Discrimination of Cheese Products Regarding Milk Species’ Origin Using FTIR, ¹H-NMR, and Chemometrics

Abstract

The present study deals with the discrimination of various European cheese products based on spectroscopic data and chemometric analysis. It is the first study that includes cheese products from Cyprus along with cheese samples from abroad and several different cheese types. Therefore, forty-nine samples were collected, freeze-dried, and measured by using spectroscopic techniques, such as FTIR (Fourier-Transform Infrared Spectroscopy) and ¹H-NMR (proton nuclear magnetic resonance). Discriminant analysis was applied, particularly OPLS-DA. All data obtained from ¹H-NMR were included, whereas, regarding the FTIR data, only the spectral subregion between 1900 and 400 cm⁻¹ was used in the extracted model. The cheese samples were classified according to the milk species’ origin. In the future, the samples of this study will be enriched for further testing with spectroscopic techniques and chemometrics.

1. Introduction

Cheese can be made from the milk of various animal species, each contributing unique flavors and textures to the final product. For instance, cow-milk cheese, like Cheddar, Gouda, etc., has a rich and creamy flavor. Goat-milk cheese, e.g., Feta, Bucheron, etc., is tangy and has an earthy taste, while sheep-milk cheese, e.g., Pecorino Romano, Roquefort, Manchego, etc., is often slightly salty. Buffalo-milk cheese is led by Mozzarella di Bufala and is creamy and indulgent. In addition, some cheeses are made from a blend of different milks, offering a complex flavor profile [1]. The milk’s composition, fat content, and protein structure vary between species, influencing the cheese’s taste, texture, and aging process. Traditional cheese-making methods also play a role in shaping the final characteristics.

Analytical techniques like spectroscopy and chromatography help to identify unique chemical fingerprints, aiding in the accurate classification of milk from different species such as cows, goats, or sheep [2,3,4,5,6]. In particular, spectroscopic methods are very promising, with many advantages compared to other analytical techniques [7,8,9]. In dairy research, various analytical techniques can be used to characterize cheese; however, spectroscopy is a valuable tool for characterizing cheese, providing insights into its composition, structure, and quality [9,10]. Nuclear magnetic resonance (NMR) employs the detection of hydrogen nuclei for the examination of the chemical composition of fats and proteins in cheese and, as a non-destructive technique, provides detailed information on molecular composition. Infrared Spectroscopy analyzes vibrations of molecular bonds to identify functional groups in proteins, fats, and carbohydrates. Its main advantages are rapid analysis and a limited cost of use. These spectroscopic techniques help cheese producers to monitor quality, authenticity, and consistency throughout the production process, contributing to the overall understanding of cheese composition and characteristics [11].

Concerning the aqueous fraction of milk and similar dairy products, the primary divisions in the ¹H-NMR spectrum encompass three key areas. A high-field (low-frequency) segment ranging from 0 to 3 ppm is linked to signals originating from aliphatic groups of amino acids, organic acids associated with lactose and galactose glycolytic pathways, the Krebs (or citric acid) cycle, and alcohols. Within a mid–low field range spanning 3.1 to 5.3 ppm, signals are attributed to the presence of lactose. Typically, minimal signals are detected in the 5.5 to 9.0 ppm range, potentially arising from amide protons of proteins or the aromatic groups found in amino acids and phenolic compounds. In addition, acyl chains are presented in proteins; i.e., side-chain H and backbone H appear between 6 and 8 and 7 and 10 ppm, respectively. Aromatic H is located at 7–8 ppm [12,13]. In the third region, the identification of formic acid, tyrosine, and phenylalanine is easily achievable. However, challenges emerge when attempting to assign the remaining spin systems, which come from olefin or ring protons [14]. Moreover, it has been proven that amino acids with -NH₂ or -NH₄⁺ groups exchange protons with a solvent such as deuterium oxide-containing water [15].

