Public RNA-seq and microarray data: reanalysis and GSEA done properly

Tens of thousands of gene expression datasets sit in public archives, free to reuse. Most reanalyses of them are wrong in one of a handful of predictable ways. This guide is about the handful.

Who it is for

You have a biological question and suspect that somebody, somewhere, has already generated the data that would answer it. You can run an analysis, or have one run for you, but you have been burned before by a result that looked exciting and turned out to be a batch effect, a mislabeled sample, or a pathway that appears in every list no matter what you feed it.

Chapters

What is in it

Finding and choosing a dataset. Where to look, how to read a study design from the metadata, and when a dataset cannot answer your question no matter how it is analyzed.
Checking the samples before trusting the labels. Sex, tissue and treatment markers that let the data prove what the metadata claims. Mislabeled samples are common, and this step catches most of them.
The difference between an array and a sequencing experiment, and why the same pipeline should not be used for both.
Differential expression that holds up. Batch effects, outliers, and the questions to ask before believing a fold change.
Pathway analysis that means something. Why enrichment of a gene list is not the same as a pathway moving, what an enrichment score does and does not measure, and how to choose gene sets that fit your biology instead of the defaults.
The positive control rule. Building a control comparison into every analysis, so that a null result carries information.
Comparing across datasets, and a worked example that replicates a published finding and then shows what it takes to break it.

Why I wrote it

My lab’s early work profiled how host cells respond to different parasite strains, and much of what we found came from careful reanalysis of expression data. The mistakes in this guide are ones I have made, caught in reviews, or seen published.