Genome-wide association studies (GWAS) usually contrast the frequencies of genetic variants between cases and controls for a large set of genetic markers distributed across the genome (see Corvin et al for an introduction to these methods Reference Corvin, Craddock and Sullivan1 ). These studies are notable for the large numbers of genetic makers (usually 500 000 to 1 million) and large sample sizes (often >10 000 participants). Since 2005, 1050 GWAS of human diseases and biometrical traits have been published. 2 These studies of 575 phenotypes implicated 2881 genetic variants at stringent levels of significance, often with compelling independent replication. Since 2007, 115 GWAS that focused on psychiatric disorders have appeared, most on Alzheimer's disease, bipolar disorder and schizophrenia, along with efforts in attention-deficit hyperactivity disorder (ADHD), autism, major depressive disorder, and licit and illicit drug use and dependence. Given these efforts, it is timely to ask what we have gained from this body of work.
The efforts of GWAS have directly or indirectly led to important changes in the conduct of genetic studies. First, large-scale collaboration has become the norm in psychiatric genetics. Because the sample sizes needed for GWAS discovery and replication are beyond the reach of single groups, multiple consortia have emerged to foster scientific discovery. As an example, the Psychiatric Genomics Consortium (PGC) 3 has approximately 300 investigators and more than 75 000 participants with GWAS data under analysis, and may be the largest consortium in the history of psychiatry.
Second, prompt sharing of results and full genomic data is now standard. This welcome change maximises progress towards understanding the genetic basis of critically important psychiatric disorders, and fosters reproducibility by allowing the independent evaluation of claims of association. For example, anyone can view or obtain the results from the PGC studies, and qualified investigators can obtain individual-level data in order to conduct additional analyses (see Appendix).
Third, as critically, uncompromising statistical rigor is now required. Genomic studies must explicitly account for the 105–106 statistical comparisons. Thresholds for declaring significance are severe but appropriate, and most journals require replication in independent samples. Genome-wide association studies can observe dozens of associations with P ∼10–6–10–7 resulting from the play of chance. Most investigators now understand that ‘intriguing biology’ is irrelevant to establishing a robust genetic association. For example, we observed multiple associations of major depressive disorder with PCLO (piccolo) with P ∼10–7. Reference Sullivan, de Geus, Willemsen, James, Smit and Zandbelt4 The biology of PCLO is fascinating (among its other functions, PCLO ‘touches’ serotonin); however, this ‘intriguing’ association did not withstand the test of replication. Reference Sullivan, de Geus, Willemsen, James, Smit and Zandbelt4,5 The community standard in human genetics now requires significance well beyond chance plus replication in independent samples. The biology of a gene does not play a role in establishing the association.
Discoveries of genetic loci
The central goal of psychiatric genetics is to discover loci that are robustly and repeatedly associated with a disorder and thereby gain insight into aetiology (Table 1). Approximately 11 copy number variants (CNVs) have been discovered (included in this review as most were discovered using GWAS technology). These CNVs are rare (<0.1% in controls), potent (odds ratios (ORs) of 4–20), and often non-specific risk factors for psychiatric disorders. Reference Sullivan, Daly and O'Donovan6 In addition, GWAS have strongly implicated common variation in approximately 30 different genomic loci for psychiatric disorders along with loci for nicotine (CHRNA3, BNDF, CYP2A6) and alcohol consumption (ADH1B, ALDH2, AUTS2). These genetic variants are relatively common (allele frequencies >5%) and subtly increase disease risk (ORs of 1.10–1.25).
Associations popular in the literature before 2007 (such as COMT, DRD3, DRD2, HTR2A, NRG1, BDNF, DTNBP1 and SLC6A4) have generally not fared well in GWAS. Reference Collins, Kim, Sklar, O'Donovan and Sullivan7 The reasons why these candidate genes did not replicate are unclear. Although they cannot be definitely excluded from playing an aetiological role, their involvement now seems far less likely. A few associations have stood the test of time, and are limited to loci with unusually large effect sizes (Alzheimer's disease–APOE, schizophrenia–22q11.21, and alcohol dependence–alcohol metabolic genes).
