Vina-CUDA: An Efficient Program with in-Depth Utilization of GPU to Accelerate Molecular Docking

As a mainstream technology in modern drug discovery, molecular docking methodologies enable precise and efficient identification of lead compounds within large chemical repositories to improve drug development efficiency and reduce costs. The exponential growth of chemical databases has substantially expanded drug discovery resources while improving the identification rates of true positives in lead compounds. However, this rapid expansion poses significant challenges for existing docking tools to efficiently screen lead compounds from these massive chemical libraries. In this study, we proposed Vina-CUDA, which leverages GPU hardware features to optimize and accelerate the core algorithm of the popular tool AutoDock Vina at three aspects, computational capability, memory access, and resource utilization, significantly improving docking efficiency. A hybrid parallel optimization strategy integrating task and computational parallelism was implemented, accompanied by systematic code and data structure optimization, to maximize GPU resource utilization and enhance computational efficiency. Building upon this, we developed its derivatives, QuickVina2-CUDA and QuickVina-W-CUDA, as well as a user-friendly multi-GPU docking framework to utilize multi-GPU resources to accelerate large-scale virtual screening tasks. The performance and docking accuracy of Vina-CUDA and its derivatives were evaluated under five chemical databases. Results showed that, compared to baseline programs, Vina-CUDA with RILC-BFGS optimization algorithm achieved average and maximum accelerations of 3.71× and 6.89× across five databases, while QuickVina2-CUDA and QuickVina-W-CUDA achieved average speedups of 6.19× and 1.46×, respectively, without compromising docking accuracy. Furthermore, Vina-CUDA and its derivatives demonstrated comparable performance to baseline docking programs in docking, scoring, and ranking power, with excellent scalability and portability.