Applications of Mixture Experiments

Download as pdf or txt
Download as pdf or txt
You are on page 1of 19

Received: 6 January 2020 Revised: 1 April 2020 Accepted: 7 April 2020

DOI: 10.1002/cem.3246

SPECIAL ISSUE - REVIEW

Applications of mixture experiments for response surface


methodology implementation in analytical methods
development

Silvana M. Azcarate1,3 | Licarion Pinto2 | Héctor C. Goicoechea1,4

1
Consejo Nacional de Investigaciones
Científicas y Técnicas (CONICET), Godoy
Abstract
Cruz 2290, CP C1425FQB Buenos Aires, This review presents applications of response surface methodology (RSM)
Argentina when mixture experiments are involved for the optimization in the field of
2
Departamento de Química Fundamental,
analytical methods development. Several critical issues such as sort of designs,
Universidade Federal de Pernambuco
UFPE, Recife, PE50670-901, Brazil modeling with least squares or artificial neural networks, and multiple
3
Facultad de Ciencias Exactas y Naturales, response optimization are discussed. The results of a literature survey of the
Universidad Nacional de La Pampa, and works reported up to 2019 are presented. Finally, an illustrative example pro-
Instituto de Ciencias de la Tierra y
Ambientales de La Pampa (INCITAP), Av.
viding the necessary information to carry out this kind of work is discussed. A
Uruguay 151 (6300), Santa Rosa, La list of the most popular software available to apply RSM is also presented.
Pampa, Argentina
4 KEYWORDS
Laboratorio de Desarrollo Analítico y
Quimiometría, Facultad de Bioquímica y analytical methods, chemometrics, mixture design of experiments, optimization
Ciencias Biológicas, Universidad Nacional
del Litoral, Ciudad Universitaria, 3000
Santa Fe, Argentina

Correspondence
Héctor C. Goicoechea, Consejo Nacional
de Investigaciones Científicas y Técnicas
(CONICET), Godoy Cruz 2290 CP
C1425FQB, Buenos Aires, Argentina.
Email: [email protected]

Funding information
Agencia Nacional de Promoción Científica
y Tecnológica, Grant/Award Number:
PICT 2017-0340; Consejo Nacional de
Investigaciones Científicas y Técnicas,
Grant/Award Number: PIP-2015 Nº 0111;
ANPCyT, Grant/Award Number: 0340;
CONICET, Grant/Award Number: 0111

1 | INTRODUCTION

As Freddy Adams from University of Antwerp pointed out: “Analytical chemistry has played a major role in the success
of chemistry as a central discipline; it contributes fully together with other aspects of chemistry to the well-being of
humankind.”1 According to Shoji Motomizu, from Okayama University, “… the aims/roles of analytical chemistry are:
(1) to develop the methodology for getting information on the truth and the mystery in nature in chemistry, and (2) to
enhance intellectual assets through discovery of novel substances”.2 Henceforth, development of analytical methods is

Journal of Chemometrics. 2020;e3246. wileyonlinelibrary.com/journal/cem © 2020 John Wiley & Sons, Ltd. 1 of 19
https://doi.org/10.1002/cem.3246
2 of 19 AZCARATE ET AL.

a common task in a wide variety of activities. This assignment demands enormous efforts owing to high quality require-
ments of regulatory agencies, high cost of reagents, and large amounts of variables affecting the analytical process,
among others. At this point, the term optimization refers to improving the performance of the analytical process, that
is, determining the conditions at which the best response is obtained.3
During the development of an analytical method, significant factors that affect the quality of the method perfor-
mance should be carefully optimized. Among the factors affecting the procedure, a large number of the so-called pro-
cess factors as temperature, time of reaction, reagent concentration, pH of the solution, sort of agitation, types and
volumes of extraction and dispersant solvents, and salt addition, and so forth, can be cited. In addition, there is a special
sort of factors that should be studied at different proportions, as a mixture of extractant solvents in a preconcentration
step of several substances with different physicochemical properties. On the other hand, the quality of the performance
of the method can be measured through responses, for example, sensitivity, percentage of recovery, and peak area in a
chromatographic method.
In this context, the response surface methodology (RSM), which is a collection of statistical techniques, becomes an
extremely useful tool for modeling and analyzing the effects of several parameters of the process under study. RSM is
an important protagonist to find out the best combination of factors that produces the optimum/a response/s. Its funda-
mental idea is to reach the optimum conditions executing the lesser number of experiments as possible. This multivari-
ate methodology is much more practical compared with the conventional univariate experimental work, which is
called “one variable at time.” The latter fact is owed that the multivariate approach is carried out from experimental
data gathered from statistical experimental designs which are built under certain requirements.4
Regrettably, design of experiments and optimization by the RSM is not as known and applied as it should be
wanted. Nowadays, a large number of reports can be found showing that the optimization of the analytical procedure
was performed by the univariate approach.5 Thus, the goal of this review is to evidence the real advantages in terms of
both the reduced experimental effort and the improved quality of information that can be obtained by following this
approach, when working with mixture designs in the field of analytical methods development. In a mixture experiment,
unlike conventional experiments, the factors are proportions of different components, and the measured response is
supposed to depend only on the relative proportions of the constituents in the mixtures and not on the total amounts in
them.6
Figure 1 shows a general scheme considering the steps commonly followed for RSM implementation when develop-
ing an analytical method. The application includes a series of consecutive steps: (a) screening study, in which an initial
large amount of factors suspected of having influence in one or several responses, is reduced by means of screening
designs; (b) building response surface designs to provide experimental data (these designs can be built for process fac-
tors or mixture experiments); (c) modeling through least squares fitting or exceptionally by artificial neural networks;
and (d) simultaneous optimization following appropriate criterion, when a large number of responses are involved (the
desirability function is the most popular tool to be applied).4

F I G U R E 1 Scheme showing the steps


generally followed for RSM implementation
when developing an analytical method
AZCARATE ET AL. 3 of 19

This review is focused on applications of mixture design and its corresponding modeling to enhance the optimiza-
tion during analytical method developments. For this purpose, a description of the theory involved and the works
reported up to 2019 were described. In addition, an illustrative example providing the necessary information to carry
out this kind of work is discussed.

2 | EXPERIMENTAL DESIGN AND RES PONS E SURFACE METHODOLOG Y

As commented before, the response surface methodology (RSM) is a strategy based on mathematical modeling of data
gathered by application of multivariate design of experiments (DOE) and experimentation, with the aim of optimization
of a given process. Interestingly, its implementation requires fewer efforts and resources than those involved in univari-
ate optimization. In the following, a description of the responses and factors, sort of designs involved in the process of
data collection, modeling the experimental data to establish the relationship among responses and variables, and multi-
ple response optimization is discussed.

2.1 | Factors and responses

Factors can be controlled and uncontrolled variables, which affect the process of an analytical method development
(see Figure 1). These factors can be numerical (as temperature and time of reaction, pH, concentration of a given
reagent, etc.) or categorical (as sonication or vortexing, light or darkness, etc.). In addition, in many applications it is
necessary the implementation of experiments involving mixtures or blends. For instance, in the field of sample pre-
treatment, especially in extraction and preconcentration steps, the enhancement of the percent recovery of a number of
substances at the same time depends on the mixture of several solvents (e.g., water, methanol, and acetonitrile).
Another example is found in mobile phase optimization in high performance liquid chromatography (HPLC) methods
development.
The response is some performance measure or quality characteristic of the system under study. Several examples
can be found in the field of analytical methods development: sensitivity (measured by the height or width of a chro-
matographic peak), precision (coefficient of variation), time of a chromatographic separation, and so forth.

