An intersection network based on combining SNP coassociation and RNA coexpression networks for feed utilization traits in Japanese Black cattle

Abstract

Genome-wide association studies (GWAS) of quantitative traits have detected numerous genetic associations, but they encounter difficulties in pinpointing prominent candidate genes and inferring gene networks. The present study used a systems genetics approach integrating GWAS results with external RNA-expression data to detect candidate gene networks in feed utilization and growth traits of Japanese Black cattle, which are matters of concern. A SNP coassociation network was derived from significant correlations between SNPs with effects estimated by GWAS across 7 phenotypic traits. The resulting network genes contained significant numbers of annotations related to the traits. Using bovine transcriptome data from a public database, an RNA coexpression network was inferred based on the similarity of expression patterns across different tissues. An intersection network was then generated by superimposing the SNP and RNA networks and extracting shared interactions. This intersection network contained 4 tissue-specific modules: nervous system, reproductive system, muscular system, and glands. To characterize the structure (topographical properties) of the 3 networks, their scale-free properties were evaluated, which revealed that the intersection network was the most scale-free. In the subnetwork containing the most connected transcription factors (URI1, ROCK2, and ETV6), most genes were widely expressed across tissues, and genes previously shown to be involved in the traits were found. Results indicated that the current approach might be used to construct a gene network that better reflects biological information, providing encouragement for the genetic dissection of economically important quantitative traits.

INTRODUCTION

Feed utilization efficiency is an important consideration for efficient beef cattle production. Genome-wide association studies (GWAS) of quantitative traits including component traits of feed efficiency have detected numerous genetic associations, but encounter difficulties in pinpointing prominent candidate genes and inferring gene networks. The association weight matrix (AWM: Fortes et al., 2010) or partial correlation and information theory (PCIT: Reverter and Chan, 2008) approach uses several GWAS results to narrow down candidate genes for traits of concern and explores their interactions by using correlations between allelic effects of SNPs (Fortes et al., 2011, 2012). The AWM-inferred network (here, referred to as SNP coassociation network) provides insights into key regulators of, and interactions among, genes affecting feed efficiency (Widmann et al., 2015). However, even such an analysis is not sufficient to provide information about the tissues expressing genes in AWM. A gene coexpression network is constructed from pairwise correlations between gene expression measures. Network analysis of coexpression patterns across multiple tissues is useful for identifying tissue-specific signatures and gene regulatory modules (Miller et al., 2010; Yang et al., 2014). However, transcriptome data from many tissues are difficult to collect in individuals used for GWAS. One way to address this issue is to reuse gene expression data from public databases (Rung and Brazma, 2013). Thus, a network integrating the properties of both SNP coassociation and coexpression/tissue-specificity would be crucial for studying feed utilization traits. In the present study, an intersection network was generated by combining a SNP coassociation network constructed from 7 feed utilization traits in Japanese Black cattle and a RNA coexpression network inferred using publicly available gene expression data from cattle tissues, and its properties were investigated.

MATERIALS AND METHODS

Performance tests of bulls were according to the protocol approved by the Wagyu Registry Association. All protocols for collecting DNA samples from the bulls were reviewed and approved by the Shirakawa Institute of Animal Genetics Animal Care and Use Committee.

Phenotype and Genotype Data

Phenotypic data from 974 performance-tested young Japanese Black bulls were used to examine 7 traits: roughage intake (kg), concentrate intake (kg), TDN intake (kg), roughage residual feed intake (RFI) (kg), concentrate RFI (kg), TDN RFI (kg), and ADG (kg/day). The initial ages of the bulls tested were 6 to 7 mo, and the length of the testing period was 112 d following an adaptation period of 20 d. The bulls received ad libitum feeding of roughage and water, whereas they were fed concentrate restricted at a rate of 1.0% to 1.3% of BW (kg) per day. During the test, amounts of feed offered and refused and BW were recorded for each animal in a pen, and intake was calculated as the difference between offered and refused feed. An outline of the testing scheme including the calculation of the 3 RFIs was described in some detail by Okanishi et al. (2008).

