Recognition of Conformational States of a G Protein-Coupled Receptor from Molecular Dynamic Simulations Using Sampling Techniques

Protein structures are complex and dynamic entities relevant to many biological processes. G-protein-coupled receptors in particular are a functionally relevant family of cell membrane proteins of interest as targets in pharmacology. Nevertheless, the limited knowledge about their inherent dynamics hampers the understanding of the underlying functional mechanisms that could benefit rational drug design. The use of molecular dynamics simulations and their analysis using Machine Learning methods may assist the discovery of diverse molecular processes that would be otherwise beyond our reach. The current study builds on previous work aimed at uncovering relevant motifs (groups of residues) in the activation pathway of the \(\beta 2\)-adrenergic (\(\beta _2AR\)) receptor from molecular dynamics simulations, which was addressed as a multi-class classification problem using Deep Learning methods to discriminate active, intermediate, and inactive conformations. For this problem, the interpretability of the results is particularly relevant. Unfortunately, the vast amount of intermediate transformations, in contrast to the number of re-orderings establishing active and inactive conditions, handicaps the identification of relevant residues related to a conformational state as it generates a class-imbalance problem. The current study aims to investigate existing Deep Learning techniques for addressing such problem that negatively influences the results of the predictions, aiming to unveil a trustworthy interpretation of the information revealed by the models about the receptor functional mechanics.