Genome-Wide Association and Exome Sequencing Study of Language Disorder in an Isolated Population

Abstract

BACKGROUND AND OBJECTIVE:

Developmental language disorder (DLD) is a highly prevalent neurodevelopmental disorder associated with negative outcomes in different domains; the etiology of DLD is unknown. To investigate the genetic underpinnings of DLD, we performed genome-wide association and whole exome sequencing studies in a geographically isolated population with a substantially elevated prevalence of the disorder (ie, the AZ sample).

METHODS:

DNA samples were collected from 359 individuals for the genome-wide association study and from 12 severely affected individuals for whole exome sequencing. Multifaceted phenotypes, representing major domains of expressive language functioning, were derived from collected speech samples.

RESULTS:

Gene-based analyses revealed a significant association between SETBP1 and complexity of linguistic output (P = 5.47 × 10⁻⁷). The analysis of exome variants revealed coding sequence variants in 14 genes, most of which play a role in neural development. Targeted enrichment analysis implicated myocyte enhancer factor–2 (MEF2)-regulated genes in DLD in the AZ population. The main findings were successfully replicated in an independent cohort of children at risk for related disorders (n = 372).

CONCLUSIONS:

MEF2-regulated pathways were identified as potential candidate pathways in the etiology of DLD. Several genes (including the candidate SETBP1 and other MEF2-related genes) seem to jointly influence certain, but not all, facets of the DLD phenotype. Even when genetic and environmental diversity is reduced, DLD is best conceptualized as etiologically complex. Future research should establish whether the signals detected in the AZ population can be replicated in other samples and languages and provide further characterization of the identified pathway.

What’s Known on This Subject:

Genetic underpinnings of common forms of pediatric disorders of language are heavily understudied. Recent association studies identified several tentative candidate genes. However, thus far, none of these candidates has received strong support in replication or confirmation analyses.

What This Study Adds:

We established a statistically significant association between SETBP1 and language disorders in a geographically isolated population. Whole exome sequencing convergently implicated the myocyte enhancer factor–2–regulated pathways (of which SETBP1 is part) in language disorders in this special population.

Developmental language disorder (DLD) is a prevalent neurodevelopmental disorder, with 7-10% of children^1,2 exhibiting atypical patterns of language development despite not having apparent sensorimotor/cognitive impairments or other medical conditions.³ DLD is lifelong,⁴ comorbid with other neurodevelopmental⁵ and psychiatric⁶ disorders, and associated with adverse academic⁷ and socio-emotional⁸ outcomes. It is phenotypically complex and genetically heterogeneous; although highly heritable,⁹ the etiology and pathogenesis of DLD are poorly understood.

A rare Mendelian type of DLD has been attributed to deleterious variants in the FOXP2 gene^10–12 (7q31); however, it is not associated with the disorder’s common forms.¹³ For the latter, linkage studies have identified 3 susceptibility regions: 16q24, 19q13,¹⁴ and 13q2.¹⁵ Targeted association studies implicated CNTNAP2¹⁶ (7q35; downregulated by FOXP2) and CMIP and ATP2C2¹⁷ (16q) genes in phonological memory deficits. Four genome-wide association studies (GWAS) divulged no genome-wide significant signals,^18–21 with the exception of gene-based associations for CDC2L1, CDC2L2, LOC728661, and RCAN3.¹⁹ A whole exome sequencing (WES) study of DLD in an admixed Chilean founder population suggested the involvement of a nonsynonymous single nucleotide variant (SNV) in NFXL1²²; however, its location is not in the previously identified linkage regions in this population.²³

This pattern of findings highlights the complexity of DLD’s etiology, driven by the exclusionary nature of the diagnosis, the multicomponential nature of the phenotype, and the heterogeneity of the samples studied. The main objective of the present study was to identify genetic bases of DLD in a unique population (small, geographically secluded, and with an elevated prevalence of DLD [hereafter, the AZ population]) in which genetic and environmental variability is constrained. Genetic profiles of isolated populations are characterized by restricted genetic and allelic heterogeneity, thus rendering them ideal for studying the genetic bases of complex disorders.²⁴

The study population resides in a remote cluster of villages in Russia’s rural north; it was founded in the 15th century by several nuclear families. Currently, the AZ population comprises ∼860 individuals (∼120 children aged 3–18 years). Of these, 74.6% are represented by a set of multigenerational family structures (6391 individuals), of whom 82% are interconnected through a single 11-generational pedigree. The environmental conditions in the population are relatively uniform: all children go to the same kindergarten and school, and the socioeconomic indicators such as parental education and income show little variation. The AZ population is relatively geographically isolated and is characterized by an atypically high prevalence of DLD²⁵ (ie, ∼30% compared with 9% in the control rural population). This finding suggests the presence of a shared genetic component, potentially attributable to the founder effect(s).

