DeepSSInter: Protein-protein contact prediction with a structure-aware protein language model

Accurate prediction of the interface residue-residue contacts between interacting proteins is valuable for determining the structure and function of protein complexes. Recent deep learning methods have drastically improved the accuracy of predicting the interface contacts of protein complexes. However, existing methods rely on Multiple Sequence Alignments (MSA) features which pose limitations on prediction accuracy, speed, and computational efficiency. Here, we propose a transformer-powered deep learning method to predict the inter-protein residue-residue contacts using single-sequence and structure-aware protein language models (PLM), called DeepSSInter. Utilizing the intra-protein distance and graph representations and the ESM2 and SaProt PLM, we are able to generate the structure-aware features for the protein receptor, ligand, and complex. These structure-aware features are passed into the ResNet Inception module and the Triangle-aware module to effectively produce the predicted inter-protein contact map. Extensive experiments on both homo- and hetero-dimeric complexes show that our DeepSSInter model significantly improves the performance in both accuracy and speed compared with previous state-of-the-art methods. Integrating predicted contacts significantly improves the docking performance. The DeepSSInter is available at https://github.com/huang-laboratory/DeepSSInter/.