Node similarity-based graph convolution for link prediction in biological networks.

Area of Science:

• Network biology
• Graph representation learning
• Bioinformatics

Background:

• Link prediction is crucial in network biology.
• Graph Convolutional Networks (GCNs) are increasingly used for link prediction via node embedding.
• Existing GCN methods often use degree-normalized adjacency matrices for feature propagation.

Purpose of the Study:

• To develop an improved GCN-based link prediction method using node similarity.
• To evaluate the effectiveness of various node similarity measures in GCN convolution matrices.
• To enhance the prediction of biological associations and interactions.

Main Methods:

• Proposed using node similarity-based convolution matrices in GCNs.
• Evaluated eight node similarity measures: Common Neighbors, Jaccard Index, Adamic-Adar, Resource Allocation, Hub-Depressed Index, Hub-Promoted Index, Sorenson Index, and Salton Index.
• Compared performance against standard GCNs and other link prediction algorithms on biomedical networks.

Main Results:

• Node similarity-based convolution matrices significantly improved GCN-based link prediction performance.
• Demonstrated enhanced accuracy in predicting drug-disease associations, drug-drug interactions, and protein-protein interactions.
• The proposed method, SiGraC, offers a robust alternative to traditional GCNs for biological network analysis.

Conclusions:

• Node similarity-based convolution matrices are effective for GCN-based link prediction in network biology.
• This approach offers a valuable enhancement for sophisticated machine-learning applications in biology.
• The SiGraC library provides an accessible implementation for researchers.