1Christian-Albrechts-Universityof KielInstitute of Animal Breeding and HusbandryDavid HabierNapapan PyiasatianJack DekkersRohan FernandoHabier et al. 2009 Genetics 182: 343 - 353Genomic Selection using Low-Density SNPsFaculty of Agriculture and Nutritional ScienceEstimate Estimate marker marker effectseffectsGenotypeGenotypefor >50,000 for >50,000 SNPsSNPsPhenotypePhenotypeGenotypeGenotypefor >50,000for >50,000SNPsSNPsPredict BV Predict BV from marker from marker genotypes at genotypes at early ageearly ageGenotypeGenotypefor >50,000 for >50,000 SNPsSNPsTraining dataTraining dataTraining dataPredict BV Predict BV from marker from marker genotypes at genotypes at early ageearly ageGenomic selectionGenetic Evaluation using high-density SNPsMeuwissen et al. 2001MeuwissenMeuwissenet al. 2001et al. 20010.60.70.80.91345678GenerationAccuracy23Need Low- (<380) vs. High-density panel for routine implementation?? $50 vs. $250 per animal ??‘Standard’ approach to developing Low-density panels:• Select the ‘best’ SNPs from the HD-panel• Trait and population specificProposed approach: use well-spaced Low-density SNP genotypes onselection candidates to ‘fill in’ missing HD SNP genotypesIntroductionImplementation of GSOriginal principle of Genomic Selection (GS)High-density (HD) SNP genotypes used for both• Estimation of marker effects (training)• Prediction of GS-EBV for selection candidatesNot feasible for many species4Outline  Introduction – What is ELD-GS? Methods Published results Unpublished results Criteria for loss of accuracy Factors affecting loss of accuracy of ELD-GS Precision of PDMs Simulations – Results  Conclusions & outlook35ProgenyiProgenyiSire sSire sDam dDam dConcept of Low-DensityGenomic SelectionpaternalmaternalpaternalmaternalpaternalmaternalHD-GS Î EBVi= Σ (gmik+ gpik)Sum estimates of effectsof maternal and patpaternernalalSNP allelesSNP kSNP kSNP kSNP kSNP kSNP kLD-GS Î EBVi= Σ (pmdgm+ ppdgp+ ppmsmsggmm+ ppsgp)ik dk ik dkikiksksk ik skProbability that i received damProbability that i received dam’’s maternal allele at SNP ks maternal allele at SNP kLDLD--GSGSLDLD--GSGS6Steps of proposed low-density genomic selection method:1. Estimate marker allele effects of HD-SNPs – Bayes-B2. Infer HD-SNP haplotypes of training individuals• Requires parental HD-SNP genotypes3. Trace HD-SNP alleles of selection candidatesbased on their LowD-SNP genotypes• Probability of descent of marker alleles4. Predict GS-EBV of selection candidates• Weighted sum of effects of parental HD-SNP allelesMethodsProgenyiProgenyipaternalmaternalLDLD--GSGSLDLD--GSGS47I. Estimation of HD-SNP effectsGeneral statistical model:kkkkμβδ=++∑y1x exk= # “1” alleles carried at SNP kβk= substitution effect of SNP kδk= indicator variable for SNP k to be in (=1) or out (=0) of the modelBayesB is used here, but other methods modeling disequilibrium and co-segregation, dominance or epistasis can be used also.8II. Infer HD-SNP haplotypesParent iIn the training generation, haplotypes must be inferred for males and females,ikmkpixx= maternal and paternal allele states of individual i at SNP k59III. Track HD-SNP alleles Parent iProgenyProbability of Descent of Marker alleles (PDMs)Genotyped for evenly-spaced LD-SNPspikpmikp10Estimation of PDMs MCMC sampling: Joint probabilities of sampled allele origins for adjacent ELD-SNP pairs were estimated Information from all ELD-SNPs is utilized Haplotype phases of HD-genotyped ancestors assumed known611()ˆˆlociHD kklocikpmkkkxGExBV Xbb==+∑∑IV. Prediction of GEBVs ELD-SNP genotyped offspring: HD genotyped parents:()ˆˆˆlmkociELD kkpkG xEB bxV =+∑ˆ*pppkkkxpx=ˆ*mmmkkkxpx=ˆˆ*pppkkkxpx=ˆˆ*mmmkkkxpx=Generation after training:Later generations:12Tested by SimulationPopulationGS-EBV using HighD SNPsGS-EBV using LowD SNPs10 males x 100 femalesGeneration 4-7Training data(N=1000)Generation 4Pedigree recording and genotyping starts50 males x 500 females (N=1000)Generation 1-3Population Growth(N=100 to N=1000)Generation -10Random Mating(Ne=100)Generation -60Random Mating(Ne=500)Generation -1060Genome10 chromosomes of 1 M20,000 SNPs ; 500 QTL1,000 SNPs selectedafter 1060 gener.HD SNP spacing ~ 1 cMLD SNPs at 10 or 20 cMTrait h2 = 0.5Bayes-B(Meuwissen et al. ‘01)713Results0.30.40.50.60.70.84 4.5 5 5.5 6 6.5 7GenerationAccuracyHDELD-10ELD-20TrainingELD-10+ELD-20+Accuracy of GS-EBV based on High- and Low-Density SNP genotyping (20 Replicates) 500 QTL (~220 MAF>0.01)140510152025% loss in accuracyELD-10 ELD-10+ BB-110 BB-40 FSS-0.01 FSS-110100 QTL2 3 4Generation0510152025% loss in accuracyELD-10 ELD-10+ BB-110 BB-40 FSS-0.01 FSS-110500 QTL2 3 4Generation0510152025% loss in accuracyELD-10 ELD-10+ BB-110 BB-40 FSS-0.01 FSS-1101000 QTL2 3 4Generation815Genomic Selection can beimplemented with low-density SNPgenotyping of selection candidates• Loss in accuracy limited: < 3.5% after 1 generation< 8 % after 2 generationswith 300 equally spaced SNPs(10 cM)• Loss in accuracy ~ independent of # QTL and # traits• Lower rate of fixation of panel SNPs with selection Æ slower accuracy decline• Cost effectiveness needs to be analyzed• Depends on costs of Low- vs. High-density genotyping$40 Å??Æ $180• Optimal implementation needs to be further analyzed• Which individuals to genotype – HD / LDDiscussion & Conclusions00.10.20.30.40.50.60.70.84567GenerationAccuracy16Outline  Introduction – What is ELD-GS? Methods Published results Unpublished results Criteria for loss of accuracy Factors affecting loss of accuracy of ELD-GS Precision of PDMs Simulations – Results  Conclusions & outlook917Objectives of recent work Analyze factors affecting loss of accuracy with ELD-GS  Type and extent of LD  Precision of PDMs Analyze loss of accuracy under more realistic assumptions LD based on a real pedigreeFunding from Aviagen18Criteria for loss of accuracy Accuracy of GEBVHDand GEBVELD Uncertainty in tracking HD-SNP alleles Assumption: Only precision of PDMs affects loss of accuracy Æ Correlation between GEBVHDand GEBVELD (lower bound)()ˆˆlociHD kklocikpmkkkxGExBV Xbb==+∑∑()ˆˆˆlmkociELD kkpkG xEB bxV =+∑1ˆ=b1019Factors affecting accuracy from ELD-GS Precision of PDMs HD-genotyping of parents (see previous) ELD-SNP spacing  Family structure20Simulations – Genome structure 8 chromosomes of