Advancing Fundamental Understanding of Retention Interactions in Supercritical Fluid Chromatography Using Artificial Neural Networks: Polar Stationary Phases with –OH Moieties

In the research article published recently in the Analytical Chemistry journal, the researchers from the Faculty of Pharmacy in Hradec Králové, Charles University, Czechia, and Toulouse III Paul Sabatier University, Toulouse, France, used artificial neural networks to advance a fundamental understanding of retention interactions in supercritical fluid chromatography.

The study investigates the retention behavior and stability of polar stationary phases in supercritical fluid chromatography (SFC) over time. Using silica, hybrid silica, and diol columns, the research correlates molecular descriptors of over 100 analytes with their retention, employing various CO₂-based mobile phases with and without additives. A deep learning model was trained on extensive experimental data to identify key molecular descriptors influencing retention. The study provides a comprehensive analysis of retention changes over a year of column use, suggesting explanations for these changes based on alterations to the stationary phase surface. The findings offer valuable insights for selecting appropriate stationary phases for specific analytes in SFC.

The original article

Advancing Fundamental Understanding of Retention Interactions in Supercritical Fluid Chromatography Using Artificial Neural Networks: Polar Stationary Phases with –OH Moieties

Kateřina Plachká, Veronika Pilařová, Tat’ána Gazárková, František Švec, Jean-Christophe Garrigues, and Lucie Nováková

Analytical Chemistry 2024 96 (31), 12748-12759

DOI: 10.1021/acs.analchem.4c01811

licensed under CC-BY 4.0

Selected sections from the article follow. Formats and hyperlinks were adapted from the original.

Abstract

The retention behavior in supercritical fluid chromatography and its stability over time are still unsatisfactorily explained phenomena despite many important contributions in recent years, especially focusing on linear solvation energy relationship modeling. We studied polar stationary phases with predominant –OH functionalities, i.e., silica, hybrid silica, and diol columns, and their retention behavior over time. We correlated molecular descriptors of analytes with their retention using three organic modifiers of the CO₂-based mobile phase. The differences in retention behavior caused by using additives, namely, 10 mmol/L NH₃ and 2% H₂O in methanol, were described in correlation to analyte properties and compared with the CO₂/methanol mobile phase. The structure of >100 molecules included in this study was optimized by semiempirical AM1 quantum mechanical calculations and subsequently described by 226 molecular descriptors including topological, constitutional, hybrid, electronic, and geometric descriptors. An artificial neural networks simulator with deep learning toolbox was trained on this extensive set of experimental data and subsequently used to determine key molecular descriptors affecting the retention by the highest extent. After comprehensive statistical analysis of the experimental data collected during one year of column use, the retention on different stationary phases was fundamentally described. The changes in the retention behavior during one year of column use were described and their explanation with a proposed interpretation of changes on the stationary phase surface was suggested. The effect of the regeneration procedure on the retention was also evaluated. This fundamental understanding of interactions responsible for retention in SFC can be used for the evidence-based selection of stationary phases suitable for the separation of particular analytes based on their specific physicochemical properties.

Supercritical fluid chromatography (SFC) has undergone an important transformation over the years to increase its applicability in various fields. (1) As a result, SFC evolved from a marginal method used primarily for the analysis of nonpolar compounds to the method of choice for the analysis of compounds with a wide range of polarity and physicochemical properties. (1) However, this technique is still primarily considered a research tool rather than a routine method, even though the causes of the technique’s negative reputation, including a lack of method robustness, instrument unreliability, and complex technology transfer, were already mitigated. (2) In recent years, several studies on interlaboratory validation have been carried out confirming the robustness and repeatability of SFC methods. (3−5) Nevertheless, several negative aspects related to long-term retention time (t_R) stability have also been described for SFC. (6,7)

Current state-of-the-art SFC typically uses a mobile phase containing CO₂ and organic modifiers such as methanol (MeOH) or other alcohols. The organic modifier can interact with the free acidic silanols on the silica surface of the stationary phase to form silyl ethers. (6) Silyl ether formation (SEF) reduces the number of free silanols that are no longer involved in the interactions between the stationary phase surface and analytes, causing changes in selectivity over time. In addition, the SEF reaction forms water as a byproduct, which acts as a polar additive and affects the retention and separation selectivity. SEF can be catalyzed by acid and base, commonly used as additives to the mobile phase. The reaction kinetics can also be correlated with the modifier composition. (6) The SEF is a condensation reaction that can be reversed by water. Therefore, a small percentage of water in the organic modifier, i.e., 2–5%, could shift the equilibrium toward the free silanols and mitigate the SEF. Thus, the process of regeneration, where the column is washed with a large volume of water, is suggested to reverse the SEF. (6−9) However, the SEF phenomenon is still not well understood, and more detailed studies are necessary. Furthermore, organic acids, ammonia, water, and/or buffers can be added to the mobile phase, affecting the separation of acidic/basic analytes and their peak shapes. This additive can be adsorbed on the silica surfaces of the stationary phase. (10) Its removal presents an additional problem, resulting in a change in selectivity over time, especially when different additives are used on the same column. (11−13)

