Lipase-catalyzed synthesis of dilauryl azelate ester: process optimization by artificial neural networks and reusability study

Abstract

An application of artificial neural networks (ANNs) to predict the performance of a lipase-catalyzed synthesis for esterification of dilauryl azelate ester was carried out. The central composite rotatable design (CCRD) experimental data were utilized for training and testing of the proposed ANN model. The model was applied to predict various performance parameters of the enzymatic reaction conditions, namely enzyme amount (0.05–0.45 g), reaction time (90–450 min), reaction temperature (40–64 °C) and molar ratio of substrates (AzA : LA, 1 : 3–1 : 9 mol). The incremental back propagation (IBP), batch back propagation (BBP), quick propagation (QP), genetic algorithm (GA), and the Levenberg–Marguardt (LM) algorithms were used in the network. It was found that the optimal algorithm and topology were the incremental back propagation (IBP) and the configuration with 4 inputs, 14 hidden, and 1 output nodes, respectively.

---

Introduction

Azelaic acid is a naturally occurring saturated dicarboxylic acid which, on pharmacological application, has been shown to be effective in the treatment of comedonal acnes, and inflammatory (papulopustular, nodulocystic and nodular) acne, as well as various cutaneous hyperpigmentary disorders.¹ But its usage is limited in cosmetics and pharmaceutical products because of insolubility, high melting point and large dosage requirement. In addition, due to its crystalline form which melts at high temperature (103 °C), incorporation of azelaic acid into product formulation is difficult at standard conditions. However, efficacy was only observed when amount ≥15% of azelaic acid was incorporated.^2,3 Different derivatives of azelaic acid have also been proven to be an effective topical treatment for human inflammatory skin disorders.⁴

Tamarkin and Maccabim (2004) reported that modification of azelaic acid by attachment of different steroidal hormones through covalent ester linkages and converting at least, one of the carboxylic acid groups into ester group has enhanced not only the technical property of azelaic acid, but also the functional performance of the resulting modified azelaic acid. Their studies showed that the modified azelaic acid exhibited anti-acne, sebum regulatory and skin lightening properties on the treated skin at low dosage.⁵ The antibacterial properties of some azelaic acid diesters were investigated by Charnock and collaborators.⁶ Hsieh et al. also introduced a co-drug of conjugated hydroquinone and azelaic acid to enhance topical skin targeting and decrease penetration through the skin. Their studies demonstrated that esterification of azelaic acid by hydroquinone contains two hydroxyl functionalities was allowed to formation an ester co-drug.⁷

In this work, the synthesis of dilauryl azelate ester was studied using immobilized lipase (Novozym 435) as biocatalyst has not been reported. Therefore, the optimization of process parameters was carried out based on the investigations relating to the influence of enzyme amount, reaction time, reaction temperature, and substrates molar ratio by using artificial neural network. In fact, this works aims to locate the optimum reaction conditions for lipase catalyzed synthesis of dilauryl azelate ester through CCRD as an experimental design and ANN as a statistical tool for optimization.

An artificial neural network (ANN) inspired by biological neural networks process input information using an interconnected group of artificial neurons and a connectionist approach to computation.⁸ This work is concerned with ANNs that have proven to be a powerful tool successfully used for several real world applications over the past few years.^9–14 Artificial neural networks have been used for representing non-linear functional relationships between variables. The ability of an ANN to learn and generalize the behavior of any complex and non-linear process makes it a powerful modeling tool.¹⁵ Solving and modeling the complex relation between input and output variable can be simply performed by an ANN model imitated by biological neuron processing. ANNs have been used to study a wide variety of chemical problems.¹⁶

Methodology

Materials

Novozym 435 (lipase B from Candida Antarctica) was obtained from Novo Nordisk A/S (Denmark). Azelaic acid and lauryl alcohol were purchased from Merck (Germany). n-Hexane obtained from J.T.Baker (USA) was used as the organic solvent. All other chemicals used in this study were of analytical reagent grade.

