Determining gradient conditions for peptide purification in RPLC with machine-learning-based retention time predictions

Journal of Chromatography A, 1598 (2019) 92–100

Contents lists available at ScienceDirect

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

Determining gradient conditions for peptide purification in RPLC with machine-learning-based retention time predictions Jörgen Samuelsson a,∗ , Finnur Freyr Eiriksson b,c , Dennis Åsberg a , Margrét Thorsteinsdóttir b,c , Torgny Fornstedt a,∗ a

Department of Engineering and Chemical Sciences, Karlstad University, SE-651 88 Karlstad, Sweden Faculty of Pharmaceutical Sciences, University of Iceland, Hagi, Hofsvallargata 53, 107 Reykjavik, Iceland c ArcticMass, Sturlugata 8, 101 Reykjavík, Iceland b

a r t i c l e

i n f o

Article history: Received 2 February 2019 Received in revised form 20 March 2019 Accepted 21 March 2019 Available online 29 March 2019 Keywords: Peptide Preparative Purification Retention time prediction In silico determination Machine learning

a b s t r a c t A strategy for determining a suitable solvent gradient in silico in preparative peptide separations is presented. The strategy utilizes a machine-learning–based method, called ELUDE, for peptide retention time predictions based on the amino acid sequences of the peptides. A suitable gradient is calculated according to linear solvent strength theory by predicting the retention times of the peptides being purified at three different gradient slopes. The advantage of this strategy is that fewer experiments are needed to develop a purification method, making it useful for labs conducting many separations but with limited resources for method development. The preparative separation of met-enkephalin and leu-enkephalin was used as model solutes on two stationary phases: XBridge C18 and CSH C18. The ELUDE algorithm contains a support vector regression and is pre-trained, meaning that only 10–50 peptides are needed to calibrate a model for a certain stationary phase and gradient. The calibration is done once and the model can then be used for new peptides similar in size to those in the calibration set. We found that the accuracy of the retention time predictions is good enough to usefully estimate a suitable gradient and that it was possible to compare the selectivity on different stationary phases in silico. The absolute relative errors in retention time for the predicted gradients were 4.2% and 3.7% for met-enkephalin and leu-enkephalin, respectively, on the XBridge C18 column and 2.0% and 2.8% on the CSH C18 column. The predicted retention times were also used as initial values for adsorption isotherm parameter determination, facilitating the numerical calculation of overloaded elution profiles. Changing the trifluoroacetic acid (TFA) concentration from 0.05% to 0.15% in the eluent did not seriously affect the error in the retention time predictions for the XBridge C18 column, an increase of 1.0 min (in retention factor, 1.3). For the CSH C18 column the error was, on average, 2.6 times larger. This indicates that the model needs to be recalibrated when changing the TFA concentration for the CSH column. Studying possible scale-up complications from UHPLC to HPLC such as pressure, viscous heating (i.e., temperature gradients), and stationary-phase properties (e.g., packing heterogeneity and surface chemistry) revealed that all these factors were minor to negligible. The pressure effect had the largest effect on the retention, but increased retention by only 3%. In the presented case, method development can therefore proceed using UHPLC and then be robustly transferred to HPLC. © 2019 Elsevier B.V. All rights reserved.

1. Introduction Therapeutic peptides are an important class of biopharmaceuticals that have received several recent approvals from the US

∗ Corresponding authors. E-mail addresses: [email protected] (J. Samuelsson), [email protected] (T. Fornstedt). https://doi.org/10.1016/j.chroma.2019.03.043 0021-9673/© 2019 Elsevier B.V. All rights reserved.

Food and Drug Administration [1]. Separation using reversed-phase liquid chromatography (RPLC) is a main technique for biopharmaceutical purification in both drug development and production [2,3]. However, the purification of peptides is more complicated and time consuming than that of small molecules since they are more affected by mobile- and stationary-phase conditions [4,5], leading to lower throughput [6]. Efficient method development is therefore essential for achieving adequate throughput, since the

J. Samuelsson et al. / J. Chromatogr. A 1598 (2019) 92–100

resources available for method development are often limited in the early stages of pharmaceutical projects [7]. Retention time prediction models for peptides in RPLC are well established and routinely employed in analytical applications such as shotgun proteomics to reduce false positives in mass spectrometry detection and to design targeted proteomics experiments [8]. However, the theory for preparative chromatography is much more complex than for analytical chromatography. In analytical chromatography the goal is information whereas in preparative chromatography the goal is to isolate large amounts of the desired component(s) from a complex sample mixture. Because of the limited surface capacity of the stationary phase, the column will be operated under overloaded conditions, and a further increased sample load results in a lower fraction adsorbed. Thus, non-linear conditions prevail and the eluted bands of the components at the column outlet are strongly distorted and asymmetrical. Moreover, different types of components compete with each other for the same surface, an effect which further complicates proper predictions of preparative operational conditions when starting from scaling up from analytical conditions. Therefore, this study investigates the use of expanding analytical retention time prediction models based on machine-learning, for the in silico determination of suitable solvent gradients for preparative peptide separations. Such in silico screening before the experimental tuning would considerably reduce the number of experiments in preparative method development, since only the most promising in silico results would be investigated experimentally. This would make the method development of peptide purifications more efficient and free up more resources for method optimization, resulting in methods with higher throughput and robustness. We propose here a strategy in which, for each column, three models corresponding to three different gradient slopes are constructed using a small calibration set of peptides with experimentally determined retention times. The construction of these models (or “calibration”) is a one-time task and is not repeated for new purification cases. The retention times of new peptides with known amino acid sequences can now be predicted for these three gradients on the specific column. These data are used to calculate a suitable gradient and as initial values for numerically modelling the overloaded elution profiles. Several stationary phases can then be compared, and the best gradient–stationary phase combination can be tested experimentally. This study is intended as a proof-of-concept of the in silico screening strategy using machine-learning (ML) to predict also preparative nonlinear conditions; the separation of leu-enkephalin and met-enkephalin, which differ in one amino acid, is used as model. These peptides were selected since they represent a difficult enough separation problem and since they are available in the required degree of purity for proper preparative modeling. In addition, the robustness of the retention time predictions to the trifluoroacetic acid (TFA) in the mobile phase and the conversion of the method from UHPLC to preparative HPLC are evaluated. The reasons for this are that TFA is the most common mobile phase additive in RPLC for peptides and that it is common to conduct the method development using UHPLC and then to convert the final method to preparative HPLC for the actual purification.

