Predicting the future of genetic risk prediction.

Recent genome-wide association studies (GWAS) have identified many susceptibility loci for a variety of complex traits. There is intense interest in evaluating the potential utility of these variants for individualized risk prediction for chronic diseases such as breast cancer. A recent article in this journal (1) evaluated the potential discriminatory accuracy of breast cancer risk models that use a variety of genetic variants, some of which, but not all, were found through GWAS. We comment on the methodology and conclusions of that paper and compare it with other work that has assessed the potential role of genetic variants for risk prediction and considered the possible impact of discoveries of new variants in future studies. The paper by van Zitteren et al. (1) used simulations to estimate the area under the operating characteristic curve (AUC) from breast cancer risk models based on single nucleotide polymorphisms (SNPs). From a review of meta-analyses and GWAS, they found 96 variants, 41 of which were nominally significant at the 0.05 level. They estimated the AUC as 0.67 based on these 41 variants and 0.68 based on all 96 variants. They also commented on the numbers of additional variants that would be needed to increase the AUC to a higher value, such as 0.80. The work of van Zitteren et al. is interesting and surprising in several respects. Previous studies based on SNPs associated with breast cancer in GWAS yielded estimates of AUC of 0.574 for 7 SNPs (2), 0.579 for 10 SNPs, (3) and 0.585 for 11 SNPs (4). The SNPs in these studies attained genome-wide significance (P < 10 ) and odds ratios per allele were based on independent data, thus reducing or eliminating the bias found in odds ratios in the discovery phase of GWAS, known as the "winners’s curse" (5). It is very difficult to increase the AUC from 0.58 to 0.67 (2, 6). Park et al. (7) examined the prospects for increasing the AUC based on further GWAS. They estimated the likely number and the distribution of effect sizes for yet to be discovered SNPs based on the effect sizes for known susceptibility loci and the power of the original GWAS that led to these discoveries. For breast cancer, they estimated that within the range of effect sizes of known loci, there could be a total of 67 common susceptibility SNPs, including those already identified, which is almost 4 times the number of known SNPs. However, the incremental contribution of the undetected SNPs to the genetic variance is likely to be modest as their effect sizes will tend to concentrate toward the lower end of the range. In particular, Park et al. estimated that the total of 67 susceptibility SNPs could explain a total of 17.1% of genetic variance of breast cancer and could lead to an AUC of only 0.635. Because detection of most of these additional SNPs at a genomewide significance would require very large sample sizes, they concluded that an AUC of 0.635 would be the practical upper limit of the discriminatory power of risk model for breast cancer on the basis of common susceptibility SNPs alone. Thus, it is impressive that van Zitteren et al. report an AUC of 0.67 based on 41 nominally significant variants. On the basis of a log-normal approximation of risk (8), we estimated that to achieve an AUC of 0.67, a genetic risk model will need to explain 27.9% of known heritability of breast cancer corresponding to a sibling recurrence risk of 2.0. In contrast, the 18 SNPs found in GWAS studies in Table 1 account for only 7.1% of the heritability. There has been recent speculation (9) on reasonswhy the SNPs from GWAS account for such a small portion of the theoretical heritability that one would anticipate from studies of familial aggregation of breast cancer (8, 10). Some possibilities are variants that contribute to risk but are not tagged by common SNPs in standard GWAS analyses. Such variants might include deletions, repeats, copy number variations, and uncommon and rare SNPs that are not in tight linkage disequilibrium with SNPs accessible with a GWAS platform, such as the Illumina Infinium 660K array. Thus, it is possible that van Zitteren et al. have tapped into some of the "missing heritability" to achieve an AUC of 0.67 from nominally significant variants on the basis of meta-analytic literature; some of these variants are deletions and repeats that might not be detected in GWAS. We will return to this possibility in discussing Table 1, which classifies the variants in van Zitteren et al. partly in terms of the ability of GWAS to detect them. Another possibility is that some of the variants used by van Zitteren et al. are not truly associated with breast cancer and represent chance findings with possibly exaggerated odds ratios. Many of the meta-analyses cited by van Zitteren et al. focused on a "candidate gene." Findings from many early candidate gene studies were not reproducible, and a number of authors sought to explain why (11, 12). Important reasons for these failures were the Authors' Affiliation: Division of Cancer Epidemiology and Genetics, National Cancer Institute, National Institute of Health, Rockville, Maryland