JBCS

INTRODUCTION

Hibiscus sabdariffa is popularly known as 'Roselle' or 'Flor de Jamaica' around the world. The calyxes of this plant have been employed in the preparation of medicinal teas in order to treat several health disorders.^1-3 H. sabdariffa calyxes have been reported as an important anthocyanin-containing plant material.^3-6 Other chemical constituents are included such as phenolic acids and flavonol glycosides.^3,6-8 Anthocyanin and phenolic contents are directly related to the maturity degree⁹ as well as to the seasonal changes.⁴ These compounds confer a markedly high antioxidant power to the decoctions of their calyxes.¹⁰ Spite of the biological activity and the increasing demand of the H. sabdariffa calyxes worldwide, there is lack of information on chemical variability and its effect on biological effectiveness. In Colombia, H. sabdariffa calyxes are employed to prepare both cold and hot beverages attributing them significant medicinal benefits. Due to the great demand of Roselle in Colombia, their cultivation has rapidly increased and extended to several zones of the country. However, there are not quality control procedures. Even the variability within cultivars around the country has not been analyzed.

A practical method to ensure consistency of the herbal medicines can be accomplished by employing the concept of phytoequivalence, which involves fingerprint analysis. In other words, with the help of the fingerprints, the authentication and identification of the herbal medicines can be obtained. Moreover, both evident and minor differences between the fingerprints in combination with multivariate analysis could be established. The application of chemometrics to quality assurance of herbal medicines has increased in recent years and then reviewed.¹¹ Several reports have been published regarding to the usefulness of chemometrics in natural product and food chemistry.^12-16 Among these, HPLC-DAD has been demonstrated as a useful tool to accomplish fingerprinting.

Therefore, the present work shows the chemical profiling of Colombian calyxes of H. sabdariffa by means of HPLC-DAD. Unsupervised and supervised chemometrics based on PCA, PARAFAC and OPLS-DA analyses are applied to two different data sets of chromatographic profiles in order to demonstrate similarities and/or differences throughout them. So far, there is not previously report focusing on the chemical characterization and differentiation of H. sabdariffa calyx extracts using chromatographic data and chemometrics. Moreover, the present study might be taken as the first step in a quality control strategy to Colombian samples.

EXPERIMENTAL

Reagents

Gallic acid and cyanidin 3-O-glycoside were purchased from Sigma-Aldrich Chemical Co. Folin-Ciocalteu reagent, BHT (3,5-di-tert-butyl-4-hydroxytoluene), formic acid, and ethanol were acquired from Merck Chemical Co.

Plant material and extraction

H. sabdariffa calyxes (n = 22) were randomly purchased on local markets and healthfood stores at Bogotá, Colombia in August 2012. All samples corresponded to Sudan cultivar-derived reddish calyxes cultivated in Colombia. Samples were obtained as dried raw material. Each sample was powdered and stored at 4 ºC until analyses.

A portion (0.5 g) of each sample of H. sabdariffa calyxes was independently mixed with 0.5% aqueous formic acid (7.0 mL). These mixtures were kept at room temperature in the dark during 3 h, and then filtered, washed and diluted to the mark with the same solvent into a 10.00 mL volumetric flask, and mixed thoroughly.

Total Phenolic content

Folin-Ciocalteu reagent assay was used to determine total phenolic content (TPC).¹⁷ Briefly, sample solution was diluted with water (1:25). Dilution (400 µL) was mixed with freshly, diluted (1:10) Folin-Ciocalteu reagent (800 µL), was then kept during 3 min and 7.5% Na₂CO₃ (3.0 mL) was added. After 2 h of incubation at room temperature in the dark, absorbance at 765 nm was measured. Each measurement was performed in triplicate. TPC was calculated as mg gallic acid equivalent per g dried sample (mg GAE/g DS) by employing a calibration curve.

