Molecular mechanics force fields define how the energy and forces of a molecular system are computed from its atomic positions, and enable the study of such systems through computational methods like molecular dynamics and Monte Carlo simulations. Despite progress toward automated force field parameterization, considerable human expertise is required to develop or extend force fields.
In particular, human input has long been required to define atom types, which encode chemically unique environments that determine which parameters must be assigned. However, relying on humans to establish atom types is suboptimal: the resulting atom types are often unjustified from a statistical perspective, leading to over- or under-fitting; they are difficult to extend in a systematic and consistent manner when new chemistries must be modeled or new data becomes available; and human effort is not scalable when force fields must be generated for new (bio)polymers or materials. We aim to replace human specification of atom types with an automated approach, based on solid statistics and driven by experimental and/or quantum chemical reference data. Here, we describe a novel technology for this purpose, termed SMARTY, which generalizes atom typing by using direct chemical perception with SMARTS strings, and adopting a hierarchical approach to type assignment. The SMARTY technology enables creation of a move set in atom-typing space that can be used in a Monte Carlo optimization approach to atom typing. We demonstrate the power of this approach with a fully automated procedure that is able to re-discover human-defined atom types in the traditional small molecule force field parm99/parm@Frosst. Furthermore, we show how an extension of this approach that makes use of SMIRKS strings to match multiple atoms, which we term SMIRKY, allows us to take full advantage of the advances in direct chemical perception for valence types (bonds, angles, and torsions) afforded by the recently-proposed SMIRNOFF direct chemical perception force field typing language. We assess these approaches using several molecular datasets, including one which covers a diverse molecular subset from DrugBank.