A benchmarked sample, NA12878/HG001 (v4.2.1), was purchased from NIST (https://www.nist.gov/programs-projects/genome-bottle). The NA12878 sample represents a single female genome that has been extensively characterized using multiple genomic analysis platforms (Zook et al., 2016) (PMID: 27271295). Using the custom workflow described below, NA12878 was prepared and sequenced on an Oxford Nanopore Technologies PromethION-24 at the UMN Advanced Research and Diagnostics Laboratory (ARDL). A concordance analysis was performed to determine the ability of our ONT sequencing pipeline to detect the well-characterized SNV and indel variants present in the NA12878 sample. This concordance analysis was restricted to exonic variants (n = 26,584) in annotated human genes associated with clinical phenotypes. This restriction was performed by intersecting the NA12878 VCF file with a BED file (DOI: 10.5281/zenodo.14532101) containing exonic coordinates for 5631 clinically relevant genes (Supplementary Table S1) using BEDTools (Quinlan and Hall, 2010) (PMID: 20110278). This list of 5631 clinically relevant genes is regularly updated and it is curated using data from publicly available databases such as OMIM as well as new publications in the relevant medical literature identified via periodic targeted PubMed searches. Concordance analysis was performed by comparing VCF files and identifying exact matches at the chromosome (CHROM), position (POS), reference (REF), and alternate (ALT) columns using R (v4.0.4).

Seventy-two clinical samples previously analyzed at the UMN MDL were selected for nanopore sequencing (Supplementary Table S2). These samples were chosen as a representation of diverse genetic variation. Each sample had at least one clinically relevant abnormal repeat expansion, SNV, indel, SV, or variant in a gene with a highly homologous pseudogene. The available specimens were either extracted DNA or buffy coat stored at −80°C.

For 56 samples, DNA was purified from buffy coats on an Autogen Flexstar; 16 samples had been previously extracted [Qiagen DNeasy Blood & Tissue Kit (Cat No. 69506)]. Extracted DNA was concentrated using an Eppendorf Vacufuge plus at room temperature. In most cases, 4 µg of DNA was diluted into 150 µL water and sheared by centrifugation in Covaris g-TUBEs for 30 s at 1,250 g. Sheared DNA was characterized using an Invitrogen Qubit using the 1x dsDNA BR assay, and on an Agilent Tapestation. Ideally, samples had approximately 80% of the sheared fragments between 8 kb and 48.5 kb in length. Though Tapestation fragment size did not directly correlate with N50, the Tapestation fragment size was used as a quality check to estimate optimal levels of DNA shearing prior to sequencing. Samples were prepared for sequencing using the Oxford Nanopore Ligation Sequencing kit V14 using 3 µg of sheared DNA.

Samples were sequenced on a PromethION-24 and run on a single flow cell (R10.4.1 with the E8.2 motor protein) for approximately 5 days, with daily washing and reloading. For each sample, one library was built and split into four to five ∼300 ng aliquots. One flow cell was used to sequence four library aliquots. Each aliquot was sequenced overnight. The following day, the flow cell was washed for at least an hour before a new aliquot was added. Each flow cell was run for a total of 100 h. No adaptive sequencing was used. Several versions of MinKNOW were used depending on when the samples were sequenced. The versions used were 22.12.5, 23.04.6, 23.07.8, 23.07.12, 23.11.4, and 23.11.7.

Data was processed at the Minnesota Supercomputing Institute. Default parameters were used for all software unless otherwise noted. POD5 files were processed using Dorado(v0.3.4+5f5cd02+cu118) with a minimum q-score of 9. The Dorado basecalling model used was dna_r10.4.1_e8.2_400bps_sup@v4.2.0. Methylation analysis was not performed in this study. Unaligned BAMs were aligned to the hg19 genome using Nanopore’s epi2me wf-alignment workflow(v0.3.3). Sequencing depth was calculated using Mosdepth(v0.3.3) (Pedersen and Quinlan, 2018) (PMID: 29096012) and aligned N50s were calculated using Cramino(v0.9.9) (De Coster and Rademakers, 2023) (PMID: 37171891). Read quality was assessed with NanoPlot(v1.44.0) (De Coster and Rademakers, 2023) (PMID: 37171891).

