Prediction of bitterness based on modular designed graph neural network.

Area of Science:

• Food science and sensory analysis
• Computational chemistry and cheminformatics
• Machine learning and artificial intelligence

Background:

• Bitterness is a key taste for identifying harmful substances.
• Human tasting for flavor analysis is costly and impractical for large-scale screening.
• In silico methods offer a practical alternative for bitterness prediction.

Purpose of the Study:

• To develop and evaluate advanced machine learning models for accurate in silico bitterness prediction.
• To introduce Graph Neural Networks (GNNs) as a superior alternative to traditional methods for bitterness prediction.
• To create accessible tools and resources for bitterness prediction in food science.

Main Methods:

• Development of a Hybrid Graph Neural Network (HGNN) model for bitterness prediction.
• Implementation of BitterGNNs, a GNN-based predictor incorporating HGNN and other GNN architectures.
• Validation of BitterGNNs on public datasets against established predictors like RDKFP-MLP.

Main Results:

• The BitterGNNs predictor achieved an Area Under the Curve (AUC) of 0.87 in both bitter/non-bitter and bitter/sweet evaluations.
• BitterGNNs outperformed the RDKFP-MLP predictor, which achieved AUC values of 0.86 and 0.85 respectively.
• A publicly accessible bitterness prediction website and database, TastePD, was created.

Conclusions:

• Graph Neural Networks, particularly the developed BitterGNNs model, offer highly accurate and efficient bitterness prediction.
• The BitterGNNs predictor enhances food testing methodologies and contributes to understanding the origins of bitterness.
• The TastePD database and associated code provide valuable resources for researchers and the food industry.