Recessive gene disruptions in autism spectrum disorder.

Autism spectrum disorder (ASD) affects up to 1 in 59 individuals1. Genome-wide association and large-scale sequencing studies strongly implicate both common variants2-4 and rare de novo variants5-10 in ASD. Recessive mutations have also been implicated11-14 but their contribution remains less well defined. Here we demonstrate an excess of biallelic loss-of-function and damaging missense mutations in a large ASD cohort, corresponding to approximately 5% of total cases, including 10% of females, consistent with a female protective effect. We document biallelic disruption of known or emerging recessive neurodevelopmental genes (CA2, DDHD1, NSUN2, PAH, RARB, ROGDI, SLC1A1, USH2A) as well as other genes not previously implicated in ASD including FEV (FEV transcription factor, ETS family member), which encodes a key regulator of the serotonergic circuitry. Our data refine estimates of the contribution of recessive mutation to ASD and suggest new paths for illuminating previously unknown biological pathways responsible for this condition.

We performed a systematic analysis of exome data from the Autism Sequencing Consortium (ASC)[15], representing 2,343 affected and 5,852 unaffected individuals (Supplementary Table 1, Supplementary Figure 1). We cataloged a total of 696,143 autosomal loss-of-function (LOF) events (representing 28,685 unique variants in 11,745 unique genes) that introduce a stop codon or disrupt a canonical splice site (Table 1). After excluding common variants (allele frequency (AF) >1%), there were 84,645 rare LOF events (27,648 unique alleles) for an average of ~10 LOF mutations per individual. After computational phasing, we found 298 events in the ASC (after filtering to exclude common polymorphisms) consistent with complete gene knockout (homozygous or compound heterozygous LOF mutations), affecting 266 individuals (Table 1). Affected individuals were disproportionately likely to harbor a gene knockout (62% more likely; 0.05 events per affected individual versus 0.031 per unaffected individual, p=0.0003 by random permutation testing), with the bulk of the excess arising from the rarest alleles (Table 1; Supplementary Table 2; Figure 1a). To control for possible differences in population and family structure (e.g., founder effects and/or consanguinity in the Finnish and Middle Eastern cohorts), we also normalized to background burdens of biallelic synonymous variants (see Methods and Supplementary Table 3–6). Individuals with ASD continued to exhibit higher knockout rates after normalization (0.042 biallelic LOF events per affected individual versus 0.031 per unaffected individual, p=0.008 by random permutation testing) (Supplementary Figure 2a). Based on the observed ascertainment differentials (AD) between affected and unaffected individuals (0.050 vs. 0.031, or 0.042 vs. 0.031 after normalization), these burdens predict a contribution of biallelic LOF alleles to ~1–2% of ASD cases (see Methods).

Table 1.

Patterns of biallelic LOF mutation in the Autism Sequencing Consortium

				Events per individual
Allele frequency	# of biallelic LOF events	# of unique genes involved	# of individuals harboring	Unaffected	Affected	p-value
All	237,936	636	8195	28.997	29.128	0.2
≤10%	1,646	414	1420	0.194	0.219	0.03
≤5%	621	324	559	0.068	0.096	0.0005
≤1%	298	237	266	0.031	0.050	0.003
≤0.5%	225	182	201	0.022	0.042	<0.0001
≤0.1%	170	138	146	0.016	0.032	<0.0001

Figure 1.

An excess of rare, damaging, biallelic mutation in ASD. (a) Rates of biallelic gene knockout (strict LOF) in the ASC, stratified by diagnosis and allele frequency. Rates of biallelic variation, considering (b) LOF variants paired with a damaging missense variant (NsynD4, predicted to be deleterious by at least 4 algorithms), (c) LOF or NsynD4 variants, or NsynD4 variants alone. Genes impacted by biallelic variation in cases exhibit (d) haploinsufficiency and (e) pLI score profiles consistent with recessive genes. Error bars represent the 95% confidence intervals.

