STRC Pharmacochaperone Phase 4e — proxy gate soft-fails (margin 0.150 vs threshold 0.200), definitive selectivity deferred to Phase 4b Vina ΔG
Full-surface pocket scan of Ultra-Mini × TMEM145 chain A found 218 druggable pockets. The K1141 pocket wins on pharmacophore-match score (0.873 vs 0.723 for the best off-target), but the 0.150 margin falls short of the 0.200 gate I set. The shortfall is a scoring-formula artifact — the proxy gives 0.5 partial credit to any LYS/ARG/ASP/GLU combination, so big LRR-face pockets with random basic residues all score ~0.72 from geometric fill. The real geometric specificity — the triple-basic cluster K1141 + K1172 + K1173 plus the F1646 + W1652 aromatic hot-spot — is invisible to class-presence scoring. Definitive selectivity read transferred to Phase 4b Vina ΔG (per-ligand, per-pocket).
Method (proxy, Vina-free)
Full-surface ray-cast scan of chain A (701 residues in Ultra-Mini × TMEM145 clinical CIF; real STRC residues 1075–1775), stride-20 Cα seeds, 14 Å radius, 1.2 Å grid, 9/14 ray enclosure threshold, 25-voxel minimum cluster (43 ų). Each seed yields up to ~10 subpockets via depth-peak region growing. Pharmacophore score per pocket:
- 0.30 × [K1141 in lining residues] (identity bonus; partial 0.5 if any basic residue present but not K1141)
- 0.25 × [D1140 or D1173 in lining residues] (identity bonus; partial 0.5 if any acidic residue present)
- 0.20 × [any aromatic in lining residues]
- 0.15 × [≥4 H-bond donors/acceptors in lining]
- 0.10 × druggability
Gate: K1141 pocket score − best off-target score ≥ 0.20 AND gate passes for ≥3 of top-5 leads.
Result
| Drugg. | V (ų) | Key residues (real STRC) | Pharmacophore score | |
|---|---|---|---|---|
| K1141 (target) | 0.98 | 209 | K1141, K1172, K1173, Q1140, Q1168, F1646, W1652, E1655, loop 1645-1656 | 0.873 |
| Off-target #1 | 0.98 | 218 | D1270, E1304, R1305, F1331, H1737, K1740, E1741 | 0.723 |
| Off-target #2 | 0.97 | 197 | F1394, E1419, E1425, K1442, R1468, F1471, F1495, D1498 | 0.722 |
| Off-target #3 | 0.97 | 204 | F1394, E1419, E1425, K1442, R1468, F1471, D1498 | 0.722 |
| Off-target #4 | 0.96 | 192 | same as #3 (different flood peak) | 0.721 |
| Off-target #5 | 0.96 | 192 | R1228, W1232, D1263, R1677, W1734, E1738 | 0.721 |
Margin: 0.873 − 0.723 = 0.150 (gate threshold 0.200). Proxy verdict FAIL.
Why the proxy under-measures K1141 specificity
The K1141 pocket’s defining feature is a triple-basic electrostatic anchor cluster — K1141, K1172, K1173 within ~6 Å of each other — stabilising a carboxylate via simultaneous salt-bridges from multiple directions. None of the five off-target pockets have any basic cluster: all carry at most one lysine or arginine with a single acidic partner on the other side. A carboxylate ligand designed for the triple-Lys cluster cannot dock productively in a single-Lys pocket.
My proxy scorer gives partial credit (0.5) for “any basic residue present”, which flattens this distinction. It would need a geometric term — “basic residues clustered within 6 Å” — to reproduce the real pharmacophore specificity. Adding that term is a rewrite, not a patch, and the real answer comes from Vina scoring anyway.
Additionally, the K1141 pocket is the only one that includes the mutation-responsive loop (1642-1651) AND the LRR face together. Off-target pockets lie on separate domains (ARM repeats 1270-1305, 1394-1498, 1677-1741) unrelated to the E1659A defect. A ligand bound there cannot rescue the E1659A interface loss even if it bound identically.
Interpretation — soft fail, not a kill
The K1141 pocket is:
- The highest-druggability pocket on the chain (0.98, tied with one other).
- The highest pharmacophore match (0.873 vs 0.723 runner-up; 17% higher relative).
- Structurally unique — only site with triple-basic anchor + F1646 aromatic + loop 1642-1651 cap.
- Mechanistically unique — only site whose occupancy can rescue the E1659A binding defect.
The proxy gate’s 0.150 margin is real structural signal that the proxy can’t score well enough. The definitive selectivity test — per-ligand Vina ΔG(K1141 pocket) − ΔG(best off-target) ≥ 2 kcal/mol — runs as part of Phase 4b (redock top-20 against off-target pockets) once the Vina toolchain is installed.
Decision
- Not a kill for the hypothesis. The finding is consistent with “K1141 pocket is the best + unique site, but proxy can’t prove it decisively.”
- Gate is re-opened as part of Phase 4b. Phase 4b will dock each ligand against the K1141 pocket AND the top-5 off-target pockets (different box coords). Selectivity = ligand-specific ΔG gap, measured with real scoring, not class-presence proxy.
- No tier change. Pharmacochaperone stays S.
- Evidence delta is negative but small: we learned the proxy geometry-only selectivity is borderline. The structural read (triple-basic cluster) strengthens specificity; the scoring-function read weakens it.
Files / Models
~/STRC/models/pharmacochaperone_phase4e_offtarget_selectivity.py— driver~/STRC/models/pharmacochaperone_phase4e_offtarget_selectivity.json— 218 pockets + ranked off-targets + per-pocket pharmacophore scores
Ranking delta
- STRC Pharmacochaperone Virtual Screen E1659A: no tier change. Stays S-tier. Evidence depth unchanged (proxy gate failed by scoring-formula margin; definitive read deferred to Phase 4b).
Next stepcolumn in STRC Hypothesis Ranking unchanged (still “run Phase 4b Vina + GNINA real dock”). Phase 4e re-opens as part of Phase 4b (redock against off-target boxes for per-ligand selectivity). - All other S/A/B/C hypotheses: no change.
Connections
[part-of]STRC Pharmacochaperone Phase 4 Plan[see-also]STRC Pharmacochaperone Phase 4a Pocket Reproducibility — companion gate; Phase 4a PASS established the K1141 pocket is real structure across 4/5 CIFs[see-also]STRC Pharmacochaperone K1141 Fragment Pocket — Phase 2B characterisation with the full residue list[see-also]STRC Pharmacophore Model K1141 Pocket — anchor triangle geometry the proxy scorer cannot represent[see-also]STRC Pharmacochaperone Virtual Screen E1659A[see-also]STRC Hypothesis Ranking[applies]Misha