During the past 25 years, genetic linkage-based studies have proved very effective in identifying causal genetic factors in Mendelian (single gene) disorders; causal genes for more than 1,300 dominant and recessive Mendelian diseases have been identified¹. Most common diseases and endophenotypes, however, do not exhibit Mendelian inheritance, but rather feature complex, multifactorial expression and inheritance. Although linkage-based methods have been broadly applied, these studies have had little success in identifying the allelic determinants of common disorders². In particular, there has been poor replication among studies, whereby an initial study identifies an allele (genotype) with large estimated genetic effects (relative risk) but subsequent studies fail to corroborate the results^3,4. In part, this reflects the dependence of linkage-based studies on unusually informative families (with multiple affected and unaffected individuals), which induce a bias toward rare, semi-Mendelian disease subsets in subpopulations. Reports of successful identification of genetic variants in common diseases using an approach that circumvents this limitation — genome-wide association (GWA) studies — have therefore generated considerable excitement.

Human GWA studies are based on three hypotheses: First, the common trait/common variant hypothesis proposes that the genetic architecture of complex traits consists of a limited number of common alleles, each conferring a small increase in risk to the individual^5,6; second, the brief history of most human populations precludes sufficient generations (or meioses) to create recombination events (or mutations) between closely located, common (ancient) variants; and, third, suppression of meiotic recombination (coldspots) occurs very frequently. Thus, approximately 80% of the human genome is comprised of around 10 kb regions that exhibit reduced recombination in human populations (haplotypes)⁷. Genetic variants (alleles) within haplotypes are in linkage disequilibrium (LD). This phenomenon enables much of the recombination history in a population to be ascertained by genotyping a large set of well-spaced, common (ancient) variants throughout the genome, especially if variant selection is informed by knowledge of haplotypes. During the last 10 years, more than 10 million single nucleotide polymorphisms (SNPs) have been identified⁸. Furthermore, the International HapMap project has genotyped approximately 4 million common SNPs (occurring with a minor-allele frequency of more than 5%) in human populations and has assembled these genotypes computationally into a genome-wide map of SNP-tagged haplotypes⁷. These resources, together with array technologies for massively parallel SNP genotyping and the well-established epidemiological case-control association studies have rendered GWA feasible (BOX 1, FIG. 1).

Initial genetic association studies focused on candidate loci and exhibited a lack of replication among studies^9,10. There were biological explanations for inconsistent results: unobserved, confounding biological sources of heterogeneity, including inconsistent or poorly defined measurements of the phenotype, heterogeneous genetic sources for the phenotype (genocopies), population stratification (ethnic ancestry), population-specific LD, heterogeneous genetic and epigenetic backgrounds or heterogeneous environmental influences (phenocopies). In addition, there were statistical reasons for irreproducibility, including failure to control the rate of false discoveries, model misspecification and heterogeneous bias in estimated effects among studies^11–14. Also, a frequent source of non-replication was lack of power due to the limited number of individuals genotyped and phenotyped^15,16.

In order to ameliorate poor replication, GWA experiments employ multi-tiered experimental designs with discovery, replication and biological validation stages¹⁷ (FIG. 1). Tiered designs are critical for cost-effective detection of meaningful, hypothesis-generating, genotype–phenotype associations given the large number of comparisons involved, prior probability estimates of association, sample sizes, resampling procedures and statistical significance thresholds. GWA studies also owe their statistical power to their large cohort size and high rate of SNP detection. Currently, a respected threshold for uncorrected, significant associations is P <5 × 10⁻⁷ (REFS ^18,19). Alleles with moderately less significant associations, however, are often also reported, as they might indicate loci that reach the aforementioned threshold in subsequent studies.

