Toxicological Knowledge-Guided Graph Learning for Interpretable and Generalizable Molecular Toxicity Prediction.

Abstract

Early identification of toxic liabilities is essential for improving drug safety and reducing attrition. However, structure-based deep learning models often generalize poorly to new chemical scaffolds and provide limited biological mechanistic insight. To address this, we introduce Toxicological Knowledge-guided Graph-based Learning (TKGL), a multiview framework that integrates molecular graphs with substructural alerts and biological mechanism information. TKGL leverages an attention mechanism for interpretable prioritization of toxicity-relevant substructures and biological events. Contrastive learning further aligns heterogeneous chemical and biological information spaces, while domain adaptation robustly enhances generalization under stringent scaffold-dissimilar splits. Across 37 toxicity data sets, TKGL performs favorably against 11 state-of-the-art models. Interpretability analyses confirm that TKGL identifies mechanistically meaningful features, exemplified by the discovery of two potential hepatotoxicity pathways involving p53 and PPARg. Overall, TKGL establishes a new paradigm for toxicity prediction, unifying chemical structure and biological behavior for more accurate, interpretable, and generalizable chemical safety assessment.