Chromatographic fingerprint analysis of secondary metabolites in citrus fruits peels using gas chromatography–mass spectrometry combined with advanced chemometric methods

Journal of Chromatography A, 1251 (2012) 176–187

Contents lists available at SciVerse ScienceDirect

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

Chromatographic fingerprint analysis of secondary metabolites in citrus fruits peels using gas chromatography–mass spectrometry combined with advanced chemometric methods Hadi Parastar a,∗ , Mehdi Jalali-Heravi b , Hassan Sereshti c , Ahmad Mani-Varnosfaderani b a

Department of Chemistry, Faculty of Science, University of Isfahan, Isfahan 81746-73441, Iran Department of Chemistry, Sharif University of Technology, P.O. Box 11155-3516, Tehran, Iran c Department of Chemistry, Faculty of Science, University of Tehran, Tehran, Iran b

a r t i c l e

i n f o

Article history: Received 13 March 2012 Received in revised form 30 May 2012 Accepted 3 June 2012 Available online 15 June 2012 Keywords: Chemometrics Multivariate curve resolution Chromatographic fingerprinting Classification Gas chromatography–mass spectrometry Citrus fruits

a b s t r a c t Multivariate curve resolution (MCR) and multivariate clustering methods along with other chemometric methods are proposed to improve the analysis of gas chromatography–mass spectrometry (GC–MS) fingerprints of secondary metabolites in citrus fruits peels. In this way, chromatographic problems such as baseline/background contribution, low S/N peaks, asymmetric peaks, retention time shifts, and co-elution (overlapped and embedded peaks) occurred during GC–MS analysis of chromatographic fingerprints are solved using the proposed strategy. In this study, first, informative GC–MS fingerprints of citrus secondary metabolites are generated and then, whole data sets are segmented to some chromatographic regions. Each chromatographic segment for eighteen samples is column-wise augmented with m/z values as common mode to preserve bilinear model assumption needed for MCR analysis. Extended multivariate curve resolution alternating least squares (MCR-ALS) is used to obtain pure elution and mass spectral profiles for the components present in each chromatographic segment as well as their relative concentrations. After finding the best MCR-ALS model, the relative concentrations for resolved components are examined using principal component analysis (PCA) and k-nearest neighbor (KNN) clustering methods to explore similarities and dissimilarities among different citrus samples according to their secondary metabolites. In general, four clear-cut clusters are determined and the chemical markers (chemotypes) responsible to this differentiation are characterized by subsequent discriminate analysis using counter-propagation artificial neural network (CPANN) method. It is concluded that the use of proposed strategy is a more reliable and faster way for the analysis of large data sets like chromatographic fingerprints of natural products compared to conventional methods. © 2012 Elsevier B.V. All rights reserved.

1. Introduction The awareness of consumers concerning the relation between food and health is revolutionizing in recent years. This claim is confirmed by ever-increasing uses of natural products in human life [1–4]. Citrus fruits are among the most widely used natural products in the world due to the presence of bioactive compounds, such as phenols, vitamin C and carotenoids [2,5]. However, citrus fruits are also sources of essential oils (EOs) due to their aromatic secondary metabolites which usually obtained from the peels of sweet oranges (Citrus sinensis L.), bitter oranges (Citrus aurantium L.), lemons (Citrus limon L.), bergamots (Citrus bergamia), mandarins (Citrus deliciosa Ten.) and grapefruits (Citrus paradise L.) [1,6,7]. Citrus EOs have been classified as generally recognized

∗ Corresponding author. Tel.: +98 311 7932702; fax: +98 311 6689732. E-mail address: [email protected] (H. Parastar). 0021-9673/$ – see front matter © 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.chroma.2012.06.011

as safe (GRAS) due to their antimicrobial, antifungal, antioxidant, anti-inflammatory and anxiolytic activities [1,2,5,7–10]. Due to their great importance, numerous investigations have been performed aimed at identifying the chemical composition of EOs from peels and leaves of different citrus species [6,11–18]. However, the composition and flavor quality of citrus fruits considerably depend on their cultivar, maturity, genotype, origin, climate, season and ripening stage. By considering all the differences in citrus EOs composition, it is clear that only a detailed knowledge of their secondary metabolites will lead to a proper application of their components. However, such a detailed knowledge can be only obtained by means of applying suitable extraction techniques and carefully performed chromatographic analysis [19]. Among different methods, gas chromatography combined to mass spectrometry (GC–MS) is the primary choice for the analysis of citrus EOs [6,19]. On the other side, there has recently been substantial growth of interest in characterization of chemical components of a sample

H. Parastar et al. / J. Chromatogr. A 1251 (2012) 176–187

using its chromatographic fingerprint which is a chromatogram that represents the chemical characteristics of herbal medicine. In general, samples with similar chromatographic fingerprints have similar properties [20–23]. As a result, chromatographic fingerprint analysis has potential to determine the identity, authenticity, quality, and lot-to-lot consistency of natural products [20–23]. Chromatographic fingerprints of secondary metabolites of citrus EOs usually contain a large number of components and form a very complicated system. In addition, presence of different chromatographic problems, such as baseline drift, spectral background, low signal-to-noise (S/N) ratios, retention time shifts, asymmetric peaks and co-elution (overlapped or embedded peaks), makes isolation of all detectable components much more difficult [24–26]. Multivariate chemometric methods can overcome these problems and extract informative chromatographic and spectral information from second-order data obtained from the GC–MS analysis of citrus EOs [19,24–27]. In the present contribution, a new strategy based on chemometric methods was proposed for the comprehensive chromatographic fingerprint analysis of citrus EOs. First, GC–MS was used to generate the informative chromatographic fingerprints under the optimum analytical conditions. Next, extended multivariate curve resolution-alternating least squares (MCR-ALS) [28–33] along with other chemometric methods were used to solve the common chromatographic problems occurred during the GC–MS analysis of citrus EOs. This would result in obtaining pure elution and mass spectral profiles for the secondary metabolites exist in citrus EOs as well as their relative concentrations. Then, principal component analysis (PCA) [34,35] and k-nearest neighbor (KNN) [36] as common unsupervised clustering methods were used to represent changes in chemical compositions of secondary metabolites of citrus EOs among different samples. Finally, the chemical markers (chemotype) which have the most contribution to separated clusters were determined using counter-propagation artificial neural network (CPANN) method [37–39]. 2. Experimental 2.1. Sample collection and chemicals Citrus fruits were collected randomly from healthy trees of lemon (C. limon), orange (C. Sinensis), mandarin (Citrus reticulata) and grapefruit (Citrus paradisi), cultivated under the same pedoclimatic and cultural conditions in an experimental orchard. Random sampling means that among different trees belong to the same species, three samples are selected. For this reason, different fruits were picked up and then a composite sample was chosen for analysis. The orchard was located at the Darab city in Fars province of Iran. Trees were in good vigor, disease free and without visible insect infestation. Eighteen cultivars (Table 1) were selected and their peels were analyzed. Each sample was analyzed three times. Normal hexane and anhydrous sodium sulfate (purity > 95%) were purchased from Merck (Darmstadt, Germany). Normal alkanes standards (C7 –C25 ) were purchased from Ultra Scientific (North Kingstone, USA). 2.2. Extraction of secondary metabolites of citrus fruits In separate experiments, 50.0 g of air dried peel samples were cut into small pieces and then were completely immersed in water and hydro-distilled in a full glass Clevenger-type apparatus. The extraction was carried out for 3 h. When the condensed material cooled down, the water and essential oils were separated. The oil was decanted to be used as essential oil. To improve the recovery,

