Identification and characterization of gene networks and key genes related to the high-yield production of milk in high-yield cows using meta-analysis of microarray data

Abstract

Milk yield is most important economic trait in dairy cows and understanding molecular basis and components involved in high-yield production is one of crucial steps to develop and select new breeds. In this study, we used combination of two statistical methods based on the p-value and effect sized to meta-analysis three datasets followed with construction of weighted gene co-expression network based on the expression matrix of differentially expressed genes identified in meta-analysis to detect major gene modules and hub genes. Based on the FDR cut-off<0.05 and Log₂ fold change>2 and < 0.5, we identified 1028 differentially expressed genes that were shared between the Fisher and REM method and were consistent across datasets. Molecular function analysis showed that upregulated differentially expressed genes mostly enriched to ion binding, small molecule binding, and identical protein binding while downregulated genes were enriched to catalytic activity (Bonferroni test; threshold of 0.05). Weighted gene co-expression network analysis identified three major modules associated with fatty acid metabolism, PPAR signaling pathway, insulin resistance, terpenoid backbone biosynthesis, and steroid biosynthesis. A total of 12 hub genes (one downregulated and 11 upregulated) identified from protein-protein interaction network of modules. This study could identify new differentially expressed genes related to lactation processes in high-yield-cows. Moreover, we could reveal some gene modules and hub genes in each module which are biologically more meaningful.

Highlights

• •Based on the p-value and effect size, genomic data were integrated.
• •Gene networks and major genes associated with high-yield milk were identified and characterized.
• •Weighted gene co-expression network analysis identified three major gene modules.
• •This study identified 12 core genes and their expression pattern in the PPI network.

1. Introduction

The milk yield and compositions in lactating cows are the most important economical trait that has been the aim of many studies to select or improve in dairy cows. Milk yield is a polygenic trait that is under control of many genes and loci. Detection of key genes and gene networks associated with the milk production could provide valuable molecular information to use in the selection and improvement programs. One of several organs that are involved in the process of milk production is liver. The liver metabolism is crucial in transition and during lactating for providing and balancing the energy need of dairy cows [1]. The 3 week before and after parturition is the most challenging period for dairy cows and during this period the liver role is the key to coordinate nutrient flux in lactation [2]. Transcriptional adaptation of the liver to metabolically changes involves downregulation of genes associated with oxidative phosphorylation, protein ubiquitination, and ubiquinone biosynthesis with ketosis and upregulation of genes and nuclear receptors associated with cytokine signaling, fatty acid uptake/transport, and fatty acid oxidation [3].

Several studies have explored the transcriptomic profiled of liver in dairy cows. In a study [4], it has been shown that 1063 differentially expressed genes significantly enriched into 16 biological processes and seven pathways are involved the metabolic adaptation of the transition stage. They revealed that genes related to fatty acid oxidation/metabolism, cholesterol metabolism, and gluconeogenesis have significant role [4]. In another study to identify candidate genes for milk production traits, it has showed biological processes related to metabolic and biosynthetic and signaling pathways of PPAR, AMPK and p53 [5]. In dairy cows with different level of milk production, it has been shown the association of pathways related to glycan biosynthesis and metabolism and amino acid metabolism and also biological processes related to cell-cell adhesion, multicellular organism growth, and amino acid and protein metabolism [6]. Furthermore, a few study identified key genes playing a central role in lactating in dairy cows, such as CYP7A1, APOA1, CREM, LOC522146, CYP2C87, HMGCR, FDFT1, SGLE, and CYP26A1 [4], and APOC2, PPP1R3B, PKLR, ODC1, DUSP1, LMNA, GALE, ANGPTL4, LPIN1 and CDKN1A [5].

In the last decade, only a few studies showed the key genes, biological processes, and pathways related to high-yield production of milk. Li, Liang [5] reported 147 differentially expressed genes during peak of lactation in high-yield cows related to response to stimulus, metabolic process, biological regulation, cellular component organization or biogenesis, multi-organism process, localization and identified ANGPTL4, CDKN1A, ODC1, LPIN1, and DUSP1differentially expressed genes as candidate functional genes affecting milk production traits. Gessner, Winkler [7] studied dairy cows fed a polyphenol-rich grape seed and grape marc meal extract with the high level of milk production and reported 156 up- and 51 downregulated differentially expressed genes significantly enriched to cell cycle regulation and the p53 signaling and cell cycle pathways.

