You are on page 1of 11

Trends in Analytical Chemistry, Vol.

42, 2013 Trends

Design Spaces for analytical


methods
E. Rozet, P. Lebrun, B. Debrus, B. Boulanger, P. Hubert

Since the adoption of the ICH Q8 document concerning the development of pharmaceutical processes following a Quality by
Design (QbD) approach, there have been many discussions on the opportunity for analytical method developments to follow a
similar approach. A key component of the QbD paradigm is the definition of the Design Space (DS) of analytical methods where
assurance of quality is provided. Several DSs for analytical methods have been published, stressing the importance of this
concept.
This article aims to explain what an analytical method DS is, why it is useful for the robust development and optimization of
analytical methods and how to build such a DS. We distinguish the usual mean response surface approach, overlapping mean
response surfaces and the desirability function as only they correctly define a DS. We also review and discuss recent publications
assessing the DS of analytical methods.
ª 2012 Elsevier Ltd. All rights reserved.

Keywords: Analytical method development; Design of Experiments; Design Space (DS); Desirability function; Mean response surface; Optimization;
Probability map; Quality assurance; Quality by Design (QbD); Robustness

1. Introduction
E. Rozet*, P. Lebrun, P. Hubert variables involved in formulation and
Analytical Chemistry manufacturing processes influence the
Laboratory, CIRM, University of The concept of quality by design (QbD) has
quality of the final product. Examples of
Liège, Avenue de lÕHôpital 1, been recently adopted in the pharmaceuti-
applications of QbD to pharmaceutical pro-
B36, B-4000 Liège, Belgium cal industry through several initiatives
cesses can be found in numerous case
{e.g., FDAÕs cGMP for the 21st Century [1],
B. Debrus studies in the literature [e.g., 7–18].
Process Analytical Technology (PAT) [2],
School of Pharmaceutical It has been emphasized that analytical
Sciences, University of Geneva, and the new regulatory documents, ICH Q8
procedures are also processes, and that
Switzerland [3], Q9 [4] and Q10 [5]}. The general aim is
QbD could and should be implemented for
to switch from the quality by testing (QbT)
B. Boulanger
the development of analytical procedures,
paradigm previously implemented in the
Arlenda s.a., Liège, Belgium as they are non-negligible components of
pharmaceutical industry to a development
the global pharmaceutical-product process
aimed at improving the understanding of
[19]. Several examples and discussions
the processes and the products and hence
about the development of analytical
improving product quality, process
methods following the QbD concept are
efficiency and regulatory flexibility.
available, e.g., [20–22].
QbD is not new and involves many quality
A key component of the development of
and statistical tools and methods (e.g., sta-
analytical procedure using QbD is what has
tistical designs of experiments, multivariate
been called the Design Space (DS). The aim of
statistics, Six Sigma methodologies, and
this review article is to focus on the DS of
statistical quality control). In order to raise
analytical methods. First, we summarize the
the quality of pharmaceutical products, it
general framework of QbD development of
has only recently been recognized that
analytical methods to situate the specific
increasing the testing of final products (i.e.
place of DS in this strategy. Then, we provide
QbT) is not adequate [6]. Instead, to increase
explanation of what is DS, why it is useful to
the quality of pharmaceutical products,
define the DS of analytical methods and how
*
Corresponding author. quality must be built into the products (i.e.
to correctly compute and build the DS. To
Tel.: +32 4366 4320; QbD), as already done in many other
achieve this, we also review recent publica-
Fax: +32 4366 4317; industries. It requires understanding of how
E-mail: eric.rozet@ulg.ac.be
tions assessing the DS of analytical methods.

0165-9936/$ - see front matter ª 2012 Elsevier Ltd. All rights reserved. doi:http://dx.doi.org/10.1016/j.trac.2012.09.007 157
Trends Trends in Analytical Chemistry, Vol. 42, 2013

Figure 1. Typical steps of Quality by Design development of analytical methods highlighting the place of the concept of analytical method design
space. ATP, Analytical Target Profile; CQA, Critical Quality Attribute; FMEA, Failure Mode and Effect Analysis; SST, System Suitability Tests.

2. Place of Design Space in QbD analytical method steps illustrated in Fig. 1 have to be performed. This
development section summarizes the main steps usually required.
Deeper details can be found in the exhaustive review of
To start the development of a QbD compliant analytical Vogt and Kord [19] as well as in the following docu-
method and finally reach the definition of its DS, several ments [20–23].

158 http://www.elsevier.com/locate/trac
Trends in Analytical Chemistry, Vol. 42, 2013 Trends