2. Theory 2.1. Retention time predictions based on machine learning Various approaches have been suggested for predicting the retention time of a peptide based on its amino acid sequence; these

93

can roughly be divided into (i) index-based (ii) modeling-based and (iii) machine-learning (ML)-based methods [9]. An index-based method estimates the effect of each amino acid in a sequence and gives it a retention coefficient based on the multi-linear regression of a large training set of peptides with known retention times [10,11]. A modeling-based method predicts retention times based on the physicochemical properties of the peptide and its interactions with the stationary phase, i.e., it uses a theoretical model together with information from the amino acid sequence [12,13]. An ML-based method uses a training set of peptides to estimate the parameters of a predefined mathematical model with an artificial neural network [14] or with a support vector regression (SVR) [15]. For the application studied here, the two important traits of the retention time prediction method are high retention time accuracy and a small training set. A small training set is needed since a model needs to be trained for three gradient slopes for each column. Of the three approaches outlined above, the ML-based method best fulfils these traits. ML is a field of computer science in which a program learns from data without being explicitly programmed [16]. The advantage of ML is that complex data can be processed without the need for a detailed understanding of the underlying phenomena [14]. The training is generally an iterative process in which the model predictions are progressively improved in a least squares sense. ML is used for several prediction tasks, such as classification, clustering, and, as in the present case, regression analysis and correlating the peptide sequence with the peptide’s retention time. Here, several peptides with experimentally determined retention times for particular separation conditions are used as input data for the training. The trained model can then be used to predict the retention times from the peptide sequence. Important downsides of using a purely empirical ML approach are that (i) very large data sets are required for the dynamically training of the models, often several thousands of different peptides, to obtain an adequate prediction model and that (ii) a new model must be trained if the separation conditions are changed [9,14]. Allowing training of the models using smaller peptide data sets [9]. An SVR correlates a set of dependent variables (here, retention time) to a set of independent variables (here, vectors of peptide descriptors derived from sequence data). The support vectors are based on the training data and are used to train the model to a certain tolerance [15]. An SVR approach called ELUDE is suitable for our objective, since it selects and calibrates a model from a library of pre-trained predictors, meaning that only a smaller set of peptides is required for a certain experimental condition [17]. It has 60 different predefined peptide descriptors, also called features, such as the 34 physical and chemical properties descriptors derived from the Kyte–Doolittle hydrophobicity scale [18], peptide lengths, occurrence of each amino acid, and number of polar amino acids [17]. It was found using a set of only 100 calibration peptides could give adequate predictions, however, more than 1000 unique peptides gives the best accuracy [17]. The pre-trained ELUDE model used here uses 1995 and 1985 unique peptides for training and testing, respectively [17]. We will use trypsinized plasma samples to generate the dataset for calibrating models for each experimental condition. ELUDE was selected because it is easy to use, has accuracy on par with or better than other methods [17], and requires a relatively small calibration dataset. Moruz et al. discussed the accuracy of the retention time predictions of ELUDE by statistically validating the model using 14 datasets each containing over 2000 unique peptides [17]. In that study, the authors found that the Pearson’s correlation r between observed and predicted retention times was 0.92–0.97.

94

J. Samuelsson et al. / J. Chromatogr. A 1598 (2019) 92–100

2.2. Retention and adsorption modelling The retention of a solute is dependent on the fraction of modifier in the eluent. Using linear solvent strength (LSS) theory, we could define this as: ln k = ln kw − S,

(1)

where k is the retention factor, kw the apparent solute retention factor in the eluent without modifier, S the slope constant, and  the fraction of modifier in the eluent. Using LSS theory and assuming a linear gradient, the retention time of a solute in gradient elution could be expressed as [19,20]: tR =

t0 1n (Gk0 + 1) + t0 , G

(2)

3.2. Instrumentation

with G = S

dithiothreitol, iodoacetamide, and hydrochloric acid were purchased from Sigma Aldrich (St. Louis, MO, USA) and trypsin was obtained from Promega (Madison, WI, USA). Uracil (98%), used as a void-volume marker, and TFA (≥99.0%), used as an ion-pairing reagent, were obtained from Sigma Aldrich. The mobile phase consisted of gradient-grade acetonitrile (MeCN) from VWR (Radnor, PA, USA) for the overloaded experiments, LC MS ultra MeCN from Sigma Aldrich for the analytical experiments, and water with a conductivity of 18.2 M cm from a Milli-Q Plus 185 water purification system from Merck Millipore (Darmstadt, Germany). Three types of columns were used in this study: Acquity UPLC BEH C18 (100 × 2.1 mm, 1.7 ␮m), XBridge BEH C18 (100 mm × 2.1 mm, 5 ␮m), and XSelect CSH C18 (100 × 2.1 mm, 5 ␮m), all from Waters (Milford, MA, USA).

t0 , tg

(3)

where k0 is the retention factor in the starting eluent composition, t0 the hold-up time, tg the gradient time, G the gradient steepness factor, and  the change in  during the gradient. So far only analytical retention time has been described; for overloaded elution profiles, as in preparative chromatography, we need to use an adsorption isotherm to describe the adsorption of peptide in gradient elution. In this study, the adsorption isotherm model does not need to give an actual physicochemical description of the adsorption, since we are interested only in simulating band profiles. One simple model is the Langmuir adsorption isotherm, which could be written for the components as [21]: q (C, ) =

aw e−Sa  C

(4)

1 + bw e−Sb  C

where bw is related to the association equilibrium constant using eluent without modifier, q and C are the stationary- and mobilephase concentrations, respectively, and aw is often known as the Henry constant and is identical to kw /F, where F = (1 − εt )/εt is the phase ratio and εt is the total porosity of the column. The Langmuir equation could be expanded into a two-component competitive adsorption isotherm, which for the ith compound, could be expressed as [22,23]: qi (C1 , C2 , ) =

1 + bw,1 e

aw,i e −Sb,1 

−Sa,i 

Ci

C1 + bw,2 e−Sb,2  C2

,

(5)

where the indexes 1 and 2 refer to the two compounds under investigation. Using an adsorption model that also includes the ion-pair formation would probably describe the data more accurately [24]. However, the aim of this investigation is not to determine accurate adsorption isotherm parameters but instead to use simple models that just describe the elution profiles adequately. We have previously demonstrated that trifluoroacetic acid (TFA) adsorbs to the CSH stationary phase [24]. In a later study, we compared a model accounting for the competition with TFA with a model neglecting TFA adsorption [25]. We found that the two models predicted elution profiles with the same accuracy, so the TFA adsorption is also excluded from this study. 3. Materials and methods

