Characterization of intellectual disability and autism comorbidity through gene panel sequencing

2.1 Patient selection

Patients were referred from clinical geneticists of 17 Italian public hospitals with a diagnosis of nonspecific neurodevelopmental disorder. Clinical data were collected with a standardized clinical record describing clinical and family history, clinical phenotype (auxological parameters, neurological development, physical features, and behavioral profile), and presence of associated disorders. Data from neurophysiological profiles, electroencephalograms (EEG) and brain magnetic resonance imaging (MRI) were also collected. Table 1 summarizes the clinical data of the patients, while Table S1 reports for each of 150 patients the presence of ID, ASD, epilepsy, microcephaly or macrocephaly, hypotonia, and ataxia. Written informed consent was obtained from the patient's parents or legal representative. This study was approved by the Local Ethics Committee, University-Hospital of Padova, Italy.

2.2 Gene panel selection

For the construction of an efficient and low-cost gene panel, we selected the most promising ID and/or ASD genes gathering data from public databases (AutismKB, http://autismkb.cbi.pku.edu.cn, and SFARI, https://sfari.org/resources/sfari-gene), OMIM, and PubMed. Candidate genes were extracted in particular from recent exome sequencing and meta-analysis studies (Table S2). We collected a list of 972 genes scored according to recurrence in different sources, annotated for clinical phenotype, gene function, subcellular localization, and interaction with other known causative genes. Separated lists were generated considering ASD or ID association. The extracted information was stored in a dedicated SQL database used in conjunction with the disease network construction. Using data from STRING 9.0 (Franceschini et al., 2013), a disease protein–protein interaction (PPI) network was built starting from 66 high confidence genes (intersection list), shared by both ASD and ID gene lists. Emerging features of the network were assessed by enrichment analysis with Enrich web server (Kuleshov et al., 2016). The same list was used as a training set for Endeavor gene prioritization (https://endeavour.esat.kuleuven.be/; Tranchevent et al., 2016). Hub direct interactors (i.e., genes with STRING degree score above 0.45) belonging to the top ranking prioritized list, but not included in the intersection, were also included in the most promising candidate gene list. To this list, we also added the top-ranked genes associated with ID or ASD only (i.e., genes with at least five evidence for ID or ASD; Figure S1). The final panel set resulted in a manually curated list of 74 genes, comprising selected known causative genes, top-ranked genes by gene prioritization, and genes meeting PPI network parameters (Table S3).

2.3 Gene panel sequencing

Nucleic acids were extracted from blood samples using the Wizard genomic DNA Promega Kit (Promega Corporation). Multiplex, polymerase chain reaction (PCR)-based primer panels were designed with Ion AmpliSeq™ Designer (Thermo Fisher Scientific) to amplify all exons and flanking regions (10 bp) of the 74 selected genes (Thermo Fisher Scientific). Template preparation and enrichment were performed with the Ion One Touch 2 and Ion One Touch ES System, respectively (Thermo Fisher Scientific). Read alignment to the human genome reference (hg19/GRCh37) and variant calling were performed with the Ion Torrent Suite Software v5.02 (Thermo Fisher Scientific).

