Redes convolucionais na identificação de alterações sutis em mapas de distâncias de conformações de dinâmica molecular

Abstract

One way to understand the effects of mutations on the natural mobility of a structure is through molecular dynamics simulations. This approach has been used in the study of SARS-CoV-2 variants, especially those that tend to develop a higher binding affinity for the human receptor. Due to the volume of information generated, DM trajectories are considered as large-scale data, and their acquisition (and subsequent analysis) is commonly associated with a high computational cost. Even so, these trajectories reveal subtle conformational changes inherent to the structures, creating a favorable scenario for deep learning techniques. In this sense, the present work proposes to evaluate the use of convolutional networks in predicting the impact of mutations, through data obtained from molecular dynamics simulations represented as distance maps. These data correspond to the region of the S protein known as Receptor Binding Domain, where mutations characteristic of SARS-CoV-2 variants are common. In this way, the model is able to predict whether the product of a new simulation, referring to a mutated structure, would have a greater affinity with the human receptor and, consequently, be more infective. Based on the results of the model's precision, to the different systems of the database, it was possible to correlate the changes in binding free energy for different mutations, with an estimated $\rho$ equal to 0.77. This value indicates that the model's predictions follow recently published results, regarding the affinity, or neutrality, of mutations found in the RBD of the S protein. Finally, it is also possible to observe which regions of the distance map would be contributing to the learning and to the decision of the model, through the corresponding resource map.