9 prospects for genomic selection in forage plant species

The second scheme would allow two rounds of selec-tion for key agronomic traits within a time period previously required for a single round, potentially leading to doubling of genetic g

Prospects for genomic selection in forage plant species

BE N J A M I N J HA Y E S1,2,3, NO E L O I CO G A N1,2, LU K E W PE M B L E T O N1,2,3, MI C H A E L E GO D D A R D1,2,4,

JU N P I N G WA N G2,5, GE R M A N C SP A N G E N B E R G1,2,3and JO H N W FO R S T E R1,2,3,6

1

Department of Primary Industries, Biosciences Research Division, AgriBio, the Centre for AgriBioscience, 5 Ring Road, Bundoora, Vic 3083, Australia;2Dairy Futures Cooperative Research Centre, Victorian AgriBiosciences Centre, La Trobe University Research and Development Park, Bundoora, Vic 3083, Australia;3La Trobe University, Bundoora, Vic 3086, Australia;4Faculty of Land and Environment, University of Melbourne, Parkville, Vic 3052, Australia;5Department of Primary Industries, Biosciences Research Division, Hamilton Centre, Hamilton, Vic 3300, Australia;6Corresponding author, E-mail: john.forster@dpi.vic.gov.au

With 3figures and 2 tables

Received May 31, 2012/Accepted December 14, 2012

Communicated by O A Rognli

Abstract

Genomic selection (GS) is a powerful method for exploitation of DNA

sequence polymorphisms in breeding improvement, through the

predic-tion of breeding values based on all markers distributed genome-wide.

Forage grasses and legumes provide important targets for GS

implemen-tation, as many key traits are dif ficult or expensive to assess, and are

measured late in the breeding cycle Generic attributes of forage breeding

programmes are described, along with status of genomic resources for a

representative species group (ryegrasses) Two schemes for implementing

GS in ryegrass breeding are described The first requires relatively little

modi fication of current schemes, but could lead to significant reductions

in operating cost The second scheme would allow two rounds of

selec-tion for key agronomic traits within a time period previously required for

a single round, potentially leading to doubling of genetic gain rate, but

requires a purpose-designed reference population In both schemes, the

limited extent of linkage disequilibrium (LD), which is the major

chal-lenge for GS implementation in ryegrass breeding, is addressed The

strategies also incorporate recent advances in DNA sequencing

technol-ogy to minimize costs.

Key words:single-nucleotide polymorphism— pasture — grass

— legume — sequencing — breeding programme

The international forage and turf genetics supply industry is a

significant component of the global crop seed business The

world markets for pasture species are dominated by the

consumption of grass seed in the USA, followed by northern

Europe, and temperate regions of Australasia, South America

and East Asia Perennial ryegrass (Lolium perenne L.), Italian

ryegrass (L multiflorum Lam.) and tall fescue (Festuca

arundin-acea Schreb syn L arundinaceum Schreb [Darbysh.]) are the

major species In Australia, the largest volumes of sales for

forage varieties are for perennial ryegrass (over 9000 tonnes/

year), as compared to tall fescue and white clover, both

<1000 tonnes/year

The historical rate of genetic improvement for key traits in

forage breeding (yield, quality and persistence) has been

moderate, at ca 7% per decade for perennial grasses (Wilkins

and Humphreys 2003, Gout and Jones 2006) These systems

have been based on the estimation of genetic merit from

recorded phenotypes of progeny and to a much lesser extent,

pedigree information (e.g Henderson 1984), followed by the

selection of superior individuals based on such estimated breeding values

In the genomics era, a third source of information (phenotype and pedigree being thefirst and second sources of information)

is available to improve the accuracy of estimated breeding values, in the form of DNA-based markers The most abundant class of markers is single-nucleotide polymorphisms (SNPs) In livestock and crop species, SNP arrays are now available These arrays can genotype a large number of markers in a single assay

at low cost For maize, for example, a 60 000 SNP array is available, and for wheat, a 9000 SNP array has been designed (Ganal et al 2011) Although the majority of SNPs in these assays are neutral with respect to trait variation, population-wide associations may be identified between the SNP and causal polymorphisms [quantitative trait loci (QTLs)] that affect the traits of interest These associations arise due to linkage disequi-librium (LD), which occurs when the SNP and QTL are located within a chromosome segment that traces back to a common ancestor, with limited intervening recombination (Mackay 2001) Marker-assisted selection (MAS) can be used to exploit the prop-erty of LD between markers and QTLs in breeding programmes The relative advantage of MAS over more conventional methods

of selection is proportional to the percentage of genetic variance accounted for by QTLs associated with the significant markers (Meuwissen and Goddard 1996, Spelman et al 1999)

A generalfinding from genome-wide association (GWA) stud-ies for complex traits in animals (humans and livestock) and crop plants (rice and maize) (Sanna et al 2008, Visscher 2008, Huang et al 2010, Lango Allen et al 2010, Tian et al 2011) is that the proportion of variation for a given trait explained by the significant markers is small Although individual loci controlling qualitative disease resistance and phenology traits (such asfloral induction requirements) in crop plants may account for large components of phenotypic variance (Vp), this is unlikely to be the case for most quantitative agronomic traits (such as for yield, quality and persistence) For example, a GWA study for 14 agronomic traits in rice found that the majority of traits were influenced by multiple loci with relatively small effects (Huang

et al 2010) Similar results were obtained from a GWA study of leaf architecture in maize (Tian et al 2011) For perennial ryegrass, individual SNPs in candidate genes explained small

