Identity-by-descent genomic selection using selective and sparse genotyping.

0.22 MB·English

by  ?degard, J?rgen

ØdegårdandMeuwissenGeneticsSelectionEvolution2014,46:3 Genetics
http://www.gsejournal.org/content/46/1/3 Selection
Evolution
RESEARCH Open Access
Identity-by-descent genomic selection using
selective and sparse genotyping
Jørgen Ødegård1* and Theo HE Meuwissen2
Abstract
Background: Genomic selection methods require dense and widespread genotypingdata,posing a particular
challenge if both sexes are subjectto intenseselection (e.g., aquaculturespecies). This study focuseson alternative
low-cost genomic selection methods (IBD-GS) thatuse selective genotypingwith sparse markerpanels to estimate
identity-by-descent relationships through linkage analysis. Our aim was to evaluate thepotentialof these methods
inselection programs for continuous traits measured onsibs of selection candidates ina typical aquaculture
breeding population.
Methods: Phenotypic and genomic data were generated by stochastic simulation, assuming low to moderate
heritabilities (0.10 to 0.30) for a Gaussian trait measuredon sibs ofthe selection candidates ina typical aquaculture
breeding population that consisted of100 families (100training animals and 20 selection candidates per family).
Low-density marker genotype data (~ 40 markers perMorgan) were used to trace genomic identity-by-descent
relationships. Genotyping was restricted to selection candidates from 30 phenotypically top-ranking families and
varying fractions of theirphenotypically extreme training sibs.Allphenotypeswere included inthegenetic analyses.
Classical pedigree-based and IBD-GS modelswere compared based onrealized genetic gain over one generation of
selection.
Results:Geneticgainincreasedsubstantially(13to32%)withIBD-GScomparedtoclassicalselectionandwasgreat-
est with higher heritability. Most of theextra gain from IBD-GS was obtained already bygenotyping the5% pheno-
typically most extreme sibs within the pre-selected families. Additional genotypingfurther increased genetic gains,
butthesewere small when going from genotyping20% of the extremes to all phenotyped sibs. The success of
IBD-GS withsparse and selective genotyping can be explained by the fact that within-family haplotype blocks are
accurately traced even with low-marker densitiesand that most of thewithin-family variance for normally distrib-
uted traits is capturedby a small proportion of thephenotypically extreme sibs.
Conclusions: IBD-GS was substantially more effective than classical selection, even when based onvery few
markers and combined with selective genotypingof small fractions of thepopulation.The study shows that
low-cost GS programs can be successful by combining sparse and selective genotypingwith pedigreeand linkage
information.
Background species like dairy cattle, for which traditional selection
Recently, genomic selection (GS) [1] has led to a para- schemesfocusmainlyonthemale candidates andneces-
digm shift in quantitative genetic analyses and selective sarily involve progeny testing, GS can often be imple-
breeding programs. GS was originally developed on the mented more cost-effectively by genotyping a limited
basis that all selection candidates and training animals number of male selection candidates and reducing the
are individually and densely genotyped. However, imple- need for progeny testing. In species with a higher female
mentation of GS has been hampered in many species by fecundity, selective breeding focuses on both sexes.
the cost of genotyping large numbers of animals. For Aquaculture breeding programs represent the most ex-
treme of these, in which both males and females have an
*Correspondence:[email protected] extremely high fecundity. Moreover, since some of the
1AquaGenAS,P.O.Box1240,Sluppen,NO-7462Trondheim,Norway most important aquaculture species are single spawners
Fulllistofauthorinformationisavailableattheendofthearticle
©2014ØdegårdandMeuwissen;licenseeBioMedCentralLtd.Thisisanopenaccessarticledistributedunderthetermsof
theCreativeCommonsAttributionLicense(http://creativecommons.org/licenses/by/2.0),whichpermitsunrestricteduse,
distribution,andreproductioninanymedium,providedtheoriginalworkisproperlycited.
ØdegårdandMeuwissenGeneticsSelectionEvolution2014,46:3 Page2of8
http://www.gsejournal.org/content/46/1/3
under certain conditions (e.g., Atlantic and Pacific sal- shown that relationships with reference individuals have
mon), progeny testing is uncommon and organized a much higher effect on accuracy of resulting genomic
breeding programs are often based on sib-testing (for breeding value predictions than LD per se [4]. Further-
invasive traits). Such breeding populations are usually more, IBS-GS in pedigreed populations can be used to
large, with many training animals and numerous selec- predict genomic breeding values even in the absence of
tion candidates of both sexes. Genotyping costs are thus LD, which shows that markers capture close relation-
the main obstacle for the successful large-scale imple- ships among genotyped animals, and thereby affect ac-
mentation of GS in aquaculture species. Such costs can curacy of predictions [5]. IBD-GS is expected to use this
be reduced by: (1) genotyping fewer samples (either information more accurately but at the cost of not using
fewer animals or pooled samples) and (2) genotyping LDthatmayexistbetweenmarkersandQTL.
