What if artificial intelligence could help discover the next generation of sweeteners, without the bitter aftertaste? A new study combined machine learning and molecular modeling to screen more than 50 million chemical compounds, narrowing the search to 92 promising candidates for future sugar alternatives. The findings offer a new approach to discovering sweeteners with improved taste profiles, potentially accelerating innovation in the food and beverage industry.

Researchers Dr. Shayma El-Atawneh and Prof. Amiram Goldblum at the Hebrew University of Jerusalem have developed a new computational framework that identified 92 promising candidates for future sugar substitutes from more than 50 million chemical compounds. The study, published in the scientific journal Food Chemistry, demonstrates how artificial intelligence and molecular modeling could accelerate the discovery of new sweeteners while helping overcome one of the industry's persistent challenges: unwanted bitter aftertastes.

As demand for low-calorie sugar alternatives grows, scientists are seeking sweeteners that provide pleasant taste along with favorable safety and stability profiles. Excessive sugar consumption has been linked to several health and environmental concerns. High sugar intake can increase the risk of obesity, type 2 diabetes, fatty liver disease, and cardiovascular disorders. The FDA has approved artificial sweeteners and established recommended daily amounts per kilogram for each.

However, widely used artificial sweeteners such as aspartame, saccharin, and sucralose have limitations, including bitter or metallic aftertastes.
Discovering new compounds that are sweet without undesirable characteristics remains a significant challenge.

The researchers addressed this problem using a machine-learning algorithm called Iterative Stochastic Elimination (ISE), developed by Goldblum's research group. The models provided by ISE enable the screening and scoring millions of compounds in silico in a short time, thereby greatly reducing the number of compounds for testing in vitro and in vivo. The ISE method builds the models on the basis of molecular properties to predict whether a compound is likely to taste sweet or bitter.

The researchers developed predictive models based on known sweet- and bitter-tasting compounds, enabling them to distinguish between desirable and undesirable taste characteristics. They then applied these models to screen three large chemical databases containing approximately 50.7 million molecules, identifying promising sweet candidates while filtering out compounds predicted to be bitter.

To refine their findings, the team combined machine learning with molecular docking, a computational technique that simulates how molecules interact with biological targets. In this case, the researchers examined how candidate compounds might bind to the single human sweet-taste receptor, which is composed of two protein subunits, T1R2 and T1R3. The detailed structure of this receptor was published only last year, enabling refinement by docking.

By combining taste prediction, molecular docking and chemical structure filters, the researchers narrowed the initial pool to 556 potential sweeteners. Further prioritization yielded 92 candidates, selected for their favorable predicted properties and suitability for future experimental testing.

Importantly, the analysis revealed that many of these molecules have chemical structures distinct from those of known sweeteners, suggesting opportunities to discover entirely new classes of sweet-tasting compounds. The researchers also identified several previously patented sweet-receptor modulators among the candidates, providing additional support for the effectiveness of their computational approach.

A step toward the next generation of sweeteners
Beyond identifying promising molecules, the study highlights the potential of machine learning to accelerate and target the discovery of sugar alternatives. By narrowing millions of possible compounds to a manageable selection, the approach could significantly reduce the experimental effort required to identify new sugar alternatives.

The findings may have implications for the food and beverage industry, which continues to seek low-calorie sweeteners that can mimic the taste of sugar while minimizing undesirable aftertastes.

However, the researchers emphasize that the identified compounds are still computational predictions. Laboratory experiments will be needed to confirm their sweetness, evaluate their safety and determine whether they are suitable for use in food products.

The research demonstrates that combining artificial intelligence with molecular modeling can open new avenues in the science of taste. Although the 92 candidates still require experimental validation, the study establishes a promising framework for discovering the next generation of sweeteners.