Bio Pathway Gsea

Version Compatibility

Reference examples tested with: DESeq2 1.42+

Before using code patterns, verify installed versions match. If versions differ:

• R: packageVersion('<pkg>') then ?function_name to verify parameters

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Gene Set Enrichment Analysis (GSEA)

Core Concept

GSEA uses all genes ranked by a statistic (log2FC, signed p-value) rather than a subset of significant genes. It finds gene sets where members are enriched at the top or bottom of the ranked list.

When to Use GSEA vs ORA

In benchmarks, GSEA-family methods outperform ORA by ~35% higher F1 score on simulated data. GSEA is strictly preferred for DE-derived analyses.

Prepare Ranked Gene List

Goal: Create a sorted named vector of gene-level statistics suitable for GSEA input.

Approach: Extract fold changes (or other statistics) from DE results, name by gene ID, and sort in decreasing order.

"Run GSEA on my differential expression results" → Rank all genes by expression statistic and test whether predefined gene sets cluster toward the extremes of the ranked list.

library(clusterProfiler)
library(org.Hs.eg.db)

de_results <- read.csv('de_results.csv')

# Create named vector: values = statistic, names = gene IDs
gene_list <- de_results$log2FoldChange
names(gene_list) <- de_results$gene_id

# Sort in decreasing order (REQUIRED)
gene_list <- sort(gene_list, decreasing = TRUE)

Convert Gene IDs for GSEA

Goal: Map gene symbols to Entrez IDs while preserving the ranked statistic values.

Approach: Use bitr for ID conversion, then rebuild the named sorted vector with Entrez IDs as names.

# Convert symbols to Entrez IDs
gene_ids <- bitr(names(gene_list), fromType = 'SYMBOL', toType = 'ENTREZID', OrgDb = org.Hs.eg.db)

# Create ranked list with Entrez IDs
gene_list_entrez <- gene_list[names(gene_list) %in% gene_ids$SYMBOL]
names(gene_list_entrez) <- gene_ids$ENTREZID[match(names(gene_list_entrez), gene_ids$SYMBOL)]
gene_list_entrez <- sort(gene_list_entrez, decreasing = TRUE)

Ranking Metric Selection

Goal: Choose a ranking metric that balances magnitude and significance for GSEA.

Approach: The ranking metric choice matters enormously. Match the metric to the DE tool used.

# DESeq2 Wald statistic (default recommendation)
gene_list <- de_results$stat
names(gene_list) <- de_results$gene_id
gene_list <- sort(gene_list[!is.na(gene_list)], decreasing = TRUE)

# Signed p-value (for edgeR or when Wald stat unavailable)
# Replace p=0 with small value to avoid Inf
pvals <- pmax(de_results$pvalue, 1e-300)
gene_list <- -log10(pvals) * sign(de_results$log2FoldChange)
names(gene_list) <- de_results$gene_id
gene_list <- sort(gene_list[!is.na(gene_list)], decreasing = TRUE)

Never use: shrunken log2FC from lfcShrink(type='normal') -- the prior distorts rankings. Also: lfcShrink() with type='apeglm'/'ashr' drops the stat column, so pull stat from unshrunk results(dds) if needed.

GSEA with GO

Goal: Detect coordinated expression changes across GO gene sets without requiring a significance cutoff.

Approach: Run gseGO on a ranked gene list, testing whether GO term members are enriched at the top or bottom of the list.

gse_go <- gseGO(
    geneList = gene_list_entrez,
    OrgDb = org.Hs.eg.db,
                        # BP, MF, CC, or ALL
    minGSSize = 10,
    maxGSSize = 500,
    pvalueCutoff = 0.05,
    verbose = FALSE,
    pAdjustMethod = 'BH'
)

# Make readable
gse_go <- setReadable(gse_go, OrgDb = org.Hs.eg.db, keyType = 'ENTREZID')

GSEA with KEGG

Goal: Identify KEGG pathways with coordinated expression changes across all genes.

Approach: Run gseKEGG on the ranked gene list using KEGG pathway definitions.

gse_kegg <- gseKEGG(
    geneList = gene_list_entrez,
    organism = 'hsa',
    minGSSize = 10,
    maxGSSize = 500,
    pvalueCutoff = 0.05,
    verbose = FALSE
)

# Make readable
gse_kegg <- setReadable(gse_kegg, OrgDb = org.Hs.eg.db, keyType = 'ENTREZID')

GSEA with Custom Gene Sets

Goal: Run GSEA against user-provided or non-standard gene set collections.

