3.1. Optimization of the Separation Conditions
RP HPLC is currently the technique of choice for the determination of phenolic compounds in a wide variety of food samples and related food supplements. Beyond the excellent performance of RP HPLC for the study of the vast majority of phenolic families, it has shown some limitations for dealing with the oligomeric nature of condensed tannins or PACs, and compounds with DP higher than 3 usually remain undetected. In this case, the separation mechanism mostly depends on polarity issues [
21], as there is no correspondence between size and retention behavior.
Most of the former scientific publications devoted to the assessment of the oligomeric composition of PACs in fruits and vegetables relied on NP-HPLC in which the separation behavior was related to the size of the molecules. In the NP approach, methods commonly use low polar organic solvents such as methylene chloride or toluene and a growing gradient of a more polar solvent (methanol, acetone or acetonitrile) containing a low percentage of an organic acid (acetic or formic acid), while the stationary phase consists of bare silica particles in which interactions occur among the silanol groups and the polar groups of the analytes [
14]. As the separation pattern, chromatograms clearly showed peaks or bands of compounds clustered according to their degree of polymerization, with retention times increasing from monomers to bigger oligomers.
Despite the initial success of NP separations for the characterization of oligomeric polyphenols, some drawbacks were encountered regarding the occurrence of long separations as well as the limited variety of stationary phases. The combination of silica columns with more polar organic solvent (such as acetonitrile or methanol) results in the HILIC mode has also demonstrated its great performance for the separation of different types of tannins depending on size features. Besides, apart from silica, a wide range of stationary phases with polar, weak ion exchange and zwitterionic groups have been developed for HILIC separations.
The extension of NP chromatographic separation proposed by Gu et al. to HILIC was taken as the starting point to optimize the separation of flavanol-related components occurring in the dietary supplements and nutraceutical samples analyzed [
23]. The main objective of this optimization was to simplify the chromatographic separation through a gradient with a single slope and to reduce the analysis time since, in general, chromatographic runs longer than 60 min were required NP methods.
As indicated in the experimental section, the mobile phase was generated from 1% acetic acid in acetonitrile (solvent A) and MeOH:water:acetic acid (95:3:2, v/v/v) (solvent B). Preliminary experiments indicated that PACs have a greater affinity to polar solvents; as a result, increasing the polarity of the mobile phase by increasing the percentage of solvent B accelerated their elution. Nevertheless, the use of stronger elution conditions resulted in more overlapping peaks of components. Elution limits of the most and less retained species were studied to establish the working range of solvent B, concluding that ca. 10% solvent B was necessary for the retention of monomers (i.e., the less retained species) while the solvent B percentage should be increased up to 40% to elute bigger oligomers (i.e., the most retained species).
A systematic optimization of the separation conditions was attempted based on experimental design approaches, being the experimental factors under study, their experimental domain and the optimization criterion are key aspects to be defined. Apart from a preliminary study using pure standards of each analyte, the separation performance was assessed using a cranberry extract as a representative sample. The cranberry extract was rich in a wide range of phenolic compounds, especially flavanol species which are the main targets of this paper, thus, this matrix was considered a good example to address HILIC optimization.
On the one hand, experimental factors to be considered were the initial percentage of solvent B and the gradient time (to reach 40% solvent B). In this way, the experimental plan consisted of a 2 factor at 3 level full factorial design, involving 9 runs, in which the initial percentage of solvent B was studied at 10, 15 and 20% and the gradient time at 30, 45 and 60 min. In all cases, a further cleaning and conditioning step was added after the separation stage, in which the solvent B percentage increased up to 85%, was maintained at this percentage for 5 min, and returned to the initial conditions for column stabilization. Experiments were carried out randomly to avoid the influence of systematic drifts on the results.
On the other hand, the optimization criterion relied on reaching a suitable compromise among best separation capacity (i.e., maximum number of resolved compounds) and analysis time (i.e., fastest analysis).
Figure 1 depicts the response surface corresponding to the optimization of the separation under the HILIC mode.
Figure 1a shows the number of compounds resolved represented as a function of the experimental variables (solvent B percentage and gradient time), and the maximum of this surface theoretically corresponded to the optimal conditions given by this design. As can be seen in this figure, the best separation was obtained by increasing the percentage of solvent B from 10 to 40% in 60 min.
Figure 1b shows the retention time of the last analyte peak represented as a function of the experimental variables (solvent B percentage and gradient time). The optimization criterion was mathematically expressed as a response function (desirability function) combining both contributions of resolved compounds and retention time, given by the number of analyte peaks found in the chromatogram and the retention time of the last analyte peak in the chromatogram, respectively. In order to equalize the influence of each contribution to the overall response function, in other words, minimizing the influence of the different scales, each response was normalized from 0 to 1, with 0 representing the worse results and 1 the best ones. The overall response function was then calculated as the geometric mean of number of peaks (n
peaks) and analysis time (t
R): D = (dn
peaks × dt
R)
1/2.
