A New Efcient Mixture Screening Design for Optimization of Media

Fred Rispoli
Dept. of Mathematics, Dowling College, Oakdale, NY 11769

Vishal Shah
Dept. of Biology, Dowling College, Oakdale, NY 11769 DOI 10.1021/bp.225 Published online June 16, 2009 in Wiley InterScience (www.interscience.wiley.com).

Screening ingredients for the optimization of media is an important rst step to reduce the many potential ingredients down to the vital few components. In this study, we propose a new method of screening for mixture experiments called the centroid screening design. Comparison of the proposed design with Plackett-Burman, fractional factorial, simplex lattice design, and modied mixture design shows that the centroid screening design is the most efcient of all the designs in terms of the small number of experimental runs needed and C 2009 American Institute of for detecting high-order interaction among ingredients. V Chemical Engineers Biotechnol. Prog., 25: 980985, 2009 Keywords: screening, centroid mixture design, mixture design, optimization

Introduction
Fermentative performance and physiological response of microorganisms is directly inuenced by the composition of the medium, which in turn affects the results of strain analyses and strain performance in industrial applications. This makes the process of media optimization vital during the various stages of process development. The process of developing optimal medium involves two stages. In the rst stage, the critical medium ingredients and process parameters inuencing production of the desired product are screened. The primary goal in this step is to study the effect that a particular factor exerts on the dependent variable of interest. Once the ingredients critical to the production have been screened, the second stage of media optimization is to nd the optimum concentration of each component for maximum product formation. When optimizing the fermentation conditions for a biological process, different factors involved include nutrition components for carbon, nitrogen, phosphorus, and trace elements; physico-chemical parameters such as pH, temperature, and aeration rate must also be considered. Moreover, while screening for critical media components, interactions amongst ingredients must be taken into consideration. In a practical scenario, once a target product and culture are selected, the rst step in optimizing the culture media is to select the nutritional components and physical parameters that the scientist believe would have a signicant inuence on the production. This decision on which nutrient components to screen is either based on the standard literature originating from the studies carried out using the same or related organisms, cost analysis and process economics, and personal experience of the scientists. Normally, the list of ingreCorrespondence concerning this article should be addressed to V. Shah at [email protected].
980

dients to screen is limited by ve to seven ingredients. This approach involves a risk of not including a media ingredient that may have a positive inuence on the production directly or via interaction with other ingredients. A screening method that allows a scientist to screen a larger number of ingredients is desirable. Numerous studies have been published describing media experimental design strategies for fermentation.13 However, it is agreed that statistical design of experiment methods are a better approach than a classical one-factor at a time approach. Two of the most widely used experimental designs are the Plackett-Burman design and the fractional factorial design.4 Both of these design approaches are based on basic factorial design of experiments, which considers the factors affecting the production process at two levels. A factorial experiment includes a series of values or levels of each media component chosen, and combinations of the levels of the ingredients are tested. The objective of a factorial experiment is to measure the change in the production of a product when changing the level of each ingredient while holding the levels of the other ingredients xed, as well as when changing the levels of two or more ingredients simultaneously.5 Advantages of the design include its simplicity and assessment of a large number of ingredients on the relative efciency of the production process. However, the approach has major drawbacks. The number of experiments that one would need to perform increases exponentially as one increases the number of ingredients to be screened. An experiment involving 15 media ingredients using the basic factorial method would require 215 (32,768) experiments to obtain the optimal media. Fractional factorial designs are considered to be a good alternative to a full factorial design, especially in the initial screening stage of a project. Fractional factorial designs consist of a carefully chosen subset (or fraction) of the
C 2009 American Institute of Chemical Engineers V

