Accelerate AutoDock Vina With Software-Hardware Co-design And Virtual Screening Methods

Drug discovery is a costly and time-consuming process.In the early stages of drug discovery,virtual screening is employed to identify potential drug candidates from a large compound database.Auto Dock Vina,a widely used molecular docking tool,offers high docking speed and accuracy,making it popular over the past three decades.However,with the advent of big data,compound databases have grown significantly larger.While larger databases present more opportunities for discovering better drug candidates,conducting virtual screening with Auto Dock Vina on such vast databases is extremely time-consuming and fails to meet the demands of rapid drug discovery.Current attempts to accelerate Auto Dock Vina mainly involve stacking computing resources in a simplistic manner,such as using the Virtual Flow platform.However,this approach incurs substantial resource costs and imposes high entry barriers for users,severely limiting the popularity and flexibility of Auto Dock Vina in modern drug discovery.Therefore,this thesis makes contribution to the software-hardware acceleration of Auto Dock Vina and virtual screening method in the following aspects:(1)Addressing the need for enhanced docking speed,this thesis proposes Vina-GPU,a multi-start parallel annealing-like algorithm that leverages significantly increased initial random conformations and reduced search depth for each conformation.Experiment results show that Vina-GPU obtains the same docking accuracy with Auto Dock Vina.(2)Additionally,employing the Open CL framework,this thesis presents a GPU-accelerated,high-performance parallel computing hardware implementation that achieves an average docking speed improvement of 21 times.In a real virtual screening task,Vina-GPU reduces the screening runtime from 5 days to just 9 hours.(3)At last,this thesis introduces Vina-GPU+,a powerful docking tool that provides a range of optimizations targeting the virtual screening procedure.These optimizations include implementing grid cache calculations on the GPU and optimizing redundant calculations during the virtual screening process.In a virtual screening scenario involving the RIPK1 and RIPK3 targets,Vina-GPU+ achieves over three times the acceleration compared to Vina-GPU.This thesis also provides the open-source versions of Vina-GPU and Vina-GPU+,along with a user-friendly graphical user interface,comprehensive user manual,and source code for the CUDA version of Vina-GPU.These contributions greatly advanced the field of open-source molecular docking computing.Since the introduction of the Vina-GPU heterogeneous accelerator,it has garnered significant attention and adoption by numerous prestigious universities and research institutions,such as Microsoft Research,DP Technology,Harvard University,and China Pharmaceutical University.This work has received a wealth of positive feedback on the open-source platform Git Hub,underscoring its influential contribution to the field of molecular docking.