G Model

CHROMA-460439; No. of Pages 10 ARTICLE IN PRESS

Journal of Chromatography A, xxx (xxxx) xxx

Contents lists available at ScienceDirect

Journal of Chromatography A
journal homepage: www.elsevier.com/locate/chroma

Robustness assessment in computer-assisted liquid chromatography

procedures based on desirability functions
S. Triñanes a , P. Rodriguez-Mier b , C. Cobas c , E. Sánchez c , R. Phan-Tan-Luu d,1 , R. Cela a,∗
a
Dpt. of Analytical Chemistry, Research Institute of Chemical and Biological Analysis, University of Santiago de Compostela, Spain
b
Centre for Information Technologies, University of Santiago de Compostela, Spain
c
Mestrelab Research, Spain
d
University Paul Cezanne, Marseille, France

a r t i c l e i n f o a b s t r a c t

Article history: A novel approach based on the use of desirability functions is presented for the robustness assessment
Received 28 June 2019 of liquid chromatographic separations as derived from computer-assisted methods development pro-
Received in revised form 4 August 2019 cesses. The approach is based on generally accepted hypothesis that a robust separation procedure will
Accepted 6 August 2019
be inert to small random variations of the operational variables, typically encountered in the day-to-day
Available online xxx
routine analytical practice. This means that peak positions along the chromatograms must keep stand-
still or move insignificantly when operational variables are not intentionally changed. Thus, the degree
Keywords:
of peak positions variation as evaluated from mathematical retention models can be used to assess the
Robustness
LC methods development robustness of the developed procedures before testing the actual performance experimentally. In the
Validation approach proposed, this assessment is obtained by fixing a bilateral partial desirability window around
Computer assisted separations each peak in the simulated chromatogram. The whole chromatogram robustness is characterized by an
Desirability functions overall desirability value calculated as the geometric mean of the partial desirability windows evalua-
Simulation tion. An added advantage of this approach is that the robustness value calculated is normalized between
zero and one and thus, easy to interpret. Thus, when chromatograms are simulated and small random
variations are introduced into the operational factors of the model, values for the overall desirability
close to one means that the procedure performs robustly. On the contrary, low values for the overall
desirability clearly indicated a serious lack of robustness. When used in conjunction with the Pareto
optimality approach, as shown here, this robustness assessment strategy allows testing several Pareto
front solutions before the final experimental testing which is always needed. In this way, a dramatical
reduction of the experimental effort is obtained. Although the approach is theoretically applicable to any
chromatographic separation, examples of reversed phase liquid chromatographic procedures are used
to show the performance of the proposed methodology.
© 2019 Elsevier B.V. All rights reserved.

1. Introduction
Computer-assisted methods development (CAMD) in liquid
chromatography (LC) [1,2] has been an important helping tool
for chromatographers since its establishment in the 1970s when
computers entered the analytical laboratory. Although the power
of LC was recognized from the very beginning, developing prac-
tical methods in LC appeared clearly more complicated than in
gas chromatography. More operational variables, frequently not
independent, means more capabilities for challenging samples but
∗ Corresponding author.
E-mail address: rafael.cela@usc.es (R. Cela).
1
Retired professor
also more difficulties in the development stages and during rou-
tine application. This required a significant degree of expertise
in chromatographers. Moreover, before the advent of sub-two-
micron particles and the development of the associated equipment,
LC procedures were quite slow and thus, not only expertise, but
dedication was also required. The use of computerized approaches
facilitates the work of chromatographers needing developing new
separation procedures and reduces dramatically the development
time from days or weeks to hours in routine and unique samples.
Generally, the main goal in LC method’s development is to find
the set of operational parameters that allows a complete separa-
tion of all the components in the sample or, alternatively, of those
components of interest from any other present in the samples. If a
separation procedure is developed to be used in routine analysis,
the robustness of that procedure is a critical performance charac-

https://doi.org/10.1016/j.chroma.2019.460439
0021-9673/© 2019 Elsevier B.V. All rights reserved.

