Genomic Signatures Reveal New Evidences for Selection of Important Traits in Domestic Cattle

Abstract

We investigated diverse genomic selections using high-density single nucleotide polymorphism data of five distinct cattle breeds. Based on allele frequency differences, we detected hundreds of candidate regions under positive selection across Holstein, Angus, Charolais, Brahman, and N'Dama. In addition to well-known genes such as KIT, MC1R, ASIP, GHR, LCORL, NCAPG, WIF1, and ABCA12, we found evidence for a variety of novel and less-known genes under selection in cattle, such as LAP3, SAR1B, LRIG3, FGF5, and NUDCD3. Selective sweeps near LAP3 were then validated by next-generation sequencing. Genome-wide association analysis involving 26,362 Holsteins confirmed that LAP3 and SAR1B were related to milk production traits, suggesting that our candidate regions were likely functional. In addition, haplotype network analyses further revealed distinct selective pressures and evolution patterns across these five cattle breeds. Our results provided a glimpse into diverse genomic selection during cattle domestication, breed formation, and recent genetic improvement. These findings will facilitate genome-assisted breeding to improve animal production and health.

Introduction

As one of most important farm animals, cattle are used for a variety of purposes including dairy, beef, hide production, and labor. The majority of the global cattle population can be classified into one of two groups: Humpless (taurine) and humped (indicine or zebu) cattle with dramatic phenotypic differences (Bradley et al. 1996; Bovine HapMap Consortium 2009). Earlier studies indicated that these two subspecies diverted from the last common ancestor between 275,000 and 850,000 years ago (MacHugh et al. 1997; Hiendleder et al. 2008). Each subspecies appeared to be separately domesticated with taurine cattle in the Fertile Crescent approximately 10,000–8,000 years ago and indicine cattle in the Indus Valley approximately 6,000–8,000 years ago (Loftus et al. 1994; Larson et al. 2014). A third independent domestication was proposed in Africa (Troy et al. 2001); however, a recent result argued against it (Decker et al. 2014). After these domestication events, continuous human selective breeding for desirable traits like easy management and high production has been more documented in taurine cattle than in indicine cattle. Since the early 1800’s, breed development was based on phenotype selection on coat color and polled phenotypes, and included the imposition of severe bottlenecks followed by breed expansion via artificial insemination. During the last 50 years, animal breeding based on quantitative genetics has resulted in a remarkable progress in improving production traits for milk and meat (Andersson and Georges 2004). Therefore, selection (natural and human-imposed) and nonselective forces (the demographic events and introgression) drove changes within the cattle genome. Their combined effects have created exceptional phenotypic diversity and genetic adaptation to local environment across the globe within the modern cattle breeds.

It is generally accepted that there are four mechanisms of evolutionary change: Mutation, genetic drift, gene flow or migration (demographic history), and selection. However, only selection is locus specific, while the first three forces work uniformly across the whole genome. Selection can be divided into three modes: Positive, purifying (or negative selection, eliminating a deleterious mutation), and balancing selection (including heterozygote advantage and frequency-dependent selection). Positive selection is a mode of natural selection that drives the increase in prevalence of advantageous alleles due to their favorable effects on fitness (Biswas and Akey 2006; Kelley and Swanson 2008; Oleksyk et al. 2010). Genetic hitchhiking refers changes in the frequency of an allele because of linkage with a positively selected or neutral allele at another locus. The availability of genomic data has spurred many approaches for mapping positive selection, mainly based on reduced local variability, deviations in the marker frequency, increased linkage disequilibrium (LD), and extended haplotype structure. These methods such as CLR, CMS, F_ST, EHH, iHS, and hapFLK (Tajima 1989; Fay and Wu 2000; Sabeti et al. 2002; Nielsen et al. 2005; Voight et al. 2006; Grossman et al. 2010; Fariello et al. 2013) have been widely used in human, mouse, rat, and domesticated animals like dogs, cattle, sheep, pigs, horses, and chickens (Waterston et al. 2002; Gibbs et al. 2004; Rubin et al. 2010, 2012; Kijas et al. 2012; Petersen et al. 2013). One method (di) was recently developed to identify genomic regions indicative of selection with a high degree of genetic differentiation between dog breeds (Akey et al. 2010). Distinct from F_ST, which measures the fraction of total genetic variation between two populations, the di value is defined as a function of unbiased estimates of all pairwise F_ST between one breed and the remaining breeds within a population. It is suited for detecting selection specific to a particular breed, or subset of breeds, and isolating the direction of change. It was utilized to track lineage-specific signatures of selection in the dog and horse genomes, revealing its power to detect selection acting on both newly arisen and preexisting variations (Akey et al. 2010; Petersen et al. 2013).

