Database, prediction, and antibacterial research of astringency based on large language models

Area of Science:

• Food Science
• Computational Chemistry
• Sensory Science

Background:

• Astringency, a key taste sensation, affects food quality (e.g., wine, tea) but traditional prediction methods are expensive.
• The link between astringency and antibacterial properties is largely uninvestigated, despite potential applications.

Purpose of the Study:

• To develop a computational framework for predicting molecular astringency and antibacterial activity.
• To establish the first comprehensive database of astringent molecules.
• To explore the relationship between astringency and antibacterial properties.

Main Methods:

• Created a database of 238 molecules for astringency prediction.
• Developed a Ligand-Based Prediction (LBP) framework using machine learning, deep learning, and large language models.
• Validated predictions using electronic tongue measurements and developed specific predictors for molecules and peptides.

Main Results:

• Achieved 0.95 accuracy and 0.90 AUC for astringency prediction.
• Developed antibacterial predictors with 0.92 accuracy for molecules and 0.88 for peptides.
• Discovered that 51% of astringent molecules exhibit antibacterial properties.

Conclusions:

• The developed computational approach accurately predicts astringency and antibacterial activity.
• A significant overlap exists between astringent molecules and those with antibacterial properties.
• The AstringentPD and ABPD web servers provide accessible tools for further research in food science and medicinal chemistry.