Total Anthocyanin content

The pH differential method was employed to measure total anthocyanin content (TAC).¹⁸ Briefly, test dilutions (3.0 mL) were separately prepared mixing fresh sample solution (100 µL) with buffer pH 1.0 and pH 4.5 (1:30), and the absorbance at 520 nm and 700 nm was then measured. Each measurement was performed in triplicate. TAC was calculated as mg cyanidin 3-O-glycoside equivalent per g dried sample (mg C3GE/g DS) as follows:

where A = (A_520nm - A_700nm)_{pH 1.0} - (A_520nm - A_700nm)_{pH 4.5}; MW as molecular weight for cyanidin 3-O-glycoside (C3G) (449.2 g/mol); DF = dilution factor; V = total volume of sample solution after extraction, in L; 10³ = factor for conversion from g to mg; ε (molar extinction coefficient) = 26,900 L mol^-1 cm^-1, for C3G; l = path length, in cm; and SW = sample weight used for extraction, in g.

Chromatographic Analyses

HPLC-UV-DAD analysis

HPLC profiles were recorded on a Shimadzu UFLC Prominence system. A Shimadzu Premier^® C₁₈ column (150 x 4.6 mm; 5 µm) was used (0.6 mL/min flow rate; 30 ºC column oven temperature). 1% formic acid in acetonitrile and 1% formic acid in water were employed as mobile phases A and B, respectively, in gradient elution as follows: 0 min, 0% A; 7 min, 20% A; 10 min, 25% A; 13 min, 0% A; 20 min, 0% A. Data were acquired each 1.2 nm between 200 - 900 nm. The monitoring wavelength was selected to 520 nm. Extracts were diluted in 1:20 ratio and then 50 µL were injected.

HPLC-ESI-MS analysis

HPLC- MS profiles were recorded on a Liquid Chromatograph Mass Spectrometer LCMS-2020 (Shimadzu Corp, Japan) system equipped with a single quadrupole analyzer and an electrospray ion source (ESI). A Synergi RP C₁₈ (Phenomenex) (150 x 4.6 mm; 4 µm) was employed (1.0 mL/min flow rate; 30ºC column oven temperature). 1% formic acid in acetonitrile and 1% formic acid in water were used as mobile phases A and B, respectively, in gradient elution as follows: 0 min, 0% A; 10 min, 10% A; 11 min, 10% A; 15 min, 30% A; 17 min, 0% A. Extracts were diluted in 1:20 ratio and then 5 µL were injected. ESI was simultaneously operated in positive and negative ion mode (scan 100 - 800 m/z; 250 ºC CDL temperature; 1.2 kV detector voltage; 1.5 L min^-1 nebulizing gas flow rate; 9.0 L min^-1 drying gas flow rate).

Data Analysis

Chromatographic data were divided in two different data sets. The first one, absorbance at 520 nm as function of time was exported as 2D ASCII files for each sample and a matrix of 1876 x 22 points was then built. The second one, response surface obtained for each sample by HPLC-DAD was exported as 2D ASCII files and a three-way array was then built (951 x 491 x 22). All data analysis were performed using MATLAB R2013a (The Mathworks, Inc.) on a PC (Core i5 processor; 8 GB RAM; on Microsoft Windows 7).

Two-way array raw data were submitted to chromatographic peak alignment as first step of pretreatment prior to the statistical analysis. Correlation optimized warping (COW) algorithm was employed to the alignment process.¹⁹ COW algorithm from Warping Toolbox supplied by The Quality and Technology Website was used to this process.²⁰ Normalization and Pareto Scaling processes²¹ were then applied on the data set by means of algorithms coded by authors.

Principal Component Analysis (PCA)²² was implemented on the two-way array after corresponding above-mentioned pre-treatment. Results were presented as score plots. The PCA algorithm included into Eigenvector PLS Toolbox for Matlab was employed.²³

Parallel factor analysis (PARAFAC)²⁴ was individually performed on each sample in the three-way array by means of the algorithm supplied by The Quality and Technology Website.²⁰ Prior to analysis, core consistency diagnostic was implemented in order to provide an appropriate number of components. Three components were demonstrated to be able to describe the PARAFAC model according to the core consistency diagnostic tool and the corresponding residual sum of squares. No restrictions on PARAFAC model were taken into account in the present research. Resultant first loadings were used to build a final two-way array (951 x 22) which was then submitted to PCA as described above. In addition, an OPLS-DA model was constructed for classification of the samples using the PARAFAC-derived loadings. TPC and TAC ranges were used to define the classes (class 1: 12.5-25.0; class 2: 25.0-30.3 for TPC values; class 1: 7.0-12.0; class 2: 0.8-7.0 for TAC values). A single-Y OPLS model using TAC as Y variable was also constructed from the PARAFAC-derived loadings. OPLS-DA and OPLS models were created using SIMCA 14 software (Umetrics Inc.).

