Assessing the readiness of Oxford Nanopore sequencing for clinical genomics applications

Performant analysis pipeline for ONT WGS data

ONT remains less mature than other sequencing platforms in the availability of standardized end-to-end analysis pipelines that integrate clinically relevant tools. In contrast, DRAGEN is the leading analysis platform for Illumina data, with version 4.2 onward including gene-specific callers relevant for pharmacogenomics (CYP2D6 and CYP2B6 in addition to the generic pharmacogenomics caller), RH blood type, and disease (SMN1 for SMA, GBA for Gaucher's disease, HB1/HB2 for alpha-thalassemia, CYP21A2 for congenital adrenal hyperplasia, and LPA for cardiac disease risk) (Behera et al. 2025). Therefore, a scalable and accurate workflow must integrate variant-calling methodologies along with upstream basecalling, alignment steps, and quality metrics to ensure high-quality results (Supplemental Note S2).

We have consolidated a comprehensive ONT WGS pipeline (Supplemental Fig. S1) that integrates improvements from our primary and secondary analysis workflows, along with additional callers for CNV and STR detection. We implemented concurrent real-time basecalling using a NVIDIA A100 tower connected to a PromethION 48 ONT sequencing instrument and the latest Dorado base-caller (https://github.com/nanoporetech/dorado), extracting both DNA sequence and 5mC methylation signals during basecalling. We then perform, in the cloud (a high-performance computing platform), mapping to the reference genome sequence using Sentieon-accelerated minimap2 (https://www.sentieon.com/), followed by variant calling with Clair3 for SNVs and indels (Zheng et al. 2022), Sniffles2 for SVs (Smolka et al. 2024), Spectre for CNVs (https://github.com/fritzsedlazeck/Spectre), and Straglr for STRs (see Methods; Supplemental Note S2; Chiu et al. 2021).

A high-quality diverse set of 17 WGS reference samples sequenced on multiple HTS technologies

We analyzed 17 Coriell reference samples to evaluate the performance of high-throughput sequencing (HTS) technologies in accurately detecting genetic variants. The samples were ordered from the Coriell Institute for Medical Research, which provides well-characterized reference materials with publicly available genetic truth sets. Cell lines were used instead of extracted DNA to minimize fragmentation and ensure high-quality input for long-read ONT sequencing (see Supplemental Note S1).

We selected two types of Coriell reference samples. The first comprised a well-characterized parent–offspring trio (GM24143, GM24149, and GM24385) that has been extensively whole-genome-sequenced by the community, with publicly available truth sets for SNVs and indels across the genome (Supplemental Table S2; Zook et al. 2019). The second type consisted of 14 samples from individuals affected by specific diseases, for which the pathogenic genetic variants are well documented (Supplemental Table S2). In each case, the disease-causing variant was expected to be present only in the corresponding sample. This design allowed us to evaluate the ability of each sequencing platform to detect not only small variants (SNVs and indels) but also larger forms of variation such as CNVs and STRs.

We performed ∼30× WGS for each of the 17 Coriell cell lines using long-read ONT (R9 and R10 chemistries) and short-read technologies (Illumina and MGI). The R9 chemistry, used since 2016, was discontinued in 2024 and replaced by R10, which provides improved accuracy and read length (Ni et al. 2023). This design resulted in 68 unique ID samples across platforms (17 samples × 4 platforms). To further assess intra- and interrun variability, we sequenced two samples (GM27631 and GM03620) two additional times on ONT, resulting in triplets: GM27631a, GM27631b, GM27631c and GM03620a, GM03620b, GM03620c. Thus, although Illumina and MGI each produced 17 runs (one per sample), the additional replicates on the ONT platforms increased their total to 21 runs corresponding to 17 unique IDs. Altogether, this yields a total of 76 runs/samples: 17 (Illumina) + 17 (MGI) + 21 (ONT R9) + 21 (ONT R10).

We evaluated the quality of all WGS data set and confirmed that all 76 samples met high-quality standards (Supplemental Tables S1, S3). These thresholds were defined based on the UK Biobank (Sudlow et al. 2015), The 1000 Genomes Project (1000 Genomes Project Consortium et al. 2015), and GATK Best Practices (DePristo et al. 2011). Overall, we generated >100 Gb of WGS data per sample across all three platforms (Supplemental Fig. S2; Supplemental Table S3).

