Interpretable machine learning for optimization of ultrasound-assisted polysaccharide extraction using natural deep eutectic solvents: From yield enhancement to structural and antioxidant activity validation.

Traditional methods for polysaccharide extraction are inefficient and environmentally harmful. To develop an environmentally friendly extraction process, we combined natural deep eutectic solvents (NADES) with ultrasonic-assisted extraction and introduced machine learning models to optimize the extraction parameters and achieve a prediction accuracy of 95.24%. The optimal extraction parameters were determined as follows: NADES water content 54.65%, NADES to Astragalus membranaceus leaf ratio (NAR) 29.75 mL/g, extraction time 30 min, ultrasound power 240 W, with an Astragalus membranaceus leaf polysaccharide (AMP) yield of 175.40 mg/g. NADES was characterized by FT-IR and NMR, verifying the successful synthesis of NADES. After extraction, the structure of AMP was characterized by UV, FT-IR, molecular weight, particle size, zeta potential and SEM. Then, through the monosaccharide composition analysis, the AMP contains six types of monosaccharides. Through DFT simulation calculations, the binding stability of each monosaccharide component in AMP with H2O / NADES was compared. These results confirmed that NADES has a strong binding affinity and interaction stability with all monosaccharides, theoretically verifying the excellent extraction performance of NADES. Thermal and rheological analysis further supported its potential applications. In antioxidant assays, the scavenging rates of inhibitory DPPH and hydroxyl radicals exceeded 50%, demonstrating that NADES effectively preserves biological activity. This intelligent optimization model offers an efficient, green strategy for natural product extraction, reducing prediction errors to within 5% and establishing a comprehensive framework for green extraction process optimization, with great practical value for the green chemical industry.