Systematic evaluation of long-read and short-read sequencing in neurological disorders diagnosis: a direct comparison study of 310 patients

Abstract

Genetic neurological disorders are highly heterogeneous, and many are driven by complex variants that challenge short‑read genome sequencing (srGS). Long‑read genome sequencing (lrGS) has recently shown promise, but head‑to‑head evaluations are limited. A direct comparison study of srGS and lrGS was performed for 310 families with undiagnosed neurological disorders from the Hong Kong Genome Project to assess their diagnostic performance, technical capabilities and costing. Genome sequencing showed an overall diagnostic yield of 22.6% (n = 70/310, 77 variants). lrGS and srGS showed comparable variant detection rates at 92.2% (n = 71/77) and 96.1% (n = 74/77), respectively. lrGS solely confirmed three repeat expansions with phased methylation data and enhanced accuracy in two complex structural variants. srGS detected six variants in homopolymeric regions that were missed by lrGS. Costing analysis revealed a comparable unit cost per sample for srGS and lrGS with USD1,545.97 and USD1,580.00, respectively. This study demonstrated that srGS excels in detecting variants near homopolymers while lrGS offers better resolution for complex variants and enables the simultaneous analysis of methylation and phasing. Although lrGS is not yet capable of fully replacing srGS, we anticipate that the continued technical improvement would position lrGS as the first‑tier genomic test for neurological disorders.

Introduction

Genetic neurological disorders (NDs) encompass a broad range of conditions caused by dysfunction of the central and/or peripheral nervous systems, many of which warrant genetic evaluation when a hereditary etiology is suspected^1–3. Accurate molecular diagnosis is complicated by extensive phenotypic overlap between ND subtypes^4,5. In addition, pronounced genetic heterogeneity exists in NDs⁶ so short‑read genome sequencing (srGS) has recently been proposed as a front‑line diagnostic modality for many NDs^6–8.

Over the past decade, advances in genetic testing have revolutionized the diagnostic and research landscape in clinical genomics. srGS is currently the cornerstone with high diagnostic yield, well-established pipelines, high accuracy and cost efficiency^9–12. However, a critical limitation of srGS lies in its inability to resolve complex genomic regions and structural variants (SVs) due to the inherent constraints of short sequencing reads^13–15. In recent years, long-read genome sequencing (lrGS) technologies, such as Oxford Nanopore Technologies (ONT) nanopore sequencing and Pacific Biosciences (PacBio)’s HiFi sequencing, have demonstrated exceptional promise in addressing these limitations. lrGS offers unprecedented capabilities in resolving complex genetic variants^13–15. Specifically, lrGS provides longer read lengths for more accurate read alignment in challenging genomic regions, such as low-complexity or highly repetitive regions, and regions with high sequence homology. lrGS enables the simultaneous detection of haplotype-resolved complex SVs, short tandem repeat (STR) expansions and DNA methylation patterns, offering a more comprehensive assessment for ND diagnosis.

Despite its potential, the application of lrGS in the diagnosis of NDs has been limited by several factors^16,17. The PacBio sequencing platform offers relatively lower throughput compared to Illumina srGS and ONT lrGS. On the other hand, while the frequent updates in the ONT platform have brought significant improvements in the sequencing accuracy, it still requires intensive computation resources, particularly in GPUs for base-calling, and has difficulties in determining the exact length of homopolymers^17–20. Higher running cost of lrGS has also been cited as an additional limiting factor. It is further complicated by the absence of robust large-scale comparative evidence between lrGS and srGS, leaving critical gaps in understanding their relative clinical utility^14,16,17,21. Addressing these evidence gaps will provide the necessary data to support the appropriate use of lrGS technologies in facilitating genetic ND diagnoses. This study aims to optimize testing strategies and support the integration of lrGS by leveraging experience from the Hong Kong Genome Project (HKGP), the first large-scale genome sequencing project in the Hong Kong Special Administrative Region of China.

Results

Participants

Probands and families were recruited through referring clinicians from local public hospitals between 2021 and 2024. The inclusion criteria for the probands, which encompassed diverse neurological symptoms, included but were not limited to epilepsy, dyskinesia, ataxia or spinocerebellar ataxia, neuropathy, spastic paraplegia, neurodegeneration or Parkinsonism. In this study, pediatric probands with NDs often presented with neurodevelopmental delay, which accounted for 42.6% of the total cohort (n = 132/310). Figure 1 depicts an overview of the parallel sequencing, variant calling and interpretation pipeline using srGS and lrGS while the details are described in the Methods section.

Fig. 1. Overview of the analytic workflow.

A schematic diagram illustrating the sequencing, variant calling and variant prioritization pipeline. HKGP, Hong Kong Genome Project; SNV, single nucleotide variant; indel, small insertions-deletions; SV, structural variant; CNV, copy number variant; STR, short tandem repeat; gnomAD, Genome Aggregation Database; GS, genome sequencing; MAF, minor allele frequency; QC, quality control.

Cohort demographics

This cohort included 310 affected families with undiagnosed NDs from the HKGP, with a total of 850 individuals sequenced collectively. Amongst the 310 probands, 168 are males and 142 are females, aged from 2 days to 73 years, with 92.6% being of Chinese ethnicity. The cohort included 47 singletons (15.2%), 11 duos (3.5%), 225 trios (72.6%) and 27 other family structures (8.7%). Prior to recruitment into the HKGP, 166 probands had undergone conventional genetic testing (including karyotyping, array comparative genomic hybridization, multiplex ligation-dependent probe amplification and Fragile X test), targeted sequencing and/or ES (Supplementary Fig. 1). The 310 probands were categorized into seven disease groups based on their overall phenotypes (Fig. 2A). Family histories of matched phenotypes were recorded in 28% of the families, including 15% with first-degree relatives.