The average genome coverage in ONT R9 samples (∼45×) was approximately one-third higher than that in R10 (∼33×) (Supplemental Fig. S3A), which reflects the known lower throughput of R10 flow cells (Ni et al. 2023) evident in our data (median yield: 103 Gb for R10 vs. 141 Gb for R9). Although ONT achieved higher mean coverage (∼40×) than SRS (∼35×), the proportion of bases with >10× coverage was slightly lower (92.0–94.0% for ONT vs. 95.8–99.3% for SRS), reflecting greater local variability. This variability stems from ONT reads extending into repetitive, subtelomeric, and pericentromeric regions where SRS often shows no coverage. As a result, ONT yields low (but nonzero) coverage in these areas, whereas SRS exhibits complete dropouts, inflating apparent dispersion when averaged genome-wide. Regions showing the largest platform differences were enriched near centromeres and telomeres (Supplemental Fig. S3D).

To assess whether this variability affects functionally important regions, we analyzed the well-characterized GM24385 sample and quantified the proportion of bases with >10× coverage across medically relevant genes (MRGs), protein-coding regions, and nonrepetitive regions. Across all subsets, ONT coverage was uniform and comparable to SRS (Supplemental Table S4). In MRG, ONT R9 and R10 achieved 99.8% and 99.2% of bases covered at >10×, closely matching Illumina (99.7%) and MGI (99.3%). Similar uniformity was observed genome-wide, confirming that ONT's variability arises mainly from repetitive regions, without affecting coverage in clinically relevant loci.

We also compared yield and read quality across platforms. Illumina and ONT R9 showed the smallest yield variation (93–131 Gb and 122–163 Gb, respectively), whereas MGI (103–171 Gb) and ONT R10 (72–198 Gb) were more variable. In Supplemental Table S3, we can also see how read accuracy is still much lower in ONT compared with SRS, as previously reported (Amarasinghe et al. 2020). Although we observed close to 90% of SRS sequencing bases with quality scores >Q30, a similar proportion of ONT sequencing bases only achieved >Q10. We could detect the improvement in read accuracy in R10 relative to R9 (median: ∼86% and ∼80% bases >Q10, respectively). This improvement is consistent at increasing quality thresholds as well as when comparing average values (Supplemental Fig. S4). In addition to increased read accuracy in R10, we also detected larger read lengths in that chemistry compared with R9 (Supplemental Fig. S5). Although we found >95% of mapping rate across all platforms, both ONT chemistries achieved the highest (>99%). Given these very high mapping rates across samples, yield translated into the proportionally expected effective mapping coverage values.

In summary, compared with SRS, ONT (R9 and R10) shows (1) lower read accuracy, (2) slightly higher variation in coverage driven by repetitive and pericentromeric regions, and (3) higher mapping rate; between the two ONT chemistries, R10 shows (1) lower yield, (2) improved read quality, and (3) longer reads.

In terms of the number of variants called per sample, we observed platform-specific differences (see Supplemental Fig. S6). Both SRS technologies resulted in about 5 million variants per Coriell sample (about 4 million and 1 million SNVs and indels, respectively). Both ONT chemistries resulted in a higher number of variants (R9: total = 5.5 million, SNV = 4.5 million, indel = 1 million; R10: total = 5.7 million, SNV = 4.5 million, indel = 1 million). These differences reflect intrinsic platform characteristics and the higher sensitivity of LRS, which can also increase false-positive (FP) rates and thus require stricter postcalling filtering.

We next performed principal component analysis (PCA) to the SNV/indel call sets to assess data consistency across samples and detect potential outliers or batch effects. The first two principal components (PC1–PC2, 20.58% variance) reflected the ancestry structure of the 17 Coriell reference samples. As shown in Supplemental Figure S7A, the single African American sample and the South American family trio (left and bottom data points, respectively) separated from the remaining European-descent samples. Each individual was represented by four data points, corresponding to sequencing performed with ONT (R9 and R10), Illumina, and MGI, which clustered closely together (Supplemental Fig. S7B). PC1 and PC2 also grouped family members, whereas PC4 distinguished short- and long-read data sets (6.47% variance), with minimal separation between ONT R9 and R10 (Supplemental Fig. S8). Although this shows the platform's influence on genotype calls, it remains minor and has no impact on the subsequent analyses. Additionally, no outliers were observed in any of the PCs, confirming the absence of problematic samples and the robustness of the data set.

