AADR sample labels explained: .AG, .TW, .SG, .DG, .HO and no suffix

Samples in the Allen Ancient DNA Resource carry a short code after a dot: Loschbour.AG, I0231.TW, Kostenki14.SG, Han.DG. The words before the dot are archaeology. The suffix is laboratory history: it records how that genome's data were produced. .AG and .TW are two capture reagents, .SG is shotgun sequencing read one allele at a time, .DG is shotgun sequencing deep enough to read both alleles, and .HO is a present-day person typed on a genotyping array. A label with no suffix means something different depending on which file you are reading, and that is the case that catches most people out.

It matters because qpAdm can mistake a difference in laboratory method for a difference in ancestry. Below: each suffix decoded, the three meanings of "no suffix", and the selection rules that follow from them.

The suffixes at a glance#

In AADR v66.p1, the current release, the 23,104 individuals of the 2M dataset split into 11,241 .AG, 6,737 .SG, 3,800 .DG and 688 .TW, plus 498 with combined suffixes such as .AG.TW, 136 with rarer capture tags and four reference genomes. Two words in the table need unpacking first.

Pseudo-haploid. In most ancient genomes a given position is covered by one read or a few, too few to tell a heterozygous site from a homozygous one. So AADR does not guess: at each position it draws one read at random and records that allele. The file stores the result as a homozygous call, which is why a pseudo-haploid sample never shows a heterozygous genotype.

Diploid. A genome sequenced deeply enough has many reads at each position, so both chromosomes can be read and real genotypes called, heterozygotes included.

Capture: .AG and .TW#

Ancient bone usually yields far more microbial DNA from the soil and the grave than human DNA. Capture is the answer: short synthetic baits, each matching a known variable position in the human genome, are mixed with the sequencing library, the fragments that stick are kept and the rest are washed away. Sequencing then lands almost entirely on the targeted positions, which is what made it affordable to study hundreds of people instead of a handful.

.AG, Agilent capture. The original reagent, the "1240k" baits made on Agilent's platform and introduced in 2015, targets 1,233,013 SNPs. It produced most of the ancient genomes published since, and in v66.p1 it is the largest data type by far. Releases before 2024 did not mark it at all, which is the first meaning of "no suffix" below.

.TW, Twist capture. In 2021 Twist Bioscience released a commercial assay, Twist Ancient DNA, that targets 1,352,535 SNPs: the 1240k positions plus extra content. Rohland and colleagues compared it with the 1240k reagent on 27 ancient libraries in 2022 and found it gave the highest coverage, the most even coverage across targets, and almost no bias toward enriching one allele over the other relative to shotgun sequencing. The AADR documentation states the practical consequence: Twist capture "reduces bias when co-analyzed with shotgun data". The 1240k reagent, by contrast, favours one allele over the other at a share of its positions, which is the root of most suffix-related caution.

Combined suffixes. Some people were captured both ways, and their data were pooled into one genome, still one read per position. I0231.AG.TW is a Yamnaya man from Samara whose Agilent and Twist data sit together under one ID.

Rarer capture tags. About 140 individuals carry tags from specialised bait sets: .BY ("Big Yoruba"), .AA ("Archaic Admixture"), .EC (exome), .WGC (whole-genome capture) and .IM (immune capture). Each comes from its own reagent, so treat it as its own data class rather than as .AG.

Shotgun: .SG and .DG#

Shotgun sequencing skips the fishing. The whole library is sequenced, so reads fall anywhere in the genome, human or not, and the human ones cover the panel's positions in proportion to how much was sequenced. No baits means no bait bias, which is why shotgun data are the yardstick capture reagents are measured against.

.SG, shotgun, pseudo-haploid. AADR calls these the same way as capture data, one random read per position. 6,737 individuals in v66.p1 are .SG.

.DG, shotgun, diploid. Sequenced deeply enough to call both alleles. These are the field's reference individuals: 3,726 of the 3,800 .DG genomes in v66.p1 are present-day people from high-coverage projects (French.DG, Han.DG, Mbuti.DG), and 74 are exceptionally preserved ancients such as Loschbour.DG, Ust_Ishim.DG and Stuttgart.DG.

