One Reference Was Never Enough
No. 12 · 2026
Reference Genomes · Graph Genomics · Clinical NGS
One Reference Was Never Enough
For twenty years, nearly every human genome analyzed in a lab was compared against one haploid sequence — a mosaic of a handful of donors, dominated by a single one. The Human Pangenome replaces that assumption with a graph of many genomes, and it quietly changes what your variant calls are actually measured against.
Every alignment you have ever run made a hidden assumption: that the reference genome on disk is a fair stand-in for the sample in the tube. GRCh38 has been that reference for most of clinical genomics, and it works remarkably well — until it doesn't. It is a single haploid sequence: a mosaic stitched from the clone libraries of about twenty anonymous donors, yet dominated by just one of them — the library known as RP11 accounts for roughly 70% of it. Whatever its origins, it can represent only one version of each locus. Anything a patient carries that isn't on that path has to be inferred, forced to fit, or missed. The Human Pangenome Reference Consortium's 2023 draft is the first serious attempt to remove that assumption from the pipeline entirely.1
This isn't a niche academic upgrade. The choice of reference is upstream of everything — mapping, variant calling, annotation, interpretation — and reference bias is the kind of error that never announces itself. Reads that don't match get thrown away or mismapped silently, and the variant that would have explained the phenotype was never a candidate. For labs running rare-disease, oncology, or pharmacogenomic panels, that is a coverage problem hiding inside a clean-looking VCF.
01What "single-reference bias" actually costs
A linear reference forces one representation of each locus. Where human populations genuinely diverge — highly polymorphic immune loci, segmental duplications, expansions, complex structural rearrangements — a single path can't hold the variation, so short reads pile up ambiguously or fail to map. The bias also has an ancestry dimension. GRCh38 is one haploid path dominated by a single donor's haplotypes, so it can hold only a sliver of the diversity that exists across populations. The samples that diverge most from it — disproportionately those from groups under-represented in reference and variant resources — carry more sequence the reference simply doesn't contain, and accumulate more reference-mismatch artifacts as a result. That is a scientific problem and an equity problem at the same time.
The pangenome answers both by refusing to pick one sequence. Instead of a line, it builds a graph: shared sequence forms a common backbone, and where individuals differ, the graph branches into parallel paths — one per observed allele — before rejoining. A read no longer has to match "the" reference; it matches whichever path it belongs to.
02What the draft pangenome actually added
The consortium's first draft assembled 47 phased, diploid genomes — 94 haplotypes — from ancestrally diverse individuals, most with African, Latin American, and South and East Asian ancestry, and deliberately not centered on European donors.1 Relative to GRCh38, it contributed roughly 119 million base pairs of new euchromatic sequence and 1,115 gene duplications, with about 90 million of those bases coming from structural variation that the linear reference never represented.1 These are not obscure corners of the genome; a substantial fraction sits in medically relevant, structurally complex regions.
Using the draft pangenome cut small-variant discovery errors by 34% and more than doubled the structural variants recovered per haplotype — on the same short-read data.
That result is the one worth internalizing. When the HPRC re-analyzed short-read data against the graph instead of GRCh38, small-variant discovery errors fell by 34% and the number of structural variants detected per haplotype rose by 104%.1 No new sequencing, no longer reads — only a better reference. Structural variation is exactly where single-reference pipelines were weakest, so a doubling in SV recovery is a direct answer to the failure mode described in our earlier piece on the variants short reads miss. Release 2, announced in May 2025, extended the resource to more than 200 individuals, roughly a fivefold expansion of the panel.1
03The toolchain: build the graph, map to it, genotype from it
Adopting a pangenome is a pipeline decision, not a download. The ecosystem splits cleanly into three jobs — constructing the graph, aligning reads to it, and calling or genotyping variants — and different tools own each stage. You rarely build the graph yourself; the HPRC publishes reference graphs, and most labs consume those.
