March 2012 AIPL Update - PowerPoint Presentation

March 2012 AIPL Update

Topics Genomics overview April 2012 changes Accounting for bias Selection index adjustments Cow adjustments Calving traits update

Whole-genome selection Use many markers to track inheritance of chromosomal segments Estimate the impact of each segment on each trait Combine estimates with traditional evaluations to produce genomic evaluations ( GPTA ) Select animals shortly after birth using GPTA Very successful worldwide

Illumina genotyping arrays BovineSNP50 54,001 SNPs (version 1) 54,609 SNPs (version 2) 45,187 SNPs used in evaluation BovineHD 777,962 SNPsOnly BovineSNP50 SNPs used >1,700 SNPs in databaseBovineLD6,909 SNPsAllows for additional SNPs BovineSNP50 v2 BovineLD BovineHD

Reliabilities for young Holsteins * *Animals with no traditional PTA in April 2011 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 40 45 50 55 60 65 70 75 80 Reliability for PTA protein (%) Number of animals 3K genotypes 50K genotypes

Date SNP Estimation* Young animals** All animals Bulls Cows  Bulls Heifers   04-10 9,770 7,415 16,007    8,630 41,822 08-10 10,430 9,372 18,652 11,021 49,475 12-10 11,293 12,825 21,161 18,336 63,615 04-11 12,152 11,224 25,202 36,545 85,123 08-11 16,519 14,380 29,090 52,053 112,042 09-11 16,812 14,415 30,185 56,559 117,971 10-11 16,832 14,573 31,865 61,045 124,315 11-11 16,834 14,716 32,975 65,330 129,855 12-11 17,288 17,236 33,861 68,051 136,436 01-12 17,681 17,418 35,404 74,072 144,575 02-12 17,710 17,679 36,597 80,845 152,831 *Traditional evaluation **No traditional evaluation Genotyped Holsteins

What’s a SNP genotype worth? For the protein yield (h 2 =0.30), the SNP genotype provides information equivalent to an additional 34 daughters Pedigree is equivalent to information on about 7 daughters

And for daughter pregnancy rate (h 2 =0.04), SNP = 131 daughters What’s a SNP genotype worth?

High density update HD tests before and after GBR, ITA 342 HD animals with 636,967 SNPs 1,074 HD animals with 636K or 311K 1,510 HD animals with 311,725 SNPsREL gain above 50K for the 3 tests-0.5% decrease in REL using 342 HD+0.4% increase using 1,074 HD+1.0% increase using 1,510 HD

Trait a Bias b b REL (%) REL gain (%) Milk (kg) −64.3 0.92 67.1 28.6 Fat (kg) −2.7 0.91 69.8 31.3 Protein (kg) 0.7 0.85 61.5 23.0 Fat (%) 0.0 1.00 86.5 48.0 Protein (%) 0.0 0.90 79.0 40.4 PL (months) −1.8 0.98 53.0 21.8 SCS 0.0 0.88 61.2 27.0 DPR (%) 0.0 0.92 51.2 21.7 Sire CE 0.8 0.73 31.0 10.4 Daughter CE −1.1 0.81 38.4 19.9 Sire SB 1.5 0.92 21.8 3.7 Daughter SB − 0.2 0.83 30.3 13.2 a PL = productive life, CE = calving ease and SB = stillbirth. b 2011 deregressed value – 2007 genomic evaluation. Holstein prediction accuracy

April 2012 changes Genotypes from GGP included Revised weights used to combine information for genotyped animals Reduces evaluations of top young animals compared with older animals Reliabilities changed only slightly, but would decrease if DGV weights were reduced more Some regressions lower than expected, and the revised weights helped bring those into compliance with validation tests

April 2012 changes, cont’d Reliabilities revised to agree more precisely with observed reliabilities from truncated data Published reliabilities for young Holstein animals were adjusted: -3 percentage points for yield traits +3 to 6 percentage points for fitness traits-7 to 10 percentage points for calving traits+1 percentage point for type traits

