Refinement of molecular dynamics ensembles using experimental data and flexible forward models

Abstract: A novel method combining maximum entropy principle, the Bayesian-inference of ensembles approach, and the optimization of empirical forward models is presented. Here we focus on the Karplus parameters for RNA systems, which relate the dihedral angles of $\gamma$, $\beta$, and the dihedrals in the sugar ring to the corresponding $^3J$-coupling signal between coupling protons. Extensive molecular simulations are performed on a set of RNA tetramers and hexamers and combined with available nucleic-magnetic-resonance data. Within the new framework, the sampled structural dynamics can be reweighted to match experimental data while the error arising from inaccuracies in the forward models can be corrected simultaneously and consequently does not leak into the reweighted ensemble. Carefully crafted cross-validation procedure and regularization terms enable obtaining transferable Karplus parameters. Our approach identifies the optimal regularization strength and new sets of Karplus parameters balancing good agreement between simulations and experiments with minimal changes to the original ensemble.