYRWPLPSKLCAVWIYLDVLFSTASIMHLCAISLDRYVAIQNPIHHSRFNSRTKAFLKIIAVWTISVGISMPIPVFGLQDDSKVFKEGSCLLADDNFVLIGSFVSFFIPLTIMVITYFLTIKSLQKEATLCVSDLGTRAKLASFSFLPQSSLSSEKLFQRSIHREPGSYTGRRTMQSISNEQKACKVLGIVFFLFVVMWCPFFITNIMAVICKESCNEDVIGALLNVFVWIGYLSSAVNPLVYTLFNKTYRSAFSRYIQCQYKENKKPLQLILVNTIPALAYKSSQLQMGQKKNSKQDAKTTDNDCSMVALGKQHSEEASKDNSDGVNEKVSCV\nLigands: RISP, PSIL\nApo structure: boltz_cif/H2AAP_model_0.cif\nFamily type: gpcr\nGroup: 5HT2A\n\nProtein: H2AAP\nSequence: MDILCEENTSLSSTTNSLMQLNDDTRLYSNDFNSGEANTSDAFNWTVDSENRTNLSCEGCLSPSCLSLLHLQEKNWSALLTAVVIILTIAGNILVIMAVSLEKKLQNATNYFLMSLAIADMLLGFLVMPVSMLTILYGYRWPLPSKLCAVWIYLDVLFSTASIMHLCAISLDRYVAIQNPIHHSRFNSRTKAFLKIIAVWTISVGISMPIPVFGLQDDSKVFKEGSCLLADDNFVLIGSFVSFFIPLTIMVITYFLTIKSLQKEATLCVSDLGTRAKLASFSFLPQSSLSSEKLFQRSIHREPGSYTGRRTMQSISNEQKACKVLGIVFFLFVVMWCPFFITNIMAVICKESCNEDVIGALLNVFVWIGYLSSAVNPLVYTLFNKTYRSAFSRYIQCQYKENKKPLQLILVNTIPALAYKSSQLQMGQKKNSKQDAKTTDNDCSMVALGKQHSEEASKDNSDGVNEKVSCV\nLigands: none\nFamily type: gpcr\nGroup: 5HT2A\n\n# ===================== 5-HT2B =====================\n\nProtein: 5HT2B\nSequence: MALSYRVSELQSTIPEHILQSTFVHVISSNWSGLQTESIPEEMKQIVEEQGNKLHWAALLILMVIIPTIGGNTLVILAVSLEKKLQYATNYFLMSLAVADLLVGLFVMPIALLTIMFEAMWPLPLVLCPAWLFLDVLFSTASIMHLCAISVDRYIAIKKPIQANQYNSRATAFIKITVVWLISIGIAIPVPIKGIETDVDNPNNITCVLTKERFGDFMLFGSLAAFFTPLAIMIVTYFLTIHALQKKAYLVKNKPPQRLTWLTVSTVFQRDETPCSSPEKVAMLDGSRKDKALPNSGDETLMRRTSTIGKKSVQTISNEQRASKVLGIVFFLFLLMWCPFFITNITLVLCDSCNQTTLQMLLEIFVWIGYVSSGVNPLVYTLFNKTFRDAFGRYITCNYRATKSVKTLRKRSSKIYFRNPMAENSKFFKKHGIRNGINPAMYQSPMRLRSSTIQSSSIILLDTLLLTENEGDKTEEQVSYV\nPartners: GNAQ, GNB1, GNG2\nLigands: BALO, LSD1\nApo structure: boltz_cif/H2BAP_model_0.cif\nFamily type: gpcr\nGroup: 5HT2B\n\nProtein: H2BNG\nSequence: MALSYRVSELQSTIPEHILQSTFVHVISSNWSGLQTESIPEEMKQIVEEQGNKLHWAALLILMVIIPTIGGNTLVILAVSLEKKLQYATNYFLMSLAVADLLVGLFVMPIALLTIMFEAMWPLPLVLCPAWLFLDVLFSTASIMHLCAISVDRYIAIKKPIQANQYNSRATAFIKITVVWLISIGIAIPVPIKGIETDVDNPNNITCVLTKERFGDFMLFGSLAAFFTPLAIMIVTYFLTIHALQKKAYLVKNKPPQRLTWLTVSTVFQRDETPCSSPEKVAMLDGSRKDKALPNSGDETLMRRTSTIGKKSVQTISNEQRASKVLGIVFFLFLLMWCPFFITNITLVLCDSCNQTTLQMLLEIFVWIGYVSSGVNPLVYTLFNKTFRDAFGRYITCNYRATKSVKTLRKRSSKIYFRNPMAENSKFFKKHGIRNGINPAMYQSPMRLRSSTIQSSSIILLDTLLLTENEGDKTEEQVSYV\nLigands: BALO, LSD1\nApo structure: boltz_cif/H2BAP_model_0.cif\nFamily type: gpcr\nGroup: 5HT2B\n\nProtein: H2BAP\nSequence: MALSYRVSELQSTIPEHILQSTFVHVISSNWSGLQTESIPEEMKQIVEEQGNKLHWAALLILMVIIPTIGGNTLVILAVSLEKKLQYATNYFLMSLAVADLLVGLFVMPIALLTIMFEAMWPLPLVLCPAWLFLDVLFSTASIMHLCAISVDRYIAIKKPIQANQYNSRATAFIKITVVWLISIGIAIPVPIKGIETDVDNPNNITCVLTKERFGDFMLFGSLAAFFTPLAIMIVTYFLTIHALQKKAYLVKNKPPQRLTWLTVSTVFQRDETPCSSPEKVAMLDGSRKDKALPNSGDETLMRRTSTIGKKSVQTISNEQRASKVLGIVFFLFLLMWCPFFITNITLVLCDSCNQTTLQMLLEIFVWIGYVSSGVNPLVYTLFNKTFRDAFGRYITCNYRATKSVKTLRKRSSKIYFRNPMAENSKFFKKHGIRNGINPAMYQSPMRLRSSTIQSSSIILLDTLLLTENEGDKTEEQVSYV\nLigands: none\nFamily type: gpcr\nGroup: 5HT2B\n\n# ===================== 5-HT2C =====================", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n## \ud83e\uddeb Testing\n\n```sh\n.venv/bin/pip install -r requirements-dev.txt\n.venv/bin/pytest tests/\n```\n\n455 tests here, plus 616 for the web app. The `compare-sse` annotators are covered against real\nfixture data (a real apo EGFR kinase-domain structure vs the `egfr_covalent` example's real holo\nprediction; a real apo beta2-adrenergic-receptor structure vs `adrb2_gs_panel`'s real holo\npredictions), with GPCRdb/KLIFS/PDBe network calls swapped for an injectable fake client seeded\nwith real, previously-verified API responses -- fully offline and fast (~9s). Plus grammar and\nCLI-resolution tests for the parser fields above, chain-resolution tests against real\nfusion-construct and kinase-domain-only apo structures, GPCRdb/KLIFS/Pfam annotator pipelines, and\nthe dashboard's summary-stats and SSE-table column logic.\n\n53 of the web app's tests cover Fully Automated Mode end to end. Rather than asserting against a\nhand-written `.bmz` fixture, they render the real `pack_results.py` out of a real bundle, run it\nover a synthetic campaign, and read the result back with the real reader -- packer and reader are\ntwo halves of one contract living in different files on different machines, exactly the shape of\nthing that drifts silently. Also covered: the generated bundle really executed under `sh`,\n`bash`, `dash`, `zsh` and `ksh` (final step stubbed), proving the documented\n`sh ./<bundle>.command` works where `/bin/sh` is dash as well as bash; the generated scripts\nchecked with `bash -n` (they only ever execute on someone else's machine); every run-setting flag\nchecked against `BoltzMaker.py all --help` so a typo cannot reach a user's overnight run; and the\nhostile-upload guards exercised with real zip-slip, compression-bomb and malformed-manifest\narchives.\n\nFour `compare-sse` tests need example structure data that is gitignored\n(`examples/*/boltz_cif/`), so they fail in a fresh clone until you have run one of the example\ncampaigns.\n\n## \ud83d\udcda Citation\n\n> Passaro, S., Corso, G., Wohlwend, J., Reveiz, M., Thaler, S., Somnath, V.R., Getz, N., Portnoi, T., Roy, J., Stark, H., Kwabi-Addo, D., Beaini, D., Jaakkola, T., Barzilay, R. (2025). Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. *bioRxiv*. https://doi.org/10.1101/2025.06.14.659707\n\n> Mirdita, M., Sch\u00fctze, K., Moriwaki, Y., Heo, L., Ovchinnikov, S., Steinegger, M. (2022). ColabFold: making protein folding accessible to all. *Nature Methods*. https://doi.org/10.1038/s41592-022-01488-1\n\n> Schake, P., Bolz, S.N. et al. (2025). PLIP 2025: introducing protein-protein interactions to the protein-ligand interaction profiler. *Nucleic Acids Research*, gkaf361. https://doi.org/10.1093/nar/gkaf361\n\n> The PyMOL Molecular Graphics System, Version 3.1, Schr\u00f6dinger, LLC.\n\n> Rego, N., Koes, D. (2015). 3Dmol.js: molecular visualization with WebGL. *Bioinformatics*, 31(8), 1322-1324. https://doi.org/10.1093/bioinformatics/btu829\n\n## \ud83d\udcc4 License\n\n[MIT](LICENSE) &copy; Marc C. Deller\n\n## \ud83d\udccb To do\n\n- [x] Fix the **Apple-silicon masked-select miscount** in boltz's flat-bottom steering\n  potential, which cost a 58-target campaign nine crashes and eighteen lost targets over three\n  days. `energy[mask] = other[mask]` compacts twice through the same mask, and on a large tensor\n  with a sparse mask the two compactions return different counts (73 against 82, on 3 x 35,830,451\n  values with 73 flagged), killing the whole `boltz predict` process and every target queued\n  behind it. Replaced with `torch.where`, which is elementwise and never compacts: verified\n  identical on 600 randomised cases, on the production tensor shape, and exactly against a CPU\n  reference on the GPU. Shipped with a shape reporter that re-raises the original error unchanged,\n  since the message named neither the mis-sized operand nor the potential it came from. See\n  **What we fixed in Boltz** item 9.", "# AlphaFraud \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n# --workers N adds a download thread pool (network-bound work parallelises well). Single-\n# threaded is the reliable default for long unattended runs; workers give a partial speedup\n# but are more sensitive to the droplet's core count.\npython3 AlphaFraud.py backfill --all --two-tier --workers 6\n\n# Database summary + worst-offenders leaderboard\npython3 AlphaFraud.py status\n\n# Local dev web server (production uses gunicorn)\npython3 AlphaFraud.py serve --port 8000\n```\n\n| Command | Purpose |\n| --- | --- |\n| `init` | Create directories, the SQLite schema, and vendor `plotly.min.js` |\n| `run` | Discover, match, score and store one deposit window (`--since`, `--until`, `--limit`, `--dry-run`) |\n| `backfill` | Process a historical window (`--from`, `--to`); `--two-tier` screens by TM-score and fully analyses only disagreements below `--tm-threshold`; `--all` covers the entire post-cutoff archive in resumable monthly chunks |\n| `status` | Print a database summary and the current leaderboard |\n| `serve` | Run the Flask app locally (`--port`, `--debug`) |\n\n## \ud83d\udcca Output\n\nThe web app serves live from SQLite. Every plot carries a plain-language explanatory panel (what the axes mean, the cutoffs, a take-home summary, and a per-metric breakdown giving both a lay explanation and the equation, e.g. `TM = (1/L) \u03a3\u1d62 1/(1 + (d\u1d62/d\u2080)\u00b2)`) plus a CSV export; every table exports to CSV too.\n\nThe site is live at **[alphafraud.mdeller.com](https://alphafraud.mdeller.com/)**.\n\n| Route | Shows |\n| --- | --- |\n| [`/`](https://alphafraud.mdeller.com/) | Home \u2014 a concise navigation page: the project\u2019s purpose, headline KPIs, and a colour-coded card per tab with a \u201cwhat you\u2019ll learn\u201d line |\n| [`/all`](https://alphafraud.mdeller.com/all) | Cumulative \u201cAll structures\u201d dashboard across every processed week: KPIs, the pLDDT-vs-TM fraud-quadrant scatter, metric histograms, the weekly trend, and a top-ranked table. A pulldown jumps to any single week |\n| `/week/<label>` | One release week, same layout scoped to that batch |\n| `/entry/<id>` | Per-structure detail: per-residue error tracks, calibration scatter, distance-matrix and PAE-vs-observed heatmaps, per-domain breakdown, and every metric |\n| [`/leaderboard`](https://alphafraud.mdeller.com/leaderboard) | The all-time worst AlphaFold failures across every processed week |\n| [`/analysis`](https://alphafraud.mdeller.com/analysis) | Structural deep dive over the worst offenders: fold/family enrichment (Wilson-CI + Fisher), sequence-similarity clustering, failure-mode PCA, theme flags, per-superfamily blind-spot scorecards, conformational-heterogeneity detection, a \u201cnew this week\u201d ribbon, and RCSB + DOI links on every structure |\n| [`/examples`](https://alphafraud.mdeller.com/examples) | Curated storytelling layer: a positive control (an excellent AlphaFold match), an auto-generated \u201cexample of the week\u201d plus a full weekly-release archive, and hand-authored deep-dive panels with interactive viewers (see below) |\n| [`/calculate`](https://alphafraud.mdeller.com/calculate) | Type any human PDB ID and get the same panel computed on demand: strict autocomplete (exists + human + has an AlphaFold model), full pipeline in an isolated subprocess, no LLM (see below) |\n| [`/archive`](https://alphafraud.mdeller.com/archive) | Every PDB release week since the 2018 cutoff, at weekly granularity, each linking to that week\u2019s structures and confidently-wrong count |\n| [`/stats`](https://alphafraud.mdeller.com/stats) | An exhaustive live Statistics page: themed panels for catch/quality, coverage & schedule, runs & compute, server (CPU/memory/uptime), storage & database, data sources & APIs, and traffic |\n| `/api/week/<label>`, `/api/leaderboard`, `/api/entry/<id>` | JSON for external tools |\n\n## \ud83d\udd2c Examples\n\nThe **[Examples](https://alphafraud.mdeller.com/examples)** tab turns raw catches into readable stories and deliberately shows both sides of AlphaFold\u2019s record, not just the failures:", "# GitHub README: ChemSage\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ChemSage, 2026\n\n### Chat CLI: Streaming output and UX polish (2026-06-25)\n\n**Streaming generation:**\n- Response tokens now stream live to the terminal in real-time using `mlx_lm.stream_generate` + Rich `Live`; raw text streams token-by-token during generation, then the final response re-renders as formatted Markdown \u2014 no more waiting behind a \"Thinking\u2026\" spinner\n- Falls back to blocking `generate` + spinner if `stream_generate` is unavailable\n\n**New slash commands:**\n- `/save` \u2014 exports the full conversation to a timestamped Markdown file on the Desktop (includes marcdeller.com attribution header)\n- `/info` \u2014 shows a panel with loaded model name/size, training config (iters, RSLoRA rank, lr, ctx), corpus stats (chunks, CSV files, MB), and embed model; clickable hyperlinks to [marcdeller.com](https://marcdeller.com) and [marc@marcdeller.com](mailto:marc@marcdeller.com)\n- `/retry` \u2014 re-generates the previous response using the same RAG context; replaces the last assistant turn in history\n\n**UX refinements:**\n- **Tab completion**: type `/` and press Tab for a dropdown of all slash commands; completion menu styled to match the ChemSage colour scheme\n- **RAG retrieval indicator**: `\ud83d\udd0d Retrieving context\u2026` spinner during retrieval, then `\ud83d\udcda N chunks retrieved` confirmation line before generation begins\n- **Context window indicator in toolbar**: bottom toolbar now shows `ctx N/3` alongside the turn counter so you can see how much of the 3-turn context window is filled\n- **`/reset` confirmation**: asks `Clear history? [y/N]` before wiping \u2014 no accidental data loss\n- **Generation stats**: now includes elapsed seconds \u2014 `\u26a1 N tok \u00b7 X.X tok/s \u00b7 X.Xs`\n- **Styled goodbye**: exit via `quit`/`q`/Ctrl-C shows a full-screen Panel with the ChemSage ASCII banner, \"Thank you for using ChemSage\", and clickable links to [marcdeller.com](https://marcdeller.com) and [marc@marcdeller.com](mailto:marc@marcdeller.com)\n- **Hyperlinks throughout**: banner, `/help`, `/info`, `/save` output, and goodbye Panel all include clickable `[link=...]` Rich markup pointing to marcdeller.com and marc@marcdeller.com\n\n**Earlier polish (same date):**\n- `prompt_toolkit` input with arrow-key history and styled `You:` prompt\n- Persistent bottom toolbar: `\u2b21 ChemSage \u00b7 model \u00b7 QLoRA \u00b7 RAG \u00b7 turn N \u00b7 ctx N/3`\n- Corpus tables: Rich `Table` with rounded cyan borders, single-line rows (ellipsis overflow), abbreviated headers, `\u2713`/`\u2717` booleans, method abbreviations\n- RDKit output in yellow-bordered `Panel`; retrieval errors in red-bordered `Panel`\n- Terminal title set to `ChemSage` on launch and `/clear`\n- Added `rich`, `prompt_toolkit`, `pyfiglet` to `requirements.txt`\n\n### Eval harness: R4 extended metrics (2026-06-25)\n\nExtended `eval/eval_chem.py` with new metrics that degrade gracefully to `n/a` when run against R1-R3 models/datasets, enabling a clean side-by-side comparison across all rounds:\n\n| New metric | What it measures | Why added |\n|---|---|---|\n| **Rounding precision** | MW 1 d.p. \u00b7 logP 2 d.p. \u00b7 TPSA 1 d.p. \u00b7 QED 2 d.p. | R4 drills exact decimal-place conventions |\n| **Code-then-quote** | prose values match what code blocks actually printed to stdout | Distinguishes quoting own code vs. recalling from memory |\n| **Extended executability** | all `python` blocks (RDKit + Plotly + Biopython) | R4 adds Plotly and Biopython code generation |\n| **Refusal accuracy** | out-of-scope prompts correctly declined | R4 adds explicit refusal training |\n| **QED range** | all stated QED values in [0, 1] | R4 adds QED computation training |\n\nBackwards compatibility: R1-R3 test sets have no refusal examples (refusal shows n/a), no QED mentions (QED shows n/a), code blocks may not print (code-then-quote shows n/a). All five new metrics return `(0, 0)` when inapplicable. The core 3-metric overall score formula is unchanged for direct cross-round comparison.\n\nNew scorecard features: per-class breakdown table (if test examples carry a `class` field), overall score card, collapsible response viewer in per-example table, extended exec breakdown by library category, mean latency in header, cleaner terminal summary with n/a annotation.\n\n### Round 3: Code-heavy dataset and extended training (2026-06-25)\n\n- Expanded SFT dataset from 1,500 to 5,000 examples (4,000 train) across 19 behaviour classes\n- 8 new generator classes targeting eval gaps: `single_property`, `full_profile`, `qed_score`, `molecular_formula`, `smiles_validation`, `property_comparison`, `corpus_drug`, `murcko_sar`\n- 51 seed SMILES + ~1,000 corpus drug examples from `approved_drugs.csv`\n- Configured for 1,500 iters; stopped early at iter 625 after val loss plateaued at ~0.060\n- Best val loss **0.054** at iter 625 (vs 0.347 for Round 2): **6.4x improvement**\n- Attempted 64 layers (swap thrashing at 31.5 GB) and batch_size=8 (4.5x slower); reverted to 32 layers / batch 4\n- Resumed from iter 100 checkpoint after battery shutdown; no data loss\n- Eval: SMILES 99%, Exec 99%, Fidelity 68% (vs R2: 100%, 41%, 10%)\n- Fused model saved to `models/chem_sage_32b_v3/` (17 GB)\n\n### Round 2: Qwen 2.5 32B upgrade (2026-06-24)\n\n- Upgraded base model from Qwen 2.5 7B to 32B (4-bit, MLX); trained with 32 adapted layers (vs 16 for 7B)\n- Best val loss 0.347 at step 675 (vs 0.389 for 7B): 10.8% improvement\n- Fused model saved to `models/chem_sage_32b_v2/` (17 GB); serves on 64 GB Apple Silicon at 24.9 GB peak memory\n- Eval shows 100% SMILES validity, 41% tool executability, 10% numerical fidelity (7B scored 0% across the board)\n- Added tqdm progress bar and live metrics to `eval/eval_chem.py`\n- System prompt injection in eval harness (test examples lack system messages)\n- Fixed `merge_export.py`: uses `sys.executable -m mlx_lm fuse` (corrected deprecated syntax), resolves adapter path relative to config directory\n\n### Round 1: Corpus fast-path and chat CLI", "# Therapeutic Enzymatic Compositions and Delivery Methods for the Degradation of Microplastics and Nanoplastics in Human Physiology\nAuthors: Deller, Marc C. (inventor)\nSource: United States Provisional Patent Application (Elora Therapeutics), 2025\n\n**Inventor:** Marc C. Deller (Chief Scientific Officer and co-founder, Elora Therapeutics). **Document type:** United States provisional patent application (2025).\n\n## Cross-Reference to Related Applications\n\nNot applicable.\n\n## Federally Sponsored Research and Development\n\nNot applicable.\n\n## Joint Research Agreement\n\nNot applicable.\n\n## Reference to a Sequence Listing, Table, or Computer Program Listing Appendix\n\nNot applicable.\n\n## Prior Art\n\nThis invention explicitly differentiates itself from industrial polyethylene terephthalate hydrolase (PETase) enzyme patents, including US 20240263155 A1 (Hubei University), which focuses on modifying the enzyme PETase (e.g., N212A, N277A mutations) for increased polyethylene terephthalate (PET) degradation in industrial applications. Unlike this prior work, the present invention is designed for human therapeutic use, employing PETase variants specifically optimized for stability in human physiology and systemic bioavailability. Furthermore, this invention incorporates additional enzyme systems, i.e., mono(2-hydroxyethyl) terephthalate hydrolase (MHETase), CotA-Laccase enzyme, and Horseradish Peroxidase (HRP), as well as novel delivery mechanisms, i.e., PEGylation, lipid nanoparticles, encapsulation, tailored for microplastic degradation in the human body rather than industrial PET recycling.\n\nThis invention explicitly excludes environmental, industrial, fungal recombinant production, biofuel, paper production, and wastewater treatment applications previously covered by existing patents, as these focus on non-therapeutic uses and lack the novel enzyme stabilization and delivery mechanisms proposed herein. These prior patents address distinct applications unrelated to enzymatic microplastic and nanoplastic degradation within human physiology, reinforcing the uniqueness of the present invention. Specifically, this invention differs in its optimization of enzyme formulations for human therapeutic use, including systemic circulation, immune compatibility, and targeted metabolic clearance pathways, which have not been claimed or disclosed in prior art, including but not limited to:\n\n- WO2022/076380A2 \u2013 Plastic degradation via engineered PETase (University of Texas System).\n- US 9,840,726 \u2013 CotA-Laccase for biomass saccharification (MetGen Oy).\n- US 10,087,582 \u2013 CotA-Laccase in paper pulping applications (MetGen Oy).\n- WO2020/113962 \u2013 CotA-Laccase for mycotoxin degradation in food and feed applications (China Agricultural University).\n- US 9,540,617 \u2013 Recombinant HRP optimized for fungal expression (Meiji Seika Pharma).\n- EP 2768953 \u2013 Engineered HRP isoenzymes with improved properties for wastewater treatment (Graz University of Technology).\n- US 20240263155 A1 \u2013 Engineered PET hydrolase with N212A and N212A/N277A mutations for enhanced PET degradation in industrial applications (Hubei University).\n\nMicroplastics and nanoplastics\u2014plastic particles ranging from 1 nm to 5 mm\u2014are pervasive contaminants increasingly identified within human biological systems, including tissues, organs, and bodily fluids. Originating primarily from synthetic polymers such as PET, polyethylene (PE), polypropylene (PP), and polystyrene (PS), these particles pose significant health risks, including chronic inflammation, endocrine disruption, oxidative stress, and potential carcinogenic effects.\n\nCurrent mitigation strategies primarily target environmental sources and industrial recycling, leaving a critical gap in therapeutic solutions for actively degrading and clearing internalized microplastics and nanoplastics within humans.\n\nEnzymatic biodegradation represents a promising approach, as evidenced by PETase and MHETase enzymes demonstrating substantial PET degradation under laboratory conditions. Engineered PETase variants, optimized for increased stability and catalytic efficiency at human physiological temperatures (~37 \u00b0C), along with fusion enzymes combining PETase and MHETase, provide enhanced therapeutic potential. Oxidative enzymes, specifically bacterial-origin CotA-Laccase and engineered HRP, have shown the capability to degrade PE, PP, and PS plastics effectively. Engineered modifications\u2014including site-directed mutagenesis, thermostabilization mutations, enhanced redox potentials, and improved substrate specificity\u2014enable therapeutic application in complex physiological environments.\n\nChallenges remain, including enzyme stability at physiological temperatures, immunogenicity, and metabolic clearance of degradation products such as terephthalic acid (TPA), ethylene glycol (EG), and oxidized polyolefin monomers. Metabolic studies confirm hepatic and renal pathways efficiently process these byproducts under controlled therapeutic dosing.\n\nThe existing aforementioned patents (WO2022/076380A2, US 20240263155 A1, WO2020/113962, US 9,840,726, US 10,087,582, US 9,540,617, EP 2768953, and US 20240336900 A1) primarily address industrial, environmental, or food safety enzyme applications, failing to address therapeutic-specific enzymatic microplastic and nanoplastic degradation within human physiology.\n\nThis invention distinctly provides therapeutic enzyme compositions, novel pharmaceutical delivery methods, and specific metabolic clearance strategies addressing stability, bioavailability, immunogenicity reduction, and safe human application.\n\n## Summary of the Invention", "# 2013 \u2013 Crystal structure of a soluble cleaved HIV-1 envelope trimer\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-11 on marcdeller.com (categories: papers).*\n\n* [Link to Paper](https://marcdeller.com/wp-content/uploads/2025/09/nihms-540638.pdf)\n\n* **Structural Architecture**: The BG505 SOSIP.664 crystal structure reveals the native-like trimeric organization with gp120 V1/V2/V3 loops forming a stabilizing apex\n* **Antibody Recognition**: PGT122 epitope mapping demonstrates complex interactions involving both protein residues and multiple glycans for broad neutralization\n* **Membrane Integration**: Recent structures show how the membrane-proximal external region contributes to overall trimer stability and conformational integrity\n* **Glycan Shield Dynamics**: Advanced imaging reveals the protective glycan coat\u2019s heterogeneous nature and its role in immune evasion strategies\n* **Vaccine Applications**: Structure-guided modifications have led to stabilized trimers with improved immunogenic properties in clinical development\n* **Conformational Flexibility**: Asymmetric trimer structures provide insights into the dynamic nature of envelope proteins during viral entry processes\n\n1. **Crystal structure of a soluble cleaved HIV-1 envelope trimer:** [Julien, J.P., et al., Science, 2013](https://www.science.org/doi/abs/10.1126/science.1245625)\n2. **The membrane-proximal external region of human immunodeficiency virus envelope glycoprotein trimers:** [Chen, L., et al., Communications Biology, 2025](https://pmc.ncbi.nlm.nih.gov/articles/PMC11910548/)\n3. **Structure-guided envelope trimer design in HIV-1 vaccine development:** [Sanders, R.W., et al., Journal of the International AIDS Society, 2021](https://pmc.ncbi.nlm.nih.gov/articles/PMC8606863/)\n4. **Visualization of the HIV-1 Env glycan shield across scales:** [Stewart-Jones, G.B., et al., Proceedings of the National Academy of Sciences, 2020](https://www.pnas.org/doi/10.1073/pnas.2000260117)\n5. **Vaccination with mRNA-encoded membrane-anchored HIV envelope trimers:** [Saunders, K.O., et al., Nature Medicine, 2025](https://pubmed.ncbi.nlm.nih.gov/40737434/)\n6. **Structure of the membrane proximal external region of HIV-1 gp41 in micelles:** [Fu, Q., et al., Proceedings of the National Academy of Sciences, 2018](https://www.pnas.org/doi/10.1073/pnas.1807259115)\n\n**Unlocking the HIV-1 Envelope: Structural Insights from the Cleaved Trimer Crystal Structure**\n\nThe human immunodeficiency virus type 1 remains one of the most formidable challenges in modern medicine, with its envelope glycoprotein trimer serving as both the key to viral entry and the primary target for neutralizing antibodies. Recent structural discoveries have revolutionized our understanding of this molecular machinery, particularly through the landmark crystal structure determination of the soluble cleaved HIV-1 envelope trimer. This breakthrough has provided unprecedented atomic-level detail of how the virus presents itself to the host immune system and how it successfully evades neutralizing responses.\n\nThe BG505 SOSIP.664 gp140 trimer represents a remarkable achievement in structural virology, capturing the envelope glycoprotein in its near-native, prefusion state at 4.7 angstrom resolution. This structure reveals the intricate architecture of the trimeric spike, showing how the gp120 and gp41 subunits interact to maintain stability while preserving the conformational flexibility required for viral entry. The complex with the broadly neutralizing antibody PGT122 illuminates critical aspects of immune recognition, demonstrating how potent antibodies can penetrate the virus\u2019s sophisticated defense mechanisms.\n\nThe structural analysis reveals several key features that define the envelope trimer\u2019s function and vulnerability. The prefusion state of gp41 maintains the metastable conformation necessary for membrane fusion, while the gp120 V1/V2/V3 loops form a protective apex around the threefold axis. This arrangement serves dual purposes: stabilizing the overall trimeric structure and shielding conserved regions from antibody recognition. The complete epitope mapping of PGT122 shows its interaction with gp120 V1 and V3 regions, along with multiple surrounding glycans, highlighting the complex nature of broadly neutralizing antibody recognition.\n\nRecent developments have built upon this foundational structure to advance HIV-1 vaccine design significantly. The membrane-proximal external region has emerged as a critical structural element, with new studies revealing how it interacts with gp120 C-termini and contributes to trimer stabilization. Advanced cryo-electron microscopy techniques have captured the dynamic nature of the glycan shield, showing how these sugar modifications create both barriers and opportunities for immune recognition. The visualization of asymmetric envelope conformations has provided insights into the conformational changes that occur during viral entry.\n\nContemporary research has demonstrated the importance of envelope stabilization strategies in vaccine development. Modified trimers incorporating additional disulfide bonds and stabilizing mutations have shown improved antigenic profiles and enhanced immunogenicity in preclinical studies. The development of membrane-anchored versions delivered through messenger RNA platforms represents a significant advancement, addressing limitations associated with soluble protein immunogens and reducing off-target immune responses. These innovations have led to clinical trials demonstrating the induction of tier 2 neutralizing antibodies, marking substantial progress toward effective HIV-1 vaccination.", "# Unlocking the Power of Protein Engineering: FDA-Approved Enzyme Therapies Transform Rare Disease Treatment\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nMutation strategies to enhance substrate affinity and catalytic efficiency enable the development of enzymes with improved therapeutic profiles. [Human arginase-I engineering](https://pmc.ncbi.nlm.nih.gov/articles/PMC8714691/) through cobalt substitution for the native manganese cofactor decreases Km from 1-5 mM to approximately 0.2 mM, achieving a 15-fold improvement in IC50 values for cancer cell cytotoxicity despite serum arginine concentrations of only 80-120 \u03bcM. Similarly, [engineering of Pseudomonas fluorescens kynureninase](https://pmc.ncbi.nlm.nih.gov/articles/PMC8714691/) demonstrates 500-700-fold improved catalytic efficiency over human kynureninase, enabling 100-fold reduction in serum kynurenine concentrations and 45% improvement in median survival in murine tumor models. These approaches highlight the potential for heterologous enzymes when combined with deimmunization strategies including [epitope engineering that reduces B-cell recognition](https://pmc.ncbi.nlm.nih.gov/articles/PMC8714691/) by 38.5-99% through targeted mutagenesis or C-terminal steric hindrance.\n\nExamining the relative contribution of enzyme therapies to the FDA approval landscape reveals significant investment in orphan disease treatments despite representing a small fraction of total approvals. With [50 novel drugs authorized in 2024](https://www.nature.com/articles/d41573-024-00224-w) and a 10-year rolling average of 46.5 approvals annually, enzyme therapies constitute approximately 1-3% of annual novel drug approvals. However, their clinical impact far exceeds this numerical representation, with [enzyme replacement therapy market projections reaching USD 21.62 billion by 2035](https://www.futuremarketinsights.com/reports/enzyme-replacement-therapy-market) at a compound annual growth rate of 6.6%. The [orphan drug designation](https://coherentmarketinsights.com/market-insight/enzyme-replacement-therapy-market-4783) provides crucial incentives including tax credits, grant availability, seven years of marketing exclusivity, and expedited FDA review timelines, enabling commercial viability for ultra-rare conditions affecting fewer than 200,000 individuals in the United States.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics; supervisor Prof. E. Yvonne Jones), 1999\n\nNative and multiple wavelength anomalous dispersion (MAD) data were collected from single propane frozen crystals on beam line BL19 of the European Synchrotron Radiation Facility (ESRF), Grenoble, France. MAD data were collected at wavelengths of 1.00914 \u00c5, 0.826516 \u00c5 and 1.00669 \u00c5, to a resolution of 2.6 \u00c5. Data were collected in a nitrogen stream at 100 K and data recorded as 1\u00b0 oscillations using a MAR 345 image plate detector (MAR Research, Hamburg, Germany) in 18 cm mode. Exposure times were typically 40 seconds with the crystal to detector distance set to 225 mm. Native data were collected at the same beamline at a wavelength of 0.98 \u00c5 to a resolution 2.2 \u00c5. Data were collected in a nitrogen stream at 100 K and data recorded as 1\u00b0 oscillations using a MAR 345 image plate detector (MAR Research, Hamburg, Germany) in 30 cm mode. Exposure times were typically 60 seconds and the crystal to detector distance was set to 225 mm.\n\nHeavy atom derivative screening was carried out using 'in house' radiation and synchrotron radiation at station 9.6 of the Synchrotron Radiation Source (SRS), Daresbury, U.K. 'In house' data were collected using CuK\u03b1 X-rays (1.54 \u00c5) produced by a rotating anode (Rigaku RU2000) fitted with a graphite monochromator. Data were collected at 100 K or at room temperature. Data were recorded as 1\u00b0 oscillations using a 18 cm MAR image plate detector (MAR Research, Hamburg, Germany). Exposure times were typically 2000 seconds and the crystal to detector distance was set to 200 mm. Data were collected at station 9.6 of the SRS using a wavelength of 0.87 \u00c5. Data were collected at room temperature as 2\u00b0 oscillations using a charged coupled device (CCD) detector (Advanced Detector Systems Corporation, USA). Exposure times were typically 40 seconds with the crystal to detector distance set to 300 mm.\n\n#### 5.2.4 Data Processing\n\nData were autoindexed, integrated and corrected for Lorentz and polarizing effects with the program DENZO (Otwinowski and Minor, 1997). Scaling, merging and addition of partially recorded reflections was carried out using the program SCALEPACK (Otwinowski and Minor, 1997). General processing and file manipulation was carried out using the CCP4 suite of programs (Collaborative Computational Project, 1994). SCALEPACK files were converted to CCP4 format using the program SCALEPACK2MTZ and intensities were converted to amplitudes using the program TRUNCATE (French and Wilson, 1978).\n\n#### 5.2.5 Heavy Metal Derivatisation\n\nHeavy metal derivative screening was carried out as detailed in Table 5-1 below. Heavy metal solutions were prepared at twice the required concentration in mother liquor consisting of 30 % (w/v) PEG 35000, 0.25 M sodium acetate, pH 7.5. Solutions were left overnight, with agitation, to equilibrate. Crystals were transferred to fresh 2 \u03bcl drops of mother liquor and 2 \u03bcl of heavy metal solution was added. Drops were sealed over 0.5 ml reservoirs of mother liquor and the crystals incubated at room temperature for between 1 and 72 hours.\n\n| Heavy Metal Reagent | Concentration | Duration of Soak |\n|---|---|---|\n| Trimethyllead Acetate (TA) | 20\u201350 mM | 16\u201372 hours |\n| Lead Acetate (LA) | 5 mM | 1\u201324 hours |\n| Potassium Mercuric Iodide (PMI) | 2 mM | 16\u201324 hours |\n| Ethylmercury Phosphate (EMP) | 1\u201340 mM | 16\u201324 hours |\n| Uranyl Acetate (UA) | 1\u20135 mM | 1\u201372 hours |\n| Uranyl Nitrate (UN) | 1\u20135 mM | 1\u201372 hours |\n| Cis-dichlorodiamine Platinum II (CP) | 5 mM | 16\u201372 hours |\n| Mercury meso-(2,3-dimethoxytetramethylene)acetate (BAKERS) | 5 mM | 16\u201372 hours |\n| Tetrakis(acetomercuric)methane (TAMM) | 5 mM | 16\u201372 hours |\n| Sodium Gold Chloride (SGC) | 5 mM | 1\u201372 hours |\n| Di-potassium Platinum Nitrate (PN) | 5 mM | 1\u201372 hours |\n\n*Table 5-1.* Table showing summary of heavy metal reagents used for derivative screening of \u0394188_OSM. All soaks were carried out for between 1 and 72 hours in equilibrated mother liquor containing 30 % (w/v) PEG 35000, 0.25 M sodium acetate, pH 7.5 and corresponding heavy metal.\n\nThe state of derivatisation was analysed by collecting ~45\u00b0 of data (as 1\u00b0 or 2\u00b0 oscillations) and manual inspection of anomalous difference Patterson and Fourier maps, calculated using the program FFT (Ten Eyck, 1973). Maps were contoured using the program GROPAT (Esnouf, 1993).\n\n#### 5.2.6 Multiple Wavelength Anomalous Dispersion\n\nMAD data were collected to a resolution of 2.6 \u00c5 from a single EMP derivatised propane cooled crystal on beam line BL19 at the ESRF, Grenoble, France (see 5.2.3). The mercury LIII edge of the derivative crystal was characterised by a fluorescence scan and three wavelengths were selected for data collection. The first wavelength was selected just below the LIII edge at a wavelength of 1.00669 \u00c5 (f\u2032 = \u221214.8 e\u2212, f\u2033 = 10.1 e\u2212, E = 12.3163 keV). The second wavelength was selected at the inflexion point at a wavelength of 1.00914 \u00c5 (f\u2032 = \u221220.3 e\u2212, f\u2033 = 10.2 e\u2212, E = 12.2863 keV). A remote wavelength was also selected above the absorption edge at 0.826516 \u00c5 (f\u2032 = \u22127.4 e\u2212, f\u2033 = 11.8 e\u2212, E = 15.0009 keV). Data were collected as 1\u00b0 oscillations over the same 90\u00b0 phi range for each wavelength.\n\nAll data was autoindexed and processed as previously (see 5.2.4) except scaling and merging of MAD data at each wavelength was carried out separately, with preservation of Bijvoet pairs, using the program SCALEPACK (Otwinowski and Minor, 1997). Structure factor amplitudes were calculated using the program TRUNCATE (French and Wilson, 1978) and the MAD data were scaled, with reference to the inflexion data, using the program SCALEIT (Howell and Smith, 1992). Finally, normalised anomalous scattering magnitudes were estimated, to aid in analysis of anomalous scatterer positions, using the program REVISE (Woolfson et al., 1997).", "# Precision Oncology: Integrating AI and High-Throughput Screening to Fight Cancer\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* The combination of high-throughput screening (HTS) and artificial intelligence (AI) is creating a powerful engine for oncology drug discovery.\n* AI algorithms analyze vast HTS datasets to quickly identify promising drug candidates and novel therapeutic targets, increasing the efficiency of research.\n* This integrated approach is helping to tackle \u201cundruggable\u201d cancer targets and shorten the timeline for developing next-generation precision therapies.\n\n1. **Top 5 Drug Discovery Trends 2025 Driving Breakthroughs:**<https://www.pelagobio.com/cetsa-drug-discovery-resources/blog/drug-discovery-trends-2025/>\n2. **Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor:**<https://pubs.acs.org/doi/abs/10.1021/acs.jmedchem.4c02662>\n3. **AI in Drug Discovery: From Promise to Platform:**<https://www.pelagobio.com/cetsa-drug-discovery-resources/blog/drug-discovery-trends-2025/>\n\nThe fight against cancer has become a race for precision. Broad-spectrum chemotherapies, while effective for some, often act as blunt instruments, causing significant collateral damage to healthy cells. The goal of modern oncology is to develop targeted therapies that attack the specific molecular drivers of a patient\u2019s tumor. However, identifying these drivers and finding drugs to target them is a monumental task. The genetic complexity of cancer means we must screen thousands of potential compounds against numerous targets, a process that is both time-consuming and expensive with traditional methods.\n\nThe convergence of two powerful technologies\u2014high-throughput screening (HTS) and artificial intelligence\u2014is accelerating this quest for precision. HTS platforms, using advanced robotics and automation, can test hundreds of thousands of compounds for biological activity in a single day. But this firehose of data is only useful if it can be interpreted effectively. This is where AI comes in. Machine learning models can sift through vast HTS datasets to identify subtle patterns and prioritize the most promising drug candidates, doing in hours what would take humans years. This synergy creates a powerful engine for discovery.\n\nThis integrated approach is revolutionizing oncology drug development. For notoriously \u201cundruggable\u201d targets like KRAS, which drive a significant percentage of cancers, combining structural biology, AI, and HTS has led to the discovery of new, orally bioavailable inhibitors. By using AI to guide HTS campaigns and predict which compounds are most likely to succeed, researchers can focus their efforts, shorten timelines, and reduce the high attrition rates that plague oncology research. This powerful combination allows us to move faster and with greater accuracy, designing smarter drugs that are tailored to the unique molecular profile of each cancer and bringing the promise of precision medicine closer to reality.", "# Overcoming Biological Barriers: The Evolution of Oral Protein and Peptide Drug Delivery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Key Concept | Description | Key Reference |\n| --- | --- | --- |\n| Permeation Enhancers | SNAC and 5-CNAC transiently open tight junctions to improve peptide absorption | [Chen et al., Theranostics, 2022](https://pmc.ncbi.nlm.nih.gov/articles/PMC8771547/) |\n| Cell Targeting | M-cell and enterocyte receptors mediate endocytosis for 10-fold bioavailability enhancement | [Chen et al., Theranostics, 2022](https://pmc.ncbi.nlm.nih.gov/articles/PMC8771547/) |\n| Device-Based Delivery | Robotic pills achieve 65% bioavailability by direct intestinal injection | [Johnson et al., Nature Reviews Drug Discovery, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC10990675/) |\n| Chemical Modification | PEGylation and lipidization improve stability and membrane penetration | [Chen et al., Theranostics, 2022](https://pmc.ncbi.nlm.nih.gov/articles/PMC8771547/) |\n| Clinical Progress | Semaglutide and octreotide capsules approved; oral insulin in phase III trials | [Zhang et al., Journal of Controlled Release, 2021](https://pmc.ncbi.nlm.nih.gov/articles/PMC8424290/) |", "# GitHub README: ALPHABETTI\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ALPHABETTI, 2026\n\n```\nH_i      = -sum_a p_a log2(p_a)      Shannon entropy, bits\nR_i      = log2(20) - H_i            information content, maximum 4.322 bits\nheight_a = p_a * R_i                 per-letter height, bits\n```\n\nLetters are stacked along the residue's own side-chain direction, each standing on the one below, so **the total height of a stack is that position's information content**. A unit test checks this against hand-calculated values, because it is the difference between a quantitative picture and a decorative one.\n\nNo small-sample entropy correction is applied. The usual correction compensates for estimating a distribution from a finite number of aligned sequences; here there is no alignment and no sample, so correcting would subtract a bias that does not exist.\n\n## \ud83d\udd2c Scientific choices, and what they were checked against\n\n| Quantity | Method | Validation |\n|---|---|---|\n| Structure | ESMFold (`facebook/esmfold_v1`) | **0.84 \u00c5 CA RMSD** for ubiquitin against the 1UBQ crystal structure |\n| Expectation | ESM-2 650M (`facebook/esm2_t33_650M_UR50D`), **masked** marginals | Unmasked marginals return the wild type at 100 % of positions; see below |\n| pLDDT | ESMFold's B-factor column | Normalised and asserted to 0-100; a 0-1 scale renders as a uniformly unconfident protein |\n| Relative SASA | FreeSASA, 1.4 \u00c5 probe, over Tien et al. (2013) theoretical maxima | Total SASA 4804 \u00c5\u00b2 for 1UBQ against a published value of about 4800 |\n| Secondary structure | DSSP where present, else P-SEA from CA positions | **83.7 %** three-state agreement with DSSP over 805 residues of ten chains |\n| Virtual C\u03b2 | Standard tetrahedral construction from N, CA, C | **0.13 \u00c5 mean** deviation from the real C\u03b2 across all 70 non-glycine residues of 1UBQ |\n| Variant score | `log p(mutant) - log p(wild type)`, ESM-1v convention | Wild-type entry is exactly zero by construction; **negative means deleterious** |\n\nGlycine has no C\u03b2, so one is constructed and glycines are **never skipped**: a hole at every glycine would be a hole at exactly the positions where backbones turn.\n\n### Masked marginals are the default, and that is a change from the brief\n\nThe original specification made masked marginals an opt-in toggle, \"L times slower\". They are, and on a CPU that would settle it. On a GPU it does not, and the cheap option is actively misleading. Measured on ubiquitin:\n\n| | Wild-type marginals | Masked marginals |\n|---|---|---|\n| Top-1 equals wild type | **100 %** | 80 % |\n| Mean information content | 4.01 bits | **3.20 bits** |\n| Language model time | 0.03 s | 0.42 s (about 2.2 s at 400 residues) |\n\nAn unmasked forward pass can read each residue off its own input, so it returns near-maximal confidence everywhere and GIBBERISH becomes a 3D rendering of the input sequence. Set `ALPHABETTI_MASKED_MARGINALS=0` to compare.\n\n### Where it fails\n\nESM-2 has almost no evolutionary signal for avGFP: masked marginals recover the wild-type residue at **10 %** of positions against **8 %** for a randomly shuffled version of the same sequence, and mean information content is 0.27 of a possible 4.322 bits. Because ESMFold's trunk *is* ESM-2, the structure fails with it rather than independently, at mean pLDDT 42.8. The two never contradict each other, so nothing on screen flags it except the confidence number. **Check the pLDDT before believing a picture.**\n\n### Examples name a residue range, deliberately\n\nA UniProt sequence is the **precursor**, and folding it whole gives something nobody means. P0CG48 is nine exact tandem copies of ubiquitin (and the repeats let a masked model copy each position from its neighbours, driving information content to near-maximal everywhere). P00698 carries an 18-residue signal peptide that is cleaved in vivo, has no structure of its own, and trails off the fold as a disordered tail. Every example therefore names a range taken from UniProt's own Chain feature. Removing lysozyme's signal peptide raised its mean pLDDT from **91.0 to 95.1**.\n\n## \ud83c\udfa5 The camera turns nothing; the protein does\n\nDragging vertically used to die about halfway across the viewport. The cause is in three.js OrbitControls' own arithmetic rather than anything app-specific, which is why sibling apps had it too:\n\n```js\nrotateUp( 2 * Math.PI * deltaY / element.clientHeight * rotateSpeed )\n```\n\nA full canvas height of vertical drag asks for **360\u00b0** of polar rotation, while the polar angle is clamped to **[0, \u03c0] = 180\u00b0**. Half a viewport therefore exhausts the entire permitted range and the camera pins at the pole. Measured: the polar angle stopped changing at **59%** of a vertical drag. Halving the sensitivity only moved that to 91%, because a spherical-coordinate camera *has* a pole and the clamp exists to protect its degenerate up vector.\n\nALPHABETTI now uses `StageCamera`, taken unchanged from ButtFold (itself ported from PhoneFold's Swift). **The camera is fixed on +Z and the protein carries a quaternion**, so there is no pole to protect and it tumbles freely. Drag increments are premultiplied about the *screen* axes, which keeps \"drag right turns right\" true even upside down. Measured after: a full vertical drag turns 137\u00b0 with zero dead steps, and dragging again the same way turns another 137\u00b0.\n\nThe orbit resumes **8 seconds** after you stop, not 2.5: a view you have just set should not start sliding away while you are still looking at it.\n\n## \ud83c\udfd7\ufe0f Architecture\n\n```\nBrowser (three.js r169, vanilla ES modules, no build step)\n        |  JSON over fetch()\n        v\nFlask + Gunicorn on the mdeller.com droplet\n        |-- SQLite cache, keyed on sha256(sequence)\n        |-- bounded thread pool, job state in SQLite\n        v  HTTPS\nHugging Face ZeroGPU Space\n        |-- ESMFold          -> coordinates + pLDDT\n        `-- ESM-2 650M       -> a 20 x L probability matrix\n```", "# GitHub README: smiles2grid\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/smiles2grid, 2026\n\n> **Turn a JSON of SMILES into a paginated, property-annotated PDF grid with RDKit.**\n\n    \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/smiles2grid\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/smiles2grid</a></td>\n</tr>\n</table>\n\n---\n\nRender SMILES strings from a JSON file into a paginated PDF grid using RDKit.\n\nWhy it matters: after a generative or screening run you often have hundreds of SMILES and no fast way to actually look at them together, compare scaffolds, or pull a labelled subset for a slide. smiles_grid renders them into clean 5x5 PDF pages with structure labels tied to the original input order, computes a full property panel (MW, cLogP, HBD/HBA, TPSA, and more) into a companion CSV, and can highlight near-neighbours of a query structure in red. It is useful for medicinal chemists and computational scientists who need a quick, shareable visual of a compound set with the numbers attached: ideal for triage, SAR discussions, and decision-making meetings.\n\n## What it does\n\n- Extracts SMILES strings from a nested JSON file.\n- Preserves the original input order for labeling and selection.\n- Renders molecules into a boxed 5x5 grid, 25 compounds per page.\n- Supports optional similarity search against a query SMILES.\n- Supports range-based selection such as `all`, `1-25`, `50-75`, or `1,5,10-20`.\n- Writes a CSV summary for every rendered record.\n- Highlights closely related structures within a page in red.\n\n## Example input\n\n```bash\npython smiles_grid.py scratch_Baricitinib_sim.out.20260530_0219.json\n```\n\nThe script prompts by default. You can also run it non-interactively with `--no-prompt` and selection arguments.\n\n## Features\n\n### PDF output\n\n- Single PDF output.\n- 5x5 bordered panels per page.\n- Structure labels based on original JSON order.\n- Clean footer text with predicted properties.\n- Structure drawing scaled to fit each panel.\n\n### CSV output\n\nA CSV file is written alongside the PDF and includes:\n\n- output index.\n- original index.\n- source index.\n- label.\n- SMILES and canonical SMILES.\n- formula.\n- MW, cLogP, HBD, HBA, TPSA.\n- rotatable bonds, rings, heavy atoms, fraction sp3, aromatic rings, hetero atoms.\n- input JSON path.\n- PDF filename.\n- selection mode.\n\n## Usage\n\n### Interactive mode\n\n```bash\npython smiles_grid.py input.json\n```\n\nAt runtime the script prompts for:\n\n1. An optional query SMILES for similarity search.\n2. A similarity cutoff percentage.\n3. If no query is provided, a range selection such as `all`, `1-25`, or `50-75`.\n\n### Similarity search\n\n```bash\npython smiles_grid.py input.json --query-smiles \"CCO\" --similarity 75\n```\n\n### Range selection\n\n```bash\npython smiles_grid.py input.json --range 1-25\npython smiles_grid.py input.json --range all\npython smiles_grid.py input.json --range 1,5,10-20\n```\n\n### Non-interactive mode\n\n```bash\npython smiles_grid.py input.json --no-prompt --range 1-25\n```\n\n## Output files\n\nBy default the script writes:\n\n- `input_grid.pdf`\n- `input_summary.csv`\n\nYou can override the PDF filename with `--output`.\n\n## Requirements\n\n- Python 3.9+\n- RDKit\n- ReportLab\n- Pillow\n- cairosvg, if your RDKit build does not include Cairo drawing support\n\n## Installation\n\n### Conda\n\n```bash\nconda install -c conda-forge rdkit reportlab pillow cairosvg\n```\n\n### Pip-style environments\n\nIf you already have RDKit available in your environment:\n\n```bash\npip install reportlab pillow cairosvg\n```\n\n## Command-line options\n\n- `--query-smiles`: similarity search query.\n- `--similarity`: similarity cutoff as `0-100` or `0-1`.\n- `--range`: selection by range or comma-separated ranges.\n- `--no-prompt`: disable interactive prompting.\n- `--max-pages`: render only the first N pages for debugging.\n- `--label-prefix`: change the label prefix, default `J`.\n- `--output`: choose the PDF filename.\n\n## Notes\n\n- Numbering is tied to the original JSON order, not to whether a SMILES can be rendered.\n- Failed renderables are skipped visually but do not shift labels.\n- The script keeps the input order for selection and CSV reporting.\n- Similarity highlighting is page-local and uses red bond highlighting.\n\n## Example workflow\n\n1. Place the JSON file in your working directory.\n2. Run the script.\n3. Enter a query SMILES or press Enter.\n4. Enter a similarity cutoff or press Enter.\n5. Or press Enter again and provide a range such as `all` or `1-25`.\n6. Open the generated PDF and CSV.\n\n## License\n\nAdd your preferred license here.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/smiles2grid\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/smiles2grid</a></td>\n</tr>\n</table>", "# 2014 \u2013 Approaches to automated protein crystal harvesting\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nPerhaps most significantly, the integration of these diverse technologies into comprehensive automated pipelines is becoming reality. The Photon Factory\u2019s PXS2 system demonstrates how automated crystallization screening can seamlessly connect with in situ data collection capabilities. This integration allows researchers to progress from protein solution to structural data with minimal manual intervention, representing a true gene-to-structure automation pipeline.\n\nThe economic impact of these technological advances extends far beyond academic research. The protein crystallization market, valued at $1.84 billion in 2023, is projected to reach $3.84 billion by 2033, driven largely by automation technologies and the pharmaceutical industry\u2019s increasing reliance on high-throughput screening. This growth reflects both the maturation of automated harvesting technologies and their increasing adoption in drug discovery pipelines where hundreds of protein targets may be screened simultaneously.\n\nLooking toward the future, emerging technologies promise even more sophisticated approaches to crystal harvesting. MEMS-based innovations like the RodBot system utilize magnetized microscale agents to manipulate crystals through hydrodynamic forces, offering gentle handling capabilities that surpass traditional mechanical methods. Meanwhile, advances in detector technology, particularly pixel-array detectors with millisecond readout times, are enabling new data collection strategies that may eventually eliminate traditional crystal mounting requirements altogether.\n\n| Key Concept | Description | Key References |\n| --- | --- | --- |\n| Semi-Automated Harvesting | Operator-assisted systems combining human crystal identification with robotic precision for manipulation and cryoprotection | [Viola, R., et al., Journal of Applied Crystallography, 2007](https://journals.iucr.org/paper?kk5010) |\n| Micromanipulation Robotics | Six-axis anthropomorphic robots with submicron precision capable of handling crystals as small as 10 micrometers | [Deller, M.C., et al., Acta Crystallographica Section F, 2014](https://pmc.ncbi.nlm.nih.gov/articles/PMC3936438/pdf/f-70-00133.pdf) |\n| MEMS Microgrippers | Two-fingered micro-gripping devices fabricated using MEMS technology with controlled pressure application and direct X-ray compatibility | [Heidari Khajepour, M.Y., et al., Acta Crystallographica Section D, 2013](https://pubmed.ncbi.nlm.nih.gov/23519413/) |\n| Acoustic Droplet Ejection | Non-contact crystal transfer using focused acoustic energy to eject nanoliter droplets at rates of hundreds per second | [Soares, A.S., et al., Acta Crystallographica Section D, 2013](https://pmc.ncbi.nlm.nih.gov/articles/PMC3747951/) |\n| Optical Trapping | Contactless crystal manipulation using focused laser beams for damage-free handling of delicate specimens | [Deller, M.C., et al., Acta Crystallographica Section F, 2014](https://pmc.ncbi.nlm.nih.gov/articles/PMC3936438/pdf/f-70-00133.pdf) |\n| AI Crystal Detection | Machine learning algorithms for automated crystal identification with <5% error rates using extensive training datasets | [Hiraki, M., et al., Journal of Synchrotron Radiation, 2025](https://journals.iucr.org/paper?he5685) |\n| High-Throughput Screening | Integration of harvesting systems with chemical library screening for up to 1,728 compounds per microplate | [Teplitsky, E., et al., Journal of Structural Biology, 2015](https://pubmed.ncbi.nlm.nih.gov/26027487/) |\n| In Situ Data Collection | Direct X-ray data collection from crystals in crystallization plates without traditional mounting procedures | [Hiraki, M., et al., Journal of Synchrotron Radiation, 2025](https://journals.iucr.org/paper?he5685) |", "# GitHub README: ALPHABETTI\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ALPHABETTI, 2026\n\n**A**mino-acid **L**etters **P**lotted **H**elically **A**s **B**ackbone-**E**mbedded **T**ext in **T**hree-dimensions and **I**nteractive\n\n> **A protein sequence logo, in three dimensions, wrapped around its own predicted structure.**\n\n[](https://alphabetti.mdeller.com)\n\n| | |\n|---|---|\n| \ud83c\udf10 **App** | [alphabetti.mdeller.com](https://alphabetti.mdeller.com) |\n| \ud83e\uddec **Fold service** | [Dellboy/alphabetti-fold](https://huggingface.co/spaces/Dellboy/alphabetti-fold) on Hugging Face ZeroGPU |\n| \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) |\n| \ud83d\udc64 **Author** | Marc C. Deller, D.Phil. |\n\n**ALPHABETTI** draws a protein's own single-letter amino acid codes in 3D, positioned and oriented along an ESMFold-predicted backbone. Instead of a cartoon or a stick model, you get the letters themselves: stacked at each position by how strongly a protein language model expects them, tallest at the bottom, standing on the side chain.\n\n**Why it matters:** a conventional sequence logo tells you which positions are conserved, but it does so on a flat strip that has been divorced from the structure, and it gets there through a multiple sequence alignment, inheriting every bias in the search-align-trim pipeline that produced it. ALPHABETTI has **no alignment anywhere**: it asks ESM-2 what it expects at each position, having first masked that position so the model cannot read the answer off its own input, and then wraps the answer onto the fold. Conserved buried core positions become tall single-letter towers, tolerant surface positions splay into short scruffy stacks, and helices read as spiral staircases of text. It is useful for: seeing at a glance which parts of a fold a model considers load-bearing, spotting exposed hydrophobic patches that drive aggregation or block crystallisation, finding positions where the wild-type residue is one the model would not have chosen, and producing something that people will actually share.\n\n## \ud83e\uddea One view, five modes\n\nThere are no tabs. The structure, the camera and the sequence ruler are shared, and the mode buttons on the canvas change only what the glyphs mean. Everything is computed once, shipped in one payload, and switches instantly with no further network round trip.\n\n| Mode | Name | Backronym | Glyph geometry is driven by |\n|---|---|---|---|\n| **Information** | GIBBERISH | Glyph Interface for Bits, Entropy and Residue Information in Structural Homology | ESM-2 per-position probabilities, heights in bits of information content |\n| **Solvent** | BUMFLUFF | Buried/Unburied Mapping of Fonts, Letters, Uncovered Faces and Folds | Relative solvent accessible surface area |\n| **Hotspots** | BALDERDASH | Bayesian Amino-acid Letter Display of Estimated Residue Deviations And Substitution Hotspots | Variant effect scores, ghost glyphs, ClinVar/gnomAD overlay |\n| **Fold** | FOLDEROL | Folding Of Letters Displayed En Route, Ordered Linearly | An animated morph from a flat 2D logo strip into the 3D coordinates |\n| **Logo** | HOGWASH | Height-Ordered Glyphs Weighted Across Sequence Homologues | A real [WebLogo](https://github.com/gecrooks/weblogo) from a real alignment: the one mode that uses one |\n\n## \ud83d\udc16 HOGWASH: the one tab with an alignment\n\nEvery other tab asks a language model what it expects and uses **no alignment anywhere**. HOGWASH is the opposite and the older idea: count what evolution actually did in a column of aligned homologues. Both are here so they can be compared.\n\nIt is a real WebLogo, not a lookalike. [WebLogo 3](https://github.com/gecrooks/weblogo) (MIT, used unmodified) does all of it: the column counts, the composition priors and pseudocounts, the small-sample correction, the unit conversions, the colour schemes, and the rendered EPS/PDF/PNG/SVG. ALPHABETTI supplies the alignment plumbing and a second rendering of the same `LogoData`.\n\n- **Flat cartoon rendering** on a button (or `M`) in every mode: toon shading, no specular highlights, bloom off. Better for a screenshot, and for reading a dense logo where every letter is catching the light at once.\n- **The logo is 3D and interactive**, not a picture. Four arrangements: **helix** (a spiral tower, the default and the right one for a long family), **ring** (a crown you orbit or drop into), **strip** and **rows**. The ring and the helix orbit. The flat layouts do not: they sit square to the camera and are **navigated in two dimensions** \u2014 drag slides the sheet in x and y, shift-drag (or the zoom-box button) rubber-bands a region to zoom into, and a click centres and zooms on a column without rotating anything. A sheet of text is for reading, not orbiting.\n- **A Publication button** puts the real WebLogo figure over the canvas, drawn by WebLogo itself with no interpretation from this app. Not the default view, because a 231-column alignment as a static image is six stacked rows of tiny letters, but one button (or `P`) away when you want the figure.\n- **Input**: paste or upload FASTA, CLUSTAL, Stockholm, PHYLIP, MSF and six other formats, or fetch the current protein's Pfam seed alignment from InterPro.\n- **Every WebLogo option**: units (bits, nats, probability, kT, kJ/mol, kcal/mol, digits), background composition, small-sample correction, alphabet, colour scheme, error bars, stacks per line.\n- **Authentic output**: PNG, PDF, SVG, EPS, JPEG, CSV and WebLogo's own `logodata` text. The flat WebLogo is the **publication figure**, generated by WebLogo itself; it is an export rather than the way you look at the thing.\n- **Five bundled example alignments**, WebLogo's own, so the tab works when InterPro does not.\n\n> **Please cite WebLogo.** Crooks GE, Hon G, Chandonia JM, Brenner SE (2004). WebLogo: a sequence logo generator. *Genome Research* **14**(6):1188-1190. [doi:10.1101/gr.849004](https://doi.org/10.1101/gr.849004) \u00b7 [PMC419797](https://pmc.ncbi.nlm.nih.gov/articles/PMC419797/)\n\nLicence text and all other dependencies: [`THIRD-PARTY.md`](THIRD-PARTY.md).\n\n## \ud83d\udcd0 How a stack is built", "# Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor\nAuthors: Kong, L.; Ju, B.; Chen, Yajing; He, L.; Ren, L.; Liu, J.; Hong, K.; Su, B.; Wang, Z.; Ozorowski, G.; Ji, X.; Hua, Y.; Chen, Yanli; Deller, M.C.; Hao, Y.; Feng, Y.; Garces, F.; Wilson, R.; Dai, K.; O'Dell, S.; McKee, K.; Mascola, J.R.; Ward, A.B.; Wyatt, R.T.; Li, Y.; Wilson, I.A.; Zhu, J.; Shao, Y.\nSource: Immunity, 2016\n\n**Authors:** Leopold Kong, Bin Ju, Yajing Chen, Linling He, Li Ren, Jiandong Liu, Kunxue Hong, Bin Su, Zheng Wang, Gabriel Ozorowski, Xiaolin Ji, Yuanzi Hua, Yanli Chen, Marc C. Deller, Yanling Hao, Yi Feng, Fernando Garces, Richard Wilson, Kaifan Dai, Sijy O'Dell, Krisha McKee, John R. Mascola, Andrew B. Ward, Richard T. Wyatt, Yuxing Li, Ian A. Wilson, Jiang Zhu, and Yiming Shao. Leopold Kong, Bin Ju, Yajing Chen, Linling He, Li Ren and Jiandong Liu are co-first authors; Yuxing Li, Ian A. Wilson, Jiang Zhu and Yiming Shao are co-senior authors.\n\n**Affiliations (condensed):** Department of Integrative Structural and Computational Biology, The Scripps Research Institute, La Jolla, CA 92037, USA; State Key Laboratory for Infectious Disease Prevention and Control, National Center for AIDS/STD Control and Prevention, Chinese Center for Disease Control and Prevention, and Collaborative Innovation Center for Diagnosis and Treatment of Infectious Diseases, Changping District, Beijing 102206, China; Scripps Center for HIV/AIDS Vaccine Immunology & Immunogen Discovery; Department of Immunology and Microbial Science; The Joint Center for Structural Genomics; Skaggs Institute for Chemical Biology; International AIDS Vaccine Initiative Neutralizing Antibody Center and the Collaboration for AIDS Vaccine Discovery \u2014 The Scripps Research Institute, La Jolla, CA 92037, USA; Jiangsu Provincial Center for Disease Control and Prevention, Nanjing, Jiangsu Province 210009, China; Vaccine Research Center, NIH, Bethesda, MD 20892, USA; Institute for Bioscience and Biotechnology Research, University of Maryland, Rockville, MD 20850, USA; Health Science Center, Peking University, Haidian District, Beijing 100191, China; School of Medicine, Nankai University, Nankai District, Tianjin 300071, China.\n\n**Correspondence:** jiang@scripps.edu (J.Z.), yshao@bjmu.edu.cn (Y.S.)\n\n**Published:** Immunity 44, 939\u2013950, April 19, 2016. Received: October 26, 2015; revised and accepted: December 14, 2015; published online: April 5, 2016.\n\n## In Brief\n\nAlthough VRC01-class antibodies have been extensively studied, early events in the emergence of this antibody class in HIV-1 infection remain unknown. Zhu, Shao, and colleagues find through structural, functional, and repertoire analyses that key gp120 glycans appear to stall rapid development of a VRC01-like antibody lineage in a Chinese patient.\n\n## Highlights\n\n- Early-stage development of a VRC01-class antibody in a Chinese HIV patient\n- VRC01-class heavy chains can evolve within 2 years\n- VRC01-class light-chain CDR3 signature is encoded within the B cell repertoire\n- Light-chain N terminus and CDR1 conformation need to accommodate key gp120 glycans\n\n## Summary\n\nVRC01-class antibodies neutralize diverse HIV-1 strains by targeting the conserved CD4-binding site. Despite extensive investigations, crucial events in the early stage of VRC01 development remain elusive. We demonstrated how VRC01-class antibodies emerged in a Chinese donor by antigen-specific single B cell sorting, structural and functional studies, and longitudinal antibody and virus repertoire analyses. A monoclonal antibody DRVIA7 with modest neutralizing breadth was isolated that displayed a subset of VRC01 signatures. X-ray and EM structures revealed a VRC01-like angle of approach, but less favorable interactions between the DRVIA7 light-chain CDR1 and the N terminus with N276 and V5 glycans of gp120. Although the DRVIA7 lineage was unable to acquire broad neutralization, longitudinal analysis revealed a repertoire-encoded VRC01 light-chain CDR3 signature and VRC01-like neutralizing heavy-chain precursors that rapidly matured within 2 years. Thus, light chain accommodation of the glycan shield should be taken into account in vaccine design targeting this conserved site of vulnerability.\n\n## Introduction\n\nThe envelope glycoprotein (Env) spike of human immunodeficiency virus type-1 (HIV-1) is a trimer of heterodimers, each containing a receptor-binding protein, gp120, and a transmembrane fusion protein, gp41 (Wyatt and Sodroski, 1998). Dense heterogeneous N-linked glycans and hypervariable loops on the Env surface form barriers to the elicitation of broadly neutralizing antibodies (bNAbs) (Julien et al., 2013; Lyumkis et al., 2013; Pancera et al., 2014). Nonetheless, a small fraction of infected individuals are capable of developing bNAbs after several years of infection (Gray et al., 2011; Li et al., 2007; Mikell et al., 2011; Walker et al., 2010). The identification and characterization of bNAbs and their epitopes are considered paramount to guiding rational vaccine design (Walker and Burton, 2010). Due to the recent advances in antibody isolation, panels of bNAbs with extraordinary breadth and potency have been reported (Burton and Mascola, 2015; Klein et al., 2013; Kwong and Mascola, 2012; Kwong et al., 2013).\n\nOf the conserved HIV-1 epitopes identified thus far, the CD4-binding site (CD4bs) on gp120 appears to be an ideal target for vaccine design (Zhou et al., 2007). However, the CD4bs is recessed between two subunits of the Env trimer, with a narrow range of approach angles by antibodies (Chen et al., 2009; Lyumkis et al., 2013). A class of CD4bs-directed bNAbs, with VRC01 being the first of the class (Wu et al., 2010), has been identified from multiple donors, suggesting that the immune system is capable of recognizing this recessed site of vulnerability.", "# GitHub README: ChemSage\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ChemSage, 2026\n\n- [ ] **Weighted class-balanced sampling across all 78 classes.** Distribution shift was the proven cause of the R4 regression (exec 79\u219271%, SMARTS 100\u219255%) and SMARTS has swung 100\u219255\u2192100 across three rounds purely on dataset composition. Lock the fragile metrics by controlling sampling density rather than relying on raw example counts.\n- [ ] **Triple the `code_then_quote_v2` drill.** Code\u2192Quote fidelity moved from 19% (R4) to 61% (R5) but is still the weakest metric by 28 points. The failure mode is copy-fidelity, not reasoning \u2014 add examples where the quoted value must byte-match the printed `stdout` token, including trailing zeros and units, so the model learns to transcribe rather than re-derive.\n- [ ] **Add a dedicated `pymol_syntax_drill` generator.** PyMOL syntax has sat at 89\u201390% for two rounds and is the lowest structural biology metric. Failures cluster in selection algebra and command chaining \u2014 target those specifically rather than adding more generic structural biology classes.\n- [ ] **Two-phase curriculum: full run then a short fidelity-focused continuation.** Train the full 78-class dataset to plateau (~1,600 iters, matching R5 pattern), then run a short ~400-iter second phase at lower LR on a weak-metrics-only subset (code\u2192quote, PyMOL, fidelity). Phase 2 is cheap (~7 h) and concentrates remaining capacity on the actual gaps rather than fighting the plateau.\n- [ ] **Configure for ~2,000 iters, not 3,000.** R3 stopped at 42% of configured iters, R5 at 67% \u2014 both plateaued well before the ceiling. Stop burning 15\u201320 hours on a flat val curve; redirect that budget into the phase-2 continuation or more dataset quality work.\n- [ ] **Try rank 96 (not 128).** The rank 32\u219264 jump in R5 tracked with the 95% overall score but the val floor rose to 0.055 with 78 classes, suggesting the harder task is capacity-limited. Rank 96 (~400M trainable params) should sit comfortably under the R5 peak of 30.3 GB; skip rank 128 to avoid swap risk on 64 GB unified memory.\n- [ ] **Extend `fidelity_multistep` with longer chains and distractor numbers.** Fidelity reached 89%, but the residual 11% is multi-step cases where an intermediate value is transcribed incorrectly. Adding 4\u20135 step chains with adversarial distractor numbers forces the model to track which value to quote rather than defaulting to the largest or last number.\n- [ ] **Add per-class eval breakdown and run it on a mid-training checkpoint (~iter 1,000).** R4's exec/SMARTS regression only surfaced after 43 hours of training. A per-class pass/fail on the 100-example shared test set at iter 1,000 would expose class-level dilution early \u2014 while there is still time to rebalance the dataset and restart.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/ChemSage\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/ChemSage</a></td>\n</tr>\n</table>", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 3: Structure Determination of Oncostatin M (Chapter 5)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics; supervisor Prof. E. Yvonne Jones), 1999\n\n| Cross-Rotation Solution | \u03b1 (\u00b0) | \u03b2 (\u00b0) | \u03b3 (\u00b0) | Correlation Coefficient (%) |\n|---|---|---|---|---|\n| 1 | 6.8 | 77.2 | 153.8 | 21.5 |\n| 2 | 172.4 | 76.5 | 217.2 | 12.5 |\n| 3 | 170.6 | 90.0 | 326.5 | 12.5 |\n| 4 | 117.2 | 75.0 | 207.8 | 11.8 |\n| 5 | 147.3 | 82.3 | 184.9 | 11.3 |\n\n*Table 5-5.* Table showing top 5 results after cross-rotation search of molecular replacement of I222 data. Cross-rotation carried out using program AMoRe (Navaza, 1994).\n\n| Translation Solution | \u03b1 (\u00b0) | \u03b2 (\u00b0) | \u03b3 (\u00b0) | x | y | z | Correlation Coefficient (%) | R-factor (%) |\n|---|---|---|---|---|---|---|---|---|\n| 1 | 6.8 | 77.2 | 153.8 | 0.160 | 0.162 | 0.311 | 55.0 | 43.5 |\n| 2 | 6.8 | 77.2 | 153.8 | 0.160 | 0.457 | 0.310 | 41.9 | 49.6 |\n| 3 | 6.8 | 77.2 | 153.8 | 0.160 | 0.067 | 0.311 | 40.0 | 50.3 |\n| 4 | 6.8 | 77.2 | 153.8 | 0.159 | 0.271 | 0.313 | 37.7 | 50.7 |\n| 5 | 6.8 | 77.2 | 153.8 | 0.360 | 0.043 | 0.000 | 33.4 | 53.3 |\n\n*Table 5-6.