Non-Targeted NMR Method to Assess the Authenticity of Saffron and Trace the Agronomic Practices Applied for Its Production

applications the omics analytical tools to food traceability, quality control, and authenticity Abstract: The development of analytical methods aimed at tracing agri-food products and assessing their authenticity is essential to protect food commercial value and human health. An NMR-based non-targeted method is applied here to establish the authenticity of saffron samples. Speciﬁcally, 40 authentic saffron samples were compared with 18 samples intentionally adulterated by using turmeric and safﬂower at three different concentration levels, i.e., 5, 10, and 20 wt%. Statistical processing of NMR data furnished useful information about the main biomarkers contained in aqueous and dimethyl sulfoxide extracts, which are indicative of the presence of adulterants within the analyzed matrix. Furthermore, a discrimination model was developed capable of revealing the type of agronomic practice adopted during the production of this precious spice, distinguishing between organic and conventional cultivation. The main objective of this work was to provide the scientiﬁc community involved in the quality control of agri-food products with an analytical methodology able to extract useful information quickly and reliably for traceability and authenticity purposes. The proposed methodology turned out to be sensitive to minor variations in the metabolic composition of saffron that occur in the presence of the two adulterants studied. Both adulterants can be detected in aqueous extracts at a concentration of 5 wt%. A lower limit of detection was observed for safﬂower contained in organic extracts in which case the lowest detectable concentration was 20%.


Introduction
In recent years, globalization of the food supply chain has occurred at a considerable level, and accompanying it, also an escalation of food fraud events. According to the last technical report drawn up by the Food and Agriculture Organization (FAO), it is estimated that the cost of food fraud in the global food industry is approximately 30 billion EUR every year [1]. To face the high economic damage and to prevent potential high risk for human health, the scientific community is working hard to develop new technologies and new analytical approaches to achieve complete traceability of the food chain. In this context, the combination of blockchain and artificial intelligence is becoming a valuable tool to trace the several stages within the food chain, including all processes from the acquisition of raw materials to production, consumption, and disposal [2]. Nevertheless, one of the main challenges is to assess food features by detecting variations in the metabolic composition of the product under investigation. The "traditional" approach relies on a targeted analytical method, by which a few a priori known compounds are detected and, eventually, quantified. However, this approach may fail when unknown substances are indicative of either accidental or intentional alteration of the food product. In this contest, the non-targeted analytical method is becoming a more valuable tool for quality control and authenticity assessment [3].
Over the last decades, the use of an NMR-based analytical approach for food quality and authentication has significantly increased [4][5][6]. The distinctive ability of NMR spectroscopy compared to other analytical techniques is that of providing reproducible spectral data in terms of relative signal intensity, regardless of the spectrometer features [7,8]. Recently, our research group contributed to the harmonization of NMR data production by differently configured spectrometers [9,10]. A further attractive feature of the NMR-based analytical approach relies on the possibility of collecting a large amount of data related to the same sample. Such an NMR feature allows the entire metabolic profile (fingerprint) of a food product to be obtained rapidly and reproducibly [11][12][13][14][15]. Thus, variations of food fingerprints derived from possible alterations can be easily unveiled by NMR analysis at any stage of the control process in the supply chain. As a consequence, NMR data can be exploited, through Distributed Ledger Technologies such as blockchain, to trace food products and certify their quality, origin, and authenticity.
In this work, we report on the application of the non-targeted NMR method toward the detection of adulterated saffron, which represents one of the most expensive spices on the market (up to 20.000 €/kg). This very high price, although justified by the hard manual labor together with its limited production (around 150,000 flowers are needed to produce 1 kg of saffron) [16], makes saffron very susceptible to economically motivated adulterations (EMAs). One of the most common strategies for adulterating saffron is to add natural phytochemicals able to confer organoleptic and/or commercial traits like those naturally belonging to saffron. Safflower (Carthamus tinctorius L.) and turmeric (Curcuma longa L.) represent the two most frequent contaminants found in adulterated saffron on the market [17].
Currently, the quality of saffron is determined by specifications described within the ISO 3632-2:2011 [18] through the application of a typical targeted analytical method. According to ISO 3632, saffron can be classified into three categories of quality (I, II, III) depending on the concentration of the three known metabolites, namely crocins, picrocrocin, and safranal, contained in this spice and responsible for its typical taste and color. Such metabolites are quantified by a spectrophotometric analytical protocol. Nevertheless, it has been reported that this analytical protocol may not be able to detect amounts of safflower or turmeric lower than 20 wt% [19], as well as other kinds of adulterants in low concentration [20]. Further reported studies explored the application of alternative analytical methods to ascertain the presence of contaminants in saffron under the threshold value of 20% through the use of MALDI mass spectrometry [21], diffuse reflectance infrared Fourier transform spectroscopy [22], Raman spectroscopy [23], and electronic nose [24]. In this context, NMR spectroscopy coupled with chemometric analysis has been reported as a suitable approach for the quality assessment and authentication of this precious spice, being able to detect a wide range of adulterants, from synthetic dyes to herbal materials [25][26][27][28][29][30][31][32][33][34][35][36][37][38][39][40][41]. However, to date, no official NMR analytical protocol is available to detect levels of contaminants in saffron below the threshold level of 20 wt%. In the present work, we explored the capability of the NMR-based non-targeted method to detect contamination of saffron samples with a threshold level of 5 wt% of safflower (Carthamus tinctorius L.) and turmeric (Curcuma longa L.). Aiming at identifying the metabolites which can be exploited as biomarkers to detect the adulteration taken place, the study was conducted on saffron samples extracted both in water and in DMSO-d 6 . As a result, a pool of metabolites indicative of the contamination of the saffron samples was identified. In addition, during the present study, the developed non-targeted method was exploited to obtain useful insights into the agronomical practice adopted during the cultivation of the saffron under investigation, allowing discrimination between organic saffron and the conventional type.