fewer marker loci per sample (e.g., genotyping by se- Use of the IBD-GS method has the potential to reduce
quencing regions of the genome). However, the chal- marker density, while maintaining the efficiency of GS
lenge is to find a balance between genotyping costs and and taking the entire heritable variance into account.
accuracy of selection. Pooled genotyping will be investi- Yet, the number of animals to be genotyped can lead to
gated in a future study and will not be further discussed a significant cost. Thus, in this study, our aim was to in-
here. The current study focuses on the use of combined vestigate whether IBD-GS can be effectively combined
selective and sparse genotyping in GS, which can greatly with selective genotyping of selection candidates and
reduce genotyping costs, but the potential advantage training animals. From this perspective, the advantage of
of GS over classical pedigree-based selection in these IBD-GS over IBS-GS is that the accuracy of predicted
conditionsremainsuncertain. breeding values depends on the number of genotyped
The methods suggested by Meuwissen et al. [1] imply and phenotyped known relatives, rather than on geno-
that identity-by-state (IBS) information from all available typing of the whole population. Hence, the IBD-GS
marker loci is used, i.e., a given marker allele is assumed method may provide accurate genomic EBV (estimated
to have the same statistical effect on phenotype across breeding values) forspecifically targeted familiesbypref-
different familial backgrounds. Thus, all marker alleles erentially genotyping high-ranking families (based on
that are IBS are assumed to have the same effect, which phenotypic/pedigree-based pre-selection). Furthermore,
can be explained by the existence of linkage disequilib- thegenotypingoftraininganimalscouldfocusonthein-
rium (LD) between the marker and one or more quanti- dividuals that are most informative for prediction of
tative trait loci (QTL). However, Luan et al. [2] showed Mendelian deviationsfrom themid-parentmeans,which
small differences in accuracies of genetic evaluations of are the phenotypically most extreme sibs (in both direc-
dairy bulls when using genomic identity-by-descent tions); the intermediate sibs mainly provide information
(IBD) relationships within the known pedigree (traced about mid-parent means (which are easily estimated
by markers) instead of the IBS relationships suggested with classical methods), and are thus less informative for
by Meuwissen et al. [1]. In the following, these two prediction of Mendelian deviations from the parental
methods of GS will be termed IBD-GS and IBS-GS, re- means.
spectively. Assuming base animals are unrelated, IBD- The aim of the study was to quantify whether a low-
GS uses only information from known, and thus close cost IBD-GS scheme could provide a significant increase
relatives, which typically share long chromosomal seg- in genetic gain for traits evaluated on sibs of selection
ments IBD. In comparison, the (LD-based) haplotype candidates, compared to classical selection schemes, as
blocks found across a population are much shorter [3], applied in aquaculture breeding populations. The low-cost
both for domesticated breeds and, even more so, for approachcombinessparsemarkerpanels(orgenotypingby
large wild populations. Tracing large IBD-blocks within sequencing)withselectivegenotypingofasubsampleofthe
a known pedigree does not require dense markers, selection candidates and the phenotypically most extreme
which makes IBD-GS less sensitive to marker density, trainingsibsofthesecandidates.
comparedtothemore widely usedIBS-GS.
Although IBS-GS does not use the pedigree explicitly Methods
(and thus IBD information), it does use the available Simulation
population structure, i.e., IBD relationships necessarily The QMSim software [6] was used to simulate all data-
imply sharing of marker alleles (but not vice versa). sets.Allscenarioswere replicated50times.
