Assessment of Potential Clinical Role for Exome Sequencing in Schizophrenia

Abstract

Background

There is increasing evidence that certain genetic variants increase the risk of schizophrenia and other neurodevelopmental disorders. Exome sequencing has been shown to have a high diagnostic yield for developmental disability and testing for copy number variants has been advocated for schizophrenia. The diagnostic yield for exome sequencing in schizophrenia is unknown.

Method

A sample of 591 exome-sequenced schizophrenia cases and their parents were screened for disruptive and damaging variants in autosomal genes listed in the Genomics England panels for intellectual disability and other neurological disorders.

Results

Previously reported disruptive de novo variants were noted in SETD1A, POGZ, SCN2A, and ZMYND11. Although the loss of function of ZMYND11 is a recognized cause of intellectual disability, it has not previously been noted as a risk factor for schizophrenia. A damaging de novo variant of uncertain significance was noted in NRXN1. A previously reported homozygous damaging variant in BLM is predicted to cause Bloom syndrome in 1 case and 1 case was homozygous for a damaging variant in MCPH1, a result of uncertain significance. There were more than 400 disruptive and damaging variants in the target genes in cases but similar numbers were seen among untransmitted parental alleles and none appeared to be clinically significant.

Conclusions

The diagnostic yield from exome sequencing in schizophrenia is low. Disruptive and damaging variants seen in known neuropsychiatric genes should not be automatically assumed to have an etiological role if observed in a patient with schizophrenia.

Introduction

It is now well established that deletions and duplications at specific chromosomal locations, termed copy number variants (CNVs), can have a substantial effect on the risk of developing schizophrenia and are found in 3% of people with schizophrenia who have normal IQ and 7% of those with borderline IQ.^1,2 Because of their clinical implications, it has been advocated that testing for these pathogenic CNVs should be a routine investigation for patients diagnosed with schizophrenia.³ Identifying a CNV with a substantial effect on risk could help by providing the patient with an explanation of why they have become unwell and might improve compliance with treatment, as well as having implications for relatives.⁴ As well as CNVs, polygenic risk factors and non-genetic risk factors, there is now good evidence that rare coding variants contribute to schizophrenia risk.⁵ There is strong statistical evidence that schizophrenia cases have an excess burden of rare variants predicted to damage the function of particular sets of genes, eg, those related to synaptic functioning.^6,7 However, only a small number of individual genes are strongly implicated. Currently, we would argue that there is good evidence to support the claim that variants causing loss of function (LOF) of SETD1A, RBM12, and NRXN1 can substantially increase the risk of schizophrenia.^8–10 However, these are extremely rare and are collectively found in fewer than 1% of cases.

It has become apparent that the relationship between genotype and phenotype is often more complex than had originally been supposed and that variants which had been thought to have a simple Mendelian effect might manifest reduced penetrance.¹¹ In the context of psychiatric disorders, an important general theme which emerges from the findings to date is that a variant which confers increased risk of schizophrenia may also confer increased risk of intellectual disability (ID) or other neurodevelopmental disorder, suggesting a form of pleiotropy in that processes which disrupt neurodevelopment may lead to ID, schizophrenia, both, or neither. This applies to both CNVs and, eg, LOF variants of SETD1A, which are found both in schizophrenia cases and in subjects diagnosed with the developmental disorder.⁸ In general, we find that rare, damaging coding variants in schizophrenia cases affect genes implicated in other neurodevelopmental disorders.¹² Another example is a recently reported case with a de novo variant in SCN2A who initially presented with infantile-onset seizures, autistic features, and episodic ataxia but then developed psychotic symptoms as an adult.¹³

In contrast to schizophrenia, several hundred genes are implicated as risk factors for ID, in the sense that it is known that specific variants in these genes can produce a phenotype, which includes ID. Given the pleiotropy described earlier, if one detected in a patient with schizophrenia a variant known to cause ID then one might reasonably conclude that in this case, the variant was likely making an etiological contribution to the schizophrenia. Testing for variants in known ID genes, rather than only the handful of currently known schizophrenia genes, might produce a worthwhile diagnostic yield and this approach has been implemented in a pilot study incorporated in the 100 000 Genomes Project, a diagnostic service implemented by the UK National Health Service. Here, patients with schizophrenia accompanied by other features suggestive of a genetic etiology undergo whole-genome sequencing and are then checked for variants in genes known to be causative of ID and other neurological conditions. In children with severe developmental disorders, exome sequencing can produce an overall diagnostic yield of about 40%.¹⁴ We thought it would be of interest to gain a better understanding of what the likely yield might be for patients with schizophrenia by applying the proposed diagnostic testing protocol to a sample of research subjects for whom exome sequencing data were already available.

