Identification of candidate genes and metabolites associated with lactation performance in Kazakh mares using blood multi-omics and machine learning

Abstract

Background

The Kazakh mare, an indigenous breed of Xinjiang, exhibits strong adaptability to arid and cold conditions while maintaining relatively stable milk production under low-input extensive farming systems. However, its genetic improvement has been constrained by traditional management practices.

Results

In this study, we monitored milk yield and milk composition over a 105-day lactation period and recorded 15 phenotypic traits. Milk yield was significantly correlated with body length, teat diameter, and teat length. The high-yield (HY) versus low-yield (LY) and high-fat (HF) versus low-fat (LF) comparisons identified 286 and 627 differentially expressed genes (DEGs), respectively. Several candidate genes were identified, including PMP22, FAM83A, HSD17B3, AGPAT4, SLC50A1, and ERBB3, which were associated with pathways including PI3K-Akt signaling, MAPK signaling, and triglyceride metabolism. High-yield mares showed metabolic differences characterized by enrichment of pathways related to the tricarboxylic acid (TCA) cycle, suggesting altered energy and intermediary metabolism involving carbohydrates, lipids, and amino acids. Metabolites associated with these differences included glycerone, α-D-glucose, D-galactose, glycerol, L-histidine, and anserine.

In addition, machine learning analysis identified GLDC as a candidate gene potentially associated with milk fat percentage, possibly through its association with histidine. However, this relationship requires further validation.

Conclusion

Using peripheral blood samples, this study integrated differential expression analysis, mixed linear models, and machine learning approaches to identify candidate genes and metabolites associated with lactation performance in Kazakh mares. The results provide preliminary insights into molecular and metabolic features associated with lactation traits in this breed. These findings may serve as a reference for future molecular breeding and nutritional studies.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12864-026-13131-w.

Introduction

Globally, dairy mare resources are clustered in Eurasian countries such as Russia, Kazakhstan, Kyrgyzstan, and China [1]. Kazakhstan has the largest mare population, while in China, Xinjiang is the main production region, followed by Inner Mongolia [2]. Mare’s milk has gained attention for its distinctive nutritional profile, closely resembling human milk in composition, including fat (0.40% ~ 1.90%), protein (1.70% ~ 2.20%), lactose (5.80% ~ 7.00%), and total solids (9.30% ~ 9.85%) [3]. Its essential amino acid (EAA) composition also meets FAO/WHO criteria for ideal protein [4]. As China’s consumer market evolves, mare’s milk products are increasingly valued for their superior nutrition, driving growth in the premium dairy sector and offering strategic potential for resource development and utilization.

Blood serves as a key integrative medium reflecting the physiological states of multiple organs. Through the systemic transport of hormones, metabolites, and signaling molecules, it plays a central role in regulating mammary gland development and lactation [5]. As a minimally invasive and easily accessible biological specimen, blood offers stable molecular components that hold promise as biomarkers for elucidating the genetic architecture of lactation traits. Recent advances in integrated transcriptomic and metabolomic analyses have provided valuable insights into the molecular regulation of lactation and nutritional requirements in species such as dairy cattle [6–9]. However, the genetic basis of lactation in equids remains largely unexplored. This knowledge gap highlights the urgent need for in-depth studies on mare milk production, which are essential for the genetic improvement and strategic utilization of dairy mare resources. Machine learning (ML) techniques have proven highly effective in identifying phenotype-associated molecular features, including key regulatory genes [10] and functional microbial communities [11]. To address the challenges of high dimensionality, small sample sizes, multicollinearity, and limited generalizability, methods such as least absolute shrinkage and selection operator (LASSO) regression [12, 13], random forest (RF) regression [14, 15], and support vector regression (SVR) [16, 17] offer distinct advantages. Although ML has made notable progress in biomedical fields such as clinical decision support and personalized medicine, its application in agricultural biosciences remains limited. In particular, the use of ML to predict and interpret complex traits like milk yield and composition in dairy livestock, is still at an early stage. Thus, incorporating ML methods into phenotypic studies of dairy animals not only offers practical benefits but also provides innovative insights into the molecular regulatory networks of lactation traits.

The Kazakh mare, an economically significant equine breed in Xinjiang, is predominantly distributed in the northern Tianshan Mountains, the margins of the Junggar Basin, and the western Altai Mountains. Renowned for its tolerance to arid and cold environments [2, 18], this breed maintains stable milk production under extensive grazing conditions, with daily yields ranging from 7.00 to 9.00 kg in low-producing individuals and 13.00 to 15.00 kg in high-producing individuals. Although it provides a basic source of income for local herders, its overall productivity remains inferior to that of specialized dairy breeds, and the traditional pastoral management system imposes substantial constraints on systematic genetic improvement. To address these challenges, the present study integrates transcriptomic and metabolomic profiling with multi-omics association analyses and machine learning approaches. Specifically, we aim to elucidate the molecular differences among Kazakh mares with divergent lactation phenotypes, identify key candidate genes and metabolites, and construct a gene–metabolite–phenotype interaction network. This integrative framework reveals the regulatory mechanisms underlying lactation performance and provides theoretical foundations as well as candidate molecular targets for marker-assisted selection to improve milk-related traits in the Kazakh horse breed.

Materials and methods

Experimental design

This study selected 30 healthy lactating Kazakh mares from a private farm operated by the Akekule Horse Industry Company in Altay City, Xinjiang, China. All mares were in their second to fifth parity. Lactation performance of the entire herd was systematically monitored from 30 to 105 days postpartum during the 2024 lactation season (May to September). For multi-omics analysis, 24 mares at peak lactation (approximately 100–110 days postpartum) were selected from the monitored population. Based on the on-site lactation performance data collected in 2024, these 24 mares were divided into two comparisons: a high-yield versus low-yield (HY vs. LY) comparison and a high-fat versus low-fat (HF vs. LF) comparison, with 12 mares in each group. The average daily milk yield was 7.92 ± 0.56 kg in the HY group and 4.53 ± 1.08 kg in the LY group. In the milk fat percentage comparison, the HF group showed an average milk fat percentage of 2.09% ± 0.25%, while the LF group showed 0.84% ± 0.06%. To evaluate the representativeness of the lactation performance in the present study population, historical monitoring data from the same farm between 2021 and 2023 were also referenced. Detailed information on each mare’s parity, sampling year, and exact days postpartum is provided in Table 1.

Table 1.

Summary of mare information and lactation data

Mares	Chip ID	Lactation days	Parity	Sampling time
HY1	900,002,299,900,574	103	3	June 29, 2024
HY2	900,002,266,601,673	110	5	June 29, 2024
HY3	900,115,000,156,271	100	5	June 29, 2024
HY4	900,002,266,601,579	103	5	June 29, 2024
HY5	900,002,299,900,546	109	4	June 29, 2024
HY6	900,002,299,900,572	103	3	June 29, 2024
LY1	900,002,266,601,566	100	2	June 30, 2024
LY2	900,002,266,601,569	103	4	June 30, 2024
LY3	900,002,266,601,580	103	3	June 30, 2024
LY4	900,002,299,900,557	109	3	June 30, 2024
LY5	900,002,266,601,101	104	5	June 30, 2024
LY6	900,002,299,900,562	109	2	June 30, 2024
HF1	900,002,299,900,568	109	2	July 1, 2024
HF2	900,002,299,900,580	103	5	July 1, 2024
HF3	900,002,266,601,678	103	3	July 1, 2024
HF4	900,002,299,900,570	102	3	July 1, 2024
HF5	900,002,266,601,671	102	2	July 1, 2024
HF6	900,002,266,601,548	109	4	July 1, 2024
LF1	900,002,299,900,552	100	4	July 3, 2024
LF2	900,002,266,601,222	105	3	July 3, 2024
LF3	900,002,266,601,568	101	5	July 3, 2024
LF4	900,002,266,601,561	109	5	July 3, 2024
LF5	900,002,266,601,668	100	3	July 3, 2024
LF6	900,002,266,601,679	100	4	July 3, 2024

Animal management

During the experimental period, all mares were kept under a standardized feeding and management system. Each mare was provided with 28 kg of alfalfa hay daily (7 kg × 4 feedings) and 6 kg of concentrate daily (3 kg × 2 feedings). The detailed ingredient composition of the concentrate diet is shown in Table 2. To minimize variation in feed intake, mares were housed individually in stalls equipped with dedicated feed troughs. Following the daytime milking sessions, mares and foals were turned out to pasture in the evening for free grazing and exercise. All procedures throughout the experiment were conducted in strict accordance with the policies and guidelines of the Animal Ethics Committee of Xinjiang Agricultural University.

