Protein fold classification with Grow-and-Learn network

Abstract

Protein fold classification is an important subject in computational biology and a compelling work from the point of machine learning. To deal with such a challenging problem, in this study, we propose a solution method for the classification of protein folds using Grow-and-Learn (GAL) neural network together with one-versus-others (OvO) method. To classify the most common 27 protein folds, 125 dimensional data, constituted by the physicochemical properties of amino acids, are used. The study is conducted on a database including 694 proteins: 311 of these proteins are used for training and 383 of them for testing. Overall, the classification system achieves 81.2% fold recognition accuracy on the test set, where most of the proteins have less than 25% sequence identity with the ones used during the training. To portray the capabilities of the GAL network among the other methods, comparisons between a few approaches have also been made, and GAL's accuracy is found to be higher than those of the existing methods for protein fold classification.