Thirty-seven thousand cells,
and four samples
In the treated tumour the lymphocyte fraction jumped from 2% to 21%. The number is large, the direction is consistent, and it happens in only one of the four cells of the grid. How far should you trust that?
This course answers the question with the GEO raw data (GSE268555) — that compositions are tied to one another, that testing on cells drives p down of its own accord, and that with four samples the smallest p you can reach is 0.25.
How you'll learn
QC, clustering and naming are what /dataset teaches. Here the map is finished (19 cell types) and 20 lessons go on a much harder question — are these four conditions really different?
When neutrophils drop from 49% to 18%, everything else rises without moving at all. You strip that illusion away with the CLR and see whether a difference survives it.
Hold the proportion fixed, cut only the cell count, and p travels from 0 to 0.9. And an honest permutation that shuffles samples bottoms out at p = 0.25 — a wall in the design, not in the arithmetic.
The cytotoxicity score goes up 7.4-fold overall and down if you look at T cells alone. You take the illusion composition created apart, in numbers.