177

Table 1 Citrus samples included in this study. Sample

Name

Abbreviation

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

Lemon 1 Lemon 2 Lemon 3 Lemon 4 Lemon 5 Lemon 6 Lemon 7 Lemon 8 Orange 1 Orange 2 Orange 3 Orange 4 Orange 5 Mandarin 1 Mandarin 2 Mandarin 3 Grapefruit 1 Grapefruit 2

L1 L2 L3 L4 L5 L6 L7 L8 O1 O2 O3 O4 O5 M1 M2 M3 G1 G2

the essential oil was taken up in n-hexane, dried over anhydrous sodium sulfate until the last traces of water were removed and stored in a dark and air-tight sealed vial at 4 ◦ C and then 0.5 ␮L of each sample was used for GC–MS analysis. 2.3. GC–MS conditions GC–MS analyses were performed using a HP-6890 GC system coupled with a 5973 network mass selective detector (MSD) and equipped with a HP5-MS capillary fused silica column (60 m, 0.25 mm I.D., 0.25 ␮m film thickness). The oven temperature program initiated at 40 ◦ C, held for 1 min then rose at 3 ◦ C/min to 250 ◦ C, held for 20 min. Other operating conditions were as follows: carrier gas, He (99.999%), with a flow rate of 1 mL/min; injector temperature, 200 ◦ C; split ratio of 1:20. An oil sample of 0.5 ␮L was injected in the split mode injection. Mass spectra were taken at 70 eV. The m/z values were recorded in the range of m/z 20–350 amu. 2.4. Identification of secondary metabolites The identification of individual components was based on (i) comparison of mass spectral fragmentation patterns with those stored in the NIST Mass Spectral Library built up using pure substances and the mass spectra from the literature and (ii) comparison of the GC retention indices (RIs) on HP-5MS columns, determined relative to the retention time of a series of n-alkanes with linear interpolation, with those of authentic compounds (using a homologous series of n-alkanes (C7 –C25 ) in this work) and literature data. 2.5. Chemometric analysis Fig. 1 shows the general framework of the strategy proposed in this work for the comprehensive chromatographic fingerprint analysis of secondary metabolites of citrus fruits. First, the data for each sample was exported into the MATLAB environment as comma separated values (CSV) file. The total ion chromatograms (TICs) of EOs are very complicated due to the large number of constituents of these mixtures. In addition, the efficiency of MCR techniques is increased when the number of chemical components and artifacts like baseline drift and noise remain under certain level in a data matrix. It means when more components included in the data matrix, the probability of occurrence of artifacts in chromatographic and mass spectrometric dimensions increases. Therefore, the overlaid TICs for eighteen samples were segmented to desired number of chromatographic segments using local rank analysis

178

H. Parastar et al. / J. Chromatogr. A 1251 (2012) 176–187

Fig. 1. Overall data analysis procedure for the analysis of raw GC–MS fingerprints of secondary metabolites in citrus fruits peels.

and zero component regions. Evolving factor analysis (EFA) [40] and fixed-size moving window EFA (FSMW-EFA) [41] were used to obtain a rank map of data. Using local rank mapping, which splits the GC–MS data into small parts and then measures their ranks, chromatographic segmentation was performed. However, how to split the GC–MS data matrix is uncertain. The main rationale behind the techniques based on local rank analysis is to simply use the separation ability of the chromatographic efficiency without losing the information from the spectral mode or to figure out the whole rank-map in the chromatographic direction. If the rank-map can be quickly and clearly figured out, the information for resolving efficiently the concentration profiles of all the chemical species in the system is at hand. Then, each segment for different samples (e.g. eighteen samples) was augmented column-wise with elution times (or scan numbers) as columns and m/z values as rows (i.e. the common dimension of data between different samples). Some preprocessing methods, such as baseline/background correction using congruence analysis and least squares fitting [42], denoising using morphological score [43] and smoothing using Savitzky–Golay filter [44], were used for some data sets. Baseline correction using congruence analysis and least-square fitting developed by Liang and Kvalheim [42] is a very powerful technique for removing the baseline drift in the chromatographic data. By using this method the local rank analysis of zero component regions can provide sufficient information for performing linear regression with respect to the retention time and finally correcting the baseline. In denoising step, homoscedastic noise in data matrix was reduced. For this purpose, morphological score method [43] was used which was able to discriminate the signal from the noise. Therefore, deleting the mass spectral channels due to the noise would be helpful in reducing noise in the whole signal. Finally, the GC–MS data was smoothed using the Savitzky–Golay filter [44] to transform heteroscedastic noise to homoscedastic one. This filter uses a moving polynomial fit of any order, n, and the size of the filter consists of (2n + 1) points. More details on the performance of these methods for the analysis of GC–MS data can be found in our previous works [19,45–49]. After data pretreatment, the data must be arranged in bilinear way and then, analyzed by MCR-ALS. However, before these steps, determining the number of chemical components (N) in each chromatographic region was very important. In this study, the singular values from singular value decomposition (SVD) [35] and the change in the lack of fit (LOF) of MCR-ALS model by adding more components were used for chemical rank determination.

MCR-ALS is one of the more popular MCR methods which is based on alternating least squares (ALS) [28,29,50,51]. MCR-ALS like other MCR methods is based on bilinear decomposition of mixed signal into the contribution of pure components profiles. The general MCR bilinear decomposition for GC–MS data is as follow: X = CST + E

(1)

which X(I,J) is GC–MS data matrix with I elution time points and J m/z values. The C(I,N) is the matrix of resolved elution profiles and ST (N,J) is the matrix containing resolved mass spectra for N chemical components. Also, E(I,J) is the unmodeled part of the data by MCR. In MCR-ALS, MCR bilinear model (Eq. (1)) is solved for C and ST , using an iterative algorithm based on two constrained linear least squares steps: C = X(ST ) T

+

(2)

+

S = (C) X where (C)+

(3) and (ST )+

are the pseudoinverse of the ST

and C matrices, respectively. ALS optimization requires an initial estimate of the concentration, C, or of the spectra, ST , profiles, which can be easily obtained using different methods or from the ‘purest” data samples or variables (e.g. Simple-to-use interactive self-modeling mixture analysis (SIMPLISMA) [52]). One of the most important features of MCR-ALS is its potential to extension to the analysis of different samples. The extended MCR-ALS bilinear decomposition is as follow:

