Identification of genomic regions associated with phenotypic variation between dog breeds using selection mapping.

Identification of Genomic Regions Associated with
Phenotypic Variation between Dog Breeds using
Selection Mapping
Amaury Vaysse1., Abhirami Ratnakumar2., Thomas Derrien1, Erik Axelsson2, Gerli Rosengren Pielberg2,
Snaevar Sigurdsson3, Tove Fall4, Eija H. Seppa¨la¨5, Mark S. T. Hansen6, Cindy T. Lawley6, Elinor K.
Karlsson3,7, The LUPA Consortium, Danika Bannasch8, Carles Vila`9, Hannes Lohi5, Francis Galibert1,
Merete Fredholm10, Jens Ha¨ggstro¨m11, A˚ke Hedhammar11, Catherine Andre´1, Kerstin Lindblad-Toh2,3,
Christophe Hitte1, Matthew T. Webster2*
1 Institut de Ge´ne´tique et De´veloppement de Rennes, CNRS-UMR6061, Universite´ de Rennes 1, Rennes, France, 2 Science for Life Laboratory, Department of Medical
Biochemistry and Microbiology, Uppsala University, Uppsala, Sweden, 3 Broad Institute of Harvard and Massachusetts Institute of Technology, Cambridge, Massachusetts,
United States of America, 4 Department of Medical Epidemiology and Biostatistics, Karolinska Institute, Stockholm, Sweden, 5 Department of Veterinary Biosciences,
Research Programs Unit, Molecular Medicine, University of Helsinki and Folkha¨lsan Research Center, Helsinki, Finland, 6 Illumina, San Diego, California, United States of
America, 7 FAS Center for Systems Biology, Harvard University, Cambridge, Massachusetts, United States of America, 8 Department of Population Health and
Reproduction, School of Veterinary Medicine, University of California Davis, Davis, California, United States of America, 9 Department of Integrative Ecology, Don˜ana
Biological Station (CSIC), Seville, Spain, 10 Faculty of Life Sciences, Division of Genetics and Bioinformatics, Department of Basic Animal and Veterinary Sciences, University
of Copenhagen, Frederiksberg, Denmark, 11 Department of Clinical Sciences, Swedish University of Agricultural Sciences, Uppsala, Sweden
Abstract
The extraordinary phenotypic diversity of dog breeds has been sculpted by a unique population history accompanied by
selection for novel and desirable traits. Here we perform a comprehensive analysis using multiple test statistics to identify
regions under selection in 509 dogs from 46 diverse breeds using a newly developed high-density genotyping array
consisting of .170,000 evenly spaced SNPs. We first identify 44 genomic regions exhibiting extreme differentiation across
multiple breeds. Genetic variation in these regions correlates with variation in several phenotypic traits that vary between
breeds, and we identify novel associations with both morphological and behavioral traits. We next scan the genome for
signatures of selective sweeps in single breeds, characterized by long regions of reduced heterozygosity and fixation of
extended haplotypes. These scans identify hundreds of regions, including 22 blocks of homozygosity longer than one
megabase in certain breeds. Candidate selection loci are strongly enriched for developmental genes. We chose one highly
differentiated region, associated with body size and ear morphology, and characterized it using high-throughput
sequencing to provide a list of variants that may directly affect these traits. This study provides a catalogue of genomic
regions showing extreme reduction in genetic variation or population differentiation in dogs, including many linked to
phenotypic variation. The many blocks of reduced haplotype diversity observed across the genome in dog breeds are the
result of both selection and genetic drift, but extended blocks of homozygosity on a megabase scale appear to be best
explained by selection. Further elucidation of the variants under selection will help to uncover the genetic basis of complex
traits and disease.