Fig. 2. Diagnostic summary of the cohort.

A Diagnostic yield by disease. B Variant detection performance of srGS and lrGS. PKD, paroxysmal kinesigenic dyskinesia.

Overall diagnoses in the cohort

The WGS findings from all individuals are provided in Supplementary Data 2. For analysis, we primarily focused on diagnosing genetic causes of neurological symptoms such as epilepsy, neuropathy, and movement disorders. Variants affecting genes associated exclusively with isolated developmental delay and intellectual disability were not reported. Diagnostic findings were identified in 70 out of 310 families (77 pathogenic/likely pathogenic (P/LP) variants), resulting in an overall diagnostic yield of 22.6%. Additionally, nine variants of uncertain significance (VUSs) were identified, with affected individuals exhibiting phenotypes consistent with the associated diseases. Across the seven ND categories, there were no significant differences in the diagnostic yield across them (p = 0.75) (Fig. 2A). The diagnostic yield of probands with pediatric-onset diseases (23.7%) was not significantly different from that of adult-onset diseases (20%) (p = 0.74). Additionally, there was no significant difference in diagnostic yield across all family structures, which ranged from 18% to 28% (p = 0.90). In 55 probands diagnosed with dominant NDs, we identified 38 de novo variants. Seven of the probands inherited the variant from affected parents. Additionally, ten probands inherited P/LP variants from unaffected parents, indicating a potentially reduced penetrance. In these cases, diagnoses were confirmed only if the probands exhibited consistent phenotypes aligned with the gene-associated disease. Of the 166 probands who underwent previous genetic testing, 40 (24%) received pathogenic diagnoses within this cohort (Supplementary Fig. 1 and Supplementary Data 2).

Evaluation of variant calling performances of srGS and lrGS

Here, we conducted a comprehensive analysis of the variant calling performance of ONT lrGS compared to Illumina srGS. Both pipelines met standard quality‑control metrics comparable to published recommendations (Supplementary Data 3)^22,23. To benchmark each platform, we utilized the GIAB sample NA24631 (HG005) and the NIST v.4.2.1 benchmark regions, which include clinically relevant, difficult-to-map areas. Our benchmarking results indicate that both technologies performed well, with ONT lrGS demonstrating superior performance in SNV calling (srGS F1 score: 0.991; lrGS F1 score: 0.997), while Illumina srGS outperformed lrGS in indel (<50 bp) calling (srGS F1 score: 0.991; lrGS F1 score: 0.916) (Supplementary Data 4). After removing variants with GQ ≤ 20, we found that lrGS identified a distinct set of variants (average ~134,000 across the whole genome and ~900 in coding regions), consistent with previous findings¹⁷. Notably, since we omitted the VQSR step in srGS to detect mosaic variants with lower quality, we observed that the number of unique variants was eight times greater by srGS calling. However, this resulted in only about 300 rare coding variants requiring clinical analysis, which is fewer than the ~500 rare coding variants identified by lrGS (Table 1, Supplementary Fig. 2A).

Table 1.

Comparison of SNV/indel calling performance between srGS and lrGS

Variant type	Illumina srGS only			ONT lrGS only			Intersection
	All	Coding	Rare coding	All	Coding	Rare coding	All	Coding	Rare coding
Total SNV and indel	1,215,644 ± 69,620	2995 ± 261	318 ± 59	134,398 ± 13,071	933 ± 115	498 ± 73	3,783,284 ± 63,781	11,044 ± 244	420 ± 42
SNV	483,690 ± 42,460	2421 ± 229	255 ± 54	121,563 ± 11,883	905 ± 112	481 ± 72	3,532,616 ± 47,518	10,750 ± 208	410 ± 41
Indel	731,954 ± 36,752	574 ± 52	63 ± 12	12,835 ± 2544	28 ± 8	17 ± 5	250,667 ± 27,375	294 ± 36	10 ± 4

GQ, genotype quality; srGS, short-read genome sequencing; lrGS, long-read genome sequencing; ONT, Oxford Nanopore Technologies; SNVs, single-nucleotide variants. Values are presented as mean ± standard deviation and rounded to nearest integer.

A more pronounced disparity between srGS and lrGS was noted in the total number of SVs detected, with lrGS identifying ten times more deletions and insertions (Table 2). Furthermore, lrGS exhibited a reduction in false positive calls for duplications and inversions. After excluding common SVs documented in gnomAD SVs v2.1, the mean number of rare coding SVs decreased to approximately 100 for both srGS and lrGS. Subsequent removal of common SVs with minor allele frequency (MAF) ≥ 1% detected in the local population (HKGP participants) further reduced the count to six, underscoring the importance of population-specific genomic data in enhancing the diagnostic utility of GS (Table 2, Supplementary Fig. 2B)²⁴.

Table 2.