RESULTS AND DISCUSSION

Chemical analysis

Chemical characterization of twenty-two Colombian samples of H. sabdariffa was developed by measuring total phenolic and total anthocyanin contents (Table 1). Wide range on TPC and TAC values can be observed indicating high composition variation by the Colombian calyxes. Samples 13, 15, 16 and 10 resulted to exhibit high contents while samples 17, 21, 6 and 19 were found to have the lowest values. However, clear grouping was not detected by means of Tukey test at 95% level. Samples 1 and 13 (corresponded to the same crop and same commercial batch) exhibited closely related TPC values (the same group according to the Tukey test), while samples 11 and 14 showed significant different values although these are also from the same commercial batch. TAC values were more disperse than TPC for above-mentioned samples demonstrating marked variability even those pertaining to the same production batch of H. sabdariffa calyxes. Similarly, samples 2 and 22 showed statistically significant differences on TPC and TAC in spite of belonging to the same crop but different batches. Nevertheless, TPC values were in agreement to those reported to calyxes of H. sabdariffa cultivated in other parts of the world.^4,5,9,10 Positive correlation between TPC and TAC values was exhibited by Pearson correlation coefficient (0.792 at 95% level). This result was expected because anthocyanins are a group of phenolic compounds.

Chromatographic profiling

HPLC-UV-DAD profiles were monitored at 520 nm in order to detecwt anthocyanin-related compounds. This analysis was useful to demonstrate that there are two main anthocyanins in Colombian H. sabdariffa calyxes (Figure 1). These findings are in agreement with those previously reported to H. sabdariffa calyxes around the world.^4,7,8

Figure 1. HPLC profiles (520 nm) of the extracts of Colombian Hibiscus sabdariffa calyxes. Individual raw data (a), superimposed raw data (b) and pre-processed data by using COW, normalization, and Pareto scaling (c)

Statistical analysis

Data pre-processing

In order to improve the quality of the data as well as decrease their inconsistency and noise, three pre-processing algorithms were applied on the raw data. The alignment was accomplished by COW¹⁹ and the reference chromatographic profile was arbitrarily assigned as sample 1. The adequate parameters such as slack variable and segment length were achieved by means of the algorithm package supplied by The Quality and Technology Website.²⁰ Figure 1a shows the aligned, normalized chromatographic profiles of the twenty-two Colombian Roselle extracts. Figure 1b and 1c (before and after peak alignment, respectively) indicated a perfect correction of the retention time shifting into LC-derived profiles after warping processing by COW. Additionally, Pareto scaling was applied to eliminate the unfavorable effects between LC profiles (e.g., variation due to injection of uneven concentrations).

Multivariate analysis to the two-way array (single-wavelength data)

PCA was applied on the pre-processed chromatographic data (two-way array) and the differences were then established. The PCA-derived score plot (Figure 2a) shows that samples 17 and 19 were located distant from others demonstrating significant difference of these samples from the rest. However, the above mentioned samples were located in remote positions on the score plot so that these could be defined as outliers in the used model. Therefore, Hotelling's T2 and Q residuals were analyzed. It let us to designate the outliers. Samples 12, 17 and 19 resulted as outliers and they were then discarded for further analyses. Figure 2 shows the PCA-derived score plots (PC1-PC2) of the normalized, aligned data before (a) and after (b) removal of the outliers. The two first PC on the constructed model were able to explain 85.78% and 83.36%, respectively, total variance of the variable set. The existence of these outliers could be a consequence of alterations in the plant conditions involving soil type, ecological environment, weather, collecting season, among others. Once the outliers were discarded, samples were remained in similar distribution on the PC score plane but relative grouping from Hierarchical Clustering Analysis (HCA) was changed. PCA score plots showed explicit differentiation between some samples (at least five groups). Thus, samples 12, 13, 15, 16, 20 and 10 resulted spatially far from samples 6, 2, 7 and 9 (together with 17 and 19), demonstrating a direct relationship with their TPC values. The first group exhibited high TPC values while the second one had low TPC values. TPC for groups A, B and C were not significantly different (at 95% level).