2.4 Variant filtering

An in house pipeline was built to create a database of genetic variants identified in our cohort and to annotate them with features provided by ANNOVAR, that is allelic frequency (AF) in control cohorts, variant interpretation from InterVar automated, ClinVar report, pathogenicity predictions, and conservation scores. Detected variants were ranked for their frequency in the gnomAD (Lek et al., 2016), 1000G (Genomes Project Consortium et al., 1000, 2015), and ExAC (Kobayashi et al., 2017) databases, as well as in our in house database of 150 patients. We excluded single nucleotide variants (SNVs) found more than twice in our cohort or reported with an AF higher than expected for the disorder, which has been calculated to be <0.002% and <0.45% for autosomal dominant and recessive genes, respectively (Piton, Redin, & Mandel, 2013). Variants reported as risk factors for autism in the literature were nevertheless considered for further segregation analysis even if their frequency in the control populations exceeded the ID incidence. Rare variants were ranked for their pathogenicity prediction considering the consensus among 12 computational methods. ANNOVAR provides predictions from SIFT (Sim et al., 2012), the Polyphen-2 (Adzhubei, Jordan, & Sunyaev, 2013) HDIV and HVAR versions, LRT (Chun & Fay, 2009), Mutation Taster (Schwarz, Cooper, Schuelke, & Seelow, 2014), MutationAssessor (Reva, Antipin, & Sander, 2011), FATHMM (Shihab et al., 2014); PROVEAN (Choi & Chan, 2015), MetaSVM (Dong et al., 2015), MetaLR (Dong et al., 2015), M-CAP (Jagadeesh et al., 2016), fathmmMKL (Shihab et al., 2013, 2014) and CADD (Kircher et al., 2014). Conservation was evaluated with the scoring schemes GERP++ (Davydov et al., 2010), PhyloP (Pollard, Hubisz, Rosenbloom, & Siepel, 2010) and SiPhy (Garber et al., 2009). Intronic or synonymous variants near the exon-intron junction were also evaluated in silico for their impact on splicing using Human Splicing Finder (Desmet et al., 2009). The Integrated Genome Viewer platform (Robinson et al., 2011) has been used to exclude sequencing or alignment errors around selected SNVs.

2.5 Variant validation and functional assays

Selected variants were validated by Sanger sequencing. Segregation analysis was performed in the patient relatives when DNA samples were available. For apparent de novo variants, paternity and maternity were confirmed by the inheritance of rare detected variants in parental samples. In other cases, pedigree concordance was checked using polymorphic microsatellite markers of chr15q described in (Giardina et al., 2008). For maternally inherited X-linked variants, the X-inactivation pattern of the mother was evaluated on the highly polymorphic androgen receptor (ARA locus) at Xq11-q12, as described in (Bettella et al., 2013). The X-inactivation was classified as random (ratio < 40:60) or significantly skewed (ratio ≥ 80:20).

Analysis of the transcripts was performed to confirm putative splicing variants. RNA was extracted from patient peripheral blood leukocytes and reverse-transcription polymerase chain reaction was performed using random primers. cDNA was used as a template in nested PCR reactions with specific primers to amplify the regions containing the mutation. PCR products were tested on 1.5% agarose gel and sequenced.

2.6 Variant classification

A clinical interpretation of selected variants was first evaluated using InterVar (Li & Wang, 2017), providing an automated variant interpretation based on 18 criteria published by the American College of Medical Genetics and Genomics (ACMG). The InterVar web server has a manual adjustment step that allows reviewing the automated interpretation by selecting appropriate criteria according to additional information and knowledge about the involved domain.

According to the ACMG recommendations, we classified filtered candidate variants into five categories (pathogenic, likely pathogenic, uncertain significance, likely benign, benign) based on multiple lines of evidence (conservation, allele frequency in population databases, computational inferences of variant effect, mode of inheritance, X-inactivation pattern, and disease segregation; Figure 1). Assuming that a patient phenotype is consistent with a Mendelian disorder, variants be classified as pathogenic are considered causative, that is, responsible for the phenotypic manifestations of the carrier patient. Likely pathogenic variants instead require further investigation to be classified as pathogenic/causative variants, for example, segregation analysis and/or functional analysis. Rare or novel variants predicted as pathogenic and altering genes conferring an increased risk of autism have been classified as possible contributing factors if inherited from parents reported as healthy, as they alone are not sufficient to cause the disease (Table S5). The criteria used to classify the variants (Richards et al., 2015) are reported for both causative and likely pathogenic variants ( Table 2 and 3). All the causative and likely pathogenic variants have been submitted to the LOVD database.

