Comparative evaluation of yak ghee and crossbred yak ghee: lipid composition, volatile flavour constituents, and microbial community dynamics

Abstract

Yak ghee (YG) and crossbred yak ghee (YD), traditional fermented dairy products central to the diet of pastoralists in China's Tibetan regions, are distinguished by their lipid composition, volatile profiles, and microbial communities. Nevertheless, comprehensive assessments of quality variations in yak ghee sourced from distinct milk origins are scarce. This research systematically contrasted YG and YD concerning their lipid constituents, volatile compounds, and microbial populations to delineate the factors contributing to quality heterogeneity. Lipidomics analysis revealed 43 distinct lipids, with seven glycerides identified as potential biomarkers. A total of 22 volatile compounds were detected, and 3-hydroxy-2-butanone was identified as a characteristic constituent. More than 20 microbial taxa were identified, with Lactobacillus and Lactococcus designated as key differential genera. The observed quality disparities were found to correlate with lipid degradation patterns and microbial metabolism, thereby establishing a foundation for breed identification and quality assurance.

Highlights

• •Comparative lipidomics revealed 43 differential lipids between yak and crossbred yak ghee.
• •Seven glyceride-dominant lipids serve as potential biomarkers for discrimination.
• •3-Hydroxy-2-butanone was identified as a key aroma compound in Gannan ghee.
• •Microbial community shifts were associated with lipid degradation and quality differences.

1. Introduction

The yak is a large mammal primarily found in the Tibetan pastoral regions of the Qinghai-Tibet Plateau in China. Its unique physiological environment and feeding practices confer a high-quality nutritional profile to yak milk, with notably elevated levels of fat, protein, and minerals, which has earned it the title of “natural concentrated milk” (Fan, Wanapat, Yan, & Hou, 2020; Kalwar et al., 2023). This characteristic also makes it an ideal raw milk source for developing dairy products in the Tibetan region. Ghee, a traditional fermented dairy product derived from yak milk, is produced through preheating, defatting, fermentation, and kneading, yielding a nutrient-rich product (Elbarbary, Jin, & Wang, 2024; Jiang et al., 2024; Liang et al., 2025).

Ghee contains high levels of conjugated linoleic acid, omega-3 fatty acids, and other mono- or polyunsaturated functional lipids, making it a primary dietary nutrient source for herders in the Tibetan region (Kalwar et al., 2023; Li et al., 2011; Mamet et al., 2023; Zhang et al., 2020). Moreover, the health benefits of ghee in promoting human well-being have been extensively investigated (Kalwar et al., 2023; Singh, Tyagi, Singh, & Pandey, 2020). For instance, it has been shown to enhance digestion, modulate immune function, and slow ageing processes (Kataria & Singh, 2024; Yang et al., 2024). Therefore, since lipids are the primary constituents of ghee, establishing characteristic lipid profiles and exploring their potential bioactivities are essential for the development of ghee and its derivatives.

Due to their unique genetic background, semi-wild free-grazing system, and seasonal forage availability (Marquardt et al., 2018), yaks show relatively low milk yields (230–400 kg annually, approximately 10% of that of conventional dairy cows). Thus, the increasing market demand for yak dairy products contrasts sharply with the limited volume available for commercialisation. More recently, crossbreeding between yaks and cattle breeds such as Yellow cattle and Jersey has produced hybrids with high-altitude adaptability comparable to purebred yaks. These hybrids exhibit higher milk yields even under low stocking density at lower altitudes (Marquardt et al., 2018) and demonstrate superior milk yield and fatty acid content than purebred yaks (Chen et al., 2025). Moreover, maternal yak milk traits are partially retained yet modified in hybrid offspring. The milk obtained from these hybrids has partially alleviated the supply shortages caused by limited yak milk production (Chen et al., 2025) and has become an important substitute for traditional yak milk in products such as Qula and ghee.

However, ghee quality and flavour are influenced not only by the milk source (Kataria & Singh, 2025) but also by lipid composition and microbial metabolism. Lipidomic analyses have shown that yak milk contains higher levels of unsaturated fatty acids than Simmental milk (Liu et al., 2025). Although Lactococcus, Streptococcus, and Lactobacillus are known to participate in ghee fermentation, they may release extracellular enzymes (e.g., lipases) during storage and transport, accelerating lipid oxidation and hydrolysis and thereby inducing discolouration, nutritional deterioration, and rancid flavours (Liu et al., 2022). The existing studies on this topic have mainly focused on physicochemical (Tian et al., 2023) and oxidative properties (Abid et al., 2017), whereas the combined effects of lipid and microbial diversity on flavour differences between yak ghee and crossbred yak ghee remain unclear.

Lipidomics, a high-throughput analytical approach for comprehensive characterisation of complex lipidomes, has allowed detailed identification of lipid species and structural features (Liu & Rochfort, 2023). It has been widely employed for targeted quantitative analyses of milk fat from different species and lactation stages, including human milk (Zhang et al., 2021), bovine milk (Li et al., 2020; Li et al., 2020), and ghee (Cong et al., 2025; Feng et al., 2025). These studies have demonstrated the reliability of lipidomics in distinguishing ghee derived from various milk sources. Therefore, we hypothesise that integrating gas chromatography (GC)-mass spectrometry (MS), lipidomics, and high-throughput sequencing can elucidate the heterogeneity in ghee from different milk sources and facilitate biomarker identification and quality regulation strategies.

The objectives of this study were as follows: (1) comprehensively identify and quantify the lipid compositions of yak and crossbred yak ghee; (2) compare lipid profiles using multivariate analysis to screen potential lipid biomarkers; (3) characterise key volatile compounds and dominant genera using GC–MS and high-throughput sequencing; and (4) integrate microbial, volatile, and lipidomic data to clarify their interrelationships. These findings will complement existing quality studies and provide insights for process optimisation of traditional ghee.