Enzymatic esterification of azelaic acid and analysis of samples

To optimize the reaction conditions, enzymatic reactions were usually performed in screw-capped glass vials on a shaker. In a typical reaction, a certain amount of azelaic acid and proportional amount of lauryl alcohol were added into 5 ml solvent (n-hexane) contained in screw-capped glass vial. The enzyme amount was added to the mixture was sealed and shaken at 150 rpm in a horizontal water bath shaker.

The enzymatic reaction was terminated by adding 5 ml of ethanol–acetone (50 : 50, v/v). The catalyst was then separated by filtration, and the residual free acid into the final mixture sample was measured by standard NaOH (0.1 N) titration. The value of reacted acid was calculated from the amounts obtained for the test (with enzyme) and control (without enzyme) samples. The amount of ester formation was expressed as equivalent to the acid conversion.^17,18 The conversion percentage was calculated from consumption of NaOH for test and control. The conversion was calculated using (eqn (1)).¹⁹ The ester formation was confirmed by thin-layer chromatography (TLC) using chloroform : dichloromethane (95 : 5 v/v) solvent system. Further identification for ester formation was carried out by gas chromatography/mass spectroscopy GC-MS on a Shimadzu (model GC 17A; model MS QP5050A, Japan) and FTIR (Perkin Elmer, model 1650, UK) instruments.

For purification of the product, after termination of the reaction, the enzyme was filtered and the solvent removed by evaporator under reduced pressure. Product in the remaining mixture were separated via silica gel (Kieselgel 60, Merck, particle size 0.063–0.200 mm) column chromatography (15 cm × 20 mm) using a chloroform/dichloromethane (95/5, v/v) mixture as eluent. A sample made up of 1 : 1 (w/w) ratio of silica gel and the free solvent reaction mixture was deposited at the top of the column previously equilibrated with chloroform/dichloromethane (95/5, v/v) mixture. Five millilitre fractions were collected and tested using thin layer chromatography to identify the product-rich portions. Such fractions were pooled and the solvent evaporated by a rotary evaporator. The purity of product was then checked with TLC and GC before FTIR analysis.

where V_control and V_test are volumes of NaOH solution required to neutralize the excess acid for control (without enzyme) and test (with enzyme) experiments, respectively.

The ANN description

Artificial neural networks that contain input, hidden and output layers are mathematic free functionalization of the complicated practical process. The layers which consist of several nodes, are connected by multilayer normal feed-forward or feed-back connection formula.²⁰ The hidden layer could be more than one parallel layer however the single hidden layer is universally suggested. The connection is that the nodes of particular layer are connected to the nodes of the next layer. The nodes are simple artificial neurons which simulate the behavior of biological neural networks. The nodes of input layer are qualify by sending data via the special weights to the nodes of hidden layer and then to the output layer.^20–23 The qualification is carried out by associated weights during learning process by well known learning algorithms.

The learning process

In the learning process, the weights are calculated by the weighted summation (eqn (2)) of received data from the former layer and transfer to next layer.²⁴ The number of hidden nodes is obtained by trial and error training calculation which is examined from one to n nodes. In the process, the output of the hidden nodes in turn, acts as input to final (output) layer's nodes which undergoes similar or different transformation.where S is summation, b is bias, I_i is the ith input to hidden neuron and W_i is the weight associated with I_i. The bias shifts the space of the nonlinearity properties.

The universal learning algorithms are IBP, BBP, QP, GA and LM while the multilayer is the nodes' connection type.²⁵ The usual transfer function is the logarithmic sigmoid for both hidden and output layers that is bounded from (0–1).²⁶ The sigmoid bounded area is used to normalize the input and output data that is provided by the software scalling. The scaled data are passed into the first layer, propagated to hidden layer and finally meet the output layer of the network. Each node in hidden layer or in output layer firstly acts as a summing junction which modifies the inputs from the previous layer using the following equation:

