Journal of Food Engineering 78 (2007) 836–845

www.elsevier.com/locate/jfoodeng

Modeling and optimization I: Usability of response

surface methodology
_
Deniz Basß, Ismail H. Boyacı *

Department of Food Engineering, Faculty of Engineering, Hacettepe University, Beytepe Campus, TR-06532 Ankara, Turkey

Received 5 April 2005; accepted 17 November 2005

Available online 7 February 2006

Abstract

Response surface methodology (RSM) is the most popular optimization method used in recent years. There are so many works based
on the application of RSM in chemical and biochemical process. On the other hand, few articles were published about the limitation and
usability of it. In this paper, we looked at some of the RSM articles published during the last few years. We tried to identify common
mistakes made in the application and the limitations of RSM. We asked ourselves two important questions. These questions are ‘‘Can
RSM be used for optimization of all chemical and biochemical processes without any limitation?’’ and ‘‘Is RSM usable for other pur-
poses (determination of reaction kinetics, stability or evaluation of kinetic constants etc.) in addition to optimization?’’. We were able to
answer these questions based on the observations obtained from reviewed articles. We believe that the answers will be helpful for
researchers, who will use RSM in their future studies.
Ó 2005 Elsevier Ltd. All rights reserved.

Keywords: Response surface methodology; Biochemical process; Usability of RSM

1. Introduction processes in which a response of interest is influenced by

several variables and the objective is to optimize this
It is important to improve the performance of the sys- response. RSM has important application in the design,
tems and to increase the yield of the processes without development and formulation of new products, as well as
increasing the cost. The method used for this purpose is in the improvement of existing product design. It defines
called optimization. There is a parameter change in the gen- the effect of the independent variables, alone or in combina-
eral practice of determining the optimal operating condi- tion, on the processes. In addition to analyzing the effects of
tions while keeping the others at a constant level. This is the independent variables, this experimental methodology
called one-variable-at-a-time technique. The major disad- generates a mathematical model which describes the chem-
vantage of this technique is that it does not include interac- ical or biochemical processes (Anjum, Tasadduq, &
tive effects among the variables and, eventually, it does not Al-Sultan, 1997; Myers & Montgomery, 1995).
depict the complete effects of the parameters on the process. RSM has been very popular for optimization studies in
In order to overcome this problem, optimization studies can recent years. Some examples of the RSM applications per-
be carried out using response surface methodology (RSM). formed for optimization of biochemical process are hydro-
RSM is a collection of statistical and mathematical tech- lysis of pectic substrates, enzymatic synthesis of fatty
niques useful for developing, improving, and optimizing esters, lipase-catalyzed incorporation of docosahexanenoic
acid (DHA) into borage oil, alkaline protease production
from Bacillus mojavensis in a bioreactor, butylgalactoside
*
Corresponding author. Tel.: +90 312 297 71 00; fax: +90 312 299 21
synthesis by-galactosidase from Aspergillus oryzae,
23. biotransformation of 2-phenylethanol to phenylacetalde-
_
E-mail address: [email protected] (I.H. Boyacı). hyde in a two-phase fed-batch system, lipase catalyzed

0260-8774/$ - see front matter Ó 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jfoodeng.2005.11.024
 _
D. Basß, I.H. Boyacı / Journal of Food Engineering 78 (2007) 836–845 837

esterification reactions, pectinase usage in pretreatment of It is possible to separate an optimization study using
mosambi juice for clarification, cholesterol oxidase produc- RSM into three stages. The first stage is the preliminary
tion by Rhodococcusequi no. 23, phytase production by work in which the determination of the independent
Pichia anomala and determination of reaction parameters parameters and their levels are carried out. The second
for damaged starch assay (Beg, Sahai, & Gupta, 2003; stage is the selection of the experimental design and the
Boyacı, Williams, & Köksel, 2004; Çelik, Bayraktar, & prediction and verification of the model equation. The last
Mehmetoğlu, 2004; Ismail, Linder, & Ghoul, 1999; Lee, one is obtaining the response surface plot and contour plot
Chen, & Chou, 1998; Manohar & Divakar, 2004; Panda of the response as a function of the independent parameters
& Naidu, 1999; Rai, Majumbar, DasGupta, & De, 2004; and determination of optimum points. Details about each
Senanayake & Shahidi, 2002; Vohra & Satyanarayana, stage are given below.
2002). RSM is used not only for optimization but also
for determination of kinetic constants and investigation 2.1. Preliminary work: determination of independent
of enzyme stability and kinetics. Beg, Saxena, and Gupta variables and their levels
(2002) used RSM for determination of kinetic constants
for an alkaline protease from Bacillus mojavensis. Stability Chemical and biochemical processes are affected by
and kinetics of b-1,3-glucanase from Trichoderma harzia- numerous parameters. Because it is not possible to identify
num were investigated using RSM. Also simple enzyme the effects of all parameters, it is necessary to select the
kinetics of alcohol dehydrogenase was evaluated using parameters that have major effects. Screening experiments
RSM (Rana, Theodore, Naidu, & Panda, 2003). are useful to identify the independent parameters. Factorial
Nowadays, it is clearly seen that RSM has been widely designs may be used for this purpose. After the identifica-
applicable for different purposes in chemical and biochem- tion of the important parameters, the direction in which
ical processes. Although RSM has limitations, some studies improvements lie is determined and the levels of the param-
have been carried out without considering these limita- eters are identified. Determination of these levels is impor-
tions. In this article, we reviewed some of the RSM studies tant because the success of optimization process directly
published recently and tried to answer two important ques- relates to these levels. Mistakenly chosen levels result in
tions about usability of RSM. First, ‘‘can RSM be used for an unsuccessful optimization.
optimization of all chemical and biochemical processes The units of the independent variables differ from one
without any limitation?’’ and second, ‘‘is RSM usable for another. Even if some of the parameters have the same
other purposes (determination of reaction kinetics, stability units, not all of these parameters will be tested over the
or evaluation of kinetic constants etc.) in addition to opti- same range. Since parameters have different units and/or
mization?’’. These questions were answered at the end of ranges in the experimental domain, the regression analysis
the article based on the observations obtained from should not be performed. Instead, one must first normalize
reviewed articles. the parameters before performing a regression analysis.
Each of the coded variables is forced to range from 1
2. Theory to 1, so that they all affect the response more evenly, and
so the units of the parameters are irrelevant. Commonly
RSM consists of a group of mathematical and statistical used equation for coding is seen below:
techniques that can be used to define the relationships
between the response and the independent variables. x  ½xmax þ xmin =2
X ¼ ð2Þ
RSM defines the effect of the independent variables, alone ½xmax  xmin =2
or in combination, on the processes. In addition to analyz-
where x is the natural variable, X is the coded variable and
ing the effects of the independent variables, this experimen-
xmax and xmin are the maximum and minimum values of the
tal methodology also generates a mathematical model. The
natural variable.
graphical perspective of the mathematical model has led to
the term Response Surface Methodology.
The relationship between the response and the input is 2.2. Selection of the experimental design, and prediction
given in Eq. (1): and verification of model equation
g ¼ f ðx1 ; x2 ; . . . ; xn Þ þ e ð1Þ Ten important properties that should be taken into
where g is the response, f is the unknown function of re- account in choosing a response surface design is given by
sponse, x1, x2, . . . , xn denote the independent variables, also Myers and Montgomery (1995). Some computer packages
called natural variables, n is the number of the independent offer optimal designs based on the special criteria and input
variables and finally e is the statistical error that represents from the user. These designs differ from one other with
other sources of variability not accounted for by f. These respect to their selection of experimental points, number
sources include the effects such as the measurement error. of runs and blocks.
It is generally assumed that e has a normal distribution After selection of the design, the model equation is
with mean zero and variance. defined and coefficients of the model equation are
838 _
D. Basß, I.H. Boyacı / Journal of Food Engineering 78 (2007) 836–845

