A total of 175 fresh milk samples were used in this study, obtained from 2 dairy farm studies conducted in Los Lagos Region, Chile, between 2018 and 2020 as follows.

Set 1: Between October 2018 and March 2019, 133 bulk tank milk samples were collected from a dairy farm in Fresia, Los Lagos Region, Chile (41 ° 28′18 ″ S 72 ° 56′12 ″ W). During this period, the tested dairy farm had 43 lactating Holstein Friesian cows with an average production of 21 L/cow/day. Their diet was composed of grazed pasture [ Lolium Perenne (54%), Lolium Trifolium (34%), Holcus Lanatus (7%), Trifolium repens (2%) and 3% of other species] and 2 kg of a pelleted concentrate (2 kg/cow/day).

Set 2: Between August 2019 and February 2020, 42 milk samples were collected from milk collection trucks at a dairy processing facility in La Araucanía region. The sample collection period represented the pasture-grazing period of the year, and all milk sampled from the trucks was produced in pasture-grazing based dairies.

All milk samples collected were kept at 4°C during transport to the NIR spectroscopy laboratory of INIA-Remehue. Each fresh milk sample was registered by NIRS and subsequently stored at -80 ° C until FA analysis by gas chromatography was carried out.

The derivatization of milk fat was performed with 3 mL of fresh milk at room temperature using the double fat extraction method of chloroform and methanol 1:1 (Kramer et al., 2008). For FA methylation, a base-catalyzed methylation procedure with sodium methoxide was used (i.e. 0.5N methanolic base #33080, Supelco Inc., Bellefonte. PA) as described by Cruz-Hernandez et al. (2006). Prior to methylation, 1 mL of internal standard was added to the sample for FA quantification (1 mg/mL of 23:0 methyl ester, n-23-M, Nu-Chek Prep Inc., Elysian. MN. USA). The contents in FA methyl esters (FAME) were expressed as g per 100 g of FAME quantified and as mg of FAME per 100 mL of fresh milk.

The FAME were analyzed using a GC equipped with a flame ionization detector (GC-2010 Plus; Shimadzu®. Kyoto. Japan), a capillary column (SP-2560; 100 m × 0.25 mm (i.d.) with 0.2-μm film thickness; Supelco Inc., Bellefonte. PA. USA) and an ionic liquid capillary column (SLB-IL111; 100 m × 0.25 mm (i.d.) with 0.2-μm film thickness; Supelco Inc., Bellefonte. PA. USA) to confirm the identification of several biohydrogenation intermediates such as CLA isomers and other trans FA (Delmonte et al., 2011). The samples were analyzed with two GC temperature programs that plateaued at 175 °C and 150 °C (Kramer et al., 2008). Hydrogen was used as carrier gas in both columns, with a constant flow rate of 1 mL/min. and the injector and detector temperatures were set at 250 °C.

For peak identification, one reference standard (GLC463), individual FAMEs (21:0. 23:0. 26:0), and a CLA mixture (9c,11t-/8t,10c-/11c,13t-/10t,12c-/8c,10c-/9c,11c-/10c,12c-/11c,13c-/11t,13t-/10t,12t-/9t,11t-/8t,10t-18:2; UC-59M) were used, all of which were obtained from Nu-Chek Prep Inc. (Elysian, MN, USA). Isomerized mixtures of linoleic (18:2n−6) and linolenic (18:3n−3) acids were purchased from Sigma-Aldrich, and branched-chain FA (BCFA) were identified using a bacterial FAME mixture from Matreya (Pleasant Gap, PA, USA).

Atherogenic (AI) and Thrombogenic (TI) indexes were calculated as in Ulbricht and Southgate (1991), as follows:

For spectral analysis, 175 fresh milk samples in reflectance mode were scanned using NIR spectroscopy (MPA-FT NIR, Bruker Optik GmbH, Ettlingen, Germany). Spectral data were transformed to absorbance (A) according to the equation: A = log10 (1/R), where R is the reflectance obtained at each wavenumber from 12.000-4.000 cm^-1 (NIR region) with 16 cm⁻¹ resolution and 64 scans (Figure 1). Partial least-squares regression (PLSR) with leave-one-out (LOO) cross validation was performed to fit predictive models using chemometrical software OPUS version 6.5 (Bruker Optik GmbH, Melvyn Becerra Cia. Ltda). For the cross-validation method, the selection of validation samples was based on spectral information as described by Conzen, (2006).

