Supervised learning is an accurate method for network-based gene classification.

Area of Science:

• Genetics
• Computational Biology
• Bioinformatics

Background:

• Assigning human genes to functions, diseases, and traits is a major challenge in genetics.
• Computational methods, including supervised learning and label propagation, leverage molecular interaction networks for gene attribute prediction.
• Supervised learning's broad performance across networks and gene classification tasks, compared to label propagation, remains largely unexamined.

Purpose of the Study:

• To comprehensively benchmark supervised learning for network-based gene classification.
• To evaluate supervised learning against label propagation and node embeddings across diverse prediction tasks and networks.
• To determine the efficacy of supervised learning for prioritizing genes associated with biological functions, diseases, and traits.

Main Methods:

• Benchmarking supervised learning and label propagation on hundreds of diverse gene classification tasks.
• Utilizing multiple molecular interaction networks for evaluation.
• Comparing performance using stringent evaluation schemes and against node embeddings derived from node2vec.

Main Results:

• Supervised learning on a gene's full network connectivity significantly outperforms label propagation.
• Supervised learning achieves high prediction accuracy by effectively capturing local network properties.
• Supervised learning on the full network is superior to using node embeddings for gene classification.

Conclusions:

• Supervised learning is a highly accurate and robust approach for network-based gene classification.
• The method efficiently prioritizes genes linked to diverse functions, diseases, and traits.
• Supervised learning should be integrated as a standard tool in network-based gene classification workflows.