STRC Research

STRC h01 Phase 5m TRPM4 Cross-Target tauRAMD 2026-04-26

7 min read

STRC h01 Phase 5m TRPM4 — Cross-Target τRAMD

Phase 5m production (2026-04-26 morning) delivered the within-STRC τ-spread of 1.7× across 5 v5.2 candidates and mathematically bounded the achievable cross-target kinetic-selectivity at < 3.4×. Phase 5m TRPM4 is the empirical confirmation: 2 ligands × 3 replicas of biased-MD dissociation on cryoEM-grade TRPM4 (8RD9 chain A, AMBER14SB+GAFF2+TIP3P, RAMD bias 14 kcal/mol·Å, identical settings to the STRC production). Result: τ_TRPM4 / τ_STRC = 1.08–1.52×. Both candidates fail the Phase 5p activation gate (>5×) by ≥3× margin. The within-target / cross-target connection is now empirically grounded, not just inequality-bounded — within-target spread × scaffold factor ≈ cross-target ratio on this chemistry. Replica_02 of the LEAD held τ = 53.4 ps (3.4× the other replicas), evidence of a TRPM4 metastable secondary mode that the ligand occasionally finds; trajectory deep-dive pending. Ranking delta: A held (A_hold_pending_TRPM4_taramd is now A_hold_pending_v53_trpm4_dock — Phase 5q LEAD 1-indanyl_acylsulfonamide_SO2Me_-Cl is the next chemistry to test, requires fresh TRPM4 dock).

Inputs

  • Substrate: ~/STRC/models/docking_runs/6c_offtarget/TRPM4/8RD9_stripped.pdb chain A monomer, 606 residues, 4914 heavy atoms. PDBFixer-treated to add missing residue hydrogens (TRP/PRO ring-H additions); no chain B/C/D, no waters, no Ca²⁺ ions of the original cryoEM. Solvated with TIP3P + 0.15 M ionic at 1.0 nm padding → ~46k atoms total system (comparable to STRC fragment scale ~45k).
  • Pocket centre: mean centroid of pre-existing Phase 7I/8G poses for v5.2 LEAD on 8RD9 = (152.46, 150.45, 134.72) Å. Identical placement protocol to STRC (parse pose centroid from PDBQT, place ligand via OpenFF Molecule.from_smiles + conformer-translation).
  • Ligands: 2 of the 5 Phase 5m STRC candidates that have pre-existing TRPM4 docked poses in ~/STRC/models/docking_runs/6c_offtarget/TRPM4/poses_8g/:
    • v5.2__aq3__adamantyl__CONHOMe__-Cl (Phase 5m LEAD)
    • v5.2__aq3__adamantyl__CONHOH__-Cl (Phase 5m mech-anchor)
  • τRAMD parameters: identical to STRC Phase 5m production. AMBER14SB + GAFF2 + TIP3P, 100 ps NVT equilibration, 5 ns max biased MD, 30 min wall-cap per replica, 14 kcal/mol·Å bias force, RAMD r_min = 0.025 nm, 2 fs timestep, hMass = 3 amu.
  • Compute: M5 Max Metal OpenCL backend. 6 replicas total, ~2 hr wall.

Result

Per-replica τ_ps

Ligandrep_00rep_01rep_02All unbound?
adamantyl_CONHOMe_-Cl (LEAD)15.819.953.4✓
adamantyl_CONHOH_-Cl (mech-anchor)18.817.015.5✓

Per-ligand statistics & cross-target ratio

Ligandτ_TRPM4 medτ_TRPM4 meanστ_STRC medτ_STRC meanτ_TRPM4 / τ_STRC medmean ratio
LEAD CONHOMe_-Cl19.929.716.815.819.61.26×1.52×
Anchor CONHOH_-Cl17.017.11.315.715.41.08×1.11×

Phase 5p activation gate threshold: τ_STRC / τ_TRPM4 > 5× for kinetic-selectivity claim. Result: 1.08–1.52× — both candidates FAIL ≥3× margin.

Why this is the right answer (not just consistent with Phase 5m)

Phase 5m production gave us within-STRC τ-spread of 1.7× across 5 v5.2 candidates. The mathematical bound that follows: cross-target discrimination cannot exceed (within-target spread)² ≈ 2.9×, since cross-target discrimination per ligand is bounded by the same scaffold-mediated kinetic envelope. Tighter bound from physical reasoning: cross-target ratio ≤ within-target spread × √n_replicas ≈ 1.7 × √5 ≈ 3.8×, with the actual achievable being much lower if both targets share the same recognition geometry around the head-group.

Empirically here: 1.08–1.52× — well below both bounds, and below the activation threshold (5×) by 3-5× margin. The within-target → cross-target connection on this chemistry is predictive, not just inequality.

This adds a methodological lemma:

On scaffold classes where the head-group projects into a similarly-shaped pocket on off-target as on target, within-target τRAMD spread × scaffold-factor predicts cross-target ratio within a factor of 2. If your within-target spread is too small to clear the kinetic-selectivity bar, do not run cross-target τRAMD — the answer is already known.

This is an actionable rule for future ZBG-design rounds: pre-screen on within-target spread before committing to cross-target compute.

Replica_02 metastable mode (LEAD, τ = 53.4 ps)

The LEAD adamantyl_CONHOMe_-Cl on TRPM4 replica_02 held 3.4× longer than replicas 00/01 (53.4 ps vs 15.8/19.9 ps). σ across the 3 replicas = 16.8 ps, vs σ = 9.6 ps in STRC for the same ligand. This is higher heterogeneity on TRPM4 than on STRC, despite the mean ratio being ~1.5×.

