SilicoTox: interpretable machine learning for molecular hazard prioritization, with analysis of database provenance bias and benchmarking across nine defined toxicity endpoints

Toxic liabilities cause costly late-stage attrition in drug discovery, and ligand-based machine learning is widely used to flag them early, but models are usually compared on headline accuracy alone, which obscures what that accuracy measures. Our contribution is not a new predictor but a controlled separation of three contributions to accuracy: molecular representation, learning algorithm and training-data provenance. One pipeline (2,071 features: Morgan fingerprint, twenty descriptors and pan-assay interference (PAINS), Brenk and National Institutes of Health (NIH) alert counts; scaffold-based partitioning; calibration and thresholding confined to validation data) was applied to two model classes. The provenance model separates hazard-annotated compounds from approved therapeutics (8,545 molecules, four public databases). Its label encodes database provenance, not a toxicological endpoint, so it demonstrates provenance bias and supports hazard prioritization only. Molecular size, its leading Shapley contributor, separates the sources (280 versus 358 Da); we bound this at about 0.03 in the area under the curve (AUC). On an external comparison set (838 molecules, same databases, also used for model selection) it reached AUC 0.902, measuring recovery of provenance rather than toxicity prediction. Nine endpoint models retrained on protocol-defined datasets represent toxicity prediction (cross-validated AUC 0.630-0.858). Replacing the engineered representation with a path fingerprint cost 0.048 AUC in every contrast, significant after multiplicity correction in five of eight; exchanging tree ensembles changed performance by 0.002. Following OECD principles for (Q)SAR validation, we report balanced accuracy, Matthews correlation coefficient, confusion matrices, calibration, an applicability domain and atom-level attribution. Data are openly deposited.