2.7 In silico analysis of candidate variants

Canonical protein sequences were retrieved from UniProt (Apweiler et al., 2004) and protein domains predicted by InterPro (Mitchell et al., 2018). To evaluate conservation, orthologous sequences were downloaded from OMA Browser (Schneider, Dessimoz, & Gonnet, 2007) and aligned with MAFFT (Katoh & Standley, 2013). When available, crystal structures were retrieved from PDB (Rose et al., 2017). Structures of protein domains were modeled with MODELLER (Alva, Nam, Söding, & Lupas, 2016) (automatic best template selection), using templates predicted by HHpred (Alva et al., 2016). Structure of the CASK L27 domain and its complex with SAP97 have been analyzed using Mistic2 (Colell, Iserte, Simonetti, & Marino-Buslje, 2018) to evaluate covariation between residues. Structures were manually explored with Pymol (Janson, Zhang, Prado, & Paiardini, 2017) or UCSF Chimera (Pettersen et al., 2004).

Disorder content and the presence of short linear motifs for protein interactions were assessed combining MobiDB (Potenza, Di Domenico, Walsh, & Tosatto, 2015) and ELM (Gibson, Dinkel, Van Roey, & Diella, 2015), using the interactive exploration tool ProViz (Jehl, Manguy, Shields, Higgins, & Davey, 2016).

3.1 Gene panel description

The computational approach adopted to select panel genes includes genes recurrently mutated in ID or ASD conditions, genes shared among ID and ASD disease networks, and genes directly connected to ID/ASD. Of the 74 selected genes, 42 are FMRP targets, 21 postsynaptic proteins, and 16 chromatin modifiers. The majority of the selected genes are associated with autosomal dominant diseases; of these, DEAF1, PTEN and RELN are also associated with a recessive disorder. Moreover, the panel includes 8 genes associated with autosomal recessive disorders and 19 genes associated with X-linked diseases. Twenty genes are associated with nonspecific ID, while 40 genes are responsible for defined genetic syndromes. Seven genes have been found to confer autism susceptibility (SHANK2, CNTNAP2, RELN, CHD8, NLGN3, NLGN4X, and PTCHD1). At the time of panel design, for 11 of the selected genes there was only scant evidence in the literature about their association with ID and/or ASD (ASH1L, NTNG1, KATNAL2, MIB1, MTF1, MYH10, PTPN4, TANC2, TBR1, TRIO, and WAC). Two more genes associated with an OMIM ID have been meanwhile reclassified and the disease association confirmation is pending (HDAC4 and KIRREL3; Table S3).

3.2 Data output and quality

Our strategy allowed generating 263 reads per amplicon with 97% of them on target. The mean target region coverage for all patients ranged between 67 and 612. About 94% of target regions were covered at least 20x and 93% of amplicons had no strand bias. The uniformity of amplicon coverage resulted being 89%. A small proportion of targeted regions was weakly covered (<20×) throughout all patients. These are mainly first exons or GC-rich regions representing a well-known burden in an amplicon-based strategy. Specifically, exon 4 of MECP2, exon 5 of ARX, exon 1 of FMR1, part of exon 21 of SHANK3 have a read depth <10×. Depending on phenotype manifestations, otherwise, these regions have been covered by Sanger sequencing, as they could not be analyzed reliably.

