Cross-study pathway meta-analysis

Runs pre-ranked GSEA separately in every study and combines the pathway-level evidence across studies. This answers whether a pathway is consistently perturbed across cohorts, which is more robust than enrichment of a single pooled gene list when studies differ in platform, depth or gene coverage (Shen & Tseng 2010).

Usage

mx_pathway_studies(
  de_results,
  db = c("Hallmarks", "Reactome", "KEGG", "WikiPathways", "GO_BP"),
  gene_sets = NULL,
  min_size = 15,
  max_size = 500,
  weights = "equal",
  seed = 1L,
  ...
)

Arguments

de_results: A named list of per-study DE data.frames or metaXpressStudy objects after mx_de_all.
db: Gene set collection passed to mx_gene_sets. Ignored when gene_sets is supplied.
gene_sets: Optional named list of gene sets (e.g. from mx_gene_sets or a GMT file) whose IDs match meta_table$gene_id.
min_size, max_size: Gene set size limits after restricting to genes in the analysis. Defaults: 15 and 500.
weights: Per-study weights for the Stouffer combination: "equal" (default), "sample_size" (square root of the per-study sample size from n_case / n_control), or a numeric vector with one weight per study. Equal weights are the default because GSEA p-values of a large cohort are already far more extreme, and sample-size weights would let one cohort dominate.
seed: Integer seed for the fgsea permutation sampling, for reproducibility. The global RNG state is restored afterwards.
...: Further arguments passed to fgsea::fgsea().

Value

A data.frame with one row per gene set: pathway_id, n_studies, mean_NES, n_up / n_down (studies with positive / negative NES), n_sig (studies with per-study padj <= 0.05), meta_z, meta_pvalue, meta_padj, followed by one NES_<study> column per study. Per-study fgsea tables are in attr(x, "per_study").

Details

For each study, genes are ranked by the signed z-score derived from the per-study test (the stat column from mx_de when present). Each pathway's two-sided GSEA p-value is converted to a signed z-score using the sign of its NES, and the z-scores are combined by weighted Stouffer's method, so pathways enriched in opposite directions cancel out. This is the p-value-combination variant of MAPE (Shen & Tseng 2010) with directional combination.

References

Shen, K. & Tseng, G.C. (2010) Meta-analysis for pathway enrichment analysis when combining multiple genomic studies. Bioinformatics, 26(10), 1316–1323. doi:10.1093/bioinformatics/btq148

Korotkevich, G. et al. (2016) Fast gene set enrichment analysis. bioRxiv. doi:10.1101/060012

Examples

if (requireNamespace("fgsea", quietly = TRUE)) {
  set.seed(2)
  genes <- paste0("G", 1:1000)
  sets  <- list(UP_SET = genes[1:40], NULL_SET = genes[501:560])
  de <- lapply(1:3, function(s) {
    lfc <- rnorm(1000, c(rep(0.8, 40), rep(0, 960)), 0.3)
    p <- 2 * pnorm(-abs(lfc / 0.3))
    data.frame(gene_id = genes, log2FC = lfc, pvalue = p,
               padj = p.adjust(p, "BH"))
  })
  names(de) <- c("S1", "S2", "S3")
  mx_pathway_studies(de, gene_sets = sets, min_size = 10)
}
#> Running per-study GSEA on 3 studies...
#>   pathway_id n_studies  mean_NES n_up n_down n_sig    meta_z  meta_pvalue
#> 1     UP_SET         3 3.0669911    3      0     3 17.732418 2.356787e-70
#> 2   NULL_SET         3 0.4940344    2      1     0  1.166798 2.432918e-01
#>      meta_padj   NES_S1     NES_S2   NES_S3
#> 1 4.713575e-70 3.037499  3.1581902 3.005284
#> 2 2.432918e-01 1.102117 -0.7690123 1.148999