High-Resolution Mass Spectrometry for Detailed Lipid Profile and Chemometric Discrimination of X-Ray Irradiated Mozzarella Cheese

TL;DR

This study uses high-resolution mass spectrometry to analyze lipid changes in mozzarella cheese after X-ray irradiation and identifies potential markers for detecting irradiation.

Contribution

The study introduces a novel analytical workflow combining lipidomics and chemometric modeling to detect irradiation markers in cheese.

Findings

X-ray irradiation did not form new lipid molecules in mozzarella cheese.

Chemometric modeling identified specific lipids as potential markers for irradiation.

An artificial neural network was developed to model omics data patterns related to irradiation.

Abstract

Ionizing radiation is a non-thermal sanitization technique used in the food field to eliminate bacteria, molds, insects and other microbes, resulting in delayed spoilage and extended shelf life. In this work, mozzarella cheese was irradiated with X-rays at a dose of 3.0 kGy, and irradiation-induced lipid modifications were evaluated through a comprehensive analysis of the mozzarella lipid fingerprint. To this aim, an optimized microwave-assisted extraction method associated with UHPLC-Q-Orbitrap-MS analysis was used for reliable and accurate lipid identification in the controls and in irradiated samples. The outcomes demonstrated that the X-ray dose employed in this investigation did not cause the formation of new lipid molecules. However, lipidomic chemometric modeling, including partial least squares-discriminant analysis, enabled the discrimination of irradiated versus non-irradiated samples and the selection of five ceramides, eight hexosyl ceramides, four sphingomyelins, one phosphatidylethanolamine, one cholesterol ester, ten oxidized triacylglycerols, and one oxidized diacylglycerol as potential markers of treatment. Finally, an artificial neural network was developed to accurately model the entire pattern in omics data in relation to the treatment. This developed analytical workflow allows for expanding knowledge on the effects of this technology and could have interesting applications in food safety traceability and control plans.

Figures

Full text

1. Introduction

Mozzarella is a “pasta filata” cheese, manufactured by stretching the curd from pasteurized cow or buffalo milk in hot water to get a particular texture [1]. This fresh soft cheese is the most exported and consumed dairy product worldwide, and it is produced in several countries, e.g., the U.S., Brazil, Argentina, Australia, Germany, and Italy. Italy is one of the most important producers with 152.558 tons exported worldwide in 2024 [2,3]. Traditional Italian mozzarella is a high-moisture food (about 60% water content) with about 20% fat and 19% protein [4] characterized by a fresh, milky flavor and a soft consistency and a shelf life ranging from a few days to twenty days, depending on the moisture level, manufacturing procedures and storage conditions [5]. Mozzarella cheese, due to its physicochemical properties (e.g., high moisture and pH close to neutrality) may represent a suitable substrate for pathogen growth such as Listeria monocytogenes, Staphylococcus aureus, Escherichia coli and Salmonella spp., especially if the proper temperature conditions are not respected during storage or the supply chain [6]. Furthermore, this type of cheese is usually stored in trays or bags with a covering liquid which is a brine with variable composition (i.e., water, lactic or citric acid, NaCl, CaCl_2_). The covering liquid allows for retention of water content in the cheese, while maintaining a very soft consistency and preserving the particular freshness and flavor of this cheese, but at the same time, it promotes its perishability [7].