The first step is to define the intended purpose of the resolution value between a critical pair, the acceptable
analytical method. This has been called the Analytical value for tailing peaks, the maximum acceptable value
Target Profile (ATP) [20–22,24] which is similar to expressed in RSD for the repeated analysis of a standard
the Quality Target Product Profile (QTPP) for solution, the minimum value of the determination coeffi-
pharmaceutical-process. This ATP will be a set of cri- cient (R2) of a standard curve, and so on. We recommend
teria that defines what will be measured, in which reading the following documents as examples of how to
matrix, over what concentration range, and the re- obtain such information from the application of the DoE
quired performance criteria of the method, together with methodology and the resulting analytical DS [36,37].
specifications for these last ones. These performance Another methodology that can be implemented as a
criteria can be called Critical Quality Attributes (CQAs) control strategy for analytical methods is to use statistical
of the analytical method to keep a nomenclature close to control charts such as Shewhart X  R ones [32]. Out of
the one used in pharmaceutical-process development. control methods can be detected efficiently and corrective
Examples of ATP for the interested readers are available actions realized by following the daily performance of
[20–22,24]. analytical methods on such charts.
The method under development will then follow a risk
assessment [19]. The analytical method can be decom- 3. What is Design Space of analytical methods
posed in a flow-chart highlighting the main steps of the
procedure from sample preparation to data analysis. This As stated above, the main interest of this work is to focus
allows identifying parameters that should be studied on the concept of DS applied to analytical methods. In
during the risk assessment. ICH pharmaceutical-development guideline Q8 [3], the
Typically, the first step in the risk assessment is to build DS is defined as ‘‘the multidimensional combination and
a fishbone diagram (or Ishikawa diagram) to identify interaction of input variables (e.g., material attributes) and
further potential factors and relate them to the require- process parameters that have been demonstrated to provide
ment included in the ATP [25]. This diagram classifies assurance of quality’’. Therefore, the multidimensional
risks in groups related to instrumentation, materials, combination and interaction of input variable corre-
methods, chemicals and reagents, measurements, human sponds to a subspace, so-called the DS, where assurance
factors, environmental issues (e.g., laboratory tempera- of quality has been proved. The DS is necessarily
ture, relative humidity, and light) [26,27]. encompassed within the experimental domain, which is
Having defined the risk factors, they can be ranked and the multidimensional space formed by the factor ranges
prioritized using dedicated approaches {e.g., Failure Mode used during method development. The main concept
and Effect Analysis (FMEA) [27–29]}. The high noise risk lying behind the ICH Q8 definition of DS is assurance of
factors (e.g., analysts, equipment, days of analysis, bat- quality (also known as quality-risk management). This
ches of reagent or of materials, and sample batches) can document also states that ‘‘Working within the design
subsequently be evaluated experimentally using Mea- space is not considered as a change’’.
surement System Analysis approaches [30–32] to study In the framework of method development, DS can be
their impact on the repeatability and the reproducibility of considered as a zone of theoretical robustness as no drastic
the method, and eventually corrective measures may be changes in the levels of the CQAs of the method should be
implemented to reduce their impact. observed. Hence, to define an analytical DS, a wisely se-
High-risk instrumental parameters can also be assessed lected number of factors, also called critical process
experimentally using statistical Design of Experiments (DoE) parameter (CPP)-operating factors (e.g., gradient time in
methodology. These parameters are the critical process chromatography, and concentration of reagents in
parameters (CPPs). From the DoE results and their inter- immunoassays) that impact on the analytical technique
pretation, the DS of the analytical method will be obtained. under development have to be studied simultaneously.
We provide further information concerning the con- Usually, the CPPs are obtained from a risk analysis and a
cept of DS for analytical methods in the remaining sec- prioritization strategy, as explained in Section 2. The
tions of this article, as it is its core aim. Formal analytical juxtaposition of univariate studies of the involved factors
method validation is then needed [19,33-35]. cannot determine an analytical DS properly, as potential
Nonetheless, the development of QbD analytical meth- interactions of the factors involved cannot be studied.
ods does not end with the DS. A control strategy of the These interactions may result in synergic or antagonistic
method has to be implemented to assure that the method effects on the responses measured and therefore modify
will perform as intended on a routine basis. Here also, the delimitation of the DS.
elements from the MSA and/or from the DS can be used to The analytical DS is finally a multivariate domain of
select responses that have to be monitored at each ana- input factors ensuring that critically chosen responses are
lytical run. These responses that will be implemented in included within predefined limits with an acceptable level
the control strategy are known as system-suitability tests of probability. The analytical method is therefore no longer
or validity tests. They can be the definition of a minimum- defined by a single point in the space of its operating

http://www.elsevier.com/locate/trac 159
Trends Trends in Analytical Chemistry, Vol. 42, 2013