We considered whether these findings might be extended to incorporate the impact of missense variation. Missense variation occurs much more frequently than truncating mutations, with a small yet significant subset likely to be damaging. We evaluated potential missense contributions of two different types: all missense/nonsynonymous variants (Nsyn), and damaging missense/nonsynonymous variants (NsynD4, defined as missense changes classified as damaging by 4 or more predictive algorithms, representing 4% of all Nsyn events). Individuals with ASD exhibited an excess of biallelic events specifically involving a LOF allele in trans with a NsynD4 allele (Figure 1b), an effect which persisted after normalization to synonymous rates (Supplementary Figure 2b). Similarly, considering biallelic events involving either rare LOF or NsynD4 alleles, individuals with ASD demonstrated an excess of biallelic variation (p<10−4) with or without normalization (Supplementary Figure 2c, Figure 1c). Biallelic NsynD4 alleles were also found in excess in cases (Figure 1c) but this did not persist after normalization (p=0.07, Supplementary Figure 2c).

Alleles contributing to biallelic LOF knockouts in cases were not in excess in the heterozygous state (Supplementary Figure 3). Genes impacted by biallelic mutation also scored low with respect to haploinsufficiency[16] and gene constraint[17] (Figure 1d–e, Supplementary Table 7), compared to known dominant and de novo ASD and ID genes[18]. These data support a recessive risk model, as opposed to the additive contributions of alleles acting co-dominantly.

ASD exhibits a striking male bias, and females with ASD exhibit a disproportionate burden of de novo and inherited SNVs and CNVs, supporting a female protective/male susceptibility effect[19,20]. We asked whether this also applies to rare recessive mutations. Unaffected males and females did not differ in burdens of LOF knockout (Figure 2a, Supplementary Figure 4a). Females with ASD, however, exhibited a larger excess of biallelic LOF mutations (Figure 2a, Supplementary Figure 4a). Similar patterns were observed when the analysis was extended to incorporate damaging missense variants, either alone (NsynD4) (Figure 2b, Supplementary Figure 4b) or grouped with LOF mutation (LOF or NsynD4) (Figure 2c, Supplementary Figure 4c). As observed before, LOF alleles were not in excess in the heterozygous state (Supplementary Figure 3). From these patterns we conclude that the female protective/male susceptibility effect extends to recessive mutations as well.

Figure 2.

Biallelic mutations in ASD: effects of sex. (a) Rates of biallelic gene knockout (strict LOF) in the ASC, stratified by diagnosis and sex. (b, c) Rates of biallelic variation stratified by sex, considering (b) LOF or damaging missense (LOF or NsynD4) variants or (c) damaging missense variants alone (NsynD4). Error bars represent the 95% confidence intervals.

The ascertainment differentials observed led us to estimate the total contribution of biallelic mutation to ASD to be 3–5% (2% from LOF mutations, 1–3% from missense mutations). Furthermore, based on the ascertainment differentials, identifiable biallelic LOF or damaging missense alleles appeared to contribute to ~10% of female ASD cases in the ASC (Supplementary Figure 4b), suggesting that the category of biallelic mutation could be one of the largest genetic contributors to ASD in females.

To address potential biological implications of these results, we examined the genes in which biallelic mutations were found. We first focused on the category of strict biallelic LOF mutations, estimating from the observed ascertainment differentials that ~25–40% of those found in ASD patients (0.019 out of 0.05; or 0.011 out of 0.042 after normalization to synonymous burdens) contribute to their condition. A total of 109 unique genes were knocked out (biallelic LOF) in affected individuals in our dataset. We reasoned that the most relevant of these genes would be those that are selected against biallelic inactivation in the general population. We therefore developed a filter to remove any gene that is also knocked out in one or more of 60,706 individuals sequenced as a part of ExAC (see Methods)[17]. To test its effectiveness, we examined its impact on observed patterns and burdens of gene knockouts in ASD. Overall, 131 of the 227 (58%) gene knockouts observed in the ASC were also observed in ExAC. Application of this filter reduced the overall burdens of observed biallelic knockout by half but left the ascertainment differentials intact (Figure 3a, Supplementary Figure 5). This significantly increased the observed proportional excess of biallelic LOF mutation: after filtering, on average, individuals with ASD were ~2.6 times as likely to have biallelic LOF mutations as individuals without ASD (p=0.002) (3.9X in females, p=0.001, and 2.6X in males, p=0.01) (Figure 3a).