Methods

Population and Sample

Altogether, 474 AZ individuals donated DNA. When considered in combination with first-degree relatives, 405 of these donors represented 79 nuclear and extended pedigrees (N = 1152; range, 3–474; median, 6). Of these, 359 underwent phenotyping and constituted the GWAS sample: 124 children (62 male subjects; age, 5.33–17.92 years) and 235 adults (102 male subjects; age, 18.83–83.42 years). A total of 149 were classified as affected (DLD) and 210 as typically developing individuals (Supplemental Information).

Phenotyping

Phenotyping was performed by clinical linguists using elicited semi-structured speech samples. These samples were scored by using previously described phenotyping procedures²⁵ to produce 5 quantitative phenotypes representing the major facets of DLD: phonetic/prosodic characteristics (eg, phonological omissions, misarticulations); well-formedness (rate of grammatical/lexical errors); complex structures (frequency of complex syntactic structures); mean length of utterance in words; and semantic/pragmatic errors (rate of errors in sentence meaning). Age-adjusted z scores were computed by using data from healthy control subjects from the comparison population to determine impairment status (ie, a z score below –1). Individuals were classified as overall DLD if they met the impairment criterion for ≥2 facets. Principal component analysis revealed that the 5 phenotypes formed 2 independent components: linguistic errors (phonetic/prosodic characteristics, well-formedness, and semantic/pragmatic errors) and syntactic complexity (complex structures and mean length of utterance in words).

Single Nucleotide Polymorphism Genotyping

The DNA extracted from peripheral blood (n = 384) or saliva/buccal swabs (n = 21) underwent quality control (QC) assessment for purity and degradation after standard collection, storage, and extraction procedures recommended by the manufacturers (Qiagen N.V. [Hilden, Germany] and DNA Genotek, Inc [Ottawa, ON, Canada]), and prepared at a concentration of 50 ng/µL.

Samples were genotyped at the Yale Center for Genome Analysis using HumanCNV 370k-Duo (n = 315) or 610k-Quad (n = 90) BeadChips (Illumina, Inc, San Diego, CA). Language status and gender distributions across the plates were not statistically different from random. Allele calling was performed by using the GenCall algorithm in GenomeStudio version 2011.1.

Samples and markers underwent QC review with GenomeStudio and SNP & Variation Suite (SVS) version 7.7.8 (GoldenHelix, Inc, Bozeman, MT). Samples with call rates >95% and verified gender were retained. A total of 223 580 autosomal single nucleotide polymorphisms (SNPs) common to 2 genotyping platforms were retained after QC so that the GenCall score was >0.30, the call rate was >95%, and minor allele frequency was >1%.

Whole Exome DNA Sequencing

Four subpedigrees were chosen for WES based on the results of complex segregation analysis²⁶ that suggested possible Mendelian transmission. From these subpedigrees, 12 severely affected individuals were selected. Nine control non-AZ individuals without DLD from the same geographical region also underwent sequencing.

Exome capture was completed by using NimbleGen EZ Exome SeqCap v2 (Roche NimbleGen, Madison, WI). One microgram of fragmented genomic DNA was used to prepare the library using the manufacturer’s protocol (Supplemental Information). The bar-coded libraries were sequenced by using Illumina’s HiSeq 2500 platform, producing 75-bp paired-end reads that were aligned to the hg19 human genome build using NovoAlign (http://www.novocraft.com; Novocraft Technologies Sn Bhd, Selangor, Malaysia). Variant calling was performed jointly for all samples by using the HaplotypeCaller algorithm in GATK.

Genetic Association Analysis

All of the quantitative trait loci association analyses were performed within the AZ sample for a set of 5 quantitative phenotypes. SNP-based association analysis of age- and gender-adjusted quantile-normalized phenotypes was performed by using mixed linear modeling (MLM) as implemented in GEMMA²⁷ version 0.94. MLM tests for genetic association of SNPs with quantitative traits were performed under the additive model while controlling for sample structure estimated directly from data as a genetic relatedness matrix. MLM can be considered an example of the de-correlation approach to family-based data, and we chose it as our analytical framework for several reasons. First, although a number of transmission-based approaches (eg, family-based association testing, FBAT) have been developed, their use in a large complex multigenerational pedigree is problematic and computationally intensive in the presence of missing genotypic or phenotypic data, requiring splitting the larger pedigree into smaller units; this approach, coupled with conditioning on the founders’ genotypes, can lead to a loss of power. Second and most importantly, comparative studies suggest that decorrelation approaches (and MLM among them) tend to have higher (or at least comparable) statistical power than transmission approaches even in large and complex pedigrees.²⁸

All 5 phenotypes were first used in a multivariate MLM analysis. Two multivariate MLMs were then fitted: 1 that modeled the genetic effects on the indicators of linguistic errors and the second that used syntactic complexity. We performed gene-based association analyses as implemented in KGG3 software²⁹ version 3.0 by using the hybrid set–based test.

