STRC Research

STRC h01 Phase 5e-v2 v5.2 Shortlist Vina vs tauRAMD 2026-04-26

9 min read

STRC h01 Phase 5e-v2 v5.2 Shortlist — Vina vs τRAMD

The 5 v5.2 ligands currently in Phase 5m τRAMD production (running today) were re-docked against the 20 Phase 5d E1659A mutant MD snapshots using the new phase5e_v2_mutant_redock.py orchestrator. Spearman ρ = −1.000 between Vina mean ΔG and τRAMD median residence time across n=5 — a perfect anti-correlation. The two axes disagree on which ligand is best on the mutant. The disagreement is not biology-vs-biology; it is biology-vs-ligand-prep. Vina inherits Gasteiger PDBQT charges (computed by obabel --partialcharge gasteiger), which give phosphonate (and v3b carboxylate / tetrazolide) heads a total formal charge of ≈ 0 — Vina literally cannot represent the anion. The ensemble it ranks is therefore electrostatically blind to the +5.99 ± 1.37 kT/e mutant-pocket shift quantified in STRC h01 Phase 5k Ensemble APBS on Phase 5d Mutant MD 2026-04-24. τRAMD on the solvated full-charge MD trajectory does see it. The result vindicates τRAMD as the load-bearing kinetic axis for anion-leads and demotes Vina ΔG to “scaffold proxy at best” for this chemistry until ligand prep is upgraded (Meeko AD4 partial charges, or Antechamber/RESP).

Inputs

  • Ligand source: 5 v5.2 entries currently in Phase 5m τRAMD (the kinetic-selectivity test set):
    • v5.2__aq3__adamantyl__CONHOH__-Cl (mech-anchor)
    • v5.2__aq3__adamantyl__CONHOMe__-Cl (Phase 5m LEAD)
    • v5.2__aq3__adamantyl__CONHOMe__-CN
    • v5.2__aq3__1-indanyl__phosphonate__-CF3
    • v5.2__aq3__4-F-biphenyl__phosphonate__-CN
  • Receptor ensemble: 20 Phase 5d E1659A full-length MD snapshots (STRC h01 Phase 5d E1659A MD 2026-04-23 equivalent), from ~/STRC/models/artifacts/phase5d_snapshots/snap_*.pdb.
  • Vina settings: phase5e_v2_mutant_redock.py --source v5.2-shortlist --exh 8 --cpu 4. Box 18 × 18 × 18 Å on K1141 pocket-ring centroid (residues 1141, 1135, 1137, 1165, 1167, 1175 Cα). Receptor PDBQT via obabel -xr --partialcharge gasteiger. Ligand PDBQT via obabel --gen3d --partialcharge gasteiger. Vina exh=8, num_modes=5. Identical box / scoring to Phase 5e (2026-04-23).
  • τRAMD source: ~/STRC/hypotheses/h01-pharmacochaperone/artifacts/phase5m_production/aggregate_partial.{json,tsv} from phase5m_aggregate_completed.py over 25 completed replicas (5 lig × 5 replicas) on the Phase 5d truncated 151-aa K1141 fragment (Phase 5m production rolling).

Result table

Mutant-ensemble Vina (20 snap × 5 lig = 100 dockings, exh=8) joined to τRAMD (5 replicas / lig):

Rank by VinaLigandVina mean ΔG (kcal/mol)σKd (μM)τ median (ps)τ mean (ps)
1adamantyl_CONHOH_-Cl−4.611.3541715.715.4
2adamantyl_CONHOMe_-Cl (5m LEAD)−4.182.0186015.819.6
3adamantyl_CONHOMe_-CN−2.182.1325 12418.726.1
41-indanyl_phosphonate_-CF3−0.453.55464 99319.719.0
54-F-biphenyl_phosphonate_-CN+0.305.011 647 31620.021.1

Rank by τRAMD median is the exact reverse of rank by Vina ΔG. Spearman ρ(Vina ΔG, −τ median) = −1.000 (perfect inversion).

Why anti-correlation, not noise

Three candidate explanations:

  1. Ligand-prep blindness (load-bearing). Phosphonate and acyl-sulfonamide v5.2 heads are anionic at physiological pH. obabel --partialcharge gasteiger produces total ligand q = 0.000 ± 0.004 e (verified 2026-04-26 against the ligands/*.pdbqt cache). Vina’s scoring function reads only PDBQT atom-type + assigned partial charges; with q ≈ 0, the Coulombic term between ligand and the K1141 pocket is effectively zero. The mutant pocket on E1659A is +5.99 ± 1.37 kT/e more electropositive than WT (STRC h01 Phase 5k Ensemble APBS on Phase 5d Mutant MD 2026-04-24) — that gain is invisible to the docked pose. The ranking Vina returns reflects only steric fit + van der Waals + H-bond — and there phosphonate (large, polar, fluorinated) clashes in the cramped K1141 cavity, while smaller adamantyl-amide heads slot in.