predicted. The model used in RSM is generally a full qua- Note that C is a square matrix (Penny & Lindfield, 1999).
dratic equation or the diminished form of this equation. After the regression coefficients were obtained, the esti-
The second order model can be written as follows: mated response could be easily calculated using model
X
k X
k XX equation. Usually the behavior of the system is unknown
y ¼ b0 þ bj X j þ bjj X 2j þ bij X i X j ð3Þ so one must check whether the model fits well to the exper-
j¼1 j¼1 i<j imental data. For verification of the model adequacy, sev-
where b0, bi, bii and bij are regression coefficients for inter- eral techniques are carried out. Some of these techniques
cept, linear, quadratic and interaction coefficients respec- are residual analysis, scaling residuals, prediction error
tively and Xi and Xj are coded independent variables. sum of squares (PRESS) residuals, and testing of the lack
The matrix notation of the model is given in Eq. (4): of fit. The overall predictive capability of the model is com-
monly explained by the coefficient of determination (R2),
y ¼ Xb þ e which is calculated from PRESS. On the other hand, one
2 3 2 3 2 3 2 3
y1 1 x11 x12   x1k b0 e1 must not forget that the R2, alone, is not a measure of
6y 7 61 x x22   x2k 7 6b 7 6e 7 the model’s accuracy. R2 is a measure of the amount of
6 27 6 21 7 6 17 6 27
6 7 6 7 6 7 6 7 the reduction in the variability of response obtained by
6  7 6      7 6  7 6  7
6 7¼6 7 6 7þ6 7 ð4Þ using the repressor variables in the model. However, a
6  7 6      7 6  7 6  7
6 7 6 7 6 7 6 7 large value of R2 does not necessarily imply that the regres-
6 7 6 7 6 7 6 7
4  5 4      5 4  5 4  5 sion model is a good one. Adding a variable to the model
yn 1 xn1 xn2   xnk bk en will always increase R2, regardless of whether the addi-
|fflffl{zfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl
ffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflffl{zfflffl} |fflffl{zfflffl} tional variable is statistically significant or not. Thus, it is
y X b e
possible for models that have large values of R2 to yield
The system of equations given above is solved using the poor predictions of new observations or estimates of the
method of least squares (MLS). Method of least squares is mean response (Myers & Montgomery, 1995). If you plot
a multiple regression technique and this method can be experimental results versus model results, you should
summarized as follows. obtain a straight line passing the origin with the angle of
In MLS, it is assumed that random errors are identically 45 °C [y = x]. But in some cases you might obtain such a
distributed with a zero mean and a common unknown var- line [y = a + bx]. This line would also have a regression
iance and they are independent of each other. The differ- analysis with R2 = 1.000 however it is totally insignificant.
ence between the observed and the fitted value (^y ) for the These types of errors can be eliminated by using absolute
ith observation ei ¼ y i  ^y i is called the residual and is an average deviation (AAD) analysis, which is a direct method
estimate of the corresponding ei. Our criterion for choosing for describing the deviations.
the bj estimates is that they should minimize the sum of the The AAD is calculated by the following equation:
squares of the residuals, which is often called the sum of (" #, )
squares of the errors and is denoted by SSE. Thus, X p
AAD ¼ ðjy i; exp  y i;cal j=y i; exp Þ p  100 ð12Þ
X n X
SSE ¼ e2i ¼ ðy i  ^y i Þ2 ð5Þ i¼1

i¼1 where yi,exp and yi,cal are the experimental and calculated
The residuals may be written as responses, respectively, and p is the number of experimen-
e ¼ y  Xb ð6Þ tal run. Evaluation of R2 and AAD values together should
be better to check the accuracy of the model. R2 must be
and the SSE is close to 1.0 and the AAD between the predicted and ob-
T
SSE ¼ eT e ¼ ðy  X bÞ ðy  X bÞ ð7Þ served data must be as small as possible. The acceptable
Differentiating the SSE with respect to b, we get a vector of values of R2 and AAD values mean that the model equa-
partial derivatives, as follows: tion defines the true behavior of the system and it can be
used for interpolation in the experimental domain. One
o
ðSSEÞ ¼ 2X T ðy  X bÞ ð8Þ must be careful about extrapolating beyond the region con-
ob taining the original observations. It is quite possible that a
Equating this derivative to zero we have Xb = y and this model that fits well in the region of the original data will no
over-determined system of equations could be solved di- longer fit well outside that region.
rectly to obtain the coefficients of b by the following:
X TX b ¼ X Ty ð9Þ 2.3. Graphical presentation of the model equation and
determination of optimal operating conditions
The formal solution of these equations is then
1
b ¼ ðX T X Þ X T y ¼ CX T y ð10Þ The visualization of the predicted model equation can be
where obtained by the response surface plot and contour plot.
1
The response surface plot is the theoretical three-dimen-
C ¼ ðX T X Þ ð11Þ sional plot showing the relationship between the response
 _
D. Basß, I.H. Boyacı / Journal of Food Engineering 78 (2007) 836–845 839

