Detection and quantification of extra virgin olive oil adulteration by means of autofluorescence excitation-emission profiles combined with multi-way classification

Author’s Accepted Manuscript DETECTION AND QUANTIFICATION OF EXTRA VIRGIN OLIVE OIL ADULTERATION BY MEANS OF AUTOFLUORESCENCE EXCITATION-EMISSION PROFILES COMBINED WITH MULTI-WAY CLASSIFICATION Isabel Durán Merás, Jaime Domínguez Manzano, Diego Airado Rodríguez, Arsenio Muñoz de la Penala Peña

PII: DOI: Reference:

www.elsevier.com/locate/talanta

S0039-9140(17)31039-1 https://doi.org/10.1016/j.talanta.2017.09.095 TAL17998

To appear in: Talanta Received date: 15 May 2017 Revised date: 27 September 2017 Accepted date: 30 September 2017 Cite this article as: Isabel Durán Merás, Jaime Domínguez Manzano, Diego Airado Rodríguez and Arsenio Muñoz de la Penala Peña, DETECTION AND QUANTIFICATION OF EXTRA VIRGIN OLIVE OIL ADULTERATION BY MEANS OF AUTOFLUORESCENCE EXCITATION-EMISSION PROFILES COMBINED WITH MULTI-WAY CLASSIFICATION, Talanta, https://doi.org/10.1016/j.talanta.2017.09.095 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting galley proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

DETECTION AND QUANTIFICATION OF EXTRA VIRGIN OLIVE OIL ADULTERATION BY MEANS OF AUTOFLUORESCENCE EXCITATIONEMISSION PROFILES COMBINED WITH MULTI-WAY CLASSIFICATION

Isabel Durán Merás*a, b, Jaime Domínguez Manzanoa, Diego Airado Rodríguezc, Arsenio Muñoz de la Peñaa, b

a

Department of Analytical, University of Extremadura, 06006 Badajoz, Spain

b

Research Institute on Water, Climate Change and Sustainability (IACYS), University of Extremadura, 06006 Badajoz, Spain c

Department of Science and Mathematics Education, University of Extremadura, 06006 Badajoz, Spain

*Corresponding author: [email protected]

1

ABSTRACT Within olive oils, extra virgin olive oil is the highest quality and, in consequence, the most expensive one. Because of that, it is common that some merchants attempt to take economic advantage by mixing it up with other less expensive oils, like olive oil or olive pomace oil. In consequence, the characterization and authentication of extra virgin olive oils is a subject of great interest, both for industry and consumers. This paper reports the potential of front-face total fluorescence spectroscopy combined with second-order chemometric methods for the detection of extra virgin olive oils adulteration with other olive oils. Excitation-emission matrices (EEMs) of extra virgin olive oils and extra virgin olive oils adulterated with olive oils or with olive pomace oils were recorded using front-face fluorescence spectroscopy. The full information content in these fluorescence images was analyzed with the aid of unsupervised parallel factor analysis (PARAFAC), PARAFAC supervised by linear discriminant analysis (LDA-PARAFAC), and discriminant unfolded partial leastsquares (DA-UPLS). The discriminant ability of LDA-PARAFAC was studied through the tridimensional plots of the canonical vectors, defining a surface separating the established categories. For DA-UPLS, the discriminant ability was established through the bidimensional plots of predicted values of calibration and validation samples, in order to assign each sample to a given class. The models demonstrated the possibility of detecting adulterations of extra virgin olive oils with percentages of around 15% and 3% of olive and olive pomace oils, respectively. Also, UPLS regression was used to quantify the adulteration level of extra virgin olive oils with olive oils or with olive pomace oils.

Keywords:

Front-face

fluorescence,

Olive

discrimination

2

oils,

Adulteration,

Chemometric

1-Introduction Virgin olive oil is characterized by very pleasant sensory properties as well as important health benefits because of its singular composition of unsaturated fatty acids, polyphenols, vitamin E, carotenoids, sterols and chlorophylls. Consequently, it is an economically important product in Mediterranean countries. The International Olive Council (IOC) distinguishes between extra virgin olive oil (EVOO) and virgin olive oil (VOO). In both cases, olive oil is obtained from the fruit of the olive tree (Olea europaea L.) exclusively by means of mechanical, such as pressure, or other physical procedures under conditions, particularly thermal conditions, not leading to significant alterations in the oil. These oils are not submitted to any other treatments than washing, decantation, centrifugation and filtration. The differences between EVOO and VOO mainly lie in their free acidity. In the case of EVOO, free acidity is not higher than 0.8 grams per 100 grams, while for VOO it uses to be below 2 grams per 100 grams. Other commercially available olive oils are: refined olive oil (ROO); olive oil (OO); and olive pomace oil (OPO). Refined olive oil is obtained from virgin olive oils submitted to several refining methods, which do not lead to alterations in the initial glyceridic structure. Olive oil is a blend of refined olive oil and virgin olive oil suitable for consumption without further processing, and olive pomace oil is obtained by treating olive pomace with solvents or by means of other physical treatments, excluding reesterification processes. Extra virgin olive oil is the one possessing the highest quality in terms of aromaticity and flavor, but its production is very limited, and the high demand from consumers makes it susceptible to be adulterated with cheaper seed oils or even with other olive oils of lower quality. In this sense, and due to their low prices, refined olive oil, olive oil or olive pomace oil are sometimes used to adulterate olive oil of better quality, such as virgin and extra virgin olive oil. Similarly, due to lower market prices, other edible vegetable seed oils such as soybean, corn, canola, cotton, sunflower, peanut and almond are likely to be used as illicit adulterants of olive oil. In this scenario, rapid and robust analytical methodology able to detect adulteration is important and very welcome for purposes of quality control and labeling olive oils of high quality. Different methods have been described in the literature dealing with the detection of olive oils adulterated with other edible oils. In the past decade, analytical methods based on liquid and gas chromatography, capillary electrophoresis, and 3