where y_i is the input of the network to j node in hidden layer, i is number of nodes, x_i are the output of previous layer while w_ij are the weights of connection between the ith node and jth node. The bias associates with node j that is presented by b_j. The main aim of the process is to find the weights for minimizing the error of RMSE which is obtained from difference between network prediction and actual responses.where n is number of the points, y_pi is the predicted values and y_ai is the actual values.²⁷ Therefore, the learning process with an algorithm is continued until finding the minimum RMSE which is called topology. To avoid random correlation due to the random initialization of the weights, learning of a topology is repeated several times. As a result, the topology with the lowest RMSE is selected to compare with other nodes' topologies. Therefore, the topologies for the n numbers of hidden layer for the considered algorithms are obtained in same way. Finally the topologies of the algorithms are compared to select the provisional model by maximum R² (eqn (5)), minimum RMSE and absolute average deviation (AAD) (eqn (6)).where n is the number of points, y_pi is the predicted value, y_ai is the actual value, and y_m is the average of actual values.²⁸

Incremental back propagation algorithm

Gradient descent backpropagation algorithm: this algorithm is one of the most popular training algorithms in the domain of neural networks. It works by measuring the output error, calculating the gradient of this error, and adjusting the ANN weights (and biases) in the descending gradient direction. Hence, this method is a gradient descent local search procedure.²⁹ This algorithm includes different versions such as: (a) standard or incremental backpropagation (IBP): the network weights are updated after presenting each pattern from the learning data set, rather than once per iteration³⁰ (b) batch backpropagation (BBP): the network weights update takes place once per iteration, while all learning data pattern are processed through the network³¹ (c) quick propagation (QP): QP is a heuristic modification of the backpropagation algorithm. It is proved much faster than IBP for many problems. QP is also defined as: mixed learning heuristics without momentum, learning rate optimized during training.³²

The back propagation algorithm of learning a multilayer perceptron neural network (MLP – MultiLayers Perceptron) was described for the first time in 1969 by Arthur E. Bryson.³³ With the version created in 1986 by David E. Rumelhart, Geoffrey E. Hinton and Ronald J. Williams, the back propagation algorithm has gained popularity and interest to neural network field. Batch backpropagation is an optimization technique where the updating of the weights is based on the cyclic use of the entire learning set. Unlike the batch back propagation, the incremental backpropagation is based on weights updating after the utilization of each pattern of the learning set.

Enzyme reusability study

The reusability study of Novozym 435 was carried out in term of percentage conversion for the synthesis of dilauryl azelate ester. The esterification experiments were conducted for 8 cycles of 6 h that more than 45 h working reaction time under the optimal conditions, as follows: enzyme amount (0.14 g), reaction time (360 min), reaction temperature (46 °C), and molar ratio (1 : 4.1). At the end of each reaction, the enzyme was separated from the product by filtration and rinsed with n-hexane for removing the organic phase that surrounded enzyme particles.

Results and discussion

Modeling process

The topologies of the algorithms. In this research, the examined neural network includes four inputs, including enzyme amount, reaction time, reaction temperature, and molar ratio of substrates, while the conversion (%) of dilauryl azelate ester was only the node in output layer. It should be noted that the experimental data of central composite design were divided into two set, 20 of the data sets were used as the training set and the remaining 5 data set were used as the test set (Table 1). The structure of the hidden layer was determined by examining a series of topologies with varied node number from 1 to 15 for each algorithm. The model learning was performed for testing data set to determine minimum value of RMSE as error function. The performance was 10 times repeated for each node to avoid random correlation due to the random initialization of the weight.³⁴

Table 1 The experimental design that consists of training and testing data sets, each row indicates an experiment while the columns present the composition of the synthesis of dilauryl azelate ester