Copy number variant (CNV) association analysis was performed in the FBAT³⁰ framework. Samples underwent additional CNV-specific QC (Supplemental Information). CNVs were identified by using a univariate Copy Number Analysis Method algorithm as implemented in SVS with a minimum of 5 markers per segment. Permutation testing was used to identify cut-points, and average segment intensity was used in the analyses.

Homozygosity mapping was completed by using the runs of homozygosity (ROH) detection algorithm in SVS. The minimum size was set to 250 kb and 25 SNPs, allowing for up to 1 heterozygote and 5 missing genotypes. The maximum gap between SNPs was 100 kb, and the minimum density was 18 kb. The total length of ROHs in 5 different length brackets was log-transformed to ensure normality before analysis. ROH association and burden analyses were performed by using univariate linear and logistic regression in R (R Foundation for Statistical Computing, Vienna, Austria).

WES data were analyzed by using a set of annotation and filtering tools. This analysis assumed that the most severely affected individuals in the AZ population from familial substructures with suggestive evidence for Mendelian transmission could provide additional information about the genetic architecture of DLD in the sample by focusing on: (1) the coding variants in candidate genes highlighted in the larger GWAS sample; or (2) the disruptive coding variants in other genes that could be conferring additional DLD risk in a subsample of the AZ population. Thus, we focused on coding sequence variants that were frequent among severely affected AZ probands (present in at least 4 of the 12 affected AZ individuals) but were not present in the control sample of 9 exomes. We then excluded variants observed in >5% of the National Heart, Lung, and Blood Institute Go Exome Sequencing Project (https://esp.gs.washington.edu) and the 1000 Genomes project (http://www.1000genomes.org/) Phase 1 exomes. We then retained only those variants that were located within the genes associated with any of the phenotypes in the GWAS (P < .05 for gene-level tests), disruptive frameshift variants, and variants prioritized by eXtasy³¹ based on the fusion of the information about their pathogenicity, haploinsufficiency predictions, and similarity to other genes linked to related phenotypes (Supplemental Information).

P values were corrected by using either standard or adjusted Bonferroni procedures (Supplemental Information). The study protocol was approved by the Yale University (New Haven, CT) and Northern State Medical University (Arkhangelsk, Russia) internal review boards.

Results

Genome-wide SNP Associations

No single SNP reached genome-wide statistical significance (Fig 1). Table 1 lists the top 10 nominally significant SNPs for each analysis. For linguistic errors, the strongest association (P = 5.35 × 10⁻⁷) was for rs3787751 (21q22), located in the noncoding region of the HLCS (holocarboxylase synthetase) gene, involved in the biotinylation of apocarboxylases. Holocarboxylase synthetase deficiency syndrome (MIM#253270) is characterized by neurologic, developmental, and metabolic abnormalities in infancy.³² For syntactic complexity, the top 10 SNPs included 4 SNPs (rs378968, rs3789867, rs2480933, and rs2482078) located in intronic regions of the TNC gene on chromosome 9q33 (Fig 2). TNC codes for an extracellular matrix protein implicated in cochlear development³³ and autosomal dominant deafness (MIM#615629). The univariate GWAS analyses produced similar results (Supplemental Information).

FIGURE 1.

Manhattan plots of P values for three multivariate GWAS analyses. Top row - MLM analysis of all five phenotypes; Middle row - MLM analysis of linguistic errors; Bottom row - MLM analysis of syntactic complexity.

TABLE 1.

Top 10 Nominally Significant SNP Associations for Each of the 3 Multivariate GWAS Analyses

