Introduction

Obsessive-compulsive disorder (OCD) is a common and impairing disorder characterized by persistent, intrusive thoughts and/or repetitive, ritualized behaviors. OCD is a heritable condition with an estimated heritability of 47% in a large twin study [1] and a heritability based on common single nucleotide polymorphisms (SNPs) of 16–28% [2, 3]. However, replicated specific genetic risk factors for OCD have yet to be identified. Several disorders co-occur with OCD such as anxiety disorders, mood disorders, anorexia nervosa (AN), tics, among others [4,5,6] and these disorders share genetic risk with OCD [7, 8].

For many psychiatric disorders, including OCD, it is thought that their genetic risk is continuously distributed in the general population, contributing to varying levels of symptom expression [9]. This is in line with the observation that obsessive-compulsive symptoms (OCS) are relatively common in the population. For example, 13 to 38% [10] of all adult individuals experience OCS, with even higher rates in younger individuals [11, 12], but only 1 to 2% meet the criteria for a clinical diagnosis of OCD13. Nevertheless, these subclinical symptoms can result in substantial distress and interference, even in individuals not meeting diagnostic criteria for the disorder [14].

Studies showing a shared genetic risk between OCS and diagnosed OCD support the hypothesis that clinical OCD may represent the extreme end of a continuous distribution of symptoms [12, 15] and that by considering sub-clinical OCS data we can increase the population available for study [16]. Like clinical OCD, OCS measured as a quantitative trait are heritable, with estimates of 30–74% from twin studies (total heritability; [17,18,19,20,21]) and a SNP-based heritability of 7–16% from genome-wide association studies (GWAS; [12, 15, 22]). Studies of quantitative OCS have also identified genome-wide significant variants. A study by den Braber et al. [22] first reported a genome-wide significant SNP (rs8100480) in MEF2BNB for OCS in the Netherlands Twin Register (NTR; N = 6931), although this was not replicated in a relatively small clinical sample of patients with OCD. When OCS in the NTR sample were meta-analyzed with diagnosed OCD in the Psychiatric Genomics Consortium (PGC) sample (N = 17,992), no genome-wide significant variants were identified [12]. A study by Burton et al. [15] reported a genome-wide significant hit (rs7856850) in PTPRD for OCS in the Spit for Science sample (N = 5018), which was also associated with OCD in a meta-analysis of independent clinical OCD and control samples (N = 11,980). Together these studies suggest that cohort- and community-based samples may be useful for identifying genetic risk for not only OCS but also OCD.

Here we present the results of the largest GWAS meta-analysis of OCS to date, combining the results from various large-scale population-based cohorts from Sweden, the Netherlands, England, and Canada that assessed OCS with a variety of questionnaires. To further characterize our GWAS results, we conducted gene-based and gene-set analyses, as well as genetic correlation analyses with 97 other traits. Further, polygenic risk score (PRS) analysis allowed us to assess the probabilistic susceptibility of OCS using the combined risk measure of variants associated with educational attainment and several psychiatric disorders that often co-occur and genetically correlate with OCD, such as depression (DEP; 15–41% comorbidity rate [23,24,25]), schizophrenia (SCZ; 8–26% comorbidity rate [26]), autism spectrum disorder (ASD; 17% comorbidity rate [27]), and attention-deficit hyperactivity disorder (ADHD; 6–21% comorbidity rate [23]). We further assessed the validity of using OCS in population-based samples as a proxy for clinical OCD diagnosis by comparing the OCS GWAS to the latest GWAS of OCD [2]. We assessed the association of the OCD polygenic risk score (PRS) with OCS in our samples and compared the genetic correlation patterns of OCS and OCD with other traits and disorders.

Methods

Cohorts & obsessive-compulsive symptom measures

Individuals included in this study stem from seven different European-ancestry cohorts, including four cohorts from the Swedish Twin Registry (STR; [28,29,30]), namely CATSS18 [31], CATSS24 [31], STAGE, YATSS, and one each from the Netherlands Twin Register (NTR; [32]), Spit for Science (SfS; [33, 34]), and TwinsUK [35]. The cohorts are predominantly population-based twin cohorts, except SfS, with a mean age between 10 and 57 years (see Table 1). CATSS is a prospective, longitudinal study of all twins born in Sweden since 1992. Here, we used data measured at age 18 (CATSS18), and/or age 24 (CATSS24), selecting only one measurement time point per individual (preferring the measurement at age 24 over age 18 if both measurements were completed as the CATSS24 cohort employed the Obsessive-Compulsive Inventory Revised (OCI-R) which was also used by other cohorts). Data from NTR [12] and SfS [15] were included in previous GWASs. See Supplementary Material for more detailed cohort descriptions.

