Milk-based multi-omics reveals distinct quality signatures associated with Astragalus polysaccharide supplementation in late lactation goats Authors: Shanshan Han, Peng Wang, Yunhong Hu, Qiang Zhang, Ke Wan, Xiaoyu Wang Published: 2026 ABSTRACT AbstractLate lactation is frequently accompanied by changes in goat milk composition and volatile-related quality, which may reduce its commercial value. This study investigated whether Astragalus polysaccharide (APS), a plant-derived bioactive polysaccharide from Astragalus membranaceus (Fisch.) Bunge, is associated with milk quality responses distinct from those induced by probiotic supplementation. Late lactation Saanen dairy goats were assigned to probiotic or APS supplementation for 30 days, and milk samples were analyzed using conventional composition testing, electronic nose (E-nose) profiling, proteomics, untargeted metabolomics, and lipidomics. APS supplementation reduced milk urea nitrogen by 6.20 mg/dL and decreased milk fat percentage, while electronic nose (E-nose) profiling revealed a distinct sensor-level volatile fingerprint dominated by W5S, W2W, and W1S responses. At the molecular level, APS generated a broader metabolomic response than probiotic supplementation, with 904 versus 587 differential metabolites, and was associated with a milk proteomic signature featuring higher ATP citrate lyase abundance (2.52-fold). Lipidomics further showed selective remodeling of milk lipid subclasses, including increased phosphatidylethanolamine, phosphatidylcholine, and ceramide abundance by approximately 1.4-fold, 1.5-fold, and nearly 5-fold, respectively. Together, these milk-based data indicate that APS supplementation was associated with a distinct quality signature involving nitrogen-use indicators, volatile-related sensor profiles, lipid-related protein features, and lipid subclass redistribution. The findings support APS as a potential non-microbial nutritional strategy for modulating late lactation goat milk quality.Keywords: Goat milk, Astragalus polysaccharide, Late lactation, Electronic nose, Multi-omics, Lipidomics Introduction Goat milk is increasingly recognized globally as a functional food matrix, owing to its distinct protein profile, low allergenicity, and a wealth of bioactive constituents (Alkaisy et al., 2023; Clark & Mora García, 2017). Compared to bovine milk, it is characterized by a higher proportion of small milk fat globules, as well as an enrichment of short-chain, medium-chain, and polyunsaturated fatty acids (Kiiru et al., 2018). Despite these nutritional merits, the industrial scalability of goat dairy remains limited by the pronounced seasonal fluctuation in milk quality. This challenge is particularly pronounced during late lactation, a period where the natural physiological decline in milk yield coincides with a marked deterioration in sensory attributes (Saviard et al., 2026). Late lactation milk often exhibits an increased milk fat content, which may be associated with altered lipid metabolic pathways, including peroxisome proliferator-activated receptor (PPAR)-related lipid metabolic signaling, as reported in dairy ruminants (Y. Fan et al., 2020). However, consumers often display reluctance toward late lactation milk because of its intensified goaty off-flavor and poor oxidative stability, both of which significantly reduce its commercial value (Sakha & Ahmadi, 2024; Xi et al., 2023). Because these sensory changes are closely linked to milk compositional and lipid-derived chemical features, strategies that modulate milk metabolism and molecular composition may be more suitable than post-processing approaches alone for addressing late lactation quality deterioration. Such efforts also align with the broader need for reliable goat milk quality control, including authentication methods for detecting adulteration in goat milk and milk from other dairy species (Ait El Alia et al., 2026a, 2026b, 2026c). Nutritional management and functional feed additives are pivotal for sustaining milk yield and quality while reducing reliance on antibiotics during late lactation. Within this context, extensive research demonstrates that targeted dietary strategies can modulate metabolic efficiency and milk composition in dairy ruminants. Approaches ranging from optimizing dietary protein and lipid profiles to incorporating alternative forage resources, microalgae, oilseeds, functional fatty acids, and botanical supplements have been shown to influence milk yield, milk physicochemical traits, fatty acid composition, antioxidant-related properties, and systemic metabolic responses (AlSuwaiegh et al., 2022; Boukrouh et al., 2023; Boukrouh et al., 2025; Boukrouh et al., 2026; Kyriakaki et al., 2023; Plata-Pérez et al., 2022). In particular, recent studies in goats have shown that alternative forage resources such as Sulla flexuosa and microalgae supplementation can modify milk quality-related traits, especially fatty acid profiles, while meta-analytic evidence in dairy cows further supports microalgae as a dietary strategy for improving milk fatty acid composition (Boukrouh et al., 2023; Boukrouh et al., 2025; Boukrouh et al., 2026). These findings establish a nutritional basis for exploring next generation dietary interventions to stabilize caprine milk quality under late lactation physiological stress. Among these interventions, microbial and probiotic feed additives, such as rumen-native microbes, Lactobacillus spp., Bifidobacterium spp. and Streptococcus thermophilus, have been widely explored as nutritional strategies for improving ruminant health and milk production efficiency (Kholif et al., 2024; Marinho et al., 2024). These additives may contribute to host metabolic homeostasis by supporting intestinal barrier function, modulating immune responses, and promoting the production of microbial metabolites such as short-chain fatty acids (SCFAs) (Ayyat et al., 2024; Cappellozza et al., 2025; Nalla et al., 2022). However, the effects of probiotics are often strain-, dose-, diet-, and management-dependent, and their practical application may be influenced by microbial viability, storage stability, and the indirect nature of gut-mediated regulation (Leistikow et al., 2022). Therefore, alternative non-microbial bioactive feed additives may provide complementary opportunities for regulating milk quality under late lactation physiological stress. Astragalus polysaccharides (APS), representative bioactive plant macromolecules derived from Astragalus membranaceus (Fisch.) Bunge, have attracted attention because of their stability, immunomodulatory properties, and potential metabolic regulatory effects (Liang et al., 2024; Tang et al., 2022). Beyond their reported roles in intestinal barrier protection and anti-inflammatory regulation, existing evidence suggests that APS and other Astragalus-derived interventions may influence systemic metabolic and antioxidant responses in a model-dependent manner (Luo et al., 2020; Zeng et al., 2020). For example, APS intervention in heat-stressed dairy cows has been associated with changes in serum metabolites and pathways related to glucose metabolism, amino acid metabolism, glutathione metabolism, and prolactin signaling, suggesting a possible role in systemic metabolic adaptation under stress (Zeng et al., 2020). In addition, Astragalus membranaceus supplementation has been reported to improve oxidative stability in Cashmere goats, further supporting its antioxidant-related potential of Astragalus-derived interventions in ruminants (Luo et al., 2020). Nevertheless, these previous studies were conducted in different physiological contexts and biological matrices, and the effects of APS supplementation on late lactation goat milk quality