BERT4Bitter

BERT4Bitter predicts bitter peptides from amino acid sequences using a BERT-based model (bidirectional encoder representations from transformers) to enable computational screening for drug development and nutritional research.

Key Features:

• BERT-Based Model: Applies a BERT-based model (bidirectional encoder representations from transformers) to predict bitterness directly from amino acid sequences without requiring structural information.
• Predictive Performance: Achieves accuracy of 0.861 in cross-validation and 0.922 in independent tests, demonstrating superior performance to conventional machine learning models.
• Empirical Benchmarking: Outperforms existing methods on independent datasets with an 8.0% improvement in accuracy and a 16.0% improvement in Matthews correlation coefficient.

Scientific Applications:

• Drug development: Facilitates rapid screening and identification of novel bitter peptides to inform lead selection and formulation in drug discovery.
• Nutritional research: Enables large-scale prediction of peptide bitterness to support studies of taste, food formulation, and nutritional component screening.

Methodology:

The method trains a BERT-based model on labeled peptide sequences to learn sequence patterns associated with bitterness and predicts bitterness for new peptides solely from their amino acid composition without using structural data.