Synergistic ultrasound- and vortex-assisted extraction of lobetyolin from Codonopsis tangshen using a supramolecular ternary deep eutectic solvent: yield optimization and antioxidant activity evaluation

Abstract

Herein, a supramolecular ternary deep eutectic solvent coupled with in-situ ultrasound-assisted extraction (STDES-ISUSE) was developed to extract lobetyolin (LBT) in Codonopsis tangshen. The hydroxypropyl-β-cyclodextrin: lactic acid: glycerol (HP-β-CD: LA: Gly) system stood out to extract LBT among the twenty formulations tested. The optimal extraction conditions—water content of 46% (v/v), liquid/solid ratio of 29 mL/g, ultrasound time of 34 min, ultrasound power of 240 W, and vortex time of 10 min—were established via a hybrid response surface methodology-artificial neural network-genetic algorithm (RSM-ANN-GA) paradigm. Under these conditions, the LBT yield reached 0.613 ± 0.011 mg/g, significantly exceeding that of conventional methods. Twenty compounds were identified via UHPLC-Q-TOF-MS, encompassing two phenolic acids, two polyacetylenic glycosides, five fatty acids, two glycosides, one iridoids, and one alkaloids, while a tailored HPLC method enabled precise LBT quantification. Mechanistic investigations, combining FT-IR corroboration of STDES fabrication, SEM visualization, and MD simulations, collectively revealed that the STDES-ISUSE approach potentiates extraction through synergistic cell-wall disintegration and enhanced LBT solubilization. In vitro antioxidant testing further substantiated the bioactivity of the extracts. In summary, this work delivers a green, high-efficiency extraction protocol for bioactive targets from C. tangshen, offering a practical route toward sustainable phytochemical recovery and the broader implementation of green extraction practices in natural product research.

1. Introduction

Codonopsis tangshen Oliv. (Campanulaceae), a plant species native to China with both medicinal and edible values, is officially listed in the Chinese Pharmacopoeia as a dried root–based herbal medicine [1], [2]. Traditionally, it has been employed to invigorate the spleen, tonify the lung, nourish blood, and promote fluid production. The plant is predominantly cultivated in the mountainous regions of Chongqing, Guizhou, and Sichuan provinces [3]. Phytochemically, the root is rich in LBT and polysaccharides, which are primarily responsible for its multifaceted bioactivities, including antioxidant, anti-inflammatory, and hypolipidemic effects. Beyond its therapeutic applications, C. tangshen has also been integrated into Chinese culinary practices for millennia, serving as a daily nourishment ingredient—a testament to its longstanding dual role as both medicine and food [4], [5].

Methanol, ethanol, and acetonitrile remain the workhorse solvents for isolating active compounds from plant sources. Yet, their inherent drawbacks—ecotoxicity, non-renewability, and associated health hazards—have spurred intensive research into greener alternatives [6], [7]. Deep eutectic solvents (DESs), prepared simply via complexation between a hydrogen-bond acceptor (HBA) and a hydrogen-bond donor (HBD) in stoichiometric ratios, have garnered substantial interest owing to their low cost, ease of synthesis, favorable environmental profile, high biosafety, and recyclability [8], [9]. Compared with traditional solvent systems, DESs not only deliver comparable or enhanced extraction yields but also significantly reduce processing costs and environmental impact. A further evolution is represented by TDESs, wherein a third constituent is introduced to strengthen the hydrogen-bonding scaffold, thereby improving solvent stability and extraction robustness [10], [11]. For instance, Yu et al. [12] employed a TDES (choline chloride/1,4–butanediol/ascorbic acid) coupled with microwave assistance for the simultaneous extraction of iridoids and phenolic acids from Eucommia ulmoides leaves, achieving a 1.4–fold enhancement over conventional extraction under optimized conditions. Zhao et al. [13] extracted total flavonoids from Pueraria lobata using a TDES consisting of choline bitartrate, lactic acid, and glycerol. Under optimization by an artificial neural network, the protocol afforded a total flavonoid yield of 223.79 ± 1.89 mg/g, which substantially surpassed that obtained with conventional organic solvents, thereby underscoring the considerable promise of TDES as a sustainable extraction medium.

Lobetyolin (LBT), a C₁₄-polyacetylene glycoside and a hallmark bioactive component of C. tangshen, is an amphiphilic molecule composed of a hydrophilic β-D-glucopyranosyl moiety linked to a hydrophobic polyacetylene aglycone, and thus displays good solubility in routine extraction solvents, ranging from water to common organic solvents. Under ultrasound cavitation, these solvents penetrate and agitate the plant matrix efficiently, producing mechanical disruption of cell walls that triggers the efflux of intracellular LBT [14], [15]. Here, we exploit this principle through a novel dual-mode in-situ extraction strategy that advances sustainable processing—offering the combined merits of speed, reproducibility, straightforward operation, and superior yield. The extraction efficacy is primarily dictated by cavitation dynamics: ultrasound pressure fluctuations impose alternating rarefaction and compression cycles in the liquid medium, leading to the formation and catastrophic collapse of microbubbles. This concerted physicochemical cascade acts synergistically to promote exhaustive release of phytochemicals from the cellular architecture [16], [17].

Conventional RSM is incapable of adequately resolving the inherent nonlinearity and higher–order interactions in extraction systems, by which its predictive reliability is undermined. An integrated RSM–ANN–GA framework was thus established to circumvent this deficiency [18]. Within this hybrid framework, RSM serves as the workhorse for experimental design and preliminary model construction, ANN undertakes the task of delineating complex nonlinear response surfaces, and GA performs a global traversal of the parameter hyperspace for optimal search. This tripartite orchestration substantially elevates both the predictive fidelity and process optimization efficacy for multifaceted extraction systems [19].