Fourier-Transform Infrared Spectroscopy (FTIR) is another powerful analytical technique used for assessing the authenticity and quality of cheese. FTIR, combined with chemometrics, offers a robust and versatile approach to the assessment of cheese quality and authenticity, providing valuable information for both research and industry applications [10,11]. The MIR region, approximately 4000–400 cm⁻¹, provides quantitative compositional information about molecules. In the spectroscopic region of 1840–950 cm⁻¹, the FTIR spectrum can have bands due to the significant presence of lactose in dairy products [16]. In greater detail, a spectrum derived from a dairy product exhibits distinctive bands as follows: 3700–3200 cm⁻¹, indicative of -OH and -NH stretching in proteins; 3000–2800 cm⁻¹, associated with C-H stretching in fatty acids; 1750–1650 cm⁻¹, reflective of -C=O in fatty acids and esters; and 1650–1450 cm⁻¹, involving both -C=O and –NH of amides I and II in proteins. These vibrations result from various combinations within the peptide bonds and the secondary structure of casein protein. Paradkar and Irudayaraj [17] emphasized the importance of the 3200–2800 cm⁻¹ range for characterizing cholesterol. The amide I vibration, primarily associated with the stretching of C=O bonds, occurs between 1650 and 1550 cm⁻¹. Simultaneously, the amide II vibration, originating from the deformation of N–H bonds and the stretching of C-N bonds, is observed in the region from 1550 to 1450 cm⁻¹. Additionally, it is found at 1460–1150 cm⁻¹ for esters and aliphatic chains of fatty acids, and at 1200–800 cm⁻¹ for -C=O from polysaccharides and C=C stretching of acids [18,19]. Concerning milk composition, Elbassbasi et al. [20] noted the presence of bands from amino acid side-chain vibrations attributed to tyrosine at approximately 1515 cm⁻¹, phenylalanine at about 1498 cm⁻¹, and proline at around 1454 and 1438 cm⁻¹. Chemometrics plays a crucial role in classifying cheeses by analyzing huge loads of chemical data. During chemometric modeling, multivariate data analysis takes place. The application of chemometric methods usually helps to identify key variables in distinguishing between different milk types. Variable selection is another important step that shows the most relevant spectral regions for predicting the desired properties. Moreover, cross-validation is necessary to ensure the robustness of calibration models by validating them on subsets of the data. By combining chemometrics and spectroscopy, a comprehensive database of authentic sample spectra can be created for comparison with unknown samples. Model updating, thus periodically updating calibration models, is important to account for variations in raw materials or production processes. For regulatory compliance, validation of the chemometric models to meet regulatory requirements may take place [21].

The problem of food adulteration is noteworthy, even in cases where it does not directly impact public health. Adulterating food has consequences for manufacturers’ profits and the raw materials used in production. Extensive scrutiny has been directed at fraudulent practices in dairy products [16,21]. One prevalent form of adulteration is the substitution of more valuable milk, like sheep’s or goat’s milk, with lower-value cow’s milk. Hence, the precise identification of the milk species is imperative, especially for products like pure sheep or goat cheeses. This illicit practice yields economic advantages but leads to unethical competition involving the compromise of product quality [6].

This study is related to a previous one conducted by Tarapoulouzi and Theocharis [22]. The goal of this study was to obtain the FTIR and ¹H-NMR spectra of various cheese products of different geographical origins and use chemometrics to classify the samples based on milk species’ origin. In addition, another aim was to check if both spectroscopical techniques are important or if only one could serve this purpose. Subsequently, such an outcome could be used as a tool against counterfeiting and adulteration.

2. Materials and Methods

2.1. Cheese Samples

Forty-nine cheese samples of different species origin, type of cheese, and country of provenance were bought from local grocery stores in Cyprus and are presented in

Table 1. Information about the cheese samples.