Please cite this article in press as: S. Triñanes, et al., Robustness assessment in computer-assisted liquid chromatography procedures
based on desirability functions, J. Chromatogr. A (xxxx), https://doi.org/10.1016/j.chroma.2019.460439
G Model
CHROMA-460439; No. of Pages 10 ARTICLE IN PRESS
2 S. Triñanes et al. / J. Chromatogr. A xxx (xxxx) xxx
teristic because the procedure should perform equally day after day
until the substitution of the analytical column by a new one. Conse-
quently, CAMD software includes facilities to assess the robustness
of the developed procedures before the final experimental test-
ing. For example, Drylab, one of the oldest and probably the most
widespread commercial software package for CAMD assess the
robustness of separation procedures developed by estimating the
distribution of values of the critical resolution (the resolution of the
worst resolved pair of peaks in the chromatogram) when the opera-
tional variables are subjected to small random variations around set
point, by simulating the corresponding chromatograms. Because
the critical resolution is the criteria used by this software for meth-
ods development, a separation procedure should be robust when
the proportion of simulations under the critical resolution require-
ments, usually taken as 1.5, is low [3]. Moreover, the significance
of factors can be estimated through the regression coefficients
magnitude and sign by calculating a regression model of these oper-
ational factors in that simulation process. Although the procedure
is straightforward, some additional precautions should be taken in
practice. On one side, a conservative decision should be adopted
because, given certain variability, at the limit between satisfactory
or not critical resolution, there is only 50% of probability perfor-
mance for the procedure. This means that the acceptable working
space is a region well inside the optimum region in the resolution
map. Another more subtle detail is considering that critical res-
olution can be set by any pair of peaks in the chromatogram so,
if crosses of peaks are taking place in the optimal resolution map
region, that region should be divided carefully into subregions asso-
ciated to each pair of peaks limiting the resolution and thus, the
variation limits in the robustness assessment should be defined
accordingly [4].
For strategies based on multi-objective empirical models (e.g.
Fusion QbD software from S-Matrix [5], which is currently the prin-
cipal representative), the acceptable working region is defined by
the overlay of the response surfaces corresponding to the chosen
optimization criteria. Eventually, a region will meet the optimal
values of all criteria and thus, the robustness should be assessed
inside that workable region. Fusion QbD used conventional process
capability indexes by comparing the inherent ± 3sigma variability
of the process with the quality limits imposed. According to that
concept, the capability index is a measure of the process variabil-
ity and the capability of the process to achieve the expected results.
Thus, to estimate the robust working space, the optimization objec-
tives defining the acceptable performance region are combined
with the robustness capability indexes required after developing
Monte Carlo simulations to estimate the variability of responses
around the point selected as optimal [6]. Overlapping the responses
and robustness regions defines the final acceptable space where the
experimental verification design must be executed to validate the
simulation results.
In this paper, a novel approach to assess the robustness of solu-
tions adopted from CAMD processes is presented. In this approach,
the robustness assessment is applied to the selected optimal elution
procedure and not to a region of the response surface model. More-
over, the innerness of the whole chromatogram when operational
variables are subjected to small random variations is considered as
the measure of the robustness and not a particular measure (e.g.
the critical resolution), by defining bilateral desirability functions
[7] around each peak in the chromatogram. The overall desirability
is then calculated and used to estimate the robustness of the proce-
dure. To the best of our knowledge, this is the first time this concept
is applied in the field of chromatography and CAMD. The proposed
strategy was implemented as a routine that can be combined with
CAMD software to perform the whole process of method develop-
ment and robustness assessment. A pair of case studies are shown
to demonstrate the strategy and its practical application.
Please cite this article in press as: S. Triñanes, et al., Robustness assessment in computer-assisted liquid chromatography procedures
based on desirability functions, J. Chromatogr. A (xxxx), https://doi.org/10.1016/j.chroma.2019.460439
2. Experimental section
2.1. Reagents and samples
Solvents (acetonitrile and methanol) of gradient LC grade were
purchased from Merck (Darmstadt, Germany), formic acid (VWR,
Hypersolv, Llinars del Vallés, Spain) and ultrapure water (prepared
by means of a Millipore MilliQ-Gradient A10 facility), were used
to prepare mobile phases. Active pharmaceutical ingredients (API)
as well as their impurities and degradants are proprietary informa-
tion. Case study A was composed by an API and four impurities. Case
study B was formed by another API with eight impurities. Sample
solutions and mobile phases were filtered by means of 0.20 and 0.22
microns Millex LG and Millex GV membranes and MF-Millipore
0.45 microns MCE membranes respectively.
Samples of case study A were processed with a Kinetex phenyl-
hexyl (Phenomenex® ), 2.6 ␮m, 4.6 x 50 mm, core shell analytical
column. Samples of case study B were processed with a fully porous
Xbridge C18 (Waters® ), 5 ␮m, 4.6 x 75 mm column. Aqueous mobile
phases in both cases were acidified to pH 3.0 with formic acid, and
the organic modifier was constituted by mixtures of acetonitrile
(ACN) and methanol (MeOH), with pure solvents as the range limits.
In all cases, sample volume injection was 1.0 ␮L.
2.2. Apparatus and equipment
Experiments with the case study A were carried out in an Agilent
1290 II Infinity LC system composed by a G7104 quaternary flexible
pump, a G7167B multisampler, a G7116B multicolumn tempera-
ture controller and a G7117B diode array detector (DAD), all from
Agilent TechnologiesTM (Santa Clara, CA, USA). The instrument was
controlled by the Open Lab ChemStation CDS. Experiments with
case study B were carried out in a Waters (Milford, MA, USA),
Alliance 2695 separations module attached to a Waters 2996 DAD.
The system was controlled by the WatersTM , Empower-2 CDS.
2.3. Software
Retention model calculations were developed with the Mestre-
lab ResearchTM MChrom Scout v 1.52 computer assisted methods
development software [8], developed by one of the authors (R.
Cela). This software allows building retention maps in reversed-
phase liquid chromatography for sample mixtures of known
composition as well as for samples of unknown composition. Only
the modules associated to trackable samples were used in this
study. Separations were developed by resorting to standard Hoke
D6 [9] experimental designs, and optimized using multi-objective
Pareto optimality (MOPO), although conventional approaches (res-
olution maps and/or chromatographic response functions, CRFs)
can also be applied. These strategies can be developed in modes
that go from fully automated to fully interactive processes. A rou-
tine for the robustness assessment was developed and added inside
the Mchrom Scout software platform. This will be the only routine
described and discussed here in sufficient detail. Other features of
this software can be found in the Mestrelab web page [8] and can
be provided by the corresponding author upon request.