remain insufficiently defined. In particular, it is still unclear whether APS is associated with coordinated changes in conventional milk composition, volatile-related sensory fingerprints, and milk-derived proteomic, metabolomic, and lipidomic signatures, or whether these responses differ from those induced by probiotic supplementation. Clarifying this issue is important because late lactation milk deterioration may involve not only changes in bulk compositional traits but also alterations in lipid molecular composition, oxidative stability-related features, and flavor-associated chemical characteristics. To address this gap, we applied a longitudinal milk-based multi-omics framework to compare the effects of APS and probiotic supplementation on late lactation goat milk quality. Conventional milk composition, E-nose sensor responses, milk proteomic profiles, untargeted metabolomic features, and lipidomic remodeling were evaluated before and after a 30-day supplementation period. We hypothesized that APS would generate a milk quality response distinct from that induced by probiotic supplementation, reflected by coordinated changes in conventional composition and milk-derived molecular signatures. Materials and methods Animal management and experimental design All animal procedures in this study were conducted in strict accordance with the ARRIVE (Animal Research: Reporting of In Vivo Experiments) guidelines and institutional livestock animal welfare requirements, and were approved by the Animal Ethics Committee of the Bio-Agriculture Institute of Shaanxi (Approval No. 2025010). The study was carried out using privately owned lactating dairy goats under commercial farm management conditions. Healthy adult lactating female dairy goats in late lactation were used in this study. As only female animals were included, the influence of sex on the results was not applicable. Thirty clinically healthy Saanen dairy goats in late lactation, defined according to farm production records as 260 ± 15 days in milk, were recruited from a commercial farm in Xi'an, China. These goats had balanced parity (average parity 3) and initial milk yield (1.2 ± 0.3 kg/day). The animals were housed in ventilated free-stall barns with ad libitum access to fresh water and standardized bedding management to minimize environmental variance. Eligibility and group allocation were based on available production and health records, including days in milk, parity, baseline milk yield, mastitis history, and recent antibiotic treatment. Goats with a history of mastitis or antibiotic treatment within the preceding three months were excluded before group allocation. A randomized longitudinal before–after intervention design was employed to evaluate milk quality responses to dietary supplementation. Goats were numbered based on parity and baseline milk yield, and subsequently randomly assigned to treatment groups using a computer-generated sequence of random numbers. After a 7-day acclimatization period, eligible goats were randomized into two comparable groups according to parity and baseline milk yield: a Probiotic Group (PRO; n = 15) and an APS group (APS; n = 15). The PRO group received a basal total mixed ration (TMR) supplemented with 5 g/head/day of a composite microbial agent (L. plantarum, B. subtilis, and S. cerevisiae; > 1 × 109 CFU/g, Shaanxi Xincheng Datang Bio-Tech, China). The APS group received the same basal TMR supplemented with 5 g/head/day of a commercial feed-grade Astragalus polysaccharide powder with 55% polysaccharide purity and Astragaloside IV > 0.1% (GMP certificate No. (2023) 27004; veterinary drug production license No. 27058; batch No. 250403Q; Shaanxi Xincheng Datang Bio-pharmaceutical Co., Ltd., China). The supplementation level of 5 g/head/day was selected based on the preliminary feeding period, during which both supplements were completely consumed without apparent feed refusal or adverse clinical signs, and was considered practical for daily administration under commercial farm conditions. The trial lasted for 30 days, during which all supplements were top-dressed onto a small portion of concentrate at morning feeding to ensure complete ingestion. The basal diet (Yinchuan Kangdi Ruminant Animal Nutritional Department Ji Co., Ltd., China) was formulated to meet the nutritional requirements recommended by the NRC (2007). The ingredient composition, forage-to-concentrate ratio, and nutrient profile of the basal TMR are provided in Supplementary Table S1. Sample collection protocols Milk samples were collected at 06:00 on Day 0 (serving as Baseline) and Day 30 (serving as Endpoint). For compositional analysis, approximately 50 mL of fresh milk was collected into preservative-treated tubes, stored at 4 °C, and immediately analyzed for fat, protein, lactose, dry matter, milk urea nitrogen (MUN), and somatic cell count (SCC) using a CombiFoss FT+ analyzer (Foss Electric, Denmark). For E-nose and multi-omics analyses, milk samples were collected using standardized hygienic sampling procedures. Teats were thoroughly sanitized with 0.5% iodine solution and dried with single-use paper towels prior to milking. After discarding the first three strips of milk, milk was divided into separate aliquots for E-nose, proteomic, metabolomic, and lipidomic analyses. Aliquots for E-nose analysis were collected in sterile tubes, whereas aliquots for proteomic, metabolomic, and lipidomic analyses were collected into sterile, RNase-free cryotubes. All aliquots were prepared immediately at the farm, snap-frozen in liquid nitrogen, stored at −80 °C, and transported to the analytical laboratory on dry ice. Each aliquot was thawed once for the corresponding analysis. Milk samples used for E-nose and multi-omics analyses were coded before instrumental detection, and treatment information was withheld from the analytical personnel. Electronic nose analysis Olfactory profiles were digitized using a PEN3-Plus electronic nose (Airsense Analytics GmbH, Schwerin, Germany) equipped with 10 metal oxide semiconductor (MOS) sensors. The sensor array consisted of W1C, W5S, W3C, W6S, W5C, W1S, W1W, W2S, W2W, and W3S, covering broad responses to aromatic compounds, nitrogen oxides, hydrogen, alkanes, methane, sulfur-containing compounds, terpenes, alcohols, and aliphatic compounds. Milk samples (15 mL) were incubated in 50 mL headspace vials at 20 °C for 30 min to reach volatile equilibrium. Before each measurement sequence, the system was purged with filtered clean air until the baseline stabilized. Sensor arrays were operated with the following settings: cleaning, 180 s; zeroing, 5 s; preparation, 60 s; and measurement, 100 s at a flow rate of 400 mL/min. Measurements were conducted under consistent indoor laboratory conditions, and each sample was analyzed in triplicate. Sensor responses were exported as relative response values (G/G0) for subsequent statistical analysis. No additional post hoc drift-correction algorithm was applied. Label-free proteomics analysis Protein extraction and digestion were performed following a DIA-based quantitative proteomics workflow. Milk protein samples were lysed using SDT buffer (4% SDS, 100 mM Tris-HCl, pH 7.6), and protein concentrations were determined using a BCA Protein Assay Kit. After reduction with DTT and alkylation with iodoacetamide, proteins were digested with trypsin at a trypsin-to-protein ratio of 1:50 (w/w) at 37 °C for 15–18 h. The resulting peptides were desalted using SPE MCX cartridges, reconstituted in 0.1% formic acid, and spiked with iRT calibration peptides for retention-time alignment. For each LC-MS/MS run, 400 ng of peptides was injected. Peptides were analyzed using a Vanquish Neo UHPLC system coupled to an