Genotyping of SNPs from 600 bulls was performed using Illumina BovineSNP50 BeadChip (50K; Illumina Inc., San Diego, CA), whereas the remaining 374 were genotyped with GeneSeek Bovine GGP Super LD BeadChip (LD; GeneSeek, Lincoln, NE). The LD SNP genotypes of these 374 bulls were imputed using BEAGLE 3.3.2 (Browning and Browning, 2007), together with 50K data from 506 animals as haplotype references. For the Japanese Black cattle, imputation accuracies of SNP genotypes from low-density panels using BEAGLE 3.3.2 are available (Uemoto et al., 2015; Ogawa et al., 2016), and the accuracy from LD to 50K reported by Uemoto et al. (2015) was 0.96. The observed and imputed 50K genotype data were merged. SNPs were excluded if they did not meet one of the following criteria: minor allele frequency ≥ 0.01, genotype call rate ≥ 0.95, P ≥ 0.001 from Hardy–Weinberg equilibrium test, or being located on autosomes (UMD3.1). Genotype information on 37,732 autosomal SNPs remained after filtering.

Transcriptome Data

To generate the gene coexpression network, we used expression profile data retrieved from the Bovine Gene Atlas (Harhay et al., 2010) within the National Center for Biotechnology Information (NCBI)’s Sequence Read Archive (SRP002394). These data derived from bovine samples (n = 95) obtained from multiple breeds, sexes, developmental stages, and tissues. The original files for SRP002394 were converted to FASTQ files using the SRA Toolkit v 2.3.4-2 (https://www.ncbi.nlm.nih.gov/sra/docs/toolkitsoft/). We used FastX (http://hannonlab.cshl.edu/fastx_toolkit/) to remove low quality reads (Quality Value < 20 at 80% probability) from the FASTQ files. Then, poor quality bases (Quality Value < 20) were trimmed off from both 5′ and 3′ ends using PRINSEQ (Schmieder and Edwards, 2011). After preprocessing, sequence data were mapped to UMD3.1 (http://bovinegenome.org/?q=node/61) using TopHat2 (Kim et al., 2013), with default options and parameters, and annotated on the bovine draft genome sequence using the RefSeq database (Release 57). We obtained 26,325 gene expression profiles from those 95 samples by calculating read density as reads per kilobase of exon model per million mapped reads (RPKM) (Mortazavi et al., 2008), using the “DEGseq” package of R (Wang et al., 2010) with default parameters.

Samples derived from cell lines, tumors, and dairy cattle were excluded because the present study focused on beef cattle. Genes with RPKM < 0.2 were also excluded, resulting in 19,491 gene expression profiles from 85 samples (Supplementary Material S1). Samples spanned 2 beef-cattle breeds and 15 tissues defined by BRENDA Tissue Ontology (Gremse et al., 2011) at 3 developmental stages (180-d fetus, juvenile, and adult). The coefficient of variation (CV) of expression levels across samples was used to measure tissue specificity (Tang et al., 2010). Genes with large expression variability among tissues are considered tissue-specific. The CV of each gene was calculated as the standard deviation of its expressions in the 85 samples divided by its mean expression. Genes with a CV > 1 were considered tissue-specific.

GWAS and Construction of AWM

The single-marker approach was used for GWAS of each trait. Trait data were analyzed using the following mixed linear model:

$$y=Xb+w_i{\alpha }_i+Zu+e,$$

where $y$ is the vector of records; $b$ is the vector of overall mean, fixed discrete effects (test station, year, and month) at the beginning of performance testing, and fixed continuous effects (body condition score [linear and quadratic], age at the beginning of performance testing [linear and quadratic], and the degree of inbreeding [linear]); $w_i$ is the vector of allele count on SNP i from individual j (0, 1, or 2) in the jth element; ${\alpha }_i$ is the scalar of SNP i’s additive effects; $u$ is the vector of random additive genetic effects; $e$ is the vector of random residuals; and $X$ and $Z$ are incidence matrices. Vectors $u$ and $e$ are assumed to follow a multivariate normal distribution with mean and covariance structure of the following:

$$\text{E}\left[\begin{array}{c}u\\ e\end{array}\right]=\left[\begin{array}{c}0\\ 0\end{array}\right]\text{ and Var}\left[\begin{array}{c}u\\ e\end{array}\right]=\left[\begin{array}{cc}A{\sigma }_u^2& 0\\ 0& I{\sigma }_e^2\end{array}\right],$$

where ${\sigma }_u{}^2$ is the additive genetic variance; ${\sigma }_e{}^2$ is the residual variance; $A$ is the additive relationship matrix; and I is the identity matrix. Pedigree depth was 5 generations and contained 6,355 animals. Likelihood ratio tests were performed for each SNP-trait combination. Computation was performed in Qxpak 5.05 (Pérez-Enciso and Misztal, 2011).