In summary, the present study was undertaken with the following objectives: (i) to construct a STDES as an integrated extraction platform for LBT from C. tangshen; (ii) to screen and optimize the extraction conditions—including STDES composition, water content, liquid–to–solid ratio, ultrasound time, ultrasound power, and vortex time—through a hybrid RSM–ANN–GA approach; (iii) to comprehensively annotate the bioactive components in the resulting extracts using UHPLC–Q–TOF–MS; (iv) to elucidate the underlying solvent–matrix interaction mechanisms via FT–IR, SEM, and MD simulations; and (v) to evaluate the in vitro antioxidant activity of the optimized extracts. By fulfilling these objectives, this work not only establishes a green and high–efficiency extraction protocol for LBT but also provides a mechanistic and compositional foundation for the expanded utilization of C. tangshen in functional foods and nutraceuticals [20].

2. Experimental

2.1. Materials

The dried roots of C. tangshen (batch No. 20250822) were collected from Nongba Township, Wuxi County, Chongqing, China, and taxonomically authenticated by Prof. Chen Shaocheng at Chongqing University of Education. Reference standard LBT was sourced from Chengdu Biopurify Phytochemicals Ltd. Additional reagents, including 1,1-diphenyl-2-picrylhydrazyl (DPPH), 2,2′-azinobis(3-ethylbenzothiazoline-6-sulfonic acid) diammonium salt (ABTS), hydroxypropyl-β-cyclodextrin, choline chloride, betaine, glycerol, lactic acid, acetic acid, urea, malic acid, and ascorbic acid, were acquired from Shanghai Yuanye Biotechnology Co., Ltd.

2.2. TDES fabrication and preliminary screening

Twenty ternary systems were formulated from various HBA–HBD pairs. Each mixture was maintained at 70 °C with constant agitation for 1–2 h to ensure complete homogenization, with deionized water added whenever necessary. Table 1 provides the complete compositional data for all prepared TDESs [21], [22].

Table 1.

TDES compositions.

No.	Abbreviation	HBA	HBD	Molar ratio
1	ChB:AA:Gly	Choline bitartrate	Acetic acid : Glycerol	1:2:2
2	ChB:EG:LA	Choline bitartrate	Ethylene glycol : Lactic acid	1:2:2
3	ChB:EG:LA	Choline bitartrate	Ethylene glycol : Lactic acid	1:2:2
4	ChB:PG:LA	Choline bitartrate	Propylene glycol : Lactic acid	1:2:2
5	ChB:BG:LA	Choline bitartrate	Butylene glycol : Lactic acid	1:2:2
6	ChB:U:LA	Choline bitartrate	Urea : Lactic acid	1:2:2
7	ChB:PG:Gly	Choline bitartrate	Propylene glycol : Glycerol	1:2:2
8	ChB:LA:Gly	Choline bitartrate	Lactic acid : Glycerol	1:2:2
9	ChB:U:Gly	Choline bitartrate	Urea : Glycerol	1:2:2
10	ChCl:U:Gly	Choline chloride	Urea : Glycerol	1:2:2
11	ChCl:LA:U	Choline chloride	Lactic acid : Urea	1:2:2
12	ChCl:AA:U	Choline chloride	Acetic acid : Urea	1:2:2
13	ChCl:LA:Gly	Choline chloride	Lactic acid : Glycerol	1:2:2
14	ChCl:MA:Gly	Choline chloride	Malic acid : Glycerol	1:2:2
15	Bet:Gly:U	Betaine	Glycerol : Urea	1:2:2
16	Bet:MA:U	Betaine	Malic acid : Urea	1:2:2
17	Bet:LA:Gly	Betaine	Lactic acid : Glycerol	1:2:2
18	HP-β-CD:U:LA	Hydroxypropyl-β-cyclodextrin	Urea : Lactic acid	1:15:15
19	HP-β-CD:PG:Gly	Hydroxypropyl-β-cyclodextrin	Propylene glycol : Glycerol	1:15:15
20	HP-β-CD:LA:Gly	Hydroxypropyl-β-cyclodextrin	Lactic acid : Glycerol	1:15:15

A preliminary assessment was then carried out to rank the extraction efficacy of the twenty TDES candidates. In each run, 3.0 g of dried C. tangshen powder was combined with 90 mL of the TDES (supplemented with 40% v/v water), vortexed for 12 min, and exposed to ultrasound irradiation for 30 min. The resulting mixtures were immediately filtered, and the clear filtrates were retained for further use.

2.3. Analytical procedures

2.3.1. HPLC assay

Prior to HPLC analysis, the TDES extract was appropriately pretreated to remove the viscous DES matrix and prepare the sample for injection. Briefly, an aliquot of the supernatant obtained after extraction (Section 2.2) was diluted with HPLC-grade methanol at a volume ratio of 1:5 (extract/methanol), vortexed for 2 min, and then centrifuged at 12,000 rpm for 10 min at 25 °C to precipitate any insoluble residues. The resulting supernatant was filtered through a 0.22 μm nylon syringe filter prior to injection.

HPLC separation of LBT was conducted on an Agilent 1220 system using the gradient program in Table 2, with a flow rate of 0.6 mL/min at 30 °C. Chromatographic separation was achieved on an SB–C18 column (5 μm, 250 × 4.6 mm), with 20 μL injections and UV detection at 266 nm (Fig. 1).

Table 2.

HPLC gradient elution.

Time(min)	Acetonitrile (%)	0.2 % phosphoric acid (%)
0	20	80
10	20	80
40	75	25

Fig. 1.

The standard solution (top) and test sample solution (bottom); the peak labeled A correspond to LBT.

LBT yields (%)$=\frac{\text{mean masse of LBT (mg)}}{\text{mean masses of C. tangshen powder}}\text{×100\%}$

2.3.2. UHPLC-Q-TOF-MS-Based Metabolite Profiling

The chemical constituents of the C. tangshen extracts were profiled on an Agilent 1290 UHPLC system interfaced with an Agilent 6546 Q–TOF mass spectrometer. The separation was conducted on a Waters ACQUITY UPLC HSS T3 column (2.1 × 100 mm, 1.8 μm) thermostated at 30 °C, using the gradient program specified in Table S5 with a 1 μL injection [23].

