Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261015T141849Z-rare-cell-7k-8k-panels/diffusion/dataset/server_1/model8/plans/rare_gene_context_expansion_20261015/research/ilc_dc/REPORT.md

@@ -0,0 +1,59 @@
+# Human ILC and rare DC gene-context evidence
+
+Research date: 2026-10-15 (Asia/Seoul). Scope: PBISC-Diffusion Model 8 candidate gene-context expansion; literature evidence supplier only. No models, frozen panels, training, inference, donor selection, or manuscript claims were changed.
+
+## Decision
+
+Use gene-context expansion to retain **joint lineage, alternative-lineage and state programs**. The literature does not support a rule that 8,000 genes is intrinsically better than 6,000 or 7,000. Whether these candidates improve rare-population recovery remains an empirical question on independent donors and cohorts.
+
+`evidence.csv` contains **7,076 evidence rows and 4,349 distinct source gene strings** from nine primary papers. This is a searchable catalog, not 4,349 required additions. There are **89 unique core candidates**; supporting and optional tissue/state sets overlap. Exact category counts are in `summary.json`. `core` means a strong candidate for retaining expression context, not a sufficient classifier or an expression requirement in every cell. Symbols remain as published for the integrator to reconcile with the canonical 20,097-gene universe and Ensembl identifiers.
+
+`protected_sentinels.csv` identifies17 narrow ILC functional/receptor candidates to retain before broad DC DE catalogs: NCR2, RORC, AHR, IL23R, CCR6, IL22, IL17A, IL17F, CSF2, TNFSF11, TNFSF4, IL1RL1, IL17RB, IL13, IL5, PPARG, AREG. This budget decision preserves the ability to assess rare ILC function; it does **not** promote tissue/activated evidence to universal blood identity. IL17F evidence is human stimulated ILC3 protein production, explicitly distinct from observed resting PBMC RNA. `core_panels.csv` retains101 source rows for the89 core symbols, including contrast directions.
+
+The main evidence was actually downloaded: Mazzurana Data S1/S2, Lim RNA/ChIP overlap and RNA differential tables, Villani Tables S1–S16, Bourdely Table S1, Maier Tables S1/S2, and Dutertre's supplementary workbook. Source copies and hashes are retained. There is no invented universal FDR cutoff. Source statistics, directions, sheet names and Excel row numbers are recorded where available.
+
+## Blood-focused identity and boundary decisions
+
+| Population | Positive context worth retaining | Essential alternative interpretation | Scope and evidence |
+|---|---|---|---|
+| ILC2 | GATA3, MAF, HPGD, HPGDS, MBOAT2, IL9R; blood PTGDR2, S1PR2, CCR2, TNFSF10, TNFRSF19, CD200R1 | GATA3 also occurs in T cells; PTGDR2/CRTH2 can decrease after activation. CD3/TCR evidence must be evaluated jointly. | Human blood/tonsil/lung RNA, Mazzurana Fig.4 and exact module50. |
+| Circulating ILCP / naïve ILC | IL7R, KIT, CD7, IL1R1 phenotype; ID2, TOX, TCF7, GATA3, ETS1, RORA, ZBTB16, NOTCH1 program | KIT+ does not imply mature ILC3. TCF7, IL7R and CD7 are shared with other lymphocytes. Precursor TFs vary by developmental stage. | Adult blood functional lineage potential and bulk RNA/ChIP, Lim Figs.1–2; blood B_nILC RNA in Mazzurana DataS1. |
+| Putative blood ILC1 | Retain IL7R/ILC context and TBX21/IFNG type-1 context, alongside alternative T/NK programs | **No definitive pure circulating ILC1 RNA gate is supplied.** Mazzurana's B_ILC1/EOMES+ ILC1 cells often had rearranged TCR and T-cell-like transcription. | Blood label requires explicit ontology review; tissue ILC1 findings cannot be transferred automatically. |