Biotechnol. Prog., 2009, Vol. 25, No. 4

981

experimental runs of a full factorial design. The subset is chosen so as to expose information about the most important features of the problem studied, while using a fraction of the effort of a full factorial design in terms of experimental runs and resources. The disadvantages of fractional factorial experiments includes the large number of experiments that would need to be performed (also an exponential increase similar to full factorial designs) and an inability to capture more than three way interaction amongst ingredients. Software such as Statistica allows a maximum of 11 ingredients to be screened, whereas Minitab (Release 14) allows up to 15 or 16 ingredients to be screened with resolution III or IV. For a 15 factor design with resolution IV, one would need to perform 128 runs. On the other hand, the Plackett-Burman design is a very efcient screening design when only main effects are of interest and all interactions amongst ingredients are negligible when compared with the few important main effects. However, should the interactions between ingredients be signicant, they are confounded by a Plackett-Burman design. The main advantage of the Plackett-Burman design is that it requires very few runs. Typically, statistical software such as Statistica or Minitab will design Plackett-Burman experiments for up to 15 or 16 ingredients, and the number of runs required is 20 or less. The Simplex Lattice design has been widely used for both screening and nding polynomial models in pharmaceutical, geology, petroleum, food, and tobacco industries. In this design, one identies a set of points that are evenly distributed over the design space which is a simplex for most mixture problems. In general, to obtain sufcient data for a polynomial model of degree m with q ingredients, we allow ingredi1 2 ent i to assume the values xi 0, m , m, , m m, where i 1, 2, , q. The values xi are called proportion variables since xi gives the proportion of ingredient i in the mixture being considered. The {q,m}-simplex lattice design consists of all possible combinations (mixtures) of the ingredients that satisfy q m 1 Pq i1 xi 1. This leads to a design that requires m design points. A disadvantage of this is that the design includes blends that consist only of m components, where m is the order of the model. For example, if one explores a 5-component system by applying a second order model (i.e., m 2), it would only consider mixtures of up to two components in a {5,2} simplex lattice design. Thus, any interaction involving three or more ingredients is ignored by the design. For q and m relatively small, simplex lattice designs are simple to construct and simple formulas to calculate regression coefcients are known. Consequently, they are often used to nd the optimal region of a response.5 The coefcients of linear terms give the effect of individual components on the magnitude of the response. However, the interpretation of higher order coefcients is not as straight forward. As a screening technique with q [ 5, the simplex lattice design requires too many runs to detect interactions of three or more variables. For example, a simplex lattice design is not efcient for q 15 ingredients because a level m 1 design requires 15 runs and only looks at extreme point vertices. A level m 2 design requires 120 runs and does not detect interactions with more than two ingredients. A level m 3 requires 680 runs and looks at blends of two or three ingredients. But this number of runs is not acceptable in most cases since too many resources are required which is

why we are screening in the rst place. Moreover, software is needed to generate all the permutations needed to obtain the design points. Simplex lattice designs above level 3 with 15 variables are not included in popular statistical software because the number of runs required is too large. Hence, for a process involving interactions of three or more ingredients, this method is not practical for screening. In prior research,6 we proposed a modied simplex lattice design that is based on the simplex lattice with additional constraints. In the modied simplex lattice, there are just 2 three levels for the input variables given by {0, 1 q, q}. In addition, exactly one proportion variable is set to zero in each run, exactly one proportion variable is set to 2 q, and the remaining are set to 1 . This gives a total of q ( q 1) design q points that are distributed throughout the design space. Compared with the simplex lattice, the number of runs is greatly reduced. For example, when q 15 ingredients, the modied simplex lattice requires 210 runs, a signicant reduction from 680. The modied simplex lattice is simple to construct, however, there are no known formulas to calculate the regression coefcients. In this study, we describe a new centroid screening mixture design for the screening of media ingredients that is efcient in requiring fewer runs and also able to detect the interaction amongst media ingredients.

