BPS2026 – Protein binding landscapes with enhanced sampling-driven coarse-graining

Coarse-graining (CG) of proteins enables the simulation of large biomolecular assemblies over long timescales, yet a long-standing challenge remains in accurately reproducing binding-free energies. Conventional bottom-up CG methods, such as force-matching and relative entropy minimization (REM), often train CG models without sufficient sampling of protein binding events, and therefore fail to capture the correct all-atom (AA) thermodynamics of protein association. Conversely, ad hoc models can be tuned to fit binding affinities but lack rigorous microscopic foundations. To bridge this gap, we introduce an enhanced sampling-driven coarse-graining (ESCG) framework that directly incorporates binding potentials of mean force (PMFs) into CG parameterization. By combining PMF matching and REM on reweighted ensembles, ESCG yields models that reproduce both structural statistics and binding free-energy landscapes with systematic fidelity. We demonstrate this method on the barnase-barstar system, showing that ESCG accurately recovers AA PMFs and binding affinities, with the CG PMF further validated by independent enhanced sampling methods. Moreover, our method remains robust even in the absence of atomistic enhanced sampling, as it can directly reproduce a top-down or experimentally derived reference PMF, thereby substantially reducing the computational cost. We illustrate this capability with an Ezrin-membrane binding system, demonstrating the method’s effectiveness, efficiency, and broad generalizability. Our results establish ESCG as a useful strategy for constructing thermodynamically faithful CG models of protein association, paving the way for applications to complex biomolecular assemblies.