Effect of a Dairy Cow’s Feeding System on the Flavor of Raw Milk: Indoor Feeding or Grazing

The flavor of fresh, raw milk is considered to be the key to maintaining the quality of dairy products, and is very crucial in affecting a consumer’s choice. To better understand the differences in flavor of fresh milk between feeding patterns, we conducted the following study. Twelve Holstein cows reared in pure grazing mode and twelve reared intensively in medium to large farms were selected from the Xinjiang Uygur Autonomous Regions at the same time, and the flavor of their raw milk was analyzed. Aroma profiles and taste attributes were assessed by electronic nose and electronic tongue, respectively, and volatile flavor compounds were characterized and quantified by Headspace-Solid Phase Microextraction/Gas Chromatography-Mass Spectrometry. Thirteen volatile compounds were identified in the indoor feeding pattern and 12 in the grazing; most of them overlapped. W1S, W2S and W5S were the main contributing sensors of the electronic nose for the overall assessment of the aroma profile. Raw milk from grazing had more intense astringency, bitterness, sourness and richness in taste compared to indoor feeding. Different dietary conditions may contribute to a variety of aroma profiles. Oxime-, methoxy-phenyl-, octadecanoic acid, furfural and dodecanoic acid were the key volatile flavor compounds of grazing. Meanwhile, raw milk from indoor feeding patterns was unique in 2-nonanone, heptanoic acid and n-decanoic acid. All three detection techniques were valid and feasible for differentiating raw milk in both feeding patterns, and the compounds were significantly correlated with the key sensors by correlation analysis. This study is promising for the future use of metabolic sources of volatile organic compounds to track and monitor animal feeding systems.


Introduction
Sensory attributes play a key role in determining consumer acceptability for dairy products. Compared with other dairy products, such as UHT milk, yoghurt, butter or cheese, the type and concentration of aroma compounds in fresh milk are very low [1]. The organoleptic characteristics of dairy products are directly influenced by the flavor of raw milk. Fresh cow's milk has a distinctive yet subtle delicate flavor, which is easily affected. Potential factors in this study involved type of forage [2,3], patterns of feeding [4] and animal breeds [5,6]. In addition, process conditions [7][8][9][10], types of products [11], packing materials [12] and storage [13] were also key influencing factors.
For raw milk, the feeding pattern is a direct and critical element. Volatile flavor compounds derived from the diet that potentially transfer directly from forage or act as substrates eventually accumulate in milk [14]. Grazing systems are associated with increased product quality and greater global sustainability [15][16][17]. For consumers, the grazing model may signify more freedom, better health and pristine conditions. Therefore, they are increasingly interested in choosing dairy products produced from the milk of pasture-fed

E-Tongue
The taste properties of samples were assessed by SA 402B (Intelligent Sensor Technology Co., Ltd., Atsugi, Japan), which has 6 sensors that indicated 9 taste attributes, including bitterness, sourness, umami, saltiness, astringency, sweetness, aftertaste-bitterness, aftertaste-astringency and richness (aftertaste-umami). The taste attribute values were relative outputs using artificial saliva (e-tongue reference solution) as a standard, which test simulates the state of the human mouth when only saliva was present. Sensors and reference electrodes were pre-activated for at least 24 h. The activation of the sensor for sweetness and the detection of the sample were independent. Samples were appropriately diluted and filtered at room temperature. Each sample was assayed three times.

Statistical Analysis
The qualitative analysis of volatile organic compounds was carried out using MS combined with the retention index. The NIST 14 database was used for MS to identify unknown compounds. The compound's retention index was determined by measuring the retention time of C7-C40 n-alkanes. The concentration of volatile organic compounds was estimated semi-quantitatively using the internal standard. An independent samples t-test (p < 0.05) was performed using SPSS 22.0. Principal component analysis (PCA), Partial Least Squares Discriminant Analysis (PLS-DA) and orthogonal partial least squares-discriminant analysis (OPLS-DA) were performed using Metabo Analyst 4.0 (https://metaboanalyst.ca/) (accessed on 26 April 2023) and SIMCA 14.1. OmicStudio tools were used for correlation analysis (https://www.omicstudio.cn/tool) (accessed on 26 April 2023).

