Uncovering obsessive-compulsive disorder risk genes in a pediatric cohort by high-resolution analysis of copy number variation

Participants were recruited from four academic child psychiatry sites: The Hospital for Sick Children, McMaster University, University of Michigan, and Wayne State University. Subjects were enrolled via clinics (site clinics, other mental health providers and primary care physicians), the internet (e.g., www.​umhealthresearch​.​org at University of Michigan), hospital and community bulletin boards, and paid and public service advertisements in local media.

All enrolled individuals (164 females and 143 males) had symptoms first identified before age 18 (mean, 7.9 ± 3.5 years). Respective institutional ethics review boards approved all procedures. Informed consent was provided by capable adolescents. For younger children, parents or other legal guardians provided written informed consent, and the children gave verbal assent prior to participating in the study. Criteria for diagnosis are in the Additional file 1.

Our unrelated population control data were from three cohorts: Cooperative Health Research in the Region Augsburg (KORA) [21], the collaborative genetic study of nicotine dependence (COGEND), [22] and the Ontario Population Genomics Platform (OPGP) [23]. The same quality control procedures and CNV calling algorithms applied to our subjects had been applied to these controls.

We genotyped all 174 trios and 58 additional unrelated probands using the CytoScan HD array, and 75 unrelated probands using the OMNI 2.5 M array. We employed multiple algorithms to call CNVs from the CytoScan HD and OMNI 2.5 M microarray data. We defined a stringent set of variants wherein each variant was called by at least two algorithms (Additional file 2: Table S1) [24]. We defined the ancestry and relatedness of the samples using PLINK [25] following methods previously described in other studies (see Additional file 1) [26, 27]. This information was used to exclude related probands or controls and detect any sample mismatches.

To define rare CNVs, we first computed frequency using a pooled set of stringent CNVs from cases and controls, matching for ancestry, platform, and sex (Additional file 1: Figure S1). We then removed those CNVs present at >0.5 % frequency, using the 50 % reciprocal overlap criteria [11, 28]. Next, we required that CNVs overlap a region that is at least 75 % copy number stable according to the CNV map of the human genome [28]. Finally, we used a cut-off of 15 kb and 10 probes for any remaining CNVs, yielding a comprehensive and high-quality list of rare CNVs from our case and control cohorts. We first ascertained CNVs from all samples, including individuals of non-European or mixed ancestry and utilized the entire control set. We then restricted our analysis to cases of European ancestry (for the subsequent gene set analysis). Those CNVs listed in Table 1 are not present in population controls unless stated otherwise in the table. We validated de novo events and possible risk variants in proband and parental samples (when available) using a secondary confirmation method (a SYBR-green-based real-time quantitative PCR (qPCR) or a TaqMan real-time PCR assay) [27, 29]. Primer sequences are available upon request from SWS. Twenty-five of 28 CNVs of interest were validated (including all in Table 1 and 4/5 that were tagged as being de novo). Suspected mosaicism in Case D was confirmed using digital droplet qPCR (Additional file 1: Figures S2 and S3). All coordinates refer to human genome assembly build GRCh37, hg19.

We analysed the burden of all CNVs, and of deletions and duplications independently. We also looked at CNV burden in the following size ranges: >15 kb (all), 15–100, 100–500, >500 kb, and >1 Mb. A binomial logistic regression analysis determined whether the differences in gene rates (the average number of genes intersected by CNVs per subject) between cases and controls were the result of a biological mechanism or statistical chance. We corrected for the number of CNVs and the total length of CNVs in the subject in our model.

We tested gene set enrichment using the rare variants detected for individuals of European ancestry (Additional file 1: Figure S4). This analysis focused on events impacting protein-coding exons, and considered deletions and duplications separately. The analysis consists of a case-control burden test with correction for the different platforms and also for global CNV burden modelled as total CNV gene count by subject (i.e., penalization for large CNVs that may drive results in an unspecific way). The gene sets tested were comprised of genes associated with neurological functions (Additional file 3: Table S2) [11, 30]. Gene set burden enrichment was performed using a logistic regression analysis deviance test corrected for gene count and global burden (Additional file 1). The genotyping platform used for CNV calling was modelled as a covariate. We undertook the analysis twice; first without exclusion criteria, and second, by excluding neurodevelopmental risk loci previously described [20, 31, 32]. This provided an opportunity to see whether any enrichment signal was being driven by loci already recognized. For final results, we utilized 15 % as the Benjamini-Hochberg false discovery rate significance threshold.