⎡

X1

⎤

⎡

C1

⎤

⎡

E1

⎤

⎢X ⎥ ⎢C ⎥ ⎢ ⎥ ⎢ 2 ⎥ ⎢ 2 ⎥ T ⎢ E2 ⎥ ⎥ = ⎢ ⎥ S + ⎢ ⎥ = Caug ST + Eaug ⎥ ⎢ ⎥ ⎢ . ⎥ ⎣ ... ⎦ ⎣ ... ⎦ ⎣ .. ⎦

Xaug = ⎢ ⎢

Xk

Ck

(4)

Ek

where Xaug (KI,J) is the column-wise augmented GC–MS data matrix for K samples. Although finding a unique solution is difficult when only the information of the original data matrix is provided, the use of constraints can decrease significantly this indeterminacy. The common constraints for augmented data matrix are non-negativity (to avoid negative values in both resolved elution profiles), unimodality (to avoid the presence of more than one peak maximum in the resolved elution profiles), spectral normalization (to avoid intensity ambiguity in the resolved profiles), component correspondence (to determine the presence or absence of each component in different samples and assure their correct correspondence) and selectivity (to introduce known chemical information about components into the model) [51,53–55]. The selectivity

H. Parastar et al. / J. Chromatogr. A 1251 (2012) 176–187

constraint is associated with the concept of local rank, i.e. how the number and distribution of components vary locally along the data set. The selectivity constraint holds for concentration and spectral windows where only one component is present (i.e. local rank equals to one) and suppresses the ambiguity linked to some of the profiles in the system. In addition, for the regions that components are absent (i.e. local rank equals to zero), the concentration values set to zero. The final results of extended MCR-ALS for each chromatographic segment are an augmented matrix containing the pure elution profiles for each secondary metabolite in different samples (Caug ). Also, these results contain one matrix containing the pure mass spectral profiles for secondary metabolites (ST ). A more detailed discussion about the MCR-ALS method can be found elsewhere [28,29,31,50]. This type of data arrangement and analysis has flexibility to handle the retention time shifts for target compounds in different samples without need to correct it before analysis. In addition, it has the advantage of modeling the baseline in some cases instead of its correction. To get relative concentration of each secondary metabolite in different samples from Caug , proper transformation of this augmented matrix is needed. To do this, each column in augmented elution profiles matrix was appropriately refolded to give a new matrix. The summation of this refolded matrix gives the corresponding relative concentrations of the secondary metabolites in different samples. Statistical parameters of lack of fit (LOF) and variance explained (R2 ) were used to evaluate the best obtained extended MCR-ALS model for each chromatographic segment [31]. To identify the components giving rise to the mass spectral profiles obtained for each extended MCR-ALS model, the ‘NIST MS Search 2.0’ software was used. After obtaining the best extended MCR-ALS model for each chromatographic segment, PCA and KNN were used to cluster data into clear-cut groups. PCA [34,35] is an unsupervised learning technique used to determine hidden structure (associations) in large data sets and to determine those volatiles which were most differentiating. This helps to identify inherent patterns in the data in an unbiased way and to highlight the similarities and dissimilarities (differences) amongst samples (citrus fruits). It also helps to identify those volatiles which are most differentiating within the entire data set. KNN [36] is a another clustering technique by which each sample is classified according to the majority vote of its k-nearest neighbors, where k is an odd number (e.g. 1, 3 and 5). Based on the class label of the majority of the sample’s KNN, the sample is assigned to a class in the data set. PCA and KNN methods were performed by using the final relative areas of the well resolved peaks to seek for similarity and dissimilarity of different samples of citrus fruits based on the secondary metabolites. Then, a supervised discriminant analysis algorithm was used for finding the effective variables responsible for clear-cut clusters. The method of counter propagation artificial neural network (CPANN) [37–39] was used for this purpose. After training the network, the samples were separated correctly on the map. The average values of the weights of the variables responsible for each clear-cut cluster were obtained. These values were used to develop different chemotypes (chemical markers) for similar samples. 2.6. Data analysis and software An Enhanced ChemStation G1701 DA version D.00.01.27 was used for the data collection and conversion to ASCII format. Data analyses were performed on a Pentium-based HP-Compaq personal computer. Some of the chemometric techniques were coded and were executed by authors in MATLAB 7.4 (Mathworks, Natick, USA) for windows. MCRC software v. 1.0 [32] was used for preprocessing, chemical rank determination, local rank analysis and MCR analysis. Additionally, PCA and KNN methods of PLS Toolbox v. 3.5

179

(Eigenvector Research Inc., WA, USA) were used for multivariate clustering of datasets. The CPANN toolbox [39] was used for supervised classification purposes. The library searches and spectral matching of the resolved pure components were conducted on the NIST MS database (National Institute of Standards and Technology, Gaithersburg, MD). 3. Results and discussion The variation of the composition of secondary metabolites of citrus fruits might be attributed to two major factors: genetics and type of environment (soil, cultural practices and weather) [12]. Here, a thorough comparative study of the composition of peels of eighteen samples of citrus fruits has been carried out because all trees were grown in the same conditions of soil, climate and cultural practices. In addition, the extraction and GC–MS conditions were identical for all samples and therefore, the influence of environmental and technical parameters on the chemical composition of the EOs were negligible [12]. 3.1. Chromatographic fingerprint segmentation and preprocessing The chromatographic landscape of the overlaid TICs of chromatographic fingerprints of eighteen citrus fruits samples is depicted in Fig. 2a. In Fig. 2b the mean TIC of eighteen samples is demonstrated to better visualization of detailed pattern of GC–MS data. As it can be seen the chromatograms of these samples are very complex with a large number of peaks together with common chromatographic problems, such as baseline/background contribution, low S/N, different types of noise, retention time shifts and co-elution (i.e. overlapped and imbedded peaks). Therefore, the use of chemometric methods to overcome these problems and obtain proper information about the identity and quantity of the constituents in each sample is mandatory [19,24,26,27]. On the other hand, the chromatographic fingerprints (i.e. the chemical composition) are different for different samples. Therefore, a proper segmentation of the chromatograms to speed up the data analysis and to obtain reliable qualitative and quantitative information is an important step in chromatographic data analysis. However, due to the presence of a large amount of shifts in retention times between different samples and presence of a large number of peaks in each sample, a proper segmentation of all samples is difficult [19,27]. The use of local rank analysis methods of EFA [40] and FSMW-EFA [41] as a data microscope, which gives a clear map of the rank of the data matrix in different regions (elution time points) for each sample, was used to determine zero component regions (where no chemical components elute with local rank equals to zero). These regions were used for proper segmentation of TICs. On the other side, presence of mass spectral dimension in the GC–MS data can help in dividing the chromatograms to a small number of chromatographic segments in the presence of large retention time shifts. Finally, after precise inspection of local rank maps obtained by local rank analysis methods and the mass spectra of constituents, the overlaid TICs for eighteen samples were divided to forty-five segments. Once the whole dataset was divided to desired number of chromatographic segments, each segment for eighteen samples and their replicates was column-wise augmented, with m/z values as columns and elution times (or scan numbers) as rows, in the multi-set augmented matrix according to the strategy in Fig. 1 and Section 2.5. In order to develop a comprehensive, easy and fast method for the analysis of this augmented matrix in one shot, two approaches were followed. In the first approach, it was tried to handle the chromatographic problems, such as baseline/background contribution and retention time shifts during the MCR analysis instead of their