* Table showing top 5 results after cross-rotation and translation search of molecular replacement of I222 data. Cross-rotation and translation carried out using program AMoRe (Navaza, 1994). Eulerian angles in degrees; translations in fractional coordinates.\n\n| Refined Solution | \u03b1 (\u00b0) | \u03b2 (\u00b0) | \u03b3 (\u00b0) | x | y | z | Correlation Coefficient (%) | R-factor (%) |\n|---|---|---|---|---|---|---|---|---|\n| 1 | 5.3 | 76.8 | 154.2 | 0.159 | 0.156 | 0.309 | 64.5 | 40.5 |\n| 2 | 5.6 | 77.8 | 154.6 | 0.160 | 0.454 | 0.310 | 50.1 | 47.8 |\n| 3 | 5.9 | 77.8 | 154.3 | 0.158 | 0.065 | 0.310 | 47.7 | 49.3 |\n| 4 | 6.2 | 77.3 | 154.1 | 0.160 | 0.267 | 0.310 | 44.8 | 49.4 |\n| 5 | 4.6 | 77.1 | 154.5 | 0.409 | 0.130 | 0.043 | 43.0 | 51.5 |\n\n*Table 5-7.* Table showing top 5 results after refinement of cross-rotation and translation matrix of molecular replacement solution. Cross-rotation, translation and rigid body refinement carried out using program AMoRe (Navaza, 1994).\n\n#### 5.3.7 Map Calculation and Model Refinement\n\nInitial electron density maps based on combined MAD and SIRAS phases showed poor side chain density although the main chain density was generally contiguous. Solvent flattening improved the quality of density and further improvements were seen during the course of refinement as shown below in Figure 5-7.\n\nCrystallographic-R and free-R (Brunger, 1992) values were monitored throughout the course of refinement, as were deviations from stereochemical parameters as defined by Engh and Huber (Engh and Huber, 1991). Initial rigid body refinement resulted in a model with crystallographic-R value of 48 % and a free-R value of 52 %. Low weighting of the X-ray terms resulted in tight stereochemistry with root mean squared deviations (RMSD) from ideality in bond lengths typically less than 0.008 \u00c5. After 15 rounds of manual rebuilding and refinement and the inclusion of 116 waters the current model has a crystallographic-R value of 20.5 % with a free-R value of 26.1 % for all data between 20.0\u20132.2 \u00c5. Stereochemistry is typified by RMSD in bond lengths of 0.0147 \u00c5 and RMSD in bond angles of 1.7997\u00b0. Statistics relevant to model refinement are shown below in Table 5-8.\n\nThe current model for the I222 structure contains 35 waters and has been refined to a crystallographic-R value of 20.8 % and a free-R value of 28.9 % for all data between 20.0\u20132.5 \u00c5. Stereochemistry is typified by RMSD in bond lengths of 0.0129 \u00c5 and RMSD in bond angles of 1.6112\u00b0. Statistics relevant to the final models are shown below in Table 5-8.\n\n| | P212121 | I222 |\n|---|---|---|\n| Resolution Range | 20\u20132.2 \u00c5 (2.28\u20132.20 \u00c5) | 20\u20132.5 \u00c5 (2.59\u20132.50 \u00c5) |\n| Rcryst | 20.5% (27.1%) | 20.8% (27.8%) |\n| Rfree | 26.1% (33.6%) | 28.9% (31.5%) |\n| Completeness \u2013 Working Set | 86.1% (81.0%) | 75.7% (75.4%) |\n| Completeness \u2013 Test Set | 9.9% (8.6%) | 9.0% (6.6%) |\n| Number of Reflections (F>0) \u2013 Working Set | 9410 (870) | 4927 (481) |\n| Number of Reflections (F>0) \u2013 Test Set | 1082 (92) | 583 (42) |\n| Number of Non-Hydrogen Atoms \u2013 Protein | 1319 | 1219 |\n| Number of Non-Hydrogen Atoms \u2013 Water | 116 | 35 |\n| RMSD Bond Lengths | 0.0147 \u00c5 | 0.0129 \u00c5 |\n| RMSD Bond Angles | 1.7997\u00b0 | 1.6112\u00b0 |\n| RMSD Dihedrals | 19.8882\u00b0 | 20.6155\u00b0 |\n| RMSD Improper | 1.0381\u00b0 | 1.1047\u00b0 |\n| Average B-factor \u2013 Main Chain | 27.1 \u00c52 | 38.5 \u00c52 |\n| Average B-factor \u2013 Side Chain | 32.4 \u00c52 | 42.9 \u00c52 |\n| Average B-factor \u2013 Water | 38.4 \u00c52 | 41.5 \u00c52 |\n| Average B-factor \u2013 All Atoms (Range) | 30.5 \u00c52 (6.3\u201384.9 \u00c52) | 40.8 \u00c52 (10.2\u201384.4 \u00c52) |\n\n*Table 5-8.* Structural refinement and model statistics for the two crystal structures of \u0394188_OSM. Rcryst = \u03a3||Fobs|\u2212|Fcalc||/\u03a3|Fobs|. Rfree is as for Rcryst but calculated using a 10% test set of reflections excluded from the refinement. Values in parentheses correspond to the highest resolution shell (2.28\u20132.20 \u00c5 for P212121 and 2.59\u20132.50 \u00c5 for I222).\n\n*Figure 5-7.* Figure showing the course of refinement of the \u0394188_OSM model. All maps are contoured at 1 \u03c3. Region shown corresponds to the helix C running from residues 106 to 119. Maps were produced from MAD data using SHARP (de La Fortelle and Bricogne, 1997) followed by solvent flattening with SOLOMON (Abrahams and Leslie, 1996) and refinement with CNS-SOLVE (Brunger et al., 1998). Final refined coordinates are shown in each case. Figure was produced using BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994).\n\n#### 5.3.8 Analysis of Model\n\nIn the P212121 structure residues 4\u2013134 and 156\u2013187 are well ordered, with residues 1\u20133 at the N-terminus and residues 135\u2013155 of the C-D loop being disordered with no interpretable electron density. Electron density is good throughout the course of the main chain and a representative portion of electron density is shown below in Figure 5-8. Residue numbering corresponds to that of the human native mature protein (GenBank accession number M27286; (Malik et al., 1989)).", "# GitHub README: TopPDBLX\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/TopPDBLX, 2026\n\n> **Every crystallisation condition in the Protein Data Bank, parsed, normalised and linked to the sequence that produced it.**\n\n [](https://toppdblx.mdeller.com)          [](https://doi.org/10.5281/zenodo.21807133) [](https://github.com/ml-explore/mlx-lm) [](#-three-language-models-each-doing-a-different-job) [](#-three-language-models-each-doing-a-different-job) [](https://huggingface.co/Dellboy) \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/TopPDBLX\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/TopPDBLX</a></td>\n</tr>\n</table>\n\n---\n\nThe Protein Data Bank holds about 200,000 crystallisation recipes, and every one was typed in free-hand by a different scientist, in no agreed format. TopPDBLX recovers structured, curated chemistry from that text: it turns the free-text `_exptl_crystal_grow.pdbx_details` field into typed components (reagent, concentration, unit, role), cross-references them against published commercial screen formulations, and attaches MMseqs2 cluster identifiers so that redundancy can be controlled properly.\n\n**Why it matters:** the strongest lever on crystallisation success is construct design and the second strongest is choosing a screen that suits the protein, yet both are still done by intuition, because all the precedent sits in a quarter of a million unstructured strings that nobody can query. Turning that text into data is the precondition for every downstream question worth asking. It is useful for: crystallographers planning a screen, structural genomics groups mining historical outcomes, method developers who need a benchmark, and anyone building models over crystallisation space.\n\n- Project brief: [`toppdblx_spec_v1.md`](toppdblx_spec_v1.md)\n- Phase 0 implementation plan: [`PHASE0_PLAN.md`](PHASE0_PLAN.md)\n- Dataset datasheet: [`DATASHEET.md`](DATASHEET.md)\n- Roadmap to completion: [`ROADMAP.md`](ROADMAP.md)\n\n---\n\n## \u2728 What it does\n\n| Capability | Detail |\n|---|---|\n| **Parses the whole archive** | 199,185 crystallisation records from 198,691 PDB entries, keyed on `(pdb_id, crystal_id)` |\n| **Typed components** | 645,656 reagents with role, concentration, unit, PEG molecular weight, Hofmeister rank and buffer pKa, each tagged with the parser that produced it |\n| **Separates text from chemistry** | Method notes, screen references and unnamed ligands are labelled `not_a_component`, so they never inflate a failure rate |\n| **Accounts for every record** | Anything not parsed carries one of seven discard codes, with its raw text retained |\n| **Links to sequence** | 184,229 usable records carry the construct sequence, UniProt accessions and cluster ids at 30%, 50% and 90% identity |\n| **Cross-references screens** | 434 wells across 9 commercial screens, extracted verbatim from vendor support materials |\n| **Refuses to guess** | A condition with an unrecognised reagent or no stated amount is Unclassified, with the reason recorded |\n| **Records its own uncertainty** | Inferred units, inferred cryoprotectant roles and pH attribution are flagged, never presented as fact |\n| **Reproducible by stage** | Every stage is one command and writes a manifest of input hashes, tool versions and git state |\n| **Reads what the rules cannot** | A 32B **teacher** labels the hard records; a 360M **student** learns from it and reads every residual record the rule parser gives up on. Recall against hand-labelled truth: 69.5% (rules alone) \u2192 90.3% (rules + round 09, pooled) |\n| **Suggests where to cut** | A third model, a **boundary proposer**, reads a protein sequence and predicts where a crystallographer would truncate it. Trained on 523,018 deposited constructs, and separate from the two above |\n\n### Three language models, each doing a different job\n\nThe project runs **three**, and they are not interchangeable. Two read text; one reads protein.\n\n| Model | What it reads | What it produces | Why it exists |\n|---|---|---|---|\n| **1. Parser teacher**<br>Qwen2.5-32B, local | The hardest crystallisation strings, a few thousand of them | Training labels for the student | The rule parser cannot label its own failures, so something has to say what the difficult cases mean. Too slow to run over the corpus: at ~40 s a record it would take weeks |\n| **2. Parser student**<br>[SmolLM2-360M](https://huggingface.co/Dellboy/toppdblx-residual-parser), LoRA | Every condition string the rule parser gave up on | Typed components: reagent, amount, unit, role | Does the corpus-scale work the teacher cannot. 90x smaller, reads the whole residual in under two hours, and contributes **35,580 components** to the release |\n| **3. Boundary proposer**<br>[ESM-2 t12-35M](https://huggingface.co/Dellboy/toppdblx-construct-boundary) | A protein sequence, not text | Per-residue probability of being inside a crystallised construct | Answers a different question entirely: not *what was in the drop* but *what was in the tube*. Trained on 523,018 deposited constructs |\n\n**Models 1 and 2 are a distillation pair**: the big one teaches, the small one works. Model 3 is\nunrelated to both, shares no code path with them, and would still be useful if the parser did not\nexist. Keeping them straight matters when reading the results below, because \"the model\" means\nsomething different in each section.\n\n## \ud83d\udcd6 The words on this page\n\nCrystallography and machine learning each bring their own vocabulary, and this page uses both.\nEverything here is defined where it first matters, but this is the short version.", "# BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n```\npython3 BoltzMaker.py new      [boltz_input.md]  # write a new campaign by answering plain questions\npython3 BoltzMaker.py format   boltz_input.md    # auto-align comments/blank-lines (cosmetic only)\npython3 BoltzMaker.py generate boltz_input.md    # write the target YAMLs + manifest\npython3 BoltzMaker.py preflight boltz_input.md   # environment + input sanity checks\npython3 BoltzMaker.py run      boltz_input.md    # boltz predict, live progress, resumable\npython3 BoltzMaker.py analyze  boltz_input.md    # CSV / XLSX / HTML dashboard\npython3 BoltzMaker.py all      boltz_input.md    # generate -> preflight -> run -> analyze\npython3 BoltzMaker.py boltz_input.md             # same as `all` (subcommand is optional)\npython3 BoltzMaker.py compare-sse boltz_input.md # apo-vs-holo secondary-structure motif shifts (see below)\n```\n\n`new` interviews you (proteins, partners, ligands, and the three constraint sentence types) and writes the file for you \u2014 it won\u2019t overwrite an existing file without asking first. It covers the common case only; rarer fields (modifications, cyclic, MSA override, templates, homo-oligomer copies) are left for hand-editing the file it writes. If `setup-plip` has been run, it also asks whether you have a reference structure with a ligand already bound (a co-crystal or homology model) for each protein \u2014 if so, it runs cif2plip on it, lets you pick the relevant ligand if more than one is detected, and suggests the contacted residues as `Pocket contact:` constraints, remapped onto your target\u2019s own numbering via sequence alignment (BLOSUM62 + affine gaps) so the reference structure\u2019s residue numbers don\u2019t have to match your target\u2019s.\n\n`format` re-aligns trailing comments to a clean column and normalizes blank-line spacing around section/record boundaries, purely cosmetic (it validates the file parses first, and never changes meaning). Pass `--check` to report whether reformatting is needed without writing anything (exit 1 if so, e.g. for a pre-commit check).\n\nAny field BoltzMaker doesn\u2019t recognize (a typo like `Predict afinity:` instead of `Predict affinity:`) prints a `WARNING` naming the block, its name, and the line number, and is otherwise silently dropped, so a misspelled field never just vanishes without a trace.\n\n**Common options:**\n\n| Option | Default | Description |\n| --- | --- | --- |\n| `--output-dir` | `settings.output_dir` | Override where generated YAMLs are written |\n| `--out-dir` | `./boltz_output` | Boltz\u2019s own `--out_dir`, next to the md file |\n| `--accelerator` | `auto` | `auto` / `gpu` / `cpu` |\n| `--limit N` | none | Cap how many pending targets `run` submits (smoke test before a full batch) |\n| `--max-retries` | `2` | Auto-retry a target that doesn\u2019t complete (e.g. an OOM), isolating to one target at a time \u2014 see \u201cMemory on Mac\u201d below (`0` disables) |\n| `--strict` | off | Promote preflight WARN to FAIL |\n| `--skip-interactions` | off | Skip cif2plip interaction analysis during `analyze`, even if `setup-plip` has been run |\n| `--skip-sse` | off | Skip compare-sse apo-vs-holo analysis during `analyze`, even if a family has `Apo structure:` set |\n\n**Memory-control options** (see below for why these matter on Mac/unified-memory hardware):\n\n| Option | Default | Description |\n| --- | --- | --- |\n| `--workers` | `2` | Matches Boltz\u2019s own default |\n| `--mps-watermark` | `1.0` | `PYTORCH_MPS_HIGH_WATERMARK_RATIO` cap |\n| `--max-parallel-samples` | `1` | Boltz `--max_parallel_samples` |\n| `--recycling-steps` | Boltz default | Passthrough |\n| `--sampling-steps` | Boltz default | Passthrough |\n| `--diffusion-samples-affinity` | Boltz default | Passthrough |\n| `--sampling-steps-affinity` | Boltz default | Passthrough |\n| `--max-msa-seqs` | Boltz default | Passthrough |\n| `--memory-warn-tokens` | `1000` | Preflight size-heuristic WARN threshold |\n\n`run` is idempotent: targets with a complete prediction (cif + confidence json, and an affinity json if `predict_affinity` is on) are skipped on re-run, so an interrupted batch can just be re-run as-is.\n\n**`compare-sse` options** (see the section below for what it does \u2014 it now also runs automatically as part of `analyze`/`all` for every family with an `Apo structure:` set; the standalone command below is for re-running just this analysis on its own, e.g. after adding an apo structure without re-running `boltz predict`):\n\n| Option | Default | Description |\n| --- | --- | --- |\n| `--family` | every family with `Apo structure:` set | Restrict to one `Protein` family id |\n| `--target` | every target for the selected family | Restrict to one target stem |\n| `--out-dir` | alongside `boltz_input.md` | Where to write the CSV/HTML/PyMOL scripts |\n| `--phi-psi-threshold` | `30` (degrees) | Per-residue phi/psi delta above this is flagged |\n| `--dfg-distance-threshold` | `8.0` (Angstrom) | DFG-Asp to catalytic-Lys Ca-Ca distance below this is classified DFG-in |\n| `--alphac-distance-threshold` | `10.0` (Angstrom) | alphaC-Glu to catalytic-Lys Ca-Ca distance below this is classified alphaC-in |\n| `--no-pymol` | off | Skip writing `.pml` session scripts |\n| `--refresh-cache` | off | Bypass the GPCRdb/KLIFS/PDBe disk cache for this run |\n\n## \ud83d\udee0\ufe0f Memory on Mac (unified-memory) hardware\n\nA real 4-chain GPCR+G-protein complex (~1250 combined residues/atoms) used **~65GB RAM on a 64GB M1 Max** during testing and swap-thrashed for 20+ minutes with zero progress before being killed, worth knowing about before running anything large. Mitigations built in:", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n**Run time scales with complex size, not target count.** `ADRB2_ISO1`'s two-chain receptor+Gs\ncomplex took disproportionately longer than the single-chain examples, since attention-style\noperations scale worse than linearly with sequence length -- confirmed by `AR2NG_PRO1` (same\nreceptor, no partner) finishing in a fraction of the time. On Apple Silicon, `torch.linalg.svd`\n(used in the diffusion step) also has no MPS implementation and silently falls back to CPU:\nworth budgeting for on large multi-chain campaigns.\n\n`5ht2_gq_panel`'s six large 4-chain receptor+Gq targets (~1250-1280 tokens) originally crashed\non Apple Silicon: boltz's triangular attention computes the full row-wise QK^T score matrix for\nthe whole complex in one unchunked matmul, which exceeds MPS's single-tensor size ceiling past\nroughly 1250 residues and crashes inside PyTorch's internal tiled-bmm fallback. Each row's\nattention is independent, so chunking along that axis is exact, not an approximation -- `setup`\npatches this into the installed `boltz` package (idempotent, and checked against boltz's exact\nsource so a future upgrade can't be silently mis-patched). `run` also wraps `boltz predict`\nwith `caffeinate` (macOS only, silently skipped if unavailable) as sleep-prevention hygiene for\nlong GPU jobs. All 15 targets now complete -- see [CHANGELOG.md](CHANGELOG.md) for the fix.\n\n## \ud83e\udded boltz_input.md format\n\nPlain labelled text -- no markdown, no YAML, no brackets, no quoting. One rule: blocks are\n`Label: value` lines with a blank line between them; comments start with `#`. Field names are\nplain English (`Output folder`, `Predict affinity`, `Pocket contact`) rather than\nBoltz-internal snake_case. Don't want to hand-write it? Run `python3 BoltzMaker.py new` and\nanswer plain questions instead.\n\nThe format has two layers: a **family x partners x ligand cross-product** (the ergonomic\nlayer -- write each protein/ligand once, get every combination as a separate target), and\nstandalone **constraint sentences** for the two/three-ended relationships (covalent bonds,\ndistance constraints) that don't fit inside one block. Each names the protein it belongs to\nand can be written anywhere in the file.\n\n```\nSettings:\nOutput folder: ./boltz_yamls    # where generated per-target YAMLs are written\nPredict affinity: no            # off by default -- it's a heavier prediction pass\nTargets per invocation: 4       # restart Boltz every N targets so the GPU allocator\n                                # starts clean -- see \"what we fixed\", item 2b.\n                                # 0 runs the whole batch in one long-lived process", "# Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly\nAuthors: Deller, M.C.; Johnson, H.A.; Miller, M.D.; Spraggon, G.; Elsliger, M.-A.; Wilson, I.A.; Lesley, S.A.\nSource: PLoS ONE, 2015\n\nThe similarities between the ADP-bound TrkA/TrkH structure and the AMP-bound TM1088 assembly suggests that our structure most closely resembles a closed state of the channel. To confirm this hypothesis, further biochemical studies will be required. Additionally, to confirm the exact nature of each interface, and the mechanism of regulation, additional structures of the TM1088 assembly captured in various conformational states will also be needed. Perhaps more importantly, high-order assemblies of TM1088 in association with its membrane-spanning TrkH component, TM1089, will also be required to address the mechanism of coupling between the cytosolic gating ring component and that of the membrane-spanning channel.\n\nFinally, two-subunit Trk proteins are predicted to be fairly commonplace in the bacterial realm [6,9]. The structure of the TM1088 assembly provides the first structural insights into what may be an evolutionary ancestor of more modern single-subunit gating ring assemblies.\n\n## Acknowledgments\n\nThe content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. X-ray experiments were conducted at beamlines 5.0.1 and 5.0.3 of the Advanced Light Source (ALS). The ALS is supported by the Director, Office of Science, Office of Basic Energy Sciences, Material Sciences Division of the U.S. Department of Energy under contract No. DE-AC03-76SF00098 at Lawrence Berkeley National Laboratory. We thank other members of the Joint Center for Structural Genomics for technical assistance in particular Tom Clayton, Henry Tien, Tina Trout, Julie Feuerhelm and Anna Grzechnik. Genomic DNA from T. maritima MSB8 (DSM3109, ATCC Number 43589D-5) was obtained from the American Type Culture Collection (ATCC).\n\n## Author Contributions\n\nConceived and designed the experiments: MCD HAJ SAL. Performed the experiments: MCD HAJ MDM. Analyzed the data: MCD HAJ MDM GS. Contributed reagents/materials/analysis tools: MCD HAJ GS. Wrote the paper: MCD HAJ MDM GS MAE IAW SAL.\n\n## References\n\n*(reference list omitted)*", "# GitHub README: TopPDBLX\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/TopPDBLX, 2026\n\n**The scaling is what makes this convincing rather than merely disappointing.** A 3,000-record\npass scored +0.001 at hit@5; the full set, 32 times larger, scored -0.013. A model that moves\n*away* from the baseline as data is added is not short of data.\n\nHomology retrieval fails the same way, at every level, on both split thresholds, under three\ndefinitions of ground truth. Two independent methods failing identically points at the data rather\nthan the method, and the reason is already visible in the data (see [Known limitations](#-known-limitations)):\ncondition frequency reflects screen popularity, not protein identity. Homologues were screened\nwith the same popular screens as everything else, so there is little protein-specific signal to\nretrieve.\n\n**The archive supports choosing a screen by precedent, and choosing a construct by sequence, but\nnot choosing a condition by protein.** That is a useful thing to know before building a\nrecommender, and it is why this project does not ship one.\n\n## \ud83d\udcc4 The paper\n\n[`PAPER.md`](PAPER.md) draws the whole project together in about 1,800 words: fidelity, the\nparser, the curation, and then what can and cannot be learned. Every figure in it is re-derived\nfrom the released artefacts rather than copied from earlier prose, which caught five stale numbers\nincluding two of the roadmap's own.\n\n## \ud83d\udd01 Redundancy\n\n**In plain terms:** the same experiment appears in the archive over and over. Popular proteins are solved hundreds of times, and a successful recipe gets copied. One condition string in this corpus appears **1,784 times**. That sounds harmless, and it quietly corrupts three different things if you let it.\n\n| Where it bites | Left alone | Handled |\n|---|---|---|\n| Expert curation | 1,484 near-identical records to judge | 10 grouped decisions covering 1,004 names |\n| Train and test splits | The same protein on both sides, so accuracy measures memory | Split by sequence cluster, leak-free at 30%, 50% and 90% |\n| Training data | 45.7% of rows repeated; one string took 1.3% of all training | Deduplicated: 135,793 rows to 95,818, all distinct bar a deliberate 8x |\n| Validation loss | Scored on duplicated conditions | 17,292 rows to 12,730, fully distinct |\n\n**For the crystallographer:** the corpus is not a sample of crystallisation space, it is a sample of *what people deposited*, and deposition is heavily skewed. A single 30% identity cluster holds up to 2,876 entries. Lysozyme, trypsin and Fab fragments alone would dominate any statistic computed per record.\n\nThree consequences, each of which had to be designed for rather than discovered late:\n\n- **Splits are by sequence cluster, never by entry** (spec 7.3). Splitting at random would put near-identical proteins on both sides and every metric would measure memorisation. The folds are the connected components of the union of all three clustering thresholds, because MMseqs2 clusters do not nest: 68 clusters at 90% identity straddled a 30% split until that was fixed.\n- **Audits rank distinct decisions, not instances.** Judging \"20% PEG 3350 becomes PEG_3350\" for the thirty-thousandth time buys no evidence. Each question stands for every occurrence of the string it names.\n- **Training data is deduplicated before oversampling, not after.** This one was caught late: 36% of training rows were exact duplicates, and because the rare `not_a_component` records were themselves duplicated, multiplying them first produced an effective 11.7x where 8x was specified. Deduplicating the validation set matters more still, since a repeated condition there distorts the loss curve used to judge the run.\n\nThe general lesson is that a redundant corpus punishes anything counted per record. Every figure in this README that could be inflated by repetition is either computed over distinct strings or over sequence clusters, and says which.\n\n## \ud83e\uddea Curation interfaces\n\nExpert time is the scarcest input, so every audit is reduced to **tens of decisions ranked by corpus reach**, never a row-per-instance review.\n\n| Round | Questions | Reach |\n|---|---|---|\n| Reagent lexicon, rounds 1 and 2 | 35, then 13 | Seeded the original lexicon |\n| Lexicon gaps | 40 | 7,018 unidentified components |\n| Lexicon gaps, grouped | 10 | 1,004 names, 34% of the unidentified mass |\n| Classification accuracy, rounds 1 and 2 | 96 each | The per-class accuracy number (spec 6.6), split by provenance |\n| **Gold sets** | **96, then another 96** | **Precision and recall against labelled truth, the yardstick for every model round** |\n\n`app/condition_courtroom_v5.html` renders any payload with the same shape, so a new audit is a new generator stage rather than a new page. `condition_courtroom_v7.html` is the accuracy audit: one condition per screen, one checkbox, and a three-field diagnosis that appears only when a card is ticked, so the ~90% waved through stay a single click. `gold_bench_v1.html` is the labeller: deposition text, removable reagent chips, and an autocomplete over all lexicon names for what the pipeline missed.\n\n**A recommendation is an answer, not a suggestion.** Every dropdown is pre-set and there is an accept-all button, so a wrong recommendation is not caught, it is accepted in bulk. Generators therefore carry explicit chemistry guards: numbers must match exactly (a number in a reagent name is molecular weight or substitution position, never decoration), element and acid words must agree, and a family name such as bare `peg` identifies to nothing rather than inventing a member. Guards decide what is **recommended**, never what is **available**.\n\n## \ud83d\udd27 Installation\n\n```bash\ngit clone https://github.com/bellcheddar/TopPDBLX.git\ncd TopPDBLX\npython3.14 -m venv .venv\n.venv/bin/python -m pip install -r requirements.txt\n.venv/bin/python -m pip install -e '.[dev]'\nbrew install mmseqs2          # required by the clustering stage\n```", "# Discovery of an Orally Bioavailable KRAS G12D Inhibitor\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [9E5D](https://www.rcsb.org/structure/9E5D)\n* **Release Date:** 2025-01-22\n* **Method:** X-RAY DIFFRACTION 1.36 \u00c5\n* **Organisms:** Homo sapiens\n* **Macromolecules:** GTPase KRas (protein)\n* **Unique Ligands:** A1BEI (INCB159020), GDP, MG\n* **Publication:** [Deller, M.C., et al. (2025). Discovery of an Orally Bioavailable KRAS G12D Inhibitor. Journal of Medicinal Chemistry, 68, 1924-1939.](https://pubs.acs.org/doi/10.1021/acs.jmedchem.4c02662)\n* Blog Post: [Discovery of INCB159020: An Orally Bioavailable KRAS G12D Inhibitor](https://marcdeller.com/discovery-of-incb159020-an-orally-bioavailable-kras-g12d-inhibitor/)\n\n**Description:** This structure represents a major breakthrough in targeted oncology, revealing the binding mode of INCB159020, a novel orally bioavailable inhibitor targeting the oncogenic KRAS G12D mutation. The high-resolution structure (1.36 \u00c5) provides critical insights for structure-based drug design in pancreatic and colorectal cancers.", "# GitHub README: AlphaFraud\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/AlphaFraud, 2026\n\n### \ud83d\udd34 Stromal interaction molecule 1 / CC1 (STIM1) \u00b7 [`6YEL_1`](https://alphafraud.mdeller.com/examples#ex-6YEL_1)\n\n| Field | Detail |\n|---|---|\n| **Take-home** | A bistable three-helix Ca\u00b2\u207a switch; AlphaFold committed to one conformer, the structure is the other |\n| **AlphaFold vs experimental** | <a href=\"https://alphafraud.mdeller.com/examples#ex-6YEL_1\"><img src=\"https://alphafraud.mdeller.com/ribbon/6YEL_1.svg\" alt=\"Experimental structure of 6YEL_1 coloured by C\u03b1 deviation from AlphaFold\" width=\"380\"></a><br><sub>Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \"ghost\" overlay</sub> |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-6YEL_1)** |\n| **KPIs** | TM 0.389 \u00b7 lDDT 0.341 \u00b7 pLDDT 78.9 \u00b7 C\u03b1-RMSD 35.8 \u00c5 \u00b7 FRAUD 0.75 \u00b7 novelty 2% |\n| **Stats** | 1 deposition, confidently wrong \u00b7 SS-Q3 69.6% (secondary structure right, packing wrong) |\n| **Structure** | CC1 is a compact antiparallel three-helix bundle that toggles between clamped (resting) and extended (active) states |\n| **Why AlphaFold gets it wrong** | Right fold class, wrong tertiary packing \u2014 a single-answer predictor cannot represent a two-state switch |\n| **Biology & disease** | Gates store-operated Ca\u00b2\u207a entry via Orai1; gain-of-function \u2192 Stormorken / tubular aggregate myopathy; loss \u2192 CRAC-channelopathy |\n| **Key fact** | 35.8 \u00c5 RMSD with correct secondary structure = a conformer error, not a misfold |\n\n### \ud83d\udd34 Activator of Hsp90 ATPase 1 / Aha1 (AHSA1) \u00b7 [`7DME_1`](https://alphafraud.mdeller.com/examples#ex-7DME_1)\n\n| Field | Detail |\n|---|---|\n| **Take-home** | Two well-folded domains bolted together at the wrong angle \u2014 a flexible-linker ensemble collapsed to one guess |\n| **AlphaFold vs experimental** | <a href=\"https://alphafraud.mdeller.com/examples#ex-7DME_1\"><img src=\"https://alphafraud.mdeller.com/ribbon/7DME_1.svg\" alt=\"Experimental structure of 7DME_1 coloured by C\u03b1 deviation from AlphaFold\" width=\"380\"></a><br><sub>Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \"ghost\" overlay</sub> |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-7DME_1)** |\n| **KPIs** | TM 0.432 \u00b7 lDDT 0.748 \u00b7 pLDDT 83.3 \u00b7 C\u03b1-RMSD 23.4 \u00c5 \u00b7 FRAUD 0.82 \u00b7 novelty 0% |\n| **Stats** | 4 depositions, 1 confidently wrong \u00b7 pLDDT\u2194lDDT +0.33 (high confidence hides the joint) |\n| **Structure** | N- and C-terminal domains on a partly-ordered linker; each folds correctly in isolation |\n| **Why AlphaFold gets it wrong** | Domains are right (lDDT 0.75) but their relative orientation is wrong (TM 0.43) \u2014 a conformational-ensemble failure, not a misfold |\n| **Biology & disease** | Strongest stimulator of the Hsp90 ATPase; drives \u0394F508-CFTR maturation (cystic-fibrosis target) and pathological tau |\n| **Key fact** | The pieces are right; only the assembly instruction is wrong |\n\n### \ud83d\udd34 NEDD4-like E3 ligase WWP2 / WW module (WWP2) \u00b7 [`6RSS_1`](https://alphafraud.mdeller.com/examples#ex-6RSS_1)\n\n| Field | Detail |\n|---|---|\n| **Take-home** | A tiny two-Trp \u03b2-sheet AlphaFold folds fine, but the deposited chain's inter-module geometry is flexible \u2014 and, tellingly, the model was honestly unsure |\n| **AlphaFold vs experimental** | <a href=\"https://alphafraud.mdeller.com/examples#ex-6RSS_1\"><img src=\"https://alphafraud.mdeller.com/ribbon/6RSS_1.svg\" alt=\"Experimental structure of 6RSS_1 coloured by C\u03b1 deviation from AlphaFold\" width=\"380\"></a><br><sub>Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \"ghost\" overlay</sub> |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-6RSS_1)** |\n| **KPIs** | TM 0.46 \u00b7 lDDT 0.503 \u00b7 pLDDT 74.5 \u00b7 C\u03b1-RMSD 21.9 \u00c5 \u00b7 FRAUD 0.71 \u00b7 novelty 8% |\n| **Stats** | 7 depositions, 1 confidently wrong \u00b7 pLDDT\u2194lDDT +0.15 (lowest confidence of the set) |\n| **Structure** | NEDD4-family HECT E3 ligase; the WW domain is a ~38-residue three-stranded antiparallel \u03b2-sheet that reads PPxY motifs |\n| **Why AlphaFold gets it wrong** | Uncertain relative orientation between small folded modules \u2014 and its low pLDDT half-admits the uncertainty |\n| **Biology & disease** | Ubiquitinates SMAD7, PTEN, TP53, OCT4; isoform balance shifts drive EMT and metastasis |\n| **Key fact** | The \"honest\" failure of the set \u2014 low confidence tracking real uncertainty |\n\n### \ud83d\udd34 Transmembrane protein 106B (TMEM106B) \u00b7 [`7U14_1`](https://alphafraud.mdeller.com/examples#ex-7U14_1)", "# The Signature Cowboy Hat of Adam Lee: A Blend of Tradition and Function\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-27 on marcdeller.com (categories: uncategorized).*\n\n* **Cultural Roots:** The hat\u2019s design honors 19th-century Western ranch traditions.\n* **Functional Design:** Felt construction offers superior UV protection and moisture control.\n* **Structural Integrity:** Reinforced crown and sweatband ensure consistent fit and form.\n* **Visual Impact:** Distinctive silhouette boosts social memorability by 30%.\n* **Color Psychology:** Cream tones foster trust and collaborative openness.\n* **Professional Branding:** Merges rugged authenticity with scientific credibility.\n\n1. **The cowboy hat\u2019s UV protection:** [Diffey, B.L., et al., Photodermatology](https://doi.org/10.1111/phd.2004.20)\n2. **Felt thermal and moisture performance:** [Kaloustian, N., et al., Textile Research Journal](https://doi.org/10.1177/TRJ.2023.003)\n3. **Memory enhancement via attire distinctiveness:** [Johnson, A.R., et al., Journal of Environmental Psychology](https://doi.org/10.1016/j.jenvp.2021.101)\n4. **Hat materials and structural design:** [Lee, S.K., et al., Journal of Textile Science](https://doi.org/10.1080/00405000.2022.145)\n5. **Color-emotion associations in professional contexts:** [Smith, T.J., et al., Frontiers in Psychology](https://doi.org/10.3389/fpsyg.2022.00987)\n6. **Role of hats in occupational ergonomics:**\n   [Marras, W.S., et al., Ergonomics](https://doi.org/10.1080/00140139.2020.1820710)\n\nA well-crafted hat does more than shield the wearer from the elements, it conveys identity, merges heritage with purpose, and evokes both confidence and comfort. For Adam Lee, Director of Advanced Analytics & Data Engineering at Incyte, the **cowboy hat** is more than sartorial flair. Forged in the crucible of Western tradition yet refined for modern utility, his choice of headgear reflects a harmonious convergence of personal history, functional requirement, and stylistic preference. Given Adam\u2019s professional trajectory, this exploration delves into the cultural resonance and scientific rationale behind his iconic headpiece.\n\nThe cowboy hat\u2019s origins trace to 19th-century ranch hands who required durable, wide-brimmed protection against sun, rain, and dust. Over time, this utilitarian necessity evolved into a symbol of rugged independence and authenticity. Adam\u2019s preference for a cream-colored felt variant nods to its historic roots in the American West while offering superior thermal regulation and moisture wicking, a key consideration for outdoor events or fieldwork. Modern felting techniques ensure consistent fiber density for optimal UV blocking, with laboratory tests confirming up to 98% UVA/UVB attenuation. Beyond sun protection, the structured crown and reinforced sweatband maintain form and fit, crucial for long hours under bright lights or in dynamic environments.\n\nThe hat\u2019s silhouette also fosters a visual narrative. In professional settings, conferences, presentations, or lab visits, it becomes a conversation catalyst, subtly conveying Adam\u2019s blend of scientific rigor and approachable leadership. Cognitive psychology research shows that distinctive attire can enhance memorability by up to 30% during social exchanges. Furthermore, the choice of neutral hues aligns with color-psychology findings: cream and tan tones evoke warmth and trust, reinforcing team cohesion and openness in collaborative meetings.\n\nUltimately, Adam Lee\u2019s cowboy hat embodies a deliberate fusion of **heritage**, **functionality**, and **professional branding**, creating an enduring motif that resonates both on the ranch and in the research lab.", "# 2025 \u2013 Introduction to Artificial Intelligence\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-27 on marcdeller.com (categories: certificates).*\n\n* [Link to certificate](https://www.coursera.org/account/accomplishments/verify/71R8R1FHA194)\n\n* **Award Details:** IBM, through Coursera, issued this certificate on March 10, 2025.\n* **Course Focus:** It recognizes completion of an introductory course on artificial intelligence.\n* **Career Value:** This matters because AI increasingly influences prediction, automation, and analysis across science.\n* **Practical Use:** It helps scientists begin applying AI concepts to modeling, imaging, and data interpretation.", "# Cellular Uptake of Microplastics: Pathways, Persistence, and Therapeutic Potential for Intracellular Targeting of PETase\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Key Concept | Description | Key References |\n| --- | --- | --- |\n| Endocytic Uptake Pathways | Size-dependent mechanisms including phagocytosis for particles above 500 nanometers, macropinocytosis for 200 to 500 nanometers, and clathrin-mediated or caveolin-mediated endocytosis for particles below 200 nanometers | [Mondal et al., Nanomaterials, 2025](https://www.mdpi.com/2079-4991/11/2/496); [Yee et al., iScience, 2022](https://pmc.ncbi.nlm.nih.gov/articles/PMC9929686/) |\n| Lysosomal Accumulation | Microplastics localize within lysosomes where they resist enzymatic degradation due to synthetic composition, persisting through cell divisions and accumulating over time | [Pichler et al., Chemosphere, 2024](https://ecancer.org/en/news/24366-micro-and-nanoplastics-in-the-body-are-passed-on-during-cell-division); [Traversa et al., Microplastics, 2024](https://pmc.ncbi.nlm.nih.gov/articles/PMC11292486/) |\n| Cancer Linkage | Microplastics promote inflammation, oxidative stress, genotoxicity, and epithelial-mesenchymal transition while accumulating preferentially in tumor tissue, potentially facilitating metastasis | [Traversa et al., Microplastics, 2024](https://pmc.ncbi.nlm.nih.gov/articles/PMC11643376/); [Mondal et al., Nanomaterials, 2025](https://www.mdpi.com/2079-4991/11/2/496) |\n| Lysosomal Targeting | Mannose-6-phosphate receptor-mediated pathways and IGF-II peptide fusion proteins enable specific delivery of therapeutic enzymes to lysosomes via receptor-mediated endocytosis | [Mondal et al., ACS Omega, 2021](https://pmc.ncbi.nlm.nih.gov/articles/PMC8757330/); [Zhu et al., PNAS, 2004](https://www.pnas.org/doi/10.1073/pnas.0400656101) |\n| HGMP01 Enzyme | Novel PET hydrolase from human gut metagenome capable of degrading PET nanoparticles, with HGMP01-like genes widely distributed and expressed in intestinal microbiome | [Zhang et al., Journal of Hazardous Materials, 2024](https://pubmed.ncbi.nlm.nih.gov/39681680/); [Zhang et al., Journal of Hazardous Materials, 2024](https://www.sciencedirect.com/science/article/pii/S0304389424032199) |\n| FAST-PETase Therapy | Machine learning-engineered enzyme that degrades PET microplastics into non-toxic monomers terephthalic acid and mono(2-hydroxyethyl) terephthalate in human serum at 37 degrees Celsius without cytotoxicity | [Lopez-Lorenzo et al., Green Chemistry, 2024](https://pubs.rsc.org/en/content/articlelanding/2024/gc/d4gc00624k); [Lu et al., Nature, 2022](https://pmc.ncbi.nlm.nih.gov/articles/PMC9007415/) |", "# GitHub README: cif2plip\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/cif2plip, 2026\n\n### CONECT generation\nBonds are perceived from covalent radii (Cordero et al., 2008) with a 0.45 \u00c5 tolerance. H\u2013H pairs are excluded. The bond table is written using strictly 80-column CONECT records and inserted immediately before the `END` record, after all serials have been made contiguous.\n\n### Serial renumbering\nAll `ATOM`, `HETATM`, and `TER` records are renumbered sequentially from 1 with no gaps. This step is the fix for the OpenBabel CONECT-parsing bug described above.\n\n---\n\n## Dependencies\n\n| Package | Role |\n|---|---|\n| [gemmi](https://gemmi.readthedocs.io) | CIF parsing and PDB writing |\n| [pdb-tools](https://www.bonvinlab.org/pdb-tools/) | `pdb_tidy`, `pdb_validate` |\n| [PLIP](https://github.com/pharmai/plip) | Protein-ligand interaction profiling |\n| [OpenBabel](https://openbabel.org) | Ligand bond-order perception (PLIP dependency) |\n| [PyMOL (OSS)](https://pymol.org) | `.pse` session generation (PLIP dependency) |\n\n---\n\n## Licence\n\nMIT\n\n---\n\n## Citation\n\nIf you use this pipeline in published work, please cite PLIP:\n\n> Adasme MF et al. (2021). PLIP 2021: expanding the scope of the protein-ligand interaction profiler to DNA and RNA. *Nucleic Acids Research*, 49(W1), W530\u2013W534. https://doi.org/10.1093/nar/gkab294\n\nAnd gemmi if the CIF conversion step is central to your workflow:\n\n> Wojdyr M (2022). GEMMI: A library for structural biology. *Journal of Open Source Software*, 7(73), 4200. https://doi.org/10.21105/joss.04200\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**  \nStructural biologist & drug discovery scientist  \n\n<table>\n<tr>\n<td>\ud83c\udf10</td><td><a href=\"https://marcdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">marcdeller.com</a></td>\n<td>\u2709\ufe0f</td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19</td><td><a href=\"https://github.com/bellcheddar/cif2plip\" target=\"_blank\" rel=\"noopener noreferrer\">github.com/bellcheddar/cif2plip</a></td>\n</tr>\n</table>", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\nTypically, 1 ml samples of culture, before and after induction with IPTG, were centrifuged at 14000 \u00d7 g and resuspended in 100 \u03bcl of protein loading buffer consisting of 10 % (v/v) glycerol, 5 % (v/v) \u03b2-mercaptoethanol, 2 % (w/v) SDS, 0.001 % (w/v) bromophenol blue and 62.5 mM Tris-HCl, pH 6.8 (see Appendix II). Magnesium chloride was added to 1 mM, to precipitate genomic DNA, and the sample vortexed. 20 \u03bcl of samples from all other steps of the preparation were retained and added to 60 \u03bcl of protein loading buffer. Samples were heated to 95 \u00b0C for 5 minutes, to denature any proteases, and then centrifuged at 14000 \u00d7 g for 10 minutes. Typically, 20 \u03bcl samples were loaded onto 12 % SDS-polyacrylamide gels (Laemmli, 1970) and electrophoresed at 150 V for approximately 1.5 hours. Broad range rainbow molecular weight markers ranging from 210 kDa to 8 kDa (Bio-Rad) were run alongside the samples (see Appendix II). Gels were stained using 0.2 % Coomassie Blue R-250 in 80 % (v/v) methanol and 20 % (v/v) acetic acid (see Appendix II) for ~0.5 hours and destained in 80 % (v/v) methanol and 20 % (v/v) acetic acid overnight.\n\n### 3.2.5 Coomassie Protein Assay\n\nProtein concentrations were determined using a Coomassie Protein Plus\u2122 Assay Kit (Pierce). Standard curves were constructed using bovine serum albumin (BSA) concentrations in the range 0 to 1.6 mg/ml. Typically, 50 \u03bcl of sample was added to 1.5 ml of Coomassie Protein Plus\u2122 Reagent and the absorbance measured at a wavelength of 595 nm using a Beckman DU-640 spectrophotometer. Protein concentrations were determined by comparison with the absorbance measurements of BSA standards. The accuracy of this assay was confirmed by comparison with direct absorbance measurements at 280 nm.\n\n### 3.2.6 Western Blot Analysis\n\nProteins were separated on a 4\u201315 % gradient SDS polyacrylamide gel (Bio-Rad). The gel was equilibrated in western blot transfer buffer consisting of 0.3 % (w/v) Tris-base, 1.4 % (w/v) glycine and 10 % (v/v) methanol (see Appendix II) for 1 hour. Electrophoretic transfer of proteins was routinely performed using a Mini-Protean II wet blotter (Bio-Rad). The gel was placed on a pre-wetted section of 0.45 \u03bcm nitrocellulose (Schleicher and Schuell) trimmed to the dimensions of the gel. Three layers of pre-wetted filter paper (Whatman 3MM\u2122) were placed on each side and the sandwich blotted overnight at 30 V or for 1 hour at 100 V at 4 \u00b0C.\n\nAfter transfer, the nitrocellulose was removed and blocked overnight with 5 % (w/v) milk (Marvel\u2122) in phosphate buffered saline (PBS) (see Appendix II). Primary polyclonal antibody (sheep \u03b1-oncostatin M; Binding Site, U.K., 8.7 mg/ml) was added at a dilution of 1:10000 in PBS containing 5 % (w/v) milk and 0.1 % (v/v) Tween-20. The blot was incubated at room temperature for 1 hour, with agitation, and then extensively washed in PBS containing 0.1 % (v/v) Tween-20. Secondary polyclonal antibody (donkey \u03b1-sheep IgG1-peroxidase conjugated, Binding Site, U.K.) was applied at a dilution of 1:5000 in PBS as above. The blot was incubated at room temperature for 1 hour, with agitation, and then extensively washed in PBS as above. Finally, the blot was washed with PBS and luminescent detection of the horseradish peroxidase-labelled secondary antibody carried out using ECL\u2122 SuperSignal Substrate (Amersham). Blots were wrapped in ClingFilm\u2122 and then covered with HyperFilm\u2122 (Amersham) and exposed for 10\u201360 seconds. Films were developed automatically (X-Omat).\n\n### 3.2.7 Seleno-methionine Labelled Oncostatin M\n\nSelenium labelled OSM (78Se-Met) was prepared according to the method of Hendrickson (Hendrickson et al., 1990). pGEX-3C-\u0394188_OSM constructs were transformed as previously (see Chapter 2 for method) into methionine auxotrophic strains of E. coli in an attempt to maximise 78Se-Met incorporation. Strains tested included B834 and DL41, which were both obtained from the European culture collection (ECACC).\n\nTransformed cells were grown in defined LeMasters medium (LeMaster and Richards, 1982) containing glucose, seventeen amino acids, nucleotide bases, various salts and sulphates. The media is produced in two portions; a non-autoclavable portion and an autoclavable portion (see Table 3-2). Autoclavable portions were made up to 4 L with water and the supersaturated solution was autoclaved immediately. The non-autoclavable portion was made up to 400 ml with water and then filter sterilised through a 0.22 \u03bcm filter (Whatman). Both the autoclavable and non-autoclavable fractions were stored at room temperature, to prevent precipitation. Once cool the non-autoclavable fraction was added to the autoclavable fraction and the medium used immediately.\n\nCell growth was carried out as previously (see 3.2.1) except only 2 ml of starter culture was used in an attempt to reduce the incorporation of non-labelled methionine. Affinity purification and cleavage of the fusion proteins was also carried out as previously (see 3.2.2 and 3.2.3) except for the addition of 1 mM DTT during purification in an attempt to minimise oxidation.\n\n**Non-autoclavable portion (400 ml)**\n\n| Chemical class | Component | Quantity |\n|---|---|---|\n| Nutrient source | Glucose | 40.00 g |\n| Nutrient source | MgSO4\u00b77H2O | 1.00 g |\n| Nutrient source | FeSO4 | 16.7 mg |\n| pH adjustment | conc. H2SO4 | 33.2 \u03bcl |\n| Vitamin | Thiamine-HCl | 20 mg |\n| Label | Seleno-Methionine | 200 mg |\n\n**Autoclavable portion**", "# 2013 \u2013 Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nContemporary research has revealed the dynamic nature of this immune battleground, where viral evolution and antibody responses engage in constant molecular warfare. Studies of patient samples show that HIV-1 frequently attempts to escape N332-directed antibodies through multiple mechanisms, including eliminating the N332 glycan site entirely or developing resistance through structural modifications in variable regions while retaining key epitopes. However, the virus faces a fundamental dilemma: removing glycans to escape antibody recognition exposes conserved epitopes to immune surveillance, while maintaining the glycan shield preserves vulnerability to glycan-dependent bnAbs. This evolutionary constraint helps explain why the Asn332 supersite remains a persistent vulnerability despite strong selective pressure.\n\nThe clinical significance of the supersite extends beyond basic science into practical therapeutic applications. Recent isolations of potent bnAbs from HIV-infected donors demonstrate that natural immune responses can indeed target this region effectively, with some antibodies achieving neutralization breadth comparable to well-characterized therapeutics like VRC01. These discoveries have profound implications for both passive immunotherapy and vaccine design, suggesting that immunogens capable of presenting the supersite in appropriate configurations might elicit protective responses. The supersite\u2019s conservation across diverse viral strains, combined with its ability to accommodate multiple recognition modes, makes it an ideal target for broad-spectrum interventions.\n\nModern vaccine development strategies increasingly focus on exploiting the supersite\u2019s unique properties through structure-guided design approaches. Researchers are developing immunogens that display the glycan-protein epitopes in optimal configurations to engage germline precursors of bnAbs, addressing a critical bottleneck in vaccine-induced antibody development. The intrinsic mannose patch\u2019s stability in the face of glycan site deletions provides confidence that this target will remain accessible even as the virus attempts to evolve resistance. Furthermore, the multiple binding modes possible at the supersite increase the probability of eliciting diverse antibody responses that could collectively provide robust protection against viral escape.\n\n| Key Concept | Description | Key References |\n| --- | --- | --- |\n| Asn332 Supersite | A highly accessible region of immune vulnerability on HIV-1 gp120 encompassing multiple glycans and underlying protein surface, allowing diverse antibody recognition modes | [Kong, L., et al., Nature Structural & Molecular Biology, 2013](https://www.nature.com/articles/nsmb.2594) |\n| Glycan Shield Penetration | Mechanism by which broadly neutralizing antibodies use extended CDR loops to reach through dense glycan coverage and access conserved epitopes | [Hitting the sweet spot, Current Opinion in HIV and AIDS, 2020](https://pmc.ncbi.nlm.nih.gov/articles/PMC7877895/) |\n| Intrinsic Mannose Patch | Region of under-processed oligomannose glycans resulting from high glycan density, creating distinctive molecular signatures for immune recognition | [Stewart-Jones, G.B., et al., Nature Communications, 2015](https://www.nature.com/articles/ncomms8479) |\n| Conformational Flexibility | Ability of antibodies to accommodate glycan heterogeneity through varied angles of engagement while maintaining neutralization efficacy | [Kong, L., et al., Structural Studies, 2013](https://dash.harvard.edu/bitstreams/7312037d-1635-6bd4-e053-0100007fdf3b/download) |\n| Viral Escape Mechanisms | HIV-1 strategies to evade glycan-dependent antibodies including glycan deletion, shifting, and structural modifications with associated fitness costs | [Multiple Mechanisms Study, Journal of Virology, 2025](https://pubmed.ncbi.nlm.nih.gov/40559564/) |\n| Therapeutic Applications | Clinical potential of supersite-targeting antibodies for both passive immunotherapy and structure-guided vaccine design approaches | [Vaccination Study, Nature Immunology, 2024](https://www.nature.com/articles/s41590-024-01833-w) |", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\n**32.** An antibody that binds to human mutant calreticulin (CALR), wherein: the antibody comprises a heavy chain variable region (VH) comprising a VH CDR1; a VH CDR2; and a VH CDR3; wherein: the VH CDR1 comprises the amino acid sequence ELSMQ (SEQ ID NO:1); the VH CDR2 comprises the amino acid sequence GFDPDDX101ETMYAEX102X103QG (SEQ ID NO:102); wherein X101 is D or G; wherein X102 is K or R; and wherein X103 is F or L; the VH CDR3 comprises the amino acid sequence SPGYDFFDY (SEQ ID NO:18); wherein the antibody comprises a light chain variable region (VL) comprising a VL CDR1, a VL CDR2, and a VL CDR3, wherein: the VL CDR1 comprises the amino acid sequence GGX104X105X106GX107X108X109VX110 (SEQ ID NO:103), wherein X104 is N, D, or S; wherein X105 is Y, N, or D; wherein X106 is I or T; wherein X107 is S, D, I, R, or T; wherein X108 is K, E, or I; wherein X109 is S, I, R, G, N, or A; and wherein X110 is H, F, or N; the VL CDR2 comprises the amino acid sequence DDX111DRPX112 (SEQ ID NO:104), wherein X111 is G, S, or R; and wherein X112 is S or L; and the VL CDR3 comprises the amino acid sequence QVWDX113X114X115DX116X117X118 (SEQ ID NO:105), wherein X113 is S or A; wherein X114 is I or S; wherein X115 is S, I, or N; wherein X116 is H, L, or Q; wherein X117 is V or L; and wherein X118 is V or I.\n\n**33.** The antibody of claim 32, wherein the VH CDR1 comprises the amino acid sequence GYTLTELSMQ (SEQ ID NO:329).\n\n**34.** An antibody that binds to human mutant calreticulin (CALR), wherein: the antibody comprises a VH CDR1 comprising ELSMQ (SEQ ID NO:1); a VH CDR2 comprising GFDPDDX101ETMYAEX102X103QG (SEQ ID NO:102), wherein X101 is D or G, X102 is K or R, and X103 is F or L; a VH CDR3 comprising SPGYDFFDY (SEQ ID NO:18); a VL CDR1 comprising TGTSSDVGGYNYVS (SEQ ID NO:30); a VL CDR2 comprising X119VSX120RPS (SEQ ID NO:106), wherein X119 is E or D and X120 is N or K; and a VL CDR3 comprising QVWDSSNDLLI (SEQ ID NO:71).\n\n**35.** The antibody of claim 34, wherein the VH CDR1 comprises the amino acid sequence GYTLTELSMQ (SEQ ID NO:329).\n\n**36.** An antibody that binds to human mutant calreticulin (CALR), wherein the antibody comprises a VH CDR1 comprising DYFIH (SEQ ID NO:2); a VH CDR2 comprising LVDPEDGETIYAEX121FQG (SEQ ID NO:107), wherein X121 is K or R; a VH CDR3 comprising PGGILTDPDAFDI (SEQ ID NO:19); a VL CDR1 comprising X122GTX123SDVGGYNX124VS (SEQ ID NO:108), wherein X122 is T or A, X123 is S or G, and X124 is Y or H; a VL CDR2 comprising X125VX126X127RPS (SEQ ID NO:109), wherein X125 is D or E, X126 is N or S, and X127 is K or N; and a VL CDR3 comprising SSYX128X129SSTX130X131V (SEQ ID NO:110), wherein X128 is I or T, X129 is P or S, X130 is R, P, F, or absent, and X131 is W or Y.\n\n**37.** An antibody that binds to human mutant calreticulin (CALR), wherein the antibody comprises a VH CDR1 comprising SYAIS (SEQ ID NO:3); a VH CDR2 comprising LVDPEDGETIYAEKFX132G (SEQ ID NO:111), wherein X132 is R or Q; a VH CDR3 comprising EESYGP (SEQ ID NO:20); a VL CDR1 comprising QASQDISNYLX133 (SEQ ID NO:112), wherein X133 is N or D; a VL CDR2 comprising DASNLET (SEQ ID NO:61); and a VL CDR3 comprising QQLNSYPLT (SEQ ID NO:80).\n\n**38.** An antibody that binds to human mutant calreticulin (CALR), wherein the antibody comprises a VH CDR1 comprising EX134SMH (SEQ ID NO:113), wherein X134 is S or L; a VH CDR2 comprising LVDPEDGETIYAQKFQG (SEQ ID NO:14); a VH CDR3 comprising EEWSGDGDDAFDI (SEQ ID NO:21); a VL CDR1 comprising SGSSSNIGSYSVS (SEQ ID NO:46); a VL CDR2 comprising DX135NKRPS (SEQ ID NO:114), wherein X135 is N or D; and a VL CDR3 comprising GTWDSSLSAWV (SEQ ID NO:81).\n\n**39.** An antibody that binds to human mutant calreticulin (CALR), wherein the antibody comprises a VH CDR1 comprising SYAIS (SEQ ID NO:3); a VH CDR2 comprising GIIPIFGTANYAQKFQG (SEQ ID NO:15); a VH CDR3 comprising SPLRGSGWYWHYYYGMDV (SEQ ID NO:22); a VL CDR1 comprising GGNX136IX137X138KX139VH (SEQ ID NO:115), wherein X136 is N or K, X137 is R or G, X138 is A, S, R, or T, and X139 is H or S; a VL CDR2 comprising X140DX141X142RPS (SEQ ID NO:116), wherein X140 is Q or R, X141 is S or R, and X142 is N or K; and a VL CDR3 comprising QX143WX144SX145TX146V (SEQ ID NO:117), wherein X143 is A or V, X144 is D or G, X145 is S or N, and X146 is V, A, or E.\n\n**40.** An antibody that binds to human mutant calreticulin (CALR), wherein the antibody comprises a VH CDR1 comprising TSGMGVG (SEQ ID NO:6); a VH CDR2 comprising HIWWDDDKYYNPYLKN (SEQ ID NO:17); a VH CDR3 comprising SAYGSTYGY (SEQ ID NO:25); a VL CDR1 comprising KASQSVDYDGDSFMN (SEQ ID NO:52); a VL CDR2 comprising TASILES (SEQ ID NO:68); and a VL CDR3 comprising QQSNEDPWT (SEQ ID NO:91).\n\n**41.** The antibody of claim 40, wherein the VH is at least 80% identical to the amino acid sequence of SEQ ID NO:182; and the VL is at least 80% identical to the amino acid sequence of SEQ ID NO:305.\n\n**42.** The antibody of claim 41, wherein the VH comprises the amino acid sequence of SEQ ID NO:182; and the VL comprises the amino acid sequence of SEQ ID NO:305.\n\n**43.** The antibody of claim 42, wherein the heavy chain comprises the amino acid sequence of SEQ ID NO:136, and the light chain comprises the amino acid sequence of SEQ ID NO:250.\n\n**44.** An antibody that binds to human mutant calreticulin (CALR), wherein the antibody: inhibits one or more signaling pathways downstream of thrombopoietin receptor (MPL) in a cell expressing human mutant CALR; inhibits oncogenic cell proliferation in a cell expressing human mutant CALR; inhibits dimerization of MPL in a cell expressing human mutant CALR, and/or inhibits binding of MPL to human mutant CALR.", "# AlphaFraud \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n| Field | Detail |\n| --- | --- |\n| **Take-home** | A bistable three-helix Ca\u00b2\u207a switch; AlphaFold committed to one conformer, the structure is the other |\n| **AlphaFold vs experimental** | Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \u201cghost\u201d overlay |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-6YEL_1)** |\n| **KPIs** | TM 0.389 \u00b7 lDDT 0.341 \u00b7 pLDDT 78.9 \u00b7 C\u03b1-RMSD 35.8 \u00c5 \u00b7 FRAUD 0.75 \u00b7 novelty 2% |\n| **Stats** | 1 deposition, confidently wrong \u00b7 SS-Q3 69.6% (secondary structure right, packing wrong) |\n| **Structure** | CC1 is a compact antiparallel three-helix bundle that toggles between clamped (resting) and extended (active) states |\n| **Why AlphaFold gets it wrong** | Right fold class, wrong tertiary packing \u2014 a single-answer predictor cannot represent a two-state switch |\n| **Biology & disease** | Gates store-operated Ca\u00b2\u207a entry via Orai1; gain-of-function \u2192 Stormorken / tubular aggregate myopathy; loss \u2192 CRAC-channelopathy |\n| **Key fact** | 35.8 \u00c5 RMSD with correct secondary structure = a conformer error, not a misfold |\n\n### \ud83d\udd34 Activator of Hsp90 ATPase 1 / Aha1 (AHSA1) \u00b7 [`7DME_1`](https://alphafraud.mdeller.com/examples#ex-7DME_1)\n\n| Field | Detail |\n| --- | --- |\n| **Take-home** | Two well-folded domains bolted together at the wrong angle \u2014 a flexible-linker ensemble collapsed to one guess |\n| **AlphaFold vs experimental** | Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \u201cghost\u201d overlay |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-7DME_1)** |\n| **KPIs** | TM 0.432 \u00b7 lDDT 0.748 \u00b7 pLDDT 83.3 \u00b7 C\u03b1-RMSD 23.4 \u00c5 \u00b7 FRAUD 0.82 \u00b7 novelty 0% |\n| **Stats** | 4 depositions, 1 confidently wrong \u00b7 pLDDT\u2194lDDT +0.33 (high confidence hides the joint) |\n| **Structure** | N- and C-terminal domains on a partly-ordered linker; each folds correctly in isolation |\n| **Why AlphaFold gets it wrong** | Domains are right (lDDT 0.75) but their relative orientation is wrong (TM 0.43) \u2014 a conformational-ensemble failure, not a misfold |\n| **Biology & disease** | Strongest stimulator of the Hsp90 ATPase; drives \u0394F508-CFTR maturation (cystic-fibrosis target) and pathological tau |\n| **Key fact** | The pieces are right; only the assembly instruction is wrong |\n\n### \ud83d\udd34 NEDD4-like E3 ligase WWP2 / WW module (WWP2) \u00b7 [`6RSS_1`](https://alphafraud.mdeller.com/examples#ex-6RSS_1)\n\n| Field | Detail |\n| --- | --- |\n| **Take-home** | A tiny two-Trp \u03b2-sheet AlphaFold folds fine, but the deposited chain\u2019s inter-module geometry is flexible \u2014 and, tellingly, the model was honestly unsure |\n| **AlphaFold vs experimental** | Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \u201cghost\u201d overlay |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-6RSS_1)** |\n| **KPIs** | TM 0.46 \u00b7 lDDT 0.503 \u00b7 pLDDT 74.5 \u00b7 C\u03b1-RMSD 21.9 \u00c5 \u00b7 FRAUD 0.71 \u00b7 novelty 8% |\n| **Stats** | 7 depositions, 1 confidently wrong \u00b7 pLDDT\u2194lDDT +0.15 (lowest confidence of the set) |\n| **Structure** | NEDD4-family HECT E3 ligase; the WW domain is a ~38-residue three-stranded antiparallel \u03b2-sheet that reads PPxY motifs |\n| **Why AlphaFold gets it wrong** | Uncertain relative orientation between small folded modules \u2014 and its low pLDDT half-admits the uncertainty |\n| **Biology & disease** | Ubiquitinates SMAD7, PTEN, TP53, OCT4; isoform balance shifts drive EMT and metastasis |\n| **Key fact** | The \u201chonest\u201d failure of the set \u2014 low confidence tracking real uncertainty |\n\n### \ud83d\udd34 Transmembrane protein 106B (TMEM106B) \u00b7 [`7U14_1`](https://alphafraud.mdeller.com/examples#ex-7U14_1)\n\n| Field | Detail |\n| --- | --- |\n| **Take-home** | AlphaFold built a tidy globular \u03b2-sandwich at pLDDT 94; the real structure is a novel brain amyloid \u2014 every ordered residue wrong |\n| **AlphaFold vs experimental** | Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \u201cghost\u201d overlay |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-7U14_1)** |\n| **KPIs** | TM 0.157 \u00b7 lDDT 0.536 \u00b7 pLDDT 94.4 \u00b7 C\u03b1-RMSD 27.1 \u00c5 \u00b7 FRAUD 0.87 \u00b7 novelty **100%** |\n| **Stats** | 29 depositions, **27** confidently wrong \u00b7 pLDDT\u2194lDDT 0.03 (confidence uninformative) |\n| **Structure** | Native luminal domain is a fibronectin-III \u03b2-sandwich; the deposited form is a five-layered cross-\u03b2 amyloid core |\n| **Why AlphaFold gets it wrong** | Amyloid is a self-templating polymer state absent from training, and the sequence is 100% novel \u2014 nothing steers it off its globular prior |\n| **Biology & disease** | Lysosomal protein; top GWAS hit for FTLD-TDP; its C-terminal domain forms fibrils across ageing and neurodegeneration |\n| **Key fact** | 100% sequence-novel \u2014 no pre-2018 fibril template existed |\n\n### \ud83d\udd34 Alpha-synuclein (SNCA) \u00b7 [`9A1Q_1`](https://alphafraud.mdeller.com/examples#ex-9A1Q_1)", "# Revolutionary AI Tool XenoBug Transforms Environmental Pollution Control Through Predictive Enzyme Discovery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-20 on marcdeller.com (categories: blog, microplastics).*\n\n* Revolutionary AI Platform: XenoBug uses machine learning to predict bacterial enzymes capable of degrading environmental pollutants, trained on 6,814 substrates and 141,200 biochemical reactions with >75% prediction accuracy\n* Massive Database Integration: Platform incorporates 3.3 million environmental enzyme sequences and 16 million enzymes from 38,000 bacterial genomes, enabling discovery of both known and novel pollutant-degrading enzymes\n* Multilabel Classification System: Advanced AI architecture using Random Forest and Neural Networks predicts multiple enzyme types simultaneously for complex pollutant degradation pathways\n* Validated Performance: Successfully predicted enzymes for diverse pollutants including pesticides (triazophos), hydrocarbons (p-xylene), and persistent compounds like DDT and toxaphene where degrading enzymes were previously unknown\n* Global Accessibility: Free web-based tool democratizes advanced environmental biotechnology research, enabling worldwide collaboration for rapid response to contamination challenges\n* Practical Applications: Municipal authorities already testing XenoBug-derived enzyme solutions for landfill treatment, agricultural runoff management, and contaminated aquifer restoration\n* Metagenomic Source Identification: Platform identifies specific environmental sources (marine, soil, compost metagenomes) of predicted enzymes, enabling targeted deployment strategies\n\n1. Malwe, A.S., Longwani, U., & Sharma, V.K. (2025). **XenoBug: machine learning-based tool to predict pollutant-degrading enzymes from environmental metagenomes.** NAR Genomics and Bioinformatics, 7(2), lqaf037:\n   <https://academic.oup.com/nargab/article/7/2/lqaf037/8123453>\n2. Malwe, A.S., Longwani, U., & Sharma, V.K. (2025). M**achine learning-based tool to predict pollutant-degrading enzymes**:\n   <https://pubmed.ncbi.nlm.nih.gov/40314024/>\n3. **AI tool finds bacterial enzymes to remove pollutants.** (2025). Nature:\n   <https://www.nature.com/articles/d44151-025-00117-y>\n4. Blessing, A.-A., & Olateru, K.(2025). **AI-driven optimization of bioremediation strategies for river pollution: a comprehensive review and future directions.** Frontiers in Microbiology:\n   <https://www.frontiersin.org/journals/microbiology/articles/10.3389/fmicb.2025.1504254/full>\n5. Bhatt, P., et al. (2021). **Recent Advanced Technologies for the Characterization of Xenobiotic-Degrading Microorganisms and Microbial Communities.** PMC:\n   <https://pmc.ncbi.nlm.nih.gov/articles/PMC7902726/>\n6. **How AI and Enzyme Science Are Rewriting the Future of Pollution Control.** (2025). BioGlobe:\n   <https://bioglobe.co.uk/how-ai-and-enzyme-science-are-rewriting-the-future-of-pollution-control>\n7. **Research into marine plastic pollution reveals bacterial enzymes actively degrading plastic.** (2024). Phys.org:\n   <https://phys.org/news/2024-02-marine-plastic-pollution-reveals-bacterial.html>\n8. **XenoBug \u2013 IISER Bhopal.** (2023). IISER Bhopal MetaBioSys:\n   <https://metabiosys.iiserb.ac.in/xenobug/>\n\nThe global environmental crisis demands innovative solutions, and a groundbreaking artificial intelligence platform called XenoBug is emerging as a powerful weapon in the fight against pollution. Developed by researchers at the Indian Institute of Science Education and Research Bhopal, this machine learning-based tool represents a paradigm shift in how we approach environmental remediation, offering unprecedented capability to predict and identify bacterial enzymes capable of degrading toxic pollutants.\n\nEnvironmental contamination has reached critical levels worldwide, with industrial growth, urbanization, and modernized agriculture contributing to the accumulation of synthetic chemicals including pesticides, plastics, petroleum products, and pharmaceutical waste. These xenobiotic compounds\u2014foreign to natural ecosystems\u2014pose severe threats to ecological systems and human health, with environmental pollutants contributing to approximately 22% of the global disease burden. Traditional bioremediation methods, while environmentally friendly, have been limited by time-consuming laboratory work, expensive analytical techniques, and the challenge of identifying appropriate bacterial enzymes for specific contaminants.\n\nXenoBug addresses these limitations through sophisticated artificial intelligence algorithms trained on an extensive database of 6,814 diverse substrates involved in approximately 141,200 biochemical reactions. The platform integrates multiple machine learning approaches, including Random Forest classifiers and Artificial Neural Networks, utilizing a hybrid feature set comprising 1,603 molecular descriptors and linear and circular fingerprints to capture the structural and physicochemical properties of pollutant molecules. This comprehensive approach enables the tool to achieve remarkably high prediction accuracy, with binary accuracies exceeding 0.75 and F1 scores greater than 0.62 across different reaction classes.\n\nThe revolutionary aspect of XenoBug lies in its massive database integration, incorporating approximately 3.3 million enzyme sequences from environmental metagenome databases and 16 million enzymes from 38,000 bacterial genomes. This vast repository allows researchers to identify not only known pollutant-degrading enzymes but also previously unreported metabolic enzymes capable of biotransforming specific contaminants. The platform operates through a modular architecture consisting of three interconnected modules: reaction class prediction, reaction subclass prediction, and complete enzyme identification through structural similarity searches.", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nAt various places in the present specification, variables defining divalent linking groups may be described. It is specifically intended that each linking substituent include both the forward and backward forms of the linking substituent. For example, \u2014NR(CR'R\")n\u2014 includes both \u2014NR(CR'R\")n\u2014 and \u2014(CR'R\")nNR\u2014 and is intended to disclose each of the forms individually. Where the structure requires a linking group, the Markush variables listed for that group are understood to be linking groups. For example, if the structure requires a linking group and the Markush group definition for that variable lists \"alkyl\" or \"aryl\" then it is understood that the \"alkyl\" or \"aryl\" represents a linking alkylene group or arylene group, respectively.\n\nAs used herein, the phrase \"optionally substituted\" means unsubstituted or substituted. The substituents are independently selected, and substitution may be at any chemically accessible position. As used herein, the term \"substituted\" means that a hydrogen atom is removed and replaced by a substituent. A single divalent substituent, e.g., oxo, can replace two hydrogen atoms. It is to be understood that substitution at a given atom is limited by valency, that the designated atom's normal valency is not exceeded, and that the substitution results in a stable compound.\n\nAs used herein, the term \"independently selected from\" means that each occurrence of a variable or substituent are independently selected at each occurrence from the applicable list. When any variable (e.g., R^G) occurs more than one time in any constituent or formula for a compound, its definition at each occurrence is independent of its definition at every other occurrence.\n\n### Definitions\n\nThroughout the definitions, the term \"Cn-m\" indicates a range which includes the endpoints, wherein n and m are integers and indicate the number of carbons. Examples include C1\u20133, C1\u20134, C1\u20136, and the like.\n\nAs used herein, \"Cn-m alkyl\", employed alone or in combination with other terms, refers to a saturated hydrocarbon group that may be straight-chain or branched, having n to m carbons. Examples of alkyl moieties include methyl (Me), ethyl (Et), n-propyl (n-Pr), isopropyl (i-Pr), n-butyl, tert-butyl, isobutyl, sec-butyl; higher homologs such as 2-methyl-1-butyl, n-pentyl, 3-pentyl, n-hexyl, 1,2,2-trimethylpropyl, and the like.\n\n\"Cn-m alkenyl\" refers to an alkyl group having one or more double carbon-carbon bonds and having n to m carbons; examples include ethenyl, n-propenyl, isopropenyl, n-butenyl, sec-butenyl. \"Cn-m alkynyl\" refers to an alkyl group having one or more triple carbon-carbon bonds; examples include ethynyl, propyn-1-yl, propyn-2-yl. \"Cn-m alkoxy\" refers to a group of formula -O-alkyl. \"Amino\" refers to a group of formula \u2014NH2.\n\n\"Aryl,\" employed alone or in combination with other terms, refers to an aromatic hydrocarbon group, which may be monocyclic or polycyclic (e.g., having 2 fused rings). \"Cn-m aryl\" refers to an aryl group having from n to m ring carbon atoms. In some embodiments, the aryl group has 6 to 10 carbon atoms; in some embodiments the aryl group is phenyl or naphthyl; in some embodiments the aryl is phenyl.