Following the work of Fortes et al. (2010), we constructed the AWM using GWAS results on the 7 traits. The estimated allele substitution effects per trait of all 37,732 SNPs were normalized with z-scores. After normalization, SNPs associated (P < 0.05) with 3 or more traits were selected to have at least 1 moderate genetic correlation between the associated traits, considering the estimates of genetic parameters reported by Okanishi et al. (2008). We next identified neighboring genes (within 5 kb of at least 1 SNP) as candidates for feed utilization and growth. Thus, each selected SNP was annotated with corresponding candidate genes. Finally, we obtained the AWM with an (i,j)-th element that was the standardized allele substitution effect of the jth SNP for the ith trait. For the candidate genes, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses were conducted in DAVID 6.7 (Huang et al., 2009).

Inferring SNP Coassociation, RNA Coexpression, and Intersection Network

To infer the SNP coassociation network (hereafter “AN”), a correlation matrix was calculated using the AWM elements corresponding to genes shared between the AWM and final transcriptome data. Then, the PCIT algorithm (Reverter and Chan, 2008) was applied to the correlation matrix via the “PCIT” package of R (Watson-Haigh et al., 2010). The statistical test based on PCIT identifies the significant correlations that should be used as interactions in the network.

We inferred the RNA coexpression network, as follows. First, the correlation matrix (19,491 × 19,491) relevant to genes in final transcriptome data was obtained from the whole gene expression matrix (19,491 genes × 85 samples). Next, we implemented the PCIT algorithm using the correlation matrix to construct a coexpression network (hereafter “EN”).

To generate an intersection network (hereafter “IN”), we superimposed AN and EN and extracted the shared interactions between target genes (i.e., those shared across AWM and transcriptome data). A tissue-specific module was defined as a group of consecutively connected genes with identical tissue-specificity. In the IN, we focused on the highly connected hub transcription factors (TFs) to identify key regulators and extracted a hub TFs-centered subnetwork as follows: 1) the TFs present in IN were identified using AnimalTFDB (Zhang et al., 2012); 2) the network degrees of the TFs were counted; and 3) the top 3 hub TFs and their immediate neighbors in IN were extracted. The IN and its subnetwork were visualized in Cytoscape 2.8.2 (Shannon et al., 2003). For in silico validation of the TF-target relationships involved in the subnetwork, we analyzed the promoter region 5 kb upstream the predicted target genes. This analysis was performed using LASAGNA-Search 2.0 (Lee and Huang, 2013, 2014), which allows us to find TF binding sites (TFBSs) in the promoter region of genes.

Checking Scale-Free Topology

To describe the structural characteristics of a given gene network, we investigated the degree distributions of AN, EN, and IN. A network is scale-free if a randomly picked node has k connections with other nodes at a probability of power law, p(k) ~ k^−γ, where k (≥1) is the number of a node’s connections and γ is the power-law exponent (Zhang and Horvath, 2005). A straight line was fitted to the frequency distribution of p(k) ~ k^−γ on a log-log plot, and the coefficient of determination was used to quantify how well AN, EN, and IN satisfied a scale-free topology.

