Transient protein-protein interface prediction: datasets, features, algorithms, and the RAD-T predictor.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in drug discovery

Background:

• Transient protein-protein interactions (PPIs) are fundamental to biological processes and therapeutic targets.
• Existing prediction methods like docking and sequence analysis have limitations regarding structural data and pattern recognition.
• Machine learning (ML) offers a promising approach to generalize interface sites using descriptive features, but best practices are still evolving.

Purpose of the Study:

• To analyze the efficacy of ML-based PPI predictors.
• To identify areas for improvement in ML-based PPI prediction.
• To develop a novel ML-based PPI predictor.

Main Methods:

• Investigated the impact of unknown interaction sites on prediction accuracy.
• Improved data labeling by enforcing higher interface site rates per domain.
• Developed and evaluated a new ML predictor, RAD-T.
• Identified key features for predicting protein interfaces.

Main Results:

• Enhancing training data accuracy by 44% improved ML algorithm performance.
• Identified seven features with the highest predictive power across datasets.
• The new predictor, RAD-T, demonstrated a 59% increase in MCC score compared to existing methods.
• A consistent set of 10 biologically unrelated proteins showed high prediction accuracy.

Conclusions:

• Current evaluation methods may underestimate ML-based PPI predictor performance due to unidentified interaction sites.
• Improving training datasets is crucial for advancing ML-based interface prediction.
• Larger test sets of well-studied proteins or domain-specific scoring are needed to address poor interaction site identification.