Table 1 Overview of the individual cohorts.
Full size table

Several questionnaires were used to assess OCS across the cohorts. STR-CATSS18 employed the Brief Obsessive-Compulsive Scale (BOCS; [36]), STR-STAGE used a seven-item OCS instrument [1], while STR-CATSS24, STR-YATSS, and TwinsUK employed the OCI-R [37], excluding the hoarding and neutralizing sub-scales. In NTR, the Padua Inventory Revised was used [38, 39] in the form of the 12-item abbreviated and Dutch translated version [40] excluding the rumination items, leaving 9 items on checking, washing, precision, and intrusive thoughts. In SfS, parent- or self-reported obsessive-compulsive traits within the last 6 months of visiting the Ontario Science Center were assessed using 19 items from the Toronto Obsessive Compulsive Scale (TOCS), a 21-item questionnaire described elsewhere [15, 41, 42]. Two items related to hoarding were removed. To ensure reliable and valid symptom reporting, SfS participants below 12 years of age with self-reported OCS and above 16 years of age with parent-reported OCS were excluded (see Supplementary Table S1 for cohort-specific details on OCS questionnaires).

For all cohorts, individuals with one or more missing items were excluded and items were summed and standardized into a Z-score. Distributions of the raw obsessive compulsive item scores are shown in Supplementary Figs. S1–S9. The distributions of the Z-transformed sum scores are shown in Supplementary Figs. S10–S13.

Genome-wide association analysis

All participants were genotyped on SNP-arrays using DNA from saliva or blood. One part of the STR-CATSS samples was genotyped on the PsychChip genotyping array (N = 5683), another part was genotyped on the GSA genotyping array (N = 412). For the GWAS analyses STR-CATSS cohorts (CATSS18, CATSS24) were pooled over each genotyping platform (GSA, PsychChip), forming two separate CATSS datasets (STR-CATSS-GSA and STR-CATSS-PC). Each of the seven datasets (STR-CATSS-GSA, STR-CATSS-PC, STR-YATSS, STR-STAGE, NTR, SfS, and TwinsUK) underwent stringent quality-control (QC), including the removal of non-European ancestry outliers based on PCA and imputation using the Haplotype Reference Consortium (HRC; STR, NTR) or the 1000G (SfS, TwinsUK) reference sets (see the Supplementary Material for more detailed information). Together, all cohorts comprised 33,943 individuals (STR: N = 16,888, NTR: N = 8550, SfS: N = 5171, and TwinsUK: N = 3334) with complete phenotypic and genotypic information (see Table 1 and Supplementary Material for a more detailed description of each cohort).

We used GCTA-fastGWA [43] to perform a mixed-linear-model GWAS within each cohort separately. GCTA-fastGWA controls for population stratification by principal components (PCs) and for relatedness by a SNP-derived genetic relationship matrix (GRM). In STR, NTR, and TwinsUK a sparse family GRM was defined, and each GWAS analysis included the first 10 genetic PCs, sex, age, age squared, and genotyping batches as covariates. In a sparse GRM, all off-diagonal values below 0.05 are set to 0 (default), thereby capturing the same proportion of phenotypic variance as by pedigree-relatedness and accounting for the close relatedness of individuals in the data. In SfS, analyses were performed on unrelated individuals. For sibling pairs, the first enrolled sibling from each family was selected for further analysis (see Table 1 for the number of siblings included). Siblings were removed in this cohort as the high genetic resemblance between them may dominate the genetic variation covered and adjusted for by the GRM, and with only few sibling pairs this adds to the uncertainty of the estimate. Here, the GCTA-fastGWA linear mixed model was performed using a full GRM and sex, age, respondent (parent vs. child reporting), genotyping array type, PCs 1–3, and projected PCs 1–3 (see supplement for details) as covariates.