RESULTS AND DISCUSSION

GWAS Results

We performed GWAS on 7 feed utilization and growth traits using relevant phenotype data in performance-tested Japanese Black bulls. The results are depicted in Manhattan plots for ADG, feed intake (Figure 1), and RFI traits (Figure 2). Under a lenient threshold (P < 0.0001), we detected 35 SNPs for ADG and <12 SNPs for the other traits. At a less stringent significance level (P < 0.05), strong overlaps (>20%) were found between significant SNPs for roughage intake and roughage RFI, concentrate intake and concentrate RFI, and TDN intake and TDN RFI (Table 1). We detected many SNPs on BTAs 3 and 6 that were highly significantly (P < 0.0001) or suggestively (P < 0.05) associated with some or all of the 7 traits. However, such SNPs were not found on BTAs 1, 4, 7, 9, 15, 19, 20, and 25–27. The genomic regions associated with ADG and feed intake on BTA 6 found in this study agreed with those previously reported in beef cattle (Lindholm-Perry et al., 2011). At the individual SNP level, the ADG-associated SNP (ARS-BFGL-NGS-33731) previously found on BTA 8 (Lu et al., 2013) was also significantly linked to ADG, and the RFI-associated region of BTA 5 (110–111 Mb), including the SNP Hapmap39945-BTA-75021, had also been identified in a previous GWAS (Bolormaa et al., 2011), showing that the results found here are consistent with previous findings, at the region level.

Figure 1.

Manhattan plots of genome-wide association study (GWAS) results for ADG (a), roughage intake (b), concentrate intake (c), and total digestible nutrient (TDN) intake (d). The aqua and green horizontal lines represent P < 0.0001 and P < 0.05, respectively.

Figure 2.

Manhattan plots of the genome-wide association study (GWAS) results for roughage residual feed intake (RFI) (a), concentrate RFI (b), and TDN RFI (c). The aqua and green horizontal lines represent P < 0.0001 and P < 0.05, respectively.

Table 1.

Number of single nucleotide polymorphisms associated (P < 0.05) with each pair of traits in Japanese Black beef cattle, based on genome-wide associations

Trait	Roughage intake	Concentrate intake	TDN intake	Roughage RFI	Concentrate RFI	TDNRFI	ADG
Roughage intake	(2,247)	171	674	1,015	191	378	425
Concentrate intake		(2,712)	906	154	959	362	561
TDN intake			(2,659)	494	676	940	783
Roughage RFI				(2,226)	645	841	191
Concentrate RFI					(2,587)	1,002	227
TDNRFI						(2,214)	198
ADG							(3,195)

Diagonal elements between parentheses show the number of single nucleotide polymorphisms associated with each trait. TDN = total digestible nutrient; RFI = residual feed intake; ADG = average daily gain.

Supplementary Table S1 lists 31 genes that are the nearest neighbors (within 5 kb) of highly significant SNPs (P < 0.0001) for at least 1 trait. One of the detected genes was ligand-dependent nuclear receptor corepressor-like protein (LCORL), a gene that is associated with feed intake when expressed in adipose tissue in beef cattle crosses derived from matings using Angus, Hereford, Simmental, Limousin, Charolais, Gelbvieh, and Red Angus (Lindholm-Perry et al., 2013). Moreover, leucine aminopeptidase 3 (LAP3) and potassium voltage-gated channel interacting protein 4 (KCNIP4) were previously confirmed to be associated with knuckle and bicep weight in Simmental cattle, respectively (Song et al., 2016). However, we did not detect significant SNPs near growth hormone receptor (GHR) and leptin (LEP), which are both well-known candidate genes for feed efficiency and feed intake (Hill, 2012). This might be due to these genes having no polymorphisms in the current sample population of Japanese Black cattle and/or to insufficient statistical power of the analysis.

GO and KEGG Pathway Analyses for Genes in AWM

One disadvantage of the strict significance threshold in GWAS is that it limits the power to detect SNPs of modest effects. However, AWM approach using a less stringent threshold can increase power, whereas the false positives are minimized, because the probability of randomly detecting a SNP associated with multiple correlated phenotypes is lower than detecting one that is associated with a single phenotype (Rao, 2008; Reverter and Fortes, 2013). Fortunately, almost all genetic correlations between ADG, feed intake, and RFI traits in Japanese Black bulls are moderate to high (Okanishi et al., 2008). Thus, based on the criteria using a less stringent threshold (P < 0.05), we selected SNPs associated with the 7 traits to build the AWM. In total, 644 SNPs could be assigned to their closest genes within 5 kb distance. Significant GO terms and KEGG pathways (Kanehisa and Goto, 2000) in the gene enrichment analysis for the 644 genes are summarized in Table 2. They encompassed many terms associated with metabolism, ion transport, channel activity, and ion binding, similar to previous GWAS of feed efficiency (Serão et al., 2013a, 2013b). Several of the pathways in Table 2 agree with findings in previous studies on RFI. That is, microarray experiments have demonstrated that genes involved in mitogen-activated protein kinase (MAPK) signaling of the liver are differentially expressed in beef cattle with high and low RFI (Chen et al., 2011). Additionally, RFI is related to brain remodeling, as well as the regulation of energy balance and feeding via GnRH release (Perkins et al., 2014); in turn, the latter is associated with axon guidance and cell adhesion pathways (Fortes et al., 2011). These results indicate that 644 genes selected for AWM provide useful information for feed utilization and growth traits which is worth further investigation in dissecting the traits.