parameters (e.g., one value of wavelength, one value of and to review all historical information available on the
proportion of organic modifier, or one value of pH of the analytical method under development, previously devel-
mobile phase). The analytical method is identified by a oped methods that are closely related, and the literature
range of operating conditions that are defined by the and scientific information available on the subject.
analytical DS. Several authors have proposed to call ana-
lytical DS the ‘‘Method Operable Design Region’’ (MODR; 5.1. Critical quality attributes and modeled responses
[20–23]) in order to distinguish it from the process DS. The first step is to define the CQAs of the analytical method
This name, MODR, clearly states that the analytical pro- that have to be included into the ATP. These CQAs are the
cedure is defined as a multivariate range of values of responses that are measured to judge the quality of the
operating parameters where the analytical method pro- developed analytical methods. CQAs are defined as ‘‘a
vides quality outputs with adequate probability. physical, chemical, biological or microbiological property
or characteristic that should be within an appropriate
4. Why Design Space for analytical methods? limit, range, or distribution to ensure the desired product
quality’’ [3]. For separative analytical methods (e.g.,
The key meaning of QbD and hence of analytical DS is to chromatography), the CQAs can be related to the method
increase the understanding of the analytical procedures selectivity (e.g., the resolution (RS) criteria). Additional
themselves and to understand better the relationship CQAs can be the run time of the analysis, the precision of
between the analytical procedures and the capability of the analytical method, the lower limit of quantification or
the process that includes them [38]. The analytical DS, as the dosing range of the analytical method. Sometimes
defined in the ICH Q8 document, allows the inclusion of these CQAs can be directly modeled through a multivar-
the measurement uncertainty to assess whether the iate (non-)linear model. However in some situations, the
analytical procedures will provide assurance of quality. modeled responses may different from the CQAs. The
The DS obtained for an analytical procedure will ensure CQAs are obtained after the modeling of these primary
that it will be a useful analytical method, allowing users responses.
to take adequate decisions with the results generated For chromatographic methods, the usual key CQA is
from it. The exercise of defining a DS for analytical resolution of the critical pair. However, resolution de-
methods allows users to determine the critical analytical pends upon the retention factor of the two chromato-
method parameters (corresponding to the CPPs in process graphic peaks involved, so several authors [40–43] have
literature) that load heavy weights on the CQAs of the proposed to model the retention factors instead of the
analytical methods. Control over the most important resolution. The resolution can subsequently be com-
factors can hence be judiciously implemented. The ranges puted from these modeled responses. Similarly, several
over which these method parameters can be varied are commercial software packages use the solvophobic the-
known at the end of the definition of the analytical DS. ory [44,45] or the linear solvent-strength theory [46,47]
This implies that the DS of an analytical method is a to model the retention factors and then derive the
measure of its robustness. In addition, the control strat- optimal resolution, depending on several factors.
egy of the analytical procedure can be implemented by
selecting system-suitability tests that are highly corre- 5.2. Experimental factors, ranges and levels
lated to the CQAs of the analytical procedure [20–22,37]. To obtain the DS of analytical methods, the choice of the
As moving within the DS is not considered change, experimental factors and their respective range is pri-
more flexibility for the analytical methods during its mordial. From the whole experimental design region, the
routine application is possible. Hence change controls factors and the ranges that will affect the responses most
will only be required when stepping outside the DS limits. have to be chosen [48]. On this domain, sometimes
The particular case of method transfer from laboratories called the knowledge space, formal designs of experi-
to laboratories is eased [39]. This is possible since adap- ments will be performed. This investigated knowledge
tation of the analytical procedure being transferred can space is a multidimensional space that needs to be large
be performed within its DS at the receiving site. enough to create response variations. These variations
Finally, the main reason to determine the DS of an should allow users to reach the minimum requirements
analytical procedure is that it provides information on of the CQAs. An example for chromatographic methods
how often the analytical procedure will meet its is RS >1.5. The number of factor levels determines the
requirement in order to provide reliable, useful data. polynomial degree of the multi-linear equation used to
model the responses. For example, if two levels are
5. How to define the Design Space of analytical selected, the modeling equation can only be linear.
methods Generally, if no prior information about the response
variation is known, preliminary experiments should be
In order to define the DS of analytical methods, several carried out to estimate the range and the magnitude of
steps have to be performed. The starting point is to gather variation of each factor.

160 http://www.elsevier.com/locate/trac
Trends in Analytical Chemistry, Vol. 42, 2013 Trends

5.3. Design of Experiments (DoEs) and response about the method under development and optimization.
modeling It should also allow estimation of the experimental error
The naı̈ve way to perform experiments to gain knowl- and assessment of the validity of the model tested.
edge about a process or to optimize it is to perform one- DoE can be split up into two main categories: screen-
factor-at-a-time (OFAT) designs. OFAT will generally ing designs and response-surface designs
require a higher number of experiments to estimate the
factors effect with good precision and their interactions 5.3.1. Screening designs. Screening designs estimate the
can rarely be estimated. effects of factors on selected responses. When too many
DoEs provide an effective, efficient approach to eval- factors (four or more) seem to affect the responses and
uate simultaneously the effects of factors and their have been revealed by the FMEA prioritization, these
interactions, and to model and to predict the relationship designs can be used to select those having the largest
between these factors and the CQAs or responses effects on the responses. The remaining significant fac-
[49–51]. The selected DoE needs to have good statistical tors will be studied in a subsequent DoE [e.g., method
properties (e.g., orthogonality and/or rotatability), and optimization (see below)].
should maintain the number of experiments as low as In the screening category of designs, well known are
possible. The possibility of expanding the design is also the Plackett and Burman designs that study factors at
an interesting property in order to extend the range of two levels. In liquid chromatography (LC), Plackett and
values of the factors, add new factors, or increase the Burman designs are also used to estimate the robustness
model complexity when starting to acquire knowledge of an optimal separation [48,52,53].

Figure 2. Mean response surface for run-time with respect to pH and % acetonitrile obtained from the central composite design applied to the
hypothetical analytical method example. The desired maximum run-time cannot exceed 20 min. Hatched region: experimental domain where
maximum run-time is at most 20 min.

http://www.elsevier.com/locate/trac 161
Trends Trends in Analytical Chemistry, Vol. 42, 2013