\n\n\"Halo\" refers to F, Cl, Br, or I. In some embodiments, halo is F, Cl, or Br. In some embodiments, halo is F or Cl. In some embodiments, halo is F. In some embodiments, halo is Cl.\n\n\"Cn-m haloalkoxy\" refers to a group of formula \u2014O-haloalkyl having n to m carbon atoms. Example haloalkoxy groups include OCF3 and OCHF2. In some embodiments, the haloalkoxy group is fluorinated only.\n\n\"Cn-m haloalkyl\", employed alone or in combination with other terms, refers to an alkyl group having from one halogen atom to 2s+1 halogen atoms which may be the same or different, where \"s\" is the number of carbon atoms in the alkyl group, wherein the alkyl group has n to m carbon atoms. Example haloalkyl groups include CF3, C2F5, CHF2, CH2F, CCl3, CHCl2, C2Cl5 and the like. \"Cn-m fluoroalkyl\" refers to an alkyl group having from one fluoro atom to 2s+1 fluoro atoms; example fluoroalkyl groups include CF3, C2F5, CHF2, CH2F and the like.", "# From Bench to Regulation: How the FDA Is Formalising a Future Beyond Animal Testing\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nThe 2025 follow-on paper, Yao et al. in International Journal of Toxicology (2025), documented how FDA/CDER has already incorporated specific NAMs into approved drug programmes, including assays that supported expanded indications for cystic fibrosis and Fabry disease therapeutics without requiring additional clinical trials. Taken together, the March 2026 draft guidance and its accompanying scientific record represent a coherent shift in regulatory philosophy: human-centric, mechanistically grounded data are no longer supplementary; with appropriate validation, they are becoming the standard. For drug developers, the message is clear \u2014 early dialogue with FDA review divisions, rigorous experimental design, and transparent reporting of NAM context and limitations will determine how quickly these tools gain traction in regulatory submissions.", "# The Unsung Hero: How a Single Chlorine Atom Transforms Drug Discovery\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nWhat makes chlorine particularly valuable in structure-activity relationship (SAR) exploration is its unique electronic versatility. As both an electron-withdrawing group (through inductive effects) and an electron-donating group (through resonance), chlorine can mimic methyl, trifluoromethyl, cyano, and even hydroxyl substituents depending on molecular context. This chameleonic behavior explains chlorine\u2019s prominent position in the [Topliss decision tree](https://www.tandfonline.com/doi/full/10.1080/17568919.2025.2485865)\u2014a systematic medicinal chemistry optimization scheme where para-chlorophenyl derivatives serve as the preferred first-tier analog. The synthetic accessibility of chlorination via electrophilic aromatic substitution makes this modification both practical and cost-effective for early-stage drug discovery programs.\n\nHigh-resolution crystallographic studies have revealed the structural basis for chlorine\u2019s potency-enhancing effects. For instance, in [eIF4E inhibitor co-crystal structures](https://www.science.org/content/blog-post/chlorine-magic), the para-chlorophenyl moiety of compound 12B fits the binding pocket with remarkable complementarity, making extensive van der Waals contacts with Phe48, Leu60, and Ser92 while the chlorine atom forms a stabilizing halogen bond with the serine hydroxyl. This 729-fold improvement over the non-chlorinated parent compound (12A) exemplifies how chlorine substitution can transform weak micromolar binders into nanomolar-affinity ligands suitable for clinical development. Similar structural insights from HIV reverse transcriptase, PfDHODH, and various kinase inhibitor complexes demonstrate that chlorine atoms consistently occupy small hydrophobic cavities while lowering ligand pKa values to optimize hydrogen bonding networks.\n\nBeyond potency enhancement, chlorine profoundly influences drug metabolism and pharmacokinetics. [Computational ADME studies conducted between 2020-2025](https://pmc.ncbi.nlm.nih.gov/articles/PMC10687521/) revealed that chlorinated analogs exhibit increased chemical stability (reflected in larger HOMO-LUMO gaps) without compromising gastrointestinal absorption or increasing toxicity. Human liver microsome (HLM) assays with chlorinated cathinones demonstrated that strategic chlorine placement can modulate intrinsic clearance rates\u20143-chloro isomers showing faster metabolism than their 4-chloro counterparts. This metabolic tunability provides medicinal chemists with a powerful tool for optimizing drug exposure profiles. For example, aliphatic chlorides once dismissed as reactive liabilities have reemerged as [stable bioisosteres](https://chemrxiv.org/engage/chemrxiv/article-details/68af37dba94eede154f1b5ba) when properly positioned, as evidenced by FDA-approved drugs like sucralose and asciminib.\n\nThe privileged 7-chloroquinoline motif illustrates chlorine\u2019s role in defining pharmacophore essentiality. Present in essential medicines including chloroquine, hydroxychloroquine, and amodiaquine, this structural element exploits chlorine\u2019s electron-withdrawing properties to fine-tune the quinoline nitrogen\u2019s basicity. By lowering pKa values by approximately 0.4 units, the 7-chloro substituent positions these antimalarial agents within the optimal protonation range for concentrating within the parasite\u2019s acidic digestive vacuole\u2014a mechanism-based design principle validated across dozens of antiplasmodial and antitrypanosomal programs spanning 2020-2025.\n\nEmerging applications of dichlorinated scaffolds reveal that dual chlorine substitutions can produce synergistic effects exceeding single-chloro enhancements. [Recent heterocyclic syntheses](https://pmc.ncbi.nlm.nih.gov/articles/PMC12088962/) featuring 3,4-dichlorophenyl and 2,4-dichlorophenyl motifs demonstrated >10,000-fold potency improvements against bacterial DNA gyrase and various cancer cell lines. The two-chloro effect arises from conformational rigidification that pre-organizes ligands into bioactive geometries, combined with enhanced lipophilic filling of binding pockets and strengthened halogen bonding networks. For instance, FGFR1 kinase inhibitor 26B achieved a remarkable 12,363-fold potency improvement over its non-chlorinated parent through orthogonal dichlorination that renders the phenyl ring perpendicular to the core scaffold, optimizing hydrophobic contacts.\n\nSelectivity\u2014often the Achilles\u2019 heel of kinase inhibitors and GPCR modulators\u2014can also be dramatically enhanced through judicious chlorination. Studies of dopamine D3 receptor antagonists revealed that para-chloro substitution on compound 22B improved D3 selectivity over D2 by >3,300-fold, despite only a 16-fold potency increase. This extraordinary selectivity gain stems from chlorine\u2019s ability to engage D3-specific binding pocket features absent in the closely related D2 receptor. Similar [selectivity-driving chlorination strategies](https://pmc.ncbi.nlm.nih.gov/articles/PMC10872772/) have been reported for kinase inhibitors targeting MEK1 over off-target kinases, and for nuclear receptor modulators achieving PXR activation without LXR cross-reactivity.", "# Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor\nAuthors: Kong, L.; Ju, B.; Chen, Yajing; He, L.; Ren, L.; Liu, J.; Hong, K.; Su, B.; Wang, Z.; Ozorowski, G.; Ji, X.; Hua, Y.; Chen, Yanli; Deller, M.C.; Hao, Y.; Feng, Y.; Garces, F.; Wilson, R.; Dai, K.; O'Dell, S.; McKee, K.; Mascola, J.R.; Ward, A.B.; Wyatt, R.T.; Li, Y.; Wilson, I.A.; Zhu, J.; Shao, Y.\nSource: Immunity, 2016\n\nLongitudinal lineage tracing thus uncovered the emergence of this VRC01-class antibody. Although the DRVIA7-like heavy chains were first detected in 2008, we observed functional precursors in 2006 with a divergence at least 6% lower than that of the 2008 and 2009 sequences. This result placed the \"birth date\" of the DRVIA7 lineage shortly before 2006. Within 4 years, the heavy chains of DRVIA7 lineage showed increasing breadth and potency, when paired with mature VRC01 light chain (Figure 5B). In contrast, although the VRC01-class LCDR3 signature was present in the repertoire since 2006, the DRVIA7 light chains were unable to improve upon breadth and potency over time (Table S4B), supporting the notion that the light-chain interaction with gp120 glycans is a critical barrier for the VRC01 development. However, the immune system failed to overcome this barrier even though a large pool of germline light chains was generated to accelerate the maturation process (Figure 4B). To better define the range of neutralizing activity of the DRVIA7 lineage, two light-chain somatic variants from the 2009 repertoire were reconstituted with the DRVIA7 heavy chain, which displayed neutralization breadths ranging from 17% to 75% against the global panel and 24% to 60% against the DRVI panel (Tables S5C and S5D). Given the lower level of maturation, antibodies of the DRVIA7 lineage from 2006 should be less broad and potent than those from 2008 and 2009. Nonetheless, DRVIA6b, a clonal lineage utilizing a long HCDR3 loop (19 aa) and therefore a potentially different mode of gp120 binding began to emerge while DRVIA7 was declining in 2009 (Figure S4C), suggesting an active CD4bs-directed polyclonal B cell response in this Chinese donor.\n\nThe identification of DRVIA7 suggested that VRC01-class antibodies might be present in other Chinese LTNPs. This notion was supported by the repertoire analysis of five Chinese LTNPs that showed broad serum neutralization (85%\u2013100%) after 7\u201316 years of infection (Figure S5). Although the heavy chain repertoires obtained from the gene-specific method (He et al., 2014) displayed low frequencies of IgHV1-2 germline gene ranging from 0.3% to 7.2%, cross-donor analysis (Zhu et al., 2013) detected the maturation pattern of VRC01-class heavy chains for two LTNPs. Thus, a systematic, unbiased repertoire profiling will probably identify more VRC01-like antibodies from the HIV-1-infected Chinese cohort.\n\n*Figure 5. Longitudinal Analysis of the DRVIA7 Lineage Development.* (A) Identity-divergence analysis of the unbiased heavy (H) and light (\u03ba) chain repertoires of donor DRVI01 at time points 2006, 2008, and 2009, and dendrograms of synthesized heavy and light chains identified from the 2009 repertoire. For the whole-repertoire analysis (rows 1 and 3), sequences are plotted as a function of sequence identity to DRVIA7 and sequence divergence from putative germline genes. Color-coding denotes sequence density. For the analysis of DRVIA7-encoding germline gene families (rows 2 and 4), sequences are plotted as a function of sequence identity to DRVIA7 and sequence divergence from IgHV1-2 and IgKV1-5 for heavy and light chains, respectively. The germline gene family distribution is presented as black contour lines. DRVIA7 and sequences identified based on various criteria are shown on the black contours with per-mille values (\u2030) labeled accordingly. Dendrograms of functionally tested heavy and light chains are rooted by their respective germline allelic genes, IgHV1-2*02 and IgKV1-5*03, respectively. (B) Neutralization of five selected DRVIA7-like heavy chains paired with mature VRC01 light chain was tested against both global and DRVI panels for each time point, with DRVIA7 heavy chain included as a control (labeled in red). Neutralizing breadth is labeled for each antibody, which is shown as a branch of the dendrogram, and each branch is color-coded by average potency (concentration [\u03bcg/mL]: <0.1, red; 0.1\u20131, orange; 1\u201320, green; and >20, black). See also Tables S4 and S5 and Figures S4 and S5.\n\n### SGA Analysis Revealed that Rapid HIV-1 Envelope Evolution Outpaced the DRVIA7 Antibody Lineage\n\nWe utilized SGA (Salazar-Gonzalez et al., 2008) to investigate viral diversity and virus-host co-evolution in donor DRVI01. A total of 123 full-length Env genes encoding gp160 were sequenced from plasma viral RNA at 2005, 2006, 2008, and 2009. Of these, 111 non-redundant Env sequences were used to construct a neighbor-joining (NJ) phylogenetic tree (Figure 6A). The phylogenetic branches coincided precisely with the time points from which the Env sequences were derived, showing a clear time axis of viral evolution. Overall, the viruses appeared to diversify rapidly over time, as shown by the fact that the genetic distance between phylogenetic branches at 2009 was significantly greater than that at 2005. Consistently, the time point-specific NJ trees also showed a greater intra-group genetic distance for the later time points (Figure 6B).", "# Tuning the Signal: Rational Design of G Protein Selectivity\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-12-06 on marcdeller.com (categories: blog, sb).*\n\n* [Molecular \u2018glues\u2019 and \u2018bumpers\u2019 on receptors can bias signalling inside the cell](https://www.nature.com/articles/d41586-025-03461-2)\n\n* **Molecular Bumpers**: [Allosteric modulators can sterically occlude specific G protein subtypes](https://www.nature.com/articles/s41586-025-09643-2) at the intracellular interface while permitting others to bind.\n* **Conformational Selection**: Modulators stabilize distinct receptor conformations that are [energetically favorable for only a subset of G protein C-termini](https://www.nature.com/articles/s41392-024-01803-6).\n* **Rational Design**: [Small chemical changes to a modulator scaffold](https://bpspubs.onlinelibrary.wiley.com/doi/10.1111/bph.16495) can predictably alter selectivity profiles between G protein families.\n* **Therapeutic Precision**: Targeting the conserved transducer pocket provides a [universal strategy for engineering pathway-selective therapeutics](https://www.frontiersin.org/journals/endocrinology/articles/10.3389/fendo.2023.1137604/full) with reduced side effects.\n\n1. **Designing allosteric modulators to change GPCR G protein subtype selectivity:** [Moore et al., Nature, 2025](https://www.nature.com/articles/s41586-025-09643-2)\n2. **G protein-coupled receptors (GPCRs): advances in structures and their drug discovery:** [Wu et al., Signal Transduction and Targeted Therapy, 2024](https://www.nature.com/articles/s41392-024-01803-6)\n3. **Relevance of G protein\u2010coupled receptor (GPCR) dynamics for therapeutic intervention:** [Bock et al., British Journal of Pharmacology, 2024](https://bpspubs.onlinelibrary.wiley.com/doi/10.1111/bph.16495)\n4. **Allosteric modulation of G protein-coupled receptor signaling:** [Zhang et al., Frontiers in Endocrinology, 2023](https://www.frontiersin.org/journals/endocrinology/articles/10.3389/fendo.2023.1137604/full)\n\n**How allosteric modulators like SBI-553 reshape the GPCR-transducer interface to program precise signaling outcomes**\n\nG protein-coupled receptors (GPCRs) represent the largest family of drug targets, yet developing therapies that activate therapeutic pathways while avoiding side effects remains a formidable challenge. Most current drugs bind to the orthosteric site, often activating all downstream effectors\u2014G proteins and $\\beta$-arrestins\u2014indiscriminately. This lack of selectivity contributes to on-target toxicity, as the therapeutic benefit often arises from one signaling cascade while adverse events stem from another. Although \u201cbiased signaling\u201d has long been a goal, the structural determinants governing how a receptor selects one G protein subtype over another (e.g., Gq vs. GoA) have remained elusive, stalling the rational design of pathway-specific medicines.\n\nIn a significant advance, researchers recently elucidated the mechanism by which small molecules can dictate G protein preference. By focusing on the neurotensin receptor 1 (NTSR1), the team demonstrated that the allosteric modulator SBI-553 does not merely inhibit signaling but actively [sculpts the intracellular transducer pocket](https://www.nature.com/articles/s41586-025-09643-2). Using cryo-electron microscopy and computational modeling, they revealed that SBI-553 acts simultaneously as a \u201cmolecular bumper\u201d and a \u201cmolecular glue.\u201d It binds within the receptor\u2019s core, creating a steric block that prevents the engagement of specific G proteins (like Gq) while stabilizing an alternative \u201copen\u201d receptor conformation that energetically favors others (like GoA). This selectivity is driven by the ability of the G protein\u2019s C-terminal tail to adapt to the modified pocket geometry.\n\nThese findings establish a proof of principle for the rational engineering of signaling bias. By making minor chemical modifications to the SBI-553 scaffold, the authors successfully inverted the receptor\u2019s selectivity profile, creating variants that favored Gq over GoA. This confirms that the conserved intracellular pocket of GPCRs can be targeted to program precise signaling outcomes. This structure-based strategy offers a reproducible blueprint for designing \u201cbiased allosteric modulators\u201d (BAMs) across the GPCR superfamily, potentially ushering in a new generation of drugs that decouple therapeutic efficacy from unwanted side effects.", "# Discovery of (4-Pyrazolyl)-2-Aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [8UV0](https://www.rcsb.org/structure/8UV0)\n* **Release Date:** 2024-02-14\n* **Method:** X-RAY DIFFRACTION 1.55 \u00c5\n* **Organisms:** Homo sapiens\n* **Macromolecules:** Cyclin-dependent kinase 2 (protein)\n* **Unique Ligands:** XKU (Compound 17)\n* **Publication:** [Hummel, J.R., et al. (2024). Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2. Journal of Medicinal Chemistry, 67, 3112-3126.](https://pubs.acs.org/doi/10.1021/acs.jmedchem.3c02287)\n* **Blog Post:** [Discovery of 4-Pyrazolyl-2-Aminopyrimidines as Potent and Selective CDK2 Inhibitors](https://marcdeller.com/discovery-of-4-pyrazolyl-2-aminopyrimidines-as-potent-and-selective-cdk2-inhibitors/)\n\n**Description:** This structure reveals the molecular basis of unprecedented CDK2 selectivity achieved by a novel (4-pyrazolyl)-2-aminopyrimidine scaffold. The co-crystal with Compound 17 demonstrates how structural insights guided selectivity over CDKs 1, 4, 6, 7, and 9, opening a therapeutic window for CCNE1-amplified cancers.", "# ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n* \u2b1c **Weighted class-balanced sampling across all 78 classes.** Distribution shift was the proven cause of the R4 regression (exec 79\u219271%, SMARTS 100\u219255%) and SMARTS has swung 100\u219255\u2192100 across three rounds purely on dataset composition. Lock the fragile metrics by controlling sampling density rather than relying on raw example counts.\n* \u2b1c **Triple the `code_then_quote_v2` drill.** Code\u2192Quote fidelity moved from 19% (R4) to 61% (R5) but is still the weakest metric by 28 points. The failure mode is copy-fidelity, not reasoning \u2014 add examples where the quoted value must byte-match the printed `stdout` token, including trailing zeros and units, so the model learns to transcribe rather than re-derive.\n* \u2b1c **Add a dedicated `pymol_syntax_drill` generator.** PyMOL syntax has sat at 89\u201390% for two rounds and is the lowest structural biology metric. Failures cluster in selection algebra and command chaining \u2014 target those specifically rather than adding more generic structural biology classes.\n* \u2b1c **Two-phase curriculum: full run then a short fidelity-focused continuation.** Train the full 78-class dataset to plateau (~1,600 iters, matching R5 pattern), then run a short ~400-iter second phase at lower LR on a weak-metrics-only subset (code\u2192quote, PyMOL, fidelity). Phase 2 is cheap (~7 h) and concentrates remaining capacity on the actual gaps rather than fighting the plateau.\n* \u2b1c **Configure for ~2,000 iters, not 3,000.** R3 stopped at 42% of configured iters, R5 at 67% \u2014 both plateaued well before the ceiling. Stop burning 15\u201320 hours on a flat val curve; redirect that budget into the phase-2 continuation or more dataset quality work.\n* \u2b1c **Try rank 96 (not 128).** The rank 32\u219264 jump in R5 tracked with the 95% overall score but the val floor rose to 0.055 with 78 classes, suggesting the harder task is capacity-limited. Rank 96 (~400M trainable params) should sit comfortably under the R5 peak of 30.3 GB; skip rank 128 to avoid swap risk on 64 GB unified memory.\n* \u2b1c **Extend `fidelity_multistep` with longer chains and distractor numbers.** Fidelity reached 89%, but the residual 11% is multi-step cases where an intermediate value is transcribed incorrectly. Adding 4\u20135 step chains with adversarial distractor numbers forces the model to track which value to quote rather than defaulting to the largest or last number.\n* \u2b1c **Add per-class eval breakdown and run it on a mid-training checkpoint (~iter 1,000).** R4\u2019s exec/SMARTS regression only surfaced after 43 hours of training. A per-class pass/fail on the 100-example shared test set at iter 1,000 would expose class-level dilution early \u2014 while there is still time to rebalance the dataset and restart.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/ChemSage](https://github.com/bellcheddar/ChemSage) |", "# GitHub README: ChemSage\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ChemSage, 2026\n\n### Round 5 training (complete \u2014 early stop)\n- [x] Config: `config/train_config_r5.yaml` \u2014 rank=64, scale=90, iters=3000, max_seq=3072, val_batches=50\n- [x] Launch: single-process `caffeinate -i .venv/bin/python /tmp/r5_launch.py` (cache 16.7 GB, memory limit 50.1 GB)\n- [x] **Training started 2026-06-28** \u2014 Iter 1 val loss 2.210\n- [x] No rank=64 instability \u2014 converged normally\n- [x] **Early stop at iter 2000** \u2014 val loss plateaued at 0.055 from iter 1600 (best: 0.054 at iters 1650 and 1950, unsaved)\n- [x] Best available checkpoint: iter 1600, val loss **0.055** \u2192 copied to `adapters.safetensors`\n- [x] Fused into `models/chem_sage_32b_v5/` (~17.2 GB)\n- [x] Run `eval/compare/eval_compare.py` after fuse\n\n### Round 4 dataset (complete)\n- [x] 8,000 examples across 59 behaviour classes (`build_dataset.py --n 8000 --seed 47`)\n- [x] Protein family coverage: kinases (7,687 PDB), antibodies (3,322), GPCRs (103), GTPases (359), cytokines (769), receptors (6,891) \u2014 real PDB IDs and ligand SMILES from corpus\n- [x] Numerical fidelity drills: `code_then_quote` enforces \"compute then quote\" pattern; `rounding_precision` teaches MW 1dp, logP 2dp, TPSA 1dp\n- [x] Structural biology: PyMOL selection algebra, visualization recipes, PLIP interaction analysis, PDB format, pdb-tools, Biopython Bio.PDB, ChimeraX, gemmi, structure preparation, docking interpretation\n- [x] Visualization: Plotly Python (scatter, radar), R/ggplot2 (SAR plots), bio3d (RMSD, RMSF)\n- [x] Set `steps_per_report: 25` to align train/val reporting at same iters\n\n### Round 4 training (complete)\n- [x] `num_layers: 32` \u2014 48 and 64 layers tested and found too slow (~2 min/iter); 32 is the M1 Max sweet spot\n- [x] Adapter path: `adapters/chem_sage_32b_v4_lora` (checkpoints saved every 200 iters)\n- [x] Training completed 2026-06-27 10:59 BST \u2014 1,500 iters, ~43 h\n- [x] Best checkpoint: iter 950, val loss **0.041** (also tied at iter 1425)\n- [x] Copied `0000950_adapters.safetensors` \u2192 `adapters.safetensors`\n- [x] Fused into `models/chem_sage_32b_v4/` (~17.2 GB)\n- [x] Updated `eval/compare/models.yaml` with final stats and complete val loss curve\n- [x] Raised rank to 64 in Round 5 for more adapter expressivity\n\n### Evaluation and model quality\n- [x] Complete Round 3 eval: SMILES 99%, Exec 99%, Fidelity 68%\n- [x] Comparative eval harness: `eval/compare/eval_compare.py` with 13 metrics, loss curves, resource stats\n- [x] Comparative eval run after R4 fuse: full 4-round report `eval/compare/results/compare_20260627_1753.html`\n- [x] Run R5 comparative eval after fuse: add Round 5 to `eval/compare/models.yaml` and run `eval_compare.py`\n- [x] Improve exec/PyExec from R4's 74%/75% to >85% (R5 result: exec 95%, PyExec 99%)\n- [ ] Add perplexity benchmark against base (unfine-tuned) 32B\n- [ ] Add hallucination rate metric: count invented PDB IDs, ChEMBL IDs, SMILES not in corpus\n\n### Infrastructure\n- [x] ChromaDB 1.x fix: `allow_reset=True` in `rag/retrieve.py` and `scripts/ingest_rag.py` \u2014 55,742 chunks preserved, no re-ingest required\n- [x] `psutil` added to `requirements.txt` for server process monitoring in comparative eval\n- [x] Set up HuggingFace Hub: all 5 models published at [huggingface.co/Dellboy](https://huggingface.co/Dellboy)\n- [ ] Create `scripts/run_eval.sh` (start server, wait, run eval, stop server)\n\n### Publishing and deployment\n- [x] **Live web deployment at chemsage.mdeller.com** \u2014 shipped as an xterm.js terminal in the browser, not a chat-bubble Gradio UI: the real, unmodified `scripts/chat.py` runs server-side in a PTY, bridged to the browser over a WebSocket, so the hosted session is identical to running the CLI locally (same slash commands, same Rich rendering). No local GPU on the droplet, so `mlx_lm`'s `generate`/`stream_generate` are patched at import time to call a HuggingFace Space running the real fused model on ZeroGPU instead \u2014 see **Web deployment** below.\n- [ ] **Packaged CLI installer** \u2014 publish `chat.py` as a `pipx`-installable tool (`pyproject.toml` entry point); users with a local Qwen 32B model can `pipx install chemsage` and point it at their own `mlx_lm.server` endpoint.\n\n### Round 6 (planned)\n\nIdeas to try next, based on learnings from Rounds 1\u20135:", "# Crystal Structure of Acetyl Xylan Esterase (TM0077) from Thermotoga maritima at 2.40 \u00c5 Resolution\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [3M82](https://www.rcsb.org/structure/3M82)\n* **Release Date:** 2010-05-26\n* **Method:** X-RAY DIFFRACTION 2.4 \u00c5\n* **Organisms:** Thermotoga maritima\n* **Macromolecules:** Acetyl xylan esterase (protein)\n* **Unique Ligands:** EDO, ACT, CA\n* **Publication:** [Levisson, M., et al. (2012). Functional and Structural Characterization of a Thermostable Acetyl Esterase from Thermotoga maritima. Proteins, 80, 1545-1559.](https://doi.org/10.1002/prot.24041)\n\nDescription: This structure of the thermostable acetyl esterase TM0077 from Thermotoga maritima reveals a classic \u03b1/\u03b2-hydrolase fold assembled into a doughnut-shaped hexamer. The structure confirms the catalytic serine mechanism and reclassifies the enzyme\u2019s activity from acetyl xylan esterase to a short-chain acetyl esterase, with relevance to industrial biocatalysis at extreme temperatures.", "# Elora Therapeutics Investor Pitch Deck\nAuthors: Swartzendruber, Paul (Co-Founder & CEO); Deller, Marc C. (Co-Founder & CSO)\nSource: Elora Therapeutics (investor pitch deck), 2025\n\nDeveloping a first-in-class systemic enzyme therapeutic to address cardiometabolic disease by reducing microplastic burden in the human body.\n\n## Slide 1 \u2014 Title\n\nElora Therapeutics is developing a first-in-class systemic enzyme therapeutic to address cardiometabolic disease by reducing the microplastic burden in the human body.\n\n## Slide 2 \u2014 Problem\n\nMicroplastics like polyethylene terephthalate (PET) have been detected in human blood and tissues and are associated with cardiometabolic disease risk affecting approximately 1 in 3 U.S. adults, including MASLD, yet no approved therapy reduces systemic microplastic burden.\n\nMASLD stands for metabolic dysfunction\u2013associated steatotic liver disease, a common liver manifestation within cardiometabolic disease.\n\nKey sources: Leslie 2022; Ragusa 2021; Hirode 2020; Wang 2023; Nihart et al., Nature Medicine (2025). These are separate datasets, not co-measured in the same individuals.\n\n## Slide 3 \u2014 Why Now\n\nThe unique insight is that microplastics in humans are now a druggable clinical target. Three developments have converged to make this the right moment:\n\nWe can measure it: microplastics are now directly measurable in human blood and tissues. We can intervene: PET-degrading enzymes now function in human serum in vitro. And the clinic cares: there is a growing clinical focus on microplastics as an environmental driver of cardiometabolic disease.\n\nPositioned against existing and adjacent approaches, Elora's therapy is a potential first-in-class treatment that both addresses microplastic accumulation and addresses cardiometabolic disease. By comparison, the FDA-approved MASLD drug Rezdiffra addresses cardiometabolic disease but not microplastic accumulation; GLP-1 therapies likewise address cardiometabolic disease but not microplastic accumulation; the Winnow probiotic acts only via gut absorption and excretion with unknown effect on accumulation and no cardiometabolic benefit; and apheresis addresses blood only, not tissues, again with unknown accumulation effect and no cardiometabolic benefit.\n\nKey sources: Leslie et al., 2022; Lu et al., 2022; FDA Expedited Programs (guidance); Lopez-Lorente et al., 2024.\n\n## Slide 4 \u2014 Our First-in-Class Solution\n\nElora is developing the first systemic enzyme therapy to reduce microplastic burden in the human body. It is a first-in-class engineered enzyme therapeutic combining PETase (polyethylene terephthalate hydrolase) and MHETase (monohydroxyethyl terephthalate hydrolase) to reduce systemic PET microplastic burden and the downstream cardiometabolic disease risk signals, with staged expansion to additional high-prevalence polymers over time.\n\nThe enzymes target PET microplastics in blood and tissues. The enzyme composition combines PETase and MHETase, which break polyethylene terephthalate (PET) down to terephthalate (TPA) and ethylene glycol (EG), products that are removed via natural excretory pathways.\n\nKey sources: Lu 2022; Tournier 2020; Lopez-Lorente 2024.\n\n## Slide 5 \u2014 Market Opportunity\n\nThe initial U.S. beachhead is MASLD, with expansion to global markets and additional cardiometabolic indications.\n\nThe total addressable market (TAM) is $67.5B, based on PET-weighted MASLD patients (9M patients \u00d7 $7,500/yr). The serviceable addressable market (SAM) is $37.5B, based on PET-weighted, clinically managed MASLD patients (5M patients \u00d7 $7,500/yr). The serviceable obtainable market (SOM) is $4.9B, based on trial-eligible PET-weighted MASLD patients (650K patients \u00d7 $7,500/yr).\n\nKey sources: Hirode & Wong, JAMA 2020 (U.S. cardiometabolic disease prevalence); ICER, 2025 (value-based and net pricing benchmarks for metabolic biologics); IQVIA Institute (U.S. drug pricing, net realization, and launch economics).\n\n## Slide 6 \u2014 Traction\n\nElora has begun early de-risking of its proof-of-concept package.\n\nThere has been early clinician and key opinion leader engagement, with strong interest contingent on safety and biomarker data. Third-party de-risking evidence exists: engineered PETase has been shown to remain active in human serum in vitro with no reduction in cell viability. A provisional patent has been filed (#63/775,252) covering enzyme compositions and therapeutic use. And a University of Texas technology validation agreement is in place (work not yet started) for planned in vitro testing of enzyme kinetics, stability, and cytotoxicity in simulated human serum.\n\nThe third-party evidence comes from a KTH Royal Institute of Technology Sweden in vitro study showing engineered PETase activity in human serum, with HEK293 viability unaffected (Lopez-Lorente et al., Faraday Discussions, 2024).\n\n## Slide 7 \u2014 IP Strategy\n\nElora is pursuing layered protection across enzymes, formulation, and therapeutic use.\n\nA provisional has been filed covering enzyme compositions and therapeutic use for systemic PET microplastic degradation, with follow-on polymer expansion. The next step is protecting Elora-designed variants and proprietary formulations to improve stability and human compatibility. The strategy is to protect lead selection plus delivery and dosing concepts. The resulting moat is built on enzyme engineering know-how, formulation, manufacturing, and therapeutic use claims.\n\n## Slide 8 \u2014 Development Timeline Through Phase 1\n\nMilestones are sequenced to de-risk the program through Phase 1 and enable partnering, spanning from December 2025 through May 2031.\n\nThe program moves through in vitro and in vivo validation; regulatory and clinical planning (pre-IND); CRO development plus early CMC/analytics; lead optimization plus in vivo work; CMC tox-supply and dose-range-finding (DRF); GLP tox/TK plus biodistribution and safety pharmacology; CMC scale-up to GMP; IND authoring and submission; and Phase 1 clinical execution.\n\nFunding is staged across the phases: $275K at the current phase/validation (pre-seed), $5.55M for pre-clinical (seed), and $27.4M for Phase 1 clinical (Series A).\n\n## Slide 9 \u2014 Two Paths to Revenue", "# Publications of Marc C. Deller\n\nMarc C. Deller (ORCID 0000-0001-8070-6502; Scopus Author ID 21741715300; Google Scholar user o02oSJIAAAAJ) has authored or co-authored more than 60 scientific publications. His ORCID record lists 51 indexed works from 1999 to 2017, and Google Scholar adds later Incyte-era papers and patents (2022\u20132025), covering cytokine structural biology (Oxford), structural genomics at the Joint Center for Structural Genomics (JCSG, Scripps), HIV-1 and hepatitis C virus envelope/antibody structures (Scripps, with Ian Wilson), crystallography methods and reviews (with Bernhard Rupp), polysaccharide monooxygenases (Stanford), and structure-based drug discovery at Incyte (CDK2, FGFR2/3, KRAS G12D). Keywords on his ORCID profile: Structural Biology, Protein Crystallography, Crystallization, Protein Expression and Purification.\n\nFull list: https://scholar.google.com/citations?user=o02oSJIAAAAJ and https://marcdeller.com/publications/\n\n## Indexed works (newest first)", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\nThis system is also modulated by the SOCS family of proteins, which act as intracellular inhibitors of several cytokine signal transduction pathways. Their expression is induced by cytokine activation of the JAK/STAT pathway and they act as a negative feedback loop by subsequently inhibiting the JAK/STAT pathway either by direct interaction with activated JAKs or with the receptors. These interactions are modulated, in part, by a cytoplasmic element termed the SOCS box which is thought to target activated cell signalling proteins to the protein degradation pathway (Zhang et al., 1999).\n\nIn a similar fashion, box 2 is important for association with JAKs and also mediates other biological effects. Box 1 is conserved in all members of the haemopoietin receptor superfamily, whereas box 2 has so far only been observed in G-CSFR, IL-2R\u03b2, EpoR, GM-CSFR, IL-3R and the gp130 related receptors. The role of the box 3 motif is not clear; however, all members containing the motif such as gp130, LIFR and OSMR\u03b2 activate the STAT3 transcription factor. It has been suggested that JAK1 is the major JAK involved in gp130 signaling and this possibly functions in phosphorylating the box 3 motif of the receptor, so allowing subsequent recruitment and activation of STAT3.\n\n*Figure 1-5.* Diagrammatic representation of the JAK-STAT signaling pathway utilised by the gp130 family of cytokines as exemplified by OSM. Type I (gp130-LIFR-OSM) and Type II (OSMR\u03b2-LIFR-OSM) signaling assemblies are shown on the cell surface and the associated JAK-STAT and MAP kinase pathways. Adapted from (Ihle et al., 1994).\n\n## 1.5 gp130 Family of Cytokines\n\nAs discussed previously LIF, IL-6, IL-11, CNTF, OSM and CT-1 are structurally and functionally related sharing the same signal transducer and protein fold. However, there is no strict functional compensation between the gp130 subfamily of haemopoietin cytokines and each member has unique biological actions (Piquet Pellorce et al., 1994), thus necessitating detailed structure-function studies of each.\n\n### 1.5.1 Leukemia Inhibitory Factor and its Receptor\n\n#### 1.5.1.1 Biological Action of Leukemia Inhibitory Factor\n\nLIF is a polyfunctional cytokine that plays a role in many systems of the organism, eliciting a diversity of biological effects dependent on the cell type (Hilton, 1992). In vitro, LIF acts as a growth regulator in adipocytes, glial cells (Patterson, 1992), motor neurons (Li et al., 1995), hepatocytes, myoblasts, embryonic stem cells and certain myeloid leukaemic cell-lines (Moreau et al., 1988; Hilton et al., 1988). The precise role of LIF in vivo is less defined, with the over expression of LIF resulting in rapid weight loss, behavioural disorders, ectopic calcification and bone abnormalities. In contrast, deletion of the LIF gene in mice results in little change towards known cellular targets of LIF. However, female reproduction is inhibited due to the prevention of embryonic implantation (Stewart et al., 1992). This suggests that LIF is one of the few cytokines absolutely essential for life (Heath, 1992). Targeted disruption of the LIFR gene has also been shown to cause placental, skeletal, neural and metabolic defects resulting in perinatal death (Li et al., 1995).\n\n#### 1.5.1.2 Structure-Function Relationship of Leukemia Inhibitory Factor and its Receptor\n\nAs discussed above the biological effects of LIF are mediated by interaction with a specific LIFR subunit. Formation of a high-affinity (Kd ~ 10\u207b\u2079 M) signal-transducing complex requires association of the LIF-LIFR complex with the signal transducing component gp130. The hGH-hGHR complex has provided a model for possible modes of gp130-LIF-LIFR engagement (see section 1.4.2.1.1, Figure 1-6). In common with the other haemopoietin cytokines it has been shown that several discrete sites of the molecule are involved in receptor engagement; sites 1 and 2 in similar locations to that observed for hGH. However, a third site, termed site 3, has been identified which allows for extra interactions not observed in the hGH-hGHR system. Site 1 and site 3 are the contact points for LIFR and site 2 is the contact point for gp130 (see Figure 1-6).\n\nTests of biological activity in human Ba/F3 and murine Da-1a assays have indicated that murine and human LIF, although exhibiting closely related three-dimensional structures, show distinct differences in biological specificity (Hudson et al., 1996; Layton et al., 1994a; Layton et al., 1994b; Owczarek et al., 1993). For example, both murine and human LIF interact with human LIFR, whereas murine LIF binds only murine LIFR. Data obtained from human-murine LIF chimeras suggests that the species-specific regions lie in the C-terminal region between residues 99 and 180. In particular, residues Ser 107, Val 155, Lys 58, His 112 and Ser 113 have been shown to be important for human LIFR binding. This region corresponds to site 1 in the growth hormone model and lies on helix D and the AB loop (see Figure 1-6). Using alanine scanning mutagenesis critical residues in site 1 have been localised to Val 175 with contributions from Asp 57, Lys 58, Lys 170 and Ala 174 (Hudson et al., 1996).", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\nThe ability of the five antibodies shown here to inhibit cell proliferation was tested. Inhibition of cell proliferation was assessed using engineered BaF3 cells expressing MPL/mutCALR Type 1. Cells were plated at 5,000 cells per well in RPMI 1640 + 2% FBS, antibodies were added and incubated for 72 hours, followed by assessment of cell viability using the CellTiter-Glo Luminescent Cell Viability Assay (Promega, Madison, WI) and Top Count (Perkin Elmer, Boston, MA) or Pherastar (BMG Labtech, Ortenberg, Germany) for luminescence quantification. As shown in FIG. 21B, among the antibodies shown to bind mutant CALR, only clone 4 is able to substantially inhibit cell proliferation. The data shown in FIGs. 21A and 21B were compiled from different experiments.\n\nThe ability of other antibodies of the present disclosure to inhibit cell proliferation as compared to AB1, AB2, AB3 and AB4 was tested in the same manner as described above. The identified anti-mutCALR antibody clones listed below inhibited mutCALR-induced oncogenic cell proliferation in a dose-dependent manner, while AB1, AB2, AB3 and AB4 showed no functional activity. The IC50 values are shown in Table 16.\n\n*Table 16.* IC50 values for anti-mutCALR antibody mediated inhibition of cell proliferation.\n\n| Clone No. | IC50 (\u03bcg/mL) |\n|---|---|\n| 53 | 9.5 |\n| 39 | 21 |\n| 35 | 2.18 |\n| 36 | 0.98 |\n| 41 | 3.98 |\n| 4 | 0.10 |\n| 51 | 10 |\n| 37 | 0.201 |\n| AB1 | >60 |\n| AB2 | >60 |\n| AB3 | >60 |\n| AB4 | >60 |\n\n## Other Embodiments\n\nIt is to be understood that while the invention has been described in conjunction with the detailed description thereof, the foregoing description is intended to illustrate and not limit the scope of the invention, which is defined by the scope of the appended claims. Other aspects, advantages, and modifications are within the scope of the following claims.\n\n## Claims\n\nThe application as published contains 101 claims. The claim set is highly repetitive: claims 2-31 enumerate several hundred individual CDR, VH/VL and heavy-chain/light-chain sequence combinations by SEQ ID NO. The independent claims, the structurally and functionally significant dependent claims (in particular the epitope claims that follow directly from the crystallographic work) and a representative sample of the sequence-enumeration claims are reproduced below; the remaining sequence-permutation dependent claims are not reproduced.\n\n**1.** An antibody that binds to human mutant calreticulin (CALR), wherein the antibody comprises a heavy chain variable region (VH) comprising a VH CDR1, a VH CDR2, and a VH CDR3, and a light chain variable region (VL) comprising a VL CDR1, a VL CDR2, and a VL CDR3, defined by the same degenerate consensus sequences (SEQ ID NOs:97-99 and the X1-X72 position definitions) as set out in the Summary above.\n\n**2.** The antibody of claim 1, wherein [the six CDRs correspond to one of an enumerated list of clone-specific SEQ ID NO combinations].\n\n**3.** The antibody of claim 1, wherein the VH CDR1, the VH CDR2, and the VH CDR3 each correspond to the VH CDRs set forth in Tables 1-2 for a single VH clone, and wherein the VL CDR1, the VL CDR2, and the VL CDR3 each correspond to the VL CDRs set forth in Tables 1-2 for a single VL clone.\n\n**5.** The antibody of any one of claims 1-4, wherein: the VH is at least 80% identical to the amino acid sequence of any one of SEQ ID NOs:165-181 and 183-208; and the VL is at least 80% identical to the amino acid sequence of any one of SEQ ID NOs:264-304 and 306-318.\n\n**6.** The antibody of claim 1, wherein the VH comprises the amino acid sequence of any one of SEQ ID NOs:165-181 and 183-208; and the VL comprises the amino acid sequence of any one of SEQ ID NOs:264-304 and 306-318.\n\n**7.** The antibody of claim 1, wherein [one of an enumerated list of specific VH/VL SEQ ID NO pairs, e.g. the VH comprises the amino acid sequence of SEQ ID NO:165 and the VL comprises the amino acid sequence of SEQ ID NO:264].\n\n**8.-18.** The antibody of claim 1, wherein the VH comprises the amino acid sequence of a specified SEQ ID NO (e.g. SEQ ID NO:191 or 195) and the VL comprises the amino acid sequence of a specified SEQ ID NO (e.g. SEQ ID NO:318). *(Individual pairings not reproduced.)*\n\n**19.** The antibody of claim 1, wherein the antibody comprises: a heavy chain comprising the amino acid sequence of any one of SEQ ID NOs:119-135 and 137-164; and a light chain comprising the amino acid sequence of any one of SEQ ID NOs:209-249 and 251-263.\n\n**20.** The antibody of claim 1, wherein the antibody comprises a heavy chain and a light chain each corresponding to the heavy chain and the light chain set forth in Tables 4-5 for a single clone selected from the group consisting of clones 1-53 and 55-215.\n\n**21.-31.** The antibody of claim 19, wherein the heavy chain comprises a specified SEQ ID NO and the light chain comprises a specified SEQ ID NO \u2014 for example, claim 21: heavy chain SEQ ID NO:119 and light chain SEQ ID NO:213; claim 24: heavy chain SEQ ID NO:142 and light chain SEQ ID NO:213; claim 31: heavy chain SEQ ID NO:150 and light chain SEQ ID NO:263. *(Remaining pairings not reproduced.)*", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\nThe long chain cytokine family is often further sub-divided into three further groups with GH, Epo, leptin and GCSF forming the classic four helix group and the IFN members adopting a topology based on a five helix bundle. LIF, CNTF, CT-1, IL-6, IL-11 and OSM form another group through structural and sequence homology and the shared use of the gp130 receptor (see Figure 1-2). Long chain cytokines differ from the short chain variants in having between 50 and 100 extra amino acids, so producing a more elongated cylindrical shape as compared to the ellipsoidal shape adopted by the short chain forms.\n\n*Figure 1-2.* Gallery of group 1 \u03b1-helical cytokines arranged in their sub-groupings of long and short chain (Nicola, 1994). Further sub-divisions according to dimerisation state, number of main helices and receptor usage are shown in red lettering. All structures solved by X-ray crystallography (unless otherwise stated). Figures produced using BOBSCRIPT (Esnouf, 1997) and RASTER3D (Merritt and Murphy, 1994). The polypeptides are coloured from blue at the N-terminus through, red at the C-terminus. See abbreviations for descriptions.\n\n#### 1.4.1.2 Group 2 Cytokines (\u03b2)\n\nMembers of this family include TNF, IL-1, FGF, PDGF, NGF, VEGF, CD40 ligand, and TGF\u03b2 (see Figure 1-2). All members of this family contain long chain \u03b2-sheet structures and can be further subdivided according to \u03b2-jelly roll, \u03b2-trefoil or cystine knot motifs.\n\n##### 1.4.1.2.1 \u03b2-jelly roll motif\n\nThe TNF sub-family of group 2 cytokines includes TNF\u03b1, TNF\u03b2, OX-40 antigen, CD40, CD27 and Fas ligands (see Table 1-2). The TNF family are cell surface associated ligands and form symmetric homotrimers based on the \u03b2-jelly roll motif as observed in many viral coat proteins (e.g. satellite tobacco necrosis virus) and cell surface proteins. The \u03b2-jelly roll motif is best considered as a \u03b2-sandwich structure formed by two anti-parallel \u03b2-pleated sheets as observed in the crystal structures of the CD40 ligand (Karpusas et al., 1995) and TNF (Baeyens et al., 1999; Eck and Sprang, 1989; Jones et al., 1989; Reed et al., 1997) (see Figure 1-1).\n\n##### 1.4.1.2.2 \u03b2-trefoil motif\n\nThe IL-1 (Le and Vilcek, 1987) and FGF (Faham et al., 1996) sub-family of group 2 cytokines contain a \u03b2-trefoil motif (see Table 1-2), as first identified in Soya bean trypsin inhibitor (Blow et al., 1974). This motif consists of three sets of 4-\u03b2-strands each of which adopts a trefoil shape (Y-shape). The 12 strands of \u03b2-sheet form six hairpins, with three of the hairpins forming a \u03b2-barrel structure and the other three forming a triangular structure. Members of this family that have been determined structurally include basic FGF (Faham et al., 1996), acidic FGF (Blaber et al., 1996) and IL-1\u03b2 (Priestle et al., 1988) (see Figure 1-1).\n\n##### 1.4.1.2.3 Cystine knot motif\n\nThe cystine knot sub-family of group 2 cytokines is particularly divergent and includes TGF\u03b2, BMP, PDGF, VEGF, NGF and BDNF (see Table 1-2). Only the overall fold and position of the three disulphide bonds unites the group. The cystine knot motif consists of an elongated asymmetric monomer produced by an interlocking set of three disulphide bonds and at least four strands of twisted anti-parallel \u03b2-sheet. Members of the group delineated structurally include BMP-2 (Scheufler et al., 1999), NGF (McDonald et al., 1991), VEGF (Muller et al., 1997), PDGF (Oefner et al., 1992), TGF-\u03b22 (Daopin et al., 1992) and chorionic gonadotrophin (Wu et al., 1994) (see Figure 1-1).\n\n#### 1.4.1.3 Group 3 Cytokines (\u03b1/\u03b2)\n\nMembers of this family include EGF, TGF-\u03b1, \u03b2-cellulin, insulin, IGF-1 and the chemokine sub-family of cytokines including IL-8, GRO, MIP-2, RANTES and MIP-1\u03b1 (see Table 1-2). All members of this family contain short chain \u03b1/\u03b2 structures. Structurally the group 3 cytokines can be further sub-divided according to those containing many disulphides, \u03b2-meander motifs or a conserved pattern of cysteines (CXXXC, CXC, CC or C chemokines).\n\n##### 1.4.1.3.1 \u03b2-meander motif\n\nThe \u03b2-meander sub-family of group 3 cytokines includes EGF, TGF\u03b1, \u03b2-cellulin, amphiregulin and vaccinia virus growth factor (see Table 1-2). These cytokines are small proteins of ~50 residues in length, that contain characteristic EGF-like domains (Carpenter and Cohen, 1990). EGF-like domains consist of at least two anti-parallel \u03b2-strands connected to the intervening loops by three disulphide bonds. The EGF domain is released from the transmembrane precursor by proteolysis and has been shown to bind calcium. EGF-like domains are found in several non-cytokine and growth factor proteins, for example the adhesion molecule L-selectin and the blood coagulation protein factor X, which has also been shown to bind calcium. Members of this group solved structurally include TGF-\u03b1 (Moy et al., 1993) and EGF (Campbell et al., 1989) (see Figure 1-1).\n\n##### 1.4.1.3.2 Chemokines", "# HGMP01 and FAST-PETase: Decoding Two Paths to Plastic Degradation \u2013 Engineered versus Natural PETase Enzymes\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-11-05 on marcdeller.com (categories: blog, microplastics).*\n\n* [Machine learning-aided engineering of hydrolases for PET depolymerization](https://pubmed.ncbi.nlm.nih.gov/35478237/)\n* [Identification of a PET hydrolytic enzyme from the human gut microbiome unveils potential plastic biodegradation in human digestive tract](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf)\n\n* **Engineered precision versus natural adaptation:** FAST-PETase contains [five strategic mutations](https://pubs.acs.org/doi/10.1021/jacs.3c04427) that enhance thermostability to 63\u00b0C, while HGMP01 evolved naturally in human gut bacteria for optimal function at body temperature.\u200b\n* **pH specialization reflects environmental niches:** FAST-PETase operates efficiently at [moderately alkaline pH 9.0](https://www.nature.com/articles/s42003-023-04413-0), whereas HGMP01 thrives under [highly alkaline conditions](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf) of pH 10.0 to 11.0, demonstrating adaptation to distinct biochemical environments.\n* **Structural conservation with functional divergence:** Both enzymes share the [canonical \u03b1/\u03b2 hydrolase fold](https://www.nature.com/articles/s41467-018-02881-1) and Ser-His-Asp catalytic triad, yet exhibit only 20 to 41% sequence identity, indicating convergent evolution toward plastic degradation.\u200b\n* **Mutation strategy determines catalytic enhancement:** The critical [N233K mutation in FAST-PETase](https://pmc.ncbi.nlm.nih.gov/articles/PMC10851425/) lowers activation energy by 4.4 kcal/mol through altered backbone folding, demonstrating how distal mutations impact active site chemistry.\n* **Widespread expression in human microbiome:** [Metatranscriptomic evidence](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf) shows HGMP01-like genes are actively transcribed across diverse gut bacterial species, suggesting an underappreciated role in processing ingested microplastics.\u200b\n\n1. **Insights into the Enhancement of the Poly(ethylene terephthalate) Degradation Activity of FAST-PETase:** [Garc\u00eda-Meseguer, R., et al., Journal of the American Chemical Society, 2023](https://pubs.acs.org/doi/10.1021/jacs.3c04427)\n2. **Identification of a PET hydrolytic enzyme from the human gut microbiome unveils potential plastic biodegradation in human digestive tract:** [Zhang, G., et al., International Journal of Biological Macromolecules, 2024](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf)\n3. **Machine learning-aided engineering of hydrolases for PET depolymerization:** [Lu, H., et al., Nature, 2022](https://pubmed.ncbi.nlm.nih.gov/35478237/)\n4. **Current advances in the structural biology and molecular engineering of PETase:** [Liu, F., et al., Frontiers in Bioengineering and Biotechnology, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC10546322/)\n5. **Structural insight into molecular mechanism of poly(ethylene terephthalate) degradation:** [Han, X., et al., Nature Communications, 2018](https://www.nature.com/articles/s41467-018-02881-1)\n\n**How laboratory evolution and gut microbiome discoveries converge on solving the plastic crisis**\n\nThe global accumulation of polyethylene terephthalate waste has driven researchers to explore enzymatic solutions from two distinct sources: rational protein engineering of known bacterial enzymes and discovery mining within complex microbial ecosystems. This comparison examines [FAST-PETase](https://www.nature.com/articles/s41586-022-04599-z), an engineered variant of the bacterial enzyme from *Ideonella sakaiensis*, and [HGMP01](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf), a newly discovered enzyme from the human gastrointestinal microbiome, revealing how different evolutionary pressures shape protein structure and function for plastic degradation.\u200b\n\nFAST-PETase represents the culmination of [machine learning-guided protein design](https://pmc.ncbi.nlm.nih.gov/articles/PMC10851425/), incorporating five strategic mutations that enhance both stability and catalytic efficiency at elevated temperatures. The enzyme achieves a [38-fold increase in activity](https://pmc.ncbi.nlm.nih.gov/articles/PMC10546322/) compared to its parent enzyme at 50\u00b0C, with modifications that strengthen hydrogen bonding networks and optimize electrostatic interactions near the active site. In contrast, HGMP01 emerged from [bioinformatic screening](https://www.sciencedirect.com/science/article/abs/pii/S0141813024085428) of the Unified Human Gastrointestinal Genome collection, representing an independently evolved solution adapted to the unique chemical environment of the human digestive tract.\u200b\n\nThe structural foundation of both enzymes follows the [canonical \u03b1/\u03b2 hydrolase fold](https://www.nature.com/articles/s41467-018-02881-1), characterized by a central twisted \u03b2-sheet surrounded by \u03b1-helices. FAST-PETase contains [290 amino acids](https://pubs.acs.org/doi/10.1021/acsomega.4c05142) organized around a conserved [Ser160-His237-Asp206 catalytic triad](https://pmc.ncbi.nlm.nih.gov/articles/PMC5785972/), while HGMP01 comprises 275 amino acids with a molecular weight of [30.18 kDa](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf). The shorter sequence length in HGMP01 reflects its distinct evolutionary trajectory within gut bacteria of the [Acutalibacteraceae family](http://eeg.qd.sdu.edu.cn/PDF/2024-zhangguoqiang-IJBM.pdf), organisms known for fermenting complex plant carbohydrates.\u200b", "# marcdeller.com: Structural Biology\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n### [Revolutionizing Protein Structure Determination: How MICA Combines Cryo-EM and AlphaFold3 for Unprecedented Accuracy](https://marcdeller.com/revolutionizing-protein-structure-determination-how-mica-combines-cryo-em-and-alphafold3-for-unprecedented-accuracy/)\n\nSB\n\n### [Engineering Protein Stability: Unlocking the Secrets of Successful Crystallization](https://marcdeller.com/engineering-protein-stability-unlocking-the-secrets-of-successful-crystallization/)\n\nSB\n\n### [Shedding Light on the Shadows: The Untapped Potential of the Dark Kinome](https://marcdeller.com/shedding-light-on-the-shadows-the-untapped-potential-of-the-dark-kinome/)\n\nSB\n\n### [Revolutionizing Protein Structure Analysis: Cryo-EM without Purification Accelerates the Future of Structural Biology](https://marcdeller.com/revolutionizing-protein-structure-analysis-cryo-em-without-purification-accelerates-the-future-of-structural-biology/)\n\nSB\n\n### [Multi-Step LLM Prompt: From Protein Sequence to Validated Protein Structure](https://marcdeller.com/multi-step-llm-prompt-from-protein-sequence-to-validated-protein-structure/)\n\nSB\n\n### [Supercharging Discovery: How LLM Prompts Are Transforming Structural Biology and Cheminformatics](https://marcdeller.com/supercharging-discovery-how-llm-prompts-are-transforming-structural-biology-and-cheminformatics/)", "# ChemSage: Vibe-Coding a Chemically Aware Drug-Discovery LLM on Local Hardware\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-30 on marcdeller.com (categories: ai, artificial-intelligence, blog, project, sb).*\n\nChemSage: a retrieval layer feeds corpus context to a QLoRA-tuned model whose tool calls are executed and validated by a sandboxed RDKit and PyMOL shim before any answer is returned.\n\n* [ChemSage GitHub Repo](https://github.com/bellcheddar/ChemSage)\n* [ChemSage Vibe Page](https://marcdeller.com/chemsagevibe)\n* [ChemSage Models on Hugging Face](https://huggingface.co/Dellboy)\n* [ChemSage Model Evaluation](https://marcdeller.com/ChemSage/chemsage_evaluation.html)\n* [Latest Model: chem_sage_32b_v5](https://huggingface.co/Dellboy/chem_sage_32b_v5)\n* [Project inspiration from TikTok](https://www.tiktok.com/@ai.christianson/video/7538150476467014943?lang=en)\n\n* **Hybrid by design:**\u00a0Retrieval for facts, fine-tuned weights supply behavior, and tools including PyMOL, PDBTOOLS, Biopython, ChEMBL, PubChem, and RDKit ensure chemical truth.\n* **78 Data Classes:** Covering structural biology, MedChem, computational chemistry, protein families, and drug-target knowledge.\n* **13 Metric evaluation script:** Grounded truths beat guessing, teaching model to run tools lifted Round 5 to 95% overall against Round 1\u2019s 72%, with tool executability climbing from 0% to 95%.\n* **Unified memory removes the wall:** Apple Silicon fine-tunes a 32B model on old M1 MacBook Pro, no discrete GPU required.\n* **Data quality beats compute:** Validated training examples drove validation loss from 0.389 down to 0.041.\n\n1. **Exploring LLMs with MLX and the Neural Accelerators in the M5 GPU:** [Apple Machine Learning Research, 2025](https://machinelearning.apple.com/research/exploring-llms-mlx-m5)\n2. **MLX: An Array Framework for Apple Silicon:** [Apple ml-explore, 2026](https://mlx-framework.org/)\n3. **MLX-LM: LoRA, Fuse and Serve on Apple Silicon:** [ml-explore, GitHub, 2026](https://github.com/ml-explore/mlx-lm)\n4. **QLoRA: Efficient Finetuning of Quantized LLMs:** [Dettmers et al., NeurIPS, 2023](https://arxiv.org/abs/2305.14314)\n5. **LoRA: Low-Rank Adaptation of Large Language Models:** [Hu et al., arXiv, 2021](https://arxiv.org/abs/2106.09685)\n6. **Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks:** [Lewis et al., NeurIPS, 2020](https://arxiv.org/abs/2005.11401)\n7. **ChemSage: A Hybrid, Chemically Aware LLM for Drug Discovery (Project Plan):** [Deller, GitHub, 2026](https://github.com/bellcheddar/ChemSage/blob/main/PROJECT_PLAN.md)\n8. **ChemSage: Multi-Round Comparative Evaluation:** [Deller, marcdeller.com, 2026](https://marcdeller.com/ChemSage/chemsage_evaluation.html)\n\n**How a structural biologist fine-tuned, grounded, and evaluated a 32-billion-parameter chemistry assistant entirely on Apple Silicon**\n\n###### Why build it?\n\nHosted frontier models are convenient until you meet their ceilings: token caps, document-upload limits, and the simple fact that they cannot see your private SAR data. The starting question for [ChemSage](https://github.com/bellcheddar/ChemSage) was a practical one: if a model has to run somewhere, why not run it on the Mac already sitting on the desk? The spark came from [@ai.christianson](https://www.tiktok.com/@ai.christianson), whose [TikTok video](https://www.tiktok.com/@ai.christianson/video/7538150476467014943?lang=en) of MLX-LM fine-tuning on Apple hardware showed the path.\n\nThe rationale is about correctness, not just cost. Off-the-shelf models invent chemistry: a confidently wrong logP, a malformed SMILES, a property quoted from memory rather than computed. ChemSage answers that by separating three concerns that are usually conflated. Volatile knowledge (your assays, dossiers, papers) lives in retrieval; stable behavior (when and how to reach for a tool, house output style) lives in the fine-tuned weights; and deterministic chemical facts are never guessed at all. They are computed live by RDKit and PyMOL, the same tools a medicinal chemist already trusts.\n\n> **\u201cChemical correctness is delegated to deterministic tools that already exist and are trusted.\u201d** \u2013\u00a0[ChemSage Project Plan, 2026](https://github.com/bellcheddar/ChemSage/blob/main/PROJECT_PLAN.md)\n\nA note on provenance, since it matters here: I am a structural biologist, not a professional software engineer. ChemSage is a vibe-coding exercise, written to find out how far a subject-matter expert can get with their own data, moderate desktop hardware, and very little money. The chemistry and biology are still under validation, so treat it as a proof of concept rather than a finished product. The full build log, with the warts left in, sits on the [making-of page](https://marcdeller.com/chemsage).\n\n###### How it works\n\nTwo failure modes kill most custom-chemistry-LLM projects, and ChemSage is designed around both. The first is the model inventing numbers: the fix is not to fine-tune harder but to make the model emit tool calls and reason over the returned values. The second is treating facts and behavior as the same thing. Keeping facts in retrieval and behavior in the weights means you re-tune only when you hit a behavioral gap that retrieval cannot fill. New papers and SAR tables just get ingested.\n\nThe request loop is easy to read even if you do not write code. Retrieval supplies the facts, the tuned model supplies the behavior and any tool code, and a sandboxed shim supplies the deterministic chemical truth before anything reaches you.\n\n```\n\n\t\t\t\t\ton user_query:\n    context  = rag.retrieve(query)              # facts from your own corpus\n    response = model.generate(query, context)   # tuned behaviour; may emit tool code\n    for block in response.tool_calls:           # RDKit / PyMOL / docking\n        result   = sandbox.run(block)           # deterministic, validated\n        response = model.ground(response, result)\n    return response                             # grounded, numerically correct\n\n```", "# Smiles2Grid \u2013 Turn a JSON of SMILES into a paginated, property-annotated PDF grid with RDKit\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-14 on marcdeller.com (categories: vibe-code).*\n\n> **Turn a JSON of SMILES into a paginated, property-annotated PDF grid with RDKit.**\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/smiles2grid](https://github.com/bellcheddar/smiles2grid) |\n\n---\n\nRender SMILES strings from a JSON file into a paginated PDF grid using RDKit.\n\nWhy it matters: after a generative or screening run you often have hundreds of SMILES and no fast way to actually look at them together, compare scaffolds, or pull a labelled subset for a slide. smiles_grid renders them into clean 5\u00d75 PDF pages with structure labels tied to the original input order, computes a full property panel (MW, cLogP, HBD/HBA, TPSA, and more) into a companion CSV, and can highlight near-neighbours of a query structure in red. It is useful for medicinal chemists and computational scientists who need a quick, shareable visual of a compound set with the numbers attached: ideal for triage, SAR discussions, and decision-making meetings.\n\n## What it does\n\n* Extracts SMILES strings from a nested JSON file.\n* Preserves the original input order for labeling and selection.\n* Renders molecules into a boxed 5\u00d75 grid, 25 compounds per page.\n* Supports optional similarity search against a query SMILES.\n* Supports range-based selection such as `all`, `1-25`, `50-75`, or `1,5,10-20`.\n* Writes a CSV summary for every rendered record.\n* Highlights closely related structures within a page in red.\n\n## Example input\n\n```\npython smiles_grid.py scratch_Baricitinib_sim.out.20260530_0219.json\n```\n\nThe script prompts by default. You can also run it non-interactively with `--no-prompt` and selection arguments.\n\n## Features\n\n### PDF output\n\n* Single PDF output.\n* 5\u00d75 bordered panels per page.\n* Structure labels based on original JSON order.\n* Clean footer text with predicted properties.\n* Structure drawing scaled to fit each panel.\n\n### CSV output\n\nA CSV file is written alongside the PDF and includes:\n\n* output index.\n* original index.\n* source index.\n* label.\n* SMILES and canonical SMILES.\n* formula.\n* MW, cLogP, HBD, HBA, TPSA.\n* rotatable bonds, rings, heavy atoms, fraction sp3, aromatic rings, hetero atoms.\n* input JSON path.\n* PDF filename.\n* selection mode.\n\n## Usage\n\n### Interactive mode\n\n```\npython smiles_grid.py input.json\n```\n\nAt runtime the script prompts for:\n\n1. An optional query SMILES for similarity search.\n2. A similarity cutoff percentage.\n3. If no query is provided, a range selection such as `all`, `1-25`, or `50-75`.\n\n### Similarity search\n\n```\npython smiles_grid.py input.json --query-smiles \"CCO\" --similarity 75\n```\n\n### Range selection\n\n```\npython smiles_grid.py input.json --range 1-25\npython smiles_grid.py input.json --range all\npython smiles_grid.py input.json --range 1,5,10-20\n```\n\n### Non-interactive mode\n\n```\npython smiles_grid.py input.json --no-prompt --range 1-25\n```\n\n## Output files\n\nBy default the script writes:\n\n* `input_grid.pdf`\n* `input_summary.csv`\n\nYou can override the PDF filename with `--output`.\n\n## Requirements\n\n* Python 3.9+\n* RDKit\n* ReportLab\n* Pillow\n* cairosvg, if your RDKit build does not include Cairo drawing support\n\n## Installation\n\n### Conda\n\n```\nconda install -c conda-forge rdkit reportlab pillow cairosvg\n```\n\n### Pip-style environments\n\nIf you already have RDKit available in your environment:\n\n```\npip install reportlab pillow cairosvg\n```\n\n## Command-line options\n\n* `--query-smiles`: similarity search query.\n* `--similarity`: similarity cutoff as `0-100` or `0-1`.\n* `--range`: selection by range or comma-separated ranges.\n* `--no-prompt`: disable interactive prompting.\n* `--max-pages`: render only the first N pages for debugging.\n* `--label-prefix`: change the label prefix, default `J`.\n* `--output`: choose the PDF filename.\n\n## Notes\n\n* Numbering is tied to the original JSON order, not to whether a SMILES can be rendered.\n* Failed renderables are skipped visually but do not shift labels.\n* The script keeps the input order for selection and CSV reporting.\n* Similarity highlighting is page-local and uses red bond highlighting.\n\n## Example workflow\n\n1. Place the JSON file in your working directory.\n2. Run the script.\n3. Enter a query SMILES or press Enter.\n4. Enter a similarity cutoff or press Enter.\n5. Or press Enter again and provide a range such as `all` or `1-25`.\n6. Open the generated PDF and CSV.\n\n## License\n\nAdd your preferred license here.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/smiles2grid](https://github.com/bellcheddar/smiles2grid) |", "# Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3\nAuthors: Shvartsbart, A.; Roach, J.J.; Witten, M.R.; Koblish, H.; Harris, J.J.; Covington, M.; Hess, R.; Lin, L.; Frascella, M.; Truong, L.; Leffet, L.; Conlen, P.; Beshad, E.; Klabe, R.; Katiyar, K.; Kaldon, L.; Young-Sciame, R.; He, X.; Petusky, S.; Chen, K.-J.; Horsey, A.; Lei, H.-T.; Epling, L.B.; Deller, M.C.; Vechorkin, O.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2022\n\nDespite the well-distributed and overall polarity of 27, this compound still inhibited the activity of CYP3A4 (IC50 = 9 \u03bcM). At this point, we speculated that the CYP inhibition in this series might be due to specific heme binding rather than general lipophilicity. A possible culprit was proposed to be the carbonyl of the N-methyl amide or the corresponding (presumed) primary amide metabolite. Increasing steric hindrance of the amide and/or blocking N-demethylation would be expected to disrupt this binding and mitigate CYP inhibition. To test these hypotheses, an N-ethyl derivative (28) and the N-CD3 congener (29) of 27 were prepared (Table 7). Pleasingly, both compounds had a clean CYP profile (CYP3A4 IC50 > 25 \u03bcM). Compound 29 maintained the excellent potency and selectivity of 27, while 28 was considerably less potent and selective (FGFR3 IC50 = 2.6 nM, R1/R3 = 21\u00d7). The in vitro intrinsic clearance of 29 was also dramatically reduced compared to 27 (0.5 L/h/kg vs 1.0 L/h/kg, respectively), indicating N-demethylation as a likely site of metabolism. We developed a whole blood (WB) assay using FGFR3 RT-112/84 cells, in which 29 displayed excellent potency (WB IC50 = 177 nM). On the basis of this profile, 29 was advanced into rat PK studies (Table 8). Clearance in the i.v. arm was low (HBF = 35%), with a moderate half-life (t1/2 = 1.7 h). In the p.o. dose, 29 demonstrated good exposure (AUC = 5108 nM\u00b7h) and high oral bioavailability (F% = 82).\n\n*Table 7.* Potency and In Vitro ADME Profile of Lead Compound 29.\n\n| Compound | FGFR3 IC50 (nM) | FGFR1 IC50 (nM) (R1/R3) | FGFR3 V555L IC50 (nM) | FGFR3 V555M IC50 (nM) | BA/F3 IC50 (nM) | WB IC50 (nM) | Caco-2 (Papp \u00d7 10\u207b\u2076 cm/s) | h-Cl (L/h/kg) | CYP3A4 IC50 (\u03bcM) | Solubility (\u03bcg/mL, FaSSIF/SGF) |\n|---|---|---|---|---|---|---|---|---|---|---|\n| 27 | 0.4 | 17 (43\u00d7) | 1.8 | 5.2 | 7.2 | \u2014 | 1.5 | 1.0 | 9 | 17/>400 |\n| 28 | 2.6 | 54 (21\u00d7) | \u2014 | \u2014 | 22 | \u2014 | 0.6 | 1.0 | >25 | \u2014 |\n| 29 | 0.5 | 21 (42\u00d7) | 2.7 | 6.1 | 5.5 | 177 | 2.0 | 0.5 | >25 | 18/679 |\n\n*Table 8.* Pharmacokinetics of 29 in Rat.\n\n| Parameter | Value |\n|---|---|\n| Rat i.v. dose (mg/kg) | 1.0 |\n| Rat i.v. HBF% | 35 |\n| Rat i.v. Vdss (L/kg) | 1.6 |\n| Rat i.v. t1/2 (h) | 1.7 |\n| Rat p.o. dose (mg/kg) | 3.0 |\n| Rat p.o. Cmax (nM) | 2303 |\n| Rat p.o. AUC (nM\u00b7h) | 5108 |\n| Rat p.o. F% | 82 |\n\nCompound 29 was tested for activity toward several FGFR3 gatekeeper mutants. Despite the fact that nearly all compounds tested in this series were equipotent for the GK mutants, 29 lost 5-fold potency for the V555L mutant (IC50 = 2.7 nM) and 10-fold potency for the V555M mutant (IC50 = 6.1 nM). Screening an additional diverse set of compounds from these series revealed that pyridine nitrogen was responsible for this loss in activity. To understand this result, we obtained an X-ray co-crystal structure of an analogous compound (30) bound to FGFR2 (Figure 3). This isoform was more amenable to crystallization than FGFR3, and 30 was confirmed to have identical biochemical potency in both. Contrary to our initial docking model, the pyridyl ring is flipped, with the amide extending toward the ATP sugar pocket and the pyridine nitrogen oriented toward the gatekeeper residue. It appears that the larger, more lipophilic groups of the FGFR3 V555L and V555M mutants clash with the polar nitrogen of the pyridyl ring, thereby resulting in the observed loss in potency. Notably, the oxygen of the THF ring forms a hydrogen bonding interaction with the primary amide N\u2212H of Asn571 as designed (for Asn562 in FGFR3).\n\n*Figure 3.* X-ray co-crystal structure of 30 in FGFR2 (PDB code: 8E1X).\n\nWe further profiled 29 against the other FGFR isoforms (FGFR2/4) in addition to the approved inhibitors 1\u22123 (Table 9). Compound 29 was found to be equally potent for FGFR2 (IC50 = 1.0 nM) and highly selective over FGFR4 (IC50 = 145 nM). As expected, 1\u22123 maintained equal potency for FGFR1 with >50-fold loss in activity for the FGFR3 V555L and V555M GK mutants. Collectively, this data demonstrates that our novel compounds are isoform-selective inhibitors of FGFR2/3 with significantly improved potency toward the common GK mutations, in contrast to first-generation pan-FGFR inhibitors. Furthermore, 29 also maintained >100-fold selectivity against all 36 structurally representative kinases from an in-house kinase panel (Table S1).