Figure 3.

Biallelic mutations in ASD: ExAC filtration. (a) Rates of biallelic gene knockout (strict LOF) after filtration of commonly inactivated genes in ExAC. (b) Breakdown of candidate ASD genes excluded by presence of knockouts in ExAC, or by knockouts that are common and/or in unaffected individuals of the ASC. Error bars represent the 95% confidence intervals.

We applied a similar logic to biallelic missense mutations as well. We reasoned that missense alleles encountered in the homozygous state in the general population are not likely to contribute to recessive condition. ~50% of rare, biallelic damaging missense events in the ASC involved an allele encountered in the homozygous state in ExAC (Supplementary Figure 6a, Supplementary Table 8). Filtering these events allowed us to discern for the first time a significant excess of rare biallelic damaging missense variation (i.e. NsynD4) in cases compared to controls (Supplementary Figure 6a), with an ascertainment differential of ~4%. Biallelic Nsyn variants were also in slight excess (p=0.0037), which arose almost entirely from affected females (Supplementary Figure 6b).

These results indicate that the observed excess in ASD is driven by genes and alleles that are biallelically constrained in the control population, strongly supporting the biological relevance of these findings.

De novo gene disrupting mutations contribute strongly to ASD[21]. To examine possible interactions between de novo and biallelic gene disrupting events, we compared the burden of biallelic LOF events in cases with and cases without LOF de novo mutations (Supplementary Figure 10)[21]. An excess of rare biallelic LOF events is strongly evident in ASD cases lacking de novo LOF variants (Supplementary Figure 10, p=0.0006). Their contribution in ASD cases with de novo LOF variants is harder to distinguish, as the observed difference in burdens of biallelic LOF events did not achieve statistical significance (Supplementary Figure 7). These may represent orthogonal risk categories, although larger sample sizes will be required to clarify this in the future.

After ExAC filtering, a total of 57 biallelically constrained genes harbored biallelic LOF mutation in at least one ASD case. Of these, we can estimate that ~60% of these (based on the observed burdens, see Methods), and ~80% of the knockouts observed in females, likely confer ASD risk. To further enrich for relevant genes, we also set aside 13 genes knocked out in one or more unaffected individuals in our ASC dataset (Figure 3b, Supplementary Table 7), and three that failed Sanger confirmation, leaving a final set of 41 genes that were specifically knocked out in individuals diagnosed with ASD, but never in 66,557 controls (60,706 from ExAC and 5,852 unaffected individuals from ASC) (Supplementary Table 9,10).

Several of these represent genes already implicated in disease. Eight (CA2, DDHD1, NSUN2, PAH, RARB, ROGDI, SLC1A1, USH2A) were associated with recessive genetic conditions involving neurodevelopmental delay (Table 2 and Supplemental Information for clinical details). In at least six of these cases (CA2, NSUN2, PAH, RARB, ROGDI, USH2A), available medical and phenotypic records were concordant with the clinical features predicted from the genetic finding (Supplemental Information). Phenotypic data was either unavailable or incomplete, but not discordant, for other cases. In two cases, knockouts were encountered in genes that cause well-characterized autosomal recessive conditions without clear connection to autism (RFX5 and DNAI2, associated with autosomal recessive immunodeficiency and primary ciliary dyskinesia, respectively). In two other cases, knockouts were observed in genes associated with autosomal dominant disorders via recurrent gain-of-function mutations (IFITM5 and KIF22, causing osteogenesis imperfecta type V and spondyloepimetaphyseal dysplasia with joint laxity type 2, respectively). The phenotypic consequence of recessive loss-of-function mutations for these two genes is unknown. Interestingly, KIF22 is one of 25 genes in the 593kb core interval affected by 16p11.2 deletions.[22] KIF22 was proposed as a one of two genes in the interval that demonstrates dosage sensitivity, with knockdown causing axon tract and behavioral defects in zebrafish,[23] and disruption a KIF22 ortholog in Drosophila causes synapse development defects[24].

Table 2.