SNVs were identified with Clair3 as the default genotyper using the epi2me wf-human-variation workflow (v1.2.0) (Zheng et al., 2022) (PMID: 38177392). Insertions/deletions/inversions (INS/DEL/INV) were identified using NanoVar(v1.5.1) (Tham et al., 2020) (PMID: 32127024) and DeBreak (v1.0.2) (Chen Y. et al., 2023) (PMID: 36650186). SVs identified by DeBreak were limited to a minimum of 100 bp in length while NanoVar was used to identify SVs > 50 bp. INS/DEL < 50 bp were identified using the Clair3 genotyper. CNVs on the sex chromosomes were identified using CNVpytor(1.3.1) (Suvakov et al., 2021) (PMID: 34817058) with 1 kbp bins, which were also filtered by: Q0 < 0.5, p_N < 0.5, and p-values < 0.001; choosing these parameters allowed for the highest quality calls for review. Autosomal CNVs were called by QDNAseq (1 kbp bin size), which is included in the epi2me workflow (v1.10.1).

A structured filtering approach was developed to accurately identify clinically relevant SVs (Figure 1). In the first filtering step, each of the caller’s outputs were restricted to a subset of calls related to its variant calling strengths. The next filtering step involved narrowing the list of calls to variant types that could mechanistically cause Mendelian disease. For example, an inversion impacting a subset of coding exons within a gene would advance to the next filtering step, but a complete gene inversion with intergenic breakpoints would be excluded. Another filtering step was then performed to restrict the calls to genes that have a human phenotype. This list of 5,631 genes associated with clinically relevant human phenotypes is regularly updated and it is curated using data from publicly available databases such as OMIM as well as new publications in the relevant medical literature identified via periodic targeted PubMed searches (Supplementary Table S1). A final set of logic rules was applied using a BEDtools/BCFtools intersect (Quinlan and Hall, 2010; Danecek et al., 2021) (PMID: 20112078, 33590861) as the last step to accurately identify and detect a pathogenic SV.

Samples were processed with Paraphase (v3.1.1) (Chen X. et al., 2023) (PMID: 36669496) for assessment of variants in genes like PMS2 and STRC, that are known to have homologous pseudogenes. Tandem repeat expansions were called using Tandem Genotypes (v1.9.1) (Mitsuhashi et al., 2019) (PMID: 30890163) and Sniffles2 (v2.0.7) (Sedlazeck et al., 2018) (PMID: 29713083). For Tandem Genotypes, reads surrounding loci of interest were realigned using LAST (v1256) (Kielbasa et al., 2011) (PMID: 21209072). A schematic representation of the pipeline is shown in Figure 2.

All clinical analyses and reporting infrastructure in our laboratory currently use hg19 as human reference. In order to maintain consistency and ensure that results can be accurately compared to our current clinical pipelines, sequencing data generated by our ONT-based pipeline was also analyzed using hg19 as the reference genome.

Since SNVs and small indels are the most common clinically relevant genomic alterations, we compared how the ONT-based pipeline performed when compared to our current short-read based clinical pipeline. Our ONT-based long-read pipeline was able to correctly identify 26,210 of the 26,584 gold-standard NA12878 variants across exons of known annotated coding genes, which corresponds to an analytical sensitivity of 98.87% and an analytical specificity exceeding 99.99%. In comparison, when our current clinically validated short-read pipeline was compared with the NA12878 sample, it showed a comparable analytical sensitivity of 99.59% and a specificity exceeding 99.99% on the same set of variants (Table 1).

To assess and compare the analytical sensitivity and specificity for detection of SNVs and indels, the set of 26,584 gold-standard NA12878 variants was analyzed separately based on variant type. For the 25,514 SNVs in this dataset, the analytical sensitivity was 98.91% and the analytical specificity exceeded 99.99%. For the 1,070 indel calls in this dataset, the analytical sensitivity is 97.81% and the analytical specificity exceeds 99.99% (Table 1).

