Retention modelling on SPAM-traps for two-dimensional liquid chromatography

(1)

1

title

Bachelor Thesis Scheikunde

Retention modelling on SPAM-traps for two-dimensional

liquid chromatography

by

Jesper Ruiter

30, November 2020

Student number

11682876

Research institute Accountable teacher

Van ‘t Hoff Institute for Molecular Sciences (HIMS)

Dr. Bob Pirok

Research group Daily supervisor

(2)

2

Abstract

Gradient-elution Reversed-Phase Liquid Chromatography (RPLC) is the most common mode in liquid chromatography. The method development for gradient-elution RPLC is often a time-consuming process. When this mode is applied in two-dimensional liquid chromatography (2D-LC), the time to develop a suitable method is increased even more. To decrease the method-development time in the application of the Stationary-Phase Assisted Modulation (SPAM) technique for 2D-LC, retention modelling could offer a solution, yielding shorter optimization time, using fewer chemicals, and saving more money. Using retention data retrieved from an analytical column in retention modelling to develop and optimize a separation method is known. However, using retention modelling for the method development in the SPAM technique in 2D-LC is not. In this work, was investigated if using retention data obtained from an analytical column could be used for the prediction of retention parameters for a trap column in the SPAM technique. To answer this, first isocratic experiments on an analytical column and several trap columns were performed, resulting in a statistically significant difference in selectivity for a trap column over time (24 hours), among trap columns and between an analytical column and a trap column. Scanning-gradients were carried out on the analytical column and a trap column. Data analysis showed that the retention data retrieved from the analytical column fitted best to the LSS model and that the retention data retrieved from the trap column fitted best to the NKM model. Dilution flow experiments were conducted to simulate 2D-LC and using the scanning-gradient data, retention times were predicted to see if the analytes (Sudan-I, propylparaben, and benzophenone) would be trapped on a trap. This resulted in good predictions except for one false negative.

Populair wetenschappelijke samenvatting

In de analytische chemie zijn er heel veel verschillende soorten technieken, maar één van de meest bekende is hogedruk vloeistofchromatografie, afgekort HPLC. Dit is vloeistofchromatografie waarbij de mobiele fase onder hoge druk door een dicht gepakte kolom met stationaire fase deeltjes wordt gedrukt. Een nadeel van HPLC is dat erg complexe mengsels niet echt geanalyseerd kunnen worden. Een oplossing is om tweedimensionale vloeistofchromatografie (2D-LC) te doen. Dit klinkt heel ingewikkeld, maar het komt simpelweg neer op het feit dat je nu twee kolommen achter elkaar hebt en je dus twee verschillende scheidingen doet. Deze scheidingen zijn vaak gebaseerd op verschillende eigenschappen zoals: hydrofobiciteit en lading. Bij 2D-LC moet wat uit de eerste kolom komt naar de tweede kolom worden vervoerd. Dit gebeurt bij de zogeheten modulatie-interface met behulp van een 8-poorts klep. Om geen sensitiviteit te verliezen in je scheiding, kan deze overdracht worden uitgevoerd met kleine kolommetjes, ook wel trap kolommen genoemd om je stoffen even vast te houden in de trap

(3)

3 om ze vervolgens naar de 2e_{kolom te laten gaan. De toepassing van trap kolommen gebruik in 2D-LC} wordt Stationary-Phase Assisted Modulation of kortweg SPAM genoemd.

Meestal duurt het even voordat je een methode hebt ontwikkeld die goed werkt en je een goede scheiding hebt. Door 2D-LC te doen is deze methode ontwikkeling nóg moeilijker en langer dan als je normale eendimensionale vloeistofchromatografie zou doen, wat natuurlijk nog langer wordt als je de SPAM-techniek óók moet ontwikkelen. Om de methode ontwikkeling te versnellen kan er gebruik gemaakt worden van retentie modellering. Hiermee hoef je maar een paar metingen te doen en met die data kan je dan andere methodes programmeren. Dit bespaart tijd, chemicaliën en hiermee dus ook geld. In dit verslag is er gekeken of deze modellering ook mogelijk is voor de SPAM-techniek in 2D-LC door eerst te kijken of de trap kolommen en de analytische kolom vergelijkbare selectiviteiten geven voor de onderzochte stoffen (Sudan-1, propylparabeen en benzofenon) met behulp van isocratische experimenten (een constante samenstelling van je mobiele fase). Kort gezegd betekent een hoge selectiviteit een makkelijke scheiding tussen de twee stoffen. Dit was gedaan om te kijken of er veel verschil zat tussen de kolom en de trap kolommen. Daarna werden er scan gradiënten uitgevoerd (bij een gradiënt verandert de samenstelling van je mobiele fase) om te bepalen welke van de vijf meest gebruikte modellen in retentie modellering het beste past bij de data van de analytische kolom en de trap kolom. Tenslotte is er gekeken of er met behulp van die modellen juiste voorspellingen konden worden gedaan met betrekking tot de retentie op SPAM-kolommen.

(4)

4

1. Introduction

‘The numbers tell the tale’ is a saying that applies to many different techniques that analytical chemists use to obtain data. One of these is called: High-Performance Liquid Chromatography (HPLC). HPLC is not only of great importance for analytical chemistry, but it is vital for a variety of fields, such as forensic toxicology, food science, and pharmaceutical chemistry.1_{Due to the vast application of HPLC in an} extensive number of fields, improving HPLC and developing new methods are meaningful not only for academic research but for society as well.

HPLC can separate molecules based on their polarity, charge, and size depending on the method.1_{Of all the retention mechanisms available, Reversed-Phase Liquid Chromatography (RPLC)} is most commonly used.2_{In RPLC, the analytes are separated due to a difference in retention between a} non-polar stationary phase and a polar mobile phase. Non-polar compounds adsorb more onto the stationary phase than the polar compounds and, for this reason, elute later than the polar compounds, resulting in a separation between the analytes. The time it takes for all the compounds to elute can be relatively long. The time can be reduced by using a gradient in which the percentage of organic modifier in the mobile phase is increased during the measurement. The method development for gradient-elution RPLC can be time-consuming since usually, a considerable number of experiments have to be performed to find the right method.3_{To facilitate the method development, an approach called retention modelling} can be used, which allows for the prediction of optimal method parameters for a specific chromatographic system and sample. In addition, retention modelling results in faster method development.4

