Combining Online HPLC-ABTS Screening With Machine Learning to Predict the Bioactivity of Antioxidant Components in Pandani Tectorii Radix et Rhizoma

This study developed an online HPLC-ABTS method to screen antioxidant phenolic acids in Pandani tectorii radix et rhizoma (PT) from 20 geographical origins. Six common constituents were screened and assigned by standards: neochlorogenic acid, chlorogenic acid, cryptochlorogenic acid, isochlorogenic acid B, isochlorogenic acid A, and isochlorogenic acid C. Using these six components as variables, four regression models were constructed to predict ABTS radical scavenging rate: support vector regression (SVR), decision tree regression, K-nearest neighbors (KNN) regression, and extreme gradient boosting (XGBoost) regression. From 80 samples, 64 were used for model development, and 16 were reserved for hold-out validation. Model selection employed 5-fold, nested, and leave-one-out cross-validations, complemented by permutation tests and learning curve analysis. The SVR model performed best, with nested CV R 2 = 0.9298, RMSE = 4.1207, MAE = 2.2313, and an optimism gap of 0.0137. After final tuning (C = 100, gamma = 0.1, and epsilon = 0.5), it achieved R 2 = 0.9000, RMSE = 4.0602, and MAE = 2.2986 on the hold-out set. This efficient and accurate method enables simultaneous screening of antioxidant constituents and prediction of radical scavenging activity, providing a new technical tool and scientific basis for PT quality evaluation.