Enhancing link prediction in biomedical knowledge graphs with BioPathNet.

Area of Science:

• Biomedical informatics
• Network biology
• Machine learning in healthcare

Background:

• Biomedical network analysis is vital for progress, but traditional link prediction methods struggle with complexity.
• Existing representation-based learning methods have limitations in capturing intricate biological relationships.

Purpose of the Study:

• To introduce BioPathNet, a graph neural network framework utilizing path-based reasoning for improved link prediction in biomedical knowledge graphs.
• To overcome limitations of node-embedding methods by considering all relations along paths for enhanced accuracy and interpretability.

Main Methods:

• BioPathNet employs a neural Bellman-Ford network (NBFNet) for path-based reasoning in link prediction.
• It incorporates a background regulatory graph for advanced message passing and uses stringent negative sampling for precision and scalability.

Main Results:

• BioPathNet demonstrates superior or comparable performance across gene function annotation, drug-disease indication, synthetic lethality, and lncRNA-target interaction prediction.
• The framework successfully identified potential new drug indications for diseases like acute lymphoblastic leukaemia and Alzheimer's disease, with expert and clinical validation.
• It also prioritized synthetic lethal gene pairs and regulatory lncRNA-target interactions.

Conclusions:

• BioPathNet offers enhanced accuracy and interpretability in biomedical link prediction through path-based analysis.
• The framework's ability to visualize influential paths provides valuable molecular insights for researchers.
• BioPathNet facilitates biological validation and accelerates discovery in areas such as drug repurposing and gene interaction studies.