The Bulgarian sample consists mainly of a set of schizophrenia cases along with their parents who have undergone exome sequencing and who have been studied for the presence of de novo mutations and recessively acting variants.^15,16 These previous reports were gene discovery projects aiming to build evidence to identify novel genes conferring susceptibility to schizophrenia, and indeed provided some of the evidence to implicate SETD1A. We set out to use this sample as if it were a clinical sample with the aim of identifying any variants, acting dominantly or recessively, which might in a clinical context be reported as pathogenic or likely pathogenic.

Methods and Materials

This sample comprised exome sequence data from 591 Bulgarian trios, consisting of probands with schizophrenia and their parents, 5 of whom were also affected.^15,16 The authors of the original studies reported that all probands and parents provided written informed consent and that ethical committee approval was obtained from the hospitals from where the cases were recruited. The short read files were downloaded from dbGaP along with family structure and phenotype information (https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000687.v1.p1). The short read files were converted to fastq files using the fastq-dump utility of the dbGaP SRA toolkit. Reads were then aligned to the hg19 human reference sequence (build GRCh37) using Novoalign V3.02.08 (NovoCraft Technologies), duplicate reads were marked using SAMBLASTER,¹⁷ and the BAM files were sorted using Novosort V1.03.09 (NovoCraft Technologies). Genotypes were called according to GATK best practices (Broad Institute). The HaplotypeCaller module of GATK V3.6 was used to produce gVCF files and these were then combined using the CombineGVCFs module. Initial calls were made using the GenotypeGVCFs module and then single nucleotide variants (SNPs) were filtered based on accuracy estimates produced by VariantRecalibrator and indels were filtered using the VariantFiltration module and the filtering expression “QD < 2.0 || FS > 50.0 || ReadPosRankSum < -20.0” to generate a PASS entry in the information field of the resultant VCF file. Variants were excluded if they did not have a PASS in the information field and individual genotype calls were excluded if they had a genotype quality score less than 20. Heterozygote calls were excluded if 1 allele accounted for less than 0.2 of total reads. Variants were also excluded if more than 10% of cases or of pseudocontrols failed this quality threshold or if the heterozygote count was smaller than both homozygote counts in both cohorts. Variants were annotated with using variant effect predictor (VEP), PolyPhen, and SIFT.^18–20

To identify sets of genes for which there was strong evidence of involvement in neurodevelopmental disorders, such that they might be used clinically, we downloaded panels generated for the 100 000 Genomes Project from the Genomics England Panel App (https://panelapp.genomicsengland.co.uk/panels/). We downloaded version 1.13 of these panels: Familial Focal Epilepsies, Familial Genetic Generalised Epilepsies, Genetic Epilepsy Syndromes, and ID. Attention was restricted to well-supported genes (color-coded Green), which were implicated in their own right rather than being listed only through lying within a pathogenic deletion or duplication. To handle transmissions from parents in a consistent manner, attention was restricted to autosomal genes. Genes reported to have a monoallelic effect, with the addition of SETD1A, were incorporated in a list of 328 putative dominant genes. A list of 942 putative recessive genes included the dominant genes along with the panel genes reported to have biallelic effects.

The GENEVARASSOC program (https://github.com/davenomiddlenamecurtis/geneVarAssoc) was used to export variants in each gene to SCOREASSOC.^21–23 SCOREASSOC is able to use trio data to construct a sample of cases and pseudocontrols with genotypes consisting of the untransmitted alleles from the parents of each affected proband. It can also detect de novo variants, which are present in offspring but not parents and can detect compound heterozygotes where a gene contains both paternally and maternally inherited variants. Stop gained, frameshift, splice site, and transcript ablation variants were classified as disruptive whereas nonsynonymous variants characterized by SIFT as deleterious or by PolyPhen as probably damaging were classified as damaging. Version hg19 of the reference human genome sequence and RefSeq genes were used to extract variants on a gene-wise basis. The gnomAD database (http://gnomad.broadinstitute.org/) was used to obtain minor allele frequencies (MAFs) for variants in populations with different ancestries.¹¹ For analyses of dominant effects, variants were excluded if they had MAF > 0.005 in any population or if more than 6 cases or pseudocontrols were heterozygous or homozygous for the minor allele. For the analyses of recessive effects, the same exclusions were applied, but using a threshold of MAF > 0.05. We reasoned that any variants with higher allele frequencies would have been discovered in previous genome-wide association studies if they had appreciable effects on risk.

