Population-Wide Generalizability of Genome-Wide Discovered Associations

Genome-wide association studies (GWAS) have started revealing hundreds of associations between genetic variants and phenotypes ( 1 , 2 ). Most associations have robust statistical support in the populations where they have been examined. However, almost all GWAS published to date were performed in populations of European ancestry, and the replication efforts in the original publications also focused primarily on similar populations. Are these discovered associations relevant for populations of different ancestry? In this issue, a study by Yamada et al. ( 3 ) evaluated in Japanese participants 23 single-nucleotide polymorphisms (SNPs) that had been identified in the GWAS era. Of those, 16 had emerged originally from studies of European descent populations, two from African descent populations, and five from studies addressing diverse populations. Seven of these 23 SNPs (five independent loci) showed nominally statistically significant associations ( P < .05) with prostate cancer risk in the Japanese population. The other 16 SNPs were not associated with prostate cancer risk, and in five of those, the point estimates of the odds ratios (ORs) were in opposite direction than previously described. Do these results herald extensive nonreplication vs the original discoveries? Not necessarily. One can estimate the power of Yamada et al. (311 case patients and 1035 control subjects) to detect ORs similar to those previously found, given the allele frequencies observed in the Japanese population. I used the per-allele ORs found in European descent populations ( 1 , 4 , 5 ) in 21 previously evaluated SNPs for a comparison against the Japanese data ( Figure 1 ). Estimated power ( 6 ) at alpha value of .05 ranges from almost 0% for rs4962416, which has a 1% risk allele frequency in the Japanese population, to practically 100% for rs16901979, for which the previously proposed OR was large (1.79) and the risk allele frequency in the Japanese population is 19%. If we sum the power estimates for all SNPs to detect the respective ORs previously seen in European descent participants, we can estimate that if ORs were identical in the Japanese population, Yamada et al. would have found 8.5 nominally statistically significant associations among the 21—not very inconsistent with the seven statistically significant associations actually observed. Correlation of genetic effects and risk allele frequencies in populations of European descent vs the Japanese population. A ) Genetic effects, expressed in the odds ratio scale. B ) Risk allele frequencies. Data are shown for the 21 of the 23 single-nucleotide polymorphisms (SNPs) evaluated by Yamada et al. ( 3 ) that had been previously evaluated in populations of European descent. Odds ratios are per copy of risk allele based on per-allele model. The odds ratio and control group risk allele frequency in the populations of European descent is taken from the earliest entry for the respective SNP in the National Human Genome Research Institute (NHGRI) Catalog of Genome-Wide Association Studies ( 1 ) and from Haiman et al. ( 4 ) (rs13254738, rs6983561, rs7000448, and rs10090154) and Eeles et al. ( 5 ) (rs7920517) when entries were not found in the NHGRI catalog. The odds ratio and control group frequency of the same allele (the one that confers risk in populations of European descent) in the Japanese population is from the study by Yamada et al. ( 3 ). The data are shown by filled circles. The boxed number neighboring each data point shows the estimated power (in percent) at alpha value of .05 to detect the odds ratio of the Europeans also in the Japanese, given the risk allele frequency observed in the Japanese control subjects in the study by Yamada et al. Power calculations with the PS software ( 6 ) are for Fisher exact test based on allele counts; results for χ 2 were either identical or slightly larger (data not shown). For rs4962416, the statistical power was too small to be estimated for standard Fisher application, given that the minor allele frequency is only 1% (power shown here as 0%). Populations of European descent have been sampled either from Europe per se or from North America or Australia. However, when each SNP is analyzed separately, there are clear signs of divergent risks in Asian vs European descent populations. For five SNPs, the ORs differ beyond chance. For rs6983561 and rs4430796, genetic effects are strong in the Japanese population (ORs of 1.81 and 1.51), but weak and inconclusive in European descent populations (ORs of 1.16 and 1.22, respectively) ( 4 ). For rs2735839, rs1859962, and rs9364554, robust associations exist in populations of European descent ( 1 ), but no statistically significant association exits in the Japanese population, and point estimates go in the opposite direction. Several other SNPs may also have considerably different effects in the two populations, but the difference does not reach nominal statistical significance because of limited sample size. Overall correlation between ORs in the two ancestry groups is modest (Pearson correlation coefficient = .45, P = .04; Figure 1, A ). For the same variants, correlation is even weaker for the allele frequencies (correlation coefficient = .30, P = .18; Figure 1, B ). Some risk alleles may confer the same OR in both populations, but their population-specific frequencies may vary substantially. This situation (varying frequency, similar OR) has been previously well documented for candidate gene variants ( 7 ), and the same phenomenon appears now for GWAS-discovered markers, for example, rs16901979 consistently confers an OR of approximately 1.8 in both populations, but the risk allele is more common in the Japanese population than in populations of European descent (19% vs 3%). Conversely, rs721048 confers modest risk in both populations, but it is more common in populations of European descent than in the Japanese population (19% vs 4%). Data are more limited in populations of African descent, which is unfortunate, because the haplotypic structure of African populations would be ideally suited for making specific discoveries ( 8 ). Differences may be even greater in populations of African descent vs other populations ( 4 ). One of the tested SNPs (Broad11934905) originally found in African populations was monomorphic in the Japanese population. These data reinforce the notion that discovered variants in the GWAS setting are often simply population-specific markers that need far more effort to lead to functional culprits ( 9 ). However, even functional culprits may differ across populations, and some may exist only in specific populations. Population diversity could have implications about the utility of GWAS-derived information for predictive purposes. Despite high expectations, discovered variants contribute small quanta to the disease risk of each individual. Using single or very few variants for predictive purposes is inefficient and useless or even misleading for any population ( 10–12 ). The question is whether predictive ability may reach sufficient performance once many such variants are considered ( 13 ). Would results be generalizable across populations? Yamada et al. used information from the five nonoverlapping SNPs in the Japanese population to construct a predictive model in which men with six or more risk alleles (top 11%) had 6.22-fold greater odds of having prostate cancer compared with those with two or fewer alleles (lower 20% of control subjects). This discriminatory ability seems superior to what has been described in populations of European descent ( 14 ), among whom having seven SNPs resulted in an OR of approximately 3.5 for the top 1% vs the lower 1% of genetic risk and among whom was estimated that 15 SNPs would barely achieve an OR of 2.1 for the top 10% vs the population average. The higher OR found by Yamada et al. may mean that better discrimination is feasible to achieve in the Japanese population or the effect sizes of the five nominally statistically significant SNPs are inflated compared with their true values. When markers are selected based on their ability to pass a desired threshold of statistical significance, their observed effect sizes are on average larger than the true estimates ( 15 , 16 ). Understanding the exact ability of these markers to discriminate between prostate cancer case patients and control subjects has to wait for yet another independent validation sample. Ideally, understanding precise effect sizes for these variants and their combined effects will require large studies and extensive replication in diverse populations ( 17 ) and careful synthesis of all the accumulated evidence ( 18 ). Although we have found many promising markers for prostate cancer and other diseases through GWAS, and these have been replicated in similar populations, this is just the beginning. The findings need to be replicated again and then again validated.