Where the pangenome came from, and where it is going

1. 2005 Proceedings of the National Academy of Sciences

   Genome analysis of multiple pathogenic isolates of Streptococcus agalactiae: implications for the microbial "pan-genome"

   Tettelin H, Masignani V, Cieslewicz MJ, Donati C, Medini D, et al.

   Coined the term. Six genome sequences of group B Streptococcus showed a core genome covering only about 80 percent of any single isolate. The lasting part was the extrapolation: modelling predicted new genes would keep appearing after hundreds of genomes. A species has no single genome to sequence.

2. 2005 Current Opinion in Genetics & Development

   The microbial pan-genome

   Medini D, Donati C, Tettelin H, Masignani V, Rappuoli R

   Published weeks later by an overlapping group, and it turned the result into a concept. Core plus dispensable is stated as the general way to describe a bacterial species, with the species pangenome potentially orders of magnitude larger than any genome in it. Every core/shell/cloud plot descends from this.

3. 2016 Nature Communications

   The pangenome of an agronomically important crop plant Brassica oleracea

   Golicz AA, Bayer PE, Barker GC, Edger PP, Kim H, et al.

   The idea crosses from microbiology into breeding. Nearly 20 percent of B. oleracea genes showed presence/absence variation, enriched for disease resistance, flowering time and glucosinolate metabolism. The direct ancestor of livestock pangenomics: missing sequence is an agronomic problem.

1. 2015 Nature Genetics

   Improved genome inference in the MHC using a population reference graph

   Dilthey A, Cox C, Iqbal Z, Nelson MR, McVean G

   The proof of concept. Eight assembled haplotypes, the classical HLA alleles and 87,640 SNPs combined into a graph across the 4.5 Mb extended MHC, with sample genomes reconstructed as paths through it. The hardest locus in the human genome was chosen deliberately: that is where a graph should help most.

2. 2017 Genome Research

   Genome graphs and the evolution of genome inference

   Paten B, Novak AM, Eizenga JM, Garrison E

   Named the problem. The reference genome introduces pervasive reference bias because it excludes common variation, and the scattered graph approaches belong under one label. After this they could be argued about as one research programme.

3. 2018 Nature Biotechnology

   Variation graph toolkit improves read mapping by representing genetic variation in the reference

   Garrison E, Siren J, Novak AM, Hickey G, Eizenga JM, et al.

   vg turned the argument into software other people can run. Mapping reads to a variation graph is more accurate than mapping to a linear reference, and it works at gigabase scale. Most pangenome pipelines today are vg, descend from it, or were written to compete.

1. 2018 Briefings in Bioinformatics

   Computational pan-genomics: status, promises and challenges

   The Computational Pan-Genomics Consortium

   The hinge. A consortium spanning human genetics, oncology, plant breeding, microbiology and virology generalised the definition: a pangenome is any collection of genomic sequences analysed jointly or used as a reference. That absorbs both traditions into one term. It also named the string-to-graph shift, and was right.

2. 2020 Nature Reviews Genetics

   Pan-genomics in the human genome era

   Sherman RM, Salzberg SL

   Handed the concept to general genomics, working through pangenome efforts from bacteria to humans. A single reference genome is inadequate for many purposes now, not in principle. After this the burden of proof reversed.

1. Now

   The reference becomes a function of the sample

   The end point of removing reference bias is not one universal graph but a different reference per sample: take a subgraph matched to the individual's haplotypes, then align to that. It inverts thirty years of practice, and it is already practical.

   Anchored to Personalized pangenome references, Nature Methods

2. Now

   Genotyping at population scale is the killer application

   For thousands of low-coverage short-read samples, which is most of livestock and human population genetics, the graph's value is genotyping, not discovery. Giraffe made graph alignment fast enough for 5,202 genomes; COSIGT now targets complex loci from low coverage. The clearest cost-benefit case in the field.

   Anchored to Pangenomics enables genotyping of known structural variants in 5202 diverse genomes, Science

3. Soon

   Variant calling moves inside the graph

   Most pipelines still align in the graph and then leave it, projecting back before calling. Pangenome-aware callers and formal definitions of a variant inside a graph are both active work. The open question is identity: whether two labs looking at the same bubble agree they found the same variant.

   Anchored to Defining and cataloging variants in pangenome graphs, Cell Genomics

4. Open

   From representation to association

   Building the graph and genotyping from it are largely solved. Connecting a graph-derived SV to a phenotype is not. Association testing, LD and heritability were built on biallelic sites with coordinates, and multiallelic bubbles fit badly. The widest open gap, and the one nearest to livestock work.

   Anchored to Global pangenome analysis highlights the critical role of structural variation, Molecular Biology and Evolution

5. Open

   The pangenome stops being only about DNA

   Pantranscriptomes for expression, meta-pangenomics for communities that cannot be cultured, and eventually annotation defined on graph coordinates. Each extension re-opens the coordinate problem: a graph has no agreed address system, and the annotation ecosystem assumes one.

Settled

• Why "pangenome" means two different things. One tradition counts genes across a species; the other builds an alignment target. The 2018 consortium definition covers both deliberately, so a paper can be pangenomic in either sense without saying which.
• Why the graph turn came from human genetics rather than microbiology. Bacteria have too much gene turnover for a sequence graph to be the natural object; the graph was invented to solve reference bias in a species where the reference already dominated everything.
• Why crops led livestock by roughly five years. The Brassica-era argument that presence/absence variation carries agronomic traits was made on plants first, and livestock pangenomics inherited both the motivation and the methods.

Still open

• Whether a graph should be a reference at all, or only an intermediate that pipelines pass through on the way back to linear coordinates.
• Whether the community converges on one graph construction method. Construction choices still change which variants are recoverable, so the graph remains a modelling decision presented as a data structure.
• Whether the gene-content strand and the alignment strand ever genuinely reunify, or whether "pangenome" simply stays a word with two meanings that readers disambiguate from context.