Qualifying damaging or disruptive variants occurring in the dominant genes, or as homozygotes or compound heterozygotes in the recessive genes, were investigated further by consulting OMIM (https://omim.org), ClinVar,²⁴ GeneCards,²⁵ and PubMed (https://www.ncbi.nlm.nih.gov/pubmed) to determine whether there would be sufficient evidence to warrant reporting them in a clinical context. The interpretation of the clinical significance of variants remains a challenging task, which at times is inevitably subjective. When attempting this task, we endeavored to take account of the gene concerned, the related phenotypes, and previous reports of the occurrence of either the variant itself or else of variants with similar predicted effect on protein function.

This study differed from previous analyses of this dataset in that it sought to explore how an application of our current knowledge regarding genes involved in schizophrenia and ID might be used in a clinical context. In addition, although previous studies investigated de novo mutations and recessively acting variants, here we also considered single inherited variants, which might have a dominant effect. For variants of interest, we carried out more intensive consultation of the literature, as would be done in a clinical situation.

Results

The de novo variants have been reported on previously,²⁶ and in panel genes consisted of disruptive variants in POGZ, SCN2A, SETD1A, and ZMYND11 and damaging variants in ARID1B and NRXN1. Each of these occurred in a single subject and all had been confirmed by Sanger sequencing in the original study, which had also reported that both SCN2A and POGZ were established autism genes.²⁶ LOF variants in SETD1A have subsequently been recognized as a well-established risk factor for schizophrenia.^8,27 LOF of ZMYND11 has not previously been reported in schizophrenia cases but has been reported to be associated with ID and sometimes features of behavioral and/or mood disturbance, with 1 case being given diagnoses of “rapid cycling bipolar disorder, borderline personality disorder and pervasive developmental disorder, with psychosis and alcohol and drugs abuse”.²⁸ Taken together, it seems that these 4 disruptive variants would be classified as pathogenic or likely pathogenic. The damaging variant in ARID1B is 6:157495166 T>C, which is absent from gnomAD and which produces a methionine to threonine substitution at position 426 in the product of transcript ENST00000319584. It is predicted to be “deleterious_low_confidence” and “benign”. LOF mutations of ARID1B are a known cause Coffin-Siris syndrome, which comprises ID with other abnormalities but no clear psychiatric phenotype, and isolated nonsynonymous mutations have also been reported in 2 cases.²⁹ Thus, one would probably categorize this variant as being of unknown significance. The damaging variant in NRXN1 is 2:50724807 T>C, which has MAF 0.000007 in gnomAD but is absent from ClinVar. It produces a glutamic acid to glycine substitution at position 848 in the product of transcript ENST00000401669. It lies in a laminin G domain and is predicted to be “deleterious” and “probably damaging”. Given that there is evidence that disruption of NRXN1 increases risk of schizophrenia, one would probably categorize this variant as being of unknown significance but one might regard it as likely pathogenic.

