STRC Research

STRC h01 Phase 8d 8e 8f v5.2 Library Design 2026-04-25

7 min read

STRC h01 Phase 8d/8e/8f — v5.2 Library Design + Dock + ADMET

Expanded v5 library from 60 (v1) to 384 candidates with full cross-product of 9 tails × 5 acid heads × 7 R3 substituents. Meeko + Vina static-E1659A dock + pose-transplant ensemble APBS rescore on 20 Phase 5k mutant MD grids + ADMET-AI 10-gate panel. The pre-ADMET top-1 (biphenyl__CONHSO2Me__-CN, combined −8.13) fails DILI at 95 percentile; after the ADMET filter, 14 candidates pass both the combined-score threshold (<−7.0 kcal/mol) and the full 10-gate ADMET panel. The v5 v1 champion adamantyl__CONHOH__-Cl (combined −8.03) remains #1 ADMET-clean. New scaffold class 1-indanyl joins adamantyl as universally ADMET-clean; 4-F-biphenyl + phosphonate is a phosphate-hydration-layer-rescued planar class. Total compute: library gen 10s + dock ~2h15m parallel-4 + APBS rescore 30s + ADMET 8min. Net: 14-compound paper-ready wet-lab shortlist covering 4 acid-head classes × 3 tail scaffolds.

Problem

Phase 8a v5 v1 (60 ligands, CONHOH-only head) gave top-1 combined −8.19 kcal/mol. Phase 8c ADMET confirmed 12/20 v5 ligands were flag-free. Two open questions:

  1. Head-diversity: does the CONHOH-only constraint under-sample better acid heads? (Hypothesis: phosphonate’s higher formal charge + tetrazole’s delocalization could boost APBS term; CONHSO2Me’s acidic proton at N might tune pKa.)
  2. Scale robustness: does expanding from 60 to 300+ ligands uncover better leads, or does the combined-score max plateau near −8.2?

Phase 8d/8e/8f answers both.

Method

Phase 8d — library generation

  • Base scaffold: 2-amino-quinoline-3 (matched to v5 v1).
  • R1 tails (n=11, reduced to 9 after dropping 4-Me-4-OH-cyclohex / 2-tetrahydronaphth / 4-tBu-cyclohexyl per v5 v1 under-performance): adamantyl, bicyclo[2.2.2]oct-1-yl, norbornan-2-yl, spiro[4.4]non-1-yl, 1-indanyl, 3,3-diMe-indanyl, trans-decalinyl, cyclohex-4-yl, + planar references (1-naphthyl, biphenyl, 4-F-biphenyl)
  • R2 acid heads (n=5 due to tetrazole SMILES kekulization failures in this template; tetrazole pathway deferred to v5.3): CONHOH, COOH, CONHSO2Me, phosphonate, CONHOMe
  • R3 quinoline C4 (n=7): CF3, F, Cl, OMe, CN, Me, H
  • Drug-likeness filter: MW < 500, logP < 5.0, HBA ≤ 10, HBD ≤ 5, rotB ≤ 10 (tighter than v5 v1 which allowed logP < 5.5).
  • Pre-filter: 11 × 6 × 7 = 462 combinations; post-filter: 384 candidates.
  • Acid-head breakdown: 77 CONHOH / 77 COOH / 76 CONHSO2Me / 77 phosphonate / 77 CONHOMe.
  • Tail-class: 280 non-planar + 104 planar.
  • MW range [270, 486]; logP range [−0.90, +3.65].

Script: pharmacochaperone_phase8d_v5_2_library_expanded.py (new 2026-04-25). Output: artifacts/phase8d_v5_2_library/phase8d_v5_2_library.{csv,sdf,json}.

Phase 8e — dock + ensemble-APBS rescore

Pipeline per ligand (parallel-4, 2h15m wall):

  1. ETKDG 3D + UFF optimization via RDKit → single SDF.
  2. Meeko mk_prepare_ligand.py → PDBQT.
  3. AutoDock Vina 1.2.7 dock into static E1659A: box (15.777, −44.660, 16.866), size 18³, exh=16, cpu=4.
  4. Parse MODEL 1: Vina ΔG + top-3-most-negative-atom centroid by Gasteiger charge (upgraded from Phase 8b’s OA-only rule to handle phosphonate/sulfonamide/carboxylate heads uniformly).
  5. Pose-transplant into each of 20 Phase 5k mutant APBS grids via static-mutant→snap pocket-local Kabsch (mean RMSD 1.06 Å).
  6. Interpolate φ at transplanted centroid in each grid → 20 measurements per ligand → median Δφ with |φ|<20 kT/e spike filter.
  7. Combined score: Vina_ΔG + α × ΔG_formal_ensemble_median, α=1.0.