wileyonlinelibrary.com

© 2013 Blackwell Verlag GmbH

proportions (an average value of 8.5%, with a range of

4.4–17.2%) of the phenotypic variation in nutrient quality traits

associated with cell wall biosynthesis and oligosaccharide

metab-olism (L W Pembleton, J Wang, N.O I Cogan, J E Pryce,

G Ye, C Bandaranayake, M L Hand, R C Baillie, M C

Drayton, K Lawless, S Erb, M P Dobrowolski, T I

Saw-bridge, G C Spangenberg, K F Smith, and J W Forster,

sub-mitted) QTL for water-soluble carbohydrate (WSC) content in

the same species also explained a small to moderate proportion

of the genetic variance (Turner et al 2010) If the majority of

complex traits targeted for forage improvement share this

archi-tecture, MAS is unlikely to result in appreciable increases in

genetic gain, as the variance explained by the SNPs with the

largest effects will probably be small

An alternative approach is to simultaneously use all

genome-wide distributed markers to predict breeding values, in an

approach known as genomic selection (GS: Meuwissen et al

2001) GS uses a panel of markers that are sufficiently dense

such that all QTLs are expected to be in LD with at least one

marker locus The major advantage of GS over MAS is that all

of the genetic variance may potentially be captured by the

mark-ers, as marker effects do not need to exceed a significance

threshold in order to be used to predict breeding value

In order to implement GS, a reference population of

individu-als with both known genotypes and phenotypes is used to derive

a prediction equation, for the effect of each marker on the trait

This prediction equation can then be used to estimate genomic

breeding values (GEBV) for selection candidates for which

genotypes, but potentially no phenotypes, have been obtained, as

the sum of the marker allele by the effect of that marker,

summed over all loci The accuracy of GEBV (the correlation of

the GEBV with the true breeding value, which determines the

rate of genetic gain) for selection candidates is as 0.85 in

simula-tion studies (Meuwissen et al 2001) In practice, accuracy values

as large as this have not yet been reported, although accuracies

of GEBV up to 0.71 have been reported in Holstein-Friesian

dairy cattle (VanRaden et al 2009)

The potential value of GS for crop plant breeding has been

recognized (Heffner et al 2009, Janninck et al 2010) In

for-ages, yield and quality of herbage are the key traits for forage

improvement, and support feed conversion into unit quantities of

meat or milk These characters provide good candidates for GS,

as they require expensive and/or destructive late life-cycle

mea-surement Traits such as in-field persistence are also of obvious

interest The availability of GEBV for such traits could mean

lower cost breeding programmes, and accelerated rates of genetic gain This article provides a review of current knowledge relevant to the prospects for GS in forages The review high-lights the fact that LD in most forage species is very limited, in strong contrast to crop species (such as maize and wheat) for which prospects for GS have previously been evaluated, and GS strategies that accommodate these differences are required

Target Species The major pasture species for current and future use in Austra-lian pastoral agriculture are described in Table 1 The table sum-marizes some key elements of biology of these target species that are relevant to GS implementation

The majority of forages belong to either the grass (Poaceae)

or legume (Fabaceae) families Grasses and legumes are either cultivated separately, or in combination as companion species in

a mixed sward The most important temperate pasture grass spe-cies are members of the Lolium–Festuca spespe-cies complex (ryegrasses and fescues), two closely related genera residing within the Poeae tribe of the cool-season grass subfamily Pooi-deae The key cool-season legume species predominantly belong

to the genera Medicago [medics, including alfalfa (Medicago sativa L.)] and Trifolium [clovers, including white clover (Trifolium repens L.) and red clover (Trifolium pratense L.)] within the Trifolieae tribe of the Fabaceae clade Hologalegina

A key feature of the majority of temperate forage species (with the exception of subterranean clover, T subterraneum L.) is that they are obligate outbreeders Examples include perennial rye-grass, Italian ryerye-grass, tall fescue, meadow fescue (Festuca prat-ensis Huds syn L pratense), white clover and alfalfa Inhibition

of inbreeding is due to genetically controlled self-incompatibility (SI) systems, in which incompatible pollen–stigma interactions arise from specific matching of SI locus alleles in male and female reproductive tissues The implication of an obligate outbreeding habit is that forages will display much more limited LD than many crop species, for which inbreeding is possible This has a major effect on the design of GS programmes, as discussed below

Target Traits Herbage yield, persistence and quality are the primary produc-tion traits for forage species The latter is primarily controlled by variation for digestibility (associated with cellulose and lignin polymer content in cell walls) and availability of oligosaccharide

Table 1: Summary information for biological factors in fluencing feasibility of GS implementation in forage species

Attribute

Symbiotic plant –

microbe association

Fungal (Neotyphodium lolii, LpTG2)

Fungal (Neotyphodium occultans)

Fungal (Neotyphodium coenophialum, FaTG2, FaTG3)

Bacterial (Rhizobium leguminosarum)

Bacterial (Rhizobium meliloti)

Key issues for

implementation of GS

Non-extensive LD, likely high N e , development of ef ficient GBS methods required

Limited knowledge of LD and N e , technical constraints on SNP discovery and validation, development of ef ficient GBS methods

Limited knowledge of LD and N e , development of

ef ficient GBS methods capable of determining allele dosage GBS, genotyping-by-sequencing; GS, genomic selection; LD, linkage disequilibrium; SI, self-incompatibility.

carbohydrates for energy provision to the grazing animal (Cogan

et al 2005) Other aspects of quality include reduction of

bloat-ing in ruminants, associated with protein-rich leguminous diets

and ameliorated by the presence of condensed tannins Biotic

stress resistance characters include responses to viral, bacterial

and fungal pathogens and to invertebrate pests Examples include

alfalfa mosaic virus (AMV) of both lucerne and white clover

and the crown pathogen of ryegrasses (Puccinia coronata f.sp

lolii) (Dracatos et al 2010) Abiotic stress–related traits include

tolerance to waterlogging (Pearson et al 2010) and drought in

forage grasses, and to aluminium toxicity, phosphorus deficiency

and saline stress in clovers Capacity to produce sufficient

quan-tities of viable seed is important for both grass and pasture

legume varieties All of these traits are excellent targets for GS,

due to requirements for laborious, expensive, destructive or late

life-cycle measurement

Current Structure of Pasture Plant Breeding

Programmes

Many different forms of breeding programme have been

implemented for pasture plant species A generic scheme,

which describes most of the relevant features of current

com-mercially orientated breeding programmes, is depicted in

Fig 1 The majority of commercial activities have been based

on a strategy that begins with establishment of a base popula-tion of up to 10 000 individuals, followed by seed multiplica-tion within families, generating up to 100 000 individuals for mass selection Evaluation for persistence under grazing pres-sure or visual assessment (as spaced plants in the field) is used for subselection with a reduction by a factor of 100 in the number of individuals evaluated A surviving group of up

to 1000 potential parental clones undergo further assessment for key performance characteristics such as yield and persis-tence In addition, it is possible to determine parental breeding value through the use of a number of experimental methods (Vogel and Pedersen 1993) In half-sib progeny testing (HSPT), maternally derived progeny from a selected individual are evaluated in replicated phenotypic trials to minimize con-founding environmental variation, obtaining information on the general combining ability (GCA) of the parent Full-sib prog-eny testing (FPST) also measures the specific combining abil-ity (SCA) of a parent through the identification of superior pair-cross-derived families Within- and between-family selec-tion (WBFS) involves establishment of multiple polycrosses (random intermatings between selected individuals), harvest and bulking of equal seed numbers from each mother plant, and establishment and evaluation of replicated spaced-plant half-sib progeny nurseries for key production characters, with

a primary emphasis on yield

F1 Production

Seed production (F2 Production)

Selection under grazing and/or visual assessment

Varietal construction

Seed production

Multi-environment plot trials

1 Variety release

Base population

establishment

c 1000 – 10 000

Individuals

c 100 000 Individuals

Reduction in

individuals by a

factor of 100

Selective recombination

Non-selective recombination

Selective

recombination

c 1–10 Breeding Lines

Multi-environment plot trials

Less than 100 Breeding Lines

Non-selective recombination

Less than 100 Breeding Lines

Syn1 Generation

Syn2 Generation Syn0 Individuals

Fig 1: Generic scheme for a current commercial ryegrass breeding programme.