Clinically relevant gene knockouts in the ASC

Gene	Mutations	Disease relevance	Notes
CA2	hom c.232+1G>A (NM_000067)	Carbonic anhydrase deficiency	Intellectual disability, osteopetrosis, renal tubular acidosis
DDHD1	comp het c.1311–2A>T/c.2459–2A>T (NM_001160147)	Spastic paraplegia	Spastic paraplegia, mitochondrial abnormalities; only 3 cases reported
NSUN2	hom c.C1708T:p.Q570X (NM_001193455)	Autosomal recessive intellectual disability	Syndromic intellectual disability; only 7 mutations reported
PAH	hom c.C703T:p.Q235X and hom c.592_613del:p.Y198fs (NM_000277) (two separate individuals)	Phenylketonuria	~20% of individuals with PKU have autism (finding previously reported[13])
SLC1A1	hom c.G142T:p.E48X (NM_004170)	Dicarboxylic aminoaciduria	Elevated urinary glutamate and aspartate, variable neuropsychiatric symptoms; only 3 cases reported
RARB	hom c.C78A:p.C26X (NM_000965)	PDAC syndrome	Microphthalmia, pulmonary hypoplasia, diaphragmatic hernia, and cardiac defects; milder forms described
ROGDI	hom c.201–1G>T (NM_024589)	Kohlschütter-Tönz syndrome	Global developmental delay, epilepsy, spasticity, amelogenesis imperfecta
USH2A	comp het c.T12714G:p.Y4238X/ c.G6224A:p.W2075X (NM_206933)	Usher syndrome	Sensorineural hearing deficiencies, retinitis pigmentosa (finding previously reported[14])

Several of the remaining 33 genes are of strong neurobiological interest. One affected individual bore a homozygous LOF mutation (p.E48X) in SLC1A1, mutations in which have been linked to dicarboxylic aminoaciduria,[25] an extremely rare disorder with prior associations with intellectual disability[26,27] and obsessive compulsive disorder (OCD)[25]. SLC1A1 encodes a brain-expressed, activity-regulated glutamate transporter[28,29] that modulates glutamatergic neurotransmission via a variety of mechanisms[30,31,32,33,34,35] SLC1A1 knockout in mice results in altered locomotor behavior[36] and learning and memory defects.[37]

Two brothers with ASD bore homozygous stop-gain mutations in FEV, a transcription factor that is required for both development and function of serotonergic neurons (Supplementary Figure 8a). These neurons send projections that ramify broadly throughout the CNS (Supplementary Figure 8b), and have wide-ranging neuromodulatory effects on many physiological states and behaviors, including mood, aggression, anxiety, sleep, and movement[38]. Serotonergic dysfunction has been strongly implicated in neuropsychiatric conditions including ASD[39], schizophrenia and mood disorders[38]. In mice, Fev/Pet-1 has been shown to be highly expressed in the serotonergic raphe nuclei[40], and inactivation of Fev during development[41] or adulthood[42] results in depletion of serotonergic neurons and neurobehavioral changes including anxiety and aggression.[41,42] Human FEV is also expressed in the serotonergic raphe in adult brain[43,44], and we found similar results via in situ hybridization analysis of human fetal brain as well (Supplementary Figure 8c–f). The older affected brother was diagnosed with autism and had a measured IQ of 69. He exhibited severe, stereotyped aggressive and self-injurious behaviors. He had a BMI of 25.6 (99%) and mildly dysmorphic facial features. EEG showed bilateral rolandic focus with generalization. The younger affected brother was diagnosed with PDD-NOS and intellectual disability (IQ not measured), had a BMI of 17.7 (91%), and no facial dysmorphisms (further details in Supplemental Information). We are not aware of any previous report of a human loss-of-function phenotype for FEV.

Gene knockouts were also identified in FCHSD2, a cytoskeletal adaptor protein that is highly expressed in human fetal brain and is the human homolog of Drosophila nervous wreck[45], NCOA7, a brain-specific transcriptional coactivator for steroid and nuclear hormone receptors[46], and MOB1B, a downstream effector in the Hippo signaling pathway, a pathway recently implicated in regulation of brain size[47]. Many additional genes in the candidate set have abundant brain expression (BRMS1L, ELOF1, GCN1L1, NUPR1, SLC22A6) but less clear neurobiological activity.

