Optimizing general molecular mechanics force fields in chemical space using binary-encoded SMARTS

General force fields for molecular mechanics are challenging to design due to the large
chemical space the parameters must describe accurately. Recently, a new scheme of
parameter assignment using SMARTS substructure searches was constructed by hand,
where each SMARTS pattern was carefully designed. Further refinement of these
parameters is difficult due to the complexity and expressiveness of the SMARTS
grammar, and as such many SMARTS modifications only lead to marginal gains in
force field performance. Finding the correct SMARTS patterns to split into new
parameters is tedious and prone to error, and therefore requires significant human time.
This work automates the search for SMARTS patterns to derive new parameters,
alleviating the need to derive complex, possibly overfitting and nontransferable
SMARTS patterns. The parameters are selected based on their quality of fit to a QM
reference dataset. To perform the search, SMARTS patterns are first transformed to
bit maps embedded into graphs, and bits corresponding to the SMARTS information
are then iterated in a deterministic fashion. The search aims to find the most general
SMARTS patterns as possible in effort to produce general, transferable force fields.
Here we show 3 use cases of constructing a set of parameters from scratch: charges,
bonds and angles, and torsions. We then show how this work can be used as a general
SMARTS clustering tool by converting a set of assignments based on GAFF into a
force field in the SMIRNOFF format which uses SMARTS patterns as the parameter
assignment mechanism.