Topological-Similarity Based Canonical Representations for Biological Link Prediction

Area of Science:

• Systems Biology
• Network Science
• Machine Learning

Background:

• Graph machine learning is widely used for prediction tasks in systems biology.
• Node topological similarity is crucial for designing graph convolution and loss functions.
• Existing similarity measures significantly impact model performance and reliability.

Purpose of the Study:

• To propose GraphCan, a framework for computing canonical biological network representations.
• To integrate multiple node similarity measures for enhanced network analysis.
• To improve the robustness and performance of graph machine learning models in systems biology.

Main Methods:

• Developed GraphCan, a similarity-based Graph Convolutional Network (GCN) framework.
• Integrated eight diverse node similarity measures: Common Neighbor, Adamic Adar, Random Walk with Restart, Von Neumann, Resource Allocation, Hub-Depressed Index, Hub-Promoted Index, and adjacency matrix.
• Evaluated GraphCan using link prediction tasks in systems biology.

Main Results:

• GraphCan computes canonical node embeddings by integrating multiple similarity measures.
• The framework demonstrates improved robustness, especially on sparse biological networks.
• Integrated similarity measures enhance the reliability of graph machine learning models.

Conclusions:

• GraphCan provides a robust method for generating canonical representations of biological networks.
• Integrating multiple topological similarity measures is key to improving graph machine learning performance.
• The GraphCan framework offers a valuable tool for systems biology research and prediction tasks.