Recommendations for performing GSEA on topics?

[AmosFong1]

I was wondering if one of the maintainers could provide some recommendations on how to perform GSEA on fastTopics derived topics. Would it be sensible to rank genes by -log10(lfsr) * LFC (assuming lfsr is analogous to q-value)?

[pcarbo]

@AmosFong1 Any of the approaches that have been used for differential expression analyses could be used here as well; I do not have enough experience to say what is the best approach (and it may be different for different studies). In the Genome Biology paper, we performed a GSEA using the posterior mean LFC estimates. In that case, it would be testing for enrichment for the size of the changes. Alternatively, you could use the lfsrs, or another common approach is to set some threshold (e.g., based on lfsr or LFC), and test for enrichment of the gene sets above that threshold. My intuition is that the approach you are suggesting is similar to using the posterior mean LFC estimates (assuming you are shrinking the LFC estimates with shrink.method = "ash" in de_analysis()).

[AmosFong1]

Thanks @pcarbo, could you comment on why there are NA values for some lfsr? If I wanted to rank genes for GSEA based on lfsr, how should I handle NAs?

[pcarbo]

@AmosFong1 This might happen if there is little to no variance in the LFC estimates; check the output de_analysis outputs such as "lower" and "upper". Increasing control$ns can help give you more accurate estimates and sometimes eliminate some of the NAs.