Efficiency and power in genetic association studies

1
Efficiency and power in genetic association
studies

• Yixuan Chen
• 1/11/2008

2

• Paul I Wde Bakker, et al. Efficiency and power
  in genetic association studies. Nature Genetics
  37, 11 (2005).
• Nikolas Maniatis, et al. Effects of Single SNPs,
  Haplotypes, and Whole-Genome LD Maps on Accuracy
  of Association Mapping. Genetic Epidemiology 31
  179188 (2007).

3
Outline

• Motivation
• Comparison
• Conclusions
• Discussion

4
Motivation

• HapMiner project

5
HapMiner

• HapMiner project extension
• Utilize family data by applying phenotype scores
  for founder haplotypes
• Seek experimental supports
• Situations where haplotype-based association
  tests have advantages over single-marker tests

6
Association Study

• Genotyping a higher density of SNPs increases the
  fraction of sites captured through LD (linkage
  disequilibrium)
• Correlations among nearby variants (LD) can
  improve the costeffectiveness, guiding selection
  of informative tag SNPs and providing
  information about nearby variants not genotyped
• The use of multimarker haplotypes may result in
  greater efficiency

7
Efficiency and Power

• Explicitly modeled disease association studies
• Empirically assessed significance thresholds
• Carried out evaluations using empirical (rather
  than simulated) human genotype data from the
  International HapMap ENCODE Project

8
Simulation

• To simulate a case-control panel
• Designated one SNP to be causal
• Calculated an effect size such that if this SNP
  were directly tested in 1,000 cases and 1,000
  controls, power would be 95 to achieve a nominal
  P value of 0.01
• Fixed the absolute power for each putative causal
  SNP
• Rare alleles are assigned a stronger effect than
  common alleles
• Examined only common alleles (with frequency gt
  5)

9
Significance Threshold

• The significance threshold for declaring
  association was based on the empirical null
  distribution
• The statistical tests were examined in a set of
  null panels (in which no SNP is causal), with the
  maximum chi-square value exceeded in 1 of null
  panels

10
Tag SNP Selection

• Select a subset of non-redundant SNPs from the
  reference panel such that every common allele
  either is directly genotyped or has a perfect
  proxy (r2 1.0) among the tags.

11
Use Haplotype

• An identical set of tests of association with one
  degree of freedom (d.f.) are done but we allow a
  haplotype of tags to serve as surrogate for an
  untyped SNP
• In other words, if a specific multimarker
  combination (i.e., haplotype of tag SNPs) can
  serve as an effective proxy for some putative
  causal alleles, then these alleles need not be
  typed as tag SNPs (or tested as single markers).

12
Increased Efficiency
13
Relax Thresholds for Tagging

• r2 gt 0.8
• Best N rank potential tags according to the
  number of other SNPs for which they can act as a
  proxy and then to type the SNPs in this priority
  order

14
In Summary

• If a complete reference panel is available,
  multimarker haplotype tests are more efficient
  than pairwise tests, and prioritizing SNPs on the
  basis of their LD properties allows impressive
  reductions in the genotyping burden while
  maintaining excellent power.

15
Incomplete Panel

• Created a pseudo 5-kb HapMap by thinning the
  ENCODE data to achieve the spacing and frequency
  distribution of phase I HapMap
• Selected tags and designed tests using this
  incomplete resource, evaluating performance in
  simulated case-control panels where all alleles
  (not just those from the incomplete HapMap) were
  allowed to be causal

16
Undiminished Power

• Two key changes
• A much smaller set of tags was selected
• A subset of common variants had no good proxies
• Power was largely undiminished
• 95 relative power in CEU, using a set of best N
  tags

17
The Unexpected

• Tests that capture many putative causal alleles
  add the same amount to the multiple testing
  burden as do independent tests that capture only
  a single site.

18
The Best N

• The best N approach underperforms at sparser
  densities
• The best N method suffers further when applied to
  incomplete reference panels
• Therefore, where complete data is available, and
  as denser versions of HapMap become available
  (such as the pending phase II), the utility of
  the best N method should increase, particularly
  for choosing marker densities of more than one
  SNP per 10 kb

19
Conclusions

• Specified multimarker tests substantially
  increase tagging efficiency relative to
  singlemarker approaches, without loss of power.
• When selecting SNPs from very dense reference
  panels, a method such as the best N strategy,
  which ranks SNPs according to the number of
  proxies they have, allows marked reductions in
  genotyping with limited loss of power,
  substantially outperforming a method based on
  relaxing r2 thresholds.
• Sparser sets of tags selected from a pseudo phase
  I HapMap are almost as powerful as equally sized
  sets chosen from complete reference panels.

20
Linkage Disequilibrium Maps in LD Units (LDU)

• Using LDU locations for the 27 SNPs flanking the
  CYP2D6 gene on chromosome 22, the most common
  functional polymorphism within the gene was
  located at 15 kb from its true location.
• Expressing the locations of the 27 SNPs in LDU
  from the HapMap LDU map, analysis yielded an
  estimated location that is only 0.3 kb away from
  the CYP2D6 gene.
• The haplotype data provided much poorer
  localization compared to single SNP analysis.

21
Drawbacks of Haplotype

• If haplotype testing increases the degrees of
  freedom or number of tests in statistical
  analysis, it may decrease, rather than increase,
  overall power.
• Autocorrelation generated by the duplicated SNPs
  among haplosets
• Deflates the error variance for association
  mapping
• power is exaggerated and localization is poor
  compared to single SNPs

22
Discussions

• Haplotype-based methods may perform better
• If allelic heterogeneity exists (more than one
  causal alleles at a disease locus) (not
  applicable)
• Two-locus disease model
• Disease allele is not common (frequency lt 5)?
• Use tag SNPs?
• More?

23
THANKS!