Food Chemistry 143 (2014) 325–335

Contents lists available at ScienceDirect

Food Chemistry
journal homepage: www.elsevier.com/locate/foodchem

Differentiation of lemon essential oil based on volatile and non-volatile

fractions with various analytical techniques: a metabolomic approach
Florence Mehl a, Guillaume Marti a, Julien Boccard a, Benjamin Debrus a, Philippe Merle b, Estelle Delort b,
Lucie Baroux b, Vilfredo Raymo c, Maria Inés Velazco b, Horst Sommer b, Jean-Luc Wolfender a,
Serge Rudaz a,⇑
a
School of Pharmaceutical Sciences, University of Geneva, University of Lausanne, 30, Quai Ernest-Ansermet, 1211 Geneva 4, Switzerland
b
Firmenich, Corporate Research, Geneva, Switzerland
c
Simone Gatto Company, San Pier Niceto, Italy

a r t i c l e i n f o a b s t r a c t

Article history: Due to the importance of citrus lemon oil for the industry, fast and reliable analytical methods that allow
Received 21 December 2012 the authentication and/or classification of such oil, using the origin of production or extraction process,
Received in revised form 12 June 2013 are necessary. To evaluate the potential of volatile and non-volatile fractions for classification purposes,
Accepted 25 July 2013
volatile compounds of cold-pressed lemon oils were analyzed, using GC-FID/MS and FT-MIR, while the
Available online 3 August 2013
non-volatile residues were studied, using FT-MIR, 1H-NMR and UHPLC-TOF-MS. 64 Lemon oil samples
from Argentina, Spain and Italy were considered. Unsupervised and supervised multivariate analyses
Keywords:
were sequentially performed on various data blocks obtained by the above techniques. Successful data
Cold-pressed lemon oils
classification
treatments led to statistically significant models that discriminated and classified cold-pressed lemon oils
geographic origin according to their geographic origin, as well as their production processes. Studying the loadings allowed
production processes highlighting of important classes of discriminant variables that corresponded to putative or identified
GC-FID-MS chemical functions and compounds.
FT-MIR Ó 2013 Elsevier Ltd. All rights reserved.
1
H-NMR
UHPLC-TOF-MS
chemometrics
OPLS

1. Introduction oil, contains hydrocarbons, sterols, fatty acids, waxes, carotenoids,

Citrus is the most important fruit tree crop in the world, with an
annual production of approximately 102 million tons. Citrus limon
(L.) Burm. f. is considered the third most important citrus species
after orange and mandarin. Its production quantity is estimated
at 4.2 million tons. Argentina (1,400,000 t/year), Spain (750,000
t/year) and Italy (576,000 t/year) are among the major producers
of lemon (González-Molina, Domínguez-Perles, Moreno, & Gar-
cía-Viguera, 2010). Citrus essential oils are characterized by vola-
tile and non-volatile fractions that are composed of more than
200 compounds. Volatiles, which represent 85% to 99% of the entire
oil, are well characterized in the literature (Dugo & Mondello,
2010). This fraction is mainly composed of monoterpene and ses-
quiterpene hydrocarbons, their oxygenated derivatives and ali-
phatic aldehydes, alcohols and esters. The non-volatile residue of
cold-pressed lemon oils (CPLOR), which forms 1% to 15% of the
⇑ Corresponding author. Tel.: +41 22 3796336; fax: +41 22 3796808.
E-mail address: serge.rudaz@unige.ch (S. Rudaz).
coumarins, psoralens, and flavonoids (Dugo & Mondello, 2010).
Cold-pressed lemon oils (CPLOs) are widely used in the per-
fume, cosmetic, food and beverage and pharmaceutical industries
(González-Molina et al., 2010). CPLOs can be extracted by different
extraction processes: oil recovery from peel after juice extraction
(Sfumatrice), simultaneous extraction of juice and oil emulsion
from whole fruit (Food Machinery Corporation (FMC) Inline Extrac-
tor) or recovery of oil from the peel flavedo after removal from the
whole fruit by abrasion or shaving (Brown Oil Extractor, BOE and
Pelatrice). Several techniques can be employed to analyze volatile
and non-volatile fractions of CPLOs. Although chromatographic
methods, such as gas chromatography (GC) (e.g., GC-MS (gas chro-
matography - mass spectrometry), GC-FID (gas chromatography
coupled with flame - ionization detection), GC-O (gas chromatog-
raphy - olfactometry)) and liquid chromatography (LC) (e.g.,
HPLC-UV (high performance liquid chromatography coupled with
ultra-violet detection), HPLC-MS (high performance liquid chroma-
tography - mass spectrometry)) are widely used, only a few publi-
cations describe the Fourier transform-middle infrared
spectrometry (FT-MIR) and 1H nuclear magnetic resonance

0308-8146/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.foodchem.2013.07.125
326 F. Mehl et al. / Food Chemistry 143 (2014) 325–335

(1H-NMR) for the analyses of citrus cold-pressed oils (Schulz & Bar- spectroscopy has been used to qualitatively and quantitatively
anska, 2007; Sommer et al., 2003; Tomi, Barzalona, Casanova, & characterize essential oils (Rivero-Cruz, Rivero-Cruz, Rodríguez,
Luro, 2008). Cerda-García-Rojas, & Mata, 2006) but has been infrequently
GC-FID and GC-MS are the golden standard techniques for ana- used to analyze citrus oils (Tomi et al., 2008). Sommer et al.
lyzing the volatile fractions of essential oils. They are sometimes (2003) analyzed CPLORs, using HPLC-NMR and MS to identify
associated with other spectroscopic techniques, including FT- coumarins and furocoumarins. Because of its high sensitivity,
MIR, which is a versatile technique with powerful qualitative and UHPLC-TOF-MS is a well-suited technique for the analysis of
quantitative aspects. Moreover, it is a cheap, fast and reproducible the non-volatile fraction of lemon oils. The coupling of UHPLC
analytical method. Attenuated total reflectance (ATR) sampling al- with MS is particularly attractive in terms of detection, potential
lows well-resolved FT-MIR spectra of the oil to be obtained with- quantification and identification of a wide range of natural prod-
out any specific sample treatment or preparation. FT-MIR spectra, ucts (Eugster et al., 2011; Guillarme & Veuthey, 2012).
recorded in the wave number range between 650 and 4000 cm-1, As CPLOs are employed in a wide variety of products, their com-
provide a high level of specific molecular information and allow position must be carefully characterized to ensure quality, genu-
highlighting of characteristic vibrational bands that correspond ineness and absence of contaminations or to determine their
to various chemical functional groups and families (Schulz & Bara- geographic origin. In particular, present-day adulterations of citrus
nska, 2007). FT-MIR associated with chemometrics was previously oils are becoming increasingly difficult to reveal and highly selec-
used to analyze natural products such as vegetable oils and essen- tive and sensitive analytical methodologies are therefore required
tial oils (Charve, Chen, Hegeman, & Reineccius, 2011; Dupuy, (Tranchida, Bonaccorsi, Dugo, Mondello, & Dugo, 2012). Due to dif-
Molinet, Mehl, Nanlohy, Le Dréau, & Kister, 2013) to follow aging or ferences of quality between essential oils from different geographic
for authentication (e.g., determination of geographic origin, chemical origins, varieties of fruit or extraction processes, and the impor-
composition). To the best of our knowledge, the only attempt to com- tance of lemon oil in the industrial context, it is essential to devel-
bine FT-MIR and NIR for characterizing the chemical composition of op methodologies that are able to reliably differentiate and classify
citrus oils was reported by Schulz and Baranska (2007). CPLO samples.
The non-volatile fraction of CPLOs (Cold Pressed Lemon Oil The aim of this study was the development and evaluation of ana-
Residue, CPLOR) is also of great interest for the cosmetic indus- lytical methods, including sample preparation, for the discrimina-
try and human health. Furocoumarins possess phototoxic proper- tion and/or classification of lemon essential oils according to their
ties and they are unfortunately believed to cause mutagenesis, geographic origin and extraction process. A multi-approach method,
carcinogenesis, and photodermatitis (Desmortreux, Rothaupt, similar to that developed for metabolomics, was considered to dif-
West, & Lesellier, 2009). Conversely, polymethoxylated flavones ferentiate the samples (Fig. 1). The volatile and oxygen heterocyclic
display antioxidative, anti-mutagenic, and antitumor properties, fractions of the lemon oils were independently analyzed using spec-
and may therefore possess chemopreventive potential (Weber tral (i.e., FT-MIR, NMR) and separation (i.e., GC-FID, UHPLC-TOF-MS)
et al., 2005). In this study, CPLOR were analyzed by FT-MIR, techniques. The data obtained using these techniques were investi-
1
H-NMR and UHPLC-TOF-MS. 1H-NMR spectra provide informa- gated with various chemometric tools. Data sets were first explored
tion on a wide range of compounds that are present in a com- with non-supervised multivariate analysis (Principal Component
plex matrix in a single experiment. This method is simple, Analysis (PCA) and Hierachical Cluster Analysis (HCA)) to gain an
comprehensive, reproducible for both long- and short-term stud- overview and detect possible outliers. Then, orthogonal partial least
ies and requires very limited sample preparation. It has been ap- square discriminant analysis (OPLS-DA) was carried out to differen-
plied for many applications related to plant, natural product tiate classes of samples and highlight biomarkers.
quality control and metabolomics (Eugster, Guillarme, Rudaz,
Veuthey, Carrupt, & Wolfender, 2011), while high-resolution 2. Materials and methods
1
H-NMR has been used, in combination with chemometrics, to
address various issues of food authenticity and origin. This tech- 2.1. Sample set
nique is commonly used to authenticate natural products, such
as virgin olive oil, honey or citrus juice, i.e., to classify samples The sample collection consisted of 64 samples of cold-pressed
according to geographic origin and detect adulteration. 1H-NMR lemon oil obtained by industrial extraction from 3 different

