Five Ways an Ancient Genome Gets Lost

Ancient genomes are rarely lost in any dramatic sense. The sequencing files usually still exist somewhere; the paper is still online; the specimen is probably still in a museum drawer. What gets lost is quieter and more damaging: the connection between those things. A researcher downloads a genome and can no longer say with confidence which specimen it came from, which version they are using, what was removed before it was deposited, or whether the date attached to it is still the date the original authors would give.

In the first Origins article, The Sequence Is Not the Evidence, we argued that an ancient genome is the product of an evidence chain — site, context, specimen, sample, library, sequencing, processing, dataset, publication and reanalysis — and that the chain, not the sequence, is the evidence. This piece asks a more practical question: where do those chains actually break?

The answer is not speculative. Over the past few years the ancient DNA community has begun to audit its own archives, and the findings are remarkably consistent. Five kinds of break recur.

1. The individual gets a new name

A single ancient individual can accumulate identifiers the way a traveller accumulates stamps. There is the excavation or grave number assigned on site; the accession number given by the museum that holds the remains; the sample code used by the laboratory; separate identifiers for each sequencing library; sample, experiment and run accessions in a public sequence archive; the name used in the publication’s figures; and, if the individual enters a community dataset, an identifier assigned there.

None of this is careless. Each identifier makes sense within the system that created it. The problem is that the relationships between them are often recorded nowhere except in the heads of the people who created them, or in a supplementary table that may or may not survive.

Community resources work hard to manage this. The Allen Ancient DNA Resource (AADR), which by the end of 2022 compiled genome-wide data from more than 10,000 ancient individuals, distinguishes between identifiers for an individual and for a particular version of their data, and is explicit that these can change: its documentation notes that IDs “may change from data release to data release”, and that if data under two different identifiers are found to come from the same individual, “they are merged into a single Master ID”. That is exactly the right thing to do. But it means that an analysis run on one release, citing one identifier, may not map cleanly to the next — unless the history of the change has been kept.

What breaks: the identity of the individual across systems and over time.

2. The paper and the archive disagree

If identifiers are the joints of an evidence chain, the link between a publication and the public sequence archive is the joint that most often fails.

In 2024, Anders Bergström published an audit of the archived data from 42 ancient genomics studies that appeared in Nature, Science and Cell between 2021 and 2023. In 15 of those studies — more than a third — the accession numbers for samples in the archive did not match the way the samples were described in the paper. Thirty-three of the studies archived aligned reads, and 27 of those recorded a sample identifier in the file headers; in 18 cases that identifier matched nothing in the archive metadata, and in five cases a single file carried tags for more than one sample.

The ancient metagenomics community found the same problem when it built AncientMetagenomeDir, a community-curated list of published ancient microbial and environmental samples. Its authors reported “errors in some data accessions uploaded to public repositories, such as multiple sample codes assigned to different libraries of the same sample, and insufficient metadata to link accessions to specific samples reported in a study”.

A mismatch like this does not mean the data are wrong. It means a later user has to work out, by inference, which file belongs to which individual — and record the answer somewhere that may never be shared.

What breaks: the link between a scientific claim and the data that support it.

3. The raw evidence was never archived

The third break is a deletion that happens before anything reaches the archive.

Ancient DNA analysis typically aligns sequencing reads to a reference genome and keeps the reads that match. For a human population study, that is sensible: the analysis concerns human DNA. But it means that the reads that did not align — often the large majority in an ancient sample — may be discarded before the data are deposited.

Bergström’s audit found that only 20 of the 42 studies archived raw sequencing reads. Of 29 studies that used capture enrichment, only 10 deposited raw reads. All 33 studies that archived aligned reads had removed the unaligned ones, and 26 had applied further filters on mapping quality or read length. And 23 of the 42 studies did not archive data from all the samples they studied, typically omitting samples that were screened but performed poorly.

Those choices have consequences that the original authors may not have intended. The unaligned reads from an ancient human sample are exactly where DNA from ancient pathogens, oral microbes and environmental organisms is found. Some of the most important ancient pathogen discoveries have come from re-examining sequencing data generated for other purposes: in 2015, for example, Rasmussen and colleagues identified Bronze Age Yersinia pestis, the plague bacterium, by screening shotgun data from 101 ancient Eurasian individuals. When only human-aligned reads are archived, that future is closed off. Samples that “failed” are part of the evidence too: they tell later researchers about preservation, contamination and the selection that produced the published set.

What breaks: the chain back to the raw observation, and every question that was not asked at the time.

4. The context stays behind

An ancient genome without its archaeological context is a sequence from nowhere and no time. Yet context is the part of the chain most likely to be left out of the archive.

In Bergström’s audit, only three of the 42 studies provided latitude and longitude for their samples in the sequence archive, and only one provided radiocarbon laboratory codes — the identifiers that allow a date to be traced back to the laboratory that measured it. The information usually exists, but in the paper’s supplementary material: a spreadsheet, a PDF, sometimes a narrative paragraph describing a burial.