SNP	Phenotype	Chr	Min/Maj	MAF	PPC	WF	CS	MLU	SPE	P	Gene	Variant Type
rs722208	All	6q25	G/A	0.35	0.38	0.18	−0.17	0.08	0.22	3.09 × 10−6	ESR1	Intron
rs1559831	All	15q14	T/C	0.06	−0.17	−0.64	0.65	0.17	−0.34	3.81 × 10−6		Intergenic
rs6496012	All	15q14	G/A	0.06	−0.17	−0.64	0.65	0.17	−0.34	3.81 × 10−6		Intergenic
rs3787751	All	21q22	C/T	0.05	−0.16	−0.39	0.04	0.04	−0.07	4.88 × 10−6	HLCS	Intron
rs639373	All	11q23	C/T	0.38	0.30	0.08	0.19	0.05	−0.10	1.37 × 10−5	C2CD2L	Upstream
rs2180386	All	14q32	A/C	0.38	−0.35	−0.27	0.05	−0.11	−0.22	1.60 × 10−5	MEG8	Intron
rs780382	All	11p15	A/G	0.42	−0.26	−0.04	0.22	0.09	−0.04	1.89 × 10−5	SBF2	Intron
rs585149	All	11q23	T/C	0.16	0.16	0.23	−0.02	0.45	0.16	2.83 × 10−5	SIK2, PPP2R1B	UTR′-3
rs8066993	All	17q25	T/C	0.21	−0.06	−0.36	0.15	−0.21	0.00	2.90 × 10−5	RNF213	Intron
rs2183850	All	13q34	A/G	0.48	0.03	0.04	0.28	0.21	0.16	2.98 × 10−5
rs3789868	Syn. comp.	9q33	G/A	0.47	—	—	0.06	−0.34	—	3.81 × 10−6	TNC	Intron
rs3789867	Syn. comp.	9q33	T/G	0.47	—	—	0.05	−0.35	—	4.02 × 10−6	TNC	Intron
rs2480933	Syn. comp.	9q33	A/G	0.46	—	—	0.06	−0.34	—	5.51 × 10−6	TNC	Intron
rs2482078	Syn. comp.	9q33	A/G	0.42	—	—	0.05	−0.33	—	6.57 × 10−6	TNC	Intron
rs2011604	Syn. comp.	14q32	T/C	0.43	—	—	0.11	−0.30	—	1.25 × 10−5		Intergenic
rs585149	Syn. comp.	11q23	T/C	0.16	—	—	−0.01	0.45	—	1.70 × 10−5	SIK2, PPP2R1B
rs2893567	Syn. comp.	7p14-p13	G/A	0.27	—	—	0.23	−0.28	—	2.04 × 10−5	VPS41	Intron
rs1537722	Syn. comp.	9q33	T/C	0.36	—	—	−0.03	0.33	—	2.62 × 10−5
rs4796604	Syn. comp.	17q21	T/C	0.44	—	—	0.36	0.04	—	2.96 × 10−5	HAP1	Missense
rs2699376	Syn. comp.	2q14	T/C	0.19	—	—	0.07	−0.39	—	3.15 × 10−5	CNTNAP5	Intron
rs3787751	Ling. err.	21q22	C/T	0.45	−0.16	−0.39	—	—	−0.07	5.35 × 10−7	HLCS	Intron
rs10504229	Ling. err.	8q12	A/G	0.16	0.25	0.25	—	—	0.50	5.59 × 10−6	LINC00588	Intron
rs2180386	Ling. err.	14q32	A/G	0.38	−0.33	−0.27	—	—	−0.21	7.37 × 10−6	MEG8	Intron
rs12121864	Ling. err.	1p36	C/A	0.09	0.64	0.14	—	—	0.28	8.00 × 10−6		Intergenic
rs10486031	Ling. err.	7p21	G/A	0.16	0.33	−0.05	—	—	−0.32	1.06 × 10−5	DGKB	Intron
rs639373	Ling. err.	11q23	C/T	0.38	0.31	0.08	—	—	−0.09	1.21 × 10−5	C2CD2L	Upstream
rs643788	Ling. err.	11q23	T/C	0.41	0.32	0.11	—	—	−0.06	1.22 × 10−5	DPAGT1, H2AFX	Missense
rs17445063	Ling. err.	20q13	G/T	0.11	0.29	0.53	—	—	0.09	2.16 ×10−5		Intergenic
rs722208	Ling. err.	6q25	G/A	0.35	0.37	0.18	—	—	0.22	2.48 ×10−5	ESR1	Intron
rs4895970	Ling. err.	6q23	T/C	0.50	0.36	0.15	—	—	0.14	2.75 ×10−5		Intergenic

Coefficients for phonetic/prosodic characteristics (PPC), well-formedness (WF), complex structures (CS), mean length of utterance (in words) (MLU), and semantic/pragmatic errors (SPE) represent effect size estimates (regression coefficients in SD units) for the minor allele. Chr, chromosomal location (cytoband); Ling. err, linguistic errors; MAF, minor allele frequency; Min/Maj, minor/major allele; Syn. comp., syntactic complexity; — coefficient was not estimated in the analysis and is not available.

FIGURE 2.

Regional association plots for the TNC (left) and SETBP1 (right) genes and syntactic complexity phenotype. The purple diamond represents the SNP with the lowest P value in the plotted region.

Gene-based Associations

We found no genome-wide significant gene-based associations for the 5-phenotype multivariate analysis or linguistic errors. Importantly, such an association was established between SETBP1 (SET binding protein 1; 18q21) and the multivariate syntactic complexity phenotype (P = 5.47 × 10⁻⁷) (Fig 2). The nuclear protein encoded by SETBP1 binds the SET nuclear oncogene protein involved in DNA replication, apoptosis, transcription, and nucleosome assembly. Rare variants in SETBP1 are associated with Schinzel-Giedion syndrome (MIM#269150) characterized by severe developmental delays.

