Relative Burden of Large CNVs on a Range of Neurodevelopmental Phenotypes

Summary

While numerous studies have implicated copy number variants (CNVs) in a range of neurological phenotypes, the impact relative to disease severity has been difficult to ascertain due to small sample sizes, lack of phenotypic details, and heterogeneity in platforms used for discovery. Using a customized microarray enriched for genomic hotspots, we assayed for large CNVs among 1,227 individuals with various neurological deficits including dyslexia (376), sporadic autism (350), and intellectual disability (ID) (501), as well as 337 controls. We show that the frequency of large CNVs (>1 Mbp) is significantly greater for ID–associated phenotypes compared to autism (p = 9.58×10⁻¹¹, odds ratio = 4.59), dyslexia (p = 3.81×10⁻¹⁸, odds ratio = 14.45), or controls (p = 2.75×10⁻¹⁷, odds ratio = 13.71). There is a striking difference in the frequency of rare CNVs (>50 kbp) in autism (10%, p = 2.4×10⁻⁶, odds ratio = 6) or ID (16%, p = 3.55×10⁻¹², odds ratio = 10) compared to dyslexia (2%) with essentially no difference in large CNV burden among dyslexia patients compared to controls. Rare CNVs were more likely to arise de novo (64%) in ID when compared to autism (40%) or dyslexia (0%). We observed a significantly increased large CNV burden in individuals with ID and multiple congenital anomalies (MCA) compared to ID alone (p = 0.001, odds ratio = 2.54). Our data suggest that large CNV burden positively correlates with the severity of childhood disability: ID with MCA being most severely affected and dyslexics being indistinguishable from controls. When autism without ID was considered separately, the increase in CNV burden was modest compared to controls (p = 0.07, odds ratio = 2.33).

Introduction

Recent studies have implicated large, rare CNVs in a range of neurodevelopmental disorders including intellectual disability (ID) [1], [2], autism [3], [4], schizophrenia [5], [6], bipolar disorder [7], [8], epilepsy [9], [10], and attention deficit hyperactivity disorder (ADHD) [11], [12]. Several themes have emerged from these studies: first, a significant enrichment for rare CNVs in individuals with the disease compared to unaffected controls was observed, independently, for each of these disorders; second, the same recurrent CNVs are associated with different neuropsychiatric phenotypes; and third, locus heterogeneity is substantial as many distinct variants can lead to similar phenotypes.

Our understanding of the relevance of rare CNVs across a broad spectrum of neurodevelopmental disorders, varying in severity and prevalence, is limited as previous studies were restricted to the analysis of one phenotype at a time and each of such studies was performed using different CNV genotyping methodologies with distinct platform-specific biases, making comparisons difficult. We undertook a systematic analysis of 1,227 cases and 337 controls to assess the relative contribution of CNVs in three phenotypically distinct neurodevelopmental disorders. We designed a whole-genome custom microarray targeted to genomic hotspots for comparative genomic hybridization (CGH) to identify potentially pathogenic CNVs that contribute to ID, autism, and dyslexia.

Results

We analyzed 1,227 individuals ascertained for three neurodevelopmental disorders: 376 dyslexic children with a verbal IQ (VIQ) ≥90 on the Wechsler Intelligence Scale for Children [13] and dyslexia defined as poor performance and IQ-performance discrepancy in one or more of a set of standardized reading measures, 350 cases with sporadic autism from the Simons Simplex Collection (SSC), and 501 cases with ID. We used 337 NIMH control individuals for comparison. Further, based on the presence or absence of ID (full-scale IQ score cutoff of 70), autism cases were divided into those with ID (n = 97) or without ID (n = 253) (see Materials and Methods). Based on the presence of multiple congenital anomalies (MCA), individuals with ID were divided into those with ID only—i.e. idiopathic ID (n = 428)—and those with ID and MCA (n = 73).

All copy number variation analyses were performed using a custom microarray with a high probe density (∼2.6 kbp) targeted to 107 genomic hotspot regions [14] (∼251 Mbp) and a median probe spacing of ∼36 kbp in the genomic backbone (see Materials and Methods, Table S1). We used a Hidden Markov Model (HMM)-based algorithm to identify deletions and duplications. We restricted our analysis to CNVs >50 kbp in size to reduce false positive calls and validated all relevant CNVs using a second custom designed high-density array. To empirically determine the validation rate of the array at different genomic regions, we examined 118 CNVs detected in 24 samples and confirmed 117 events (>99% accuracy, see Table S2). While we were easily able to detect smaller events in the hotspot regions, the specificity of the array restricted our CNV discovery to >50 kbp in hotspot-associated regions and to >300 kbp in regions not associated with genomic hotspots (Figure S1).

