On the verge of diagnosis: Detection, reporting, and investigation of de novo variants in novel genes identified by clinical sequencing

2.1 NGS sequencing, variant calling, and analysis

Nine hundred seventy-one patient samples were referred by subspecialist pediatricians for trio clinical NGS (WES or WGS) with targeted phenotype-driven analysis in a 23-month period. Patients’ clinical records and previous testing results were reviewed prior to testing. Peripheral blood samples were provided for the proband and both parents, when available. DNA was extracted using a Chemagen (Perkin Elmer, MA) following standard procedures. NGS was prepared using the TruSeq PCR free library prep kit, with the addition of 5 cycles of PCR (Illumina, San Diego, CA). Sequencing was completed on an Illumina HiSeq 2500 or 4000 instrument (Illumina, San Diego, CA) utilizing paired end 2 × 125 base pair reads. Samples were sequenced for a mean coverage of ∼35 × (WGS) or ∼80 × (WES). Base calling was performed and required a minimum of 500,000 raw cluster density with 75% passing filter and 80% or greater of reads above Q30. If these quality checks (QC) metrics were satisfied, samples were processed through an alignment and variant detection pipeline using DRAGEN 2.0.4-2.1.3 (Miller et al. 2015); although some older samples were processed using BWA 0.7.2 and GATK 3.2-2. Postpipelining QC included a minimum of 85% of reads aligning to the human genome and a minimum of 85 Gb (WGS) and 6 Gb (WES) of data obtained after alignment is complete. Variant annotation and categorization was performed using Rapid Understanding of Nucleotide variant Effect Software (RUNES v.3.4.3–v4.2.4) as previously described (Saunders et al., 2012; Soden et al., 2014; Willig et al., 2015). Variants were filtered to a 1% minor allele frequency and prioritized by type using VIKING software, as previously described (Saunders et al., 2012; Soden et al., 2014; Willig et al., 2015), and using the American College of Medical Genetics and Genomics guidelines (Richards et al., 2015). Candidate gene lists were generated by SSAGA and/or Phenomizer using Human Phenotype Ontology (Kohler et al., 2014) terms with a cutoff at P value of 0.5. These gene lists were imported into VIKING to guide the analysis, however phenotype and OMIM filters were removed when necessary. Pathogenic, likely pathogenic, and variants of unknown significance in genes related to phenotype were reported; likely benign and benign variants were not reported. For WGS, incidental findings in the 59 genes recommended by ACMG (Kalia et al., 2017) were analyzed if requested by the family, with variant reporting limited to pathogenic and likely pathogenic. No specific CNV caller was used, however manual inspection of alignments was performed as needed. De novo variants (with parental sampled identity confirmed) were reported in GUS. These clinically reported GUS were limited to those with previously published links to a relevant human phenotype. Each GUS was submitted to GeneMatcher (https://www.genematcher.org) (Sobreira, Schiettecatte, Valle, & Hamosh, 2015). De novo variants in genes with no previous reported human phenotype were not reported. This study was approved by the CMH institutional review board.

2.2 Scoresheet for assessing clinical validity of 21 clinically reported GUSs with de novo variants

Gene–disease relationships were scored independently by two PhD molecular geneticists based on ClinGen curation categories: limited (1–6), moderate (7–11), strong (12–18), or definitive (>12 with replicative studies) published. The ClinGen Variant Curation Interface is available for public use (https://github.com/ClinGen/clincoded/wiki/GCI-Curation-Help). The category “definitive” was not applicable in this study because it is generally reserved for Mendelian disease genes that have been reported in independent studies over a period of at least 3 years (Strande et al., 2017). Points were assigned for the number of previously reported, unrelated patients with compelling variants in the respective gene, entries in Deciphering Developmental Disorders (Firth et al., 2009), SFARI (https://www.sfari.org/resource/sfari-gene/) (Larson et al., 2018; Zhang & Shen, 2017), or in large cohort exome or genome studies for NDD. Assessing case-level and case–control data was highly dependent on the inheritance pattern and phenotype reported. As all variants investigated in our cohort were de novo, the number of points for genes associated with autosomal recessive disorders was capped. Second, evidence was downgraded if the report contained different phenotypic findings or variant type inconsistent with the assumed disease mechanism (i.e., a missense variant in a gene associated only with LoF). In addition to genetic evidence, points were given for experimental evidence, gene function, cellular or model organism, protein interactions, pathway, and constraint metrics. The total number of points were used to determine the clinical validity score for each curator. The sum of points were weighted against any available contradictory evidences such as “negative” functional/animal studies, reduced penetrance, and lack of constraint in population databases (such as the residual variation intolerance score [RVIS], gene damage index [GDI], and ExAC constraint scores). Scores assigned by the two independent curators were averaged to determine the final classifications for the 21 gene–disease relationships.

