Primary Metabolite Chromatographic Profiling as a Tool for Chemotaxonomic Classification of Seeds from Berry Fruits

Chemicals and materials

The fatty acid standard Supelco® 37 Component FAME Mix was purchased from Sigma-Aldrich, Merck (Burlington, MA, USA). Sugar standards: glucose (Glc), fructose (Fru), saccharose (Sac), trehalose (Tre), maltose (Mal) and arabinose (Ara) were purchased from TCI Europe N.V. (Zwijndrecht, Belgium), turanose (Tur), rhamnose (Rham), gentiobiose (Gent), isomaltose (Ism), panose (Pan), raffinose (Raf), isomaltotriose (Ismt), maltotriose (Malt), melibiose (Mel), xylose (Xyl), celobiose (Cel), melesitose (Mels), maltotriose (Maltotri), maltotetraose (Maltotetr), maltopentaose (Maltopen), maltohexaose (Maltohex) and maltoheptaose (Maltohep) were obtained from Tokyo Chemical Industry (TCI, Tokyo, Japan). Sorbitol (Sor), erythritol (Ert), arabinitol (Arabt), mannitol (Mant) and galactitol (Glt) were purchased from Sigma-Aldrich, Merck (Steinheim, Germany). Ultrapure water (MicroPure water purification system, 0.055 μS/cm, TKA, Thermo Fisher Scientific, Niederelbert, Germany) was used to prepare standard sugar solutions and blanks. Methanol, chloroform, diethylether, acetone and n-hexane were from Merck (KGaA, Darmstadt, Germany). Syringe filters (13 mm, polytetrafluoroethylene (PTFE) membrane 0.45 μm) were purchased from Supelco (Bellefonte, PA, USA). All reagents and chemicals whose purity was not previously emphased were of analytical purity grade.

Plant material

Fruits from cultivars belonging to nine berry fruit species (strawberry, raspberry, blackberry, black currant, blueberry, gooseberry, chokeberry, cape gooseberry and goji berry) (Table 1) were picked in commercial orchards as presented inFig. S1. The data were collected from a network of berry orchards with modern and intensive production system. The locations are situated in the middle of the north temperate zone, having continental climate with warm summers and cold winters. During the experimental year, the average daily temperature was between 11 and 11.5 °C. The average annual amount of precipitation in all locations was about 650 L/m².

Before planting, soil was prepared, weed-controlled, and pH was measured. Planting distance and support was adapted to each fruit species/cultivar. All necessary agrotechnical measurements, such as pruning, fertilization and irrigation were done regularly, while weeding and pest control were done according to the Integrated pest management directive 2009/128/EC (16). All nursery plants used for planting orchards were purchased from certified nurseries who guaranteed that plant material was propagated from a known nucleus plant with inspected phytosanitary status, claiming that it is true to varietal type (accompanied with original tag).

Strawberry orchard was organized on raised double beds covered with black polyethylene foil. Blueberries were planted at raised beds mulched with agrotextile. Plots with raspberry, blackberry, currants, gooseberry, chokeberry, goji berry and cape gooseberry were prepared by clean tillage, where the space between plants, within the row, was hand weeded. All orchards had a system for drip irrigation which was used for fertigation too. In the spring, plants were fertilized with NPK fertilizer (20:20:20) to stimulate vegetative growth, while formulation NPK 11:11:33 was utilized in the cropping stage. For strawberries, foliar application of Ca fertilizers with microelements and amino acids was applied too. For blueberries, ammonium, potassium and magnesium sulphate were added in the phase of fruit setting and fruit development.

At the harvesting time, a sample of 50 fully ripened fruits was collected randomly from all around five plants/bushes (5 plants×10 fruits). Berries that were picked had typical shape and colour, were undamaged and without any signs of diseases. Seeds were manually removed from mesocarp just after picking, rinsed with tap water, air dried in the dark place for 10 days at room temperature (20 °C) and stored in paper bags until the analysis.