Following the selection of foundation (Syn0) individuals based

on the progeny performance as described above, a synthetic 1

(Syn1) population is generated This is usually achieved by

poly-crossing, less commonly by combination of F1seed from crosses

between each parent in a diallel structure The number of

foun-dation individuals may vary from as low as four for perennial

ryegrass (although higher for Italian ryegrass) to 50–100 for

polyploid species such as tall wheat grass and alfalfa (Bray and

Irwin 1999) Synthetic 2 (Syn2) populations are then obtained by

Syn1 multiplication, through seed harvest from maternal parents

fertilized within a single common pollen cloud The Syn2

popu-lations are then assessed in multiple environments for key traits,

leading to the selection of one population for commercial release

as a variety

Commercial breeding programmes that implement specific

ver-sions of this generic scheme will typically involve two cycles of

selective genetic recombination and subsequent

selection/evalua-tion within a 6- to 9-year period (Fig 1) Given the extended

time frame for these stages, there are major opportunities for

acceleration of genetic gain based on GS strategies However,

implementation of GS in commercial forage breeding

pro-grammes is likely to require major structural alterations to fully

exploit the technology In current systems, foundation clones

have rarely been retained, and details of lineage structure have

not generally been recorded, precluding pedigree-based

calcula-tion of estimated breeding values The value of individual

geno-types in this paradigm has been relatively low, limiting the

opportunity to apply strategies that depend on expensive

geno-typing analysis and/or phenotypic characterization A transition

to pedigree-based breeding (or at least the use of marker

infor-mation to capture pedigree structure) through clonal nursery

evaluation of specific genotypes is hence an important step

towards the implementation of GS in pasture plant breeding

Requirements for GS in Ryegrasses

Of the major forage species, the ryegrasses and alfalfa currently

possess the best-developed suites of genetic and genomic

resources (see Li and Brummer 2012 for a review of status in

alfalfa) However, as the genetic systems of ryegrasses are

argu-ably more similar (as outbreeding, diploid taxa) to those of

domestic animals, these species provide a potential test case for

the development of GS strategies in forages (although see Li and

Brummer 2012 for commentary on implementation of GS in

alfalfa breeding systems) In order for this objective to be

achieved, a number of key limitations must be overcome

Availability of SNPs

Genomic selection requires a panel of SNPs distributed

genome-wide which can be assayed across a large number of individuals

at reasonable cost Development of genomic resources for

peren-nial ryegrass has until recently been relatively slow compared to

major cereal species such as rice, maize, wheat and barley

Ini-tial large-scale SNP discovery for perennial ryegrass was

per-formed through functional selection of gene sequences, amplicon

generation from parents of full-sib mapping families, followed

by cloning, sequencing, alignment and validation through the

use of a Mendelian transmission test (Cogan et al 2006) The

consensus from multiple SNP discovery studies (Cogan et al

2006, Xing et al 2007, Dracatos et al 2008, 2009, Braazauskas

et al 2010, Fiil et al 2011) is that SNP frequency in ryegrass

populations typically varies between 1 per 20–150 bp DNA

sequencing of pooled gene-specific amplicons from multiple (ca 500) genotypes further suggests a high‘global’ average value

of between 1 per 20–25 bp (Cogan et al 2010)

A number of approaches have been implemented to generate larger numbers of SNPs with genome-wide distribution In one early approach, groups of candidate genes were selected and these regions were sequenced using the second-generation Roche

GS FLX platform Potential SNPs were formatted for genotyping using the Illumina GoldenGateTM

384-plex assay (Cogan et al 2010a) A similar approach based on SNP discovery in tran-scribed sequences was implemented by Studer et al (2012) An approach that allows discovery of a larger number of SNPs, with broader genome coverage, has been based on comparative ge-nomics to select exon regions that are regularly distributed across the ryegrass genome for pooled sequencing Species for genomic comparisons include bread wheat (Jones et al 2002) and Brachypodium distachyon (Cogan et al 2010b) ‘Proof-of-concept’ for this activity has come from fine-scale genetic and physical mapping of the SI loci (Shinozuka et al 2010) Com-plexity reduction methods are also useful to survey the whole gene space, using methods such as C0tfiltration and hypomethy-lation (Forster et al 2010)

Collectively, the discovery activities described above have delivered ca 20 000 predicted SNP loci, sufficient to support an initial design for an integrated high-density oligonucleotide-based genotyping chip, analogous to those used in livestock GWAS and GS studies However, the dramatic increase in power of current sequence technologies suggests that whole-genome sequencing per se, coupled with comparison between contrasted genotypes, is now a more cost-effective route to genome-wide SNP discovery Complete assembly of a complex Poaceae genome is still highly challenging for second-generation platforms (Gupta 2008) However, assembly of the gene space

as unigene contigs is now feasible based on sequencing with Illumina Hi-Seq, which has delivered ca 60 X coverage of the perennial ryegrass genome (Cogan et al 2011)

In the near future, SNP discovery and subsequent genotyping

in species such as ryegrasses are likely to converge through a transition from genotyping per se to ‘genotyping-by-sequencing’ (GBS) methods (Huang et al 2009, Elshire et al 2011) For instance, the restriction site–associated DNA (RAD) method (Baird et al 2008) provides access to an essentially unlimited set of sequence polymorphisms through second-generation sequencing of reduced complexity representations from specific genotypes (Wang et al 2011) This approach is likely to be essential for ryegrass and other forage species, in order to obtain the large numbers of markers and low cost of genotyping that are necessary given the population structures and current breed-ing practices for these crops