2.4. Process Optimization

2.4.1. Preliminary Screening and RSM Design

A series of single–factor assays was carried out using 3.0 g of C. tangshen powder per run with the optimal STDES, aiming to clarify how five independent parameters (Table S4) affect LBT recovery. For each variable tested, its level was progressively adjusted while all other experimental parameters remained unchanged.

A three–level, three–factor Box–Behnken design was adopted for response surface optimization targeting maximum compound recovery. The factors examined (Table 3) were incorporated into a quadratic model expressed as:

$$Y={\beta }_0+\sum _{i=1}^i\left({\beta }_i{x}_i\right)+\sum _{\mathit{ij}=1}^{\mathit{ij}}\left({\beta }_{\mathit{ij}}{X}_i{X}_j\right)+\sum _{1=1}^i\left({\beta }_{\mathit{ii}}{X}^2\right)$$

Table 3.

ANOVA analysis.

Source	Sum of Squares	Mean Square	F-value	p-value
Model	0.241	0.0268	133.37	< 0.0001	significant
A	0.0008	0.0008	4	0.0856
B	0	0	0.0748	0.7924
C	0.1264	0.1264	629.84	< 0.0001
AB	0.0018	0.0018	8.86	0.0206
AC	0.0001	0.0001	0.4701	0.515
BC	0.0001	0.0001	0.4726	0.5139
A2	0.0029	0.0029	14.44	0.0067
B2	0.004	0.004	19.8	0.003
C2	0.0997	0.0997	496.84	< 0.0001
Residual	0.0014	0.0002
Lack of Fit	0.0007	0.0002	1.32	0.3854	not significant
Pure Error	0.0007	0.0002
Cor Total	0.2424

where Y, β₀, β_ij, and X denote the yield, intercept, coefficients, and factors, respectively.

2.5. Hybrid RSM–ANN–GA optimization platform

The inability of RSM to adequately capture the nonlinear behavior intrinsic to ultrasound–assisted extraction, attributable to its quadratic polynomial formulation, was addressed by constructing a hybrid modeling framework that integrated RSM, ANN, and GA. A Box–Behnken design supplied the initial training dataset, which was then used to train a multilayer perceptron–based ANN for high–fidelity prediction of LBT yields. The trained ANN served as the objective function for a GA–driven global search over the entire parameter space. This synergy preserves the statistically sound experimental foundation of RSM while leveraging the ANN's superior approximation capacity and the GA's global optimization capability, yielding a more robust and accurate predictive model [23].

A BP–ANN was developed within the MATLAB R2024a environment, and its hyperparameters were optimized using the built–in GA toolbox. Training utilized the RSM–derived experimental datasets, which furnished the input–output pairs needed to characterize the extraction system—three extraction variables served as inputs, while the combined LBT yield was designated as the output response [24].

A 3–9–1 BP–ANN was designed (Fig. 2), with the hidden–layer size optimized via sensitivity testing to ensure a sound balance between model capacity and generalization. Network training relied on the Levenberg–Marquardt algorithm applied to an RSM–generated dataset, with a 70/15/15 split for training, validation, and testing purposes. Selection of the final architecture was driven by minimum MSE and maximum R on validation data. A GA post–training optimization step was then executed to fine–tune the network parameters, enhancing both predictive precision and model stability [25], [26].

Fig. 2.

Topology of the BP-ANN model.

2.6. Method validation

2.6.1. Methanol–assisted ISUSE

A 3.0 g portion of plant powder was mixed with 90 mL of methanol, and the resulting dispersion was vortexed for 12 min prior to ultrasonic irradiation (240 W, 40 kHz, 50 °C) for 30 min. Then filtered, and the filtrate was centrifuged at 4000 rpm for 10 min to obtain the clear supernatant, which was subsequently assayed by HPLC [23].

2.6.2. UAE

A 3.0 g aliquot of the powdered material was immersed in 90 mL of methanol and irradiated with ultrasound at 240 W and 40 kHz (50 °C) for 30 min. Following centrifugation at 4000 rpm for 15 min, the supernatant was separated for subsequent chromatographic analysis [27].

2.6.3. ME

Powdered plant material (3.0 g) and methanol (90 mL) were combined in a 100 mL flask and macerated statically at 25 °C for 8 h. Following centrifugation of the macerate, the clear supernatant was collected for HPLC quantification.

2.7. Mechanistic insights

2.7.1. FT–IR

FT–IR spectroscopy was employed to probe the hydrogen–bonding interactions present in the optimized STDES formulation. Spectral acquisition was performed over the range of 4000–400 cm⁻¹.

2.7.2. SEM

SEM imaging was adopted to observe the surface features of the C. tangshen residues generated by different extraction methods, with the goal of linking morphological alterations to extraction efficiency.

2.7.3. MD simulation

MD simulations enable real-time, atomic-level observation of intermolecular behavior within the extraction system. To elucidate the molecular basis for the enhanced extraction efficiency of LBT achieved with the STDES-ISUSE protocol, all-atom MD simulations were performed to compare the solvation behavior of LBT in the best-performing STDES with that in methanol. Emphasis was placed on the intermolecular forces governing solute partitioning and solvation thermodynamics, including hydrogen bonding, van der Waals interactions, and electrostatic contributions. These forces offer a microscopic mechanistic basis for the differences in extraction yields observed in wet experiments. Complete simulation parameters are given in the Supporting Information.

Radial distribution functions (RDFs), cumulative coordination numbers, and density evolution were calculated from the production trajectories using GROMACS.

The simulation setup and force-field parameterization, with topologies generated by Sobtop and partial charges assigned by Multiwfn, followed the procedure reported in a recent MD study of herb-derived phytochemical assemblies [49]. In that study, short-range Lennard-Jones (LJ-SR) and Coulomb (Coul-SR) interaction energies were likewise obtained from GROMACS trajectories using gmx energy. The detailed calculation parameters and results are provided in the Supplementary Information.