Table 2 presents the top 10 genes for the gene-based analyses. After SETBP1, the 2 next strongest associations with syntactic complexity were found for 2 genes on chromosome 11q23: PPP2R1B (P = 4.77 × 10⁻⁵), encoding a constant regulatory subunit of protein phosphatase 2A, and SIK2 (P = 5.00 × 10⁻⁵), a gene hypothesized to play a role in neuronal protection. These findings are likely driven by the top hit SNP rs585149 (P = 1.70 × 10⁻⁵), assigned to both genes and located in the 3′-UTR region of SIK2. We also found a nominally significant association of syntactic complexity with TNC (P = .0068).

TABLE 2.

Top 10 Gene-Based Associations for Each of the 3 Multivariate GWAS Analyses

Phenotype	Gene	P	Chr	Length, bp
All	ESR1	4.76 × 10−5	6q25	297 602
All	ABCG4	6.53 × 10−5	11q23	13 626
All	SBF2	6.58 × 10−5	11p15	515 542
All	PPP2R1B	7.90 × 10−5	11q23	28 566
All	SIK2	8.30 × 10−5	11q23	124 464
All	HYOU1	8.44 × 10−5	11q23	13 022
All	HINFP	8.77 × 10−5	11q23	13 534
All	SNORD113-9,-7,-8	9.05 × 10−5	14q32	72–74
All	SNORD113-4,-5	9.06 × 10−5	14q32	75–78
All	H2AFX	9.10 × 10−5	11q23	1594
Syn. comp.	SETBP1a	5.47 × 10−7	18q21	388 337
Syn. comp.	PPP2R1B	4.77 × 10−5	11q23	28 566
Syn. comp.	SIK2	5.00 × 10−5	11q23	124 464
Syn. comp.	SERPINA1	7.83 × 10−5	14q32	13 947
Syn. comp.	SPATA2	2.05 × 10−4	20q13	12 153
Syn. comp.	EIF1	2.14 × 10−4	17q21	2773
Syn. comp.	ST7-OT3	2.33 × 10−4	7q31	27 258
Syn. comp.	RNF114	2.40 × 10−4	20q13	17 510
Syn. comp.	HAP1	2.52 × 10−4	17q21	12 009
Syn. comp.	GAST	2.54 × 10−4	17q21	3645
Ling. err.	LINC00588	3.44 × 10−5	8q12	5190
Ling. err.	RP11-513O17.3	3.60 × 10−5	8q12	5387
Ling. err.	SNORD113-9,-7,-8,-4,-5	4.17 × 10−5	14q32	72–78
Ling. err.	HLCS	4.40 × 10−5	21q22	239 358
Ling. err.	SNORD113-3	4.76 × 10−5	14q32	72
Ling. err.	SNORD113-6	4.90 × 10−5	14q32	75
Ling. err.	SNORD113-2,-1	5.46 × 10−5	14q32	71–72
Ling. err.	ABCG4	5.75 × 10−5	11q23	13 626
Ling. err.	HYOU1	6.34 × 10−5	11q23	13 022
Ling. err.	HINFP	6.59 × 10−5	11q23	13 534

Chr, chromosomal location (cytoband); Ling. err., linguistic errors; Syn. comp., syntactic complexity.

^aStatistically significant after Bonferroni corrections for multiple testing.

Nominally significant associations were also established between linguistic errors and several genes (ABCG4, HYOU1, and HINFP) 7 Mb away from PPP2R1B (11q23); these genes and DPAGT1 and H2AFX were associated with the combined multivariate DLD phenotype, with the top hits being rs639373 (P = 1.21 × 10⁻⁵) and rs643788 (P = 1.22 × 10⁻⁵). The functional significance of the ABCG4 product is unknown; the DPAGT1 product is crucial for glycoprotein biosynthesis; and H2AFX encodes a histone involved in the maintenance of chromatin structure. The transcription factor encoded by HINFP plays an important role in DNA methylation. We found an association between the combined multivariate phenotype and estrogen-receptor 1 (ESR1; P = 4.76 × 10⁻⁵), with rs722208 (P = 3.09 × 10⁻⁶) as a top hit. There was a nominally significant association between linguistic errors and the HLCS gene (P = 4.40 × 10⁻⁵). Neither linguistic errors nor syntactic complexity was associated with previously identified candidate DLD genes.

CNV Analysis and Homozygosity Mapping

The multivariate FBAT CNV analysis revealed several nominally statistically significant and 1 highly statistically significant CNV. However, follow-up confirmation using real-time polymerase chain reaction (PCR) failed to substantiate the presence of these CNVs. An alternative pipeline that integrated 3 CNV detection algorithms yielded no genome-wide significant associations (Supplemental Information).

Overall, AZ-affected individuals, compared with unaffected individuals, had longer cumulative lengths of ROHs that were 250 to 500 kb long (P = .006) and 1000 to 1750 kb long (P = .004), corresponding to ∼10% and ∼1% increases in estimated autosomal homozygosity, respectively (Supplemental Fig 3). The association analysis did not reveal any ROHs that were genome-wide significantly enriched in affected individuals. None of the top 20 regions overlapped with the regions identified in the SNP analyses. Several potentially relevant identified regions are discussed in Supplemental Information.