Does diploid against pseudo-haploid bias a model? Not by itself. A pseudo-haploid genome carries one observed allele per site instead of two, so it holds less information per person, but the f4-statistics qpAdm is built on are not skewed by it. ADMIXTOOLS 2, the standard implementation, detects pseudo-haploid samples automatically and counts one allele for each, and Harney and colleagues' simulations found qpAdm accurate "when diploid calls cannot be made". The bias worth worrying about sits in the chemistry, not the ploidy.

.HO: the array suffix#

.HO marks a present-day person genotyped on the Affymetrix Human Origins array: no sequencing and no reads, just direct diploid calls at fixed positions chosen, with a clearly documented ascertainment, for population genetics (584,131 of them in recent releases, after poorly performing sites were dropped). Its people live in AADR's HO dataset, which holds 8,475 present-day individuals against 3,971 in the 1240k and 2M datasets; the extra 4,500 exist only there.

Ancestrify's reference panel is built on the 2M dataset, so you will not meet .HO labels in our Workbench or in a report's model record: the present-day references there are mostly .DG genomes.

No suffix: three different meanings#

In older AADR releases, no suffix meant capture. Up to v54.1 (November 2022, patched in March 2023), shotgun data were .SG, diploid data .DG, and capture data had no suffix at all: 7,279 of v54.1's roughly 16,500 individuals were unsuffixed, nearly all of them 1240k capture. The .AG and .TW suffixes arrived with v62 in September 2024. One Yamnaya man from Samara shows the change: in v54.1 he is I0231, and in v66 he is I0231.AG, I0231.TW, I0231.AG.TW and I0231.SG. Papers, tutorials and forum posts written for the older releases follow the old convention, so an unsuffixed ancient there is almost always what is now .AG.

In the current AADR release, the suffix left the label. AADR v66.p1 (June 2026) keeps the suffix on every Genetic ID but drops it from the population label. So Serbia_IronGates_Mesolithic holds capture, shotgun and diploid rows together: 73 rows for 38 people. Read that release by ID, never by label alone, or the same person enters a model two or three times.

On Ancestrify, no suffix marks our own additions. Our reference panel keeps the suffixed labels of AADR v66.0, so each data type stays a separate population, and adds 1,079 people from six studies published after the AADR cut-off (the "Ancestrify supplement" a report's header names): Slavic-period central and eastern Europe (Gretzinger et al. 2025), the South Caucasus (Skourtanioti, Jia et al. 2025), the Sarmatians of the Carpathian Basin (Schütz et al. 2025), Late Bronze Age central Europe (Orfanou et al. 2026), the Wielbark cemetery at Masłomęcz (Golubinski et al. 2026) and ancient Epirus and Corinthia (the APOIKIA project, Genome Biology 2026). Each of those people appears once, sequenced one way, so their labels keep the study's own group name with no suffix (Greece_Epirus_Amvrakia_Classical, for example) and their sample IDs end in .ODN, our marker. Eight of AADR's own labels also lack a final suffix (AncientGreek, LatinEtruscan and Thracian among them, 22 people in all); their sample IDs carry one, so the ID settles it.

Your own sample is the other unsuffixed entry in any model you run. It is the target, and a consumer DNA file is a different kind of data again: diploid calls from a genotyping chip, compared with the panel only at the positions both share (the merge guide covers how it is prepared).

What the suffix changes in a model#

Each production route leaves a faint fingerprint: the capture reagent's allele preference, the damage profile, diploid against pseudo-haploid encoding. qpAdm copes well with artefacts that every sample shares, and Harney and colleagues found it accurate through low coverage, heavy missing data and ancient DNA damage. It goes wrong when an artefact sits on one side of a comparison only, because the model then reads a laboratory difference as an ancestry difference. Fournier, Fulton and Reich measured the size of the effect in 2026: restricting an analysis to SNPs where platforms agree cut the spurious Z-scores by nearly an order of magnitude. Five rules follow.