Selection mapping is a powerful approach, together with genome-wide association studies, to detect candidate genes associated with quantitative traits. Selection mapping in cattle has been previously investigated using a lower density markers like BovineSNP50 array (Flori et al. 2009; Hayes et al. 2009; Qanbari et al. 2010, 2011; Stella et al. 2010; Rothammer et al. 2013). Only recently similar studies were reported based on a higher density markers like BovineHD array (Porto-Neto et al. 2013; Utsunomiya et al. 2013; Kemper et al. 2014; Perez et al. 2014). More recently, sequence-based signatures were reported in Fleckvieh (Qanbari et al. 2014). However, these studies focused on limited breeds with specific traits. Therefore, it is possible many breed-specific selection signatures remain undetected due to lack of comparison across breeds. There are a few of targeted studies of the haplotype pattern and evolution on selected gene families like Toll-like receptors in cattle (Seabury et al. 2010). However, to our knowledge, no systematic effort has been reported to investigate the haplotype pattern and evolution of positively selected genes in the cattle genome.

In this study, we investigated diverse genomic selection using high-density single nucleotide polymorphism (SNP) data of five distinct cattle breeds, including Holstein (HOL), Angus (ANG), Charolais (CHL), Brahman (BRM), and N'Dama (NDA). HOL, ANG, and CHL are taurine breeds from Europe. HOLs represent the highest-production dairy animals, originally from the Netherlands and northern Germany. Their black-and-white color was due to artificial selection by the breeders. ANG cattle, first developed in Scotland, are used in beef production. They are naturally polled (do not have horns) and solid black or red in color. CHL is a dual purpose breed (both milk and beef) originated in France, which is known for its large body size, bone structure, and white to cream coat. NDA is an indigenous local taurine breed from West Africa. With a small size and fawn coat, NDA is well known for its trypanotolerant and shows superior resistance to ticks and other parasites (http://www.ansi.okstate.edu/breeds/cattle/, last accessed December 2, 2014). BRM is a composite of several zebu breeds imported from India (Guzerat, Kankrej, Gir, and others), and was first bred in America in the 1880s for beef production with a minor taurine contribution (Decker et al. 2014). The BRM is known for its gray coat, heat tolerance, and disease resistance. We performed a genome-wide scan with the BovineHD SNP genotypes to map selection signatures among these five diverse cattle breeds. Selected regions were further independently validated by next-generation sequencing. We also conducted a genome-wide association analysis of milk production traits for 26,362 HOLs and found that regions under selection were likely functional. In addition, we investigated haplotype pattern and evolution of positively selected genes in the cattle genome.

Results

Genetic Diversity, Breed Relationship, and LD

We chose five genetically diverse and geographically distinct breeds to systematically investigate the selection signatures in cattle (supplementary table S1, Supplementary Material online), including HOL, ANG, CHL, BRM, and NDA. Several quality control steps were used to assure SNP data quality. These filters included missing data per individual, missing data per marker, and allele frequency. A total of 710,681 SNPs were kept after these QC steps. We estimated the average heterozygosity using the filtered SNPs within each breed, and results indicated that significant genetic diversities exist among five breeds. The average heterozygosities for the three European taurine breeds were just over 0.3 for HOL, ANG, and CHL whereas they were lower for the indicine breed BRM (0.25) and the African taurine breed NDA (0.22) (supplementary table S2, Supplementary Material online). The inbreeding coefficient, F, was also estimated for the five breeds. The average values of F were less than 0.1 for the three European taurine breeds, just over 0.25 for BRM and NDA (supplementary table S2, Supplementary Material online).

Multidimensional scaling (MDS) analysis revealed a typical triangular relationship among the five breeds from three continents, with the first dimension separating the taurine and indicine breeds, and the second dimension separating the European taurine from African taurine (supplementary fig. S1A, Supplementary Material online). The well-separated clusters indicated that the chosen samples were powerful to explore the genomic characteristics of these five breeds. Admixture analysis was also performed using 72,945 LD-filtered SNPs to identify ancestry and admixture level among the five breeds when the number of clusters, K, varied from 2 to 5. When K = 2, the clustering pattern reflects the primary, predomestication division of taurine from indicine cattle. When K = 3, NDA separates from the European breeds, suggesting a early divergence. When K = 5, the five breeds can be unequivocally separated into five closed endogamous breeding units with expected levels of admixture (supplementary fig. S1B, Supplementary Material online). We also used PLINK to construct a neighbor-joining tree based on pairwise nucleotide genetic distances as shown in supplementary fig. S1C, Supplementary Material online (Saitou and Nei 1987; Purcell et al. 2007). Two main features are obvious: 1) cattle from the same breed always cluster together and 2) little structure is observed in the internal branches within a breed. The patterns were consistent with the fact that modern breeds were derived from population bottlenecks with closed gene pools. Modern breed formation happened within a short period of time, and a certain amount of genetic variations exist within a breed. We also analyzed LD decay patterns in the five breeds. We found that LD in cattle genome generally decays rapidly and the decay rates could vary greatly among breeds and chromosomes, as reported previously (Bovine HapMap Consortium 2009; Espigolan et al. 2013; Qanbari et al. 2014).

Candidate Regions and Genes under Positive Selection

We used selection mapping to explore the differences among breeds selected for different purposes (milk or beef). We calculated average di values in a nonoverlapping sliding window way for all 11,651 widows of 50 neighboring SNPs. The average length and standard deviation (SD) of the 50-SNP sliding windows was 212 ± 110 kb. The genome-wide distribution of the window di values is shown in figure 1. For each breed, we defined candidate selection regions using two thresholds: The top 1% or 5% windows with highest average di values in the empirical distribution (Voight et al. 2006; Pickrell et al. 2009). Thus, a total of 117 or 583 windows were identified under these two criteria for each breed (supplementary table S3, Supplementary Material online). To keep a high specificity, our results were mainly based on the top 1%, that is, 117 windows unless stated otherwise.