Orbitrap Astral mass spectrometer (Thermo Fisher Scientific) in data-independent acquisition (DIA) mode. Chromatographic separation was performed using a trap column (300 μm × 5 mm, 3 μm) and an analytical column (150 μm × 15 cm, 1.7 μm) under a high-pressure system (up to 1450 bar). The LC gradient was programmed as follows: 0–4.0 min, 4–20% B; 4.0–5.8 min, 20–35% B; 5.8–6.2 min, 35–99% B; and 6.2–6.9 min, 99% B, with a flow rate of 2.5 μL/min. MS data were acquired with precursor scans ranging from 380 to 980 m/z at an MS1 resolution of 240,000 at m/z 200. Data-independent acquisition was performed using 299 DIA windows with a 2 m/z isolation width, and fragmentation was conducted using higher-energy collisional dissociation (HCD) at 25 eV. MS2 spectra were acquired at a resolution of 17,500 at m/z 200. Raw data acquired in data-independent acquisition (DIA) mode were processed using DIA-NN v1.8.1 against the Capra hircus UniProt protein database (uniprotkb_Capra_78255_20251112), with trypsin specificity and up to one missed cleavage allowed. Carbamidomethylation of cysteine was set as a fixed modification, while oxidation (M) and protein N-terminal acetylation were treated as variable modifications. Protein identification was controlled at a false discovery rate (FDR) of ≤1% at the precursor and protein levels. Protein abundances were calculated as relative spectral peak-area-based intensities and used for downstream analysis without additional normalization, log2 transformation, or missing-value imputation. Each biological sample was independently prepared, digested, and analyzed once, without technical duplicate or triplicate injections. Six biological replicates per group were included for downstream proteomic analysis. Untargeted metabolomics analysis Metabolites were extracted using a pre-cooled methanol/acetonitrile/water solution (2:2:1, v/v/v). Briefly, milk samples were thawed at 4 °C, mixed with extraction solvent, vortexed, subjected to low-temperature ultrasonication for 30 min, incubated at −20 °C for 10 min, and centrifuged at 14,000 ×g for 20 min at 4 °C. The supernatants were vacuum-dried, reconstituted in 100 μL acetonitrile/water (1,1, v/v), vortexed, centrifuged at 14,000 ×g for 15 min at 4 °C, and used for LC-MS analysis. Chromatographic separation was performed on a Vanquish UHPLC system equipped with an ACQUITY UPLC BEH Amide column (2.1 mm × 100 mm, 1.7 μm; Waters), coupled to an Orbitrap Exploris 480 mass spectrometer (Thermo Fisher Scientific) in both positive and negative ESI modes. The column temperature was maintained at 25 °C, the flow rate was 0.5 mL/min, and the injection volume was 2 μL. Mobile phase A consisted of water containing 25 mM ammonium acetate and 25 mM ammonium hydroxide, and mobile phase B was acetonitrile. The gradient was as follows: 0–0.5 min, 95% B; 0.5–5.0 min, 95–65% B; 5.0–5.5 min, 65–40% B; 5.5–6.0 min, 40–95% B; and 6.0–8.0 min, 95% B. The autosampler was maintained at 4 °C. Samples were analyzed in randomized order, and pooled QC samples, prepared by mixing equal aliquots from all study samples, were inserted regularly throughout the analytical sequence to monitor system stability. The ESI source parameters were set as follows: Gas1, 50; Gas2, 2; source temperature, 350 °C; spray voltage, +3500 V in positive mode and − 2800 V in negative mode. Full-scan MS data were acquired over 70–1200 m/z at a resolution of 60,000 with an accumulation time of 100 ms. MS/MS data were acquired over 70–1200 m/z at a resolution of 60,000 with an accumulation time of 100 ms, and the dynamic exclusion time was 4 s. Raw files were converted to mzXML format using ProteoWizard and processed using XCMS for peak picking and alignment. Peak alignment, retention-time correction, and peak-area extraction were performed in XCMS. Features with more than 50% missing values were removed, missing values were imputed using the K-nearest neighbor method, and features with QC RSD >50% were filtered out. Identification relied on accurate mass matching (<10 ppm) and MS/MS spectra against an in-house standard library and public databases (HMDB, METLIN). Metabolite annotation was further supported by retention time, MS/MS spectra, and collision energy matching against an in-house database, followed by manual confirmation. The annotation confidence was reported as MSI level 2 or above. The metabolomics data were used for relative abundance profiling and no internal standards were used. No batch correction, signal-drift correction, or log2 transformation was applied. Normalization or scaling was applied only within multivariate analyses where required. Each biological sample was independently prepared and analyzed once, without technical duplicate or triplicate injections. Untargeted lipidomics analysis Lipids were extracted via the methyl tert-butyl ether (MTBE) method. Briefly, 100 μL of milk sample was mixed with 200 μL of water, followed by the addition of 800 μL of MTBE. The mixture was ultrasonicated for 20 min at 4 °C, kept at room temperature for 30 min, and centrifuged at 14,000 ×g for 15 min at 10 °C. The upper organic phase was collected and dried under nitrogen. Lipid extracts were reconstituted in 200 μL of 90% isopropanol/acetonitrile, centrifuged at 14,000 ×g for 15 min, and 3 μL of supernatant was injected for LC-MS/MS analysis. Lipidomic profiling was performed using UHPLC coupled to a Q Exactive mass spectrometer (Thermo Fisher Scientific). Lipid species were separated on a CSH C18 column (2.1 mm × 100 mm, 1.7 μm; Waters) at 45 °C with a flow rate of 300 μL/min. Mobile phase A consisted of acetonitrile/water (6:4, v/v) containing 0.1% formic acid and 0.1 mM ammonium formate, and mobile phase B consisted of acetonitrile/isopropanol (1:9, v/v) containing 0.1% formic acid and 0.1 mM ammonium formate. The gradient was as follows: 0–3.5 min, 40% B; 3.5–13 min, 40–75% B; 13–19 min, 75–99% B; and 19–24 min, 40% B. Samples were maintained at 10 °C in the autosampler and analyzed in randomized order. Pooled quality-control (QC) samples were included in the analytical sequence to evaluate instrument stability and repeatability. Data were acquired in both positive and negative electrospray ionization (ESI) modes using data-dependent acquisition (DDA). The ESI source parameters were as follows: heater temperature, 300 °C; sheath gas flow rate, 45 arb; auxiliary gas flow rate, 15 arb; sweep gas flow rate, 1 arb; spray voltage, 3.0 kV; capillary temperature, 350 °C; and S-Lens RF level, 50%. The MS1 scan range was 200–1800 m/z, with an MS1 resolution of 70,000 at m/z 200. After each full MS scan, ten MS/MS spectra were acquired using higher-energy collisional dissociation (HCD), with an MS2 resolution of 17,500 at m/z 200. Raw data were processed using LipidSearch (Thermo Fisher Scientific) for lipid peak extraction and identification. The main LipidSearch parameters were precursor tolerance, 5 ppm; product tolerance, 5 ppm; and product ion threshold, 5%. Lipid abundances were reported as raw intensity-based relative values without log2 transformation or additional normalization. No isotope correction was applied in the lipidomics data processing workflow. Bioinformatics and multi-omics integration analysis For omics data, multivariate analyses were conducted using Python (v3.14.2) and R (v4.3.0). Principal component analysis (PCA) and orthogonal partial least squares discriminant analysis (OPLS-DA) were conducted to visualize group clustering and supplementation-associated molecular separation. To reduce the risk of overfitting in supervised models, OPLS-DA robustness was evaluated using 7-fold cross-validation and 200-times permutation testing. Model reliability was assessed using R2Y(cum), Q2(cum), and