Experimental
Design of experiments Five different types of experimental designs were generated in the study for screening 15 ingredients. Three of these designs, namely, the Plackett-Burman, the fractional factorial with level IV resolution, and the simplex lattice {15,3} design were generated using Minitab, Release 14. A modied simplex lattice design was created based on the constraints described elsewhere.6 Developing Centroid Screening Mixture Design. The centroid mixture design in this study is dened as follows. The design points are contained P in the q-dimensional simplex {(x1, , xq): xi  0 and q i1 xi 1} and are of three varieties given in terms of the proportion vectors (x1, x2, , xq), where each xi is the proportion of ingredient i contained in the mixture. 1 centroid design point: x (x1, x2, , xq), where xi 1 q, for i 1, 2, , q. q interior design points: For j 1, 2, , q, yj j j j 1 2 (y1 ,y2 ,,yjq ), where yji q 1, for i = j, and yj q1. q null effects design points: For j 1, 2, , q, zj j j j 1 (z1 ,z2 ,,yjq ), where zji q 1, for i = j, and zj 0. Observe that the design detects all variables occurring in interactions since at most one variable is at the zero level in each experiment. In addition, more than half of the design points are distributed throughout the interior of the design space. Unlike factorial or simplex lattice designs, the number of experiments will not grow factorially or exponentially. The number of experiments required for the centroid screening design is 2q 1, which grows linearly in q. Consequently, the design can be used with a large number of ingredients. To screen 15 ingredients in a media optimization experiment, as shown in Table 1, only 31 experiments are needed. This includes 1 centroid point whose coordinates are all 1/15 (run no. 1), 15 interior points whose coordinates

982
Table 1. The q 5 15 Experimental Design Matrix for the Centroid Screening Mixture Design Run 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 2 24 25 26 27 28 29 30 31 x1 1/15 1/8 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 0 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 x2 1/15 1/16 1/8 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/14 0 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 x3 1/15 1/16 1/16 1/8 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/14 1/14 0 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 x4 1/15 1/16 1/16 1/16 1/8 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/14 1/14 1/14 0 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 x5 1/15 1/16 1/16 1/16 1/16 1/8 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/14 1/14 1/14 1/14 0 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 x6 1/15 1/16 1/16 1/16 1/16 1/16 1/8 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/14 1/14 1/14 1/14 1/14 0 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 x7 1/15 1/16 1/16 1/16 1/16 1/16 1/16 1/8 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/14 1/14 1/14 1/14 1/14 1/14 0 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 x8 1/15 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/8 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/14 1/14 1/14 1/14 1/14 1/14 1/14 0 1/14 1/14 1/14 1/14 1/14 1/14 1/14 x9 1/15 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/8 1/16 1/16 1/16 1/16 1/16 1/16 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 0 1/14 1/14 1/14 1/14 1/14 1/14 x10 1/15 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/8 1/16 1/16 1/16 1/16 1/16 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 0 1/14 1/14 1/14 1/14 1/14 x11 1/15 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/8 1/16 1/16 1/16 1/16 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 0 1/14 1/14 1/14 1/14

Biotechnol. Prog., 2009, Vol. 25, No. 4

x12 1/15 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/8 1/16 1/16 1/16 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 0 1/14 1/14 1/14

x13 1/15 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/8 1/16 1/16 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 0 1/14 1/14

x14 1/15 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/8 1/16 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 0 1/14

x15 1/15 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/16 1/8 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 1/14 0

are all 1/16 except for one coordinate which is 2/16 (runs no. 216), and 15 null effects points whose coordinates are all 1/ 14 except for one coordinate which is 0 (runs no. 1731).

The last polynomial P6 is the only one with four-way interaction and also includes lower order terms.

