The Reference That Isn't There

Zetobit

Insight Series Services

Bioinformatics Insight Series · No. 13

The Reference That Isn't There

Why structural variants still break your pipeline — and why the problem isn't the caller.

By Kanna Nandakumar, PhD · Zetobit LLC · Lexington, KY

Short reads gave us a decade of clean SNV calling and an illusion of completeness. The illusion holds until you ask a harder question: what happened to this 4 kb region? Then the pipeline that flagged your missense variant with 99% confidence goes quiet.

Structural variants — deletions, duplications, inversions, translocations, and insertions above roughly 50 bp — account for more variable base pairs per genome than SNVs and indels combined [1]. They are also where our tooling is weakest. This piece is about why that gap persists, and what to do about it in a production setting.

Why short reads struggle

An SV is not detected directly. It is inferred from indirect evidence:

• Discordant read pairs — mates that map too far apart, too close, or in the wrong orientation.
• Split reads — a single read whose alignment breaks across two loci, marking a breakpoint.
• Read depth — a coverage drop or spike suggesting copy number change.
• Local assembly — reconstructing the region from scratch and comparing to reference.

Each signal has a blind spot. Depth alone cannot see balanced events — inversions, reciprocal translocations — because copy number is unchanged. Discordant pairs lose resolution as insert size distributions widen. Split reads require the breakpoint to fall inside a read, a hard ask at 150 bp when the flanking sequence is repetitive [2].

Figure 1. Four classes of indirect evidence for a deletion. Each is strongest outside the repeat-flanked zone (shaded) and weakest inside it — which is precisely where structural variants are enriched. No single signal is sufficient; ensembles exist because the evidence classes fail in different places.

And the repeats are the crux. Roughly half the human genome is repetitive, and SVs are enriched precisely there — segmental duplications, transposable elements, and tandem repeat arrays are the substrate for non-allelic homologous recombination [3]. The regions where SVs are most likely to occur are the regions where short reads are least able to map uniquely.

The concordance problem

Run three SV callers on the same BAM. Manta, Delly, and Lumpy will agree on large, clean deletions in unique sequence. Outside that comfort zone, pairwise concordance drops steeply — different callers, different evidence weighting, different breakpoint estimates for what is nominally the same event [4].

This is not a bug to be fixed by picking the best caller. It reflects genuine uncertainty in the underlying inference. The standard mitigation is an ensemble: run several callers, merge with something like SURVIVOR, and require support from two or more [5]. This raises precision and costs sensitivity — and the merge parameters (breakpoint tolerance, type matching, strand requirements) quietly determine your callset more than the callers do.

Your SV callset is a function of your merge settings. Document them like you document a reference build.

What long reads actually change

HiFi and ONT reads spanning 10–100 kb resolve the mapping ambiguity directly. A single read can cross a repeat array and its flanks, making the SV an alignment fact rather than a statistical inference. Sensitivity for insertions — historically the worst-detected class, because the inserted sequence is by definition absent from the reference — improves substantially [6].

This is not a free win in a clinical context:

• Cost per sample remains above short-read WGS.
• Validation burden is real. A CAP/CLIA pipeline needs established truth sets, and the SV benchmarks — GIAB Tier 1 SV regions, the Challenging Medically Relevant Genes set — cover far less of the genome than their SNV counterparts [7].
• Reference dependence persists. A long read still gets compared to something. If that something is a single linear reference, you have improved detection without improving representation.

The representation problem underneath

Every point above assumes there is a correct answer to compare against. For SVs, "the reference" is doing more work than it can bear. A linear haploid reference cannot represent an insertion carried by a large fraction of the population — that sequence simply isn't in the coordinate system. A carrier looks like an insertion; a non-carrier looks like reference; the population-level truth is a common polymorphism that the coordinate system has no way to express.

This is the argument for graph-based references, and it is why the pangenome work matters more for SVs than for SNVs [8]. Genotyping against a graph that already contains the alternate haplotypes converts a detection problem into a much easier assignment problem [9]. The tooling is maturing but not yet routine, and coordinate translation back to GRCh38 for reporting remains a genuine friction point.

Practical guidance

1. Define your SV scope explicitly. Size range, event types, and target regions. A pipeline that reports "structural variants" without qualification is overpromising.
2. Use an ensemble, and version the merge. Two-caller support is a reasonable default; record breakpoint tolerance as a pipeline parameter.
3. Separate CNV from SV calling. Depth-based CNV detection and breakpoint-based SV detection answer different questions with different failure modes. Merging them into one output field hides both.
4. Benchmark honestly. Report performance against GIAB SV truth sets within the confident regions, and state what fraction of your target those regions cover. That second number is usually the informative one.
5. Flag the dark regions. Any SV callset should ship with a companion track of where you couldn't call. Silence is not a negative result.

The takeaway

SV calling is not SNV calling with bigger coordinates. It is a different inference problem, built on indirect evidence, in the genomic neighborhoods least hospitable to short reads, against a reference that structurally cannot represent the thing being measured.

Long reads and pangenome graphs address the first and last of those honestly. Until both are routine in the clinic, the responsible position is not "we call structural variants" — it's "here is the class of structural variants we call, here is our measured sensitivity within a defined region, and here is where we are blind."

That is a less marketable sentence. It's a considerably more useful one.

References

1. Sudmant PH, Rausch T, Gardner EJ, et al. An integrated map of structural variation in 2,504 human genomes. Nature. 2015;526(7571):75–81. — Nature — 1000 Genomes structural variation map
2. Alkan C, Coe BP, Eichler EE. Genome structural variation discovery and genotyping. Nature Reviews Genetics. 2011;12(5):363–376. — Nature Reviews Genetics — SV discovery and genotyping
3. Treangen TJ, Salzberg SL. Repetitive DNA and next-generation sequencing: computational challenges and solutions. Nature Reviews Genetics. 2012;13(1):36–46. — Nature Reviews Genetics — repetitive DNA and NGS
4. Kosugi S, Momozawa Y, Liu X, et al. Comprehensive evaluation of structural variation detection algorithms for whole genome sequencing. Genome Biology. 2019;20:117. — Genome Biology — benchmarking SV callers
5. Jeffares DC, Jolly C, Hoti M, et al. Transient structural variations have strong effects on quantitative traits and reproductive isolation in fission yeast. Nature Communications. 2017;8:14061. (SURVIVOR merging toolkit) — SURVIVOR — GitHub repository
6. Sedlazeck FJ, Rescheneder P, Smolka M, et al. Accurate detection of complex structural variations using single-molecule sequencing. Nature Methods. 2018;15(6):461–468. — Nature Methods — long-read SV detection (Sniffles/NGMLR)
7. Zook JM, Hansen NF, Olson ND, et al. A robust benchmark for detection of germline large insertions and deletions. Nature Biotechnology. 2020;38(11):1347–1355. — Nature Biotechnology — GIAB SV benchmark
8. Liao WW, Asri M, Ebler J, et al. A draft human pangenome reference. Nature. 2023;617(7960):312–324. — Nature — draft human pangenome reference (HPRC)
9. Hickey G, Heller D, Monlong J, et al. Genotyping structural variants in pangenome graphs using the vg toolkit. Genome Biology. 2020;21:35. — Genome Biology — SV genotyping with vg

Bioinformatics Insight Series · No. 13 © 2026 Zetobit LLC · Lexington, KY