ToxiGuard: an AOP-guided mechanistically interpretable framework for multi-organ toxicity prediction.

Accurate prediction of organ-specific toxicity with mechanistic interpretability remains a central challenge in chemical safety assessment and translational toxicology. Although animal-based assays provide biologically relevant information, they are low throughput and increasingly constrained by ethical and regulatory considerations. In parallel, data-driven computational approaches-particularly deep learning-have achieved strong predictive performance but often lack transparency, limiting their utility for mechanistic interpretation and regulatory decision-making. Here, we present ToxiGuard, a mechanistically informed deep learning framework that embeds organ-specific Adverse Outcome Pathway (AOP) structures directly into the model architecture. By explicitly encoding Molecular Initiating Event-Key Event-Adverse Outcome (MIE-KE-AO) connectivity, ToxiGuard constrains representation learning to biologically plausible causal pathways rather than relying solely on post hoc explanation. The framework integrates molecular descriptors, functional-class fingerprints, and curated AOP networks to enable complementary interpretability at both chemical and biological levels. Across four organ toxicity endpoints-hepatotoxicity, cardiotoxicity, nephrotoxicity, and respiratory toxicity-ToxiGuard demonstrates robust and consistent predictive performance, outperforming traditional machine learning models and AOP-agnostic neural network baselines. SHAP-based analyses reveal concentrated contributions from specific physicochemical properties, molecular substructures, and mechanistic AOP components, including CAR/PXR-associated pathways in hepatotoxicity and hERG-related mechanisms in cardiotoxicity. These findings are consistent with established toxicological knowledge and support the biological plausibility of model predictions. Overall, this study demonstrates that embedding mechanistic toxicological structure into deep learning architectures can enhance both predictive reliability and interpretability. ToxiGuard provides a transparent, AOP-anchored computational approach that complements existing experimental and in silico methods for early-stage organ toxicity assessment and chemical safety evaluation.