Contrary to other techniques, for example, Capillary Electrophoresis (CE) and Gas Chromatography (GC), LC has lower plate numbers.5_{For this reason, LC is not able to cope with} complex mixtures that contain a substantial number of analytes. A solution to increase the resolving power and plate numbers is to perform two-dimensional liquid chromatography (2D-LC) in which fractions of the mobile phase from the first-dimension (1_{D) column are transferred to another column} (second-dimension, 2_D).6,7 _{These fractions subsequently undergo another separation in the}2_{D column.} However, there are some disadvantages with 2D-LC such as a decrease in detection sensitivity due to subsequent dilution of the sample, longer analysis times, and the method development is more challenging and time-consuming compared to 1D-LC.5_{An essential part of a 2D-LC setup is the} modulation interface that transfers the fractions of the 1_{D to the}2_{D column. Usually, the fraction transfer} is achieved by a valve that is equipped with storage loops. Alternatively, low-volume trapping columns, called ‘traps’ that contain a stationary phase similar to the 2_{D column can be used instead, which was} first demonstrated in 2D-LC by Vonk et al.5,8_{This modulation technique is referred to as} Stationary-Phase Assisted Modulation (SPAM) and is a form of active modulation, meaning that the volume and the concentrations of the analytes change in contrast to passive modulation. See Figure 1 for a schematic overview of the SPAM modulation interface, copied from Pirok et al.5

(6)

6 In short, the analytes are trapped and the mobile phase of the 1_{D column passes through the trap, leaving} the chromatographic system.5_{When the valve switches, the mobile phase of the}2_{D column is pumped} through the trap, eluting the trapped analytes into the 2_{D column where further separation takes place.} Using the SPAM approach the volume of the mobile phase fractions from the 1_{D column is reduced.}9 As a result, the detection sensitivity is improved, and the dilution effect is diminished. Despite these advantages, there are some disadvantages when using SPAM. Analytes can be lost during the measurement, which occurs when the analytes from the 1_{D fractions have insufficient retention on the} trap throughout the total modulation time. Furthermore, the traps potentially reduce the toughness of the system, and when the traps are non-identical a difference in performance can occur.5,10

A possible solution to solve the problems with 2D-LC and SPAM is to use retention modelling to allow for a faster method development when using 2D-LC with a SPAM interface. When using SPAM in 2D-LC, sometimes the strength of the 1_{D effluent needs to be reduced, which can be achieved by} using a dilution flow. Using analytical column data to predict analytical column retention times is possible.11,12,13_{However, using analytical column data to predict retention on a trap column is, until now,} not investigated. This is different from standard retention modelling, for the reason that these two columns have been fabricated under different circumstances and there is a major difference in length.

In this work, first, the day-to-day repeatability of a single trap, trap-to-trap repeatability of different traps, and the difference between an analytical column versus a trap column were investigated with regards to the selectivity values. Second, scanning-gradient experiments were performed on an analytical column and a trap column to examine which of the five common retention modelling models (see theory section) is the best fit for the data and if this differs between the two. Finally, using this scanning gradient data, the retention times will be predicted and compared with performed dilution flow experiments to simulate 2D-LC.

(7)

7

2. Theory

Five common models are used in retention modelling: linear-solvent-strength (LSS), adsorption (ADS), mixed-mode (MM), quadratic (QM) and the Neue-Kuss model (NKM). Each of these models will be briefly explained in the following sections for a better understanding of the different models.

2.1. Linear-solvent-strength model

The LSS model also referred to as the exponential model, is often used for RPLC and yields a log-linear equation that contains two parameters: the retention factor (k) and the fraction organic modifier (𝜑) see Equation 1.2,11

ln(𝑘) = ln(𝑘0) − 𝑆𝜑 (1)

In Equation 1 k0 describes the retention factor of the analyte when the mobile phase consists of 100% water (when 𝜑=0) and is usually extrapolated. The solvent strength parameter (S) describes the change in retention with increasing strength of the mobile phase.11,12

2.2. Adsorption model

In Normal-Phase Liquid Chromatography (NPLC) the description of the retention is based on localized surface adsorption, which is the basis for the ADS model.12,14_{Therefore, the ADS model is} often used to describe NPLC, which has the form that can be seen in Equation 2.12,15

ln(𝑘) = ln(𝑘0) − 𝑛 ln(𝜑) (2)

Where n is the ratio between the occupied surface areas of the analyte and water molecules.

2.3. Mixed-mode model

The equation for this model was established to give a quantitative description of retention in Hydrophilic Interaction Liquid Chromatography (HILIC).16 _{It is assumed that the retention mechanism} of HILIC consists of a combination of electrostatic interactions and partitioning processes.17,18_{As a} result, the best description of HILIC would be as a mixed-mode retention mechanism. For an accurate description of the retention behaviour of HILIC the following equation was proposed by Jin et al. (see Equation 3).16

ln(𝑘) = ln(𝑘0) + 𝑆1𝜑 + 𝑆2 ln (𝜑) (3)

Differing from the LSS and ADS model, the MM, QM and NK models have three parameters. Here, S1 takes the interaction of solutes with the stationary phase into account and S2 considers the interaction of solutes with the solvents.12

(8)

8

2.4. Quadratic model

The QM model is a more comprehensive form of the LSS model, because a third parameter is introduced, meaning that the QM model is capable of characterizing retention across a more considerable range of mobile phase compositions.19_{The retention can be calculated using Equation}

4.12,19

ln(𝑘) = ln(𝑘0) + 𝑆1𝜑 + 𝑆2𝜑2 (4)

Contrary to the MM model, S1 and S2 show the effect of the volume fraction of the strong solvent.

2.5. Neue-Kuss model

This is an empirical model, formulated by Neue and Kuss, in which an equation is used that describes the relationship between the composition of the solvent and the retention.20_{Moreover, the} equation can either be differentiated to obtain isocratic retention from gradient data or be integrated to predict gradient retention from isocratic data. See Equation 5 for the full equation.

ln(𝑘) = ln(𝑘0) + 2 ln(1 + 𝑆2𝜑) − 𝑆1𝜑

1 + 𝑆2𝜑 (5)

In this equation, S1 represents the slope and S2 represents the curve coefficient.12,20

3. Experimental

3.1 Sample preparation for the experiments

Four compounds were analysed: uracil, Sudan I, propylparaben and benzophenone, in this order. Storage solutions of these compounds were prepared as follows: A 50:50 Milli-Q water: acetonitrile (ACN) solution was prepared by mixing 10 mL of Milli-Q water and 10 mL of ACN. 0.0209 g of uracil, 0.0023 g of Sudan I, 0.0022 g of propylparaben and 0.0014 g of benzophenone were added to the 50:50 mixtures to obtain separate 100 ppm storage solutions. Except for the uracil for which a 1000 ppm storage solution was prepared.