Run No.	Enzyme amount (w/w %)	Reaction time (hour)	Reaction temperature (°C)	Molar ratio of substrates, AzA : LA (mol)	Conversion (%)
					Actual value	Predicted value
Training set
1	0.25	450	52	6	96.74	96.73
2	0.25	90	52	6	80.80	80.81
3	0.25	270	40	6	85.71	85.72
4	0.25	270	64	6	92.52	92.52
5	0.05	270	52	6	93.20	93.20
6	0.45	270	52	6	93.17	93.17
7	0.25	270	52	3	85.24	85.24
8	0.25	270	52	9	79.70	79.70
9	0.25	270	52	6	86.61	86.61
10	0.15	180	58	4.5	87.83	87.83
11	0.15	360	58	4.5	93.51	93.51
12	0.35	360	46	4.5	95.40	95.40
13	0.35	360	46	7.5	90.83	90.84
14	0.35	180	46	4.5	80.62	80.62
15	0.35	360	58	4.5	95.22	95.22
16	0.35	180	58	4.5	89.78	89.78
17	0.15	180	46	4.5	84.14	84.14
18	0.35	180	46	7.5	78.76	78.75
19	0.35	360	58	7.5	90.70	90.70
20	0.15	180	58	7.5	84.98	84.98
Test set
1	0.20	180	50	6	84.23	82.53
2	0.15	360	58	7.5	86.73	87.27
3	0.30	300	52	5	91.19	90.22
4	0.15	360	46	4.5	95.38	95.05
5	0.15	360	46	7.5	90.13	90.22

---

The training was identically carried out for IBP, BBP, QP, GA and LM algorithms to discover the optimized topology for each algorithm. Among the 10 time learning repetition data for each node, the minimum value of RMSE was selected and plotted versus the nodes of the algorithms' hidden layer (Fig. 1). As shown, one node of 15 topologies for each algorithm has presented the lowest RMSE which was selected as the best topology for comparison. The selected topologies were 4-14-1, 4-5-1, 4-13-1, 4-13-1, and 4-10-1 for IBP, BBP, QP, GA and LM algorithms, respectively. As Fig. 1 shown, the topology of IBP-4-14-1 presented the lowest RMSE among the other topologies was selected as provisional model for producing dilauryl azelate ester.

Fig. 1 The selected RMSE vs. node number of the dilauryl azelate ester network's hidden layer for IBP, BBP, QP, GA and LM. The lowest RMSE belong to node 14 (IBP), 5 (BBP), 13 (QP), 15 (GA), 10 (LM).

The model selection. To select the final model for dilauryl azelate ester optimum conditions, the values of RMSE, R² and AAD were relatively studied for the topologies of IBP-4-14-1, BBP-4-5-1, QP-4-13-1, GA-4-13-1 and LM-4-10-1. To calculate the R², the topologies prediction and actual values of the dilauryl azelate ester conversion% were plotted for testing data set (Fig. 2) as well as the R² calculation of training set was carried out in similar way (Fig. 3). As the scatter plots showed, IBP-4-14-1 has presented the highest R² for testing (0.965) and training (1.0) data sets. Moreover, the AAD of testing and training sets for the topologies was calculated by Table 2.

Fig. 2 Scatter plot of predicted conversion (%) value versus actual conversion (%) value by using five algorithms for testing set.

Fig. 3 Scatter plot of predicted conversion (%) value versus actual conversion (%) value by using five algorithms for training data set.

Table 2 The performance results of the optimized topologies, IBP-4-14-1, BBP-4-5-1, QP-4-13-1, GA-4-15-1, LM-4-10-1 of the dilauryl azelate ester synthesis

Learning algorithm	Architecture	Training data			Testing data
		RMSE	R^2	AAD	RMSE	R^2	AAD
GA	4-15-1	0.01715	1.000	0.01294	1.10181	0.920	1.29923
BBP	4-5-1	0.00297	1.000	0.00247	0.92118	0.950	1.20485
QP	4-13-1	0.00123	1.000	0.00033	0.85146	0.937	1.10360
LM	4-10-1	0.00055	1.000	0.00093	0.85462	0.936	1.10333
IBP	4-14-1	0.00220	1.000	0.00163	0.68859	0.965	0.83528

---

As observed, the lowest value of the AAD has also belonged to IBP-4-14-1. As a result, IBP-4-14-1 was pioneer in minimum RMSE and AAD as well as at maximum R² among the topologies for testing and training data sets.