Plant Material
The saffron stigmas (Crocus sativus L.) submitted for investigation were provided by four Italian farms. Table 1 summarizes the number of samples (n) provided by each farm together with all the information on the geographical origin, and the agronomic practice applied for their cultivation. Saffron samples were adulterated by mixing the pure saffron powder with turmeric (Curcuma longa) or safflower (Carthamus tinctorius) powders to get turmeric-adulterated samples and safflower-adulterated samples, respectively. Three levels of adulteration were obtained: 20 wt%, 10 wt%, and 5 wt%. Three replicates were prepared for each level of adulteration, obtaining a total of 18 samples of adulterated saffron.

Preparation of Organic Extracts in DMSO-d 6
An aliquot of 10 mg of each powdered sample was added to 600 µL of DMSO-d 6 , vortexed (Advanced Vortex Mixer ZX3, VELP Scientifica Srl, Usmate Velate, Italy) for 3 min at 2500 rpm, and centrifuged for 10 min at 6000 rpm (ROTOFIX 32 A, Hettich, Milan, Italy). The DMSO extracts proved to be stable within 24 h even without pasteurization (see Figure S1c in Supplementary Materials for further details). NMR tubes were filled in by an automated system for liquid handling (SamplePro Tube, Bruker BioSpin, Bruker, Billerica, MA, USA) according to the following proportion: 540 µL of the obtained extract and 60 µL of TSP/D 2 O solution [3-(trimethylsilyl)propionic-2,2,3,3-d 4 acid sodium salt in D 2 O (0.20%)].

NMR Experiment
One-dimensional 1 H NOESY NMR and 1 H NMR spectra were recorded through a Bruker Avance 400 MHz spectrometer equipped with a 5 mm inverse probe and with an autosampler (Bruker, Billerica, MA, USA). The following acquisition parameters were used to record 1  Each spectrum was acquired using TOPSPIN 2.1 software (Bruker BioSpin GmbH, Rheinstetten, Germany) under an automatic process that lasted ca. 15 min and encompassed sample loading, temperature stabilization for 5 min, tuning, matching, shimming, and 90 • pulse calibration.
NMR raw data (Free Induction Decays, FIDs) were processed using the software MestReNova 11.0 (Mestrelab Research SL, Santiago de Compostela, Spain). The FIDs were zero-filled to 128 K number of points and then underwent Fourier transformation by applying an exponential multiplication function with a line broadening of 0.1 Hz. Phase and baseline were automatically corrected, and the TSP singlet signal set at δ = 0.00 ppm was used as a chemical shift reference.

Pre-Treatment of Raw Data for the Statistical Analysis
The raw data (FIDs) relative to the 1D 1 H NOESY NMR and 1 H NMR: experiments were processed by a single operator using Mestrelab and segmented into regular-sized (0.04 ppm) intervals (buckets) in the range of [10.50, 0.50] ppm. The underlying area of each bucket was calculated and normalized to the total intensity. The areas of the buckets in the region [5.13, 4.69] ppm, corresponding to the residual water signal, were set to 0. The data matrices were imported into MetaboAnalyst 5.0, and buckets were subjected to meancentering and divided by the standard deviation of each variable (Unit Variance scaling). Multivariate statistical analyses such as Principal Component Analysis (PCA), Hierarchical Clustering Dendrogram (HCD), and Partial Least Square-Discriminant Analysis (PLS-DA) were performed. PCA and HCD were used as unsupervised approaches to get an overview of the data. PLS-DA was used as a supervised method that uses multivariate regression techniques to extract via a linear combination of original variables (X) the information that can predict the class membership (Y). The PLS regression was performed using the plsr function provided by R pls package [42]. The classification and cross-validation were performed using the corresponding wrapper function offered by the caret package [43]. To assess the significance of class discrimination, a permutation test was performed. In each permutation, a PLS-DA model was built between the data (X) and the permuted class labels (Y) using the optimal number of components determined by cross-validation for the model based on the original class assignment [44].