Fig. 1.

Genomic distribution of selection regions in five cattle breeds. The distribution of di for each 50-SNP windows across all auto chromosomes is shown for each breed. Alternating color indicate di values from adjacent chromosomes. Breeds are abbreviated as described in supplementary table S1, Supplementary Material online.

To investigate the shared or breed-specific selection regions, we plotted a Venn diagram (fig. 2) and a Circos plot (supplementary fig. S3, Supplementary Material online) (Zhang et al. 2013), and counted the numbers of overlapping windows among the five breeds. Out of 470 merged nonredundant windows, 383 windows (∼81.5%) were unique to only one breed. On the other hand, 87 windows (∼21.2%) were shared by two or more breeds, and within them 24 windows (∼5.1%) were shared by three or more breeds.

Fig. 2.

Venn diagram for shared versus breed-specific selection events among five cattle breeds. Top 1% (117) windows of the average di values were compared across five breeds.

Based on the bovine RefSeq gene annotation, there were 44, 38, 44, 39, and 23 windows, which did not overlap with any gene in HOL, ANG, CHL, BRM, and NDA, respectively (supplementary table S3, Supplementary Material online). We then focused on the left windows which overlap with genes. Out of a total of 13,775 genes, we detected 1,133 and 3,480 genes (8.23% and 25.26%) within the top 1% and 5% windows, respectively (table 1 and details in supplementary table S3, Supplementary Material online). And both of them were significantly enriched for gene content as determined in our 10,000 permutation tests. We further performed DAVID analyses on breed-specific loci for each breed separately, with genes in shared windows excluded (supplementary table S4, Supplementary Material online). Interestingly, we found a number of functional clusters that were significantly enriched among these breeds. In ANG, the top two enriched clusters were defense response and the voltage-gated potassium channel complex. In HOL, enrichments were found for lymphocyte activation, protein kinase activity, ATP binding, ubiquitin conjugation, and mitochondrial lumen and matrix. In CHL, transcription coactivator activity, calcium binding, and protein ubiquitination were overrepresented. In NDA, significantly enriched clusters were for T-cell activation, hematopoiesis, carbohydrate catabolic process, and bone, skeletal, and embryo development. Finally, BRM showed distinct, extensive and strong enrichments for 12 clusters, including cytoskeleton, adherens junction, intracellular organelle lumen, activities for arylesterase, peptidase, ligase, protein dimerization and adenyl nucleotide binding, and response to DNA damage stimulus and stress.

Table 1.

Selected Genes under Positive Selections.

Gene	Long Name	Breed	Genome Position	Traits	Cattle Studies	Studies in Other Species
Known genes in cattle
LAP3	Leucine aminopeptidase 3	CHL	chr6:38,574,590–38,600,027	Milk production traits	Flori et al. 2009
LCORL	Ligand dependent nuclear receptor corepressor-like	CHL	chr6:38,840,864–38,992,112	Body size	Flori et al. 2009	Rubin et al. 2012; Petersen et al. 2013
KIT	V-kit Hardy-Zuckerman 4 feline sarcoma viral oncogene homolog	CHL; HOL	chr6:71,796,318–71,917,430	Coat color	Stella et al. 2010	Kijas et al. 2012; Rubin et al. 2012
MC1R	Melanocortin 1 receptor	ANG; BRM; CHL; HOL	chr18:14,757,332–14,759,082	Coat color	Flori et al. 2009	Petersen et al. 2013
NUDCD3	NudC domain-containing protein 3	HOL	chr4:77,598,538–77,670,088	Mitosis	Flori et al. 2009
WIF1	WNT inhibitory factor 1	ANG; BRM; CHL	chr5:48,917,722–49,009,466	Mammalian mesoderm segmentation	Bovine HapMap Consortium et al. 2009
SAR1B	SAR1 homolog B	HOL	chr7:47,717,605–47,747,828	Milk production traits	Larkin et al. 2012
ASIP	Agouti signaling protein	ANG; HOL	chr13:64,213,312–64,239,962	Coat color		Norris and Whan 2008
LRIG3	Leucine-rich repeats and immunoglobulin-like domains protein 3	CHL	chr5:54,884,394–54,936,361	Elongated body axis		Akey et al. 2010
Known genes in other domesticated animals:
FGF5	Fibroblast growth factor 5	BRM	chr6:96,723,382–96,746,176	Coat color		Cadieu et al. 2009
OSTN	Osteocrin	CHL; HOL	chr1:76,685,699–76,721,920	An inhibitor of osteoblast		Rubin et al. 2012
NCAPG	Non-SMC condensin I complex, subunit G	CHL	chr6:38,765,969–38,812,055	Body size		Petersen et al. 2013
ABCA12	ATP-binding cassette, subfamily A (ABC1), member 12	BRM	chr2:103,520,024–103,720,887	Ichthyosis fetalis; skin keratinization, calf birth weight	Charlier et al. 2008; Cole et al. 2014	Tennessen and Akey 2011
ACTN3	Actinin, alpha 3	BRM	chr29:45,230,630–45,242,406	Skeletal and cardiac muscle		Gu et al. 2009; MacArthur et al. 2007
NPR2	Atrial natriuretic peptide receptor 2 precursor	BRM	chr8:60,368,097–60,379,016	Skeletal morphology and body size
						Kijas et al. 2012
			chr8:60,381,937–60,385,961