Table 2.

Enriched Gene Ontology (GO) terms and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways for 644 genes in the association weight matrix (AWM)

GO term/KEGG pathway	Category	P-value	AWM gene count	Gene count in each category
Glycoprotein biosynthetic process	GO_BP	6.09E-04	10	80
Glycoprotein metabolic process	GO_BP	6.19E-04	11	97
Transmembrane transport	GO_BP	0.0030	23	389
Protein amino acid glycosylation	GO_BP	0.0044	8	69
Biopolymer glycosylation	GO_BP	0.0044	8	69
Glycosylation	GO_BP	0.0044	8	69
Generation of a signal involved in cell-cell signaling	GO_BP	0.0058	6	39
Enzyme linked receptor protein signaling pathway	GO_BP	0.0084	11	138
Neurotransmitter secretion	GO_BP	0.0155	4	18
Multicellular organismal movement	GO_BP	0.0170	3	7
Musculoskeletal movement	GO_BP	0.0170	3	7
Neuropeptide signaling pathway	GO_BP	0.0177	6	51
Ion transport	GO_BP	0.0187	25	511
Biological adhesion	GO_BP	0.0214	17	306
Cell adhesion	GO_BP	0.0214	17	306
Extracellular structure organization	GO_BP	0.0217	7	73
Cation homeostasis	GO_BP	0.0219	9	115
Exocytosis	GO_BP	0.0238	6	55
Cellular calcium ion homeostasis	GO_BP	0.0332	6	60
Regulation of cell morphogenesis	GO_BP	0.0334	5	41
Calcium ion homeostasis	GO_BP	0.0375	6	62
Synaptic vesicle transport	GO_BP	0.0411	3	11
Cellular metal ion homeostasis	GO_BP	0.0446	6	65
Metal ion transport	GO_BP	0.0476	15	285
Cell-cell adhesion	GO_BP	0.0495	9	135
Calcium ion binding	GO_MF	4.39E-05	33	528
Ion binding	GO_MF	7.70E-05	99	2,439
Cation binding	GO_MF	2.45E-04	96	2,415
Metal ion binding	GO_MF	2.60E-04	95	2,387
Sialyltransferase activity	GO_MF	0.0022	5	20
Ion channel activity	GO_MF	0.0044	17	268
Substrate-specific channel activity	GO_MF	0.0049	17	271
Passive transmembrane transporter activity	GO_MF	0.0056	17	275
Channel activity	GO_MF	0.0056	17	275
α-N-acetylneuraminate α-2,8-sialyltransferase activity	GO_MF	0.0077	3	5
Manganese ion binding	GO_MF	0.0105	7	65
Transition metal ion binding	GO_MF	0.0307	62	1,698
ATP binding	GO_MF	0.0354	40	1,018
Cation:cation antiporter activity	GO_MF	0.0379	3	11
Adenyl ribonucleotide binding	GO_MF	0.0397	40	1,027
Cation channel activity	GO_MF	0.0476	10	165
O-Glycan biosynthesis	KEGG_Pathway	0.0020	6	30
Axon guidance	KEGG_Pathway	0.0026	11	114
Glycosphingolipid biosynthesis	KEGG_Pathway	0.0083	4	14
Phosphatidylinositol signaling system	KEGG_Pathway	0.0219	7	71
Lysosome	KEGG_Pathway	0.0232	9	113
MAPK1 signaling pathway	KEGG_Pathway	0.0271	15	257
Arrhythmogenic right ventricular cardiomyopathy (ARVC)	KEGG_Pathway	0.0498	6	65

¹MAPK = mitogen-activated protein kinase.

