面向虚拟筛选的GPU加速的分子对接方法

虚拟筛选是在计算机上对化合物分子进行模拟预筛选,找出容易和药物靶标结合的小分子(配体),从而降低实际实验测试次数,提高药物先导化合物的发现效率。常用的分子对接软件可以用于基于结构的虚拟筛选,寻找配体与靶标的最佳的作用模式和结合构象,并通过打分函数来筛选出潜在的配体。现有的对接软件如AutoDock Vina等在分子对接过程中需要耗费大量时间和计算资源,特别是面对大规模分子对接时,过长的筛选时间不能满足应用需求,因此,本文在最高效的QVina2对接软件基础上,提出一种基于GPU的QVina 2并行化方法QVina2-GPU,利用GPU硬件高度并行体系加速分子对接。具体包括增加初始化分子构象数量,以扩展蒙特卡罗的迭代局部搜索中线程的并行规模,增加蒙特卡罗的迭代搜索的广度以减少每次蒙特卡罗迭代搜索深度,并利用Wolfe-Powell准则改进局部搜索算法,提高了对接精度,进一步减少蒙特卡罗迭代搜索深度,最后,在NVIDIA Geforce RTX 3090平台上在公开的配体数据库上验证了QVina2-GPU的性能,实验表明在保证分子对接精度的基础上,我们提出的QVina2-GPU对Qvina2的平均加速比达到5.18倍,最大加速比达到12.28倍。

GPU accelerated molecular docking method for virtual screening

Virtual screening exploits computational platforms to simulate pre-screening in which molecules (ligands) that can readily bind to receptors are identified, reducing the number of drug tests and improving the efficiency of drug lead compounds discovery. Some commonly-used docking software have been applied to structure-based virtual screening to find the proper binding conformations and affinities. Specifically, the scoring function is employed to identify potential binders. However, the existing docking software such as AutoDock Vina are still time-consuming and computationally expensive, especially for larger scale dockings where longer execution time compromises the performance of the docking programs in the screening tasks. Hence, a GPU accelerated Molecular docking method for Virtual screening, QVina2-GPU, is proposed to parallelize the efficient QVina2 docking software, which can exploit the highly-parallel hardware architecture to speed up the process of molecule docking. Specifically, the multi-threaded parallel execution of Monte-Carlo iterative local search is enhanced by increasing initial molecule conformations. That is, the breadth of Monte-Carlo search is increased to reduce the depth of each Monte-Carlo search. Meanwhile, the rule of Wolfe-Powell is applied to upgrade the local search for better docking accuracy, which can further reduce the depth of Monte-Carlo search. Finally, QVina2-GPU was implemented with NVIDIA Geforce RTX 3090 on the public ligand database, and the experimental results demonstrate that the proposed QVina2-GPU can achieve an average acceleration of about 5.18 times and a maximum acceleration of up to 12.28 times compared to Qvina2 while guaranteeing desired docking accuracy.