2.2 | Experimental designs

2.2.1 | Screening

The screening step consists of the exploration of the variables, which could have an influence on the process. The aim
is to consider a large number of factors at the beginning following a screening design and statistical analysis of the data
(employing graphical tools and ANOVA). Finally, it is concluded which of the factors have a significant influence on
the response/s.
Screening designs involve factor variations in two levels: –1 and +1 (widest interval in which the factor can be
varied in the system under study). The most popular designs are full factorial, fractional factorial, Plackett-Burman
(PBD), and Taguchi designs (TD). The factorial fractional design is the most used and allows for the evaluation of a
large number of factors in a small number of experiments by fractioning a full factorial 2k design in a 2k-p design,
being p the number chosen to fractionate the design. As an example, if six factors should be studied, a fractional fac-
torial design with 26-2 = 16 experiments is good option against the 64 experiments required for a 26 full factorial
design.7

2.2.2 | Response surface methodology designs with independent factors

The main objective pursued in the choice of the design to model a second order response surface is performing a
reduced number of experiments. However, sometimes the analyst wishes to exclude zones from the experimental
region, reaching high quality adjusted parameters.
4 of 19 AZCARATE ET AL.

Full factorial, central composite (CCD) and Box-Behnken (BBD) are the most reported designs in the literature. The
number of experiments in a full factorial design is equal to 3k, being k the number of factors. Thus, in the case of three
factors, 27 experiments are required, although replication of the central points is recommended to evaluate a lack of fit.
On the other hand, central composite design is among the most popular options, built with two-level (–1 and +1) facto-
rial design points, axial or star points and center points (with all factors set to 0). All factors in the star points are set to
0, except for one factor whose value is set at ±α; a number that determines the location and usually varies from 1 to √k.
For three factors, the number of experiments (equal to 15) results from the equation (2k + 2 k + central points),
although several replicated points are required. Finally, the number of experiments for a Box-Behnken design is com-
puted as (2 k[k–1] + central points). For three factors and one central point design, at least 13 experiments are required,
becoming the BBD in most efficient design.

2.2.3 | Response surface methodology designs with dependent factors: Mixture designs

Simplex designs
Mixture simplex centroid designs are built in those cases in which the analyzed factors are the components of a mix-
ture, that is, their levels are not independent of one another, allowing the study of the effect of the ratio variations
among the factors. The domain is a regular figure having as many vertices as components, in a space with dimensional-
ity equal to the number of components minus one. In Figure 2A,B, graphical representations of mixture designs with
three and four components, respectively, can be appreciated. In the first case, the design is an equilateral triangle whose
vertices correspond to combinations containing 100% of a single component, each of the three sides represents a binary
mixture and the internal points correspond to ternary mixtures. Usually, these designs are augmented with extra points
in the interior of the experimental region.

Simplex designs with constraints in the component proportions: Pseudocomponents


Some mixture experiments need constraints on the constituent proportions. These are often upper- and/or lower-limits
with the following form: Li ≤xi ≤ Ui, i =1, 2, q, where Li is the lower bound and Ui is the upper bound for the ith compo-
nent, respectively. The lower bound represents a minimum proportion of the ith component, while the upper bound is
a maximum proportion of the ith component that must be present in the mixture. Generally, the constraints are set by
practical bounds on the relative proportion of components that produce working results. Equation 1 represents a con-
strained mixture problem for a three-component system:

Li ≤ xi ≤ Ui, ð1Þ

where x1 + x2 + x3 = 1, Li ≥ 0, and Ui ≤ 1 for i = 1, 2, 3. This case is represented in Figure 3A, in which L1 = 0.20,
L2 = 0.33, and L3 = 0.25. On the other hand, Figure 3B shows a three component system with the following upper limits
imposition: U1 = 0.60, U2 = 0.50, and U3 = 0.50. As it can be seen, these constrains do not modify the shape of the orig-
inal simplex design.

F I G U R E 2 Graphical representations
of mixture designs with (A) three and
(B) four components, respectively
AZCARATE ET AL. 5 of 19

F I G U R E 3 Graphical representation of
constrained designs (A) with lower limits
imposition: L1 = 0.20, L2 = 0.33, and
L3 = 0.30; and (B) with upper limits
imposition: U1 = 0.60, U2 = 0.50, and
U3 = 0.50

Owing that the original simplex form is still present, a new set of components can be defined having values from
0 to 1 over the feasible region. This fact simplifies the design building besides of the posterior modeling. These new
components are usually named pseudocomponents.
On the other hand, it is important to note that if only one upper limit had been imposed, the simplex shape would
have been lost. Figure 4A represents a design in which upper limits have been imposed for the three components:
U1 = 0.70, U2 = 0.80, and U3 = 0.60. Similar behavior will be present in those cases in which both lower and upper bou-
nds are imposed. Figure 4B shows the latter case, in which the following constraints were imposed: L1 = 0.20, L2 = 0.33
and L3 = 0.25, U1 = 0.42, U2 = 0.50 and U3 = 0.47.

Designs with constraints in the component proportions when the simplex shape has been lost: D-optimal criterion
The D-optimal design can be a useful strategy to select points for a mixture design in a constrained region in those cases
in which the simplex shape has been lost, as in those cases shown previously in Figure 4A,B. Through the point
exchange procedure, design points can be selected from a list of candidate points so that the variances of the model
regression coefficients are minimized, that is, the points are selected in such a way that the determinant of the matrix
XTX is maximized, being X the matrix design considering the model to be fit.
In order to implement the D-optimal design algorithms, a set of reasonable candidate points from which to select
the design points is required. In addition, a systematic procedure is necessary for selecting the points. The candidate
points should be selected considering the model which the experimenter wants to fit. For instance, let us consider the
example 12.2 described in Myers and Montgomery.4 This example corresponds to a shampoo formulation with lower
and upper bonds as: 0.14 ≤ x1 ≥ 0.20, 0.40 ≤ x2 ≥ 0.60 and 0.26 ≤ x3 ≥ 0.40. In this case, six of the 13 candidate points
must be selected in order to fit the quadratic model, plus several additional runs in order to estimate error and adequacy
of the model (four replicated and three distinct points), giving in this case 13 experimental combinations. For a qua-
dratic model (see below), the candidate points should include four vertices of the region, the four edge centers, the over-
all centroid, and the four axial runs that lie midway between the centroid and the vertices of the constrained region.
The design built with the software Design-Expert8 is shown in Figure 5 in terms of the pseudocomponents. This figure
shows the candidate points (stars and spots) and those selected by the D-optimal criterion (stars).

F I G U R E 4 Graphical representation of
constrained designs with (A) upper limits
imposed for the three components: U1 = 0.70,
U2 = 0.80, and U3 = 0.60; and (B) upper and
lower impositions: L1 = 0.20, L2 = 0.33 and
L3 = 0.25, U1 = 0.42, U2 = 0.55 and U3 = 0.47
6 of 19 AZCARATE ET AL.

