Mechanism-Aware Graph Neural Network for Organ-Specific Drug Toxicity Prediction with Explainable Molecular and Pathway Attribution.

Predicting drug-induced toxicity remains a central challenge in computational toxicology, particularly for organ-specific adverse effects that arise from diverse structural, biochemical, and mechanistic origins. Existing deep learning models excel at pattern recognition but often lack mechanistic interpretability, while mechanistic rule-based frameworks offer interpretability but limited predictive breadth. To bridge this divide, we developed a mechanism-aware deep learning framework that integrates Graph Neural Network (GNN) structural encodings with biologically curated mechanistic pathway features to predict toxicity across five clinically relevant endpoints: hepatotoxicity, nephrotoxicity, cardiotoxicity, mitochondrial toxicity, and hERG/QT risk. A curated dataset of 50 mechanistically diverse compounds was assembled to evaluate the model. The GNN component learns atom-level and bond-level representations from molecular graphs, while mechanistic features encode known biological drivers of toxicity such as CYP-mediated bioactivation, mitochondrial disruption, transporter interactions, electrophilicity, and hERG binding risk. The hybrid model jointly performs binary organ-specific toxicity classification and continuous toxicity-severity regression. Training curves demonstrated stable learning dynamics, with toxicity classification loss decreasing from ~0.69 to ~0.30 and severity regression loss decreasing from >2.0 to <0.40 over 50 epochs. Organ-specific validation AUROC values were high for hepatotoxicity, cardiotoxicity, mitochondrial toxicity, and hERG/QT risk (typically 0.85–1.00), with nephrotoxicity showing lower but biologically consistent performance due to pathway heterogeneity. Model explainability was assessed using atom-level importance maps and mechanism-level attribution. Acetaminophen exhibited strong contributions from CYP2E1 bioactivation and reactive metabolite formation, faithfully recapitulating its known toxicity mechanism. In contrast, cisplatin nephrotoxicity was driven primarily by structural GNN features rather than pathway flags, consistent with the drug’s mechanism (Pt(II)-mediated DNA crosslinking and proximal tubule accumulation). Across the dataset, the model produced intuitive and biologically grounded mechanistic attributions while maintaining competitive predictive accuracy. Toxicity profiles highlighted clear clustering patterns, including mitochondrial disruptors (e.g., doxorubicin, isoniazid), hERG-risk compounds (e.g., cisapride), and strong hepatotoxins (e.g., valproate, diclofenac). This integration of structure-level deep learning and biology-based mechanistic features enables high-performance prediction while preserving interpretability necessary for risk assessment, drug design, and regulatory decision-making. This study demonstrates that mechanism-aware deep learning is a promising direction for translational toxicology, offering improved organ-specific prediction, mechanistic transparency, and practical utility. The framework is extensible to larger datasets, additional endpoints, and real-world drug discovery pipelines, positioning it as a viable foundation for next-generation computational safety evaluation.