Fast binding site mapping using GPUs and CUDA

Binding site mapping refers to the computational prediction of the regions on a protein surface that are likely to bind a small molecule with high affinity. The process involves flexibly docking a variety of small molecule probes and finding a consensus site that binds most of those probes. Due to the computational complexity of flexible docking, the process is often split into two steps: the first performs rigid docking between the protein and the probe; the second models the side chain flexibility by energy-minimizing the (few thousand) top scoring protein-probe complexes generated by the first step. Both these steps are computationally very expensive, requiring many hours of runtime per probe on a serial CPU. In the current article, we accelerate a production mapping software program using NVIDIA GPUs. We accelerate both the rigid-docking and the energy minimization steps of the program. The result is a 30× speedup on rigid docking and 12× on energy minimization, resulting in a 13× overall speedup over the current single core implementation.