In the analysis of dominant genes excluding de novo variants, disruptive variants were seen on 21 occasions in pseudocontrols and 12 in cases. Thus, of disruptive variants present in the parents of schizophrenia probands, there was no tendency for them to be preferentially transmitted to the proband. The disruptive variants seen in cases are listed in table 1. Three cases have LOF mutations in ABCC9. One parent had an additional frameshift variant in ABCC9, which was not transmitted to a case. Missense mutations of ABCC9 can cause Cantu syndrome, which can include ID as part of the phenotype but no clinical effects have been reported for LOF variants, except for a single case of dilated cardiomyopathy, hence we do not think the observed variants are likely to be clinically significant.³⁰ A single subject has a frameshift variant in ASXL1. De novo stop mutations in transcript NM_015338 of this gene cause Bohring-Opitz syndrome, a severe developmental and malformation disorder characterized by profound mental retardation among other features.³¹ The variant we observe, p.(Thr1217ThrX), is predicted to form a truncated product of this transcript and lies close to one of the previously reported pathogenic variants, p. (Ser1028X). However, the normal product is only 1541 amino acids in length and all the pathogenic known variants lead to a more drastic truncation so if this finding does not represent a genotyping error, it is possible that the small truncation it causes does not have severe consequences. In any event, its clinical significance seems uncertain. Two cases have 1:245019204-GCCT>G, which is a splice site variant for some transcripts of HNRNPU but is upstream or downstream of other transcripts. A variety of types of variant causing LOF of HNRNPU cause developmental delay and seizures³² but this particular variant is classified by ClinVar as “Benign/Likely_benign.” One case has a frameshift variant in KCNT1, which codes for a sodium-activated potassium channel. Although heterozygous gain-of-function mutations in this gene can cause infantile epilepsy,³³ there is no phenotype associated with LOF. One subject had a variant in NF1, which affects the splice site of some transcripts but is upstream or intronic for others. Although mutations of NF1 can cause neurofibromatosis with learning disability,³⁴ this particular variant is absent from ClinVar and its significance is unclear. One subject has a stop gained and frameshift variant in PDE4D. Heterozygous missense variants in this gene can cause acrodysostosis and acroscyphodysplasia with ID but these seem to be through gain-of-function effects and no phenotype is associated with LOF.^35,36 One subject has a frameshift variant in PTCH1. There is a report of a mother and son with holoprosencephaly like phenotype who had a duplication of this gene but there is no evidence for a psychiatric phenotype associated with LOF.³⁷ One subject has a stop gained and frameshift variant in SAMD9. Missense variants in this gene can cause MIRAGE system, a multisystem disorder that can have ID as part of the phenotype.³⁸ These effects seem to be due to gain of function, whereas LOF variants have been reported to be associated with adult myelodysplastic syndrome but no psychiatric phenotype.³⁹ One subject has a frameshift variant in SMARCA2. De novo missense (but not LOF) variants in this gene, which is involved in chromatin remodeling and neural development, cause Nicolaides-Baraitser syndrome, which includes severe mental retardation.⁴⁰ There is a report of SNPs in SMARCA2 being associated with schizophrenia and that psychotogenic drugs reduce its expression whereas antipsychotic drugs increase it.⁴¹ Although the observation of this variant may be potentially of some scientific interest, there is not sufficient evidence to justify making any clinical interpretation of it. To summarize, more disruptive variants are seen in untransmitted parental alleles than occur in cases. Of the 12 which were seen in cases only 1, a truncating variant in ASXL1, might be viewed as possibly pathogenic and even for this variant, the case is very arguable. The others would all be viewed as benign or of uncertain significance.

Table 1.

Rare Disruptive Variants Noted in Schizophrenia Cases Along With Their VEP Annotation, Minor Allele Frequency (MAF) in Non-Finnish Europeans and ClinVar Entry

Gene	Variant	Effect	MAF	ClinVar
ABCC9	12:22028600-G>A	stop_gained	0.000004	Uncertain significance
ABCC9	12:22048250-CTG>C	splice_acceptor_variant	0.00003	Absent
ABCC9	12:22086827-A>AT	frameshift_variant	0.000004	Absent
ASXL1	20:31024165-C>CAA	frameshift_variant	0.000004	Absent
HNRNPU	1:245019204-GCCT>G	splice_donor_variant	0.0006	Benign/Likely_benign
HNRNPU	1:245019204-GCCT>G	splice_donor_variant	0.0006	Benign/Likely_benign
KCNT1	9:138671283-CA>C	frameshift_variant	0.000009	Absent
NF1	17:29645324-A>G	splice_donor_variant	0.00002	Absent
PDE4D	5:58270510-T>TCATCTATGACA	stop_gained	0.00001	Absent
PTCH1	9:98279020-C>CG	frameshift_variant	0.000009	Absent
SAMD9	7:92735007-G>GAACCTTT	stop_gained	0.000009	Absent
SMARCA2	9:2191408-T>TGTAA	frameshift_variant	0.00002	Absent

With respect to damaging variants, these were seen on 429 occasions in pseudocontrols and 467 in cases, a nonsignificant difference. 594 were absent from ClinVar and the rest were judged to be benign, likely benign or of uncertain significance, split approximately equally between pseudocontrols and cases. None was reported by ClinVar to be likely to be pathogenic when occurring as a heterozygote.

In the recessive analyses, 1 case and 1 pseudocontrol had disruptive variants in both copies of a gene. The case was described in the original report and is homozygous for 15:91304245 C>T, which causes a stop mutation in the BLM gene. LOF variants in both copies of this gene cause Bloom syndrome and in addition to schizophrenia this subject was noted to have vitiligo and epilepsy. A pseudocontrol had 2 copies of 1:97915614 C>T, which produces a splice site variant in the DYPD gene. A subject homozygous for this variant would be expected to manifest dihydropyrimidine dehydrogenase deficiency but of course the pseudocontrol does not actually consist of a single individual but rather represents the genotype of the parental alleles not transmitted to the corresponding case.

