Twin studies indicate that obsessive–compulsive disorder (OCD) is strongly influenced by additive genetic factors. Yet, molecular genetic association studies have yielded inconsistent results, possibly because of differences across studies in statistical power. Meta-analysis can yield greater power. This study reports the first comprehensive meta-analysis of the relationship between OCD and all previously examined polymorphisms for which there was sufficient information in the source studies to compute odds ratios (ORs). A total of 230 polymorphisms from 113 genetic association studies were identified. A full meta-analysis was conducted for 20 polymorphisms that were examined in 5 or more data sets, and a secondary meta-analysis (limited to the computation of mean effect sizes) was conducted for 210 polymorphisms that were examined in fewer than 5 data sets. In the main meta-analysis, OCD was associated with serotonin-related polymorphisms (5-HTTLPR and HTR2A) and, in males only, with polymorphisms involved in catecholamine modulation (COMT and MAOA). Nonsignificant trends were identified for two dopamine-related polymorphisms (DAT1 and DRD3) and a glutamate-related polymorphism (rs3087879). The secondary meta-analysis identified another 18 polymorphisms with significant ORs that merit further investigation. This study demonstrates that OCD is associated with multiple genes, with most having a modest association with OCD. This suggests a polygenic model of OCD, consistent with twin studies, in which multiple genes make small, incremental contributions to the risk of developing the disorder. Future studies, with sufficient power to detect small effects, are needed to investigate the genetic basis of OCD subtypes, such as early vs late onset OCD.
Obsessive–compulsive disorder (OCD) is characterized by distressing, time-consuming obsessions and compulsions. Obsessions are unwanted thoughts, images or urges. Compulsions are repetitive behaviors or mental acts that the person feels compelled to perform, typically with a desire to resist.1 The exact etiology of OCD is unknown, although evidence suggests that the disorder arises from complex combination of biopsychosocial factors, including genetic and environmental factors.2 A recent meta-analysis of 37 twin samples suggests that obsessions and compulsions arise from a combination of additive genetic factors and nonshared environment.3 As with other forms of psychopathology,4 OCD might be shaped by a large number of genes of modest impact, which additively combine to influence the risk for developing obsessive–compulsive symptoms or disorder.
One of the leading biological models of OCD is the frontal–striatal–thalamic model, which emphasizes the role of aberrant modulation of brain circuits involving the prefrontal cortex, regions of the basal ganglia such as the striatum and globus pallidus, and the thalamus.5 Aberrant modulation could be due to structural abnormalities and/or dysregulations in the neuromodulators or neurotransmitters (for example, serotonin, glutamate and dopamine) involved in these circuits.
There have been many molecular genetic studies of OCD but relatively few consistent findings.6 Inconsistencies might have been due to differences across studies in statistical power, which is problematic when genes with small effect sizes are being investigated. Meta-analysis, compared with the individual studies included in a meta-analysis, can be a more powerful method of evaluating effect sizes.7 Inconsistent findings across studies also might be an indication of the etiologic heterogeneity of OCD; there is evidence of sex differences and evidence of distinct subtypes of the disorder.8 Different subtypes may be etiologically different from one another, and the genes involved in OCD might be different for males compared with females. Meta-analysis can be used to investigate this issue by means of moderator (subgroup) analyses.
The purpose of this study was to conduct the first comprehensive meta-analysis of genetic association studies of OCD, including, where possible, data on all polymorphisms that have been studied so far. There have been over 100 OCD genetic association studies, and so a comprehensive meta-analysis of this substantial literature is warranted. Four previous meta-analyses have been published, but each was limited to only a single polymorphism. Azzam and Mathews9 meta-analyzed a small number of case–control (CC) and family-based (FB) studies of COMT. They found no association between either the Met or Val polymorphisms and OCD. Pooley et al.10 conducted a meta-analysis of COMT and found significant sex differences in terms of allele distribution and risk of OCD. Their study was limited because they included only CC studies, which are prone to sample stratification effects.11 Both CC and FB studies have their strengths and weaknesses.12 Lin13 and Bloch et al.14 meta-analyzed studies on the L and S polymorphisms of 5-HTTLPR and found no association between any polymorphism and OCD. A limitation of both meta-analyses was that there are 3, not 2, relatively common polymorphisms of 5-HTTLPR; La, Lg and S. Lg biologically acts like S, and so the question arises as to whether OCD is associated with La. (There are also two forms of S: Sa and Sg. The latter is very rare compared with Sa, La and Lg).15, 16 Until recently, most OCD studies had treated La and Lg as if they were identical to one another. This study sought to address these limitations and to extend the investigation by meta-analytically examining the association between OCD and many other polymorphisms that have not been previously meta-analyzed. This study is the first meta-analysis to examine all polymorphisms that have been studied so far in relation to OCD for studies in which effect sizes (odds ratios; ORs) could be computed. In addition to presenting meta-analytic findings, this study also sought to identify important gaps or limitations in the research literature. If OCD is likely to be etiologically heterogeneous and associated with genes that have small effects, then the question arises as to whether the research literature adequately addresses these issues. This involves questions of whether genetic association studies have had sufficient statistical power to detect small effects, and whether these studies have adequately investigated the question of whether genetic findings vary as a function of moderator variables, such as OCD subtypes (for example, early vs late onset OCD, where the former is more prevalent in males)8 and other potential moderators, such as particular symptom profiles.
