Accelerating MM/PBSA calculation of protein-ligand binding on Graphics Processing Units

Molecular Mechanics Poisson-Boltzmann Surface Area (MM/PBSA) methods have been widely used in calculating protein-ligand binding affinities. However, the efficiency is still limited due to the high dimensionality of typical biomolecular systems while numerous snapshots need to be processed. Here we report our efforts to develop a highly efficient MM/PBSA program on graphics processing units (GPUs) utilizing Nvidia^® CUDA. First, we implemented several linear Poisson-Boltzmann equation (PBE) solvers which are the bottlenecks of PBSA calculation onto GPUs. The preconditioned conjugate gradient (CG) and geometric multigrid (MG) solvers were found to be among the most efficient. Second, we developed higher-order-accuracy X-factor and 2^nd-order harmonic average methods with a simple 7-point stencil and implemented them on GPUs using the biconjugate gradient (BiCG) algorithm. Third, we rewrote most of the computation-intensive setup routines of PBSA into GPU kernels. The level set density calculation, the surface area non-bonded list determination, and the reaction field energy computation were the most time-consuming steps. Finally, we implemented the full GPU-supported MM/PBSA program by hooking the GPU-accelerated molecular mechanics (MM) energy calculation with the PBSA CUDA code. In addition, different matrix formats such as DIA or CSR, the unified memory, and the CUSP library were also explored to optimize the code efficiency. Impressive efficiency speedups of MM/PBSA calculation with a range of biomolecular simulations were achieved. The new GPU-accelerated MM/PBSA code is to be released via the Amber 2020 and will significantly improve the computational throughput of the protein-ligand binding affinities.