2. Materials and methods

2.1. Ghee sample collection

Yak and crossbred yak ghee samples were collected in July 2024 (summer pasture) from a single herder's grazing site in Luqu County, Gannan Tibetan Autonomous Prefecture, Gansu Province, China. To ensure biological representativeness, each type of ghee was derived from pooled milk obtained from the same herd. Milk sampling procedures were in accordance with those described by Guo et al. (2026). After collection, milk samples from each source were pooled separately. The herders then prepared ghee from the collectively collected milk samples under identical environmental conditions. For laboratory analysis, aliquots of the samples were transferred into cryovials, immediately placed in liquid nitrogen for transport to the laboratory, and subsequently stored at −80 °C until further analysis.

The inclusion criteria for all selected animals were as follows: healthy female yaks and crossbred yaks aged 5–6 years with a body weight of 210 ± 20 kg and reared under identical natural grazing conditions without supplemental feeding on the same pasture. All sampled animals were managed under uniform conditions to minimise management-related effects on milk composition.

2.2. Chemicals and reagents

Liquid chromatography (LC)-MS grade isopropyl alcohol and methanol were purchased from Fisher Scientific (Loughborough, UK). Chloroform was obtained from Sinopharm (Shanghai, China). 2-Octanol, sodium chloride, and other reagents were purchased from Shanghai Macklin Biochemical Company.

2.3. Lipidomics analysis

2.3.1. Lipid extraction

First, 0.05 g of ghee was added to a centrifuge tube. After adding 750 μL of chloroform-methanol solvent (2:1, v/v), the mixture was vortexed for 30 s, homogenised, and incubated on ice for 40 min. Next, 190 μL of ultrapure water was added, and the mixture was vortexed thoroughly and allowed to phase-separate for 10 min, which was followed by centrifugation at ambient temperature. Subsequently, 300 μL of the organic phase was transferred to a new centrifuge tube, and 500 μL of chloroform-methanol solvent was added to perform a secondary extraction. The organic phases were combined and evaporated under nitrogen or vacuum. The lipid extract was reconstituted in 200 μL of isopropanol and filtered through a 0.22-μm membrane filter to prepare the sample for LC-MS analysis.

2.3.2. Lipid analysis

Chromatographic parameters: Gradient elution was conducted on an ACQUITY UPLC® BEH C18 column (1.7 μm, 2.1 × 100 mm) maintained at 50 °C, with an autosampler temperature of 8 °C and a flow rate of 0.25 mL/min. Mobile phase A2 was acetonitrile-water (60:40, supplemented with 0.1% formic acid and 10 mM ammonium formate), while mobile phase B2 was isopropanol-acetonitrile (90:10, containing 0.1% formic acid and 10 mM ammonium formate).

MS analyses were performed using an electrospray ionisation (ESI) source, acquiring data in both positive and negative ion modes. Positive ion spray voltage was set at 3.50 kV, while negative ion spray voltage was set at 2.50 kV. The sheath gas flow was maintained at 30 arb, assist gas at 10 arb, and capillary temperature at 325 °C. The full scan resolution was configured at 35000, covering a mass range of 150–2000 m/z. Secondary fragmentation was performed by higher-energy collisional dissociation (HCD) with a collision energy of 30 eV, and dynamic exclusion was enabled to prevent redundant tandem MS (MS/MS) signals.

2.3.3. Data preprocessing

Using LipidSearch (V4.2.28), raw MS data (* .raw) were analysed using the parameters RT tolerance = 0.25 and m-Score threshold = 3 to identify lipid species on the basis of the m/z measurements and their retention time (RT) and peak intensity values. Quantitative normalisation was performed using total ion current normalisation. The annotated lipids in the dataset were classified on the basis of the composition of the fatty acid chain and the modification of the functional groups.

2.4. Volatile organic compound analysis

2.4.1. Sample preparation

In accordance with a previously described method (Liang et al., 2025), approximately 6.00 g of ghee was weighed and transferred to a 20-mL headspace vial containing 1.5 g of sodium chloride and 100 μL of 2-octanol internal standard solution (0.001 μg/μL). The vial was sealed with a screw cap. The headspace vial was incubated at 60 °C in a water bath for 30 min. Subsequently, using a manual headspace (HS)-solid-phase microextraction (SPME) device, the extraction fibre was inserted into the vial's headspace, positioning the fibre tip approximately 5 mm above the sample liquid surface. Analyte adsorption was allowed for 30 min. After adsorption, the fibre was promptly inserted into the GC–MS injection port; the analytes were desorbed at 250 °C for 5 min; and chromatographic analysis was performed.

2.4.2. GC–MS conditions

GC was performed using the HP-INNOWAX column (60.0 m × 0.25 mm, 0.5 μm) at an injection temperature of 200 °C. The temperature programme was as follows: initial temperature, 60 °C; hold for 1 min; ramp at 2 °C/min to 230 °C; final ramp at 20 °C/min to 245 °C; and hold for 5 min. In the splitless injection mode, the flow rate of the carrier gas (helium) was maintained constant.

The conditions for MS were as follows: ion source and ion monitor temperature, 230 °C and 200 °C, respectively; electron energy, 70 eV; scan mode, full scan; mass range, 50–500 m/z; solvent delay, 3 min.