Analysis for large CNV burden in neurodevelopmental phenotypes

After quality control (QC) filtering and manual curation, we obtained 5,086 CNVs in 1,395 out of 1,564 individuals (89.2%) with high-quality array CGH data (Table 1; Datasets S1, S2, S3, S4). Using these data, we compared the CNV enrichment between the multiple cohorts tested. We found a significant excess of large CNVs (>1 Mbp) in individuals with ID (p = 2.75×10⁻¹⁷, odds ratio = 13.71) or autism (p = 0.012, odds ratio = 2.99) when compared to controls analyzed on the same microarray platform (Figure 1). The frequency of large CNVs among children with dyslexia was similar to controls (p = 0.64, odds ratio = 0.94), although this might indicate a lack of statistical power in our study to detect any subtle enrichment (power >0.8 to detect 4.2% increase in burden) for large CNVs in dyslexia.

Within the neurodevelopmental disorder cohorts, a comparison showed a significantly greater large CNV burden in individuals with ID compared to autism (p = 9.58×10⁻¹¹, odds ratio = 4.58) or dyslexia (p = 3.81×10⁻¹⁸, odds ratio = 14.45). When we partitioned the ID cohort into subsets with and without MCA, we observed a significantly increased large CNV burden in individuals with ID/MCA compared to ID alone (p = 0.001, odds ratio = 2.54). This trend was also observed when individuals with autism were separated into those with ID and without ID (Figure 1), although not statistically significant (p = 0.102, odds ratio = 2.1). When compared to controls, we noted a trend for increase in large CNV burden for autism without ID (p = 0.07, odds ratio = 2.33) as well as autism with ID (p = 0.0048, odds ratio = 4.85). In addition, a gene-based analysis showed an incremental increase in the proportion of disrupted genes and average gene density per CNV with higher estimates for the ID/MCA cohort as compared to ID alone or autism (Table 1). We also note that within the cohorts no bias towards deletions or duplications was observed in relation to phenotypic severity or variability (Tables S3, S4, S5, S6, S7). Overall, our results suggest a positive correlation of the severity of the phenotype to the size and gene density of CNVs.

Rare CNVs in dyslexia, autism, and intellectual disability phenotypes

To identify rare CNVs of likely pathogenic significance, we compared the pattern of CNVs from dyslexia, autism, ID, and NIMH control cohorts to a map developed from an expanded set of 8,329 normal individuals genotyped with Illumina microarrays and to the publicly available Database of Genomic Variants [15] (see Materials and Methods). We eliminated common copy number polymorphisms and CNVs from our cases if they had a reciprocal overlap of 50% or more of their length with CNVs found in these 8,329 controls. After filtering, we compared the groups. We found a significant increase of rare CNVs in individuals with autism (35/336, 10%; p = 2.4×10⁻⁶, odds ratio = 6) or ID (69/431, 16%; p = 3.55×10⁻¹², odds ratio = 10) compared to individuals with dyslexia (6/322, 2%) (Figure 2A, Table 2). In fact, when analyzed separately, the frequency of rare CNVs in NIMH controls (6/306, 2%) was not different compared to dyslexia (p = 0.57, odds ratio = 0.94) (Table S8).

Given the high population prevalence of dyslexia [16], we then relaxed our selection to include events present at an allele frequency of <0.1% in controls (8/8,635) and identified four additional CNVs—i.e., a total of 10 CNV events (Table 3). The analysis of hotspot regions identified only one individual with dyslexia who carried a 15q11.2 BP1–BP2 deletion, which has previously been associated with ID [17], schizophrenia [18], [19], and epilepsy [20]; however, this deletion was also observed in 25/8,635 of our total control individuals. None of the seven deletions and three duplications detected in our dyslexia cohort mapped to candidate loci known to be associated with dyslexia [21].

