Cheese yield and nutrients recovery in the curd predicted by Fourier-transform spectra from individual sheep milk samples

The objectives of this study were to explore the use of Fourier-transform infrared (FTIR) spectroscopy on individual sheep milk samples for predicting cheese-making traits, and to test the effect of the farm variability on their prediction accuracy. For each of 121 ewes from 4 farms, a laboratory model cheese was produced, and 3 actual cheese yield traits (fresh cheese, cheese solids, and cheese water) and 4 milk nutrient recovery traits (fat, protein, total solids, and energy) in the curd were measured. Calibration equations were developed using a Bayesian approach with 2 different scenarios: (1) a random cross-validation (80% calibration; 20% validation set), and (2) a leave-one-out validation (3 farms used as calibration, and the remaining one as validation set) to assess the accuracy of prediction of samples from external farms, not included in calibration set. The best performance was obtained for predicting the yield and recovery of total solids, justifying for the practical application of the method at sheep population and dairy industry levels. Performances for the remaining traits were lower, but still useful for the monitoring of the milk processing in the case of fresh curd and recovery of energy. Insufficient accuracies were found for the recovery of protein and fat, highlighting the complex nature of the relationships among the milk nutrients and their recovery in the curd. The leave-one-out validation procedure, as expected, showed lower prediction accuracies, as a result of the characteristics of the farming systems, which were different between calibration and validation sets. In this regard, the inclusion of information related to the farm could help to improve the prediction accuracy of these traits. Overall, a large contribution to the prediction of the cheese-making traits came from the areas known as “water” and “fingerprint” regions. These findings suggest that, according to the traits studied, the inclusion of water regions for the development of the prediction equation models is fundamental to maintain a high prediction accuracy. However, further studies are necessary to better understand the role of specific absorbance peaks and their contribution to the prediction of cheese-making traits, to offer reliable tools applicable along the dairy ovine chain


INTRODUCTION
Sheep dairy products are common worldwide, and in the last 20 years, the global production of ovine milk and cheese has increased by 27 and 9%, respectively.Indeed, the consequent gain of the gross production value of the sheep dairy chain expanded by more than 7%, which means a potential international market of about $5 billion (FAOSTAT 2022).Nevertheless, to guarantee the environmental and economic sustainability and the future of this dairy chain, the entire sector must address climate change and water quality while providing safe products for the world.To achieve this goal, it is essential to provide monitoring tools along the dairy chain that can help to select efficient animals, limit the production waste, and reduce the cost: income ratio.At the dairy industry, the percentage cheese yield (%CY) is one of the most informative cheese-making traits that must be monitored, together with the recovery of nutrients (%REC) from milk in the curd.However, these traits cannot be routinely measured in practice and neither the simulation of cheese-making procedures through laboratory models (Wedholm et al., 2006;Cipolat-Gotet et al., 2016a) can be applied on a large number of samples (i.e., individual animal milk samples) because of the cost and the huge labor of each step (Cipolat-Gotet et al., 2013).As most of the sheep milk is used to produce cheese, cheese-making traits should be the most appropriate selection goals for dairy breeds to better address the market demand.
But the moisture content of cheese is depending especially on cheese-making procedures and on the length and condition of ripening (Cipolat-Gotet et al., 2020;Stocco et al., 2022), so it is especially the cheese solids, expressed per 100 kg of milk processed (%CY SOLIDS ) and per 100 kg of milk solids (%REC SOLIDS ) to be particularly valuable for the dairy industry.Most of the programs in sheep still select only for total milk yield per lactation while major milk components (i.e., milk fat and protein) are used only in few dairy breeds.In these latter, genetic improvement of %CY is not pursued directly, but by including percentage of milk fat and protein in selection indices (Othmane et al., 2002), as major traits affecting the variability of %CY (Fox et al., 2017).Moving along the dairy chain, also at the dairy industry the use of these traits would improve the monitoring of the efficiency of cheese-making process, giving a tool to the cheese makers to daily correct the procedures used for cheese production.Both at animal and bulk milk levels, fat and protein are routinely measured via Fourier-transform infrared (FTIR) spectroscopy, a technique used also to indirectly measure a large number of nutritional characteristics of milk (Soyeurt et al., 2006;Rutten et al., 2010;Yakubu et al., 2022) but also informative traits for the environmental impact of dairy production (Gengler et al., 2016;Bittante and Cipolat-Gotet, 2018;Coppa et al., 2022).Regarding the prediction of characteristics of milk useful for dairy industry (e.g., coagulation properties, cheesemaking), literature for the dairy cattle is massive (Bittante et al., 2014;Ferragina et al., 2015;El Jabri et al., 2019), whereas limited work has been done for small ruminants (Ferrand-Calmels et al., 2014;Ferragina et al., 2017;Stocco et al., 2021).So, the possibility to gain information on cheese-making ability of milk in a single-fast and accurate analysis could be advantageous for improving the efficiency along the entire ovine dairy chain.Therefore, the aims of the present study were to investigate the feasibility of using FTIR milk spectra to (1) predict %CY and %REC traits, (2) quantify the effect of individual farm on their prediction accuracy, and (3) assess the importance and the potential contribution of single wavelengths to these traits, using individual sheep milk samples.