Performance of SNV, indel, and SV detection across the genome

We used the parent-offspring trio (GM24143, GM24149, and GM24385) to assess SNV/indel variant-calling performance genome-wide (Supplemental Table S2). For each ONT (R9 and R10), Illumina, and MGI sample, we compared variant call sets to publicly available golden-truth variants to calculate recall, precision, and their harmonic mean or F-score (see Methods). These performance metrics were evaluated genome-wide, in exonic regions and MRGs. Across all comparisons, values were highly consistent within each sequencing platform, as shown by the narrow interquartile ranges in Figure 1, indicating robust and reproducible performance.

Figure 1.

SNV- and indel-calling performance across sequencing platforms. F-score, precision, and recall distribution for SNVs (left) and indels (right) across the different sequencing platforms. For Illumina, MGI, and ONT (both R9 and R10), displayed are performance metrics in high-confidence regions genome-wide. In addition, for ONT R9 and R10, shown are also performance metrics in exonic regions. Each boxplot encapsulates the corresponding metrics for the three Coriell samples with truth sets available (GM24143, GM24149, and GM24385).

Open in new tabDownload PowerPointLink to figure

In all three sequencing platforms, we achieved high genome-wide performance for SNV calling, with F-scores of 0.975–0.983 across all samples (Fig. 1; Table 1A; Supplemental Table S5). As expected, SNV precision (median = 0.997) exceeded recall (median = 0.963) regardless of the sequencing platform. Although Illumina exhibited slightly higher SNV precision, ONT (R9 and R10) achieved better recall, especially compared with MGI. Regardless, these values confirm robust and consistent SNV detection across platforms.

Table 1.

Summary of genome-wide SNV/indel-calling performance and detection of positive controls across sequencing platforms

Open in new tabDownload PowerPointLink to table

Accurately detecting indels is more challenging compared with SNVs (i.e., small indels <50 bp compared with single-point sequence changes). As expected, indel calling showed lower precision and worse recall across all platforms (Fig. 1; Table 1A; Supplemental Table S5). We found that the drop in the indel performance is very pronounced in ONT, with an F-score of 0.659–0.758, precision of 0.745–0.832, and recall of 0.592–0.696 in all samples. In contrast, the same metrics exceeded >0.85 in the SRS samples and were consistently higher in Illumina compared with MGI. Although ONT performed worse than Illumina and MGI, we observed a substantial improvement in indels detection with the newer chemistry compared with the previous R9.

Clinical genomic applications primarily target genetic variants with functional impact, typically restricting analyses to coding regions, namely, exons. Variants within exons are more likely to produce interpretable functional and phenotypic effects, whereas those in noncoding regions often have uncertain consequences. We therefore assessed whether ONT SNV/indel-calling performance improved in exons relative to genome-wide results. ONT showed higher indel performance in exons compared with genome-wide (Fig. 1; Supplemental Table S5), with recall increasing by 5.6% in R9 and 5.1% in R10. R10 also outperformed R9 in exons in terms of indel recall. This improvement is particularly important, as it reduces the proportion of missed true variants or false negatives (FNs), more frequent at the genome-wide level. Restricting our analysis to exons did not alter indel precision in R9, but it slightly increased that metric in R10, resulting in a lower FP rate. Nevertheless, ONT continued to miss some intergenic indels genome-wide. In contrast, restricting analyses to exons had minimal effect on the high SNV performance reported above (Fig. 1; Supplemental Table S5), with marginal recall gains and consistently high precision across both R9 and R10 data sets.

In the previous analysis, we focused on exonic regions, as these are most commonly examined in clinical genomics. Similarly, prior benchmarks have concentrated on medically important genes, such as the 73 actionable genes defined by the American College of Medical Genetics and Genomics (ACMG) (Mandelker et al. 2016; Miller et al. 2021; Mahmoud et al. 2024). Others have particularly focused on MRGs that are difficult or even impossible to characterize by SRS to highlight the potential of LRS to resolve these loci (Mandelker et al. 2016; Miller et al. 2021; Mahmoud et al. 2024). Thus, we evaluated all three sequencing technologies in a larger set of approximately 5000 MRGs (Wagner et al. 2022). SNV calling remained highly consistent across platforms (Supplemental Fig. S9, top left), underscoring their reliability for single-base variant identification. In contrast, ONT showed notable gains in both recall and precision relative to its genome-wide performance, particularly for indels (Supplemental Fig. S9, top right). Specifically, R9 and R10 each demonstrated an increase of ∼3% in F-score compared with their genome-wide benchmarks, with R10 exhibiting the greater improvement owing to its higher baseline performance. Illumina and MGI also performed robustly in this gene set, although their gains were more modest, with F-score increases of ∼1%. These findings underscore the impact of ongoing ONT refinements on variant detection in clinically relevant regions.