The analytical experiments were conducted using a Waters Acquity UPLC system, coupled to a Waters Synapt G1 QToF/MS equipped with an electrospray ionization (ESI) probe. The UPLC system consisted of a binary solvent manager, a fixed loop autosampler with a 10-␮L injection loop, and a tunable ultraviolet detector (Waters). The instrument was equipped with a 50-␮L mobile-phase mixer and a 30-␮L PEEKsil non-deactivated sample needle. The dwell volume from the pump to the detector was estimated to be 165 ␮L and the column void volume was estimated to 109 ␮L. The overloaded experiments were conducted using an Acquity H-Class Bio System (Waters) with a quaternary solvent manager, an autosampler with a 50-␮L injection loop, a still-air column oven with a mobile-phase pre-heater, and a PDA detector with a 500nL flow-cell; the PDA detector was used throughout the study. The extra-column volume from the autosampler to the detector was 19 ␮L. The dwell volume from the pump to the detector was estimated to be 350 ␮L. Analytical peptide elution profiles were recorded at 229 nm, overloaded peptide elution profiles at 290 nm, and uracil profiles at 220 nm. 3.3. Sample preparation The samples were prepared according to Gilar et al. [10]. The human serum sample (from one person) was digested as follows: 5 ␮L of human serum was washed twice with 500 ␮L of 0.1% RapiGest detergent (Waters) in 50 mM ammonium bicarbonate using an Amicon Ultra 10 K cut-off spin column (Millipore, Billerica, MA, USA). The protein was then denatured (80 ◦ C for 15 min) and 5.0 ␮L of 100 mM dithiothreitol was added to reduce the disulfide bonds (60 ◦ C for 15 min). The proteins were then derivatized in the dark with the addition of 5 ␮L of 200 mM iodoacetamide (room temperature for 30 min). Tryptic digestion was accomplished by adding 2 ␮L of 1 mg mL–1 trypsin solution in 50 mM ammonium bicarbonate. Digestion was carried out at 37 ◦ C overnight. The mixture was then acidified with 2 ␮L of hydrochloric acid (37 ◦ C for 20 min). Three samples were pooled; 100 ␮L of the pooled sample was then diluted with 150 ␮L of water and spiked with 50 ␮L of 1 ␮g mL–1 of the reference peptides (i.e., LeuEnk and MetEnk). The final sample volume was 350 ␮L. To study overloaded profiles using the inverse method, three different samples were prepared: 20 mg mL–1 of LeuEnk, 20 mg mL–1 of MetEnk, and a sample comprising 10 mg mL–1 of each peptide dissolved in 10/90 (v/v) MeCN/Water with 0.1% TFA.

3.1. Chemicals and materials 3.4. Procedure Two peptides were investigated: leu-enkephalin (LeuEnk) and met-enkephalin (MetEnk). LeuEnk was obtained from Bachem (Bubendorf, Switzerland) and MetEnk from Alfa Aesar (Karlsruhe, Germany). For the sample preparation, ammonium bicarbonate,

The UPLC coupled to the Synapt G1 QToF/MS was used for the analytical experiments. The data were collected in positive ionization mode, with MSE ranging from 50 to 1950 m/z. The capillary

J. Samuelsson et al. / J. Chromatogr. A 1598 (2019) 92–100

voltage was set to 0.8 kV, cone voltage to 24.0 V, and extraction cone voltage to 5.0 V. The source temperature was set to 120 ◦ C, desolvation temperature to 400 ◦ C, desolvation gas flow to 700 L h–1 , and cone gas flow to 50 L h–1 . The scan time was 0.4 s and the inter-scan time 0.02 s. LeuEnk was used for lock mass correction, and a 0.5-s scan was acquired every 30 s. The peptides were selected using BiopharmaLynx 1.3.5 (Waters) and the software was used to process the peptide maps and extract peptide retention time data from LC–MS chromatograms. All post-modified peptides were excluded from the list of identified peptides. If not otherwise stated, all separations were conducted at a flow rate of 0.25 mL min–1 and a column temperature of 25 ◦ C. All retention times in this study were determined at the maximal peak heights. To obtain calibration data for the elution prediction, 5-␮L injections of plasma sample were analyzed using three different gradient slopes for three different TFA concentrations in the eluent (i.e., 0.05, 0.1, and 0.15% TFA) on the XBridge and CSH columns. The responses were recorded using MS. The gradient separations were conducted first with a 5-min isocratic hold at 2% (v/v) MeCN/water followed by a 25-, 45-, or 65-min gradient to 40% (v/v) MeCN/water. After the gradient, the columns were washed with 90% (v/v) MeCN/water for 2 min followed by an 8-min re-equilibration using 2% MeCN before the next injection. To investigate the effects of pressure and temperature, samples of the pooled plasma were analyzed on the Acquity UPLC BEH C18 column using 0.1% TFA at flow rates of 0.1, 0.25, and 0.50 mL min–1 . The column temperature was controlled using a conventional stillair column oven as well as a water bath. The gradients used during the separation were scaled to the volumes so that the flow rate would not affect the effective gradient slope: First, a 1.25-mL isocratic hold at 2% (v/v) MeCN/water was followed by an 11.25-mL gradient to 40% (v/v) MeCN/water. After the gradient, the columns were washed with 0.5 mL of 90% (v/v) MeCN/water followed by a 2-mL re-equilibration using 2% MeCN before the next injection. The injection volume was 2 ␮L. The overloaded study was conducted on the CSH and XBridge columns. In all runs, 0.1% TFA was added to the eluents; 5-, 10, 20-, 30-, and 45-␮L samples of single peptides and mixtures were injected. The gradient started with a 5-min isocratic hold at 10% (v/v) MeCN/water followed by a 15-min gradient to 25% (v/v) MeCN/water for the CSH column and 30% (v/v) MeCN/water for the XBridge column. After the gradient, the column was washed with 50% (v/v) MeCN/water for 3 min followed by a 2-min reequilibration before the next injection. 3.5. Calculations 3.5.1. Retention time predictions with ELUDE Retention time predictions were made with the ELUDE software described by Moruz et al. [17]. ELUDE is freely available under the Apache License and can be run locally or via a website [17]. In short, we calibrated a pre-trained model for three different gradient slopes for each of the two different columns giving a total of six models. As a calibration set, the sequences and retention times of approximately 50 peptides (for raw data, see Tables S1 and S2 in Electronic Supplementary Material) with numbers of amino acids similar to those of the peptides of interest in the screening were employed for each experimental condition. Note that the peptides to be separated were not included in the calibration set. 3.5.2. Gradient optimization By fitting the retention times predicted for the peptides at several different gradient slopes to Eq. (2), S and k0 can be estimated for each peptide. These constants can be used to estimate the optimal

