Architecture independent absolute solvation free energy calculations with neural network potentials

Allowing atoms or molecules to disappear is a critical step in alchemical free energy simulations (FES). The necessary tricks are well understood when using force fields. Over the past few years, neural network potentials (NNPs) have seen rapid development. Their potentially higher accuracy compared to force fields makes them attractive for use in FES. Here, we outline a method for gradually decoupling atoms and molecules in systems that are fully described by NNPs. Specifically, we show that manipulating the neighbor list is equivalent to using soft-core potentials in force-field-based FES. Since constructing the neighbor list is a central step, regardless of the NNP's inner workings, our approach is agnostic to NNP architecture. We validate the correctness of our methodology by demonstrating cycle closure for a model problem and report solvation free energies obtained with the MACE-OFF23(S/M) NNP.