2.4.3. Data preprocessing

In accordance with a previously described methodology (Zhang et al., 2024), qualitative analysis and matching searches for volatile organic compounds (VOCs) were performed using the NIST 11 spectral library, and retention indices (RIs) were calculated. Using 2-octanol as the internal standard, the content of each component in the ghee sample was determined using the internal standard method. The semi-quantitative analysis formula for each component is as follows:

$$m_i=\frac{m_0\times V_0\times A_i}{V_i\times A_0}$$

In the equation, m_i and m₀ denote the mass concentrations (μg/L and μg/μL) of the unknown substance and internal standard, respectively; A_i and A₀ denote their respective peak areas; while V_i (L) and V₀ (μL) represent the volumes of the ghee sample and internal standard solution used in extraction.

2.5. High-throughput sequencing

Microbial genomic DNA was extracted using the DNeasy PowerSoil Kit (Qiagen, Germany) in accordance with the manufacturer's instructions. DNA concentration and purity were determined using a NanoDrop spectrophotometer (Thermo Fisher Scientific, USA), and integrity was verified by 1% agarose gel electrophoresis. DNA samples were diluted to a uniform concentration before amplification.

The V3-V4 hypervariable region of the bacterial 16S rRNA gene was amplified using the universal primers 338F (5′-ACTCCTACGGGAGGCAGCA-3′) and 806R (5′-GGACTACHVGGGTWTCTAAT-3′). PCR amplification was performed in a 25-μL reaction mixture containing 12.5 μL of 2× PCR Master Mix, 1 μL of each primer (10 μM), 2 μL of template DNA, and nuclease-free water to volume. Negative controls without template DNA were included to monitor contamination.

The thermal cycling conditions were as follows: initial denaturation at 98 °C for 5 min; 30 cycles of denaturation at 98 °C for 30 s, annealing at 55 °C for 45 s, and extension at 72 °C for 45 s; and final extension at 72 °C for 5 min and holding at 12 °C. PCR products were verified by 2% agarose gel electrophoresis, and target bands were purified using an Axygen Gel Extraction Kit (Axygen, USA).

Purified amplicons were used to construct sequencing libraries using the Illumina TruSeq DNA kit, and library quality was assessed using an Agilent Bioanalyzer. Libraries were quantified using the QuantiFluor dsDNA system (Promega, USA) and normalised to 2 nM. After pooling, the libraries were denatured with 0.1 mol/L NaOH and sequenced on an Illumina NovaSeq 6000 platform (2 × 250-bp paired-end).

2.6. Statistical analysis

The lipidomics analysis included six samples; high-throughput sequencing involved four samples; and GC–MS analysis was performed on three samples. The results are presented as mean ± standard error. Orthogonal partial least-squares discriminant analysis (OPLS-DA), along with community species composition and abundance metrics, were subjected to multivariate statistical analyses utilizing the R package ropls. Subsequently, linear discriminant analysis (LDA) was performed on the BioDeep cloud platform (https://www.biodeep.cn/login) to ascertain differential species influence (LDA score), and a LefSe analysis plot was generated. Screening for differential compounds was conducted based on a p-value <0.05 and a variable importance in projection (VIP) score > 1.0. Lipid biomarkers were identified utilizing the same BioDeep cloud platform. Beta-diversity analysis was performed with QIIME software, followed by calculation of weighted Bray–Curtis similarity to assess differences in community composition among samples. SIMCA 14.1 software was used to perform OPLS-DA analysis of VOCs to calculate VIP values; differential aroma components were screened on the basis of p < 0.05 and VIP ≥ 1. Statistical analyses, including one-way analysis of variance and Duncan's multiple-range test (p < 0.05), were conducted using Origin 2021 (OriginLab, Microsoft Corporation, USA) and SPSS (version 26, USA). Spearman's correlation coefficients were used to analyse correlations among differential lipids, key microorganisms, and characteristic flavours (p < 0.05).

3. Results and discussion

3.1. Lipid analysis of YG and YD samples

A comparison of the retention times and base peak chromatograms (Fig. S1) in both positive and negative ion modes between the two sample groups revealed a high degree of overlap, confirming stable detection performance of the MS system. These results validated the reliability of the analytical method used in this study. As illustrated in Fig. 1A, a total of 2243 lipids were identified in the YG and YD samples, and were categorised into 40 lipid subclasses. As shown in Figs. 1B and C, all lipids were further classified into six major classes: neutral lipids (NL), which accounted for 70% of all lipids, phospholipids (PL; 16.2%), sphingolipids (SL; 8.6%), derivatised lipids (DL; 2.8%), glycolipids (GGL; 2.2%), and fatty acylcarnitines (FA; 0.3%). These results indicated that NL are the predominant lipid component of ghee fat.

Fig. 1.

Lipid composition analysis of YG and YD. (A) Distribution map of lipid components. (B) Categories of lipid composition. (C) Percentage composition of lipid subclasses. (D) Content of significantly different lipid subclasses.

Lipids present at concentrations exceeding 1% included triglycerides (TG; 1361 types), diglycerides (DG; 160 types), phosphatidylcholine (PC; 116 types), phosphatidylethanolamine (PE; 106 types), hexosylceramides (Hex1Cers; 56 types), ceramides (Cer; 54 types), sphingomyelin (SM; 46 types), phosphatidylserine (PS; 46 types), phosphatidylinositol (PI; 37 types), monogalactosyl diglyceride (MGDG; 34 types), yeast sterol ester (ZyE; 34 types), hexosylceramide (Hex2Cer; 33 types), dimethylphosphatidic acid (BisMePA; 22 types), cardiolipin (CL; 21 types), dimethylphosphatidylethanolamine (dMePE; 17 types), and methylphosphatidylcholine (MePC; 17 types). TG, DG, and PE were the primary determinants of the lipid richness of ghee. TG and DG were the two most abundant lipid subclasses in both ghee fat samples, consistent with the findings reported by Cong et al. (2025) and Li et al. (2024) and Li, He, Zheng, Shi, and Huang (2024).

