Quantitative structure–activity relationships to predict sweet and non-sweet tastes

The aim of this work was the calibration and validation of mathematical models based on a quantitative structure–activity relationship approach to discriminate sweet, tasteless and bitter molecules. The sweet-tasteless and the sweet-bitter datasets included 566 and 508 compounds, respectively. A total of 3763 conformation-independent Dragon molecular descriptors were calculated and subsequently reduced through both unsupervised reduction and supervised selection coupled with the k-nearest neighbors classification technique. A model based on nine descriptors was retained as the optimal one for sweet and tasteless molecules, while a model based on four descriptors was calibrated for the sweetness-bitterness dataset. Models were properly validated through cross-validation and external test sets. The applicability domain of models was investigated, and the interpretation of the role of the molecular descriptors in classifying sweet and non-sweet tastes was evaluated. The classification and the performance of the models presented in this paper are simple but accurate. They are based on a relatively small number of descriptors and a straightforward classification approach. The results presented here indicate that the proposed models can be used to accurately select new compounds as potential sweetener candidates.