A human pangenome, drafted from forty-seven genomes

The reference genome that genetics has navigated by for two decades was never really human. It was a single, mostly linear string of roughly three billion letters, assembled largely from a small number of anonymous donors, and it carried the peculiar property that long stretches of it did not match any particular person very well. By 2023, that reference was known to be missing tens of thousands of base letters in any given cohort, to under-represent structural variation, and to be biased toward the few populations that contributed most of the raw material. On 10 May 2023 the Human Pangenome Reference Consortium published a replacement: a first draft of a human pangenome, signed by Wen-Wei Liao, Mobin Asri, Jana Ebler and a large consortium team, in Nature.

The draft is not one genome. It is 47 phased, diploid assemblies, from a cohort of individuals chosen for genetic diversity rather than convenience. A phased diploid assembly means that, for each of the 47 people, the two inherited copies of every chromosome are resolved separately rather than mashed into one consensus string. The authors report that each assembly covers more than 99 per cent of the expected sequence in its genome, and is more than 99 per cent accurate at both the structural and the base-pair levels. That matters because the earlier reference was, by construction, haploid and incomplete: it could not represent a gene duplicated in one population but absent in another, because it could only carry one version of that position.

The shape of the result is a graph, not a line. Where two of the 47 genomes agree, the sequences overlap; where they diverge, the map branches. A later analysis pipeline aligns a new sample to the graph, following the matching branches, instead of being forced to squeeze that sample onto a single linear backbone. The practical payoff is quantitative. The consortium reports that the draft pangenome reduces small-variant calling errors and captures large numbers of structural variants, the insertions, deletions and rearrangements that a linear reference systematically hides because they cannot live in one string.

The deeper point is what this does to the study of human genetic difference. For twenty years, a person whose genome carried a stretch absent from the reference was analysed as if they were missing a chunk of themselves. With a pangenome, that stretch is present somewhere in the graph, and the sample can be compared to the closest matching branch. Regions that were hard to align against a single backbone, including highly repetitive and medically relevant stretches of the genome, finally come into view. The draft is explicitly a draft: the consortium’s stated goal is to extend the cohort to 350 genomes by mid-2024, which is when the graph stops being a generous sample and starts being a population-weighted map.

The limitations are worth stating plainly. Forty-seven genomes, even chosen carefully, still undersample most of the world’s populations, and the tools for aligning to a graph are younger than the tools for aligning to a line, which means real labs will need time before their pipelines trust the new map. Structural variants also remain harder to call than single letters, and a more complete reference does not by itself solve the harder problem of linking variants to disease.

What the paper settles is the principle. The single-string reference was a product of the sequencing technology of the 1990s, which could only afford to read one consensus. Long-read technology in the 2020s can afford to read both copies of both parents, and the consortium has shown that the resulting graph is more accurate and more representative at the same time. The next decade of human genetics will be run against this map, not against the old one.