The conventional processing methods for cheese preservation are mainly based on the use of heat [5]; in fact, the presence of the pathogens can be effectively addressed by the high temperatures reached during milk pasteurization or the curd spinning step, when temperatures may reach 90–95 °C. In the heat-treatment processes, lipid variations can result both from chemical changes, such as oxidation of milk lipids with more intense heating [8], and from physical phenomena, for example, in stretched mozzarella, the reduction of fat content with temperature/degree of stretching [9]. Nevertheless, thermal pasteurization is not able to completely eliminate the harmful bacterial population and, in addition, post-pasteurization contamination may occur during or after the remaining cheese-making process [10,11]. Refrigeration (cold storage) or freezing (less than −18 °C) treatments are also used to improve shelf life, but Alinovi et al. [2] reported microstructural damage and different physical, textural, and sensory properties in frozen cheeses with respect to fresh samples. Furthermore, all cold and hot thermal processes are expensive techniques, involving high energy and water consumption [12]. In this context, irradiation treatment is an alternative non-thermal preservation technology which involves the use γ-rays, e-beams or X-rays to inhibit or eliminate microbial proliferation in order to enhance food safety and extend the shelf-life of perishable products, like soft cheeses [13]. To date, in accordance with the European regulatory framework [14,15], the EU authorizes irradiation only for dried aromatic herbs, spices and vegetable seasonings. Other foods may be irradiated only where permitted under specific Member State authorizations. In this context, raw-milk Camembert is authorized in some Member States up to 2.5 kGy, and in France and the Czech Republic, casein and caseinates are authorized up to a maximum dose of 3.0 kGy for microbial control. [16]. Currently, it is known and reported by our previous work [17] that specific radio-induced markers, such as, in the case of foods containing fats, 2-alkylcyclobutanones, are generated from radiolysis of triglycerides through irradiation. Nowadays, in Europe, screening and confirmatory validated or standardized protocols are available to identify whether or not food products have been treated with ionizing radiation and some procedures are chemical methods, based on determination of these molecules and free radicals. In this context, continuous scientific research is required to expand knowledge in this field to support official control plans under EU legislation. Regarding dairy products, Lacivita et al. [18] demonstrated that X-ray treatment up to 3.0 kGy can effectively reduce pathogens without significantly affecting the taste, texture, or nutritional value of the mozzarella cheese. On the other hand, the occurrence of lipid oxidation as a result of the production of free radicals during irradiation is well known [19,20]; so, despite these known studies, the lipidomic mechanisms underlying irradiation on mozzarella cheese require continuous in-depth analysis. Indeed, lipids are important constituents in mozzarella cheese and, from a chemical point of view, they are a heterogeneous pool of compounds, with fatty acyl/alkyl, sphingosine, or isoprene moieties as their hydrophobic building blocks [21]. Because of their heterogeneity, they play fundamental roles in cellular processes and in a series of physiological activities [22]. In particular, the quality and composition of cheese lipids are key determinants of their nutritional value and their overall impact on human health within a balanced diet.

Lipidomics, a specialized branch of metabolomics, focuses on the comprehensive analysis of lipids, their interactions, and their biological roles. In particular, untargeted lipidomics is an extensive approach to the study of all detectable lipids in biological systems, useful for many purposes, such as, in the food field, the assessment of the authenticity and adulteration of foods or for the evaluation of the effects related to the use of technological processes, such as irradiation [23]. This approach, thanks to chemometric tools, can also allow for the identification of new biomarkers and unknown but relevant lipids for in-depth knowledge of food irradiation treatment and for the elucidation of metabolic pathways involved [23].

The analytical technology that is generally used in lipidomics is mass spectrometry (MS), often associated with chromatographic separation, due to its high sensitivity and specificity and its ability to provide both structural and quantitative information on the analytes [24]. MS-based lipidomic workflows include several different crucial steps, such as sample preparation, lipid extraction, chromatography, MS analysis, lipid identification by dedicated tools and chemometric elaboration [21]. In the extraction step, different methods were optimized for a single lipid class or for comprehensive lipidomics [25]. Recently, automated techniques, such as accelerated solvent extraction (ASE) [24] and microwave-assisted extraction (MAE) [26], were proposed as alternatives to classic procedures because these techniques are characterized by low usage of solvents along with the ability to automate the extraction process, reducing process times and sample preparation costs while ensuring relevant extraction efficiency [13,14]. More specifically, in the MAE procedure, the solvent in contact with the sample is heated by microwave energy and the analytes are extracted on the basis of their physical–chemical properties [26]. In our previous investigation, it was demonstrated that an optimized MAE method can be used as an efficient alternative to the classic methods for lipidomics, using eco-friendly solvents with consequent increases in the green aspect of this technique [26].

Due to their high-dimensional nature, omics data require the use of appropriate statistical elaboration that includes both univariate and multivariate tests. Among multivariate tests, unsupervised analysis by principal component analysis (PCA) allows for the verification of the presence of both possible first aggregations and outliers, while supervised approaches, such as partial least squares discriminant analysis (PLS-DA) provides a classification model very useful in metabolomic studies [11,23]. In this context, artificial neural networks (ANNs) hold great promise for modeling complex data and they can be applied in every aspect of food science, especially for food safety and quality analyses, such as in omics data elaboration [27]. Moreover, this approach can achieve high classification accuracy and prediction capability, particularly when dealing with complex non-linear data [28].

In this work, the lipid fingerprint of the commercial mozzarella cheese and the evaluation of the effects of X-ray irradiation on its lipid composition were carried out. A dose of 3.0 kGy was chosen because this dose was sufficient for the purposes of the treatment [18] and also because this level is recommended to prevent rancidity and off-flavor resulting from high doses [16,20]. For this aim, an optimized MAE method [26] was used, combined with chromatography coupled to orbitrap mass spectrometry (UHPLC-Q-Orbitrap-MS) analysis and subsequent chemometric data processing. The latter involved supervised and unsupervised methods and subsequent validation of discriminant and classification models. Finally, an enrichment analysis was carried out to deepen knowledge of the metabolic pathways involving the most significant lipids in distinguishing irradiated mozzarella from non-irradiated mozzarella.