For the HPLC measurements, samples of these four compounds were prepared. The samples for analysis were made of 10 ppm for all compounds. For uracil (S1): 10 µL of the storage solution was put in an LC vial and filled up to 1 mL with Milli-Q water. The propylparaben (S5) and benzophenone (S6) samples were prepared by adding 100 µL of the storage solution in an LC vial. This was filled up to 1 mL with Milli-Q water. For the Sudan I (S4) sample: 500 µL of the storage solution was put in an LC vial, which was filled up to 1 mL with ACN.

3.2 Isocratic experiment method for the column and traps

The following method was used: the mobile phase consisted of a buffer/ACN (v/v, 95/5), which was referred to as mobile phase A, and ACN/buffer (v/v, 95/5), which was referred to as mobile phase B. The buffer was an ammonium formate buffer with a 5 mM concentration at a targeted pH of 3. This buffer was prepared by adding 0.056 g of ammonium formate and 0.193 mL of formic acid (98%) to 1 L of Milli-Q water. Each sample, as described in 3.1, was measured five times at a selected fraction of

(9)

9 organic modifier (𝜑). The selected 𝜑 values were 0.45, 0.50 and 0.55. When the measurements were performed for a selected 𝜑, a washing programme took place that consisted of 100% B for 1 minute followed by a re-equilibration period for 2 minutes with the subsequently selected 𝜑. The flow rate was 0.4 mL·min-1_{, the temperature was kept at 20 ˚C throughout the measurements and the injection volume} was 5 µL (except for the Sudan I where the injection volume was 3 µL). For the traps, each run was 2 minutes and for the column, each run lasted 10 minutes. The 95/5 mobile phase bottles were the same throughout the isocratic experiment. Note: the traps used for the isocratic experiments are labelled T4, T5 and T6. The reason for this is that the traps labelled T1, T2 and T3 started to show severe peak splitting.

3.3 Scanning-gradient experiments for the column and trap

The used mobile phase was the same as described in section 3.2. The gradients that were carried out started from 0 min to 0.5 min isocratic 95% A, followed by a linear gradient to 95% B in 3, 6 or 9 min, respectively. In all the gradient programmes, 100% B was maintained for 1 min and brought back to 95% A in 0.01 min. Mobile phase A was kept at 95% for 1.99 min before a new run was started. The flow rate was 0.4 mL·min-1_{, the temperature was kept at 20 ˚C throughout the measurements and the} injection volume was 5 µL. The sample was a mix of the S1, S4, S5 and S6 samples (described in 3.1), which was made by adding 0.3 mL of S1, S4, S5 and S6 in an LC vial. The mix sample was measured ten times for each gradient programme. The 95/5 mobile phase bottles were the same throughout the scanning-gradient experiments.

3.4 Dilution flow experiment methods for the column and trap

3.4.1 Dilution flow method 1

For the dilution flow (DF) methods two pumps were used and five different versions were made. For the 1D pump, isocratic 50% A and 50% B was used throughout all of the experiments. The mobile phase consisted of a 1 L H2O/ACN (v/v 50/50) mixture. Both lines of the 1D pump were placed in this flask. The flow rate was 0.05 mL·min-1_{. For the 2D pump, the same mobile phase was used as previously} described in section 3.2. The gradients that were carried out started from 0 min to 3 min isocratic 50%, 40%, 30%, 20%, 10%, and 0% A, respectively, followed by a linear gradient to 100% B in 3 min. 100% B was maintained for 1 min and subsequently brought back to 50%, 40%, 30%, 20%, 10% or 0% A in 0.01 min. Mobile phase A was kept at 50% for 2.99 min before a new run was started. The flow rate was 0.55 mL·min-1_{and the injection volume was 5 µL. The sample that was measured was the same} mix sample as described in section 3.3. The sample was measured ten times for each DF method version. After five measurements a blank run with ACN was performed using the method that was running at that moment. The mobile phase bottles were the same throughout this dilution flow experiment. It must be noted that this method was used for testing, leading to the development of method 2 and 3, which are described below. For this reason, there are no results reported for this method in section 4.3.1

(10)

10

The same as dilution flow method 1, but in this method, the used mobile phase in the 2D pump consisted of a buffer, which was mobile phase A and ACN, which was referred to as mobile phase B. The buffer was an ammonium formate buffer with a 5 mM concentration at a targeted pH of 3. This buffer was prepared by adding 0.056 g of ammonium formate and 0.202 mL of formic acid (95%) to 1 L of Milli-Q water. The other parameters were the same as was described in 3.4.1. As before, five versions were made of this method starting at different percentages mobile phase A similar to the DF method 1. The mobile phase bottles were the same throughout this dilution flow experiment.

This DF method is identical to DF method 2 except for the flow rates. In this method the flow rate of the 1D pump was 0.04 ml·min-1_{and the flow rate of the 2D pump was 0.56 mL·min}-1_{. The mobile} phase bottles were the same throughout this dilution flow experiment.

3.5 Data processing

The MOREPEAKS programme was used to determine the AIC values in the scanning gradient experiments for the different models and to predict the retention times in the dilution flow experiments. Microsoft Excel was used for other data processing (retention factor calculation etc.)

3.6 Chemicals

Milli-Q water (18.2 MΩ cm) was obtained from a purification system (Arium 611UV, Sartorius, Germany). Acetonitrile (ACN, LC-MS grade) was purchased from Biosolve Chemie (Dieuze, France). Formic acid (FA, 98%) and propylparaben (propyl 4-hydroxybenzoate, ≥99%) were purchased from Fluka (Buchs, Switzerland). Ammonium formate (AF, ≥99%), Sudan I (≥97%), uracil (≥99%) and benzophenone (≥99%) were obtained from Sigma Aldrich (Darmstadt, Germany).

3.7 Hardware

Experiments were performed on an Agilent 1290 series Infinity II 2D-LC system (Waldbronn, Germany) configured for one-dimensional operation. The system included a binary pump (G4220A), an autosampler (G4226A) equipped with a 20-µL injection loop, a thermostatted column compartment (G1316C), and a diode-array detector (DAD, G4212A) equipped with an Agilent Max-Light Cartridge Cell (G4212-60008), 10 mm path length, 𝑉det= 1.0 µL). The dwell volume of the system was 0.187 mL for the 1D pump and 0.0687 mL for the 2D pump. The injector needle drew and injected at a speed of 10 μL·min-1, with a 2 s equilibration time. The system was controlled using Agilent OpenLAB CDS Chemstation Edition (Rev. C.01.10). In this study a Kinetex 1.7 µm C18 100 Å 50×2.1 mm column (Phenomenex, Torrance, CA, USA) was used as well as Kinetex 1.7 µm C18 100 Å 3×2.1 mm traps (Phenomenex, Torrance, CA, USA).