Polar stationary phases with predominant –OH functionalities have been used in more than 35% of published works. (9) At the same time, the –OH functionalities of these phases are more prone to SEF. (9) The well-established linear solvation energy relationship (LSER) classification (14) sorts these columns among the polar stationary phases in two clusters (Supporting Information Figure S1): (i) nonbonded silica and hybrid silica and (ii) polar ligands bonded to the silica surface. Polar ligands also include other functional groups such as amino groups, cyano groups, and 2-ethylpyridine groups which will be discussed within the following paper. Looking at the individual parameters of the LSER equation (Supporting Information, Figure S1B), all columns discussed in this paper have strong dipole–dipole and π–π interactions (terms e and s), hydrogen bonding with acids and bases (a and b terms), and interactions with cations (d⁺ term). The main differences can be seen in the magnitude of the terms. Furthermore, the differences in retention behavior can be correlated to different physicochemical properties of –OH functionalities. Free silanols in bare silica are easily ionizable and thus affected by the mobile phase pH. Their pK_a has been estimated to be 4 – 7. (15,16) pK_a values vary for different types of silanols, including geminal, isolated, and vicinal. Some silanols can form strong hydrogen bonds with water via proton sharing, indicating a higher acidity (pK_a ≈ 2.9 – 4.6). In contrast, some silanol groups with pK_a ≈ 8.9 can be deprotonated, i.e., forming SiO^–, and be stabilized by nearby –OH. (15,16) For hybrid stationary phases, different acidities are expected as the free unreacted silanols are sterically and hydrophobically hindered by methyl groups to prevent further attack of the silica surface by the mobile phase. Additionally, a less acidic support is used. In fact, the pK_a for the first generation of hybrids (XTerra) was estimated to be ≈ 9 – 11 based on the mobile phase composition. For the diol column, the pK_a of –OH functionalities was estimated to be around 14. (15−18) Later, two other terms, sphericity (gG) and flexibility (Ff), were added to the LSER classification. (19) The positive contribution of g, indicating a higher retention of spherical molecules, and the negative contribution of f, indicating a lower retention of flexible molecules, were described. (19) However, no model taking into account also the localization of the charge and detailed parameters of the 2D and 3D analyte structure has been proposed, yet.

Our study focuses on the determination of the differences between bare silica, hybrid silica, and diol columns and increases the knowledge of SEF. Three mobile phase compositions were tested, including CO₂ with (i) neat MeOH with apparent pH ≈ 5, (ii) MeOH + 2% H₂O with apparent pH ≈ 1, and (iii) MeOH + 10 mmol/L NH₃ with apparent pH ≈ 7 – 8. (20) The tested organic modifiers were selected to cover the most commonly used SFC mobile phases. MeOH enabled us to describe the retention mechanism without interactions caused by the additive. Furthermore, the results obtained using MeOH served as a baseline for the evaluation of SEF. The use of MeOH + H₂O causes acidic apparent pH of the mobile phase similarly to other acidic additives such as formic acid. (20) Moreover, the beneficial effect of H₂O addition on retention stability in SFC has been previously reported. (5) MeOH + NH₃ was selected as the most straightforward example of an ammonium-based additive. Indeed, when using ammonium salts as additives, both ions, e.g., ammonia and formate, can affect the retention mechanism. In our study, we can be certain that all of the observed interactions are caused by either MeOH or NH₃. Furthermore, Ovchinnikov et al. showed that diethylamine and ammonium acetate caused identical changes of LSER parameters. (21) All experiments within our study were carried out under typical SFC conditions to enable easy transfer of the results. (12,22) Structures of >100 analytes were described by topological, constitutional, hybrid, electronic, and geometric descriptors. This extensive set of experimental data was used to train the artificial neural networks (ANN) simulator with deep learning toolbox which then linked the structure of the analytes to the observed retention. The aims of the study included the following: (i) a fundamental description of the retention behavior on polar stationary phases related to specific molecular features of the analytes, (ii) quantitative description of the changes in retention behavior during one year of column use and their explanation with a proposed interpretation of changes on the stationary phase surface, and (iii) the investigation of the effect of the regeneration procedure.

Experimental Section

Chemicals

Methanol (MeOH), acetonitrile (ACN), 2-propanol (IPA), and water of LC/MS grade quality were provided by VWR International (Prague, Czech Republic). Ammonia (4 mol/L) solution in MeOH for LC/MS was purchased from Sigma-Aldrich (Steinheim, Germany). Pressurized liquid CO₂ 4.5 grade (99.9995%) was purchased from Messer (Prague, Czech Republic). Most of 107 reference standards listed in Supporting Information Table S1 were purchased from Sigma-Aldrich (Prague, Czech Republic). Several standards were kindly donated by Zentiva, k.s. (Prague, Czech Republic).

