Enhancing nutritional yeast-derived meaty flavours with multi-task Bayesian optimisation.

Abstract

This study applied data-driven machine learning to optimise the generation of meaty flavour from Maillard-reaction-derived nutritional yeast hydrolysates. Conventional flavour development relies heavily on empirical expertise, limiting efficiency and generalisability. In this study, a Multi-Task Bayesian Optimisation (MTBO) framework integrating gas chromatography-mass spectrometry and descriptive sensory data was developed to accelerate flavour optimisation. A Multi-Task Gaussian Process model was used to fit relationships between input parameters (enzyme-to-substrate and protein-to-sugar ratios) and experimental outputs (volatile compound concentrations and sensory attribute scores). Bayesian Optimisation iteratively recommended new parameters for experimental validation. Model-guided optimisation resulted in a 3.0-fold increase in total pyrazine concentration and a 16.58% reduction in 2-furanmethanol in glucose-derived MRPs, alongside a 2.7-fold increase in total pyrazine concentration for xylose-derived MRPs. Sensory validation demonstrated a 55%-79% reduction in perceived bitterness and up to a 28% increase in perceived meaty flavour intensity, indicating MTBO's potential to accelerate flavour optimisation.