What the Garum Vat Left Behind

The Smelliest Sample in Archaeology

Picture the actual object first, because the paper is really about the object. A Roman fish-salting works on the Galician coast, four rectangular vats cut into rock and lined with waterproof plaster, standing empty for seventeen hundred-odd years and then excavated. Vat one still held the bottom of what had been in it: a compacted mass of crushed fish bone and vertebrae, residue from a batch of garum that was ladled off as liquid centuries ago, leaving behind what nobody bothered to take with them.

Every step that produced this material was designed, more or less by accident, to be bad for DNA. The fish were gutted and packed in salt. They were crushed, deliberately, to speed up the process. They sat in their own acids and in the digestive activity of microorganisms for weeks or months, at ambient outdoor temperature, in a Mediterranean-adjacent Atlantic climate that does bone preservation no favours. Compare that to how most ancient DNA is recovered: a buried skeleton, protected by soil, cool and dark and mercifully un-fermented. Garum residue is close to the worst realistic substrate you could hand a molecular lab and still call bone.

A team led by Gonçalo Espregueira Themudo sequenced it anyway, and published the result in Antiquity in the summer of 2025 — not as a curiosity, but as a demonstration that an entire category of archaeological material, industrial fish-processing waste, sitting in storage boxes at excavated sites across the former Roman world, might be usable for population genetics after all [1].

What Garum Actually Was

Garum is usually introduced to modern readers as a punchline: the Romans put rotten fish sauce on everything. That undersells it. It was closer to what soy sauce or fish sauce is in parts of Asian cuisine today, a fermented, intensely savoury condiment used constantly, in small amounts, across a huge range of dishes — not a garnish for the adventurous but a basic pantry item, made in enormous industrial quantities and shipped around the Mediterranean and Atlantic in amphorae as one of the most heavily traded foodstuffs of the Roman world.

Robert Curtis's standard economic history of the industry lays out the production logic: small pelagic fish, usually whichever species schooled abundantly and cheaply near a given coast, were mixed with salt at a preserving ratio and left to autolyze, their own digestive enzymes and ambient microbes breaking the flesh down over weeks in the sun, before the liquid was drawn off and graded by quality — garum and liquamen at the top, the paste left behind sold on as allec for people who could not afford the good stuff [2][3]. This was not a cottage industry. The archaeological record includes fish-salting complexes with dozens of vats standing along entire stretches of coastline, from Baetica in southern Spain up through Lusitania and Galicia to Brittany, exporting a manufactured commodity the way an olive-oil or wine estate did [4][5].

Which species went into a given vat depended almost entirely on what a coastal fleet could catch locally and cheaply, in bulk, without refrigeration or a long supply chain. That is exactly the kind of decision archival texts rarely record and bone morphology only sometimes settles. It is also exactly the kind of decision population genetics can speak to, because a fishery's catch carries the genetic signature of the population it was drawn from.

Four Vats on a Galician Headland

Adro Vello sits on the O Grove peninsula in Galicia, the north-westernmost corner of Iberia, a stretch of Atlantic coast that supported a dense chain of Roman fish-processing sites from roughly the first century AD onward. The site had already been excavated and its stratigraphy worked out before this study; what changed in 2025 was somebody deciding the sludge at the bottom of vat one was worth a DNA extraction rather than a discard bag [1].

Our own arithmetic, not the paper's: subtract those calibrated ranges from the present. Cal AD 84–394, the wide 95.4% interval, is roughly 1,632 to 1,942 years before 2026. The tighter 68.2% interval, cal AD 162–321, is roughly 1,705 to 1,864 years before 2026. Loosely rounded, that is the same "about seventeen hundred years old" figure that circulates in casual summaries of this kind of find — it sits at the young edge of the tighter interval, not its centre, but it is not a wrong rounding, just an optimistic one.

Why Nobody Expected DNA to Survive This

Ancient DNA degrades by a handful of well-understood mechanisms, and garum production hits several of them at once. The paper is explicit about the two it considers dominant here. First, the fermentation process itself creates an acidic environment, which speeds up depurination — the loss of purine bases from the DNA backbone that leaves the molecule prone to breaking at that exact point — and therefore accelerates fragmentation beyond what age alone would produce [1]. Second, whatever microbial community did the fermenting brought its own endonucleases, enzymes that cut DNA, adding a second and independent route to degradation on top of ordinary chemical decay.