• One person, one row. The same skeleton often appears two or three times under different suffixes (Loschbour.AG and Loschbour.DG; I0231.AG and I0231.SG). Use one version: never two in one population, and never one as a source with the other among the outgroups. AADR's own advice, when it does not matter which: keep the version with the most SNPs.
• One label per group. When a group comes in several data types, as Serbia_IronGates_Mesolithic does in our panel (.AG with 33 people, .SG with 26, .DG with 9), choose one, normally the largest, rather than combining them: the same people sit in two or three of them.
• One chemistry per contrast. Sources the model weighs against each other, and the outgroups meant to tell them apart, should share a data class where the panel allows. .TW and .SG behave alike; .AG is the one to watch. The AADR curators recommend the full 2M SNP set when the data share a chemistry (Twist and shotgun without 1240k, or 1240k alone) and their 1,243,531-site compatibility set when 1240k and non-1240k ancients are mixed.
• Ancients with ancients, moderns on purpose. Harney and colleagues caution against co-analysing ancient and present-day data. When a model needs present-day references, use them deliberately, typically among the outgroups, rather than as sources beside ancient genomes.
• Respect the flags. _lc (low coverage), _contam, _o or -o (outlier), kinship tags such as _brother or _1d, and Ignore_ are the curators saying a sample strained their checks; the .anno file's assessment column (Pass, Questionable, CRITICAL) says why. A model gains nothing by putting back what the curators set aside.

Labels are hypotheses too#

One level up from the suffix, the population label itself is a curator's grouping, and group labels can pool centuries or split arbitrarily. The .anno file is the referee again: check that the individuals under a label share the period and place your model assumes, and prefer site-and-period labels over umbrella groupings for source roles. Our own catalog's rule, every source population one real archaeological context and never an umbrella, exists because the difference is visible in model stability.

Where you meet suffixes on Ancestrify#

Every source in an Ancient Ancestry report's model record is printed twice: our catalog name and the exact panel label, suffix and all, so anyone with the same panel can reproduce the model (the model record guide walks through one). In the Ancestry Workbench you pick populations by the same labels, and the rules above are the ones to apply when you build a model of your own.

None of this is exotic once seen. The suffix is the dataset being honest about how each genome was made, and every rule here is one principle worn five ways: never let a laboratory difference sit where your model expects an ancestry difference.

Terms used here are defined in the glossary.

References#

• Mallick, S. & Reich, D. (2026). The Allen Ancient DNA Resource (AADR), release v66.p1, with its README and annotation files. Harvard Dataverse, doi:10.7910/DVN/FFIDCW. (Suffix definitions, counts, the Loschbour example, the compatibility SNP set.)
• Mallick, S. et al. (2024). The Allen Ancient DNA Resource (AADR): a curated compendium of ancient human genomes. Scientific Data, 11, 182.
• Fu, Q. et al. (2015). An early modern human from Romania with a recent Neanderthal ancestor. Nature, 524, 216-219. (The original Agilent in-solution capture reagent.)
• Rohland, N., Mallick, S., Mah, M., Maier, R., Patterson, N. & Reich, D. (2022). Three assays for in-solution enrichment of ancient human DNA at more than a million SNPs. Genome Research, 32, 2068-2078.
• Fournier, R., Fulton, A. P. & Reich, D. (2026). A SNP panel for co-analysis of capture and shotgun ancient DNA data. Genome Research, 36(9), 1818-1824.
• Harney, É., Patterson, N., Reich, D. & Wakeley, J. (2021). Assessing the performance of qpAdm: a statistical tool for studying population admixture. Genetics, 217(4), iyaa045.
• Patterson, N. et al. (2012). Ancient admixture in human history. Genetics, 192(3), 1065-1093. (The Human Origins array.)
• Lazaridis, I. et al. (2014). Ancient human genomes suggest three ancestral populations for present-day Europeans. Nature, 513, 409-413.