Based on the umami-transformer machine learning model: Screening of umami peptides from ultrasound-assisted enzymatic hydrolysis of shiitake mushrooms and study on their flavor mechanism.

This study aimed to optimize ultrasound-assisted enzymatic hydrolysis for umami peptide preparation from shiitake mushrooms, develop a reliable predictive screening model for umami peptides, and clarify the molecular mechanism underlying the umami taste of these peptides. Results indicated that ultrasonic pretreatment (200 W, 20 min) combined with dual-enzyme hydrolysis (complex protease and flavourzyme, at 1:1 ratio, 2400 U/g, 4 h) markedly enhanced hydrolysis efficiency. The α-amino nitrogen content reached 0.20 ± 0.00 mg/mL, with 98.31 % of the resulting peptides having molecular weights below 1 kDa. A triple-stream Umami-Transformer model was constructed by integrating peptide sequence, acid-base properties, and hydrophobicity. This model achieved an umami peptide prediction accuracy of 94 %, outperforming well-established models including Umami-BERT (90.5 %) and Umami-MRNN (93.2 %), thereby confirming its superior predictive performance. Four novel umami peptides (LPFQTE, MIPLLL, EVLPNL, and EVENVI) were successfully identified and validated. Molecular docking and molecular dynamics simulations revealed key binding sites (Asp147, Lys155, Ser217, Lys379, Asn150) on the T1R1 receptor and demonstrated that hydrogen bonding and ionic interactions are critical for stable binding and umami taste initiation. The findings provide a robust framework for the efficient production and identification of umami peptides from shiitake mushrooms, together with mechanistic insights that contribute to the development of novel flavor enhancers and functional foods.