xBitterT5: an explainable transformer-based framework with multimodal inputs for identifying bitter-taste peptides.

Area of Science:

• Food Science
• Biomedicine
• Computational Biology
• Bioinformatics

Background:

• Bitter peptides (BPs) influence taste and physiological processes, crucial for food quality and health.
• Traditional machine learning methods for BP identification face limitations with complex biological sequence data.
• Recent advances in protein language models offer improved accuracy but often lack interpretability and molecular representation.

Purpose of the Study:

• To develop a novel, interpretable framework for accurate bitter peptide identification.
• To integrate peptide sequence and molecular representations for enhanced predictive power.
• To provide mechanistic insights into the chemical basis of peptide bitterness.

Main Methods:

• Proposed xBitterT5, a multimodal framework combining BioT5 embeddings with peptide sequences and SELFIES molecular representations.
• Utilized pretrained transformer-based embeddings for advanced feature extraction.
• Incorporated both sequence and molecular string data for a comprehensive analysis.

Main Results:

• xBitterT5 achieved superior performance compared to existing methods on benchmark datasets.
• The model demonstrated residue-level interpretability, identifying key chemical substructures contributing to bitterness.
• Provided mechanistic insights beyond traditional black-box prediction models.

Conclusions:

• xBitterT5 offers a significant advancement in bitter peptide identification.
• The framework provides valuable mechanistic insights for peptide-based food and biomedical applications.
• Freely available web server and standalone version facilitate broader research accessibility.