Can you find a target
without running one experiment?
NCBI's GEO holds 8.7 million samples of gene expression data, uploaded by labs all over the world — because journals will not publish without it.
Over these 20 lessons we actually download four public series on IPF, a disease that turns lungs to scar tissue. Different machines, different years, different hospitals — so all four use different scales. We put them on one ruler, combine them, and rank the genes by hand.
And we find where the method starts lying — one study's 'healthy' lungs came from cancer patients. Python runs right inside this page.
How you'll learn
Build the URL from a GSE number, take the series matrix apart, and work out who the patients are — four series wrote it four different ways.
One series runs to 96,841; another starts at −8.3. Before you compare anything, you fix the ruler.
Hedges' g, inverse-variance weights, I², τ², Benjamini–Hochberg — each written out in twenty lines of Python.
Batch effects, publication bias, duplicate samples — and the study whose 'normal' lungs were surgical margins from cancer resections.