and the independent variables. The two-dimensional dis- empirical model that related the response to the indepen-
play of the surface plot is called contour plot and in the dent variables is used to obtain information about the pro-
contour plot, lines of constant response are drawn in the cess. With respect to these, we can say that RSM is a useful
plane of the independent variables. The contour plots help tool for the optimization of chemical and biochemical
to visualize the shape of a response surface. When the con- process.
tour plot displays ellipses or circles, the center of the system On the other hand, the major drawback of RSM is to fit
refers to a point of maximum or minimum response. Some- the data to a second order polynomial. We cannot say that
times, contour plot may display hyperbolic or parabolic all systems containing curvature are well accommodated by
system of the contours. In this case, the stationary point the second order polynomial. For example, simple enzyme
is called a saddle point and it is neither a maximum nor kinetics is defined by the Michaelis–Menten equation. In
a minimum point. These plots give useful information fact, Michaelis–Menten equation defines a curve which is
about the model fitted but they may not represent the true a rectangular hyperbola through the origin, with asymp-
behavior of the system. One must not forget that the con- totes S = Km and v = V (Cornish-Bowden, 1999). More-
tours (or surfaces) represent contours of estimated over, the effect of temperature on biochemical processes
response and the general nature of the system arises as a is described by symmetrical or nonsymmetrical bell shaped
result of a fitted model, not the true structure (Myers & curves. To explain the effect of temperature with a second
Montgomery, 1995). order polynomial is not possible, especially when the curve
The stationary point (minimum or maximum point) of a is nonsymmetrical. To overcome this, the data can be con-
second order equation is the point where the first derivative verted into another form that can be explained by the
of the function equals to zero: second order model. For example, logarithmic transforma-
tions and other linearization methods can be used for this
Let y ¼ f ðx1 ; x2 Þ and
purpose. An example of logarithmic transformation of
¼ b0 þ b1 x1 þ b2 x2 þ b11 x21 þ b22 x22 þ b12 x1 x2 ð13Þ response and independent parameter are given in the fol-
lowing equation:
The stationary point is found by computing oy/ox1 and
oy/ox2 and setting zero: ln y ¼ b0 þ b1 x1 þ b2 x2 þ b12 x1 x2 ð15Þ
oy y ¼ b0 þ b1 ðln x1 Þ þ b2 ðln x2 Þ ð16Þ
¼ b1 þ 2b11 x1 þ b12 x2 ¼ 0
ox1
ð14Þ Although these transformations may be useful, it is not
oy
¼ b2 þ 2b22 x2 þ b12 x1 ¼ 0 possible to say that these transformations give desirable
ox2 results for all systems. In addition, the transformation of
The system of equations is solved to find the values of x1 response or inputs is time consuming and sometimes it is
and x2. The calculated values of x1 and x2 indicate the difficult to know which form of the transformation is the
coded value of the independent parameters that give the best.
highest or lowest response. For optimization of the pro- Alternatively, if the system is hardly explained by a sec-
cess, the stationary point has to be determined in the tested ond order model, one should choose a smaller range of
independent parameters’ ranges. One exception to this, in independent parameters. It is possible to increase the accu-
some processes only the effects of the parameters are inves- racy of the model equation by working in a narrow range
tigated without optimization or due to the physical rea- of independent parameter but it should be remembered
sons, therefore, it is not possible to obtain the stationary that working in a narrow range reduce the possibility of
point. determination of the stationary point. Preliminary work
becomes more critical for the determination of the indepen-
2.4. Advantages and disadvantages of RSM dent parameter range.

RSM has several advantages compared to the classical 3. Review of published studies
experimental or optimization methods in which one vari-
able at a time technique is used. Firstly, RSM offers a large In this section, we reviewed some RSM studies pub-
amount of information from a small number of experi- lished in the last few years. The reviewed works were clas-
ments. Indeed, classical methods are time consuming and sified in two groups as RSM application for chemical and
a large number of experiments are needed to explain the biochemical process optimization and other applications
behavior of a system. Secondly, in RSM it is possible to of RSM.
observe the interaction effect of the independent parame-
ters on the response. Especially in biochemical processes, 3.1. RSM application for chemical and biochemical
the interaction effect of the parameters would be more crit- process optimization
ical such as synergism, antagonism, and addition. The
model equation easily clarifies these effects for binary com- RSM has been used for optimization of alkaline prote-
bination of the independent parameters. In addition, the ase production from Bacillus mojavensis (Beg et al.,
840 _
D. Basß, I.H. Boyacı / Journal of Food Engineering 78 (2007) 836–845