Figure 2. PC1-PC2 score plots of the fingerprints of the 22 samples of Hibiscus sabdariffa (a) and without outliers (b)

Multivariate analysis to the three-way array (multi-wavelength data)

In order to obtain larger information regarding the chemical composition variability to be responsible for the above-mentioned results, HPLC-UV-DAD surfaces were employed. As the first step, PARAFAC model was applied. Three components were required to explain all variance according to the residual sum of squares and the core consistency diagnostics. The first loading of the model retained chromatographic profile information, generating an extracted chromatogram from all range of analyzed wavelength and the data mining was thus achieved (Figure 3a). Observed peaks in Figure 3a represented compounds with significant UV-Vis absorption in the 250-800 nm range. As expected, peaks corresponding to the two main anthocyanin-type compounds resulted to be present. A new matrix with these data was assembled and the PCA was then performed. Two first PC on the constructed model were able to explain 54.83% and 17.91%, respectively, total variance of the variable set. Pareto scaling was also carried out before analysis. Moreover, peak alignment was evaluated; however the algorithm was demonstrated to be inadequate and it was not therefore included.

Figure 3. Extracted chromatograms from PARAFAC for representative samples of Hibiscus sabdariffa (a). PC1-PC2 score plot of the extracted chromatogram from PARAFAC loadings to 22 samples of Hibiscus sabdariffa (b). Bold numbers represent compounds responsible of the variability among data; normal numbers represent samples

Figure 3a shows the PARAFAC-derived first loading to samples 13, 8, 18 and 17, which correspond to representative high TPC, medium TPC and low TPC values-possessing samples, respectively (Table 1). Clear differences can be found in that figure regarding to the presence/absence of certain compounds as well as in the relative proportion of some of them. Score plots appear in Figure 3b. Five groups were also defined here from HCA (I-V denoted by colored ellipses). PCA from three-way data exhibited a similar number of clusters than that by two-way data. Significantly differences to samples 17 and 19 from the rest were again demonstrated (Figure 3b). In the same way, samples 10, 20, 16 and 13 were grouped in the PCA score plot (group I) being these samples those with higher TPC values. Therefore, PCA clustering is evidently supported on the chemical composition of the analyzed samples, which might be useful in further studies. Moreover, the presence of several groups with high to medium TPC let us to confirm that chemical composition of analyzed samples were very variable regarding to chemical constituents possessing UV-Vis absorption.

UV-Vis and MS on-line acquired spectra of these components were analyzed to get structural information regarding the compounds responsible of the chemical variability of the Colombian Roselle samples (denoted by bold numbers). The structural assignment of the detected compounds was proposed (shown in Table 2) by comparison of the obtained UV and MS spectra with previously reported chemical components of H. sabdariffa calyxes.^7,8 The most variable components within analyzed samples were established as anthocyanins and chlorogenic acid-related derivatives. Our analyses on the PCA score plot and the extracted chromatograms from PARAFAC model (Figure 3) let us to infer that samples with low TPC values (no grouped in PCA) are rich in compounds 1, 3 and 6. All of them were not anthocyanin-related substances (Table 2). Meanwhile, anthocyanin-type compounds (4 and 5) were present in higher quantities in samples with higher TPC (group I).

Additionally, PCA-derived first loading were compared with the input data of the three-way array (illustrated in Figure 4a). The explained variance was supported on the four main constituents along with at least the six minor ones. Variability of the analyzed profiles resulted to be affected by compounds 4 and 5 towards the negative side of the PC1-axis while compounds 1 and 6 were related towards the positive one (blue line in Figure 4a). Influence of compounds 1-10 on the PC's was analyzed and presented as a loading bar plot (Figure 4b). This plot confirmed that the first principal component (PC1) was highly, positively influenced by the compounds 1 and 6 while it was negatively influenced by peaks 4 and 5. Similarly, discrimination across the second principal component (PC2) was due to higher levels of compounds 1 and 5, influencing the PC2 in the negative direction. Compound 3 influenced the PC2 across the positive direction. Further analysis of the data reveals that samples 6, 17 and 19 contain higher levels of 1, 3, 5 and 6. Samples 17 and 19 contain lower levels of 2, 4, 7-10. Additionally, samples 13 and 10 contain higher levels of 3, 4 and 5 while 6-10 appear in lower levels. Compounds 1 and 2 were presented in very low levels or even absent in samples 13 and 10, which influenced the PC1 in positive direction.