spectroscopic methods such as Fourier transform infrared spectrometry (FTIR) and FTRaman [1-9] were published. However, few papers have been published addressing the problem of adulteration of high quality olive oils with other low quality olive oils, such as ROO, OO or OPO. Spectroscopic parameters, as the specific extinction coefficients at 232 and 270 nm (K232 and K270), the variation between both (ΔK), and the content of trans fatty acids and stigmasta-3,5-diene, both analyzed by gas chromatography, were used to determine EVOO adulteration with 2% of refined olive oil and 0.4% of pomace olive oil [10]. A linear model algorithm, based on chaotic parameters from UV-Vis scans of adulterated EVOO samples, has been proposed to quantify adulterations with low grade olive oils [11]. An artificial neural network model in combination with absorption spectral data has been used to identify the adulteration of EVOO with olive-pomace oils or with olive oils [12]. FTIR data and DA-UPLS were used to discriminate between olive oils obtained from whole olive and stone olive pastes [13]. Other spectroscopic techniques, such as synchronous or conventional fluorescence combined with chemometric methods, have been also proposed [14-16] to detect adulteration of high quality olive oil with olive oils of lower quality. Currently, one of the most promising advances is the profiling approach, which typically is not able to differentiate between analytes and neither quantify them, but it allows the rapid determination of the genuineness of olive oils based on information from multi-target screening methods [17]. In this sense, it is for instance very attractive the full information of the luminescence behavior of olive oils, obtained as excitationemission matrices (EEMs) of autofluorescence, which can be further globally analyzed. It is certain that excitation-emission matrices can only give information related to the whole fingerprinting characteristics, but this technique is easy to use, rapid and ecofriendly. Besides, the possibility of using the front-face mode, allows obtaining the whole information of the matrix avoiding physicochemical changes induced by dilution of the sample, since measurements can be directly performed on intact samples. In this sense, the fluorescence EEM of the intact sample can be considered as a kind of fingerprint, which gives information about the fluorophores naturally occurring on the sample and also about the full environment of those fluorophores in the sample matrix. Autofluorescence of intact food systems is described in the literature as a useful tool for

4

food characterization and analysis [18] and it has been proposed for instance as an effective tool for control of the fraud in the wine industry [19]. There are various chemometric methods susceptible of being combined with EEMs for exploratory analysis discrimination, classification or characterization of olive oils. For instance, discriminant unfolded partial least squares regression (DA-UPLS) was applied for the discrimination between olive oils from two different origin denomination Spanish regions [20]. Unfolded-PCA and PARAFAC were used to detect pomace olive oil in extra virgin olive oil, and PLS regression in combination with EEMs was used to quantify the grade of adulteration [21]. Clustering analysis (CA) was used to discriminate between three types of commercial Spanish olive oils used for human consumption (virgin, pure, and olive pomace oil) using EEMs [22]. The current research article deals with the implementation of different chemometric algorithms on EEMs of autofluorescence, and further comparison of their performance, for the differentiation of extra virgin olive oils from extra virgin olive oils adulterated with olive oils or olive pomace oils. The full fluorescence information contained in the EEMs of intact samples, obtained in front-face mode, has been used as second-order data. EEMs were registered in two spectral regions and further analyzed using PARAFAC (unsupervised and supervised by linear discriminant analysis) and linear discriminant analysis unfolded-partial least-squares (DA-UPLS) algorithms. Finally, UPLS regression has been applied to the quantification of the adulteration levels of extra virgin olive oils with olive oils and olive pomace oils.

2. Experimental 2.1 Samples Samples of extra virgin olive oils from different olive varieties (hojiblanca, manzanilla cacereña, cornicabra and picual), olive oils, and pomace olive oils were obtained from market. The samples were stored in the dark, at room temperature, and the EEM of each sample was registered immediately after opening each bottle. Adulterated samples of extra virgin olive oils with 33%, 16%, 10%, 5% and 3% of olive oil or olive pomace oil were used.

5

2.2 Apparatus Fluorescence measurements were made using a Fluorescence Spectrophotometer Varian Model Cary Eclipse, equipped with two Czerny-Turner monochromators (excitation and emission), a xenon light source and two photomultiplier tubes as detector, and connected to a PC microcomputer via an IEEE 488 (GPIB) serial interface. The Cary Eclipse 1.0 software was used for data acquisition. Measurements were made with a variable-angle front-face accessory, to ensure that reflected light, scattered radiation, and depolarization phenomena were minimized. The angle of incidence, defined as the angle between the excitation beam and the perpendicular to the cell surface, was set at 34º. Fluorescence measurements were recorded in a 10-mm quartz cell at room temperature. The excitation and emission monochromators slits were set at 5 and 5 nm, respectively. The photomultiplier tube sensitivity was set at 550 V, and the monochromators scan rate was set at 300 nm·min-1.

2.3 EEMs recording All excitation-emission matrices were registered in front-face mode and in two spectral regions. Tocopherols and polyphenols region: EEMs were recorded as a set of emission spectra in the range 284-500 nm, each 0.5 nm, and in the excitation range of 270-350 nm, each 5 nm. Chlorophylls and derivatives region: EEMs were recorded as a set of emission spectra in the range 640-700 nm, each 0.5 nm, and in the excitation range of 360-500 nm, each 5 nm. EEMs were registered in ASCII format and transferred to a PC for subsequent treatment and analysis.

2.4 Experimental design The considered experimental design comprises nine sets of samples. Four of them consisted on ten genuine EVOO samples plus other ten samples prepared by 6

spiking the ten genuine EVOO samples with different percentages of olive oils of different commercial brands to simulate adulteration. The other five sets of samples were prepared in the same way, but in this case the spiking to simulate adulteration was performed with different percentages of olive pomace oils of different commercial brands (Table 1). When fluorescence measurements were performed, each sample set yielded 40 EEMs, namely twenty EEMs recorded in the tocopherol and polyphenol spectral region and other 20 EEMs, corresponding to the scans in the pigments spectral region.

2.5 Data modeling All calculations carried out in this work were done in MatLab environment (MATLAB R2008a version 7.6.0.324). Routines for PARAFAC were available in the Internet thanks to Bro [23]. An useful MatLab graphic interface providing a simple means of loading the EEM data into the MatLab working space before running and analyzing them via PARAFAC and DA-UPLS was used [24, 25]. An in house MatLab routine was used for LDA calculations [26].

