Whole genome sequencing and integrative genomic analysis approach on two 22q11.2 deletion syndrome family trios for genotype to phenotype correlations

Abstract

The 22q11.2 deletion syndrome (22q11DS) affects 1:4000 live births and presents with highly variable phenotype expressivity. In this study, we developed an analytical approach utilizing whole genome sequencing and integrative analysis to discover genetic modifiers. Our pipeline combined available tools in order to prioritize rare, predicted deleterious, coding and non-coding single nucleotide variants (SNVs) and insertion/deletions (INDELs) from whole genome sequencing (WGS). We sequenced two unrelated probands with 22q11DS, with contrasting clinical findings, and their unaffected parents. Proband P1 had cognitive impairment, psychotic episodes, anxiety, and tetralogy of Fallot (TOF); while proband P2 had juvenile rheumatoid arthritis but no other major clinical findings. In P1, we identified common variants in COMT and PRODH on 22q11.2 as well as rare potentially deleterious DNA variants in other behavioral/neurocognitive genes. We also identified a de novo SNV in ADNP2 (NM_014913.3:c.2243G>C), encoding a neuroprotective protein that may be involved in behavioral disorders. In P2, we identified a novel non-synonymous SNV in ZFPM2 (NM_012082.3:c.1576C>T), a known causative gene for TOF, which may act as a protective variant downstream of TBX1, haploinsufficiency of which is responsible for congenital heart disease in individuals with 22q11DS.

Introduction

The 22q11.2 deletion syndrome (22q11DS), also known as velo-cardio-facial syndrome or DiGeorge syndrome (MIM#s 192430, 188400) occurs in 1:4000 live births and is caused by a similar sized 3 megabase (Mb) hemizygous deletion on chromosome 22q11.2 in the majority of patients (Morrow, et al., 1995; Shaikh, et al., 2000). One of the most striking features of 22q11DS is the highly variable clinical presentation or phenotype expressivity of affected individuals. A potential source of the variable expressivity could be genetic variants. It has been hypothesized that the 22q11.2 deletion could act to unmask recessive mutations or sensitize the population to genetic modifiers with small effects (McDonald-McGinn, et al.; Stalmans, et al.). A large proportion of individuals with 22q11DS have learning disabilities or psychiatric disorders. For example, approximately 25% of adults with 22q11DS have schizophrenia (SCZD; MIM# 181500), which is 30 fold higher than what occurs in the general population (Schneider, et al., 2014).

In addition to cognitive impairment and psychiatric disorders, over 70% of individuals with 22q11DS have congenital heart disease (CHD), primarily of the conotruncal type, including tetralogy of Fallot (TOF; MIM# 187500). TOF occurs in ∼35% of individuals with 22q11DS (Momma, et al., 1996) as compared to 5 in 10,000 live births in the general population. Mouse genetic studies identified Tbx1 (T-box 1; Entrez Gene ID: 21380), encoding a T-box transcription factor as one of the key genes in the 22q11.2 region as being responsible for CHD (Jerome and Papaioannou, 2001; Lindsay, 2001; Merscher, et al., 2001). However, mutation analysis of the remaining allele of TBX1 (MIM# 602054) in human subjects with varying cardiac defects failed to identify potential modifier alleles that could explain this variability (Guo, et al., 2011). This leaves open the possibility that genetic modifiers exist in other genes on 22q11.2 or elsewhere in the genome.

There are many additional clinically important features in 22q11DS that occur at a lower frequency (Cancrini, et al., 2014). Some anomalies of particular interest are immunodeficiency and autoimmune disease, including juvenile rheumatoid arthritis (JRA; MIM# 604302) (Gennery, 2013; Rasmussen, et al., 1996). JRA is seen in approximately 2% of 22q11DS patients, but 0.12% in the general population (Cobb, et al., 2014).

Personal genomics approaches to understand an individual's susceptibility to disease are now possible with next-generation sequencing methods. Whole genome sequencing (WGS) provides the potential to directly identify causative coding and non-coding variants. However, interpretation of the vast amount of data can be a limiting factor. This task can be aided by utilizing integrative genomic analysis methods, which combine knowledge from multiple data sources to identify potentially causative variants (Khurana, et al., 2013; Wang, et al., 2010). Integrative approaches have successfully been used to identify primary causative mutations in studies of various complex disorders including cancer and cardiovascular disease (Khurana, et al., 2013; Ware, et al., 2013). Most tools annotate coding variants with relatively few that are able to annotate non-coding genetic variations. However, recent efforts to characterize the non-coding genome, such as the ENCODE project and Epigenomics Roadmap Project have made significant progress towards identifying functional variants from the non-coding elements (Bromberg, 2013; Khurana, et al., 2013). These new tools and resources facilitate our ability to identify potential genetic modifiers in individuals with 22q11DS.

In this study, we performed WGS on DNA from two trios with 22q11DS in order to test integrative genomic methods are able to discover genetic modifiers that could explain variable phenotypes in these affected individuals. This work may lead the way to future larger studies to find genetic modifiers of this or other genomic disorders occurring with variable expressivity.

Materials and Methods

Subject recruitment and phenotype information

Informed consent for genetic analyses was obtained from all individuals in the study. The protocol was approved by the Internal Review Board at Albert Einstein College of Medicine (IRB: 1999-201-047). The two probands were diagnosed with 22q11DS after fluorescence in situ hybridization (FISH) testing for the 22q11.2 deletion. Complete family and medical history was obtained for two probands with de novo 22q11.2 deletions and their unaffected parents (Figure 1 and Supp. Table S1).

