Slide deck
BackSmall Variants in the Long-Read Era
Long reads and deep learning now call SNPs above 99.9% accuracy on PacBio, but that figure comes from the easy part of the genome. In homopolymers, segmental duplications and newly finished T2T regions, accuracy falls sharply, and the truth sets to measure it hardly exist.
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
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.
Kind Reference Sample What changed SNV ACGTTGA ACGCTGA One letter is swapped for another Insertion ACGTGA ACGTTGA Extra letters appear in the sample Deletion ACGTTGA ACG-TGA Letters are missing from the sample Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 20262 / 15
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
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
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
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
Contexts and challenges
Four kinds of variant, four problems
Kind Share of reads Origin Challenge Germline ~50% or 100% Inherited; in every cell Mapping it to the right place in repeats De novo ~50% New in the child; absent from both parents Needs family (trio) data Mosaic Typically <30% After conception; in some cells only Very low fractions hide among noise Somatic <1% to 100% Acquired later, for example in tumours Tumour purity and subclones blur the signal Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 20267 / 15
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
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
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
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
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
Beyond one reference
Three ways to call variants
Strategy Strength Trade-off Linear reference Mature, fast tools Reference bias; fragile in repeats Assembly-based Reaches up to 15% of the genome mapping missed, in T2T evaluations About ten times the computing; weaker on homopolymer indels Pangenome graph About 34% fewer small-variant errors than GRCh38 in the HPRC draft, on short reads Heavy computing; no stable coordinates for clinical reports Zheng, Sedlazeck & Luo · Nature Reviews Genetics · 20 Oct 202613 / 15
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
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