Novel Findings from CNVs Implicate Inhibitory and Excitatory Signaling Complexes in Schizophrenia

1Schizophrenia is a highly heritable disorder, the genetic architecture of which includes a large number of alleles spanning the full spectrum of frequencies. It has been estimated that the additive effects of common variation, as indexed by alleles represented on the platforms used in genome-wide association studies (GWASs), contribute around a quarter to a third of the total population variance in schizophrenia liability. However, the 108 genome-wide-associated common variant loci reported in the largest GWAS study to date only explain a small fraction of this contribution. An increased burden of rare mutations has also been documented in schizophrenia, taking the form of both large CNVs and single-nucleotide variants (SNVs), which often occur as de novo mutations. While several CNVs have been implicated in the disorder, no individual SNV has yet been robustly associated. The CNVs (n = 11) strongly associated with schizophrenia in the largest systematic survey to date are in general large in both size (> 500 kb) and effect (ORs 2–60), the latter being in stark contrast with the small effects conferred by common alleles (typical OR < 1.1). Approximately 2.5% of patients and 0.9% of unaffected controls carry a CNV that is strongly supported as a risk factor for schizophrenia. The pathogenic effects of these CNVs are not confined to schizophrenia; many increase risk for other disorders with a putative major neurodevelopmental component such as intellectual disability, autism spectrum disorder, and attention deficit hyperactivity disorder.

2A small number of single-gene CNVs have been associated with schizophrenia, but the only ones to be definitively implicated are deletions of NRXN1, which encodes the presynaptic cell adhesion protein neurexin 1. All other robustly associated CNVs span multiple genes making it difficult to infer the biological mechanism(s) through which they contribute to disease. Nevertheless, early pathway or gene set analyses of schizophrenia case-control CNV datasets indicated that case CNVs were enriched for synaptic and neurodevelopmental genes.

3It has been noted that these initial approaches to pathway analysis did not completely control for confounds such as CNV size. However, a study of parent-proband trios in which these factors were taken into account found that de novo CNVs in people with schizophrenia were enriched for synaptic proteins. Moreover, this was largely the result of enrichment for genes encoding members of N-methyl-D-aspartate receptor (NMDAR) and neuronal activity-regulated cytoskeleton-associated (ARC) protein complexes, both of which are known to be important for synaptic plasticity and cognitive function in rodents. When these same sets were additionally examined in large case-control datasets, case CNVs were found to contain an excess of genes from NMDAR, but not ARC, complexes. Exome sequencing studies have subsequently supported a role for both NMDAR and ARC complexes in disease. The same exome sequencing studies also found evidence of enrichment for rare disruptive and de novo point mutations among targets of fragile X mental retardation protein (FMRP), a finding that has also been reported for CNVs in a large schizophrenia case-control study.

4Here we present a detailed functional analysis of the largest schizophrenia CNV dataset for which full autosomal CNV data have been examined to date. The study is based on 11,355 cases and 16,416 controls from three separate studies: the International Schizophrenia Consortium (ISC), the Molecular Genetics of Schizophrenia (MGS), and a UK study of individuals diagnosed with schizophrenia and taking the anti-psychotic clozapine (CLOZUK). The ISC and MGS datasets were utilized in to investigate CNV enrichment for ARC and NMDAR gene sets, while no CNV gene set analyses have yet been performed in CLOZUK. Starting from the hypothesis that schizophrenia reflects perturbation of brain function and development, our primary analysis focuses on a circumscribed set of annotations that are related to CNS function and development and are based on proteomic, RNA sequencing, and functional genetic data. In order to evaluate to what extent the pathogenic effects of CNVs primarily reflect disruption of brain function, as a secondary analysis we searched more widely for additional gene set enrichments using a more comprehensive range of annotations available from large, freely accessible databases.