The Challenge of Limited LD The extent of LD is a key parameter determining the accuracy

of GS, as it determines the proportion of genetic variance that can be captured with the SNP markers, which in turn limits the accuracy of GEBV that can be achieved LD is commonly mea-sured by the r2 statistic, which may also be interpreted as the proportion of variance in a trait explained by a SNP in LD with

a QTL (Hill and Robertson 1968) The results from LD studies

in ryegrass have revealed useful LD (r2> 0.25) extending at best 1 kb (Auzanneau et al 2007, Ponting et al 2007, Xing

et al 2007, Fiil et al 2011) For example, Ponting et al (2007) demonstrated that LD, as measured by r2, decayed to <0.27

within 1000 bp in most population types (including ecotypes

and cultivars) Fiil et al (2011) also found that LD decayed

rap-idly within the representative genes, although some genes

exhib-ited much slower rates of decay

Both the extremely high frequency of SNP and rapid decay of

LD suggest a very large past effective population size (Ne) value

for perennial ryegrass The expectation of r2is 1

4Nec þ1 where Ne

is effective population size and c is the map distance between

loci in Morgans (Sved 1971) As a consequence, very large Ne

values are required to give the low r2values that observed For

comparison, in humans, LD decays to 0.27 at ca 25 kb, so Ne

for ryegrass is likely to be larger than the tens of thousands

esti-mated for the human population (e.g Tenesa et al 2007) In

cat-tle, LD decays to 0.27 at ca 50 kb, and ancestral population

sizes are in the order of 2000–3000 (Bovine Hap Map

Consor-tium 2009)

The pattern of LD in ryegrass is also very different to that

observed in crops like maize, for which LD within heterotic

groups, and even across highly diverse germplasm, is extensive

(Yan et al 2009, Van Inghelandt et al 2011), extending across

Megabase (Mb) distances This property allows accurate

estima-tion of GEBVs As an example, Riedelsheimer et al (2012)

reported accuracies of GEBV of 0.72 for biomass in maize using

a set of 56 110 SNPs, while Albrecht et al (2011) reported

accuracies of GEBV of 0.72–0.74 for grain yield in test cross

progeny, using only 1152 SNPs Extensive LD also exists in

wheat: Hao et al (2011) reported useful LD extending up to

5 Mb in populations of Chinese bread wheats, which has been

reflected in high accuracies of genomic prediction (e.g Crossa

et al 2010) In barley, moderate accuracies of GEBV have been

reported, even with limited numbers of markers and small

refer-ence populations (Li and Sillanp€a€a 2012)

The extent of LD affects GS in two ways – the number of

required SNPs, and the number of individuals that must be

geno-typed and phenogeno-typed in the reference population Meuwissen

(2009) demonstrated by simulation that to achieve accurate GS,

10*Ne*L markers are required, where L is the genome length in

Morgans (M) If Ne is in the order of 10 000s, and we assume

the ryegrass genome is ca 8 M in length (Jones et al 2001), this

implies that at least 1 million markers are necessary to

imple-ment GS in ryegrass

The second element for design of a GS experiment is to

deter-mine the number of fully phenotyped and genotyped individuals

that are required in the reference population The accuracy of

GEBVs in individuals with no phenotype of their own can be

derived deterministically and depends on number of individuals

genotyped and phenotyped in the reference population,

heritabil-ity of the trait and number of loci affecting the trait (Daetwyler

et al 2008)

r¼

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi hb

h þ 1  h2r2

r

where h ¼Th 2 b

q , T is the number of phenotyped and genotyped

animals, and q is the number of loci affecting the trait The

for-mula is not explicit because r (accuracy) appears on both sides,

but can be readily solved for r and is presented in this form in

order to make the components easier to understand

Given that little knowledge is available regarding the number

of loci affecting most agronomically important traits, equivalence

between the number of loci and the number of independent

chromosome segments in the population is a conservative

assumption, and this can be calculated from Ne and L, as 2NeL/ log (2Ne) (Goddard 2008) This deterministic prediction suggests that large reference populations are required to predict accurate GEBVs, particularly for low heritability traits (Fig 2a), and agrees well with accuracies that have been achieved in dairy cat-tle breeding experiments (Hayes et al 2009a,b) Once again, given that Ne is likely to be very large in ryegrass populations, the deterministic prediction would indicate that extremely (and impractically) large reference populations would be required to implement GS for ryegrass, in the absence of any strategy to reduce Ne and therefore increase LD Such strategies actually become a key component of GS implementation for ryegrasses

As the equation above for the accuracy of genomic predictions demonstrates, the heritability of the trait is also a key parameter determining the accuracy of genomic predictions Heritability values for target traits in forage improvement, at least, are unli-kely to limit the accuracy of GEBV Heritability values that have been reported for yield and quality traits in ryegrass and other forage species, all of which are greater than zero, and most of which are moderate to high in magnitude, are summarized in Table 2

It is important to recognize that GS may deliver gains even if marker density is not as high as suggested values, and the size of the reference population is lower than desired, provided that some structure is present within the population (such as, for example, if full-sib or half-sib matings are in common) SNP markers may also

be used to trace the pedigree of the population and the inheritance

of large chromosomal blocks from parents and other recent ances-tors Habier et al (2007) demonstrated that a considerable propor-tion of GEBV accuracy was actually derived from these sources, particularly when Ne has recently contracted, as must have occurred in most forage breeding programmes In the strategy out-lined below, an artificial reduction in Neand therefore increased extent of LD are specifically exploited

Strategies for Application of GS to Forage Breeding Programmes

As well as limited LD, a significant challenge in applying GS to pasture breeding programmes is the assembly of a useful reference population The most important phenotypes (yield and persistence) are usually evaluated in swards, given that phenotypes measured

on individual plants in the absence of competition show a rela-tively poor correlation with persistence and yield of the same genotypes in swards (Connolly 2001) Nevertheless, GS could be applied to existing ryegrass breeding programmes, such as that depicted in Fig 1 To take full advantage of the technology, how-ever, it is likely that the breeding scheme would need to be re-structured A scheme that wouldfit relatively simply into the existing schemes is first described, followed by a more radical scheme Thefirst strategy largely makes use of LD which exists within families (for which Neis very small), which is equivalent to the use of linkage information to predict GEBV The second scheme exploits both within- and between-family LD

Scheme 1: Reduction of costs through the use of GS Thefirst scheme applies GS at the stage of selecting Syn2 popu-lations for evaluation (Fig 1) The reference popupopu-lations for cal-culating GEBV for persistence and yield are actually the polycross-derived progeny sets that are established in a field trial Two additions to the standard system are required The individuals that must be genotyped are the foundation parents,

0 50 100 150 200 250

Accuracy of genomic breeding value

0 2000 4000 6000 8000

10 000

12 000

14 000

16 000

18 000

20 000

(a)

(b)

Heritability of trait

Accuracy of genomic breeding value = 0.7 Accuracy of genomic breeding value = 0.5

Fig 2: (a) Size of reference populations required for genomic selection (GS), and (b) genetic gains that can be achieved with GS as percentage of gain achieved by existing schemes (100% = gain achieved by current scheme).