2.6 Chemometric tools A brief summary of the different chemometric algorithms applied in the present work is given below:

a) PARAFAC. This algorithm permits the decomposition of three-dimensional data arrays, such as EEMs, into two-dimensional spectral profiles for both qualitative and quantitative purposes (23). If the considered data are EEMs and they are arranged in a three-way array X of dimensions I x J x K, where I, J, and K are the number of samples, number of emission wavelengths, and number of excitation wavelengths, respectively, PARAFAC attempts to decompose the whole dataset into three matrices, called A (scores), B, and C (loadings) with elements ain, bjn, ckn, respectively, where n indicates the component number. An element of X is given by

7

∑

(1)

where xijk is the fluorescence intensity for sample i at the emission wavelength j and excitation wavelength k and eijk indicates an element of the array E, which collects the variability not accounted by the model. For a given component n, the elements ain, bjn, and ckn are arranged in the score vector an (whose elements are directly proportional to its concentration in each sample), and the loading vectors bn and cn, which estimate its emission and excitation profiles. The array of EEMs data is fitted to equation 1 by least-squares. The score values represent samples differences and similarities and it is an important aid to look for sample patterns. b) Linear discriminant analysis (LDA) is a well-known chemometric method that calculates a surface separating the established categories [26]. The criterion used to establish the discrimination function is to maximize the ratio of variance between categories to variance within categories [26, 27]. Categories are supposed to follow a multivariate normal distribution and be linearly separated. From a mathematical point of view, if A is the score matrix of PARAFAC, and Y is the I × g dummy matrix of binary digits representing the group assignments (I is the number of samples and g is the number of categories), then the best representation is obtained when the ratio of the between-groups variance Bc matrix and the within-groups variance Wc matrix is maximized. In principle, Y might be of size I x (g-1); however, the mathematical expressions are simplified if a redundant I x g matrix is employed for group assignment [26]. Suitable expressions for the matrices Bc and Wc are [28]: (

)

(

)

(

) (

(2) )

(3)

The canonical variate (CV) scores contain the successively maximized ratios between-groups variance/within-groups variance. They are obtained by PCA of the matrix (Wc- 1Bc) and projection of the data matrix A matrix onto the first loadings (those explaining the higher amount of variance). The samples are then plotted on a two -or three- dimensional space defined by the first CV scores for each sample.

8

c) DA-UPLS. Although the mathematical foundations of U-PLS were originally developed for multivariate calibration purposes [29], its application to the classification of samples has been extensively reported [30-32]. The main difference between U-PLS and discriminant analysis U-PLS (DA-UPLS) consists in the building of the dependent variable y. For model calibration purposes, the variable y contains concentrations values. For discriminant analysis purposes, y contains a coding integer representing the class label of the samples. PLS regression is conducted between the instrumental response in X block (built with the unfolded original second-order matrix data) and the class label in y block using training samples, and the optimal number of latent variables is chosen based on the error rate by cross-validation [33]. The final model for A latent variables is used to predict the class label in the test set according to: v

(4)

where ytest is the label class predicted, ttestT are the scores of test samples obtained by projection of xtest onto the training loadings and v is the vector of the regression coefficients. In the ideal case scenario, the calculated values of ytest - for two classes of samples - are 1 or 2; in practice, ytest values are often close to 1 and 2. Therefore, in order to assign a test sample to a given class, it is necessary to establish thresholds for the ytest predicted values. The threshold is defined as the value that minimizes the number of false positives and false negatives [34].

3. Results and Discussion The focus of our studies was the use of autofluorescence excitation-emission matrices (EEMs), as fluorescence fingerprints of genuine extra virgin olive oils, and of extra virgin olive oils adulterated with olive oils or with pomace olive oils, for the detection and quantification of possible adulterations and frauds. Typically, the detection of extra virgin olive oil adulteration requires complex and time-consuming sample pre-treatment procedures, which are not suitable for on-site analysis. An alternative approach, which is the one covered by this paper, is to use the intrinsic fluorescence (autofluorescence) of the olive oils, without pre-treatment of the samples, for the detection and quantification of the level of extra virgin olive oil 9

adulterations. For that, two dimensional excitation-emission fluorescence spectra were obtained, and multiway data analysis was applied to handle these complex fluorescence signals. With this purpose, PARAFAC, PARAFAC supervised by linear discriminant analysis (LDA) and discriminant unfolded partial least-squares (DA-UPLS) were used.

3.1. General fluorescent behavior of olive oil. Description of the EEMs In the extra virgin olive oil, two different fluorescence emission zones, containing the main fluorescent features of the assayed oils, can be visualized. The first one is located in the excitation range between 270-310 nm and in the emission range between 300-340 nm, and it is attributed to the tocopherol and polyphenol components [22, 35-37]. The other band, with emission at 660-690 nm (excitation between 370-420 nm), is characteristic of the fluorescent pigments of extra virgin olive oil, mainly chlorophylls and pheophytins [38]. For our purpose, and with the object of avoiding the signal corresponding to the Rayleigh scatter, the EEMs of the olive oil samples were registered in front-face mode, independently for each spectral region (Figure 1). It was also explored the possibility of registering the whole matrix at once for each sample, which would imply the further pretreatment of the registered matrices in order to remove dispersion signals. However it was found that Rayleigh signals were not located exactly in the same wavelengths areas in all the matrices and also a big area of the matrices should be removed to be sure that all dispersion signals were avoided, with the subsequent loss of information. This was the main reason for exploring both areas independently. Figure 1 (a, b, c) shows the EEMs of intact extra virgin olive oil (EVOO) (a), olive oil (b) and olive pomace oil (c), at the wavelength interval corresponding to tocopherols and polyphenols emission. The differences in the EEMs between EVOO and olive oil result in a substantial decreases in the band of tocopherols and polyphenols (emission at 330-340 nm and excitation at 270-310 nm, left bottom corner of the landscape), as a consequence of the refined treatment, and in the appearance of an additional fluorescence band, related with the oxidation products, with emission at 370440 nm and excitation at 310-335 nm [39]. Olive pomace oil exhibits even higher fluorescent intensity with the appearance of a broad band at slightly higher excitation

10

wavelengths, namely, the main typical band for olive pomace oil is located at emission 380-480 nm and excitation 320-350 nm [21, 40]. The EEMs obtained in the green pigments region are shown in Figure 1 (d, e, f). In this region, only EVOO presents an intense band with excitation at 360-420 nm and emission at 660-680 nm. In olive oil and olive pomace oil this band has a substantially lower intensity, and is practically not detectable. This fact reveals a decrease in the green pigments concentration as a consequence of the refining process. These pigments are known to be photosensitive and simply the exposure of the oil to the light could explain the decrease in their concentration. To sum up, it can be said that this first approach to the EEMs of different oils reveals important differences in the fluorescent features of samples of different quality. Differences consist in the presence of different bands in the tocopherol and polyphenols spectral region and in substantial changes of intensity in the pigments spectral region, according to the quality of the oil. These differences are the basis for the development of the characterization methodology presented in this article.

