FT-MIR Analysis of Water-Soluble Extracts during the Ripening of Sheep Milk Cheese with Different Phospholipid Content

The purpose of this work was to study the suitability of the water-soluble extracts (WSE) of semi-hard sheep milk cheese for analysis by diffuse reflectance Fourier transform mid-infrared spectroscopy (FT-MIR) and the development of classification models using discriminant analysis and based on cheese age or phospholipid content. WSE was extracted from three types of sheep milk cheeses (full-fat, reduced-fat and reduced-fat fortified with lyophilized sweet sheep buttermilk) at various stages of ripening from six to 168 days and lyophilized. The first model used 1854–1381 and 1192–760 cm−1 regions of the first-derivative spectra and successfully differentiated samples of different age, based on changes in the water-soluble products of ripening biochemical events. The second model used the phospholipid absorbance spectral regions (3012–2851, 1854–1611 and 1192–909 cm−1) to successfully discriminate cheeses of markedly different phospholipid content. Cheese WSE was found suitable for FT-MIR analysis. According to the results, a fast and simple method to monitor cheese ripening based on water-soluble substances has been developed. Additionally, the results indicated that a considerable amount of phospholipids migrates to the cheese WSE and that FT-MIR can be a useful tool for their assessment.


Introduction
Infrared spectroscopy is based on the measurement of infrared light absorption by a sample, corresponding to the vibrations of the chemical bonds in its molecules [1,2]. The main spectral regions used in the analysis of dairy products are the near-infrared (NIR, 14,000-4000 cm −1 ) and the mid-infrared (MIR, 4000-400 cm −1 ) regions [1,3]. MIR spectroscopy exhibits higher signal and structural sensitivity and higher repeatability compared to NIR analysis [4]. Fourier transform mid-infrared spectroscopy (FT-MIR) is a fast, non-complicated and non-destructive analytical procedure which is advantageous for the in-depth study of protein structure [2,5].
It has been used in many studies to determine composition-i.e., cheddar cheese [6], Swiss cheese [7]-or chemical parameters in several kinds of cheeses-i.e., Emmental cheese [8], soft cheeses, [9]-or short-chain free fatty acid content in Swiss cheese [10], or for the evaluation of flavour quality in cheddar [6,11] and Swiss cheese [12] or in the rapid estimation of microbial counts in Feta cheese [13]. Different manufacturing conditions have been related with distinct FT-MIR spectral reflectance in pecorino cheese [14], cream cheese [15] and cheese produced from high pressure homogenized or untreated goat milk [16]. The successful determination of the geographic origin has been reported for Emmental [17,18], gruyère and L'Etivaz cheeses [19]. Furthermore, FT-MIR could be a reliable tool for the determination of authenticity of cheeses-as reported for grated Parmigiano Dairy 2021, 2 532 WSN the effect of residual casein and the fat on the spectra is excluded. In addition, the present study aimed to investigate the effect of different phospholipid contents of cheese on the spectra. For this purpose, three types of cheese used in our previous study [42] were analyzed; that is, full-fat, reduced-fat and reduced-fat fortified with sweet sheep buttermilk powder.

Cheese Samples and Water-Soluble Extracts
Semi-hard full-fat (FF) and reduced-fat (RF) sheep milk cheeses manufactured at various stages of ripening from six to 168 days were used for water-soluble extraction and analyses. Cheese making and ripening conditions have been reported by Sakkas et al. [42]. The cheese codes were: "A" (FF; 22.37 ± 0.890 g fat/100 g cheese), "B" (RF; 12.52 ± 0.620 g fat/100 g cheese) and "C" (RF fortified at 1% of cheese milk with lyophilized sweet sheep buttermilk; 11.67 ± 0.481 g fat/100 g cheese). The first step of cheese watersoluble extract (WSE) preparation was the homogenization of 20 g cheese in 180 g distilled water at 9500 rpm for 2 min by means of Ultra Turrax (IKA-Werke GmbH & Co. KG, Staufen im Breisgau, Germany). Next, the cheese slurry was heated at 40 • C for 1 h and rehomogenized at 9500 rpm for two min. After immediate cooling, the slurry was centrifuged at 10,000× g, for 10 min, at 4 • C, and the supernatant was filtered through Whatman No 1 filter paper at a temperature lower than 6 • C. WSE filtrate was lyophilized at ≤−40 • C for 48-72 h using an ALPHA I-12 freeze-dryer (Martin Christ Gefriertrocknungsanlagen GmbH, Osterode am Harz, Germany), prior to FT-MIR analysis.

