Single-Cell Long-Read RNA Sequencing: PacBio vs Nanopore, Tools, and Practical Advice

Why Single-Cell Long-Read RNA Sequencing?

Standard short-read single-cell RNA-seq (scRNA‑seq) has revolutionised biology by revealing cellular heterogeneity. However, short reads typically cover only one end of a transcript (3′ or 5′), making it impossible to determine which full-length isoform a read came from. This is a critical gap, because alternative splicing means that >95% of multi‑exon genes produce multiple isoforms—on average 14.6 per gene (Bhatia et al., 2025). Many diseases, including cancer and neurodevelopmental disorders, are driven by specific isoform switches that bulk and short-read scRNA‑seq cannot reliably capture.

Long-read sequencing by Pacific Biosciences (PacBio) or Oxford Nanopore Technologies (ONT) reads entire transcript molecules in one pass. When applied at single-cell resolution, it lets you see which isoforms are expressed in each cell, discover novel splice variants, and phase genetic variants to individual transcript molecules for allele‑specific analysis. Until recently, high error rates, low throughput, and cost kept this technology out of reach for many labs. A flurry of recent benchmarks now provides the practical, evidence‑based guidance needed to adopt it effectively.

Platform Comparison: PacBio vs Oxford Nanopore for Single Cells

The table below summarises the key performance differences observed in single‑cell experiments. Data are drawn primarily from Deng et al. (2025), with annotations from other studies where noted.

Important note: You et al. (LongBench) found that when the very same 10x cDNA library was split and sequenced on both PacBio and ONT, the choice of sequencer had minimal impact on pseudo‑bulk gene quantification in single‑cell data. This suggests that many of the differences seen in Deng et al. may stem from the specific library preparation workflows or analysis pipelines used, not solely the sequencing technology itself. For isoform‑level and allele‑specific analyses, however, PacBio’s advantages remain consistent across studies.

Sequencing Quality and Data Characteristics

PacBio HiFi reads deliver base‑level accuracy on par with standard short‑read sequencing, while ONT reads currently have a higher single‑base error rate. In the Deng et al. benchmark, PacBio circular consensus sequencing (CCS) reads showed quality scores around Q30 (one error per 1,000 bases), whereas ONT reads were mostly below Q20 (one error per 100 bases). Despite this, both technologies produce reads that span entire transcripts, which helps compensate for random errors during alignment and quantification.

PacBio recovers full‑length transcripts more completely, but its standard library preparation can lose very short transcripts (<1 kb). Deng et al. reported that nearly all PacBio reads were full‑length when processed with the Longbow tool, while only about half of ONT reads qualified as full‑length. However, the MAS‑ISO‑seq protocol (PacBio) concatenates multiple cDNA molecules into one HiFi read, which are later computationally split; this increases throughput but requires careful bioinformatic segmentation. Zajac et al. found that MAS‑ISO‑seq reads had a higher mapping rate (99.4–99.7%) than Illumina short reads (85–94%) because the protocol removes artefacts like template‑switching oligo (TSO) concatemers. Importantly, the size‑selection step in PacBio’s standard workflow depletes fragments below ~1 kb, meaning very short transcripts may be underrepresented.

ONT single‑cell reads are shorter and show stronger 3′ bias. In LongBench bulk data (You et al., 2025), ONT PCR‑cDNA reads had median lengths of 0.59–0.67 kb—the shortest among the protocols compared. ONT dRNA (not compatible with single cells) produced the longest individual reads and the lowest non‑exonic fraction, while ONT cDNA had the most pronounced drop‑off in coverage for long transcripts.

Gene Detection and Cell Type Identification

Both platforms can recapitulate the major cell types identified by short‑read scRNA‑seq, and they sometimes detect rare populations missed by short reads when cell numbers are limited. In Deng et al., lung and brain cell types from the 500‑cell libraries were successfully recovered by both PacBio and ONT, and the long‑read platforms even identified a rare muscle population that short‑read NGS missed. With larger libraries (5,000 cells), short reads captured cell types more efficiently, but long‑read‑based clustering remained robust.