For each of the seven resulting GWAS summary statistics, variants were filtered on minor allele frequency (MAF) > 1%, and imputation-quality (INFO) score >0.8. For strand ambiguous (A/T and C/G) SNPs, those with a MAF ≥ 0.4 were removed, while the frequencies of those with a MAF < 0.4 were compared to frequencies in the HRC reference [44]. Same strand orientation was assumed if the frequencies matched (i.e., the minor allele was the same in both data sets). All SNPs with matching frequencies were retained while mismatched SNPs (i.e., the minor allele was different in either dataset), assumed to be reported on different strands, were flipped according to the orientation reported in the HRC reference. Following removal of poorly genotyped SNPs, all datasets were aligned to the HRC-reference.

With these harmonized datasets, we conducted an inverse variance weighted meta-analysis utilizing METAL [45], a tool included within the Rapid Imputation for COnsortias PIpeLIne (Ricopili; [46]). To identify any residual population stratification or systematic technical artifacts, we inspected the genomic control factor (Lambda and Lambda1000). The genome-wide significance threshold was set to p < 5 × 10−8.

Gene-based analyses (MAGMA/FUMA)

We performed gene-analysis and gene-set analysis using Multi-marker Analysis of GenoMic Annotation (MAGMA) [47] v1.08 as implemented in Functional Mapping and Annotation of Genome-Wide Association Studies (FUMA; [48]) v1.3.7. To test genetic associations at the gene level for the combined effect of SNPs in or near protein coding genes, we applied default settings (SNP-wise model for gene analysis and competitive model for gene-set analysis). Gene-based p-values were computed by mapping SNPs to their corresponding gene(s) based on their position in the genome. Positional mapping was based on ANNOVAR annotations, and the maximum distance between SNPs and genes was set to 10 kb (default). A multiple regression model was employed while accounting for linkage disequilibrium (LD) between the markers. The 1000 Genomes phase 3 reference panel [49], excluding the MHC region, was used to adjust for gene size and LD across SNPs. Using the result of the gene-based analysis (gene level p-values), competitive gene-set analysis was performed with default parameters: 15,496 gene sets were tested for association. Gene sets were obtained from MsigDB v7.0 (see www.gsea-msigdb.org for details), including ‘Curated gene sets’ consisting of nine data resources including KEGG, Reactome, and BioCarta, and ‘GO terms’ consisting of three categories (biological processes, cellular components, and molecular functions).

Heritability and cross-disorder analyses

Heritability estimates

Heritability estimates of each cohort were extracted from the GCTA association output. GCTA uses the restricted maximum likelihood (REML) approach [50] to estimate heritability in the GRM that is supplied to correct for relatedness in the linear association test. For the twin cohorts (STR, NTR, and TwinsUK), this means that heritability was based on the sparse GRM. For SfS, with no related individuals, the heritability was based on the full GRM. For all heritability estimates, the same covariates as in the GWAS analyses were used. We also calculated the SNP-based heritability of the OCS meta-analysis using LD score regression (LDSC; [51]). LDSC bases its calculation of SNP-based heritability on the estimated slope from the regression of the SNP effect from the GWAS on the LD score.

Cross-trait genetic correlations

With LDSC [51] we calculated genetic correlations between OCS and 97 traits, including psychiatric, substance use, cognition & socio-economic status, personality, neurological, autoimmune, cardiovascular, anthropomorphic, and fertility phenotypes (see Supplementary Table S4 for a list of the source GWASs used). 15 of the included traits pertain to neuroticism and include the neuroticism sum score, worry- and depressive subclusters, as well as individual neuroticism items. The genetic correlation is based on the estimated slope from the regression of the product of Z-scores from the two GWASs of interest on the LD score. It represents the genetic correlation between two traits based on all polygenic effects captured by the included SNPs. Because imputation quality is correlated with LD score, and low imputation quality generally yields lower test statistics, imputation quality is a confounder for LD score regression. We therefore filtered on INFO > 0.9, if INFO was available, and MAF > 0.01. The SNPs from the European HapMap3 [52] were used as a reference. We further compared the genetic correlation patterns between OCS and OCD with all other 96 traits, to identify concordant and discordant patterns. OCD results were based on the publicly available summary statistics from the PGC [2].

