Metabolite Profiling of Microwave-Assisted Sargassum fusiforme Extracts with Improved Antioxidant Activity Using Hybrid Response Surface Methodology and Artificial Neural Networking-Genetic Algorithm

Sargassum fusiforme (SF) is a popular edible brown macroalga found in Korea, Japan, and China and is known for its health-promoting properties. In this study, we used two sophisticated models to obtain optimized conditions for high antioxidant activity and metabolite profiling using high-resolution mass spectrometry. A four-factor central composite design was used to optimize the microwave-assisted extraction and achieve the maximum antioxidant activities of DPPH (Y1: 28.01 % inhibition), ABTS (Y2: 36.07 % inhibition), TPC (Y3: 43.65 mg GAE/g), and TFC (Y4: 17.67 mg CAE/g), which were achieved under the optimized extraction conditions of X1: 47.67 %, X2: 2.96 min, X3: 139.54 °C, and X4: 600.00 W. Moreover, over 79 secondary metabolites were tentatively identified, of which 12 compounds were reported for the first time in SF, including five phenolic (isopropyl 3-(3,4-dihydroxyphenyl)-2-hydroxypropanoate, 3,4-dihydroxyphenylglycol, scopoletin, caffeic acid 4-sulfate, and cinnamoyl glucose), two flavonoids (4’,7-dihydroxyisoflavone and naringenin), three phlorotannins (diphlorethohydroxycarmalol, dibenzodioxin-1,3,6,8-tetraol, and fucophlorethol), and two other compounds (dihydroxyphenylalanine and 5-hydroxybenzofuran-2(3H)-one) being identified for the first time in optimized SF extract. These compounds may also be involved in improving the antioxidant potential of the extract. Therefore, optimized models can provide better estimates and predictive capabilities that would assist in finding new bioactive compounds with improved biological activities that can be further applied at a commercial level.


Introduction
Phenolic compounds, which exhibit numerous biological and pharmacological activities, are among the most interesting in plant fractions [1]. Phenolic compounds from red, green, and brown seaweed have numerous health-promoting properties, including antioxidant, anticancer, antiviral, and anti-inflammatory activities [2]. Among brown seaweeds, Sargassum fusiforme (SF) belongs to the family Sargassaceae and is widely found in the coastal areas of Asia, especially in Korea, China, and Japan [3]. In Korea and Japan, like other marine organisms, SF is consumed as a sea vegetable because it possesses numerous nutritional components, such as proteins, minerals, phenolic compounds, dietary fibers, and polysaccharides [4,5]. The potential applications of these bioactive moieties in the pharmaceutical, food, and chemical industries have led to growing interest in the development and optimization of extraction techniques to isolate bioactive moieties from natural sources, as extraction is the most important step that plays a significant role in the quantity and quality of the result [6][7][8].
Recent advancements in extraction techniques have led to the development of efficient and new approaches for extracting natural bioactive components with improved yields and lower extraction times, temperatures, and solvent use. Numerous non-conventional techniques, such as ultrasound-assisted extraction (UAE) [9,10], pressurized liquid extraction (PLE) [11], supercritical fluid extraction (SFE) [12], and microwave-assisted extraction (MAE) [13], fulfill these requirements. In comparison with other techniques, MAE has its advantages, such as low cost and lower extraction time; thus, it can be categorized as "green technology" according to environmental policies [14,15]. The extraction time, temperature, technique, solid-to-liquid ratio, and solvent concentration are significant variables that, individually or in combination, alter the extraction efficiency. Optimizing the extraction conditions using a traditional methodology (one factor at a time) is a difficult, laborious, and time-consuming procedure. Empirical methods based on statistical or artificial intelligence methodologies have therefore been applied to overcome this problem [10]. RSM is a sophisticated statistical and mathematical approach used in the engineering field to optimize and improve conditions by analyzing the interactions between independent variables and response variables [16]. In RSM, a non-linear equation is used along with other designs, that is, the central composite design (CCD) or Box-Behnken design, to fit the experimental responses. Artificial neural networking (ANN) is a computational and mathematical approach that works based on the principles of the human brain. ANN, in comparison with RSM, is considered a superior and powerful tool because of its ability to adapt itself from the observations and predict the outcome after training on the experimental responses [17]. Moreover, optimized extraction conditions in combination with analytical tools such as high-resolution liquid chromatography-mass spectrometry (HR-LC-MS) have enabled the tentative identification of bioactive compounds in different species.
Previously, there was limited literature available that only focused on the optimization of polysaccharides of SF using RSM [18] without using ANN and green extraction technology, whereas in another trial, Li et al. [19] extracted and identified the phlorotannin fraction without using non-conventional extraction techniques and also without sophisticated statistical models. Therefore, in order to overcome the gaps, the present study was designed to optimize MAE-SF extracts using two sophisticated statistical tools (RSM and ANN) to obtain optimal conditions for exhibiting higher antioxidant activity and to determine the presence of bioactive compounds in SF using high-resolution tandem mass spectrometry (HR-LC-MS).