VOLATILES NON-VOLATILES
SPECTRAL METHOD 0,2
SPECTRAL METHOD
FT-MIR FT-MIR
Cold Pressed Cold Pressed Oil Residue
0,2
10 to 16 mg 0,1

Lemon Oil
0
0,1
4 000 3000 2000 1000

400 μL
SPECTRAL METHOD
0
4 000 3000 2000 1000

1H-NMR
Dissolution in CDCl3
at 10 mg/mL

Reducing procedure :
SEPARATIVE METHOD SEPARATIVE METHOD
GC-FID - Slow reduction of pressure
Abundance

280000 to avoid bumping

260000

240000
- Full vacuum for 4 hours UHPLC-TOF-MS
220000

200000
Dissolution in MeOH
180000

160000
at 1mg/mL
140000

120000

5.00 10.00 15.00 20.00 25.00 30.00 35.00 40.00 45.00 50.00 55.00 Time min

Fig. 1. Methodology used for the comprehensive characterization of volatile and non-volatile fractions of CPLOs.
 F. Mehl et al. / Food Chemistry 143 (2014) 325–335
geographic origins: Italy (n=46), Argentina (n=13) and Spain (n=5).
Italian lemon fruits were harvested from November 2010 to May
2011, while Argentinian lemon fruits were harvested in May and
June 2011. Spanish fruits were harvested from November 2010 to
May 2011. The samples were obtained from industrial suppliers
and were a mix of several cultivars. Because samples were ob-
tained from industrial suppliers who bought lemon fruits from dif-
ferent farmers, it was impossible to obtain monovarietal CPLOs.
Argentinian samples were obtained from Eureka, Lisbon, and
Genova cultivars, Spanish samples were obtained from Fino and
Verna cultivars, and Italian samples were obtained mainly from
Feminello but also from Commune, Santa Teresa, and Sfusato cul-
tivars. In this work, the geographic origin denomination did not
consider differences among the cultivars.
To distinguish between the different extraction processes,
emphasis was placed on the Italian sample set. The Italian sample
set consisted of 46 samples extracted with 4 main different pro-
cesses: BOE (Brown Oil Extractor) (n=11), Sfumatrice (n=11), Pelat-
rice (n=11), and FMC (FMCÓ extractor) (n=13).
The most common method is the FMC-type extractor, which is
widely used in orange processing worldwide. In this method, the
oil and juice are simultaneously extracted by a cutter tube that
pierces the fruit and removes the juice while the oil is extracted.
The separate juice and oil products are then piped away for further
processing and finishing (Reeve, 2005).
The Brown Oil Extractor (BOE or Brown) process is particularly
common in the United States and South America, and rare in Eur-
ope. It differs from FMC because the oil is extracted from lemons,
first, by gently puncturing the flavedo, or peel, with thousands of
stainless steel needle points. The oil sacks are then ruptured,
releasing the oil, which is subsequently captured in a water spray.
Next, this water spray is centrifuged to separate the water and to
polish and finish the oil (Reeve, 2005).
The third main method of processing involves Pelatrice-type
rasping equipment. This process is similar to the Brown process,
but rather than using stainless steel needle points to pierce the
fruit and remove the oil, the fruit is subjected to rolling disc graters
that allow the flavedo to be rasped (Reeve, 2005).
The last method, Sfumatrice, involves extracting the oil from
the half peels without juice. Peels fall into the oil extracting sec-
tion and are taken by a rotating drum that rolls and presses
them against a stationary profile. This causes a repeated defor-
mation of the half peel and a breakage of the oil sacks. Oil is col-
lected by a water rain provided through a set of sprayers. The
mixture of water and essential oil falls into a tank and is then
conveyed to the following stages of finishing and separation (Re-
eve, 2005).
The volatile and non-volatile fractions of the oil were subjected
to various analytical methods, as described above.
2.2. FT-MIR
Three spectra of each sample (one for the CPLOR) were re-
corded, using a PerkinElmer Spectrum One spectrometer equipped
with a DTGS detector, an Ever-Glo source, and a KBr/germanium
beam splitter. Samples were deposited without any preparation
on an ATR cell equipped with a diamond crystal. Spectra were re-
corded between 4000 cm-1 and 650 cm-1 with a nominal resolution
of 4 cm-1 and 16 co-added accumulations. Air was taken as refer-
ence for the background spectrum before each sample. Between
spectra, the ATR plate was cleaned in situ by scrubbing with etha-
nol solution. A background spectrum was regularly collected and
compared to the previous background spectrum to verify cleanli-
ness. This procedure was also used as instrument qualification.
327
2.3. GC-FID and GC-MS
The CPLO was injected into a GC 6890 (Agilent) equipped with a
double injector and two DB-1 capillary columns (60 m  0.25 mm,
film thickness 0.25 lm, J&W 122-1062). One column was con-
nected to an MS 5973N (Agilent) for identification, and the other
to an FID for quantitation. The temperature programme was as fol-
lows: 50 °C for 5 min, increased to 120 °C at a rate of 3 °C/min, then
increased to 250 °C at a rate of 5 °C/min, 5 min isothermal, then in-
creased to 300 °C at a rate of 15 °C/min, and then 20 min isother-
mal; split ratio, 1:50; injection volume, 0.2 ll; injector and
detector temperatures, both 250 °C; carrier gas, helium at constant
flow rates of 1.8 ml/min and 2.1 ml/min, respectively. The CPLO
was diluted 1:10 in dichloromethane before being submitted to
GC-FID-MS analysis. For all GC-MS analyses, mass spectra were
generated by EI at 70eV at a scan ranging from m/z: 29-250 during
the first 20 min and then at a scan ranging from m/z: 29-450.
Compounds were identified by analyzing and comparing mass
spectra and linear retention indices obtained with pure references.
Linear retention indices (LRIs) were determined after injection of a
series of n-alkanes (C5-C31) under similar conditions. GC-MS peaks
were identified and integrated using HP Chemstation software and
MS and LRI library (internal, Wiley and Nist05).
Quantitative analysis was performed using methyl octanoate as
internal standard and the response factor of each compound. Their
percentages were obtained from the FID area corrected with the
use of the response factors, previously measured with pure ana-
lytes under the same conditions or calculated according to de Saint
Laumer, Cicchetti, Merle, Egger, and Chaintreau (2010). The
relative response factors (RRFs) were calculated as follows:
RRF = (mcompound  areaISTD)/(mISTD  areacompound), where
mcompound and areacompound are the mass and corresponding
GC peak area of the analyte and mISTD and areaISTD are the mass
and GC peak area of the internal standard.
2.4. NMR
To analyze the non-volatile fraction of the cold-pressed lemon
oil, 400 ll (340 mg) of each sample was evaporated under vac-
uum at 40°C for 6 h (Genevac HT-4X, Ipswich, UK). CPLORs were
dissolved in CDCl3 at a constant concentration of 10 mg/ml and
transferred to 5 mm NMR tubes. 1H-NMR spectra were recorded
on a BrukerÒ 500 MHz AVANCE I NMR spectrometer at
500.13MHz (Bruker BioSpin AG, Fällanden, Switzerland) equipped
with a BBO probe. For each sample, 64 scans with 64 K data points
were collected, using a spectral width of 8,000 Hz with a relaxation
delay of 20 s, an acquisition time of 4.09 s and a pulse width of 11
ls (90° pulse). All spectra were recorded at a constant temperature
of 298 K. Each spectrum was imported into MestReNova software
7.1 (Mestrelab Research, Santiago, Spain) for data pre-processing.
CPLOR 1H NMR fingerprints were aligned to the residual signal of
CDCl3 at 7.26 ppm. The Whittaker smoother algorithm, in full auto
mode, was applied for baseline correction. The region from 7.25
ppm to 7.28 ppm was removed before data reduction, using a
bin width of 0.04 ppm. Spectra were then normalized to total area
and exported as .txt files to MS ExcelÒ prior to data analysis.