Genes under Positive Selection

We identified genes from positive selection regions and then compared them with published literature. We categorized them into three classes. Class 1, genes reported before in cattle: We detected many genes which have been previously identified under positive selection in cattle (table 1; Bovine HapMap Consortium 2009; Flori et al. 2009; Hayes et al. 2009; Stella et al. 2010; Larkin et al. 2012). We found two well-known genes related to coat color, that is, KIT in CHL and HOL and MC1R in ANG, BRM, CHL, and HOL. We also identified NUDCD3 in HOL and WIF1 in ANG, BRM, and CHL, which may be involved in cell division and developmental pathways (Zhou et al. 2006; Cai et al. 2009). In addition, LAP3 (leucine aminopeptidase 3) and LCORL (ligand-dependent nuclear receptor corepressor-like) were detected in CHL, which are associated with milk production traits and body size, respectively (Flori et al. 2009; Pryce et al. 2011). We further detected gene SAR1B (SAR1 homolog B) under selection in HOL, which was briefly reported to be associated with milk production and disease resistance in dairy cattle (Larkin et al. 2012). Class 2, genes reported before in other domesticated animals: We discovered some genes which were previously reported to be under positive selection in other domesticated species but not in cattle. These included coat color related genes (ASIP and FGF5), body size-related genes (NPR2, NCAPG, and LRIG3), and skeletal muscle-related gene (OSTN). Class 3, genes reported bin human: We further detected that the following bovine genes, whose human ortholog genes were previously reported to before under selection in human populations. For example, signals of selective sweeps of NTRK2, CNGA3, ETFA, ISL2, KCND2, and SYN3 were found in human populations, according to geographic locations (Pickrell et al. 2009). Evidence for positive selection of IRAK3, LLPH, TMBIM4, BTBD11, GCK, AKAP3, and CISH were reported in African populations (Granka et al. 2012; Jarvis et al. 2012). Four blood pressure related genes (CLCN6, KIAA2013, FBXO2, and FAM117A) were identified in a human genome-wide association study (Newton-Cheh et al. 2009). Promoter regions of many neural- and nutrition-related genes (RPL37A and GLMN) also experienced positive selection during human evolution (Haygood et al. 2007).

When we relaxed the threshold from 1% to 5%, we obtained more previously reported bovine genes. For example, GHR (growth hormone receptor) on chr20 has effects on milk yield and composition in dairy cow (Blott et al. 2003; Viitala et al. 2006) and body size in dog (Rimbault et al. 2013, Wiener and Wilkinson 2011). Other additional genes are related to muscle formation and fatty (ACTC1, FABP3, and HBEGF) (Gerbens et al. 1997; Rahman et al. 2010; Stella et al. 2010; Qanbari et al. 2012); mammary gland and daily production (ITFG1, ITGB3, RTN4, SULT1E1, and TMCC3) (Lemay et al. 2009; Barreiro and Quintana-Murci 2010; Stella et al. 2010; Larkin et al. 2012); and immunity and nervous system (CD79A, NEUROD6, and SPOCK1) (Kosiol et al. 2008; Bovine HapMap Consortium 2009; Gautier et al. 2009). If we include published results from other species, we found more genes related to coat color (KRT71 and RSPO2) (Cadieu et al. 2009), body size (BMP2, IGF1) (Sutter et al. 2007; Kijas et al. 2012; Petersen et al. 2013), nervous system (MBP, SMO, and TLX3) (Axelsson et al. 2013), and immune process (CCR2) (Kijas et al. 2012).

Validation by Next-Generation Sequencing

To validate high-density SNP array results, we carried out additional analyses based on next-generation sequencing. Because LAP3 was reported before to be under positive selection in HOL (Flori et al. 2009) and we also identified LAP3 in CHL and CHL is a dual purpose breed (both milk and beef), we chose to investigate selection signature in HOL near the LAP3 gene. Using SNP calls generated from next-generation sequencing (NGS) data, we checked the 24.1-kb region (92 SNPs) flanking the LAP3 locus in HOL and ANG. We plotted MAF and calculated F_ST between HOL and ANG to look for high degrees of breed differentiation. Their trends agreed well with our di results derived from the BovineHD array. As expected, there was an excess of high F_ST values among the 92 SNPs, although F_ST values varied to a certain degree (fig. 3). The elevated F_ST was predominantly due to SNPs located near LAP3, with allele frequencies that were generally higher in HOL than in ANG. This elevated F_ST revealed and confirmed the signature of selection near the LAP3 gene.

Fig. 3.

MAF and F_ST plot for LAP3 genes between HOL and ANG using 92 SNPs extracted from next-generation sequencing.

