Highly Efficient Synthesis of an Emerging Lipophilic Antioxidant: 2-Ethylhexyl Ferulate

Ferulic acid in ester form has shown a stronger ability in ameliorating certain pathological conditions and inhibiting lipid oxidation. In present study, a solvent-free and reduced pressure evaporation system was developed for lipase-catalyzed synthesis of 2-ethylhexyl ferulate (2-EF) from ferulic acid and 2-ethylhexanol. A Box-Behnken design with response surface methodology (RSM) and artificial neural network (ANN) was selected to model and optimize the process. Based on the yields of 2-EF, reaction temperature was shown to be the most important process factor on the molar conversion among all variables. The residual values and the coefficient of determination (R2) calculated from the design data indicated that ANN was better than RSM in data fitting. Overall, the present lipase-catalyzed approach for 2-EF synthesis at low reaction temperature in a reduced pressure evaporation system shows high 2-EF production efficiency. Notably, this approach can reduce the enzyme denaturation and ferulic acid oxidation that usually occur during long-term biosynthetic operations at high temperature.


Introduction
Ferulic acid (FA) belongs to the family of phenolic acids and is very abundant in grains, fruits and vegetables. Studies have shown that FA displays excellent antioxidant [1], ultraviolet-absorbing activities [2,3] and potential health benefits against cardiovascular problems, inflammatory diseases and cancer [4]. However, the solubility of FA in hydrophobic solvents is low, thus limiting its utilization in the pharmaceutical, cosmetic and food industries. In order to increase utilization of FA, lipophilization is needed, where enzymatic esterification with alcohols has proved to be a promising approach [5][6][7][8]. In recent years, many studies have report that the enzymatic synthesis of hydrophobic derivatives of FA can increase its oil-solubility [9][10][11], and some of them have been applied for preparation of emulsion-based creams due to their anti-inflammatory activity [12]. FA esters, the hydrophobic derivatives of FA, have also demonstrated other bioactivities such as antioxidant [13], anticonvulsant [14], synergistic photobactericidal effects [15], and antithrombotic activities [16], as well as inhibition of the atherogenic effects of leptin [17]. Interestingly, FA esters possess better inhibitory effect on lipid oxidation than FA [18].
Enzymatic biosynthesis of these compounds has attracted much attention in recent years, and it may be a better approach in comparison to conventional chemical processes [19][20][21][22][23]. It has been

Results and Discussion
The present study applied RSM and ANN for experimental design of enzymatic synthesis of 2-EF in a solvent-free system. The molar ratio of FA and 2-ethylhexanol (the reactants) was 1:129, and the independent parameters selected for the experimental design were: reaction time (X 1 ), enzyme amount (X 2 ), and reaction temperature (X 3 ). Their ranges and levels are given in Table 1.

