Nature Genetics 37, 1299 - 1300 (2005)
doi:10.1038/ng1205-1299
Power tools for human geneticsLeonid KruglyakLeonid Kruglyak is at the Lewis-Sigler Institute for Integrative Genomics and Department of Ecology and Evolutionary Biology, Carl Icahn Laboratory, Princeton University, Princeton, New Jersey 08544, USA. leonid@genomics.princeton.edu Genome-wide association studies have the potential to identify systematically the contributions of common genetic variants to human disease, but the tools to carry out such studies have been incomplete. Assessments of the HapMap resource suggest that the tools are now in hand, but care is required in their use for study design, analysis and interpretation.Humans show a great deal of heritable interindividual variation in appearance, physiology, psychology, behavior and disease susceptibility. Geneticists and the general public alike want to understand the connection between this variation and variants in DNA sequence. In principle, this can be accomplished, at least for common variants, by comparing the frequency of every variant in individuals with and without a trait (or alternatively by comparing the trait distribution in individuals with and without a variant). In practice, although many of the  7 million variants that exist with frequencies above 5% in the human population1 have been identified2, current genotyping technologies are limited to testing 100,000–500,000 variants in a collection of individuals in any given association study. Fortunately, the fact that nearby variants are correlated owing to a lack of historical recombination between them (such correlations are known as linkage disequilibrium) offers a shortcut alternative to genotyping every variant. As a result, the presence or absence of a variant at one site provides information about neighboring variants, allowing one variant to serve as a proxy for several others and reducing the total number of genotyping assays (Fig. 1). Several hundred thousand proxies can capture most common variants through strong correlations (the exact number depends on the study population and the criteria for proxy choice)2,
3. But not just any proxy set of this size will do: variation in history among human populations and in recombination rates across the genome means that efficient proxy selection requires empirical measurement of linkage disequilibrium in multiple populations for a dense set of markers covering the genome4,
5. This has now been accomplished by the International HapMap Consortium, as reported in Nature2. Two related studies in Nature Genetics, by de Bakker et al.6 and Zeggini et al.7, assess the utility of this resource for genome-wide association studies.
 | | |
What is the HapMap?
Most human sequence variation is in the form of single-nucleotide polymorphisms (SNPs), and these are the markers of choice for association studies. The International HapMap Consortium genotyped more than one million SNPs (roughly one per 3 kb) in 269 individuals of four ethnic origins2. A second phase of the project is slated to increase the density of genotyped SNPs to one per kilobase. de Bakker et al.6 and Zeggini et al.7 used combinations of empirical and simulated data to show that the current HapMap resource allows selection of reference panels with 250,000–500,000 SNPs that reliably capture a high fraction (70–80%) of all common variants (those with an allele frequency above 5%). The fraction captured is considerably lower in a panel of individuals of African origin, but this should be remedied when data from the second phase become available2. Thus, the HapMap project has largely succeeded in achieving its principal goal. Although the resource was not designed to capture variants with frequency below 5%, both groups considered its ability to do so and found that some rare variants are picked up (in the range of 20%). More of these variants can be recovered by exhaustive tests of association with multimarker haplotypes, but at the cost of lowering the power to detect association with common variants owing to the increased number of statistical tests. The relative contribution of rare versus common variants to phenotypic variation is a big question mark, but early successes8 suggest that some common variants with important roles remain to be found.
How should a reference panel be assembled? The standard approach is to choose a minimal set of SNPs such that every common variant is in the panel or highly correlated with a SNP in the panel9. de Bakker et al. and Zeggini et al. show that panel size can be reduced by up to one-third if correlations between SNPs and multimarker haplotypes are considered. If this approach is pursued, a multimarker approach must also be used in association studies using the panel to avoid missing those variants correlated only with haplotypes (and not with any one SNP)7. de Bakker et al. propose a panel selection strategy in which SNPs are added to the panel in order of the number of SNPs they capture by proxy. In theory, this strategy can greatly reduce panel size while maintaining the power of an association study with a fixed number of individuals to detect a given effect. But this reduction can be achieved only by not including proxies for a substantial fraction of common variants, which means that power cannot be improved by increasing sample size. It is also worth noting that this strategy prioritizes the detection of variants strongly correlated with many others, making it difficult to identify the causal variant. In general, the average size of a group of strongly correlated SNPs is large (10–20)2. Although it is precisely this observation that enables reduction of the number of assays by proxy selection, it has the flip side that distinguishing a causal variant from those merely strongly correlated with it will often not be possible by genetics alone10.
HapMap revolutions?
The HapMap resource provides a rich data set for exploring demographic, selective and molecular forces that have shaped human sequence variation2. How will the resource be used by medical geneticists to find genetic determinants of disease, the main rationale for the project? In theory, researchers could use the principles laid out by de Bakker et al. and Zeggini et al., along with considerations of cost, desired power and prior knowledge of the genetic architecture of the disease in question, to select study-specific sets of SNPs for genotyping in the relevant case and control samples. In practice, however, this approach will be out of reach for all but a small number of the best funded and technologically most sophisticated groups. Instead, most researchers will use off-the-shelf SNP panels provided by a few commercial suppliers. It is therefore essential that such panels be carefully assembled with the aim of maximizing statistical power for a given number of SNPs (the feasible number will increase as genotyping technologies continue to improve). Equally important, each panel must be accompanied by a precise, freely available description of the selection algorithms, the identities of the SNPs and their genotypes in the HapMap samples, the fraction of variants captured at different correlation levels and the expected power of the panel for a universal benchmark. One such benchmark, modified from that used by de Bakker et al., is the power of the panel to detect, at a genome-wide P < 0.05, an association due to a causal variant (chosen without regard to its inclusion in the panel) with an effect that would be detected with 95% power at nominal P < 10-7 in 2,000 cases and 2,000 controls if tested directly (of course, the relevant parameters, including the frequency of the causal variant, should be varied to ensure robustness). Adherence to these standards will allow independent evaluation of the panels and facilitate development of software for association analysis appropriately tailored to each panel. Panel designers should also give priority to SNPs that change amino acids or otherwise alter gene function.
A new phase of human genetic research is upon us, one in which the contribution of common sequence variants to phenotypic variation will be tested systematically. This phase, relying on reference panels of common variants provided by the HapMap and similar resources11, will in turn be supplanted by one of routine resequencing of entire genomes in large collections of phenotyped individuals12,
13. Then we will finally be able to take an unbiased look at the influence of all types of DNA differences, common and rare, on heritable human traits.
REFERENCES
- | Article |
- | Article |
- | Article |
- | Article |
- | Article |
- | Article |
- | Article |
- | Article |
- | Article |
- | Article |
- | Article |
|