Citation: Vaysse A, Ratnakumar A, Derrien T, Axelsson E, Rosengren Pielberg G, et al. (2011) Identification of Genomic Regions Associated with Phenotypic
Variation between Dog Breeds using Selection Mapping. PLoS Genet 7(10): e1002316. doi:10.1371/journal.pgen.1002316
Editor: Joshua M. Akey, University of Washington, United States of America
Received February 1, 2011; Accepted July 30, 2011; Published October 13, 2011
Copyright: � 2011 Vaysse et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits
unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: This research was mainly supported by The LUPA Consortium, which is a Collaborative Research Project funded by the European Commission under
the 7th Research Framework Programme (www.eurolupa.org). The consortium consists of research groups in more than 20 European institutes, with the goal of
unraveling the genetic basis of inherited disease in dogs with relevance to human health. The funders had no role in study design, data collection and analysis,
decision to publish, or preparation of the manuscript.
Competing Interests: The authors have declared that no competing interests exist.
* E-mail: matthew.webster@imbim.uu.se
. These authors contributed equally to this work.
Introduction
There are more than 400 breeds of domestic dog, which exhibit
characteristic variation in morphology, physiology and behavior.
This astonishing phenotypic diversity has been molded by two
main phases of evolution: 1) the initial domestication from wolves
more than 15,000 years ago, where dogs became adapted to life in
closer proximity to humans and 2) the formation of distinct breeds
in the last few hundred years, where humans chose small groups of
dogs from the gene pool and strongly selected for novel and
desirable traits [1,2]. A by-product of these processes has been that
many dog breeds suffer from a high incidence of inherited
disorders [3,4].
Its unique population history makes the dog an ideal model
organism for mapping the genetic basis of phenotypic traits due
to extensive linkage disequilibrium (LD) and a reduction in
PLoS Genetics | www.plosgenetics.org 1 October 2011 | Volume 7 | Issue 10 | e1002316

haplotype diversity due to genetic drift in isolated populations [3-
5]. Another major advantage of the canine model is that much of
the variation in morphological characteristics in dogs appears to
be governed by a relatively small number of genetic variants with
large effect [6]. This is likely because novel variants with large
effects are preserved by artificial selection. This is in strong
contrast to humans where morphological variation in traits such
as height appears to be controlled by hundreds of loci with small
effects, which have proven extremely difficult to catalogue [7].
Identifying the targets of artificial selection in dog breeds is
therefore an extremely promising approach for identifying
genetic variants involved in phenotypic variation, which could
greatly facilitate the identification of similar variants and novel
molecular pathways in humans.
Several loci have now been identified that control variation in
morphological traits between dog breeds. In some cases, variation
in a trait occurs within a breed, and long blocks of LD can be used
to identify the locus responsible using genome wide association
studies (GWAS). Using this approach loci involved in traits
including size (IGF1) [8], coat type (RSPO2, FGF5, KRT71) [9]
and coat color (MITF, CBD103) [10,11] were identified in single
breeds, and it was shown that variation in these loci is also
correlated with phenotypic variation between breeds. An alterna-
tive approach, when a particular trait is shared by several breeds,
is to perform across-breed GWAS. In general, levels of LD decay
much faster between breeds, and this reduces the power to detect
association [11]. However, selection acts to fix long haplotypes
bearing the causative variant, thus increasing levels of LD between
breeds in regions under selection. Jones et al. [12] used a sparse
marker set and across-breed GWAS to identify correlations with a
number of morphological traits, such as size, height, and shape of
ears, snout and limbs, which was further refined by Boyko et al. [6]
using 80 dog breeds and ,61,000 SNPs. Across-breed GWAS
have also been used to identify an FGF4 retrogene associated with
chondrodysplasic breeds [13] and the THBS2 locus associated
with brachycephalic breeds [14].
Genomic regions with a high degree of genetic differentiation
between breeds are also indicative of selection. A large proportion
of SNPs with high FST between dog breeds are found in loci
associated with phenotypic traits such as size, ear morphology and
coat color [6]. Akey et al. [15] scanned patterns of variation in 10
dog breeds and ,21,000 SNPs using a 1 Mb sliding windows to
identify larger regions with elevated FST in particular breeds. This
scan identified many regions likely to be under selection in one or
more of the breeds in their dataset. Notably, a highly differentiated
interval in Shar-Pei on chromosome 13 contains the HAS2 gene
and is likely associated with the wrinkled skin phenotype of this
breed [15,16].
