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

Area of Science:

• Computational chemistry
• Toxicology
• Artificial intelligence in drug discovery

Background:

• Early identification of toxic liabilities is crucial for drug safety and reducing development attrition.
• Current structure-based deep learning models struggle with generalization to new chemical scaffolds and lack biological mechanistic insights.

Purpose of the Study:

• To develop an advanced AI framework, Toxicological Knowledge-guided Graph-based Learning (TKGL), for more accurate, interpretable, and generalizable toxicity prediction.
• To integrate molecular graph information with substructural alerts and biological mechanisms for enhanced toxicological assessment.

Main Methods:

• TKGL employs a multiview framework integrating molecular graphs with substructural alerts and biological mechanisms.
• An attention mechanism is utilized for interpretable prioritization of toxicity-relevant substructures and biological events.
• Contrastive learning aligns chemical and biological information spaces, while domain adaptation improves generalization on scaffold-dissimilar data.

Main Results:

• TKGL demonstrated superior performance across 37 toxicity datasets compared to 11 state-of-the-art models.
• Interpretability analyses confirmed TKGL's identification of mechanistically relevant features.
• Two potential hepatotoxicity pathways involving p53 and PPARg were discovered.

Conclusions:

• TKGL represents a new paradigm in toxicity prediction, unifying chemical structure and biological behavior.
• The framework offers more accurate, interpretable, and generalizable chemical safety assessments.
• TKGL enhances the understanding of drug-induced toxicity through mechanistic insights.