Analysis of 336 individuals from the SSC autism cohort showed that 35 individuals (10%) carried 36 rare CNVs (680 RefSeq genes, median size = 662 kbp) and about 58% (21/36) of these CNVs mapped to genomic hotspots (Table 2). Only eight of the events (all hotspot sites) associated with genomic disorders, including 22q11.2 deletion (TBX1, DiGeorge syndrome), 17p12 duplication (PMP22, Charcot-Marie-Tooth disease), and 15q11.2q13.1 duplication (UBE3A and SNRPN) (Table S9). In addition, as reported previously [3], [22], [23], the autism-associated proximal 16p11.2 deletion (TBX6) was observed in approximately 1% (3/336) of all autism cases analyzed. Interestingly, one case with a de novo 16p11.2 deletion also inherited a 2 Mbp duplication 22q11.2 (TBX1) from the mother.

Among 431 cases with ID (358 cases with ID only and 73 cases with ID plus MCA), 69 individuals carried 77 rare CNVs (2,215 RefSeq genes, median size = 1.5 Mbp) that were either of known pathogenic significance or not observed in a total set of 8,635 controls, and 32% (25/77, median size = 1.42 Mbp) of these variants localized to genomic hotspot regions (Table 2). This is a significant enrichment for rare CNVs in the ID cohort compared to autism (p = 0.019, odds ratio = 1.6) or dyslexia cohorts (p = 3.55×10⁻¹², odds ratio = 10). Interestingly, 20/77 CNVs (16 hotspot and four non-hotspot sites) mapped to a known genomic disorder site, including those associated with variable phenotypes such as 15q13.1q13.3 (CHRNA7), 16p11.2 proximal (TBX6; two cases) and distal (SH2B1) hotspots, 16p13.11 (MYH11; three cases), 17q12 (TCF2), and 3q29 (DLG1) as well as syndromic regions such as 7q11.23 (Williams syndrome), 17q21.31 (MAPT), 5q35 (Sotos syndrome), 8p23.1, 22q13 (Phelan-McDermid syndrome) [24] and 1p36 [25].

We next sought to determine whether these rare CNVs were inherited or if they arose de novo in the probands. Parental DNA samples were available to investigate inheritance for 90 out of 123 rare CNVs detected in all three disease cohorts (Table S8). In four cases, only maternal DNA was available. We find that 44/90 CNVs arose de novo and a majority (77%, 34/44) of these de novo CNVs were large (>1 Mbp). Overall, we find a greater proportion of de novo events in ID (64%, 30/47) compared to autism (40%, 14/35; p = 0.027, odds ratio = 2.6) or dyslexia (0/8; p = 0.0009, odds ratio = infinity) (Figure 2B). These data are suggestive of a general trend of increased de novo rates and CNV size with increased severity of the disorder.

Novel, rare CNVs reveal potential candidate genes

We then focused on rare CNVs involving single genes or regions of potential interest. In the dyslexia cohort, two unrelated families carried CNVs on chromosome 7q11.23 that involved the autism susceptibility candidate 2 (AUTS2, MIM# 607270). A 669 kbp duplication that included AUTS2 and WBSCR17 was transmitted from an affected father to the daughter and an approximately 84 kbp deletion was transmitted from the affected paternal grandmother through the unaffected father to the proband (Figure 3). In addition, we also identified a 354 kbp deletion encompassing AUTS2 in one individual with idiopathic ID, pervasive developmental delay, partial epilepsy, and left hemihypertrophy. An approximately 1.2 Mbp deletion encompassing the eyes shut drosophila homolog gene (EYS, MIM# 612424) on chromosome 6q12 was detected in an affected proband and several unaffected family members. Although autosomal recessive single-nucleotide mutations in EYS have been reported in patients with retinitis pigmentosa [26], [27], the role of heterozygous microdeletions involving this gene is unknown.

We also identified a 471 kbp deletion involving IMMP2L inherited by the proband from the affected mother (Figure 3). Deletions involving IMMP2L have been associated with ADHD [12], autism [28], and Tourette syndrome [29]. Recently, Pagnamenta and colleagues also reported a 594 kbp IMMP2L-DOCK4 deletion resulting in a fusion transcript and an intragenic DOCK4 deletion segregating with dyslexia [30]. Our results are best interpreted within the context of candidate gene identification in dyslexia. Although at least nine chromosomal loci are associated with dyslexia, for two of these loci the candidate genes were identified on the basis of a rare balanced chromosomal translocations disrupting ROBO1 [31] and DYX1C1/EKN1 [32], [33]. More recently, a Danish Cytogenetic Registry study of all cases with chromosomal translocations identified additional novel dyslexia candidate genes affirming the value of rare structural variants in understanding the genetics of dyslexia [34]. Our study is the first to systematically characterize rare CNVs in dyslexia and thus evaluate the contribution of rare deletions and duplications to this common genetic disorder.

