Molecular dynamics simulations of protein-ligand systems have the potential to transform drug discovery through the accurate prediction of ligand binding conformations and energies. While protein force fields only require the modeling of a relatively limited set of functional groups, general-purpose ligand force fields should provide a consistently accurate model across a large and diverse chemical space. To make the development of a general-purpose ligand force field tractable, we typically parameterize the model according to a predefined set of “atom types.” These atom types are used to group atoms in different ligand molecules (but similar chemical environments) such that they receive the same force field parameters. Exhaustive exploration of all atom type definitions is computationally infeasible, so ligand force fields have traditionally depended on human intuition. Here, we develop an automated and performant strategy to generate atom types by combining the enumeration of chemical environments and properties with Least Absolute Shrinkage and Selection Operator (LASSO) and other modern statistical methods for graph clustering and model selection. We apply our strategy by training atom type definitions using a large data set of noncovalent quantum chemical interactions and then validate the quality of the resultant force field (based on these atom types) according to out-of-sample condensed phase data. We show that the condensed phase performance is comparable to an otherwise identical force field fit with a similar number of carefully tuned human-derived atom types. We also examine the generated atom types to understand which chemical environments and properties are most crucial for the accurate representation of ligand chemistries. Our results suggest that our automated strategy for defining atom types should aid in the rapid development of more performant ligand force fields.

LASSOing the atom-typing problem: Statistical method for ligand force field model selection

2019 Fall National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.