permutation-test intercepts, with particular emphasis on negative Q2 intercepts. OPLS-DA results were interpreted as supervised exploratory evidence and were considered together with univariate statistics and biological relevance. Differential molecular features were screened according to the criteria defined for each omics layer. For metabolomic and lipidomic analyses, P values were adjusted using the Benjamini–Hochberg false discovery rate (FDR) method, and FDR-adjusted p < 0.05 was used to define statistically significant features where applicable. Spearman's rank correlation analysis was performed to construct an interaction matrix between differential molecular features and E-nose sensor response values. To reduce false positives caused by multiple testing, correlation P values were adjusted using the Benjamini–Hochberg FDR method. Significant associations in the multi-omics correlation matrix were defined as |r| > 0.6 and FDR < 0.05. For lipid–lipid correlation network analysis, a more stringent threshold of |r| > 0.8 and FDR < 0.05 was applied. Correlation-based results were interpreted as association patterns rather than evidence of direct mechanistic or causal relationships. Significant correlations were visualized using heatmaps to identify coordinated molecular features associated with volatile-related sensor responses. Statistical analysis Statistical analyses of milk composition, E-nose sensor responses, and other non-omics measurements were performed using SPSS 26.0 (USA). Prior to hypothesis testing, data normality was assessed using the Shapiro-Wilk test. Longitudinal changes (Day 0 vs. Day 30) within each group were evaluated using a paired Student's t-test. Group differences at the endpoint (PRO vs. APS) were assessed using an independent Student's t-test. Data are presented as mean ± standard deviation. Significance was declared at p < 0.05. Results and discussion Conventional milk composition responses to APS and probiotic supplementation To assess how APS and probiotic supplementation influenced conventional milk quality during late lactation, SCC and major compositional traits, including milk fat, total solids, protein, lactose, and MUN, were first evaluated (Fig. 1). SCC is widely used in dairy production as an indicator of udder health, mammary inflammatory status, and milk hygienic quality (Halasa et al., 2020; Pegolo et al., 2021), but its interpretation in goat milk requires caution because both infectious and non-infectious factors, including parity, season, milk yield, and lactation stage, can influence SCC values (Jiménez-Granado et al., 2014; Paape et al., 2007). In this study, SCC did not change significantly after either probiotic or APS supplementation (Fig. 1A), suggesting that the 30-day interventions did not markedly affect this udder health-related indicator. This stability provides a useful background for interpreting the compositional data, because the observed changes in milk fat, total solids, protein, lactose, and MUN occurred without a parallel shift in SCC. Milk composition before and after probiotic or APS supplementation. A–F, Somatic cell count (SCC; A), milk fat (B), total solids (C), milk protein (D), lactose (E), and milk urea nitrogen (MUN; F). Con A and Con B represent the baseline samples corresponding to the PRO and APS groups, respectively. Data are shown as mean ± s.d. (n = 6 biological samples per group). Baseline and endpoint values within each treatment group were compared using two-sided paired Student's t-tests. *P < 0.05, **P < 0.01 and ***P < 0.001. The two supplementation strategies produced distinct responses in bulk milk composition. Microbial feed additives have been reported to influence ruminal fermentation, feed utilization, animal performance, and milk-related outcomes in ruminants, but their responses depend on microbial species, diet, physiological stage, and management conditions (Kholif et al., 2024). In the PRO group, milk fat percentage increased significantly after supplementation (p < 0.05; Fig. 1B), accompanied by a significant increase in total solids (p < 0.05; Fig. 1C). Because milk fat is a major contributor to total solids, the parallel increase in these two traits suggests that the PRO-associated change in total solids was largely driven by the fat fraction. In contrast, APS supplementation significantly reduced milk fat percentage relative to baseline (p < 0.01; Fig. 1B), while total solids did not show a corresponding increase (Fig. 1C). Thus, unlike the PRO group, the APS group exhibited a compositional response characterized by lower milk fat percentage rather than enhanced accumulation of total milk solids. Protein and lactose showed more moderate responses. Milk protein percentage tended to increase in the APS group, although the difference was not statistically significant (+0.42%, Fig. 1D). Lactose percentage remained statistically stable after both supplementation treatments (Fig. 1E). As lactose is the main carbohydrate in milk and contributes substantially to osmotic equilibrium during milk secretion, stable lactose levels suggest that the aqueous-phase secretory characteristics of milk were largely maintained during the intervention period (Costa et al., 2019). Together, the stable lactose level and the non-significant upward tendency in protein percentage indicate that APS supplementation did not produce a broad decline in conventional milk solids, despite the reduction in milk fat percentage. MUN is commonly used as an indicator of nitrogen utilization in dairy animals, reflecting the balance among dietary protein supply, ruminal nitrogen metabolism, and fermentable energy availability (Hof et al., 1997; Zhao et al., 2025). In the present study, MUN increased markedly in the PRO group (p < 0.001; Fig. 1F), whereas APS supplementation significantly reduced MUN relative to baseline (−6.20 mg/dL, p < 0.05; Fig. 1F). Elevated MUN is generally associated with less efficient nitrogen utilization, whereas lower MUN within an appropriate physiological range may indicate improved nitrogen-use efficiency. Accordingly, the opposite MUN responses observed in the PRO and APS groups suggest that APS supplementation was associated with a more favorable nitrogen-use profile during late lactation. Collectively, the conventional milk composition data revealed two distinct response profiles (Fig. 1). Probiotic supplementation was associated with higher milk fat, total solids, and MUN, whereas APS supplementation reduced milk fat and MUN while maintaining lactose stability and showing a non-significant tendency toward higher protein percentage. These results suggest that APS supplementation produced a distinct conventional milk composition profile characterized by lower MUN and reduced milk fat percentage, providing the compositional basis for the subsequent E-nose and multi-omics analyses. E-nose sensor fingerprints of volatile-related milk profiles E-nose analysis was used to characterize supplementation-associated changes in the headspace sensor fingerprints of late lactation goat milk. The 10-sensor metal oxide semiconductor array captured broad response patterns across volatile-sensitive channels, providing an overview of the milk headspace fingerprint. Recent dairy studies have shown that E-nose profiling, especially when combined with sensory evaluation or chromatographic volatile analysis, can discriminate goat milk samples with different production backgrounds and reflect overall volatile-related milk characteristics (Balivo et al., 2023). In addition, fatty acid, volatile, and metabolite profiles of goat milk vary across lactation stages and are interrelated, supporting the relevance of volatile-related fingerprinting for interpreting