Pitfalls of machine learning models for protein-protein interaction networks

Area of Science:

• Computational biology
• Bioinformatics
• Systems biology

Background:

• Protein-protein interactions (PPIs) are crucial for biological processes, development, and disease.
• Existing computational tools for PPI prediction lack standardized, reproducible frameworks, leading to inconsistent results.
• Discrepancies between different algorithms for in silico PPI prediction remain unexplained.

Purpose of the Study:

• To develop an open-source benchmarking framework for evaluating computational PPI prediction models.
• To investigate the impact of network topology and protein connectivity on PPI prediction algorithms.
• To compare the performance of functional genomics-based and sequence-based models for PPI prediction.

Main Methods:

• Designed and implemented an open-source benchmarking framework addressing biological and statistical challenges.
• Analyzed human PPI data using both functional genomics and sequence-based models.
• Evaluated algorithm performance concerning network topology and protein hub status.
• Replicated analyses using data from human and S. cerevisiae.

Main Results:

• The developed framework facilitates reproducible benchmarking of PPI prediction models.
• Functional genomics models excel at predicting interactions involving non-hub proteins, while sequence-based models are better for hub proteins.
• Algorithm design has minimal impact on performance when using functional genomics data.
• Functional genomics-based models demonstrate superior PPI prediction capabilities across species.

Conclusions:

• The study provides a principled foundation for constructing, comparing, and applying PPI networks.
• The developed framework enhances the reliability and comparability of computational PPI prediction.
• Functional genomics data offers a robust approach for cross-species PPI prediction.