STRC h01 Phase 8c — v5 ADMET-AI Triage
After Phase 8b’s 60-ligand v5 library (non-planar tails + CONHOH + extended R3), Phase 8c screens the top-20 combined-ranked ligands through ADMET-AI (Swanson 2024, Chemprop ensemble on 41 TDC endpoints). Using the DrugBank-approved-percentile gate at 90th-pct-or-better across 10 critical safety/PK endpoints (hERG, BBB, DILI, AMES, Carcinogens, ClinTox, Bioavailability, CYP3A4, Clearance, LD50): 12/20 v5 are fully flag-free (60%) vs 0/10 v3b benchmarks (0%). The Phase 8b combined-top lead
v5__aq3__adamantyl__CONHOH__-Clis flag-free on all 10 gates — the first ADMET-clean pharmacochaperone candidate in the STRC line. Critical structural insight: all 1-naphthyl variants fail DILI at 93-95 percentile; replacing naphthyl with non-planar aliphatic cages (adamantyl, BCO, norbornyl) simultaneously addresses hERG AND DILI class liabilities.
Problem
Phase 6c identified bicyclic-aromatic-COOH as a hERG-class pharmacophore (21/28 v3b hit sub-10 µM hERG). Phase 8b proposed non-planar aliphatic tails as mitigation via logP reduction ([5.47 v3b top-1 → 1.94 v5 top-1]) and out-of-plane geometry. Phase 8c tests whether this structural change actually clears the broader ADMET panel, not just hERG.
Secondary question: does the v3b Coulomb-top-1 (nc__2-amino-quinoline-3__1-naphthyl__CONHOH__-CF3) — which Phase 5k-B identified as the best ensemble-Coulomb binder and the template for the v5 scaffold — itself pass ADMET? If yes, v5 inherits a clean starting point; if no, v5 must solve toxicology structurally.
Method
- Tool: ADMET-AI (Swanson 2024 JCIM; Chemprop message-passing ensemble on 41 TDC endpoints; DrugBank-approved-percentile scaling).
- Env:
/opt/miniconda3/envs/admet-ai/bin/admet_predict. - Input: 31 SMILES = v5 top-20 (by combined Vina+APBS score, Phase 8b) + 11 v3b references (9 top-Vina + Vina-top-1 + Coulomb-top-1).
- Gate: 10 percentile columns, threshold 90. Convention: a value >90 means the compound is worse than 90% of approved drugs on that axis (problematic). A compound with
gate_flags = 0means it sits in the top ~90% of approved drugs on all 10 axes — ADMET-clean. - Gate columns:
hERG, BBB_Martins, DILI, AMES, Carcinogens_Lagunin, ClinTox, Bioavailability_Ma, CYP3A4_Veith, Clearance_Hepatocyte_AZ, LD50_Zhu— all_drugbank_approved_percentilesuffixes.
Script: pharmacochaperone_phase8c_v5_admet_triage.py (new 2026-04-24). Output: artifacts/phase8c_admet/phase8c_admet_{raw,summary,per_source}.csv + phase8c_admet_summary.json.
Results
Per-source summary
| source | n | mean flags | n clean | frac clean | hERG mean | BBB mean | DILI mean | AMES mean |
|---|---|---|---|---|---|---|---|---|
| v5_top20 | 20 | 0.70 | 12 | 60% | 79.1 | 62.1 | 77.2 | 86.5 |
| v3b_top10 (Vina) | 9 | 1.11 | 0 | 0% | 74.9 | 45.1 | 96.8 | 81.0 |
| v3b_vina_top (#1) | 1 | 2.00 | 0 | 0% | 71.1 | 56.7 | 95.9 | 98.2 |
| v3b_coulomb_top (Phase 5k-B #1) | 1 | 2.00 | 0 | 0% | 72.0 | 50.2 | 96.9 | 96.3 |
DILI is the decisive difference: v3b averages 96-97 percentile (liability); v5 top-20 averages 77 percentile (clean). hERG averages are similar (~73-79) — v5 is marginally worse on hERG means but compensates on DILI by a much larger margin.