Figure 1.

Pedigree and family history for the P1 and P2 families. Probands (arrows) have 22q11DS (upper left quadrant solid fill). We performed WGS on the probands and parents (grey blocks). (A) Proband P1 presented with 22q11DS, tetralogy of Fallot (TOF), and schizophrenia. The P1 family pedigree shows a history of CHD and mental illness in the paternal branch. (B) Proband P2 presented with 22q11DS, juvenile rheumatoid arthritis. The P2 family pedigree shows no history of CHD. The paternal grandfather died of Alzheimer's disease and the paternal grandmother was diagnosed with breast cancer.

Proband 1 (P1) a now 32 year old male was referred for genetics evaluation at age 13, with characteristic clinical features of 22q11DS. He was the product of an uneventful pregnancy of a non-consanguineous couple. He had developmental delays in motor and speech milestones, learning disabilities, and attention deficit hyperactivity disorder (ADHD; MIM# 143465). At age 27, P1 developed psychotic symptoms, including visual hallucinations, and was diagnosed with other specified schizophrenia spectrum and other psychotic disorder (DSM-5 298.8) after review of clinical charts. He is currently being treated with venlafaxine and olanzapine. A SCID (Structured Clinical Interview for DSM Disorders) was performed in November of 2014 and the proband was diagnosed with clinical anxiety but did not present with psychosis, indicating a good response to anti-psychotic medication. A paternal great uncle had a ventricular septal defect and SCZD, while a paternal aunt died shortly after birth from pneumonia, and possibly a heart defect (Figure 1A). On physical exam at age 13, the proband was obese and had the typical facies seen in individuals with 22q11DS. A high arch and bifid uvula were medically confirmed. Fingers were tapered and the hands had multiple fine palmar creases with 5^th digit camptodactyly, as is typical with 22q11DS. He also had joint laxity. He had TOF, as diagnosed by echocardiogram, but no other cardiovascular malformations. Hypogonadism was also observed. Fluorescence in situ hybridization (FISH) confirmed a de novo 22q11.2 deletion (data not shown). Detailed phenotype information can be found in Supp. Table S1.

Proband 2 (P2) a now 15 year old female was referred for a genetics evaluation at age 8 from a craniofacial center and had a de novo 22q11.2 deletion as determined by FISH. She was the product of an uneventful pregnancy of a non-consanguineous couple. Family history was unremarkable except for unknown mental illness and cancer in the paternal grandfather and grandmother, respectively (Figure 1). She performed at an average level in elementary school. She had a history of hypernasal speech, velopharyngeal incompetence, occipitalization of the atlas and spina bifida occulta as typical in 22q11DS. At age 2, she was diagnosed with JRA and treated with methotrexate. The JRA is now in remission. She had occasional emotional lability, but met all of her milestones appropriately and performed well in her classes. On physical exam, she was found to be in the 5^th percentile for height and weight. Other than a slightly cylindrical appearance of her nose, she had normal facial features. There was no strong history of mental illness in P2. However, she is still within the risk age group for mental illness in 22q11DS. Detailed phenotype information can be found in Supp. Table S1.

DNA preparation and sequencing

Blood samples were obtained from all subjects and DNA was purified (Puregene system, Gentra Corp, Einstein Molecular Cytogenetics Core). All individuals were given internal identification codes, BM1452.001 (P1, proband 1), BM1452.100 (P1M, mother of proband 1), BM1452.200 (P1F, father of proband 1), BM1453.001 (P2, proband 2), BM1453.100 (P2M, mother of proband 2), BM1453.200 (P2F, father of proband 2). Genomic DNA was extracted from blood samples and then mechanically fragmented. After purification by electrophoresis, DNA fragments were ligated with adapter oligonucleotides to form paired-end libraries with a span size of 500bp. Libraries were sent for whole genome sequencing at Beijing Genome Institute (BGI, Beijing, China) using Illumina the HiSeq2000 sequencing system (Illumina, San Diego, CA).

Variant calling

For variant calling, GATK version v2.2 was used. In general, the sequencing reads that contained adapter sequence or high rate of low quality bases were removed. The sequence reads were aligned to the reference genome hg19 using the Burrows-Wheeler Aligner (BWA version 0.6.1) (Li and Durbin, 2009). Alignment files were converted from sequence alignment map (SAM) format to binary alignment map (BAM) format using SAMtools (version 0.1.18) (Li, et al., 2009). Duplicate reads were then marked using Picard tools (version 1.70, http://broadinstitute.github.io/picard/). The local realignment, recalibration of base quality values, and adjustment of per-Base Alignment Qualities (BAQ) were performed by components from GATK pipeline.

The Unified Genotyper was used to call SNPs or INDELs (DePristo, et al., 2011). The raw SNV calls were filtered to based on quality by depth (QD < 2), root mean square mapping quality (MQ < 40), Strand Bias (FS > 60), haplotype score (HaplotypeScore > 13), alternate vs reference read mapping quality (MQRankSum < -12.5), and alternate vs reference read position bias (ReadPosRankSum < -8). The filtering criteria for INDELs were QD < 2, ReadPosRankSum < -20, and FS > 200. To obtain phased genotypes and haplotype information, GATK was used to perform both transmission (PhaseByTransmission) and physical (ReadBackedPhasing) phasing.

