Testing the stability of Neural Network Potentials for free energy calculations

To achieve chemical accuracy in free energy calculations, it is necessary to accurately describe the system's potential energy surface and efficiently sample configurations from its Boltzmann distribution. While neural network potentials (NNPs) have shown significantly higher accuracy than classical molecular mechanics (MM) force fields, they have limited range of applicability and are significantly slower than MM potentials, often by orders of magnitude. To address this challenge, Rufa et al suggested a two-stage approach that uses a fast and established MM alchemical energy protocol, followed by reweighting the results using NNPs, known as endstate correction or indirect free energy calculation. In this study, we systematically investigate the accuracy, robustness, and efficiency of free energy calculations between a MM reference and a neural network target potential (ANI-2x) using free energy perturbation (FEP) and non-equilibrium (NEQ) switching simulation in vacuum. We investigate the influence of longer switching lengths and the impact of slow degrees of freedom on outliers in the work distribution. We compare the results to multistage equilibrium free energy calculations (MFES) for an established dataset. Our results demonstrate that free energy calculations between NNPs and MM potentials should not be performed using FEP but require NEQ switching simulations to obtain accurate free energy estimates. NEQ switching simulations between MM and NNPs are highly efficient, robust, and trivial to implement.