Introduction

Clinical gene panel and exome sequencing have transformed the diagnosis of rare Mendelian disorders, and greatly reduce repeated blood samples, cost, and time. Despite the successes, many cases remain unresolved, with reported diagnostic yields ranging from ~15% to ~60%. Negative results have been attributed to factors including incomplete knowledge of disease architecture, a focus on exonic variation, challenges in variant pathogenicity interpretation, and technical limitations influencing variant calling. Assumptions incorporated into test design and analysis pipelines can also contribute to missed diagnoses. For example, assumptions about inheritance patterns led to overlooked variants in the imprinted genes CDKN1C1 and MAGEL2 (refs. 2,3).

Incomplete consideration of alternative transcripts can also cause pathogenic variants to be missed. We recently reported a patient with epileptic encephalopathy for whom clinical gene panel testing was unrevealing.4 Research-based genome sequencing identified a de novo variant in an alternative transcript of CDKL5, a gene targeted by the clinical panel. Similarly, in a reanalysis of previously undiagnosed epilepsy patients, the Epilepsy Genetics Initiative identified three cases with de novo variants in an alternative transcript of SCN8A, an isoform that had only recently been added to the set of transcripts evaluated.5

These variants demonstrate that alternative transcripts can be disease-relevant. Here, we investigate whether these examples are isolated cases, or whether alternative isoforms may be more widely relevant to clinical sequencing. Using neonatal epilepsy as an example, we found that clinically relevant alternative transcripts are common in known disease genes. The results suggest that alternative isoforms should be assessed more routinely in assays dependent on a set of reference transcripts, including gene panel, exome, and genome sequencing, and that reanalysis or resequencing incorporating alternative transcripts should be considered for patients with negative test results.

Materials and methods

Genes and transcripts

Gene symbols and RefSeq (https://www.ncbi.nlm.nih.gov/refseq/) identifiers for the primary transcripts assessed by neonatal epilepsy clinical gene panels as of December 2017 were provided by the genetic testing companies. Genes were limited to those strongly associated with neonatal epilepsy, defined as a primary seizure condition starting in the first months of life. Genomic coordinates (hg19) of the panel transcripts and alternative transcripts associated with the neonatal epilepsy genes were extracted from RefSeq and the GENCODE v27 comprehensive data set (hg19.wgEncodeGencodeCompV27lift37), downloaded from the University of California–Santa Cruz (UCSC) Genome Browser (http://genome.ucsc.edu/). The GENCODE data were filtered for transcripts annotated as protein-coding and with complete coding regions.

Alternative coding regions

Alternative coding regions were computed for each gene by subtracting the genomic positions of the coding exons and 20 flanking bases of the panel transcript(s) from the coding exons and 20 flanking bases of the filtered GENCODE transcript(s). Evidence of expression in neonatal brain for a region was defined as ≥50% of the coding bases supported by >20 normalized reads in the fetal or infant RNA-Seq data from polyA+ transcriptomes of human dorsolateral prefrontal cortex (DLPFC), downloaded from the Lieber Institute for Brain Development (LIBD) DLPFC Development UCSC custom track hub.6 For alternative coding regions confined to intronic flanking sequence, expression was computed using the associated exonic bases.

Variants

Variants were obtained from ClinVar (https://www.ncbi.nlm.nih.gov/clinvar/) version 201711 and Human Genome Mutation Database (HGMD) Professional (Qiagen) version 2013.2, preprocessed as described7 and annotated with ANNOVAR (http://annovar.openbioinformatics.org/) version 2017-06-01. Variants were considered pathogenic if they were categorized as pathogenic or likely pathogenic in ClinVar or as disease-causing mutations (DM) in HGMD, limited to neonatal epilepsy-related disorders when the condition was provided. Relative variant deleteriousness was defined as putative loss of function (stopgain, stoploss, consensus splice site, frameshift, startloss) > nonsynonymous > synonymous. Variants were filtered for allele frequency <0.0001, using the maximum frequency from the gnomAD genomes and exomes,8 where average coverage was ≥20 or ≥50, respectively. Annotations were computed from brain-expressed transcripts, defined as transcripts for which every exon has >50% of its coding bases supported by >20 normalized reads in the fetal or infant LIBD RNA-Seq data.6

Genome sequencing

This study was approved by the Inova Institutional Review Board (IRB 15–18196). Full written informed consent was obtained from the participants, with the parents providing consent for minors. Genome sequencing methods are described in the Supplementary Methods.

