Residue-Frustration-Based Prediction of Protein-Protein Interactions Using Machine Learning

Area of Science:

• Computational biology
• Biophysics
• Machine learning in bioinformatics

Background:

• Protein-protein interactions (PPIs) are crucial for cellular functions.
• Experimental investigation of transient PPIs remains challenging.
• Computational prediction of PPIs is gaining importance due to data limitations.

Purpose of the Study:

• To explore the application of frustration, a statistical potential, for predicting protein-protein interactions.
• To develop a novel feature, highly frustrated density, for PPI prediction.
• To integrate frustration-based features with machine learning for enhanced PPI prediction.

Main Methods:

• Calculated residue pair frustration index based on stabilization energy.
• Derived highly frustrated density as a residue-frustration-based feature.
• Combined frustration-based and structure-based features with a long short-term memory (LSTM) neural network for prediction.

Main Results:

• The model achieved 75% accuracy in predicting dimers using top 2‰ residue pairs.
• Frustration-based features demonstrated significant improvement over traditional chemical features.
• Highly frustrated density effectively captures residue propensity for PPI involvement.

Conclusions:

• Frustration is a potent statistical potential for PPI prediction.
• Frustration-based features offer a viable alternative to chemical features in machine learning models for PPI.
• Statistical potentials like frustration hold significant promise for advancing PPI prediction methodologies.