Extraction procedures

High-performance thin-layer chromatography analysis

Aliquots of 1 µL of each seed lipid extract were applied to the 20 cm×10 cm HPTLC silica gel plates (Merck, Darmstadt, Germany) as 6-mm band by using Automatic TLC sampler 4 (ATS4; CAMAG, Muttenz, Switzerland). Plates were developed in the saturated twin trough chamber up to a migration distance of 90 mm, using four mobile phases: V(chloroform):V(methanol):V(acetic acid)=90:10:1 up to 25 mm, V(n-hexane):V(diethylether):V(acetone)=60:40:5 up to 70 mm, V(n-hexane):V(diethylether)=97:3 up to 85 mm and 100% n-hexane up to 90 mm. The developed plates were dried in a stream of warm air and dipped in post-chromatographic derivatization solution of cerium ammonium molybdate for 1 s with an immersion speed of 3.5 cm/s, using Chromatogram Immersion Device III (CAMAG). Derivatized plates were heated for 5 min at 110 °C on TLC Plate Heater III (CAMAG). Images were captured at 366 nm with DigiStore 2 device image analysing system in conjunction with Reprostar 3 (CAMAG). Four apertures with exposure time of 30 ms and frame of 2 mm were applied. The photos were stored as TIF files for further image processing.

Gas chromatography-mass spectrometry

Gas chromatography-mass spectrometry (GC-MS) analysis was performed using a Thermo Fisher Focus GC coupled with Polaris Q system (Carlsbad, CA, USA). The column was Thermo Scientific’s TRACE™ TR-WaxMS (30 m×0.25 mm×0.25 µm). Oven temperature program was as follows: initial temperature 50 °C for 1 min, then increased to 200 °C at 25 °C/min and finally to 230 °C at 3 °C/min for 18 min.

Injector, transfer line and ion source temperatures were 250, 260 and 260 °C, respectively. Injection volume was 1 μL and the injector was in split mode (50:1). Helium with a flow rate of 1 mL/min was used as the carrier gas. The analysis time was 35 min. Fatty acid methyl esters were identified by comparison of the retention times with previously injected mixture of the standards and the resulting mass spectra to those of injected standards and those given in the NIST database (19). Results are presented as relative content (%) obtained after integration of the total ion chromatograms.

High-performance anion-exchange chromatography with pulsed amperometric detection

Carbohydrate analysis was performed on an ICS3000 DP liquid chromatography system (Dionex, Sunnyvale, CA, USA) equipped with a quaternary gradient pump (Dionex). The carbohydrates were separated on a CarboPac® PA100 pellicular anion-exchange column (4 mm×250 mm; Dionex) at 30 °C. An ICS AS-DV 50 autosampler (Dionex) was used for sample injection (injection volume for each sample was 25 μL). The mobile phase, consisting of 600 mM sodium hydroxide (A), 500 mM sodium acetate (B) and ultrapure water (C), was used according to the following linear gradient (flow rate, 0.7 mL/min): 0–5 min, 15% A, 85% C; 5.0–5.1 min, 15% A, 2% B, 83% C; 5.1–12.0 min, 15% A, 2% B, 83% C; 12.0–12.1 min, 15% A, 4% B, 81% C; 12.1–20.0 min 15% A, 4% B, 81% C; 20.0–20.1 min 20% A, 20% B, 60% C; 20.1–30.0 min 20% A, 20% B, 60% C. Before the analysis, the system was preconditioned with 15% A and 85% C for 15 min. The pulsed amperometric detector consisted of gold as the working electrode and Ag/AgCl as the reference one. Data were collected using Chromeleon software (20) and exported to Excel for further analysis.

Data acquisition and statistical analysis

Images of the HPTLC chromatograms were processed with the ImageJ software (21). Before the signal acquisition and analysis, each image was split-filtered through red, green and blue channel. Denoising of filtered images was done using median filter with width of 2 pixels. Bandpass filter, which filters large structures down to 40 pixels and small structures up to 3 pixels, was used for removing the differences of the background intensity between images. Chromatographic signals were extracted from images in form of pixel intensity line profiles.

Pretreatment of chromatographic signals (HPTLC, GC-MS and HPAEC-PAD) consisted of warping, normalization and centring, and was performed by algorithms implemented in the PLS ToolBox software for MATLAB (22). Warping procedure was performed due to the spatial differences of the band locations. Two variable alignment algorithms were applied: correlation optimized warping (COW) and peak alignment. Normalization step was used to adjust the values measured on different scales to a notionally common scale. Weighted normalization of the chromatograms was done by standard normal variate (SNV) transformation. The data were additionally pre-processed by mean or median centring, which is the preferred option when the classification of samples is based on variables that are all measured in the same unit.