5We identified 134 gene sets relevant to various aspects of nervous system function and development, covering subcellular neuronal function, cellular physiology, cell morphology, brain region and fiber tract morphology, behavior, and brain development (Table S1). Gene sets were derived from functional studies of single genes recorded in the MGI Mammalian Phenotype (MP) database with the exception of subcellular neuronal terms, which comprised a mixture of CNS-related gene sets taken from previous studies of schizophrenia as well as sets that were curated from the proteomic literature (see Table S1 for full list of references). To constrain multiple testing, we utilized a subset of the terms available from MGI (Table S1), which represented CNS annotations postulated to be of most likely relevance to schizophrenia, while at the same time retaining broad functional coverage to allow for the emergence of novel pathophysiological clues. Directional terms such as “decreased” or “enhanced” were avoided in favor of broader categories denoted by “abnormal,” “impaired,” etc.

6Consistent with recent approaches, our analyses are based on large, rare CNVs (> 100 kb, frequency < 1%), as these are both the most robustly called and most enriched in people with schizophrenia. Gene set enrichment analysis was performed using a logistic regression model with covariates included to control for the size and total number of genes overlapping each CNV and for the source of the data (study and genotyping array used). As we were only interested in gene sets that were enriched for CNVs in cases, we used one-tailed tests.

7Of the 134 CNS-related gene sets, we first evaluated those for which there existed prior, replicated evidence of enrichment for rare mutations in schizophrenia in at least three independent studies: the NMDAR protein network, ARC protein complex, and mRNA targets of FMRP. Here, as in all subsequent analyses, we first tested for enrichment in the combined set of CNVs and then deletions and duplications separately. The results of these analyses were Bonferroni corrected for the nine tests performed. NMDAR network genes were highly enriched in CNVs overall (Pcorrected = 3.82 × 10−8), the signal primarily coming from duplications (Pcorrected = 2.26 × 10−8). The ARC gene set was not enriched for CNVs overall, but was enriched in the secondary test of deletions (Pcorrected = 0.0031), while FMRP targets displayed a modest trend toward enrichment in deletions (Pcorrected = 0.076).

8Both ISC and MGS samples were utilized in our previous study to investigate CNV enrichment for ARC and NMDAR. We therefore asked whether we would have found the above NMDAR and ARC association signals if we had performed our analysis in CLOZUK only. Restricting to CLOZUK samples, case CNVs were still enriched for both NMDAR network (Pcorrected = 1.5 × 10−6 combined, Pcorrected = 7.8 × 10−7 duplications) and ARC (Pcorrected = 0.0077 deletions) gene sets. Thus, our analysis provides fully independent evidence for the NMDAR network and ARC complexes.

9We next investigated CNV enrichment for the full list of 134 CNS-related gene sets. To evaluate whether there was evidence for a general enrichment of CNS gene sets in our data, we tested whether the numbers of sets surpassing defined enrichment p value thresholds (Puncorrected < 0.01, 0.001) were greater than expected. To do this, gene set enrichments for the full set of 134 CNS-related terms were compared to those generated from permuted data in which CNVs were randomly re-assigned among individuals, with the constraint that assignments were restricted to individuals from the same study so we could continue to allow for chip and study effects. In our analysis of deletions and duplications combined, more sets were enriched for CNVs in schizophrenia than expected under the null at all enrichment p value thresholds. This was also true when deletions and duplications were considered separately (Table 1).

10To evaluate the significance of the tests of individual gene sets in a manner that allows easy comparison with our later analyses of the much larger annotation datasets (where permutation tests were computationally prohibitive), we adjusted gene set p values for multiple testing using Bonferroni correction for the 402 CNS gene set tests (134 sets × 3 analyses) performed (Table S1). Recognizing that this is over-conservative due to annotation overlap, in Table 2 we additionally list all gene sets with an uncorrected p < 0.001 under the combined test of all CNVs. As can be seen from Table 1, given the large excess in the observed number of associated sets compared with expectation (minimum Nobs/Nexp = 50), most gene sets surpassing this threshold are likely to be true positives, even if they do not survive correction for multiple testing here. Functional processes captured by the six terms with a Bonferroni-corrected p value < 0.05 centered upon behavioral and physiological correlates of learning and related neuronal complexes.