Association Tests

To further validate the genes involved in selection for milk production, we performed association tests using 26,362 BovineSNP50 array data from HOL to test whether LAP3 and SAB1B were related to five milk production traits (milk yield [MY], fat yield [FY], protein yield [PY], fat percentage [FP], and protein percentage [PP]). Within a 652-kb region containing 12 SNPs from the BovineSNP50 array, LAP3 gene located between the sixth (Hapmap26308-BTC-057761) and seventh SNPs (Hapmap43470-BTA-114677) (fig. 4). We identified 5, 5, 3, 5, and 1 significant SNPs associated with MY, FY, PY, FP, and PP traits, respectively (-log10 P value ≥ 7 as the threshold). Based on the high-density HOL BovineHD data, we observed two large haploblocks in strong LD with the same 652-kb region. Among them, one 155 kb block from BovineHD0600010649 to BovineHD0600010693 surrounding LAP3 also contains the sixth (Hapmap26308-BTC-057761) and seventh SNPs (Hapmap43470-BTA-114677), which were significantly associated with FP (fig. 4). Similar haplotype LD patterns were found in the neighboring region of 600 kb (12 SNPs) around SAR1B (supplementary fig. S4, Supplementary Material online). We identified 2, 2, 2, 1, and 1 associated SNPs for MY, FY, PY, FP, and PP, respectively. There were three long and tightly linked neighboring haploblocks in the proximity of SAR1B. The sixth SNP (BTB-01846474) was also significantly associated with FP (supplementary fig. S4, Supplementary Material online).

Fig. 4.

Association test of LAP3 with HOL milk production traits and haplotype LD analyses in HOL, BRM, and NDA.

Haplotype Homozygosity

One genetic signature of an incomplete selective sweep is a region of extensive LD (termed extended haplotype homozygosity or EHH) and low variation on high-frequency chromosomes carrying the derived beneficial mutation relative to chromosomes with the ancestral allele. Therefore, we compared the haploblock LD patterns around LAP3 and SAR1B and found dramatic differences among the five breeds. We calculated D′ and r² between SNPs to evaluate the LD pattern in these regions. Pairwise comparisons show that strong LD within sites over short distance. For the locus near LAP3, we detected an extensive high LD level in HOL as compared with BRM and NDA (fig. 4). It is worthwhile to note that NDA display an intermediate LD level between HOL and BRM. BRM’s haploblocks seem short and not tightly linked in these regions. A similar LD pattern was found in the locus around SAR1B (supplementary fig. S4, Supplementary Material online).

The recombination rate is another important parameter, which influences the LD level and the strength and frequency of positive selection. We estimated fine-scale recombination rates based on approximate conditionals model near these genes. Although we did detect dramatic differences in recombination rates across five breeds near genes like GHR and ABCA12 (supplementary fig. S5, Supplementary Material online), all breeds showed little recombination near LAP3 and SAR1B. In summary, although the differences between HOL and BRA could be related to ancient events like subspeciation, the comparison between HOL and NDA agree well with the hypothesis that the selected allele and its haplotype could rise rapidly in frequency through genetic hitchhiking in the process of a strong selective sweep.

Haplotype Network Analysis

We further investigated haplotype pattern and evolution to further characterize selective pressures near selected genes including LAP3 and SAR1B. We obtained 131 haplotypes within the 175.7 kb haploblock region near LAP3. We detected the top four haplotypes: H1, H2, H3, and H4 with a frequency of 13.41%, 7.33%, 7.04%, and 6.12%, respectively (fig. 5). Among them, top haplotype H1 contained CHL (53.13%), HOL (1.41%), and ANG (8.84%). H2 contained only NDA (59.03%), but not any of the other four breeds. H3 contained HOL (17.18%), ANG (6.31%), and CHL (5.37%). We also observed three exclusive haplotypes for NDA. Fourteen haplotypes clustered together only in the BRM samples. Altogether, this pattern indicated separate haplotypes were clustered only for BRM or NDA, while overlapping haplotypes were identified for HOL, ANG, and CHL. It is interesting to note that most BRM haplotypes cluster exclusively together in one individual branch.

Fig. 5.

Haplotype networks of two loci. (A) LAP3 and (B) SAR1B. Each node represents a different haplotype, with the size of the circle proportional to frequency. Branch lengths are proportional to the number of nucleotide differences. Circles are color coded according to population (red: HOL, blue: ANG, yellow: CHL, green: BRM, and gray: NDA).

We obtained 95 haplotypes within the 404.0 kb haploblock region near SAR1B. The top haplotype, H1 (with frequency of 37.32%), was found in HOL (77.29%), ANG (52.54%), CHL (23.15%), and BRM (1.67%) (fig. 5). H2 (with frequency of 10.90%) included HOL (13.62%), ANG (16.74%), CHL (14.73%), BRM (1.67%), and NDA (1.14%). H3 (5.99%) had a large proportion of NDA (37.83%), ANG (1.25%), and CHL (4.82%). A similar general trend was found: Separate haplotypes were clustered only for NDA or BRM, while overlapping haplotypes were identified for HOL, ANG, and CHL.

