Quantitative Toxicophore Discovery through Interpretable Deep Learning: A Framework for Mechanism-Guided Toxicity Mitigation in Drug Safety Profiling

Abstract Unanticipated toxicity accounts for ∼30% of clinical-phase drug failures, yet despite advances in computational toxicity prediction, predictive models typically offer scores without identifying the structural culprit. Here, we introduce an interpretable deep-learning framework that couples fingerprint-descriptor fusion with task-adaptive attention and a SHAP-guided bioisosteric ablation pipeline. Trained on 9,315 compounds spanning 14 end points (Tox21 + ClinTox) under scaffold-stratified split, the attention-fusion model achieves mean test ROC-AUC of 0.739 (Tox21) and 0.953 (ClinTox). The critical advance is a structure-guided validation protocol: predictive intervention ablation─computationally replacing toxicophores with bioisosteric equivalents─demonstrates that the model’s predictions are predictively dependent on specific substructures. Halogen → methyl substitution drops predicted toxicity from 0.659 to 0.101 (paired t = 22.66, p = 6.32 × 10–10, n = 11; Cohen’s d = 6.82), providing the primary structure-guided evidence. Cross-end point SHAP correlation reaches 0.333 (MTL) and 0.334 (STL; Welch’s p = 0.82), indicating that both paradigms extract coherent mechanistic structure above the label baseline (0.172). SMARTS-based toxicophore alignment identifies a strong positive logit differential for aniline on the clinically relevant CT_TOX end point (+0.76), and predictive-intervention ablation confirms model dependence on the identified substructures (halogen → methyl: n = 11, p < 10–9; aniline → methoxy: n = 2, directional check only). External validation on 1,427 SIDER adverse-drug-reaction compounds reveals significant toxicophore enrichment (aniline: 12.6% vs 0%, Fisher exact p = 4.5 × 10–12; halogenated aromatic: 29.8% vs 0%, p = 1.6 × 10–29), remaining significant after controlling for molecular weight and Log P. Three case studies illustrate how SHAP-guided attribution translates into actionable medicinal-chemistry mitigation strategies.