Standard Solutions

Standard solutions of all reference standards were prepared by dissolving each compound in MeOH. The reference standards were then divided into 12 mixtures specific for each column and organic modifier and diluted to the final concentration of 50 μg/mL by ACN.

Analytical Instrumentation and Procedure

The experiments were carried out using an Acquity UPC² SFC system (Waters, Milford, MA, USA) equipped with a binary pump, an autosampler, a column thermostat, a back pressure regulator (BPR), and a PDA detector. The system was coupled to a single quadrupole detector (QDa, Waters) via a commercial SFC-MS dedicated pre-BPR splitter device with an additional isocratic pump for the make-up solvent delivery (Waters).

A generic gradient method was used with a mobile phase consisting of (A) CO₂ and (B) organic modifier at a flow rate of 1.5 mL/min and following gradient program: 2% B for 1 min, 2–45% B in 1–5 min, followed by 1 min of isocratic step at 45% B and 1.5 min of equilibration at initial conditions. Three organic modifiers were tested: MeOH, MeOH + 10 mmol/L NH₃, and MeOH + 2% H₂O. The column temperature was 40 °C and the BPR pressure was 13 MPa. The BPR was adjusted for each measurement sequence to avoid t_R variations due to changes in the system pressure. The BPR was manually adjusted before each sequence so that the system pressure for the blank injection overlapped the system pressure of the first sequence within 0.07 MPa. The autosampler temperature was 10 °C and the injection volume was 2 μL. Peak detection and integration was carried out using the PDA detector, with data collected in the range of 210 to 400 nm. The MS detector with electrospray ionization in positive and negative modes enabled the confirmation of each analyte. MeOH + 10 mmol/L NH₃ was used as a make-up solvent at a flow rate of 0.3 mL/min.

Columns and Regeneration Procedure

Three stationary phases with the same dimensions (100 × 3.0 mm) were tested: nonbonded silica (Zorbax HILIC Plus, Agilent Technologies, Inc., CA, USA, silica), bridged ethylene hybrid (Viridis BEH, Waters, BEH), and high density diol with pure propanediol linker (Torus Diol, Waters, diol). All columns were packed with 1.7 μm particles except for the silica column with 1.8 μm particles. Prior to the first injection, the columns were flushed with CO₂/MeOH (50/50) at 1.5 mL/min for 35 min to eliminate further retention shifts. (23) A separate column was used for each organic modifier, but the three columns were always from the same batch to mitigate interbatch variability and ensure the same retention properties. Column regeneration was carried out on an Acquity UPLC system, Waters (Milford, USA). The procedure, in agreement with previous findings and Waters Column Care & Manual Guide, (6,7) included washing with >200 column volumes of H₂O at 0.6 mL/min for 280 min, followed by >10 column volumes of IPA/H₂O (9/1, v/v) at 0.5 mL/min for 20 min, and >10 column volumes of IPA at 0.5 mL/min for 20 min.

Study Design

Eight data points were collected for each column at defined time periods: first injection (month 0), month 1 (1M), 2M, 3M, 6M, 9M, and 12M. The column was then regenerated according to the regeneration procedure, and the last data point (R) was collected. Prior to measurement at each data point, the column was flushed with the CO₂/organic modifier (55/45) at 1.5 mL/min for 15 min and then equilibrated with the CO₂/organic modifier (98/2) at 1.5 mL/min for 15 min. Blank, standard mixtures, and blank were injected within the sequence, each in triplicate. After the use, the column was washed with CO₂/MeOH (55/45) at 1.5 mL/min for 30 min (>20 column volumes) and neat CO₂ at 0.6 mL/min for 30 min. CO₂ was used as the storage solvent (6) to avoid column aging and to eliminate t_R shifts.

Data Evaluation

Raw data were processed using Empower 3 to collect t_R and peak widths at 5% of peak height. The % change in t_R over time was calculated for each analyte, column, and organic modifier (Microsoft Excel, version 2302). The 3D structures of analytes were optimized by semiempirical AM1 quantum mechanical calculations using the MOPAC application of Chem 3D Pro version 14.0 software (CambridgeSoft). A root-mean-square gradient of 0.100 was used to minimize the energy for all of the compounds. These optimized structures were then used for computing 2D and 3D molecular descriptors (CDK Descriptor Calculator, v.1.4.8). The 226 calculated molecular descriptors included topological, constitutional, hybrid, electronic, and geometric descriptors of the 2D and 3D structure of the molecule and are listed in Supporting Information Table S2 and categorized in Supporting Information Table S3. Molecular descriptors and retention factors were normalized by dividing by the maximal value.

