Graph Neural Network-Based Toxicity Prediction by Integrating Molecular Fingerprints and Knowledge Graph Features

Molecular toxicity prediction plays a crucial role in drug screening and environmental health risk assessment. Traditional toxicity prediction models primarily rely on molecular fingerprints and other structural features, while neglecting the complex biological mechanisms underlying compound toxicity, resulting in limited predictive accuracy, poor interpretability, and reduced generalizability. To address this challenge, this study proposes a novel molecular toxicity prediction framework that integrates knowledge graphs with Graph Neural Networks (GNNs). Specifically, we constructed a heterogeneous toxicological knowledge graph (ToxKG) based on ComptoxAI. ToxKG incorporates data from authoritative databases such as PubChem, Reactome, and ChEMBL, and covers multiple entities and relationships including chemicals, genes, signaling pathways, and bioassays. We then systematically evaluated six representative GNN models (GCN, GAT, R-GCN, HRAN, HGT, and GPS) on the Tox21 dataset. Experimental results demonstrate that heterogeneous graph models enriched with ToxKG information significantly outperform traditional models relying solely on structural features across multiple metrics including <i>AUC</i>, <i>F</i>1-score, <i>ACC</i>, and balanced accuracy (<i>BAC</i>). Notably, the GPS model achieved the highest <i>AUC</i> value (0.956) for key receptor tasks such as NR-AR, highlighting the critical role of biological mechanism information and heterogeneous graph structures in toxicity prediction. This study provides a promising pathway toward the development of interpretable and efficient intelligent models for toxicological risk assessment.