A linear reference is one haplotype standing in for a species. What changed between 2019 and 2026 is not that the field noticed, but that tools, assemblies and file formats improved enough to work without one. Three phases: making graphs computable, releasing plural references, then getting answers out of them.
2019 - 2020 Making graphs computable_
Aligning a read to a graph was too slow to use at genome scale, and there was no agreed pangenome file format. This period fixed both.
-
Bit-parallel sequence-to-graph alignment
Rautiainen, Mäkinen, Marschall
Generalised two classic linear-sequence algorithms to graphs, giving 3 to 20-fold speedups over the previous asymptotically optimal method. Unglamorous, and the reason graph alignment stopped being a thought experiment.
-
HUPAN: a pan-genome analysis pipeline for human genomes
Duan and colleagues
An early attempt to run pangenome logic on human data. Worth remembering because it shows the pre-graph framing - pangenome as a set of sequences to add to a reference, rather than as a graph.
-
The design and construction of reference pangenome graphs with minigraph
Li, Feng, Chu
The pragmatic design that won. Minigraph keeps the linear reference coordinate system while encoding tens of thousands of structural variants missing from it, which is exactly the compromise that let existing pipelines adopt graphs incrementally instead of all at once.
-
GraphAligner: rapid and versatile sequence-to-graph alignment
Rautiainen and Marschall
13 times faster and 3 times less memory than the state of the art for long reads on graphs. The speed is what made graph-based error correction and genotyping tractable on real datasets.
-
Pangenome Graphs
Eizenga and colleagues
The review that gave the field a shared vocabulary - bidirected graphs, paths, snarls, bubbles. Terminology is infrastructure, and a lot of later interoperability traces back to everyone using these words the same way.
-
Bovine breed-specific augmented reference graphs facilitate accurate sequence read mapping and unbiased variant discovery
Crysnanto and Pausch
The first variation-aware reference for an agricultural animal, and the source of a finding people still get wrong: graphs built from random variants do not beat the linear reference. Only pre-selected, frequency-filtered variants help.
-
Pan-Genome of Wild and Cultivated Soybeans
Liu and colleagues
Plant genomics reached graph pangenomes before animal genomics did, driven by the sheer structural divergence between crop accessions. Worth reading across the species boundary: the crop field solved several practical problems first.
-
A Graph-Based Pan-Genome Guides Biological Discovery
Tao and colleagues
Made the case that a graph is not just a better reference but a discovery instrument, by surfacing variation that mapping to a single accession had silently hidden.
-
Telomere-to-telomere assembly of a complete human X chromosome
Miga and colleagues
Not a pangenome paper, but a precondition for one. A graph of incomplete assemblies inherits every gap; closing centromeres and satellite arrays is what made the later haplotype-resolved assemblies worth graphing.
2021 - 2022 Proving the reference could be plural_
Assemblies became cheap and haplotype-resolved as the graph libraries became fast. Species got their first graphs, and the result was how much sequence the linear references had been missing.
-
Efficient dynamic variation graphs
Eizenga and colleagues
The libbdsg and libhandlegraph stack, up to an order of magnitude better than the previous vg implementation on speed and memory. The vg toolkit moved onto it. Boring, load-bearing, and the reason the 2023 human pangenome could be handled on ordinary hardware.
-
Pangenomics enables genotyping of known structural variants in 5202 diverse genomes
Sirén, Monlong and colleagues
Giraffe: a short-read mapper that threads haplotypes through the graph at roughly single-reference speed. 167,000 structural variants found by long reads, genotyped in 5,202 short-read genomes. This turned pangenomes into a way of getting SV genotypes out of data people already had.
-
Pan-genome analysis of 33 genetically diverse rice accessions reveals hidden genomic variations
Qin and colleagues
Hidden is the operative word. Rice had one of the best-studied reference genomes in agriculture and the pangenome still found substantial variation that the reference could not represent.