We also studied genes related to growth and body size (LCORL, LRIG3, FGF5, and NCAPG), coat color (KIT and MC1R), and other functions, obtaining similar results (supplementary fig. S6, Supplementary Material online). Combined with LD pattern results, the haplotype network analyses indicated 1) common overlapping haplotypes were often identified for HOL, ANG, and CHL, while separate distinct haplotypes were clustered only for NDA or BRM; 2) the high SNP diversity and differential LD patterns near some genes, suggesting that they have been under different evolution pressure in these five breeds.

Discussion

Diverse geographic adaptation and selection have contributed to the shared and population-specific phenotypes in many species. For example, environment contributed to the process of selection and evolution of bovine genes in the cattle breeds of West Africa (Gautier et al. 2009) and the Senepol breed of Caribbean (Flori et al. 2012). Previous studies have shown that purifying selection is the most dominant mode of selection on genes of the innate immune system in human and cattle (Mukherjee et al. 2009; Seabury et al. 2010). To explore the cattle breed diversity and breed-specific selection signatures, we compared the significant candidate regions among five breeds. Shared selection regions across populations were found in five cattle breeds. One possibility is that different breeds have been under similar artificial selection intention. It is interesting to note that many of these genes have also been reported in other species. One possibility is that these genes carried out similar functions and went through similar adaptations in different species. On the other hand, we also found the unique loci, which likely contain genes and variations that confer breed-specific phenotypes. For example, DAVID analysis for the genes specific to individual breeds revealed differentially enriched molecular functions, suggesting that each breed was under different selections for unique phenotypes. We detected 12 DAVID clusters showing positive selection only in BRM. These genes may be related to the fact that founding zebu breeds for BRM were developed with inadequate food supplies, insect pests, parasites, and diseases in a tropical climate. Extensive breed-specific regions agreed well with the idea that differential selection often sorts individuals into separate breeds with distinct phenotypes.

The diversity of selection signatures across the genome could provide critical insights for the understanding of gene function and evolution. Our results have revealed a series of genes involved in coat color, growth, and milk production under selection. Coat color variation in domestic animals is evidence of phenotypic adaptation in response to selection (Hubbard et al. 2010). Genes underlying such variation are of considerable interest in breed formation and tracing molecular evolution in domestic animals. For example, in the domestic pig the evolutionary pattern of MC1R demonstrated that coat color phenotypes were likely a direct result of artificial selection (Fang et al. 2009). KIT is known to play roles in coat color in many species including cattle (Hayes et al. 2010; Stella et al. 2010; Kemper et al. 2014; Qanbari et al. 2014), pig (Amaral et al. 2011; Rubin et al. 2012), horse (Haase et al. 2007; McCue et al. 2012), and sheep (Kijas et al. 2012). This gene was also reported to be related to reproduction (Koch et al. 2009), and expressed in the lactating bovine mammary gland (Flori et al. 2009; Lemay et al. 2009). Segregating quantitative trait locus around GHR affected milk composition and yield (Blott et al. 2003; Khatkar et al. 2004; Viitala et al. 2006). In this study, we found differential recombination rates near GHR in five breeds, which correlated to the distinct LD patterns in taurine (as represented by HOL) and indicine breeds (BRM).

LAP3 catalyses the removal of N-terminal amino acids and is involved in protein maturation and degradation. LAP3 (chr6:38,574,590–38,600,027) is only 544 kb downstream of to ABCG2 (chr6:37,959,536–38,030,586), which affects milk production (Cohen-Zinder et al. 2005; Olsen et al. 2005). LAP3 and nearby genes NCAPG and LCORL have also been found to be associated with direct calving ease in Piedmontese cattle (Bongiorni et al. 2012) and feed intake and growth in beef cattle (Lindholm-Perry et al. 2011). Positive selection of LAP3 has been reported in European HOLs, Chinese HOLs, and five Italian cattle breeds (Gautier et al. 2009; Pan et al. 2013; Mancini et al. 2014). The haplotype diversity near LAP3 in the haplotype network analysis reflected diverse evolutionary patterns across breeds, and no significant recombination event was observed in this gene, so we conclude that differential selection pressures are responsible for the diverse haplotype distributions across breeds. SAR1B belongs to the Sar1-ADP ribosylation factor family of small GTPases, which govern the intracellular trafficking of proteins in coat protein-coated vesicles (Schekman and Orci 1996). It is annotated by the GO term of secretary pathway as it is associated with the milk fat globule membrane in dairy cattle (Lemay et al. 2009). It was first reported under selection in HOL in a whole-genome resequencing and haplotype-phasing effort (Larkin et al. 2012). In this study, we provided additional evidence that this gene is under positive selection as it is included in a couple of major haplotypes. Our haplotype network results shed light into the evolution of the SAR1B haplotypes across breeds. For instance, we found dominant haplotypes were common across multiple breeds, and the most plausible explanation is that positive selection in multiple breeds resulted in their increased frequencies. On the other hand, breed-specific haplotypes and changes in gene frequency for particular haplotypes may represent the imprint of diverse selections for an individual breed.