RSM Model
The RSM Box-Behnken procedure was employed to fit the second-order polynomial equation to the experimental data (Table 1). Among the various treatments, the highest molar conversion (98.9%˘2.5%) was observed in treatment 12 (reaction temperature of 80˝C, enzyme amount of 1500 PLU and reaction time of 16 h), and the lowest molar conversion (42.7%˘3.3%) was observed in treatment 10 (reaction temperature of 60˝C, enzyme amount of 1000 PLU and reaction time of 8 h). The second-order polynomial equation obtained was as follows: Y "´968.26667`3.60052X 1`2 6.28208X 2`0 .039783X 3´0 .027812X 1 X 2`0 .0003X 1 The analysis of variance (ANOVA), represented in Table 2, indicates that this second-order polynomial model was highly significant and adequate to represent the actual relationship between the response (molar conversion) and the significant variables with a very small p value (<0.001) and a satisfactory coefficient of determination (R 2 = 0.9852). The ANOVA results of responses reveal an insignificant "lack of fit" for p > 0.05. Therefore, these models were adequate for prediction within the range of variables employed. Furthermore, the ANOVA results indicate that the linear term of X 1 and X 3 and the square term of X 2 2 and X 3 2 had significant (p < 0.05) influence on molar conversion.
The interaction terms had less influence (p > 0.05).  The effect of varying reaction temperature and enzyme amount on transesterification efficiency at a constant reaction time of 16 h was shown in Figure 1A. With the lowest reaction temperature (60˝C) and enzyme amount (500 PLU), a 2-EF molar conversion of 39.8% was obtained. Low molar conversions are also found in a previous report on lipase-catalyzed synthesis of ferulate esters at the temperature of 60˝C, enzyme amount of 100 mg, and time of 240 h [22]. The low molar conversion of ferulate esters can be improved by using a reduced pressure evaporation system. In the reduced pressure evaporation system, when the enzyme amount was increased to 1500 PLU and the reaction temperature was increased to 80˝C, the molar conversion of 2-EF was increased to 98.9 ( Figure 1A). The great increase in molar conversion can result from the impact of solvent-free and reduced pressure evaporation system which increases the reaction rate of lipase-catalyzed biosynthesis by elevating the substrate concentration and eliminating byproducts with low vapor pressure physical characteristics. Figure 1B shows the effects of reaction time and reaction temperature on 2-EF biosynthesis at a constant enzyme amount (1000 PLU). At the lowest reaction time (8 h) and reaction temperature (60˝C), the molar conversion was 42.7%˘3.3%. Figure 1C shows the effects of reaction time and enzyme amount on 2-EF biosynthesis at a constant reaction temperature. In this regard, enzyme amount had an insignificant effect on molar conversion, while reaction time had only a partial effect on molar conversion. Taken together, the molar conversion significantly increased when the reaction temperature and reaction time increased. As the ANOVA results shown in Table 2 indicate, reaction temperature and reaction time are important factors in Novozym ® 435-catalyzed reactions, which is in accordance with our previous works [27]. However, it has been reported that higher reaction temperature might decrease the lipase activity [21]. Thus, a reduced pressure evaporation system can be a good solution for lipase-catalyzed synthesis of FA esters in order to obtain a higher conversion at a lower reaction temperature.
Molecules 2016, 21, 478 4 of 10 molar conversion was 42.7% ± 3.3%. Figure 1C shows the effects of reaction time and enzyme amount on 2-EF biosynthesis at a constant reaction temperature. In this regard, enzyme amount had an insignificant effect on molar conversion, while reaction time had only a partial effect on molar conversion. Taken together, the molar conversion significantly increased when the reaction temperature and reaction time increased. As the ANOVA results shown in Table 2 indicate, reaction temperature and reaction time are important factors in Novozym ® 435-catalyzed reactions, which is in accordance with our previous works [27]. However, it has been reported that higher reaction temperature might decrease the lipase activity [21]. Thus, a reduced pressure evaporation system can be a good solution for lipase-catalyzed synthesis of FA esters in order to obtain a higher conversion at a lower reaction temperature.

ANN Model
ANN is a promising modeling methodology for non-linear multivariate process/reaction [28]. In this study, the data generated from the Box-Behnken design (Table 1) were employed to generate an ANN model for the enzymatic 2-EF synthesis. The proposed ANN consisted of three layers in the present work: an input layer with four neurons (reaction time, reaction temperature, and enzyme amount), a hidden layer with several neurons and an output layer containing one output neuron (molar conversion). The feed-forward neural network was trained by using incremental back propagation (IBP), batch backpropagation (BBP) and Levenberg-Marquardt algorithm (LM) or genetic algorithm (GA), respectively. All neurons from the hidden layer and output layer were calculated by a sigmoid transfer function. Different numbers of neurons, from three to six, were examined. The various architectures for ANN models were developed using NeuralPower software (Version 2.5, CPC-X Software, USA; available from Internet: http://www.brothersoft.com/neuralpower-download-21356.html), and the best one was selected based on the maximization of the R 2 value and the minimization of the RMSE value.
All results from the test were evaluated statistically based on the RMSE values. According to Figure 2, the best performance network was trained using a LM algorithm with a sigmoid transfer

