An Interpretable Machine‐Learning Framework for Multi‐Endpoint Drug‐Induced Liver Injury Classification: Evaluating the Contribution of Molecular Structure

ABSTRACT Drug‐induced liver injury (DILI) contributes to drug development failure and post‐marketing withdrawal, highlighting the need for reliable computational safety assessment. We present an interpretable AI framework for multi‐endpoint DILI classification using 801 drugs curated from LiverTox and related resources. Pipeline A evaluated phenotypic features, 1199 molecular structural descriptors, and their combinations across five endpoints: injury mechanism, pattern, severity, DILI risk, and clinical outcome. Models were assessed using repeated stratified and scaffold‐disjoint fivefold cross‐validation. Phenotypic models achieved macro‐AUCs of 0.660, 0.678, 0.696, 0.877, and 0.676, respectively, under repeated stratified validation. Structural augmentation yielded modest, endpoint‐dependent improvements for T1 under stratified validation and T5 under scaffold‐disjoint evaluation, whereas phenotypic features outperformed combined representations for T4. Pipeline B incorporated SMARTS toxicophores, molecular fingerprints, atom‐level importance mapping, pharmacophores, Murcko scaffolds, and activity‐cliff analysis. SHAP identified contributions from molecular fingerprints and graph‐topology descriptors alongside phenotypic features. However, toxicophore and scaffold associations were not statistically significant after multiple‐testing correction. Overall, the findings demonstrate the selective, complementary contributions of molecular structure to retrospective DILI classification and establish an interpretable framework for further validation in drug‐safety assessment.