Metabolic Analysis of Authentic Saffron Aqueous Extracts
The NMR analysis (1D 1 H NOESY, COSY, and TOCSY) performed on the aqueous extracts of the authentic pure saffron enabled the identification of a pool of metabolites (see Table S1 in Supplementary Materials for the full list of metabolites). The main representative classes of metabolites detectable in the aqueous extracts of authentic saffron (see Figure S2 in Supplementary Materials for further details) include organic acids (lactic, acetic, malic, fumaric, and formic acids), amino acids (alanine, asparagine, glycine), and carbohydrates (D-glucose, fructose, D-gentiobiose, and D-sucrose). Additionally, the colorless monoterpene glucosidepicrocrocin (4-[β-d-glucopyranosyloxy]-2,6,6-trimethyl-1-cyclohexene-1-carboxaldehyde (hereafter referred to as picrocrocin) has been reported as abundantly present in the aqueous extracts of saffron. It represents one of the metabolites majorly responsible for the typical bitter taste of saffron [45]. However, it has been previously reported as occurring from degradation of picrocrocin in aqueous saffron spice extracts upon thermal treatment [46]. Therefore, since the analyzed samples were subjected to a pasteurization process at 80 • C to keep the samples stable for 24 h (see Figure S1 in Supplementary Materials for stability studies), aglycone 4-hydroxy-2,6,6-trimethylcyclohex-1-enecarbaldehyde (HTCC) was expected to be detected as a degradation product. Indeed, characteristic signals as singlets ascribable to this compound were detected at 1.20, 1.23, and 9.97 ppm. Correspondingly, the saccharidic moiety, namely β-D-glucose, derived from the degradation of picrocrocin was identified by the well-resolved doublets at 5.24 and 4.65 ppm, and the doublet of doublets at 3.26 ppm. Moreover, it has been reported that the analyzed aqueous extracts contain significant amounts of a group of water-soluble carotenoids derived from crocetin (8,8 -diapo-Ψ,Ψ -carotenedioic acid) which are responsible for the coloring capacity of saffron and the use of this spice as food colorant [47]. While a pool of carotenoids derived from crocetin (8,8 -diapo-Ψ,Ψ -carotenedioic acid) has been identified through LC-ESI-MS using water (or acidified with 0.25% formic acid) and acetonitrile [48], the 1D 1 H NOESY spectrum in D 2 O, in agreement with the literature [26], contains predominantly the signals related to one of the identified derivatives, namely trans crocetin (β-D-gentibiosyl) ester, commonly appointed as crocin with the trans isomer significantly more abundant than the cis one, which has not been detected. Indeed, in our case, characteristic signals ascribable to the polyconjugated olefinic protons of trans crocetin (β-D-gentibiosyl) ester are present at 7.26, 6.42, and 6.12 ppm, while the broad singlets related to the methyl moieties can be found at 1.87 and 1.77 ppm. In addition, kaempferol, which is a flavonol under the class of flavonoids, is contained in the saffron aqueous extracts. While stigma samples have been reported to contain high levels of crocin, picrocrocin, and their derivatives, the stamens and sepals of saffron have been reported as rich in flavonols which are present under three forms: kaempferol-3-sophoroside, kaempferol-3-sophoroside-7glucoside, and kaempferol-3,7,4 -triglucoside [49]. Kaempferol can be detected in the 1D 1 H NOESY NMR spectrum of aqueous extracts of stigmas as a mixture of two glycosylated forms, which have been characterized as 7-O-glucopyranoside-3-O-sophoroside, and the 7-O-sophoroside glycosides of kaempferol [49]. In agreement with the literature [26], two doublets at 8.03 and 7.94 ppm, and, two further doublets at 6.97 and 6.89 ppm can be ascribed to these two forms of kaempferol.

