Whole genome association studies in complex diseases: where do we stand?

Abstract

Hundreds of genome-wide association studies have been performed in recent years in order to try to identify common variants that associate with complex disease. These have met with varying success. Some of the strongest effects of common variants have been found in lateonset diseases and in drug response. The major histocompatibility complex has also shown very strong association with a variety of disorders. Although there have been some notable success stories in neuropsychiatric genetics, on the whole, common variation has explained little of the high heritability of these traits. In contrast, early studies of rare copy number variants have led rapidly to a number of genes and loci that strongly associate with neuropsychiatric disorders. It is likely that the use of whole-genome sequencing to extend the study of rare variation in neuropsychiatry will greatly advance our understanding of neuropsychiatric genetics.

In recent years, hundreds of genetic association studies have sought to explore the relationship between common genetic variation and disease, biological characteristics, or drug response. The basic premise of these studies is that the diseases (or traits) are not caused by single gene variants of strong effect, such as, for instance, sickle-cell anemia or cystic fibrosis, but rather that some “manageable” number of common variants have an important influence on the trait under question. Part of the motivation for this perspective is the “common disease, common variant” (CDCV) theory^1,2 Once a genetic variant has been found to be associated, there are a number of possible uses for the information. If the effect of the genetic variant is strong enough, perhaps in combination with lifestyle or other environmental factors, it might be used to predict risk of the disease. Alternatively, the associated variant(s) may be used to try to predict response to a particular medication. Finally, if the effect size of the genetic variant is very small and thus not useful for either of these purposes, it may still be of use in identifying a disease-associated gene or genetic pathway that could illuminate disease pathophysiology or implicate new therapeutic targets. Here we review the current status of genome-wide association studies, with a particular focus on neuropsychiatric disorders.

Genome-wide association studies

Genome-wide association studies (GWAS), are a way of performing genetic association studies without prior hypotheses about which genes are likely to be involved. To do this, arrays of single-nucleotide polymorphisms (SNPs) that cover the whole genome are used. Although there are thought to be approximately 10 million common SNPs in the genome,³ it is not necessary to genotype each one of these individually to get information about most of them. This is because, due to the way that human populations have migrated and genetic variants have arisen, many of the variants are associated with each other or “linked.” Thus, in European and Asian populations, if you genotype one variant, you are gaining information about 10 to 20 other variants simultaneously. This is called “tagging” (the genotyped variants “tag” the ungenotyped, linked variants), and was brought to the genome-wide scale by the HapMap project, which has genotyped millions of common SNPs in four populations to create a detailed map of how common genetic variants relate to one another.^3-5 A significant motivation for the HapMap project was the idea that common variants make up an important part of the genetic contribution to common diseases (the CDCV hypothesis). While some theoretical arguments were marshaled in support of this hypothesis - and indeed, even before the HapMap project a handful of examples were known - there was no way to know a priori how general the CDCV hypothesis might turn out to be. For this reason, a systematic investigation of common variation was judged by much of the community (including these authors) but not all⁶ to be a sensible beginning to the study of human disease genetics. The result is that a true genome-wide study can be performed by actually genotyping as few as 300 000 to 1 million SNPs.^7,8 However, because so many tests are being performed, it is necessary to obtain a very strongly significant P value to be sure that the result is really significant. This is known as “genome-wide significance” and the consensus is that this should be about 10⁸ or less.⁹ Because the effects sizes of common variants are generally small, it is usually necessary to include a large number of subjects in the study in order to have the power to detect a genome-wide significant P value (Figure 1.)

Figure 1. The power to detect a causal variant that is perfectly tagged by a genotyped marker (assuming dominant model, minor allele frequency=0.2, frequency of disease is 1% and equa numbers of cases and controls). To have a good chance of detecting a variant with a relative risk of 1.2, about 2000 cases and controls are needed .

