bio-pathway-gsea

Build a named vector from the DESeq2 stat column, remove NAs and duplicate gene names, and sort it strictly decreasing. Then call clusterProfiler's gseGO or gseKEGG with set.seed fixed and eps=0 for exact small p-values.

Use a signed, variance-calibrated metric: the DESeq2 Wald stat or limma moderated t. For edgeR use sign(logFC) times -log10(p) clamped with pmax(p, 1e-30). Never rank by raw p-value, which erases direction, or bare log2FC, which lets low-count outliers dominate.

Use GSEA when you have a ranked statistic for all genes and no natural cutoff; it detects coordinated subtle shifts. Use over-representation analysis (enrichGO/enrichKEGG) only when you have a pre-selected unranked gene list such as GWAS hits or a screen module.

Preranked GSEA in clusterProfiler and fgsea uses gene permutation, which assumes gene independence and is anti-conservative for co-regulated sets. Report the permutation type, and when a design matrix is available use limma's camera with inter.gene.cor=NA for a correlation-honest test.

The fgseaMultilevel engine uses stochastic Monte Carlo sampling, so p-values drift without a fixed seed. Call set.seed before every run and pass seed=TRUE; also note gseKEGG queries the live KEGG API, so results change across KEGG releases.

No. nPerm is silently accepted with only warnings and downgrades the engine from fgseaMultilevel to the coarser fgseaSimple. Remove nPerm entirely and control tiny p-value resolution with the eps parameter instead.