Stack those on top of the crushing (physical shearing of the bone and whatever DNA it still held) and centuries of salt exposure (osmotic stress, and salt's own complicated relationship with nucleic acid stability), and the honest prior going into this project was that there might be nothing left worth sequencing. The reason it worked at all is the same reason ancient DNA work in general has gotten dramatically better over the last decade: shotgun sequencing does not need long, intact strands. It needs enough short fragments, however degraded, to pile up into confident coverage once you have enough of them and a reference genome to map them against.

From Crushed Bone to Sequencer

The wet-lab side of this is standard ancient-DNA protocol, chosen specifically because it tolerates degraded material. Bone samples were extracted using a buffer built around EDTA and proteinase K — EDTA to chelate the calcium that locks DNA inside the mineral matrix of bone, proteinase K to digest the collagen and other proteins holding everything together — then built into double-stranded Illumina sequencing libraries and run through shotgun sequencing on a NovaSeq X.

Shotgun sequencing, rather than targeting specific genes with primers, just sequences whatever DNA is in the tube: sardine DNA, bacterial DNA from the fermentation microbiome, environmental DNA from the vat, human DNA from centuries of handling. Only a fraction of any given sequencing run in a project like this belongs to the fish you actually care about. That fraction is the endogenous DNA percentage, and it is the first number that tells you whether an extraction was worth the reagents.

From there, mapped reads were run through ANGSD for genotype likelihoods, consensus sequence calling and SNP discovery [6], mitochondrial genomes were assembled and placed on a phylogeny with BEAST [7], and nuclear population structure was estimated with PCAngsd and NGSadmix [8][10], tools built specifically for the low- and uneven-coverage data that ancient samples produce, where you often cannot call a genotype with confidence at every site and have to work with the probability of one instead. Relationships among the mitochondrial haplotypes themselves were visualised as a median-joining network in PopART [12], the standard way to show how closely a set of mitogenomes are related without forcing them onto a fully resolved tree.

Proving the DNA Was Actually Roman

None of that matters if the DNA in the tube is modern contamination rather than the Roman-era fish. Ancient DNA has a signature that contamination does not: over centuries, cytosine bases near the ends of DNA fragments deaminate into uracil, which sequencing machinery reads as thymine, producing a characteristic excess of C-to-T substitutions concentrated at the start of sequencing reads and tapering off toward the middle. The paper reports exactly this pattern, checked with DamageProfiler [9], which is the standard software for quantifying it. A modern contaminant fragment does not carry that damage signature; an authentically old one does, reliably, because deamination is a slow chemical clock that keeps ticking as long as the molecule exists.

The Ledger: Eighteen Fragments, Twelve Genomes

Our arithmetic on the paper's own counts: 17/18 = 94.4% of extractions produced usable sequence, an unusually good hit rate for a substrate this hostile. Of those, 12/17 ≈ 70.6% had enough material to justify the deeper, more expensive sequencing needed for nuclear population genetics rather than mitochondrial genealogy alone. And the spread inside that success is enormous: the best-preserved fragment yielded a mitochondrial depth 197.6 / 1.3 ≈ 152× higher than the worst, and an endogenous DNA fraction 21 / 4 = 5.25× higher. Same vat, same species, wildly different survival — which is itself a finding, since it means preservation inside one deposit is not a single number you can assume in advance.

A Name Bones Already Half-Knew

It would be a cleaner story if the DNA revealed a species nobody suspected. It did not, quite, and the honest version of that is more interesting than a reveal would have been. Morphological analysis of the Adro Vello bones had already identified the assemblage as mostly Sardina pilchardus, European sardine, without attaching a firm percentage to that call [1]. Crushed, salted, partially digested vertebrae are hard to measure with confidence; "mostly sardine" was a reasonable read of the bone shapes available, not a number anyone could defend to a second decimal place.

Two other Roman Atlantic fish-salting sites had already done better than that, because their assemblages were less processed. At La Falaise, a site at Étel on the French Atlantic coast, vat one's fish remains were counted directly: 99.6% European sardine, with the residue split between Atlantic herring, whiting and Atlantic mackerel [1]. At Belém, in Portugal, sardine came out above 98% of recovered material, with anchovy, Atlantic horse mackerel, mackerel and Mediterranean moray making up the rest [1]. Those numbers exist because those bones were countable in a way Adro Vello's were not.

Where in the Atlantic These Fish Belonged

Here is the actual novelty, and the paper says so itself: this is about moving past species identification into population range and admixture, something no amount of staring at a vertebra can do [1]. A bone tells you what animal it came from. It cannot tell you which breeding population that animal belonged to, whether it was a local resident or a long-distance migrant, or how connected its population was to the ones along the neighbouring coast. Two sardine skeletons from opposite ends of the species' range look, under a microscope, like two sardine skeletons.

