G25 PCA viewer

Principal component analysis takes the 25 dimensions of a Global25 coordinate and finds the directions along which the samples spread the most, so the cloud can be drawn on a flat chart. This tool works in two modes. An era view projects your pasted row onto a reference space that was computed once, server-side, over that era's published samples — your point lands among ancient and modern individuals without moving any of them. A custom view fits a fresh PCA in your browser over exactly the rows you paste, which is the right tool for comparing a handful of coordinates against each other.

The two modes answer different questions, and swapping them silently is the classic mistake. In an era view the axes belong to the reference panel, so where you land is comparable across visitors and across sessions. In a custom fit the axes are recomputed from your own input — add or remove one row and every axis can rotate. A custom chart is a statement about your input set, not about any fixed reference space.

The limit to keep in view: a PCA chart shows 2 of 25 dimensions, chosen for spread, not for meaning. Two samples can sit near each other on the chart and be far apart in the dimensions the chart dropped — plot proximity is not descent, and no distance read off a PCA replaces the 25-dimensional distance the G25 distance tool computes. Everything runs in your browser; era reference spaces are public, read-only data, and the coordinates you paste are not uploaded.

Project a row onto a West Eurasian era view and suppose the point lands between the southern European cluster and the Levantine one, nearer the former. Three readings, in decreasing order of soundness. First, the neighbourhood: the clusters flanking your point are the populations whose company your coordinate keeps on these two axes, and that much is real. Second, the cline: sitting along the axis that runs between two clusters is consistent with mixture between things like them — consistent with, not evidence of, because an unsampled third population and off-plane structure produce the same picture. Third — and this is the reading to refuse — the exact pixel: a point's precise position moves with which two of twenty-five dimensions the view spends, and treating it as a measurement is how plots get over-read.

The one-step check that keeps a plot honest: any time two points look close, measure them. The distance tool computes the full 25-dimensional separation in one paste, and the classic surprise — a projected ancient sample 'on top of' a modern population in PC1 and PC2 while sitting far away in full distance — is exactly the amputation this page's method section describes. Plot for structure, measure for closeness; the two jobs do not swap.

For custom fits over pasted rows, one more discipline: caption which rows produced the axes. Add or remove one sample and every axis can rotate, so two custom plots are never comparable, including your own from different sessions. A screenshot without its input list is unreadable by construction.

Vahaduo's Global 25 Views ships a large gallery of preset projections — regional slices of the same Eurogenes space — and its Custom PCA fits fresh axes over whatever you paste, with 2D and 3D outputs. The mechanics match this tool's two modes; the differences are curation and captioning. Here the era views are the same curated panels the paid report plots against, each view stating its era and reference set; the gallery approach leaves knowing what a given preset contains to the user. For fitted plots both tools share the identical caveat: the axes belong to the input.

The deeper alternative is not another PCA tool but a different lens entirely. A projection shows where a coordinate sits; it cannot say what the coordinate is made of, and betweenness on a chart is a hypothesis rather than a decomposition. The admixture calculator turns the hypothesis into a stated fit with a residual; qpAdm turns it into a testable model with a p-value. A plot that suggests mixture and a fit that quantifies it belong in that order — and neither certifies descent, which no arrangement of dots can.

The limit this tool keeps regardless of alternative: twenty-five dimensions flattened to two, chosen for variance, not meaning. Every PCA viewer shares it; the ones that serve you well are the ones that keep saying so.