3.3 Cohort description and diagnostic yield

Our cohort of 150 individuals is enriched in males (61%). Sporadic cases of 81.3% and the remaining had a family history of neurodevelopmental disorders with siblings affected in 3.3% of cases. At the time of molecular testing, the age of the patients was ranging between 2 and 42, with a median of 11 years. Although the vast majority of patients have ID, for four individuals clinicians did not report information about the presence of an intellectual defect. In 38 cases, the level of cognitive impairment has not been evaluated. Among the patients with a cognitive impairment evaluation, a slightly higher proportion have a moderate form (25.3%) than severe (22%) and mild ones (24.6%). Ninety-three patients (62%) had both ID and ASD. Fifty-three (35.3%) had the only ID, while four (2.6%) had ASD and no information was provided about the presence of ID. Epilepsy was reported for 55 (36.6%) individuals, of which 6 had early onset epilepsy (<24 months). Thirty-nine (26%) patients with ID and ASD present also epilepsy (Table 1). MRI and EEG abnormalities were reported, respectively, for 37.3% and 24.6% of the sequenced individuals (Table 1). In most of the patients, structural pathogenic alteration of chromosomes was excluded by FISH, karyotype, or aCGH; however, in 17 of these, the aCGH analysis revealed a CNV inherited from unaffected parents or involving gene-poor regions (Table S4). Eighty-six patients had a negative Fragile-X test, and 57 resulted negative to other single gene tests, such as MECP2, CDKL5, UBE3A, and/or to the Chr15q methylation test.

3.4 Variant detection and prioritization

In coding exons or exon-intron boundaries regions of the 74 genes, we detected on average about 74 SNVs, with a range of 60–80 SNVs per patient. Overall, 202 coding or splicing SNVs passing the quality control (GQ > 30, DP > 20) and frequency filters (MAF < 1%), were observed only once in our cohort. Based on the prioritization criteria described in the Materials and Methods section, we selected about 170 variants for further analysis, 47 of which were absent from the general population (Table S5). We detected certainly causative mutations in 26 patients, leading to an overall diagnostic yield of 17.3% for the entire cohort. These rare or novel variants were predicted pathogenic by several approaches, found to be de novo or inherited from affected parents and consistent with the expected disease for the respective gene (Table 2). In other 15 patients, we identified 17 rare variants (missense, synonymous or splicing) with a putative although not established pathogenic role (Table 3). For these variants, either there were no family members available for segregation analysis, or the clinical features did not fit those that would have been expected for the respective gene. With the availability of functional studies or further sequencing data, we expect that a number of these variants will turn out to be benign, but some variants might be proven pathogenic. In particular, five of these putative pathogenic variants meet most of the ACMG criteria (p.Tyr2418Phe in ATRX, p.Gly1465Arg in CREBBP, c.870+1G>A in DEAF1, p.Val821Leu in GRIN2B, and p.Pro576Gln in SHANK2) and thus with segregation analysis we would be able to assign their causative role on the proband's phenotype. For two other cases, MR1769.01 and MR2053.01, either the digenic or autosomal recessive transmission are possible, only with functional assays supporting the pathogenic effect of the identified variants. In addition to a novel missense mutation in the FOXP1 gene, MR1769.01 carries a rare variant in CNTNAP2, which is transcriptionally regulated by FOXP1. As previously proposed by O'Roak and collaborators, we hypothesize a two-hit model for the disease risk in this patient, where a mutant FOXP1 protein leads to an amplification of the deleterious effects of p.Asp417Tyr in CNTNAP2 (O'Roak et al., 2011). In MR2053.01, we detected two variants in the GAD1 gene, which is associated with an autosomal recessive phenotype. The girl presents with clinical features consistent with the few reported GAD1 cases, for example, severe developmental delay and nonprogressive ataxia. The paternally inherited missense p.Leu284Phe maps on the pyridoxal 5′-phosphate (PLP) transferase domain and is predicted as pathogenic by several computational methods, while the maternally inherited c.83–3T>C is predicted to alter splicing mechanisms. However, we were not able to demonstrate its pathogenicity by qualitative analysis of the transcript extracted from the mother.

After segregation analysis, we were able to classify about one hundred of the selected variants as likely benign. The majority of these variants were found in genes associated with highly penetrant disorders inherited from apparently healthy parents (n = 85) or found in individuals with a causative mutation in another gene consistent with the proband phenotype (n = 7). Five variants in X-linked genes were inherited from a healthy father or the X-inactivation pattern did not support their pathogenic role (Table S5).

