Based on the results (Table 1), the main effect of PEG molecular weight was its high significant effect on partitioning of serine protease. It was observed that the maximum partition coefficient (84.2) was obtained at 8000 g·mol⁻¹ of PEG molecular mass. Meanwhile, there was a decrease in partition coefficient at higher PEG molecular weight, because, in this condition, PEG gets more compact conformation with intermolecular hydrophobic bonds and prevents enzyme migration into the top phase, thus target enzymes should be transferred to the bottom phase. Also, purification factor and yield of serine protease were decreased at 10,000 g·mol⁻¹ of PEG molecular weight. Reports have shown that when PEG chain length increases, available free volume to accommodate serine protease in the top phase decreases, resulting in a decrease of enzyme specific activity in this phase [18]. When a 6000 g·mol⁻¹ of PEG molecular weight was used to purify the serine protease, the purification factor of enzyme also was decreased. This could be due to the decreased exclusion effect of PEG, and thus, both contaminate and target enzyme could be moved from bottom phase to the top phase and decrease the purity and yield of enzyme. Therefore, the intermediate molecular weight is the best choice for the ATPS purification of serine protease. As clearly shown in Figure 1a, the partition coefficient was significantly increased in the presence of 8000 g mol⁻¹ of PEG molecular mass.

The partition coefficient and yield of serine protease were significantly influenced by the main and quadratic effects of TLL and the interaction effect of TLL and PEG molecular mass (Table 1). The partition coefficient and yield of serine protease were 18.3 and 34% at 6.25% (w·w⁻¹) of TLL. It has been reported that increase in tie-line length causes to increment the hydrophobicity of the top phase as well as interfacial tension between two phases. Based on the results, 17.4% (w·w⁻¹) TLL caused a better partitioning of serine protease into the top phase and further increase of the yield. The maximum partition coefficient and yield were 84.2 and 97.3% at TLL of 17.4% (w·w⁻¹). Rito-Palomares and Hernández [19] reported that when TLL increases, the free volume of the bottom phase decreases, thus adding protein partitioning from bottom to the top phase or to the interface. Therefore, an increase in TLL caused an increase in the protein partition coefficient, resulting in an increase in the yield of serine protease.

At higher TLL [25.6% (w·w⁻¹)], partition coefficient and yield were significantly decreased to 34 and 50.8%, respectively, which could possibly be due to reduction in the volume ratio of top phase thus serine protease activity and yield significantly decreased in this phase. Figure 1a,b indicates how significant (p < 0.05) interaction effects of TLL and PEG molecular mass influenced the partition coefficient and yield of serine protease. The figure shows that when TLL was increased from 0% (w·w⁻¹) to 17.4% (w·w⁻¹), partition coefficient and yield increased but the increase in TLL value from 17.4% (w·w⁻¹) to 25.6% (w·w⁻¹) led to a decrease in partition coefficient and yield of purified enzyme.

The main and quadratic effects of pH and interaction effect of pH and tie line length showed a highly significant (p < 0.05) effect on purification factor of serine protease (Table 1). An important parameter to control enzyme partitioning is pH. The purification factor of serine protease was significantly decreased at pH 6, perhaps because the denaturation of the enzyme was higher when the lower pH was applied. Also, at pH 6, with serine protease partitioned to the bottom phase, it showed that change of enzyme partition behavior was the effect of protein charge. When the pH of system changed, the partitioning of the serine protease was based on the net charge of the protein and surface properties other than the charge.

The isoelectric point (pI) of serine protease is about 6, thus the enzyme has negative charge above pH 7–9 and PEG tends to interact with negative charge. The results indicate that the highest purification factor (14.37) with improvement in partitioning of the enzyme was achieved at pH 7.5. According to the results, ATPS of pH 7.5 was chosen for the serine protease purification. Figure 1c shows the relationship between purification factor and interaction effect of independent variables.

The purification of the serine protease using the ATPS method showed that the main and quadratic effects of NaCl and the interaction effects of NaCl and PEG molecular mass indicated the significant (p < 0.05) effects on partitioning of the serine protease (Table 1). In general, the interaction of hydrophilic polymers and enzymes can be modified by addition of NaCl in ATPS [20]. The variation of NaCl concentration causes an electrical potential difference between two phases and results in enzyme partitioning. Based on the results, the highest yield (97.3%) of serine protease was obtained with the addition of 4.5% (w·w⁻¹) of NaCl (Figure 1d). It showed that the hydrophobic interaction between PEG and hydrophobic surface of seine protease was significantly increased at this concentration of NaCl. However, the higher concentration of the salt had a negative effect on the partitioning of the enzyme because the unequal partitioning of natural salt between two phases affected the chemical potential of the solute. Figure 1d shows how the interaction effect of NaCl and PEG molecular mass significantly affects the yield of serine protease using ATPS. In addition, Figure 2 (Line 2) showed that the purity of purified serine protease after purification.

