G3: Genes|Genomes|Genetics Early Online, published on July 16, 2014 as doi:10.1534/g3.114.011023
1
1
Fast and cost-effective genetic mapping in apple using next-generation sequencing
2
Kyle M. Gardner*, Patrick Brown§, Thomas F Cooke†, Scott Cann*, Fabrizio Costa‡, Carlos
3
Bustamante†, Riccardo Velasco‡, Michela Troggio‡ and Sean Myles*
4
5
* Department of Plant and Animal Sciences, Faculty of Agriculture, Dalhousie University, Nova
6
Scotia, Canada
7
§
Department of Crop Sciences, University of Illinois, Urbana, Illinois, USA
8
†
Department of Genetics, Stanford School of Medicine, Stanford University, Stanford,
9
California, USA
10
‡
11
Italy
Genetics and Molecular Biology Department, IASMA Research Center, San Michele all'Adige,
12
13
Data available via the Dryad Digital Repository: http://doi.org/10.5061/dryad.55t54
14
15
16
Running title: Mapping apple skin color using GBS
17
Keywords: Next-generation DNA sequencing, Genotyping-By-Sequencing, apple, Malus, QTL,
18
SNP
19
20
21
*
22
Sean Myles
23
Department of Plant and Animal Sciences
24
Faculty of Agriculture
25
Dalhousie University
26
Kentville, Nova Scotia
27
B2N 5E3
28
Canada
29
Phone: +1-902-365-8460
30
Fax: +1-902-679-2311
31
Email: [email protected]
Corresponding author:
1
© The Author(s) 2013. Published by the Genetics Society of America.
2
1
ABSTRACT
2
Next-generation DNA sequencing (NGS) produces vast amounts of DNA sequence data,
3
but it is not specifically designed to generate data suitable for genetic mapping. Recently
4
developed DNA library preparation methods for NGS have helped solve this problem,
5
however, by combining the use of reduced representation libraries with DNA sample
6
barcoding to generate genome-wide genotype data from a common set of genetic markers
7
across a large number of samples. Here we use such a method, called genotyping-by-
8
sequencing (GBS), to produce a data set for genetic mapping in an F1 population of apples
9
(Malus x domestica) segregating for skin color. We show that GBS produces a relatively
10
large, but extremely sparse, genotype matrix: over 270,000 SNPs were discovered, but
11
most SNPs have too much missing data across samples to be useful for genetic mapping.
12
After filtering for genotype quality and missing data, only 6% of the 85 million DNA
13
sequence reads contributed to useful genotype calls. Despite this limitation, using existing
14
software and a set of simple heuristics, we generated a final genotype matrix containing
15
3967 SNPs from 89 DNA samples from a single lane of Illumina HiSeq and used it to
16
create a saturated genetic linkage map and to identify a known QTL underlying apple skin
17
color. We therefore demonstrate that GBS is a cost effective method for generating
18
genome-wide SNP data suitable for genetic mapping in a highly diverse and heterozygous
19
agricultural species. We anticipate future improvements to the GBS analysis pipeline
20
presented here that will enhance the utility of next-generation DNA sequence data for the
21
purposes of genetic mapping across diverse species.
22
23
24
25
26
27
28
29
30
31
2
3
1
INTRODUCTION
2
The introduction of new high-throughput DNA sequencing technologies has dramatically
3
reduced sequencing costs and increased the pace of genomics research. One of the primary goals
4
of genomics research is to establish relationships between genotypes and phenotypes. In
5
agriculture, such genotype-phenotype associations form the basis of genomics-assisted breeding
6
programs that aim to accelerate the breeding of improved varieties. Long-lived woody perennials
7
are expensive to breed using traditional methods and therefore stand to benefit more from the
8
genomics revolution than most other agricultural species. Offspring from breeding programs can
9
be genetically screened and discarded at the seedling stage without incurring the enormous
10
expense of growing them for years to fruit-bearing maturity for evaluation (BUS et al. 2000;
11
KELLERHALS et al. 2000; DIRLEWANGER et al. 2004; DI GASPERO AND CATTONARO 2010).
12
Most forms of genetic mapping require the collection of genome-wide polymorphism data,
13
where genotypes from a set of common loci are obtained from a set of samples. While next-
14
generation DNA sequencing technologies produce vast amounts of DNA sequence data, they
15
were not designed to generate these types of genotype data. Recently, however, next generation
16
sequencing technologies have been coupled with reduced representation libraries and DNA
17
barcoding to simultaneously identify and genotype a common set of polymorphic loci across a
18
set of samples in a single experiment. The two most common of these methods include
19
Genotyping-By-Sequencing (GBS; ELSHIRE et al. 2011) and RADseq (BAIRD et al. 2008) and
20
the present study focuses on GBS. In addition to the low per sample cost, there are several
21
benefits to using sequence-based genotyping methods over microarray-based technologies
22
(MYLES 2013). For example, polymorphism discovery and genotyping is completed in a single
23
step, which not only saves time, but also reduces the ascertainment bias inherent in the process of
24
developing genotyping microarrays. Moreover, as reference genomes, alignment methods and
25
genotype calling algorithms improve, raw sequence data collected today will become more
26
valuable in the future as improved methods will enable more information to be extracted from
27
the original raw files.
