A recently-published comparison [1] of experiment vs. theory for the octanol-water
partition coefficient for a set of 18 quinazolones and related compounds provides a useful basis for examining
the accuracy of the ACD/LogP additive-constitutive algorithm. The discrepancies of logP (partition coefficient
in octanol-water) for this set of compounds were found to be unusually large. A search through the database from
which the logP values are derived showed the class of compounds with fused heterocyclic rings was under-represented.
Thus, the set of quinazolones in Ref. [1] was deemed to be a good candidate to examine system training.
Nine of the compounds in the set have the general structure designated as (I)
and nine have the general structure (II). The R1, R2, ... R6 substituents were a combination of hydro,
ethyl, methyl, methoxy, bromo, isopropyl, and tolulyl fragments.
The starting basis for the system training were the compounds (III) and (IV).
These compounds represented the simplest versions of (I) and (II) for which experimental values were known:
Once a compound is designated as a "training structure" for logP user-training
prediction, the algorithm breaks its structure down into the most important components, and then re-scales its
parameter estimation based on the newly-identified fragments. The production of fragments, the re-scaling process,
and the determination of new additive increments are automatic and reproducible, and can be applied to any
other case.
The results for the logP calculation using "generic" parameters (i.e., no specialization
in the type of structures of interest) are summarized in Table 1. The calculated logP value has an average discrepancy
from experiment of 1.15 logP units. The effect of system training with user-supplied data is also summarized in that
table. The average discrepancy has been reduced to 0.30 logP units.
|