2. Results and Discussion

2.1. Untargeted Lipid Profiles of Mozzarella Cheese

Lipid profiles of the mozzarella cheese not irradiated (NI), used as controls, and X-ray irradiated (IRR), were obtained using the untargeted approach. Lipid information was acquired by analyzing 60 MAE extracts in duplicate, in both positive and negative ionization mode, using data-dependent acquisition, for a total of 120 UHPLC-Q-Orbitrap-MS analyses. This large sample set improved the robustness of the data processing and supported the identification of lipids.

A total of 478 compounds in positive and 56 lipids in negative mode were detected, assigned to 18 subclasses (Figure 1). Specifically, in positive ion mode, in addition to cholesterol ester (ChE) as an +H−H_2_O adduct, 380 triacylglycerols (TGs) and 19 related oxidized forms (OXTG and OX2TG) as +NH_4_ or +Na and +NH_4_ or +H adducts, respectively, 42 diacylglycerols (DGs) as +Na and +NH_4_ adducts and one related oxidized form (OxDG) as an +NH_4_ adduct were identified. Moreover, one ceramide (Cer), four sphingomyelins (SMs), eight phosphatidylcholines (PCs), eight lysophosphatidylcholines (LPCs), eight phosphatidylethanolamines (PEs), three lysophosphatidylethanolamines (LPEs) and three lysophosphatidylserines (LPSs) were all measured as +H adducts. In negative ion mode, nine Cer, eight hexosyl ceramides (four Hex1Cer and four Hex2Cer) and nine SM as +HCOO adducts together with six PCs, seven PEs, three phosphoserines (PS) and four phosphoinositols (PIs), three LPCs, four LPEs, two LPSs and one lysophosphatidylinositol (LPI), as −H adducts, were identified. Detailed information on the individual lipids identified is provided in the Supplementary Materials (Table S1) accompanying this manuscript.

2.2. Characterization of Lipids Detected in Mozzarella Cheese

From the nutritional point of view, both irradiated samples and controls exhibited a high number of acylglycerols, which primarily serve as storage lipids and, as expected, many of them contained oleic acid (C18:1). This acid is considered beneficial for health, as an association has been demonstrated between high monounsaturated fatty acid intake and reduced both plasma cholesterol and LDL-cholesterol levels. Moreover, oleic acid is highly stable to oxidation with respect to polyunsaturated fatty acids (PUFAs) [29]. PUFAs were carried in many TGs and DGs; specifically, acylglycerols with C18:2 (linoleic acid ω-6), C18:3 (α-linolenic acid ω-3), 20:3 (eicosatrienoic acid), C20:4 (arachidonic acid ω-6) and 20:5 (eicosapentaenoic acid EPA ω-3) were detected. These molecules are bioactive components with biological benefits for human health: ω-3 PUFAs are known to prevent cardiovascular disease and improve immune function [29,30] and, in particular, EPA can partially inhibit the conversion of ω-6 fatty acids into harmful eicosanoids, thereby reducing cardiovascular risk and inhibiting tumorigenesis [30]. At the same time, ω-6 fatty acids are essential components of cell membranes and play a physiological role in immune regulation and inflammatory responses [31].

The polar lipid fraction was characterized by both phospholipids and sphingolipids. The first are a group of biomolecules that recently, have attracted great attention for their composition, stability, and potential health benefits, such as their ability to suppress Alzheimer’s disease and protect against hypercholesterolemia and cancer [32,33]. Moreover, some glycerophospholipids containing bioactive PUFAs were also measured, such as PC (16:0_18:2), PC (36:3), PE (16:0_18:2); PE (18:0_18:2), PE (18:1_18:2), PI (18:0_18:2) and PS (18:0_18:2). They carry out a lot of functions as binding cations and emulsion stabilizers and they also influence enzymes on the globule surface, cellular activities and transmembrane signaling [29].

Furthermore, other lipids, such as sphingolipids, as CER and HexCer, with a role as direct metabolic mediators [34], were identified. Along with SM, cholesterol and phospholipids, these are components of the milk fat globule membrane (MFGM) and their composition in the final products and by-products is affected by any treatment, such as homogenization, centrifugation or thermal treatments [30]. In cheese, milk fat can exist as fat globules, also defined pools, mainly entrapped within the protein network. These structures can be individual or aggregated, spherical but larger than typical milk fat globules, elongated (especially in pasta-filata cheese types, as mozzarella cheese) or even nonglobular forms [19]. Moreover, cheese manufacture-induced modifications of the fat pools are likely to influence digestion behavior [22].

