Comparing gene set enrichment across 3+ conditions?

Question

[penny.lane]

I typically use GSEA to look for enriched pathways between the conditions of an experiment. While it's clear how to employ it to compare a pair of conditions, I've yet to find an entirely satisfying way to compare holistically across all N conditions of an experiment and I'm wondering what solutions others have thought up. Basically what I'd like to do is identify all M gene sets that change significantly in any condition of the experiment and then visualize their enrichment across all conditions of the experiment. For example, in a drug treatment time course, is the up-regulation of a pathway at 6 hr sustained at 24 hr? Or perhaps is it further up-regulated? Etc.

What I've done in the past is choose one condition as the reference condition (e.g., 0 hr post-treatment in a time course experiment) and compare each of the other conditions (e.g., 6 hr and 24 hr post-treatment) to that reference with DESeq, feeding the results to GSEA and yielding N - 1 sets of GSEA stats (ES, NES, pval, and padj). Typically I will then heatmap the NES values for the N -1 condition comparisons for the subset of gene sets that achieved statistical significance (say padj < 0.05) in at least one comparison. This approach seems to work OK, but is deficient in a few ways: (1) changes between non-reference conditions are not comprehensively captured, (2) unclear interpretation of the NES differences between non-reference conditions, and (3) the NES generally has a gap between roughly -1 and 1, so it can swing a bit too wildly between positive and negative values, particularly when the enrichment is not significant in some of the conditions, which makes for strange visualization.

Any suggestions for a better approach to this problem using GSEA? Are there tools other than GSEA I should consider?

Answer 1

[Elucidata]

GSEA as an algorithm is dependent on comparing a control cohort with a treated cohort to calculate gene set enrichment, thus essentially one needs to define these two conditions explicitly and at a time comparison can be made between two conditions only as it is calculating enrichment based on the control condition. GSVA would be a good package to use if you know about the pathways you want to check. It calculates a variance score for each of the gene set in different samples and thus it is independent of a control-based comparison. An enrichment score is assigned to each sample in GSVA which can be plotted as a heatmap or using other visualization methods to show increased or decreased enrichment in all three (or even more cohorts/conditions).