Figure 4. Effect of the compounds 1-10 on the PC' first loading (blue line) for the sample 6 (green line) (a). Loading bar plot from PCA for extracted chromatograms (b)

Anthocyanins 4 and 5 resulted to be present in very low level in the samples with lower TPC and TAC values as mentioned above. This effect was complemented with increased quantities of 1, 3 and 6 (all weighting for the positive direction of PC1). Sample 17 also presented a very low level of 5 and high level of 6. This fact was a decisive discriminating factor along PC2 as above mentioned and therefore it was responsible for the differentiation of samples 5, 2 and 6 from 17. Finally, samples 10, 20, 13 and 16 possess higher levels of 4 than that for the rest of the samples being compound 4 the main responsible of the explained variability between them. The PC2 is strongly, negatively influenced by compounds 1 and 5 meanwhile it is positively influenced by compound 3 (Figure 4b). Therefore, high TPC and TAC measurements are directly related to high contents of compounds 4 and 5, meanwhile low TPC and TAC contents can be related to high contents of compounds 1 and 6. Moreover, intra-group differentiation (i.e., between samples 20 and 1) can be due to differences on the concentration/level of compounds 1, 3 and 5. Thus, group I is completely different from group V by compounds 1, 3, 4, 5 and 6. For their part, group II and III are mainly differentiated by compounds 3 and 5.

An explanation to the inverse relationship between level/concentration of 4 and 6 in the extracts may be inferred from their general biosynthetic pathways. In this way, the biosynthetic pathway of 4 is related to that of 6 by a simple C₆C₃ scaffold: p-coumaric acid. From this intermediate, caffeic acid is produced and subsequently it is coupled to quinic acid affording 6.²⁵ The p-coumaric acid is a fundamental intermediate to produce aromatic polyketides which are later converted in flavonoids including anthocyanins such as 4.²⁵ Thereof, compounds 4 and 6 can be inversely related due to two discriminating metabolic pathways with p-coumaric acid as a pivotal common intermediate. Nevertheless, it is not clear how is preferred one way or what is the discriminating factor for it. Similarly, differentiation across PC2 is due to variable quantities of 4 and 5. These compounds are commonly produced simultaneously, although regulation or selective metabolite production could be limited from naringenin and dihydrokaempferol as key intermediate.^26,27

Apparent discrimination of samples according to PCA from PARAFAC extracted information let us to infer that supervised methods can be suitable for characterization of the chemical variability of H. sabdariffa. In this way, three-way array-derived data was then submitted to Orthogonal Partial Least Squares - Discriminant Analysis (OPLS-DA). Such analysis was accomplished in terms of TPC and TAC values as classification variables as described in experimental section. The resulting score plots are exposed in Figure 5a and 5b, respectively. TPC demonstrated to be inappropriate classification variable for the tested samples. In other words, H. sabdariffa samples were not discriminated as function of TPC classes. In contrast, TAC was found to be a more suitable classification variable although incomplete discrimination was observed for samples 18, 11, 12, 14, 1, 8, 3, and 9. An S-plot derived from discrimination across TAC was obtained in OPLS-DA (Figure 5c) which can be related with the corresponding loading line plot (Figure 5d). Peaks responsible for discrimination in OPLS-DA are shown in red circles on Figure 5d. Samples arranged along the positive first component corresponded to those with higher TAC values while samples along the negative first component resulted to be those with lower values. Compounds 4 and 5 were demonstrated as responsible for the discrimination towards the positive direction of the first OPLS-DA component. Therefore, the presence of the two antocyanins resulted in a key factor for discrimination of Colombian H. sabdariffa samples such as it was inferred from above mentioned PCA. Similarly, the presence of the compounds 1, 3 and 6 was a decisive factor for discrimination of samples towards the negative side of the first component characterized by the lowest TAC values. OPLS-DA results were consistent with previously PCA findings. In fact, compounds 1, 3 and 6 are not flavonoids but were found in significant proportion in several samples. On the other hand, genic expression towards production of 4 and 5 was apparently inversely coupled to that for 1, 3 and 6 as above described.