Whole Exome DNA Sequencing

We identified 14 coding sequence variants, frequent in affected AZ individuals: 4 frameshift indels, 1 inframe insertion, 2 stop gain/loss, and 7 missense variants (Table 3). SNVs were predicted by polymorphism phenotyping (PolyPhen) to be possibly or probably damaging. Although any or all of these 14 variants could be implicated in the etiology of DLD in AZ, 2 sets of findings deserve special attention.

TABLE 3.

Prioritized Coding Variants Identified in WES Data

ID	Type	Variant	N	Gene (GWAS Pa)	Brain Exp.	NHLBI Freq.	rsid	Protein Change	PolyPhen	GERP
AZW1	Del (FS)	NC_000001.10:g.6727803_6727804delTC	4/12	DNAJC1 (P = .0509)	Low	0.0000	rs374290353	NP_060668.2:p.Glu115fs	NA	NA
AZW2	Ins (FS)	NC_000003.11:g.52559298_52559299insCAAA	4/12	NT5DC2 (P = .3050)	High (pn)	0.0002	rs745485897	NP_001127703.1:p.Asp378fs	NA	5.1
AZW3	Stop gain (SNV)	NC_000006.11:g.167591956C>T	7/12	TCP10L2 (P = .0198)	Medium	0.0000	rs2297462	NP_001138593.1:p.Gln195Ter	NA	NA
AZW4	Stop loss (SNV)	NC_000009.11:g.71152177A>G	4/12	TMEM252 (P = .0921)	Low/ medium	0.0030	rs147022534	NP_694969.1:p.Ter171fs	NA	0.4
AZW5	Ins	NC_000011.9:g.1651095insGGCTGTGGCTCCGGCTGTGGAGGCCTTGGCTCCGGCTGTGGG	4/12	KRTAP5-5 (P = .0163)	NA	0.0000	NA	NP_001001480.2:p.Gly9_Cys10insLeuTrpLeuArgLeuTrpArgProTrpLeuArgLeuTrpGly	NA	NA
AZW6	Del (FS)	NC_000019.9:g.13255238delG	4/12	STX10 (P = .3610)	Medium (us)	0.0018	NA	NP_003756.1:p.Leu246fs	NA	1.1
AZW7	Del (FS)	NC_000022.10:g.40161342_40161343delGA	4/12	ENTHD1 (P = .4250)	High (us)	0.0075	rs200712517	NP_689725.2:p.Pro369fs	NA	5.9
AZW8	Missense (SNV)	NC_000007.13:g.100465824G>A	4/12	TRIP6 (P = .4260)	Medium (us)	0.0205	rs2437100	NP_003293.2:p.Arg111Gln	PrD	4.5
AZW9	Missense (SNV)	NC_000008.10:g.91953077G>T	4/12	NECAB1 (P = .0246)	High (esp. temporal)	0.0001	rs115555424	NP_071746.1:p.Ala271Thr	PrD	4.4
AZW10	Missense (SNV)	NC_000009.11:g.124072992G>A	4/12	GSN (P = .8910)	Medium (us)	0.0047	rs41305623	NP_000168.1:p.Val179Met	PrD	5.3
AZW11	Missense (SNV)	NC_000011.9:g.3700876C>T	4/12	NUP98 (P = .9480)	Low/medium	0.0164	rs34446357	NP_057404.2:p.Glu1661Lys	PrD	5.6
AZW12	Missense (SNV)	NC_000011.9:g.6629343T>A	4/12	ILK (P = .1540)	Low	0.0001	rs200336608	NP_001014794.1:p.Leu53Met	PrD	3.9
AZW13	Missense (SNV)	NC_000017.10:g.56355397G>A	4/12	MPO (P = .1420)	No	0.0135	rs28730837	NP_000241.1:p.Ala332Val	PoD	4.3
AZW14	Missense (SNV)	NC_000018.9:g.25532304T>C	6/12	CDH2 (P = .0041)	High (pn)	0.0266	rs2289664	NP_001783.2:p.Asn845Se	PoD	5.5

Brain expression (Brain Exp.) information is based on the combination of the data from the BrainSpan Atlas of the Developing Human Brain (www.brainspan.org), the Human Protein Atlas (www.proteinatlas.org), and the Expression Atlas (www.ebi.ac.uk/gxa/). Del, deletion; FS, frameshift; GERP, Genomic Evolutionary Rate Profiling sequence conservation score; Ins, insertion; N, number of carriers (of 12 individuals); NA, not available; NHLBI Freq., variant frequency in National Heart, Lung, and Blood Institute European American exomes; pn, prenatally; PoD, possibly damaging; PrD, probably damaging; us, ubiquitous; rsid, Single Nucleotide Polymorphism Database submission number.

