AdmixTools 2 Lab

ADMIXTOOLS 2 is the R package formal admixture modelling is done with in published ancient-DNA research. Normally using it means installing R, compiling the package, obtaining a genotype panel and learning its interface. This page runs the real functions on our servers against the Allen Ancient DNA Resource Human Origins panel, so you can compose a model in a form and read the output — the same `qpadm()`, `qpwave()` and f-statistic routines, not an approximation of them.

The methods available here work from allele-frequency statistics rather than coordinate distances. f2, f3, f4 and D statistics measure shared drift and treeness between population sets. qpWave asks how many independent ancestry streams are needed to relate two sets. qpAdm goes further: you give it a target, a set of candidate sources (the left set) and a set of outgroups (the right set), and it estimates each source's weight with a standard error and a z-score, plus a p-value for the model as a whole. Admixture-graph fitting searches for a tree-with-admixture topology consistent with the f-statistics.

The p-value is what makes qpAdm different in kind from a coordinate fit: it can say the proposed model is incompatible with the data, and a rejected model is the method working rather than a failure of the tool. Expect rejections, and expect the right set to matter as much as the left — outgroups are what give the test its power, and a right set too small or too closely related to your sources will accept almost anything. A weight is only meaningful if it sits several standard errors away from zero.

Limits worth knowing before you start. Runs execute on a single serialized worker, so one account may create 20 runs in a rolling 24 hours and hold 2 queued or running at once — the quota exists because each run costs real CPU that everyone else is waiting on. Analyses here operate on reference populations from the panel, not on your own genotypes; modelling your own sample as the target requires it to be merged into the panel first, which is what our paid qpAdm analysis does. A free, email-verified account is needed to run, though this page and its description are public.

A sensible first run: target a Bronze Age European population from the panel, offer Anatolian farmer and steppe pastoralist proxies as the left set, and give the right set the classic spine — a deep African outgroup, an Upper Palaeolithic Eurasian or two, East Asian, Oceanian and Native American references, plus the era-appropriate contrasts (a European forager, an early Iranian-plateau population) that let the test tell your two sources apart. The output to read, in order: the p-value for the whole model, then each source's weight beside its standard error and z-score. A weight whose two-standard-error interval touches zero is a source the data cannot certify, whatever the percentage says.

Then break it on purpose, because the failures teach the method. Drop the forager contrast from the right set and watch previously distinct models start passing together — outgroups are where the power lives. Add a third source a two-source model already explains and watch its weight hug zero. Propose a target as a mixture of two populations that postdate it and watch the rejection: f-statistics have no clock, but the model discipline does. A p-value below 0.05 here is the method working, and half the value of a free sandbox is meeting rejections where they cost nothing.

The numbers behind that discipline are published and worth knowing: screening many models and keeping the best p-value picks the true model barely better than a coin flip, and adding outgroups without limit eventually rejects everything, true models included. Compose few models deliberately rather than sweeping many — the sweep is exactly the workflow whose measured false-discovery rates run worst.

The reference alternative is ADMIXTOOLS 2 in R itself — free software, full control, every dial exposed — at the cost of installation, panel assembly and format wrangling before the first statistic runs. Our R tutorial walks that road for readers who want it; this page is the same functions with the setup already done, which is the entire difference.

Illustrative DNA's AdmixLab sells a subscription environment for the same family of runs, with a guided wizard and generous daily quotas — a genuinely capable product whose facts sit on our comparison page. The structural differences: this sandbox is free with an account and runs on the public panel only, and our paid product is the opposite division of labour — not a bigger sandbox but a finished analysis, where a person composes and stress-tests the models against your own merged genotypes and publishes one that survives, with the complete record attached. Which arrangement is better depends only on whether you want the modelling as a hobby or the model as an answer.

The limit every environment shares: qpAdm on reference populations tests relationships between the panel's groups. Statements about YOUR ancestry need your genotypes merged into the panel — the step no browser tool performs — and that merge plus the hand-built search is precisely what the paid analysis is.