Major discoveries with GWAS

The success of GWAS has been very variable for different disease areas. Some diseases have found common variants with very strong effects, and managed to track these down to the causal variant. An inspiring example is an intronic variant in BCL11A that was found in two GWAS studies to associate with fetal hemoglobin (HbF) levels in healthy adults,^{10, 11} and also to modify the presentation of (3-thalassemia, and associate with HbF levels in patients with sickle-cell disease.¹¹ This finding was soon followed up with a functional study that showed that the variant associated with high HbF¹² reduced the expression of BCL11A,¹³ and that reduction of BCL11A expression caused increase in levels of gamma-globin in adult human red blood progenitor cells, which led to increased levels of HbF¹³ These findings clearly suggest that BCL11A serves as an inhibitor of HbF production and that directed repression of BCL11A could be developed as a clinical tool to ameliorate the presentation of thalassemias and sickle-cell disease. These findings in turn have led to further understanding of developmental and species-specific changes in globin regulation.¹⁴.

On the less inspirational side, however, other diseases, like hypertension, have been thoroughly and carefully investigated using huge numbers of patients and controls with very little progress.¹⁵ Here we outline some of the highest impact findings of GWAS and where (if anywhere) they have led us.

As might be expected by the laws of natural selection, there are not many common genetic variants that confer a strong predisposition to common diseases. Such variants would be expected to have been selected against, and thus maintained at low population frequencies. However, there are some phenotypes that might be expected to have dodged the purifying effects of selection. These include common diseases that do not onset until old age, and response to drugs that the body has not historically had to interact with. Accordingly, some of the strongest effects of common variants on disease have been found in association with ailments with an onset during the postreproductive years, and with drug response.

Genetic variants that affect late-onset diseases

One of the most well-known genetic risk factors is the E4 variant of the apolipoprotein E gene, ApoE, which greatly increases the risk of Alzheimer's disease (AD) and reduces the age of onset in a dose-dependent manner.^16-18 The effect of this variant is so strong that it was, in fact, discovered before the GWAS era, but it has since been confirmed as the most important predictor of lateonset AD in a number of genome -wide analyses,^19-22 one with fewer than 500 cases and controls reporting a P value of 1 x 10^-40.²¹ However, despite the definitive effects of this genetic variant on AD and the length of time that we have known about it, it is still not clear how the variant mediates its effects,²³ and it has not yet led to improved treatment.

One of the very earliest novel discoveries of GWAS was the association of an amino acid substitution in the complement factor H gene, CFH, with age-related macular degeneration, a very common form of blindness that affects the elderly. This genetic association was found with a tiny sample size: 96 cases and 50 controls, and carrying two copies of the risk variant increases the risk of illness up to 7 times.²⁴ The associated variant does itself seem to be functional, changing the binding properties of the protein, although it is not yet exactly understood how the variant contributes to disease,²⁵ nor how this can be utilized in novel treatments.

A third very strong disease-associated common genetic variant is in the LOXL1 gene in exfoliation glaucoma, another very common form of age-related blindness. The associated variant was discovered in a set of only 75 cases, and individuals homozygous for the risk haplotypes are thought to be at 700-fold increased risk of exfoliation glaucoma when compared with homozygotes of the low-risk haplotype. However, because the risk haplotype is so common, this translates to just a 2.5-fold increase risk from the population average.²⁶ The two variants contributing to the risk haplotype are both protein-coding changes, and the same variants have now been associated with disease in multiple populations,^27-40 suggesting that these are the causal variants, although the degree of penetrance, and the risk haplotype, have been reported to differ in Australia and Japan.^{28,29,35,37,38,41,42} Unfortunately, the very high frequency of the risk haplotype in the general population currently precludes these markers from being used to predict disease, but it is hoped that a better understanding of the role of LOXL1 in optical pathophysiology may lead to advances in treatment.⁴⁰

Genetic variants that affect drug response

Genetic variants affecting drug response can have very strong effects, and often occur in the genes that would be most expected to be involved.⁴³ Thus, pharmacogenetics was one of the more successful areas of genomics before the GWAS area, and a number of strong genetic influencers of drug response have been known for some time.⁴⁴ GWAS have added at least three pharmacogenetic associations of considerable strength and importance.

