BiGvCL: bipartite graph-based cross-domain contrastive learning model for the predicting drug-gene interactions

Area of Science:

• Bioinformatics
• Computational Biology
• Pharmacogenomics

Background:

• Drug-gene interactions (DGIs) are crucial for drug efficacy, toxicity, and understanding drug mechanisms.
• Current computational methods for DGI prediction often require explicit chemical or genetic features, limiting their application to novel or unannotated entities.
• There is a need for DGI prediction methods that can generalize to new drugs and genes based on broader network information.

Purpose of the Study:

• To develop and evaluate BiGvCL, a novel framework for predicting drug-gene interactions (DGIs) solely based on network topology.
• To demonstrate the effectiveness of topology-based DGI prediction for facilitating precision medicine and drug discovery.
• To assess the generalizability and biological relevance of predictions made by the BiGvCL framework.

Main Methods:

• Proposed BiGvCL framework utilizing a lightweight graph attention mechanism (GATLite) for local neighborhood aggregation.
• Employed a gated graph convolutional network (GatedGCN) to learn high-order drug-gene interactions.
• Integrated contrastive learning to improve model generalizability and performance.
• Validated the framework on DrugBank and DGIdb datasets, and performed cross-domain evaluations on OGB datasets.

Main Results:

• BiGvCL achieved competitive performance across multiple metrics compared to existing baseline methods on established DGI datasets.
• Cross-domain evaluations confirmed BiGvCL's adaptability to diverse biomedical networks.
• Ablation studies highlighted the significant contributions of the contrastive and gated mechanisms within the framework.
• Case studies and molecular docking provided evidence for the biological relevance of predicted DGIs.

Conclusions:

• BiGvCL demonstrates the potential of topology-based approaches for discovering novel drug-gene interactions, even without explicit feature data.
• The framework shows promise for advancing precision medicine and informing drug repurposing strategies.
• While reliant on network topology and transductive learning, BiGvCL offers a valuable alternative for DGI prediction in scenarios with limited feature information.