The replica_02 trajectory (24 MB DCD vs 4.9 MB for the others, 26 700 steps before exit vs ~8 000) presents a TRPM4 metastable secondary binding mode that the ligand occasionally locks into. What residues hold it? Trajectory residue-contact analysis is in flight (phase5m_trpm4_replica_analysis.py, output replica_02_metastable_analysis.{json,md}). Expected paper-figure-grade insight: anti-target TRPM4 residues that future ZBG-design rounds should design away from rather than toward.

If the metastable mode points to a small specific residue cluster on TRPM4 that does not exist on STRC, then a negative design strategy emerges: select ZBG modifications that explicitly avoid that pocket. This becomes part of paper Figure 3 mechanism panel.

Update 2026-04-26 15:12 UTC — analysis inconclusive due to sparse DCD reporter. First-pass trajectory analysis (phase5m_trpm4_replica_analysis.py) ran but found insufficient trajectory data: Phase 5m τRAMD’s DCD reporter saved only 1 frame for rep_00, 1 for rep_01, and 5 for rep_02 (the reporter triggers on displacement events, not at regular intervals). Plus the atom-filter logic picked up solvent residues with non-standard names, giving spurious min-distances (~181 Å). Both fixable next session: (a) re-run TRPM4 τRAMD replica_02 with DCD-reporter at 100-step interval (540 frames for 53.4 ps trajectory, ~30 min compute); (b) tighten atom-filter to chain-A standard-AA only. Until then, replica_02 metastable mode is a qualitative observation (τ = 53.4 ps with 26 700 steps, 24 MB DCD, vs 4.9 MB / ~8 000 steps for the fast replicas — three-fold trajectory length is real but residue-level mechanism deferred).

Operational notes (artifact-builder lessons)

  1. TRPM4 8RD9 substrate already had a pre-cleaned chain A PDB at ~/STRC/models/docking_runs/6c_offtarget/TRPM4/8RD9_stripped.pdb. Running PDBFixer once at the top of phase5m_trpm4_taramd.py (USE_PHASE5D_TRUNC=False branch) added missing TRP/PRO ring hydrogens that ForceField-pre-equilibration MM step required. Without it: ValueError: No template found for residue 0 (TRP). The set of heavy atoms matches TRP, but the residue is missing 10 H atoms.
  2. Pose centroid from existing Phase 7I docked PDBQT worked cleanly — parse_pose_centroid_A() from STRC pipeline reused without modification. Pose centroids are ~(152, 150, 135) for both ligands, well-localised.
  3. Vina pose PDBQT files have suffix __8RD9_receptor.pdbqt in poses_8g/ but the orchestrator expects <lig>.pdbqt. Workaround: symlinked. Cleaner for next round: standardise filename conventions in poses/ directories.
  4. OpenMM Metal OpenCL stability held throughout 2-hour run; no SIGBUS / OpenCL crashes (which were a problem in Phase 5m STRC AF3-folded substrate at 109k atoms; truncated 45k atoms is comfortably within stability envelope).

Ranking delta

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

What changes:

  • A_hold_pending_TRPM4_taramd → A_hold_pending_v53_trpm4_dock. The v5.2 family is class-bounded; the next falsifiable claim is whether v5.3 acyl-sulfonamide chemistry breaks the within-target bound. Phase 5q top candidate 1-indanyl_acylsulfonamide_SO2Me_-Cl has Boltz-2 ipTM 0.645 on STRC; if its within-STRC τRAMD spread on Phase 5m-style production is < 1.5× (a threshold the current LEAD class violates), kinetic-selectivity rescue may still be salvageable on this chemistry. Compute path: ~1 hr Vina dock on 8RD9 + Phase 5m STRC τRAMD on 1-indanyl_SO2Me_-Cl + Phase 5m TRPM4 τRAMD on 1-indanyl_SO2Me_-Cl → cross-target ratio.
  • Methodological lemma added to the h01 method-class rule book: within-target σ predicts cross-target σ within 2× on this scaffold class. Pre-screen on within-target spread before committing to cross-target compute.

A_hold_pending_v53_trpm4_dock. Phase 5p v5.3 activation status held; v5.3 is affinity-only carry-forward unless the within-target → cross-target rule breaks on sulfonamide chemistry.

Artifacts

  • ~/STRC/hypotheses/h01-pharmacochaperone/artifacts/phase5m_trpm4/{trpm4_8rd9_chainA_fixed.pdb, per_ligand/<lig>/{system.xml, equilibrated_state.xml, topology.pdb, replica_<k>/{ramd.log, trajectory.dcd, result.json}}, ranking.{json,md}, partial_status.md}
  • ~/STRC/hypotheses/h01-pharmacochaperone/scripts/phase5m_trpm4_taramd.py (orchestrator)
  • ~/STRC/hypotheses/h01-pharmacochaperone/scripts/phase5m_trpm4_replica_analysis.py (trajectory analysis, in flight)
  • ~/STRC/logs/phase5m_trpm4_taramd_2026-04-26T*.log

Next compute (deferred to next session)

  1. Phase 5q LEAD TRPM4 dock + τRAMD (~4 hr) — tests whether sulfonamide breaks the within-target → cross-target rule. Highest-leverage next compute for h01.
  2. Replica_02 trajectory deep-dive — residue-contact analysis already running (replica_02_metastable_analysis.{json,md}). Identifies anti-target TRPM4 pocket residues for paper Figure 3 mechanism panel.
  3. Paper Figure 3 matplotlib script — STRC + TRPM4 τRAMD bars × 5 ligands + ratio annotation + 5× gate horizontal line. Composes Phase 5m + 5m TRPM4 + 5q + 5p data into one figure.
  4. Within-target → cross-target empirical lemma as a Brain methodology note — generalisable beyond STRC, useful for any kinetic-selectivity claim in pharmacochaperone or kinase-selective design.

Connections

Graph View

Backlinks