Repeatability of Relative Free Energy Calculations in Solution with ANI-2x and MACE-OFF23.

We investigate the feasibility and challenges of using neural network potentials (NNPs) for alchemical free energy calculations, employing a single-coordinate dual-topology approach. As a model application, we compute free energy differences between tautomer pairs to predict the preferred tautomeric state in aqueous solution. A central aspect of our approach is based on energy mixing via the selective masking of interactions involving dummy atoms, enabling a smooth interpolation between tautomeric states. This methodology is independent of the specific NNP architecture and holds potential for broader application to larger alchemical transformations. We tested this framework using two well-known NNPs: ANI-2x and MACE-OFF23(small). While MACE-OFF23(small) produced converged free energy results, simulations with ANI-2x showed significant variability across repeated runs. Our analysis traced this inconsistency to slow water dynamics and the overstabilization of artificial metastable states of the solute under ANI-2x, causing difficulties in converging sampling. Although transferable NNPs offer the advantage of general applicability without system-specific parametrization, our findings emphasize the importance of evaluating their performance in the condensed phase before employing them for free energy simulations.