Adaptive Hydration-Shell ML/MM Simulations Recover Full ML Accuracy in Biomolecular Conformational Description

Residue-specific backbone conformational preferences govern the conformational ensembles of intrinsically disordered proteins, yet their accurate description using classical force fields is limited by prescribed functional forms that cannot fully represent environment-dependent many-body interactions. Machine-learned force fields (MLFFs) learn complex atomic interactions directly from quantum-mechanical reference data, providing a systematic alternative to classical parameterisations. Here, we employ a pretrained MLFF to simulate nine cationic Gly-x-Gly tripeptides in water and compare the resulting conformational ensembles against experimental NMR scalar J-couplings. To improve computational efficiency, we develop and assess hybrid Machine Learning/Molecular Mechanics (ML/MM) protocols as computationally tractable alternatives to full ML simulation. Mechanical-embedding ML/MM is found insufficient for accurate conformational sampling, as the classical treatment of peptide–solvent interactions fails to adequately represent solvation effects on backbone conformational preferences. We therefore implement an adaptive ML/MM scheme that incorporates the immediate solvation environment into the ML description and recovers near-full-ML accuracy at a reduced computational cost.