2003). The effects of casamino acids concentration, glucose the experimental data, there is no negative effect of amount
concentration, inoculum age, incubation time, and agita- of enzyme and reaction temperature on the response. How-
tion rate on response have been investigated. Firstly, the ever, we see negative effects of these independent parame-
effect of the independent parameters on protease produc- ters on response in the three-dimensional response
tion in shake flask cultures has been determined using surface plot and two-dimensional contour plot illustrations
RSM and then optimum parameters have been used in a of predicted model equation. Due to the model equation
bioreactor. A face center design with five independent vari- estimate of incorrect data at high values of enzyme amount
ables has been created. A second order polynomial has and temperature, it is hardly possible to say that the pre-
been fitted to the experimental data, the regression coeffi- dicted model equation is useable for the optimization of
cients have been obtained, and R2 is reported as 0.9473. the process. Also, in the experimental design section of
We calculated AAD value as 29.7%. It is reported that the article it is said that the logarithmic transformation
the coordinates of three factors (casamino acids, inoculum has been applied to the data before fitting to the second
age and agitation) are positive, whereas negative coordi- order polynomial equation, but we did not observe any
nates have been obtained as two factors (glucose and incu- transformation on the prediction of the model equation.
bation time). The negative effect of incubation time on Performance of pectolytic enzymes during hydrolysis of
response implies that lower incubation time cause higher pectic substrate under assay conditions were investigated
protease production. It is not possible for a biochemical using RSM (Panda & Naidu, 1999). The physical (temper-
process. Based on the data given in this paper it could be ature and pH) and chemical (volume of substrate and
said that incubation time has a negative effect on response enzyme solution) parameters have been optimized using
in the investigated incubation time range. In the second central composite design. In addition to the using substrate
part of the study, in bioreactor, it has been observed that and enzyme concentration as the independent parameters,
incubation time has a positive effect on protease produc- volumes of substrate and enzyme solutions were used.
tion between the sixth and 12th hour, and after 12 h, the Using the volume as independent parameter is a disadvan-
response is adversely affected. The range of casamino acids tage because a change in the volume of the substrate or
concentration in experimental design was 2–8 mg/ml but in enzyme causes a change in the total reaction volume and
the validation of model equation, 8 and 12 mg/ml of casa- a change in the total reaction volume changes both sub-
mino acids concentrations were used. Using of 12 mg/ml strate and enzyme concentrations. It is not possible to
may not be a correct approach for validation of the equa- obtain individual effect of parameters on response using
tion because they obtained the descriptive equation with this approach. Activities of three pectolytic enzymes
the data in low concentrations of the substrate. (Extrapola- (polymethylgalacturonase, pectinlyase and polygalacturo-
tion is needed for the estimation of response in this concen- nase) were investigated with four independent parameters.
tration.) In addition, a high yield at this value does not It is also possible to design the experiment with four inde-
mean that it is the optimum value of casamino acids con- pendent parameters in combination but researchers pre-
centration for protease production. ferred two experiment designs for physical and chemical
Senanayake and Shahidi (2002) worked on lipase cata- parameters, separately. The average of R2 and AAD of
lyzed biochemical reaction, incorporation of docosahexae- six predicted model equations were calculated using the
noic acid (DHA) into borage oil using RSM. They data given in the paper as 0.897 and 14.9% respectively.
evaluated the effects of three independent parameters, the It is possible to determine the stationary point using
amount of enzyme, reaction temperature, and reaction time RSM but it does not mean that all stationary points are
on the yield of DHA incorporation. This study is a good optimum for enzymatic reaction. For optimization of bet-
model for determination of the range of the independent ter hydrolysis of substrate, the stationary point should be
parameters. In the preliminary work, the effect of the inde- the maximum and the values of the independent parame-
pendent parameters on response was investigated by vary- ters at this stationary point are chosen as optimum operat-
ing one factor at a time while keeping the others constant. ing condition. Although the maximum response for pH
Based on this preliminary work, a three factor and three and temperature were chosen to be optimum for pectinly-
level face centered cube design has been adapted. Second ase and polygalacturonase reactions, the level of substrate
order full quadratic model was fitted and R2 was deter- and enzyme volumes where these enzyme activities are min-
mined as 0.94. AAD value was calculated as 3.1%. The imum were determined as optimum level of independent
relation between the dependent parameter and the indepen- parameters. It is confusing that the selection of one station-
dent parameters was explained using a second order poly- ary point was made at maximum and the other at mini-
nomial equation. As it is seen from the plots in the mum for optimization of the same response.
article representing the effects of the amount of enzyme Vohra and Satyanarayana (2002) worked on statistical
and incubation temperature on DHA incorporation, optimization of the medium components by RSM to
DHA incorporation increases only up to a certain extent enhance phytase production by Pichia anomala. The indi-
with an increase in amount of enzyme and reaction temper- vidual and interaction effects of glucose concentration, beef
ature, and then the response does not change when the extract concentration and inoculum density on phytase and
amount of enzyme and reaction temperature increase. In biomass productions were investigated using central
 _
D. Basß, I.H. Boyacı / Journal of Food Engineering 78 (2007) 836–845 841