28
Despite difficulties in experimental design, due to self incompatibility and high heterozygosity,
29
there are a wide variety of apple genetic maps constructed from bi-parental crosses. Most of
30
these linkage maps have been built with low throughput genetic markers such as microsatellites
31
(CELTON et al. 2009; FERNÁNDEZ-FERNÁNDEZ et al. 2012) and AFLPs (LIEBHARD et al. 2003;
3
4
1
KENIS AND KEULEMANS 2005) resulting relatively low marker density across assembled linkage
2
groups. Recently there has been a shift towards the high throughput identification of single
3
nucleotide polymorphisms (SNPs) in apple spurred on by decreasing DNA sequencing costs and
4
the availability of an apple (Malus x domestica, cultivar ‘Golden Delicious’) reference genome
5
(VELASCO et al. 2010). Chagne et al. (2012a) detail the creation of a SNP genotyping microarray
6
that assays 8,000 SNPs discovered from low coverage sequencing of 27 cultivars. To date, the
7
apple 8K SNP array has been used to create saturated linkage maps in bi-parental cross
8
populations (ANTANAVICIUTE et al. 2012) and to perform genomic selection (KUMAR et al. 2012)
9
and genome-wide association (KUMAR et al. 2013) in diverse breeding material. While SNP
10
arrays are widely used, the high levels of polymorphism in many agricultural species like apples
11
often results in unreliable or useless genotype calls because of highly variable probe-sequence
12
hybridization (e.g. MILLER et al. 2013). In addition, the ascertainment bias inherent in the design
13
of SNP genotyping microarrays results in only a small fraction of the queried loci being
14
polymorphic in any given bi-parental cross (MICHELETTI et al. 2011). For example, only about a
15
third of the SNPs on the apple 8K SNP array were observed to be polymorphic in a ‘Royal
16
Gala’בGranny Smith’ segregating population (CHAGNE et al. 2012a).
17
It is evident that GBS offers several advantages over competing technologies and is quickly
18
becoming the genotyping method of choice in many agricultural systems (POLAND AND RIFE
19
2012; MYLES 2013). For example, GBS has been recently used for a variety of applications
20
including saturating an existing genetic map in rice (SPINDEL et al. 2013); creating high density
21
genetic maps in wheat and barley (POLAND et al. 2012b); performing genomic selection in wheat
22
(POLAND et al. 2012a); ordering of a draft genome sequence in barley (CONSORTIUM 2012;
23
MASCHER et al. 2013a); and characterizing germplasm diversity in maize and switchgrass (LU et
24
al. 2013; ROMAY et al. 2013). Almost all GBS studies to date have focussed on inbred lines,
25
because genotype calling in highly heterozygous crops using next-generation DNA sequence
26
data requires more data and is far more complicated. The present study addresses this issue by
27
presenting a pipeline for GBS SNP calling in apples and follows recently published work on
28
GBS workflows developed for other heterozygous crops like grape (BARBA et al. 2013) and
29
raspberry (WARD et al. 2013). Using a single lane of Illumina HiSeq data, we identify a robust
30
set of SNPs and use them to generate a saturated genetic linkage map of the apple genome and
31
map a major QTL for apple skin colour in an F1 population.
4
5
1
MATERIALS AND METHODS
2
3
Population description and phenotyping
4
The ‘Golden Delicious x Scarlet Spur’ population investigated here is planted at the experimental
5
orchard of the Foundation Edmund Mach (FEM) in San Michele all’Adige, Italy. Each individual
6
progeny is represented by a single tree, grafted on M9 rootstock and planted in 2003. The
7
population has been grown and maintained following standard agronomical practice for fruit
8
thinning, canopy pruning, chemical fertilization and disease control. Due to the large variation in
9
ripening time among progeny, phenotyping was repeated three times during the harvesting
10
season. Skin color was scored as a binary trait: trees had apples with either yellow/green skin
11
(like Golden Delicious) or apples with red skin (like Scarlet Spur).
12
13
GBS library construction
14
GBS libraries were constructed using the two-enzyme modification of the original GBS protocol
15
(ELSHIRE et al. 2011; POLAND et al. 2012). DNA was extracted using commercial extraction kits.
16
Restriction/ligation reactions were performed in 96-well plates using 500 ng of DNA from each
17
individual, digestion with HindIII-HF and MspI (New England Biolabs, Ipswich, MA), and
18
0.1M and 10M of A1 and A2 adapters per well, respectively. Libraries were pooled, size-
19
selected on a 1% agarose gel, column-cleaned using a PCR purification kit (Qiagen, Valencia,
20
CA), and amplified for 12 cycles using Phusion DNA polymerase (NEB). Average fragment size
21
was estimated on a Bioanalyzer 2100 (Agilent, Santa Clara, CA) using a DNA1000 chip
22
following a second column-cleaning, and library quantification was performed using PicoGreen
23
(Invitrogen, Carlsbad, CA). Pooled libraries were adjusted to 10 nmol and sequenced with 100
24
bp, single-end reads on the HiSeq2000 (Illumina, San Diego, CA).
25
26
SNP calling
27
We created a custom bioinformatics pipeline, employing custom python scripts and existing
28
software, to process raw GBS sequence data from a single lane of an Illumina HiSeq sequencer
29
into SNP genotype tables (Figure S1). DNA barcode deconvolution and basic sequence quality
30
filtering was carried out using a custom python program (barcode_splitter.py). This program
31
splits the raw Illumina fastq file into 96 separate fastq files based on the barcode sequences
5
6
1
associated with each sample while filtering out reads containing any ambiguous bases in the
2
barcodes or restriction site remnants immediately following the barcode sequence. All sequences
3
successfully passing these basic filters were then scanned for the presence of an additional
4
restriction site remnant and, if present, the read was trimmed accordingly. Reads were also
5
trimmed if the sequence contained the common (A2) GBS adapter, indicating that the genomic
6
fragment sequenced was less than 90-100 bp in size.
7
The separate fastq files for each DNA sample were then aligned to version 1.0 of the apple
8
reference genome (VELASCO et al. 2010) using bwa (LI AND DURBIN 2009), allowing a maximum
9
of 4% sequence mismatch. Alignments were converted to the SAM format, then merged and
10
sorted into one master binary alignment file (BAM format) with SAMtools 0.1.18. (LI et al.