-
The Human Pangenome Project: a global resource to map genomic diversity
Wang, Antonacci-Fulton and colleagues, Human Pangenome Reference Consortium
The consortium stating its case: the current reference is a linear composite of merged haplotypes from more than 20 people, one of whom contributed most of the sequence. Sets out a graph-based, telomere-to-telomere goal, and puts ethical frameworks and global partnership in the design rather than the acknowledgements.
-
A cattle graph genome incorporating global breed diversity
Talenti and colleagues
Only 8 percent of cattle are in Europe, yet European breeds dominated the genomic resources. This graph added 116.1 Mb of sequence absent from the Hereford reference, most of it African, and improved mapping rates and SV calling agreement. The clearest demonstration in livestock that reference bias is also a representation-equity problem.
-
ODGI: understanding pangenome graphs
Guarracino and colleagues
Tooling for inspecting and manipulating graphs rather than building them. The field had reached the point where you needed to look at a graph to debug it, which is a sign of maturity.
-
From telomere to telomere: the transcriptional and epigenetic state of human repeat elements
Hoyt and colleagues
Showed what a complete assembly buys: satellite arrays, composite repeats and retroelement transductions that were simply absent from GRCh38. These are exactly the regions where pangenome graphs become tangled and multiallelic, and where they matter most.
2023 The year the reference changed_
A draft human pangenome, the two construction methods behind it, and the finding that the method chosen changes which variants you get.
-
A draft human pangenome reference
Liao and colleagues, Human Pangenome Reference Consortium
The centre of gravity for everything since. Whatever its population gaps, this is the artefact that made graph-based analysis a default option rather than a bespoke project, and every 2024-2026 method paper benchmarks against it.
-
Pangenome graph construction from genome alignments with Minigraph-Cactus
Hickey and colleagues
The method behind the HPRC graph. Combines minigraph's structural backbone with base-level alignment, which is what makes small variants and structural variants representable in one object.
-
Unbiased pangenome graphs
Garrison and colleagues
The pggb alternative: build the graph from all-to-all alignment with no reference privileged at all. Philosophically the cleaner answer, practically the more expensive one, and the two approaches still coexist because they fail differently.
-
Graph construction method impacts variation representation and analyses in a bovine super-pangenome
Leonard, Crysnanto, Mapel, Bhati, Pausch
The paper most worth pressing on colleagues. Built the same bovine data with pggb, cactus and minigraph: 221k structural variations recovered, only 61 percent common to all three. Your graph construction method is a analysis parameter, not an implementation detail, and almost nobody reports it as one.
-
A pangenome reference of 36 Chinese populations
Gao and colleagues
The direct answer to the HPRC's population coverage. Establishes the pattern the field is now in: not one pangenome, but several, with the interesting questions at the seams between them.
-
A sheep pangenome reveals the spectrum of structural variations and their effects on tail phenotypes
Li, Gong and colleagues
The livestock result that made the case to breeders. 130.3 Mb of non-reference sequence, and critically: nearly half the SVs show low to moderate linkage disequilibrium with surrounding SNPs, and most cannot be tagged by the ovine 50K chip. Then a causal 168-bp insertion for the long-tail trait. That is the whole argument for SV genotyping in one paper.
-
Chromosome-level genome assembly of Korean native cattle and pangenome graph of 14 Bos taurus assemblies
Jang and colleagues
528 Mb of non-reference sequence across 14 assemblies, 62 Mb of it Hanwoo-specific. A concrete measure of how much a single-breed reference costs you in a species with deep population structure.
-
Accurate and fast graph-based pangenome annotation and clustering with ggCaller
Horsfield, Tonkin-Hill, Croucher, Lees
A reminder that the bacterial pangenome literature is a separate tradition using the same word. Here the graph is a de Bruijn graph over a population and the payoff is consistent gene calling across thousands of genomes.
2024 - 2026 From representation to inference_
Genotyping, calling and association. The gains are clear for discovery and interpretation, much smaller for prediction.
-
Personalized pangenome references
Sirén and colleagues
The insight that a pangenome contains variants irrelevant to your sample, and that those cause false mappings. Sampling a personalised subgraph from k-mer counts cut small-variant genotyping errors fourfold versus GATK and made short-read SV genotyping of known variants competitive with long-read discovery. One of the largest single accuracy jumps in this whole list.
