Development and benchmarking of an open, self-consistent force field for proteins and small molecules from the open force field initiative.

Abstract

Contemporary protein force fields model canonical proteins well, but parameters for covalently modified residues are challenging to develop. Additionally, protein-specific parameters are derived separately from those for small molecules, limiting the transferability of small molecule parameters to protein contexts. A flexible force field that provides self-consistent parameters for proteins and small molecules would advance research in drug design and molecular mechanisms of protein structure and function. The Open Force Field (OpenFF) Initiative develops open, reproducible force fields for atomistic simulations, delivering automated infrastructure and systematic methodology for force field fitting and validation as well as version-controlled force field releases. OpenFF force fields assign parameters using direct chemical perception via the SMIRKS cheminformatics language, providing coverage of a large chemical space with fewer parameters than conventional atom-typed force fields. The most recent force field release, Sage (OpenFF 2.0.0), provides valence parameters trained against quantum chemical (QC) data and Lennard-Jones parameters trained against condensed phase properties of pure liquids and binary mixtures. Sage achieves competitive accuracy on benchmarks of the energies and geometries of QC minima, solvation free energies, and protein-ligand binding free energies. Here we describe the extension of Sage to proteins, resulting in a self-consistent force field that can simulate canonical and covalently modified proteins and small molecules. We used protein-specific SMIRKS to train proper torsions to two-dimensional QC scans of the backbone and sidechain dihedrals of capped peptides. The parameters were validated using benchmarks of published NMR observables for small peptides, folded proteins, and disordered proteins. Beyond force field parameters, the OpenFF software infrastructure now supports loading proteins from PDB files, iterating over hierarchies such as residues, and exporting parametrized systems to common molecular dynamics formats.