Measuring effects Experimental data It would be very impractical to generate laboratory data to evaluate six different designs of experiments owing to the absence of information before hand on the type of inuence ingredients may have on a microorganism. We would have to carry out thousands of experiments for different organisms just to obtain different types of interactions amongst the ingredients. Thus, to compare the designs, we use synthetic data produced by polynomials having different characteristics concerning interaction. Six different test polynomials having unique characteristics were designed and are given in Figure 1. Each of these polynomials possesses unique characteristics and is given in order of complexity. Observe that P1 represents a process where only main effects are important with no interaction amongst the ingredients. In P2, there are rst and second order terms, but no variable occurs in more than one term. So, P2 represents a process with a combination of main effects and interactions of two ingredients. However, no ingredient interacts with more than one other ingredient. The polynomial P3 has rst and second order terms and several variables occur in more than one term, indicating that an ingredient could interact with more than one ingredient. P4 has rst, second, and third order terms with variables that occur in more than one term. In particular, P4 involves a small number of main effects, plus two-way and three-way interactions with ingredients. The polynomial P5 is the sum of P1 and P4, which involves numerous main effects in addition to two-way and three-way interactions with ingredients. One approach to screening is to choose an experimental design, obtain linear regression coefcients, and then consider the multiple correlation coefcient R value obtained from the linear regression computation. If R is sufciently large, then sort the regression coefcients to obtain a ranking of the proportion variables. This method may lead to the wrong conclusion. For a good value of R (i.e., above 0.900), implies that one has found good predictors for the response, which does not necessarily imply that a proper ranking of the proportion variables has been obtained. For example, the Plackett-Burman and the fractional factorial designs yield R above 0.900 for both P3 and P5 (Table 2 for R values). The coefcients obtained from these linear regression computations fail to predict the rank of the proportion variables correctly. This is partly due to the fact that with mixture experiments, there is an additional constraint that proportion variables must sum to one. So, even though a regression coefcient is positive, the centroid effect of the corresponding variable may be negative. As an alternative, sensitivity analysis based on the cent1 roid (1 q, , q) of the design space was used to determine the effect of an ingredient on the response. To measure the effect of a proportion variable xi, we dene vectors Ui and Vi as follows. To obtain Ui, start with the centroid and add a small amount denoted by d to the ith coordinate. Since the d sum of the coordinates of Ui must be 1, we then subtract q 1 from the remaining coordinates of Ui. In geometric terms, one can think of obtaining Ui by starting at the centroid and

Biotechnol. Prog., 2009, Vol. 25, No. 4

983

Figure 1. Six test polynomials used to generate synthetic data.

then taking a small step in the direction of the vertex representing ingredient i, where d represents the size of the step. Formally, Ui (u1, u2, , uq) satises ui 1 q d, and uj 1 d 1 , for j = i , where 0 \ d  . q q 1 q

The vector Vi is dened similarly, only now, we subtract d d from the ith coordinate of Vi, and then add q 1 to the remaining coordinates of Vi. In geometric terms, Vi is obtained by starting at the centroid and taking a step of size

Table 2. Results from a Comparison of Screening Designs for 15 Ingredients Experimental Design Plackett-Burman (20 runs) Polynomial P1 P2 P3 P4 P5 P6 P1 P2 P3 P4 P5 P6 P1 P2 P3 P4 P5 P6 P1 P2 P3 P4 P5 P6 P1 P2 P3 P4 P5 P6 Correct Rank Yes No No No No No Yes No No No No No Yes Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Correct Signs Yes No No No Yes No Yes Yes No No Yes No Yes Yes No No Yes No Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Identical Effects Yes No No No Yes No Yes Yes Yes Yes No No Yes Yes Yes Yes Yes No Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Multiple R 1.00 0.980 0.979 0.964 0.994 0.871 1.00 0.831 0.945 0.822 0.971 0.622 1.00 0.913 0.960 0.941 0.999 0.820 1.00 0.998 0.999 0.999 0.999 0.999 1.00 0.999 0.999 0.999 0.999 0.999

Fractional Factorial Resolution IV (128 runs)

Simplex Lattice Level m 3 (680 runs)

Modied Simplex Lattice (210 runs)

Centroid Screening Design (31 runs)

