Introduction
Candidate gene studies have been a major focus in schizophrenia (SCZ) research with the SZGene database listing more than 1400 studies since 1965 (Allen et al. Reference Allen, Bagade, McQueen, Ioannidis, Kavvoura, Khoury, Tanzi and Bertram2008). By contrast, there are around 2200 PubMed citations for ‘schizophrenia randomized controlled trials’. Until 5 years ago, genetic studies could investigate only an extremely small proportion of the genome because of genotyping and cost limitations. Investigators thus had to focus on a limited number of genetic markers, genes and samples. In hypothesis-driven candidate gene studies, targets were selected based on ideas about pathophysiology or gene location under a linkage peak (Cichon et al. Reference Cichon, Craddock, Daly, Faraone, Gejman, Kelsoe, Lehner, Levinson, Moran, Sklar and Sullivan2009). For most biomedical disorders (including SCZ), the results of these studies were inconsistent or confusing (Ioannidis et al. Reference Ioannidis, Ntzani, Trikalinos and Contopoulos-Ioannidis2001). It is unclear whether this reflects poor choices of candidate genes or inadequate assessment (i.e. small samples or incomplete coverage of common genetic variation).
Genotyping and cost improvements now permit routine assessment of a million or more genetic variants distributed across the genome (Beaudet & Belmont, Reference Beaudet and Belmont2008). Genome-wide association studies (GWAS) can extract information from most common genetic variants in the genome either through direct assessment of single nucleotide polymorphisms (SNPs) or indirectly through linkage disequilibrium (LD) between genotyped SNPs and unmeasured but correlated genetic variants. Despite the advantages of genome-wide genotyping (Hindorff et al. Reference Hindorff, Sethupathy, Junkins, Ramos, Mehta, Collins and Manolio2009), stringent adjustment for multiple comparisons is required, which necessitates the use of large sample collections.
Ten GWAS for SCZ have been published (Lencz et al. Reference Lencz, Morgan, Athanasiou, Dain, Reed, Kane, Kucherlapati and Malhotra2007; O'Donovan et al. Reference O'Donovan, Craddock, Norton, Williams, Peirce, Moskvina, Nikolov, Hamshere, Carroll, Georgieva, Dwyer, Holmans, Marchini, Spencer, Howie, Leung, Hartmann, Moller, Morris, Shi, Feng, Hoffmann, Propping, Vasilescu, Maier, Rietschel, Zammit, Schumacher, Quinn, Schulze, Williams, Giegling, Iwata, Ikeda, Darvasi, Shifman, He, Duan, Sanders, Levinson, Gejman, Cichon, Nothen, Gill, Corvin, Rujescu, Kirov, Owen, Buccola, Mowry, Freedman, Amin, Black, Silverman, Byerley and Cloninger2008; Shifman et al. Reference Shifman, Johannesson, Bronstein, Chen, Collier, Craddock, Kendler, Li, O'Donovan, O'Neill, Owen, Walsh, Weinberger, Sun, Flint and Darvasi2008; Sullivan et al. Reference Sullivan, Lin, Tzeng, van den Oord, Perkins, Stroup, Wagner, Lee, Wright, Zou, Liu, Downing, Lieberman and Close2008; Kirov et al. Reference Kirov, Zaharieva, Georgieva, Moskvina, Nikolov, Cichon, Hillmer, Toncheva, Owen and O'Donovan2009; Need et al. Reference Need, Ge, Weale, Maia, Feng, Heinzen, Shianna, Yoon, Kasperaviciute, Gennarelli, Strittmatter, Bonvicini, Rossi, Jayathilake, Cola, McEvoy, Keefe, Fisher, St Jean, Giegling, Hartmann, Moller, Ruppert, Fraser, Crombie, Middleton, St Clair, Roses, Muglia, Francks, Rujescu, Meltzer and Goldstein2009; Purcell et al. Reference Purcell, Wray, Stone, Visscher, O'Donovan, Sullivan and Sklar2009; Shi et al. Reference Shi, Levinson, Duan, Sanders, Zheng, Pe'er, Dudbridge, Holmans, Whittemore, Mowry, Olincy, Amin, Cloninger, Silverman, Buccola, Byerley, Black, Crowe, Oksenberg, Mirel, Kendler, Freedman and Gejman2009; Stefansson et al. Reference Stefansson, Ophoff, Steinberg, Andreassen, Cichon, Rujescu, Werge, Pietilainen, Mors, Mortensen, Sigurdsson, Gustafsson, Nyegaard, Tuulio-Henriksson, Ingason, Hansen, Suvisaari, Lonnqvist, Paunio, Borglum, Hartmann, Fink-Jensen, Nordentoft, Hougaard, Norgaard-Pedersen, Bottcher, Olesen, Breuer, Moller, Giegling, Rasmussen, Timm, Mattheisen, Bitter, Rethelyi, Magnusdottir, Sigmundsson, Olason, Masson, Gulcher, Haraldsson, Fossdal, Thorgeirsson, Thorsteinsdottir, Ruggeri, Tosato, Franke, Strengman, Kiemeney, Melle, Djurovic, Abramova, Kaleda, Sanjuan, de