In addition to the quality control measures used in GATK, we removed variants with genotype quality score ≤ 20, read depth ≤ 15, and alternate allele ratio (depending on genotype): homozygous reference ≥ 0.15, heterozygous alternate allele ratio ≤ 0.3 or ≥ 0.7 and homozygous alternate ≤ 0.85 (Patel, et al., 2014). Variant call format (VCF version 4.1) genotype calls were uploaded to the Database of Genotypes and Phenotypes (dbGaP, http://www.ncbi.nlm.nih.gov/gap) and BAM files were submitted to the Sequence Read Archive (SRA) (http://www.ncbi.nlm.nih.gov/sra, Study Accession: phs000837.v1.p1).

Copy number detection for 22q11.2 deleted region

Deletion size was determined by copy number variant (CNV) analysis of the 22q11.2 region. The chromosome 22 was extracted from aligned sequencing reads. Three tools (CNVnator version 0.2.2, BreakDancer version 1.1_2011_02_21, and EDRS version 1.06) were used with default settings to call structural variants with consensus between two out of three methods required to call a structural variant (Abyzov, et al., 2011; Chen, et al., 2009b; Zhu, et al., 2012). Bedtools (version 2.16.2) was used to identify CNVs with 50% reciprocal overlap between the three methods (Quinlan and Hall, 2010).Variant annotation and filtering

The basic functional annotation of SNVs/INDELs was performed by ANNOVAR version 2014Jul14 (Wang, et al., 2010). We used the table annotation script to annotate variants with RefGene, dbSNP (version 141), Exome Sequencing Project (ESP6500, 2013/05 release), 1000 Genomes project (1000 Genomes, 2012/10 release), dbNSFP v2.3 (for PolyPhen, Sift, LRT, and MutationTaster annotations), and Segmental Duplications (Adzhubei, et al.; Chun and Fay; Kumar, et al.; Liu, et al., 2011; Liu, et al., 2013; Schwarz, et al.; Setti, et al.). The DNA and amino acid changes were annotated using HGVS mutation nomenclature (den Dunnen and Antonarakis). We excluded variants in segmental duplications because of the high rate of false positive calls in these areas. Novel variants were defined as not being previously reported in the dbSNP 141, 1000 Genomes, and ESP6500. Rare variants were defined as those with a MAF ≤ 0.05 in the 1000 Genomes Project using all populations.

For coding sequence analysis, we only considered variants with possible deleterious effects including nonsynonymous SNVs, frame-shift INDELs and SNVs/INDELs that affect stop codons and splice sites (2-bp from exon-intron boundary). The cutoffs for predicting pathogenic scores are SIFT ≤ 0.05, PolyPhen2 HDIV ≥ 0.453, LRT = Deleterious, and MutationTaster = “disease causing” or “disease causing automatic”. Nonsynonymous SNVs were considered deleterious if variants were predicted pathogenic by at least two out of the four algorithms; we created an aggregate Deleterious Prediction Count (DPC), ranging from 0-4, indicating the number of deleterious predictions for a given variant.

We analyzed rare non-coding SNVs with MAF ≤ 0.05 in the 1000 Genomes Project database using the FunSeq (v0.1) standalone tool (Khurana, et al., 2013). FunSeq annotates non-coding variants with data from the ENCODE database, Gencode project, and protein-protein interaction information and regulatory networks. It generates a score ranging from 0-6, with 6 indicating maximum deleterious effect, and a score ≥4 indicating potential deleterious effects.

Genes with compound heterozygous mutations were identified by filtering based on strict genotype quality and inheritance criteria (Supp. Figure S1). Rare variants, MAF ≤ 0.05 were selected in the 1000 Genomes Project dataset that had genotype phasing information. Variants were filtered based on predicted pathogenicity, selecting variants predicted to be pathogenic by two or more prediction methods. Genes were selected with two or more pathogenic alleles that were located on different chromosomes.

A pipeline was implemented using Snakemake (v2.5.2), a flexible bioinformatics workflow engine (Koster and Rahmann, 2012). Scripts used for analysis are available on GitHub (https://github.com/jhchung/22q11DS_WGS_Trio_Analysis).

Candidate gene prioritization

The ToppGene suite (https://toppgene.cchmc.org/) was used to perform Gene Ontology (GO), pathway enrichment, and candidate gene prioritization for coding SNVs and INDELs (Chen, et al., 2009a). The ToppFun tool was applied, using the default parameters to annotate target genes for GO Biological Process and curated pathway terms. The ToppGene tool was used to prioritize candidate genes using known disease causing genes as training gene sets. The following training parameters were used all with FDR correction, p-value cutoff = 0.05 and Gene limits between 1 and 1500: GO: Molecular Function, GO: Biological Process, GO: Cellular Component, Human Phenotype, Mouse Phenotype, Pathway, Pubmed, Interaction, Transcription Factor Binding Site, and Disease. Test parameters were set to the default values.

For training sets, disease genes in humans and mouse models were downloaded from the OMIM (http://www.ncbi.nlm.nih.gov/omim) and MGI Human-Mouse Disease Connection databases (http://www.informatics.jax.org/humanDisease.shtml), respectively. The training sets for TOF, SCZD, and JRA can be found in Supp. Table S2.

Sanger sequencing

We validated SNVs and INDELs identified by WGS using Sanger sequencing of PCR amplified products. PCR primers were designed using Primer3Plus web based tool (http://primer3plus.com/). Target product size ranges were set to 150-500 base pairs with default parameters for all other settings. Primers used for PCR amplification can be found in Supp. Table S3.The PCR assays were performed using FastStart High Fidelity PCR system (Roche: 0478292001) Touchdown PCR with 70 °C and 58 °C starting and final Tm, respectively, was used for amplification (Korbie and Mattick). Products were purified using a PCR Purification Kit (Qiagen: 28104) and sequenced on 3730 DNA Analyzer (Applied Biosystems) at the Albert Einstein College of Medicine Genomics Core Facility. Chromatograms were visualized using SnapGene Viewer 2.7.1 (GSL Biotech, http://www.snapgene.com/products/snapgene_viewer/).

