Prediction of bitterant and sweetener using structure-taste relationship models based on an artificial neural network

• Published in: Food research international 2022-03, Vol.153, p.110974-110974, Article 110974
• Main Authors: Bo, Weichen, Qin, Dongya, Zheng, Xin, Wang, Yue, Ding, Botian, Li, Yinghong, Liang, Guizhao
• Format: Article
• Language: English
• Subjects: Aversive Agents; Bitterant prediction, Sweetener prediction; Convolutional neural networks (CNN); Multi-layer perceptron (MLP); Neural Networks, Computer; Structure-taste relationship; Sweetening Agents; Taste
• ISSN: 0963-9969; 1873-7145
• DOI: 10.1016/j.foodres.2022.110974

[Display omitted] •The artificial neural network models were proposed to predict bitterants/sweeteners.•The multi-layer perceptron-Fingerprint model exhibited the best prediction results.•The convolutional neural network model automatically extracted molecular features.•Molecular hydrophobicity, weight and charge could identify bitterants/sweeteners. Identifying the taste characteristics of molecules is essential for the expansion of their application in health foods and drugs. It is time-consuming and consumable to identify the taste characteristics of a large number of compounds through experiments. To date, computational methods have become an important technique for identifying molecular taste. In this work, bitterant/non-bitterant, sweetener/non-sweetener, and bitterant/sweetener are predicted using three structure-taste relationship models based on the convolutional neural networks (CNN), multi-layer perceptron (MLP)-Descriptor, and MLP-Fingerprint. The results showed that all three models have unique characteristics in the prediction of bitterant/non-bitterant, sweetener/non-sweetener, and bitterant/sweetener. For the prediction of bitterant/non-bitterant, sweetener/non-sweetener, and bitterant/sweetener, the MLP-Fingerprint model exhibited a higher predictive AUC value (0.94, 0.94 and 0.95) than the MLP-Descriptor model (0.94, 0.84 and 0.87) and the CNN model (0.88, 0.90 and 0.91) by external validation, respectively. The MLP-Descriptor model showed a distinct structure-taste relationship of the studied molecules, which helps to understand the key properties associated with bitterants and sweeteners. The CNN model requires only a simple 2D chemical map as input to automate feature extraction for favorable prediction. The obtained models achieved accurate predictions of bitterant/non-bitterant, sweetener/non-sweetener and bitterant and sweetener, providing vital references for the identification of bioactive molecules and toxic substances.