The partition coefficients for neutral species of 18 quinazolones and
related compounds [1], as calculated by the ACD/Labs algorithm, were found to have a larger than usual deviation from experimental
values, with an average deviation of 1.1 - 2.2 logP units, depending on the presence of a ketone or methoxy substituent on
the pyrimidine ring. Modification of the logP prediction algorithm, to permit "system training" on two representative
species, was found to lead to a four-fold improvement in the accuracy of prediction; the average discrepancy dropped to
0.30 - 0.32 logP units.
The system training method was then extended to two sets of logD predictions.
In the first case [2], the average discrepancy predicted for a set of 18 phenylpiperidine analogs was found to be 0.94 for logD
calculations with the "generic" set of parameters. When system training was implemented using a single compound (structure VII),
the accuracy was improved and the average discrepancy dropped to 0.38.
In the second case [3] a set of 10 phenylquinolines had an average logD discrepancy of
1.39 logD units. With system training - again using a single compound (structure VIII), the average discrepancy dropped to 0.33
logD units.
For both the neutral and the dissociative partition coefficients, it is evident
that system training based on the simplest analog of a set of compounds will ensure greater accuracy. The greatest source of
error then becomes that of the experimental value. This suggests that experiment need only consider the highly accurate
determination of a single point, for the simplest analog. The remainder of the partition coefficients in the set of analogs
can be predicted with an uncertainty of ±0.3 units, even for highly challenging compounds.
|