PacBio tends to detect more genes per cell than ONT, but the practical impact on cell annotation is modest. Deng et al. reported a median of ~900 genes per cell for PacBio versus ~700 for ONT in the 5,000‑cell lung sample. Despite this gap, UMAP embeddings and cluster assignments were broadly similar. LongBench’s single‑cell data (You et al.) showed that pseudo‑bulk gene expression from the same cDNA library correlated well between platforms, and the sequencer choice had minimal effect. Zajac et al. observed that PacBio MAS‑ISO‑seq detected slightly fewer genes per cell than Illumina when using default filtering, but unfiltered PacBio data gave pseudo‑bulk correlations of R = 0.92–0.97 with Illumina, underscoring the importance of bioinformatic choices (see the Bioinformatics Tools section).

Cell barcode (CB) identification—the step that assigns each read to its cell of origin—is more accurate with PacBio. Deng et al. found that PacBio achieved near‑perfect whitelist matching (even better than NGS), while ONT correctly matched less than 60% of CBs. This difference did not prevent cell type identification, but it may affect the recovery of rare cells. Hamraoui et al., working solely with ONT data, showed that tools like BLAZE (used with FLAMES) and wf‑single‑cell achieve high precision (0.91) and recall (0.99) for CB detection against a short‑read truth set, mitigating this issue.

A note on empty droplets and mitochondrial genes: Zajac et al. found that the standard PacBio Iso‑Seq pipeline filters out mitochondrial genes entirely. If mitochondrial content matters for your quality control, you will need to recover this information separately. They recommend using BLAZE for empty droplet detection to avoid including low‑quality cells that would normally be flagged by high mitochondrial percentages.

Isoform Discovery and Validation

This is the core reason many labs turn to long reads—and where PacBio shows a clear advantage in validation rate. In Deng et al., ONT data actually identified a larger number of distinct isoform types per gene, but PacBio produced a far greater proportion of full splice match (FSM) transcripts—meaning the entire read matched a known isoform without gaps or mismatches. More importantly, when the authors experimentally validated novel isoform predictions by RT‑PCR and Sanger sequencing, 11 out of 15 PacBio‑specific novel isoforms were confirmed, compared to only 3 out of 14 ONT‑specific ones. This suggests that while ONT can generate many isoform candidates, many are likely technical artefacts. If your study depends on discovering and reporting new isoforms, this validation gap is critical.

Bioinformatic tools can improve isoform recovery from ONT data, but they cannot fully close the accuracy gap. The Hamraoui et al. tool benchmark on ONT single‑cell data found that Bambu achieved high precision and recall (F1 = 73% on R9.4.1 chemistry) for novel isoform discovery, thanks to its splice‑alignment correction algorithm. On the newer R10.4.1 chemistry, Isosceles (paired with Sicelore 2.1) reached an F1 of 70%. In real tumour data, Bambu and IsoQuant reported the largest numbers of transcripts (93% and 77% of them novel, respectively), indicating high sensitivity but also a risk of false positives. This aligns with the bulk‑data findings from Dong et al., where even the best tools (StringTie2 for precision, bambu for sensitivity) still produced many artifactual isoforms and required orthogonal validation.

Filtering matters: over‑filtering can discard genuine biology. Zajac et al. focused on PacBio’s standard filtering, which removes “low‑coverage / non‑canonical” (LCNC) isoforms. That filter eliminated 45% of reads and 82% of isoforms on average, and it disproportionately affected long transcripts (>4,000 bp). In some samples, filtering caused cells from the two platforms to form separate clusters, driven by genes where up to 60% of isoforms were classified as LCNC. Yet when the authors checked those same molecules in the matched Illumina data, 17% were still counted as legitimate gene expression. This has direct relevance for cancer studies: non‑canonical transcripts can encode tumour‑specific antigens, and removing them wholesale could discard meaningful signals. The take‑home message is to treat default filters as a starting point, validate critical isoforms with an orthogonal method, and consider sample‑specific filtering thresholds.

Differential transcript usage (DTU) remains a hard problem. Dong et al. evaluated five DTU tools on ONT bulk data and found poor concordance with Illumina—only 14.4% of DTU transcripts were detected by both platforms. No tool achieved a satisfactory balance between power and false discovery control. While single‑cell‑specific DTU tools are being developed (FLAMES, Isosceles, scisorseqr), cautious interpretation of DTU results is warranted regardless of platform.

