A working index of structural variation and pangenome tools

Five things that moved while pipelines stood still

Most of the value of keeping a list like this is catching the moments when a default quietly stopped being the default. Five of those happened recently enough that a pipeline written two years ago is now wrong in at least one of them.

1. The GIAB structural variant benchmark you are validating against is deprecated. NIST now distributes an HG002 v5.0q benchmark covering small variants and structural variants together, built on the v1.1 T2T-HG002 assembly from the Q100 effort. In the same release note, the long-standing Tier 1 SV v0.6 benchmark is listed as deprecated, along with the tandem-repeat v1.0, XY v1.0 and the HG002 small-variant v4.2.1 sets. A great many methods papers — including ones published this year — still report F1 against v0.6. That is not fraud, but it is a stale ruler, and if you are choosing a caller on the strength of those numbers you should know the ruler changed.

2. Canu is retired, by its own authors. The v2.3 release notes say, in as many words, that the sequencing technology has moved on and that Canu is all but obsolete. There will not be another release. It still works on CLR data; it should not appear in a 2026 recommendation.

3. PacBio's own recommended HiFi SV caller is no longer pbsv. The nf-core/pacvar documentation now states that sawfish — which does joint SV and copy-number calling with local haplotype modelling — is the recommended SV caller for HiFi data. pbsv is still maintained, but a pipeline that lists pbsv and not sawfish is a year behind.

4. ARS-UCD2.0 exists, and the field has not moved onto it. The current cattle reference at NCBI and Ensembl has been ARS-UCD2.0 since 2023. But 1000 Bull Genomes and a large share of 2025–2026 cattle GWAS still run on ARS-UCD1.2. Plan for a dual-coordinate world, budget a liftover step, and state your coordinate system in the first paragraph of every methods section. This is the single most common source of quiet coordinate errors I see.

5. Short-read SV calling has almost no living software left. Manta's last release was 2019, CNVnator 2020, smoove 2021, GRIDSS2 2022. Delly is the only tool in that category still shipping — and its recent work is long-read oriented. If your SV discovery rests on short reads and a 2019 caller, the caller being old is the smaller of your two problems; see the section on calling below.

Assembly

For a phased diploid mammalian genome from HiFi plus Hi-C or trio data, hifiasm remains the first thing I reach for and the thing I would defend in review. Verkko earns its extra cost when contiguity is the deliverable rather than a means — and its documentation is honest that its Hi-C and Pore-C integration was tested mostly on human and primate data, which is worth remembering before you point it at a ruminant with a very different repeat landscape.

The interesting development is on the ONT side. A February 2026 Nature paper describes near-telomere-to-telomere assembly from nanopore simplex reads with a built-in error-correction step, reporting more complete chromosomes than Verkko with HERRO-corrected reads. The honest reading is that it wins on contiguity and loses on consensus quality — homopolymers are still homopolymers — so an extra polishing step is needed. If your project is contiguity-first and budget-constrained, ONT-only is now a real option. If you are going to call small variants off the assembly, it is not.

Building the graph

The Minigraph-Cactus versus PGGB question used to be a genuine coin flip. In 2026 it is less balanced, and not because of a paper: Cactus has shipped four minor or major releases since September 2025, while pggb, wfmash, seqwish and smoothxg have all been quiet for over a year. Release cadence is not quality, but for a pipeline you intend to run again next year it is not nothing either.

The trade-off itself has not changed. Cactus documentation is explicit that Minigraph-Cactus depends on a predetermined reference which is never clipped and never self-aligned — that reference is structurally privileged, and it is the coordinate basis for the VCF. PGGB's all-versus-all alignment avoids that bias, at the cost of parameter tuning that the PGGB documentation itself concedes is problem-specific. If a reviewer is going to ask you about reference bias, know which of these two answers you are giving before you start.

A comparison published in late 2025 across several crop species is worth reading precisely because it complicates the story: minigraph-only graphs were roughly five times smaller but covered about 83% of the input assemblies against over 94% for Minigraph-Cactus and PGGB — meaning genes genuinely go missing. And on simulated structural variants, the assembly-based caller SVIM-asm scored highest, ahead of SyRI, Minigraph-Cactus and PGGB in that order. Graphs are the right tool for genotyping a population. They are not automatically the best tool for discovering SVs in a handful of assemblies, and it is worth being clear with yourself about which of those two jobs you are doing.

Three Cactus gotchas that cost people days: soft-masking in the input is ignored, so masking has to be handled through the run's own options; alternate loci and patch scaffolds in the reference must be excluded or the output is meaningless; and small contigs are dropped during chromosome assignment unless you pass the permissive contig filter, which matters a great deal for fragmented livestock assemblies.

Genotyping a cohort

Giraffe is the scalable option and the one whose documentation actually tells you how to run it in production: pass the distance, zipcode and minimizer indexes explicitly rather than relying on autodetection, and make sure your graph carries haplotype paths as W-lines — the older P-line heuristics can silently give wrong results. A 30× human-scale sample lands somewhere in the one-to-ten-hour range on 32 threads, which is what makes cohort-scale work tractable at all. As of late 2025 there is a long-read Giraffe that is more than an order of magnitude faster than GraphAligner on graphs with hundreds of haplotypes, at accuracy comparable to minimap2, which removes most of the reason to keep GraphAligner in a new pipeline.

