Protein Fold Classification using Graph Neural Network and Protein Topology Graph

Abstract

Protein fold classification reveals key structural information about proteins that is essential for understanding their function. While numerous approaches exist in the literature that classifies protein fold from sequence data using machine learning, there is hardly any approach that classifies protein fold from the secondary or tertiary structure data using deep learning. This work proposes a novel protein fold classification technique based on graph neural network and protein topology graphs. Protein topology graphs are constructed according to definitions in the Protein Topology Graph Library from protein secondary structure level data and their contacts. To the best of our knowledge, this is the first approach that applies graph neural network for protein fold classification. We analyze the SCOPe 2.07 data set, a manually and computationally curated database that classifies known protein structures into their taxonomic hierarchy and provides predefined labels for a certain number of entries from the Protein Data Bank. We also analyze the latest version of the CATH data set. Experimental results show that the classification accuracy is at around 82% - 100% under certain settings. Due to the rapid growth of structural data, automating the structure classification process with high accuracy using structural data is much needed in the field. This work introduces a new paradigm of protein fold classification that meets this need. The implementation of the model for protein fold classification and the datasets are available here https://github.com/SuriDipannitaSayeed/ProteinFoldClassification.git Author summaryClassification of protein structures is traditionally done using manual curation, evolutionary relationship, or sequence comparison-based methods. Applying machine learning and deep learning to protein structure classification is a comparatively new trend that holds great promises for automating the structure classification process. Advance deep learning technique like Graph Neural Network is still unexplored in this respect. SCOP and CATH are two traditional databases that provide the hierarchical taxonomic classification of protein structures. This work provides a novel computational approach that classifies protein folds in SCOP and CATH with graph neural network, performing a graph classification task.