(11)

11

4. Results and discussion

4.1 Isocratic experiment on the analytical column and trap columns

4.1.1. Day-to-day repeatability of the trap columns

First, the repeatability of a single trap column over time was determined, by calculating the selectivity values of the traps. If there is no statistically significant difference in selectivity between the runs this would suggest that there is day-to-day repeatability for a trap column. The selectivity (α) is a quantitative measure for the difference in the behaviour of two analytes in a specific phase system.21 The α is by definition ≥ 1 and a high α indicates a good separation of the two analytes. The stationary phase material of the traps is similar to that of the column that was used. For this reason, it was expected that the selectivity values of the traps were similar to those of the column.

In Table 1 the calculated selectivities of two sequences of T4 at 𝜑=0.45 are shown. One run consisted of five measurement repeats. The second run was performed 24 hours later for all traps. In the supporting information, all the calculated selectivities for all the traps at the selected percentages 𝜑 are tabulated. In Figure 2, the selectivities of T4 at a 𝜑 of 0.45 are presented with box and whisker plots. For the other 𝜑s see the supporting information.

Table 1: Calculated selectivities of run 1 and 2 of T4 at 0.45𝜑.

T4 run 1 T4 run 2 T4 run 1 T4 run 2 T4 run 1 T4 run 2

Run (0,45𝜑) α1,2 α1,2 α2,3 α2,3 α1,3 α1,3 1 11.998 12.921 2.546 2.532 4.712 5.103 2 11.875 13.067 2.510 2.553 4.731 5.117 3 11.639 13.071 2.483 2.576 4.688 5.074 4 11.844 13.012 2.506 2.557 4.725 5.090 5 11.838 13.004 2.486 2.554 4.761 5.091

(12)

12 From the different α values in Figure 2 it can be seen that the values of the first run differ from the second run. To be sure that the α values vary statistically between runs, t-tests were carried out with Excel to determine whether the difference between the two values is statistically significant. First, an F-test was performed to see if the variances are assumed to be equal or unequal. Based on this result the t-test assuming (un)equal variances was carried out. The results of all the t-t-tests of all the traps can be seen in the supporting information. For T4 only the α2,3 of run 1 + run 2 at 0.50𝜑 showed a statistically insignificant result. Here run 1 indicates day 1 and run 2 indicates day 2. For T5 only the α1,3 of run 1 + run 2 at 0.50𝜑 and for T6 only the α2,3 of run 1 + run 2 at 0.55𝜑 showed a statistically insignificant result. From Figure 2 it can be seen that the α1,2 and α1,3 values of the different runs are not overlapping, indicating different values. For the α2,3 values, there is some overlap, but according to the

t-tests, the selectivities between compounds 2 and 3 were mostly statistically significant for the same

trap. Therefore, it was concluded that the retention on trap columns was not repeatable from day-to-day, because the difference in selectivities was mostly statistically significant.

4.1.2. Trap-to-trap repeatability

The next step was to investigate the trap-to-trap repeatability of the trap columns. This was again carried out by calculating the selectivity values. The trap columns should contain a similar stationary phase and are considered to be equal by the manufacturer, which is why it was expected that two different traps showed similar selectivities. However, in 4.1.1. it was concluded that there was no day-to-day repeatability of a single trap column (it must be mentioned that trap columns are normally used

Figure 2: Box and whisker plot showing the day-to-day repeatability of T4 at 0.45𝜑. The boxes are based on 5 values for each α and run, respectively. The blue boxes represent the selectivity values of T4 run 1 and the orange boxes represent the selectivity values of T4 run 2. The boxes represent 50% of the data ranging from the first to the third quartile.

(13)

13 as pre-columns to protect the analytical column, which is why selectivity is not something that the trap columns are tested on). This means that there also should be no trap-to-trap repeatability. In Table 2 the calculated selectivities from the first run of T4 and T5 at 𝜑=0.45 are shown. In the supporting information, all the calculated selectivities for all the comparisons between traps at the selected percentages 𝜑 are tabulated. In Figure 3, the difference in selectivities of T4 and T5 at a 𝜑 of 0.45 are presented with box and whisker plots. Note that in Figure 3 the values of the second run are also taken into account. Please see the supporting information for the results of the other 𝜑s.

Table 2: Calculated selectivities from run 1 of T4 and T5 at a 𝜑 of 0.45𝜑.

T4 run 1 T5 run 1 T4 run 1 T5 run 1 T4 run 1 T5 run 1

Run (0,45𝜑) α1,2 α1,2 α2,3 α2,3 α1,3 α1,3 1 11.998 12.162 2.546 2.462 4.712 4.940 2 11.875 12.177 2.510 2.464 4.731 4.942 3 11.639 12.150 2.483 2.454 4.688 4.950 4 11.844 12.170 2.506 2.477 4.725 4.912 5 11.838 12.052 2.486 2.450 4.761 4.918

As previously described in section 4.1.1., the same approach was used when two different traps were compared with each other. Using t-tests in Excel it was investigated if there was a significant difference between the two values. The results from the t-tests of all the traps can be seen in the supporting information. For T4 and T5 the α1,2 at 0.50𝜑, α1,2 at 0.55𝜑 and the α2,3 at 0.55𝜑 showed a statistically insignificant result. But from Table S68 in the supporting information can be seen that when

Figure 3: Box and whisker plot showing the trap-to-trap repeatability of T4 and T5 at 0.45𝜑. The boxes are based on both runs of the traps including 10 values for each α, respectively. The blue boxes represent the selectivity values of T4 (of both days) and the orange boxes represent the selectivity values of T5 (of both days). The boxes represent 50% of the data ranging from the first to the third quartile.

(14)

14 the second run was compared between T4 and T5, only the α1,2 at 0.55𝜑 showed an insignificant result. When comparing T4 with T6 for the first run only the α2,3 at 0.50𝜑 gave an insignificant result. On the contrary, in the second run, the α1,2; α2,3 and α1,3 at 0.55𝜑 all gave insignificant results. Comparing T5 with T6 for the first run yielded a significant difference between selectivities at all 𝜑s. In the second run, only the α1,3 at 0.50𝜑 and α1,3 at 0.55𝜑 had an insignificant difference. From Figure 3 it can be seen that the variance of T4, most notably, at the α1,2 is greater compared to T5. Moreover, the variances of the different selectivities are either not or only partially overlapping. This partial overlap could be explained by the t-tests that showed a statistically insignificant result. When comparing T4 with T6 (see supporting information) it can be seen that the variance of T4 is also greater than T6, but there is more overlap compared to T4 and T5. T6 has a larger variance compared to T5 and the variances of these two traps showed more overlap.

Based on these results it can be assumed that there is no pattern when the different traps show an insignificant result. Even between runs, different outcomes of the t-test are observed and when T5 and T6 were compared the results of the t-tests of the first run all indicated a statistically significant difference between the two trap columns. Thus, it was concluded that the retention on trap columns was not repeatable from trap-to-trap, because the difference in selectivities was mostly statistically significant.