| Tool | Stage | What it does |
|---|---|---|
| Minigraph-Cactus | Construction | Builds base-level graphs from whole-genome assemblies at the scale of dozens to hundreds of haplotypes; produces the VCF and indexes needed downstream.2 |
| PGGB | Construction | Reference-free graph builder using all-versus-all alignment; symmetric, no single genome privileged as the backbone.1 |
| minigraph | Construction | Fast SV-level graph builder; the structural scaffold Minigraph-Cactus refines to base resolution.5 |
| vg Giraffe | Mapping | Short-read mapper that aligns to thousands of embedded haplotypes at near single-reference speed; reduces reference bias.3 |
| GraphAligner | Mapping | Sequence-to-graph aligner for long reads against the same graph reference. |
| PanGenie | Genotyping | k-mer, alignment-free genotyping across variant classes; >4× faster than mapping-based methods at 30× coverage with better concordance.4 |
A representative short-read workflow today builds on the HPRC graph, maps with vg Giraffe, and calls small variants with a graph-aware DeepVariant while genotyping known structural variants with PanGenie.3,4 The economics are not prohibitive: Giraffe was used to genotype 167,000 structural variants across 5,202 short-read genomes at roughly $1.50 of compute per sample — production scale, not a research demonstration.3
04Where it earns its place clinically
Structurally complex, medically important loci
The genes that resist linear references are often the ones clinicians most want to read cleanly: HLA and KIR immune haplotypes, pharmacogenes with pseudogene paralogs such as CYP2D6, and repeat or segmental-duplication regions behind several rare diseases. Parallel allele paths give these loci a native representation instead of a lossy projection onto one sequence.
Rare disease and the "negative" exome
A meaningful share of undiagnosed cases are structural or sit in reference-poor regions. Doubling SV recovery per haplotype is precisely the kind of gain that reclassifies a previously negative case — the reference bias was hiding the causal allele, not the sequencing.
Equity as a quality metric
Because the panel was chosen for ancestral diversity, patients from historically underrepresented populations gain the most: fewer false reference mismatches and more of their sequence actually represented. For a lab, that is not a values statement — it is a reduction in ancestry-correlated error, which is a validation and accuracy concern.
05What isn't solved yet
Three frictions are real. First, representation: VCF was designed for a linear coordinate system and still struggles to express variation nested inside a large insertion, so graph-native calls often need projection back onto a linear backbone for reporting and for the annotation databases clinicians rely on. Second, validation burden: a clinical lab changing its reference is changing the most fundamental input to every result, which means concordance studies, truth-set benchmarking, and documentation before anything reaches a patient report. Third, tooling maturity: annotation, coverage QC, and visualization ecosystems remain more complete for linear references, so a graph pipeline may still round-trip through GRCh38 coordinates at the reporting stage.
None of this argues against adoption. It argues for staging it: use the pangenome where its advantages are largest — SV-heavy panels, structurally complex loci, ancestrally diverse cohorts — and keep a validated linear path for the rest until the surrounding tooling closes the gap. The reference stopped being a fixed constant the moment there were 47 genomes in it. Treating that as a deliberate pipeline decision, rather than a default no one revisits, is the whole point.
Thinking about a graph-based reference?
Zetobit designs and validates pangenome-aware NGS workflows — Giraffe mapping, SV genotyping, and CAP/CLIA-grade concordance studies — for clinical, biotech, and pharma teams weighing the move off single-reference pipelines.
Talk through your pipeline →Publications
- Liao, WW., Asri, M., Ebler, J., et al. A draft human pangenome reference. Nature 617, 312–324 (2023). doi:10.1038/s41586-023-05896-x
- Hickey, G., Monlong, J., Ebler, J., et al. Pangenome graph construction from genome alignments with Minigraph-Cactus. Nature Biotechnology 42, 663–673 (2024). doi:10.1038/s41587-023-01793-w
- Sirén, J., Monlong, J., Chang, X., et al. Pangenomics enables genotyping of known structural variants in 5202 diverse genomes. Science 374, abg8871 (2021). doi:10.1126/science.abg8871
- Ebler, J., Ebert, P., Clarke, W.E., et al. Pangenome-based genome inference allows efficient and accurate genotyping across a wide spectrum of variant classes. Nature Genetics 54, 518–525 (2022). doi:10.1038/s41588-022-01043-w
- Li, H., Feng, X., Chu, C. The design and construction of reference pangenome graphs with minigraph. Genome Biology 21, 265 (2020). doi:10.1186/s13059-020-02168-z