Cross-phenotype polygenic risk score analyses

To further explore the genetic relationship between OCS in each of our datasets and other (psychiatric) phenotypes, we calculated polygenic risk-scores (PRS) based on large-scale GWAS summary statistics of OCD (Ncases = 2688, Ncontrols = 7037; [2]), DEP (Ncases = 170,756, Ncontrols = 329,443; excluding 23andMe [53]), SCZ (Ncases = 53,386, Ncontrols = 77,258; [54]), ASD (Ncases = 18,381, Ncontrols = 27,969; [55]), ADHD (Ncases = 19,099, Ncontrols = 34,194; [56]), and educational attainment (EA; Ntotal = 245,621; [57]) using PRSice2 [58] and evaluated their association with OCS in our cohorts (STR-CATSS-GSA, STR-CATSS-PC, STR-STAGE, STR-YATSS, NTR, SfS, and TwinsUK). We pre-selected p-value thresholds based on the best performing thresholds reported in the primary publications (EA: P = 1; ADHD, ASD, OCD, SCZ: P = 0.1; DEP: P = 0.5). The PRS scores were calculated as the weighted sum of the risk allele dosages.

For STR, NTR, and TwinsUK, we employed a generalized estimating equation (GEE) [59] in R to evaluate the relationship between the PRS scores and OCS scores in each cohort. The GEE analysis takes into account the resemblance within clusters, accounting for the relatedness in the datasets. Robust standard errors (sandwich-corrected) and Z-scores are reported. For SfS we applied linear regression, as conducted within the PRSice2 pipeline, to evaluate the relationship between PRS and OCS. Contributions of the PRSs were measured through comparison of the R² of the full model (including the PRS, and covariates) minus the null model (including only covariates). The same covariates that were previously included in the respective GWASs were used. We combined the PRS estimates across all target datasets using an inverse variance weighted meta-analysis using the metagen package in R. Cochran’s Q test [60] and Higgin’s I² [61, 62] were used to examine a possible heterogeneity in PRS estimates across the cohorts. Q is calculated as the weighted sum of the squared differences between individual cohort effects and the pooled effect across cohorts, with the weights being those used in the pooling method. The I² statistic describes the percentage of variation across studies that is due to heterogeneity rather than sample variation and does, unlike Q, not inherently depend on the number of measures included in the meta-analysis. We calculated a fixed effects model to evaluate the association of each PRS with OCS, regardless of observed heterogeneity. We further calculated a random effects model if there was substantial observed heterogeneity across study sites (I² > 0.5 and/or PQ < 0.05).

Comparability of cohorts

To identify if the summary statistics from any of the included cohorts substantially deviated from the rest, we performed leave-one-out (LOO) GWAS meta-analyses and subsequently used those datasets to conduct a set of sensitivity analyses. First, we performed sign-test analyses on the top SNPs (inclusion threshold of p = 0.0001, p = 0.00001, and p = 0.000001) using the replication module of the RICOPILI pipeline. Sign-tests allow for quantification of the number of genomic regions that are independent across the different p-value thresholds and identification of how many genomic regions within the replication study have the same direction of effect as the discovery. The output (in the form of a ratio) provides an estimate of the percentage of genomic regions with the same direction of effect between any two datasets. A sign-test is a binomial test with the null hypothesis = 0.5, with a ratio > 0.5 indicating a positive sign test (convergence), while a ratio < 0.5 indicates divergence. We conducted two sets of sign-tests, one comparing the direction of effect for each pair-wise combination of cohorts, and one comparing each LOO meta-analysis with the respective sample that was left out. While fluctuations in the sign-tests across different p-value thresholds are expected, depending on the true association of each SNP with the phenotype, we mainly aimed to assess whether a specific cohort markedly deviated from the rest.

Following the same procedure as for the cross-trait PRS analyses described above (see previous method-section on cross-trait PRS analyses for details), we conducted LOO PRS analyses to evaluate the relationship between the PRS scores of each LOO GWAS and standardized OCS scores in the left-out cohort. Further, we conducted genetic correlation analyses between each LOO OCD meta-analysis (LOO_NTR, LOO_SfS, LOO_STR, and LOO_TwinsUK) and the same set of 97 phenotypes as described earlier to explore a possible heterogeneity in correlation patterns depending on the included OCS cohorts.