Comparison of SV calling performance between srGS and lrGS

Variant type	Illumina srGS only			ONT lrGS only			Intersection
	All	Rare coding	Rare coding + MAF <1% in HKGP	All	Rare coding	Rare coding + MAF < 1% in HKGP	All	Rare coding	Rare coding + MAF <1% in HKGP
Total	1503 ± 352	93 ± 36	7 ± 3	116,856 ± 1095	75 ± 8	6 ± 3	5619 ± 603	1 ± 2	0.08 ± 0.4
DEL	598 ± 210	20 ± 9	4 ± 3	5512 ± 415	34 ± 5	2 ± 1	3765 ± 291	0.2 ± 0.6	0.06 ± 0.3
DUP	486 ± 54	32 ± 13	3 ± 2	5 ± 2	1 ± 1	0.4 ± 0.7	5 ± 2	0.04 ± 0.25	0.02 ± 0.2
INS	205 ± 104	0.3 ± 0.5	0.1 ± 0.3	11,261 ± 699	35 ± 4	4 ± 2	1824 ± 331	0.2 ± 0.4	0.00 ± 0.05
INV	215 ± 45	40 ± 16	1 ± 1	11 ± 3	4 ± 2	0.3 ± 0.6	25 ± 6	0.9 ± 1.4	0.00 ± 0.00

HKGP Hong Kong Genome Project, srGS short-read genome sequencing, lrGS long-read genome sequencing, ONT Oxford Nanopore Technologies, SVs structural variants, DEL deletion, DUP duplication, INS insertion, INV inversion. Values are presented as mean ± standard deviation and rounded to nearest integer.

Evaluation of lrGS performance in homopolymeric region

A comparison analysis between High accuracy (HAC) and Super accurate (SUP) basecalling models was conducted in three samples to evaluate the performance of lrGS in homopolymeric regions, using srGS results as reference (Supplementary Fig. 5). The HAC model requires fewer computational resources and enables faster basecalling, whereas the SUP model is more computationally intensive but provides higher accuracy. These basecalling models are independent of the methylation-calling model and do not affect its performance. Both HAC and SUP models showed a sensitivity of around 85% when compared to srGS, with a remarkable drop beyond size 8bp. To assess the potential risk of lrGS missing SNVs or indels in ND individuals, we analyzed pathogenic or likely pathogenic variants involving homopolymeric regions (≥5 bp) reported in ClinVar. Of the 1,479 genes included in our gene list for NDs, 12.3% of genes (n = 182/1,479) were found to contain 391 known pathogenic or likely pathogenic SNVs (Supplementary Data 5), underscoring the necessity of accurately resolving homopolymeric sequences.

Among the 70 solved cases in the current cohort, Illumina srGS alone diagnosed 88.6% (n = 62/70), while the remaining 11.4% (n = 8/70) were diagnosed solely by ONT lrGS (Supplementary Data 2). This latter group included seven probands with complex SVs and STRs, as well as one singleton proband requiring phasing information. Conversely, ONT lrGS diagnosed 91.4% (n = 64/70) of the 70 probands, with 8.6% (n = 6/70) diagnosed only by Illumina srGS, comprising six probands with pathogenic variants in homopolymeric regions. In total, 77 variants were detected across the 70 families. The variant detection performance of lrGS and srGS was comparable at 92.2% (n = 71/77) and 96.1% (n = 74/77), respectively.

Comparison of diagnostic performance between srGS and lrGS

Among the 77 variants, 64 (83.1%) were identified by both sequencing technologies (Fig. 2B). In contrast, the remaining 13 variants (16.9%) showed differing levels of discordance between srGS and lrGS (Table 3). For instance, lrGS enabled the confirmation of three large repeat expansions in NOTCH2NLC [MIM: 618025] with phased DNA methylation profiles, which is an analysis not feasible with srGS.

Table 3.

Discordant variants detected by srGS and lrGS

Diagnosis confirmed by	Proband	Gene/ variant	Variant detection performance	Description
srGS	HKGP005330-1	NM_145239.3(PRRT2):c.649dup p.(Arg217ProfsTer8)	Not detected by lrGS	Homopolymeric region lrGS: genotype qualities (5); allelic depth (10;7)
	HKGP008340-1	NM_145239.3(PRRT2):c.649dup p.(Arg217ProfsTer8)	Not detected by lrGS	Homopolymeric region lrGS: variant not called
	HKGP012805-1	NM_145239.3(PRRT2):c.649dup p.(Arg217ProfsTer8)	Not detected by lrGS	Homopolymeric region lrGS: genotype qualities (7); allelic depth (12;13)
	HKGP002559-1	NM_145239.3(PRRT2):c.649dup p.(Arg217ProfsTer8)	Not detected by lrGS	Homopolymeric region lrGS: genotype qualities (6); allelic depth (7;8)
	HKGP005191-1	NM_005859.5(PURA):c.159dup p.(Leu54AlafsTer147)	Not detected by lrGS	Homopolymeric region lrGS: variant not called
	HKGP002404-1	NM_004333.6(BRAF):c.770 A > G p.(Gln257Arg)	Not detected by lrGS	Homopolymeric region lrGS: variant not called
lrGS	HKGP000204-1	NM_001364013.2(NOTCH2NLC):c.[-164GGC[15]];[-164GGC[127]]	Not detected by srGS	Large repeat expansion with allele-specific methylation data
	HKGP000245-2	NM_001364013.2(NOTCH2NLC):c.[-164GGC[21]];[-164GGC[108]]	Not detected by srGS	Large repeat expansion with allele-specific methylation data
	HKGP000503-1	NM_001364013.2(NOTCH2NLC):c.[-164GGC[23]];[-164GGC[96]]	Not detected by srGS	Large repeat expansion with allele-specific methylation data
	HKGP002555-1	NM_024577.4:c.3676-8 G > A p.?; NM_024577.4:c.3143 T > C p.(Leu1048Pro)	Detected by both srGS and lrGS. Phasing information not provided by srGS	Phasing of variants in recessive gene in singleton-based analysis
	HKGP001566-1	NM_001940.4(ATN1):c.[1449CAG[21];1449CAG[73]]	More precisely detected by lrGS	Large repeat expansion
	HKGP013149-1	NR_002717.3(ATXN8OS):n.[1103CTG[9];CTG[310]]	More precisely detected by lrGS	Large repeat expansion
	HKGP000668-1	NC_000002.12(SCN1A):g.166073530_166073531insN[351]	More precisely detected by lrGS	Mobile element insertion
	HKGP002750-1	seq[GRCh38] der(10)t(5;10)(p14.1;q26.3)	More precisely detected by lrGS	Complex SV (chromosomal translocation)

