Accelerating a Molecular Docking Application by Leveraging Modern Heterogeneous Computing Systems

Abstract: In drug development, molecular docking methods aim at characterizing the binding of a drug-like molecule to a protein. In a typical drug development process, a docking task is repeated millions of time, which makes optimization efforts essential. In particular, modern heterogeneous architectures, such as GPUs, allow for significant acceleration opportunities. AutoDock-GPU, a state-of-the-art GPU-accelerated molecular docking software, estimates the geometrical conformation of a docked ligand-protein complex by minimizing an energy-based scoring function. Our profiling results indicated that a reduction operation, which is performed several millions times in a single docking run, limits performance in AutoDock-GPU. Thus, we proposed a method to accelerate the block-level sum reduction of four-element vectors by using matrix operations. We implemented our method to make use of the high throughput capabilities offered by NVIDIA Tensor Cores to perform matrix operations. We evaluated our approach by designing a simple benchmark, and achieved a 4 to 7-fold runtime improvement compared to the original method. We then integrated our reduction operation into AutoDock-GPU and evaluated it on multiple chemical complexes on three GPUs. This evaluation allowed to assess the possibility to use half-precision reduction operations in parts of AutoDock-GPU code, without detrimental effects on the simulation result. In addition, our implementation achieved an average 27% improvement on the overall docking time during a real-world docking run.