Professional Documents
Culture Documents
Journal of Chromatography A
journal homepage: www.elsevier.com/locate/chroma
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 also more difficulties in the development stages and during rou-
tine application. This required a significant degree of expertise
Computer-assisted methods development (CAMD) in liquid in chromatographers. Moreover, before the advent of sub-two-
chromatography (LC) [1,2] has been an important helping tool micron particles and the development of the associated equipment,
for chromatographers since its establishment in the 1970s when LC procedures were quite slow and thus, not only expertise, but
computers entered the analytical laboratory. Although the power dedication was also required. The use of computerized approaches
of LC was recognized from the very beginning, developing prac- facilitates the work of chromatographers needing developing new
tical methods in LC appeared clearly more complicated than in separation procedures and reduces dramatically the development
gas chromatography. More operational variables, frequently not time from days or weeks to hours in routine and unique samples.
independent, means more capabilities for challenging samples but 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
∗ Corresponding author.
components of interest from any other present in the samples. If a
separation procedure is developed to be used in routine analysis,
E-mail address: rafael.cela@usc.es (R. Cela).
1
Retired professor 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 2. Experimental section
until the substitution of the analytical column by a new one. Conse-
quently, CAMD software includes facilities to assess the robustness 2.1. Reagents and samples
of the developed procedures before the final experimental test-
ing. For example, Drylab, one of the oldest and probably the most Solvents (acetonitrile and methanol) of gradient LC grade were
widespread commercial software package for CAMD assess the purchased from Merck (Darmstadt, Germany), formic acid (VWR,
robustness of separation procedures developed by estimating the Hypersolv, Llinars del Vallés, Spain) and ultrapure water (prepared
distribution of values of the critical resolution (the resolution of the by means of a Millipore MilliQ-Gradient A10 facility), were used
worst resolved pair of peaks in the chromatogram) when the opera- to prepare mobile phases. Active pharmaceutical ingredients (API)
tional variables are subjected to small random variations around set as well as their impurities and degradants are proprietary informa-
point, by simulating the corresponding chromatograms. Because tion. Case study A was composed by an API and four impurities. Case
the critical resolution is the criteria used by this software for meth- study B was formed by another API with eight impurities. Sample
ods development, a separation procedure should be robust when solutions and mobile phases were filtered by means of 0.20 and 0.22
the proportion of simulations under the critical resolution require- microns Millex LG and Millex GV membranes and MF-Millipore
ments, usually taken as 1.5, is low [3]. Moreover, the significance 0.45 microns MCE membranes respectively.
of factors can be estimated through the regression coefficients Samples of case study A were processed with a Kinetex phenyl-
magnitude and sign by calculating a regression model of these oper- hexyl (Phenomenex® ), 2.6 m, 4.6 x 50 mm, core shell analytical
ational factors in that simulation process. Although the procedure column. Samples of case study B were processed with a fully porous
is straightforward, some additional precautions should be taken in Xbridge C18 (Waters® ), 5 m, 4.6 x 75 mm column. Aqueous mobile
practice. On one side, a conservative decision should be adopted phases in both cases were acidified to pH 3.0 with formic acid, and
because, given certain variability, at the limit between satisfactory the organic modifier was constituted by mixtures of acetonitrile
or not critical resolution, there is only 50% of probability perfor- (ACN) and methanol (MeOH), with pure solvents as the range limits.
mance for the procedure. This means that the acceptable working In all cases, sample volume injection was 1.0 L.