2.8. Antioxidant activity

2.8.1. DPPH

According to the research of Iftikhar [28], sample solutions spanning 0.02–0.1 mg/mL were prepared, and 2 mL aliquots were combined with an equal volume of methanolic DPPH (0.15 mmol/L). After 30 min of incubation in the dark at ambient temperature, the absorbance was recorded at 517 nm. The percentage of radical scavenging was derived from the following formula:

$$\text{S}\text{cavenging rate}\left(\text{\%}\right)=\left[\frac{\text{Ac-As}}{\text{Ac}}\right]\text{*100\%}$$

where Ac = control absorbance and As = sample absorbance.

2.8.2. ABTS

According to the research of Shu [29], ABTS⁺ was generated from 7 mmol/L ABTS and 2.45 mmol/L potassium persulfate (1:1, v/v) after 16 h incubation at room temperature in the dark. The stock was then diluted with ethanol to an absorbance of 0.70 ± 0.02 at 734 nm prior to use. For each measurement, 3 mL of the freshly diluted ABTS⁺ reagent was combined with 0.5 mL of sample (0.02–0.1 mg/mL), and the reaction was allowed to proceed for 6 min in the dark. The final absorbance was then read at 734 nm, and the scavenging percentage was derived from the following equation:

$$\text{S}\text{cavenging rate}\left(\text{\%}\right)=\left[\frac{\text{A}\text{0}\text{-A}\text{1}}{\text{A}\text{0}}\right]\text{*100\%}$$

where A₀ = blank absorbance and A₁ = sample absorbance.

2.8.3. Ferric ion reducing Assay

The reducing power of the extracts was evaluated following the method of Iqbal et al [30] using a concentration gradient of 0.02–0.1 mg/mL. Sample (3 mL), phosphate buffer (0.2 mol/L, pH 6.6, 2 mL), and 1% potassium ferricyanide (2 mL) were mixed and incubated at 50 °C for 30 min. The reaction was stopped with 10% trichloroacetic acid (2 mL) and centrifuged (3000 rpm, 10 min). Supernatant (2 mL) was then mixed with 0.1 mL of 0.1% ferric chloride and 2 mL of deionized water. After 10 min, the absorbance was measured at 700 nm.

2.9. Statistical evaluation

The significance of the RSM-derived models was assessed via one–way analysis of variance (ANOVA). All experimental results are expressed as the mean ± standard deviation (SD).

3. Results

3.1. Identification of the most effective TDES

A comparative evaluation of the twenty TDES formulations (each containing 40% v/v water) was carried out under identical extraction conditions to identify the most effective solvent system for LBT recovery from C. tangshen. As shown in Fig. 3, the extraction efficiencies varied markedly across the tested formulations, underscoring the pronounced sensitivity of the extraction process to the specific composition of the TDES. Among all candidates, the HP–β–CD:LA:Gly system (DES–20) consistently afforded the highest LBT yield, significantly surpassing the other nineteen formulations. In contrast, conventional methanol extraction yielded only 0.34 ± 0.012 mg/g under the same conditions, a value that aligns well with the classical “like–dissolves–like” solubility principle—given LBT's moderate polarity, it exhibits limited solubility in pure organic solvents. The superior performance of DES–20 can be tentatively attributed to the synergistic hydrogen–bonding network established among HP–β–CD, lactic acid, and glycerol, which likely enhances both the solubilization and mass transfer of the target analyte. Based on these findings, DES–20 was screened as STDES for subsequent optimization and mechanistic studies [31], [32].

Fig. 3.

Extraction efficiency of TDES.

3.2. Single-factor experiments

Water content is a critical parameter governing TDES performance, as it modulates solvent viscosity and, consequently, the mass–transfer efficiency during extraction. For the optimal STDES identified above, we anticipated that incremental water addition would reduce viscosity and thereby improve LBT recovery. To test this hypothesis, we varied the water content across a preset range and monitored the resulting changes in extraction yield (Fig. 4) [6], [33]. A clear nonlinear correlation emerged between LBT yield and water content when the latter was varied systematically from 20% to 60% (Fig. 4A). In the lower range (20–40%), increasing water addition led to a progressive rise in extraction efficiency, with the maximum observed at 40%. This improvement is readily explained by the viscosity–lowering effect of water: as the solvent became less viscous, mass transfer accelerated, allowing more efficient release of LBT from the plant matrix. Beyond this optimal point, however, further elevation of water content to 50–60% produced a substantial drop in yield. The most plausible explanation for this decline lies in an excessive dilution of the system, which weakens the intermolecular hydrogen–bonding network essential for solute–solvent interactions and thus compromises the TDES's ability to dissolve the target compounds [29].

Fig. 4.

Single factor experiments.

The liquid/solid ratio was investigated over a range of 10 to 50 mL/g (Fig. 4B) to determine its impact on LBT yield. Extraction efficiency improved consistently as the ratio advanced from 10 to 30 mL/g, with the peak recovery observed at the latter point. This increase was consistent with enhanced mass transfer and improved solvent accessibility to the plant matrix at higher solvent volumes. Below the optimum, insufficient solvent likely limited solute diffusion and reduced extraction completeness. When the ratio exceeded 30 mL/g, however, a progressive loss of extraction efficiency was noted. This downward trend appeared to stem from two concurrent phenomena: the enhanced leaching of competing phytochemicals at high solvent loads, and the attenuation of ultrasound energy transmission through the larger liquid volume, which weakened the disruptive effect on cell walls and ultimately lowered the extraction yield [34], [35].

Ultrasound time is another critical parameter affecting extraction outcome, as it regulates cavitation–induced cell wall rupture and the corresponding mobilization of LBT from the plant tissue. The yield profile obtained across different sonication durations (Fig. 4C) revealed a progressive enhancement as the treatment time extended from 10 to 30 min, with the peak recovery recorded at the latter point. Extension of the irradiation beyond this optimal duration, however, resulted in a considerable reduction in extraction efficiency [36]. A plausible explanation for this decrease involves thermal degradation of the target analyte, since sustained cavitational activity generates localized hotspots that may destabilize heat–sensitive LBT molecules during prolonged sonication [37].

