Optimization of self-emulsifying drug delivery system of cefuroxime axetil

Overcoming solubility problems is the greatest challenge during formulation of poorly soluble active pharmaceutical ingredients (API’s) into oral solid dosage forms. Different formulation approaches were used to surpass this problem and enhance their solubility in the gastrointestinal (GI) fluids, in order to achieve a faster dissolution and better absorption, which will directly influence their therapeutic effect. In this paper, an evaluation of the potential of a self-emulsifying drug delivery system (SEDDS) to improve the solubility of the active ingredient cefuroxime axetil (CA) was done. Screening of the solubility of the API in different excipients was done, and Tween 80, PEG 400, and Olive oil as a surfactant, co-solvent, and oil, respectively, were chosen as the most convenient system constituents. An optimal self-emulsification and solubilization ability of this system was assessed using mixture experimental design statistical tools based on the response surface methodology (RSM). The prepared CA-SEDDS were evaluated for droplet size (d10, d50, d90 in μm), droplet size distribution (Span factor), and absorbance. As a complementary approach, for better representation of the non-linear relationship between the formulation compositions and the observed dispersion characteristics an artificial neural network (ANN) was used. Optimal formulation that consists of 10% (w/w) Tween 80 as surfactant, 80% (w/w) PEG 400 as co-solvent and 10% (w/w) Olive oil, was obtained. Both, mixture experimental design and ANN were combined for a comprehensive evaluation of CA-SEDDS and the obtained results suggested that formulation of SEDDS is a useful approach for improving the solubility of the CA.


Introduction
The lipophilic nature of the active pharmaceutical ingredients (API's) is a major disadvantage for their oral administration. It determines their solubility in the gastrointestinal (GI) fluids, consequently the speed of their dissolution and absorption, therefore directly affecting their therapeutic effect. It has been estimated that up to 70% of the new chemical entities have poor water solubility, along with 40% of poorly soluble drugs that are already on the market (Perioli and Pagano, 2012). These API`s are classified as class II or class IV compound according to the Biopharmaceutical Classification systems (BCS), having poor water solubility and good permeation or poor water solubility __________________ * etrajanovska@alkaloid.com.mk Maced. pharm. bull., 66 (2) 67 -79 (2020) and poor permeability, respectively. The solubility and permeability of the API`s are mostly affected by GI conditions such as pH, pKa, volume and amount of surfactants in the gastric fluids, molecular size, the lipophilicity of the API, and its affinity to influx and efflux transporter proteins. GI environment is influenced by the presence or absence of food, which can change the volume, pH, or viscosity of the GI fluid, additionally can influence the gastric emptying time and the amount of secreted bile, hence affecting the API solubility (Kamboj and Rana, 2016). Co-administration of poorly soluble drugs with a meal rich in fat can have an impact on the bioavailability by enhancing the API solubilization, affecting dose linearity as well as showing inter-and intrasubject variability (Porter et al., 2008). Furthermore, for poorly water-soluble API`s progressive increase of the dose is usually needed in order to achieve the therapeutic concentration of the drug upon oral administration, which in most cases can cause drug toxicity and can affect patients' compliance. Additionally, high dose loading might cause more problems during the manufacturing process and an increase in process cost.
Innovative formulation approaches have been explored to overcome this problem (Kawabata et al., 2011). Several formulation strategies aiming to improve the solubility of the API in water were developed, such as micronization, solid dispersions, complexation, and microencapsulation (Bahloul et al., 2014). Moreover, in recent years, attention has been focused on the formulation of lipid solutions, emulsions and emulsion concentrates, which can be prepared as physically stable formulations suitable for encapsulation of poorly water soluble drugs. Self-emulsifying drug delivery systems (SEDDS) are one of the formulation techniques that can be used to overcome solubility problems. These systems are defined as isotropic multi-component drug delivery systems composed of oil, surfactant, and co-surfactant/co-solvent, which form an emulsion in the presence of water. When orally administrated, these systems with the help of the gentle peristaltic agitation spontaneously self-emulsify upon contact with the gastrointestinal fluids forming fine oil-in-water emulsions. Formed emulsion droplets can have different sizes ranging from few nanometers to several micrometers. Fast self-emulsification provides a high surface area for interaction between the formulation and gastrointestinal fluids, therefore, offering better absorption and a reproducible plasma concentration (Chakraborty et al., 2009;Porter et al., 2008).