Table 2: Heritability of traits in some forage species Unless otherwise indicated, heritability is narrow sense

Days first anthesis to harvest 0.23 Baert et al (2010) 12 varieties, sown in 3 replicates a-Linolenic acid content 0.79

Crush et al (2006) 20 half-sib families in a open

pollination pool

Bonos et al (2005) 54 cultivars or experimental populations Grey leaf spot resistance 0.95 1 Easton et al (2002)1 Partial diallele cross and their

twelve parent clones

Endophyte-derived alkaloid content 0.58 –0.72 Tall fescue Baert et al (2010) 3 varieties sown in three replicates a-Linolenic acid content 0.39

White clover Baert et al (2010) 3 varieties sown in three replicates a-Linolenic acid content 0.28

Annicchiarico et al (1999) 16 parents with polycross progeny Dry matter yield 0.52  0.29

1 Broad-sense heritability.

the Syn1 progeny and the Syn2 individuals Both the Syn1

prog-eny and potential synthetic parents can be genotyped with

mark-ers of very low density, due to restricted recombination in the

limited number of generations (only one in the case of the Syn1)

between these individuals and the parents Due to the limited

recombination between foundation lines and Syn2 individuals,

GS may be envisaged as evaluating each foundation line haploid

chromosome and then selecting the Syn2 progeny that have the

best combination of these haploid chromosomes for further

eval-uation The reduction in number of Syn2 progeny that can be

tested in varietal trials depends on the heritability of yield and

persistence While these parameters are not well understood, the

heritability of potential component traits (such as root system

size for persistence) is moderate in value (Table 2) (Crush et al

2006) The extra accuracy that can be obtained from the

applica-tion of GS within a family, as opposed to selecapplica-tion only on

parental average, has been determined by Hayes et al (2009a)

Using a deterministic formula, it was reported that if 1000 F1

full-sib individuals are measured for the trait, and h2= 0.1, the

accuracy of GS would be 0.62, as opposed to 0.49 from

selec-tion on parental average alone (Hayes et al 2009a) With this

accuracy, the number of varieties taken through to field

evalua-tions could be reduced from 100 to 27, with a low probability of

losing the most superior population at this step This number

was obtained by performing a small simulation with an accuracy

of 0.62 and by determining the maximum number of individuals

required to be selected on GEBV in order to ensure that the

indi-vidual with the highest true breeding value was always part of

the selected group, with 1000 replicates As the reference

popu-lation expands over various cycles of breeding, and more and

more foundation parents are evaluated, it should be possible to

implement GS at the F1stage, so as to accurately select

individ-uals within this group as parents to enter the progeny testing

step, and so further reduction in programme costs

Scheme 2: Towards doubling the rate of genetic gain with

GS

A more radical GS scheme would omit the progeny test stage

completely, and cycle through multiple rounds of selection and

mating based on GEBV alone In this type of scheme, multiple

cycles of selection and crossing take place until the predicted

yield and quality measurements exceed those of the best current

cultivar by a desired amount, at which stage field trials are

implemented This scheme requires a reference population that

has been generated for the sole purpose of enabling GS

How much additional gain is possible with this scheme?

Genetic gain per year is given byDG ¼irrgL , where i is the

inten-sity of selection, r is the accuracy of GEBV or traditional

breed-ing value, rg is the genetic standard deviation, and L is the

generation interval If the progeny testing step is omitted, it

would be possible to halve the generation interval from 9 to

4.5 years Provided the accuracy of GEBV was the same as the

accuracy of progeny testing, this would double the rate of

genetic gain In practice, this is unlikely, as very large reference

populations would be required to achieve the same accuracy as

progeny testing (which is ca 0.8 given the current scale of

test-ing) (Fig 2a) Using the formula for genetic gain as stated

above, and with a halving of the generation interval, the

accu-racy of GEBV would have to exceed 0.4 before the GS scheme

would result in larger gain than for current schemes (Fig 2b) If

the accuracy of GEBV was 0.6, the GS scheme would increase

genetic gain by 50%

In practice, in the first few rounds of the new scheme, addi-tional gains may be limited, as the accuracy of the GEBV will likely be low, until reference information accumulates for target traits The actual size of the requisite reference population is dif-ficult to determine the absence of heritability estimates for per-sistence-related traits, but previously stated, values for potentially correlated characters are encouraging

The other attraction for such a GS scheme is that it is closed,

in that Ne would be reduced, and therefore the extent of LD increased, as compared to the broader ryegrass gene pool This means that as the scheme continues, the contribution of between-family information to the accuracy of GEBV will increase However, as reduced Ne increases the rate of inbreed-ing, simultaneous selection and mating to maximize gain with an acceptable level of inbreeding would be recommended, and this can be also achieved through the use of genomic information (Kemper et al 2012)

Given these generic features of the second scheme, specific details for the establishment of the reference population are now presented:

1 The scheme begins with the evaluation of available elite germ-plasm This is achieved by the establishment of a spaced-plant in-field nursery incorporating individuals (c 1000) from multi-ple elite germplasm sources, with a moderate level of clonal replication (e.g fourfold), to obtain c 4000 ramets (Fig 3) As many current elite varieties of perennial ryegrass have a restricted base (Guthridge et al 2001), being derived from a small number of parents (4–6), the inclusion of commercial germplasm will predispose towards a reduction of Nefrom the outset Importantly, this base population should include all varieties or cultivars which will subsequently be used for breeding activities Phenotypic evaluation for traits such as yield and quality is performed on the replicated genotypes

2 Selection of the 150 best-performing plants is performed based on the nursery evaluation data A series of paired crosses between these selected superior individuals is made

In order to allow more accurate prediction of breeding values between progeny sets, and to further reduce Ne, 50 genotypes are designated as‘bigamous’ parents, as each is mated to two separately selected ‘monogamous’ parents This procedure obtains 50 pairs of F1 families, each pair being half-sibs based on the common bigamous parent (Fig 3, box B) GBS

of the 150 parents is performed based on a complexity reduc-tion sequencing method (e.g Baird et al 2008, Elshire et al 2011) to identify genome-wide SNP variation within the parental group (Fig 3, box B)