Table 2.

Composition of supplemental concentrates

Ingredients	Content(%)
Corn	51.52
Soybean meal	18.08
Bran	26.78
Lysine	1.20
Vitamin	1.90
Salt	0.52
Total	100

Mares are supplemented with 3 kg twice a day; once at 7:00 am and once at 13:00

Sample collection

Standardized machine milking was performed four times daily at 10:00, 13:00, 16:00, and 19:00. Milk yield from each milking session was immediately measured using a portable electronic scale (± 0.01 kg accuracy), and daily milk yield was calculated as the sum of the four milkings. Blood samples were collected from the jugular vein 30 min prior to the morning milking. Milk samples (80 mL per milking) were collected during each of the four daily milking sessions. Both blood and milk samples were collected at 10-day intervals throughout the 105-day monitoring period. For transcriptomic analysis, whole blood was immediately mixed with TRIzol reagent at a 1:3 ratio, thoroughly mixed, and snap-frozen in liquid nitrogen. For metabolomic analysis, blood samples were centrifuged at 2,350 × g for 10 min at 4 °C. All samples were ultimately transferred to a − 80 °C ultra-low temperature freezer for long-term storage.

Milk composition (fat, protein, and lactose percentages) was analyzed using a MilkoScan™ FT3 Milk Analyzer (Foss Electric, Denmark) on thawed samples in the laboratory.

In addition, fifteen linear body conformation traits were measured for each mare one hour before the morning milking, while the mares were restrained in a stock. Body height, body length, chest girth, cannon bone circumference, chest width, maximum head width, forequarter width, barrel circumference, hindquarter width, rump width, udder length, udder width, udder depth, teat diameter, and teat length were recorded using standardized procedures with a measuring stick and tape.

RNA sequencing, raw data filtering, and read mapping

RNA concentration and purity were assessed using NanoDrop and Qubit fluorometers, while total RNA quantity and integrity were evaluated with an Agilent Bioanalyzer. Samples were deemed qualified if total RNA yield exceeded 0.6 µg and RNA integrity number (RIN) values were greater than 6. A total of 24 RNA-seq libraries (HY1 ~ HY6, LY1 ~ LY6, HF1 ~ HF6, LF1 ~ LF6) were constructed from qualified samples by Novogene Bioinformatics Technology Co., Ltd. (Beijing, China). Poly(A) + mRNA was enriched using magnetic oligo(dT) beads, followed by fragmentation. First-strand cDNA synthesis was performed using random hexamer primers and dNTPs under enzymatic catalysis, and second-strand cDNA synthesis completed the library construction. After quantification and quality control of the libraries, paired-end sequencing (2 × 150 bp) was conducted on the Illumina NovaSeq 6000 platform. Raw sequencing reads were processed using Fastp (v0.23.2) to remove low-quality sequences (Q20 < 95%), adapter contamination, and reads with > 5% ambiguous bases (N content), yielding high-quality clean reads. These were aligned to the horse reference genome (GCF_002863925.1_EqCab3.0_novel) using HISAT2 (v2.2.1). Gene-level quantification was performed with FeatureCounts (v1.5.0-p3), and gene expression levels were normalized and reported as FPKM (Fragments per kilobase of transcript per million mapped reads).

Metabolite extraction and annotation

Plasma samples were separated using a Vanquish ultra-high-performance liquid chromatography (UHPLC) system (Thermo Fisher Scientific) equipped with a Hypersil GOLD C18 column. Chromatographic separation was performed at a column temperature of 40 °C with a flow rate of 0.2 mL/min. In positive ion mode, mobile phase A consisted of 0.1% formic acid in water, and mobile phase B was methanol. In negative ion mode, mobile phase A was a 5 mmol/L ammonium acetate aqueous solution, while mobile phase B remained methanol. To enable comprehensive interpretation of the identified metabolites, functional annotation was performed using multiple public databases, including the Kyoto Encyclopedia of Genes and Genomes (https://www.genome.jp/kegg/, accessed November 24, 2024), the Human Metabolome Database (https://hmdb.ca/metabolites/, accessed November 24, 2024), and LIPID Maps (http://www.lipidmaps.org/, accessed November 24, 2024).

Correlation analysis between lactation performance and body conformation traits

To investigate the relationship between lactation performance and body conformation traits in Kazakh mares, a series of statistical analyses were conducted. Initially, milk yield and milk fat percentage were analyzed using one-way analysis of variance (ANOVA) in GraphPad Prism software (v9.4.1) to assess overall comparison differences. Subsequently, Student’s t-tests (P < 0.05) were applied to identify statistically significant differences in lactation performance between specific comparisons. After confirming the presence of significant comparison differences, Pearson correlation analysis was performed using R software to evaluate the linear associations between lactation performance indicators (milk yield and milk fat percentage) and 15 body conformation traits.

Identification of DEGs

Prior to DEGs screening, principal component analysis (PCA) was conducted to evaluate sample clustering and variability. Hierarchical clustering of gene expression profiles was then performed to provide a global overview of gene expression patterns among the experimental comparisons. Differential expression analysis between HY vs. LY and HF vs. LF comparisons were conducted using the DESeq2 package (v1.50.2). Genes with an absolute |log₂ fold change| ≥ 1, P < 0.005 and P_adj < 0.2 were considered differentially expressed. Identified DEGs were further analyzed using Venn diagrams to identify overlapping genes across comparisons.

Functional enrichment analysis of DEGs

Based on Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) database annotations, functional enrichment analysis of DEGs in Kazakh mares was conducted using the R package clusterProfiler. All detected genes were used as the background dataset, and up-regulated and down-regulated DEGs were analyzed independently. Significantly enriched GO terms and KEGG pathways were identified using a threshold of P_adj < 0.05. pathway-based functional clustering was further performed to identify biological processes and pathways associated with lactation performance.

Comparative analysis with animal QTL database

To identify potential functional genes, DEGs identified in this study were cross-referenced with lactation-related genes reported in the Animal QTL database (https://www.animalgeno.org/cgi-bin/QTLdb/, accessed November 30, 2024), including loci associated with milk yield and milk composition traits in cattle, sheep, goats, and pigs. Genes overlapping between the DEGs and lactation-associated QTL genes were retained for subsequent analyses.

To assess the association between these overlapping genes and lactation traits in Kazakh mares, including milk yield, milk fat percentage, milk protein percentage, lactose percentage, and the yields of milk fat, protein, and lactose, mixed linear models (MLMs) were fitted using the lme4 package in R. The models were specified as: Y = Xβ + Zγ + ε, where Y represents the lactation phenotype, β denotes fixed effects including gene expression level (gene expression levels were log-transformed before analysis), parity, age, and lactation days, γ represents random effects accounting for individual mares, and ε denotes the residual error. This modeling framework was used to control for individual heterogeneity and potential confounding factors. Data processing, visualization, and statistical analyses were conducted in R. Graphical outputs, including p-value distribution plots and linear regression plots, were generated using ggplot2, while ggbeeswarm was used to visualize data distributions. Data restructuring and formatting were performed using reshape2, dplyr, and tidyr, and graphical scaling and color schemes were adjusted using scales and RColorBrewer, respectively.

To account for multiple testing, p-values obtained from the mixed linear models were adjusted for false discovery rate (FDR) using the Benjamini-Hochberg (BH) method implemented in the p.adjust() function of the stats package in R. Genes with adjusted p-values (P_adj < 0.05) for at least one lactation trait were considered candidate genes. These genes were then subjected to GO functional annotation and KEGG pathway enrichment analysis using clusterProfiler, in order to elucidate their potential biological functions. Protein-protein interaction (PPI) network analysis was performed using the STRING database (https://string-db.org/, accessed November 10, 2024), integrating information from genomic neighborhood, gene co-occurrence, fusion events, co-expression, experiments, curated databases, and text mining. Interactions with a confidence score > 0.40 were considered reliable and included in the PPI network. Hub genes were identified using the CytoHubba plugin in Cytoscape (v3.9.1), applying the Maximum Clique Centrality (MCC) algorithm. The top 10 genes ranked by MCC score were selected as hub genes.