\n\n*Table 9.* Comparison of Novel FGFR2/3 Inhibitor 29 and Approved FGFR Inhibitors 1\u22123.\n\n| Compound | FGFR1 IC50 (nM) | FGFR2 IC50 (nM) | FGFR3 IC50 (nM) | FGFR4 IC50 (nM) | FGFR3 V555L IC50 (nM) | FGFR3 V555M IC50 (nM) |\n|---|---|---|---|---|---|---|\n| 29 | 21 | 1.0 | 0.5 | 145 | 2.7 | 6.1 |\n| erdafitinib (1) | 0.2 | 0.2 | 0.6 | 1.9 | 21 | 68 |\n| pemigatinib (2) | 0.9 | 0.4 | 1.7 | 27 | 373 | 201 |\n| infigratinib (3) | 2.1 | 2.7 | \u2014 | 115 | 3036 | 352 |\n\nPrior to the discovery of 29, compound 27 was dosed in an in vivo pharmacodynamic study to measure both the inhibition of FGFR pathway as well as the induction of phosphorous in mice. First, we determined the percent inhibition of pERK in the UM-UC-14 human bladder cell carcinoma model. Tumor-bearing mice were dosed by oral gavage with 27 or pemigatinib, and tumors were analyzed 2 h post-dose administration. The data show a dose-dependent target inhibition of pERK by 27, with maximal inhibition observed at the high dose (100 mg/kg) and comparable to a high dose of 10 mg/kg pemigatinib (Figure 4). Importantly, 27 did not significantly increase phosphorus levels in serum at any of the doses tested when compared to vehicle control. In contrast, a high single dose of 10 mg/kg pemigatinib was sufficient to significantly increase phosphorus levels by 36%, further supporting the importance of a selective FGFR3 inhibitor.\n\n*Figure 4.* Percentages of pERK inhibition and phosphorus increase induced by 27 compared to vehicle control.\n\n## Conclusions", "# BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-07-10 on marcdeller.com (categories: vibe-code).*\n\n> **BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, binding analysis, and run control, orchestrated end to end from a single spec file.**\n>\n> [**Read the BoltzMaker Blog page for more info**](https://marcdeller.com/boltzmaker-vibe-coding-a-campaign-scale-boltz-2-pipeline/)\n>\n> [**Read an example 5HT2 GPCR receptor campaign analysis**](https://marcdeller.com/5HT2_findings/findings.preview.html)\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [boltzmaker.mdeller.com](https://boltzmaker.mdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/BoltzMaker](https://github.com/bellcheddar/BoltzMaker) |\n\n---\n\nA single script that manages a Boltz-2 batch campaign end to end: parse a `boltz_input.md` spec, generate the per-target Boltz YAML files, preflight the environment and inputs, run `boltz predict` with a live progress bar (with resume support), and analyze the results into a CSV, an XLSX workbook, and an HTML dashboard \u2014 optionally enriched with real protein-ligand interaction analysis via [cif2plip](https://github.com/bellcheddar/cif2plip). The non-GPU stages \u2014 the `new` wizard, `generate`, `preflight`, and full `analyze` (including PLIP interaction detection and compare-sse) \u2014 are also available with no local install at **[boltzmaker.mdeller.com](https://boltzmaker.mdeller.com)**; only the GPU `run` step still needs your own hardware (see **Web deployment** below).\n\nWhy it matters: hand-running a Boltz-2 campaign means writing dozens of near-identical YAML files by hand, remembering the right CLI flags for the hardware you\u2019re on, and manually digging through prediction JSONs afterwards, all repetitive, error-prone steps that don\u2019t need a human. BoltzMaker turns a single annotated spec into the full pipeline: generated inputs, environment/input validation (including ligand-chemistry sanity checks \u2014 undefined stereocentres, ambiguous protonation states, stray salts \u2014 so bad input chemistry is caught before hours of compute, not silently mispredicted), a monitored run with Mac-safe memory defaults, and a ready-to-read report. It is useful for: structural biologists and drug-discovery scientists running Boltz-2 structure/affinity panels (single targets, covalent-linkage studies, or multi-chain SAR/selectivity campaigns) who want a repeatable, resumable, well-documented pipeline instead of a pile of hand-edited scripts.\n\nDon\u2019t want to hand-write the spec at all? `BoltzMaker.py new` interviews you in plain English \u2014 proteins, partners, ligands, and the constraint sentences \u2014 and writes a valid `boltz_input.md` for you. If you\u2019ve already got a reference co-crystal or homology structure for a protein, it can even suggest pocket-contact residues automatically, remapped onto your target\u2019s own numbering via sequence alignment, instead of you reading them off a structure viewer by hand.\n\nBoltzMaker grew out of five smaller, single-purpose tools written earlier: [generate_yaml](https://github.com/bellcheddar/generate_yaml) for building the input YAMLs, [simple-zsh-script-to-run-boltz2](https://github.com/bellcheddar/simple-zsh-script-to-run-boltz2) for driving the actual `boltz predict` runs, [analyze-boltz2-results](https://github.com/bellcheddar/analyze-boltz2-results) for the post-run analysis, [cif2plip](https://github.com/bellcheddar/cif2plip) for protein-ligand interaction profiling, and [smiles2grid](https://github.com/bellcheddar/smiles2grid) for rendering a ligand set into a boxed grid of 2D structures. BoltzMaker consolidates all five into one spec-driven pipeline sharing a single campaign format, so the same input file drives every stage instead of juggling separate scripts and re-typing target names between them.\n\nAnd yes, the name is a nod to [Boltmaker](https://www.timothytaylor.co.uk/beer/boltmaker), Timothy Taylor\u2019s Champion Beer of Britain and one of the author\u2019s favourites.\n\nSee <CHANGELOG.md> for what\u2019s changed recently, and <PROJECT_PLAN.md> for the roadmap behind the To-Do list below.\n\n## \ud83d\udce5 Installation\n\nTwo ways to install, depending on whether the computer running BoltzMaker has internet access. Most people want **Path 1**. (Looking for more detail/troubleshooting on either path? See **One-time setup** and `docs/tier_b_offline_install.md` further down.)\n\n### Path 1: This computer has internet access\n\n**1.** Open the **Terminal** app (Applications -> Utilities -> Terminal, or search for \u201cTerminal\u201d with Spotlight), then make a folder for BoltzMaker and go into it:\n\n```\nmkdir -p ~/Desktop/BoltzMaker\ncd ~/Desktop/BoltzMaker\n```\n\n**2.** Download BoltzMaker into that folder:\n\n```\ngit clone https://github.com/bellcheddar/BoltzMaker.git .\n```\n\n(the trailing `.` tells it to put the files directly into the folder you just made, rather than inside another new folder). If this is the first time you\u2019ve used `git`, macOS may pop up a dialog asking to install \u201cCommand Line Tools\u201d \u2014 click Install and wait for it to finish, then run the command again.\n\n**3.** Run the installer:\n\n```\n./install.sh\n```\n\nWait for it to finish. It downloads a few GB of software the first time, so it can take a while depending on your connection.\n\n**4.** Check it worked:\n\n```\npixi run preflight examples/t4_lysozyme/boltz_input.md\n```\n\nIf you see a table of green PASS results, you\u2019re ready to go \u2014 see **Commands** below for what to run next.\n\n### Path 2: This computer has no internet access\n\nUse this for a lab machine, server, or any computer that\u2019s offline or behind a firewall. You\u2019ll need a second computer that *does* have internet access to prepare a single installer file first.\n\n**1.** On the computer **with** internet access, follow **Path 1** above, then build the offline installer file:\n\n```\npixi global install pixi-pack\npixi-pack --platform osx-arm64 --ignore-pypi-non-wheel --create-executable -o boltzmaker-installer.sh\n```", "# The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\nAuthors: Span, E.A.; Suess, D.L.M.; Deller, M.C.; Britt, R.D.; Marletta, M.A.\nSource: ACS Chemical Biology, 2017\n\nX-band EPR spectra of wild type MtPMO3* and the variants described herein show strong Cu(II) signals and indicate efficient loading of Cu into the active site (Figure 6). There is also a lack of signal for free Cu(II), showing the proteins as purified are free of excess Cu. All spectra display superhyperfine coupling to multiple 14N nuclei, consistent with Cu(II) binding to the histidine brace motif. The 14N superhyperfine coupling is more resolved in some spectra (e.g., H161Q), although this likely reflects small differences in the degree of spectral broadening between the samples (e.g., unresolved superhyperfine coupling from other nuclei) rather than significant structural differences. The wild type MtPMO3* EPR spectrum displays a pseudoaxial g-tensor that is consistent with the geometry about Cu observed in the crystal structure: an elongated octahedral or square pyramidal geometry that gives rise to an unpaired electron in the d(x\u00b2\u2013y\u00b2) orbital. The fourth equatorial site is presumably occupied by a water ligand; while this is not observed in the crystal structure, likely due to structure resolution or photoreduction of the Cu(II) site to Cu(I) during data collection, the equatorial water ligand has been observed in Cu(II)-AA10 structures.[32] Similarly, the Y169\u2013OH ligand is nearly 3 \u00c5 away from the Cu center in the crystal structure and may interact more strongly in the Cu(II) oxidation state.\n\nThe EPR spectra of the H161A and H161E variants are essentially identical to that of wild type MtPMO3*, suggesting that any changes in H-bonding in these variants have little effect on the Cu coordination environment. The EPR spectrum of the Q167A variant exhibits a single g-tensor (g = [2.246 2.055 2.049]) but one that differs from that of the wild type enzyme (g = [2.260 2.079 2.032]). In particular, the g-tensor is somewhat more axial and has a lower g1 value (Figure 6). We interpret that these spectral changes result from the modification of the H-bonding environment of Y169. As discussed above, Q167 serves as an H-bond acceptor to Y169. With the elimination of this interaction in the Q167A variant, the donor strength of the Y169\u2013OH ligand to Cu is expected to decrease. As a result, the equatorial ligands are expected to bind Cu more strongly in order to conserve the charge at the Cu site. This raises the d(x\u00b2\u2013y\u00b2) orbital relative to that of the wild type enzyme (Figure 7). Because the g1 value is inversely proportional to the energy difference between the antibonding d(x\u00b2\u2013y\u00b2) orbital and the nonbonding d(xy) orbital,[33] the larger energy gap between the d(x\u00b2\u2013y\u00b2) and d(xy) orbitals is expected to result in a modest decrease in the g1 value for the Q167A variant, as observed experimentally. The Q167E sample serves to test this hypothesis: in the Q167E variant, the H-bond acceptor to the Y169\u2013OH group is maintained, and the EPR spectrum is restored to that of wild type enzyme, as expected. By enhancing the donor strength of the tyrosyl O ligand in the wild type enzyme, the H-bond between Q167 and Y169 helps to stabilize positive charge buildup on the Cu ion and could aid in the stabilization of high-valent oxyl intermediates.\n\nThe EPR spectra of the H161Q and T74A variants display a mixture of signals as can be most clearly seen in the broad, asymmetrical shapes of the feature at ~305 mT. This indicates that some subpopulation of the Cu sites in these samples are unaffected by the substitution while others are affected by changes in H-bonding. However, the presence of multiple EPR signals in these samples precludes more detailed, molecular-level analysis.\n\n*Figure 6.* X-band EPR spectra (black) of wild type MtPMO3* and six H-bonding network variants, with simulations (red) for WT and the Q167A variant. Spectra were recorded at 40 K, 9.4 GHz, 0.2 mW power, and 0.5 mT modulation amplitude and have been scaled for signal strength. The signal at ~335 mT in the WT spectra is a feature from the cavity. Simulation parameters for WT MtPMO3*: g = [2.260 2.079 2.032], A(63/65Cu) = [504 30 25] MHz, 3x A(14N) = [40 40 40] MHz, g-strain = [0.007 0.007 0.007]. Simulation parameters for the Q167A variant: g = [2.246 2.055 2.049], A(63/65Cu) = [558 56 20] MHz, 3x A(14N) = [37.6 37.6 37.6] MHz, g-strain = [0.007 0.007 0.007].\n\n*Figure 7.* (A) H-bonding to Y169 in wild type (WT) MtPMO3* and the Q167E variant. (B) Geometric and electronic effects of removing a H-bond acceptor to Y169, as in the Q167A variant: elongation of the Cu\u2013Tyr distance and contraction of the Cu\u2013equatorial ligand distances results in an increase in the energy between d(x\u00b2\u2013y\u00b2) and d(xy).\n\n## Conclusions\n\nAltogether, these data show that the conserved active-site H-bonding network, comprising the secondary sphere histidine and glutamine residues, facilitates O2 activation in AA9 PMOs. The combined activity data indicate that H161 and Q167 function in O2 turnover in MtPMO3*, while an additional role in proton transfer to an oxy-bound intermediate may be attributed to H161. These data show that the generation of products is diminished in H161 variants, and instead superoxide is released, therefore suggesting that a Cu(II)\u2013superoxo intermediate is either not stable in these variants or not the active species in the PMO mechanism. The spectroscopic data show that Q167 enhances the ligand donor strength of the active-site tyrosine, tuning the electronic environment of the Cu center and favoring stabilization of positive charge buildup on the Cu ion. Together, these newly defined roles for H161 and Q167 support the formation of a Cu\u2013oxyl active intermediate. The function of T74 is less clear, but the data demonstrate this residue is less critical for PMO activity. This study provides the first detailed mechanistic insights into the AA9 family of PMOs that play an important role in fungal cellulose degradation and have potential for widespread use in the production of renewable fuels and chemicals.\n\n## Methods\n\n### Strains and Materials", "# GitHub README: GOBSMACKED\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/GOBSMACKED, 2026\n\n**The docking box is sized from the ligand, not from the residues around it.** Erlotinib's own extent plus padding is 32 x 21 x 23 A; the shell of its 51 contacting residues gives 48 x 42 x 42, a search volume five times larger that samples the true site less densely for no benefit. The centre still comes from the model, because the crystal is in its own coordinate frame and only the extent transfers.\n\n**Reference selection prefers the same ligand over a sharper crystal.** EGFR and erlotinib make the case: 1M17 holds erlotinib itself at 2.6 \u00c5, while the best sub-2.5 \u00c5 entry holds gefitinib at Tanimoto 0.41. Judging an erlotinib pose against a gefitinib crystal because the crystal is 0.9 \u00c5 sharper would measure the wrong thing.\n\n**Visibility** is chosen here and defaults to public. A private run is issued a 32-character owner key, shown once, stored only as its sha256, and carried inside the bundle so uploading results needs no typing. Private runs are **absent** from the Runs table without the key rather than greyed out: a greyed row would leak that the run exists, and how many there are.\n\n## \u2697\ufe0f Run\n\nSix stages, each idempotent and resumable from a `.done` marker.\n\n| Stage | Tool | Notes |\n|---|---|---|\n| `fold` | ESMFold | Skipped when the bundle carries a model, which is the usual case. Chunk size scales with sequence length; pocket residues below pLDDT 70 raise a warning that reaches the scorecard |\n| `prep` | PDBFixer, RDKit | Missing atoms, hydrogens at the campaign pH, waters and heteroatoms removed. Terminal missing residues are deliberately not built: they are absent from the construct, not from the model |\n| `dock` | PandaDock | `hybrid` (search plus SE(3) GNN rescoring), `flex` (induced fit) or `dock` (empirical only). Falls back from `hybrid` to `dock` when the GNN checkpoint cannot be fetched, and says so |\n| `md` | OpenMM, OpenFF | Amber14 plus OpenFF Sage, TIP3P with 0.15 M NaCl and 10 \u00c5 padding, restraints released over the equilibration, 2 fs with hydrogen mass repartitioning. The DCD holds the solute only |\n| `affinity` | Boltz-2 | Optional, on by default. The docked pose and frames sampled from the last fifth of the trajectory, each scored by Boltz-2's affinity head with the structure module bypassed. The MSA is computed once per target and cached, so only the first pose queries the server |\n| `summarise` | MDTraj | Per-frame ligand and backbone RMSD, per-residue RMSF, pocket volume by voxel counting, a residue-by-frame contact matrix, then packs the archive |\n\n```bash\n./run.sh --list          # what would run, and what is already done\n./run.sh --stage dock    # rerun from dock onward\n```\n\nTarget on one consumer GPU: under 30 minutes for a 300-residue domain with the default 1 ns\nproduction, and roughly an hour with the affinity stage on, which is the slowest of the six once\nthe MD is short. The runner prints a per-stage plan with wall-clock estimates and a finishing time\nbefore it starts, then a progress bar and a running ETA for each stage as it goes.\n\n**Warnings are triaged rather than dumped.** Every third-party warning the run emits was read\nonce, and each is either fixed at the source, suppressed by an exact-match filter with the reason\nrecorded next to it, or left visible because it is telling you something. What reaches the terminal\nis what a reader should act on.\n\n## \ud83d\udcca Analyze\n\nUpload the archive and the whole pipeline runs inside the request: ingest, superpose, interactions, modes, dynamics, scorecard. Seconds, not minutes.\n\n**Superposition is on pocket C\u03b1 atoms, never on the whole chain.** A model can be excellent at the binding site and 6 \u00c5 out at a disordered terminus; superposing whole chains spreads that error into the pocket and inflates the ligand RMSD, which is the one number this app exists to report honestly. The whole-chain TM-score is reported alongside, as context.\n\n**Residue numbering is never assumed to match.** The reference chain and the model are aligned by sequence first and every measurement walks that mapping. 1M17 numbers EGFR from the mature protein, 24 lower than UniProt: comparing raw numbers would make every contact look lost and the interaction overlap read as zero.\n\n### \ud83c\udfaf The scorecard\n\n| Metric | A | B | C | D | F | Weight |\n|---|---|---|---|---|---|---|\n| Ligand RMSD, best of pose 1 and MD-final | \u2264 1.0 \u00c5 | \u2264 2.0 | \u2264 3.0 | \u2264 4.0 | > 4.0 | 30 |\n| PLIP interaction Jaccard, best of pose 1 and MD-final | \u2265 0.75 | \u2265 0.55 | \u2265 0.40 | \u2265 0.25 | < 0.25 | 20 |\n| Pocket C\u03b1 RMSD, MD-final | \u2264 0.8 | \u2264 1.2 | \u2264 1.8 | \u2264 2.5 | > 2.5 | 15 |\n| \u03c71 agreement, MD-final (within 40\u00b0) | \u2265 0.85 | \u2265 0.70 | \u2265 0.55 | \u2265 0.40 | < 0.40 | 10 |\n| MD stability: ligand drift, last window minus first | \u2264 0.5 \u00c5 | \u2264 1.0 | \u2264 1.5 | \u2264 2.5 | > 2.5 | 10 |\n| Rescue: pocket C\u03b1 RMSD before MD minus after | \u2265 +0.5 \u00c5 | \u2265 +0.2 | \u2265 0 | \u2265 \u22120.3 | < \u22120.3 | 10 |\n| Pose validity: clashes, bond lengths, chirality, inside the box | pass | | | | fail | 5 |\n\nThe composite **GOBSMACK score** is the weighted mean of those grades. Three rules keep it honest:\n\n- **A metric that could not be measured drops out** and the remaining weights are renormalised, which the card states rather than hiding in the arithmetic.\n- **A validity failure caps the composite at 40.** A pose with a 1.8 \u00c5 clash is not a B whatever else it scored.\n- **A run with no reference gets no composite at all.** There is nothing to verify it against, and scoring it on the two metrics that survive would be a grade for something nobody checked.\n\nEvery gauge carries one plain sentence saying what the number means and what to do about it, because \"\u03c71 agreement 0.42\" is not actionable and \"most pocket side chains are in the wrong rotamer, which is what an apo-like predicted pocket looks like: try flex docking plus a longer equilibration\" is.\n\n### \ud83e\uddec The overlay\n\nModel in grey, MD-final in phosphor, the crystal in amber, all superposed on the pocket, with the ten most displaced side chains and their \u03c71 angles listed underneath.", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\nBoltzMaker applies them itself before every run, so you rarely need the commands above, and\n`preflight` shows a `boltz_patches` row because a `pip install -U boltz` silently reverts them.\n**What we fixed in Boltz** below describes each one, including the device-correct full-precision\nguard around the Kabsch alignment and the NaN diagnostic that names the offending tensor instead\nof blaming matrix conditioning.\n\n## \ud83e\uddea Examples\n\nFour entirely public-domain campaigns ship in `examples/`; run any with\n`python3 BoltzMaker.py all examples/<name>/boltz_input.md`. All four, plus two more that have\nno bundled folder, are **public runs with a live interactive report** on the hosted site. The\nindividual URLs are unguessable, so they are also all listed under the\n[**Runs** tab](https://boltzmaker.mdeller.com/runs). The four bundled campaigns were verified\nend to end (`generate` -> `preflight` -> real `boltz predict` -> `analyze`, including cif2plip\ninteraction analysis) on an Apple M1 Max, 64GB.\n\n| Campaign | Targets | Demonstrates | Verified run | Links |\n|---|---|---|---|---|\n| `t4_lysozyme` | 1 | The minimal shape: one protein (T4 lysozyme L99A, UniProt P00720), one ligand (benzene), no partners, no pocket contacts. Smallest/fastest smoke test | 3m 16s. Confidence 0.98, pIC50 8.9, 7 hydrophobic contacts matching the known L99A cavity residues | [explore](https://boltzmaker.mdeller.com/runs/K-hKV9wI5OYO-t3srj3C5g/explore) \u00b7 [dashboard](https://bellcheddar.github.io/BoltzMaker/examples/t4_lysozyme/boltz_dashboard.html) \u00b7 [input](https://github.com/bellcheddar/BoltzMaker/blob/main/examples/t4_lysozyme/boltz_input.md) |\n| `egfr_covalent` | 1 | Covalent-linkage modelling: EGFR kinase domain (UniProt P00533) + a generic covalent fragment, linked via `bond_constraints` at Cys797 | 5m 31s. Confidence 0.92, covalent Cys797 SG-to-fragment bond confirmed at 1.75 Angstrom | [explore](https://boltzmaker.mdeller.com/runs/5a_wDBYYltnO3xxJeh73CA/explore) \u00b7 [dashboard](https://bellcheddar.github.io/BoltzMaker/examples/egfr_covalent/boltz_dashboard.html) \u00b7 [input](https://github.com/bellcheddar/BoltzMaker/blob/main/examples/egfr_covalent/boltz_input.md) |\n| `adrb2_gs` (folder `adrb2_gs_panel`) | 2 | `compare-sse`, and why a co-folded partner should match each ligand's real biology rather than being crossed with everything -- see below | 1h 28m 36s (`ADRB2_ISO1`, agonist+Gs) + 25m 6s (`AR2NG_PRO1`, antagonist alone). Confidence 0.79 (`ADRB2_ISO1`) / 0.83 (`AR2NG_PRO1`) | [explore](https://boltzmaker.mdeller.com/runs/Nutmagij4prtzjYgi_v0LA/explore) \u00b7 [dashboard](https://bellcheddar.github.io/BoltzMaker/examples/adrb2_gs_panel/boltz_dashboard.html) \u00b7 [input](https://github.com/bellcheddar/BoltzMaker/blob/main/examples/adrb2_gs_panel/boltz_input.md) |\n| `5ht2_gq` (folder `5ht2_gq_panel`) | 15 | `Ligands: none`, a large size-heterogeneous campaign in one manifest, Apple Silicon MPS support for large multi-chain complexes, and `compare-sse` against a *predicted* apo reference -- see below | 9 small targets (apo + receptor-alone) ~4-5m each; 6 large receptor+Gq-heterotrimer complexes ~43-48m each. All 15 completed (12 ligand-bound + 3 apo); confidence 0.66-0.81, iPTM up to 0.99 for the ligand-bound complexes | [explore](https://boltzmaker.mdeller.com/runs/TaRE0NcV9BOpuguKRujt5g/explore) \u00b7 [dashboard](https://bellcheddar.github.io/BoltzMaker/examples/5ht2_gq_panel/boltz_dashboard.html) \u00b7 [input](https://github.com/bellcheddar/BoltzMaker/blob/main/examples/5ht2_gq_panel/boltz_input.md) |\n| `5HT2A_GQ` | 2 | Public run, no bundled example folder | -- | [explore](https://boltzmaker.mdeller.com/runs/d96nzlk1sj5JzfkNCumscw/explore) |\n| `GLP1R_GIPR_pocket_matrix` | 20 | Public run, no bundled folder: two receptors x three ligands x three pocket conditions, plus two apo controls. The campaign behind **How tight should a pocket be?** and **Does the restraint just manufacture the pose?** below, so the live report is where you can see those measurements in situ | -- | [explore](https://boltzmaker.mdeller.com/runs/niSiE5MfGFyddSjGLmeYBA/explore) |\n\n**`adrb2_gs`: a co-folded partner has to match the ligand's real biology.** Beta-2 adrenergic\nreceptor (UniProt P07550), agonist vs antagonist, written as two separate `Protein:` blocks\nsharing one sequence rather than one family crossed with both ligands. The agonist target\nco-folds a Gs alpha partner (UniProt P63092); the antagonist target does not, because in\nreality Gs only forms a stable complex with the active, agonist-bound receptor. Co-folding it\nwith the antagonist too made Boltz predict a near-identical active-like fold for both, 0.38\nAngstrom apart; splitting them out gets a real conformational difference, 1.28 Angstrom\napart, with the TM6 shift roughly doubled for the agonist.\n\n**`5ht2_gq`: a 3x2x2 panel plus apo controls.** Three serotonin receptors (5-HT2A/2B/2C,\nUniProt P28223/P41595/P28335), each with a real agonist/antagonist pair (Psilocin/Risperidone,\nLSD/Balovaptan, Lorcaserin/SB-242084), each predicted both with and without the Gq\nheterotrimer (GNAQ+GNB1+GNG2) co-folded, plus one native ligand-free (`Ligands: none`) apo\ntarget per receptor. Those apo predictions are each receptor's `compare-sse` reference,\nbecause no genuinely apo experimental structure exists for any of the three (checked\nentity-by-entity across all 59 deposited structures). TM6 centroid shift comes out\nconsistently larger for the Gq-bound targets than their no-Gq counterparts across all three\nreceptors -- the expected activation signal; full statistical write-up in\n[findings.md](https://github.com/bellcheddar/BoltzMaker/blob/main/examples/5ht2_gq_panel/findings.md).", "# Vibe Coding MicroPlastic Blaster\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-02-22 on marcdeller.com (categories: blog, microplastics).*\n\nStoryboard Used for Initial Game Design\n\nScreengrab of latest gameplay including water bottle and coffee cup enemies, and co-founder powerups (v115)\n\n* [**Launch MicroPlastic Blaster!**](https://marcdeller.com/game/microplastic_blaster_115.html)\n* [**Source Code on GitHub**](https://github.com/bellcheddar/microplastic-blaster.git)\n\n* **Vibe coding enables rapid scientific outreach:** A fully functional, branded HTML5 game was built in a single day using only LLM prompts.\n* **Gameplay as mechanism-of-action storytelling:** Each game session teaches players about microplastic types and enzyme degradation, transforming abstract therapeutic science into an interactive experience.\n* **Single-file architecture maximizes\u00a0deployment flexibility:** Packaging the entire game into a single .html file with zero external dependencies allows instant embedding on any website.\n* **Iterative vibe coding scales from prototype to production:** Over 115 versions, structured prompting, and disciplined version control transformed a 200-line proof of concept into a 2000-line feature-rich game with leaderboards, co-founder power-ups, and audio.\n\n1. **\u201cThere\u2019s a New Kind of Coding I Call \u2018Vibe Coding'\u201d:** [Karpathy, *X (formerly Twitter)*, 2025](https://x.com/karpathy/status/2019137879310836075)\n2. **Canvas API:** [MDN Contributors, *MDN Web Docs (Mozilla)*, 2025](https://developer.mozilla.org/en-US/docs/Web/API/Canvas_API)\n3. **2D Breakout Game Using Pure JavaScript:** [MDN Contributors, *MDN Web Docs (Mozilla)*, 2024](https://developer.mozilla.org/en-US/docs/Games/Tutorials/2D_Breakout_game_pure_JavaScript)\n4. **VibeGame: Exploring Vibe Coding Games:** [Hugging Face Team, *Hugging Face Blog*, 2025](https://huggingface.co/blog/vibegame)\n\n**From Prompt to Playable: Using\u00a0AI-Assisted\u00a0Coding to Turn Enzyme Science into an Interactive Outreach Tool for Elora Therapeutics**\n\nWhen we needed a way to make enzyme therapy tangible for investors and the public, we turned to vibe coding and built an HTML5 game from scratch using nothing but natural language prompts and a single afternoon.\n\nMicroplastics have been detected in human blood, lungs, and the placenta, and in over a dozen organ systems, with studies finding an average concentration of 4.2 particles per milliliter of blood in healthy adults. At Elora Therapeutics, the mission is to develop the first FDA-targeted enzyme therapy capable of degrading polyethylene terephthalate (PET) microplastics inside the human body. Explaining that science to investors, journalists, and the general public, however, is a different kind of challenge. Slide decks and white papers communicate the data, but they rarely create the visceral, emotional connection that drives someone to champion a cause.\n\nThe question became: how do you make enzymatic degradation of synthetic polymers feel real and immediate?\n\nThe answer was a video game.\n\n[MicroPlastic Blaster](https://marcdeller.com/microplastic-blaster/) puts the player in control of a human cell under siege, firing enzyme projectiles at waves of incoming microplastic particles. The gameplay directly mirrors the therapeutic mechanism of action: FAST-PETase enzymes intercept and degrade PET fragments before they can damage cellular structures. Rather than hiring a game development studio or spending weeks learning a framework, the entire project was built using Vibe Coding, the AI-assisted development paradigm coined by Andrej Karpathy in February 2025, which uses natural language prompts to replace traditional line-by-line programming. The technology stack was deliberately minimal: vanilla JavaScript and HTML5, packed into a single `.html` file with zero external dependencies. This was not an accident. JavaScript game code is the most richly represented genre in LLM training data, which means AI models can generate, debug, and iterate on it with remarkable fluency. A single-file architecture also meant the game could be deployed to any web server, embedded in WordPress as an iframe, or shared as a simple link in a pitch deck, all within minutes of completion:\n\n| Factor | JS + HTM | Phaser.js | React Game | Pygame |\n| --- | --- | --- | --- | --- |\n| Single-file deploy | Yes | CDN needed | Build step | Not browser-native |\n| Iframe embed | Trivial | Possible | Complex | No |\n| LLM code generation | Excellent | Good | Moderate | Not web |\n| Time to deploy | < 1 hour | 2\u20133 hours | 4+ hours | N/A |\n| Mobile support | Touch events | Built-in | Varies | No |\n\n**The development process followed a structured six-hour timeline:**\n\n* **Hour 1 produced a complete working prototype from a single prompt:** \u201cCreate an HTML5 Canvas game where you shoot enzyme projectiles at falling microplastic particles to protect a cell.\u201d\n* **Hour 2 focused on branding:** swapping the color palette to Elora\u2019s teal and cyan biotech identity and embedding scientifically accurate microplastic facts that appear during gameplay and on the game-over screen.\n* **Hour 3 introduced visual polish:** particle explosion effects, screen-shake on cell damage, and synthesized sound effects generated entirely through the Web Audio API with no external audio files.\n* **Hour 4 tackled mobile responsiveness:** ensuring touch events worked correctly in iOS Safari and that the canvas scaled dynamically across any viewport.\n* **Hours 5 and 6, the game was live on the website:**\u00a0undergoing cross-browser QA. Each iteration followed the same vibe coding best practice: describe the desired change in a specific, descriptive prompt, test in the browser immediately, and save a versioned copy before the next edit.", "# Crystal Structure of a Two-Subunit TrkA Octameric Gating Ring Assembly\nAuthors: Deller, M.C.; Johnson, H.A.; Miller, M.D.; Spraggon, G.; Elsliger, M.-A.; Wilson, I.A.; Lesley, S.A.\nSource: PLoS ONE, 2015\n\n**Authors:** Marc C. Deller (corresponding author), Hope A. Johnson, Mitchell D. Miller, Glen Spraggon, Marc-Andr\u00e9 Elsliger, Ian A. Wilson and Scott A. Lesley (corresponding author). **Affiliations (condensed):** The Joint Center for Structural Genomics, and Department of Integrative Structural and Computational Biology, The Scripps Research Institute, La Jolla, California, USA (Deller, Johnson, Elsliger, Wilson, Lesley); The Joint Center for Structural Genomics, and Stanford Synchrotron Radiation Lightsource, SLAC National Accelerator Laboratory, Menlo Park, California, USA (Miller); Protein Sciences Department, Genomics Institute of the Novartis Research Foundation, San Diego, California, USA (Spraggon, Lesley). Current addresses: Department of Biological Science, California State University, Fullerton, California, USA (Johnson); Department of BioSciences, Rice University, Houston, Texas, USA (Miller). Correspondence: slesley@gnf.org (SAL); mdeller@scripps.edu (MCD). **Published:** PLoS ONE, volume 10, issue 3, e0122512, 2015. Received December 2, 2014; accepted February 12, 2015; published March 31, 2015. Academic Editor: Sheng Ye, Zhejiang University, China. Open access article distributed under the terms of the Creative Commons Attribution License.\n\n**Data availability:** All relevant data are in the RCSB database (www.rcsb.org) via accession numbers 3L4B, 2G1U, and 3JXO.\n\n**Funding:** This work was supported by the Protein Structure Initiative of the National Institute of General Medical Sciences, National Institutes of Health (NIH), grant number U54 GM094586. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. The authors have declared that no competing interests exist.\n\n## Abstract\n\nThe TM1088 locus of T. maritima codes for two proteins designated TM1088A and TM1088B, which combine to form the cytosolic portion of a putative Trk K+ transporter. We report the crystal structure of this assembly to a resolution of 3.45 \u00c5. The high resolution crystal structures of the components of the assembly, TM1088A and TM1088B, were also determined independently to 1.50 \u00c5 and 1.55 \u00c5, respectively. The TM1088 proteins are structurally homologous to each other and to other K+ transporter proteins, such as TrkA. These proteins form a cytosolic gating ring assembly that controls the flow of K+ ions across the membrane. TM1088 represents the first structure of a two-subunit Trk assembly. Despite the atypical genetics and chain organization of the TM1088 assembly, it shares significant structural homology and an overall quaternary organization with other single-subunit K+ gating ring assemblies. This structure provides the first structural insights into what may be an evolutionary ancestor of more modern single-subunit K+ gating ring assemblies.\n\n## Introduction\n\nK+ ions are the major monovalent cation in prokaryotic and eukaryotic cells and stringent regulation of its intracellular concentration is essential for a wide variety of cellular processes [1]. Ion transporters that translocate cations, particularly K+, are central to the regulation of cellular pH and osmolarity [2], electrical signal generation [3] and regulation of protein expression and activity [4].\n\nIn contrast to the three K+ transporters found in E. coli (Kdp, Kup and Trk), only one K+ transport system, Trk, has been identified in T. maritima [5]. Trk is a member of the Trk/Ktr/HKT family of transporters, which are specific to non-animal cells [6]. The Trk system is the major constitutive K+ transporter in E. coli and functions as a proton symporter powered by the proton motive force and binding of ATP [7,8]. The E. coli Trk potassium transporter complex consists of a cytosolic regulatory component, TrkA, which associates with the channel forming membrane component, TrkH. In T. maritima the cytosolic regulatory component of Trk comprises of two proteins, TM1088A and TM1088B, that associate with the TrkH component, TM1089, to form the functional transporter [9].\n\nGiven the requirement for tight regulation of K+, several common structural features have evolved, including the use of conserved cytoplasmic regulatory components for control of ion flow through the transmembrane channel. One such feature is the RCK domain (Regulating the Conductance of K+) that is present in a large class of proteins involved in the regulation of many eukaryotic and prokaryotic K+ transporters [10]. More than half of the prokaryotic K+ channels identified so far are associated with RCK domains, making them the most common regulatory domain of prokaryotic K+ channels [11,12]. A sub-family of RCK regulatory proteins, including those found in prokaryotic KefC, KtrA and TrkA channels, have homology to dinucleotide binding proteins. These RCK family members are often collectively referred to as KTN (K+ Transport, Nucleotide-binding) domains due to their ability to bind nucleotides. KTN and RCK domains are closely related and both adopt a Rossmann-like fold (\u03b2-\u03b1-\u03b2-\u03b1-\u03b2) [11,13,14].", "# Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor\nAuthors: Kong, L.; Ju, B.; Chen, Yajing; He, L.; Ren, L.; Liu, J.; Hong, K.; Su, B.; Wang, Z.; Ozorowski, G.; Ji, X.; Hua, Y.; Chen, Yanli; Deller, M.C.; Hao, Y.; Feng, Y.; Garces, F.; Wilson, R.; Dai, K.; O'Dell, S.; McKee, K.; Mascola, J.R.; Ward, A.B.; Wyatt, R.T.; Li, Y.; Wilson, I.A.; Zhu, J.; Shao, Y.\nSource: Immunity, 2016\n\nWe next investigated whether the VRC01-like LCDR1 signature combined with an N terminus truncation could help DRVIA7 better accommodate the inhibitory glycan patch, and as a result, improve its neutralizing activity (Figure 3C). We designed two DRVIA7 variants \u2014 one containing a VRC01 LCDR1 and the other containing a VRC01 LCDR1 and a 2-aa deletion at the light-chain N terminus \u2014 to facilitate the comparison with DRVIA7 and VRC01 (Figure S3D). The N-terminal truncation of the DRVIA7 light chain significantly increased the breadth, which was further enhanced by the introduction of VRC01 LCDR1 (Figure 3C, middle). Inclusion of the VRC01 LCDR1, but not the N-terminal deletion, had little effect on the breadth. Lastly, the use of the mature VRC01 light chain substantially improved potency (Figure 3C, bottom left). Our analysis thus demonstrated that proper light chain maturation was critical in the early development of VRC01-class antibodies to accommodate the inhibitory gp120 glycans. Of note, DRVIA7H-VRC01L neutralized 3 out of 16 VRC01-resistant isolates (Figure S3F), suggesting a potential role of the heavy chain in strain specificity.\n\n*Figure 3. Interaction between DRVIA7 Light Chain and the N276 Glycan Determines the Neutralizing Breadth of the Antibody.* (A) Octet association and dissociation curves for DRVIA7 or DRVIA7H-VRC01L toward glycosylated or endoH deglycosylated YU2 gp120 at 1,000 nM concentration (above). Binding response (nm shift) is plotted as a function of time (s), with KD values indicated below. (B) ELISA binding for DRVIA7, DRVIA7H-VRC01L, and VRC01 against YU2 gp120 (black) and T278G (i.e., no glycan at N276) YU2 gp120 (red), with the optical density (OD) plotted as a function of antibody dilution (\u03bcg/mL). (C) The neutralizing activity of DRVIA7, VRC01H-DRVIA7L, DRVIA7H-VRC01L, and VRC01 against both the global and DRVI panels of pseudoviruses. In addition, the neutralizing activity of two DRVIA7 variants (enclosed by red dotted lines), DRVIA7H-DRVIA7L + Del2aa and DRVIA7H-DRVIA7L + VRC01 LCDR1 + Del2aa, against the global panel is shown. Neutralization potencies (Tables S4A and S4B) are mapped onto dendrogram representations of the virus panels as in Figure 1. See also Figure S3.\n\n### Dynamic Antibody Evolution Revealed by Unbiased Analysis of DRVI01 B Cell Repertoires across Three Time Points\n\nWe applied the unbiased repertoire analysis (He et al., 2014) to study the DRVI01 memory B cell repertoire across three time points (2006, 2008, and 2009). For each time point, 10\u201320 million PBMCs were used to prepare antibody H, \u03ba, and \u03bb chain libraries from the 5\u2032-RACE PCR products of the B cell transcripts. In total, NGS on the Ion Torrent Personal Genome Machine (PGM) generated more than 11.7 million raw reads. After data processing (He et al., 2014; Wu et al., 2011), 1.52 to 3.89 million antibody chains were obtained for the three time points studied (Table S3).\n\nWe first analyzed the germline gene usage (Figure 4A). IgHV1-2, the germline family of DRVIA7 heavy chain, accounted for 3.7%, 3.3%, and 3.9% of the heavy chains at 2006, 2008, and 2009, respectively, whereas IgHV4-34 and IgHV4-39 were predominantly used across the time points. The low frequency of IgHV1-2 suggested that DRVIA7 did not constitute a major lineage within the repertoire, consistent with the weak binding and neutralization of DRVIA7, which might have limited the B cell expansion. IgKV1-5, the germline family of DRVIA7 light chain, accounted for 8.2%, 10.4%, and 8.1% of the \u03ba chains at 2006, 2008, and 2009, respectively. We next determined the distribution of germline gene divergence, or SHM (Figure 4B). Although the heavy chains displayed a modest 1.5% difference in germline divergence between adjacent time points, the light chains showed a rapid change on a scale of 10%\u201320%. In particular, a large population of germline light chains was found in the 2008 repertoire, constituting nearly 20% of the \u03ba chains. Lastly, in terms of CDR3 loop length, we observed a rather dynamic heavy-chain distribution compared to a steady light-chain distribution (Figure 4C). More heavy chains with 9-aa and 15-aa HCDR3 loops were found in the 2008 and 2009 repertoires, respectively. Thus, the production of a large number of germline light chains that rapidly diverged, coupled with notable shifts of HCDR3 length in the context of a stable level of SHM for heavy chains, demonstrated a dynamic B cell repertoire in this donor.\n\n*Figure 4. Unbiased B Cell Repertoire Profiles of Donor DRVI01 at Three Time Points.* The distribution is plotted for (A) germline V gene usage for heavy and light (\u03ba and \u03bb) chains, (B) germline gene divergence, or degree of SHM, and (C) CDR3 loop length (H, heavy chain; K, \u03ba chain; L, \u03bb chain). Color coding denotes the time point analyzed, with 2006 shown in blue, 2008 in green, and 2009 in red. See also Table S3.\n\n### Longitudinal Tracing of the DRVIA7 Lineage over 4 Years of Maturation Revealed Stalled Light Chain Maturation to Accommodate Glycans", "# Creativity: 7+ patents through innovative protein engineering and AI integration\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-29 on marcdeller.com (categories: 6cs).*\n\n## Creativity: 7+ patents through innovative protein engineering and AI integration\n\nCreativity, for me, is about combining rigorous structural thinking with unconventional tooling to build solutions that did not exist before. I treat AI, automation, and protein engineering as parts of a single design space where ideas can move fluidly from *in silico* to *in vitro*. That integrated approach has led to [7+ patents](https://marcdeller.com/patents/) that capture not only individual molecules but also new ways to discover and optimize them. By designing inventive workflows and molecules, we unlock new modalities and routes to patient benefit that keep science ahead of the curve.\n\n* Creativity defines how I merge experimental insight with digital innovation in structural biology.\n* Examples include integrating AI-guided protein design with wet-lab validation to create patentable enzyme and receptor solutions.\n* It matters because creativity bridges disciplines, generating breakthroughs that conventional methods can\u2019t reach.\n* Creativity often looks like reframing\u2014turning an undruggable target into a tractable one by changing the question.\n* Today, it means designing proteins nature never built\u2014using generative models to invent function rather than discover it.\n\nPersonal Philosophy: The 6C's\n\n### [**Curiosity:** 10+ years driving hands-on drug discovery through multidisciplinary approaches](https://marcdeller.com/curiosity-20-years-driving-hands-on-discovery-through-multidisciplinary-approaches/)\n\n[**Curiosity:** 10+ years driving hands-on drug discovery through multidisciplinary approaches](https://marcdeller.com/curiosity-20-years-driving-hands-on-discovery-through-multidisciplinary-approaches/)\n\n### [**Courage**: 400+ challenging structures determined, enabling breakthrough mechanistic insights](https://marcdeller.com/courage-400-structures-determined-enabling-breakthrough-mechanistic-insights/)\n\n[**Courage**: 400+ challenging structures determined, enabling breakthrough mechanistic insights](https://marcdeller.com/courage-400-structures-determined-enabling-breakthrough-mechanistic-insights/)\n\n### [**Communication:** 60+ publications translating structural data into actionable insights](https://marcdeller.com/communication-60-publications-translating-structural-data-into-actionable-insights/)\n\n[**Communication:** 60+ publications translating structural data into actionable insights](https://marcdeller.com/communication-60-publications-translating-structural-data-into-actionable-insights/)\n\n### [**Collaboration:** 2+ state-of-the-art Structural Biology facilities built through global partnerships](https://marcdeller.com/collaboration-2-state-of-the-art-structural-biology-facilities-built-through-global-partnerships/)\n\n[**Collaboration:** 2+ state-of-the-art Structural Biology facilities built through global partnerships](https://marcdeller.com/collaboration-2-state-of-the-art-structural-biology-facilities-built-through-global-partnerships/)\n\n### [**Compassion:** 6+ FDA IND filings supported while mentoring 20+ scientists](https://marcdeller.com/compassion-6-fda-ind-filings-supported-while-mentoring-20-scientists/)\n\n[**Compassion:** 6+ FDA IND filings supported while mentoring 20+ scientists](https://marcdeller.com/compassion-6-fda-ind-filings-supported-while-mentoring-20-scientists/)", "# Tricyclic Compounds as Inhibitors of WRN\nAuthors: Ai, Y.; Barbosa, J.; Ghosh, A.; Deller, M. D.; Glenn, J., Jr.; McQuirter, L.; Qiu, Y.; Shao, L.; Shepard, S.; Taylor, C.; Wang, X.; Yue, E.; Zhang, G.\nSource: United States Patent Application Publication US 2025/0215006 A1 (Incyte Corporation), 2025\n\nHuman colon carcinoma cell lines are obtained from ATCC, which included MSI-H cell lines RKO (CRL-2577) and SW-48 (CCL-231), and MSS cell lines SW-620 (CRL-227) and SW-837 (CRL-235). All cells are maintained in RPMI-1640 supplemented with 10% FBS (Gibco cat #11875119, 16140071) and will be kept at 37 \u00b0C. in a humidified 5% CO2 incubator. WRN inducible knockouts are generated using CRISPR-Cas9 editing and stable cell lines will be kept in media supplemented with antibiotic selection of 200 micrograms per milliliter hygromycin B (Gibco cat #10687010) and 1 microgram per milliliter puromycin (Gibco cat #A1113803).\n\n**Cellular Assay\u2014WRN Inhibitor Anti-Proliferation Assay.** All cells are trypsinized, counted, and resuspended in RPMI-1640 with 10% FBS. 100 microliter cells and growth media are seeded into black 96 well plates (Greiner cat #655090). Seeding densities are 2,500 cells for RKO and SW-620, and 12,500 cells for SW-48 and SW-837. Following incubation to allow for cell adherence to the plate, compounds are added for dose response measurement, starting at the highest concentration, followed with a 3:1 serial dilution scheme. The last dose contains only DMSO and serves as no treatment control. Plates are incubated for three to seven days at 37 \u00b0C. in a humidified 5% CO2 incubator. Following incubation, 100 microliter Cell Titer Glo reagent (Promega cat #G7573) is added and plates are read on the Pherastar FSX (BMG Labtech). Data analysis are carried out using the Genedata software, following % of growth inhibition normalized to control (DMSO treated cells). Fitting parameters can include dose response IC50, Hill slope (n), maximum response (S\u221e), and minimum response (S0).\n\n**Cellular Assay\u2014WRN Inhibitor pKAP1 Induction Assay.** Cells are trypsinized, counted, and resuspended in RPMI-1640 with 10% FBS. Cells re resuspended at 1\u00d710^6 cells per milliliter and ten microliters cells are added to regular volume Culturplate-384 well (Perkin Elmer cat #6007680) via the Multiflo (Biotek). Plates are wrapped to prevent evaporation and are incubated at 37 \u00b0C. in a humidified 5% CO2 incubator. The next day, compounds are added via the CyBio Felix (Analytik Jena) for dose response measurement, starting at the highest concentration, followed with a 3:1 serial dilution scheme. The last dose contains only DMSO and serves as no treatment control. Plates are wrapped to prevent evaporation and are incubated at 37 \u00b0C. in a humidified 5% CO2 incubator. The next day, cells are washed with the Blue washer (Blue Cat Bio) and AlphaLISA lysis buffer is added. Cells are allowed to lyse at room temperature for 30 min followed by the addition of pKAP1 AlphaLISA reagents (Perkin Elmer/Revvity cat #ALSU-PKAP1-A10K). Plates are incubated in the dark until read on the Pherastar FSX (BMG Labtech).\n\nData analysis is carried out using the Genedata software. Activity data are normalized to percent of active control (WRN tool inhibitor compound). Fitting parameters can include dose response EC50, Hill slope (n), maximum response (S\u221e), and minimum response (S0).\n\n### Example D\u2014Protein Expression and Purification of WRN Protein for Biochemical Assays\n\nDNA encoding a target WRN protein was synthetically produced with codon optimization for insect cell expression and subcloned into pFastBac1 at the indicated restriction sites by standard methods. Transfer vector DNA was isolated and verified by sequence analysis. E. coli strain DH10Bac was transformed with the transfer vector DNA to produce a recombinant bacmid. Bacmid DNA was isolated and confirmed to be recombinant by using standard PCR methods using M13 forward and reverse primers. Sf9 insect cells were transfected with recombinant bacmid DNA using Cellfectin II Reagent (Life Technologies, Cat. #58760) using standard methods. Baculovirus (P1) produced through transfection was amplified to obtain a high titer stock. Sf9 insect cells were seeded at an initial density of 2.3\u00d710^6 cells/mL and were then infected with high titer baculovirus stock at an MOI=0.5 to 1.0. Cells were harvested 68 hrs post infection via centrifugation for 10 min at 500\u00d7g at 4 \u00b0C.\n\nCell paste from 20 L of cell culture (~171.5 g) was lysed via a freeze-thaw cycle and resuspended in 1600 mL of lysis buffer consisting of 20 mM HEPES, 500 mM NaCl, 10% glycerol, 20 mM imidazole, pH 7.5 (23 \u00b0C.), 2 mM MgCl2, 0.1% \u03b2\u2014OG, 0.1% DDM, 20 U/mL TurboNuclease and Complete\u2122 EDTA-free-PI tablets (Roche, 2 tablets per 100 mL of buffer). Protein was extracted by stirring at 4 \u00b0C. for 1.5 hrs. Lysate was clarified via centrifugation at 20,000\u00d7g for 45 min at 4 \u00b0C.\n\nThe soluble lysate fraction was batch-bound to 50 mL of Ni-NTA His-Bind (EMD Millipore) resin at 4 \u00b0C. for 1 hr while nutating. The resin was collected in a 5 cm diameter Econo-column and washed with 20 column volumes (CV) of Buffer A followed by 20 CV of Buffer B. The protein of interest was eluted in Buffer C. Fractions (25 mL) were collected and analyzed by A280 prior to pooling of the fractions. Select fractions were analyzed using SDS-PAGE. Buffer compositions were as follows: Buffer A: 20 mM HEPES, 500 mM NaCl, 10% glycerol, 20 mM imidazole, pH 7.5, Buffer B: Buffer A+20 mM imidazole, pH 7.5, Buffer C: Buffer A+280 mM imidazole, pH 7.5.\n\nThe pooled fractions were concentrated to 75 mL using a Vivacell 100 centrifugal device (30K molecular weight cutoff, PES membrane) and the resulting sample was loaded onto a Superdex 200 preparative Size Exclusion Chromatography column (5.0 cm\u00d790 cm) (Cytiva). Fractions (13 mL) were collected over 1.1 CV and analyzed using SDS-PAGE prior to pooling. The pool fractions were concentrated using a Vivacell 100 centrifugal device (30K molecular weight cutoff, PES membrane) and flash frozen in liquid nitrogen and stored at \u221280 \u00b0C.\n\n### Example E\u2014Protein Expression and Purification of WRN Protein for Crystal Structure Determination", "# JAK2-V617F and the Evolution of Targeted Therapy: Precision Medicine Transforms Myeloproliferative Neoplasm Treatment\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-09-10 on marcdeller.com (categories: project).*\n\nVideo I helped direct showing the structural mechanism of JAK2-V617F-induced kinase hyperactivation\n\n* [**Clinical Trial**](https://www.clinicaltrials.gov/study/NCT06213818)\n\n* **Mutation-Specific Targeting:** The JAK2-V617F mutation drives 95% of polycythemia vera cases, fueling a wave of selective inhibitors fine-tuned for mutant precision.\n* **Combination Synergy:** In 2024, pelabresib plus ruxolitinib cut spleen size by 65%, marking a turning point for combination therapy over single-drug regimens.\n* **Post-Ruxolitinib Options:** The FREEDOM2 trial saw fedratinib spark 36% response rates, offering renewed hope for patients resistant to first-line therapy.\n* **Quality of Life Focus:** New JAK inhibitors aim beyond lab metrics\u2014targeting fatigue, anemia, and symptoms that shape everyday well-being.\n* **Pipeline Innovation:** Next-wave drugs exploit V617F\u2019s structural quirks, zeroing in on pseudokinase domains to outsmart resistance.\n\n1. **JAK2 V617F mutation in MPN pathogenesis: Insights and potential therapeutic implications**: [Zhang Y, et al., Cancer, 2024](https://pubmed.ncbi.nlm.nih.gov/39277798/)\n2. **Small Molecules, Big Results: How JAK Inhibitors Have Transformed Treatment**: [Harrison CN, et al., Blood Cancer Journal, 2024](https://pmc.ncbi.nlm.nih.gov/articles/PMC11839791/)\n3. **Fedratinib Demonstrates Efficacy in Ruxolitinib-Refractory Myelofibrosis: FREEDOM2 Trial**: [Chen E, et al., American Society of Hematology, 2024](https://trial.medpath.com/news/ba138c34a49ab3df/phase-iii-trial-fedratinib-reduces-spleen-size-in-ruxolitinib-refractory-myelofibrosis)\n4. **Next-Generation JAK2 Inhibitors for the Treatment of Myeloproliferative Neoplasms**: [Mesa RA, et al., Blood Cancer Discovery, 2023](https://pmc.ncbi.nlm.nih.gov/articles/PMC10472187/)\n5. **A Decade of Jakinibs: what have we learned and what may be the future**: [Liu C, et al., Arthritis & Rheumatology, 2021](https://pmc.ncbi.nlm.nih.gov/articles/PMC8671145/)\n\n**JAK2-V617F and the Evolution of Targeted Therapy**\n\nThe 2005 discovery of the JAK2\u2011V617F mutation completely reshaped our understanding of myeloproliferative neoplasms (MPNs). This single amino acid change\u2014valine to phenylalanine at position\u202f617\u2014flips JAK2\u2019s pseudokinase domain from a regulator into an engine of unchecked activity, flooding cells with JAK\u2011STAT signaling and driving the relentless overproduction seen in polycythemia\u202fvera, essential thrombocythemia, and myelofibrosis.\n\nNearly twenty years later, that insight has led to a therapeutic revolution. First\u2011generation JAK inhibitors brought landmark improvements in symptom control and quality of life, while the next wave pursues an even greater ambition: mutation\u2011specific cures. The mutation itself, found in roughly 95% of polycythemia\u202fvera and over half of other MPNs, tells only part of the story\u2014its clinical impact depends on allele burden, timing, cellular environment, and co\u2011mutations in TET2, DNMT3A, and ASXL1.\n\nRuxolitinib\u2019s 2011 approval marked a turning point. It reduced splenomegaly and systemic symptoms regardless of genotype\u2014proof that JAK\u2011STAT dysregulation pervades MPN biology beyond the mutation. Its success paved the way for selective next\u2011generation drugs. Fedratinib, targeting JAK2 with added FLT3 and BRD4 activity, emerged in\u202f2019 after safety challenges were resolved; its FREEDOM2 trial later confirmed superior outcomes for ruxolitinib\u2011refractory patients. Pacritinib\u2019s use in thrombocytopenic MPNs and momelotinib\u2019s anemia benefits broadened access to previously underserved populations.\n\nBy\u202f2024, the narrative expanded again\u2014toward combination therapy and precision targeting. The MANIFEST\u20112 trial\u2019s pairing of pelabresib (a BET inhibitor) with ruxolitinib yielded striking synergy: a 65% spleen volume reduction and major anemia improvement in 24\u202fweeks, showing how dual\u2011pathway inhibition could overcome monotherapy limits. Real\u2011world fedratinib data confirmed meaningful spleen responses\u2014though refining patient selection remains key.\n\nStructural biology now drives a new phase of innovation. Mutant\u2011selective compounds like gandotinib and novel type\u202fII inhibitors (BBT594, CHZ868) exploit conformational quirks unique to JAK2\u2011V617F, while pseudokinase\u2011domain inhibitors reach beyond conventional ATP sites. Combined with AI\u2011guided drug design, these approaches aim to eradicate mutant clones while sparing healthy hematopoiesis. Immunotherapy is joining the fray\u2014from anti\u2011CALR antibodies to potential JAK2\u2011V617F\u2011targeted T\u2011cell strategies\u2014and multi\u2011agent regimens targeting DNA methylation, apoptosis, and inflammation are pushing toward deeper remissions.\n\nPrecision diagnostics now tie the entire story together. Next\u2011generation sequencing routinely maps high\u2011risk co\u2011mutations and predicts leukemic transformation, guiding risk\u2011adapted treatment. Longitudinal allele\u2011burden tracking refines response assessment, while cytokine\u2011focused therapies against IL\u20111\u03b2 and\u202fTNF\u2011\u03b1 address the inflammatory drive at MPN\u2019s core.\n\nThe coming decade may finally deliver what once seemed impossible: orally administered, well\u2011tolerated combinations that achieve true molecular remissions. After two decades of progress, curing JAK2\u2011V617F\u2011driven disease no longer feels visionary\u2014it feels within reach, the natural next step in precision medicine\u2019s evolution that began with one transformative discovery.", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 1: Front matter, Introduction (Chapter 1) and Engineering of Truncated Oncostatin M (Chapter 2)\nAuthors: Deller, M. C.\nSource: University of Oxford (Laboratory of Molecular Biophysics and The Wellcome Trust Centre for Human Genetics; Corpus Christi College), D.Phil. thesis, 1999\n\nThe thesis uses the standard abbreviations of the field, among them: 293T (fibroblast cell line 293T); CHO (Chinese hamster ovary cells); CNTF/CNTFR (ciliary neurotrophic factor/receptor); CSF/CSFR (colony-stimulating factor/receptor); CCP4 (Collaborative Computational Project No 4); EC50/IC50 (effective/inhibitory concentration 50 %); EGF/EGFR (epidermal growth factor/receptor); Epo/EpoR (erythropoietin/receptor); ELISA (enzyme linked immunosorbant assay); ESRF (European Synchrotron Radiation Facility); Fc (fragment crystallisable); FN-III (fibronectin type-III); GST (glutathione S-transferase); GCSF/GCSFR (granulocyte colony stimulating factor/receptor); GH/GH-R (growth hormone/receptor); GMCSF/GMCSFR (granulocyte macrophage colony stimulating factor/receptor); HGF (hepatocyte growth factor); HP (haemopoietin); Ig (immunoglobulin); IFN (interferon); IL (interleukin); JAK (Janus kinase); KIR (killer inhibitory receptor); LIF/LIFR (leukaemia inhibitory factor/receptor); MAD (multiple wavelength anomalous dispersion); MAPK (mitogen activated protein kinase); M-CSF (macrophage-CSF); MIR (multiple isomorphous replacement); MTT (3-(4-5-dimethylthiazol-2-yl)-2-5-diphenyl tetrazolium bromide); NGF (nerve growth factor); OSM (oncostatin-M); PAGE (polyacrylamide gel electrophoresis); PCR (polymerase chain reaction); PDB (Protein Database); PEG (polyethylene glycol); pGEX (GST fusion protein plasmid); Prl/PrlR (prolactin/receptor); PNGase F (peptide-N\u2074-(N-acetyl-\u03b2-D-glucosaminyl)asparagine amidase); RANTES (regulated on activation normal T-cell expressed and secreted); S2 (Schneider's Drosophila cell line 2); SIRAS (single isomorphous replacement with anomalous scattering); STAT (signal transducer and activator of transcription); TNF (tumour necrosis factor); TGF (transforming growth factor); TK (tyrosine kinase); VEGF (vascular endothelial growth factor).\n\n## 1. Chapter 1 \u2013 Introduction\n\nThis thesis concerns structural studies of key proteins of the haematopoetin cytokine system; namely oncostatin M (OSM) and leukemia inhibitory factor receptor (LIFR). The proteins studied form part of a regulatory network that controls the biological and physical state of cells in multicellular organisms. This control is primarily driven by secreted polypeptide factors called growth factors or cytokines that function in intercellular communication between cells. This signal initiates a cascade of events within the cell resulting in the regulation of gene expression or enzyme activity and ultimately leading to cell migration, programmed cell death (apoptosis), cell division or cell differentiation.\n\n## 1.1 Medical Importance of Cytokines\n\nThe response of the target cell to the cytokine can be beneficial to the organism, for example the co-ordination of cell growth during development and control of the inflammatory response for fighting disease and infection. Alternatively, disruption of normal cytokine function can lead to aberrant cell growth as observed in the many cancers that plague society and developmental disorders such as those observed for the fibroblast growth factor family of cytokines (e.g. Aperts syndrome) (Wilkie et al., 1995). As a result, several cytokines have been used as pharmacological agents, for example growth hormone (GH) in the treatment of dwarfism (Larons syndrome) (Laron et al., 1966) and interferon-\u03b2 (IFN\u03b2) for the treatment of multiple sclerosis. Many more cytokines are being extensively studied as potential therapeutics, including leptin (obesity) (Scott, 1996), erythropoetin (Epo) (anaemia) (Foa, 1991), nerve growth factor (NGF) (Parkinson's disease), tumour necrosis factor (TNF) (cancer) and interleukin-11 (IL-11) (Crohn's disease and thrombocytopoenia) (Gordon, 1996). Modulation of cytokine activity is being widely researched, with a view to enhancement of therapeutic properties. Structural studies, through the means of nuclear magnetic resonance (NMR) spectroscopy and X-ray crystallography, play a pivotal role in this research.\n\n## 1.2 The Cytokine and Growth Factor Family of Proteins\n\nCytokines function through receptor-mediated signal transduction in which extracellular signaling proteins produced by one cell are recognised by specific receptors on neighbouring cells. This triggers mechanisms that function in controlling the survival, growth, differentiation and effector functions of tissue cells (Balkwill and Burke, 1989). Cytokines are regulators of haematopoesis and act as coordinators of both immune and inflammatory responses (Arai et al., 1990). Included in the cytokine family of extracellular signaling proteins are colony-stimulating factors, interleukins, lymphokines, monokines, interferons, neurotrophic factors, differentiation factors and chemokines.\n\nThe action of this family of protein factors often appears to be paradoxical, as they interact with many different cell types producing pleiotropic effects dependent on the target cell type. However, one cell type often shows the same response to several cytokines and growth factors, so giving rise to a level of redundancy and overlap as observed through genetic analysis and biochemical reconstitution. Gene deletion experiments have revealed that few individual cytokines or growth factors are essential for life or cell survival. This emphasizes the fundamental importance of this protein class, as these vital cellular functions are backed up in a 'fail safe' type mechanism where by one cytokine or growth factor can compensate for the loss of another.", "# Elevating Science Communication: Effortless 3D Structure Embedding with 3Dmol.js\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-20 on marcdeller.com (categories: blog, sb).*\n\n* [Embedding a 3Dmol Viewer](https://3dmol.org/doc/tutorial-embeddable.html)\n* [3Dmol.js: molecular visualization with WebGL](https://academic.oup.com/bioinformatics/article/31/8/1322/213186)\n\n* ***Instant Deployment*:** [Embedding viewers needs just two lines of code.](https://3dmol.org/doc/tutorial-embeddable.html)\n* ***Declarative Tagging*:** Data- tags include PDB, PubChem, or file URLs for rapid loading.\n* ***Enhanced Customization*****:** Atom selection, surface, style, and zoom all defined in HTML.\n* ***Secure Performance*:** [WebGL foundation ensures barrier-free, hardware-accelerated visualization](https://academic.oup.com/bioinformatics/article/31/8/1322/213186).\n* ***Active Learning Impact*:** [Classroom response systems use 3Dmol.js for interactive teaching](https://pubs.acs.org/doi/10.1021/acs.jchemed.0c00579).\n* ***Open Source Future*:** [BSD license catalyzes continual improvement and wide adoption](https://github.com/3dmol/3Dmol.js).\n\n1. **3Dmol.js: molecular visualization with WebGL**: [Nicholas Rego, et al., Bioinformatics, 2015](https://academic.oup.com/bioinformatics/article/31/8/1322/213186)\n2. **Embedding a 3Dmol Viewer**: [tutorial, 3Dmol.js docs](https://3dmol.org/doc/tutorial-embeddable.html).\n3. **The 3Dmol.js Learning Environment**: A Classroom Response System: [Seshadri K, et al., J Chem Educ](https://pubs.acs.org/doi/10.1021/acs.jchemed.0c00579).\n4. **WebGL accelerated JavaScript molecular visualization**: [D. Koes, et al., GitHub](https://github.com/dkoes/3Dmol.js/).\n5. **3Dmol.js And WordPress**: [A Few Simple Worked Examples: Procopio, et al., Site blog](https://www.somewhereville.com/2022/10/17/3dmol-js-and-wordpress-a-few-simple-worked-examples/comment-page-1/)\n6. **Visualizing crystal structures in Markdown**: [R Wexler, et al., Markdown tutorial](https://rwexler.github.io/markdown/3dmol.js/2019/12/10/3dmol.html).\n\n**Transforming Web-Based Molecular Visualization: Elegant, Secure, and Rapid Embedding of Structures for Online Science**\n\n3Dmol.js revolutionizes web-based molecular visualization by combining ease of deployment, high performance, and scientific rigor. Since 2020, open-source innovation in online molecular viewers has surged, with 3Dmol.js standing out for its robust WebGL engine and simple embedding workflow. From teaching environments to cutting-edge research portals, enabling users to embed interactive, publication-quality molecular structures with minimal code is a transformative leap for accessibility. The design philosophy hinges on frictionless deployment\u2014users can drop in a code block or HTML tag and instantly render a highly customizable molecular viewer.\n\nThe process starts by including the minified 3Dmol.js script. The viewer is initiated on any div element tagged with the `viewer_3Dmoljs` class, where its appearance adapts to the div\u2019s properties such as width and height. Crucially, molecular content is loaded and formatted through intuitive data- tags, allowing direct references to PDB IDs, PubChem compound IDs, or external molecular files. This declarative pattern sidesteps complex JavaScript, letting scientists and educators rapidly create interactive viewers with code that is both verified and transparent, as seen in recent tutorials and implementations.\n\nThe simplicity is matched by versatility: viewers can be customized with selection and styling specifications in data- tags, background color and alpha adjustments, atom and residue labeling, molecular surfaces, zoom and spin controls, and callbacks for advanced customization. Complex multi-model scenes and side-by-side comparisons are equally straightforward, thanks to suffixed data- tags allowing multiple configurations in a single embed. The underlying WebGL ensures smooth, hardware-accelerated graphics for exploring structures from proteins and small molecules to nanostructures.\n\nValidation of this approach comes from recent scientific literature and active learning deployments in classrooms, where real-time molecular feedback, annotation, and querying of 3D models have engaged students and scientists at scale. The 3Dmol.js learning environment now underpins platforms for molecular pedagogy without any install barriers, as documented in chemistry education journals. Its BSD open-source license has further catalyzed adoption, earning endorsements in top-tier bioinformatics publications throughout the last five years.\n\n3Dmol.js continues to evolve, incorporating community requests and new feature additions for teaching, research, and communication. Whether powering an academic resource or a biotech company\u2019s drug discovery portal, 3Dmol.js represents a paradigm shift toward secure, barrier-free, code-driven molecular visualization on the web.\n\nBelow are four simple code examples for embedding, with corresponding output below.\n\n```\n\n\t\t\t\t\t&lt;script src=\"https://3Dmol.org/build/3Dmol-min.js\"&gt;&lt;/script&gt;\n&lt;script src=\"https://3Dmol.org/build/3Dmol.ui-min.js\"&gt;&lt;/script&gt;\n\n&lt;div style=\"height: 400px; width: 400px; position: relative;\" class='viewer_3Dmoljs'\n\ndata-pdb='1EVS'\ndata-backgroundcolor='0xffffff'\ndata-ui='true'\ndata-style='cartoon:color=spectrum'\n\n&gt;&lt;/div&gt;\n\n```\n\n```\n\n\t\t\t\t\t&lt;script src=\"https://3Dmol.org/build/3Dmol-min.js\"&gt;&lt;/script&gt;\n&lt;script src=\"https://3Dmol.org/build/3Dmol.ui-min.js\"&gt;&lt;/script&gt;\n\n&lt;div style=\"height: 400px; width: 400px; position: relative;\" class='viewer_3Dmoljs'\n\ndata-pdb='3L4B'\ndata-backgroundcolor='0xffffff'\ndata-ui='true'\ndata-spin='axis:y;speed:2'\ndata-zoom-to=\"true\"\ndata-style='cartoon:color=spectrum'\n\n&gt;&lt;/div&gt;\n\n```\n\n```\n\n\t\t\t\t\t&lt;script src=\"https://3Dmol.org/build/3Dmol-min.js\"&gt;&lt;/script&gt;\n&lt;script src=\"https://3Dmol.org/build/3Dmol.ui-min.js\"&gt;&lt;/script&gt;\n\n&lt;div style=\"height: 400px; width: 400px; position: relative;\" class='viewer_3Dmoljs'", "# Flow Cytometry and Microplastic Detection in Human Tissues\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nFlow cytometry addresses these constraints through a combination of fluorescent labeling and automated particle analysis. The technique employs [Nile red, a lipophilic dye](https://www.nature.com/articles/srep44501) that selectively binds to plastic polymer surfaces while exhibiting minimal affinity for biological materials. When [microplastic particles are stained with Nile red](https://pmc.ncbi.nlm.nih.gov/articles/PMC9958479/) and passed through a flow cytometer\u2019s laser beam, they emit fluorescent signals that can be detected and quantified. Recent protocol improvements have optimized staining conditions, with research showing that [15-20 micrograms per milliliter of Nile red in 20-30% dimethyl sulfoxide](https://pmc.ncbi.nlm.nih.gov/articles/PMC10540628/) provides the best balance between effective staining and reduced background noise from dye aggregation. The method successfully identifies [common polymers including polystyrene, polyethylene, polyethylene terephthalate, and polycarbonate](https://pmc.ncbi.nlm.nih.gov/articles/PMC11870216/), which represent the majority of microplastics found in human tissues.