Tratopedia
Side N繁中
Settings

Text size

Language

Theme

High contrast

Version

v1.230.2

The release this page was built from. It is what the service worker caches under.

Variant calling · Sequencing technology · Genomics · Machine Learning · Bioinformatics · Research BriefReview published 20 October 2026

A review in Nature Reviews Genetics

Small Variants in the Long-Read Era

A small variant is a one-letter change or a short insertion or deletion, under 50 base pairs, between a person’s genome and the reference. Short-read sequencing found most of them and missed the ones inside long repeats. A review by Zhenxian Zheng, Fritz Sedlazeck and Ruibang Luo argues that long reads and deep learning have changed how these variants are found, and that the headline accuracy now hides where the remaining errors are.

A field moved to deep learning, and a review takes stock Published online in Nature Reviews Genetics on 20 October 2026; every figure below is the review’s own unless marked.

  • >99.9%SNP F1 for deep-learning callers on PacBio HiFi, in high-confidence regions
  • ~99%The same on Oxford Nanopore, where indels remain the weak spot
  • <20%Indel performance in homopolymers longer than 11 bp

What is established

Long reads of 10–20 kb (PacBio HiFi) and over 100 kb (Oxford Nanopore) map into repeats that short reads of 150–300 bp cannot, and neural-network callers now lead on both platforms.

What is not

How accurate any caller is in centromeres, segmental duplications and long tandem repeats. The review calls it poorly characterised, because the truth sets there cover one sample.

Eight years from the first deep-learning caller to the review Dates are publication years of the papers the review cites, read at PubMed where marked.

  1. 2018DeepVariant treats a pile of aligned reads as an image and classifies candidate sites with a convolutional network; the review calls it the first widely adopted deep-learning caller.
  2. 2019PacBio’s circular consensus reads (HiFi) are shown to combine long length with high accuracy; Clairvoyante, a multi-task network for single-molecule data, appears the same year.
  3. 2020–2021Clair (recurrent) and PEPPER-Margin-DeepVariant (haplotype-aware, for nanopore reads) follow.
  4. 31 Mar 2022The Telomere-to-Telomere consortium publishes T2T-CHM13, a gapless human genome adding nearly 200 million bases that the reference had left unfinished (PubMed).
  5. 2022Clair3 combines a fast summary model with a detailed read-level model; Zheng and Luo are among its authors.
  6. 10 May 2023The Human Pangenome Reference Consortium’s draft cuts small-variant discovery errors from short-read data by 34% against GRCh38 (PubMed).
  7. 2025A clinical laboratory validates HiFi genome sequencing for small variants, reporting SNV recall above 99.9% in GIAB high-confidence regions (PubMed); Oxford Nanopore’s Dorado adds a self-attention variant caller.
  8. 20 Oct 2026The review is published online in Nature Reviews Genetics.

Calling is solved where it was easy, and open where it matters The review’s central claim, and what complicates it.

Zheng, Sedlazeck and Luo argue that small variant calling is no longer one task but a family of them. A germline variant sits at about 50% or 100% of reads, a de novo one needs both parents’ data, a mosaic one is usually under 30%, and a somatic one in a tumour can be anywhere from under 1% to 100%. Long reads help all four by phasing: reads are sorted by the parent they came from, so a real variant gathers on one haplotype while sequencing errors stay scattered. The headline figures, they write, come from GIAB high-confidence regions; in a PacBio data set about 75% of discordant indel calls sit in or beside homopolymer runs, and on the complete T2T-CHM13 reference the best caller’s recall fell to 95% for SNPs and 85% for indels in low-mappability and segmental-duplication regions.

  • 93%Of newly benchmarked T2T-CHM13 variants that lie in highly complex regions with little ground truth
  • 1Sample, HG002, behind the T2T truth sets, which is also used for training
  • ~1.7×Higher raw-read indel error for Oxford Nanopore than PacBio in a T2T-HG002 benchmark

What complicates it: the authors build the tools

Luo’s group develops the Clair family that the review discusses throughout, and Zheng is first author of Clair3. The review says so through its own reference list rather than a statement, and its comparisons are of designs, not a ranking of products.

And one figure reads differently at its source

The review says about 8% of GRCh38, nearly 200 million bases, was inaccessible to short-read mapping. The T2T paper it cites describes that 8% as the part of the genome the reference had left unfinished. The two overlap, but they are not the same measurement.

Three ways to call, and what each costs The review’s comparison of strategies, checked against the primary papers it cites where they could be read.