Materials and methods
Relevant studies were identified by systematically searching, up to 1 March 2012, MEDLINE, PsychINFO and EMBASE. All three databases were searched because they are not entirely overlapping in content.17 Search terms were (obsessive–compulsive or OCD) combined with (gene, genes, genetic, genotype, allele* or polymorphism*). Asterisks denote the use of wild cards. References cited in all identified articles, and references cited in relevant review articles and book chapters, were also examined. For all journals that had published genetic association studies of OCD, their on-line ‘in press’ (or ‘online first’) publications were also inspected for potentially suitable studies.
A total of 179 potentially relevant studies were located, yielding 113 studies that were suitable for meta-analysis. A complete list of included and excluded studies appears in the appendix of Supplementary Materials. Studies were included if they used either CC or FB association designs in which cases (probands) met DSM-III, DSM-III-R or DSM-IV criteria for OCD. Studies were included only if it was possible to compute an effect size (OR) for allele frequency for a given polymorphism. Studies were excluded if their samples were subsumed within other, larger studies included in the meta-analysis, or if the sample consisted of form of OCD that may not be representative of OCD in general (for example, OCD studied exclusively in people with Velocardiofacial syndrome). Samples consisting entirely of probands with hoarding symptoms were excluded from meta-analysis. This was to limit the number of probands who would be diagnosed with hoarding disorder (as currently proposed for DSM-V; www.dsm5.org) rather than diagnosed with OCD. As currently proposed for DSM-V, people with prominent hoarding symptoms would be diagnosed as having hoarding disorder. OCD would be diagnosed when a person has hoarding symptoms that do not dominate the clinical picture; that is, when hoarding co-occurs with prototypic obsessive–compulsive symptoms (for example, checking, washing) and hoarding is not the primary symptom in terms of severity.
Analyses were conducted using the Comprehensive Meta-Analysis program, version 2.2050.18 ORs from individual studies were weighted according to the inverse variance method, in which more precise studies (that is, those with smaller variances and typically larger sample sizes) received greater weighting.7 Random effects modeling was used to compute mean ORs because it was unclear whether the true effect size for a given polymorphism is the same across all studies or subgroups. This type of analysis was indicated, given the evidence suggesting that OCD is etiologically heterogeneous.8 Effect sizes might vary as a function of a range of variables, including study design (CC vs FB designs) and demographic variables such as sex and racial background.
Given the number of statistical analyses conducted in this study, the α level for statistical significance was set at 0.01 (two-tailed) instead of the conventional 0.05, in order to reduce type I error without unduly inflating type II error. (With regard to type I and II errors when multiple tests are conducted, a stringent correction in type I error (for example, a Bonferroni correction) increases the number of type II errors. This is of particular concern when small effects are being investigated because statistical power is reduced when type II error is increased. Some investigators have argued against the use of Bonferroni and related corrections, because of their effects on inflating type II error.19, 20 Lieberman and Cunningham21 argued for a more liberal α level than that imposed by a Bonferroni correction because ‘type I errors are self-erasing because they will not replicate’ (p. 423). But an α level of 0.05 would seem unduly liberal given that 20 polymorphisms were examined in the main meta-analysis in this article. A commonly used alternative is to set α at 0.01,22, 23 which in this study lies between a Bonferroni-corrected α for the 20 polymorphisms (0.003) and 0.05. For the tests of mean ORs of these polymorphisms, this approach was supplemented by comparing the obtained P-values with cutoff values computed from the Holm–Bonferroni method,24, 25 which controls type I error (at α=0.05) without inflating type II error as much as the Bonferroni method. Holm–Bonferroni corrections did not alter the main conclusions in this article.)