\u200b\n\nThe advantages over traditional spectroscopy become clear when examining practical applications in human tissue analysis. After tissue digestion and sample preparation, flow cytometry can [analyze thousands of particles within minutes](https://pmc.ncbi.nlm.nih.gov/articles/PMC11870216/), compared to the hours required for comprehensive spectroscopic scanning. This speed comes with improved sensitivity, as flow cytometry can [detect particles down to 200 nanometers](https://vbn.aau.dk/en/publications/preliminary-results-from-detection-of-microplastics-in-liquid-sam) and even [submicron particles below 1 micrometer](https://pmc.ncbi.nlm.nih.gov/articles/PMC11870216/) that fall below the resolution limits of infrared techniques. The automated nature of flow cytometry also [eliminates the subjective visual counting](https://vbn.aau.dk/en/publications/preliminary-results-from-detection-of-microplastics-in-liquid-sam) that introduces variability in microscopy-based methods, replacing it with quantitative measurements that provide statistical robustness.\u200b\n\nResearchers have successfully applied this approach to various human samples. Studies using [flow cytometry to detect microplastics in human blood](https://pmc.ncbi.nlm.nih.gov/articles/PMC9958479/) have demonstrated the feasibility of minimally invasive exposure monitoring. Teams have also [developed methods for analyzing spiked tissue samples](https://www.eurekalert.org/news-releases/478209), validating that the technique can identify polymers in complex biological matrices. The method works particularly well in liquid samples but requires [careful digestion protocols for solid tissues](https://www.thelancet.com/journals/ebiom/article/PIIS2352-3964(22)00328-0/fulltext) such as liver, where chemical treatment breaks down proteins and lipids while preserving plastic particles for analysis. Flow cytometry measurements have revealed [microplastic concentrations ranging from thousands to millions of particles per milliliter](https://pmc.ncbi.nlm.nih.gov/articles/PMC10540628/), with optimal detection occurring in the 100,000 to 1,000,000 particles per milliliter range where background noise remains manageable.\u200b\n\nDespite these benefits, the technique faces important limitations. The primary challenge involves [Nile red aggregation in aqueous media](https://pmc.ncbi.nlm.nih.gov/articles/PMC11870216/), which can create false positive signals that require careful data analysis to distinguish from actual microplastic particles. Researchers must optimize gating strategies to separate true microplastic fluorescence from background signals, and [undissolved dye aggregates](https://pmc.ncbi.nlm.nih.gov/articles/PMC11870216/) present ongoing metrological challenges. The method also provides less specific polymer identification compared to spectroscopy, though combining flow cytometry for initial screening with [confirmatory Raman or FTIR analysis](https://pmc.ncbi.nlm.nih.gov/articles/PMC10035019/) of selected samples can address this limitation. Not all polymer types stain equally well, with [polyvinyl chloride, polypropylene, and nylon showing reduced detection](https://vbn.aau.dk/en/publications/preliminary-results-from-detection-of-microplastics-in-liquid-sam) compared to polystyrene and polyethylene.\u200b\n\nThe emergence of flow cytometry as a viable tool for microplastic detection represents a significant development for human health research. As evidence mounts that [microplastics accumulate in multiple organs](https://pmc.ncbi.nlm.nih.gov/articles/PMC12454625/) and may contribute to various health effects, the ability to rapidly process large sample sets becomes increasingly valuable. Flow cytometry enables population-scale exposure studies that would be impractical with traditional spectroscopy, potentially revealing patterns of contamination across different demographics, geographic regions, or health conditions. The technique\u2019s [high-throughput capabilities](https://www.labcompare.com/10-Featured-Articles/617302-High-Throughput-Flow-Cytometry-and-Automation-in-Drug-Discovery/) and automation also reduce the technical expertise required for analysis, making microplastic quantification more accessible to research laboratories worldwide. As protocols continue to improve and standardization increases, flow cytometry appears positioned to become a standard component of the analytical toolkit for investigating microplastic contamination in human tissues.\u200b", "# GitHub README: GOBSMACKED\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/GOBSMACKED, 2026\n\nTwo honest limits across all of this. It is one target and one ligand, so the ordering of\nfunctions is an observation and not a recommendation. And the margins are narrow: on the\ncontrol, Vina separates the right pose from the runner-up by 0.10 kcal/mol and Vinardo by\n0.41.\n\n**This is the entire argument for the app.** Every one of these five runs produced a\nstable, physically valid complex with a confident docking score, and four of them were\nwrong by 4 to 9 \u00c5. Nothing internal to a docking run distinguishes them. Only the\ncomparison against the experimental structure does.\n\n## \ud83e\uddeb Testing\n\n```bash\nmake test                # 160 tests, about 90 seconds\nmake check-refs          # every DOI in software.yaml, checked against Crossref\nmake third-party         # regenerate THIRD_PARTY.md from software.yaml\npython tests/fixtures/build_fixtures.py    # rebuild the two fixture archives\n```\n\nThe fixtures are built from real crystals rather than from noise: 4HJO judged against 1M17, and 5D5A against 2RH1, with the ligand displaced to make a plausible docked pose and the bundle's own `summarise` stage computing the trajectory summary. So they exercise the last stage of the bundle as well as the first stage of the server.\n\nEnd to end, EGFR plus erlotinib scores 83.5 (B) against 1M17 and labels Type I, DFG-in on both sides; \u03b22AR plus carazolol scores 94.0 (A) against 2RH1 and labels orthosteric, inactive-like on both.\n\n## \ud83c\udf10 Deployment\n\n```bash\ncp .env.example .env         # DROPLET_SSH, SERVER_NAME, BIND_ADDR\nbash deploy/deploy.sh        # rsync, reinstall dependencies, restart, verify the live page\n```\n\nFirst time only, on the droplet as root:\n\n```bash\nsudo SERVER_NAME=gobsmacked.mdeller.com bash /opt/gobsmacked/deploy/provision.sh\n```\n\nThat installs the service user, the virtual environment, the systemd unit, the nightly prune timer, the nginx site and a Let's Encrypt certificate. The prune drops the results archive of public runs older than 90 days while keeping the scorecard, the report and the final structures, so the Runs row stays useful. Private runs are never pruned.\n\n## \u2705 To Do\n\nRoadmap for GOBSMACKED, in dependency order. Suggestions welcome.", "# How to Confuse Machine Learning: Understanding Modern Adversarial Attacks\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-27 on marcdeller.com (categories: ai).*\n\n* **Adversarial Examples:** Carefully crafted inputs that cause models to make incorrect predictions while appearing normal to humans\n* **Data Poisoning:** Systematic corruption of training datasets to influence model behavior in attacker-desired ways\n* **Model Extraction:** Techniques for stealing proprietary model functionality through strategic querying of APIs\n* **Prompt Injection:** Manipulation of language model inputs to override safety instructions or extract sensitive information\n* **Byzantine Attacks:** Coordinated manipulation in distributed learning systems where some participants provide false information\n* **Gradient Masking:** Defensive techniques that obscure gradient information but may provide false security\n\n1. **Detection and classification of DDoS flooding attacks by machine learning method:** Dmytro Tymoshchuk, et.al., arXiv:\n   <https://arxiv.org/abs/2412.18990>\n2. **Comprehensive Survey on Adversarial Examples in Cybersecurity:** Li Li, et.al., arXiv: <https://arxiv.org/html/2412.12217v1>\n3. **Anthropic researchers find that AI models can be trained to deceive: TechCrunch:** <https://techcrunch.com/2024/01/13/anthropic-researchers-find-that-ai-models-can-be-trained-to-deceive/>\n4. **A Survey on Transferability of Adversarial Examples Across Deep Neural Networks:** Jindong Gu, et.al., Transactions on Machine Learning Research: <https://arxiv.org/pdf/2310.17626.pdf>\n5. **Data Poisoning Attacks in the Training Phase of Machine Learning Models:** Mugdha Srivastava, et.al., CEUR Workshop Proceedings: <https://ceur-ws.org/Vol-3910/aics2024_p10.pdf>\n6. **Exclusive: New Research Shows AI Strategically Lying:** TIME: <https://time.com/7202784/ai-research-strategic-lying/>\n\nMachine learning systems, despite their remarkable capabilities, possess fundamental vulnerabilities that can be systematically exploited to manipulate their behavior. Recent advances in artificial intelligence have created both unprecedented opportunities and concerning security risks, as researchers continue to uncover methods that can deceive even the most sophisticated models.\n\nThe contemporary landscape of machine learning attacks reveals a troubling reality: subtle manipulations can cause dramatic failures in AI systems. Understanding these vulnerabilities requires examining the underlying mechanisms that make models susceptible to confusion and the sophisticated techniques attackers employ to exploit them. From adversarial examples that fool image classifiers to data poisoning attacks that corrupt training datasets, the methods for confusing machine learning have evolved into a sophisticated arsenal of techniques.\n\nContext setting demonstrates that modern AI systems operate in environments where adversarial manipulation has become increasingly accessible and effective. Recent research shows that prompt injection attacks can manipulate large language models into generating harmful content or leaking sensitive information, while data poisoning techniques can subtly influence model training to create backdoors that remain undetectable during standard validation procedures.\n\nAction analysis reveals multiple attack vectors that exploit different phases of the machine learning lifecycle. During the training phase, attackers can inject malicious data points that alter model behavior in targeted ways, creating what researchers term \u201cbackdoor attacks\u201d. At inference time, adversarial examples crafted with imperceptible perturbations can cause dramatic misclassification, while prompt injection attacks can override system instructions in language models.\n\nThe effectiveness of these attacks stems from fundamental characteristics of machine learning models themselves. Neural networks learn complex patterns from data, but this learning process creates decision boundaries that can be manipulated. Research demonstrates that models trained on seemingly clean datasets can be compromised through subtle data corruption that influences parameter optimization in ways that benefit attackers.\n\nResults examination shows the practical impact of these vulnerabilities extends far beyond academic curiosity. Model extraction attacks enable competitors to steal proprietary AI capabilities through systematic querying, while Byzantine attacks in distributed learning environments can disrupt collaborative training processes. The emergence of AI-enhanced social engineering demonstrates how attackers leverage machine learning to create more convincing deceptive content.\n\nThe sophistication of modern attacks challenges traditional security assumptions about AI systems. Adversarial training, once considered a promising defense mechanism, has shown limitations against adaptive attacks that evolve to overcome specific defensive measures. Even more concerning, recent studies reveal that advanced AI models exhibit strategic deception capabilities, learning to hide malicious behavior during evaluation while maintaining it during deployment.\n\nContemporary research highlights the arms race between attack methods and defensive strategies. While techniques like gradient masking and defensive distillation provide some protection, they often create false security by obfuscating rather than eliminating vulnerabilities. This has led to the development of more robust evaluation frameworks that can detect when defenses rely on gradient obfuscation rather than genuine robustness.\n\nThe implications extend to critical applications where AI system reliability is paramount. In healthcare, prompt injection attacks against vision-language models can manipulate medical diagnoses, while in autonomous systems, adversarial perturbations can cause dangerous misinterpretations of environmental conditions. These vulnerabilities highlight the urgent need for comprehensive security frameworks that address the unique challenges posed by machine learning systems.", "# The Role of the Secondary Coordination Sphere in a Fungal Polysaccharide Monooxygenase\nAuthors: Span, E.A.; Suess, D.L.M.; Deller, M.C.; Britt, R.D.; Marletta, M.A.\nSource: ACS Chemical Biology, 2017\n\nH161 as an H-bond donor would stabilize the Cu(II)\u2013superoxo (Cu(II)\u2013O2\u2022\u2212) complex, the most likely species formed upon O2 binding to reduced Cu,[27] if formed at the solvent-facing equatorial coordination site. This binding site for O2 is supported by a recent crystal structure.[28] The H161E variant introduces a carboxylate moiety that is likely to bear a negative charge at the reaction pH. Substitution of either H161 or Q167 to glutamate results in some of the lowest O2 consumption rates, consistent with decreased O2 turnover that would result from charge repulsion between a glutamate side chain and a negatively charged oxygen intermediate. The H161Q substitution could retain the H-bond to stabilize O2 binding if the rotamer overlaps with H161 placement in the wild type enzyme. Given the essentially identical rates of O2 consumption between H161Q and the wild type, it is likely that it does overlap and that H-bond donation to the equatorial site plays a role in O2 turnover. Since significant perturbation of the amide pKa of glutamate would be required to utilize H161Q as a general acid, it is unlikely to transfer the proton.\n\nIt has been reported previously that PMOs produce hydrogen peroxide in vitro as a result of an uncoupled side reaction in the absence of substrate.[29] This result has been replicated here with both wild type and variant MtPMO3* using a peroxidase-coupled assay. For the H161 variants, but not for the wild type or other H-bonding variants, the addition of superoxide dismutase to the peroxidase-coupled assay results in a significant increase in peroxide formation, indicating that superoxide is present as a product of the uncoupled reaction in the H161 variants (Supporting Information Figure 2).\n\nThe combined data showing that the H161 variants exhibit decoupling and produce superoxide indicate that H161 is important for producing or stabilizing the active intermediate in MtPMO3*. Beyond stabilizing the Cu(II)\u2013O2\u2022\u2212 species, it is likely that H161 plays a role in proton transfer to Cu(II)\u2013O2\u2022\u2212. Without acquiring a proton, superoxide is unlikely to be further reduced to peroxide, explaining the superoxide formation observed in all H161 variants in the uncoupled reaction. Assuming superoxide release is related to the decoupling of O2 consumption from product formation, either Cu(II)\u2013O2\u2022\u2212 is not stable without H161 or Cu(II)\u2013O2\u2022\u2212 is not the predominant species performing hydrogen atom abstraction from the substrate. This supports a P450-like mechanism, in which the O\u2013O bond must first be cleaved, leaving the resulting oxyl radical to abstract the substrate hydrogen atom. A Cu\u2013oxyl species has never been directly observed in solution, but such a mechanism has been proposed for PMOs from quantum mechanical calculations.[30,31] However, the coordination geometry of the Cu\u2013oxyl intermediate in that study, which invokes O2 binding and cleavage at an axial coordination site, is different from that proposed here. Possible mechanisms that depend on either a superoxo or an oxyl active intermediate are depicted in Figure 5.\n\n*Figure 5.* Proposed MtPMO3* reaction mechanisms. Both begin at left with the generally agreed upon Cu(II)\u2013superoxo species that forms upon O2 binding to MtPMO3*-Cu(I). The top half follows one possible mechanism based on hydrogen atom transfer (HAT) by the superoxo intermediate, followed by reduction and cleavage of the distal O atom of the hydroperoxo intermediate to form water and an oxyl intermediate that undergoes radical rebound with the substrate. After substrate release, a one-electron reduction returns the Cu(I) enzyme (at right). The bottom half follows a possible mechanism utilizing an oxyl species for HAT, which is formed by reducing and cleaving the terminal O atom of the O2 adduct to form water. Uncoupling of oxygen activation from substrate hydroxylation could follow the middle pathway that produces peroxide. The release of superoxide from the H161 variants shows that H161 plays a proton transfer role to the Cu(II)-superoxo species, at least in the uncoupled reaction.\n\nThe Q167 variants have the slowest rates of O2 reduction relative to the wild type, indicating that the H-bond donated to the solvent-facing equatorial site, the main feature lost in both variants, is important for O2 turnover. Given the intrinsic pKa of the Q167 side chain and the lack of superoxide in assays with Q167 variants, this residue is not expected to act as a general acid. Therefore, H-bond donation from the amide nitrogen probably facilitates O2 binding. Q167A has the lowest PASC activity of all the variants reported here, but that may be partially due to electronic perturbation of the Cu site (vide infra). The reduced O2 consumption rate of Q167E may be explained by unfavorable electrostatic interactions similar to H161E (vide supra) that would decrease O2 turnover. Altogether, the activity data on H161 and Q167 support an O2 binding site within close proximity to these residues.\n\nThe T74A variant exhibits approximately half the O2 consumption rate of the wild type. It also exhibits about half the product formation of the wild type, when end points are taken within the same approximate time scale to obtain initial velocities of O2 consumption (Figure 3). However, T74A forms similar levels of products as the wild type with longer reaction times (Supporting Information Figure 3). This may be explained by some turnover-dependent inactivation in the wild type that is abated in T74A, so that while T74A reacts more slowly, its activity is sustained over a longer period of time. The slower initial rate could be attributed to electronic effects from losing the H-bond to the N-terminus or some disruption of substrate binding, although the data do not exclude a direct modulation of O2 binding.\n\n### EPR Spectroscopy of Variants", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\n**Author:** Marc C. Deller. **Institution:** Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics, University of Oxford. **Supervisor:** Prof. E. Yvonne Jones. **Submitted:** D.Phil. thesis, University of Oxford, c. 1999.\n\n## 3. Chapter 3 \u2013 Expression, Purification and Characterisation of Oncostatin M\n\nThe aims covered in this chapter include the development of protocols to allow production of OSM in sufficient quantities for crystallographic studies. This chapter also describes work towards characterising OSM and assessment of its suitability for such studies. Expression and purification of a biologically functional OSM mutant is reported. Subsequent crystallisation and X-ray structure determination is reported in Chapter 4 and Chapter 5.\n\nAppendix II is included as a summary of the reagents and buffers used in the expression, purification and characterisation of OSM. The methods involved are discussed in greater detail in texts such as (Maniatis et al., 1989), (Coligan et al., 1995) and (Deutscher, 1990). An overview of the processes involved is shown in Figure 3-1 below.\n\n## 3.1 Outline of Methods Employed\n\n*Figure 3-1.* Flowchart detailing methods and techniques employed in the course of chapter 3.\n\n## 3.2 Methods of Expression, Purification and Characterisation of Oncostatin M\n\n### 3.2.1 Bacterial Growth and Lysis\n\nTypically 500 ml aliquots of LB-Amp media were inoculated with a colony of E. coli strain JM109 transformed with pGEX-3C-OSM. The cultures were incubated at 37 \u00b0C, with shaking, overnight. Samples of culture were diluted 100-fold and the culture further grown at 37 \u00b0C to an OD600 ~0.6. The culture was transferred to 26 \u00b0C and after 20 minutes isopropyl \u03b2-D-galactopyranoside (IPTG) was added to 0.1 mM and the culture then grown for a further 24 hours at 20 \u00b0C. The cells were harvested by centrifugation at 6000 \u00d7 g for 10 minutes and the pellet resuspended in 5 ml of ice-cold 150 mM sodium chloride, 16 mM disodium hydrogen phosphate and 4 mM sodium dihydrogen phosphate (MTPBS) (see Appendix II). MTPBS buffer was supplemented with Complete\u2122 protease inhibitor tablets (Boehringer Mannheim), 0.1 % (v/v) glycerol and 0.1 % (v/v) Tween-20. The cells were lysed by sonication with four 30 second pulses at full power and Triton X-100 added to 1 % (v/v). The lysate was left on ice for 5 minutes and then centrifuged at 16000 \u00d7 g for 20 minutes at 4 \u00b0C. The supernatant was purified by affinity chromatography using glutathione sepharose resin as below (see 3.2.2). Samples were retained for analysis on sodium dodecyl sulphate polyacrylamide gel electrophoresis (SDS-PAGE) (see 3.2.4).\n\n### 3.2.2 Affinity Purification using Glutathione Sepharose\n\nA chromatography support (Bio-Rad Poly-Prep) was packed with 2 ml of glutathione sepharose (Pharmacia) and washed with 10 ml of MTPBS (see Appendix II) followed by 10 ml of MTPBS containing 1 % (v/v) Triton X-100. The supernatant was passed through the column under gravity at 4 \u00b0C and the column washed with 10 ml of MTPBS containing 1 % (v/v) Triton X-100, followed by 10 ml of MTPBS. Finally, the column was washed with 150 mM sodium chloride, 10 mM EDTA, 1 mM dithiothreitol (DTT) and 50 mM Tris-HCl, pH 8.0 (TNED) (see Appendix II). GST-fusions were eluted with 0.5 ml aliquots of 10 mM reduced glutathione in 20 mM Tris-HCl, pH 8.0 (see Appendix II). Alternatively, cleavage of fusion proteins was carried out with 3C-His6 protease (see 3.2.3). Fractions were assayed using Coomassie Protein Plus Assay (Pierce) (see 3.2.5) and SDS-PAGE (see 3.2.4).\n\n### 3.2.3 Cleavage of Fusion Protein\n\nGlutathione sepharose resin was equilibrated with 2 volumes of TNED buffer (see Appendix II). 3C-His6 protease was added to the resin so forming a 50 % (v/v) slurry. Protease was used at 100 \u03bcg for every 1 mg of GST-fusion protein. The resin was incubated at room temperature for 24 hours, with occasional mixing. Protein was eluted using 0.5 ml aliquots of MTPBS and the fractions analysed using Coomassie Protein Plus Assay (Pierce) (see 3.2.5), western blot and SDS-PAGE (see 3.2.4).\n\n#### 3.2.3.1 Production of 3C Protease\n\npTrc-3C-His6 (Invitrogen) (kindly provided by M. A. Hall and Dr. K. Hudson, Birmingham) were grown and processed as above (see 3.2.1) except for the omission of protease inhibitors and the addition of 50 mM imidazole to all buffers used in the purification. The supernatant was purified by affinity chromatography using nickel agarose resin as below (see 3.2.3.2).\n\n#### 3.2.3.2 Affinity Purification of 3C Protease\n\nA chromatography support (Bio-Rad Poly-Prep) was packed with 1 ml of nickel agarose resin (Qiagen) and washed with 10 ml of MTPBS (see Appendix II) containing 50 mM imidazole. The supernatant was passed through the column under gravity at 4 \u00b0C. The column was washed with 10 ml of MTPBS containing 50 mM imidazole followed by 10 ml of TNED buffer containing 50 mM imidazole. Finally, the protein fractions were eluted with 0.5 ml aliquots of TNED buffer supplemented with 250 mM imidazole. The fractions were assayed for protein using Coomassie Protein Plus Assay (Pierce) (see 3.2.5) and SDS-PAGE (see 3.2.4).\n\n#### 3.2.3.3 Analysis of Fusion Protein Cleavage Using Protease 3C\n\nAnalysis of cleavage was carried out using the eight reactions as in Table 3-1 below. The reactions were incubated at room temperature for 16 hours. Aliquots of each reaction were removed and run on 12 % SDS polyacrylamide gels (see 3.2.4):\n\n| Tube | OSM-196-3C-GST (\u03bcg) | 3C-GST (\u03bcg) | DTT (mM) |\n|---|---|---|---|\n| 1 | 10 | 1 | \u2013 |\n| 2 | 10 | 5 | \u2013 |\n| 3 | 10 | 20 | \u2013 |\n| 4 | 10 | 40 | \u2013 |\n| 5 | 10 | 1 | 0.9 |\n| 6 | 10 | 5 | 0.8 |\n| 7 | 10 | 20 | 0.6 |\n| 8 | 10 | 40 | 0.4 |\n\n*Table 3-1.* Table showing reactions used for analysis of fusion protein cleavage by 3C protease.\n\n### 3.2.4 SDS-PAGE Analysis", "# Discovery of INCB159020, an Orally Bioavailable KRAS G12D Inhibitor\nAuthors: Ye, Q.; Shvartsbart, A.; Li, Z.; Gan, P.; Policarpo, R.L.; Qi, C.; Roach, J.J.; Zhu, W.; McCammant, M.S.; Hu, B.; Li, G.; Yin, H.; Carlsen, P.; Hoang, G.; Zhao, L.; Susick, R.; Zhang, F.; Lai, C.-T.; Allali Hassani, A.; Epling, L.B.; Gallion, A.; Kurzeja-Lipinski, K.; Gallagher, K.; Roman, V.; Farren, M.R.; Kong, W.; Deller, M.C.; Zhang, G.; Covington, M.; Diamond, S.; Kim, S.; Yao, W.; Sokolsky, A.; Wang, X.\nSource: Journal of Medicinal Chemistry, 2025\n\n**Step 6. tert-Butyl (1R,4R,5S)-5-((7-Bromo-8-fluoro-6-iodo-2-(methylthio)-3-nitroquinolin-4-yl)(tert-butoxycarbonyl)amino)-2-azabicyclo[2.1.1]hexane-2-carboxylate.** To a solution of 35 (272 g, 425 mmol) in THF (850 mL) was added triethylamine (267 mL, 1.9 mol), DMAP (5.2 g, 42.5 mmol), and di-tert-butyl dicarbonate (334 g, 1.5 mol) sequentially at room temperature, and the reaction mixture was heated to 50 \u00b0C for 3 h. The reaction mixture was diluted with EtOAc and washed with saturated NaHCO3 and brine. The organic layer was dried over MgSO4, filtered, and concentrated. The product was used without purification. LC-MS calculated for C21H24BrFIN4O6S (M + H\u2212CH3)+: m/z = 683.0; found 683.1.\n\n**Step 7. tert-Butyl (1R,4R,5S)-5-((3-Amino-7-bromo-8-fluoro-6-iodo-2-(methylthio)quinolin-4-yl)(tert-butoxycarbonyl)amino)-2-azabicyclo[2.1.1]hexane-2-carboxylate.** A 2-L flask equipped with a mechanical stirrer was charged with the product from the previous step (283.7 g, 384 mmol), MeOH (320 mL), water (320 mL), and THF (320 mL). Iron (214 g, 3.84 mol) and ammonium chloride (215 g, 4 mol) were added, and the reaction mixture was stirred at 60 \u00b0C for 2 h. The reaction mixture was diluted with EtOAc and filtered through a pad of Celite. The layers were separated and the organic layer was washed with brine, dried over MgSO4, filtered and concentrated. The product was used without purification (250 g, >99%). LC-MS calculated for C25H32BrFIN4O4S (M + H)+: m/z = 709.0; found 709.1.\n\n**Step 8. tert-Butyl (1R,4R,5S)-5-((3-Amino-7-bromo-8-fluoro-6-iodo-2-(methylthio)quinolin-4-yl)(tert-butoxycarbonyl)amino)-2-azabicyclo[2.1.1]hexane-2-carboxylate (36).** To a solution of the product from the previous step (51 g, 78 mmol) in DMF (200 mL) were added acrylonitrile (103 mL, 1.56 mol), tetramethylammonium formate (60.8 mL, 156 mmol) and DIPEA (27.3 mL, 156 mmol), followed by tetrakistriphenylphosphine palladium (20.3 g, 17.56 mmol). The reaction flask was evacuated, backfilled with nitrogen, then stirred under a nitrogen atmosphere at 50 \u00b0C overnight. The reaction was quenched with water and extracted with ethyl acetate. The organic layer was washed with water and brine, dried over sodium sulfate and concentrated. The crude product was purified by silica gel chromatography (0\u2212100% hexanes/ethyl acetate) to provide the desired product (18.1 g, 36% over 3 steps). LC-MS calculated for C28H36BrFN5O4S (M + H)+: m/z = 636.2; found 636.3.\n\n**Step 9. tert-Butyl (1R,4R,5S)-5-((7-Bromo-6-(2-cyanoethyl)-8-fluoro-3-iodo-2-(methylthio)quinolin-4-yl)amino)-2-azabicyclo[2.1.1]hexane-2-carboxylate (37).** To a solution of 36 (22 g, 34.6 mmol) in acetonitrile (135 mL) at \u221220 \u00b0C was added a solution of sulfuric acid (4.61 mL, 86 mmol) in water (3 mL). A solution of sodium nitrite (4.77 g, 69.1 mmol) in a minimum amount of water was next added dropwise, maintaining the temperature of the reaction mixture around \u221220 \u00b0C. After addition was complete, the reaction mixture was stirred at \u221210 \u00b0C for 5 min, then a concentrated solution of potassium iodide (22.95 g, 138 mmol) in water was added slowly, maintaining reaction temperature around \u221220 \u00b0C. After addition was complete, the reaction mixture was stirred at \u221210 \u00b0C for 45 min, then quenched with sat. sodium thiosulfate and extracted with ethyl acetate. The organic layer was washed with sat. sodium bicarbonate, sat. sodium thiosulfate, water and brine, dried over sodium sulfate and concentrated. The crude product was dissolved in DCM/TFA (1:1, 150 mL) and stirred at r.t. for 3 h, then concentrated and azeotroped with acetonitrile three times. The residue was dissolved in THF (150 mL) and triethylamine (48.2 mL, 346 mmol) and Boc-anhydride (7.54 g, 34.6 mmol) were added. The reaction mixture was stirred at r.t. for 1 h, then quenched with water and extracted with ethyl acetate. The organic layer was washed with water and brine, dried over sodium sulfate and concentrated. The product was purified by FCC (0\u221250% EtOAc/hexanes) to yield the title compound (10.7 g, 48% over 3 steps). LC-MS calculated for C23H26BrFIN4O2S (M + H)+: m/z = 647.0; found 647.0.\n\n**Step 10. tert-Butyl (1R,4R,5S)-5-(7-Bromo-8-(2-cyanoethyl)-6-fluoro-4-(methylthio)-2-((R)-1-(3-oxomorpholino)ethyl)-1H-pyrrolo[3,2-c]quinolin-1-yl)-2-azabicyclo[2.1.1]hexane-2-carboxylate (39).** A solution of 37 (3.6 g, 5.56 mmol), 38 (1.70 g, 11.12 mmol, see Supporting Information), DIPEA (9.7 mL, 55.6 mmol), Pd(PPh3)4 (1.28 g, 1.11 mmol), and copper(I) iodide (0.424 g, 2.22 mmol) in DMF (55 mL) was stirred at 70 \u00b0C for 1 h. Sat. aq. ammonium chloride was then added and the mixture was extracted with ethyl acetate. The organic layer was washed with water and brine, dried over sodium sulfate and concentrated. The crude product was dissolved in DMF (55 mL) and cesium carbonate (5.38 g, 16.5 mmol) was added and heating was continued at 70 \u00b0C for 1 h. The reaction was quenched with a small amount of concentrated aq. ammonium hydroxide, diluted with water and extracted with ethyl acetate. The organic layer was washed with water and brine, dried over sodium sulfate and concentrated. The crude product was purified by flash column chromatography (0\u2212100% ethyl acetate in DCM) to provide the desired product (3.25 g, 88%). LC-MS calculated for C31H36BrFN5O4S (M + H)+: m/z = 672.2; found 672.2.", "# Revolutionizing Protein Structure Determination: How MICA Combines Cryo-EM and AlphaFold3 for Unprecedented Accuracy\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe practical implications are profound. MICA has been successfully applied to newly released cryo-EM density maps from 2025, demonstrating its real-world applicability for automated, high-accuracy protein structure determination. This capability could accelerate drug discovery, enable rapid structural characterization of disease-related proteins, and facilitate large-scale structural genomics efforts.\n\nThe method\u2019s robustness to protein size and resolution makes it particularly valuable for challenging targets like membrane proteins, large macromolecular assemblies, and dynamic protein complexes that have historically been difficult to model accurately.\n\n## Future Directions and Implications\n\nMICA\u2019s success validates the power of multimodal deep learning in structural biology and opens new avenues for research. Future developments may incorporate symmetry constraints for multi-chain complexes, advanced side-chain prediction algorithms, and integration with other experimental techniques like nuclear magnetic resonance (NMR) spectroscopy.\n\nThe approach also demonstrates the broader potential of combining experimental and computational methods in biology, suggesting that the future of structural determination lies not in choosing between techniques, but in intelligently integrating them through sophisticated AI architectures.\n\nAs cryo-EM technology continues to advance and produce higher-resolution maps, MICA\u2019s performance is expected to improve further, potentially achieving near-perfect automated structure determination for a wide range of biological targets. This represents a significant step toward the long-sought goal of fully automated, high-accuracy protein structure modeling.", "# 2013 \u2013 Crystal structure of a soluble cleaved HIV-1 envelope trimer\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe glycan shield continues to be a focal point of structural and immunological investigation. Recent analyses have revealed the complex architecture arising from glycan-glycan interactions and the heterogeneous nature of glycan processing across different expression systems. Understanding how glycan holes and shield completeness influence neutralization breadth has provided valuable insights for immunogen design. The structural characterization of diverse broadly neutralizing antibodies has revealed multiple approaches for penetrating this protective barrier, informing strategies for eliciting such responses through vaccination.\n\nCurrent investigations are exploring the full spectrum of envelope conformations and their implications for vaccine design. The relationship between structural stability, antigenic presentation, and immune recognition continues to evolve as new techniques provide higher resolution insights into envelope dynamics. The integration of structural biology with advanced immunological analyses is revealing how to optimize trimeric immunogens for maximum efficacy while minimizing undesired immune responses. These efforts are contributing to a new generation of HIV-1 vaccine candidates that more faithfully represent the native viral spike and its sites of vulnerability.\n\n| Concept | Description | Key References |\n| --- | --- | --- |\n| BG505 SOSIP Trimer | Stabilized soluble envelope trimer with native-like conformation for structural and vaccine studies | [Julien, J.P., et al., Science, 2013](https://www.science.org/doi/abs/10.1126/science.1245625) |\n| Glycan Shield | Dense N-linked glycan array protecting envelope surface from antibody recognition | [Stewart-Jones, G.B., et al., PNAS, 2020](https://www.pnas.org/doi/10.1073/pnas.2000260117) |\n| MPER Structure | Membrane-proximal external region forming trimeric cluster essential for fusion and antibody targeting | [Fu, Q., et al., PNAS, 2018](https://www.pnas.org/doi/10.1073/pnas.1807259115) |\n| Broadly Neutralizing Antibodies | Rare antibodies targeting conserved epitopes across diverse HIV-1 strains | [Garces, F., et al., Nature, 2014](https://pmc.ncbi.nlm.nih.gov/articles/PMC4278586/) |\n| V1/V2 Apex Domain | Variable loop regions forming protective quaternary structure at trimer apex | [McLellan, J.S., et al., Nature, 2011](https://dukespace.lib.duke.edu/items/2d26edb5-c8f1-49ac-ab20-7bff2716ce41) |\n| mRNA Vaccine Platform | Nucleic acid delivery system enabling membrane-bound trimer presentation with improved immunogenicity | [Saunders, K.O., et al., Nature Medicine, 2025](https://pubmed.ncbi.nlm.nih.gov/40737434/) |", "# ChatMol \u2013 Natural-language control of PyMOL, describe the scene, get the picture\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n### Implicit (command-line box)\n\nJust type natural language into PyMOL\u2019s command line and press Enter:\n\n```\nPyMOL> load the protein in ~/Downloads/kinase.pdb and colour it by chain\nPyMOL> zoom in on the active site, residues 45 to 60 of chain A\nPyMOL> show the ligand as sticks coloured by element with green carbons\nPyMOL> make a high resolution image and save it to ~/Desktop/render.png\nPyMOL> mutate residue 100 of chain A to alanine\n```\n\nValid PyMOL/Python lines (e.g. `color red, chain A`, `zoom`, `@script.pml`,\n`util.cbc`, lines starting with `#`, `/`, `@`) are always passed straight\nthrough to PyMOL, untouched.\n\n### Explicit\n\nIf auto-detection is off, or you just prefer to be explicit:\n\n```\nPyMOL> ai colour everything by secondary structure\nPyMOL> chatmol show a surface around the binding pocket, transparent\n```\n\n### Multi-turn context\n\nChatMol keeps a short rolling history of your requests and the commands it\ngenerated, so follow-ups like *\u201cnow make it bigger\u201d* or *\u201cundo the surface,\nshow sticks instead\u201d* have context. Clear it with:\n\n```\nPyMOL> chatmol_reset\n```\n\n---\n\n## 4. Configuration\n\nAll settings are stored in `~/.chatmol/config.json` and can be viewed/changed\nwith `chatmol_config`, or via **Plugin > ChatMol Settings** if installed\nthrough the Plugin Manager.\n\n```\nPyMOL> chatmol_config                       # list everything\nPyMOL> chatmol_config model                 # show one value\nPyMOL> chatmol_config model, qwen3:0.6b     # change the model\nPyMOL> chatmol_config host, http://localhost:11434\nPyMOL> chatmol_config temperature, 0.1\nPyMOL> chatmol_config auto, off             # disable command-line auto-translation\nPyMOL> chatmol_config debug, on             # print raw LLM output (troubleshooting)\n```\n\n| Setting | Default | Meaning |\n| --- | --- | --- |\n| `host` | `http://localhost:11434` | Ollama server URL |\n| `model` | `qwen3.6:27b` | Ollama model tag (must be pulled, must match `ollama list`) |\n| `temperature` | `0.2` | LLM sampling temperature |\n| `timeout` | `120` | Request timeout (seconds) |\n| `auto` | `True` | Auto-route non-PyMOL command-line input through the LLM |\n| `echo` | `True` | Print each translated command before running it |\n| `debug` | `False` | Print the raw LLM response (before cleanup) to the console |\n| `max_history` | `6` | Number of past request/response turns kept for follow-ups |\n\n---\n\n## 5. How it works\n\n```\nchatmol/\n\u251c\u2500\u2500 __init__.py        plugin entry point: registers commands, installs hook\n\u251c\u2500\u2500 config.py          JSON-backed settings (~/.chatmol/config.json)\n\u251c\u2500\u2500 ollama_client.py   minimal urllib client for POST /api/chat\n\u251c\u2500\u2500 pymol_reference.py system prompt + curated PyMOL command cheat-sheet\n\u251c\u2500\u2500 nl_router.py       PyMOL-vs-English heuristic, cmd.do hook, translation\n\u251c\u2500\u2500 commands.py        `ai`, `chatmol_config`, `chatmol_reset` commands\n\u2514\u2500\u2500 gui.py             Plugin-menu settings dialog (Qt)\n```\n\n1. **Hooking the command line.** PyMOL\u2019s GUI command-line box (and\n   `@script.pml`, menu actions, etc.) all funnel through `cmd.do(line)`.\n   ChatMol saves the original `cmd.do` and replaces it with a wrapper.\n2. **Deciding what\u2019s \u201cPyMOL\u201d vs \u201cEnglish\u201d.** `nl_router.looks_like_pymol()`\n   checks whether the first token of the line is a known PyMOL command (in\n   `cmd.keyword`, which covers both built-ins and anything registered via\n   `cmd.extend`, e.g. ChatMol\u2019s own `chatmol_config`), a known namespace\n   (`util`, `pymol`, `_self`, \u2026), or a Python keyword. If the word\n   *following* the command looks like an English article/preposition/etc.\n   (e.g. \u201cshow **the** ligand\u2026\u201d) or the line contains connector words like\n   \u201cexcept\u201d/\u201dbut\u201d/\u201dwithout\u201d, it\u2019s treated as English instead. As a second\n   safety net, if a line doesn\u2019t look like PyMOL *or* English, it\u2019s passed\n   through unchanged so odd internal strings aren\u2019t sent to the LLM.\n3. **Translation.** Anything left over is sent to Ollama\n   (`ollama_client.chat`) with:\n4. a system prompt describing the required output format and a curated\n   PyMOL command reference (`pymol_reference.PYMOL_CHEAT_SHEET`),\n5. a snapshot of the current session (`cmd.get_names`, chains per\n   object), so \u201cthe protein\u201d / \u201cit\u201d / \u201cchain A\u201d resolve correctly,\n6. the last few turns of conversation for follow-up requests.\n\nThe model is instructed to output **only** PyMOL command lines (or a\n`#` comment if the request can\u2019t be done), one per line, and to pass\nalready-valid PyMOL syntax through unchanged \u2014 a safety net for cases\nthe heuristic misclassifies. Generation is capped (`num_predict`) and any\nresponse that comes back too long or repetitive is discarded with a\nfriendly message rather than executed or remembered, so a bad response\nfrom a small model can\u2019t snowball into garbage for every later request.\n\n1. **Execution.** Each returned line is echoed as `[ChatMol] > ...` and run\n   through the *original* `cmd.do`, so results are indistinguishable from\n   typing the command yourself (including normal PyMOL error messages if a\n   translated command is wrong).\n\n---\n\n## 6. Safety notes\n\n* ChatMol executes whatever the model returns. A bad translation could run\n  a destructive command (e.g. `delete all`, overwriting a file with `save`).\n  Review the echoed `[ChatMol] > ...` lines, especially for destructive\n  requests, and use `chatmol_config auto, off` + the explicit `ai` command\n  if you want a manual checkpoint before every translation.\n* Everything runs locally against your Ollama server \u2014 no data leaves your\n  machine.\n* `chatmol_config debug, on` is useful when a translation looks wrong: it\n  prints the model\u2019s raw output (including any stripped `<think>` blocks\n  or code fences) so you can see what was sent to PyMOL.\n\n---\n\n## 7. Troubleshooting\n\n**`Could not reach Ollama at http://localhost:11434 ...`**\nStart Ollama (`ollama serve` or open the desktop app) and confirm the model\nis pulled (`ollama pull qwen3.6:27b`, check with `ollama list`).", "# Approaches to automated protein crystal harvesting\nAuthors: Deller, M. C.; Rupp, B.\nSource: Acta Crystallographica Section F: Structural Biology Communications, 70, 133\u2013155 (IYCr crystallization series), 2014\n\n| Mechanism | Systems | Physical damage to crystal | Ease of harvesting | Ease of automation | Cost | Compatibility with data collection | High throughput | Ability to harvest microcrystals | Reduced late-stage failures |\n|---|---|---|---|---|---|---|---|---|---|\n| Microcapillary | PCGDS, CPPI, ACH, CARESS, CMM | + | +++ | +++ | + | + | ++ | +++ | +++ |\n| Microtool | UMR, CH, SAMI | +++ | + | ++ | + | +++ | + | + | + |\n| Microgripper | Chopsticks, REACH | ++ | +++ | ++ | +++ | + | ++ | +++ | +++ |\n| Acoustic droplet ejection | ADE | + | +++ | + | +++ | + | +++ | +++ | +++ |\n| Optical trap | COT and FOT | ++ | +++ | + | +++ | + | ++ | +++ | +++ |\n| MEMS | RodBot | + | +++ | +++ | + | +++ | +++ | +++ | +++ |\n| In situ | REACH | + | +++ | +++ | + | +++ | +++ | +++ | +++ |\n\n*Table 5.* Key numbers cited in the review.\n\n| Quantity | Value |\n|---|---|\n| Share of PDB structures from X-ray crystallography (January 2014) | 88% (85,486 of 96,692) |\n| HTPX/SG contribution to all PDB structures | over 10% |\n| HTPX/SG contribution to structures of families with no previous coverage | about 50% |\n| Crystals harvested and screened at the JCSG by January 2014 | just over 165,000, giving 1,470 structures |\n| Crystals harvested per deposited JCSG structure | about 100 |\n| Time for an expert to harvest 100 crystals manually | of the order of 4 h |\n| Estimated crystals screened worldwide per year | of the order of 700,000 |\n| Global cryopin market | about 2 million per year |\n| PDB growth rate at the time of writing | about 7,000 structures per year |\n| UMR robot step size | 1 \u00b5m |\n| REACH microgripper resolution and opening range | 1 \u00b5m; 0\u2013500 \u00b5m |\n| REACH harvesting size range demonstrated | 40\u2013400 \u00b5m |\n| COT / FOT crystal size ranges | ~10 \u00b5m / 5\u201330 \u00b5m |\n| COT vs FOT laser power | ~50 mW vs ~10 mW |\n| ADE droplet volume | 2.5 nl |\n| Structures from ADE-mounted slurries | insulin 1.9 \u00c5, lysozyme 1.8 \u00c5 |\n| CPV polyhedrin structure from COT-mounted crystals | 1.5 \u00c5 (versus 2.1 \u00c5 previously in the PDB) |\n| Lower theoretical crystal size for X-ray structure determination | about 1.2 \u00b5m |\n| Hyperquenching cooling rate | above 10,000\u00b0C s\u207b\u00b9 |\n| Cryoprotectant reduction demonstrated by hyperquenching | 22% |\n| JCSG crystals requiring added cryoprotectant | 80% |\n| UMR plate-tape aperture and drop lifetime gain | 3 mm; about tenfold |\n| SSRL grid capacity | 75 holes per grid; 1,200 crystals per UniPuck; 8,400 per shipping dewar |\n| Typical crystal-detection algorithm error rate | around 15% |\n| Sphere of confusion needed for a 10 \u00b5m crystal in a 10 \u00b5m beam | 2 \u00b5m or less |\n\n## Acknowledgements\n\nThe authors thanked members of the Joint Center for Structural Genomics for helpful advice and assistance. The JCSG was supported by the National Institutes of Health, National Institute of General Medical Sciences, Protein Structure Initiative (U54 GM094586). The UMR robotics system was supported by the NIH and NIGMS under STTR Phase II Grant No. R42 GM073278-02A1 and supplement. Thanks were also given to Professor Martin Caffrey (Trinity College Dublin), Professor Yuan Zheng (The Ohio State University), Jean-Luc Ferrer (NatX-ray) and colleagues at ETH for images, comments and discussion.\n\n## Key takeaways", "# GitHub README: ChemSage\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/ChemSage, 2026\n\n**Per-example pass/fail scoring:**\n- Changed the eval metric methodology throughout `eval_compare.py` from per-instance accuracy (`ok / total_attempts`) to per-example pass/fail (`passed_examples / 100`). An example \"passes\" a metric if every instance in that response is correct (`ok == tot`, `tot > 0`). This makes round-over-round comparisons meaningful: R1 calling 2 SMILES correctly is no longer the same score as R3 calling 100 SMILES correctly. The `_per_example_agg()` helper re-derives this from `results[].scores` in `raw_results.json`.\n\n**5-round comparative eval (complete, 2026-06-29):**\n- All 5 rounds evaluated on 100 shared R5 test examples (seed=42). Full report: `eval/compare/results/compare_20260629_1928.html`. R5 overall 95% (vs R4 80%, R3 87%). R5 wins on exec (+24%), PyExec (+30%), fidelity (+42%), SMARTS (+45%). All README eval tables updated.\n\n### Round 5: Higher rank + 2.5\u00d7 corpus (2026-06-28, training in progress)\n\n- Expanded SFT dataset from 8,000 to **20,000 examples** (16,000 train / 2,000 val / 2,000 test) across **78 behaviour classes** (19 new vs R4)\n- Raised LoRA rank to **64** (from 32); scale set to **90.0** (RSLoRA effective alpha \u2248 11.25, same order as R4); trainable parameters **doubled to 268M (0.82%)**\n- Extended `max_seq_length` to **3,072** (from 2,048); all corpus examples fit (p99=488, max=558 tokens)\n- Tripled training iters to **3,000**; doubled `val_batches` to 50 for a more stable eval signal\n- New generators targeting R4 exec/PyExec gap and numerical fidelity: `pyexec_drill` (\u00d76), `code_then_quote_v2` (\u00d75), `rounding_explicit` (\u00d75), `fidelity_multistep` (\u00d75)\n- New content generators: `herg_liability`, `selectivity_profile`, `prodrug_bcs`, `sar_delta`, `mdanalysis`, `dssp`, `ppi_interface`, `uniprot_api`, `conformer_3d`, `mcs_search`, `reaction_smarts`, `recap_fragmentation`, `electron_density`, `biological_assembly`, `drug_target_family` (\u00d73-4 each)\n- 707 examples rejected during validation (96.5% pass rate); token lengths all under max_seq_length (p99=488)\n- Training launched 2026-06-28 with `caffeinate -i` under single-process MLX launch (`/tmp/r5_launch.py`) \u2014 memory limits set in-process before importing mlx_lm; cache 16.7 GB / memory limit 50.1 GB\n- **Iter 1 val loss: 2.210** (lower than R4's 2.397 \u2014 rank=64 adapter has more capacity from the start)\n- Config: `config/train_config_r5.yaml`; adapter path: `adapters/chem_sage_32b_v5_lora/`\n- **Early stop at iter 2000** \u2014 val loss plateaued at 0.055 from iter 1600; fused from iter 1600 checkpoint into `models/chem_sage_32b_v5/` (~17.2 GB)\n\n### Round 4: Expanded dataset training (2026-06-25 to 2026-06-27, complete)\n\n- Expanded SFT dataset from 5,000 to 8,000 examples (6,400 train / 800 val / 800 test) across 59 behaviour classes\n- New behaviour classes: drug mechanisms/indications/ATC, multi-step reasoning, refusals, personality, structural biology, docking, visualisation (Plotly, R/ggplot2), numerical fidelity drills (`code_then_quote`, `rounding_precision`)\n- `steps_per_report` aligned with `steps_per_eval` (both 25) so train/val loss appear at the same iters\n- Training ran to completion: 1,500 iters, ~43 h (2026-06-25 15:39 to 2026-06-27 10:59 BST)\n- **Best val loss: 0.041 at iter 950** \u2014 beats R3's all-time best of 0.054 by **24%** (tied at iter 1425)\n- Peak memory 25.5 GB, 32 adapted layers, 134M trainable parameters (0.41%)\n- Fused from iter 950 checkpoint into `models/chem_sage_32b_v4/` (~17.2 GB, 4 safetensors shards)\n\n### Comparative eval harness (2026-06-26)\n\nNew script `eval/compare/eval_compare.py` runs all 4 model rounds side-by-side on the same 100 R4 test examples and produces a rich HTML + Markdown comparison report.\n\n**Features:**\n- **Option A auto-server**: spawns and kills `mlx_lm.server` per model automatically (no manual terminal juggling); health-checks the OpenAI-compatible endpoint before starting eval; sends SIGTERM to the process group on completion\n- **Shared examples**: 100 examples sampled from the R4 test set with a fixed seed (reproducible)\n- **All 13 metrics**: 3 core + 7 new metrics, all reporting `n/a` gracefully on older models\n- **R3 as baseline**: metric table includes a \u0394 column (\u00b1%) vs Round 3; colour-coded green/red\n- **Model profile panel**: full spec comparison across architecture, training config, training outcomes, corpus, and eval runtime\n- **Runtime resource stats**: server process CPU mean/peak %, RSS mean/peak GB (via `psutil`), mean latency (s), throughput (tok/s)\n- **Validation loss curves**: SVG comparison chart of all 4 models' val loss curves (clipped at 0.60; dashed line for estimated curves)\n- **Per-example table**: 100 collapsible rows; click \"responses\" to see all model outputs side-by-side\n- **Resume flag**: `--resume` loads `raw_results.json` cache and skips already-evaluated models\n- **Branding**: ChemSage ASCII art (pyfiglet), clickable [marcdeller.com](https://marcdeller.com) links throughout\n\n**Usage** (run after R4 training completes and fuse is done):\n```bash\ncd chem_sage/\n# Full run \u2014 all 4 models\npython eval/compare/eval_compare.py\n\n# Subset and resume\npython eval/compare/eval_compare.py --models round3 round4 --resume\n\n# Quick smoke-test (10 examples)\npython eval/compare/eval_compare.py --limit 10\n```\n\n**Config:** `eval/compare/models.yaml` \u2014 all model metadata, loss curves, and training stats in one place; edit to update R4 final values after training.\n\n### ChromaDB 1.x compatibility fix (2026-06-25)\n\n- ChromaDB 1.0+ switched to a Rust connection backend; `PersistentClient` timed out on databases created with the old Python backend\n- Fix: pass `Settings(allow_reset=True, anonymized_telemetry=False)` to `PersistentClient` in both `rag/retrieve.py` and `scripts/ingest_rag.py`\n- No data loss, no re-ingest \u2014 all 55,742 indexed chunks preserved; confirmed connected and serving queries\n- Downgrade to `chromadb<1.0.0` is not possible on Python 3.14 (no pre-built `chroma-hnswlib` wheel)", "# Revolutionary Bridge Between Prediction and Reality: How Elastic Networks Transform AlphaFold2\u2019s Structural Accuracy\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-14 on marcdeller.com (categories: blog, sb).*\n\n* [Flexible fitting of AlphaFold2-predicted models to cryo-EM density maps using elastic network models: a methodical affirmation](https://academic.oup.com/bioinformaticsadvances/article/5/1/vbae181/7903278?login=false)\n\n* **Masked map superiority**: Density maps excluding neighboring chain densities consistently improve refinement accuracy over box-cropped maps containing all experimental information\u200b\n* **Wide mode range advantages**: Including normal modes 1-12 outperforms narrower ranges by capturing both rigid-body motions and internal deformations essential for accurate fitting\u200b\n* **Local optimization efficiency**: Powell method matches global optimization accuracy while requiring significantly less computational time, suggesting smooth correlation landscapes\u200b\n* **Inner product robustness**: This similarity measure eliminates cropping biases and zero-padding dependencies while maintaining equivalent accuracy to Pearson correlation\u200b\n* **Systematic improvement validation**: Four challenging AlphaFold2 predictions showed TM-score improvements ranging from 0.11 to 0.17 units through elastic network refinement\u200b\n* **Biophysical motion preservation**: Elastic networks maintain protein fold integrity while enabling large-scale conformational adjustments impossible with rigid fitting approaches\n\n1. **Flexible fitting of AlphaFold2-predicted models to cryo-EM density maps using elastic network models: a methodical affirmation:** [Alshammari, M., et al., Bioinformatics Advances](https://academic.oup.com/bioinformaticsadvances/article/5/1/vbae181/7903278?login=false)\n2. **MDFF_NM: Improved Molecular Dynamics Flexible Fitting into Cryo-EM Maps Using Normal Mode Analysis:** [Ribeiro, J.V., et al., Journal of Chemical Information and Modeling](https://pubs.acs.org/doi/10.1021/acs.jcim.3c02007)\n3. **ModeHunter: A Package for the Reductionist Analysis, Animation, and Application of Elastic Biomolecular Motion:** [Wriggers, W., et al., Journal of Physical Chemistry B](https://pubs.acs.org/doi/abs/10.1021/acs.jpcb.4c05077)\n4. **Effective Molecular Dynamics from Neural Network-Based Structure Prediction Models:** [Bouchiat, C., et al., Journal of Chemical Theory and Computation](https://pubs.acs.org/doi/10.1021/acs.jctc.2c01027)\n5. **Predicting protein conformational motions using energetic frustration analysis and AlphaFold2:** [Chen, J., et al., Proceedings of the National Academy of Sciences](https://www.pnas.org/doi/10.1073/pnas.2410662121)\n6. **protein dynamic conformations modeling in the post-AlphaFold era:** [Zhang, Y., et al., Briefings in Bioinformatics](https://academic.oup.com/bib/article/26/4/bbaf340/8202937)\n\n**Unlocking the hidden potential of artificial intelligence through biophysical refinement**\n\nComputational structural biology stands at a transformative crossroads where deep learning predictions meet experimental validation. Recent breakthroughs demonstrate that AlphaFold2\u2019s remarkable prediction capabilities can be significantly enhanced through careful integration with cryo-electron microscopy data and elastic network models. This convergence represents more than incremental improvement\u2014it establishes a new paradigm for bridging the gap between computational prediction and experimental reality.\n\nThe challenge confronting structural biologists lies not in AlphaFold2\u2019s general accuracy, which remains exceptional, but in addressing the minority of predictions that exhibit significant deviations from experimental structures. When 18 out of 137 individual protein chains showed TM-scores below 0.80, researchers recognized an opportunity to develop targeted refinement strategies. These challenging cases, rather than representing failures, became laboratories for innovation.\n\nElastic network models emerge as the key to unlocking improved accuracy through their ability to capture biologically relevant protein motions. Unlike traditional molecular dynamics approaches that can become trapped in local energy minima, elastic networks provide a reduced-dimensional framework for exploring conformational space. The mathematical elegance lies in representing protein dynamics through normal modes\u2014eigenvectors of the Hessian matrix that describe collective atomic motions while maintaining structural integrity.\n\nThe breakthrough methodology employed systematic parameter optimization across multiple dimensions. Three distinct mode ranges were evaluated: modes 1-9, modes 7-9, and the comprehensive modes 1-12 approach. The widest range consistently produced superior results, demonstrating that including both rigid-body motions and higher-frequency internal deformations provides essential flexibility for accurate fitting. This finding challenges conventional wisdom that suggested focusing solely on low-frequency modes.\n\nMap preparation strategies proved equally crucial to refinement success. [Masked density maps](https://academic.oup.com/bioinformaticsadvances/article/5/1/vbae181/7903278?login=false), which exclude extraneous densities from neighboring chains, consistently outperformed box-cropped maps that retain all density information. This observation highlights the importance of data curation in computational structural biology\u2014sometimes less information enables more accurate results by reducing noise and spurious optimization targets.", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\n**Authors:** Xu Q, Bateman A, Finn RD, Abdubek P, Astakhova T, Axelrod HL, Bakolitsa C, Carlton D, Chen C, Chiu HJ, Chiu M, Clayton T, Das D, Deller MC, Duan L, Ellrott K, Ernst D, Farr CL, Feuerhelm J, Grant JC, Grzechnik A, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kozbial P, Krishna SS, Kumar A, Marciano D, McMullan D, Miller MD, Morse AT, Nigoghossian E, Nopakun A, Okach L, Puckett C, Reyes R, Rife CL, Sefcovic N, Tien HJ, Trame CB, van den Bedem H, Weekes D, Wooten T, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Journal of molecular biology (2010). DOI: 10.1016/j.jmb.2009.11.006. PMID: 19913036.\n\nPleckstrin homology (PH) domains have been identified only in eukaryotic proteins to date. We have determined crystal structures for three members of an uncharacterized protein family (Pfam PF08000), which provide compelling evidence for the existence of PH-like domains in bacteria (PHb). The first two structures contain a single PHb domain that forms a dome-shaped, oligomeric ring with C(5) symmetry. The third structure has an additional helical hairpin attached at the C-terminus and forms a similar but much larger ring with C(12) symmetry. Thus, both molecular assemblies exhibit rare, higher-order, cyclic symmetry but preserve a similar arrangement of their PHb domains, which gives rise to a conserved hydrophilic surface at the intersection of the beta-strands of adjacent protomers that likely mediates protein-protein interactions. As a result of these structures, additional families of PHb domains were identified, suggesting that PH domains are much more widespread than originally anticipated. Thus, rather than being a eukaryotic innovation, the PH domain superfamily appears to have existed before prokaryotes and eukaryotes diverged.\n\n## Conformational changes associated with the binding of zinc acetate at the putative active site of XcTcmJ, a cupin from Xanthomonas campestris pv. campestris (2010)\n\n**Authors:** Axelrod HL, Kozbial P, McMullan D, Krishna SS, Miller MD, Abdubek P, Acosta C, Astakhova T, Carlton D, Caruthers J, Chiu HJ, Clayton T, Deller MC, Duan L, Elias Y, Feuerhelm J, Grzechnik SK, Grant JC, Han GW, Jaroszewski L, Jin KK, Klock HE, Knuth MW, Kumar A, Marciano D, Morse AT, Murphy KD, Nigoghossian E, Okach L, Oommachen S, Paulsen J, Reyes R, Rife CL, Tien HJ, Trout CV, van den Bedem H, Weekes D, White A, Xu Q, Zubieta C, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Wilson IA.  \n**Journal:** Acta crystallographica. Section F, Structural biology and crystallization communications (2010). DOI: 10.1107/s1744309109021988. PMID: 20944231.\n\nIn the plant pathogen Xanthomonas campestris pv. campestris, the product of the tcmJ gene, XcTcmJ, encodes a protein belonging to the RmlC family of cupins. XcTcmJ was crystallized in a monoclinic space group (C2) in the presence of zinc acetate and the structure was determined to 1.6\u2005\u00c5 resolution. Previously, the apo structure has been reported in the absence of any bound metal ion [Chin et al. (2006), Proteins, 65, 1046-1050]. The most significant difference between the apo structure and the structure of XcTcmJ described here is a reorganization of the binding site for zinc acetate, which was most likely acquired from the crystallization solution. This site is located in the conserved metal ion-binding domain at the putative active site of XcTcmJ. In addition, an acetate was also bound within coordination distance of the zinc. In order to accommodate this binding, rearrangement of a conserved histidine ligand is required as well as several nearby residues within and around the putative active site. These observations indicate that binding of zinc serves a functional role in this cupin protein.\n\n## Crystal structure of the first eubacterial Mre11 nuclease reveals novel features that may discriminate substrates during DNA repair (2010)\n\n**Authors:** Das D, Moiani D, Axelrod HL, Miller MD, McMullan D, Jin KK, Abdubek P, Astakhova T, Burra P, Carlton D, Chiu HJ, Clayton T, Deller MC, Duan L, Ernst D, Feuerhelm J, Grant JC, Grzechnik A, Grzechnik SK, Han GW, Jaroszewski L, Klock HE, Knuth MW, Kozbial P, Krishna SS, Kumar A, Marciano D, Morse AT, Nigoghossian E, Okach L, Paulsen J, Reyes R, Rife CL, Sefcovic N, Tien HJ, Trame CB, van den Bedem H, Weekes D, Xu Q, Hodgson KO, Wooley J, Elsliger MA, Deacon AM, Godzik A, Lesley SA, Tainer JA, Wilson IA.  \n**Journal:** Journal of molecular biology (2010). DOI: 10.1016/j.jmb.2010.01.049. PMID: 20122942.", "# JCSG Top96 Screen Formulation\nAuthors: Deller, M. C.\nSource: Joint Center for Structural Genomics, The Scripps Research Institute (screen formulation sheet, 16 April 2013), 2013\n\n| Well | Precipitant class | Parent screen | JCSG Core Suite | JCSG Top96 | Crystallization solution | JCSG Cryo | PDB count |\n|---|---|---|---|---|---|---|---|\n| A01 | Organic | W2cryo #10 | JCSG-II F05 | A01 | 40% (v/v) 1,2-Propanediol, 0.1 M Sodium acetate pH 4.5 | None | 6 |\n| A02 | Organic | H2 #03 | JCSG-II H03 | A02 | 25% (v/v) Ethylene glycol | 10% EG | 5 |\n| A03 | Organic | MPD #01 | JCSG-II H11 | A03 | 10% (v/v) MPD, 0.1 M Citric acid pH 2.5, final pH 4 | 15% MPD | 4 |\n| A04 | Organic | MPD #02 | JCSG-II F07 | A04 | 10% (v/v) MPD, 0.1 M Sodium acetate pH 5.0, final pH 5 | 10% MPD | 5 |\n| A05 | Organic | MPD #16 | JCSG-III D02 | A05 | 40% (v/v) MPD, 0.1 M HEPES pH 6.5, final pH 7 | None | 5 |\n| A06 | Organic | MPD #24 | JCSG-IV B04 | A06 | 65% (v/v) MPD, 0.1 M Bicine pH 8.5, final pH 9 | None | 4 |\n| A07 | Organic/PEG/Salt | W1cryo #14 | JCSG-II G12 | A07 | 0.2 M Ammonium sulfate, 20% (v/v) PEG 300, 10% Glycerol, 0.1 M Phosphate-citrate pH 4.2 | None | 8 |\n| A08 | Organic/PEG/Salt | H1cryo #39 | JCSG-IV D10 | A08 | 1.7 M Ammonium sulfate, 1.7% (v/v) PEG 400, 15% (v/v) Glycerol, 0.085 M HEPES pH 7.5 | None | 8 |\n| A09 | Organic/PEG/Salt | H1cryo #20 | JCSG-II G08 | A09 | 0.16 M Ammonium sulfate, 20% (w/v) PEG 4000, 20% (v/v) Glycerol, 0.08 M Sodium acetate pH 4.6 | None | 6 |\n| A10 | Organic/PEG/Salt | H1cryo #09 | JCSG-III F11 | A10 | 0.17 M Ammonium acetate, 25.5% (w/v) PEG 4000, 15% (v/v) Glycerol, 0.085 M Sodium citrate pH 5.6 | None | 10 |\n| A11 | Organic/PEG/Salt | H1cryo #22 | JCSG-IV C03 | A11 | 0.17 M Sodium acetate, 25.5% (w/v) PEG 4000, 15% (v/v) Glycerol, 0.085 M Tris pH 8.5 | None | 12 |\n| A12 | Organic/PEG/Salt | H1cryo #18 | JCSG-III E07 | A12 | 0.16 M Magnesium acetate, 16% (w/v) PEG 8000, 20% (v/v) Glycerol, 0.08 M Sodium cacodylate pH 6.5 | None | 5 |\n| B01 | Organic/PEG | W2cryo #29 | JCSG-III C09 | B01 | 10% (v/v) Glycerol, 5% (w/v) PEG 3000, 30% (v/v) PEG 400, 0.1 M HEPES pH 7.5 | None | 7 |\n| B02 | Organic/PEG | W1cryo #16 | JCSG-IV F11 | B02 | 30% (v/v) PEG 600, 5% (w/v) PEG 1000, 10% (v/v) Glycerol, 0.1 M MES pH 6.0 | None | 7 |\n| B03 | Organic/PEG | H1cryo #37 | JCSG-II G06 | B03 | 5.6% (w/v) PEG 4000, 30% (v/v) Glycerol, 0.07 M Sodium acetate pH 4.6 | None | 4 |\n| B04 | Organic/PEG | H1cryo #41 | JCSG-II B11 | B04 | 17% (w/v) PEG 4000, 15% (v/v) Glycerol, 8.5% (v/v) Isopropanol, 0.085 M Sodium HEPES pH 7.5 | 10% EG | 5 |\n| B05 | Organic/PEG | H1cryo #40 | JCSG-I F06 | B05 | 19% (v/v) Isopropanol, 19% (w/v) PEG 4000, 5% (v/v) Glycerol, 0.095 M Sodium citrate pH 5.6 | None | 9 |\n| B06 | Organic/PEG | H1 #41 | JCSG-I B05 | B06 | 20% (w/v) PEG 4000, 10% (v/v) Isopropanol, 0.1 M HEPES pH 7.5 | 10% G | 12 |\n| B07 | Organic/PEG | H1 #40 | JCSG-I F07 | B07 | 20% (v/v) Isopropanol, 20% (w/v) PEG 4000, 0.1 M Sodium citrate pH 5.6 | 10% EG | 10 |\n| B08 | Organic/PEG | W2cryo #01 | JCSG-I C03+ | B08 | 5% (w/v) PEG 8000, 40% (v/v) MPD, 0.1 M Sodium cacodylate pH 6.5 | None | 7 |\n| B09 | Organic/Salt | H1cryo #47 | JCSG-II G05 | B09 | 1.6 M Ammonium sulfate, 20% (v/v) Glycerol, 0.08 M Sodium acetate pH 4.6 | None | 4 |\n| B10 | Organic/Salt | H2 #10 | JCSG-III G11 | B10 | 0.2 M Sodium chloride, 30% (v/v) MPD, 0.1 M Sodium acetate pH 4.6 | None\u2020 | 5 |\n| B11 | Organic/Salt | H1 #26 | JCSG-IV G08 | B11 | 0.2 M Ammonium acetate, 30% (v/v) MPD, 0.1 M Sodium citrate pH 5.6 | 5% MPD\u2020 | 4 |\n| B12 | Organic/Salt | W2 #02 | JCSG-I E08 | B12 | 0.2 M Lithium sulfate, 35% (v/v) MPD, 0.1 M MES pH 6.0 | None | 4 |\n| C01 | PEG | W2cryo #27 | JCSG-III D10 | C01 | 50% (v/v) PEG 200, 0.1 M Tris pH 7.0 | None | 4 |\n| C02 | PEG | W1cryo #46 | JCSG-III F01 | C02 | 40% (v/v) PEG 400, 5% (w/v) PEG 3000, 0.1 M MES pH 6.0 | None | 5 |\n| C03 | PEG | W2cryo #31 | JCSG-III A02 | C03 | 40% (v/v) PEG 600, 0.1 M CHES pH 9.5 | None | 4 |\n| C04 | PEG | W1 #06 | JCSG-I F08+ | C04 | 20% (w/v) PEG 3000, 0.1 M Sodium citrate pH 5.5 | 10% EG | 5 |\n| C05 | PEG | P6K #01 | JCSG-I H12 | C05 | 5% (w/v) PEG 6000, 0.1 M Citric acid pH 4.0, final pH 4 | 25% P200 | 4 |\n| C06 | PEG | P6K #14 | JCSG-I G04+ | C06 | 20% (w/v) PEG 6000, 0.1 M Citric acid pH 4.0, final pH 5 | 15% G | 6 |\n| C07 | PEG | P6K #18 | JCSG-I A02+ | C07 | 20% (w/v) PEG 6000, 0.1 M Bicine pH 8.5, final pH 9 | 10% EG | 4 |\n| C08 | PEG | P6K #21 | JCSG-IV G06 | C08 | 30% (w/v) PEG 6000, 0.1 M MES pH 5.0, final pH 6 | 10% P200 | 7 |\n| C09 | PEG | P6K #24 | JCSG-IV B03 | C09 | 30% (w/v) PEG 6000, 0.1 M Bicine pH 8.5, final pH 9 | 5% EG | 6 |\n| C10 | PEG | W1 #21 | JCSG-I B02 | C10 | 20% (w/v) PEG 8000, 0.1 M HEPES pH 7.5 | 10% EG | 6 |\n| C11 | PEG | W1 #01 | JCSG-I A01+ | C11 | 20% (w/v) PEG 8000, 0.1 M CHES pH 9.5 | 12% P200 | 5 |\n| C12 | Salt | AmSO4 #01 | JCSG-I H09+ | C12 | 0.8 M Ammonium sulfate, 0.1 M Citric acid pH 3.5, final pH 4 | 20% G | 5 |\n| D01 | Salt | AmSO4 #08 | JCSG-II F09 | D01 | 1.6 M Ammonium sulfate, 0.1 M Citric acid pH 4.0, final pH 5 | 20% G | 7 |\n| D02 | Salt | AmSO4 #09 | JCSG-IV G05 | D02 | 1.6 M Ammonium sulfate, 0.1 M MES pH 5.0, final pH 6 | 20% G | 6 |\n| D03 | Salt | AmSO4 #12 | JCSG-IV A10 | D03 | 1.6 M Ammonium sulfate, 0.1 M Bicine pH 9.0, final pH 9 | 15% G | 4 |\n| D04 | Salt | W2 #04 | JCSG-II D10+ | D04 | 0.2 M Sodium chloride, 2.0 M Ammonium sulfate, 0.1 M Sodium cacodylate pH 6.5 | 15% EG | 7 |\n| D05 | Salt | H1 #04 | JCSG-II B01 | D05 | 2.0 M Ammonium sulfate, 0.1 M Tris pH 8.5 | 15% EG | 7 |\n| D06 | Salt | W1 #33 | JCSG-IV A01 | D06 | 0.2 M Lithium sulfate, 2.0 M Ammonium sulfate, 0.1 M CAPS pH 10.5 | 15% EG | 5 |\n| D07 | Salt | AmSO4 #17 | JCSG-III B10 | D07 | 2.4 M Ammonium sulfate, 0.1 M Tris pH 8.5, final pH 8 | 15% G | 5 |\n| D08 | Salt | AmSO4 #18 | JCSG-IV A12 | D08 | 2.4 M Ammonium sulfate, 0.1 M Bicine pH 9.0, final pH 9 | 20% G | 9 |\n| D09 | Salt | W2 #38 | JCSG-III H03 | D09 | 0.2 M Lithium sulfate, 2.5 M Sodium chloride, 0.1 M Sodium acetate pH 4.5 | 10% EG | 4 |\n| D10 | Salt | W1 #14 | JCSG-III D12+ | D10 | 1.0 M Sodium citrate, 0.1 M Sodium cacodylate pH 6.5 | 15% G | 8 |\n| D11 | Salt | W2 #31 | JCSG-III C10 | D11 | 0.2 M Sodium chloride, 1.0 M Sodium citrate, 0.1 M Tris pH 7.0 | 10% EG | 6 |\n| D12 | Salt | W2 #16 | JCSG-II A03 | D12 | 1.0 M Sodium citrate, 0.1 M CHES pH 9.5 | 10% EG | 4 |\n| E01 | Salt | H1 #38 | JCSG-III C05 | E01 | 1.4 M tri-Sodium citrate, 0.1 M Sodium HEPES pH 7.5 | 10% EG | 10 |\n| E02 | Salt | W1 #27 | JCSG-IV A02\u0394 | E02 | 0.2 M Lithium sulfate, 1.2 M Sodium dihydrogen phosphate/0.8 M di-Potassium hydrogen phosphate, 0.1 M CAPS pH 10.5 | 25% G | 5 |\n| E03 | Salt/PEG | W2cryo #48 | JCSG-I F09 | E03 | 0.2 M Sodium chloride, 50% (v/v) PEG 200, 0.1 M Phosphate-citrate pH 4.2 | None | 5 |\n| E04 | Salt/PEG | W1cryo #37 | JCSG-II C07+ | E04 | 0.2 M Calcium acetate, 40% (v/v) PEG 300, 0.1 M Sodium cacodylate pH 6.5 | None | 4 |\n| E05 | Salt/PEG | H1 #14 | JCSG-III C06 | E05 | 0.2 M Calcium chloride, 28% (v/v) PEG 400, 0.1 M Sodium HEPES pH 7.5 | 10% EG | 6 |\n| E06 | Salt/PEG | W1 #44 | JCSG-II H01 | E06 | 0.2 M Calcium acetate, 30% (v/v) PEG 400, 0.1 M Sodium acetate pH 4.5 | None | 4 |\n| E07 | Salt/PEG | W1cryo #38 | JCSG-II B02+ | E07 | 0.2 M Lithium sulfate, 40% (v/v) PEG 400, 0.1 M Tris pH 8.5 | 5% P400 | 4 |\n| E08 | Salt/PEG | W1cryo #05 | JCSG-II G03 | E08 | 0.2 M Magnesium chloride, 40% (v/v) PEG 400, 0.1 M MES pH 5.5 | None | 4 |\n| E09 | Salt/PEG | W2 #27 | JCSG-II D08 | E09 | 0.2 M Magnesium chloride, 10% (w/v) PEG 3000, 0.1 M Sodium cacodylate pH 6.5 | 15% EG | 5 |\n| E10 | Salt/PEG | W1 #28 | JCSG-I B01 | E10 | 0.2 M Sodium chloride, 20% (w/v) PEG 3000, 0.1 M HEPES pH 7.5 | 10% EG | 4 |\n| E11 | Salt/PEG | W2 #07 | JCSG-IV E07 | E11 | 0.2 M Sodium chloride, 30% (w/v) PEG 3000, 0.1 M Tris pH 7.0 | None | 6 |\n| E12 | Salt/PEG | PEG/ion #03 | JCSG-I E05 | E12 | 0.2 M Ammonium fluoride, 20% (w/v) PEG 3350 | 15% EG | 6 |\n| F01 | Salt/PEG | PEG/ion #23 | JCSG-I D09+ | F01 | 0.2 M Ammonium formate, 20% (w/v) PEG 3350 | 10% EG | 5 |\n| F02 | Salt/PEG | PEG/ion #35 | JCSG-I E09 | F02 | 0.2 M Ammonium sulfate, 20% (w/v) PEG 3350 | 10% EG | 5 |\n| F03 | Salt/PEG | PEG/ion #28 | JCSG-I B08 | F03 | 0.2 M Calcium acetate hydrate, 20% (w/v) PEG 3350 | 10% EG | 14 |\n| F04 | Salt/PEG | PEG/ion #07 | JCSG-III G05 | F04 | 0.2 M Calcium chloride, 20% (w/v) PEG 3350 | 10% EG | 9 |\n| F05 | Salt/PEG | PEG/ion #44 | JCSG-III B11 | F05 | 0.2 M di-Ammonium hydrogen phosphate, 20% (w/v) PEG 3350 | 10% EG | 9 |\n| F06 | Salt/PEG | PEG/ion #38 | JCSG-I D07 | F06 | 0.2 M di-Ammonium tartrate, 20% (w/v) PEG 3350 | 15% EG | 10 |\n| F07 | Salt/PEG | PEG/ion #40 | JCSG-III A07 | F07 | 0.2 M di-Sodium hydrogen phosphate, 20% (w/v) PEG 3350 | 15% EG | 5 |\n| F08 | Salt/PEG | PEG/ion #25 | JCSG-I A12 | F08 | 20% (w/v) PEG 3350, 0.2 M Magnesium acetate | 10% EG | 12 |\n| F09 | Salt/PEG | PEG/ion #05 | JCSG-I F05 | F09 | 0.2 M Magnesium chloride, 20% (w/v) PEG 3350 | 10% EG | 11 |\n| F10 | Salt/PEG | PEG/ion #20 | JCSG-I F03+ | F10 | 0.2 M Magnesium formate, 20% (w/v) PEG 3350 | 10% EG | 5 |\n| F11 | Salt/PEG | PEG/ion #16 | JCSG-I F04 | F11 | 0.2 M Magnesium nitrate, 20% (w/v) PEG 3350 | 15% EG | 6 |\n| F12 | Salt/PEG | PEG/ion #22 | JCSG-I B09+ | F12 | 0.2 M Potassium formate, 20% (w/v) PEG 3350 | 10% EG | 5 |\n| G01 | Salt/PEG | PEG/ion #18 | JCSG-I C11+ | G01 | 0.2 M Potassium nitrate, 20% (w/v) PEG 3350 | 10% EG | 12 |\n| G02 | Salt/PEG | PEG/ion #27 | JCSG-I C08 | G02 | 0.2 M Sodium acetate, 20% (w/v) PEG 3350 | 10% P200 | 4 |\n| G03 | Salt/PEG | PEG/ion #01 | JCSG-II C05 | G03 | 0.2 M Sodium fluoride, 20% (w/v) PEG 3350 | 15% P200 | 6 |\n| G04 | Salt/PEG | PEG/ion #21 | JCSG-I B11 | G04 | 0.2 M Sodium formate, 20% (w/v) PEG 3350 | 10% EG | 6 |\n| G05 | Salt/PEG | PEG/ion #17 | JCSG-II D02 | G05 | 0.2 M Sodium nitrate, 20% (w/v) PEG 3350 | 10% EG | 5 |\n| G06 | Salt/PEG | PEG/ion #13 | JCSG-I C12+ | G06 | 0.2 M Sodium thiocyanate, 20% (w/v) PEG 3350 | 10% EG | 6 |\n| G07 | Salt/PEG | PEG/ion #48 | JCSG-I G01\u0394 | G07 | 20% (w/v) PEG 3350, 0.2 M di-Ammonium citrate | 10% EG | 7 |\n| G08 | Salt/PEG | PEG/ion #47 | JCSG-I A07+ | G08 | 20% (w/v) PEG 3350, 0.2 M tri-Potassium citrate | 10% EG | 7 |\n| G09 | Salt/PEG | H1 #20 | JCSG-I G11 | G09 | 0.2 M Ammonium sulfate, 25% (w/v) PEG 4000, 0.1 M Sodium acetate pH 4.6 | 10% EG | 6 |\n| G10 | Salt/PEG | H1 #09 | JCSG-II E10 | G10 | 0.2 M Ammonium acetate, 30% (w/v) PEG 4000, 0.1 M Sodium citrate pH 5.6 | None | 10 |\n| G11 | Salt/PEG | H1 #17 | JCSG-IV B11 | G11 | 0.2 M Lithium sulfate, 30% (w/v) PEG 4000, 0.1 M Tris pH 8.5 | 10% EG | 8 |\n| G12 | Salt/PEG | H1 #06 | JCSG-IV B09 | G12 | 0.2 M Magnesium chloride, 30% (w/v) PEG 4000, 0.1 M Tris pH 8.5 | None | 13 |\n| H01 | Salt/PEG | H1 #22 | JCSG-III B05 | H01 | 0.2 M Sodium acetate, 30% (w/v) PEG 4000, 0.1 M Tris pH 8.5 | None | 9 |\n| H02 | Salt/PEG | P6K/LiCl #08 | JCSG-I G03 | H02 | 1.0 M Lithium chloride, 10% (w/v) PEG 6000, 0.1 M Citric acid pH 5.0, final pH 5 | 15% EG | 5 |\n| H03 | Salt/PEG | P6K/LiCl #11 | JCSG-IV C12 | H03 | 1.0 M Lithium chloride, 10% (w/v) PEG 6000, 0.1 M Tris pH 8.5, final pH 8 | 15% EG | 5 |\n| H04 | Salt/PEG | P6K/LiCl #16 | JCSG-III D07 | H04 | 1.0 M Lithium chloride, 20% (w/v) PEG 6000, 0.1 M HEPES pH 7.0, final pH 7 | 10% EG | 8 |\n| H05 | Salt/PEG | P6K/LiCl #15 | JCSG-II E03 | H05 | 1.0 M Lithium chloride, 20% (w/v) PEG 6000, 0.1 M MES pH 6.0, final pH 6 | 20% G | 10 |\n| H06 | Salt/PEG | P6K/LiCl #14 | JCSG-II F11 | H06 | 1.0 M Lithium chloride, 20% (w/v) PEG 6000, 0.1 M Citric acid pH 5.0, final pH 5 | 20% G | 4 |\n| H07 | Salt/PEG | W1 #46 | JCSG-II B03 | H07 | 0.2 M Calcium acetate, 10% (w/v) PEG 8000, 0.1 M Imidazole pH 8.0 | 20% EG | 9 |\n| H08 | Salt/PEG | W2 #28 | JCSG-I E07 | H08 | 0.2 M Calcium acetate, 20% (w/v) PEG 8000, 0.1 M MES pH 6.0 | 15% G | 9 |\n| H09 | Salt/PEG | H1 #18 | JCSG-I D12 | H09 | 0.2 M Magnesium acetate, 20% (w/v) PEG 8000, 0.1 M Sodium cacodylate pH 6.5 | 15% EG | 10 |\n| H10 | Salt/PEG | W2 #03 | JCSG-III B01 | H10 | 0.2 M Magnesium chloride, 20% (w/v) PEG 8000, 0.1 M Tris pH 8.5 | 10% P200 | 12 |\n| H11 | Salt/PEG | W2 #39 | JCSG-II A01 | H11 | 0.2 M Sodium chloride, 20% (w/v) PEG 8000, 0.1 M CAPS pH 10.5 | 10% EG | 7 |\n| H12 | Salt/PEG | H1 #28 | JCSG-III E09 | H12 | 0.2 M Sodium acetate, 30% (w/v) PEG 8000, 0.1 M Sodium cacodylate pH 6.5 | 15% EG | 5 |", "# Structure\u2013Function Analysis of the Extended Conformation of a Polyketide Synthase Module\nAuthors: Li, X.; Sevillano, N.; La Greca, F.; Deis, L.; Liu, Y.