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, 2.2. Apparatus and equipment
if crosses of peaks are taking place in the optimal resolution map
region, that region should be divided carefully into subregions asso- Experiments with the case study A were carried out in an Agilent
ciated to each pair of peaks limiting the resolution and thus, the 1290 II Infinity LC system composed by a G7104 quaternary flexible
variation limits in the robustness assessment should be defined pump, a G7167B multisampler, a G7116B multicolumn tempera-
accordingly [4]. ture controller and a G7117B diode array detector (DAD), all from
For strategies based on multi-objective empirical models (e.g. Agilent TechnologiesTM (Santa Clara, CA, USA). The instrument was
Fusion QbD software from S-Matrix [5], which is currently the prin- controlled by the Open Lab ChemStation CDS. Experiments with
cipal representative), the acceptable working region is defined by case study B were carried out in a Waters (Milford, MA, USA),
the overlay of the response surfaces corresponding to the chosen Alliance 2695 separations module attached to a Waters 2996 DAD.
optimization criteria. Eventually, a region will meet the optimal The system was controlled by the WatersTM , Empower-2 CDS.
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 2.3. Software
of the process with the quality limits imposed. According to that
concept, the capability index is a measure of the process variabil- Retention model calculations were developed with the Mestre-
ity and the capability of the process to achieve the expected results. lab ResearchTM MChrom Scout v 1.52 computer assisted methods
Thus, to estimate the robust working space, the optimization objec- development software [8], developed by one of the authors (R.
tives defining the acceptable performance region are combined Cela). This software allows building retention maps in reversed-
with the robustness capability indexes required after developing phase liquid chromatography for sample mixtures of known
Monte Carlo simulations to estimate the variability of responses composition as well as for samples of unknown composition. Only
around the point selected as optimal [6]. Overlapping the responses the modules associated to trackable samples were used in this
and robustness regions defines the final acceptable space where the study. Separations were developed by resorting to standard Hoke
experimental verification design must be executed to validate the D6 [9] experimental designs, and optimized using multi-objective
simulation results. Pareto optimality (MOPO), although conventional approaches (res-
In this paper, a novel approach to assess the robustness of solu- olution maps and/or chromatographic response functions, CRFs)
tions adopted from CAMD processes is presented. In this approach, can also be applied. These strategies can be developed in modes
the robustness assessment is applied to the selected optimal elution that go from fully automated to fully interactive processes. A rou-
procedure and not to a region of the response surface model. More- tine for the robustness assessment was developed and added inside
over, the innerness of the whole chromatogram when operational the Mchrom Scout software platform. This will be the only routine
variables are subjected to small random variations is considered as described and discussed here in sufficient detail. Other features of
the measure of the robustness and not a particular measure (e.g. this software can be found in the Mestrelab web page [8] and can
the critical resolution), by defining bilateral desirability functions be provided by the corresponding author upon request.
[7] around each peak in the chromatogram. The overall desirability The basic idea in this robustness assessment procedure follows
is then calculated and used to estimate the robustness of the proce- a suggestion of Prof. Dr. Roger Phan Tan Luu, who implemented
dure. To the best of our knowledge, this is the first time this concept a related procedure to assess the robust region in a response
is applied in the field of chromatography and CAMD. The proposed surface as obtained by experimental designs in his Design of Experi-
strategy was implemented as a routine that can be combined with ments (DOE), Nemrodw TM software [10]. This idea was elaborated
CAMD software to perform the whole process of method develop- and adapted to the problem of chromatographic separations and
ment and robustness assessment. A pair of case studies are shown computer-assisted methods development and optimization by R.
to demonstrate the strategy and its practical application. Cela, using the following approach:
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 3
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
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
diction errors for test points are very high as compared to errors in the optimal Pareto front. This is done instead of prejudging the
for calibration runs, the retention model might lack reliability and final solution on the basis of decisions taken at the beginning of the
thus, the robustness assessment could not be reliable. optimization process, when no information about the final results is
Once the retention model has been developed, the optimiza- available, as frequently occurs when using CRFs or resolution maps.
tion procedure can start in the search of a satisfactory separation The inspection of the experimental chromatograms developed
of the sample components. To develop this process, we used Pareto to calibrate the retention model (see the retention times on table
optimality [15,16]. This approach has the inherent advantage of S1), gave the information needed to define the objectives and con-
considering the separation as a true multi-objective optimization straints for Pareto optimization process:
process, where the objectives of the chromatographer as well as the Objective 1. Critical resolution. The goal is to maximize the crit-
constraints of the problem, are selected as independent dimensions ical resolution no matter the pair of peaks producing that critical
without worrying about equations, combinations of objectives, value.