TG plays a crucial role in fatty acid transport and serves as the primary form of energy storage in animals (Chu, Wang, & Xie, 2025). The TG content in yak milk exceeds 90%, which is higher than that in other species; however, the relative proportions of other lipid subclasses vary depending on the source species. This variation is likely a key factor contributing to the differences in lipid profiles between YG and YD (Hewelt-Belka, Garwolinska, Mlynarczyk, & Kot-Wasik, 2020; Li, Li, Kang, et al., 2020; Li, Li, Wu, et al., 2020; Wang et al., 2023; Wang et al., 2023). For instance, Feng et al. (2025) found that the relative levels of LPC and LPE in Jersey ghee were higher than those in Holstein ghee, indicating differences in functional lipids across ghee from different dairy sources.

As shown in Fig. 1D, further analysis revealed significant differences (p < 0.05) in the relative levels of BisMePA, CmE, LPC, MGDG, and PE and highly significant differences (p < 0.01) in the relative levels of AcHexCmE, AcHexSiE, CL, LPE, and PC in YG. In contrast, the relative levels of LPC and LPE in YD were significantly higher than those in YG. In this study, the relative levels of PC, PE, and CL in YG were all significantly higher than those in YD. PL undergoes autooxidation due to its structural linkage with unsaturated fatty acids, while PE exhibits a faster oxidation rate than PC due to its higher capacity for chelating pro-oxidant metal ions. Thus, YG may be more susceptible to rancidity than YD.

3.2. Multivariate statistical analysis of lipids

Multivariate statistical analysis, including principal component analysis (PCA), revealed substantial similarity in lipidomic profiles and concentrations between YG and YD, with all samples within the 95% confidence ellipse. However, the distinct separation along the first principal component (PC1) highlighted the notable differences in lipid composition between YG and YD. Feng et al. (2025) also confirmed the differences in lipid composition between Holstein and Jersey ghee by using this method. Parallel samples within each group showed varying degrees of separation along the PC2 axis, indicating strong reproducibility and representativeness.

Subsequent OPLS-DA analysis indicated that lipids such as TG, SQDG, Cer, and PE made more significant contributions to classification. Fig. 2B illustrates clear regional heterogeneity in YG and YD samples, reflecting compositional divergence. The S-plot (Fig. 2C) highlights key lipid biomarkers, with the variables at the extremes indicating significant differential lipids for targeted metabolomic screening. Additionally, we validated the predictive accuracy and goodness-of-fit of the OPLS-DA model through seven cycles of cross-validation and 200 response-order tests. As shown in Fig. 2D, the R² and Q² values (0.98; 0.11) confirmed the reliability of the analytical model without overfitting. Li, Guo, et al. (2024) and Li, He, et al. (2024) and Yang et al. (2024) subsequently adopted this methodology in their studies on lipid differences in yak colostrum/mature milk and the lipid characteristics of ghee from different altitudes, further affirming the scientific validity of this model.

Fig. 2.

(A) Principal component analysis of YG and YD. (B) OPLS-DA analysis of YG and YD. (C) OPLS-DA S-PLOT diagram of YG and YD. (D) Permutation test plot of YG and YD. (E) Heatmap of differentially expressed lipid subclasses in YG and YD. (F) Heatmap of differentially expressed lipid structures. (G) Functional enrichment bar chart of differentially expressed lipids.

3.3. Screening of differentially expressed lipids between YG and YD

As shown in Fig. S2, a total of 667 lipids exhibited significant differences between YG and YD. A total of 32 differentially expressed lipid classes were identified across 40 lipid subclasses. Fig. 2E illustrates the significant differences in lipid subclass content between YG and YD. Specifically, BiotinyPE, PG, Cer, PI, dMePE, Hex2Cer, LdMePE, LPE, LPC, and SQDG were enriched in YD, while the levels of CL, SM, PE, DG, AcHexSiE, PC, StE, MGDG, PIP, AcHexCmE, BisMePA, and PS were higher in YG. Yang et al. (2024) also found a higher proportion of PL in yak ghee. These components provide essential lipid nutrition for Tibetan herders in high-altitude regions and help protect against stress-induced damage from the plateau environment (Liu et al., 2022). In comparison with the findings reported by Feng et al. (2025), TG, DG, PE, and PC were common lipids across different milk sources of ghee, while variations in other lipid subclasses may have been influenced by altitude or grazing cycles (Liu et al., 2022). For instance, secondary metabolites in pasture grasses (e.g., polyphenols, terpenoids) have been shown to affect milk fat composition (Cui et al., 2016).

In Fig. 2F, Cer, DG, PC, PE, PS, SM, and TG showed significant interspecies variability between YG and YD, and this variability was primarily distinguished by acyl chain unsaturation levels ranging from zero to ten double bonds, with most variations occurring between zero and five. Notably, the degree of unsaturation is correlated with lipid metabolic regulation: in YD, lipids (DLs) with more than five double bonds were generally downregulated across subclasses, except for Cer and PS, which were upregulated, while DLs with fewer than five double bonds were predominantly upregulated. YD showed an increase in ceramide-upregulated DLs, while DG, TG, PC, and PE exhibited balanced regulation patterns.

These results suggest that PL and NL are key discriminators in lipidomic profiles between the two ghee phenotypes. Additionally, the results highlighted the predominance of medium- and long-chain fatty acids (including both saturated and unsaturated fatty acids), with a small proportion of short-chain fatty acids (C < 6) mainly from the AcHexSiE, DG, and some PC and TG classes. Oxidative degradation of unsaturated fatty acids is the main cause of lipid rancidity (Li et al., 2025; Li et al., 2025). Lipid peroxidation occurs mainly through autooxidation and enzymatic mechanisms facilitated by endogenous and microbial lipases.