Validation of DEGs by quantitative real-time PCR (qRT-PCR)

12 DEGs were randomly selected for validation via qRT-PCR, including 6 from the HY vs. LY comparison and 6 from the HF vs. LF comparison. Total RNA was reverse-transcribed to cDNA using the PrimeScript RT Reagent Kit (TaKaRa, Cat. No. 6210). GAPDH served as the internal reference gene. Primers were designed using Primer Premier 5.0 (Table 3). PCR reactions were performed in a 15 µL system containing 7.5 µL 2× qPCR Mix, 1.5 µL of forward and reverse primers, 2.0 µL cDNA, and 4.0 µL nuclease-free water. The amplification protocol included an initial denaturation at 95 ℃ for 30 s, followed by 40 cycles of 95 ℃ for 15 s and 60 ℃ for 30 s, with fluorescence signal collected at 0.5 ℃ increments. Relative gene expression levels were calculated using the 2^−ΔΔCt method and log₂-transformed prior to analysis.

Table 3.

qRT-PCR primer sequence information

Gene	Primer sequence (5’-3’)	Amplicon length(bp)	Accession No.
OXTR (Oxytocin Receptor)	S: TGGGACATCACCTTCCGCTT A: CAGGTAGGTGGAGGCGAACAT	93	XM_023620041.2
CAMK2B (Calcium/Calmodulin Dependent Protein Kinase II Beta)	S: GCCAAGAACAGCAAGCCAATC A: GTACTGCGTGAGGCGGATGT	102	XM_070265546.1
ERBB3 (Erb-B2 Receptor Tyrosine Kinase 3)	S: ACCAGGGTTAGAGGAAGAGGATG A: CCAGGACAGAACTGAGACCCAC	116	XM_023643764.2
GPM6A (Glycoprotein M6A)	S: TGCTCTATGCTGGAGTTGCC A: GATGGCACCAGTTGTGAAGAA	227	XM_070252863.1
ASAH1 (N-Acylsphingosine Amidohydrolase 1)	S: GAAACCTTTAACGGTGAATCTGG A: CCAAGATCCATTCTAGGACACCC	179	XM_005606374.4
KLRA1 (Killer Cell Lectin Like Receptor A1)	S: CAGGCAAATTCAATGAAGACCAC A: CCGCTGTCAATCCATTTCCAC	238	XM_005610809.4
PDE4D (Phosphodiesterase 4D)	S: AGACCTGAGCAACCCGACAA A: GGCATCAGGATGGACGAGGT	226	NM_001242504.1
RGS6 (Regulator Of G Protein Signaling 6)	S: GGCTACGTCTTTCCCATCTCAG A: GAGATAGATGGCATAGTCGGTGTT	132	XM_023628026.1
DST (Dystonin)	S: CATCATTCACAAATACAGGCCG A: ACTGATGCCTTCGCCACCTT	232	XM_023624277.1
CTSH (Cathepsin H)	S: GAACCCTACTTTCTGCCTGTGAC A: TCCACAGCAACCAGGTCTTCA	296	NM_001434490.1
SSTR1 (Somatostatin Receptor 1)	S: TGCCTGTGCTACGTGCTCATC A: TGCTCTGCGAACACGTTGAC	179	XM_023638714.2
ADCY5 (Adenylate Cyclase 5)	S: GTGTTTCCTGCTGCTGACCTT A: CTGGCTGGCACTGATGTTGT	241	XM_001917066.6
GAPDH (Glyceraldehyde-3-Phosphate Dehydrogenase)	S: ATGGTGAAGGTCGGAGTAAACG A: CATGGGTGGAATCATACTGAAACA	154	NM_001163856.1

Metabolomics data analysis

Metabolomics data were preprocessed using the metaX tool, followed by multivariate statistical analysis. For unsupervised analysis, PCA was applied to assess the natural clustering tendencies among sample comparisons. For supervised analysis, orthogonal partial least squares discriminant analysis (OPLS-DA) was employed to calculate Variable Importance in Projection (VIP) scores, and metabolites with VIP > 1 were retained for further analysis. Univariate analysis was performed using Student’s t-test (P < 0.05) to identify statistically significant differences between comparisons.

Metabolites meeting both criteria (VIP > 1 and P_adj < 0.05) were considered significantly differential and subjected to KEGG pathway enrichment analysis (P_adj < 0.05) using MetaboAnalyst (https://www.metaboanalyst.ca/, accessed November 20, 2024). To gain a comprehensive understanding of inter-pathway and metabolite-pathway relationships, enriched pathways and associated metabolites were imported into the R environment. KEGG pathway topologies were constructed using clusterProfiler, visualized with ggplot2 and enrichplot, and metabolite-pathway interaction maps were generated using pathview. Statistical significance of differences in the abundance of relevant metabolites was assessed using Student’s t-tests in GraphPad Prism.

Integrated transcriptomic and metabolomic analysis

Based on KEGG pathway enrichment analysis, significantly regulated pathways (P_adj < 0.05) that were co-enriched by both DEGs and DAMs were identified, and the corresponding DEGs and DAMs involved in these pathways were extracted for further analysis. Subsequently, Mantel tests were performed using the vegan package (v2.5-7) to assess global correlations among transcriptomic, metabolomic, and phenotypic data matrices. Data integration and transformation were conducted using dplyr (v1.0.8), and correlation networks between DEGs and DAMs were constructed using ggcor (v0.9.4). Prior to machine learning modeling and feature selection, multi-omics and phenotypic data were standardized and preprocessed. Data cleaning and integration were performed using tidyverse (v1.3.1), followed by log₂ transformation of gene expression levels and metabolite concentrations. Dimensionality reduction was conducted using PCA implemented in FactoMineR (v2.4) with visualization via factoextra (v1.0.7). Collinearity among variables was evaluated using Hmisc (v4.6-0) and corrplot (v0.92) based on Pearson correlation coefficients. Highly collinear variables (|r| > 0.9) were identified, and representative features were retained for subsequent modeling.

For each phenotypic trait, including milk yield, milk fat percentage, milk protein percentage, lactose percentage, and the yields of milk fat, protein, and lactose, three high-dimensional regression models were constructed: LASSO regression, Random Forest regression, and SVR. In the model design, stratified random split was used to divide the 24 samples into a training set and an independent test set. The training set consisted of 20 samples (approximately 83.33%) and was used for model construction, hyperparameter tuning, and feature selection, while the independent test set included 4 samples (approximately 16.70%) and was reserved exclusively for preliminary, unbiased evaluation of final model performance. This partitioning strategy balances the need to maximize training information under small-sample conditions with the requirement to preserve an independent validation subset, thereby reducing the risk of optimistic bias in model assessment. To further mitigate the potential randomness associated with a single data split, all model tuning and feature importance evaluations were conducted within the training set using 5-fold cross-validation. Hyperparameter optimization, feature ranking, and stability assessment were based on aggregated results across cross-validation folds, ensuring that model inference was not driven by any individual partition.

The modeling workflow was implemented in R using the tidyverse and caret packages (v6.0-94), with parallel computing enabled via doParallel (v1.0.17). LASSO regression was performed using glmnet (v4.1-8), with the regularization parameter α set to 1. The optimal λ value was determined through five-fold cross-validation on the training set. Random Forest regression was conducted using randomForest (v4.7-1.1), with the mtry parameter optimized via five-fold cross-validation. Support Vector Regression (SVR) was implemented using e1071 (v1.7-14), employing a radial basis function (RBF) kernel and evaluating combinations of kernel bandwidth (σ = 0.01, 0.1, 1) and penalty parameter (C = 0.1, 1, 10). Model performance was evaluated using root mean squared error (RMSE) and coefficient of determination (R²). Feature importance was assessed using vip (v0.3.2) to rank and visualize variables by relative importance in each model. Core features were defined as those consistently ranked among the top contributors across all three models. Pearson correlation coefficients were then calculated between these core features and phenotypic traits, retaining pairs with |r| > 0.7 and P < 0.05 for downstream analysis. Based on these associations, gene-metabolite-phenotype interaction networks were constructed and visualized using Cytoscape.

