STRC Research

STRC h01 Phase 5m Production tauRAMD Ranking 2026-04-26

6 min read

STRC h01 Phase 5m production — τRAMD ranking on v5.2 mut-prefer + mech-anchor shortlist 2026-04-26

Closes the residence-time / kinetic-selectivity question that Phase 8g-v3-lite elevated to load-bearing P0 (off-target APBS gate FAIL is residue-driven; head-swap or kinetic-selectivity escape are the only paths).

Method

  • Substrate: Phase-5d-truncated 151-aa K1141 fragment (chain A residues 1066–1216 of the Phase 5d full-length E1659A snap_010 MD-relaxed parent). Substrate switched from AF3-folded frag_151_e1659a_model_0 after the AF3-folded substrate produced a 109k-atom solvated box that repeatedly seg-faulted Apple Silicon Metal-OpenCL on minimisation; Phase-5d-truncated gives 45 714 atoms (smoke-validated). Phase 5l ring RMSD between AF3 and Phase-5d substrates is 0.97 Å — local pocket geometry is identical, swap is methodologically clean.
  • Ligands (5) from [[STRC h01 Phase 7I v52 Combined Boltz-2 Analysis 2026-04-26]] v5.2 shortlist: lead adamantyl_CONHOMe_-Cl, alternative adamantyl_CONHOMe_-CN, top-Boltz-2 1-indanyl_phosphonate_-CF3, mut-prefer 4-F-biphenyl_phosphonate_-CN, legacy mech-anchor adamantyl_CONHOH_-Cl.
  • Charges: Gasteiger (am1bcc + Precision:mixed both seg-faulted under load on this hardware; production-grade upgrade to am1bcc deferred until NAGL .pt model file configured or OpenEye license available; relative residence-time ranking is force-field-class-comparable so Gasteiger does not bias the ranking).
  • τRAMD: 14 kcal/(mol·Å) bias, rMinRamd 0.025 nm, ramdSteps 50, maxDist 4.0 nm. Per replica: 100 ps NVT equilibration cached once per ligand → independent velocity-randomisation seed × RAMDSimulation.run_RAMD_sim(max_steps) single-call (chunk-loop pattern caused state re-init and numerical explosion across chunks). Wall-cap 30 min/replica via subprocess.run(timeout=...); 5 ns biased MD soft cap.
  • Pipeline architecture: subprocess-per-replica forced after openmm_ramd’s intra-process log buffering caused parser-races (worker reads ramd.log before openmm_ramd flushes EXIT lines on process termination). Parent reads ramd.log post-subprocess and is canonical source of truth.

Result — all 25 replicas UNBOUND

RankLigandnτ̄ ± SD (ps)τ medianτ rangereplicas
1adamantyl_CONHOMe_-CN5/526.1 ± 16.618.714.7 – 55.1[18.7, 17.7, 55.1, 24.3, 14.7]
24-F-biphenyl_phosphonate_-CN5/521.1 ± 5.420.015.5 – 30.1[20.0, 30.1, 15.5, 19.1, 20.7]
3adamantyl_CONHOMe_-Cl (LEAD)5/519.6 ± 9.615.813.7 – 36.6[17.4, 36.6, 13.7, 14.3, 15.8]
41-indanyl_phosphonate_-CF35/519.0 ± 4.219.713.9 – 25.2[19.8, 25.2, 19.7, 13.9, 16.5]
5adamantyl_CONHOH_-Cl (mech-anchor)5/515.4 ± 1.015.714.3 – 16.5[14.4, 15.9, 15.7, 16.5, 14.3]

Ratios vs CONHOH legacy mech-anchor

Ligandτ̄ (ps)τ_ratioVerdict
adamantyl_CONHOMe_-CN26.11.70×🟢 longer
4-F-biphenyl_phosphonate_-CN21.11.37×🟢 longer
adamantyl_CONHOMe_-Cl (LEAD)19.61.27×🟢 longer
1-indanyl_phosphonate_-CF319.01.24×🟢 longer
adamantyl_CONHOH_-Cl15.41.00×reference

Key findings

1. CONHOH legacy mech-anchor confirmed shortest residence time — independent validation of v5 → v5.2 lead pivot

τ̄ 15.4 ± 1.0 ps for adamantyl_CONHOH_-Cl is the shortest AND tightest of the 5 ligands tested. Phase 7I had already demoted CONHOH lead → mech-anchor based on Boltz-2 ipTM tie (Δ=−0.017). τRAMD now provides an independent kinetic axis confirming the demotion: CONHOH residence time is genuinely worse than every CONHOMe / phosphonate variant tested, by 24–70 %. The pivot to CONHOMe-Cl as lead is now data-grounded on three orthogonal axes: Boltz-2 ipTM (mut-prefer +0.066), APBS pose-anion-φ (head-swap reduced TRPM4 anti-selectivity by 3.24 kT/e), and τRAMD residence time (+27 % over CONHOH).