3 Seed from each of the 100 full-sib families is harvested for the establishment of mini-swards for the evaluation of target traits under competition Although harvesting of mini-swards

is not a perfect substitute for persistence under grazing, mechanical cutting has been found to reasonably well mimic the effects of intense grazing, such that moderate to high cor-relations are obtained between rankings for each regime (Camlin and Stewart 1975) After establishment by close sow-ing, phenotypic evaluation for yield, quality and other traits such as disease resistance, as appropriate, is performed on the mini-sward as an individual unit (Fig 3, box C) For compo-nents of vegetative persistence under grazing or cutting, either the family-specific mini-swards or the spaced-plant nursery from step 1 may be retained as separate entities for longer-term evaluation, in addition to establishment of the new popu-lations in each cycle of selection

4 The resulting performance data are used for selection of the

10 most superior F1 families (10% selection) Retained seed

from each family is germinated, and 100 individuals from

each family are then used to regenerate the spaced-plant

nurs-ery, once again at a size of 1000 genotypes (Fig 3, box D)

5 Genotypic analysis of the selected F1 individuals from round

1 of the process can be performed using lower-density SNP

assays (e.g 384-plex Illumina GoldenGateTM

assays), as impu-tation can be used to infer their genotypes at the SNPs

identi-fied in step 2 Habier et al (2009) demonstrated that this

process can be achieved with high accuracy if the parents

have been genotyped for high-density markers, as for this

example (step 2) In parallel, phenotypic evaluation of

selected individuals is performed as previously described for

the base population From use of the imputed genotypic data

and individual plant performance information, GEBV

predic-tion equapredic-tions for key traits are derived (Fig 3, box E)

6 Selection of the best-performing individuals from the second

round of spaced-plant assessment identifies another set of 150

parents for crossing, subsequent mini-sward evaluation and

return to the spaced-plant nursery In each successive round,

the GEBV prediction equation is refined until the convergence

between predicted and actual performance is sufficiently close

to omit the phenotypic assessment steps, from which point

onwards, selection is based solely on genotypic data

7 Multiple GEBV-based selection and breeding cycles are per-formed until the values for a proportion of genotypes exceed that of current varieties by the necessary amount, at which point variety development is initiated At any stage, a subset

of elite individuals from the selected cohort can be diverted into polycrossing in order to obtain restricted-base synthetic populations for agronomic evaluation

8 If novel genotypes with desirable alleles are identified outside the scope of the breeding system, individuals from this germplasm can be added to the clonal nursery and should be genotyped with the densely spaced markers to allow accurate introgression

One of the key objectives of this proposed scheme is to enhance the significance of individual genotypes during breeding practice, which have traditionally been negligible compared to the population as a whole Retention, intensive phenotypic char-acterization and use in a fully recorded (SNP genotype) pedigree structure will deliver at least part of this worth

Statistical methodology The statistical method that is used to derive the SNP prediction equation for the calculation of GEBV in the above scheme could affect the accuracy of these estimates Published methods for

M 1 B M 2

(c) (d)

(e)

Genomic selection

Update prediction equation

Multi-site environment trials

Fig 3: Proposed genomic selection (GS)-based breeding programme for forage species Establishment of the scheme requires the following steps: (a)

a spaced-plant clonal nursery, under no competition of c 1000 genotypes with fourfold replication Individual phenotypic assessment is performed for yield and quality This leads to the selection of 150 elite parental genotypes (b) The 150 selected parents are comprehensively genotyped, probably using a GBS analysis ( >50 000 SNPs) Fifty bigamous (B) parents are crossed to 100 monogamous (M) parents to generate 50 pairs of half-sib fami-lies (c) Resulting half-sib families are sown under close spacing as ‘mini-swards’ at typical pasture sowing rates Phenotypic analysis is performed for yield and quality, with the data developing and re fining the GS equation Elite families are also identified and are used to create the second generation

of a spaced-plant clonal nursery, as in (a) All individuals entering the new clonal nursery are genotyped at lower resolution to characterize inherited parental blocks This process is repeated until the GS equation has achieved suf ficient accuracy (>0.4, see text) to remove the need for the ‘mini-sward’ assessment, at which point the half-sib families are germinated, genotyped and elite plants identified to create the subsequent generation of the spaced-plant clonal nursery, as in (a) Multiple cycles of breeding and GS can take place without the mini-sward step, reducing the generation interval The ‘mini-swards’ will, however, be required to periodically update the prediction equation, particularly as most of the information will be from within-family and will therefore erode rapidly Generation of a cultivar for extensive trialling (d) may occur from any of the generations of the clonal plant nursery when performance of the elite individuals is predicted to exceed current varieties by a desired amount GBS, genotyping-by-sequencing.

deriving the prediction equation differ in the prior assumptions

made of the distribution of SNP effects, which in turn reflects

the distribution of QTL effects and LD between SNPs and QTL

The prior assumptions of the SNP effects can range from: a large

number of small and normally distributed effects (SNPBLUP); a

t distribution of effects with many small effects but a small

num-ber of moderate to large effects (BAYESA); many SNPs with

zero effects and a few SNPs following a t distribution of effects

(BAYESB); and a double exponential distribution of effects

(BAYESSIANLASSO) These methods are reviewed in more

detail in the study by de Los Campos et al (2012)

In real-world data, methods that assume very many small

effects which follow a normal distribution perform well For

example, Verbyla et al (2009) observed little difference in

GEBV accuracy for production traits in Australian dairy cattle

when some of the above methods were compared The only

exception was percentage fat content, for which trait a known

mutation that explains c 40% of the variance has been described

(Grisart et al 2002) For this trait, a method with the prior

assuming the effects of many SNPs were zero, and a small

pro-portion had moderate to large effects, outperformed the rest

For the issues addressed in this study, a method is required

for effective capture of particularly linkage information, as well

as any LD information The relative contributions of these two

factors are likely to be large and small, respectively, within the

family structure that has been created A moderately

straightfor-ward method capable of capturing linkage information is to

pre-dict breeding values using a genomic relationship matrix, in

place of the pedigree-derived relationship matrix (e.g Habier

et al 2007, Goddard 2008, VanRaden et al 2009, Hayes et al

2009a,b) This model can be shown to be equivalent to

predic-tion of individual SNP effects and calculapredic-tion of GEBV as the

sum of these effects, provided the SNP effects are assumed to be

normally distributed

The model is (ignoring fixed effects that should be fitted in

practice):