srGS, short-read genome sequencing; lrGS, long-read genome sequencing; SV, structural variant.

The current study confirmed a total of six small variants, including five indels located within homopolymeric regions of PURA and PRRT2, and one SNV within a homopolymeric region of BRAF, by srGS alone. lrGS failed to detect the PURA (Fig. 3A and Supplementary Fig. 6A), BRAF (Supplementary Fig. 3E and 6L) and one of the PRRT2 (Supplementary Fig. 3D and 6K) variants. Manual inspection of the variant in Integrative Genome Viewer also failed to identify the variants (Supplementary Fig. 3A–C and 6H–J).

Fig. 3. Illustrative variants highlighting the strengths and limitations of ONT lrGS.

A Visualization of srGS and lrGS results in the Integrative Genome Viewer (IGV) showed that lrGS failed to detect a 1bp duplication in a homopolymeric region within PURA. B lrGS successfully confirmed the precise breakpoints and identified a de novo pathogenic unbalanced translocation between 5p14.1 and 10q26.3, resulting in a 25 Mb terminal duplication of 5p15.33, encompassing 80 protein-coding genes, and a 3.48 Mb terminal deletion of 10q26.3, encompassing 30 protein-coding genes. srGS, short-read genome sequencing; lrGS, long-read genome sequencing.

In a proband HKGP002750-1, with significant global developmental delay, both srGS and lrGS identified a de novo unbalanced translocation. This event involved a 25 Mb terminal duplication of 5p15.33, encompassing 80 protein-coding genes, and a 3.48 Mb terminal deletion of 10q26.3, encompassing 30 protein-coding genes [t(5;10)(p14.1;q26.3)] (Fig. 3B and Supplementary Fig. 6B)²⁵. Both CNVs were classified as pathogenic according to the ACMG/ClinGen technical standards for CNV interpretation²⁶. In srGS, Manta failed to resolve the distinct breakpoint in chromosome 10. Hence, srGS alone was unable to pinpoint whether any coding genes were affected at the rearrangement sites. On the other hand, lrGS clearly showed the precise breakpoints located in intergenic regions. The enhanced resolution of complex SVs by lrGS improves SV calling performance substantially.

The ability of srGS to precisely resolve large STR expansions is limited by its short-read length. For example, in proband HKGP001566-1, presenting with epilepsy and mild intellectual disability, ExpansionHunter identified a pathogenic repeat expansion of 84 repeats (95% CI: 64–113) in ATN1 [MIM: 607462] (Supplementary Fig. 4A and 6M). The confidence interval of this estimate spans the repeat ranges associated with both adult-onset (48–65 repeats) and juvenile-onset (≥65 repeats) forms of ATN1-associated dentatorubral-pallidoluysian atrophy (DRPLA) [MIM: 125370], thereby introducing uncertainty in the prognosis²⁷. In contrast, lrGS identified 73 repeats, with greater confidence enabled by reads spanning the entire repeat region, thereby supporting a diagnosis of juvenile-onset DRPLA.

DNA methylation analysis has become an important tool in the diagnosis of rare diseases. Long-read technologies enable direct detection of base modification with haplotype-resolved data. Using phased methylation analysis with lrGS, three probands, HKGP000204-1, HKGP000245-2 and HKGP000503-1, were confirmed to have neuronal intranuclear inclusion disease (NIID) [MIM: 603472]. The diagnosis required the identification of a pathogenic repeat expansion (≥60 repeats) in the 5’-UTR of NOTCH2NLC, in addition to hypomethylation of the expanded region^28,29. (Fig. 4A) In proband HKGP000204-1, methylation analysis revealed hypomethylation of both haplotypes at the NOTCH2NLC locus (Fig. 4B and Supplementary Fig. 6C), supporting the molecular diagnosis of NIID. Conversely, in another proband, HKGP013204-1, who did not have phenotypes of NIID despite the detection of expansion (720 repeats), methylation analysis showed that the haplotype carrying the expansion was hypermethylated (Fig. 4C and Supplementary Fig. 6D). Previous studies have suggested that hypermethylation suppresses the transcription of the toxic expanded allele^28,29. As a result, the diagnosis of NIID was not supported in this proband.

Fig. 4. lrGS provides allele-specific methylation data for diagnosing NIID.