FIGURE 5 Representation of the D-optimal design built for a


shampoo formulation with lower and upper bonds given by:
0.14 ≤ x1 ≥ 0.20, 0.40 ≤ x2 ≥ 0.60 and 0.26 ≤ x3 ≥ 0.40.4 Candidate
points (stars and spots) and those selected by the D-optimal criterion
(stars)

Mixture-process and mixture-mixture designs


When performing and experiment with mixtures, sometimes process variables can affect the blending properties. For
example, optimum electrophoretic conditions were set by experimental design and optimization combining process var-
iables (voltage and pH) and a mixture of background electrolyte (BGE) salts.9 The design was built by combining a sim-
plex lattice design (seven combinations of the mixture components) with a factorial design (including nine
combinations of the two process variables). The whole design needed 63 experiments. For the mixtures, the amounts of
three different salts used to prepare the BGE were varied using the simplex lattice design, in the range of 0 to
15 mmol L−1, with the restriction that no pure salt should be used and with a maximum concentration of 10 mmol L−1.
This design consists of seven combinations: one central point, three points in the interior of the simplex, and the last
three points in the center of the edges. As can be appreciated in Figure 6, the factorial process design consisted on nine
combinations of pH (between 7.00 and 9.00) and voltage (between 15 and 25 kV).
The combined design with 63 points was used to obtain the experimental responses and fitting a combined model
that includes both the mixture and the process variables. The usual approach is to fit the complete combined model
(see below), which would be very useful for predicting the response of a particular mixture formulation at combinations
of the process variables that are not factorial design points.
Another sort of combined design can be built for a combination of two mixture designs as shown in Figure 7. An
interesting example was presented in the Almeida and Scarminio10 in which a crossed mixture design was built for the
optimization of mobile phase and for solvent extraction media. First, a simplex-centroid design with three axial mixture

F I G U R E 6 Experimental
mixture-process design scheme of
the optimization presented in
Alcaráz et al.9 The red dots
represent the factorial design points,
and the green dots represent the
mixture design points
AZCARATE ET AL. 7 of 19

F I G U R E 7 Graphical representation of a
combined design built for a combination of two
mixture designs

points was used to investigate the influence of the different solvent proportions on extract preparation, based on that
solvents from different groups in the triangle have different selectivity characteristics. Second, considering that the type
and concentration of the organic modifiers are of prime importance in HPLC, a mixture of water and two organic sol-
vents was carried out using a simplex-centroid design involving mixtures with the component proportions illustrated in
Figure 7. The crossed design leads to 70 different combinations.

2.3 | Modeling

2.3.1 | Classical least squares fitting

The RSM involves building a mathematical model describing the relationship among each response and factors. The
model can be a first-, second-, or higher-order polynomial function, which is fit employing the data collected after
experimentation according to the experimental design chosen for that purpose. The second-order model describing the
relation among one response (y) and two factors (x1 and x2) can be defined by the following regression model:

y = b0 + b1 x 1 + b2 x 2 + b1− 2 x 1 x 2 + b1 − 1 x 1 2 + b2 − 2 x 2 2 + e, ð2Þ

where b0 is the overall mean effect, b1 and b2 represent the effects of the factors x1 and x2, b1-2 is the effect of the interac-
tion between both factors and b1-1, b2-2 correspond to the quadratic terms and e corresponds to the error of the fit.
It should be remarked that the fitted model must properly describe the data relationship to make accurate
predictions inside the experimental region. In the case of two factors are optimized, a response surface can be rep-
resented in a three-dimensional space. If more than two factors are optimized, the graphical representation is
made for two of them, maintaining the others at constant values. In addition, a contour plot representation, con-
sisting of iso-response lines, is a useful tool to see how is the behavior of the response inside the experimental
region.
The significance of the coefficients should be investigated by mean of ANOVA test. The model can be considered
acceptable when the regression is significant, and a nonsignificant lack of fit is obtained for the selected confidence
level (α = 0.05). After the test application, the analysis of the residual plots, as well as both the coefficient of determina-
tion (R2) and the adjusted coefficient of determination (Radj2) evaluations, is required. The latter represent the percent-
age of variance explained by the model. Finally, the fulfillment of the ANOVA assumptions should be checked by the
8 of 19 AZCARATE ET AL.

following tools: (a) the normal probability plot, which indicates that the residuals follow a normal distribution; (b) the
plot of residuals versus the ascending predicted response values, which allows checking homogeneity of the variance;
and (c) the plot of residuals versus the runs to test independency. Whether absence of normality or heterocedasticity in
residuals is detected, transformation of the response can be applied. Another alternative is modeling with non-
parametric methods, as artificial neural networks (ANN), which will be shown below.

2.3.2 | Least squares fitting of mixture experiments: Scheffé model

The polynomial model used for fitting mixture designs was developed by Scheffé,11 who modified some terms from the
polynomial expression presented above (Equation 2) in order to exclude the constraint originated in the correlated vari-
ables. Equation 3 shows the canonical form of the quadratic model:

X
q XX
y= bi x i qi< j bij x i x j , ð3Þ
i=1

whereq the parameter bi represents the expected response to the pure mixture xi = 1, xj = 0, j 6¼ i. The term given by
P
y= bi x i represents the response when blending is strictly additive, and there are not interactions between the com-
i=1
ponents of the mixture. The quadratic term bijxixj represents the excess response over the linear model due to the inter-
action between two components, and this effect is often called synergism (or antagonism).4 Figure 8 shows a typical
response surface plot obtained with a quadratic three factors model.
In a combined design as those shown in Figures 6 and 7, two models have to be taken into account, one for each
mixture being considered or one for the process factors model and one for the mixture design. In this case, a product
model can be implemented with the two groups of components x and z.12 This combined model can be represented by
Equation 4

y = f ðx Þ × gðzÞ + e, ð4Þ

where y is the response, the functions f(x) and g(z) are separated polynomial models that represent the two models, and
e is a zero-mean random variable with variance independent of x and z.
For the case of mixture-mixture designs, and considering both functions f(x) and g(z) as quadratic, the qua-
dratic × quadratic product model will be represented by the following equation:

F I G U R E 8 Graphical representation of a response surface plot


obtained with a quadratic three factors model
AZCARATE ET AL. 9 of 19

XX XX XX XX
y= bi j x i z j + bi jl x i z j zl + bik j x i x k z j + bik jl x i x k z j zl , ð5Þ
i j i j<l i<k j i<k j<l

where i, k = 1, … I, i < k and j, l = 1, … J, j < l. The parameters b's are products of the regression coefficients of xi and zj.
This particular product model demands to fit 36 parameters b.

2.3.3 | Artificial neural networks fitting

An ANN constitutes a nonparametric modeling which mimics the human brain. Among the large number of ANNs
reported and employed for a wide variety of tasks, the multilayer feed-forward networks or multilayer perceptron
(MLP) networks are the most used for prediction, classification, and optimization purposes.13 The typical architec-
ture consists of three layers of neurons or nodes, which constitute the basic computing units. The input layer has a
number of neurons equal to the number of factors under study. The hidden layer has a variable number of active
neurons (in some occasions, more than one hidden layer can be used). The output layer has a unit for the response.
These neurons are linked hierarchically: The outputs of one layer of neurons are used as inputs for the next layer
and so on. The input and hidden layers commonly have a sigmoid function in its nodes. On the other hand, a lin-
ear function is generally used in the output layer. The number of hidden layers and the neurons in each of them
must be optimized to attain satisfactory fitting ability, which is associated with an acceptable predictive ability. It is
important to note that there is a main difference between modeling by LS regression or ANNs: For the latter, it is
not necessary to know the exact form of the function on which the model should be built, that is, an implicit model
is built.
Radial basis functions (RBF) is another popular ANN, which has a single hidden layer of neurons incorporating
Gaussian transfer functions, and a linearly activated output layer. In comparison with an MLP network, an RBF pre-
sents some advantages such as faster training stage as well as robustness towards noisy data.14

2.4 | Multiple response optimization

In those cases, in which the optimization process involves several responses, an overall solution must be found inside
the experimental region. The desirability function, presented by Derringer and Suich in 1980, is the most popular tool
among the several approaches that have been presented for solving multiple response optimizations.15 The latter func-
tion is based on the idea that the quality of a process having several responses is undesirable if one of them lies outside
of an accepted bound. Thus, the function searches working conditions to guarantee agreement with the criteria of all
the involved responses, providing the best compromise values for those responses.
In order to implement this multiple response approach, the first phase is to compute an individual desirability
function di (ŷi) for each response ŷi(x). This is done using previously fitted models (by LS or ANN). Several opti-
mization criteria are usually implemented: maximization, minimization, reaching a target value or a range of
values. To satisfy the latter criteria, different functions are built, making the desirability values vary within an
acceptable range of response values established by the range Ui–Li, where Ui and Li are the upper and lower
acceptable values, respectively. For example, if a mixture of solvents is assayed to attain the maxima extraction of
several substances, a maximization function, as the described in Equation 6, should be computed for each
response:
2 3
0 if y^i ðx Þ < Li
 
6 y^ ðx Þ − Li s 7
di ðy^i ðx ÞÞ = 6 i
4 U i − Li if Li ≤ y^i ðx Þ ≤ U i 7
5 ð6Þ
1 ^
if yi ðx Þ < U i

where s is a weight establishing the importance of the closeness of ŷi to the maximum.