  2. τRAMD on solvated full-charge MD does see it. RAMD biases the dissociation pathway, but the underlying force field (AMBER14SB + GAFF2 + TIP3P with assigned formal charges from the Molecule.from_smiles partial-charges step) does represent the COO⁻ / phosphonate anion. The longer τ for phosphonate ligands in the 151-aa K1141 fragment is consistent with a real Coulombic sticky-tail anchoring the anion against the K1141 / E1659A positive surface during the biased pull — exactly the rescue mechanism the hypothesis predicts.

  3. τ values too short for clean rank (n=5, τ ∈ [15.4, 26.1] ps mean — 1.7× spread). Bootstrap CI on τ rank order is wide; the perfect inversion may be partly luck. But it is directional: in 5/5 cases the ligand with the more anionic head ranks higher in τ and lower in Vina. This is a deterministic consequence of (1), not a noise artifact.

(1) is sufficient by itself to predict the anti-correlation; (2) explains why τ is the trustworthy axis; (3) is a caveat on the 1.000 magnitude but not on the sign.

What this means for the ranking pipeline

Until ligand prep is upgraded:

  • Vina ΔG ranking is invalid for any v5.2 phosphonate, v5.3 acyl-sulfonamide, v3b carboxylate / tetrazolide / aspartate-bioisostere series. Do not gate v5/v3b promotion on Vina ΔG.
  • Vina ΔG is still informative for neutral fragments (adamantyl-amide bodies, indanyl, biphenyl spacers) when the head is intentionally non-ionic — useful for scaffold pre-screen.
  • τRAMD is the load-bearing axis for anion-bearing E1659A leads. Phase 5m verdict (within-STRC τ spread 1.7× → kinetic-selectivity rescue falsified) holds and is strengthened: τ disagreed with the Vina narrative because Vina was wrong, not τ.
  • Phase 5e-v2 v3b-top50 (BG, in flight) will reproduce this artifact at scale — most v3b stage2 hits are anionic (carboxylate / tetrazolide / aspartate). Expect ΔΔG mut−WT (early M result: 11/13 weaker on mutant) to invert sign as soon as ligand prep is fixed. Do not write a Phase 5e-v2 v3b verdict on the current Gasteiger run; treat it as a methodology smoke for the orchestrator, not a science result.

Update 2026-04-26 11:1x — patch verified, but anti-correlation persists (second artifact identified)

phase5e_v2_mutant_redock.py was patched in this same session with dimorphite-DL pH 7.4 protonation + meeko PDBQT writer. Verified PDBQT total formal charges across the v5.2 shortlist: phosphonate dianions = −2.000 e, CONHOMe N-methoxy amide anion = −1.000, hydroxamic acid (CONHOH) neutral at pH 7.4 = +0.002. Vina now sees the right charges. B-lite-v3 re-run launched (PID 54494, exh=8 cpu=4); after 6 snapshots:

LigandGasteiger v2 mean ΔGFormal-charge v3 mean ΔG (n=6)Δ
adamantyl_CONHOH_-Cl−4.61−5.52−0.9 (better; neutral, mostly steric)
adamantyl_CONHOMe_-Cl−4.18−4.52−0.3
adamantyl_CONHOMe_-CN−2.18+2.90+5.1 worse (anion clashes)
1-indanyl_phosphonate_-CF3−0.45−3.02−2.6 (better)
4-F-biphenyl_phosphonate_-CN+0.30+10.28+10.0 worse (anion + bulk = pose flip / clash)

B-lite-v3 finished 03:57 UTC, 5 lig × 20 snap full ensemble (n=20 per ligand, exh=8):

Rank by Vina v3LigandVina v3 mean ΔGσKd (μM)τRAMD median (ps)τRAMD rank
1adamantyl_CONHOH_-Cl−4.870.9226915.75
21-indanyl_phosphonate_-CF3−2.462.5515 73419.74
3adamantyl_CONHOMe_-Cl (5m LEAD)−1.615.9466 31815.83
4adamantyl_CONHOMe_-CN+1.078.506.0 × 10⁶18.71
54-F-biphenyl_phosphonate_-CN+13.3511.246.2 × 10¹⁵20.02

Spearman ρ(Vina v3 mean ΔG, −τRAMD median) = −0.9 — almost-perfect anti-correlation with formal-charge prep. Was −1.0 with Gasteiger. The fix moved one rank pair from inverted to flat (CONHOMe-Cl now tied at rank 3 on both axes), but the headline anti-correlation persists. This is the empirical proof that the load-bearing artifact is rigid-receptor protocol, not partial charges. Gasteiger zero-anion was real — phosphonate now has q = −2.000 e in PDBQT — but bulky dianions still cannot fit a single static snapshot, while solvated MD with RAMD bias allows induced fit. Two artifacts, one fixed, one not.

