Prediction of binding sites of GPCRs based on 3D convolutional neural networks

Abstract

AbstractG protein-coupled receptors are a class of receptor proteins located on the cell membrane, regulating the perception and response of cells to various external signals. Identifying the binding sites of G protein-coupled receptors plays a crucial role in understanding their allosteric modulation mechanisms. However, obtaining the crystal structure of the complex through experimental means and subsequently identifying the binding sites require substantial resources. With the development of computer-aided computation, deep learning can effectively predict the binding sites between proteins and ligands. This study predicted the binding sites of G protein-coupled receptors based on 3D convolutional neural network. A total of 108 G protein-coupled receptors recorded in the scPDB database were collected for this study, and a 3D convolutional neural network model was established based on these three-dimensional structures. Firstly, the PDB file of the protein is voxelized and segmented into different channels according to the type of atoms in a certain region. Then, 3D convolutional neural network is employed to predict the binding site through traversal, and the optimal voxel box size and model parameters are determined based on performance evaluation metrics. The established 3D convolutional neural network accurately predicts the binding site of G protein-coupled receptors, with an accuracy as high as 0.942, precision of 0.678, and recall rate of 0.532. Additionally, using the backpropagation algorithm, the gradients of the input data are calculated, and the importance of input elements on the final classification result is analyzed.