95

gradient separation. In this study, we used ELUDE to estimate the retention times of LeuEnk and MetEnk at three different gradient slopes (see section 3.4). For the gradient optimization, the starting and ending fractions of MeCN in the eluent were optimized to give a baseline separation of the two peptides while keeping the gradient run to 15 min. The starting conditions were selected so that the retention factor would be very high, this being achieved at 10% (v/v) MeCN in the eluent for both columns. To achieve acceptable productivity, the later-eluting peptide should elute later than halfway through the gradient run but before 85% of the gradient program run has been completed. 3.5.3. Calculations of overloaded elution profiles To calculate the elution profiles, we need to use a column model. In this study, the equilibrium-dispersive model was used [21,22]. The orthogonal collocation on finite elements method [26] was used to discretize the spatial derivatives of the equilibriumdispersive model, and the Adams–Moulton method implemented in the variable-coefficient ordinary differential equation solver (VODE) procedure [27] was used to solve the system of ordinary differential equations. The gradient version of the Langmuir adsorption isotherm was used, i.e., Eq. (4). The adsorption isotherm parameters were estimated using the inverse method by means of a modified least squares Marquardt method [28]. The nonlinear UV-detector response (R) for the overloaded elution profiles was converted to concentration (C) by fitting the detector responses of six different column loads (i.e., 5-, 10-, 15, 20-, 30-, and 45-␮L injections) for each condition to Eq. (6) so that the injected mass equals the eluted mass [29]: C = m1 log

m − R 2 m2

+ m3 R,

(6)

where m1 , m2 , and m3 are constants used in the calibration curve. Observe that we used only single-component data in the calibration, in other words only LeuEnk or only MetEnk data. As initial guesses for the inverse problem, we used the predicted linear data from ELUDE (see section 3.5.2 above) to obtain acceptable initial guesses for aw and Sa . Sb and bw cannot be estimated using retention time prediction methods and were set to 10 and 100, respectively. The column efficiency was set to 1000. As experiments for the model calibrations, three 10-, 20-, and 45␮L injections of 20 mg mL–1 LeuEnk or MetEnk were used on each column (see section 3.4 for more experimental details). The inverse method gave the following adsorption isotherm parameters on the XBridge column: aw = 9018, bw = 22.26 L g–1 , Sa = 38.82, and Sb = 19.57 for MetEnk, and aw = 28885, bw = 7.623 L g–1 , Sa = 35.94, and Sb = 7.342 for LeuEnk, both at a flow rate of 0.25 mL min–1 and a column temperature of 25 ◦ C. Observe that these adsorption parameters were used only to estimate elution profiles and not for the thermodynamic characterization of the separation system. Extrapolating adsorption parameters can often result in unrealistic adsorption constants that still could describe the elution profiles well, as in this case [30].

4. Results and discussion Section 4.1 presents the proposed strategy for determining the solvent gradient in silico for the preparative separation of peptides, which is tested on two different columns using 0.1% TFA in the eluent. Section 4.2 investigates the effects of using different TFA contents in the eluent on retention and selectivity. Section 4.3 investigates possible complications with the conversion from UHPLC to HPLC due to temperature gradients and pressure differences.

96

J. Samuelsson et al. / J. Chromatogr. A 1598 (2019) 92–100

Fig. 1. Schematic of the steps in the new in silico approach (at left) compared with the conventional experimental approach (at right).

4.1. In silico screening A conventional experimental approach to finding a suitable solvent gradient is compared with the suggested in silico strategy in Fig. 1. Normally, two or three broad gradients are screened experimentally to find suitable gradient settings, while the proposed in silico approach instead uses a model to predict retention times and LSS theory to estimate a suitable gradient. First, we calibrate an already trained SVR model, ELUDE [17], using peptides from a tryptic-digested plasma sample (see section 3.5.1 for details). As input in the calibration phase, the peptide sequence and retention are used, and the calibration data are presented in Tables S1 and S2 in the Electronic Supplementary Material. The calibrated model predicts retention times just from the peptide sequence. To model the gradient dependency, retention time models are calibrated for three gradient slopes for each of the stationary phases (see section 3.5.2 for more details). Second, the retention of the target peptides is estimated for the three different gradient slopes and these retention times are used to estimate S and kw in Eq. (1). Third, a suitable gradient for conducting purification is numerically determined (see section 3.5.2 for details). We found that the average absolute relative errors in retention time for the predicted gradient were 4.2% and 3.7% for MetEnk and LeuEnk, respectively, on the XBridge C18 column and 2.0% and 2.8% for the CSH C18 column for the three different gradients used to calibrate ELUDE. Once the gradient is determined, actual experiments are needed to optimize the separation, since the retention prediction model can only predict the retention time and not the overloaded elution profiles; note the crossover from the in silico part to the experimental part in Fig. 1. In this case, the separations of LeuEnk and MetEnk were investigated on two different columns, XBridge C18 and CSH C18. These are the target peptides and were not part of the calibration set for the prediction model. The starting and ending fractions of MeCN for the gradient are taken as adjustable parameters. The flow rate was set to 0.25 mL min–1 and the gradient time to 15 min. The starting condition was selected so that the isocratic retention factor would be very high, this being achieved with 10% (v/v) MeCN/water for both columns. To achieve acceptable productivity, the late-eluting peptide should be eluted later than halfway through the gradient run but before 85% of the gradient program has been completed. Fig. 2 presents the predicted retention times of different ending fractions of MeCN in the gradient separation. Fig. 2aI and bI show that an ending fraction of approximately 30% (v/v) MeCN/water is acceptable for the XBridge column and that 25% (v/v) MeCN/water

is acceptable for the CSH column. For these conditions, the predicted retention time for the later eluting compound is 12.0 min and 12.2 min on the XBridge and CSH column, respectively. To verify the predicted gradient separation, overloaded experiments were conducted using a mixture of LeuEnk and MetEnk on both columns. The overloaded separation is presented in Fig. 2aII for the XBridge column and Fig. 2bII for the CSH column. The dashed lines in Fig. 2aII and bII are the elution times predicted using ELUDE. Good agreement between predicted and experimental retention times was found on the XBridge column for both peptides as well as for MetEnk (first eluting) on the CSH column. On the CSH column, LeuEnk (last eluting) was slightly underestimated, indicating that the separation was slightly better than predicted. Overloaded elution profiles are modelled using a column model and an adsorption isotherm model; in this case, we used the equilibrium dispersive column model and the Langmuir adsorption isotherm model, Eq. (4). To determine the adsorption isotherm parameters, we iteratively minimize the difference between experimental and calculated elution profiles, which is called the inverse method [31]. We have previously investigated several ways to determine the adsorption isotherm using only gradient condition data [21–23]. In these studies, we demonstrated that adsorption isotherms could be estimated using the inverse method just from overloaded elution zones under gradient conditions. However, even the rather simple Langmuir adsorption isotherm, for example, Eq. (4), results in equation systems that take several hours to solve numerically. To speed up the process and make it more feasible, good initial values for the isotherm parameters are needed. Here, we use the data from the retention time predictions to calculate initial values of the linear terms aw and Sa . However, the nonlinear terms Sb and bw cannot be estimated using retention time prediction methods and were, after some testing, set to 10 and 100, respectively. All adsorption isotherm parameters were treated as free variables in the inverse method (see section 3.5.3 for more details). Fig. 3a and b present experimental and calculated elution profiles for MetEnk and LeuEnk on the XBridge column. The simple Langmuir adsorption isotherm fits the experimental data very well. However, to describe a two-component separation, competitive adsorption isotherms are needed, as in Eq. (5). Fig. 3c presents elution profiles of a mixture of LeuEnk and MetEnk, and the singlecomponent adsorption isotherm parameters determined in Fig. 3a and b are used in Eq. (5). In this case, the single-component isotherm parameters describe the elution profiles of the twocomponent separation rather well. Similar results were observed for the CSH column (see Electronic Supplementary Material, Fig.

