iBitter-GRE: A Novel Stacked Bitter Peptide Predictor with ESM-2 and Multi-View Features

Area of Science:

• Biochemistry
• Computational Biology
• Bioinformatics

Background:

• Accurate identification of bitter peptides is crucial for scientific research.
• Existing sequence-based models for bitter peptide prediction have limitations in performance.
• There is a need for improved computational tools to identify bitter peptides.

Purpose of the Study:

• To develop a novel and accurate predictive tool, iBitter-GRE, for identifying bitter peptides.
• To enhance the predictive performance beyond current sequence-based models.
• To provide a user-friendly web server for researchers.

Main Methods:

• Utilized ESM-2 and traditional descriptors for feature extraction, capturing physical and biochemical properties.
• Employed a stacking approach to integrate multiple machine learning learners.
• Analyzed feature contributions using SHAP values for interpretability.

Main Results:

• iBitter-GRE demonstrated higher accuracy compared to existing bitter peptide identification methods.
• Validation by domain experts confirmed the model's ability to identify key biochemical characteristics.
• The developed web server (http://www.bioai-lab.com/iBitter-GRE) is now available for public use.

Conclusions:

• iBitter-GRE represents a significant advancement in bitter peptide identification accuracy.
• The tool effectively leverages both deep learning and traditional descriptors.
• iBitter-GRE is a valuable resource for accelerating bitter peptide discovery and research.