Multimodal Learning of Protein–Protein Interactions for Accurate Binding Affinity Prediction

Protein–protein interactions are central mediators in many biological processes. Accurately predicting the protein–protein binding affinity is crucial for guiding the modulation of these interactions, thereby playing a significant role in therapeutic development and drug discovery. However, existing methods often rely solely on sequence information or oversimplified structural representations, overlooking critical details such as side-chain interactions. In this work, we introduce a multimodal framework that integrates both 1D protein sequence information and 3D structural information (including residue, backbone atom, and side-chain atom) to construct a comprehensive representation of protein complex. By leveraging this rich and hierarchical representation, our model effectively captures the geometric and physicochemical information on protein complexes. Extensive experimental evaluation demonstrates that our approach delivers competitive predictive performance compared to existing affinity prediction methods.