Statistical Analysis to Reveal Adulterated Saffron Aqueous Extracts Univariate Analysis
The binned 1D 1 H NOESY NMR spectra data constituted 58 (samples) by 227 (spectra bins) data matrix. Upon normalization to constant sum and auto-scaling, the obtained data were subjected to a univariate analysis method providing a preliminary overview of features that are potentially significant in discriminating the samples under study (see Figure S3 in Supplementary Materials for a visual inspection of 1D 1 H NOESY NMR spectra of authentic saffron aqueous extracts and adulterated ones). The unpaired two-sample Wilcoxon test (also known as the Wilcoxon rank-sum test or Mann-Whitney test) as a non-parametric alternative to the unpaired two-samples t-test was used to compare two independent groups of samples [50]. As a result, 72 significant features selected by t-tests with a p-value threshold of 0.05 were detected (see Table S2 in Supplementary Materials for the full list of the top 50 features identified by the Wilcoxon rank-sum test). Among these, the most important variables were related to the spectral regions containing signals ascribed to crocin at 1.79, 6.15, 6.47, and 5.63 ppm, glucose at 3.67 ppm, and formic acid at 8.43 ppm (Figure 1). While variables related to crocin and formic acid were positively correlated to the pure saffron samples (see Supplementary Materials, Figure S4a-e), the variable related to glucose was positively related to the adulterated samples (see Supplementary Materials, Figure S4f). Hierarchical Clustering An agglomerative hierarchical cluster analysis was performed using the hclust function in package stat, to detect similarities among the samples under investigation. Selecting the opportune similarity measure and clustering algorithm, each sample begins as a

Hierarchical Clustering
An agglomerative hierarchical cluster analysis was performed using the hclust function in package stat, to detect similarities among the samples under investigation. Selecting the opportune similarity measure and clustering algorithm, each sample begins as a separate cluster and the algorithm proceeds to combine them until all samples belong to one cluster. Specifically, using Pearson as distance measure and average as a clustering algorithm (clustering uses the centroids of the observations), it was possible to obtain useful information on the metabolites that most characterized the authentic saffron samples compared to the adulterated ones, and vice versa. As a result, the investigated samples clustered into two main groups as observed in the obtained dendrogram (see Figure S5 in Supplementary Materials for further details). The first cluster was composed of pure saffron samples, while the second group contained adulterated samples. Interestingly, the samples which were adulterated with safflower (indicated with S) clustered separately from the samples adulterated with turmeric (indicated with T). Moreover, within each group of adulterated samples, a clear similarity according to the level of adulteration (5, 10, and 20 wt%) was noticeable.
When the results of the hierarchical analysis were analyzed through a heatmap, further information could be extrapolated ( Figure 2). In agreement with the results provided by the univariate analysis, the adulterated samples were characterized by higher levels of saccharides, like sucrose (3.59 ppm), fructose (buckets at 3.99 and 4.03 ppm, respectively), and glucose (buckets at 3.67 and 5.39 ppm, respectively, containing also the overlapping signals related to the gentiobiose bound to crocetin). In particular, the samples of saffron previously adulterated with turmeric presented a higher content of these saccharides. On the other hand, the samples adulterated with safflower were characterized by a higher level of succinic acid (bucket at 2.55 ppm). The samples of authentic saffron presented chiefly predominance of crocins (buckets at 1.79, 1.87, 5.63, 6.15, 6.43, and 6.47 ppm) and a shift of the formic acid signal (buckets at 8.43 ppm). Aiming at developing a model to discriminate the adulterated saffron samples from the authentic ones, PLS-DA was applied to the data matrix obtained from the binned and Aiming at developing a model to discriminate the adulterated saffron samples from the authentic ones, PLS-DA was applied to the data matrix obtained from the binned and pre-processed spectral data. The model presented eight components, though the first seven components were selected as the optimal number of components for classification according to a 10-fold cross-validation method. To estimate the predictive ability of the model, the performance measure, Q2, was calculated via cross-validation (CV). In each CV, the predicted data were compared with the original data, and the sum of squared errors was calculated. The prediction error was then summed over all samples (Predicted Residual Sum of Squares or PRESS). For convenience, the PRESS was divided by the initial sum of squares and subtracted from 1 to resemble the scale of the R2. A model characterized by good predictions and without overfitting should have low PRESS or high Q2, as in the case of the PLS-DA model developed here (see Figure S6 in Supplementary Materials for further details). The PLS-DA model was validated through a permutation test upon passing a set of 1000 permutations and testing the prediction accuracy during training. As a result, the observed statistic p-value was lower than 0.001, confirming the good predictive performance of the developed model (see Figure S7 in Supplementary Materials for further details). The first four components exerted the highest contribution with an explained variance of 62.5% (see Figure S8 in Supplementary Materials for further details). Indeed, the PLS-DA scores plot between Component 1 and Component 2 clearly shows two clusters along Component 1 with an explained variance of 14.4% (Figure 3a). The analysis of the Variable Importance in Projection (VIP) was performed to obtain insights into the variables, and, thus, the metabolites, which mostly contributed to the observed distribution of saffron samples in the PLS-DA scores plot. Indeed, VIP represents a weighted sum of squares of the PLS loadings considering the amount of explained Y-variation in each dimension. As a result, a pool of 66 variables with VIP > 1 on the Component 1 was identified (see Table S3 in Supplementary Materials for the full list). A deeper look into the top 15 VIPs (Figure 3b), revealed that crocin represented the main metabolite contributing to the clustering of the authentic saffron samples which were characterized by a higher content of this compound. Such a result agrees with the information obtained from the univariate data analysis of the same aforementioned samples. On the other hand, the group of adulterated saffron samples presented a higher content of saccharides, like fructose and glucose.
A visual inspection of the spectra of authentic and adulterated saffron samples confirmed the results of the class discrimination since some noticeable variations were observed in the spectral regions containing the signals ascribed to the aforementioned metabolites. Specifically, the content of crocin was relatively higher in the samples of authentic saffron (Figure 4, blue spectrum), while fructose was more abundant in the samples of saffron adulterated with safflower ( Figure 4, red spectrum). Significant variations were observed also for the spectral regions containing the signals assigned to the organic acids like formic acid and succinic acid. While for formic acid a slight shift of the corresponding signal was detected, in the case of succinic acid a remarkable abundance of this metabolite in both the adulterated samples (Figure 4, red and green spectra) was observed compared to the authentic ones.

