On the Traceability of the Hazelnut Production Chain by Means of Trace Elements

The production chain of hazelnuts has been studied by analyzing three sets of samples produced in purity from three different pools of hazelnuts of cultivar “Tonda Gentile Trilobata”, “Tonda Gentile Romana” and “Mortarella”, all cultivated in Italy. From each pool, five processed products were obtained: roasted hazelnuts, hazelnut paste, hazelnut cream, Gianduja paste and Gianduiotto paste. After pre-treatment by means of dry ashing, all samples from each cultivar, including raw hazelnuts, were then analyzed by means of Inductively Coupled Plasma–Optical Emission Spectroscopy (ICP-OES) and Inductively Coupled Plasma–Mass Spectrometry (ICP-MS). A good discrimination was obtained among the different chain stages according to the distribution of the trace elements, as expected. More interesting was the discrimination among the different cultivars: it was possible to distinguish the samples produced from the respective cultivar by means of specific chemical markers, particularly Mo and Ni.


Introduction
The concept of traceability is currently very important in the food market, and the contribution of analytical chemistry has become relevant thanks to the development of methods that can support the actual traceability systems based as the monitoring of material flows. The possibility of tracing the production chain of foodstuffs on a chemical basis is interesting both for consumers, who can be aware of the provenance of the products they purchase, and for producers, who can guarantee protection from fraud. Traceability can be obtained by analytical methods provided that reliable chemical descriptors are identified, providing an ideal fingerprint that is maintained through all the steps of the entire production chain.
This possibility is, however, limited by the availability of reliable markers that could be used to follow the chain from the initial stages to the final product: this can be easier in cases where raw materials are little or not at all modified, such as for most fruit and vegetable products, as it is shown in a recent study on extra virgin olive oil [1]. It can be indeed more difficult in cases where the production chain involves many steps that deeply modify raw materials with additions of other materials and/or technological modifications: examples on wine [2,3] and milk [4] can illustrate this point. On the biological side, DNA-based tools, recently reviewed by Böhme et al. [5], such as DNA barcoding, can constitute a powerful resource in this field. However, very few studies exist in the field of analytical chemistry dealing with this topic.
With particular regards to the confectionery industry, one of the most relevant foodstuffs is provided by hazelnuts (Corylus avellana L.). It is well known that the "Tonda Gentile Trilobata" (TGT) variety cultivated in Piemonte (Italy) is considered perhaps the best in the world as far as sensorial characteristics and technological properties are concerned. Moreover, it is characterized by a higher price if compared to other cultivars and its use is, therefore, prone to commercial fraud. There is, therefore, an interest in verifying its real use in processed products that are mostly products mixed with cocoa (pralines, pastas, bars, etc.) and other confectionery ingredients. The precise identification of the cultivar in raw hazelnuts is consolidated, as it can be seen in recent works [6,7]. The link between soil and fruit, based on the distribution of lanthanides, has been established in previous traceability studies [8]. As far as processed products are involved, however, there are no complete studies based on chemical descriptors. Locatelli et al. [9] studied the effect of roasting on the possibility of recognizing different cultivars, particularly TGT, in hazelnuts isolated from commercial products. Torello Marinoni et al. [10] were able to follow the production chain of hazelnut from the fruit up to processed products, exploiting specific types of biological descriptors such as nuclear SSR (simple sequence repeats), SNP (single nucleotide polymorphism), InDels (insertions and deletions) and chloroplast markers. The request of recognizing the use of TGT in ready-to-eat confectionery products using methods based on analytical chemistry is at present unaddressed.
In this work, the production chain of hazelnuts has been studied by analyzing three sets of samples produced in purity from three different pools of hazelnuts. The hazelnuts pools were of cultivar "Tonda Gentile Trilobata" (TGT), "Tonda Gentile Romana" (TGR) and "Mortarella" (MOR). From each pool, five processed products were obtained: roasted hazelnuts, hazelnut paste, hazelnut cream (a spreadable cream made from a mixture of hazelnut paste and cocoa), Gianduja paste (the same as hazelnut cream but richer in cocoa) and Gianduiotto paste (similarly to Gianduja paste but used to prepare a typical Piemontese chocolate candy called Gianduiotto). All five products from each cultivar and raw hazelnuts, as well a total of 18 samples, were then analyzed by means of Inductively Coupled Plasma-Optical Emission Spectroscopy (ICP-OES) and Inductively Coupled Plasma-Mass Spectrometry (ICP-MS), attempting to verify whether the determined distribution of the trace elements was able to recognize the original cultivars used in the processed products.