Results

Genome-wide association results

The final OCS GWAS meta-analysis was based on 33,943 individuals with complete phenotypic and genomic data available and included 6,232,765 associations of autosomal SNPs. No significant inflation was observed (λ = 1.027, λ1000 = 1.001, LDSC intercept = 1.0047, see Supplementary Fig. S14 for a QQ-plot). No SNP reached genome-wide significance (p < 5 × 10−8; see Fig. 1 for a Miami-plot including the Manhattan-plot of the GWAS in the upper panel). The SNP with the lowest p-value was rs113538937 (p = 6.36 × 10−8) on chromosome 4 (see Supplementary Fig. S15 for a regional association plot and forest plot). The region tagged by this SNP spans 207.6 kb (LD R² > 0.6) and entails the genes SH3BP2, ADD1, MFSD10, NOP14-AS1, NOP14, GRK4, HTTAS, and HTT. Another 26 independent SNPs with a p-value < 1 × 10−5 were identified (see Supplementary Table S2 for a list of association results).

Fig. 1: Miami plot of the association results from the GWAS meta-analysis (upper panel) and of the gene-wide association analysis (lower panel) of OCS.
figure 1

The y-axes represent -log10 p-values for the association of SNPs/genes with OCS. The x-axis represents chromosomes 1 to 22. In the upper plot, the p-value threshold for genome-wide significance (p = 5 × 10−8) is represented by the horizontal red line, suggestive significance (p = 1 × 10−5) by the blue line. In the lower panel, Bonferroni-corrected gene-wide significance (p = 2.708 × 10−6) is represented by the horizontal red line, suggestive gene-wide significance (p = 1 × 10−3) is indicated by the blue horizontal line.

Full size image

Gene and gene-set analyses

Gene-based tests were conducted to test whether any protein-coding gene carries a load of common variation associated with OCS. Using MAGMA v1.08 within FUMA v1.3.7, input SNPs were mapped to 18,464 protein coding genes. No gene reached the Bonferroni-corrected significance threshold (p = 0.05/18,464 = 2.71 × 10−6) in the gene-based test (see Fig. 1 for a Miami-plot including the Manhattan-plot of the gene-based test in the lower panel, and Supplementary Fig. S14 for a QQ-plot). Further, 15,496 gene sets (Curated gene sets: 5500, GO terms: 9996) from MsigDB v7.0 were tested for association. One curated gene set (Mootha Gluconeogenesis) reached Bonferroni-corrected significance (Ngenes = 28, Beta = 0.7872, SE = 0.1611, Pbon = 0.008).

Heritability and cross-disorder analyses

Heritability estimates

For the twin cohorts, the additive genetic variance of OCS, estimated based on the sparse genetic relatedness matrix from the association tests in GCTA, ranged between 0.33 (NTR) and 0.43 (TwinsUK), with estimates for the STR cohorts in between: STR-CATSS: 0.35 (GSA chip sub-sample), STR-CATSS: 0.41 (PsychChip sub-sample), STR-YATSS: 0.39, STR-STAGE: 0.39. Note that these heritability estimates based on the twin cohorts are largely driven by the twin resemblance (~0.5 between DZ twins and siblings, 1.0 for MZ twins, and 0 between unrelated individuals). The heritability for SfS, only including unrelated individuals, was 0.083 (SE = 0.053, P = 0.0516). The SNP-based heritability estimate of the GWAS meta-analysis using LDSC resulted in a total observed scale h² of 0.041 (SE = 0.0144, Z = 2.85, P = 0.0044).