Νο.	Label	Type of Cheese	Animal Origin of Milk	Country of Provenance	Νο.	Label	Type of Cheese	Animal Origin of Milk	Country of Provenance
1	C1	CHEDDAR	Cow	New Zealand	25	K10	KEFALOTYRI	Goat and Sheep	Cyprus
2	C2	CHEDDAR	Cow	Somerset, England	26	K11	KEFALOTYRI	Cow	Greece
3	C3	CHEDDAR	Cow	Somerset, England	27	K12	KEFALOTYRI	Cow	Cyprus
4	C4	CHEDDAR	Cow	Davidstow, United Kingdom	28	K13	KEFALOTYRI	Goat and Sheep	Greece
5	C5	CHEDDAR	Cow	England	29	K14	KEFALOTYRI	Sheep	Greece
6	C6	CHEDDAR	Cow	Ireland	30	Τ1	PECORINO TOSCANO	Sheep	Italy
7	C7	CHEDDAR	Cow	England	31	Τ2	PECORINO TOSCANO FRESCO	Sheep	Italy
8	C8	CHEDDAR	Cow	Ireland	32	Τ3	PECORINO TOSCANO STAGIONATTA	Sheep	Italy
9	C9	CHEDDAR	Cow	Ireland	33	Τ4	FRICO CHERVETTE	Goat	Holland
10	C10	CHEDDAR	Cow	Ireland	34	Τ5	FRICO CHERVETTE	Goat	Holland
11	C11	CHEDDAR	Cow	England	35	Τ7	PECORINO ROMANO	Sheep	Italy
12	C12	CHEDDAR	Cow	United Kingdom	36	Τ14	PECORINO ROMANO	Sheep	Italy
13	C13	CHEDDAR	Cow	Netherlands	37	Τ22	HAVARTI	Cow	Denmark
14	C14	CHEDDAR	Cow	Somerset, England	38	Τ23	KEFALOGRAVIERA	Goat and Sheep	Greece
15	C15	CHEDDAR	Cow	United Kingdom	39	Τ24	KEFALOGRAVIERA	Goat and Sheep	Greece
16	K1	KEFALOTYRI	Cow	Denmark	40	Τ25	EMMENTAL	Cow	Germany
17	K2	KEFALOTYRI	Sheep	Cyprus	41	Τ26	MANCHEGO	Sheep	Spain
18	K3	KEFALOTYRI	Cow	Greece	42	Τ27	MOZZARELLA	Cow	Denmark
19	K4	KEFALOTYRI	Goat and Sheep	Greece	43	Τ29	REGATO	Cow	Belgium
20	K5	KEFALOTYRI	Goat and Sheep	Greece	44	Τ30	GOUDA	Cow	Ireland
21	K6	KEFALOTYRI	Cow	Greece	45	Τ31	QUARK	Cow	Latvia
22	K7	KEFALOTYRI	Goat and Sheep	Cyprus	46	Τ32	EMMENTAL	Cow	France
23	K8	KEFALOTYRI	Goat and Sheep	Mytilene, Greece	47	Τ36	PECORINO CAGGIANO	Sheep	Italy
24	K9	KEFALOTYRI	Goat and Sheep	Cyprus	48	Τ37	CREAM SPREAD CHEESE	Cow	Austria
					49	Τ38	QUESO DE CABRA	Goat	Spain

. Cheddar cheese (C1–C15) is only made from cow milk, and it comes mainly from Ireland and the United Kingdom. Two Cheddar samples came from New Zealand and the Netherlands. Kefalotyri cheese (K1–K14) is made from cow or goat and sheep milk, and all the samples were made in Cyprus or Greece, and only one in Denmark. The Pecorino Toscano samples (T1–T3) were from Italy with sheep milk. The Frico Chervette (T4–T5) samples came from the Netherlands, and they were made of goat milk. The Pecorino Romano samples (T7, T14) from Italy were made of sheep milk. One Havarti cheese sample (T22) from Denmark was made of cow milk. Both Kefalograviera samples (T23–T24) originated from Greece and were made of goat and sheep milk. Two Emmental samples (T25, T32) were made of cow milk and originated from Germany and France, respectively. T26 is a Manchego cheese produced using sheep milk in Spain. T27 is a Mozzarella cheese from Denmark, made of cow milk. Respectively, T29–T31 are a Regato, a Gouda, and a Quark cheese from Belgium, Ireland, and Latvia, all made of cow milk. T36 is a Pecorino Caggiano cheese from Italy made of sheep milk. A cream cheese spread was included with code T37, made of cow milk in Austria. Lastly, T38 is a Queso De Cabra cheese made of goat milk in Spain.

Therefore, the samples contained either cow-origin milk or goat- and sheep-origin milk. Based on their label, none of the samples had a mixture of the three milk species, to keep the dataset simple.