E-Nose Analysis
Slight changes in aroma have resulted in the discrepancy between sensors' responses [35]. According to the sensitive functional group corresponding to each sensor, the main character- istic odor composition of the sample can be inferred [17]. The response curves are presented in Figure 1. During enrichment, the initial response of the W5S sensor (red) 6 (which has a sensitivity to nitrogen oxides) was highest for both sets of samples, with grazing milk exceeding 12 and indoor feeding pattern raw milk below 2. As the enrichment time increases, the response of the W2S sensor (blue) (which has a sensitivity to alcohols, ketones and aldehydes) was highest during the final stabilization phase, with the indoor feeding group exceeding 3 and the grazing group below 3. The response of the W3S sensor (orange) was prominent (which has a sensitivity to long-chain alkanes), and the response was close to W1S.

E-Nose Analysis
Slight changes in aroma have resulted in the discrepancy between sensors' responses [35]. According to the sensitive functional group corresponding to each sensor, the main characteristic odor composition of the sample can be inferred [17]. The response curves are presented in Figure 1. During enrichment, the initial response of the W5S sensor (red) 6 (which has a sensitivity to nitrogen oxides) was highest for both sets of samples, with grazing milk exceeding 12 and indoor feeding pattern raw milk below 2. As the enrichment time increases, the response of the W2S sensor (blue) (which has a sensitivity to alcohols, ketones and aldehydes) was highest during the final stabilization phase, with the indoor feeding group exceeding 3 and the grazing group below 3. The response of the W3S sensor (orange) was prominent (which has a sensitivity to long-chain alkanes), and the response was close to W1S. The data of 110-115 s were converted into a radar map (Figure 2A). It was observed that the odor profile of raw cow's milk between grazing and indoor feeding was significantly different. At the beginning of the detection, the highest enrichment was from the W5S sensor (red). The response of the W5S sensor was over 12 for raw milk in the grazing mode and below 2 for raw milk in the housed mode. Therefore, it was indicated that the grazing raw milk had a higher nitrogen oxide. As the enrichment time increased, the W2S sensor (light blue) reached the highest response and continued to grow, eventually plateauing at a high level. These results were similar to the GC-MS test; the raw milk from the grazing pattern has a high abundance of nitrogen oxides. The response values of W1W and W2W sensors less than 1 indicated a lower function [35,36]. In addition, the response value of most sensors under grazing was low in the whole test process, except the W5S sensor. The PCA plot explained 88% of the variances ( Figure 2B). These results showed that the aroma profile was well represented by sensors of the e-nose. The data of 110-115 s were converted into a radar map (Figure 2A). It was observed that the odor profile of raw cow's milk between grazing and indoor feeding was significantly different. At the beginning of the detection, the highest enrichment was from the W5S sensor (red). The response of the W5S sensor was over 12 for raw milk in the grazing mode and below 2 for raw milk in the housed mode. Therefore, it was indicated that the grazing raw milk had a higher nitrogen oxide. As the enrichment time increased, the W2S sensor (light blue) reached the highest response and continued to grow, eventually plateauing at a high level. These results were similar to the GC-MS test; the raw milk from the grazing pattern has a high abundance of nitrogen oxides. The response values of W1W and W2W sensors less than 1 indicated a lower function [35,36]. In addition, the response value of most sensors under grazing was low in the whole test process, except the W5S sensor. The PCA plot explained 88% of the variances ( Figure 2B). These results showed that the aroma profile was well represented by sensors of the e-nose. The score scatter plot showed that samples from Holstein cattle reared intensively in an indoor feeding pattern were more consistent and clustered closely, while the distribution of raw milk in grazing were more dispersed ( Figure 3A). The biplot ( Figure 3B), as an overlay of the score and load plots, was also straightforward in terms of seeing the correlation with the sensors. The samples in group A were grazing feed, and the main sensors  The score scatter plot showed that samples from Holstein cattle reared intensively in an indoor feeding pattern were more consistent and clustered closely, while the distribution of raw milk in grazing were more dispersed ( Figure 3A). The biplot ( Figure 3B), as an overlay of the score and load plots, was also straightforward in terms of seeing the correlation with the sensors. The samples in group A were grazing feed, and the main sensors associated were the W2W sensor, W1W sensor and W5S sensor, while the W6S sensor was the closest to the housed samples, indicating that for housed cows, their raw milk was the most sensitive to this sensor. The score scatter plot showed that samples from Holstein cattle reared intensively in an indoor feeding pattern were more consistent and clustered closely, while the distribution of raw milk in grazing were more dispersed ( Figure 3A). The biplot ( Figure 3B), as an overlay of the score and load plots, was also straightforward in terms of seeing the correlation with the sensors. The samples in group A were grazing feed, and the main sensors associated were the W2W sensor, W1W sensor and W5S sensor, while the W6S sensor was the closest to the housed samples, indicating that for housed cows, their raw milk was the most sensitive to this sensor.

