v2 validator: 3 hits PASS, methodology calibration documented
3-LLM consensus validator now PRODUCTION (v2)
The Spark Chai-1 saturator surfaces protein-ligand pair predictions continuously, and a Modal serverless A100 backend re-runs each high-score hit on Boltz-2 with 5 independent diffusion samples + the dedicated affinity head + a pose-geometric pocket check. Verdicts post automatically to _modal_validations/ after every saturator tick.
Three findings re-validated 2026-04-27
| Hit | Spark Chai-1 (1 sample, screening) | Modal Boltz-2 v2 (5 samples + affinity + pose) | Verdict |
|---|---|---|---|
| ROCK2_KD × Fasudil | iPTM 0.86 | iPTM 0.98, affinity_prob 0.71, 20/20 ligand atoms in pocket | PASS_consensus_strong — recapitulates the clinically validated ROCK inhibitor profile (Fasudil approved Japan 1995) |
| MuSK_KD × Fasudil | iPTM 0.69 | iPTM 0.98, affinity_prob 0.33, 20/20 ligand atoms in pocket | PASS_borderline_weak_binder — new NMJ-axis cross-talk: Fasudil weakly engages MuSK kinase domain in addition to ROCK1/2. Hypothesis: a fraction of Fasudil's potential SMA effect could route through NMJ in addition to actin-pathway ROCK inhibition. Needs orthogonal validation. |
| LIMK2_KD × 4-AP | iPTM 0.64 | iPTM 0.97, affinity_prob 0.31, 7/7 ligand atoms in pocket | PASS_borderline_weak_binder — adds LIMK2 to the existing 4-AP multi-target profile (CORO1C, NCALD, SMN1/2, UBA1, Kv1.2 already documented in the 4-AP-Research package). Weak binder in the µM range, consistent with 4-AP's known multi-target nature. |
Methodology calibration finding
A first version of the validator was raising false-FAIL on real binders. Empirical proof:
- CORO1C × 4-AP — known REAL binder per the existing 4-AP-Research package (DiffDock +0.251, 100 ns MD stable, FEP favourable, mutant R→A abolishes binding) — returned
iPTM = 0.000across 5 Modal Boltz-2 samples in v1. - ROCK2 × Fasudil — clinically validated ROCK inhibitor — same
iPTM = 0.000in v1.
Boltz-2 v2.2's iPTM metric is a protein-protein interface confidence; for ligands smaller than ~30 heavy atoms there is not enough surface to score a reliable interface PTM. The fix is to enable Boltz-2's dedicated affinity head via properties: - affinity: binder: B in the input YAML, which produces calibrated affinity_pred_value (pIC50-like) and affinity_probability_binary outputs, and to add a pose-geometric pocket check (count ligand heavy atoms within 5 Å of any protein heavy atom in the output mmCIF).
After the fix:
- iPTM came up from 0.0 to ~0.97 on the same input
- All three hits above were re-classified from FAIL to PASS
Architecture
- NIM (cloud API): esmfold + molmim + boltz2 saturator, ~30k calls/day
- Dell RTX Pro 6000 (free 90-day demo): Boltz-2 PPI saturator, 4400+ pair predictions
- NVIDIA DGX Spark GB10 (owned, 128 GB unified): Chai-1 ligand saturator + Qwen 35B local LLM endpoint, 621 unique target × ligand pairs in current rotation
- Modal A100 serverless (free $30/mo): on-demand v2 validation of any saturator hit with
aggregate_score ≥ 0.5, ~6¢ per 5-sample validation
The validator runs every 30 minutes; verdicts (PASS_consensus_strong / PASS_borderline_weak_binder / FAIL_no_binding / AMBIGUOUS_needs_AF3) are written to research-data/SMA/_modal_validations/<date>/<target>__<ligand>.json and ingested into the platform Postgres on the next tick.
What this does NOT change
The existing 4-AP-Research package (CORO1C-led, 5 SMA targets, DiffDock + 100 ns MD + FEP + mutant + selectivity panel) remains the canonical 4-AP evidence. The v2 validator confirms structural binding (7/7 atoms in pocket) and quantifies the affinity range as weak-µM — the exact regime expected for a multi-target compound, never claimed otherwise.
Pharma rule
External claims still require ≥ 2 independent backends in agreement plus AF3 manual upload as the third opinion when the AF3 server's CCD-only ligand whitelist permits it. Where AF3 server cannot accept the ligand (e.g. 4-AP, Fasudil, LIMKi3), the third opinion is replaced by a 100 ns MD simulation.