Figure 1c shows the effect of initial methanol percentage and gradient time on the separation of flavanol-related components expressed as a desirability function D. The best conditions were 10–40% in 60 min. Then, the gradient profile was further modified to shorten the analysis time by reducing those initial and final chromatographic ranges without significant peaks of analytes (no peaks were detected with retention times higher than 30 min). Hence, the separation chosen consisted of increasing the solvent B percentage from 10 to 25% in 30 min; after that period, cleaning and conditioning steps were applied.
Under the selected conditions, sample chromatograms recorded by UV absorption spectroscopy at 280 nm resulted in complex patterns with a great diversity of peaks and multiple overlapping; peaks detected corresponded to a wide range of phenolic compounds, including hydroxybenzoic and hydroxycinnamic acids, stilbenes, flavonoids, et cetera. Nevertheless, in order to simplify the chromatographic profiles and focused on flavanols and other structurally related families, data was recorded by FLD at λexc 276 and λem 316 nm. The higher selectivity and sensitivity achieved by this detection technique led to cleaner chromatograms in which many interferences remained undetectable.
3.2. Analysis of the Sample Extracts
Nutraceuticals and food supplements were analyzed in triplicate, from the injection of three independent sample extracts which were analyzed randomly in order to minimize the influence of possible instrumental drifts in the results. Besides, the QC was analyzed every 10 sample injections to assess repeatability of the chromatographic method. Regarding compositional issues, apart from catechin, epicatechin, procyanidin A2, procyanidin B2 and procyanidin C1, which were confirmed from commercial standards, other relevant peaks were observed in the chromatograms. Peaks with retention times higher than 14 min were attributed to bigger oligomers, including clusters of other trimers, tetramers, et cetera. Unfortunately, standards for these species were not available so they could only be tentatively deduced from MS measurements. For instance, B-type oligomers exhibited m/z values for the [M-H]
− ions of 865 (trimers), 1155 (tetramers) 1445 (pentamers) and 1735 (hexamers); when A-type links occurred in these molecules, the m/z values decreased in 2 units per link [
24,
25]. In the range, 3.8–8.6 min, peaks were attributed to hydroxybenzoic acids (e.g., gallic acid, vanillic acid and gentisic acid) and flavanones (hesperidin, naringenin, etc.) that may be abundant in some of the samples. These compounds have a fluorescent moiety similar to that of flavanols so they were also detected by FLD at the excitation and emission wavelengths used. The individual resolution of these compounds was not attempted here as our description was mainly focused on flavanols.
Table S1 shows some relevant method features estimated from the series of QCs. As can be seen, the values of the relative standard deviation in the retention time, expressed as a percentage (RSD%), were lower than 1%, thus indicating a good repeatability of the chromatographic events throughout the series of analyses. In the case of peak areas, RSD% values of catechin, epicatechin, procyanidin A2, procyanidin B2 and procyanidin C1 ranged from 1.3 to 6.7%. For the rest of the peaks, the variability was lower than 10% except for the bigger oligomers that exhibited higher RSD% values, possibly due to the low intensities of the corresponding peaks.
As an example,
Figure 2 shows some representative FLD chromatograms of cranberry, raspberry, grapevine, red grape, antiox products and artichoke samples. For samples manufactured with purified cranberry extracts, epicatechin and procyanidin A2 were the most remarkable compounds although some oligomeric species were also detected. In some cases, cranberry extracts were combined with other medicinal plants, thus resulting in more complex chromatograms in terms of compound diversity. Raspberry extracts were rich in catechin, epicatechin and procyanidin B2. In the case of the grapevine sample, the most relevant peaks corresponded to catechin, epicatechin, procyanidin B2 and some trimer and tetramer species. Red grape samples were abundant in procyanidin C1. All the artichoke samples displayed simpler profiles, catechin and epicatechin being the principal species. Antiox products exhibited the most complex profiles in accordance with the variety of ingredients that contain them, including red berries, pomegranate, grapevine, plum, strawberry, et cetera. Since differences in the composition of the samples were significant, the compositional profiles were exploited as a source of potential descriptors to characterize and discriminate the nutraceutical samples.