ANN Model
ANN is a promising modeling methodology for non-linear multivariate process/reaction [28]. In this study, the data generated from the Box-Behnken design (Table 1) were employed to generate an ANN model for the enzymatic 2-EF synthesis. The proposed ANN consisted of three layers in the present work: an input layer with four neurons (reaction time, reaction temperature, and enzyme amount), a hidden layer with several neurons and an output layer containing one output neuron (molar conversion). The feed-forward neural network was trained by using incremental back propagation (IBP), batch backpropagation (BBP) and Levenberg-Marquardt algorithm (LM) or genetic algorithm (GA), respectively. All neurons from the hidden layer and output layer were calculated by a sigmoid transfer function. Different numbers of neurons, from three to six, were examined. The various architectures for ANN models were developed using NeuralPower software (Version 2.5, CPC-X Software, USA; available from Internet: http://www.brothersoft.com/ neuralpower-download-21356.html), and the best one was selected based on the maximization of the R 2 value and the minimization of the RMSE value.
All results from the test were evaluated statistically based on the RMSE values. According to Figure 2, the best performance network was trained using a LM algorithm with a sigmoid transfer function, which exhibited a lower and stable RMSE. It is noted that the curve of the RMSE levels off at five hidden neurons. Therefore, the best ANN model in this study is determined to be a multilayer feed forward connection, trained by a LM algorithm using a sigmoid transfer function that consists of a 3-6-1 topology, as shown in Figure 3. This learning was acquired with RMSE = 0.79, R 2 = 0.99. at five hidden neurons. Therefore, the best ANN model in this study is determined to be a multilayer feed forward connection, trained by a LM algorithm using a sigmoid transfer function that consists of a 3-6-1 topology, as shown in Figure 3. This learning was acquired with RMSE = 0.79, R 2 = 0.99.  According to Figure 4, the reaction time and reaction temperature displayed great impact on the conversion with a relative importance of approximate 39% and 46%, respectively. These results are in accordance with the ones obtained with RSM. at five hidden neurons. Therefore, the best ANN model in this study is determined to be a multilayer feed forward connection, trained by a LM algorithm using a sigmoid transfer function that consists of a 3-6-1 topology, as shown in Figure 3. This learning was acquired with RMSE = 0.79, R 2 = 0.99.  According to Figure 4, the reaction time and reaction temperature displayed great impact on the conversion with a relative importance of approximate 39% and 46%, respectively. These results are in accordance with the ones obtained with RSM. According to Figure 4, the reaction time and reaction temperature displayed great impact on the conversion with a relative importance of approximate 39% and 46%, respectively. These results are in accordance with the ones obtained with RSM. According to Figure 4, the reaction time and reaction temperature displayed great impact on the conversion with a relative importance of approximate 39% and 46%, respectively. These results are in accordance with the ones obtained with RSM.

Comparison between RSM and ANN and Optimization for Experimental Condition
The distribution of the residual values obtained from the experimental data of Box-Behnken design experiments minus the related computed RSM and ANN data are shown in Figure 5. It was observed that the deviations of residual values showed a relatively small variation in the ANN model when compared with the RSM model. This suggested that ANN was better than RSM in data fitting.

