Structural Variants: Why Finding One, Placing It, and Genotyping It Are Three Different Problems
Bioinformatics Insight Series
Structural Variants: Why Finding One, Placing It, and Genotyping It Are Three Different Problems
An SV caller's F1 score answers one question: was an event detected near where the truth set says one exists. It does not answer whether the breakpoints are where the VCF places them, or whether the genotype is right — and in multi-sample data those can diverge from detection by fifty percentage points.
A structural variant call has more moving parts than a SNV call, and they fail independently.
For a SNV, "detected" and "characterized" are nearly the same thing. The position is a single base, the alleles are explicit, and the genotype falls out of the allele fraction. Get the call right and you have got the whole object.
An SV call is a composite claim: an event of some type exists, spanning approximately this interval, with breakpoints at these coordinates, present in this many copies. A pipeline can be right about the first part and wrong about the rest, and the standard accuracy metric will not distinguish those cases.
What SV benchmarking actually measures
The convention is worth stating explicitly, because it is the source of the gap.
SV comparison tools count a call as a true positive when it falls within a specified size and distance of a truth-set call. Truvari's typical settings, for example, require something like 95% sequence and size similarity within a distance window — but the window is a window, not an exact match. A call placed hundreds of base pairs from the true breakpoint, with the right type and approximately the right size, scores as a true positive.
That is a reasonable design decision. The alternative — demanding exact breakpoint agreement — would fail calls that are substantively correct, and in repetitive sequence exact agreement may not even be well defined. As the TT-Mars authors put it, in repetitive regions there may be multiple placements of breakpoints that have equal support for a variant, or similar placements that depend on parameters for scoring alignments.
But the consequence is direct: an SV F1 score is a detection metric. It tells you an event was found in roughly the right place, of roughly the right size and type. Breakpoint precision and genotype correctness are separate properties that the headline number does not report.
Genotyping diverges from detection, sometimes enormously
The clearest evidence comes from work on multi-sample SV genotyping, where the two metrics were reported side by side.
Genotypers were run on SVs merged across 47 diverse HPRC genomes and evaluated against assembly-derived baselines. The results:
- kanpig — genotype concordance 84.9%, present-SV F1 0.948
- SVJedi — genotype concordance 55.4%, F1 0.817
- cuteSV-fc — genotype concordance 34.1%, F1 0.611
- sniffles-fc — genotype concordance 33.8%, F1 0.605
Look at the third row. An F1 of 0.611 is a mediocre-but-usable detection result. A genotype concordance of 34.1% means roughly two-thirds of the genotypes are wrong. Both numbers describe the same tool on the same data, and they support very different conclusions about whether the output is fit for a cohort analysis.
Single-sample germline genotyping on curated benchmarks is considerably better — one method reported genotype concordance around 87.8% in GIAB Tier 1 regions, and an assembly-derived HG002 call set reached 99.3% concordance against GIAB v1.1. The gap opens with multi-sample merging, and the reason is structural: merging across samples creates neighborhoods of nearby, partially redundant SV representations that genotypers handle with widely varying success.
One detail from that work explains why this is not a marginal problem. In the GIAB v1.1 benchmark, 15,685 SVs — 43% of the set — have at least one neighbor within 1 kb, whereas the older v0.6 benchmark had no Tier 1 SVs with a neighbor within that distance. And 92.2% of the v1.1 SVs with a neighbor are heterozygous, compared with 65.5% of those without. Adjacency and heterozygosity travel together, and both make genotyping harder.
Figure 1. Benchmark tolerance windows accept calls placed near, but not at, the true breakpoint. For most purposes that is fine. Where a breakpoint sits relative to an exon, a regulatory element, or a domain boundary, a shift within the tolerance window can change the predicted functional consequence while the call continues to score as a true positive.
Breakpoint placement is not just noisy — it is biased
The natural assumption is that breakpoint imprecision is random jitter that averages out. It doesn't.
A study of breakpoint placement across haplotypes found that small allelic variants near an SV shift where the breakpoint gets placed, and that the effect is systematic. Long tracts of homology — such as SVs mediated by transposable elements — increase both the likelihood of imprecise calls and the distance they are shifted. Tandem duplications were the most affected class, with 14% of TDs placed at different locations across haplotypes. Overall, the breakpoint inconsistencies characterized affect roughly 5% of the SVs called in a human genome.
The framing in that paper's title is the part worth carrying: this is described as a source of ancestral bias. Because the small variants that shift breakpoint placement differ in frequency between populations, the placement error is not uniformly distributed across samples. In a cohort analysis, that means the same underlying SV can be recorded at different coordinates in different individuals for reasons correlated with ancestry — and the same authors note that while graph genome methods normalize SV calls across many samples, the resulting breakpoints are sometimes incorrect.
This connects to a failure mode covered earlier in this series. A technical artifact that correlates with ancestry, entering an association analysis, is the structure that manufactures confident false positives.
Where the distinction becomes clinical
For many research uses — SV burden, rough size distributions, population catalogues — detection-level accuracy is genuinely sufficient. Clinical interpretation is where the three questions come apart.
Whether a deletion removes an exon or sits entirely within an intron is a breakpoint-resolution question. So is whether a duplication is in tandem and disrupts a reading frame, or is inserted elsewhere and leaves the gene intact. ACMG guidance on germline SV detection notes that most genomic breakpoints disrupting protein-coding genes fall in introns, because introns are so much larger than exons — so the breakpoint usually sits in sequence where its exact position determines the consequence rather than being obvious from the affected gene alone.
Genotype matters for a different reason: zygosity determines inheritance interpretation. A homozygous deletion in a recessive gene and a heterozygous one are different findings, and a genotype error flips that conclusion without changing anything about detection.
