The National Birth Defects Prevention Study (NBDPS) was a population-based study of over 30 major structural birth defects, which sought to identify environmental and genetic factors associated with these conditions. Details of the study methods and population have been outlined previously (Reefhuis et al., 2015; Yoon et al., 2001). Briefly, sites in 10 U.S. states were included as part of the NBDPS, including Arkansas, California, Georgia, Iowa, Massachusetts, New Jersey, New York, North Carolina, Texas, and Utah. Birth defect surveillance programs from these states ascertained children with eligible defects among pregnancies with estimated dates of delivery between October 1997 and December 2011.

All liveborn children diagnosed with BA were considered for inclusion. First, a board-certified clinical geneticist at each NBDPS site reviewed clinical information abstracted from medical records to verify eligibility (Rasmussen et al., 2003). Consistency across centers was established by a clinical geneticist who performed the final classification of each child diagnosed with BA. Children with known syndromes, chromosomal, or single-gene disorder etiologies were excluded. Next, children with BA were classified as isolated (no additional major birth defects or additional related birth defects only) or multiple (one or more additional major birth defects in an unrelated organ system).

Enrolled mothers completed a computer-assisted telephone interview 6 weeks–2 years after their estimated date of delivery. Following the interview, they were asked to collect buccal cell specimens from themselves, their child (if living), and the child’s father (if available) (Reefhuis et al., 2015). Mothers who participated in the NBDPS with a previous child, those who could not complete the interview in English or Spanish, or those who were incarcerated or otherwise did not have custody of their child at the time of recruitment were excluded.

There were 315 women with eligible pregnancies affected by BA, of whom 216 completed the telephone interview (Reefhuis et al., 2015). Of these, 115 mothers, 105 children, and 95 fathers provided buccal cell specimens. Similarly, 1513 parents from other birth defect groups from the NBDPS with buccal cell specimens for ES were selected as controls (Jenkins et al., 2019). As described previously (Jenkins et al., 2019), two different types of cytobrushes were used to collect specimens during phases of the study: “wet brushes” (1997–2003) (cytobrushes packaged in closed plastic tubes preventing air drying [Cyto-Pak Cytosoft Brushes CP-5B, Medical Packaging Corporation, Camarillo, CA]) and “dry brushes” (2003–2011) (cytobrushes packaged in open paper-backed peel pouches [Cytology Brush Pack CYB-1, Medical Packaging Corporation]). Informed consent was obtained for all participants providing buccal cell specimens, and the study protocol for the NBDPS was approved by the institutional review board at each NBDPS site.

Specimen processing, sequencing, and alignment were performed in collaboration with the National Institutes of Health Intramural Sequencing Center (NISC) at the National Human Genome Research Institute (NHGRI) and the University of Washington Center for Mendelian Genomics. Detailed procedures have been previously described (Jenkins et al., 2019). Specifically, due to DNA quality and quantity considerations, 64 child–parent trio specimens from only dry brushes were subjected to ES. Buccal specimens with adequate DNA amounts (≥200 ng assessed by quantitative real-time polymerase chain reaction [PCR] targeting the RNaseP gene) were sent to the NISC at the NHGRI for processing and sequencing. The ES capture kit used at NISC was a standard commercially available kit, the NimbleGen Seq-Cap EZ Exome+UTR Library (Version 3.0) and covered 96 Mb (Roche NimbleGen, 2013). The DNA was sheared mechanically, and targeted fragments were captured by probe hybridization and amplified before sequencing (Jenkins et al., 2019). NISC generated read lengths of 126 bases on an Illumina HiSeq 2500 instrument. Paired-end reads generated approximately 250 base pairs (bp) of sequence from each fragment in the library. A total of 38 million paired-end 126 bp reads were targeted and as many as 48 libraries were pooled and sequenced across as many lanes as needed to achieve the targeted number of reads (938 million read pairs or 76 million reads pre-library); thus, 5–6 libraries were run per lane. Image analysis and base calling were performed using the Illumina Genome Analyzer Pipeline software (version 1.18.64.0) with default parameters. In preparation for ES, ten trios and one father failed during sequencing library preparation due to bad reagents from contaminated library kits, and there was insufficient DNA quantity available for a second library rebuild. The final sample of BA cases sequenced comprised 54 child–parent trios and one child–mother duo. Cases were predominantly isolated BA (n = 50), whereas five had additional major birth defects.

