Classification of tastants: A deep learning based approach

Area of Science:

• Computational chemistry
• Molecular modeling
• Cheminformatics

Background:

• Predicting molecular taste is vital for food, flavor, and pharmaceutical industries.
• G protein-coupled receptors mediate the sensation of basic tastes: sweet, bitter, and umami.
• Developing in-silico methods accelerates the design and screening of novel tastants.

Purpose of the Study:

• To develop and evaluate deep learning models for classifying sweet, bitter, and umami molecules.
• To explore the utility of molecular descriptors and graph neural networks for taste prediction.
• To apply explainable AI techniques to understand model predictions and demonstrate practical applications.

Main Methods:

• Curated an extensive dataset of 1466 bitter, 1764 sweet, and 238 umami tastants.
• Trained a deep neural network (DNN) using molecular descriptors and a graph neural network (GNN).
• Addressed class imbalance using specialized sampling techniques and employed Shapley Additive Explanations (SHAP) for model interpretability.

Main Results:

• Both DNN and GNN models achieved comparable performance in taste prediction.
• The GNN model demonstrated the ability to learn representations directly from molecular structures.
• SHAP analysis provided insights into the DNN model's predictions, enhancing understanding.
• Models were successfully applied to screen tastants from a large food database.

Conclusions:

• Developed effective in-silico deep learning models for predicting molecular taste (sweet, bitter, umami).
• Highlighted the advantage of GNNs in learning from molecular structure without handcrafted features.
• The study provides a powerful computational tool for accelerating tastant design and discovery.