3.2. EEMs of adulterated extra virgin olive oil samples Figure 2 shows the EEMs in the spectral zone of tocopherols and polyphenols of an extra virgin olive oil sample adulterated by spiking it with different percentages of olive oil (Figure 2a), or olive pomace oil (Figure 2b). It can be seen that, as the grade of adulteration with olive oil increases (Figure 2a), a narrow fluorescence band, located between 310 and 335 nm in the excitation range, and between 370 and 440 nm in the emission range appears. When adulteration is simulated by spiking extra virgin olive oil with olive pomace oils (Figure 2b), the main fluorescence band is located at the same wavelengths, but now, this band is broader and more intense than in the case of adulteration with olive oil. With respect to the green pigments wavelength region, it is observed that as the grade of adulteration of extra virgin olive oil with olive oil (Figure 2c), or with olive pomace oil (Figure 2d) increases, the fluorescence of this region (emission at 660-680 nm and excitation at 360-420 nm) notably decreases revealing that only EVOO

11

contributes to the fluorescence emission in this spectral region. As stated above, EVOO contains higher concentration of pigments that olive oil or olive pomace oil.

3.3. Classification according to the chemometric methods 3.3.1. PARAFAC: number of components, profiles and score values PARAFAC algorithm was applied in the nine sets of samples, and in the two spectral regions. Results are depicted in Figure 3. In all cases, non-negative constraints for the resolved profiles for all modes were applied. This was done in order to obtain a realistic solution, and in accordance with a priori knowledge (i.e. spectral values are positive). The first models were constructed considering the eighty samples belonging to calibration sets from one to four (Table 1), this is EVOO samples and EVOO samples adulterated with OO. The spectral region corresponding to tocopherols and polyphenols was treated independently from the region corresponding to the pigments. Thus, the fluorescence data were arranged in two independent 3D arrays, the first one of dimensions 80⨯433⨯17 (samples⨯emission⨯excitation) in the tocopherols and polyphenols spectral region and the second one with dimensions 80⨯121⨯29, corresponding to the pigments spectral region. The corresponding PARAFAC models were constructed and the obtained results are depicted in Figure 3. The selection of the optimum number of PARAFAC components (N) is usually performed on the basis of two different criteria: core consistency diagnostics (CORCONDIA) and the analysis of residuals. CORCONDIA is a diagnostic tool that considers the PARAFAC internal parameter known as core consistency. Working under this criterion involves observing the changes in the core consistency parameter as the number of trial fluorescent components is increased and the optimal value of N is selected as the largest tested value for which the core consistency is larger than ≈50 % [41, 42]. It has been noted that the core consistency analysis is a tool based in some data structural assumptions and may fail in certain circumstances [25, 43]. For this reason, we prefer to employ other useful technique based in the consideration of the PARAFAC residual error, i.e., the standard deviation of the elements of the array eijk in eq (1), because it is more intuitively appealing than the rather complex core consistency, and has already shown 12

to be a good indicator of having reached the correct value of N [44]. Thus, a reasonable choice for N is the smallest number of components for which the residual error is not statistically different than the instrumental noise. As it can be observed in Figure 3, in this study, it was found that three factors could be reliably extracted for EVOO samples adulterated with olive oils in both spectral regions. The same procedure was followed with the one hundred samples corresponding to sample sets from five to nine (Table 1). In this case data were arranged in two 3D arrays, the first one of dimensions 100⨯433⨯17 (samples⨯emission⨯excitation) in the tocopherols and polyphenols spectral region, and the second one with dimensions 100⨯121⨯29, corresponding to the pigments spectral region. The number of factors are also three in both spectral regions, as it can be deducted by observation of the variation of the residual error (Figure 3). As stated above, three spectral components are obtained in each spectral region, and when adulteration is performed by spiking with OO or OPO. However, when analyzing the tocopherol and polyphenol region, it is observed that the order in which those components are calculated is different, depending on the type of oil employed to simulate adulteration. The obtained first component presents roughly the same spectral features in the case of adulteration with OO and OPO, with maximum excitation around 320 nm and a broad emission band in the range 370-420 nm. The emission profile of this first component matches with fluorescence emission of oxidation products, as well as its excitation spectra at wavelengths above 310 nm [45], however in its excitation profile, it is possible to find a contribution of tocopherols at wavelengths below 310 nm [22], where a clear shoulder is observed. The observation of the calculated score values for this first component, also support the contribution of tocopherols to the first component, since score values are lower for adulterated samples than for EVOO. Thus, the decrease in those score values might be related to the dilution with OO or OPO of the tocopherols naturally present in EVOO. The shape of the second component obtained for the set of EVOO samples and EVOO samples adulterated with OO is quite similar to the shape of the third one obtained in the case of EVOO samples and EVOO samples adulterated with OPO. However, the spectral features of these components are not very well defined. Their excitation profiles present a continuous increase with some trivial local maxima and their emission profiles present only a drastic decrease between 380 and 410 nm. These components could be related to the presence of degradation 13