Flucloxacillin-induced liver injury

Idiosyncratic drug reactions are the most common cause of liver failure in the US.⁴⁵ Flucloxacillin is an antibiotic drug commonly used to treat Staphylococcus aureus infections, but it has a relatively high incidence of causing liver injury (6.1 per 100 000 users) in comparison with other antibiotics such as penicillin.⁴⁶ This has previously led to restrictions on its use.⁴⁶ A GWAS was performed on 51 patients with flucloxacillin-induced liver injury and 487 controls, in which a huge signal was seen for a missense polymorphism in the HCP5 gene (P= 8.710^-33)⁴⁷. Through linkage disequilibrium, the association was traced to the HLA-B*5701 allele, the presence of which increased the likelihood of flucloxacillin-induced liver injury by 80 times.⁴⁷ Since the general frequency of the associated allele in the European population is only about 5%, and it was present in 84% of cases, this variant could potentially be used to screen out people at high risk of liver injury before flucloxacillin is prescribed. However, due to the rarity of the hepatoxicity, this would result in a high false-positive rate. A proposed alternative is to use the genotyping of this variant as a diagnostic marker in suspected cases of hepatoxicity so that the patient can be rapidly switched to alternative antibiotics.⁴⁷

Statin-induced myopathy

Taking statin therapy to reduce the levels of low-density lipoprotein cholesterol has been shown to reduce the likelihood of cardiovascular events, such as heart attack and stroke.⁴⁸ Occasionally, however, statins, particularly at high doses, can cause serious myopathy, which may lead to hospitalization or death.⁴⁹ In August 2008 a GWAS that included only 85 cases and 90 controls revealed a SNP in the SLCOIBI gene, which accounted for more than 60% of cases of myopathy⁵⁰ Carrying one C at this locus increases the risk of statin-induced myopathy by 4.5 times, and CC homozygotes have a 17fold greater risk than TT homozygotes. This has been suggested as a genetic test to identify vulnerable individuals before offering high-dose simvastatin therapy⁵¹

Hepatitis-C treatment response

One of the most recent, and perhaps the most clinically significant, of any GWAS to date is the association of a SNP close to the IL28B gene with response to treatment for hepatitis C.⁵² In this study, Ge et al focused on who is cured by treatment, and found that the good response genotype is associated with a greater than 80% chance of clearance in European- Americans, while the poor response genotype is associated with only about a 30% chance. A follow-up study found that the polymorphism also influences natural clearance of hepatitis C and shows very sharp geographic differentiation.⁵³ This suggests that the variant may be common in the population because the “good response” allele conferred protection against one or more viruses and hence was positively selected. This variant is a very good candidate to use as a pharmacogenetic predictor of treatment response before beginning hepatitis C treatment, since the procedure is long and often associated with adverse effects.⁵⁴

The major histocompatibility complex

Setting aside the old-age or pharmacogenetic associations, many of the strongest reported GWAS associations of common variants with common disease involve markers in the major histocompatibility complex (MHC). These associations are too extensive to discuss in detail in this review, but include autoimmune diseases, infectious diseases, neuropsychiatric disorders, and variability in normal traits such as height.⁵⁵ A number of hypotheses have been put forward to explain why variants conferring disease risk at this locus have been maintained at high frequency in the population. One suggestion is that the disease-associated variants have been selected for because they confer resistance to particular infectious agents, either now or historically. An alternative hypothesis is that each locus that confers risk for one common disease is maintained at high frequency because it confers protection against one or more other common diseases. For example, the HLA gene DQB1*0602, which encodes the β chain for the HLA class II molecule DQ6, is protective against diabetes,⁵⁶ but a strong risk factor for narcolepsy⁵⁷ and multiple sclerosis.⁵⁸

GWAS in neuropsychiatry

Neuropsychiatric traits have been among the most disappointing GWAS results. Despite many GWAS, most associated variants have either not withstood significance correction for multiple testing, or else have failed to replicate. In general, where replicable effects have been found, they have required very large sample sizes and the effects have been small.