MATERIALS AND METHODS
All the dairy ewes involved in this study were reared in commercial private farms and were not subjected to any invasive procedures.Milk samples used for the analyses were collected during routine milking.

Data Collection
Individual milk was collected from 121 ewes reared in 4 different farms of the regional territory of Sardinia, Italy.At each of the 4 farms, ewes were sampled in a single day (one sampling day for each farm).Details of animals and farms were previously reported by Cipolat-Gotet et al. (2016a).Briefly, ewes were of the Sarda dairy breed, between the first and eighth parity, and 65 to 220 d in milking, milked twice a day (at 0600 and 1600 h) using manually operated milking machines; farms were conducted by following semi-extensive methods.The samples were collected in plastic bottles, immediately refrigerated at 4°C and analyzed within 24 h after collection.Information on farming systems and coagulation ability of the ovine milk of the area is available in previous studies (Pazzola et al., 2014;Vacca et al., 2015).

Milk Analyses
For each sample, an aliquot of 100 mL was used to achieve milk composition at the laboratory of the Regional Farmers Association of Sardinia (ARAS, Oristano, Italy).Fat, protein, lactose, TS, and pH were analyzed using a MilkoScan FT6000 (Foss Electric, Hillerød, Denmark), according to FIL-IDF recommendations calibrated for ovine milk (ISO 9622 and IDF 141:2013).

Model Cheese-Making
Another aliquot of 500 mL/ewe was processed at the Laboratory of the DAFNAE Department of the University of Padova (Legnaro, Italy) using a bench-top model cheese-making method.This procedure mimics the production of a fresh, uncooked and artisanal cheese and allows the assessment of %CY and %REC traits Cipolat-Gotet et al. (2016a).Briefly, each sheep milk sample (500 mL) was heated to 35°C for 30 min, inoculated with a starter culture [freeze-dried formulation of thermophilic lactic bacteria (Delvo-Tec TS-10A DSL; DSM Food Specialties, Delft, the Netherlands), solubilized with skim milk before use and concentrated to 8-fold].The acidified milk was mixed with a rennet solution at 1.2% wt/vol (Hansen Standard 215, 215 international milk clotting units (IMCU)/mL; with 80 ± 5% chymosin and 20 ± 5% pepsin; Pacovis Amrein AG, Bern, Switzerland; diluted in distilled water to yield 0.051 IMCU/mL −1 of milk).Coagulation time was assessed by the visual observation of gelation with the aid of a spoon.After coagulation, each curd obtained from the individual samples was cut into cubes of about 0.5 cm 3 .Curds were then separated from the Pazzola et al.: CHEESE-MAKING TRAITS PREDICTED VIA SHEEP MILK SPECTRA whey, drained, and pressed, for 18 h at room temperature, using a weight of 1 kg.At the end of the process, each curd and the corresponding whey were weighed and analyzed to achieve composition.Dry matter, fat, and protein of the curd were measured by a FoodScan (Foss, Hillerød, Denmark) calibrated according to the following reference methods: DM (ISO 8968-1:IDF 20-1; determination of the TS content); fat (ISO 1735:IDF 5; gravimetric method); protein (ISO 5534:IDF 4; determination of nitrogen content, Kjeldahl).Dry matter, fat, protein, and lactose of the whey were analyzed by a MilkoScan FT2 (Foss, Hillerød, Denmark) calibrated according to the following reference methods: TS (ISO 2920:IDF 58; determination of DM); fat (ISO 1854:IDF 59; gravimetric method); protein (ISO 8968-1:IDF 20-1; determination of nitrogen content, Kjeldahl); lactose (ISO 5765-1:IDF 79-1; enzymatic method; IDF, 2013: ISO 9622 andIDF 141:2013).The energy of milk and whey was computed on the basis of the standard values in kilojoules per gram by the NRC (2001; fat: 38.89 kJ × g −1 ; protein: 23.90 kJ × g −1 ; lactose: 16.53 kJ × g −1 ).The following cheese-making traits were obtained from the weights (g) and chemical compositions of milk, curd, and whey (Cipolat-Gotet et al. 2016b): (1) 3 %CY traits, namely the yield of fresh curd (%CY CURD ), DM (%CY SOLIDS ) and water retained in curd (as the difference between the milk processed and the whey obtained, %CY WATER ) calculated as the respective percentage of weight of curd, DM, and water out of the processed milk weight; (2) 4 %REC traits, namely the recovery of protein (%REC PROTEIN ), fat (%REC FAT ), solids (%REC SOLIDS ), and energy (%REC ENERGY ) calculated as the respective percentage of the nutrient or energy in the curd out of the corresponding nutrient or energy in the processed milk.