J. Samuelsson et al. / J. Chromatogr. A 1598 (2019) 92–100

97

Fig. 2. (aI ) Predicted retention times for the gradient elution of the peptides LeuEnk and MetEnk for different gradient conditions and ending fractions of MeCN in the eluent and using the XBridge column. (aII ) Corresponding elution profiles for 10-, 20-, and 45-␮L injections of a mixture of 10 mg mL–1 MetEnk (first eluting) and 10 mg mL–1 LeuEnk. (bI and bII ) Corresponding results for the CSH column. The vertical dashed lines in (aII ) and (bII ) are the predicted analytical retention times from (aI ) and (bI ), whereas the dotted lines are estimated from the experimental retention of LeuEnk and MetEnk (see Section 3.5 for more details).

Fig. 3. Comparison between calculated and experimental elution profiles from 10-, 20-, and 45-␮L injections of (a) MetEnk (20 mg mL–1 ), (b) LeuEnk (20 mg mL–1 ), and (c) a mixture of MetEnk and LeuEnk (10 mg mL–1 each) using the XBridge column. In the calculations, the efficiencies were set to 2000 and 250 subdomains were used.

S1). A better model fit could be achieved using elution profiles of a mixture of LeuEnk and MetEnk instead of expanding singlecomponent adsorption isotherm data; however, this approach is more complicated because more adsorption isotherm parameters need to be determined and a way needs to be found of translating the response to concentration in the overlapping elution zones. The selectivity for LeuEnk and MetEnk were roughly the same on both the XBridge and CSH columns according to the retention time predictions and experimental results (see Fig. 2). Selectivity is a very important factor to optimize in separations. Many times during the preparative screening phase several columns are tested to find suitable selectivity for the separation under investigation [32]. To investigate whether the selectivity is more or less the same on both columns for a larger set of peptides, the retentions of 139 tryptic-digested plasma peptides on the CSH and XBridge columns using 0.1% TFA in the eluent are compared (see Fig. 4). The XBridge C18 column is a hybrid C18 column and the XSelect CSH C18 column is a charged-surface hybrid C18 bonded-phase column in which the CSH particles are prepared by derivatizing bare bridged ethylene hybrid (BEH) particles with a weakly basic ionizable silane before the ligand (C18) bonding and the end-capping steps [33,34]. As can be seen from Fig. 4, the peptides have slightly smaller retention volumes on the CSH column, averaging 0.46 mL (RSD% 29.97%), than on the XBridge column, probably because the peptides are slightly repelled by the positive charges on the CSH column surface. For the most part, there are no major differences in selectivity between the columns, just a constant retention shift. Only a few

98

J. Samuelsson et al. / J. Chromatogr. A 1598 (2019) 92–100

Fig. 4. Difference between peptide gradient retention volumes on XBridge and CSH columns. The dashed line is a reference line indicating identical retention volumes; flow rate 0.25 mL min–1 , gradient time 45 min, and 0.1% TFA in the eluent. For raw data, see Table S3 in the Electronic Supplementary Material.

Fig. 5. Peptide retention dependence on the TFA level in the mobile phase for five peptides using the XBridge and CSH columns. The 0.1% TFA level is used as the reference for both types of columns. For raw data, see Table S4 for CSH and Table S5 for XBridge in the Electronic Supplementary Material.

peptides display a difference in selectivity, though analyzing these peptides reveals no clear pattern. This suggests that the retention time predictions from ELUDE give an indication of the selectivity one could expect of the two tested columns. Let us return to the normal, experimental approach and compare the accuracy achieved if S and kw in Eq. (1) are determined experimentally with three gradient experiments (same gradient settings as used for ELUDE). Using these retention data, the retention time for the preparative gradient was calculated (see vertical dotted lines in Fig. 2aII and bII ). Better agreement with the overloaded elution profiles is observed, but the difference is quite small, and we argue that the results of the in silico approach are good enough for a suitable preparative gradient and for selecting between the two stationary phases. We argue that these results, although limited to two separation cases, serve as a proof-of-concept to show that the in silico strategy is possible and worth further investigation.

necessary if the TFA concentration is changed. This is especially true if a smaller fraction of TFA is to be used in the eluent. Although TFA reduces the MS signal, preparative HPLC is commonly run with UV detection, so TFA concentrations above 0.1% are not unrealistic.

4.2. Influence of TFA on retention times The TFA concentration affects the retention time of peptides, especially at low concentrations, since the dependence follows a saturation-type relationship [35,36], and we recently demonstrated that the shape of overloaded peaks can be substantially improved by increasing the TFA concentration [24]. We therefore investigated the error expected from retention time predictions when the TFA concentration differs between the sample and the training set. This difference in TFA could be unintentional (e.g., due to experimental error) or intentional (e.g., to improve the peak shape). In Fig. 5, the retention times of approximately 50 peptides are compared for 0.05%, 0.1%, and 0.15% TFA on the XBridge and CSH columns. The retention times increases with increased TFA concentration, which is a well-established trend [35,37]. A larger difference between 0.05% and 0.1% TFA versus between 0.1% and 0.15% TFA is indicated, agreeing with a saturation-type relationship. For the XBridge column, the retention times increase approximately by 1.0 min, corresponding to 1.3 in retention factor, resulting in small errors if the TFA concentration is changed from 0.05% to 0.15% after the peptide prediction step. The CSH column gives slightly larger retention time differences when the TFA content is changed; on average 2.6 times larger compared to the XBridge column (for an increase in TFA from 0.05% to 0.15%). Therefore, recalibration of the peptide prediction model might be

