Semantic knowledge improves molecular machine learning for chemical toxicity prediction

Quantitative structure-activity relationship (QSAR) modeling predicts chemical activity from structural and physico-chemical molecular properties. Beyond molecular structure, a growing body of semantic knowledge-often represented as knowledge graphs, in which biomedical entities are linked by interpretable relations reflecting scientific evidence-can inform computational toxicology. Here, we assess adding semantic knowledge to graph neural network (GNN)-based QSAR modeling to improve assay- and endpoint-level toxicity prediction, formulated as binary classification of active/inactive compounds. Both approaches we investigate use GNNs to learn molecular representations and augment them with semantic knowledge. Across publicly available in silico toxicity datasets, including Tox21 assays and external endpoints, semantic models improve discrimination while maintaining calibration comparable to structure-only QSAR models. Model explanations show that semantic knowledge helps the GNN encoder focus on substructures with known toxicity effects. Integrating mechanistic semantic context with molecular structure offers a route to more accurate and interpretable toxicity models that support chemical safety assessment.