6/3/16

postGSf90 - ssGWAS

Ignacio Aguilar
Ins;tuto Nacional de Inves;gación Agropecuaria
INIA Las Brujas, Uruguay
[email protected]

Classical GWAS
• Test single marker one at ;me
• Simple linear regression for the SNP effects
• Other fixed effects can be fit (contemporary
group)
• Polygenic breeding value can be used

1
 6/3/16

Genomic Selec;on
• Considers all gene;c associa;ons derived
from markers
• Methods (Bayes A, Bayes B, Bayes Lasso)
provide solu;ons to SNP effects

• Then Genome-Wide Associa;on Analysis

(GWAS) can be performed
• Accounts for popula;on stra;fica;ons and
cryp;c relatedness

Non-Genotyped individuals with

Phenotype
• Recorded informa;on from non-genotyped individuals
can not easily be incorporated in Single marker
regression and Bayesian methods

• Although can incorporate informa;on by accumula;ng

data from rela;ves , e.g. EBV

• But problems with

– heterogeneity from different sources
– Loss of informa;on
– bias
– Computa;onal cost with MCMC method for large number
of genotypes and makers

2
 6/3/16

Single-step GBLUP
• Integrates all available informa;on
– Phenotypes
– Genotypes
– Pedigree
• Limita;on of ssGBLUP, in contrast to BayesX
methods
– infinitesimal model i.e. same variance for all SNP
effects

ssGWAS
• Combining methods
– Unequal variances
– Use all available informa;on like in ssGBLUP

• Improve Accuracy of es;ma;on of GEBV

– For breeding and selec;on
– Accuracy for es;ma;on of SNP effects for GWAS

3
 6/3/16

SNP variances
• Zhang et al. 2010, presented a method to
es;mate weights for SNP variances without
sampling i.e. non MCMC methods

• SNP weights: func;on of squares of SNP effects

• Incorporate weights into genomic rela;onship

matrix

• Similar approach by Sun et al., 2012 PlosOne

BUT both approaches can not u;lize
phenotypes of un-genotyped individuals !!!

SNP weights
• SNP weights derived from SNP effects

• Zhang et al., 2010 PlosOne

• Matrix D, diagonal matrix, with un-equal

variances for each SNP

4
 6/3/16

Equivalent Model
VanRanden et al 2009; Goddard, 2009; Habier el al 2007

Model that estimate SNPs effects

y = µ + Za + e, var(a) = Dσ a2 u = Za

Model that estimate Breeding Values

y = µ + u + e, var(u) = Gσ u2 , G = ZDZ '/ k

Genomic Information ≈ genomic relationship

Simple conversion between :
Breeding values and SNP effects
u = Za
a = DZ '(ZDZ ')−1 u Stranden & Garrick, 2009

Equivalent Model
• SNP effects from GEBV’s (Henderson, 1973;
Strandén and Garrick, 2009):
σ u2
uˆ = 2 DZ ' G −1aˆ g = DZ ' [ ZDZ ' ]−1 aˆ g
σa
Differen'al weight to each SNP

• Also, for each SNP effect (i-th):

σˆ u2,i = uˆi2 2 pi (1 − pi )

5
 6/3/16

Compu;ng algorithm
• Denote t as an itera;on number and i as the i-th SNP

1. t=0, D(t)=I, G(t)=ZD(t)Z’λ

2. Compute â gby ssGBLUP

−1
3. Calculate uˆ( t ) = λD( t ) Z ' G( t ) aˆ g

4. Calculate d i*( t +1) = uˆi2( t ) 2 pi (1 − pi ) for all SNPs (Zhang et al., 2010)
tr ( D( 0 ) )
5. Normalize D( t +1) = D(*t +1)
tr ( D(*t +1) )

6. Calculate G( t +1) = ZD( t +1) Z ' λ

7. t=t+1

8. Exit , or loop to step 2 or 3

6
 6/3/16

Weighted ssGBLUP
WssGBLUP(Wang et al., 2012)

Weights = I

GEBV

SNP
effects
SNP
weights

G matrix

- Gives more weight to important markers

Simulated data
1. QMSim
2. Simple model: y = 1 µ + Zaa + e
3. 10 QTLs w. 3000 SNP markers on 2 chromosomes
4. N = 15,800
Ng= 1500
5. h2=0.5, all due to QTLs (No Polygen)
6. 10 replica;ons

7
 6/3/16

Methods

1. Single marker model: WOMBAT

2. BayesB using de-regressed proofs : GENSEL

3. ssGBLUP: S1 & S2

Manhapan plot of S2

8
 6/3/16

Manhapan plot of BayesB

Manhapan plot of WOMBAT

9
 6/3/16

Accuracy of (G)EBVs

EBVs DP
BLUP 0.81 0.77
(0.01) (0.01)
it1* it2 it3 it4 it5 it6 it7 it8
ssGBLUP 0.87 0.89 0.88 0.88 0.88 0.87 0.87 0.87
(0.01) (0.01) (0.01) (0.02) (0.02) (0.02) (0.02) (0.02)
†
NW c=0.1
BayesB_DP 0.88 0.88
(0.02) (0.02)
1