Furthermore, the fingerprint showed some lysophospholipids, such as LPCs (LPC (16:0), LPC (18:1) and LPC (18:2)), LPEs (LPE (16:0), LPE (18:1), LPE (18:2)), LPI (18:0) and LPSs (LPS (18:0), LPS (18:1) and LPS (18:2)). These molecules are an essential component of lipoproteins with important roles as technologic and/or oxidative markers but also in the regulation of lipid transport and they are important for the physiological functions of the organism for the absorption of nutrients [35].

Regarding oxide lipids linked to oxidative phenomena, thanks to the dedicated “Oxid. GPL” database integrated into the LipidSearch™ v4.2.2.7 software, it was possible to evaluate oxidative modifications in phospholipids, triacylglycerols, diacylglycerols and fatty acids. Oxidized species are annotated with specific labels: “+O” indicating hydroxyl addition; “+OX” referring to epoxide formation and “CHO”, “COOH” or “COOCH_3_” marking terminal aldehyde, carboxylic acid or methyl ester groups, respectively. A total of 19 OxTG species, annotated as “+O”, “+2O”, “+OO”, “+CHO” were identified, and only TG(18:1+O_18:1_18:1) was also detected in its non-oxidized form. All OxTG species were classified as grade “A” or “B”. One OxDG (DG(18:2+OO_9:0)) was also detected. It should be noted that oxidative lipidomics is still an emerging field with no consolidated guidelines for annotation; for this reason, only the most reliable identifications, supported by manual MS/MS inspection, were retained.

Figure 1 shows the qualitative lipid fingerprint (in terms of both number and type) of the control, which remained unmodified after irradiation, indicating that no variations occurred in the lipid components between IRR and NI samples, confirming that the nutritional lipid profile remained unchanged up to the dose level studied. However, differences in the relative abundance of lipids were observed, and subsequently considered for chemometric analysis.

2.3. Chemometrics

2.3.1. Univariate Analysis

Univariate analysis, using a volcano plot, was conducted for exploratory purposes and performed to assess the differences between NI and IRR (Figure 2). This plot shows the presence of numerous lipids clustered in a cloud below the threshold value (−log_10_ p-value < 1.3) suggesting that these molecules did not exhibit significant differences in the ANOVA test (p-value > 0.05) between the two conditions, with mean value deviations close to zero. On the other hand, some lipids located above the volcano plot threshold, with p-value < 0.05 in the ANOVA test, belonging to the classes of sphingolipids, glycerophospholipids, ChE, and oxidized TG and DG, revealed significant differences/modifications between the irradiated and control samples (Figure 2). However, this approach did not allow for the capture of all of the information that can be extracted from the entire lipid fingerprint; therefore, in this study, the processing was further explored using multivariate analysis.

2.3.2. Multivariate Analysis

Multivariate analysis was also performed to reveal patterns, clusters, and discriminant molecules, as well as to obtain discriminant models. In particular, an unsupervised PCA was first applied to both the negative and positive datasets (Figure 3) to assess the presence of initial sample aggregations and potential outliers. However, the resulting PCA did not show clear groupings between treated and untreated samples, suggesting that this unsupervised approach was not suitable for discriminating between the two conditions.

Subsequently, PLS-DA supervised models were built to discriminate among the two conditions. By applying the normalized orthogonal distance and normalized Hotelling’s T^2^ distance with a significance level of 0.01, only one outlier out of 120 was identified in the positive dataset, as it was beyond the critical limit (Figure S1). Consequently, this sample was excluded from the discriminant elaborations.

To reliably assess the predictive performance of these models, a double cross-validation algorithm was applied, iterated 200 times, ensuring the stability and robustness of the estimated parameters [20,23]. The models obtained showed good performance in terms of Q^2^, DQ^2^, accuracy, sensitivity, specificity, Area Under the Receiver Operating Characteristic Curve (AUROC), and root mean squared error of cross-validation (RMSECV) (Table 1), calculated as reported in our previous work [23]. Moreover, the selection of the optimal number of latent variables (#LV) was performed within the cross-validation procedure. This procedure was previously described by Tomaiuolo et al. [23]. Briefly, the entire dataset was divided into m segments. At each iteration, m–1 segments (training set) were used for model optimization, while the remaining segment (test set) was used to calculate diagnostic parameters. The training set was subjected to an internal cross-validation procedure, during which the number of latent variables was optimized by maximizing the Efficiency Index (I_eff_), calculated according to the method described in our previous study [23]. In the present work, this procedure was repeated 20 times with m = 8, resulting in a total of 160 estimated #LV values. The mode of the 160 calculated #LV values was used for the final model. The optimal number of latent variables (#LV) was six for positive and five for negative acquisition mode. This approach enabled us to determine the values of R^2^, obtained from the entire dataset, which were 0.810 in the positive mode and 0.731 in the negative mode, as well as the respective Q^2^ values derived from the double cross-validation (Table 1). The R^2^ and Q^2^ within each model were not markedly different from one another, indicating that the models were not overfitted.