The network of IBP-4-14-1. Fig. 4 shows the network of IBP-4-14-1 as final model for dilauryl azelate ester which consists of input, hidden and output layers.^35,36 The input layer with 4 nodes (enzyme amount, reaction time, reaction temperature and molar ratio of substrates) is the distributor for the hidden layer with 14 nodes which were determined by learning process. The input data of hidden nodes are calculated by weighted summation (eqn (2)).³⁷ Then the output data of hidden layer are transferred to output layer (conversion%) by using log-sigmoid function (eqn (7)).³⁸where f(x) is the hidden output neuron. As a result, IBP-4-14-1 was used to determine the optimum values of the input variables of the dilauryl azelate ester synthesis.

Fig. 4 Schematic representation of a multilayer perceptron feedforword network of ANN based on IBP consisting of four inputs, one hidden layer with 14 nodes and one output.

Description of the dilauryl azelate ester synthesis. In the modeling process, the optimized topologies for different learning algorithms were determined by training and testing data set. The comparison were carried out to define the best relative topology with optimum R², RMSE and AAD which selected as provisional model for more evaluation. The adequacy of the selected model (IBP-4-14-1) was evaluated by testing and validation data sets. Table 3 shows four new trials combinations of experimental factors which do not participate in the training data set. The comparison between experimental observed and predicted data shows excellent agreement for the response (conversion%). As a result of the process, the network of IBP-4-14-1 was selected to navigate the dilauryl azelate ester synthesis. The navigation has contained optimization of the effective variables as well as the importance of them. Table 4 shows the predicted optimum values of were enzyme amount, reaction time, reaction temperature, and molar ratio of substrates, which experimentally performed to obtain actual conversion (95.38%). The ANN result confirmed the validity of the model, and the experimental value was determined to be quite close to the predicted value (95.05%), implying that the empirical model derived from central composite rotatable design can be used to adequately describe the relationship between the independent variables and the response.

Table 3 Validation set predicted by selected model IBP-4-14-1 for dilauryl azelate ester synthesis

Run No.	Enzyme amount (w/w %)	Reaction time (hour)	Reaction temperature (°C)	Molar ratio of substrates, AzA : LA (mol)	Conversion (%)
					Actual value	Predicted value
1	0.15	180	46	7.5	83.91	81.67
2	0.35	180	58	7.5	87.9	86.61
3	0.20	360	52	6.5	90.99	91.55
4	0.16	250	55	5.5	87.17	88.13

---

Table 4 Optimum conditions derived by RSM and ANN for dilauryl azelate ester synthesis

Independent variables				Conversion (%)
Enzyme amount (g)	Reaction time (min)	Reaction temperature (°C)	Molar ratio of substrates, AzA : LA (mol)	Actual value	Predicted value by ANN	Predicted value by RSM	RSE (%) by ANN	RSE (%) by RSM
0.14	360	46	1 : 4.1	95.38	95.05	96.23	0.35	0.88

---

Graphical optimization of the variables. The validated model (IBP-4-14-1) simulated the interaction of the effective variables (enzyme amount (g), reaction time (min), reaction temperature (°C) and molar ratio of substrates (mol/mol)) on the conversion% of dilauryl azelate ester without further requirement of function and equation knowledge. The simulations are the effect of nonlinear relationship of two variables on the conversion% of dilauryl azelate ester which graphically presents in three dimensional plots (3D plots).

Fig. 5a shows the effect of enzyme amount and reaction time on the percentage conversion of dilauryl azelate ester. Response surface plot was generated with a reaction temperature fixed at 46 °C and molar ratio of substrate 1 : 4.1 mol (AzA : LA) for obtaining the interaction of varying enzyme amount (0.05–0.45 g) and the reaction time (90–450 min). The result showed that when increasing the reaction time from 90 to 450 min, the conversion% of dilauryl azelate ester rapidly was increased from 82 to 96%. However, enzyme amount was not significant affect on the percentage of conversion at any given amounts from 0.05 to 0.45 g. High percentage conversion (95.38%) was achieved by increasing the reaction time and reducing the amount of enzyme, which is important from the economic viewpoint since the cost of enzyme is expensive.