To better understand the source of ONT-specific indel errors, we examined whether homopolymer-associated artifacts contributed to the performance gap. Excluding homopolymer regions, both genome-wide and within the 5000 MRG set, led to mean F-score improvements of 24.9% and 18.9% for R9 and R10 genome-wide and 27.4% for R9 and 17.9% for R10 within MRGs (Supplemental Fig. S10). In contrast, Illumina and MGI exhibited minor variation, maintaining F-scores above 0.90 regardless of homopolymer inclusion. These results suggest that a substantial fraction of ONT-specific indel errors are attributable to homopolymer-associated artifacts, consistent with previous reports of signal compression effects (Lang et al. 2020; Delahaye and Nicolas 2021), and demonstrate that excluding such regions can improve variant detection in clinical contexts.

Building on this, we examined GM24385 across 273 clinically CMRGs with established truth sets to evaluate performance in complex loci. Consistent with the genome- and exome-wide results (Fig. 1), ONT (R9 and R10) performed comparably to Illumina and outperformed MGI when calling SNVs (Supplemental Fig. S9). Although ONT showed slightly lower precision than the two SRS platforms, recall was higher, even relative to Illumina. As previously reported (Wagner et al. 2022; Mahmoud et al. 2024), detecting indels within CMRGs is particularly challenging for SRS, as denoted by the decrease of 0.07 in precision relative to the genome-wide performance. In contrast, the impact in ONT is lower, with only a drop of 0.03 in precision. Again, the improvement of R10 relative to R9, especially in recall, is also noticeable in CMRG.

We also used the well-characterized GM24385 sample to compare SVs between LRS and SRS, for which SV truth sets exist. ONT achieved high average precision genome-wide (0.910 for R9 and 0.906 for R10), comparable to Illumina (0.916) and MGI (0.953) (Supplemental Table S6A). On the other side, the average recall in ONT was more than twice that of Illumina and nearly five times that of MGI (0.627, 0.644, 0.258, and 0.132 for ONT R9, ONT R10, Illumina, and MGI respectively). Consequently, the F-scores for ONT (0.743 and 0.753 for R9 and R10) were nearly double that of Illumina (0.403) and about three times higher than that MGI (0.231). The advantage of ONT became even more pronounced in CMRGs (Supplemental Table S6B). Precision for both LRS and SRS, along with the recall values for Illumina and MGI, exhibited consistency with genome-wide performance (Supplemental Table S6B), in which recall and F1-values approached 0.90, whereas precision across all platforms remained consistent with genome-wide results. Performance metrics were highly similar between ONT R9 and R10.

Performance detection of disease-causing mutations

We extended the evaluation of the performance of the three sequencing technologies to detect disease-causing mutations in 14 Coriell reference samples. These samples contained 11 mutations (three SNVs, three indels, two CNVs, and three STRs) linked to nine different diseases of high clinical relevance, for example, breast cancer, cystic fibrosis (CF), or DMD (Supplemental Table S2). In each of these samples, we assessed the presence or absence of the associated known mutation, for which genomic coordinates and changes in the genome reference sequence are publicly available.

Both SRS and LRS precisely detected the expected three disease-causing SNVs evaluated (Table 1B; Supplemental Table S7). In GM27630, mental retardation autosomal-dominant (MRD40) is caused by a de novo c.2127T > G (p.Tyr709*) pathogenic dominant mutation, which should therefore be absent in the parents (GM27631 and GM27632). By analyzing the trio samples from each sequencing platform, we confirmed, as expected, a single copy of the pathogenic variant in the proband and its absence in both parents.