We also expanded our scope to incorporate biallelic missense mutation. There were 409 distinct genes that harbored biallelic, damaging missense mutations in cases but not in controls (Supplementary Figure 9, Supplementary Table 11). Only one gene was found that was independently hit in multiple families: AMT, an established cause of nonketotic hyperglycinemia (NKH). Hypomorphic mutations in AMT have previously been reported to cause ASD without NKH’s characteristic metabolic abnormalities. No other gene was hit more than once, so larger sample sizes would be necessary to establish any of these genes as ASD genes on this basis alone (Supplementary Table 11, Supplementary Figures 10 and 11). However, there were thirteen cases in which affected individuals bore biallelic mutations in alleles of established medical significance (in ClinVar as Pathogenic or Likely Pathogenic) (Table 3, Supplementary Table 12, Supplementary Information). All were diagnostic of genetic disorders with known neurodevelopmental consequences including classic metabolic diseases, mitochondrial depletion syndrome, and other syndromic conditions. All together, combined with the previously described eight LOF cases, biallelic mutations in known genes constituted ~1% of our cohort (21 out of 2,343 affected individuals), underscoring the importance of clinical screening for monogenic recessive conditions in this patient population.

Table 3.

Biallelic mutations identified in cases and controls previously reported as pathogenic or likely pathogenic in ClinVar.

Gene	Condition	Sample identifier	Sex	Mutation	ClinVar (HGMD)	Variant pathogenicity	Damaging Predictions
Found in 13 cases
ASS1	Citrullinemia type I	SKUSE5080161	M	hom c.1168G>A:p.G390R(NM_000050.4)	6329 (CM900037)	Pathogenic	10 / 10
ATP7B	Wilson disease	80001103644	M	het c.3207C>A:p.H1069Q(NM_000053.3)	3848 (CM930059)	Pathogenic	11 / 11
				het c.4087T>C:p.S1363P(NM_000053.3)	Same site as p.S1363F (ClinVar 188908, HGMD CM992829)	(p.S1363F=Pathogenic)	12 / 12
B3GALNT2	Muscular dystrophy with brain malformations	80001103921	F	het c.1368+1G>A(NM_152490.4)	na	(Likely Pathogenic)	na
				het c.979G>A:p.D327N(NM_152490.4)	424766 (CM144415)	Likely Pathogenic	10 / 12
CA2	Osteopetrosis, autosomal recessive 3	AU-10501	F	hom c.232+1G>A(NM_000067.2)	288909	Pathogenic	na
COQ8A	Coenzyme Q10 deficiency	AU-13901	M	hom c.1805C>G:p.P602R(NM_020247.4)	214050 (CM141297)	Pathogenic	11 / 12
GBE1	Glycogen storage disease, type IV and Adult polyglucosan body disease	AU-24601	F	hom c.986A>G:p.Y329C(NM_000158.3)	371439 (CM128776)	Pathogenic	11 / 11
GLRA1	Hereditary hyperekplexia	AU-12001	M	hom c.277C>T:p.R93W(NM_000171.3)	225379 (CM104312)	different missence LP in ClinVar (225379)	10 / 11
HADH	Deficiency of 3-hydroxyacyl-CoA dehydrogenase	09C86928	M	hom c.676T>C:p.Y226H(NM_005327.4)	212734 (CM064046)	Likely pathogenic	11 / 11
MYO15A	Deafness, autosomal recessive 3	SAGA-96	F	het c.5515C>T:p.Q1839X(NM_016239.3)	na	(Likely Pathogenic)	na
				het c.8183G>A:p.R2728H(NM_016239.3)	228276 (CM117958)	Pathogenic	10 / 10
NAGLU	Sanfilippo syndrome B	AU-16201	F	homc.934G>A:p.D312N(NM_000263.3)	437446 (CM113463)	Pathogenic	12 / 12
PAH	Phenylketonuria	AU-16201	F	het c.782G>A:p.R261Q(NM_000277.1)	582 (CM910287)	Pathogenic	12 / 12
				het c.842C>T:p.P281L(NM_000277.1)	589 (CM910292)	Pathogenic	11 / 11
POLG	Alpers-Huttenlocher syndrome, Childhood myocerebrohepatopathy spectrum	AU-16201	F	hom c.3151G>A:p.G1051R(NM_002693.2)	13501 (CM040472)	Pathogenic	11 / 11
USH2A	Usher Syndrome, Type 2A	09C96107	M	het c.12714T>G:p.Y4238X(NM_206933.2)	48405 (CM134383)	Pathogenic	na
				het c.6224G>A:p.W2075X(NM_206933.2)	48554 (CM134384)	Pathogenic
Found in 3 controls
MCCC2	3-methylcrotonyl-CoA carboxylase deficiency	AU-13303	F	hom c.295G>C:p.E99Q(NM_022132.4)	1920 (CM010910)	Pathogenic	10 / 10
PAH	Phenylketonuria	AU-16503	F	hom c.898G>T:p.A300S(NM_000277.1)	92751 (CM920555)	Pathogenic	12 / 12
		AU-16502	M	hom c.898G>T:p.A300S(NM_000277.1)	92751 (CM920555)	Pathogenic