A Proposed molecular mechanism of NIID. Pathogenic STR in the 5’ UTR of the NOTCH2NLC gene leads to the production of a toxic polyG peptide in the upstream open reading frame. Hypermethylation of the expanded allele silences transcription, preventing expression of toxic peptide. B lrGS analysis revealed hypomethylation of the expanded NOTCH2NLC alleles in the proband, confirming the diagnosis of NIID. (C) Conversely, lrGS detected hypermethylation of the expanded NOTCH2NLC allele, leading to the rejection of the NIID diagnosis. NIID, neuronal intranuclear inclusion disease; lrGS, long-read genome sequencing. Visualization of allele-specific methylation patterns was performed using IGV and modbamtools.

Costing analysis

Leveraging the population-scale HKGP with an annual sequencing throughput of over 10,000 GS, the large sample volume supports a reagent rental approach for both srGS and lrGS sequencing platforms, in which the cost of the sequencer was built into that of the reagent kits. The estimated cost per sample was USD1,545.97 for srGS on NovaSeq^TM 6000, compared to USD1,580.00 for lrGS on PromethION 24. The cost for each stage of the GS for each sequencing system is shown in Supplementary Data 6.

The sequencing consumables were the most expensive component for srGS, accounting for 72.2% of total costs (before overheads). The other major cost contributor was genome curation and reporting, accounting for 11.1% of the total costs. For lrGS, sequencing consumables were also the most expensive component, constituting 61.8% of total costs (before overheads). Data storage (9.6%), library preparation (9.5%) and genome curation and reporting (8.2%) were the other major cost contributors to the total costs. While sample reception, DNA extraction and quality control (QC) were identical in both srGS and lrGS, the wet-lab (sequencing library preparation and sequencing) portion of the srGS workflow was more established, streamlined and compatible with automation, reducing manual intervention and keeping labor costs lower ($192.00, 14.9% of the total costs for srGS) compared to lrGS that relied primarily on manual manipulation ($285.27, 21.7% of the total costs).

The full results of the sensitivity analysis of srGS and lrGS are presented in Supplementary Data 7. As the ONT lrGS workflow stabilizes and becomes more efficient over time, it is expected that the manual intervention in the library preparation and QC, data processing and storage would decrease, matching the levels seen in srGS. AI-based variant prioritization and interpretation tools may benefit the curation process for both srGS and lrGS workflows, reducing the manual input and increasing efficiency. The first scenario analysis compared the effects of increasing or decreasing manual involvement in the most laborious steps of the workflows by up to 50% while keeping all other components constant, resulting in a change in total srGS costs of 5.2% and lrGS costs of 10.8%.

Innovations and enhancements in sequencing technologies have continued to bring the cost of sequencing down. The second scenario analysis compared the current costs to the latest Illumina NovaSeq^TM X Plus system for srGS and estimated future costs of lrGS, where the sequencing reagent cost would decrease by 50% while the other components remained the same. The future total costs would be lowered by 34.8% to $1008.15 for srGS and 29.6% to $1112.65 for lrGS, respectively, approaching the realization of the $1000 genome milestone (Supplementary Data 7).

Discussion

We conducted a comprehensive comparison between srGS and lrGS in a large cohort of probands with NDs (n = 310), combined with a detailed costing analysis. While previous investigations consistently indicated that lrGS could enhance diagnostic yield, they were often smaller in scale or only included ES/srGS-negative cases (Supplementary Data 8)^14,16,17,21. This study underscores the importance of understanding specific strengths and limitations of the two sequencing methods and provides valuable insights in selecting the appropriate sequencing platform for diagnosing NDs.

Regarding sample requirements, consistent DNA quality is a critical determinant of downstream sequencing performance for both srGS and lrGS. We use the QIAsymphony platform for automated DNA extraction from blood samples to ensure consistent DNA integrity and purity. This helps support high-quality sequencing output for downstream analysis.

lrGS identified approximately 57% more rare coding SNVs and indels than srGS (Table 1). This improvement may be attributed to the enhanced ability of lrGS in detecting variants in difficult-to-map regions^30–32. Pathogenic variants located in these regions have traditionally been considered technically challenging to detect using srGS³⁰. Studies have shown that lrGS resolved SNVs in these regions with higher accuracy^32–34.

On the other hand, homopolymeric regions remain a major challenge for the ONT lrGS platform^17,18,32,35. It is important to note that all six SNVs that lrGS failed to detect in the current study were located within homopolymeric regions, while srGS accurately identified them. Further analysis on pathogenic and likely pathogenic variants involving homopolymeric regions (≥5bp)⁶, reported in ClinVar indicated that a notable proportion of ND-associated genes harbor variants within these regions. Collectively, these observations indicate that, at present, lrGS cannot fully replace srGS for genetic diagnosis of NDs yet. That being said, lrGS performance is improving rapidly, with substantial improvements in indel calling accuracy using the latest R10 flow cells and enhancements in base-calling models and variant calling tools^19,24,36. In this study, we observed improved performance with the SUP model over the HAC model, along with increased computational resource requirements, consistent with previous reports^24,33. Recent advancements in variant callers and the combined use of different callers have further reduced genotyping errors^34,37.

