Research On Fast Molecular Docking Method Based On GPU Acceleration And Local Search Optimization

In recent years,drug virtual screening has been widely used in the medicinal chemistry industry,which mainly uses computer technology to simulate the interaction between targets and compounds for pre-screening.The drug virtual screening is designed to identify the small molecules(ligands)binding to the target drug with high probability,which can boost the efficiency of discovering the drug lead compounds by reducing the number of physical and chemical experiments,and lowering the labor and material cost.Commonly-used molecular docking software is developed for the structure-based virtual screening,which can explore the optimized interaction model by estimating the interaction force between the ligand and the target with the scoring function.In this way,the potential ligand can be screened out.The current mainstream molecular docking software such as Auto Dock Vina consumes ignificant time and computing resources during the molecular docking process,which limits its use in large-scale molecular docking application.Based on the most popular QVina2 docking software,we propose a method of parallelization of molecule docking software QVina2,QVina2-GPU,based on CPU-GPU heterogeneous architecture,which can accelerate the docking software with the powerful parallel computing capability of GPU.Specifically,the number of initialization molecule conformations is increased to meet the requirement of the parallelizing the threads to execute the iterative search of fast heuristic Monte Carlo.Therefore,the breadth of the iterative search of Monte Carlo is increased to reduce the depth of each iteration of Monte Carlo search.We conducted multiple comparable experiments in the public ligand database on the NVIDIA Geforce RTX 3090 platform,and the experimental results show that the maximum speedup ratio of QVina2-GPU to QVina2 reaches 5.03 times,and the maximum Monte Carlo time speedup ratio reaches 31.23 times.Considering the local search with Armijo criterion in QVina2-GPU proposed above,the expense of more search iterations and search time will be incurred due to the fact that step size is too small.We propose a method of parallelization of molecule docking,QVina2-GPU-W,with the Wolfe-Powell criterion,where the Wolfe-Powell criterion is utilized to improve the efficiency of the local search in QVina2-GPU.Therefore,the docking accuracy and efficiency are both enhanced by further reducing the fast heuristic Monte Carlo iterative search depth and Search time.Finally,we validated the superiority of QVina2-GPU-W in the public ligand database on the NVIDIA Geforce RTX 3090 platform.The experimental results show that the maximum speedup ratio of QVina2-GPU-W with the Wolfe-Powell criterion to QVina2 reaches 12.28 times.The maximum Monte Carlo time speedup ratio reaches 60.28 times.In summary,with the help of GPU-accelerated and local search optimized technology,the acceleration performance and docking accuracy of our proposed QVina2-GPU-W are verified by these above multiple experiments.