Inter-chain residue-residue distances prediction method for protein complexes using deep learning

Protein-protein interactions are crucial for understanding cellular life activities. Predicting inter-chain residue-residue distances provides critical information for analyzing protein-protein interactions as well as for modeling flexible, dynamic, and ultra-large-scale protein complex structures. Existing inter-chain residue-residue distances prediction methods primarily focus on inter-chain information, neglecting inter-domain information. Based on the dataset of inter-chain domain-domain interactions that we constructed, a deep learning method for predicting inter-chain residue-residue distances in protein complexes (i.e., DPIC) is proposed. First, for the input monomer sequences, multiple sequence alignment for each monomer sequence and paired multiple sequence alignment for the complex sequence are constructed. Meanwhile, the monomer structures are predicted using AlphaFold2. Second, the sequence features, multiple sequence alignment features, and structural features are constructed. An ensemble deep learning network that combines multi-column convolutional neural network modules with triangular interaction modules is designed to predict inter-chain residue-residue distances in protein complexes. Finally, the results of the CASP13-15 dimers test set show that the inter-chain residue-residue contact prediction precision of DPIC outperforms mainstream methods, such as DeepInter and CDPred, with Top L/10 precision improved by 7.87% and 13.74%, respectively. Furthermore, the results of modeling large protein complexes indicate that the inter-chain residue-residue distances predicted by DPIC contribute to the assembly of large protein complexes. The web server of DPIC can be accessed at hrefhttp://zhanglab-bioinf.com/DPIChttp://zhanglab-bioinf.com/DPIC.