lrGS outperformed srGS in calling complex SVs, owing to the long-read length and improved mapping in difficult-to-map genomic regions. Most SVs identified by srGS failed to resolve precise breakpoints and required further investigation and validation using lrGS^13,38. The ability of lrGS to resolve complex SVs provides valuable diagnostic insights, particularly in copy-number-neutral SVs where one of the breakpoints disrupts a morbid gene. Constrained by the short-read length, srGS struggles to characterize many STR expansions accurately^39–42. In contrast, lrGS overcomes the above challenges and allows haplotype-resolved DNA methylation profile, as illustrated in the ATN1 and NOTCH2NLC diagnoses^28,29,39,43. The overall allele-specific methylation profiling also provides diagnostic and prognostic values in other STR expansion disorders^44–46. Recent work by our group and others has shown that the single ONT-lrGS workflow enables comprehensive characterization of the 3.3 kb D4Z4 repeat unit that is associated with facioscapulohumeral muscular dystrophy (unpublished data). Additionally, global genomic DNA methylation patterns can provide insights into the underlying pathogenic mechanisms and have emerged as a powerful diagnostic tool^47–49. These remarkable capabilities of lrGS play an important role in understanding NDs. Lastly, although we included predominantly trios in the current study, lrGS offers the capability to phase variants in cases where parental samples are not available. Although 77.4% of patients remained undiagnosed in the present study, the comprehensive data generated by lrGS provide opportunities for future reanalysis as additional disease-causing genes and evidence are discovered, potentially leading to new diagnoses.

A lack of large reference population databases like the Genome Aggregation Database (gnomAD) for lrGS data is one of the major obstacles for efficient variant filtering. Projects like the Human Pangenome Reference Consortium⁵⁰, Genomics England and other studies are building long-read sequencing datasets^34,51. Contributing to the global effort, HKGP is assembling a lrGS database of the local population, including the data described in the current study. As more long-read sequencing data become available from large-scale population resources, comprehensive variant analysis is anticipated to be streamlined in the future.

One of the most cited limitations of long-read sequencing is the high cost compared to short-read sequencing^31,52,53. This study compared the cost of srGS to lrGS in individuals with NDs using a detailed costing approach. Our results show that the unit costs (per sample) of srGS and lrGS are comparable, at USD1,545.97 (Illumina NovaSeq™ 6000) and USD1,580.00 (ONT PromethION 24), respectively, which are generally lower than those reported and perceived previously for both sequencing platforms. Consistent with previous reports for srGS^54–57, our findings indicate that the main cost driver for both srGS and lrGS is the sequencing consumables, accounting for 72.2% and 61.8% of total costs.

Innovations and enhancements in sequencing technologies have brought down the cost of sequencing. Compared to the sequencers used in the previous reports^54–56, the sequencing capacity of the new Illumina NovaSeq^TM X Plus instrument has increased by at least 3-fold. Based on our estimates, the new NovaSeq^TM X Plus system reduces the cost of srGS by more than 50%. Relative to sequencing laboratories with low-to-medium sample volumes that typically process hundreds to a few thousand samples annually, the high-throughput scale provided by population-scale genome projects like HKGP enables more efficient resource allocation and economies of scale, thereby achieving lower average sequencing costs. Although the cost of sequencing consumables varies across countries and may not be directly generalizable, laboratories processing high sample volumes typically benefit from volume-based discounts (ranging from 10-20%) and maintain more consistent turnaround time, contributing to enhanced operational efficiency. The costs presented in this study were estimated in the HKGP setting and may not be directly generalizable to single-center settings. Nonetheless, the estimates provide a practical reference for diagnostic laboratories planning genomic testing services for inherited NDs.

It should be noted that, in addition to ONT, PacBio is another major long-read genome sequencing platform, with its own strengths and limitations. In this study, we evaluated only the ONT lrGS platform. Some of the findings could be applicable to the PacBio platform, particularly regarding its advantages in resolving complex SVs and STRs, as well as its capability to phase variants in singleton samples and detect DNA methylation. However, remarkable differences exist between PacBio and ONT concerning variant calling accuracy, cost and throughput. In practice, few groups have the resources to evaluate both technologies in parallel across large clinical cohorts. Based on published studies, PacBio HiFi generally provides higher read accuracy, in some settings even exceeding that of Illumina srGS, which enables more accurate variant calling in challenging regions such as homopolymers and may help overcome the key limitation of ONT lrGS^21,38. This raises the possibility that PacBio could serve as an alternative to srGS in some diagnostic settings¹⁶. However, PacBio also has several limitations, including lower throughput and higher sequencing cost per genome compared with Illumina⁵⁸. Given the respective advantages and disadvantages of lrGS platforms (ONT and PacBio) and srGS (Illumina), clinical laboratories may select the most appropriate technology according to their diagnostic objectives, available resources, and the disease categories being investigated.

One limitation of this study is that the observed performance of srGS and lrGS depends, in part, on the versions of the bioinformatic tools used, and performance may change as these tools are updated. Another limitation is that our current analytical pipeline does not systematically capture promoter variants or intronic SVs with potential regulatory or splicing effects.

In summary, we demonstrated that while srGS excels in detecting variants near homopolymeric regions, lrGS offers better resolution for complex SVs, large STRs, and uniquely enables the simultaneous analysis of DNA methylation and variant phasing. Although lrGS is not yet capable of fully replacing srGS, our findings highlight its significant potential in diagnosing specific neurological diseases enriched for complex variants. Given that ONT lrGS is continuously improving its sequencing technology and computational methods to enhance its performance in homopolymeric regions, we anticipate that lrGS could become a first-tier test for genetic NDs in the near future.