Within the autism cohort, several novel deletions and duplications involving neurologically-relevant genes were identified. A 5 Mbp de novo deletion involving FOXP1 on chromosome 3p14.1 was identified in an individual with features of idiopathic autism (full-scale IQ = 75). An additional 6.6 Mbp de novo deletion overlapping FOXP1 was also identified in an individual with idiopathic ID (Figure 4). A review of the DECIPHER database revealed a similar-sized deletion disrupting FOXP1 in an individual with developmental delay, sensorineural deafness, hypotonia, club foot, and dislocation of hip. Recently, FOXP1 was implicated in autism, ID, and language impairment [35], [36], [37]. It is believed that FOXP1interacts with FOXP2 and CNTNAP2, both implicated in speech disorders and autism [38], [39]. The overlapping 1.16 Mbp region of the deletion common to both autism and ID indicates a potential involvement of FOXP1 in pathways related to both of these disorders. Other variants involving functionally relevant genes include 7q36.2 deletion and duplication (DPP6), 17q23.3 duplication (SCN4A), and 17q21.32 duplication (WNT3 and WNT9B).

Our analysis of the ID cohort was enriched for singleton events often involving genes related to developmental or neurological functions including SYNPR, GABRA, AUTS2, FOXP1, FKBP6, COBL, and FMR1. However, within the same cohort we were also able to detect novel overlapping deletions (3.5 Mbp and 9 Mbp) on chromosome 9p24 in two unrelated individuals (Figure 5A). Both cases exhibited clinical features of ID and Pervasive Developmental Delay-Not Otherwise Specified. The distal breakpoints of these deletions map to segmental duplications while the proximal end maps within a high density of repeat elements. A survey of this region in the DECIPHER database [40] revealed about 15 cases with overlapping deletions. Variable clinical presentations and heterogeneity of deletion breakpoints preclude further genotype-phenotype correlation studies for this region (Figure S2). We also identified a nonrecurrent 6q16 deletion (chr6: 100,383,567-103,310,184) that potentially narrows the critical region for this recently described Prader-Willi-like syndrome [41] to approximately 2.9 Mbp. The refined critical region contains only five genes including the obesity-associated SIM1 [42] and the autism-associated GRIK2 [43] (Figure 5B). About 70% of children with 6q16 deletion manifest obesity [41]; however, our case with the smaller deletion, encompassing SIM1, showed no evidence of obesity at 10 years of age (Table S10).

Multiple large CNVs associate with phenotypic severity

We find that 8/69 (11.6%) cases in the ID cohort carried more than one large, rare CNV and all of these individuals presented with severe clinical features (Tables S6, S7, S8). A striking difference (p = 0.008, odds ratio = 11.2) in multiple CNV rates was also observed when the ID cohort was divided into those with severe MCA (44%) and those with idiopathic ID (6.7%). Notable examples are co-occurrences of a 3.4 Mbp 16p13.11 duplication and a 3.3 Mbp deletion on chromosome 4q25 involving PITX2 in a case with features of Rieger syndrome [44] (Figure 6) and a 17p13.3 deletion (YWHAE, Miller-Dieker syndrome) and 3q29 duplication (DLG1) in a child with cryptorchidism, ventricular septal defect, and seizures. The 3q29 duplication is a recurrent interstitial rearrangement [45], [46] potentially mediated by flanking segmental duplications of high sequence identity (27 kbp size, 96% identity). The 17p13.3 deletion is a previously reported nonrecurrent rearrangement associated with Miller-Dieker syndrome [47], [48]. In contrast, only 1/35 (2.9%) autism cases and none of dyslexia individuals carried another large CNV. This observation suggests that the severity of the phenotypes can be influenced by more than one large, rare CNV co-occurring in the same individual. During this analysis we considered the possibility of a derivative chromosome representing an unbalanced translocation possibly creating the impression of multiple CNVs in our cases. We carefully reviewed available chromosomal analysis data (G-banded karyotyping or FISH) for each of the individuals with two hits reported in our study. We did find one case with two hits where apparent CNVs represent a derivative chromosome inherited from a balanced translocation carrier parent (Table S8D).