Other types of screening designs are two-level frac- However, most of the time, there are no theoretical
tional factorial ones, which generally do not allow models that include all the factors that may influence the
understanding of a process under investigation if it may responses and the analytical CQAs. In this case, empir-
include interactions and higher order effect terms. ical models can be fitted on the data obtained to link the
However they are very useful in selecting the most responses and the factors studied. This is usually per-
important factors that influence the selected responses of formed by fitting multiple linear equations of adequate
the analytical method under investigation. polynomial degree, related to the number of factors se-
lected. In some situations, it may also be required to fit
5.3.2. Response-surface designs. The second category of non-linear models.
DoE corresponds to designs used to predict and to optimize
the responses [54,55]. These DoEs are three-level, full fac- 5.4. Are mean response surfaces the Design Space?
torial designs, central composite designs, and Box-Benkhen When the experiments have been performed using a
[56] and Doehlert designs [57]. D-optimal designs can also response-surface design, the resulting model is generally
be selected in order to answer particular requirements (e.g., illustrated and interpreted through a mean response
constraints on the levels of factors, or specific models) [58]. surface or contour plot, as shown in Fig. 2. The mean
These designs are aimed at understanding the process response surface depicts the behavior of the response
under investigation. It involves understanding the rela- measured with respect to the range of factors assessed
tionship between the factors to assess the behavior of the and to the model fitted. Fig. 2 shows the mean response
response, and the effects on the response. These designs are surface for the maximum run-time of a hypothetical
used to find the combination of factors that predict the chromatographic method used to illustrate further key
optimal response with good precision. issues concerning analytical DS. For this example, the
More than two levels of each factor are usually re- true model that links the response modeled run-time (y)
quired in order to fit quadratic or higher order terms to the factors proportion of acetonitrile in the mobile
{e.g., when pH is a factor in LC, it may be required to phase (ACN) and pH of the aqueous buffer (pH) is sup-
study pH up to the third-order term: pH+pH2+pH3 [59]}. posed known:
Response-surface designs are key tools to define the DS
y ¼ b0 þ b1  ACN þ b2  pH þ b11  ACN 2 þ b22  pH 2
of analytical methods. They study a large experimental
domain, understanding the behavior of the responses ð1Þ
and the CQAs with respect to the studied factors, and 0 1
they provide a model to predict the value of the CQAs b0 ¼ 20
B C
within the range of these levels of factors. B b1 ¼ 2 C
B C
with B¼B
B b2 ¼ 3 C
C
5.3.3. Response modeling. The modeling of the responses B C
@ b11 ¼ 1 A
can be realized in two main ways. The first involves a
theoretical or mechanistic model that connects some of b22 ¼ 1
the factors to the responses {e.g., realized with software To represent the reality of method optimization, a
available to optimize chromatographic methods using the rotatable central composite design was defined involving
solvophobic theory or linear solvent-strength theory 13 experiments, as shown in Table 1. Values of maxi-
[44–47]}. mum run-time were then sampled from the following
normal distribution:
Table 1. Central composite design used as example for defining N(y, r), with r = 1 at the experimental conditions
the Design Space proposed by the DoE in Table 1.
The response-surface model defined in Equation (2) is
Experiment pH Acetonitrile % Run-time (min.)
then fitted to the data obtained and the corresponding
1 0 0 20.15 response surface is built, as illustrated in Fig. 2:
2 0.7071068 0.7071068 15.98
3 0 0 20.68 y^ ¼ b0 þ b1  ACN þ b2  pH þ b12  ACN  pH
4 0 1 14.68
5 0 0 19.62
þ b11  ACN 2 þ b22  pH 2 ð2Þ
6 0 1 20.94
It is generally thought that the DS can be obtained by
7 0 0 20.03
8 1 0 22.10 searching the range of values of the factors that show that
9 0.70710678 0.7071068 19.03 the response (or CQA) meets a pre-defined criteria (e.g., Rs
10 0.7071068 0.70710678 18.37 >1.5 or run time <20 min). The hatched region of Fig. 2
11 0.70710678 0.70710678 21.60 shows the values of the two factors having a run time less
12 0 0 18.67
than 20 min. It would be tempting to define this multi-
13 1 0 16.29
variate region as the DS for this hypothetical example.

162 http://www.elsevier.com/locate/trac
Trends in Analytical Chemistry, Vol. 42, 2013 Trends

Figure 3. Probability map showing the true probability of having a maximum run-time of at most 20 min applied to the hypothetical analytical
method example. Hatched region: experimental domain where the true probability to have a run-time of maximum 20 min is at least 90%.

However, the mean response surfaces do not give any is a region where process parameters ‘‘have been dem-
guarantee that the responses (or CQAs) will attain the onstrated to provide assurance of quality’’.
defined criteria with high probability [60]. Indeed, they By contrast, the hatched region of Fig. 3 shows the
only represent a region where the response is observed values of the factors pH and ACN that ensure that the
on average. In other words, there is one chance in two true probability of having a maximum run-time of
that the response will be on this mean response surface. 20 min is at least 90%, so using mean response surfaces
Fig. 3 shows the true probability to meet the criteria alone does not provide any assurance of quality.
required for the CQA maximum run-time <20 min. This In addition, when several CQAs are measured simul-
true probability can be obtained only because the true taneously, a desirability function or overlap in mean
underlying model is known here. As can be seen on response surfaces is generally used to find the optimal
Fig. 3, the probability to meet the specification is only conditions to reach simultaneously the predefined per-
about 50% at the edge of the supposed DS, where the formance criterion required for these responses (or CQAs)
maximum run-time is <20 min. It means that, when [61,62]. Here again, using such a classical methodology
observing the mean response surface corresponding to does not give any guarantee about the probability of
20 min of run-time, there is effectively one chance in two achieving them jointly at the mean optimal conditions
to reach the objective. It can be seen from Fig. 3 that this [63]. For example, when using two overlapping mean
supposed DS does not give a guarantee that the specifi- response surfaces that have each only 50% probability to
cation is reached. Indeed, for the contour corresponding attain the desired response (or CQA) level, the probability
to 20 min of run time, many conditions have a less than to reach simultaneously the required levels of the two
50% chance to reach this maximum run-time. This is far responses (or CQAs) is p = 0.5 · 0.5 = 0.25 = 25 sup-
from the DS requirement of the ICH Q8 that states that DS posing that they are independent! In this situation, the

http://www.elsevier.com/locate/trac 163
Trends Trends in Analytical Chemistry, Vol. 42, 2013

