Link Prediction Based on Subgraph Learning in Biological Networks

Area of Science:

• Computational Biology
• Network Science
• Bioinformatics

Background:

• Link prediction (LP) using graph neural networks (GNNs) is vital for understanding biological networks (BNs).
• Existing GNN-based LPs struggle with BN complexity, heterogeneity, dynamic changes, asymmetry, hierarchical modularity, and high computational costs.

Purpose of the Study:

• To propose a novel GNN-based LP model, termed LCS, to address the limitations of current methods in biological networks.
• To improve the accuracy, efficiency, and generalization performance of link prediction in complex biological networks.

Main Methods:

• LCS employs a subgraph-based GNN approach to manage BN heterogeneity, asymmetry, and modularity.
• A dynamic local subgraph extraction (SE) mechanism using heat kernel diffusion and the Chopper algorithm enables linear-time SE, mitigating subgraph explosion.
• Diversity regularization constraints are imposed to reduce computational complexity and enhance generalization.

Main Results:

• LCS demonstrates significant improvements over state-of-the-art LP methods on four biological network benchmarks.
• The model effectively addresses the challenges of complex and dynamic biological network structures.
• Enhanced efficiency and generalization performance were observed compared to existing approaches.

Conclusions:

• LCS offers a robust and efficient solution for link prediction in complex biological networks.
• The proposed subgraph-based GNN approach with dynamic SE mechanism advances the field of network biology.
• The open-source implementation facilitates further research and application.