Behind every stacked-bar ancestry figure in a population-genetics paper, and upstream of every hobbyist calculator on GEDmatch, sits one family of software: STRUCTURE (Pritchard, Stephens & Donnelly, 2000) and its fast successor ADMIXTURE (Alexander, Novembre & Lange, 2009). Knowing what the academic tool actually estimates — and what it deliberately does not — clears up most of the confusion the consumer versions inherit.
The model, in one paragraph#
ADMIXTURE assumes every genome in the dataset is a mixture of K ancestral populations, each defined by its own allele frequencies at every marker. Given genotypes for hundreds or thousands of individuals, it estimates both things at once: the K allele-frequency profiles (the components) and every individual's mixing proportions across them, by maximising the likelihood of all the data jointly. Nothing is anchored in advance; the components emerge from the sample. That joint estimation is the defining property — and the first thing lost downstream.
Unsupervised, supervised, and the hobbyist third mode#
Unsupervised is the default just described: components are inferred from the data. Change the sample — add fifty Sardinians — and every component can shift, because the components belong to the dataset, not the world.
Supervised mode fixes some individuals as reference members of designated ancestral populations and estimates only the remaining proportions. It answers a narrower question and depends entirely on the chosen references being what the labels claim.
The hobbyist calculators are a third thing: a past ADMIXTURE run's component frequencies, frozen, with single uploads projected onto them one kit at a time. Your GEDmatch result did not re-run ADMIXTURE; it fitted your file against profiles somebody computed in 2012. That is why a published K plot cannot be reproduced by uploading a kit anywhere, why components never update, and why every frozen calculator carries its training panel's biases — the calculator effect — permanently. The mechanics of that consumer layer are covered in what an admixture calculator is.
Choosing K, and what K is not#
The papers pick K with cross-validation — try a range, keep the K that predicts held-out genotypes best — and typically show several K values side by side because no K is true. Each K is a resolution, not a hypothesis: at K=3 Europe is one cline, at K=6 the familiar farmer/forager/steppe structure appears, at K=12 regional slices emerge whose boundaries follow sampling as much as history. The tutorial every reader of bar plots should know (Lawson, van Dorp & Falush, 2018) demonstrates the sharp edge: very different demographic histories — a real three-way mixture versus an unsampled ghost population versus plain bottlenecks — can produce identical ADMIXTURE plots. The bars cannot distinguish them. That is not the software failing; it is the model class being descriptive rather than testable.
Descriptive versus testable#
ADMIXTURE has no null hypothesis. It always returns proportions summing to one; there is no p-value under which the K-component story could fail; a beautiful bar plot is compatible with many histories. The field's answer is to pair it with methods that do test: f-statistics and qpAdm, which propose an explicit source-and-outgroup model and return a p-value that can reject it, with standard errors per source. The two families answer different questions on purpose — ADMIXTURE describes structure, qpAdm tests models — and the published literature uses them in exactly that order: bar plots to see, formal models to argue. The consumer translation of that pairing is a calculator for exploring and a formal model for defending, which is precisely the split between our free Lab tools and the paid qpAdm analysis.
Running it yourself#
ADMIXTURE is free academic software: a Linux/macOS command-line tool taking PLINK-format
genotypes (admixture mydata.bed 6 is a run at K=6, --cv adds cross-validation). The practical
obstacles are dataset assembly — merging your kit with reference panels position by position,
handling strand flips and missingness — and compute time at scale. If the goal is understanding
rather than pipeline-building, the same conceptual territory is browsable interactively: our
AdmixTools 2 Lab exposes the testing family (f-statistics, qpWave, qpAdm)
against a curated ancient panel, which is where the arguing happens anyway.
Terms used here are defined in the glossary.
References#
- Alexander, D. H., Novembre, J. & Lange, K. (2009). Fast model-based estimation of ancestry in unrelated individuals. Genome Research, 19(9), 1655–1664.
- Pritchard, J. K., Stephens, M. & Donnelly, P. (2000). Inference of population structure using multilocus genotype data. Genetics, 155(2), 945–959.
- Lawson, D. J., van Dorp, L. & Falush, D. (2018). A tutorial on how not to over-interpret STRUCTURE and ADMIXTURE bar plots. Nature Communications, 9, 3258.
