A Flexible, Gpu - Powered Fast Multipole Method For Realistic Biomolecular Simulations In Gromacs

The calculation of electrostatic interactions is typically the computational bottleneck of molecular dynamics (MD) simulations and thus decisive for the overall simulation performance. Further, biomolecules typically contain many sites whose electrostatic charge distribution changes over time, e.g. due to uptake and release of protons or tautomerism at protonatable sites, or electron transfer between redox-active cofactors. Besides computational efficiency, a physically accurate electrostatics treatment therefore has to account for this variability, too, thus aggravating the bottleneck. Taking advantage of the computational power of GPUs through innovative algorithms, high-throughput simulations of large systems become feasible. To that aim, we have developed a fast multipole method (FMM) for the rapid computation of the electrostatic interactions that are required for a lambda-dynamics treatment of the interconversion between the different site forms during the simulation. The tree data structure used by the FMM allows one to include alternative charge distributions of the different protonation forms without requiring redundant computations. Therefore, our FMM enables efficient computation of electrostatic forces and interaction energies between large numbers of titratable sites with a small, nearly constant computational overhead. For taking full advantage of GPUs, the FMM implementation ensures that the computational work is evenly distributed among the large number of GPU processing units. To this aim, our parallel GPU-implementation optimally matches hardware and algorithmic requirements, resulting in very good scaling properties.