More complex analyses
The basic analytic model used in most GWAS is very simple and considers single genetic markers in isolation. This simple model is not optimal given empirical data that psychiatric disorders are polygenic, and analyses of sets of markers could provide further insight by better reflecting the fundamental genetic architecture. Of the ways in which multiple markers can be analysed in combination, three have been the focus of particular effort (Table 1). First, large sets of genetic markers can be used to estimate heritability. Unlike twin or family studies whose assumptions continue to be criticised, these approaches yield assessments of heritability based directly on the genome. For schizophrenia and bipolar disorder, these results confirm that substantial proportions of the variance in liability (about 25%, or approximately a third to a half of the heritability) are accounted for by the current generation of genotyping arrays. Reference Visscher, Brown, McCarthy and Yang8 Thus, the critical assumption that has driven a generation of genetic studies now seems particularly secure.
|Finding, location or SNP||Band, nearest gene or finding||Class||Disorder|
|chr2:50.1-51.2||2p16.3||Rare, potent||ASD, SCZ|
|chr15:30.9-33.5||15q13.3||Rare, potent||ADHD, ASD, SCZ|
|chr16:29.5-30.2||16p11.2||Rare, potent||ASD, SCZ|
|chr17:34.8-36.2||17q12||Rare, potent||ASD, SCZ|
|chr22:18.7-21.8||22q11.21||Rare, potent||ASD, SCZ|
|rs610932||MS4A cluster||Common, subtle||AD|
|rs2075650||APOE, TOMM40||Common, notably strong||AD|
|miR-137 network||Common variation||SCZ|
|Cholesterol, innate immune||Common variation||AD|
|Calcium signalling||Common variation||BIP|
|Post-synaptic signalling||Rare CNVs||BIP, SCZ|
|Increased, multiple studies||Rare CNVs||ASD, SCZ|
|Increased||Rare CNVs||ADHD, BIP, MDD|
|Increased||Common variation||SCZ, BIP, MDD|
CNV, copy number variation; SNP, single nucleotide polymorphism; AD, Alzheimer's disease; ADHD, attention-deficit hyperactivity disorder; ASD, autism spectrum disorder; BIP, bipolar disorder; MDD, major depressive disorder; SCZ, schizophrenia.
a. See Tables 2 and 3 in Sullivan et al Reference Sullivan, Daly and O'Donovan6 for full citations. The CNV and SNP findings meet genome-wide significance in large samples. Most are likely secure but some may not stand the test of time. The heritability, pathway and burden results have replicated in multiple samples and/or represent consistent results from different analytical methods. Genomic locations are NCBI Build 37/UCSC hg19.
Second, pathway analyses evaluate whether associations in predefined sets of genes have smaller P-values than expected. Reference Sullivan, Daly and O'Donovan6 These analyses have provided new ideas about the biology of these disorders as hypotheses for future work, and include: micro-RNA miR-137 and multiple genes containing binding sites for miR-137 in schizophrenia; calcium signalling in bipolar disorder; cholesterol metabolism and the innate immune response in Alzheimer's disease; and post-synaptic signalling in schizophrenia and bipolar disorder (via CNVs).
Third, it is possible to compare the ‘burden’ of a type of genetic variant between cases and controls. Reference Sullivan, Daly and O'Donovan6 Individuals with autism or schizophrenia have consistently been reported to have a greater burden of rare CNVs than controls, and similar findings may hold for ADHD, bipolar disorder and major depressive disorder. For common genetic variation, people with schizophrenia, bipolar disorder and major depressive disorder have greater burden of risk alleles than controls. For schizophrenia, this finding is highly replicable and highly significant (P<10–25). Reference Sullivan, Daly and O'Donovan6
Genome-wide association studies have provided real data about the genetic basis of psychiatric disorders. Genetic architecture refers to the number of loci conferring risk for a disorder and their frequencies, effect sizes, modes of action and interactions with other genetic loci and environmental factors. In the many, mostly philosophical debates on this topic in the past century, two extreme views have been articulated: psychiatric disorders are caused by rare mutations of strong effect with most cases having a different causal variant v. a causal model consisting of the cumulative effects of many common variants of relatively subtle effects.