The basic idea in this robustness assessment procedure follows
a suggestion of Prof. Dr. Roger Phan Tan Luu, who implemented
a related procedure to assess the robust region in a response
surface as obtained by experimental designs in his Design of Experi-
ments (DOE), Nemrodw TM software [10]. This idea was elaborated
and adapted to the problem of chromatographic separations and
computer-assisted methods development and optimization by R.
Cela, using the following approach:
G Model
CHROMA-460439; No. of Pages 10 ARTICLE IN PRESS
S. Triñanes et al. / J. Chromatogr. A xxx (xxxx) xxx
Fig. 1. Schematic view of the individual bilateral desirability function associated to
each peak in the proposed robustness assessment strategy. See text for details about
the values taken for points b1 to b4 .
a) Robustness is a goal of the experimenter and thus, for multi-
objective problems can be expressed in terms of desirability
functions [7,11,12]. Chromatography separations are nearly
always multi-objective optimization problems. Moreover, in
chromatography, we depart from a particular elution program
(and thus, from a particular chromatogram) and not from a DOE
model, so the idea of assessing the robustness by means of desir-
ability functions need important adaptations.
b) A robust elution program is the one providing chromatograms
insensitive to small random variations in the day-to-day work
(e.g. small errors in column temperature control, pH adjustment
and solvent composition, etc.). The separation will be robust if
the retention times and shapes of peaks are not affected by these
small errors. Thus, we can define a bilateral desirability func-
tion for each peak in the chromatogram as the one schematized
in Fig. 1. This bilateral desirability function (a desirability win-
dow), encapsulates the optimized position of each peak. Inner
limits (b2 and b3 limits in Fig. 1), correspond to the confidence
bands estimated from the retention model built. Within these
limits, simulated retention times are considered under the reten-
tion model accuracy, so these limits cannot be reduced. External
limits (b1 and b4 ), can be enlarged interactively, and thus the
robustness assessment can be tailored in some way by the chro-
matographer goals and expertise to adapt his/her particular
problem. As an example, Fig. 2 shows the desirability windows
defined for a particular chromatogram represented as bands in
the auxiliary scale under the chromatogram plot. This graphic
representation helps in configuring the robustness assessment
study. Be aware that inner limits can be quite different for the dif-
ferent peaks in the retention model and also external limits can
be relaxed or not in the view of the critical separation between
peaks. Moreover, peak’s individual desirability functions can be
weighted to highlight the importance of the critical pairs when
needed.
c) By definition [7], desirability functions provide a normalized
response in the range 0-1. For bilateral functions, the functions
become zero when the peak goes outside the defined desirability
window and one when it is inside the optimum prediction band
foreseen by the retention model. When any peak departs from its
optimum location the value for its desirability function decreases
and is calculated according to a transition function. As shown in
Fig. 1, linear functions have been adopted in our implementa-
tion, although a variety of transition functions can be found in
the literature [12].
d) The whole chromatogram robustness can be characterized by
an overall desirability value, calculated as the geometric mean
of all the individual desirability functions. Consequently, also
the robustness value for the whole chromatogram takes values
between zero and one, providing a very intuitive and practical
Please cite this article in press as: S. Triñanes, et al., Robustness assessment in computer-assisted liquid chromatography procedures
based on desirability functions, J. Chromatogr. A (xxxx), https://doi.org/10.1016/j.chroma.2019.460439
3
measure of the degree of inertness of the peak’s positions when
the procedure is subjected to small random variations.
e) Following Monte Carlo simulations, operational variables in the
retention model can be modified by small random variations.
A chromatogram is simulated for each set of operational val-
ues, the individual desirability values and the overall desirability
calculated and registered. Then, robustness evaluation tools are
applied.
f) Two evaluation tools were implemented to assess the separation
process robustness. The first one is a sensitivity analysis [13,14],
and the second is a plot of the width and symmetry of the allow-
able region around the optimum as a function of the variation
of the critical operational variables as depicted by the sensitivity
analysis. These tools will be shown in the results and discussion
section as applied to the case studies.
3. Results and discussion
In order to show the performance of the proposed robustness
assessment process two case studies were selected. In both cases
the problem consists in the mixture of an API and its impurities
and degradants. Case study A is a quite simple separation problem
including only four impurities most of them easily resolvable. Case
study B considers a more challenging separation problem including
eight impurities, some o which are difficult to resolve both from the
API and from the small neighbouring peaks.
3.1. Case study A
As indicated in section 2.1, that sample was processed with a
phenyl-hexyl core-shell analytical column, using a pseudo-ternary
mobile phase composed by mixtures of ACN and MeOH as the
organic modifier and acidified ultrapure water. In the procedure
development the following four variables were considered: gra-
dient time (in the range 10–30 min), initial and final modifier
percentages (ranging from 15 to 95%) and the proportion of ACN
in the organic modifier binary mixture (from 0 to 100%). Conse-
quently, the procedure development using Mchrom Scout involved
a total of eight gradient runs, six to calibrate the model and two as
testing points to validate the retention model. The number of test
points can be increased, and these points will never be included
in the model calibration processes, thus providing a conventional
external validation strategy. The experimental plan and the regis-
tered retention times for peaks have been summarized in table S1.
Quality assessment plots for the retention model obtained have
been reproduced in figures S1 and S2, indicating a satisfactory
model.
Figure S1 compares the predicted retention times with the
experimental ones, showing the goodness of fit. Good retention
models are characterized by points close to the diagonal in the
plot. Data points are differently depicted for calibration runs (empty
symbols) and for validation runs (filled symbols), allowing appre-
ciate that both types of symbols are equally distributed along the
diagonal which is a signal of model predictive reliability. Figure
S2 provides further indication of prediction errors distribution dis-
persion and symmetry. Confidence bands (5% significance level) for
that prediction errors have been summarized in Table 1. This table
indicate the estimates for prediction errors (mean, standard devia-
tion, minimum and maximum values) for each peak in the samples
and the 95% confidence bands calculated in each case. These values
will be used later to build the desirability windows in the robust-
ness assessment. Although the use of test points in Mchrom Scout is
not mandatory, the robustness option is allowed only when having
included at least two test points, because otherwise the prediction
errors estimation cannot be considered reliable. Moreover, if pre-
G Model
CHROMA-460439; No. of Pages 10 ARTICLE IN PRESS
4 S. Triñanes et al. / J. Chromatogr. A xxx (xxxx) xxx