Hence, in a domesticated population under artificial se-
lection, differences in IBS relationships between individ- Genome
uals are largely explained by close relationships, rather The genome was assumed to consist of 20 chromosomes,
than by relationships due to common ancestors prior to each 100 cM long, with 1200 potential marker loci and 80
the known base population. For IBS-GS, it has also been potential QTL. Mutation rate for both markers and QTL
ØdegårdandMeuwissenGeneticsSelectionEvolution2014,46:3 Page3of8
http://www.gsejournal.org/content/46/1/3
was set to 3.0*10-5. To simplify the simulation and statis- Table1Descriptivestatisticsofthesimulationschemes
tical analysis, only recurrent allele mutations were allowed, Numberofchromosomes 20
andthereforeonlybi-allelicloci(similartosinglenucleotide
Lengthperchromosome(Morgan) 1.0
polymorphisms or SNPs) existed. Loci (both marker and
Basepopulation
QTL) with minor allele frequencies (MAF) below 5% (in
Numberofgenerations 5000
the founder population of the pedigree) were deleted. The
QTLeffectsweresampledfromagammadistributionwith Mutationrate,markers 3.0*10-5
shapeparameter0.4. Mutationrate,QTL 3.0*10-5
Effectivepopulationsize 500
Population
Evaluatedpopulation
A base population with effective population size Ne=
Numberofgenerations 3
500 was simulated for 5000 generations in order to
Genotypedmarkersperchromosome* ~40
achieve mutation-drift balance. At generation 5000,
~4000 marker loci (~200 per chromosome) and~240 to NumberofsegregatingQTLperchromosome* ~240-280
280 QTL segregated with a MAFabove 5%. From gener- Geneticvariance* 0.3
ation 5000, 100 males and 100 females were randomly Residualvariance 2.7or0.7
selected and each male was randomly mated with a sin-
Heritability* 0.1or0.3
gle female, resulting in 100 families, each consisting of
Traininganimalspergeneration 10,000
120 full-sibs (12 000 individuals per generation). The
Selectioncandidatespergeneration 2000
same population structure was used to simulate the
threesubsequentgenerations(5001to5003).Phenotypes Sirespergeneration 100
and genotypes (for a fraction of the animals) were stored Damspergeneration 100
only for generations 5001 to 5003, which were used in Familiespergeneration 100
the subsequent genetic analysis. No selection was ap-
Selectioncandidatesperfamily 20
plied within thesegenerations.
Traininganimalsperfamily 100
*Frombasepopulationgeneration5000(basepopulationinthestatistical
Datastructure
analyses).
Breeding values were defined as the sum of all QTL ef-
fects for each individual. The QTL effects were rescaled always retained) from the complete dataset, i.e., ~40
such that the total additive genetic variance (variance of markers/Morgan (M). These markers were then used to
breeding values summed over all QTL) was equal to trace genomic relationships within the known pedigree,
0.30 for all replicates (in generation 5001). Standardized as described in the next section (considering parents
random residuals for phenotypes were sampled independ- from generation 5000 as base animals). To mimic a situ-
entlyfromastandardnormaldistributionandsubsequently ation with selective genotyping, marker genotypes were
scaledbytheappropriateenvironmentalstandarddeviation. stored only from parents (generations 5001 and 5002)
Phenotypesweredefinedasthesumofthebreedingvalues and individuals of the phenotypically best (based on the
andthescaledrandomresiduals.Twoscenariosweresimu- average phenotype of the 100 training sibs) 30 of 100
lated,onewithheritabilityh2=0.10andonewithheritabil- families in the last generation (generation 5003). It was
ity h2=0.30. Datasets with different heritabilities were considered unlikely that candidates from the remaining
generated from the same simulation datasets by changing 70 families would be selected (even with marker data,
theresidualstandarddeviationusedforscalingofthestan- which was confirmed by later results). Furthermore,
dardized residuals. Hence, breeding values and genetic within the best 30 families, marker genotypes were
variance were identical in the parallel replicates across sce- stored for all selection candidates (20 per family) and for
narioswithdifferentheritabilities. varying fractions of the phenotyped training sibs (the
Of the 120 sibs in each family, 100 sibs were randomly top and bottom 5 to 20%, or all sibs). Thus, the total
selected as training animals (phenotyped and were not number of genotyped individuals ranged from 900 to
available for selection) and the remaining 20 were chosen 3600 out of 12 000 individuals (per generation). The dif-
as selection candidates (non-phenotyped validation ani- ferent genotypingstrategiesaredescribedinTable2.
mals). Descriptive statistics of the data sets are given in
Table1. Tracingrelationships
Markers were traced through the pedigree by linkage
Genotyping analysis using the linkage disequilibrium multi-locus it-
Genotypingwasperformedbysamplingevery5thmarker erative peeling (LDMIP) method [7] based on informa-
(the first and last marker of each chromosome were tion from all genotyped marker loci and the entire
ØdegårdandMeuwissenGeneticsSelectionEvolution2014,46:3 Page4of8
http://www.gsejournal.org/content/46/1/3
Table2Genotypingstrategiesforthedifferentbreeding IBD-based genomic relationship matrix, calculated as de-
programscenarioswithclassicalpedigree-basedselection scribed above. The true variance components were as-
(PED)andIBD-GS sumedknown.