SNP Coassociation and RNA Coexpression Networks

Figure 3 presents a flowchart of the steps used for network inference. The AN was inferred by the PCIT algorithm (Reverter and Chan, 2008). The PCIT is not a fixed-threshold method and applies local thresholds to every trio of genes to capture weak correlations. Thus, significance by the PCIT depends on gene content. To remove the influence of genes without transcriptome data, we applied PCIT to the 544 genes shared between the 644 genes for AWM and the 19,491 expressed genes from transcriptome data (second step on the left hand side of Figure 3). The AN comprised the 544 genes connected via 16,526 interactions (Supplementary Material S2). The EN was also inferred using PCIT, based on 19,491 gene expression patterns across the 85 samples from the Bovine Gene Atlas (Harhay et al., 2010). The EN comprised 540 genes connected via 7,284 interactions (Supplementary Material S3). Although all of the 544 target genes had significant correlations (or interactions) in the AN, only 540 of the 19,491 genes were significantly correlated in the EN. The small percentage of genes with significant correlations between expression levels suggested an absence of correlation between the majority of gene pairs, or insufficient power to detect significant correlations in the expression dataset.

Figure 3.

Flowchart of the steps involved in network inference. AWM (association weight matrix); PCIT (partial correlation and information theory); AN (SNP coassociation network); EN (RNA coexpression network); IN (intersection network).

Generating the Intersection Network via Superimposing 2 Networks

Although some SNP coassociation interactions indicated probable relationships between the TFs and their target genes, as demonstrated previously (Fortes et al., 2010), the presence of significant coassociations does not guarantee the existence of real regulatory interactions at the transcriptional level. Thus, we generated the IN from AN and EN to increase the reliability of detecting transcriptional regulation, by extracting shared interactions between EN and AN (last step in Figure 3). The resulting IN comprised 377 genes connected through 807 interactions (Figure 4). Supplementary Table S2 presents the results of gene enrichment analysis in AN, EN, and IN, evidencing that the significant GO terms and KEGG pathways were similar to those presented in Table 2.

Figure 4.

Overall view of the intersection network (IN). Node colors correspond to the tissue specificity of genes: gland (red), nervous system (blue), reproductive system (yellow), muscular system (green), other tissues (aqua), and nontissue-specific genes (white). (A) Tissue-specific module for nervous system. (B) Tissue-specific module for gland. (C) Tissue-specific modules for reproductive and muscular system.

The IN contained 4 tissue-specific modules with over 3 genes from the nervous system, reproductive/muscular systems, and glands (Figure 4). The largest module contained 6 genes (DZANK1, CDH4, SEMA4D, PCDH9, GPC5, and LOC100296463) predominantly expressed in the nervous system. In mouse hippocampus, the DZANK1 protein binds to Dlgap1, which encodes a scaffold protein involved in postsynaptic density (Li et al., 2016). Furthermore, PCDH9-knockout mice show cognitive impairments (Bruining et al., 2015), and GPC5 expression in the mouse brain is correlated with brain development (Saunders et al., 1997). LOC100296463 has significant homology to human HS3ST4, a gene implicated in synaptic transmission and neurotransmitter receptor activity based on functional prediction analyses (Debette et al., 2015). Both CDH4 and SEMA4D genes are annotated to KEGG pathways and GO terms involved in the regulation of axon guidance.

Five genes (LOC523049, LOC100847965, LOC787250, SPATA17, and KCNT2) were present in the second largest module, related to the reproductive system. SPATA17 is associated with human spermatogenesis and meiosis (Miyamoto et al., 2009). KCNT2 encodes a K⁺ channel that is regulated through cell-volume changes in Xenopus laevis oocytes (Tejada et al., 2014). LOC523049, LOC100847965, and LOC787250 are of unknown function.

The module related to the muscular system included FGF9, CILP, LOC511494, and ZNF407. Loss of FGF9 expression in mouse muscle leads to limb deformity (Hung et al., 2007). Down-regulation of CILP expression has been identified in humeri from splotch-delayed mouse embryos (Rolfe et al., 2014). Expressed in smooth muscle tissue, human ZNF407 encodes a zinc finger protein (http://www.proteinatlas.org/).

Finally, the gland module contained 3 genes (LOC100848745, KCNN3, and NOS1AP). KCNN3 belongs to the KCNN family of potassium channels that modulate insulin secretion in mouse β cells (Zhang et al., 2005). NOS1AP polymorphisms are associated with type 2 diabetes and the gene potentially affects calcium handling in human pancreatic β cells (Chu et al., 2010).