Fig. 2. Helping tool associating the individual desirability windows to chromatogram peaks. Block limits in the auxiliary time scale correspond to the external desirability
window limits for each peak (see text for details).

Table 1
Confidence bands for prediction errors in the retention model derived for Case study A.

Standard Mean
Peak Count Mean Minimum Maximum
deviation
Confidence Confidence Confidence
interval (±) interval Lower interval Upper
limit limit

API 8 0.0175 0.0299 −0.0022 0.0828 0.0207 −0.0032 0.0382

Imp A 8 0.0532 0.0818 −0.0467 0.1560 0.0567 −0.0034 0.1099
Imp B 8 0.0158 0.0510 −0.0590 0.1123 0.0353 −0.0196 0.0511
Imp C 8 0.0132 0.0312 −0.0369 0.0690 0.0216 −0.0084 0.0348
Imp D 8 0.0071 0.0170 −0.0142 0.0330 0.0118 −0.0047 0.0189

diction errors for test points are very high as compared to errors
for calibration runs, the retention model might lack reliability and
thus, the robustness assessment could not be reliable.
Once the retention model has been developed, the optimiza-
tion procedure can start in the search of a satisfactory separation
of the sample components. To develop this process, we used Pareto
optimality [15,16]. This approach has the inherent advantage of
considering the separation as a true multi-objective optimization
process, where the objectives of the chromatographer as well as the
constraints of the problem, are selected as independent dimensions
without worrying about equations, combinations of objectives,
weighting coefficients or penalty functions. This avoids the intro-
duction of bias characteristic of weighted sum functions developed
some decades ago to measure the quality of the experimental and
simulated chromatograms reducing the need of chromatographer
expertise at the same time. Additionally, and very importantly,
selection of candidate optimal solutions is made at the end of the
optimization process by choosing amongst the solutions located
in the optimal Pareto front. This is done instead of prejudging the
final solution on the basis of decisions taken at the beginning of the
optimization process, when no information about the final results is
available, as frequently occurs when using CRFs or resolution maps.
The inspection of the experimental chromatograms developed
to calibrate the retention model (see the retention times on table
S1), gave the information needed to define the objectives and con-
straints for Pareto optimization process:
Objective 1. Critical resolution. The goal is to maximize the crit-
ical resolution no matter the pair of peaks producing that critical
value.
Objective 2: API and impurity A should show a resolution better
than 1.5. In the priming chromatograms API and impurity A peaks
are frequently overlapped or even crossed, so this pair of peaks
appeared critical to develop a good separation
Objective 3: Impurities B and C should also exhibit a resolu-
tion better than 1.5. Also, peaks from impurities B and C appeared
frequently overlapped, thus defining another critical pair.

Please cite this article in press as: S. Triñanes, et al., Robustness assessment in computer-assisted liquid chromatography procedures
based on desirability functions, J. Chromatogr. A (xxxx), https://doi.org/10.1016/j.chroma.2019.460439
G Model
CHROMA-460439; No. of Pages 10 ARTICLE IN PRESS
S. Triñanes et al. / J. Chromatogr. A xxx (xxxx) xxx 5

Fig. 3. Optimal solutions selected from the Pareto fronts for robustness testing (top simulated and bottom experimental) in case study A. (a) Selected solution from the
Pareto front of isocratic elutions; (b) selected solution from the Pareto front of linear gradients.

