In-silico prediction of sweetness of sugars and sweeteners

Three quantitative models for the prediction of sweetness of 103 compounds were developed. These compounds include 29 sugars and 74 common sweeteners, whose sweetness is in the range of 22–300,000 and the molecular weight is from 122 to 1287. The molecules were represented by three descriptors. On the basis of the Kohonen’s Self-Organising Neural Network (KohNN) map, the whole data set was split into a training set including 58 compounds and a test set including 45 compounds. Then, logSw was predicted by using a Multi-Linear Regression (MLR) analysis, an Artificial Neural Network (ANN) analysis and a Support Vector Machine (SVM) regression analysis. For the test set, the correlation coefficient of 0.925, 0.932 and 0.943 for the MLR, ANN and SVM were achieved, respectively.

► A dataset of 103 compounds including sugars and sweeteners was used.
► Molecules were represented by calculated descriptors according to their structures.
► Three molecular descriptors were selected for building the sweetness QSAR model.
► The selected descriptors contain the information of 2D molecular bond distances.
► Three accurate QSAR models have been built for predicting sweetness.