Where the results are sufficient to afford the ability to judge, the answer is that both common and rare variants have roles. This general conclusion applies to Alzheimer's disease (rare mutations with very strong effects (APP, PSEN1, PSEN2) and ten common variants of far more subtle effects) and bipolar disorder and schizophrenia (although no Mendelian-like mutations have been identified, rare CNVs play a role along with multiple common variants). For autism spectrum disorders, the rare variant catalogue is more complete (rare Mendelian syndromes with autistic features like Rett syndrome, karyotype abnormalities in approximately 5% of individuals, and CNVs in 5–10% of individuals). A recent series of papers in Nature used exome sequencing to identify just three candidate genes (SCN2A, KATNAL2, CHD8) containing rare de novo variants. Indeed, the hypothesis that autism results only from many different Mendelian-like mutations could be rejected. The role of common variation in autism is currently unknown as the available GWAS samples are small by current standards.
Common variation has been implicated for alcohol and nicotine consumption. For ADHD, several rare CNVs have been reported. For anorexia nervosa, obsessive–compulsive disorder, post-traumatic stress disorder and Tourette syndrome, the published data are sparse, and the roles of common and rare variation are unknown. For all of these disorders, GWAS sample sizes are not large by current standards, and considerably smaller than the sample sizes that were required to identify robust and replicable findings for other biomedical diseases.
For major depressive disorder, a large GWAS mega-analysis (approximately 19 000 participants) failed to identify findings of genome-wide significance. In context, nearly all GWAS with sample sizes above 11 000 participants identified at least one genome-wide significant finding.
What have we learned?
For most researchers, the rare v. common variant debate is settled. Unsurprisingly, where there are reasonable amounts of data, the answers contain elements of both models. Although some have wished for these disorders to conform to a classical medical genetic model whereby the aetiology of complex psychiatric disorders would resolve into a series of highly penetrant mutations, such hopes have now been demonstrated to be inconsistent with results for Alzheimer's disease, autism and schizophrenia. The ‘many Mendelians’ model now seems to be very unlikely.
Psychiatric disorders are polygenic. The rare variant and the common variant results indicate that many different loci are involved in Alzheimer's disease, bipolar disorder, schizophrenia, autism spectrum disorders and drug consumption. A parsimonious hypothesis is that these variants encode or regulate multicomponent biological pathways. Reference Sullivan9 Several intriguing hypotheses have emerged from GWAS data that suggest novel mechanisms underlying these disorders.
There are more discoveries to be made. From the experiences of other biomedical disorders for which GWAS have been conspicuously successful (for example type 2 diabetes mellitus or inflammatory bowel disease), we can confidently project that larger studies will yield more robust and replicable findings. To our knowledge, this is the first time in the history of psychiatry where there is a clear path to increasing our fundamental understanding of these disorders.
Genome-wide results for PGC analyses for ADHD, bipolar disorder, major depressive disorder and schizophrenia can be freely downloaded (https://pgc.unc.edu) and visualised in a genomic regional context (www.broadinstitute.org/mpg/ricopili). Full results and individual data are available by application to the NIMH Repository (www.nimhgenetics.org), the database of Genotypes and Phenotypes (www.ncbi.nlm.nih.gov/dbgap), and/or the Wellcome Trust (www.wtccc.org.uk). The top findings for published GWAS are also available (www.genome.gov/26525384).
We are indebted to the tens of thousands of individuals who chose to participate in the work summarised here, and to the hundreds of scientific colleagues who have donated thousands of person-hours to these efforts.