y¼ 1nl þ Zg þ e where y is a vector of phenotypes, l is the mean, 1nis a vector

of 1s, Z is a design matrix allocating records to breeding values,

g is a vector of breeding values, and e is a vector of random

normal deviates~ Nð0; r2

eÞ Then, g = Wu where ujis the effect

of the jth SNP, and VðgÞ ¼ WW0r2 Elements of matrix W are

wij for the ithplant and jth SNP, where wij= 02pj if the plant

is homozygous 11 at the jthSNP, 12pjif the plant is

heterozy-gous and 22pjif the plant is homozygous 22 at the jth SNP

The diagonal elements of WW′ will be Pm

j ¼12pjð1  pjÞ where

m is the number of SNPs If WW′ is scaled such that

G¼PnWW n 0

i¼1wii

, then VðgÞ ¼ Gr2 Then, breeding values for both

phenotyped and non-phenotyped individuals can be predicted by

solving the equations:

g^

h i

¼ Z0Zþ G1r2

e

r2

Z0y

Genomic selection implementation on this basis is attractive,

as all that may be required are to replace the average relationship

matrix with the genomic relationship matrix in the existing

genetic evaluation The method is also very attractive for

popula-tions that lack good pedigree records (as is the situation for

rye-grass breeding programmes), in that the genomic relationship

matrix will capture this information, at least among the geno-typed individuals For real-world data, this method has been shown to be at least as effective as other methods for many traits (VanRaden et al 2009) In addition, Hayes et al (2009a,b) demonstrated that the method is suitable for capture of linkage information

Conclusions Although forage species have been relatively undeveloped in terms of molecular breeding compared to other major crop plants, GS implementation has the potential to deliver major advances, mainly through the capacity to complete multiple selection rounds within time periods conventionally used for sin-gle rounds This outcome is possible if accurate GEBVs can be predicted for important traits such as yield, quality and persis-tence in swards Good information on these traits is currently obtained only 5 years into the breeding cycle However, if GS is

to be implemented in forage species, a number of challenges must be overcome The relative deficiencies of DNA marker resources, and influence of polyploid genome structures for some species, constitute the first challenge Barriers to marker avail-ability will rapidly disappear as GBS becomes less expensive, while enhanced methods for genotypic analysis of polyploid ge-nomes have also been developed (Gidskehaug et al 2011) The two major remaining challenges are the very limited extent of LD in forage species such as ryegrasses, and restricted opportunities to implement GS in current breeding programmes

In this review, the first factor is addressed through the use of both linkage information within families to increase the accuracy

of GEBV prediction, and, in the longer-term, reduction of Ne in populations by breeding from elite varieties Undesirable corre-lated effects of inbreeding depression under such schemes could

be managed through the incorporation of measures of genomic diversity into the selection criteria (e.g Pryce et al 2012) To address the second factor, a breeding scheme has been proposed which permits GS to accelerate genetic gain through the reduc-tion in generareduc-tion interval However, implementareduc-tion of this scheme would require restructuring of current breeding systems

Acknowledgements The authors acknowledge the support from the Victorian Department of Primary Industries Research in genomics-assisted breeding of temperate forage species has been funded by Dairy Australia, the Geoffrey Gardiner Dairy Foundation and Meat and Livestock Australia through the Molecu-lar Plant Breeding and Dairy Futures Cooperative Research Centres.

References Albrecht, T., V Wimmer, H J Auinger, M Erbe, C Knaak, M Ouzu-nova, H Simianer, and C C Sch €on, 2011: Genome-based prediction

of testcross values in maize Theor Appl Genet 123, 339 —350 Annicchiarico, P., E., Piano, and I Rhodes, 1999: Heritability of, and genetic correlations among, forage and seed yield traits in Ladino white clover Plant Breeding 118, 341 —346.

Auzanneau, J., C Huyghe, B Julier, and P Barre, 2007: Linkage dis-equilibrium in synthetic varieties of perennial ryegrass Theor Appl Genet 92, 837 —847.

Baert, J., M Vandewalle, J De Riek, J De Boever, V Fievez, and C Van Waes, 2010: Variation and heritability of a-linolenic acid content and rumen escape protein fraction in fodder grass and clover In: S Barth, and D Milbourne (eds), Breeding Strategies for Sustainable Forage and Turf Grass Improvement, 181 —188 Springer, Berlin.

Baird, N A., P D Etter, T S Atwood, M C Currey, A L Shiver, Z.

A Lewis, E U Selker, W A Cresko, and E A Johnson, 2008:

Rapid SNP discovery and genetic mapping using sequenced RAD

markers PLoS ONE 3, e3376.

Bonos, S A., D., Rush, K Hignight, S Langlois, and W A Meyer,

2005: The effect of selection on gray leaf spot resistance in perennial

rye grass Intl Turfgrass Soc Res J 10, 501 —507.

Braazauskas, G., I Pasakinskiene, T Asp, and T L€ubberstedt, 2010:

Nucleotide diversity and linkage disequilibrium in five Lolium perenne

genes with putative role in shoot morphology Plant Sci 179,

194 —201.

Bray, R A., and J A G Irwin, 1999: Medicago sativa L (lucerne) cv.

Hallmark Austr J Exp Agric 39, 643—644.

Camlin, M S., and R H Stewart, 1975: Reaction of Italian ryegrass

cul-tivars under grazing as compared with cutting Grass Forage Sci 30,

121—130.

Cogan, N O I., K F Smith, T Yamada, M G Francki, A C

Vec-chies, E S Jones, G C Spangenberg, and J W Forster, 2005: QTL

analysis and comparative genomics of herbage quality traits in

perennial ryegrass (Lolium perenne L.) Theor Appl Genet 110,

364 —380.

Cogan, N O I., R C Ponting, A C Vecchies, M C Drayton, J.

George, M P Dobrowolski, T I Sawbridge, G C Spangenberg, K.

F Smith, and J W Forster, 2006: Gene-associated single nucleotide

polymorphism (SNP) discovery in perennial ryegrass (Lolium perenne

L.) Mol Genet Genomics 276, 101 —112.

Cogan, N O I., R C Baillie, M C Drayton, K Savin, C Anderson,

G C Spangenberg, and J W Forster, 2010: Massively-parallel

sequencing technology permits direct detection of novel SNP

haplo-types in perennial ryegrass (Lolium perenne L.) Molecular Breeding

of Forage and Turf 2010, Abstract P-6, 51 —52 Buenos Aires,

Argentina.

Cogan, N O I., R C Baillie, M L Hand, M C Drayton, P Tian, T.

Webster, R Chapman, G C Spangenberg, and J W Forster, 2010a:

Accelerated SNP discovery in perennial ryegrass based on pooled

amplicon resequencing Molecular Breeding of Forage and Turf 2010,

Abstract P-4, 47 —48 Buenos Aires, Argentina.

Cogan, N O I., M L Hand, T I Sawbridge, R C Baillie, and J W.

