CrossPPI: A Cross-Fusion Based Model for Protein–Protein Binding Affinity Prediction

Many biological processes depend on protein–protein interactions (PPIs), which are particularly important across biology, medicine, and biotechnology. It is essential to accurately predict the binding affinity between protein pairs to prioritize candidate interactions in large-scale studies and expedite drug discovery. The application of cross-attention mechanisms between ligand and receptor protein sequences is often neglected in current computational models, limiting their capacity to accurately represent inter-protein dependencies. In this study, we introduce CrossPPI, a novel deep learning framework that integrates structural and sequential features of interacting proteins to improve binding affinity prediction. To model intricate interactions between protein pairs, CrossPPI uses a transformer-based cross-fusion module and a dual-view feature-extraction approach that combines Graph Attention Networks (GATs) and Convolutional Neural Networks (CNNs).

On the test dataset of 300 protein–protein pairs, CrossPPI achieved a Pearson correlation coefficient (PCC) of 0.7616, a Spearman correlation coefficient (SCC) of 0.7644, a mean absolute error (MAE) of 1.2869, and a root mean square error (RMSE) of 1.6824, indicating its ability to predict the binding affinity of two proteins.

The results highlight CrossPPI’s capability to predict inter-protein binding affinities by leveraging an attention-based integration of sequence and structural features.