STRC Research

STRC h01 Phase 5 MD Ensemble Rescoring 2026-04-23

10 min read

STRC h01 Phase 5 MD Ensemble Rescoring 2026-04-23

TL;DR. Phase 5a OpenMM MD pipeline validated on local Mac (62 ns/day OpenCL Metal on 164 k-atom Ultra-Mini × TMEM145 chain A; 2 ns production delivered 20 snapshots). Phase 5b ensemble re-docking of Phase 4b top-5 leads + diflunisal positive control against each snapshot → all f_PC estimates at [L]=10 μM are BELOW 0.10, far below NORMAL-rescue threshold of 0.50. Best binder (diflunisal positive control) at f_PC = 0.083; best lead (naphthalene-2-COOH) at f_PC = 0.047. Phase 4b single-structure rankings were over-optimistic by ~0.5-1 kcal/mol: ensemble MD-relaxed receptor reveals the K1141 pocket is LESS favorable for carboxylate ligands than the static snapshot suggested. RED-LIGHT: current shortlist cannot deliver monotherapy NORMAL rescue for Misha via h01; need second-wave virtual screen (DrugBank FDA + ZINC22 carboxylate tranche + non-carboxylate scaffolds) OR de novo design. h01 tier A held; next-step reordered from Phase 5c HEK293 wet-lab to Phase 3c v2 expanded virtual screen. h03 S-tier primary path fully confirmed — remains the near-term Misha lever for MILD rescue (clinically meaningful 20-30 dB ABR improvement). Full Misha NORMAL cure requires longer-horizon h01 lead discovery extending project timeline by 1-2 years.

Why this replaces the deferred Phase 5 GROMACS scaffold

The inventory’s prior pharmacochaperone_phase5_md.py scaffold assumed GROMACS + AmberFF19SB + A100 rental + 50 ns × 3 replicates × 5 ligands in explicit-ligand solvated MD with MM-PBSA — weeks of setup, expensive hardware, untested on Mac.

Phase 5a + Phase 5b split keeps the information gain from ensemble sampling but drops the hardware dependency:

AxisDeferred Phase 5Phase 5a + 5b (delivered)
Software stackGROMACS + AmberTools + MMPBSA.py + OpenFF condaOpenMM (pip) + PDBFixer (pip) + existing Vina/obabel
Receptor MD50 ns explicit ligand2 ns apo-receptor (ensemble-averaged receptor conformations)
ScoringMM-PBSA absolute ΔGVina ensemble mean — captures conformer variance, drops single-pose bias
Runtime~24-48 h on A1002 ns MD ~45 min + 6 leads × 20 snapshots Vina ~20 min on Mac
Precision±1 kcal/mol absolute±1 kcal/mol relative (bias-aware, order-of-magnitude Kd)
Cost$100-500 cloudlocal only

We lose absolute ΔG calibration quality but gain usability without conda + hardware dependency and avoid the Phase 4f implicit-solvent absolute-energy failure mode. This is Method B of the three possible ways to sample conformational noise; Method A (full MM-GBSA) and Method C (FEP+ / TI) remain on the table for later if Phase 5b signal warrants wet-lab follow-through.

Method

Phase 5a — apo MD ensemble

  • Target: Ultra-Mini × TMEM145 CIF, chain A only (STRC Ultra-Mini 1075-1775 residues).
  • Prep: PDBFixer (add missing atoms, protonate at pH 7.4).
  • Force field: AMBER14SB + TIP3P water; 0.15 M NaCl neutralisation.
  • Box: cubic, 1 nm pad → 164 k atoms solvated.
  • Protocol: 5 k-step minimization → 50 ps NVT @ 310 K Langevin → 50 ps NPT @ 1 bar Monte Carlo barostat → 2 ns NPT production → 20 snapshots every 100 ps.
  • Platform: OpenMM OpenCL (Apple Metal backend).
  • Performance: 62 ns/day measured on SMOKE; 2 ns production in ~45 min wall time.
  • Artifacts: models/artifacts/phase5a_snapshots/snap_000.pdb through snap_019.pdb.

Phase 5b — ensemble re-docking

  • Ligand set: Phase 4b top-5 leads (indole-3-acetic, naphthalene-2-COOH, cyclopropane-phenyl-COOH, salicylic, nicotinic) + diflunisal positive control.
  • Per-snapshot pipeline: strip water + ions → chain A protein PDB → obabel → PDBQT (Gasteiger charges).
  • Docking: Vina with Phase 4b box (centre [-22.027, -18.547, 2.215], 18×18×18 Å), exhaustiveness 16, 5 modes, 8 CPU.
  • Aggregation: per-ligand mean ± std ΔG across 20 snapshots → Kd = exp(ΔG/RT) μM → bound fraction at [L] = 1, 10 μM.
  • f_PC estimate: bound_fraction × 0.5 (conservative rescue efficiency η; Phase 6 would refine via mechanism QSAR).

