The probability of edge existence due to node degree: a baseline for network-based predictions

Area of Science:

• Computational Biology
• Network Science
• Bioinformatics

Background:

• Biomedical discovery tasks like gene function prediction and drug repurposing rely on network edge prediction.
• Node degree distribution varies across biological networks, potentially biasing predictions.

Purpose of the Study:

• To introduce a network permutation framework to quantify the influence of node degree on edge prediction accuracy.
• To differentiate performance contributions from node degree versus specific network connections.

Main Methods:

• Developed a network permutation framework to isolate the effect of node degree.
• Generated degree-only features to assess baseline prediction performance.
• Decomposed overall performance into degree-attributable and network-specific components.

Main Results:

• Node degree often accounts for a substantial portion of prediction performance in biomedical networks.
• Performance attributable to specific network connections is frequently a minor component.
• The influence of node degree can lead to nonspecific or misleading predictions.

Conclusions:

• Researchers must account for node degree bias when interpreting edge prediction results.
• The proposed permutation approach provides a crucial baseline for evaluating prediction specificity.
• Open-source Python package `xswap` is available for implementing these methods.