late lactation milk quality (Liao et al., 2024). Radar charts showed clear differences in the overall sensor response patterns between the two supplementation strategies (Fig. 2A and B). In the PRO group, the endpoint profile remained broadly similar to its baseline profile, with only a modest expansion of the sensor response polygon (Fig. 2A). By contrast, the APS group showed a more pronounced expansion, particularly in W5S, W2W, and W1S (Fig. 2B). According to the nominal sensor sensitivities of the PEN3-Plus array, these sensors broadly respond to nitrogen oxide-related compounds, sulfur-containing/aromatic compounds, and methane-related or broad-range volatile classes, respectively. This pattern indicates that APS supplementation was associated with a stronger shift in the volatile-related sensor fingerprint than probiotic supplementation. E-nose profiles of milk before and after supplementation. A,B, Radar plots of the ten E-nose sensor responses for PRO versus Con A (A) and APS versus Con B (B). C, Principal component analysis (PCA) of the sensor response profiles. D, PCA loading plot showing the contribution of individual sensors. Sensor labels refer to the nominal sensitivity ranges of the PEN3-Plus array. Con A and Con B represent the baseline samples corresponding to the PRO and APS groups, respectively. The PCA plot further confirmed the divergence of E-nose response profiles among groups (Fig. 2C). The APS endpoint samples shifted clearly away from their baseline distribution and formed a distinct cluster, whereas the PRO endpoint samples were positioned closer to the central region of the score plot. This separation suggests that APS supplementation produced a more distinguishable volatile-related sensor profile during late lactation. Because E-nose analysis is based on cross-reactive sensor responses and multivariate pattern recognition, this separation should be interpreted as a difference in overall headspace fingerprint rather than direct identification of individual volatile compounds (Balivo et al., 2023). Loading vector analysis identified the sensor channels contributing most strongly to group separation (Fig. 2D). W5S was the dominant contributor, followed by W2W and W1S, indicating that nitrogen oxide-sensitive, sulfur/aromatic-sensitive, and broad-range volatile-sensitive channels were the principal drivers of the APS-associated E-nose profile. The co-occurrence of lower MUN in the APS group and a strong W5S contribution suggests that nitrogen-related compositional indicators and nitrogen oxide-sensitive sensor responses changed in parallel. However, the E-nose data support interpretation at the sensor-pattern level and should not be used alone to assign specific volatile compounds or aroma pathways. Proteomic landscape of milk molecular responses Milk proteomic profiling revealed distinct supplementation-associated response patterns between the PRO and APS groups. In the PCA score plot, APS endpoint samples were clearly separated from their corresponding baseline samples, whereas PRO endpoint samples remained closer to their baseline counterparts (Fig. 3A). This pattern suggests that APS supplementation produced a stronger shift in the milk proteome than probiotic supplementation during the 30-day intervention. The proteome mirror view further showed that APS and PRO supplementation generated both shared and treatment-specific differentially abundant proteins (Fig. 3B), indicating that the two nutritional strategies affected the milk proteome through distinguishable molecular response patterns. Changes in the milk proteome following probiotic and APS supplementation. A, PCA of milk protein profiles. B, Mirror plot of differentially abundant proteins in PRO versus Con A and APS versus Con B. C, KEGG enrichment of proteins with increased or decreased abundance in APS relative to Con B. D, Heatmap of the top 50 differentially abundant proteins. Heatmap values are protein-wise Z-scores. Six biological samples per group were included in the proteomic analysis. Con A and Con B represent the baseline samples corresponding to the PRO and APS groups, respectively. Functional enrichment analysis provided insight into the pathway-level characteristics of the APS-associated proteomic response. Proteins with increased abundance after APS supplementation were mainly annotated to starch and sucrose metabolism, aminoacyl-tRNA biosynthesis, glycolysis/gluconeogenesis, nucleotide metabolism, and sphingolipid signaling-related pathways (Fig. 3C). In contrast, proteins decreased after APS supplementation were enriched in cholesterol metabolism, O-glycan biosynthesis, glycosaminoglycan degradation, glycosaminoglycan biosynthesis, and mannose-type O-glycan biosynthesis pathways. These enrichment results should be interpreted as pathway annotations of differentially abundant milk proteins rather than direct evidence of intracellular pathway activation. Nevertheless, the results suggest that APS supplementation was associated with coordinated changes in milk proteins related to carbohydrate metabolism, amino acid utilization, nucleotide metabolism, glycan remodeling, and lipid-associated processes. The heatmap of the top 50 differentially abundant proteins further showed two major protein modules with opposite abundance patterns between APS and PRO samples (Fig. 3D). The PRO-associated module contained proteins such as fibrinogen gamma chain (FGG), complement component 9 (C9), biglycan (BGN), galectin 3 binding protein (LGALS3BP), lipoprotein lipase (LPL), fibronectin 1 (FN1), and polymeric immunoglobulin receptor (PIGR), which are linked to extracellular matrix organization, immune-related processes, transport, or lipid handling. In contrast, the APS-associated module included heat shock protein 90 beta family member 1 (HSP90B1), calreticulin (CALR), protein disulfide isomerase family A member 3 (PDIA3), protein disulfide isomerase family A member 4 (PDIA4), calnexin (CANX), asparagine synthetase (ASNS), hypoxanthine phosphoribosyltransferase 1 (HPRT1), lactate dehydrogenase A (LDHA), and glyceraldehyde-3-phosphate dehydrogenase (GAPDH), indicating enrichment of proteins related to protein folding, cellular metabolic processes, amino acid biosynthesis, nucleotide metabolism, and glycolysis. This reciprocal pattern supports the presence of treatment-specific milk protein signatures and provides a molecular context for the divergent milk quality responses observed after APS and probiotic supplementation. Milk proteins, amino acid-related pathways, carbohydrate metabolism, and lipid-associated processes can contribute to the pool of biochemical precursors relevant to dairy quality and flavor-related traits, particularly through downstream interactions among proteins, amino acids, carbohydrates, and lipids in dairy matrices (Zheng et al., 2021). The APS-associated enrichment of carbohydrate metabolism, amino acid-related pathways, and sphingolipid-associated features therefore provides a proteomic framework for interpreting the metabolite and lipid changes described below. Accordingly, these findings are best interpreted as supplementation-associated milk protein signatures rather than direct evidence of mammary intracellular metabolic regulation. Targeted inspection of lipid metabolism-related proteins further refined this pattern (Fig. 4). Compared with its baseline control, the APS group showed higher abundance of acetyl-CoA carboxylase alpha (ACACA), fatty acid synthase (FASN), and ATP citrate lyase (ACLY), with fold changes of 1.49, 1.70, and 2.52, respectively, whereas stearoyl-CoA desaturase (SCD) showed only a modest change. ACACA, FASN, and ACLY are closely related to acetyl-CoA supply, malonyl-CoA formation, and