This document provides an in-depth comparison of Gene Set Enrichment Analysis (GSEA) and Over-Representation Analysis (ORA) to help you understand when to use each method. --- ## GSEA (Gene Set Enrichment Analysis) ### Method Overview **What it does:** Tests whether genes in a pathway tend to be at the TOP or BOTTOM of your ranked gene list (by fold change or test statistic). **Input:** ALL genes ranked by expression change (no cutoff applied) **Question answered:** "Are genes in this pathway coordinately up- or down-regulated?" **Statistical approach:** Uses a running enrichment score that walks down the ranked gene list, increasing when a gene is in the pathway, decreasing when not. Computes significance via permutation testing. ### Strengths - No arbitrary significance cutoffs needed - Detects subtle but coordinated changes across many genes - Works well even when few genes pass strict thresholds - Provides direction (activated vs suppressed pathways) - More sensitive to modest changes across entire pathways - Uses all available information from the experiment ### Limitations - Requires ranking metric (fold changes or test statistics) - Computationally more expensive (permutation testing) - Results can be harder to interpret for non-bioinformaticians - Requires careful selection of ranking metric ### Output Interpretation - **Positive NES (Normalized Enrichment Score):** Pathway is ACTIVATED (genes tend to be upregulated) - **Negative NES:** Pathway is SUPPRESSED (genes tend to be downregulated) - **Leading edge genes:** Core genes driving the enrichment signal - **FDR < 0.05:** Statistically significant after multiple testing correction ### Best Ranking Metrics (in order of preference) 1. **Test statistic (t-statistic, Wald statistic)** - BEST, accounts for both effect size and significance 2. **Signed -log10(pvalue)** - Good alternative if test statistic unavailable 3. **Log2 fold change** - Avoid if possible, ignores statistical significance --- ## ORA (Over-Representation Analysis) ### Method Overview **What it does:** Tests whether your significant genes overlap with a pathway more than expected by chance. **Input:** A discrete list of significant genes (e.g., padj < 0.05, |log2FC| > 1) **Question answered:** "Do my significant genes contain more pathway X members than expected?" **Statistical approach:** Uses Fisher's exact test / hypergeometric test to compare observed vs. expected overlap between gene list and pathway. ### Strengths - Simple and intuitive - Fast computation - Good for validating GSEA findings - Works with any gene list (not just DE results) - Easy to explain to collaborators ### Limitations - Requires arbitrary cutoffs (which genes are "significant"?) - Loses information about genes just below threshold - Must analyze UP and DOWN genes separately to preserve direction - Less sensitive to subtle coordinated changes - Can be biased by gene list size ### Output Interpretation - **GeneRatio:** k/n (genes in pathway / total query genes) - **BgRatio:** M/N (genes in pathway in background / total background genes) - **p.adjust < 0.05:** Statistically significant enrichment - **Higher GeneRatio** = stronger enrichment - **Must run separately for up/down genes** to preserve directionality ### Critical ORA Requirements 1. **Specify background genes:** Use all tested genes, not all known genes 2. **Separate up/down genes:** Don't combine opposite directions 3. **Appropriate thresholds:** padj < 0.05, |log2FC| ≥ 1 (adjust as needed) --- ## When to Use Which Method | Scenario | Recommended Method | Rationale | | --- | --- | --- | | **Default / no preference stated** | GSEA | More sensitive, no arbitrary cutoffs | | User has full DE results with fold changes | GSEA | Can use all information | | User wants to detect subtle coordinated changes | GSEA | Designed for this | | User asks "what pathways are activated/suppressed" | GSEA | Provides directionality via NES | | User has a specific gene list (no fold changes) | ORA | GSEA requires ranking | | User wants quick validation of GSEA results | ORA | Complementary approach | | User explicitly requests ORA | ORA | Honor user preference | | Few genes pass significance threshold (<50) | GSEA | ORA underpowered with few genes | | User wants to compare to Enrichr results | ORA | Enrichr uses ORA method | | Very large gene list (>1000 significant genes) | Both | GSEA for sensitivity, ORA for validation | --- ## Decision Flowchart ``` User requests enrichment analysis │ ▼ Do they have full DE results (with fold changes)? │ ┌────┴────┐ │ │ YES NO │ │ ▼ ▼ Use GSEA Use ORA (default) (gene list only) │ ▼ Did user specifically request ORA? │ ┌───┴───┐ │ │ YES NO │ │ ▼ ▼ Run both Run GSEA only ``` --- ## Comparing Results Between Methods When you run both GSEA and ORA, you may see: ### Concordant Results (expected) - Pathways significant in both GSEA and ORA - This is the strongest evidence for pathway enrichment ### GSEA significant, ORA not significant - Pathway has modest but coordinated changes across many genes - Few genes pass strict ORA thresholds - Still biologically meaningful ### ORA significant, GSEA not significant - Strong signal in a few genes, but not coordinated across pathway - May indicate a subset of pathway is affected - Less common than scenario above ### Neither significant - Pathway not affected in your experiment - Or: wrong database, wrong species, poor gene identifier mapping --- ## Common Pitfalls ### GSEA Pitfalls 1. ❌ Using fold change alone for ranking (ignores significance) 2. ❌ Not removing duplicate genes 3. ❌ Using too few permutations (use ≥10,000) 4. ❌ Filtering genes before ranking (use ALL genes) ### ORA Pitfalls 1. ❌ Not specifying background genes (biases results) 2. ❌ Combining up and down genes (loses biological meaning) 3. ❌ Using all known genes as background (wrong denominator) 4. ❌ Filtering by keyword instead of statistics --- ## References 1. **GSEA original paper:** Subramanian A, et al. (2005) Gene set enrichment analysis: A knowledge-based approach for interpreting genome-wide expression profiles. PNAS. DOI: 10.1073/pnas.0506580102 2. **Method comparison:** Geistlinger L, et al. (2021) Toward a gold standard for benchmarking gene set enrichment analysis. Brief Bioinform. DOI: 10.1093/bib/bbz158 3. **Best practices:** Reimand J, et al. (2019) Pathway enrichment analysis and visualization of omics data using g:Profiler, GSEA, Cytoscape and EnrichmentMap. Nat Protoc. DOI: 10.1038/s41596-018-0103-9