Although a large number of loci under selection have now
been identified, the genetic basis of much of the phenotypic
variation in dog breeds and particularly behavioral traits remains
unexplained. One drawback of previous studies is the use of SNP
arrays with relatively low coverage of the genome. With the
development of a new high-density array it is now possible to
examine the dog genome at much higher resolution, allowing a
comprehensive characterization of regions under selection.
Genetic variants under selection in dogs can be loosely divided
into two categories: 1) those that control variation in common
traits such as size and ear carriage, which segregate across many
breeds [6,8] and 2) those that encode rare traits that present in
one or a few breeds, such as brachycephaly, chondrodysplasia
and skin wrinkling [13,14,16].
Here we implement a variety of approaches to identify both these
types of loci. In cases where a common trait has been identified, it is
possible to search for genotype-phenotype correlations. We attempt
to identify both behavioral and morphological traits that vary
between breeds using across-breed GWAS. We also use FST statistics
to identify additional SNPs that have high variability in frequency
between breeds. These methods identify known loci and indicate
new regions that may be involved in common trait variation.
The action of selection can potentially be identified by
examining patterns of variation in individual breeds in order to
detect the characteristic signature of selective sweeps. This
signature is characterized by the presence of long haplotypes, a
skew in allele frequency, reduced heterozygosity, and elevated
population differentiation. A large number of statistical
methods have been developed to detect sweeps based on these
different patterns [17-22]. The formation of dog breeds
occurred during an extremely brief evolutionary time, and
likely involved rapid fixation of haplotypes under strong
artificial selection. Under this scenario, simulations suggest
that statistics based on FST and differences in heterozygosity are
likely to be most powerful. [23]. Furthermore, dog breeds are
known to be characterized by extensive LD and limited
haplotype diversity, including long blocks of homozygosity,
which reflect the action of population bottlenecks and selective
breeding. This suggests that tests based on allele frequency
spectrum and haplotype length will be of limited applicability,
as many genomic regions are essentially devoid of genetic
variation. We therefore base our approach to identify selective
sweeps on pairwise comparisons of both FST and heterozygosity
between breeds.
The presence of long blocks of homozygosity in the dog genome
[1,11] is likely to reflect the action of both selection and genetic
drift. We therefore conduct extensive coalescent simulations in
order to distinguish between these processes. These simulations
incorporate a realistic model of dog population history under
neutrality to provide null distributions to compare with the real
data. We also conduct a comprehensive characterization of SNP
variation in a 3 Mb region encompassing several loci with extreme
population differentiation that are associated with at least two
morphological traits.
Author Summary
There are hundreds of dog breeds that exhibit massive
differences in appearance and behavior sculpted by tightly
controlled selective breeding. This large-scale natural
experiment has provided an ideal resource that geneticists
can use to search for genetic variants that control these
differences. With this goal, we developed a high-density
array that surveys variable sites at more than 170,000
positions in the dog genome and used it to analyze genetic
variation in 46 breeds. We identify 44 chromosomal regions
that are extremely variable between breeds and are likely to
control many of the traits that vary between them,
including curly tails and sociality. Many other regions also
bear the signature of strong artificial selection. We
characterize one such region, known to associate with
body size and ear type, in detail using ‘‘next-generation’’
sequencing technology to identify candidate mutations
that may control these traits. Our results suggest that
artificial selection has targeted genes involved in develop-
ment and metabolism and that it may have increased the
incidence of disease in dog breeds. Knowledge of these
regions will be of great importance for uncovering the
genetic basis of variation between dog breeds and for
finding mutations that cause disease.