+| Mature ILC3 | RORC, AHR, IL23R, KIT, IL1R1, CCR6, NCR2, NRP1, TOX, IKZF2, TNFSF11, TNFSF4; broader exact modules3/13 | **No universal blood ILC3 requirement is supplied.** Blood KIT+ ILCP can lack mature ILC3 cytokines/TF protein. IL22/IL17A absence does not reject a precursor. | Primarily colon, lung and tonsil; optional tissue programs for a PBMC-first panel. |
+| NK vs ILC | NK/cytotoxic module28 includes NKG7, GNLY, PRF1, KLRD1, GZMB, FGFBP2, FCGR3A, EOMES, S1PR5 | NKG7, TBX21 or IFNG alone do not identify NK. Cytotoxic T cells share this module. | Exact human RNA module; compare with TCR/CD3 and ILC programs. |
+| T vs ILC | CD3D, CD3E, CD3G, CD4, CD8A/CD8B, CD5, CD6, CD27, LEF1, BCL11B; inspect TCR constant/rearrangement evidence when available | CD3 surface negativity and sparse RNA non-detection do not establish absence of a T-cell program. TCR rearrangement is not recoverable from a few ordinary expression markers. | Mazzurana Figs.6–7 and exact module32; V(D)J evidence is an orthogonal assay. |
+| cDC1 | CLEC9A, XCR1, CADM1, plus ranked WDFY4/BATF3/IRF8 context | THBD/CD141 alone is shared with other DC populations. | Villani blood TableS2; broad tumor DC1 list is optional. |
+| pDC | TCF4, IRF7, TLR7, SLC15A4, PACSIN1, DERL3, LAMP5, SCAMP5, SLA2, PTCRA, PTPRCAP; ranked GZMB/CLEC4C context | IL3RA/CD123 and CLEC4C/CD303 also occur in pDC-like ASDC. TCF4 here is E2-2, not TCF7L2. | Villani Figs.4–5, TableS2. |
+| ASDC / AXL+SIGLEC6+ DC / tDC | AXL, SIGLEC6, SIGLEC1, CD22; PPP1R14A, DAB2, S100A10, FAM105A, MED12L, ALDH2, LTK | ASDC spans pDC-like and cDC-like profiles. AXL alone is not specific. Do not equate all tDC with conventional pre-cDC. | Villani exact top-ten purity program; Sulczewski human Fig.4i–k and methods. |
+| cDC2 vs DC3 vs monocyte | Shared DC: CD1C, FCER1A, CLEC10A, CD74; DC3/monocyte-shared CD14, S100A8/A9/A12, VCAN; monocyte C5AR1, FCAR, SOD2, RBP7 | CD14/S100 alone do not establish DC3. CD163, BTLA and CD5 protein can be informative despite poor RNA detection. Require a monocyte reference. | Bourdely Fig.1 and exact blood bulk contrasts; Dutertre blood/tissue boundary validation. |
+| mregDC / LAMP3+ mature DC state | LAMP3, CCR7, FSCN1, CD40, CD200, CD274, PDCD1LG2, CCL19, CCL22, RELB, IL12B and exact human TableS2 program | Maturation/regulation is a state accessible from DC lineages; **not synonymous with inflammatory blood DC3**. | Maier human NSCLC Fig.4/TableS2, optional tissue/state evidence. |
+
+The table is a compact interpretation of the catalog, not an extra unsourced gene list. Full source membership is in `evidence.csv`; long programs are not reproduced here. Some supporting candidates are shared, low-specificity, sex-linked, housekeeping or activation genes. Their inclusion is evidence for source expression, not evidence that they improve PBISC.
+
+## What each source contributes
+
+- **Mazzurana et al., 2021**, [Cell Research](https://www.nature.com/articles/s41422-020-00445-x), DOI `10.1038/s41422-020-00445-x`: blood-specific selected transcripts for ILC2/B_nILC, exact WGCNA ILC2/NK/T/ILC3 modules, and tissue-context caveats. The DataS1 DE_modules header explicitly uses permissive input (absolute log2FC >0.1 and nominal p<0.1). It does not mean every module member is FDR-significant or independently specific. Author-selected transcript sheets use the **named gene column**; row names such as CD3D2 are not separate genes. ILC2 activation includes IL1RL1/IL17RB/IL13 and reduced CRTH2; these were labeled optional tissue/state.
