Admixture mapping has been proposed as an alternative to traditional linkage and association studies and in theory holds great promise for selected traits 
. In the first application of this strategy to mapping complex traits, Zhu et al. performed a genome-wide admixture mapping of hypertension in an African-American sample based on a set of microsatellite markers designed for traditional linkage analysis and identified two regions on chromosome 6 and 21 that potentially harbored hypertension susceptibility genes 
. Reich et al. 
also reported a locus on chromosome 1 associated with multiple sclerosis by admixture mapping. More recently, linkage analysis followed by fine mapping identified a chromosome region harboring genes for susceptibility to prostate cancer 
and this locus was subsequently replicated with admixture mapping 
. In this report we present additional evidence from genome-wide admixture mapping for hypertension in 1743 unrelated African Americans using a highly informative SNP panel 
. This panel extracted on average 89% of the information of ancestry (measured by SIC) 
for the African-American sample and as a consequence the results from both case-only and case-control tests were reasonably robust. Our estimate of the average information for ancestry is higher than was found in the original report for this SNP panel 
. This may be explained by the fact that we assumed a 50%–50% mixture when no SNPs are genotyped while Smith et al. assumed a 79%–21% mixture. Under the latter assumption the estimated average information was 65%, slightly less than the 71% suggested by Smith et al. 
. The estimated ancestral allele frequencies in the parental populations are close to the corresponding observed frequencies in contemporary African and European populations, suggesting that our model fits the data well. We observed 6 regions with consistent excess of African or European ancestry in both case-only and case-control analyses, including the region on chromosome 6 identified in our previous study 
. However, simulation studies indicate that we did not find any region that reaches genome-wide significance in this study. Although the regions on chromosome 6 and 21 identified in our previous study are large (37cM and 30 cM, respectively), our simulations indeed suggest that the present result is unlikely to be due to chance. It is interesting that we observed stronger evidence but less significant association with the African-derived allele on chromosome 6 than on chromosome 21. Several reasons may explain this outcome. 1) The power of admixture mapping is dependent on the underlying genetic model, with the recessive mode of inheritance being more powerful than additive or dominant modes 
. A recessive effect for the T allele of missense SNP rs2272996 in VNN1 gene best explains the evidence, suggesting it is reasonable to observe better admixture mapping evidence on chromosome 6 than 21. 2) The power of admixture mapping is dependent on the disease allele frequency difference between two ancestral populations, while the power of association is dependent on the linkage disequilibrium between the underlying disease variant and the test marker, which is the ancestry in this case, and the disease allele frequency in the admixed population. We performed a power analysis by assuming that the missense SNP rs2272996 is the true causative variant. We further assumed that the odds ratio of the TT genotype vs others in African ancestry population is 1.57, as estimated in . Since we failed to observe evidence of association in the European American population, the odds ratio was simply placed at 1.0, which leads to the estimate of the relative population risk ratio of 1.38. Under these assumptions, our power analysis suggests that, post hoc, we had 38% power to detect a region reaching the genome wide significance with the sample size in the current study. Thus, our study clearly has relatively low power to find similar locus. 3) A spurious finding is still possible because of the statistical fluctuation despite our previous report. Further association analysis is necessary to rule out the possibility that this is a false positive finding.
We performed our analysis based on a method which directly maximizes the likelihood function from the hidden Markov Model using the EM algorithm, allowing for uncertainty in model parameters, such as the allele frequencies in the parental populations 
. Simulations suggested that this method can perform as well as the widely used Bayesian MCMC method STRUCTURE for the data generated from various population admixture models.
However, we believe the results can be regarded with greater confidence if different approaches lead to similar results, as suggested by others 
. We thus performed similar analysis using STRUCTURE, although we are also aware that other similar MCMC based software has become available 
. The difference between ADMIXPROGRAM and STRUCTURE includes the approach to estimation of a large number of parameters, eg, ADMIXPROGRAM directly maximizing the likelihood function based on EM algorithm while STRUCTURE uses a Gibbs sampling scheme. The other difference involves the transition probability in the Hidden Markov model. ADMIXPROGRAM uses the transition probability derived from a continuous-gene-flow model and STRUCTURE uses linkage model by assuming that chunks of chromosomes are derived from ancestral populations and the breakpoints between successive chunks occur randomly 
. Despite the substantial difference of the two approaches, the results are almost identical, providing further confidence that our results are not biased of the selection of the statistical method.