To ask the population question, the team compared their ancient mitogenomes and nuclear SNPs against 108 modern reference sardine mitogenomes spanning the species' known range, grouped into four broad geographic clusters used in prior sardine population genomics: a Western group, a Central group covering Morocco, Iberia and Atlantic France, a Mediterranean group, and a Canary Islands group [11]. Running principal component analysis with PCAngsd on this combined dataset, PC1 captured 7.3% of total genetic variance and PC2 a further 1.2% [1][8].

Our arithmetic: PC1 plus PC2 together account for only 7.3% + 1.2% = 8.5% of the variance in the dataset. That is a small number by the standards of, say, human population genetics, where the first couple of axes in a similarly structured sample can carry far more. It is not a weakness in this analysis; it is a known feature of small pelagic marine fish, which tend to have enormous population sizes and correspondingly weak, slow genetic drift between regions, so real, meaningful population structure can and does sit on top of comparatively little total variance [11]. Low percentages here are compatible with the clustering being real.

And it was real: on that PCA and in NGSadmix admixture modelling, the Adro Vello samples fell inside the modern Central group, alongside today's Iberian, Moroccan and Atlantic French sardines — the geographically appropriate cluster for a sardine caught off the Galician coast [1][10]. One sample stood out. Specimen 1003 showed a minor genetic contribution from the Eastern/Mediterranean side of the species' range, a small signal of gene flow that the rest of the assemblage did not share.

A Population More Isolated Than Today's

The comparison that makes this more than a species-confirmation exercise is what happens when you set the ancient cluster against modern sardines sampled from the same part of the Atlantic. The paper's own description of the result is direct: the Adro Vello samples are "much less admixed" than their modern western Iberian counterparts [1]. Roughly, the Roman-period population reads as more genetically homogeneous, more internally consistent, less mixed with neighbouring groups than the sardines swimming off the same coast today.

Sit with what that implies rather than skating past it. Two broad explanations are on the table, and the paper does not need to pick one for the observation to matter. One is that the modern Central Atlantic sardine population has become more connected to its neighbours over the last seventeen-odd centuries — more gene flow in, whether from long-distance larval drift, from changes in ocean currents, or simply from a larger, more mobile modern population mixing more freely across a wider range. The other is more directly about people: a small-boat, oar-and-sail Roman fishery working a limited stretch of coastline without refrigeration would have depended on whatever sardine schools were locally, reliably present, for a fishery that had to convert live catch into salted fish within days. A locally resident, less externally admixed population is exactly the kind of stock such a fishery would both find and, over generations of harvest, help keep locally rooted.

What This Says About How Rome Fished

Take the population result seriously as an economic clue rather than only a biological one. The scale of the Atlantic garum industry — the chain of cetariae running from Baetica up through Lusitania and Galicia to Brittany [4] — has always been read primarily from the built infrastructure: how many vats, how large, how they were plastered and drained. This paper adds a genetic constraint on the other half of that system, the part upstream of the vat: the fish themselves, and by extension the boats and crews that caught them.

A genetically distinct, comparatively closed Central Atlantic sardine population being harvested at Adro Vello is consistent with a fishery built around short-range, repeated exploitation of a nearby, locally persistent stock rather than one drawing opportunistically on whatever schools happened to be passing through from further afield. That is a modest claim, and the paper does not push it further into a full economic argument [1] — but it is a genuinely new kind of evidence for an old question in Roman economic history: was large-scale ancient fishing sustainable pressure on a stable local resource, or something closer to sequential exploitation of whatever showed up? A less-admixed ancient population looks more like the former.

What the Paper Doesn't Claim

Worth being precise about the limits here, because it is easy to over-read a result like this. The study does not reconstruct Roman fishing routes, does not quantify how large the historical Central Atlantic sardine population was, and does not claim the modern increase in admixture, if that is what it is, was caused by any specific event. It is a single vat, at a single site, with seventeen usable ancient genomes set against a modern reference panel. The authors' own framing is appropriately restrained: this is a demonstration that fish-salting residue is a viable, previously overlooked source of population-genetic data, with one worked example of what that data can show [1].

That restraint is, if anything, the more useful message for anyone thinking about what comes next. The claim is not "we solved Roman sardine population history." It is "this kind of material, previously written off, is usable," which is a much larger and more durable claim, because it is not about one vat.

