Reliability of Genomic Predictions for North American Dairy Bulls

2007

Paul VanRadenPaul VanRaden11, Curt Van Tassell, Curt Van Tassell22, George Wiggans, George Wiggans11, Tad , Tad SonstegardSonstegard22, Jeff O’Connell, Jeff O’Connell11, Bob Schnabel, Bob Schnabel33, Jerry Taylor, Jerry Taylor33, and , and Flavio SchenkelFlavio Schenkel44, , 1Animal Improvement Programs Lab and 2Bovine Functional Genomics Lab, USDA Agricultural Research Service, Beltsville, MD, USA, 3U. Missouri, Columbia, 4U. Guelph, ON, [email protected]

2008

Reliability of Genomic Predictions Reliability of Genomic Predictions for North American Dairy Bullsfor North American Dairy Bulls

FASS annual meeting, July 2008 (2) Paul VanRaden200

8

Sequencing and GenotypingSequencing and Genotyping

Cattle genome sequenced in 2004• 30 chromosome pairs (including X,Y)• 3 billion letters from each parent

Illumina BovineSNP50 BeadChip• 58,000 genetic markers in 2007• 38,416 used in genomic predictions• Current cost < $250 per animal


8

Experimental DesignExperimental Design

Compute genomic evaluations and parent averages from 2003 data• 3576 older Holstein bulls born 1952-1998

Compare ability to predict daughter deviations in 2008 data• 1759 younger bulls born 1999-2002

Test results for 27 traits: 5 yield, 5 health, 16 conformation, and Net Merit


8

Genotyped Animals (n=6005)Genotyped Animals (n=6005)As of April 2008As of April 2008

0

200

400

600

800

100019

50

1970

1990

1992

1994

1996

1998

2000

2002

2004

2006

2008

Year of Birth

Nu

mb

er o

f A

nim

als

Predictor

Predictee

Young


8

Genomic MethodsGenomic Methods

Direct genomic evaluation• Inversion for linear prediction, REL• Iteration for nonlinear prediction

Combined genomic evaluation• 3 x 3 selection index combining

direct genomic PTA, traditional PA or PTA, and subset PA or PTA by REL


8

Marker Effects for Net MeritMarker Effects for Net Merit


8

Significance Tests for Net MeritSignificance Tests for Net Merit


8

Major Gene on Chromosome 18Major Gene on Chromosome 18Net Merit, Productive Life, Calving Ease, Stature, Strength, Rump WidthNet Merit, Productive Life, Calving Ease, Stature, Strength, Rump Width


8

Marker Effects for MilkMarker Effects for Milk


8

Marker Effects for Final ScoreMarker Effects for Final Score


8

Linear and Nonlinear PredictionsLinear and Nonlinear Predictions

Linear model• Infinitesimal alleles model in which

all loci have non-zero effects

Nonlinear models• Model A: infinitesimal alleles with a

heavy-tailed prior• Model B: finite locus model with

normally-distributed marker effects• Model AB: finite locus model with a

heavy-tailed prior


8

Nonlinear and Linear Regressions Nonlinear and Linear Regressions for marker allele effectsfor marker allele effects


8

RR22 for linear and nonlinear predictions for linear and nonlinear predictions

ModelTrait Linear A B ABNet Merit 28.2 28.4 27.6 27.6Milk 47.2 48.5 46.7 47.3Fat 41.8 44.2 41.5 43.6Protein 47.5 47.0 46.8 46.6Fat % 55.3 63.3 57.5 63.9Protein % 51.4 57.7 51.4 56.6Longevity 25.6 27.4 25.4 26.4Somatic cell 37.3 38.3 37.3 37.6Days open 29.5 29.0 29.4 29.2


8

RR22 vs. Reliability vs. Reliability

Adjust the observed genomic R2

• Daughter deviations contain error– Divide by REL of 2008 deviationsDivide by REL of 2008 deviations

• Parents are selected– Add difference of PA RAdd difference of PA R22 from expected from expected

Adjust theoretical genomic REL• Genotypes contain a few errors • QTLs are located between SNPs


8

RR22 and Reliabilities for Traditional and and Reliabilities for Traditional and Genomic PredictionsGenomic Predictions

Squared corr (x100)

Reliability

Traditional Genomic Genomic

Trait PA Genomic PA Realized Gain

Net Merit 11 28 30 53 23

Milk 28 49 35 58 23

Fat 15 44 35 68 33

Protein 27 47 35 57 22

Fat % 25 63 35 78 43

Protein % 28 58 35 69 34


8

RR22 and Reliabilities for Traditional and and Reliabilities for Traditional and Genomic PredictionsGenomic Predictions

Squared corr (x100)

Reliability

Traditional Genomic Genomic

Trait PA Genomic PA Realized Gain

Longevity 17 27 27 45 18SCS 23 38 30 51 21Fertility 20 29 25 41 16S.calf ease 27 29 28 31 3D.calf ease 14 22 25 40 15Final score 23 36 24 42 18


8

Reliability Gains for Proven BullsReliability Gains for Proven Bulls

Bulls included in test had:• >10 daughters in August 2003• >10% increase in reliability by 2008• Numbers of bulls in test ranged from

104 to 735 across traits• Predicted the change in evaluation

Significant increase in R2 (P < .001) for 26 of 27 traits


8

Net Merit by Chromosome for O ManNet Merit by Chromosome for O ManTop bull for Net MeritTop bull for Net Merit