Principal component analysis (PCA) and partial least square-discriminant analysis (PLS-DA) were performed by PLS ToolBox (22). PCA was carried out by using a singular value decomposition algorithm and a 0.95 confidence level for Q and T2 Hotelling limits for outliers. The PLS-DA was performed using the SIMPLS algorithm without forcing orthogonal conditions to the model in order to condense Y-block variance into first latent variables. The optimal complexity and reliability of the resulting models was ensured by fourfold cross-validation (CV) which employed Venetian blinds resampling strategy. In order to assess predictive performance of PLS-DA models, i.e. to determine the degree of accuracy and sensitivity, the entire data set of 45 samples was divided into two subsets, a model building and optimization set consisting of 33 randomly selected objects, and testing set composed of 11 objects. The same subsets of berry samples were used for all chromatographic methods. The quality of the models was monitored with the following parameters: (i) R²_cal, the cumulative sum of squares of the Ys explained by all extracted components, R²_CV, the cumulative fraction of the total variation of the Ys that can be predicted by all extracted components, and R²_PRED, the cumulative fraction of the total variation of the Ys that can be predicted by test components, and (ii) root mean square errors of calibration (RMSEC), root mean square errors of cross-validation (RMSECV) and root mean square errors of prediction (RMSEP).

Primary metabolite profiles

Lipid profiling

Complete lipid characterization of analysed berry seed cultivars was performed by HPTLC fingerprinting of the main lipid classes (polar lipids, monoacylglycerols, sterols, diacylglycerols, free fatty acids, triacylglycerols and sterol esters, retention factor R_F≈0.11, 0.22‒0.32, 0.42, 0.48, 0.62, 0.75 and 0.92, respectively,Fig. 1) in order to screen the contribution of those involved in fundamental metabolic processes (free fatty acids and acylglycerols). Due to the complex composition of berry seed extracts, optimization of HPTLC conditions was done regarding the resolution of the low, medium and high polarity lipid compounds. Optimal results with vast number of sharp bands, maximum separation efficiency and lowest background noise for all analysed species and all lipid classes were obtained using multiple one-dimensional chromatographic systems (listed in the experimental part). The zones on the chromatograms were identified based on the knowledge of the polarity of each lipid class, as well as by comparison with the previously published chromatograms (23).

Fig. 1 High-performance thin-layer chromatograms of lipids in berry seeds of: a) strawberry, b) blackberry (1–3) and black currant (4–15), and c) raspberry (1–4), gooseberry (5 and 6), chokeberry (7), goji berry (8), cape gooseberry (9) and blueberry (10–14) cultivars

Visual examination of the HPTLC chromatograms (Fig. 1) and the corresponding plots of chromatographic signals (Fig. S2) indicates the uniformity of lipid profiles within each fruit species, but also reveals differences in the chemical composition depending on the botanical origin. The same lipid classes are present in all samples, but different compounds are present in various amounts within each class. Therefore, the presence or absence of zones in a particular group of samples was noted, as well as the variability of the signal intensity of the compounds at the same R_F values​. The chromatograms of the strawberry samples showed high intensity zones or peaks at R_F=0.75 that correspond to triglycerides and well separated medium intensity zones at R_F=0.62 that correspond to free fatty acids (Fig. 1a, samples 1–16). Black currant samples (Fig. 1b, samples 4–15, pixel intensity 1846) and raspberries (Fig. 1c, samples 1–4, pixel intensity 2008) have more intense and characteristic zones at lower R_F values between 0.1 and 0.3 which are considered as polar lipids. Raspberry seed samples (Fig. 1c, samples 1-4) have also intense zones that correspond to free fatty acids (R_F=0.62, pixel intensity 6246). The zones at the highest R_F values (0.92) that correspond to sterol esters are the most intense in the goji and cape gooseberry samples (Fig. 1c, samples 8 and 9, intensity of pixels 5660 and 6240, respectively), while the same zones are hardly noticeable in the blueberry samples (Fig. 1c, samples 10–14, 973 pixels). The chromatogram obtained for chokeberry seed extract (Fig. 1c, band 7) has more zones than the rest of the analysed samples. The chromatogram profiles of goji and cape gooseberry lipids (Fig. 1c, bands 8 and 9) are similar to raspberry samples (Fig. 1c, bands 1–4), while the gooseberry (Fig. 1c, bands 5 and 6) are similar to blueberries (Fig. 1c, bands 10–14).