We could not find studies that report comprehensive meta-analysis of transcriptomic data of liver in dairy cows with the focus on the milk production. However, a few meta-analyses are recently published, such as meta-analysis to the understanding of the relationships between energy balance in early lactation and cow performance [8], meta-analysis to model the prediction of milk yield in dairy cows [9], and meta-analysis to evaluate the pattern of differential gene expression in the liver of cows under negative energy balance and under subclinical and clinical ketosis [10].

This study's strategy was to use the combination of two methods based on the p-value and effect size and the WGCNA approach. The strategy was to use datasets from samples that showed the high amount of milk production compared to their control conditions, and then extract common genes between datasets, and finally performing microarray meta-analysis. The microarray data of high-yield cows downloaded from the Gene Expression Omnibus database were used to carry out meta-analysis of expression data to identify new DEGs related to high-yielding in the liver of dairy cows. The WGCNA of identified meta-DEGs was used to construct gene co-expression network and discover major gene modules of the network, and identify hub genes in each module. Gene ontology analysis and the Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis were used to explore biological contributions and biological pathways of DEGs and major modules.

2. Materials and methods

2.1. Datasets collection

Gene expression studies were identified by searching the Gene Expression Omnibus (GEO) repository (https://www.ncbi.nlm.nih.gov/gds/) and ArrayExpress database (https://www. ebi.ac.uk/arrayexpress/). The query: ((“cattle"[MeSH Terms] OR “Bos taurus"[Organism] OR cow [All Fields]) AND dairy [All Fields] AND (“lactation"[MeSH Terms] OR “breast feeding"[MeSH Terms] OR lactation [All Fields])) AND “Bos taurus"[porgn] was used to find and collect GEO datasets through October 2023. The strategy was to select studies that used treatments to increase milk yield. The studies were selected to be included in meta-analysis if they had the following criteria: (1) case-control studies with high-yield milk production case; (2) availability of raw data; (3) data based on Affymetrix; and (4) data came from liver tissue. The meta-analysis was performed based on guidelines provided in the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement (http://www.prisma-statement.org/).

2.2. Preprocessing and quality control

The normalization was performed using Average algorithm (RMA) available in the R-package affy [11,12] and quality was checked using the R-package ArrayQualityMetrics [13], and the R-package COMBAT package was used to correct batch-specific variations.

2.3. Meta-analysis and identification of DEGs

The R-package metaDE [14] was used to analyze and identify DEGs across studies. At the first stage, we used variance cut-off = 0.05 to eliminate duplicate probe sets and probes with very low expression. Then, to reduce variability came from different studies, we only analyzed genes that were common across datasets and had same expression pattern (MVperc = 0). The moderated t-test and the corrected p-value using false discovery rate (FDR) were used to identify meta-DEGs [15] and those genes were considered DEGs that showed FDR cut-off<0.05 and Log₂ fold change>2 and < 0.5 [16]. Based on the some recently used methods for microarray meta-analysis and data integration [[16], [17], [18]] and guidelines for statistical considerations for microarray meta-analysis [19,20], this study used two methods. The first approach was Fisher Z-test developed by Zaykin [21] and the second approach was Fixed Effect Method (REM) developed by [22]. The Fisher method combines data from different studies by calculating p-value, and the REM method used FDR to identify DEGs. To determine up- or downregulation pattern of gene expression, right and left-sided p-value were used.

2.4. Weighted gene co-expression network analysis

The matrix of meta-DEGs was used as input to The R-package WGCNA [23] in order to construct a scale-free co-expression network and identify significant gene modules. The WGCNA uses the dynamic tree cut (DTC) algorithm based on the topological overlap matrix (TOM) and the corresponding dissimilarity (1-TOM) value. The following formula was used to calculate TOM:

$${TOM}_{i,j}=\frac{\sum _u{a}_{iu}{a}_{uj+}{a}_{ij}}{\min \left(k_i,k_j\right)+1-a_{ij}}+k_i=\sum _u{a}_{iu}$$

Where, row index u (u = 1, …,m) represents sample measurements.

