Global25 admixture calculator

A Global25 coordinate is 25 numbers that locate one genome in a genetic reference space built from ancient and modern samples. This tool treats your row as a point in that space and searches for the weighted mixture of reference populations whose combined point sits closest to it. The search is a Monte-Carlo fit in the nMonte tradition: it distributes weight across the sources in the calculator you pick, keeps the combinations that reduce the gap, and reports the leftover gap as a fit distance. A smaller fit distance means the mixture lands nearer your coordinate, nothing more than that.

You choose the source panel, and that choice does more to shape the answer than the arithmetic does. A curated calculator is a deliberate, era-scoped set of sources; you can also paste your own. The tool ranks your closest single populations by plain Euclidean distance across all 25 dimensions, so you can see which sources are near you before you ask for a mixture of them. Everything runs in your browser: coordinates you paste are not uploaded, and nothing is stored on our servers.

What it cannot do: a coordinate fit has no p-value and cannot reject a model. It will always return percentages, including from a source panel that contains nobody your ancestors ever met. The fit distance is the only signal that the answer is poor, and a plausible-looking breakdown from a badly chosen panel is the most common way to read a real number wrongly. It also cannot distinguish two source populations that sit close together in the reference space; their weights trade off almost freely. For a method that can test a model and refuse it, see Ancient Ancestry, our qpAdm analysis, which works from allele-frequency statistics rather than coordinate distances.

Suppose an era calculator returns: source A 52.4%, source B 31.2%, source C 12.0%, source D 4.4%, fit distance 0.0182. The first read is the fit itself. On a curated, era-scoped panel of three to eight sources, a scaled fit at or under about 0.020 is excellent territory, 0.020 to 0.035 is the ordinary range for real individuals, and past roughly 0.045 either a relevant source is missing from the panel or the coordinate is not well described by that era's populations at all. 0.0182 says the geometry closed well, and says nothing more than that.

Then the percentages, in order of trustworthiness. The two large components are the model's actual statement, and their ratio is more stable than either number alone. The 12.0% is meaningful if dropping that source visibly worsens the fit; the 4.4% is close to the floor where quantisation and noise live. This tool hands out weight in steps of 0.2% and reports every source the search kept, small ones included; the result card only folds shares of 3% or less into Other so the large components read first, and you can show any of them on its own. What no percentage carries is an error bar, because a coordinate fit has none to give, so treat differences of a point or two as texture.

The one comparison that is never valid: a fit distance from this panel against a fit distance from a different-sized panel. Adding sources always improves the number, whether or not the added sources belong in your history. In our own calibration a clean 5-source model at 0.0110 'improved' to 0.0068 as it was buried under 27 sources and stopped meaning anything. When two panels disagree, the smaller defensible panel is usually the one telling the truth.

Vahaduo's Admixture JS runs the same family of arithmetic as this tool, with the opposite philosophy about sources: it is bring-your-own-spreadsheet, which is maximum freedom and zero curation. That freedom is genuinely useful for testing your own hypotheses; it is also how most bad models get built, because nothing warns you when a pasted panel mixes eras, duplicates a signal across near-identical sources, or omits the ancestry the target actually needs. The curated calculators here are the opposite trade: panels assembled and stress-tested era by era, at the cost of freedom a power user may miss. Many people sensibly use both.

nMonte, the R script this whole tradition descends from, adds batch runs and dials (penalty terms above all) at the cost of running R yourself and managing sheet discipline by hand. Results agree with this tool's for the same inputs, because the underlying mathematics is the same; the differences are convenience and reproducibility, not correctness.

GEDmatch's classic calculators answer from a different era of method entirely: fixed allele-frequency components trained around 2012, before nearly all of today's ancient genomes existed. They remain readable instruments for within-calculator comparisons, but their components blend ancestries the current reference record separates, and they cannot be pointed at a dated era at all. What none of these alternatives changes: every tool in this family is an optimiser that cannot say no. When a component needs to be tested rather than fitted, that is qpAdm's job, not any calculator's.