-40

-20

0

20

40

60

80

X 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30

Chromosome

NM

$

NM$


8

SNPs on X ChromosomeSNPs on X Chromosome

Each animal has two evaluations:• Expected genetic merit of daughters• Expected genetic merit of sons• Difference is sum of effects on X• SD = .1 σG, smaller than expected

Correlation with sire’s daughter vs. son PTA difference was significant (P<.0001), regression close to 1.0


8

X, X, YY, , Pseudo-autosomalPseudo-autosomal SNPs SNPs

487 SNPs

35 SNPs

0 SNPs

35 SNPs


8

Clones and Identical TwinsClones and Identical Twins21HO2121, 21HO2125, 21HO2100, CAN6139300, CAN613930321HO2121, 21HO2125, 21HO2100, CAN6139300, CAN6139303

Traditional Genomic

Bull Dtrs NM$ REL NM$ REL

Triton - ETN 98 -363 82 -371 91

Triad - ETN 26 -306 68 -370 91

Trey - ETN 108 -395 83 -371 91

Loyalty 108 -185 78 -196 87

Lauriet 83 -203 76 -196 87


8

Value of Genotyping More Value of Genotyping More SNPSNP9,604 (10K), 19,208 (20K), and 38,416 (40K) SNP9,604 (10K), 19,208 (20K), and 38,416 (40K) SNP

REL of PA

Genomic REL

Trait 10K 20K 40K

Net Merit $ 30 48 50 53

Milk yield 35 53 56 58

Fat yield 35 64 66 68

Protein yield 35 54 56 57

Productive Life 27 38 41 45

SCS (mastitis) 30 45 47 51

Dtr Preg Rate 25 37 39 41


8

Value of Genotyping More BullsValue of Genotyping More Bulls

Bulls R2 for Net Merit

Predictor Predictee PA Genomic Gain

1151 251 8 12 4

2130 261 8 17 9

2609 510 8 21 13

3576 1759 11 28 17


8

SimulatedSimulated Results ResultsWorld Holstein PopulationWorld Holstein Population

15,197 older and 5,987 younger bulls in Interbull file

40,000 SNPs and 10,000 QTLs

Provided timing, memory test

Reliability vs parent average REL• REL = corr2 (EBV, true BV) • 80% vs 34% expected for young bulls• 72% vs 30% observed in simulation


8

Brown Swiss ResultsBrown Swiss Results

Nearly all proven bulls genotyped• Data from 225 bulls born before 1999 • Predict 118 bulls born during or after 1999

Gains in young bull reliability • Expected to be 1% to 3%• Actual gains were about 2% for yield• Little or no gain for other traits

Cooperation with Europe is needed


8

Jersey GenotypesJersey Genotypes

Same experimental design• DNA available for 766 bulls • Total of 594 genotyped as of June• Results not available yet

Gains in reliability expected to be proportional to number of bulls genotyped


8

Expected vs Observed ReliabilityExpected vs Observed ReliabilityHolsteinsHolsteins

Reliability for predictee bulls • Traditional PA: 27% average across traits• Genomic: 63% expected vs. 50% observed• Observed range 78% (fat pct) to 31% (SCE)• PTA regressions .8 to .9 of expected

Multiply genomic daughter equivalents by .7 to make expected closer to observed • For example, 16 * .7 = 11• Include polygenic effect, less than 5%


8

Genetic ProgressGenetic ProgressHolsteinsHolsteins

Assume 60% REL for net merit• Sires mostly 2 instead of 6 years old• Dams of sons mostly heifers with 60% REL

instead of cows with phenotype and genotype (66% REL)

Progress could increase by >50%• 0.37 vs. 0.23 genetic SD per year• Reduce generation interval more than

accuracy


8

Genetic Evaluation AdvancesGenetic Evaluation Advancesand increases in genetic progressand increases in genetic progress

Year Advance % Gain

1935 Daughter-dam comparison 100

1962 Herdmate comparison 50

1973 Records in progress 10

1974 Modified cont. comparison 5

1977 Protein evaluated 4

1989 Animal model 4

1994 Net merit, PL, and SCS 50

2008 Genomic selection >50


8

How Related are Relatives?How Related are Relatives?CorrectionCorrection

Example: Full sibs • Share 50% ± 5% of their DNA on

average (in cattle)• SD 3.5% reported previously was low• For any diploid species, general

formula is 50% ± 50% / [2(C + L)].5, where C is number of chromosomes and L is genome length in Morgans


8

ConclusionsConclusions

Genomic predictions significantly better than parent average (P < .0001) for all 26 traits tested

Gains in reliability equivalent on average to 11 daughters with records • Analysis used 3576 historical bulls • April data included 5285 proven bulls

High REL requires many genotypes


8

AcknowledgmentsAcknowledgments

Genotyping and DNA extraction:• BFGL, U. Missouri, U. Alberta, GeneSeek, GIFV, and

Illumina

Computing: • AIPL staff (Mel Tooker, Leigh Walton, etc.)

Funding: • National Research Initiative grants

– 2006-35205-16888, 2006-35205-167012006-35205-16888, 2006-35205-16701• Agriculture Research Service• Contributors to Cooperative Dairy DNA Repository

(CDDR)


8

CDDR ContributorsCDDR Contributors

National Association of Animal Breeders (NAAB, Columbia, MO)• ABS Global (DeForest, WI)• Accelerated Genetics (Baraboo, WI)• Alta (Balzac, AB)• Genex (Shawano, WI)• New Generation Genetics (Fort Atkinson, WI)• Select Sires (Plain City, OH)• Semex Alliance (Guelph, ON)• Taurus-Service (Mehoopany, PA)

Reliability of Genomic Predictions for North American Dairy Bulls

Documents