MechTox: an audited benchmark for ordinal drug-induced liver injury prediction with uncertainty and mechanism-oriented interpretation

Drug-induced liver injury (DILI) predictions are sensitive to cohort definition, molecular representation, model selection, and uncertainty assumptions. We developed MechTox as an audited benchmark/workflow for ordinal DILI concern prediction and reserve MechTox-NN for the dual-branch neural model. From a pool of 1,336 DILIrank source records, we resolved 1,216 compound identities, where 1,081 records had usable standardized structures, 674 had verified transcriptomic profiles, and 672 were dual-modal records; the primary three-class cohort comprised 512 compounds (147 vNo, 234 vLess, and 131 vMost), and 160 accepted ambiguous records were retained separately. The primary performance was evaluated through five repeats of five outer scaffold folds and five grouped inner folds. The pooled outer quadratic-weighted kappa (QWK) was 0.287 for Morgan fingerprint logistic regression, 0.232 for the ordinal RDKit descriptor model, 0.058 for late-average structure–pathway fusion, and 0.252 for MechTox-NN. A separate compound-bootstrap sensitivity based on repeat-averaged predictions resulted in a median QWK of 0.296 [0.207, 0.377] for MechTox-NN and was not treated as the confidence interval for the pooled estimate. The canonical single-sample gene-set enrichment analysis served as the primary pathway representation, and a custom signed rank-integral score was evaluated only as a sensitivity comparator. Source-shift cohorts and empirical prediction-set containment were used to characterize the behaviors under source composition changes and were interpreted descriptively. In the held-out permutation-calibrated mechanism analysis, cationic amine dealkylation alone met the family-wise support criterion; this finding was interpreted as an exploratory association. Thus, MechTox provides a documented uncertainty-aware framework for protocol-specific ordinal DILI benchmarking and mechanism-oriented hypothesis generation.