Frutos, Bramon, Vassos, Fraser, Ettinger, Picchioni, Walker, Toulopoulou, Need, Ge, Yoon, Shianna, Freimer, Cantor, Murray, Kong, Golimbet, Carracedo, Arango, Costas, Jonsson, Terenius, Agartz, Petursson, Nothen, Rietschel, Matthews, Muglia, Peltonen, St Clair, Goldstein, Stefansson and Collier2009; Athanasiu et al. Reference Athanasiu, Mattingsdal, Kahler, Brown, Gustafsson, Agartz, Giegling, Muglia, Cichon, Rietschel, Pietilainen, Peltonen, Bramon, Collier, Clair, Sigurdsson, Petursson, Rujescu, Melle, Steen, Djurovic and Andreassen2010). Given that some GWAS had larger samples and more comprehensive genotyping than typical for the candidate gene literature, GWAS may be better placed to capture true associations than earlier studies. Indeed, GWAS have yielded highly significant and replicated associations for SCZ, including genetic variation in the major histocompatibility complex (MHC) region and within TCF4 and ZNF804A (O'Donovan et al. Reference O'Donovan, Craddock, Norton, Williams, Peirce, Moskvina, Nikolov, Hamshere, Carroll, Georgieva, Dwyer, Holmans, Marchini, Spencer, Howie, Leung, Hartmann, Moller, Morris, Shi, Feng, Hoffmann, Propping, Vasilescu, Maier, Rietschel, Zammit, Schumacher, Quinn, Schulze, Williams, Giegling, Iwata, Ikeda, Darvasi, Shifman, He, Duan, Sanders, Levinson, Gejman, Cichon, Nothen, Gill, Corvin, Rujescu, Kirov, Owen, Buccola, Mowry, Freedman, Amin, Black, Silverman, Byerley and Cloninger2008; Purcell et al. Reference Purcell, Wray, Stone, Visscher, O'Donovan, Sullivan and Sklar2009; Shi et al. Reference Shi, Levinson, Duan, Sanders, Zheng, Pe'er, Dudbridge, Holmans, Whittemore, Mowry, Olincy, Amin, Cloninger, Silverman, Buccola, Byerley, Black, Crowe, Oksenberg, Mirel, Kendler, Freedman and Gejman2009; Stefansson et al. Reference Stefansson, Ophoff, Steinberg, Andreassen, Cichon, Rujescu, Werge, Pietilainen, Mors, Mortensen, Sigurdsson, Gustafsson, Nyegaard, Tuulio-Henriksson, Ingason, Hansen, Suvisaari, Lonnqvist, Paunio, Borglum, Hartmann, Fink-Jensen, Nordentoft, Hougaard, Norgaard-Pedersen, Bottcher, Olesen, Breuer, Moller, Giegling, Rasmussen, Timm, Mattheisen, Bitter, Rethelyi, Magnusdottir, Sigmundsson, Olason, Masson, Gulcher, Haraldsson, Fossdal, Thorgeirsson, Thorsteinsdottir, Ruggeri, Tosato, Franke, Strengman, Kiemeney, Melle, Djurovic, Abramova, Kaleda, Sanjuan, de Frutos, Bramon, Vassos, Fraser, Ettinger, Picchioni, Walker, Toulopoulou, Need, Ge, Yoon, Shianna, Freimer, Cantor, Murray, Kong, Golimbet, Carracedo, Arango, Costas, Jonsson, Terenius, Agartz, Petursson, Nothen, Rietschel, Matthews, Muglia, Peltonen, St Clair, Goldstein, Stefansson and Collier2009). A lack of congruity has been noted between the hypothesis-driven candidate genes for SCZ and the best findings from GWAS. This may be typical for biomedical diseases, where results from large GWAS infrequently correspond to a priori candidate genes (www.genome.gov/gwastudies; Hindorff et al. Reference Hindorff, Sethupathy, Junkins, Ramos, Mehta, Collins and Manolio2009).
For SCZ, there are multiple reasons for the lack of overlap between GWAS and candidate gene studies. A key possibility is that prior hypotheses about the genetics of SCZ are incorrect. However, alternative explanations require exploration before accepting such an important conclusion. First, current GWAS chips provide coverage for most but not all of the genome (International HapMap Consortium, 2005), so particular regions and non-SNP genetic variation may be covered poorly. Second, power may be insufficient. Although GWAS tend to have large sample sizes by historical standards, the necessity to adjust for around a million statistical tests could result in low power. If that is the explanation, support for the hypotheses underpinning previous candidate genes might be obtained by a more systematic analysis of the GWAS data for evidence for over-representation of smaller p values than expected by chance (Holmans et al. Reference Holmans, Green, Pahwa, Ferreira, Purcell, Sklar, Owen, O'Donovan and Craddock2009). Third, individually rare genetic variants of strong effect might also be missed by GWAS (although these would also go undetected by most prior candidate gene studies).