Chemicals and Reagents
Chemicals, reagents, and standards were procured from Sigma-Aldrich (St. Louis, MO, USA). Anhydrous sodium carbonate and acetic acid were acquired from Ducksan Pure Chemicals (Ansan, Republic of Korea), while sodium hydroxide and aluminum chloride hexahydrate were procured from Junsei Chemical (Tokyo, Japan) and Samchun Pure Chemical (Pyeongtaek, Republic of Korea), respectively. Acetic acid and anhydrous sodium carbonate were purchased from Ducksan Pure Chemicals (Ansan, Republic of Korea).

Sample Collection and Preparation
SF was collected from the coastal areas of Korea in mid-May 2021 and confirmed by a specialist at the Department of Oceanography, Kyungpook National University, Korea. Samples were washed, dried for four days at 37 • C, ground to obtain powder (approximately 800 g), and stored at −20 • C in the polyethylene bags for further analysis.

MAE Procedure
MAE was performed following the protocol of Kashif et al. [20] using a microwaveassisted extractor (MARS 6; CEM, Matthews, NC, USA) fitted with 50 mL quartz vessels with some minor modifications. Briefly, the SF sample was extracted three times by adjusting the microwave power (150-750 W) and time. Ten g of SF powder was placed in 500 mL Erlenmeyer flasks, and aqueous ethanol (150 mL) was added at different concentrations (25-75%) 30 min before extraction. Extraction was performed using CCD (Supplementary  Table S1). After extraction, flasks were allowed to cool at room temperature for 1 h and were filtered (Whatman No. 1; Schleicher & Schuell, Keene, NH, USA), followed by concentration to dryness in a rotary evaporator (Tokyo Rikakikai, Tokyo, Japan) at 40 • C and 50 rpm. In addition, samples were lyophilized using a freezer dryer and stored at −20 • C in Falcon tubes (15 mL) for further analysis.

Antioxidant Activities
Folin-Ciocalteu reagent was used to assess the total polyphenolic content (TPC) using gallic acid as a reference curve. It is a rapid and easy procedure based on the association between the detected color change and reactive reagent attenuation as a result of phenolic compounds present in the plant matrix. Total flavonoid content (TFC) was determined using the aluminum chloride colorimetric procedure [21]. Similarly, the radical scavenging activity of the SF extracts was assessed using 2,2-diphenyl-1-picrylhydrazyl (DPPH) and 2,2-azino-bis-3-ethylbenzothiazoline-6-sulfonic acid (ABTS) [22]. The RSM was implemented to assess the best possible combination of independent parameters to maximize the output of the dependent variables. The dependent variables DPPH, ABTS, TPC, and TFC were calculated in triplicate, and mean values were employed for regression analysis, as shown in Table 1. Experimental results were analyzed using Design Expert software to predict the optimal extraction conditions, and a second-order polynomial model Equation (1) was used to fit the model between independent and dependent variables. Various other tests, including the coefficient of variation (CV), adjusted coefficient of determination (Adj. R 2 ), lack of fit, and coefficient of determination (R 2 ), were employed to validate the RSM model [23].
where y represents the predicted variables; β 0 , β i , β ii , and β ij represent the regression coefficients of the intercept coefficient, linear, quadratic, and second-order terms, respectively; X i and X j are the independent variables (j and I range from 1-z), z is the number of factors (z = 4); ε j is the error. The Design Expert software package (11 STAT-EASE, Minneapolis, MN, USA) was used for graphical, regression, and experimental designs. ANOVA was used to determine the adequacy of the fitted model, the significance of the independent variables, their association, and the statistical significance of the regression coefficients.

