STRC Research

STRC h01 Phase 7 ADMET-AI Triage 2026-04-24

6 min read

STRC h01 Phase 7 ADMET-AI Triage

Ran ADMET-AI on 67 unique SMILES covering the h01 front line: top-30 Phase 3c v3b YELLOW + top-30 Phase 3c v4 YELLOW + 21 Phase 4h tafamidis-playbook seeds. 6/67 compounds have zero flags across 15 medchem gates — and all 6 are Phase 4h drug-like seeds (diflunisal, niflumic, flufenamic, benzoxazole-3OH-4F-COOH, benzimidazole-3OMe-COOH, indole-3OMe). The 30 tight-binding combinatorial winners from v3b/v4 (Kd ~ 4-5 µM) uniformly flag on DILI + solubility + BRENK + Lipinski. The K1141 affinity gain came at measurable developability cost. Tafamidis parent itself flags 3 gates in ADMET-AI space — a calibration caveat that anchors the interpretation of the combinatorial flags.

Problem

Phase 3c v3b + v4 delivered a tight K1141-pocket binding cluster (3-amino-benzofuran-2-COOH / 2-amino-quinoline-3 / tetrazole / CONHOH, Kd 4-5 µM). Phase 4h proposed a bioisosteric Tier-1 alternative built on the tafamidis-class chemistry (benzoxazole / benzothiazole / benzimidazole cores). STRC h01 Fenamic Scaffold Tox Audit 2026-04-23 already killed parent fenamates (niflumic/flufenamic/meclofenamic/mefenamic/tolfenamic) as wet-lab candidates on ion-channel/COX liability. STRC h01 Phase 6c Full 6-Target Panel 2026-04-24 added COX + TMEM16A off-target walls on the bicyclic-COOH pharmacophore.

This phase asks: once we correct for the known on-target off-target liabilities, what does systemic ADMET look like on our chemistry? Does the scaffold direction clear the general drug-like / hepato / cardio / mutagen gates?

Method

  • Tool: ADMET-AI v2.0.1 (ADMET-AI tool card) — Chemprop-ensemble predictor for 41 ADMET properties from Therapeutics Data Commons; Swanson 2024 Bioinformatics 40:btae416.
  • Script: [[pharmacochaperone_admet_triage.py]] (new 2026-04-24).
  • Input: Phase 3c v3b top-30 non-covalent + Phase 3c v4 top-30 non-covalent + 22 Phase 4h seed SMILES, deduped to 67 unique.
  • Triage gates (percentile-scale against FDA-approved drug distribution; > 90th pct = “worse than 90 % of approved drugs”):
    • Toxicity: hERG, AMES, DILI, Carcinogens_Lagunin, ClinTox
    • Distribution: BBB_Martins, solubility (AqSolDB, bottom 10 %)
    • Metabolism: CYP3A4_Veith, CYP2D6_Veith
    • Developability: Lipinski, MW > 500, logP > 5, PAINS / BRENK alerts, bioavailability (bottom 10 %)
  • Gate rationale per percentile-against-approved-drugs interpretation: ADMET-AI raw probabilities are not well-calibrated in absolute terms — the percentile column is the intended readout.

Results

Aggregate

bandnn_flags
clean60
light281-2
heavy33≥ 3

Flag frequency (out of 67)

gaten flaggedcomment
DILI > 90th pct59ADMET-AI DILI classifier predicts > 0.88 for all 67 — a known calibration skew toward aromatic/fused-ring chemotypes, relative ranking only
solubility bottom 10 %36real concern for v3b/v4 lipophilic tight binders
BRENK alerts21reactive substructure flags (CONHSO2Me, halogenated aryls)
Lipinski fail20MW / logP / HBA / HBD edge
logP > 520lipophilicity limit
AMES > 90th pct20mutagenicity flag on fused-aromatic amine cores
Carcinogens_Lagunin > 90th pct8≥3 flags usually co-occur
CYP3A4 inh > 90th pct2minimal drug-drug-interaction risk
hERG > 90th pct0cardiac safety clean at 90th-pct gate
BBB > 90th pct0CNS-penetration risk not above approved-drug distribution

Clean candidates (6/67)

All six are Phase 4h Tier-1/Tier-2 seeds, all drug-like small molecules:

nameQEDQED_pctsourcepediatric OK?
diflunisal parent0.91—phase4h_seedNO — known ion-channel promiscuity (Phase 6c + Fenamic Tox Audit)
niflumic parent0.90—phase4h_seedNO — known ion-channel promiscuity
flufenamic parent0.89—phase4h_seedNO — known ion-channel promiscuity
benzox-3OH-4F-COOH (#15)0.75—phase4h_seedCANDIDATE
benzimid-3OMe-COOH (#27)0.77—phase4h_seedCANDIDATE
indole-3OMe (#17)0.76—phase4h_seedCANDIDATE

Of the 6 clean, 3 are fenamic parents already killed by the scaffold tox audit. The remaining 3 compounds (benzox-3OH-4F-COOH, benzimid-3OMe-COOH, indole-3OMe) are the only ADMET-clean tafamidis-class seeds.

