Fast training of bespoke SMIRNOFF-format molecular mechanics force fields using machine learning potentials

Molecular mechanics force fields enable the thorough sampling required for applications such as alchemical binding free energy calculations for computer-aided drug design. However, parameters from transferable molecular mechanics force fields can be unreliable, and training accurate molecule-specific parameters against quantum mechanical (QM) reference data is slow. Here, we present presto, a Python package for fast training of bespoke SMIRKS-Native Open Force Field (SMIRNOFF) format force fields. presto generates training data using high-temperature molecular dynamics with well-tempered metadynamics. It iteratively fits all valence parameters to energies and forces from transferable machine learning potentials (MLPs). By leveraging fast MLP evaluations and training on the GPU using PyTorch (via the packages smee and descent), the fully-automated workflow completes in around 15 minutes for a 50-atom molecule on a single GPU. presto reduces torsion scan errors by a factor of three relative to Open Force Field Sage 2.3.0 on 400 molecules from TorsionNet500. It also achieves torsion scan performance comparable to direct QM fitting on a benchmark dataset of fragmented drug-like molecules, and relative conformer energies of 510 molecules from the set of Folmsbee and Hutchison are significantly improved compared to both Sage and espaloma. presto allows simultaneous fitting of congeneric series with partially shared parameters. It produces comparable relative binding free energy performance to OpenFF 1.3.1 (Parsley) on systems where Parsley already performs well, and may improve performance in cases where initial parameters are particularly poor.