2.2. Sample Preparation

The lyophilization protocol replicated the procedure detailed by Tarapoulouzi et al. [5] and Tarapoulouzi and Theocharis [22] and was carried out using a Christ (Osterode, Germany) Alpha 1–2 freeze drier. The condenser temperature measured 233 K, and the drying chamber’s final pressure reached 3 Pa. To freeze-dry 5 g of each cheese sample, a 5 h treatment was necessary. The resulting residue underwent homogenization using a drum grater (Ghizzoni mod. GS electric, Retsch, Haan, Germany) and was stored in sealed plastic containers at room temperature until spectroscopic measurements were conducted.

2.3. Spectroscopy Measurements

FTIR and ¹H-NMR equipment, along with their corresponding parameters, were employed in accordance with previously described methods [5,6]. The FTIR spectra were recorded twice using a Shimadzu Fourier Transform-8900 Spectrometer instrument (Kyoto, Japan), equipped with a potassium bromide (KBr) beam splitter. Therefore, samples were recorded as compressed KBr pellets, and 20 scans were averaged at a standard resolution of 8 cm⁻¹ within the 4000–400 cm⁻¹ range. To minimize interference from carbon dioxide and water vapor, the samples were measured against an air background. To obtain ¹H-NMR spectra, as detailed by Tarapoulouzi and Theocharis [22], we dissolved 80 mg of the cheese sample in 600 μL of deuterium oxide. After filtration, 500 μL was then transferred into the NMR tube. The measurements of ¹H-NMR spectra were taken by using a BRUKER AVANCE 300 Ultrashield Fourier spectrometer (Bruker BioSpin, Rheinstetten, Germany). The spectral acquisition characteristics were a 90° ¹H-pulse angle, an acquisition time of 2.9 s, 32 K data points, and a 1 sec relaxation delay time. The bin size was 0.01 ppm. Sixteen scans were recorded in about 10 min at 25 °C and applied with a spectral width of 12.0 ppm. The water peak in the NMR spectra (~4.9 ppm) was excluded and not taken into account during the data analysis.

2.4. Data Analysis

A multivariate data analysis took place using SIMCA (version 15.0.2., Umea, Umetrics, Sweden). At the beginning of the data analysis, a preliminary principal component analysis (PCA) took place for data overview, detecting outliers and groups among the observations. Group classification took place based on 95% confidence intervals. The model fit was evaluated by using R2, and Q2 was used to assess model predictability. More particularly, the Q2 value of the orthogonal projections to latent structures discriminant analysis (OPLS-DA) model was an indication of the robustness of classification. The orthogonal components have a subscript of, e.g., to1 for the first orthogonal component in X. The data preprocessing involved centering and auto-scaling to unit variance, adhering to the default settings in SIMCA. We employed the supervised method of OPLS-DA, aiming to identify distinctive patterns among the predefined groups. To evaluate the OPLS-DA model’s accuracy, a misclassification table was generated to assess the percentage of correct classifications. In the chemometric analysis of the data (variables), it was discerned that the entirety of the data from ¹H-NMR was significant, whereas only the subregion of 1900–400 cm⁻¹ from the FTIR held importance. Permutation testing and the ROC curve were utilized to evaluate the model’s performance. Permutation testing serves as a statistical measure of significance for predictive power in cross-validation. In this process, the X data remain unaltered, while the Y data undergo random permutation to assume a different sequence. Subsequently, the model is fitted to the permuted Y data, and cross-validation metrics, R2Y and Q2Y, are calculated to assess the effectiveness of the derived model. Variable influence was assessed using VIP (variable importance in the projection) graphs during OPLS-DA modeling [23].

3. Results and Discussion

3.1. ¹H-NMR Measurements

¹H-NMR revealed characteristic peaks in the cheese samples, as can be seen in Figure 1. As anticipated, the NMR spectrum indicates the prevalent presence of the lactate resonance signal at 1.33 ppm (a) across all samples. Lactate emerges as the predominant component, given its status as the primary product resulting from the fermentation of lactose during the cheese-making process [24]. It is also predominant in the peak at 5.21–5.26 ppm due to the presence of a CH anomeric of α-glucose and a CH anomeric of α-galactose. These signals likely originate from the metabolism of lactose, affirming the widely accepted notion that glucose was the favored microbial substrate among the two end products resulting from lactose hydrolysis [25]. That is why the signal of α-galactose at 5.26 ppm (b) is lower than that of β-galactose at 4.58–4.60 ppm (c) due to the CH axial proton. The results of a study by Litopoulou-Tzanetaki [26] suggested that lactic acid bacteria survived throughout the ripening of Kefalotyri cheese; thus, they catalyze the hydrolysis of lactose to α-glucose and α-galactose. Contrastingly, in Cheddar cheese, lactic acid bacteria die early, with no presence of CH anomerics of α-glucose and CH anomerics of α-galactose.