Our study extends prior estimates of the contribution of recessive mutations to ASD[13,14] and proves that the female protective / male susceptibility effect in ASD applies to biallelic mutations as well. Expanding these efforts to larger ASD cohorts should offer a path forward to delineate novel neurobiological mechanisms in autism.

METHODS

Whole exome data analyses

Whole-exome sequencing data were gathered as part of the Autism Sequencing Consortium[48]. We focused our analysis on nine datasets generated using the Illumina sequencing platform, Supplementary Table 1). Alignment and variant calling were performed as previously described. Low-quality variants were filtered with the following criteria: (1) SNPs with GQ<20, (2) SNPs not passing the standard PASS filter, and (3) all indels (due to excessive false positives leading to spurious frameshift calls). Variant annotation was performed using ANNOVAR[49] to add allele frequencies, functional predictions, conservation, and gene annotations including OMIM and HGMD[50] disease associations. Allele frequencies were also calculated for each contributing cohort and across the entire collection of ASC samples analyzed. Variants were categorized by their predicted impact into synonymous, nonsynonymous (missense), or loss-of-function (altering a canonical splice site or resulting in a stopgain). Additional classification of damaging missense variants (NsynD4) was performed for variants predicted to be deleterious by at least 4/6 algorithms (SIFT[51], PolyPhen2_HDIV[52], PolyPhen2_HVAR, MutationTaster[53], MutationAssessor[54], and LRT[55]).

Phasing was performed using Beagle 4[56] on the ASC dataset. VCFs were subdivided by chromosome and cohort prior to processing with Beagle employing default parameters.

Biallelic burden calculations

Allele frequency filtration was performed as described in the text using available public databases (1000G, EVS, ExAC). For rare variant calculations at maxAF 0.1%, 0.5%, 1%, 5%, and 10%, additional cohort-specific filters (cohort-specific AF of 5% or 10%) were applied to account for allelic variants that may be prevalent in a population but are not well represented in existing control databases. Burdens of homozygous and compound heterozygous variants were then counted for each cohort and for the entire ASC, stratifying by affected/unaffected status, gender, and mutational category (LOF, Nsyn, NsynD4, Syn).

To control for potential differences in underlying family and population structure of some of the constituent cohorts, we also measured rates of synonymous biallelic variation. In two cohorts (from the Middle East and Finland, respectively), rates of biallelic synonymous variation in affected individuals were higher than rates of biallelic synonymous variation in unaffected individuals (Supplementary Table 3). Therefore, for these cohorts, we normalized our LOF, NsynD4, and Nsyn burden calculations to the background rate of biallelic synonymous variation, to isolate the impact of biallelic LOF variation independent of background homozygosity in these cohorts.