After ES at NISC, Binary Alignment Map files were sent to the University of Washington Center for Mendelian Genomics for reprocessing. Reads were aligned to human reference (hg19hs37d5) using BWA-MEM (Burrows-Wheeler Aligner v0.7.10) (Li et al., 2009). Read data from a flow-cell lane were treated independently for alignment and quality control (QC) purposes in instances where merging of data from multiple lanes was required (e.g., for DNA sample multiplexing). Read-pairs not mapped within ±2 standard deviations (SD) of the average library size (~150 ± 15 bp for exomes) were removed. All aligned read data were subject to the following steps: (a) “duplicate removal” (Picard MarkDuplicates v1.111); (b) indel realignment (the Genome-Analysis-Toolkit (GATK) IndelRealigner v3.2–2); and (c) base quality recalibration (GATK BaseRecalibrator v3.2–2). Variant detection and genotyping were performed using the HaplotypeCaller tool from GATK v3.2. Following GATK best practices, variant quality score recalibration was performed. Variants flagged as low quality or potential false positives (quality score ≤ 50, long homopolymer run >4, quality by depth <5, or within a cluster of single nucleotide polymorphisms) were excluded.

Following specimen sequencing and alignment, we performed all subsequent ES data QC. In particular, all variants underwent additional genotype, variant, and sample QC prior to case–control analyses. We first removed variants with mean genotype depth < 10 reads. Variants were removed if they were multi-allelic, failed the Hardy–Weinberg equilibrium (HWE) check at p < 10⁻⁶ or had a call rate <0.99. GATK best practices (McKenna et al., 2010) for variant prioritization were applied, whereby variants with heterozygosity values >54.69 were removed. Next, the variant quality score recalibration pipeline in GATK v4.1.2 was implemented, utilizing seven informative annotation profiles (Quality by Depth, Mapping Quality, MQRankSum, ReadPos-RankSum, Fisher strand, Strand odds ratio, and InbreedingCoeff) to quantify the quality of all variants. Single nucleotide variants (SNVs) and insertions/deletions (INDELs) were evaluated independently using unique annotation profiles as recommended by GATK. Top-quality SNVs and INDELs were selected at the 99.7 and 99.0 percentile, respectively. Lastly, only uniquely mapped variants with a 100-mers mappability score of one were evaluated (Karimzadeh et al., 2018).

Sample QC involved filtering on mean sample’s genotype depth, number of variants, number of singletons, inbreeding coefficient, heterozygous-to-homozygous ratio, transition-to-transversion ratio, and missingness. Specifically, individuals were excluded if any of the evaluated metrics fell beyond ±6 SD from the sample mean. Sample kinship for genetic relatedness and genetic ancestry estimation were evaluated with KING v2.2 (Manichaikul et al., 2010) and PRIMUS v1.9 (Staples et al., 2013), respectively. Individuals identified to be duplicates, related at second degree or closer with cases, or parents of children with BA were excluded from the final sample for the case–control association analysis.

ES data was available for 54 child–parent trios. In each trio, variants were identified using Platypus v0.8.1 (Rimmer et al., 2014), and variant annotation was conducted using ANNOVAR (Wang et al., 2010) for information on variant type, alternate allele frequency (AAF) in the gnomAD v2.1.1 population database (Karczewski et al., 2020), and multiple in silico predictions of variant deleteriousness that included rare exome variant ensemble learner (REVEL) (Ioannidis et al., 2016) and phred-scaled combined annotation dependent depletion (CADD) (Kircher et al., 2014) scores. To identify potential pathogenic variants, we prioritized rare de novo (novel or gnomAD AAF <0.0001), homozygous (novel or gnomAD AAF <0.001), and compound heterozygous (novel or gnomAD AAF <0.001) variants.

Differences in case and control groups by demographic characteristics were compared using the Pearson χ² test. Common and rare variants were evaluated separately. Specifically, we utilized the single variant score test in Rvtests (Zhan et al., 2016) to identify common variants (minor allele frequency [MAF] ≥0.05) associated with BA. The association of rare variants (MAF <0.05) was evaluated using a gene-based approach. To prioritize for PAVs, we evaluated rare missense and rare synonymous variants independently since we expected synonymous variants to be unassociated. Furthermore, in the gene-based association analyses, we only evaluated genes with (1) at least two variants in the overall study sample and (2) at least one variant present in children with BA. Principal components (PCs) were calculated using PLINK v1.9 (Purcell et al., 2007) to capture unmeasured ancestry structure in the study population. We conducted the sequence kernel-based association test (SKAT) for a combined effect of rare variants using the SKAT v2.0.1 package in R v4.1.1 (Ionita-Laza et al., 2013; Lee et al., 2012). Briefly, SKAT is a region-based test for the joint effects of the individual variant score test statistic. Within a prespecified genomic region of multiple rare variants, SKAT performs a multiple regression approach directly regressing a phenotype on genetic variants and covariates, and SKAT p-values for the association are computed analytically (Lee et al., 2012; Wu et al., 2011). All analyses were applied using the efficient resampling method for the inclusion of extremely rare variants (Lee et al., 2016). Assuming an additive genetic model, all statistical models for common and rare variants were adjusted for sex and the first five PCs to account for population stratification. Quantile–quantile plots and genomic inflation factors were evaluated for signs of genomic inflation. Raw association p-values were corrected for multiple testing using the Bonferroni correction approach, and statistically significant findings were defined at the corrected p < 0.05.