Publication bias—the selective publication of significant results—was assessed by Egger's test,26 for which significant t-values indicate that the obtained results are biased or unrepresentative of the universe of possible studies. Publication bias was examined only for polymorphisms in which the weighted mean OR was significantly different from 1. For each polymorphism, heterogeneity of ORs was assessed by I2, which represents the percentage of variability in effect sizes that is due to true heterogeneity (that is, in excess of that due to random error).
According to current guidelines,7 a full random effects meta-analysis is recommended only when there are five or more data sets per effect size (that is, five or more data sets per polymorphism). Accordingly, a full meta-analysis (for example, effect size analysis, Egger's test, I2, mixed effects moderator analyses and other analyses reported in the appendix of Supplementary Information) was conducted only for polymorphisms for which there were five or more data sets. Moderator analyses were limited by the information provided in the source studies. These analyses were confined to the analysis of effects of design (CC vs FB), sex, race (Asian vs Caucasian) and sample age (child–adolescent (<17 years) vs adult (17 or older)).
Sample age is a potentially important moderator variable because it can be used as a proxy measure of age of onset of OCD. (Too few studies reported age of onset for inclusion of this variable in the meta-analysis). Early vs later onset OCD are empirically defined subtypes of the disorder.8 Child–adolescent samples consist of early onset cases whereas adult samples likely consist of a mix of early vs later onset. The age cutoff of 17 years, which was dictated by the nature of the age groups in the association studies, comes close to the cutoff of 21 years, which is the empirically defined optimal cutoff for distinguishing early from late onset OCD.8
For polymorphisms that were investigated in fewer than five data sets, a secondary meta-analysis was conducted, limited to the computation of mean ORs for each polymorphism. This can provide useful preliminary information about whether a given polymorphism is a promising candidate for further investigation.
Characteristics of probands included in the meta-analyses
For the studies included in the main and secondary meta-analyses, the mean number of OCD probands was 113 (range 9–459). Regarding the descriptive features of the probands, the following weighted means were computed (weighted by sample size) for those studies presenting relevant data: males=52%, age at assessment=33 years, age at OCD onset=17 years, lifetime history of a tic disorder=17%, prevalence of OCD among first-degree relatives of probands=25%, mean score of probands at the time of assessment on the Yale–Brown Obsessive–Compulsive Scale=24 (indicating moderate-to-severe symptoms). The proportion of probands reporting particular obsessive–compulsive symptoms was as follows: checking=69%, contamination/cleaning=46%, symmetry/ordering=42%, hoarding=18% (a given proband may have had more than one symptom).
Mean effect sizes
Most (94%) controls in CC studies in the main and secondary meta-analyses were in Hardy–Weinberg equilibrium, and the pattern of results did not change when studies not in equilibrium were excluded (see Appendix). Twenty polymorphisms were examined in the main meta-analysis (Table 1). These were related to the regulation of serotonin (5-HTT, 5-HTTLPR, HTR1B, HTR2A and HTR2C), glutamate (SLC1A1), dopamine (DAT1, DRD2, DRD3 and DRD4), catecholamines (COMT and MAOA) and neurotrophin (BDNF). For HTR2A, the meta-analysis combined data from two polymorphisms (rs6311 and rs6313). These were combined because they are in complete linkage disequilibrium with one another,27 which means that the findings for one polymorphism are identical to those of the other. The weighted mean ORs for rs6311 and rs6313 did not significantly differ from one another; χ2(df=1)=2.13, P>0.10.
Table 1 shows that the mean ORs were significant for three polymorphisms; 5-HTTLPR (coded as triallelic; La vs Lg+S), HTR2A and COMT. Table 1 also shows nonsignificant trends for two dopamine-related polymorphisms (DAT1 and DRD3) and a glutamate-related polymorphism (rs3087879).
Forest plots of ORs for triallelic 5-HTTLPR and HTR2A appear in Figures 1 and 2, respectively. The size of the squares corresponds to the weighting assigned to each study and the width of the diamonds corresponds to the 99th percentile confidence interval of the weighted mean OR. Egger's test was nonsignificant for the triallelic 5-HTTLPR, t(6)=0.39, P>0.10, and for HTR2A, t(17)=0.86, P>0.10. This suggests that it was unlikely that the mean ORs for these polymorphisms were biased by the selective publication of studies obtaining significant results. Egger's tests for COMT are reported later in this article, with regard to the analyses of significant sex differences.