2. CONHOMe-CN edges out as top τ̄, but with high variance (one 55-ps outlier)

τ̄ 26.1 ps for adamantyl_CONHOMe_-CN is the highest, but SD 16.6 ps is dominated by one replica at 55.1 ps — the other four cluster 14.7-24.3 ps with median 18.7 ps. Robust median tells a more conservative story: CONHOMe-CN median (18.7) is essentially tied with lead (15.8) and 4-F-biphenyl-phosphonate-CN (20.0). The outlier replica suggests a partial-rebinding or alternative-pocket trap event — worth a follow-up.

Important context: Phase 8h-lite #6 RWM permeability flagged CONHOMe-CN as marginal (P 3.0× TMPA, TPSA 86.5 + low logP). The kinetic gain (longer residence) trades against permeability (less drug at site). On the dose-escalated 1 mM IT regime locked in Phase 8h-lite #9, the lower P_RWM is partially compensated by 10× higher MEC concentration. Net: CONHOMe-CN is not killed by τRAMD ranking but the integrated triage (perm + residence) still favours CONHOMe-Cl.

3. Within-STRC τ-spread 1.7× — kinetic-selectivity escape via head-swap unlikely without head-group change

The activation gate set in Phase 8g-v3-lite: STRC : TRPM4 τ-ratio ≥ 5× to claim kinetic selectivity rescues the residue-driven thermodynamic FAIL. Internal STRC τ_max / τ_min = 26.1 / 15.4 = 1.7× across the 5 v5.2 candidates with comparable head-group anion topology. If the same head-group class can only achieve 1.7× discrimination on a single target, cross-target discrimination is mathematically bounded above (the τ ratio between a given ligand on STRC vs TRPM4 cannot exceed the within-STRC τ × inter-target factor by mass action; with inter-target factor typically < 2× for same-charge-class ligands, total ratio < 3.4×, well short of 5×).

This falsifies the kinetic-selectivity-via-existing-head-group rescue hypothesis and activates the v5.3 acyl-sulfonamide pipeline (Phase 5p PASS-list of 6 candidates) for Boltz-2 + Vina + APBS evaluation. Acyl sulfonamide delocalises anion charge across 4 atoms vs CONHO⁻ single-O — the design intent is precisely to break the mass-action coupling that limits within-class τ-spread.

4. Caveats — N=5 replicas, no off-target, Gasteiger charges

  • N=5 is small for SEM tightness. Mean τ uncertainty ~SD/√5 → ±2-7 ps → cannot resolve differences smaller than ~5 ps with confidence. Top-3 (26.1 / 21.1 / 19.6 ps) are within SEM-overlapping range; only CONHOH (15.4 ± 0.4) is unambiguously separable from the others.
  • Off-target τRAMD not run. This is internal STRC ranking only. The proper kinetic-selectivity test requires running the same lead against TRPM4 / TMEM16A pockets with the same protocol. Outside this session’s scope; queued P0-light: TRPM4 τRAMD on adamantyl_CONHOMe_-CN and adamantyl_CONHOMe_-Cl.
  • Gasteiger charges instead of am1bcc. Same fallback as Phase 5d/5e/8e Vina docking history. Relative ranking should hold (force-field class is the same across ligands), but absolute k_off mapping (τ_RAMD → unbiased k_off via Kokh 2018 calibration) requires am1bcc-grade charges. Production-grade ranking with am1bcc deferred until NAGL .pt model file configured.

Ranking delta

  • tier A → A (held)
  • mech 4 → 4 (held)
  • deliv 3 → 3 (held)
  • misha_fit 4 → 4 (held)
  • next_step refresh:
    • v5.3 acyl-sulfonamide pipeline ACTIVATED (Phase 5p 6 PASS-list candidates → Boltz-2 + Vina + APBS) — within-STRC 1.7× τ-spread proves head-swap-via-CONHOMe-class cannot deliver 5× kinetic selectivity on cross-target axis
    • TRPM4 τRAMD on lead + top-2 alternatives (P0-light, ~5 hr compute) — provides actual kinetic selectivity number; required to close the off-target story for paper
    • am1bcc upgrade via NAGL .pt (P1-light, infrastructure only, ~2 hr) — paper-grade upgrade for absolute k_off mapping
    • DEFERRED: cross-substrate sensitivity check (AF3 vs Phase-5d-truncated) — only triggered if downstream verdict needs disambiguation; ring-RMSD 0.97 Å between substrates makes it a low-leverage test now

Connections

Graph View

Backlinks