Results

Correlation analysis between lactation performance and body conformation traits

In the HY vs. LY comparison, mares in the HY group had a significantly higher milk yield than those in the LY group (P = 0.0001) (Fig. 1A). Similarly, in the HF vs. LF comparison, the HF group showed a significantly higher milk fat percentage than the LF group (P < 0.0001) (Fig. 1B). Further correlation analysis indicated positive associations between milk yield and body conformation traits. Significant correlations (P < 0.05) were observed specifically with body length, teat diameter, and teat length. In contrast, milk fat percentage showed a tendency for negative associations with the majority of the 15 body conformation traits assessed. However, none of these correlations reached statistical significance (Fig. 1C).

Fig. 1.

Associations between lactation performance and body conformation traits in Kazakh mares. A Difference in milk yield between the HY and LY groups. B Difference in milk fat percentage between the HF and LF groups. C Pearson correlation analysis among milk yield, milk fat percentage, and selected body and udder conformation traits. Correlations with |r| > 0.7 and P < 0.05 are highlighted. Data in (A) and (B) were analyzed using Student’s t-test. Statistical significance: ns, not significant (P > 0.05); *P < 0.05; **P < 0.01

Transcriptome sequencing results overview

Transcriptome sequencing was performed on 24 samples across two comparisons. For HY vs. LY, raw reads per sample ranged from 40.0 to 58.3 million (total 81.39 Gb), yielding 38.8–56.1 million clean reads per sample (78.92 Gb total) after quality filtering, with average GC content of 50.87%, Q20 of 97.97%, Q30 of 94.25%, and a mean mapping rate of 93.04% (Table 4). For HF vs. LF, raw reads ranged from 41.4 to 47.6 million, yielding 39.6–46.3 million clean reads per sample after filtering, with average GC content of 50.92%, Q20 of 97.64%, Q30 of 93.62%, and a mean mapping rate of 92.36% (Table 5).

Table 4.

Summary of HY vs. LY comparison transcriptome data

Sample	Raw reads (pairs)	Clean reads (pairs)	Raw bases (Gb)	Clean bases (Gb)	Q20 (%)	Q30 (%)	GC (%)	Total map (%)
HY1	44,741,490	43,437,194	6.71	6.52	98.05	94.47	51.32	93.70
HY2	47,316,580	46,062,124	7.10	6.91	98.12	94.55	51.83	93.98
HY3	42,219,116	41,170,850	6.33	6.18	98.00	94.29	50.71	93.58
HY4	45,268,268	43,762,456	6.79	6.56	97.81	93.84	51.00	93.0
HY5	46,417,190	44,937,740	6.96	6.74	97.67	93.49	51.13	93.55
HY6	49,878,738	48,754,248	7.48	7.31	97.97	94.3	50.88	92.67
LY1	42,800,070	41,555,372	6.42	6.23	98.01	94.34	51.16	93.62
LY2	58,261,260	56,060,808	8.74	8.41	97.75	93.83	48.46	91.70
LY3	42,736,286	41,283,270	6.41	6.19	98.01	94.37	50.69	91.15
LY4	42,883,390	41,578,066	6.43	6.24	98.09	94.51	51.05	93.13
LY5	40,000,966	38,836,780	6.00	5.83	98.07	94.44	51.45	93.68
LY6	40,119,834	38,659,460	6.02	5.80	98.11	94.58	50.79	92.81

Table 5.

Summary of HF vs. LF comparison transcriptome data

Sample	Raw reads (pairs)	Clean reads (pairs)	Raw bases (Gb)	Clean bases (Gb)	Q20 (%)	Q30 (%)	GC (%)	Total map (%)
HF1	47,623,630	46,320,676	7.14	6.95	97.98	94.46	51.90	92.68
HF2	44,210,864	42,937,466	6.63	6.44	97.82	94.00	51.79	92.59
HF3	43,999,654	42,440,070	6.60	6.37	97.81	93.98	51.09	91.92
HF4	41,407,792	40,169,418	6.21	6.03	98.00	94.44	51.61	93.48
HF5	46,798,064	45,441,206	7.02	6.82	97.61	93.54	50.94	91.27
HF6	44,904,392	43,500,108	6.74	6.53	97.53	93.36	51.41	92.70
LF1	45,109,260	43,764,366	6.77	6.56	97.76	93.84	49.66	92.04
LF2	41,760,726	40,390,494	6.26	6.06	97.66	93.51	47.53	93.60
LF3	43,143,730	42,322,970	6.47	6.35	97.10	92.31	51.01	92.85
LF4	46,160,326	45,033,794	6.92	6.76	96.85	91.87	51.10	91.72
LF5	41,597,536	39,599,766	6.24	5.94	97.59	93.67	51.19	90.68
LF6	42,116,922	40,121,655	6.34	6.05	96.58	93.12	51.91	92.66

Identification of DEGs

PCA revealed clear separation between the HY and LY groups, as well as between the HF and LF groups, confirming the suitability of the samples for subsequent differential expression analyses (Fig. 2A, B). Hierarchical clustering heatmaps (Fig. 2C, D) further revealed distinct differences in gene expression patterns between HY and LY, as well as between HF and LF, suggesting that the identified DEGs may be associated with lactation-related traits in Kazakh mares. In the HY vs. LY comparison, a total of 286 DEGs were identified, comprising 198 upregulated and 88 downregulated genes (Fig. 2E). Detailed information on the DEGs is provided in Supplementary Table S1. In the HF vs. LF comparison, 627 DEGs were identified, including 424 upregulated and 203 downregulated genes (Fig. 2F). Detailed information on the DEGs is provided in Supplementary Table S2. Analysis of group-specific DEGs revealed nine genes expressed exclusively in the HY vs. LY comparison (8 in HY only and 1 in LY only; Fig. 2G). Similarly, 11 genes were expressed exclusively in the HF vs. LF comparison (8 in HF only and 3 in LF only; Fig. 2H). Most of these genes encode hypothetical or poorly annotated proteins, potentially indicating novel roles in specific biological processes. Venn diagram analysis revealed limited overlap of DEGs between the two comparisons (Fig. 2I). 13 genes were commonly upregulated, and 2 were commonly downregulated in both comparisons. Additionally, 2 genes were upregulated in HY vs. LY but downregulated in HF vs. LF, while 1 gene exhibited the opposite pattern (downregulated in HY vs. LY but upregulated in HF vs. LF).

Fig. 2.

Identification and visualization of DEGs. A PCA of samples based on gene expression profiles in the HY vs. LY comparison. B PCA of samples based on gene expression profiles in the HF vs. LF comparison. C Hierarchical clustering heatmap of gene expression profiles in the HY vs. LY comparison. D Hierarchical clustering heatmap of gene expression profiles in the HF vs. LF comparison. E Volcano plot of DEGs in the HY vs. LY comparison. F Volcano plot of DEGs in the HF vs. LF comparison. G Hierarchical clustering heatmap of group-specific DEGs in the HY vs. LY comparison. H Hierarchical clustering heatmap of group-specific DEGs in the HF vs. LF comparison. I Venn diagram illustrating the overlap of upregulated and downregulated DEGs between the HY vs. LY and HF vs. LF comparisons

Functional enrichment analysis

In the HY vs. LY comparison, GO enrichment analysis revealed that upregulated DEGs were significantly enriched in 26 terms (P_adj < 0.05), primarily involving transport processes, signal transduction, and metabolic regulation, whereas downregulated DEGs were primarily enriched in immune-related functions and amino acid metabolism (Fig. 3A, C). Consistently, KEGG pathway analysis showed that upregulated DEGs were enriched in signaling pathways involved in ErbB, Wnt, and cAMP and hormone regulation (e.g., insulin, prolactin, and GnRH signaling), while downregulated DEGs were mainly associated with immune-related and amino acid metabolism pathways (Fig. 3B, D).

Fig. 3.