3.4. Structural-functional characteristics of differentially expressed lipids between YG and YD

During fermentation, β-oxidation of unsaturated fatty acids generates acetyl-coenzyme A (CoA) for the tricarboxylic acid (TCA) cycle, whereas saturated fatty acids undergo β-oxidation to form β-keto acids, which decarboxylate into odd-chain methyl ketones and are subsequently reduced to secondary alcohols by reductases (Feng et al., 2021). The reactive oxygen species generated by lipid oxidation decrease the antioxidant capacity of clarified ghee during storage (Shi et al., 2022). Differential lipid functional enrichment analysis using the LION lipid ontology database identified significant pathways (p < 0.05) from differentially expressed lipid modules detected by Leiden community detection; these pathways are visualised with log10|FDFq| values in Fig. 2G. The top five enriched pathways, namely, plasma membrane, ceramides, sphingomyelin (SP), N-sphingomyelin (ceramide) (SP0201), and lipid signalling, were associated with membrane structure integrity, cellular signalling, and lipid-mediated flavour precursors. This was also observed in the study by Chu et al. (2025). Enrichment was most pronounced for fatty acids with 13 carbon atoms, followed by those with 16–18 carbon atoms, and included polyunsaturated fatty acids, highlighting their potential contributions to ghee flavour development and oxidative stability.

3.5. Screening of lipid markers between YG and YD groups

Building on previous research, we further examined group-specific differential lipid molecules (SDLs) on the basis of the screening criteria (VIP > 1, p < 0.05, |fold change (FC)| > 1.8) (Li, Guo, et al., 2024; Li, He, et al., 2024). As shown in Fig. 3A, 43 SDLs were selected, with TG (15:0–10:4–18:1), TG (10:0–11:4–14:0), PE (18:1–22:6), and TG (4:0–10:0–22:6) being significantly upregulated in YG. Conversely, TG (4:0–17:0–18:2), PE (32:2e), PE (14:1e-18:2e), PE (14:1e-18:3e), and PE (32:3e) were significantly upregulated in YD. Thus, PE and TG lipids are key SDLs distinguishing YG from YD. These lipids are primarily biosynthesised through pathways such as sphingolipid and glycerolipid metabolism (Feng et al., 2025).

Fig. 3.

Lipid Marker Analysis Between YG and YD (A) Heatmap of differential lipid molecules between YG and YD. (B)Random Forest Model Score Map. (C) Differential marker levels between YG and YD. (D) OPLS-DA plot of markers between YG and YD. (E) Marker replacement test plot between YG and YD.

To further identify the lipid biomarkers distinguishing YG and YD, feature analysis of the top 20 SDLs ranked by p-value was performed using a random forest model (Gu et al., 2023). The absolute mean Shap score value (feature contributions) from the test set samples was then used as the benchmark. A higher value indicated a greater contribution of the SDLs to the model, signifying their importance. Fig. 3B shows that seven SDLs with feature contributions ≥0.015 were ultimately selected as potential lipid biomarkers. Fig. 3C illustrates the quantitative differences in these seven potential lipid biomarkers between YG and YD. YG exhibited significantly higher quantitative values for these lipids, consistent with the trend observed in FC values and the biomarker screening-validation patterns reported by Li, Guo, et al. (2024) and Li, He, et al. (2024).

Additionally, the fatty acid types of potential lipid biomarkers were concentrated in PL and NL. The significant differences in these two lipid classes may also be key factors distinguishing the YG and YD ghee varieties. Subsequently, an OPLS-DA classification model for the seven potential lipid biomarkers was established using the methodology reported by Li, Guo, et al. (2024) and Li, He, et al. (2024). The OPLS-DA model in Fig. 3D enhanced discrimination by integrating grouping variables and eliminating unrelated noise, effectively distinguishing the differences between YG and YD. Permutation testing (Fig. 3E, n = 999) confirmed model robustness, with R² = 0.39 and Q² = −0.25. These results further confirmed the reliability of these candidate biomarkers in distinguishing YG from YD. Wang, Fan, et al. (2023) and Wang, Liu, et al. (2023) and Li, Guo, et al. (2024) and Li, He, et al. (2024) successively screened potential lipid biomarkers from yak milk of different species (human and ewe) and lactation stages. Their results identified two TAGs, one DAG, and one FFA as biomarkers for distinguishing human and ovine colostrum, and three TGs and one CE for distinguishing yak colostrum from mature milk. This analytical method provided a scientific basis for the authentication and detection of bovine breed-specific ghee.

3.6. VOC analysis

Volatile aroma compounds and their relative concentrations in ghee from different milk sources were analysed using combined HS-SPME and GC–MS techniques. A total of 23 VOCs were identified (Table 1), including 10 alcohols, three ketones, and seven carboxylic acids, such as isobutyric, isovaleric, n-pentanoic, n-hexanoic, caproic, n-decanoic, and n-butyric acids. The alcohols identified included n-pentanol, n-octanol, (2R,3R)-(−)-2,3-butanediol, isopentanol, n-hexanol, 1-octen-3-ol, n-heptanol, 2-ethylhexanol, 3-methyl-2-hexanol, and phenethyl alcohol. Among these, 2,3-butanediol is a primary compound produced by lactic acid bacteria during pyruvate catabolism, and it contributes to the sweet and buttery flavours present in most ghee samples (Liang et al., 2025). Ketones, particularly 3-hydroxy-2-butanone, 2-heptanone, and 2-nonanone, play a pivotal role in shaping the characteristic aroma. Additionally, trans-caryophyllene (an alkene) and the antioxidant BHT were identified. Notably, YG exhibited higher levels of alcohols and ketones than YD, resulting in a more complex flavour profile.

