Machine learning-assisted green extraction of polysaccharides from Syringa oblata Lindl leaf residue using natural deep eutectic solvents: process optimization, structural characterization, and bioactivities

The high-value utilization of underutilized plant resources and waste materials is of considerable significance for the sustainable resource development, agricultural by-product valorization, and the circular economy. In this study, machine learning was integrated with natural deep eutectic solvent (NADES), using Syringa oblata Lindl ( S. oblata ) leaf residue polysaccharides (SOLP) as a case study, to establish a green and efficient ultrasound-assisted NADES extraction strategy. Based on multiple machine learning models, the polysaccharide extraction yield and key process parameters were accurately predicted and optimized. The optimal extraction conditions predicted by the model were as follows: NADES water content of 51.4%, NADES to S. oblata leaf residue ratio (DSR) of 42.9 mL/g, ultrasonic time of 41.7 min, and ultrasonic power of 250.0 W. Among the machine learning models, the XGB model exhibited the highest prediction accuracy (test set, R 2 > 0.92). SHAP analysis further revealed that DSR, water content, ultrasonic time, and ultrasonic power contributed 58.57%, 21.31%, 12.82%, and 7.29% to the polysaccharide yield, respectively. Density functional theory (DFT) calculation was performed to further explore the NADES extraction mechanism. The results demonstrated that the binding energy between NADES-6 and SOLP was significantly better than that in the traditional solvent. The formation of hydrogen-bond interactions promoted dissolution and release of polysaccharides, thereby enhancing extraction efficiency. Structural characterization showed that SOLP was mainly composed of galacturonic acid and galactose, with minor amount of rhamnose, glucuronic acid, glucose, and arabinose. In addition, SOLP exhibited ABTS and hydroxyl radical scavenging activities. Overall, this study establishes a data-driven research framework integrating machine learning, DFT analysis, and NADES-based extraction. In addition to improving polysaccharide extraction efficiency and prediction accuracy, this study also provides a promising strategy for the high-value utilization of natural polysaccharides derived from underutilized plant resources and agricultural residues.