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

Area of Science:

• Computational toxicology
• Cheminformatics
• Bioinformatics

Background:

• Traditional molecular toxicity prediction models lack accuracy and interpretability due to limited biological mechanism consideration.
• Existing methods often rely solely on molecular structural features, hindering generalizability.

Purpose of the Study:

• To develop a novel molecular toxicity prediction framework integrating knowledge graphs and Graph Neural Networks (GNNs).
• To construct a heterogeneous toxicological knowledge graph (ToxKG) incorporating diverse biological data.
• To evaluate the performance of various GNN models using the Tox21 dataset.

Main Methods:

• Construction of a heterogeneous toxicological knowledge graph (ToxKG) using ComptoxAI, integrating data from PubChem, Reactome, and ChEMBL.
• Systematic evaluation of six GNN models (GCN, GAT, R-GCN, HRAN, HGT, GPS) on the Tox21 dataset.
• Comparison of GNN performance with and without ToxKG enrichment.

Main Results:

• Heterogeneous graph models enriched with ToxKG information significantly outperformed traditional structure-based models.
• Key metrics including AUC, F1-score, ACC, and BAC showed substantial improvements.
• The GPS model achieved a top AUC of 0.956 for NR-AR receptor tasks, demonstrating the value of biological mechanisms.

Conclusions:

• Integrating biological mechanisms via heterogeneous knowledge graphs and GNNs is crucial for accurate molecular toxicity prediction.
• This framework offers a promising direction for developing interpretable and efficient intelligent toxicological risk assessment tools.