A Deep Learning-Based Method for Uncovering GPCR Ligand-Induced Conformational States Using Interpretability Techniques

There is increasing interest in the development of tools for investigating the protein ligand space. Understanding the underlying mechanisms of G protein-coupled receptors (GPCR) in the ligand-binding process is of particular interest due to their role in pharmacoproteomics. In this work, we propose the study of GPCR ligand-induced conformational variations from Molecular Dynamics (MD) simulations using Deep Learning (DL)-based methods. We devise and train a Convolutional Neural Network (CNN) for classifying the states for both ligand-free structure and the bound of agonists in the beta 2-adrenergic receptor. We also study the transformation of MD data into an interaction network matrix to further improve and facilitate the analyses without significant loss of information. Our method introduces a framework for the study of the effect of ligand-receptor binding activity that includes a novel analysis based on interpretability algorithms, allowing the quantification of the contribution of individual residues to structural re-arrangements.