Allele-Specific Expression

Long reads enable phasing of variants and detection of isoform‑level allelic imbalance—capabilities that are largely lost with short reads. In the F1 hybrid mouse experiment by Deng et al., PacBio detected over 120,000 SNP sites in gene regions versus ~70,000 for ONT, and assigned a higher fraction of reads to a specific parental allele (~45% vs <40%). Both platforms showed balanced overall allelic expression at the gene level, but 71 genes displayed no allelic imbalance at the gene level yet showed significant isoform‑level bias—a phenomenon completely invisible to short‑read methods. For example, the Sftpa1 gene had equal total expression from both alleles, but the C57BL/6J allele preferentially expressed the Sftpa1‑201 isoform while the DBA/2C allele favoured Sftpa1‑202. This was confirmed by Sanger sequencing.

LongBench (You et al.) similarly found that long reads phased the majority of heterozygous variants—unlike short reads, which could phase only a small fraction—and that PacBio detected the most allele‑specific expression and splicing events. ONT direct RNA (not available for single cells) identified the highest proportion of known exonic somatic mutations, but for single‑cell applications, PacBio currently offers the richest allele‑resolved data.

Cost and Experimental Design

Budget matters, and the premium for PacBio over ONT is about 30% per cell in the most directly comparable study. Deng et al. calculated per‑cell sequencing costs for the 5,000‑cell libraries as $0.25 for Illumina NGS, $0.52 for ONT, and $0.69 for PacBio. While absolute costs vary by institution and scale, the relative ranking is informative: both long‑read platforms cost roughly 2–3 times more than short‑read scRNA‑seq. LongBench also notes that PacBio is costlier than ONT cDNA. The lower price of ONT may allow you to sequence more cells or deeper, which can partially compensate for lower per‑cell sensitivity.

Cell input size influences the trade‑offs. Deng et al. tested both 500‑cell and 5,000‑cell libraries. For cell‑type discovery and differential expression analysis, they recommend loading more cells (up to ~5,000) rather than maximising per‑cell depth, because having more cells yields more cell‑type‑specific markers. At very shallow per‑cell depths, long‑read correlation with short‑read data drops below short‑read self‑correlation, so some depth is still needed.

You don’t need a matching short‑read library anymore. Hamraoui et al. explicitly state that hybrid tools requiring paired short‑read data (Sicelore, Snuupy) are now outperformed by long‑read‑only methods. That simplifies experimental design and reduces costs further.

Bioinformatics Tools for Single-Cell Long-Read Data

Processing single‑cell long‑read data involves several steps: extracting cell barcodes and unique molecular identifiers (UMIs), demultiplexing reads into cells, aligning to the genome, quantifying genes and isoforms, and filtering artefacts. The Hamraoui et al. benchmark is the most comprehensive evaluation of these tools for ONT data, while Zajac et al. and Dong et al. provide additional insights.

For ONT single‑cell data, wf‑single‑cell (part of the EPI2ME suite; Bhatia et al., 2025) is the most balanced pipeline (Hamraoui et al., 2025). It performed best at UMI error correction (F1 = 0.80 on R9.4.1 chemistry), maintained the lowest quantification error under high UMI duplication, gave the best cell‑type separation, and delivered solid isoform quantification (lowest RMSE). It is recommended when you need reliable gene‑ and isoform‑level accuracy.

Bambu excels at novel isoform discovery and pseudo‑bulk correlation, but lacks UMI correction. In Hamraoui et al., Bambu showed high precision and recall for novel isoforms (F1 = 73% on R9.4.1) and achieved the highest per‑gene correlation with ground truth at pseudo‑bulk level (~0.95). However, its missing UMI correction means gene expression can be overestimated in the presence of PCR duplicates, especially on R9.4.1 chemistry. On the newer, higher‑quality R10.4.1 chemistry, Bambu improved markedly. If your primary aim is cataloguing novel isoforms and you are using R10.4.1 flow cells, Bambu is a strong candidate.

Sicelore 2.1 + Isosceles provides conservative, precise transcript assignment. Sicelore 2.1’s strict exon‑structure matching gave the best read‑to‑transcript assignment (F1 = 0.88). When coupled with Isosceles for novel transcript detection, it reduces false‑positive exons compared to using wf‑single‑cell alone. This combination is useful if you want to minimize false discoveries.

