STRC Research

ADMET-AI

3 min read

ADMET-AI

Chemprop-based ensemble predictor for 41 ADMET properties (absorption, distribution, metabolism, excretion, toxicity) from Therapeutics Data Commons (TDC). Open-source Python package, runs on CPU in seconds per molecule.

What It Does

Predicts, from SMILES alone:

  • Physicochemical: molecular weight, logP, logS (aqueous solubility), TPSA, rotatable bonds, HBA/HBD, Lipinski/Veber flags.
  • Absorption: Caco-2 permeability, HIA (human intestinal absorption), PAMPA, oral bioavailability.
  • Distribution: BBB (blood-brain barrier), PPB (plasma protein binding), VD (volume of distribution).
  • Metabolism: CYP1A2 / CYP2C9 / CYP2C19 / CYP2D6 / CYP3A4 inhibition + substrate.
  • Excretion: half-life, clearance.
  • Toxicity: hERG, AMES mutagenicity, DILI (drug-induced liver injury), carcinogenicity, LD50, skin/eye irritation, Tox21 panel (NR-AR, NR-AhR, etc.).

Reference: Swanson et al. 2024 Bioinformatics 40:btae416. Models trained on TDC benchmarks; 41 classifiers + regressors ensembled.

Install

conda create -n admet-ai python=3.11 -y
conda run -n admet-ai pip install admet-ai

Verified 2026-04-24: pip install pulls torch 2.11.0, chemprop 2.2.3, admet-ai 2.0.1. ~2.5 GB disk, ~30 s per SMILES on CPU (M5 Max).

How to Use

CLI

conda run -n admet-ai admet_predict \
  --data_path input.csv \
  --save_path output.csv \
  --smiles_column smiles

Input CSV must have a smiles column. Output CSV has 41 prediction columns appended.

Python

from admet_ai import ADMETModel
model = ADMETModel()
preds = model.predict(smiles=["CCO", "c1ccccc1O"])
# preds is a pandas DataFrame with 41 property columns

STRC Research Usage

  • Phase 4h+ medchem triage — post-Vina filter on top-30 v3b/v4 YELLOW candidates. Uses: CYP3A4 inhibition (drug-drug interactions), hERG (cardiac safety — FDA-mandatory), BBB (CNS penetration — cochlear targets care about opposite), PPB (free-fraction at K1141), DILI + carcinogenicity (pediatric indication safety floor), logP/logS/TPSA (physicochemical developability).
  • Complements Phase 6c cochlear off-target panel — ADMET-AI gives systemic-PK context; Phase 6c gives on-target selectivity. Both needed for Phase 8 Stage-1 nomination.
  • Does NOT cover: cochlear-specific PK (RWM permeability, scala media distribution) — those live in Phase 4 plan’s compartmental model. ADMET-AI is general mammalian systemic PK.

Known Limitations

  • Training data dominated by drug-like space (MW 200-600, logP 1-5). Predictions for fragments (<200 MW) or lipid-conjugates (>600 MW) are less reliable.
  • Binary classifiers (hERG, BBB) output probabilities; use 0.3 / 0.7 as soft flags, not hard thresholds.
  • Does not resolve enantiomers; racemic predictions by default.
  • PPB and VD predictions are human-oriented; rodent translation requires species correction.

Connections

Graph View

Backlinks