Abstract
<jats:p>One solution to the insufficient sampling obtained in biomolecular simulations is coarse-graining, combining multiple atoms into single beads which interact through “effective” interactions, and implicit solvation, in which the solvent is not treated explicitly but modulates the effective interactions. Implicit solvation and coarse-graining drastically reduce the number of beads and inter-bead interactions to be simulated and this enables sufficient sampling. However, graphicsprocessing units (GPUs), which have increasingly become the norm in molecular dynamics (MD) simulations, have many cores, and each GPU often cannot be split over multiple simulation jobs. This leads to an underutilization of the GPU resources when the simulation involves only a few beads. To overcome this underutilization and to enhance sampling further, we introduce the simple but powerful idea of simulating multiple non-interacting replicas (NIR; identical copies of the system to be simulated) within one simulation box. We benchmark the performance of this technique using MD simulations of coarse-grained structure-based models of ubiquitin (76 beads) and the SARS-CoV-2 spike protein (1248 beads) performed using OpenSMOG/OpenMM on two different low to midrange desktop NVIDIA GPU cards. We show that the NIR technique leads to more sampling per compute time through the better utilization of GPU resources. Additionally, the method of NIR increases sampling efficiency without affecting dynamics and kinetics and thus, data-unbiasing is not required. The technique is hardware and software agnostic and so, should be useful for a wide range of coarse-grained implicit-solvation MD simulations which would otherwise underutilize available computational resources.</jats:p>