Computational Prediction of Interactions Between SARS-CoV-2 and Human Protein Pairs by PSSM-Based Images

Abstract

Identifying protein-protein interactions is essential to predict the behavior of the virus and to design antiviral drugs against an infection. Like other viruses, SARS-CoV-2 virus must interact with a host cell in order to survive. Such interaction results in an infection in the host organism. Knowing which human protein interacts with the SARS-CoV-2 protein is an essential step in preventing viral infection. In silico approaches provide a reference for in vitro validation to protein-protein interaction studies by finding interacting protein pair candidates. The representation of proteins is one of the key steps for protein interaction network prediction. In this study, we proposed an image representation of proteins based on position-specific scoring matrices (PSSM). PSSMs are matrices that are obtained from multiple sequence alignments. In each of its cells, there is information about the probability of the occurrence of amino acids or nucleotides. PSSM matrices were handled as gray-scale images and called PSSM images. The main motivation of the study is to investigate whether these PSSM images are a suitable protein representation method. To determine adequate image size, conversion to grayscale images was performed at different sizes. SARS-CoV-2-human protein interaction network prediction based on image classification with siamese neural network and Resnet50 was performed on PSSM image datasets of different sizes. The accuracy results obtained with 200x200 size images and siamese neural network as 0.915, and with 400x400 size images and Resnet50 as 0.922 showed that PSSM images can be used for protein representation.