Selection Mapping in Dogs
PLoS Genetics | www.plosgenetics.org 2 October 2011 | Volume 7 | Issue 10 | e1002316

Results
High-density canine array design and evaluation
Our first goal was to develop a high-density, high-accuracy
mapping array with uniform SNP coverage across the whole
genome. Since the SNP map from the canine genome project,
although containing .2.8 million SNPs at fairly even coverage,
still contained gaps, we first performed targeted resequencing
within 1,555 regions that lie within intervals .40 kb containing no
known SNPs in unique sequence. We performed Roche Nimble-
Gen array capture to enrich these regions followed by sequencing
using the Illumina Genome Analyzer on 4 pools containing
multiple samples of a single dog breed (Irish Wolfhounds, West
Highland White Terrier, Belgian Shepherds and Shar-Pei) and
one pool of wolf samples. In total, we discovered 4,353 additional
high-quality SNPs using this method. We selected SNPs from this
improved map to form the ‘‘CanineHD’’ array panel. We
generated an initial panel of 174,943 SNPs that were included
on the array of which 173,622 (99.2%) give reliable data. These
loci are distributed with a mean spacing of 13 kb and only 21 gaps
larger than 200 kb. Loci with unreliable SNP calls, potentially due
to copy number polymorphism, were not included in the analysis.
In total, 172,115 are validated for SNP genotyping and 1,547 are
used only for probe intensity analyses. This is a significant
improvement compared with the largest previously existing array,
which has 49,663 well performing SNPs, with a mean spacing of
47 kb and 1,688 gaps larger than 200 kb. Figure S1 shows the
distribution of SNPs in 100 kb windows across the genome. The
improvement in coverage is particularly striking on the X
chromosome, where .75% of 100 kb windows contain no SNPs
on the previous array, but ,5% of windows do not contain SNPs
on the CanineHD array.
Of all the SNPs on the array, 0.9% are novel SNPs discovered
by the targeted resequencing experiment. The remaining SNPs
have been previously described: 65.1% of them were present in a
comparison of the boxer reference genome with a previously
sequenced poodle, 21.7% were present in alignments of low
coverage sequencing reads from various dog breeds to the boxer
reference genome, 25.4% were present within the boxer reference
and 1.2% were present in alignments of wolf and/or coyote
sequencings with the reference boxer genome. There is therefore a
bias in the way that SNPs were ascertained: all of them were
identified in a comparison involving the boxer reference assembly.
However this has not had a great impact on the number of SNPs
polymorphic in different breeds (see below). The array was initially
evaluated using 450 samples from 26 breeds termed the
‘‘Gentrain’’ dataset. Within this dataset, average call rates were
99.8% and reproducibility and Mendelian consistency were both
.99.9%. A subset of 24 samples generated by whole genome
amplification (WGA) of 12 blood and 12 cheek swab samples
produced slightly lower call rates (blood-WGA 99.3%; buccal-
WGA 98.9%). Probe intensities from the array can also be used to
analyze copy number polymorphisms, although this is not
evaluated here.
Dataset construction
To perform a broader analysis of canine breed relationships and
selective sweeps, we constructed a larger dataset consisting of
unrelated samples from the Gentrain dataset, and unrelated
control dogs genotyped for disease gene mapping studies from
multiple breeds as part of the LUPA consortium. This dataset,
which we refer to here as the ‘‘full LUPA genotype dataset’’
consists of 509 dogs from 46 diverse breeds and 15 wolves,
genotyped on the CanineHD array. These include 156 dogs from
13 breeds derived from LUPA control dogs and 353 dogs from 33
breeds from the Gentrain dataset (See Table S1 for full details). A
subset of this dataset, referred to here as the ‘‘reduced LUPA
genotype dataset’’ is made up of all the samples in the 30 breeds
(plus wolf) with more than 10 samples in the full dataset (471
samples in total).
Table 1 shows patterns of polymorphism in the reduced LUPA
genotype dataset. In total, 157,393 SNPs on the array were
polymorphic (90% of SNPs on the array). A mean of 119,615
SNPs (69%) were polymorphic within a single dog breed. Hence
although there is a bias in the way that SNPs were ascertained,
there is a substantial amount of variation within all breeds
surveyed. On average 39 SNPs were polymorphic only in one
breed, although this figure shows large variation between breeds.
A subset of 1,471 SNPs were variable in wolves but not within any
dog breed. However, most of these SNPs were originally
discovered by comparisons of sequences from different dog breeds,
which suggests that they are also variable between (but not within)
dog breeds.
