DmcFlavor: Compound-informed prediction and interpretation of integrated multi-flavor profiles in foods and herbs

Food flavor is inherently multidimensional and arises from the complex combined effects of diverse compounds; however, dedicated computational tools for predicting integrated multi-flavor profiles of foods are still lacking. In this study, DmcFlavor was developed as the first framework for predicting and interpreting integrated multi-flavor profiles of foods. DmcFlavor employs Transformer-based modeling to perform both compound-level and mixture-level flavor prediction across five flavor attributes: bitter, pungent, salty, sour, and sweet. DmcFlavor was applied to tomato ripening analysis, revealing dynamic changes in integrated flavor profiles during ripening and quantitatively identifying major contributors to sourness and sweetness. It was further applied to characterize integrated flavor patterns across 18 food categories, including fishes, fruits, herbs and spices, and vegetables, and to explore potential flavor-health effect associations. This framework may facilitate integrated food multi-flavor evaluation, support flavor optimization in the food industry, and promote the systematic exploration of food therapy.