E-Tongue Analysis
The reference solution consists of KCL and tartaric acid, so the tasteless point was −13 for sour tastes and −6 for salty tastes, and it was used as a benchmark. When the value is below, it means that the sample has no taste. An independent sample t-test was carried out for each taste attribute value to determine significant differences, whereas a p < 0.05 value was regarded as significant. Two groups showed significant correlations in all flavor attributes, except for sourness and umami ( Table 2). Potentiometric difference between each sensitive electrode and the Ag/AgCl reference electrode in the equilibrium state was recorded as the response signal. The potential change of the sample was recorded and transformed into taste data by these sensors using set conversion factors. The results of the e-tongue can be observed on the load diagram ( Figure 4A). The two types of feeding practices had a more pronounced effect on the taste of raw milk. Two groups were clustered into one category and clearly differentiated, which was consistent with the electronic nose. The PLS-DA model was used for the discriminant analysis of taste attributes. The contribution rate of the two principal components was 82.7%. Therefore, it was indicated that the model has good discriminative and predictive ability. Raw milk samples were clearly separated into different areas with different feeding patterns. Through the PLS-DA model, four taste attributes with VIP > 1 were screened, which indicated that bitterness, astringency, aftertaste-B and richness had great differences in raw milk with different feeding patterns ( Figure 4B). The distance represented the degree of contribution in the biplot plots ( Figure 4C). From the distribution of taste attributes of the samples in the figure, the taste of raw milk in grazing mode was mainly concentrated in richness, astringency, bitterness, sweetness and sourness. In contrast, the taste of raw milk in the indoor feeding mode was mainly concentrated in salty, umami and bitter aftertaste and aftertaste astringency. The results of the e-tongue can be observed on the load diagram ( Figure 4A). The two types of feeding practices had a more pronounced effect on the taste of raw milk. Two groups were clustered into one category and clearly differentiated, which was consistent with the electronic nose. The PLS-DA model was used for the discriminant analysis of taste attributes. The contribution rate of the two principal components was 82.7%. Therefore, it was indicated that the model has good discriminative and predictive ability. Raw milk samples were clearly separated into different areas with different feeding patterns.
Through the PLS-DA model, four taste attributes with VIP > 1 were screened, which indicated that bitterness, astringency, aftertaste-B and richness had great differences in raw milk with different feeding patterns ( Figure 4B). The distance represented the degree of contribution in the biplot plots ( Figure 4C). From the distribution of taste attributes of the samples in the figure, the taste of raw milk in grazing mode was mainly concentrated in richness, astringency, bitterness, sweetness and sourness. In contrast, the taste of raw milk in the indoor feeding mode was mainly concentrated in salty, umami and bitter aftertaste and aftertaste astringency.
OPLS-DA can filter out the between-group differences more accurately. The direction of the horizontal coordinate showed the differences in groups. The within-group differences have been observed on the vertical coordinate. In terms of the distribution of samples ( Figure 5A), within-group variation was more pronounced for raw milk collected in the grazing mode. The arrangement of the samples was more dispersed, indicating that their diet had some inconsistency in the free-feeding state. In contrast to raw milk obtained in the indoor feeding pattern, which has a more concentrated distribution, raw milk obtained through an outdoor feeding pattern may be similar in terms