Identification of umami peptide using experimental designs is time-consuming and expensive thus limiting large-scale discovery in food and nutritional studies. To overcome this issue, we introduce a deep learning model of precise umami peptides prediction based on a hybrid feature representation, which incorporates sequence-derived, physicochemical, and contextual embedding features. The given approach was tested on a benchmark dataset and the BIOPEP-UWM one. Transformer-based model reported a 98 and 95 percent accuracy, whereas the RNN model reported an accuracy of 96 and 93 percent on the two datasets. The combination of hybrid features invariably enhanced the performance of models and this proves the power of wholesome sequence representation in classifying umami peptides.This framework provides an efficient computational tool to accelerate umami peptide discovery. • Introduces a deep learning-based framework for umami peptide prediction using hybrid structural, physicochemical, and sequence-based features. • Employs CTDC, KNNProtein, and transformer embeddings to enhance peptide representation and classification accuracy. • Utilizes SMOTE to address dataset imbalance for more robust model generalization. • Achieves high predictive performance across benchmark and BIOPEP-UWM datasets using RNN and Transformer models. • Provides a scalable computational tool for accelerating umami peptide discovery in food and nutrition research.
Jiang et al. (Sat,) studied this question.