products produced as a consequence of the treatment of OO and OPO, since their score values are higher in adulterated oils than in EVOO. Lastly, the excitation and emission profiles for the third component of the model for EVOO samples and EVOO samples adulterated with OO, resemble to the excitation and emission profiles of the second component of the PARAFAC model calculated with EVOO samples and EVOO samples adulterated with OPO. The observation of the score values for these two spectral components reveals that higher values are obtained in the case of samples adulterated with OPO than in the case of samples adulterated with OO. Thus, these components are probably related to oxidation and degradation products commonly present in OPO, as a consequence of the OPO elaboration processes. Those oxidation and degradation products are more important in OPO than in OO, due to the more drastic elaboration process, and because of that this component is the second one in the case of OPO adulteration and the third one in the case of OO adulteration. On the other hand, in the pigments region, the excitation and emission loadings are very similar, regardless of the oil used in the spiking to simulate adulteration. Also, the obtained three spectral components in each case, are calculated in the same order by the PARAFAC algorithm. This might be related with the weak presence of green pigments in olive oils and in olive pomace oils as a consequence of its elaboration processes. Therefore, the pigments in the adulterated EVOO are provided by the EVOO presents in the mixtures. In both adulterations, the first component shows excitation and emission maxima at 418 and 675 nm, respectively, and these wavelengths and the profile of the excitation spectra may be attributed to the contribution of chlorophyll a. The second component, with excitation maximum located at a shorter wavelength, 390 nm, and the emission maxima at slightly higher wavelength, 678 nm, might be attributed to the presence of pheophytin a. The third component presents an emission maximum at 655 nm, and it can be associated with the chlorophyll b. The excitation and emission wavelengths obtained by PARAFAC for this pigments are slightly different that those reported in the bibliography, where the fluorescence information has been obtained in pure solvents. The spectra obtained are those due to the green pigments present in an intact olive oil matrix, a non-polar and high viscosity medium.

3.3.2. LDA-PARAFAC 14

An approach that often improves the screening capability of PARAFAC consists on submitting the scores calculated by PARAFAC to supervised LDA. In our case, and with the object of improving the limits of differentiation between adulterated olive oils, LDA was applied on the matrix of scores of PARAFAC, to obtain the canonical vectors. A tridimensional plot allows to represent the data of canonical vectors in clusters groups based on the coded values assigned to each sample according to its category. Then, the classification is performed by location of the training samples as points whose coordinates are the canonical vector scores. Each set of samples, containing both adulterated and genuine EVOO samples, has been independently treated in order to assess the discrimination ability of the LDAPARAFAC. Figure 4 shows the tridimensional plots of the canonical vector score values, obtained in the tocopherols and polyphenols spectral region, for the calibration sets of EVOO adulterated with different percentages of olive oils. Besides, and with the object of facilitating the visualization of pairwise comparison, each plot includes the projections of the 95% confidence level ellipses over the three planes defined by the corresponding axes [46]. If the confidence ellipses do not overlap, it can be concluded that the assigned categories are different. According to the results, as the percentage of the olive oil used in the adulteration increases, the ellipses separation increased, allowing the discrimination of adulterations higher than 16%. With respect to the adulteration with olive pomace oils, Figure 5 shows the tridimensional plots of the three score values obtained for the three components and for three adulteration levels, 33%, 10% and 3%. In this case, the model was able to differentiate between pure EVOO and EVOO adulterated with 3% of olive pomace oil or higher. With respect to the pigments spectral region, and in accordance with what has been previously indicated, the fluorescence in this zone is due almost entirely to the presence of EVOO in the mixtures, and the adulterants do not add any differential information. Only intensity differences are observed in this area as a consequence of adulteration, but not changes in the shape of the EEMs, thus the discriminant power of this spectral region is supposed to be poorer. In the adulteration with olive oils, an adequate classification was obtained when the percent of this adulterant is of 16% or higher, similarly to the results found in the tocopherols and polyphenols region. The 15

results achieved in the adulteration with olive pomace oils show an adequate differentiation when the presence of olive pomace oil is of 33% or higher.

3.3.3. DA-UPLS With this algorithm we proceed in the same way, with two independent spectral zones. The optimization of wavelength ranges in the tocopherols and polyphenols region, provided the following excitation and emission intervals: 270-300 nm and 320400 nm, respectively. The estimation of the number of optimum latent variables was carried out by the leave-one-sample out-cross validation approach [47] using fourteen samples as training set for each adulteration type. Under these conditions, two factors were found for EVOO adulterated with olive oils, and five factors for EVOO adulterated with olive pomace oils. To test the discriminant ability of DA-UPLS, each set of Table 1was divided in a training subset containing twelve samples and in a test subset containing the remaining eight samples. In both cases samples were coded according to code values for each category (0 and 1 for EVOO and for EVOO adulterated with different percentages of olive oils or olive pomace oils, respectively). In Figure 6, the predicted versus nominal code values for different levels of adulteration with olive oils and with olive pomace oils are shown. The confidence interval of both categories, marked as horizontal bands, was estimated as the product of the calculated standard deviations of the results of each one of the training samples and the Student tvalue (95% confidence interval) with n-1 degrees of freedom for each category. The prediction ability of this algorithm is considered deficient when a significant overlapping between the populations takes place. As it can be seen in Figure 6a, for EVOO samples adulterated with different levels of olive oils, and in Figure 6b, for EVOO adulterated with different percentages of olive pomace oils, all samples are clearly predicted and classified. The limits for an acceptable discrimination were adulteration levels higher than 16% for olive oils and 5% for olive pomace oils. As it can be seen in the DA-UPLS plots, the results are similar to the obtained with LDA-PARAFAC except in the differentiation of

16

adulteration of pomace olive oils, as in this case the algorithm is not able to distinguish adulterations lower than 5%.

3.4. Validation and quantification of olive oil or olive pomace oil adulteration levels Using the previously optimized models by means of DA-UPLS, UPLS was applied in an attempt of quantifying the adulteration level. The models were validated by predicting two new set of samples prepared as according with Table 2. The U-PLS predictions are very good in all the validation samples, and the statistical parameters were evaluated through the relative error of prediction (REP%) and the root mean square error of prediction (RMSEP). These values were similar for both adulterations with olive oils and with olive pomace oils, 1.8 and 15.4%, and 1.3 and 15.5%, respectively. On the other hand, a set of 10 blind samples were processed by the optimized DA-UPLS and UPLS models. A simply visualization of the contour plots of the EEMs obtained in front-face mode, Figure 2a and 2b, allow us to detect the type of adulteration, with olive oils or with olive pomace oils. This first visual discrimination was corroborated when DA-UPLS was applied, and all blind samples were correctly classified according with the oil used for their adulteration. After this, the level of adulteration was predicted by UPLS regression and the results were very satisfactory (Figure 7).