Results

Two trio families with de novo 22q11.2 deletion

Two unrelated 22q11DS probands with discordant phenotypes, P1 and P2, and their unaffected parents, were ascertained for genetic research and performed WGS on all six individuals (Figure 1A and 1B, grey shaded area). Subject P1, a male, presented with intellectual disability, behavioral findings, TOF, and hypogonadism. Additionally, he was diagnosed at age 27 with other specified schizophrenia spectrum disorder (DSM-V 298.8). Subject P2, a female, had only mild clinical findings including velo-pharyngeal insufficiency and resolved JRA (Supp. Table S1). The de novo 3 Mb deletion was determined in both probands by WGS (Figure 2A). The deletions were on the maternal allele for P1 and on the paternal allele for P2.

Figure 2.

The 22q11.2 deletion and validation of common variants on remaining allele. (A) De novo deletions identified from CNV calling on the 22q11.2 region using WGS data. Red sections indicate the high confidence deletions; pink segments indicate low copy repeat regions where there are low quality alignments and inaccurate CNV endpoint predictions. (B) A UCSC browser view showing rs4680 in COMT. P1 carries a single A allele. (C) Sequencing of rs4680 in family 1 using Sanger sequencing. P1 and the father (P1F) are homozygous for the A allele while the mother (P1M) is heterozygous G/A. The variant is highlighted in yellow. (D) Sequencing of rs4680 in family 2 shows P2 was hemizygous for the G allele, the mother (P2M) was heterozygous G/A and the father (P2F) was homozygous A.

Whole genome sequencing results and variant prioritization pipeline

We then analyzed the WGS of the two family trios to identify potential coding and non-coding single nucleotide variants (SNVs) and small insertions and deletions (INDELs). Each sample had a mean depth of 35.8X, covering 99.6% of the human reference genome with at least 20X coverage over 90.6% of the genome. Among all six samples, we detected 5,588,483 SNVs and 782,608 INDELs (Supp. Table S4). Each sample had on average 3,176,477 SNVs and 350,038 INDELs, which is comparable to previously reported WGS data from the 1000 Genomes Project (Genomes Project, et al., 2012). A total of 98.2% of SNVs and 86.92% of INDELs were previously reported in dbSNP version 141, the 1000 Genomes Project (2012/10 release), or the Exome Sequencing Project version 6500 (Supp. Table S4).

In order to prioritize the variants that might be pathogenic, we developed an analytical pipeline shown in Figure 3. The pipeline utilized several established tools to annotate coding and non-coding variants, filter the variants, and prioritize candidate variants using integrative genomic analysis methods. We used ANNOVAR to annotate variants with population frequency, phylogenetic conservation scores, gene regions, and exonic functions. Nonsynonymous variants were predicted to be deleterious if there was consensus between two out of four prediction algorithms. We next separated coding and non-coding variants for downstream analysis. We also filtered variants based on minor allele frequency to identify de novo, novel, and rare (MAF ≤ 0.05). For non-coding variants, we selected rare variants and used FunSeq to identify potentially functional non-coding variants. After filtering, we used ToppGene to prioritize the genes affected by the target variants based on functional similarity with established disease causing genes.

Figure 3.

Variant prioritization pipeline. After variant calling, variant annotation was performed using ANNOVAR to identify coding and noncoding variants. ANNOVAR was also used to annotate minor allele frequency data from dbSNP, 1000 Genomes, and Exome Sequencing Project databases. Predicted deleterious variants were selected if they were predicted to be deleterious by two out of four prediction methods (MutationTaster, Polyphen2, LRT, and SIFT). For noncoding variants, FunSeq was used to identify potentially deleterious variants, followed by allele frequencies and DPCs to identify candidate genes for variant prioritization. ToppGene was used to prioritize the candidate “target genes” using known disease specific genes as “training sets.”

Analysis of the remaining allele of chromosome 22q11.2

We first searched for SNVs or INDELs on the remaining allele of chromosome 22q11.2 since the 22q11.2 deletion can unmask deleterious alleles on the remaining chromosome. The 3 Mb region on chromosome 22q11.2 includes flanking low copy repeats (LCRs), also known as segmental duplications, (chr22:18656000-21792000; Figure 2A) (Bittel, et al., 2009). The LCR sequences were not included in the WGS analysis. In total, we identified 1,068 variants in P1 and 970 variants in P2 on the remaining allele of 22q11.2, the majority of which were non-coding.

