Prediction of protein-protein interactions using stacked auto-encoder

Abstract

Protein-protein interactions (PPIs) play essential roles in understanding the protein functions and the corresponding pathways which are involved in various biological processes, as well as help in understanding the cause and growth of diseases. Several computational methods such as Support Vector Machine and decision tree are popularly used along with the experimental methods to address the PPIs problem. Such algorithms consider different protein features, including protein sequence, genomes, protein structure, function, topology of the PPIs network, and those that combine multiple aspects. Nowadays, Deep learning (DL) algorithms are successfully used in solving problems in different domains. So, in this paper, we have used stacked auto-encoder as one of the DL methods in solving the problem of PPIs. This model takes the input 92-length feature vector, which is the integration of features extracted from the protein sequence using different methods. The feature vector consists of evolutionary features obtained by PSI-BLAST algorithm, predicted structural properties obtained by SPIDER2, and seven physicochemical properties of amino acids. The key novelty of the current study lies in extracting useful features to solve the PPI problem. The results obtained by our method of feature extraction are compared with those obtained by other feature extraction methods such as Autocovariance and Conjoint-triad, and our proposed feature extraction method is found to be relatively more accurate. © 2021 John Wiley & Sons, Ltd.