4. Conclusions The possibility of obtaining analytical signals in a nondestructive fashion, not altering the composition of the matrix, and measuring fluorophores in their native environment, is highly valuable for the quality control of olive oil samples. The results of this work provide full fluorescence data which can be directly used as a visualization tool, and once combined with various multivariate analysis algorithms, can be correlated with adulteration of extra virgin olive oils. On the basis of the obtained results, it can be concluded that LDA-supervised PARAFAC shows a high discrimination power between non-adulterated and adulterated extra virgin olive oils 17

with olive oils or olive pomace oils. A satisfactory discrimination is obtained for adulterations higher than 16% with olive oils and higher than 3% with olive pomace oils. DA-UPLS allows the detection of adulteration with 16% of olive oils, and 5% of olive pomace oils. Besides, UPLS was used to quantify the grade of adulteration, concluding that this algorithm has a high potential to be applied in the study of adulterations. These studies are a promising demonstration of the potential of autofluorescence of intact extra virgin olive oil samples, in combination with second order chemometric analysis, for detecting possible adulterations with olive oils of lower quality. Rapidness is among the main advantages of the developed methodology, allowing a fast overall view of the possible adulteration of extra virgin olive oils.

Acknowledgments The authors are grateful to Ministerio de Economía, Industría y Competitividad of Spain (Project CTQ2017-82496-P) and Gobierno de Extremadura (GR15090-Research Group FQM003, and IB16058 and IB16022), all co-financed by European FEDER funds, for financially supporting this work.

18

References 1.

H. Jabeur, A. Zribi, J. Makni, A. Rebai, R. Abdelhedi, M. Bouaziz, Detection of Chemlali Extra-Virgin Olive Oil Adulteration Mixed with Soybean Oil, Corn Oil, and Sunflower Oil by Using GC and HPLC, J. Agric. Food Chem. 62 (2014) 48934904.

2.

M.G. Bagur-González, E. Pérez-Castaño, M. Sánchez-Viñas, D. GázquezEvangelista, Using the liquid-chromatographic-fingerprint of sterols fraction to discriminate virgin olive from other edible oils, J. Chromatogr. A 1380 (2015) 6470.

3.

Y. Yang, M.D. Ferro, I. Cavaco, Y. Liang, Detection and Identification of Extra Virgin Olive Oil Adulteration by GC-MS Combined with Chemometrics, J. Agric. Food Chem. 61 (2013) 3693-3702.

4.

A. Rohman, Y.B.C. Man, Fourier transform infrared (FTIR) spectroscopy for analysis of extra virgin olive oil adulterated with palm oil, Food Res. Int. 43 (2010) 886-892.

5.

M. Vasconcelos, L. Coelho, A Barros, J.M.M.M. de.Almeida, Study of adulteration of extra virgin olive oil with peanut oil using FTIR spectroscopy and chemometrics, Cogent Food Agric. 1 (2015) 1018695.

6.

E.C. López-Díez, G. Bianchi, R. Goodacre, Rapid Quantitative Assessment of the Adulteration of Virgin Olive Oils with Hazelnut Oils Using Raman Spectroscopy and Chemometrics, J. Agric. Food Chem. 51 (2003) 6145-6150.

7.

V. Baeten, J.A. Fernández Pierna, P. Dardenne, M. Meurens, D.L. GarcíaGonzález, R. Aparicio-Ruiz, Detection of the Presence of Hazelnut Oil in Olive Oil by FT-Raman and FT-MIR Spectroscopy, J. Agric. Food Chem. 53 (2005) 6201-6206.

8.

L. Sánchez-Hernández, M.L. Marina, A.L. Crego, A capillary electrophoresis– tandem mass spectrometry methodology for the determination of non-protein amino acids in vegetable oils as novel markers for the detection of adulterations in olive oils, J. Chromatogr. A 1218 (2011) 4944-4951.

9.

L. Sánchez-Hernández, M. Castro-Puyana, M L. Marina, A.L. Crego, Determination of betaines in vegetable oils by capillary electrophoresis tandem mass spectrometry –application to the detection of olive oil adulteration with seed oils, Electrophoresis 32 (2011) 1394-1401.

10. H. Jabeur, A. Zribi, M. Bouaziz, Extra-Virgin Olive Oil and Cheap Vegetable Oils: Distinction and Detection of Adulteration as Determined by GC and Chemometrics, Food Anal. Methods 9 (2016) 712-723. 11. J.S. Torrecilla, E. Rojo, J.C. Dominguez, F. Rodríguez, A Novel Method To Quantify the Adulteration of Extra Virgin Olive Oil with Low-Grade Olive Oils by UV−Vis, J. Agric. Food Chem. 58 (2010) 1679-1684. 12. R. Aroca-Santos, J.C. Cancilla, G. Matute, J.S. Torrecilla, Identifying and Quantifying Adulterants in Extra Virgin Olive Oil of the Picual Varietal by 19

