Identifying Large Scale Conformational Changes in Proteins Through Distance Maps and Convolutional Networks

Abstract

AbstractConformational changes in protein structures are strongly correlated with functional changes. Some conformational modifications may be easily noticeable, others are more subtle. In this work, we model the problem of protein conformation classification through its representation as images that illustrate the interatomic distance matrices. We aim then to discover if a convolutional neural network would be able to identify these conformational changes only from the distance patterns in these maps. Hence, this work presents the development of a model based on convolutional neural networks, capable of identifying large scale conformational changes in proteins. As a case study, we used the S protein from SARS-CoV-2, a protein known for its function in the infection of human cells through a conformational change to binding to the human cell receptor. Initially, we intend to identify large-scale conformations, such as states where the S protein trimers are closer together (closed) or further away (open). The proposed classifier achieved a satisfactory performance after cross validation, reaching an average accuracy in validation of 90.58%, with an error of 22.31%. The model was also able to successfully distinguish both classes (open and closed states for S protein) achieving a precision of 84.32% and a recall of 89%. In the test, the accuracy of the model reached 71.79%, with an error rate of 28.2%. Precision and recall reached 68.18% and 78.94%, respectively. For future work, we want to evaluate the ability of such model to identify even more subtle conformational changes, as well as those caused by point mutations that occur in virus variants.KeywordsConformational changesDistance mapsConvolutional networks