Sensitivity, specificity, and accuracy for both positive and negative models were all above 0.9. Taking IRR as the “positive” class for classification, the corresponding parameter ranges, calculated using the Wilson score method [36] with a 0.95 confidence level, were as follows: in the positive model, sensitivity ranged from 0.914 to 0.924, specificity from 0.981 to 0.985, and accuracy from 0.949 to 0.954; in the negative model, sensitivity ranged from 0.982 to 0.986, specificity from 0.915 to 0.925, and accuracy from 0.949 to 0.955. These values confirm the very good performance of both models.

To estimate model stability and variability, the bootstrap algorithm was used. This algorithm was applied to the entire dataset using non-stratified sampling, and the model was built using the optimal number of #LV previously determined during the double cross-validation. According to this approach, with n objects, an n-dimensional set is obtained by randomly sampling n objects with replacement from the original set. As a consequence, some objects may appear multiple times in the bootstrap sample whereas others may not appear. The latter constitute the out-of-bag test set while the set of objects that are selected only once constitutes the test set on which the model is built. This process was repeated 10,000 times, and diagnostic statistics were calculated by summarizing the results obtained from all iterations. The good values obtained for sensitivity, specificity and accuracy (Table 2) were also confirmed by the values of their range. Similarly to previous elaboration, considering IRR as the “positive” class for classification, the parameter ranges calculated at a 0.95 confidence level were as follows: in the positive model, sensitivity ranged from 0.915 to 0.917, specificity from 0.986 to 0.987, and accuracy from 0.951 to 0.953, while in the negative model, sensitivity ranged from 0.983 to 0.984, specificity from 0.907 to 0.909, and accuracy from 0.945 to 0.947.

In addition, a permutation test, within the PLS-DA, was carried out to confirm that the validation results of the classification models were not achieved by chance. The average number of misclassified samples during validation was chosen according to Szymańska et al. [37]. In this test, the p-value was calculated considering the number of permuted models that generated a number of misclassified samples less than or equal to the average number of those in the validation phase [23] (Figure 4A,B). For both positive and negative datasets, a p-value close to zero was obtained, confirming that the discriminating ability of the models was not determined by random phenomena. This result was confirmed by the mean Q^2^ values obtained from the permuted models in both positive and negative modes, which were −0.135 and −0.129, respectively. Figure 4C,D show the relationship between the actual R^2^ and the permuted R^2^ in blue, and the relationship between the actual Q^2^ and the permuted Q^2^ in red. Therefore, the outcomes confirmed that the obtained PLS-DA models were reliable and well suited to assess the differences between the two sample groups.

PLS analysis enables the selection of relevant variables (lipids) based on VIP (Variable Importance in Projection) scores. For this investigation, the VIP score limits were set at 1.4 and 1.2 for positive and negative acquisitions, respectively. The choice of these thresholds was driven by the aim of obtaining a model with good performance and, at the same time, minimizing the number of influent molecules used in the model. In fact, diagnostic parameters were estimated at increasing threshold levels until a significant change in model performance was observed. The models were tested as described above (double cross-validation, bootstrap, and permutation test) using only the selected variables. Table 3 lists the positive and negative VIP along with their corresponding VIP scores and log2 fold change (log2FC) values, while Table S2 shows the details of their fragmentation patterns.

The VIP calculated in positive mode and the relative value of log_2_ FC (Table 3) highlighted the increase in the oxidative phenomena in irradiated samples with the consequent formation of oxidized triacylglycerols (Figure 5).

As known, X-ray irradiation can promote the formation of highly reactive radicals through water radiolysis [23]. These species lead to the formation of alkoxyl and peroxyl radicals, prototypical intermediates, which subsequently degrade into secondary oxidation products, such as alcohols, ketones, epoxides, aldehydes and hydrocarbons, compounds responsible for the sensory defects, including off-odors. Regarding this, in our previous work, we investigated X-ray irradiation-induced sensorial and volatolomic changes in mozzarella cheese by HS-SPME/GC-MS associated with chemometrics. The results highlighted the increase in several volatile organic compounds (hydrocarbons, alcohols, aldehydes, ketones, and disulfides) in irradiated samples, suggesting the formation of radiolytic products likely associated with oxidation of lipids and proteins [20].