Constraint 1: All sample components should elute before 10 min
to avoid long chromatograms inadequate for routine analysis.
Notice that this definition of objectives and constraints would
guide the search of the optimization algorithm (an auto-adaptive
differential evolution algorithm [17,18] which will not be detailed
here at all) to find chromatograms with the highest possible value
for the critical resolution and, where the selected pairs of peaks
appear well separated, but only those with a total runtime under
10 min will be considered feasible and thus retained into the Pareto
front. Notice that this selection of objectives, which in principle may
appear as redundant, are not combined in any way so are affecting
independently the evaluation of the simulated chromatograms but
follow exactly the goals of the chromatographer. Moreover, accord-
ing to Pareto rules, any chromatogram superior to the remaining
ones in any of these objectives will be retained and the imposed
constraint would compensate the tendency of the first objective
to favour slow gradients. Thus, finally, the optimal Pareto front
will contain all simulated solutions showing any net advantage as
regards to any of the objectives being feasible in terms of runtime.
Now the chromatographer will take the more satisfactory solutions
by exploring this Pareto front, applying eventually not only the
defined objectives but any other criteria, even subjective, helping
the development of the desired separation.
As an example, the plot of the Pareto front for linear gradient
elutions is shown in the figure S3. Each point represents a solution
in the Pareto front, and the corresponding chromatogram can be
viewed by selecting any of these points. Fig. 3 shows the simulated
chromatograms for two selected solutions (one from the Pareto
front of isocratic elutions (a), using a 48% of modifier formed by
10:90 ACN:MeOH, and the other (b), taken from the Pareto front
of linear gradients, corresponding to a gradient from 29.8 to 95% of
modifier (pure ACN in that case), in 35 min. Both solutions provided
satisfactory separations with all components in samples well sep-
arated. Experimental verifications of these optimal solutions, also
shown in Fig. 3, suggested retaining the gradient program and not
the isocratic one because the peaks for API and impurity D appeared
with poor symmetry characteristics. The comparison of these veri-
fication runs, as well as those used to calibrate the retention model,
clearly show that the presence of MeOH in the mobile phase was
responsible of the tailing for API and, especially, for impurity D
peaks. Thus, the chromatogram in Fig. 3(b), corresponding to the
linear gradient using pure ACN as modifier, produced clearly bet-
ter shapes for peaks. In consequence, the linear gradient procedure
was selected as optimal and submitted to robustness assessment.
As detailed in section 2.3, the first step in these processes consist
in the configuration of the individual desirability window for each
peak in the chromatogram. The program takes the values from the
confidence bands in prediction errors (Table 1), to provide a first
guess of these windows. Then, the user may enlarge some of them
if needed. Fig. 4 shows the configuration of individual desirability
windows for the case study, clearly showing that no overlapping
exists between desirability windows. The data corresponding to
that window configuration has been reproduced in Table S2. Now,
the system can start the Monte Carlo simulation processes in order
to assess the potential robustness of the chosen optimal solution.
The results of that processes are shown in Figs. 5 and 6. Fig. 5
shows the sensitivity analysis where the influence of each oper-
ational variable in the overall desirability is plotted in terms of a
normalized interval around the optimum value (the position of the
peaks in the elution program for the optimum). As can be expected,
the desirability grows when intervals of variation are reduced. Fre-
quently, the influence intervals are not symmetrical, so this plot
gives a clearer picture of the limitations of the robustness for the
separation when operational variables change. Although the sen-
sitivity analysis plot may provide an estimation of the variation
intervals allowed for each variable, the essential value of this plot is
in providing a clear indication about the importance of variables in
the robustness behaviour of the elution program. In that case study,
we see that only the initial modifier percentage appears influenc-
ing the results of the gradient program significantly. Moreover, as
can be expected, the system suffers more when the initial modifier
percentages increases.
Once the critical operational factors have been identified, we can
estimate the robust workable space around the optimum by plot-

Please cite this article in press as: S. Triñanes, et al., Robustness assessment in computer-assisted liquid chromatography procedures
based on desirability functions, J. Chromatogr. A (xxxx), https://doi.org/10.1016/j.chroma.2019.460439
G Model
CHROMA-460439; No. of Pages 10 ARTICLE IN PRESS
6 S. Triñanes et al. / J. Chromatogr. A xxx (xxxx) xxx

Fig. 4. Robustness desirability windows associated with the linear gradient solution corresponding to case study A.

Fig. 5. Sensitivity analysis for the case study A.

Please cite this article in press as: S. Triñanes, et al., Robustness assessment in computer-assisted liquid chromatography procedures
based on desirability functions, J. Chromatogr. A (xxxx), https://doi.org/10.1016/j.chroma.2019.460439
G Model
CHROMA-460439; No. of Pages 10 ARTICLE IN PRESS
S. Triñanes et al. / J. Chromatogr. A xxx (xxxx) xxx 7

Fig. 6. Estimation of the probability of good overall desirability values for the case study A as a function of the critical variable (initial percent of modifier) and gradient
time. Coloured cells correspond to regions above the desirability limits established. White cells are, otherwise for desirability values under the limit. See text for operational
variable ranges.