Figure 5. Supervised analyses to the three-way array (multi-wavelength data). OPLS-DA score plot using TPC classes (A); OPLS-DA score plot using TAC classes (B); OPLS-DA-derived S-plot using TAC classes (C); OPLS-DA-derived S-line using TAC classes (D); Single-Y OPLS score plot using TAC as Y variable (E); OPLS-derived Observed vs. Predicted plot using TAC as selected response (F)

Finally, single-Y OPLS model was also constructed with the three-way-derived data in order to demonstrate the usefulness of chemometric tools in differentiation of Colombian H. sabdariffa samples. The corresponding score plot is shown in Figure 5e. Samples 20, 13, 16 and 15 were clearly discriminated from samples 7, 5, 2 and 4 according to TAC values. It can be seen that only highest TAC values bring to complete discrimination. Medium to low TAC values were measured for samples with very close chromatographic profiles. Nevertheless, a very high positive correlation was observed between predicted and observed TAC values from OPLS (Figure 5f). Only samples 3, 8 and 9 exhibited a deviated behavior from the model.

CONCLUSIONS

In this study, HPLC fingerprints of Colombian H. sabdariffa calyxes were developed. The use of total chromatographic raw data matrix afforded a holistic strategy for examining the possible differences between samples in a particular production batch and even among different crops around Colombia when it is coupled to using PARAFAC analysis. The present strategy can result improved in comparison to the existing authentication protocols based on few chemical markers commonly ignoring minor compounds. This work illustrates how the use of chemometrics methods combined with hyphenated techniques offers an excellent analysis way for complex mixtures and enables exploratory comparisons. The product composition distinctions in Colombian H. sabdariffa calyxes may be translated in significant quality differences of individual products. Therefore, the importance of easy and effective quality assurance protocols for H. sabdariffa calyxes from Colombia was thus evidenced. Furthermore, the present protocol is sound and indeed the method has the advantage of being simple and cheap, hence virtually affordable by any lab for the analysis and authentication of H. sabdariffa calyxes.

ACKNOWLEDGEMENTS

Authors thank MU Nueva Granada by financial support. The present work is a product derived by the Project IMP-CIAS-1567 financed by Vicerrectoría de Investigaciones at UMNG - Validity 2014.

REFERENCES

1. Amin, I.; Emmy, H. K. I.; Halimatul, S. M. N.; Food 2008, 2, 1.

2. Hopkins, A. L.; Lamm, M. G.; Funk, J. L.; Ritenbaugh, C.; Fitoterapia 2013, 85, 84. DOI: http://dx.doi.org/10.1016/j.fitote.2013.01.003 PMID: 23333908

3. Borrás-Linares, I.; Fernández-Arroyo, S.; Arráez-Roman, D.; Palmeros-Suárez, P. A.; Del Val-Díaz, R.; Andrade-Gonzáles, I.; Fernández-Gutiérrez, A.; Gómez-Leyva J. F.; Segura-Carretero, A.; Ind. Crops Prod. 2015, 69, 385. DOI: http://dx.doi.org/10.1016/j.indcrop.2015.02.053

4. Juliani, H. R.; Welch, C. R.; Wu, Q.; Diouf, B.; Malainy, D.; Simon, J. E.; J. Food Sci. 2009, 74, S113. DOI: http://dx.doi.org/10.1111/j.1750-3841.2009.01076.x PMID: 19323768

5. Tsai, P. J.; McIntosh, J.; Pearce, P.; Camden, B.; Jordan, B. R.; Food Res. Int. 2002, 35, 351. DOI: http://dx.doi.org/10.1016/S0963-9969(01)00129-6

6. Sáyago-Ayerdi, S. G.; Arranz, S.; Serrano, J.; Goñi, I.; J. Agric. Food Chem. 2007, 55, 7886. DOI: http://dx.doi.org/10.1021/jf070485b PMID: 17705439

7. Fernández-Arroyo, S.; Rodríguez-Medina, I. C.; Beltrán-Debón, R.; Pasini, F.; Joven, J.; Micol, V.; Segura-Carretero, A.; Fernández-Gutiérrez, A.; Food Res. Int. 2011, 44, 1490. DOI: http://dx.doi.org/10.1016/j.foodres.2011.03.040