Results

Neonatal epilepsy genes have brain-expressed coding regions that are not evaluated by clinical tests

The genomic positions sequenced by gene panel tests are limited to the exons and flanking sequences of a set of reference transcripts. We determined a set of “alternative coding regions,” defined as genomic regions of a gene that would be newly sequenced by consideration of alternative transcripts, using neonatal epilepsy genes as an example. We first generated a list of transcripts sequenced by three representative clinical gene panel tests, the Invitae Epilepsy Panel (189 genes), the EpilepsyNext panel from Ambry Genetics (100 genes), and the Fulgent NeoNatal Epilepsy panel (276 genes), from data kindly provided by the companies. All three companies confirmed that the provided transcripts are the primary reference transcripts for these genes in both their gene panel and exome tests. The combined set of transcripts from the three panels has 292 genes and 305 transcripts (Supplementary Table S1). Most of the genes (96%) are represented by a single transcript, and 13 genes (4%) are represented by two transcripts.

To determine the alternative coding regions, we subtracted the genomic positions of the coding exons and flanking bases of the panel transcripts from those of transcripts from GENCODE9 (Supplementary Figure S1). The GENCODE data set has 1372 quality-filtered transcripts for the 292 panel genes, with 1–39 transcripts per gene (median 3). Most of the genes (85%) have alternative transcripts, consistent with the ubiquity of alternative splicing.10 The alternative coding regions were then limited to those transcribed in fetal or infant brain to prioritize sequences more likely to be relevant for neonatal epilepsy, resulting in 147 regions (Supplementary Table S2). Eighty-nine genes (30%) have at least one alternative coding region (range 1–6). The regions are 1–801 nucleotides long (median 74) and encompass a total of 15,713 genomic positions, of which 11,369 are exonic coding bases (72%) and 4,344 are in flanking sequences. The regions are distributed throughout the length of the encoded proteins: 19% encode alternate N-termini, 58% encode alternate C-termini, and a partially overlapping 50% are middle regions.

The set of alternative coding regions includes exons from transcripts previously shown to be expressed in brain, including alternative isoforms of CACNA1A, CDKL5, DNM1, SCN2A, and SCN3A (see Additional References). Alternative coding regions were also found for two bicistronic loci, MOCS1 and MOCS2, each of which encodes two overlapping open reading frames, of which only one is in the set of transcripts assessed by the clinical panels. Alternative exon 5A from SCN8A, the location of recently identified pathogenic variants,5 was excluded because both isoforms are assessed by the Invitae panel.

Variants in the alternative coding regions can be disease-relevant

To determine whether the alternative coding regions may be clinically relevant, we asked whether any known pathogenic variants are located in these regions. Although these regions are not routinely examined by exon-based clinical tests, variants may have been identified using other methods. We found 16 pathogenic variants located in alternative coding regions of 5 genes, CACNA1A, GFAP, MOCS1, MOCS2, and STXBP1 (Table 1a). Of the 15 published variants, the reported impact is consistent with the alternative transcript annotation, and 6 variants have functional data supporting an effect on protein function or expression (Supplementary Table S3). These examples confirm that variants in alternative coding regions can be disease-relevant.

Table 1 Pathogenic variants
Full size table

Alternative transcripts may alter reporting of variants detected by panel transcripts

In addition to identifying previously undetected variants, assessment of alternative transcripts could alter variant reporting by providing alternative annotations of the sequenced variants. Alternative annotations could impact interpretation of pathogenicity, variant prioritization, and hypothesized mechanisms of disease pathogenesis. To explore this effect, we searched for reported variants that are reannotated as loss-of-function variants based on alternative, brain-expressed, transcripts. We identified four pathogenic variants, in the genes CDKL5, ATRX, GLB1, and MOCS2, and no benign variants or variants of uncertain significance (VUS) (Table 1b). The accuracy of the predicted protein changes is unknown, but the variant in ATRX was shown to impact splicing, an effect not predicted by either the panel or alternate transcript annotations. These results suggest that alternative transcripts could alter variant reporting, consistent with studies demonstrating dependence of annotations on the set of reference transcripts,11 and that, like all variant annotations, the predicted impact should be interpreted cautiously.