The contribution of mutations to ASD risk was estimated using the following approach. First the ascertainment differential (AD) is calculated as the difference between the burden of biallelic mutations in cases and in controls and represents the estimated fraction of mutations that contribute to ASD (for example, an AD of 0.01 suggests that 1% of cases have a biallelic mutation contributing to ASD risk. The fraction of observed biallelic mutations contributing to disease is then calculated as AD divided by the total mutation burden in cases. For example, given a burden of 0.025 biallelic LOF events in cases and 0.01 biallelic LOF events in controls, the AD is 0.015 and the percentage of mutations identified which contribute to disease risk is 0.015 divided by 0.025, or 60%.

For burden comparisons, significance was assessed by permutation testing using 10,000 random assignments of diagnostic categorization to individuals within the study. A significance threshold of p<0.05 (i.e., without correction for multiple testing) was employed since the allele frequency cutoffs and variant annotation classes (LOF, LOF+NSynD4, LOF or NSynD4, NSynD4, NSyn) analyzed are not discrete. We do note that, all ExAC-filtered biallelic burden results (i.e., excess biallelic LOF, NSynD4, and NSyn mutations, summarized in Figure 3 and Supplementary Figures 5 and 6) were significant even with application of maximally conservative Bonferroni correction (e.g., p<0.01 assuming five fully independent hypotheses).

ExAC filtration

To generate a list of genes commonly inactivated by biallelic LOF mutation, we analyzed whole exome data from 60,706 individuals compiled by the Exome Aggregation Consortium (ExAC)[17]. We applied ANNOVAR to systematically re-annotate all variants from ExAC against RefSeq complete protein coding genes. Since phasing was unavailable, we focused on genes bearing putative homozygous LOF variants (any variant in stopgain, frameshift, or canonical splicing change). Homopolymer expansions / contractions in canonical splice regions were ignored, as expansions and contractions in these regions are common and have little impact on splicing. With these parameters, a total of 2986 genes were found to be inactivated by homozygous stopgain, frameshift, or canonical splice site mutation in ExAC. We also used this re-annotated dataset to generate a list of ExAC alleles encountered in the homozygous state, to identify missense and other mutations that are more likely to be clinically benign. For the final LOF candidate genes presented in Supplementary Tables 9 and 10, contributory alleles were also confirmed to be absent in the homozygous state in the gnomAD dataset[17].

Variant confirmation

Sanger sequencing was employed to confirm the accuracy of exome variant calls for a subset of LOF and NsynD4 variants with an empirical validation rate of 96% (107 of 114 alleles tested). These included all LOF alleles reported here for which DNA samples were available. If DNA samples were not available, we note the presence of allelic variants in known databases and/or whether predicted genotypes follow Mendelian segregation patterns in parental exome data (where available) as additional support.

Human tissue collection and in situ hybridization for FEV

Human embryonic and fetal tissues were obtained from the MRC-Wellcome Trust Human Developmental Biology Resource (www.hdbr.org)[57] following appropriate consent from the women involved and adhering to the relevant HTA guidelines with approval from Research Ethics Committee NRES Committee North East - Newcastle & North Tyneside 1.

RNA probes were produced by amplifying human embryonic cDNA using a forward primer to FEV (NM_017521, nt. 1286–1305) tagged with a T7 RNA promoter sequence and a reverse primer (NM_017521, nt. 1862–1843) tagged with a SP6 RNA promoter sequence (Supplementary Table 13). The resulting fragment was in vitro transcribed using the DIG RNA Labeling Kit (Roche).

The in situ hybridizations were performed as previously described[48], except that 60ng RNA probe in 100ul ULTRAhyb (Ambion) per slide was used for the hybridization and the post-hybridization washes were modified: slides were washed for 10 minutes in 5 X SSC, twice in 2 X SSC, 50% formamide/1 X SSC, 1 X SSC and finally in 0.1X SSC. The hybridization incubation and all post-hybridization steps were performed at 63.5°C. Three human fetal brain samples were used: a Carnegie Stage (CS) 23 brain (12959), a CS23 head (12011), and a 10 post conception week hindbrain (sample 12721). Findings described were confirmed in at least two specimens. Expression patterns were analyzed relative to spatial models provided by the HuDSeN human gene expression spatial database project (http://www.hudsen.org).[58]

Reporting Summary

Further information on research design is available in the Life Sciences Reporting Summary linked to this article.