In addition, external validation was performed with a random subset of 10 samples which were not included in the calibration and cross validation method.

The software OPUS was used to apply different preprocessing to spectra:vector normalization (VN), multiplicative scatter correction (MSC), straight line subtraction (SLS), first derivative (FD), and second derivative (SED). Outliers were identified and removed during the calibration process to improve precision and model performance.

The criteria used in choosing the best prediction model considered: i) low root mean square error of cross validation (RMSECV), ii) high coefficients of determination in cross-validation (R²cv), iii) root mean square error of estimation (RMSEE); iv) residual predictive deviation (RPD: ratio between the standard deviation -SD- of the reference values and the error of prediction), and v) number of factors (Conzen, 2006).

In this study the criteria for selection were the lowest RMSECV, the lowest number of PLS factors and the highest RPD. Small error of cross validation was desired, as this would result in greater RPD values, and a better prediction model.

The fatty acid methyl ester contents in the milk of the 2 sets of samples used in this study are expressed as % of total FAME (g/100g) and mg FAME per 100 mL of milk and shown in Table 1. The ratio of n-6 to n-3 was 1.69 ± 0.5 with a range between 1.14 and 5 g/100g of FAME. The TI and AI were 3.09 ± 0.3 and 2.86 ± 0.4 g/100g of FAME, respectively, with a range of 2.16 to 3.88 and 1.96 to 4.07 for TI and AI, respectively.

NIR spectral features. Average absorbance of NIR spectra for liquid milk presented two bands with maxima at 7500-6400 cm^-1 and 5400-4900 cm-1 related to O-H first overtone and O-H combination band (Figure 1).

NIR calibration. The calibration statistics for individual FA and groups of FA and indexes used to evaluate the nutritional health properties of food (considering the potential negative or positive effects) in fresh bovine milk are detailed in Table 2. The coefficient of determination in the calibration sets fluctuated between 0.76 and 0.95. The RPD values varied between 2.1 and 4.3. The number of PLS factors values varied between 2 and 10. All pre-preprocessing methods used were different and in accordance to the main functional groups of FA and health related indexes. The relation between NIRS prediction and composition obtained by the reference methods for all main functional groups of FA and health indicators are shown in Figures 2a and 2b.

External validation. Table 3 shows the means of the residuals and the Root Mean Standard Error (RMSE) obtained from the external validation. The levels of significance obtained were between 0.09 for 10 trans-18:1 and 0.88 for LA. Therefore, no differences between the spectroscopic and chromatographic method were detected.

Regarding the FA composition of milk, the FAME contents in this study were in the same range reported previously for Holstein Friesian’s milk from Chilean dairies (Morales et al., 2015; Vargas-Bello et al., 2015), in milk from dairy processing facilities in Southern Chile (Pinto et al., 2002), and in pasture-grazing based dairies from other regions (Nantapo et al., 2014). For instance, total SFA (66.6 g/100 g of FAME) was in the same range reported for grazing animals by Morales et al. (2015); 67.19 g/100 g, and for TMR-fed Holstein Friesian cows of the control group of Vargas-Bello et al. (2015); 68.2 g/100 g.

Our reported milk FAME content is not comparable to other studies where diet relies on a total mixed ration based on preserved forages and grains because diet and fresh forage inclusion in the diet have a major impact on the FA profile of milk (Sun and Gibs, 2012). Although our study only included milk samples of Holstein Friesian cows, greater variances and range of milk FA observed in other studies could be explained, in lesser magnitude, by the inclusion of other dairy breeds. In this sense, Coppa et al. (2014) reported milk FA composition for a heterogeneous productive system that included five different dairy cow breeds present in northwest Italy, with total milk SFA and total CLA (64.28 and 0.93 g/100 g of FA, respectively) below our reported values (66.7 and 1.12 g/100 g FA, respectively). On the contrary, this last study reported greater values for MUFA, PUFA, and n-6/n-3 (29.63; 5.17; 3.42 g/100 g of FA, respectively) than our study, which could be explained by the use of TMR and a more diverse set of cow genetics that included Jersey cows.