weighting coefficients or penalty functions. This avoids the intro- Objective 2: API and impurity A should show a resolution better
duction of bias characteristic of weighted sum functions developed than 1.5. In the priming chromatograms API and impurity A peaks
some decades ago to measure the quality of the experimental and are frequently overlapped or even crossed, so this pair of peaks
simulated chromatograms reducing the need of chromatographer appeared critical to develop a good separation
expertise at the same time. Additionally, and very importantly, Objective 3: Impurities B and C should also exhibit a resolu-
selection of candidate optimal solutions is made at the end of the tion better than 1.5. Also, peaks from impurities B and C appeared
optimization process by choosing amongst the solutions located 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 fication runs, as well as those used to calibrate the retention model,
to avoid long chromatograms inadequate for routine analysis. clearly show that the presence of MeOH in the mobile phase was
Notice that this definition of objectives and constraints would responsible of the tailing for API and, especially, for impurity D
guide the search of the optimization algorithm (an auto-adaptive peaks. Thus, the chromatogram in Fig. 3(b), corresponding to the
differential evolution algorithm [17,18] which will not be detailed linear gradient using pure ACN as modifier, produced clearly bet-
here at all) to find chromatograms with the highest possible value ter shapes for peaks. In consequence, the linear gradient procedure
for the critical resolution and, where the selected pairs of peaks was selected as optimal and submitted to robustness assessment.
appear well separated, but only those with a total runtime under As detailed in section 2.3, the first step in these processes consist
10 min will be considered feasible and thus retained into the Pareto in the configuration of the individual desirability window for each
front. Notice that this selection of objectives, which in principle may peak in the chromatogram. The program takes the values from the
appear as redundant, are not combined in any way so are affecting confidence bands in prediction errors (Table 1), to provide a first
independently the evaluation of the simulated chromatograms but guess of these windows. Then, the user may enlarge some of them
follow exactly the goals of the chromatographer. Moreover, accord- if needed. Fig. 4 shows the configuration of individual desirability
ing to Pareto rules, any chromatogram superior to the remaining windows for the case study, clearly showing that no overlapping
ones in any of these objectives will be retained and the imposed exists between desirability windows. The data corresponding to
constraint would compensate the tendency of the first objective that window configuration has been reproduced in Table S2. Now,
to favour slow gradients. Thus, finally, the optimal Pareto front the system can start the Monte Carlo simulation processes in order
will contain all simulated solutions showing any net advantage as to assess the potential robustness of the chosen optimal solution.
regards to any of the objectives being feasible in terms of runtime. The results of that processes are shown in Figs. 5 and 6. Fig. 5
Now the chromatographer will take the more satisfactory solutions shows the sensitivity analysis where the influence of each oper-
by exploring this Pareto front, applying eventually not only the ational variable in the overall desirability is plotted in terms of a
defined objectives but any other criteria, even subjective, helping normalized interval around the optimum value (the position of the
the development of the desired separation. peaks in the elution program for the optimum). As can be expected,
As an example, the plot of the Pareto front for linear gradient the desirability grows when intervals of variation are reduced. Fre-
elutions is shown in the figure S3. Each point represents a solution quently, the influence intervals are not symmetrical, so this plot
in the Pareto front, and the corresponding chromatogram can be gives a clearer picture of the limitations of the robustness for the
viewed by selecting any of these points. Fig. 3 shows the simulated separation when operational variables change. Although the sen-
chromatograms for two selected solutions (one from the Pareto sitivity analysis plot may provide an estimation of the variation
front of isocratic elutions (a), using a 48% of modifier formed by intervals allowed for each variable, the essential value of this plot is
10:90 ACN:MeOH, and the other (b), taken from the Pareto front in providing a clear indication about the importance of variables in
of linear gradients, corresponding to a gradient from 29.8 to 95% of the robustness behaviour of the elution program. In that case study,
modifier (pure ACN in that case), in 35 min. Both solutions provided we see that only the initial modifier percentage appears influenc-
satisfactory separations with all components in samples well sep- ing the results of the gradient program significantly. Moreover, as
arated. Experimental verifications of these optimal solutions, also can be expected, the system suffers more when the initial modifier
shown in Fig. 3, suggested retaining the gradient program and not percentages increases.