11For the combined analysis of all CNVs, after the NMDAR complex (Pcorrected = 1.71 × 10−6) the next most highly associated term was the GABAA receptor complex (Pcorrected = 0.0012). Conditional analyses revealed these two signals to be essentially independent (see conditional analysis of GO and MGI below; see also Supplemental Experimental Procedures). Thus not only do we confirm, as noted above, the involvement of proteins involved in plasticity of the major excitatory system of the CNS, we also provide the first strong genetic evidence for an etiological role in the disorder for proteins affiliated with the major inhibitory system in the CNS, namely GABAA receptor complexes.

12Tables 3 and 4 list all gene sets with Puncorrected < 0.001 for enrichment within deletions and duplications, respectively. After Bonferroni correction for the 402 CNS gene set tests, there were 14 terms with p value < 0.05 for enrichment in case deletions and 7 terms for case duplications. Enrichment for duplications was largely confined to behavioral and subcellular neuronal gene sets; terms associated via deletions extended over behavior, cellular physiology, subcellular complexes, and development. Deletions were most highly enriched for components of PSD-95 protein complexes. PSD-95, a major postsynaptic scaffolding protein at glutamatergic synapses, interacts with a wide range of channels and receptors including NMDARs. It is notable that although the NMDAR and PSD-95 complexes are functionally related and have overlapping membership, the observations of strong (ORs > 3) and highly significant enrichments (both Pcorrected < 10−7) for these sets relate to duplications and deletions, respectively. These findings are therefore based on sets of completely independent CNVs and, as such, provide extremely robust support for an etiological role for the disruption of glutamatergic signaling in schizophrenia.

13Current data provide strong support for 11 CNV loci in schizophrenia: 6 deletions and 5 duplications (Table S2). Removing CNVs overlapping these known loci, we re-calculated CNS gene set enrichment. Deletion, duplication, and combined analyses all retained an excess of associated terms (Table 1); of the 14 gene sets enriched for deletions, 5 remained nominally associated (Puncorrected < 0.05) when known loci were removed, as did 5 of the 7 terms enriched for duplications (Table S3).

14To identify genes contributing most to gene set enrichment we calculated single gene association p values. This was done in the same manner as our gene set enrichment analyses, but with each “set” restricted to a single gene. For each CNS term with a Bonferroni-corrected p value < 0.05 we then extracted all genes with an uncorrected single gene p value < 0.05 (Tables S4, S5, and S6). To obtain significance at the level of an individual gene, there must be multiple observations of CNVs at the same region. It is therefore unsurprising that established recurrent CNV risk loci account for many such findings. Moreover, as recurrent CNVs are large, these frequently overlap multiple genes and contribute to multiple sets (e.g., del22q11; Table S5). It should be noted that some CNVs also hit multiple genes within a single set, but as each CNV only contributes once to the regression model (see Experimental Procedures), co-localization of set members does not inflate the significance of the set-based enrichment. A number of nominally associated genes lying outside established loci are well known to be important for neuronal signaling. These include the glutamate transporter SLC1A1, a recently reported candidate CNV locus for schizophrenia; GABAergic (GABRD also reported in ) and nicotinic receptors (CHRNA4); synaptic scaffolding proteins DLG2, DLGAP1, and SHANK2; and key elements of the presynaptic vesicle release machinery PCLO and NSF.

15Some of the nominally associated genes have been linked to Mendelian disorders with neurological symptoms, including Walker-Warburg syndrome, a congenital muscular dystrophy with brain and eye abnormalities (ISPD OMIM: 614643, POMK OMIM: 615249); nocturnal frontal lobe epilepsy type 1 (CHRNA4 OMIM: 600513); generalized epilepsy (GABRD OMIM: 613060); spastic paraplegia 51, an autosomal recessive developmental disorder with severe intellectual disability (AP4E1 OMIM: 613744); and Batten disease, an autosomal recessive neurodegenerative condition (CLN3 OMIM: 204200).

16We next determined whether any other gene sets had evidence for enrichment in case CNVs that was independent of the association signal captured by our primary CNS-related terms. For this we drew upon both the MGI MP database, from which we had derived most of our CNS-related gene sets, and the widely used Gene Ontology (GO) terms. As the MP database contains an extensive range of physiological, behavioral, and morphological phenotypes, but little of the low-level molecular function annotation present in GO, these two classifications are to an extent non-redundant and complementary.

17We first identified a “minimal set” of terms capturing most of the enrichment signal arising from CNS-related gene sets. Taking the CNS terms surviving Bonferroni correction, we added the most significant term as a covariate to the regression model and recalculated gene set enrichment for each of the remaining terms. The term with the most significant residual enrichment was then added to the model, and the process repeated until there was no residual association (Puncorrected < 0.05) in the remaining CNS annotations. Three terms were required to capture CNS gene set enrichment in the combined analysis of duplications and deletions: NMDAR network, GABAA receptor complex, and abnormal behavior (see Supplemental Experimental Procedures). This indicates that there are independent enrichment signals in the NMDAR network and GABAA receptor complex gene sets, as reported above. Conditioning on the above three CNS terms, no other GO or MP term survived Bonferroni correction (Tables S7 and S8).

18CNS gene set enrichment for deletions was captured by three terms: PSD-95 complex, abnormal fear/anxiety-related behavior, and abnormal neural plate morphology (see Supplemental Experimental Procedures). Conditioning on these terms, no MP or GO term survived Bonferroni correction (Tables S7 and S8). Three terms captured CNS enrichment in duplications: associative learning, NMDAR network, and GABAA receptor complexes (see Supplemental Experimental Procedures). Once again there was no evidence of additional gene set enrichment in either MP or GO annotations following Bonferroni correction (Tables S7 and S8).

19Case CNVs > 100 kb were both larger (Pdel = 6.8 × 10−14, Pdup = 0.37) and overlap (“hit”) a greater number of genes (Pdel = 1.3 × 10−16, Pdup = 3.4 × 10−5) than those found in controls. However, even after conditioning on CNV size, the number of genes hit was strongly and independently associated (Pdel = 1.0 × 10−5, Pdup = 2.9 × 10−5), whereas after conditioning on number of genes, the effect of size was much weaker and was restricted to deletions (Pdel = 0.0073, Pdup = 0.49). The number of genes hit is therefore a better predictor of case-control status than CNV size.

20We next investigated whether the relationship between number of genes hit and case-control status could be entirely attributed to genes within the disease-associated CNS annotations. To test this, we combined all the CNS annotations that had a Bonferroni-corrected p value < 0.05 to create a single associated CNS set (CNSSZ). We did this separately for deletions and duplications. The number of CNSSZ genes hit by a CNV was a highly significant predictor of case-control status for both deletions and duplications (Pdel = 1.1 × 10−21, Pdup = 1.7 × 10−12). Each was at least five orders of magnitude more significant than the corresponding analyses based on total number of genes hit (see previous paragraph). As expected this CNSSZ term remained highly significant when conditioned on total number of genes hit in any category (Pdel = 7.7 × 10−7, Pdup = 1.1 × 10−9), but the converse was not the case; conditioning on CNSSZ there was little evidence of any remaining effect of the total number of genes hit by each CNV (Pdel = 0.21, Pdup = 0.053).

21Finally, we investigated whether associated CNS gene sets were also enriched in de novo non-synonymous (NS) mutations. To constrain both the number and size of the gene sets tested, we collapsed the “minimal set” of terms capturing most of the CNS enrichment signal (see analysis of GO and MGI above) into a single gene set for each of our analyses (combined, deletion only, duplication only). Gene sets capturing the CNS enrichments for deletions and duplications were associated with de novo NS mutations observed in individuals with schizophrenia. This was not entirely due to ARC and NMDAR network genes (which we have previously described to be enriched for de novo NS mutations; ) (Table 5). No enrichment was found when the analysis was repeated in a corresponding set of mutations from unaffected individuals (Table 5). When all 21 CNS terms with Pcorrected < 0.05 (Tables 2–4) were tested individually, over half were nominally enriched for de novo NS mutations (cf. none for mutations from controls). While none survive correction for multiple testing (Table S9), this has to be interpreted in the context of the very weak enrichment in schizophrenia for de novo NS mutations, and therefore low power to robustly detect gene set enrichment. These findings independently support the broader relevance of the gene sets we have identified in the present study, but larger studies of de novo mutations will be required for finer-scale dissection.

22We have performed a detailed, functionally informed analysis of large, rare CNVs from 11,355 schizophrenia cases and 16,416 controls. The results provide strong, novel evidence implicating disruption of inhibitory GABAergic modulation of neuronal signaling in schizophrenia and robustly confirm, and extend, the genetic evidence implicating disruption of excitatory glutamatergic signaling (Figure 1). It is clear, however, that these neuronal complexes do not entirely account for the enrichment of CNVs in cases given the independent enrichments seen in behavioral and neurodevelopmental gene sets. This suggests that subcellular processes beyond those currently ascribed to GABAergic and glutamatergic complexes remain to be identified.

23We found no evidence that the pathogenic effects of CNVs reflect biological processes other than those directly relevant to brain function. This conclusion follows from the absence of additional gene set enrichments after conditioning on the CNS sets, and is further supported by the observation that the number of genes hit by a CNV in the disease-associated CNS pathways was a better predictor of whether a CNV occurred in a case or a control than total number of genes hit. This contrasts with recent findings based on common polymorphisms, where independent enrichments were found in enhancer elements that were active in both CNS and immune tissues, although a more recent analysis by that group suggests that most, if not all, of the signal is captured by the CNS enhancers (http://biorxiv.org/content/early/2015/01/23/014241).

24While 11 CNV loci have been strongly associated with schizophrenia to date, our results indicate that more associated CNVs remain to be identified. The association between number of genes hit and pathogenicity suggests that, in many instances, looking for a single gene explanation for CNV pathogenicity may not be fruitful. Instead it indicates that pathogenicity depends upon the total burden of relevant genes hit by a CNV, or that different single genes are implicated in different individuals depending upon their genetic and environmental context—the larger the CNV, the greater the probability that a critical pathway or process will be sufficiently impaired. It should be noted that the presence of multiple hits in the same CNV does not artificially inflate the significance of our enrichment tests, as each CNV only contributes once to the analysis.

25There were strong, independent associations in postsynaptic complexes derived from glutamatergic synapses: NMDAR complex genes were enriched in case duplications, while PSD-95 and to a lesser extent ARC complexes were enriched in deletions. When these findings are combined with existing evidence from de novo CNVs, from rare SNVs and indels, and more recently from GWASs and common alleles, the relevance of altered glutamatergic signaling to schizophrenia etiology and pathophysiology seems to be beyond any reasonable doubt.

26While models of schizophrenia based upon NMDAR hypofunction have a long history, the genetic data now indicate that the glutamatergic contribution to schizophrenia encompasses a much wider range of cellular processes converging upon synaptic information processing and plasticity (Figure 1). This is clearly inconsistent with hypotheses in which deficits in glutamatergic signaling primarily reflect disruption via neuromodulatory pathways. Genetic evidence for disruption of neuromodulators is so far restricted to dopamine, with a genome-wide significant GWAS signal localized to DRD2. Both serotonergic 5-HT2C and nicotinic α7 receptor complexes were tested here, and neither was found to be strongly associated (Table S1).

27We also find novel, independent evidence for disruption of GABAA receptor complexes in schizophrenia. Deficits in GABAergic signaling have long been hypothesized to contribute to schizophrenia pathophysiology alongside perturbation of dopaminergic and glutamatergic systems. Evidence supporting a direct involvement of GABA has as yet not been compelling, being drawn from imaging studies and animal models of putative intermediate phenotypes, or post-mortem expression studies in small samples where reverse causality or confounding cannot be excluded. Here we find case CNVs to be enriched for components of GABAA receptor complexes (Tables 2–4), with conditional analyses revealing the GABAA association signal to be independent of that seen for NMDAR complex genes. Our results indicate that abnormalities in GABAergic signaling play a direct pathogenic role in schizophrenia and cannot be entirely attributed to secondary effects of NMDAR dysfunction.

28GABAA receptor complex enrichment was strongest among duplications, where the most highly associated genes were α5, β3, and δ receptor subunits (Table S10). The genes encoding α5 and β3 subunits are found within the Angelman/Prader-Willi locus, while the δ subunit has been mapped to the critical region for the 1p36 deletion syndrome: a relatively common CNV associated with a range of neurodevelopmental outcomes that has recently been identified as a candidate locus for schizophrenia. The remaining autosomal GABAA receptor genes largely cluster within two loci on chromosomes 4 and 5, neither of which displayed evidence of enrichment (Table S10). The GABAA enrichment signal for deletions was driven by NRXN1, which encodes for the presynaptic cell adhesion protein neurexin 1, common to both GABAergic and glutamatergic synapses.

29Multiple GABAA receptor subtypes exist, each with a unique set of functional properties and a distinct spatiotemporal expression profile (reviewed by ). In contrast to the β3 subunit, which is common to many receptor subtypes, α5 and δ subunits occur in distinct, mainly extrasynaptic populations of receptors responsible for tonic inhibition. This indicates that the contribution of GABAergic signaling to schizophrenia may not be primarily synaptic, although the presence of CNVs in NRXN1 and GHPN, encoding neurexin 1 and the synaptic GABA receptor scaffolding protein gephyrin (Table S10), suggests that perturbation of synaptic GABAergic signaling may also play a role.

30Tonic inhibition in the hippocampus, a process to which both α5- and δ-containing GABAA receptors contribute, alters the induction of long-term potentiation. Moreover, prolonged activation of NMDARs has been shown to reduce cell surface expression of δ-containing receptors. Thus, there are potential functional connections between our findings of enrichment for CNVs in GABAergic and postsynaptic glutamatergic complexes in schizophrenia. At the behavioral level, perturbation of NMDAR and tonic GABA signaling both lead to alterations in associative learning. This hierarchy of functionally related processes, from subcellular complexes to behavioral learning via cellular signaling and plasticity, encapsulates the elements of gene set enrichment common to case duplications and deletions (Tables 2–4).

31Our findings are consistent with a considerable body of non-genetic literature; it is nearly 40 years since the cognitive deficits seen in schizophrenia were first proposed to reflect dysfunctional associative learning, with later hypotheses suggesting perturbed synaptic plasticity as the source of this dysfunction. Our identification of independent genetic associations in glutamatergic and GABAergic complexes is particularly relevant to proposals that alteration in the ratio of excitatory to inhibitory transmission (E/I balance) underlies the behavioral deficits seen in schizophrenia (reviewed by; see also ). Discussions typically focus upon the oscillatory properties of neuronal networks, fundamental to efficient information transfer and the coordination of neuronal assemblies. Deficits in gamma rhythms have been reported in schizophrenia (reviewed by ), and while multiple mechanisms contribute to the generation of these rhythms, deficits have primarily been hypothesized to result from the altered firing of GABAergic interneurons. Interestingly, the frequency of hippocampal gamma oscillations is sensitive to the balance between NMDAR-dependent excitation and GABA δ subunit-dependent tonic inhibition of interneurons, linking these hypotheses directly to our findings.

32Although less discussed in relation to schizophrenia, E/I balance also plays a role in the development and maintenance of stable perceptual and motor representations (reviewed by ). During early post-natal development unbalanced excitatory input drives activity-dependent plasticity, shaping emerging networks of synaptic connections in response to the environment. As networks mature and inhibitory elements are progressively integrated, E/I inputs become correlated making internal representations resistant to further modification. In the adult brain, activation of neuromodulatory systems (including acetylcholine, dopamine, and serotonin) can alter E/I balance, allowing sensory circuits to again become sensitive to environmental input. Unlike the experience-dependent modifications that occur during development, plasticity in adult sensory cortex is generally transient (see ). Genetic associations potentially link the main elements of these processes to schizophrenia: E/I signaling and synaptic plasticity through our analyses, dopaminergic signaling via the DRD2 GWAS locus noted above. Disruption of such processes may potentially play a role in the developmental trajectory of schizophrenia, or in the manifestation of transient perceptual alterations during psychotic episodes.

33In conclusion, our analyses support and extend previous studies indicating a contribution to schizophrenia from complexes central to the induction (NMDAR) and maintenance (ARC) of synaptic plasticity and provide strong novel evidence for the involvement of inhibitory modulation (GABA) of synaptic signaling (Figure 1). Perturbation of these processes is likely to have a widespread impact on brain function, and only a subset of genetic lesions within these systems may be compatible with a schizophrenia phenotype. The identification of the mechanisms by which disruption of these processes by genetic mutation leads to psychopathology will doubtless require experimental studies in model systems of high construct validity. The strength of genetic evidence converging on a plausible and coherent set of biological processes provides firm foundations upon which such studies can now proceed.

34Case and control CNVs were derived from three samples: CLOZUK, the ISC, and the MGS. A full description of these samples, the arrays they were genotyped on, and CNV calling procedures can be found in the Supplemental Experimental Procedures or in the original publications. Briefly, CLOZUK samples were genotyped on several Illumina arrays. In order to limit any bias in detecting specific CNVs between the CLOZUK cases and controls, done on different arrays, for CNV calling we used only the 520,766 probes common to all these arrays. In the CLOZUK sample CNVs were called with PennCNV. All MGS samples were genotyped on Affymetrix 6.0 arrays, and approximately equal proportions of ISC cases and controls were genotyped on either Affymetrix 6.0 or Affymetrix 5.0 arrays (see Supplemental Experimental Procedures). In the MGS and ISC samples CNVs were detected using Birdsuite. Only samples with a European ancestry were retained for analysis. Rigorous quality control was performed to remove low-quality samples (full details presented in the Supplemental Experimental Procedures), resulting in 5,745 cases and 10,675 controls in CLOZUK, 2,214 cases and 2,556 controls in MGS, and 3,395 cases and 3,185 controls in ISC retained for analysis. Taking CNV calls from samples which passed quality control in each study, CNVs were joined if the distance separating them was less than 50% of their combined length. CNVs were excluded if they overlapped low copy repeats by more than 50% of their length, or had a probe density < 1 probe/20 kb. CNVs with a frequency > 1% or identified as false positives by an in silico median Z score outlier method were also removed. Z score validation was not performed for the ISC study as we did not have access to the raw intensity data. Following QC, genes overlapping CNVs were identified using genomic locations for the appropriate build of the human genome: Build 35 of the human genome for ISC, Build 36 for MGS, and Build 37 for CLOZUK. Studies were then collated, and CNVs < 100 kb in size and/or covered by < 15 probes were removed prior to analysis. Differences in genotyping chip, CNV calling, and genome build between studies are controlled for in our enrichment analyses through the “chip” and “study” covariates; while between-study differences may reduce power to identify true positives, they do not increase the rate of false positives. See Supplemental Experimental Procedures for further details.

