Abstract
Consuming foods with high antioxidant capacity is considered beneficial to health, and predicting the antioxidant capacity of food components is important. In the 2,2-diphenyl-1-picrylhydrazyl (DPPH) assay, multiple reactions occur simultaneously, and because the experimental conditions are not standardized across studies, quantitative prediction of DPPH activity is difficult. In this study, we qualitatively and quantitatively predicted the DPPH activity of phenols in food using data obtained under unified experimental conditions and machine learning. We measured DPPH activity of 96 compounds to create a dataset comprising measurements of 274 compounds, including values previously reported by our laboratory. The classification model implemented using LightGBM showed high performance, achieving an accuracy of 0.88 and an F1 score of 0.86. The support vector regression model satisfied the Golbraikh-Tropsha criteria, with an R(2)(test) of 0.70, RMSE(test) of 0.44, q(2) of 0.61, and RMSE(validation) of 0.46. Furthermore, the chemical validity of the prediction was confirmed by comparing the results of the machine learning model with those of previous studies. This method provides a basis for the quantitative prediction of DPPH activity of numerous phenolic compounds in foods and is expected to contribute to the elucidation of the antioxidant capacity of foods.