Prediction of octanol/water partition coefficient (K_OW) with algorithmically derived variables

A statistical model was developed with algorithmically derived independent variables based on chemical structure for prediction of octanol/water partition coefficients (K_ow) measured for more than 4,000 chemicals. The procedure first classified the chemicals into 14 groups based on the number of hydrogen bonds, and then best‐subsets, multiple‐regression analysis was used to predict K_ow within groups. In addition, a training set/test set approach was used to provide an independent evaluation of the sensitivity of the model to the number of chemicals and variables used within each group. In general, the explained variation (r²) was higher and the standard error of the estimates (see) lower in the training sets as compared with the test set groups, whereas analyses of the combined data sets were generally intermediate. Explained variation among the 14 groups, using the combined data sets, ranged from 63 to 90%, and see ranged from 0.37 to 0.78 in logarithmic units. Plots of the residuals indicated a normal scatter. These results are similar to reported error rates in other models.