Milk Spectra Storing, Editing, and Chemometric Models
For each milk sample, the FTIR spectrum over the range from wavenumbers 5,011 to 925 × cm −1 was collected and recorded as absorbance (A) using the transformation A = log (1/T), where T is the transmission.Two spectra were collected for each sample, and the results were averaged before data analysis.
Before spectra analysis, the raw absorbance values of every wavelength in the FTIR spectra of the milk samples (Figure 1A) were centered and standardized (Figure 1B) to a null mean and a unit sample variance.This procedure reduces the noise along the spectrum mainly due to the water content of milk and allows to boost wavenumbers with potential significant information.To detect outliers, Mahalanobis distances were calculated by means of the Mahalanobis function im-plemented in the R software (R Core Team, 2013).No samples were discarded because all the spectra showed a distance value lower than the mean ± 3 standard deviations.Indeed, the spectra were not subjected to any other mathematical pretreatment.
A Bayesian linear regression was used to predict the %CY and %REC traits.All phenotypes were regressed to the 1,060 FTIR standardized wavelengths under the following model: where μ is the overall mean, x ij are the wavelengths, β j are the regression coefficients, and e i the residual with ( ) The BayesB model implemented in the BGLR R package was adopted (de los Campos and Perez-Rodriguez, 2015) as described in Ferragina et al. (2017).

Cross-Validation and Leave-One-Out Validation Procedures.
The accuracy of the model and the prediction equation were assessed by a random cross-validation (CRV) and a leave-one-out validation procedure.In the random CRV, a calibration set (CAL; 80% of the total records) was used to build the equation, and a validation set (VAL; 20% of the total records) was used to test the model.The samples in CAL and VAL were randomly assigned.To account for the individual milk samples variability, the procedure was repeated 10 times for each trait and results were presented averaging the 10 replicates.
To quantify the effect of the individual farms on predictive ability of the model, a leave-one-out validation procedure was applied, predicting the cheese-making traits per each farm in validation at a time.Sheep milk samples from 3 farms were used to build the equation (CAL), and the remaining samples belonging to one farm were used to test the model (VAL).The procedure was repeated 4 times, so that all farms were individually evaluated.
The models' performance was assessed by measures of root mean squared error (RMSE), and coefficient of determination (R 2 ), in both CAL and VAL (R 2 CAL and R 2 VAL , respectively) and by residual predictive deviation, difference between measured and predicted values (bias), and slope.
For each trait, the equation coefficients corresponding to each wavelength (n = 1,060), the Pearson correlations between each milk absorbance at a given wavelength, and the corresponding measured value of the trait were assessed.The results were plotted along the spectral range and divided into 5 spectral regions, as reported by Ferragina et al. (2017): short-wave infrared (SWIR, often reported in literature as near-infrared), transition between SWIR and mid-infrared (SWIR-MWIR), first and second mid-infrared (MWIR1 and MWIR2, respectively) and transition between midand long-infrared (MWIR-LWIR).
Further, the potential contribution of the single wavelengths to the prediction of the investigated trait was defined as the product of the standardized absorbance of each spectrum and the corresponding regression coefficient (β).Then, the percentage contribution of each spectral region was calculated as the sum of the dif-ference, in absolute value, between the maximum and minimum potential contribution of each wavelength falling within the specific region, to the contribution of the entire spectrum, multiplied by 100.