The main purpose of this study was to compare hypothesis-driven candidate genes for SCZ from SZGene (Allen et al. Reference Allen, Bagade, McQueen, Ioannidis, Kavvoura, Khoury, Tanzi and Bertram2008) with the largest SCZ GWAS published to date (International Schizophrenia Consortium, 2008; Purcell et al. Reference Purcell, Wray, Stone, Visscher, O'Donovan, Sullivan and Sklar2009). First, we characterized the hypothesis-driven candidate gene studies. Second, we conducted quantitative comparisons to determine whether the set of hypothesis-driven candidate genes were either enriched for lower p values in the International Schizophrenia Consortium (ISC) results or contained markers with predictive power for SCZ. Finally, we performed a qualitative comparison of the most studied hypothesis-driven candidate genes with the ISC results.
Method
Hypothesis-driven candidate genes
Candidate genes for SCZ were drawn from SZGene (Allen et al. Reference Allen, Bagade, McQueen, Ioannidis, Kavvoura, Khoury, Tanzi and Bertram2008) (courtesy of Dr L. Bertram, 11 September 2009). SZGene included studies evaluating associations between a genetic variant and SCZ and published in a peer-reviewed, English-language journal (Allen et al. Reference Allen, Bagade, McQueen, Ioannidis, Kavvoura, Khoury, Tanzi and Bertram2008). Studies were identified through PubMed, bibliographies, and tables of contents. To ensure that the list was not ‘contaminated’ by the results of GWAS, SZGene entries from GWAS were removed, as were genes studied only subsequent to identification in GWAS. As only ISC autosomal results were available, 15 chromosome X genes were dropped. The list of autosomal hypothesis-driven candidate genes for SCZ contained 732 genes from 1374 studies (see Supplementary Table S1, available online).
The purpose in creating this list was to enumerate genes thought to be potentially etiological for SCZ by researchers in the field. The quality of the individual studies was variable. However, our interest was not in the study results per se (which can be strongly impacted by quality) but rather in the choice of a gene.
Samples
The ISC study is described elsewhere (International Schizophrenia Consortium, 2008; Purcell et al. Reference Purcell, Wray, Stone, Visscher, O'Donovan, Sullivan and Sklar2009). In brief, we studied 3322 European cases with DSM-IV or ICD-10 SCZ and 3587 controls from seven sites. All work was approved by institutional review boards. After complete description of the study to the subjects, written informed consent was obtained. Genotyping was performed on DNA extracted from blood using Affymetrix 5.0 or 6.0 arrays. Genotypes were called using Birdsuite (Korn et al. Reference Korn, Kuruvilla, McCarroll, Wysoker, Nemesh, Cawley, Hubbell, Veitch, Collins, Darvishi, Lee, Nizzari, Gabriel, Purcell, Daly and Altshuler2008) and imputation conducted using Beagle (Browning & Browning, Reference Browning and Browning2007, Reference Browning and Browning2009) against the HapMap3 CEU data resulting in 1 948 385 autosomal SNPs with direct or imputed genotypes. The primary analysis was logistic regression of disease state on the imputed allele dosages with inclusion of covariates to adjust for population stratification effects. The Genetic Association Information Network (GAIN) study genotyped 1230 SCZ cases and 1136 controls of European ancestry on the Affymetrix 6.0 array (Shi et al. Reference Shi, Levinson, Duan, Sanders, Zheng, Pe'er, Dudbridge, Holmans, Whittemore, Mowry, Olincy, Amin, Cloninger, Silverman, Buccola, Byerley, Black, Crowe, Oksenberg, Mirel, Kendler, Freedman and Gejman2009).
Statistical analysis
We first explored the set of hypothesis-driven candidate genes using a variety of descriptive approaches (SAS Institute Inc., 2004, 2005). Quanto was used for power calculations (Gauderman, Reference Gauderman2002; Gauderman & Morrison, Reference Gauderman and Morrison2006). We used the Database for Annotation, Visualization, and Integrated Discovery (DAVID; Dennis et al. Reference Dennis, Sherman, Hosack, Yang, Gao, Lane and Lempicki2003; Sherman et al. Reference Sherman, Huang, Tan, Guo, Bour, Liu, Stephens, Baseler, Lane and Lempicki2007; Huang et al. Reference Huang, Sherman and Lempicki2009) to characterize hypotheses about the pathophysiology of SCZ reflected in the hypothesis-driven candidate gene list. DAVID identifies Gene Ontology (GO) biological pathways (Harris et al. Reference Harris, Clark, Ireland, Lomax, Ashburner, Foulger, Eilbeck, Lewis, Marshall, Mungall, Richter, Rubin, Blake, Bult, Dolan, Drabkin, Eppig, Hill, Ni, Ringwald, Balakrishnan, Cherry, Christie, Costanzo, Dwight, Engel, Fisk, Hirschman, Hong, Nash, Sethuraman, Theesfeld, Botstein, Dolinski, Feierbach, Berardini, Mundodi, Rhee, Apweiler, Barrell, Camon, Dimmer, Lee, Chisholm, Gaudet, Kibbe, Kishore, Schwarz, Sternberg, Gwinn, Hannick, Wortman, Berriman, Wood, de la Cruz, Tonellato, Jaiswal, Seigfried and White2004) with chance-corrected over-representation of a given gene list. To account for overlap between pathways, we used the annotation cluster feature in DAVID to focus on higher-level clusters of similar biological processes.