Meta-DEGs were clustered using TOM and analyzed with DTC algorithm to identify gene modules. The modules with at least 100 genes and cut height of 40 were chosen for further analyzing and identifying hub genes.

2.5. Gene ontology (GO) analysis

The g:Profiler [9] was used to annotate identified DEGs, hub genes, and key modules. A list of Affy-bovine ids was passed to the gene id converter tool and converted to Ensembl Gene. Affy-bovine ids that did not matched to any Ensembl ids were ignored. The statistical domain scope set to only annotated genes and the Bonferroni test (threshold of 0.05) was used to identify significant molecular function, biological process and KEGG pathways.

3. Results

3.1. Data mining and quality control

Based on the search strategy and selection criteria, three datasets consist of expression data with 126 samples generated from liver tissues of high-yield cows were downloaded for further analysis (Table 1). The quality analysis of samples detected 16 outliers based on the distances between arrays, boxplots, and MA plots (Supplementary file 1). We removed 55 samples from GSE94794 that were not from liver tissue and totally 58 samples were included in the QC. After removing outliers, PCA and density plot showed no outlier (Fig. 1). The remaining 58 sample consist of 28 controls and 28 treatments were merged. The matching and filtering of 24128 genes removed 2354 genes (9.7 %) and resulted in 21774 common genes.

Table 1.

Characteristics of the individual datasets used in the study.

Accession number	Platform	Organism	Tissue	Sample size	Treatment
GSE56547	Affymetrix Bovine Genome Array	Bos taurus	Liver	11	Energy balance
GSE87391	Affymetrix Bovine Genome Array	Bos taurus	Liver	10	Dairy cows fed conjugated linoleic acid
GSE94794	Affymetrix Bovine Genome Array	Bos taurus	Liver	160	Lactating cows vs non-lactating

Fig. 1.

Quality control of microarray data after normalizing with RMA algorithm; (A) PCA analysis of expression datasets after batch effect correction and outlier removal; (B) Density plot of expression datasets after outlier removal. We used COMBAT package [16] with parametric adjustment to correct the batch effect and the R-package ArrayQualityMetrics [15] to remove outliers.

3.2. Identification of DEGs using combined meta-analysis

Using two meta-analysis methods, three datasets consist of 58 samples were analyzed. The first approach (Fisher) showed 1701 significant genes at p-value of 0.01 and 4139 significant genes at 0.05. Based on the FDR and fold change (FC) > 2 and < 0.5, the Fisher identified 210 and 664 DEGs at 0.01 and 0.05, respectively (Supplementary file 2). Right- and left-sided p-value at 0.05 showed 389 upregulated and 255 downregulated genes (Fig. 2). The second approach that was based on the combination of effect size (REM method) identified 926 DEGs (p-value of 0.01) and 2378 DEGs (p-value of 0.05). At FDR of 0.01 and 0.05 and FC > 2 and < 0.5, REM detected 47 and 139 DEGs, respectively (Supplementary file 3).

Fig. 2.

Heatmap of some DEGs identified using the Fisher method. The expression pattern is represented in the Log₂ Fold Change (Log FC). Each row represents a DEGs. The number 0 represents control and the number 1 represents treatment groups. The color scale ranges from red (the lowest expression value) to green (the highest expression value).

The Fisher and REM approaches identified new 83 and 840DEGs that were not detected when the analysis performed on the individual studies (moderated t-test, p < 0.05). Nevertheless, 1089 genes in GSE56547, 762 genes in GSE87391, and 885 genes in GSE94794 were detected as DEGs when the analysis was performed on induvial studies (Fig. 3-A). In total 688 genes were detected by both the Fisher (24.44 % of identified DEGs) and REM (51.26 % of identified DEGs) (p-value of 0.05). Based on the FDR of 0.05, 77 DEGs were common between the Fisher and REM. The most significant genes were 76 DEGs that were detected by both the Fisher and REM at p < 0.05 and FDR<0.05 (Fig. 3-B). In total 1028 DEGs were commonly detected by the Fisher and REM (p < 0.05 and/or FDR<0.05).