Ultrasound power constitutes a key operational parameter governing cavitation intensity, which in turn determines the effectiveness of ultrasound–assisted extraction. The relationship between power input and LBT recovery was examined over a range of 180–300 W (Fig. 4D). Extraction yield exhibited a positive response to increasing power up to 240 W, where the maximum value was recorded. Operation at power levels exceeding this optimum, however, yielded no appreciable improvement in extraction performance. This observation pointed to the onset of cavitation saturation, a state in which further energy input no longer promotes additional cell wall rupture or mass transfer enhancement. The plateau observed at higher power settings was likely due to energy dissipation, as excessive ultrasound irradiation attenuates cavitation efficiency within the liquid medium [29], [33].

Vortex mixing boosts extraction efficiency through improved solvent–matrix dispersion and accelerated dissolution of target bioactives. The influence of mixing duration on LBT recovery was evaluated over a range of 4–12 min (Fig. 4E). Extraction yield displayed a progressive increase with extending vortex time up to 10 min, after which the recovery stabilized despite continued mixing. The initial upward trend reflected the rapid and efficient release of LBT from the plant matrix during the early stage of vortex treatment. The subsequent plateau observed at prolonged mixing times was most likely due to competing dissolution of non–target matrix constituents, which altered the partitioning equilibrium of the target compound and thereby restricted any further improvement in extraction performance [16].

3.3. RSM

To optimize the extraction conditions, a three–factor, three–level Box–Behnken experimental design was constructed, with Design–Expert software managing the generation of the design matrix and subsequent regression analysis. The resulting experimental layout, along with the measured LBT yields for each run, is provided in Table S5. The quadratic model obtained from the BBD was subjected to ANOVA (Table 3), which yielded a highly significant overall regression (p < 0.001) [38], [39]. The model's strong explanatory power was supported by an R² of 0.9942 and an adjusted R² of 0.9867, demonstrating its ability to effectively describe the relationship between the independent variables and the extraction response. Moreover, the lack–of–fit test was statistically insignificant (p > 0.05), reinforcing the model's adequacy and its utility for predictive purposes.

Pairwise interactions between water content (A), ultrasound time (B), and liquid/solid ratio (C) all exerted positive synergistic effects on LBT extraction efficiency, implying that appropriate combinations of these variables could boost the recovery. The quadratic model in terms of coded factors was expressed as follows:

Y=0.5520-0.0100A-0.0014B+0.1257C+0.0211AB-0.0049AC+0.0049BC-0.0262A²-0.0307B²-0.1539C²

Fig. 5 illustrated the three–dimensional response surfaces constructed for the pairwise interactions AB, BC, and AC. All three plots showed a concave curvature with downward openings, confirming the existence of a maximum predicted yield within the investigated parameter space [40]. The gradient of each response surface provided a measure of the relative influence exerted by the corresponding variable pair on LBT recovery [41]. A comparison of the three surfaces revealed that the BC interaction exhibited the steepest gradient, indicating that the combined effect of ultrasound time and liquid/solid ratio dominated the extraction process. Optimization through Design–Expert generated the following recommended settings: 36.43% water content, 28.70 min sonication, and 33.50 mL/g solvent–to–solid ratio. At these levels, the model predicted a maximum LBT yield of 0.578 mg/g.

Fig. 5.

Response surface plots.

3.4. RSM-ANN-GA analysis

A three-layer BP-ANN, with three extraction parameters as inputs and LBT yield as the output, was developed from the RSM experimental dataset [33]. Input–to–hidden activation was achieved via tansig, with a linear function for the hidden–to–output stage. The Levenberg–Marquardt algorithm (trainlm) was applied for training over 1000 epochs (25 iterations per epoch) until an MSE of 0.0001 was reached. The network achieved its best validation score at epoch 2, yielding an MSE of 0.006245 (Fig. S1). A correlation coefficient (R) exceeding 0.90 (Fig. 6) confirmed the model's robust predictive capacity.

Fig. 6.

Regression plot of the BP-ANN model.

Optimization criteria from the RSM–ANN–GA hybrid model were established with the three independent variables bounded as follows:

30 ≤ X₁ ≤ 50, 20 ≤ X₂ ≤ 40, 20 ≤ X₃ ≤ 40

The hybrid approach gave the following theoretical optima: 46.10% water, 34.33 min sonication, and 29.41 mL/g liquid‑to‑solid ratio. Under these conditions, the maximum yield was predicted to be 0.619 mg/g. To facilitate convenient implementation and ensure experimental repeatability, these values were subsequently modified to 46% water, 29 mL/g, and 34 min for the confirmatory runs.

3.5. Validation experiment

Validation experiments were carried out under the optimal conditions derived from both the RSM model and the hybrid RSM–ANN–GA model, with minor adjustments made to accommodate laboratory accessibility. For the RSM–optimized protocol, the conditions were slightly modified to 36% water content, 33 mL/g liquid/solid ratio, and 29 min ultrasound time. For the RSM–ANN–GA approach, the parameters were set to 46% water content, 29 mL/g liquid/solid ratio, and 34 min ultrasound time. The results are summarized in Table 5.

Table 5.

Method validation

Analyte	LOD (μg/ mL)	LOQ (μg/mL)	Intraday precision	Interday precision	Stability	Repeatability	Recovery
LBT	0.45	1.34	0.96	2.24	1.77	1.82	99.94

Both optimization strategies demonstrated strong predictive capability for LBT recovery, as evidenced by the close agreement between the predicted and experimental values. The RSM–optimized method yielded an average recovery of 0.576 mg/g with a RSD of 2.30%. Furthermore, the RSM–ANN–GA hybrid model afforded a higher average recovery of 0.613±0.011 mg/g with a lower RSD of 1.80%, indicating that the combined approach offered superior accuracy and stability.