Barcode and UMI extraction tools matter. BLAZE (used with FLAMES) and wf‑single‑cell achieved the best balance of precision and recall for cell barcode detection. For initial demultiplexing, flexiplex (used by Bambu) is also reliable. scNanoGPS is not recommended for large‑scale studies due to long runtime and lower precision.

For differential expression at the gene level, standard tools (DESeq2, edgeR, limma‑voom) work well on long‑read counts (Dong et al.). A depth of about 10 million ONT reads per bulk sample recovered >50% of differentially expressed transcripts (DTEs) found in the full dataset; single‑cell datasets typically require fewer total reads but per‑cell depth will be much lower. At the isoform level, differential usage analysis is still immature, and results should be validated with an independent method.

Handle filtering with care. Zajac et al. showed that PacBio’s default LCNC filter can remove biologically relevant transcripts, especially long ones. For cancer studies, it may be better to start with unfiltered or minimally filtered data and validate interesting candidates experimentally, rather than relying on default cutoffs.

Practical Recommendations: Which Should You Choose?

Your choice depends on your primary research question, budget, and tolerance for false positives.

Choose PacBio (MAS‑ISO‑seq / Kinnex) if:

• Novel isoform discovery and validation are central to your study.
• Allele‑specific expression or haplotype‑phased variants are key outcomes.
• You need high per‑cell gene sensitivity and are willing to pay a moderate premium.
• You plan to follow up candidates with RT‑PCR and need a high validation rate (Deng et al., 11/15 validated).

Choose Oxford Nanopore (PCR‑cDNA) if:

• Your primary goal is cell‑type classification and gene‑level differential expression, with isoform information as a secondary benefit.
• Budget is a major constraint, and you aim to sequence many cells.
• You are comfortable trading some accuracy for lower cost, and you can use tools like wf‑single‑cell or Bambu to maximise data quality.
• You are working with the latest R10.4.1 chemistry, which narrows the accuracy gap for UMI‑aware tasks.

Regardless of platform:

• Use a long‑read‑only pipeline; paired short‑read data is no longer necessary (Hamraoui et al.).
• Include an empty‑droplet detection step (BLAZE) to avoid low‑quality cells, especially on PacBio where mitochondrial genes are not available for QC.
• Be cautious with default isoform filtering; validate critical novel isoforms with an orthogonal approach.
• For differential isoform usage, interpret results conservatively and consider experimental validation.

A quick decision guide:

Limitations and Caveats

The benchmarks described here, while comprehensive, have boundaries:

• Most studies used cancer cell lines or mouse tissues. Performance may differ for primary human samples, low‑input specimens, or tissues with complex isoform landscapes not represented in these models.
• Technology is moving fast. The Hamraoui et al. benchmark includes both R9.4.1 and R10.4.1 ONT chemistries; R10.4.1 improved UMI clustering and base quality, and ongoing updates (e.g., PacBio Revio; Zajac et al., 2025) will shift the balance. Check for updated benchmarks before finalising your design.
• Single‑cell sequencing depth is still limiting. Both LongBench and Deng et al. note that deeper per‑cell sequencing is needed to fully sample the transcriptome. Many isoforms present in bulk data are missed in single cells simply due to sampling.
• Tool benchmarks are platform‑specific. Hamraoui et al. evaluated tools only on ONT data. The top‑performing tools for PacBio single‑cell data (e.g., the Iso‑Seq pipeline from SMRT Link, or emerging community tools) were not directly compared. Transferring recommendations across platforms should be done with caution.
• Differential transcript usage tools remain immature. None of the DTU methods assessed by Dong et al. achieved a reliable balance of sensitivity and specificity, and this field is still evolving.

Frequently Asked Questions

Q: What is the biggest advantage of single‑cell long‑read sequencing over short‑read scRNA‑seq?
A: It lets you see full‑length transcript isoforms—the actual mRNA molecules a cell produces—rather than just gene‑level counts from short fragments. This means you can identify which splice variants are expressed, discover new isoforms, detect fusion transcripts, and link genetic variants to the isoform they reside on (allele‑specific expression). Short reads cannot reliably do any of these.