Statistical Analysis to Reveal Adulterated Saffron Organic Extracts
To reveal the adulteration through the study of the less hydrophilic metabolites, a spectroscopic non-targeted method was applied also to the saffron samples obtained by extraction with DMSO-d 6 (see Section 2.3.3 for further details). The determination of the metabolic composition of authentic saffron was achieved by NMR analysis (1D 1 H NOESY, COSY, and TOCSY), and the results were compared to spectral data reported in the literature [27,32]. Less hydrophilic metabolites could be detected, like safranal, glycosylated forms of kaempferol, linoleic acid, and further fatty acids (see Figure S9 and Table S4 in Supplementary Materials for further details). A data matrix (57 samples by 245 spectra bins) was obtained from the processing (normalization to the sum and auto-scaling) of the recorded 1 H NMR spectra (see Figure S10 in Supplementary Materials for a visual inspection of 1 H NMR spectra of extracts in DMSO-d 6 of authentic and adulterated saffron samples) and was, next, subjected to univariate analysis (t-test) and unsupervised multivariate analysis (PCA) in analogy to the statistical elaboration performed on the corresponding aqueous extracts. As a result, a less noticeable separation between the authentic samples and the adulterated ones could be extrapolated. Indeed, discrimination between these groups of samples could be detected only along the PC5 with a relatively low (4.5%) explained variance (see Figure S11 in Supplementary Materials for further details). A more noticeable clustering of these two classes of samples could be detected through supervised multivariate data analysis, namely OPLS-DA, as depicted in the corresponding scores plot (Figure 5a). Such a model was characterized by one predictive component and four orthogonal ones, and it was successfully validated upon passing the permutation test (see Figures S12 and S13 in Supplementary Materials for further details). As illustrated in Figure 5b, the top ten variables with VIP[t] > 1.70 on Component 1 (t score) included buckets related to the compounds belonging to the class of curcuminoids (indicated in Figure 5b as Cur), namely curcumin, demethoxycurcumin, and bisdemethoxycurcumin. These compounds are natural phenols and they are often employed to color foods and medicines, because of the typical yellow color derived from them [51]. They were previously isolated from turmeric and characterized via NMR [52]. The adulterated samples presented a relatively higher content of curcuminoids (Cur), along with linoleic acid (Lin. Ac.), picrocrocin (Picr), and safranal (Safr).
OR PEER REVIEW 9 of 16 group of adulterated saffron samples presented a higher content of saccharides, like fructose and glucose.
(a) (b) A visual inspection of the spectra of authentic and adulterated saffron samples confirmed the results of the class discrimination since some noticeable variations were observed in the spectral regions containing the signals ascribed to the aforementioned metabolites. Specifically, the content of crocin was relatively higher in the samples of authentic saffron (Figure 4, blue spectrum), while fructose was more abundant in the samples of saffron adulterated with safflower ( Figure 4, red spectrum). Significant variations were observed also for the spectral regions containing the signals assigned to the organic acids like formic acid and succinic acid. While for formic acid a slight shift of the corresponding The analysis of the hierarchical clustering heat map performed on the normalized data (distance measurement, Euclidean; clustering method, Ward) helped to identify the spectral regions most susceptible to changes in the metabolic composition induced by the two different types of adulteration, namely by addition of turmeric vs. safflower ( Figure 6). It is interesting to note that the adulteration through turmeric could be spotted even at a very low concentration, i.e., 5 wt%. Indeed, diagnostic buckets were those related to crocin (1.97 ppm), picrocrocin (1.76 ppm), curcuminoids (7.07, 7.01, 7.33, and 6.05 ppm), and safranal (5.97 ppm). On the other hand, the samples adulterated with safflower could be detected only at higher levels of concentration, like 20 wt%. In this case, the spectral regions containing the signals of linoleic acid (at 0.85 ppm) and crocin (at 7.33 ppm which overlaps the curcuminoid signals) could be adopted as diagnostic metabolites to detect the adulteration with safflower.
A visual inspection of the spectra of authentic and adulterated saffron samples confirmed the results of the class discrimination since some noticeable variations were observed in the spectral regions containing the signals ascribed to the aforementioned metabolites. Specifically, the content of crocin was relatively higher in the samples of authentic saffron (Figure 4, blue spectrum), while fructose was more abundant in the samples of saffron adulterated with safflower ( Figure 4, red spectrum). Significant variations were observed also for the spectral regions containing the signals assigned to the organic acids like formic acid and succinic acid. While for formic acid a slight shift of the corresponding signal was detected, in the case of succinic acid a remarkable abundance of this metabolite in both the adulterated samples (Figure 4, red and green spectra) was observed compared to the authentic ones.