Discussion

Initial discoveries of significant enrichment of rare CNVs for ID and autism led to testing the CNV basis for other behavioral and neurodevelopmental disorders of varying population frequency and severity, such as schizophrenia, ADHD, epilepsy, bipolar disorder, and Tourette syndrome. However, comparisons between these studies have been difficult due to differences in study design, insufficient sample sizes, and lack of detailed phenotype information. In this study, we compared 1,564 individuals (cases and controls) on a single platform of relatively modest density with the same type of detection bias. We utilized the duplication architecture of the human genome to custom design a DNA oligonucleotide microarray enriched for genomic hotspots, i.e., regions flanked by high-identity segmental duplications. This array has an advantage over several other commercial arrays in that there is a 25-fold enrichment for recurrent events in the genomic hotspots compared to the rest of the genome [49]. Therefore, fewer samples are required to identify several unrelated individuals with the same pathogenic mutation. We find that our array has a comparable diagnostic yield of 16% for the ID cohort compared to other clinical chromosomal microarray studies reported in the literature (Figure S3).

In strong agreement with previous studies, our data suggest that multiple, rare CNVs contribute to the etiology of autism and ID. In contrast, we find no increase in large pathogenic CNVs in individuals with dyslexia compared to controls. Notwithstanding, our analysis revealed novel regions of potential relevance to the etiology of dyslexia. Two unrelated children (2/322, 0.6%) with dyslexia carried CNVs encompassing AUTS2, both inherited from a parent. While the phenotype of dyslexia segregated with the AUTS2 duplication in the first family (Figure 3A), in the second family the deletion was inherited from affected grandmother through unaffected father (Figure 3B). This could be due to a phenomenon described as “compensation”, where some adults that reported difficulties with reading in childhood no longer evidence signs of dyslexia [50]. Previous studies of de novo chromosomal translocations and inversions identified breakpoints within AUTS2 in individuals with autism and/or ID phenotypes [51], [52], [53], [54], [55]. More recently, unique AUTS2 deletions and duplications were observed in Juvenile Myoclonic Epilepsy [10] and ADHD [11], [12]. It is interesting to note that ADHD and dyslexia are frequently comorbid and may have shared genetic risk factors [56], [57], [58]. When our study is taken together with recent CNV studies of ADHD [11], [12], AUTS2 CNVs were observed in 5/2,306 combined cases and 3/46,947 unscreened controls (p = 1.12×10⁻⁵, odds ratio = 33.9), indicating that AUTS2 might have an important role in pathways related to cognition. While the function of AUTS2 is still unclear, it is strongly expressed in fetal and adult brains, particularly in the frontal, parietal, and temporal lobes [59]. Interestingly, AUTS2 and the 7q11.2 region were identified as having the strongest statistical signal for positive selection in early modern humans as compared to the Neanderthal genome [60], indicating that AUTS2 might be important for a specialized human function such as cognition.

The CNV profile we observed in individuals with dyslexia was essentially the same as that in control individuals. This is not surprising if we take into consideration that all the subjects in our dyslexia sample had a VIQ above the 25%ile and the mean VIQ of the cohort was 110 (2/3 standard deviations above the general population mean), and given that we have shown that the CNV profile correlates with the severity of ID. The genes involved in dyslexia are likely to affect more specialized cognitive functions, may not adversely affect general intelligence, and may be more amenable to discovery with high-density arrays capable of detecting single gene or single exon CNVs or SNP microarrays that can leverage SNP allele frequency information in addition to signal intensity. In addition, all hybridization-based platforms fail to detect copy number neutral changes, such as balanced chromosomal rearrangements and inversions. This is particularly germane to dyslexia where a large number of candidate genes have been identified through mapping of translocation breakpoints [21], [34].