Evolutionary relationships between dog breeds
We used the CanineHD array to investigate breed relationships
by constructing a neighbor-joining tree [24] of raw genetic
distances in the full LUPA genotype dataset (Figure 1). Three main
features are obvious: 1) Dogs from the same breed almost
invariably cluster together. This reflects the notion that modern
breeds are essentially closed gene pools that originated via
population bottlenecks. 2) Little structure is obvious in the internal
branches that distinguish breeds. This is consistent with the
suggestion that all modern dog breeds arose from a common
population within a short period of time and that only a very small
proportion of genetic variation divides dog breeds into subgroups.
3) The internal branches leading to boxer and wolf are longer than
those leading to other breeds. The long boxer branch can be
explained by the fact that a large proportion of the SNPs were
assayed by comparing boxer with other breeds, which implies that
the dataset is enriched for SNPs that differ between boxer and
other breeds. The longer wolf branch probably reflects more
distant relatedness.
Some breeds show a tendency to group together in the tree,
such as breeds of retrievers, spaniels, setters, and terriers.
However, the length of the internal branches leading to these
clusters is only a small fraction of the average total length of
branches in these clusters, which indicates that genetic variation in
dogs is much more severely affected by breed creating bottlenecks
than it is by historical origins of various breeds, although detailed
analysis of these data has power to reveal their historical origins
[25]. The most obvious clustering of breeds is exhibited by two
wolf hybrids: Sarloos and Czechoslovakian wolf dog, which exhibit
a closer relationship to the wolf than other breeds as predicted by
their known origin [26]. The German shepherd also clusters with
this group, although this is likely to be a result of its close
relationship with the Czechoslovakian wolf dog, rather than with
wolf. The tree is consistent with previous studies and supports the
accuracy and reliability of the array. Although the long boxer
branch likely reflects SNP ascertainment bias on the array, the tree
reflects extensive polymorphism both within and between breeds.
This suggests the SNP ascertainment scheme is not problematic
and that the array is well suited for both within and across breed
gene mapping.
We performed coalescent simulations modeling the ascertain-
ment bias, sample size, and inferred recombination rate in the true
dataset (see Materials and Methods) in order to predict the
expected patterns of genetic diversity that we expect to observe
Selection Mapping in Dogs
PLoS Genetics | www.plosgenetics.org 3 October 2011 | Volume 7 | Issue 10 | e1002316

within and between breeds in the absence of selection. The
bottleneck population sizes at breed creation used in the
simulations are presented in Table S2. The decay of LD in the
simulated data closely matches the real decay in LD (Figure S2).
Across-breed GWAS: morphological traits
To identify genetic variation associated with common traits that
vary among breeds, we performed across-breed GWAS using the
full LUPA genotype dataset. A list of traits and their variation
between breeds is in Table S3. Each sample was given a value
corresponding to the standardized breed phenotype for the trait
under study. We performed quantitative association studies for size
and personality traits whereas other traits were binary coded. For
each GWAS, we assayed genome-wide significance by permuting
the phenotype of each breed, assigning each dog of the same breed
with identical phenotype values. The true significance of genotype-
phenotype correlation at each SNP was compared with the
maximum permuted value of all SNPs across the array in order to
estimate genome-wide significance (see Materials and Methods).
This permutation procedure corrects for the extreme population
substructure present in dog breeds.
Using this method we were able to replicate several known
associations. We first performed a GWAS comparing 4 breeds with
furnishings (a coat type with moustache and eyebrows [9])
compared to 42 without them. Genome-wide significant associa-
tions were observed at 3 SNPs distributed located between 10.42 -
11.68 Mb on chromosome 13. The most strongly associated SNP is
at 11,678,731 (Pgenome,0.001), 44 kb from the causative SNP
previously identified in RSPO2 [9]. We next scanned the genome
for associations with size, using weight in kilograms as a proxy (data
taken from [8]; see Table S3). The most strongly associated SNP
was located on chromosome 15 at 44,242,609 (Pgenome = 0.004),
which is within the IGF1 gene, previously implicated in size
variation [8]. Genome-wide significant associations (Pgenome,0.05)
were observed at 7 SNPs within an interval between 44.23 - 44.44
Mb. In addition, we observed an association within a previously
Table 1. Levels of genetic variation in breeds with 10 or more samples.
