Prediction of protein-protein interaction with cosine matrices

Abstract

The protein-protein interaction prediction problemis one of the unsolved fundamental problems of bioinformatics and structural biology. A wide range of machine learning approaches has been developed, relying on prediction of protein protein interaction interface. In this study we have tried a different approach to the problem. It relies on prediction of molecule centers displacement directions and their relative ro tation. We present a novel protein structure representation with cosine matrices. These matrices can be considered as successors of widely used distance maps. They have useful properties such as rotation/shift invariance and self-correcting behavior. We developed a fully convolutional neural network architecture, which is able to predict dimer complexes (both homodimer and heterodimer). The model allowed to achieve 51% of correct predictions (59% for homodimers and 45% for heterodimers) for a test set of 5,854 complexes and 10 angstrom RMSD threshold.