of volatile flavor and behave more consistently. Biplot overlapped the score and loading plots ( Figure 5B). The distribution distances of the compounds and samples reflected their correlation, including 3,4-dimethylpentanol, acetic acid, furfural, hexanol, 2-furanmethanol, methoxyphenyl oxime, undecanoic acid and dodecanoic acid. The closer proximity of these compounds to the grazing samples may represent that they are unique to grazing as distinct components. While for raw milk samples obtained in the indoor feeding pattern, the abundance of 1-octen-3-ol, heptanoic acid, nonanoic acid, n-decanoic acid, 9-decenoic acid, butanoic acid, octanoic acid, hexanoic acid and 2-nonanone were likely to be higher.
The correlation between volatile compounds and main intelligent sensory signals can be seen in Figure 6. Similar studies have also been conducted in pine mushrooms [35], cocoa bean shells [37] and pine nuts. Volatile flavor compounds were plotted on the Y axis, while the X axis had the main sensors of the electronic nose and key taste attributes of the electronic tongue. Most compounds were significantly correlated with the key sensors from the e-nose and the e-tongue. The contributions of both W1S and W2S sensors to sample differentiation were consistent, and they were negatively correlated to dodecanoic acid, Oxime-, methoxy-phenyl-, furfural, octadecanoic acid and nonanoic. By contrast, the response of the W5S sensor was positively correlated with dodecanoic acid and octadecanoic acid, but negatively correlated with nonanoic acid. The correlation results between the main taste properties of electronic tongue and volatiles were as follows: bitterness and astringency were positively correlated with dodecanoic acid, 1-hexanol, 2-pentanol and 3,4-dimethylpentanol; all three taste properties were negatively correlated with 1-octen-3-ol. These results showed that intelligent sensory devices were similar to the GC-MS. Moreover, it indicated that the e-nose could discriminate grazing and indoor feeding pattern by responding specifically to volatile compounds from raw milk. The correlation of intelligent sensory technology and the main volatile flavor compounds could further distinguish the samples and explain the difference between these methods. It could be critical for identifying or tracing raw milk from different feeding patterns. The correlation between volatile compounds and main intelligent sensory signals can be seen in Figure 6. Similar studies have also been conducted in pine mushrooms [35], cocoa bean shells [37] and pine nuts. Volatile flavor compounds were plotted on the Y axis, while the X axis had the main sensors of the electronic nose and key taste attributes of the electronic tongue. Most compounds were significantly correlated with the key sensors from the e-nose and the e-tongue. The contributions of both W1S and W2S sensors to sample differentiation were consistent, and they were negatively correlated to dodecanoic acid, Oxime-, methoxy-phenyl-, furfural, octadecanoic acid and nonanoic. By contrast, the response of the W5S sensor was positively correlated with dodecanoic acid and octadecanoic acid, but negatively correlated with nonanoic acid. The correlation results between the main taste properties of electronic tongue and volatiles were as follows: bitterness and astringency were positively correlated with dodecanoic acid, 1-hexanol, 2-pentanol and 3,4-dimethylpentanol; all three taste properties were negatively correlated with 1-octen-3-ol. These results showed that intelligent sensory devices were similar to the GC-MS. Moreover, it indicated that the e-nose could discriminate grazing and indoor feeding pattern by responding specifically to volatile compounds from raw milk. The correlation of intelligent sensory technology and the main volatile flavor compounds could further distinguish the samples and explain the difference between these methods. It could be critical for identifying or tracing raw milk from different feeding patterns.