The first priority was to identify nonsynonymous, potentially deleterious, variants in the coding regions because hemizygous deletions could uncover damaging variants on the remaining allele. On the remaining allele of 22q11.2, there were five and four nonsynonymous or frameshift variants in P1 and P2, respectively (Table 1). We found that P1 was hemizygous for the alternate alleles of common nonsynonymous SNVs rs4680 in catechol-O-methyltransferase (COMT; MIM# 116790, Figure 2B), encoding an enzyme that regulates catecholamine levels and rs450046 in proline dehydrogenase (oxidase) 1 (PRODH; MIM# 606810), encoding an enzyme that converts proline to glutamine (Table 1), both connected to brain and behavioral findings in 22q11DS (Graf, et al., 2001; Paterlini, et al., 2005). Proband P2 was hemizygous for the minor alleles of common nonsynonymous SNVs in armadillo repeat gene deleted in velocardiofacial syndrome (ARVCF; MIM# 602269), purigenic receptor P2X, ligand-gated ion channel, 6 (P2RX6; MIM# 608077), DGCR2, and an INDEL in clathrin, heavy polypeptide-like 1 (CLTCL1; MIM# 601273) (Table 1). Both probands carried the same alleles in DGCR2 and CLTCL1. The proband P2 did not carry any previously reported disease associated variants on the remaining allele of 22q11.2. However, we identified a common SNV (rs2277838; NM_001159554.1:c.647G>A:p.R216H) which was predicted to be damaging (DPC = 2) in P2RX6, encoding an ATP gated ion channel expressed in several tissues, of largely unknown function (Table 1). We did not identify any rare or novel coding variants on the remaining allele of 22q11.2 in either P1 or P2.

Table 1.

Nonsynonymous SNVs and frameshift INDELs on the remaining allele of chromosome 22q11.2. Listed are all nonsynonymous or frameshift variants found on the remaining allele of chromosome 22q11.2. All variants were called as homozygous alternate. We used an estimated location of chr22:18656000-21792000 to define the typically deleted region. Pathogenic prediction indicates the number of methods that categorize a variant as pathogenic; score can be between 0-4. No pathogenic prediction is calculated for frameshift INDEL or splice altering variants which are automatically categorized as pathogenic.

Subject	Chr	Position	Ref	Alt	rsID	Gene	DNA change	AA change	Pathogenic score
P1	chr22	18901004	C	T	rs450046	PRODH	NM_001195226:c.1238G>A	p.R413Q	0
P1	chr22	19026613	A	G	rs2072123	DGCR2	NM_001173533:c.1295T>C	p.V432A	0
P1	chr22	19951271	G	A	rs4680	COMT	NM_007310:c.322G>A	p.V108M	0
P1	chr22	21354970	C	G	rs426938	THAP7	NM_030573:c.343G>C	p.A115P	0
P1	chr22	19189003	-	C	rs11386977	CLTCL1	NM_001835:c.601dup3>G	p.V1201fs	frameshift insertion
P2	chr22	19026613	A	G	rs2072123	DGCR2	NM_001173533:c.1295T>C	p.V432A	0
P2	chr22	19959473	C	T	rs165815	ARVCF	NM_001670:c.2717G>A	p.R906Q	0
P2	chr22	21377650	G	A	rs2277838	P2RX6	NM_001159554:c.647G>A	p.R216H	2
P2	chr22	19189003	-	C	rs11386977	CLTCL1	NM_001835:c.601dup3>G	p.V1201fs	frameshift insertion

To identify potentially causal non-coding variants on the remaining allele of 22q11.2, we used the FunSeq software tool, which integrates a variety of annotations to predict disease causing potential (Khurana, et al., 2013). For P1, there were no SNVs on the intact chromosome that were predicted to be deleterious. In contrast we identified a predicted functional variant, rs6269, in P2 (FunSeq score = 4), which was located in the promoter region of the soluble COMT isoform. In agreement with the software prediction, studies have found the G allele is significantly associated with promoter hypomethylation levels (Schreiner, et al., 2011). This SNV is also part of a common haplotype associated with higher COMT enzyme activity and mRNA stability (Halleland, et al., 2009).

De novo variants outside of the 22q11.2 region

We next searched for de novo variants outside of the 22q11.2 deleted region that were predicted to be pathogenic to make genotype to phenotype correlations. We identified 48 and 32, de novo variants in P1 and P2, respectively. Of these, there was one nonsynonymous SNV in P1 with a DPC ≥ 2 but none in P2. P1 had a heterozygous de novo SNV in ADNP2 (NM_014913.3:c.2243G>C; ADNP homeobox 2; Entrez Gene ID: 22850) resulting in a p.R748P amino acid change predicted to be deleterious by 3 different methods. The variant was confirmed by Sanger sequencing (Figure 4A). The variant was located in the last exon of ADNP2 in a zinc finger, C2H2-like domain and is located at the beginning of a beta-sheet, as predicted by the SABLE web server (http://sable.cchmc.org/) (Figure 4C) (Adamczak, et al., 2005). There is a previously reported SNV (rs377757808) at the same nucleotide position but with a different substitution (NP_055728:c.2243G>A:p.R748Q) compared to the one we identified.

Figure 4.

De novo SNV in ADNP2 found in P1. (A) Sanger sequencing of the de novo variant (NM_014913:c.2243G>C, highlighted in yellow) confirmed that P1 is heterozygous G/C while both parents are homozygous for G allele. (B) UCSC genome browser views of the de novo SNV in ADNP2. The NHLBI Exome Sequencing Project identified another variant at this position, rs377757808, however, the previously reported alternate allele is A while P1 had C allele. Sequence conservation analysis shows a high degree of conservation among the 46 vertebrate species. (C) The G allele results in p.R748P amino acid change. SABLE prediction of protein secondary structure of ADNP2 suggests the de novo mutation affects the arginine (red amino acid letter) at the start of a predicted β-sheet (green arrow). The first row shows amino acid sequence, second row shows graphical representation of secondary structure (SS) prediction, third row show SS prediction confidence, fourth row shows relative solvent accessibility (RSA), fifth row shows RSA prediction confidence. SS legend: blue line = coil, red zigzag = α and other helices, green arrow = β-strand or bridge.