FT-MIR Spectroscopy
FT-MIR spectra of WSE were obtained using a Thermo Nicolet 6700 Fourier transform infrared spectrometer (Thermo Electron Corporation, Madison, WI, USA) with a DRIFTS micro sampling cup and a deuterated triglycine sulphate (DTGS) detector (Spectra-Tech Inc., Stamford, CT, USA). Three spectra were collected (from three sub-samples) and an average spectrum was assessed for each WSE sample with a resolution of 4 cm −1 and a total of 100 scans. Recording and further processing of spectra were carried out by means of the OMNIC software (OMNIC 8.2.0.387, Thermo Fisher Scientific Inc., Waltham, MA, USA). The Savitzky-Golay algorithm (5-point moving, second-degree polynomial) was used for spectra smoothing and the baseline of all spectra was corrected accordingly (second-degree polynomial, 20 iterations).

Discriminant Analysis
Initially the samples were placed according to the class to which they belong (actual class). Then, the software calculated an average spectrum for each class and classified the spectra in a class (calculated class) according to the Mahalanobis distance from each averaged spectrum of each class. The standard samples are used for the calibration and the validation samples for internal validation. The standard and validation samples are samples of all classes and their usage as "standard" or "validation" samples is determined by the software. Finally, the models were validated externally (external validation) by a set of external validation samples of all classes other than those used for the calibration and the internal validation. Ten principal components (PC) were used according to Eigen analysis and the Performance Index (PI) was calculated. A high PI value shows a robust model.

Discrimination Model of Ripening Stage
The first-derivative spectra of 1854-1381 cm −1 and 1192-760 cm −1 spectral regions were used for discriminant analysis of samples with TQ Analyst software (version 8.0.0.245, Thermo Fisher Scientific Inc.). The samples were classified in five actual classes, according to their age and coded as "6d" (6 days), "28d" (28 days), "56d" (56 days), "112d" (112 days) and "168d" (168 days). Forty-three samples (seven samples of class 6d, nine samples of class 28d, 13 samples of class 56d, nine samples of class 112d and five samples of class Dairy 2021, 2 533 168d) were used by the software as calibration standards and 15 samples (one sample of class 6d, five samples of class 28d, two samples of class 56d, six samples of class 112d and one sample of class 168d) were used as validation standards. Finally, the calibrated method was validated by an external validation set of six different samples of all classes.

Discrimination Model of Phospholipid Content
The spectral regions of 3012-2851 cm −1 , 1854-1611 cm −1 and 1192-909 cm −1 were used for discriminant analysis of the spectra of B and C cheese WSEs with TQ Analyst software. Cheese A was excluded due to its nearly two-fold higher fat content compared to that of cheeses B and C. This would probably result in a quite higher content of free short-and medium-chain fatty acids in the WSE of cheese A compared to that of B and C cheese WSEs. Similarly to phospholipids, fatty acids absorb at 3012-2851 and 1854-1611 cm −1 regions and they could complicate the development of the discrimination model by contributing to phospholipid absorbance. Instead, WSEs of cheeses B and C were used, which are both RF cheeses and they are expected to contain similar levels of short-and medium-size fatty acids [42]. The samples were classified in two actual classes, B and C. Twenty-six samples (12 samples of class B and 14 samples of class C) were used by the software as calibration standards and 10 samples (four samples of class B and six samples of class C) were used as validation standards. Finally, the calibrated method was validated by an external validation set of six different samples of both classes. According to Eigen analysis, 10 PC were used.

Spectral Regions Related to Cheese Ripening
The products of the main biochemical events of cheese ripening, e.g., peptides, free amino acids, organic acids, free short-and medium-chain fatty acids and their esters, and new substances resulting from their catabolism, e.g., alcohols, ketones, aldehydes, lactones, and amines, are mostly water-soluble; thus they are included in cheese WSE [11,26,27].
In the present study, cheese ripening assessment was based on spectral differences of samples in the regions that all the above substances absorb. Therefore, we focused on 1854-1381 and 1192-760 cm −1 spectral regions, which include the "fingerprint" region and the carbonyl group and amide absorption, among others. Figure 1 shows typical FT-MIR spectra of the five different classes of lyophilized cheese WSE in the 2000-700 cm −1 spectral region. Table 1 shows that all the peaks appeared in 1854-1381 and 1192-760 cm −1 regions and their assignments.  arginine (side chain) [43] 1605-1600 NH 3 + (asymmetric deformation) free amino acids [44] COO − (asymmetric stretching) free amino acids [44] 1549-1537 Amide II: N-H (bending) and C-N (stretching) peptides, proteins [35,45] COO − (asymmetric stretching) acidic amino acids (side chains) [33] NH 3 + (symmetric deformation) free amino acids [44] phenyl nucleus aromatic amino acids [