Fig. 5 (a) Response surface plot shows the effect of the enzyme amount, reaction temperature and their interaction on the synthesis of dilauryl azelate ester. Other variables are constant; reaction temperature 46 °C and molar ratio of substrate 1 : 4.1 mol (AzA : LA). (b) Response surface plot shows the effect of the molar ratio of substrate, reaction time and their interaction on the synthesis of dilauryl azelate ester. Other variables are constant; enzyme amount 0.14 g and reaction temperature of 46 °C. (c) Response surface plot shows the effect of the reaction temperature, reaction time and their interaction on the synthesis of dilauryl azelate ester. Other variables are constant; enzyme amount 0.14 g and molar ratio substrates at 1 : 4.1 mol (AzA : LA).

Fig. 5b represents the effect of varying molar ratio of substrates (AzA : LA, 1 : 3–1 : 9 mol) and reaction time (90–450 min), and their mutual interaction on dilauryl azelate synthesis at enzyme amount of 0.14 g and reaction temperature of 46 °C. It was shown that the maximum conversion of dilauryl azelate ester (95.38%) was obtained when the amount of substrates molar ratio was reached to the ratio of 1 : 4.1 mol (AzA : LA) and by increasing the reaction time higher than 360 min. Additionally, it was observed that at the ratios higher than 1 : 5 mol (AzA : LA), the conversion percentage of dilauryl azelate ester decreased. This may due to out of the critical molar ratio, competing alcohol binding reduce formation of the acyl–enzyme complex and thereby results in decrease of alcoholysis.³⁹

As shown in Fig. 5c, the effect of reaction temperature and reaction time were interpreted in range of 40–64 °C and 90–450 min, respectively with the enzyme amount fixed at 0.14 g and molar ratio substrates at 1 : 4.1 mol (AzA : LA). Fig. 5c shows that the higher percentage conversion of dilauryl azelate ester (95.38%) was obtained at optimal reaction temperature range between 46–50 °C and reaction time higher than 360 min. These results were similar in those most reviewed papers, which is Novozym 435 was optimally used at temperature of 40 °C and 60 °C. Higher reaction temperatures tended to induce enzyme inactivation due to denaturation processes.³⁹

Importance of effective variables. According to Fig. 6, all selected operating parameters are strongly influential on the conversion (%) of dilauryl azelate ester. Thus, any of the variables studied in this work cannot be neglected for process analysis. The results indicate that molar ratio of substrates with a relative importance of 27.93% is the most influential parameter on the synthesis of dilauryl azelate ester. The order of relative importance of the input variables on the synthesis of dilauryl azelate ester was as follows: molar ratio of substrates > enzyme amount > reaction temperature > reaction time.

Fig. 6 The relative importance of dilauryl azelate ester synthesis of input variables consist of molar ratio of substrates, enzyme amount, reaction temperature, and reaction time.

Enzyme reusability and operational stability. A catalyst by definition should be recovered and reused. The efficiency of a catalyst is determined by studying its reusability.⁴⁰ The reusability experiment of Novozym-435 was carried out in term of percentage conversion for the synthesis of dilauryl azelate ester. Fig. 7 shows that Novozym 435 retains high activity even after 8 cycles of 6 h under optimal conditions, so that, during more than 45 h of working reaction time, the result had a 4.75% drop in percentage conversion. This may be due to the shear effect on the immobilized enzyme particle cause by agitator in a reaction system. Shear effect may destroy the immobilization system of Novozym 435 because of increasing in the collision of enzyme particle with the wall of the screw-capped glass vial and cause the enzyme activity to decrease from one cycle to another in the reusability study. Meanwhile, the mass loss of the enzyme in the recycling process may also lead to decrease in reaction percentage conversion of dilauryl azelate ester.⁴¹

Fig. 7 Reusability of Novozym 435 within 48 h working reaction time. All cycles were done in optimum conditions (enzyme amount: 0.14 g, reaction time: 6 h, reaction temperature: 46 °C, and substrates molar ratio: 1 : 4.1 mol) scale.

