VISG – LARGE DATASETS

Literature Review

Introduction – Genome Wide Selection

Aka Genomic SelectionSet of Markers

• 10,000’s - enough to capture most genetic information

‘Training set’ of animals• phenotyped and genotyped• representative of industry

Predictor• Over-specified – e.g. 10000 variables, 1000 individuals• Robust model selection required

Application• Predict in selection candidates

– Maybe no phenotypes– Maybe no pedigrees

Introduction – Genome Wide SelectionPrediction Methods

• Stepwise Regression• gBLUP

– Fit all markers as a random effect– gi ~ N(0,g

2)

• BayesA– gi ~ N(0,gi

2)– prior : gi

2~ S/2 (choose S and )

• BayesB– similar to BayesA, except– proportion of effects are zero

Most investigations compare theseMany variations (sometimes with the same name)

Literature

Dairy applications review (Hayes et al., 2009)GWS in crops (Heffner, Sorrells, Jannick, 2009)Prediction in unrelateds (Meuwissen, 2009) Marker panels (Habier, 2009)Phenotypes (Harris & Johnson, unpub)+ ...

Issues

National evaluationsLong term gainsLD or relationship trackingMultiple breedsDistance from Training to ApplicationMarker Panels (subsets)Phenotypes (EBV-based)Non-additive effectsComputing requirements

Methods

gBLUP almost as good as Bayes(A) (dairy)• Interpretation(?): many genes of small effect

Bayes methods better at using real LD (vs relatedness)Bayes(B) advantage greater with

• Higher marker density• Higher Training Application distance• Smaller Training set

Mixture of 2 normals ~ BayesBPartial Least SquaresMachine LearningHaplotype methods not used in practice yet

Marker Panels

Evenly spaced panels• Track inheritance from parents (both SNP-chipped)• Will work with new traits

Lasso methods popular• Shrinks small effects to zero

Other

Combining marker and other information• Phenotype info, parent info• Index methods; ‘blending’• Important for seamless national evaluations

Computing strategies• Tricks to reduce computation• Approximation rather than Iterative (MCMC) methods

Online resources

Conferences• Statistical Genetics of Livestock for the Post-Genomic Era.

UW-Madison, May, 2009. http://dysci.wisc.edu/sglpge/index.html

• QTL/MAS Workshops. 2008: http://www.computationalgenetics.se/QTLMAS082009: http://www.qtlmas2009.wur.nl/UK/

Courses• Whole Genome Association and Genomic Selection.

September 1-8, 2008, Salzburg, Austria. http://www.nas.boku.ac.at/12100.html?&L=0

•  Use of High-density SNP Genotyping for Genetic Improvement of Livestock . Iowa State, June, 2009. http://www.ans.iastate.edu/stud/courses/short/

Toy example

• 5 SNP / 1000 individuals• y = mu + SNP1 + e

– mu = 10– SNP1 substitution effect = 10 / p = 0.5– Var(e) = 1

• 1 block / 1000 iterations• Runs in ~ 5 secs

200 400 600 800 1000

02

46

810

iteration

b.sa

mp

200 400 600 800 1000

1.0

1.5

2.0

2.5

3.0

iteration

z.sa

mp

200 400 600 800 1000

89

1011

12

iteration

mu.

sam

p

200 400 600 800 1000

02

46

810

14

iteration

var.b

.sam

p

0 2 4 6 8 10

0.0

0.4

0.8

b 1

N = 900 Bandw idth = 0.09915

Den

sity

0 2 4 6 8 10

05

1015

20

b 2

N = 900 Bandw idth = 0.02

Den

sity

0 2 4 6 8 10

020

4060

b 3

N = 900 Bandw idth = 0.00561

Den

sity

0 2 4 6 8 10

05

1020

b 4

N = 900 Bandw idth = 0.01463D

ensi

ty

0 2 4 6 8 10

0.0

0.5

1.0

1.5

b 5

N = 900 Bandw idth = 0.2309

Den

sity

3

z 1

020

060

0

1 2

z 2

020

060

0

1 2

z 3

020

060

0

1 2 3

z 40

200

600

1

z 5

020

060

0

8 9 10 11 12

0.0

0.2

0.4

mu

N = 900 Bandw idth = 0.1564

Den

sity

0 2 4 6 8 10 12 14

0.00

0.10

0.20

var.b 2

N = 900 Bandw idth = 0.3301

Den

sity

200 400 600 800 1000

0.0

0.2

0.4

0.6

0.8

1.0

iteration

p

0.5 0.6 0.7 0.8 0.9 1.0

01

23

45

6

p 1

N = 900 Bandwidth = 0.0169

Den

sity

0.0 0.1 0.2 0.3 0.4 0.5

05

1015

20

p 2

N = 900 Bandwidth = 0.007388

Den

sity

0.0 0.1 0.2 0.3 0.4

02

46

8

p 3

N = 900 Bandwidth = 0.01327

Den

sity