April 2012 changes, cont’d Traditional PL evaluations for females less than 48 months of age not used in the genomic evaluation Eliminated large differences between genomic and traditional PL for some bulls Traditional DPR evaluations for genotyped cows less than 36 months of age now excluded from genomic evaluationSimilar edit for CCR

Some sources of bias Pre-selection bias Affects domestic and international evaluations Preferential treatment of bull dams Results in inflated PTA

Expected value of Mendelian sampling no longer equal to 0 Key assumption of animal models References: Patry, Ducrocq 2011 GSE 43:30 Vitezica et al 2011 Genet Res (Camb) pp. 1–10. Bias from pre-selection

Bulls born in 2008, progeny tested in 2009, with daughter records in 2012, were pre-selected: 3,434 genotyped vs. 1,096 sampled Now >10 genotyped per 1 marketed Potential for bias: 178 genotyped progeny 32 sons progeny tested Pre-selection bias now beginning

1-Step to incorporate genotypes Flexible models, many recent studies Foreign data not yet included Multi-step GEBV, then insert in AM Same trait (Ducrocq and Liu, 2009)Or correlated trait (Mantysaari and Stranden, 2010; Stoop et al, 2011)Foreign genotyped bulls includedNational methods to reduce bias

Multi-step genomic methods Direct genomic value (DGV) Sum of effects for 45,187 genetic markers Does not include polygenic effect (USA) Does include the polygenic effect (CAN, others) Model: y = Xb + Zg + poly + eCombined genomic evaluation (GPTA)Include phenotypes not used in estimating DGVSelection index includes 3 terms per animal: (DGV + poly), traditional PTA, and subset PTAGPTA = w1 (DGV + poly) + w2 PTA + w3 SPTA

Combined GPTA GPTA = w 1 (DGV + poly) + w 2PTA + w3SPTA(DGV + poly) = contribution from SNP effectsPTA = contribution from the traditional evaluationSPTA = subset PTA estimated using pedigree relationships among the genotyped animalsTerms combined using theoretical weights based on reliabilitiesWeights average 0.99 for DGV, 0.12 for EBV, and -0.11 for SBV

Selection index examples Dam not genotyped, low GREL GPTA = .99 (DGV+poly) + .41 PTA - .40 SPTADam not genotyped, high GRELGPTA = .99 (DGV+poly) + .11 PTA - .10 SPTADam is genotypedGPTA = 1.00 (DGV+poly) + .00 PTA - .00 SPTA

Proposal: Shift weight from DGV to SPTA Dam not genotyped, low GREL GPTA = .90 (DGV+poly) + .41 PTA - .31 SPTADam not genotyped, high GRELGPTA = .90 (DGV+poly) + .11 PTA - .01 SPTADam is genotyped GPTA = .90 (DGV+poly) + .10 PTA - .00 SPTA

Results of shifting DGV weight Similar to adding more polygenic variance but easier computation Some genomic REL higher with .90 weight, but lower if < .80 weightRegressions of future on past data higher if DGV weight lowerHighest animals have lower GPTAs with .90 weight

Convert and exchange DYD g National GEBV and DYD g unbiased Can’t deregress GEBV without G Exchange similar to simple GMACEOther countries need DYD anywayDeregress, reregress EBVs in MACECountries deregress MACE EBVAvoid bias by exchanging DYDg International bias reduction

6,743 bulls with no USA daughters Corr (National EBV, MACE EBV) .77 before adding foreign data .995 after adding foreign data Few foreign bulls in JE reference population, so hard to test gain in REL of young bull GEBV Foreign data in 1-Step: results

Evaluation Regression Squared Correlation Parent Average .73 .436 Multi-Step GEBV .75 .520 1-Step GEBV .85 .520 Expected .93 1-Step vs multi-step Data cutoff in August 2008

Holstein convergence much slower JE took 11 sec / round including G HO took 1.6 min / round including GJE needed ~1000 roundsHO needed >5000 roundsAll-breed model without genomicsReplace software used since 1989Correlations >.995 with traditional AMPreliminary larger analyses