Statistical Analysis to Reveal Adulterated Saffron Organic Extracts
To reveal the adulteration through the study of the less hydrophilic metabolites, a spectroscopic non-targeted method was applied also to the saffron samples obtained by extraction with DMSO-d6 (see Section 2.3.3 for further details). The determination of the metabolic composition of authentic saffron was achieved by NMR analysis (1D 1 H NO-ESY, COSY, and TOCSY), and the results were compared to spectral data reported in the recorded 1 H NMR spectra (see Figure S10 in Supplementary Materials for a visual inspec tion of 1 H NMR spectra of extracts in DMSO-d6 of authentic and adulterated saffron sam ples) and was, next, subjected to univariate analysis (t-test) and unsupervised multivariate analysis (PCA) in analogy to the statistical elaboration performed on the corresponding aqueous extracts. As a result, a less noticeable separation between the authentic samples and the adulterated ones could be extrapolated. Indeed, discrimination between these groups of samples could be detected only along the PC5 with a relatively low (4.5%) ex plained variance (see Figure S11 in Supplementary Materials for further details). A more noticeable clustering of these two classes of samples could be detected through supervised multivariate data analysis, namely OPLS-DA, as depicted in the corresponding scores plo (Figure 5a). Such a model was characterized by one predictive component and four or thogonal ones, and it was successfully validated upon passing the permutation test (see Figures S12 and S13 in Supplementary Materials for further details). As illustrated in Fig  ure 5b, the top ten variables with VIP[t] > 1.70 on Component 1 (t score) included buckets related to the compounds belonging to the class of curcuminoids (indicated in Figure 5b as Cur), namely curcumin, demethoxycurcumin, and bisdemethoxycurcumin. These com pounds are natural phenols and they are often employed to color foods and medicines because of the typical yellow color derived from them [51]. They were previously isolated from turmeric and characterized via NMR [52]. The adulterated samples presented a rel atively higher content of curcuminoids (Cur), along with linoleic acid (Lin. Ac.), picro crocin (Picr), and safranal (Safr). The analysis of the hierarchical clustering heat map performed on the normalized data (distance measurement, Euclidean; clustering method, Ward) helped to identify the spectral regions most susceptible to changes in the metabolic composition induced by the two different types of adulteration, namely by addition of turmeric vs. safflower ( Figure  6). It is interesting to note that the adulteration through turmeric could be spotted even a a very low concentration, i.e., 5 wt%. Indeed, diagnostic buckets were those related to

