Can Simulations be Recycled to Benchmark RNA Force Fields

Abstract

Molecular dynamics is a computationally intensive but powerful technique used to simulate chemical systems at atomic resolution. It approximates the energetic favorability of the system's conformations with a force field, then samples that to simulate a thermodynamic ensemble. This can be inaccurate in two ways. It can inadequately sample the system's conformations—a failure to converge. Alternatively, the force field can incorrectly estimate the system's energetic preferences. One cannot estimate the systematic error due to the force field without sampling enough to have low statistical error, but this process is computationally expensive. Our objective is to establish whether benchmark simulations might be recycled to allow force field developers to iterate. Using the Boltzmann equation, one can estimate how probable a conformation generated from a simulation with one force field would be with another, allowing one to reweight the sampled data and estimate what the average would have been. Since this approach is only accurate when the two models cover similar regions of conformation space, reweighting has seemed inviable for many systems of interest given historically available computational power. Here we assay the convergence of reweighted estimates of Nuclear Overhauser Effects for small, flexible RNAs in long simulations using FF99.OL3 and RNA.ROC in AMBER. We reweighted simulations performed with one force field to estimate the ensemble generated by the other. Six replicas of each type of simulation have been performed to permit error estimation. This provides a controlled test of how much simulation it takes to converge a reweighting estimate, since we have both experimental data and a mature ensemble simulated with the estimating force field to benchmark estimated quantities.