We also evaluated three indels causing CF (556delA, del508, and 557delT), with delF508 present in three Coriell samples and GM07830 carrying two CF-associated mutations (Supplemental Table S2). We confidently detected the three phenotype-causing indels in all four sequencing technologies/chemistries. Our multiplatform high-quality sequencing approach detected a likely 1 bp annotation error in the GM07829 reference sample. Although the Coriell Institute for Medical Research reports a CF-causing adenine (A) deletion at Chr 7: 117,531,049 bp, corresponding to NM_000492.3(CFTR):c.424delA (p.Ile142fs) (Zielenski et al. 1991), we detected a deletion of the immediately downstream thymine (T) in all four WGS samples for GM07829 (Chr 7: 117,531,050 bp) (Supplemental Table S7). We ruled out mapping or variant-calling artifacts and confirmed the T deletion through visual inspection in the Integrative Genomics Viewer (IGV) (Robinson et al. 2011) (Fig. 2A). Sanger sequencing further validated the T deletion at Chr 7: 117,531,050 bp (Fig. 2B), providing orthogonal confirmation of the genotypes predicted by all HTS technologies. Of note, the detected indel has the same functional impact on the CFTR protein, fully consistent with the CF phenotype of the GM07829 sample.

Figure 2.

Across-platforms HTS and Sanger orthogonal validation uncover a 1 bp annotation error in GM07829. (A) IGV snapshots supporting the absence of the 556delA indel (Chr 7: 117,531,049 bp) annotated in Coriell for GM07829 but instead a 1 bp downstream T deletion (Chr 7: 117,531,050 bp) with the same detrimental functional effect. (B) Orthogonal validation of the HTS results through Sanger sequencing.

Open in new tabDownload PowerPointLink to figure

We accurately genotyped the CNV responsible for DMD in GM04099. Both SRS and LRS detected a heterozygous deletion of exons 49–52 in the DMD gene. For LRS, we assessed CNV detection with two different variant callers. The Sniffles2 v2.2 SV caller detected the expected drop in coverage corresponding to the deletion but incorrectly classified it as homozygous rather than the expected heterozygous genotype. In contrast, the updated Sniffles2 v2.5.3 correctly identified the deletion as heterozygous. Similarly, the latest version of the Spectre CNV caller implemented in our pipeline successfully identified this pathogenic deletion (Supplemental Table S7). Sequencing read alignments for GM04099 in both ONT R9 and R10 further supported the expected heterozygous genotype, with normal coverage in the flanking regions (∼30× for ONT R10) and approximately half coverage within the deletion region (Supplemental Fig. S11). These findings ruled out abnormal coverage as a confounding factor in the genotype call and indicate that the discrepancies observed with Sniffles2 reflect algorithmic rather than sequencing limitations.

Furthermore, the deletion of exons 7 and 8 in the SMN1 gene illustrates how ONT is catching up in sequencing accuracy and software developments with SRS. This variant is particularly relevant in the clinical context owing to its association with spinal muscular atrophy (SMA), an autosomal recessive neuromuscular disorder characterized by degeneration of α-motor neurons in the spinal cord, leading to progressive muscle weakness and atrophy (Wang and Lunn 2008). In most cases, SMA results from homozygous deletions in SMN1; however, accurate detection is complicated by the nearly identical paralog SMN2, which differs by a single nucleotide in exon 7 and modulates disease severity (Feldkötter et al. 2002). Illumina's DRAGEN platform includes a dedicated caller to resolve SMN1/SMN2 variation, which we also implemented in its open-source version (Chen et al. 2020) for MGI. Additionally, we introduced an alpha version of an SMN1 and SMN2 caller, named Sillago, specifically developed by ONT (Oxford Nanopore Technologies 2024). As expected, across all sequencing technologies evaluated, we correctly genotyped the SMA GM10684 sample as homozygous for a deletion of exons 7 and 8 in SMN1, while identifying two normal copies of SMN2 (Supplemental Table S7). Notably, Sillago provided concordant results for this sample on both R9 and R10 flow cells, suggesting promising performance in resolving this complex locus.

