2020 Virtual AIChE Annual Meeting

(346ax) Open Force Field Initiative: New Strategies for Parameterizing Non-Bonded Interactions

Authors

Madin, O. - Presenter, University of Colorado Boulder
Boothroyd, S., University of Colorado Boulder
Shirts, M., University of Colorado Boulder
The Open Force Field Initiative (OpenFF) is a collaborative effort from academia and industry focused on delivering better biomolecular force fields through innovative science and data curation. One of the major areas we aim to improve is the treatment of van der Waals interactions. These interactions are crucial to modeling many biologically relevant processes, like protein folding, protein-ligand binding, and membrane transport. Modeling of these interactions is difficult; they are generally fitted against condensed-phase property data, which creates a “black box” problem where the relationship between model parameters and fitting targets is not entirely clear.

The development of the OpenFF Evaluator software package, coupled with datasets sourced from the NIST ThermoML database, has accelerated our fitting process dramatically. This software provides a framework for streamlined property estimation, fitting and benchmarking. Heat of vaporization is a common fitting target for vdW models, but it is problematic, due to differences in polarization between the liquid and gas phase, as well as limited data availability. Using the Evaluator software, we fit vdW interactions against combinations of pure density, heat of vaporization, mixture density, and enthalpy of mixing for a variety of organic species. Results show that fitting against mixture properties can yield improvements in vdW modeling, and that enthalpy of mixing data may be an effective replacement for heat of vaporization.

We also explore methods to rationally choose models for vdW interactions. In addition to continuous choices about models, there are also discrete choices of functional forms: number of atom types, combination rules, and vdW functional forms. We demonstrate the use of Bayesian inference as a tool to rationally select between different models, or levels of complexity. For a simple 2CLJQ fluid model, Bayesian inference gives evidence that, for selected physical properties, a quadrupole interaction is unnecessary. We then discuss strategies and progress towards building surrogate models that will enable this analysis for more complex biomolecular force fields.