4.1.3. The selectivities between the analytical column and trap columns

The final part of the isocratic experiment was the comparison of the selectivity of trap columns with the selectivity of an analytical column. t-tests were carried out to see if there was a statistically significant difference in selectivity between the traps and the column. According to literature, the traps contain a similar stationary phase as the column used in the second dimension.5_{Because the stationary} phases are similar, it was expected that the selectivities were as well. Parameters that were kept the same were: pH, temperature, mobile phase, and the same samples. The main difference between the column and the traps is that the column is larger compared to the traps (50 mm vs 3 mm).

In Table 3 the calculated selectivities from the first run of T4 and the column are presented. See the supporting information for all the calculated selectivities for all the comparisons between the traps and the column at the selected percentages 𝜑. In Figure 4 the difference in selectivity between T4 and the column at a 𝜑 of 0.45, 0.50 and 0.55, respectively are presented with box and whisker plots.

Table 3: Calculated selectivities from the first run of T4 and the column at 0.45𝜑.

T4 run 1 column T4 run 1 column T4 run 1 column

Run (0,45𝜑) α1,2 α1,2 α2,3 α2,3 α1,3 α1,3 1 11.998 17.715 2.546 2.931 4.712 6.045 2 11.875 17.744 2.510 2.932 4.731 6.053 3 11.639 17.773 2.483 2.937 4.688 6.051 4 11.844 17.723 2.506 2.926 4.725 6.056 5 11.838 17.764 2.486 2.935 4.761 6.052

(15)

15

F-tests followed by the appropriate t-tests were conducted the see whether the selectivities of

the traps vary significantly with those of the column. The results from these t-tests can be seen in the supporting information. All the t-tests for all the selectivity values between T4 and the column resulted in a statistically significant difference, which is visible in the box-and-whisker plot. The other traps, T5 and T6, showed the same result. The boxes in Figure 4 are based on 5 measurement repeats obtained in a single day for both the analytical column and trap column. The variance of the analytical column is small compared to the trap column. In addition, there is no overlap between the selectivity values, which is also observed for the other trap columns. The analytical column had a larger volume of stationary phase compared to the traps, which means that the extra volumes, such as the tubing volume and the dwell volumes are relatively larger for the trap column compared to the analytical column. For this reason, it was assumed that there would be a difference in retention times between the trap column and the analytical column. The retention factor can be calculated with this formula: 𝑘 = 𝑡𝑅

𝑡0− 1 The length

of the analytical column is larger than that of the trap column, but this difference in length alone is not enough to explain the observed differences in retention. It is assumed that the ratio stationary phase: length is different for the trap column, due to a potential difference in porosity (ε).

From the obtained results can be concluded that there is a statistically significant difference in selectivity between the traps and the column.

Figure 4:Box and whisker plot showing the difference in selectivity of the column and T4 at 0.45𝜑. The boxes are based on 5 values for each α, respectively. The blue boxes represent the selectivity values of T4 run 1 and the orange boxes represent the selectivity values of the analytical column. The boxes represent 50% of the data ranging from the first to the third quartile.

(16)

16

4.2 Scanning-gradient experiments

There are several different models proposed for retention modelling in section 2: LSS, ADS, MM, QM and the NKM.22_{For a more elaborate explanation of each model the reader is referred to the} theory section 2. To determine if the retention data ‘fits’ a retention model, the Akaike Information Criterion (AIC) is commonly used, see Equation 6.23

𝐴𝐼𝐶 = 2𝑝 + 𝑛 [ln (2𝜋∙𝑆𝑆𝐸

𝑛 ) + 1] (6)

The AIC is a ‘goodness-of-fit’ indication in which a more negative value indicates a better fit of the model for the analysed data. T5 was chosen to conduct the scanning gradients on because this trap showed the smallest variance according to the isocratic results.

4.2.1 Scanning-gradients on the trap column (T5)

In Figure 5 the AIC values of T5 are shown for the different models, taking into account all repeats that were obtained with three measured gradient slopes (3, 6 and 9 minutes, respectively), for each of the compounds that were analysed during the scanning-gradient experiment. From Figure 5 the conclusion can be drawn that all the models have negative AIC values and, therefore, the data fitted all of the models to some extent. The positive AIC value for uracil in the ADS model can be explained by the reason that uracil does not have retention in RPLC and for this reason, can sometimes not be modelled. In the supplementary information (Table S76) the actual values can be seen. For the compounds of interest (Sudan-1, propylparaben and benzophenone since uracil was used as a t0 marker) the lowest AIC values were obtained for the NKM model, meaning that the retention data retrieved from the trap column best fitted the NKM model, according to this plot. The AIC values of the LSS model

-20 -15 -10 -5 0 5 10 15 LSS ADS MM QM NKM AIC

Uracil Sudan-1 Propylparaben Benzophenone

Figure 5: The AIC values of T5 for the different models using all repeats that were obtained with

three measured gradient slopes (3, 6 and 9 minutes). The compounds that were analysed during the scanning gradient experiment are also shown and indicated with different colours.

(17)

17 only differ slightly from those of the NKM model. The QM model had the least negative AIC values and is thus considered to be the worst fit according to Figure 5.

4.2.2 Scanning-gradients on the analytical column

In Figure 6 the AIC values of the column are shown for the different models, taking into account all repeats that were obtained with three measured gradient slopes (3, 6 and 9 minutes, respectively), for each of the compounds that were analysed during the scanning gradient experiment. In contrast with the result of T5, the LSS model yielded the most negative AIC values for the compounds of interest, which suggests that the retention data obtained from the analytical column fitted best to the LSS model. In the supplementary information (Table S77) The AIC values can be seen. The ADS, MM and QM models all yielded negative AIC values as well, but these were less negative compared to the LSS model. For the NKM model, positive AIC values were obtained, indicating a poor fit of the retention data from the analytical column with the NKM model.

The trap column and analytical column data were best described by different models. The trap column data was only marginally better described by the NKM model instead of the LSS model. Whereas the analytical column data was best described by the LSS model and worst described by the NKM model. The fact that the worst model for the analytical column data is the best model for the trap column data is quite an interesting result.

For the final step, a model had to be selected and based on these results the LSS model was chosen to continue the project with. The reason for this was because the LSS model fitted the data the

-20 -15 -10 -5 0 5 10 15 20 25 LSS ADS MM QM NKM AIC

Uracil Sudan-1 Propylparaben Benzophenone

Figure 6: The AIC values of the column for the different models using all repeats that were obtained

with three measured gradient slopes (3, 6 and 9 minutes). The compounds that were analysed during the scanning gradient experiment are also shown and indicated with different colours.