To identify key molecular descriptors linking the structure of analytes to their retention on different stationary phases, ANN were created using the neural network simulator in Matlab R2023a with the deep learning toolbox V.23.2 (The MathWorks, Inc., Massachusetts, USA) and a sigmoid activation function, a back-propagation learning algorithm with 500 learning cycles. These ANN were structured with an input layer connected to the 226 molecular descriptors and an output layer linked to the retention factor (k′) of each analyte. After 500 training cycles, the weights assigned to each input neuron were extracted, and the key molecular descriptors, with weights greater than 1.5 in absolute value, were examined. The higher the weight assigned by the ANN, the more the descriptor affects retention. (24) The differences related to the organic modifier used were determined. The standard deviation (SD) of the molecular descriptor weights at each data point were calculated and correlated with the observed changes in retention behavior. The molecular descriptors with the largest changes in weight over time were determined and used to describe changes in the stationary phase surface over time.

To determine the adequacy of the regeneration procedure, the % error between the t_R at the first injection and after the regeneration were calculated: %-error = (t_R at the first injection – t_R after regeneration)/t_R at the first injection. The effectiveness of the regeneration procedure used was obtained by comparing two % differences between: (1) at the first injection and at 12M versus (2) at the first injection and after regeneration. If the (2) % difference is lower than (1), then it means that the regeneration procedure resulted in a t_R closer to the first injection than the t_R observed at 12M.

Results and Discussion

...

Analytical Chemistry 2024 96 (31), 12748-12759: Figure 1. (A) Overlay of chromatograms for selected compounds analyzed on the BEH column using MeOH + 10 mmol/L NH3 at different data points and retention time shifts over time on selected stationary phases using (B) methanol, (C) MeOH + 10 mmol/L NH3, and (D) MeOH + 2% H2O as organic modifier, expressed as %-difference: less than 0.5% (dark blue), 0.5–1.0% (light blue), 1.0–2.0% (yellow), 2.0–5.0% (light red), and over 5.0% (dark red).

...

Conclusions

ANN have been used for the first time to comprehensively define compound properties expressed as molecular descriptors responsible for retention in SFC, specifically on polar stationary phases with predominant –OH functionalities. The key molecular descriptors affecting the retention to the highest extent were defined separately for three different organic modifiers. Overall, the retention behavior on all tested columns could be correlated with the pK_a of the respective –OH functionalities and the apparent pH of the SFC mobile phase based on the organic modifier used. For the first time, we also quantitatively described the changes in the interactions when using 10 mmol/L NH₃ and 2% water as additives compared to pure methanol as an organic modifier. For the hybrid silica column, a high retention of analytes with acidic groups, H bond donor groups, –NH, and a negative charge was observed. The coverage of the molecular surface by the negative charge and its localization played a crucial role in the retention behavior. Changing the organic modifier resulted in significant changes in molecular descriptor weights, especially when using MeOH + NH₃ compared to pure MeOH. Even stronger effect of additive was observed on the silica column. Several important molecular descriptors including the presence of keto oxygen, the number of H bond acceptors, lipophilicity, the ratio of heavy atoms in the framework to the total number of heavy atoms in the molecule, negative surface, and especially –NH groups, and the number of basic groups, played an important role in describing retention behavior on silica using different organic modifiers. For the diol column, the number of H bond acceptor decreased the retention, while it increased with increasing negative surface area and number of basic groups. Here, the retention was significantly less affected by changing the organic modifier.

The detailed description of retention interactions enables the selection of a suitable organic modifier for increasing and/or decreasing retention of particular analytes. This fundamental understanding of interactions responsible for retention in SFC can be used for the separation of analytes based on their properties. Thus, a lower number of time-consuming experiments will be necessary for the development of SFC methods, further increasing the environmental friendliness of the SFC technique.

The best stability of t_R over one year of use was observed for a diol column with –OH functionalities not prone to SEF. For the BEH column, mostly increased t_R values were observed. The addition of NH₃ and/or H₂O to the mobile phase further stabilized the retention. However, a strong decrease in retention was observed for acidic compounds, in contrast to a strong increase in retention for alkaline compounds. The highest instability of t_R was observed on the silica column with a predominant decrease in t_R over time. This can be correlated with the possibility of SEF, as in this case, less –OH is available for the interactions, resulting in lower retention. In addition, the need for longer equilibration was noted when using an organic modifier with an additive on a silica column. A combination of SEF and additive adsorption on the stationary phase surface was responsible for the column aging over time when using MeOH + NH₃. The regeneration procedure used did not have a significant positive effect on the k′ but had a positive effect on peak width, especially on the BEH column. Nevertheless, the regeneration procedure did not meet the expectations, as it did not return the stationary phase to its original state, and the results obtained at the first injection could not be reproduced.