Interpretable Machine Learning Integrates Flavor Classification with Source Attribution for Low-Salt Soy Sauce Optimization

Low-salt soy sauce (LSS) supports healthier diets but suffers flavor defects compared with high-salt and koikuchi soy sauce (HSS, KSS). In this study, an XGBoost-SHAP-MMVec model integrating a sensory-volatiles-microbiota triomics data set was developed to guide LSS flavor enhancement. Quantitative descriptive analysis and olfactometry identified 6-sensory terms and 29-odor active volatiles. XGBoost-SHAP (accuracy >0.90) revealed contributors: sauce (furanones), rancid (linear fatty aldehydes), fruit (short-chain ethyl and acetate esters), roast (3-methylpyrazine derivatives), flower (norisoprenoids), and grain (Ehrlich-derived aldehydes), confirmed by recombination/omission tests. Microbiota profiling identified 12-microbiota (with abundance and co-occurrence connectivity advantages), which were used for MMVec training to reveal volatiles-microbiota associations. A five-genus consortium (Lactiplantibacillus, Staphylococcus, Bacillus, Zygosaccharomyces, Aspergillus) capturing three soy sauce flavor profiles was confirmed by fermentation trials. Model-guided optimization yielded the enhanced LSS retained grain and flower aromas while incorporating HSS-like sauce and KSS-like roast and fruity signatures. This research demonstrates a strategy for targeted flavor enhancement in fermented foods.