Prediction Accuracy of Cheese-Making Traits from Sheep Milk
Descriptive statistics for %CY and %REC traits (mean ± SD) and results from the cross-validation procedure using individual sheep milk FTIR spectra are summarized in Table 1.The prediction statistics in calibration were reliable for %CY traits, with average high values of R 2 CAL (0.77-0.89) and low of RMSE CAL (0.43-1.16).The prediction statistics for %REC traits were similar for %REC SOLIDS and %REC ENERGY (R 2 CAL = 0.81 and 0.72; RMSE CAL = 1.81 and 1.78, respectively), moderate for %REC PROTEIN (R 2 CAL = 0.48, RM-SE CAL = 1.49), and poor for %REC FAT (R 2 CAL = 0.21, RMSE CAL = 1.36).
The %CY traits showed different prediction accuracies, with the RMSE VAL of 1.61, 0.41, 1.24, and R 2 VAL of 0.75, 0.91, 0.59, respectively for %CY CURD , %CY SOLIDS , and %CY WATER (Table 1).It is important to consider that the calibration accuracy obtainable with a secondary technique, such as infrared spectroscopy, cannot be greater than the repeatability of the reference measures that is affected by several factors such as the sampling, instrumental features, and the operator.If compared with the statistics reported by Ferragina et al. (2015) for the prediction of %CY CURD in bovine species, our values were slightly higher in terms of both R 2 VAL and RMSE VAL (0.71 and 1.44, respectively).This is probably due to the different quality and technological properties of the milk produced and the consequent different average of the cheese-making traits between the 2 species (Bittante et al., 2022) as well as to the variability of the data set used.
Regarding the %REC traits, the best prediction accuracies were observed for %REC SOLIDS (R 2 VAL = 0.86; RMSE VAL = 1.60) and %REC ENERGY (R 2 VAL = 0.74; RMSE VAL = 1.69).The %REC PROTEIN had a very low R 2 VAL (0.23) and showed the highest RMSE VAL (1.94), but the lowest accuracy in terms of R 2 VAL was for %REC FAT (0.13).As expected, models with poor performance in the calibration data set, as in the case of the %REC PROTEIN and %REC FAT , produced even poorer predictions in the validation.In this regard, the variability of the reference data in the calibration set and the repeatability of the predicted trait are the most important factors affecting the prediction accuracy (Grelet et al., 2021).Moreover, whereas models predicting the main milk components (i.e., fat, protein) are effective, because the absorbance response is linear to the concentration of the molecules, on the other hand, in the case of minor components (e.g., fatty acids) or indirect traits (e.g., coagulation, cheese-making phenotypes) this may not be true, as the signal in the spectral data is associated with traits having a relationship with the targeted trait.In this case, predictive models may be partly or totally ineffective, because they are predicting molecules having low or no interaction between the chemical element and the infrared rays (Ferragina et al., 2015;Grelet et al., 2021;Stocco et al., 2021).The FTIR spectroscopy is not expected to detect differences in the recovery of nutrients indeed, but rather in the major components affecting the variability of %REC traits.For example, in ovine milk the %REC PROTEIN and %REC FAT are largely affected by milk protein and to a lesser (e.g., %REC PROTEIN ) or no extent (e.g., %REC FAT ) by fat (Vacca et al., 2019).Moreover, in particular %REC FAT is much more dependent on physical properties (e.g., fat globule size, curd-firming rate, and curd cutting; Fagan et al., 2007;Cipolat-Gotet et al., 2013) than on chemical composition, which could explain the low accuracy of the prediction models.If compared with results from other species, the largest differences in prediction accuracies are mainly ascribed to %REC PROTEIN and %REC FAT .In fact, R 2 VAL and RMSE VAL values are much lower than the prediction accuracy provided by Ferragina et al. (2015) in dairy cattle (R 2 VAL = 0.65 and 0.28, and RMSE VAL = 1.44 and 3.13, respectively for %REC PROTEIN and %REC FAT ), and by Stocco et al. (2023) in dairy goats (R 2 VAL = 0.62 and 0.34, and RMSE VAL = 1.55 and 3.28, respectively for %REC PROTEIN and %REC FAT ).The dissimilarities observed among these comparisons could be explained by the fact that the relationships among the milk nutrients and their recovery in the curd are different across species, due to the different quantity and proportion of milk components and their physical characteristics.For example, Nicolaou et al. (2010), investigating the use of FTIR spectroscopy for the detection of milk from different species, found different degree of absorption in the CH 2 band, which is related to the acyl chain on fatty acid, as well as for C-O bands related to protein and lactose.The absorption values were higher in sheep milk spectrum compared with that of cow and goat.
In the CRV, the slope closest to one (1.03) as well as the lowest bias (0.02) were observed for %CY CURD .Average values of bias ranged from −0.16 of %REC PROTEIN to 0.11 of %CY WATER (Table 1), whereas their variability for %REC PROTEIN and %REC FAT was quite high (−0.16 and 0.08, respectively; Table 1).Probably this result was much more linked to the different phenotypic variability of these %RECs than their average values.Indeed, it is recognized that %REC PROTEIN is less variable than %REC FAT , and this leads to underfit the predicted data.Similarly, %CY WATER is usually characterized by a higher variability with respect to %CY SOLIDS (Cipolat-Gotet et al., 2013) and, for this reason, the former had the highest bias among %CY traits values were for %REC PROTEIN and %REC FAT (1.27 and 1.58, respectively), suggesting that the degree to which it is possible to predict them is higher than the other traits.These findings further justified additional analyses, as the farm factor is expected to be important for the variability of milk quality (Vacca et al., 2019).