PED IBD5 IBD10 IBD20 IBDall
Numberoffamilies 100 100 100 100 100
Validation
Selectioncandidatesperfamily 20 20 20 20 20 To evaluate the IBD-GS method, it was compared to
Numberofpre-selectedfamilies 0 30 30 30 30 classical pedigree-based selection methods. Animals were
Genotypedselectioncandidates 0 600 600 600 600 ranked based on their predicted breeding values from the
PED and IBD-GS models, respectively. Predicted breeding
Genotypedtraininganimals 0 5+5 10+10 20+20 100
perfamily values (either PED or IBD-GS) were used to select 200
Totalnumbergenotyped 0 900 1200 1800 3600 parents for the next generation. Genetic gain in the
first generation is expected to equal the average gen-
Numberofselectednewparents 200 200 200 200 200
etic level of the parents deviated from the population
Minimumnumberofparental 20 - - - -
averageinthepreviousgeneration.Hence,efficiencyof
families
classical pedigree-based selection vs. IBD-GS using se-
Onlygeneration5003isconsidered.
lectivegenotypingwasassessedbycomparingtheaver-
age true breeding values of the selected parents (from
pedigree (using parents from generation 5000 as base
generation5003) for thetwo models. Since none of the
population). The output was used to calculate IBD prob-
selection candidates had their own phenotype available
abilities at each genotyped marker locus [8] and the
(only sibs with data) classical selection based on PED
genome-wide IBD relationship matrix was produced by
implied family selection, i.e., all 200 selected individ-
averaging over all loci. The IBD matrix also included
uals would originate from only 10 families (with 20
IBD probabilities of non-genotyped animals, using infor-
candidates each) unless restrictions are imposed.
mation from relatives to estimate their genotypes [7].
Hence, for classical selection, the number of selected
Non-genotyped non-parents were consequently assumed
offspring was restricted to 10 per family, and future
to have standard pedigree relationships with all their
parents were thus randomly sampled from the 20
sibs, but not necessarily with more distant relatives
top-ranking families. GS allowed individual selection
(as their relationships are influenced by the relationships
also for non-phenotyped (albeit genotyped) animals. In
among genotyped ancestors). Animals not related through
practice, one would then select the best animals based
the pedigree were considered unrelated, regardless of their
on EBVacross families. The lowest-ranking sibs within
markergenotypes.
the best families are probably outperformed by the
highest-ranking sibs among the second-best families,
Statisticalmodels
which implies that future parents are selected from a
Two statistical models, pedigree-based (PED) and IBD-
wider range of familial backgrounds. Hence, for GS,
GS, were applied to all datasets and scenarios to esti-
there was no need to put any restrictions on selection.
mate BLUP (Best Linear Unbiased Prediction) breeding
Results showed that future parents were indeed se-
values with the DMU software package [9]. Both models
lectedfromnearlyall30pre-selectedfamilies.
hadthefollowinggeneral characteristics:
Selection based on the two models was compared
y ¼1μþZaþe based on the expected rate of inbreeding, which was ap-
proximatedas[10]:
where y is a vector of all phenotypes, μ is the overall
mean, a is a vector of additive genetic breedi(cid:2)ng val(cid:3)ues of Xk
all animals included in the pedigree, eeN 0;Iσ2e is a ΔF≈1 c2;
vector of random residuals, σ2e is the residual variance 2 i¼1 i
and Z is an appropriate incidence matrix. The two
models only differed in their distributional assumptions
where ci is the relative genetic contribution of parent i
forthe additive breeding values:
(appearing as parent for generation t) to generation t+1,
(cid:2) (cid:3)
PED:aeN 0;Aσ2a ; andk=200(totalnumberofparentspergeneration).
Thetwomodelswerealsoassessedforbiasbyregression
(cid:2) (cid:3)
IBD‐GS: aeN 0;GIBDσ2a ; of true breeding values on predicted breeding values. A
regression coefficient equal to 1 indicates proper scal-
whereσ2 istheadditivegeneticvariance,Aisthepedigree- ing, < 1 indicates inflation and>1 indicates deflation
a
based numerator relationship matrix, and GIBD is the ofthepredictions.
ØdegårdandMeuwissenGeneticsSelectionEvolution2014,46:3 Page5of8
http://www.gsejournal.org/content/46/1/3
Results frequently deviated more from the expected regression
IBD-GS led to a substantial increase in genetic gain coefficientthan theIBD-GS model.
over one generation, compared to classical selection
(Figure 1), especially with higher heritability. For a mod- Discussion
erately heritable trait (h2=0.3), the increase in genetic Geneticgain
gain with IBD-GS compared to classical pedigree-based Thestudyshows that for traitsthatare evaluatedonsibs
selection was substantial, with relative increases of 23, of selection candidates, substantial increases in genetic
28, 31 and 32% by genotyping the phenotypic most gain can be achieved through IBD-GS using sparse
extreme (top/bottom) 5, 10, 20%, and all training sibs, marker genotyping on small sub-samples of carefully se-
respectively. Hence, most of the extra gain with IBD-GS lected individuals, e.g., 900 to 1800 animals per gener-
was already obtained by genotyping the top and bottom ation from a population of 12 000 individuals. Using this
5% training sibs for the 30 pre-selected families. For a approach, only the best 30% of families were genotyped
low heritable trait (h2=0.1), the extra gain with IBD-GS (for both selection candidates and training animals). In
was lower (13 to 21%) but, again, most of the extra gain most cases, no candidates were selected from the lowest
was achieved with genotyping of the top and bottom 5% ranking preselected family. Hence, genotyping more
oftrainingsibs ofthepre-selected families. families would probably not change selection decisions
Although restrictions with respect to inbreeding were forthis populationstructure.