Table 1.

Comparison of Volatile Flavour Compound Content (Unit:μg/kg).

Flavour compounds	RI	Polar	YD (mean ± standard deviation)	YG (mean ± standard deviation)
Pentanol	1153	1250	27.06 ± 0.52a	15.18 ± 1.24b
3-hydroxy-2-butanone	1284	1203	653.63 ± 1.49a	199.87 ± 1.54b
1-Octanol	1538	1557	9.49 ± 2.80	ND
Isobutyric acid	1562	1570	31.20 ± 0.76a	8.30 ± 0.89
(2R,3R)-(−)-2,3-Butanediol	1515	1543	425.65 ± 1.42a	111.68 ± 1.73b
3-Methylbutanoic acid	1712	11,666	194.93 ± 1.72a	29.03 ± 1.17b
Pentanoic acid	1816	1733	11.66 ± 1.14a	4.40 ± 0.62b
1-Hexanoic acid	1841	1846	1121.01 ± 1.55a	262.61 ± 1.60b
Butylated hydroxytoluene	1973	1909	179.65 ± 1.70a	30.21 ± 2.39b
Octanoic acid	2079	2060	99.17 ± 0.82a	18.25 ± 3.21b
Decanoic acid	2251	2276	11.74 ± 0.90	ND
2-Heptanone	1095	1182	ND	59.28 ± 1.28
3-Methyl-1-butanol	1110	1209	ND	73.58 ± 2.85
1-Hexanol	1267	1355	ND	5.34 ± 0.41
2-Nonanone	1319	1390	ND	30.47 ± 2.15
Oct-1-en-3-ol	1390	1450	ND	3.39 ± 0.32
Heptan-1-ol	1396	1453	ND	2.83 ± 0.70
2-ethyl-1-hexanol	1443	1491	ND	5.01 ± 0.51
2-Hexanol, 3-methyl-	1304	1331	ND	11.77 ± 0.23
trans-Caryophyllene	1612	1595	ND	4.22 ± 0.35
Butyric Acid	1650	1625	ND	236.10 ± 1.81
Phenethyl alcohol	1978	1906	ND	3.64 ± 1.11

Note: ND indicates not detected; different letters on the same data points indicate significant differences at the p < 0.05 level.

Quantitative analysis revealed that 3-hydroxy-2-butanone was the most abundant aroma-active compound in traditional ghee from Luqu, Gannan. This finding, which has not been previously documented by Feng et al. (2025) or Liang et al. (2025), indicates the potential of 3-hydroxy-2-butanone as a characteristic flavour compound in local ghee. Furthermore, this finding suggests that, beyond processing methods, the milk source significantly contributes to the final aroma, providing a new theoretical perspective on the mechanisms underlying ghee aroma characteristics. Ketones are primarily formed by the oxidation of unsaturated fatty acids, amino acid degradation, or microbial metabolic processes (Zhang et al., 2024; Abdelaziz Elbarbary et al., 2025). Ketones possess distinctive flavours with low thresholds, often displaying a mild fruity aroma (Liang et al., 2025).

Previous studies (Li, Zhang, & Wang, 2012; Yang et al., 2023) have shown that ketones are commonly used as markers to assess oxidative flavours in dairy products, especially milk. For example, the oxidation of linoleic acid predominantly generates 2-heptanone, enhancing the flavour and sweetness of ghee. Aldehydes primarily result from the oxidation of yak milk fat and amino acid conversion, typically contributing pleasant grassy, fatty aromas and fruity flavours (Yang et al., 2023). Alcohols contribute to lipid oxidation and aldehyde reduction. For example, 2,3-butanediol is a key compound produced by lactic acid bacteria during pyruvate catabolism, exhibiting sweet and buttery flavours (Liang et al., 2025). A recent study identified acids as the primary flavour components in Gannan ghee, suggesting that they may serve as substrates for the microbial breakdown of amino acids into esters, thus influencing ghee flavour profiles (Liang et al., 2025). However, this hypothesis requires further validation in future studies.

3.7. Multivariate statistical analysis of VOCs

Fig. 4A presents the OPLS-DA score plot for the aroma components of the two ghee types. A distinct separation between the groups was observed, and the 95% confidence interval ellipses showed no overlap, indicating that intergroup differences substantially exceeded intra-group variations. As shown in Fig. 4B, the results of 200 permutation tests indicated that the Q2 regression line intersects the vertical axis below, confirming that the model was not overfit and had strong predictive capability. These findings, therefore, can be used to distinguish ghee derived from different milk sources.

Fig. 4.

(A) OPLS-DA score plot of ghee VOCs. (B) Model cross-validation diagram. (C) VIP analysis of key compounds. (D) Boxplots of α-diversity indices for YG and YD. (E) PCoA plots of β-diversity for YG and YD. (F) Species composition at the phylum level for YG and YD. (G) Species composition at the genus level for YG and YD. (H) LEfSe plot for YG and YD. (I) KEGG Metabolic Pathway Entry Diagram.

To further assess the role of specific aroma compounds in distinguishing the two types of ghee, four aroma components were identified on the basis of statistical criteria (p < 0.05, VIP > 1) (Fig. 4C). These compounds included n-hexanoic acid, 3-hydroxy-2-butanone, (2R,3R)-(−)-2,3-butanediol, and n-butyric acid. C3-C6 short-chain ketones such as 3-hydroxy-2-butanone are recognised as aromatic compounds abundant in dairy products, contributing distinctive aromas to ghee (Peterson & Reineccius, 2003).

