Background
AutoDock Vina is one of the most widely used molecular-docking tools — in the CASF-2016 benchmark it achieved the best docking power of all tools evaluated. Modern drug discovery increasingly means screening huge compound libraries, but Vina’s search algorithm is inherently serial, and before this work there was no successful report of parallelising it on GPUs. Speed-ups came only from stacking more machines (e.g. the VirtualFlow platform), which is costly and hard to access.
Approach
- Re-designed search for parallelism. Vina-GPU builds on a modified Monte Carlo search with simulated annealing: it greatly raises the number of initial random conformations and reduces the search depth of each thread, so thousands of threads can explore in parallel.
- Local optimisation. The classic BFGS optimiser refines ligand conformations during docking.
- Heterogeneous OpenCL implementation that runs the docking across thousands of GPU cores.
Results
On large benchmark tests, Vina-GPU reaches an average 21× and maximum 50× speed-up over the original AutoDock Vina, with comparable docking accuracy — bringing large-scale virtual screening within reach of a single PC, a station server or a cloud instance.
Publication
Shidi Tang, Ruiqi Chen, Mengru Lin, Qingde Lin, Yanxiang Zhu, Ji Ding, Haifeng Hu, Ming Ling, Jiansheng Wu. “Accelerating AutoDock Vina with GPUs.” Molecules, 27(9): 3041, 2022. doi:10.3390/molecules27093041
See also the follow-up hardware work: Vina-FPGA.