Since the majority of de novo variants were in non-coding regions, we used the FunSeq software tool to prioritize these non-coding variants (Khurana, et al., 2013). Only one de novo variant was predicted to be pathogenic in either proband. In P1, we identified a predicted functional de novo SNV (NM_144508.4:g.-9092G>T, FunSeq score = 4) in the promoter region of cancer susceptibility candidate 5 (CASC5; MIM# 609173). We validated this variant by Sanger sequencing (Supp. Figure S2). The CASC5 gene encodes for a protein that acts as a scaffold in the formation of kinetochore-microtubule attachments during chromosome segregation. Mutations in CASC5 have also been found in subjects with microcephaly, however its role in development is not well understood (Genin, et al., 2012). P2 had no de novo non-coding variants prioritized by FunSeq.

Novel predicted deleterious SNV and INDELs

Novel inherited variants could act as second hit modifiers altering risk for clinical findings in 22q11DS because the probands have the 22q11.2 deletion, which were not present in parental samples. P1 had 49 novel predicted deleterious variants (Supp. Table S5) and P2 had 44 novel predicted deleterious variants (Supp. Table S6). In P1, 19 variants were inherited from the mother (P1M) and 28 were inherited from the father (P1F) (Supp. Figure S3A). In P2, 20 variants were inherited from the mother (P2M) and 21 variants were inherited from the father (P2F) (Supp. Figure S3B). There were no genes with novel deleterious variants shared between probands. Interestingly, in P2, we found a novel predicted deleterious variant in zinc finger protein, multitype 2 (ZFPM2; MIM# 603693, NM_012082.3:c.1576C>T) which results in a p.R526W amino acid change. We validated this variant by Sanger sequencing (Figure 5A). ZFPM2 is a known candidate gene for TOF and will be described in more detail below.

Figure 5.

Novel inherited SNV in ZFPM2 found in P2. (A) A novel heterozygous SNV in ZFPM2 (NM_012082:c.1576C>T, yellow) was confirmed by Sanger sequencing. The variant alternate T allele was inherited from the father. (B) UCSC browser view shows that the SNV is in an evolutionarily conserved region of ZFPM2. (C) Schematic of ZFPM2 protein and known previously reported CHD related mutations (De Luca, et al.; Huang, et al.; Pizzuti, et al.; Tan, et al.). Circles show previously published mutations, red circles show mutations that occur in 2 or more subjects. Red diamond indicates position of novel mutation.

Rare compound heterozygous coding mutations

We next searched for genes that had a compound heterozygous inheritance pattern for rare predicted deleterious variants; which might implicate loss or reduced function of both copies of the gene. We searched for rare variants with a MAF ≤ 0.05 in the 1000 Genomes Project database with a DPC ≥ 2. We identified 5 genes with compound heterozygous mutations in P1 (Supp. Table S5) and 9 genes in P2 (Supp. Table S6). Among the genes affected in P1, the genes, FERM, RhoGEF (ARHGEF) and pleckstrin domain protein 1 (FARP1; MIM# 602654) and Reelin (RELN; MIM# 600514) are important for neuronal development (Cheadle and Biederer, 2012; Jossin, 2004). Among the genes affected in P2, DLG1 (discs large, Drosophila, homolog of, 1; aka DLGH1 and SAP97; MIM# 601014) is a candidate gene for autoimmune disorders (Zanin-Zhorov, et al., 2012).

Non-coding variant prioritization

The majority of variants identified by WGS were non-coding SNVs and INDELs. We selected rare variants with a MAF ≤ 0.05 in the 1000 Genomes Project database and used the FunSeq program to identify candidate genes with potentially damaging non-coding variants (Khurana, et al., 2013). A total of ∼120,000 rare inherited non-coding SNVs were present in each subject. In P1, there were 48 SNVs with a predicted functional effect (FunSeq score ≥ 4; Supp. Table S5), which were predicted to affect 23 genes. In P2, there were 51 SNVs with possible functional effects (Supp. Table S5), which were predicted to affect 29 genes.

Target genes for gene set enrichment and prioritization

We collected the potentially deleterious variants and their associated genes into “target gene” sets in P1 and P2. The target genes consisted of genes with (1) de novo and novel predicted deleterious variants, (2) rare compound heterozygous deleterious variants, and (3) rare non-coding predicted deleterious variants (Supp. Figure S4, dark grey node). We excluded variants on the remaining allele of chromosome 22q11.2 because they did not meet our filtering criteria. There were 76 target genes in P1 (Figure 6A) and 82 target genes in P2 (Figure 6B). The majority of the affected genes were unique to each subject and only 6 genes with rare, non-coding variants were shared between P1 and P2 (Supp. Figure S5). For each proband, we analyzed three sets of target genes: coding (Figure 6, red bars), non-coding (Figure 6, green bars), and combined coding and non-coding genes.

Figure 6.

Summary of potentially deleterious target variants. (A) Candidate functional DNA variants and genes in P1. (B) Candidate variants and gene in P2. Genes in blue indicate top prioritized candidates from variant filtering and ToppGene analysis. Circular tracks from outside to inside: gene symbols, ideogram of chromosome with cytogenetic bands, chromosome name, potentially deleterious coding variants (red bars), potentially deleterious non-coding variants (green bars). Circular plots were created using the Circos software tool.

