Umami-BERT: An interpretable BERT-based model for umami peptides prediction.

Area of Science:

• Biotechnology
• Bioinformatics
• Food Science

Background:

• Umami peptides enhance flavor and nutrition, driving demand for novel sources.
• Current screening methods for umami peptides are inefficient, necessitating advanced approaches.
• Identifying umami peptides is crucial for food innovation and nutritional enhancement.

Purpose of the Study:

• To develop a deep learning (DL) model for rapid and efficient screening of umami peptides.
• To create a model that predicts umami peptide potential directly from amino acid sequences.
• To gain insights into the molecular characteristics of umami peptides.

Main Methods:

• Developed the Umami-BERT model using a two-stage training strategy with Bidirectional Encoder Representations from Transformers (BERT) and inception network.
• Employed attention mechanisms during pre-training on extensive bioactive peptide sequences.
• Utilized the UMP789 dataset for re-training and umami peptide prediction.

Main Results:

• The Umami-BERT model achieved high accuracy (93.23% on balanced, 95.00% on unbalanced datasets) and MCC (0.78, 0.85).
• Demonstrated superior performance compared to existing models in predicting umami peptides.
• Identified Alanine (A), Cysteine (C), Aspartate (D), and Glutamic acid (E) as key amino acid contributors.

Conclusions:

• Umami-BERT offers a powerful tool for large-scale screening of candidate umami peptides.
• The model provides novel insights into the sequence patterns underlying umami taste.
• Facilitates further exploration and discovery of functional peptides in food science.