There have been some notable success stories, however. Two GWAS have revealed strong and replicable genetic influences on restless legs syndrome (RLS), a condition characterized by an unpleasant and irresistible urge to move the legs, particularly while resting and during the evening and night. Both studies, one on Icelandic individuals and one on a more mixed European cohort, implicated BTBD9. ^59,60 The European study also found an association with two other loci: MEIS1 and a locus encompassing MAP2K5/LBXCOR1 .⁶⁰ The associations with MEIS1 and BTBD9 were quickly replicated in two subsequent studies,^61,62 but the MAP2K5/LBXCOR1 appears to be weaker, showing a borderline significance in one study only⁶² Although the risk associated with MEIS1 and BTBD9 (ranging from 1.5 to 3.7^59,60,62,63 ,606263) is substantially lower than those described above, they do appear to be real and highly significant risk factors for RLS. Nevertheless, the biology underlying the associations remains unclear. The associated variants do not appear to have any obvious function, and a thorough search for putative functional variants in all coding exons and across intron-exon boundaries revealed no obviously causal variant.⁶⁴

Another positive GWAS finding in neuropsychiatry is with narcolepsy, a disorder that causes disrupted sleep patterns, with the patient often feeling excessively tired during the day, and suffering sudden sleep attacks. PreGWAS studies had connected the disorder to an MHC class II antigen called HLA-DQB1*0602, and about 85% of narcoleptics carry this antigen.⁶⁵ However, there remained unexplained heritability. Very recently, a GWAS study was done on 807 cases and 1074 controls, all positive for HLA-DQB1*0602. A significant association of three SNPs in the T cell receptor alpha locus was found, which was then replicated in the same study in 1057 further cases and 1104 controls.⁶⁶ Further analysis showed a single SNP was responsible for the association, although it is not clear whether this variant is itself causal or how it may contribute to disease. This association is of particular interest because it adds considerable weight to the view that narcolepsy is an autoimmune disease, and as such, it would be the first autoimmune disease to be associated with a T-cell receptor locus. This finding also opens up the possibility of immunotherapy as a future treatment for narcolepsy.

Other neuropsychiatric diseases for which definite, replicated effects of common SNPs have been found include schizophrenia, associated with MHC markers, NRGN and TCF4 (12 945 cases and 34 591 controls, ORs=1.24, 1.15,1.23),^67,68 bipolar disorder, associated with ANK3 and CACNA1C (4 387 cases and 6 209 controls, ORs=1.45 and 1.18) ⁶⁹, and autism, associated with SNPs at 5p14.1 (3 101 family members, 204 cases and 6 941 controls, OR=1.19).^70,71 However, all of these were discovered with very large sample sizes and account for very little of the very high heritability of these conditions.

Rare variants

Although studies of common variation in neuropsychiatric disease may be underwhelming, the opposite is true for rare variation. Although the SNP chips used for GWAS comprise only polymorphisms that are reasonably common (~≥5%), their data can be used to find other types of non-SNP variants - specifically copy number variants (CNVs) - with much lower frequency. CNVs are duplications or deletions of large stretches of DNA - ranging in size from just a few hundred base pairs to many megabases. To detect such variants, the intensity data from the SNP chips is examined to determine whether particular stretches of SNPs are less intense than expected (or absent), which would indicate a deletion, or more intense than expected, which suggests a duplication.⁷² Because the CNVs are identified on an individual-by-individual basis, very rare CNVs, even those present in a single individual, can be found. This has allowed us for the first time to examine the role of rare variation in common disease (albeit just a tiny fraction of the total amount of rare variant in a cohort). The majority of investigations of copy number variation to date have been in neuropsychiatric disease and, happily, they have led immediately to real, replicable and very strong associations. A summary of CNVs recently strongly associated with neuropsychiatric disease is shown in Table I.