In addition, it was observed (and confirmed by chemometrics) that proline (d), methionine (e), citric acid (f), and formic acid (g) at 2.00, 2.23, 2.80, and 8.40 ppm, respectively, are different in cow and goat and sheep cheese. Moreover, in heavily salty cheese, there are relatively high levels of lactic acid at 1.42 ppm (h) and citric acid compared to other less-salty cheese samples. In contrast, the content of acetic acid at 1.93 ppm (i) can change according to the salt content, and this is the reason explaining its absence from some cheese samples of this study, like the spreadable, soft cheeses [27]. Leucine, isoleucine, and valine in the range 0.95–1.03 ppm (j) are in higher proportion in the cow-origin cheese samples. The quality of cheese is significantly influenced by the up-regulation and determination of amino acids, which are directly affected by the diet of the animals.

The potential of NMR as a method for characterizing food products appears promising. Despite its inherent low sensitivity, NMR is limited to gathering data on the most abundant elements in foods, including sugars, amino acids, and carbohydrates. Nevertheless, the extensive and detailed information that NMR can provide significantly compensates for its lower sensitivity. Special attention should paid be towards the region of weak signals ranging from 2.2 to 3.1 ppm, as these signals originate from trace compounds present in milk [12]. It is noteworthy that many authors [12,13] reported these signals with complementary measurements, suggesting that they could be further investigated with other techniques and the aid of chemometrics.

3.2. FTIR Measurements

Cow and goat and sheep cheese products have different relative peak intensities, especially in subregions 3000–2800, 1750–1650, 1650–1450, 1460–1150, and 1200–800 cm⁻¹. As seen in Figure 2, cow-origin products have more absorptions at (1) 2924 cm⁻¹, (2) 1746 cm⁻¹, (3) 1634 cm⁻¹, (4) 1460 cm⁻¹, and (5) 1163 cm⁻¹, which could be characterized as (1) strong, (2) strong, (3) medium, (4) medium, and (5) strong absorptions. However, goat-and-sheep-origin products have (1) medium, (2) medium, (3) strong, (4) strong, and (5) medium intensities, respectively. The terms strong and medium absorption refer to the intensity or magnitude of the absorption bands observed in the infrared spectrum.

For a more thorough examination of the spectroscopical data, multivariate data analyses (particularly principal component analysis and discriminant analysis) were employed to observe the differentiation and categorization of the samples.

3.3. Data Analysis

After binning, centering, and unit-variance scaling applications, the OPLS-DA analysis took place for the dataset of spectroscopic measurements. The number of components used was four, with one predictive component and four orthogonal components. Regarding ¹H-NMR, all of the data obtained were found to be necessary for the extracted model; however, the most important region from the FTIR spectra was 1900–400 cm⁻¹, as in our previous studies (Tarapoulouzi and Theocharis [6] and Tarapoulouzi et al. [5]).

As can be seen from Figure 3, only two groups were obtained—(1) cow-origin and (2) goat-and-sheep-origin samples—as was initially expected. Even though most of the samples were semi-hard and hard cheese types, there were no outliers in the model.

The outcomes presented in Figure 3 and

Table 2. Misclassification table from OPLS-DA modeling regarding species’ origin of milk for all the data obtained from ¹H-NMR spectra and the region 1900–400 cm⁻¹ from FTIR; 1—cow origin, 2—goat-and-sheep-origin samples.

	Members	Correct	1	2
1	27	100%	27	0
2	22	100%	0	22
Total	49	100%	27	22
Fisher’s prob.	4.5 × 10−14

are satisfactory, as samples from cow origin are clearly positioned on the left side of the graph, while samples from goat and sheep origins are consistently shown on the right side. In addition, 100% correct classification was obtained for the two classes regarding the species origin. It must be mentioned that it still remains a great challenge to discriminate goat and sheep samples when both categories fall in the same group, as has also happened in previous studies [5,6].

