Chemistry Reference
In-Depth Information
are given the connectivity formulae of a few compounds and are given a few
months to produce a list of predicted crystal structures ranked in order of likelihood
of appearance. The crystal structures are separately determined by X-ray diffraction
and are secreted until all predictions are submitted. Predictions are considered
successful if the experimental structure falls within the first three of the predicted
ones. The results are instructive and, when carefully analyzed, confirm the above
premises.
In the first test [
55
] candidate molecules were small and rigid (Fig.
10
) and
crystals were confined to belong in one of the most common space groups for
organic compounds. Computational methods were mostly restricted to simple
atom-atom formulations, i.e., quite unsuitable for an accurate task, or to more
accurate distributed multipole methods. A relatively high number of hits were
obtained, but were not distributed according to the accuracy of the procedure: the
software written by the present author scored one of the most accurate hits
even though prescriptions were not followed by the users who actually employed
a hybridized, uncalibrated force field; the energy ranking depended on fractions of
a kJ mol
1
. It was later disclosed that the corresponding compound possesses a
second, more stable polymorph that had not been detected in any of the predictions.
The relative enthusiasm spurred by the results of the first test were quenched by
those of the second test, when very few hits were recorded in spite of massive
updating and presumed amelioration of the procedures on the basis of the previous
experience. Things proceeded in ups and downs in the third test, but in the fourth
one a participant group scored an impressive four hits out of four targets [
56
]. The
apparently decisive factor was the adoption of an algorithm for exhaustive structure
search, together with an ab initio DFT dispersion-corrected procedure for both
intra- and intermolecular energies, at the price of an equivalent of several hundred
thousand CPU hours on massive parallel hardware. In the latest test (Price et al.
2010, personal communication), the same group scored four hits out of six targets,
missing what was thought the most difficult one, a giant molecule with at least six
torsional degrees of freedom (Fig.
11
). Unexpectedly, however, two other groups
working with much less ambitious computational means were successful with this
most elusive prey, but missed the much easier targets. One participant (B.P. van
Eijck) has a constant record of one or two hits in each trial using mostly highly
O
OH
Fig. 10 Two molecules
proposed as targets for crystal
structure prediction in the first
blindfold test
C
S
N