composite design. In addition, the effects of glucose and using the experimental design and observed values of the
beef extract concentration were optimized by one variable dependent variables. Although the article includes a well-
at a time method and the results of the two optimization organized work and a high quality description of the study,
methods were compared with each other. In the five level there is a contradiction between the experimental data and
(a  1, 0, +1 and +a) central composite design, the alpha the graph of Brix increase. As can be seen from the table
value depends on the number of the independent variables which is given in the paper, representing the observed val-
and is calculated by 2k/4, where k is the number of indepen- ues of response variables, the response, °Brix increase, has
dent variables. For example, in a design with three inde- been observed as 17.1876 °Brix/g at coded temperature: 1
pendent parameters, the alpha is 1.682. The alpha value (50 °C), coded concentration: 1 (45 °Brix) and coded
must be the same for all independent parameters in a time: 1 (60 min). °Brix increase is 7.7658 °Brix/g at coded
design. In this study, different alpha values have been used temperature: +1 (50 °C), coded concentration: 1
for each independent parameter. Alpha values for glucose (45 °Brix) and coded time: +1 (120 min). Moreover, the
concentration, beef extract concentration, and inoculum response measured at the center point is 13.11 °Brix/g.
size are 1.25, 1.53, and 2.5, respectively. The reason why These results indicate that an increase in temperature and
they applied three different alpha values was not mentioned time causes a decrease in °Brix increase at minimum con-
in the text. Different alpha values totally change the archi- centration (coded: 1; uncoded: 45 °Brix). However, from
tecture of the experimental design and this may cause unex- the figure representing the response surface and contour
pected results from the statistical analysis of the data. plots for °Brix increase, it is clearly seen that increase in
Model equations for phytase and biomass productions time and temperature causes an increase in response. The
have been predicted using the experimental design and reason may be that the contours (or surfaces) represent
responses. The R2 and AAD values have been determined contours of estimated response and the general nature of
as 0.985, 0.925 and 5.6%, 4.4% for phytase and biomass the system arises as a result of a fitted model, not the true
productions, respectively. The highest biomass production structure (Myers & Montgomery, 1995).
is obtained at 5.5% glucose concentration, 2% beef extract The synthesis of isopropyl palmitate has been developed
concentration and 1.5% inoculums density. It has been and optimized using RSM (Aracil, Garcia, Sanchez, &
reported that inoculum density does not have a significant Matinez, 1999). The effects of the independent parameters
effect on the responses. Another result obtained from the (temperature and catalyst concentration) on acid conver-
experimental study is that phytase production is not sion have been investigated at different reaction times (30,
directly related to biomass production. The maximum phy- 60, 90 and 120 min) using a full factorial design with five
tase production is obtained at the lowest glucose and beef levels of independent parameters. Coefficients of linear
extract concentration as 140 U/g dry weight. There is no and nonlinear model have been calculated and all terms
information about the behavior of phytase production of both models have been found significant. Then four-sec-
below these concentrations. It may be possible to receive ond order full quadratic model equations have been pre-
higher production below these concentrations. Similarly, dicted for the coded and actual values and coefficient of
the highest phytase production is obtained at lowest glu- determination values of predicted models are reported as
cose, and beef extract concentrations in one variable at a 0.995, 0.988, 0.996 and 0.997. AAD values were calculated
time method. Phytase production at lower glucose and beef as 2.22%, 1.78%, 1.68% and 1.74%. These values indicate
extract concentration has not been investigated. A compar- that the model equations are adequate to describe the
ison has been drawn between the results of the two meth- experimental designs. On the other hand, stationary point
ods (RSM and one variable at a time method). It is not (maximum value) has not been observed in this work. It
possible to compare these results without obtaining full is reported that the maximum palmitic acid conversion
optimization graphs and without working in the same inde- (87%) is achieved at the maximum temperature (70 °C)
pendent parameter range. and maximum initial catalyst concentration (7 wt.%)
Osmotic dehydration of cantaloupe has been optimized within the range studied. For that reason, it is not possible
using desired function methodology by Corzo and Gomez to say that this optimization process is completed. A fur-
(2004). The effects of three independent parameters (tem- ther study is needed.
perature, concentration, and time) on three responses Lee et al. (1998) optimized production of cholesterol
(mass loss, water loss, and Brix increase) have been inves- oxidase by Rhodococcusequi no. 23 by using RSM. Choles-
tigated by means of five levels, central composite design. terol concentration (g/l), yeast extract (g/l) and Tween 80
The biggest difference of this work from the other RSM (ml/l) have been chosen as the independent parameters.
studies is the calculation of an individual desirability func- Central composite design has been applied and a second
tion that varies from 0 to 1 (the highest score is 1 and the order polynomial model equation has been fitted to the
lowest is 0) for each response. For all the desired functions, experimental data. Coefficient of determination value is
a total desired function has been defined. A high value of given as 0.928. We calculated AAD as 2.04%. Optimal con-
the total desired function is called as the best function of ditions were found to be 2.30 g/l cholesterol, 8.18 g/l yeast
the system, which is considered as the optimum solutions extract and 4.10 ml/l Tween 80 and the maximum response
of the system. Three model equations have been predicted were estimated using predicted model equation as
842 _
D. Basß, I.H. Boyacı / Journal of Food Engineering 78 (2007) 836–845

0.251 unit/ml. Optimum values of the independent param- have fitted a second order polynomial to responses, xylitol
eters have been determined in the investigated range of crystallization yield and purity degree. After neglecting the
parameters. In the figure of contours which displays choles- insignificant terms, a linear equation for xylitol crystalliza-
terol oxidase activity as a function of cholesterol concen- tion yield and a second order equation not including the
tration and yeast extract, there is a negative effect of interaction term for the purity degree (PD) was obtained.
cholesterol concentration above 2.3 g/l. The same effect is The R2 and AAD values were determined as 0.9604,
seen in the figure of contours of cholesterol oxidase activity 0.8763 and 4.49%, 1.80% for xylitol crystallization yield
as a function of cholesterol concentration and Tween 80. and purity degree, respectively. They compared the con-
Two reasons may cause this behavior. First reason is that tour plots of the two responses together and obtained
the cholesterol concentration may have an inhibitory effect two regions. In the first region, a PD of 0.97 was obtained
on cholesterol oxidase enzyme at high cholesterol concen- with a yield of 0.54 and an initial xylitol concentration of
tration. The second one is that the model equation may 730 g/l. On the other hand, in the second one a PD of
not be adequate to estimate the response at a high choles- 0.99 was obtained with a yield of 0.25 and an initial xylitol
terol concentration. The average of six responses at center concentration of 582 g/l. The cooling temperature for the
values of the independent parameters is 0.242 ± first region was 6 °C and 2.4 °C for the second. A
0.005 units/ml. The response at high cholesterol concentra- higher PD that enables to achieve near total purity crystals
tion (yeast extract: 8 g/l (coded: 0) and Tween 80: 4.0 ml/l and a cooling temperature near zero which lowers the cost
(coded: 0)) was measured as 0.240 units/ml. Based on the are desirable, therefore, that the second region was selected
comparison of these responses, it is difficult to say that to be the optimal.
there is an inhibitory effect of cholesterol concentration
above the 2.3 g/l. 3.2. Other applications of RSM
Nawani and Kapadnis (2005) optimized the medium of
components and environmental factors for chitinase pro- Gobbetti et al. (1999) investigated the effects of temper-
duction. They performed the study on strains of Streptomy- ature, pH and NaCl on the peptidase activities of non-
ces sp. NK1057, NK528 and NK951, which produce starter lactic acid bacteria (NSLAB) using RSM. The
chitinase. For identification of the effect of each variable, effects of independent parameters were investigated in the
a 2-level fractional factorial design with eight independent ranges of parameters valid in the ripening of cheeses. In
parameters was applied. After the identification of the this study, the aim was not to optimize the parameters,
parameters for each strain, a central composite design only the effects of them on responses were investigated.
(CCD) with four parameters for NK1057 and a CCD with For this reason, determination of minimum or maximum
three parameters for both Streptomyces sp. NK528 and response was not needed. The individual and interaction
Streptomyces sp. NK951 were applied. The R2 values were effects of temperature, pH and NaCl were determined suc-
reported and AAD values were calculated as 0.9669 and cessfully for three enzymes (aminopeptitase-type N, amino-
2.55% for NK1057, 0.899 and 5.80% for Streptomyces sp. peptitase-type A and imminopeptidase) of 11 strains of
NK528 and 0.881 and 8.82% for Streptomyces sp. NK951. NSLAB. A 5-level central composite design was performed
Optimum parameters were determined using predicted for each of the three responses, full quadratic model equa-
model equation and then validation of the model equations tions were predicted and then the terms which were not sig-
was done with the responses obtained experimentally. The nificant (p > 0.05) were eliminated in the final model.
errors between predicted and observed responses were cal- Coefficients of determination of the model equations varied
culated to be 5.1%, 10.6% and 2.9% for the strains between 0.846 and 0.994. We could not calculate the AAD
Streptomyces sp. NK1057, NK528 and NK951, respec- values because the experimental results were not given in
tively. Finally, it was reported that the use of RSM the text. Similar to other RSM studies, an acceptation of
increased the production by 29.0% for NK1057, 9.3% for all relations between the independent parameters and the
NK528 and 28.0% for NK951. It is possible to say that this response could be explained with second order polynomial
study is a successful application of RSM. On the other equation, which was valid in this work. It is better to verify
hand, optimum value of ammonium sulphate concentration this acceptation with a preliminary work or testing the
for Streptomyces sp. NK1057 (0.04 g/l) and for Streptomy- results of predicted model equation with the experimental
ces sp. NK528 (0.05 g/l) and optimum temperature for results.
Streptomyces sp. NK951 (41 °C) were determined out of Andersson and Adlercreutz (1999) evaluated the simple
the investigated range of the independent parameters. It enzyme kinetics using RSM. The two parameters (pH and
should be better to determine the optimum values of the cyclohexanone concentration) with central composite cir-
independent parameters falling in the investigated range. cumscribed design (CCCD) were performed with a total
Faveri, Torre, Perego, and Converti (2004) studied com- number of 11 experiments. A second order equation was
bined effects of initial xylitol supersaturation value and obtained using coded values of pH and coded-logarithmic
cooling temperature on xylitol crystallization from syn- transformation of substrate concentration and actual value
thetic solutions using RSM. They applied a 32 full-factorial of reaction rate. Concentration of the substrate in the
design and optimized the operational conditions. They design was log transformed in accordance with Eq. (17):
 _
D. Basß, I.H. Boyacı / Journal of Food Engineering 78 (2007) 836–845 843
 