+- **Lim et al., 2017**, [Cell](https://www.sciencedirect.com/science/article/pii/S009286741730199X), DOI `10.1016/j.cell.2017.02.021`: an exact **398-gene RNA/ChIP overlap** from mmc3.xlsx, plus adult blood ILCP phenotype and functional potential. This is an ILC commitment program against CD34+ HSC, not a 398-gene discriminator against mature T/NK. The full 5,160-row RNA table is downloaded but not imported wholesale. The article calls the overlap Table S3; neighboring workbook numbering is inconsistent, so provenance includes exact filename/header, not table number alone.
+- **Villani et al., 2017**, [Science](https://pmc.ncbi.nlm.nih.gov/articles/PMC5775029/), DOI `10.1126/science.aah4573`: actual TableS2 ranked blood DC signatures and TableS3/S6 contrasts. Whole published rank lists are preserved as supporting, including ranks below the article's emphasized AUC≥0.85 group. No claim is made that all rows pass that AUC threshold. DC4 is kept as a study-specific monocyte-boundary label. The initial discovery draw and later validation donors should not be confused with a large independent atlas.
+- **Bourdely et al., 2020**, [Immunity](https://pmc.ncbi.nlm.nih.gov/articles/PMC7445430/), DOI `10.1016/j.immuni.2020.06.002`: separate blood DC3 versus CD5−/CD5+ cDC2 and monocyte contrasts, with original directions and adjusted p-values. TLR-activated contrasts are **optional_state**. In-vitro developmental and humanized-mouse tables were not promoted to universal blood identity. The publisher workbook is Strict OOXML, so the extraction reads original XML cells; it does not modify the source.
+- **Dutertre et al., 2019**, [Immunity](https://www.sciencedirect.com/science/article/pii/S1074761319303346), DOI `10.1016/j.immuni.2019.08.008`: corroborating CD88/CD89 versus FcεRIα/HLA-DQ boundary and CD5/CD163/CD14 subdivision. The full workbook is downloaded. Its large SLE/control expression table and numbered cluster tables were **not** turned into required candidate lists: figure-sheet names are inconsistent with their content, and explicit cluster/contrast mapping would be needed. SLE-specific use must be flagged as endpoint-informed if the same disease finding becomes an evaluation target.
+- **Maier et al., 2020**, [Nature](https://www.nature.com/articles/s41586-020-2134-y), DOI `10.1038/s41586-020-2134-y`: exact **human** DC1/DC2/mregDC scoring lists from TableS2 rows16–18. Mouse rows are excluded. All are optional_tissue because the source is lung/NSCLC; the long lists include broad shared and proliferative programs. They are available for context and independent state evaluation, not universal PBMC requirements.
+- **Sulczewski et al., 2023**, [Nature Immunology](https://www.nature.com/articles/s41590-023-01545-7), DOI `10.1038/s41590-023-01545-7`: updated tDC/ASDC ontology. Human blood tDC cultured for two days gained a CD5+ DC2 phenotype, while mechanistic lineage-tracing evidence is largely murine. Human viral/CpG IL-1β evidence is explicitly protein/state evidence. No mouse core-transcriptome list was converted into human genes.
+- **Bernink et al., 2013**, [Nature Immunology](https://www.nature.com/articles/ni.2534), DOI `10.1038/ni.2534`: tissue ILC1 type-1 function and source-specific NK-negative phenotype. These seven rows rely on the verified abstract/figure description and are labeled at that evidence level. No unseen supplement genes are claimed.