Ghee's aroma characteristics are not solely defined by the composition of aroma components. The odour activity value (OAV) is a quantitative measure of an aroma's significance in evaluating its contribution to the overall aroma profile of ghee and plays a critical role in identifying key aroma components. On the basis of the studies by Zhang et al. (2024) and van Gemert (2011), the OAV was calculated to assess the individual aromas' contribution to the overall aroma profile of ghee. Table S1 shows that isoamyl alcohol and 3-hydroxy-2-butanone had OAV values greater than 1, indicating a significant contribution to ghee's flavour profile. In contrast, 2-heptanone, 2-nonanone, 1-octen-3-ol, n-heptanol, 3-methyl-2-hexanol, isobutyric acid, n-butyric acid, isovaleric acid, and n-hexanoic acid showed OAV values between 0.1 and 1. Thus, while these compounds had a lesser impact on the flavour profile, they still contributed to the flavour characteristics of ghee.

3.8. Microbial community diversity

In the α-diversity analysis, the Chao1 and richness indices reflect microbial community abundance, while the Shannon and Simpson indices reflect its diversity. Higher values indicate increased bacterial species richness and abundance. As shown in Fig. 4D, the Shannon, Chao1, Simpson, and richness indices for YD were significantly higher than those for YG. Fig. 4E presents the structural similarities in the microbial communities of YG and YD, as calculated using the Bray-Curtis distance. Axis 1 accounts for 98% of the variance, while Axis 2 accounts for 1.4%. YG and YD exhibited distinct microbial distribution separation, and the combined explained variance exceeded 98%, indicating substantial differences in their overall microbial composition. Feng et al. (2025) also found that different milk sources resulted in distinct microbial compositions in clarified ghee.

Species stack plots primarily highlight intergroup differences in species composition. QIIME2 (2019.4) allows for visualization of microbial community distribution across taxonomic levels in the samples. As shown in Figs. 4F and G, YG and YD showed differences in microbial abundance at both the phylum and genus levels. Notably, the microbial communities shared similar compositional categories, but their relative species abundances varied. At the phylum level, the dominant phyla by average relative abundance were Ascomycota (YG: 84.44% vs. YD: 47.59%) and Proteobacteria (YG: 12.85% vs. YD: 4.18%). The findings reported by Feng et al. (2025) suggest that the phylum Basidiomycota dominates microbial diversity in ghee, which is consistent with the results of this study. At the genus level, Lactococcus (YG: 71.32% vs. YD: 42.72%) and Acinetobacter (YG: 7.35% vs. YD: 30.52%) showed relatively high average abundances, although YG had lower overall relative abundances than YD. Studies by Feng et al. (2025) and Liu et al. (2022) both identified Lactobacillus as highly abundant in fresh ghee.

The higher levels of protein, fat, and nutrients in plateau yak milk may promote the growth and metabolic activity of lactic acid bacteria. Lactobacillus, Streptococcus, and Lactococcus—acidogenic microorganisms extensively documented in fermented dairy products—play key roles in the development of flavour profiles (Liu et al., 2024). These bacteria can also reduce the pH of the milk matrix through metabolic activity, thereby inhibiting the growth of harmful microorganisms and contributing to extended shelf life. During fermentation, microorganisms secrete various enzymes to hydrolyse or oxidise lipids, ultimately converting them into small molecules such as aldehydes, alcohols, ketones, and esters, which lead to changes in food flavour (Feng et al., 2021; Wang, Fan, et al., 2023; Wang, Liu, et al., 2023). Chu et al. (2025) indicated that microorganisms facilitate the oxidation and metabolism of lipids such as triglycerides (TG) and phospholipids (PL).

3.9. Linear discriminant analysis effect size

Linear discriminant analysis effect size (LEfSe) was conducted to identify the taxa exhibiting significant compositional differences among groups. As shown in Fig. 4H, genus-level LEfSe analysis identified 20 dominant genera (LDA > 4, p < 0.05), including Lactococcus, Leuconostoc, Streptococcus, Lactobacillus, Roseburia, and Pseudomonas. Among them, Lactococcus, Lactobacillus, and Streptococcus were closely associated with fermentation processes. This observation was consistent with the findings reported by Liu et al. (2022).

However, Feng et al. (2025) reported that Gluconobacter was also dominant in ghee, which could be attributed to differences in milk source and feeding conditions. Furthermore, Feng et al. (2025) demonstrated that Lactococcus and Lactobacillus contribute to the release of VOCs and the production of free fatty acids during ghee processing. Chen, Huang, Yu, and Tian (2020) further showed that Lactobacillus esterases hydrolyse di- and monoglycerides to release short-chain fatty acids and catalyse ester formation from glycerides and alcohols through transferase reactions. Collectively, these findings substantiate the positive contributions of Lactobacillus, Lactococcus, and Streptococcus to ghee quality through lipid metabolism.

PICRUSt predicts functional potential on the basis of characteristic gene sequences. In this study, PICRUSt was integrated with the Kyoto Encyclopedia of Genes and Genomes (KEGG) database to annotate the functional genes of the bacterial community in ghee. As shown in Fig. 4I, the major enriched pathways included amino acid, carbohydrate, cofactor and vitamin, and lipid metabolism. Lactobacillus, Lactococcus, and Streptococcus may influence the flavour profile of ghee mainly through amino acid and carbohydrate metabolism. Disruption of the bacterial community structure may impair these metabolic functions.

3.10. Correlation analysis of differential lipids, dominant microorganisms, and key flavour compounds

To elucidate the interrelationships among SDLs, dominant microbial genera, and characteristic flavour compounds in YG and YD, a correlation network was constructed using Spearman's correlation analysis (|r| > 0.8, p < 0.05) (Figs. 5A, B). Notably, correlation analysis reveals statistical associations rather than causal relationships; therefore, the contributions of specific microorganisms to individual volatile compounds cannot be conclusively established solely on the basis of these results. In addition, a conceptual framework linking lipid composition, microbial metabolism, and volatile formation has been proposed to provide a more integrated interpretation of the results.

Fig. 5.

Correlation Analysis of Characteristic Flavour Compounds with Differential Lipids and Key Microorganisms (A) Correlation between characteristic flavour compounds and differential lipids. (B) Correlation between characteristic flavour compounds and key microorganisms. Note: Orange indicates SDLs; pink indicates differential VOCs; blue-green indicates key microorganisms; blue lines denote negative correlations; pink lines denote positive correlations; line thickness indicates correlation strength. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 5A illustrates the representative correlations among glycerides, phospholipids, and VOCs. A total of 12 VOCs and 20 significantly differentiated lipids (SDLs) showed strong correlations, involving phosphatidylethanolamine (PE), phosphatidylcholine (PC), and triglyceride (TG) species, including PE (14:1e_18:3), LPC (20:5), PC (34:3e), PC (36:3e), PC (36:4e), and multiple TGs. Overall, phospholipids, particularly PE and PC, showed stronger associations with VOCs than TGs. For instance, PE (36:3e) and PE (18:1_22:6) were positively correlated with 2-heptanone (p < 0.05), whereas TG (43:7) and PE (14:1e_18:2) were negatively correlated with n-hexanol, amyl alcohol, and isobutyric acid.

These correlation patterns were consistent with established mechanisms of lipid oxidation. Previous studies have demonstrated that lipid oxidation is a major source of aldehydes and ketones (Chu et al., 2025), and that DG and PE lipids rich in polyunsaturated fatty acids may serve as essential precursors for the formation of key volatile flavour compounds in fish oil (Li, Bai, et al., 2025; Li, Yuan, et al., 2025). In this study, the YG group showed higher unsaturated lipid content than the YD group, which likely accounted for the richer flavour profile of YG, particularly its significantly higher levels of ketones and aldehydes than in YD. Nevertheless, these mechanistic links remained inferential and required experimental validation.

Correlation analysis between key microorganisms and VOCs (Fig. 5B) identified 12 VOCs significantly associated with 14 genera, including Enterobacter_B, Acinetobacter, Chryseobacterium, Enterococcus_B, Gluconobacter_A, Kaistella, Klebsiella, Lactobacillus, Lactococcus_A, Leuconostoc_B, Macrococcus_B, Pseudomonas_E, Rothia, and Streptococcus. Notably, Pseudomonas_E was negatively correlated with most key VOCs. However, Spearman correlation analysis only showed statistical associations and did not support causal inference; therefore, these negative correlations cannot be interpreted as evidence that Pseudomonas_E directly inhibits or reduces volatile formation, even though it is widely recognised as a major spoilage microorganism in milk fat. In contrast, alcohols and ketones were more strongly correlated with Gluconobacter_A, particularly 2-heptanone and 3-hydroxy-2-butanone, whereas acids showed closer associations with Klebsiella and Macrococcus_B.

When integrated with the lipid-VOC network (Fig. 5A), these associations indicate that specific microbial genera and lipid classes co-vary with characteristic volatiles. Genera such as Lactobacillus and Leuconostoc_B were correlated with acids and alcohols, consistent with their known fermentative metabolism. Overall, while specific VOCs were correlated with FFAs and genera such as Bacillus and Lactobacillus, the present network analysis provided associative rather than causal evidence. Future studies combining isotope tracing, targeted lipidomics, and functional assays are required to clarify the mechanistic links among lipid classes, microbial metabolism, and volatile formation in ghee.

4. Conclusion

YG and YD showed notable differences in terms of lipid composition, VOCs, and microbial communities. Lipidomics identified 43 differentially expressed lipids, with seven potential lipid biomarkers for each ghee type: DG(16:1_6:0), LPE(22:5), SM(t40:1), TG(18:1_14:0_20:5), TG(20:4e_16:0_18:1), TG(20:4e_18:0_18:1), and TG(4:0_10:0_22:6). Additionally, amplicon sequencing and GC–MS analysis highlighted the roles of Lactococcus, Gluconobacter, and Lactobacillus in ghee aroma production, with 3-hydroxy-2-butanone emerging as a key flavour compound. Correlation analysis suggested that flavour differences may result from lipid oxidation and microbial metabolic activity. These results provide deeper insights into the factors influencing ghee quality. Future studies incorporating flavoromics and metabolomics could track changes over time, facilitate the development of freshness grading models, and improve quality control strategies.

CRediT authorship contribution statement

Xinyun Yu: Writing – original draft, Visualization, Data curation, Conceptualization. Yongqiang Cao: Supervision, Project administration, Methodology. Yiheng Li: Writing – review & editing, Supervision, Formal analysis, Conceptualization. Ruiyun Wang: Supervision, Funding acquisition. Longlin Wang: Visualization, Software, Formal analysis, Data curation. Juan Song: Visualization, Investigation. Jinliang Zhang: Validation, Methodology. Chengrui Shi: Validation, Methodology. Mailin Song: Investigation. Weibing Zhang: Writing – review & editing, Project administration, Funding acquisition. Fazheng Ren: Resources, Project administration, Funding acquisition. Pengcheng Wen: Writing – review & editing, Resources, Project administration, Funding acquisition.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

Data will be made available on request.