v5 top-20 per-ligand triage
| clean | flags | hERG | BBB | DILI | ligand |
|---|---|---|---|---|---|
| ✓ | 0 | 80 | 65 | 73 | adamantyl__CONHOH__-Cl (Phase 8b combined #1) |
| ✓ | 0 | 82 | 68 | 73 | 4-Me-4-OH-cyclohex__CONHOH__-CF3 |
| ✓ | 0 | 83 | 73 | 66 | bicyclo[2.2.2]oct-1-yl__CONHOH__-CF3 |
| ✓ | 0 | 82 | 68 | 71 | adamantyl__CONHOH__-F |
| ✓ | 0 | 80 | 64 | 73 | norbornan-2-yl__CONHOH__-CF3 |
| ✓ | 0 | 73 | 54 | 69 | bicyclo[2.2.2]oct-1-yl__CONHOH__-CN |
| ✓ | 0 | 76 | 51 | 76 | adamantyl__CONHOH__-CN |
| ✓ | 0 | 84 | 65 | 72 | adamantyl__CONHOH__-OMe |
| ✓ | 0 | 81 | 74 | 64 | bicyclo[2.2.2]oct-1-yl__CONHOH__-F |
| ✓ | 0 | 81 | 66 | 83 | 1-indanyl__CONHOH__-CF3 |
| ✓ | 0 | 82 | 66 | 71 | adamantyl__CONHOH__-CF3 |
| ✓ | 0 | 76 | 56 | 80 | 3,5-diMe-phenyl__CONHOH__-CF3 |
| ✗ | 1 | 81 | 73 | 66 | bicyclo[2.2.2]oct-1-yl__CONHOH__-Cl |
| ✗ | 1 | 76 | 48 | 88 | biphenyl__CONHOH__-CF3 |
| ✗ | 2 | 78 | 62 | 93 | 1-naphthyl__CONHOH__-H |
| ✗ | 2 | 69 | 38 | 95 | 1-naphthyl__CONHOH__-CN |
| ✗ | 2 | 74 | 56 | 94 | 1-naphthyl__CONHOH__-Cl |
| ✗ | 2 | 77 | 61 | 93 | 1-naphthyl__CONHOH__-F |
| ✗ | 2 | 82 | 60 | 67 | spiro[4.4]non-1-yl__CONHOH__-CN |
| ✗ | 2 | 86 | 73 | 76 | 3,3-diMe-indanyl__CONHOH__-Cl |
Every 1-naphthyl variant fails DILI (93-95 percentile) — naphthyl-containing carboxamides have a known hepatotoxicity signal in ADMET-AI’s DILI Chemprop model (consistent with naphthalene metabolite formation via naphthoquinone electrophiles). All 4 naphthyl v5 entries and 5 of the top-10 v3b entries flag on this axis.
Adamantyl is universally clean: all 5 adamantyl variants in the v5 top-20 pass all 10 gates. This is the cleanest sub-series in the entire library.
Combined with Phase 8b scoring
The v5 Phase 8b top-1 adamantyl__CONHOH__-Cl (combined −8.19 kcal/mol) is also ADMET top-1 (flags=0). No tradeoff between potency and safety for this compound.
The Vina-best v5 1-naphthyl__CONHOH__-Cl (Vina −7.04, #4 combined) fails DILI at 94 — do not pursue.
Phase 8c final shortlist for wet-lab candidate selection:
v5__aq3__adamantyl__CONHOH__-Cl— combined #1, ADMET-clean, logP 1.94v5__aq3__bicyclo[2.2.2]oct-1-yl__CONHOH__-CF3— combined #17, ADMET-clean, most constrained cage, logP 2.06v5__aq3__norbornan-2-yl__CONHOH__-CF3— combined #6, ADMET-clean, logP 1.53, smallest 3D tailv5__aq3__adamantyl__CONHOH__-F— combined #5, ADMET-clean, logP 1.67, fluoro R3 for metabolic stabilityv5__aq3__adamantyl__CONHOH__-CN— combined #8, ADMET-clean, logP 1.27, polar R3 for solubility
Interpretation
- v5 design hypothesis is validated experimentally (in-silico): replacing planar naphthyl/biphenyl tails with non-planar aliphatic cages makes a measurable improvement in ADMET at 60% clean rate vs 0% for v3b.
