Exploring Protocols to Build Reservoirs to Accelerate Replica Exchange MD Simulations

Replica Exchange Molecular Dynamics (REMD) is a widely used enhanced sampling method for accelerating biomolecular simulations. During the past two decades, several variants of REMD have been developed to further improve the rate of conformational sampling of REMD. One such variant, Reservoir REMD (RREMD), was shown to improve the rate of conformational sampling by around 5-20x. Despite the significant increase in sampling speed, RREMD methods have not been widely used due to the difficulties in building the reservoir and also due to the code not being available on the GPUs.

In this work, we ported the AMBER RREMD code onto GPUs making it 20x faster than the CPU code. Then, we explored protocols for building Boltzmann-weighted reservoirs as well as non-Boltzmann reservoirs, and tested how each choice affects the accuracy of the resulting RREMD simulations. We show that, using the recommended protocols outlined here, RREMD simulations can accurately reproduce Boltzmann-weighted ensembles obtained by much more expensive conventional REMD simulations, with at least 15x faster convergence rates even for larger proteins (>50 amino acids) compared to conventional REMD.