Structural feature-based machine learning benchmarking for protein interface prediction.

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in biochemistry

Background:

• Accurate prediction of protein-protein interaction (PPI) interfaces is crucial for understanding molecular mechanisms and developing targeted therapies.
• Existing methods may lack efficiency or interpretability for large-scale applications.

Purpose of the Study:

• To develop and validate a machine learning pipeline for predicting interface residues in protein complexes.
• To identify key features and optimal algorithms for accurate and efficient interface prediction.
• To assess the impact of structural context on prediction performance.

Main Methods:

• Benchmarking six machine learning algorithms on a dataset of 1311 homodimers.
• Employing recursive feature elimination to identify a minimal set of predictive features.
• Developing structurally stratified models for different protein types (α-helical, β-strand, membrane).
• Comparing model performance against established pipelines like ColabFold.

Main Results:

• Multilayer perceptron and XGBoost achieved high predictive accuracy (Matthews correlation coefficient > 0.93).
• A reduced set of six biologically meaningful features retained strong predictive power (MCC > 0.90).
• Structurally specialized models showed comparable or improved accuracy, especially with the reduced feature set.
• Feature-driven models performed competitively against ColabFold.

Conclusions:

• Compact, structure-aware machine learning models offer a scalable, interpretable, and biologically informed approach to protein interface prediction.
• Structural context is vital for accurate interface prediction, and specialized models can enhance performance.
• This pipeline has implications for large-scale structural analysis, drug target identification, and protein engineering.