Extended ADMET Panel: 15+ Endpoints in Every Report

Every time ToxScreen scores a compound, ADMET-AI runs a forward pass that computes 49+ molecular properties. Until now, ToxScreen surfaced only four of them in the user-facing report: hERG, CYP3A4, CYP2D6, and CYP2C9 — the Composite Toxicity Index (CTI) panel. Everything else was computed and immediately discarded.

That was a waste. Starting with this release, every ToxScreen report now includes a supplementary section with 15+ additional ADMET endpoints across five categories: extended CYP, absorption, distribution, excretion, and safety flags. The endpoints were always being computed; we are now surfacing them.

In the same release, every endpoint — core CTI and supplementary alike — now carries a Tanimoto Applicability Domain badge indicating how far the query compound sits from the model's training distribution.

The Five Supplementary Categories

Extended CYP

The CTI panel already covers CYP3A4, CYP2D6, and CYP2C9. The extended panel adds:

• CYP1A2 — metabolises caffeine, theophylline, and a range of dietary and herbal compounds. CYP1A2 induction is a primary mechanism of herb–drug interaction for St. John's Wort and other botanicals.
• CYP2C19 — metabolises proton pump inhibitors, clopidogrel (prodrug activation), and several antidepressants. Polymorphic; poor metabolisers carry ~3% prevalence in European populations, higher in some Asian populations.

For natural products and botanical extracts — a common ToxScreen use case — CYP1A2 and CYP2C19 are often the first-line interaction risks, making these additions directly relevant to herb-drug interaction assessments.

Absorption

• HIA (human intestinal absorption) — fraction absorbed after oral dosing.
• Oral bioavailability ≥20% — binary flag: is the compound likely to achieve at least 20% oral bioavailability?
• Caco-2 permeability — rate of passive transcellular transport across the intestinal epithelium.
• P-gp substrate — binary flag for P-glycoprotein efflux, which limits CNS penetration and drives some multi-drug resistance phenotypes.

Distribution

• BBB penetration — blood-brain barrier crossing probability. Critical for CNS candidates; equally critical for flagging CNS side-effect risk in peripherally targeted compounds.
• PPB (%) — plasma protein binding percentage. High PPB compresses free fraction and can cause displacement interactions with co-administered drugs.
• Volume of distribution (Vdss, L/kg) — extent of tissue distribution at steady state.

Excretion

• Microsomal clearance (T½ min) — in vitro half-life in microsomal assay. A short T½ predicts rapid hepatic clearance and poor oral exposure.
• Half-life — predicted systemic half-life, relevant to dosing interval design.

Safety Flags

These four endpoints require explicit caveats and are presented with that context in the report.

• Ames mutagenicity (ICH M7) — binary probability. >0.5 = AMES POSITIVE. ICH M7 is notable in that it explicitly accepts QSAR/ML predictions as a substitute for the bacterial reverse mutation bioassay, making this the one endpoint in the panel where a computational flag has direct regulatory weight. A positive call flags the compound for ICH M7 follow-up assessment; it does not constitute a submission-grade result on its own.
• DILI (drug-induced liver injury) — this endpoint is intentionally excluded from the CTI panel because the underlying model is a known over-predictor. Caffeine scores 0.875 and aspirin scores 0.71 — both well-validated non-DILI compounds. The DILI score is surfaced in the supplementary section with a calibration caveat and should be treated as a signal for follow-up rather than a call.
• Carcinogenicity — binary probability from ADMET-AI's rodent carcinogenicity endpoint. Same caveat as DILI: presented as a screening flag, not a regulatory conclusion.
• Skin sensitization — binary probability. Relevant for topical formulations and occupational exposure assessments.
• LD50 (mg/kg) — acute oral toxicity estimate. Useful for rough hazard classification but carries wide confidence intervals at the compound level.

Applicability Domain Badges

Every endpoint in the report — CTI and supplementary — now carries a Tanimoto-AD badge:

The practical effect: when you submit a structurally novel natural product, ToxScreen will now tell you explicitly that the CYP1A2 prediction is OUT OF DOMAIN rather than returning a number that looks authoritative. This is the same AD-abstention philosophy that governs the core CTI panel, extended uniformly across all 15+ endpoints.

A Note on DILI

The DILI endpoint warrants a direct statement. We are surfacing it in the supplementary section because it is useful as a screening signal, but we are not including it in the Composite Toxicity Index because the model over-predicts in well-characterised ways. Caffeine — DILI score 0.875 — is not a liver toxin at normal use levels. Aspirin — DILI score 0.71 — causes GI bleeding via a well-understood COX-mediated mechanism, not hepatotoxicity. If we included raw DILI scores in the CTI, the index would penalise caffeine more harshly than it penalises a genuine hERG blocker.

The DILI score is shown with a calibration caveat in the report. Use it to flag compounds for follow-up rather than as a binary pass/fail gate.

Ames and ICH M7

The Ames mutagenicity flag is the most directly regulatory-relevant endpoint in the supplementary panel. ICH M7 — the international guideline on mutagenic impurities — explicitly states that QSAR/ML predictions are acceptable as a bioassay substitute for the Ames test, provided the model is fit-for-purpose. ToxScreen's ADMET-AI Ames model meets the structural conditions for ICH M7 use, but any compound flagged AMES POSITIVE should be reviewed in a proper ICH M7 assessment before a regulatory submission. The computational flag is a triage tool; the guideline defines the full acceptance criteria.

Report Formats

All 15+ supplementary endpoints are now included in every report format: HTML (interactive), PDF, Excel, and JSON. The JSON schema change is backward-compatible — the supplementary block is a new top-level key and existing integrations will not break. API users who parse the supplementary_admet key will find the full endpoint table; those who do not will see no change in the existing cti block.

What This Is and What It Is Not

To be direct about the scope: these are ML-predicted properties from ADMET-AI, calibrated where possible and tagged with applicability-domain badges. They are not wet-lab measurements. They are not regulatory submissions. They are a structured, machine-generated prior that a medicinal chemist can act on: run the flagged compounds through in vitro assays, check the OUT OF DOMAIN predictions first, and use the Ames flag to triage for ICH M7 assessment.

The supplementary panel increases the information density of a ToxScreen report from 4 endpoints to 19+, all from a single compound submission. The compute cost is zero — ADMET-AI was already computing these properties on every call. We are surfacing what was already there.

Published by the EstimaBio / ToxScreen team. Questions or corrections? Email info@toxscreen.ai.