(18)

18 best for the column. In addition, for T5 the LSS model showed the second-best AIC values and therefore a good fit of the trap data.

In section 3.3 it was described that the final percentage organic modifier in the scanning-gradient method was 95%. However, since 95/5 bottles of mobile phase were used the actual final percentage was 90.5%. This was taking into account when obtaining the scanning-gradient data that was used for the predictions carried out in section 4.3.

4.3 Dilution flow experiments

Using the MOREPEAKS programme, the retention times of the different dilution flow experiments were predicted based on the scanning gradient data using the LSS model. With the retention data obtained in the scanning-gradient experiments, the retention times of the analytes were predicted. Column scanning-gradient data was used to predict retention on the column and the trap. Scanning-gradient data of T5 was used for predicting the retention times on T5. A detailed description of the methods can be found in the experimental section. Within each method, there were six versions indicated with DF followed by a number. The difference between these DF versions is the starting percentage water, of which an overview is given in Figure 7 and 8.

0 20 40 60 80 100 120 0 2 4 6 8 10 12 % wat er in mob ile p h ase time (min) DF0 DF1 DF2 DF3 DF4 DF5

(19)

19

4.3.1 Using column and T5 scanning gradient data for the prediction of retention times on the column and T5

In Table 4 and 5 results of DF4 and DF5 of the dilution flow methods 2 and 3, respectively are shown for the analytical column and trap column. Table 4 and 5 show that the prediction outcomes for dilution flow versions (DF4 and DF5) are accurate for the analytical column. For T5 there is some deviation in the predicted retention time and the measured retention time but what is more relevant is the outcome of the prediction. The goal of this experiment was to trap S4, S5 and S6, which in this instance meant that the compounds needed to elute after the 3 minutes isocratic part of the method. In reality, it is not 3 minutes because these 3 minutes do not include the dwell time and dead volume of the trap column and analytical column.

The prediction can have, therefore, two outcomes. The first one is that the model predicts retention of fewer than 3 minutes or ‘not trapped’. The second one is that retention of more than 3 minutes is predicted or ‘trapped’. When the outcome of the prediction matched that of the outcome that was observed in the measurements the cells are highlighted in green, if not than the cells are highlighted in red. The column scanning gradient data was based on all repeats that were obtained with three measured gradient slopes (3, 6 and 9 minutes). The T5 scanning gradient data was also based on all repeats that were obtained with three measured gradient slopes (3, 6 and 9 minutes) for T5.

0 20 40 60 80 100 120 0 2 4 6 8 10 12 % wat er in mob ile p h ase time (min) DF0 DF1 DF2 DF3 DF4 DF5

(20)

20

Table 4: Measured and predicted retention times of the analysed compounds of the column and T5 for

method 2. Prediction based on the column and T5 scanning gradient data including all measurement repeats obtained with three measured gradient slopes (3, 6 and 9 minutes) with the LSS model. Green indicates that the predicted outcome matched the measured outcome. Red indicates if the predicted outcome does not match the measured outcome.

Method 2 column on column

tR (min) S1 tR (min) S4 tR (min) S5 tR (min) S6

DF4, predicted 0.1691 5.7011 4.5311 4.9803 DF4, measured 0.5528 5.6133 4.4364 4.8781 DF5, predicted 0.1853 5.8175 4.7843 5.1749 DF5, measured 0.5929 5.7138 4.681 5.0649 Method 2 T5 on T5

DF4, predicted 0.01519 4.3724 0.3998 2.7806

DF4, measured 0.3129 4.346 1.5419 3.4904

DF5, predicted 0.01651 4.5892 3.7365 4.0061

DF5, measured 0.3154 4.5455 3.7403 4.0034

Table 5: Measured and predicted retention times of the analysed compounds of the column and T5 for

method 3. Prediction based on the column and T5 scanning gradient data including all measurement repeats obtained with three measured gradient slopes (3, 6 and 9 minutes) with the LSS model. Green indicates that the predicted outcome matched the measured outcome. Red indicates if the predicted outcome does not match the measured outcome.

Method 3 column on column

DF4, predicted 0.1695 5.6919 4.5452 4.9845 DF4, measured 0.6486 5.5777 4.4275 4.857 DF5, predicted 0.1891 5.8085 4.7934 5.1771 DF5, measured 0.7182 5.6897 4.6709 5.044 Method 3 T5 on T5

DF4, predicted 0.01523 4.3812 0.5498 3.2856

DF4, measured 0.3813 4.3487 1.6327 3.5334

DF5, predicted 0.01682 4.5964 3.7594 4.0262

DF5, measured 0.3839 4.5478 3.7488 4.0051

From Table 4 and 5, the conclusion can be drawn that using the analytical column scanning-gradient data accurate predictions of the retention times can be made for DF4 and DF5. In addition, all

(21)

21 of the predicted outcomes (trapped or not trapped) all matched those of the measured outcomes. When the T5 scanning gradient data was used to predict retention on T5 the times deviated more severely compared to the column. From the previous section regarding the poor repeatability of the trap columns, this can be expected. However, the prediction outcomes were almost all correct, except for the DF4 prediction in both method 2 and 3. This is not ideal, because the model predicts that the analyte will not have enough retention and therefore will not be trapped. In the measurement, the analyte was trapped. For the final application, this false negative did not matter.

4.3.2 Using retention modelling for the prediction of retention times on T5

In Table 6 and 7, the results for dilution flow methods 2 and 3, respectively are shown. The column scanning gradient data was based on all repeats that were obtained with three measured gradient slopes (3, 6 and 9 minutes). Each method consisted of six versions indicated by DF and ranging from DF0 to DF5. In the tables, the average measured retention times and the predicted retention times from the retention modelling (LSS model) are tabulated. The x indicates that no separate peak was measured for this compound, meaning that it eluted with t0.

Table 6: Measured and predicted retention times of the analysed compounds of T5 for method 2.

Prediction based on the column scanning gradient data including all measurement repeats obtained with three measured gradient slopes (3, 6 and 9 minutes) with the LSS model. Green indicates that the predicted outcome matched the measured outcome. Red indicates if the predicted outcome does not match the measured outcome.

Method 2 column on trap

DF0, predicted 0.01500 0.3917 0.02493 0.06754 DF0, measured 0.3243 x x x DF1, predicted 0.01500 0.9447 0.05017 0.1655 DF1, measured 0.3420 0.8728 x x DF2, predicted 0.01500 2.3072 0.1394 0.4459 DF2, measured 0.3433 x x 0.5029 DF3, predicted 0.01503 4.1132 0.4555 1.2496 DF3, measured 0.3129 4.0654 0.5472 1.1265 DF4, predicted 0.01522 4.5803 1.5750 3.4971 DF4, measured 0.3129 4.3460 1.5419 3.4904 DF5, predicted 0.01667 4.8383 3.8800 4.2432 DF5, measured 0.3154 4.5455 3.7403 4.0034

(22)

22

Table 7: Measured and predicted retention times of the analysed compounds of T5 for method 3.