We then compared hypothesis-driven candidate genes for SCZ with ISC GWAS results to assess whether the hypothesis-driven candidate gene list had over-representation of smaller ISC p values than expected by chance. These analyses were conducted using ALIGATOR (Holmans et al. Reference Holmans, Green, Pahwa, Ferreira, Purcell, Sklar, Owen, O'Donovan and Craddock2009) and InRich (Lee et al. Reference Lee, O'Dushlaine, Thomas, Holmans and Purcell2011). These programs use different algorithms to assess whether GWAS findings are over-represented for small p values with reference to a predefined set of genes (i.e. a pathway). ALIGATOR uses permutation to account for variable numbers of SNPs per gene, different patterns of LD between SNPs (within the same gene), and varying gene sizes. We considered SZGene hypothesis-driven candidate genes as a ‘pathway’ and used ALIGATOR to estimate the probability that this list contained an over-representation of smaller ISC GWAS p values. The ISC GWAS results were input to ALIGATOR, which assigned these SNPs to UCSC hg18 RefSeq genes (Pruitt et al. Reference Pruitt, Tatusova and Maglott2005). We determined the significance threshold (generally 0.002–0.004) that designated the top 5% of all genes as ‘significant’ (Holmans et al. Reference Holmans, Green, Pahwa, Ferreira, Purcell, Sklar, Owen, O'Donovan and Craddock2009). The key statistical comparison is akin to a 2×2 table of whether a gene is in the top 5% by whether a gene is a member of a pathway. Assessing significance is complex because of violation of independence assumptions. ALIGATOR uses an SNP-based permutation algorithm to create a reference distribution for a pathway. InRich controls LD between genes by comparing a gene set of interest to LD-independent regions. Using the same significance thresholds as in ALIGATOR, we identified LD-independent significant regions from the ISC dataset using the clump function within PLINK (r 2=0.5 over 1 Mb). We then used InRich to determine whether the candidate gene set showed enrichment for these regions.
Polygenic score analysis was conducted as described in the ISC GWAS paper by Purcell et al. (Reference Purcell, Wray, Stone, Visscher, O'Donovan, Sullivan and Sklar2009). We used 14 811 SNPs that were genotyped in both the ISC and SCZ GAIN samples and that mapped to candidate genes. We made a polygenic profile based on the risk alleles within these SNPs in the ISC data, and used this profile to create a polygenic score for each individual within the SCZ GAIN sample (an independent target sample). We used logistic regression between the score of each individual in SCZ GAIN and their case/control status.
Results
Characteristics of hypothesis-driven candidate gene studies of SCZ
Table 1 describes hypothesis-driven candidate genes from SZGene (Allen et al. Reference Allen, Bagade, McQueen, Ioannidis, Kavvoura, Khoury, Tanzi and Bertram2008). There were 732 autosomal genes from 1374 hypothesis-driven candidate gene studies (Supplementary Table S1). These genes were studied from one to 81 times. Most genes (563; 76.9%) were investigated in one (60.9%) or two studies (16.0%). Although replication is important in human genetics, there is little capacity to evaluate both false-positive and false-negative findings. Two-thirds of genes were assessed by ⩽3 markers and had a median SNP density of 15.4 kb/SNP. The median numbers of cases (191) and controls (214) were modest.
Table 1. Characteristics of hypothesis-driven candidate gene studies and the ISC GWAS
ISC, International Schizophrenia Consortium; GWAS, genome-wide association studies; SNP, single nucleotide polymorphism; IQR, interquartile range; n.a., not applicable; DAVID, Database for Annotation, Visualization, and Integrated Discovery.
a Values given as median (range) for candidate gene studies (biased due to subject overlap across publications) or actual number for ISC.
b See text for assumptions.
c For ISC, assuming gene boundaries expanded by ±10 kb and SNP density <20 kb/SNP.
We used pathway analysis to characterize the hypotheses that guided candidate gene selection. The 732 hypothesis-driven candidate genes were entered into DAVID to assess the GO biological processes to which these genes belonged. The top four annotation clusters consist of biological processes involved in synaptic transmission, neuronal development and morphogenesis, regulation of synaptic transmission, and response to chemical stimuli (Table 1 and Supplementary Table S2). The full list reflects a diversity of ideas about SCZ etiology; as expected, cluster enrichment scores were particularly strong for candidate genes selected under the hypotheses that SCZ is a disease of the synapse and/or a neurodevelopmental process.
We next evaluated completeness and coverage for the hypothesis-driven candidate genes. First, we estimated statistical power to detect association. Even for relatively large studies (i.e. samples sizes at the 90th percentiles of n case=537 and n control=628), and a liberal correction for multiple comparisons (α=0.005, 10 markers), the power was 48% to detect genetic effects typical for GWAS of human diseases (median genotypic relative risk of 1.28 and median minor allele frequency of 0.29 for disease associations with p<5×10−8) (Hindorff et al. Reference Hindorff, Sethupathy, Junkins, Ramos, Mehta, Collins and Manolio2009). Second, we assessed genomic coverage. The 732 hypothesis-driven candidate genes represent 3.7% of RefSeq autosomal genes (Pruitt et al. Reference Pruitt, Tatusova and Maglott2005). Marker coverage can be assessed only generally, but included only small proportions of common genetic variation. Finally, of all genes comprising pathways in the top four DAVID annotation clusters, only 6.7% had ever been studied. Although these pathways may be overinclusive, the main hypotheses guiding selection of hypothesis-driven candidate genes were evaluated incompletely.