Cross-disorder genetic correlations

We estimated genetic correlations using LDSC between the current OCS GWAS and OCD as well as 97 other phenotypes, including psychiatric, personality, psychological, substance-use, neurological, cognition, socioeconomic status, autoimmune, cardio-vascular, anthropomorphic, and fertility traits. Of these, 24 exceeded the FDR-corrected significance threshold (Fig. 2). As expected, OCS were most strongly associated with OCD out of all psychiatric disorders, followed by anxiety, DEP, major depressive disorder (MDD), SCZ, and AN. Significant positive genetic correlations were also observed for all 15 neuroticism phenotypes (14 sub-items and the neuroticism total score). Higher correlations emerged for all worry-related items, and slightly lower correlations for all depressive-related items. A significant positive correlation was further observed for “cigarettes per day” and “tiredness”, while “subjective well-being” yielded a significant negative correlation (see Fig. 2, and Supplementary Table S4). We further compared the genetic correlation patterns of OCS and OCD with all other 97 traits, to identify concordant and discordant patterns. There was a clear relationship between the two correlation-patterns of OCS and OCD (see Fig. 3), with 66 traits showing concordant directions of correlation (i.e., both correlations above 0 or both below 0), and 30 traits showing a discordant direction of correlation (i.e., one correlation above 0 and one below 0). Especially strong concordance was observed for the psychiatric and personality traits. Of the significant correlations with OCS, only ‘neuroticism loneliness’ (non-significant for OCD) and ‘cigarettes per day’ (significant for OCD) showed a different direction of effect with OCD (both positive with OCS, but negative with OCD).

Fig. 2: Genetic correlations between OCS and a variety of other traits and disorders.
figure 2

Genetic correlations (rg) between OCS and a broad range (N = 97) of other phenotypes, assembled into 11 groups (psychiatric, substance, cognition/socioeconomic status (SES), personality, psychological, neurological, autoimmune, cardiovascular, anthropomorphic, fertility, and other). Error bars represent 95% confidence intervals; red circles indicate significant association after FDR correction for multiple testing.

Full size image
Fig. 3: Comparing concordance and discordance between genetic correlation estimates of OCS and OCD.
figure 3

Comparison between genetic correlation estimates (rG) of OCS and OCD for 97 other phenotypes, color-coded according to 11 groups (psychiatric, substance, cognition/socioeconomic status (SES), personality, psychological, neurological, autoimmune, cardiovascular, anthropomorphic, fertility, and other). On the x-axis the genetic correlation estimates for OCD are displayed, on the y-axis for OCS. Green quadrants indicate concordance, red quadrants discordance in the direction of genetic correlations between OCS and OCD.

Full size image

Polygenic risk score analyses

We calculated a range of PRSs, based on publicly available summary statistics of OCD, DEP, SCZ, ASD, ADHD, and EA and examined their relationship with OCS in each cohort. GEE (STR, NTR, TwinsUK) and linear regression analyses (SfS) revealed significant (Bonferroni-corrected p < 0.05/7 = 0.0071) associations between OCS and PRS based on OCD, DEP, SCZ, and EA, but not consistently across all target datasets (see Supplementary Table S4). In the meta-analysis, summarizing the PRS results across all target cohorts, the OCD (Pfixed = 3.06 × 10−5) and SCZ (Prandom = 3.69 × 10−6) PRS showed significant associations with OCS. The DEP PRS also showed a significant association with OCS, however, there was significant heterogeneity across the four cohorts (PQ = 0.0020), and the random effects model failed to reach significance (Prandom = 0.0801). The other traits’ PRSs (ASD, ADHD, EA) were not significantly associated with OCS in the PRS meta-analysis (All meta-analyzed results are shown in Table 2).

Table 2 The table shows the PRS results for OCS (Leave-one-out), obsessive-compulsive disorder (OCD), depressive disorder (DEP), schizophrenia (SCZ), autism-spectrum disorder (ASD), attention-deficit hyperactivity disorder (ADHD), and educational attainment (EA), meta-analyzed across all target datasets for pre-selected p-value thresholds (Pthreshold).
Full size table

Comparability between cohorts

No genome-wide significant heterogeneity was observed in the OCS GWAS meta-analysis (see Supplementary Fig. S16 for Manhattan plot and QQplot). The LOO PRS meta-analysis showed a significant positive association between the LOO PRS and OCS (Pfixed = 5.03 × 10−7) (see Table 2 for the PRS meta-analysis results and Supplementary Table S4 for individual results for each target cohort).

Given the low power for the main GWAS analysis, the power for LOO GWAS analyses and signtests between partial analyses is even lower, making it difficult to draw any definitive conclusions from the results. While no individual study markedly stands out from the rest, results of the sign tests analyses fluctuate with estimates ranging from 0.28 to 0.68 (and between 0 and 1 for P = 1 × 10−6). See Supplementary Figs. S17–S20 for Manhattan-plots and QQ-plots of the LOO GWASs and Supplementary Tables S5 and S6 for sign-test results.