4.3. Scale-up from UHPLC to preparative HPLC UHPLC offers much shorter run times than does HPLC and is therefore often preferred in method development [38]. However, preparative HPLC is always conducted at a much lower pressure than is UHPLC. Converting a method from UHPLC to HPLC is not always straightforward, since complications can arise from stationary-phase differences [39,40], temperature gradients due to viscous heating [41,42], and thermodynamic effects due to the different pressures [41,43]. Here, we evaluate each of these factors for the method conversion from UHPLC to HPLC. To investigate the difference in packing materials, an XBridge column packed with 5-␮m particles was compared with an Acquity UPLC BEH C18 column of the same length but packed with 1.7-␮m particles. Properties such as surface coverage, silanol activity, and stationary-phase porosity may differ between particle sizes [41,42]. The same peptide separation was conducted on both columns using a low flow rate to avoid pressure and temperature effects. We found very little difference between the two phases (see Fig. S1 in Electronic Supplementary Material). Viscous heating arises when the mobile phase percolates through the column at a high pressure and high flow rate, which is the case in UHPLC [44]. Keeping the column in a conventional still-air column oven will result in an axial temperature gradient in which the outlet is warmer than the inlet [45]. Fig. 6 presents the retentions of several peptides separated using flow rates of 0.1 mL min–1 and 0.5 mL min–1 and with the temperature controlled using a conventional still-air column oven or a water bath. When the column was operated at a pressure drop of 900 bar, we noted a longitudinal temperature gradient of 8 ◦ C (see Fig. S2 in Electronic Supplementary Material). The observed difference in retention between systems temperature controlled in water (without an axial temperature gradient) and in still air (with an axial temperature gradient) is due to the temperature gradient. In this case, when we compare the retention volume with and without an axial temperature gradient, there is almost no difference (see Fig. 6). Potential shifts in retention due to viscous heating are therefore negligible in the experimental system under investigation.

J. Samuelsson et al. / J. Chromatogr. A 1598 (2019) 92–100

Fig. 6. Retention volumes of peptides separated using different flow rates on the Acquity UPLC BEH C18 column with eluent containing 0.1% TFA. The column is temperature regulated using a water bath or still air (see symbols). The gradient volume was 11.25 mL (see section 3.4 for more details). For raw data, see Table S6 in the Electronic Supplementary material.

To study the effect of pressure while minimizing viscous heating, the column was immersed in a water bath, since the heat transfer from the surface to the surroundings is then very efficient and the internal column temperature can be maintained within ± 1 ◦ C [42]. Inspecting the two separations conducted in the water bath (see Fig. 6) reveals only a small increase (approximately 3%) in retention volume when comparing separations conducted at 0.1 mL min–1 ˜ bar) and 0.5 mL min–1 (900 ˜ bar). An increase in retention of (100 this magnitude will have little impact on the scale-up. Fallas et al. studied the pressure effect on the retention of small solutes under isocratic conditions and found differences of up to 50% [46]. In that study, the authors speculated that the pressure effects would probably be smaller in gradient elution because small changes in the content of organic modifier produce much larger retention changes than do pressure changes. To strengthen this argument, we can use Eq. (2) and assume that the pressure will have the most influence on the retention factor in the starting eluent composition (k0 ) and only a slight influence on the gradient steepness factor (G). In this case, the XBridge column with MetEnk had G and k0 values of 0.22 and 469, respectively, while the column with LeuEnk had G and k0 values of 0.21 and 874 (45-min gradient run). Using these values and if the isocratic elution were to increase by 25% due to pressure, the retention would then increase by 4.6% and 4.1% for MetEnk and LeuEnk, respectively. Using a steeper gradient (here, 15 min), the gradient steepness becomes 0.65 and 0.64 for MetEnk and LeuEnk and the estimated retention shift decreases to 3.5% and 3.1% for MetEnk and LeuEnk, respectively. This indicates that the pressure effects on a solute decrease with increasing gradient steepness and are more severe in systems operating under isocratic conditions. To recap, the presented results indicate that scale-up from UHPLC to HPLC using gradient elution is robust and that no difficulties should be expected from pressure and temperature effects. 5. Conclusion Preparative liquid chromatography is a purification method of rising importance in chemical and pharmaceutical industry especially for biological molecules such as peptides, nucleotides, mRNA and therapeutic proteins, instead of the small molecules used in current and yesterday’s bestsellers. However, the operational conditions in preparative peptide chromatography are much more complex than for analytical chro-

99

matography; nonlinear conditions prevail resulting in complex adsorption behavior and overloaded band profiles often having a sharp front and diffuse rear. Here, we suggested a machine-learning based - in silico approach - to determine proper gradient conditions in preparative chromatography utilizing selected tools from analytical linear chromatography. An machine-learning algorithm with support vector regression called ELUDE [17] was used to predict the retention times of peptides based on their amino acid sequences. A pre-trained model was used, and it was calibrated with as few as approximately 50 peptides with known retention times for three gradient slopes and two stationary phases, giving us the possibility of predicting retention times for new peptides under these six conditions. A suitable gradient for each column was then calculated using LSS theory to separate peptides of interest. The advantage of this strategy is that no prior experimental screening is necessary once one-time calibration of the stationary phases has been done. Knowledge of the most promising stationary phases and gradients can be obtained in silico, reducing the number of required experiments. The workload of calibrating the ELUDE models and the accuracy of the retention time predictions were found to be acceptable and the calculation of a suitable gradient from the LSS theory to be straightforward. We demonstrated how the parameters calculated from LSS theory could be used to obtain initial values for some of the adsorption isotherm parameters needed to undertake numerical calculations, using the inverse solver, of the overloaded elution profiles of the peptides. The agreement between experimental and calculated elution profiles was acceptable for further optimization of the purification method. We also investigated how the retention times are affected by TFA concentration in the eluent in different phase systems. The goal was to investigate whether one needs to recalibrate the prediction model when the TFA concentration in the eluent changes. We found that the CSH column gave much larger retention time shifts, on average 2.6 times larger, than did the XBridge column when changing the TFA concentration in the eluent from 0.05% to 0.15%, so recalibration will likely be necessary with the CSH column. Method conversion from UHPLC to HPLC with peptides in gradient elution was found to be robust and no complications were observed from viscous heating, elevated pressure, or stationary phase differences. This study presents proof-of-concept of the strategy of using retention time prediction models as an in silico screening step in developing RPLC purification methods for peptides. The strategy would be most suitable for chromatographers in laboratories that conduct many preparative peptide separations under similar separation conditions but have limited time for method development, for example, during early drug discovery in the pharmaceutical industry.

Acknowledgements We are grateful to Martin Gilar and Fabrice Gritti at Waters (Milford, MA, USA) for the kind gift of all columns used in this study and for valuable scientific discussions. This work was supported by the Swedish Knowledge Foundation via the KKS SYNERGY 2016 project “BIO-QC: Quality Control and Purification of New Biological Drugs” (grant number 20170059); the Swedish Research Council (VR) via the project “Fundamental Studies on Molecular Interactions aimed at Preparative Separations and Biospecific Measurements” (grant number 2015-04627); and the ÅForsk Foundation via the project “Quality control of next generation biological based medicines” (grant number 17/500).