Forster, 2010b: Comparative genomics in the grass family Poaceae for

structured genome-wide SNP discovery in perennial ryegrass (Lolium

perenne L.) Molecular Breeding of Forage and Turf 2010, Abstract

P-5, 49 —50 Buenos Aires, Argentina.

Cogan, N O I., J Wang, C Inch, D Pickett, R Elton, R C Baillie, M.

C Drayton, H Shinozuka, and J W Forster, 2011: Genomic tools for

pasture plant breeding Plant and Animal Genome XIX, Abstract

W335 San Diego, California, USA.

Connolly, V., 2001 Breeding improved varieties of perennial ryegrass.

Teagasc – Project Report – 3495, ISBN 1-84170-199-8 17.

Crossa, J., G de Los Campos, P Perez, D Gianola, J Burgue~no, J L.

Araus, D Makumbi, R P Singh, S Dreisigacker, J Yan, V Arief,

M Banziger, and H J Braun, 2010: Prediction of genetic values of

quantitative traits in plant breeding using pedigree and molecular

markers Genetics 186, 713 —724.

Crush, J R., H S Easton, and J E Waller, 2006: Genetic variation

in perennial ryegrass for root pro file characteristics In: C E Mercer

(ed.), Breeding for Success: Diversity in Action, Proceedings of the

13th Australasian plant breeding conference, Christchurch, New

Zealand, April 2006, 1073 —1077 New Zealand Grassland

Associa-tion, Dunedin.

Daetwyler, H D., B Villanueva, B , and J A Woolliams, 2008:

Accu-racy of predicting the genetic risk of disease using a genome-wide

approach PLoS ONE 3, e3395.

Dracatos, P M., N O I Cogan, M P Dobrowolski, T I Sawbridge, G.

C Spangenberg, K F Smith, and J W Forster, 2008: Discovery and

genetic mapping of single nucleotide polymorphisms in candidate

genes for pathogen defence response in perennial ryegrass (Lolium

perenne L.) Theor Appl Genet 117, 203—219.

Dracatos, P M., N O I Cogan, T I Sawbridge, A R Gendall, K F.

Smith, G C Spangenberg, and J W Forster, 2009: Molecular

charac-terisation and genetic mapping of candidate genes for qualitative dis-ease resistance in perennial ryegrass (Lolium perenne L.) BMC Plant Biol 9, 62.

Dracatos, P M., N O I Cogan, P J Keane, K F Smith, and J W Forster, 2010: Biology and genetics of crown rust disease in

ryegrass-es Crop Sci 50, 1605 —1624.

Easton, H S., G C Latch, B A Tapper, and O J Ball, 2002: Ryegrass host genetic control of concentrations of endophyte-derived alkaloids Crop Sci 42, 51 —57.

Elgersma, A., 1990: Heritability estimates of spaced-plant traits in three perennial ryegrass (Lolium perenne L ) cultivars Euphytica 51,

163 —171.

Elshire, R J., J C Glaubitz, Q Sun, J A Poland, K Kawamoto, E S Buckler, and S E Mitchell, 2011: A robust, simple genotyping-by-sequencing (GBS) approach for high diversity species PLoS ONE 6, e19379.

Fiil, A., I Lenk, K Petersen, C S Jensen, K K Nielsen, B Schejbel,

J R Andersen, and T L €ubberstedt, 2011: Nucleotide diversity and linkage disequilibrium of nine genes with putative effects on flowering time in perennial ryegrass (Lolium perenne L.) Plant Sci 180,

228 —237.

Forster, J W., N O I Cogan, J Wang, G Ye, C Bandaranayake, M.

L Hand, R C Baillie, M C Drayton, K Lawless, M P Dobrowol-ski, S Erb, and K F Smith, 2010: Development and implementation

of molecular genetic tools for enhancement of herbage quality and other agronomic traits in perennial ryegrass (Lolium perenne L.) Molecular Breeding of Forage and Turf 2010, Abstract, 18 —19 Bue-nos Aires, Argentina.

Ganal, M W., G Durstewitz, A Polley, A Berard, E S Buckler, A Charcosset, J D Clarke, E M Graner, M Hansen, J Joets, M C Le Paslier, M D McMullen, P Montalent, M Rose, C C Sch€on, Q Sun, H Walter, O C Martin, and M Falque, 2011: A large maize (Zea mays L.) SNP genotyping array: development and germplasm genotyping, and genetic mapping to compare with the B73 reference genome PLoS ONE 6, e28334.

Gidskehaug, L., M Kent, B J Hayes, and S Lien, 2011: Genotype call-ing and mappcall-ing of multisite variants uscall-ing an Atlantic salmon iSelect SNP array Bioinformatics 27, 303 —310.

Goddard, M E., 2008: Genomic selection: prediction of accuracy and maximisation of long term response Genetica 136, 245 —257 Gout, M., and S Jones, 2006: Estimates of the annual sales of proprie-tary cultivars in Australia, and the value of pastures to the livestock and cropping industries Presentation made to Pastures Australia A-delaide, Australia, February 2006.

Grisart, B., W Coppieters, F Farnir, L Karim, C Ford, P Berzi, N Cam-bisano, M Mni, S Reid, P Simon, R Spelman, M Georges, and R Snell, 2002: Positional candidate cloning of a QTL in dairy cattle:

iden-ti fication of a missense mutation in the bovine DGAT1 gene with major effect on milk yield and composition Genome Res 12, 222—231 Gupta, P K., 2008: Single-molecule sequencing technologies for future genomics research Trends Biotechnol 26, 602—611.

Guthridge, K M., M P Dupal, R K €olliker, E S Jones, K F Smith, and J W Forster, 2001: AFLP analysis of genetic diversity within and between populations of perennial ryegrass (Lolium perenne L.) Euphytica 122, 191 —201.

Habier, D., R L Fernando, and J C Dekkers, 2007: The impact of genetic relationship information on genome-assisted breeding values Genetics 177, 2389 —2397.

Habier, D., R L Fernando, and J C M Dekkers, 2009: Genomic selec-tion using low-density marker panels Genetics 182, 343 —353 Hao, C., L Wang, H Ge, Y Dong, and X Zhang, 2011: Genetic diversity and linkage disequilibrium in Chinese bread wheat (Triti-cum aestivum L.) revealed by SSR markers PLoS ONE 18, e17279.

Hayes, B J., P M Visscher, and M E Goddard, 2009a: Increased accu-racy of arti ficial selection by using the realized relationship matrix Genet Res 91, 47—60.

Hayes, B J., H D Daetwyler, P Bowman, G Moser, B Tier, R Crump, M Khatkar, H Raadsma, and M E Goddard, 2009b: