Gene Set Testing

Gene-set testing

Three tests are offered. They answer different questions, so they can disagree without contradiction.

Test	Hypothesis type	Question	Uses
ORA (over-representation)	Competitive	Are the significant genes over-represented in the set, relative to all kept genes?	The current DE hit list (up, down or both)
GSEA (preranked)	Competitive	Are the set's genes concentrated at one end of the genome-wide ranking?	All genes ranked by moderated t
FRY	Self-contained	Is the set, as a whole, differentially expressed between the groups?	Per-sample expression of the set's genes

A competitive test compares the set with the other genes; a self-contained test compares the set with no change. A self-contained test can be significant for a set that is no more affected than the average gene, which is often the case when a large fraction of the transcriptome responds.

ORA

One-sided hypergeometric test of the overlap between the hit list and each set. The universe is all genes kept after filtering. FDR by Benjamini–Hochberg across the tested sets.

GSEA

Weighted Kolmogorov–Smirnov running enrichment score with weight p = 1 (hits weighted by |t|), computed on the moderated t ranking. Significance by gene-set permutation with a seeded generator (500 permutations for up to 200 sets, 250 for up to 1,000 sets, 100 above that). The normalized enrichment score (NES) divides by the mean of same-sign null scores. The smallest attainable p-value is 1/(permutations + 1), which limits FDR for large libraries. The leading edge is the set of member genes up to the running-score extremum.

FRY

The fast rotation test of limma (fry, default settings), a closed-form approximation to ROAST. It reports a directional p-value (coordinated shift up or down) and a mixed p-value (any perturbation, including mixed directions). Results match limma to machine precision.

Set size limits

Only sets with between 5 and 2,000 members present in the filtered data are tested by default. Change the limits with set size.

Supporting displays

Display	Content
Ridgeline	log₂FC distribution within each top set
Barcode plot	Positions of set members in the ranking, with the enrichment worm (limma tricubeMovingAverage)
Set-vs-rest ECDF	Cumulative distributions of the statistic for members and non-members, with the Kolmogorov–Smirnov distance
Jaccard heatmap	Pairwise overlap of top sets, to identify redundant results
Leading-edge matrix	Which genes drive which GSEA sets

Enrichment across contrasts

With three or more groups, Enrichment across all contrasts runs FRY independently for every pairwise contrast, with BH correction within each contrast. The heatmap shows signed −log₁₀ FDR for the 25 sets with the strongest evidence; the CSV export contains all sets.

Cluster enrichment

On the Heatmap tab, when rows are ordered by k-means, each cluster is tested by ORA against the loaded library. Because the clusters were selected from the data, these results are descriptive.

References

Subramanian A et al. (2005) PNAS 102:15545. Wu D et al. (2010) Bioinformatics 26:2176. Wu D and Smyth GK (2012) Nucleic Acids Res 40:e133.