A chemical language model for molecular taste prediction

Area of Science:

• Food Science
• Computational Chemistry
• Cheminformatics

Background:

• Predicting molecular taste from chemical structure is a significant challenge in food science.
• Existing models often focus on binary taste classification, limiting their scope.

Purpose of the Study:

• To develop a chemical language model for predicting molecular taste from chemical structure.
• To create the first model capable of parallel predictions across multiple taste categories (sweet, bitter, sour, umami).

Main Methods:

• Developed FART (Flavor Analysis and Recognition Transformer), a transformer-based chemical language model.
• Trained FART on the largest public dataset of molecular tastants to date (15,025 compounds).
• Utilized gradient-based visualization for model interpretability and identification of key structural features.

Main Results:

• FART achieved over 91% accuracy in parallel taste prediction across four categories.
• The model outperformed previous state-of-the-art binary classifier models.
• Identified key structural elements responsible for specific taste properties.

Conclusions:

• FART provides a powerful tool for rapid taste prediction from molecular structure.
• The model accelerates flavor compound development and enables systematic exploration of taste chemistry.
• Released model and dataset empower the food science community.