Hypothesis-driven candidate gene studies and the ISC GWAS
The ISC GWAS had 3322 cases, 3587 controls and 1 948 385 genotyped and imputed autosomal SNPs. The sample size was 1.36 times greater than the largest hypothesis-driven candidate gene study. Statistical power was 89% to detect a genetic effect corresponding to that typical for SNPs implicated in human disease GWAS (Hindorff et al. Reference Hindorff, Sethupathy, Junkins, Ramos, Mehta, Collins and Manolio2009) including adjustment for multiple comparisons (α=5×10−8). The ISC reported 1 948 385 associations, which exceeds the total associations in the SZGene database by over 180-fold. The mean marker density over the genome was 1 SNP/1.6 kb. In comparison to HapMap (r27, phases I+II+III), ISC markers assessed 79.0% of known common variants present in individuals of European ancestry either directly or indirectly with r 2 ⩾0.8.
We next investigated coverage and gene size (using strict gene boundaries, ±0 kb). The 732 hypothesis-driven candidate genes were markedly larger than the remaining 18 891 autosomal RefSeq genes (median 33.5 v. 20.5 kb, Wilcoxon p=4×10−20). Importantly, ISC SNP densities were similar in hypothesis-driven candidate genes in comparison to all other autosomal genes (median 1360 v. 1379 bases/SNP, Wilcoxon p=0.25). A sizable fraction of autosomal RefSeq genes had no ISC SNPs within their boundaries (18.7%). As expected, genes with no SNPs were markedly smaller (median 4.1 v. 28.2 kb, Wilcoxon p≈0). As the 732 SCZ candidate genes were larger, they were significantly less likely to have no coverage than the remaining RefSeq genes (10.0% v. 19.0%, p=5×10−11). Although this generation of GWAS chips provides partial genomic coverage of common variation, hypothesis-driven candidate genes for SCZ had better coverage than other RefSeq genes.
Do the ISC GWAS data support hypothesis-driven candidate genes as significant contributors to SCZ risk?
There were no SNPs within the gene boundaries (±0 kb) for 73 hypothesis-driven candidate genes and no SNPs within expanded gene boundaries (±10 kb) for 27 genes (Supplementary Table S3). Hypothesis-driven candidate genes with no ISC coverage were excluded from enrichment analyses. Of the ~1.9 million ISC SNPs, 56 981 mapped within hypothesis-driven candidate genes (±0 kb) and 65 803 mapped within expanded gene boundaries (±10 kb).
We assessed whether hypothesis-driven candidate genes were over-represented for smaller ISC GWAS p values using ALIGATOR. The central comparison was whether there was an over-representation of the top 5% of significant genes in the hypothesis-driven candidate gene list in comparison to the remaining annotated genes. In Table 2, there was a nominally significant over-representation of smaller p values in the ISC data for the full set of hypothesis-driven candidate genes but these values did not survive multiple testing correction. In addition, there was no evidence for over-representation of small ISC p values in the most studied hypothesis-driven candidate genes (⩾3 times, 23.1% of the total).
Table 2. Testing for over-representation of smaller ISC GWAS p values in hypothesis-driven candidate genes for SCZ and genes corresponding to major hypotheses
ISC, International Schizophrenia Consortium; GWAS, genome-wide association studies; SCZ, schizophrenia; MHC, major histocompatibility complex; LD, linkage disequilibrium; DAVID, Database for Annotation, Visualization, and Integrated Discovery.
Empirical p values (from ALIGATOR unless otherwise noted) testing for over-representation of smaller ISC GWAS p values in a given gene list in comparison to that expected by chance (10 000 permutations). Single nucleotide polymorphisms (SNPs) were mapped to strict (±0 kb) or expanded (±10 kb) gene boundaries. SNP thresholds to select the top 5% of genes varied from 0.002 to 0.004.
Pathway analysis can be complex in regions such as the MHC that have extensive disequilibrium between genes (Stenzel et al. Reference Stenzel, Lu, Koch, Hampe, Guenther, De La Vega, Krawczak and Schreiber2004). When we repeated the ALIGATOR analysis after excluding genes in the MHC region, there was no evidence for over-representation of smaller p values (p≈0.48), indicating that the marginal enrichment was due to bias. We repeated the pathway analysis using InRich, which may be more robust than ALIGATOR in regions of high LD. InRich evaluates regions defined by LD. We tested the full candidate gene dataset, using the same significance thresholds as in ALIGATOR, and found no evidence for enrichment of significant findings in the candidate genes (Table 2). Therefore, the pathway analyses are consistent with the null hypothesis because all p values were non-significant or marginal and would not survive correction for multiple testing, and because removal of the MHC region and analysis with a program that corrects for LD indicates the results are a false positive resulting from extensive LD in the MHC region.
