Accelerating AutoDock VINA with GPUs

Abstract

AutoDock VINA is one of the most-used docking tools in the early stage of modern drug discovery. It uses a Monte-Carlo based iterated search method and multithreading parallelism scheme on multicore machines to improve docking accuracy and speed. However, virtual screening from huge compound databases is common for modern drug discovery, which puts forward a great demand for higher docking speed of AutoDock VINA. Therefore, we propose a fast method VINA-GPU, which expands the Monte-Carlo based docking lanes into thousands of ones coupling with a largely reduced number of search steps in each lane. Furthermore, we develop a heterogeneous OpenCL implementation of VINA-GPU that leverages thousands of computational cores of a GPU, and obtains a maximum of 403-fold acceleration on docking runtime when compared with a quad-threaded AutoDock VINA implementation. In addition, a heuristic function was fitted to determine the proper size of search steps in each lane for a convenient usage. The VINA-GPU code can be freely available at https://github.com/DeltaGroupNJUPT/VINA-GPU for academic usage.

Simulation and Machine Learning Methods for Ion-Channel Structure Determination, Mechanistic Studies and Drug Design

“…Under the logic that screening of larger libraries could generally lead to higher quality results ( Gorgulla et al, 2020 ), a variety of programs and algorithms have been developed for ultra-high throughput virtual screening. Generally, the ideas for these approaches include optimization in algorithm engineering to enable high-performance computing ( Alhossary et al, 2015 ; Hassan et al, 2017 ) and the usage of graphics processing units ( Santos-Martins et al, 2021 ; Shidi et al, 2022 ), the development of highly-organized workflows ( Gorgulla et al, 2020 ), the application of deep learning which trained QSAR models on docking scores of subsets of the library ( Gentile et al, 2020 ), and the improvement in searching efficiency in a fragment-based manner with the idea partial similar with dynamic programming ( Sadybekov et al, 2022 ). These methods have pushed the limit of virtual screening throughput to libraries with more than tens of billion compounds, and may potentially help the drug discovery of various target classes including ion channels.…”