Conclusions

The ANN-based design of experiment was used for the synthesis of dilauryl azelate ester. The optimization of process parameters was carried out based on the investigations relating to the influence of enzyme amount, reaction time, reaction temperature, and molar ratio of substrates by using artificial neural networks. To obtain the qualified network, different algorithms such as IBP, BBP, QP, GA and LM were learned by using training, and testing data sets. The results of the learning program were obtained by five best topologies; IBP-4-14-1, BBP-4-5-1, QP-4-13-1, GA-4-13-1 and LM-4-10-1. The performance of the topologies was optimized by RMSE, AAD and R². The topology (IBP-4-14-1) with the lowest RMSE, AAD and the highest R² was selected as provisional network of the synthesis of dilauryl azelate ester for validation test. The results of the validation confirmed high predictability of the model. The validated model determined the optimum values and relative importance of the effective variables. The importance of the variables which include molar ratio of substrates, 27.93%, enzyme amount, 26.33%, reaction temperature, 25.65% and reaction time, 20.09% showed none of the variables were neglectable in this work. In conclusion, ANN is an efficient quantitative tool which is able to model the effective input variables to predict the conversion% of dilauryl azelate ester enzymatic reaction.⁴²

Acknowledgements

The financial assistance provided by Universiti Putra Malaysia under the Research University Grant Scheme (RUGS), is gratefully acknowledged.