In addition, we evaluated whether the list of historical candidate genes, as a whole, were making a significant contribution to risk of SCZ by evaluating the polygenic score profile. This approach has provided support for an important polygenic contribution to SCZ (Purcell et al. Reference Purcell, Wray, Stone, Visscher, O'Donovan, Sullivan and Sklar2009). We created a polygenic score profile for SNPs that mapped to historical candidate genes using the ISC data. We then applied this score to an independent dataset (SCZ GAIN, n=1230 cases and 1136 controls). Independent SCZ cases did not have greater risk scores than controls based on these historical candidate genes (p=0.92).
Many hypothesis-driven candidate genes were selected from two overarching hypotheses: SCZ as a synaptic or a neurodevelopmental disorder (Table 1). These hypotheses have been incompletely assessed. We used pathway analysis of the ISC data to assess these hypotheses in far more detail than has been possible previously. The set of 4808 genes that comprise the synaptic cluster 1 from DAVID did not show over-representation of lower ISC p values (p≈1). This list may be overinclusive and the candidate genes studied might be more refined and appropriate to SCZ; however, the subset of 222 cluster 1 genes investigated in a hypothesis-driven candidate gene study did not have over-representation of smaller ISC p values (p≈1). Similarly, genes in DAVID cluster 2 (neurodevelopment) did not show enrichment for lower ISC p values for the full set (4834 genes) or the subset evaluated in a candidate gene study (401 genes, all p values non-significant).
Qualitative comparisons
Pathway analysis considers sets of genes in aggregate. Negative aggregate results could miss true over-representation of small p values in a small number of hypothesis-driven candidate genes. Table 3 a depicts the ISC findings for the 10 most-studied hypothesis-driven candidate genes. There was inadequate coverage for two small genes (DRD4 and APOE), and the remainder had good SNP densities but weak ISC results, with none surviving a liberal gene-wise Bonferroni correction. Supplementary Fig. S2 depicts these genes and highlights regions of conspicuous attention in the literature (COMT/val58met, DRD3/ser9gly, DRD2/Taq1A, HTR2A/T102C, NRG1/HapICE, BDNF/val66met, DTNBP1, and SLC6A4/HTTLPR). The ISC results do not implicate common genetic variation in these genes. Although the region containing SLC6A4 shows no signal, the widely studied promoter polymorphism HTTLPR was not directly genotyped and neighboring SNPs are in low LD (Konneker et al. Reference Konneker, Crowley, Quackenbush, Keefe, Perkins, Stroup, Lieberman, van den Oord and Sullivan2010).
Table 3 a. ISC results for the 10 most studied genes from SZGene
ISC, International Schizophrenia Consortium; SNP, single nucleotide polymorphism; n.a., not applicable.
We also investigated the 35 ISC SNPs with associations <5×10−8 and all were located at chr6:31.58–32.77 mb in the MHC region (Purcell et al. Reference Purcell, Wray, Stone, Visscher, O'Donovan, Sullivan and Sklar2009). These SNPs map to 10 genes (Table 3 b), five of which had not previously been the subject of a candidate gene study. Four genes had been studied 1–5 times and one extensively (NOTCH4). The strong caveat for Table 3 b is the extensive LD in the MHC region (Supplementary Fig. S1); these genome-wide significant SNPs could reflect one or a few risk variants, which may or may not be a candidate gene.
Table 3 b. Genes with genome-wide significant ISC results and studies from SZGene
ISC, International Schizophrenia Consortium; SNP, single nucleotide polymorphism; MHC, major histocompatibility complex.
Discussion
The main aim of this study was to evaluate the hypothesis-driven candidate gene literature for SCZ with respect to a large GWAS dataset. Hypothesis-driven candidate gene studies have been a major approach to the molecular etiology of SCZ. However, we now can perform analyses of several orders of magnitude more detailed than were possible even 5 years ago. We wanted to determine whether the systematic investigations now allowed by GWAS supported this body of work in aggregate and the degree to which GWAS empirical results support the overarching concepts that influenced candidate gene selection. We emphasize that we did not conduct meta-analyses of the findings of hypothesis-driven candidate gene studies as this has been done elsewhere (Allen et al. Reference Allen, Bagade, McQueen, Ioannidis, Kavvoura, Khoury, Tanzi and Bertram2008).
We acknowledge the advantages of hindsight. The hypothesis-driven candidate gene literature, a body of work to which the present authors have contributed, contains numerous studies that were state of the art when they were performed and represent considerable effort by teams of investigators. GWAS will undoubtedly be subject to similar scrutiny that performed here for candidate gene studies. Although the ISC study is the largest and most comprehensive of the SCZ GWAS published to date, it was not ideal. Statistical power was high by historical standards, but we now know that greater power is desirable to detect the small genotypic relative risks characteristic of SCZ. In addition, coverage was not necessarily ‘genome-wide’, as some important regions had inadequate genotyping and rare genetic variation was poorly assessed. With these caveats in mind, several observations of the historical candidate gene literature emerged from our analyses.
