FAPD: An Astringency Threshold and Astringency Type Prediction Database for Flavonoid Compounds Based on Machine Learning

Astringency is a puckering or velvety sensation mainly
derived
from flavonoid compounds in food. The traditional experimental approach
for astringent compound discovery was labor-intensive and cost-consuming,
while machine learning (ML) can greatly accelerate this procedure.
Herein, we propose the Flavonoid Astringency Prediction Database (FAPD)
based on ML. First, the Molecular Fingerprint Similarities (MFSs)
and thresholds of flavonoid compounds were hierarchically clustering
analyzed. For the astringency threshold prediction, four regressions
models (i.e., Gaussian Process Regression (GPR), Support Vector Regression
(SVR), Random Forest (RF), and Gradient Boosted Decision Tree (GBDT))
were established, and the best model was RF which was interpreted
by the SHapley Additive exPlanations (SHAP) approach. For the astringency
type prediction, six classification models (i.e., RF, GBDT, Gaussian
Naive Bayes (GNB), Support Vector Machine (SVM), k-Nearest Neighbor
(kNN), and Stochastic Gradient Descent (SGD)) were established, and
the best model was SGD. Furthermore, over 1200 natural flavonoid compounds
were discovered and built into the customized FAPD. In FAPD, the astringency
thresholds were achieved by RF; the astringency types were distinguished
by SGD, and the real and predicted astringency types were verified
by t-Distributed Stochastic Neighbor Embedding (t-SNE). Therefore,
ML models can be used to predict the astringency threshold and astringency
type of flavonoid compounds, which provides a new paradigm to research
the molecular structure–flavor property relationship of food
components.