Effect of the Farm on the Prediction Accuracy of Cheese-Making Traits
The results from the leave-one-out validation procedure are reported in Table 2.The R 2 VAL values ranged from 0.33 to 0.75 among %CY traits, and from 0.07 to 0.67 for the %REC traits, with very large variability of the range within each trait.As above mentioned, these results confirm that the variability previously found in the CRV procedure is actually due also to the individual farm.The RMSE VAL ranged from 0.61 to 1.85 among the %CY traits, and from 1.67 to 2.33 for the %REC traits, again with large variability of the interval range within each single trait.Apparently, it seems that the leave-one-out validation model is not working properly for some traits (e.g., %REC PROTEIN , %REC FAT ), but the poor prediction accuracies here are the result of flock specificities (e.g., diets, management), which are different between calibration and validation sets.Our results agree with previous studies investigating crossvalidation and external validation strategies (Roberts et al., 2017;Wang and Bovenhuis, 2019;Eskildsen et al., 2021).These findings confirm that the possibility to add information on zootechnical and environmental characteristics related to the farms is fundamental to not incur large predictions errors.Hence, to make those models usable in routine, model robustness should be updated by adding new samples until covering an acceptable proportion of samples when tested under real field conditions.

Importance and Potential Contribution of Single Wavelengths for the Prediction of Cheese-Making Traits
In the Figure 2 are reported the prediction equation and simple correlation coefficients for the absorbance of each wavelength of the milk FTIR spectrum for the traits having the best predictive performances (Table 1; %CY CURD , %CY SOLIDS , and %REC SOLIDS , respectively).In all the 3 cheese-making traits, the SWIR region was characterized by a constant positive correlation, but still not consistent with any high regression coefficient.This finding agrees with the results found by Stocco et al. (2023) investigating the prediction accuracy of cheese-making traits using FTIR milk spectra from dairy goats.Indeed, the SWIR region is not known to contain any absorbance peak specific to a chemical bond found in milk (Karoui et al., 2010), so it would be expected that the contribution of this region to the prediction of cheese-making traits would be close to zero.However, plots showing the potential contribution (e.g., regression coefficients multiplied by the absorbances) of each wavelength for the prediction of cheese-making Pazzola et al.: CHEESE-MAKING TRAITS PREDICTED VIA SHEEP MILK SPECTRA traits (Figure 3) evidenced that this region accounted for about 10%, 8% and 15% respect to the potential contribution of the entire spectrum for the prediction of %CY CURD (Figure 3A), %CY SOLIDS (Figure 3B), and %REC SOLIDS (Figure 3C), respectively.So, further research is needed to evaluate the value of this region of the milk spectrum.
In the SWIR-MWIR region, also known as "water" region, several regression coefficients were recorded across the 3 cheese-making traits, with wavelength peaks close to 0.5 in absolute value for %REC SOLIDS (Figure 2C).On the contrary, Stocco et al. (2023) reported an almost null correlation and regression coefficients in this region for goat milk.A possible interpretation of this difference could be linked to the proportion of dry matter respect to the total fresh curd, which is similar between the 2 species, but the quantity of nutrients (and of curd obtained) is higher in sheep compared with the goat.Accordingly, the signals caused by the peaks are stronger in sheep spectrum compared with the goat.Probably, the weakness of these signals in the goat spectrum causes an increase in the noise in turn, that affects the overall prediction accuracy of the model.It is interesting to notice that, in all the 3 cheese-making traits, the region containing most of the peaks was the SWIR-MWIR region (Figure 