Gene set enrichment analysis explained: ORA and GSEA
Enrichment analysis asks whether the genes that changed in your experiment share a biological theme. Two methods dominate, and they answer slightly different questions.
The idea
A list of 400 changed genes is hard to interpret one gene at a time. Enrichment analysis compares the list with gene sets, which are groups of genes known to work together, such as a signalling pathway or a Gene Ontology term. If far more of your genes fall in a set than chance would predict, that set is enriched.
Over-representation analysis (ORA)
ORA starts from a list of significant genes that you define with cutoffs. For each gene set it counts the overlap with your list and asks how likely an overlap that large would be by chance, using a hypergeometric test (equivalent to a one-sided Fisher's exact test).
- Strengths: simple, fast and easy to explain. The result names the specific genes behind each enriched set.
- Weaknesses: the answer depends on your cutoffs, and it ignores how strongly each gene changed.
g:Profiler, Enrichr, DAVID and ShinyGO run ORA, and so does ExpressionBench DE Results.
Gene set enrichment analysis (GSEA)
GSEA uses no cutoff. It ranks every gene, typically by fold change or test statistic, then asks whether the members of a gene set cluster near the top or bottom of the ranking.
- Strengths: it can detect a pathway in which many genes shift modestly in the same direction, none of them individually significant.
- Weaknesses: results depend on the ranking metric, and it is slower and harder to explain.
Which to use
Use ORA when you have a clear list of changed genes and want to know what they have in common. Use GSEA when few genes pass your cutoffs, or when you expect subtle, coordinated shifts. Many papers report both.
Reading the results
- Use the adjusted p-value (FDR). Thousands of gene sets are tested, so the same multiple-testing problem applies as with genes.
- Look at the overlap size. A set enriched on the strength of three genes is fragile.
- Expect redundancy. Related terms share genes, so one biological signal often appears as a dozen similar rows.
- Enrichment is a lead, not proof. It tells you where to look next.
Why tools disagree
Run the same list through two tools and the results often differ. The usual causes are different database versions, different background gene lists, and different rules for which set sizes to test. Record all three when you report results. The guide to background genes covers the one that matters most.
Try it on your own results. The DE Results tool on ExpressionBench turns one results table into a volcano plot, a gene table and pathway enrichment. It is free, and your data stays in your browser.
Open the tool