+- **Konya et al., 2018 (online2017)**, [JACI](https://pubmed.ncbi.nlm.nih.gov/28433688/), DOI `10.1016/j.jaci.2017.01.045`: human mucosal ILC3 IL-17F functional sentinel. Verified primary abstract and ELISA methods show stimulated tonsil/gut ILC3 production and vitamin-D-associated suppression. Only IL17F was added from this source; an unseen microarray gene list was not inferred.
+
+## Use in 7k/8k candidate construction
+
+1. Resolve source gene strings against the canonical panel universe, preserve Ensembl identities, and review old aliases rather than creating duplicate biological features. Protein names are not gene symbols: CD127=IL7R, CD117=KIT, CRTH2=PTGDR2, CD88=C5AR1, CD89=FCAR, CD141=THBD, CD303=CLEC4C. Broad HLA-DQ or TCR family descriptions must not be invented as single genes.
+2. First inspect missing blood core anchors and **complete complementary discriminating programs**. A missing T-cell alternative marker can matter as much as a positive ILC marker. Then rank source-supported blood candidates by program coverage and specificity. Source count must count independent studies, not repeated contrast rows.
+3. Treat optional tumor/tissue/activation programs as a separate catalog. Do not consume the 2,000-gene expansion budget merely because a DE sheet is long. Disease programs can characterize a known lineage but cannot replace lineage identification.
+4. Retain relevant genes in the generative context even when they will be used for evaluation. Instead of deleting them, preregister independent annotation evidence and endpoint sets. A classifier built from a particular paper's markers and scored on the same markers measures **panel fidelity**, not independently validated rare-population recovery.
+5. Audit classifier label availability and mapping before claims: does it resolve ILCP from ILC3, ASDC from pDC, inflammatory DC3 from mregDC, and T-like ILC1 from T cells? A larger context cannot create a missing ontology. Do not force unresolved cells into a named rare class.
+
+## Remaining limits
+
+- No live corpus expression, donor coverage, model ablation, or classifier validation was executed in this research supplier task. The integrator owns canonical availability and final 7k/8k files.
+- RNA dropout and low RNA/protein concordance mean that “not detected” is not “absent.” Use multi-gene likelihood/module evidence and uncertainty; true surface-negative lineage gates need protein or orthogonal evidence.
+- Mature blood ILC3 and pure blood ILC1 remain ontology-sensitive. Tissue signatures and precursor functionality cannot substitute for evidence of those exact populations in the paired PBMC samples.
+- Source programs can contain contamination or shared biology. Especially broad DC1/DC2 tumor lists and CD5− DC2 antibody-associated genes need biological adjudication before interpretation.
+- No archive delta was published by this subagent. The root coordinator is responsible for the standing Model 8 archive workflow for this completed research bundle.

paper-releases/pbisc-diffusion/updates/20261015T141849Z-rare-cell-7k-8k-panels/diffusion/dataset/server_1/model8/plans/rare_gene_context_expansion_20261015/research/ilc_dc/SEARCH_LOG.md

@@ -0,0 +1,58 @@
+# Search and retrieval log
+
+Date: 2026-10-15 (Asia/Seoul). Search scope: human blood/PBMC ILC1/2/3/ILCP, NK/T disambiguation, pDC/cDC1/cDC2/DC3/ASDC, mregDC. Primary publications and their own supplements were preferred. No GPU, paid API, generation or external messages were used.
