Long-read sequencing reads DNA in single pieces tens of thousands of letters long, instead of the 100 to 300 letters typical of the short-read machines that produce most of the world’s genomes. The longer pieces let researchers see repeats, large rearrangements and the parent-of-origin of each variant that short reads blur or lose. In published rare-disease studies, adding long reads has solved roughly 7 to 17 percent of cases that short-read genome sequencing could not.

That range comes with caveats, and the technology is not simply “better.” This explainer works through what long reads change, with numbers from 2025 papers, and where short reads still win.

Why short reads leave blind spots

Short-read sequencing chops DNA into small fragments, reads each one accurately, and then maps the fragments back onto a reference genome. That works well for single-letter changes in unique stretches of DNA. It fails when the fragment itself could belong to many places.

Picture a jigsaw puzzle where a whole region is the same blue sky. Each piece fits dozens of spots. Genomes contain such regions: tandem repeats, segmental duplications and near-identical gene copies. Large deletions, insertions and inversions are hard to call for a related reason, since the evidence is a subtle change in how fragments align rather than a read that crosses the whole event.

A long read that spans an entire repeat, together with unique sequence on both sides, simply shows how long the repeat is and where it sits. The price has historically been higher cost per genome, lower throughput and higher raw error rates, though accuracy has improved a great deal; I am leaving out specific cost figures because the ones I found were years old.

What long reads resolve: five categories

Category Why short reads struggle What long reads add
Structural variants (changes over 50 base pairs) Inferred indirectly from alignment patterns The breakpoints and inserted sequence are read directly
Tandem repeat expansions Reads are shorter than the repeat Repeat length and internal composition
Near-identical gene copies Reads map to multiple places Sequence unique to each copy anchors the read
Phasing Two variants cannot be assigned to one parental copy unless close together Variants on the same long read sit on the same chromosome
Methylation Needs a separate chemical treatment Detectable from the same read on some platforms

The Genome Research review by Del Gobbo and Boycott lists real examples of each. Long reads uncovered expanded repeats in FGF14, NOTCH2NLC and SAMD12 in people with neurological disease, resolved variants in hard-to-map genes such as IKBKG, HYDIN and PKD1, and phased pairs of variants where parental DNA was unavailable. In one case, a repeat expansion in DIP2B was spotted partly because of its associated hypermethylation.

The clinical numbers: how many extra diagnoses?

The 2025 Genome Research review, which draws on more than 50 studies, puts the additional yield after a negative short-read genome at 7 to 17 percent. It does not offer one pooled figure, and the individual studies are small and differ in how patients were selected. Examples it cites:

  • A 2024 study of 96 people with undiagnosed rare disease found likely causal variants in 16, with 7 of them (7.3 percent) detectable only by long reads.
  • A 2024 study of 232 people in 93 families reported 13 percent new diagnoses and a further 4.3 percent with candidate structural variants.
  • A 2023 study of 34 families, all negative on short-read exome sequencing, reached a diagnosis in 13 families, 8 of which (23.5 percent) were not detectable by exome sequencing at all.

The smallest early studies span from none in five patients to 60 percent in ten, which shows mainly how unstable percentages are at that scale. A fair reading: long reads solve a meaningful minority of cases that other tests leave open, and the exact fraction depends on who is tested and how thoroughly they were studied before.

What 1,019 genomes show about normal variation

Finding a variant in a patient is only useful if you know how common it is in healthy people. Two Nature papers in 2025 supplied much of that background. One team sequenced 1,019 people from 26 populations of the 1000 Genomes Project and, according to the abstract record, identified more than 100,000 sequence-resolved biallelic structural variants and genotyped about 300,000 multiallelic tandem repeats. The authors describe it as a resource for prioritizing variants in patient genomes.

A second consortium produced near-complete assemblies of 65 people (130 haplotypes). They closed 92 percent of the gaps left in earlier assemblies, fully assembled 1,246 centromeres, and resolved 1,852 complex structural variants. Each person carried about 26,000 structural variants. The University of Washington summary adds that centromere arrays varied up to 30-fold in length, and that the team untangled the SMN1 and SMN2 genes, which matter in spinal muscular atrophy.

These reference catalogs make clinical interpretation less guessy. They also feed back into short-read analysis: the consortium reports that pairing its data with a pangenome reference improved genotyping from short reads.

The two main technologies, in plain terms

Two companies dominate. Pacific Biosciences reads a single DNA molecule many times around a circular loop and combines the passes into one high-accuracy “HiFi” read, usually in the range of ten to twenty thousand letters. Oxford Nanopore threads DNA through a tiny protein pore and infers each letter from changes in electrical current. Its reads can be far longer, in some cases hundreds of thousands of letters, and the same signal carries methylation information.

The trade-offs are practical. Nanopore devices range from a pocket-sized unit to large bench instruments, and runs can be read as they happen. HiFi data tend to be prized for base-level accuracy. Either way, the analysis is where much of the effort goes: alignment tools, variant callers and the pangenome references mentioned above all have to cope with sequence that a single linear reference never contained.

A worked example: why a repeat is so hard to count

Imagine a disease gene with a three-letter unit, GAA, repeated. Healthy people might carry a few dozen copies. A person with the disease might carry several hundred. A short read of 150 letters holds at most fifty copies of that unit, so a read lying entirely inside the repeat looks the same whether the repeat is long or enormous. Short-read tools can estimate expansion size from the number of such reads, but cannot see the ends, check whether the repeat is interrupted by other letters, or tell which copy of the chromosome it sits on.

A long read that begins in unique sequence, runs through the whole repeat and ends in unique sequence on the other side answers all three questions in one pass. That, in outline, is why the FGF14 and NOTCH2NLC findings listed in the review came from long reads.

The review also notes that long-read RNA sequencing has revealed splice-altering variants, for example in MFN2 and CLPB, where a gene’s messenger RNA is spliced in an unexpected way. That points to a wider use: reading whole transcripts rather than reassembling them from fragments.

Where short reads still win

  • Established workflows. Clinical short-read pipelines, interpretation databases and reimbursement rules are mature. Long-read equivalents are catching up but are not uniform across labs.
  • Throughput and cost per sample. For large screening projects, short reads remain the practical default.
  • Small, single-letter changes. Short reads are very accurate at these, and they are the largest class of known disease variants.
  • Sample demands. Long reads need long, intact DNA, which not every stored clinical sample provides.

Short reads are not obsolete. The two are complementary, and some labs run both.

Questions people ask

Is long-read sequencing more accurate?

It depends on the platform and the type of variant. Modern instruments have narrowed the gap, but accuracy per base and accuracy per variant call are different questions. The structural variant and repeat advantage is real. A blanket claim of higher accuracy is not supported by the studies above.

Can I order it?

Availability varies by country, lab and insurer, and I did not find a current, authoritative list. A genetics clinician or genetic counselor is the person to ask about whether a repeat test would make sense for a specific diagnostic question.

Does it complete the human genome?

A near-complete reference has been built from long reads, but the 2025 work shows how much remains to catalog. With 39 percent of chromosomes assembled end to end in the 65-person set, many people’s genomes still have unresolved regions.

What to watch

Three things will decide how fast long reads move into routine care: prospective trials that compare long-read-first against standard testing, shared interpretation resources for structural variants and repeats, and evidence on the cost per extra diagnosis. Until those arrive, the picture is promising and incomplete. This article is general information, not medical advice.

Sources