We further calculated genetic correlations between each LOO OCS meta-analysis and the same 97 phenotypes (described above) to compare the individual influence of each cohort on the overall correlation estimates. When not considering the individual sub-items of neuroticism, the LOO GWAS excluding SfS showed 13 significant correlations, the GWAS without TwinsUK showed eight significant correlations, the GWAS without STR showed two significant correlations, while the GWAS excluding NTR did not significantly correlate with any of the traits (see Supplementary Fig. S21). As sample sizes for the LOO GWAS meta-analyses varied (LOO STR: N = 17,055; LOO NTR: N = 25,393; LOO SfS: N = 28,772; LOO TwinsUK: N = 30,609), it was expected that the power to detect significant correlations for each LOO GWAS differs. For almost all genetic correlations, each LOO GWAS meta-analysis showed the same direction of effect. For all correlations with psychiatric disorders and neuroticism phenotypes, the LOO analysis excluding NTR showed slightly higher estimates (but also larger confidence intervals).

Discussion

This is the first meta-analysis aiming to identify the genetic underpinnings of OCS in the general population. Although we could not replicate previous findings, two SNPs previously associated with OCS reached suggestive significance in this meta-analysis: the SNP found by den Braber et al. (rs8100480, p = 2.56 × 10−8; [22]) had a p-value of 0.0155 in the current study, while the SNP reported in Burton et al. (rs7856850, p = 2.48 × 10−8; [15]) had a p-value of 0.00029. The SNP with the lowest p-value (p = 6.36 × 10−8) in the current meta-analysis was rs113538937 on chromosome 4, tagging eight genes which have previously been associated with alcohol use (HTT), smoking (GRK4, NOP14, NOP14-AS1, ADD1, SH3BP2), worry (HTT), and measures of socioeconomic status and education (HTT, GRK4, NOP14, NOP14-AS, ADD1). In the gene-based tests, no gene achieved genome-wide significance. One gene-set reached Bonferroni-corrected significance. Expression of genes in this set is high at sites of insulin-mediated glucose disposal [63]. Dysregulated insulin signaling in the etiology of OCS has been suggested by a previous study describing a genetic overlap of OCS with an insulin signaling-related trait in children and adolescents [64]. Epidemiological studies similarly support the role of dysregulated insulin signaling in OCD and OCS. Specifically, patients diagnosed with OCD were found to have a significantly higher risk of developing type 2 diabetes compared to population controls [65].

Our results show that OCS share genetic risk with OCD. Polygenic risk for OCD was associated with OCS; and OCS and OCD case/control status showed a substantial genetic correlation (rG = 0.72, p = 0.0007). This is comparable to estimates reported in recent studies (rG = 0.61, p = 0.017 [12]; rG = 0.83, p = 0.07; [15]). Notably, OCS did show the highest correlation coefficient with OCD (rG = 0.72), but the strength of this genetic correlation was statistically not significantly different from that with anxiety (rG = 0.62), DEP (rG = 0.51), SCZ (rG = 0.34), or AN (rG = 0.30), as confidence intervals overlapped. While we did not exclude participants with any of these diagnoses, the most likely explanation is that these common comorbidities of OCD share genetic underpinnings with OCS. We also observed a high concordance in direction and strength of the correlation patterns of OCS and OCD with other phenotypes. The concordance between the OCD and OCS GWASs indicate that the genetic variation captured by our symptom based GWAS in the general population reflects that of the disorder OCD. However, given the present sample size, these values must also be interpreted with caution. A fact that is further illustrated by our PRS analyses: scores calculated based on the most recent SCZ GWAS more accurately predicted OCS than scores calculated based on the most recent OCD GWAS, which is considerably underpowered. This suggests that larger cohorts are needed for the accurate estimation of these associations.

