Enter a molecule. Anima returns instant cardiotoxicity and liver-injury triage — a calibrated risk, a confidence flag, and the nearest known compounds as rationale. Open, non-profit, no animals.
Example — troglitazone, an antidiabetic withdrawn in 2000 for liver toxicity. Anima flags the liver, not the heart. (Astemizole, a withdrawn antihistamine, returns the inverse: cardiac-HIGH, liver-LOW.)
Your SMILES is standardized and featurized with RDKit — Morgan fingerprints, physicochemical descriptors, and structural-alert counts.
A LightGBM cardiotoxicity model (124k matched hERG compounds) and a RandomForest liver model (3.5k human-DILI labels) — the architectures that won an exhaustive, honest bake-off.
A calibrated probability, a risk band, and an applicability-domain flag that says off-domain — low confidence for chemistry unlike the training set. Plus nearest known compounds, so you can see the reasoning.
Honest isn’t a tagline — it’s the method. Every number here is scaffold-split and error-barred, and a scientific-governance review is on the record correcting our own early mistakes.
Scaffold cross-validated on 123,950 matched compounds — at the upper end of published hERG performance, with an applicability-domain flag for chemistry outside the training set.
A three-part review — cheminformatics, toxicology, ML methodology — scrutinized the corpus and methods, and caught an inflated early result (leakage from missing-not-at-random assay coverage) before it shipped.
Human liver injury is largely idiosyncratic — host genetics, dose, immune response. Across five model families and every data lever, structure alone tops out near 0.75. Anima tells you which is which, honestly.
Anima ships two versioned, provenance-tracked, openly-licensed corpora — with dataset cards and cross-source quality control. Reuse them, extend them, build on them.
123,950 matched compounds — ChEMBL, TDC, PubChem qHTS, hERGCentral, BindingDB — 98.7% cross-source agreement on the IC50 core.
3,560 human liver-injury labels — FDA DILIrank/DILIst, LiverTox, curated literature — with graded severity, harmonized with provenance.
Sign in to run the Screen online — hosted, account-based access is rolling out in early access. Prefer local? Run the real tool today in two commands:
# build the model bundles (once)
python anima/train_models.py
# then open http://127.0.0.1:5000
python anima/app.pyAnima is a non-profit research prototype. This is a triage signal to prioritize testing — not a wet-lab replacement, not medical or clinical advice. Off-domain predictions are flagged low-confidence.