Molecular simulations offer insight into the structures and functions of biomolecules and are used to help design drugs and materials, but their predictions remain unreliable, and improvements will require more accurate force fields. Though valuable, today’s leading force fields represent difficult starting points for improvement, because it is not clear which aspects are already well-optimized and which are not. Our long-term goal is a systematic approach to developing new force fields from the ground up.<br/>Here we introduce a SMIRKS-based Force Field format, which allows parameter assignment via chemical substructure search instead of conventional atom typing. Different chemical perception can be set for different classes of parameter; e.g., nonbondeds, bonds, angles, and dihedrals. This approach provides two major benefits. First, it can simplify how force fields are constructed and reduce the number of parameters required, while enabling greater generality. For example, a SMIRKS-based force field will allow recognizing and adding a needed torsion type without adding a new atom type. Second, SMIRKS perception rules can be changed and tested on-the-fly, so they enable iterative sampling and optimization of the chemical perception criteria, in order to best and most parsimoniously reproduce the reference data.<br/><br/>More broadly, this work provides a new foundation for systematic force field development. We are developing a Bayesian framework that will automatically generate force fields with types and parameters consistent with the reference data, given a preselected functional form. This approach will also assign uncertainties to the parameters, and these may be used to estimate the uncertainties in simulation results.

Toward more accurate force fields through direct chemical perception

253rd National Meeting (2017)

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