The nutritional and health indexes PUFA/SFA, n-6/n-3, AI, and TI, are commonly used to evaluate the nutritional value and effects of edible products on consumer health. In general, a ratio of dietary PUFA to SFA above 0.45 and a ratio of n-6/n-3 below 4.0 are expected to reduce the risk of diseases such as coronary heart disease and cancer (Simopoulos, 2002). Furthermore, the low PUFA/SFA ratio (0.04 g/100g FAME) reported in this study was due to the high SFA content in the two sets of milk samples analyzed. The n-6/n-3 ratio obtained in the current study (1.69 g/100 g of FAME) is lower than the ratio reported by Morales et al. (2015), most likely because of differences in pasture botanical composition. Indeed, some authors have indicated that the PUFA/SFA ratio may not be adequate to evaluate the nutritional value of dietary fat, as it ignores the effects of MUFA and also, some SFA have no effect on plasma cholesterol (Orellana et al., 2009).

Milk atherogenic and thrombogenic indices in cattle from different breeds and feed and management systems have been previously reported. Kuczyńska et al. (2012) and Pilarczyk et al. (2015) reported AI of 2.1 and 2.5 in dairy cows from a pasture-based and a total mixed-ration-based dairy system, respectively, values which are slightly lower than those reported in our study. Nantapo et al. (2014) reported milk AI values of ​​4.08 and 5.13 for Jersey and Friesian x Jersey cows, respectively. In Tarentaise and Montbeliarde cows, Ferlay et al. (2006) reported an AI of 3.14 and 3.43. On the other hand, the TI reported in the present study is within the range of that reported by Vargas-Bello et al. (2015), and below that of Thanh and Suksombat (2015), who reported 4.11.

NIR spectral features. Average absorbance of NIR spectra for liquid milk presented small bands corresponding to FA and fat contents and appeared at 8900-7450 cm^-1 and 6000-5300 cm^-1, associated with the first and second overtones from C-H stretching vibration of methyl (-CH₃), methylene (-CH₂-), and ethenyl (-CH=CH-). In addition, the absorption bands of bovine liquid milk used in this study were similar to those reported by Coppa et al. (2010) and Llano-Suaréz et al. (2018).

NIR calibration results. For model evaluation, Williams (2014) proposed R² y RPD as parameters that serve to classify NIR models into excellent or good, when R² fall above 0.91 or between 0.9 and 0.82, respectively. However, when R² falls between 0.81 and 0.66, the model will predict approximate values, and only simple discrimination of low, medium and high when R² is between 0.65 and 0.5. The RPD value is a measure of comparison between the standard error of the predicted values with the deviation of the references data, and therefore will evaluate the NIR calibration model (Williams, 2014; Nuñez et al., 2016). Therefore, RPD above 4 and 3 are considered excellent and good, respectively, while values between 2.9 and 2 are acceptable for detection, between 2 and 1.5 acceptable for discrimination between low and high concentration, and below 1.5 are not useful.

Regarding R_CV ² statistics, DPA, ∑CLA, SFA, MUFA, PUFA, n-6, AI, PUFA/SFA and LA, showed good predictive capacity, with R_CV ² values between 0.80 and 0.85. The R_CV ² for EPA, n-3, n-6/n-3, TI, UFA, UFA/SFA, HFA, HFA/UFA, h/H, LNA, LA/LNA, 10 trans, 11 trans, and ∑ odd and branched FA were very acceptable with values from 0.7 to 0.8. The ratio MUFA/SFA had a R_CV ² of 0.6. Therefore, this model is considered not suitable to be used; however, discriminations between high, medium and low concentrations could still be made. Regarding our data, RPD values ranged from 1.6 to 2.5, and were in accordance with those detailed above when comparing R_CV ² values.