3).Indeed, it accounted from 25% (%CY SOLIDS ) to 30% (%REC SOLIDS ) respect of the total potential contribution of the entire spectrum on the traits' prediction, with the most contributing signal at 3,576 wavenumbers × cm −1 for both %CY CURD and %REC SOLIDS .The same wavelength did not show the highest regression coefficient (Figure 2) evidencing the importance of further deepening the relationships between the spectrum peaks and the studied trait.In the case of %CY SOLIDS , the most contributing signal was at 3,260 wavenumbers × cm −1 .In addition, as abovementioned, the SWIR-MWIR is often referred to as the main "water" region, so our findings clearly evidence that the inclusion of wavelengths related to the water is fundamental for the prediction of these traits.Similarly, Toledo-Alvarado et al. ( 2018) reported a significant association between cows' pregnancy status and wavelengths included within the SWIR-MWIR region.Wang and Bovenhuis (2018) found wavelengths included between 1,619 and 1,674 cm −1 and 3,073 to 3,667 wavenumbers × cm −1 characterized by a polymorphism in the DGAT1 gene, known for being associated with milk quality traits (Lu et al., 2020).On the contrary, Bittante and Cipolat-Gotet (2018) in their study on the prediction of methane emission from milk spectra did not observe important signals in either of the 2 water absorption regions of the FTIR spectrum.In that study, the most important signals were concentrated in the MWIR-1 and the MWIR-LWIR regions.So, it may be assumed that the importance of considering the "water" regions depends on the target traits to be predicted and specific applications.For example, using the milk spectra after removing the "water" regions we observed lower R 2 CAL and RMSE CAL values (data not shown) compared with the results in CAL procedure by using the entire spectrum (Table 1) for all the cheese-making traits, except for %CY SOLIDS and %REC SOLIDS , which fitting statistics remained basically identical (data not shown).This means that the information contained in  the "water" areas are most likely useful for the prediction of the recovery of single milk components (e.g., fat, protein) rather than their sum (e.g., TS).
The MWIR-1 region of the milk spectrum is very informative, as it contains the major absorbance peaks for C-H, C = O, C-N, and N-H bonds.The highest regression coefficient for the 3%CY traits were at 2,975 cm −1 , around 2,377 to 2,280 cm −1 and at 1,720 wavenumbers × cm −1 (Figure 2).These areas are linked to the C = C (i.e., to fat), C = N (i.e., amines) and C = O (i.e., aldehydes and ketones) bonds, respectively (Stuart, 2004).In particular, the 2,975 wavenumbers × cm −1 is associated with a specific vinyl bond (i.e., C = CH 2 ; Bittante and Cecchinato, 2013) accountable for the oxidation flavor of milk (Hoff et al., 1959).This region accounted for about 33, 36, and 29% respect to the total potential contribution of the entire spectrum for the prediction of %CY CURD , %CY SOLIDS , and %REC SOLIDS , respectively (Figure 3).In particular, the highest contribution here for %CY CURD and %REC SOLIDS was observed at 2,978 wavenumbers × cm −1 , wavelength not previously associated with traits of interests and therefore not yet mentioned in other studies.However, this wavelength is near to others associated with vinyl bond type (e.g., C = CH 2 ; Bittante and Cecchinato, 2013) in bovine milk, and comprised within a group of signals related to other phenotypes useful for the dairy industry (Ferragina et al., 2015), fatty acids, and enteric traits (Bittante and Cipolat-Gotet, 2018).In the MWIR-2 region the 2 highest peaks were at 1,659 wavenumbers × cm −1 for %CY CURD (Figure 2A) and at 1,650 wavenumbers × cm −1 for %REC SOLIDS (Figure 2C), associated with amide I (i.e., 3-turn helix; Stuart, 2004) and dienes bonds (i.e., conjugated C-C; Bittante and Cecchinato, 2013), respectively.This region accounted only for about 4% on the total potential contribution for the prediction of all the 3 cheese-making traits (Figure 3).