A comparison of rare de novo CNV rates for autism shows that our estimates (4%, 14/336) fall within a range of 4–10% reported previously by other large-scale, high-density array studies [4], [61], [62], [63]. This suggests that no platform-specific bias exists for large variants and also that the contribution of large CNVs is consistent across all studies for autism. We find a significantly greater enrichment for large CNVs, higher de novo rates, and a higher frequency of two rare CNV hits in individuals with ID-associated phenotypes compared to autism or dyslexia. This observation is exemplified by the fact that individuals with autism with ID have more large CNVs than those with autism only. We also find a significant difference between individuals with autism versus those with dyslexia. Sanders and colleagues recently analyzed 1,124 SSC families affected with autism spectrum disorder (ASD). Using stepwise linear models, they evaluated the relationship between intellectual functioning, sex, and the number of genes within rare, de novo CNVs. While the number of genes affected correlated with the size of the de novo CNV, the authors did not find a strong correlation of the Autism Diagnostic Observation Schedule (ADOS) combined severity score (p = 0.25, R² = 0.005) or full-scale IQ (p = 0.02, R² = 0.08) with the size of the CNV. In contrast, we considered all large CNVs (common and rare, de novo and transmitted) identified in a relatively smaller sample size and essentially bifurcated the autism cohort using a full-scale IQ score cutoff of 70. There was also a greater enrichment of two hits in the ID cohort (11.6%) compared to the autism cohort (2.8%). In fact, one individual carrying a 16p11.2 deletion with autism and features of ID also has a maternally inherited 22q11.2 duplication (TBX1) providing further evidence for the two-hit hypothesis we previously proposed for severe developmental delay [64]. Further, the frequency of two hits was even more striking when only individuals with ID/MCA were considered (44%), albeit the number of cases is few. We believe these data provide support for an incremental effect of CNV size and number on the severity of phenotypic outcome.

Our experimental design is biased towards interrogating hotspot regions in the human genome. A comparison to recently reported studies [61], [63] suggests that the majority of false-negative calls will reside within non-hotspot regions due to a lack of probe coverage (<10 probes). While the detection power of our array increases with the size of the variant, we would certainly miss smaller and intragenic CNVs, for example in autism candidate genes such as NRXN1 [65], [66], CACNA1C, SLC4A10, MAGI1 [63], SYNGAP1, DLGAP2 [62], NLGN1, ASTN2 [67], and exonic copy number variants in ASPM, DPP10, CNTNAP2, A2BP1, PCDH9 [68], and PTCHD1 [69]. While we find no excess of large CNVs in dyslexia, there is still the possibility that large CNVs are relevant in some familial cases of the disease as well as occasional sporadic cases. Further studies are warranted for a more detailed analysis of all the three neurodevelopmental cohorts using high-resolution arrays and next-generation exome and/or whole-genome sequencing. While it can be difficult to compare data derived from different microarrays, there is value to multiple array platforms and cross-platform validation. The depositing of the resulting data into publicly available databases will facilitate the continued elucidation of recurring clinically significant CNV and genotype-phenotype correlations.

Materials and Methods

Ethics statement

Patients from each of study cohort were recruited after appropriate human subjects approval and informed consent. Informed consent was also obtained to publish photographs.

Patient ascertainment based on severity of phenotypes

DNA samples were obtained from cases ascertained for three neurodevelopmental disorders of varying severity: (1) ID/developmental delay and MCA, (2) dyslexia or reading impairment, and (3) idiopathic autism. We defined severity of clinical features based on presence or absence of ID (IQ<70) for the autism group and congenital malformation for the ID group. Our dyslexia cohort had no ID or congenital malformation cases; as an IQ≤90 and the presence of congenital malformations were exclusion criteria. Individuals with idiopathic autism were partitioned into those with autism and ID (IQ<70) and those without ID (IQ>70). For the ID cohort, those individuals with brain malformations, gross craniofacial dysmorphology, cardiac defects, and neurological deficits were separated into an ID plus MCA (ID/MCA) group. Thus, in the order of severity, the ID/MCA cohort is considered the most severe, followed by ID only, autism with ID, autism without ID, dyslexia, and normal controls. However, we note that although the individuals with dyslexia do not have ID, they have severe impairments in core phonological measures leading to significantly reduced reading abilities despite normal IQ (IQ≥90). Detailed descriptions of each of the cohorts are given below.

Ascertainment of individuals with dyslexia or reading disability

For the dyslexia subject set, children were considered eligible for the study if they met researcher-defined criteria based on test scores from a standardized battery of tests. DNA samples were obtained from two cohorts. The first cohort included probands aged 6 to 16 from 198 families who were initially ascertained at the University of Washington (UW) multidisciplinary Learning Disability Center (UWLDC) under protocols approved by the UW Institutional Review Board. For the UWLDC cohort, probands were required to have a prorated VIQ at or above 90 (≥25%ile) on the Wechsler Intelligence Scale for Children – 3^rd edition [13], with performance below the age-specific population mean and at least one standard deviation below the VIQ on one or more out of 10 research measures of reading, writing, or spelling. As a group, on average, probands met the impairment criteria between 6 and 7 measures. As expected by ascertainment requirements, the average VIQ of probands in this cohort was 110 (≥75%ile) [70], [71]. Siblings older than 6.5 years were invited to participate, and additional family members were added using a sequential sampling strategy to extend pedigrees through family members with the most extreme impairment values on the same 10 research measures. Detailed recruitment and evaluation procedures for the UWLDC cohort were described earlier [50], [72] (see Table S4A).