35Proteomic studies used to derive subcellular terms are listed in Table S1. For terms analyzed in, the processed gene sets analyzed in that study were re-used here. Gene sets for all other subcellular terms were extracted from the relevant studies and mapped to human coding genes. GO annotations were taken from NCBI gene2go (ftp://ftp.ncbi.nih.gov/DATA), using Homo sapiens annotations only. MP ontology and gene annotations were downloaded from the Mouse Genome Informatics (MGI) online resource (http://www.informatics.jax.org). Genes were mapped to human using the file HOM_MouseHumanSequence.rpt, also downloaded from MGI. For further details see Supplemental Experimental Procedures.

36For each gene set, the numbers of genes “hit” by case and control CNVs were compared; a gene was counted as being hit by a CNV if the CNV overlapped any part of its length. To overcome biases related to gene and CNV size, and to control for differences between studies and genotyping chips, the following logistic regression models were fitted to the combined set of CNVs:(a) logit (pr(case)) = study + chip + CNV size + total number of genes hit(b) logit (pr(case)) = study + chip + CNV size + total number of genes hit + number of genes hit in gene set

37(a) logit (pr(case)) = study + chip + CNV size + total number of genes hit

38(b) logit (pr(case)) = study + chip + CNV size + total number of genes hit + number of genes hit in gene set

39Comparing the change in deviance between models (a) and (b), a one-sided test for an excess of genes in the gene set being hit by case CNVs was performed. For further details and a full description of the approach taken for multiple testing correction (outlined in the main text), see Supplemental Experimental Procedures.

40Case-control status was permuted 1,000 times, status being shuffled between CNVs from the same study and genotyping chip (“Affymetrix 5.0,” “Affymetrix 6.0,” or “Illumina”). Enrichment analyses were performed in each permuted dataset; the proportion of datasets in which the number of terms with p < Pthr equaled or exceeded that of the true data being used as the empirical p value for an excess of associated terms at the threshold Pthr.

41From CNS gene sets with Pcorrected < 0.05, a subset was identified that captured the association signal in all other terms. GO and MGI terms were then analyzed using the enrichment test outlined above, but with this “minimal” set of terms added as covariates to the regression models (see main text and Supplemental Experimental Procedures).

42To investigate whether gene set enrichment was solely driven by CNVs at loci well supported by current data (Table S2), we removed all CNVs overlapping these loci and re-ran the enrichment analysis as above. See Supplemental Experimental Procedures for further details.

43This was performed in an identical manner to gene set enrichment analysis, but with each “gene set” here comprising a single gene. The term “number of genes hit in gene set” in model (b) thus becomes a binary variable.

44CNV size and number of genes hit (either total or CNSSZ) were regressed against CNV case-control status under a logistic regression model. Covariates were included for study and genotyping chip (as in enrichment test, above). See Supplemental Experimental Procedures for further details.

45NS de novo variants found in individuals with schizophrenia were taken from, consisting of variants identified in four separate studies. These were analyzed for gene set enrichment using the dnenrich software (http://bitbucket.org/statgen/dnenrich). NS de novo variants found in unaffected individuals were also taken from and analyzed in an identical fashion. These consisted of healthy controls and unaffected siblings collated from six separate studies.

46A.J.P., M.C.O., and M.J.O. led the study and interpreted the findings. A.J.P. collated gene sets with the assistance of J.H., D.H.K., and P.H. A.J.P. designed and implemented the analytic approach, with P.H. providing guidance on correction for multiple testing. J.T.R.W. led recruitment of the CLOZUK sample. E.R. and G.K. were responsible for all CNV calling in all samples. K.D.C., J.L.M., and S.A.M. coordinated the genotyping of the CLOZUK sample. A.J.P., assisted by M.C.O. and M.J.O., wrote the first draft of the manuscript, which was then commented on by other authors.