ClinGen scores were compared to SFARI Gene scoring, if available. Briefly, SFARI scores assess the strength of evidence linking candidate genes to ASD (https://www.sfari.org/resource/sfari-gene/), considering genotypes observed in ASD cohorts, functional studies and animal models, and expert opinion/curation from the ASD scientific community. Gene–disease relationships fall into seven possible categories: S (syndromic, genes predisposing to autism in the context of a syndromic disorder such as Fragile X), category 1 (high confidence with genome-wide statistical significance between cases and controls, with independent replication), category 2 (strong: statistical significance between ASD cases and controls), category 3 (suggestive: relatively small studies of candidate genes, using either common or rare variant approaches), category 4 (minimal evidence: relatively small studies of candidate genes with accessory evidence), category 5 (hypothesized but untested: genes that have been implicated solely by evidence in model organisms or other functional nature), and category 6 (genes that have been tested in a ASD cohort, but the weight of the evidence argues against a role in ASD). SFARI curation relies on expert objectivity and complex assessment in the context of genetic heterogeneity in ASD.

3.1 Patient characteristics

Nine hundred seventy-one patient samples were submitted for clinical NGS in a 23-month period (Figure 1). The overall diagnostic rate was 21% (207/971), with de novo variants accounting for 43% (89/207) of positive cases. An additional 26 (3.4%) patients had heterozygous de novo variants in GUS. Five of these were excluded for lack of a published human phenotype and were not clinically reported (Figure 1). In total, 21 de novo variants in GUS were clinically reported in 20 patients, including ATG5, BPTF, CUL3, EEF2, GCC2, KDM3B, KDM5B, KIAA0100, KMT2C, MYADML, NLGN4Y, PDE10A, REST, RORA, RYR3, SBNO2, TBLIXR1, TRIP4, UBE2H, VPS4A, and ZNF666. Eight of 21 (38%) are predicted to cause LoF (Figure 1 and Table 2). The average number of variants reported in these patients was 10 (Table 1). One patient also had a pathogenic variant in PKD2, consistent with a diagnosis of Polycystic kidney disease; six patients were heterozygous carriers for one (4/6) or two (1/6) autosomal recessive diseases (Table 1). The diagnostic odyssey prior to NGS included multiple blood and urine samples, often involving cerebrospinal fluid punctures, comparative genomic hybridization (CGH) array, multiplex ligation-dependent probe amplification (MLPA^®), other NGS panel testing, and biochemical testing (Supporting Information Table S1). The mean age at testing was 6 years (ranging from 3 months to 18 years) with a male/female ratio of 2:1 (Table 1, Supporting Information Table S2). Etiologic testing was performed for a range of clinical concerns and by a variety of subspecialists. The majority of clinical NGS reports (57%) were ordered by geneticists, followed by neurologists (24%), 9% from perinatalogists, and one from gastroenterology (Supporting Information Table S2). A significant number of patients with de novo variants in GUS (95%, 19/20) had NDD, consistent with previous reports of high de novo variant rates in this group of disorders (Table 1, Supporting Information Table S1).

3.2 ClinGen scoring

To reassess the potential clinical validity of 21 de novo variants reported in GUS, two PhD molecular geneticists independently evaluated each gene using the ClinGen scoring system for gene–disease assertions (Strande et al., 2017). A comparison of points assigned by the two independent curators was completed and final classifications were determined for each gene (Table 2, Supporting Information Tables S3 and S4). Overall there was agreement between the two sets of scores, however, one minor discrepancy between curator classifications was noted for one gene (ATG5), which was classified at the border of limited versus moderate evidence (6.5:9); in this case, one curator weighted the sum of points to account for the inconsistences in the mode of inheritance reported (autosomal recessive vs. de novo AD). The linkage disequilibrium, functional studies, and knockout and knock-in mouse models supported the role of impaired autophagy in neurodegenerative diseases such as spinocerebellar ataxia but were inconsistent for mode of inheritance and the phenotype reported (global developmental delays, ptosis, craniosynostosis, lactic acidosis, epileptic spasms and focal T2 hyperintensity). The scoring system was influenced by the timing of publication, with significant gaps between independent publications suggesting uncertainty and newer published reports lacking time to refute claims. However, broader NGS-based testing led to more convincing publications as compared to linkage and/or candidate gene sequencing.

3.3 Correlation with OMIM, Orphanet, GeneMatcher, and SFARI gene entries

A minority (43%; 9/21) of reported GUSs had OMIM phenotypes (Table 2), most of which (8/9, 89%) had a creation date months or even years following the date of the clinical report (Table 2). Notably, there was significant overlap with Orphanet status (7/9 [78%] OMIM genes versus 8/21 [38%] total GUSs), with only one in OMIM but absent from Orphanet, and the other in Orphanet and not OMIM (Table 2). Consistent with the observation of OMIM phenotype entry date, 13 of 21 (62%) GUS in this study had publications near or after the time of report, 11 of which scored as having moderate or strong evidence. In addition, we compared the number of GeneMatcher hit(s), with ClinGen scores for the 21 GUS.