11
2009). SNP calling was carried out using the genome analysis toolkit (MCKENNA et al. 2010) on
12
the BAM file using a minimal set of filters that required a called SNP to have a locus quality
13
score of at least 30 given a prior probability of heterozygosity of 0.01.
14
15
SNP filtering and segregation analysis
16
SNP calls were filtered for quality by restricting the marker set to biallelic SNPs, requiring
17
genotype calls at each SNP to have a depth of coverage of six reads in each sample,
18
implementing a minor allele frequency threshold of ≥0.2, and limiting missing genotype data to a
19
maximum of 20% per SNP. All filtering was carried out using vcftools 0.1.10 (DANECEK et al.
20
2011) and the final SNP genotype tables were output into PLINK format (PURCELL et al. 2007).
21
The final SNP table contained 3967 SNPs and all of these SNPs were used to map apple color
22
using a simple chi-squared test (see below). The construction of the genetic map, however,
23
required further filtering based on the segregation pattern of the genotypes in parents and
24
offspring.
25
Linkage mapping in apple F1 populations is referred to as a double pseudo-testcross and only
26
three genotype combinations in the parental lines are informative for the construction of a
27
genetic map: when one parent is heterozygous and the other is homozygous (i.e. either AA x AB
28
or AB x AA) and when both parents are heterozygous (i.e. AB x AB). Only SNPs following
29
these segregation patterns in the parents were retained. Each SNP was subsequently tested for
30
the expected segregation ratios in the F1 progeny: heterozygous in Golden Delicious (1:1),
31
heterozygous in Scarlet Spur (1:1), and heterozygous in both Golden Delicious and Scarlet Spur
6
7
1
(1:2:1). SNPs deviating from these ratios according to a chi-squared test (P < 0.01) were not
2
included in map construction. Finally, progeny genotypes inconsistent with Mendelian
3
inheritance from the parental genotypes were set to missing. After implementing these filters,
4
2436 SNPs remained for linkage map construction.
5
6
Linkage map construction
7
A single composite linkage map was constructed using JoinMap 4.0 (STAM 1993) by combining
8
the backcross type (Aa x aa) markers segregating within each parental background and intercross
9
type (Aa x Aa) markers segregating within both parental backgrounds. Markers suspected of
10
being incorrectly phased by JoinMap, due to high pairwise linkage LOD and spurious
11
recombination fractions above 0.5, had their allele codes switched manually, but were dropped
12
from further analysis if phasing problems persisted. To increase mapping efficiency, pairs or
13
groups of loci with identical genotypes (i.e. complete linkage) were identified and a single
14
marker was chosen to represent the group.
15
Linkage groups were constructed, and ordered, with a linkage LOD of at least 6.0, a minimum
16
recombination fraction of 0.35, and a jump threshold of 5. SNPs exhibiting “suspect linkage” to
17
several loci within their assigned group (as determined by JoinMap), poorly fitting markers
18
within an ordered group, and markers that greatly inflated the linkage group size, were dropped
19
from the final mapping. All map distances were calculated using the Kosambi mapping function.
20
After applying these filters, 1994 SNPs remained within the genetic linkage map.
21
22
Mapping skin color
23
For association analysis based on the physical coordinates obtained from the reference genome,
24
the full set of 3967 SNPs were tested for association with skin color measurements using the
25
case/control single marker analysis in PLINK 1.04 (PURCELL et al. 2007) which employs a chi-
26
squared test of independence of allele frequencies between cases and controls with one degree of
27
freedom at each marker independently. For linkage map based QTL analysis, R/qtl 1.26-14
28
(BROMAN et al. 2003) was used with the single binary QTL interval mapping model, scanning
29
the linkage map at 1 cM steps for the presence of a significant QTL. A significance threshold of
30
3.05 was determined by permutation tests on 1000 randomizations of the trait data (CHURCHILL
31
AND DOERGE
1994). The PLINK genotype files, JoinMap input files and phenotype data are
7
8
1
available from the Dryad Digital Repository: http://doi.org/10.5061/dryad.55t54.
2
3
RESULTS
4
Alignment of GBS reads to the reference genome
5
GBS of a single plate of 96 DNA samples yielded 85,129,960 100bp reads using Illumina’s
6
HiSeq 2000 sequencing technology. The high proportion of reads beginning with a barcode
7
sequence (98.2%) and containing a restriction site remnant (99%) indicated that the library
8
preparation was effective and the data were of high quality. In addition, there were very low
9
occurrences of chimeric reads (~1.0%) and of reads containing downstream adapter sequences
10
(~1.2%).
11
After aligning each of the 96 samples’ reads to the ‘Golden Delicious’ v1.0 reference genome,
12
seven of the samples were found to have a relatively low numbers of reads (<150 000) that
13
uniquely aligned to the genome and these samples were subsequently dropped from further
14
analysis. In the remaining 89 samples (87 F1 progeny and 2 parents) there was an average of
15
973,896 (sd = 609,869) reads per sample and an average of 628,085 reads (sd = 393,153)
16
uniquely aligned to the reference genome (Figure 1). Despite the wide range of read counts
17
across samples, there was relative uniformity in the proportion of reads successfully mapped
18
across samples with an average of 63.5% (sd = 2.5%).
