Molecular dynamics (MD) simulations have been used as a computational microscope to investigate and understand phenomena related to protein dynamics. However, since MD simulations are run using 10−15s time steps due to being limited by the fastest motions in the system (e.g., the vibration of bonds) and protein dynamics ranges from 10−6s to longer, there is a major timescale barrier between MD simulations and protein dynamics. Though “enhanced sampling methods,” a class of computational tools,can help accelerate the exploration of the protein conformational spaceand dynamics compared to conventional MDsimulations, obtaining accurate thermodynamic and kinetic properties of large systems (>200,000 atoms) from MD simulations in a computationally tractable period is still challenging. To tackle this issue, we developeda novel enhanced sampling method called “parallelized Gaussian accelerated molecular dynamics” (ParGaMD), whichruns many short Gaussian accelerated molecular dynamics (GaMD) simulations over multiple GPUs in parallel by using the weighted ensemble method (WE) framework. Although GaMD has demonstrated accelerated sampling for many systems by adding a boost potential to the system, GaMD takes weeks to run for large systems since the Amber MD GPU enginepmemd.cuda, in which GaMD is mainly implemented, does not parallelize well over multiple GPUs. By using the WE framework, we were able to overcome this bottleneck and sample more efficiently along the chosen collective variables, which enables ParGaMD to be more powerful than GaMD itself.ParGaMDcansignificantlyacceleratethesamplingofdifferent conformational states and dynamics of any protein, biomolecule, and material system, which will benefit the wider scientific community. We will demonstratethat ParGaMD is able to achieve close to linear scaling on multiple GPUsover two popular MD engines (AMBER and OpenMM) and significantly reduce the needed walltime for MD simulationsof variousbiomolecular systems.
