A chef tweaking a recipe relies on experience built from thousands of dishes, adjusting each version until taste and aroma blend into the perfect flavour. “It is effective, but what works for one ingredient doesn’t always carry over to another, and there are often far too many possible combinations to test by hand,” said Jie Hong Chiang, a Senior Scientist at the A*STAR Singapore Institute of Food and Biotechnology Innovation (A*STAR SIFBI).
Seeing how machine learning has sped up discovery and optimisation processes like drug design, Chiang and team set out to apply the same strategy to flavour development in alternative proteins. Machine learning models can learn from previous results to guide the next recipe adjustment, making recreating rich, meaty flavours less of a guessing game.
A*STAR SIFBI researchers led the work on aroma chemistry and sensory testing, focusing on nutritional yeast as the protein source. Meanwhile, colleagues at the A*STAR Institute of Advanced Intelligence and Computing (A*STAR IAIC) built a Multi-Task Bayesian Optimisation (MTBO) framework to model how enzyme-to-substrate and protein-to-sugar ratios shape flavour.
They took gas chromatography-mass spectrometry measurements of nine volatile compounds, along with sensory scores for meaty odour, meaty flavour, savouriness and bitterness. “Every experimental set helps build a picture of how the enzyme and sugar concentrations shape the flavour, with each new clue narrowing down the conditions to test next,” said Michelle Choy, a Research Officer at A*STAR SIFBI.
While chemical aroma data was measured per batch, taste testing was more time- and resource-intensive and could only be performed occasionally, leading to comparatively less sensory data to work with. To address this, the A*STAR IAIC researchers treated each aroma compound and each taste attribute as separate learning tasks within a shared model, emphasising the multi-task nature of their framework.
“This way, the model could still learn something useful about the taste profiles from the smaller pool of sensory data while drawing on the richer chemical data at the same time, instead of being thrown off by the imbalance,” explained Chiang.
After six rounds, the new formulations produced up to a threefold increase in the total pyrazine concentration, which are associated with roasted, nutty meat aromas, across various samples. Sensory panels also rated the optimised products 55 to 79 percent less bitter. “What surprised us was how much of that came down to a single compound,” Choy said, adding that bitterness has remained a longstanding obstacle in enhancing the appeal of meat alternatives.
The bitter-taste culprit, 2-furanmethanol, was cut by around 16 percent in glucose-based flavourings based on MTBO model suggestions. Removing that undesirable note turned out to be a more impactful change than boosting meat-like aroma compounds, the researchers found.
Beyond nutritional yeast, the team sees potential to adapt the strategy to other protein sources. Building effective models would require an initial dataset, alongside identification and preliminary optimisation studies of the most relevant conditions and ingredient compounds that contribute to the flavour.
The A*STAR-affiliated researchers contributing to this research are from the A*STAR Singapore Institute of Food and Biotechnology Innovation (A*STAR SIFBI) and A*STAR Institute of Advanced Intelligence and Computing (A*STAR IAIC).