In this study, the amorphous form of cefuroxime axetil (CA) was used as a model active substance. CA is a second-generation cephalosporin that has a broad spectrum of antibacterial activity against Gram-positive and Gram-negative microorganisms by inhibiting their bacterial enzymes necessary for cell-wall synthesis (peptidoglycan synthesis), thereby causing cell death. According to the BCS, CA is classified as a class IV drug, meaning it is a drug molecule that has poor water solubility and poor permeability upon oral administration. CA is stable in the acidic conditions of the stomach, afterwards when absorbed from the GI tract it is hydrolyzed by esterases to cefuroxime. Literature data show that bioavailability of CA after oral administration is ranging from 30% in fasted to 50% in fed states, hence is hard to establish the optimal oral dosage schedule (Gorajana et al., 2015).
The objective of our study was development and optimization of CA-SEDDS using mixture experimental design and artificial neural networks (ANNs) in order to achieve better solubility and oral bioavailability of the API.
All used excipients were of pharmaceutical grade and were used as received. Optimization of CA-SEDDS Mixture experimental design. To investigate the influence of the chosen excipients in the system on the emulsification efficacy (evaluated by droplet size, droplet size distribution (Span factor) and absorbance), experimental runs were designed using software Design expert (Design-Expert V8 trial, Stat-Ease, Inc., Minneapolis, USA) following mixture simplex lattice design. In this design, the total amount of the components was held constant, while the proportions of the mixture components were changed. The total of the components was 100% (Smix=1 g). A preliminary assessment of the mixtures' ability to form a self-emulsifying system was performed to define the levels of the constraints for the independent variables. The established levels of the constraints were: for Factor A (PEG 400) 40 to 80%, Factor B (Tween 80) 10 to 50 %, and Factor C (Olive oil) 10 to 50 %. Coded and actual values for the levels of the independent formulation variables (factors) are given in Table 1, while the designed experiments are presented in Table 2. Twenty experimental runs were done in random order to increase the predictability of the model. The responses, droplet size (d 10, d 50 , d 90 in µm), droplet size distribution (Span factor), and absorbance (Abs) were analyzed and fitted in appropriate models. The most convenient model was selected based on the statistical parameters provided by analysis of variance (ANOVA) and polynomial equations were generated for each response. The same software was used to plot 3D response surface graphs of the predicted values and desirability levels for each variable.  The structure of the multilayer perceptron, a type of an artificial neural network, employed in the study. It consists of three layers, an input layer of three nodes representing the three input factors, a hidden layer of five nodes, and an output layer of five nodes representing the five variables of interest.

Solubility of CA
Artificial neural networks. It is known that any arbitrary function can be fitted with some artificial neural networks (ANN), such as the multilayer perceptron (MLP). Therefore, the scikit-learn machine learning library in Python was employed, where the "sklearn.neural_network.MLPRegressor" class allows solving regression problems using MLP. As shown in Fig.  1, a three-layer network was constructed where the inputs were the three factors in normalized ranges from 0 to 1, for each of the factors. For example, Factor A can vary between 40% and 80%, so 80% is mapped into 1 and 40% into 0, while the other values fall in between. A similar normalization is done for factors B and C, where 50% is mapped into 1 and 10% into 0. In this way, the ANN can be trained better. The hidden layer consisted of five nodes with a logistic activation function. The outputs were the five variables of interest, droplet size (d 10, d 50 , and d 90 in µm), droplet size distribution (Span factor), and absorbance. The training was performed with backpropagation using the LBFGS solver. The initial weights in the ANN were randomly assigned. The other settings of the ANN were: constant learning rate of 0.001, stopping tolerance of 0.0001 during 10 iterations, 200 maximum iterations, 15000 maximum function calls, L2 regularization penalty coefficient of 0.0001, and shuffled samples in each iteration. Most of these settings were left to the default values/options as they provided good results.

