Sweetness Prediction Based on Chemo-Physical Parameters | Apr-2026
- Yuval Klein
- Apr 16
- 1 min read
Abstract
Predicting the relative sweetness of compounds remains challenging due to the subjectivity and variability inherent in human sensory panels. This study presents a machine-learning-based algorithm that forecasts relative sweetness values (RSV, sucrose = 1.00) using only readily calculable two-dimensional physicochemical descriptors, including molecular size and geometry, hydration-related indices, and related parameters. The model employs an ensemble of gradient-boosting decision trees, with features standardized and sweetness targets log-transformed to handle the five-order-of-magnitude range. Sucrose is anchored exactly at 1.0000 through weighted training and a post-prediction calibration offset, ensuring consistent reference scaling.
The algorithm was evaluated on a diverse set of 54 sweeteners spanning sugars, polyols, terpenic glycosides, peptides, and synthetic high-intensity compounds. Predictions achieve a training R² of 0.9969 (log scale), a median absolute percentage error of 6.3%, and a mean error of 12.8%, with all compounds falling within a factor of 2 of literature consensus values or ranges. The dominant driver is a hydration-related index, explaining much of the observed potency ordering. Limitations include inability to distinguish stereoisomers with identical 2D profiles. Despite this, the model offers reliable ranking and magnitude estimation across five orders of magnitude.
By providing deterministic, reproducible RSV estimates free from human sensory variability, this sucrose-anchored approach enables faster virtual screening of novel sweeteners, supports formulation in low-calorie foods and pharmaceuticals, and paves the way for a more standardized, objective sweetness scale. Future extensions incorporating 3D conformational and charge-state-aware descriptors are expected to further enhance accuracy and domain coverage.
Comments