Methods

Recruitment, sample collection and gDNA extraction

This study was conducted in compliance with all relevant ethical regulations and the principles of the Declaration of Helsinki. Ethics approvals were granted from research ethics committees in Hong Kong, comprising the Central Institutional Review Board (IRB) (HKGP-2021-001, HKGP2022-001) and The University of Hong Kong/Hospital Authority Hong Kong West Cluster (UW 21-413, UW 23-289). All participants received pre-test genetic counseling and provided informed written consents following the unique three-tier consent and assent model designed by the HKGI⁵⁹. Informed consent was obtained from all individuals involved in this study. Clinical details of the probands were collected in a standardized format based on Human Phenotype Ontology (HPO) terms, along with family history, previous genetic testing and clinical testing information. The HKGP follows in-house standard operating protocols for whole blood sample collection and transfer from Partnering Centers to the Hong Kong Genome Institute (HKGI) laboratory. Genomic DNA (gDNA) was extracted from 600 µL of whole blood using the QIAsymphony SP system and QIAsymphony SP DNA Mini Kit (Qiagen)⁶⁰.

Illumina short-read GS and data analysis

The detailed workflows for sequencing and data analysis of srGS have previously been described⁶⁰. Briefly, 1 µg of gDNA was used in PCR-free GS library construction using the KAPA HyperPlus Kit and KAPA Unique Dual-Indexed adapter (Roche), following the instructions provided by the manufacturer. The srGS libraries were sequenced on the Illumina NovaSeq^TM 6000 sequencer (Illumina), to a mean coverage of ≥30X. Then, sequence reads were aligned to the GRCh38 reference genome by BWA (v0.7.17), and duplicates were removed by Picard (v2.27.4). Variant calling was performed by GATK HaplotypeCaller (v4.2.6.1), INSurVeyor (v1.1.0), Manta (v1.6.0), CNVkit (v0.9.9), and ExpansionHunter (v3.1.2). Variant Quality Score Recalibration (VQSR) was not implemented to avoid the removal of clinically important mosaic variants with low genotype quality (GQ) and read depth. Variants were annotated using Ensembl Variant Effect Predictor (VEP) (v110.1), BCFtools (v1.13), and in-house tools⁶⁰.

Oxford Nanopore Technologies long-read GS and data analysis

In brief, 2.5 µg of gDNA was sheared to an average DNA fragment size of 10–15 kb using the HG-600 Geno/Grinder^® 2010 Automated Tissue Homogenizer and Cell Lyser (Cole-Parmer SPEX SamplePrep). Sequencing libraries were prepared using the Ligation Sequencing Kit V14 (Oxford Nanopore Technologies, SQK-LSK114) and NEBNext^® Companion Module v2 (New England Biolabs, M6630), following the instructions provided by the manufacturer. The lrGS libraries were sequenced using the R10.4.1 flow cells on the PromethION 24 system to achieve a mean coverage of ≥30X. Dorado (v7.3.9) was used for real-time base-calling in the high accuracy (HAC) model⁶⁰. Reads with a qscore of ≥10 were retained by Samtools (v1.16.1)⁶¹. Pass-filter reads were processed through the EPI2ME Labs Human Variation Workflow (v1.8.3). In the workflow, reads were aligned to the GRCh38 reference genome using Minimap2 (v2.24 r1122)⁶². Variant calling was performed by Clair3 (v1.0.4)³⁶; Sniffles2 (v2.0.1)⁶³; QDNAseq (v1.34.0)⁶⁴ with a bin size of 100 kb; and Straglr (v1.4.1)⁴¹. NanoPlot (v1.41.6) was used to assess sequencing data quality⁶⁵. Variants were annotated with Ensembl VEP (V110.1)⁶⁶. Variant phasing was performed by WhatsHap⁶⁷, while DNA methylation was detected using the default wf-human-variation pipeline.

Benchmarking and evaluation of the performance of variant calling

The performance of the pipeline was validated using the HG005 reference sample from the Genome in a Bottle (GIAB) project. For SNVs and indels <50 bp, we used the benchmark regions defined by NIST v.4.2.1 for HG005⁶⁸, RTG Tools version 3.12.1, to determine whether the lrGS and srGS variant calls should be evaluated. For the truth variant callset, we use the corresponding VCF file provided by NIST v.4.2.1 for HG005⁶⁹. To determine a variant seen in our query callsets to be counted as a true positive, we used the definition of “genotype match” provided by Krusche et al. (2019)⁷⁰, which requires that the query variant has to match the truth variant in genotype and allele. This definition is implemented in RTG Tools version 3.12.1, which was used in this study for the calculation.

Subsequently, the variant detection performance of srGS and lrGS was compared for SNVs and SVs using trios. Variants with a GQ < 20 were excluded from the analysis. Rare coding SNVs and indels are defined as MAF < 1% in gnomAD (both v2.1.1 and v3.1.2) with high or moderate impacts predicted by Ensembl VEP. Additionally, rare coding SNVs and indels with Mendelian inheritance error (MIE) were counted. Variants were excluded based on the following criteria: (1) the proband had fewer than 5 reference reads; (2) fewer than 5 alternative reads; (3) an allelic ratio below 0.2 (0.3 for indels); or (4) any proband or parent had an allelic depth of less than 10 (less than 15 for indels). For SVs, variants that failed SV caller-specific quality filters were excluded. Rare coding SVs are defined as MAF < 1% in gnomAD SVs v2.1 and affect ≥1 exon of any MANE-select transcripts. SV intersection is performed through the sveval package^17,71.

