Assessment of Binding Affinity in the Complexes of ACE2 with RBD of the S Protein of SARS-CoV Using Convolutional Neural Networks

Abstract

The experimentally obtained structures of 48 ACE 2 receptor complexes with RBD of the S protein of the SARS-CoV and SARS-CoV-2 coronaviruses (including mutant forms of the latter) were evaluated, for which the dissociation constants were calculated. To predict the binding affinity, the ProBAN neural network algorithm developed by the authors earlier was used, as well as a number of other Gibbs free energy estimation algorithms: Prodigy, FoldX, DFIRE, and RosettaDock. A comparison of the evaluation results showed that ProBAN demonstrated the best prediction quality (Pearson correlation coefficient was 0.56 and the mean absolute error was 0.7 kcal/mol). The results we obtained suggested a better quality of affinity prediction for other protein–protein complexes as well. Information about the studied complexes and the prediction results are available in the repository at the link: https://github.com/EABogdanova/ProBAN_RBD-ACE2.