2.5. UHPLC-TOF-MS
An aliquot of each sample prepared for the NMR spectroscopy
was dissolved in methanol at 1mg/ml to obtain UHPLC-PDA-TOF-
MS profiles. MS data were acquired on a Micromass-LCT Premier
Time of Flight (TOF) mass spectrometer (Waters, MA, USA) with
an electrospray interface and coupled with an Acquity UPLC
328 F. Mehl et al. / Food Chemistry 143 (2014) 325–335
system (Waters, MA, USA). ESI conditions were as follows: capil-
lary voltage 2800 V, cone voltage 40 V, MCP detector voltage
2400 V, source temperature 120 °C, desolvation temperature
300 °C, cone gas flow 20 l/h, desolvation gas flow of 800 l/h. Detec-
tion was performed in positive ion mode (PI) with a m/z range 100–
1000 Da and a scan time of 0.30 s in W-mode. The MS was cali-
brated, using sodium formate, and leucine enkephalin was used
as an internal reference at 2 lg/ml and infused through the Lock
Spray™ probe at a flow rate of 10 ll/min by means of a second
LC pump (Shimadzu LC-10ADvp, Duisburg, Germany).
The separation was obtained using a short UPLC BEH Phenyl
Acquity column (50  1.0 mm i.d., 1.7 lm, Waters, MA, USA).
The mobile phase consisted of 0.1% formic acid (FA) in water
(phase A) and 0.1% FA in methanol (phase B). The linear gradient
programme was as follows: 35% to 95% B over 4.0 minutes, held
at 95% B for a further 1.1 min, then returned to initial conditions
(35% B) in 0.1 min for 1.1 minutes of equilibration before next anal-
ysis. The flow rate was 0.4 ml/min; column temperature was kept
at 60°C. The UV spectra were recorded from 200 to 450 nm with a
resolution of 1.2 nm.
UHPLC-TOF-MS acquisitions were pre-processed, using Marker-
Lynx 4.1 software (Waters, MA, USA) for mass signal extraction and
alignment from 0 to 5 minutes with m/z values from 100 to 1000
Da with the following parameters: noise elimination level was set
at 25.00, minimum peak intensity at 400 counts, a mass window of
0.05 and a retention time window of 0.10 minutes; isotopic peaks
were excluded and no internal standard was used.
2.6. Data treatment
Pre-treatment of FT-MIR data (Standard Normal Variate, SNV),
multivariate analysis by principal component analysis (PCA) and
hierarchical clustering analysis (HCA), as well as orthogonal pro-
jection to latent structures discriminant analysis (OPLS-DA), were
performed with SIMCA-P+ 12Ó (Umetrics, Sweden). Data were
UV- (unit variance) or Pareto-scaled and a seven-fold procedure,
using a default option of the software for cross-validation, was
applied. PCA models were evaluated according to the percentage
of explained initial variance R2X. OPLS models were evaluated by
the goodness of fit R2Y and the prediction ability Q2.
The typical structure of treated data obtained by various meth-
ods was transformed into a data table in which the values are the
features, peak areas or signal intensities at a given min, m/z, cm-1
or ppm. Multivariate data analyses were performed on each data-
set to differentiate samples according to their geographic origin
and extraction process. A typical data-mining workflow begins
with unsupervised analysis to provide a preliminary overview of
the dataset. Principal components analysis (PCA) was initially con-
ducted to select the principal components (PCs) that explained at
least 95% of the initial data variance (Jackson, 2003). This step
mainly consisted of data reduction because the non-selected PCs
corresponded to non-informative systematic variability. HCA was
then performed using scores of the samples, i.e., their coordinates
in the new subspace spanned by the selected PCs, as described by
Boccard et al. (2007). HCA was performed, using Euclidean dis-
tances and Ward’s aggregation method (Ward, 1963). This step is
useful to assess the samples’ distribution and detect potential out-
liers. In some cases, this facilitates a first understanding of the rela-
tionships between the samples. The next step consists of
supervised methods to take advantage of prior information for
the data set analysis. With this end in view, OPLS-DA classification
models were computed to separate the predictive from the orthog-
onal sources of variation of the datasets and highlight discriminant
variables (Bylesjö, Rantalainen, Cloarec, Nicholson, Holmes, & Try-
gg, 2006).
To facilitate interpretation of results obtained with spectral
methods, specific graphic representations were implemented.
Average spectra were colored according to the Variable Influence
on Projection (VIP) (Fig. 2 and Fig. 3), with the most important vari-
ables colored red. This graphic representation is useful to interpret
the result of the multivariate analysis in terms of spectral region
and therefore of chemical functions. The VIP values reflect the
importance of variables in supervised models, both with respect
to Y, i.e. its correlation with the responses, and with respect to X,
i.e. the projection quality (Wold, Ruhe, Wold, & Dunn, 1984).
3. Results and discussion
3.1. Methodology
The complete sample set consisted of samples of cold-pressed
lemon oils (CPLOs) from three of the most important supplier
countries of the world: Argentina, Italy and Spain.
To comprehensively analyze the CPLOs, volatile and non-vola-
tile fractions of the oils were separately subjected to spectroscopic
and separation techniques, following the methodology presented
in Fig. 1.
3.2. Spectral analytical techniques
3.2.1. FT-MIR
3.2.1.1. FT-MIR of CPLO. Three spectra of each sample were re-
corded, and the average spectra computed to build a data table
of 64 observations x 3350 variables. A SNV pre-treatment was per-
formed on this matrix before performing PCA with UV scaling. The
two first principal components represented 55.6% of the initial var-
iance. Based on the PCA score plot and the PCA/HCA dendrogram,
no clear clustering could be highlighted according to the three geo-
graphic origins. The sample set was considered homogeneous be-
cause no specific outlier was detected.
An OPLS-DA model was then built to differentiate the three
groups of samples, and the results are presented in Table 1. The
predictive score plot (Fig. 2.A) shows three well separated groups
(R2Y(cum) = 0.942), and this significant model was efficient for pre-
diction (Q2Y(cum) = 0.679). To evaluate the variable’s contribution,
the average spectrum was coloured according to the VIP. The load-
ings plots are shown in Fig. 2 (C and D). The first predictive latent
variable was mainly discriminant between the Italian (positive
score) and Argentinian samples (negative score), while the second
predictive component was useful to distinguish Spanish samples
(positive score) from the others (negative score). Colours coding,
according to VIP, allowed vibrational bands corresponding to dis-
criminant variables to be highlighted. Argentinian samples were
characterized from the others using the limonene (1658, 1645,
886, 797 cm-1) and sabinene contents (1653 and 861 cm-1), Span-
ish samples using terpinen-4-ol (836, 950, 1071, 1093, 1179, 2727
cm-1) and camphor (1745, 1449, 1220, 1300, 1094, 1079, 1046,
936, 926 cm-1), and Italian samples using a-pinene (1470, 1221,
1166, 1101, 1063, 953, 928, 787 cm-1) and a-terpinene (1461,
1360, 1235, 1170, 1042, 1021, 988, 823 cm-1).
3.2.1.2. FT-MIR of cold pressed lemon oil residues. A first study of the
obtained results showed that FT-MIR spectra of CPLOs were very
similar to the spectrum of limonene and IR spectra having intense
vibrational bands at 886 cm-1, as well as a group of bands between
2800 and 3000 cm-1. These bands were absent in the CPLOR spectra
in which a large band at 3400 cm-1, corresponding to OH vibration,
and bands at 1700 cm-1, corresponding to C-O vibration, were
dominant.
 F. Mehl et al. / Food Chemistry 143 (2014) 325–335 329

A D
40

35

30
Spain 50

40
Spain
25

20
30
Argentina
15 20

10
10

Cold Pressed Lemon Oil CPLO Residue

t[2]

t[2]
5

0
Italy 0

-5
-10

-10

-15 Argentina 0,2

-20

0,18 -20 -30 Italy

-25
-40

B E
-50 -40 -30 -20 -10 0 10 20 30 40 50
-70 -60 -50 -40 -30 -20 -10 0 10 20 30 40 50 60 70 t[1]

0,13
t[1]
0,15
Average spectrum of Italian samples Average spectrum of all samples
0,08 0,1

0,03
0,05

-0,02

0
-0,07

-0,05
-0,12

-0,1
-0,17

Average spectrum of Argentinian samples Average spectrum of all samples

-0,22 -0,15
4000
3900
3800
3700
3600
3500
3400
3300
3200
3100
3000
2900
2800
2700
2600
2500
2400
2300
2200
2100
2000
1900
1800
1700
1600
1500
1400
1300
1200
1100
1000

4000
3900
3800
3700
3600
3500
3400
3300
3200
3100
3000
2900
2800
2700
2600
2500
2400
2300
2200
2100
2000
1900
1800
1700
1600
1500
1400
1300
1200
1100
1000
900
800
700

900
800
700
600
0,19
Average spectrum of Spanish samples C 0,2
Average spectrum of all samples F
0,15
0,14

0,09 0,1

0,04
0,05

-0,01
0

-0,06

-0,05
-0,11

-0,1
-0,16

Average spectrum of Italian and Argentinian samples Average spectrum of all samples
-0,21 -0,15
4000
3900
3800
3700
3600
3500
3400
3300
3200
3100
3000
2900
2800
2700
2600
2500
2400
2300
2200
2100
2000
1900
1800
1700
1600
1500
1400
1300
1200
1100
1000

4000
3900
3800
3700
3600
3500
3400
3300
3200
3100
3000
2900
2800
2700
2600
2500
2400
2300
2200
2100
2000
1900
1800
1700
1600
1500
1400
1300
1200
1100
1000
900
800
700

900
800
700
600
Average spectrum VIP from 0 to 1 VIP from 1 to 1,2 VIP from 1,2 to 1,5 VIP from 1,5 to 1,7 VIP from 1,7 to 2 VIP from 2 to 2,5

Fig. 2. OPLS-DA scores plot (A and D) and average spectra coded from VIP and loadings for cold-pressed oil data (B (pq1) and C (pq2)) and cold-pressed oil residue data (E
(pq1) and F (pq2)).

A SNV signal filter pre-treatment was performed on the 64  and Spanish samples were vibrational bands at 3540-3440, 1750-
3400 data matrix obtained from CPLOR FT-MIR analysis. UV-scaled 1730, 1580-1575, 1130-1050, 990-980, and 750 cm-1.
PCA was performed, and the two first principal components repre- The loadings that were associated with the second predictive
sented 67.4% of the initial variance. Two samples were identified as component discriminated Italian samples (1450-1430, 1045-965,
being on the borderline of the Hotelling’s T2 95% confidence ellipse and 900-885 cm-1) from Argentinian, and Spanish samples
but were retained to maintain a consistent dataset common to all (2935-2920, 2855-2850, and 1472-1460 cm-1).
analytical techniques. PCA and HCA results showed that no rele- Italian samples were characterized by ketones, esters, alde-
vant clustering of samples according to the geographic origin could hydes (carbonyl band at 1750-1730 cm-1) and furocoumarins
be obtained based on these unsupervised methods. For that pur- (ether (1130-1050 cm-1), furan (750 cm-1), and aromatic bands
pose, OPLS-DA modelling was carried out, based on the sample (1580-1575 cm-1)). Argentinian samples were characterized by ali-
class labels (Table 1). Even if three clusters could be highlighted phatic compounds, such as terpenes with bands at 2970-2800,
on the predictive score plot (Fig. 2), the obtained model had a mod- 2935-2920, 1460-1440, and 800 cm-1. Spanish samples were puta-
erate predictive ability (R2Y(cum) = 0.796, Q2(cum) = 0.546). The tively characterized by aliphatic compounds that were mainly
groups of Italian and Spanish samples were separated from the terpenes (2980-2800 and 1100-980 cm-1) (Schulz, Schrader, Quili-
groups of Argentinian samples according to the first predictive la- tzsch, & Steuer, 2002).
tent variable (Fig. 2.D). The second predictive component was dis-
criminant for Italian/Argentinian and Spanish samples. Some
discriminant vibrational bands could be highlighted due to average 3.2.2. 1H-NMR
spectra with bands that were coloured according to the VIP No signal filter pre-treatment was performed on the 64 x 257
(Fig. 2.E and F). Vibrational bands are positive or negative, accord- data matrix obtained with 1H-NMR spectra of the total dataset.
ing to the sign of the loadings. PCA was performed with UV scaling and HCA calculated by Ward’s
By analyzing the loadings associated with the first predictive la- method, using 12 PCs carrying 95% of the total data variance. The
tent variable, Argentinian samples could be characterized by vibra- two first principal components represented 52.7% of the initial var-
tional bands at 2970-2800, 1460-1440, 1430-1425, and 800 cm-1. iance. A multivariate unsupervised analysis of 1H-NMR data did
On the same component, the characteristic loadings of Italian not allow the differentiation of samples according to their origin.
330 F. Mehl et al. / Food Chemistry 143 (2014) 325–335

Average spectrum of Argentinian samples

0,8

0,6

0,4

0,2

-0,2

-0,4
Spain
-0,6

Average spectrum of Italian samples -0,8

-1
10,1

9,7

9,3

8,9

8,5

8,1

7,7

7,2

6,8

6,4

6,0

5,6

5,2

4,8

4,4

4,0

3,6

3,2

2,8

2,4

2,0

1,6

1,2

0,8

0,4

0,0

t[2]
1
Argentina
Average spectrum of Spanish samples Italy
0,8

0,6

0,4

0,2
t[1]

-0,2

-0,4

-0,6

-0,8
Average spectrum of Argentinian and Italian samples
-1
10,1

9,7

9,3

8,9

8,5

8,1

7,7

7,2

6,8

6,4

6,0

5,6

5,2

4,8

4,4

4,0

3,6

3,2

2,8

2,4

2,0

1,6

1,2

0,8

0,4

0,0

Fig. 3. VIP and loadings (pq[1] and pq[2] vertically) of coded average 1H-NMR spectra.

Table 1
Results of OPLS-DA models for differentiation according to the geographic origin and extraction process.

Analytical technique Differentiation according to the geographic origin (n = 64) Differentiation according to the extraction process (n = 46)
R2X(cum) R2Y(cum) Q2(cum) LVs R2X(cum) R2Y(cum) Q2(cum) PCs
Spectral methods FT-MIR 0.759 0.942 0.679 2+5 non-significant
FT-MIR⁄ 0.951 0.796 0.546 2+7 non-significant
1
H-NMR⁄ 0.658 0.938 0.742 2+6 non-significant
Separative methods GC-FID 0.898 0.937 0.881 2+4 0.736 0.971 0.708 3+8
UHPLC-TOF-MS⁄ 0.636 0.939 0.750 2+5 0.663 0.882 0.490 3+5
⁄
CPLOR analyses.

A supervised OPLS-DA model was then built to take advantage were discriminated by signals at 1.4, 1.7, 2.2 to 2.3, 5.7, 5.9, 7.4,
of the class labels information (Table 1). Satisfactory model fit and 10.0 ppm, while Italian ones were characterized by signals at
(R2Y (cum) = 0.938) and predictive ability (Q2(cum) = 0.742) were 1.1, 2.1, 5.0, 5.5, 7.0, 7.2, 7.6, 8.0, and 8.2 ppm. Spanish samples
obtained. The three groups of samples could easily be distin- were characterized by signals at 1.3 to 1.4, 2.4, 5.5, 7.2, and 8.2
guished in the predictive OPLS-DA score plot (Fig. 3); Italian and ppm. These highly discriminant chemical shifts of the different
Argentinian samples were discriminated according to the first la- groups were mainly related to the same patterns. Signals between
tent variable, while Spanish samples were differentiated from 1 and 2 ppm could be assigned to methyl groups, while signals at
Argentinian and Italian ones along the second predictive compo- approximately 5.5 ppm could be related to unsaturated protons of
nent. VIP- and loadings-coded average 1H-NMR spectra are pre- aliphatic side chains. Chemical shifts between 7 and 8.2 ppm are
sented in Fig. 3. According to these results, Argentinian samples typical of aromatic protons. Taken together, these signals could
 F. Mehl et al. / Food Chemistry 143 (2014) 325–335
be assigned characteristic patterns of coumarin- and furocouma-
rin-related structures, which were confirmed by a careful inspec-
tion of 1H-NMR spectra in full resolution (Sommer et al., 2003).
Interestingly, signals above 8 ppm, that were characteristic of
Italian and Spanish samples, could be attributed to the proton in
position 4 of the pyranone moiety, typical of an O-substituent in
position 5 (Waight, Razdan, Qadri, & Harkar, 1987) of furocouma-
rins, thus indicating the abundance of these derivatives in these
two groups.
3.3. Discussion on spectral methods’ results
Unsupervised multivariate analyses of data obtained with spec-
tral methods did not allow good clustering of the samples accord-
ing to their geographic origin. In other words, this modality
represents neither a factor of major variability nor the first obser-
vable difference between the investigated samples. Because signif-
icant OPLS-DA models that were able to differentiate and classify
cold-pressed lemon oil samples according to their country of pro-
duction were obtained, minor discriminant features are however
present and could be captured with spectral analytical techniques.
Studying the loadings allows highlighting those features as chem-
ical families of compounds that are characteristic of each geo-
graphic origin.
Because of the high quality and value of the Italian cold-pressed
lemon oil for the flavour industry, the study then focussed on their
extraction processes. Data obtained from the sub sample set con-
sisting of 46 samples were subjected to multivariate analysis to
highlight differences between the four main methods used to ex-
tract the juice and oil from lemons. To investigate this aspect, Ital-
ian samples obtained from these four industrial extraction
processes, i.e., BOE (n=11), FMC (n=13), Sfumatrice (n=11), and Pel-
atrice (n=11), were considered an independent data sub-set. Clas-
sical multivariate data analyses, including unsupervised and
supervised methods, were applied to the spectral data, but no sat-
isfactory and significant model could be built to distinguish the
samples according to the extraction processes (Table 1).
3.4. Analytical separation techniques
3.4.1. General
Non-separation (or spectral) analytical methods enable the
rapid and high-throughput screening of samples, but they have
limited metabolite identification capabilities. Therefore, the use
of separation techniques (GC, LC or UHPLC) with conventional
or MS detection could significantly enhance the discrimination
power due to the identification of a wider range of features in
the sample. Notably, an important distinction could be made be-
tween separation and non-separation analytical techniques when
working with metabolomics. Spectral techniques, which yield re-
sponses that correspond to combinations of multiple features
that give a common signal, are mainly used for fingerprinting,
chemotaxonomy, or simple classification studies. For more com-
plex cases, investigation with the appropriate separation tech-
niques should be considered.
3.4.2. GC-FID
3.4.2.1. Differentiation. While the volatile composition part of
lemon essential oils has already been exhaustively studied by
GC-FID and GC-MS (Dugo & Mondello, 2010), only few studies were
published on the differentiation and classification of lemon oils
(Lota, de Rocca Serra, Tomi, Jacquemond, & Casanova, 2002). Di Vaio
et al. (2010) discriminated 18 lemon cultivars by their content in
some terpenes in the peel ethanol extract (Di Vaio, Graziani,
Gaspari, Scaglione, Nocerino, & Ritieni, 2010). However, to the best
of our knowledge, only one study has dealt with the differentiation
331
of lemon oil samples according to the geographic origin. This
work, published recently by Kostadinovic et al., (2010) is based
on GC analyses, using the relatively low number of 6 samples
(Kostadinovic, Stefova, Nikolova, Nedelcheva, Martinez, Lorenzo,
et al., 2010).
3.4.2.2. Differentiation according to geographic origin. Data obtained
by GC-FID analysis consisted of a 64  216 data matrix, which was
subjected to unsupervised multivariate analysis given that authors
of previous studies. Among these 216 variables, only 22 remained
unknown but were unambiguously labelled due to their EI-MS
spectra and retention indices (A list of identified compounds is
provided as supplementary material).
PCA was performed with Pareto scaling, selected to emphasize
the most intense variable changes. The first three PCs represented
74.6% of the initial variance and the PCA/HCA dendrogram showed
three principal clusters: a first large cluster containing almost all
Italian samples, a second cluster with the majority of the Argentin-
ian samples and a third one containing Spanish samples and some
Italian, and Argentinian oils (Data not shown). These results high-
lighted the difference of geographic origin as the principal source
of variability. Thus, it could be considered that the variability
was captured by the most intense variables of the GC data set.
An OPLS-DA model was then built, to provide deeper insight
into this source of variability (Table 1). The model was highly
significant and presented a good model fit (R2Y(cum) = 0.937)
and a high predictive ability (Q2(cum) = 0.881). Based on these
results, it can be hypothesized that the volatile fraction of the
lemon oil contains unambiguous information to clearly differenti-
ate samples according to their origin. Because of the high accuracy
of the technique, discriminant volatile compounds were easily
identified. Compound identification was achieved using GC-MS
analysis and retention index calculation. GC-MS and GC-FID
analyses were under strictly identical experimental conditions,
which allowed us to compare both chromatograms. GC-EI-MS
analysis enables the use of MS libraries to identify peaks.
Calculating retention indices allows confirmation of the
compounds identified by GC-MS analysis. Moreover, GC-FID allows
accurate quantification, using an internal standard, which could be
of interest if a quality control process is established.
Three binary OPLS-DA models were then built to more readily
highlight discriminant compounds with a two-class comparison,
and relevant loadings were selected, based on S-plots.
To highlight the contribution of discriminant variables when
two discriminant models were simultaneously considered, the cor-
relation vector corr(tp,X) from the predictive component (first la-
tent variable) of each model was extracted to build a shared and
unique structure plot (SUS-plot, Fig. 4). This representation aims
to evaluate the contributions of variables to different classification
models by comparison with a common reference (Wiklund et al.,
2007)
Spanish samples were mainly characterized by their limonene,
camphor, and 4-terpineol contents, Argentinian oils by their
sabinene, limonene, and cis-sabinene hydrate contents, and Italian
samples by their a-thujene, a-pinene, a-terpinene, sesquiterpenes
(i.e., b-caryophyllene), and furocoumarins (i.e., oxypeucedanin)
contents. These results are consistent with those obtained with
FT-MIR data. This result is particularly interesting because it
indicates that an easy-to-use, but less sensitive, spectroscopic
technique can be related to a sensitive, but time-consuming,
analytical separation technique. For establishing a quality control,
this result could be used to study the chemical composition and
origin of the CPLO.
3.4.2.3. Differentiation according to the extraction process. Various
attempts to differentiate Italian samples (n=46), according to their
332 F. Mehl et al. / Food Chemistry 143 (2014) 325–335

UHPLC-TOF-MS
Non-volatile part of the oil
Terpenes GC-FID
Spain Volatile part of the oil

30
0,8
Spain Sesquiterpenes

25
20
15
Furocoumarins

10
5
t[1]
0
1,0
0,6
ITALY

-5
ITALY ...
200

Spain

-10
0,9 150

-15
-20
100

0,8

25

20

15

10

-5

-10

-15

-20

-25
50
to[1]

0,7
0,4

to[1]
-50

-100
0,6
-150

0,5 -200

-250
ITALY
0,2 0,4 -300 -250 -200 -150 -100 -50 0 50
t[1]
100 150 200 250 300 350 400

0,3

0,2

-0,0 ARGENTINA ITALY 0,1

0,0
ITALY Spain
-0,1
-0,2
-0,2
25

-0,3
20
A

15
IN

10

300
-0,4
-0,4
Furocoumarins
NT

250
5

ITALY ITALY

200
to[1]

-0,5
GE

150
-5

100
Coumarins
-10
AR

-0,6

50
-15

t[1]
ITALY

0
-20

-50
-0,6 -25 -0,7

-100
ARGENTINA
-35 -30 -25 -20 -15 -10 -5 0 5 10 15 20 25 30 35

-150
t[1]

-0,8

ARGENTINA

-200
-0,8 -0,6 -0,4 -0,2 -0,0 0,2 0,4 0,6 0,8

-250
...
-0,9

-300
350

300

250

200

150

100

50

-50

-100

-150

-200

-250

-300
to[1]

-0,9 -0,8 -0,7 -0,6 -0,5 -0,4 -0,3 -0,2 -0,1 0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8

Fig. 4. OPLS-DA score plots of models of Italy vs. Argentina and Italy vs. Spain and the SUS-plots obtained by these models from UHPLC-TOF-MS and GC-FID data.

extraction process with unsupervised multivariate analysis (UV 3.4.3. UHPLC-TOF-MS

scaled), were unsuccessful. However, a significant classification 3.4.3.1. General. In metabolomics studies, access to fast chromatog-
model was obtained with OPLS-DA (R2Y(cum) = 0.971 and raphy methods that provide high separation resolution and high
Q2(cum) = 0.708) (Fig. 5). UV scaling was used to enhance the repeatability is a key to generating valuable datasets. In this re-
importance of less intense variables. Four OPLS-DA models were spect, the introduction of ultra-high pressure liquid chromatogra-
then built to highlight the relevant variables that characterized phy (UHPLC) systems, operating at very high pressures and using
each class: (i) models comparing Pelatrice vs. others, (ii) BOE vs. sub-2 lm packing columns, has allowed a remarkable decrease
others, (iii) FMC vs. others and (iv) Sfumatrice vs. others. The four in analysis time and increase in peak capacity, sensitivity and
extraction processes were well separated and formed subgroups. reproducibility compared with conventional HPLC. This technology
Samples extracted using the BOE and Pelatrice techniques were has rapidly been widely accepted by the analytical community and
closely grouped, while samples extracted using Sfumatrice and is being gradually applied to various fields of plant analysis
FMC constituted a second cluster. This result was consistent with (Eugster et al., 2011), especially in metabolomics. Oxygen hetero-
the different principles of each process. BOE and Pelatrice pro- cyclic components of lemon oil have been mainly studied using
cesses involve rasping or puncturing the peel before extracting HPLC-UV and HPLC-MS for oxygen heterocyclic components
the juice while, in FMC and Sfumatrice, the oil is extracted from characterization and quantification (Desmortreux etal., 2009; Dugo
the lemon peel simultaneous with, or after, the extraction of the et al., 2010; Frérot & Decorzant, 2004; Tranchida et al., 2012). To
juice. our knowledge, there are no published studies on the non-volatile
Characteristic compounds of the FMC extraction process were fraction of lemon oil analyzed by UHPLC-TOF-MS, which could
citronellyl propanoate, hexanol, and cis-piperitol. Nonanoic acid, potentially be a method of choice to characterize the non-volatile
decanol, 1,4-p-menthadien-7-al, perillyl acetate, b-copaene, ner- fraction of the CPLO, as it can detect even low abundant
olidol, caryophyllene, and tetradecanal were related to the BOE components.
process. The Pelatrice process was characterized by hexanal,
trans-2-hexenal, geraniol, nootkatone, valencene, trans-b-ocim-
ene, c-terpinene, and geranyl acetate. Finally, lemon oils ex- 3.4.3.2. Differentiation according to geographic origin. Unsupervised
tracted by Sfumatrice contained mainly heptanal, myrcene, multivariate analysis (PCA and HCA with Pareto scaling) was per-
phellandrene, iso-citral, 4-terpineol, citroptene, neral, and gera- formed on the UHPLC-TOF-MS (positive mode) data, which con-
nial. The concentrations of characteristic compounds were higher sisted of a 64  1393 data matrix. No clear clustering of the
in one oil, but they could also be present in other oils. Previ- samples could be observed according to their origin. Both ioniza-
ously, the influence of the extraction technologies on lemon oil tion modes were assessed; however, the positive ionization mode
chemical composition has scarcely been reported (Verzera, Dugo, was found to be more informative than the negative ionization
Mondello, Trozzi, & Cotroneo, 1999). According to Verzera et al., mode; ten times more features were detected than in the negative
(1999) the Pelatrice process produces lemon oils with richer mode. Due to the high sensitivity and resolution of UHPLC-TOF-MS,
oxygenated compound content, an observation that is consistent several features were detected. This large number of features could
with our results. Additionally, they reported a poorer oxygenated also be explained by the presence of sodium adducts and complex
compound content in oils obtained by the Sfumatrice process ions ([2M+Na]+, [2M+H]+). To study the most significant and rele-
and more pleasant odour notes. Again, these considerations are vant features, the data were Pareto-scaled. A supervised multivar-
consistent with our results because the Sfumatrice samples con- iate analysis was then carried out, and a significant OPLS-DA model
tained key odour components, such as neral and geranial. was obtained (R2Y(cum)= 0.939, Q2(cum)= 0.750).
 F. Mehl et al. / Food Chemistry 143 (2014) 325–335 333

GC-FID
Volatile part of the oil

FMC
Pelatrice

Sfumatrice
BOE Pelatrice BOE Sfumatrice FMC

UHPLC-TOF-MS
Non-volatile part of the oil

Sfumatrice FMC Sfumatrice

FMC
BOE Pelatrice

Pelatrice BOE

Fig. 5. OLPS-DA score plots and OPLS-DA/HCA dendrograms obtained by GC-FID and UHPLC-TOF-MS data.

Three binary OPLS-DA models (Pareto scaling) were then built
to determine which variables were discriminant for each class,
namely, Argentina vs. Italy, Italy vs. Spain and Spain vs. Argen-
tina. Italian lemon oil samples were considered a common refer-
ence, and a SUS-plot was built, using the two models that
involved Italy (Italy vs. Argentina and Italy vs. Spain) for discrim-
inant variable detection (Fig. 4). Most discriminant features
could be attributed to furocoumarin derivatives, based on high
resolution mass spectra and retention time relative to reference
compounds provided by Firmenich SA. Italian samples were
characterized by their high bergamottin content, while imperato-
rin and byakangelicol were found in the highest quantities in
Spanish and Argentinian oils, respectively. These results are con-
sistent with the chemical information obtained from the analysis
of CPLOR by FT-MIR and NMR. A study is underway to clearly
identify these discriminant compounds from Italian and Argen-
tina samples by 1H-NMR and UHPLC-TOF-MS.
3.4.3.3. Differentiation according to the process. Focussing on the
Italian sample set, a four-class OPLS-DA model (Pareto scaling)
334 F. Mehl et al. / Food Chemistry 143 (2014) 325–335
was built and a satisfactory fit (R2Y(cum) = 0.882) with a poor pre-
dictive ability (Q2Y(cum) = 0.490) was obtained. The four classes of
extraction process were well separated, as shown by the OPLS-DA/
HCA dendrogram (Fig. 5).
The BOE- and Pelatrice-extracted samples were close to each
other, but they were separated from FMC- and Sfumatrice-ex-
tracted samples, as previously observed by GC-FID analysis. There-
fore, the information that was obtained from the separation
techniques is independent of the nature of the studied sample (vol-
atile/non-volatile).
Four OPLS-DA models were built to detect the characteristic
variables of each class, namely, Pelatrice vs. others, BOE vs. oth-
ers, FMC vs. others and Sfumatrice vs. others. According to the S-
plot of each OPLS-DA, the furocoumarin heraclenin was mainly
found in the Pelatrice-extracted samples, while bergamottin
was present in high quantities in the Sfumatrice-processed sam-
ples. Additionally, the FMC-extracted samples were characterized
by hydroxymethylherniarin, and the BOE-extracted samples dis-
played the highest unsaturated fatty acid content.
4. Conclusion
The results obtained in this study showed that spectral meth-
ods, coupled with chemometric tools, were able to differentiate
and classify samples of lemon oils according to their geographic
origin (Italy, Argentina or Spain). Spectral approaches (FT-MIR
and 1H-NMR) are highly reproducible and fast compared to separa-
tion techniques (GC-FID and UHPLC-TOF-MS). The latter are sensi-
tive and, in some cases, allow direct compound identification.
Moreover, because 1H-NMR spectra are directly related to the
absolute quantity of compounds, quantitative NMR spectroscopy
has been applied to a mixture of compounds to determine the rel-
ative content of each metabolite of interest (Wishart, 2008).
Information obtained, using separation techniques treated with
chemometric tools, permitted the differentiation of Italian samples
according to their extraction process. The speed with which FT-
MIR, 1H-NMR spectra and UHPLC-TOF-MS profiles can be obtained,
particularly by automated processes, enables the examination of
many samples. Because the richness of information often results
in high spectral complexity, the use of multivariate analysis is
mandatory to study large numbers of spectra and extract meaning-
ful information.
The study of loadings revealed that the GC-FID-MS, UHPLC-TOF-
MS and FT-MIR results were consistent. It was shown that studying
the non-volatile fractions of cold-pressed lemon oils lead to very
good differentiation and classification of samples, regarding the
geographic and extraction process modalities. The study of load-
ings highlighted chemical families or compounds that were charac-
teristic of each quality of lemon oils. It could be demonstrated that
Italian CPLOs were enriched in a-thujene, a-pinene, a-terpinene,
sesquiterpenes (i.e., b-caryophyllene) and furocoumarins (i.e.,
bergamottin). Spanish and Argentinian CPLOs were characterized
by their terpene contents, such as limonene, but differed in their
imperatorin, and byakangelicol contents. Moreover, Spanish CPLOs
contained more camphor and 4-terpineol, while Argentinian CPLOs
contained more sabinene and cis-sabinene hydrate. This work con-
stitutes an important improvement in cold-pressed lemon oil qual-
ity and authenticity analysis. To improve classification models, it
would be necessary to obtain more samples from Argentina and
Spain. The variety of spectral and separation techniques and the
consistency of the results allowed identification of many important
biomarkers from a wide range of volatile and non-volatile com-
pounds of various chemical families.
Acknowledgement
The authors acknowledge Dr Laurent Wünsche for his valuable
scientific advice.
Appendix A. Supplementary data
Supplementary data associated with this article can be found, in
the online version, at http://dx.doi.org/10.1016/j.foodchem.2013.
07.125.
References
Boccard, J., Grata, E., Thiocone, A., Gauvrit, J.-Y., Lantéri, P., Carrupt, P.-A., Wolfender,
J.-L., & Rudaz, S. (2007). Multivariate data analysis of rapid LC-TOF/MS
experiments from Arabidopsis thaliana stressed by wounding. Chemometrics
and Intelligent Laboratory Systems, 86(2), 189–197.
Bylesjö, M., Rantalainen, M., Cloarec, O., Nicholson, J. K., Holmes, E., & Trygg, J.
(2006). OPLS discriminant analysis: combining the strengths of PLS-DA and
SIMCA classification. Journal of Chemometrics, 20(8–10), 341–351.