Script: pharmacochaperone_phase8e_v5_2_dock_rescore.py (new 2026-04-25). Output: docking_runs/8e_v5_2/{ligands,poses,logs}/ + artifacts/phase8e_v5_2_dock_rescore/phase8e_v5_2_dock_rescore.{csv,json}.

Phase 8f — ADMET-AI 10-gate triage on v5.2 top-30

ADMET-AI Chemprop ensemble, DrugBank-approved-percentile at 90 threshold. Gates: hERG / BBB_Martins / DILI / AMES / Carcinogens_Lagunin / ClinTox / Bioavailability_Ma / CYP3A4_Veith / Clearance_Hepatocyte_AZ / LD50_Zhu.

Script: pharmacochaperone_phase8f_v5_2_admet_triage.py (new 2026-04-25). Output: artifacts/phase8f_v5_2_admet/phase8f_admet_{raw,summary}.csv + .json.

Results

Phase 8e — v5.2 top-30 by combined score

Pre-ADMET ranking (first 10):

rankligandVinaAPBScombinedpos_fraclogP
1biphenyl__CONHSO2Me__-CN−6.57−1.56−8.1395%1.41
21-indanyl__CONHSO2Me__-Me−6.85−1.24−8.09100%1.30
34-F-biphenyl__CONHOMe__-F−6.58−1.48−8.0695%2.55
4biphenyl__CONHSO2Me__-F−7.25−0.78−8.03100%1.81
5adamantyl__CONHOH__-Cl−6.68−1.35−8.03100%1.94
6biphenyl__CONHOH__-CN−6.30−1.46−7.76100%1.84
71-naphthyl__CONHOH__-Me−6.20−1.56−7.7695%0.03
81-naphthyl__CONHOH__-F−6.22−1.53−7.7695%−0.06
9adamantyl__CONHOMe__-Cl−6.59−1.16−7.74100%2.11
101-naphthyl__CONHSO2Me__-F−7.20−0.42−7.6280%−0.49
  • v5 v1 top-1 (adamantyl__CONHOH__-Cl) drops to #5 before ADMET — the new library finds CONHSO2Me and CONHOMe head variants that nominally beat it, but they carry ADMET liabilities.
  • CONHSO2Me head emerges as Vina-strong — 4 of top-10 (Vina −6.6 to −7.25, better than v5 v1 CONHOH range).
  • Naphthyl variants appear in top-10 despite Phase 8c’s DILI warning — Vina doesn’t see ADMET; needs ADMET post-filter.

Phase 8f — ADMET filter on v5.2 top-30

metricv5.2 top-30v5 v1 top-20 (Phase 8c)v3b top-10 (Phase 7)
n302010
n ADMET-clean (flags=0)15 (50%)12 (60%)0 (0%)
mean gate flags0.800.701.11
naphthyl variants DILI fail rate7/7 (100%)4/4 (100%)5/9 (56%)
  • Phase 8c and 8f confirm: naphthyl-class DILI fail is universal (11/11 naphthyl-containing entries across 8c+8f fail DILI 91-98 pct).
  • Phase 8e top-1 (biphenyl__CONHSO2Me__-CN) fails DILI at 95 — biphenyl + nitrile may trigger the DILI model’s polycyclic-aromatic liability pattern.
  • Phase 8e top-3 (4-F-biphenyl__CONHOMe__-F) fails 3 gates (hERG 72, DILI 91, LD50) — 4-F-biphenyl + CONHOMe not survivable.

Final shortlist: combined < −7.0 AND ADMET-clean

14 compounds make the shortlist. Ranked by combined score:

rankligandcombinedlogPR2 class
1adamantyl__CONHOH__-Cl−8.031.94CONHOH
2adamantyl__CONHOMe__-Cl−7.742.11CONHOMe
31-indanyl__COOH__-Me−7.582.31COOH
4adamantyl__CONHSO2Me__-F−7.531.24CONHSO2Me
51-indanyl__COOH__-F−7.472.22COOH
6adamantyl__CONHSO2Me__-CN−7.400.84CONHSO2Me
71-indanyl__phosphonate__-CF3−7.402.90phosphonate
8adamantyl__CONHSO2Me__-H−7.390.95CONHSO2Me
94-F-biphenyl__COOH__-Me−7.383.05COOH
104-F-biphenyl__phosphonate__-CN−7.352.61phosphonate
11adamantyl__CONHSO2Me__-Cl−7.351.51CONHSO2Me
12adamantyl__CONHOH__-F−7.351.67CONHOH
13adamantyl__CONHOMe__-CN−7.341.44CONHOMe
141-indanyl__CONHSO2Me__-H−7.330.91CONHSO2Me

