Water model optimization

Problem Statement

OpenFF force fields descend from AMBER, and therefore the recommended water model for use so far has been TIP3P. Many flaws of TIP3P have been documented; for example, it underestimates the density of water and overestimates the self-diffusion coefficient (Vega and Abascal, 2011). The water model is known to substantially impact the calculation of simulation properties; for example, salt-effects varied by over 7 kcal/mol between different water models in host-guest binding enthalpy calculations (Gao et al., 2015). In benchmarking a force field candidate, Chapin Cavender furthermore observed helices unwinding in simulations of GB3 in TIP3P water, which did not occur in simulations using OPC water. Finally, systematic errors in simulated properties using the recently re-fit Sage force field (Boothroyd et al., 2023) suggest that improvements in the water model may improve performance on physical parameters. Therefore, our goal is to co-optimize a water model and accompanying ion parameters for release with the Rosemary protein force field.

Background and considerations

How many sites?

A number of water model architectures exist (e.g. 3-site, 4-site, 5-site models). It is unclear which architecture would work best for OpenFF force fields. We note that OPC is a 4-site model; however, a co-optimized 3-site model may be of sufficient quality as well. In general OpenFF would prefer to release a co-optimized 3-site model, for reasons of computational expense and simplicity. However, this project should investigate at least a 4-site model to see whether the additional site is necessary. There is a possibility of releasing and maintaining two lines of force field: one with virtual sites and a 4-site model, and one without virtual sites and a 3-site model.

Water properties

A water model comprises both bonded/valence parameters (angles, bonds) and nonbonded parameters (LJ, electrostatics).

Valence

One direction we can follow is to choose either the angle/bond parameters of TIP3P or another existing water model. Another is to re-fit our own. For example, we could follow the workflow set out by Izadi et al. when building the OPC water model.

Ion parameters

Ions can be parameterised following the methods of Joung and Cheatham (2008) or Li et al. (2020), that is: fitting to free energies from Schmid or Marcus. Currently OpenFF Sage only supports monovalent ion parameters; however, divalent ions are common in the biomolecular simulations we expect that Rosemary will apply to. Depending on our available computational resources we may only fit to monovalent ions; however, we should eventually look towards including common divalent ions as well.

Computational resources

UC Irvine

A water model re-fit to the same properties as in the DEXP paper ( ) took ~21 hours on HPC3 for a single iteration. This was 21 hours on the standard queue, with work_queue_workers running on the free-gpu queue. It is therefore likely that fitting to a wider array of properties, even up to the total used in the TIP3P-FB paper, is doable fairly cheaply on free resources at UCI.

Milestones

Actions

Notes

References