GO and KEGG enrichment analysis of DEGs. GO enrichment was based on level 2 terms, and KEGG pathways were summarized at level 1 categories. A GO enrichment of upregulated DEGs in the HY vs. LY comparison. B KEGG enrichment of upregulated DEGs in the HY vs. LY comparison. C GO enrichment of downregulated DEGs in the HY vs. LY comparison. D KEGG enrichment of downregulated DEGs in the HY vs. LY comparison. E GO enrichment of upregulated DEGs in the HF vs. LF comparison. F KEGG enrichment of upregulated DEGs in the HF vs. LF comparison. G GO enrichment of downregulated DEGs in the HF vs. LF comparison. H KEGG enrichment of downregulated DEGs in the HF vs. LF comparison

In the HF vs. LF comparison, GO enrichment analysis revealed that upregulated DEGs were significantly enriched in 27 terms (P_adj < 0.05), primarily related to transport, signal transduction, and immune regulation-related GO terms, whereas downregulated DEGs were associated with immune response, metabolic processes, and stress-related functions (Fig. 3E, G). KEGG analysis further revealed that upregulated DEGs were enriched in immune-related pathways, while downregulated DEGs were mainly involved in MAPK and hormone-related signaling pathways (e.g., estrogen signaling) (Fig. 3F, H).

Comparative analysis of DEGs with genes in the animal QTL database

To identify functionally relevant candidate genes underlying lactation performance, DEGs from the HY vs. LY and HF vs. LF comparisons were cross-referenced with the Animal QTL Database. This integrative analysis identified 43 overlapping genes associated with milk production-related traits across multiple livestock species, including cattle, sheep, goats, and pigs (Fig. 4A). These traits predominantly involved milk yield and milk composition parameters, such as milk fat, protein, and lactose traits.

Fig. 4.

Integration analysis of DEGs with QTL-associated genes from the Animal QTL Database. A UpSet plot illustrating the overlap between DEGs and trait-associated genes in the Animal QTL Database. Pink bars represent the number of overlapping genes for each specific lactation trait, while blue dots and connecting lines indicate shared genes across multiple traits. BP_adj obtained from mixed linear model fitting. C Summary of MLM analysis results, indicating associations between individual genes and lactation phenotypes (Est.: regression coefficient; SE: standard error; R²: coefficient of determination). D, E Pathway enrichment analysis of selected genes. Heatmaps show gene expression levels normalized as log₁₀-transformed FPKM values, and bubble plots display log₂ (fold change) values for each gene

A mixed linear model identified 17 genes significantly associated with lactation traits after multiple testing correction (P_adj < 0.05) (Fig. 4B). The model showed strong explanatory power for most genes (R² > 0.85), with moderate fits for ERBB3 (R² = 0.67), B4GALNT2 (R² = 0.76), and DMPK (R² = 0.61). Among genes negatively associated with milk yield or component yields, PMP22 was significantly associated with milk yield, milk protein yield, and lactose yield (P_adj < 0.05), while FAM83A and HSD17B3 were negatively associated with milk protein and lactose yields. Conversely, PDGFRB showed highly significant positive associations with milk yield, milk protein yield, and lactose yield (P_adj < 0.01); B4GALNT2 was positively associated with milk yield; and HTR1B was positively associated with milk protein and lactose yields. Composition-specific associations included AGPAT4 with milk fat yield, VIPR2 with milk fat percentage, GRIA4 and SLC50A1 with lactose percentage, RFX4 and EPOR negatively with lactose percentage, HPN positively with milk protein percentage, and SPATC1 and ERAS negatively with milk protein percentage (P_adj < 0.05) (Fig. 4C). Functional enrichment analysis (P_adj < 0.05) revealed significant enrichment in pathways related to cell growth, metabolism, and signal transduction (Fig. 4D, E). Notably, PDGFRB and ERBB3 were enriched in PI3K-Akt, JAK-STAT, and MAPK pathways; EPOR in PI3K-Akt; VIPR2 and HTR1B in cAMP signaling; AGPAT4 in phospholipase D signaling and glycerolipid metabolism; GRIA4 in glutamate receptor signaling; and HPN in exopeptidase-related pathways. Hub gene analysis using the CytoHubba MCC algorithm identified ERBB3, PDGFRB, GRIA4, and HPN as central nodes (Fig. S1A, B). ERBB3 exhibited high connectivity in both the HY vs. LY (Degree = 14) and HF vs. LF (Degree = 24) networks. DMPK and ERAS were more central in the HF vs. LF network (Fig. S1C, D).

Validation of DEGs by qRT-PCR

To validate the reliability of the RNA-seq results, twelve DEGs were selected for qRT-PCR analysis. These included OXTR, CAMK2B, ERBB3, GPM6A, ASAH1, and KLRA1 in the HY vs. LY comparison, and CTSH, PDE4D, RGS6, DST, SSTR1, and ADCY5 in the HF vs. LF comparison. The qRT-PCR results showed expression trends highly consistent with those from the RNA-seq analysis (Fig. 5A, B), thereby confirming the reliability of the transcriptomic data. These findings further support the robustness and reproducibility of the RNA-seq-based differential expression analysis in this study.

Fig. 5.

Validation of selected DEGs by qRT-PCR. Statistical significance: ns, not significant (P > 0.05). A Relative expression levels of selected DEGs in the HY vs. LY comparison, as determined by RNA-seq and qRT-PCR. B Relative expression levels of selected DEGs in the HF vs. LF comparison, as determined by RNA-seq and qRT-PCR

Identification of DAMs

PCA in positive ion mode revealed clear separation of plasma metabolite profiles between the HY and LY groups, as well as between the HF and LF groups (Fig. 6A, B), supporting the suitability of these comparisons for subsequent metabolomic analyses. Model validation by permutation tests in orthogonal partial least squares discriminant analysis (OPLS-DA) demonstrated good predictive performance (HY vs. LY: R² = 0.97, Q² = 0.95; HF vs. LF: R² = 0.84, Q² = 0.76) without overfitting (Fig. 6C, D).

Fig. 6.

Multivariate analysis and identification of DAMs in positive ion mode. A PCA score plot of plasma samples from the HY and LY groups. B PCA score plot of plasma samples from the HF and LF groups. C OPLS-DA score plot of plasma samples from the HY and LY groups. D OPLS-DA score plot of plasma samples from the HF and LF groups. E Volcano plot of DAMs in the HY vs. LY comparison. F Volcano plot of DAMs in the HF vs. LF comparison. G Venn diagram illustrating the overlap of upregulated and downregulated DAMs between the HY vs. LY and HF vs. LF comparisons. H Hierarchical chemical classification of DAMs in the HY vs. LY comparison (superclass level; bubble size represents VIP value). I Hierarchical chemical classification of DAMs in the HF vs. LF comparison (superclass level; bubble size represents VIP value)

Using criteria of VIP > 1 and P_adj < 0.05, a total of 100 DAMs were identified in the HY vs. LY comparison (61 upregulated and 39 downregulated; Fig. 6E). Detailed information on the DAMs is provided in Supplementary Table S3. In the HF vs. LF comparison, 29 DAMs were identified (13 upregulated and 16 downregulated; Fig. 6F). Detailed information on the DAMs is provided in Supplementary Table S4. Venn diagram analysis showed only one overlapping DAM between the two comparisons (Fig. 6G). This metabolite exhibited opposite regulation patterns: significantly upregulated in the HY vs. LY comparison (fold change = 1.08, VIP = 1.49, P_adj = 0.0029) but downregulated in the HF vs. LF comparison (fold change = 0.18, VIP = 1.75, P_adj = 0.0444), indicating divergent metabolic mechanisms underlying milk yield and milk fat percentage. Hierarchical classification revealed distinct metabolite profiles between the comparisons. In the HY vs. LY comparison, DAMs were predominantly lipids and lipid-like molecules (24 species, 24.0%) and organic acids and derivatives (23 species, 23.0%), followed by organoheterocyclic compounds (14 species, 14.0%), organic oxygen compounds (11 species, 11.0%), phenylpropanoids and polyketides (8 species, 8.0%), and benzenoids (7 species, 7.0%). Less abundant classes included nucleosides and nucleotides (4 species, 4.0%), organic nitrogen compounds (4 species, 4.0%), alkaloids and derivatives (2 species, 2.0%), and single representatives from lignans, organosulfur compounds, and mixed metal/non-metal compounds (1 species each, 1.0%) (Fig. 6H). In contrast, in the HF vs. LF comparison, 41.4% of DAMs (12 species) remained unannotated, likely due to limitations in current metabolite databases. Among annotated DAMs, organic acids and derivatives were most prevalent (8 species, 27.6%), followed by organoheterocyclic compounds (4 species, 13.8%), organic oxygen compounds (2 species, 6.9%), benzenoids (1 species, 3.4%), and lipids (1 species, 3.4%) (Fig. 6I).