StrategyStrengthTrade-off
Linear referenceMature, fast toolsReference bias; fragile in repeats and structural variation
Assembly-basedReaches up to 15% of the genome that mapping had called inaccessible, in T2T evaluationsAbout ten times the computing; weaker on homopolymer indels
Pangenome graphAbout 34% fewer small-variant errors than GRCh38 in the HPRC draft, measured on short readsHeavy computing; no stable coordinates for clinical reports
  • Networks

    Three designs, three limits

    From the review’s Table 1

    • Convolutional networks (DeepVariant, Clair3, PEPPER) read the read pile as an image; a fixed window makes distant links hard.
    • Recurrent networks (Clair, Clair3, Medaka) read the sequence in both directions; they are sequential and slow.
    • Self-attention (Dorado variant, ClairS-TO) weighs every position at once; its cost grows quickly with length.
  • Material

    Each kind of sample brings its own noise

    DNA, bulk RNA, single-cell RNA

    • Long-read DNA: systematic errors in repeats and homopolymers, and bias from gaps in the reference.
    • Bulk RNA: allele-specific expression shifts the expected fraction, and RNA editing can pass for a variant.
    • Single-cell RNA: sparse coverage and allelic dropout, with ambient RNA and doublets adding false positives.

The bottleneck moves from calling a variant to trusting it What the review settles, what it leaves open, and what it proposes.

Settled

In easy, high-confidence regions, long-read deep-learning callers are accurate enough for clinical use, and a clinical laboratory has validated HiFi sequencing on that basis.

Open

Accuracy where it matters clinically. A pathogenic indel in an RPGR homopolymer or a paralogue-confounded variant in SMN1 barely moves an overall F1 score, yet can decide whether a patient receives a diagnosis.

  • Read an accuracy figure with its regions. A number above 99.9% describes GIAB high-confidence regions, not the whole genome.
  • Ask which reads a pangenome result was measured on. The HPRC draft’s 34% comes from short-read data.
  • Treat hard-region benchmarks as provisional. They rest on one sample that callers are also trained on, so no held-out test yet exists.

Quick notes

  1. Nature Reviews Genetics · Review article

    Small variant calling in the era of long-read sequencing

    A guide for beginners and specialists to the review by Zhenxian Zheng, Fritz J. Sedlazeck and Ruibang Luo

    Published online 20 Oct 2026 · doi:10.1038/s41576-026-01020-71 / 15

  2. Background

    What a small variant is

    The reference genome is like a published edition of a book, and your genome a copy with a few spelling differences: single-letter swaps (SNVs) or short insertions and deletions (indels) under 50 base pairs.

    KindReferenceSampleWhat changed
    SNVACGTTGAACGCTGAOne letter is swapped for another
    InsertionACGTGAACGTTGAExtra letters appear in the sample
    DeletionACGTTGAACG-TGALetters are missing from the sample

    Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 20262 / 15

  3. Background

    Why small variants matter

    Landmark catalogues such as 1000 Genomes, gnomAD and UK Biobank were built with short reads, which struggled in repetitive DNA and left hundreds of thousands of small variants unresolved.

    Differences between people

    They account for many of the differences between individuals.

    Human evolution

    They carry the record of how human populations are related.

    Precision medicine

    Knowing a patient’s variants helps match diagnosis and care to the person.

    Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 20263 / 15

  4. Background

    Where short reads fall short

    Think of a jigsaw with many identical blue-sky pieces. Short pieces cannot show where they belong, so pipelines discard ambiguous alignments rather than risk a false call.

    • 150–300 bpLength of a short read; a repeat longer than that can match many places
    • ~8%Of the genome, nearly 200 million bases, left unfinished in the reference until T2T-CHM13

    Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 20264 / 15

  5. Long reads

    Long reads arrive

    Longer reads are bigger puzzle pieces and easier to place in a repetitive picture.

    • 150–300 bpShort reads: accurate, but too short to cross long repeats
    • 10–20 kbPacBio HiFi: over 99.9% accurate, reading one circular molecule many times
    • >100 kbOxford Nanopore: ultralong reads, with a higher error rate, especially in repeats

    Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 20265 / 15

  6. Long reads

    The shared variant-calling workflow

    It works like proofreading. Long reads change two parts of the recipe: reads can be sorted by parent (phasing), and the classifier must learn long-read error patterns.

    Align

    Place each read on the reference genome, or represent it relative to it.

    Find candidates

    Flag positions where reads differ from the reference, often after local reassembly.

    Classify

    A model decides from local evidence whether each candidate is a true variant or an artefact.

    Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 20266 / 15

  7. Contexts and challenges

    Four kinds of variant, four problems

    KindShare of readsOriginChallenge
    Germline~50% or 100%Inherited; in every cellMapping it to the right place in repeats
    De novo~50%New in the child; absent from both parentsNeeds family (trio) data
    MosaicTypically <30%After conception; in some cells onlyVery low fractions hide among noise
    Somatic<1% to 100%Acquired later, for example in tumoursTumour purity and subclones blur the signal

    Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 20267 / 15

  8. Contexts and challenges

    Phasing, the long-read advantage

    Before phasing

    A faint variant looks like random noise.