A second, pioneering GWA study examined age-related macular degeneration (AMD)³⁶ (TABLE 1). The discovery phase sought associations of 105,980 SNPs with AMD in 96 cases and 50 control individuals. Despite the small cohort size, SNPs in the complement factor H (CFH) gene, including an nsSNP, showed significant association with AMD. Replication was not performed, but subsequent studies have replicated associations of CFH alleles with AMD^37–40. Of all common diseases examined by GWA to date, AMD is unique in that a single haplotype explains 61% of the genetic variance, conferring a homozygous odds ratio of 7.4. To put this in perspective, this is of a similar magnitude to the classic associations of HLA-B27 with anterior uveitis/ankylosing spondylitis and HLA alleles with type 1 diabetes mellitus (T1DM). Complement pathway dysregulation was a novel, unexpected association with AMD. Subsequent studies have shown an association of AMD with two additional members of the alternative complement pathway (factor B (CFB) and C3)^41,42. These findings, together with biological validation studies, have led to the initial development of new AMD therapies, based upon complement inhibition.

In the past year, technical challenges associated with GWA were largely overcome, genotyping costs were decreased and a significant number of studies have used SNP genotyping arrays in larger population groups to produce replicated associations between individual SNP alleles and common diseases.

As well as identifying novel associations, GWA studies have confirmed several susceptibility genes that were previously established by linkage analysis in large pedigrees. For example, a GWA study of late-onset Alzheimer’s disease (LOAD) identified the well-established APOE-susceptibility allele⁹⁰. This association was also replicated in a study that genotyped 17,343 putative functional cSNPs²³.

A remaining problem with large GWA studies is the cost of genotyping, but one study provided evidence that sample pooling strategies might help to overcome this issue. In a GWA study of bipolar disorder, investigators created 39 pools, containing DNA from 2,672 individuals⁹¹. These pools were used for both discovery and replication experiments. Pools were individually genotyped for 555,235 SNPs and normalized allele frequencies were inferred from intensity data. Replicates were assayed for each pool. Thirty-seven SNPs showing allele frequency differences in both cohorts were individually genotyped and one SNP retained a significant association. The aforementioned WTCCC study also studied bipolar disorder, identifying an association at 16p12 (REF. ¹⁸). One locus, for glutamate receptor, metabotropic 7 (GRM7), showed association in both studies.

Encouragingly, most associated haplotype intervals identified to date are sufficiently small to feature a single gene. In large measure, this reflects the use of several, outbred populations in confirmatory, fine-mapping studies. Even when the association is within a single gene, the predisposing variant might affect an adjacent gene, as in adult lactose intolerance⁹⁹. Although some association intervals have been found to contain a single, unequivocally functional gene variant, the causality of alleles has been established in only a minority of cases. Causal alleles identified to date do not yet show much difference in genetic mechanism from those identified in Mendelian disorders; this could reflect ascertainment bias^1,83,84.

Many genes identified by GWA were not candidate genes previously, highlighting the hypothesis-informing value of genetic studies. Already, there are examples of potentially tractable therapeutic targets that had not previously been considered in a disease or trait. As yet, the confluence of associated genes into biological networks and pathways is at an early stage. In part, this reflects scant or incorrect annotation of many genes. There appears to be a significant conservation of associations of common alleles between human populations. Thus, to date, it appears that GWA studies are fulfilling expectations with regard to the elucidation of molecular mechanisms underpinning poorly understood, common diseases.

In the few informative studies reported to date, endophenotypes have been highly instructive in dissecting the network or pathway that is perturbed by an individual allele, which affects a complex trait. It is particularly exciting to see the application of multi-mode endophenotypes, such as combinations of psychological testing, brain imaging and gene expression⁸⁷. This is clearly an area of potential opportunity.

The cost of enrolling the very large cohorts that are needed to discover and validate alleles with small effect sizes has hitherto precluded the collection and integration of rich, accurate clinical metadata. It is likely that future studies will use a much greater stratification of traits than the phenotypically crude studies reported so far. Recent GWA studies of breast cancer provide a good example of the added genetic complexity that can be revealed by trait stratification^19,85,86. In addition, following replication of associations with categorical traits, it is anticipated that targeted genotypic examination of many endophenotypes will be highly instructive in the dissection of the role of individual alleles in disease pathogenesis.