Bias in cow evaluations Top cows over-evaluated compared to top bulls Parent averages over-estimate eventual evaluations of bulls Unreasonable estimates of SNP effects in PAR reflect sex effect Adjustment of evaluations of genotyped cows implemented April 2010 Adjustment made genotyped cows not comparable to non genotyped cows

Genomic evaluation Deregressed traditional evaluations used for estimation of SNP effects Predictor population consists of animals with both genotypes and traditional evaluations Cows can be predictors Increases size of predictor population Requires that cow and bull evaluations be comparable

Adjustment of cow evaluations US industry requested adjustment of all cow evaluations to restore comparability Desirable to leave estimates of genetic trend unchanged Variability of cow evaluations to be reduced Industry requested proposal in February for possible implementation in April 2011 Industry partners collaborated in developing and distributing information on the new adjustment

Method Adjustment for Milk, Fat, and Protein only Mendelian Sampling (MS) = PTA - PA Deregressed Value = MS/R DE cow = DEtot – DEpaR = DEcow/(DEtot + k)SD of Deregressed Values of cows and bulls comparedAdjVar = SDbull/SDcow Varies with reliability

Mean adjustment Calculate mean PA by birth year Adj Mean = factor*(PA – PA mean )HO factor = -0.434Devadj = AdjVar*Deregressed Value + AdjMean+ PAHO AdjVar = 0.3165 + 1.433 * RcowRcow = DEcow /(DEcow + k)PTAadj = R*Dev adj + (1-R)*PAnew PAnew includes PTAadj of dam

PTA milk for cows born in 2005 ADJ No ADJ Difference - 1500 - 1000 - 500 0 500 1000 1500 -1000 -760 -520 -280 -40 200 440 680 920 PA PTA

Effect of reliability -400 -300 -200 -100 0 100 200 300 400 500 600 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 PTA Reliability ADJ No ADJ Difference

Additional adjustment for genomics DGV – direct genomic value, sum of SNP effects Further adjust so Mean PTA = Mean DGV All cow adjustment not able to remove bias in cows selected to be genotyped Reduce adjusted PTA by following amounts Holstein Jersey & Brown Swiss Milk (lb) 169.7 165.9 Fat (lb) 8.3 6.4 Protein (lb) 4.2 5.8

Possible further applications Additional breeds Ayrshire, Guernsey, Milking Shorthorn Additional traits Large SNP effects on PAR observed in type and functional traits Limitation with One-Step method Adjustment between traditional and genomic evaluations not possible

Cow adjustment summary Improved adjustment of cow evaluations for use in genomic evaluations Accommodation of different population being genotyped with 3K chip Improved comparability of evaluations of genotyped and non genotyped cows Foreign cow evaluations not used in estimation of SNP effects

Calving traits topics Proposed changes Multiple-trait evaluation Interbull trend validation Future research

Calving traits Sire-maternal grandsire threshold model y = HY + YS + PS + Ys + Ym + s + m + e All parities combined into a single trait Low heritabilities, 2 to 8%

Interbull trend validation Ensures evaluation results are in line with expectations Must pass every two years The US was failing the method 3 trend test If we don’t pass, we get kicked out

How do we fix it? Wiggans et al. (2007) tried multiple-trait linear models Poor correlations w/other countries Did not implement New approach – first and later parities evaluated separately and blended into a single PTA Method may be too simple

Results Bulls ranked similarly Reliabilities for bulls with little data decreased Good correlations with MACE for high-reliability bulls Poor correlations with test run results

Decision We don’t have a good explanation for the drop in correlations May be way reliabilities are blended Errors found in trend-testing code Both models now pass validation Continue with current model until we figure our correlation issue

Ongoing research Is there a link between use of sexed semen and stillbirths? What about effects of age at first calf on calving traits? What’s wrong with our correlations?

Overall conclusions No changes made to calving traits evaluations All data are important We are working on biases in both cow and bull evaluations 1-step looks promising if we can get the calculations done

Questions?