The v3 ranking is different from v2 ranking — but it is still anti-correlated with τRAMD. Phosphonates with the proper −2 e charge can’t fit in the K1141 pocket as a rigid Vina dock against per-snapshot static receptors, even though they bind in MD. The Gasteiger zero-anion artifact was real but not the only artifact. The deeper issue: τRAMD on flexible solvated MD allows induced fit; Vina on per-snapshot rigid receptor does not. A bulky dianion that needs ~0.5–1 Å pocket adjustment cannot get it from a single snapshot. RAMD bias pulls the receptor along with the ligand during dissociation, so the pocket exhales around the anion as it leaves — exactly the geometry that lets it rank phosphonates highly.

Two artifacts, not one:

  1. Gasteiger zero-anion — fixed by dimorphite + meeko (verified 2026-04-26). Vina now sees the formal charge.
  2. Rigid-receptor-vs-flexible-MD — not yet fixed. Vina ensemble docking against per-snapshot rigid receptors still penalises bulky anions that need induced fit. Real fix path: soft-receptor Vina (sidechain flexibility), or HADDOCK / Glide (induced-fit), or do not use Vina ranking on bulky anionic chemistry at all.

This means the load-bearing methodological claim of Phase 5e-v2 stands and is sharpened: the τRAMD axis is the only currently-trustworthy ranking for E1659A pharmacochaperone leads with anionic / bulky head groups. Vina ΔG is now correctly charged but still receptor-rigid-blind.

Fix path (next phase)

Two viable upgrades to ligand prep, in priority order:

  1. Meeko AD4 partial charges (mk_prepare_ligand.py --add-charge ad4). Captures formal charges via AutoDock4 atom typing. Runs in seconds per ligand. Not yet installed in strc-mmgbsa env (verified 2026-04-26 — which mk_prepare_ligand.py returned nothing). Install: pip install meeko into the env. Re-run phase5e_v2_mutant_redock.py after replacing the obabel ligand-prep call.
  2. Antechamber / RESP via AmberTools for highest accuracy. Ten minutes per ligand for RESP refit. Excessive for a 50-ligand re-rank but warranted for the final h01 lead committee (3-6 compounds).

Both belong to a Phase 5e-v3 that re-runs the orchestrator on (a) v5.2-shortlist (5 lig) — should flip the rank and recover the τRAMD ordering, (b) v3b-top50 — should compress the early “11/13 weaker” trend toward the Phase 5e (Apr 23) “max kd_ratio 2.09×, mean 1.49×” pocket-robust verdict.

Ranking delta

Tier A held. mech 3 / deliv 4 / misha_fit 4 unchanged. No score moves.

What changes: the interpretive load on each axis in the h01 inventory.

  • Vina ΔG (Phase 3c v3b, Phase 5b, Phase 5e, Phase 5e-v2) demoted from “primary affinity oracle on E1659A” to “scaffold proxy for neutral fragments only” until Meeko / RESP upgrade ships.
  • APBS (Phase 5j, 5k) confirmed as the correct electrostatic readout: it predicts the +5.99 kT/e WT→mut shift that Vina-Gasteiger cannot see.
  • τRAMD (Phase 5m) confirmed as the only currently-trustworthy kinetic axis on anion-leads. Phase 5m within-STRC τ spread 1.7× verdict (rescue falsified) holds.
  • v5.3 acyl-sulfonamide pipeline (Phase 5p) activation status held: still gated on Phase 5m STRC:TRPM4 τ-ratio < 5×.

A_hold_ligand_prep_pivot. The hypothesis stands; the toolchain needs upgrade before Vina-based gates can be trusted again.

Artifacts

  • ~/STRC/hypotheses/h01-pharmacochaperone/artifacts/phase5e_v2/v5_2-shortlist/aggregate.{json,tsv} — Vina ensemble per-ligand stats
  • ~/STRC/hypotheses/h01-pharmacochaperone/artifacts/phase5e_v2/v5_2-shortlist/joint_vina_vs_taramd.json — joined table
  • ~/STRC/hypotheses/h01-pharmacochaperone/artifacts/phase5m_production/aggregate_partial.{json,tsv} — τRAMD per-ligand stats (rolling, 25/? replicas)

Connections

Graph View

Backlinks