Prediction of Umami Peptides Based on a Large Language Model of Proteins

Umami peptides possess unique characteristics, making their study highly significant. To better understand umami peptides, this research systematically investigates them using protein language models. First, we collected IC 50 and K d data to construct a protein–peptide affinity model and combined it with protein–peptide docking techniques to explore the affinity relationships between umami peptides, non-umami peptides, and taste receptors. The results indicate that umami peptides exhibit stronger affinity to umami receptors compared to non-umami peptides but show no significant difference in affinity to bitter receptors. Subsequently, we systematically gathered 972 umami peptides and 608 non-umami peptides, developing the largest data set of umami peptides to date. Using protein language models combined with molecular docking and affinity prediction results, we constructed the most accurate umami peptide prediction model, achieving an accuracy of 82% and an area under the curve (AUC) of 0.87. Finally, we developed a user-friendly website for umami peptide analysis, UmamiMeta, accessible at https://hwwlab.com/Webserver/umamimeta, providing a convenient tool for the research and application of umami peptides.