GWA studies show significant potential to redefine disease classification. In some cases, GWA studies are identifying molecular factors that enable patient stratification and might prove useful in personalized medicine. Cancers provide the clearest examples of this to date. In other cases, exemplified by IBD, GWA studies are pointing to common molecular underpinnings in diseases that were believed to be distinct. In restless leg syndrome, replicated associations have provided concrete evidence that the phenotype represents a bona fide neurological disorder^100,101. In mental illness, there is great anticipation that GWA studies will provide an objective, molecular revision of disease categorization.

As noted above, phenotypes studied to date have been crude. The use of endophenotypes is expected to increase significantly. In particular, biomarker phenotypes are anticipated to become widely used. These will probably include gene expression, proteomic, metabolomic and imaging biomarkers. As determinants of complex traits are identified, genetic stratification will become possible, potentially reducing the genetic complexity of traits and enabling the identification of additional association signals. An example of this was the recent use of periodic limb movements and serum ferritin levels in GWA studies of restless leg syndrome¹⁰⁰. An area of substantial future interest for the pharmaceutical industry will be pharmacogenetic GWA studies to identify markers for patient stratification in clinical trials. Comprehensive pharmacogenetic information will, in turn, facilitate the practice of personalized medicine. Pharmacogenetic GWA studies and early adoption of personalized therapy are likely to be used in the selection of expensive or chronic medications in life threatening conditions or where the therapeutic index is narrow or adverse event concerns are high, such as cancer chemotherapy.

Despite the current excitement, GWA studies have only been able to account for a small proportion of the expected genetic variance in complex traits^24,102. This is not surprising given current limitations. First, current GWA studies are designed to identify common risk alleles that are predicted to be important in complex disorders under the common disease/common alleles hypothesis^5,6. Increasing evidence suggests that some complex disorders and traits, such as schizophrenia, hypercholesterolaemia and body mass, are genetically heterogeneous^104–107. The genetic basis of such diseases is more likely to conform to the common trait/rare variant hypothesis, which proposes that many rare variants exist, with substantial allelic heterogeneity at causal loci^58,108,109. The GWA approach is unable to detect susceptibility loci that harbour numerous, individually rare (recent), polymorphisms. Instead, a resequencing approach will be needed to identify rare alleles. Encouragingly, massively parallel sequencing methods provide a potential solution^102,107,110, suggesting disease-specific rare alleles and recent mutations that provide supplementary genotyping array content. Second, a proportion of the genome cannot effectively be examined on the basis of tag SNP genotypes. Approximately 20% of the genome is comprised of recombination hotspots that are not amenable to LD-based approaches⁷. Alternatively, at recombination coldspots, haplotype blocks might be too large for unambiguous identification of causal loci. The extent of the effect of genomic copy number variation (CNV) on association signals is not yet clear, although recent genotyping arrays do provide CNV information. Insufficient numbers of cases will be available for GWA studies of many orphan diseases, uncommon disease complications or adverse events. For some common diseases, these considerations could obfuscate a substantial proportion of the genetic variance. Supplementation of genotyping array content reflective of CNV regions should, however, circumvent some of these limitations. Use of adaptive statistical methods and resampling strategies might also circumvent the need for thousands of affected individuals in studies of orphan diseases⁸³.

GWA successes are creating substantial need for downstream genetics, biochemistry and cell biology efforts to confirm the biological relevance of genotype–phenotype associations and to elucidate the underlying mechanisms of disease. This is especially true of association signals in gene deserts or alleles without apparent functional consequence. Translation of the fruits of GWA studies to clinical practice will require the derivation of predictive models of the genetic architecture of complex traits that evaluate with much greater precision the contributions of factors such as epistasis, genocopies, phenocopies and penetrance.