References

1. A. Fitton and K. L. Goa, Drugs, 1991, 41, 780–798.
2. G. Webster, J. Am. Acad. Dermatol., 2000, 43, S47–S50.
3. J.-F. Hermanns, L. Petit, C. Piérard-Franchimont, P. Paquet and G. Piérard, Dermatology, 2002, 204, 281–286.
4. M. Nazzaro-Porro, Dermatology in Five Continents, Springer, Verlag Berlin Heidelberg, 1988.
5. D. Tamarkin, US Pat., 20,040,191,196, 2004.
6. C. Charnock, B. Brudeli and J. Klaveness, Eur. J. Pharm. Sci., 2004, 21, 589–596.
7. P.-W. Hsieh, S. A. Al-Suwayeh, C.-L. Fang, C.-F. Lin, C.-C. Chen and J.-Y. Fang, Eur. J. Pharm. Biopharm., 2012, 81, 369–378.
8. S. Haykin, Neural Networks: A Comprehensive Foundation, Mac millian, New York, 1994.
9. M. Lam, Decis. Support Syst., 2004, 37, 567–581.
10. R. Ghazali, A. Jaafar Hussain, N. Mohd Nawi and B. Mohamad, Neurocomputing, 2009, 72, 2359–2367.
11. P. J. Lisboa and A. F. Taktak, Neural Network, 2006, 19, 408–415.
12. M. Egmont-Petersen, D. de Ridder and H. Handels, Pattern Recogn., 2002, 35, 2279–2301.
13. J. J. Lahnajärvi, M. I. Lehtokangas and J. P. Saarinen, Neurocomputing, 2004, 56, 345–363.
14. M. W. Craven and J. W. Shavlik, Future Generat. Comput. Syst., 1997, 13, 211–229.
15. D. Bingöl, M. Hercan, S. Elevli and E. Kılıç, Bioresour. Technol., 2012, 112, 111–115.
16. F. Despagne and D. L. Massart, Analyst, 1998, 123, 157R–178R.
17. N. Chaibakhsh, M. Abdul Rahman, S. Abd-Aziz, M. Basri, A. Salleh and R. Rahman, J. Ind. Microbiol. Biotechnol., 2009, 36, 1149–1155.
18. H. R. Fard Masoumi, M. Basri, A. Kassim, D. Kuang Abdullah, Y. Abdollahi, S. S. Abd Gani and M. Rezaee, Sci. World J., 2013, 2013, 1–9.
19. S. C. Lau, H. N. Lim, M. Basri, H. R. F. Masoumi, A. A. Tajudin, N. M. Huang, A. Pandikumar, C. H. Chia and Y. Andou, PLoS One, 2014, 9, 1–10.
20. A. Ghaffari, H. Abdollahi, M. Khoshayand, I. S. Bozchalooi, A. Dadgar and M. Rafiee-Tehrani, Int. J. Pharm., 2006, 327, 126–138.
21. M. G. Moghaddam, F. B. H. Ahmad, M. Basri and M. B. A. Rahman, Electron. J. Biotechnol., 2010, 13, 3–4.
22. S. Aber, A. Amani-Ghadim and V. Mirzajani, J. Hazard. Mater., 2009, 171, 484–490.
23. G. D. Garson, AI Expet, 1991, 6, 46–51.
24. M. Khare and S. S. Nagendra, Artificial neural networks in vehicular pollution modelling, Springer, 2007.
25. D. Salari, N. Daneshvar, F. Aghazadeh and A. Khataee, J. Hazard. Mater., 2005, 125, 205–210.
26. Y. Abdollahi, A. Zakaria, A. H. Abdullah, H. R. F. Masoumi, H. Jahangirian, K. Shameli, M. Rezayi, S. Banerjee and T. Abdollahi, Chem. Cent. J., 2012, 6, 88.
27. H. R. F. Masoumi, M. Basri, A. Kassim, D. K. Abdullah, Y. Abdollahi and S. S. A. Gani, J. Surfactants Deterg., 2014, 17, 287–294.
28. Y. Abdollahi, N. A. Sairi, M. K. Aroua, H. R. F. Masoumi, H. Jahangirian and Y. Alias, J. Ind. Eng. Chem., 2014, 25, 168–175.
29. D. Rumelhart, G. Hinton and R. Williams, Nature, 1986, 323, 533–536.
30. A. P. Plumb, R. C. Rowe, P. York and M. Brown, Eur. J. Pharm. Sci., 2005, 25, 395–405.
31. M. T. Hagan, H. B. Demuth and M. H. Beale, Neural network design, Pws Boston, 1996.
32. A. Ghaffari, H. Abdollahi, M. R. Khoshayand, I. Soltani Bozchalooi, A. Dadgara and M. Rafiee-Tehrani, Int. J. Pharm., 2006, 327, 126–138.
33. A. E. Bryson, Y.-C. Ho and G. M. Siouris, IEEE Transactions on Systems, Man, and Cybernetics, 1979, 9, 366–367.
34. M. Kasiri, H. Aleboyeh and A. Aleboyeh, Environ. Sci. Technol., 2008, 42, 7970–7975.
35. W.-H. Ho, K.-T. Lee, H.-Y. Chen, T.-W. Ho and H.-C. Chiu, PLoS One, 2012, 7, 1–9.
36. H.-Y. Shi, K.-T. Lee, H.-H. Lee, W.-H. Ho, D.-P. Sun, J.-J. Wang and C.-C. Chiu, PLoS One, 2012, 7, 1–6.
37. J. Zhang, A. Morris, E. Martin and C. Kiparissides, Chem. Eng. J., 1998, 69, 135–143.
38. L. Aijun, L. Hejun, L. Kezhi and G. Zhengbing, Acta Mater., 2004, 52, 299–305.
39. M. B. A. Rahman, N. Chaibakhsh, M. Basri, R. N. Z. R. A. Rahman, A. B. Salleh and S. M. Radzi, J. Chem. Technol. Biotechnol., 2008, 83, 1534–1540.
40. S. Karmee, Energy Sources, Part A, 2015, 37, 536–542.
41. H. R. F. Masoumi, M. Basri, A. Kassim, D. K. Abdullah, Y. Abdollahi, S. S. A. Gani and M. Rezaee, J. Ind. Eng. Chem., 2014, 20, 1973–1976.
42. Y. Abdollahi, A. Zakaria and N. A. Sairi, Clean: Soil, Air, Water, 2014, 42, 1292–1297.

---