Gene set enrichment analysis of target genes

Gene set enrichment analysis was performed on the target genes identified in P1 and P2 with Gene Ontology (GO) Biological Process terms using the ToppFun software tool. Proband P1 had significant enrichment of chromatin organization (GO: 0006325, FDR q-value = 4.304e-03) and the p53 signaling pathway (PW: 0000718, FDR q-value = 1.199e-2) (Supp. Table S7). P2 also had significant enrichment for GO terms and pathways related to chromatin and nucleosome assembly, however, the associations in P2 were driven by non-coding variants affecting several histone genes (Supp. Table S8).

Candidate gene prioritization for major phenotypes

We next prioritized the target genes based upon sets of well-established human disease genes using the ToppGene tool which prioritizes and ranks a target gene set based on functional annotation similarity with training set genes made up of known disease genes (Chen, et al., 2009a). For the training gene sets, we used the OMIM and MGI Human-Mouse: Disease Connection databases for genes with high confidence connections with the observed phenotypes (Supp. Table S2).

Cognitive and psychiatric candidate gene prioritization

There were 36 known genes that are associated with SCZD (we chose schizophrenia because P1 was diagnosed with schizophrenia spectrum disorder and it is the most common and severe behavioral defects in individuals with 22q11DS), in the OMIM and MGI databases, which we used as the training set (Supp. Table S2). Many of these genes are shared with other behavioral or neurocognitive deficits. We performed ToppGene prioritization on the 76 target genes in P1 and 82 target genes in P2. When testing all coding and non-coding target genes, 51.3% (39/76) of the target genes in P1 and 8.54% (7/82) target genes in P2 had significant functional similarity with SCZD training set genes (Fisher's exact test p-value = 2.13e-9) (Table 2, Supp. Table S9). This enrichment seems to be driven by variants that affect coding genes (17/53 in P1 vs. 3/53 in P2; Fisher's exact test p-value = 8.77e-04) since there was no significant enrichment when testing only non-coding target genes (Fisher's exact test p-value = 0.442) (Table 2). These results suggest P1 has a burden of deleterious coding variants in the behavioral/cognitive related genes. The top ranked genes in P1 included neurotrophic tyrosine kinase, receptor, type 2 (NTRK2; MIM# 600456), RELN, histone deacetylase 1 (HDAC1; MIM# 601241), and kinesin family member 5C (KIF5C; MIM# 604593) (Figure 6A, blue gene labels on chromosomes 9, 7, 1, and 2, respectively). We validated these variants by Sanger sequencing (Supp. Figure S2).

Table 2.

ToppGene prioritization results for coding, non-coding and all target gene analysis. Significant gene columns indicate the number of genes with significant functional similarity scores after multiple testing correction (FDR) for the number of input genes tested. There was significant enrichment of SCZD prioritized genes in P1 compared to P2 when testing all target genes and coding genes. There was also significant enrichment of TOF genes in P Significant p-values (≤ 0.05) are shown in bold. P-values are Fisher's exact test p-values comparing the proportion of significant genes in P1 vs. P2.

Target gene set	Phenotype	P1 significant	P2 significant	P1 input	P2 input	P-value
All genes	SCZD	39	7	76	82	2.13e-09
	TOF	21	5	76	81	4.17e-04
	JRA	13	7	76	82	0.15
Coding	SCZD	17	3	53	53	8.77e-04
	TOF	8	8	53	52	1
	JRA	7	3	53	53	0.319
Non-coding	SCZD	1	0	23	29	0.442
	TOF	11	8	23	29	0.157
	JRA	3	8	23	29	0.308

TOF candidate gene prioritization

There are six well-established human disease genes listed in the OMIM and MGI databases responsible for TOF (Supp. Table S2). When we analyzed coding and non-coding target genes together, there was a significant enrichment of prioritized genes in P1 (21/76 target genes) as compared to P2 (5/81 target genes) (Fisher's exact test p-value = 4.17e-04, Table 2). The top prioritized target gene in P1 was HDAC1 (Figure 6A, blue gene label on chromosome 1). This variant was validated by Sanger sequencing (Supp. Figure S2). P1 was heterozygous G/C for a rare non-coding SNV (rs187821252; NC_000001.11:g.33234860G>C) in HDAC1 that was predicted to affect its putative distal regulatory module (FunSeq score = 4; Supp. Table S10).

Interestingly, proband P2, who does not have any cardiac defects, has a novel heterozygous SNV in ZFPM2. Mutations in this gene are associated with TOF in OMIM (Figure 5C). This gene was mentioned above when discussing novel, predicted deleterious SNVs and it was also identified by gene set enrichment. The SNV found in P2 is located proximal to the 5th zinc finger domain (Figure 5C).

Juvenile rheumatoid arthritis candidate gene prioritization

JRA is a relatively rare phenotype associated with 22q11DS (Keenan, et al., 1997; Sullivan, et al., 1997). There were two genes linked to JRA in the OMIM and MGI databases (Supp. Table S2). There was no significant difference in the number of target genes prioritized by ToppGene between P1 and P2 when using coding, non-coding, or all target gene sets (Table 2). We focused on P2 because they presented with JRA. P2 had rare compound heterozygous mutations in DLG1 (Figure 6, blue gene label on chromosome 3; Supp. Table S11); rs34492126 and rs1802668 inherited from the father and mother, respectively. Both variants were validated using Sanger sequencing (Supp. Figure S6). DLG1 encodes a member of the membrane-associated guanylate kinase (MAGUK) protein family that regulates T-cell activation and is dysregulated in individuals with rheumatoid arthritis (Humphries, et al., 2012; Knowlton, et al., 2009; Zanin-Zhorov, et al., 2012).