Finally, some rare or novel inherited variants with strong pathogenicity predictions were classified as possible contributing factors. These variants occur on genes known to confer autism susceptibility (e.g. SHANK2, RELN, CNTNAP2) or that have been reported in the literature in individuals with the very mild phenotype (e.g. TRIO and SLC6A1; Table S5).

3.5 Disease-causing variants

We identified likely causative variants in 26 patients (17%) and 18 different genes (Table 2). More than one mutation was found in seven genes (ANKRD11, CASK, EHMT1, GRIN2B, MECP2, SHANK3, and TRIO); SHANK3 resulted to be the most mutated gene. Among the identified variants, 6 were in X-linked genes and 12 in genes associated with autosomal dominant conditions. Twenty variants were absent from the gnomAD database, while five were known pathogenic mutations of the MECP2, SATB2, SHANK3, and GRIN2B genes.

Most of the mutations were de novo; one patient carried two of them. In 20 cases, paternity and maternity were established, while for two others, the parent DNA samples were not available. We can assume that the two truncating variants in ANKRD11 and MECP2 should be de novo, since they involve genes that are associated with highly penetrant disorders. Furthermore, p.Arg168X variant in the MECP2 gene is a recurrent pathogenic variant associated with Rett syndrome. We also identified four inherited causative variants, one maternally inherited variant in the TRIO gene, which is associated with an autosomal dominant disorder, and three maternally inherited variants in X-linked genes (ATRX, GRIA3, and RAB39B). X-inactivation analysis was consistent with the phenotype expression. The RAB39B variant is thought to be also responsible for the mild phenotype reported in the mother, clinically re-evaluated after the molecular finding.

3.6 In silico structural analysis may help predict mutation effects

Fifteen causative variants (one splicing, six frameshift, and eight stop codon variants) are predicted to result in truncated proteins if escaping the nonsense-mediated mRNA decay. The SHANK3 splicing variant has been shown to functionally impair mRNA splicing, producing an aberrant transcript containing an additional 77 bp intronic sequence (Li et al., 2018). Nine out of the 12 missense mutations were predicted pathogenic by the majority of computational methods, while pathogenicity predictions were discordant for three variants (Table 2). Among the nine variants with the strong prediction of pathogenicity, p.Arg696His in GRIN2B has previously reported as pathogenic and has been shown to alter the Agonist Binding Domain reducing channel activity (Swanger et al., 2016). In silico analysis of two other missense mutations, DYRK1A p.Lys174Asn and EHMT1 p.Gly1193Arg allowed us to classify them as loss of function mutations. Lysine 174 maps to the catalytic pocket of the DYRK1A kinase domain and alters the electrostatic surface of the domain, which is important for nucleotide binding (Figure 2). Glycine 1193 maps on the EHMT1 SET domain, which is necessary for methylation of lysine-9 in the histone H3 N-terminus, and is buried in the rigid structure of the domain. A substitution of Glycine 1193 with an arginine residue should result in the unfolding of the domain core (Figure 3).

We also hypothesized a predictive impact on the protein function of two variants with discordant pathogenicity: p.Phe193Leu in RAB39B (Figure S2) and p.Tyr387His in CASK (Figure 4). Phenylalanine 193 maps in the RAB39B hypervariable C-terminal tail, which mediates interactions with effector proteins for proper intracellular targeting (Chavrier et al., 1991). This nonconservative substitution may disrupt a functional motif involved in protein interaction and cause a mislocalization of RAB39B, as shown for the protein mutated at position p.Gly192Arg close by (Mata et al., 2015).

