Umami-Transformer: A deep learning framework for high-precision prediction and experimental validation of umami peptides.

In food field, both identification of umami peptides and their sensory evaluation are limited by low efficiency of traditional methods and subjectivity of human-based assessments. To overcome these issues, Umami-Transformer was developed by integrating Transformer architecture with eight physicochemical descriptors. A classification accuracy of 0.965, an F1 score of 0.903 and a Matthews correlation coefficient of 0.889 were obtained. All dipeptides to pentapeptides were examined, four peptides with top prediction scores and strong docking affinities (DD, DDE, DDED, and DDEDD) were synthesized. Sensory and electronic tongue analyses confirmed umami and saltiness of DDE (1 mg/mL) and DDED (1 mg/mL), which surpassed 3 mg/mL monosodium glutamate. Molecular docking studies revealed the presence of Asp/Glu residues at either the N-terminus or C-terminus of umami peptides enhances their interaction with the umami receptor, thereby eliciting umami taste perception. Theoretical modeling is bridged with practical applications of taste optimization, resulting in significant cost savings.