While the National Institute of Standards and Technology (NIST) does have available samples with well-characterized SVs, these samples were not used to assess the analytical detection abilities of our pipeline. The available samples from healthy individuals do not have enough CNV or SV calls in coding regions of clinically relevant genes to accurately evaluate the analytical sensitivity and specificity of our pipeline.

Prior to testing our pipeline’s ability to detect a diverse spectrum of clinically relevant genetic variation, we assessed the functionality of several variant detection tools on a pilot set of clinical samples. Clair3 accurately identified all SNVs and small indels in the pilot dataset. Therefore, Clair3 was included in our pipeline as the default genotyper for small genetic alterations.

With respect to detecting SVs, we systematically evaluated the functionality of several available breakpoint callers (Supplementary Table S3). In our initial assessment of SV breakpoint callers, we evaluated eight clinical samples with known complex SVs, including a combination of deletions, duplications, inversions, and/or translocations. Across the eight samples with complex SVs, there were 29 known breakpoints. Of the tools evaluated, NanoVar was the most sensitive tool for breakpoint identification, with a detection rate of 93.1%. While DeBreak’s SV detection rate (86.2%) was lower than NanoVar’s, DeBreak was superior in characterizing the type of rearrangement occurring at a particular breakpoint rather than assigning it a generic classification such as “breakpoint not determined.” In contrast, SVIM and Sniffles2 displayed significantly inferior SV detection performance, with SV breakpoint detection rates of 58.6% and 6.9%, respectively. Therefore, based on the performance of the callers tested within our clinical sequencing environment (non-targeted capture with a target minimum genome-wide coverage of 30x), DeBreak was selected as the pipeline’s primary SV caller (for SVs >100 bases), with Nanovar implemented specifically to identify SV calls between 50-100 bases in size.

For detection of primarily larger SVs and CNVs that are not mediated by detectible breakpoints, we assessed two read-depth based callers, QDNAseq and CNVpytor. We found that QDNAseq performed best for autosomes, while CNVpytor performed best for sex chromosomes.

As clinically significant repeat expansion variants are a unique subset of SV that are notoriously difficult to accurately characterize by short-read NGS, we also assessed the functionality of a tool specifically designed to detect and quantify repeat expansion content. The Tandem Genotypes tool performed well for all loci assessed except for the pentanucleotide repeat expansions in RFC1 that cause cerebellar ataxia, neuropathy, vestibular areflexia syndrome (CANVAS). In search of a bioinformatics tool capable of identifying clinically relevant variants specifically in RFC1, we begin by reevaluating the previously discarded SV callers. We found that Sniffles2 worked particularly well for characterizing RFC1 repeat expansions. Therefore, this tool was added to the pipeline specifically for this purpose.

While several components of the pipeline such as Clair3, CNVpytor, and NanoVar were able to identify variants in some genes with highly homologous pseudogenes, they were not successful in certain genes such as PMS2 or STRC. As such, particular components of the Paraphase package were assessed to determine the functionality of this tool in parsing out haplotypes for genes like PMS2 and STRC. Since Paraphrase successfully distinguished between variants in highly homologous regions of genes such as PMS2 from pseudogenes such as PMS2CL, this tool was included in our pipeline with the intention of employing it only for particular loci such as PMS2 and STRC where other pipeline components were unsuccessful.

All analyzed clinical samples had a genome-wide coverage of approximately 30X; the median N50 was 10,294 bp (Range = 970–15,855 bases). The sequencing metrics are shown in Figure 3. We assessed the performance of long-read sequencing on a cohort of 72 samples from our clinical laboratory. This sample set contained a total of 167 clinically relevant variants that were assessed in this study (Supplementary Table S2; Figure 4). This variant pool consisted of 80 SNVs, 26 indels, 32 SVs, and 29 repeat expansions, which included 14 variants in genes with highly homologous pseudogenes.