In this work, the detection of oxidized lipids provides important information to understand the impact of X-ray irradiation treatment. In particular, given that OxTG, as well as the OxDG, were detected both in non-irradiated and irradiated samples, it is reasonable to assume that the X-ray dose used in this study was not sufficient to generate new oxidized lipids and that previous processing steps, such as pasteurization or enzymatic activities, may have contributed to their formation. Nonetheless, quantitative differences between irradiated and non-irradiated samples were observed, and, as previously reported, several of these oxidized species were included in the list of VIP-selected variables, useful for potential markers of X-ray treatment. Among the VIP scores calculated in positive mode, the increase in cholesterol ester could hypothetically suggest potential involvement of the enzymic activity of cholesterol esterase during digestion, because the main function of this lipase is to hydrolyze dietary cholesterol esters into absorbable free (unesterified) cholesterol and free fatty acids.

On the other hand, the decrease in almost the entirety of VIP scores calculated in negative mode (Figure 6), particularly for sphingolipids, suggests that radiation-induced modifications probably affect the fat globules/pools, present in the casein network of the curd [38]. Furthermore, bioactive lipids (phospholipids and sphingolipids) included in fat globules are involved in biological functionality that could also be influenced by X-rays.

2.4. Lipid Enrichment Analysis

The above considerations were further detailed by a functional enrichment analysis of lipids more interesting for discriminating the treated samples from the controls. In this investigation, a LION enrichment analysis was performed using a target list-mode approach in which variables with high VIP scores were compared against the list of all identified lipids.

The first step was to convert LION names in clear and unambiguous LION ID. LION links each lipid to its chemical group, backbone, physical properties, and likely assigned cellular part [39]. Concerning oxidized di- and triacyclglycerols, no match was found with LION nomenclature; therefore, these lipids were not considered. Subsequently, a LION enrichment graph (Figure 7) was generated and this revealed the significantly over-represented categories (FDR < 0.05).

As shown in Figure 7, the enrichment was mainly driven by sphingolipids (SPs) with a d18:1 backbone, in particular SM and Cer. These lipid species are commonly associated with the plasma membrane. With regards to lipid classification, the figure suggests that these molecules are part of sphingolipids, in particular, phosphosphingolipids (including SM), Cer and neutral glycosphingolipids (Hex1Cer and Hex2Cer). They are characterized by a headgroup with positive or neutral charge, long fatty acids chain and negative intrinsic curvature. Moreover, they perform the function of lipid-mediated signaling and membrane component and are part of the cellular component, specifically, the plasma membrane and endosome/lysosome. This information suggests that X-ray irradiation primarily affected the lipids involved in cell membrane organization and signaling function [40].

2.5. Artificial Neural Network (ANN) Model

Along with PLS-DA, we used an artificial neural network (ANN), as an alternative supervised chemometric tool. The ANN required training to allow the optimization of model parameters and predictive capabilities were verified through a validation process, which consisted of randomly splitting the initial dataset into a training dataset, in this study, 75%, and a validation dataset, 25%. The validation set was used to identify the best network on the basis of the network’s error performance and to stop training if overfitting occurred.

The number of neurons in the hidden layer is a critical point since too many hidden neurons may lead to low training error, but poor generalization performance due to overfitting [28]. Therefore, in this elaboration, the ANN model was tested with a number of neurons in the hidden layer, from two to ten, to select the most accurate classification. Based on the calculated accuracy and specificity values, three neurons were empirically determined as the optimal value. The validation process was repeated 1000 times and the model produced excellent performances (Table 4) able to detect the lipid fingerprint modified by X-rays, with a sensitivity, specificity and efficiency close to one. Also, for ANN, taking into account IRR as our “positive” classification, the corresponding ranges for the parameters, with a confidence level of 0.95, were 0.981–0.985 for sensitivity, 0.998–0.999 for specificity and 0.989–0.992 for accuracy in the positive model, while in the negative model, these values were 0.973–0.978 for sensitivity, 0.988–0.992 for specificity and 0.981–0.984 for accuracy.

The comparison of the two supervised analyses, ANN and PLS-DA, showed that ANN models, based on the entire lipid fingerprint, achieved better performance. However, PLS-DA enabled the identification of specific molecules acting as potential treatment markers, offering insights into the metabolic effects of irradiation.