These tissue-specific modules likely have an important function related to feed efficiency and growth, such as ion channel and axon guidance which were significant pathways in the gene set analysis (Table 2). For example, the nervous system has been referred as an important regulator of feeding behavior, and some effects of restricted feeding on reproductive function have also been reported (Blank and Desjardins, 1985). However, further studies are required to understand the exact relationships between these genes and relevant traits.

Network Topological Properties

The robustness of biological networks (e.g., gene coexpression, protein-protein interaction, and metabolic networks) is associated with their natural exhibition of scale-free behavior (Jeong et al., 2000; Crucitti et al., 2003; Barabási and Oltvai, 2004; Khanin and Wit, 2006). A scale-free network has many low-degree genes and few high-degree genes. Therefore, we investigated the degree distributions of AN, EN, and IN to determine their structural characteristics. The coefficients of determination of the 3 networks were 0.34, 0.60, and 0.84, respectively. These values quantify how well the networks satisfy a scale-free topology, with the threshold being >0.8 (Zhang and Horvath, 2005). Thus, IN was the only scale-free topology network.

Scale-free topology in biological networks seems to result from evolutionary events such as gene duplication and loss (Aloy and Russell, 2004; Sun and Kim, 2011). Gene coexpression, protein-protein interactions, and metabolic networks all have scale-free topology (Barabási and Oltvai, 2004; Khanin and Wit, 2006), as do SNP coassociation networks (Fortes et al., 2012; Widmann et al., 2013). However, our analyses showed that AN is actually the least scale-free. One possible explanation for this discrepancy is that AWM did not account for all relevant genes, leading to missing gene interactions that affect trait expression. Specifically, AWM includes only the nearest gene to selected SNPs, excluding other adjacent genes even if they were in linkage disequilibrium (LD) with significant SNPs. This limitation may be more serious for Japanese Black cattle, because this breed has relatively high whole-genome LD levels (McKay et al., 2007; Ogawa et al., 2014). Thus, AN was likely to contain numerous false positive interactions. Construction from excess interactions and a small fraction of genes probably caused AN to be the least scale-free.

Hub Gene-Centered Subnetwork

Highly connected hub genes in scale-free networks tend to be essential in biological systems (Barabási and Oltvai, 2004; Yu et al., 2007), and their topological importance suggests that they have considerable biological significance. By exploiting IN’s scale-free properties, we found 3 important, topologically centered hub TFs through data from AnimalTFDB (Zhang et al., 2012). The 3 TFs were unconventional prefoldin RPB5 interactor (URI1), ρ-associated protein kinases 2 (ROCK2), and ETS variant 6 (ETV6) (triangles in Figure 5); their respective degrees were 16, 15, and 13, whereas the average degree of IN was 4.28. The subnetwork comprised 32 genes with 43 interactions and contained the top 3 hub TFs along with their immediate neighbors in IN (Figure 5). Most subnetwork genes were widely expressed, except for 3 nervous system-specific genes (CPNE3, ELAVL2, and RAB7B) and 1 gland-specific gene (SV2B).

Figure 5.

Subnetwork extracted from the top 3 hub transcription factors (TFs) and their neighboring genes. Node colors correspond to gene tissue specificity: gland (red), nervous system (blue), and nontissue-specific genes (white). Node shapes indicate gene classification: triangle (TF) and circle (other genes).

The TF with the highest degree was URI1, known to suppress androgen receptor transcriptional activation (Mita et al., 2011). Androgen is associated with feed efficiency in turkeys (Weppelman, 1984). ROCK2, with the second highest degree, is involved in regulating the actin cytoskeleton pathway. Although ROCK2 has no known links with feed utilization or growth traits, actin-cytoskeleton regulation is an important component of feed efficiency and compensatory gain in cattle (Rolf et al., 2011; Keogh et al., 2016). Finally, the TF with the third highest degree was ETV6, associated with pig growth (Puig-Oliveras et al., 2014) and human height (Gudbjartsson et al., 2008).