Results and Discussion
The 54 samples (3 cultivar × 6 products × 3 independent replicates) were analyzed by means of ICP-OES and ICP-MS analysis and yielded the data reported in Table S1. All 46 elements determined were above the LOQ (Limit of Quantification) values for the instruments used in this work. Elemental data were subjected to multivariate analysis in order to evaluate potential groupings; the dataset comrpised 54 samples × 46 variables, which were transformed into z-scores before analysis. A reduction in dimensionality by means of PCA yielded the results shown in Figure 1. The first two PCs accounted for the 63.83% of the total explained variance and showed no clusterization of the samples according to cultivar but rather according to the product type ( Figure 1a): the samples can be divided into three main groups:   The observed clustering can be attributed to the increasing contribution of cocoa, absent in hazelnuts-only products, but added to hazelnut cream, Gianduja and Gianduiotto paste (the last two in larger percentages), as it emerges from the content of the two main ingredients (hazelnuts and cocoa) of the different products (Table 1). In fact, as it emerges from the loadings (Figure 1b), the elements positively contributing to this difference are Al, Co, Li, Na and Si, which, according to the literature data (Table 2), are higher in cocoa than in hazelnuts. In particular, silicon in cocoa beans is estimated to be up to 300 times higher than in hazelnuts [11]; as for aluminum, in the study by Stahl et al. [12], the authors found that cocoa powder was among the foodstuffs with the higher Al content, which is possibly due to the contribution of growing soil. The observed clustering can be attributed to the increasing contribution of cocoa, absent in hazelnuts-only products, but added to hazelnut cream, Gianduja and Gianduiotto paste (the last two in larger percentages), as it emerges from the content of the two main ingredients (hazelnuts and cocoa) of the different products (Table 1). In fact, as it emerges from the loadings (Figure 1b), the elements positively contributing to this difference are Al, Co, Li, Na and Si, which, according to the literature data (Table 2), are higher in cocoa than in hazelnuts. In particular, silicon in cocoa beans is estimated to be up to 300 times higher than in hazelnuts [11]; as for aluminum, in the study by Stahl et al. [12], the authors found that cocoa powder was among the foodstuffs with the higher Al content, which is possibly due to the contribution of growing soil. To resume, it seems that a good correlation exists between the mineral content of our samples and their cocoa powder %.
Regarding to the information about the separation of the three cultivars, it seems mostly related to PC3 and PC4 (Figure 1c), accounting for about 20% of the variance: The samples appear, in fact, quite clustered according to the cultivar with TGT samples at negative scores on PC3, TGR samples at positive scores on both PC3 and PC4 and MOR samples at intermediate cores on PC3. Looking at the corresponding loading plot (Figure 1d), Ni, Mo, Cs and Rb appeared as the variables mostly related to this separation (Ni and Mo more present in TGT samples, while Cs and Rb more present in TGR samples).

Grouping by Process Stage
The effect of processing was then investigated by looking for chemical markers able to discriminate among the different processed products. Applying LDA to the elemental data with variable selection in forward search (T to-enter = 3, F to-remove = 2), seven variables were selected as significant (α < 0.05). A good classification was obtained, as expected, and all samples were correctly classified both in fitting and in cross-validation (Monte Carlo with elimination of the 20% of the objects of each class at each iteration; 100 iterations); the elements selected by the forward search procedure were Y, Zn, Cu, Si, Li, S and Co. The calculated canonical functions allowed the obtainment of the F1 vs. F2 diagrams reported in Figure 2a, where hazelnuts-only products (Cluster 1) obtained positive scores on both F1 and F2, Hazelnut cream (Cluster 2) obtained positive scores on F1 and negative ones on F2 and Gianduja and Gianduiotto pastes (Cluster 3) obtained negative scores on F1. Looking at the corresponding loadings (Figure 2b), Group 1 and 2 are characterized by high values of Cu, S Zn and Y and low levels of Li, Si and Co if compared to Cluster 3, but Cluster 2 shows a larger concentration of Y and lower concentration of Cu, S and Zn, while Cluster 1 has an opposite behavior. Cluster 3 shows instead a high content of Li, Si and Co and low content of the other elements (Cu, S, Zn and Y). As explained before, the presence of some of these descriptors in the discrimination model is due to their different contents in hazelnut or cocoa powder.

Grouping by Cultivar
After discriminating according to the type of product/production chain stage, the possibility of discriminating products was verified according to the original hazelnut cultivars, i.e., "Tonda Gentile Trilobata" (TGT), "Tonda Gentile Romana" (TGR) and "Mortarella" (MOR). This is the most important and useful discrimination, as it would allow the verification of the origin of the raw materials even at the end of the production chain. In particular, verifying the possibility of discriminating between products obtained from TGT hazelnuts and products obtained from other hazelnuts was important.
The preliminary results of the reduction in dimensionality by means of PCA, shown in Figure 1, were unsatisfactory from the point of view of the discrimination conducted among cultivars, because, as explained before, the samples were clustered, independently from the cultivar, into three groups: (1) hazelnut-only products; (2) hazelnut cream samples; (3) Gianduja pastes/Gianduiotto pastes. Therefore, LDA was again used in order to achieve a better classification, with variable selection in forward search (T to-enter = 3, F to-remove = 2). Six variables were selected as significant (α < 0.05). The calculated canonical functions allowed obtaining F1 vs. F2 diagrams reported in Figure 3a, which show very good clustering in which all samples were grouped according to the original cultivar.