These variants confer considerable risk, but they are not completely penetrant. Although the specific variants are very rare in the general population, they are occasionally seen in controls (Table I) , and where families have been examined, the variants are often inherited from unaffected or only mildly affected parents.^73-77 Additionally, as can be seen in Table I , many of the variants have been associated with more than one neuropsychiatric condition. This is consistent with the characteristics of neuropsychiatrically-associated rare variants that were found before the GWAS era, such as DISC1 in schizophrenia, which associated with a range of phenotypes from psychiatrically normal to suicide, recurrent major depression, and schizophrenia.⁷⁸ It seems that these variants, rather than predisposing to a specific neuropsychiatric condition, may strongly confer some sort of “neural vulnerability,” the ultimate manifestation of which depends on other interacting genetic and environmental factors. Because, to date, the only rare variants that we have been able to associate with neuropsychiatric illness are very large deletions and duplications, it is not clear whether this lack of specificity will be a general rule, or is somehow related to the size of the lesion. However, there is some evidence from the associations with common SNPs that this is a characteristic of the disease rather than the size of the associated variant. For instance, bipolar-associated common variants in CACNA1 C may also confer risk of depression and schizophrenia.⁷⁹

Table I. Copy number variants (CNV) strongly associated with neuropsychiatric disorders.. Frequencies are given only when the CNV was found in a large case-control study design. *Controls may not have been carefully screened for neuropsychiatric illness. NR, not reportec; Dup, cuplication; Del, Deletion .

CNV	Copy # linked to disease	Seen in schizophrenia patients?	Seen in autism patients?	Seen in epilepsy patients?	Sseen in patients with mental retardation?	Other disorders	Reported in controls*?	Lead candidate genes
1q21.1	Deletion and duplication	Deletion (0.23-0.29)82-84	Yes75,94	Yes77,83,95, unpublished data	Yes95,96	Congenital heart disease77,97,98, micro-and macrocephaly,75,77 neuroblastoma,99, other75	Deletion, 0.02% (8/41, 199),84 frequency for dup unclear	HYDIN paralog77; GJA84
15q11	Deletion	Yes (0.61%)84-100	Yes101	No	Yes102	Deletions of this region cause Angelman and Prader Willi syndrome	Yes, (0.19%) (79/41,194) (0.19%)	CYFIP184
15q13.1	Duplication and deletion	Yes81,83,84,103	NR	NR	NR	NR	no data	APBA2
15q13.3	Deletion	Yes (0.17%-0.27%)83,84	Yes (0.31%)84,86	Yes (1-1.3%)73,74,110	Yes73,106	Various including mild developmental delay, heart defects73	Yes, 0.02% (8/39,800)74,84	CHRNA783,84
16p11.2	Deletion and duplication	Yes94	Yes (0.6%, del only,107 1% dup + del94)	Yes110,111	Yes76,108	Various neuropsychiatric and developmental76,108	Yes, 0.01% del, 0,03% dup94	SEZ6L2110,111
16p13.11	Deletion	Yes81	Yes112	Yes (0.6%)unpublished data	Yes (0.5%)87,111	NR	NR (0/3313),87,96 unpublished data but inherited from unaffected parents87	NDEI81
CNTN4	Deletion and duplication	NR	Yes85,114,115	NR	NR	NR	no data
NRXN1	Deletion	Yes (0.19%)81-83,120-122	Yes85,116-118	NR	Yes122	NR	Yes, 0.04% (17/42054)115

The future for neuropsychiatric genetics

There are two, not incompatible, possible directions for neuropsychiatric genetics research. One approach is to continue searching for common variants of small effect size using much larger cohorts in the tens or hundreds of thousands. This has been suggested as a future direction for schizophrenia genetics.⁸⁰ Although this will require a considerable effort, there are already established worldwide collaborations for schizophrenia,^68,80 so very large collections should be achievable in the relatively near future. The disadvantages of this approach are that if such huge sample sizes are needed to discover them, the effect sizes of the associated variants must be very small (Figure 1), and they will be present at a similar frequency in unaffected controls. This makes further study of the effects of the variants very difficult or impossible. However, proponents of this approach correctly suggest that although the associated variant may have a very small effect, the gene it is in may have a big impact on disease when targeted by novel pharmaceuticals.