3.6. Benchmarking against conventional techniques

The superiority of the STDES–ISUSE method over conventional extraction techniques for LBT recovery from C. tangshen was confirmed through a side–by–side comparative evaluation. The data presented in Fig. 7 demonstrated that STDES–ISUSE consistently produced markedly higher yields than maceration (ME) while requiring substantially less processing time. When compared with the other approaches tested, STDES–ISUSE also yielded superior recoveries relative to ISUSE (0.34 ± 0.012 mg/g) and UAE (0.30 ± 0.019 mg/g). Notably, the STDES–ISUSE protocol combines rapidity, high efficiency, and eco–friendliness, positioning it as a highly competitive strategy for the sustainable recovery of bioactive compounds from C. tangshen.

Fig. 7.

Extraction efficiency of LBT by four different methods.

3.7. MS analysis

The chemical profile of the C. tangshen extracts was established using UHPLC–Q–TOF–MS. To the best of our knowledge, this study constitutes the first systematic phytochemical interrogation of C. tangshen extracts via this analytical platform, with the objective of pinpointing its bioactive components [42]. Fig. 8 displays the total ion chromatograms acquired under positive and negative ionization conditions. Putative identification of the detected peaks was accomplished by cross-referencing the MS/MS fragmentation data against the Pesticide Compound Database (PCDL) and previously published spectral information. In total, 20 compounds were assigned, spanning six chemical categories: phenolic acids (2), polyacetylenic glycosides (2), fatty acids (5), glycosides (2), iridoids (1), and alkaloids (1) (Table 4). Among the identified peaks, peak 1 was unequivocally characterized as LBT. The wealth of spectral data presented here constitutes a valuable chemical resource for advancing bioactivity research and facilitating product innovation related to C. tangshen.

Fig. 8.

UHPLC–Q–TOF–MS chromatograms of C. tangshen extracts under positive (A) and negative (B) ionization.

Table 4.

UHPLC-Q-TOF-MS results.

Peak	RT(min)	Model	Compound	Molecular formula	Theoretical (m/z)	Major fragment ion
1	3.126	+	Codonopsine	C14H21NO4	267.1471	161.0596,118.0863,88.0756,58.0652
2	4.693	+	L-Tryptophan	C8H14NO5	204.0872	188.0706,146.0600,118.0651,91.0540
3	10.485	+	(2R)-6-[2-[2-(2-methoxyethoxy)ethoxy]ethoxy]-2-[(4-methylidenepyrrol-2-yl)methyl]hex-3-ynoic acid	C19H27NO6	365.1838	266.1387,177.0547,88.0756
4	11.815	+	ethyl 2-[4-acetyl-N-(2,2-diethoxyethyl)anilino]acetate	C18H27NO5	337.1889	205.0860,161.0597,88.0756
5	12.113	+	diethyl 2-[2,2-dimethoxyethyl(1-phenylethyl)amino]propanedioate	C19H29NO6	367.1995	266.1388,221.0810,177.0546,88.0756
6	13.561	+	hexyl-gentiobioside	C18H34O11	426.2101	163.0600,145.0496,127.0390,85.0284
7	13.799	+	Tangshenoside I	C29H42O18	678.2371	193.0861,161.0597,133.0648,105.0701
8	14.633	+	Tert-butyl 2-[3-(3-morpholin-4-yl-2-oxopropyl)phenoxy]acetate	C19H27NO5	349.1889	250.1438,205.0859,161.0596,88.0755
9	16.161	+	Hexyl pentosyl glucoside	C17H32O10	396.1995	287.1462,205.0009,140.0100
10	16.439	-	Hexyl β-D-Glucopyranoside	C12H24O6	264.1573	263.1492,207.0311,101.0250,59.0139
11	16.875	+	2-[4-[(2S)-2-acetyloxy-3-(tert-butylamino)propoxy]phenyl]ethyl acetate	C19H29NO5	351.2046	251.1471,162.0630,88.0755
12	17.411	-	Lobetyolinin	C26H38O13	558.2312	467.1786,323.0987,179.0562,89.0245
13	21.043	+	Lobetyolin	C20H28O8	396.1784	199.1119,155.0855,128.0621,93.0334
14	22.889	+	6‴-trans-p-Coumaroyl tangshenoside I	C38H48O20	824.2739	655.1845,493.1317,161.0596
15	28.943	+	9,12,13-trihydroxy-10,15-octadecadienoic acid	C18H32O5	328.225	281.1368,194.1173,105.0690
16	30.571	+	9,12,13-trihydroxy-10-octadecenoic acid	C18H34O5	330.2406	277.2162,195.1381,155.1066,81.0697
17	37.500	+	N-(1,3-dihydroxyhexadecan-2-yl)acetamide	C18H37NO3	315.2773	298.2742,280.2636,60.0443
18	40.101	+	9,10-dihydroxy-12-octadecenoic acid	C18H34O4	314.2457	279.2317,161.1331,95.0856,81.0697,67.0542
19	44.988	-	9-hydroxy-10,12-octadecadienoic acid	C18H32O3	296.2351	277.2175,195.1392,171.1034,113.0972
20	45.962	+	13-Oxo-9E,11E-octadecadienoic acid	C18H30O3	294.2195	235.1686,179.1429,93.0697,71.0854

3.8. Method validation

For accurate quality assessment of the extracts, we conducted quantitative measurements under the optimized conditions determined from the preceding extraction studies. The analytical method was validated through calibration curves established with the authentic LBT standard. The curves yielded high determination coefficients (R²), confirming that the target analyte displayed good linearity across the concentration intervals examined [43].

A linear regression equation of Y = 20110x + 42.648 (R² = 0.9999) was obtained for LBT. The LBT demonstrated wide linear response ranges between 20 and 100 μg/mL, indicating excellent linear relationships. The LOD and LOQ values for LBT were calculated as 0.45 μg/mL and 1.34 μg/mL. These findings validated the HPLC method as sufficiently sensitive and fit for purpose in quantifying the target analytes [44].