Our data generally agree well with the common beliefs that NDA and the ancestral founding breeds for BRM have a long history of local geographical adaptation while the three European taurine breeds (HOL, ANG, and CHL) went through recent strong artificial selection for milk and meat production. We found that LD patterns estimated using high-density SNP array offer detailed LD characteristics across breeds. The diverse LD patterns may indicate changes in genomic regions in different populations after selection for a particular allele, the various timescales of cattle domestication and breed formation, or specific characteristics of traits under selection.

Potential Pitfalls

Our study had detected many genes previously reported to be under positive selection. However, some genes were not detected by the di statistics, and this situation was reported before (Petersen et al. 2013). One example was the RXFP2 gene (ENSBTAG00000015132, chr12:29234959–29280832), which was found in polled sheep and cattle (Kijas et al. 2012; Allais-Bonnet et al. 2013) but not detected in our studies. Another example was the 547 kb polled locus in the 1.5- to 2.0-Mb region on chromosome 1 in cattle (Drogemuller et al. 2005; Medugorac et al. 2012; Seichter et al. 2012). We did find some marginally significant differences in terms of pairwise F_ST value and minor alleles frequency among five breeds (supplementary figs. S7 and S8, Supplementary Material online) near these loci, but they did not achieve significance due to different study designs, samples, and statistical methods. Window size and target locus length could also affect the detection. For example, genes with limited numbers of SNPs (such as ASIP) need further analysis. Another example is LAP3, which only contains 13 SNPs from BovineHD array. Luckily, with F_ST test using 92 NGS-based SNPs and LAP3 haplotype network analysis, our results revealed direct evidence of positive selection signals involved in milk production trait in HOL.

This study applied the di statistic to detect signatures of positive selection using common, moderate-frequency SNP data. It is noted that LD as measured by r² depends on allele frequencies and the SNPs included on the Illumina Bovine arrays, which were selected based on their allele frequency and genome coverage (Pritchard et al. 2000; Matukumalli et al. 2009). With sequencing costs decreasing rapidly, high-quality, high-coverage whole-genome sequence information will make it feasible to study positive selection free of ascertainment bias.

While this study was being completed, Kemper et al. reported that selection for complex traits leaves little or no classic signatures of selection (Kemper et al. 2014). The authors attributed this to the fact that “selection response is caused by weak selection at many sites across the genome, probably for previously segregating variants”. Several lines of evidences (multiple independent studies, different samples and methods, NGS data, and haplotype network analyses) suggest our results are probably not analytical artifacts. It is well known, even in human population genomics studies, that different samples and methods often make it difficult to compare the results from different selection mapping efforts (Kelley et al. 2006; Akey 2009; Hohenlohe et al. 2010).

Conclusion

We carried out a genome-wide scan of selection signature using high-density SNP array across five diverse cattle breeds. Our analysis identified multiple genes under positive selection that are related to coat color, milk production, growth and body size, and other functions. Even though some genes have been previously reported, we confirmed the positive selections using high-density SNP arrays and more diverse samples. We further validated them (LAP3 and SAR1B) with next-generation sequencing and association tests. Haplotype diversity and network analyses revealed distinct selective pressures and evolution patterns across these five cattle breeds. Our results are unique in indicating that diverse genomic selections during speciation, domestication, breed formation, and recent genetic improvement could contribute to the cattle breed diversity.

Materials and Methods

Samples and SNP Data Quality Control

We retrieved a total of 169 samples genotyped on Illumina BovineHD BeadChip (Illumina Inc., San Diego, CA) from five cattle breeds, including HOL, ANG, CHL, BRM, and NDA. On the BovineHD Genotyping BeadChip, SNP markers were uniformly distributed throughout the cattle genome with a median interval less than 3 kb. We used PLINK 1.7 (Purcell et al. 2007) and custom in-house R scripts to convert file formats, manage data, and control quality. To minimize close relationships and genome sharing among samples, we removed closely related individuals when the pi-hat value (an identity-by-descent or IBD estimation) was more than 0.4. All chosen samples showed a genotyping success rate of more than 99%. The SNPs were filtered with MAF ≥ 0.05 and geno ≥ 0.1 (i.e., only SNPs with a 90% genotyping rate or higher). A total of 710,681 autosomal SNPs passed these filters. We then used 581,820 informative autosomal SNPs across breeds to calculate the di values.

Population Structure and Phylogenetic Analysis

MDS analysis was conducted using a total of 72,945 SNPs after LD-based pruning (>0.2). Pairwise genetic distances (4 dimensions) were used to identify the relationship between populations with PLINK (command line arguments: -mds -plot 4). Population structure was examined using Structure 2.3 (Pritchard et al. 2000; Falush et al. 2003). Each analysis was performed using 5,000 replicates and 2,000 burn-in cycles under admixture and correlated allele frequencies models. Genetic distance (D) between pairwise combination of individuals was calculated using PLINK (Purcell et al. 2007), where D = 1-(IBS2+0.5IBS1)/N: IBS1 and IBS2 are the number of loci that share either one or two alleles identical by state (IBS), respectively and the N is the number of loci (Stevens et al. 2011). Neighbor-joining phylogenetic trees were built based on pairwise genetics distance using PHYLIP 3.69 (http://www.phylip.com/, last accessed December 2, 2014). The phylogenetic trees were visualized with Figtree 1.3.1 (http://tree.bio.ed.ac.uk/software/figtree/, last accessed December 2, 2014).