On the other hand, in the case of the objective is reaching a target value Ti, the function can be described by
Equation 7:
10 of 19 AZCARATE ET AL.

2 3
0 if y^i ðx Þ < Li
6  S 7
6 y^i ðx Þ − Li
if Li < y^i ðx Þ < T i 7
6 T i − Li 7
6 7
di ðy^i ðx ÞÞ = 6
6 1 if y^i ðx Þ = T i 7 7, ð7Þ
6  t 7
6 y^i ðx Þ − U i
if T i < y^i ðx Þ < U i 7
4 Ti − Ui 5
0 if y^i ðx Þ < U i

where s and t is a weights establishing the importance of the closeness of ŷi to target value. As can be appreciated in
both Equations 6 and 7, the individual desirability takes values ranged from 0 to 1. If di (ŷi) = 0, the response is unac-
ceptable, but if di (ŷi) = 1, the response is totally desirable. Intermediate values of di (ŷi) indicate more or less desirable
responses. Interestingly, factor levels might also be included in the optimization procedure, for example, completing an
extraction procedure in the lesser time of work.
Once individual desirability functions are computed, they are combined in a global desirability function (D), which
allows to find out the best combined responses, through Equation 8:

!P1
 P1 ri Y
n ri
D= dr11 × dr22 × …… × drnn = drii , ð8Þ
i=1

where ri is the importance, which allows to indicate if one or more responses should be more carefully taken into
account than others. For example, if a mixture of solvents is assayed to attain the maxima extraction of several sub-
stances, and one of them presents a low sensitivity during the measurement step, an exponent equal to 5 can be set for
this particular substance, whereas the exponent can be set equal to 1 for the rest of them.
An inspection of Equation 8 allows to observe that if D reaches a value different from zero, all the responses can be
considered to have a desirable value. On the contrary, if di (ŷi) = 0, D will be zero, thus, at least one of the responses is
completely undesirable. It is important to consider that the goal of an optimization procedure is to find a good set of
conditions to fulfill all the requirements, but not to get a D value equal to 1.

2.5 | Software

For implementation of RSM, several programs are accessible to the analysts. All the programs allow the utilization of
different statistical tools and graphical options for screening analyses and DOE-optimization applications. The follow-
ing can be mentioned: Design Expert (http://www.statease.com/), Minitab (http://www.minitab.com/), SPSS (www.
ibm.com/analytics/us/en/technology/spss/), JMP (https://www.jmp.com/), R: The R Project for Statistical Computing
(https://www.r-project.org/), Six Sigma (https://www.isixsigma.com/six-sigma-software/), Unscrumbler (http://www.
camo.com/), NCSS (http://www.ncss.com/), Cornerstone (www.camline.com/), and SAS (https://support.sas.com/rnd/
app/qc/qc/qcdesign.html).

3 | LITERATURE SEARCH

Compared with factorial design, mixture design is still under-explored for optimization of analytical methods. Interest-
ingly, a search on the web of science using the terms “Factorial Design” and “Mixture Design” revealed that 17929 and
11796 papers (including 159 and 31 reviews), respectively, were reported. In addition, restricting the search by using
the Boolean operator “AND” with the term “Analytical methods,” 119 and 18 papers using the term “Factorial Design”
and “Mixture Design,” respectively, were found. The search was performed in all fields since 2000 up to 2019.
Figure 9A shows the reports presented year by year corresponding to the general situation, whereas Figure 9B displays
the particular situation regarding the development of analytical methods.
The publications using mixture experimental regards the analysis of food,16–19 drugs,20–25 biological samples,26–29
fingerprints,10,30–32 among others.33–37 The literature searches reveals that there are many situations where mixture
experiments are applicable, being the great majority of these works related to the sample preparation
AZCARATE ET AL. 11 of 19

F I G U R E 9 (A) Number of the reports presented during the


period of 2000-2019, year by year, corresponding to factorial designs
(dark gray bars) and mixture designs (white bars). (B) Number of the
reports presented during the period of 2000-2019, year by year, but
restricting the search to analytical methods

procedure10,17,19,30–32,35,36 or to the liquid chromatography mobile phase optimization.10,20,21,25,27,29,31,38 Most of the


mixture experiments requires a single simplex design to optimize one or multiple responses of the studied
system.23,24,39–46 However, several complex situations require innovative strategies to solve cases such as the application
of constraints18,19,29,32,35,37,47 or more than one mixture design (mixture-mixture design).10,28,30,31,48 In addition, process
(independent) and mixture (dependent) factors were also simultaneously optimized.16,20,21,34 Others strategies that use
spherical coordinate approaches have been used to optimize mixture and process factors by transforming mixture fac-
tors into independent variables parameters.27,38

3.1 | Mixture designs without constraint

In this section, those works that do not apply any constraint to restrict the mixture design space to optimize analytical
methods are presented.
Passari et al45 performed an analysis on Mikania laevigata leaves to access the seasonal and solvent composition
effects and evaluated their interaction on secondary metabolites using a mixture design. The solvent composition con-
sisted of a mixture of four solvents: ethanol, acetone, dichloromethane, and chloroform. They were varied using a
simplex-centroid mixture design. As responses, the masses of the crude extract and the fraction of neutral, organic,
basic, polar, fiber, and residue composition collected in spring, summer, autumn, and winter seasons were evaluated.
Zheng et al49 performed extractions in rats tissue samples (brain, heart, liver, kidney, muscle, and pancreas) with
different proportions of methanol, chloroform, and water assisted by a special cubic simplex lattice design. The ana-
lyzed tissue was added with the mixture defined by the experimental design, vortexed, centrifuged and the supernatant
was collected and lyophilized. This extract was reconstituted with a buffer prepared in D2O and analyzed with a 1H
NMR. The metabolite profiles were identified and quantified using two-dimensional 1H-1H correlated spectroscopy and
13 1
C- H heteronuclear single quantum coherence. The authors concluded that there is an optimum solvent proportion
to extract metabolites of each one of the analyzed tissues.
12 of 19 AZCARATE ET AL.

Pauli, Scarminio, and Tauler50 used a simplex centroid where the four solvents were chosen according to Snyder's
solvent-selectivity triangle to evaluate the metabolites profiles extracted from Camellia sinensis L. leaves of first and sec-
ond harvests from 2011 and 2012, using a multiway strategy in addition to the mixture design. Searching for a better
discrimination the authors used a liquid chromatography coupled to ultraviolet diode array detector and an electrospray
ionization mass spectrometer (LC-DAD-ESI/MS) and data fusion strategies to evaluate the differences between the sec-
ondary metabolites. By the analysis of the chemometric models obtained from the fused data, the authors found that
ethanol and binary mixture of ethyl acetate and dichloromethane with ethanol results in extraction of a higher number
of peaks, providing more information and better discrimination efficiency among the samples.
Finally, the work presented by Teglia et al40 will be discussed below as literature example.