Approach: Load a GMT file and use the generic GSEA function with TERM2GENE mapping.

# Read GMT file (Gene Matrix Transposed)
gene_sets <- read.gmt('msigdb_hallmarks.gmt')

gse_custom <- GSEA(
    geneList = gene_list_entrez,
    TERM2GENE = gene_sets,
    minGSSize = 10,
    maxGSSize = 500,
    pvalueCutoff = 0.05
)

MSigDB Gene Sets

Goal: Run GSEA using curated gene set collections from the Molecular Signatures Database.

Approach: Retrieve gene sets via msigdbr, format as TERM2GENE data frame, and run GSEA.

# Use msigdbr package for MSigDB gene sets
library(msigdbr)

# Hallmark gene sets
hallmarks <- msigdbr(species = 'Homo sapiens', category = 'H')
hallmarks_t2g <- hallmarks[, c('gs_name', 'entrez_gene')]

gse_hallmark <- GSEA(
    geneList = gene_list_entrez,
    TERM2GENE = hallmarks_t2g,
    pvalueCutoff = 0.05
)

# Other categories: C1 (positional), C2 (curated), C3 (motif), C5 (GO), C6 (oncogenic), C7 (immunologic)

Understanding Results

# View results
head(gse_go)
results <- as.data.frame(gse_go)

# Key columns:
# - NES: Normalized Enrichment Score (positive = upregulated, negative = downregulated)
# - pvalue: Nominal p-value
# - p.adjust: FDR-adjusted p-value
# - core_enrichment: Leading edge genes

Interpreting NES (Normalized Enrichment Score)

Correct interpretation order:

1. Check FDR first. Use FDR < 0.25 (Broad Institute recommendation) or FDR < 0.05 (common in publications). High |NES| with non-significant FDR is meaningless.
2. Use NES for prioritization among significant results.
3. Examine the leading edge genes to understand what drives the signal.

NES caveats: Very large gene sets (> 500 genes) can achieve high |NES| even randomly. Very small sets (< 10 genes) can be driven by a single outlier. Always cross-check with minGSSize/maxGSSize filtering.

Leading Edge Interpretation

The core_enrichment column contains the "leading edge" genes -- those driving the enrichment signal. These appear before the enrichment peak in the ranked list.

• High leading edge count, concentrated at the extreme of the ranked list: Strong, trustworthy enrichment. The pathway's genes are coordinated at one end.
• Low leading edge count: Enrichment may be driven by 1-2 extreme outlier genes, not coordinated pathway regulation. Inspect the individual genes.
• The leading edge genes are the most biologically actionable output of GSEA -- use them for downstream analysis (pathway visualization, network analysis).

Key Parameters

Export Results

Goal: Save GSEA results and extract leading edge genes for downstream analysis.

Approach: Convert enrichment object to data frame, export to CSV, and parse core_enrichment for driving genes.

results_df <- as.data.frame(gse_go)
write.csv(results_df, 'gsea_go_results.csv', row.names = FALSE)

# Get leading edge genes for a term
leading_edge <- strsplit(results_df$core_enrichment[1], '/')[[1]]

Duplicate Gene Handling

Duplicate gene IDs in the ranked list will bias enrichment scores. After ID conversion, some genes may map to multiple IDs. Always deduplicate:

# Remove duplicates -- keep the entry with the largest absolute value
gene_list <- gene_list[!duplicated(names(gene_list))]

# Or more carefully, keep the most extreme signal per gene:
gene_df <- data.frame(id = names(gene_list), val = gene_list)
gene_df <- gene_df[order(-abs(gene_df$val)), ]
gene_df <- gene_df[!duplicated(gene_df$id), ]
gene_list <- setNames(gene_df$val, gene_df$id)
gene_list <- sort(gene_list, decreasing = TRUE)

Notes

• Must be sorted - gene list must be sorted in decreasing order
• Named vector - names are gene IDs, values are statistics
• No arbitrary cutoffs - uses all genes, not just significant ones
• NES sign matters - positive = upregulated enrichment
• Leading edge - core_enrichment contains driving genes
• FDR threshold - Broad Institute recommends FDR < 0.25 for GSEA (more lenient than ORA's 0.05) because GSEA is a competitive test with less power
• No duplicates - deduplicate the ranked list after ID conversion

Related Skills

• go-enrichment - Over-representation analysis for GO
• kegg-pathways - Over-representation analysis for KEGG
• enrichment-visualization - GSEA plots, ridge plots