Laboratories that have built clinical SV calling treat resolution as a design requirement rather than a bonus. One clinical breakpoint-analysis method is described as detecting deletions, tandem duplications and inversions over 50 bp with resolution of one to a few base pairs — and clinical genome sequencing work reports refining breakpoints at base-pair resolution and clarifying the structural configuration of intragenic duplications as additions beyond what prior testing provided. Those are characterization outcomes, not detection outcomes.
| Property | What it depends on | What breaks if it's wrong |
|---|---|---|
| Presence and type | Read-pair, split-read and depth signal above noise | The variant is missed or invented entirely |
| Breakpoint position | Split-read support; local assembly; repeat context | Predicted consequence, exon involvement, frame |
| Size and boundaries | Same, plus how the caller resolves ambiguous placement | Dosage interpretation; overlap with regulatory elements |
| Genotype | Allele-balance modelling; neighboring SV handling | Inheritance mode, carrier status, cohort frequencies |
| Representation | Merge and normalization choices across samples | Whether the same event matches itself between samples |
What to report and check
Report genotype concordance separately from F1
They are different metrics answering different questions, and the multi-sample results show they can diverge by fifty points. A validation document that reports only detection F1 has characterized only detection.
State the benchmark's matching tolerance
An F1 computed with a 500 bp window and one computed with a 50 bp window are not comparable numbers. The tolerance is part of the measurement, in the same way a benchmark region is.
Treat breakpoint precision as a stratified metric
Distance from called to true breakpoint, reported by SV type and by repeat context, is far more informative than a single score. Tandem duplications and transposable-element-mediated events are where placement degrades most, and they degrade predictably.
Confirm breakpoints before clinical interpretation
Where a reported consequence depends on exactly which exon a breakpoint falls in, the breakpoint needs orthogonal support — local assembly, long reads, or targeted confirmation. A caller's coordinates are an estimate whose precision varies by context, and the VCF does not carry that context.
Be careful with merged multi-sample call sets
Merging is where representation problems concentrate. In one analysis, collapsing redundant representations reduced 561,735 discovered variants to 257,323 — a 54% reduction — before genotyping even began. How that collapse is performed determines whether the same event is recognized as the same event across samples.
Match the metric to the use
For SV burden, detection accuracy is what matters. For a cohort association study, genotype accuracy is what matters. For clinical reporting, breakpoint resolution is what matters. A single number cannot serve all three, and choosing the wrong one is how a pipeline gets validated for a purpose it wasn't measured against.
The short version
SV benchmarking counts a call correct when it lands within a tolerance window of a truth-set event, which makes F1 a detection metric. Breakpoint placement is separately fallible — systematically so, with 14% of tandem duplications placed differently across haplotypes — and genotyping is separately fallible too, with multi-sample concordance ranging from 33.8% to 84.9% across genotypers whose detection F1 spanned a much narrower range. Report all three.
A composite claim reported as a single result
The recurring theme in this series is a number that describes less than it appears to. The SV version has a particular structure: the output is genuinely a composite of several claims, and the accepted metric evaluates one of them.
That is not a flaw in the metric. Detection is the hardest part of SV calling and the right thing to optimize first, and the tolerance windows exist because exact breakpoint comparison is not always well posed. The field's benchmarking tools are thoughtfully built and the newer assembly-based truth sets are a substantial advance.
The gap is between what a benchmark measures and what a downstream analysis assumes. An annotation pipeline reads breakpoint coordinates as if they were exact. A cohort analysis reads genotypes as if they were determined. A clinical report reads exon involvement as if it followed from detection. None of those assumptions is tested by the number in the validation summary — and for structural variants, unlike SNVs, the distance between finding something and knowing what it is remains substantial.
Zetobit builds and validates CAP/CLIA-compliant NGS pipelines, including SV workflows with breakpoint-resolution metrics, genotype concordance reporting, and orthogonal confirmation where clinical interpretation depends on exact placement. If you are validating an SV pipeline or reconciling call sets across cohorts, we're happy to talk.
References
- English AC, Dolzhenko E, Ziaei Jam H, et al. K-mer analysis of long-read alignment pileups for structural variant genotyping (kanpig). Nature Communications 16 (2025). doi:10.1038/s41467-025-58577-w.
- Chen X, Harris H, Sedlazeck FJ, et al. Small allelic variants are a source of ancestral bias in structural variant breakpoint placement. (2023).
- English AC, Menon VK, Gibbs RA, et al. Truvari: refined structural variant comparison preserves allelic diversity. Genome Biology 23:271 (2022). doi:10.1186/s13059-022-02840-6.
- Jiang T, Liu S, Cao S, et al. TT-Mars: structural variants assessment based on haplotype-resolved assemblies. bioRxiv preprint (2021); later Genome Biology 22:293. Preprint cited for the breakpoint-degeneracy discussion.
- Linthorst J, Nelson W, Alkan C, et al. NPSV: a simulation-driven approach to genotyping structural variants in whole-genome sequencing data. GigaScience 10(7):giab046 (2021). doi:10.1093/gigascience/giab046.
- Saunders CT, Holt JM, Baker DN, et al. Sawfish: improving long-read structural variant discovery and genotyping with local haplotype modeling. Bioinformatics 41(4):btaf136 (2025). doi:10.1093/bioinformatics/btaf136.
- Sanchis-Juan A, Souche E, Laurie S, et al. Structural variant calling and clinical interpretation in 6224 unsolved rare disease exomes. European Journal of Human Genetics (2025). doi:10.1038/s41431-024-01637-4.
- Raca G, Astbury C, Behlmann A, et al. Points to consider in the detection of germline structural variants using next-generation sequencing: a statement of the American College of Medical Genetics and Genomics (ACMG). Genetics in Medicine 25(2):100316 (2023). doi:10.1016/j.gim.2022.09.017.