Analytical precision was evaluated via intra–day (six replicate injections within one day) and inter–day (once daily for six days) tests. Both yielded RSD values below 2%, confirming the method's high reproducibility. Stability testing of the extracts at 25 °C over 12 h—with sampling at 0, 3, 6, 9, and 12 h—also produced RSDs under 2%, demonstrating satisfactory short–term stability under the tested conditions. The repeatability of the method was assessed through six replicate measurements, which yielded an RSD of 1.82% and attested to the procedure's satisfactory reproducibility. Recovery experiments were performed by fortifying 3.0 g of plant matrix with LBT standard at known concentrations. The recoveries (Table 5) confirmed the method's reliability and accuracy.

3.9. Extraction mechanism

3.9.1. FT-IR

The STDES was formed by pairing a single HBA with two HBDs. The FT–IR spectra of the pristine components and the resulting TDES were contrasted to probe the intermolecular interactions and structural evolution accompanying solvent assembly.

Fig. 9 displayed characteristic bands attributable to O–H stretching (broad envelope at 3500–3200 cm⁻¹), C–H vibrations (3000–2500 cm⁻¹), and C–O stretching (1112–1000 cm⁻¹) [45]. In the STDES spectrum, the C–N stretch of hydroxypropyl–β–cyclodextrin appeared at 1230 cm⁻¹, representing a clear shift from its original position. The ternary system showed a much broader O–H band with a blue shift versus the neat precursors. Overall, these spectral features were consistent with the establishment of extensive hydrogen–bonding networks and confirmed the successful fabrication of the TDES [25].

Fig. 9.

FT–IR characterization.

Beyond the chemical interactions evidenced by FT–IR spectroscopy, the physical effects induced by ultrasonic irradiation prove equally critical in promoting solvent–matrix contact and augmenting extraction performance. Rueangsri et al. [15] demonstrated that lower ultrasonic frequencies (e.g., 12 kHz) facilitate the generation and collapse of larger cavitation bubbles at the material surface, thereby generating intense micro–jets and pronounced shear forces that effectively disrupt plant cell walls and enhance mass transfer efficiency. In contrast, higher frequencies yield smaller cavitation bubbles with shortened oscillation cycles and attenuated cavitation intensity, which may compromise extraction yields. These observations collectively underscore the exigency of optimizing ultrasonic parameters—particularly frequency—in DES–based extraction systems, wherein the synergistic interplay between chemical solvation (governed by hydrogen bonding) and physical disruption (driven by cavitation) dictates the overall extraction outcome.

3.9.2. SEM

A dual–mode extraction strategy integrating vortex mixing with ultrasonication was employed to improve LBT recovery. Vortex action at the interface promoted solvent spreading and triggered the early efflux of analytes from the cells. Concurrently, the applied ultrasound produced cyclic pressure fluctuations in the medium, triggering cavitation events that released mechanical energy capable of rupturing cell walls. These two physical processes operated in a complementary fashion: the vortex–driven convective flow improved solvent distribution, while the cavitational action from ultrasound facilitated extensive disintegration of the tissue structure. This cooperative effect enabled substantially greater liberation of intracellular components, as evidenced by the prominent morphological alterations visible in the SEM micrographs (Fig. 10A).

Fig. 10.

SEM micrographs (A) STDES–ISUSE, (B) ISUSE, (C) UAE, and (D) ME.

Crystalline cellulose constitutes the principal scaffolding material of plant cell walls, and its disruption represents a critical step toward liberating intracellular metabolites. The STDES containing hydroxypropyl–β–cyclodextrin proved distinctly more effective at decomposing this structural polymer than did the conventional solvent systems. This superiority stemmed from the capacity of the DES components to form non–covalent associations with cellulose chains, which competitively weakened the existing hydrogen–bonded framework and lowered the degree of crystallinity within the microfibrillar architecture [46]. The ensuing mechanical vulnerability of the wall matrix was evident in the SEM micrographs (Fig. 10A–D), where specimens receiving STDES–ISUSE displayed far greater structural disintegration than those processed by standard techniques. The extent of this physical breakdown paralleled the observed increments in phytochemical yield, substantiating the proposed function of the ternary DES in coordinating both matrix dismantling and solute mobilization [46].

3.9.3. MD analysis

MD simulations were undertaken to rationalize the marked improvement in LBT recovery conferred by the STDES–ISUSE protocol. The computational campaign juxtaposed the optimally performing STDES against methanol, scrutinizing their respective solvation behaviors toward LBT—the signature phytochemical marker of C. tangshen. Particular attention was devoted to the intermolecular forces that govern solute partitioning and solvation thermodynamics, chiefly hydrogen bonds, van der Waals contacts, and electrostatic contributions. The simulated trajectories unequivocally demonstrated that the STDES established a substantially denser and more persistent solvation environment around LBT than its methanolic counterpart. These in silico findings furnished a compelling molecular–scale narrative for the superior extraction outcomes realized in the wet–lab experiments [47].

The MD simulations uncovered distinctly different solvation patterns for LBT when dissolved in the STDES versus methanol. Over the 100–ns simulation, LBT remained well dispersed throughout the STDES medium, whereas substantial solute clustering emerged in the methanolic environment (Fig. 11A, B)—a finding that echoes the experimentally determined solubility advantage of the STDES. Deeper analysis of the interaction dynamics reinforced this interpretation: the STDES supported a higher population of LBT–solvent hydrogen bonds, and these non–covalent contacts exhibited appreciably greater stability, as demonstrated by the sustained high hydrogen–bond counts (Fig. 11C) and the prolonged bond lifetimes (Fig. 11D). These observations rationalize the superior extraction efficacy of the STDES on mechanistic grounds, attributed to its thermodynamically enhanced solvation capacity [48].

Fig. 11.

MD snapshots of LBT in STDES (A) and methanol (B), with hydrogen–bond statistics (C) and lifetime profiles (D).