Figure 4. Probability map giving the estimated probability of having a run-time of at most 20 min obtained from the experiments defined by the
central composite design applied to the hypothetical analytical method example. Hatched region: the Design Space showing that the probability
of having a run-time of maximum 20 min is at least 90%.

level of quality, p, decreases according to the power of p, Several options can be implemented to reach this
the number of responses simultaneously studied (i.e. requirement. Bayesian modeling [64], Monte-Carlo
p = 0.5p). Similar pitfalls are present when using mean simulations [65] or bootstrapping techniques [66] can be
predicted global desirability functions, which ignore the performed to include uncertainty of the parameters of
correlation between the various responses, and neglect the models and to estimate the probability of meeting the
the prediction uncertainty and the uncertainty of model- specifications imposed on the CQAs.
parameter estimates [64]. Fig. 4 shows the DS obtained using a Bayesian approach
as proposed by Peterson et al. [63] or Lebrun et al. [7]. As can
5.5. Are probability maps the Design Space? be seen by the hatched region of Fig. 4, if it is required that
As the previous approaches related to mean response the specification over the CQA should be met with a proba-
surfaces do not provide any guarantee that the bility of at least 90%, the DS is far less than the one obtained
analytical method can meet the specifications with re- with the mean response surface. This means the specifica-
spect to the investigated CQAs with high probability, tion ‘‘run-time of at most 20 minutes’’ will be met in 90% of
other approaches should be implemented. These ap- runs. This measure of the assurance of quality is the prob-
proaches should take into account the model-parameter ability associated with the DS. We should also note that the
uncertainty and should provide information about how DS obtained on Fig. 4 is very close to the true probability map
often the specifications will be met. This is essential, of this hypothetical example, as shown in Fig. 3.
since ICH Q8 clearly requires it for a level of assurance Hence, when defining an analytical DS, the method-
guaranteeing that the specifications will be met. ology that is used should take into account the uncer-

164 http://www.elsevier.com/locate/trac
Trends in Analytical Chemistry, Vol. 42, 2013
Table 2. Details of the analytical Design Spaces published in the scientific literature

Analytical method Analytes Matrix Modeling approach Design Space type Ref.
UHPLC-UV Impurities and degradation product Tablets Chromatographic theory Mean response surface [70]
of ethinylestradiol
UHPLC-UV Dienogest, estradiol, ethinylestradiol, Cleaning validation Chromatographic theory Mean response surface [70]
finasterid, gestodene, levonorgestrel, samples
norethisterone acetate
UHPLC-UV 1-naphtol, duloxetine, related Spiked and stressed Chromatographic theory Mean response surface [70]
impurities and degradation products capsule samples
UHPLC-UV Bicalutamide and related impurities Tablets Chromatographic theory Mean response surface [70]
HPLC-UV Paracetamol, 4-hydroxy-3-methoxy na Empirical linear model and Mean response surface [71]
benzyl alcohol, DL-mandelic acid, Chromatographic theory
phthalic acid, p-hydroxyphenylacetic
acid, vanillic acid, m-
hydrophenylacetic acid, isovanillic
acid, benzyl alcohol and impurities
HPLC-UV Phthalic acid, vanillic acid, Synthetic mixture Empirical linear model and Mean response surface [72]
isovanillic acid, aspirin, furosemide, Chromatographic theory
doxepin, terbinafin, atorvastatin,
clopidogrel and related impurities
HPLC-UV Phthalic acid, vanillic acid, Synthetic mixture Chromatographic theory Mean response surface [73]
isovanillic acid, anthranilic acid,
vanillin, syringaldehyde, ferulic acid,
ortho vanillin, benzoic acid
UHPLC-UV 2 Active Pharmaceutical Ingredients Eye drop solution Chromatographic theory Mean response surface [74]
and 9 impurities
HPLC-UV 19 antimalarial drugs Synthetic mixture Empirical linear model Monte-Carlo Probability map [59]
HPLC-UV Diflunisal, Granisetron, Nifedipine, Synthetic mixture Empirical linear model Monte-Carlo Probability map [75]
Phenytoine, Sulfinpyrazone
HPLC-UV D9-tetrahydrocannabinol, D9- DifferentCannabis Empirical linear model Monte-Carlo Probability map [67]
tetrahydrocannabinolic acid A, products
cannabidiolic acid, cannabigerolic
acid, cannabidiol, cannabigerol,
cannabinol, D8-
tetrahydrocannabinol
http://www.elsevier.com/locate/trac

HPLC-UV Tertiary alkaloids Strychnos Empirical linear model Monte-Carlo Probability map [68]
usambarensis leaves
HPLC-UV Aprophine alkaloids Leaves of Empirical linear model Monte-Carlo Probability map [69]
Spirospermum
penduliflorum
Thouars
HPLC-UV Sulfide, sulfone, sulindac, E-sulindac Drug substance Empirical linear model Monte-Carlo Probability map [76]
HPLC-UV 9 unknown compounds Drug product Empirical linear model Monte-Carlo Probability map [77]
HPLC na na Empirical linear model Bayesian Probability map [64,78]

na: no data available.