A total of 80 clinically relevant SNVs were assessed by our ONT-based pipeline, all of which were present at a variant allele frequency (VAF) consistent with germline variation (VAF of ∼50% for heterozygous calls and 100% for homozygous calls). All 80 SNVs were accurately identified by Clair3 (Figure 2), indicating that our pipeline has an SNV detection concordance for germline variants of 100% (95% CI: 95%–100%) (Figure 4; Table 1).

We assessed 26 clinically significant indels using our ONT-based pipeline (Figure 4). The indel sample set consisted of insertions, deletions, and delins variants ranging from one to 36 bases in size. Clair3 accurately detected 25 of the 26 indels, including seven that fell within homopolymer stretches (4–8 bases in length) and three in genes with highly homologous pseudogenes.

Several examples of specific indels assessed using our pipeline are shown in Figure 5. Figure 5A depicts a complex indel in ABCD1 that was detected and accurately resolved by our pipeline. The only indel variant that was not detected by our pipeline was a complex indel in TRHR (Figure 5B). This variant, c.1137_1152delinsTTTTGTGGCAGGTGCTTGGCTGCCTGCCACAGGCAA, includes a loss of 16 bases and a gain of 36 bases. This TRHR variant was the largest independent indel variant assessed in this study and it was not called by Clair3. The next-largest indel variant assessed was BRCA1 c.2472_2490delinsGTAT, which resulted in a loss of 19 bases and a gain of four bases and was accurately called by Clair3.

Figure 5C shows a complex structural variant in GBA that contains a 55-base deletion. This GBA deletion is part of a complex recombinant allele that also includes several SNVs. As such, the recombinant allele was considered a SV in this study and was not included as an independent indel variant for data analysis. Notably, while Clair3 did not detect the 55-base GBA deletion, it was accurately detected by the breakpoint-based caller, NanoVar.

Since NanoVar accurately detected the 55-base deletion component of the GBA recombinant allele and Clair3 accurately detected a 19-base indel in BRCA1, but all pipeline components failed to detect the 36-base TRHR indel, a potential limitation of the current pipeline may be detection of variants between 20 and 55 bases in size, based on this dataset. Given the missed indel call in TRHR, the overall indel detection concordance of our pipeline was 96.15% (95% CI: 81%–99%) (Table 1).

Thirty-two clinically relevant SVs were assessed by our pipeline. A combination of breakpoint-based callers (NanoVar and DeBreak) and read depth-based callers (QDNAseq and CNVpytor) was required to accurately call and detect SVs, including copy number variants (CNVs). The combined output from the four callers for SVs and CNVs resulted in an average of about 56,000 raw calls per patient. This number does not indicate unique calls, but rather the sum of all calls across the four callers, including low quality calls that are subsequently filtered out. This combined total also does not account for overlapping calls made by multiple variant calling tools. After implementing a structured filtering strategy, the average number of relevant calls per sample requiring further evaluation and manual review was reduced to about 62 (Figure 1). All 32 SVs were successfully identified by our pipeline, indicating an SV detection concordance of 100% (95% CI: 89%–100%) (Figure 4; Table 1).

The validation cohort included 29 clinically significant repeat expansion variants. Our pipeline accurately detected and resolved all 29 repeat expansions, demonstrating a detection concordance of 100% (95% CI: 88%–100%) (Table 1). Tandem Genotypes software correctly identified clinically significant repeat expansions in ATXN1, ATXN3, ATXN7, DMPK, FGF14, FMR1, FXN, and HTT. Figure 6 depicts the wild type and expanded pathogenic CAG repeat sequences detected in a case of autosomal dominant Spinocerebellar ataxia type 1. All genes with expanded repeats in protein coding regions were accurately detected and sized. Tandem Genotypes also detected larger non-coding expansions, including FXN expansions greater than 3,000 nucleotides in length.

Included in these 29 pathogenic repeats were six RFC1 expansions, all of which were accurately detected by Sniffles2 but were missed by Tandem Genotypes. Importantly, manual review also showed that the expansion had converted from wild type sequence AAAAG to pathogenic sequence AAGGG.