Accuracy of SNP effects

1 Table 3. Average correlations (standard deviations) between QTL effects and sum of cluster of

2 m SNP effects using ssGBLUP

S1* 1† 2 4 8 16 40
it1 0.53 (0.07) 0.68 (0.05) 0.79 (0.03) 0.81 (0.02) 0.80 (0.03) 0.62 (0.08)
it2 0.46 (0.07) 0.66 (0.05) 0.78 (0.02) 0.82 (0.02) 0.81 (0.02) 0.63 (0.08)
it3 0.43 (0.07) 0.64 (0.05) 0.77 (0.02) 0.81 (0.02) 0.80 (0.02) 0.62 (0.08)
it4 0.42 (0.07) 0.63 (0.05) 0.77 (0.02) 0.81 (0.02) 0.80 (0.02) 0.62 (0.08)
it5 0.41 (0.07) 0.63 (0.05) 0.76 (0.02) 0.80 (0.02) 0.79 (0.02) 0.61 (0.08)
it6 0.41 (0.07) 0.62 (0.05) 0.75 (0.02) 0.80 (0.02) 0.79 (0.02) 0.61 (0.07)
it7 0.41 (0.07) 0.62 (0.05) 0.75 (0.02) 0.80 (0.02) 0.79 (0.02) 0.61 (0.07)
it8 0.41 (0.07) 0.62 (0.05) 0.75 (0.02) 0.80 (0.02) 0.79 (0.02) 0.60 (0.07)
S2 1 2 4 8 16 40
it1 0.53 (0.07) 0.68 (0.05) 0.79 (0.03) 0.81 (0.02) 0.80 (0.03) 0.62 (0.08)
it2 0.44 (0.09) 0.65 (0.06) 0.77 (0.03) 0.82 (0.03) 0.81 (0.02) 0.63 (0.06)
it3 0.41 (0.08) 0.62 (0.05) 0.75 (0.03) 0.79 (0.03) 0.79 (0.03) 0.65 (0.06)
it4 0.40 (0.07) 0.61 (0.05) 0.73 (0.03) 0.77 (0.03) 0.78 (0.03) 0.64 (0.06)
it5 0.40 (0.07) 0.60 (0.05) 0.72 (0.04) 0.76 (0.04) 0.77 (0.04) 0.64 (0.06)
it6 0.40 (0.07) 0.60 (0.05) 0.72 (0.04) 0.75 (0.04) 0.76 (0.04) 0.63 (0.06)
it7 0.40 (0.07) 0.60 (0.05) 0.72 (0.04) 0.75 (0.04) 0.76 (0.04) 0.63 (0.06)
it8 0.40 (0.07) 0.60 (0.05) 0.71 (0.04) 0.75 (0.04) 0.76 (0.04) 0.63 (0.06)
*
3 S1: update weights for SNP effects but not for GEBVs; S2: update weights for both GEBVs
4 and SNP effects in each iteration.
†
5 Number of SNPs (i.e. m ranges from 1 to 40) in each cluster.

10
 6/3/16

Variances explained by segments

• ISU propose to present results from GWAS
using variance explained by windows of
adjacent SNP

• Fan et al 2011, Onteru et al 2011, Peters el al

2012, etc.

• Poten;ally use of bootstrap to get significance

of detected QTL

Windows Variances
Z u

a = Zu for only SNP in segment

a = EBV derived from segment

Get sample variance Var(a)

from genotyped individuals

11
 6/3/16

postGSf90
• Genomic POST processing program

• Extract SNP effects from solu'ons aqer

genomic evalua;ons (GLBUP and ssGLBUP)

• Calculate variance explained by segments

postGSf90 par files

1) Parameter files:
(1) BLUPF90 (and preGSf90)
(2) postGSf90

2) OPTIONs:
BLUPf90 / PreGSf90:
OPTION SNP_file marker.geno.clean
OPTION saveGInverse
OPTION weightedG w # A vector with length = M
postGSf90:
OPTION SNP_file marker.geno.clean
OPTION ReadGInverse
OPTION chrinfo mapfile #format: snpID chr pos
OPTION weightedG w
# OPTION Manhattan_plot
3) Document:
hpp://nce.ads.uga.edu/wiki/doku.php?id=readme.pregsf90#gwas_op;ons_postgsf90

12
 6/3/16

POSTGSF90 Op;ons

POSTGSF90 Op;ons

13
 6/3/16

Output files from POSTGSF90

14
6/3/16

15
 6/3/16

Model to study genetic of heat stress

• Performance data + weather data (Ravagnolo & Misztal, 2000)

Variation in slopes
could be due to
Milk yield

genetic component

F(Heat Index)

Single-Step GWAS Heat Stress

• Mul;ple-Trait Test-Day model heat tolerance
• ~ 90 millions records, ~ 9 millions pedigrees
• ~ 3,800 genotyped bulls
• Compu;ng ;me
– Complete evalua;on ~ 16 h
Milk yield no Heat stress Heat stress

16
 6/3/16

Variance explained Heat Stress

Milk yield no Heat stress

Heat stress

17