Supplementary material is fragile. It is often unstructured, rarely versioned and not always preserved as well as the main article. More broadly, a well-known study by Vines and colleagues, which requested the data behind 516 biology papers published between 1991 and 2011, found that where authors reported the status of their data, the odds of a dataset still existing fell by 17% for every year since publication, and the odds of finding a working email address for its authors fell by 7% a year. Context that lives only with the authors, or only in an attachment, has a half-life.

What breaks: the link between the genome and the place, time and burial it came from.

5. The record changes, but the copies don’t

The final break is subtler, because it happens when things go right.

Ancient DNA records are corrected. A contaminated sample is flagged; two individuals turn out to be one; a sample is reassigned to a different site. Dates change too: radiocarbon measurements are recalibrated when the international calibration curves are updated — the current Northern Hemisphere curve, IntCal20, was published in 2020 — and archaeological dates can be revised as new evidence emerges. The AADR states plainly that it relies on “ongoing curation of the dataset as well as feedback from the user community” to identify errors, “which we then seek to correct in subsequent releases”. Where an individual has been published more than once, it typically includes only the best-quality version.

That is good practice. But corrections propagate forward, not backward. A study published last year may have used an earlier release; a supplementary table in a third paper may have copied the old date; a derived dataset may preserve a merged identifier as two individuals. None of these copies knows that the source has changed. Unless each analysis records exactly which version of which record it used, a later reader cannot tell whether a result reflects the current state of knowledge or a superseded one.

What breaks: the link between a result and the version of the evidence it was built on.

What the five breaks have in common

Look closely and none of these is primarily a missing piece of data. The sequences exist. The identifiers exist. The dates exist. What is missing is the relationship between them: this library came from that sample; this accession corresponds to that individual in the paper; this file was filtered in this way; this date was calibrated with that curve; this analysis used that release.

Metadata can describe each object perfectly and still fail to capture those relationships. That is the distinction we drew in the first Origins article between metadata, which describes a thing, and provenance, which describes what happened to it. The five breaks above are provenance failures.

It is also worth being clear about what is already being done. The community has responded to these problems with care and considerable effort. The AADR curates identifiers and corrections across releases. AncientMetagenomeDir provides structured, validated metadata linking published samples to archive accessions. The Poseidon framework packages genotype data together with archaeological and laboratory context in a standard, machine-readable form. Bergström’s audit ends with six practical recommendations for authors, including archiving all sequencing reads rather than only aligned ones, providing accurate metadata linking samples, libraries and runs, supplying coordinates and dating information, and archiving data from low-coverage and negative experiments.

Ancient BioChain is intended to work alongside these efforts, not to replace them. What a provenance layer adds is a way of recording the relationships themselves — as events that can be traced, versioned and queried — so that when an identifier changes, a date is revised or a file is re-filtered, the history travels with the evidence instead of being lost between systems. Our white paper, A Provenance Layer for Ancient DNA, sets out how that model works and proposes a minimum provenance record designed to close each of the five breaks described here.

What we will do next

Over the coming months, Origins will put this into practice by following a single, well-published ancient individual from excavation record to community dataset, documenting every identifier, every transformation and every point at which the chain has to be reconstructed by inference. The breaks we find will be as informative as the links.

A genome is rarely lost all at once. It is lost one broken link at a time.

References

1. Bergström, A. (2024) ‘Improving data archiving practices in ancient genomics’, Scientific Data, 11, 754. https://doi.org/10.1038/s41597-024-03563-y
2. Fellows Yates, J.A., Andrades Valtueña, A., Vågene, Å.J., et al. (2021) ‘Community-curated and standardised metadata of published ancient metagenomic samples with AncientMetagenomeDir’, Scientific Data, 8, 31. https://doi.org/10.1038/s41597-021-00816-y
3. Mallick, S., Micco, A., Mah, M., et al. (2024) ‘The Allen Ancient DNA Resource (AADR): a curated compendium of ancient human genomes’, Scientific Data, 11, 182. https://doi.org/10.1038/s41597-024-03031-7
4. Rasmussen, S., Allentoft, M.E., Nielsen, K., et al. (2015) ‘Early divergent strains of Yersinia pestis in Eurasia 5,000 years ago’, Cell, 163(3), pp. 571–582. https://doi.org/10.1016/j.cell.2015.10.009
5. Reimer, P.J., Austin, W.E.N., Bard, E., et al. (2020) ‘The IntCal20 Northern Hemisphere radiocarbon age calibration curve (0–55 cal kBP)’, Radiocarbon, 62(4), pp. 725–757. https://doi.org/10.1017/RDC.2020.41
6. Schmid, C., Ghalichi, A., Lamnidis, T.C., et al. (2024) ‘Poseidon – A framework for archaeogenetic human genotype data management’, eLife (reviewed preprint RP98317). https://elifesciences.org/reviewed-preprints/98317
7. Vines, T.H., Albert, A.Y.K., Andrew, R.L., et al. (2014) ‘The availability of research data declines rapidly with article age’, Current Biology, 24(1), pp. 94–97. https://doi.org/10.1016/j.cub.2013.11.014