-
Haplotype-aware sequence alignment to pangenome graphs
Chandra, Gibney, Jain
Treats a read as an imperfect mosaic of known haplotypes and penalises implausible recombinations, borrowing the logic of genotype imputation. Also proves hardness bounds, which is rarer than it should be in this field.
-
Global pangenome analysis highlights the critical role of structural variants in cattle improvement and identifies a unique event as a novel enhancer in IGFBP7+ cells
Dai, Zhao and colleagues, with Liu Yang and George E. Liu
2,409 cattle from 82 breeds through a pangenome graph. The number that matters for anyone running SNP arrays: about 40 percent of SVs are not tagged by nearby SNPs at r-squared above 0.6. Those orphan SVs are invisible to conventional genotyping, and one of them turns out to be an IGFBP7 enhancer regulating fat deposition. Liu Yang is an author on this one.
-
From genotype to phenotype with 1,086 near telomere-to-telomere yeast genomes
Loegler and colleagues
The cleanest existing test of what pangenome variation buys an association study, because yeast lets you do it at species scale. Including structural variants and indels improved heritability estimates by 14.3 percent on average over SNPs alone, and SVs were more often associated and more pleiotropic. A realistic ceiling estimate for what SV-aware GWAS gains elsewhere.
-
Pangenome-aware DeepVariant
Asri and colleagues, UCSC and Google
Feeds pangenome haplotypes into the caller itself rather than only the aligner, cutting errors by up to 25.5 percent versus linear-reference DeepVariant. Signals where this is heading: the graph stops being a preprocessing step and becomes part of the model.
-
Culture-independent meta-pangenomics enabled by long-read metagenomics reveals associations with pediatric undernutrition
Minich and colleagues
Complete metagenome-assembled genomes at scale make microbial pangenome GWAS possible from stool. Included because it shows the concept escaping genomics proper into clinical microbiome work.
-
Defining and cataloging variants in pangenome graphs
Salehi Nowbandegani, Zhang, Hu, Li, O'Connor
Settles what a variant even means inside a graph, via a reference tree spanning the nodes. 29.6 million variants in the HPRC graph, of which 11.7 percent have a reference allele GRCh38 does not contain - unreachable in principle by a linear pipeline. The definitional work the field skipped for six years.
-
COSIGT: population-scalable genotyping of complex loci from low-coverage sequencing data using pangenome graphs
Bolognini and colleagues, Human Technopole
Matches read-depth shape rather than counts, so complex loci become callable at 1-2X. This is the paper that could change livestock economics, where cohorts are sequenced shallow on purpose.
What is actually settled, and what is not_
Settled
- Linear references lose real sequence, and the amount is species-specific and large. 116 Mb in global cattle, 130 Mb in sheep, 528 Mb across 14 taurine assemblies. This is not in dispute any more.
- Graph construction is an analysis choice with consequences. Three methods on the same bovine data agreed on only 61 percent of structural variants. Reporting your construction method is not optional.
- A meaningful fraction of structural variation is untaggable by SNPs. Roughly 40 percent in global cattle, nearly half in sheep. Array-based genotyping cannot see it, whatever the array density.
- Personalisation and haplotype-awareness beat raw graph size. Sampling the subgraph relevant to your sample outperforms throwing the whole pangenome at it.
Still open
- Prediction accuracy. Discovery and interpretation improve substantially; routine genomic prediction gains remain marginal. Whether that is a ceiling or a tooling problem is open.
- Population representation. Every pangenome is a sample of haplotypes, and the sampling is not neutral. A pangenome of a different population is a second, differently biased object.
- Coordinates and interchange. There is still no agreed way to say where a variant is in a graph, which blocks the cross-study aggregation that made SNP meta-analysis powerful.
- Functional annotation. Variant discovery has run far ahead of knowing what any of it does. This is the rate-limiting step for breeding or clinical decisions.