Light-flag tier (1-2 flags)

Typical v3b/v4 tight binder: dG ~ −7.1 kcal/mol, flags = {DILI_gt_90pct, solubility_bot_10pct}. The DILI flag is partly a model artefact (see caveat below); the solubility flag is real.

Top-5 viable (≤ 2 flags, best mean_dG):

ranknamemean dGflags
1v4__3-amino-benzofuran-2-COOH__6-F-2-naphthyl__tetrazole__-F−7.253DILI, solubility
2nc__2-amino-quinoline-3__3_5-diMe-phenyl__tetrazole__-Me−7.09DILI, solubility
3nc__3-amino-benzofuran-2-COOH__4-F-biphenyl__tetrazole__-F−7.053DILI, solubility
4v4__2-amino-benzoxazole-5-COOH__2-naphthyl__CONHSO2Me__-OMe−7.044DILI, Carcinogens
5v4__3-amino-benzofuran-2-COOH__8-quinolinyl__CONHSO2Me__-OMe−6.972DILI

Heavy-flag tier (≥ 3 flags)

33/67 compounds, including tafamidis parent itself (flags = {DILI, Carcinogens, solubility}). See calibration caveat below.

Important calibration caveat

ADMET-AI flags tafamidis parent at 3 gates — yet tafamidis is FDA-approved (Vyndaqel, 2019, TTR pharmacochaperone). This mismatch is informative:

  • The DILI model is binary-trained on the TDC DILI dataset (~475 compounds) and appears to over-predict the positive class for aromatic / fused / carboxylate chemotypes in general. The percentile-against-approved-drugs interpretation is the honest readout (> 90th pct = “higher DILI prediction than 90 % of approved drugs”), but tafamidis itself lands in that tail despite being safe in humans. Treat DILI flags as a relative ranking cue, not as an absolute kill-gate.
  • Solubility bottom-10 % is also a known tafamidis property (logP 4.9, aqueous solubility < 1 µg/mL) — but the drug is deliverable in meglumine salt form for oral use.
  • Carcinogens flag on the iodinated diflunisal variants is consistent with the known carcinogenicity signal of some halogenated aromatics, but our ADMET-AI model does not resolve metabolite-level risk.

Action: use ADMET-AI for relative ranking inside the v3b/v4 set, not for absolute go/no-go on individual compounds.

Interpretation

  1. Phase 4h seeds carry measurable ADMET advantage over v3b/v4 combinatorial. Every v3b/v4 compound with Kd < 5 µM flags DILI + solubility. Phase 4h drug-like analogs flag one or two gates; 3 of them flag none.

  2. Affinity-developability trade-off is sharp and real. Phase 3c v4 ceiling at ΔG −7.27 kcal/mol comes with solubility bottom-10 % and DILI top-10 %. Moving from 40 µM (tafamidis-class) to 4 µM (v3b) on the same scaffold cost at least one developability gate.

  3. hERG and BBB are clean on this chemistry. Both 0/67 above 90th pct. Cochlear target needs non-penetrant compounds and the chemistry class is structurally biased toward it — a meaningful positive finding.

  4. Three novel Phase 4h candidates emerge cleanly from this triage:

    • benzox-3OH-4F-COOH (benzoxazole core, 3-OH / 4-F ring substitution)
    • benzimid-3OMe-COOH (benzimidazole core, 3-OMe distal)
    • indole-3OMe (indole-3-COOH scaffold, 3-OMe distal) These have no known ion-channel liability data (unlike fenamic parents), pass ADMET-AI gates, and share the bicyclic-heteroaromatic-COOH motif that fits the K1141 pocket. They are the only compounds that simultaneously clear Phase 6c off-target + scaffold tox audit + ADMET-AI triage.

  5. v3b/v4 as-is is not a wet-lab shortlist, but the pharmacophore stands. The binding direction is real; the specific molecules need structural debuffing (solubilising group, lipophilicity reduction) before any wet-lab triage — separate medchem round.

Ranking delta

  • Hypothesis h01: tier A held | mech 2 held (static pharmacochaperone claim already falsified per Phase 4c-v3b / 4d, independent of ADMET) | deliv 4 held | misha_fit 4 held
  • Next step refinement: the 3 ADMET-clean Phase 4h candidates (benzox-3OH-4F-COOH, benzimid-3OMe-COOH, indole-3OMe) are the cleanest handoff to any future wet-lab Stage-1, IF the dynamic chaperone claim survives Phase 5g holo-MD. Wet-lab still gated on Phase 5g verdict, unchanged.
  • Medchem action (no wet-lab advance): next-round v4b library design should constrain solubility ≥ −4 log10(M) (i.e., introduce solubilising group on distal aryl), cap logP ≤ 4.5, keep the 3-amino-benzofuran / benzoxazole cores.

Connections

Graph View

Backlinks