In brief
A team at Hebrew University used machine learning and molecular docking to screen more than 50 million molecules, narrowing the pool to 92 potential sugar alternatives. The models predicted taste profiles and receptor binding, aiming to reduce bitter aftertastes. Further laboratory validation is required before commercial use can be confirmed.
JERUSALEM – In a push to find healthier alternatives to table sugar, researchers at Hebrew University of Jerusalem announced on Sunday that an artificial‑intelligence system sifted through more than 50 million chemical molecules to identify 92 promising sweet‑tasting candidates. The study, published in the journal Food Chemistry, demonstrates how computational methods can speed up the discovery of new sugar substitutes while avoiding unwanted bitter or metallic aftertastes.
The team first assembled predictive models trained on known sweet and bitter compounds. Using these models, they evaluated three large chemical databases, in which the researchers logged a total of 50,700,000 molecules. The machine‑learning framework flagged those predicted to taste sweet and excluded the ones likely to be bitter, reducing the potential list to 556 candidates.
Next, the researchers applied molecular docking simulations. The approach modeled how each compound might fit into the human sweet‑taste receptor – a protein complex whose detailed structure was published the previous year. The docking calculations helped to prioritise compounds that showed favorable binding interactions while maintaining a low predicted bitterness signature.
After a series of refinements that combined taste‑prediction scoring with receptor‑binding scores, the research group reportedly narrowed the list to 92 molecules. According to the authors, many of those candidates possess chemical backbones distinct from existing commercial sweeteners, potentially opening the door to novel classes of sugar alternatives.
Methodology and Predicted Properties
Key to the approach was a dual‑model system that segregated desirable sweetness from undesirable bitterness. By feeding the framework thousands of molecules with known palatal qualities, the scientists calibrated the model to recognise subtle structural motifs that contribute to taste. The docking step used the 3‑D structure of the human taste receptor, which had been solved by crystallography last year, allowing the AI to simulate realistic ligand‑receptor interactions.
The combination of machine‑learning predictions and docking was designed to flag candidates that not only taste sweet but also bind effectively to the receptor. This step may reduce the number of experimental trials needed, since compounds that lack sufficient receptor affinity are unlikely to elicit a sweet sensation in humans.
Future Work and Uncertainties
While the computational pipeline identified 92 noteworthy compounds, the authors emphasised that these are preliminary findings. The molecules still require laboratory synthesis, sensory testing, and safety evaluations before they can be considered viable food additives. In particular, there is uncertainty about whether the predicted low bitterness will hold true in human taste tests, and if the compounds will meet regulatory standards for consumption.
The study’s authors noted that the novelty of the chemical structures might pose additional challenges in manufacturing and formulation. Therefore, bridging the gap from in silico promise to practical sweeteners will demand multidisciplinary collaboration among chemists, food technologists, and regulatory experts.
Thus, while the AI‑driven discovery marks a compelling advance in sweetener research, it leaves notable uncertainties about the real‑world performance and commercial viability of the highlighted candidates.
