1. Xiaole Shirley Liu STAT115/STAT215/
Haplotypes and GWAS
Xiaole Shirley Liu
STAT115/STAT215/

2. Haplotype Haplotype block: a cluster of linked SNPs
Haplotype boundary: blocks of sequence with strong LD within blocks and no LD between blocks, reflect recombination hotspots
Association studies using haplotype is more accurate than using individual SNPs
Haplotype size distribution
STAT115

3. SNP Profiling [C/T] [A/G] T X C [A/C] [T/A] Tagging SNPs:
Possible haplotype: 24
In reality, a few common haplotypes explain 90% variations
Tagging SNPs:
SNPs that capture
most variations
in haplotypes
removes
redundancy
Redundant
STAT115

4. SNP Genotyping One SNP at a time or genome-wide (SNP array) 2.5kb
0.30
STAT115

5. 40 Probes Used Per SNP Allele call Signal AA, BB, AB Theoretically
1A+1B, 2A, 2B
But could
have 1A+3B
Amplified!
STAT115

6. Haplotype Inference
Genotyping only tells an individual is e.g. Aa BB Cc, but it doesn’t tell whether haplotype is: ABC + aBc, or ABc + aBC
Haplotype can often be inferred if parental genotype is known
Similar to blood typing, e.g. F: A, M: AB, C: B  F: , M: , C:
Otherwise, look at the population genotypes, infer common haplotypes
STAT115

7. Haplotype Inference Clark’s Algorithm
Construct haplotypes from unambiguous individuals
Remove samples that can be explained as combinations of haplotypes discovered already
Propose haplotype that would explain most remaining
Iterate 2 & 3 until finish
STAT115

8. Haplotype Inference Clark’s Algorithm
Construct haplotypes from unambiguous individuals
Remove samples that can be explained as combinations of haplotypes discovered already
Propose haplotype that would explain most remaining
Iterate 2 & 3 until finish
Disadvantages:
Depend on # of ambiguous subjects
Cannot get started when n is small
STAT115

9. EM and Gibbs Sampling in Motif Finding
Problem
Observe: sequence S
Unknown: motif θ and site location A (alignment), but given one, can infer the other
EM and Gibbs Sampler
Initialize random motif θ
Iterate:
Given θ and sequence S, update site location A
Given A and S, update θ
EM updates by weighted average
Gibbs sampling updates by sampling
STAT115

10. Statistical Model for Haplotype
T T A C C --- 1
T T A C G --- 2
T T A G C --- 3
T T A G G --- 4
T T C C C --- 5
T T C C G --- 6
T T C G C --- 7
T T C G G --- 8
Haplotype
Frequency 4 2 5 3 1 6 7 8
Haplotype Pool 1 6
Each individual’s two haplotypes are treated as random draws from a pool of haplotypes with certain frequencies that can satisfy the genotyping
STAT115

11. Haplotype Inference EM and Gibbs Sampler
Observe genotype Y, estimate haplotype pair Z for each individual and haplotype frequency 
Initialize haplotype frequencies
Iteration:
Estimate Z given Y, 
Estimate  given Y, Z
STAT115

12. Haplotype Inference EM and Gibbs Sampler
Observe genotype Y, estimate haplotype pair Z for each individual and haplotype frequency 
Initialize haplotype frequencies
Iteration:
Estimate Z given Y, 
Estimate  given Y, Z
STAT115

13. Haplotype Inference Partition-Ligation
When #SNP is big, # possible haplotypes is too big, so divide and conquer
Consider an inferred sub-haplotype as one allele
STAT115

14. Hapmap of Human Genome
HapMap: catalog of common genetic variants in human
What are these variants
Where do they occur in our DNA
How are they distributed within populations and between populations around the world
Goals:
Define haplotype “blocks” across the genome
Enable unbiased, genome-wide association studies
STAT115

15. 1000 Genomes Projects
Characterization of human genome sequence variation
Foundation for investigating the relationship between genotype and phenotype
Break
STAT115