Preparation of CA-SEDDS
CA-SEDDS's were prepared by adding 125 mg CA into 1 g of the previously prepared mixtures (Table 2). Subsequently, the mixtures were vortexed for a period of 1 min. (IKA Vortex 1, Germany) and mixed for 24 hours on a magnetic stirrer (150 rpm, 37 °С; IKA C-MAG HS Magnetic Stirrer, Germany) to assure a complete dissolution of the CA.

Characterization of CA-SEDDS
Droplet size and droplet size distribution analysis. Dispersions were prepared by mixing 500 µL of CA-SEDDS with 150 mL 0.1 M HCl (37 °С, 150 rpm, 15 min.; IKA C-MAG HS Magnetic Stirrer, Germany). Analysis of droplet size (d 10 , d 50 , d 90 , µm) and droplet size distribution (Span factor) was performed with laser diffraction method using Mastersizer 3000 (Malvern Instruments Ltd., UK) equipped with wet dispersion cell (Hydro MV, Malvern Instruments Ltd., UK). All the measurements were repeated five times. Before the analysis, the index of refraction of the samples was evaluated (RI=1.339). The measurements were done under the following conditions: water was used as a dispersant, mixing speed was set as 2520 rpm and the obscuration ranged from 1.5 -2%. Droplet size is expressed as dv and droplet size distribution as Span factor: where d 10 , d 50 , and d 90 are droplet sizes expressed in µm, for the corresponding cumulative diameters of 10, 50, and 90% of the droplets. Span factor is a measure of the distribution width of the measured droplets in relation to the mean diameter. A high value for this parameter indicates wide droplet size distribution (Simonoska Crcarevska et al., 2008).

Solubility of CA
Digestible oils such as sunflower oil, soya bean, almond oil, and olive oil are preferred over synthetic/semisynthetic oils. These oils provide better bioavailability for lipophilic APIs by improving their solubilization through lipids digestion (Porter et al., 2008). After performing the solubility study, among the oils screened olive oil showed the highest solubility for CA and was chosen as an oil phase in the CA-SEDDS. According to the results of the solubility study, Tween 80 showed the highest solubility for CA among the surfactants. This hydrophilic surfactant has an HLB value of 15 and is considered superior in forming fine and uniform emulsion droplets (Garg et al., 2017). Additionally, it has been reported that Tween 80 with its nonionic nature helps in providing better emulsion stability over a wide range of pH and ionic strength (Gupta et al., 2011). It is expected that this surfactant will reduce the interfacial energy by creating a layer around the emulsion droplets and will provide a mechanical barrier to coalescence; therefore, it will prevent the precipitation of the active substance in the GIT. In order to create a layer with sufficient flexibility that can easily decrease the bending stress of the interface, incorporation of co-solvent in the system is recommended (Parmar et al., 2011). PEG 400 showed the highest solubility between the evaluated co-solvents and was included in the formulation.
Mixture experimental design. In the view of the fact that the conventional "trial-and-error" method is laborintensive and time-consuming, mixture experimental design statistical tools based on the response surface methodology (RSM) have been used for the optimization of SEDDS (Desai et al., 2008). This tool was employed to investigate the effect of the independent variables (oil, surfactant, and co-solvent ratio) on the responses: droplet size (d 10, d 50 , d 90 in µm), droplet size distribution (Span factor), and absorbance. The optimization study aimed to investigate the significance of independent variables and their interactions since it has been reported that these are the major factors influencing the in vitro dispersion of CA-SEDDS formulation (Tang et al., 2012;Yeom et al., 2015). The chosen responses are considered critical attributes of the system influencing the solubility of poorly soluble drugs.  The obtained results for the dependent variables of all 20 formulations generated with the mixture simplex lattice design are presented in Table 3. Systems showing lower average size, span factor, and higher absorbance were preferred. Mathematical models and their polynomial equations describing the relationship between the system components were established for the studied responses.