+
+## Search families and decisions
+
+| Search | Primary results used | Decision |
+|---|---|---|
+| human blood innate lymphoid cells single-cell RNA-seq; tissue-specific transcriptional imprinting | Mazzurana 2021, Lim 2017; Björklund 2016 identified | Download Mazzurana blood and tissue tables; Lim precursor program. Björklund tonsil dataset is relevant but was not independently extracted; Mazzurana reuses tonsil data so do not count it as independent corroboration. |
+| Villani2017 single-cell blood dendritic cells supplementary tables | Villani 2017 | Retrieve original TableS1–S16 workbook and inspect named sheets before extracting. |
+| CD163 DC3 Bourdely Dutertre monocyte CD88 CD89 | Bourdely 2020, Dutertre 2019 | Separate resting blood, activated blood, cultured cells and tissue; preserve pairwise direction. |
+| conserved dendritic-cell regulatory program limits antitumour immunity | Maier 2020 | Download TableS2, select species=human only; mark all tumor/lung lists optional. |
+| ASDC transitional dendritic cells pre-cDC | Sulczewski 2023 | Use primary 2023 evidence to qualify pre-DC synonym and human-versus-mouse findings. |
+| human type1 ILC inflamed mucosal tissue | Bernink2013 | Use verified abstract/figure descriptions for tissue support; no unseen full transcriptome list. |
+| human tonsil ILC3 IL17F; vitamin D IL-23 receptor pathway | Konya2018 (online2017), PubMed28433688 | Primary abstract and publisher ELISA methods explicitly support human IL17F functional output. One protein/state row added; no unverified RNA signature. |
+
+Additional discovered sources (not evidence-list suppliers) included human fetal ILC development, 2024 lymph-node/BM/spleen ILC atlas, lung ILC in humanized mice, and integrative ILC reanalyses. They were not imported as universal blood evidence because of age/tissue/species context, secondary reanalysis, or lack of independently retrieved program definitions. Reviews were used only to find primary publications, not to manufacture gene lists.
+
+## Successful source retrievals
+
+| Local source | Exact retrieval endpoint | Verified content |
+|---|---|---|
+| PMC8089104.xml | https://www.ebi.ac.uk/europepmc/webservices/rest/PMC8089104/fullTextXML | Mazzurana full JATS, supplement names and stated source MD5s. |
+| Mazzurana_DataS1_alt.xlsx | https://media.springernature.com/full/springer-static/esm/art%3A10.1038%2Fs41422-020-00445-x/MediaObjects/41422_2020_445_MOESM10_ESM.xlsx | 4,934,137 bytes; named sheets, 4,410 module memberships. |
+| Mazzurana_DataS2_alt.xlsx | https://media.springernature.com/full/springer-static/esm/art%3A10.1038%2Fs41422-020-00445-x/MediaObjects/41422_2020_445_MOESM11_ESM.xlsx | 416,063 bytes; cluster-specific contrasts. Retrieved but not required for broad blood list. |
+| Mazzurana_supp.zip | https://www.ebi.ac.uk/europepmc/webservices/rest/PMC8089104/supplementaryFiles | Official supplement bundle; independently available source copy. |
+| Villani.html | https://pmc.ncbi.nlm.nih.gov/articles/PMC5775029/?pdf=true | Full article and exact supplement URLs. |
+| Villani_Tables1-16.xlsx | https://pmc.ncbi.nlm.nih.gov/articles/instance/5775029/bin/NIHMS910854-supplement-Supplementary_Tables_1-16.xlsx | 867,260 bytes; actual ZIP/XLSX parsed. Public download page required its own small SHA-256 proof-of-work cookie; followed the site's JavaScript protocol, no credentials. |
+| PMC7445430.xml | https://www.ebi.ac.uk/europepmc/webservices/rest/PMC7445430/fullTextXML | Bourdely full JATS and original publisher identifier. |
+| Bourdely_TableS1.xlsx | https://ars.els-cdn.com/content/image/1-s2.0-S1074761320302326-mmc2.xlsx | 408,823 bytes; 11 named Strict-OOXML sheets. Parsed directly from original XML because installed openpyxl omitted Strict sheets. |
+| Bourdely_supp_europe.zip | https://www.ebi.ac.uk/europepmc/webservices/rest/PMC7445430/supplementaryFiles | Official supplement bundle. |
+| Maier.html; Maier_nature.html | https://pmc.ncbi.nlm.nih.gov/articles/PMC7787191/?pdf=true ; https://www.nature.com/articles/s41586-020-2134-y | Full article and official attachment links. |