Among the clinical samples assessed, there were 14 clinically significant variants in genes with highly homologous pseudogenes (Figure 7). These variants included SNVs, SVs, and indels in STRC, CHFR1, CHFR3, SBDS, GBA, PKD1, and SORD. Importantly, several of these variants could not be detected by short-read NGS and required specialized supplementary assays. The ten SNV and indel variants in genes with highly homologous pseudogenes were detected accurately by the Clair3 component of the ONT-based pipeline. The four SVs were called by Clair3, CNVpytor, and/or Paraphase (Supplementary Table S2). As such, our pipeline was able to detect all these variants in genes with highly homologous pseudogenes, including those that required supplemental methods previously, resulting in a detection concordance of 100% (95% CI: 78%–100%) (Table 1). Thus, detection of variants in genes with highly homologous pseudogenes is a notable strength of the ONT-based pipeline when compared to short-read NGS.

Since our laboratory routinely offers sequence analysis of PMS2 by long-range PCR (LR-PCR), we focused specifically on our pipeline’s ability to differentiate between variants in PMS2 and its pseudogene, PMS2CL. Paraphase (Chen X. et al., 2023) (PMID: 36669496) is a computational tool specifically designed for genotyping a small set of genes with homologous pseudogenes, including PMS2. LR-PCR results were used as a validation tool for providing orthogonal confirmation of results when possible. A key advantage of Paraphase is its ability to infer distinct PMS2 and PMS2CL haplotypes. Five of the 72 samples had orthogonal LR-PCR data available for PMS2 comparison with ONT outputs. Therefore, we elected to perform a concordance assessment specifically for Paraphase calls with a VAF > 0.2 and a quality score >1. In four of the five samples, all calls that satisfied these criteria were concordant between the nanopore and LR-PCR outputs. Of note, the concordance for other calls, such as those within non-coding homopolymer stretches greater than 10 bases long was lower. Typically, variants within these deep intronic non-coding homopolymer stretches are not expected to be clinically relevant.

Importantly, in one of the five samples, Paraphase identified an intronic block of 18 variants that were not called by LR-PCR. Notably, this is a region of the gene where variants are not analyzed and reported clinically. Paraphase accurately established these calls as being present in PMS2 despite having substantial overlap with PMS2CL. Since nanopore sequencing and Paraphase allow for reads to be “laddered” together to infer distinct haplotypes, this tool allowed us to conclude that the calls made by Paraphase but not LR-PCR represented a true gene conversion of PMS2 such that it contains a large but clinically insignificant region of PMS2CL.

In four cases where short-read NGS findings were not sufficient for a molecular diagnosis in the proband, our ONT-based pipeline was able to accurately detect and resolve variants leading to a complete clinical diagnosis.

In the first case, our current short-read NGS clinical pipeline identified what appeared to be a single heterozygous FANCA deletion (exons 1–23), in a 6-year-old male with a clinical presentation consistent with Fanconi anemia. A second pathogenic variant in FANCA was not detected. In contrast, our ONT-based sequencing pipeline detected two distinct FANCA deletions in trans (exons 1–11 and exons 12–23), with a 140 base pair overlap (Figures 8A,B), allowing for a complete molecular diagnosis for the proband.

The second case is a 62-year-old-male with neurologic symptoms consistent with ataxia. Prior testing via a short tandem repeat panel was negative for this patient, albeit before the discovery of Spinocerebellar ataxia 27B (SCA27B) (Pellerin et al., 2023; Rafehi et al., 2023) (PMID: 36516086, 37267898). ONT-based testing detected a pathogenic GAA repeat expansion in FGF14, indicating a molecular diagnosis of SCA27B for this proband. Subsequently, this patient received FGF14 short tandem repeat testing at an external laboratory via his clinical care team and the results were consistent with our finding.

In the third case, panel-based NGS testing of PKD1, PKD2 and PKHD1 performed in a 12-year-old female with Caroli’s syndrome identified two heterozygous pathogenic variants in the PKD1 gene. It was unclear whether these variants were present in cis or in trans. ONT-based sequencing clearly showed that these variants were present in cis as the result of a gene conversion event that incorporated pseudogene sequence into the PKD1 gene (Figure 9). It has been shown that the presence of two PKD1 variants in trans is associated with more severe disease and poorer prognosis (Ali et al., 2015) (PMID: 25880449). Therefore, establishing the cis phase of the variants in this proband was vital for informed clinical management and provided a more accurate estimate of disease progression.