Breed Abbreviation No. Samples Seg. sites Private seg. sites
Belgian Tervuren BeT 12 115,154 0
Beagle Bgl 10 115,254 16
Bernese Mountain Dog BMD 12 106,152 15
Border Collie BoC 16 127,491 7
Border Terrier BoT 25 108,344 15
Brittany Spaniel BrS 12 130,115 11
Cocker Spaniel CoS 14 126,118 19
Dachshund Dac 12 131,372 5
Doberman Pinscher Dob 25 112,627 19
English Bulldog EBD 13 111,720 19
Elkhound Elk 12 127,066 82
English Setter ESt 12 121,196 24
Eurasian Eur 12 120,360 6
Finnish Spitz FSp 12 109,510 20
Gordon Setter GoS 25 134,615 12
Golden Retriever Gry 11 112,144 10
Greyhound GRe 14 128,907 45
German Shepherd GSh 12 108,614 11
Greenland Sledge Dog GSl 12 102,899 19
Irish Wolfhound IrW 11 92,718 61
Jack Russell Terrier JRT 12 137,837 12
Labrador Retriever LRe 14 129,951 23
Newfoundland NFd 25 127,503 13
Nova Scotia Duck Tolling Retriever NSD 23 118,387 36
Rottweiler Rtw 12 107,022 15
Schipperke Sci 25 126,530 21
Shar-Pei ShP 11 124,828 93
Standard Poodle StP 12 132,289 123
Yorkshire Terrier TYo 12 129,768 388
Weimaraner Wei 26 111,958 21
Wolf Wlf 15 118,256 1,471
Total - 471 157,393 -
doi:10.1371/journal.pgen.1002316.t001
Selection Mapping in Dogs
PLoS Genetics | www.plosgenetics.org 4 October 2011 | Volume 7 | Issue 10 | e1002316

defined region on chr10 (11,169,956 bp; Pgenome = 0.036). The SNP
at chr10:11,169,956 is about 500kb upstream of HMGA2, which
has been established to be associated with body size variation in
other species [27–29].
The frequency of the SNP (chr15:44,242,609) most strongly
associated to size shows a steady decline according to the size of
the breed. However, the differences in allele frequency at the SNP
chr10:11,169,956 are more striking, as one allele appears at very
low frequencies in all breeds apart from a number of very small
breeds (Yorkshire Terrier, Border Terrier, Jack Russell Terrier,
Schipperke), where it is at or close to fixation (Figure S3). Hence,
there appears to be relatively continuous variation in frequency in
a variant affecting IGF1 between breeds, whereas a variant
upstream of HMGA2 appears to have been fixed in a subset of
small breeds but shows little variation in allele frequencies in other
breeds.
Dog breeds show extreme variation in ear morphology ranging
from pricked ears to low hanging dropped ears. We performed a
GWAS using 12 breeds with pricked ears and 15 breeds with
dropped ears. Within an interval between 10.27 - 11.79 Mb, 23
SNPs had genome-wide significant associations (Pgenome,0.05;
Figure 2). The most strongly associated SNP was chr10:
11,072,007 (Pgenome,0.001), which lies between the HGMA2
and MSRB3 genes. This region has been associated with ear type
and body size in previous studies [6,12]. Using the CanineHD
array, we are able to type SNPs at a much higher density in the
associated region. There is also large variation between dog breeds
in degree of tail curl. We classified breeds in our dataset into 11
Figure 1. Neighbor-joining tree constructed from raw genetic distances representing relationships between samples. More than
170,000 SNPs were genotyped in 46 diverse dog breeds plus wolves using the CanineHD array. The boxer branches are longer, which likely represents
the influence of ascertainment bias, as the SNPs were discovered from sequence alignments involving the boxer reference sequence.
doi:10.1371/journal.pgen.1002316.g001
Selection Mapping in Dogs
PLoS Genetics | www.plosgenetics.org 5 October 2011 | Volume 7 | Issue 10 | e1002316