An Evaluation of Force Field Accuracy for the Mini-Protein Chignolin using Markov State Models

All-atom molecular dynamics (MD) simulations can provide detailed insight into a molecule's conformational ensemble in solution. While molecular force fields are parameterized to accurately model a protein's potential energy surface, it remains challenging in practice to evaluate how well force fields can capture ensemble-averaged experimental observables, since it requires simulation of the complete folding landscape. In this work, we employ massively parallel molecular simulations, performed using the Folding@home distributed computing platform, to investigate the ability of nine force fields (AMBER14SB, AMBER99, AMBER99SB, AMBER99SB-ildn, AMBER99SBnmr1-ildn, CHARMM22*, CHARMM27, CHARMM36 and OPLS-aa) with TIP3P explicit solvent to accurately reproduce experimental observables for chignolin, a beta-hairpin mini-protein with an experimental folding time of ~600 ns. From over 200 µs of aggregate trajectory data, we constructed Markov state models (MSMs) to obtain estimates of thermodynamic and kinetic properties of chignolin in each force field. Quantitative assessment of the force fields was performed by comparing predicted and experimental folded populations, and the statistical agreement between predicted and experimental solution-state NMR observables. This work highlights the utility of MSM approaches for force field evaluation, and provides a baseline for future studies using Bayesian inference methods to evaluate and parameterize force fields.