Fig. 3.

Venn diagram of co-occurrence of DEGs identified by combined meta-analysis; (A) co-occurrence of DEGs across datasets, Fisher, and REM methods; (B) common DEGs identified by Fisher and REM methods based on p-value of 0.05 and FDR of 0.05.

Table 2 indicates the result of GO analysis of 1028 DEGs based on the expression pattern. The upregulated DEGs were significantly enriched into six molecular function and five biological processes (p < 0.05). GO analysis enriched the downregulated DEGs into two molecular function and one biological process (p < 0.05).

Table 2.

Gene ontology analysis of up- and downregulated DEGs identified using the meta-analysis.

	Source	Term name	Term id	Adjusted p-value	Gene number
Upregulated	GO:MF	Oxidoreductase activity	GO:0016491	3.42E-10	38
	GO:MF	Amide binding	GO:0033218	0.008234	13
	GO:MF	Identical protein binding	GO:0042802	0.012504	43
	GO:MF	Ion binding	GO:0043167	0.029319	88
	GO:MF	Small molecule binding	GO:0036094	0.030799	53
	GO:MF	Carboxy lyase activity	GO:0016831	0.036385	5
	GO: BP	Small molecule metabolic process	GO:0044281	9.21E-21	71
	GO: BP	Plasma lipoprotein particle organization	GO:0071827	3.44E-05	8
	GO: BP	Cholesterol transport	GO:0030301	0.000515	10
	GO: BP	Acetyl CoA metabolic process	GO:0006084	0.005624	6
	GO: BP	Cholesterol homeostasis	GO:0042632	0.00676	8
Downregulated	GO:MF	Catalytic activity	GO:0003824	0.001478	63
	GO:MF	Heme binding	GO:0020037	0.005421	8
	GO: BP	Oxoacid metabolic process	GO:0043436	4.07E-10	28

3.3. Identification of gene modules and hub genes using WGCNA

To get more insight on biological systems and detect biologically more significant hub genes, the DEGsidentified by both the Fisher and the REM approaches were further analyzed using WGCNA(WGCN). WGCN analysis of DEGs expression matrix based on the minimum size of 30 and TOM type of unsigned resulted in three significant modules including MEturquoise (277 genes), MEblue (162 genes), and MEbrown (136 genes) (Fig. 4 and Supplementary file 4). A list of DEGs clustered in each module was analyzed for significant KEGG pathways (p < 0.05). MEturquoise enriched into five, MEblue into one, and MEbrown into two KEGG pathways (Fig. 5).

Fig. 4.

Cluster dendrogram of meta-DEGs. The hierarchical average linkage clustering was employed to identify gene co-expression modules.

Fig. 5.

KEGG analysis of significant modules that shows several genes significantly enriched to KEGG pathways in the turquoise module at the threshold of 0.05. A list of DEGs identified using the meta-analysis and assigned to the turquoise module was analyzed in The g:Profiler to identify significant KEGG pathways.

To discover hub genes in each module, the protein interaction (PPI) network was constructed and the genes with the most connections were considered as hub genes. We used “degree” value as threshold to determine significant hub genes (degree score >8). Of three modules, 12 hub genes were detected in MEturquoise and no hub genes with the defined threshold were in MEblue and MEbrown (Fig. 6 and Table 3).

Fig. 6.

PPI network of the turquoise module. The hub genes were illustrated as highlighted boxes. The PPI were constructed using a list of DEGs from MEturquoise.

Table 3.

List of hub genes identified using weighted gene co-expression network analysis.