19
20
SNP calling
21
Considering only the successfully mapped reads from 89 samples, SNPs were discovered and
22
genotypes were called by analyzing the single master alignment file with GATK (MCKENNA et
23
al. 2010). After employing a minimal set of initial quality filters (see Methods), 273,835 SNPs
24
were identified. However, the resulting genotype matrix was extremely sparse: more than 75% of
25
the 273,835 SNPs contained >50% missing genotypes (Figure 2a). Restricting the analysis to
26
SNPs with <20% missing genotype data drastically reduced the number of SNPs to 30,393. It is
27
likely that many false negative genotype calls still exist in the resulting genotype table since
28
confidently calling heterozygotes in a highly heterozygous diploid species like apple requires a
29
relatively high depth of sequence coverage compared to genotype calling in inbreds. For the set
30
of 30,393 SNPs, the number of genotype calls at various sequence depth thresholds is shown in
31
Figure 2b. Despite the observation that many genotype calls are supported by >20 reads, the
8
9
1
number of SNPs for downstream analysis drops rapidly as the minimum depth of coverage
2
threshold is increased from 1 to 10 while implementing a missing data threshold of 20% (Figure
3
2c). We further reduced the number of markers to 3967 SNPs by applying a minimum depth of
4
coverage threshold of six sequence reads for a genotype call (i.e. any genotype with fewer than
5
six supporting sequence reads was set to missing) and then re-filtering for SNPs with <20%
6
missing genotypes. The choice of six reads per genotype as a threshold was arbitrary: it was
7
chosen as a tradeoff between increased confidence in genotype calls and the number of SNPs
8
retained for mapping. This set of 3967 SNPs was used to map skin color using a chi-squared test.
9
To create the genetic map, a final set of 2436 SNPs were retained after filtering for Mendelian
10
inconsistencies and segregation distortion (see Methods).
11
12
Genetic map construction
13
The final set of 2436 SNPs was separated into the three segregation types: 884 heterozygous in
14
Golden Delicious, 1044 heterozygous in Scarlet Spur, and 508 heterozygous in both parental
15
varieties. Once this set of SNPs was imported into JoinMap 4.0, 442 segregated identically to at
16
least one other SNP (i.e. no recombinants observed) and were dropped from further mapping.
17
The remaining 1994 SNPs were successfully grouped into 17 linkage groups (Figure S2), as
18
expected for the apple genome, at a conservative LOD threshold of six. Only 35 loci were found
19
to be unlinked to any group, suggesting a high degree of saturation of the assembled mapping
20
groups. Once ordered, the linkage groups spanned 1272 cM, with individual linkage groups
21
ranging from 56 cM to 96 cM. Across linkage groups, the average marker density was high with
22
a marker found every 0.68 cM (±0.13 SD).
23
24
QTL mapping for apple skin color
25
Single marker analysis using the final set of 3967 SNPs identified 15 SNPs significantly
26
associated with skin color after applying a conservative Bonferroni correction for multiple
27
testing (P < 1.26x10-5; Figure 3). These SNPs cluster within a single interval on chromosome 9
28
(bounded by coordinates 29,305,493 and 33,701,563). There was also a cluster of SNPs on the
29
distal portion of chromosome 13 that showed a suggestive association with skin color, but all fell
30
short of the corrected significance threshold (Figure 3).
31
Interval mapping using 1994 SNPs detected one highly significant QTL for skin color on the
9
10
1
distal portion of chromosome nine, at position 89.5 cM, near SNP 9_30303238 (Figure 4). The
2
“2 LOD” confidence interval for this QTL extended to a 14 cM region around the point estimate
3
of the QTL position.
4
5
DISCUSSION
6
A major consideration when conducting GBS experiments is the choice of restriction enzyme
7
used to generate the reduced representation library (RRL). This choice determines the tradeoff
8
between the number of fragments in the library and the sequencing depth of the fragments. There
9
are several GBS library preparation methods currently in use, including the original single
10
enzyme ApeKI protocol described in Elshire et al. (2011); a double digest employing enzymes
11
with differing restriction site lengths (Pst I/Msp I; POLAND et al. 2012b); and a multi-step
12
procedure that combines the double restriction enzyme digest with selective PCR amplification
13
(SONAH et al. 2013). While there was ultimately no explicit test of library fragment composition
14
in the present study, using the HindIII-MspI enzyme combination followed by size selection we
15
observed a very small proportion of sequence reads that required adapter trimming (1.2%),
16
suggesting that most fragments were larger than 100bp and were thus suitable for GBS.
17
However, the number of DNA sequence reads across individual samples varied by more than an
18
order of magnitude (0.15 million to 3 million; Figure 1), and we ultimately excluded seven
19
samples’ sequence data from analyses due to low read counts. These observations highlight the
20
sensitivity of GBS to DNA sample uniformity. Recent GBS library preparation methods have
21
been shown to improve the uniformity in read counts across samples (SONAH et al. 2013) and
22
further improvements are expected.
23
24
Due to the apple’s high heterozygosity and ancestral polyploidy, there are major challenges in the
25
assembly of its genome. It is estimated that the assembled portion of the current apple reference
26
assembly represents only 71% of its ~750Mb genome (VELASCO et al. 2010). Despite this
27
constraint, we restricted our analysis to sequences aligning to this assembly in order to enable a
28
comparison of our inferred linkage groups to the physical map. Approximately 40% of the
29
sequence reads did not map to the assembled genome and were thus excluded from further
30
analysis (Figure 1). These excluded reads are a combination of DNA sequences from the ~29%
31
of the genome that is not represented in the current assembly, DNA sequences that mapped to
10
11
1
multiple locations, and sequences that map to a unique position but with too many mismatches
2
(>4 mismatches per 100bp). As improvements are made to the genome assembly and to DNA
3
sequence length and quality, we anticipate significant improvements in the mapping step of the
4
GBS analysis pipeline. It is worth noting that a potential alternative is to avoid the use of a
5
reference genome altogether and to use a SNP calling pipeline that does not rely on a reference
6
genome (LU et al. 2013).
7
8
There are an enormous number of possible paths one can take from a DNA sequence file to a
9
SNP genotype table and it is known that alignment and genotype calling parameters have a
10
strong effect on the resulting quantity and quality of genome-wide SNP data (MASCHER et al.