To further characterize the structural organization of the simulated systems, RDFs and cumulative coordination numbers were calculated for the oxygen atoms of LBT. In the STDES system, the first-peak height of the RDF between oxygen atoms of LBT (approximately 275 in g(r) units) was lower than that in the methanol system (approximately 307), and the cumulative coordination number was also lower. This indicates a weaker self-aggregation tendency and more uniform dispersion of LBT in the STDES (Figs. S2-S5). The density evolution curves for both systems (Figs. S6 and S7) remained stable throughout the 100 ns production run, with average values of approximately 1193.76 kg/m³ for STDES and 840.50 kg/m³ for methanol, confirming that the simulations were fully equilibrated [50], system stability was assessed by monitoring structural and thermodynamic quantities over a 100 ns production run; the convergence of energy, density, and RDFs observed here indicates that both systems were fully equilibrated. These structural analyses are presented in Supplementary Figs. S2–S7.

Nonbonded interaction energies were calculated for each system to quantify the intermolecular affinities that define the solvation environment described above. The short-range Coulomb (Coul-SR) and Lennard-Jones (LJ-SR) terms were extracted from the production trajectories (Figs. S8-S13; Table S6). For System 1 (STDES), all total interaction energies between the individual components and water remained negative throughout the 100 ns production run, indicating net attraction (Fig.12). The largest absolute values were found for lactic acid-water (-117,826 kJ/mol) and glycerol-water (-110,447 kJ/mol), which can be attributed to the high numbers of hydroxyl and carboxyl groups (900 molecules each) that establish strong hydrogen-bond networks with water. HP-β-CD-water (-34,782 kJ/mol) ranked intermediate, reflecting electrostatic interactions between water and the hydroxyl groups at the rim of its cavity. LBT-water (-2,068 kJ/mol) exhibited the smallest absolute value, in line with the low LBT loading (10 molecules) and the hydrophobic character of the solute. In every pair, electrostatic contributions exceeded van der Waals terms. For System 2 (methanol), the LBT-methanol interaction energy was -4,833 kJ/mol (electrostatic -2,932; van der Waals -1,901 kJ/mol; Table S6; Figs. S12-S13). This value was also negative and electrostatically dominated, pointing to favorable solvation compatibility between LBT and methanol. Since the interaction-energy objects are not the same in the two systems (component-water pairs versus the solute-solvent pair), absolute values were not compared directly. Instead, the advantage of STDES is supported by the structural and dynamic evidence presented above, together with the experimental extraction yields. Details on the interaction energy of the methanol system are available in Figs. S12-S13 and Table S6.

• Antioxidant activity

Fig. 12.

Interaction energies between each component and water in System 1 (STDES).

DPPH⁺ and ABTS⁺ scavenging, together with ferric reducing activity, were used as three parallel assays to assess the antioxidant profile of the extracts. Over the 0.02–0.1 mg/mL concentration range, the extracts neutralized both radicals in a strictly dose–dependent fashion (Fig. 13A, B). Maximal intervention at 0.1 mg/mL produced scavenging ratios of 87.39 ± 0.055% toward DPPH⁺ and 97.12 ± 1.24% toward ABTS⁺, attesting to the extracts' potent free–radical quenching capability. The reference compound, ascorbic acid, maintained >90% inhibition across both assay systems, confirming the robustness of the experimental setup [51], [52].

Fig. 13.

Antioxidant evaluation of the extracts: (a) DPPH, (b) ABTS, and (c) reducing power.

The electron–donating capacity of the extracts, evidenced by their proficiency in mediating the Fe³⁺→Fe²⁺ conversion, was likewise corroborated (Fig. 13C). At 0.1 mg/mL, the optical density attained 0.437 ± 0.0050, reflecting appreciable reducing potential. Viewed collectively, these data attest to the considerable antioxidative prowess of the C. tangshen extracts and substantiate their prospective utility as naturally derived antioxidant entities in functional food formulations or phytotherapeutic preparations.

4. Conclusion

In this work, a STDES–ISUSE platform was successfully developed as a sustainable and highly efficient strategy for the recovery of LBT from C. tangshen. Through systematic screening of twenty ternary DES candidates, the hydroxypropyl–β–cyclodextrin–based system was identified as the optimal formulation, owing to its superior hydrogen–bonding capacity and solvation performance. A hybrid RSM–ANN–GA optimization framework was established, which demonstrated significantly enhanced predictive accuracy and robustness compared to conventional RSM alone, highlighting the advantage of machine learning–assisted approaches in complex phytochemical extraction processes. Comparative evaluations confirmed that the STDES–ISUSE strategy substantially outperformed conventional extraction techniques, achieving superior yields within considerably shorter processing times. Comprehensive UHPLC–Q–TOF–MS analysis enabled the confident identification of multiple bioactive constituents, including polyacetylenic glycosides. Mechanistic investigations via FT–IR spectroscopy, scanning electron microscopy, and molecular dynamics simulations collectively revealed that the enhanced extraction performance originates from the synergistic interplay of physical disruption and chemical solvation. Antioxidant assays further corroborated the potent free radical scavenging and reducing capacities of the obtained extracts.

Collectively, this study establishes a green, reproducible, and highly efficient extraction protocol that not only advances the sustainable recovery of bioactive compound from C. tangshen but also provides a mechanistic and compositional foundation for its expanded application in functional foods and nutraceutical formulations. By marrying the green solvent properties of supramolecular ternary DESs with the enhanced transport efficiency of ISUSE, this strategy affords a practical, sustainable alternative with direct relevance to natural product research and phytopharmaceutical development.

CRediT authorship contribution statement

Cheng Liu: Writing – original draft, Investigation, Funding acquisition. Yiquan Zhou: Formal analysis. Jin Guo: Investigation. Yijia Chen: Investigation. Xia Zhao: Investigation. Zhiqiang Chen: Software. Xin Zhao: Investigation. Xiaochuan Zou: Investigation. Fangyuan Gong: Software. Ailing Lei: Investigation. Shuxin He: Investigation. Xin Ma: Investigation. Xinyu Liu: Investigation. Jie Lei: Writing – review & editing, Methodology.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following are the Supplementary data to this article: