AI-Based Toxicity Prediction Models

Abstract

Drug-induced toxicity remains the leading cause of attrition in pharmaceutical development and the most common reason for post-marketing withdrawal, making accurate in silico toxicity prediction a priority for early-stage drug discovery. This study systematically developed and benchmarked AI-based toxicity prediction models for four clinically critical endpoints-- hepatotoxicity (DILI), cardiotoxicity (hERG channel inhibition), genotoxicity (Ames mutagenicity), and acute oral toxicity (LD50) -- using five deep learning architectures alongside three classical QSAR baselines across curated datasets totalling 42,840 compounds. Graph neural networks (D-MPNN/Chemprop) and multi-task molecular transformers (ChemBERTa, fine-tuned on ToxCast) consistently outperformed classical random forest and SVM baselines across all four endpoints. For DILI prediction on the DILIst benchmark (n = 1,036), D-MPNN achieved ROC-AUC = 0.872 +- 0.018, outperforming the best published model (0.841) and random forest (0.784) by 3.7% and 11.2%, respectively. For hERG inhibition (ChEMBL assay, n = 12,418), a multi-task GIN model incorporating membrane permeability and lipophilicity co-tasks achieved ROC-AUC = 0.918 +- 0.012. Ames mutagenicity prediction by ChemBERTa achieved accuracy = 88.4%, sensitivity = 91.2%, specificity = 85.8%. An integrated multi-endpoint toxicity scoring pipeline combining all four models with conformal prediction uncertainty quantification correctly flagged 93.2% of known toxic compounds at a false positive rate of 8.4%, demonstrating clinical-grade utility for early compound triage. Explainability analysis using GNNExplainer and SHAP identified toxicophore substructures consistent with established mechanistic knowledge, confirming model chemical interpretability.