Response droplet size ( The quadratic model was suggested as the best fitting model for the responses by comparing the statistical parameters such as p-value, lack of fit, squared correlation coefficient (R 2 ), adjusted R 2 , predicted R 2 , and predicted residual sum of squares (PRESS) ( Table 4).
A positive sign of the coefficients in front of the factors indicates a positive correlation between the response of interest and the variables studied, while the negative term stands for negative proportional dependence. The influence of the independent variable on the response corresponds with the coefficient value. Oneway ANOVA (p<0.05) showed that factors A, B, C, AB, AC have a statistically significant effect on d 10 and d 50 , and A, B, C, AB, AC, and BC on the response d 90. From Eq. 1, 2, and 3 it can be observed that dispersion droplet size is positively correlated with the increase of the PEG 400; Tween 80, Olive oil as well as PEG 400*Tween 80 and PEG 400*Olive oil, however, there is a negative proportional relationship with Tween 80*Olive oil.
Droplet size and absorbance of the emulsion are affected by the composition and physicochemical properties of both, the oil and the aqueous phase. A prerequisite for longer shelf life is the formation of smaller droplets since this type of emulsion resists better to gravitational separation, such as sedimentation or creaming (Wooster et al., 2008). Also, the droplet size is one of the critical characteristics influencing the in vivo behavior of the emulsion because this parameter determines the rate and the extent of drug release. Formation of smaller droplets provides a larger interfacial area for the partitioning of the drug between the SEDDS and the GI fluids (Rahman et al., 2011).
Literature data regarding the influence of the formulation variables on the parameter droplet size are in favor of the obtained results. A thermodynamically stable emulsion is formed with a reduction of the interfacial energy between the two phases after the surfactant is adsorbed on the surface of the oil droplets to form the interface film. This film decreases the coalescence of the droplets with a distribution of the hydrophobic groups of the surfactant in the hydrophobic phase on the interface and its` hydrophilic groups in the hydrophilic phase. The concentration of the surfactant forming the interface film is crucial to the stability, droplet size, and droplet size distribution of the formed emulsion (Zheng et al., 2020). Although an increase of the amount of surfactant leads to a formation of smaller droplets due to stabilization when located at the oil-water interface, above a certain Макед. фарм. билт., 66 (2) 67 -79 (2020) surfactant concentration it can cause water penetration into the oil droplet, which can result in a breakdown of oil droplets and a formation of bigger droplets (Li and Chiang, 2012;Parmar et al., 2011). It has been assumed that the presence of an excess amount of the surfactant can cause an increase in the rate of Ostwald ripening in the emulsion, most probably by micellar dissolution which promotes the development of attractive forces between the oil droplets (Agrawal et al., 2017;Wooster et al., 2008).
Moreover, the droplet size is influenced by the presence of the co-solvent and the oil. Higher amounts of olive oil promote the formation of bigger droplets due to its higher viscosity compared to the water, due to the significant decrease of the droplet break-up rate (Tang et al., 2012). Hence, co-solvent PEG 400 is needed to increase the viscosity of the continuous phase and to promote a reduction of the emulsion average droplets size (Wooster et al., 2008). The co-solvent also increases the interfacial fluidity with penetration into the surfactant film and formation of void space between surfactant molecules (Sharma et al., 2018). Co-solvent synergistically interacts with the surfactant and forms a mechanical barrier against coalescence of the drops thus influencing the emulsion stability (Parmar et al., 2011).
To provide a deeper insight into the way in which component proportions affect the studied responses the effects of the components' ratio on the studied responses are shown in their respective 3D response plots in Fig. 2.
From the plot presented as Fig. 2 the interactive mutual relationship existing between the used excipients and their influence on the droplet size can be observed. The blue region represents the optimal concentration of the excipients needed to obtain smaller droplets.