Q: Which platform is more accurate, PacBio or Oxford Nanopore?
A: For single‑cell RNA‑seq, PacBio HiFi reads have higher base accuracy (≈Q30 vs. <Q20 for ONT), better cell barcode recovery, and a much higher validation rate for novel isoforms (73% vs. 21% in Deng et al.). However, ONT data can still be used effectively for cell typing and gene quantification, especially with the right bioinformatics tools and the latest R10.4.1 chemistry.

Q: Do I still need to sequence the same library with both short and long reads?
A: No. Hamraoui et al. showed that long‑read‑only analysis pipelines now outperform hybrid methods that rely on paired short‑read data. You can process single‑cell long‑read data independently, simplifying your experiment and reducing cost.

Q: How much does single‑cell long‑read sequencing cost?
A: In the most recent comparative study (Deng et al., 2025), per‑cell costs were $0.52 for ONT and $0.69 for PacBio, compared to $0.25 for Illumina short‑read sequencing. These numbers are for a specific library size and chemistry; actual costs will vary with scale and institution. As a rough guide, budget 2–3 times the cost of a standard short‑read scRNA‑seq experiment.

Q: Can I detect RNA modifications with single‑cell long‑read sequencing?
A: Currently, direct RNA sequencing on ONT can detect RNA modifications (You et al., 2025), including m6A (Chen et al., 2025), but it requires high input and is not compatible with single‑cell methods. Single‑cell long‑read protocols use cDNA, which erases RNA modifications. If you need modification data from single cells, you will need a parallel direct RNA experiment on bulk or enriched populations.

Q: What bioinformatics pipeline should I use for ONT single‑cell data?
A: The Hamraoui et al. benchmark recommends wf‑single‑cell as the most balanced tool for both gene‑ and isoform‑level analyses. If novel isoform discovery is your main goal and you are using R10.4.1 chemistry, Bambu offers high sensitivity. Use BLAZE or flexiplex for initial barcode demultiplexing.

Conclusion

Single‑cell long‑read RNA sequencing has matured to the point where it can be a routine part of a transcriptomics study, provided you align your platform and analysis choices with your research goals. PacBio delivers higher accuracy and more reliable novel isoform calls, while Oxford Nanopore offers a lower‑cost entry point that still recovers rich isoform‑level information. The growing toolkit of specialised bioinformatics pipelines—led by wf‑single‑cell and Bambu—means you no longer need matched short‑read data or deep computational expertise to get started. As with any evolving technology, careful pilot experiments and validation remain essential. Not sure which approach suits your project? Our team can help you design the right experiment and analysis strategy.

References

1. Deng E, Shen Q, Zhang J, Fang Y, Chang L, Luo G, Fan X. Systematic evaluation of single‑cell RNA‑seq analyses performance based on long‑read sequencing platforms. J Adv Res. 2025;71:141‑153. doi:10.1016/j.jare.2024.05.020
2. You Y. et al. Benchmarking long‑read RNA‑sequencing technologies with LongBench: a cross‑platform reference dataset profiling cancer cell lines with bulk and single‑cell approaches. bioRxiv. 2025. doi:10.1101/2025.09.11.675724
3. Hamraoui et al. A systematic benchmark of bioinformatics methods for single‑cell and spatial RNA‑seq Nanopore long‑read data. bioRxiv. 2025. doi:10.1101/2025.07.21.665920
4. Zajac N et al. Comparison of single‑cell long‑read and short‑read transcriptome sequencing via cDNA molecule matching: quality evaluation of the MAS‑ISO‑seq approach. NAR Genom Bioinform. 2025;7(3):lqaf089. doi:10.1093/nargab/lqaf089
5. Dong, X., Du, M.R.M., Gouil, Q. et al. Benchmarking long‑read RNA‑sequencing analysis tools using in silico mixtures. Nat Methods 20, 1810–1821 (2023). doi:10.1038/s41592-023-02026-3
6. Chen, Y., Davidson, N.M., Wan, Y.K. et al. A systematic benchmark of Nanopore long‑read RNA sequencing for transcript‑level analysis in human cell lines. Nat Methods 22, 801–812 (2025). doi:10.1038/s41592-025-02623-4
7. Bhatia S, Field MA, Hebbard L, Schmitz U. Bioinformatics frameworks for single‑cell long‑read sequencing: unlocking isoform‑level resolution. Briefings in Bioinformatics. 2025;26(6):bbaf655. doi:10.1093/bib/bbaf655