Corroborating previous research indicating that highly connected TFs tend to be broadly expressed across tissues (Ravasi et al., 2010), the top 3 hub TFs were not tissue-specific. In the subnetwork, the only TF with TFBS information was ETV6. Thus, we examined the promoter regions of the genes that might be targeted by ETV6 to identify corresponding TFBSs (Lee and Huang, 2013, 2014). Eight of the 13 predicted targets, namely, CPNE3, NTRK2, LNX1, TRAPPC9, RAB7B, LOH12CR1, CYYR1, and LMO3, had at least 1 TFBS at P < 0.001. According to DAVID UP_KEYWORDS annotations, 20 of the 32 genes in the subnetwork were related to phosphoproteins (ACSBG1, ANKRD54, DHCR7, ETV6, GPD2, LOH12CR1, LPHN3, MPP7, MSH2, NR1H4, NR5A1, NTRK2, PAK1, RAB7B, ROCK2, SPATA6, SPSB1, TRAPPC9, URI1, and ZDHHC20). This finding suggests that neighboring hub-TF genes may be involved in TF targeting and upstream signaling. As it was beyond the scope of this study, we did not validate these interactions in the subnetwork, but doing so is a valuable direction for future research. The subnetwork also contained prominent candidate genes that are associated with feed efficiency and growth. For instance, USP46 controls mouse feeding behavior (Imai et al., 2013), whereas LMO3 and NTRK2 have physical functions linked to obesity (Xu et al., 2003; Lindroos et al., 2013). These results suggest that the subnetworks obtained are related to feed utilization and growth traits, and therefore, the genes and interactions in the subnetworks should be further investigated for elucidating the biological mechanisms underlying these traits.

On the Current Approach

The GWAS methods often have a difficulty in pinpointing prominent candidate genes at loci detected within a region of high LD. Our current intersection approach would be useful for prioritizing candidate genes through the consideration of tissue specificity beyond the AWM/PCIT approach. In addition to identifying interactions as is the case with AWM approach, the current approach, also using gene expression information, is less reliant on LD. Comparing gene expression data among many tissues allows us to distinguish tissue-specific genes from housekeeping genes. Moreover, correlation of expression profiles across several tissues might suggest common regulatory pathways. The IN obtained exhibited scale-free behavior, and this characteristic is extremely important to find hub components. Previous references suggest that hub-TFs in IN are prominent candidates for feed efficiency (Weppelman, 1984; Mita et al., 2011; Rolf et al., 2011; Puig-Oliveras et al., 2014; Keogh et al., 2016).

Gene expression data can provide useful insights into the mechanisms underlying phenotypic traits. The tissue-specific modules identified with the current approach were consistent with known functions relevant to their respective tissues. The module of reproductive systems suggested relationships between feed utilization efficiency and reproductive tissues, which was somewhat unexpected as reproductive phenotypes were not used in the present GWAS. However, there are some evidences linking reproductive traits to feed efficiency as described below.

The approach presented here is also applicable to quantitative phenotypes with antagonistic correlations, because AWM depends only on the absolute magnitudes of the correlations and not on their directions. Undesirable genetic relationships between feed utilization efficiency and female fertility have been reported (Arthur et al., 2005; Basarab et al., 2007; Crowley et al., 2011). Therefore, a simultaneous analysis of feed utilization and reproductive traits using an approach like the one presented here might have a great potential to elucidate the biological processes underlying the relationships between them.

It should be noted that the transcriptome data used here were retrieved from a public database, containing samples from multiple breeds, and that expression data in each tissue were obtained from a single sample. Therefore, the expression profiles reported here might not reflect those adequately in Japanese Black cattle. If transcriptome data originated from Japanese Black population become available, it would be possible to infer a more precise network. Multiple samples per tissue will also allow examining individual differences in gene expression.

Functional genomic selection using knowledge on gene functions and interactions has recently received attention (Snelling et al., 2013). In the model for genomic BLUP, Gao et al. (2017) considered gene interaction effects based on gene annotation data, and Fang et al. (2017) used transcriptome data to focus on the genes and regulatory elements likely to affect phenotypic traits. Studies of this type have shown that the use of biological information can improve the accuracy of the predictions, and therefore, our current approach may provide useful information for functional genomic selection.

SUPPLEMENTARY DATA

Supplementary data are available at Journal of Animal Science online.