A positive correlation was observed between high-scoring GUS and number of GeneMatcher hits, with those with > 10 GeneMatcher hits all scoring as strong or moderate high clinical validity. In addition, 76% (16/21) of our GUS had at least one GeneMatcher hit, of which 37% (6/16) resulted in collaborations to functionally assess GUS, further phenotyping and potential publication.

Interestingly, for the seven genes with SFARI entries, there was strong agreement with ClinGen scoring in six (86%) (Table 2). The discrepant gene, RORA, had a minimal SFARI score but strong ClinGen score due to a very recent publication (Guissart et al., 2018). All four genes with SFARI scores of “high” or “strong” scored as “high” or “moderate high” for ClinGen, and two with a “minimal” SFARI ranking had A “limited” ClinGen classification. No disease–gene associations were scored as “definitive” due to lack of replicative studies over 3 years (Table 2, Supporting Information Tables S3 and S4).

3.4 Other metrics

We used three different published metrics to assess if a GUS is subject to strong selection against variation: (a) the RVIS (Petrovski, Wang, Heinzen, Allen, & Goldstein, 2013), (b) the GDI (Itan et al., 2015), and (c) the ExAC constraint metric (Lek et al., 2016). The RVIS predicts if a gene is more intolerant to variation (i.e., likely to be disease causing) (Petrovski et al., 2013). The RVIS was associated with strong or moderate high clinical validity score in nine of 12 (75%) GUS. However, three genes (CUL3, REST, and TBL1XR1) scored as strong or moderate high clinical validity but have a RVIS between ∼40–97%, indicating high tolerance to variation. The GDI, derived from a genome-wide, gene-level metric of the mutational damage in the general population, has been shown to be an efficient gene-level approach for filtering out false positive variant (Itan et al., 2015), thereby serving as a potential indicator of the relative biological indispensability (low GDI) or redundancy (high GDI) of a given human gene. However, little correlation was observed between the GDI and either RVIS or ClinGen scores (Supporting Information Table S3). Therefore, GDI was not helpful in assessing the clinical validity of the GUSs reported in our cohort.

Finally, we compared ExAC gene-level constraint metrics to ClinGen scores obtained for the 21 GUS. For LoF, three classes of genes with respect to tolerance to LoF variation are assumed. For genes with haploinsufficiency as the disease model, the closest the LoF constraint score (pLI) is to 1, the less tolerant the gene is to LoF (Lek et al., 2016). For missense variants, ExAC gene-level constraint metrics provide a Z score for the deviation of observed counts from the expected number, with positive Z scores indicating intolerance to variation (less variants than expected in the population) and negative Z scores indicating tolerance to variation (more variants than expected in the population) (Lek et al., 2016). The 21 clinically reported GUSs with de novo variants had an average pLI score of 0.7 (range: 0–1) and Z score of 2.5 (range: –1.26 to 5.66). Six of eight (75%) GUS reported with LoF are predicted to be highly intolerant to LoF (pLI: 0.91–1) and five of six (83%) were classified as moderate (1) or strong (4). The two remaining GUS with LoF variants, predicted to tolerate LoF (pLI: 0–0.02), had ClinGen scores of either moderate or limited. In the 12 GUS reported with de novo missense variants, 11 of 12 (9%) had an available Z score, and nine of 11 (81%) were predicted intolerant to variation with an average Z score of 3.15 (range: 0.75–5.66). Of those, five of nine (55%) were classified as moderate (2) or strong (3), and four of nine (45%) had limited association with human disease. In agreement with previous published data, genes encoding for protein in core biological processes and pathways have high constraint metrics for both type of variants (Supporting Information Table S3), providing evidence of putative association with severe Mendelian diseases (Lek et al., 2016). Based on ExAC pLI scores, ∼3,230 genes are highly intolerant to LoF, with less than 30% of those yet to be associated with a human phenotype. The remaining genes may represent good candidates for uncharacterized severe Mendelian diseases or incompatibility to life due to dosage sensitivity. Finally, two of 21 (9%) GUS could not be assessed using these metrics. The first, MYADML2, had no ExAC constraint metrics available, and the second, PDE10A, has very high ExAC constraint scores but the variant in question is a (c.1091+7A > G) intronic substitution with unknown impact on protein. Although good correlation is observed between scores generated by algorithms such as RVIS, GDI, and ExAC, they should be used with caution, particularly in a clinical setting.