The term pangenome was coined in 2005 for a gene-counting problem in bacteriology. The graph was built a decade later for a read-mapping problem in human genetics. A 2018 consortium review widened the definition to cover both, which is why the two get confused. That history is in the companion collection on where the pangenome came from.
Two things, one word
A pangenome is a collection: assembled genomes plus a catalogue of what is present in which of them. Outputs are core and dispensable gene sets, presence/absence tables, and sequence missing from the reference. The unit is a gene or a locus.
A pangenome graph is a data structure: those genomes embedded as paths, so a read can be aligned against all of them at once. The output is an alignment; genotypes and calls are derived from it. The unit is a read.
The question that resolves it_
What is the unit of the answer? A gene or a locus in genomes already assembled: a collection. A sample reachable only through its reads: a graph.
When a collection is enough
- The question is gene content. Core, shell and cloud partitions, gene-family presence/absence, repertoire size. Made on gene models, not graph nodes.
- Describing diversity, not genotyping new samples. A species description or breed survey delivers a catalogue.
- Few assemblies. Below roughly ten divergent assemblies, a graph buys little. The benefit scales with the number of unassembled samples it will serve.
- Downstream tools are fixed. Breeding pipelines, browsers and annotation need coordinates on a familiar reference.
When only a graph works
- SV genotypes from thousands of short-read samples. The strongest case, and the one that matters for livestock. Reads carrying a non-reference allele misalign, so the variant is invisible in the samples that carry it. Giraffe genotyped known SVs across 5,202 human genomes (doi:10.1126/science.abg8871).
- Reference bias is the problem itself. Immune regions, segmental duplications, populations far from the reference. The first demonstration chose the MHC (doi:10.1038/ng.3257).
- Discovery beyond the reference. A graph carries alleles the reference lineage never had.
- Haplotype structure matters. A per-sample reference can be sampled from the graph before alignment (doi:10.1038/s41592-024-02407-2).
What the graph costs
- Coordinates. The largest cost. A graph has no agreed address system, so results must be projected back, and projection is lossy in the complex regions the graph was built for.
- It is a modelling choice. Same assemblies, different builder, different variant set (doi:10.1186/s13059-023-02969-y).
- Complexity compounds. More divergent genomes tangle repeat-rich regions, and the usual fix is to mask the loci that motivated the build.
- Maintenance. Graph pipelines need specialist knowledge, and that does not end at publication.
The decision table
| Question | Use | Why |
|---|---|---|
| Genes missing from the reference | Collection | Needs gene models, not read alignments |
| Size of the gene repertoire | Collection | A counting problem |
| SV genotypes in thousands of samples | Graph | The one case nothing else covers |
| SNP calls in divergent samples | Graph | The original, measurable gain |
| One hard locus: MHC, a CNV cluster | Locus graph | Build it where the problem is |
| SV to trait in a breeding population | Both | Graph genotypes, linear coordinates |
| New assembly with a diversity description | Collection | A graph adds cost, not an answer |
| Annotation transfer, regulatory analysis | Linear reference | No graph annotation standard yet |
What pipelines actually do
Most production work treats the graph as an intermediate, not a reference: build it from assemblies, align short reads to it, genotype in graph space, then project onto linear coordinates so the rest of the pipeline still works.
So it is a sequence, not a choice. Build the collection first. It is cheaper, publishable on its own, and it answers the question the graph decision depends on: how much sequence is missing from the reference in this population.
Five questions first
- How many samples will go through this that will not be assembled? If none, the graph has no one to serve.
- How far is the population from the current reference?
- What consumes the output, and does it accept anything but linear coordinates?
- Would a linear SV caller on long reads reach the same answer for less?
- Who maintains this in two years?
A synthesis of abstracts indexed in PubMed, not a benchmark. Nothing here was tested on data. Quantitative claims belong to the cited paper.