Fatty acid profiling

Qualitative description of differences between fatty acid profiles of plant species is based on the presence and intensity of the fatty acid signals (Fig. 2 andTable S1). Fatty acid profiles within the same fruit species showed specific pattern, but different compared to other species. The most abundant fatty acid in the analysed berry seed oils was linoleic acid (t_R=12.07 min, 44.36‒80.09%), followed by oleic (t_R=11.48 min, 12.71‒26.38%) and α-linolenic (t_R=12.90 min, 0.53‒31.42%) acids (Table S1). Oomah and Ladet (24) previously reported that 96% of the total fatty acid composition of raspberry seed oil consists of these three fatty acids. Linoleic acid was the most prevalent in cape gooseberry and goji berry samples, oleic acid in gooseberry, blackberry, chokeberry and blueberry samples, and α-linolenic in strawberry and blueberry seeds (Fig. 2a). Literature data (9,25) confirmed low content of α-linolenic acid in goji (1.1%) and chokeberry (0.34%) fruit, similar to samples of goji, gooseberry, chokeberry and cape gooseberry (0.53‒1.41%) analysed in this study (Fig. 2b). Parker et al. (26) showed that cranberry seed oil contains about 22% α-linolenic acid, indicating that the fruit seeds may be a potential dietary source of this acid.

Fig. 2 Fatty acid methyl ester chromatograms of analyzed berry species: a) strawberry, black currant, blueberry, blackberry and raspberry cultivars, and b) cape gooseberry, goji berry, chokeberry and gooseberry cultivars

Higher abundance of γ-linolenic acid (t_R=12.46 min) was noticed in black currant and gooseberry samples (14.63 and 12.26%, respectively) (Table S1) confirming that seed oil from Grossulariaceae (genus Ribes) family are rich in γ-linolenic acid (from 18.3% in gooseberry to 30.5% in black currant) (27). On the contrary, it was not found in strawberry, raspberry and blackberry seed oil, a typical feature of the Rosaceae family. This is an indicator of the differences in fatty acid desaturase profiles of Rosaceae and Grossulariaceae (Ribes) species within the same superorder Rosanae (11).

Only in the samples of black currant cis-11-eicosenoic acid was found (t_R=14.19 min, 2.46%). Heneicosapentaenoic acid (21:5, n−3) (t_R=13.34 min) is an unsaturated fatty acid that has been identified in gooseberry and black currant samples, by comparison of the resulting mass spectra to the NIST database (19), as well as 6-octadecanoic acid (t_R=11.87 min) detected only in black currant.

All analysed berry seeds contained low amounts of saturated fatty acids, namely palmitic (t_R=9.30 min, 3.07‒8.56%) and stearic (t_R=11.23 min, 1.14‒11.47%) acids (Table S1). Palmitic acid was the major saturated fatty acid detected in gooseberry seeds and pulp by Ramadan and Mörsel (28). Arachidic acid (t_R=13.86 min) showed weak signals present only in strawberry, blackberry and goji berry seed samples (0.50‒0.77%) (Table S1).

Primary metabolites in chemotaxonomic classification

In order to explore a multivariate space of chromatographic profiles and look for possible patterns between samples and chromatographic signals, PCA was applied on each data set separately. The matrix consisted of 45 objects (tested samples), 1200 variables representing the intensities of pixels along the 1200 length lines obtained by digitization of lipid chromatograms (using green channel of profiles as the most informative one), 2384 variables representing intensities of analytical signals on fatty acid chromatograms, and 1797 variables representing intensities of analytical signals on sugar chromatograms.

All PCA models are graphically presented as mutual projections of factor scores and their loadings for the first two principal components (PCs) (Fig. 3).