For samples with notable CNV findings (Table 1), we selected those 10 families for which adequate sample material was available from the proband and both parents for additional exome sequencing. This was conducted using the Illumina HiSeq2500, following exome capture with Agilent SureSelect Human All Exon V5 target enrichment kit. We aligned sequences to the hg19 reference using Burrows-Wheeler Aligner (BWA) v0.5.9, and used the Genome Analysis Toolkit (GATK) v1.1–28 to detect single nucleotide and small insertion and deletion variants. Rare variants were defined as those with a frequency of <0.01 in 1000 genomes, National Heart, Lung, and Blood Institute (NHLBI) Exome Variant Server and Exome Aggregation Consortium (ExAc datasets; http://​exac.​broadinstitute.​org/​). A Residual Variation Intolerance Score (RVIS) [33] and Probability of Loss of Function (LoF) Intolerance score (pLI) [34] were retrieved for each variant discovered. We confirmed the presence of all variants in Table 2 in proband and parental samples using Sanger sequencing [35].

To explore the contribution of rare CNVs to OCD, we used a comprehensive CNV calling procedure with experimental validation of relevant variants. CNVs from 307 probands passed quality control, including 174 for whom DNA from both parents was also genotyped. Of these 307 probands, 259 were of European ancestry. Data from 1773 European controls were treated as those from the cases.

We uncovered 729 rare CNVs from these European cases (Additional file 2: Table S1), and 5182 from the European controls. This represented a mean of 2.81 and 2.92 rare CNVs per individual, respectively. We considered the number of genes intersected by CNVs in cases compared with controls (Additional file 1: Table S3), but found no statistically significant difference in this respect when looking at all CNVs, or at deletions and duplications independently. We also looked for differences in a particular CNV size range but found no greater global burden for rare CNVs in cases compared with controls (Additional file 1: Table S3).

We screened 174 parent-child trios of all ancestries for rare de novo CNVs. In four of these probands (2.3 %) (Table 1, cases A–D), we uncovered a single de novo CNV (sometimes in addition to inherited variants).

Various genomic syndromes have been observed in neurodevelopmental disorders including autism, schizophrenia, and epilepsy [13]. We sought to identify in our cohort rare CNVs that have been noted in some of these syndromes using the DECIPHER database containing expert-curated microdeletions and microduplications involved in developmental disorders [36]. We also compared our results with a list of OCD and ASD risk genes [1, 11, 32, 37]. We identified four individuals (Table 1, cases E–H), each with a CNV whose breakpoints were compatible with those in DECIPHER. We identified other CNVs in genes with possible relevance to OCD (Table 1, cases I–M), in gene targets of the enriched gene sets, particularly the fragile X mental retardation protein (FMRP) (cases N–P), and in a new plausible candidate, BTBD9 (cases Q–R).

In 10 of the trios where we discovered a potentially relevant CNV (i.e., de novo, or involving gene(s) associated with neurological function), we followed our CNV analysis with exome sequencing of the proband and parents. We focused our analysis on any de novo mutations or any predicted loss-of-function mutations in genes considered to play a role in neurological function. From these families, we identified six de novo variants and seven rare inherited loss-of-function mutations in genes with a neurological role (Table 2). We subsequently obtained RVIS and pLI scores for each of the genes to show how able they are to tolerate functional genetic variation and loss-of-function mutations, respectively. Only one of the genes identified from exome sequencing, AFF2, had an RVIS score in the tenth percentile (among the most intolerant genes) and a pLI score of above 0.9 (a score considered to be suggestive of pathogenic LoF mutation) [34].

Our initial analysis sought to identify gene set enrichment, in cases compared with ethnically matched controls, from CNVs that impact a gene’s coding sequence (Table 3; Additional file 1: Table S4 and Additional file 4: Table S5). We found such enrichment for CNVs in FMRP target genes (nominal p = 1.85 × 10⁻⁰³), both when we included or excluded known neurodevelopmental risk loci from the analysis. When such loci were included, cases were 1.75 times more likely to have one FMRP target impacted by a CNV than were controls, and 4.38 times more likely to have two FMRP targets impacted (Additional file 4: Table S5). Much of the signal seemed to be driven by duplications (nominal p = 0.014), though the limited sample size and false discovery rate (FDR) of 36 % cautions against over-interpretation. When considering all loci and a more relaxed FDR threshold (0.275) [38], cases were enriched with CNVs for genes involved in nervous system development (human neural function or pathway, nervous system development (GO)) (nominal p = 0.019) and the broader group encompassing genes involved in neurological function (human neural function or pathway, union, inclusive (GO, KEGG, NCI, Reactome)) (nominal p = 0.020). Similar results were generated when removing CNVs at known neuropsychiatric disease loci.