What We'd Want to See Next

Three things follow fairly directly from a result shaped like this one. First, more vats: Adro Vello is one site on one coast with one radiocarbon date range; the Atlantic cetariae chain runs for hundreds of kilometres and several centuries, and each additional assemblage sequenced turns a single data point into an actual time series of population structure. Second, other species: sardine dominated this vat and the two comparison sites, but garum production used whatever was locally abundant, and mackerel, herring and anchovy residues sitting in the same storage boxes are, by this paper's logic, equally sequenceable. Third, and most directly testable, more samples per site: with only seventeen ancient genomes against 108 modern ones, the admixture comparison is suggestive rather than saturated, and a larger ancient panel would say with more confidence whether "much less admixed" is a small-sample artefact or a durable population-level difference.

None of that requires new technology. It requires somebody with access to the right storage boxes deciding, the way this team did, that industrial food waste from two thousand years ago is worth a proteinase K digest.

References

1. 1.

   Themudo, G. E., Fernández-Fernández, A., Valle Abad, P., Rodríguez Nóvoa, A. A., Fernández-Rodríguez, C., González-Gómez de Agüero, E., da Fonseca, R. R. & Campos, P. F. (2025). Roman Atlantic garum: DNA confirms sardine use and population continuity in north-western Iberia. Antiquity 99(406), 1049–1064. doi:10.15184/aqy.2025.73

2. 2.

   Curtis, R. I. (1991). Garum and Salsamenta: Production and Commerce in Materia Medica. Brill. doi:10.1163/9789004377264

3. 3.

   Grainger, S. (2020). The Story of Garum: Fermented Fish Sauce and Salted Fish in the Ancient World. Routledge. ISBN 978-1-315-26982-5

4. 4.

   Bernal-Casasola, D., Expósito, J. A. & Díaz, J. J. (2018). The Baelo Claudia paradigm: the exploitation of marine resources in Roman cetariae. Journal of Maritime Archaeology 13(3), 329–351. doi:10.1007/s11457-018-9209-z

5. 5.

   Marzano, A. (2013). Harvesting the Sea: The Exploitation of Marine Resources in the Roman Mediterranean. Oxford University Press.

6. 6.

   Korneliussen, T. S., Albrechtsen, A. & Nielsen, R. (2014). ANGSD: analysis of next generation sequencing data. BMC Bioinformatics 15, 356. doi:10.1186/s12859-014-0356-4

7. 7.

   Bouckaert, R., Vaughan, T. G., Barido-Sottani, J., Duchêne, S., Fourment, M., Gavryushkina, A., Heled, J., Jones, G., Kühnert, D., De Maio, N., Matschiner, M., Mendes, F. K., Müller, N. F., Ogilvie, H. A., du Plessis, L., Popinga, A., Rambaut, A., Rasmussen, D., Siveroni, I., Suchard, M. A., Wu, C.-H., Xie, D., Zhang, C., Stadler, T. & Drummond, A. J. (2019). BEAST 2.5: an advanced software platform for Bayesian evolutionary analysis. PLOS Computational Biology 15, e1006650. doi:10.1371/journal.pcbi.1006650

8. 8.

   Meisner, J. & Albrechtsen, A. (2018). Inferring population structure and admixture proportions in low-depth NGS data. Genetics 210, 719–731. doi:10.1534/genetics.118.301336

9. 9.

   Neukamm, J., Peltzer, A. & Nieselt, K. (2021). DamageProfiler: fast damage pattern calculation for ancient DNA. Bioinformatics 37, 3652–3653. doi:10.1093/bioinformatics/btab190

10. 10.

    Skotte, L., Korneliussen, T. S. & Albrechtsen, A. (2013). Estimating individual admixture proportions from next generation sequencing data. Genetics 195, 693–702. doi:10.1534/genetics.113.154138

11. 11.

    da Fonseca, R. R., Campos, P. F., Rey-Iglesia, A., Barroso, G. V., Bergeron, L. A., Nande, M., Tuya, F., Abidli, S., Pérez, M., Riveiro, I., Carrera, P., Jurado-Ruzafa, A., Santamaría, M. T. G., Faria, R., Machado, A. M., Fonseca, M. M., Froufe, E. & Castro, L. F. C. (2024). Population genomics reveals the underlying structure of the small pelagic European sardine and suggests low connectivity within Macaronesia. Genes 15, 170. doi:10.3390/genes15020170

12. 12.

    Leigh, J. W. & Bryant, D. (2015). POPART: full-feature software for haplotype network construction. Methods in Ecology and Evolution 6, 1110–1116. doi:10.1111/2041-210X.12410