imposed only for the classical selection scheme, the un- Most of the potential advantage of IBD-GS relative to
restricted IBD-GS scheme had lower average rates of classical selection was already obtained when the 5% top
inbreeding (1.04 to 1.13% vs. 1.25%) and new parents and bottom training sibs from the pre-selected families
were recruited from more (26 to 29 vs. 20) of the 30 were genotyped. Furthermore, there was little practical
pre-selected families(Table3). difference in response to selection between genotyping
The average regression coefficients across replicates the 20% top and bottom vs. all training sibs. However,
(standard deviations) of true breeding values on pre- when analyzing real data, extreme observations may be
dicted breeding values are in Table 4. Across replicates, artifacts and selective genotyping strategies using small
the pedigree-based model had an average regression co- and highly selected samples are likely be more vulner-
efficient close to the expectation (1.0) and was thus un- able to such errors. Hence, careful quality control of the
biased. IBD-GS was also unbiased when all members of data is particularly important when combined with se-
preselected favorable families were genotyped (IBDall). lective genotyping. Genotyping larger fractions (i.e. 10 to
However, selective within-family genotyping of pheno- 20%) would increaserobustnessoftheIBD-GS analysis.
typically extreme sibs resulted in slight, albeit significant
bias at the highest heritability, in terms of inflated Populationstructureandfamilyproduction
variance of genomic EBV (Table 4). Nevertheless, when In this study, both models required pedigree information.
looking at single replicates, the pedigree-based model For aquaculture species for which tagging is impossible at
1.4
v.) 1.2
e
d
d. 1
n
a
st
n. 0.8
e Heritability = 0.1
g
n 0.6 Heritability = 0.3
G (i
a 0.4
elt
D
0.2
0
PED IBD5 IBD10 IBD20 IBDall
Figure1GeneticgaininGS-IBDandPEDselectionschemes.Geneticgainingeneticstandarddeviationunitsforclassicalpedigree-based
selectionschemes(PED)andGS-IBDusingsparseandselectivegenotypingfordifferentnumbersofgenotypedphenotypicallyextremesibs
(IBD#=genotypingthe#phenotypicallybestandworsttrainingsibsofeachfamily,IBDall=genotypingalltrainingsibs).
ØdegårdandMeuwissenGeneticsSelectionEvolution2014,46:3 Page6of8
http://www.gsejournal.org/content/46/1/3
Table3Averagenumberofcontributingfamiliesand Multi-traitselection
estimatedrateofinbreedingforclassicalpedigree-based This study considered selection on a single trait but real
selection(PED)andIBD-GS selection programs are often multi-trait and can include
Heritability 0.10 0.30 more than one trait measured on sibs of selection candi-
Numberof ΔF(%) Numberof ΔF(%) dates. In such cases, selection of the families to be geno-
families families typed would be based on a total merit index. In some
PED 20.0 1.25 20.0 1.25 cases, traits measured on sibs are recorded on different
IBD5 26.1 1.13 28.6 1.07 subsets of training sibs. This would still allow the pro-
posed strategy for within-family selective genotyping to
IBD10 27.0 1.12 28.9 1.06
be used for each trait separately, but would require
IBD20 27.4 1.11 29.2 1.04
genotyping more animals. If multiple traits are recorded
IBDall 27.6 1.11 29.1 1.05
on the same group of training sibs, sibs for genotyping
could be selected based on their total merit index.
young ages, it may be necessary to separate families until These animals are not necessarily the most phenotypic
they reach tagging size, which increases the costs of family extremes for each trait but they are expected to be the
production and creates physical limitations in the number most extreme with respect to the aggregated genotype,
offamiliesthatcanbeproduced,andpotentiallyintroduces which is what we want to predict and genetically
commonenvironmentalfamily(tank)effects.Still,genome- improve.