fatty acid biosynthesis, and previous studies have linked these enzymes and related acetyl-CoA-generating pathways to milk fat synthesis in dairy animals and caprine mammary epithelial models (Lalonde et al., 2024; Mu et al., 2021; Tian et al., 2022; Xu et al., 2018). Thus, the higher abundance of ACACA, FASN, and ACLY points to an APS-associated enrichment of lipid-synthesis-related protein features in milk. Abundance of lipid metabolism-related proteins in milk. Left, Z-score heatmap of proteins related to fatty acid synthesis (SCD, ACACA, FASN and ACLY), NADPH generation (IDH1, G6PD and PGD), and lipid uptake (LPL). Right, log2 fold changes for the same proteins in PRO versus Con A and APS versus Con B; numbers indicate fold-change ratios. P < 0.05, P < 0.01 and P < 0.001. n = 6 biological samples per group. APS supplementation also increased the abundance of proteins linked to reducing-equivalent supply, including isocitrate dehydrogenase 1 (IDH1), glucose-6-phosphate dehydrogenase (G6PD), and 6-phosphogluconate dehydrogenase (PGD), with fold changes of 1.30, 2.21, and 2.29, respectively (Fig. 4). These enzymes contribute to NADPH-generating pathways that can support lipid biosynthesis (Gelman et al., 2018; Laliotis et al., 2010). Together with the changes in ACACA, FASN, and ACLY, this pattern indicates enrichment of lipid metabolism-related protein features in the APS milk proteome. In contrast, LPL abundance decreased markedly in the APS group, with a fold change of 0.26, suggesting a shift in lipid-processing-related protein composition rather than a generalized increase in all lipid-associated proteins. The PRO group showed smaller changes in the same protein set, including mild increases in G6PD, PGD, and LPL and a decrease in IDH1, further supporting the distinct proteomic response induced by APS. Metabolomic fingerprinting of milk molecular responses Metabolomic profiling identified 904 differential metabolites in the APS group and 587 in the PRO group (VIP > 1 and FDR-adjusted p < 0.05) (Fig. 5), indicating that APS supplementation was associated with a broader milk metabolomic response than probiotic supplementation. Among the detected differential features, lipid-related metabolites and organic acids represented prominent responsive categories, suggesting that APS supplementation influenced metabolite pools related to lipid-derived and central‑carbon-associated milk chemistry. These changes are relevant to late lactation goat milk quality because fatty acid profiles, volatile compounds, and metabolites in goat milk vary across lactation stages and show interconnected patterns (Liao et al., 2024). The representative metabolomic OPLS-DA model showed R2X(cum), R2Y(cum), and Q2(cum) values of 0.647, 0.997, and 0.983, respectively, and 200-times permutation testing yielded negative Q2 intercepts (−0.5718 and − 0.3621), indicating no obvious overfitting of the supervised metabolomic model. Metabolomic profiles of milk following probiotic and APS supplementation. A, Orthogonal partial least-squares discriminant analysis (OPLS-DA) score plots for APS versus Con B and PRO versus Con A in positive and negative ion modes. B, Mirror volcano plot of differential metabolites in PRO versus Con A and APS versus Con B. Grey points indicate metabolites that did not meet the differential-analysis criteria. C, Chemical classes of differential metabolites. Differential metabolites were defined by VIP > 1 and Benjamini–Hochberg FDR-adjusted P < 0.05. OPLS-DA models were evaluated by sevenfold cross-validation and 200 permutation tests. Con A and Con B represent the baseline samples corresponding to the PRO and APS groups, respectively. The APS-associated enrichment of organic acids, including tricarboxylic acid cycle-related intermediates such as citrate and malate, indicates changes in milk organic-acid composition rather than direct evidence of intracellular mitochondrial flux. Organic acids can contribute to the chemical background of dairy matrices and may interact with amino acid-, carbohydrate-, and lipid-derived compounds during downstream quality formation. In goat dairy systems, integrated metabolomic and lipidomic analyses have shown that changes in organic acids, peptides, fatty acids, and lipid species can be associated with sensory-quality variation (Jia et al., 2021; R. Zhang & Jia, 2023). Therefore, the organic-acid shifts observed here provide a milk-level metabolic context for interpreting the concurrent proteomic and lipidomic responses. In addition to organic acids, the broader APS-associated metabolomic response suggests changes in the pool of potential flavor- and quality-related precursors. Milk proteins, carbohydrates, amino acids, and lipids provide biochemical substrates for downstream flavor-related reactions and quality traits in dairy matrices (Zheng et al., 2021). Food and dairy metabolomics studies further support the value of metabolite profiling for connecting precursor composition with volatile-related and sensory-quality differences (Diez-Simon et al., 2019; Jia et al., 2021; Zhang & Jia, 2023). Accordingly, the metabolomic profile observed after APS supplementation is best interpreted as a milk-derived molecular fingerprint reflecting altered precursor availability, rather than direct proof of a specific metabolic pathway being restored or activated. Together with the E-nose and proteomic results, these metabolomic data suggest that APS supplementation produced a coordinated shift in milk molecular composition, providing a metabolic basis for the lipidomic remodeling described below. Lipidomic remodeling of milk lipid composition Lipidomic profiling identified 2553 lipid molecules spanning 42 subclasses, providing a high-resolution map of goat milk fat composition. Fig. 6 illustrates the global lipid landscape, in which triacylglycerols (TGs) remained the dominant lipid class, accounting for more than 88% of total detected lipid signal. This distribution is consistent with the known predominance of TGs in mammalian milk lipids (Mitina et al., 2020). Although the overall TG-dominant architecture was maintained, the relative composition of non-TG lipid subclasses differed among groups (Fig. 6B). In particular, APS supplementation was associated with a 1.12-fold increase in total TG signal compared with its baseline control, together with visible changes in selected polar lipid and sphingolipid subclasses. Because lipidomics reflects molecular-intensity profiles rather than bulk fat concentration, these results indicate lipid compositional remodeling rather than an absolute increase in total milk fat. The representative lipidomic OPLS-DA model showed R2X(cum), R2Y(cum), and Q2(cum) values of 0.589, 0.999, and 0.987, respectively, and 200-times permutation testing yielded a negative Q2 intercept (−0.1086), indicating no obvious overfitting of the supervised lipidomic model. Lipid-class composition of milk. A, Relative contribution of triacylglycerols (TGs) and other detected lipids. B, Relative composition of non-TG lipid subclasses after normalization to total non-TG lipid intensity. Con A and Con B represent the baseline samples corresponding to the PRO and APS groups, respectively. DG, diacylglycerol; CerP, ceramide phosphate; PE, phosphatidylethanolamine; SM, sphingomyelin; Hex1Cer, hexosylceramide; SPH, sphingosine; PC, phosphatidylcholine; WE, wax ester; MG, monoacylglycerol; ZyE, zymosterol ester. The heatmap of the top 30 medium-chain TG species further highlighted the specificity of the APS-associated lipid response (Fig. 7). Most selected TG species containing medium-chain acyl moieties, including C6:0, C8:0, and C10:0-containing combinations, showed higher relative abundance in the APS group