Automated Training of ReaxFF Reactive Force Fields for Energetics of Enzymatic Reactions

Computational
studies of the reaction mechanisms of various enzymes
are nowadays based almost exclusively on hybrid QM/MM models. Unfortunately,
the success of this approach strongly depends on the selection of
the QM region, and computational cost is a crucial limiting factor.
An interesting alternative is offered by empirical reactive molecular
force fields, especially the ReaxFF potential developed by van Duin
and co-workers. However, even though an initial parametrization of
ReaxFF for biomolecules already exists, it does not provide the desired
level of accuracy. We have conducted a thorough refitting of the ReaxFF
force field to improve the description of reaction energetics. To
minimize the human effort required, we propose a fully automated approach
to generate an extensive training set comprised of thousands of different
geometries and molecular fragments starting from a few model molecules.
Electrostatic parameters were optimized with QM electrostatic potentials
as the main target quantity, avoiding excessive dependence on the
choice of reference atomic charges and improving robustness and transferability.
The remaining force field parameters were optimized using the VD-CMA-ES
variant of the CMA-ES optimization algorithm. This method is able
to optimize hundreds of parameters simultaneously with unprecedented
speed and reliability. The resulting force field was validated on
a real enzymatic system, ppGalNAcT2 glycosyltransferase. The new force
field offers excellent qualitative agreement with the reference QM/MM
reaction energy profile, matches the relative energies of intermediate
and product minima almost exactly, and reduces the overestimation
of transition state energies by 27–48% compared with the previous
parametrization.