Abstract
Protein-protein interactions (PPI) are crucial in disease and infection pathways. Peptide-protein interactions (PepPI) provide insights into the peptide-based drug discovery process. This paper presents an approach that explores and exploits novel features related to a protein-protein or peptide-protein pair's backbone. The paper introduces a signal-processing treatment of features by leveraging the Gaussian mixture model (GMM)-based analysis of the protein's and peptide's backbone features. Specifically, we propose a structure-based feature set composed of statistical features and regression model features. These structurebased features are derived from a protein's or a peptide's backbone inter alpha-carbon atomic distances and the regression of these distances against backbone dihedral angles using a deep neural network. In this setting, we present a comparative study with two approaches: (a) A signal-processing treatment of these features along with a Multi-layer Perceptron (MLP)-based protein-protein or peptide-protein interaction prediction model, (b) A Convolutional Neural Network (CNN)-based approach that directly uses the structure-based features derived from protein and peptide backbones and other physicochemical and sequence-related features. We also present a novel computational approach to developing protein-protein interaction datasets from experimentally determined structures of protein complexes. Our results with both approaches are encouraging in comparison with the state-of-the-art methods. Thus our work provides a context for representation learning for protein or peptide structure-based features, especially in the absence of co-evolutionary information on protein-protein or peptide-protein association.