wideSNPmarkers(asinIBD-GS)canbeusedtotracepar-
entage of individuals with high accuracy, even with com- Biasbyselectivegenotyping
munal rearing e.g. [11,12]. However, when within-family When genotyping is restricted to phenotypically extreme
selective genotyping is to be used, prior knowledge of the sibs, estimates of genetic variation may be inflated,
pedigree is required, i.e., favoring separate rearing of fam- resulting in biased EBV. However, results from a previ-
ilies.Fortunately,separaterearingisalreadycommonprac- ous study indicate that when non-genotyped sibs are in-
tice in many existing aquaculture selection programs for cluded in the analysis through pedigree information (as
traitsthatareevaluatedonsibsofselectioncandidates(e.g., in the current study), accurate estimates of genetic vari-
in Salmonidae), and thus the structure necessary to apply ance canstillbeobtainedforlargerselectivelygenotyped
selective genotyping exists. In contrast, for mass-spawning samples [13]. In the current study, the true genetic vari-
species, for which artificial stripping is difficult to apply, e. ance was assumed known (i.e., BLUP), which reduces
g., gilthead seabream (Sparus aurata), communal rearing the risk of inflated variation of EBV. Still, a statistically
may be the only practical alternative. Furthermore, mass- significant but small bias in terms of inflated variance of
spawningspeciestypicallyhaveratherchaoticfamilystruc- EBV was detected with IBD-GS, but the limited magni-
tures (many small families, and uneven contributions of tude of the inflation indicates that the bias would not be
parents), which makes within-family selective genotyping a problem in practice. The bias detected was of similar
difficultto apply,evenwhenfamilybackgroundcanbede- magnitude as previously reported for other GS ap-
duced (e.g., through an initial parentage assignment based proaches e.g. [14,15]. Furthermore, the pedigree-based
on microsatellites). Alternatively, parentage testing and model (which was unbiased when averaged over repli-
sparse genotyping for IBD-GS may be combined into one cates) showed larger between-replicate variation. Hence,
genotyping test but then additional saving from selective for real data, pedigree-based evaluation may actually
genotypingisnotbeingused. give larger scaling errors than IBD-GS using selective
genotyping.
Table4Average(β)andstandarddeviation(SD)across
replicatesofregressioncoefficientsoftruebreeding Inbreeding
valuesonpredictedbreedingvaluesforclassical Although there were no restrictions on inbreeding and
pedigree-basedselection(PED)andIBD-GS geneticgain wasfaster, inbreedingrates were systematic-
Heritability 0.10 0.30 ally lower with IBD-GS than with pedigree-based selec-
β SD β SD tion schemes. Here, the expected rate of inbreeding was
estimated over only one generation of selection. In the
PED 0.963 0.198 0.991 0.126
longer term, differences between the two methods are
IBD5 0.964 0.155 0.961* 0.088
expected to be more substantial and in favor of IBD-GS
IBD10 0.965* 0.145 0.947*** 0.085
because pedigree-based selection for traits measured on
IBD20 0.971 0.145 0.948*** 0.080 the sibs will favor co-selection of relatives. Furthermore,
IBDall 0.993 0.146 1.020 0.086 imposed recruitment of parents of less favorable EBV, as
*P<0.05forβ<1;**P<0.01forβ<1;***P<0.001forβ<1. was done for the pedigree-based program, can reduce
ØdegårdandMeuwissenGeneticsSelectionEvolution2014,46:3 Page7of8
http://www.gsejournal.org/content/46/1/3
inbreedingin the short term but the long term contribu- explained by the additive QTL effects, (3) all alleles
tion of such parents may be low. In the IBD-GS scheme originate from a single base population, which implies
parents were solely selected based on EBV, which is ex- consistent population-wide LD (no segmentation of the
pected to give more even long-term genetic contribu- population),and(4) marker andQTL allelesfollow simi-
tions of the selected parents compared to a classical lar frequency distributions. In practice, these assump-
selection scheme imposing selection of genetically infer- tions do not necessarily apply: many populations have
iorparents. mixed origins, QTL and markers do not necessarily fol-
low the same distributions (e.g., QTL loci are often
Geneticvariability subjected to selection) and non-additive (dominance
It has been stated that GS is particularly useful for traits and epistatic) genetic effects exist. Through sexual
with low heritability [16], which is indeed likely to be reproduction and recombination, non-additive effects
true for traits that are recorded directly on selection are often not passed on to future generations, which
candidates as well as on sibs. With classical selection reduces their relevance for selective breeding. How-
for traits of high heritability, individual phenotypes will ever, epistatic effects between linked loci are expected
compensate for the lack of precision achieved with sib tobemoresustainedovergenerationsandevidencefor
data. For a trait measured only on sibs, the extra genetic epistatic interactions between linked loci has been re-
gain from IBD-GS over pedigree-based selection is more ported for different species [18,19]. Within families,
likely to increase with increasing heritability. This is be- haplotypes of linked loci are likely inherited as one
cause, with pedigree-based selection, the EBV of all se- block, i.e., as if they were a single locus [20]. Similarly,
lection candidates is set equal to their mid-parent EBV, transgenerational epigenetic inheritance [21], such as
regardlessof heritability,whileIBD-GSallowsestimation chromatin marking (including methylation), can also
of Mendelian deviations from this mean, even without contribute to inherited differences and thus contribute
individual phenotypes. In the current study, mid-parent to the heritable (“additive”) variation, i.e., the variance
means are expected to be accurately estimated in both available for selective breeding. Such heritable factors,
models, even for traits with low heritability because of may partly explain the so-called “missing heritability
thelargenumberofphenotypedoffspring(100perfamily), problem” that is typically observed with IBS relationships
while accuracies of the estimated Mendelian sampling [22]. These effects of local epistasis and transgenerational
deviations from these means (predicted in IBD-GS) are epigenetics (chromatin marking) may, however, be effect-
expected to be greater with high heritability. As a conse- ively captured by genomic IBD relationships. Thus, the
quence,theextragainfromIBD-GSincreaseswithincreas- IBD-GS model is expected to take such effects into
ing heritability of the traits evaluated on sibs. The greater account.