Variant prioritization pipeline and tertiary analysis

An automated analytic pipeline was implemented to filter and prioritize rare variants that can potentially explain the proband’s and/or the family’s primary clinical indication. Variants were considered as benign and removed if the MAF was ≥1% in any subpopulations on gnomAD v2.1.1 and v3.1.2 for SNVs and indels, or Database of Genomic Variants (DGV) and gnomAD SVs v2.1 for CNVs and SVs. SNVs and indels with low predicted impact were defined by the following criteria: they were non-coding variants, had a SpliceAI score <0.2 and lacked pathogenic or likely pathogenic submissions on ClinVar. These variants were excluded from further analysis. Implementation of Mastermind, a genomic association search engine, allowed automatic search for related publications at the variant level⁷². In addition, the number of genes encompassed by each CNV or SV event was annotated. For SVs, an in-house pipeline was employed to annotate breakpoints affecting protein-coding genes. Deletion, duplication and insertion events that did not overlap with any coding exons were excluded from analysis. Additionally, inversions and translocations with both breakpoints located in intergenic regions were also excluded. The variants were then prioritized using a phenotype-driven approach by selecting panels related to neurological symptoms from the Genomics England PanelApp and PanelApp Australia. Genes only associated with isolated NDD are not selected. (Supplementary Data 1).

Blinded tertiary analyses for srGS and lrGS were performed independently by two groups of equally experienced genome curators in-parallel. Candidate variants and their pathogenicity were classified using up-to-date American College of Medical Genetics (ACMG) guidelines and specifications from ClinGen Variant Curation Expert Panels.

Costing analysis

A detailed costing study was performed using the srGS and lrGS workflow of the HKGP during a six-month period between January and June 2024. The costing analysis was carried out in line with the methods outlined by Drummond et al. ⁷³, using a cost estimation method that identifies all resources required by the HKGI laboratory to conduct genome sequencing, data analytics and genome curation to yield a molecular diagnosis.

Data was collected on resource utilization and unit cost for each step in the workflow, consisting of equipment, consumables, labor and overheads. Item costs were calculated in 2024 values by multiplying the volume of use by the unit cost. Equipment costs were calculated based on prorated usage, which was based on a 5-year depreciation schedule and maintenance costs. Consumables costs included laboratory supplies and reagents costs. Labor costs were calculated by multiplying the time (in minutes) for each task by the midpoints of staff salary ranges. A working year was assumed to be 45 weeks, and a working week was assumed to be 40 h. An overhead cost of 20% was added to the total costs, which include items such as general administration, electricity, cleaning and rent. Cost data were reported in Hong Kong dollars (HKD) and converted to US dollars (USD) (exchange rate of HKD1.00 = USD0.1282; 2024 average rate).

The srGS costs were based on the Illumina NovaSeq^TM 6000 sequencing platform, where 24 sequencing libraries were batched and sequenced in one S4 flow cell. The lrGS costs were based on the ONT PromethION 24 sequencing platform, where one sequencing library was sequenced in one R10.4.1 flow cell. For both srGS and lrGS sequencing platforms, a reagent rental approach was employed, in which the cost of the sequencer was built into that of the reagent kits.

Sensitivity analysis was performed to evaluate the effects of alternative cost assumptions on the overall costs. Two major components in GS were varied: (1) labor cost, and (2) sequencing reagent cost, where percentage ranges were applied to assess the impact on the total cost estimates.

Statistical analysis

Differences in diagnostic yield between subgroups with respect to clinical and demographic characteristics were compared using logistic regression. The significance level was set at P < 0.05 for all analysis.

Author contributions

D.L.S.C., H.Y.F.T., and A.H.Y.T. contributed equally to this work. Conceptualization: B.H.Y.C.; Data curation: D.L.S.C., H.Y.F.T., A.H.Y.T., W.M.; S.T.F.C, S.P.Y.H, A.M.C.X., and D.O.; Formal analysis: D.L.S.C., H.Y.F.T., A.H.Y.T., W.M., D.L.H.Y., C.C.Y.C., S.C.Y.Y., and J.S.L.K.; Investigation: D.L.S.C., H.Y.F.T., A.H.Y.T., W.M., A.T.W.C., and B.H.Y.C.; Methodology: D.L.S.C., H.Y.F.T., A.H.Y.T., W.M.; Resources: C.S.L., H.M.L., S.Y.Y.P., S.S.W.C., Y.G., L.T.L., N.S.Y.T, C.C.F.L, W.L.W., R.Y.W.L., J.K.N.W.,A.T.W.C., and B.H.Y.C.; Supervision: A.T.W.C. and B.H.Y.C.; Writing: D.L.S.C., H.Y.F.T., A.H.Y.T., and W.M.; All authors reviewed and approved the manuscript.

Data availability

All results are included as supplementary files (Supplementary Information and Supplementary Data), which can be found with this article online.

Code availability

The in-house SV annotation script can be found at https://doi.org/10.5281/zenodo.19492233.

Competing interests

B.H.Y.C. is Deputy Editor of npj Genomic Medicine. B.H.Y.C. was not involved in the journal’s review of, or decisions related to, this manuscript. All other authors declare no competing financial or non-financial interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41525-026-00593-w.