Comparison between RSM and ANN and Optimization for Experimental Condition
The distribution of the residual values obtained from the experimental data of Box-Behnken design experiments minus the related computed RSM and ANN data are shown in Figure 5. It was observed that the deviations of residual values showed a relatively small variation in the ANN model when compared with the RSM model. This suggested that ANN was better than RSM in data fitting. In order to validate and test the extrapolative capability of both the ANN and RSM models, a completely new set of five experiments was conducted on an experimental range which does not belong to the design data sets. The experimental and predicted values of the response for both the ANN and RSM models are given in Table 3. The performance of the newly constructed ANN and RSM models were statistically measured, in terms of AAD and RMSE values. The RMSEs for RSM and ANN were found as 7.322 and 5.152. These results indicate that the RSM prediction has a greater deviation than the ANN prediction. The AAD for RSM and ANN was found to be 10.046 and 9.430. From the results, it was observed that predictive power of ANN was found to be more powerful than RSM.  In order to validate and test the extrapolative capability of both the ANN and RSM models, a completely new set of five experiments was conducted on an experimental range which does not belong to the design data sets. The experimental and predicted values of the response for both the ANN and RSM models are given in Table 3. The performance of the newly constructed ANN and RSM models were statistically measured, in terms of AAD and RMSE values. The RMSEs for RSM and ANN were found as 7.322 and 5.152. These results indicate that the RSM prediction has a greater deviation than the ANN prediction. The AAD for RSM and ANN was found to be 10.046 and 9.430. From the results, it was observed that predictive power of ANN was found to be more powerful than RSM. Ridge max analysis can be used to determine optimum operating conditions by computing the estimated ridge of the maximum response for increasing the radius from the center of the original design. As shown in Table 4, the optimum condition for lipase-catalyzed synthesis of 2-EF determined by the ridge max analysis of RSM was reaction time of 23 h, reaction temperature of 71˝C and enzyme amount of 1422 PLU. A verification experiment performed at this condition obtained a molar conversion of 99.43%. Notably, higher molar conversion (99.74) was observed when the experiment was performed using the ANN-modeled optimal conditions (Table 4).

Evaluation of Enzyme Reusability
To examine the enzyme reusability, the ability of reused immobilized lipase for 2-EF synthesis was investigated under optimum conditions ( Figure 6). After reaction, the mixture was filtered through a 0.45 µm membrane. The immobilized lipase was then recovered from the membrane, washed three times with fresh reaction media and dried before reuse in the next batch. After the immobilized lipase was reused five times, the molar conversion of 2-EF still remained at about 90.4%. This result showed that the immobilized lipase remained stable through long-term 2-EF exposure. Thus, the result demonstrated that the lipase could be effectively applied for 2-EF synthesis in the reduced pressure evaporation system and that the stability was high enough to permit reuse.  (Table 4). Table 4. The optimal trial obtained from RSM and ANN models.

Evaluation of Enzyme Reusability
To examine the enzyme reusability, the ability of reused immobilized lipase for 2-EF synthesis was investigated under optimum conditions ( Figure 6). After reaction, the mixture was filtered through a 0.45 μm membrane. The immobilized lipase was then recovered from the membrane, washed three times with fresh reaction media and dried before reuse in the next batch. After the immobilized lipase was reused five times, the molar conversion of 2-EF still remained at about 90.4%. This result showed that the immobilized lipase remained stable through long-term 2-EF exposure. Thus, the result demonstrated that the lipase could be effectively applied for 2-EF synthesis in the reduced pressure evaporation system and that the stability was high enough to permit reuse.

Materials
Immobilized lipase Novozym ® 435 (10,000 U/g; propyl laurate units (PLU)) from Candida antarctica B (EC3. 1 After incubation, liquid samples were withdrawn from the reaction mixture for analysis of 2-EF by high-performance liquid chromatography (HPLC). The sample was diluted with hexanol/ 2-methyl-2-butanol (1:150), and was injected (30 μL) into a HPLC (Hitachi L-7400; Tokyo, Japan) equipped with UV detector and Thermo C18 column (250 mm × 4.6 mm, Agilent, Santa Clara, CA, USA). Separations were carried out by a gradient elution with 0.1% acetic acid and methanol, the flow rate was set to 1.0 mL/min, and 2-EF was detected under UV light at 325 nm as shown in Figure 7. Calibration curves were prepared from the FA and 2-EF standards based on the peak areas. The molar conversion was defined as (mmol of 2-EF production per mmol of initial FA) × 100%.

Box-Behnken Design
A Box-Behnken design for 15 experimental runs was employed in present study. To avoid bias, the 15 runs were performed in a totally random order. The variables and their levels selected for the study of 2-EF biosynthesis were: reaction temperature (60-80 °C), enzyme amount (500-1500 PLU) and reaction time (8-24 h), which were coded as shown in Table 5. Table 1 shows the independent After incubation, liquid samples were withdrawn from the reaction mixture for analysis of 2-EF by high-performance liquid chromatography (HPLC). The sample was diluted with hexanol/ 2-methyl-2-butanol (1:150), and was injected (30 µL) into a HPLC (Hitachi L-7400; Tokyo, Japan) equipped with UV detector and Thermo C18 column (250 mmˆ4.6 mm, Agilent, Santa Clara, CA, USA). Separations were carried out by a gradient elution with 0.1% acetic acid and methanol, the flow rate was set to 1.0 mL/min, and 2-EF was detected under UV light at 325 nm as shown in Figure 7. Calibration curves were prepared from the FA and 2-EF standards based on the peak areas. The molar conversion was defined as (mmol of 2-EF production per mmol of initial FA)ˆ100%. After incubation, liquid samples were withdrawn from the reaction mixture for analysis of 2-EF by high-performance liquid chromatography (HPLC). The sample was diluted with hexanol/ 2-methyl-2-butanol (1:150), and was injected (30 μL) into a HPLC (Hitachi L-7400; Tokyo, Japan) equipped with UV detector and Thermo C18 column (250 mm × 4.6 mm, Agilent, Santa Clara, CA, USA). Separations were carried out by a gradient elution with 0.1% acetic acid and methanol, the flow rate was set to 1.0 mL/min, and 2-EF was detected under UV light at 325 nm as shown in Figure 7. Calibration curves were prepared from the FA and 2-EF standards based on the peak areas. The molar conversion was defined as (mmol of 2-EF production per mmol of initial FA) × 100%.

Box-Behnken Design
A Box-Behnken design for 15 experimental runs was employed in present study. To avoid bias, the 15 runs were performed in a totally random order. The variables and their levels selected for the study of 2-EF biosynthesis were: reaction temperature (60-80 °C), enzyme amount (500-1500 PLU) and reaction time (8-24 h), which were coded as shown in Table 5. Table 1 shows the independent (1) (2)

Box-Behnken Design
A Box-Behnken design for 15 experimental runs was employed in present study. To avoid bias, the 15 runs were performed in a totally random order. The variables and their levels selected for the study of 2-EF biosynthesis were: reaction temperature (60˝C-80˝C), enzyme amount (500-1500 PLU) and reaction time (8-24 h), which were coded as shown in Table 5. Table 1 shows the independent factors (xi), levels and experimental design. Each experimental point was carried out in duplicate. Statistical parameters mentioned in present study including coefficient of determination (R 2 ), root-mean-square error (RMSE), and absolute average deviation (AAD) were calculated by following Equations (1) to (3): where Y pre is the predicted 2-EF yield (by either RSM or ANN), Y exp is the observed 2-EF yield and Y m is the average 2-EF yield.
where Y pre and Y exp are the predicted and experimental response, respectively, and n is the number of experiments.
AAD " « ř n i"1`ˇY exp´Ypreˇ{ Y pren ffˆ1 00 (3) where Y pre and Y exp are the predicted and experimental response, respectively, and n is the number of experiments.

Conclusions
The present study demonstrates for the first time that immobilized C. antarctica lipase (Novozym ® 435) is able to synthesize 2-ethylhexyl ferulate (2-EF). A Box-Behnken design with RSM and ANN was selected to model and optimize the biocatalysis. Both RSM and ANN models for the 2-EF synthesis were built, and ANN was better than RSM in data fitting and prediction. The optimum synthesis condition modeled by ANN was in a reaction time of 24 h, reaction temperature of 78˝C and enzyme amount of 1485 PLU. The conversion of 2-EF under this optimum condition was verified to be 99.74%. The reduced pressure evaporation system combined with solvent free reaction conditions can decrease the reaction time and reaction temperature for ferulic acid ester synthesis.