benefit of IBD-GS for traits with higher heritability is One of the main advantages of the IBD-GS method is
evident from our study; the classical pedigree-based model that it can effectively use sparse marker genotypes and
showedsimilargeneticgains(6%different)fortraitswitha selective genotyping. In contrast, IBS-GS methods are
lowormoderateheritability,whileforIBD-GS,geneticgain intended for dense marker data [1] and sparse genotyp-
wassubstantiallygreater(15to16%)fortraitswithmoder- ing would require a sparse-to-dense imputation step
ateversuslowheritability. (i.e., dense marker data must be available for preceding
This study considered genetic gain over only one generations). Furthermore, IBD-GS is well suited for the
generation,startingfromapreviouslyunselectedpopula- jointanalysisofdatafromnon-genotypedandgenotyped
tion. In practice, selection is usually continued over sev- individuals (e.g., as in selective genotyping), which is
eral generations,and the Bulmer effect would (aslong as likely to introduce bias in IBS-based analyses. The IBD-
selection continues) reduce between-family variation in GS method is thus easier to implement for analysis of
the population [17]. As a consequence, the ability to use selectively genotyped and sparse marker data and of
within-family (Mendelian) genetic variation will be more phenotypic data that do not necessarily fit the idealized
important for future generations (and more so if restric- conditions that are typically assumed in stochastic simu-
tions are imposed on inbreeding). Thus, because clas- lationstudies.
sical sib-selection does not use within-family variation
the relative advantage of GS is expected to increase if Conclusions
selectivebreedingisapplied overmultiple generations. The IBD-GS was substantially more effective than clas-
sical selection, even when based on very few markers,
Sensitivitytoassumptions and when combined with selective genotyping of small
Genomic selection simulation studies are typically based proportions of the population, i.e., 900 to 1800 out of 12
on several idealized assumptions, including: (1) QTL ef- 000 individuals. The study shows that a low-cost GS
fects are strictly additive; (2) all heritable variation is program can be successfully performed by combining
ØdegårdandMeuwissenGeneticsSelectionEvolution2014,46:3 Page8of8
http://www.gsejournal.org/content/46/1/3
sparse and selective genotyping with pedigree and linkage 15. LuanT,WoolliamsJA,LienS,KentM,SvendsenM,MeuwissenTHE:The
information.Thus,theIBD-GSmethodissuitableforcost- AccuracyofGenomicSelectioninNorwegianRedCattleAssessedby
Cross-Validation.Genetics2009,183:1119–1126.
effective genomic evaluation using sparse markers, for se-
16. CalusMPL,MeuwissenTHE,deRoosAPW,VeerkampRF:Accuracyof
lective genotyping of training animals, and for phenotypic genomicselectionusingdifferentmethodstodefinehaplotypes.
datathatdonotnecessarilyfittheidealizedconditionsthat Genetics2008,178:553–561.
17. BulmerMG:Theeffectofselectionongeneticvariability.AmNat1971,
aretypicallyassumedinstochasticsimulationstudies. 105:201–211.
18. GaertnerBE,ParmenterMD,RockmanMV,KruglyakL,PhillipsPC:More
Competinginterests thanthesumofitsparts:acomplexepistaticnetworkunderliesnatural
Theauthorsdeclarethattheyhavenocompetinginterests. variationinthermalpreferencebehaviorinCaenorhabditiselegans.
Genetics2012,192:1533–1542.
19. DuthieC,SimmG,Doeschl-WilsonA,KalmE,KnapPW,RoeheR:Epistatic
Authors'contributions
analysisofcarcasscharacteristicsinpigsrevealsgenomicinteractions
JØhadtheoriginalidea,performedthesimulationsandstatisticalanalyses
betweenquantitativetraitlociattributabletoadditiveanddominance
andwasmainresponsibleforwritingthemanuscriptandTHEMwas geneticeffects.JAnimSci2010,88:2219–2234.
responsiblefordevelopingthelinkageanalysismethodology(LDMIP
20. HaigD:DoesheritabilityhideinepistasisbetweenlinkedSNPs?EurJ
software).Allauthorsreadandapprovedthefinalmanuscript. HumGenet2011,19:123–123.