Grouping by Cultivar
After discriminating according to the type of product/production chain stage, the possibility of discriminating products was verified according to the original hazelnut cultivars, i.e., "Tonda Gentile Trilobata" (TGT), "Tonda Gentile Romana" (TGR) and "Mortarella" (MOR). This is the most important and useful discrimination, as it would allow the verification of the origin of the raw materials even at the end of the production chain. In particular, verifying the possibility of discriminating between products obtained from TGT hazelnuts and products obtained from other hazelnuts was important.
The preliminary results of the reduction in dimensionality by means of PCA, shown in Figure 1, were unsatisfactory from the point of view of the discrimination conducted among cultivars, because, as explained before, the samples were clustered, independently from the cultivar, into three groups: (1) hazelnut-only products; (2) hazelnut cream samples; (3) Gianduja pastes/Gianduiotto pastes. Therefore, LDA was again used in order to achieve a better classification, with variable selection in forward search (Tto-enter = 3, Ftoremove = 2). Six variables were selected as significant (α < 0.05). The calculated canonical functions allowed obtaining F1 vs. F2 diagrams reported in Figure 3a, which show very good clustering in which all samples were grouped according to the original cultivar. All 54 samples were correctly classified in fitting and very good performances were reached in cross-validation (Monte Carlo with elimination of the 20% of the objects of each class at each iteration; 100 iterations) with NER% = 94.82%. The performance indexes are reported in Table 3.

Grouping by Cultivar
After discriminating according to the type of product/production chain stage, the possibility of discriminating products was verified according to the original hazelnut cultivars, i.e., "Tonda Gentile Trilobata" (TGT), "Tonda Gentile Romana" (TGR) and "Mortarella" (MOR). This is the most important and useful discrimination, as it would allow the verification of the origin of the raw materials even at the end of the production chain. In particular, verifying the possibility of discriminating between products obtained from TGT hazelnuts and products obtained from other hazelnuts was important.
The preliminary results of the reduction in dimensionality by means of PCA, shown in Figure 1, were unsatisfactory from the point of view of the discrimination conducted among cultivars, because, as explained before, the samples were clustered, independently from the cultivar, into three groups: (1) hazelnut-only products; (2) hazelnut cream samples; (3) Gianduja pastes/Gianduiotto pastes. Therefore, LDA was again used in order to achieve a better classification, with variable selection in forward search (Tto-enter = 3, Ftoremove = 2). Six variables were selected as significant (α < 0.05). The calculated canonical functions allowed obtaining F1 vs. F2 diagrams reported in Figure 3a, which show very good clustering in which all samples were grouped according to the original cultivar. All 54 samples were correctly classified in fitting and very good performances were reached in cross-validation (Monte Carlo with elimination of the 20% of the objects of each class at each iteration; 100 iterations) with NER% = 94.82%. The performance indexes are reported in Table 3. All 54 samples were correctly classified in fitting and very good performances were reached in cross-validation (Monte Carlo with elimination of the 20% of the objects of each class at each iteration; 100 iterations) with NER% = 94.82%. The performance indexes are reported in Table 3. The selected variables were Mn, Ni, Zr, Mo, Ba and W. This behavior seems to partly reflect the elemental content of the hazelnut component. In fact, the application of LDA to a dataset containing the hazelnut-only products (data in Figure S1) highlighted that Mo, Ni and Sr were the more efficient elements in the discrimination among the three cultivars, with a minor role of Ce, La, Mn and Rb. We may hypothesize that the addition of cocoa powder and other ingredients does not influence the original content of these elements, while it somewhat influences the original content of Ba and Cs. Following the indications of LDA, Mo and Ni were selected among the variables with the highest discriminating power and their 2D plot (Figure 4) allowed obtaining a good separation among the three cultivars.

MOR
93.97% 95.17% 96.61% The selected variables were Mn, Ni, Zr, Mo, Ba and W. This behavior seems to partly reflect the elemental content of the hazelnut compo nent. In fact, the application of LDA to a dataset containing the hazelnut-only product (data in Figure S1) highlighted that Mo, Ni and Sr were the more efficient elements in th discrimination among the three cultivars, with a minor role of Ce, La, Mn and Rb. We ma hypothesize that the addition of cocoa powder and other ingredients does not influenc the original content of these elements, while it somewhat influences the original conten of Ba and Cs. Following the indications of LDA, Mo and Ni were selected among the var iables with the highest discriminating power and their 2D plot (Figure 4) allowed obtain ing a good separation among the three cultivars. In the end, therefore, Mo, and Ni worked efficiently as the chemical markers of th original hazelnut cultivars.
Similar results were achieved after removing the information about the type of prod uct, operating a centering of the data according to the three types of products classes sep arately.

