BERT4Bitter: a bidirectional encoder representations from transformers (BERT)-based model for improving the...

Area of Science:

• Bioinformatics
• Computational Biology
• Peptide Science

Background:

• Experimental identification of bitter peptides is costly and slow.
• The post-genomic era necessitates automated methods for discovering new bitter peptides.

Purpose of the Study:

• To develop and present BERT4Bitter, a Bidirectional Encoder Representation from Transformers (BERT)-based model for predicting bitter peptides.
• To provide an automated and efficient tool for identifying novel bitter peptides.

Main Methods:

• Utilized a BERT-based deep learning architecture.
• Trained and validated the model on peptide amino acid sequences.
• Compared performance against traditional machine learning models.

Main Results:

• BERT4Bitter achieved high accuracy: 0.861 (cross-validation) and 0.922 (independent tests).
• Outperformed existing methods by 8.0% in accuracy and 16.0% in Matthews coefficient correlation.
• Demonstrated effectiveness and robustness on an independent dataset.

Conclusions:

• BERT4Bitter is the first BERT-based model for bitter peptide identification.
• The model offers a rapid and effective tool for screening bitter peptides.
• Potential applications include drug development and nutritional research.