NO	Probe ID	Enseble ID	FDR	Expression	Gene name	Description
1	BT.19850.1.S1_AT	ENSBTAG00000017258	2.13E-04	Down	ACSL3	Acyl CoA synthetase long chain family member 3
2	BT.4798.1.S2_AT	ENSBTAG00000055207	2.53E-18	Up	SCD	Stearoyl CoA desaturase
3	BT.23212.1.S1_AT	ENSBTAG00000003068	6.35E-02	Up	MSMO1	Methylsterol monooxygenase 1
4	BT.5449.1.S1_AT	ENSBTAG00000016465	2.53E-18	Up	DHCR7	7-dehydrocholesterol reductase
5	BT.23182.1.S1_AT	ENSBTAG00000003948	2.53E-18	Up	FDPS	Farnesyl diphosphate synthase
6	BT.20584.1.S1_AT	ENSBTAG00000009231	4.28E-02	Up	NSDHL	NAD(P) dependent steroid dehydrogenase-like
7	BT.11465.1.S1_AT	ENSBTAG00000007840	2.53E-18	Up	HMGCR	3-hydroxy-3-methylglutaryl-CoA reductase
8	BT.22763.2.S1_A_AT	ENSBTAG00000011839	2.09E-02	Up	HMGCS1	3-hydroxy-3-methylglutaryl-CoA synthase 1
9	BT.22285.2.S1_AT	ENSBTAG00000018936	2.01E-02	Up	LSS	Lanosterol synthase
10	BT.3884.1.S1_AT	ENSBTAG00000012059	3.76E-02	Up	MVD	Mevalonate diphosphate decarboxylase
11	BT.12640.1.S1_AT	ENSBTAG00000013303	2.53E-18	Up	ACSS2	Acyl CoA synthetase short chain family member 2
12	BT.22207.1.S1_AT	ENSBTAG00000017819	6.05E-02	Up	PMVK	Phosphomevalonate kinase

4. Discussion

This study integrated three datasets with common 21774 genes, followed with meta-analysis of common genes to identify DEGs with consistent expression across datasets, gene modules, and hub genes relate to milking in high-yield cows. After quality control and excluding outliers, we used genes that were present in all datasets and removed gene with very high/low expression and un-expressed genes to lower the possibility of false positive results. Instead of using one specific meta-analysis method, this study's strategy was to combine two meta-analysis methods based on p-value and effect size. The genes detected as DEGs in this work had common gene expression pattern between studies with different conditions and treatments. Previous reports indicated that using combination of meta-analysis methods results in stringent relevant genes by excluding false positives [19,24]. Although it may exclude important genes [24]. Marot, Foulley [25] proved that in meta-analysis and combination of data from different microarray sources, p-value based methods have better sensitivity and effect-size based methods are more conservative. Fisher method based on the p-value combination could detect more DEGs compared to the REM method based on the effect-size combination (Supplementary file 2 and 3). Only a few genes identified as DEGs was shared between the Fisher and REM (Fig. 3). The Fisher method that works based on p-value combination, identify significant genes that are significant in more than one study [26]. The Fisher method is the best to combine data from different platforms [26]. The REM method, most frequently used method in meta-analysis, is the best when the heterogeneity is high between studies [26].

The type and number DEGs identified by meta-analysis compared to those identified in each individual dataset was significantly different. A hundred of DEGs identified DEGs in individual studies were not found by meta-analysis while about nine hundred DEGs only identified by meta-analysis. Meanwhile, the share of genes identified as DEGs between meta-analysis methods (Fisher and REM) and individual studies was different. For example, 401 DEGs was common between Fisher, REM, and GSE94794, or 698 genes in GSE56547 found differentially expressed by Fisher only, etc. (Fig. 3). One of the main purposes of using meta-analysis in data expression studies is to enhance the statistical power and detect more relevant genes [27]. In this regard, this study found 1028 genes that were detected as DEGs by both Fisher and REM and these genes were used to construct gene co-expression network to identify major modules and hub genes associated with high-yield lactation. Usually microarray studies have limited sample size and in such cases meta-analysis offers more reliable results and statistically better power by combining multiple datasets from multiple studies [25].

The analysis of expression pattern in identified DEGs showed that the number of upregulated genes is greater than the number of downregulated genes. Molecular function analysis of upregulated genes showed that the most of upregulated DEGs is enriched to ion binding, small molecules binding, identical protein binding, oxidoreductase activity, and small number enriched to carboxy lyase activity. This is comparable to molecular function of downregulated DEGs that enriched in catalytic activity and a few numbers enriched to heme binding (Table 2). Analysis of biological process for upregulated DEGs showed five biological processes, particularly small molecule metabolic process and cholesterol transport, compared to downregulated DEGs that enriched to oxoacid metabolic process (Table 2). It has been previously reported that in lactating dairy cows, the milk yield is considerably reduced when cows are subjected to negative energy balance with low plasma total cholesterol [28,29]. Moreover, previously has been shown that genes involved in the oxoacid metabolic process are essential in liver metabolic process in milking cows [[30], [31], [32]], however the expression pattern was not determined. These results are consistent with our findings and confirm the strategy robustness to identify meta-DEGs.

