Building a Kokumi Database and Machine Learning-Based Prediction: A Systematic Computational Study on Kokumi...

Area of Science:

• Food Science and Technology
• Computational Chemistry
• Sensory Science

Background:

• Kokumi, a taste sensation of fullness and thickness, traditionally requires laborious analysis.
• Emerging computational methods offer efficient strategies for molecular taste prediction.
• Developing predictive models is crucial for accelerating the discovery of kokumi compounds.

Purpose of the Study:

• To comprehensively analyze, predict, and screen kokumi compounds using computational approaches.
• To categorize kokumi compounds based on molecular characteristics and predict their taste profiles.
• To identify novel kokumi active compounds through large-scale virtual screening.

Main Methods:

• Categorization of 285 kokumi compounds into five groups based on molecular features.
• Prediction of kokumi/non-kokumi and multi-flavor compositions using six structure-taste relationship models (MLP-E3FP, MLP-PLIF, MLP-RDKFP, SVM-RDKFP, RF-RDKFP, WeaveGNN).
• High-throughput virtual screening of over 100 million molecules, followed by toxicity and similarity filtering.

Main Results:

• The WeaveGNN model achieved an AUC of 0.94 for kokumi/non-kokumi prediction, outperforming other models.
• The MLP-E3FP model demonstrated high predictive performance (AUC 0.94, MCC 0.74) for multi-flavor prediction.
• Successful identification of kokumi active compounds via extensive virtual screening.

Conclusions:

• Computational models, particularly WeaveGNN and MLP-E3FP, show high proficiency in predicting kokumi molecules.
• The developed platform, KokumiPD, provides a valuable resource for kokumi compound analysis and prediction.
• This study significantly advances the efficiency of kokumi compound discovery and application.