Impact of alternative transcripts on patient data

To examine the potential impact of alternate transcripts on patient data, we reannotated publicly available exome results from 337 probands diagnosed with epileptic encephalopathy (epi4kdb.org). Although these data are themselves limited by a set of reference transcripts, we identified three rare protein-coding variants in alternative coding regions and no variants with potentially more deleterious annotations (Table 2a). We also analyzed genomes from 44 probands with congenital disorders, including 1 patient with neonatal epilepsy,4 and identified three rare protein-coding variants in alternative transcripts from the epilepsy panel genes, the pathogenic CDKL5 variant in the epilepsy patient and two VUS in patients without epilepsy (Table 2b, c). These results suggest that consideration of alternative transcripts can improve detection of pathogenic variants without introducing a large number of VUS.

Table 2 Reportable variants in patient data
Full size table

Discussion

Clinical gene panel and exome sequencing have provided molecular diagnoses for many rare disease patients, but for some patients these tests are nonexplanatory. Ongoing efforts to identify overlooked pathogenic variants include novel disease gene discovery and analysis of regulatory variants. The results presented here reaffirm that incomplete representation of alternative transcripts also causes pathogenic variants to be missed, and suggest that more complete evaluation of protein-coding regions in known disease genes will increase diagnostic yields.

Recent publications have highlighted the importance of reanalysis of sequencing data from initially uninformative exome tests.12,13,14,15 Our study underscores this conclusion, and suggests that assessment of alternative transcripts should be part of the re-evaluation. Because many of the alternative coding regions identified here are fully captured by commonly used exome capture kits (48% by Illumina TruSeq and 67% by Agilent SureSelect), initial re-evaluation may require only computational reanalysis without additional sequencing.

Our analysis identified a set of alternative coding regions for 292 neonatal epilepsy genes, but this set is not expected to be comprehensive. Determination of the regions relied on a database of reference transcripts that is incomplete,6 and regions may have been excluded due to low read counts in the expression data. The set of alternative coding regions is also likely to include false positives, including regions resulting from computational errors such as incorrect transcript mapping to the reference genome, and regions that are not relevant to seizure disorders. The regions may also include segments difficult to sequence by short-read technologies, such as the polyglutamine repeat region of CACNA1A.

In this study, we focused on neonatal epilepsy but consideration of alternative transcripts is likely to benefit clinical testing for a broad range of diseases. Alternative splicing occurs in a wide variety of tissues and cell types16 and affects ~95% of multiexon genes.10 Pathogenic variants affecting alternative exons have been identified in the gene ACTG2 for the smooth muscle disorder megacystis–microcolon–intestinal hypoperistalsis syndrome, in SCN5A for the cardiovascular disorder congenital long-QT syndrome, and in ABCA4 for the retinal disorder Stargardt disease (see Additional References).

Currently there is no standardized method for selecting transcripts for clinical tests. Each company individually defines a set of primary transcripts based on sources such as HGMD (Qiagen), Alamut (Interactive Biosoftware), and literature review, or selects the longest transcript. Efforts underway to more fully characterize the human transcriptome across cell types and developmental stages6,17,18 and to curate clinically relevant exons19 will aid the detection and evaluation of variants in alternative transcripts. Variants presented here support the disease relevance of some alternative transcripts. Incorporating alternative transcripts into sequencing tests will likely yield data useful for determining additional disease-relevant regions.

This study has important implications for clinical practice. Although it is unknown how many attainable diagnoses are missed due to the nonassessment of alternative transcripts, our results indicate that clinicians should consider genetic tests that assess multiple isoforms, particularly for patients with negative test results for whom a positive result was expected. Adding additional transcripts may also introduce VUS requiring further evaluation, but ongoing efforts to fully characterize the transcriptome will help resolve the uncertain results, yielding additional pathogenic variants, increasing diagnosis rates, and ensuring a more complete genetic evaluation.