Trends
165
Trends Trends in Analytical Chemistry, Vol. 42, 2013

tainty of the model parameters, the correlation of the initiate a regulatory post-approval change process. It is
responses studied and a measure of the assurance of then of core importance to demonstrate that the level of
attaining the quality target. quality required for the CQAs can be met with high
probability. Methodologies to achieve this aim are
6. Design Space of analytical methods available and can alleviate the false impression that risk
has been mitigated by using mean response surfaces or
There are several examples of DS of analytical methods similar approaches. Indeed, assurance of quality requires
in the literature, as can be seen in Table 2. All the a metric to measure how often quality will be achieved.
examples involved liquid chromatographic methods. Evidently, obtaining the DS of analytical methods is not
Most are conventional HPLC and four examples involve the end of the story. In particular, a control strategy of
UHPLC. The domains of application are nonetheless not the analytical method will have to be defined in order to
limited to the pharmaceutical industry. assure and to monitor its daily performance. The meth-
Of the 16 examples collected, three [67–69] aim to odology implemented to define the analytical DS will also
define the DS of analytical methods applied to analysis in help in defining an adequate control strategy. Finally,
plant materials. The DSs were obtained by using the the achievement of a DS is meaningless in itself if the DS
mean response-surface methodology in eight cases [70– is not completely included in a quality system that has a
74], hence failing to provide the demonstration of global risk-management plan.
‘‘assurance of quality’’ required in the ICHQ8 definition
of DS. Seven other cases used Monte-Carlo simulations
[59,75–77], and only one DS applied a Bayesian ap- Acknowledgments
proach [64,78]. The authors are very grateful to the anonymous
As all the examples reported involved chromato- reviewers for providing important comments that led to
graphic methods, the CQAs measured related to the significant improvements of this article. A research grant
ability of the methods to separate the various components from the Belgium National Fund for Scientific Research
of the samples analyzed. The resolution was used most (FRS-FNRS) to E. Rozet is gratefully acknowledged.
[64,70–74,78], while another one reported was separa-
tion (the difference between the time of the beginning of References
the second peak minus the time of the end of the previous [1] US Food and Drug Administration (FDA), Department of Health
and Human Services, Pharmaceutical Quality for the 21st Century
peak) [59,67–69,75–77]. Other CQAs that were used to
A Risk-Based Approach Progress Report, May 2007 (http://www.
determine the DS were total run time [59,64,78], signal- fda.gov/AboutFDA/CentersOffices/CDER/ucm128080.html).
to-noise ratio and tailing factor [64,78]. These last [2] US Food and Drug Administration (FDA), Guidance for industry
examples defined a DS that could simultaneously comply PAT-A framework for innovative pharmaceutical manufacturing
with specifications assigned to each individual CQA (e.g., and quality assurance, FDA, Washington, DC, USA, 2004.
[3] International Conference on Harmonization (ICH) of Technical
a resolution of the critical peak pair of at least 1.5 and a
Requirements for Registration of Pharmaceuticals for Human Use,
maximum run time of 15 min). Topic Q8 (R2): Pharmaceutical Development, ICH, Geneva,
All these examples showed that a DS can be built for Switzerland, 2009.
analytical methods. They also show the two main vi- [4] International Conference on Harmonization (ICH) of Technical
sions about DS: Requirements for Registration of Pharmaceuticals for Human Use,
Topic Q9: Quality Risk Management, ICH, Geneva, Switzerland,
(1) pseudo-DS based on mean response surface that only
2005.
gives information on the mean predicted quality; and, [5] International Conference on Harmonization (ICH) of Technical
(2) DS that accounts for uncertainty and correlation Requirements for Registration of Pharmaceuticals for Human Use,
and provides a level of assurance of quality, as Topic Q10: Pharmaceutical Quality System, ICH, Geneva, Swit-
defined by ICH Q8 document. zerland, 2008.
[6] R.A. Lionberger, S.L. Lee, L. Lee, A. Raw, L.X. Yu, AAPS J. 10
The fact that only DSs concerning chromatographic
(2008) 268.
methods were found in the literature does not imply that [7] P. Lebrun, F. Krier, J. Mantanus, H. Grohganz, M. Yang, E. Rozet,
the DS is restricted to such techniques. DS can be obtained B. Boulanger, B. Evrard, J. Rantanen, Ph. Hubert, Eur. J. Pharm.
for immunoassays or other bio-assays (e.g., PCR). Biopharm. 80 (2012) 226.
However, none was reported in the scientific literature. [8] H. Wu, M. White, M.A. Khan, Int. J. Pharm. 405 (2011) 63.
[9] T.R.M. De Beer, M. Wiggenhorn, A. Hawe, J.C. Kasper, A.
7. Conclusion Almeida, T. Quinten, W. Friess, G. Winter, C. Vervaet, J.P.
Remond, Talanta 83 (2011) 1623.
[10] D. am Ende, K.S. Bronk, J. Mustakis, G. OÕConnor, C.L. Santa
The DS requirement of the ICH Q8 states that the DS is a Maria, R. Nosal, T.J.N. Watson, J. Pharm. Innov. 2 (2007) 71.
region where process parameters ‘‘have been demon- [11] J. Harms, X. Wang, T. Kim, X. Yang, A.S. Rathore, Biotech. Progr.
strated to provide assurance of quality’’. An additional 24 (24) (2008) 655.
opportunity for the DS for analytical procedures is the [12] A. Baldinger, L. Clerdent, J. Rantanen, M. Yang, H. Grohganz,
possibility to move inside the DS without the need to Pharm. Dev. Technol. 17 (2012) 389.