Genome-Wide LD Estimation

To estimate genome-wide levels of LD in five breeds, pairwise LD measures for all retained SNPs were performed using the PLINK “-ld” option. We used the default window size of 1 Mb, and set “–ld-window-r2” to 0 in order to get all pairs reported. The LD decay along genomic distance was fitted by smooth.spline function in R (version 2.15.1).

Selection Mapping

We measured the locus-specific divergence in allele frequencies for each breed based on unbiased estimates of pairwise F_ST, that is, di, as described previously (Akey et al. 2010). Briefly, for each of 581,820 SNPs, we calculated the expected value and SD of F_ST between breeds i and j. For each breed, di was averaged over the SNPs contained within nonoverlapping windows of 50 SNPs. There were a total of 11,651 windows in the cattle genome. The top 1% (117) or 5% (583) regions with highest average di scores were used in the downstream analyses. Custom Perl scripts were used to annotate regions with RefSeq, Ensemble, and human-mapped genes. We performed enrichment analysis using DAVID Functional Annotation Tools (Huang et al. 2009). Only clusters with enrichment scores more than 1 (P value < 0.10) were considered.

Validation with Next-Generation Sequencing

Individual HOL and ANG samples were sequenced using the Illumina HiSeq2000 platform, with library preparation and sequence generation according to the manufacturer's protocols. Sequences were aligned against the UMD 3.1 reference genome and read-pairing was performed with the Burrows-Wheeler Aligner (BWA) v. 0.7.3a-r367 (Li and Durbin 2010). The SAM files were converted to BAM files and unmapped reads were removed with SAMtools v. 0.1.18 (Li et al. 2009). File sorting, marking and removal of duplicates, and indexing were performed with Picard v. 1.88 (http://broadinstitute.github.io/picard/, last accessed December 2, 2014). The resulting files were realigned and recalibrated with the Genome Analysis Toolkit v. 2.3-3 (GATK) (McKenna et al. 2010; DePristo et al. 2011) using a test data set containing known variant sites on LD, BovineHD and BovineSNP50 arrays as well as validated variant sites obtained from dbSNP (http://www.ncbi.nlm.nih.gov/SNP/, last accessed December 2, 2014). Default settings were used for BWA, Samtools, Picard, and GATK. SNPs and indels were called separately using the GATK Unified Genotyper within a 200 bp window surrounding the targeted intervals. Multisample calling was employed to detect SNPs in all samples. Variants were selected using the variant quality score recalibration threshold of 2.30 (99% of known high-quality SNPs identified). SNPs were filtered if their call rates were less than 80% across all HOL and ANG samples. F_ST and MAF frequency plot for LAP3 genes across HOL and ANG used 92 SNPs extracted from the GATK calls. F_ST between two breeds were calculated by using weighted analysis of variance in Genepop (Rousset 2008).

Association Tests for Milk Production Traits

We chose to confirm two genes, LAP3 and SAR1B, under positive selection in a genome-wide association analysis involving milk production traits for HOLs. We retrieved and analyzed BovineSNP50 BeadChip (Illumina Inc.) data for 26,362 HOLs. Predicted transmitting abilities (PTAs) for five milk production traits (MY, FY, PY, FP, and PP) were used in association tests (Xu et al. 2014a). PTAs are estimates of an animal’s additive genetic merit for a trait. Target regions were detected using line regression after FDR correction to estimate the significant associated SNPs as described previously (Xu et al. 2014b).

LD, Recombination rate, and Haplotype Network Analysis

Genotypes for the SNPs in each gene were extracted as unphased information. LD parameters (D′ and r²) and LD blocks were computed using the Haploview 4.2 program (Barrett et al. 2005). Recombination rates were estimated using PHASE 2.1 under the approximate conditionals model for varying recombination rate (Li and Stephens 2003; Crawford et al. 2004). To explore the diversity of haplotypes and evolutionary relationships cross populations, haplotypes and their frequencies were estimated separately for each breed using PHASE 2.1 (Stephens et al. 2001; Crawford et al. 2004). To obtain reliable results, we employs an iterative scheme to perform inference with 10,000 iterations and 10,000 burn-ins, also we increased the number of iterations of the final run of the algorithm using option -X 100, for detailed see http://stephenslab.uchicago.edu/instruct2.1.pdf, last accessed December 2, 2014. We constructed haplotype networks for functional genes such as LAP3, SAR1B, LCORL, KIT, MC1R, and MIF1. For LAP3, SAR1B, and MC1R, we extended to include the full LD block length. Phylogenetic relationships among the identified haplotypes were inferred through a median-joining network analysis by Network 4.6.12 (http://www.fluxus-engineering.com/, last accessed December 2, 2014).

Availability of Data

SNP genotype data and population genetics and evolutionary analysis results are available upon request for research purposes.

Supplementary Material

Supplementary figures S1–S8 and tables S1–S4 are available at Molecular Biology and Evolution online (http://www.mbe.oxfordjournals.org/).