Construction of weighted gene co-expression network based on the expression matrix of identified meta-DEGs resulted in three significant modules related to high-yield milking (Fig. 4). Gene module is a set of genes with similar expression pattern and its identification is cornerstone of interpretation of genome-wide data and gain comprehensive understanding from systems biology perspective [33]. KEGG analysis of modules indicated metabolic pathways as the major biological pathways associated with identified modules. In more details, fatty acid metabolism, PPAR signaling pathway, insulin resistance, terpenoid backbone biosynthesis, and steroid biosynthesis were significantly enriched for the modules related to lactation process in high-yield cows (Fig. 5). Fatty acids metabolism is crucial in liver for high milk production [34]. It has been proved that fatty acids are energy source during milking and provide a better energy supply for dairy cows [34,35]. PPAR signaling pathway and its regulatory role in fat metabolism in liver and other organs such as mammary glands has been studied [36]. The PPARs expression correlate with high level of fatty acids metabolism [36] and previously reported that PPAR signaling pathway is one of important regulatory pathways in dairy cows [5]. The involvement of insulin resistance pathway [37], terpenoid backbone biosynthesis [38], and steroid biosynthesis [39,40]in processes of the milk production has been previously shown and here we revealed major modules related to these pathways as well as their contribution in the high-yield production of milk.

This work also identified 12 hub genes (highly connected genes) within modules by constructing PPI network (Fig. 6 and Table 3). We think that these genes are key components of lactation processes in liver of high-yield cows. The expression pattern of 11 hubs were up-regulation and only ACSL3 had downregulated pattern. It has been shown that ACSL3 are involved in fatty acid metabolism in liver and its main physiological function is to promote synthesis of lecithin and the cytosolic lipid droplets [41] through regulating activity of several lipogenic transcription factors [42]. The SCD is key enzyme in synthesis of monounsaturated fatty acids [43]. The physiological function of other hub genes proved by wet lab experiments includes MSMO1, DHCR7, FDPS, NSDHL, HMGCR, HMGCS1, LSS, MVD, ACSS2, and PMVKthat are involved in lipid metabolism and cholesterol homeostasis [[44], [45], [46], [47]].

The strength of this study was to use well-designed strategy in using combined meta-analysis powered by including only genets that were present in all datasets with similar expression pattern. Moreover, the findings should be considered in a light of two limitations. The first is the availability of datasets that passed selection criteria and quality control was limited to three datasets. The second, despite confirmation of the findings with experimental published articles, it is good practice to use real-time polymerase chain reaction to further validate the results. In this study to reach robust results and to overcome heterogeneity across datasets, the stringent setting was used that may exclude number of biologically important genes.

5. Conclusion

This work used the meta-analysis based on the combination of two statistical methods to study datasets related to lactation process in high-yield cows. Our study identified differentially expressed genes related to high-yield production of milk and could detected three major gene modules and hub genes. These findings extend of understanding of biological processes involved in lactation processes from systematic perspective. Moreover, identified differentially expressed genes and specially hub genes could be potential candidates for breeding programmers and even could be used as a biomarker in marker-assisted breeding.

CRediT authorship contribution statement

Mahdi Rahmatzadeh: Writing – original draft, Formal analysis, Data curation, Conceptualization. Reza Shokri-Gharelo: Writing – review & editing, Software, Methodology, Formal analysis, Conceptualization. Morteza Derakhti-Dizaji: Software, Methodology. Asghar Bazzaz: Software, Methodology, Investigation. Bizhan Mahmoudi: Writing – review & editing, Supervision, Methodology.

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.

Appendix A. Supplementary data

The following are the Supplementary data to this article:

Data availability

Data will be made available on request.