Fingerprinting for Determining the Agronomic Practice Adopted for Saffron Cultivation
The non-targeted NMR approach applied to the spectroscopic data derived from the analysis of the aqueous extracts of pure saffron allowed extrapolation of insightful information related to the agronomic practice adopted for this spice cultivation. In the PC1/PC2 scores plot (see Supplementary Materials, Figure S11), a clear clustering of the samples was observed when the regular-sized (0.04 ppm) spectral buckets were subjected to the unsupervised PCA. Through a more in-depth examination of the characteristics of the sample, including the geographical origin, the producer, and the agronomic practice adopted for cultivation (see Table S5 in Supplementary Materials for the complete list of available information relating to the investigated samples), the agronomic practices were found to be the most influencing factor on samples distribution. The supervised PLS-DA (see Figures S15 and S16 in Supplementary Materials for further details on the model overview and model validation) was conducted to obtain more information regarding changes in metabolic composition associated with the agronomic practice employed. The analysis of the Variable Importance in Projection, VIP (see Table S5 in Supplementary Materials for the full list of 78 variables with VIP > 1), along with the visual inspection of the 1D 1 H NOESY NMR spectra allowed the selection of a pool of 15 spectral regions acting as diagnostic variables towards the discrimination between the saffron samples obtained through conventional cultivation compared to those derived from organic cultivation (Figure 7b). Appl. Sci. 2022, 12,  . On the other hand, the samples adulterated with safflower could be detected only at higher levels of concentration, like 20 wt%. In this case, the spectral regions containing the signals of linoleic acid (at 0.85 ppm) and crocin (at 7.33 ppm which overlaps the curcuminoid signals) could be adopted as diagnostic metabolites to detect the adulteration with safflower. Figure 6. Results of the hierarchical clustering heat map performed on the normalized data (distance measure using Euclidean, and clustering algorithm using Complete). Each colored cell on the map corresponds to a concentration value in the data table, with samples in columns and metabolites in rows. The belonging class of samples is indicated as red and green squares for adulterated saffron (AS) and authentic one (PS), respectively.

Fingerprinting for Determining the Agronomic Practice Adopted for Saffron Cultivation
The non-targeted NMR approach applied to the spectroscopic data derived from the analysis of the aqueous extracts of pure saffron allowed extrapolation of insightful information related to the agronomic practice adopted for this spice cultivation. In the PC1/PC2 scores plot (see Supplementary Materials, Figure S11), a clear clustering of the samples was observed when the regular-sized (0.04 ppm) spectral buckets were subjected to the unsupervised PCA. Through a more in-depth examination of the characteristics of the sample, including the geographical origin, the producer, and the agronomic practice adopted for cultivation (see Table S5 in Supplementary Materials for the complete list of available information relating to the investigated samples), the agronomic practices were found to be the most influencing factor on samples distribution. The supervised PLS-DA (see Figures S15 and S16 in Supplementary Materials for further details on the model overview and model validation) was conducted to obtain more information regarding changes in metabolic composition associated with the agronomic practice employed. The analysis of the Variable Importance in Projection, VIP (see Table S5 in Supplementary Materials for the full list of 78 variables with VIP > 1), along with the visual inspection of the 1D 1 H NOESY NMR spectra allowed the selection of a pool of 15 spectral regions acting as diagnostic variables towards the discrimination between the saffron samples obtained through conventional cultivation compared to those derived from organic cultivation (Figure 7b). The hierarchical clustering heatmap performed on the normalized data (distance measure, Euclidean; Clustering method, Ward) clearly shows that the aqueous extracts of saffron cultivated according to an organic agronomic practice were characterized by a relatively higher content of sugars, like sucrose and gentiobiose, along with organic acids, as succinic and malic acids. On the other hand, it was found that the two most typical metabolites of saffron, namely picrocrocin and crocin, were contained majorly in those samples derived from a conventional agronomic practice (Figure 8). The hierarchical clustering heatmap performed on the normalized data (distance measure, Euclidean; Clustering method, Ward) clearly shows that the aqueous extracts of saffron cultivated according to an organic agronomic practice were characterized by a relatively higher content of sugars, like sucrose and gentiobiose, along with organic acids, as succinic and malic acids. On the other hand, it was found that the two most typical metabolites of saffron, namely picrocrocin and crocin, were contained majorly in those samples derived from a conventional agronomic practice (Figure 8).
The hierarchical clustering heatmap performed on the normalized data (distance measure, Euclidean; Clustering method, Ward) clearly shows that the aqueous extracts of saffron cultivated according to an organic agronomic practice were characterized by a relatively higher content of sugars, like sucrose and gentiobiose, along with organic acids, as succinic and malic acids. On the other hand, it was found that the two most typical metabolites of saffron, namely picrocrocin and crocin, were contained majorly in those samples derived from a conventional agronomic practice (Figure 8). Figure 8. Results of the hierarchical clustering heat map performed on the normalized data (distance measure using Euclidean, and clustering algorithm using Ward). Each colored cell on the map corresponds to a concentration value in the data table, with samples in columns and metabolites in rows. The belonging class of samples is indicated as red and green squares for saffron cultivated according to organic farming (B) and conventional (C), respectively.