ANN
ANN is a statistical machine learning technique that predicts nonlinear relationships between dependent variables and input parameters [24]. The model was constructed using the deep learning toolbox MATLAB R2020a (Mathworks, Minneapolis, MA, USA). The model was trained, tested, and validated using the experimental outputs obtained from the 27 experimental runs ( Table 1). The multilayer perceptron topology is comprised of three layers. The first (input) layer contains four independent parameters, and the middle layer comprises 16 neurons to achieve the optimal conditions to obtain the maximum output results. The third (output) layer contains four dependent variables, as shown in Figure 1A. The hit and trial technique was adopted to train the data and to lower the mean square error (MSE) calculated from Equation (2). The feed-forward and cascade feed-forward networks with the Broyden-Fletcher-Goldfarb-Shanno (BFGS) and Levenberg-Marquardt backpropagation (trainlm) algorithms were used [21,25]. The model exhibits a higher coefficient of determination (R 2 ) calculated from Equation (3), a lower standard error of prediction (SEP) from Equation (4), absolute average deviation (AAD) from Equation (5), and root mean squared error (RMSE) from Equation (6), which are considered optimal ANN models.
where Y Exp is the experimental data, Y ANN is the predicted value, and N is the number of samples.

Process Optimization Using RSM-GA, ANN-GA and RSM-DF
To optimize the MAE process, three different techniques were applied to the developed models, i.e., RSM-GA, RSM-DF (desirability functions), and ANN-GA. Data from the developed models (ANN and RSM) were subjected to GA (as initial input) using the optimization toolbox of MATLAB R2020a (Mathworks, Inc., Minneapolis, MA, USA). In GA, the RSM and ANN data were added as a fitness function to achieve the highest results for all the target responses. The initial population size, mutation fraction, evolutionary algebra, and crossover percentage were selected based on the current situation, while the remaining parameters were left at their default settings. The objective function was trained using the model data generated by the RSM-ANN, and the GA was implemented by maximizing the problems [26]. The RSM-DF was achieved by using the "Design Expert" software.

Intracellular Reactive Oxygen Species (ROS) Measurement
ROS generation as a result of cellular oxidative stress was assessed using the DCFH-DA method following the protocol of [28]. Briefly, RAW 264.7 cells were seeded for 24 h and then treated with samples at different concentrations. After 1 h, the cells were treated with 2, 2 -azobis (2-amidinopropane) dihydrochloride (AAPH) and incubated again for 30 min. With the help of PBS, cells were washed followed by treatment with DCFH-DA (25 µM), the plate was placed in the incubator at 37 • C for 30 min. The fluorescence intensity was determined at both the excited (485 nm) and emitted states (528 nm) using a Victor fluorescence microplate reader (PerkinElmer).

Western Blotting and Cell Lysate Preparation
Cell lysates were prepared following the standard methodology and then treated with 5X SDS-PAGE (3M Science, Seoul, Republic of Korea) sample buffer, and proteins were denatured by heating it at 95 • C for 10 min. Proteins were separated according to their molecular weight using 10% SDS gel electrophoresis, followed by the transfer of bands to nitrocellulose membranes for 2 h. A 5% bovine serum albumin (BSA) was added to the membrane and incubated overnight at 4 • C with the first antibody, followed by incubation with a secondary antibody (anti-goat IgG). Bands were detected using a chemiluminescence system (PerkinElmer) [29].

Compounds Identification Using ESI-MS/MS
The SF extract was analyzed to profile the secondary metabolites by following the protocol described by Choi et al. [21] using a Q-Executive Orbitrap mass spectrometer (Thermo Fisher Scientific, San Jose, CA, USA). Briefly, the SF sample was inserted at a rate of 15 µL/min using a 500 µL graduated syringe (Hamilton, Reno, NV, USA) and a syringe pump (Model 11, Harvard, Holliston, MA, USA) in the ESI source. The other parameters of ESI-MS (negative mode) were set as follows: the flow rates of auxiliary gas, sheath gas, and seep gas were set at 0, 5, and 0, respectively, spray voltage was set at 4.20 kV, capillary temperature was set at 320 • C, mass resolution of 140,000 (full width at half maximum, FWHM), S-lens at Rf level, and automatic gain control of 5 × 10 6 . Three different normalized collision energies (10, 30, and 40 eV) were applied stepwise using the same instrument [30].

Data Processing
The mass spectral data was analyzed with Xcalibur 3.1 with Foundation 3.1 (Thermo Fisher Scientific, Rockford, IL, USA). The detected m/z peaks were identified by comparing the exact masses of monoisotopic (negative mode) masses and ESI-MS/MS breakage patterns from an online database and as in-house MS/MS database. The online databases included FooDB [31], METLIN [32], and CFM-ID 4.0 [33].