Droplet size distribution (Span factor). In emulsion formulation, beside the droplet size, droplet size distribution plays a crucial role in determining its stability. The Span factor describes the uniformity of the droplet sizes. Lower values for this parameter indicated monodisperse dispersion, which has good kinetic stability, opposite to higher values that are an indication for a less stable polydisperse system. From the results  Table 3 are in the range from 2.07 to 2.43. After comparing the statistical parameters presented in Table 4 the relationship of the examined variables and Span factor (Eq. 4) was best described by the linear model and the following polynomial equation in terms of coded factors was derived: Span factor = +2.24A + 2.14B + 2.37C [Eq. 4] From Eq. 4 it can be observed that droplet size is positively correlated with the increase of the PEG 400; Tween 80 and Olive oil. These obtained results are in accordance with the literature data, similar as for the parameter droplet size, i.e. increase of surfactant and oil concentration may lead to droplets coalescence, since the dissolution rate of the surfactant as well as its interface adsorption are directly affected by their concentration. All evaluated formulations have narrow globule size distribution therefore indicating their robustness to any apparent instability such as flocculation, coalescence, creaming, or phase separation. On that account, these systems may be considered thermodynamically stable.
The relationship between the studied system components is also presented in the respective 3D response plot presented in Fig. 3. The optimal concentrations of the excipients for which the lowest span factor values are predicted are presented as a blue-colored region. Absorbance. Considering that microemulsion can be defined as a system of water, oil, and surfactant (or amphiphile) which is an optically isotropic and thermodynamically stable solution, the appearance of the emulsion can be considered as one of the most important parameter influencing its overall quality (Parmar et al., 2011). The structure and composition of the emulsion, which depends on the size, concentration, and spatial distribution of droplets, affect the general emulsion appearance (Katsouli et al., 2018). Therefore, emulsion appearance is closely correlated by its interaction with the electromagnetic radiation in the spectrum visible region, such as reflecting, transmitting, adsorbing, or scattering the light (Park et al., 2013). Bigger droplets have smaller scattering efficiency and the light easily penetrates the emulsion, therefore the absorbance (Abs.) of the emulsion decreases with increasing of the droplet size and decreasing of the droplet concentration (Park et al., 2013).
The obtained results for the absorbance of the prepared CA-SEDDS are presented in Table 3. The results for absorbance from all 20 formulations ranged from 0.2813 to 1.3801. The relationship of the examined variables and the logarithmic function of absorbance can be described as by a quadratic model and the polynomial equation (Eq. 5) in terms of coded factors is: As it can be observed from Eq. 5 droplet size is negatively correlated with the increase of the PEG 400; Olive oil as well as PEG 400*Tween 80 and PEG 400*Olive oil, however, there is a synergistic proportional relationship with Tween 80 and Tween 80*Olive oil. A pvalue of <0.05 showed that factors A, B, C, AB, AC have a statistically significant effect on the response absorbance . Therefore, with the increase of the concentration of the surfactant Tween 80 an increase of the absorbance is observed, meaning smaller droplets are formed. On contrary, the increase of the other two excipients, PEG 400 and Olive oil promote the formation of bigger droplets, hence a decrease of absorption has been observed. High values for absorbance suggest that the dispersion is isotropic with uniform small droplets, which in turn indicates that the formulation possesses a large surface area for release of the drug thus favoring enhanced absorption and increased oral bioavailability. Even though the results suggest that a higher amount of the surfactant is necessary for the formation of smaller droplets, as discussed for the parameter droplet size, the concentration of this excipient should be carefully optimized since high concentrations can also promote recoalescence of the droplets and opposite effect (Wooster et al., 2008).
A three-dimensional surface plot (Fig. 4) from which data on absorbance within the limits of the experimental region can be extrapolated also exemplified the relationship.
The blue-colored region represents the possible optimal concentration range of the excipients to obtain a higher absorbance value.  3D surface ternary plots of response Absorbance.

