Research On Molecular Blind Docking Based On Graph Neural Network And GPU Acceleration

With the rapid development of computer technology,computer-aided drug design is widely used in drug research and development.In the actual application of molecular docking,without knowledge of the docking pocket,the blind docking mode is presented to find the best binding mode of receptor and ligand on the whole surface of the protein target.In this case,the common molecular docking software takes quantities of time to find the best binding mode for the blind docking mode.Therefore,how to quickly and accurately deal with blind docking has become an important issue in large-scale drug virtual screening.In this paper,in order to improve QVina-W,the software of molecular blind docking,we propose a GPU-accelerated and multi-thread collaborative blind docking method,QVina-W-GPU,which utilizes GPU parallel processing capability to reduce the search depth of Monte Carlo’s iterative local search by increasing the number of initial random ligand conformations.Meanwhile,the octree is implemented in Open CL by constructing global memory structure to further improve the efficiency of inter-process communication.On this basis,a CPU-GPU heterogeneous architecture is designed to achieve higher parallelism and acceleration capability.Finally,multiple experiments were carried out on the PDBBind public dataset,and the results show that our proposed GPU-accelerated and multi-thread collaborative molecular blind docking method accelerates computations for molecular docking while guaranteeing the accuracy of molecular docking in the blind docking mode.In order to obtain the best binding mode of ligand and receptor more accurately,a scoring function optimization algorithm based on graph neural network is proposed.First,we construct a protein-ligand interaction graph,and learn the representation for the ligand atoms and ligand-protein interactions to extract the node features in the graph.Hence,the binding affinity prediction model of protein-ligand can be constructed by regression algorithm and graph attention network.Finally,we use the trained prediction model to optimize the molecular conformation of the output of QVina-W-GPU,and the experimental results show that our proposed algorithm can effectively enhance the docking accuracy of QVina-W-GPU and meet the actual requirements for molecular docking of large-scale compound databases.