Statistical Analysis
The final results were analyzed using MATLAB and Design Expert 11 software. All data is presented as the mean ± standard deviation of at least three independent experiments (n = 3), where each experiment had a minimum of three replicates for each sample. Differences were considered significant at p < 0.001, p < 0.01, and p < 0.05. Table 1 shows a comparison between dependent response variables (DPPH, ABTS, TPC, and TFC) and predicted outcomes of RSM and ANN for all 27 runs obtained after MAE according to CCD as shown in Figure 1C. The DPPH, ABTS, TPC, and TFC contents of SF extracts varied in the ranges of 7.31-32.74% inhibition, 25.36-42.61% inhibition, 20.48-36.08 mg GAE/g, and 6.94-18.95 mg CAE/g, respectively. The ANOVA results are depicted in Table 2. For a well-fitted model, R 2 values should be close to 1, and, in our trial, all four variables, that is, DPPH (0.95), ABTS (0.96), TPC (0.95), and TFC (0.97), exhibited higher R 2 values, which shows that our model fits the second-order polynomial equation well. Similarly, other factors, including the CV, regression coefficient (β), F-value, adjusted correlation factor (R 2 ), adequate precision, and lack of fit (probability), also contributed to the significance of the model. In the current study, lack of fit (0.09-0.35), F-values (2.18-9.58), Adj R 2 (0.89-0.94), adequate precision (12.92-19.08), and CV (4.81-8.90) indicated that the fitness of the RSM model outcomes are in line with Simic et al. [6], as they checked the total polyphenolic compounds in the chokeberries using RSM and ANN. Likewise, Ameer et al. [26] also checked the parameters, including (R 2 , CV, adeq. Precision) during the recovery of stevioside and rebaudioside-A from Stevia rebaudiana leaves using MAE, which also supports our results. High F-values (p < 0.05) indicate that the lack of fit is non-significant, and our RSM model had higher F-values, showing that it was well-fitted, and these results are similar to Choi et al. [21]. The level of significance is expressed as *** p < 0.001, ** p < 0.01, and * p < 0.05.

Effect of MAE-SF Parameters on DPPH and ABTS
To better understand the association between the independent and response variables, contour plots were produced, as shown in Figure S1. As shown in Table 2, linear terms of X 3 exerted negative effects on ABTS and DPPH. In quadratic terms, X 2 2 and X 3 2 hurt DPPH, whereas, for ABTS, all four quadratic terms showed significant negative effects. In terms of interactions, X 1 X 2 , X 2 X 4 , and X 3 X 4 had negative effects on both ABTS and DPPH. The maximum activity (32.74 mg GAE/g) for DPPH was calculated at run 25 (X 1 : 30%, X 2 : 4 min, X 3 : 110 • C, X 4 : 600 W), whereas the ABTS highest activity (42.61 mg CAE/g) was recorded at run 11 (X 1 : 50%, X 2 : 3 min, X 3 : 130 • C, X 4 : 750 W). The lack of fit for DPPH (0.1138) and ABTS (0.0981) and the coefficients of regression of 0.9543 and 0.9608 for DPPH and ABTS, respectively, suggest that the RSM model fits well. The fitted second-order polynomial equations for DPPH (% inhibition) and ABTS (% inhibition) are shown in Equations (7) and (8) as follows:

Effect of MAE-SF Parameters on TPC and TFC
Similar to DPPH and ABTS, contour plots were generated for TPC and TFC to determine the relationship between the independent variables by varying the conditions. TPC (36.06 mg GAE/g) and TFC (16.85) exhibited the highest activities in runs 16 and 12, respectively (Table 1). ANOVA results show that TPC had significant effects in terms of linear (X 1 , X 3 , X 4 ), interaction (X 1 X 3 , X 2 X 3 , X 2 X 4 ), and quadratic relationships (X 1 2 , X 2 2 , X 3 2 , X 4 2 ). The TFC exhibited significant effects in terms of linear (X 1 , X 2 , X 3 , X 4 ), interaction (X 1 X 2 , X 1 X 3 , X 2 X 3 , X 2 X 4 , X 3 X 4 ), and quadratic relationships (X 2 2 , X 3 2 ) ( Table 2). Likewise, the lack of fit for TPC (0.1025) and TFC (0.3556), along with the coefficient of determination for TPC (0.9533) and TFC (0.9745), show that our model could be fitted well. The relationship between TPC, TFC independent, and response variables can be determined using Equations (9) and (10), as follows:

ANN Modeling
The ANN architecture topology for MAE-SF conditions is shown in Figure 1. For a well-developed model, the testing and training errors should be minimal, and the epochs should be kept optimal. Lower epoch levels result in an under-fitted model, whereas high epochs result in model over-fitting. In this study, an ANN topology of 4-16-4 was found to be sufficient, with the least MSE and a higher R 2 for better model reliability and precision, as shown in Figure S2. Figure 1B Figure 1F-I shows a comparison between the experimental response and both the RSM and ANN predictions. Our outcomes are inconsistent with the results reported by Ameer et al. [26] for the recovery of bioactive components from Stevia rebaudiana leaf extracts using ultrasoundassisted extraction conditions with a hybrid RSM-ANN-GA model. Table 3 presents the model validation parameters. For an optimal model, the values of RMSE, SEP, and AAD should be lower, whereas R 2 should be higher. In our study, for all four response variables, SEP, RMSE, and AAD values were lower, whereas R 2 was higher in the ANN model than in the RSM model, which validated the superiority of the ANN model.

Process Optimization and Model Validation
The data generated by the RSM and ANN models was used to optimize the extraction conditions by using a genetic algorithm. A scalar technique was used in the objective function instead of using second-or first-order derivatives [34]. GA is adopted by maximizing the outputs of RSM and ANN in the objective function. For the fitness function, the rank method was used, whereas a stochastic uniform was utilized for the selection method. In GA with constraint dependence, 90 population sizes were initially taken, while for the survival of the next generation, the elite count in the reproduction option was put on default. Moreover, for genetic variation and also to achieve new individual formation, constraint dependent mutation and crossover functions were applied (Table S3). A Lagrangian solver was adopted for better accuracy as the nonlinear constraint algorithm, whereas Design Expert software was adopted to achieve the optimal condition by keeping the independent variables in their maximum range. Table 4 Table S2.

Cell Viability, ROS Scavenging, and Antioxidant Enzyme Activity
SF extracts obtained after optimizing the MAE conditions did not exhibit any RAW 264.7 cellular toxicity up to 300 µg/mL (Figure 2A). Figure 2B shows that ROS production was enhanced by LPS induction; however, SF extracts scavenged ROS in a dose-dependent manner (approximately two-fold). Similarly, SF extracts also increased antioxidant enzyme activity, that is, catalase (CAT), glutathione peroxidase-1 (GPx-1), and superoxidase-1 (SOD-1) levels in a concentration-dependent pattern ( Figure 2C). Wang et al. (2022) stated that fermented fucoidan isolated from SF attenuates ROS levels and improves SOD levels in zebrafish induced by H 2 O 2 [35], which is consistent with our results.

Metabolite Profiling of SF Extracts Using High-Resolution LC-MS/MS
Metabolite profiling of the optimized SF extract (Supplementary Table S2) was conducted in the negative mode of HRLC-ESI-MS/MS. As shown in Table 5, 79 secondary metabolites were tentatively verified, and 12 of them were reported for the first time in SF, which were as follows.

Phenolic Acids
During collisions, phenolic acids exhibit certain breakage patterns and loss-specific molecules, such as methyl (15 Da) and carboxyl (44 Da) [36]. Based on the fragmentation pattern, compounds 1-2 were previously identified as gallic acid and vanillic acid, respectively [21]. Compound 3, reported in Dasya sp., yielded two daughter ions at m/z 179.03 (isopropanol group) and m/z 123.04 by cleaving the α, β carbon-carbon bond, and was identified as isopropyl 3-(3,4-dihydroxyphenyl)-2-hydroxypropanoate [37]. Compound 4 (3,4-dihydroxyphenylglycol) yielded a precursor ion at m/z 169.0506, which was further fragmented into two major daughter ions at m/z 151.03 (losing H 2 O) and 123.00 (losing CO). Similarly, compound 5 was previously reported in Codium and Grateloupia sp., and produced a parent ion at m/z 191.0352 and was fragmented further by losing methyl , followed by further cleavage into a caffeic acid ion (135.04 Da) by losing SO 3 (80 Da). Fragmentation of phenolic acid glycoside started at the glycosidic link and yielded a sugar moiety (−162 Da), which tentatively confirmed the presence of compound 7 (cinnamoyl glucose) in SF by having a characteristic ion at m/z 303.08 [38]. Compounds 3-7 have been reported for the first time in SF extracts.