Optimization
The Design Expert 8 software was also used to combine all responses in one measurement and predict the optimum levels of the independent variables, through the desirability function. The set goals were minimal values for the responses droplet size and droplet size distribution and maximal values for the absorbance. For the crossvalidation of the model, the predicted and experimental results for the proposed CA-SEDDS and the percent of the relative error are presented in Table 5 and graphically in Fig. 5. The highest percentage of the relative error of 7.79% was observed between the predicted and experimental values obtained for the response d 10 , and the lowest for the response Span factor, a value of -0.89%.

Artificial neural network and comparison with mixture experimental design
Benefits from using some standard ANNs, such as multilayer perceptron (MLP) have been considered (Medarević et al., 2016). When it comes to formulation optimization, RSM has been widely used since it enables fitting linear and non-linear models such as cubic and quadratic. Additionally, RSM allows selection of optimal solution/s through mapping of the response by contour plot or 3D response graphs as shown in Fig. 2 and 3. However, the prediction of responses obtained using polynomial equations is oftentimes limited in describing the complex relationship between the variables and responses (Kundu et al., 2015). As a complementary approach, an ANN has been used. ANNs are biologically inspired models, which mimic the operation of the nervous systems, and allow a certain function to be represented as a black-box, i.e. it is not important how the results are obtained as long as it provides suitable outputs for the given inputs. MLP is a type of ANN which is relatively simple but allows representation of any arbitrary function, including non-linear relationships. First, the MLP needs to be trained with some observed data using some training method, in order to tune its intrinsic parameters. Then, it can be used to predict the outputs for any given inputs (Amani et al., 2008: Shalaby et al., 2014.   The same data presented in Table 3 and used for the mixture experimental design were also used for training the ANN. Because ANN uses the same input space as the mixture experimental design, it can be considered as a complementary approach. Thus, the weights of the ANN are fitted to model the training data and can be then used for predicting the outcome for any given input factors. To test the predictive power of the trained ANN, seven test formulations were randomly chosen, for which both experimental and predicted values were obtained. The results of these additional test experiments are given in Table 6, along with the values predicted with the ANN and the mixture experimental design. The predictive power of the model was compared using the following two standard measures, the root mean squared error and the coefficient of determination. The root mean squared error is defined as , where Y pred is the value predicted with the corresponding model (ANN or DoE), and Y exp is the experimentally observed value. The coefficient of determination R 2 is calculated as , where SS res is the residual sum of squares, and SS reg is the explained sum of squares.
From the obtained calculations, it can be noticed that R 2 is slightly higher when using ANN for all output variables, especially for parameters absorbance and span factor. Regarding the RMSE very similar results were obtained for both models for parameter droplet size, lower value was obtained with mixture experimental design for the span factor, and with ANN for the absorbance. The ANN has 40 parameters (weights), while the mixture experimental design uses a total of 27 parameters, hence, it can be expected that the ANN could provide a better fit to the training data and a better prediction.
Even though the RMS is a very efficient method and provides a good insight into the quantitative relationships between the independent and dependent variables, its main limitation is that can be used for the assumption of quadratic non-linear correlation only. Therefore, the appropriate use of this tool is particularly dependent on the design space and requires a good knowledge of the system before defining the constraints of the variables. On the other hand, ANN can be easily used for any form of non-linearity, hence, allowing bigger freedom regarding the choice of the design space (Desai et al., 2008). Both approaches can be used as modeling tools during formulation development and optimization and provide certain benefits when applied together.

Conclusion
In this study, CA-SEDDS was successfully optimized to improve the solubility of the CA. Combined, mixture experimental design and ANN were employed for Макед. фарм. билт., 66 (2) 67 -79 (2020) better understanding and optimization of the formulation through evaluation of the dependence of the concentration of the used excipients and the quality characteristics of the dispersion such as droplet size, droplet size distribution and absorbance. A nonlinear relationship between the independent variables, and dispersion droplet size and absorbance was observed, while the dependence of the droplet size distribution was described with a simple linear relationship. After validation of the predictive power of the models, an optimal formulation with desired characteristics that consists of 10% (w/w) Tween 80 as a surfactant, 80% (w/w) PEG 400 as co-solvent, and 10% (w/w) as oil, was obtained.