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

Area of Science:

• Computational toxicology
• cheminformatics
• Deep learning in drug discovery

Background:

• Traditional animal testing for chemical safety is limited by throughput and ethical concerns.
• Current deep learning models offer predictive power but lack mechanistic interpretability for regulatory use.
• Bridging the gap between predictive performance and biological understanding is crucial for chemical safety.

Purpose of the Study:

• To develop a deep learning framework, ToxiGuard, that embeds mechanistic toxicological knowledge, specifically Adverse Outcome Pathways (AOPs), into its architecture.
• To enhance the transparency and mechanistic interpretability of computational toxicity prediction models.
• To improve the accuracy and reliability of organ-specific toxicity predictions for chemical safety assessment.

Main Methods:

• Developed ToxiGuard, a deep learning framework integrating organ-specific AOP structures (Molecular Initiating Event-Key Event-Adverse Outcome connectivity) into the model architecture.
• Integrated molecular descriptors, functional-class fingerprints, and curated AOP networks for multi-level interpretability.
• Validated the framework across four organ toxicity endpoints: hepatotoxicity, cardiotoxicity, nephrotoxicity, and respiratory toxicity.

Main Results:

• ToxiGuard achieved robust and consistent predictive performance across multiple organ toxicity endpoints, outperforming traditional machine learning and AOP-agnostic deep learning models.
• SHAP analyses identified key contributions from physicochemical properties, molecular substructures, and specific AOP components (e.g., CAR/PXR for hepatotoxicity, hERG for cardiotoxicity).
• Model predictions were consistent with established toxicological knowledge, supporting biological plausibility.

Conclusions:

• Embedding mechanistic toxicological structure, such as AOPs, into deep learning models significantly enhances both predictive reliability and interpretability.
• ToxiGuard offers a transparent, AOP-anchored computational approach that complements existing experimental and in silico methods for early-stage toxicity assessment.
• This framework supports informed regulatory decision-making in chemical safety evaluation by providing mechanistically interpretable predictions.