Flavonoids
Compound 8, previously known as 3 -O-methylcatechin, produced an ion peak at m/z 303.0868 with characteristic peaks at m/z 271.07 and m/z 163.05, corresponding to a loss of 30 Da (CH 3 O) and 109 Da (C 6 H 5 O 2 ), respectively, identified for the first time in this species. Compound 9, previously known as 4',7-dihydroxyisoflavone, had a precursor ion peak at m/z 253.0500 and fragment peaks at m/z 225.05 and m/z 197.06, owing to the loss of two CO molecules [M-H-2CO], followed by the removal of water molecules from the B ring at m/z 143.03. Similarly, compound 10 (naringenin) was previously found in Ulva intestinalis and tentatively identified by Choi et al. [21]. Naringenin exhibits various biological properties, including anti-inflammatory and anticancer activities [39]. Compounds 9 and 10 were also reported for the first time in SF-optimized extracts.

Carboxylic Acid, Fatty Acids, Sugars, and Amino Acids
Compounds 21-35 were previously reported in a variety of seaweeds and were identified as carboxylic acids based on their fragmentation patterns. Compounds 36-66 were fatty acids, compounds 67-69 were amino acids, and compounds 70-74 were sugars [30,[46][47][48][49][50][51], as shown in Table 5. Additionally, compounds 75 and 76 (phenols), and compound 77 (alcohol) were identified. Compound 78 was previously identified in Ascophyllum nodosum as dihydroxyphenylalanine, based on its fragmentation pattern. Dihydroxyphenylalanine is a large neutral amino acid with a catecholamine precursor that has a neuroprotective effect, especially against Parkinson's disease [52]. Compound 79 (5-hydroxybenzofuran-2(3H)-one) yielded a monoisotopic mass at m/z 149.0238, which represented an aromatic alcohol group. To the best of our knowledge, compounds 78 and 79 have been tentatively reported for the first time for SF.

Conclusions
In this study, the effects of MAE parameters on the antioxidant potential of Sargassum fusiforme were investigated using two sophisticated techniques: RSM and ANN-GA. A comparative overview of both statistical models based on SEP, RMSE, R 2 , and AAD revealed that the ANN-GA models were superior to the RSM model. The optimized MAE conditions (X 1 : 47.67%, X 2 : 2.96 min, X 3 : 139.54 • C, and X 4 : 600.00 W) exhibited maximum response for four dependent variables: Y 1 :28.01% DPPH inhibition; Y 2 :36.07% ABTS inhibition; Y 3 :43.65 mg GAE/g TPC; Y 4 :17.67 mg CAE/g TFC content. The optimized extracts showed no toxicity to RAW 264.7 cells up-to 300 µg/mL and had increased CAT, GPx-1, and SOD-1 levels in a dose-dependent manner. After optimizing the MAE conditions, optimized extracts were analyzed by HR-LC-MS and identified 79 secondary metabolites; among them, 12 new bioactive compounds (five phenolic compounds, two flavonoids, three tannins, and two other compounds) were also tentatively reported for the first time in optimized SF extracts. The current investigation may not only provide an alternative statistical technique, but it could also support the preferred extraction technology for identifying important bioactive components that can be used in broad commercial applications as promising ingredients for the development of functional foods and nutraceuticals and in the cosmetic industry, shedding light on the bright side of human health. However, several other aspects that might affect the extractability of bioactive components, including kinetic behavior, type of solvent, pH, ultrasound frequency, and quantitative measurement of bioactive components, should be investigated, as this study provides the basis for further future trials.
Supplementary Materials: The following supporting information can be downloaded at: https: //www.mdpi.com/article/10.3390/antiox11112246/s1. Table S1: Independent process variables with experimental ranges and levels for MAE of SF, Table S2: Comparison between optimized model and experimental values, Table S3. Setting parameters of genetic algorithm used in the optimization of process for SF, Figure S1: The three-dimensional (3D) response surface plots of MAE-SF extraction condition displaying the influence of independent parameters (Ethanol Concentration, Time, Temperature, and power) on dependent variables (DPPH radical-scavenging activity, ABTS, TPC, and TFC) as a function of significant interaction factors for RSM, Figure S2: Regression of experimental and predicted values in ANN model of MAE for SF using the training, testing and validation datasets to optimized the extraction conditions.