Improving Conformational Ensembles of Folded Proteins in GoMartini

The Martini coarse-grained (CG) force field enables efficient simulations of biomolecular systems but cannot reliably maintain folded protein structures. To stabilize proteins during simulation, Martini is typically combined with structure-based force fields such as elastic network models (ENMs) or Go̅ models. While these approaches preserve global folds and capture protein flexibility, their ability to reproduce conformational dynamics remains unclear. Here, we evaluate Martini 3 combined with ENMs or Go̅ models on three folded proteins and show that both approaches struggle to sample the conformational space observed in atomistic simulations, even when uniform interaction strengths or equilibrium bond distances are adjusted. This limitation arises from the assumption of a uniform interaction network, in which all Go̅-bonds are assigned the same ϵ value, and therefore have the same potential depth. To overcome this, we present a fully automated, perturbation-based optimization approach for Go̅ networks, PoGo̅, that iteratively refines a nonuniform Go̅ network against a precomputed atomistic free-energy landscape in essential conformational space. Moreover, we demonstrate that our approach can also be used to optimize ENMs. In both cases, convergence is rapid and yields CG ensembles in close agreement with reference atomistic simulations. As a cross-validation, the optimization also improves the root-mean-square fluctuation profile.