Deep Learning for Protein-Protein Contact Prediction Using Evolutionary Scale Modeling (ESM) Feature

Protein-protein interactions (PPIs) are essential for various biological processes, and their binding sites provide important information for cell function and drug design. Traditional experimental methods for identifying these sites are expensive and time-consuming, prompting the emergence of computational forecasting tools. However, the performance of these tools tends to be limited due to single experimental training data and other limitations. We introduce a new hybrid deep neural network that fuses global sequence features with local influences, improving the traditional sequence sliding window strategy. Our method integrates a distance matrix between residue pairs to guide structural neighborhood screening, thereby increasing confidence in the results. In order to characterize the properties of proteins more comprehensively, we introduce protein language models into regular features, and design different attention mechanism models for feature learning in different dimensions. Experimental results show that our model performance has reached an advanced level and is ahead of other competing methods in several indicators.