Absorption Spectroscopy and Nonlinear Modeling, J. Agric. Food Chem. 63 (2015) 5646-5652. 13. M. De Luca, D. Restuccia, M.L. Clodoveo, F. Puoci, G. Ragno, Chemometric analysis for discrimination of extra virgin olive oils from whole and stoned olive pastes, Food Chem. 202 (2016) 432-437. 14. K.I. Poulli, G.A. Mousdis, C.A. Georgiou, Synchronous fluorescence spectroscopy for quantitative determination of virgin olive oil adulteration with sunflower oil, Anal. Bioanal. Chem. 386 (2006) 1571-1575. 15. A. Dankowska, M. Małecka, W. Kowalewski, Discrimination of edible olive oils by means of synchronous fluorescence spectroscopy with multivariate data analysis, Grasas y Aceites 64 (2013) 425-431. 16. T. Mu, S. Chen, Y. Zhang, H. Chen, P. Guo, F. Meng, Portable Detection and Quantification of Olive Oil Adulteration by 473-nm Laser-Induced Fluorescence, Food Anal. Methods 9 (2016) 275-279. 17. R. Aparicio, M.T. Morales, R. Aparicio-Ruiz, N. Tena, D.L. García-González, Authenticity of olive oil: Mapping and comparing official methods and promising alternatives, Food Res. Int. 54 (2013) 2025-2038. 18. J. Christensen, L. Norgaard, R. Bro, S.B. Engelsen, Multivariate autofluorescence of intact food systems, Chem. Rev. 106 (2006) 1979–1994. 19. D. Airado-Rodríguez, I. Durán-Merás, T. Galeano-Díaz, J.P. Wold, Front-face fluorescence spectroscopy: A new tool for control in the wine industry, Journal of Food Composition and Analysis 24 (2011) 257-264. 20. F. Guimet, J. Ferré, R. Boqué, Study of oils from the protected denomination of origin «Siurana» using excitation-emission fluorescence spectroscopy and threeway methods of analysis, Grasas y Aceites 56 (2005) 292-297. 21. F. Guimet, J. Ferré, R. Boqué, Rapid detection of olive–pomace oil adulteration in extra virgin olive oils from the protected denomination of origin «Siurana» using excitation–emission fluorescence spectroscopy and three-way methods of analysis, Anal. Chim. Acta 544 (2005) 143-152. 22. F. Guimet, R. Boqué, J. Ferré, Cluster analysis applied to the exploratory analysis of commercial spanish olive oils by means of excitation-emission fluorescence spectroscopy, J. Agric. Food Chem. 52 (2004) 6673-6679. 23. R. Bro, Multiway analysis in the food industry. Models, algorithms and applications. https://www.researchgate.net/publication/2407346_Multiway_analysis_in_the_foo d_industry_Models_algorithms_and_applications 24. A.C. Olivieri, H.L. Wu, R.Q. Yu, MVC2: A MATLAB graphical interface toolbox for second-order multivariate calibration, Chemom. Intell. Lab. Syst. 96 (2009) 246-251. 25. A. Olivieri, G.M. Escandar, Practical Three-Way Calibration. Elsevier; 2014. 20

26. E.K. Kemsley, A genetic algorithm (GA) approach to the calculation of canonical variates (CVs), TrAC Trends Anal. Chem. 17 (1998) 24-34. 27. L.A. Berrueta, R.M. Alonso-Salces, K. Héberger, Supervised pattern recognition in food analysis, J. Chromatogr. A 1158 (2007) 196-214. 28. A.F. Arruda, H.C. Goicoechea, M. Santos, A.D. Campiglia, A.C. Olivieri, Solid−Liquid Extraction Room Temperature Phosphorimetry and Pattern Recognition for Screening Polycyclic Aromatic Hydrocarbons and Polychlorinated Biphenyls in Water Samples, Environ. Sci. Technol. 37 (2003) 1385-1391. 29. U.G. Indahl, The geometry of PLS1 explained properly: 10 key notes on mathematical properties of and some alternative algorithmic approaches to PLS1 modelling, J. Chemom. 28 (2014) 168-180. 30. M. Barker, W. Rayens, Partial least squares for discrimination, J. Chemom. 17 (2003) 166-173. 31. S.S. Ouertani, G. Mazerolles, J. Boccard, S. Rudaz, M. Hanafi, Multi-way PLS for discrimination: Compact form equivalent to the tri-linear PLS2 procedure and its monotony convergence, Chemom. Intell. Lab. Syst. 133 (2014) 25-32. 32. R.G. Brereton, G.R. Lloyd, Partial least squares discriminant analysis: taking the magic away, J. Chemom. 28 (2014) 213-225. 33. M.R. de Almeida, D.N. Correa, W.F.C. Rocha, F.J.O. Scafi, R.J. Poppi, Discrimination between authentic and counterfeit banknotes using Raman spectroscopy and PLS-DA with uncertainty estimation, Microchem. J. 109 (2013) 170-177. 34. D. Ballabio, V. Consonni, Classification tools in chemistry. Part 1: linear models. PLS-DA, Anal. Methods 5 (2013) 3790-3798. 35. K. Papadopoulos, T. Triantis, C.H. Tzikis, A. Nikokavoura, D. Dimotikali, Investigations of the adulteration of extra virgin olive oils with seed oils using their weak chemiluminescence, Anal. Chim. Acta 464 (2002) 135-140. 36. M. Zandomeneghi, L. Carbonaro, C. Caffarata, Fluorescence of Vegetable Oils: Olive Oils, J. Agric. Food Chem. 53 (2005) 759-766. 37. R. Cheikhousman, M. Zude, D.J.R. Bouveresse, C.L. Léger, D.N. Rutledge, I. Birlouez-Aragon, Fluorescence spectroscopy for monitoring deterioration of extra virgin olive oil during heating, Anal. Bioanal. Chem. 382 (2005) 1438-1443. 38. T. Galeano Díaz, I. Durán Merás, C.A. Correa, B. Roldán, M.I. Rodríguez Cáceres, Simultaneous Fluorometric Determination of Chlorophylls a and b and Pheophytins a and b in Olive Oil by Partial Least-Squares Calibration, J. Agric. Food Chem. 51 (2003) 6934-6940. 39. N.B. Kyriakidis, P. Skarkalis, Fluorescence spectra measurement in olive oil and other vegetable oils, J. AOAC Int. 83 (2000) 1435-1439. 40. E. Sikorska, I. Khmelinskii, M. Sikorski, Analysis of Olive Oils by Fluorescence Spectroscopy: Methods and Applications, in B. Dimitrios (Eds), Olive Oil21