Fig. 3 Principal component analysis: a) score plot, b and c) loading plots for lipid profile, d) score plot, e and f) loading plots for fatty acid profile, g) score plot, and h and i) loading plots for sugar profile of seeds from berry fruit cultivars. Number of 1200 variables represent the intensities of pixels along the 1200 length lines obtained by digitization of lipid chromatograms (using green channel of profiles as the most informative one), 2384 variables represent intensities of analytical signals on fatty acid chromatograms, and 1797 variables represent intensities of analytical signals on sugar chromatograms. Intensities of signals were transformed in RF and tR scale in order to determine the most influential variables

Comparative study of chromatographic approaches for metabolite profiling and sample classification

Comparison of the suitability of the studied chromatographic methods for chemotaxonomic classification of seeds from cultivated berry fruits was analysed based on the statistical parameters of PLS-DA classification models. Given the limited number of botanical species with satisfactory number of samples, only black currant, strawberry, blueberry and raspberry seeds were evaluated.

All classification models demonstrate good statistical performance (Table 2,Fig. S3,Fig. S4,Fig. S5), with the coefficient of determination ranging from R²=0.4734, for the blueberry seeds based on HPAEC sugar profiles, to R²=0.9940, for the raspberry seeds and GC-MS profiles of fatty acids. Although the best chemotaxonomic classification was achieved based on the GC-MS data, with average R²=0.9058 and 0% misclassification error (Table 2), it is closely followed by PLS-DA models based on HPTLC profiles of lipids (with R² in the range of 0.7171–0.958). However, mathematical model based on HPAEC profiles of sugar components was slightly worse (R² ranging from 0.4734 to 0.8801, with arithmetic mean 0.7392). We believe that this is not a result of shortcomings of HAEC-PAD technique, but probably a result of huge variations of sugar content, even within each particular cultivar, which further prevented the best differentiation of classes due to the overlapping ranges.

Although GC-MS provides the best overall performance of PLS-DA models, these models are slightly better for the separation of strawberry and raspberry seeds from the rest (averaged calibration, cross-validation and prediction R² values for these two classes are 0.9440, and 0.9726, respectively), while in the case of black currant and blueberry seeds, the HPTLC profiles result in PLS-DA models with slightly better separation/statistical parameters (averaged R² values are 0.9011 and 0.8350, respectively).

Such good results of GC-MS, in terms of providing the most suitable chromatographic fingerprints for PLS-DA models, go along with the fact that MS coupled to HPLC and GC is the most widely used analytical tool for profiling highly complex mixtures of primary metabolites. However, GC-MS analysis requires extensive sample preparation and derivatization procedures, while the retention of primary metabolites in the mostly used reversed-phase LC mode is very low (2). Despite continuous development of these sophisticated techniques, such multianalyte methods require fine tuning of a multitude of parameters, they need highly trained analysts and suffer from high costs of equipment and reagents. Therefore, the HPTLC, which gives equally suitable classification models, can be a method of choice for the first stage of analysis in generation of data for biomarker discovery. The main advantages of HPTLC method is its simplicity, flexibility, accessibility and low cost. With development of high-performance adsorbent layers and sophisticated instrumentation for sample application, chromatogram development, derivatization and chromatogram evaluation, its employment in fingerprint analysis is significant, as this study also confirms. Additionally, the main advantage of HPTLC method in the analysis of primary metabolites is a possibility of application of wide range of solvents of different polarity combined in multi-one-dimensional chromatography system.

Having in mind that all PLS-DA models resulted in statistically sound chemotaxonomic classification of berry seeds, it is clear that an approach based on combining advanced data analysis with non-targeted chromatographic profiling was able to overcome the principal challenges associated with untargeted metabolite profiling (30). First drawback of the untargeted strategy, which comprises complex protocols and time required to reduce the extensive datasets generated from chromatographic profiles into smaller set of manageable signals, was overcome by the application of simple dimensionality reduction to few latent variables (PCA and PLS-DA). The difficulties in identifying and characterizing unknown metabolites, as a second drawback of the untargeted strategy, were solved by using the entire chromatographic signal as a unique chemical fingerprint of the extract, without identification of individual peaks. The third drawback of the untargeted strategy is the bias towards detection of highly abundant metabolites. Reducing the dominance of highly abundant signals over those present in smaller quantity has been achieved by adequately processing the data prior to multivariate analysis.