For the second cohort, 178 children aged 5 to 12 were recruited from a special K-6 school for students with dyslexia or via their direct relatives in the Atlanta area (The Schenck School, Atlanta, GA). For this cohort, children were required to have a psychological battery of tests completed by a licensed psychologist and usually have a diagnosis of a reading disability. Based on strong verbal comprehension score, perceptual reasoning score, Peabody picture vocabulary test, or other cognitive tests that measure intelligence, these children have average to above-average intelligence. Both cohorts were composed of individuals with >90% Caucasian ethnicity with an approximately equal number of males and females. Except for ADHD, children with other psychiatric and neurological disorders, moderate to severe receptive language disorders, developmental disabilities, or other conditions known to affect cognition were excluded based on parental questionnaire. Clinical details are shown in Table S4B.

Ascertainment of individuals with features of autism with or without ID

For the autism cohort, families were identified through the SSC (www.sfari.org) [73]. The Simons Foundation-funded SSC includes families with no more than one child with autism ascertained through 12 data collection sites across North America. Of the 350 individuals included in this study 297 (85%) are of Caucasian ethnicity. Inclusion criteria in the collection requires that the child with autism meet ASD criteria on the ADOS [74], on the Autism Diagnostic Interview, Revised (ADI-R) [75], and meet expert clinical judgment. Nonverbal IQ estimate must also be greater than 35. Children with significant hearing, vision, or motor problems, significant birth complications (e.g. extended NICU stay), or with a diagnosis of ASD-related disorders, such as Fragile X, were excluded. Children with a relative (up to third degree) with ASD or sibling who showed ASD-related symptoms were also excluded. Diagnostic evaluations, cognitive assessment, and phenotypic characterization were conducted at each site with data collection, data entry, and data validation methods standardized across sites to ensure reliability of sample collection. We further partitioned the autism cohort into those associated with ID (average full scale IQ = 49) consisting of 97 cases (73 males, 24 females; median age, 12 years) and those without ID (average full scale IQ = 98.9) comprising 253 cases (228 males, 25 females; median age, 11 years and 11 months). Clinical details are shown in Table S5.

Ascertainment of individuals with intellectual disability with or without congenital malformation

The idiopathic ID cohort was selected from individuals admitted to the IRCCS Associazione Oasi Maria Santissima and screened for ID according to the Diagnostic and Statistical Manual of Mental Disorders-IV-Text Revision (DSM-IV-TR) criteria. This cohort consists of 428 cases (153 females, 275 males; median age, 15 years) of Caucasian ethnicity with idiopathic ID and previously excluded for common causes of ID, including Fragile X syndrome, trisomies 21 and 13. In addition, classical genomic syndromes such as Smith-Magenis, DiGeorge, Prader-Willi/Angelman, and Williams syndromes, if recognized by clinical evaluation, were followed up for confirmation using targeted multiplex ligation-dependent probe amplification and excluded. We note that cases with phenotypic variability that escape clinical detection might not have been excluded. Typically, idiopathic cases of ID with no classical constellation of clinical features suggestive of a known disorder or those with mild to moderate ID without significant congenital malformation were included in this cohort. Clinical details are shown in Table S6.

Individuals with features of ID with MCA not necessarily assigned to a specific syndrome were evaluated and recruited at the University of Torino. This cohort consists of 73 individuals (32 females and 41 males) of Caucasian ethnicity with a median age of 2 years at diagnosis. Clinical features of these individuals included brain malformations, craniofacial dysmorphology, and neurological deficits along with variable ID (Table S7). Informed consent was obtained from all the subjects included in both the studies.

Ascertainment of normal controls

The control cohort consisted of 337 DNA samples obtained from the Rutgers University Cell and DNA Repository (www.rucdr.org). These individuals were ascertained by the NIMH Genetics Initiative [76] through an online self-report based on the Composite International Diagnostic Instrument Short-Form (CIDI-SF) [77] and screened specifically for eight mental health disorders, including major depression, bipolar disorder, and psychosis, but were not screened for dyslexia and therefore not ideal for such comparisons. Those who did not meet DSM-IV criteria for major depression, denied a history of bipolar disorder or psychosis, and reported exclusively European origins were included [78], [79].