To facilitate the interpretation of the ORs in Table 1, they were converted to Cohen's d, where values of 0.2, 0.5 and 0.8 represent small, medium and large effect sizes, respectively.28 These values of d correspond to ORs of 1.437, 2.477 and 4.268. Table 1 shows that, in terms of d, all effects were small, even those that attained statistical significance. With a greater number of studies, and/or larger sample sizes in the studies, more of the mean ORs in Table 1 may have reached statistical significance. However, these would have represented very small effects.
Moderator (subgroup) analyses
The values of I2 (Table 1) suggests that for each polymorphism there was a good deal of heterogeneity in ORs, which raises the question of whether ORs significantly varied as a function of moderator variables. Moderator (subgroup) analyses were conducted for each polymorphism in the main meta-analysis. Method effects (CC vs FB designs) were nonsignificant for all 20 polymorphisms; χ2(df=1)<6.19, P>0.01. Sample age was not significantly related to ORs for any polymorphism, although there were two nonsignificant trends: for HTR1B, there was a trend (P=0.03) for ORs of adults (1.410) to be larger than those of children and adolescents (0.652). For HTR2C, there was a trend (P=0.05) for ORs of children and adolescents (1.550) to be larger than those of adults (1.087). For all 20 polymorphisms, there were no significant differences between White and Asian samples; χ2(df=1)<1.00, P>0.10.
Significant sex effects in ORs were obtained for COMT (χ2(df=1)=9.40, P<0.002) and for MAOA (χ2(df=1)=24.17, P<0.001). The COMT Met allele was associated with OCD only in males (Figure 3) and the MAOA T allele was associated with OCD only in males (Figure 4). In terms of Cohen's d, the effect size in males was small for COMT (0.241) and large for MAOA (0.582). For COMT for males, Egger's test was nonsignificant; t(7)=0.33, P>0.10. For MAOA for males, Egger's test was also nonsignificant; t(1)=0.26, P>0.10.
Mean ORs were computed for 210 polymorphisms that were examined in fewer than five data set (see Appendix for a complete list). Of these, 18 were significant (P<0.01). Polymorphisms significantly associated with OCD were those related to trophic factors (BDNF polymorphisms apart from the one examined in the main meta-analysis, as well as NGFR and NTRK2), GABA (GABRB3), glutamate (GRIK2), serotonin (HTR2A C516T), bradykinin (BDKRB2), acetylcholine (CHMR5, CHRNA1), glycine (GLRB), ubiquitin (involved in the normal functioning of cells; UBE3A), immunologic factors (TNFA) and myelinization (OLIG).
Mean ORs (and 99th percentile confidence intervals) for the 18 significant polymorphisms were as follows: BDKRB2 (rs8016905): 1.420 (1.004–2.011); BDNF (C11756G): 2.060 (1.194–3.557); BDNF (rs2049046): 1.477 (1.037–2.103); CHMR5 (rs2702285): 1.622 (1.008–2.611); CHRNA1 (rs3755485): 1.745 (1.181–2.581); GABRB3 (rs4304994): 1.426 (1.049–1.937); GLRB (rs11930311): 1.533 (1.070–2.198); GRIK2 (rs1556995): 4.364 (1.033–18.434); HTR2A (C516T): 3.247 (1.464–7.200); NGFR (rs3785931): 1.429 (1.029–1.984); NTRK2 (rs2378672): 4.215 (1.456–12.202); OLIG (rs1059004): 3.245 (1.074–9.798); OLIG (rs762178): 4.944 (1.567–15.597); OLIG (rs9653711): 4.510 (1.153–17.648); TNFA (rs361525): 2.980 (1.299–6.838); UBE3A (rs12899875): 1.705 (1.118–2.598); UBE3A (rs2526025): 1.619 (1.094–2.395); UBE3A (rs7175651): 1.800 (1.171–2.766). These findings may serve as useful leads for future investigation, although it should be cautioned that most of these results were based on only a single study, and so the reliability (replicability) of the findings remains to be determined.
OCD has been the subject of over 100 genetic association studies, involving the investigation of >200 polymorphisms. The present study is the first comprehensive meta-analysis of this substantial body of research. Findings indicated that OCD is associated with multiple genes, which is consistent with twin studies showing that OCD is shaped by additive genetic factors; that is, by multiple genes that incrementally increase the odds of developing the disorder.3 This study showed that OCD is associated with polymorphisms involved in serotonin modulation (HTTLPR and HTR2A) and, for males, polymorphisms involved in catecholamine regulation (COMT and MAOA). Sex differences for COMT is consistent with other research showing that the expression of this polymorphism is influenced by estrogens.10 Karayiorgou et al.29 speculated that MAOA may be associated with X-linked markers, which might account for the sex differences of MAOA and OCD. Three other polymorphisms had nonsignificant trends toward significance; DAT1, DRD3 and SLC1A1 (rs3087879). These results were based on 6–8 data sets, which may have been insufficient for detecting small effects. Such findings merit further investigation to determine whether they are reliably associated with OCD.