Prediction based on the column scanning gradient data including all measurement repeats obtained with three measured gradient slopes (3, 6 and 9 minutes) with the LSS model. Green indicates that the predicted outcome matched the measured outcome. Red indicates if the predicted outcome does not match the measured outcome.

Method 3 column on trap

DF0, predicted 0.01500 0.3917 0.02493 0.06754 DF0, measured 0.3935 x x x DF1, predicted 0.01500 0.9595 0.05095 0.1683 DF1, measured 0.4123 0.9222 x x DF2, predicted 0.01500 2.3852 0.1453 0.4630 DF2, measured 0.4144 3.2460 x x DF3, predicted 0.01503 4.1457 0.4869 1.3225 DF3, measured 0.3857 4.0606 0.6053 1.1845 DF4, predicted 0.01525 4.5938 1.7237 3.6264 DF4, measured 0.3813 4.3487 1.6327 3.5334 DF5, predicted 0.01702 4.8464 3.9310 4.2694 DF5, measured 0.3839 4.5478 3.7488 4.0051

From Table 6 and 7 the retention time for uracil was not accurately predicted in both methods and a much lower value was predicted compared to the actual measured value. For the compounds of interest (S4, S5 and S6) the prediction was most accurate for S4 from DF1 onwards in both methods. For S5 and S6, the predicted values are frequently lower compared to the measured value. An interesting observation is that when the method started at a lower percentage of organic modifier, the prediction became more accurate. Both in method 2 and 3 the DF5, where the initial percentage organic modifier was 0%, predictions are close to the actual measured retention time. The observation that for method 2 S4 did have a measured retention time at DF1 and not at DF2 was most likely due to the precipitation of the Sudan-I in the mix sample. A new mix sample was made and replaced during the analysis of method 2 after which the S4 signal was no longer absent.

This dilution flow experiment started with a 3 minute isocratic part followed by a gradient part and the goal was to ‘trap’ the analytes of interest, which in this case meant that the compounds needed to elute after 3 minutes (not including dwell time and dead volume) because then they would elute of the column/trap with the gradient. This was realised in DF5 for both methods where S4, S5 and S6 all eluted after 3 minutes. The prediction needs to give a proper indication whether an analyte will be trapped or not, which is why it is not a major issue that the predictions for the early DF methods are a bit off compared to the measured values. These early predicted retention times do not resemble the measured retention times well, but the outcome of the prediction (trapped or not trapped) mostly matches

(23)

23 the outcome that was experimentally observed, which for the early DF variants is that compounds S5 and S6 will not be trapped.

There was one case in which the outcome of the prediction was wrong (indicated with red in

Table 7) in method 3 with DF2 it was anticipated that S4 would not be trapped, but the measured value

was 3.2460 minutes meaning that it would have been trapped. It must be noted that a false prediction is not ideal, but in this case, the false prediction led to the good outcome of trapping the analyte so for the final application it did not matter. In this case rather a false negative is observed, meaning that the analyte is trapped even though this was not predicted, than a false positive. A false positive would mean that the model predicts that an analyte will be trapped, but in reality, it is not, which would be bad. Fortunately, this was not observed in these measurements. The cases where the prediction outcome matched that of the measured outcome are highlighted green in Table 6 and 7. Of the 48 predictions that were performed in which analytical column scanning-gradient data was used to predict the retention times for the trap column, 47 of the 48 prediction outcomes (trapped or not trapped) matched the measurement outcome and only one false negative was observed.

Because the traps used in SPAM often contain similar stationary phase material to the column used it would be useful if accurate predictions can be made using the more reliable column instead of the trap. Based on these results there is a good chance that this is indeed possible because in all the predictions no false positives were observed. The observation of the false negative could indicate that more scanning-gradient data is needed to have a hundred percent accuracy with the prediction outcomes. In the supporting information, S78 can be seen in which the other models (ADS, MM, QM and NKM) were used to predict the retention times of DF5 of method 3. From these results, it can be concluded that, except for the NKM model, other models can give similar predictions. This indicates that method development for different LC modes could also be possible.

4.3.3. Accuracy of the retention modelling

To determine how accurate the retention modelling was, method 3 was further investigated. The data used was a collection of 30 data points, including all scanning gradient data that were obtained. By predicting the retention times using the LSS model based on the scanning gradient data for just one measurement insight was obtained for the spread that the model gave for each measurement. This is summarised in Figure 9 for compound S4 in method 3. In Figure 9 the spread of the retention times of compound S4 for the different dilution flow variants can be seen to be substantially small. For the other compounds, please see the supporting information, similar spreads were obtained indicating that the spread of the model is low.

(24)

24

5. Conclusion

In conclusion, it was determined using isocratic experiments that there was a statistically significant difference in selectivities for a single trap over time (24 hours), between similar traps, and between a column and a trap, which indicates that care should be taken when comparing these. Next, based on scanning-gradient experiments, the AIC values for the fit of the retention data on the different retention modelling models were determined. It was found that the NKM model resulted in the most negative AICs for a trap column, whereas for the analytical column the LSS model yielded the most negative AIC values. The LSS model was chosen to carry out the predictions with because the data retrieved from the analytical column fitted best for the LSS model and second-best for the trap column. Finally, using retention modelling, predictions of the retention times, based on the scanning-gradient data of the analytical column, for trap columns were carried out and compared with the measured values in the dilution flow experiments. It was found that for these methods it was possible to predict the outcome in 97.9% of the measurements using the LSS model and in the cases, it did not it was a false negative, meaning that the model predicts that an analyte will not be trapped, but in reality is it is trapped, 2.1% of the measurements, without complications for the application. The spread of the predicted retention times was also determined and was found to be low. In future research, it should be investigated if the predicted retention times can be used for different LC-modes, such as HILIC or NPLC, if this is applicable for more compounds, and if the other retention models could be used to accurately predict the outcome of trapping an analyte since in this work this was only slightly examined, by predicting the retention times for DF5 of method 3 using the other models.

(25)

25

Acknowledgements

This project could not have been completed without the help of others and I would like to thank the following people for helping me during this project:

Ing. Tom Aalbers, for giving me the lab tour at the beginning of the project to familiarise me with the labs.

Mimi den Uijl, for her great supervision, advice, and coffee card throughout the project.

Pascal Camoiras Gonzalez, for helping me get familiar with the HPLC systems and with the measurements.