The Tyr387 residue maps on the N-terminal L27 domain, which mediates hetero- and homo-dimerization of the CASK protein. Comparing L27 domain sequence from different proteins, the 387 position is one of the most conserved and connected in the structure, denoting its important structural role (Figure 4a,b). However, at that position histidine is more frequent than tyrosine, thus the p.Tyr387His should not have a dramatic effect on the L27 domain structure (Figure S3). However, when considering more related homologs of human CASK, the tyrosine 387 is highly conserved, presenting mostly tyrosine or, in few cases, phenylalanine compared to other L27 domains; this might indicate that tyrosine is important for the protein function and that a change to histidine might be deleterious, particularly in human and nearest species. We hypothesize that this substitution may result in a hypomorphic mutant protein with a reduced ability to form dimers (Figure 4), which is consistent with the phenotype of the boy, hemizygous for the CASK mutation. Indeed, individuals carrying hypomorphic mutations have been reported with an X-linked intellectual disability with or without nystagmus and additional clinical features (Hackett et al., 2010).

3.7 Genotype–phenotype correlation

One of the major criteria to assign the pathogenicity of the variants is the patient phenotype correlation with the classical syndrome associated with the corresponding gene. Most of the identified causative mutations (n = 22) correlate with the previously reported phenotype of the corresponding gene (Table 2). For instance, the two known MECP2 mutations were found in two girls with a suspected Rett syndrome. The p.Arg168* variant found in MR414.01 patient was missed in previous single gene testing since it was in a mosaic state. The p.Arg294* variant found in MR2145.01 patient was identified by Sanger sequencing of the MECP2 region not covered by the 74-gene panel. A previous panel for Rett-Angelman spectrum disorders also missed the variant. In addition, for the two cases carrying a de novo mutation in the EHMT1 gene, the patient phenotypes were consistent with a Kleefstra Syndrome (KS). Both patients presented with core symptoms of the disease, including a moderate to the severe ID with absent speech, and hypotonia. However, a Kleefstra syndrome was suspected for the characteristic carp-shaped mouth, only in MR2243.01 carrying the frameshift mutation, while MR2166.01 was originally suspected to have Smith–Magenis Syndrome probably due to brachycephaly, which is also a common characteristic of KS.

Nonetheless, in other cases, the probands lacked some peculiar clinical features of the associated syndromes that prevented the geneticists to formulate a hypothesis about a suspected syndrome. As an example, the two patients carrying a truncating mutation in the ANKRD11 gene, both lacked, macrodontia of the upper central incisors, craniofacial features, and skeletal anomalies typical of the KBG syndrome, but both have a mild ID, behavioral issues, and hearing loss (Sirmaci et al., 2011).

On the other hand, in some cases, the detected mutations were found in genes never associated with particular phenotypic traits. In particular, we observed macrocephaly in three individuals carrying pathogenic mutations in three different genes (ATRX, GRIN2B, and TRIO) previously associated with microcephaly (Table 4). Interestingly, the de novo p.Arg696His in GRIN2B had been previously reported in a girl with a significant phenotypic overlap with our patient (MR 2019.01), including developmental delay, poor speech, ID, ASD with stereotypic behavior (Swanger et al., 2016). However, our patient presents macrosomia and marked macrocephaly. There is only one report in literature associating GRIN2B with macrocephaly, in a case with an ~2 Mb interstitial deletion in 12p13 involving the entire GRIN2B gene, in addition to other genes (Morisada et al., 2016). Furthermore, we observed that the other two cases presenting with macrocephaly (MR 2276.01 and MR 984.01), carrying the maternally inherited ATRX and the de novo TRIO mutations, also carried other rare inherited CNVs or sequence variants (Tables S4 and S5). Since inherited from healthy parents these alterations were classified as benign or of uncertain significance.

4.1 CAGI competition

At the completion of our study and before submitting for publication, genetic data and corresponding phenotypes of the tested individuals have been provided for a challenge at the Critical Assessment of Genome Interpretation (CAGI). CAGI is a worldwide blind test to assess computational methods for predicting phenotypic impacts of genomic variations. The outcome of the predictive competition using the ID-ASD gene panel data set is presented elsewhere. However, here we provide the original data with few updates that can be used to train and/or to test their own computational tools/approach aiming to predict comorbid phenotypes from genetic variants in a subset of NDDs genes.