11
2013b). Although established software does exist for SNP genotype calling from GBS data
12
(BRADBURY et al. 2007), our goal here was to demonstrate that, with simple heuristics applied
13
together with standard software packages, one can generate SNPs of sufficient quality and
14
quantity to be of utility for genetic mapping. Regardless of what tools are used, it is evident that
15
GBS generates a sparse genotype matrix due to uneven sequence coverage across samples and
16
sites: a large number of SNPs are discovered, but genotypes for these SNPs are most often
17
generated from only a small proportion of the DNA samples. For example, over 250,000 SNPs
18
were discovered in the present study, but when SNPs with >20% missing data are excluded, the
19
number of SNPs remaining for analysis is reduced to ~30,000 (Figure 2A). Moreover, the
20
number of SNPs decreases as the depth of coverage filter is increased (Figure 2B,C). We chose
21
arbitrarily to include only genotypes with ≥6 supporting reads as a trade-off between genotype
22
quality and quantity. Using these thresholds, we obtain a set of 3967 SNPs derived from only
23
6% of the sequence reads. Thus, in the end, 94% of the sequence we collected in this experiment
24
was ignored. Most GBS studies to date have focused on genotyping inbred lines, which requires
25
far less data and is statistically far simpler than genotype calling in highly heterozygous species
26
like apple or in polyploids, which are common among species of agricultural importance. GBS
27
SNP calling pipelines designed specifically for highly heterozygous and polyploid species that
28
take haplotype phasing and imputation into account promise to significantly enhance the utility
29
of GBS (e.g. LU et al. 2013; WARD et al. 2013).
30
31
Despite the high proportion of sequence reads discarded due to filtering, there remained a
11
12
1
sufficient number of markers to perform genetic mapping of apple skin color with a modest
2
sample size. By mapping skin color, we have intentionally focused on a trait with a simple
3
genetic architecture, which deviates from most QTL studies which focus on more complex traits.
4
Skin color is known to be controlled almost entirely by a single locus of large effect (i.e.
5
effectively Mendelian) and we leverage this knowledge here to verify the utility and power of
6
GBS to identify this known locus. With 3967 SNPs genotyped in 89 samples, a simple chi-
7
squared test for association revealed a single significant peak from position 29.3Mb to 33.7Mb
8
on chromosome 9 (Figure 3). This peak overlaps with the R2R3 MYB transcription factor gene
9
at position 32.8Mb on chromosome 9 known to regulate apple skin color (TAKOS et al. 2006;
10
ESPLEY et al. 2007; LIN-WANG et al. 2011; ZHU et al. 2011; CHAGNE et al. 2013). In addition, a
11
set of 1994 SNPs were used to produce a well-saturated linkage map spanning 1272 cM (Figure
12
S2). This map size is consistent with the sizes of apple genetic maps from previous studies, for
13
example 1538 cM (Khan et al. 2012); 1005 cM (CHAGNE et al. 2012b); 1143 cM (Han et al.
14
2011). On average, the genetic map has a marker every 0.68cM, which is in line with the marker
15
density achieved using the apple 8K SNP array (0.5cM/marker in ANTANAVICIUTE et al. 2012;
16
0.88cM/marker in CHAGNE et al. 2012a). Using this genetic map, interval mapping revealed
17
one highly significant QTL for skin color centered at position 89.5 cM (Figure 4; Figure S3), in
18
agreement with the results from the chi-squared test for association. Thus, with a set of simple
19
heuristics for calling genotypes from GBS data, a saturated genetic map can be generated and
20
QTL mapping can robustly identify genotype-phenotype associations.
21
22
In the present study, the parents of the F1 population were included only once and had average
23
sequence coverage (Figure 1). However, since inclusion of a SNP in the genetic map relied on
24
accurate genotype calls from both parents, it may be advisable to sequence the parents of
25
mapping populations to a higher depth, i.e. include them multiple times in the plate of samples.
26
Moreover, the genetic map presented here was constructed using only SNPs that mapped to the
27
anchored portion of the reference genome in order to allow comparison between physical and
28
genetic map positions. By including the unanchored portion of the genome sequence during the
29
initial alignment stage, it is likely that additional SNPs could be identified and placed on the
30
genetic map.
31
12
13
1
For 364 SNPs (18.3% of all mapped SNPs), the linkage group assignments conflicted with the
2
predicted chromosomal locations according to the reference genome. Antanaviciute et al. (2012)
3
reported a similar proportion (13.7%) using genotype data from the apple 8K SNP array. The
4
most likely reasons for these conflicts between genetic and physical maps are the high frequency
5
of paralagous genomic regions in the apple genome and the incorrect anchoring of sequences
6
during the assembly of the reference genome. Such conflicts may obscure association signals
7
and complicate the interpretation of genetic mapping results. For example, although below the
8
conservative threshold for significance, we detected a signal of association on chromosome 13
9
when SNPs were positioned according to the physical map (Figure 3). However, this same block
10
of SNPs that physically map to chromosome 13 were subsequently found to genetically map
11
close to the QTL for skin color on linkage group 9 (Figure 4). This demonstrates that caution is
12
warranted when relying on the physical map coordinates of the current reference genome
13
sequence.
14
15
It is worth noting that our modest mapping population size of 87 F1 offspring likely often
16
prevented the ordering algorithm from finding the correct order for SNPs that were closer than
17
~2-5 cM. Over larger distances, however, the estimated map order of SNPs was generally in
18
agreement with physical coordinates from the reference genome (Fig. S2).
19
20
The present study demonstrates that a Mendelian trait (skin color) can be mapped in an apple F1
21
population using GBS data from a single lane of Illumina next-generation sequencing.
22
Considering the currents costs of acquiring SNP data from the apple 8K genotyping microarray,
23
we estimate that a similar quantity and quality of apple SNP data can be achieved using GBS
24
with a 10-100x decrease in cost. Because of the rapid improvements in DNA sequencing
25
technology, we anticipate that genotyping efforts will increasingly favor the use of GBS or
26
similar methods over the use of microarrays. The present study uses standard software tools and
27
simple heuristics to generate biological insights from GBS data. However, it is clear that
28
methodological developments required for analyzing GBS data lag far behind the technology
29
developed to generate GBS data. In order to maximize the utility of next-generation DNA
30
sequence in the future, there is a clear need for improved computational and statistical tools to
31
extract as much information as possible from the raw data, and to phase, impute, order and
13
14
1
genetically map large sets of genetic markers.
2
3
ACKNOWLEDGMENTS
4
This article was written, in part, thanks to funding from the Canada Research Chairs program
5
and the National Sciences and Engineering Research Council of Canada.
6
7
REFERENCES
8
9
Antanaviciute, L., F. Fernandez-Fernandez, J. Jansen, E. Banchi, K. Evans et al., 2012
10
Development of a dense SNP-based linkage map of an apple rootstock progeny
11
using the Malus Infinium whole genome genotyping array. BMC Genomics 13:
12
203.
13
Baird, N. A., P. D. Etter, T. S. Atwood, M. C. Currey, A. L. Shiver et al., 2008 Rapid SNP
14
Discovery and Genetic Mapping Using Sequenced RAD Markers. PLoS ONE 3:
15
e3376.
16
Barba, P., L. Cadle-Davidson, J. Harriman, J. Glaubitz, S. Brooks et al., 2013 Grapevine
17
powdery mildew resistance and susceptibility loci identified on a high-resolution
18
SNP map. Theoretical and Applied Genetics: 1-12.
19
Bradbury, P. J., Z. Zhang, D. E. Kroon, T. M. Casstevens, Y. Ramdoss et al., 2007
20
TASSEL: software for association mapping of complex traits in diverse samples.
21
Bioinformatics 23: 2633-2635.
22
23
24
Broman, K. W., H. Wu, Ś. Sen and G. A. Churchill, 2003 R/qtl: QTL mapping in
experimental crosses. Bioinformatics 19: 889-890.
Bus, V., C. Ranatunga, S. Gardiner, H. Basset, E. Rikkerink et al., 2000 Marker-
25
assisted selection for pest and disease resistance in the New Zealand apple
26
breeding program. Acta Horticulturae 538: 541-547.
27
Celton, J. M., D. S. Tustin, D. Chagné and S. E. Gardiner, 2009 Construction of a dense
28
genetic linkage map for apple rootstocks using SSRs developed from Malus
29
ESTs and Pyrus genomic sequences. Tree Genetics & Genomes 5: 93-107.
30
Chagne, D., R. N. Crowhurst, M. Troggio, M. W. Davey, B. Gilmore et al., 2012a
14
15
1
Genome-Wide SNP Detection, Validation, and Development of an 8K SNP Array
2
for Apple. PLoS ONE 7: e31745.
3
Chagne, D., C. Krieger, M. Rassam, M. Sullivan, J. Fraser et al., 2012b QTL and
4
candidate gene mapping for polyphenolic composition in apple fruit. BMC Plant
5
Biology 12: 12.
6
Chagne, D., K. Lin-Wang, R. V. Espley, R. K. Volz, N. M. How et al., 2013 An Ancient
7
Duplication of Apple MYB Transcription Factors Is Responsible for Novel Red
8
Fruit-Flesh Phenotypes. Plant Physiology 161: 225-239.
9
10
11
12
13
14
15
16
Churchill, G. A., and R. Doerge, 1994 Empirical Threshold Values for Quantitative Trait
Mapping. Genetics 138: 963-971.
Consortium, T. I. B. G. S., 2012 A physical, genetic and functional sequence assembly
of the barley genome. Nature 491: 711-716.
Danecek, P., A. Auton, G. Abecasis, C. A. Albers, E. Banks et al., 2011 The variant call
format and VCFtools. Bioinformatics 27: 2156-2158.
di Gaspero, G., and F. Cattonaro, 2010 Application of genomics to grapevine
improvement. Australian Journal of Grape and Wine Research 16: 122-130.
17
Dirlewanger, E., E. Graziano, T. Joobeur, F. Garriga-Calderé, P. Cosson et al., 2004
18
Comparative mapping and marker-assisted selection in Rosaceae fruit crops.
19
Proceedings of the National Academy of Sciences of the United States of
20
America 101: 9891-9896.
21
Elshire, R. J., J. C. Glaubitz, Q. Sun, J. A. Poland, K. Kawamoto et al., 2011 A Robust,
22
Simple Genotyping-by-Sequencing (GBS) Approach for High Diversity Species.
23
PLoS ONE 6: e19379.
24
Espley, R. V., R. P. Hellens, J. Putterill, D. E. Stevenson, S. Kutty-Amma et al., 2007
25
Red colouration in apple fruit is due to the activity of the MYB transcription factor,
26
MdMYB10. The Plant Journal 49: 414-427.
27
Fernández-Fernández, F., L. Antanaviciute, M. M. Dyk, K. R. Tobutt, K. M. Evans et al.,
28
2012 A genetic linkage map of an apple rootstock progeny anchored to the Malus
29
genome sequence. Tree Genetics & Genomes 8: 991-1002.
30
Han, Y., D. Zheng, S. Vimolmangkang, M. A. Khan, J. E. Beever et al., 2011 Integration
15
16
1
of physical and genetic maps in apple confirms whole-genome and segmental
2
duplications in the apple genome. Journal of Experimental Botany.
3
4
5
Kellerhals, M., L. Gianfranceschi, N. Seglias and C. Gessler, 2000 Marker-assisted
selection in apple breeding. Acta Horticulturae 521: 255-266.
Kenis, K., and J. Keulemans, 2005 Genetic linkage maps of two apple cultivars (Malus ×
6
domestica Borkh.) based on AFLP and microsatellite markers. Molecular
7
Breeding 15: 205-219.
8
Khan, M. A., Y. Han, Y. F. Zhao, M. Troggio and S. S. Korban, 2012 A Multi-Population
9
Consensus Genetic Map Reveals Inconsistent Marker Order among Maps Likely
10
Attributed to Structural Variations in the Apple Genome. PLoS ONE 7: e47864.
11
Kumar, S., D. Chagné, M. C. A. M. Bink, R. K. Volz, C. Whitworth et al., 2012 Genomic
12
Selection for Fruit Quality Traits in Apple (Malus×domestica Borkh.). PLoS ONE
13
7: e36674.
14
Kumar, S., D. Garrick, M. Bink, C. Whitworth, D. Chagne et al., 2013 Novel genomic
15
approaches unravel genetic architecture of complex traits in apple. BMC
16
Genomics 14: 393.
17
18
Li, H., and R. Durbin, 2009 Fast and accurate short read alignment with Burrows–
Wheeler transform. Bioinformatics 25: 1754-1760.
19
Li, H., B. Handsaker, A. Wysoker, T. Fennell, J. Ruan et al., 2009 The Sequence
20
Alignment/Map format and SAMtools. Bioinformatics 25: 2078-2079.
21
Liebhard, R., M. Kellerhals, W. Pfammatter, M. Jertmini and C. Gessler, 2003 Mapping
22
quantitative physiological traits in apple (Malus X domestica Borkh.). Plant
23
Molecular Biology 52: 511.
24
Lin-Wang, K. U. I., D. Micheletti, J. Palmer, R. Volz, L. Lozano et al., 2011 High
25
temperature reduces apple fruit colour via modulation of the anthocyanin
26
regulatory complex. Plant, Cell & Environment 34: 1176-1190.
27
Lu, F., A. E. Lipka, J. Glaubitz, R. Elshire, J. H. Cherney et al., 2013 Switchgrass
28
Genomic Diversity, Ploidy, and Evolution: Novel Insights from a Network-Based
29
SNP Discovery Protocol. PLoS Genet 9: e1003215.
30
Mascher, M., G. J. Muehlbauer, D. S. Rokhsar, J. Chapman, J. Schmutz et al., 2013a
16
17
1
Anchoring and ordering NGS contig assemblies by population sequencing
2
(POPSEQ). The Plant Journal 76: 718-727.
3
Mascher, M., S. Wu, P. S. Amand, N. Stein and J. Poland, 2013b Application of
4
Genotyping-by-Sequencing on Semiconductor Sequencing Platforms: A
5
Comparison of Genetic and Reference-Based Marker Ordering in Barley. PLoS
6
ONE 8: e76925.
7
McKenna, A., M. Hanna, E. Banks, A. Sivachenko, K. Cibulskis et al., 2010 The
8
Genome Analysis Toolkit: a MapReduce framework for analyzing next-generation
9
DNA sequencing data. Genome Research 20: 1297-1303.
10
Micheletti, D., M. Troggio, A. Zharkikh, F. Costa, M. Malnoy et al., 2011 Genetic
11
diversity of the genus Malus and implications for linkage mapping with SNPs.
12
Tree Genetics & Genomes 7: 857.
13
Miller, A., N. Matasci, M. K. Aradhya, B. H. Prins, G. Y. Zhong et al., 2013 Vitis
14
phylogenomics: hybridization intensities from a SNP array outperform genotype
15
calls. PLoS ONE submitted.
16
17
18
19
20
Myles, S., 2013 Improving fruit and wine: what does genomics have to offer? Trends in
Genetics 29: 190-196.
Poland, J., J. Endelman, J. Dawson, J. Rutkoski, S. Wu et al., 2012a Genomic Selection
in Wheat Using Genotyping-By-Sequencing. The Plant Genome in press.
Poland, J. A., P. J. Brown, M. E. Sorrells and J.-L. Jannink, 2012b Development of
21
High-Density Genetic Maps for Barley and Wheat Using a Novel Two-Enzyme
22
Genotyping-by-Sequencing Approach. PLoS ONE 7: e32253.
23
24
25
Poland, J. A., and T. W. Rife, 2012 Genotyping-by-Sequencing for Plant Breeding and
Genetics. Plant Gen. 5: 92-102.
Purcell, S., B. Neale, K. Todd-Brown, L. Thomas, M. A. R. Ferreira et al., 2007 PLINK:
26
A Tool Set for Whole-Genome Association and Population-Based Linkage
27
Analyses. Am J Hum Genet 81: 559-575.
28
Romay, M., M. Millard, J. Glaubitz, J. Peiffer, K. Swarts et al., 2013 Comprehensive
29
genotyping of the USA national maize inbred seed bank. Genome Biology 14:
30
R55.
17
18
1
Sonah, H., M. Bastien, E. Iquira, A. Tardivel, G. Légaré et al., 2013 An Improved
2
Genotyping by Sequencing (GBS) Approach Offering Increased Versatility and
3
Efficiency of SNP Discovery and Genotyping. PLoS ONE 8: e54603.
4
Spindel, J., M. Wright, C. Chen, J. Cobb, J. Gage et al., 2013 Bridging the genotyping
5
gap: using genotyping by sequencing (GBS) to add high-density SNP markers
6
and new value to traditional bi-parental mapping and breeding populations.
7
Theoretical and Applied Genetics 126: 2699-2716.
8
9
10
Stam, P., 1993 Construction of integrated genetic linkage maps by means of a new
computer package: JoinMap. The Plant Journal 5: 739-744.
Takos, A. M., F. W. Jaffé, S. R. Jacob, J. Bogs, S. P. Robinson et al., 2006 Light-
11
Induced Expression of a MYB Gene Regulates Anthocyanin Biosynthesis in Red
12
Apples. Plant Physiology 142: 1216-1232.
13
14
15
Velasco, R., A. Zharkikh, J. Affourtit, A. Dhingra, A. Cestaro et al., 2010 The genome of
the domesticated apple (Malus x domestica Borkh.). Nature Genetics 42: 833.