166 http://www.elsevier.com/locate/trac
Trends in Analytical Chemistry, Vol. 42, 2013 Trends

[13] S. Garcia Munoz, S. Dolph, H.W. Ward II, Comput. Chem. Eng. 34 [44] C. Horváth, W. Melander, I. Molnár, J. Chromatogr. 125 (1976)
(2010) 1098. 129.
[14] J. Huang, G. Kaul, C. Cai, R. Chatlapalli, P. Hernandez Abad, K. [45] S. Heron, A. Tchapla, J. Chromatogr., A 556 (1991) 219.
Ghosh, A. Nagi, Int. J. Pharm. 382 (2009) 23. [46] L.R. Snyder, J.W. Dolan, Adv. Chromatogr. 38 (1998) 115.
[15] S. Adam, D. Suzzi, C. Radeke, J.G. Khinast, Eur. J. Pharm. Sci. 42 [47] P. Jandera, Adv. Chromatogr. 43 (2004) 1.
(2011) 106. [48] Y. Vander Heyden, A. Nijhuis, J. Smeyers-Verbeke, B.G.M.
[16] T. Lipsanen, O. Antikainen, H. Raikkonen, S. Airaksinen, J. Vandeginste, D.L. Massart, J. Pharm. Biomed. Anal. 24 (2001)
Yliruusi, Int. J. Pharm. 345 (2007) 101. 723.
[17] S. Verma, Y. Lan, R. Gokhale, D.J. Burgess, Int. J. Pharm. 377 [49] V. Czitrom, Am. Statist. 53 (1999) 126.
(2009) 185. [50] D. Brynn Hibbert, J. Chromatogr., B (2012) (DOI: 10.1016/
[18] J. Huang, C. Goolcharran, K. Ghosh, Eur. J. Pharm. Biopharm. 78 j.jchromb.2012.01.020).
(2011) 141. [51] T. Lundstedt, E. Seifert, L. Abramo, B. Thelin, A. Nystrom, J.
[19] F.G. Vogt, A.S. Kord, J. Pharm. Sci. 100 (2011) 797. Pettersen, R. Bergman, Chemom. Intell. Lab. Syst. 42 (1998) 3.
[20] M. Schweitzer, M. Pohl, M. Hanna-Brown, P. Nethercote, P. [52] B. Dejaegher, Y. Vander Heyden, J. Chromatogr., A 1158 (2007)
Borman, G. Hansen, K. Smith, J. Larew, Pharm. Tech. 34 (2010) 138.
52. [53] R.D. Marini, P. Hubert, E. Rozet, Y. Vander, Heyden, E. Ziemons,
[21] J. Ermer, Eur. Pharm. Rev. 16 (2011) 16. B. Boulanger, A. Bouklouze, J. Crommen, J. Pharm. Biomed. Anal.
[22] P. Nethercote, P. Borman, T. Bennett, G. Martin, P. McGregor, 44 (2007) 640.
Pharm. Manufact. April (2010) 37. [54] M.A. Bezerra, R.E. Santelli, E.P. Oliveira, L.S. Villar, L.A. Escaleira,
[23] P. Borman, J. Roberts, C. Jones, M. Hanna-Brown, R. Szucs, S. Talanta 76 (2008) 965.
Bale, Sep. Sci. 2 (2010) 1. [55] S.L.C. Ferreira, R.E. Bruns, E.G.P. da Silva, W.N.L. dos Santos,
[24] E. Rozet, E. Ziemons, R.D. Marini, B. Boulanger, Ph. Hubert, Anal. C.M. Quintella, J.M. David, J.B. de Andrade, M.C. Breitkreitz, I.C.S.
Chem. 84 (2012) 106. Fontes Jardim, B.B. Neto, J. Chromatogr., A 1158 (2007) 2.
[25] K. Ishikawa, What is Total Quality Control? The Japanese Way, [56] G.E.P. Box, D.W. Behnken, Technometrics 2 (1960) 455.
Prentice-Hall, Englewood Cliffs, NJ, USA, 1985, pp. 63–64. [57] D.H. Doehlert, Appl. Statist. 19 (1970) 231.
[26] L. Zhou, J. Socha, F.G. Vogt, S. Chen, A.S. Kord, Am. Pharm. Rev. [58] A.C. Atkinson, R.D. Tobias, J. Chromatogr., A 1177 (2008) 1.
13 (2010) 74. [59] B. Debrus, P. Lebrun, J. Mbinze Kindenge, F. Lecomte, A. Ceccato,
[27] P.J. Borman, P. Nethercote, M.J. Chatfield, D. Thompson, K. G. Caliaro, J. Mavar Tayey Mbay, B. Boulanger, R.D. Marini, E.
Truman, Pharm. Technol. 31 (2007) 142. Rozet, P. Hubert, J. Chromatogr., A 1218 (2011) 5205.
[28] D. Stamatis, Failure Modes and Effects Analysis, Second Edition, [60] J.J. Peterson (http://www.pharmamanufacturing.com/articles/
ASQ Quality Press, Milwaukee, WI, USA, FMEA from Theory to 2010/097.html).
Execution, 2003. [61] E. Harrington, Ind. Qual. Control 21 (1965) 494.
[29] J.F. van Leeuwen, M.J. Nauta, D. de Kaste, Y.M.C.F. Odekerken- [62] G. Derringer, R. Suich, J. Qual. Technol. 12 (1980) 214.