3.2 | Mixture designs with constraints

Moreira et al19 used a simplex-centroid design to evaluate the digestion efficiency of cashew nuts for quantification of
copper, iron, and manganese. The design presented by the authors show both a lower limit containing 1.0 mL of each
component and an upper limit of 8 mL, resulting in a total of 10 mL of the digestion solution (see Equation 1). For the
optimization, instead of applying the desirability function, the authors normalized the signal of each metal with the
maximum emission signal of each one of them.51
Pinho et al18 optimized the solvent mixture for a solid–liquid extraction with low temperature purification for
extraction of pesticides on tomato samples and posterior analysis by gas chromatography. A constraint was imposed
based on a previous study defining a regular pseudocomponent simplex space. A linear model was used to adjust the
recovery values and the optimum condition was located in one of the vertices. The obtained recovery and limit of quan-
tification values attained showed that the method fulfilled the requirements of local regulatory agencies and presented
better recovery values than the one obtained by matrix solid phase dispersion.
Andrade et al37 used a simplex centroid with pseudocomponent design to define a homogeneous ternary
solvent system using a N,N-dimethylformamide (DMF)-ethanol-water mixture, for the determination of molybde-
num in steels by adsorptive stripping voltammetry. The proportion of the ternary system was optimized to
intensify the signal of the molybdenum, but one constraint was imposed to guarantee the electrolytic medium
andthe homogeneity of the mixture solution. After the mixture solution optimization, it was evaluated the pH,
a-benzoinoxime concentration, and the accumulation time on the signal of molybdenum. Instead of using experi-
mental design for a multivariate analysis, these three parameters were analyzed using a univariate study, which
did not allow the identification of interaction between these factors. The authors performed a quantification of
molybdenum in three certified samples and compared the results with the one obtained with a reference
instrument.

3.3 | Mixture design with constraints obtained by D-optimal criterion

As was commented before, to build the design when irregular space are defined by the imposition of constraints,
the D-optimal strategy can be a useful option.52 Azcarate et al35 proposed a mixture design based on the latter crite-
rion to access a ternary solution that was used for a sample preparation method for wear metals determination in
lubricating oil by microwave induced plasma with optical emission spectrometry. The experimental region was
restricted, requiring an irregular design to fulfill the constraints imposed by the experiment. One model was built
for each one of the 18 analyzes metals, using linear, quadratic, special cubic, or cubic models. All of them were sig-
nificant without lack of fit. According to the optimization, the metals required different optimum ternary mixture,
thus the Derringer's function was used to access a proportion that recovered a large amount of all the metals
simultaneously.
Chen et al32 used ionic liquid microemulsion with water to extract hydrophilic and lipophilic components from Flos
Chrysanthemi. For this purpose, the proportion of water, oil, and surfactant was optimized by a D-optimal design,
owing that all the constituents should be present and different proportions of each constituent of the ternary mixture
were required, which lead to an irregular design space. The responses were related to six substances, which were simul-
taneously optimized using the Derringer's function. Finally, to demonstrate the extraction efficiency, they compared
the results with others extraction solvents.
AZCARATE ET AL. 13 of 19

Xie et al47 proposed the use of a D-optimal design to monitor the production of 19 metabolites extracted from
Aspergillus flavus 73 and Leclercia adecarboxylata WT16. The mixture design was used to evaluate the proportion of
methanol, dichloromethane, acetonitrile, and ethyl acetate. Although the authors proposed the use of a D-optimal
design, the studied region did not require any constraint. So a simplex centroid design would lead the same result,
but with only 15 experiments instead of 20, being still possible to evaluate a special cubic model as the one used by
the authors.

3.4 | Mixture-mixture and mixture-process designs

Crossed-mixture designs are used to optimize simultaneously two mixture experiments. In the context of the develop-
ment of analytical methods, this strategy was mostly used to evaluate the interaction between the extraction solvent
and mobile phase mixture for the analysis in a liquid chromatography system.10,31 Almeida and Scarminio10 used a
mixture-mixture design to evaluate the quality of the chromatographic fingerprint for the characterization of green tea.
The authors found that to develop an efficiency fingerprint method, it is important to optimize not only the extraction
solvent but also the mobile phase ratio, because different constituents were extracted for each ternary mixture solvent
and the number of peaks observed vary when different mobile phases are employed. In another work, Delaroza and
Scarminio31 used the same strategy that results in similar conclusion, but in this paper, the authors applied a constraint
to the mixture design used to optimize the mobile phase.
On the other hand, there are cases where it is important to evaluate parameters that are dependent (mixture) and
independent (factorial) for the same analytical method development.16,20,21,34 Santos et al16 optimized an ultrasound
assisted extraction procedure employing the latter design. Initially, they evaluated three acids using a mixture design,
and then the optimized solution was used to evaluate three independent factors from ultrasound extraction: time,
concentration, and particle size. The authors found a condition that satisfied their needs, but it was not possible to
evaluate the interaction between these factors, which may be important. Næs, Færgestad, and Cornell34 compared
the results with the design evaluated separately and when they are evaluated simultaneously in a same modeling.
Only in the last case, it was possible to identify the importance of the interaction of one independent with the depen-
dent factors. Næs's work does not proposed an optimization of an analytical method but performed an interesting
comparison with the strategy of modeling simultaneously dependent and independent factors and pointed the advan-
tages of this strategy.
Orlandini et al20,21 proposed the simultaneous optimization of process variables and mixture components. In
both works, they used Quality by Design principles to access the optimum parameters to develop a capillary electro-
phoresis method searching for the highest resolution in the smallest time of analysis. The authors highlighted the
importance of this strategies once there are significant interactions between the parameters of both mixture and
process experimental designs, which were necessary to find the region where the risk of failure of the
developed method is low. In the first work, the authors explored the full analysis of all parameters and theirs inter-
actions.20 However, the main disadvantage of this approach is the requirement high number of experiments (62) to
be made. In the other work,21 the authors evaluated the same strategy for a screening analysis. They pointed out
that the main advantage of using this approach for screening analysis is the reduced number of experiments. In this
case, they used 17 experiments for the screening analysis and 36 for the RSM after remove one nonsignificant
parameter.
Alcaraz et al9 developed a capillary electrophoretic method for the determination of six quinolones in water sam-
ples. This work was commented above in “Mixture-process and mixture-mixture designs” section.

4 | I L L U S T R A T I V E L I T E R A T U R E EX A M P L E

In this section, the development of a pretreatment process for extraction of several compounds present in poultry litter
will be shown and discussed.40 This application will be explained and illustrated to furnish guidelines for the different
steps in developing a mixture design, including data treatments, and model validation. In the mentioned work, the
simultaneous extraction efficiency of the 21 veterinary drugs was enhanced by implementation of a simplex-centroid
design with combinations of the three components of the extracting mixture: methanol (MeOH), acetonitrile (ACN),
and sodium phosphate buffer 10 mmol L–1 pH = 3.50.
14 of 19 AZCARATE ET AL.