3.3. Sample Characterization by Principal Component Analysis and Related Methods
A comprehensive characterization of dietary supplements and nutraceuticals was tackled under both profiling and fingerprinting approaches. In the first strategy, data consisted of concentrations of given analytes determined for the set of samples. Strictly speaking, the quantification can only be done in the case of analytes for which commercial standards were available, namely, catechin, epicatechin, procyanidin A2, procyanidin B2 and procyanidin C1. In a broader sense, however, peak areas of non-quantified compounds can also be used to enrich the data matrix. Further data pretreatment consisting of autoscaling was applied to equalize the differences in the magnitude and amplitude of the different scales regardless of concentrations or areas that were used.
Principal component analysis (PCA) was applied as an exploratory method to characterize samples regarding the type of ingredients, such as berries, grape-related products, artichoke, et cetera. The PCA profiling model performed considering all detected peaks showed some sample clustering according to type of product (
Figure S2). Cranberry samples were mainly discriminated from artichoke samples while raspberry, grape and antiox samples showed some overlapping; this finding was attributed to the influence of phenolic acids and other flavonoids (different to flavanols) on the model. In order to try to achieve a better separation among groups of samples with similar features, another PCA model was attempted using flavanols only. PCA results (see
Figure 3) showed that similar samples were grouped in compact groups, with cranberries on the left, artichokes in the center and the other classes mainly on the right side. Red grape-based samples appeared dispersed since, although grape was the main ingredient, they also contained other plant extracts (grape alone, grape + pomegranate and grape + polygonum + onion). It should be pointed out that QCs were tightly clustered in the center of the model, thus demonstrating the stability and reproducibility of the chromatographic data throughout the entire series of measurements as well as the suitability of the PCA. It was also deduced that PC2 mainly distinguished raspberry and grapevine samples (top right corner) from red grape and antiox samples (bottom).
The map of loadings (
Figure 3b), procyanidin A2 and trimer 1 (A type trimer), on the left side, were well correlated variables (r
2 = 0.68), being the most characteristic descriptors of cranberry-based samples. Procyanidin B2, on the top right corner, was especially abundant in raspberry and grapevine preparations. Procyanidin C1, on the middle right, was plentiful in red grape samples. Epicatechin was ubiquitous to all the samples so it was found close to the center of the graph. Catechin was more characteristic of the antiox combinations.
The application of supervised methods for class modeling based on PLS-DA used class information to achieve a more efficient model. Results shown in
Supplementary Materials (Figure S3) indicated that five latent variables were used to establish a PLS-DA model. Plots of scores and loadings provided similar conclusions to PCA with regards to the main class descriptors, although, in this case, the groups were better defined.
Sample characterization based on chromatographic fingerprints required an exhaustive data preprocessing in order to obtain reliable results. First, although the reproducibility of the chromatographic data was noticeable, in general, retention time of analytes increased slightly throughout the series of injections. This peak shifting phenomenon was more remarkable for compounds with higher retention times. Hence, chromatograms were aligned using representative peaks as a reference, such as the epicatechin peak which was common to all the samples. After synchronization, trends associated with the working time were fully corrected and the influence of this factor on the results became negligible, as it was deduced from the study of the behavior of QCs.
The selection of the most appropriate chromatographic time range was another important issue that affected the performance of the description models. When using the whole chromatogram, various time ranges containing useless data (e.g., holdup time, cleaning step and column stabilization) were kept in the dataset. In contrast, all relevant compounds were located within the working range 3.5–25.5 min, from phenolic acids to oligomers. In this case, as occurred in the profiling approach, the PCA model revealed that samples were mainly grouped according to their classes although some different samples were found in the same areas (see
Figure S4). Various shorter ranges were investigated, and the best results were obtained when working in the time window from 6.2 to 16.0 min, which included the most important flavanolic compounds.
Chromatograms were further smoothed according to a Savitzky–Golay filter (order 2 and 15-point window). Finally, chromatograms were scaled by vector normalization to equalize the influence of each sample regardless the overall concentration of analytes. Normalization was applied by dividing each absorbance value throughout the chromatogram by the vector norm, in other words, square root of the sum of squares.
PCA results from the filtered dataset are depicted in
Figure 4. As can be seen, cranberry and artichoke samples were found in compact groups on the top and left parts of the graph, respectively, which suggested that the composition of these samples was quite homogenous within each class. Some raspberry samples as well as antiox and red grape products were distributed on the right side of the graph. The study and interpretation of the loading vectors here were more complex. In this regard, catechin seemed to have a high weight in PC1 while procyanidin A2 and epicatechin mainly influenced PC2 (see
Figure 4b).
In the
Supplementary Materials, Figure S5 summarizes the PLS-DA results, showing a more marked sample structuration according to the type of products; in this case, the optimum number of latent variables established from Venetian blinds cross validation was 7, which again indicated the great complexity of the FLD chromatographic fingerprints.