^aP value for the HYST gene-based analysis (multivariate, all phenotypes).

First, multiple individuals in the AZ population carried coding sequence variants in genes that regulate neural development or are highly expressed in the brain; that is, a frameshift insertion in NT5DC2 (3p21.1) and missense SNVs in NECAB1 (8q21.3) and ILK (11p15.4). NT5DC2 has been implicated in schizophrenia³⁴ and borderline personality disorder.³⁵ NECAB1 is a member of the neuronal calcium-binding family of proteins essential to Ca²⁺-mediated signaling and is highly expressed in the temporal lobe.³⁶ The protein encoded by ILK is 1 of the key regulators of neural stem cell astrocytic differentiation³⁷ and neurite outgrowth.³⁸ We also found that 7 (58%) of 12 individuals in the AZ population carried a known missense variant in CDH2 (18q12) that was found only at a 2% frequency in the 1000 Genomes data set. CDH2 codes for a major cadherin that is widely expressed prenatally in neural stem cells and supports their differentiation and migration,³⁹ regulating the laminar organization of the cortex.⁴⁰ Moreover, 7 of 12 AZ individuals carried a stop-gain variant in TCP10L2 (6q27). It is unknown whether TCP10L2 codes for a functional protein; it is highly similar to TCP10L, a primate-specific transcription factor thought to evolve via segmental duplication⁴¹ from TCP10L2 or TCP10.

Second, a missense SNV in TRIP6 (7q22.1) and a frameshift deletion in ENTHD1 (22q13) indicate commonalities between the genetic pathways identified through GWAS and WES. TRIP6 is a transcription factor that has been identified as a regulator of postnatal neural stem cell maintenance in the subventricular zone.⁴² ENTHD1 codes for ENTH domain-containing protein 1. ENTH domain-containing proteins are involved in synaptic vesicle endocytosis at nerve terminals at the crucial stages that precede synapse formation.⁴³ Importantly, TRIP6 interacts with and ENTHD1 is upregulated by the same family of genes, myocyte enhancer factor–2 (MEF2), labeled MEF2A-D. MEF2 are transcription factors implicated in muscle and central nervous system differentiation. In addition to ENTHD1, MEF2 targets in human neural stem cells include SETBP1, TNC, and DKGB (3 genes highlighted by our GWAS), as well as individual genes (BDNF, DMD, and NCAM2) and gene families (cadherins, contactins, semaphorins, and serpins) implicated in (a)typical central nervous system development. A targeted formal analysis of gene list enrichment using the Enrichr tool⁴⁴ suggested that, combined, GWAS and WES hits in this population are indeed enriched for MEF2 targets (for MEF2A, P = 1.28 × 10⁻⁶) (Supplemental Information), providing support to this hypothesis.

Our WES analysis also revealed the presence of 2 heterozygous missense mutations in SETBP1, carried by 2 (rs3744825) and 1 (rs1064204) sequenced AZ individual, respectively. Both were common (for European ancestry, minor allele frequency >10% in National Heart, Lung, and Blood Institute exome database) known SNPs, projected to be tolerated according to 5 different functional prediction algorithms.

Replication

We interrogated the main loci highlighted in the GWAS or WES analyses of DLD in the AZ population in an independent sample (n = 372) of children at risk for developmental disorders of language (spoken and written) by using teachers’ ratings of student’s spoken and written language skills as the main phenotype (details are given in the Supplemental Information). Association analysis controlled for age and gender and was performed by using EMMAX,⁴⁵ a MLM algorithm implemented in SVS.

Both main findings were replicated. First, a significant gene-based association was found between language scores and SETBP1 (P = .009360). The top signal originated at exm1383999/rs11082414 (P = .000359), a missense SNP located within exon 4 of SETBP1 that explained 3.41% of the variance in children’s language skills. Predicted to be tolerated according to sorting intolerant from tolerant/PolyPhen, this SNP may play a role in the regulation of expression of SETBP1. The analysis of the Braineac⁴⁶ brain expression quantitative trait loci database suggested that it differentiates levels of SETBP1 expression in the brain, including the cerebellar cortex, hippocampus, and temporal cortex.

Second, genes nominally associated (at P < .05) with teacher ratings of students’ spoken and written language skills were enriched for MEF2A targets (P = .0007024), replicating the finding from the discovery cohort.