Possible explanations for our low NIR prediction for individual FA and some ratios may include: i) Low FA variability in the sample sets used in this study (Table 1) which may have limited their prediction from NIR spectra. The generation of a successful statistical model requires wider data sets, or data sets covering a wide range of concentrations. When random samples are used for the purpose of calibration, as was the case in this study, performance may be constrained by narrow data sets. Also, ii) the complexity of the aqueous matrix of liquid milk (minerals in solution, proteins in a colloidal dispersion, and lipids in emulsion), hinder the NIR analysis (Marinori et al., 2013). The highwater content of milk in the fresh state could limit the detection capacity of other constituents, since the absorption bands of water in NIRS are strong. Calibration models could be improved by the use of a bigger set of samples and with a greater range in concentration of analyzed FA.

Our results suggest that water content can often mask NIR signals and generate a limited predictive model (Reeves, 2000). It is necessary to work on reducing the difference between RMSEE and RMSECV values, which will serve to strengthen the calibration models (González-Sáiz et al., 2007).

The R² _CV and RPD obtained in this study for SFA, MUFA, PUFA, n-6, n-3, n6/n3, AI, LNA and 10 trans were higher than those reported by Núñez-Sanchez et al. (2016), in calibration models using dry and fresh samples of goat milk measured in reflectance and transflectance mode, respectively. The R² _CV obtained by Andueza et al. (2013) in samples of dry goat milk were higher than our reported R² _CV, except for n6 (0.26), LA (0.42), LNA (0.49) and ∑odd and branched fatty acids (0.38). In addition, our study obtained greater R² _CV and RPD values than those reported by Núñez-Sanchez et al. (2020) for EPA, DPA, MUFA, n-6, n-3 and h/H in thawed ewe’s milk samples.

As compared to our study, the models generated by Coppa et al. (2010) with samples of fresh cow’s milk measured in transflectance mode obtained lower R² coefficients in set validations for: ∑CLA (0.60); PUFA (0.65); n-6 (0.00); n-3 (0.20) and ∑odd and branched FA (0.57). Similarly, Coppa et al. (2014), using samples of fresh cow’s milk in reflectance mode, reported validation R² values for ∑CLA (0.71); MUFA (0.79); PUFA (0.77), n-6 (0.43); n-3 (0.72) and n-6/n-3 (0.66). In addition, our study obtained greater RCV2 and RPD values than those reported by Llano-Suárez et al. (2018) for SFA, MUFA, and PUFA in liquid milk samples.

In addition, only some authors have determined AI and TI by NIRS: i) Nuñez-Sanchez et al. (2016) in oven-dried goat milk samples in reflectance mode (AI=0.77; TI=0.74) and liquid goat milk samples in transflectance mode (AI=0.68; TI=0.75); ii) Llanos-Suárez et al. (2018) in liquid milk samples using a near infrared handheld spectrometer (AI=0.85; TI=0.89); and iii) Nuñez-Sanchez et al. (2020) in thawed ewe’s milk samples previously oven-dried on glass fiber filters (AI=0.75; TI=0.87). In comparison to these reports, our calibration models presented greater R² _CV (AI=0.95; TI=0.93) and RPD (AI=4.3; TI=3.6) values.

External validation. Finally, the robustness of our model was checked with an independent subset of 10 milk samples which were not included initially in the calibration set of samples. Briefly, these samples were analyzed both by GC, and by NIR, and resulting NIR prediction values were compared to values obtained by GC with a Student t-test for paired values. There were no differences between the predicted values by NIRS and the values reported by gas cromatography obtained for this set of 10 samples which were not included in the calibration model (P=0.05). Therefore, it can be concluded that the NIRS method provides significantly identical data to the reference data for individual FA, groups of FA and health indexes.

Therefore, the development of a fast, reliable method to routinely monitor milk FA individual or as FA groups (PUFA, MUFA, etc) could be applied on a larger scale in the dairy industry in order to promote farmers, and improve dairy systems regarding factors which affect milk’s FA composition. The use of NIRS as a rapid method provides an opportunity for the routine quantification of healthy milk FA such as omega-3 and CLA, with applications in the dairy industry for food labeling, and at the farm level for management of the dairy cow’s diet.