In addition to these 2 subjects, a further 9 cases and 4 pseudocontrols had a damaging and/or disruptive variant in both copies of a gene. The variants that were seen in cases are listed in table 2. In a clinical context, none could be interpreted as likely to be pathogenic. In addition, certain homozygous genotype calls may be inaccurate because it seems unlikely that one would observe a homozygote for a variant with allele frequency less than 0.001 in an outbred population. Again, in a clinical situation, any potentially pathogenic variants would need to be confirmed by Sanger sequencing. The observation of a subject homozygous for a probably damaging variant in MCPH1, a gene in which recessively acting variants can cause microcephaly, is possibly of some interest because of reports of variants in this gene being associated with schizophrenia and bipolar disorder.^42–44 However, the effects of this particular variant are too uncertain for this finding to have any value in a clinical context.

Table 2.

Homozygote and Compound Heterozygote Genotypes of Damaging and/or Disruptive Variants Observed in Schizophrenia Cases

Gene	Haplotype	Variant	Transcript	Effect	SIFT	PolyPhen	ClinVar	MAF
AGA	Paternal	4:178360811-G>T	ENST00000264595	L105I	Tolerated (0.06)	probably_damaging(0.954)	Benign	0.012
	Maternal	(Homozygous)
FH	Paternal	1:241669398-T>C	ENST00000366560	Y270C	Deleterious (0.03)	benign(0.243)	Uncertain significance	0.0003
	Maternal	(Homozygous)
LRP2	Paternal	2:170042245-T>C	ENST00000263816	N3205D		probably_damaging(0.999)	Absent	0.00002
	Maternal	(Homozygous)
MCPH1	Paternal	8:6371240-C>G	ENST00000325203	L386F	Tolerated (0.15)	probably_damaging(0.972)	Absent	0.003
	Maternal	(Homozygous)
MTRR	Paternal	5:7885945-A>G	ENST00000264668	I372M	Deleterious (0.03)	benign(0.208)	Absent	0.0003
	Maternal	(Homozygous)
PCCB	Paternal	3:136002730-C>T	ENST00000469217	P219S	deleterious_ low_confidence (0.01)	probably_damaging(0.998)	Uncertain significance	0.007
	Maternal	(Homozygous)
RELN	Paternal	7:103234828-G>C	ENST00000428762	I1217M	Deleterious (0)	possibly_damaging(0.643)	Conflicting interpretations of pathogenicity: Likely benign3 or Uncertain significance3	0.004
	Maternal	7:103138569-G>A	ENST00000428762	T2933I	Deleterious (0.01)	possibly_damaging(0.493)	Likely benign	0.0002
SCN3A	Paternal	2:165950891-C>T	ENST00000283254	R1510H	Deleterious (0.01)	probably_damaging(0.977)	Absent	0.000008
	Maternal	2:166003301-G>A	ENST00000283254	S540F	Deleterious (0.04)	probably_damaging(0.986)	Likely benign	0.002
SLX4	Paternal	16:3639713-G>A	ENST00000294008	S1309F	Deleterious (0.01)	probably_damaging(0.962)	Uncertain significance	0.0004
	Maternal	(Homozygous)

For the compound heterozygotes, both the paternal and maternal allele are listed or else the genotype is noted to be homozygous. A transcript is given for which the stated SIFT and PolyPhen annotations were output. Also, shown are the ClinVar entry and minor allele frequency (MAF) in non-Finnish Europeans.

Discussion

Exome sequencing of 591 schizophrenia cases and their parents and examining genes known to cause neuropsychiatric phenotypes identifies 4 cases with a disruptive de novo variant which would be categorized as likely pathogenic, in the genes SETD1A, POGZ, SCN2A, and ZMYND11. Two further cases have damaging de novo variants in ARID1B and NRXN1 which would probably be classified as of unknown significance, although it could be argued that the NRXN1 variant was likely pathogenic. An additional 12 subjects have inherited a disruptive variant in a known susceptibility gene but we argue that in a clinical context all of these would be classified as of unknown significance. Importantly, 21 of such disruptive variants were seen in parents but not transmitted to the schizophrenic proband. Likewise, there were 467 damaging variants that were inherited by cases but another 429 variants that were seen in parents but not transmitted. All of these appeared to be of benign or of unknown significance. With respect to recessively acting variants, we observed the previously reported homozygous variant disruptive of BLM but no other variants, which would be reported as likely pathogenic. The BLM finding is probably not of relevance to the diagnosis of schizophrenia, as psychosis is not a recognized phenotypic component of Bloom syndrome but should be regarded as an incidental finding with implications for management, eg, heightened surveillance for neoplasia.⁴⁵ The only other finding of note is a case homozygous for a damaging variant in MCPH1 but in a clinical context, this would simply be reported as of unknown significance.