Table 2 reduced to a minimum in order to highlight the differences and

Data for injections of Case study A sample in three different days.
Name Ret time (min) USP resolution
API 0.838 ± 0.002 5.96 ± 0.07
Imp A 1.471 ± 0.001 5.25 ± 0.06
Imp B 1.947 ± 0.001 3.06 ± 0.03
Imp C 2.244 ± 0.001 6.09 ± 0.07
Imp D 2.913 ± 0.003 – 0.83 ± 0.02
ting the region above a certain desirability limit as a function of the
values taken by these critical factors. Fig. 6 plots that region for a ≥
0.85 desirability limit, which is a reasonable level in practice. Here
we can see that the procedure may resist changes of ± 1% in the
initial modifier and of ±2 min in the gradient time without appre-
ciably degrading its performance. These values are larger than those
that can be expected on well performing laboratories equipped
with modern instruments. Of course, robustness assessment using
CAMD procedures do not substitute the final experimental verifi-
cation if positive conclusions are reached. Table 2 shows the data
obtained in the injections of the sample developed in three different
days preparing fresh mobile phases each time. These data confirm
that the optimal elution program chosen for Case study A can be
applied conveniently in routine analysis.
3.2. Case study B
Case study B is similar to the one described before although more
challenging. In that case, the API samples can contain eight impuri-
ties and/or degradants. Since the process to develop a satisfactory
elution program and to assess the robustness are the same dis-
cussed for the Case study A, the descriptions in that case will be
new facts associated to this Case study.
USP tailing As indicated in section 2.1, these samples were processed with a
1.63 ± 0.02 conventional Xbridge C18 analytical column, using again a pseudo-
1.10 ± 0.08 ternary mobile phase formed by mixtures of ACN and MeOH as
0.99 ± 0.01 the organic modifier and acidified ultrapure water. In this case,
1.01 ± 0.05 the same operational variables taken into account for Case study
A were considered (initial and final modifier percent in the range
5–40%, gradient time in the range 15–45 min and proportion of ACN
in the modifier in the range 0–100%). In addition, the temperature
in the range 35–55 ◦ C was included in the retention model. Con-
sequently, a total of five operational variables were included in
the model so 14 runs (12 to calibrate and two or more for vali-
dation purposes) are needed to build the model. Model runs and
the registered retention times for peaks have been summarized in
table S3. Quality assessment plots for the retention model obtained
have been reproduced in figures S4 and S5, indicating a satisfac-
tory model. Confidence bands for prediction errors in the retention
model have been summarized in Table 3.
Again, Pareto optimality was applied to search good solutions
for that separation. Resorting to Pareto methods is fully justified
when the separation problems are more and more challenging
because the built-in advantages of this method (easiness of objec-
tives definition and enhanced objectivity in the selection of optimal
solutions), are more evident when the difficulty of the separation
problems makes it hard to take decisions a priori. In that case, only
two objectives were defined: (1) critical resolution and (2) total elu-
tion time, and a constraint regarding the total elution time which
will be limited to 15 min to avoid excessively slow gradients. Notice
that in Pareto optimality the same parameter (e.g. total runtime in
that case), can be used both as an objective and as a constraint
in a fully independent way. As an example of the results obtained,

Please cite this article in press as: S. Triñanes, et al., Robustness assessment in computer-assisted liquid chromatography procedures
based on desirability functions, J. Chromatogr. A (xxxx), https://doi.org/10.1016/j.chroma.2019.460439
G Model
CHROMA-460439; No. of Pages 10 ARTICLE IN PRESS
8 S. Triñanes et al. / J. Chromatogr. A xxx (xxxx) xxx

Table 3
Confidence bands for prediction errors in the retention model derived for Case study B.

Standard Mean
Peak Count Mean Minimum Maximum
deviation
Confidence Confidence Confidence
interval (±) interval Lower interval Upper
limit limit

API 14 0.0692 0.1353 −0.1457 0.3198 0.0708 −0.0017 0.1400

Imp A 14 0.0816 0.2261 −0.2609 0.4287 0.1184 −0.0368 0.2000
Imp B 14 0.0939 0.2629 −0.2241 0.5194 0.1377 −0.0438 0.2316
Imp C 14 0.0924 0.2371 −0.1835 0.4740 0.1242 −0.0318 0.2167
Imp D 14 0.0733 0.2358 −0.2096 0.4470 0.1235 −0.0502 0.1968
Imp E 14 0.0450 0.1227 −0.1179 0.3108 0.0643 −0.0193 0.1093
Imp F 14 0.0139 0.1773 −0.3603 0.2580 0.0929 −0.0790 0.1068
Imp G 14 0.0102 0.0586 −0.0643 0.1576 0.0307 −0.0205 0.0409
Imp H 14 0.0755 0.2259 −0.1892 0.4603 0.1183 −0.0429 0.1938

Fig. 7. Experimental verification of gradient elutions selected for robustness assessment in case study B. See text for elution programs in gradients (a) and (b).