To further demonstrate the prospective clinical utility of this technology, we analyzed a single case that had not been definitively resolved with prior short-read sequencing. The proband was a one-year-old male with a clinical diagnosis of osteopetrosis. Prior to referral to our institution, the patient had short-read whole genome sequencing performed at an external laboratory. This testing identified a single heterozygous c.346C>T (p.Gln116*) nonsense variant in the TCIRG1 gene. A second clinically significant TCIRG1 variant was not identified by the external testing. The patient was referred to our institution for potential bone marrow transplantation based on a presumed diagnosis of autosomal recessive TCIRG1-related osteopetrosis. Upon receiving a new sample from the proband and both parents, targeted testing of the parental samples revealed that the c.346C>T (p.Gln116*) variant was paternally inherited. Short-read whole genome sequencing performed at MDL provided evidence of a maternally inherited potential structural variant, but the exact nature of the variant could not be resolved. This sample was sequenced with ONT and a novel Alu insertion was identified at chr11:67816810 (hg19). This Alu insertion was located 49 bases from the exon 15/intron 15 splice boundary. Both breakpoints of this novel Alu insertion were confirmed by Sanger sequencing. The finding of a novel Alu insertion in trans with a pathogenic nonsense variant allowed for a highly probable molecular diagnosis in this case (Figure 10). This insertion was classified as a variant of uncertain clinical significance by strict application of the ACMG criteria (Richards et al., 2015) (PMID: 25741868). This Alu insertion is an example of a variant that is difficult to fit into an interpretive framework that is very specifically designed for sequence variants and small indels. However, in consensus conversations with the clinical providers caring for this patient, we strongly believe that this variant represents a diagnostic finding, which was reflected on the final report. Overall, we chose a conservative formal interpretation and classification given the lack of functional evidence regarding this novel intronic insertion. Of note, because this Alu insertion variant could not be formally classified as clinically significant, this case was not included in the table of clinically significant variants.

All 80 SNVs evaluated in clinical samples were accurately detected, thus showing a 100% detection concordance for clinically relevant SNVs (Figure 4; Table 1). To stress test the pipeline and assess the ability of our platform to detect variants at non-germline VAFs, we utilized an incidental finding of polycythemia vera in one of the samples studied. The JAK2 c.1849G>T (p.Val617Phe) variant, observed at a VAF of 24% in peripheral blood on prior short-read testing, was not detected by the variant callers in our pipeline. Manual review of the BAM file confirmed that the variant was present in sequencing reads for this sample. The inability of our pipeline to detect this mosaic variant is not surprising, given that the pipeline was optimized for detecting germline mutations. Therefore, reliable detection of constitutional mosaic findings may be limited with the current ONT-based pipeline.

In several cases with complex SVs, although a diagnostic finding had been reported based on short-read NGS or other techniques, a full characterization of these complex variants could not be completed using existing clinical methods. In these cases, our ONT-based pipeline helped us to clarify the exact genes involved in a particular SV and better understand the ways these complex SVs altered cellular biology and contributed to proband phenotypes. These SVs were detected using a combination of Clair3, breakpoint-based callers, and read-depth-based callers. We were able to develop an effective SV call filtering strategy that eliminated a majority of the superfluous calls, allowing us to focus on reviewing those with the highest likelihood of being clinically relevant, thus allowing for a high detection concordance of clinically relevant SVs (Figure 1). This filtering strategy is critical for establishing a workflow that can be practically implemented in clinical laboratories. A combination of both read-depth and breakpoint-based callers was used; Clair3 is our default genotyper, while the primary SV callers are NanoVar and DeBreak. DeBreak could accurately detect SVs between 100 kb and 361 kb, but large deletions and duplications (>750 kb) mediated by repetitive elements and those extending to the chromosome centromere or telomere could not be routinely detected by this caller. NanoVar calls were only used to bridge the gap between the maximum for Clair3 (50 bases) and the minimum for DeBreak (100 bases). In contrast, the general purpose SV caller, Sniffles2, showed poor overall performance in detection of complex SVs (Supplementary Figure S3) and was therefore not included as a primary SV caller in our pipeline. It is unclear as to exactly why Sniffles2 did not perform as well in detecting SV breakpoints. This could be due to the use of complex multi-component SVs intended to challenge the limitations of the SV-calling tools as our pilot dataset. Our data therefore suggests that while Sniffles2 may work well for detecting simple deletions and duplications, it showed less than optimal performance for detection of the complex SVs included in our dataset.