Spectral Regions Related to Phospholipids
Sakkas et al., 2021 [42] proved that buttermilk addition in the cheese milk significantly increased (p < 0.05) the phospholipid content of cheese. Phospholipid basic structure includes a glycerol, linked to two fatty acids and one phosphate group [39]. We focused on three spectral regions to discriminate samples with different phospholipid content, that is the 3012-2851 cm −1 where the acyl chain of lipids absorbs, the 1854-1611 cm −1, where the carbonyl group of lipids absorbs, and the 1192-909 cm −1, where the phosphate group absorbs. Typical FT-MIR full spectra of cheese B and C WSE are presented in Figure 2. Table 2 shows all the peaks which appeared in the three aforementioned spectral regions and could be attributed to the functional groups of phospholipids, among other substances. The peaks at 2978-2971 and 2883-2870 cm −1 have been assigned to the CH 3 asymmetric [50] and symmetric stretching [8,44], respectively, while the peak at 2936-2920 cm −1 has been attributed to the CH 2 asymmetric stretching [8,44]. The peak at 1720-1714 cm −1 has been associated with the stretching of the carbonyl group [51]. The absorption at 1160-1150 cm −1 can be assigned to the PO 2 − symmetric stretching [51] and the P=O stretching [44], while at1096-1093 cm −1 , can be attributed to the P-O-C asymmetric stretching [51] and the PO 2 − symmetric stretching [13,44]. The peaks at 1129-1126, 1053-1050 and 1014-998 cm −1 have been associated with the P-O-C asymmetric stretching [44,51], while the peak at 932-929 cm −1 can be assigned to the P-O stretching [44].

Discriminant Analysis
Discriminant analysis of the samples was carried out by means of TQ Analyst software, which is a commercial and thus easy-to-use software and enables the direct external validation of a method using a spectrum test-set. The differences between spectra of different classes were limited and not distinct. Therefore, the discrimination of different classes could not be based on individual peaks and multivariate analysis was performed. After the separation of samples in the actual classes, the software applied the Mahalanobis algorithm to calculate the distances and create new calculated classes. Then each sample was inserted in a calculated class.  PC Scores measure the accuracy of the representation of each standard in the method by each PC used for method calibration. The calculated score values represent the multidimensional distance of each standard estimated on a PC. Figure 4 shows the PC Scores for the three-dimensional plot of the five classes of samples.  PC Scores measure the accuracy of the representation of each standard in the method by each PC used for method calibration. The calculated score values represent the multidimensional distance of each standard estimated on a PC. Figure 4 shows the PC Scores for the three-dimensional plot of the five classes of samples.  PC Scores measure the accuracy of the representation of each standard in the method by each PC used for method calibration. The calculated score values represent the multidimensional distance of each standard estimated on a PC. Figure 4 shows the PC Scores for the three-dimensional plot of the five classes of samples.  PC Scores measure the accuracy of the representation of each standard in the method by each PC used for method calibration. The calculated score values represent the multidimensional distance of each standard estimated on a PC. Figure 4 shows the PC Scores for the three-dimensional plot of the five classes of samples.  PC Scores measure the accuracy of the representation of each standard in the method by each PC used for method calibration. The calculated score values represent the multidimensional distance of each standard estimated on a PC. Figure 4 shows the PC Scores for the three-dimensional plot of the five classes of samples.  PC Scores measure the accuracy of the representation of each standard in the method by each PC used for method calibration. The calculated score values represent the multidimensional distance of each standard estimated on a PC. Figure 4 shows the PC Scores for the three-dimensional plot of the five classes of samples.  PC Scores measure the accuracy of the representation of each standard in the method by each PC used for method calibration. The calculated score values represent the multidimensional distance of each standard estimated on a PC. Figure 4 shows the PC Scores for the three-dimensional plot of the five classes of samples. PC Scores measure the accuracy of the representation of each standard in the method by each PC used for method calibration. The calculated score values represent the multidimensional distance of each standard estimated on a PC. Figure 4 shows the PC Scores for the three-dimensional plot of the five classes of samples. PC Scores measure the accuracy of the representation of each standard in the method by each PC used for method calibration. The calculated score values represent the multidimensional distance of each standard estimated on a PC. Figure 4 shows the PC Scores for the three-dimensional plot of the five classes of samples.

Discrimination Model of Phospholipid Content
Twenty-four calibration and eight validation standards were classified correctly (92.3 and 80% respectively), the PI was 82.7% and five external validation standards were classified correctly (83.3%). The distances of cheeses B varied from 0.599 to 1.003 in calculated B class and the respective variation was 0.679 to 1.065 for cheeses C ( Figure 5).