4.1 | Mixture design building

Herein, the optimal proportion of three components of a mixture employed in poultry litter sample preparation was
evaluated in order to improve the simultaneous extraction efficiency of the 21 veterinary drugs for their determination
through liquid chromatography with fluorescence and UV detection. The chromatographic procedure optimization has
been described in detail in a previous work.40
In general terms, a mixture experimental design is applied when input variables are strongly correlated, and the
response varies as a function of the relative proportions of its components.53 Otherwise, the design space must be
defined by the low and high levels of each factors being usually from 0 (the factor is not present) to 1 (only that factor is
present) in the mixture.
Considering the aims of this study, the simplex-lattice mixture design (SLD) is a proper design to apply in order to
assess the optimal composition of this extraction mixture. Thus, the SLD is built from different combinations of the
three components of the extracting mixture, that is, MeOH, ACN, and sodium phosphate buffer 10 mmol L–1 pH = 3.50.
In this case, the obtained experimental region results in a triangular grid of runs, which can be observed in Figure 9.
Several spots can be seen in the figure, indicating different experimental points. Moreover, each star point indicates the
replicates, which were made in order to evaluate lack of fit of the model. All the experiments and the different levels of
each studied factor are displayed in Table 1.
To carry out each experiment, 5.0 mL of each solvent combination were added to 1.00 g of PL containing the
analytes. After injection into the HPLC system, the extraction efficiency of each analyte was considered as response and
calculated by comparing the areas of the extracted analytes with those obtained for the standard solutions. The experi-
ments were performed in randomized order to ensure the independence of the results and minimize the effects of
uncontrolled factors.

4.2 | Modeling

From the obtained data matrix, including the responses corresponding to the different levels of each factor, it is possible
the estimation of a mathematical model which defines the functional form of the relationship between the response (y)
and the independent variables (the factors). Thus, 21 individual polynomial equations were fitted. In this stage, it is nec-
essary to evaluate the data variability inasmuch as it is assumed in analysis of variance that the variances in the

TABLE 1 Experiments corresponding to the lattice-mixture design

Factors, k
a
Run MEOH, % ACN, % Buffer, %a
1 0 1 0
2 1 0 0
3 0.167 0.167 0.667
4 0.667 0.167 0.167
5 0.167 0.667 0.167
6 0.5 0 0.5
7 0 0 1
8 0.333 0.333 0.333
9 0 0.5 0.5
10 0.5 0.5 0
11 0 1 0
12 0.5 0.5 0
13 1 0 0
14 0 0 1
a –1
Buffer phosphate 10 mmol L pH = 3.50.
AZCARATE ET AL. 15 of 19

different experimental conditions are homogeneous and that they are uncorrelated with the means. If the distribution
of values within each experimental condition is skewed, and the means are correlated with the standard deviations,
then the obtained results can be erroneous. In order to stabilize the variances and to reduce or eliminate the correlation
between the means and standard deviations, different transformations were applied on some responses allowing a bet-
ter fit to the system. The type of transformation required for each response is indicated in Table 2.
Then, every response was fitted to a unique model (linear, quadratic, cubic, or especial cubic), applying the RSM.
The coefficients were calculated by backward multiple regression and validated by ANOVA tests. The model fittings for
the SLD responses resulting from the mixture are summarized in Table 2. ANOVA was carried out to justify the signifi-
cance and adequacy of the regression model fit. Statistical significance of the model was established at p ≤ 0.05. The
obtained model F-values for each response indicate that the models were highly adequate and significant. Model valida-
tion by the ANOVA test application produced p-values below 0.05 with lack of fit between 0.1242 and 0.8512, being not
significant. Besides, the variation coefficients (ranged between 0.95% and 14.97%) alongside adjusted R2 (between
0.0625 and 0.8512) values indicated that the models can satisfactorily explain the data variability (data not shown).

4.3 | Mixture optimization procedure

Subsequently, in order to confirm the optimum conditions for the percentages of each solvent in the mixture, an optimi-
zation phase was performed. Thus, model fittings for the responses of the SLD were used to establish the optimum mix-
ture for the simultaneous extraction of the 21 veterinary drugs. For this optimization stage, each individual response
was maximized mining to find the optimal mixture for the maxima response values (see Equation 6).

TABLE 2 Model fitting for each response

ANOVA p Valuea

Response Model Transformation Model Lack of Fit Recoveryb Standard Deviationb


% TMP Cubic Inverse square (lambda = 0.5) <0.0001 0.4464 55.9 5.0
% IMI Cubic None <0.0001 0.1648 76.2 0.3
% CLE Quadratic Logarithm 0.0002 0.3285 109.7 2.7
% ENR Quadratic None <0.0001 0.1648 25.5 1.7
% DIF Quadratic Logarithm (lambda = 0.711) <0.0001 0.1242 60.6 1.3
% CLT Lineal Logarithm (lambda = 0.331) <0.0001 0.4153 23.4 1.8
% CRP Quadratic Inverse square (lambda = –0.5) <0.0001 0.5770 69.0 1.1
% CFT Lineal None <0.0001 0.1764 55.6 2.2
% PRT Cubic Logarithm <0.0001 0.3032 111.9 4.6
% FLU Lineal Inverse square (lambda = –0.5) 0.027 0.2353 24.6 1.8
% PNS Quadratic Logarithm <0.0001 0.4500 53.0 1.9
% MBT Quadratic None <0.0001 0.3125 78.4 0.9
% ALB Quadratic None <0.0001 0.7280 36.9 0.4
% FEN Quadratic None <0.0001 0.7114 59.2 1.8
% DIA Quadratic Power (lambda = –1.44) <0.0001 0.3746 108.2 3.7
% FNX Lineal None 0.0420 0.7256 93.1 3.8
% DNC Quadratic Logarithm <0.0001 0.4895 33.9 1.9
% IBU Cubic Inverse square <0.0001 0.6007 50.3 1.4
% DIC Lineal None 0.0004 0.2956 63.2 0.9
% PRO Special cubic None 0.0063 0.0625 40.9 2.8
% BET Cubic Inverse square (lambda = –0.5) <0.0001 0.8512 36.1 1.3
a
p values less than 0.05 indicate significance.
b
Result obtained with the suggested optimal condition.
16 of 19 AZCARATE ET AL.

Following the conditions previously discussed, the optimization procedure was carried out. Thus, the 21 contour
plots obtained for each response were simultaneously analyzed and the overlaying response surface through the global
desirability function (D) was acquired. The joint acceptability region as a function of the proportions of MeOH, ACN,
and phosphate buffer 10 mmol L–1 pH = 3.50 is shown in Figure 10 by a response surface contour plot for yield for the
21 analytes.
The global D produced a maximum value of 0.611 for a mixture of MeOH and phosphate buffer 10 mmol L–1
pH = 3.50, whose proportions were 33 and 67 v/v, respectively. The individual response values corresponding to the
value of D and their respective confidence intervals are depicted in Table 2. According to the performed fitting, the
acquired desirability is highly acceptable, taking into account the large number of responses being simultaneously opti-
mized (Figure 11).

F I G U R E 1 0 Experimental region of the simplex-lattice


mixture design for three component indicating experimental points
(red points) and replicates (star points)

F I G U R E 1 1 Response surface contour plot curve for the


desirability function combining the whole extraction yields
AZCARATE ET AL. 17 of 19

4.4 | Prediction at the optimum

Afterward, the suggested values during the optimization procedure should be experimentally corroborated. Thus, addi-
tional experimental confirmation in the optimal conditions were assessed through four replicates. The obtained experi-
mental results in the optimal conditions were compared statistically with the theoretical results in order to validate of
predictive model. The recoveries values for all responses were in close agreement with the predicted results presenting
errors lowers than 10% (data not shown).

5 | C ON C L US I ON S

In the present review, the application of mixture experiments for optimization of analytical methods was revised and
discussed. It could be seen that several strategies were published for many applications, where mixture experiments
were evaluated for a single or multiple responses optimization, and using constraints when necessary.
Interestingly, mixture experiments are still under-explored for optimization of analytical methods; although, an
important number of publications regarding the analysis applications in different analytical areas such as food, drugs,
bioanalytical, fingerprint determinations, among others, have been reported. The literature survey showed that two
main fields have been explored: sample preparation procedures and liquid chromatography mobile phase optimization.
It is expected that the use of RSM for mixture experiments grows in the future.

