A multi-endpoint ADMET panel, running in your browser

Type any molecule (SMILES) and get a 13-endpoint ADMET panel — absorption, distribution, metabolism and toxicity — each with calibrated uncertainty and an applicability-domain flag. Featurisation (RDKit ECFP) and inference both run live in your browser — no server. Models are benchmarked on the Therapeutics Data Commons ADMET group (scaffold splits); each row shows its held-out test metric.

Honest scope: ECFP + small-MLP baselines on open data — not a production ADMET model, and I have not worked in pharmacokinetics. It shows the methods (multi-task modelling, calibration, applicability domain, reproducible benchmarking) transferring to molecular data. A Chemprop D-MPNN would score higher on the hard endpoints. Reference bar: ADMET-AI. Code: github.com/mlavanga/adme-lab.