Tobias, for learning with me at the beginning of the project.

Dr. Bob Pirok, for helping me with the data analysis and answering any questions I had. The Analytical Chemistry group, for providing me with a place to conduct my research. Jan van Maarseveen, for being my second examiner.

Supporting information

See the supporting information document.

References

1: Paré, J. R. J.; Bélanger, J. M. R. In Techniques and instrumentation in Analytical Chemistry, Paré, J. R. J.; Bélanger, J. M. R.; Elsevier Science Publisher BV: Amsterdam, 1997; 18, pp 37–59.

2: Snyder, L. R.; Dolan, J. W.; Gant, J. R. Gradient elution in high-performance liquid chromatography : I. Theoretical basis for reversed-phase systems. J. Chromatogr. A 1979, 165, 3–30.

3: den Uijl, M. J.; Schoenmakers, P. J.; Stoll, D. R.; van Bommel, M. R.; Pirok, B. W. J. Measuring and using scanning-gradient data for use in liquid-chromatography method-optimization. unpublished results

4: Pirok, B. W. J.; Pous-Torres, S.; Ortiz-Bolsico, C.; Vivó-Truyols, G.; Schoenmakers, P. J. Program for the interpretive optimization of two-dimensional resolution J. Chromatogr. A 2016, 1450, 27–37. http://dx.doi.org/10.1016/j.chroma.2016.04.061

5: Pirok, B. W. J.; Stoll, D. R.; Schoenmakers, P. J. Recent Developments in Two-dimensional Liquid Chromatography: Fundamental improvements for Practical Applications. Anal. Chem. 2019, 91, 240– 263. https://doi.org/10.1021/acs.analchem.8b04841

6: Pirok, B. W. J.; Gargano, A. F. G.; Schoenmakers, P. J. Optimizing separations in online comprehensive two-dimensional liquid chromatography. J. Sep. Sci. 2018, 41, 68–98. DOI: 10.1002/jssc.201700863

7: Gargano, A. F. G.; Duffin, M.; Navarro, P.; Schoenmakers, P.J. Reducing Dilution and Analysis Time in Online Comprehensive Two-Dimensional Liquid Chromatography by Active Modulation. Anal.

Chem. 2016, 88, 1785–1793. DOI: 10.1021/acs.analchem.5b04051

8: Vonk, R. J.; Gargano, A. F. G.; Davydova, E.; Dekker, H. L.; Eeltink, S.; de Koning, L. J.; Schoenmakers, P. J. Comprehensive Two-Dimensional Liquid Chromatography with Stationary-Phase-Assisted Modulation Coupled to High-Resolution Mass Spectrometry Applied to Proteome Analysis of

(26)

26 9: Baglai, A.; Blokland, M. H.; Mol, H. G. J.; Gargano, A. F. G.; van der Wal, S.; Schoenmakers, P. J. Enhancing detectability of anabolic-steroid residues in bovine urine by actively modulated online comprehensive two-dimensional liquid chromatography – high-resolution mass spectrometry. Anal.

Chim. Acta 2018, 1013, 87–97. https://doi.org/10.1016/j.aca.2017.12.043

10: Pirok, B. W. J.; Abdulhussain, N.; Aalbers, T.; Wouters, B.; Peters, R. A. H.; Schoenmakers, P. J. Nanoparticle Analysis by Online Comprehensive Two-Dimensional Liquid Chromatography combining Hydrodynamic Chromatography and Size-Exclusion Chromatography with intermediate Sample Transformation. Anal. Chem. 2017, 89, 9167–9174. DOI: 10.1021/acs.analchem.7b01906

11: Roca, L. S.; Schoemaker, S. E.; Pirok, B. W. J.; Gargano, A. F. G.; Schoenmakers, P.J. Accurate modelling of the retention behaviour of peptides in gradient-elution hydrophilic interaction liquid chromatography. J. Chromatogr. A 2020, 1614, 460650. https://doi.org/10.1016/j.chroma.2019.460650 12: Pirok, B. W. J.; Molenaar, S. R. A.; van Outersterp, R. E.; Schoenmakers, P. J. J. Chromatogr. A Applicability of retention modelling in hydrophilic-interaction liquid chromatography for algorithmic optimization programs with gradient-scanning techniques. 2017, 1530, 104–111. https://doi.org/10.1016/j.chroma.2017.11.017

13: den Uijl, M. J.; Schoenmakers, P. J.; Pirok, B. W. J.; van Bommel, M. R. Recent applications of retention modelling in liquid chromatography. J. Sep. Sci. 2020, 1–27. DOI: 10.1002/jssc.202000905 14: Česla, P.; Vaňková, N.; Křenková, J.; Fischer, J. Comparison of isocratic retention models for hydrophilic interactionliquid chromatographic separation of native and fluorescently labeledoligosaccharides. J. Chromatogr. A 2016, 1438, 179–188. DOI: 10.1016/j.chroma.2016.02.032 15: Snyder, L. R.; Poppe, H. Mechanism of solute retention in liquid-solid chromatography and the role of the mobile phase in affecting separation. J. Chromatogr. A 1980, 184, 363–413.

16: Jin, G.; Guo, Z.; Zhang, F.; Xue, X.; Jin, Y.; Liang, X. Study on the retention equation in hydrophilic interaction liquid chromatography. Talanta 2008, 76, 522–527.

DOI: 10.1016/j.talanta.2008.03.042

17: Alpert, A. J. Hydrophilic-interaction chromatography for the separation of peptides, nucleic acids and other polar compounds. J. Chromatogr. A 1990, 499, 177–196.

18: Hemström, P.; Irgum, K. Hydrophilic interaction chromatography. J. Sep. Sci. 2006, 29, 1784–1821. DOI 10.1002/jssc.200600199

19: Schoenmakers, P. J.; Billiet, H. A. H.; Tijssen, R.; de Galan, L. Gradient selection in reversed-phase liquid chromatography. J. Chromatogr. A 1978, 149, 519–537.

20: Neue, U. D.; Kuss, H. Improved reversed-phase gradient retention modeling. J. Chromatogr. A 2010,

1217, 3794–3803. DOI:10.1016/j.chroma.2010.04.023

21: Skoog, D. A.; Holler, F. J.; Crouch, S. R. Principles of Instrumental Analysis, seventh edition; CENGAGE learning: Boston, 2017; pp 701–702.

22: Nikitas, P.; Pappa-Louisi, A. Retention models for isocratic and gradient elution in reversed-phase liquid chromatography. J. Chromatogr. 2009, 1216, 1737–1755. DOI:10.1016/j.chroma.2008.09.051 23: Akaike, H. IEEE Trans. Automat. Contr. A New Look at the Statistical Model Identification. 1974,