Sequence Based Prediction of Protein-Protein Interactions Via Siamese Neural Networks

Abstract

Proteins are complex macromolecules and participate in nearly every process within the living cells. They generally make physicochemical connections and complex structures called protein-protein interactions (PPIs) to carry out their specific functions. The PPIs play essential roles in cellular processes and regulate various cellular functions such as signal transduction, recognition of foreign molecules, and immune response. Additionally, they have a high potential for drug discovery applications, treatment design, and understanding of disease mechanisms. Therefore, it is crucial to identify PPIs rapidly and accurately. In this study, we aim to investigate the performance of the convolutional Siamese neural network approach for the prediction of PPIs by only using the sequence information of proteins. We encoded protein sequences using three different protein representation methodologies: Binary Representation, Auto Covariance (AC), and Position Specific Scoring Matrices (PSSM). Results show that the PSSM method gives better accuracy than the other two encoding methods. Also, we have presented that the implemented convolutional Siamese neural network approach improves sequence-based PPI prediction.