We next assessed trinucleotide repeat expansions leading to disease phenotypes (Supplemental Table S2) using dedicated STR callers implemented in our analysis pipeline. Expansion Hunter (Dolzhenko et al. 2017) was employed for Illumina and MGI, whereas Straglr (Chiu et al. 2021) was applied for ONT data. Our data set included a sample affected by myotonic dystrophy type 1 (DM1), an autosomal-dominant disorder marked by muscle weakness, myotonia, cataracts, and cardiac and endocrine involvement, typically appearing in the 20 sec to 40 sec (Brook et al. 1992; Udd and Krahe 2012). DM1 is caused by expansion of more than 50 CTG repeats in the 3′ UTR of the DMPK gene (Brook et al. 1992). Specifically, the GM03990 sample was reported to have between 50 and 80 CTG repeats. In our analysis, we observed 64, 57, 78, and 79 copies using Illumina, MGI, ONT R9, and ONT R10, respectively (see Fig. 3). Similarly, Huntington disease (HD) is an autosomal-dominant, progressive neurodegenerative disorder characterized by chorea, cognitive decline, and psychiatric symptoms, with onset typically between 30 and 50 years of age (Walker 2007). HD results from an expansion of more than 36 CAG repeats in exon 1 of the HTT gene. The GM03620 sample in our study is expected to present 60 copies of such repeat. Illumina detected 55 repeats and MGI identified 68, whereas ONT R9 and R10 found 58 and 60 copies (Fig. 3). Altogether, for both GM03990 and GM03620, in all sequencing technologies we predicted the STR copy number in the pathogenic range for DM1 and HD, respectively (Supplemental Table S7).

Figure 3.

Detection of STR expansions across sequencing platforms for genes FMR1, HTT, and DMPK in the 14 Coriell samples. The top panel shows the detection of CTG repeats in the DMPK gene in which expansions greater than 50 repeats are associated with myotonic dystrophy type 1. The middle panel displays the detection of CAG repeats in the HTT gene, with expansions beyond 36 repeats, characteristic of Huntington's disease. And finally, the bottom panel illustrates the detection of CGG repeats in the FMR1 gene, in which expansions exceeding 200 repeats are indicative of fragile X syndrome.

Open in new tabDownload PowerPointLink to figure

The third STR evaluated was a CGG expansion in the FMR1 gene that causes FXS when exceeding 200 copies (Supplemental Table S2; Hagerman et al. 2017). SRS failed to precisely call the pathogenic expansion in the GM09145 sample. Illumina and MGI detected 176 and 111 copies of the CGG repeat, respectively (Fig. 3; Supplemental Table S7), both below the greater-than-200 threshold defining the disease state. This highlights the inherent limitations of SRS in accurately detecting STR, mainly owing to average read length and the extensive size of these specific repeat sequences. In contrast, ONT detected more than 200 copies of this CGG repeat in GM09145 using both R9 and R10 chemistries (Fig. 3), although considerable variation was observed between them (864 and 573 copies, respectively) (Fig. 3).

Because the FXS-linked CGG repeat is relatively long, we hypothesized it could also be detected by Sniffles2 (Smolka et al. 2024), the SV caller in our ONT pipeline. Indeed, Sniffles2 identified the CGG repeat in FMR1 as an CGG-based insertion, consistent with Straglr estimates, detecting 842 repeats for ONT R9 and 772 for ONT R10 (Fig. 3). This approach using different callers for variant detection, highlighted the robustness of LRS in identifying variants confidently.

In summary, we correctly identified the expected genotype in 13 out of the 14 (92.86%) Coriell positive control samples across all four sequencing platforms, corresponding to 54 out of all the 56 genotypes (96.43%) overall. Moreover, each of the 11 disease-causing mutations we evaluated (Supplemental Table S2) was detected exclusively in its corresponding positive control and absent from all others, yielding a genotyping accuracy near 100% (Supplemental Table S7).

Finally, for ONT only we tested the consistency within (intra) and between (inter) sequencing runs using two of the Coriell positive controls: GM27631 (c.2127T > G in MRD40) and GM03620 (CAG repeat in HTT. Intrarun tests assess the stability of sequencing results within a single batch, detecting potential variability during the process. Interrun tests, on the other hand, extend this quality assessment across different runs, reagents, or instruments to ensure reproducibility under varying conditions. In the intrarun tests, we sequenced each of the two samples twice within the same run. The c.2127T > G mutation in GM27631 was successfully detected in both intrarun replicates for both R9 and R10. Similarly, the predicted CAG repeat numbers in GM03620 were consistent between the intrarun replicates for R9 (58 and 58 repeats) and R10 (59 and 60 repeats). For the interrun analysis, we sequenced the same two samples in two separate sequencing runs. The c.2127T > G mutation was detected in all interrun replicates, and the STR repeat numbers were again consistent across runs (R9: 58 and 58; R10: 59 and 60). Altogether, our results demonstrated satisfactory intra- and interrun replicability of ONT sequencing.