180

H. Parastar et al. / J. Chromatogr. A 1251 (2012) 176–187

Fig. 2. (a) TICs for the secondary metabolites in eighteen citrus samples and (b) the mean TIC for GC–MS data of eighteen samples.

correction before the MCR analysis. The second approach was used in some cases that the first approach did not give proper results. In this case, it was needed to correct the baseline/background contribution and to reduce the effects of different types of noise before column-wise augmentation and MCR analysis. Congruence analysis and least square fitting method developed by Liang and Kvalheim [42] was used for the baseline/background correction. In addition, morphological score method [43] and Savitzky–Golay filter [44] were used to reduce homoscedastic and heteroscedastic noise in the data, respectively. All of these methods were straightforwardly performed by MCRC software as a chemometric tool for the analysis of chromatographic data [32]. Fig. 3 shows an example of the outputs of MCRC software on the performance of baseline correction and noise reduction methods on a chromatographic segment. Fig. 3a depicts a chromatographic segment before correction and Fig. 3b shows the corrected data after applying morphological score (for homoscedastic noise reduction), Savitzky–Golay filter (for heteroscedastic noise reduction) and Liang–Kvalheim methods (for baseline/background correction), respectively. It is clear from this figure that the results are promising and these methods could significantly reduce the effects of chromatographic artifacts.

3.2. Extended MCR-ALS performance on chromatographic segments As mentioned before, each chromatographic segment for eighteen samples can be arranged in a column-wise augmented matrix with elution time points as rows and m/z values as columns (Fig. 1). Once this matrix in arranged for each segment and preprocessed (in some cases), it was analyzed by extended MCR-ALS under proper constraints including non-negativity (both elution time and mass spectral modes), unimodality (elution time mode), spectral normalization, selectivity (in elution time mode) and component correspondence. Additionally, SIMPLISMA [52] was used as a method to estimate the initial values of spectra to start ALS optimization algorithm. In addition, since determining the number of components is critical in the resolution procedure, therefore, the singular values from SVD [35] and the changes in the LOF values by adding more components to the MCR-ALS model were used as criteria to determine the right number of components in the augmented matrix. Fig. 4 shows an example of the results obtained by extended MCR-ALS for the chromatographic segment between 20.5 and 20.9 min. Fig. 4a shows the eighteen TICs for a one-component

H. Parastar et al. / J. Chromatogr. A 1251 (2012) 176–187

181

Fig. 3. Performance of baseline/background correction method (congruence analysis and least squares fitting) and noise reduction methods (morphological score and Savitzky–Golay filter) on a chromatographic segment taken from sample L4 using MCRC software. (a) Raw chromatographic segment and (b) chromatographic segment after pretreatment.

system with a large amount of retention time shifts. It can be seen from Fig. 4b and c that extended MCR-ALS can efficiently handle the retention time shifts and obtain proper elution and mass spectral profiles for the augmented matrix (with LOF = 4.6%). This is important to note that the resolved elution profile for this chromatographic segment was an augmented matrix contains a

single elution profiles in eighteen samples. However, for clarity only the mean elution profile after doing post-processing to obtain the relative concentration of this component in different samples is shown in Fig. 4b. The relative concentration of 1R-␣-pinene in eighteen samples was obtained and is shown in Fig. 4d. The reverse match factor (RMF) (which is the square root of the normalized dot

Fig. 4. Extended MCR-ALS results for TICs of a chromatographic segment of eighteen citrus samples between 20.5 and 20.9 min contain a large amount of retention time shifts. (a) Overlaid TICs for eighteen samples, (b) Resolved elution profile, (c) resolved mass spectrum and (d) relative concentration of resolved component in eighteen samples.

182

H. Parastar et al. / J. Chromatogr. A 1251 (2012) 176–187

product between the MCR-ALS resolved mass spectrum and the possible standard mass spectrum candidate from NIST MS library [56]) was 954 for the resolved mass spectrum (1R-␣-pinene) which confirms the validity of the proposed strategy. Fig. 5a demonstrates a more complex chromatographic segment with a large number of components, baseline/background contribution, low S/N ratio (i.e. high level of noise), retention time shift and co-elution. This example is devoted to show how extended MCR-ALS can help to obtain proper elution and mass spectral profiles for all components present, where most of the chromatographic problems exist. According to the proposed strategy in this work, the GC–MS data for this chromatographic region in eighteen samples as well as their replicates arranged in a column-wise augmented data matrix with m/z values as common mode. This bilinear data structure allows keeping mass spectral mode the same among different chromatographic runs, but allowing each chromatographic run to be described in elution shape and retention time by a different set of elution profiles, even if they belong to the same compounds in different runs. In addition, in this example it was tried to resolve the data using MCR-ALS without using pre-processing methods. In other words, baseline/background contribution, different type of noise and retention time shifts dealt with during MCR analysis. Therefore, the MCR-ALS model was performed on a raw data arranged in a column-wise augmented data matrix with SIMPLISMA initial values of mass spectral profiles under proper constraints of non-negativity, unimodality, spectral normalization and component correspondence. It should be pointed out that SVD analysis showed that at least twelve significant components (eleven chemical components as well as one baseline/background contribution) could exist in this data matrix which then confirmed by LOF value of MCR-ALS model. Fig. 5b and c shows the MCR-ALS resolved elution and mass spectral profiles, respectively. As extended MCR-ALS is based on bilinear model assumption, so, it can model the retention time shift without need to be corrected before analysis. Also, in this case the baseline/background contribution was modeled by adding one more component to the MCR-ALS model. The LOF value for extended MCR-ALS model with twelve components was 8.98% which is acceptable according to the noise level in the data. In this case also for the sake of brevity the mean elution profiles for eighteen samples are shown. Modeling the baseline/background contribution in this case is an interesting aspect of the proposed strategy. As it can be seen from Fig. 6b, six components out of twelve are below the baseline level which means their detection using conventional methods are impossible. The resolved mass spectra in Fig. 6c were identified using the NIST MS database. According to the MCR-ALS resolved mass spectral matching with the standard mass spectra in the library, the identified components were respectively germaceren B (RMF = 964), elemol (RMF = 933), endo-1-bourbonanol (RMF = 968), spathulenol (RMF = 979), cedren-13-ol, 8- (RMF = 821), veridiflorol (RMF = 957), ␤-guaiene (RMF = 878), germacrene-d-4ol (RMF = 910), cubenol (RMF = 897), isospathulenol (RMF = 910) and ␥-cadinol (RMF = 858). Most of the components were identified with high RMF values. This example shows one of the most important features of the MCR-ALS for comprehensive analysis of multi-set chromatographic data in the presence of chromatographic artifacts. The proposed strategy shown in Fig. 1 was used for all chromatographic segments and the elution and mass spectral profiles for each secondary metabolite in eighteen samples were extracted. In addition, the relative concentration of each secondary metabolite modeled in different samples (integrated peaks matrix) was obtained from the augmented matrix contains the resolved elution profiles by proper transformation (Section 2.5) [56]. In this way, a large data set was obtained for all identified secondary metabolites in eighteen citrus samples. For some samples (e.g. L4–L7) the

numbers of components were more than 150. However, For the sake of brevity, only the main secondary metabolites are reported. Table 2 shows the chemical name and formula for the thirtyseven main identified secondary metabolites in eighteen citrus samples. In this table, the average retention times for identified components and corresponding kovat’s retention indices (RIs) on HP5-MS column are reported. The values of RIs were used for further confirmation of the identified metabolites. However, in most cases identification was based on the comparison of the MCR-ALS resolved mass spectra of metabolites with those of the standards stored in NIST MS database and the average of RMF values are shown in the Table 2. Also, the relative peak areas for these secondary metabolites averaged over three replicates are reported. As it can be seen from Table 2, secondary metabolites in citrus fruits peel oils belong almost exclusively to hydrocarbons (monoterpenes and sesquiterpenes), olefins and oxygenated ones. In addition, limonene was always the main secondary metabolite in all citrus EOs. The results in Table 2 which were obtained using a detailed MCR analysis are only a part of all the identified secondary metabolites for eighteen citrus samples. In other words, the components in Table 2 were common identified metabolites with noticeable relative concentrations which were relevant for subsequent cluster analysis. 3.3. Multivariate clustering of citrus samples by their secondary metabolites After finding the best fit extended MCR-ALS model, the elution and mass spectral profiles of each detectable component were obtained. In addition, as it is mentioned before, the integrated peak areas for each component in different samples were also obtained. Preliminary studies using PCA and KNN showed that only thirtyseven secondary metabolites shown in Table 2 have greater effect on the clustering of eighteen citrus samples. Therefore, these main citrus secondary metabolites in different samples were considered for subsequent analyses. In order to examine this data set, PCA and KNN methods were employed. The raw data set was clearly useless for clustering the secondary metabolites of citrus EOs. Therefore, the data was subsequently autoscaled as pretreatment step to remove biasing due only to feature size [36]. Fig. 6a and b depicts the score and loading plots for eighteen samples and thirty-seven variables, respectively. The PC1 and PC2 explained 41.11% and 22.76% of the total variance, respectively. It can be seen from Fig. 6a that the data points, representing individual citrus samples, form three clear-cut groups on the map according to their GC–MS data. However, it was not clear that samples L4 and L5 are in the same group with L6 and L7 or not. According to the score plot (Fig. 6a), most of the samples (twelve out of eighteen) are included in cluster II. The cluster I composed of two samples G1 and G2. The last cluster (cluster III) composed of four samples of L4–L7. Fig. 6b demonstrates the 2D loading plot for thirty-seven secondary metabolites. This plot reveals the influence of individual secondary metabolites in differentiating between citrus samples when the entire data is considered. Those secondary metabolites that account for maximum variance in the data set are given more weight or loading. However, exploiting information about the effective variables on differentiating clear-cut clusters is very difficult from Fig. 6b and needs more exploration. In order to make sure about clear-cut clusters of eighteen samples, the effective variables on their differentiations and to explore more detailed pattern of the secondary metabolites of citrus samples, KNN method was used. Fig. 7a and b shows the cluster dendrograms of eighteen citrus samples and thirty-seven variables, respectively. It should be noted that mahalanobis distances were calculated here and PCA was used for preprocessing. In this case mahalanobis distance on

Table 2 Main identified secondary metabolites in citrus EOs, their retention times (RTs), retention indices (RIs), RMF values and peak areas obtained using MCR-ALS. No. Chemical namea

Formula

RTb

RIc

RMFd Peak arease (×107 ) L1

a b c d e

1-Hexanol l-Phellandrene 1R-␣-pinene Camphene Sabinene 2-␤-Pinene ␤-Myrcene dl-Limonene trans-␤-Ocimene ␥-Terpinene 1-Octanol Linalool oxide Linalool Nonanal Limonene oxide Citronellal 4-Terpineol ␣-Terpineol cis-Carveol trans-Carveol ␤-Citronellol Nerol ␤-Citral l-Carvone Geraniol Linalyl acetate E-citral Perilla alcohol Citronellyl acetate Neryl acetate Geraniol acetate Caryophyllene trans-␤-Farnesene ␣-Humulene Germacrene D ␥-Elemene Nerolidol

C6 H14 O C10 H16 C10 H16 C10 H16 C10 H16 C10 H16 C10 H16 C10 H16 C10 H16 C10 H16 C8 H18 O C10 H18 O2 C10 H18 O C9 H18 O C10 H16 O C10 H18 O C10 H18 O C10 H18 O C10 H16 O C10 H16 O C10 H20 O C10 H20 O C10 H16 O C10 H14 O C10 H18 O C12 H20 O2 C10 H16 O C10 H16 O C12 H22 O2 C12 H20 O2 C12 H20 O2 C15 H24 C15 H24 C15 H24 C15 H24 C15 H24 C15 H26 O

17.09 20.29 20.73 21.55 22.83 23.12 23.58 26.13 26.68 27.50 27.84 28.06 29.33 29.48 31.41 31.99 33.46 34.18 35.40 35.45 35.67 35.78 36.50 36.67 36.81 37.01 37.78 39.17 41.29 41.84 42.68 44.98 45.91 46.44 47.55 48.06 50.11

750.8 876.1 891.8 920.1 962.2 971.4 985.8 1060.7 1075.9 1098.0 1107.0 1112.8 1145.1 1148.8 1195.1 1208.4 1241.2 1256.8 1282.4 1283.4 1287.9 1290.2 1304.7 1308.1 1310.9 1314.8 1329.9 1356.2 1394.7 1404.4 1418.9 1457.2 1472.1 1480.5 1497.7 1505.5 1536.0

990 984 981 985 982 989 990 988 991 978 993 978 981 926 922 937 963 985 965 967 992 987 980 985 984 993 984 967 910 978 993 988 956 982 988 928 983

L3

L4

0.01 0.06 0.52 0.01 0.11 0.52 0.13 9.82 0.14 5.52 7.34 34.80 0.01 0.05 0.18 0.57 0.01 1.39 5.77 18.00 0.43 1.48 22.9 84.00 0.33 17.8 18.7 53.80 15.5 606.0 637.0 1130.0 0.01 25.00 2.65 22.90 0.01 0.17 0.67 188.00 0.01 0.43 1.18 1.01 1.08 1.12 1.91 7.39 1.68 14.90 4.61 541.00 0.25 0.28 0.18 0.45 0.01 0.59 0.07 0.01 0.11 1.54 0.19 1.75 0.18 0.49 1.11 11.50 0.86 1.66 3.46 58.40 0.01 0.19 0.40 1.12 0.01 1.45 0.18 0.58 0.11 0.22 0.09 1.83 0.83 0.54 0.12 11.10 1.54 0.15 2.19 19.80 0.01 0.14 1.09 1.29 0.81 0.12 0.57 0.06 0.01 0.07 0.01 132.00 1.97 0.17 2.76 25.50 0.01 0.05 0.57 0.21 0.09 2.66 0.06 0.51 0.80 4.37 0.98 10.10 0.30 0.19 1.43 18.50 0.36 3.36 3.83 7.23 0.01 0.49 0.29 1.63 0.01 0.54 0.24 1.01 0.01 8.87 0.27 12.70 0.01 1.23 0.50 0.57 0.01 0.10 0.15 0.21

L5

L6

L7

L8

O1

O2

O3

O4

O5

M1

M2

M3

G1

G2

0.02 0.01 0.01 0.44 0.01 0.02 0.02 0.66 0.22 0.09 0.02 0.03 0.04 0.01 8.17 6.81 4.81 0.04 0.08 0.04 0.03 0.04 0.02 6.03 0.03 0.02 5.96 0.10 29.20 32.00 23.70 5.74 12.70 4.03 8.48 5.33 5.54 25.10 8.63 4.54 24.90 21.90 0.56 1.75 1.25 0.01 0.08 0.03 0.06 0.03 0.03 0.16 0.06 0.03 0.01 0.12 14.50 0.01 0.01 3.63 11.20 2.42 2.18 2.48 2.09 5.44 1.99 1.54 10.70 8.84 72.10 242.00 184.00 0.57 0.88 0.19 1.34 0.17 0.14 11.60 1.40 0.16 0.92 0.85 45.40 11.70 9.73 17.10 47.40 14.20 28.30 15.90 16.20 42.60 27.0 13.70 88.60 74.40 965.0 337.0 301.0 583.0 1220.0 544.0 914.0 562.0 571.0 1030.0 880.0 563.0 1830.0 1630.0 18.70 4.63 2.99 1.59 0.86 0.25 0.89 0.32 0.14 1.30 2.74 1.93 5.24 5.38 162.0 133.0 86.20 0.19 1.08 0.38 0.17 0.30 0.18 116.00 0.23 0.24 1.16 0.98 0.52 0.01 2.36 0.96 0.01 2.87 0.27 17.20 13.40 2.15 0.90 0.06 0.14 0.32 6.08 0.01 0.01 0.39 0.01 0.01 4.41 0.01 0.01 0.01 1.32 0.01 11.60 8.19 451.0 6.26 3.58 1.89 21.40 15.10 1.47 2.70 10.00 34.10 4.03 9.09 20.70 16.20 0.23 0.01 1.06 0.18 0.01 0.45 0.36 4.86 4.20 1.68 0.14 0.05 0.76 0.48 0.01 0.01 0.15 0.23 0.12 0.46 0.06 0.51 0.35 1.19 0.09 0.11 0.53 1.25 1.70 1.48 0.72 0.53 3.05 1.17 0.04 0.39 0.38 1.18 0.05 4.05 3.25 2.23 9.43 14.80 6.75 0.21 2.11 0.91 0.15 0.53 0.35 4.72 0.38 0.59 3.20 2.24 47.70 21.10 10.10 1.54 4.01 2.46 0.55 1.13 2.33 11.50 1.90 2.49 0.21 13.00 0.25 0.20 0.16 0.05 0.33 0.22 0.25 0.20 0.15 0.15 0.36 0.37 0.31 0.20 0.16 0.08 0.23 0.21 0.90 0.13 0.17 0.14 0.03 0.36 0.25 0.27 2.87 2.97 1.48 2.69 1.61 0.26 2.86 1.14 0.05 0.53 0.36 1.00 0.11 8.51 0.82 1.02 9.26 5.85 4.21 0.55 0.83 0.77 0.15 0.34 0.43 0.80 0.20 1.01 1.50 0.19 16.20 31.90 12.80 2.36 4.62 2.81 0.01 0.76 1.16 0.32 0.10 0.57 10.10 5.56 1.02 0.01 0.07 0.32 0.98 0.40 0.09 0.41 0.23 0.81 1.66 0.45 1.55 1.43 0.16 9.40 6.50 0.76 0.97 0.71 0.14 0.41 0.58 0.14 0.53 0.14 3.81 2.09 109.0 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 2.33 0.03 0.01 0.01 20.80 41.00 16.80 3.00 6.02 3.52 0.20 0.96 1.48 0.40 0.12 0.70 13.00 7.10 0.15 0.11 0.11 0.33 0.08 0.07 0.19 0.08 0.12 0.11 0.14 0.16 1.33 0.10 0.41 0.03 0.11 0.10 0.13 0.08 0.04 0.08 0.02 0.05 0.13 0.05 0.01 0.01 8.16 1.58 1.07 0.50 0.08 0.06 0.28 0.07 0.03 0.07 0.58 0.01 0.33 0.30 15.10 2.85 1.45 0.87 0.10 0.06 1.83 0.01 0.06 0.07 1.42 0.03 2.79 3.12 5.53 14.50 9.90 1.23 0.65 0.13 2.88 0.24 0.09 0.14 0.84 0.24 15.00 13.90 1.26 1.63 1.37 0.03 0.10 0.15 0.09 0.03 0.02 0.11 0.01 0.11 0.27 0.35 1.37 2.44 1.87 0.22 0.11 0.01 0.31 0.04 0.05 0.18 0.07 0.15 2.52 2.19 9.81 2.16 1.77 4.00 0.21 0.06 3.65 0.12 0.16 1.06 0.17 0.42 4.08 3.63 0.69 1.20 0.90 0.52 0.01 0.01 0.18 2.95 0.05 0.22 0.03 0.07 1.56 1.33 0.01 0.20 0.16 0.01 0.28 0.02 0.01 0.01 0.02 0.03 0.21 0.05 0.67 0.61

H. Parastar et al. / J. Chromatogr. A 1251 (2012) 176–187

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37

L2

Identification was made according to comparison of resolved mass spectra with those of standards in MS Library Database. Average retention times (min) for main identified components. Kovats retention indices in HP-5MS column in reference to normal alkanes. Reverse match factor for the resolved MCR-ALS mass spectrum and the standard mass spectrum in NIST MS database. Average peak areas for three times repetitions.

183

184

H. Parastar et al. / J. Chromatogr. A 1251 (2012) 176–187

Fig. 5. Example of the results obtained using extended MCR-ALS. (a) TICs of the chromatographic segment for eighteen citrus samples between 53.0 and 54.0 min, (b) resolved MCR-ALS elution profiles for twelve components and (c) resolved mass spectral profiles.

three PCs could explain 81.4% of the variance. Fig. 7a suggests the existence of four principal clusters (I–IV) within the citrus samples. These results are better than PCA ones (Fig. 6a). The results of KNN revealed that clusters I and II are the same as PCA. However, samples L6 and L7 are different from L4 and L5 and they considered as

a separate cluster according to the mahalonobis distances. Therefore, cluster III (by PCA) is separated to two different clusters III (L6 and L7) and IV (L4 and L5). Fig. 7b shows the dendrogram of thirty-seven most abundant secondary metabolites of eighteen samples with 99.7% of explained

Fig. 6. PCA projection analysis of peak areas for thirty-seven secondary metabolites in eighteen samples obtained using extended MCR-ALS. (a) Score plot and (b) loading plot.

H. Parastar et al. / J. Chromatogr. A 1251 (2012) 176–187

185

Fig. 7. KNN dendrograms of (a) eighteen citrus samples and (b) thirty-seven identified secondary metabolites using extended MCR-ALS.

variance for three PCs. According to this figure, the distributions of thirty-two variables (out of thirty-seven) are more or less similar for all samples. In contrast, the distances between five variables (secondary metabolites) ␤-pinene (6), limonene (8), ␥-terpinene (10), linalool (13), and linalylacetate (26) are considerable and therefore, they have different information inside. Therefore, these variables were considered for subsequent discriminant analysis to determine which variable(s) make each cluster different to others. By considering the results of both PCA and KNN methods, we concluded that the samples belong to four clear-cuts clusters. In addition, five variables of ␤-pinene (6), limonene (8), ␥-terpinene (10), linalool (13), and linalylacetate (26) were considered for subsequent analysis.

3.4. Chemotypes of clear-cuts citrus samples groups After dividing the eighteen citrus samples to four distinct clear-cut groups according to their secondary metabolites, the determination of the main chemical components responsible for each class is important from food chemistry and pharmaceutical points of view. According to the formalization of European Union in 2006 with the adoption of the regulation REACH, chemotype (sometimes chemovar) is defined as a chemically distinct component(s) in a plant, with differences in the composition of the secondary metabolites [12]. Minor genetic and epigenetic changes with little or no effect on morphology may produce large changes in the chemical phenotype.

Fig. 8. The main secondary metabolite(s) (chemotype) responsible for each clear-cut groups. (a) Cluster I, (b) cluster II, (c) cluster III, and (d) cluster IV. Variables 6 = ␤-pinene, 8 = dl-limonene, 10 = ␥-terpinene, 13 = linalool and 26 = linalyl acetate (see text for more information).

186

H. Parastar et al. / J. Chromatogr. A 1251 (2012) 176–187

The CPANN method as a supervised classification method was used for the development of the chemotype of each clear-cut cluster. The data matrix of eighteen samples with five variables was used as input for training an 11 × 11 network. The elements of the output vector were assigned to be 1, 2, 3, or 4 according to the clusters of samples (I–IV). After training the network, the samples were separated correctly on the map. The average values of the weights of the variables for the neurons of the map were used for chemotype development. Fig. 8 shows the main chemical components (secondary metabolites) responsible for each clear-cut groups. It can be seen from Fig. 8a that in the cluster I limonene is the main component which makes G1 and G2 samples different to others. Cluster II was contained the samples with ␥-terpinene/limonene as the main secondary metabolites (Fig. 8b). Cluster III is differed by high contents of ␤-pinene/␥-terpinene/limonene (Fig. 8c). Cluster IV was contained of linalyl acetate/linalool/␥-terpinene/limonene/␤-pinene as the main components (Fig. 8d). The results showed four major distinguished chemotypes for citrus fruits peel oils. The compositions of citrus fruits EOs have been the subject of numerous studies [11–14,16,57–59]. In most of these studies only different cultivars of one type of citrus fruits (e.g. lemon, orange, mandarin and grapefruit) have been studied. However, in this work chemical compositions of four types of citrus fruits were simultaneously studied. It is important to note that the four distinguished chemotypes in this work were in agreement with previous studies for separate types of citrus fruits [11–14,16,57–59]. In summary, our results in association with literature data provide a new insight into the chemical variability of citrus fruits peel oils. In addition, each cultivar can be distinguished from the other by their peel oil pattern. Obviously, such proposition should be confirmed with a broader set of samples. Nevertheless, they represent a valuable contribution for the determination of the botanical origin of the natural products with the aid of multivariate chemometric methods. Furthermore, differences in chemical compositions of the studied citrus fruits peel oils reflect genetically differences among studied cultivars. The found chemotypes could be considered as potential sources of different EO pattern which could be used in a wide variety of applications, mainly in the food, pharmaceutical, perfume and cosmetic industries. Finally, we have reached the conclusion that reliable information can be obtained from the chromatographic fingerprint analysis of secondary metabolites in citrus EOs with the help of chemometric methods.

4. Conclusions Analysis of chromatographic fingerprints of complex mixtures such as natural products suffers from fundamental problems such as baseline/background contribution, low S/N peaks, asymmetric peaks, retention time shifts, and co-elution (overlapped and embedded peaks). These problems can severely affect quantification and identification of the components exists in these samples. In other words, the common method for the analysis of GC–MS fingerprints of secondary metabolites in natural products is integration of those peaks that can be identified by their mass spectra and then data exploration using PCA. However, in this case chromatographic problems can severely affect the peak integration and identification and therefore, final results. This means that the development of a reliable method requires combination of GC–MS with advanced chemometric methods. Therefore, in this work, a new strategy based on combination of multivariate curve resolution and multivariate clustering methods along with other chemometric methods was proposed for comprehensive analysis of whole chromatographic fingerprints of secondary metabolites in citrus fruits

peels on account of the advantages of their pharmacological activities and low toxicity. Extended MCR-ALS was used not only for fast extraction of elution and mass spectral profiles of the components exist in column-wise augmented data matrix of chromatographic segments but also for handling baseline/background contribution and retention time shifts among different samples. The obtained relative concentrations of the secondary metabolites in eighteen citrus samples were classified to four clear-cut groups using PCA and KNN methods and then, using CPANN method the main chemical components (chemotypes) responsible for differentiation of each class were determined. Inspection of the results confirmed the achievement of more reliable results with time and work saving from the GC–MS fingerprints of secondary metabolites in citrus samples. Acknowledgments Hadi Parastar would like to thank the University of Isfahan for the financial support of this project. References [1] M. Viuda-Martos, Y. Ruiz-Navajas, J. Fernández-López, J. Perez-Álvarez, J. Food Saf. 28 (2008) 567. [2] E. González-Molina, R. Domínguez-Perles, D.A. Moreno, C. García-Viguera, J. Pharm. Biomed. Anal. 51 (2010) 327. [3] S. Siddique, M. Shafique, Z. Parveen, S.J. Khan, R. Khanum, Pharmacologyonline 2 (2011) 499. [4] S. Roowi, A. Crozier, J. Agr. Food Chem. 59 (2011) 12217. [5] R. Guimarães, L. Barros, J.C.M. Barreira, M.J. Sousa, A.M. Carvalho, I.C.F.R. Ferreira, Food Chem. Toxicol. 48 (2010) 99. [6] P.Q. Tranchida, I. Bonaccorsi, P. Dugo, L. Mondello, G. Dugo, Flavour Frag. J. 27 (2012) 98. [7] K. Fisher, C. Phillips, Trends Food Sci. Technol. 19 (2008) 156. [8] M.P. Salas, G. Céliz, H. Geronazzo, M. Daz, S.L. Resnik, Food Chem. 124 (2011) 1411. [9] A.Y. Wang, M.Y. Zhou, W.C. Lin, Food Chem. 124 (2011) 958. [10] F.C.F. Pimenta, N.A. Correia, K.L.G.D. Albuquerque, D.P. de Sousa, M.R.D. de Rosa, M.B.F. Pimenta, M.F.F.M. Diniz, R.N. de Almeida, J. Med. Plant Res. 6 (2012) 342. [11] S.B. Mitiku, M. Sawamura, T. Itoh, H. Ukeda, Flavour Frag. J. 15 (2000) 240. [12] M.L. Lota, D. De Rocca Serra, F. Tomi, C. Jacquemond, J. Casanova, J. Agr. Food Chem. 50 (2002) 796. [13] T. Barboni, F. Luro, N. Chiaramonti, J.M. Desjobert, A. Muselli, J. Costa, Food Chem. 116 (2009) 382. [14] N.T. Lan-Phi, T. Shimamura, H. Ukeda, M. Sawamura, Food Chem. 115 (2009) 1042. [15] K. Hosni, N. Zahed, R. Chrif, I. Abid, W. Medfei, M. Kallel, N.B. Brahim, H. Sebei, Food Chem. 123 (2010) 1098. [16] A. Bermejo, M.J. Llosá, A. Cano, Food Sci. Technol. Int. 17 (2011) 55. [17] P. Dugo, I. Bonaccorsi, C. Ragonese, M. Russo, P. Donato, L. Santi, L. Mondello, Flavour Frag. J. 26 (2011) 34. [18] H. Omori, K. Nakahara, K. Umano, Flavour Frag. J. 26 (2011) 396. [19] M. Jalali-Heravi, H. Parastar, Talanta 85 (2011) 835. [20] X.H. Fan, Y.Y. Cheng, Z.L. Ye, R.C. Lin, Z.Z. Qian, Anal. Chim. Acta 555 (2006) 217. [21] P. Xie, S. Chen, Y.z. Liang, X. Wang, R. Tian, R. Upton, J. Chromatogr. A 1112 (2006) 171. [22] C.J. Xu, Y.Z. Liang, F.T. Chau, Y.V. Heyden, J. Chromatogr. A 1134 (2006) 253. [23] G. Alaerts, N. Matthijs, J. Smeyers-Verbeke, Y. Vander Heyden, J. Chromatogr. A 1172 (2007) 1. [24] J.M. Amigo, T. Skov, R. Bro, J. Coello, S. Maspoch, Trends Anal. Chem. 27 (2008) 714. [25] T. Skov, R. Bro, Anal. Bioanal. Chem. 390 (2008) 281. [26] J.M. Amigo, T. Skov, R. Bro, Chem. Rev. 110 (2010) 4582. [27] J.M. Amigo, M.J. Popielarz, R.M. Callejón, M.L. Morales, A.M. Troncoso, M.A. Petersen, T.B. Toldam-Andersen, J. Chromatogr. A 1217 (2010) 4422. [28] R. Tauler, B. Kowalski, S. Fleming, Anal. Chem. 65 (1993) 2040. [29] R. Tauler, Chemometr. Intell. Lab. Syst. 30 (1995) 133. [30] R. Tauler, A. Smilde, B. Kowalski, J. Chemometr. 9 (1995) 31. [31] J. Jaumot, R. Gargallo, A. De Juan, R. Tauler, Chemometr. Intell. Lab. Syst. 76 (2005) 101. [32] M. Jalali-Heravi, H. Parastar, M. Kamalzadeh, R. Tauler, J. Jaumot, Chemometr. Intell. Lab. Syst. 104 (2010) 155. [33] H. Parastar, M. Jalali-Heravi, R. Tauler, Trends Anal. Chem. 31 (2012) 134. [34] S. Wold, K. Esbensen, P. Geladi, Chemometr. Intell. Lab. Syst. 2 (1987) 37. [35] K.H. Esbensen, P. Geladi, in: D.S. Brown, R. Tauler, B. Walczak (Eds.), Comprehensive Chemometrics, Elsevier, Oxford, 2009, p. 211. [36] K.H. Liland, Trends Anal. Chem. 30 (2011) 827. [37] J. Zupan, M. Noviˇc, J. Gasteiger, Chemometr. Intell. Lab. Syst. 27 (1995) 175. [38] J. Zupan, M. Noviˇc, I. Ruisánchez, Chemometr. Intell. Lab. Syst. 38 (1997) 1.

H. Parastar et al. / J. Chromatogr. A 1251 (2012) 176–187 [39] D. Ballabio, V. Consonni, R. Todeschini, Chemometr. Intell. Lab. Syst. 98 (2009) 115. [40] M. Maeder, Anal. Chem. 59 (1987) 527. [41] H.R. Keller, D.L. Massart, Anal. Chim. Acta 246 (1991) 379. [42] Y.Z. Liang, O.M. Kvalheim, A. Rahmani, R.G. Brereton, Chemometr. Intell. Lab. Syst. 18 (1993) 265. [43] H. Shen, L. Stordrange, R. Manne, O.M. Kvalheim, Y. Liang, Chemometr. Intell. Lab. Syst. 51 (2000) 37. [44] A. Savitzky, M.J.E. Golay, Anal. Chem. 36 (1964) 1627. [45] M. Jalali-Heravi, H. Parastar, H. Sereshti, Anal. Chim. Acta 623 (2008) 11. [46] M. Jalali-Heravi, H. Parastar, H. Ebrahimi-Najafabadi, J. Chromatogr. A 1216 (2009) 6088. [47] M. Jalali-Heravi, H. Parastar, Chemometr. Intell. Lab. Syst. 101 (2010) 1. [48] M. Jalali-Heravi, H. Parastar, H. Ebrahimi-Najafabadi, Anal. Chim. Acta 662 (2010) 143.

[49] [50] [51] [52] [53] [54] [55] [56]

187

M. Jalali-Heravi, H. Parastar, H. Sereshti, J. Chromatogr. A 1217 (2010) 4850. R. Tauler, D. Barcelo, Trends Anal. Chem. 12 (1993) 319. R. Tauler, J. Chemometr. 15 (2001) 627. W. Windig, J. Guilment, Anal. Chem. 63 (1991) 1425. A. De Juan, R. Tauler, Anal. Chim. Acta 500 (2003) 195. A. de Juan, R. Tauler, Cr. Rev. Anal. Chem. 36 (2006) 163. A. de Juan, R. Tauler, J. Chromatogr. A 1158 (2007) 184. ´ M. Jalali-Heravi, S. Diez, J.M. Bayona, R. Tauler, Anal. H. Parastar, J.R. Radovic, Chem. 83 (2011) 9289. [57] M.G. Chisholm, M.A. Wilson, G.M. Gaskey, Flavour Frag. J. 18 (2003) 106. [58] M. Viuda-Martos, Y. Ruiz-Navajas, J. Fernández-López, J.A. Pérez-Álvarez, J. Essent. Oil Bear. Plants 12 (2009) 236. [59] M.W. Cheong, S.Q. Liu, J. Yeo, H.K. Chionh, K. Pramudya, P. Curran, B. Yu, J. Essent. Oil Res. 23 (2011) 34.