Evaluation of different biological data and computational classification methods for use in protein interaction...

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Protein-protein interactions (PPIs) are fundamental to biological processes.
• High-throughput methods for PPI detection in yeast are often incomplete and error-prone.
• Supervised learning offers a promising approach to integrate diverse biological data for PPI prediction.

Purpose of the Study:

• To systematically evaluate the utility of various biological data sources and feature encodings for PPI prediction.
• To compare the performance of different classifiers across three distinct PPI tasks: physical interaction, co-complex, and pathway co-membership.
• To identify the most influential features for accurate PPI prediction.

Main Methods:

• Assembled a comprehensive set of biological features and varied their encoding.
• Employed six classifiers: Random Forest (RF), RF k-NN, Naïve Bayes, Decision Tree, Logistic Regression, and Support Vector Machine.
• Utilized RF's Gini index for feature importance assessment and evaluated accuracy with top-ranked features.

Main Results:

• Co-complex relationship prediction was generally easier than physical interaction or pathway co-membership.
• The Random Forest classifier consistently performed as a top-tier model across all feature sets and tasks.
• Gene expression emerged as the most critical feature for all prediction tasks, irrespective of encoding.
• Yeast-2-hybrid data was not a top-ranking feature under any tested condition.

Conclusions:

• Supervised learning, particularly with Random Forest, is effective for PPI prediction.
• Feature importance is task-dependent and influenced by data encoding.
• Gene expression data holds significant predictive power for various types of protein interactions.