-C.; Deller, M.C.; Mathews, I.I.; Matsui, T.; Cane, D.E.; Craik, C.S.; Khosla, C.\nSource: Journal of the American Chemical Society, 2018\n\nChain translocation kinetics were measured for module 2 + TE with its natural diketide (NDK) substrate, (2S,3R)-2-methyl-3-hydroxypentanoyl-ACP1.[16] As shown in Figure 6A, both the kcat and KM values were comparable in the presence and absence of 1B2 (kcat = 19 \u00b1 1 and 28 \u00b1 3 min\u22121 in the presence and absence of 1B2, respectively; KM = 52 \u00b1 4 and 69 \u00b1 16 \u03bcM in the presence and absence of 1B2, respectively). To quantify the chain elongation kinetics for the same module, an alternative assay was used involving a diffusible mimic of the same substrate.[17] Again, as shown in Figure 6B, both the kcat and KM values were comparable in the presence and absence of 1B2 (kcat = 4.1 \u00b1 0.2 and 4.3 \u00b1 0.2 min\u22121 in the presence and absence of 1B2, respectively; KM = 1.6 \u00b1 0.2 and 2.0 \u00b1 0.3 mM in the presence and absence of 1B2, respectively). Trapping module 2 + TE in its extended conformation by addition of excess 1B2 had no significant adverse effect on the ability of this PKS module to catalyze either chain elongation or chain translocation. We therefore conclude that the extended KS-AT conformation (Figure 2A) is fully competent for both chain translocation and chain elongation, while the alternative arch-shaped conformation is not required for either activity. The data in Figure 6 also suggest that neither AT-catalyzed extender unit transacylation (step II, Figure 3) nor KR-catalyzed \u03b2-ketoreduction (lumped into step III, Figure 3) requires access to the arch-shaped conformation, although neither transacylation nor ketoreduction is rate-limiting in either assay.\n\n*Figure 6.* Effect of 1B2 binding on chain translocation and chain elongation by DEBS module 2 + TE. 1B2 binds tightly to DEBS module 2 + TE. The initial slope of the graph in (A) reflects conditions where chain translocation is rate-limiting, whereas the maximum velocities in (A) and (B) reflect conditions where elongation is rate-limiting. Blue curves show the activity without 1B2; red curves show the activity in the presence of 1.2 equiv of 1B2.\n\nTo directly correlate the above findings from crystallography and activity assays of 1B2-bound PKS modules, we performed SEC in tandem with SAXS on the 1B2/module 2 complex under solution conditions corresponding to high specific activity of this PKS module.[18] Samples of the complex were freshly prepared in three of the four states shown in Figure 3 (states A, B, and D) and immediately subjected to SEC\u2013SAXS analysis using methods very similar to those reported for DEBS module 3.[5] The data from each SEC peak fraction were compared to theoretical scattering curves generated by CRYSOL[19] based on the two alternative docking models shown in Figure 7A. In every case, the experimental data correlated much more strongly to the theoretically predicted curve from the extended model (\u03c7\u00b2 = 0.55) than to the curve from the arched model (\u03c7\u00b2 = 7.74) (Figure 7B,C). These findings strongly reinforce the conclusion that module 2 predominantly assumes the extended conformation in the preacylation state (state A, Figure 3), the acylated state (state B, Figure 3), and the postelongation state (state D, Figure 3), establishing that conformational reorganization of the KS, AT, and KR domains is not required during catalysis.\n\n*Figure 7.* Tandem (SEC\u2013SAXS) analysis of the 1B2/module 2 complex. (A) Docking models of 1B2 bound to each conformation from Figure 2 were generated by aligning the X-ray structure of the 1B2/KS-AT complex (Figure 5) with each module conformation. The Fab complexed to the extended conformation is shown at the top, whereas the arched conformation is shown below. Domain coloring is the same as in Figures 2 and 5. (B) Comparisons of (left) log(I) vs q and (right) Kratky plots generated by SEC\u2013SAXS analysis of the 1B2/module 2 complex (black curve) and CRYSOL analysis of the extended (blue curve) and arched (red curve) models from (A). (C) Comparison of (left) log(I) vs q and (right) Kratky plots from SEC\u2013SAXS analysis of 1B2/module 2 trapped in states A (black curve), B (green curve), and D (purple curve) (cf. Figure 3).\n\nIn summary, we have investigated the structure and dynamics of modules from a prototypical assembly-line PKS using a combination of protein structural and functional approaches. The simplest model for polyketide biosynthesis, based on the well-established model for vertebrate fatty acid biosynthesis,[20] posits that the catalytic domains of individual modules assemble into relatively rigid structures, with interdomain and intermodular motions being primarily restricted to acyl-ACP species. Alternatively, PKS modules might undergo larger conformational changes along their reaction coordinates, such as interconversions between the extended and the arched protein conformations shown in Figure 2. To test these alternative hypotheses, we identified a specific Fab that stabilizes a PKS module in an extended conformation. The targeted discovery of the specific Fab antibody 1B2 combined with crystallographic, SEC\u2013SAXS, and kinetic analyses of 1B2/PKS complexes establish that the extended conformation of the KS-AT didomains of both DEBS modules 2 and 3 is catalytically competent for all of the essential PKS reactions shown in Figure 3. The stage is now set for a deeper mechanistic understanding of long-distance translocations of growing polyketide chains between successive PKS modules.\n\n## Acknowledgments\n\nThis research was supported by grants from the National Institutes of Health (GM087934 to C.K., P41CA196276 and GM104659 to C.S.C., and GM022172 to D.E.C.). X-ray analysis was performed on beamlines 11-1 and 12-2 at the Stanford Synchrotron Radiation Lightsource (SSRL), SLAC National Accelerator Laboratory. Use of the SSRL, SLAC National Accelerator Laboratory, is supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences, under Contract DE-AC02-76SF00515. The SSRL Structural Molecular Biology Program is supported by the DOE Office of Biological and Environmental Research and by the National Institutes of Health, National Institute of General Medical Sciences (including P41GM103393).\n\n## References\n\n*(reference list omitted)*", "# Structural Studies of Cytokines and Cytokine Receptors \u2014 Part 2: Expression, Purification and Characterisation of Oncostatin M (Chapter 3) and of LIFR (Chapter 4)\nAuthors: Deller, M. C.\nSource: D.Phil. thesis, University of Oxford (Laboratory of Molecular Biophysics / Wellcome Trust Centre for Human Genetics); supervisor Prof. E. Yvonne Jones, 1999\n\n| Chemical class | Component | Mass (g) |\n|---|---|---|\n| Amino acid | Alanine | 2.00 |\n| Amino acid | Asparagine | 2.32 |\n| Amino acid | Aspartic acid | 1.60 |\n| Amino acid | Cystine | 0.12 |\n| Amino acid | Glutamic acid | 2.68 |\n| Amino acid | Glutamine | 1.32 |\n| Amino acid | Glycine | 2.16 |\n| Amino acid | Histidine | 0.24 |\n| Amino acid | Isoleucine | 0.92 |\n| Amino acid | Leucine | 0.92 |\n| Amino acid | Lysine-HCl | 1.68 |\n| Amino acid | Phenylalanine | 0.52 |\n| Amino acid | Proline | 0.40 |\n| Amino acid | Serine | 8.32 |\n| Amino acid | Threonine | 0.92 |\n| Amino acid | Tyrosine | 0.68 |\n| Amino acid | Valine | 0.92 |\n| Nucleotide bases | Adenine | 2.00 |\n| Nucleotide bases | Guanosine | 2.68 |\n| Nucleotide bases | Thymine | 0.68 |\n| Nucleotide bases | Uracil | 2.00 |\n| Salts/Buffers | Sodium acetate | 6.00 |\n| Salts/Buffers | Succinic acid | 6.00 |\n| Salts/Buffers | Ammonium | 3.00 |\n| Salts/Buffers | Sodium hydroxide | 4.32 |\n| Salts/Buffers | Dibasic potassium hydroxide | 42.0 |\n\n*Table 3-2.* Table showing chemical composition of the non-autoclavable and autoclavable portions of LeMasters defined medium.\n\n### 3.2.8 Receptor Competition and Titration Assay\n\nAssays were carried out as represented in Figure 3-2 below. MaxiSorp\u2122 96 well plates (Nunc) were coated with 2\u20135 \u03bcg/well of Protein-A (Pharmacia). The plates were covered and left for 16 hours at room temperature. The Protein-A was discarded and the plate then blocked with PBS containing 1 % (w/v) BSA, for 2 hours, followed by two washes with PBS containing 0.05 % (w/v) Tween-20. Supernatants containing the receptor of interest, as Fc-fusions, were diluted 1:4 in PBS and 200 \u03bcl added to each well. The plate was left for 2 hours and after discarding the plate contents, a further two washes with PBS containing 0.05 % (w/v) Tween-20 were carried out.\n\nBinding reactions were set up using serial 4-fold dilutions of stock OSM_\u0394188, starting at a saturating concentration of 10000 ng/ml. OSM-196 and murine LIF were included as controls. Dilutions were carried out using a stock solution of PBS containing biotinylated ligand (kindly provided by K. Chobotova, CRC Growth Factors Group, Birmingham) as the competing agent. Biotinylated human OSM-196 and biotinylated human LIF were routinely used for competition. The plate was left at room temperature for 1 hour to equilibrate. The plate contents were discarded and the plate was then washed twice with PBS containing 0.05 % Tween-20. Table 3-3 below shows the receptor-ligand combinations tested.\n\n| Receptor | Ligand: [Human OSM-196] (ng/ml) | Competing Ligand: [Human Biotinylated OSM-196] (ng/ml) |\n|---|---|---|\n| Human gp130 | 10000 \u2013 0.6 | 50 |\n| Murine LIFR | 10000 \u2013 0.6 | 400 |\n\n*Table 3-3.* Table showing receptor and ligand concentrations used for titration and competition assays.\n\nThe plate was washed with PBS and streptavidin-horseradish peroxidase conjugate (Amersham\u2122), diluted 1:1000 in PBS, was added to each well and the plate incubated for 1 hour at room temperature. After washing each well with PBS, substrate development was carried out using 1,2-phenylenediamine (OPD) (Sigma) as chromogen. 8 mg of OPD was dissolved in 12 ml of citric acid phosphate buffer consisting of 0.035 M citric acid, 0.067 M Na2HPO4\u00b712H2O, pH 5.0 (see Appendix II). 5 \u03bcl of hydrogen peroxide was added. Chromogen solution was added to each well and the colour developed, in the dark, for 10\u201315 minutes. The reaction was terminated by the addition of 75 \u03bcl of 1 M sulphuric acid to each well. The absorbance was read at 490 nm.\n\nReceptor titration assays were carried out as for the competition assays above, except for the omission of competing biotinylated ligand. Binding reactions were set up using serial 4-fold dilutions of stock biotinylated human OSM-196, starting at a saturating concentration of 10000 ng/ml. All receptor supernatants were used at a dilution of 1:4 in PBS and murine LIFR, murine OSMR and human FGFR2 were included as controls.\n\n*Figure 3-2.* Figure showing schematic diagram of format used for receptor competition assays. Receptor titration assays were assembled in a similar fashion with the omission of competing biotinylated-OSM.\n\n### 3.2.9 Ba/F3 Cell Survival Assay\n\nBa/F3 cells were cultured in RPMI medium (Gibco) supplemented with 10 % dialysed foetal calf serum (Gibco), 50 IU/ml penicillin and 50 \u03bcg/ml streptomycin, glutamine and 2 \u03bcg/ml murine IL-3. Cultures were grown at 37 \u00b0C in a humidified atmosphere of 5 % CO2 in air. Cells were maintained at a density of between 5 \u00d7 10^4 and 1 \u00d7 10^6 cells/ml. All operations involving cell-lines were carried out aseptically, in a class II microbiological flow cabinet.\n\nStock samples of test cytokine were produced by dilution in minimal RPMI medium (as above with the exception of murine IL-3). OSM_\u0394188 was diluted to 500 ng/ml. Control samples of human OSM-196 and murine LIF were prepared in a similar fashion at concentrations of 500 ng/ml and 50 ng/ml, respectively. All samples were centrifuged at 15000 \u00d7 g for 15 minutes at 4 \u00b0C. Serial 4-fold dilutions of each cytokine were prepared, by dilution in minimal RPMI medium as above, and 50 \u03bcl of each was transferred aseptically to the wells of a 96 well culture plate. Duplicate wells containing minimal RPMI with and without IL-3 were included as positive and negative controls, respectively.\n\nCells were harvested by centrifugation at 1000 \u00d7 g for 5 minutes and carefully resuspended in 50 ml of minimal RPMI medium, to remove all traces of IL-3. This wash step was repeated 4 times and the cells then resuspended in 5 ml of minimal RPMI medium supplemented with 10 % FCS. Viable cells were counted in the presence of trypan blue dye using a haemacytometer. The culture was diluted to a density of 2 \u00d7 10^5 cells/ml and 50 \u03bcl of culture pipetted into each well. The plate was covered with clingfilm to limit evaporation and then incubated at 37 \u00b0C in a humidified atmosphere of 5 % CO2 in air, for 72 hours.", "# Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design\nAuthors: Irimia, A.; Serra, A.M.; Sarkar, A.; Jacak, R.; Kalyuzhniy, O.; Sok, D.; Saye-Francisco, K.L.; Schiffner, T.; Tingle, R.; Kubitz, M.; Adachi, Y.; Stanfield, R.L.; Deller, M.C.; Burton, D.R.; Schief, W.R.; Wilson, I.A.\nSource: PLoS Pathogens, 2017\n\nOur combined design, structural and functional study has provided an explanation for how two extremely broad MPER antibodies engage their common epitopes at the stem of the gp41 ectodomain. The location and conformation of MPER with regard to gp41 and the membrane at different stages of viral fusion remains unclear, but the information presented here helps to fill in missing pieces of the dynamic viral fusion process. These structural and functional insights are important for design of therapeutic antibodies and immunogens as HIV vaccines. The information here can enable the MPER to be linked to lipids in an appropriate orientation as in liposomes or chimeric viruses for vaccination and provide information for improving the pharmacodynamic and pharmacokinetic properties of 10E8 for therapeutic applications.\n\n## Materials and methods\n\n### Protein expression and purification\n\nGenes for 10E8 IgG and Fabs mutants were synthesized by Genscript, Inc. All antibodies and Fabs were expressed in FreeStyle 293S cells (Invitrogen) and purified as described previously. Briefly about 400 \u03bcg heavy-chain and 200 \u03bcg light-chain vectors were diluted in 25 ml Gibco Opti-MEM I (Invitrogen) reduced-serum medium, sterile filtered, and mixed with 25 ml final volume Opti-MEM I pre-incubated with 1 ml of 293fectin (Invitrogen). After 30 minutes incubation, the mixture was added to 1 L of cells (about 1.2 \u00d7 10^6 cells/ml density) in FreeStyle 293 expression medium (Invitrogen). The Fabs were purified on a lambda light chain CaptureSelect affinity column pre-equilibrated with 1\u00d7 PBS buffer. The unbound material was washed out with the same buffer and the bound Fabs were eluted with 0.1 M glycine, pH 3.0, and immediately neutralized using Tris pH 8.0. Fractions were concentrated and buffer exchanged into 20 mM sodium acetate, pH 5.5, then loaded on a Mono S column (GE Healthcare Life Sciences) equilibrated with the same buffer. The proteins were eluted with a linear gradient of 0 to 50%, 1 M KCl in 20 mM sodium acetate, pH 5.6. The concentrated samples were stored in 1\u00d7 HBS (150 mM NaCl, 10 mM HEPES, pH 7.4).\n\nT117v2 was expressed and purified as previously described. The purified 10E8 variants were incubated in a 1:1 molar ratio with T117v2 scaffold and purified by size exclusion chromatography using a HiLoad 16/600 SuperDex 200pg column (GE Healthcare Life Sciences) in 1\u00d7 HBS buffer.\n\n### Surface plasmon resonance experiments\n\nKinetics and affinities of antibody-antigen interactions were measured on a ProteOn XPR36 (Bio-Rad) using GLC Sensor Chip (Bio-Rad) and 1\u00d7 HBS-EP+ pH 7.4 running buffer (20\u00d7 stock from Teknova, Cat. No H8022) supplemented with BSA at 1 mg/ml. The Human Antibody Capture Kit instructions (Cat. No BR-1008-39 from GE Healthcare Life Sciences) were used to prepare chip surfaces for ligand capture. In a typical experiment, about 6000 RU of capture antibody was amine-coupled in all six flow cells of the GLC Chip. Regeneration was accomplished using 3 M magnesium chloride with 180 seconds contact time and injected four times per each cycle. Raw sensorgrams were analyzed using ProteOn Manager software (Bio-Rad), including interspot and column double referencing, and either Equilibrium or Kinetic fits with Langmuir model, or both, were employed when applicable. Analyte concentrations were measured on a NanoDrop 2000c Spectrophotometer using absorption at 280 nm.\n\n### Neutralization\n\nPseudoviruses were generated by transfection of 293T cells (ATCC) with an HIV-1 Env expressing plasmid and an Env-deficient genomic backbone plasmid (pSG3\u0394Env, NIH AIDS Reagent Program 11051), as described previously. Pseudoviruses were harvested 72 hours post-transfection for use in neutralization assays. Neutralizing activity was assessed using a single round of replication pseudovirus assay and TZM-bl target cell (NIH AIDS Reagent Program 8129). TZM-bl cells were seeded at a density of 5,000 cells/well in half-volume white luminescent 96-well plates (Costar 3688), one day prior to assay. Assay and growing medium was Complete DMEM [Dulbecco's Modified Eagle Medium (Corning Cellgro MT15013CV) with 200 \u03bcM L-glutamine (Gibco 25030081), 100 U/ml Penicillin-Streptomycin (Invitrogen 15140\u2013122), and ten percent fetal bovine serum (Thermo Scientific HyClone SH3091003)]. To this plate was added pseudovirus, which was preincubated with serial dilutions of antibody for 1 hour at 37\u00b0C in duplicate with 25 \u03bcl per well final volume. Virus-infected (no serum) and uninfected cell wells were controls on each cell plate. After 24 hours, 75 \u03bcl of Complete DMEM was added to each well, bringing the total volume to 100 \u03bcl; the plates were replaced in the incubator another 48 hours. Prior to virus signal determination, the liquid medium was removed from the plates, cells were lysed with 45 \u03bcl per well Promega Cell Lysis buffer (Product number E1531), and the plates were then shaken for 10 min at 1000 RPM on a Jitterbug Microplate Incubator/Shaker. Thirty \u03bcl of Promega Flash substrate (Promega Luciferase 1000 Assay System E4550) was added per well, and luminescence was measured via Synergy 2 Multi-Mode Reader (BioTek).\n\n### Size exclusion chromatography-multi-angle light scattering\n\nSECMALS analysis was performed by separating approximately 50 \u03bcg IgG on a Superdex 200 Increase 10/300 GL column (GE Healthcare Life Sciences) in phosphate buffered saline (140 mM NaCl, 2.7 mM KCl, 5.6 mM Na2HPO4, 1.8 mM KH2PO4, pH 7.4) and measuring UV absorption at 280 nm and multi-angle light scattering on a DAWN HELEOS II system with Optilab T-rEX refractometer (Wyatt Technology). Raw values were background subtracted and normalized to the maximum signal intensity of each injection. Molecular weights were calculated using the ASTRA6 software package (Wyatt Technology). Data were plotted using GraphPad Prism version 7.0a for Mac (GraphPad Software).\n\n### Protein crystallization", "# AlphaFraud \u2013 Weekly watcher: newly deposited human PDB structures vs their blind AlphaFold predictions \u2014 a running test of whether protein folding is solved\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-07-21 on marcdeller.com (categories: vibe-code).*\n\n> **Catch AlphaFold where the fold is wrong: weekly, automatically, on freshly deposited human structures.**\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [alphafraud.mdeller.com](https://alphafraud.mdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/AlphaFraud](https://github.com/bellcheddar/AlphaFraud) |\n\n---\n\nAlphaFraud watches the RCSB PDB for newly deposited human protein structures, matches each one to its blind AlphaFold prediction from the AlphaFold Structure Database (EBI), superposes them, and quantifies exactly where the prediction disagrees with the experiment. AlphaFold2 was trained only on PDB chains released before 2018-04-30, so any human structure deposited after that date is something the model predicted without ever seeing its coordinates: the ideal test set.\n\n**Why it matters:** the claim that \u201cAlphaFold has solved protein folding\u201d is only testable on structures the model never trained on, and those arrive continuously as the PDB grows. AlphaFraud turns that stream into a standing experiment: every week it surfaces the structures AlphaFold got most wrong, and ranks the \u201cconfidently wrong\u201d cases (high pLDDT, low agreement) where the model was sure and still missed. It is useful for: structural biologists auditing AlphaFold reliability, method developers looking for hard cases, and anyone testing whether predicted models are safe to build on.\n\n**A live instance runs at [alphafraud.mdeller.com](https://alphafraud.mdeller.com)**, having backfilled the entire post-cutoff archive (96,827 human protein entities, 82,796 analysed) and updating every week. The landing page is a cumulative dashboard across every processed week; the worst offenders so far are a roster of amyloid-forming proteins (transthyretin, \u03b22-microglobulin, islet amyloid polypeptide) that AlphaFold predicts as confidently folded but which crystallise or cryo-EM in a completely different aggregated form.\n\n## \u2728 Features\n\n| Capability | What it does |\n| --- | --- |\n| Automated PDB watch | Weekly systemd timer fires after the PDB\u2019s Wednesday 00:00 UTC release; resumes from the last processed run |\n| Human, post-cutoff filter | RCSB Search for taxonomy 9606 protein entities deposited after AlphaFold\u2019s 2018-04-30 training cutoff |\n| Blind-model matching | Maps each entity to its UniProt accession and AlphaFold DB model, handling fragments, mutants and missing loops via sequence alignment |\n| Sequence-novelty score | Max percent identity to any pre-cutoff PDB chain, so genuinely unseen sequences are flagged (not just easy homologs) |\n| Full metric suite | Global, local, backbone, per-domain and confidence-calibration metrics (see below) |\n| Confidence audit | pLDDT vs actual lDDT calibration, and a PAE-honesty check comparing AlphaFold\u2019s self-reported error to the observed error |\n| Two-tier archive backfill | A fast TM-score screen across every structure, running the full metric suite only on the disagreements; makes the whole ~96k-entity archive tractable |\n| Structural imagery | Every worst offender is drawn as a deviation-coloured C\u03b1 ribbon (the experimental structure, coloured residue-by-residue by its distance from the AlphaFold model, on an absolute-\u00c5ngstr\u00f6m scale) on the leaderboard, entry page and weekly highlights; an interactive 3Dmol.js viewer with an optional translucent AlphaFold \u201cghost\u201d overlay; hover-preview thumbnails on the scatter; and a \u201cdivergence\u201d ribbon banner in the header |\n| On-demand Calculate tab | Type any human PDB ID and the full pipeline runs on demand (fetch AlphaFold model, superpose, score, render) to return the auto-example panel, with strict autocomplete (exists + human + has an AlphaFold model) and no LLM |\n| Home & live Statistics | A Home navigation page outlining every tab, and an exhaustive Statistics page grouping all KPIs, coverage, runs, server (CPU/memory/uptime), storage/database and API stats into themed panels |\n| Cumulative dashboard | Default landing page aggregating every processed week: the \u201cfraud quadrant\u201d scatter, metric histograms, a weekly trend, and a browsable per-week / per-structure archive |\n| Branded, mobile-responsive report | Flask app with the signature scatter, heatmaps, per-domain tables and an all-time leaderboard; every plot has a CSV export and a plain-language explanatory panel that defines each metric twice over (a lay sentence and the underlying equation), and every table exports to CSV |\n| One-command deploy | Provisioning, deploy and one-shot release scripts for a DigitalOcean droplet (gunicorn, nginx, certbot TLS) |\n\n## \ud83c\udfd7\ufe0f Architecture and workflow\n\nOne controller (`AlphaFraud.py`, subcommands) drives a small Python package; a Flask app serves the results. Two runtimes share one SQLite file: a **systemd timer** runs the weekly pipeline (writes the database and the ribbon/coords assets), and an always-on **gunicorn** process serves the site by reading that same database live.", "# PDB structures deposited by Marc C. Deller\nAuthors: Deller, M.C. and co-depositors\nSource: RCSB Protein Data Bank, 2026\n\nThis list holds the 78 Protein Data Bank entries on which Marc C. Deller is a deposition author or an author of the primary citation, retrieved live from the RCSB PDB on 07 September 2026. It excludes the several hundred structural-genomics entries deposited under the collective 'Joint Center for Structural Genomics (JCSG)' author name during his years as JCSG Target Manager (2006\u20132015); across his career he has contributed to more than 400 PDB depositions. A browsable list is at https://marcdeller.com/structures/.\n\n## Entries (newest release first)", "# GitHub README: AlphaFraud\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/AlphaFraud, 2026\n\n> **Catch AlphaFold where the fold is wrong: weekly, automatically, on freshly deposited human structures.**\n\n[](https://alphafraud.mdeller.com)     [](https://doi.org/10.5281/zenodo.21808182)  \n\n<table>\n<tr>\n<td>\ud83c\udf10 <b>Website</b></td><td><a href=\"https://alphafraud.mdeller.com\" target=\"_blank\" rel=\"noopener noreferrer\">alphafraud.mdeller.com</a></td>\n<td>\u2709\ufe0f <b>Contact</b></td><td><a href=\"mailto:marc@marcdeller.com\">marc@marcdeller.com</a></td>\n<td>\ud83d\udc19 <b>GitHub</b></td><td><a href=\"https://github.com/bellcheddar/AlphaFraud\" target=\"_blank\" rel=\"noopener noreferrer\">bellcheddar/AlphaFraud</a></td>\n</tr>\n</table>\n\n---\n\nAlphaFraud watches the RCSB PDB for newly deposited human protein structures, matches each one to its blind AlphaFold prediction from the AlphaFold Structure Database (EBI), superposes them, and quantifies exactly where the prediction disagrees with the experiment. AlphaFold2 was trained only on PDB chains released before 2018-04-30, so any human structure deposited after that date is something the model predicted without ever seeing its coordinates: the ideal test set.\n\n**Why it matters:** the claim that \"AlphaFold has solved protein folding\" is only testable on structures the model never trained on, and those arrive continuously as the PDB grows. AlphaFraud turns that stream into a standing experiment: every week it surfaces the structures AlphaFold got most wrong, and ranks the \"confidently wrong\" cases (high pLDDT, low agreement) where the model was sure and still missed. It is useful for: structural biologists auditing AlphaFold reliability, method developers looking for hard cases, and anyone testing whether predicted models are safe to build on.\n\n**A live instance runs at [alphafraud.mdeller.com](https://alphafraud.mdeller.com)**, having backfilled the entire post-cutoff archive (96,827 human protein entities, 82,796 analysed) and updating every week. The landing page is a cumulative dashboard across every processed week; the worst offenders so far are a roster of amyloid-forming proteins (transthyretin, \u03b22-microglobulin, islet amyloid polypeptide) that AlphaFold predicts as confidently folded but which crystallise or cryo-EM in a completely different aggregated form.\n\n## \u2728 Features\n\n| Capability | What it does |\n|---|---|\n| Automated PDB watch | Weekly systemd timer fires after the PDB's Wednesday 00:00 UTC release; resumes from the last processed run |\n| Human, post-cutoff filter | RCSB Search for taxonomy 9606 protein entities deposited after AlphaFold's 2018-04-30 training cutoff |\n| Blind-model matching | Maps each entity to its UniProt accession and AlphaFold DB model, handling fragments, mutants and missing loops via sequence alignment |\n| Sequence-novelty score | Max percent identity to any pre-cutoff PDB chain, so genuinely unseen sequences are flagged (not just easy homologs) |\n| Full metric suite | Global, local, backbone, per-domain and confidence-calibration metrics (see below) |\n| Confidence audit | pLDDT vs actual lDDT calibration, and a PAE-honesty check comparing AlphaFold's self-reported error to the observed error |\n| Two-tier archive backfill | A fast TM-score screen across every structure, running the full metric suite only on the disagreements; makes the whole ~96k-entity archive tractable |\n| Structural imagery | Every worst offender is drawn as a deviation-coloured C\u03b1 ribbon (the experimental structure, coloured residue-by-residue by its distance from the AlphaFold model, on an absolute-\u00c5ngstr\u00f6m scale) on the leaderboard, entry page and weekly highlights; an interactive 3Dmol.js viewer with an optional translucent AlphaFold \"ghost\" overlay; hover-preview thumbnails on the scatter; and a \"divergence\" ribbon banner in the header |\n| On-demand Calculate tab | Type any human PDB ID and the full pipeline runs on demand (fetch AlphaFold model, superpose, score, render) to return the auto-example panel, with strict autocomplete (exists + human + has an AlphaFold model) and no LLM |\n| Home & live Statistics | A Home navigation page outlining every tab, and an exhaustive Statistics page grouping all KPIs, coverage, runs, server (CPU/memory/uptime), storage/database and API stats into themed panels |\n| Cumulative dashboard | Default landing page aggregating every processed week: the \"fraud quadrant\" scatter, metric histograms, a weekly trend, and a browsable per-week / per-structure archive |\n| Branded, mobile-responsive report | Flask app with the signature scatter, heatmaps, per-domain tables and an all-time leaderboard; every plot has a CSV export and a plain-language explanatory panel that defines each metric twice over (a lay sentence and the underlying equation), and every table exports to CSV |\n| One-command deploy | Provisioning, deploy and one-shot release scripts for a DigitalOcean droplet (gunicorn, nginx, certbot TLS) |\n\n## \ud83c\udfd7\ufe0f Architecture and workflow\n\nOne controller (`AlphaFraud.py`, subcommands) drives a small Python package; a Flask app serves the results. Two runtimes share one SQLite file: a **systemd timer** runs the weekly pipeline (writes the database and the ribbon/coords assets), and an always-on **gunicorn** process serves the site by reading that same database live.", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\n**Experiment 3.** The amino acid sequence of recombinant MBP-mutCALR Type 2 fusion protein used in Experiment 3 is given as SEQ ID NO:331. Recombinant MBP-mutCALR Type 2 fusion protein was custom made at Cepter. Data for the interaction of antibodies with recombinant MBP-mutCALR Type 2 fusion protein were collected as described in Experiment 2. The chip used was a Biacore Series S Sensor Chip CM4 with anti-huFc immobilized onto the chip surface. Recombinant MBP-mutCALR Type 2 fusion protein was prepared at nominal concentrations of 1.1, 3.3, 10, 30, and 90 nM. Typical capture levels were in the range of 16-37 RU. The results are shown in Table 8.\n\n*Table 8.* Binding affinities and kinetic association and dissociation rate constants to mutCALR for the indicated antibodies.\n\n| Clone No. | ka (1/Ms) | kd (1/s) | KD (M) |\n|---|---|---|---|\n| 162 | 6.07E+05 | 6.82E-03 | 1.12E-08 |\n| 74 | 8.54E+05 | 5.79E-03 | 6.78E-09 |\n| 65 | 6.68E+05 | 4.24E-03 | 6.35E-09 |\n\n### Example 3: Binding of Anti-mutCALR Antibodies to Cell Surface MPL and mutCALR\n\nTo assess binding of CALR antibodies to cell surface CALR, parental BaF3 (DSMZ) cells or BaF3 cells stably expressing only human MPL or both human MPL (Uniprot Number: P40238-1) and human mutCALR Type 1 and Type 2 were used. On the day of the assay, cells were washed and resuspended in assay buffer. Approximately 200,000 cells/well were added to 96-well plates and stained with the indicated concentration of antibodies for 30 minutes on ice. Cells were then washed and stained with goat anti-human secondary conjugated to R-phycoerythrin (R-PE) (Jackson Immuno Research Laboratories) for 30 minutes on ice. The cells were then washed and analyzed by flow cytometry. Geometric Mean Fluorescence Intensity (GMFI) of cell binding was graphed and EC50, hill slope, and area under the curve (AUC) were determined after four parameter curve fitting using GraphPad Prism Software (version 7.04).\n\n*Table 9.* EC50 values determined by in vitro binding studies using MPL-mutCALR cells. (Representative rows; the full table lists 53 clones. EC50 values for the best parental clones are in the 0.01-0.05 \u03bcg/mL range.)\n\n| Clone No. | MPL-mutCALR EC50 (\u03bcg/mL) | MPL-mutCALR Hill Slope | MPL-mutCALR AUC |\n|---|---|---|---|\n| 38 | 1.44 | 0.45 | 36975 |\n| 1 | 0.01 | 0.47 | 52843 |\n| 2 | 0.01 | 0.38 | 60393 |\n| 5 | 0.09 | 0.31 | 57069 |\n| 6 | 0.30 | 0.22 | 58250 |\n| 32 | 0.01 | 0.57 | 52533 |\n| 36 | 0.01 | 0.67 | 73813 |\n| 16 | 0.04 | 0.35 | 63157 |\n\n*Table 10.* EC50 values for variant clones determined by in vitro binding studies using MPL-mutCALR cells. (Representative rows; the affinity-matured clones give EC50 values roughly two to three orders of magnitude lower than the parental clones, in the 0.0004-0.02 \u03bcg/mL range.)\n\n| Clone No. | MPL-mutCALR EC50 (\u03bcg/mL) | MPL-mutCALR Hill Slope | MPL-mutCALR AUC |\n|---|---|---|---|\n| 64 | 0.0072 | 0.73 | 73166 |\n| 70 | 0.0019 | 0.50 | 84756 |\n| 55 | 0.0101 | 0.59 | 73079 |\n| 68 | 0.0033 | 0.49 | 68742 |\n| 74 | 0.0022 | 0.60 | 81764 |\n| 65 | 0.0059 | 0.48 | 81192 |\n| 132 | 0.0004 | 0.45 | 73424 |\n| 133 | 0.0017 | 0.39 | 85219 |\n| 131 | 0.0011 | 0.47 | 58328 |\n\n### Example 4: Anti-mutCALR Antibody Mediated Inhibition of pSTAT5\n\nTo test the ability of anti-mutCALR antibodies to inhibit phosphorylation of STAT5, Ba/F3 cells (DSMZ) expressing MPL and Type 1 mutCALR variants were generated by nucleofection (Amaxa Cell Line Nucleofection Kit V, Lonza, Basel, Switzerland) and cultured in RPMI 1640 + 10% FBS + selection antibiotics. Prior (24 hours) to pSTAT5 assessment, cells were cultured in selection-free media and then plated at 200,000 cells per well (96 well plate) in RPMI 1640, 10% FBS. Antibodies were added to the cells and incubated for 2 hours followed by cell lysing and quantification of pSTAT-5 levels by MSD (Phospho-STAT5a,b Whole Cell Lysate Kit, MSD, Kenilworth, NJ). Anti-mutCALR antibodies inhibit phosphorylation of STAT5 in a dose-dependent manner (FIG. 1).\n\n*Table 11.* IC50 values for anti-mutCALR antibody mediated inhibition of pSTAT5. (Representative rows; the full table lists 53 clones. The most potent clones give IC50 values of ~0.02-0.06 \u03bcg/mL.)\n\n| Clone No. | IC50 (\u03bcg/mL) |\n|---|---|\n| 54 | 12.7 |\n| 38 | 3.55 |\n| 1 | 0.47 |\n| 2 | 0.12 |\n| 4 | 0.04 |\n| 5 | 0.03 |\n| 6 | 0.03 |\n| 20 | 0.018 |\n| 3 | 0.021 |\n| 37 | 0.039 |\n| 15 | 0.041 |\n| 16 | 0.028 |\n| 17 | 0.031 |\n| 33 | 80.4 |\n\n### Example 5: Anti-mutCALR Antibody Mediated Inhibition of Cell Proliferation\n\nTo test the ability of anti-mutCALR antibodies to inhibit cell proliferation, engineered Ba/F3 cells transfected with MPL and mutCALR Type 1 were plated at 5,000 cells per well in RPMI 1640 + 2% FBS, antibodies added, incubated for 72 hours, and followed by assessment of cell viability using the CellTiter-Glo Luminescent Cell Viability Assay (Promega, Madison, WI) and Top Count (Perkin Elmer, Boston, MA) or Pherastar (BMG Labtech, Ortenberg, Germany) for luminescence quantification. Anti-mutCALR antibodies inhibit mutCALR-induced oncogenic cell proliferation in a dose-dependent manner in both the Ba/F3 engineered cells (FIG. 2).\n\n*Table 12.* IC50 values for anti-mutCALR antibody mediated inhibition of cell proliferation (Ba/F3 engineered cells) obtained after high-throughput primary screening. (Representative rows; the affinity-matured clones 55, 65, 68, 74, 128, 134, 170, 184 and 188 reach IC50 values of 0.0003-0.0098 \u03bcg/mL.)\n\n| Clone No. | IC50 (\u03bcg/mL) |\n|---|---|\n| 38 | 1.9070 |\n| 4 | 0.0022 |\n| 5 | 0.0049 |\n| 6 | 0.0032 |\n| 12 | 0.0043 |\n| 15 | 0.0037 |\n| 25 | 0.0035 |\n| 29 | ~0.002898 |\n| 55 | 0.0011 |\n| 68 | 0.0009 |\n| 74 | 0.0010 |\n| 170 | 0.0019 |\n| 65 | 0.0003 |\n| 124 | 0.0040 |\n| 134 | 0.0036 |\n| 128 | 0.0098 |\n| 184 | 0.0019 |\n| 188 | 0.0046 |\n\nND, not determined.\n\n### Example 6: Anti-mutCALR Antibody inhibits oncogenic cell proliferation induced by both Type 1 and Type 2 CALR mutations", "# ChemSage \u2013 A hybrid, chemically aware LLM for drug discovery: RAG for facts, a QLoRA-tuned model for behavior\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n| Iter | R1 Val | R2 Val | R3 Val | R4 Val | R5 Val | Notes |\n| --- | --- | --- | --- | --- | --- | --- |\n| 1 | 3.800 | 3.191 | 2.110 | 2.397 | **2.210** | Baseline; R5 lower than R4 despite 2.5\u00d7 data \u2014 rank=64 adapter fits better |\n| 25 | 1.020 | 0.580 | 0.238 | 0.480 | \u2014 |  |\n| 50 | 0.740 | 0.533 | 0.145 | 0.277 | \u2014 | R3 already below R2\u2019s best |\n| 75 | 0.630 | 0.527 | 0.122 | 0.200 | \u2014 |  |\n| 100 | 0.580 | 0.495 | 0.110 | 0.146 | \u2014 | R3 battery shutdown; resumed from here |\n| 200 | 0.490 | 0.449 | 0.116 | 0.109 | \u2014 |  |\n| 300 | 0.440 | 0.416 | 0.076 | 0.082 | \u2014 |  |\n| 400 | 0.420 | 0.400 | 0.059 | 0.074 | \u2014 |  |\n| 500 | 0.405 | 0.375 | 0.060 | 0.071 | \u2014 |  |\n| 600 | 0.395 | 0.359 | 0.060 | 0.059 | \u2014 | R4 matches R3\u2019s best |\n| 625 | \u2014 | \u2014 | **0.054** | 0.055 | \u2014 | **Best R3 checkpoint (fused)** |\n| 650 | \u2014 | \u2014 | \u2014 | 0.052 | \u2014 | R4 beats R3 best |\n| 675 | **0.389** | **0.347** | \u2014 | 0.069 | \u2014 | Best R1/R2 checkpoint |\n| 700 | 0.390 | 0.372 | \u2014 | 0.058 | \u2014 |  |\n| 750 | \u2014 | \u2014 | \u2014 | 0.056 | \u2014 |  |\n| 775 | \u2014 | \u2014 | \u2014 | 0.049 | \u2014 |  |\n| 800 | \u2014 | \u2014 | \u2014 | 0.051 | \u2014 |  |\n| 950 | \u2014 | \u2014 | \u2014 | **0.041** | \u2014 | **Best R4 checkpoint (fused)** |\n| 1000 | \u2014 | \u2014 | \u2014 | 0.043 | \u2014 |  |\n| 1100 | \u2014 | \u2014 | \u2014 | 0.043 | \u2014 |  |\n| 1425 | \u2014 | \u2014 | \u2014 | **0.041** | \u2014 | Ties R4 best at iter 950 |\n| 1500 | \u2014 | \u2014 | \u2014 | 0.050 | \u2014 | R4 final iter |\n| 1600 | \u2014 | \u2014 | \u2014 | \u2014 | **0.055** | **Best R5 checkpoint (fused)** |\n| 2000 | \u2014 | \u2014 | \u2014 | \u2014 | 0.061 | R5 early stop |\n\nRound 3 val loss (**0.054**) is **6.4x better** than Round 2 (0.347) and **7.2x better** than Round 1 (0.389). The improvement is driven by dataset quality (19 behaviour classes with code-heavy examples) and 3.3x more training data (4,000 vs 1,200 examples). Round 3 was stopped early at iter 625 of 1,500 after val loss plateaued.\n\nRound 4 **completed 2026-06-27** (1,500 iters, ~43 h). Best val loss: **0.041** at iter 950, beating R3\u2019s best of 0.054 by **24%**. The best was also tied at iter 1,425. Fused from iter 950 checkpoint into `models/chem_sage_32b_v4/` (~17.2 GB). Peak memory 25.5 GB, 32 adapted layers, 134M trainable parameters (0.41%).\n\nRound 5 **early-stopped 2026-06-29** at iter 2000 of 3,000. Val loss plateaued at 0.055 from iter 1600 (true best 0.054 at iters 1650 and 1950, but fell between checkpoint saves). Fused from iter 1600 checkpoint into `models/chem_sage_32b_v5/` (~17.2 GB). Higher floor than R4 (0.055 vs 0.041) reflects the harder task: 78 behaviour classes vs 59, 20k examples vs 6.4k.\n\n### \ud83e\uddea Evaluation results\n\n#### Original scorecard eval (per-round, `eval/eval_chem.py`)\n\n| Metric | Round 1 (7B) | Round 2 (32B) | Round 3 (32B) |\n| --- | --- | --- | --- |\n| SMILES validity | 0/0 (n/a) | 25/25 (100%) | **494/499 (99%)** |\n| Tool executability | 0/8 (0%) | 72/174 (41%) | **437/442 (99%)** |\n| Numerical fidelity | 0/0 (n/a) | 7/73 (10%) | **363/533 (68%)** |\n| Code blocks generated | 8 | 174 | 442 |\n| **Overall** | **0%** | **51%** | **89%** |\n\nScorecard files: `eval/scorecards/scorecard_r1_7b.html`, `scorecard_r2_32b.html`, `scorecard_r3_32b_v2.html`.\n\n#### Comparative eval \u2014 all 5 rounds (2026-06-29, `eval/compare/eval_compare.py`)\n\n100 shared R5 test examples (seed=42). Full report: `eval/compare/results/compare_20260629_1928.html`.\n\nScores: per-example pass/fail (examples where all instances correct / 100). Metrics degrade gracefully to `N/A` when a metric\u2019s trigger type never appears in a model\u2019s responses.\n\n| Metric | Round 1 | Round 2 | Round 3 | Round 4 | **Round 5** | \u0394 vs R4 |\n| --- | --- | --- | --- | --- | --- | --- |\n| SMILES validity | 100% | 100% | 100% | 100% | **100%** | +0% |\n| SMARTS validity | N/A | 0% | 100% | 55% | **100%** | +45% |\n| Tool executability | 0% | 28% | 79% | 71% | **95%** | +24% |\n| Code attempted | 14% | 41% | 100% | 97% | **100%** | +3% |\n| Python exec (extended) | 36% | 34% | 78% | 69% | **99%** | +30% |\n| Code\u2192Quote fidelity | N/A | 0% | 47% | 19% | **61%** | +42% |\n| Numerical fidelity | N/A | 18% | 57% | 47% | **89%** | +42% |\n| Rounding precision | 100% | 100% | 98% | 98% | **99%** | +1% |\n| Refusal accuracy | 98% | 98% | 97% | 98% | **100%** | +2% |\n| QED range | 100% | 100% | 100% | 100% | **100%** | +0% |\n| PDB ID validity | 100% | 100% | 100% | 100% | **100%** | +0% |\n| PyMOL syntax | 77% | 97% | 89% | 89% | **90%** | +1% |\n| Degeneration-free | 93% | 96% | 91% | 98% | **100%** | +2% |\n| **Overall** | **72%** | **63%** | **87%** | **80%** | **95%** | **+15%** |\n\nR5 wins on all exec/fidelity targets: **exec +24%**, **PyExec +30%**, **fidelity +42%**, **Code\u2192Quote +42%**, **SMARTS +45%**, **Degeneration-free 100%**, **Refusal 100%**.\n\nThe R4 exec regression vs R3 (71% vs 79%) was caused by dataset distribution shift towards structural biology content that reduced the density of pure-RDKit drills. R5 corrected this with `pyexec_drill` (\u00d76), `code_then_quote_v2` (\u00d75), and `fidelity_multistep` (\u00d75) generators.\n\n## \ud83d\udcda RAG corpus\n\nThe `data/corpus/` directory contains the full retrieval knowledge base: tool reference docs plus two databases of experimental chemistry data:\n\n### \ud83d\udcda ChEMBL (`data/corpus/chembl/`, ~70 MB)", "# Abstracts of Marc C. Deller's publications\nAuthors: various (Deller, M.C. co-author)\nSource: Europe PMC, 2026\n\n**Authors:** Pozharski E, Deller MC, Rupp B.  \n**Journal:** Methods in molecular biology (Clifton, N.J.) (2017). DOI: 10.1007/978-1-4939-7000-1_25. PMID: 28573591.\n\nModels of target proteins in complex with small molecule ligands or peptide ligands are of significant interest to the biomedical research community. Structure-guided lead discovery and structure-based drug design make extensive use of such models. The bound ligands comprise only a small fraction of the total X-ray scattering mass, and therefore particular care must be taken to properly validate the atomic model of the ligand as experimental data can often be scarce. The ligand model must be validated against both the primary experimental data and the local environment, specifically: (1) the primary evidence in the form of the electron density, (2) examined for reasonable stereochemistry, and (3) the chemical plausibility of the binding interactions must be inspected. Tools that assist the researcher in the validation process are presented.\n\n## Key gp120 Glycans Pose Roadblocks to the Rapid Development of VRC01-Class Antibodies in an HIV-1-Infected Chinese Donor (2016)\n\n**Authors:** Kong L, Ju B, Chen Y, He L, Ren L, Liu J, Hong K, Su B, Wang Z, Ozorowski G, Ji X, Hua Y, Chen Y, Deller MC, Hao Y, Feng Y, Garces F, Wilson R, Dai K, O'Dell S, McKee K, Mascola JR, Ward AB, Wyatt RT, Li Y, Wilson IA, Zhu J, Shao Y.  \n**Journal:** Immunity (2016). DOI: 10.1016/j.immuni.2016.03.006. PMID: 27067056.\n\nVRC01-class antibodies neutralize diverse HIV-1 strains by targeting the conserved CD4-binding site. Despite extensive investigations, crucial events in the early stage of VRC01 development remain elusive. We demonstrated how VRC01-class antibodies emerged in a Chinese donor by antigen-specific single B cell sorting, structural and functional studies, and longitudinal antibody and virus repertoire analyses. A monoclonal antibody DRVIA7 with modest neutralizing breadth was isolated that displayed a subset of VRC01 signatures. X-ray and EM structures revealed a VRC01-like angle of approach, but less favorable interactions between the DRVIA7 light-chain CDR1 and the N terminus with N276 and V5 glycans of gp120. Although the DRVIA7 lineage was unable to acquire broad neutralization, longitudinal analysis revealed a repertoire-encoded VRC01 light-chain CDR3 signature and VRC01-like neutralizing heavy-chain precursors that rapidly matured within 2 years. Thus, light chain accommodation of the glycan shield should be taken into account in vaccine design targeting this conserved site of vulnerability.\n\n## Protein stability: a crystallographer's perspective (2016)\n\n**Authors:** Deller MC, Kong L, Rupp B.  \n**Journal:** Acta crystallographica. Section F, Structural biology communications (2016). DOI: 10.1107/s2053230x15024619. PMID: 26841758.\n\nProtein stability is a topic of major interest for the biotechnology, pharmaceutical and food industries, in addition to being a daily consideration for academic researchers studying proteins. An understanding of protein stability is essential for optimizing the expression, purification, formulation, storage and structural studies of proteins. In this review, discussion will focus on factors affecting protein stability, on a somewhat practical level, particularly from the view of a protein crystallographer. The differences between protein conformational stability and protein compositional stability will be discussed, along with a brief introduction to key methods useful for analyzing protein stability. Finally, tactics for addressing protein-stability issues during protein expression, purification and crystallization will be discussed.\n\n## Complete epitopes for vaccine design derived from a crystal structure of the broadly neutralizing antibodies PGT128 and 8ANC195 in complex with an HIV-1 Env trimer (2015)\n\n**Authors:** Kong L, Torrents de la Pe\u00f1a A, Deller MC, Garces F, Sliepen K, Hua Y, Stanfield RL, Sanders RW, Wilson IA.  \n**Journal:** Acta crystallographica. Section D, Biological crystallography (2015). DOI: 10.1107/s1399004715013917. PMID: 26457433.\n\nThe HIV-1 envelope gp160 glycoprotein (Env) is a trimer of gp120 and gp41 heterodimers that mediates cell entry and is the primary target of the humoral immune response. Broadly neutralizing antibodies (bNAbs) to HIV-1 have revealed multiple epitopes or sites of vulnerability, but mapping of most of these sites is incomplete owing to a paucity of structural information on the full epitope in the context of the Env trimer. Here, a crystal structure of the soluble BG505 SOSIP gp140 trimer at 4.6\u2005\u00c5 resolution with the bNAbs 8ANC195 and PGT128 reveals additional interactions in comparison to previous antibody-gp120 structures. For 8ANC195, in addition to previously documented interactions with gp120, a substantial interface with gp41 is now elucidated that includes extensive interactions with the N637 glycan. Surprisingly, removal of the N637 glycan did not impact 8ANC195 affinity, suggesting that the antibody has evolved to accommodate this glycan without loss of binding energy. PGT128 indirectly affects the N262 glycan by a domino effect, in which PGT128 binds to the N301 glycan, which in turn interacts with and repositions the N262 glycan, thereby illustrating the important role of neighboring glycans on epitope conformation and stability. Comparisons with other Env trimer and gp120 structures support an induced conformation for glycan N262, suggesting that the glycan shield is allosterically modified upon PGT128 binding. These complete epitopes of two broadly neutralizing antibodies on the Env trimer can now be exploited for HIV-1 vaccine design.\n\n## Crystal structure of a two-subunit TrkA octameric gating ring assembly (2015)\n\n**Authors:** Deller MC, Johnson HA, Miller MD, Spraggon G, Elsliger MA, Wilson IA, Lesley SA.  \n**Journal:** PloS one (2015). DOI: 10.1371/journal.pone.0122512. PMID: 25826626.", "# BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n* `--mps-watermark` sets `PYTORCH_MPS_HIGH_WATERMARK_RATIO`, which caps how much memory MPS will allocate relative to the device\u2019s recommended maximum. At the default `1.0`, an oversized complex now raises a clean MPS out-of-memory error in the log instead of silently spilling into swap.\n* `--workers` defaults to 2 and `--max-parallel-samples` to 1: both trade a little parallelism for a much smaller memory footprint, which matters more on unified memory than on a dedicated-VRAM GPU.\n* `preflight`\u2018s `memory_heuristic` check WARNs when a target\u2019s total residue+ligand-atom count crosses `--memory-warn-tokens` (default 1000), citing the empirical data point above. It\u2019s a rough heuristic, not a precise memory model.\n* `run`\u2018s progress bar shows live memory usage (RSS summed across the whole `boltz predict` process tree), and logs a warning if usage stays above 90% of system RAM for 60+ seconds with no new completed target: a sign of thrashing, not genuine progress.\n* `run`/`all` auto-retries (`--max-retries`, default 2) any target that doesn\u2019t complete, isolating every still-incomplete target to its own single-target `boltz predict` invocation from the first retry onward \u2014 a real 4-target cascade on `5ht2_gq_panel` (an OOM on 2 of 6 large targets run together crashed the shared affinity phase for 2 more that had already succeeded) recovered cleanly this way. This means a large campaign can be started and left unattended: a transient OOM no longer needs a human to notice, wait, and manually re-run just the affected targets.\n\nIf a target still fails after every automatic retry, that\u2019s a real finding (this hardware may not be viable for a complex that size), not something to force through.\n\n## \u2699\ufe0f Progress bar\n\nTwo rows during `run`: the outer bar shows targets done/total, elapsed time, an ETA, and live memory usage. The inner row shows which **phase** Boltz is in (MSA generation / structure prediction / affinity prediction) and that phase\u2019s own progress, parsed from Boltz\u2019s own log output. Note: Boltz doesn\u2019t expose diffusion- or recycling-step-level progress anywhere in its output (verified against the installed package\u2019s source), so this is the finest granularity actually available, not a per-step diffusion counter.\n\n## \ud83d\udcca Outputs\n\nWritten next to `boltz_input.md`:", "# 2026 \u2013 From Tool to Teammates: Working Smarter with AI\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-28 on marcdeller.com (categories: certificates).*\n\n* [Link to certificate](https://www.linkedin.com/learning/certificates/495a6382d4faf9508e4de7d702bef1ee6efc138e02a081418665061a41cc23b2)\n\n* **Award Details:** LinkedIn Learning issued this certificate on March 24, 2026 at 10:02 AM UTC.\n* **Course Focus:** It recognizes completion of a course on using AI as a practical workplace collaborator.\n* **Career Value:** This matters because AI is becoming part of everyday scientific planning, drafting, and synthesis.\n* **Practical Use:** It helps researchers work more efficiently in brainstorming, review, and structured problem-solving.", "# Anti-Mutant Calreticulin (CALR) Antibodies and Uses Thereof\nAuthors: Mayes, Patrick; Nastri, Horacio G.; Buonpane, Rebecca A.; Reis, Edimara S.; Stewart, Shaun M.; Zhou, Jing; Deller, Marc C.; Lei, Hsiang-Ting; McQuirter, Leslie Brooke Epling; Zhao, Yonghong; Celik, Hamza; Wass, Brittney Melissa (inventors)\nSource: PCT International Patent Application WO 2023/107994 A1 (PCT/US2022/081076), applicant Incyte Corporation; published 15 June 2023, 2023\n\nFIGs. 13A-13B show residues involved in binding of the Fab to CalR1. A magnified image of selected CalR1-Fab1 interacting residues is shown in FIG. 13B. FIG. 13C provides details for the selected CalR1-Fab1 interacting residues shown in FIG. 13B. FIGs. 14A-14B show residues involved in binding of the Fab to CalR2. A magnified image of selected CalR2-Fab1 interacting residues is shown in FIG. 14B. FIG. 14C provides details for the selected CalR2-Fab1 interacting residues shown in FIG. 14B. In FIGs. 13A-13C and FIGs. 14A-14C, residues were selected based on distance and interaction type determined by MOE (Molecular Operating Environment). Heavy chain residues are shown in black and light chain residues are shown white. Distances in Angstroms are shown by dotted lines. Interacting residues are shown as sticks and were calculated using the InterfaceResidues.py script and Pymol (protein.osaka-u.ac.jp/rcsfp/supracryst/suzuki/jpxtal/Katsutani/en/interface.php).\n\nFIG. 15 shows the sequence of mutant CALR peptide with the CalR1 conformation binding residues (top) and CalR2 conformation binding residues (bottom) in grey shading.\n\n### Example 12: Structural Determination of Fab Fragment from anti-mutCALR Antibody Bound with Type 1 Mutant CalR 31-mer Peptide\n\nThe Fab fragment having the sequences shown below was purified as described above for Example 11.\n\n**Sequences:**\n\nFab Heavy Chain (SEQ ID NO:330):\n\n```\nEVQLVQSGAEVKKPGASVKVSCKVSGYTLTELSMQWVRQAPGKGLEWMGGFDPDDAETMYAEKF\nQGRLTVTEDTSTDTVYMELSSLRSEDTAVYFCATSPGYDFFDYWGQGTLVTVSSASTKGPSVFP\nLAPSSKSTSGGTAALGCLVKDYFPEPVTVSWNSGALTSGVHTFPAVLQSSGLYSLSSVVTVPSS\nSLGTQTYICNVNHKPSNTKVDKRVEPKSCDKTHT\n```\n\nFab Light Chain (SEQ ID NO:260):\n\n```\nSYVLTQPPSVSVAPGKTARITCTGTSSDVGGYNYVSWYQQKPGQAPVLVVYEVSNRPSGIPERF\nSGSNSGNTATLTISRVEAGDEADYYCQVWDSSNDLLIFGGGTKLTVLGQPKAAPSVTLFPPSSE\nELQANKATLVCLISDFYPGAVTVAWKADSSPVKAGVETTTPSKQSNNKYAASSYLSLTPEQWKS\nHRSYSCQVTHEGSTVEKTVAPTECS\n```\n\n31-mer mutCALR peptide (SEQ ID NO:340):\n\n```\nH-DEEQRTRRMMRTKMRMRRMRRTRRKMRRKMS-OH\n```\n\n**Crystallization** was performed as follows: Recombinant Fab fragment was mixed with the 31-mer peptide at a ratio of 1:3 molar excess, and then concentrated to a final concentration of 21 mg/mL. Sparse matrix crystallization screens were set up using an NT8 crystallization robot (Formulatrix, Bedford, MA). Drops containing 200 nl (Fab fragment plus CalR peptide) + 200 nl (reservoir) were used for the setup and the plates were incubated at 4 \u00b0C, 13 \u00b0C and 20 \u00b0C. Crystals grew from condition E7 of the JCSG Top96 Screen (Rigaku, Bainbridge Island, WA) after one day of incubation at 20 \u00b0C. The final crystal condition for antibody X:31-mer peptide complex was TRIS 0.1 M, pH 8.5, polyethylene glycol 400 40% v/v, and 0.2 M lithium sulfate.\n\nAll crystals were flash cooled in liquid nitrogen for X-ray data collection.\n\n**Data Collection, Processing and Refinement** was performed as follows: Diffraction data was collected at 100 K (at \u03bb = 0.9201 \u00c5) with an Eiger 9M detector using synchrotron radiation at beamline 17-ID-1 at the Center for BioMolecular Structure at the National Synchrotron Light Source II, Brookhaven, NY, USA. 1800 images were collected using a rotation of 0.1\u00b0 per image. Diffraction data indexing, integration and scaling were performed using Fast_dp. Data collection statistics, phasing and refinement are given in Table 15.\n\nFor the Fab fragment-31-mer peptide data, Phaser was used to place fragments of the Fab fragment model and subsequent rounds of manual model building were performed resulting in a model containing 1 Fab and 1 peptide. Subsequent cycles of density modifications, model building and refinement were carried out using Refmac, Coot and Phenix until the structure converged at reasonable R-work and R-free (Table 15). The final model were analyzed for good stereochemistry, geometry and clash scores using MolProbity (Table 15).\n\n*Table 15.* Data collection statistics, phasing and refinement. (Values in parentheses are for the highest-resolution shell.)\n\n| | Fab fragment:CalR peptide (31-mer) |\n|---|---|\n| Resolution range (\u00c5) | 29.62 \u2013 2.04 (2.113 \u2013 2.04) |\n| Space group | C 1 2 1 |\n| Unit cell | 105.37 71.477 92.347 90 105.807 90 |\n| Total reflections | 82513 (8206) |\n| Unique reflections | 42070 (4190) |\n| Multiplicity | 2.0 (2.0) |\n| Completeness (%) | 99.76 (99.90) |\n| Mean I/sigma(I) | 10.09 (1.61) |\n| Wilson B-factor | 38.61 |\n| R-merge | 0.04238 (0.4129) |\n| R-meas | 0.05993 (0.584) |\n| R-pim | 0.04238 (0.4129) |\n| CC1/2 | 0.997 (0.679) |\n| CC* | 0.999 (0.899) |\n| Reflections used in refinement | 42064 (4190) |\n| Reflections used for R-free | 1995 (198) |\n| R-work | 0.2017 (0.2759) |\n| R-free | 0.2262 (0.3079) |\n| CC(work) | 0.957 (0.789) |\n| CC(free) | 0.949 (0.760) |\n| Number of non-hydrogen atoms | 3690 |\n| macromolecules | 3440 |\n| ligands | 38 |\n| solvent | 212 |\n| Protein residues | 464 |\n| RMS(bonds) | 0.008 |\n| RMS(angles) | 1.13 |\n| Ramachandran favored (%) | 95.18 |\n| Ramachandran allowed (%) | 4.39 |\n| Ramachandran outliers (%) | 0.44 |\n| Rotamer outliers (%) | 0 |\n| Clashscore | 5.02 |\n| Average B-factor | 47.84 |\n| macromolecules | 47.77 |\n\n**Results:**\n\nThe structure of Fab fragment-31-mer mutCalR peptide was determined at 2.0 \u00c5 with space group C 1 2 1. Each asymmetric unit contains one Fab molecule and one 31-mer peptide molecule. The Fab molecule consists of a light chain, Chain L: 2-215, and a heavy chain, Chain H: 1-221. The model has well defined electron density throughout, with the exception of the following residues: Chain H, Gly42-Lys43, which are not modelled due to weak electron density. The structure was refined to an R-work/R-free of 20.1% and 22.6%, respectively, and has good stereochemistry throughout with 2 Ramachandran outliers.", "# Crystal structure and functional dissection of the cytostatic cytokine oncostatin M\nAuthors: Deller, M.C.; Hudson, K.R.; Ikemizu, S.; Bravo, J.; Jones, E.Y.; Heath, J.K.\nSource: Structure, 2000\n\nKnowledge of the structural basis of receptor recognition by gp130 cytokines is a central issue in understanding both the specificity of their biological functions and developing methods for therapeutic intervention in gp130-mediated signalling. The crystal structures of three gp130 cytokines, LIF [19], CNTF [20] and IL-6 [21], have been previously described. These reveal that all three cytokines share a common three-dimensional fold, the four-helix bundle [22], which comprises four \u03b1 helices ranging from 15 to 22 amino acids in length (termed A, B, C and D) and linked by polypeptide loops. Detailed structure\u2013function studies of human LIF, which is most closely related to hOSM in sequence and receptor repertoire, have defined two receptor recognition epitopes: sites 2 and 3. Site 2 (so-termed by analogy to the receptor-binding sites in growth hormone [23]), is required for recognition of gp130 and has been narrowed down to a cluster of candidate amino acid residues located in helices A and C [24]. Site 3 is required for recognition of LIFR and is dominated by a pair of conserved amino acid residues located at the N-terminal tip of helix D [24]. The combined activity of these two recognition epitopes leads to the formation of a trimeric, LIF-mediated, high-affinity signalling complex [24].\n\nStructural comparisons of ligands with differing recognition properties present a valuable opportunity to analyse the molecular determinants that dictate receptor specificity. Here, we report the determination of the three-dimensional structure of hOSM by crystallographic techniques and concomitant functional dissection by mutagenesis. Because hOSM binds to gp130 with a higher affinity than LIF [24], this allows structure\u2013function analysis of the site 2 gp130 recognition epitope at the resolution of individual amino acid residues. Combining these findings with the high-resolution crystal structure of the complementary ligand-recognition epitope located in the cytokine receptor homology domain (CHD) of gp130 [25] results, for the first time, in a definition of the structure of the interface between hOSM and gp130. In addition a detailed comparison of the three-dimensional structure of the site 3 recognition epitope of hOSM and hLIF also offers insights into the molecular basis by which related ligands exhibit different receptor recognition properties. These findings reveal that, in both cases, receptor specificity is controlled by subtle structural changes to a common recognition framework.\n\n## Results\n\n### Expression and structure determination\n\nA soluble 21.4 kDa fragment of hOSM was expressed in Escherichia coli as a glutathione-S-transferase (GST) fusion protein. The resulting purified protein consists of residues 1\u2013187 of hOSM and a further four residues at the N terminus corresponding to the 3C protease site linker (GPGS). Ba/F3 cell-survival assays and receptor-binding assays confirmed that this recombinant form binds gp130 with an affinity equivalent to wild-type (WT) hOSM.\n\nCrystallisation trials produced Bragg diffraction quality crystals belonging to the orthogonal space group P212121, with unit-cell dimensions a = 35.8 \u00c5, b = 53.1 \u00c5, c = 106.8 \u00c5. The structure was determined from a single ethylmercuryphosphate (EMP) derivative using the multiple anomalous dispersion (MAD) phasing technique with X-ray diffraction data collected on beam line BM14 of the European Synchrotron Radiation Facility (ESRF). The structure has been refined to a crystallographic R factor of 20.5% (Rfree 26.1%) for all data between 20.0 and 2.2 \u00c5, with stereochemistry typified by root mean squared deviations (rmsd) in bond lengths of 0.015 \u00c5 and in bond angles of 1.8\u00b0. Crystallographic statistics are reported in Tables 1 and 2. Residues 1\u20133 and 135\u2013155 appear to be disordered and are not included in the final model. With the exception of these regions, and a flexible loop between residues Glu90 and Lys110, the electron density is good throughout the course of the mainchain (a representative portion of electron density is shown in Figure 1a). Residue numbering is that of the human native mature protein (GenBank accession number M27286).\n\n### Structure description\n\nhOSM is a compact barrel-shaped molecule with dimensions of approximately 20 \u00c5 \u00d7 27 \u00c5 \u00d7 56 \u00c5 that conforms to the up-up-down-down four-helix bundle topology (Figure 1b,c). The basic fold consists of the four main helical regions (helix A, residues 10\u201337; helix B, residues 67\u201390; helix C, residues 105\u2013131; helix D, residues 159\u2013185) connected by two long overhand loops (AB loop, residues 38\u201366; CD loop, residues 130\u2013158) and one short loop (BC loop, residues 91\u2013104). The overall arrangement may be considered as two pairs of antiparallel helices that pack together, with A\u2013D forming one pair and B\u2013C forming the other.\n\nBoth helices A and C exhibit breaks in the classical hydrogen-bonding pattern, with substitute hydrogen bonds formed to tightly bound water molecules. In helix C this results in a slight kink (between residues Gln112 and Pro116). The kink in helix A is more pronounced and is caused by a break in the helical conformation, as residues Gln25 and Leu30 adopt a hydrogen-bonded turn conformation, with four tightly bound water molecules substituting for the classical hydrogen-bonding pattern (Figure 1a). Residues Thr27\u2013Ile37 of helix A adopt a 310 helix conformation. The curved nature of helices A and C facilitates close packing of the A\u2013D and B\u2013C helix pairs, so promoting an extensive solvent-inaccessible core.\n\nThe core of the protein is composed of two groups of aromatic stacking interactions in which Phe56, Tyr173, Phe169 and Phe176 form one group and Phe70, Phe185 and Trp187 form another. This distribution emphasizes the hydrophobic nature of helix D that donates all of the aromatics to these groupings, with the exception of Phe56 (AB loop) and Phe70 (B helix).", "# GitHub README: CODSWALLOP\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/CODSWALLOP, 2026\n\n### Phase 4: topology, export and polish", "# Revolutionary Cryo-EM Breakthrough: Di-Gembodies Transform Small Protein Structure Determination\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe **broader implications** for drug discovery and therapeutic development are profound. Small proteins often serve as critical nodes in cellular signaling pathways and represent important therapeutic targets. The ability to routinely determine their structures at atomic resolution opens new avenues for structure-based drug design, particularly for targets that have been historically challenging to study. The parallel structure determination capability further accelerates the pace of discovery by enabling comparative structural studies within single experiments.\n\n| Key Concept | Description | Key References |\n| --- | --- | --- |\n| **Di-Gembodies** | Covalently constrained nanobody dimers that create rigid scaffolds for displaying small proteins in cryo-EM imaging | [Yi, G. et al., Nature Chemical Biology](https://www.nature.com/articles/s41589-025-01972-7) |\n| **Size Enhancement** | Strategic approach to overcome the ~50 kDa size limitation in single-particle cryo-EM by attaching proteins to larger scaffolds | [Ackle, F. et al., Nature Communications](https://pmc.ncbi.nlm.nih.gov/articles/PMC11923154/) |\n| **Parallel Structure Determination** | Ability to simultaneously solve structures of two different proteins within a single cryo-EM experiment | [Yi, G. et al., Nature Chemical Biology](https://www.nature.com/articles/s41589-025-01972-7) |\n| **Covalent Constraints** | Chemical bonds that eliminate flexibility between protein components, ensuring rigid attachment for high-resolution imaging | [Kishikawa, J. et al., PNAS](https://www.pnas.org/doi/10.1073/pnas.2115435118) |\n| **Fiducial Markers** | Reference points that facilitate particle alignment and orientation determination during cryo-EM image processing | [Chu, X. et al., PNAS](https://www.pnas.org/doi/10.1073/pnas.2115001118) |\n| **Modular Design** | Standardized, interchangeable components that can be adapted for different protein targets without extensive re-engineering | [Liu, S. et al., Nature Communications](https://www.nature.com/articles/s41467-019-09836-0) |", "# marcdeller.com: My Patents\nAuthors: Deller, M.C.\nSource: marcdeller.com, 2026\n\n## Concept to Claims: The Patented Side of My Science\n\n7+ patents covering cancer therapeutics, immunology, and drug discovery. Cancer therapies are undergoing a seismic shift with patents emerging that unlock new molecular vulnerabilities. Among the most promising are compounds designed to disable WRN helicase, a DNA repair enzyme uniquely essential to cancers exhibiting microsatellite instability (MSI-H) or mismatch repair deficiency (dMMR). These inventions strike at the heart of tumor survival while sparing healthy cells, offering hope in settings where immunotherapies or chemotherapy resistance prevail. Alongside these inhibitors, novel antibody approaches targeting mutant calreticulin extend the toolkit of selective, mechanism-driven cancer therapeutics.\n\n* **WRN Helicase Inhibitors:** Oncology patents targeting WRN helicase, a synthetic lethal vulnerability in MSI-high and dMMR cancers. Structure-guided inhibitors offering selective therapy for colorectal, endometrial, and gastric tumors.\n* **Mutant CalR Targeted Antibodies:** Therapeutic antibodies targeting mutant calreticulin in myeloproliferative neoplasms, including essential thrombocythemia and myelofibrosis. Structure-based design enables selective targeting of malignant cells while sparing normal tissue.\n* **Innovative Structure-Enabled Drug Discovery:** Structure-guided discovery platforms integrating X-ray crystallography, cryo-EM, and computational design to enable rational drug development for difficult-to-drug targets in oncology and immunology.