- Naphthyl DILI liability is the dominant safety problem in v3b — even if we had picked a v3b compound with less hERG (e.g. by adjusting substituent), the naphthyl core still flags DILI at 93-97 percentile. The v5 tail swap solves both.
- The combined-optimization objective did not reward toxic compounds. v5 top-1 by combined score is also v5 top-1 by ADMET cleanness. Phase 5k’s Coulomb-aware ranking is self-consistent with safety when paired with the right structural filter.
- Adamantyl is the scaffold to pursue. 5/5 of all adamantyl-containing v5 ligands are clean across all 10 ADMET axes. This should be the core of v5.2 expansion.
- Deliv upgrade path opens. The pre-Phase-8 position was
deliv 3(intracochlear route required due to Phase 6c hERG + DILI). v5 top-1 clears both on in-silico ADMET — the systemic oral route is now at least open for consideration pending wet-lab ADME validation. Not a finaldeliv 4, but a genuine upgrade path.
Limitations
- In-silico ADMET. ADMET-AI predictions are Chemprop neural-net models; experimental ADME studies (Caco-2, hepatocyte clearance, hERG patch-clamp, Ames strain-specific) must confirm before clinical planning. Model generalization on non-planar adamantyl scaffolds is untested in the published TDC training data; v5’s novelty is a double-edged sword here.
- Percentile scale, not absolute probability. A compound at hERG 80th percentile is “in the worst 20% of approved drugs” on hERG, which does NOT mean “20% chance of hERG block in human heart.” The 90th-pct gate is a conservative threshold; adamantyl at hERG 80 still deserves patch-clamp confirmation.
- DILI model is controversial. Known to over-flag aromatic polycyclic scaffolds. The 93-95 pct signal on naphthyl variants is consistent with classical naphthalene metabolite concerns but may over-penalize naphthalenes that have compensating deactivation (e.g., fluorine blocking the 1,2-naphthoquinone pathway). Nonetheless, the class liability signal is strong enough to justify avoiding naphthalenes in v5.
- 10 gates is a subset. ADMET-AI produces 41 endpoints; I gated on 10 critical ones for the cochlear-drug-development context. A broader analysis (e.g., P-gp efflux, BCRP, UGT metabolism) would refine the ranking.
- No off-target selectivity re-tested on v5. Phase 6c was on v3b. Phase 8f (scheduled) will re-run off-target panel on v5 top-5 directly.
Ranking delta
- Hypothesis h01: tier A held | mech 3 held | deliv 3 held but deliv-4 upgrade path OPENED (ADMET-clean compound exists; systemic route not pre-closed) | misha_fit 4 held (still requires wet-lab confirmation for upgrade)
- Next-step update:
v5__aq3__adamantyl__CONHOH__-Clis the first wet-lab-ready pharmacochaperone candidate — Vina + APBS + ADMET all clean.- v5.2 library design priority: grow adamantyl-centric series (Phase 8d scheduled). De-prioritize all naphthyl variants regardless of Vina ranking (DILI class liability).
- Phase 8f off-target panel (scheduled) will re-verify hERG + cochlear-channel selectivity on v5 top-5; if clean, intracochlear requirement may become optional.
- Combined chemistry principle for STRC pharmacochaperone:
2-amino-quinoline-3 core + adamantyl (or BCO/norbornyl) tail + CONHOH head + Cl/F/CN at C4. This is the active forward template for v5.2 medchem.
Connections
[part-of]h01 hub[builds-on]STRC h01 Phase 8 v5 Library Coulomb-Aware Design 2026-04-24 — took v5 top-20 as input[supersedes]STRC h01 Phase 7 ADMET-AI Triage 2026-04-24 — that phase flagged v3b-scale DILI/hERG liability; Phase 8c validates that v5 structural change resolves both[complements]STRC h01 Phase 6c hERG Extension 2026-04-24 — direct experimental hERG docking vs predictive Chemprop; both agree that naphthyl tails are liable[see-also]ADMET-AI[source]Swanson 2024 ADMET-AI J. Chem. Inf. Model.