Discussion

Our goal was to develop an efficient analytical WGS analysis pipeline to define the genetic variation of single individuals with 22q11DS. We specifically chose two, unrelated individuals with 22q11DS that have discordant phenotypes. In this way, they serve as a model for genotype to phenotype correlations. Our focus was on the main clinical findings observed in each affected individual, including the brain and heart. Using this analytical pipeline, we prioritized common variants on the non-deleted allele of 22q11.2 as well as novel and rare deleterious variants (Figure 6), which can then be tested in much larger datasets for replication and model systems for functional effects in the future.

Enrichment of schizophrenia related genes

Subject P1 suffered from cognitive as well as behavioral problems during adolescence and was diagnosed with psychotic disorder at 28 years of age. Family history of several family members on the paternal side were reported to be affected by mental illness, including SCZD, indicating that P1 could carry inherited risk factors. We found P1 carried the disease associated variants in COMT and PRODH on the remaining allele of chromosome 22q11.2. These variants have been extensively studied as possible risk factors for cognitive and/or behavioral phenotypes (Carmel, et al., 2014; Prasad, et al., 2008; Radoeva, et al., 2014). A de novo heterozygous mutation was found in ADNP2, in P1, predicted to be pathogenic. This variant may disrupt protein secondary structure based on SABLE in silico predictions. Although there is little known about the function of ADNP2 it may have a potential role in brain development similar to its paralog Activity-dependent neuroprotective protein (ADNP; MIM# 611386) (Dresner, et al., 2011; Kushnir, et al., 2008). Several of the top prioritized target genes in P1, have previously been associated with psychiatric disorders. For example, the top four prioritized coding target genes (NTRK2, RELN, HDAC1, and KIF5C) have been previously associated with cognitive disorders or psychiatric illness, including SCZD (Bergen, et al., 2014; Otnaess, et al., 2009; Ovadia and Shifman, 2011; Sharma, et al., 2008). As a whole, we identified an enrichment of SCZD candidate genes in P1 compared to P2, which is in line with recent studies on SCZD shown affected individuals have increased burden of genetic variants (Ahn, et al.; Chan, et al.; Kong, et al.). These findings highlight the possible importance of these genes as modifiers of the mental illness phenotype seen in P1.

Identification of TOF candidate genetic modifiers

TOF is a complex conotruncal heart defect that requires surgical repair for survival (Lee, et al., 2014; Starr, 2010). When we performed ToppGene analysis on all coding and non-coding target genes, there was an enrichment of TOF prioritized genes in P1 with TOF as compared to P2, with a normal heart. The top prioritized gene for P1 was HDAC1. In mice, Hdac1 regulates Nkx2-5, and important transcription factor for heart development and it genetically interacts with Gata4, another important heart transcription factor to repress downstream genes (Liu, et al., 2009; Stefanovic, et al., 2014). We also found that P1 target genes were enriched for the p53 pathway, with candidate deleterious variants in three genes (HDAC1, RCHY1, and ATM). Haploinsufficiency of TBX1, encoding a T-box transcription factor is one of the strongest candidate genes for TOF in humans with 22q11DS (Jerome and Papaioannou, 2001; Lindsay; Merscher, et al., 2001). Of interest, suppression of p53 was able to partially rescue CHD caused by the loss of Tbx1 during mouse heart development (Caprio and Baldini, 2014).

Surprisingly, P2, who had no heart anomalies, carried a novel predicted deleterious SNV in a known human TOF causal gene, ZFPM2. The ZFPM2 protein acts a co-factor for dosage sensitive GATA transcription factors including GATA4, during embryonic heart development in mouse models (Fossett, et al., 2001; Garg, et al., 2003; Pu, et al., 2004; Svensson, et al., 1999). Mutations in ZFPM2 have been previously reported in patients with sporadic TOF (Figure 5C) (De Luca, et al., 2011; Pizzuti, et al., 2003). Gene profiling revealed that inactivation of Tbx1 in mice results in ectopic expression of Gata4 and Zfpm2, in the mesoderm of the pharyngeal apparatus referred to as the second heart field (Liao, et al., 2008). Taken together, it is possible that a loss of function variant on one allele of ZFPM2 may normalize expression of GATA4 or other downstream genes, suppressing the phenotype.

Challenges of sequencing annotation and variant prioritization

We aimed to test the utility of currently available tools for identifying candidate genetic modifiers in a personal genomics manner. Our pipeline incorporated computational tools and recommended workflows for the interpretation and prioritization of coding and non-coding variants to create a personal genomics view of individuals with 22q11DS. Using this pipeline, we were able to identify several candidate genes for the various phenotypes seen in these two probands with 22q11DS. However, it is likely that there were some false negative (Type II error) and false positive (Type I error) findings due to stringent filtering criteria, incomplete annotation, or inaccurate functional predictions. Functional validation is a critical step in establishing the role of a variant and its target gene in the etiology of a phenotype (Casanova, et al., 2014; MacArthur, et al., 2014).

Predicting the functional consequences of coding and non-coding variants remains a significant challenge. For example, well known variants in COMT (rs4680) and PRODH (rs450046) have been shown to affect protein function, but the majority of prediction algorithms categorize these variants as benign. The interpretation of non-coding variants is especially challenging, however efforts by the ENCODE Project and other similar projects have greatly advanced our knowledge of functional elements in the non-coding genome.

We have performed the first comprehensive analysis of coding and non-coding SNVs and INDELs in 22q11.2DS patients. Despite the limitations of this study, this work suggests that it may be possible to take a personal or individual, genomics approach in identifying genetic modifiers for 22q11DS.