Potential pathogenic variants identified in the child–parent trio analysis and their inheritance patterns were further validated by an orthogonal DNA-sequencing method. Target amplicons were amplified from genomic DNA using conventional PCR (HotStarTaqDNA polymerase, QIAGEN), and PCR amplification products were analyzed by Sanger sequencing using established methods.

Our initial population included 1568 individuals, including 50 isolated BA cases, 5 cases with multiple defects, and 1513 controls, that underwent ES with 754,935 variants available prior to QC. Following sample QC, 17 controls did not pass the inclusion threshold, and 32 controls failed relatedness checks, which included two duplicated controls and 30 controls related at second–degree or closer to children with BA. Similarly, following variant QC, we excluded 536,571 (71.1%) variants–34,999 (4.6%) variants failed the HWE threshold p < 10⁻⁶; 426,983 (56.6%) had a call rate <0.99; 71,289 (9.4%) failed GATK best practices; and 3300 (0.4%) had a 100-mers mappability score less than one. The final dataset for the child–par ent trio analysis included 54 child–parent trios. For the case–control analysis following QC, ES data on 55 cases, 1481 unrelated controls, and 218,364 SNVs and INDELs were selected.

Overall, half of the children with BA (n = 28, 50.9%) and controls (n = 740, 50.0%) were male (Table 1). Among those with BA, 54.5% were of European ancestry (n = 30), 20.0% of African ancestry (n = 11), and 10.9% of Asian ancestry (n = 6) based on PRIMUS v1.9 genetic ancestry estimation. In comparison, while the majority (n = 1023, 69.1%) of the controls were also of European ancestry, 7.8% (n = 115) and 5.5% (n = 82) were of African and Asian ancestry, respectively.

Rare de novo, homozygous, and compound heterozygous PAVs were prioritized in 54 BA child–parent trios. Overall, a total of 42 rare de novo PAVs were identified in 27 (50.0%) children with BA, of which five (11.9%) were loss-of-function variants (Table S1). However, no de novo PAVs were recurrent across more than one trio. A novel de novo stop-gain variant in the Notch receptor 2 (NOTCH2) gene (NM_024408.4:c.5194C > T,p.Gln1732Ter) was confirmed by Sanger sequencing (Table 2). Moreover, we identified two children with BA with compound heterozygous variants in the polycystic kidney disease 1 like 1 gene (PKD1L1), NM_138295.5:c.8485G > C(p.Glu2829Gln) / NM_138295.5:c.7552G > A(p.Ala2518Thr) and NM_138295.5: c.6473 + 2_6473 + 3del / NM_138295.5:c.731C > T(p.Pro244Leu) (Table 2). All variants in NOTCH2 and PDK1L1 were identified among children with isolated BA and orthogonally confirmed with Sanger sequencing. Additional homozygous (n = 1 in two children with BA) and compound heterozygous (n = 4 in six children with BA) variants are outlined in Tables S2 and S3, respectively.

Overall, 78,316 rare PAVs in 6919 genes and 48,642 rare synonymous variants in 6206 genes passed QC. The gene-based testing identified a significant association between BA and IFRD2 (p = 3.75 × 10⁻⁶; Bonferroni corrected p = 0.03) (Table 3). The IFRD2 gene-based association test was based on 21 rare PAVs, of which three variants–NM_006764.5: c.1016C > T(p.Ser339Phe), NM_006764.5:c.427G > A(p.Gly143Ser), and NM_006764.5:c.791G > A(p.Arg264Gln)–had a p < 0.05 based on the single variant Score test (Table S4). Among cases that carried PAVs in IFRD2, 50.0%, 12.5%, 25.0%, 12.5%, and 0.0% were of White, Hispanic, Black, Asian, and Mixed ancestry, respectively, while 27.2%, 18.4%, 18.4%, 8.8%, and 27.2% of controls were of White, Hispanic, Black, Asian, and Mixed ancestry, respectively (Fisher exact test p = 0.3). No other genes met the gene-based Bonferroni corrected p < 0.05 threshold. The gene-based genomic inflation factor in the missense rare PAVs analysis was observed at 1.13 (Figure S1). Additionally, no significant associations were identified among rare synonymous variants (Figure S2).

In the analysis of common variants, 15,944 SNVs and INDELs at MAF ≥0.05 were evaluated (Figure 1, Figure S3). While no variants were significant at the Bonferroni corrected threshold (p < 3.14 × 10⁻⁵), the top three hits included two synonymous variants–NM_001845.6:c.3189A > T(p.Arg1063=) and NM_001845.6: c.3183G > A(p.Gly1061=)–in COL4A1 and a missense variant–NM_007374.3:c.421C > A(p.His141Asn)–in SIX6 (Supplementary Table 5). In the gene-based association analysis, there was no statistical association for COL4A1 (p = 0.3). No rare PAVs were present for SIX6.