Drinking from the fire hose--statistical issues in genomewide association studies.

T he past 3 months have seen the publication of a series of studies examining the inherited genetic underpinnings of common diseases such as prostate cancer, breast cancer, diabetes, and in this issue of the Journal, coronary artery disease (reported by Samani et al., pages 443–453). These genomewide association studies have been able to examine interpatient differences in inherited genetic variability at an unprecedented level of resolution, thanks to the development of mi-croarrays, or chips, capable of assessing more than 500,000 single-nucleotide polymorphisms (SNPs) in a single sample. This " SNP-chip " technology capitalizes on a catalogue of common human genetic variations that is provided by the HapMap Project, which was made possible by the completion of the consensus human-genome sequence. 1 The amount of data in these studies is four to five orders of magnitude greater than that in the previous generation of case– control studies, which tested only a handful of variants, often in a specific candidate gene. This unprecedented volume poses unusual statistical challenges for the analysis, display, and interpretation of the data. The chief strength of the new approach is that it permits an " agnostic " genomewide comparison of gene-variant prevalence between cases and controls, obviating the need for guessing which genes are likely to harbor variants affecting risk. Most of the robust associations seen in this type of study have not been with genes previously suspected of being related to the disease. Some of these associations have been found in regions not even known to harbor genes, such as the 8q24 region , in which multiple variants have been found to be associated with prostate cancer. 2 Such findings promise to open up new avenues of research, through both the discovery of new genes relevant to specific diseases and the elucidation of new genetic mechanisms (e.g., the mechanism explaining why a region without known gene-coding loci would be associated with a disease). The chief strength of the new approach also contains its chief problem: with more than 500,000 comparisons per study, the potential for false positive results is unprecedented. One proposed solution is to adopt the approach conventionally used in much medical research — choosing a stringent P value at which statistical significance will be declared. To address the 500,000 or more comparisons , a Bonferroni approach can be used; for example, one can divide the commonly used P value of 0.05 by 500,000 …