Discrimination Model of Phospholipid Content
Twenty-four calibration and eight validation standards were classified correctly (92.3 and 80% respectively), the PI was 82.7% and five external validation standards were classified correctly (83.3%). The distances of cheeses B varied from 0.599 to 1.003 in calculated B class and the respective variation was 0.679 to 1.065 for cheeses C ( Figure 5).

Discrimination Model of Ripening/Storage Stage
Most compositional parameters and all proteolysis indices of the cheeses of the present study assessed by standard methods were significantly (p < 0.05) affected by cheese age [42]. These findings agree with the present FT-ΜIR analyses results and the successful discrimination of cheese WSE according to days of ripening. The discrimination of WSE was based on the absorbance regions of peptides and amino acids, among others. The nitrogenous components of cheese WSE mainly consist of residual whey proteins-which are not hydrolyzed during ripening-and of varying amounts of casein fragments-that is, large, medium, and small-size peptides and amino acids. Therefore, peptide bonds were expected to change with age in WSE, resulting in changes in the spectral regions where Amides I and II absorb. Similarly, the protein and peptide catabolism that takes place during cheese ageing would have increased the absorbance of COOand NH3 + bands, resulting in changes in their spectral regions. All the aforementioned absorbance regions were included in the spectral regions used for the discriminant analysis of WSE samples, in addition to the absorbance regions of the other water-soluble products of rip-  PC Scores measure the accuracy of the representation of each standard in the by each PC used for method calibration. The calculated score values represent th dimensional distance of each standard estimated on a PC. Figure 4 shows the PC for the three-dimensional plot of the five classes of samples.  PC Scores measure the accuracy of the representation of each standard in by each PC used for method calibration. The calculated score values represen dimensional distance of each standard estimated on a PC. Figure 4 shows th for the three-dimensional plot of the five classes of samples.  PC Scores measure the accuracy of the representation of each standar by each PC used for method calibration. The calculated score values repr dimensional distance of each standard estimated on a PC. Figure 4 show for the three-dimensional plot of the five classes of samples. presentation of each standard in the method calculated score values represent the multiated on a PC. Figure 4 shows the PC Scores ses of samples.

Discrimination Model of Ripening/Storage Stage
Most compositional parameters and all proteolysis indices of the cheeses of the present study assessed by standard methods were significantly (p < 0.05) affected by cheese age [42]. These findings agree with the present FT-MIR analyses results and the successful discrimination of cheese WSE according to days of ripening. The discrimination of WSE was based on the absorbance regions of peptides and amino acids, among others. The nitrogenous components of cheese WSE mainly consist of residual whey proteins-which are not hydrolyzed during ripening-and of varying amounts of casein fragments-that is, large, medium, and small-size peptides and amino acids. Therefore, peptide bonds were expected to change with age in WSE, resulting in changes in the spectral regions where Amides I and II absorb. Similarly, the protein and peptide catabolism that takes place during cheese ageing would have increased the absorbance of COO − and NH 3 + bands, resulting in changes in their spectral regions. All the aforementioned absorbance regions were included in the spectral regions used for the discriminant analysis of WSE samples, in addition to the absorbance regions of the other water-soluble products of ripening biochemical events. Consequently, the water-soluble extraction of cheeses is suitable for FT-MIR analysis and the developed statistical model is a reliable method for the discrimination of cheeses at different stages of ripening.
FT-MIR analysis of cheese water-soluble fractions has been reported by a few research groups. Subramanian et al., [6,11] discriminated cheddar cheese according to flavour quality and Kocauglu-Vurma et al., [12] differentiated Swiss cheeses by different manufacturers and production regions and developed predictive models for sensory evaluation. Koca et al., [10] analyzed Swiss cheese acidified water extract to determine short-chain free fatty acids and compared the results with that of direct cheese analysis. The measurement of cheese extract resulted in higher correlation with gas chromatography data and better calibration models compared to direct cheese measurement, in which the non-removal of high fat and protein contents caused interferences. Moreover, the authors suggested that an extraction procedure resolves the problem of cheese heterogeneity in sampling. Martín-del-Campo et al., [35] analyzed the defatted suspensions and their acid soluble nitrogen extracts of Camembert-type cheeses ripened up to 27 days. They discriminated samples of different age and determined the days of ripening with one day error; however, they reported no improvement of the analysis results using cheese fractions. Subramanian et al. [27] and Chen et al., [37] performed FT-MIR analysis of water/chloroform/ethanol extracts of cheddar and Swiss cheeses, respectively, and developed models for cheese classification according to their age, based on differences in the 1800-900 cm −1 spectral region. The former research group also developed prediction models of cheese age and of the levels of amino acids and organic acids.
They reported good correlation with cheese age and gas chromatography and HPLC data and concluded that cheese extraction improved the signal from water-soluble compounds.