ACK NO WLE DGE MEN TS


The authors are grateful to CONICET (Consejo Nacional de Investigaciones Científicas y Técnicas, Project PIP-2015
No. 0111) and ANPCyT (Agencia Nacional de Promoción Científica y Tecnológica, Projects PICT 2017-0340) for finan-
cial support.

ORCID
Héctor C. Goicoechea https://orcid.org/0000-0001-7145-0082

R EF E RE N C E S
1. Adams F. From nano to yocto, and beyond? Talanta. 2011;85(3):1230-1232. https://doi.org/10.1016/j.talanta.2011.06.064
2. Motomizu S. Increasing demands on analytical chemistry and chemical analysis. Talanta. 2011;85(5):2251-2253. https://doi.org/10.1016/
j.talanta.2011.08.025
3. Araujo PW, Brereton RG. Experimental design III. Quantification. TrAC - Trends Anal Chem. 1996;15(2):85-89. https://ac-els-cdn-com.
scd-rproxy.u-strasbg.fr/0165993695000860/1-s2.0-0165993695000860-main.pdf?_tid=8d8118f5-e17e-4b7b-9137-8d131037c49d&acdnat=
1527235351_8b935c1205cdb0f95b0c731def2cc955
4. Myers RH, Montgomery DC. Response Surface Methodology: Process and Product Optimization Using Designed Experiments. Third
ed. New York: Wiley; 2009.
5. Leardi R. Experimental design in chemistry: a tutorial. Anal Chim Acta. 2009;652(1-2):161-172. https://doi.org/10.1016/j.aca.2009.06.015
6. Bruns RE, Scarminio IS, Neto BB. Statistical design—chemometrics. First Ed Amsterdam: Elsevier. 2006. https://doi.org/10.1007/
springerreference_6379
7. Vera Candioti L, De Zan MM, Cámara MS, Goicoechea HC. Experimental design and multiple response optimization. Using the desir-
ability function in analytical methods development. Talanta. 2014;124:123-138. https://doi.org/10.1016/j.talanta.2014.01.034
8. Design Expert 8.0.5, Stat-Ease Inc. 2010.
9. Alcaráz MR, Vera-Candioti L, Culzoni MJ, Goicoechea HC. Ultrafast quantitation of six quinolones in water samples by second-order
capillary electrophoresis data modeling with multivariate curve resolution-alternating least squares. Anal Bioanal Chem. 2014;406(11):
2571-2580. https://doi.org/10.1007/s00216-014-7657-3
10. de Almeida AA, Scarminio IS. Statistical mixture design optimization of extraction media and mobile phase compositions for the charac-
terization of green tea. J Sep Sci. 2007;30(3):414-420. https://doi.org/10.1002/jssc.200600321
11. Scheffé H. Experiments with mixtures. J R Stat Soc Ser B. 1958;20(2):344-360. https://doi.org/10.1111/j.2517-6161.1958.tb00299.x
12. Dingstad G, Egelandsdal B, Næs T. Modeling methods for crossed mixture experiments—a case study from sausage production. Chemom
Intel Lab Syst. 2003;66(2):175-190. https://doi.org/10.1016/S0169-7439(03)00031-5
13. Zupan J, Gasteiger J. Neural Networks in Chemistry and Drug Design. second ed. Weinheim: Wiley-VCH; 1999.
14. Haykin S. Neural Networks. A Comprehensive Foundation. Second ed. Prentice-Hall, Upper Saddle River, NJ; 1999.
15. Derringer G, Suich R. Simultaneous optimization of several response variables. J Qual Technol. 1980;12(4):214-219. https://doi.org/10.
1080/00224065.1980.11980968
18 of 19 AZCARATE ET AL.

16. Santos WPC, Castro JT, Bezerra MA, Fernandes AP, Ferreira SLC, Korn MGA. Application of multivariate optimization in the develop-
ment of an ultrasound-assisted extraction procedure for multielemental determination in bean seeds samples using ICP OES. Microchem
J. 2009;91(2):153-158. https://doi.org/10.1016/j.microc.2008.10.001
17. Santos HM, Coutinho JP, Amorim FAC, et al. Microwave-assisted digestion using diluted HNO3 and H2O2 for macro and microele-
ments determination in guarana samples by ICP OES. Food Chem. 2019;273(December 2017):159-165. https://doi.org/10.1016/j.
foodchem.2017.12.074.
18. de Pinho GP, Neves AA, de Queiroz MELR, Silvério FO. Pesticide determination in tomatoes by solid-liquid extraction with purification
at low temperature and gas chromatography. Food Chem. 2010;121(1):251-256. https://doi.org/10.1016/j.foodchem.2009.11.080
19. Moreira LS, Chagas BC, Pacheco CSV, et al. Development of procedure for sample preparation of cashew nuts using mixture design and
evaluation of nutrient profiles by Kohonen neural network. Food Chem. 2019;273:136-143. https://doi.org/10.1016/j.foodchem.2018.
01.050
20. Orlandini S, Pasquini B, Stocchero M, Pinzauti S, Furlanetto S. An integrated quality by design and mixture-process variable approach
in the development of a capillary electrophoresis method for the analysis of almotriptan and its impurities. J Chromatogr A. 2014;1339:
200-209. https://doi.org/10.1016/j.chroma.2014.02.088
21. Orlandini S, Pasquini B, Caprini C, et al. A comprehensive strategy in the development of a cyclodextrin-modified microemulsion elec-
trokinetic chromatographic method for the assay of diclofenac and its impurities: mixture-process variable experiments and quality by
design. J Chromatogr A. 2016;1466:189-198. https://doi.org/10.1016/j.chroma.2016.09.013
22. Rozet E, Lebrun P, Hubert P, Debrus B, Boulanger B. Design spaces for analytical methods. TrAC - Trends Anal Chem. 2013;42:157-167.
https://doi.org/10.1016/j.trac.2012.09.007
23. Imam SS, Aqil M, Akhtar M, Sultana Y, Ali A. Optimization of mobile phase by 3 2 -mixture design for the validation and quantification
of risperidone in bulk and pharmaceutical formulations using RP-HPLC. Anal Methods. 2014;6(1):282-288. https://doi.org/10.1039/
c3ay41562g
24. Beg S, Katare OP, Singh B. Formulation by design approach for development of ultrafine self-nanoemulsifying systems of rosuvastatin
calcium containing long-chain lipophiles for hyperlipidemia management. Colloids Surfaces B Biointerfaces. 2017;159:869-879. https://
doi.org/10.1016/j.colsurfb.2017.08.050
25. Zhang X, Hu C. Application of quality by design concept to develop a dual gradient elution stability-indicating method for cloxacillin
forced degradation studies using combined mixture-process variable models. J Chromatogr A. 2017;1514:44-53. https://doi.org/10.1016/j.
chroma.2017.07.062
26. Nano RMW, Bruns RE, Ferreira SLC, Baccan N, Cadore S. Statistical mixture design development of digestion methods for Oyster tissue
using inductively coupled plasma optical emission spectrometry for the determination of metallic ions. Talanta. 2009;80(2):559-564.
https://doi.org/10.1016/j.talanta.2009.07.025
27. Teglia CM, Gil García MD, Galera MM, Goicoechea HC. Enhanced high-performance liquid chromatography method for the determina-
tion of retinoic acid in plasma. Development, optimization and validation. J Chromatogr A. 2014;1353:40-48. https://doi.org/10.1016/j.
chroma.2014.01.013
28. Didier C, Etcheverrigaray M, Kratje R, Goicoechea HC. Crossed mixture design and multiple response analysis for developing complex
culture media used in recombinant protein production. Chemom Intel Lab Syst. 2007;86(1):1-9. https://doi.org/10.1016/j.chemolab.2006.
07.007
29. Kanthiah S, Kannappan V. D-optimal mixture design optimization of an HPLC method for simultaneous determination of commonly
used antihistaminic parent molecules and their active metabolites in human serum and urine. Biomed Chromatogr. 2017;31(8):1-12.
https://doi.org/10.1002/bmc.3932
30. Borges CN, Bruns RE, Almeida AA, Scarminio IS. Mixture-mixture design for the fingerprint optimization of chromatographic mobile
phases and extraction solutions for Camellia sinensis. Anal Chim Acta. 2007;595(1-2):28-37. https://doi.org/10.1016/j.aca.2007.02.067
31. Delaroza F, Scarminio IS. Mixture design optimization of extraction and mobile phase media for fingerprint analysis of Bauhinia varie-
gata L. J Sep Sci. 2008;31(6-7):1034-1041. https://doi.org/10.1002/jssc.200700389
32. Chen J, Cao J, Gao W, Qi L, Li P. Environmentally friendly ionic liquid-in-water microemulsions for extraction of hydrophilic and lipo-
philic components from Flos Chrysanthemi. Analyst. 2013;138(20):5933-5941. https://doi.org/10.1039/c3an00957b
33. Bortoloti JA, Bruns RE, De Andrade JC, Vieira RK. Split-plot design optimization for trace determination of lead by anodic stripping
voltammetry in a homogenssseous ternary solvent system. Chemom Intel Lab Syst. 2004;70(2):113-121. https://doi.org/10.1016/j.
chemolab.2003.09.004
34. Næs T, Færgestad EM, Cornell J. A comparison of methods for analyzing data from a three component mixture experiment in the pres-
ence of variation created by two process variables. Chemom Intel Lab Syst. 1998;41(2):221-235. https://doi.org/10.1016/S0169-7439(98)
00056-2
35. Azcarate SM, Langhoff LP, Camiña JM, Savio M. A green single-tube sample preparation method for wear metal determination in lubri-
cating oil by microwave induced plasma with optical emission spectrometry. Talanta. 2019;195:573-579. https://doi.org/10.1016/j.
talanta.2018.11.045
36. Costa VC, Silva EGP, Lima DC, et al. Use of mixture design with minimal restrictions to optimize an extraction procedure employing
diluted acids assisted by ultrasound and microwave for nutrient element determination in vegetal samples. J Braz Chem Soc. 2018;29(6):
1189-1198. https://doi.org/10.21577/0103-5053.20170214
AZCARATE ET AL. 19 of 19

