Atom/fragment contribution method for estimating octanol-water partition coefficients

Abstract

Atom/fragment contribution values, used to estimate the log octanol-water partition coefficient (log P) of organic compounds, have been determined for 130 simple chemical substructures by a multiple linear regression of 1120 compounds with measured log P values. An additional 1231 compounds were used to determine 235 "correction factors" for various substructure orientations. The log P of a compound is estimated by simply summing all atom/fragment contribution values and correction factors occurring in a chemical structure. For the 2351 compound training set, the correlation coefficient (r2) for the estimated vs measured log P values is 0.98 with a standard deviation (SD) of 0.22 and an absolute mean error (ME) of 0.16 log units. This atom/fragment contribution (AFC) method was then tested on a separate validation set of 6055 measured log P values that were not used to derive the methodology and yielded an r2 of 0.943, an SD of 0.408, and an ME of 0.31. The method is able to predict log P within +/- 0.8 log units for over 96% of the experimental dataset of 8406 compounds. Because of the simple atom/fragment methodology, "missing fragments" (a problem encountered in other methods) do not occur in the AFC method. Statistically, it is superior to other comprehensive estimation methods.