8. Rodríguez-Medina, I. C.; Beltrán-Debón, R.; Micol, V.; Alonso-Villaverde, C.; Joven, J.; Menéndez, J. A.; Segura-Carretero, A.; Fernández-Gutiérrez, A.; J. Sep. Sci. 2009, 32, 3441. DOI: http://dx.doi.org/10.1002/jssc.200900298 PMID: 19750503

9. Christian, K. R.; Jackson, J. C.; J. Food Compos. Anal. 2009, 22, 663. DOI: http://dx.doi.org/10.1016/j.jfca.2009.05.007

10. Awe, F. B.; Fagbemi, T. N.; Ifesan, B. O.; Badejo, A. A.; Food Res. Int. 2013, 52, 490. DOI: http://dx.doi.org/10.1016/j.foodres.2013.01.021

11. Gad, H. A.; Ahmady, El- S. H.; Abou-Shoer, M. I.; Al-Azizi, M. M.; Phytochem. Anal. 2012, 24, 1. DOI: http://dx.doi.org/10.1002/pca.2378 PMID: 22678654

12. Liang, X.; Wu, H.; Su, W.; Food Anal. Methods 2014, 7, 197. DOI: http://dx.doi.org/10.1007/s12161-013-9618-4

13. Liu, H.; Lai, H.; Jia, X.; Liu, J.; Zhang, Z.; Qi, Y.; Zhang, J.; Song, J.; Wu, C.; Zhang, B.; Xiao, P.; Phytomedicine 2013, 20, 1135. DOI: http://dx.doi.org/10.1016/j.phymed.2013.05.001 PMID: 23880331

14. Farag, M. A.; El-Ahmady, S. H.; Elian, F. S.; Wessjohann, L. A.; Phytochemistry 2013, 95, 177. DOI: http://dx.doi.org/10.1016/j.phytochem.2013.07.003 PMID: 23902683

15. Acevska, J.; Stefkov, G.; Petkovska, R.; Kulevanova, S.; Dimitrovska, A.; Anal. Bioanal. Chem. 2012, 403, 1117. DOI: http://dx.doi.org/10.1007/s00216-012-5716-1 PMID: 22318699

16. Peng, L.; Wang, Y.; Zhu, H.; Chen, Q.; Food Chem. 2011, 125, 1064. DOI: http://dx.doi.org/10.1016/j.foodchem.2010.09.079

17. Bernal, F. A.; Cuca-Suárez, L. E.; Yamaguchi, L. F.; Coy-Barrera, E. D.; Rec. Nat. Prod. 2013, 7, 152.

18. Lee, J.; Durst, R. W.; Wrolstad, R. E.; J. AOAC Int. 2005, 88, 1269. PMID: 16385975

19. Nielsen, N-P. V.; Carstensen, J.; Smedsgaard, J.; J. Chromatogr. A. 1998, 80, 517.

20. http://www.models.life.ku.dk/algorithms#C, accessed October 2013.

21. van den Berg, R. A.; Hoefsloot, H. C.; Westerhuis, J. A.; Smilde, A. K.; van der Werf, M. J.; BMC Genomics 2006, 7, 142. DOI: http://dx.doi.org/10.1186/1471-2164-7-142 PMID: 16762068

22. Beebe, K. R.; Pell, R. J.; Seasholtz, M. B.; Chemometrics: A practical guide, John Wiley and Sons, Inc.: New York, 1998.

23. Bro, R.; Chemometr. Intell. Lab. Syst. 1997, 38, 149. DOI: http://dx.doi.org/10.1016/S0169-7439(97)00032-4

24. http://www.eigenvector.com/software/pls_toolbox.htm, accessed March 2014.

25. Dewick, P. M.; Medicinal natural products: A biosynthetic approach; 3rd ed. John Wiley and Sons, Inc.: Chichester, 2009.

26. He, F.; Mu, L.; Yan, G-L.; Liang, N-N.; Pan, Q-H.; Wang, J.; Reeves, M. J.; Duan, C. Q.; Molecules 2010, 15, 9057. DOI: http://dx.doi.org/10.3390/molecules15129057 PMID: 21150825

27. Holton, T. A.; Cornish, E. C.; Plant Cell 1995, 7, 1071. DOI: http://dx.doi.org/10.1105/tpc.7.7.1071 PMID: 12242398