The Genom of Homo sapiens.pdf

398 GOLDSTEIN, CAVALLERI, AND AHMADIpopulations. Interestingly, the size of blocks variedacross populations reflecting differing population histories,with European and Chinese populations having, onaverage, larger block sizes than African-American andYoruban samples (maximum block size 173 kb vs. 94 kb)(Gabriel et al. 2002).TAGGING METHODOLOGY ANDRELATED ISSUESAs noted, the first use of the term “tagging” was by Johnsonand colleagues, who suggested the term htSNPs. In thecase of all 9 genes examined in the Johnson paper, 2–5htSNPs per gene were sufficient to tag the common haplotypes.That is, instead of typing the full set of 122 SNPS,close to the same haplotypic variation could be captured bytyping a subset of 34 htSNPs (Johnson et al. 2001).Tagging common haplotypes is only one of many possibleways to select a subset of SNPs that retain as muchinformation as possible about the other SNPs. Broadlyspeaking, the approaches that have been evaluated can bedivided into two groups (Weale et al. 2003): those basedon maximizing the haplotype diversity present in the taggingset compared to the tagged set (diversity based) andthose based on establishing as high an association as possiblebetween the “tagging” and “tagged” set (associationbased). To avoid the close identification with haplotypediversity in the selection of tags, some have suggestedthat tags be referred to as tSNPs rather than htSNPs (see,e.g., Weale et al. 2003).The primary motivation for tSNP selection is their applicationin LD-based gene mapping. For this reason, atSNP selection criterion focused on the r 2 measure of LDseems the most directly relevant because it allows quantificationof the loss of power in typing the tSNPs insteadof all the SNPs. Pritchard and Prezeworski showed thatfor two biallelic loci, power scales with r 2 , such that typingthe associated marker with n/r 2 individuals wouldhave approximately the same power as n individuals inwhich the causative variant itself was typed, where r 2 isthe association between the two variants (Pritchard andPrzeworski 2001). This finding has been extended byChapman et al. (2003) to include generalized r 2 , includinghaplotype r 2 (see below).MULTIMARKER VERSUS PAIR-WISEAPPROACHESThere still remains the question of how to define the r 2value. Relying on pair-wise measures is straightforward,but may be inefficient. This is because pairs of SNPs willonly have high pair-wise association when their minor allelefrequencies are very closely matched, thus meaningthat SNPs which exhibit a full range of frequencies willneed to be selected as tags. This can be overcome if combinationsof the tSNPs are used to predict the other SNPs.One approach to this is to use the haplotype r 2 value(Chapman et al. 2003; Goldstein et al. 2003; Weale et al.2003; and the D. Clayton Web site: http://wwwgene.cimr.cam.ac.uk/clayton/software).Haplotype r 2 is defined as the proportion of variance ina “tagged” SNP of interest that is explained by an analysisof variance based on the G haplotypes formed by theset of tSNPs.Yi = x i1 b 1 + x i2 b 2 + .... + x iG b GWhere Yi is the predicted state of the tagged SNP of intereston the ith chromosome, x i1 ...x iG are indicator variablesfor the G haplotypes, and b 1 ...b G are coefficients estimatedby standard least squares from the observed data.This approach is more efficient in the sense of requiringfewer tags because it relies on combinations of haplotypesgenerated by tagging SNPs to predict the state oftagged SNPs. These combinations are identified by selectingthe appropriate coefficients in a linear regression.The haplotype r 2 criterion therefore appears an appropriatemeasure if one of the aims is to reduce the number oftags that must be typed in phenotyped material (e.g.,cases and controls).The haplotype r 2 approach focuses on the prediction ofhaploid allelic state (0 or 1) on the basis of the tSNP haplotypeobserved on a given chromosome. As such, it doesnot explicitly address the issue of haplotype inference inphenotyped individuals. One approach that simultaneouslyconsiders both aspects is from Stram et al. (2003), who definea coefficient of determination for predicting the haplotypesobserved in an individual based on the tSNP configuration.At present, it is hard to predict which approachesto tSNP selection will prove the most useful in practice.BLOCK-BASED AND BLOCK-FREESELECTION OF TAGSAlthough the discovery of the block-like nature of LDand its effect on haplotype distribution inspired the ideaof tags for given haplotypes (Johnson et al. 2001), the useof tSNPs in no way depends on blocks of LD. Indeed,even if there is such a block structure, it is not apparentthat tag selection should make reference to blocks. Asnoted above, the early suggestions for the definition ofhtSNPs did not address this issue directly. More recently,however, we have argued that block-based identificationof tags will always be less efficient (sometimes considerably)than methods that select across block boundaries.This is because tagging within blocks limits the effectiverange of a set of tags, and means that cross-block associationscannot be exploited (Goldstein et al. 2003). For thisreason, we advocate the selection of tSNPs across largecontiguous sequence stretches, independently of any underlyingblock structure in the region. Even so, somequestions remain in this approach. For example, computationalissues make it difficult to select across very largeregions without some sort of subdivision. In addition, selectingacross large regions may result in a set of tSNPsthat are not optimized for specific subregions, as for example,a candidate gene (Goldstein et al. 2003). These arejust some of the issues that will need to be addressed inorder to develop appropriate, efficient strategies forgenome-wide tSNP selection.