16. Association Studies Association between genetic markers and phenotype
E.g. Cystic Fibrosis ~70% of Cystic Fibrosis patients have a deletion of 3 base pairs resulting in the loss of a phenylalanine amino acid at position 508 of the CFTR gene
Especially, find disease genes, SNP / haplotype markers, for susceptibility prediction and diagnosis

17. SNPs in Pharmacogenomics
Warfarin and CYP2C9:
SNPs in Pharmacogenomics
Warfarin anticoagulant drug; CYP2C9 gene metabolizes warfarin.
A patient requiring low dosage warfarin compared to normal population, has an odd ratio of 6.21 for having  1 variant allele
Subgroup of patients who are poor metabolisers of warfarin are potentially at higher risk of bleeding
Aithal et al., 1999, Lancet.

18. Influences individual decisions on life styles, prevention, screening, and treatment

19. Genome-Wide Association Studies
Quality Control
Unusual similarity between individual
Wrong sex
Trio has non-Mendelian inheritance
Genotyping quality
Two strategies:
Family-based association studies
Population-based case-control association studies

20. Quality Control: SNP calls
% SNP called
SNP calls from all the samples at a locus
Good calls!
Bad calls!

21. Family-based Association Studies
Look at allele transmission in unrelated families and one affected child in each
Like coin toss, likelihood of fair coin
A a A a

22. TDT: Transmission Disequilibrium Test
Only heterozygote parents matters, calculate observed over expected
Could also compare allele frequency between affected vs unaffected children in the same family
Break

23. Case Control Studies
SNP/haplotype marker frequency in sample of affected cases compared to that in age /sex /population-matched sample of unaffected controls

24. From Genotyping to Allele Counts

25. Test Significant Associations
Expected:
( ) * ( ) / ( ) = 49
( ) * (86+296) / ( ) = 321
2 = 27.5, 1df, p < 0.001

27. Association of Alleles and Genotypes of rs1333049 (‘3049) with Myocardial Infarction
2 (1df)
P-value
Cases
2,132 (55.4)
1,716 (44.6)
55.1
1.2 x 10-13
Controls
2,783 (47.4)
3,089 (52.6)
Allelic Odds Ratio = 1.38
OR = 1, no disease association
OR > 1, allele C increase risk of disease
OR < 1, allele C decrease risk of disease
Samani N et al, N Engl J Med 2007; 357:

28. Multiple hypotheses testing?
GWAS Pvalues

29. GWAS Pvalues for Type II Diabetes
Bonferroni correction: most common, typically p < 10-7 or 10-8
Manhattan Plot
How many SNPs were done?
McCarthy et al, Nat Rev Genetics, 2008

30. Reproducibility of Association Studies
Most reported associations have not been consistently reproduced
Hirschhorn et al, Genetics in Medicine, 2002, review of association studies
603 associations of polymorphisms and disease
166 studied in at least three populations
Only 6 seen in > 75% studies

31. Size Matters
Visscher, AJHG 2012

32. How to Improve Statistical Power?
Without increasing samples?
Test association of disease with haplotypes instead of individual SNPs
Also reduce genotyping errors
Split samples:
First half narrow down promising SNPs / haplotypes
Second half refining hits (much fewer multiple hypotheses)
Increase sample size: precision medicine initiative cohort ~ 1 million volunteers

33. Manolio et al., Clin Invest 2008
P < 9.9 × 10–7 (P<=10-6)
Manolio et al., Clin Invest 2008 33

34. Summary Haplotype inference
Clarks: resolve unambiguous first, propose new haplotypes to maximize explanation
EM & Gibbs: iteratively infer haplotype frequency and individuals’ haplotypes
Tagging SNPs and GWAS
Family based association studies: TDT transmitted allele to affected child
Case control studies: X-sq (allele frequency difference in case and controls) and OR
STAT115

35. Acknowledgement Jun Liu & Tim Niu Cheng Li & Yuhyun Park
Kenneth Kidd, Judith Kidd and Glenys Thomson
Joel Hirschhorn
Greg Gibson & Spencer Muse
Jim Stankovich
Teri Manolio
David Evans
Guodong Wu
Stefano Monti
Bo Li