PCA in negative ion mode revealed clear separation of plasma metabolite profiles between the HY and LY groups, as well as between the HF and LF groups (Fig. 7A, B), supporting appropriate group separation for subsequent metabolomic analyses. OPLS-DA further confirmed robust model performance (HY vs. LY: R² = 0.96, Q² = 0.85; HF vs. LF: R² = 0.85, Q² = 0.63) with clear metabolic differences between groups (Fig. 7C, D). Using the criteria of VIP > 1 and P_adj < 0.05, 163 DAMs were identified in the HY vs. LY comparison (102 upregulated and 61 downregulated; Fig. 7E). Detailed information on the DAMs is provided in Supplementary Table S3. In the HF vs. LF comparison, 29 DAMs were identified (27 upregulated and 2 downregulated; Fig. 7F). Detailed information on the DAMs is provided in Supplementary Table S4. Venn diagram analysis identified three overlapping DAMs between the comparisons (Fig. 7G). Notably, these showed divergent regulation: uric acid was upregulated in the HY vs. LY comparison (fold change = 2.72, VIP = 1.79, P_adj = 0.0055) but downregulated in the HF vs. LF comparison (fold change = 1.26, VIP = 1.68, P_adj = 0.0321); kynurenic acid was downregulated in the HY vs. LY comparison (fold change = 0.62, VIP = 1.46, P_adj = 0.0220) but upregulated in the HF vs. LF comparison (fold change = 3.09, VIP = 1.66, P_adj = 0.0303); and inosine 5’-diphosphate was downregulated in the HY vs. LY comparison (fold change = 0.26, VIP = 2.04, P_adj = 0.0217) but upregulated in the HF vs. LF comparison (fold change = 2.07, VIP = 1.67, P_adj = 0.0239).

Fig. 7.

Multivariate analysis and identification of DAMs in negative ion mode. A PCA score plot of plasma samples from the HY and LY groups. B PCA score plot of plasma samples from the HF and LF groups. C OPLS-DA score plot of plasma samples from the HY and LY groups. D OPLS-DA score plot of plasma samples from the HF and LF groups. E Volcano plot of DAMs in the HY vs. LY comparison. F Volcano plot of DAMs in the HF vs. LF comparison. G Venn diagram illustrating the overlap of upregulated and downregulated DAMs between the HY vs. LY and HF vs. LF comparisons. H Hierarchical chemical classification of DAMs in the HY vs. LY comparison (superclass level; bubble size represents VIP value). I Hierarchical chemical classification of DAMs in the HF vs. LF comparison (superclass level; bubble size represents VIP value)

Hierarchical chemical classification highlighted distinct metabolite profiles. In the HY vs. LY comparison, DAMs were predominantly organic acids and derivatives (34 species, 20.9%) and lipids and lipid-like molecules (33 species, 20.2%), followed by organic oxygen compounds (29 species, 17.8%), benzenoids (25 species, 15.3%), phenylpropanoids and polyketides (19 species, 11.7%), and organoheterocyclic compounds (17 species, 10.4%). Minor classes included nucleosides and nucleotides (4 species, 2.5%), organic nitrogen compounds (1 species, 0.6%), and organohalogen compounds (1 species, 0.6%) (Fig. 7H). In contrast, the HF vs. LF comparison was dominated by lipids and lipid-like molecules (14 species, 48.3%), with 27.6% unclassified (8 species), followed by organoheterocyclic compounds (4 species, 13.8%), nucleosides and nucleotides (3 species, 10.3%), and single representatives from phenylpropanoids and polyketides and organic acids and derivatives (1 species each, 3.4%) (Fig. 7I).

Metabolite-pathway interaction enrichment analysis

Analysis of the metabolite-pathway interaction network revealed a highly interconnected metabolic network in the HY vs. LY comparison. A total of 41 DAMs were significantly enriched in 15 metabolic pathways (P_adj < 0.05), primarily involving energy metabolism, amino acid metabolism, and lipid metabolism. The energy metabolism network was centered on the tricarboxylic acid (TCA) cycle, with key intermediates (2-oxoglutarate, citrate, succinate, and cis-aconitate) forming a tightly integrated energy supply system. This suggests that enhanced energy metabolism may support increased milk yield during peak lactation. Furthermore, interface metabolites such as glycerone, alpha-D-glucose, D-galactose, and glycerol linked glycolysis/gluconeogenesis, glycerolipid metabolism, and galactose metabolism, thereby connecting energy, carbohydrate, and lipid metabolic processes (Fig. 8A). Notably, among the seven key metabolites in energy metabolism, all except cis-aconitate exhibited significantly higher abundance in the HY group than in the LY group. Similarly, among the four key metabolites associated with lipid metabolism, all except glycerone were significantly more abundant in the HY group (Fig. 8B).

Fig. 8.

Metabolite-pathway interaction network analysis and abundance of key metabolites. A Metabolite-pathway interaction network for DAMs in the HY vs. LY comparison. B Abundance differences of key metabolites between the HY and LY groups. C Metabolite-pathway interaction network for DAMs in the HF vs. LF comparison. D Abundance differences of key metabolites between the HF and LF groups. Statistical significance: ns, not significant (P > 0.05); *P < 0.05; **P < 0.01

In contrast, the metabolic network in the HF vs. LF comparison was less interconnected, with only six pathways significantly enriched by the 13 DAMs (P_adj < 0.05), primarily associated with lipid metabolism. Key metabolites, including L-serine, 2-phospho-D-glycerate, taurine, and arachidonate, linked pathways such as glycolysis/gluconeogenesis, arachidonic acid metabolism, glycine, serine, and threonine metabolism, and biosynthesis of unsaturated fatty acids (Fig. 8C). The abundances of these key metabolites were significantly higher in the HF group than in the LF group (Fig. 8D).

Machine learning-based prioritization of candidate features associated with lactation traits in Kazakh mares

To explore molecular features potentially associated with lactation performance in a data-driven manner, machine learning models were applied to evaluate candidate molecular features across multiple lactation-related traits. Prior to modeling, all variables, including phenotypic traits, gene expression levels, and metabolite abundances, were standardized using z-score normalization.To ensure model parsimony and mitigate multicollinearity, redundant features with high pairwise correlations (|r| > 0.9) or low variance were removed, resulting in the exclusion of specific phenotypic traits, genes, and metabolites (Fig. S2C, D, G; Tables S5). Given the high dimensionality of the omics data relative to the sample size, PCA was performed to mitigate the risk of overfitting through dimensionality reduction. In the DEG dataset, the first three principal components (PCs) accounted for 84% of the total variance (Fig. S2E). PCA score-loading analysis revealed that PC1 effectively discriminated samples according to milk yield and milk fat percentage. Key contributors to milk yield differentiation included LOC100050670, COL8A1, and COL6A2, whereas GLDC showed a stronger association with milk fat percentage (Fig. S2F).

Similarly, in the DAM dataset, the first three PCs explained 83% of the total variance (Fig. S2H). PC1 primarily captured the variation associated with milk yield and milk fat percentage, while PC2 reflected within-group variation. Loading analysis indicated that 2-phospho-D-glycerate was strongly associated with the low milk yield group, whereas the majority of metabolites, including histamine, correlated more closely with variations in milk fat percentage. Notably, histamine exhibited a higher contribution in the high milk fat percentage group (Fig. S2I).

To prioritize molecular features associated with each lactation trait, we compared feature importance rankings across multiple models using both the training and testing datasets. For milk yield (Fig. 9A), milk fat percentage (Fig. 9B), and milk protein percentage (Fig. 9C), the models achieved R² values exceeding 0.90 in both the training and testing sets. Corresponding RMSE and MAE values were calculated for each model. By integrating features consistently selected across the three models, we identified features associated with milk yield (5 features), milk fat percentage (3 features), and milk protein percentage (1 feature).

Fig. 9.