Figure S6 shows the optimal Pareto fronts for linear and curved gra-
dients. Notice that the critical resolution scale (y axis in the plots in
figure S6) indicated clear difficulties in obtaining a good separation
of all components in the samples. Among Pareto front solutions
for the different elution modes (MChrom Scout produced inde-
pendent Pareto fronts for isocratic, linear, curved, multilinear and
stepwise gradients by default), two chromatograms were selected
to be tested about robustness. These chromatograms are shown
in Fig. 7. Chromatogram in Fig. 7(a) corresponds to a linear gra-
dient from 6 to 40% of modifier (a binary mixture of ACN:MeOH
42:58) executed in 10 min, maintaining the temperature at 35 ◦ C.
Chromatogram in Fig. 7(b) corresponds to a curved gradient (using
a curve 7 in the Waters programmer), from 5 to 40% of modifier
(ACN:MeOH 69:31) executed in 10 min at 37 ◦ C.
To assess the robustness of both elution procedures, the cor-
responding chromatograms are chosen, and desirability windows
configured departing from the confidence bands for prediction
errors in the retention model (Table 3). Inner window limits are the
same for any chromatogram simulated with this retention model
no matter the gradient shape, but external limits can be defined
by the chromatographer after viewing the chromatogram appear-
ance. In Fig. 8(a) the configuration of peak’s desirability windows
as developed by the routine is shown for the linear and curved gra-
dient in Fig. 7. It can be appreciated that windows for peaks 3,4 and
5 (Impurities C, D and E respectively) are overlapping, specially in
the curved gradient, so the resolutions are critical if retention of
these peaks change. Fig. 9 shows the sensitivity analysis for both
elution programs. Although the tendency of the curves appears sim-
ilar with the gradient time and the initial percentage of modifier
as the critical factors, it is evident that the linear gradient results
less affected by the changes in these factors than the curved gradi-
ent. The remaining factors, of which only temperature needs to be
considered, affect identically both elution programs. A similar con-
clusion can be derived from Fig. 10 that compares the robust regions
for an overall desirability above 0.85. In both elution programs we
see that this region is not circular but elliptical. This means that
the procedures behave more satisfactorily from the point of view
of robustness when both critical factors move in the same direction
but not when both factors changed in opposite ways. In practice,
this means that the true robust region is a circle representing the
worse situation with changes in gradient time of about ±0.3 min
and initial modifier percent of about ±0.6% for the linear gradient
but only ±0.3 min and ± 0.3% in the case of the curved gradient.
Thus, the linear gradient procedure appears slightly more reliable
for routine analysis.

Please cite this article in press as: S. Triñanes, et al., Robustness assessment in computer-assisted liquid chromatography procedures
based on desirability functions, J. Chromatogr. A (xxxx), https://doi.org/10.1016/j.chroma.2019.460439
G Model
CHROMA-460439; No. of Pages 10 ARTICLE IN PRESS
S. Triñanes et al. / J. Chromatogr. A xxx (xxxx) xxx 9

Fig. 8. Configuration of desirability windows corresponding to the selected optimal elution programs in Fig. 7.

Fig. 9. Sensitivity analysis for the selected optimal elution programs of Fig. 8 using the desirability windows approach described. (a) Linear gradient of Fig. 8(a); (b) curved
gradient of Fig. 8(b).

Fig. 10. Acceptable working region (overall desirability ≥ 0.7), corresponding to selected elution programmes in case study B as a function of the critical variables.

Please cite this article in press as: S. Triñanes, et al., Robustness assessment in computer-assisted liquid chromatography procedures
based on desirability functions, J. Chromatogr. A (xxxx), https://doi.org/10.1016/j.chroma.2019.460439
G Model
CHROMA-460439; No. of Pages 10 ARTICLE IN PRESS
10 S. Triñanes et al. / J. Chromatogr. A xxx (xxxx) xxx

Table 4
Data for injections of Case study B sample in three different days using the linear and curved gradients considered in the robustness testing.

Test 1 (linear gradient) Test 2 (curved gradient)

Name
Ret time (min) USP resolution USP tailing Ret time (min) USP resolution USP tailing

API 4.44 ± 0.23 2.09 ± 0.41 1.21 ± 0.01 5.83 ± 0.02 5.29 ± 0.00 1.19 ± 0.02
Imp A 4.15 ± 0.28 2.81 ± 0.38 1.11 ± 0.05 4.92 ± 0.03 2.75 ± 0.42 1.11 ± 0.01
Imp B 5.59 ± 0.40 1.38 ± 0.65 1.80 ± 1.09 7.03 ± 0.08 1.49 ± 0.45 1.37 ± 0.33
Imp C 5.36 ± 0.30 6.42 ± 0.09 1.01 ± 0.08 6.71 ± 0.03 5.07 ± 0.12 1.06 ± 0.01
Imp D 5.93 ± 0.28 1.88 ± 0.51 1.04 ± 0.02 7.39 ± 0.03 1.66 ± 0.15 1.05 ± 0.00
Imp E 6.66 ± 0.32 4.46 ± 0.18 1.13 ± 0.01 8.14 ± 0.03 4.46 ± 0.11 1.13 ± 0.01
Imp F 7.29 ± 0.19 4.38 ± 0.94 1.14 ± 0.01 8.78 ± 0.02 4.93 ± 0.08 1.12 ± 0.02
Imp G 7.77 ± 0.22 3.28 ± 0.16 1.10 ± 0.02 9.20 ± 0.02 3.51 ± 0.03 1.07 ± 0.00
Imp H 3.79 ± 0.27 1.11 ± 0.04 4.42 ± 0.06 1.04 ± 0.11

In any case, to confirm these observations, both gradient pro-
grammes were executed experimentally during several days and
the results have been summarized in Table 4. Results appear quite
similar in both cases with the peak of impurity B as the critical
component in the sample both in terms of resolution and peak tail-
ing, thus arousing doubts about the application in routine analysis
of any of those selected procedures, as can be expected from the
results of the CAMD robustness assessment.
4. Conclusions
the Xunta de Galicia (Conecta-Peme program, reference IN852A-
2013/75-0). S.T and P.R-M, acknowledge their contracts to the
Spanish Ministry of Economy and Competitivity in the frame of the
above-mentioned program.
Appendix A. Supplementary data
Supplementary material related to this article can be found, in
the online version, at doi:https://doi.org/10.1016/j.chroma.2019.
460439.