984

Biotechnol. Prog., 2009, Vol. 25, No. 4

d in the opposite direction of the vertex representing ingredient i. That is, Vi (v1, v2, , vq) satises vi 1 q d, and vj d 1 q q1, for j = i. Given a polynomial P(x) P(x1, x2, , xq), we obtain the centroid effect of a variable xi on P(x) by E(xi, P) P(Ui) P(Vi). Since we are interested in small changes around the centroid, a convenient choice for d is 1 1 2 either 1 q or 2q. When we set d q, we have uj q, uj q 2 1 qq1, for j = i, vi 0, and vj q1, for j = i. The centroid effect is intended to measure the change in the variable xi while keeping the ratios of the other variables constant. For example, suppose that q 15 and we wish to calculate the centroid effect of the proportion variable x1 on 2 13 13 13 P5, using d 1 q. Then U1 (15 ; 210 ; 210 ; ; 210), V1 1 1 1 (0; 14 ; 14 ; ; 14), and the centroid effect of x1 is given by E(x1, P5) P5(U1) P5(V1) 7.12. Observe that for all j u = 1, and all k = j with k = 1, ukj 1, and this holds for the vector V1 as well. By keeping the ratios constant, we are assuming that the small difference between uj and vj will have a negligible effect on the response compared with the relatively large difference between ui and vi. We can also demonstrate that for q [ 10, there is at most a 2% difference between uj and vj for j = i. First note that d 1 d 2d |uj vj| j1 q q1 q q1j q1. We also know 1 2 that d  q, which implies that |uj vj|  qq2 1  100. In the 2 example given above, we have that for j = 1, |uj vj| 210 \ 1%.

Results and Discussion

We now compare the ve designs of experiments Plackett-Burman, the fractional factorial with level IV resolution, the simplex lattice {15,3} design, the modied simplex lattice design, and the centroid screening design based on the below mentioned criteria. a). Ability to rank the ingredients in the right order based on its centroid effects on the response. b). Ability to correctly determine which ingredients have positive effect, negative effect, or virtually no effect. c). Ability to determine if two ingredients have identical effects on the response. d). The level of interaction detected. e). The multiple regression correlation coefcient R obtained from the computation of linear regression coefcients. f). The number of experimental runs to be performed. To compare the designs, we performed the following two steps using q 15 hypothetical ingredients. Step 1: For each of the polynomials given in Figure 1, we determined the centroid effect for all of the input variables. This was used to rank the variables and determine the sign of the effect (positive or negative) for each input variable. This information was taken to be the true results. Step 2: For each of the ve designs discussed above, and for each of the six polynomials given in Figure 1, experimental output P(x) is generated by substituting the values of the xi from the design into the polynomial. For example, when one substitutes the values of ingredients x1 to x15 as described in Run 1 of the centroid screening design (Table 1) into polynomial P1, we obtain the output of 1 1 1 P1(15 ; 15 ; ; 15 ) 0.0667. Multiple regression analysis was then performed using the xi as X input variables, and the output values of the polynomials P1, , P6 as Y input variables with the constant set at zero. Using regression software, linear regression coefcients were obtained for each design-