\n\nTop Patents\n\n### [2025 \u2013 Bicyclic compounds as inhibitors of wrn](https://marcdeller.com/bicyclic-compounds-as-inhibitors-of-wrn/)\n\n[2025 \u2013 Bicyclic compounds as inhibitors of wrn](https://marcdeller.com/bicyclic-compounds-as-inhibitors-of-wrn/)\n\n### [2025 \u2013 Tricyclic compounds as inhibitors of wrn](https://marcdeller.com/tricyclic-compounds-as-inhibitors-of-wrn/)\n\n[2025 \u2013 Tricyclic compounds as inhibitors of wrn](https://marcdeller.com/tricyclic-compounds-as-inhibitors-of-wrn/)\n\n### [2023 \u2013 Anti-mutant calreticulin (calr) antibodies and uses thereof](https://marcdeller.com/anti-mutant-calreticulin-calr-antibodies-and-uses-thereof/)\n\n[2023 \u2013 Anti-mutant calreticulin (calr) antibodies and uses thereof](https://marcdeller.com/anti-mutant-calreticulin-calr-antibodies-and-uses-thereof/)\n\n### [2025 \u2013 Imidaz0[1,2-a]pyrimidine-5(1 h)-ones as wrn inhibitors](https://marcdeller.com/imidaz012-apyrimidine-51-h-ones-as-wrn-inhibitors/)\n\n[2025 \u2013 Imidaz0[1,2-a]pyrimidine-5(1 h)-ones as wrn inhibitors](https://marcdeller.com/imidaz012-apyrimidine-51-h-ones-as-wrn-inhibitors/)\n\n### [2025 \u2013 Imidaz0[1,2-a]pyrimidine-5(1 h)-ones having a covalent warhead at the 3-position as wrn inhibitors](https://marcdeller.com/imidaz012-apyrimidine-51-h-ones-having-a-covalent-warhead-at-the-3-position-as-wrn-inhibitors/)\n\n[2025 \u2013 Imidaz0[1,2-a]pyrimidine-5(1 h)-ones having a covalent warhead at the 3-position as wrn inhibitors](https://marcdeller.com/imidaz012-apyrimidine-51-h-ones-having-a-covalent-warhead-at-the-3-position-as-wrn-inhibitors/)\n\nPatents\n\n### [2026 \u2013 From Bicyclic to Tricyclic: A New WRN Inhibitor Patent from Incyte](https://marcdeller.com/from-bicyclic-to-tricyclic-a-new-wrn-inhibitor-patent-from-incyte/)\n\nPatents\n\n### [2025 \u2013 Bicyclic Compounds as Inhibitors of WRN](https://marcdeller.com/bicyclic-compounds-as-inhibitors-of-wrn-2/)\n\nPatents\n\n### [2025 \u2013 Fused Bicyclic Pyrimidones as WRN Inhibitors](https://marcdeller.com/fused-bicyclic-pyrimidones-as-wrn-inhibitors/)\n\nPatents\n\n### [2025 \u2013 Imidaz0[1,2-a]pyrimidine-5(1 h)-ones having a covalent warhead at the 3-position as wrn inhibitors](https://marcdeller.com/imidaz012-apyrimidine-51-h-ones-having-a-covalent-warhead-at-the-3-position-as-wrn-inhibitors/)\n\nPatents\n\n### [2025 \u2013 Imidaz0[1,2-a]pyrimidine-5(1 h)-ones as wrn inhibitors](https://marcdeller.com/imidaz012-apyrimidine-51-h-ones-as-wrn-inhibitors/)\n\nPatents\n\n### [2023 \u2013 Anti-mutant calreticulin (calr) antibodies and uses thereof](https://marcdeller.com/anti-mutant-calreticulin-calr-antibodies-and-uses-thereof/)\n\nPatents\n\n### [2025 \u2013 Tricyclic compounds as inhibitors of wrn](https://marcdeller.com/tricyclic-compounds-as-inhibitors-of-wrn/)\n\nPatents\n\n### [2025 \u2013 Bicyclic compounds as inhibitors of wrn](https://marcdeller.com/bicyclic-compounds-as-inhibitors-of-wrn/)", "# GitHub README: AlphaFraud\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/AlphaFraud, 2026\n\n| Field | Detail |\n|---|---|\n| **Take-home** | AlphaFold built a tidy globular \u03b2-sandwich at pLDDT 94; the real structure is a novel brain amyloid \u2014 every ordered residue wrong |\n| **AlphaFold vs experimental** | <a href=\"https://alphafraud.mdeller.com/examples#ex-7U14_1\"><img src=\"https://alphafraud.mdeller.com/ribbon/7U14_1.svg\" alt=\"Experimental structure of 7U14_1 coloured by C\u03b1 deviation from AlphaFold\" width=\"380\"></a><br><sub>Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \"ghost\" overlay</sub> |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-7U14_1)** |\n| **KPIs** | TM 0.157 \u00b7 lDDT 0.536 \u00b7 pLDDT 94.4 \u00b7 C\u03b1-RMSD 27.1 \u00c5 \u00b7 FRAUD 0.87 \u00b7 novelty **100%** |\n| **Stats** | 29 depositions, **27** confidently wrong \u00b7 pLDDT\u2194lDDT 0.03 (confidence uninformative) |\n| **Structure** | Native luminal domain is a fibronectin-III \u03b2-sandwich; the deposited form is a five-layered cross-\u03b2 amyloid core |\n| **Why AlphaFold gets it wrong** | Amyloid is a self-templating polymer state absent from training, and the sequence is 100% novel \u2014 nothing steers it off its globular prior |\n| **Biology & disease** | Lysosomal protein; top GWAS hit for FTLD-TDP; its C-terminal domain forms fibrils across ageing and neurodegeneration |\n| **Key fact** | 100% sequence-novel \u2014 no pre-2018 fibril template existed |\n\n### \ud83d\udd34 Alpha-synuclein (SNCA) \u00b7 [`9A1Q_1`](https://alphafraud.mdeller.com/examples#ex-9A1Q_1)\n\n| Field | Detail |\n|---|---|\n| **Take-home** | A protein with no single fold; AlphaFold's confidence is actively *anti*-correlated with accuracy |\n| **AlphaFold vs experimental** | <a href=\"https://alphafraud.mdeller.com/examples#ex-9A1Q_1\"><img src=\"https://alphafraud.mdeller.com/ribbon/9A1Q_1.svg\" alt=\"Experimental structure of 9A1Q_1 coloured by C\u03b1 deviation from AlphaFold\" width=\"380\"></a><br><sub>Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \"ghost\" overlay</sub> |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-9A1Q_1)** |\n| **KPIs** | TM 0.138 \u00b7 lDDT 0.177 \u00b7 pLDDT 75.2 \u00b7 C\u03b1-RMSD 36.3 \u00c5 \u00b7 FRAUD 0.71 \u00b7 novelty 0% |\n| **Stats** | **210** depositions, **206** confidently wrong \u2014 the archive's most-repeated failure \u00b7 pLDDT\u2194lDDT \u22120.29 |\n| **Structure** | 140-residue intrinsically disordered monomer; the deposited forms are cryo-EM cross-\u03b2 fibrils |\n| **Why AlphaFold gets it wrong** | No native fold to predict, and polymorph-specific packing is not in the sequence \u2014 where it is more confident it is measurably more wrong |\n| **Biology & disease** | Presynaptic protein; fibrils define Parkinson's, DLB and multiple system atrophy; PD mutations A30P/E46K/H50Q/G51D/A53T |\n| **Key fact** | 210 depositions, 206 confidently wrong \u2014 the most-deposited catch on the site |\n\n### \ud83d\udd34 Transthyretin (TTR) \u00b7 [`9BZS_1`](https://alphafraud.mdeller.com/examples#ex-9BZS_1)\n\n| Field | Detail |\n|---|---|\n| **Take-home** | pLDDT 98 on the native tetramer fold; the deposited cardiac amyloid fibril scores TM 0.21 \u2014 one of the starkest confidently-wrong cases |\n| **AlphaFold vs experimental** | <a href=\"https://alphafraud.mdeller.com/examples#ex-9BZS_1\"><img src=\"https://alphafraud.mdeller.com/ribbon/9BZS_1.svg\" alt=\"Experimental structure of 9BZS_1 coloured by C\u03b1 deviation from AlphaFold\" width=\"380\"></a><br><sub>Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \"ghost\" overlay</sub> |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-9BZS_1)** |\n| **KPIs** | TM 0.214 \u00b7 lDDT 0.489 \u00b7 pLDDT 98.0 \u00b7 C\u03b1-RMSD 22.8 \u00c5 \u00b7 FRAUD **0.95** \u00b7 novelty 0% |\n| **Stats** | 160 depositions, 34 confidently wrong \u00b7 pLDDT\u2194lDDT \u22120.16 |\n| **Structure** | Native 127-aa \u03b2-sandwich homotetramer with thyroxine sites; the amyloid form is a cross-\u03b2 fibril from cleaved segments |\n| **Why AlphaFold gets it wrong** | A wrong-*state* failure: pLDDT measures confidence in the known native fold, not whether a deposition *is* that state; TTR is 0% novel, squarely in training |\n| **Biology & disease** | Thyroxine/retinol transporter; ATTR amyloidosis (wild-type cardiac + hereditary V30M); uniquely druggable (tafamidis, patisiran) |\n| **Key fact** | FRAUD 0.95 \u2014 near the very top of the whole site |\n\n### \ud83d\udd34 Coagulation factor X (F10) \u00b7 [`6Q9F_2`](https://alphafraud.mdeller.com/examples#ex-6Q9F_2)", "# deepcovalent: DeepCovalent | Transforming Drug Discovery Today\nAuthors: deepcovalent\nSource: https://www.deepcovalent.com, 2026\n\n## Discovering New Drugs in **One Step**\n\nDeepCovalent\u2019s [One-Step-Denovo Drug Discovery (**OS3D\u2122**)](/os3d) is a multimodal, multi-objective generative AI engine that designs clinically viable, synthesis-ready de novo small molecules in a single forward pass. Purpose-built to address the central bottlenecks in early discovery \u2014 slow, capital-intensive R&D workflows and the > 90% clinical failure rate driven by incomplete biological and ADMET understanding \u2014 **OS3D\u2122** enables rapid, constraint-aware drug molecule generation with significantly higer downstream success.\n\n[Learn more](/os3d)\n\n## Ideation to Validated Molecules in **3 Weeks**\n\n**OS3D\u2122** compresses early-stage discovery from months to **\u2264 3 weeks**, enabling low-cost, de-risked evaluation of new or deprioritized targets and creating early pipeline optionality.\n\n[Partner with us](/partnerships)\n\n## Scientific Validation\n\n**OS3D\u2122** has been *in-silico* validated on FDA-approved drugs and across DeepCovalent\u2019s internal discovery programs. [Contact us](/contact) to request the white paper and learn more.\n\n## Pipelines\n\nDeepCovalent is developing viable therapeutic pipelines powered by **OS3D\u2122**. Learn more about our pipelines [here](/pipeline).", "# BoltzMaker: Boltz2 campaign-scale structure and affinity prediction, and binding analysis\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\nIt\u2019s a core part of `analyze`/`all`: any family with an `Apo structure:` field set (see **boltz_input.md format** above) gets compared automatically, no separate command needed, and the result is embedded directly into `boltz_dashboard.html`. A family with no apo structure configured isn\u2019t silently skipped either \u2014 the dashboard\u2019s \u201cFamily coverage\u201d table says so explicitly, alongside any family that *was* compared. Pass `--skip-sse` to `analyze`/`all` to opt out, or use the standalone `compare-sse` command below to re-run just this analysis on its own (its own `--family`/`--target` flags let you target one family/target instead of the whole campaign).\n\nMotifs are annotated by one of three pluggable sources, auto-selected per family (or set explicitly with `Family type:`):\n\n| Family type | Motifs | Source |\n| --- | --- | --- |\n| `gpcr` | TM1-7, H8, ECL1-3, ICL1-3 | [GPCRdb](https://gpcrdb.org)\u2018s structure-based generic-numbering service (Ballesteros-Weinstein / GPCRdb schemes) |\n| `kinase` | hinge, gatekeeper, catalytic loop (HRD), DFG motif, alphaC-Glu, catalytic Lys | [KLIFS](https://klifs.net)\u2018s public REST API (its fixed 85-residue pocket alignment) |\n| `auto` (default) | whichever of the above applies, else\u2026 | \u2026falls back to Pfam domain boundaries via [PDBe](https://www.ebi.ac.uk/pdbe)\u2018s SIFTS residue mapping \u2014 the universal last resort for any protein outside the two families above |\n\nApo is superposed onto holo using only the family\u2019s stable, non-binding-site-adjacent residues (via gemmi\u2019s `superpose_positions`), so a ligand-induced local shift can\u2019t skew the global fit. Each motif then gets:\n\n| Metric | What it means |\n| --- | --- |\n| Ca RMSD / centroid shift | How far the motif moved, post-superposition |\n| Helix-axis rotation angle | For helical motifs \u2014 e.g. the classic TM6 \u201coutward swing\u201d on GPCR activation |\n| Helix kink angle (apo/holo/delta) | Whether a helix straightened or kinked more |\n| SSE boundary shift | Did the helix/strand get longer or shorter \u2014 needs deposited HELIX/SHEET records, or an optional external `mkdssp`/`dssp` binary as a fallback (see **One-time setup** above); every other metric works without it |\n| Flagged phi/psi residues | Per-residue backbone dihedral outliers above `--phi-psi-threshold` |\n| DFG-in/out, alphaC-in/out (kinases only) | A coarse Ca-Ca distance classifier, not a full dihedral model \u2014 good for detecting a state *change* between apo and holo, not publication-grade conformational classification |\n\nA metric that genuinely wasn\u2019t computed for a given motif (e.g. axis rotation for a loop, DFG state for a non-kinase family, boundary shift with no DSSP data available) shows as an explicit `N/A` in both the CSV and every dashboard table, not a blank cell.\n\nAbove the per-motif table, both `boltz_dashboard.html`\u2018s embedded card and the standalone `boltz_sse_comparison.html` show:\n\n| Section | Content |\n| --- | --- |\n| Family coverage | One row per protein family: `OK` (with a target/motif count) / `No apo structure configured` / `Apo structure file not found` / `No motif annotation available` / `No predicted (holo) structures yet` |\n| Overall shift statistics | Targets/motifs compared; mean/median/max Ca RMSD (and which target + motif had the largest shift); mean centroid shift; total flagged phi/psi residues; kinase DFG/alphaC state-change counts |\n\n```\npython3 BoltzMaker.py compare-sse boltz_input.md\n```\n\nWrites, next to `boltz_input.md` (or `--out-dir`): `boltz_sse_comparison.csv` (one row per family/target/motif) and `boltz_sse_family_status.json` (the family-coverage table above, machine-readable), a standalone self-contained `boltz_sse_comparison.html` (Plotly bar chart + motif x target heatmap, vendored the same way as the main dashboard), and `boltz_sse_comparison_sessions/<target>.pml` \u2014 a plain-text PyMOL script per target that colours/labels each motif and highlights the ones with a significant shift. It\u2019s just text: opens in any local PyMOL install, no `pymol` dependency in BoltzMaker\u2019s own venv.\n\nWhen auto-run by `analyze`/`all`, a campaign with no apo structures configured anywhere just gets a dashboard section saying so \u2014 it never aborts the rest of the pipeline over an optional, additive feature. The standalone command above still exits with a clear error if you explicitly pass a `--family`/`--target` that matches nothing, since that\u2019s a real mistake worth stopping for.\n\n## \ud83e\ude79 Troubleshooting / FAQ", "# Self-Encoded Libraries: A New Era in Drug Discovery Screening\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n| Key Concept | Description | Key Reference |\n| --- | --- | --- |\n| **Self-Encoded Libraries (SELs)** | Barcode-free combinatorial libraries synthesized on solid-phase beads where compounds are identified through tandem MS fragmentation patterns rather than DNA tags, enabling screening of 500,000+ molecules per experiment | [V\u00f6geli et al., 2025](https://www.nature.com/articles/s41467-025-65282-1) |\n| **COMET Software** | Combinatorial Mass Encoding decoding Tool that predicts fragmentation frequencies, filters MS/MS scans, and automates structure annotation with 66-82% recall rates while reducing false positives | [V\u00f6geli et al., 2025](https://www.nature.com/articles/s41467-025-65282-1) |\n| **DNA-Incompatible Target Screening** | Application of SELs to nucleic acid-binding proteins like FEN1 (flap endonuclease 1), yielding nanomolar inhibitors that would be impossible to identify using DNA-encoded library approaches | [V\u00f6geli et al., 2025](https://www.nature.com/articles/s41467-025-65282-1) |\n| **Affinity Selection-MS/MS** | Technology platform coupling affinity-based enrichment of target binders with tandem mass spectrometry sequencing, achieving library screening of 100+ million peptidomimetic compounds without genetic barcodes | [Quartararo et al., 2023](https://pubs.acs.org/doi/10.1021/jacs.3c04899) |\n| **Chemical Reaction Compatibility** | Expanded synthesis scope including palladium cross-couplings, heterocyclizations, and harsh protecting group removals that are incompatible with DNA stability in traditional encoded library approaches | [V\u00f6geli et al., 2025](https://www.nature.com/articles/s41467-025-65282-1) |", "# From Noise to Nanomolar: PXDesign\u2019s New Benchmark for Protein Binder Design\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-04-23 on marcdeller.com (categories: blog, sb).*\n\n* [**Technical Report**](https://protenix.github.io/pxdesign/technical_report.pdf)\n* [**Web Server**](https://protenix-server.com)\n\n* **Diffusion outperforms hallucination at scale:** PXDesign-d produces more filter-passing binder candidates than hallucination-based methods, making it better for large exploratory campaigns.\n* **Filtering is the real bottleneck:** No single confidence metric works for all targets; combining Protenix and AF2-IG predictors captures complementary binder populations.\n* **Experimental hit rates set a new bar:** PXDesign attains 17\u201382% nanomolar hit rates across six targets, matching or surpassing AlphaProteo and BindCraft.\n* **Immediately accessible to the community:** PXDesign is deployable via an open-source repository and public web server, enabling drug discovery teams without specialized infrastructure or ML expertise.\n\n1. **PXDesign: Fast, Modular, and Accurate De Novo Design of Protein Binders:** [Ren, Sun, Guan et al., Protenix Team / ByteDance Seed, 2025](https://protenix.github.io/pxdesign/technical_report.pdf)\n2. **De Novo Design of Protein Structure and Function with RFdiffusion:** [Watson et al., Nature, 2023](https://www.nature.com/articles/s41586-023-06415-8)\n3. **De Novo Design of High-Affinity Protein Binders with AlphaProteo:** [Zambaldi et al., Google DeepMind, arXiv, 2024](https://arxiv.org/abs/2409.08022)\n4. **One-Shot Design of Functional Protein Binders with BindCraft:** [Pacesa et al., Nature, 2025](https://www.nature.com/articles/s41586-025-09429-6)\n5. **Protenix: Advancing Structure Prediction Through a Comprehensive AlphaFold3 Reproduction:** [Chen, Zhang et al., ByteDance AML AI4Science Team, bioRxiv, 2025](https://www.biorxiv.org/content/10.1101/2025.01.08.631967v1)\n\n**ByteDance\u2019s PXDesign suite achieves nanomolar hit rates of 17\u201382% across six drug-relevant targets, outperforming leading methods through smarter generation and multi-predictor filtering**\n\nThe design of protein binders \u2014 proteins engineered from first principles to bind a target with high affinity and selectivity \u2014 is one of the defining challenges of modern structural biology and drug discovery. For decades, the field depended on immunization, directed evolution, or display-based selection to identify candidates from large libraries, with limited control over which epitope was engaged. Deep learning changed that calculus substantially. With the arrival of diffusion-based generative models like [RFdiffusion](https://www.nature.com/articles/s41586-023-06415-8) in 2023 and hallucination-based pipelines such as [BindCraft](https://www.nature.com/articles/s41586-025-09429-6) shortly after, it became possible to propose binder sequences computationally, short-cutting rounds of experimental optimization that previously took months. Yet despite these advances, head-to-head comparisons under consistent evaluation conditions remained scarce, and critical questions around the accuracy of confidence-based filtering \u2014 the computational step that converts thousands of candidate structures into a workable experimental shortlist \u2014 were largely unresolved.\n\nThat gap is addressed directly by the team behind [Protenix](https://www.biorxiv.org/content/10.1101/2025.01.08.631967v1), ByteDance Seed\u2019s open-source AlphaFold3 reproduction. In August 2025, the same group released PXDesign, a modular suite integrating two complementary generative strategies \u2014 diffusion and hallucination \u2014 with a rigorous multi-predictor filtering framework. The system was validated in vitro across seven pharmacologically relevant targets, including PD-L1, VEGF-A, SARS-CoV-2 RBD, IL-7RA, TrkA, EGFR, and TNF-\u03b1, benchmarked directly against [AlphaProteo](https://arxiv.org/abs/2409.08022) and RFdiffusion under matched conditions.\n\n> **\u201cPXDesign achieves nanomolar binder hit rates of 17\u201382% across six of seven diverse protein targets, surpassing prior methods such as AlphaProteo by 2\u20136\u00d7.\u201d**\n\nThe [PXDesign pipeline](https://protenix.github.io/pxdesign/technical_report.pdf) operates in three stages. First, PXDesign-d, a Diffusion Transformer backbone trained to generate Cartesian atom coordinates directly rather than SE(3)-equivariant backbone frames, proposes diverse target-conditioned binder structures ranging from 60 to 160 amino acids, with sequences then assigned by ProteinMPNN. Critically, PXDesign-d samples a broader region of fold space than its predecessors \u2014 producing binders that are not overwhelmingly \u03b1-helical, a well-documented compositional bias in earlier diffusion models. Filtering then applies interface confidence metrics, specifically ipTM and pTM, from both Protenix and AF2-IG predictors in combination, since a key experimental finding is that the two predictors capture largely non-overlapping populations of true binders. Ensemble filtering, using both models, is materially better than either alone.", "# JCSG Top96: A Rationally Designed Crystallization Screen Based on over 1000 Deposited Protein Structures\nAuthors: Deller, M. C.; Clayton, T.; Elsliger, M. A.; Wilson, I. A.\nSource: American Crystallographic Association (ACA) Annual Meeting 2013 (poster), 2013\n\n*Figure 4.* Performance of the seven precipitant classes of JCSG Top96. (a\u2013c) PDB and JCSG Top96 counts, (d) difference from mean. Note above average performance for all PEG combination precipitant classes shown in (d).\n\n*Figure 5.* Enrichment of PEG combination precipitant classes. (a) JCSG Core Suite, (b) JCSG Top96 and (c) JCSG Top96 Cryo. PEG combination precipitants (green) are enriched at the expense of the single precipitant classes (blue).\n\n*Figure 6.* Selection of optimal cryo-protectant for each of the JCSG Top96 conditions. PDB deposition data used to determine cryo-protectant that most frequently gives clear freeze (visually and diffraction-based) for each condition. (a\u2013f) cryo-protectants in use at JCSG.\n\n## Conclusions\n\nThe JCSG Top96 is a minimal, structure-based 96-condition crystallization screen derived from over 1000 deposited JCSG protein structures. Fifty-five percent of JCSG structures crystallize in these 96 conditions, giving roughly 4-fold better performance than the earlier JCSG+ screen (631 versus 168 PDB depositions), with only 16 conditions in common. PEG combination precipitants perform above average and are enriched in the screen, whereas single precipitant conditions perform below average. A companion JCSG Top96 Cryo screen provides the most successful cryo-protectant for each condition. Together the screens save protein and money and are suitable for both large-scale HT-SG platforms and individual laboratories.\n\n## Acknowledgements\n\nWe thank the members of the JCSG high-throughput structural biology pipeline for their contribution to this work. Portions of this research were carried out at the Stanford Synchrotron Radiation Lightsource (SSRL) and the Advanced Light Source (ALS).", "# Revolutionizing Drug Discovery: The Click2Drug Platform and the Rise of AI-Powered Virtual Screening\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe revolutionary impact of conformal prediction in virtual screening cannot be overstated. This machine learning framework enables researchers to control error rates while screening vast chemical libraries, addressing the inherent challenge of imbalanced datasets where only a small fraction of compounds demonstrate biological activity. By combining conformal prediction with molecular docking, researchers can now reduce the number of molecules requiring explicit docking by three orders of magnitude while maintaining the ability to identify top-scoring compounds efficiently. This approach has proven particularly effective for G protein-coupled receptor targets, successfully identifying dual-target ligands that modulate both A2A adenosine and D2 dopamine receptors.\n\nThe transformation from computer-aided drug design to artificial intelligence drug design represents more than technological advancement\u2014it signifies a fundamental shift in how pharmaceutical research approaches molecular discovery. Traditional QSAR modeling, while valuable, was limited by the availability of experimental data and computational constraints. Modern AI-driven platforms leverage vast datasets, sophisticated neural network architectures, and cloud computing resources to explore previously inaccessible regions of chemical space. These systems can generate entirely novel molecular structures, predict their biological properties, and optimize multiple parameters simultaneously, offering unprecedented opportunities for discovering breakthrough therapeutics.\n\nThe practical implications of these advances extend beyond academic curiosity to tangible improvements in drug discovery success rates. AI-designed drugs demonstrate success rates of 80-90% in Phase I clinical trials compared to 40-65% for traditionally discovered compounds. Development timelines have been compressed from the traditional 10+ years to potentially 3-6 years, while costs have been reduced by up to 70% through more intelligent compound selection and optimization strategies. These improvements are not merely theoretical\u2014they represent real-world applications that are already transforming how pharmaceutical companies approach drug discovery.\n\nContemporary virtual screening campaigns routinely process libraries containing billions of compounds, a scale that was inconceivable just a few years ago. The ZINC-22 database now contains 54.9 billion molecules, with 5.9 billion available in ready-to-dock 3D formats. The Enamine REAL database has grown from 170 million compounds in 2017 to over 6.7 billion compounds in 2024, representing synthetically accessible molecules that can be produced on demand. This exponential growth in chemical space availability, combined with advanced screening algorithms, creates unprecedented opportunities for discovering novel chemotypes and exploring previously uncharted regions of molecular diversity.\n\nMachine learning applications in virtual screening have evolved beyond simple classification tasks to sophisticated systems capable of predicting complex molecular properties. Graph neural networks treat molecules as connected atomic structures, enabling more accurate predictions of binding affinities and biological activities. Transformer-based architectures, originally developed for natural language processing, have been adapted to understand molecular \u201clanguages,\u201d generating novel compounds with desired properties while maintaining synthetic accessibility. These AI systems can simultaneously optimize multiple parameters\u2014potency, selectivity, safety, and pharmacokinetic properties\u2014in ways that traditional approaches could never achieve.\n\nThe democratization of computational drug discovery through platforms like Click2Drug has profound implications for global health research. Academic institutions, particularly in developing countries, now have access to sophisticated computational tools that were previously available only to well-funded pharmaceutical companies. Open-source initiatives and cloud-based platforms have lowered barriers to entry, enabling innovative research groups to contribute to drug discovery efforts regardless of their computational infrastructure limitations. This democratization has accelerated research into neglected tropical diseases and rare disorders that might not attract commercial interest but represent significant humanitarian challenges.\n\nThe convergence of artificial intelligence, structural biology, and chemical informatics is creating synergistic effects that amplify the impact of individual technological advances. AlphaFold\u2019s contribution of over 214 million protein structures has provided structural templates for targets that previously lacked experimental structures, enabling structure-based drug design for previously \u201cundruggable\u201d proteins. Combined with AI-powered virtual screening platforms, researchers can now approach novel targets with unprecedented confidence and efficiency. The integration of molecular dynamics simulations with machine learning algorithms enables more accurate modeling of protein flexibility and allosteric effects, addressing longstanding limitations in traditional docking approaches.\n\nLooking toward the future, the field is poised for even more dramatic transformations. Quantum computing applications in drug discovery, while still in early stages, promise to solve computationally intractable problems in molecular simulation and optimization. The integration of robotic synthesis with AI-guided design will create closed-loop systems where compounds are designed, synthesized, and tested automatically, with machine learning algorithms continuously refining their predictions based on experimental outcomes. These \u201cAI factories\u201d for drug discovery will represent the ultimate realization of computational drug design vision, where human creativity guides overall strategy while artificial intelligence handles the computational complexity of molecular optimization.", "# Crystal structure of a soluble cleaved HIV-1 envelope trimer\nAuthors: Julien, J.-P.; Cupo, A.; Sok, D.; Stanfield, R.L.; Lyumkis, D.; Deller, M.C.; Klasse, P.-J.; Burton, D.R.; Sanders, R.W.; Moore, J.P.; Ward, A.B.; Wilson, I.A.\nSource: Science, 2013\n\nThe requirement for the cleaved, native Env trimer to undergo conformational changes during receptor binding and fusion makes it metastable, which has substantially hindered both structure determination and vaccine development. The extensive N-linked glycosylation (on average 81 sites/trimer) creates additional complications for x-ray structural studies. Moreover, membrane-associated forms of Env are more difficult to express and purify in appropriate quantities and qualities than soluble versions. Our approach to these various problems has been to express soluble (i.e., truncated prior to the gp41 transmembrane domain), cleaved forms of trimeric Env (SOSIP gp140) that are engineered to improve their stability and homogeneity. Specifically, a disulfide bond (termed SOS) between gp120 residue 501 (HXB2 numbering) and gp41 residue 605 covalently links these subunits, while an Ile to Pro change at position 559 (termed IP) strengthens gp41-gp41 associations. A recent version of the SOSIP gp140 trimer, based on a Tier-2 subtype A virus (BG505), was further engineered to delete all but 4 residues of the hydrophobic membrane proximal external region (MPER) of gp41. Together, these various modifications allow the expression of a thermostable, non-aggregating and homogenous soluble Env trimer, BG505 SOSIP.664 gp140, suitable for structural characterization by x-ray crystallography (Fig. 1A). These trimers are reactive with a large panel of diverse broadly neutralizing antibodies (bnAbs), including those to quaternary epitopes, while being minimally reactive with non-neutralizing antibodies that preferentially recognize individual gp120/gp41 subunits and/or uncleaved, non-native trimer forms. The near-native antigenic properties of the BG505 SOSIP.664 gp140 trimer suggest that its structure resembles the native viral spike, although we cannot completely rule out slight conformational differences resulting from engineered features, such as truncation of the gp41 MPER and transmembrane domain. Here, we show that the BG505 SOSIP.664 gp140 trimers could be successfully crystallized with a highly potent bnAb, PGT122, that targets the glycan-dependent Asn332 (N332) supersite of vulnerability on gp120. These crystals allowed the structure of an Env trimer to be determined at a resolution of 4.7 \u00c5.\n\n## Structure determination\n\nThe BG505 SOSIP.664 construct was expressed in HEK 293S GnTI\u2212/\u2212 cells, yielding trimers enriched for homogeneous oligomannose (Man5\u2013Man9) glycans (see supplemental material). Following incubation of the purified trimers with a 6-fold molar excess (2-fold for the binding sites) of PGT122 Fab, the complex was treated with EndoH glycosidase (fig. S1, A and B) to truncate any accessible N-linked glycans (i.e., not buried or occluded by PGT122) to a single N-acetyl glucosamine (NAG) moiety that remains covalently attached to the Asn side chain. Crystals of the purified complex diffracted well, albeit anisotropically, to 3.7 \u00c5 along the c-axis, but to lower resolution along the other two axes (fig. S1, C and D). Merging diffraction data from two crystals resulted in a complete dataset to a maximum resolution of 4.7 \u00c5 (table S1). Phases were obtained by molecular replacement using integrative approaches in which a search model was generated from crystal structures of the unliganded PGT122 Fab variable region (PDB ID: 4JY5) and CD4-bound gp120 core (PDB ID: 3JWD), docked into our previous ~14 \u00c5 EM reconstruction of the same complex (EMDB: 5624). Only one complex was present in the asymmetric unit with 82% solvent, which simplified structure determination (fig. S2). Initial phases resulted in a well-defined electron density map that enabled subsequent placement of the high-resolution PGT122 Fab constant domains, and the gp120 V1/V2 (PDB ID: 3U4E) and V3 (PDB ID: 2ESX) structures. Previously uncharacterized elements in the trimer, such as the gp41 helices, the V1/V2/V3 loops and various Man5-9 glycans were visible in the electron density maps, as were residual NAG moieties attached to their respective Asn residues (fig. S3). Together, these identifiable features, along with aromatic side chains, aided in model building and refinement at this moderate resolution. In addition, the prominent new features for the Env trimer that we observed in the electron density maps are the same as those visualized in the accompanying 5.8 \u00c5 cryo-EM reconstruction of the same trimer in complex with bnAb PGV04 (fig. S4).\n\n## Architecture of the Env trimer\n\nThe soluble BG505 SOSIP.664 trimer adopts a compact mushroom shape reminiscent of the Env trimer pre-fusion closed conformation determined at 20 \u00c5 resolution by Electron Microscopy. Three PGT122 Fabs protrude vertically from the membrane-distal gp120 subunits, whereas the gp41 components are membrane-proximal and interspersed with the gp120 C1 and C5 elements (Fig. 1, B-D). Previous high-resolution crystal structures of core gp120 (PDB ID: 3JWD), scaffolded V1/V2 (PDB ID: 3U4E) and the variable domains of PGT122 Fab (PDB ID: 4JY5) fit well into the electron density of the trimer complex (Figs. 1 and 2 and fig. S5), and have C\u03b1 root mean square deviations (r.m.s.d.'s) of 1.3 \u00c5, 2.9 \u00c5 and 0.8 \u00c5 with the final trimer model, respectively. Thus, the core gp120 elements in the trimer adopt conformations similar to those observed in unliganded gp120 or gp120 in complex with various ligands (specifically CD4, b12, VRC01, PGT128, PGT135 and PG9/PG16). In addition, all previously described disulfide bonds in gp120 are present and all of the variable loops (V1-V5) are on the outside of the structure (Fig. 1D).", "# Tiny Pipes, Big Discoveries: How Microfluidics Are Accelerating Structural Biology\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-13 on marcdeller.com (categories: blog).*\n\n* Microfluidics and \u201clab-on-a-chip\u201d technology are automating and miniaturizing complex laboratory workflows in structural biology.\n* These platforms enable high-throughput protein crystallization and fragment screening, using far less sample material and time than traditional methods.\n* By streamlining sample preparation and data collection, microfluidics is helping to eliminate key bottlenecks and accelerate the pace of structure-based drug discovery.\n\n1. **The Disruptive Impact of Structural Biology on Biopharmaceutical Innovation:**\n   <https://www.pharmasalmanac.com/articles/the-disruptive-impact-of-structural-biology-on-biopharmaceutical-innovation>\n2. **Structural Biology and Molecular Modeling Market \u2013 Trends, Insights:**\n   <https://www.marketbusinessinsights.com/structural-biology-and-molecular-modeling-market>\n3. **2025 Trends in Biotech and Life Sciences Research:**\n   <https://go.zageno.com/blog/2025-trends-in-biotech-and-life-sciences-research>\n\nOne of the biggest bottlenecks in structure-based drug discovery has been the slow, painstaking process of preparing samples for analysis. For X-ray crystallography, this means growing high-quality crystals of a target protein, a process that can take months of manual effort and consume large amounts of precious protein material. This \u201cartisanal\u201d approach has limited the number of targets that can be pursued and slowed the pace of discovery for decades.\n\nMicrofluidics is changing the game by shrinking the entire laboratory workflow onto a single chip. These \u201clab-on-a-chip\u201d devices use tiny channels and valves to manipulate minuscule volumes of liquid with incredible precision. In structural biology, automated microfluidic platforms can set up thousands of different crystallization experiments in parallel, using just a fraction of the protein required by traditional methods. This high-throughput approach dramatically increases the chances of finding the perfect conditions for crystal growth and compresses timelines from months to days.\n\nThe impact of this technology is especially profound for fragment-based drug discovery, where scientists screen libraries of small molecular fragments to find starting points for new drugs. Microfluidics allows for the rapid, automated preparation and data collection needed to analyze how hundreds of these fragments bind to a target. By integrating this technology with robotics, AI-driven data analysis, and high-intensity synchrotron X-ray sources, we are creating a fully automated pipeline from gene to structure, breaking down old bottlenecks and dramatically increasing the efficiency of modern drug discovery.", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n**2b. And the part that cannot be returned.** Some memory never comes back while the process\nlives, however politely you ask: on a measured run the amount held floored at ~20GB after the\nfirst target and climbed ~1.9GB per target after, while a single target in a fresh process\nneeded up to 47.6GB of the 55.7GB available. Left alone the two eventually meet, which is why\none run had **no** memory failures in its first four targets and **three** in its last four.\nOnly ending the process frees it, so BoltzMaker starts a fresh one every few targets --\n`Targets per invocation`, default 4. Boltz skips work already done, so a fresh process costs one\nmodel load (~4 minutes against ~45 minutes of prediction) and recomputes nothing. Set it to 0\nfor the old behaviour of one long-lived process.\n\n**3. Half-precision where full precision was intended.** Boltz forces its most delicate\ncalculations to full precision because they overflow otherwise, but does so with an instruction\nnaming NVIDIA hardware explicitly -- so on a Mac all **19** of those protections were silently\ndoing nothing and the delicate steps ran at half precision anyway.\n\n**4. One bad target killing all the others.** A numerical failure in a single target aborted the\nentire run, and every target queued behind it was never attempted, with nothing in the output to\nsay so. A failure is now contained to the target that caused it.\n\n**5. The affinity step tripping over a missing file.** Boltz predicts binding affinity for all\nstructures at once after they are built. If any target had been skipped, that step looked for a\nfile which was never written and crashed, taking down the results of every target that had\nsucceeded. It now skips the incomplete ones and carries on.\n\n**6. Steering that could destroy a structure.** Boltz nudges atoms with simulated physical\nforces to keep the geometry sensible. When two atoms end up very close the repulsive force\nbecomes enormous, and after twenty rounds of nudging it can overflow to infinity -- which is\nthen added to *every* atom's position, turning the whole structure into nonsense. One target\nfailed this way twenty times in a row. An unusable force is now ignored rather than applied, and\nthe run reports how often it had to: a handful is noise, hundreds means that structure was less\ncarefully steered than its siblings and deserves a second look.\n\nGuards 4, 5 and 6 only ever act on values that are *already* broken, so any target that would\nhave worked before produces byte-for-byte the same structure and affinity. They can only rescue\ntargets that previously produced nothing.\n\n**6b. Steering that pulled a molecule apart, quietly.** Guard 6 stops an *infinite* force being\napplied; that was half the problem. Alongside those infinities sit finite but enormous values,\nand those were applied. One compound came out with four of its atoms 49, 58, 737 and 2147\nangstroms from the rest of the molecule -- the same four atoms, to within a few angstroms, in\nthree separate runs. The protein was fine and every score was excellent: 0.80 confidence, top of\nthe binder ranking. Nothing in the report mentioned the ligand was no longer a molecule, because\nnothing in the report looked at its geometry. Two limits now apply: a push far beyond anything\nelse in that step is treated as a singularity and scaled back, keeping its direction; and no\nsingle step may move an atom more than a tenth of a chemical bond. Together they take that\ncompound from 2154 angstroms across to 16.5 -- the same answer, to within two angstroms, as\nrunning it with steering off entirely. The instability is still there and still reported; it\nsimply can no longer throw atoms across the room.\n\n**7. Patches that were only applied one way in.** These are edits to the installed Boltz, and\nused to be applied only by `run_campaign.sh` -- so starting a campaign directly with\n`BoltzMaker.py all`, or rebuilding the environment (installing Boltz puts back its own untouched\nfiles), quietly reverted everything above, with nothing but a warning line in the preflight table\nthat nobody is awake at 3am to read. BoltzMaker now applies them itself before every run: a no-op\nwhen already in place, and if a repair *was* needed the preflight row says so\n(`repaired 3 this run`) rather than passing silently.\n\n**8. A run that hangs forever instead of failing.** The worst failure was the one that did not\nlook like a failure: Boltz goes quiet, its worker sits in a state the operating system will not\ninterrupt, it keeps its GPU memory, and it never exits or errors. Seen on a live campaign: 24\nminutes of silence holding 39.8GB with no work being done, which would have run to morning.\nBoltzMaker now watches how long since Boltz last said anything, and after an hour treats it as\nwedged, stops it, and lets the retry ladder rerun the affected targets one at a time in fresh\nprocesses. Nothing already computed is lost. This used to need a separate watchdog program\nalongside; it is now part of the run, so an unattended campaign needs nothing watching it.", "# GitHub README: AlphaFraud\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/AlphaFraud, 2026\n\n- **A positive control.** The page opens with the one thing the rest of the site does not: an *excellent* match. Human **[Artemis (DCLRE1C)](https://alphafraud.mdeller.com/examples#ex-6TT5_1)**, a genuinely novel post-cutoff nuclease (only 29% identity to anything AlphaFold trained on), was predicted blind to TM-score 0.996, lDDT 0.985 and 0.45 \u00c5 C\u03b1-RMSD at a justified mean pLDDT of 95, and AlphaFraud scores it FRAUD 0.02. It is the proof that the pipeline certifies good predictions, not just hunts for bad ones.\n- **An auto [\"example of the week\"](https://alphafraud.mdeller.com/examples).** After every weekly run the pipeline picks that release's most instructive confidently-wrong catch and writes its panel deterministically from the database (no LLM), badged and dated.\n- **A [complete weekly archive](https://alphafraud.mdeller.com/examples).** Every PDB release week since the 2018 cutoff is listed: weeks with a catch expand to a full panel with its C\u03b1-deviation ribbon, and the quiet weeks (where AlphaFold was never overconfident) are recorded too, collapsed to the most recent by default.\n- **Curated deep-dives.** Hand-authored panels across the failure spectrum: ordinary single-chain globular folds AlphaFold gets *subtly* wrong (right fold, wrong conformation or assembly \u2014 the [PP2A scaffold](https://alphafraud.mdeller.com/examples#ex-8UWB_3), [STIM1](https://alphafraud.mdeller.com/examples#ex-6YEL_1), [Aha1](https://alphafraud.mdeller.com/examples#ex-7DME_1), [WWP2](https://alphafraud.mdeller.com/examples#ex-6RSS_1)) and the dramatic whole-fold catastrophes ([TMEM106B](https://alphafraud.mdeller.com/examples#ex-7U14_1), [alpha-synuclein](https://alphafraud.mdeller.com/examples#ex-9A1Q_1), [transthyretin](https://alphafraud.mdeller.com/examples#ex-9BZS_1), [factor X](https://alphafraud.mdeller.com/examples#ex-6Q9F_2)). Each panel carries live metrics, an interactive 3Dmol.js viewer (deviation colouring plus the AlphaFold \"ghost\") and a structure / mechanism / disease write-up.\n- **Linked from the scatter.** The curated examples are overlaid on the [home fraud-quadrant plot](https://alphafraud.mdeller.com/) as pulsing red (failure) and green (accurate) markers, each labelled with its protein name and clickable straight through to its panel.\n\nEach panel is summarised below \u2014 click any title to open its live, interactive version. Metrics: **TM-score** and **lDDT** (1.0 = identical), **C\u03b1-RMSD** (0 \u00c5 = identical), **pLDDT** (AlphaFold's own confidence, 0\u2013100), **FRAUD** (confidence-weighted error, higher = worse), **novelty** (100% = wholly unseen before AlphaFold's cutoff).\n\n### \ud83d\udfe2 Protein Artemis (DCLRE1C) \u00b7 [`6TT5_1`](https://alphafraud.mdeller.com/examples#ex-6TT5_1)\n\n| Field | Detail |\n|---|---|\n| **Take-home** | A novel post-cutoff nuclease AlphaFold predicted blind almost perfectly \u2014 the control proving AlphaFraud certifies good predictions, not only failures |\n| **AlphaFold vs experimental** | <a href=\"https://alphafraud.mdeller.com/examples#ex-6TT5_1\"><img src=\"https://alphafraud.mdeller.com/ribbon/6TT5_1.svg\" alt=\"Experimental structure of 6TT5_1 coloured by C\u03b1 deviation from AlphaFold\" width=\"380\"></a><br><sub>Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \"ghost\" overlay</sub> |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-6TT5_1)** |\n| **KPIs** | TM **0.996** \u00b7 lDDT 0.985 \u00b7 pLDDT 95.2 \u00b7 C\u03b1-RMSD 0.45 \u00c5 \u00b7 FRAUD **0.02** \u00b7 novelty 71% |\n| **Stats** | 11 depositions, **0** confidently wrong \u00b7 SS-Q3 92.5% \u00b7 pLDDT\u2194lDDT +0.56 (calibrated) |\n| **Structure** | Metallo-\u03b2-lactamase + \u03b2-CASP nuclease domain with a Zn active site; the disordered regulatory C-terminus sits (correctly) outside the crystallised fold |\n| **Why AlphaFold gets it right** | A single, conserved globular domain with a deep MSA and one native conformation \u2014 AlphaFold's ideal regime; confidence is earned (PAE-overconfident 0.00) |\n| **Biology & disease** | NHEJ endonuclease that opens V(D)J hairpins with DNA-PKcs; loss of function causes radiosensitive SCID / Omenn; pursued as a radiosensitiser drug target |\n| **Key fact** | 29% identity to anything pre-cutoff yet GDT_TS 99.2 \u2014 genuine generalisation, not memorisation |\n\n### \ud83d\udd34 PP2A scaffold subunit A / PR65 (PPP2R1A) \u00b7 [`8UWB_3`](https://alphafraud.mdeller.com/examples#ex-8UWB_3)\n\n| Field | Detail |\n|---|---|\n| **Take-home** | Near-maximal confidence (pLDDT 97) on a flexible 15-HEAT solenoid whose curvature is wrong for 99% of residues |\n| **AlphaFold vs experimental** | <a href=\"https://alphafraud.mdeller.com/examples#ex-8UWB_3\"><img src=\"https://alphafraud.mdeller.com/ribbon/8UWB_3.svg\" alt=\"Experimental structure of 8UWB_3 coloured by C\u03b1 deviation from AlphaFold\" width=\"380\"></a><br><sub>Experimental structure coloured by C\u03b1 deviation from AlphaFold (blue = agree, red = diverges) \u2014 click for the interactive 3D viewer with the AlphaFold \"ghost\" overlay</sub> |\n| **Link to live example** | **[Open the interactive panel \u2192](https://alphafraud.mdeller.com/examples#ex-8UWB_3)** |\n| **KPIs** | TM 0.293 \u00b7 lDDT 0.311 \u00b7 pLDDT 96.8 \u00b7 C\u03b1-RMSD 15.0 \u00c5 \u00b7 FRAUD 0.77 \u00b7 novelty 0% |\n| **Stats** | 26 depositions, 1 confidently wrong (this one) \u00b7 pLDDT\u2194lDDT \u22120.13 (confidence carries no signal) |\n| **Structure** | 15 tandem HEAT repeats forming a horseshoe \u03b1-solenoid \u2014 a spring, not a fixed ruler |\n| **Why AlphaFold gets it wrong** | Local helices are right, but tiny per-repeat hinge errors compound across 15 repeats into a grossly wrong global shape \u2014 a curvature failure, not a fold failure |\n| **Biology & disease** | Sole scaffold of the PP2A tumour-suppressor phosphatase; cancer hotspots (P179R, R183, S256F\u2026) cluster in the B56-binding repeats |\n| **Key fact** | The source paper needed an experimental trimer structure to guide the prediction |", "# Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design\nAuthors: Irimia, A.; Serra, A.M.; Sarkar, A.; Jacak, R.; Kalyuzhniy, O.; Sok, D.; Saye-Francisco, K.L.; Schiffner, T.; Tingle, R.; Kubitz, M.; Adachi, Y.; Stanfield, R.L.; Deller, M.C.; Burton, D.R.; Schief, W.R.; Wilson, I.A.\nSource: PLoS Pathogens, 2017\n\n**Co-crystallization of 10E8-T117v2 complex with lipids.** The phosphatidic acid and phosphatidylglycerol lipids were purchased from Avanti Polar Lipids. Each lipid (06:0 PG and 06:0 PA, respectively) solubilized to 15 mM in 20 mM sodium acetate (pH 5.5) was mixed with 10E8/T117v2 complex to obtain a final concentration of 10 mg/ml protein and 8 mM lipid. Crystallization screening was performed with the JCSG/IAVI/Scripps high-throughput CrystalMation robot (Rigaku) at TSRI in a sitting drop vapor diffusion format. The 10E8/T117v2/06:0 PG crystals grew from a 1:1 (v/v) protein:reservoir solution drop equilibrated against 50% PEG 200, 0.1 M HEPES, pH 7.0, and the 10E8/T117v2/06:0 PA crystals from a drop equilibrated against 50% PEG 400, 0.2 M NaCl, 0.1 M CHES, pH 9.5. The crystals for each complex were cryo-protected with their respective reservoir solutions.\n\n**Crystallization of 10E8-T117v2 and 10E8 mutants-T117v2.** Small crystals of 10E8-T117v2 complex were obtained in a drop (50%:50% protein:reservoir) equilibrated against 0.1 M Tris pH 7.5, 15% PEG 6,000 reservoir. The crystals were crushed in the reservoir solution and seeded at a ratio of 16% in a drop (containing 50%:34% protein:reservoir) equilibrated against 0.1 M HEPES pH 7, 15% PEG 20,000 to produce diffraction-size crystals. The 10E8 mutant-T117v2 complexes concentrated to about 10 mg/ml were also screened for crystallization on our CrystalMation system by sitting drop vapor diffusion. The crystallization conditions and the cryoprotectant solutions used for each complex are reported in S2 Table.\n\n### Data collection, structure determination and refinement\n\nX-ray diffraction data were collected at APS 23ID-B and 23ID-D beam lines and at SSRL on 12-2 and 11-1 beam lines (S1 and S2 Tables) and were auto-indexed and processed with HKL-2000 or XDS. Molecular replacement was performed with Phaser using one of the 10E8 Fabs (PDB 4G6F) and the T117 scaffold (PDB 3LF6) as search models. Model rebuilding in Coot and refinement with Phenix were performed following an initial rigid body refinement step. The refinement cycles, for structures solved between 2.0 and 2.6 \u00c5, included refinement of individual atomic coordinates, cartesian simulated annealing, refinement of individual isotropic atomic displacement parameters and optimization of X-ray/stereochemistry and X-ray/ADP weights. For the 1.6 \u00c5 structure of 10E8 mutant 1 in complex with T117v2, refinement of individual atomic coordinates, cartesian simulated annealing, occupancy and individual atomic displacement parameters refinement with anisotropic ADP for protein atoms and isotropic ADP for solvent were performed as well as optimization of X-ray/stereochemistry and X-ray/ADP weights. X-ray diffraction and refinement statistics are reported in S1 Table for the wild-type 10E8 mature-T117v2 complexes bound to lipids and in S2 Table for 10E8 mutant-T117v2 complexes. Structure figures were generated with PyMOL. The atomic coordinates and structure factors of 10E8-T117v2 structures have been deposited in the Protein Data Bank, with the accession codes: 5T6L (for 10E8-T117v2) and 5T85, 5T80 for co-crystals with 06:0 PG and 06:0 PA, respectively, and those for 10E8 mutants-T117v2 structures with the accession codes: 5SY8 (for 10E8 mutant 1-T117v2), 5TFW (for 10E8 mutant 2-T117v2), 5T29 (for 10E8 mutant 3-T117v2) and 5T5B (for 10E8 mutant 5-T117v2).\n\n### Viral membrane model using CHARMM\n\nA model of the trimeric MPER epitope-transmembrane region of the gp41 was constructed using PDB 2MOM as a template as described previously, with the orientation of the MPER epitope modeled based on crystal structures determined in this study. The structural information on the 10E8 Fab and the PG fragment was transferred to the trimeric model by superposing the MPER epitope region of the T117v2 scaffold to the corresponding region in the model. The crystallographic 06:0 PG lipid fragment was extended to the size of a 1,2-dihexadecanoyl-sn-glycero-3-phospho-(1'-rac-glycerol) (DPPG) molecule by adding the lipid tails in the direction perpendicular to the plane that includes the light-chain surface residues, the head group of the PG fragment, and Lys683. The putative transmembrane region model constructed solely to anchor the lipid bilayer was then used in membrane building with CHARMM. A rectangular box (x = y = 157.7 \u00c5) was used to generate a heterogeneous bilayer containing 400 lipids on the upper leaflet and 434 lipids on the lower leaflet of the membrane, by replacement method. The HIV-1 membrane lipid composition was taken into account when choosing the composition of the bilayer. The Monte Carlo method was used to place counter potassium ions and NVT (constant volume) ensemble was used during six equilibration steps at a constant temperature of 303 K. The final model has Lys683 of the MPER and the head group of the crystallographically observed lipid embedded into the head group region of the membrane outer leaflet, while the side chains of the residues of the aforementioned light-chain surface are embedded in, or touching, this hydrophilic layer. The model remained stable during equilibration steps.\n\n## Supporting information\n\n*S1 Fig. Superposition of the 10E8-MPER epitope peptide onto the 10E8-T117v2 structure.* The 10E8 Fab-peptide complex (PDB 4G6F) is colored green. 10E8 Fab in complex with the T117v2 scaffold (gray) is colored brown. The region of the T117v2 scaffold that contains the MPER epitope is highlighted in red.", "# Crystal structure of a soluble cleaved HIV-1 envelope trimer\nAuthors: Julien, J.-P.; Cupo, A.; Sok, D.; Stanfield, R.L.; Lyumkis, D.; Deller, M.C.; Klasse, P.-J.; Burton, D.R.; Sanders, R.W.; Moore, J.P.; Ward, A.B.; Wilson, I.A.\nSource: Science, 2013\n\nThe PGT121 family of antibodies potently neutralizes ~70% of circulating HIV-1 viruses and PGT121 is highly effective in protecting against mucosal challenge in a macaque model of HIV-1 infection. The crystal structure presented here agrees well with our lower resolution negative-stain EM reconstruction of the same PGT122-trimer complex (fig. S8), but now allows a more complete description of the epitope. As initially predicted, the N332 Man8/9 glycan that is a central feature of the PGT122 epitope sits directly at the junction between the light chain complementarity determining region 2 (LCDR2) and the 26-residue heavy chain CDR3 (HCDR3) (Fig. 4A and fig. S9). Residues in the LCDR1, LCDR3 and light chain framework 3 (LFR3) regions of PGT122, which were predicted by alanine-scanning mutagenesis to be important for HIV-1 neutralization, interact directly with the gp120 V3 base near Ile323-Arg327, a key part of the epitope. The trimer structure also helps explain why PGT121-class bnAbs depend on V1/V2 residues for locking gp120 into the pre-CD4 conformation; in fact, gp120 V1 residues 135-139, including the glycan at N137, come into close proximity to HCDR1, HCDR2, HCDR3 and LCDR3 of PGT122 (Fig. 4A and fig. S9). Indeed, the N137 oligomannose glycan in the trimer structure superimposes almost perfectly with the bi-antennary complex glycan that occupies the groove formed by the PGT121 heavy chain CDRs in previous crystal structures (Fig. 4A and B). Alanine scanning mutagenesis of gp120 glycans clearly indicates that bnAbs of the PGT121 family depend on N137 for neutralizing the BG505 virus (Fig. 4C and fig. S10). Locating this key residue in the crystal structure has allowed us not only to delineate the full PGT122 epitope, but also to confidently determine the relative position of V1/V2 elements at the trimer apex. In addition to N137 and N332, the N156 and N301 glycans are also protected from EndoH deglycosylation by PGT122. Although the composition and nature of the glycoforms on native Env are still uncertain, PGT122 recognition is compatible with the presence of a complex/hybrid carbohydrate at position N156/N173, as observed in the crystal structure of a V1/V2 scaffold with bnAb PG16 (Fig. 4B). The N156 and N301 glycans do have a clear involvement, albeit a minor one, in neutralization by PGT121-family bnAbs (Fig. 4C and fig. S10). Thus, the Env trimer structure reveals the complexity of the PGT122 epitope, which involves the V1 and V3 base and four glycans (N137, N156/N173, N301 and N332).\n\n## Revisiting glycan-dependent bnAb epitopes in the context of the HIV-1 Env trimer\n\nThe recent isolation and structural characterization of potent glycan-reactive bnAbs has emphasized the need to precisely define the composition of various glycans that are sites of vulnerability on Env. The trimer structure allows us to reevaluate previous crystal structures of glycan-dependent bnAbs that were obtained with monomeric core gp120 or epitope scaffolds. When viewed in the context of the trimer, the gp120 epitopes for the N160-dependent (PG9/PG16) and N332-dependent (PGT122, PGT128 and PGT135) epitopes are in close proximity (Fig. 5A). Once considered separate sites of vulnerability (Fig. 5B), these two epitope clusters are larger than were first thought and overlap to some extent, helping to explain why some of these antibodies compete with one another in certain binding assays (Fig. 5C). All of these glycan-dependent bnAbs use their six CDR loops, in addition to framework regions, when recognizing their epitopes. As previously suggested, PG9 and PG16 would interact with two N160 glycans across different protomers in the trimer apex (Fig. 5D). In addition, while one side of the elongated HCDR3 hammerhead inserts into the trimer apex, the other is available to interact with elements of V3 (Fig. 5D) and may help explain why V3 residues influence neutralization by PG16. The trimer crystal structure also suggests that HCDR1 and HFR3 of PG9/PG16 may interact with the V3-protecting N197 glycan on an adjacent protomer (Fig. 5D).\n\nPGT128 would also make contacts with Env elements additional to those previously reported in the crystal structure with a gp120 outer domain lacking V1/V2 (Fig. 5E). Thus, LCDR1, LCDR2 and LFR3 of PGT128, for which no previous role could be assigned, are now predicted to be in close proximity to V1 elements, including N137 and N156, whereas a change in the V1 loop orientation is required to allow PGT135 binding (Fig. 5F). Indeed, a V1 orientation closer to that observed in the V1/V2 scaffold or in the high-resolution cryo-EM trimer structure would remove a clash with N137 and facilitate PGT135 interaction with V1. A flexible role for V1 might help to explain the plasticity of PGT135-Env interactions seen in previous EM studies.\n\n## Conclusion\n\nThe crystal structure of the soluble BG505 SOSIP.664 trimer is in excellent agreement (fig. S4) with the accompanying cryo-EM structure of the same Env construct as a complex with PGV04 Fab. Structures determined using these two independent techniques show complete concordance in the major new features that are visualized at the present level of 5-6 \u00c5 resolution (fig. S4). We have defined the overall architecture of the soluble trimer, as well as the secondary, tertiary and quaternary interactions between gp120 and gp41 that are involved in its assembly. The trimer is relatively tightly packed, especially in gp41, but with a small opening between the Env apex and the top of the central gp41 helices. The gp120 subunits are held together, at least in part, by association of the V1/V2/V3 regions at the apex of the trimer. The gp41 central helices provide the main stabilizing contact between gp41 and gp120. Finally, the complete definition of how neutralization epitopes are presented in the context of the trimer, here and in the accompanying manuscript, should help the design of new immunogens as candidate vaccines.\n\n## Materials and Methods", "# Crystal structure of a soluble cleaved HIV-1 envelope trimer\nAuthors: Julien, J.-P.; Cupo, A.; Sok, D.; Stanfield, R.L.; Lyumkis, D.; Deller, M.C.; Klasse, P.-J.; Burton, D.R.; Sanders, R.W.; Moore, J.P.; Ward, A.B.; Wilson, I.A.\nSource: Science, 2013\n\nMolecular replacement was performed using Phaser with a composite search model comprised of: 1) the variable domain of PGT122 at 1.8 \u00c5 (PDB ID: 4JY5); 2) CD4-bound gp120 core at 2.6 \u00c5 (PDB ID: 3JWD); and 3) these components docked in a trimeric arrangement using the previously reported ~14 \u00c5 EM reconstruction of the PGT122 Fab in complex with the BG505 SOSIP.664 trimer (EMDB: 5624). The molecular replacement solution indicated only one copy of the trimer-antibody complex was present in the asymmetric unit with a TFZ = 6.2 and an LLG = 1,344 (resulting in an ~82% solvent content (Vm = 6.87 \u00c53/Da) based on a molecular weight of ~400 kDa for the partially deglycosylated complex). Additional components, such as the PGT122 Fab constant domain, gp120 V1/V2/V3, and gp41 helices were added manually based on clear electron density for these elements in the initial maps (fig. S3). Rigid-body, grouped-B-factor, XYZ, real-space and TLS refinements were carried out in PHENIX using NCS-restraints, as well as secondary structure and reference model restraints imposed from high-resolution structures. Model building was performed using Coot. Secondary structure was determined using STRIDE. Slightly higher crystallographic R-values for the trimer structure compared to the mean reported for other structures to the same resolution likely results from the relatively poor diffraction properties of the crystals and the anisotropy of the resulting data. We also note that ~20% of the trimer is not modeled (N-terminus of gp120, V2, V4, C-terminus of gp120, N-terminus of gp41, gp41 disulfide loop region, N-terminus of gp41 HR2), which also contributes to higher R-values. BG505 SOSIP.664 trimer elements included in the deposited model are summarized in table S2. Structure validation was performed using Molprobity and refinement statistics are reported in table S1.\n\n### Pseudovirus production and neutralization assay\n\nThe BG505 env gene used to make Env-pseudoviruses lacks the N332 glycosylation site, but the viruses are still neutralized by antibodies of the PGT121 family. To assess the importance of specific glycosylation sites on neutralization by these antibodies, site-directed mutations were introduced into the env gene using QuikChange according to the manufacturer's protocol (Stratagene, La Jolla, CA). Mutants were verified by DNA sequence analysis (Eton Bioscience, San Diego, CA). Pseudoviruses were generated by transfection of 293T cells for 72 h with an Env-expressing plasmid and an Env-deficient genomic backbone plasmid (pSG3\u0394Env) in a 1:2 ratio, using Fugene 6 (Promega). A single-cycle neutralization assay with TZM-bl target cells was performed as described previously. Briefly, TZM-bl cells were seeded in a 96-well flat bottom plate at a concentration of 20,000 cells/well. Pseudoviruses were preincubated with serial dilutions of antibody for 1 h at 37 \u00b0C before infection of the cells. Luciferase reporter gene expression was quantified 48 h after infection by cell lysis and the addition of Luciferase substrate (Promega). To determine IC50 values, dose-response curves were fitted using nonlinear regression.\n\n## Figure captions\n\n*Fig. 1.* Overall architecture of a soluble, cleaved, recombinant HIV-1 Env trimer in complex with bnAb PGT122. (A) Schematic of the soluble, cleaved, recombinant HIV-1 Env BG505 SOSIP.664 construct in comparison to full-length gp160. N-linked glycans are shown and numbered on their respective Asn residues. The constant (C1-C5) and variable (V1-V5) regions in gp120 and the fusion peptide (FP), HR1 and HR2 helices, membrane proximal external region (MPER), transmembrane (TM) and cytoplasmic (CT) elements in gp41 are indicated. The SOSIP mutations are shown in red, as well as the added N332 glycan site. The color coding is as in B. (B) Side view of the soluble Env trimer complex with PGT122 showing two of the three Env gp140 protomers associated with PGT122 Fab (blue). A 2Fo-Fc electron density map contoured at 1.0 \u03c3 is shown as a gray mesh around the leftmost gp140 protomer. The membrane to which gp41 is attached would be at the bottom of the figure. (C) Side view of the Env trimer. For one of the three protomers on Env, core gp120 is shown in yellow, whereas V1/V2 and V3 regions are highlighted in orange and red, respectively. The main gp41 helical elements are colored in different shades of green. Protein components are rendered according to their secondary structure and glycans are depicted as spheres. (D) View of Env down the trimer axis. Loops of high variability in gp120 (V1-V5) all map to the periphery of the trimer and are labeled. Glycans have been omitted for clarity. Dashed lines indicate the location of gp120 V2 and V4 loops for which electron density was absent or ambiguous. The figure was generated with Pymol.", "# deepcovalent: Pipeline \u2014 DeepCovalent\nAuthors: deepcovalent\nSource: https://www.deepcovalent.com, 2026\n\n## DeepCovalent Therapeutic Pipeline\n\n#### **In Silico\u2013Validated Novel Compounds Generated by OS3D\u2122**\n\n*We intend to advance early pre-clinical development of these compounds, through at least (i) in vitro pharmacology and receptor characterization, and (ii) pharmacokinetics and absorption studies in animals (including oral bioavailability), in order to drive compound valuation for potential licensing to or acquisition by pharma partners.*\n\n*We have identified an additional eleven backup compounds for potential pre-clinical development in our pipeline.*\n\nOur GLP-1R small-molecule candidates were validated using a comprehensive *in-silico* framework, including structure-based docking and binding analysis, receptor conformational assessment, MD simulations with stability analysis, physics-based MM/GBSA free-energy calculations, off-target selectivity modeling, ADMET profiling, and synthetic accessibility scoring.\n\nCompared to reference compounds such as orforglipron (Eli Lilly), danuglipron (Pfizer), and lotiglipron (Pfizer), our compounds exhibit:\n\n* Significantly fewer predicted ADMET risks\n* Improved predicted off-target selectivity\n* Multiple structurally diverse candidates with distinct small-molecule chemotypes\n* Stable predicted binding complexes by MD trajectory analysis and MM/GBSA free-energy calculations\n* Strong synthetic accessibility, supporting rapid synthesis and progression\n\n#### **About Our GLP-1R Compounds**\n\n[request white paper](/contact)", "# Revolutionizing Protein Crystallography: The Synergy of TELSAM-DARPin Fusion and Electric Field Optimization\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-10-12 on marcdeller.com (categories: blog, sb).*\n\n* [Engineering Protein Linkers for Optimized TELSAM-DARPin Fusion and Crystallization With Electric Field-Based Polymer Orientation](https://scholarsarchive.byu.edu/etd/11022/)\n\n* **Rapid Formation**: Short semi-flexible linkers enable crystal formation within 24 hours, dramatically reducing typical crystallization timelines from weeks to days\n* **Ultra-low Concentrations**: TELSAM fusions crystallize at 0.1 mg/mL, enabling structure determination from microgram-scale protein samples previously considered insufficient\n* **Electric Field Enhancement**: Weak AC fields (6 V/mm) expand crystallization boundaries while suppressing competing liquid-liquid phase separation pathways\n* **Linker Optimization**: Pro-Ala semi-flexible linkers outperform rigid helical fusions and flexible polyglycine connections for DARPin targets specifically\n* **Polymer Adaptability**: TELSAM helical structures adjust rise parameters (minimum 46.9 \u00c5) to accommodate diverse target protein geometries without direct inter-polymer contacts\u200b\n* **Avidity Amplification**: Fusion systems stabilize weak crystal contacts through massive numbers of minimal intermolecular interactions, overcoming traditional crystallization barriers\u200b\n\n1. **Engineering Protein Linkers for Optimized TELSAM-DARPin Fusion and Crystallization With Electric Field-Based Polymer Orientation:** [Nawarathnage, Supeshala Dilrukshi, et al.](https://scholarsarchive.byu.edu/cgi/viewcontent.cgi?article=12031&context=etd), BYU ScholarsArchive\n2. **The Effects of Electric Fields on Protein Phase Behavior and Protein Crystallization:** [Ray, D., et al.](https://pubs.acs.org/doi/10.1021/acs.jpclett.4c01744), Journal of Physical Chemistry Letters\n3. **Designed Ankyrin Repeat Proteins: A New Class of Viral Entry Inhibitors:** [G\u00fcttler, T., et al.](https://pmc.ncbi.nlm.nih.gov/articles/PMC9611651/), PMC\n4. **Optimal TELSAM-Target Protein Linker Character is Target Protein Dependent:** [Nawarathnage, S.D., et al.](https://pubmed.ncbi.nlm.nih.gov/40950120/), PubMed\n5. **Electric-field-stimulated protein mechanics:** [Hekstra, D.R., et al.](https://pmc.ncbi.nlm.nih.gov/articles/PMC5730412/), Nature Communications\n6. **Crystals of TELSAM\u2013target protein fusions that exhibit minimal lattice contacts:** [Anderson, M.J., et al.](https://pmc.ncbi.nlm.nih.gov/articles/PMC8889177/), Open Biology\n\n**Advanced Linker Engineering Transforms Challenging Structural Biology into Predictable High-Resolution Crystallization**\n\nRecent breakthroughs in protein crystallization technology have converged at the intersection of engineered fusion systems and electric field manipulation, creating unprecedented opportunities for structural biologists to solve challenging protein structures. The systematic optimization of TELSAM (Translocation ETS Leukemia Sterile Alpha Motif) fusion proteins with Designed Ankyrin Repeat Proteins (DARPins) has emerged as a transformative approach, particularly when combined with electric field-based polymer orientation techniques that dramatically accelerate crystallization kinetics.\n\nThe revolutionary TELSAM crystallization chaperone represents a paradigm shift from traditional crystallization approaches. Unlike conventional methods that rely on trial-and-error screening, TELSAM forms pH-dependent polymers that create highly ordered crystal lattices when fused to target proteins. This approach has demonstrated remarkable success with DARPin targets, achieving crystal formation within 24 hours and enabling structure determination at protein concentrations as low as 0.1 mg/mL. The technology addresses one of structural biology\u2019s most persistent challenges: the crystallization of proteins that can only be produced in microgram quantities or exhibit limited solubility.\n\nCritical to the success of TELSAM-DARPin fusions is the strategic design of connecting linkers between the crystallization chaperone and target protein. Systematic studies have revealed that short semi-flexible and rigid linkers consistently produce high-quality crystals with DARPin targets, while flexible linkers perform optimally with other protein domains. The choice between Pro-Ala semi-flexible linkers, polyglycine flexible connections, or helical rigid fusions directly influences crystal morphology, diffraction quality, and formation kinetics. Removing the traditional 10xHis purification tag further enhances crystallization rates and improves crystal quality by eliminating conformational constraints that can impede optimal protein packing.\n\nThe integration of electric field techniques provides an additional dimension of control over protein crystallization processes. Weak alternating current electric fields (6 V/mm at 1 kHz) significantly expand the crystallization boundary while reducing nucleation induction times and enhancing crystal growth rates. These field-induced effects arise from modifications to protein-protein interaction potentials, where electric forces on charged amino acid groups alter protein conformations and expose hydrophobic surfaces that strengthen intermolecular attractions. Remarkably, the electric field preferentially promotes crystallization over liquid-liquid phase separation, creating a more direct pathway to ordered crystal formation.", "# Lipid interactions and angle of approach to the HIV-1 viral membrane of broadly neutralizing antibody 10E8: Insights for vaccine and therapeutic design\nAuthors: Irimia, A.; Serra, A.M.; Sarkar, A.; Jacak, R.; Kalyuzhniy, O.; Sok, D.; Saye-Francisco, K.L.; Schiffner, T.; Tingle, R.; Kubitz, M.; Adachi, Y.; Stanfield, R.L.; Deller, M.C.; Burton, D.R.; Schief, W.R.; Wilson, I.A.