STRC Research

STRC h01 Phase 9e Blind Neural Docking AF3 DiffDock 2026-04-27 SKELETON

5 min read

STRC h01 Phase 9e — Blind Neural Docking (SKELETON)

Placeholder. Independent AI co-folding test: do multiple unrelated neural docking models naturally place the lead ligand in the K1141 pocket of the mutant but not WT, without any seeded coordinates? No production calculation yet.

Question

Given only the protein sequence (or AF-derived structure) and ligand SMILES, and no docked starting pose:

  1. Does AlphaFold 3 co-fold place the ligand in the proposed pocket for the mutant?
  2. Does DiffDock independently agree?
  3. Does Boltz-2 agree?
  4. Does Uni-Mol Docking agree?
  5. Differential: does the AI consensus place the ligand in the pocket for the mutant but not WT (or with materially better confidence in mutant)?

This is the closest in-silico analogue to “an unbiased experimentalist who’s never heard our hypothesis” — the AI models were not trained on STRC and have no knowledge of our docking/MD pipeline.

Method

For each of {AF3, DiffDock, Boltz-2, Uni-Mol}, for each lead ligand × {WT, mutant}:

  1. Submit protein sequence (AF3, Boltz-2) or AF-structure (DiffDock, Uni-Mol) + ligand SMILES.
  2. No pocket constraint, no seeded coordinates, no template ligand.
  3. Generate top-N predicted complexes (typically N=5).
  4. Score: distance from predicted ligand COM to known K1141 pocket centroid; predicted ipTM/pae/confidence per model.

Tools and access:

  • AF3 server: 23+ jobs already completed for STRC variants (per reference_strc_af3_data memory). API access established. Server limit ~50 jobs/day.
  • DiffDock: open-source GitHub (gcorso/DiffDock). GPU run. Free.
  • Boltz-2: already integrated in Phase 5q v5.3 pipeline. Re-use installation.
  • Uni-Mol Docking: open-source (deepmodeling/Uni-Mol). GPU run.

Why orthogonal

  • Different model families: AF3 (transformer + diffusion), DiffDock (E(3)-equivariant diffusion), Boltz-2 (Boltz-style architecture), Uni-Mol (SE(3)-transformer + 3D pretraining). No shared training set or architecture.
  • Different from physics-based: pure data-driven; no force field; no scoring function tuned on docking benchmarks (DiffDock and AF3 are co-folders trained primarily on PDB structural data).
  • Massive validation if positive: if the AI ensemble agrees and shows differential preference for the mutant → the pocket is real, the pose is plausible, and the mechanism is consistent with everything the AI ensemble learned from PDB.
  • Massive concern if negative: if no AI co-folder places ligand in K1141 pocket for either variant → either pocket is non-canonical (cryptic, induced-fit) and AI cannot find it, or the docked pose is artifactual.

Inputs needed

  • Protein sequence: STRC (Q7RTU9) WT and E1659A mutant.
  • Lead ligands from Phase 5q top-5 (SMILES + curated InChI).
  • Hardware: AF3 server free; DiffDock + Uni-Mol need GPU (M5 Max MPS verify, fall back to remote A100).
  • Wall time per system: AF3 ~30 min/job; DiffDock ~5 min/job; Boltz-2 ~10 min/job; Uni-Mol ~5 min/job.

Smoke test (1-day, theoretical)

Single lead ligand × {WT, mutant} × all 4 models:

  1. Submit 8 jobs total (4 models × 2 variants).
  2. Score predicted ligand COM distance from K1141-Cα.
  3. Visual sanity check on top-1 pose per model.

Smoke pass: at least 2 of 4 models place ligand within 8 Å of K1141-Cα for the mutant; and show qualitatively different pose distribution from WT.

Smoke fail: all 4 models place ligand far from pocket in both variants → either the pocket is genuinely cryptic and not co-fold-findable, or the lead ligand has no AI-detectable affinity. Reconsider lead choice or accept this is a “negative-orthogonal-check” data point.

Production protocol (theoretical)

  • 5 lead ligands × 4 models × 2 variants = 40 jobs.
  • N=5 predictions per job → 200 predicted complexes total.
  • Per ligand × variant aggregate: fraction of (predictions, models) that place ligand in pocket; mean confidence (ipTM, pae for AF3; DiffDock confidence; Boltz-2 ipTM); pose-clustering across models.
  • Cross-model agreement matrix.

Pass criteria

  • PRIMARY: ≥ 2 of 4 models place lead ligand within 8 Å of K1141-Cα in mutant (top-1 prediction) for ≥ 3 of 5 leads.
  • SECONDARY: model confidence (ipTM or DiffDock confidence) is materially higher in mutant than WT for ≥ 1 lead × ≥ 2 models.
  • TERTIARY: predicted pose clusters across models (median pairwise RMSD < 4 Å on lead ligand) — different model families converging on the same pose is strong validation.
  • FAIL state: no model places ligand in pocket for either variant → cryptic-pocket explanation needed (link to Phase 5c/5c-Mutant) or lead ligands lack AI-detectable affinity.

Known artifacts and risks

  • Training-set leakage: AF3 may have seen STRC homologues. Document whether the predicted pose is novel or re-using a templated complex.
  • AF3 confidence calibration on small molecules: ipTM is well-calibrated for protein-protein but less so for protein-ligand; treat as ordinal rather than absolute.
  • DiffDock’s pocket-blindness: when run blind (no pocket given), DiffDock can scatter poses across the full surface; aggregate over N=20-40 predictions for a stable signal.
  • Sequence-based co-folders (AF3, Boltz-2): structure quality depends on MSA depth — STRC has limited mammalian homologues, MSA may be shallow.
  • Mutation effect: a single E→A point mutation may not change predicted structure enough for AF-class models to predict differential ligand binding (their resolution on micro-environment changes is debated). If both variants give identical poses, this is a known limitation, not a refutation.

References (canonical)

  • Abramson et al. 2024. Nature 630:493 — AlphaFold 3.
  • Corso et al. 2023. ICLR — DiffDock.
  • Wohlwend et al. 2024. bioRxiv (Boltz model paper).
  • Zhou et al. 2023. NeurIPS — Uni-Mol.
  • Lu et al. 2024. Nature Methods — review of AI docking benchmarks.

Status

  • 2026-04-27 — skeleton created. AF3 + Boltz-2 already integrated. DiffDock + Uni-Mol install: TBD. No phase-9e runs performed.

Ranking delta

  • No change. Skeleton only.

Connections

Graph View

Backlinks