Strategies for genomewide familybased association analysis James Degnan1, Jessica Su1, Cliona Molony

1
Strategies for genome-wide family-based
association analysis James Degnan1, Jessica
Su1, Cliona Molony2, Eric Schadt2, Benjamin
Raby3, and Christoph Lange1 1Harvard School
of Public Health, 2Genetics, Rosetta
Inpharmatics, 3Harvard Medical School
2
Outline

• Introductionthe data
• Marker-phenotype combinations
• Screening algorithm
• a. Univariate case
• b. Multivariate case (FBAT-PC)
• Simulations
• Conclusions

3
The Data (from Rosetta)

• 287 Individuals from 15 CEPH pedigrees
• (grandparents,parents,children) with family
  sizes between 13 and 17.
• 2322 SNP markers (also 270 microsats not being
  used) on each individual
• 23,380 gene expressions measured on more than
  half of individuals (167), some missing values.
  Phenotypes do not include disease.

4
Strategy

• To find marker-phenotype associations, we want to
  look at a subset of the
• 2322 x 23,380 54 million marker-phenotype
  combinations.
• To look for cis-acting SNPs, consider only SNPs
  within 1.0 Mb of a gene. Define distance of SNP
  to gene to be the minimum of the two distances
  from the marker to each of the two ends of the
  gene, or 0 if the SNP is in the gene.

5
Marker-phenotype combinations
6
Screening Algorithm (Van Steen et al., 2005)

• Compute conditional power based on conditional
  mean model for each marker-phenotype combination.
• Rank the combinations based on either conditional
  power or heritability. (Which is better?)
• Consider a marker-phenotype combination to be
  significant if its FBAT p-value is less than
  alpha and it has been screened.

7
Best SNPS sorted by informative families and
heritability
8
(No Transcript)
9
Best SNPS sorted by conditional power only
10
Computational issues

• The amount of time it takes to run each iteration
  of PBAT is highly variableit depends on the
  pattern of missingness in the marker data,
  particular for the parents.
• PBAT takes the same amount of time for each
  marker, independent of the phenotype. If a
  marker proves too computationally difficult, we
  have to skip all phenotypes near that marker (16
  on average, but up to 70).

11
Reconstructability of parental markers in CEPH
pedigrees
12
Method for speeding-up screen

• Pedigrees with missing parental markers can take
  very long to analyze when parental genotypes
  cannot be inferred.
• A modification of the screening algorithm is to
  only keep track of the top K (say, 500)
  marker-phenotype combinations.
• For each marker-phenotype, if the parents have
  missing values that cannot be reconstructed, an
  estimate of the conditional power can be obtained
  by filling in the missing parental genotypes.
• If the power is low compared to the top K
  power-values, then this marker-phenotype
  combination can be skipped otherwise it is
  computed normally, and the list of the K most
  powerful combinations is updated.
• The end result should be the same list of the K
  most powerful marker-phenotype combinations,
  although many combinations will not have had the
  conditional power computed using the original
  (missing data).

13
FBAT-PC

• In this approach, there is one overall phenotype
  constructed for each marker. This phenotype is
  the linear combination of expression values
  within 1.0 Mb of the marker that maximizes
  heritability.
• In this approach, screening is done on the set of
  markers, rather than the set of marker-phenotype
  combinations, thus greatly reducing the power
  values to be screened.

14
FBAT-PC Results (all models)
All Three models Obs Marker Allele
Freq model infofam pvalue
power 1 TSC0616740 2 0.49510
0 6 0.93620 0.022409 2
TSC0301105 2 0.39306 0 6
0.36774 0.018697 3 TSC0849086
1 0.25515 0 5 0.78279
0.010796 4 TSC0050595 2
0.41500 0 6 0.10615
0.008893 5 TSC1143002 1
0.17403 1 5 0.76342
0.008672 6 TSC0606794 1
0.45939 0 5 0.39329
0.007477 7 TSC0057231 1
0.27228 0 6 0.59724
0.007053 8 TSC0448850 2
0.36022 0 6 0.39794
0.005986 9 TSC0057231 1
0.27228 1 5 0.57135
0.005553 10 TSC0078180 1
0.36585 0 5 0.23947
0.005377 11 TSC0202163 1
0.48276 0 7 0.03660
0.004496 12 TSC0686093 1
0.24384 0 6 0.62200
0.004218 13 TSC0026698 2
0.35294 0 6 0.39119
0.004015 14 TSC1082055 2
0.36000 1 5 0.12454
0.003550 15 TSC0287131 2
0.43284 0 6 0.99423
0.003488 16 TSC1027993 2
0.45098 0 7 0.45889
0.003411 17 TSC1106988 2
0.48477 0 7 0.20923
0.003209 18 TSC0336892 2
0.31281 0 6 0.85698
0.003146 19 TSC0722384 2
0.42822 1 6 0.74845
0.003107 20 TSC0113708 1
0.24755 0 6 0.86111
0.003094
15
FBAT-PC results (additive and dominant models)
Obs Marker Allele Freq model
infofam pvalue power 1
TSC1143002 1 0.17403 1 5
0.76342 .008671941 2 TSC0057231
1 0.27228 1 5 0.57135
.005553376 3 TSC1082055 2
0.36000 1 5 0.12454
.003549725 4 TSC0722384 2
0.42822 1 6 0.74845
.003106552 5 TSC0930497 1
0.18687 1 6 0.43951
.002934976 6 TSC0300223 2
0.20297 1 5 0.02937
.002838104 7 TSC1021966 1
0.37685 1 6 0.87593
.002657709 8 TSC0127388 1
0.36318 1 5 0.16591
.002651893 9 TSC0507796 2
0.41707 1 5 0.80357
.002615662 10 TSC0130164 1
0.39791 1 5 0.78679
.002544931 11 TSC0221720 2
0.33659 1 6 0.29030
.002276393 12 TSC0873967 1
0.36683 1 5 0.63685
.002158055 13 TSC1085444 2
0.29268 1 5 0.32480
.002111099 14 TSC0925231 1
0.37745 1 5 0.72580
.002102141 15 TSC0483362 2
0.32250 1 5 0.03309
.002047252
16
FBAT-PC interpretation

• Although no SNPs achieved genome-wide
  significance, a few SNPs were promising despite
  the low number of informative families
• When most promising SNPs based on FBAT-PC were
  analyzed using univariate approach, the lowest
  p-value genes sometimes had low power.

17
Univariate analysis of most promising SNP based
on FBAT-PC
18
Simulations

• Simulations were performed based on simulating
  200 trios, 1000 markers, and 10 gene expression
  values for each marker. 100 markers were
  considered genuinely cis-acting on 1 of the 10
  genes near the marker (so 1 of the 10,000
  marker-phenotype combinations are based on true
  cis-acting SNPs).
• Allele frequencies were drawn from a
  Unif(0.1,0.5) distribution for each marker, and
  phenotypes were drawn from a N(aX,1.0)
  distribution, where X is the number of A alleles,
  and a is the effect size assuming an additive
  model.

19
Power of PBAT screen as a function of the number
of markers
20
References

• Van Steen, K., et al. Genomic screening and
  replication using the same data set in
  family-based association. Nature Genetics 37
  683-691. 
• Monks, S.A., et al. Genetic inheritance of gene
  expression in human cell lines. Am. J. Hum. Gen.
  75 1094-1105