Discussion
The purpose of this study was to explain flavor differences under different feeding patterns. The odor summarized profile was assessed by the electronic nose. The taste scor was evaluated by the electronic tongue, and volatile flavor compounds were detected by HS-SPME/GC-MS. The results demonstrate that the ability of volatile profiling and intel ligence sensory techniques can distinguish raw milk produced from grazing versus indoo feeding systems.

Discussion
The purpose of this study was to explain flavor differences under different feeding patterns. The odor summarized profile was assessed by the electronic nose. The taste score was evaluated by the electronic tongue, and volatile flavor compounds were detected by HS-SPME/GC-MS. The results demonstrate that the ability of volatile profiling and intelligence sensory techniques can distinguish raw milk produced from grazing versus indoor feeding systems.
The response of the W5S sensor in the grazing pattern far exceeded that in the pasturefed mode. According to the sensitivity of the sensors, we hypothesized that the grazing pattern has a higher abundance of nitrogen oxide in raw milk. GC-MS results confirmed that oxime, methoxy-phenyl-was the characteristic component detected in the grazing group. Oxime-, methoxy-phenyl-was once detected in yogurt, ultra-pasteurized milk and cheese [38,39]. It also has been reported to be a characteristic component of mushrooms and mountain plants [35]. Furthermore, the W5S sensor response was positively correlated with the levels of dodecanoic and octadecanoic acids. Consistent with previous studies that compare with a confinement system, farmers' bulk milk samples of grazing pattern contained more octadecanoic acid and dodecanoic acid [40].
2-ethyl-1-hexanol has a "chemical/cleaning agent" aroma, which was detected in raw milk over refrigerated storage time and might be due to the packaging and refrigeration environment [41]. 1-Octen-3-ol was the common flavor active substance in milk and dairy products. The odor has a low threshold, which was often described as mushroom and grass. GC-MS results showed that the abundance of 1-octen-3-ol was positively correlated with the response of the W2S sensor, which may be due to the high content of silage in the forage on the farm.
Aldehydes could provide significant aromas due to their lower odor threshold, either pleasant or rancid [38]. W1S was thought to be more associated with volatiles containing methyl. The response values of W2S and W1S sensors were negatively correlated for 2furanmethanol and furfural. 2-Furanmethanol and furfural were typical Maillard reaction products and have been observed in formula and heated milk, introducing caramel flavor and richness [42]. The high abundance of aldehydes may be the direct or indirect cause of sweetness in raw milk from grazing feeding patterns.
By means of an algorithm inherent to the electronic tongue, we obtained scores for the taste attributes in raw milk. The VIP scores, which were based on the PLS-DA model, indicated that bitterness, astringency, bitter aftertaste and richness represented the main taste differences. Compared to grass silage, maize gave a higher sweet odor and less boiled milk, saltness and metallic flavor [43]. Consistent with our research, Holstein cows on concentrated pasture, which have a high proportion of silage in their feed, scored higher for freshness and saltiness in raw milk. On the other hand, cows in the grazing mode, where the forage was dominated by maize straw, had a more prominent raw milk performance in terms of bitterness, richness and sweetness.
In our research, most of the volatile fatty acids are high in indoor feeding, with little or no detection in raw milk from the grazing pattern. In previous studies, large amounts of ethyl hexanoate were detected in cheese made from milk obtained from grazing cows, but no acetic acid was detected [21]. In this study, the bitterness and astringency of raw milk were positively correlated with the abundance of acetic acid. Hexanoic acid had been detected in raw milk from pasture, and ethyl hexanoate was detected from grazed feeding patterns, with the ethyl ester having a low odor threshold and possibly contributing a significant odor. However, the correlation between ethyl hexanoate and taste was not significant. The same conclusion has occurred in cheese [44]. Acids may result from lipolysis, lactose and degradation of amino acids, which were responsible for the rancid and bitter taste in milk [38,39]. Acid was the most dominant volatile substance category in milk and was thought to be associated with diet-related levels [30]. In general, long-chain fatty acids played a small role in flavor based on their high perceived threshold [45].
Ketones originated from the degradation of amino acids, which was an important component of milk products. High levels of ketones may have an adverse effect on flavor. 2-Nonanone was one of the characteristic components in the indoor feeding pattern. Silage has a limited effect on aldehydes, whereas ketones were more influenced by diet and ripening [38,46]. Based on the correlation between volatile compounds and main intelligent sensory signals, a combination of methods was used to judge the organoleptic quality of raw milk from different feeding patterns, and the expected results were achieved. Different dietary conditions may be the key factor for odor profiles. Some aromatics, such as oxime-, methoxyphenyl-, octadecanoic acid, furfural and dodecanoic acid, were the critical volatile compounds in raw milk of grazing. Meanwhile, raw milk of indoor feeding pattern was unique in 2nonanone, heptanoic acid and n-decanoic acid. All three detection techniques were valid and feasible for differentiating raw milk in both feeding patterns, and the compounds were significantly correlated with the key sensors by Person correlation analysis. This study will contribute to tracking and monitoring animal feeding systems by metabolic sources of VOC in the future.
Experimental results showed that there was large variability in the odor and taste of raw milk under different feeding patterns. Additionally, the milk yield of cows was lower from grazing than from indoor feeding. Grazing-fed cows might use body reserves for milk production, which had an impact on milk composition and further led to differences in milk taste. We believed that there were seasonal variations in the diet of grazing animals. Therefore, milk flavor should be continuously monitored in subsequent studies to compensate for the limitations imposed by seasonal differences.

Conclusions
These findings suggest that the volatile characteristics of raw milk from indoor feeding and grazing patterns differ significantly. Dietary conditions of cows may be the crucial element for different aroma profiles. Grazing milk achieved a significantly higher intensity for bitterness, richness and sweetness. W1S, W2S and W5S sensors played a major role in differentiating feeding systems. Some aromatic components, which included oxime-, methoxy-phenyl-, octadecanoic acid, furfural and dodecanoic acid, were the key volatile compounds in raw milk of grazing. Meanwhile, the raw milk under the indoor feeding pattern was unique in 2-nonanone, heptanoic acid and n-Decanoic acid. The results of the volatile compounds, odor profile and taste assessment were consistent in distinguishing between raw milk from two different feeding patterns. It will provide a comprehensive and accurate analysis method to study the effect of feeding patterns on raw milk flavor, which could be important for tracing or identifying raw milk under different feeding patterns.