A second argument in favor of proceeding with GWAS in very large samples is that neuropsychiatric researchers have long expressed concern that clinical diagnostic criteria do not reflect the biological underpinnings of the disease, and that diseases such as schizophrenia may in fact represent multiple different disorders with different genetic contributors. Thus, only with very large sample sizes would one expect to obtain sufficient numbers of any one genetically homogenous subgroup to obtain a genome-wide significant association. However, as discussed above, all genetic variants that have been associated with neuropsychiatric disease so far seem to be very nonspecific. Where they are found in multiple patients with a single diagnosis (eg, schizophrenia), they do not segregate patients into any clear diagnostic categories either by disease presentation or drug response. Additionally, they tend to associate with multiple neuropsychiatric conditions.

The alternative approach is to further investigate the role of rare variants in neuropsychiatric disease. To date, the only type of rare variation that has been identifiable on a genome -wide scale has been large CNVs, and already we have found many strong associations.^81-87 It is likely that when we can identify the totality of rare variation in an individual using whole-genome sequencing, many more rare variants will be found to be definitely associated with neuropsychiatric illness. Fortunately, this is rapidly becoming a reality, and the first sequencing studies in neuropsychiatric illness are already underway. For confirmation and follow-up, this approach will definitely benefit from very large cohorts collected for GWAS, but the ideal discovery samples will be rather different. With this approach, we hope to find variants with very large effect sizes and high penetrance. This means that it will be much more straightforward to understand how the variants exert their effects and what genetic and environmental factors influence them. To do this, the priority will be patients and relatives that can be reapproached for further study after potentially causal variants have been identified. Additionally, since initial sequencing attempts will be expensive, it is worth, at first at least, selecting patients who are most likely to carry highly penetrant genetic variants. These include severely ill, treatment-resistant patients⁸⁸ and patients with a strong family history of mental illness. Thus, this approach benefits from close collaboration between geneticists and psychiatrists and a thorough understanding of each sequenced patient and his or her relatives.

Although it is hoped that whole -genome sequencing will lead swiftly to a clearer understanding of neuropsychiatric disease, there are many challenges ahead. Not least is a very well-characterized psychiatrically normal control cohort. And, as with any new technology, there are considerable technical challenges, such as the use of whole-genome data to identify copy number variation. However, software is constantly developing and it is doubtful that these will be limiting factors for long.^{89 -92} There are also “genomic” challenges: there are many regions of the genome on which we tend not to focus, such as remote enhancer regions, upstream open reading frames, and chromatin binding sites, which are likely to be functional and affected by rare variation. However, using Mendelian diseases as a model, it is reasonable to expect that many of the most important variants will be in or very close to exons.⁹³ Thus, neuropsychiatric geneticists should be able to gorge themselves on the lowhanging fruit for some time to come.

In summary, there have been many GWAS success stories in which common variants have been found to associate definitely with complex diseases. In most cases, however, the mechanism underlying the association is not well understood, and they have not yet led to strong predictive tests or to novel treatments. Neuropsychiatric disease, in particular, has so far benefited little from large-scale analysis of common variants. Use of GWAS data to examine rare copy number variants, however, rapidly led to multiple strong and highly penetrant associations with neuropsychiatric illness. However, the associated variants are not completely penetrant and tend to be associated with multiple neuropsychiatric conditions. Detailed studies of patients and their relatives will be necessary to understand what factors affect the manifestation of the phenotype. Despite this recent success, we can still only account for a very small amount of the heritability of neuropsychiatric conditions. Further investigation of rare variation using whole-genome sequencing is likely to significantly advance the field.