Rapid Screening of Sweeteners through Cluster Analysis of Taste Attributes and Machine Learning

In this study, an electronic nose (E-tongue) and molecular docking were utilized to analyze the taste profiles of 12 representative sweeteners and their binding characteristics with sweet taste receptors. Principal component analysis (PCA) and cluster analysis (CA) categorized the sweeteners into 4 groups based on their taste profiles: Group 1, represented by glucose, exhibited a clean sweet taste; Group 2, represented by rebaudioside A, had significant astringency; Group 3, represented by mogroside Ⅲ, showed a pronounced sourness; and Group 4, represented by sucralose, possessed pronounced bitterness. According to the differences in binding affinity to the hT1R2 and hT1R3 receptors, the sweeteners were also classified into four categories. Next, partial least squares regression, random forest regression, and support vector regression were used to build unimodal prediction based on the data of E-tongue and molecular docking, separately. Feature-level fusion was carried out on these models to construct a bimodal prediction model connected by molecular features. Maltitol and isomalt oligosaccharide were identified as sugar substitutes with taste profiles most similar to sucrose, which was further verified by sensory evaluation. This study provides a new reliable modeling strategy for the rapid screening of sweeteners with a clean sweet taste.