| ---
|
| license: cc-by-4.0
|
| tags:
|
| - structure-prediction
|
| - protein-ligand
|
| - zika
|
| - openbind-challenge
|
| ---
|
|
|
| # OpenBind Blind Challenge 1 — Method Report
|
|
|
| **Track:** Structure Prediction
|
| **Target:** Zika virus NS2B-NS3 protease
|
| **Compounds:** 400 / 400 submitted
|
|
|
| ---
|
|
|
| ## Summary
|
|
|
| Poses were generated by **MCS-guided template transfer followed by protein-aware
|
| constrained minimisation**. No GPU was used for the submitted poses. Co-folding
|
| (Boltz-2, OpenFold3) was evaluated and rejected on measured performance; its results
|
| are reported below as negative findings.
|
|
|
| The method is conditional rather than general: where a close structural analogue
|
| exists among deposited ZIKV protease complexes it is accurate, and where one does not
|
| it fails. That bimodality is quantified below rather than averaged away.
|
|
|
| ## Pipeline
|
|
|
| 1. **Receptor ensemble.** All ZIKV NS2B-NS3 entries were retrieved from the RCSB by
|
| sequence search on the 168-residue NS3 construct and superposed on 29HN (apo,
|
| closed, 1.774 A, P4(3)22 — the challenge crystal form) by NS3 Calpha, with residue
|
| correspondence established by sequence alignment rather than residue number.
|
| 195 of 200 members superposed within 1.0 A (median Calpha RMSD 0.253 A).
|
| The pocket has one meaningful degree of freedom, a bimodal Tyr161 rotamer, so two
|
| receptors are carried: 7H1Z (majority rotamer) and 7HOQ (minority).
|
|
|
| 2. **Template library.** 129 orthosteric ligand poses extracted in the common frame,
|
| disjoint from the validation set. Peptidomimetics, covalent adducts and macrocycles
|
| were excluded as unrepresentative of the blinded compound set.
|
|
|
| 3. **Template selection.** For each compound, an ECFP4 prefilter over the library
|
| followed by `rdFMCS` (ringMatchesRingOnly, completeRingsOnly, CompareElements /
|
| CompareOrderExact). The donor maximising **MCS coverage** — matched atoms divided by
|
| query heavy atoms — is chosen.
|
|
|
| 4. **Pose transfer.** Constrained embedding (ETKDGv3 with `coordMap`) pins the MCS
|
| atoms onto the donor's crystallographic coordinates; MMFF relaxation with position
|
| constraints on the pinned atoms completes the remainder. Pinned atoms land a median
|
| 0.002 A from their template positions.
|
|
|
| 5. **Protein-aware minimisation.** The MMFF relaxation above has no protein term, so
|
| peripheral atoms are drawn back into the receptor. Every protein heavy atom within
|
| 6 A is added to the force field as a fixed extra point and flat-bottom distance
|
| restraints push any too-close pair out to 0.80 x (sum of van der Waals radii), while
|
| MCS atoms are held to 0.30 A of slack. Median displacement 0.32 A.
|
|
|
| 6. **Assembly.** Receptor chains A (NS2B) and B (NS3) plus the ligand as a single
|
| `LIG` residue in chain C, waters stripped, explicit CONECT records. The receptor is
|
| chosen per compound to match the donor's Tyr161 rotamer.
|
|
|
| ## Validation
|
|
|
| Tuning and evaluation used a **held-out benchmark of 41 public ZIKV NS2B-NS3 holo
|
| structures**, disjoint from the template library, scored with OpenStructure
|
| (`SCRMSDScorer`, `LDDTPLIScorer`) and PoseBusters under the challenge's conjunction:
|
| BiSyRMSD < 1.5 A **and** lDDT-PLI > 0.8 **and** all PoseBusters checks. No leaderboard
|
| feedback was used at any point.
|
|
|
| A calibration of the three gates under rigid displacement showed that **lDDT-PLI binds
|
| before BiSyRMSD**: at 1.25 A displacement the RMSD gate still passes 100% while
|
| lDDT-PLI passes about a third. The effective accuracy requirement is nearer 1.1 A than
|
| the nominal 1.5 A.
