Trimer Protein-Protein Complex Interface Interacting Residue Pairs Prediction Using Deep Learning Approach

Abstract

Protein-protein interactions are significant for many biological processes. Identifying the residues on the interface is the first step to understand it at the structural level. However, accurate computational prediction of interacting residue pairs in protein-protein complexes is a challenged task. Inspired by theory of deep learning, the author proposed an innovation in predicting interacting residue pairs in trimers based on Long Short-Term Memory networks. In this method, we used three theoretically proposed geometric features to describe a residue to express its 3D structure. In trimers, a pair of residues from different protein chain whether interact not only depends on property themselves, but also on surrounding residues concerned. And the influences of surrounding residues related to the distance. Therefore, we used sequence-based Long Short-Term Memory to build our prediction model. Taken together information from both protein chains in a protein complex, we trained and tested our model on Protein-Protein Docking Benchmark version 5.0. In general, for each trimer in the test set, if we provided 25 pairs as interacting residue pairs, we can at least obtain one correct prediction on 100% protein-protein complex trimers. Compared with other four machine learning methods, our method has greatly improved the predictive accuracy.