the isocratic one because the peaks for API and impurity D appeared Once the critical operational factors have been identified, we can
with poor symmetry characteristics. The comparison of these veri- 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.
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.
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
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- dient in Fig. 7. It can be appreciated that windows for peaks 3,4 and
dients. Notice that the critical resolution scale (y axis in the plots in 5 (Impurities C, D and E respectively) are overlapping, specially in
figure S6) indicated clear difficulties in obtaining a good separation the curved gradient, so the resolutions are critical if retention of
of all components in the samples. Among Pareto front solutions these peaks change. Fig. 9 shows the sensitivity analysis for both
for the different elution modes (MChrom Scout produced inde- elution programs. Although the tendency of the curves appears sim-
pendent Pareto fronts for isocratic, linear, curved, multilinear and ilar with the gradient time and the initial percentage of modifier
stepwise gradients by default), two chromatograms were selected as the critical factors, it is evident that the linear gradient results
to be tested about robustness. These chromatograms are shown less affected by the changes in these factors than the curved gradi-
in Fig. 7. Chromatogram in Fig. 7(a) corresponds to a linear gra- ent. The remaining factors, of which only temperature needs to be
dient from 6 to 40% of modifier (a binary mixture of ACN:MeOH considered, affect identically both elution programs. A similar con-
42:58) executed in 10 min, maintaining the temperature at 35 ◦ C. clusion can be derived from Fig. 10 that compares the robust regions
Chromatogram in Fig. 7(b) corresponds to a curved gradient (using for an overall desirability above 0.85. In both elution programs we
a curve 7 in the Waters programmer), from 5 to 40% of modifier see that this region is not circular but elliptical. This means that
(ACN:MeOH 69:31) executed in 10 min at 37 ◦ C. the procedures behave more satisfactorily from the point of view
To assess the robustness of both elution procedures, the cor- of robustness when both critical factors move in the same direction
responding chromatograms are chosen, and desirability windows but not when both factors changed in opposite ways. In practice,
configured departing from the confidence bands for prediction this means that the true robust region is a circle representing the
errors in the retention model (Table 3). Inner window limits are the worse situation with changes in gradient time of about ±0.3 min
same for any chromatogram simulated with this retention model and initial modifier percent of about ±0.6% for the linear gradient
no matter the gradient shape, but external limits can be defined but only ±0.3 min and ± 0.3% in the case of the curved gradient.
by the chromatographer after viewing the chromatogram appear- Thus, the linear gradient procedure appears slightly more reliable
ance. In Fig. 8(a) the configuration of peak’s desirability windows for routine analysis.
as developed by the routine is shown for the linear and curved gra-
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.
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- the Xunta de Galicia (Conecta-Peme program, reference IN852A-
grammes were executed experimentally during several days and 2013/75-0). S.T and P.R-M, acknowledge their contracts to the
the results have been summarized in Table 4. Results appear quite Spanish Ministry of Economy and Competitivity in the frame of the
similar in both cases with the peak of impurity B as the critical above-mentioned program.
component in the sample both in terms of resolution and peak tail-
ing, thus arousing doubts about the application in routine analysis Appendix A. Supplementary data
of any of those selected procedures, as can be expected from the
results of the CAMD robustness assessment. Supplementary material related to this article can be found, in
the online version, at doi:https://doi.org/10.1016/j.chroma.2019.
4. Conclusions 460439.
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