MedGraphNet: Leveraging Multi-Relational Graph Neural Networks and Text Knowledge for Biomedical Predictions

Area of Science:

• Computational Biology
• Bioinformatics
• Network Medicine

Background:

• Complex interactions between genetic, molecular, and environmental factors influence diseases.
• Current methods struggle to integrate diverse multi-relational biological data, hindering the discovery of novel risk genes and drugs.

Purpose of the Study:

• To develop a multi-relational Graph Neural Network (GNN) model, MedGraphNet, for inferring relationships among drugs, genes, diseases, and phenotypes.
• To leverage text knowledge embeddings for robust data integration and improved model generalizability.

Main Methods:

• Developed MedGraphNet, a multi-relational GNN model.
• Initialized nodes using informative embeddings from existing text knowledge.
• Integrated diverse biological data including drugs, genes, diseases, and phenotypes.

Main Results:

• MedGraphNet matches or outperforms traditional single-relation approaches, especially for isolated or sparsely connected nodes.
• Demonstrated generalizability to external datasets with high accuracy in identifying disease-gene and drug-phenotype associations.
• Successfully inferred drug side effects and identified relevant factors for Alzheimer's disease, validated by literature.

Conclusions:

• Integrating multi-relational data with text knowledge using MedGraphNet enhances biomedical predictions and facilitates drug repurposing.
• MedGraphNet offers a powerful tool for uncovering complex biological relationships and advancing precision medicine.
• The model's ability to infer novel associations highlights its potential for accelerating drug discovery and development.