polynomial pair. Based on the regression coefcients, the centroid effects were calculated and used to predict the ranking of the input variables and the sign of the input variables. Table 2 summarizes the results on the comparison of ve experimental designs using six polynomials. Plackett-Burman and the fractional factorial designs led to the correct ranking of the variables only in the case of P1, the test polynomial with just linear terms. This was expected from the 20 run Plackett-Burman design since it is intended to only detect main effects. However, the 128 run fractional factorial design did not lead to better results either. The regression coefcients produced from the Plackett-Burman also failed in identifying proportion variables that have identical effects such as x5 and x6 in P2. Whereas, the fractional factorial was able to detect variables with identical effects such as x5 and x6 in P2, and the pairs x3, x4, and x5, x6 in P3. For P5, the fractional factorial method led to identical effects for x1 and x2 which is not true. Moreover, for P6, the fractional factorial design led to identical effects for x5 and x6, which is also not true, and did not detect the fact that x5 and x7 produce identical effects. It should be noted that the fractional factorial design used in this study is with resolution IV because the software used for generating the design would only allow maximum IV resolution. Also, at higher level of resolution such as resolution V and VI, the number of experiments one would need to perform will go up and high levels of interactions will not be captured. The simplex lattice design produced regression coefcients which obtained the correct ranking for the rst ve test polynomials. For polynomials P3 and P4, there were errors in the sign of the centroid effects. Moreover, when we consider the test polynomial P6, the simplex lattice design does not produce the correct rankings or signs for this case. In addition, the design produced linear regression coefcients that predict the effects of eight proportion variables are identical which is not true. The modied simplex lattice and the centroid screening design are able to produce linear regression coefcients that correctly rank the proportion variables for all of the six test polynomials. In addition, the regression coefcients produce the correct signs, identify any variables that have identical effects, and do not predict identical effects when it is not true. Clearly, both designs can detect high order interactions; i.e., interactions involving four ingredients or more. However, the modied simplex lattice required 210 runs, whereas the centroid screening design required just 31. Although it should be noted that the proposed centroid screening design may not be adequate for studying mixtures rich in any one ingredient. The design points are not uniformly distributed throughout the entire design space. In particular, design points located close to the simplex vertices are not included. Instead, the design points are located closer to the center. Satish et al.7 presented an augmented simplex-centroid design with three ingredients. This design is appropriate to study mixtures rich in one or more ingredient. The limitation of the augmented simplex-centroid design is the requirement of 2q 1 experiments to screen q ingredients. For screening 15 ingredients, 32,768 experiments would have to be carried out. McConkey et al.8 proposed a fractional simplex-centroid design to estimate main effects and signicant interactions amongst ingredients while keeping the number of runs required relatively small. However, to generate the design, one would require an algorithm to nd the subset of design points and the described algorithm unless included in the future DOE software may be hard to run. In comparison, the

Biotechnol. Prog., 2009, Vol. 25, No. 4

985
cline production with K-carrageenan immobilized. Streptomyces aureofaciens. 1997;21:314320. Lee S-L, Chen W-C. Otimization of medium composition for the production of glucosyltransferase by Aspergillus niger with response surface methodology. Enzyme Microb Technol. 1997;21: 436440. Liu B-L, Tzeng Y-M. Optimization of growth medium for the production of spores from Bacillus thuringiensis using response surface methodology. Bioprocess Biosystems Engineering. 1998; 18:413418. Broedel SE Jr, Papciak SM, Jones WR. The selection of optimum media formulations for improved expression of recombinant proteins in E. coli. Technial Bull. 2001;2:18. Cornell J. Experiments with Mixtures: Designs, Models, and the Analysis of Mixture Data. New York: Wiley; 2002. Rispoli F, Shah V. Mixture design as a rst step for optimizatios of fermentation medium for cutinase production. J Ind Microbiol Biotechnol. 2007;34:349355. Satish L, Lakshmi GS, Rao ChS, Brahmaiah P, Prakasham RS. Mixture design as a rst step for improved glutaminase production in solid-state fermentation by isolated Bacillus sp. RSPGLU. Lett Appl Microbiol. 2008;47:256262. McConkey BJ, Mezey PG, Dixon DG, Greenberg BM Fractional simplex designs for interaction screening within complex mixtures. Biometrics. 2000;56:824832.

design described in this study is easy to design and can be applied in the laboratories and industries with ease. Moreover, the design proposed by McConkey et al.8 will only capture interactions of a predetermined order and will ignore interactions of order higher than the predetermined order. For example, with 15 ingredients, one would need to carry out 105 experiments and will be able to capture only predetermined interactions of the third level. In conclusion, the centroid screening design can be applied to efciently screen ingredients for optimization of fermentation medium but can also be applied within the eld of toxicology, pharmacology, and industrial processes involving mixtures.

2.

3.

4. 5. 6.

Acknowledgments
The study was funded by National Science Foundation (Grant # CBET 0828292). The authors also acknowledge help of Ms. Sultania Makrides in carrying out certain statistical analysis carried out in the study.
7.

8.

Literature Cited
1. Teruel MLA, Gontier E, Bienaime C, Saucedo JEN, Barbotin J-N Response surface analysis of chlortetracycline and tetracyManuscript received Nov. 20, 2008, and revision received Mar. 23, 2009.