Identification of gene features using machine learning for lactation traits. A-D Model performance and feature importance for milk yield (A), milk fat percentage (B), milk protein percentage (C), and lactose percentage (D) using three machine learning algorithms

Machine learning-based identification of metabolites associated with lactation traits in Kazakh mares

For milk yield, the three models achieved R² values exceeding 0.90 in both training and test datasets, with low prediction errors (RMSE and MAE). In contrast, models for milk fat percentage, milk protein percentage, and lactose percentage exhibited R² values ranging from 0.55 to 0.90, with acceptable prediction errors. By integrating features consistently selected across models, we identified one feature associated with milk yield (Fig. 10A), two with milk fat percentage (Fig. 10B), one with milk protein percentage (Fig. 10C), and one with lactose percentage (Fig. 10D).

Fig. 10.

Identification of metabolites features using machine learning for lactation traits. A-D Model performance and feature importance for milk yield (A), milk fat percentage (B), milk protein percentage (C), and lactose percentage (D) using three machine learning algorithms

Construction of the DEGs-DAMs-phenotype correlation network

Pearson correlation analysis was performed between lactation traits and DEGs/DAMs with high loadings to construct a gene–metabolite–phenotype network. In this network, histamine exhibited the highest degree within the metabolite subnetwork, while DEGs showed comparable connectivity. Network topology indicated that both DEGs and DAMs were predominantly associated with milk fat percentage (Fig. 11). Integration with machine learning feature selection (LASSO, Random Forest, and SVR) further identified GLDC and histamine as robust candidates consistently selected across all three models for milk fat percentage. Pearson correlation confirmed that both GLDC (r = 0.85, P = 0.0005) and histamine (r = 0.91, P = 0.0039) were strongly positively correlated with milk fat percentage, and they were also significantly correlated with each other (r = 0.73, P = 0.0072).

Fig. 11.

Construction of the DEGs-DAMs-phenotype correlation network

Discussion

As an important economic livestock breed in Xinjiang, Kazakh horses have attracted growing attention due to the unique nutritional value of their milk. Elucidating the genetic and metabolic characteristics underlying milk production traits is essential for genetic improvement of quantitative traits such as milk yield in Kazakh mares. Blood plays a pivotal role in milk formation. As a circulatory carrier, it transports oxygen, glucose, amino acids, hormones, and other metabolic substrates to mammary tissue, supplying essential precursors for milk synthesis [19]. Consequently, integrated analyses of blood transcriptomes and plasma metabolomes can, to a certain extent, provide insights into blood-mediated regulatory signals associated with lactation traits. Nevertheless, the mammary gland, as the direct organ responsible for milk synthesis and secretion, generally offers more specific and mechanistic insights into core lactation processes through its tissue-specific gene expression and metabolic dynamics. While direct omics analysis of mammary tissue would greatly facilitate the identification of key regulatory factors, obtaining such biopsies from Kazakh mares poses substantial practical challenges, as the procedure is invasive and may disrupt lactation behavior. Therefore, considering animal welfare and practical feasibility, blood was selected as the primary biological material in this study. It serves as a minimally invasive proxy that captures systemic physiological changes closely linked to mammary gland function, enabling the preliminary screening of lactation-related genes and metabolites. By integrating transcriptomic and metabolomic data with machine learning-based feature selection, this study successfully identified a set of candidate genes and metabolites significantly associated with milk yield and milk composition traits in Kazakh mares, and constructed a preliminary gene–metabolite–phenotype interaction network. These findings establish candidate targets for future marker-assisted selection.

Body measurements and udder morphology are considered useful phenotypic indicators of milk production performance in livestock. In this study, we assessed 15 phenotypic traits related to body conformation and udder structure. Milk yield was significantly correlated with body length, teat diameter, and teat length. In contrast, milk fat percentage was not significantly associated with any of the assessed traits. These results are consistent with previous findings linking milk yield with body length, teat diameter, and teat length in dairy cow [20] and goat [21].

Lactation traits are classical quantitative traits influenced by polygenic inheritance and environmental factors. The Animal QTL Database contains numerous chromosomal regions and molecular markers associated with lactation traits in dairy animals. In this study, we cross-referenced the 913 DEGs identified in peripheral blood (286 from the HY vs. LY comparison and 627 from the HF vs. LF comparison) with known QTLs related to lactation traits. This analysis identified 43 overlapping genes. To explore the potential associations between these genes and lactation performance in Kazakh mares, we further performed association analyses between gene expression levels and lactation traits using mixed linear models. Several genes showed statistically significant associations. For example, PMP22 expression was negatively associated with milk, protein, and lactose yields, which is consistent with previous findings in Holstein cows [22–27]. HSD17B3, an enzyme involved in estradiol synthesis [28–30], exhibited a negative association with protein and lactose yields in Kazakh mares, differing from the positive correlations reported in cattle [31, 32]. This interspecies difference may be related to distinct selection pressures: intense selection for high milk production in dairy cattle versus selection for environmental adaptability in Kazakh mares [33–35]. Similarly, B4GALNT2 was positively associated with lactose yield, consistent with findings in cattle and sheep [36–38]. HTR1B showed positive associations with milk and protein yields, aligning with reports in dairy cattle [39–42]. AGPAT4 was positively associated with milk fat percentage, in line with bovine studies [43, 44]. VIPR2 displayed positive correlations with milk fat traits [45, 46], while SLC50A1 was positively associated with lactose-related traits [47–52]. In addition, ERBB3 showed a positive association with milk fat percentage, consistent with its known involvement in mammary development and lipid metabolism in other species [53, 54]. It should be noted that, as these findings are based on peripheral blood transcriptomics rather than mammary gland tissue, the identified genes should currently be regarded as preliminary blood-based candidate markers associated with lactation traits in Kazakh mares, rather than as direct regulators of milk synthesis.

Beyond these individual associations, several genes identified in the present study (including ERBB3, GRIA4, and others) occupied central positions in the protein–protein interaction network. These genes have previously been reported in lactation-related gene networks or associated with lactation traits in multiple livestock species, including sheep [55], goats [56], mares [57], and buffaloes [58]. Functional enrichment analysis showed that the hub genes were enriched in signaling pathways previously implicated in lactation, such as the PI3K-Akt, JAK-STAT, MAPK, cAMP, phospholipase D, and glycerolipid metabolism pathways [59–61]. These highly conserved pathways are known to play important roles in mammary gland development, milk component synthesis, and energy metabolism by integrating nutritional and hormonal signals across mammals, including rodents, ruminants, and equids [62–71].

We then compared the association patterns observed in Kazakh mares with those reported in intensively selected dairy cattle. Despite substantial differences in lactation physiology and milk composition between the two species (equine milk typically contains lower fat and protein but higher lactose), several candidate genes identified in this study (e.g., PMP22, B4GALNT2, HTR1B, and AGPAT4) showed association directions similar to those in dairy cattle. This similarity suggests that certain molecular associations with lactation traits may be partially conserved across species. In contrast, the opposite association pattern for HSD17B3 may indicate species-specific differences, potentially reflecting adaptive variations in breeding objectives, reproductive strategies, and energy allocation between Kazakh mares and dairy cattle.

The milk yield and milk fat percentage of Kazakh mares appear to be associated with distinct metabolic patterns in peripheral blood. High-yield mares showed broad enrichment of energy and substrate metabolism pathways, including the tricarboxylic acid (TCA) cycle, pentose phosphate pathway, and glycolysis/gluconeogenesis, along with elevated levels of representative metabolites such as citrate and (S)-malate. In contrast, milk fat percentage was linked to more localized metabolic alterations, characterized by significantly higher concentrations of specific metabolites, including L-serine and arachidonate. These observations suggest that milk yield is associated with broader systemic metabolic features in blood, whereas milk fat percentage may involve more restricted metabolite associations, primarily related to lipid metabolism. This divergence likely reflects the differing physiological demands of the two traits. Lactation is an energy-intensive process [72] that requires substantial metabolic precursors and ATP to support mammary epithelial cell function and milk synthesis [73, 74]. The blood metabolic profiles observed in high-yield mares are consistent with these demands. Notably, several TCA cycle intermediates (citrate, (S)-malate, succinate, and 2-oxoglutarate) were enriched, potentially indicating increased mitochondrial activity to meet elevated ATP requirements [75]. Concurrent enrichment of the pentose phosphate pathway may support the production of NADPH and D-ribose-5-phosphate, which are essential for fatty acid biosynthesis and nucleotide metabolism [76, 77]. In addition, alterations in glycolysis/gluconeogenesis, glyoxylate and dicarboxylate metabolism, butanoate metabolism, and ascorbate and aldarate metabolism may facilitate the availability and interconversion of key carbon sources such as glucose [78]. The enrichment of bridging metabolites (e.g., glycerol, glyceraldehyde, α-D-glucose, and D-galactose) further points to dynamic interactions between carbohydrate and lipid metabolic pathways [79]. Collectively, these coordinated changes are likely to support lactose biosynthesis [80, 81]. As lactose is the primary osmotic regulator of milk volume, elevated lactose synthesis is expected to drive increased milk secretion [82]. Overall, high-yield mares were characterized by a broadly enriched, TCA cycle-centered metabolic network in blood that integrates carbohydrate, lipid, and amino acid metabolism to supply energy and carbon skeletons required for lactation.