21. JablonkaE,RazG:Transgenerationalepigeneticinheritance:prevalence,
Acknowledgements mechanisms,andimplicationsforthestudyofheredityandevolution.
ThisstudywaspartlyfinancedbytheResearchCouncilofNorwaythrough QRevBiol2009,84:131–176.
projectsno.186862and225181.Dr.NinaSanti,AquaGenisacknowledged 22. VisscherPM,MedlandSE,FerreiraMAR,MorleyKI,ZhuG,CornesBK,
forvaluablescientificinputtothestudy.Twoanonymousreviewersare MontgomeryGW,MartinNG:Assumption-freeestimationofheritability
gratefullyacknowledgedfortheirvaluablecomments. fromgenome-wideidentity-by-descentsharingbetweenfullsiblings.
PLoSGenet2006,2:e41.
Authordetails
1AquaGenAS,P.O.Box1240,Sluppen,NO-7462Trondheim,Norway. doi:10.1186/1297-9686-46-3
2DepartmentofAnimalandAquaculturalSciences,NorwegianUniversityof Citethisarticleas:ØdegårdandMeuwissen:Identity-by-descent
LifeSciences,P.O.Box5003,NO-1432Ås,Norway. genomicselectionusingselectiveandsparsegenotyping.Genetics
SelectionEvolution201446:3.
Received:15May2013Accepted:3December2013
Published:20January2014
References
1. MeuwissenTHE,HayesBJ,GoddardME:Predictionoftotalgeneticvalue
usinggenome-widedensemarkermaps.Genetics2001,157:1819–1829.
2. LuanT,WooliamsJA,ØdegårdJ,DolezalM,Roman-PonzeSI,BagnatoA,
MeuwissenTHE:Theimportanceofidentity-by-stateinformationforthe
accuracyofgenomicselection.GenetSelEvol2012,44:28.
3. HabierD,FernandoRL,DekkersJCM:Genomicselectionusinglow-density
markerpanels.Genetics2009,182:343–353.
4. WientjesYCJ,VeerkampRF,CalusMPL:Theeffectoflinkage
disequilibriumandfamilyrelationshipsonthereliabilityofgenomic
prediction.Genetics2013,193:621–631.
5. HabierD,FernandoRL,DekkersJCM:Theimpactofgeneticrelationship
informationongenome-assistedbreedingvalues.Genetics2007,
177:2389–2397.
6. SargolzaeiM,SchenkelFS:QMSim:alarge-scalegenomesimulatorfor
livestock.Bioinformatics2009,25:680–681.
7. MeuwissenTHE,GoddardME:Theuseoffamilyrelationshipsandlinkage
disequilibriumtoimputephaseandmissinggenotypesinuptowhole-
genomesequencedensitygenotypicdata.Genetics2010,185:1441–1449.
8. FernandoR,GrossmanM:Markerassistedselectionusingbestlinear
unbiasedprediction.GenetSelEvol1989,21:467–477.
9. MadsenP,JensenJ:DMU:auser'sguide.Apackageforanalysingmultivariate
mixedmodels.Version6,release5.0.Tjele,Denmark:UniversityofAarhus,Faculty
ofAgriculturalSciences(DJF),DeptofGeneticsandBiotecnology;2010.
10. RobertsonA:Inbreedinginartificialselectionprogrammes.GenetRes Submit your next manuscript to BioMed Central
1961,2:189–194. and take full advantage of:
11. HauserL,BairdM,HilbornR,SeebLW,SeebJE:Anempiricalcomparison
ofSNPsandmicrosatellitesforparentageandkinshipassignmentina
• Convenient online submission
wildsockeyesalmon(Oncorhynchusnerka)population.MolEcolResour
2011,11(Suppl1):150–161. • Thorough peer review
12. HayesB,SonessonAK,GjerdeB:EvaluationofthreestrategiesusingDNA • No space constraints or color figure charges
markersfortraceabilityinaquaculturespecies.Aquaculture2005,
250:70–81. • Immediate publication on acceptance
13. ØdegårdJ,MeuwissenT:Estimationofheritabilityfromlimitedfamily • Inclusion in PubMed, CAS, Scopus and Google Scholar
datausinggenome-wideidentity-by-descentsharing.GenetSelEvol2012,
• Research which is freely available for redistribution
44:16.
14. VitezicaZG,AguilarI,MisztalI,LegarraA:Biasingenomicpredictionsfor
populationsunderselection.GenetRes2011,93:357–366. Submit your manuscript at
www.biomedcentral.com/submit