Charve, J., Chen, C., Hegeman, A. D., & Reineccius, G. A. (2011). Evaluation of
instrumental methods for the untargeted analysis of chemical stimuli of orange
juice flavour. Flavour and Fragrance Journal, 26(6), 429–440.
de Saint Laumer, J.-Y., Cicchetti, E., Merle, P., Egger, J., & Chaintreau, A. (2010).
Quantification in Gas Chromatography: Prediction of Flame Ionization Detector
Response Factors from Combustion Enthalpies and Molecular Structures.
Analytical Chemistry, 82(15), 6457–6462.
Desmortreux, C., Rothaupt, M., West, C., & Lesellier, E. (2009). Improved separation
of furocoumarins of essential oils by supercritical fluid chromatography. Journal
of Chromatography A, 1216(42), 7088–7095.
Di Vaio, C., Graziani, G., Gaspari, A., Scaglione, G., Nocerino, S., & Ritieni, A. (2010).
Essential oils content and antioxidant properties of peel ethanol extract in 18
lemon cultivars. Scientia Horticulturae, 126(1), 50–55.
Dugo, G., & Mondello, L. (2010). Citrus Oils: Composition, Advanced Analytical
Techniques, Contaminants, and Biological Activity. Boca Raton: CRC Press.
Dugo, P., Ragonese, C., Russo, M., Sciarrone, D., Santi, L., Cotroneo, A., & Mondello, L.
(2010). Sicilian lemon oil: Composition of volatile and oxygen heterocyclic
fractions and enantiomeric distribution of volatile components. Journal of
Separation Science, 33(21), 3374–3385.
Dupuy, N., Molinet, J., Mehl, F., Nanlohy, F., Le Dréau, Y., & Kister, J. (2013).
Chemometric analysis of mid infrared and gas chromatography data of
Indonesian nutmeg essential oils. Industrial Crops and Products, 43, 596–601.
Eugster, P. J., Guillarme, D., Rudaz, S., Veuthey, J.-L., Carrupt, P.-A., & Wolfender, J.-L.
(2011). Ultra High Pressure Liquid Chromatography for Crude Plant Extract
Profiling. Journal of AOAC International, 94(1), 51–70.
Frérot, E., & Decorzant, E. (2004). Quantification of Total Furocoumarins in Citrus
Oils by HPLC Coupled with UV, Fluorescence, and Mass Detection. Journal of
Agricultural and Food Chemistry, 52(23), 6879–6886.
González-Molina, E., Domínguez-Perles, R., Moreno, D. A., & García-Viguera, C.
(2010). Natural bioactive compounds of Citrus limon for food and health.
Journal of Pharmaceutical and Biomedical Analysis, 51(2), 327–345.
Guillarme, D., & Veuthey, J.-L. (2012). UHPLC in Life Sciences (Vol. 16). Cambridge:
The Royal Society of Chemistry.
Jackson, J. E. (2003). A User’s Guide to Principal Components. New York: John Wiley
and Sons.
Kostadinovic, S., Stefova, M., Nikolova, D., Nedelcheva, D., Martinez, N., Lorenzo, D.,
Dellacassa, E., & Mirhosseini, H. (2010). Multivariate analysis discrimination of
various cold-pressed lemon oils from different geographical regions. Journal of
Food, Agriculture & Environment, 8(2), 117–121.
Lota, M.-L., de Rocca Serra, D., Tomi, F., Jacquemond, C., & Casanova, J. (2002).
Volatile Components of Peel and Leaf Oils of Lemon and Lime Species. Journal of
Agricultural and Food Chemistry, 50(4), 796–805.
Reeve, D. (2005). Citrus focus. A Cultivated Zest. The lemon’s origins, production
and processing. Perfumer & Flavorist, 30(3), 32–35.
Rivero-Cruz, B., Rivero-Cruz, I., Rodríguez, J. M., Cerda-García-Rojas, C. M., & Mata, R.
(2006). Qualitative and Quantitative Analysis of the Active Components of the
Essential Oil from Brickellia veronicaefolia by Nuclear Magnetic Resonance
Spectroscopy. Journal of Natural Products, 69(8), 1172–1176.
Schulz, H., & Baranska, M. (2007). Identification and quantification of valuable plant
substances by IR and Raman spectroscopy. Vibrational Spectroscopy, 43(1),
13–25.
Schulz, H., Schrader, B., Quilitzsch, R., & Steuer, B. (2002). Quantitative Analysis of
Various Citrus Oils by ATR/FT-IR and NIR-FT Raman Spectroscopy. Applied
Spectroscopy, 56(1), 117–124.
Sommer, H., Bertram, H.-J., Krammer, G., & KIindel, G. (2003). HPLC-NMR: a
powerful tool for the identification of non-volatiles in lemon peel oils. Perfumer
& Flavorist, 28(1), 16.
Tomi, F., Barzalona, M., Casanova, J., & Luro, F. (2008). Chemical variability of the leaf
oil of 113 hybrids from Citrus clementina (Commun)  Citrus deliciosa (Willow
Leaf). Flavour and Fragrance Journal, 23(3), 152–163.
 F. Mehl et al. / Food Chemistry 143 (2014) 325–335
Tranchida, P. Q., Bonaccorsi, I., Dugo, P., Mondello, L., & Dugo, G. (2012). Analysis of
Citrus essential oils: state of the art and future perspectives. A review. Flavour
and Fragrance Journal, 27(2), 98–123.
Verzera, A., Dugo, P., Mondello, L., Trozzi, A., & Cotroneo, A. (1999). Extraction
technology and lemon oil composition. Italian Journal of Food Science, 11(4),
361–370.
Waight, E., Razdan, T., Qadri, B., & Harkar, S. (1987). Chromones and coumarins from
Skimmia laureola. Phytochemistry, 26(7), 2063–2069.
Ward, J. H. (1963). Hierarchical Grouping to Optimize an Objective Function. Journal
of the American Statistical Association, 58(301), 236–244.
Weber, B., Hartmann, B., Stöckigt, D., Schreiber, K., Roloff, M., Bertram, H.-J., &
Schmidt, C. O. (2005). Liquid Chromatography/Mass Spectrometry and Liquid
Chromatography/Nuclear Magnetic Resonance as Complementary Analytical
335
Techniques for Unambiguous Identification of Polymethoxylated Flavones in
Residues from Molecular Distillation of Orange Peel Oils (Citrus sinensis).
Journal of Agricultural and Food Chemistry, 54(2), 274–278.
Wiklund, S., Johansson, E., Sjostrom, L., Mellerowicz, E. J., Edlund, U., Shockcor, J. P.,
Gottfries, J., Moritz, T., & Trygg, J. (2007). Visualization of GC/TOF-MS-Based
Metabolomics Data for Identification of Biochemically Interesting Compounds
Using OPLS Class Models. Analytical Chemistry, 80(1), 115–122.
Wishart, D. S. (2008). Quantitative metabolomics using NMR. Trends in Analytical
Chemistry, 27(3), 228–237.
Wold, S., Ruhe, A., Wold, H., & Dunn, I. W (1984). The Collinearity Problem in Linear
Regression. The Partial Least Squares (PLS) Approach to Generalized Inverses.
SIAM Journal on Scientific and Statistical Computing, 5(3), 735–743.