In summary, carrying out exome sequencing has yielded a clinically significant incidental finding in 1 case and there are an additional 4, or possibly 5, cases with a variant, which could be regarded as likely pathogenic. The use of trios rather than single cases means that it is possible to recognize that there is little or no tendency for disruptive and damaging variants in these genes to be seen at an increased frequency in schizophrenia cases. They are not more common among cases than among the untransmitted parental alleles. We regard this as an important finding. One could easily envisage a clinical situation in which a patient with schizophrenia underwent sequencing and a disruptive or damaging variant was detected in a gene well established to cause a neuropsychiatric disorder. The sequencing might have been requested by a clinician or have been purchased from a commercial provider by the patient and their family. Our results caution that such a variant should not be regarded as likely pathogenic for schizophrenia unless there is very strong and specific evidence to support this interpretation. The vast majority of such variants should be regarded as simply being of unknown significance.

Our results also suggest that there may be considerable value in sequencing parents alongside cases, especially when there is no family history of schizophrenia. If a variant with ambiguous significance is inherited from a parent with a normal phenotype one may feel more confident that it is unlikely to be clinically relevant. Conversely, if a de novo variant is observed in a gene with a neuropsychiatric phenotype one may be more likely to interpret it as likely pathogenic.

Our study differs from an actual clinical scenario in a number of ways. We have not considered sex chromosome variants. We have not validated variants using Sanger sequencing, although this was previously done for the de novo mutations. We have used exome sequence data and it is possible that whole-genome sequencing would provide more complete coverage of all genes as well as providing information about non-coding variants, CNVs, and karyotype abnormalities.

We note a related study, which has taken a similar approach but using a different panel of genes, consisting of 37 genes known to cause inborn errors of metabolism in which psychosis could occur as the first presenting feature.⁴⁶ An exome sequenced sample of 2545 schizophrenia cases and 2545 controls was investigated. Overall, there was a 50% enrichment of rare disruptive variants in cases and although this finding is scientifically interesting no variants were identified for which one could make a clinical interpretation that they were likely pathogenic with respect to the schizophrenia phenotype.

The diagnostic yield for exome sequencing of known neuropsychiatric genes in this sample is about 1%, lower than the expected yield for testing for pathogenic CNVs of 3% or more and much lower than the yield of 40% obtained in severe developmental disorders. We can expect that the situation may change if and when more genes are identified as having an etiological role in schizophrenia and if better methods can be developed for predicting the effects of specific variants. The main conclusion of this investigation is a negative one. It is not the case that a disruptive or damaging variant in a gene known to have a neuropsychiatric phenotype should be viewed as likely to be pathogenic when seen in a patient with schizophrenia and hence the diagnostic yield from exome sequencing is currently low.

Funding

Work at Cardiff was supported by Medical Research Council (MRC) Centre (G0800509) and Program Grants (G0801418), the European Community’s Seventh Framework Programme (HEALTH-F2-2010241909 (Project EU-GEI)). Work at the Icahn School of Medicine at Mount Sinai was supported by the Friedman Brain Institute, the Institute for Genomics and Multiscale Biology and National Institutes of Health grants R01HG005827 (S.M.P.) and R01MH071681 (P.S.). Work at the Wellcome Trust Sanger Institute was supported by The Wellcome Trust (WT089062 and WT098051). The recruitment of the trios in Bulgaria was funded by the Janssen Research Foundation.

The Bulgarian Trio Sequencing study is an accumulation of exome sequencing performed and/or funded by the Broad Institute, Cardiff University, Icahn School of Medicine at Mount Sinai, and the Wellcome Trust Sanger Institute. Work at the Broad Institute was funded by Fidelity Foundations, the Sylvan Herman Foundation and philanthropic gifts from Kent and Liz Dauten, Ted and Vada Stanley, and an anonymous donor to the Stanley Center for Psychiatric Research. The authors declare they have no conflicts of interest.