+| Maier_TableS1.xlsx | https://media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41586-020-2134-y/MediaObjects/41586_2020_2134_MOESM2_ESM.xlsx | Sample/analysis metadata, not a marker list. |
+| Maier_TableS2.xlsx | https://media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41586-020-2134-y/MediaObjects/41586_2020_2134_MOESM3_ESM.xlsx | Explicit human rows16–18 and mouse rows, species verified before export. |
+| Lim2017_author.pdf | https://www.limmunity.com/_files/ugd/a41686_3e3a774a659e4e6295a6059f9c2a0b65.pdf | Author-hosted article, local text extracted to verify Figs.1–2 and context. |
+| Lim2017_mmc2.xlsx | https://ars.els-cdn.com/content/image/1-s2.0-S009286741730199X-mmc2.xlsx | Full ILCP/CD34 RNA contrast, retained as source only. |
+| Lim2017_mmc3.xlsx | https://ars.els-cdn.com/content/image/1-s2.0-S009286741730199X-mmc3.xlsx | Exact header says398 RNA-Seq/ChIP-Seq common elements; all398 verified. |
+| Lim2017_mmc4.xlsx; Lim2017_mmc5.xlsx | https://ars.els-cdn.com/content/image/1-s2.0-S009286741730199X-mmc4.xlsx ; https://ars.els-cdn.com/content/image/1-s2.0-S009286741730199X-mmc5.xlsx | Cloning/tissue metadata; no candidate genes extracted. |
+| Dutertre2019_mmc2.xlsx | https://ars.els-cdn.com/content/image/1-s2.0-S1074761319303346-mmc2.xlsx | 714,380 bytes; scRNA, SLE/control and subset contrasts. Kept as source, not bulk-imported because cluster/figure mapping requires further adjudication. |
+| Transitional2023.html | https://pmc.ncbi.nlm.nih.gov/articles/PMC10683792/?pdf=true | Human tDC culture and stimulation paragraphs inspected separately from mouse mechanism. |
+
+## Failed or limited access
+
+- Web-reader access to several PMC pages returned recaptcha; the public article `?pdf=true` HTML endpoint and EuropePMC fullTextXML were successful alternatives.
+- Legacy `static-content.springer-cdn.com` did not resolve locally; the publisher's current `media.springernature.com` endpoint worked.
+- EuropePMC fullTextXML for the author manuscripts PMC5775029 and PMC7787191 returned500. Their public PMC full HTML was retrieved.
+- Direct PMC attachment requests initially returned1817-byte “Preparing to download” HTML, not Excel. These were **not accepted as successful source files**. Villani retrieval succeeded after the site's published download verification.
+- EuropePMC Villani supplementaryFiles returned a small XML error response, not a ZIP. This response is retained as a failed retrieval receipt and is not a source table.
+- The browser connector lacked its configured Chrome executable; no installation or persistent browser configuration change was made.
+- Dutertre full HTML returned403 locally; indexed primary article text and official supplementary workbook were available. Guessed mmc3/mmc4 XLSX endpoints returned404 and contributed no genes.
+- Bernink2013 full supplementary tables were not retrieved. Seven rows remain abstract/figure-description level, explicitly marked; no unverified genes were supplied from its supplement.
+- Konya2018 microarray supplement was not retrieved. The single IL17F row is primary abstract/methods-level functional protein evidence, not resting-blood RNA evidence.
+
+## Extraction and quality checks
+
+`build_evidence.py` deterministically rebuilds the catalog from original source bytes plus source-localized curated rows. Check actual source file sizes, SHA-256/MD5, per-source counts and extracted program membership in `source_manifest.json`, `summary.json`, `verification.json` and `SHA256SUMS`.
+
+No global significance threshold is imposed across different assays. Exact author program membership is distinguished from ranked DE tables, article statements, protein phenotypes and abstract-level evidence. The 398-gene ILCP overlap is verified as398; source human/mouse tags and positive/negative contrast signs are preserved. Tissue-only lists are optional, and broad supporting lists are not required additions.