    After phasing

    Reads sorted by parent of origin: the true variant shows on one haplotype, while errors stay scattered.

    Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 20268 / 15

  9. Contexts and challenges

    Headline accuracy hides hard regions

    A homopolymer is a run of one letter, such as AAAAAAAA. Counting it exactly is hard, so indels go wrong there, and accuracy above 99.9% is measured mostly in easy, high-confidence regions.

    • ~75%Of discordant PacBio indel calls fall in or next to homopolymer runs
    • <20%Indel performance in long homopolymers (>11 bp), against ~70% in typical repeats
    • ~1.7×Higher raw-read indel error for ONT than PacBio in a T2T-HG002 benchmark

    Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 20269 / 15

  10. Contexts and challenges

    Different sequencing types, different noise

    Long-read DNA

    Systematic errors in repeats and homopolymers, and bias from gaps in the reference; phasing helps.

    Bulk RNA

    Full-length transcripts, but coverage follows expression, and allele-specific expression and RNA editing confuse callers.

    Single-cell RNA

    Links each variant to a cell, with sparse coverage, allelic dropout, ambient RNA and doublets.

    Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 202610 / 15

  11. Deep learning

    Deep learning leads the field

    • >99.9%SNP F1 on PacBio in high-confidence regions
    • ~99%SNP F1 on Oxford Nanopore; indels remain the weak spot

    Summary-based input

    Counts per position, as in Clair3’s pileup model: fast and robust, but loses per-read links.

    Read-level input

    A picture of the read pile, as in DeepVariant and Clair3’s full-alignment model: detailed, but heavier to compute.

    Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 202611 / 15

  12. Deep learning

    Three network designs

    A CNN is like looking at a photograph, an RNN like reading a line of text, and self-attention like taking in the whole page at once.

    CNN

    DeepVariant, Clair3, PEPPER: treats the read pile as an image. Limit: a fixed window makes distant links hard.

    RNN

    Clair, Clair3, Medaka: reads the sequence forwards and backwards. Limit: slow and sequential; signal fades over long spans.

    Self-attention

    Dorado variant, ClairS-TO: weighs every position at once. Limit: cost grows quickly with length.

    Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 202612 / 15

  13. Beyond one reference

    Three ways to call variants

    StrategyStrengthTrade-off
    Linear referenceMature, fast toolsReference bias; fragile in repeats
    Assembly-basedReaches up to 15% of the genome mapping missed, in T2T evaluationsAbout ten times the computing; weaker on homopolymer indels
    Pangenome graphAbout 34% fewer small-variant errors than GRCh38 in the HPRC draft, on short readsHeavy computing; no stable coordinates for clinical reports

    Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 202613 / 15

  14. Beyond one reference

    Complete references expose blind spots

    A pathogenic indel in an RPGR homopolymer, or a paralogue-confounded variant in SMN1, barely moves the overall F1 score yet can decide whether a patient receives a diagnosis.

    • 93%Of newly benchmarked T2T-CHM13 variants lie in highly complex regions with little ground truth
    • 95% / 85%Best-caller recall for SNPs / indels in low-mappability and segmental-duplication regions with HiFi, against >99.9% in easy regions
    • HG002The one sample behind these truth sets, also used for training, so no held-out test exists

    Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 202614 / 15

  15. Outlook

    The bottleneck moves from calling a variant to understanding it.

    • Learn from vast unannotated data with semi-supervised models.
    • Design callers for pangenome graphs, backed by assembly.
    • Build mosaic and somatic callers on haplotype-resolved foundations.
    • Make long-read RNA and single-cell calling routine as costs fall.

    Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 202615 / 15

  • Zhenxian Zheng, Fritz J. Sedlazeck and Ruibang Luo, ‘Small variant calling in the era of long-read sequencing’, Nature Reviews Genetics, published online 20 Oct 2026, doi:10.1038/s41576-026-01020-7, with its Crossref record
  • Nurk and colleagues, ‘The complete sequence of a human genome’, Science 376, 2022
  • Liao and colleagues, ‘A draft human pangenome reference’, Nature 617, 2023
  • Hammond and colleagues, ‘Analytical validation of germline small variant detection using long-read HiFi genome sequencing’, Genome Research 35, 2025. The three cited papers were read at PubMed. All read 21 Oct 2026.

Reliable

Reliable, but with an interestWorth believing and reliableWeak on both, and used with careHonest, but thin

Worth believing

Each point is one source used, placed by two judgements on a six-step scale. Points carry no names: the chart shows the shape of the evidence, not a verdict on any publisher.

Versions

This document is rewritten when what it says has to change. Every version stays published at its own address.

  1. v0001 current

The current version is also at latest/.