Evaluation of Force-Field Calculations of Lattice Energies on a Large Public Dataset, Assessment of Pharmaceutical Relevance, and Comparison to Density Functional Theory

Crystal
lattice energy is a key property affecting the ease of
processing pharmaceutical materials during manufacturing, as well
as product performance. We present an extensive comparison of 324
force-field protocols for calculating the lattice energies of single
component, organic molecular crystals (further restricted to Z′ less than or equal to one), corresponding to a
wide variety of force-fields (DREIDING, Universal, CVFF, PCFF, COMPASS,
COMPASSII), optimization routines, and other variations, which could
be implemented as part of an automated workflow using the industry
standard Materials Studio software. All calculations were validated
using a large new dataset (SUB-BIG), which we make publicly available.
This dataset comprises public domain sublimation data, from which
estimated experimental lattice energies were derived, linked to 235
molecular crystals. Analysis of pharmaceutical relevance was performed
according to two distinct methods based upon (A) public and (B) proprietary
data. These identified overlapping subsets of SUB-BIG comprising (A)
172 and (B) 63 crystals, of putative pharmaceutical relevance, respectively.
We recommend a protocol based on the COMPASSII force field for lattice
energy calculations of general organic or pharmaceutically relevant
molecular crystals. This protocol was the most highly ranked prior
to subsetting and was either the top ranking or amongst the top 15
protocols (top 5%) following subsetting of the dataset according to
putative pharmaceutical relevance. Further analysis identified scenarios
where the lattice energies calculated using the recommended force-field
protocol should either be disregarded (values greater than or equal
to zero and/or the messages generated by the automated workflow indicate
extraneous atoms were added to the unit cell) or treated cautiously
(values less than or equal to −249 kJ/mol), as they are likely
to be inaccurate. Application of the recommended force-field protocol,
coupled with these heuristic filtering criteria, achieved an root
mean-squared error (RMSE) around 17 kJ/mol (mean absolute deviation
(MAD) around 11 kJ/mol, Spearman’s rank correlation coefficient
of 0.88) across all 226 SUB-BIG structures retained after removing
calculation failures and applying the filtering criteria. Across these
226 structures, the estimated experimental lattice energies ranged
from −60 to −269 kJ/mol, with a standard deviation around
29 kJ/mol. The performance of the recommended protocol on pharmaceutically
relevant crystals could be somewhat reduced, with an RMSE around 20
kJ/mol (MAD around 13 kJ/mol, Spearman’s rank correlation coefficient
of 0.76) obtained on 62 structures retained following filtering according
to pharmaceutical relevance method B, for which the distribution of
experimental values was similar. For a diverse set of 17 SUB-BIG entries,
deemed pharmaceutically relevant according to method B, this recommended
force-field protocol was compared to dispersion corrected density
functional theory (DFT) calculations (PBE + TS). These calculations
suggest that the recommended force-field protocol (RMSE around 15
kJ/mol) outperforms PBE + TS (RMSE around 37 kJ/mol), although it
may not outperform more sophisticated DFT protocols and future studies
should investigate this. Finally, further work is required to compare
our recommended protocol to other lattice energy calculation protocols
reported in the literature, as comparisons based upon previously reported
smaller datasets indicated this protocol was outperformed by a number
of other methods. The SUB-BIG dataset provides a basis for these future
studies and could support protocol refinement.