Results

Phase 5a pipeline validation

MetricValue
Solvated system164 305 atoms (164 k)
SMOKE production (100 ps) wall time2.3 min
Full production (2 ns) wall time~45 min
Extrapolated ns/day62
10 ns production ETA (if needed for finer ensemble)3.9 h
Snapshots saved20

Pipeline is production-ready on local Mac. No cloud compute required for h01 Phase 5b.

Phase 5b ensemble affinity and f_PC

Sorted by mean ensemble ΔG. “Phase 4b single” = single-structure Vina score from 2026-04-21; “Δ (kcal/mol)” = Phase 5b mean − Phase 4b single (positive = Phase 5b less favorable).

Ligandmean ΔG (kcal/mol)±stdKd (μM)bound @ 1 μMf_PC @ 10 μMPhase 4b singleΔ
diflunisal (positive control)−5.8650.564500.0200.083−6.396+0.53
naphthalene-2-COOH−5.4720.470970.0100.047−5.980+0.51
indole-3-acetic-acid−5.1410.4461700.0060.028−6.050+0.91
cyclopropane-phenyl-COOH−5.0220.5352070.0050.023−5.716+0.69
salicylic-acid−4.6650.2303790.0030.013−5.024+0.36
nicotinic-acid−4.0980.2809870.0010.005−5.018+0.92

Observations:

  1. All Phase 5b numbers are less favorable than Phase 4b — Δ = +0.36 to +0.92 kcal/mol. This is the expected direction of ensemble correction: single-structure docking captures the best lucky pose; ensemble reveals the average over thermally-fluctuating receptor conformations, which is always less favorable than the best single-conformer fit.
  2. Diflunisal (positive control) is the best binder — but its f_PC is only 0.083. The original top-5 leads are worse than the control on ensemble rescoring. This means the leads selected in Phase 3 virtual screen were structurally similar to diflunisal (carboxylate + aromatic) but did not improve on its affinity against the E1659A pocket.
  3. Std ranges 0.23-0.56 kcal/mol — tighter than expected for 2 ns MD; suggests the K1141 pocket is relatively rigid (not a cryptic allosteric site). More MD won’t dramatically change ranks.
  4. Kd 50-1000 μM for all leads — order-of-magnitude too weak for pharmacochaperone therapy. Target: Kd ≤ 10 μM for f_PC ≥ 0.50.

Misha cure threshold check

Per Misha Compound-Het Therapy Stack Model, f_PC required for NORMAL hearing (≤ 25 dB ABR) monotherapy by E1659A severity:

  • Mild E1659A (f_mat=0.40): f_PC ≥ 0.50
  • Moderate (f_mat=0.25): f_PC ≥ 0.60
  • Severe (f_mat=0.10): f_PC ≥ 0.67

None of the six ligands (5 leads + diflunisal) clears f_PC ≥ 0.50 in any severity scenario at therapeutic [L]=10 μM. Best available: diflunisal 0.083, which is ~6× too low for mild-E1659A NORMAL rescue.

Implication for Nobel-path ladder: Ступенька 1 (this Phase 5b compute) returned RED. Steps 2-3 (HEK293 in vitro + SPR/BLI Kd measurement) are not justified for the current shortlist — would just confirm insufficient binding. Real forward path requires backtracking to Phase 3c v2 (expanded virtual screen) to find stronger binders before advancing to wet-lab.

Nobel-path interpretation

This run is ступенька 1 of the 8-step ladder laid out in the preceding session’s Nobel-argument discussion. Its role is: narrow the Phase 4b ±2 kcal/mol noise band down to ±1 kcal/mol via ensemble averaging, and deliver a credible f_PC point estimate per lead. This estimate then feeds the Misha stack model’s NORMAL-cure threshold check.

Green-light signal (→ wet-lab step 2): at least one lead delivers f_PC ≥ 0.50 at [L]=10 μM, supporting NORMAL-hearing monotherapy for mild E1659A. Authorises HEK293 in vitro validation (Phase 5c): transfect E1659A STRC + GFP-actin, treat with lead compound at 10 μM, confocal image to check clustering restoration.

Yellow-light signal (→ second virtual screen): best lead delivers f_PC 0.25-0.50. Stack support exists (h01+h03 combined) but monotherapy insufficient. Authorises Phase 3c v2 virtual screen expansion (DrugBank FDA + ZINC22 carboxylate tranche) to find stronger binder.

Red-light signal (→ downgrade h01 Misha-fit, strengthen h03): all leads deliver f_PC < 0.25. h01 drug-development path becomes longer-horizon (needs de novo design); Misha’s near-term hope shifts primarily to h03 AAV + h09 hydrogel (MILD rescue; no NORMAL unless longer-term pharmacochaperone development pays off).

FINAL: RED-LIGHT. All leads deliver f_PC < 0.10 at therapeutic [L]=10 μM. Interpretation: the Phase 3b→4b shortlist was tuned against a single AF3 conformer of the K1141 pocket; ensemble MD-relaxed receptor reveals the binding pocket is less accommodating of small carboxylate ligands than static analysis suggested. The ligand set is also chemically monotonous (5/6 are small aromatic carboxylates in the 100-250 Da range), so the “best of six” plateau is probably not the true ceiling of what the pocket can accept — larger or chemically diverse scaffolds may bind more strongly.