SNP-based heritability for OCS in the current sample of 4.1% was significant, but lower than previous studies (7–16%; [12, 15, 22]), also standing in stark contrast with heritability estimates observed in twin studies (37–41%; [66]). This may be explained by the heterogeneity across the eight cohorts. Specifically, these samples differed in their ascertainment (twins vs. community-based samples), instruments employed, as well as age range of participants. While no specific pattern emerged in our compatibility analyses, the sign tests indicated some level of heterogeneity. Further evidence for heterogeneity was found in the leave-one-out analysis (Supplementary Fig. S3), supporting the view that there were some differences in cohorts with respect to the genetic correlation with specific variables. For leave-STR-out, the correlations with IQ and some related phenotypes were lower, whereas for leave-NTR-out, the correlations with neuroticism were increased. For most phenotypes, however, the leave-one-out genetic correlations did not yield strong or significant change as evidenced from the confidence intervals. In the PRS analyses, results deviated mostly for the SfS sample, which is the youngest and only non-twin sample included. It also uses the only questionnaire that is coded from strengths to weaknesses (from −3 meaning far less often than average to +3 meaning far more often than average), which could have led to further differences between the cohorts.

In line with previous research reports, we found a lower SNP-based heritability for OCS [12, 15] than for clinical OCD (0.28–0.37) [67, 68]. The reason for the disparity in SNP heritability between traits and diagnosis is unclear but could be explained by the fact that clinical diagnosed OCD represents the extreme of the OCS distribution in the general population. The conversion to liability scale SNP-h2 [50] may also have contributed to uncertainty in the estimates, as the uncertainty of a population prevalence close to zero may improperly inflate the correction factor from observed to liability scale. Previous studies of related psychiatric disorders, such as SCZ, or MDD, revealed that the genetic risk is higher for more severe and chronic cases [69, 70]. In addition to this consideration of impairment, differences including informant (parent/self vs. clinician), type of measurement (categorical vs. quantitative), and timing (cross-sectional vs. lifetime symptoms) could contribute to the divergence in SNP-based heritability estimates of OCS and OCD.

The present study had several limitations. First, despite being the largest meta-analysis of OCS including, to our knowledge, all samples currently available worldwide, the sample size is still relatively small for estimating heritability and detecting specific significant genetic markers. The meta-analysis also currently lacks the integration of non-European samples. The present results thus call for replication in, and extension to larger and more diverse cohorts. Second, previous research has suggested that OCS dimensions (e.g. contamination, checking, harm or symmetry) may be etiologically heterogeneous [12, 65, 66, 71]. As such, future studies might aim to identify the genetic underpinnings of specific OCS dimensions. This could also tackle some of the heterogeneity issues that may have caused imperfect overlap between cohorts, lower SNP-h2, and reduced genetic correlation with clinical OCD. Third, no information was available on the presence of common clinical comorbidities in all samples. This precluded detailed analysis of the identified genetic overlap. Future studies in larger cohorts should also investigate in more detail how OCS relate to other phenotypes, for instance addiction and personality disorders. Fourth, the observation that patients often establish non-random relationships with persons affected by the same or another mental disorder [72], might extend to people with OCS and contribute to the observed genetic correlations of OCS with anxiety, DEP, SCZ and AN. However, the LD-score method does not investigate the impact of assortative mating [73]. Therefore, assessing the degree to which this phenomenon may have influenced the genetic correlation estimates was beyond the scope of the present study. Future investigations of larger data sets for OCS and other psychiatric disorders are needed to refine the analysis of shared and specific genetic risk as well as communalities and specificities of the respective disorders. Finally, for many of the samples included in our study the distribution of quantitative OCS scores followed a right-tailed distribution, with most of the subjects reporting low OC symptom scores (see Supplementary Material for more detail). Although we captured substantial variability that would not have been present in more conventional case-control analyses, less variability amongst individuals with lower scores may have limited our power to identify positive genetic associations. The right-tailed distribution is due to the fact that most OCS scales are designed to screen for disorder in a clinical setting, rather than capture population variation in symptoms. The one exception in our study was the Spit for Science cohort, where the TOCS measure was explicitly designed for capturing variation in non-clinical samples [41, 42]. Despite these limitations, we argue for the inclusion of data based on conventional screening measures in order to achieve the necessary sample sizes for genomic studies, as they have been widely adopted and still capture greater variability in OCS compared with studies based on dichotomized case/control samples.

To summarize, OCS have a significant polygenic contribution and share genetic risk with diagnosed OCD, supporting the hypothesis that OCD represents the extreme end of widely distributed OCS in the population.