Prediction of sweetness by multilinear regression analysis and support vector machine

The sweetness of a compound is of large interest for the food additive industry. In this work, 2 quantitative models were built to predict the logSw (the logarithm of sweetness) of 320 unique compounds with a molecular weight from 132 to 1287 and a sweetness from 22 to 22500000. The whole dataset was randomly split into a training set including 214 compounds and a test set including 106 compounds, represented by 12 selected molecular descriptors. Then, logSw was predicted using a multilinear regression (MLR) analysis and a support vector machine (SVM). For the test set, the correlation coefficients of 0.87 and 0.88 were obtained by MLR and SVM, respectively. The descriptors found in our quantitative structure-activity relationship models are prone to a structural interpretation and support the AH/B System model proposed by Shallenberger and Acree.