How ToxiPred works
A self-contained, CPU-only pathway-bottleneck neural model for predicting drug cytotoxicity across 24 human tissue lineages and the dominant toxicity mechanism.
What it predicts
For every (drug × tissue) pair, ToxiPred returns a Drug Sensitivity Score (DSS) on the 0-50 scale used by PharmacoDB, and - for hits that pass the toxicity gate (DSSmean > 5.0) - the top-3 pathway contributions and a labelled mechanism family. Inhibition at fixed anchor doses (0.1 µM, 1 µM, 10 µM) is also tracked internally and used to assign each drug to a dose-response class.
Atlas vs neural inference
Each query is canonicalized via RDKit (with stereochemistry preserved) and matched against the 53,720-drug atlas. Exact matches return measured PharmacoDB DSS values (atlas). Novel chemistry is passed through a DMPNN-based pathway-feature imputer, blended with Tanimoto-nearest atlas neighbors, and scored by a 5-seed pathway-bottleneck NN ensemble (NN · τ x) - where the Tanimoto similarity τ to the nearest atlas neighbor is reported as a confidence indicator.
Mechanism scoring rule
For each predicted hit, the drug-tissue interaction is decomposed into Hadamard pathway contributions. The drug-specific deviation from the atlas baseline (Δ × √SD), combined with the negative-median p-value, defines a score per pathway. The top 3 pathways are reported together with the dominant mechanistic family (Apoptosis, Stress/UPR, Cell-cycle, Lipid/Sterol, Metabolism, DNA-repair, Other).
Held-out validation
Public toxicity benchmarks are used only for evaluation, never for training. On a held-out subset, ToxiPred reaches AUROC 0.866 (liver apoptosis), 0.812 (liver steatosis), 0.804 (liver DNA-damage), and 0.710 (ClinTox CT_TOX) - all from the same per-tissue DSS predictions.
What it is not
ToxiPred predicts tissue cellular liability and dose-anchored inhibition - not direct clinical adverse-event probability. Predictions on novel chemistry are mechanistic hypotheses, not clinical truth. The model is intended for prioritisation, mechanism exploration, and triage, not for regulatory decisions.