For the main meta-analysis, sample age (adults vs children and adolescents) was used as a proxy for age of onset; that is, adults likely consisted of a mix of early and late onset OCD, whereas child–adolescent samples consisted of early onset OCD. Sample age was a nonsignificant moderator variable. However, there were trends for HTR1B and HTR2B, which merit further investigation to determine whether they reach statistical significant if a larger number of studies are included. The secondary meta-analysis, which focused on polymorphisms examined in fewer than five data sets, found significant results for 18/204 polymorphisms. Such findings offer potentially useful leads for further investigation.
Data from the main and secondary meta-analyses provided the opportunity for evaluating the statistical power of OCD genetic association studies. The average study, with a mean of 113 OCD probands, had insufficient power to detect small effects (that is, ORs in the vicinity of 1.437). The power of the average study to detect such effects was <0.45. In association studies of psychiatric or general medical conditions, ORs linking a disorder and polymorphic locus typically range between 1.100 and 1.500.30 The average study included in the present meta-analyses lacked sufficient power to detect such effects. In future research, studies should be sufficiently powered to detect such small effects (for example, ORs as small as 1.100). The continued publication of studies underpowered to detect small effects may hamper efforts to understand the genetic basis of OCD. That is, null results from underpowered studies of a given polymorphism may dissuade researchers from further investigating that polymorphism, thereby leading to the failure to detect genes that make small, incremental contributions to the risk of developing OCD.
This study identified a number of important gaps in the research literature. For the main meta-analysis, there was evidence of substantial heterogeneity of effect sizes, which was not attributable to basic differences in study design (CC vs FB designs). Although this study found evidence that heterogeneity was partially due to sex differences, heterogeneity may also arise from pooling etiologically distinct forms of OCD. Advances in understanding the genetic basis of OCD may require the investigation of specific subtypes of the disorder, particularly the investigation of whether genetic profiles differ across the major schemes for subtyping OCD (early vs late onset; tic related vs unrelated).8 Further research is also needed to determine whether different classes of obsessive–compulsive symptoms (for example, checking, washing, symmetry and hoarding) differ in their genetic profiles. Twin research suggests that all these symptoms have a common genetic influence, but are also influenced by genes that are specific to each symptom subtype.3 Progress in understanding OCD may also be made by conducting genetic association studies of endophenotypes associated with the disorder, such as neuropsychological variables (for example, performance on tasks assessing set shifting or response inhibition).31
A limitation of this study was that some polymorphisms could not be meta-analyzed, because some studies did not report sufficient information for the computation of ORs for their findings.32, 33 Research into the genetics of OCD would be facilitated if the raw data of individual studies was made available for meta-analytic evaluations, which should include comprehensive descriptive information on the probands (for example, demographic features, age of onset, lifetime history of tic disorders, predominant types of OCD symptoms). Such detailed archival information would facilitate research into the question of whether particular subtypes of OCD are associated with specific polymorphisms. A further limitation of this study is that there were insufficient data for meta-analyzing haplotypes. Several studies have identified haplotypes associated with OCD,15, 33 but the replicability of such findings remains to be determined.
In summary, this study is the first comprehensive meta-analysis of OCD genetic association studies. This study indicated that OCD is associated with a number of polymorphisms, particularly those related to serotonin regulation and, more broadly, to catecholamine modulation. There was also evidence of sex differences. Progress in genetic association studies may be likely to occur if future studies are (a) sufficiently powered to detect small effect sizes, (b) designed to investigate potentially important moderator variables (for example, those defined by age of onset, comorbid tic or particular types of obsessive–compulsive symptoms), and (c) provide full information on nonsignificant results (as Supplementary Materials) so that future meta-analyses can pool such results in the search for small effects. Although there is a need to investigate novel polymorphisms in relation to OCD, it is equally important to determine whether the many nonsignificant results in the literature were artifacts of insufficient power to detect small effects.
I am grateful to the following investigators for providing unpublished details of their research, which facilitated the completion of this study: Beatriz Camarena, Sîan MJ Hemmings, Humberto Nicolini, Dan J Stein, Jeremy M Veenstra-VanderWeele and Susanne Walitza.
About this article
Supplementary Information accompanies the paper on the Molecular Psychiatry website (http://www.nature.com/mp)
BMC Medical Genomics (2017)