Figure 8.
Results of the hierarchical clustering heat map performed on the normalized data (distance measure using Euclidean, and clustering algorithm using Ward). Each colored cell on the map corresponds to a concentration value in the data table, with samples in columns and metabolites in rows. The belonging class of samples is indicated as red and green squares for saffron cultivated according to organic farming (B) and conventional (C), respectively.

Conclusions
In this study, a non-targeted NMR screening was performed on a group of authentic saffron samples and on a group of saffron samples intentionally adulterated by the addition of safflower and turmeric at different concentrations. The elaboration of NMR data of aqueous extracts by chemometrics allowed the discrimination between authentic and adulterated saffron samples, detecting the presence of both adulterants at a concentration value equal to 5 wt%. When a similar study was performed on the same samples extracted in DMSO-d 6 , the presence of turmeric was detected at a concentration level of 5%, while it was possible to identify the presence of safflower only at higher levels of concentration, i.e., 20 wt%. To the best of our knowledge, to date, no non-targeted NMR-based method has been reported in the literature capable of detecting the presence of contaminants at these concentration levels. Furthermore, through the same analytical approach, it was possible to obtain crucial information on the type of cultivation practice adopted for the production of saffron. The results reported in this study can help to enhance the application of non-targeted NMR-based metabolomics in the control processes of the agri-food supply chain with the final goal to valorize sustainable farming practices and ensure traceability and authenticity of food products.
Supplementary Materials: The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/app12052583/s1. Figure S1: Effect of pasteurization step on the stability of aqueous extracts of pure saffron. Table S1: List of metabolites contained in the aqueous extracts of authentic saffron samples and identified by 1D 1H NOESY NMR. Figure S2: Typical 1D 1 H NOESY NMR spectrum of aqueous extracts of authentic saffron. Figure S3: 1D 1 H NOESY NMR spectra of aqueous extracts of authentic saffron, authentic turmeric, authentic safflower, saffron adulterated with turmeric, and saffron adulterated with safflower. Table S2: List of the top 50 features identified by the Wilcoxon rank-sum test applied to the aqueous extracts of authentic and adulterated saffron samples. Figure S4: Trend of the top six most significant variables according to the Wilcoxon rank sum test applied to the aqueous extracts of authentic and adulterated saffron samples. Figure S5: Results of the hierarchical analysis applied to the spectral data obtained from the aqueous extracts of authentic and adulterated saffron samples. Figure S6: 10-fold cross-validation method applied to the PLS-DA model built on the spectral data obtained from the aqueous extracts for discriminating between authentic and adulterated saffron samples. Figure S7: Permutation test applied to the PLS-DA model built on the spectral data obtained from the aqueous extracts for discriminating between authentic and adulterated saffron samples. Figure S8: Overview of the PLS-DA model built on the spectral data obtained from the aqueous extracts for discriminating between authentic and adulterated saffron samples. Table S3: List of the top 66 variables importance in projection (VIPs) related to the PLS-DA model built on the spectral data obtained from the aqueous extracts for discriminating between authentic and adulterated saffron samples. Figure S9: Typical 1 H NMR spectrum of extracts of authentic saffron in DMSO-d 6 . Figure S10: 1 H NMR spectra of extracts of authentic saffron, authentic turmeric, authentic safflower, saffron adulterated with turmeric, and saffron adulterated with safflower in DMSO-d 6 . Table S4: List of metabolites contained extracts of authentic saffron samples in DMSO-d 6 and identified by 1 H NMR. Figure S11: Results of PCA applied to the spectral data obtained from extracts of authentic saffron and adulterated saffron in DMSO-d 6 . Figure S12: Overview of the OPLS-DA model built on the spectral data obtained from the extracts in DMSO-d 6 for discriminating between authentic and adulterated saffron samples. Figure S13: Permutation test applied to the OPLS-DA model built on the spectral data obtained from the extracts in DMSO-d 6 for discriminating between authentic and adulterated saffron samples. Table S5: List of features characteristic of each sample of authentic saffron under investigation. Figure S14: Results of PCA performed on the spectral data obtained from the aqueous extracts of authentic saffron. Figure  S15: Overview of PLS-DA built on the spectral data of the aqueous extracts of authentic saffron samples for discriminating according to the agronomic practice: organic vs. conventional. Figure S16: 10-fold cross-validation method applied to the PLS-DA model built on spectral data of the aqueous extract of authentic saffron samples for discriminating according to the agronomic practice: organic vs. conventional. Table S6