Even if 100% correct classification was obtained for the OPLS-DA model, the R2X(cum), R2Y(cum), and Q2(cum) are not equal to 1. R2X(cum) = 0.999 and R2Y(cum) = 0.970 are very close to 1; however, Q2(cum) = 0.836 is a bit lower. The literature has highlighted the challenge of establishing definitive limits for parameters associated with good predictability, as this predictability is heavily influenced by the dataset’s characteristics [23,28]. For instance, determining an acceptable threshold for Q2 depends greatly on the number of observations included. In recent years, several SIMCA PLS-DA or OPLS-DA models have been reported with Q2 values below 0.4, or even below 0.3 [29,30]. To validate models with seemingly poor predictability, permutation tests are commonly employed. These tests involve comparing the Q2 obtained for the original dataset with the distribution of Q2 values calculated when original Y values are randomly assigned to individuals, as illustrated in Figure 4. Green circles represent R2, while blue squares represent Q2. The values in the top right corner, where correlation = 100%, correspond to the R2 and Q2 values for the actual dataset. The intersection provides an estimate of the overfitting phenomenon. As expected, the initial Y variable demonstrates significantly higher Q2 and R2 values compared to the permuted counterparts. The results of the permutation test indicate that the model effectively accounts for 100% of the overall variance in the data array.

ROC curves were generated for the model, and it shows excellent performance with area under the curve (AUC) values for both classes equal to one, as shown in Figure 5. These findings robustly affirm that the observed group separations in the OPLS-DA score scatter plot were statistically significant, and the predictability was not a result of overfitting the data. Subsequently, the extracted OPLS-DA model presented in Figure 3 is considered a good database for future use and predictions of unknown cheese samples while investigating species’ origin.

In the study by Tarapoulouzi and Theocharis [6], only three types of cheese, namely Cheddar, Kefalotyri, and Halloumi, were studied by using the same spectroscopic techniques and chemometric method. The fact that three types of cheese were compared in that study and they were classified into three groups (i.e., the type of cheese) was the first finding and the initiator of the current study. Thus, in this study, an increase in the types of cheese samples took place as well as taking out the Halloumi cheese group to test Kefalotyri and Cheddar with various other cheese products not made in Cyprus. Therefore, the use of various cheese types gave discrimination of the samples based on the milk species’ origin. Subsequently, it can be concluded that the number of different types of samples in a model is very important. For instance, a few different types of samples in a study may discriminate based on the types of samples. However, if various types of samples are used without a balanced number in each category, the discrimination may occur in terms of another parameter of the set (here, the milk species origin), which has fewer variations and discrepancies (here, cow or goat and sheep).

Considering all the outcomes, it would be appropriate for future studies if all the new types of cheese in this study were enriched with up to 10 samples per category (like Cheddars) and tested with the same spectroscopic techniques. Moreover, it is still a great challenge for our group to discriminate goat and sheep samples, which always end up in the same group after chemometric treatment. It also seems necessary to use one type of cheese and avoid any cow-origin samples to focus on the milk species (goat and sheep).

4. Conclusions

In the present study, discrimination of cheese products regarding the milk species’ origin took place by using a combination of FTIR, ¹H-NMR, and chemometrics. The cheese samples had different geographical origins, nature, and different manufacturing processes. It is important that they were classified based on the parameter of the milk species origin, highlighting the importance of spectroscopic techniques regarding this parameter. This is the first study to include cheese products from Cyprus along with cheese samples from abroad and several different cheese types, and this is the initiation of other similar studies in the future. Both the FTIR and ¹H-NMR data were very important in this study, and thus, future studies will focus on these spectroscopic techniques. Chemometrics, and particularly discriminant analysis, was a significant tool for data treatment and the presentation of the results of this study. The samples of this study will be enriched for further testing with spectroscopic techniques and chemometrics. A future challenge will also be the discrimination of goat and sheep samples, since both categories fall into the same group in this study as well as our previous ones. Subsequently, such an outcome could be used as a tool against counterfeiting and adulteration.