3. Materials and Methods

3.1. Chemicals and Reagents

Sodium sulfate (Na_2_SO_4_) of analytical grade, ammonium formate (NH_4_HCO_2_), isopropanol (IPA), water (H_2_O), acetonitrile (ACN), and formic acid (HCO_2_H) of LC/MS grade, were provided by Carlo Erba Reagents (Cornaredo, Milan, Italy). Ethyl acetate of analytical grade was purchased from Fluka (Buchs, Switzerland). Ethanol, and methanol (MeOH) of LC/MS grade were provided by Merck Life Science S.r.l. (Darmstadt, Germany) and EMD Chemicals (Gibbstown, NJ, USA). Pierce LTQ Velos ESI Positive and Negative Ion Calibration Solutions were provided by Thermo Fisher Scientific (Waltham, MA, USA). The 1,2,3-tripelargonoyl-glycerol (trinonanoin, 9:0-9:0-9:0-TAG) and the deuterated lipid internal standards, Equisplash™ Lipidomix^®^ 100 mg L^−1^, were purchased from Merck Life Science S.r.l. (Darmstadt, Germany). Stock standard solution and working standard solution of trinonanoin, 10,000 mg L^−1^ in CHCl_3_/ MeOH (1:1, v/v) and 1000 mg L^−1^ in MeOH/CHCl_3_ (4:1, v/v), respectively, were prepared and stored at 4 ^°^C (±2 °C).

3.2. Irradiation Treatment

X-ray irradiation treatment was performed at the National Reference Laboratory for treatment of irradiated foods and their ingredients (Istituto Zooprofilattico Sperimentale della Puglia e della Basilicata, Foggia, Italy). Mozzarella cheese samples of 100 g made from cow milk (65% of water, 16.5% of protein, 16% of total lipids, 1.0% carbohydrates and 0.7% of salt) and packaged in plastic bags, were purchased from local markets, then stored at 4 °C. The mozzarella samples, packaged in their bags with the covering liquid were placed intact into 500 mL carbon fiber tubes with a diameter of 80 mm for the treatment. Irradiation was carried out at ambient temperature of 20 °C using a low-energy X-ray irradiator (RS-2400, Radsource Inc., Suwanee, GA, USA) operating at the voltage of the X-ray tube of 150 kV and a current of 45 mA. The average dose absorbed by the samples under X-ray irradiation was estimated with an alanine/electron paramagnetic resonance dosimetry system. The calibration of the absorbed dose determined by the alanine dosimeter, based on the amplitude of the corresponding ESR signal, was performed using a reference calibrated ionization chamber (Radcal Inc., Monrovia, CA, USA) calibrated by accredited laboratories. The uncertainty of the value of the delivered dose was around 5% while the inhomogeneity of the dose imparted due to the sample size is estimated at around 15% compared to the average value. For this investigation, a dose level of 3.0 kGy at a dose rate of approximately 2.0 kGy h^−1^ was used. Aliquots of non-irradiated mozzarella cheese were used as controls.

3.3. Microwave-Assisted Extraction (MAE) Procedures

MAE was performed using an ETHOS-ONE microwave system with a 100 mL Teflon vessel (Milestone s.r.l., Sorisole, Bergamo, Italy). In this study, an MAE method, previously optimized in our work [26], was applied to 30 aliquots of irradiated samples (IRR) and 30 aliquots of controls (NI), subsequently analyzed in duplicate for a total of 120 runs. Briefly, 12 mL of ethanol/ethyl acetate, in a ratio of 1:2 v/v was used as the solvent mixture and added to 0.5 g of homogenized mozzarella sample (controls and irradiated samples). A total of 20 µL of the working standard solution of trinonanoin 10,000 mg L^−1^ was spiked as the internal standard. A microwave temperature program was set to heat the sample to 65 °C in 15 min; then, the temperature was held for 18 min and finally reduced to 25 °C. The maximum extraction power was set to 700 W. The extract was collected in a 50 mL falcon and evaporated to dryness at 40 °C under nitrogen flow using an automated solvent evaporation system TurboVap^®^ II (Biotage AB, Uppsala, Sweden). All dried extracts were suspended in MeOH/CHCl_3_ (4:1, v/v) to obtain a final lipid concentration of 2000 mg L^−1^, centrifuged at 400 RCF for 10 min at 4 °C (±2 °C), and then the supernatant was analyzed with two repetitions by UHPLC-Q-Orbitrap-MS.

