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

Kokumi is a subtle sensation characterized
by a sense of fullness,
continuity, and thickness. Traditional methods of taste discovery
and analysis, including those of kokumi, have been labor-intensive
and costly, thus necessitating the emergence of computational methods
as critical strategies in molecular taste analysis and prediction.
In this study, we undertook a comprehensive analysis, prediction,
and screening of the kokumi compounds. We categorized 285 kokumi compounds
from a previously unreleased kokumi database into five groups based
on their molecular characteristics. Moreover, we predicted kokumi/non-kokumi
and multi-flavor compositions using six structure–taste relationship
models: MLP-E3FP, MLP-PLIF, MLP-RDKFP, SVM-RDKFP, RF-RDKFP, and WeaveGNN
feature of Atoms and Bonds. These six predictors exhibited diverse
performance levels across two different models. For kokumi/non-kokumi
prediction, the WeaveGNN model showed an exceptional predictive AUC
value (0.94), outperforming the other models (0.87, 0.90, 0.89, 0.92,
and 0.78). For multi-flavor prediction, the MLP-E3FP model demonstrated
a higher predictive AUC and MCC value (0.94 and 0.74) than the others
(0.73 and 0.33; 0.92 and 0.70; 0.95 and 0.73; 0.94 and 0.64; and 0.88
and 0.69). This data highlights the model’s proficiency in
accurately predicting kokumi molecules. As a result, we sourced kokumi
active compounds through a high-throughput screening of over 100 million
molecules, further refined by toxicity and similarity screening. Lastly,
we launched a web platform, KokumiPD (https://www.kokumipd.com/),
offering a comprehensive kokumi database and online prediction services
for users.