First, hypothesis-driven candidate gene studies had poor statistical power by contemporary standards. Even for a relatively large candidate gene study with power-favorable multiple comparison correction, the power would have been poor to detect the genetic effects typical for GWAS of human diseases. As the genetic effects for SCZ may be smaller than for other human diseases (Purcell et al. Reference Purcell, Wray, Stone, Visscher, O'Donovan, Sullivan and Sklar2009; Shi et al. Reference Shi, Levinson, Duan, Sanders, Zheng, Pe'er, Dudbridge, Holmans, Whittemore, Mowry, Olincy, Amin, Cloninger, Silverman, Buccola, Byerley, Black, Crowe, Oksenberg, Mirel, Kendler, Freedman and Gejman2009; Stefansson et al. Reference Stefansson, Ophoff, Steinberg, Andreassen, Cichon, Rujescu, Werge, Pietilainen, Mors, Mortensen, Sigurdsson, Gustafsson, Nyegaard, Tuulio-Henriksson, Ingason, Hansen, Suvisaari, Lonnqvist, Paunio, Borglum, Hartmann, Fink-Jensen, Nordentoft, Hougaard, Norgaard-Pedersen, Bottcher, Olesen, Breuer, Moller, Giegling, Rasmussen, Timm, Mattheisen, Bitter, Rethelyi, Magnusdottir, Sigmundsson, Olason, Masson, Gulcher, Haraldsson, Fossdal, Thorgeirsson, Thorsteinsdottir, Ruggeri, Tosato, Franke, Strengman, Kiemeney, Melle, Djurovic, Abramova, Kaleda, Sanjuan, de Frutos, Bramon, Vassos, Fraser, Ettinger, Picchioni, Walker, Toulopoulou, Need, Ge, Yoon, Shianna, Freimer, Cantor, Murray, Kong, Golimbet, Carracedo, Arango, Costas, Jonsson, Terenius, Agartz, Petursson, Nothen, Rietschel, Matthews, Muglia, Peltonen, St Clair, Goldstein, Stefansson and Collier2009), nearly all hypothesis-driven candidate gene studies were underpowered. Given what we now know about the genetic architecture of SCZ, a typical candidate gene study requires sample sizes of around 11 000 cases plus controls for a single marker, 17 500 subjects for 10 markers, and 24 000 subjects for 100 markers (see Supplemental Methods online). Future association studies of SCZ should use similarly realistic power calculations.
Moreover, we demonstrated that hypothesis-driven candidate gene studies generally had poor coverage of common genetic variation. With the resources provided by the HapMap and 1000 Genomes projects, coupled with decreases in genotyping costs, researchers can ensure that future genotyping covers the majority of common and rare variation present in their samples.
We were surprised to realize that most genes in the hypothesis-driven candidate gene literature for SCZ had been assessed only once or twice (76.9%). Given the importance of replication in genetic studies of complex traits (Chanock et al. Reference Chanock, Manolio, Boehnke, Boerwinkle, Hunter, Thomas, Hirschhorn, Abecasis, Altshuler, Bailey-Wilson, Brooks, Cardon, Daly, Donnelly, Fraumeni, Freimer, Gerhard, Gunter, Guttmacher, Guyer, Harris, Hoh, Hoover, Kong, Merikangas, Morton, Palmer, Phimister, Rice, Roberts, Rotimi, Tucker, Vogan, Wacholder, Wijsman, Winn and Collins2007), evaluation of false positives and false negatives is not possible. Positive findings from one or two studies cannot be viewed as secure (particularly given the distorting potential of publication bias) and, conversely, negative findings may not provide confident exclusion.
If power and coverage are low, we can anticipate false negatives (i.e. true susceptibility loci with non-significant candidate gene findings). For example, GWAS and replication efforts support TCF4 as a risk locus for SCZ (Stefansson et al. Reference Stefansson, Ophoff, Steinberg, Andreassen, Cichon, Rujescu, Werge, Pietilainen, Mors, Mortensen, Sigurdsson, Gustafsson, Nyegaard, Tuulio-Henriksson, Ingason, Hansen, Suvisaari, Lonnqvist, Paunio, Borglum, Hartmann, Fink-Jensen, Nordentoft, Hougaard, Norgaard-Pedersen, Bottcher, Olesen, Breuer, Moller, Giegling, Rasmussen, Timm, Mattheisen, Bitter, Rethelyi, Magnusdottir, Sigmundsson, Olason, Masson, Gulcher, Haraldsson, Fossdal, Thorgeirsson, Thorsteinsdottir, Ruggeri, Tosato, Franke, Strengman, Kiemeney, Melle, Djurovic, Abramova, Kaleda, Sanjuan, de Frutos, Bramon, Vassos, Fraser, Ettinger, Picchioni, Walker, Toulopoulou, Need, Ge, Yoon, Shianna, Freimer, Cantor, Murray, Kong, Golimbet, Carracedo, Arango, Costas, Jonsson, Terenius, Agartz, Petursson, Nothen, Rietschel, Matthews, Muglia, Peltonen, St Clair, Goldstein, Stefansson and Collier2009). However, a TCF4 CAG repeat was studied for association with SCZ in three studies (Vincent et al. Reference Vincent, Petronis, Strong, Parikh, Meltzer, Lieberman and Kennedy1999; Bowen et al. Reference Bowen, Guy, Cardno, Vincent, Kennedy, Jones, Gray, Sanders, McCarthy, Murphy, Owen and O'Donovan2000; McInnis et al. Reference McInnis, Swift-Scanlanl, Mahoney, Vincent, Verheyen, Lan, Oruc, Riess, Van Broeckhoven, Chen, Kennedy, MacKinnon, Margolis, Simpson, McMahon, Gershon, Nurnberger, Reich, DePaulo and Ross2000). All reported negative results, which may have led to the inappropriate exclusion of TCF4 from consideration.