Verivication with Samples Prepared in Laboratory
A first attempt to evaluate the suitability of the method was made by preparing tw samples of Gianduja pastes and two samples of Gianduiotto pastes with variable amount of hazelnuts of known origin (Table 4). All four paste samples have features somewha intermediate among the three cultivars, with samples Gianduja 1 and Gianduiotto 1 pos sessing a higher percentage of Piemonte hazelnuts.  In the end, therefore, Mo, and Ni worked efficiently as the chemical markers of the original hazelnut cultivars.
Similar results were achieved after removing the information about the type of product, operating a centering of the data according to the three types of products classes separately.

Verivication with Samples Prepared in Laboratory
A first attempt to evaluate the suitability of the method was made by preparing two samples of Gianduja pastes and two samples of Gianduiotto pastes with variable amounts of hazelnuts of known origin (Table 4). All four paste samples have features somewhat intermediate among the three cultivars, with samples Gianduja 1 and Gianduiotto 1 possessing a higher percentage of Piemonte hazelnuts. Table 4. Samples of Gianduja and Gianduiotto pastes prepared with mixtures of hazelnuts of known origin. The column "Hazelnuts %" indicates the total amount of hazelnuts in the samples.
For every product, three independent replicates were collected and treated for a total of 54 samples.

Sample Treatment
All samples were subjected to dry ashing supported by microwave irradiation by using a Milestone (Sorisole, Italy) Pyro 260 microwave ashing system. Dry ashing was chosen as pre-treatment instead of acid digestion, because it allows treating a higher amount of sample (usually 10× with respect to acid digestion), maximizing the content of trace elements in the solutions to be analyzed. Its major drawback is the loss of volatile elements (e.g., As, Cd, Hg and Pb) but we evaluated that they were not relevant descriptors for traceability.
For dry ashing, 20 g of each sample was weighed and placed in a porcelain capsule. The heating cycle was as follows: room temperature to 150 • C in 10 ; hold at 150 • C for 20 ; up to 500 • C in 20 ; hold at 500 • C for 30 ; up to 750 • C in 10 ; hold at 750 • C for 30 ; up to 900 • C (raw hazelnuts, roasted hazelnuts, hazelnut paste and hazelnut cream) or to 1000 • C (Gianduja and Gianduiotto paste) in 10 ; hold at 900 • C (raw hazelnuts, roasted hazelnuts, hazelnut paste and hazelnut cream) or to 1000 • C (Gianduja and Gianduiotto paste) for 30 . The higher final temperature for Gianduja and Gianduiotto paste was needed in order to obtain a more efficient incineration. The resulting ash was completely dissolved in 2.0 mL of ultrapure concentrated nitric acid and taken up to 45 mL with HPW in a polypropylene tube.

ICP-OES Analysis
The determination of major and minor elements was carried out with a Spectro (SPEC-TRO Analytical Instruments GmbH, Kleve, Germany) Genesis ICP-OES simultaneous spectrometer with axial plasma observation. Instrumental parameters were as follows: pump speed, 2.0 mL/min; RF generator, 40 MHz; RF, 1300 W; plasma power, 1400 W; plasma gas outlet, 12 L/min; auxiliary gas flow rate, 0.90 L/min; nebulizer flow rate, 0.96 L/min. The elements determined were the following (in parentheses the ƒ ⊂ of acquisi- hazelnuts. It was possible to discriminate among the various stages of the production chain according to the content of Y, Zn, Cu, Si, Li, S and Co and, what it is more interesting, among the different cultivars used as raw materials according to the content of Mo, Ni, Ba, Mn, Zr and W. Further experiments are needed in order to address to different points: (1) verify the possibility of discrimination among the cultivars including other cultivars and increasing the number of samples for each cultivar; (2) explore more exhaustively the possibility to discriminate the cultivars also in presence of mixtures of different cultivars.
Supplementary Materials: The following supporting information can be downloaded at: https:// www.mdpi.com/article/10.3390/molecules27123854/s1, Figure S1: LDA score plot of the 27 samples of hazelnut-only products, F1 vs. F2 for the classification according to the cultivar: score plot of the canonical variables; Table S1: Concentrations of analytes (mg/Kg) determined by means of ICP-OES and ICP-MS; Table S2: LOD and LOQ for the elements determined with ICP-OES and ICP-MS.