100

J. Samuelsson et al. / J. Chromatogr. A 1598 (2019) 92–100

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. 03.043. References [1] C. Morrison, Fresh from the biotech pipeline—2015, Nat. Biotechnol. 34 (2016) 129–132, http://dx.doi.org/10.1038/nbt.3472. [2] M.-I. Aguilar, HPLC of Peptides and Proteins—Methods and Protocols, Humana Press, Totowa, New Jersey, 2004. [3] J.M. Conlon, Purification of naturally occurring peptides by reversed-phase HPLC, Nat. Protocols 2 (2007) 191–197, http://dx.doi.org/10.1038/nprot.2006. 437. [4] H.J. Issaq, K.C. Chan, J. Blonder, X. Ye, T.D. Veenstra, Separation, detection and quantitation of peptides by liquid chromatography and capillary electrochromatography, J. Chromatogr. A 1216 (2009) 1825–1837, http://dx. doi.org/10.1016/j.chroma.2008.12.052. [5] S. Fekete, J.-L. Veuthey, D. Guillarme, New trends in reversed-phase liquid chromatographic separations of therapeutic peptides and proteins: theory and applications, J. Pharm. Biomed. Anal. 69 (2012) 9–27, http://dx.doi.org/10. 1016/j.jpba.2012.03.024. [6] A.T. Hanke, M. Ottens, Purifying biopharmaceuticals: knowledge-based chromatographic process development, Trends Biotechnol. 32 (2014) 210–220, http://dx.doi.org/10.1016/j.tibtech.2014.02.001. [7] H. Schmidt-Traub (Ed.), Preparative Chromatography: Of Fine Chemicals and Pharmaceutical Agents, Wiley-VCH, 2005. [8] V.I. Babushok, I.G. Zenkevich, Retention characteristics of peptides in RP-LC: peptide retention prediction, Chromatographia 72 (2010) 781–797, http://dx. doi.org/10.1365/s10337-010-1721-8. [9] L. Moruz, L. Käll, Peptide retention time prediction, Mass Spectrom. Rev. 36 (2017) 615–623, http://dx.doi.org/10.1002/mas.21488. [10] M. Gilar, A. Jaworski, P. Olivova, J.C. Gebler, Peptide retention prediction applied to proteomic data analysis, Rapid Commun. Mass Spectrom. 21 (2007) 2813–2821, http://dx.doi.org/10.1002/rcm.3150. [11] O.V. Krokhin, R. Craig, V. Spicer, W. Ens, K.G. Standing, R.C. Beavis, J.A. Wilkins, An improved model for prediction of retention times of tryptic peptides in ion pair reversed-phase HPLC its application to protein peptide mapping by off-line HPLC-MALDI MS, Mol. Cell Proteomics 3 (2004) 908–919, http://dx. doi.org/10.1074/mcp.M400031-MCP200. [12] I.A. Tarasova, A.A. Goloborodko, T.Y. Perlova, M.L. Pridatchenko, A.V. Gorshkov, V.V. Evreinov, A.R. Ivanov, M.V. Gorshkov, Application of statistical thermodynamics to predict the adsorption properties of polypeptides in reversed-phase HPLC, Anal. Chem. 87 (2015) 6562–6569, http://dx.doi.org/10. 1021/acs.analchem.5b00595. ˛ A. Cimochowska, P. Juszczyk, K. Wi´sniewska, Z. [13] R. Kaliszan, T. Baczek, Grzonka, Prediction of high-performance liquid chromatography retention of peptides with the use of quantitative structure-retention relationships, Proteomics 5 (2005) 409–415, http://dx.doi.org/10.1002/pmic.200400973. ´ M.S. Lipton, [14] K. Petritis, L.J. Kangas, P.L. Ferguson, G.A. Anderson, L. Paˇsa-Tolic, K.J. Auberry, E.F. Strittmatter, Y. Shen, R. Zhao, R.D. Smith, Use of artificial neural networks for the accurate prediction of peptide liquid chromatography elution times in proteome analyses, Anal. Chem. 75 (2003) 1039–1048, http:// dx.doi.org/10.1021/ac0205154. [15] A.A. Klammer, X. Yi, M.J. MacCoss, W.S. Noble, Improving tandem mass spectrum identification using peptide retention time prediction across diverse chromatography conditions, Anal. Chem. 79 (2007) 6111–6118, http://dx.doi.org/10.1021/ac070262k. [16] G. Hackeling, Mastering Machine Learning With Scikit-learn, Packt Publishing, 2014. [17] L. Moruz, D. Tomazela, L. Käll, Training, selection, and robust calibration of retention time models for targeted proteomics, J. Proteome Res. 9 (2010) 5209–5216, http://dx.doi.org/10.1021/pr1005058. [18] J. Kyte, R.F. Doolittle, A simple method for displaying the hydropathic character of a protein, J. Mol. Biol. 157 (1982) 105–132, http://dx.doi.org/10. 1016/0022-2836(82)90515-0. [19] L.R. Snyder, J.W. Dolan, High-Performance Gradient Elution: The Practical Application of the Linear-Solvent-Strength Model, Wiley, 2007. [20] U.D. Neue, Peak capacity in unidimensional chromatography, J. Chromatogr. A 1184 (2008) 107–130, http://dx.doi.org/10.1016/j.chroma.2007.11.113. [21] D. Åsberg, M. Le´sko, M. Enmark, J. Samuelsson, K. Kaczmarski, T. Fornstedt, Fast estimation of adsorption isotherm parameters in gradient elution preparative liquid chromatography. I: the single component case, J. Chromatogr. A 1299 (2013) 64–70, http://dx.doi.org/10.1016/j.chroma.2013. 05.041. [22] D. Åsberg, M. Le´sko, M. Enmark, J. Samuelsson, K. Kaczmarski, T. Fornstedt, Fast estimation of adsorption isotherm parameters in gradient elution preparative liquid chromatography II: the competitive case, J. Chromatogr. A 1314 (2013) 70–76, http://dx.doi.org/10.1016/j.chroma.2013.09.003. [23] M. Le´sko, D. Åsberg, M. Enmark, J. Samuelsson, T. Fornstedt, K. Kaczmarski, Choice of model for estimation of adsorption isotherm parameters in gradient elution preparative liquid chromatography, Chromatographia 78 (2015) 1293–1297, http://dx.doi.org/10.1007/s10337-015-2949-0.