Across-platforms concordance

Beyond comparisons to reference truth sets, we also examined systematic differences in variant calling between sequencing platforms. We noticed that for ONT, both R9 and R10, an excess of SNPs/indels around the centromeres is called compared with other parts of the chromosome (Supplemental Fig. S12). That is also visible for SRS to a lesser extent and only in certain chromosomes. ONT also shows a higher proportion of variants in the region of Chromosome 6 that harbors the human leukocyte antigen (HLA). These trends likely reflect inaccuracies in the GRCh38 reference genome within repetitive and structurally complex regions, for which read misalignment is frequent. Although SRS typically filters these reads using stringent mapping quality thresholds, long-read technologies retain more of them, potentially contributing to the observed increase in variant calls.

We then assessed genotype concordance across sequencing platforms by calculating the Jaccard index between variant call sets derived from the same Coriell sample, separately for SNVs and indels (Methods) (Supplemental Fig. S13). In SNVs, we detected 78.5%–92.0% genotype sharing between any platform and chemistry compared (median = 82.4%); we evaluated about 4 million sites genome-wide. The two SRS technologies showed the highest concordance (90.1%) followed by 87.2% between ONT chemistries (Supplemental Fig. S12; Supplemental Table S8). The overall concordance between LRS and SRS was the lowest, ranging from 79.2% to 83.0%. Similarity, the concordance between ONT and Illumina was higher compared with between ONT and MGI. Unexpected similarity values between the newer ONT chemistry R10 and SRS were lower compared with R9 and SRS. As expected, similarity patterns in indels were lower across all platform comparisons (median: 59.9%; range: 50.8%–82.6%). Intra-SRS comparisons still showed the highest similarity (81.2%) but were ∼10% lower compared with SNVs (90.6%). Any ONT-based comparison fell below 70% similarity, even the R9 versus R10, which for SNVs was >86% in any sample. Contrary to the SNV results, in which R9 demonstrated better concordance with SRS than R10, the R10–SRS comparison for indels exhibited higher similarity (57.8%–61.7%) relative to the R9–SRS comparison (55.2% in both cases). Altogether, these results reinforce that accurate calling of indels is more challenging than for SNVs, especially for ONT.

Analysis pipeline for each HTS platform

Assessment of genome-wide and regional coverage uniformity

Coverage metrics were calculated from the aligned CRAM files using mosdepth v0.3.10 (Pedersen and Quinlan 2018). Coverage was computed at single-base resolution (1 bp bins), excluding secondary and supplemental alignments to ensure that each read contributes only once to the depth calculation.

To further evaluate coverage uniformity across functionally relevant portions of the genome, we quantified the proportion of bases with >10× coverage within three genomic subsets, MRGs, protein-coding regions, and nonrepetitive regions, using the GM24385 reference sample. This targeted analysis allowed us to determine whether genome-wide coverage variability extended into clinically or functionally important regions.

PCA

Beyond the standard QC metrics, PCA emerges as a powerful tool for uncovering underlying patterns and discrepancies in sequencing data. This step presents an additional layer of QC by assessing consistency across samples sequenced on multiple platforms, including Illumina, MGI, ONT R9, and ONT R10. PCA is a statistical procedure that transforms a set of possibly correlated variables into a set of linearly uncorrelated variables called principal components, in which the first component captures the greatest variance, and each subsequent one explains progressively less, while remaining orthogonal to the previous. In this study, PCA was applied as an additional QC measure to assess the consistency of sequencing data across sequencing platforms using 17 Coriell samples sequenced on each. After the analysis, the two main components were visually summarized in Supplemental Figure S7.

SNV/indel variant-calling performance evaluation

Evaluating variant-calling accuracy is essential in genomic research and clinical diagnostics to ensure reliable identification of genetic variants. This involves comparing the variants identified by our caller (test set) against a reference or “golden truth” for a given set of samples. The metrics commonly used for this assessment are recall, precision, and F-score.

Pairwise genotype concordance between platforms

We measured genotype concordance between sequencing platforms using the Jaccard index, a statistical measure for comparing the similarity and diversity of sample sets. The Jaccard index was calculated from the variant call sets in pairs of sequencing samples from the same Coriell specimen. This provided a quantitative measure of the proportion of shared variants relative to the total number of unique variants, thereby reflecting the similarity and concordance of the variant call sets across platforms.

To ensure high-confidence data, variants that passed all QC filters from the VCF were first retained. These filtered variants were then categorized into SNVs and indels to facilitate independent analysis of each type. The Jaccard index was subsequently computed for the specified pair-sets using BEDTools (Quinlan and Hall 2010).