3.4. LC-MS Analysis and Lipid Identification

Samples were first subjected to chromatographic separation of the lipid mixture, followed by mass spectrometry analysis using an Ultimate 3000 UHPLC system coupled to a Q Exactive Focus Orbitrap Mass Spectrometer (Thermo Fisher Scientific, Waltham, MA, USA) equipped with a heated electrospray ionization (HESI) source. Chromatographic conditions and mass spectrometry parameters are reported in Table 5. A procedural blank (quality assurance, QA) was employed to ensure experimental reliability, and it was also used for feature search and alignment during LipidSearch™ data processing. Quality control (QC) was performed by spiking the sample matrix with a mixture of Equisplash™ Lipidomix^®^ and trinonanoin, which was injected every ten runs, to monitor signal intensity, retention time, relative standard deviation of peak area (RSD%), and mass accuracy. Stability and reproducibility of chromatographic runs and MS acquisitions were confirmed when the RSD% of deuterated lipids in QC samples was below 20% (Table S3) [24]. In addition, 150 µL of six individual samples were pooled to generate a representative pooled sample, which was injected at the beginning of the analytical batch to condition the chromatographic system [23]. Successively, UHPLC-Q-Orbitrap-MS data were processed by LipidSearch^TM^ v4.2.2.7 software (Thermo Fisher Scientific, Waltham, MA, USA). Detailed software parameters are reported in Table 1. For oxidized lipids, the “Oxid. GPL” option was included in the LipidSearch™ database, and their identification was further supported using FreeStyle™ v1.6 software (Thermo Fisher Scientific, Waltham, MA, USA).

3.5. Statistical Analysis

To correct for batch effects, the data were normalized using the total peak area. Subsequently, log transformation was applied to prepare the data for an unsupervised statistical analysis. Multivariate statistical elaboration was performed using both unsupervised, PCA and supervised models, including PLS-DA and ANN.

ANN-based algorithms for the discriminative analysis of lipids are still less commonly applied to food samples differing by technological treatments. An ANN is composed of interconnected nodes to obtain behavior similar to biological neurons [41]. Each node receives input data and processes them through parametric mathematical functions, generating an output that is transferred to the subsequent nodes. The number of input nodes corresponds to the number of variables, in this case, the lipids. The processed information is then propagated to a set of nodes that form one or more internal hidden layers. Nodes in each layer elaborate the signals received from the previous layer and transmit the resulting information to the following one, until it reaches the output layer, which returns the final external response.

Chemometric analyses were performed using Rstudio 15 May 2023 Build 524 4.2.3. (R Development Core Team, Vienna, Austria, 2020), using in-house routines, partly based on the mdatools and neuralnet packages [42,43]. Metabolomic analysis, specifically, a functional enrichment of lipids, was conducted using the LION package (version 0.1.0) equipped with a large lipid ontology database, encompassing classifications of lipid properties, functions, and subcellular components [39].

4. Conclusions

In this study, high-resolution mass spectrometry combined with an appropriate multivariate analysis was applied to mozzarella cheese to obtain its lipid fingerprint and subsequently to evaluate possible radio-induced modifications after X-ray irradiation at a dose of 3.0 kGy.

Both irradiated and non-irradiated mozzarella cheese samples were characterized by a total of 478 lipids, including glycerolipids, sphingolipids, phospholipids, and lysophospholipids, many of which contain bioactive compounds with high nutritional value. These qualitative results indicate that the nutritional lipid profile remained substantially unchanged at the irradiation dose investigated.

Furthermore, supervised chemometric tools enabled the development of validated multivariate models allowing clear discrimination between irradiated and non-irradiated samples. Among the evaluated models, both PLS-DA, with double cross-validation, and ANN achieved high performance. The latter, based on the entire lipid fingerprint, showed the highest discrimination capability with a sensitivity, specificity and efficiency close to one. However, PLS-DA allowed the identification of 30 VIP as potential treatment markers. In this regard, X-ray irradiation was associated with an increase in oxidized acylglycerols, reflecting irradiation-induced oxidative phenomena, as well as a decrease in specific lipid species, including ceramides, hexosylceramides, sphingomyelins and phosphatidylethanolamines. These molecular changes mainly affected the structural organization of the lipid globules derived from milk and present in cheese. In conclusion, this study has an exploratory and scientific purpose. The findings advanced the current knowledge in the field of food irradiation and confirmed the value of the omics-based approach combined with chemometric tools to understand the effects of technological treatments on food. Moreover, the identification of new potential irradiation markers could support food safety traceability and control plans. Further investigations on different samples of mozzarella by shape, brand, and production technology are necessary to validate and then confirm the potential lipid markers identified.