As the first step in our SV filtering strategy, we restricted each of the caller’s outputs to a subset of calls related to its variant calling strengths. Our second filtering step entailed identification of variants that were most likely to have an effect on Mendelian disease; another filtering step was then performed to restrict the calls to genes that have a human phenotype (Supplementary Table S1). A final set of logic rules was then applied for accurate detection of a pathogenic SV. This list is not patient- or sample-specific; therefore, this strategy can be adapted for any clinical scenario. We recognize that there are unique complex structural variants, so we maintain all raw outputs for possible manual review in cases with high levels of clinical suspicion.

A 55-base deletion, which was part of the complex SV in GBA was not detected by Clair3; but was detected by the breakpoint-based caller NanoVar (Figure 5C). However, the 36-base TRHR indel (Figure 5B) was missed both by Clair3 and NanoVar. Manual review of the BAM file showed that this TRHR indel was present, indicating that the miss was not due to an error in sequencing. Unfortunately, there were no other SVs in the clinical dataset between 20 and 55 base pairs in length that could be assessed to further narrow the size window of variants this pipeline has difficulty detecting. Thus, reduced capacity to detect indels between 20 and 55 bp may be a potential limitation of our current bioinformatics pipeline that will need to be addressed in future optimization of this clinical workflow.

For analysis of repeat expansions, one specific challenge is that the mere presence of an expanded repeat does not always correlate with pathogenicity. In addition to size of the repeat expansion, nucleotide content is often critical for predicting disease penetrance, severity, age of onset in affected individuals as well as meiotic instability. Using CANVAS as an example we demonstrated that our pipeline was not only able to detect clinically relevant expanded repeats in RFC1, but also provided nuanced information about the content. Two key components are required to confer molecular pathogenicity in CANVAS: the repeat content must change from an AAAAG repeat sequence to one of several pathogenic pentanucleotide repeat sequences and the repeat length must expand from a nonpathogenic size (typically 11–200 repeats) to greater than 400 repeats (Delforge et al., 2024) (PMID: 38627134). Indeed, manual review of the BAM files from samples with known pathogenic RFC1 expansions showed that the reads containing the pathogenic repeats had altered from wild type pentanucleotide sequence AAAAG, to pathogenic pentanucleotide repeat sequence, AAGGG, indicating that the genotyping pipeline was able to accurately measure the size as well as content of this region for clinical diagnostic purposes. Importantly, we demonstrate that Tandem Genotypes was able to identify all the tested repeat expansions in our cohort but was unable to accurately detect all the pathogenic RFC1 repeats tested. Expanded RFC1 repeats, implicated in CANVAS, were accurately detected only by Sniffles2. Interestingly, This demonstrates the need to incorporate multiple different callers into a clinical pipeline for detection of clinically relevant variants to achieve high confidence in detection of clinically relevant repeat expansions.

As indicated in Figure 7, our pipeline was able to detect all variants in genes with highly homologous pseudogenes. The majority of these variants were detected by Clair3, Nanovar, and CNVpytor. The Paraphase tool was used to assist with variant detection in a specific subset of genes requiring a haplotype based approach, such as PMS2 and STRC, which is in alignment with the UMN MDL’s currently available test offerings. When the laboratory eventually transitions to using a newer reference genome (discussed below), the available Paraphase loci offerings supported for that reference will be evaluated for potential clinical implementation.