Discrimination Model of Phospholipid Content
The present study aimed to establish the spectral regions of phospholipid absorbance in cheese extract and to develop a discriminating method based on samples with different phospholipid content. As mentioned before, phospholipids are amphipolar molecules, which emerges the question whether they totally or merely migrated to water phase during cheese water extraction. Moreover, cheese extract also contains phosphopeptides [52] which absorb light at the phosphate group region (1192-909 cm −1 ) and free short-and medium-chain fatty acids which absorb light at the acyl chain (3012-2851 cm −1 ) and carbonyl group (1854-1611 cm −1 ) regions. The absorption of those substances at the same regions with phospholipids means that they may contribute to the signal produced by phospholipid absorption, causing difficulties in the development of the discrimination model. In order to address the aforementioned difficulties, samples of markedly different phospholipid content, namely cheeses B and C, were used for the development of the discrimination model. Sakkas et al. [42] have evaluated the phospholipid content in cheeses B and C by a direct method and they reported that it was significantly higher in fat, cheese and dry matter of cheese C compared to that of cheese B. In particular, the phospholipid contents of cheeses B and C at 28d were 378 ± 4 and 569 ± 11 mg/100 g fat, 43 ± 1 and 67 ± 1 mg/100 g cheese and 94 ± 1 and 150 ± 3 mg/100 g dry matter, respectively. Regarding the age of 112 d, the respective phospholipid contents of cheeses B and C were 401 ± 5 and 579 ± 6 mg/100 g fat, 52 ± 1 and 69 ± 1 mg/100 g cheese and 107 ± 2 and 151 ± 1 mg/100 g dry matter. This difference was due to the addition of sweet buttermilk in cheese milk C since buttermilk is considered a rich source of phospholipids due to its high content in milk fat globule membrane material [39,42,53].
Our results showed that the discrimination model based on spectral regions related to phospholipids is robust and can differentiate WSE from cheeses with different phospholipid content. It resulted in a 92.3% correct classification of calibration standards, 80% correct classification of validation standards, 82.7% PI and 83.3% correct classification of external validation standards. Therefore, the discrimination was due to differences in phospholipid absorption among samples, indicating that a significant portion of phospholipids shifts to the aqueous part during cheese water extraction.
Moreover, the suitability of cheese WSE for FT-MIR assessment of cheese phospholipid content is demonstrated similarly to the discrimination model developed for the ripening stage. Further research is needed to investigate the application of FT-MIR to phospholipid analysis, aiming to replace the current time-and cost-consuming analytical techniques and to assess the effectiveness of the present method in the differentiation of cheeses with more variable phospholipid contents.

Conclusions
Two FT-MIR robust models, double validated, were developed in this study for the discrimination of semi-hard sheep milk cheeses according to their ripening time or phospholipid content, by analyzing a simple lyophilized cheese water-extract (WSE). Double validation (internal; external) increases the robustness of the model and therefore its validity. The results of both models showed that cheese WSE is suitable for FT-MIR cheese analysis because: (i1) it is simple, (i2) it overcomes the interferences of fat and protein and the sampling difficulties resulting from cheese heterogeneity, (i3) it does not use organic solvents, and (i4) analysis is focused on water-soluble substances. The novelty of the present work is the use of FT-MIR spectra of the water-soluble extract for the assessment (1) of the ripening of a sheep milk cheese, and (2) of the phospholipid content of cheese. The rapid, plain and low-cost model developed in this work for the differentiation of cheeses by their ripening time could be very useful for scientific research and the cheese industry. It could help to replace the time consuming and expensive analytical methods used currently for the monitoring of ripening, which is an important element of the quality control of cheese. By the second developed model, cheeses with different phospholipid content were discriminated and it was shown that a considerable part of phospholipids migrates to the aqueous phase during cheese extraction. Further research for the development of a quantitative or measurement method based on FT-MIR analysis of cheese WSE would be of great interest considering the importance of dairy phospholipids from a technological and dietary point of view. The same holds true for the investigation of the implementation of FT-MIR in phospholipid analysis, which could replace the current time intensive and costly analytical techniques. The major strength of the present study is the robustness of the developed models. In this respect, their application in the analysis of various cheese samples is a future research perspective.

Data Availability Statement:
The datasets generated for this study are available on request to the corresponding author.