37. De Andrade JC, De Almeida AM, Coscione AR, Aleixo LM. Determination of molybdenum in steel by adsorptive stripping voltammetry
in a homogeneous ternary solvent system. Analyst. 2001;126(6):892-896. https://doi.org/10.1039/b101068a
38. Morris V, Hughes J, Marriott P. Spherical coordinate representations of solvent composition for liquid chromatography method develop-
ment using experimental design. J Chromatogr A. 2003;1008(1):43-56. https://doi.org/10.1016/S0021-9673(03)01021-5
39. de Albuquerque Mendes MK, dos Santos Oliveira CB, Veras MD, et al. Application of multivariate optimization for the selective extrac-
tion of phenolic compounds in cashew nuts (Anacardium occidentale L.). Talanta. 2019;205:120100. https://doi.org/10.1016/j.talanta.
2019.06.100
40. Teglia CM, Peltzer PM, Seib SN, Lajmanovich RC, Culzoni MJ, Goicoechea HC. Simultaneous multi-residue determination of twenty
one veterinary drugs in poultry litter by modeling three-way liquid chromatography with fluorescence and absorption detection data.
Talanta. 2017;167:442-452. https://doi.org/10.1016/j.talanta.2017.02.030
41. Jin H, Lao YM, Zhou J, Zhang HJ, Cai ZH. Simultaneous determination of 13 carotenoids by a simple C18 column-based ultra-high-
pressure liquid chromatography method for carotenoid profiling in the astaxanthin-accumulating Haematococcus pluvialis.
J Chromatogr A. 2017;1488:93-103. https://doi.org/10.1016/j.chroma.2017.01.088
42. Pauli ED, Barbieri F, Garcia PS, et al. Detection of ground roasted coffee adulteration with roasted soybean and wheat. Food Res Int.
2014;61:112-119. https://doi.org/10.1016/j.foodres.2014.02.032
43. Ladeira C, Sanches F, Lima D, et al. Multi-response optimisation of the extraction solvent system for phenolics and antioxidant activities
from fermented soy flour using a simplex-centroid design. Food Chem. 2016;197:175-184. https://doi.org/10.1016/j.foodchem.2015.10.124
44. Passari LMZG, Scarminio IS, Bruns RE. Experimental designs characterizing seasonal variations and solvent effects on the quantities of
coumarin and related metabolites from Mikania laevigata. Anal Chim Acta. 2014;821:89-96. https://doi.org/10.1016/j.aca.2014.03.003
45. Passari LMZG, Scarminio IS, Marcheafave GG, Bruns RE. Seasonal changes and solvent effects on fractionated functional food compo-
nent yields from Mikania laevigata leaves. Food Chem. 2019;273:151-158. https://doi.org/10.1016/j.foodchem.2017.12.075
46. Justo R, Afonso S, Bruns RE, Scarminio IS. Open column, reversed-phase high-performance liquid chromatography with diode array
detection and chemometric strategy for investigation of metabolic fingerprints of complex systems. Anal Methods. 2014;6(24):9567-9574.
https://doi.org/10.1039/c4ay01985g
47. Xie H, Wang X, Zhang L, et al. Monitoring metabolite production of aflatoxin biosynthesis by orbitrap fusion mass spectrometry and a
D-optimal mixture design method. Anal Chem. 2018;90(24):14331-14338. https://doi.org/10.1021/acs.analchem.8b03703
48. Didier C, Forno G, Etcheverrigaray M, Kratje R, Goicoechea H. Novel chemometric strategy based on the application of artificial neural
networks to crossed mixture design for the improvement of recombinant protein production in continuous culture. Anal Chim Acta.
2009;650(2):167-174. https://doi.org/10.1016/j.aca.2009.07.051
49. Zheng H, Ni Z, Cai A, et al. Balancing metabolome coverage and reproducibility for untargeted NMR-based metabolic profiling in tissue
samples through mixture design methods. Anal Bioanal Chem. 2018;410(29):7783-7792. https://doi.org/10.1007/s00216-018-1396-9
50. Pauli ED, Scarminio IS, Tauler R. Analytical investigation of secondary metabolites extracted from Camellia sinensis L. leaves using a
HPLC-DAD-ESI/MS data fusion strategy and chemometric methods. J Chemometr. 2016;30(2):75-85. https://doi.org/10.1002/cem.2772
51. Costa LM, Ferreira SLC, Nogueira ARA, Nóbrega JA. Use of factorial design for optimization of microwave-assisted digestion of lubricat-
ing oil. J Braz Chem Soc. 2005;16(6 A):1269-1274. https://doi.org/10.1590/S0103-50532005000700027
52. Hibbert DB. Experimental design in chromatography: A tutorial review. J Chromatogr B. 2012;910:2-13. https://doi.org/10.1016/j.
jchromb.2012.01.020
53. Montgomery DC. Design and Analysis of Experiments. fifth edit. New York: Arizona State University, John Wiley., Inc.; 2000.

How to cite this article: Azcarate SM, Pinto L, Goicoechea HC. Applications of mixture experiments for
response surface methodology implementation in analytical methods development. Journal of Chemometrics.
2020;e3246. https://doi.org/10.1002/cem.3246

You might also like