Several publications 
have demonstrated that admixture mapping can be seriously biased by background LD between adjacent SNPs when this phenomenon is not properly considered. We thus examined the background LD in European Americans and dropped the SNPs that are in strong LD. Since the maker panel was initially selected to minimize background LD, eliminating additional selected markers based on this characteristic had a very limited effect. We further repeated the analyses using different sets of SNPs: (eg, odd or even SNPs , those with adjacent distance>1 cM) and the results on chromosome 6 were essentially the same. We note, however, that more information might have been obtained by keeping those SNPs that were in LD when using the Markov Hidden Markov Model method recently proposed by Tang et al.
We next followed up the region on chromosome 6 by genotyping 51 missense SNPs in 36 genes spaced across the 18.3 Mb region. Searching for disease variants by testing all functional SNPs has been advocated by Risch 
and Risch and Botstein 
and this strategy successfully identified the sixth type 1 diabetes locus by examining all functional SNPs across the genome 
. We adapted this strategy in the chromosome 6 region and identified a missense SNP rs2272996 (or N131S) in the VNN1 gene significantly associated with hypertension in African Americans after adjusting for multiple comparisons; this association was replicated in Mexican Americans. A non-significant association in the opposite direction was observed in European Americans. (No correction was made for multiple comparisons in Mexican Americans and European Americans since we tested only rs2272996 in these two populations.) Further analysis also indicates that this SNP accounts for most of the evidence of excess of African ancestry observed in this region for African Americans. Thus, the association of SNP rs2272996 to hypertension is unlikely to be due to chance alone, although additional replication studies in independent studies are necessary. It should also be noted that our search of causative variants was not comprehensive and other susceptibility variants may well exist in this region.
We were also puzzled by the direction of the risk estimate associated with allele T in different populations. Theory suggests that the divergent risk relationship at this locus in ancestral populations will increase the power of admixture mapping method. However, the different patterns of risk in these populations cannot be explained with the data currently available, but could reflect gene-gene or gene-environmental interactions 
. It is possible that this locus was under different selective pressure in different populations, however, the evolutionary processes that have molded susceptibility to chronic cardiovascular disease have not been defined. Whether the at-risk alleles arose under positive selection or a neutral-equilibrium model therefore cannot be determined. The suggestion has been made that susceptibility variants for hypertension may have been under selection pressure due to the climate adaptation, however most of the loci identified under this assumption have not been replicated in association analyses 
. Under this model, the variants are unlikely to be deleterious and could be common. Voight et al.
proposed a method to detect signals of very recent positive selection in the human genome using HapMap data and the VNN1 gene is located in one of the regions with strong positive selection pressure in YRI sample. VNN1 belongs to the vanin family of proteins, including secreted and membrane-associated proteins which have been reported to participate in hematopoietic cell trafficking and to possess pantetheinase activity which may play a role in oxidative-stress response 
. It has been recently suggested that increased oxidative stress may antedate hypertension and contribute to its pathogenesis 
. More recently, VNN1 has been suggested as a novel gene for cardiovascular disease risk, strongly associated with expression levels of several lipid metabolism/CVD-risk genes, 
although the underlying pathway of VNN1 and other CVD-risk genes remains unknown.
In summary, we conducted a genome-wide search for susceptibility loci for hypertension using admixture mapping in African Americans. Further association studies identified a missense SNP rs2272996 in the VNN1 gene that may explain the evidence identified through admixture analysis. While requiring further confirmation, this finding demonstrates the potential for the admixture approach and suggests it could be a tool for use in studies to define genes affecting selected complex traits. In addition, our study only surveyed the non-coding SNPs on the region on chromosome 6. A more comprehensive assessment of variants in this region will be required to provide assurance that the causative variants associated with hypertension have been identified.