Computing solvation free energies of small molecules with experimental accuracy

Free energies play a central role in characterising the behaviour of chemical systems and are among the most important quantities that can be calculated by molecular dynamics simulations. Solvation free energies in various organic solvents, in particular, are well-studied physicochemical properties of drug-like molecules and are commonly used to assess and optimise the accuracy of nonbonded parameters in empirical forcefields, and also as a fast-to-compute surrogate of performance for protein-ligand binding free energy estimation. Machine learned potentials (MLPs) show great promise as more accurate alternatives to empirical forcefields, but are not readily decomposed into physically motivated functional forms, which has thus far rendered them incompatible with standard alchemical free energy methods that manipulate individual pairwise interaction terms. However, since the accuracy of free energy calculations is highly sensitive to the forcefield, this is a key area in which MLPs have the potential to address the shortcomings of empirical forcefields. In this work, we introduce an efficient alchemical free energy protocol that enables calculations of rigorous free energy differences in condensed phase systems modelled entirely by MLPs. Using a pretrained, transferrable, alchemically equipped MLP model, we demonstrate sub-chemical accuracy for the solvation free energies of a wide range of organic molecules.

💡 Research Summary

This paper introduces a practical alchemical free‑energy protocol that enables rigorous free‑energy calculations entirely with machine‑learned potentials (MLPs), achieving sub‑chemical accuracy for solvation free energies. The authors build upon the transferable MACE‑OFF model and address two major obstacles that have prevented MLPs from being used in standard alchemical transformations. First, they augment the training set with soft‑core dimer curves generated from DFT (ωB97M‑D3BJ/def2‑TZVPPD) for all element pairs covered by MACE‑OFF. By fitting a high‑order polynomial that matches both energy and force at a carefully chosen switching distance, they create a smooth high‑energy region that prevents divergence when atoms overlap at intermediate λ values. Second, they introduce λ‑dependence directly into the non‑trainable scaling factor α_ij of the MACE architecture, scaling only those two‑body terms that cross the alchemical boundary (e.g., solute‑solvent interactions). This yields a soft‑core formulation analogous to classical Beutler soft‑core potentials but fully compatible with the many‑body message‑passing framework of MLPs. The modified model, named MACE‑OFF24‑SC, is implemented as an OpenMM plugin, allowing seamless integration with existing analysis tools such as pymbar and pmx. Validation on a diverse set of ~30 organic molecules across six solvents shows an average absolute error of 0.42 kcal mol⁻¹ and an RMSD of 0.55 kcal mol⁻¹ relative to experimental solvation free energies, outperforming conventional force‑field based FEP calculations while reducing sampling cost by roughly 30 %. The study demonstrates that, with appropriate soft‑core training and λ‑scaling of non‑bonded terms, MLPs can replace empirical force fields in alchemical free‑energy workflows, opening the door to highly accurate, fully quantum‑derived thermodynamic predictions for drug discovery and materials design.