Furthermore, it is possible that the major hypotheses that drove the selection of many candidate genes are incorrect. SZGene candidate genes were selected for many different reasons and some resulted from genome-wide linkage screens (most notably NRG1 and DTNBP1) (Stefansson et al. Reference Stefansson, Sigurdsson, Steinthorsdottir, Bjornsdottir, Sigmundsson, Ghosh, Brynjolfsson, Gunnarsdottir, Ivarsson, Chou, Hjaltason, Birgisdottir, Jonsson, Gudnadottir, Gudmundsdottir, Bjornsson, Ingvarsson, Ingason, Sigfusson, Hardardottir, Harvey, Lai, Zhou, Brunner, Mutel, Gonzalo, Lemke, Sainz, Johannesson, Andresson, Gudbjartsson, Manolescu, Frigge, Gurney, Kong, Gulcher, Petursson and Stefansson2002; Straub et al. Reference Straub, Jiang, MacLean, Ma, Webb, Myakishev, Harris-Kerr, Wormley, Sadek, Kadambi, Cesare, Gibberman, Wang, O'Neill, Walsh and Kendler2002). However, the ISC GWAS results did not lend support for common variation contributing to SCZ, either for candidate genes from the literature as a whole or for the specific pathways from which candidate genes were frequently selected. For the full set of hypothesis-driven candidate genes, there was nominally significant support for an over-representation of small ISC p values. However, the effect was marginal, and the results were not significant when corrected for potential bias caused by LD between genes. We found no support for an aggregate effect of hypothesis-driven candidate genes contributing to SCZ risk using a risk profile generated from the SNPs within these genes. This pattern of results is not consistent with robust or notable collective contribution of common variation within the hypothesis-driven candidate genes to SCZ based on the ISC data. However, it is possible that subsets of the heterogeneous list of historical candidate genes are enriched for smaller ISC p values. We thus tested the two overarching ‘meta-hypotheses’ that have been highly influential: notions of SCZ as a disease of the synapse and as a neurodevelopmental disease. To our knowledge, these two larger-scale ideas have not been tested for empirical support in aggregate. We found no evidence to support a genetic basis for these two hypotheses in perhaps the most comprehensive analysis yet conducted. In addition, we specifically evaluated eight of the 10 most studied historical candidate genes and the ISC GWAS results provide no support for common genetic variation associated with SCZ. We note that the strongest ISC GWAS findings were in the MHC region. Genes from the expanded MHC region do appear in the hypothesis-driven candidate gene literature. Most notably, NOTCH4 had genome-wide significant SNPs in the ISC data and was highly studied (24 times; both positive and negative studies) in the candidate gene literature. However, given the high LD in the region (Supplementary Table S1), we cannot localize the MHC signal more specifically. We cannot therefore either directly validate or exclude NOTCH4 as being involved in SCZ susceptibility.
Finally, more generally, there are now numerous guidelines for candidate gene studies (Chanock et al. Reference Chanock, Manolio, Boehnke, Boerwinkle, Hunter, Thomas, Hirschhorn, Abecasis, Altshuler, Bailey-Wilson, Brooks, Cardon, Daly, Donnelly, Fraumeni, Freimer, Gerhard, Gunter, Guttmacher, Guyer, Harris, Hoh, Hoover, Kong, Merikangas, Morton, Palmer, Phimister, Rice, Roberts, Rotimi, Tucker, Vogan, Wacholder, Wijsman, Winn and Collins2007; Pearson & Manolio, Reference Pearson and Manolio2008). If these guidelines are followed at all stages of the scientific process (from study design through review), the published literature will better reflect the genetic architecture of SCZ.
No single study can disprove a meta-hypothesis in psychiatry, and our conclusions are bounded by the statistical power of the ISC sample. However, it is notable that the ISC GWAS results do not support enrichment of SCZ susceptibility loci within the candidate genes. These results suggest, but do not prove, that many traditional ideas about the genetic basis of SCZ may be incorrect. Indeed, the singular advantage of genomic surveys is that they are unbiased by prior knowledge and can yield novel and unexpected findings. Given current knowledge of the genetic architecture of SCZ and the capacity to assess common and rare variations across the genome, it is possible that the next few years will lead to marked changes in major hypotheses about the genetic basis of SCZ.
Note
Supplementary material accompanies this paper on the Journal's website (http://journals.cambridge.org/psm).
Acknowledgments
Funding for this project was from the US National Institutes of Health, who had no role in the design, execution, analysis, and manuscript preparation. This project was funded by R01 MH077139 to Dr P. F. Sullivan.
Declaration of Interest
None.