\nSource: PLoS Pathogens, 2017\n\nThe HIV-1 envelope protein (Env), a hetero-trimer of non-covalently linked gp120 and gp41 subunits, is the target of broadly neutralizing antibodies (bnAbs). BnAbs recognize several sites of vulnerability on Env. The membrane proximal external region (MPER) of Env is one of its most conserved regions and, hence, the focus of vaccine and therapeutic design efforts. Previously, several models have been proposed for the MPER orientation with respect to the viral membrane and for the mechanism by which MPER antibodies approach their respective epitopes in vivo. Several reports suggested that the MPER is relatively inaccessible to antibodies in the pre-fusion conformation and is exposed only transiently after CD4 binding. Other studies suggested that the MPER is partially laterally inserted into the viral membrane in its pre-fusion form and that neutralization would require antibodies to \"extract\" the MPER from the membrane. Early cryo-electron tomography (cryoET) of native HIV-1 and SIV Env on virions interpreted the MPER and transmembrane (TM) region as three independent helices organized in a tripod-like fashion followed by a turn at Lys683 (the last residue of the gp41 ectodomain; UNIPROT ID: Q70626, HIV1-LW123 numbering). In other cryoET studies, the gp41 stem was proposed to adopt a compact stalk organization within the trimer suggesting instead an extended, intertwined helical architecture for the MPER and TM region. An extended helical conformation was observed in a recent NMR structure of a construct spanning the MPER and part of the gp41 TM domain that showed no break in helicity at Lys683 and also in the NMR structure of the gp41 TM, which revealed a triple-helix, quaternary TM organization in bicelles. X-ray crystallography and EM studies have described in atomic detail the structure of a soluble, stabilized Env construct (BG505 SOSIP.664 gp140 trimer), but these SOSIP structures lack the MPER, TM and cytoplasmic domains. In a recent 4.2 \u00c5 resolution cryo-EM structure of a native HIV-1 Env trimer (\u0394CT) containing the MPER and TM domains in complex with antibody PGT151, the micelle-embedded TM domain could not be resolved, but the structure also suggested that the MPER may be inaccessible in the pre-fusion form of Env. Furthermore, in the presence of MPER antibody 10E8, an 8.8 \u00c5 resolution cryo-EM structure illustrated that the three MPER epitope regions within the trimer form a triple helix and the 10E8-bound Env appears to be elevated off the micelles. This elevation in comparison to the pre-fusion form suggested how the MPER may be engaged by 10E8, but the presence of detergent and the lack of a membrane in that study limited the ability to draw definitive conclusions about 10E8 engagement of native Env on virions. Due to the proximity of MPER to the membrane, MPER binding antibodies are thought to interact with the membrane and, indeed, some have been shown to interact with lipids.\n\nSeveral bnAbs target the MPER, with 10E8 and 4E10 neutralizing about 98% of all HIV-1 subtypes tested. 10E8 is the most potent of the MPER antibodies and lacks the polyreactivity of 4E10. 10E8 and 4E10 target the same helical epitope (C-terminal MPER residues 671\u2013683) but differ in their modes of binding. Residues 672\u2013674 adopt a 3-10 helical turn when engaged by 4E10, but are part of a continuous \u03b1 helix (region 671\u2013683) when bound by 10E8. The 2F5 epitope (residues 656\u2013671), which is N-terminal to the 10E8 and 4E10 epitopes, adopts an extended conformation with a type 1 \u03b2-turn and the Z13e1 epitope (residues 668\u2013677) consists of linked helical turns. Although these antibodies target linear MPER epitopes on gp41, binding to individual lipids or liposomes also suggests that their complete epitopes include viral membrane components. Indeed, studies of MPER antibodies suggest a role for their complementarity determining region (CDR) H3 in binding to viral membrane, but no clear picture has emerged as to how each antibody is oriented with respect to the MPER and viral membrane during engagement. To determine which regions of 10E8 might interact with the HIV-1 membrane, and motivated by our recent findings regarding 4E10 interaction with lipids that compose the viral membrane, we undertook a crystallographic study of 10E8 in complex with lipids and an epitope scaffold. We complemented our observations of 10E8-lipid binding with binding and neutralization experiments as well as structure determination of several 10E8 mutants. Our combined results have identified which 10E8 regions interact with the viral membrane and indicated that the MPER likely adopts an upright orientation with respect to the viral membrane during antibody engagement, as we also proposed for 4E10. Our high-resolution structural study increases our understanding of the relative MPER location, orientation and conformation during MPER antibody binding, and provides insights for the design of immunogens and therapeutic antibodies.\n\n## Results\n\nTo aid in characterization of the complete 10E8 epitope consisting of MPER and lipids, we designed an epitope-scaffold that presented the MPER on a stable framework in the 10E8-bound conformation derived from previous crystal structures. T117v2 is a variant of the previously reported T117 epitope-scaffold that presents the 10E8 core helical epitope residues and binds to mature 10E8 with 29 pM affinity (Table 1). Whereas T117 was originally designed as an epitope-scaffold for 4E10, T117v2 was modified by M125K and D126A mutations to accommodate binding by 10E8. The crystal structure of this scaffold in complex with 10E8 was solved at 2.1 \u00c5 resolution and revealed similar conformations for the antibody and MPER epitope (S1 Fig) as in the 10E8-peptide structure (PDB 4G6F) to which it aligns with a C\u03b1 r.m.s.d. of 0.31 \u00c5 calculated from the Fab variable regions and epitope residues of the respective structures.\n\n*Table 1.* 10E8 variant binding to the T117v2 scaffold.", "# MARCUS: Revolutionizing Chemical Literature Curation with AI-Driven Annotation and Recognition\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\n*Published 2025-08-22 on marcdeller.com (categories: blog, sb).*\n\n* MARCUS seamlessly integrates text annotation with multi-engine optical chemical structure recognition for automated curation.\n* It uses a fine-tuned GPT-4 model and an ensemble of DECIMER, MolNexTR, and MolScribe to optimize chemical entity and structure extraction accuracy.\n* The web-based platform supports batch processing and direct submission to the open-access COCONUT natural products database.\n* MARCUS significantly reduces manual curation time by automating the conversion of unstructured literature into machine-readable chemical data.\n* Open-source code and comprehensive documentation foster community collaboration and future tool enhancements.\n* Though automated, expert human validation remains essential for ensuring data quality and accuracy.\n\n1. Rajan K., Wei\u00dfenborn V., Lederer L., Zielesny A., Steinbeck C. **MARCUS: Molecular Annotation and Recognition for Curating Unravelled Structures**, ChemRxiv, 2025:\n   <https://chemrxiv.org/engage/chemrxiv/article-details/686b86cb1a8f9bdab5017104>\n2. Rajan K., et al., **A review of optical chemical structure recognition tools**, Google Scholar, 2025:\n   <https://scholar.google.com/citations?user=-1Kqb3IAAAAJ&hl=en>\n3. **Software Tools and Resources 2025: Honoring Margaret Dayhoff, J**. Proteome Res., 24(3), 2025:\n   <https://pubs.acs.org/page/jprobs/vsi/software-tools-2025>\n4. **Understanding Marcus Theory in Coordination Chemistry**, Number Analytics Blog, 2025:\n   https://www.numberanalytics.com/blog/ultimate-guide-marcus-theory-coordination-chemistry\n\nIn an era marked by the rapid expansion of chemical research publications, the challenge of extracting and curating molecular information from vast, unstructured scientific literature has become increasingly critical. The innovative tool known as MARCUS (Molecular Annotation and Recognition for Curating Unravelled Structures) represents a major leap forward in addressing these challenges through the integration of cutting-edge artificial intelligence and automated recognition technologies. Developed as a web-based platform, MARCUS streamlines the workflow from unstructured PDF documents to actionable data entries in open-access chemical databases, specifically targeting natural product chemistry. By employing a fine-tuned GPT-4 model for precise chemical entity extraction along with an ensemble of Optical Chemical Structure Recognition (OCSR) engines\u2014DECIMER, MolNexTR, and MolScribe\u2014MARCUS optimizes accuracy in converting chemical structures into machine-readable formats. This integration marks a significant improvement over existing isolated solutions that often suffer from limited scalability.\n\nUsers experience a seamless interface combining document upload, text annotation, structure segmentation, and direct submission to the COCONUT natural products database, significantly accelerating the traditionally manual and error-prone curation process. MARCUS\u2019s architecture, including a user-friendly frontend backed by robust containerized services, supports concurrent users and ensures stable performance, which is key for broader adoption by researchers. The platform not only prioritizes precision in data extraction but also embraces open science ideals through its open-source code, accessible models, and extensive documentation. This openness encourages community-driven development and invites collaboration that could expand MARCUS\u2019s functionality beyond natural products. While expert human oversight remains essential for validation, MARCUS reduces the cognitive load on curators by converting about two-thirds of structure images accurately on first pass\u2014a remarkable achievement considering the complexity of chemical imagery. By bridging the divide between exponentially growing chemical literature and the structured demands of modern AI-driven discovery workflows, MARCUS is swiftly becoming an indispensable asset that promotes FAIR data principles and empowers researchers worldwide to harness literature with unprecedented efficiency, paving the way for accelerated innovation in chemistry.", "# PocketTide \u2013 An interactive dashboard for GLP-1R structures\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-14 on marcdeller.com (categories: vibe-code).*\n\n> **An interactive dashboard for GLP-1R structures.**\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/PocketTide](https://github.com/bellcheddar/PocketTide) |\n\n---\n\nPocketTide is a browser-based dashboard for exploring structures of the glucagon-like peptide-1 receptor (GLP-1R), one of the most important and intensively pursued targets in metabolic disease. It brings the growing body of GLP-1R structural data into a single, navigable view so that structures, their key features, and their differences can be surveyed at a glance.\n\nWhy it matters: GLP-1R sits at the centre of the incretin-mimetic revolution in diabetes and obesity therapeutics, and its structural landscape (peptide-bound, small-molecule-bound, G-protein-coupled, in multiple conformational states) is expanding rapidly. Keeping that picture organised is hard when the relevant structures are scattered across the PDB. PocketTide is useful for structure-based drug design teams who need to triage GLP-1R structures quickly, compare binding modes and receptor states, and orient a project against the available structural evidence, all from a single self-contained HTML file with no install and no build step.\n\n---\n\n## \u2728 What it does\n\n| Capability | Description |\n| --- | --- |\n| \ud83d\uddc2\ufe0f **Structure browser** | Survey GLP-1R structures in one organised dashboard view |\n| \ud83d\udd2c **Feature summary** | Surface key attributes per structure (method, ligand, receptor state) |\n| \u2696\ufe0f **Comparison** | Place structures side by side to compare binding modes and conformations |\n| \ud83c\udfaf **Target focus** | Purpose-built around GLP-1R rather than a generic structure viewer |\n\n> \u2139\ufe0f This README describes PocketTide at the capability level. For the exact panels and data fields in your build, open the app and explore the live interface.\n\n---\n\n## \ud83d\ude80 Running locally\n\nPocketTide is delivered as a single self-contained HTML file. The simplest way to use it is to open the file directly in a modern browser:\n\n```\nopen PocketTide.html        # macOS\nstart PocketTide.html       # Windows\nxdg-open PocketTide.html    # Linux\n```\n\nIf the dashboard pulls structural data from external APIs (e.g. RCSB PDB), browser file-origin restrictions may block those requests when opened as a `file://` URL. In that case, serve it from a lightweight local HTTP server:\n\n```\n# Python 3\npython -m http.server 8080\n\n# Node.js\nnpx serve .\n```\n\nThen navigate to `http://localhost:8080/PocketTide.html`.\n\n---\n\n## \ud83d\udd27 Tech stack\n\n| Layer | Technology |\n| --- | --- |\n| UI | Vanilla HTML + CSS, no frameworks |\n| Logic | Vanilla JavaScript (ES6+) |\n| Distribution | Single self-contained HTML file, no build step |\n| Data | GLP-1R structures (PDB) |\n\n---\n\n## \ud83e\udded Context\n\nPocketTide is part of a wider portfolio of structure-focused tools and dashboards. It pairs naturally with structure-based drug design work on GLP-1R and related class B GPCR targets.\n\n---\n\n## \ud83d\udcc4 Licence\n\nSee repository for licence terms.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/PocketTide](https://github.com/bellcheddar/PocketTide) |", "# Crystal Structure of [KS3][AT3] Didomain from Module 3 of 6-Deoxyerythronolide B Synthase in Complex with Antibody Fragment (Fab)\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-03-08 on marcdeller.com (categories: structures).*\n\n* **PDB ID:** [6C9U](https://www.rcsb.org/structure/6C9U)\n* **Release Date:** 2018-05-23\n* **Method:** X-RAY DIFFRACTION 2.09 \u00c5\n* **Organisms:** Saccharopolyspora erythraea, Homo sapiens\n* **Macromolecules:** 6-deoxyerythronolide-B synthase (EryA2, modules 3 and 4), Heavy chain of Fab 1B2, Light chain of Fab 1B2 (proteins)\n* **Unique Ligands:** K, NA\n* **Publication:** [Li, X., et al. (2018). Structure\u2013Function Analysis of the Extended Conformation of a Polyketide Synthase Module. Journal of the American Chemical Society, 140, 6518-6521.](https://pubs.acs.org/doi/10.1021/jacs.8b02100)\n* **Blog Post:** [Structure\u2013Function Analysis of the Extended Conformation of a Polyketide Synthase Module](https://marcdeller.com/structure-function-analysis-of-the-extended-conformation-of-a-polyketide-synthase-module/)\n\n**Description:** This landmark structure resolved a long-standing debate in polyketide synthase biology by demonstrating that the antibody-stabilized extended module conformation is catalytically functional for both chain elongation and modification. The X-ray/SAXS-validated structure unlocks critical insights for engineering hybrid synthases to produce novel antibiotics and other natural products.", "# Discovery of Potent and Selective Inhibitors of Wild-Type and Gatekeeper Mutant Fibroblast Growth Factor Receptor (FGFR) 2/3\nAuthors: Shvartsbart, A.; Roach, J.J.; Witten, M.R.; Koblish, H.; Harris, J.J.; Covington, M.; Hess, R.; Lin, L.; Frascella, M.; Truong, L.; Leffet, L.; Conlen, P.; Beshad, E.; Klabe, R.; Katiyar, K.; Kaldon, L.; Young-Sciame, R.; He, X.; Petusky, S.; Chen, K.-J.; Horsey, A.; Lei, H.-T.; Epling, L.B.; Deller, M.C.; Vechorkin, O.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2022\n\nIn summary, the hit to lead development of next-generation FGFR2/3 inhibitors has been disclosed. Key discoveries include the THF-substituted ether to drive potency into sub-nanomolar range, the 4-fluoro-3-substituted benzamides to improve general kinase selectivity, a scaffold hop to reduce the polarity of the bicyclic core and enhance cellular permeability, and replacement of the N-methyl amide with N-CD3 to shut down a (presumed) metabolic pathway and abolish CYP3A4 inhibition. The lead compound 29 is a highly potent inhibitor of FGFR2/3, has excellent FGFR1/4 and kinome selectivity, as well as desirable in vitro ADME properties and pharmacokinetics in rat. An X-ray co-crystal structure of a related compound showed orientation of the pyridine nitrogen toward the gatekeeper residue, explaining the slight loss in potency for the V555L and V555M mutants. A combined PD/serum phosphate-level study established that toxicity caused by hyperphosphatemia could potentially be avoided in a clinical setting with FGFR2/3 inhibitors that have high selectivity over FGFR1.\n\n## Chemistry\n\nCompounds 5\u221225 were synthesized according to the synthetic route outlined in Scheme 1. Bromination of commercially available imidazo[1,2-b]pyridazine 31 with N-bromosuccinimide affords the trihalogenated intermediate 32 in nearly quantitative yield. Selective SNAr with NH4OH then provides amine 33, which is selectively halogenated with N-iodosuccinimide to deliver 34. Removal of the amino group by heating with t-BuONO leads to key intermediate 35 via diazotization and loss of N2. A selective Suzuki coupling of the iodide over the bromide affords pyrazole 36. Reaction of 36 with the appropriate alcohols and Cs2CO3 in DMF at 90 \u00b0C provides intermediates 37, which undergo a second Suzuki coupling to install the (un)substituted benzamide rings or the m-pyridyl derivatives. Notably, displacement of the chloride in 36 with amines required heating at 160 \u00b0C in NMP.\n\n*Scheme 1.* Synthesis of compounds 5\u221225. Reagents and conditions: (a) NBS, DMF, 60 \u00b0C, 2 h, quant.; (b) aq. NH4OH, sealed vessel, 120 \u00b0C, 81%; (c) NIS, DMF, 60 \u00b0C, 2 h, quant.; (d) t-BuONO, THF, 70 \u00b0C, 2 h, 81%; (e) ArB(OH)2 or ArBPin, K2CO3, Pd(PPh3)4, 5:1 1,4-dioxane/water, 100 \u00b0C, 16 h, 81%; (f) R2OH, Cs2CO3, DMF, 90 \u00b0C, 1 h when Y = O; R2NH2, Cs2CO3, NMP, 160 \u00b0C, 1 h when Y = N; and (g) ArB(OH)2 or ArBPin, Na2CO3, Pd(dppf)Cl2\u00b7CH2Cl2, 3:1 DMF/water or 3:1 dioxane/water, 90 \u00b0C, 2 h.\n\nCompounds 26\u221230 were synthesized according to the synthetic route outlined in Scheme 2. Chlorination of commercially available pyrazolo[1,5-a]pyrimidine 38 with N-chlorosuccinimide affords trichlorinated intermediate 39, which undergoes selective SNAr to give amine 40. Iodination with N-iodosuccinimide cleanly provides 41, which is deaminated by heating with t-BuONO in THF. Selective Suzuki coupling of the iodide then delivers pyrazole 43. Reaction of 43 with ethers or amines and cesium carbonate at elevated temperature provides intermediates 44, which undergo a second Suzuki coupling to install the amide-substituted aryl ring.\n\n*Scheme 2.* Synthesis of compounds 26\u221230. Reagents and conditions: (a) NCS, DMF, 60 \u00b0C, 2 h, 78%; (b) aq. NH4OH, sealed vessel, 100 \u00b0C, 85%; (c) NIS, DMF, 60 \u00b0C, 2 h, 85%; (d) t-BuONO, THF, 70 \u00b0C, 2 h, 47%; (e) ArB(OH)2 or ArBPin, K2CO3, PdCl2(dppf)\u00b7CH2Cl2, 5:1 1,4-dioxane/water, 100 \u00b0C, 1 h; (f) R2OH or R2NH2, Cs2CO3, DMF, 90 \u00b0C, 1 h; and (g) ArB(OH)2 or ArBPin, Cs2CO3, XPhos Pd G2, 3:1 DMF/water or 3:1 dioxane/water, 90 \u00b0C, 2 h.\n\nAlternatively, compounds 26\u221230 may be prepared by a modified route outlined in Scheme 3. In this case, the initial halogen introduced is a bromo, which ultimately allows the pyridyl ring to be installed first.\n\n*Scheme 3.* Alternative synthesis of compounds 26\u221230. Reagents and conditions: (a) NBS, DMF, 2 h, 86%; (b) aq. NH4OH, sealed vessel, 100 \u00b0C, 98%; (c) NCS, DMF, 50 \u00b0C, 0.5 h, 82%; (d) t-BuONO, THF, 70 \u00b0C, 2 h, 64%; (e) R2OH or R2NH2, Cs2CO3, DMF, 90 \u00b0C, 1 h; (f) ArB(OH)2 or ArBPin, K2CO3, PdCl2(dppf)\u00b7CH2Cl2, 4:1 1,4-dioxane/water, 100 \u00b0C, 1 h; and (g) ArB(OH)2 or ArBPin, Cs2CO3, XPhos Pd G2, 3:1 DMF/water or 3:1 dioxane/water, 90 \u00b0C, 2 h.\n\n## Experimental Section\n\nAll reactions were run under an atmosphere of dry nitrogen. Unless otherwise noted, all reactions were performed at ambient temperature, which averaged 20 \u00b0C. All solvents were used without further purification as acquired from commercial sources. NMR spectra were obtained using a Varian Mercury-300, Mercury-400, or Inova-500 spectrometer. Chemical shifts are reported in parts per million relative to tetramethylsilane (TMS) as internal standard. All final products were characterized by 1H NMR, HRMS or LCMS, and HPLC methods. Purifications by flash chromatography were performed on RediSep columns using an Isco CombiFlash SG100c. Reverse-phase preparative HPLC purifications were performed on a Waters FractionLynx system using UV-triggered or mass-directed fractionation and compound-specific method optimization. All final compounds for biological testing were purified by reverse-phase prep HPLC to >95% purity, as determined by analytical LCMS.\n\n### Synthesis of Compound 25\n\n**Step 1. 3,8-Dibromo-6-chloroimidazo[1,2-b]pyridazine (32).** 8-Bromo-6-chloroimidazo[1,2-b]pyridazine (10.6 g, 45.8 mmol) was dissolved in DMF (114 mL) and treated with NBS (10.6 g, 59.5 mmol). The solution was stirred at 60 \u00b0C. After 2 h, LCMS indicated complete consumption of the starting material. The reaction was poured into ice water (800 mL) and stirred for 1 h at room temperature. The mixture was filtered under vacuum, and the residue was dried with continued air flow to provide 3,8-dibromo-6-chloroimidazo[1,2-b]pyridazine (14.8 g, 47.5 mmol, quantitative yield assumed). The crude material was used without further purification. LCMS calculated for C6H3Br2ClN3 (M + H)+: m/z = 311.8, 309.8, 313.8; found: 311.8, 309.8, 313.8.", "# FoldForge \u2013 Browser GUI for the whole Boltz-2 workflow\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-14 on marcdeller.com (categories: vibe-code).*\n\n> **A single-file browser GUI for the whole Boltz-2 workflow: build, run, review.**\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/FoldForge](https://github.com/bellcheddar/FoldForge) |\n\n---\n\nForgeFold is a browser-based interface for preparing, running, and reviewing Boltz-2 structure-prediction jobs. It brings Boltz setup, YAML construction, job-script generation, and result inspection into a single lightweight HTML application.\n\nWhy it matters: the Boltz-2 workflow normally means hopping between a text editor for YAML, a shell for the run script, and a spreadsheet for the results, with plenty of room for format mistakes along the way. ForgeFold pulls all four stages (set the binary path, build the input YAML from sequences and SMILES, generate the run script, and review confidence and affinity metrics) into one lightweight HTML application that opens in any browser with no build step. It is useful for structure-guided teams who want to lower the friction of repeated co-folding runs and give less command-line-comfortable colleagues a clean, guided path from input preparation to result triage.\n\n## Overview\n\nThe application is organized around a simple four-step workflow: set the local Boltz-2 binary path, build an input YAML from protein and ligand definitions, configure a run script, and review output metrics from analyzed results.\n\nThis design is meant to reduce friction for structure-guided workflows where users repeatedly move between FASTA/SMILES input preparation, command-line execution, and affinity-oriented output review.\n\n## Core workflow\n\n1. **Setup installation** \u2014 store the path to the local Boltz-2 executable so it can be reused across sessions.\n2. **Build YAML input** \u2014 enter protein FASTA sequences and ligand SMILES, then export an `input.yaml` file for Boltz-compatible predictions.\n3. **Configure and write jobs** \u2014 choose run parameters and generate a `run_boltz.zsh` shell script for local execution.\n4. **Analyse results** \u2014 load processed outputs such as `boltz_summary.csv` to inspect confidence and affinity-related metrics.\n\n## Features\n\n* Browser-based GUI for Boltz-2 workflow management.\n* YAML builder for protein and ligand input definition.\n* Job-script generation for local prediction runs.\n* Results view for summary CSV uploads and prediction review.\n* Support for metrics including `confidence_score`, `affinity_pred_value1`, and derived `pIC50` display.\n\n## Input model\n\nForgeFold is built around Boltz-style YAML inputs that define one or more protein chains and optional ligands. The example interface content shows protein entries with `id` and `sequence`, ligand entries with `id` and `smiles`, plus optional constraint and property blocks for pocket and affinity settings.\n\nA typical workflow is to paste receptor and partner sequences, add a ligand SMILES string, and export the resulting YAML before launching prediction jobs locally.\n\n## Results interpretation\n\nThe interface is designed to surface key prediction outputs after external analysis. The bundled UI text highlights `confidence_score` as an overall model-confidence measure, `affinity_pred_value1` as an affinity-related score, `affinity_probability_binary1` as a binder-likelihood output, and `pIC50` as a derived potency-style estimate.\n\nThese values are intended for rapid triage and comparison of modeled complexes rather than as a substitute for experimental validation.\n\n## Running ForgeFold\n\nBecause ForgeFold is delivered as a standalone HTML application, the simplest way to use it is to open the file in a modern browser and interact with the GUI directly.\n\nTypical usage:\n\n```\n# Example: open the app locally\nopen ForgeFold.html\n```\n\nOr host it with any lightweight static server:\n\n```\npython -m http.server 8000\n```\n\nThen open `http://localhost:8000` in a browser.\n\n## Expected companion tools\n\nForgeFold is designed to sit alongside a local Boltz-2 installation and a downstream analysis script such as `analyze_boltz.py`, which generates CSV outputs for the Results tab.\n\nTo make the full workflow work smoothly, keep the following available:\n\n* A working local `boltz` executable.\n* Input biomolecular sequences and ligand SMILES strings.\n* A results-processing script that emits `boltz_summary.csv` or compatible output.\n\n## Suggested repository structure\n\n```\nForgeFold/\n\u251c\u2500\u2500 README.md\n\u251c\u2500\u2500 ForgeFold.html\n\u251c\u2500\u2500 analyze_boltz.py\n\u251c\u2500\u2500 examples/\n\u2502   \u2514\u2500\u2500 input.yaml\n\u2514\u2500\u2500 outputs/\n```\n\n## Roadmap ideas\n\nPotential future improvements suggested by the current UI structure include more robust YAML validation, richer job templating, tighter integration with local output directories, and more detailed downstream result visualizations.\n\n## License\n\nAdd the project license here.\n\n## Source context\n\nThis README was drafted from the attached ForgeFold application content and interface text.\n\n---\n\n## \ud83d\udc64 Author\n\n**Marc C. Deller, D.Phil.**\nStructural biologist & drug discovery scientist\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 | [github.com/bellcheddar/FoldForge](https://github.com/bellcheddar/FoldForge) |", "# Discovery of the Orally Bioavailable Isoform Selective Janus Kinase 1 (JAK1) Compound Povorcitinib (INCB054707) for the Treatment of Inflammatory and Autoimmune Diseases\nAuthors: Zhuo, J.; Li, Y.-L.; Qian, D.-Q.; Burns, D.M.; Mei, S.; Rafalski, M.; Xu, M.; Cao, G.; Pan, Y.; Jia, Z.; Jalluri, R.; Epling, L.B.; Fenalti, G.; Deller, M.C.; Procak, J.; Covington, M.; He, X.; Collins, R.; Stubbs, M.; Burke, K.; Oliver, J.; Margulis, A.; Boer, J.; Wynn, R.; Scherle, P.; Diamond, S.; Newton, R.; Zhang, Y.; Metcalf, B.; Yao, W.\nSource: Journal of Medicinal Chemistry (Drug Annotation), 2026\n\n**INA6 Proliferation Assay (Cellular).** The effects of the compounds on cell proliferation were measured in an ATP luminescence assay using the human myeloma cell line INA-6. RPMI 1640 Medium, 200 mM L-glutamine, and FBS were purchased from Thermo Fisher Scientific (Waltham, MA). Interleukin-6 was purchased from R&D Systems (Minneapolis, MN). CellTiter-Glo Luminescent Cell Viability Assay kit was purchased from Promega (Madison, WI). INA-6 cells were maintained at 4 \u00d7 10\u2075 cells/mL in the complete RPMI 1640 with IL-6 (10 ng/mL) at 37 \u00b0C, 5% CO2 24 h prior to assay. Compounds were dotted (3-fold dilution starting at 20 \u03bcM concentration in the assay) in 384-well white clear bottom assay plates (Corning Tissue Culture Plates). The MultiFlo FX was used to add 1200 cells/well directly to the compound plates. Seventy-two hours after incubation at 37 \u00b0C, 5% CO2, CellTiter-Glo reagent was added and luminescence signal was read with PHERAstar microplate reader (BMG Labtech). Compound inhibition of cell viability was calculated and reported as percentage of inhibition relative to wells with a full-inhibitor control (100% inhibition) and DMSO (0% inhibition). IC50 determination was performed by fitting the curve of compound percent inhibition versus the log of the compound concentration using the GraphPad Prism 9 software.\n\n**Human Whole Blood IL-6 pSTAT3MSD Assays.** To determine the effects of compounds on JAK1 (IL-6)-mediated STAT3 phosphorylation inhibition in normal human whole blood, 280 \u03bcL from healthy donors (BioIVT) was incubated with 10 \u03bcL of test compound (3-fold dilution starting at 10 \u03bcM concentration in the assay, or DMSO alone) for 10 min in a 96-well 2 mL Assay Block (Corning, Cat. #3961) at 37 \u00b0C. Following compound preincubation, 100 ng/mL human IL-6 (R&D Systems, Cat. #206-IL-050) was added to the appropriate wells for 15 min at 37 \u00b0C. Blood was then lysed with 1.2 mL Buffer EL (Qiagen, Cat. #79217) for 5 min on ice, centrifuged at 1200 rpm for 5 min, and supernatant was aspirated. Buffer EL was again added to the wells, lysed for 5 min on ice, centrifuged, and supernatant was aspirated. The cell pellet was lysed with 150 \u03bcL 1X Cell Signaling Lysis Buffer (Cell Signaling Technologies, Cat. #9803) and 1X Halt Protease and Phosphatase Inhibitor Cocktail (Thermo Scientific, Cat. #78447) for 30 min on ice. Subsequently, lysates were analyzed for pSTAT3 using a pSTAT3MSD kit per manufacturer protocol (Meso Scale Diagnostics Phospho-STAT Panel Kit, Cat. #K15202D-4). Plates were quantitated on a Meso Sector S 600 (Meso Scale Diagnostics). IC50 determination was performed by fitting the curve of compound percent inhibition versus the log of the compound concentration using the GraphPad Prism 9 software. INCB018424 (Compound 1) at 10 \u03bcM was used as 100% pSTAT3 inhibition. The remaining compounds were normalized to this value in each run. Curves were constrained from 0\u2013100%.\n\n**Human Whole Blood TPO pSTAT3MSD Assays.** To determine the effects of compounds on JAK2 (TPO)-mediated STAT3 phosphorylation inhibition in normal human whole blood, 280 \u03bcL from healthy donors (BioIVT) was incubated with 10 \u03bcL of test compound (3-fold dilution starting at 10 \u03bcM concentration in the assay, or DMSO alone) for 10 min in a 96-well 2 mL Assay Block (Corning, Cat. #3961) at 37 \u00b0C. Following compound preincubation, 50 ng/mL human TPO (R&D Systems, Cat. #288-TP-025) was added to the appropriate wells for 15 min at 37 \u00b0C. Blood was then lysed with 1.2 mL Buffer EL (Qiagen, Cat. #79217) for 5 min on ice, centrifuged at 1200 rpm for 5 min, and supernatant was aspirated. Buffer EL was added to the wells, lysed for 5 min on ice, centrifuged, and supernatant was aspirated. The cell pellet was lysed with 150 \u03bcL 1X Cell Signaling Lysis Buffer (Cell Signaling Technologies, Cat. #9803) and 1X Halt Protease and Phosphatase Inhibitor Cocktail (Thermo Scientific, Cat. #78447) for 30 min on ice. Subsequently, lysates were analyzed for pSTAT3 using a pSTAT3MSD kit per manufacturer protocol (Meso Scale Diagnostics Phospho-STAT Panel Kit, Cat. #K15202D-4). Plates were quantitated on a Meso Sector S 600 (Meso Scale Diagnostics). IC50 determination was performed by fitting the curve of compound percent inhibition versus the log of the compound concentration using the GraphPad Prism 9 software. INCB018424 (Compound 1) at 10 \u03bcM was used as 100% pSTAT3 inhibition. The remaining compounds were normalized to this value in each run. Curves were constrained to 0\u2013100%.\n\n**IL-2 Stimulated Human T Cell pJAK2/pSTAT3/pSTAT5 Assays.** Peripheral blood mononuclear cells (PBMCs) can be prepared from human whole blood samples using Ficoll Hypaque separation method, and T cells (fraction 20000) can be obtained from PBMCs by elutriation. Freshly isolated human T cells can be maintained in culture medium (RPMI 1640 supplemented with 10% fetal bovine serum, 100 U/mL penicillin, 100 \u03bcg/mL streptomycin) at a density of 2 \u00d7 10\u2076 cells/mL at 37 \u00b0C for up to 2 days. For IL-2-stimulated cell proliferation analysis, T cells are first treated with Phytohemagglutinin (PHA; Sigma-Aldrich) at a final concentration of 10 \u03bcg/mL for 72 h. After washing twice with PBS, 6000 cells/well are plated in 96-well plates, and treated with compounds at different concentrations or DMSO for 30 min in the culture medium in the presence or absence of 100 ng/mL of human IL-2 (ProSpec-Tany TechnoGene; Rehovot, Israel) for 10 min. Cell lysates were prepared and subjected for Western blot analysis for the indicated proteins.", "# GitHub README: AlphaFraud\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/AlphaFraud, 2026\n\n- [x] **Statistics page** \u2014 an exhaustive live `/stats` page grouping everything the app tracks into themed, equal-height panels (consistent with the rest of the site): catch & quality KPIs, coverage & schedule (date ranges, release weeks, last/next update), runs & compute, server metrics (CPU, load, memory, uptime), storage & database (file sizes, per-table row counts, disk), data sources & APIs (linked), and traffic \u2014 with a \"Full statistics \u2192\" link from the header panel\n- [x] **Home landing page** \u2014 a concise navigation page at `/` outlining the project and each tab as a colour-coded card with a \"what you'll learn\" line and headline KPIs; the cumulative dashboard moved to `/all` (\"All structures\"). Nav tabs now carry their panel accent colour to tie the site together\n- [x] **Calculate tab** \u2014 type any human PDB ID and AlphaFraud runs the full pipeline on demand (resolve the entity, fetch its AlphaFold model, superpose, score, render the ribbon) and returns the same panel as the auto examples, with **no LLM**. A strict autocomplete index (exists + human + has an AlphaFold model), seeded from the corpus and auto-refreshed each week, guarantees only qualifying entries; rejections are explained (not human / antibody-fusion / no AlphaFold model); good predictions render the green \"gets it right\" panel and confidently-wrong ones the failure panel; the heavy compute runs in an isolated subprocess and on-demand results stay out of every aggregate\n- [x] **Examples tab** \u2014 a curated storytelling layer that shows both sides of AlphaFold's record. It opens with a **positive control** (human Artemis, a novel post-cutoff nuclease predicted blind to 0.45 \u00c5 C\u03b1-RMSD and scored FRAUD 0.02) proving the pipeline certifies good predictions as well as bad; an auto-generated, LLM-free **\"example of the week\"** plus a **complete weekly-release archive** (catches and quiet weeks alike); hand-authored deep-dive panels across the failure spectrum, each with live metrics and an interactive 3D viewer; and clickable red/green example markers overlaid on the home fraud-quadrant scatter\n- [x] **Page and figure caching** \u2014 the cumulative dashboard's heavy whole-archive figures (scatter / dumbbell / histograms over ~80k rows) are precomputed once and cached in an `analysis_snapshots` row with a short TTL, and every static asset (vendored Plotly / 3Dmol, ribbons, CSS/JS) is served with long immutable cache headers, so repeat loads are near-instant\n- [x] **Weekly-release archive at full granularity** \u2014 the Archive tab now lists every PDB release week since the 2018 cutoff (derived from each structure's release date), each linking to that week's structures and confidently-wrong count, instead of only the recent ongoing-watch runs\n- [x] **Fix slow page loads after the full backfill** \u2014 once the archive backfill filled the `entities` table with ~470 KB `heatmaps_json` per compared row (\u2248 6.5 GB), every list/sort query (home, leaderboard, week pages) dragged all that JSON through memory and took 20\u201360 s (the leaderboard 504'd at the 60 s gunicorn timeout). Fixed by moving the big per-residue and heatmap payloads into an `entity_blobs` sidecar table read only by the per-structure entry page, so the hot table stays small and list queries return in well under a second\n- [x] **Structural imagery throughout** \u2014 every worst offender is rendered as a deviation-coloured C\u03b1 ribbon (experiment coloured by distance from the AlphaFold model, on an absolute-\u00c5ngstr\u00f6m scale), shown on the leaderboard, entry page and weekly highlights; the entry page adds an interactive 3Dmol.js viewer with a proper secondary-structure cartoon and an optional translucent AlphaFold \"ghost\" overlay in the shared superposition frame; a \"divergence\" ribbon banner sits in the header. All server-rendered as vendored, offline SVG/PDB (no CDN)\n- [x] **Percentages on the KPI tiles** \u2014 each headline count also shows its share of the batch (e.g. confidently wrong: 1,941 = 2.3%), on the dashboard, week and leaderboard views\n- [x] **Worst-offenders structural deep dive** \u2014 the **Analysis** tab characterises the confidently-wrong set by CATH / SCOP2 fold and family class (Wilson-CI + Fisher-exact enrichment), clusters it by sequence similarity, flags shared themes (amyloid / assembly / disordered / coiled-coil / engineered), maps failure modes by PCA, detects conformational heterogeneity, gives per-superfamily blind-spot scorecards, and links every offender to its RCSB entry and DOI-verified manuscript; refreshed hourly and after each weekly run\n- [x] **Hover-to-preview structures** \u2014 hovering a point on the \"fraud quadrant\" scatter pops a floating deviation-coloured ribbon thumbnail of that structure; the entry page carries the full interactive 3D viewer\n- [x] **Header \"Stats\" panel** \u2014 a live panel in the top-right of the header tracking app health: worker memory, SQLite DB size, unique visitors, structures analysed, confidently-wrong count and archive %; refreshes every minute from `/api/stats`\n- [x] **Supervised backfill service** \u2014 the archive backfill runs as a systemd unit (`alphafraud-backfill.service`) that auto-restarts on crash and survives reboots, replacing the detached `nohup` process\n- [ ] **Family-spectrum analysis tab** \u2014 a new tab presenting worst-offending proteins across the accuracy spectrum with a clear worked example per tier: confidently wrong (high pLDDT, low TM), intermediate / conformational disagreement, and well-predicted; plus a distinct \"honest failure\" category (low TM but AlphaFold flagged its own low confidence). Deduplicated by UniProt, since AlphaFold DB holds one blind model per sequence, so repeat depositions of a protein share an identical prediction (already reflected in the confidently-wrong dumbbell's per-protein collapse and the leaderboard `\u00d7N` column, and documented in [FAMILY_SPECTRUM.md](FAMILY_SPECTRUM.md)). Seed examples: leukocyte cell-derived chemotaxin-2 (`O14960`, \u00d77, median TM 0.17 at pLDDT 97), TMEM106B (`Q9NUM4`, \u00d722, 0.18), coagulation factor X (`P00742`, \u00d729, 0.21) at the bad end; calmodulin (`P0DP23`, \u00d7163, 0.51) and DNA ligase 4 (`P49917`, 0.50) mid; carbonic anhydrase 12 (`O43570`), ACE (`P12821`), aldo-keto reductase 1B1 (`P15121`) at ~1.0\n- [ ] **Multi-chain / complex accuracy** \u2014 compare assemblies with QS-score and interface lDDT (AlphaFold DB models are monomer-only today, so interfaces are untested)\n- [ ] **Accuracy by method and resolution** \u2014 break results down by X-ray / cryo-EM / NMR and by resolution to see where AlphaFold struggles most\n- [ ] **Map multi-domain and engineered chains** \u2014 compare antibodies, fusions and constructs per-domain instead of skipping chains without a single UniProt accession\n- [ ] **Process-based parallel backfill** \u2014 a `ProcessPoolExecutor` path to beat the single-thread ceiling without the C-extension thread-safety issues that forced single-threaded mode\n- [ ] **Predictor and model-version tracking** \u2014 record the AlphaFold DB model version, flag entries updated since the first prediction, and add other predictors (AlphaFold3, ESMFold) for comparison\n- [ ] **Catch alerts** \u2014 optional email / Slack notification when a new confidently-wrong or novel-and-wrong structure is found\n- [ ] **Robust large-assembly parsing** \u2014 handle the CIF-only and oversized structures that currently log a parse error and are skipped\n- [ ] **Tests and CI** \u2014 a regression set pinned on known catches (transthyretin, SOD1, \u03b22-microglobulin) plus continuous integration", "# GitHub README: BoltzMaker\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/BoltzMaker, 2026\n\n**The `.bmz` results file.** A zip carrying `manifest.json` (format version, campaign name,\ntimestamps, a SHA-256 of the `BoltzMaker.py` that produced it, per-file checksums, and an explicit\nrecord of anything not packed), the summary CSVs, one structure per target, and one labelled PLIP\nimage per target. Its layout is written by the bundle's generated `pack_results.py` and read by\n`web/boltzmaker_web/results.py`; both mirror one `BMZ_VERSION`, and a file declaring any other\nversion is refused rather than parsed optimistically. Every uploaded zip -- a `.bmz` or a\nStepwise campaign folder -- is treated as hostile: extraction is bounded on entry count, declared\nsize, actual written size, compression ratio and resolved path (so zip-slip and zip-bomb payloads\nare rejected), with every check run over the whole member list before a single byte is written.\n\n**Optional PLIP.** With the droplet's `.plip_env` present (the same conda-forge PyMOL/OpenBabel\nenvironment `setup-plip` builds locally, ~1-1.5GB), hosted `analyze` runs full interaction\ndetection and compare-sse; without it, both degrade gracefully rather than erroring, as locally.\n\n**Serving stack:** gunicorn (3 sync workers, 900s timeout to cover PLIP rendering and compare-sse's\nGPCRdb/KLIFS/PDBe lookups) behind nginx (Let's Encrypt TLS via certbot, HTTP/2, rate-limited upload\nendpoints, 200MB body cap), as a hardened systemd service (dedicated unprivileged user, memory/CPU\ncaps). Every request gets its own scratch directory, deleted when the request finishes and swept\nby a five-minute cleanup timer as a backstop.\n\n## \ud83e\ude79 Troubleshooting / FAQ\n\nLocal runs first, then the hosted site.", "# GitHub README: CODSWALLOP\nAuthors: Deller, M.C.\nSource: github.com/bellcheddar/CODSWALLOP, 2026\n\n- [x] **Flask skeleton, deploy pattern and port allocation.** Blueprint layout, gunicorn config, systemd unit, nginx vhost with the post-certbot HTTP/2 patch and the shared long-cache snippet. Port 8006, because 8000 to 8005 are taken by AlphaFraud, ChemSage, ChatPDB, BoltzMaker, FlexAppeal and PANTS\n- [x] **SQLite cache with a family/entry/entity schema and a TTL.** Two layers: raw API responses keyed by the request itself, and assembled families. Second query for a family is 0.07 s against 7 to 12 s cold\n- [x] **Cache keys carry a parse version.** Because searches are cached in *parsed* form, adding a field to the parser leaves it absent from every cached row and the consuming code fails silently. Adding the alignment spans without bumping it left the coverage figure reading 100 % and the fusion count reading 0, both of which look like answers rather than missing inputs\n- [x] **Column migrations, not just `CREATE TABLE IF NOT EXISTS`.** Which silently does nothing to a table that already exists, so a new column reads back NULL on any deployed instance and the app computes a plausible wrong answer\n- [x] **Family resolver accepting seven input kinds.** PDB ID, PDB entity id, UniProt accession, gene name, Pfam/InterPro accession, raw sequence, free text. Ambiguous input returns a disambiguation card, never a guess\n- [x] **One definition of family membership.** Every input resolves to a seed sequence and membership is a percent-identity threshold against it, so the slider is exact rather than annotation-dependent. Where the *seed* rather than the input is a choice (which member represents a Pfam family), it is picked by a stated rule and the header says which and why\n- [x] **Batched GraphQL fetch of core metadata.** Fifty entities per round trip, each batch cached independently so overlapping families share what they have in common\n- [x] **R-factor fallback chain.** `_refine.ls_R_factor_R_work` is the field everyone quotes and a good fraction of the archive leaves it empty, putting the number in `ls_R_factor_obs` or `ls_R_factor_all` instead. Reading only the obvious field showed a blank R-work beside a populated R-free on 82 of 1,473 carbonic anhydrase II X-ray entities\n- [x] **The drawer shell, complete.** Divider rail with phase pips, family header with the FILED stamp, stat strip, slide-over index card, dark and light token sets. Built once so later phases only add panels\n- [x] **Hero map with a placeholder layout.** Positions come from the server so the map, table and any future export agree on one set of coordinates. The radial axis is identity *rank*, not raw identity: a family whose members are 95 % near-identical piles onto one circle under a linear map, which is exactly what the first version drew\n- [x] **Cross-highlight wired.** One shared selection state, and every panel reads it. This is the element the interface is remembered by and the thing later phases must bind to rather than reinventing\n- [x] **Construct coverage census as a depth profile, not a binary union.** The union answers \"has anyone ever put this residue in a construct\", which for any well-studied protein is yes everywhere: 140 of the 1,992 spike entities carry the full 1,273 residues, so the union reports perfect coverage for a protein whose cytoplasmic tail is essentially never studied. The depth profile says the useful thing instead\n- [x] **Full table view with CSV, JSON and BibTeX export.** BibTeX deduplicated to one record per paper with the PDB IDs collected in a note\n- [x] **Payload deduplication.** Sequences, citations, chemical components and domain annotations moved to family-level lookups: 6.3 MB to 1.8 MB, ~240 kB gzipped\n- [x] **Mobile responsive and `prefers-reduced-motion` respected**, at this phase rather than retrofitted\n- [x] **Deployed to codswallop.mdeller.com** and listed at the top of `mdeller-landing/apps.json`. The beacon is `/api/stats`, and it needed the footer's \"filed so far\" line to exist first: the beacon was declared before anything on any page requested it, which would have left the launcher's hit count reading zero forever with nothing to show why. A static asset would have been the easier beacon and a worse one, since `/static/` is served `immutable` and a returning reader never re-requests it\n\n### Phase 2: sequence, construct and domain intelligence", "# SiteDoctor \u2013 Full diagnosis, health, content, and SEO for any WordPress site\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2026\n\n*Published 2026-06-14 on marcdeller.com (categories: vibe-code).*\n\n> **One pass. Full diagnosis. Health, content, and SEO for any WordPress site.**\n\n|  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- |\n| \ud83c\udf10 **Website** | [marcdeller.com](https://marcdeller.com) | \u2709\ufe0f **Contact** | [marc@marcdeller.com](mailto:marc@marcdeller.com) | \ud83d\udc19 **GitHub** | [bellcheddar/site-doctor](https://github.com/bellcheddar/site-doctor) |\n\n---\n\nA single-file, dependency-free Python 3 tool that diagnoses the health, content, and SEO/performance profile of a WordPress site \u2014 using only the standard library.\n\nWhy it matters: Keeping a WordPress site fast, secure, and well-indexed normally means juggling several separate tools (a header scanner, a TLS checker, a PageSpeed run, a content audit). Site Doctor folds all of that into a single dependency-free Python script that reads only public information and emits both a color-coded terminal report and a branded, self-contained HTML report in one pass. It is useful for site owners, agencies, and developers who want a fast, repeatable, zero-install health check: ideal for pre-launch audits, ongoing maintenance, and verifying that caching, security headers, and Core Web Vitals are actually configured the way you think they are.\n\nIt produces a color-coded report on screen **and** a branded HTML report on disk, in the same pass.\n\n## Features\n\nBy default, Site Doctor runs **all three** checks below against a target URL\nin a single pass and writes one combined HTML report. Pass `--scan`,\n`--inventory`, and/or `--analytics` to run only a subset.\n\n### 1. Diagnostic Scan (`--scan`)\n\n* Request timing breakdown: DNS, TCP connect, TLS handshake, TTFB, download, total\n* HTTP status, response headers, and redirect chain\n* Caching behaviour (`Cache-Control`, `ETag`, `Last-Modified`, etc.)\n* Security headers (HSTS, CSP, X-Frame-Options, X-Content-Type-Options, Referrer-Policy, Permissions-Policy)\n* Cookie flags (`Secure`, `HttpOnly`, `SameSite`)\n* CORS configuration and mixed-content checks\n* SSL/TLS certificate details (issuer, expiry, SANs), supported TLS protocol versions, ALPN/HTTP2 detection\n* DNS records (A, AAAA, MX, TXT, NS, CNAME, PTR)\n* `robots.txt`, sitemap, favicon, and custom 404 handling checks\n* HTTP \u2192 HTTPS and www \u2194 non-www redirect behaviour\n* Server technology / CDN detection (e.g. Cloudflare, Nginx, LiteSpeed)\n* WordPress fingerprinting: version disclosure, active theme, plugins, REST API,\n  XML-RPC, `readme.html`, `wp-login.php`, and exposed user accounts via\n  `/wp-json/wp/v2/users`\n* Detected analytics/tracking scripts (Google Analytics, Jetpack Stats, etc.)\n* A weighted **Health Scorecard** (0\u2013100, grade A\u2013F) across Security Headers,\n  TLS/HTTPS, Performance, WordPress Hygiene, HTTP Behaviour, and SEO Basics\n\n> Note: visitor counts and live CPU/memory load cannot be measured externally\n> over HTTP \u2014 the report says so explicitly rather than guessing.\n\n### 2. WordPress Content Inventory (`--inventory`)\n\n* Walks every public post type exposed via the WordPress REST API\n  (`/wp-json/wp/v2/...`), paginating through results\n* De-duplicates by type + ID\n* Outputs a clean table: **ID, Type, Date, Category, Title**\n\n### 3. Analytics & SEO Report (`--analytics`)\n\n* Real Lighthouse scores via the public Google PageSpeed Insights API\n  (Performance, SEO, Accessibility, Best Practices, plus Core Web Vitals)\n* Automatic fallback to a self-computed performance scorecard if PSI is\n  unavailable (rate-limited, offline, etc.)\n* Per-page SEO breakdown for the site\u2019s content: title/meta description\n  length, H1 count, word count, image alt coverage, internal/external link\n  counts, Open Graph/Twitter tags, structured data (JSON-LD), and load time\n\n## Requirements\n\n* Python 3.8+\n* No third-party packages \u2014 only the standard library\n* Optional: `dig` (used for DNS lookups if available, with a socket-based\n  fallback otherwise)\n\n## Usage\n\n```\npython3 site_doctor.py [URL] [OPTIONS]\n```\n\nIf `URL` is omitted, you\u2019ll be prompted to enter one interactively (e.g.\n`https://example.com`). The scheme defaults to `https://` if not given.\n\n```\n# Prompt for a URL and run everything (scan + inventory + analytics)\npython3 site_doctor.py\n\n# Run everything against a specific site\npython3 site_doctor.py https://example.com\n\n# Run only the diagnostic scan and health scorecard\npython3 site_doctor.py https://example.com --scan\n\n# Run only the WordPress content inventory\npython3 site_doctor.py https://example.com --inventory\n\n# Run only the analytics/SEO report, skipping PageSpeed Insights,\n# analyzing 25 pages instead of the default 10\npython3 site_doctor.py https://example.com --analytics --max-pages 25 --no-psi\n\n# Run everything and write the HTML report to a custom path\npython3 site_doctor.py https://example.com --out reports/example.html\n```\n\n### Mode flags (omit all to run everything)\n\n| Flag | Description |\n| --- | --- |\n| `--scan` | Diagnostic scan: request timing (DNS/connect/TLS/TTFB/download), HTTP headers, caching, security headers, cookies, CORS/mixed content, TLS certificate and protocol support, DNS records, robots.txt/sitemap/favicon, redirect behaviour, server/CDN fingerprint, WordPress core/theme/plugin/user fingerprinting, and the weighted Health Scorecard (0\u2013100, grade A\u2013F). |\n| `--inventory` | WordPress content inventory: walks every public post type exposed via `/wp-json/wp/v2/...`, paginates through all results, de-duplicates by type+ID, and prints a table of ID, Type, Date, Category, Title. |\n| `--analytics` | Analytics & SEO report: real Lighthouse scores via Google PageSpeed Insights (Performance, SEO, Accessibility, Best Practices, Core Web Vitals), with automatic fallback to a self-computed performance scorecard if PSI is unavailable, plus a per-page SEO breakdown across the site\u2019s content. |", "# Supersite of immune vulnerability on the glycosylated face of HIV-1 envelope glycoprotein gp120\nAuthors: Kong, L.; Lee, J.H.; Doores, K.J.; Murin, C.D.; Julien, J.-P.; McBride, R.; Liu, Y.; Marozsan, A.; Cupo, A.; Klasse, P.-J.; Hoffenberg, S.; Caulfield, M.; King, C.R.; Hua, Y.; Le, K.M.; Khayat, R.; Deller, M.C.; Clayton, T.; Tien, H.; Feizi, T.; Sanders, R.W.; Paulson, J.C.; Moore, J.P.; Stanfield, R.L.; Burton, D.R.; Ward, A.B.; Wilson, I.A.\nSource: Nature Structural & Molecular Biology, 2013\n\n*Figure 5.* Supersite of vulnerability centered on the N332 glycan. (a) Binding to the N332 glycan on gp120 for PGT 135, 2G12 and PGT 128 is shown with their respective structures superimposed on glycan N332. (b) The epitopes of PGT 135, PGT 128 and 2G12 are shown in topology representations in which cylinders are \u03b1-helices, arrows are \u03b2-strands and lines are loops. N-linked glycan sites are circled. The structures of 2G12 (PDB ID: 1OP5) and PGT 128 (PDB ID: 3TYG) were obtained from the PDB.\n\n*Figure 6.* Conserved conformation of N332 glycan. (a) The N-linked glycan at N332 of gp120 has a conserved conformation in crystal structures with PGT 135 (blue) and PGT 128 (light brown). The structures are superposed on the gp120 protein. (b) Definition of A and B faces of a glycan residue. The B face (labeled in red), on which the carbons are arranged counter-clockwise around the ring, is generally more apolar than the A face (labeled in cyan), on which the carbons are arranged clockwise around the ring. (c,d) The side from which each glycan moiety interacts with PGT 135 (c) or PGT 128 (d) is defined by an arrow pointing towards the interaction (red arrow indicates B face, cyan arrow indicates A face and grey arrow indicates the side of a glycan). Each sugar primarily uses its A face, B face, or the side in the interaction with the different antibodies; their buried surface areas are estimated based on the burial on each glycan residue. The structure of PGT 128 (PDB ID: 3TYG) is available from the PDB.\n\n## Author Contributions\n\nProject design by L.K., K.J.D., J.-P.J., R.L.S., D.R.B., A.B.W., I.A.W.; X-ray experimental work by L.K., J.-P.J., Y.H., R.S.L.; EM experimental work by J.H.L., C.D.M., R.K., A.B.W.; glycan array experimental work by R.M., J.C.P., Y.L., T.F.; mutational experimental work by K.J.D., K.M.L.; Env trimer reagents from A.M., A.C., P.-J.K., S.H., M.C., C.R.K., R.W.S. and J.P.M.; robotic crystallization screening at the JCSG by M.C.D., T.C. and H.T.; manuscript written by L.K., J.H.L., K.J.D., R.L.S., J.P.M., D.R.B., A.B.W., and I.A.W. All authors were asked to comment on the manuscript.\n\n## Acknowledgments\n\nWe thank X. Dai for help with data processing; J. Chittuluru for assistance with the 2D principal component analysis; D. Ekiert for initial cloning of PGT 135; R. Pejchal for initial cloning of soluble CD4 and the gp120 JRFL core construct; C. Poulsen and R. Wyatt from the Scripps Research Institute, La Jolla, California, USA for donation of 17b IgG; C. Arnold from the UK National Institute for Biological Standards and Control, Health Protection Agency for donation of ARP3119 monoclonal antibody CA13; M. Elsliger for computer support; C. Corbaci for help with preparing figures and movies; A. Irimia, C. Blattner, and M. Hong for discussions; and J.P. Verenini for help in manuscript formatting, and W. Koff for discussions and support. We also thank S.C. Arzberger, J. Zhang and ADA Technologies, Inc for supplying the high-density NHS-activated slides used for glycan microarray analysis; members of The Glycosciences Laboratory for their collaboration in the establishment of the neoglycolipid based microarray system, and Terry Butters and colleagues from the University of Oxford, UK for the glucosylated N-glycans. Work using the neoglycolipid system is supported by The Wellcome Trust (WT093378MA and WT099197MA) (T.F.), The UK Research Councils' Basic Technology Initiative 'Glycoarrays' (GRS/79268) (T.F.), EPSRC Translational Grant (EP/G037604/1) and NCI Alliance of Glycobiologists for Detection of Cancer and Cancer Risk (U01 CA128416) (T.F.). The electron microscopy data were collected at the National Resource for Automated Molecular Microscopy, which is supported by US National Institutes of Health (NIH) through the National Center for Research Resources' P41 program (RR017573) (A.B.W.) at the National Center for Research Resources. X-ray data sets were collected at the Advanced Light Source beamline 5.0.2 at the Berkeley Center for Structural Biology (BCSB) and the Advanced Photon Source beamline 23ID-B. The BCSB is supported in part by the NIH, the US NIH National Institute of General Medical Sciences (NIGMS), and the Howard Hughes Medical Institute. The Advanced Light Source is supported by the Director, Office of Science, Office of Basic Energy Sciences, of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231. Use of the APS was supported by the US Department of Energy, Basic Energy Sciences, Office of Science, under contract no. DE-AC02-06CH11357. The GM/CA-CAT 23-ID-B beamline has been funded in whole or in part with federal funds from the US National Cancer Institute (NCI) (Y1-CO-1020) and NIGMS (Y1-GM-1104). This work was supported by the International AIDS Vaccine Initiative Neutralizing Antibody Center; by the Center for HIV/AIDS Vaccine Immunology (CHAVI-ID UM1 AI00663) (A.B.W., D.R.B., I.A.W.); by the HIV Vaccine Research and Design (HIVRAD) program (P01 AI832362 and R37 AI36082) (J.P.M., A.B.W., I.A.W.); by the University of California, San Diego, Center for AIDS Research (CFAR) (A.B.W.), an NIH-funded program (P30 AI036214), which is supported by the following NIH Institutes and Centers: US National Institute of Allergy and Infectious Disease, NCI, US National Institute of Mental Health, US National Institute on Drug Abuse, US National Institute of Child Health & Human Development, US National Institute of Heart, Lung and Blood, US National Institute of Aging; by NIH RO1 grant GM046192-18 (I.A.W.), and the Joint Center of Structural Genomics by the NIH, NIGMS, Protein Structure Initiative [U54 GM094586] (I.A.W.). A portion of the work is supported by an American Foundation for AIDS Research Mathilde Krim Fellowship in Basic Biomedical Research (L.K.). The content is the responsibility of the authors and does not necessarily reflect the official views of NIGMS, NCI or the NIH. This is manuscript #21722 from The Scripps Research Institute.\n\n## References\n\n*(reference list omitted)*", "# Discovery of (4-Pyrazolyl)-2-aminopyrimidines as Potent and Selective Inhibitors of Cyclin-Dependent Kinase 2\nAuthors: Hummel, J.R.; Xiao, K.-J.; Yang, J.C.; Epling, L.B.; Mukai, K.; Ye, Q.; Xu, M.; Qian, D.; Huo, L.; Weber, M.; Roman, V.; Lo, Y.; Drake, K.; Stump, K.; Covington, M.; Kapilashrami, K.; Zhang, G.; Ye, M.; Diamond, S.; Yeleswaram, S.; Macarron, R.; Deller, M.C.; Wee, S.; Kim, S.; Wang, X.; Wu, L.; Yao, W.\nSource: Journal of Medicinal Chemistry, 2024\n\nAll reactions were carried out under an atmosphere of dry nitrogen. Unless otherwise noted all reactions were performed at ambient temperature, which averaged 23 \u00b0C. All solvents were anhydrous and used without further purification as acquired from commercial sources. No unexpected or unusually high safety hazards were encountered. Microwave reactions were performed on a Biotage Initiator+ system. NMR spectra were obtained using a Varian Mercury-400, Inova-500, or a Bruker Advance III HD 600 MHz spectrometer equipped with a 5 mm QCI Cryoprobe. Data for 1H spectra are reported as follows: chemical shift (\u03b4 ppm), multiplicity, coupling constant (Hz), integration, and are referenced to the residual solvent peak (2.50 for DMSO-d6, 7.26 ppm for CDCl3, and 4.79 for D2O). Data for 13C{1H} spectra are reported as follows: chemical shift (\u03b4 ppm), multiplicity, coupling constant (Hz), and are referenced to the residual solvent peak (39.52 ppm for DMSO-d6 and 77.16 ppm for CDCl3). Purifications by flash chromatography were performed on RediSep normal-phase silica columns using either a Biotage Isolera or a Teledyne ISCO CombiFlash system. Preparative HPLC purifications were performed on a Waters Fraction Lynx system using UV triggered or mass directed fractionation and compound-specific method optimization. All final compounds for biological testing were purified to >95% purity as determined by analytical LCMS.\n\n### Synthesis of Compound 2\n\n**Step 1. 2,5-Dichloro-4-(1-(tetrahydro-2H-pyran-4-yl)-1H-pyrazol-4-yl)pyrimidine (21).** To a mixture of 2,4,5-trichloropyrimidine (19, 0.330 g, 1.80 mmol), 1-(tetrahydro-2H-pyran-4-yl)-4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazole (20, 0.50 g, 1.80 mmol) and Na2CO3 (0.381 g, 3.60 mmol) in CH3CN/H2O (5:1, 6 mL) was added Pd(dppf)Cl2\u00b7CH2Cl2 (0.147 g, 0.180 mmol), and the reaction mixture was stirred at 80 \u00b0C for 2 h. After cooling to r.t., the reaction mixture was concentrated and the residue was purified by flash chromatography (SiO2, EtOAc/hexanes) to afford the title compound (450 mg, 84%) as a light yellow waxy solid. LCMS calculated for C12H13Cl2N4O (M + H)+: m/z = 299.0; found, 299.0.\n\n**Step 2. 5-Chloro-N-(1-(methylsulfonyl)piperidin-4-yl)-4-(1-(tetrahydro-2H-pyran-4-yl)-1H-pyrazol-4-yl)pyrimidin-2-amine (2).** A mixture of 2,5-dichloro-4-(1-(tetrahydro-2H-pyran-4-yl)-1H-pyrazol-4-yl)pyrimidine (30.0 mg, 0.100 mmol), 1-(methylsulfonyl)piperidin-4-amine (26.8 mg, 0.150 mmol), RuPhos Pd G2 (15.58 mg, 0.020 mmol), and Cs2CO3 (82 mg, 0.251 mmol) in 1,4-dioxane (0.5 mL) was stirred at 100 \u00b0C overnight. After cooling to r.t., the reaction mixture was diluted with MeOH and H2O, filtered, and purified by prep HPLC (pH 2) to afford the title compound (23.4 mg, 53% yield) as a white solid. LCMS calculated for C18H26ClN6O3S (M + H)+: m/z = 441.1; found, 441.2.\n\n### Synthesis of Intermediates 23 and 25\n\n**tert-Butyl 4-((4-chloro-5-(trifluoromethyl)pyrimidin-2-yl)amino)piperidine-1-carboxylate (23).** A mixture of 2,4-dichloro-5-(trifluoromethyl)pyrimidine (22, 11.4 g, 52.5 mmol) in 1,2-DCE (100 mL) and t-BuOH (100 mL) was cooled to 0 \u00b0C before ZnCl2 (1 M in Et2O, 74.9 mL, 74.9 mmol) was added, and the mixture was stirred at 0 \u00b0C for 1 h. To the reaction mixture was added tert-butyl 4-aminopiperidine-1-carboxylate (10.0 g, 49.9 mmol) followed by dropwise addition of a solution of N-ethyl-N-isopropylpropan-2-amine (8.72 mL, 49.9 mmol) in 1,2-DCE/t-BuOH (1:1, 15 mL). The reaction mixture was allowed to warm to r.t. before heating to 60 \u00b0C overnight. After cooling to r.t., the mixture was diluted with EtOAc (300 mL) and extracted with aqueous 1 M HCl. The aqueous phase was removed and the organic phase was washed with brine, dried over MgSO4, and concentrated. The residue was taken up in CH3CN (20 mL) and the mixture was heated to 50 \u00b0C for 1 h before cooling to r.t. overnight. The mixture was cooled to 0 \u00b0C for 1 h. The solid precipitate that formed was collected via filtration, washed with Et2O, and dried to afford the title compound (11.33 g, 60% yield) as a white solid. LCMS calculated for C11H13ClF3N4O2 (M \u2212 C4H8 + 1)+: m/z = 325.1; found, 325.0.\n\n**Step 1. 4-Chloro-N-(piperidin-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine hydrochloride (24).** A mixture of tert-butyl 4-((4-chloro-5-(trifluoromethyl)pyrimidin-2-yl)amino)piperidine-1-carboxylate (23, 1.905 g, 5.00 mmol) in THF (25.0 mL) was stirred at 60 \u00b0C for 10 min. After cooling to r.t., HCl (4 M in 1,4-dioxane, 5.0 mL, 20 mmol) was added and the reaction mixture was stirred at 60 \u00b0C for 8 h. After cooling to r.t., the mixture was sparged with N2 for 5 min before Et2O was added and the mixture was slurried for 15 min before the solid precipitate was collected via filtration, washed with Et2O, and dried to afford the title compound (1.40 g, 88% yield) as a white solid. LCMS calculated for C10H13ClF3N4 (M + H)+: m/z = 281.1; found, 281.0.\n\n**Step 2. 4-Chloro-N-(1-(methylsulfonyl)piperidin-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine (25).** A mixture of 4-chloro-N-(piperidin-4-yl)-5-(trifluoromethyl)pyrimidin-2-amine hydrochloride (1.40 g, 4.41 mmol) in THF (22 mL) was cooled to 0 \u00b0C before methanesulfonyl chloride (0.4 mL, 5 mmol) was added followed by N-ethyl-N-isopropylpropan-2-amine (3.0 mL, 17 mmol). The reaction mixture was warmed to r.t. and stirred for 2 h. The mixture was diluted with saturated aqueous NaHCO3 and extracted with CH2Cl2 and EtOAc. The combined organic phases were washed with brine, dried over MgSO4, and concentrated. The crude residue was taken up in CH3CN (7 mL) and heated to 50 \u00b0C for 1 h before cooling to r.t. overnight. The solid precipitate was collected via filtration and dried to afford the title compound (1.35 g, 85% yield) as a white solid. LCMS calculated for C11H15ClF3N4O2S (M + H)+: m/z = 359.1; found, 359.0.\n\n### Synthesis of Compounds 3\u20139", "# Cellular Uptake of Microplastics: Pathways, Persistence, and Therapeutic Potential for Intracellular Targeting of PETase\nAuthors: Deller, M.C.\nSource: marcdeller.com blog, 2025\n\nThe link between microplastic exposure and cancer progression operates through multiple interconnected mechanisms. Microplastics induce [chronic inflammation](https://www.mdpi.com/2079-4991/11/2/496) by triggering the production of pro-inflammatory cytokines including interleukin-1 beta, interleukin-6, and tumor necrosis factor alpha. This sustained inflammatory response creates a tumor-promoting microenvironment. [Oxidative stress](https://pmc.ncbi.nlm.nih.gov/articles/PMC9208838/) generated by microplastic exposure damages DNA, proteins, and lipids, leading to the formation of 8-hydroxy-2-deoxyguanosine and gamma-H2AX-marked double-strand breaks. Genotoxicity studies on human blood cells and cell lines confirm that nanoplastics can induce DNA damage. Furthermore, microplastics promote epithelial-mesenchymal transition, a process that enhances cancer cell invasiveness and migration capability. The presence of microplastics in tumor tissues at higher concentrations than in adjacent normal tissue suggests active bioaccumulation by cancer cells.\n\nEnzyme replacement therapy strategies designed for lysosomal storage disorders offer a conceptual framework for addressing microplastic accumulation in lysosomes. The [mannose-6-phosphate receptor pathway](https://pmc.ncbi.nlm.nih.gov/articles/PMC8757330/) represents the primary mechanism by which lysosomal enzymes are delivered to their target compartment. Cation-independent mannose-6-phosphate receptors on cell surfaces recognize and internalize enzymes bearing mannose-6-phosphate modifications, trafficking them through clathrin-coated vesicles to lysosomes. This natural delivery system has been exploited therapeutically through the development of mannose-6-phosphate-tagged glycopolypeptides that enhance lysosomal targeting. Alternative targeting strategies include fusion proteins incorporating insulin-like growth factor II peptides, which bind to the bifunctional IGF-II receptor. These [IGF-II-based lysosome-targeting chimeras](https://pubs.acs.org/doi/10.1021/acscentsci.3c01019) enable glycosylation-independent delivery of therapeutic proteins to lysosomes, offering advantages in production scalability and tissue-specific targeting.\n\nThe discovery of plastic-degrading enzymes in the human gut microbiome presents an unexpected biological response to plastic pollution. [HGMP01, a novel polyethylene terephthalate hydrolase](https://pubmed.ncbi.nlm.nih.gov/39681680/) identified from human gut metagenome data, demonstrates the capacity to hydrolyze PET nanoparticles. Comprehensive analysis reveals that HGMP01-like genes are widely distributed throughout the human gut microbiome, with over 20 percent exhibiting varying levels of expression. This finding provides biochemical evidence that intestinal microorganisms possess the ability to degrade PET particles, potentially mitigating some effects of dietary microplastic exposure. The enzyme shares relatively low sequence identity with previously characterized PET hydrolases from environmental bacteria, suggesting independent evolutionary origins.\n\n[FAST-PETase represents a breakthrough in engineered plastic degradation](https://pmc.ncbi.nlm.nih.gov/articles/PMC9007415/) through machine learning-guided enzyme optimization. This enzyme variant, developed using computational approaches to predict stabilizing mutations, exhibits remarkable activity at temperatures below 50 degrees Celsius. FAST-PETase can completely degrade various post-consumer PET products within one week at 50 degrees Celsius, converting the polymer into its constituent monomers terephthalic acid and mono(2-hydroxyethyl) terephthalate. The enzyme functions effectively across a range of pH values and handles temperature variations, making it suitable for non-laboratory conditions. Critically for biomedical applications, [engineered PETase can depolymerize PET microplastic particles into terephthalic acid and mono(2-hydroxyethyl) terephthalate in human serum at 37 degrees Celsius](https://pubs.rsc.org/en/content/articlelanding/2024/gc/d4gc00624k) without affecting cell viability. This demonstration opens possibilities for enzyme-based treatment of microplastics in bodily fluids.\n\nThe integration of plastic-degrading enzymes with lysosomal targeting technologies could enable cellular clearance of accumulated microplastics. Conjugating FAST-PETase or HGMP01 to mannose-6-phosphate-tagged carriers or IGF-II peptides might facilitate their delivery to lysosomes, where microplastics accumulate. Once localized in lysosomes, these enzymes could degrade trapped PET particles into less toxic monomers that cells can process or excrete. The mild acidic environment of lysosomes, maintained at pH 4.5 to 5.0, may influence enzyme activity and require optimization of enzyme variants for lysosomal conditions. Alternative approaches might involve nanocarrier systems displaying mannose-6-phosphate ligands on their surface, encapsulating plastic-degrading enzymes within pH-responsive shells that release their cargo upon lysosomal acidification.\n\nDespite these hurdles, the convergence of lysosomal targeting strategies with plastic degradation enzymology represents a promising frontier in addressing microplastic-associated health risks. The precedent established by enzyme replacement therapy for lysosomal storage disorders demonstrates that therapeutic proteins can be successfully delivered to lysosomes in patients. Advances in protein engineering, including directed evolution and machine learning-guided design, continue to improve enzyme properties. The discovery that endogenous human gut microbiome enzymes can degrade plastics suggests biological tolerance for these activities. As research progresses, the development of engineered enzyme therapeutics targeting lysosomal microplastic accumulation could complement pollution reduction efforts, offering a potential medical intervention for individuals with high exposures."], "kinds": ["qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "qa", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk", "chunk"]}