Rombouts, M.T. Oldenhof, M.J. Vredenbregt, D.M. Barends, J. [63] J.J. Peterson, K. Lief, Stat. Biopharm. Res. 2 (2010) 249.
Pharm. Biomed. Anal. 50 (2009) 1085. [64] J.J. Peterson, J. Qual. Technol. 36 (2004) 139.
[30] P.J. Borman, M.J. Chatfield, I. Damjanov, P. Jackson, Anal. Chim. [65] M.A. Herrador, A.G. Asuero, A.G. Gonzalez, Chemom. Intell. Lab.
Acta 703 (2011) 101. Syst. 79 (2005) 115.
[31] J. Mantanus, E. Ziemons, P. Lebrun, E. Rozet, R. Klinkenberg, B. [66] C. Davison, D.V. Hinkley, Bootstrap methods and their applica-
Streel, B. Evrard, P. Hubert, Talanta 80 (2010) 1750. tion, Cambridge University Press, Cambridge, UK, 1997.
[32] D.C. Montgomery, Introduction to Statistical Quality Control, [67] B. De Backer, B. Debrus, P. Lebrun, L. Theunis, N. Dubois, L.
Sixth Edition., Wiley, Hoboken, NJ, USA, 2009 p. 52. Decock, A. Verstraete, P. Hubert, C. Charlier, J. Chromatogr., B
[33] International Conference on Harmonization (ICH) of Technical 877 (2009) 4115.
Requirements for registration of Pharmaceuticals for Human Use, [68] I. Nistor, M. Cao, B. Debrus, P. Lebrun, F. Lecomte, E. Rozet, L.
Topic Q2 (R1): Validation of Analytical Procedures: Text and Angenot, M. Frederich, R. Oprean, P. Hubert, J. Pharm. Biomed.
Methodology, ICH, Geneva, Switzerland, 2005. Anal. 56 (2011) 30.
[34] A. Bouabidi, E. Rozet, M. Fillet, E. Ziemons, E. Chapuzet, B. [69] M.H. Rafamantanan, B. Debrus, G.E. Raoelison, S. Uverg-Rats-
Mertens, R. Klinkenberg, P. Hubert, J. Chromatogr., A 1217 imamang, E. Rozet, P. Hubert, J. Quetin-Leclercq, J. Pharm.
(2010) 3180. Biomed. Anal. 62 (2012) 23.
[35] E. Rozet, R.D. Marini, E. Ziemons, B. Boulanger, P. Hubert, J. [70] S. Fekete, J. Fekete, I. Molnár, K. Ganzler, J. Chromatogr., A 1216
Pharm. Biomed. Anal. 55 (2011) 848. (2009) 7816.
[36] Y. Vander Heyden, M. Jimidar, E. Hund, N. Niemeijer, R. Peeters, [71] I. Molnár, H.-J. Rieger, K.E. Monks, J. Chromatogr., A 1217
J. Smeyers-Verbeke, D.L. Massart, J. Hoogmartens, J. Chromatogr., (2010) 3193.
A 845 (1999) 145. [72] K.E. Monks, H.-J. Rieger, I. Molnár, J. Pharm. Biomed. Anal. 56
[37] J.J. Peterson, M. Yahyah, Stat. Biopharm Res. 1 (2009) 441. (2011) 874.
[38] A. Bouabidi, E. Ziemons, R. Marini, C. Hubert, M. Talbi, A. [73] I. Molnar, K.E. Monks, Chromatographia 73 (2011) S5.
Bouklouze, H. Bourichi, M. El Karbane, B. Boulanger, P. Hubert, E. [74] K. Monks, I. Molnar, H.-J. Rieger, B. Bogati, E. Szabo, J.
Rozet, Anal. Chim. Acta 714 (2012) 47. Chromatogr., A 1232 (2012) 218.
[39] W. Dewé, B. Govaerts, B. Boulanger, E. Rozet, P. Chiap, P. Hubert, [75] P. Lebrun, B. Govaerts, B. Debrus, A. Ceccato, G. Caliaro, P.
Chemom. Intell. Lab. Syst. 85 (2007) 262. Hubert, B. Boulanger, Chemom. Intell. Lab. Syst. 91 (2008) 4.
[40] P.J. Schoenmakers, H.A.H. Billet, R. Tijssen, L. DeGaan, J. [76] F. Krier, M. Brion, B. Debrus, P. Lebrun, A. Driesen, E. Ziemons, B.
Chromatogr. 149 (1978) 519. Evrard, P. Hubert, J. Pharm. Biomed. Anal. 54 (2011) 694.
[41] L.R. Snyder, J.W. Dolan, J.R. Grant, J. Chromatogr. 165 (1979) 3. [77] B. Debrus, P. Lebrun, A. Ceccato, G. Caliaro, E. Rozet, I. Nistor, R.
[42] A.P. Schellinger, P.W. Carr, J. Chromatogr., A 1077 (2005) 110. Oprean, F.J. Rupérez, C. Barbas, B. Boulanger, P. Hubert, Anal.
[43] R. Cela, E.Y. Ordoñez, J.B. Quintana, R. Rodil, J. Chromatogr., A Chim. Acta 691 (2011) 33.
(2012) (DOI: 10.1016/j.chroma.2012.07.081). [78] R. Rajagopal, E. del Castillo, J. Oper. Res. Soc. 58 (2007) 779.

http://www.elsevier.com/locate/trac 167

You might also like