Table 2 shows the estimated regression coefficient, R² and adjusted R² as long as p-value of regression for each response variable (partition coefficient, purification factor and yield) is given. As shown in the table, the R² values are obtained for all response variables higher than 0.8 (0.993–0.997) and p-value of regression is less than 0.05. Therefore, according to a criterion for validation of model presented by Joglekar and May [21], the models are suitably and accurately employed for predicting the response variable as functional of independent variables. In addition, the comparison between experimental values and those predicted is shown in Figure 3a–c. It is seen there is overall closeness between experimental and predicted data, and based on Montgomery [22] the closeness between these variables indicates the adequacy of the regression equations.

Mango fruits were purchased from local market (Selangor, Malaysia) in slightly under ripe commercial maturity stage with the brix of 14. All chemicals and reagents used were analytical grade. Dextran T500 (average molecular weight of 500,000 g·mol⁻¹), polyethylene glycol (PEG) 4000, 6000, 8000, 10,000 and 12,000 (g·mol⁻¹), bovine serum albumin (BSA), azocasein and Bicinchoninic acid solution were supplied by Sigma Chemical Co, (St. Louis, UK). Trichloroacetic acid (TCA), 99%, di-sodium hydrogen anhydrous and sodium hydrogen phosphate monohydrate were purchased from Merck (Darmstadt, Germany).

Mango peel was obtained from 100 g fresh mangoes that had been washed with double distilled water. The peel was cut into small cubes (3 mm × 3 mm) and then blended with 50 mL sodium phosphate hydrogen at pH 7.5 by using a Waring commercial laboratory blender 32BL79 (Torrington, CT, USA) for 2 min at high speed at 4 °C. Crude enzyme extract was produced by putting the resultant homogenate through a cheesecloth filter and then centrifuging at 8000 g for 15 min at 4 °C [17].

Different phase systems were prepared for the selection of the suitable bottom phase and top phase [23] and then samples from each phase were taken and analyzed for the serine protease activity. A 50% (w·w⁻¹) polymer by mass was prepared as the concentration of PEG stock solutions. This was followed by the preparation of phase systems in 15 mL graduated centrifuge tubes by weighing an appropriate amount of “dextran T-500” and PEG with different molecular weights (4000–12,000 g·mol⁻¹), NaCl (−2.5–11.5% w·w⁻¹), pH (4.5–10.5) and 20% (w·w⁻¹) crude feedstock at room temperature. Then the appropriate amount of distilled water was added to the mixture to obtain a final mass of 10 g system. Thorough mixture of the content by gentle agitation for equilibration and then phase separation was achieved using centrifugation at 4000× g for 10 min. Then serine protease activity and protein concentration of top phase and bottom phase were determined [24,25].

RSM was used to determine the effects of four independent variables in ATPS, namely PEG molecular weight (4000–12,000 g·mol⁻¹, x₁), TLL (−3.42–35.27% (w·w⁻¹), x₂), NaCl (−2.5–11.5%, (w·w⁻¹), x₃) and pH (4.5–10.5, x₄) on partition coefficient (Y₁), purification factor (Y₂) and yield (Y₃) of the purified serine protease from mango peel. Central composite design (CCD) was the method employed in the assessment. The assessment of 30 supernatants on the basis of a CCD was done involving 16 factorial points, eight axial points (±α) and six center points (Table 3). There was a six-time repetition of the center point to determine the possibility of pure error. To determine the regression coefficient and statistical significance of the models, response surface analysis was carried out. An overall response surface model is as follows:

From the above equation, Y is response calculated by the model; β₀ is a constant; β_i, β_ii and β_ij are linear, squared and interaction coefficients, respectively. The estimated regression coefficients of the response surface models for the serine protease purification, along with the corresponding R² and adjusted R² values are presented in Table 2. The significant p-value of regression and R² with no less than 0.80 indicates a good model fit. As shown in the Table, the p-value of regressions is less than 0.05 and R² and adjusted R² values from 0.993 to 0.997 and 0.991 to 0.993 respectively, were obtained for all the response variables. Thus, the models of response surface were suitably and accurately used for predicting high variation percentage (≥80%) of the properties of purified serine protease as functional to the purification variables. Larger values of absolute t-value and smaller values of p-value indicate that the variables will be more significant (p < 0.05). Only the significant (p < 0.05) independent variable effects were included in the reduced model. The terms found to be statistically non-significant (p > 0.05) were removed and the experimental data were refitted to only the significant (p < 0.05) independent variable effects so as to achieve the final reduced model. It is worth noting that in a situation where some of the variables are not significant, they will still be kept in the final reduced model, as in the case of a quadratic or interaction term involving this significant (p < 0.05) variable [29,30]. The design matrix of the experiment, the analysis of the data and the optimizing procedure were all done using the Minitab v.14 statistical package (Minitab Inc., PA, USA).

Numerical and graphical optimizations were performed to determine the optimum levels of independent variables. Numerical optimization was performed to determine the exact optimum level of independent variables, using a response surface optimizer (Minitab v.14). Graphical optimization (using three-dimensional response surface plot) can be employed to explain the response models and to be able to better comprehend the implications of interaction between independent variables and response variable. In addition, a comparison was made between the experimental values and predict data to determine the adequacy of the response regression equations as indicated by Montgomery [22].