Discussion

We established a genome-wide association between syntactic complexity and the SETBP1 gene in the AZ sample and then replicated it in an independent sample. SETBP1 is relatively large (388 337 bp), has 2 isoforms, and is expressed widely. Although little is known about its function, it is implicated in several neurodevelopmental conditions: SETBP1 haploinsufficiency is documented in expressive DLD^47–49 and intellectual disability.⁵⁰ Moreover, several tentative SNP associations were found between syntactic complexity and TNC that encodes tenascin, an extracellular matrix glycoprotein involved in neural development; TNC-deficient mice exhibit structural and functional cortical abnormalities, including atypical neuronal density and abnormal dendrite morphology.⁵¹ However, the combined multivariate phenotype was also nominally associated with ESR1, a nuclear hormone receptor involved in regulation of gene expression, cell proliferation, and differentiation. Estrogen is involved in synaptogenesis, regulates neurotransmission, and modulates the activity of all types of neural cells.⁵² This finding is intriguing given the male bias in incidence of DLD and the recent report of associations between early postnatal gender hormone concentrations and later language development.⁵³

Our WES highlighted 14 coding variants in a set of genes implicated in neural development and/or differentiation. Intriguingly, 2 of the WES-identified genes (ENTHD1 and TRIP6) and 3 of the GWAS-identified genes (SETBP1, TNC, and DKGB) interact with or are regulated by the MEF2 transcription factors. MEF2 isoforms are widely expressed in neural cells,⁵⁴ and their activity is regulated by extracellular factors (eg, in neurons via neurotrophin stimulation or Ca²⁺ influx after the release of neurotransmitters). MEF2 targets show enriched expression in the central nervous system and implicate multiple signaling pathways, rendering MEF2 as a key regulator of activity-dependent synapse development.⁵⁵ The complex transcriptional program of MEF2 results in the restriction of excitatory synaptic transmission via the reduction of the number of excitatory neurons, elimination of glutamatergic synapses,⁵⁶ and postsynaptic differentiation of neurons (dendrite morphogenesis).⁵⁷

The cascade of events regulated by the transcriptional activity of MEF2 is critical for learning and memory.^58,59 A recent electrophysiological study partially attributed the DLD phenotype in the AZ population to atypicalities in the functioning of neural circuits that support attention and memory⁶⁰ that were linked to syntactic complexity. It is plausible they at least partially stem from the dysregulation of common genetic pathways that orchestrate neural development.

This dysregulation can take multiple forms. Given the partial convergence of the results from the GWAS and WES, we hypothesized that the DLD phenotype in the AZ population emerged as the result of the interaction between common genetic variants that conferred background DLD susceptibility and rare variants that altered the development of language and memory circuits against that background. This extension of the threshold-dependent response model suggests that common variants in several genes (eg, SETBP1, TNC) formed the probabilistic landscape(s) of DLD vulnerability, and that coding variants in multiple different genes (eg, regulated by MEF2 such as such ENTHD1 and TRIP6 or other genes important for neural development such as CDH2 or NECAB1) conferred the critical amount of vulnerability and pushed this landscape into a critical state.

Finally, we established a higher rate of autosomal ROH burden among the affected AZ individuals compared with unaffected AZ individuals; this finding is not surprising given the isolated nature of the population and the role of ROHs in several developmental disorders.⁶¹ However, no single specific ROH was strongly associated with DLD. In addition, there was little overlap between the genetic loci identified in the GWAS analyses of the 2 multivariate phenotypes; this outcome raises an interesting hypothesis that the 2 global facets of DLD may be relatively independent at the level of their molecular neurobiology.

Our study has several limitations. First, it has a small sample size. Although it was modest for a GWAS study, however, the sample size was almost one-half of the total AZ population. Second, the unique nature of the population poses a complex issue for future research seeking to replicate these signals in other samples. Although we replicated the association finding for SETBP1 and the enrichment findings for GWAS-highlighted DLD genes for MEF2 targets in an independent sample of children at risk for a related disorder, further molecular and analytical studies in larger samples are necessary to better characterize the joint contribution of common and rare variants in the identified genes to DLD susceptibility and decipher the molecular pathways they affect.

Conclusions

This study presented a set of novel candidate genes and coding DNA sequence variants contributing to DLD phenotypes in the AZ population; the chief findings from this population have been replicated in an independent sample. Overall, the findings suggest that multiple genes (including a novel genome-wide significant candidate SETBP1) and genetic pathways (including the suggested MEF2-regulated pathway) are involved in DLD. This study underlines the complexity of the genetic architecture of DLDs and illustrates that even in populations with reduced genetic and environmental diversity, DLD is best conceptualized as a polygenic and etiologically complex disorder.

Glossary

ACES

Academic Competence Evaluation Scales

CNV

copy number variant

DLD

developmental language disorder

GWAS

genome-wide association study

HWE

Hardy-Weinberg equilibrium

MEF2

myocyte enhancer factor–2

MLM

mixed linear modeling

PCR

polymerase chain reaction

PolyPhen

polymorphism phenotyping

QC

quality control

ROH

runs of homozygosity

SNP

single nucleotide polymorphism

SNV

single nucleotide variant

SVS

SNP & Variation Suite

WES

whole exome sequencing