Long-Read Sequencing Explained: Nanopore vs PacBio and Why Read Length Matters

Quick answer: Long-read sequencing reads single molecules end to end — tens of thousands of bases per read instead of 150 — using one of two chemistries: Oxford Nanopore, which measures ionic-current changes as DNA threads through a protein pore, and PacBio, which watches fluorescent nucleotides being incorporated in real time through a zero-mode waveguide. Short reads force you to reconstruct genomes through repeats you cannot cross; long reads span them directly. With PacBio HiFi reads now carrying roughly 99.9% raw accuracy and Nanopore simplex reads passing Q20, the choice today is about read length, output scale, methylation detection and budget — not about whether long reads are accurate enough.

The companion to next-generation sequencing and genome assembly on this blog: NGS explains the machines and assembly explains what the reads are for — this one explains what changes when the reads get long.

Why read length is the whole argument

A genome is mostly repetitive: transposons, segmental duplications, rDNA arrays, and genes in near-identical paralog families. Short reads (100–150 bp) are shorter than most of these landmarks, so a read from inside a repeat matches many places equally well and the assembler or mapper can only guess. That guess is why short-read assemblies shatter into thousands of contigs, why structural variants are under-detected, and why mapping quality collapses in exactly the regions disease research cares about.

A 15 kb read, by contrast, usually starts in unique sequence, runs clean through the repeat, and ends in unique sequence on the other side. The repeat is no longer a puzzle; it is internal context. Everything downstream — contiguity, phasing, structural variant calling, isoform resolution — improves for that one geometric reason, not because of any clever software.

The two chemistries

Oxford Nanopore sequences by voltage. A motor protein ratchets a single DNA strand through a protein pore (the current R10.4.1 generation) embedded in a membrane; as bases pass the pore’s narrowest point they perturb the ionic current, and basecalling models convert that raw current signal into a sequence. Because the signal is electrical rather than optical, the device is small — a USB-sized MinION produces data — data streams in real time, and the raw signal retains information that optical methods throw away, most importantly modified-base calls such as 5mC methylation, which is detected directly without bisulfite treatment (the contrast with the bisulfite route in the methylation guide matters experimentally).

PacBio sequences by fluorescence. Thousands of identical copies of a fragment are fixed in a zero-mode waveguide — a well narrower than the wavelength of the light used — so only the bottom-most incorporated nucleotide is illuminated. Each dNMP carries a fluorescent tag that is cleaved on incorporation and flashes before the next base is read. The payoff of watching one molecule for a long time is circular consensus: the same fragment is read over and over as the polymerase goes around the circle, and the multiple passes are aligned into a HiFi (high-fidelity) read with a quality the single pass could not reach.

Accuracy: what changed between 2015 and now

Early long reads carried 10–15% error, mostly indels in homopolymer runs — useful for scaffolding, unusable for calling variants. That era is over:

Two practical consequences. First, HiFi reads are accurate enough that human variant calling from long reads is now routine rather than approximate. Second, Nanopore accuracy is high enough for most assembly, metagenomics and 16S work without polishing, though homopolymer regions still deserve a sceptical eye. Consensus accuracy also rises with depth: even a noisy platform gives excellent consensus sequence when coverage is high, because the errors are largely uncorrelated between reads.

What long reads change in practice

• De novo assembly. Assemblers such as hifiasm and Flye resolve repeats that shatter short-read assemblies, producing near-chromosome contigs and dramatically fewer scaffolds. The assembly metrics that matter are still the pairing of contiguity with completeness checks covered in the assembly guide — long reads improve N50, they do not excuse skipping BUSCO.
• Structural variants and phasing. A deletion, inversion or translocation shows up as one read spanning the breakpoint — no paired-end inference required. Because each long read covers many heterozygous sites, haplotypes assemble directly instead of being statistically imputed, which is what makes clinical long-read genomes compelling.
• Transcript isoforms. PacBio Iso-Seq and direct RNA sequencing read whole transcripts, so alternative splicing events are observed molecules rather than computational reconstructions from 150 bp fragments.
• Metagenomics and amplicons. Full-length 16S (~1.5 kb) or ITS on a Nanopore resolves species and even strain level that short amplicons cannot, compressing the whole read-to-taxonomy chain of the 16S workflow into one pass per molecule.
• Direct modification detection. Both platforms report base modifications from raw signal, removing the conversion-bias step from methylation studies entirely.

What short reads still win

Long reads did not make short reads obsolete. Short reads remain cheaper per gigabase, offer the most mature variant-calling pipelines, and dominate RNA-seq quantification where depth — not length — decides sensitivity for differential expression. Established cohort-scale studies still run on short reads because the reference pipelines, the validation data, and the statistical priors all live there. The common modern answer is hybrid: long reads for structure and assembly, short reads for depth and cheap replication, or simply HiFi-only when the budget allows. Coverage expectations also differ — human HiFi genomes typically run 20–30×, while microbial assemblies can be convincing at far lower depth.

Working with long reads downstream

The file formats change shape rather than identity: FASTQ as usual, then unaligned BAM or uBAM, aligned BAM with supplementary records aplenty. Three notes that trip beginners:

• Alignment needs a long-read-aware mapper. Minimap2 with the right preset (map-hifi for HiFi, map-ont for Nanopore) is the default; standard short-read settings will fail in ways that look like data problems. The read mapping guide covers the fields — expect many supplementary alignments, because one long read can span a rearrangement.
• Quality control is different. Read-length distributions, N50 of the reads themselves, and per-read quality plots (NanoPlot, or samtools stats after alignment) replace per-base FastQC tiles; the trimming mindset from NGS QC carries over, the specific plots do not.
• Duplicates and overlap matter less; depth matters more. Long-read datasets are smaller in record count and larger in bases; filtering tools such as filtlong or length/quality cutoffs do the job that adapter trimming and quality filtering did for short reads.

Choosing between them

Ask three questions. Do I need to resolve repeats, structural variants, phases or isoforms? If no, short reads are still the cheapest defensible answer. Do I need methylation or real-time answers (adaptive sampling, outbreak surveillance)? Nanopore is the only route. Do I need the highest-accuracy genome-scale variant calls with long-range context? HiFi on PacBio is the standard answer. Many projects now sequence long for assembly and structure, then revert to short reads only where depth economics demand it.

Key takeaways

• Read length beats read count for repeats, structural variants, phasing and isoforms — the geometry of spanning a repeat is the entire argument.
• Nanopore reads electrical signal through a pore in real time; PacBio reads fluorescence in a zero-mode waveguide and folds many passes into one HiFi read.
• HiFi is ~Q30 raw and Nanopore R10.4.1 ~Q20 or better: both are analysis-grade now, with homopolymers the residual weak spot.
• Long reads detect 5mC methylation directly from raw signal — no bisulfite, no conversion bias.
• Short reads still win on cost per gigabase and pipeline maturity; hybrid designs remain common.
• Downstream needs long-read-aware tools: minimap2 presets, read-level QC, and expectations of supplementary alignments.

Challenge

Take a public dataset — a finished bacterial isolate from the Nanopore public collection or a human HiFi sample on SRA — and compare it head-to-head with a short-read run of the same organism. Assemble each, then report contig count, N50, and BUSCO completeness side by side. The number of contigs almost always tells the story in one line, and seeing it yourself settles the short-vs-long question faster than any table.

Want guided practice turning sequencing data into a defensible assembly or variant call? Ampersand Academy offers one-to-one mentoring in bioinformatics — part of Mahadhi, a development and digital marketing company.

FAQ