PanGenie is what I have used for cohort genotyping from a breed-specific graph, and version 4 removed the old 254-haplotype ceiling that made large panels awkward. Its defining property is also its limitation: it genotypes variants that are already in the panel. It will never find something new, and every quality question about your results is really a quality question about your graph.

Two developments worth tracking rather than adopting immediately. Pangenome-aware DeepVariant reports single-digit-percent error reduction genome-wide and around a fifth fewer errors in segmental duplications above 99% identity — the gains are concentrated exactly where you would want them. And a 2025 method called The Great Genotyper attacks the problem from the other end, indexing thousands of samples in a coloured de Bruijn graph so that re-genotyping a new variant set across the whole cohort takes hours rather than months. If you re-genotype the same population repeatedly — which most breeding programmes do — that is the shape of the answer, even if the index build is expensive.

Calling SVs without a graph

Long reads

The most useful recent result about long-read SV calling is not about any one caller. A December 2025 benchmark across nine of them found that between roughly 77% and 91% of erroneous calls fall inside low-complexity regions that make up only about 1.2% of the genome, and that some callers miss around half of structural variants of 2 kb and above. The mechanism is mundane: minimap2 places gaps inconsistently across reads in long low-complexity stretches, and simple clustering approaches inherit that inconsistency. Two things follow. Stratify every benchmark you run by low-complexity status, because an unstratified F1 mostly measures how a caller behaves in 1% of the genome. And prefer callers that do haplotype-aware or local reassembly work over ones that cluster alignment signatures — that distinction predicted performance better than tool age did.

Assemblies

One PAV detail that is easy to miss and expensive to discover: a reference without alternate loci is essential. Alt loci, decoys and patches produce unexpected results and lose sensitivity. The same warning applies more broadly than PAV, and it is the second most common coordinate mistake I see after the ARS-UCD version confusion.

Short reads

The stronger statement is the one to internalise. A cattle structural variation review published in November 2025 puts the numbers plainly: short-read methods detect only 30–70% of structural variants, with false discovery rates reaching 85%, because read-pair, read-depth and split-read signals cannot resolve repetitive or structurally complex regions — which is precisely where structural variants are enriched. If you have short reads and no graph, you are not doing SV discovery. You are doing SV genotyping against whatever catalogue you can borrow, and it is more honest to describe it that way.

Benchmarking and QC

Beyond the deprecated-benchmark problem covered above, the practical advice is that graph quality assessment is the step people skip and then pay for. A graph can be structurally fine by every summary statistic and still be wrong for your purpose — over-collapsed in a segmental duplication you care about, or carrying a contig assigned to the wrong chromosome. Node and edge counts, path coverage, and a growth curve across added haplotypes take minutes and catch most of it.

Imputation and association

The route that works for population-scale SV association in livestock is unglamorous and now fairly well established: genotype SVs from the graph on the few thousand animals with short-read whole genomes, build a combined SNP–SV reference panel, impute SV dosages onto the array-genotyped population with Beagle, filter on imputation R², and run a mixed model. In dairy cattle that yields roughly 87% of variants surviving at R² ≥ 0.8, and R² ≥ 0.8 has become the de facto reporting threshold in this literature.

Two honesty notes that belong in any paper doing this. Structural variants show substantially stronger genome-wide enrichment than SNPs across most complex traits — and they add a fraction of a percent to the genetic variance explained. Both are true. Reporting only the first is how this subfield loses credibility.

Workflow and environments

The only strong claim I will make in this section is that the environment specification matters more than the workflow engine. A Snakemake pipeline with a pinned, locked environment is reproducible; a beautifully written Nextflow pipeline that resolves bioconda::sniffles at runtime is not, and will one day silently pull a version that was withdrawn.

Cattle-specific resources

The ARS-UCD point deserves repeating because it is the one that silently ruins analyses. Two collaborators can both say "the cattle reference" and mean different coordinate systems, and nothing will fail loudly — you will simply get a QTL in the wrong place. Put the accession, not the name, in your methods.

The stack I would actually run

For a livestock structural variation project starting today, aiming at trait association in a population that only has short reads and array genotypes:

1. Assemble the reference panel with hifiasm, phased with Hi-C or trio data. Enough animals to saturate within the breed; adding a second breed to the same graph is a different decision with different consequences.
2. Build a breed-specific graph with Minigraph-Cactus, on a no-alt reference, with masking and contig filtering handled explicitly.
3. Assess the graph before trusting it — odgi statistics, panacus growth curves, and a look at a region you already know the answer for.
4. Genotype the short-read cohort with PanGenie or vg giraffe, and treat disagreement between them as information rather than a nuisance.
5. Impute to the array-genotyped population with Beagle 5.5 against a combined SNP–SV panel; report per-variant R² and filter at 0.8.
6. Associate with GCTA MLMA or regenie depending on trait type, then fine-map, and report the variance explained alongside the enrichment.
7. Benchmark against HG002 v5.0q if you are validating methodology, stratified by low-complexity region, and say out loud if you are reporting against the deprecated v0.6.

Last reviewed September 2026. Versions and maintenance status were checked at that time and will drift. Corrections are welcome and will be credited.