x  c1 concentration and temperature. The yield of protease
log10 ðyÞ ¼ log10 ð17Þ
c2  x enzyme has been taken as the dependent parameter.
At this point, it should be useful to remember enzyme
where x substrate concentration (mM), c1 = 0.5 and
kinetics. It is well known that simple enzyme kinetics obey
c2 = 200. In a previous study, different transformations of
the Michaelis–Menten kinetics and is explained by a non-
the substrate concentration were evaluated and the best re-
linear equation (Eq. (10)):
sult was obtained with Eq. (17). But it is not correct to
name Eq. (17) as log transformation. Because each side V S
m¼ ð20Þ
of the equation contains log term and it is possible to can- Km þ S
cel both the log term and then Eq. (18) is obtained:
where m is the reaction rate, S is the substrate concentra-
x  c1
y¼ ð18Þ tion, Km Michaelis–Menten constant and V is the limiting
c2  x reaction rate (Cornish-Bowden, 1999). Depending on the
After the investigation of the coded values and the logarith- presence of inhibitor in reaction medium or sensitivity of
mic transformation of the right side of Eq. (18), it was ob- the enzyme to the substrate or the product, different equa-
served that Eq. (19) was used as the transformation tions are predicted to explain the physical situation
equation: (Ladero, Santos, Garcia, & Garcia-Ochoa, 2001). Eq. (11)
  is the linearized form of the Michaelis–Menten equation,
x  c1
y ¼ log10 ð19Þ and the plot of 1/m versus 1/S is called as the Lineweaver–
c2  x
Burk plot:
where y is the transformed form of the substrate concentra- 1 Km 1 1
tion. We also calculated the coefficients of the model equa- ¼  þ ð21Þ
m V S V
tion using the data given in the article. The coefficients of
the model equation obtained and the ones given in the pa- Lineweaver–Burk plots work best when substrate con-
per as well as their R2 and AAD values are seen in Table 1. centration covers the range of 0.25–5Km. Within this range,
Also the given model equation was used to predict the good linearity is observed. The greatest percent error is
reaction rate. Unfortunately, there is a big difference likely to be associated with the velocity at low substrate
between the experimental values and the predicted one. concentration. Unfortunately, in the reciprocal plot, the
The AAD value was calculated as 84.6%. It indicates that lowest values of [S] correspond to the highest values
the model equation is not adequate to explain the experi- of [1/S] and because of the linear regression, these data
mental design. The model obtained from RSM has been points are weighted more heavily in the analysis (Copeland,
used to evaluate the kinetic constants. The obtained model 2000).
equation has been fitted to the Michaelis Menten equation Beg et al. (2002) fitted the experimental data to a second
using the fixed pH values and kinetic constants Vmax and order polynomial equation beside of the Michaelis–Menten
Km have been evaluated. Nonlinear curve fitting is one of equation. Coefficient of determination value was given as
the best approaches for the estimation of kinetic constants 0.9278 and AAD was calculated as 16.55%. For determina-
of Michaelis–Menten equation. However, the inadequate tion of the kinetic constants, enzymatic reaction rate was
model equation obtained from RSM decreases success of estimated using the model equation at 20 different reaction
the work for determination of kinetic constants. conditions (range of casein concentration: 0.2–1.0 mg/ml
The kinetic constants of an alkaline protease from and range of reaction temperature: 45 and 60 °C). Experi-
Bacillus mojavensis has been determined using RSM (Beg mental reaction rates were also obtained at these reaction
et al., 2002). The independent parameters are casein conditions. There is no comparison between estimated
and experimentally obtained data in the paper. We com-
pared them with each other at a constant temperature
and the R2 and AAD values were determined as 0.681,
Table 1
Regression coefficients of the model equation calculated and the ones 0.808, 0.778 and 0.694 and 22.3%, 3.5%, 2.8% and 5.7%
given in the article for 45, 50, 55 and 60 °C, respectively. Kinetic constants
Term Regression coefficients were determined using the estimated and experimentally
obtained data and then the constants were compared to
Calculated using Calculated using the log Given in
actual value transformation of the variable the paper each other. Although the Km values determined are within
the range of 0.0250–0.0371 mg/ml, the casein concentra-
b0 3.05 0.48 0.35
b1 1.97 0.48 0.63 tions were investigated within the range of 0.2–1.0 mg/ml.
b2 1.01 0.32 0.38 It should have been investigated in the range of 0.25–
b11 0.17 0.16 0.14 5Km (0.006–0.180 mg/ml). In addition, it is clearly seen in
b22 0.83 0.27 0.32 the response surface and contour plots given in the article
b12 0.85 0.10 0.03
that there is an inhibition effect of substrate concentration
R2 0.97 0.94 0.96
on the reaction rate. For that reason, it is not possible to
AAD % 39.4 22.9 84.6 use simple enzyme kinetics to explain this behavior.
844 _
D. Basß, I.H. Boyacı / Journal of Food Engineering 78 (2007) 836–845