|
|
|
| ### Head-to-head, 41-complex benchmark
|
|
|
| | method | passes all three gates |
|
| |---|---|
|
| | Boltz-2, rank-0 sample | 24.4% |
|
| | Boltz-2, oracle over 5 samples | 34.1% |
|
| | Boltz-2 with pocket restraints | 19.5% |
|
| | Template transfer | 34.1% |
|
| | **Template transfer + protein-aware minimisation** | **39.0%** |
|
|
|
| Boltz-2's 34.1% requires selecting the correct sample out of five, which its confidence
|
| score does not support (Spearman -0.59 against RMSD). The template method emits one
|
| pose, so its figure needs no selector.
|
|
|
| ### Performance is conditional on template precedent
|
|
|
| | MCS coverage | benchmark pass rate | compounds in this submission |
|
| |---|---|---|
|
| | >= 0.8 | 88% (15/17) | 220 |
|
| | 0.6 - 0.8 | 9% | 116 |
|
| | < 0.6 | 0% | 64 |
|
|
|
| MCS coverage is a usable *a priori* selector (Spearman -0.80 against BiSyRMSD), known
|
| before any compute is committed. Median coverage across the submission is 0.818.
|
|
|
| ## Negative results
|
|
|
| Reported because they constrain what the remaining gap is, not as asides.
|
|
|
| - **Pocket restraints degrade co-folding.** A paired run (identical seed and inputs;
|
| the unrestrained YAML is a strict text prefix of the restrained one) restraining the
|
| ligand to within 6 A of 25 pocket residues moved Boltz-2 from 25.9% to 19.0%.
|
| `minimum_distance_to_protein` failures rose from 44 to 69 of 205 models: the
|
| restraint drives the ligand too deep.
|
| - **Co-folding does not complement template transfer.** Stratified by MCS coverage,
|
| Boltz-2's oracle is 18% below 0.8 and 59% above it — the same compounds are hard for
|
| both. A hybrid adds little.
|
| - **Free docking is the worst option tested.** On the 22 low-coverage benchmark
|
| complexes, smina produced 0/22, scoring zero on *both* accuracy gates with a
|
| PoseBusters pass rate of 31.8%, against 4.5% and 54.5% for template transfer.
|
| - **Anchored search** (MCS core pinned, 60 conformers of the remainder ranked by the
|
| smina scoring function) gave 2/22 versus 1/22 — a sound mechanism but not a solution
|
| at this sample size.
|
|
|
| ## Limitations
|
|
|
| - Roughly 180 of the 400 compounds fall below 0.8 MCS coverage, where no method tested
|
| here exceeds ~10%. They are chemically novel rather than small (median 23-25 heavy
|
| atoms, Tanimoto 0.25-0.46 to their nearest donor).
|
| - 56% of benchmark poses still exceed the 1.5 A RMSD gate after minimisation; the
|
| method is not generally accurate.
|
| - Minimised poses remain less physically clean than co-folded ones (PoseBusters 71.8%
|
| versus 78.1% for plain Boltz-2).
|
| - The improvement from minimisation (+2 compounds of 39) is not statistically
|
| significant on its own (paired exact test, p = 0.25); it is adopted on mechanism and
|
| on the absence of any regression, not on the delta.
|
| - Performance depends on an unusually dense structural prior for this target (196
|
| liganded entries, one dominant crystal form, a rigid pocket). It would not transfer
|
| to a target with few co-crystals.
|
|
|
| ## Data and code declarations
|
|
|
| - **Proprietary data: none.** Every structure used is a public RCSB deposition and
|
| every compound SMILES came from the challenge distribution.
|
| - **Training data leakage:** not applicable to the submitted poses, which come from
|
| template transfer rather than a learned model. For the co-folding comparisons,
|
| benchmark structures were checked against each engine's training cutoff by
|
| *deposition* date (the conservative choice); all 42 fall outside the Chai-1, AF3 and
|
| OpenFold3 windows. Benchmark co-folding runs used `--use-templates false` to prevent
|
| inference-time retrieval of the reference structures.
|
| - **Software:** RDKit, OpenStructure, PoseBusters, gemmi, smina, Boltz-2 (v2.2.1),
|
| OpenFold3-preview — all open source.
|
|
|