Additionally, CNV data from 8,329 additional cell line and blood-derived controls were used to assess the frequency of our putative pathogenic CNVs in a larger population of neurologically normal individuals. These data were derived primarily from genome-wide association studies of non-neurological phenotypes. Although these data were not ascertained specifically for neurological disorders, they consist of adult individuals providing informed consent. Specifically, datasets from the following sources were included in our analysis: Human Genome Diversity Project [49], [80]; National Institute of Neurological Disorders and Stroke (NINDS) (dbGaP accession no. phs000089) [49], [81]; Pharmacogenomics and Risk of Cardiovascular Disease (PARC/PARC2) [82], [83]; parents of asthmatic children courtesy of Stephanie London [49]; Fred Hutchinson Cancer Research Center (prerelease data provided courtesy of Aaron Aragaki, Charles Kooperberg, and Rebecca Jackson as part of an ongoing genome-wide association study to identify genetic components of hip fracture in the Women's Health Initiative); InCHIANTI (data provided by InCHIANTI study of aging, www.inchiantistudy.net) [49], [84]; and the Wellcome Trust Case Control Consortium phase 2 (National Blood Service) [7]. All samples were genotyped on Illumina arrays using methodology described previously [49] [85] and either natively processed in hg18 or re-mapped after CNV calling (NINDS and PARC) to hg18 using the UCSC LiftOver tool (http://genome.ucsc.edu).

Array CGH and analysis

We designed custom targeted hotspot v1.0 arrays comprised of 135,000 probes (by Roche NimbleGen) with higher density probe coverage (median probe spacing 2.6 kbp) in the genomic hotspots (regions flanked by segmental duplications) and a lower probe density in the genomic backbone (median probe spacing 36 kbp). All microarray hybridization experiments were performed as described previously [86], using a single unaffected male (GM15724 from Coriell) as reference. All validation experiments were performed using two custom array designs: (1) a custom targeted 4×180 K Agilent chip with median probe spacing of 2 kbp in the genomic hotspots and whole-genome backbone coverage of one probe every 36 kbp (Agilent Technologies) and (2) a custom targeted 3×720 K NimbleGen or 2×400 K Agilent chip with median probe spacing of 500 bp in the genomic hotspots and probe spacing of 14 kbp in the genomic backbone.

All arrays were analyzed by mapping probe coordinates to the human genome assembly Build 36 (hg18). Using chromosome-specific means and standard deviations, normalized log intensity ratios for each sample were transformed into z-scores. These z-scores were then classified as “increased”, “normal”, or “decreased” in copy number using a three-state HMM. The HMM was applied using HMMSeg [87]. For each sample, HMM state assignments of probes were merged into segments if consecutive probes of the same state less than 50 kbp apart. If two segments of the same state were separated by an intervening sequence of ≤5 probes and ≤10 kbp, both segments and intervening sequence were called as a single variant. Further, we employed stringent QC measures and empirically estimated post-HMM filtering thresholds (absolute z-score >1.5 and >10 probes) to increase the specificity of our experiments. With these filtering criteria, we were able to thoroughly scan HMM outputs for CNV events and manually check the validity of each call by examining the normalized log intensity ratios across a chromosome. For the Agilent arrays, data analysis was performed following feature extraction using DNA analytics with ADM-2 setting according to the manufacturer's instructions. All CNVs calls were visually inspected in the UCSC genome browser.

First, we carried out validation on 24 samples from the developmental delay cohort and confirmed 117/118 HMM-inferred calls with a validation rate of 99.15%. Next, we validated 84 CNVs from an independent set of cases both validated using fluorescence in situ hybridization (FISH) and locus-specific custom high-density arrays [64] (Girirajan and Eichler, unpublished). We also validated all 44 calls from the autism cohort, 22 calls from the dyslexia cohort, and 78 calls from the developmental delay cohort. In addition, in an analysis of 517 individuals with epilepsy using this array design, 61/63 CNVs were validated on a different array platform [10].

Supporting Information

Attachment 1

Attachment 2

Attachment 3

Attachment 4

Attachment 5

Attachment 6

Attachment 7

Attachment 8

Attachment 9

Attachment 10

Attachment 11

Attachment 12

Attachment 13

Attachment 14

Attachment 15

Attachment 16

Attachment 17