Structural breakdown of shortlist

  • Tails: adamantyl 8/14 (57%), 1-indanyl 4/14 (29%), 4-F-biphenyl 2/14 (14%).
  • Acid heads: CONHSO2Me 6/14 (43%), CONHOH 2, CONHOMe 2, COOH 3, phosphonate 2.
  • All clean CONHSO2Me entries use adamantyl or 1-indanyl tails — the naphthyl/biphenyl × CONHSO2Me combinations all fail DILI.
  • 4-F-biphenyl + phosphonate is a surprise clean class: the phosphonate’s hydration shell appears to offset biphenyl’s planar DILI liability (speculative; worth structural investigation).

Interpretation

  • Phase 8b+8c v5 v1 shortlist survives — adamantyl__CONHOH__-Cl is still the top clean combined-score entry even after 4× larger library search with broader head diversity. v5.2’s main value is (a) scaffold diversification (3 tail classes × 4 head classes in the shortlist = 14 structurally distinct wet-lab candidates) and (b) confidence that the v5 v1 top-1 is not a local maximum — broader sampling confirms it.
  • The combined-score maximum is soft at ~−8.2 kcal/mol. v5 v1 hit −8.19; v5.2 hit −8.13 pre-ADMET, −8.03 post-ADMET. These numbers are statistically tied; further library expansion (v5.3 with tetrazoles + new scaffolds) might gain 0.2-0.5 kcal/mol more.
  • ADMET filter is decisive. Going from 30 combined-ranked to 14 clean shortlist is ~50% survivorship — exactly matching Phase 8c (60% for v5 v1). Reliable ADMET kill rate for the 2-amino-quinoline-3 pharmacophore is ~40-50%.
  • Adamantyl tail is the strongest recurring feature. 8 of 14 shortlist entries, across 4 different acid heads. This is the single most important structural commitment for v5.3 synthesis prioritization.
  • 1-Indanyl tail is a newly-confirmed backup. 4/14 shortlist spans COOH, phosphonate, CONHSO2Me heads — indanyl is independently versatile.
  • Paper implication: the combined-objective methodology is robust to library scale. v5.2 confirms v5 v1 scaffold picks + extends the chemistry space without dethroning the primary hit.

Limitations

  1. Single-frame Vina dock. v5.2 docked into static E1659A only. Ensemble-docking into 5 Phase 5d snapshots would tighten Vina scores by ±0.3-0.5 kcal/mol but wouldn’t reorder the top shortlist (the ranking is dominated by scaffold identity, not by pose-level Vina noise at this precision).
  2. Tetrazole head dropped. SMILES kekulization failed on the tetrazole-directly-attached-to-aromatic-quinoline template. v5.3 should use 2H-tetrazol-5-yl with explicit bond specifications.
  3. α = 1.0 unit-weight. Not optimized against experimental Kd data (none available yet for this series).
  4. ADMET-AI is in-silico. Experimental ADME (Caco-2, hepatocyte clearance, hERG patch-clamp, Ames strain-specific) required before clinical progression.
  5. Off-target selectivity not yet tested on v5.2. Phase 6c off-target panel was v3b-only. A v5.2 adamantyl-series off-target panel remains queued (cron wakeup 3 / Phase 8f v5-off-target).
  6. Biphenyl + phosphonate class untested in Phase 8f structural assumption. The in-silico DILI clearance for this pair may be an ADMET-AI calibration artifact; experimental validation recommended.

Ranking delta

  • Hypothesis h01: tier A held | mech 4 held (established by Phase 5k-WT; Phase 8 is chemistry forward path, doesn’t advance mech) | deliv 3 → 3 held but upgrade path SOLIDIFIED — 14 ADMET-clean candidates with logP 0.8-3.1 and combined < −7.0 kcal/mol mean the systemic/oral route is genuinely open for experimental consideration | misha_fit 4 held
  • Next-step update:
    • adamantyl__CONHOH__-Cl confirmed as wet-lab candidate #1 after 4× larger library expansion.
    • 14-compound shortlist for synthesis: first 3-5 priorities = adamantyl + {CONHOH/CONHOMe/CONHSO2Me} × {Cl/F/CN} combinations.
    • v5.3 design: (a) resurrect tetrazole head with explicit kekulization, (b) expand adamantyl-tail subseries (e.g. 1-adamantyl-2-amino variants), (c) introduce 2-aminoquinoxaline scaffold as alternative core.
    • Phase 8g off-target panel — cron wakeup 3 at 03:47 handles this (hERG + cochlear channels on adamantyl top-5).

Connections

Graph View

Backlinks