An OpenMM-BasedML/MM–MMGBSA Workflow for End-pointProtein–Ligand Binding Energy Ranking

Machine-learning force fields provide a practical route to introduce quantum-derived potential-energy surfaces into molecular dynamics simulations, but their value in end-point protein–ligand binding energy workflows remains unclear. Here, we developed an OpenMM-based ML/MM–MMGBSA workflow by combining ligand-only mechanically embedded ML/MM hybrid simulations and MMPBSA.py analysis, and evaluated it on 199 protein–ligand complexes from eight targets in the JACS benchmark set. The results showed that the effect of ML/MM sampling on MM/GBSA ranking was strongly target-, MLFF-, and sampling-time-dependent; selected improvements were observed in several systems, but simply extending the production length did not consistently improve end-point ranking. Pairwise sign accuracy analysis further showed that trajectory-based MM/GBSA provided useful ranking performance relative to single-structure Prime MMGBSA, although alchemical RBFE methods remained more robust. Overall, this workflow provides a practical and reproducible framework for integrating ML/MM conformational sampling with end-point binding free-energy estimations, while highlighting the need for MLFFs compatible with broader chemical space and explicit-solvent biomolecular environments, improved hybrid simulation embedding schemes such as electrostatic or polarizable embedding, and better consistency between sampling and scoring models.