Manohar and Divakar (2004) summarized their study only reviewed a limited number of them. We observed a
about the applications of surface plots and statistical designs number of mistakes in some of the reviewed articles and
to selected lipase catalyzed esterification reactions. Studies most of these mistakes were similar. They were discussed
including Box–Behnken, Central Composite and Placket– separately below.
Burman designs were presented. In addition, thermal stability In most of the RSM studies reported, there is not
of lipase was investigated. Although designs given in this enough preliminary work about the range of independent
paper are the most common types of RSM designs used for parameters. In most of the optimization studies, there is
optimization of chemical and biochemical process, the R2 val- no optimum point due to unsuitable range of independent
ues were reported as 0.565, 0.74, 0.718, 0.743 and 0.89 for the parameters. The maximum (or minimum) value of the
reaction of 2-O-stearyl lactate, m-cresyl acetate, 2-O-lauroyl response without stationary point has been given as the
lactate and a-terpinyl propionate, respectively. In accordance optimum point. It is not correct to call these types of stud-
with these, we can say that second order equation hardly ies as optimization. If RSM studies were carried out with-
explains the actual behavior of the lipase catalyzed reactions. out determination of stationary point, the aim of the study
High pressure inactivation kinetics of amylase in apple and selection of the range of independent parameters
juice was investigated using RSM (Riahi & Ramaswamy, should be explained clearly. The verification of the pre-
2004). Five levels of three independent parameters (pres- dicted equation should be made. The physical mean of
sure, temperature and time) were used in the experimental the estimated results should be discussed and suitability
design. The effect of these parameters on inactivation of of the estimated response to the expected result should be
amylase at three different pH value (3.0, 3.5 and 4.0) was evaluated.
tested and three model equations were predicted. It was The other mistake, which may be the biggest one, is fit-
reported that the model equations were highly significant ting the all-experimental data to the second order polyno-
(p < 0.0001) with R2 of 0.99 and coefficient of variations mial equation. RSM is based on the use of a second
of 5.00%, 3.42% and 3.78% at pH 3.0, 3.5 and 4.0, respec- order equation. Actually, it is possible to predict a model
tively. We supposed there is a mistake in model equations. equation with a higher degree than the second order. But
The model equations were based on coded values. If the most of the packed program produced for the application
model equations were used for calculation, the responses of RSM use second order model equation. By using math-
(inactivation, %) at center points (X1u = 0, X2u = 0 and ematically packed programs, we can analyze the experi-
X3u = 0) should be 2.44%, 1.78% and 2.56% for pH 3.0, mental data using an equation higher than the second
3.5 and 4.0, respectively. (Because all terms in the model order. At this point, care must be taken in choosing the
equations except b0 are equal zero at center point for coded order of the polynomial. If the order is low, the polynomial
values.) However, its measured values (average) are 58.3%, fit will not capture the trends in the data. If too large an
40.3% and 30.3% for pH 3.0, 3.5 and 4.0, respectively. order is chosen, the fitted curve can have peaks and valleys
There is no relation between these values. as it goes through most of the data points, thereby produc-
Rana et al. (2003) studied the stability and kinetics of ing significant error when the derivatives, dy/dx, are gener-
b-1,3-glucanase from Trichoderma harzianum. A central ated at the data (Fogler, 1998). Actually using a second
composite design with two independent variables, pH and order equation in the optimization study is an easy
temperature, was applied. The estimated responses were approach because second order equation has only one sta-
compared with the experimental values. R2 has been tionary point. If a packed program, which use second order
reported as 0.6899 and AAD value was calculated as equation, is used, the preliminary work of the study
20.63%. These results show that the second order polynomial becomes more important. The trend of the response as a
equation hardly explains the true behavior of the system. In function of the independent parameters should be observed
addition, enzyme deactivation kinetics was investigated and and it should be searched whether the trend of the response
the ratio of the specific activity of the final state (a1) was in the investigated range of parameter is suitable to explain
determined at different combinations of pH and temperature with second order equation or not. If not, it is necessary to
using experimental and model data. The a1 values obtained do some further work, such as the transformation of the
for two different sources were compared and the AAD values dependent or independent parameters into a suitable form
were calculated as 55.1%. In this study, the maximum or changing the range of the independent parameter into a
response was obtained as 0.479 U at the minimum value of narrow range or another range, in which the trend of the
temperature (41 °C) and pH (3.6). But it does not mean that response could be explained with this equation.
this point is optimum. A further study is needed to obtain In the beginning of the article we asked two questions.
stationary point for b-1,3-glucanase activity within the At this point we can answer them. First question is that
investigated ranges of the independent parameters. ‘‘Can RSM be used for optimization of all chemical and
biochemical processes without any limitation?’’. The
4. Discussion answer to this question is no. RSM is only usable for the
changes, which can be described with a second order poly-
A wide application of RSM in chemical and biochemical nomial equation. The second question is ‘‘Is RSM usable
processes was observed in the literature. In this paper, we for other purposes (determination of reaction kinetics,
 _
D. Basß, I.H. Boyacı / Journal of Food Engineering 78 (2007) 836–845 845

