E-state modeling of corticosteroids binding affinity validation of model for small data set

Data for 31 steroids binding to the corticosteroid binding globulin (CBG) were modeled using E-state molecular structure descriptors and a kappa shape index. Both E-state and hydrogen E-state descriptors appear in the model in atom-level and atom-type descriptors. A four-variable model is obtained that is statistically satisfactory: r(2) = 0.81, s = 0.51; r(press)(2) = 0.72; s(press) 0.62. Structure interpretation is given for each variable in the model. A leave-group-out (LGO) approach to model-validation is presented in which each observation is removed from the data set three times in random groups of 20% of the whole data set. The average of the resulting predicted values constitutes consensus predictions for these data for which r(LOO)(2) = 0.70. These collective results support the claim that the E-state model may be useful for prediction of pK binding values for new compounds.