Improving protein-protein interaction site prediction using deep residual neural network.

Area of Science:

• Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• Protein-protein interactions (PPIs) are crucial for biological mechanisms and drug discovery.
• Experimental identification of PPI sites is costly and time-consuming.
• Computational methods offer a faster alternative for PPI site prediction.

Purpose of the Study:

• To develop a novel deep learning-based method, D-PPIsite, for enhanced sequence-based PPI site prediction.
• To improve the accuracy and efficiency of identifying protein-protein interaction sites.

Main Methods:

• Utilized four sequence-driven features: position-specific scoring matrix, relative solvent accessibility, position information, and physical properties.
• Employed a deep learning module with convolutional, squeeze and excitation, and fully connected layers.
• Integrated multiple prediction models using a mean ensemble strategy to prevent local optima.

Main Results:

• D-PPIsite achieved an average accuracy of 80.2% and precision of 36.9% on five independent datasets.
• The method covered 53.5% of all PPI sites.
• Demonstrated a significantly higher average Matthews correlation coefficient (0.330) compared to existing methods.

Conclusions:

• D-PPIsite offers a highly accurate and efficient computational approach for predicting protein-protein interaction sites.
• The developed predictor can accelerate PPI-related research and aid in drug development.
• A standalone version of D-PPIsite is available for academic use.