stability or evaluation of kinetic constants etc.) in addition Cornish-Bowden, A. (1999). Fundamentals of enzyme kinetics. London:
to optimization?’’. The answer to this question is again no. Portland Press Ltd.
Corzo, O., & Gomez, E. R. (2004). Optimization of osmotic dehydration
Responses at the minimum and maximum values of the of cantaloupe using desired function methodology. Journal of Food
independent parameter are critical for the study related Engineering, 64, 213–219.
with reaction kinetics, stability or others. Indeed, a small Çelik, D., Bayraktar, E., & Mehmetoğlu, Ü. (2004). Biotransformation of
mistake at this point totally changes the behavior of reac- 2-phenylethanol to phenylacetaldehyde in a two-phase fed-batch
tion kinetics. Unfortunately, RSM is weak at these values system. Biochemical Engineering Journal, 17, 5–13.
Faveri, D. D., Torre, P., Perego, P., & Converti, A. (2004). Optimization of
of the independent parameters. We observed a similar mis- xylitol recovery by crystallization from synthetic solutions using
take in the reviewed article about enzyme kinetics. It is well response surface methodology. Journal of Food Engineering, 61, 407–412.
known that the reaction kinetics is separated based on the Fogler, H. S. (1998). Elements of chemical reaction engineering. New
order of reaction rate such as elementary (zero, first, sec- Jersey: Prentice-Hall Inc.
ond, etc. order reaction) or nonelementary reaction kinet- Gobbetti, M., Lanciotti, R., Angelis, M. D., Corbo, M. R., Massini, R., &
Fox, P. F. (1999). Study of the effects of temperature, pH and NaCl on
ics. In addition, these kinetic models may not be enough the peptidase activities of non-starter lactic acid bacteria (NSLAB) by
to explain the true behavior of some reaction and nonlinear quadratic response surface methodology. International Dairy Journal,
kinetic models are generated such as Michaelis–Menten 9, 865–875.
kinetics. Although there are so many kinetic models, it is Ismail, A., Linder, M., & Ghoul, M. (1999). Optimization of butylgalac-
not correct to fit all experimental kinetic data to only a sec- toside synthesis by b-galactosidase from Aspergillus oryzae. Enzyme
and Microbial Technology, 25, 208–213.
ond order equation and then predict the kinetic constants Ladero, M., Santos, A., Garcia, J. L., & Garcia-Ochoa, F. (2001). Activity
based on this model equation. Unfortunately, in this type over lactose and ONPG of a genetically engineered b-galactosidase
of a study we do not have a chance to change the range from Escherichia coli in solution and immobilized: Kinetic modelling.
of independent parameters similar to optimization work. Enzyme and Microbial Technology, 29, 181–193.
We must take into account all behaviors of the response Lee, M.-T., Chen, W.-C., & Chou, C.-C. (1998). Maximization of
cholesterol oxidase production by Rhodococcusequi no. 23 by using
in a wide range of independent parameters. Only when response surface methodology. Biotechnology and Applied Biochemis-
we find a suitable transformation of parameter into an try, 28, 229–233.
explainable form with second order equation, we can use Manohar, B., & Divakar, S. (2004). Applications of surface plots and
RSM in determination of reaction kinetics, stability or statistical designs to selected lipase catalysed esterification reactions.
evaluation of the kinetic constants. Process Biochemistry, 39, 847–853.
Myers, R. H., & Montgomery, D. C. (1995). Response surface method-
ology: Process and product optimization using designed experiments.
References New York: John Wiley & Sons, Inc.
Nawani, N. N., & Kapadnis, B. P. (2005). Optimization of chitinase
Andersson, M., & Adlercreutz, P. (1999). Evaluation of simple enzyme production using statistics based experimental design. Process Bio-
kinetics by response surface modelling. Biotechnology Techniques, 13, chemistry, 40, 651–660.
903–907. Panda, T., & Naidu, G. S. N. (1999). Performance of pectolytic enzymes
Anjum, M. F., Tasadduq, I., & Al-Sultan, K. (1997). Response surface during hydrolysis of pectic substances under assay conditions: a
methodology: A neural network approach. European Journal of statistical approach. Enzyme and Microbial Technology, 25, 116–124.
Operational Research, 101, 65–73. Penny, D. J., & Lindfield, G. (1999). Numerical methods using MATLAB.
Aracil, J., Garcia, T., Sanchez, N., & Matinez, M. (1999). Enzymatic New Jersey: Prentice-Hall Inc.
synthesis of fatty esters. Part II. Optimization studies. Enzyme and Rai, P., Majumbar, G. C., DasGupta, S., & De, S. (2004). Optimizing
Microbial Technology, 25, 591–597. pectinase usage in pretreatment of mosambi juice for clarification by
Beg, Q. K., Sahai, V., & Gupta, R. (2003). Statistical media optimization response surface methodology. Journal of Food Engineering, 64, 397–403.
and alkaline protease production from Bacillus mojavensis in a Rana, D. S., Theodore, K., Naidu, G. S. N., & Panda, T. (2003). Stability
bioreactor. Process Biochemistry, 39, 203–209. and kinetics of b-1,3-glucanse from Trichoderma harzianum. Process
Beg, Q. K., Saxena, R. K., & Gupta, R. (2002). Kinetic constants Biochemistry, 39, 149–155.
determination for an alkaline protease from Bacillus mojavensis using Riahi, E., & Ramaswamy, H. S. (2004). High pressure inactivation kinetics
response surface methodology. Biotechnology and Bioengineering, of amylase in apple juice. Journal of Food Engineering, 64, 151–160.
78(3), 289–295. Senanayake, S. P. J. N., & Shahidi, F. (2002). Lipase-catalyzed incorpo-
Boyacı, I. H., Williams, P. C., & Köksel, H. (2004). A rapid method for ration of docosahexaenoic acid (DHA) into borage oil: optimization
the estimation of damaged starch in wheat flours. Journal of Cereal using response surface methodology. Food Chemistry, 77, 115–123.
Science, 39, 139–145. Vohra, A., & Satyanarayana, T. (2002). Statistical optimization of the
Copeland, R. A. (2000). Kinetics of single substrate enzyme reactions. medium components by response surface methodology to enhance
ENZYMES: A practical introduction to structure, mechanism and data phytase production by Pichia anomala. Process Biochemistry, 37,
analysis. New York: John Wiley & Sons Inc. 999–1004.