Constituents, Quality, Health Properties and Bioconversion, In TechOpen, Chapter 4, 2012. 41. R. Bro, PARAFAC tutorial and applications, Chemom. Intell. Lab. Syst. 38 (1997) 149-171. 42. R. Bro, H.A.L. Kiers, A new efficient method for determining the number of components in PARAFAC models, J. Chemom. 17 (2003) 274-286. 43. G.M. Escandar, H.C. Goicoechea, A. Muñoz de la Peña, A.C. Olivieri, Secondand higher-order data generation and calibration: A tutorial, Anal. Chim. Acta 806 (2014) 8-26. 44. A. Muñoz de la Peña, A. Espinosa Mansilla, D. González Gómez, A.C. Olivieri, H.C. Goicoechea, Interference-Free Analysis Using Three-Way Fluorescence Data and the Parallel Factor Model. Determination of Fluoroquinolone Antibiotics in Human Serum, Anal. Chem. 75 (2003) 2640-2646. 45. F. Guimet, J. Ferre, R. Boque, F.X. Rius, Application of Unfold Principal Component Analysis and Parallel Factor Analysis to the Exploratory Analysis of Olive Oils by Means of Excitation-Emission Matrix Fluorescence Spectroscopy. Anal. Chim. Acta 515 (2004) 75-85. 46. A. Muñoz de la Peña, N. Mujumdar, E.C. Heider, H.C. Goicoechea, D. Muñoz de la Peña, A.D. Campiglia, Nondestructive Total Excitation–Emission Fluorescence Microscopy Combined with Multi-Way Chemometric Analysis for Visually Indistinguishable Single Fiber Discrimination, Anal. Chem. 88 (2016) 2967-2975. 47. D.M. Haaland, E.V. Thomas, Partial least-squares methods for spectral analyses. 1. Relation to other quantitative calibration methods and the extraction of qualitative information, Anal. Chem. 60 (1988) 1193-1202.

22

Captions Figure 1. Contour plots of EEMs recorded from extra virgin olive oil, olive oil and olive pomace oil, in the tocopherols and polyphenols spectroscopic region (a, b, c): excitation range (mode C) = 270-350 nm, emission range (mode B) = 280-500 nm, and in the pigments spectroscopic region (d, e, f): excitation range (mode C) = 360-500 nm, emission range (mode B) = 640-700 nm. Figure 2. Contour plots of EEMs of fluorescence matrices of an extra virgin olive oil adulterated with different levels (5%, 16% and 33%) of olive oil (a), and olive pomace oil (b) in tocopherols and polyphenols wavelength region, excitation range (mode C) = 270-350 nm and emission range (mode B) = 280-500 nm; (c) and (d) in pigments wavelength region, excitation range (mode C) = 360-500 nm, emission range (mode B) = 640-700 nm. Figure 3. Standard deviation of residual as a function of the number of components and PARAFAC excitation and emission profiles for adulterations with olive oil and with olive pomace oil, in both spectral regions. Figure 4. Clusters of LDA-PARAFAC, on the tocopherols and polyphenols spectral region, for the discrimination between different EVOO and EVOO adulterated with different amounts of olive oils. □ EVOO; ○ EVOO adulterated with OO in different percentages. The three dimensional projection of the 95% confidence ellipse of the data is included. Figure 5. Clusters of LDA-PARAFAC, on the tocopherols and polyphenols spectral region, for the discrimination between different EVOO and EVOO adulterated with different amounts of olive pomace oils. □ EVOO; ○ EVOO adulterated with OPO in different percentages. The three dimensional projection of the 95% confidence ellipse of the data is included. Figure 6. Plot of DA-UPLS predicted versus nominal code values for 20 samples ( 12 calibration samples, and 8 validation samples) for the discrimination between EVOO (blue symbols) and adulterated EVOO with different percentages of OO (a) and OPO (b) (red symbols), on the tocopherols and polyphenols spectral region. The geometrical symbols correspond to training samples and the cross symbols to validation samples. Figure 7. Plot for predicted adulteration percentages in blind samples as a function of the nominal values (the solid line is the perfect fit) by UPLS. Adulterant: (■) olive oils (OO); (●) olive pomace oils (OPO).

23

24

25

26

27

28

29

Graphical Abstract

Table 1.- Experimental design Calibration sets of EVOO adulterated with olive oils (OO) Calibration set

EVOO

OO

1

10 samples

10 samples of EVOO adulterated with 33% of different olive oils

2

10 samples

10 samples of EVOO adulterated with 16% of different olive oils

3

10 samples

10 samples of EVOO adulterated with 10% of different olive oils

4

10 samples

10 samples of EVOO adulterated with 5% of different olive oils

Calibration sets of EVOO adulterated with olive pomace oils (OPO) Calibration set

EVOO

OPO

30

5

10 samples

10 samples of EVOO adulterated with 33% of different olive pomace oils

6

10 samples

10 samples of EVOO adulterated with 16% of different olive pomace oils

7

10 samples

10 samples of EVOO adulterated with 10% of different olive pomace oils

8

10 samples

10 samples of EVOO adulterated with 5% of different olive pomace oils

9

10 samples

10 samples of EVOO adulterated with 3% of different olive pomace oils

Table 2.- Validation statistical results for the quantitative analysis of the percentage of adulteration level in two EVOO samples adulterated with olive virgin and olive pomace oils, by UPLS.

31

Adulterant oil EVOO 1

Olive virgin oil

Actual (% v:v) 33.3 16.7

Predicted (% v:v) 31.0 16.1

Recovery%

33.3 16.7

32.7 18.0

98.4 107.7

EVOO 2 Reca ± SD RMSEPb REPc (%)

EVOO 1

Olive pomace oil

EVOO 2

99 ± 6 1.3 15.5 33.3 16.7 10.0 5.0 7.0

32.4 20.0 11.6 4.6 6.4

97.2 119.9 116.1 91.3 91.4

33.3 16.7 10.0 5.0 7.0

29.9 18.9 8.3 3.5 6.5

89.9 113.2 83.5 70.6 92.8

Reca ± SD RMSEPb REPc (%) a Rec: average recovery b

93.1 96.3

96 ± 17 1.8 15.4

∑ RMSEP: root mean square prediction = √

(

̂)

, yi = actual concentration, ̂ i =

predicted concentration, N = number of samples in the test set. c

REP: relative error of prediction =

√∑

concentration in the prediction set.

32

(

̂ )

,

= is the mean of the added

Highlights 

Front-face autofluorescence of olive oils combined with LDA-PARAFAC and DA-UPLS



LDA-PARAFAC discriminates between non and adulterated extra virgin olive oils



Discrimination is satisfactory for adulterations higher than 16% with olive oils



With olive pomace oils, discrimination is possible for levels higher than 3%



UPLS allows to quantify the grade of adulteration

33