In contrast, milk fat synthesis primarily depends on de novo fatty acid synthesis in the mammary gland and the uptake and utilization of circulating lipids [83, 84]. Previous studies have shown that the number of metabolites significantly associated with milk fat percentage is often limited [75, 85–87], suggesting that this trait may not necessitate widespread changes in global metabolite levels. Instead, milk fat synthesis and secretion appear to rely more on the expression and activity of specific genes and rate-limiting enzymes [88]. Our findings regarding genes such as AGPAT4 and VIPR2 are consistent with this perspective. Moreover, milk fat percentage involves complex lipid classes (e.g., sphingolipids, cholesterol, and phospholipids) [89, 90] that are frequently difficult to detect and annotate using conventional untargeted metabolomics. These potentially important yet unannotated lipid species may partly account for the high proportion of unidentified metabolites observed in the milk fat percentage-related analyses. In summary, future studies should prioritize optimizing experimental designs, increasing sample sizes to minimize individual variability, and employing targeted lipidomics approaches. Such efforts will help elucidate the relationships among blood metabolic processes, regulatory pathways, and lactation traits in Kazakh mares.

In this study, conventional differential analysis, mixed linear models, and machine learning–based feature selection were applied as complementary components of a hierarchical framework to dissect lactation regulation. Differential analysis identified genes and metabolites showing significant changes across contrasting lactation phenotypes, and cross-referencing with the Animal QTL Database ensured that the candidate factors were supported by prior genetic evidence. MLMs were then used to validate associations between these candidates and continuous lactation traits while accounting for individual heterogeneity and confounding factors. Importantly, the MLMs served for robust association testing rather than for additional feature selection. Machine learning models were applied to the integrated multi-omics dataset as an exploratory feature selection approach, aiming to identify candidate features that showed stable patterns across different models rather than to establish predictive or causal relationships. Consequently, factors such as GLDC and histamine, although not among the most significant hits in differential analysis, were repeatedly selected across models, suggesting that they may represent stable candidate factors associated with lactation traits rather than drivers of extreme group differences. Given the relatively small sample size, particularly the limited independent test set, these results should be considered preliminary findings requiring further validation in larger independent cohorts.

GLDC is a key enzyme in glycine metabolism and plays a central role in folate-mediated one-carbon metabolism, contributing to nucleotide, protein, and lipid biosynthesis as well as mitochondrial function [91]. Previous studies have indicated that GLDC may be associated with the PI3K/Akt/mTOR signaling pathway, which is involved in cellular anabolic processes and the generation of lipid synthesis precursors such as acetyl-CoA [92]. Given the close metabolic connections between one-carbon metabolism and amino acid metabolism, GLDC may also be indirectly related to histidine metabolism. However, the direct involvement of GLDC in histidine metabolism remains unclear. Histidine serves as the precursor of histamine, a bioactive amine that has been reported to be involved in mammary gland physiology and lactation. Previous studies in cattle have shown that histamine administration is associated with increased milk fat percentage, possibly through enhanced uptake of circulating fatty acids by the mammary gland [93, 94]. In addition, public database evidence (ProteomicsDB: https://www.proteomicsdb.org; Human Protein Atlas: https://www.proteinatlas.org) indicates that GLDC is expressed in mammary gland tissues in humans and mice, suggesting potential tissue relevance. However, it should be noted that the identification of GLDC in this study was primarily based on integrative multi-omics analysis and machine learning-based feature selection, which does not necessarily imply a direct causal role in milk fat synthesis. Collectively, these findings identify GLDC and histamine as candidate factors associated with milk fat percentage in Kazakh mares, warranting further functional validation.

Despite these findings, several limitations should be acknowledged. First, although blood serves as a practical surrogate for systemic physiological status, it does not fully reflect tissue-specific regulatory processes occurring in the mammary gland, which may limit the mechanistic depth of our conclusions. Second, the relatively small sample size, limited by the dispersed and small-scale pastoral management practices in Xinjiang, may reduce statistical power and generalizability. Third, the cross-sectional design captures a single time point during peak lactation, precluding the assessment of dynamic changes in gene expression and metabolite profiles across the lactation cycle. Fourth, although machine learning approaches enabled feature selection, the predictive performance of the identified markers requires further validation in larger independent cohorts. Future studies incorporating mammary tissue analysis, longitudinal sampling, and functional validation (e.g., knockout or overexpression models) will be necessary to confirm the regulatory roles of the candidate genes and metabolites identified in this study.

Conclusion

This study provides a multi-omics perspective on the molecular basis of lactation performance in Kazakh mares, identifying key genes and metabolites associated with milk yield and milk composition. Among these, GLDC and histamine were consistently detected across multiple analytical approaches, suggesting stable associations with milk fat percentage and potential roles in lipid-related metabolic regulation. In contrast, genes such as AGPAT4, B4GALNT2, and SLC50A1 exhibited trait-specific associations, highlighting the complex and heterogeneous genetic architecture underlying different milk production traits. Collectively, these findings advance our understanding of the molecular and metabolic determinants of lactation traits in Kazakh mares and provide a preliminary resource for molecular breeding and nutritional regulation strategies.

Abbreviations

DEGs

Differentially expressed genes

DAMs

Differentially accumulated metabolites

ML

Machine learning

LASSO

Least absolute shrinkage and selection operator

RF

Random forest

SVR

Support vector regression

GO

Gene ontology

KEGG

Kyoto encyclopedia of genes and genomes

MLM

Mixed linear model

PPI

Protein-protein interaction

MCC

Maximum clique centrality

PMP22

Peripheral myelin protein 22

HSD17B3

17β-hydroxysteroid dehydrogenase type 3

B4GALNT2

β-1,4-N-acetylgalactosaminyltransferase II

HTR1B

5-hydroxytryptamine receptor 1B

AGPAT4

1-acylglycerol-3-phosphate o-acyltransferase 4

VIPR2

Vasoactive intestinal peptide receptor 2

SLC50A1

Solute carrier family 50 member 1

ERBB3

Erb-B2 receptor tyrosine kinase 3

Authors’ contributions

Conceptualization, C.M. and J.M.; Data curation, C.M and J.M.; Formal analysis, Y.Z. and J.M.; Funding acquisition, Y.Z. and J.M.; Investigation, C.M. and J.W.; Methodology, C.M. and J.W.; Project administration, J.M.; Resources, Y.Z.; Software, C.M. and P.L.; Visualization, C.M.; Writing - original draft, C.M.; Writing - review & editing, C.M., X.Y., W.R., X.X. and J.M.

Funding

This study was supported by the Major Science and Technology Project of the Xinjiang Uygur Autonomous Region (2022A02013-1); the Study on the Effect of Mare Milk-Derived Peptides on High-Fat Diet-Induced Metabolic Disorders and Its Regulatory Mechanism (XJ2026G118); the project “Construction of an Equine Milk Probiotic Library and R&D and Demonstration of Functional Equine Milk Products” (2024B02013-2); and the Xinjiang Uygur Autonomous Region Dairy Industry Technology System Project (XJARS-11).

Data availability

Raw reads of Transcriptomic sequencing of blood are available at CNCB. GSA submission information: CRA028574. https://ngdc.cncb.ac.cn/gsa/browse/CRA028574.

Declarations

Ethics approval and consent to participate

The experimental protocol was approved by the Institutional Animal Care and Use Committee of Xinjiang Agricultural University (Urumqi, China) under approval number 2023020.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.