Structure-taste relationships for disubstituted phenylsulfamate tastants using classification and regression tree (CART) analysis

Forty-two new disubstituted phenylsulfamates have been synthesized, and 30 of these have been combined with 40 already available from earlier work to create a training database of 70 compounds. On the basis of panel taste data these were divided into three categories, N (nonsweet), N/S (nonsweet/sweet), and (S) sweet, and a "sweetness value" or weighting was also calculated for each compound. Using these 70 compounds as a training set and a series of nine predictors derived from Corey-Pauling-Koltun (CPK) models, calculated from the PC SPARTAN PRO program and Hammett σ values taken from the literature, a classification and regression tree analysis (CART) was carried out leading to a regression tree that correctly classified 62 of the 70 compounds (89% overall correct classification). The tree's predictive ability varies for the different taste categories, and for nonsweet compounds it is virtually 100%; for nonsweet/sweet compounds it is 66%, and for sweet compounds it is ≈75%. This tree correctly predicted taste categories for 10 compounds from a test set of 12 randomly selected from among the 42 new compounds (83% correct classification). Therefore, it can be used with a good degree of confidence to predict the tastes of disubstituted phenylsulfamates. For the design of new sweeteners, appropriate values or ranges of the descriptors are derived.