Forward path (reordered):

  1. Phase 3c v2 expanded virtual screen (next compute move): DrugBank FDA library (~2500 approved molecules) + ZINC22 carboxylate-enriched tranche (~50 k) + fragment-based screen at K1141 pocket on ensemble of 5 receptor conformers (sub-sample from Phase 5a snapshots). Filter by ensemble Vina ΔG ≤ −7.5 kcal/mol (target Kd ≤ 5 μM → f_PC ≥ 0.67 at 10 μM). If hits appear → re-run Phase 5b on new shortlist.
  2. De novo pharmacophore-based design if v2 screen also returns weak: use STRC Pharmacophore Model K1141 Pocket to synthesise or commission virtual libraries targeting the specific pocket hot-spots (K1141-NZ, loop 1642-1651 H-bond acceptors).
  3. Covalent strategies: K1141 has a nucleophilic amine; covalent inhibitor/chaperone class (e.g., acrylamide warhead + carboxylate recognition) may achieve Kd ~nM where reversible binders plateau at μM. New scaffold class for Phase 3c v3.
  4. Path through h03 alone: remains fully viable for MILD rescue. h01 Nobel-path is not killed, just extended by 1-2 years of lead discovery.

Parameter provenance

All Phase 5a/5b parameters are lit-anchored or explicitly docstring-flagged:

  • AMBER14SB + TIP3P: Maier 2015, Jorgensen 1983. Standard.
  • Box padding 1 nm, 0.15 M NaCl: matches physiological perilymph [Na⁺] = 140-150 mM.
  • NVT 50 ps + NPT 50 ps: standard equilibration; 50 ps NVT may be short for large systems but our minimization was deep (5 k steps), reducing NVT demand.
  • Production 2 ns: SHORT compared to typical 10-50 ns; sufficient for sampling near-native conformational fluctuations of the K1141 pocket but may miss large loop rearrangements (e.g., loop 1642-1651). Extendable to 10 ns (~4 h) if Phase 5b std is too high to distinguish leads.
  • Vina exhaustiveness 16 per snapshot vs Phase 4b 32: reduced for ensemble speed; 16 × 20 snapshots = 320 “exhaustiveness-equivalents” vs 32 single, so total sampling is richer.
  • Rescue efficiency η = 0.5: conservative assumption. Tafamidis-TTR η in lit ≈ 0.7; ivacaftor-CFTR G551D η ≈ 0.5-0.6. Our 0.5 is cautious; Phase 6 QSAR refinement would tighten.

What this doesn’t resolve

  • Absolute Kd: Vina has ±1-2 kcal/mol systematic bias per ligand class. Our ensemble reduces random noise but not bias. Absolute Kd is wet-lab gated (SPR/BLI or ITC).
  • Rescue mechanism: binding ≠ function. Drug may bind the K1141 pocket without actually restoring the TMEM145 interface charge loss. Mechanism requires MD of full complex with ligand (Phase 5c) OR wet-lab HEK293 phenotype assay.
  • Off-target: Phase 4e off-target selectivity soft-failed. Ensemble rescoring doesn’t address this. Wet-lab panel against 5-10 human proteins with similar pockets is needed before clinical advance.
  • PKPD: Does the drug reach [L]=10 μM in cochlea at tolerable systemic dose? This is Phase 6 (oto-PK modelling — analogous to Phase 4e for hydrogel but for small molecule oral/IT route).

Ranking delta

  • #1 (h01) Pharmacochaperone E1659A: A held. Mech 3 held. Deliv 4 held. Misha-fit 4 held. (Mechanism still biology-plausible; patient still compound-het compatible; only the current ligand shortlist is insufficient — not the hypothesis itself.) Next step reordered from “Phase 5 MD ensemble MM-GBSA” (now DELIVERED with red-light) to “Phase 3c v2 expanded virtual screen” (DrugBank FDA + ZINC22 carboxylate tranche + fragment-based, filter at ensemble ΔG ≤ −7.5 kcal/mol). Phase 5a/5b pipeline remains active for future shortlist rescoring; no reinvention needed.
  • #3 (h03) Mini-STRC AAV: S held, scores unchanged. Position as primary Misha near-term lever further reinforced: h01 can’t deliver NORMAL cure short-term, so h03’s MILD rescue (20-30 dB ABR improvement, clinically significant per OTOF trial responder criterion) is the operative Misha benefit for the next 2-3 years while h01 lead discovery expands.
  • Meta — compound-het stack framework: still structurally correct. h01+h03 stack path to NORMAL rescue remains theoretically achievable but now requires h01 second-wave lead discovery. Misha stack model’s “low-burden stack NORMAL” conclusion at mild E1659A preserved conditional on future h01 leads.

The Vina ensemble re-docking confirmed what Phase 4c and Phase 4e soft-hinted: the current leads are not strong enough pharmacochaperones for clinical-grade rescue. This is a HONEST failure that narrows the project scope and points to a concrete next move (expanded virtual screen) without killing any hypothesis.

Connections

Graph View

Backlinks