John-Burden/Open_Fold_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.