A novel procedure for assessment of robustness of solutions pro- References

duced in CAMD reversed phase liquid chromatography procedures,
based on the use of bilateral desirability functions, demonstrated its
applicability to limit the experimental work in the development of
separation procedures aimed at routine analysis. The implemen-
tation of the approach allows an easy application of the concept
as a part of the development process. The total experimental
effort, apart from the runs for calibrating and testing the reten-
tion model, is limited to the verification of the optimal solutions
selected, and the robustness predicted for the chosen elution pro-
gram. Only elution programs that appear satisfactory and passing
the CAMD robustness assessment need to be tested experimentally
which reduced the development process to hours instead of days
or weeks. The application of this desirability approach allows the
assessment of the robustness for the whole sample components
and not for a single characteristic of the chromatogram (e.g. the
critical resolution). Additionally, the proposed strategy allows the
chromatographer influencing the robustness goals by relaxing or
not the default configuration based on the prediction errors confi-
dence bands and thus introducing a flexibility degree in robustness
assessment. Pareto optimization provides additional advantages in
the process of selecting the candidate solutions to be tested and
thus improved reliability of the overall CAMD process.
Declaration of Competing Interest
The authors declare that they have no known competing finan-
cial interests or personal relationships that could have appeared to
influence the work reported in this paper.
Acknowledgements
This study has been supported by the Spanish Ministry of
Economy and Competitivity (Retos-Colaboración program, refer-
ence RTC-2015-3812-2) and European FEDER funds, as well as by
[1] R. Cela, E.Y. Ordoñez, J.B. Quintana, R. Rodil, Chemometric-assisted aethods
development in reversed-phase liquid chromatography, J. Chromatogr. A
1287 (2013) 2–22.
[2] R. Cela, S. Triñanes, Computerized optimization of liquid chromatography
separations, in: Encyclopaedia of Analytical Chemistry, John Wiley & Sons,
2017, http://dx.doi.org/10.1002/9780470027318.a9594), On Line.
[3] Drylab 4 User’s Manual, Molnár-Institute, Berlin, Germany, 2012, pp. 141–145.
[4] I. Molnár, H.-J. Rieger, K.E. Monks, Aspects of the “Design Space” in high
pressure liquid chromatography method development, J. Chromatogr. A 1217
(2010) 3193–3200.
[5] http://www.smatrix.com/.
[6] LC Methods Development User’s Guide, Chapter 8. Robustness Simulator,
Fusion QbD v. 9.7, S-Matrix, 2015, CA, USA.
[7] G. Derringer, R. Suich, Simultaneous optimization of several response
variables, J. Qual. Technol. 12 (1980) 214–219.
[8] http://mestrelab.com/software/mchrom-scout/.
[9] A.T. Hoke, Economical second-order designs based on irregular fractions of
the 3n factorial, Technometrics 16 (3) (1974) 375–384.
[10] Nemrod W. Génération de matrices d’experiences en fonction des objectifs et
traitement des reponses experimentales. D. Mathieu, J. Nony, R.
Phan-Tan-Luu, A. Beal, Nemrodw, Marseille, France. http://www.nemrodW.fr.
[11] E. del Castillo, D.C. Montgometry, D.R. McCarville, Modified desirability
functions for multiple response optimization, J. Qual. Technol. 28 (1996)
337–345.
[12] N.R. Costa, J. Lourenço, Z.L. Pereira, Desirability function approach: a review
and performance evaluation in adverse conditions, Chemometr. Intell. Lab.
Syst. 107 (2011) 2234–2244.
[13] A. Saltelli, M. Ratto, T. Andres, F. Campolongo, J. Cariboni, D. Gatelli, M. Saisana,
S. Tarantola, Global Sensitivity Analysis: The Primer, John Wiley & Sons, 2008.
[14] F. Pianosi, K. Beven, J. Freer, J.W. Hall, J. Rougier, D.B. Stephenson, T. Wagener,
Sensitivity analysis of environmental models: a systematics review with
practical workflow, Environ. Model. Softw. 70 (2016) 214–232.
[15] K. Deb, Multi-Objective Optimization Using Evolutionary Algorithms, John
Wiley & Sons, Ltd, Chichester, 2001.
[16] R. Cela, J.A. Martinez, C. Gonzalez-Barreiro, M. Lores, Multi-objective
optimization using evolutionary algorithms: its application to HPLC
separations, Chemometr. Intell. Lab. Syst. 69 (2003) 137–156.
[17] R. Storn, K. Price, Differential Evolution – A Simple and Efficient Adaptive
Scheme for Global Optimization over Continuous Spaces,
ftp.ICSI.Berkeley.edu/pub/techreports/l995/tr–9 5 012.ps.z.
[18] T. Robic, B. Filipic, et al., DEMO: differential evolution for multiobjective
optimization, in: C.A. Coello Coello (Ed.), EMO 2005, LNCS 3410,
Springer-Verlag, Berlin Heidelberg, 2005, pp. 520–533.

Please cite this article in press as: S. Triñanes, et al., Robustness assessment in computer-assisted liquid chromatography procedures
based on desirability functions, J. Chromatogr. A (xxxx), https://doi.org/10.1016/j.chroma.2019.460439