CONCLUSIONS
The high prediction performance of %CY SOLIDS and %REC SOLIDS justify its practical application within the ovine dairy industry.Performances for the remaining traits were lower, but still useful for the monitoring of the milk processing in the case of %CY CURD , and %REC ENERGY , but insufficient for %REC PROTEIN and %REC FAT .The leave-one-out validation procedure showed more modest prediction accuracies, as a result of farm specificities (e.g., diets, management, environment), which were different between calibration and validation sets.In this respect, the inclusion of information on farming system could help to improve the prediction accuracy of these traits.Overall, the largest contribution to the prediction of the cheese-making traits came from the SWIR-MWIR and MWIR-LWIR spectrum regions, supporting the idea that the "water" region is important to keep a high prediction accuracy at times.Further studies are necessary to better understand the role of specific absorbance peaks and their contribution to the prediction of cheese-making traits.
Figure1.Plots showing the absorbance of raw (A) and standardized (B) milk spectra (solid black line represents the average, whereas the 2 dotted black lines represent the maximum and the minimum of each wavelength, respectively).Spectral regions are identified as shortwavelength infrared (SWIR), short-and mid-wavelength infrared (SWIR-MWIR), MWIR-1, MWIR-2, and mid-and long-wavelength infrared (MWIR-LWIR).
Pazzola et al.: CHEESE-MAKING TRAITS PREDICTED VIA SHEEP MILK SPECTRA Table 1.Descriptive statistics of cheese-making traits 1 obtained by a model cheese-making procedure and results (mean ± SD) from the cross-validation procedure using Fouriertransform infrared (FTIR) spectra of individual sheep milk samples Item Cheese yield (%CY) Milk nutrients recovery in the curd (%REC) %CY) weight of fresh curd, curd solids, and curd water as percentage of weight of milk processed (%CY CURD , %CY SOLIDS , and %CY WATER , respectively); milk nutrients recovery in the curd (%REC) weight of the curd components protein, fat, TS, and energy (%REC PROTEIN , %REC FAT , %REC SOLIDS , and %REC ENERGY , respectively) to the same component in milk, multiplied by 100. 2 Interval = minimum and maximum of observed values.3 N CAL = number of samples in calibration; R 2 CAL = coefficient of determination in calibration; RMSE CAL = root mean square error in calibration; N VAL = number of samples in validation; SD VAL = standard deviation in validation; R 2 VAL = coefficient of determination in validation; RMSE VAL = root mean square error in validation; RPD = residual predictive deviation.

Figure 2 .Figure 3 .
Figure 2. Prediction equation coefficients obtained using the BayesB method (solid green line) and simple correlation coefficients (CORR; dashed black line) for the absorbance of each wavelength of the milk FTIR spectrum for the prediction of cheese-making traits (A, %CY CURD = fresh cheese yield; B, %CY SOLIDS = cheese yield in TS; C, %REC SOLIDS = recovery of TS).Spectral regions are identified as short-wavelength infrared (SWIR), short-and mid-wavelength infrared (SWIR-MWIR), MWIR-1, MWIR-2, and mid-and long-wavelength infrared (MWIR-LWIR).

Table 2 .
Pazzola et al.: CHEESE-MAKING TRAITS PREDICTED VIA SHEEP MILK SPECTRA Prediction statistics (mean ± SD) of cheese-making traits 1 obtained by a model cheese-making from the leave-one-out validation procedure using Fourier-transform infrared (FTIR) spectra of individual sheep milk samples Cheese yields (%CY) weight of fresh curd, curd solids, and curd water as percentage of weight of milk processed (%CY CURD , %CY SOLIDS , and %CY WATER , respectively); milk nutrients recovery in the curd (%REC) weight of the curd components protein, fat, TS, and energy (%REC PROTEIN , %REC FAT , %REC SOLIDS , and %REC ENERGY , respectively) to the same component in milk, multiplied by 100.