than in the other groups. Medium-chain fatty acid-containing lipids are characteristic components of goat milk and contribute to the pool of lipid-derived precursors relevant to goat milk flavor attributes after lipolysis (Chen et al., 2021; Liao et al., 2024; Y. Zhang et al., 2023). Recent goat milk studies further support that lipid compositional changes are closely associated with flavor-related quality variation in goat dairy matrices (Li et al., 2024; Liao et al., 2024). In the present study, the enrichment of selected medium-chain TG species indicates that APS supplementation altered TG molecular composition within the milk lipidome, particularly at the level of chain-length distribution. Selected triacylglycerol species in milk. Heatmap of 30 medium-chain TG species across the four groups. Columns represent individual biological samples and rows represent TG species containing medium-chain acyl groups, including C6:0, C8:0 and C10:0. Values are row-scaled Z-scores, with red and blue indicating higher and lower relative abundance, respectively. The white vertical line separates the probiotic arm (Con A and PRO) from the APS arm (Con B and APS). n = 6 biological samples per group. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Subclass-level analysis of non-TG lipids revealed a coordinated redistribution of membrane-associated lipid features (Fig. 8). Compared with its baseline control, APS supplementation increased the abundance of phosphatidylethanolamine (PE, approximately 1.4-fold) and phosphatidylcholine (PC, approximately 1.5-fold). PC and PE are major polar lipids of the milk fat globule membrane (MFGM) and contribute to the interfacial organization and physical stability of milk fat globules (Contarini & Povolo, 2013;Liu et al., 2020; Lopez et al., 2019; Nie et al., 2024). The higher abundance of PE and PC therefore suggests an APS-associated enrichment of MFGM-related phospholipid features. In addition, ceramide (Cer) increased markedly, by nearly 5-fold, whereas sphingomyelin (SM) showed a relative decrease in the APS group (Fig. 8). Since SM, Cer, and other sphingolipid species are minor but functionally important components of milk lipid organization and the MFGM environment, this inverse SM–Cer pattern is best interpreted as sphingolipid compositional remodeling within the milk matrix rather than direct evidence of sphingolipid pathway activation (Liu et al., 2020; Lopez et al., 2019; Nie et al., 2024). Relative abundance of major non-TG lipid subclasses. Relative abundance of 12 non-TG lipid subclasses normalized to total non-TG lipid intensity. Data are shown as mean ± s.d. (n = 6 biological samples per group). Comparisons shown in the figure were assessed using two-sided Student's t-tests. P < 0.05, P < 0.01 and P < 0.001. Con A and Con B represent the baseline samples corresponding to the PRO and APS groups, respectively. DG, diacylglycerol; CerP, ceramide phosphate; PE, phosphatidylethanolamine; SM, sphingomyelin; Hex1Cer, hexosylceramide; SPH, sphingosine; PC, phosphatidylcholine; WE, wax ester; MG, monoacylglycerol; ZyE, zymosterol ester; Cer, ceramide; PS, phosphatidylserine. The lollipop plot of the top 20 differential lipids further highlighted the specificity and magnitude of the APS-associated lipidomic response (Fig. 9). In the APS versus baseline comparison, several phospholipid and ceramide species showed strong positive log₂ fold changes, including phosphatidic acid PA(39:2), phosphatidylcholine PC(18:0e/22:4), phosphatidylglycerol PG(18:2/20:4), lysophosphatidylcholine LPC(18:1e), ceramide Cer(t16:1/26:1), Cer(t16:0/26:1), and phosphatidylserine PS(18:0e/18:1)-1. By contrast, multiple sphingomyelin species, including SM(d30:1), SM(d32:1)-2, SM(d34:2), SM(d36:2), SM(d14:0/20:4), SM(d35:1)-1, and SM(d39:2), showed strong negative log₂ fold changes. The PRO versus baseline comparison showed fewer lipid species with large-magnitude changes, suggesting a milder lipidomic response to probiotic supplementation. Together with the subclass-level results, these species-level changes indicate that APS supplementation was associated with selective remodeling of phospholipid and sphingolipid features in milk. Lollipop plots showing log2 fold changes of the top 20 differential lipid species in APS versus Con B (left) and PRO versus Con A (right). Opaque circles indicate significant changes, whereas pale circles indicate non-significant changes. P values were adjusted using the Benjamini–Hochberg method. FDR < 0.05, FDR < 0.01 and FDR < 0.001. Con A and Con B represent the baseline samples corresponding to the PRO and APS groups, respectively. Taken together, Fig. 6, Fig. 7, Fig. 8, Fig. 9 indicate that APS supplementation was associated with selective remodeling of the milk lipidome, involving a modest increase in total TG signal, enrichment of selected medium-chain TG species, increased MFGM-related phospholipids such as PE and PC, and a shift in sphingolipid composition characterized by lower SM and higher Cer. These lipidomic changes complement the proteomic patterns described above and support a milk-level molecular signature underlying the distinct compositional response observed after APS supplementation. Their potential relevance to lipid stability and volatile-related milk properties warrants further validation using absolute lipid quantification, targeted volatile analysis, and direct assessment of MFGM structure. Integrated multi-omics associations among milk protein, lipid, metabolite, and E-nose features To further explore the relationships among supplementation-associated milk molecular features, we performed an integrated correlation analysis linking proteomic, lipidomic, metabolomic, and E-nose variables (|r| > 0.6 and FDR < 0.05; Fig. 10). Network-based multi-omics integration provides a useful framework for interpreting coordinated relationships across different data layers, while such networks should be viewed primarily as association-based rather than causal evidence (Bodein et al., 2022). The resulting network showed that the APS-associated milk profile was characterized by coordinated variation among lipid-related proteins, polar lipid subclasses, selected metabolite features, and volatile-related sensor responses. Rather than indicating a single linear regulatory pathway, this pattern suggests that APS supplementation was associated with a coherent milk-level molecular signature involving lipid metabolism-related proteins and lipid subclass redistribution. Correlation network linking milk proteins, lipids and E-nose responses. Significant Spearman correlations among selected protein features, lipid species and E-nose sensors are shown. Red and blue edges indicate positive and negative correlations, respectively, and edge width reflects |r|. Associations were retained at |r| > 0.6 and FDR < 0.05. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Several lipid-related proteins showed strong associations with milk lipid species. G6PD and FASN were positively correlated with PE- and Cer-related lipid features (r > 0.85), whereas these correlations were weaker or absent for storage lipid-dominated features. This pattern is consistent with the proteomic and lipidomic results described above, in which APS supplementation was associated with higher abundance of lipid-synthesis-related proteins and enrichment of selected polar lipid subclasses. The preferential association with PE- and Cer-related features is noteworthy because polar lipids and sphingolipids, including phosphatidylethanolamine, phosphatidylcholine, sphingomyelin, and ceramide-related species, contribute to milk lipid organization and the milk fat globule membrane environment (Contarini & Povolo, 2013; Liu et al., 2020; Lopez et al., 2019; Nie et al., 2024). In contrast, LPL was negatively correlated with Cer(d17:1/16:0)-2 (r = −0.96), suggesting an inverse association between this milk lipid-processing protein and specific ceramide species. Together, these correlations indicate a milk-derived co-abundance pattern linking lipid-related proteins with selected polar lipid and ceramide features. The integration of the lipidomic and E-nose datasets further indicated that selected polar lipid species were associated with volatile-related sensor responses. For example, PC(18:0e/22:4) and oxidized Cer(d41:0 + O) were positively correlated with W5S- and W2W-related sensor responses (r = 0.83), linking milk polar lipid features with nitrogen oxide-sensitive and sulfur/aromatic-sensitive sensor channels. Recent goat milk studies have shown that E-nose profiling can discriminate milk samples with different production backgrounds by capturing overall headspace fingerprints, especially when interpreted together with complementary sensory or volatile analyses (Balivo et al., 2023). Therefore, these correlations support interpretation at the coordinated-pattern level between milk lipid composition and sensor responses, rather than compound-level aroma identification. This interpretation is biologically plausible because lipid oxidation and lipid-derived reactions are recognized contributors to volatile formation in dairy matrices, whereas the polar-lipid environment of milk fat globules may provide a physicochemical context for lipid-derived volatile-related features (Clarke et al., 2021; Lopez et al., 2019; Nie et al., 2024). The integration of these datasets supports a model in which APS supplementation is associated with coordinated milk-level remodeling across lipid-related proteins, metabolite features, polar lipid species, and volatile-related E-nose signals. This integrated signature is consistent with the conventional milk composition results, the broader metabolomic response observed in the APS group, and the lipidomic enrichment of PE, PC, and Cer subclasses. Therefore, the multi-omics data provide a hypothesis-generating framework suggesting that APS may modulate late lactation goat milk quality through coordinated changes in nitrogen-use indicators, lipid-related molecular features, and volatile-related sensor fingerprints. Integrated interpretation and future perspectives The integrated milk-based data indicate that APS supplementation was associated with coordinated changes in conventional milk composition, E-nose sensor responses, and milk proteomic, metabolomic, and lipidomic features during late lactation. Compared with probiotic supplementation, APS produced a distinct milk-quality response characterized by reduced MUN and milk fat percentage, stable lactose, a tendency toward higher milk protein percentage, and coordinated shifts in lipid-related molecular signatures. These findings suggest that APS may serve as a promising non-microbial nutritional strategy for modulating late lactation goat milk quality. From an integrated perspective, the APS-associated response appears to reflect milk-level compositional remodeling involving nitrogen-use indicators, volatile-related sensor fingerprints, lipid metabolism-related protein features, and lipid subclass redistribution. This interpretation is supported by the lower MUN observed in the APS group, the distinct E-nose response pattern dominated by W5S, W2W, and W1S, the enrichment of lipid-synthesis-related milk protein features, and the lipidomic remodeling of medium-chain TGs, phospholipids, and sphingolipids. Together, these findings indicate that APS supplementation was associated with a coordinated milk-quality signature rather than an isolated change in a single compositional trait. Several aspects of the study design should be considered when interpreting these findings. The absence of a basal-diet-only control group limits the separation of supplementation-associated responses from natural temporal variation during late lactation. In addition, because the omics analyses were conducted using milk rather than mammary tissue, the molecular signatures observed here should be regarded as milk-derived associations rather than direct evidence of intracellular mammary regulation. Similarly, E-nose profiling provides sensor-level volatile fingerprints rather than identification of individual aroma compounds. Accordingly, the present study provides an association-based multi-omics framework for understanding APS-related milk quality responses, while targeted validation will be needed to clarify the underlying biological mechanisms. Future studies incorporating untreated controls, larger independent cohorts, targeted volatile analysis, absolute lipid quantification, and mammary tissue-based validation will be important for confirming the treatment-specific effects of APS and clarifying the biological processes underlying the observed milk-quality responses. Conclusion This study demonstrates that APS supplementation was associated with a distinct milk-quality response in late lactation goats compared with probiotic supplementation. APS reduced MUN and milk fat percentage while maintaining lactose stability and showing a tendency toward higher milk protein percentage. E-nose profiling revealed an APS-associated sensor-level volatile fingerprint, mainly characterized by W5S, W2W, and W1S responses. Milk proteomics, metabolomics, and lipidomics further showed coordinated molecular changes, including enrichment of lipid-synthesis-related protein features, a broader metabolomic response, and selective remodeling of medium-chain TGs, MFGM-related phospholipids, and sphingolipid subclasses. Together, these milk-based findings suggest that APS may modulate late lactation goat milk quality through coordinated changes in nitrogen-use indicators, volatile-related sensor profiles, and lipid-related molecular features. Overall, APS represents a promising plant-derived, non-microbial nutritional strategy for modulating the quality profile of late lactation goat milk. CRediT authorship contribution statement Shanshan Han: Writing – original draft, Methodology, Funding acquisition, Formal analysis, Data curation. Peng Wang: Visualization, Supervision, Software. Yunhong Hu: Visualization, Validation. Qiang Zhang: Visualization, Validation. Ke Wan: Resources, Investigation, Funding acquisition. Xiaoyu Wang: Writing – review & editing, Supervision, Resources, Methodology, Funding acquisition. Funding sources This work was supported by the Science and Technology Program of Shaanxi Academy of Sciences (program No. 2026K-17), the National Natural Science Foundation of China (323B2058), Qingdao Municipal Natural Science Foundation (program No. 24-4-4-zrjj-108-jch) and the Science and Technology Department of Shaanxi Province (program No. 2025NC-YBXM-122, 25NJSYB00048). 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. Acknowledgements The authors would like to thank all members of our laboratory for their assistance. We extend our gratitude to Shaanxi Xincheng Datang Bio-Tech for their kind support in our field experiments. We are also grateful to Tgene Biotech (Shanghai) Co., Ltd. for providing technical support with the multi-omics detection and analysis. We thank Xia Ming from Xinshi Subdistrict, Lintong District, for assistance in locating the sampling source. Footnotes Supplementary data to this article can be found online at https://doi.org/10.1016/j.fochx.2026.104395. Supplementary data Supplementary material Data availability Data will be made available on request. References Associated Data Supplementary Materials Supplementary material Data Availability Statement Data will be made available on request.