[24] D. Åsberg, A. Langborg Weinmann, T. Leek, R.J. Lewis, M. Klarqvist, M. Le´sko, K. Kaczmarski, J. Samuelsson, T. Fornstedt, The importance of ion-pairing in peptide purification by reversed-phase liquid chromatography, J. Chromatogr. A 1496 (2017) 80–91, http://dx.doi.org/10.1016/j.chroma.2017.03.041. [25] D. Åsberg, M. Le´sko, T. Leek, J. Samuelsson, K. Kaczmarski, T. Fornstedt, Estimation of nonlinear adsorption isotherms in gradient elution RP-LC of peptides in the presence of an adsorbing additive, Chromatographia 80 (2017) 961–966, http://dx.doi.org/10.1007/s10337-017-3298-y. [26] K. Kaczmarski, M. Mazzotti, G. Storti, M. Mobidelli, Modeling fixed-bed adsorption columns through orthogonal collocations on moving finite elements, Comput. Chem. Eng. 21 (1997) 641–660, http://dx.doi.org/10.1016/ S0098-1354(96)00300-6. [27] P. Brown, G. Byrne, A. Hindmarsh, VODE: a variable-coefficient ODE solver, SIAM J. Sci. Stat. Comput. 10 (1989) 1038–1051, http://dx.doi.org/10.1137/ 0910062. [28] R. Fletcher, A Modified Marquardt Sub-routine for Non-linear Least Squares, Atomic Energy Research Establishment, Harwell, UK, 1971. [29] C. Wenda, A. Rajendran, Enantioseparation of flurbiprofen on amylose-derived chiral stationary phase by supercritical fluid chromatography, J. Chromatogr. A 1216 (2009) 8750–8758. [30] A. Felinger, A. Cavazzini, G. Guiochon, Numerical determination of the competitive isotherm of enantiomers, J. Chromatogr. A 986 (2003) 207–225, http://dx.doi.org/10.1016/S0021-9673(02)01919-2. [31] P. Forssén, R. Arnell, T. Fornstedt, An improved algorithm for solving inverse problems in liquid chromatography, Comp. Chem. Eng. 30 (2006) 1381–1391, http://dx.doi.org/10.1016/j.compchemeng.2006.03.004. [32] S. Andersson, H. Nelander, K. Öhlén, Preparative chiral chromatography and chiroptical characterization of enantiomers of omeprazole and related benzimidazoles, Chirality 19 (2007) 706–715, http://dx.doi.org/10.1002/chir. 20375. [33] F. Gritti, G. Guiochon, Adsorption behaviors of neutral and ionizable compounds on hybrid stationary phases in the absence (BEH-C18) and the presence (CSH-C18) of immobile surface charges, J. Chromatogr. A 1282 (2013) 58–71, http://dx.doi.org/10.1016/j.chroma.2013.01.024. [34] F. Gritti, G. Guiochon, Effect of the ionic strength on the adsorption process of an ionic surfactant onto a C18-bonded charged surface hybrid stationary phase at low pH, J. Chromatogr. A 1282 (2013) 46–57, http://dx.doi.org/10. 1016/j.chroma.2013.01.027. [35] M. Shibue, C.T. Mant, R.S. Hodges, Effect of anionic ion-pairing reagent concentration (1–60 mM) on reversed-phase liquid chromatography elution behaviour of peptides, J. Chromatogr. A 1080 (2005) 58–67, http://dx.doi.org/ 10.1016/j.chroma.2005.02.047. [36] Y. Chen, A.R. Mehok, C.T. Mant, R.S. Hodges, Optimum concentration of trifluoroacetic acid for reversed-phase liquid chromatography of peptides revisited, J. Chromatogr. A 1043 (2004) 9–18, http://dx.doi.org/10.1016/j. chroma.2004.03.070. [37] M. Gilar, H. Xie, A. Jaworski, Utility of retention prediction model for investigation of peptide separation selectivity in reversed-phase liquid chromatography: impact of concentration of trifluoroacetic acid, column temperature, gradient slope and type of stationary phase, Anal. Chem. 82 (2010) 265–275, http://dx.doi.org/10.1021/ac901931c. [38] M.W. Dong, K. Zhang, Ultra-high-pressure liquid chromatography (UHPLC) in method development, TrAC 63 (2014) 21–30, http://dx.doi.org/10.1016/j.trac. 2014.06.019. [39] D. Åsberg, M. Le´sko, J. Samuelsson, K. Kaczmarski, T. Fornstedt, Method transfer from high-pressure liquid chromatography to ultra-high-pressure liquid chromatography. I. A thermodynamic perspective, J. Chromatogr. A 1362 (2014) 206–217, http://dx.doi.org/10.1016/j.chroma.2014.08.051. [40] P. Petersson, M.R. Euerby, Characterisation of RPLC columns packed with porous sub-2 ␮m particles, J. Sep. Sci. 30 (2007) 2012–2024, http://dx.doi.org/ 10.1002/jssc.200700086. [41] D. Åsberg, J. Samuelsson, M. Le´sko, A. Cavazzini, K. Kaczmarski, T. Fornstedt, Method transfer from high-pressure liquid chromatography to ultra-high-pressure liquid chromatography. II. Temperature and pressure effects, J. Chromatogr. A 1401 (2015) 52–59, http://dx.doi.org/10.1016/j. chroma.2015.05.002. [42] D. Åsberg, M. Chutkowski, M. Le´sko, J. Samuelsson, K. Kaczmarski, T. Fornstedt, A practical approach for predicting retention time shifts due to pressure and temperature gradients in ultra-high-pressure liquid chromatography, J. Chromatogr. A 1479 (2017) 107–120, http://dx.doi.org/10. 1016/j.chroma.2016.11.050. [43] D. Åsberg, J. Samuelsson, T. Fornstedt, A fundamental study of the impact of pressure on the adsorption mechanism in reversed-phase liquid chromatography, J. Chromatogr. A 1457 (2016) 97–106, http://dx.doi.org/10. 1016/j.chroma.2016.06.036. [44] D.V. McCalley, The impact of pressure and frictional heating on retention, selectivity and efficiency in ultra-high-pressure liquid chromatography, TrAC 63 (2014) 31–43, http://dx.doi.org/10.1016/j.trac.2014.06.024. [45] J. Kostka, F. Gritti, G. Guiochon, K. Kaczmarski, Modeling of thermal processes in very high pressure liquid chromatography for column immersed in a water bath: application of the selected models, J. Chromatogr. A 1217 (2010) 4704–4712, http://dx.doi.org/10.1016/j.chroma.2010.05.018. [46] M.M. Fallas, U.D. Neue, M.R. Hadley, D.V. McCalley, Investigation of the effect of pressure on retention of small molecules using reversed-phase ultra-high-pressure liquid chromatography, J. Chromatogr. A 1209 (2008) 195–205, http://dx.doi.org/10.1016/j.chroma.2008.09.021.