Ward, J., J. Bhangoo, F. Fernandez-Fernandez, P. Moore, J. Swanson et al., 2013
16
Saturated linkage map construction in Rubus idaeus using genotyping by
17
sequencing and genome-independent imputation. BMC Genomics 14: 2.
18
19
20
21
Zhu, Y., K. Evans and C. Peace, 2011 Utility testing of an apple skin color MdMYB1
marker in two progenies. Molecular Breeding 27: 525-532.
FIGURE LEGENDS:
22
23
Figure 1: Results of alignment of GBS reads to the apple reference genome. For each sample,
24
the number of reads mapped and unmapped to the reference genome is shown. The read counts
25
for the parents of the F1 mapping population, Golden Delicious and Scarlet Spur, are indicated.
26
27
Figure 2: SNP and genotype counts from GBS data. a) Cumulative count of SNPs identified
28
across varying missing data thresholds. Over 200,000 SNPs are called with a very liberal
29
missing data threshold of 90%, but only 30,393 SNPs remain if only SNPs with <20% missing
30
data are retained. b) The number of genotypes called at increasing levels of sequencing depth,
31
after retaining only SNPs with <20% missing data. c) The number of SNPs retained at increasing
18
19
1
minimum thresholds of sequence depth while retaining only SNPs with <20% missing data. Here
2
we choose a minimum depth of coverage of six reads. Thus only SNPs with at least six
3
supporting reads and <20% missing genotypes were retained, which results in a set of 3967
4
SNPs.
5
6
Figure 3: Manhattan plot of a single marker association analysis for apple skin color. Each of the
7
3967 SNPs is plotted according to its physical position from the ‘Golden Delicious’ reference
8
genome and the –log10P value of the single marker association test. The horizontal dotted line
9
represents the Bonferonni-corrected P value significance threshold. The vertical dotted line
10
represents the location of the MYB transcription factor gene known to be responsible for skin
11
color variation.
12
13
Figure 4: Result of QTL analysis across the linkage group corresponding to chromosome 9 of
14
the apple genome. The left panel indicates the genetic map positions in cM of each of the SNPs
15
or groups of SNPs. Each SNP’s ID indicates its physical position according to the reference
16
genome, i.e. the physical coordinate it was assigned through alignment and SNP calling (e.g.
17
SNP 9_ 449878 mapped to position 449878 on chromosome 9 of the ‘Golden Delicious’ v1.0
18
reference genome). Note that many SNPs genetically mapping to the linkage group
19
corresponding to chromosome 9 are assigned to other chromosomes according to the reference
20
genome (e.g. SNP 13_30028140). The right panel displays the LOD scores from a QTL analysis
21
for skin color for markers that segregate in the Scarlet Spur genetic background. The horizontal
22
dashed line represents the significance threshold determined by permutation. LOD scores across
23
all linkage groups are shown in Figure S3.
24
25
Figure S1: A custom GBS analysis pipeline. Each box represents a file type, indicated by its file
26
extension, within the pipeline. Starting with a single .fastq file from the Illumina HiSeq 2000
27
sequencing machine and ending with a single set of PLINK files, the software packages and
28
custom scripts used to move from one file type to another are indicated beside each arrow.
29
Details of each step of the pipeline are provided in the Materials and Methods.
30
31
Figure S2: A composite genetic linkage map of the Malus x domestica genome. The map was
19
20
1
constructed from 1994 SNPs discovered and genotyped using GBS, including SNPs from all
2
three pseudo-testcross segregation types, in a Golden Delicious x Scarlet Spur F1 population.
3
4
Figure S3: Genome-wide LOD scores for apple fruit skin color. Each SNP is represented by a
5
single vertical line at the bottom of the plot. Only a single peak associated with skin color was
6
identified on linkage group 9 and this peak overlaps with the known apple skin color locus. The
7
horizontal dashed line represents the significance threshold determined by permutation.
20
Samples
Golden Delicious
Scarlet Spur
0
0.5
Mapped reads
Unmapped reads
1
1.5
2
Number of reads (x106)
2.5
3
Number of SNPs (x103)
5
15
10
0
0
Numver of SNPs (x103)
100
50
150
200
0
0
C
20
40
60
80
% missing data threshold
20
B
Number of genotypes (x105)
1
2
3
4
A
0
5
15
10
Depth of coverage
≥ 20
1
2 3 4 5 6 7 8 9
Depth of coverage threshold
10
10
-log10(P)
5
0
1
2
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Chromosome
cM
0.0
0.4
6.2
11.1
12.2
15.6
17.0
18.7
24.9
27.6
30.0
31.2
32.2
39.3
42.0
47.3
48.4
50.9
61.1
62.2
63.5
64.8
68.2
SNP ID
9_449878
15_3053874
9_915256 9_1802938
9_1803011
9_3221624 9_3221666
9_3221587 9_2791412
9_2800476
9_4450836 9_4450788
9_4737127
17_5459801
9_7225289 9_7225255
9_6818612 9_6818654
9_6818629
9_7724677 9_7724711
9_8004830 9_8817984
9_8008858
9_11011468 9_11011477
9_11144484 9_10948214
9_11861461 9_11861487
9_14020431 9_11854830
9_11854872
9_12356682 9_12382858
9_11972111 9_13486530
9_14043993 9_11864742
9_11418261 9_13486520
9_11972601 9_11972177
13_30028140 13_30241091
13_29528827 13_30028151
13_29528742 15_32421237
13_29500157 13_29966516
13_29528761 13_29528773
13_30241120
13_30470668 13_30470742
9_17523469
9_17515708
13_28845546
78.1
9_32245617 9_32245563
9_32642558
89.5
9_30303212 9_30303228
9_30303238
94.0
9_31183359
97.7
9_33513085 9_33513052
9_33701563
0
5
LOD
10
15