Interpreting Enrichment Results

Contents

• ORA vs GSEA: the statistics
• The background universe (ORA)
• Multiple-testing correction
• Reading GSEA output
• Reducing redundant terms
• Significance vs relevance
• Reproducibility checklist
• Publication table template
• Common misinterpretations

ORA vs GSEA: the statistics

ORA asks: among my k hits (out of a background of N genes), are more in gene set S (size K) than expected by chance? This is a hypergeometric / Fisher's exact test. It depends entirely on the threshold used to define hits and on the background N. Good when there is a clear, strong hit list.

GSEA asks: walking down the fully ranked list of all tested genes, is gene set S concentrated near the top (or bottom)? It uses a weighted Kolmogorov– Smirnov-like running sum; significance comes from permutations. No arbitrary threshold; sensitive to coordinated, modest shifts across many genes. Better when effects are broad/subtle or when a hit list would be very short or very long.

Rule of thumb: a discrete hit list → ORA; a ranked table with per-gene scores → GSEA. They answer different questions and can legitimately disagree.

The background universe (ORA)

The background (the "domain" / universe) is the set of genes that could have appeared as a hit. For RNA-seq that is the set of expressed/tested genes, not all ~20,000 protein-coding genes. Using too large a background makes ordinary housekeeping categories look significant — the most common way ORA results mislead.

• GSEApy 1.3.1 routes explicit online background lists through Speedrichr; standard Enrichr uses its default background. g:Profiler accepts domain_scope="custom" and background; local **gp.enrich()** with a pinned GMT and explicit gene list makes the universe directly inspectable.
• The background must contain the query, use the same ID mapping as the query and library, and represent the actual selection opportunity (tested genes, not a separately chosen set that makes results smaller). Record N, k, K and overlap after mapping/intersection. Length, abundance and detection-dependent selection can bias ORA even with the right universe; consider a bias-aware null.

Multiple-testing correction

• Benjamini–Hochberg (FDR) — used for Enrichr/gseapy ORA Adjusted P-value. Classic GSEA FDR q-val instead uses normalized enrichment-score permutation distributions. GSEApy 1.3.1 multilevel GSEA uses BH; state the backend.
• g:SCS — g:Profiler's default; accounts for the correlated structure of GO and overlapping terms. It is a different error-control procedure, not an FDR estimate; avoid a universal claim that it is superior to BH.
• Bonferroni controls family-wise error and does not require independent tests; it can be conservative with dependent/overlapping terms.

Define the testing family before looking at results. Classic GSEApy 1.3.1 scopes FDR by library prefix (library__term), while multilevel BH covers all submitted terms. Local GSEApy ORA applies BH to positive-overlap terms only. g:Profiler returns adjusted p_value and source-specific domain metadata. None of these automatically corrects across every cluster, contrast, library and threshold you tried. Report those families and avoid selecting the most favorable run.

Reading GSEA output

• NES (normalized enrichment score) — the headline metric; normalized for set size and the chosen null. Compare cautiously within the same method/run; it is not a fold-change or an interchangeable activity scale across cohorts. Sign = direction (positive = enriched at the top of your ranking, e.g., up in the test condition).
• FDR q-val — significance; filter on this. GSEA recommends < 0.05 for gene-set permutations (including preranked analyses); the exploratory < 0.25 convention applies to phenotype permutations. State the chosen threshold.
• Leading-edge genes (Lead_genes) — the subset of genes that drive the signal: hits at/before the positive peak or at/after the negative trough. Report these; they are the concrete biology and are useful for overlap/redundancy analysis.

Reducing redundant terms

GO and large pathway sets return many overlapping terms describing the same biology. Don't list 40 near-duplicates. Options:

• Enrichment map — graph with terms as nodes and edges weighted by gene overlap (Jaccard/overlap coefficient); cluster it and label clusters. gseapy: gp.enrichment_map(...); render with networkx (see the networkx skill).
• Leading-edge / gene overlap clustering — group terms sharing most genes; keep one representative per group.
• Parent terms / semantic similarity — collapse child GO terms to a parent; REVIGO-style reduction by semantic similarity.
• Report a representative term per cluster plus the count of related terms.

Significance vs relevance

• Check the overlap count, not just the p-value. "Term enriched, padj=0.01" with 2 genes out of a 1500-gene set is rarely meaningful.
• Watch gene-set size: tiny sets reach significance with few genes; huge, generic sets ("metabolic process") are uninformative — the min_size/max_size filters (15–500) exist for this reason.
• Assess power relative to the universe and term size; there is no universal minimum/maximum hit-list length. GSEA requires an appropriate complete ranking, which cannot be reconstructed from a hit list alone.

Reproducibility checklist

• Record exact library names and versions/date (Enrichr/GO libraries drift).
• Record the background used (or state the default).
• For GSEA, record permutation_num, seed, min_size, max_size, weight, and the ranking metric (e.g., DESeq2 stat).
• State the organism and gene-ID namespace.
• Save the full results table, not just the filtered top hits.

Publication table template

Report a compact, reviewer-friendly table:

For ORA use Odds Ratio + Overlap (k/K); for GSEA use NES + leading-edge size. Note method, library version, background, and correction in the legend.

Common misinterpretations

• "Enriched pathway X" does not mean pathway X is activated — ORA is direction-agnostic unless you split up/down lists; GSEA NES sign gives rank direction, not necessarily activation (sets can contain inhibitors).
• Overlapping significant GO terms are not independent findings.
• Absence of enrichment ≠ absence of biology (power, annotation gaps, wrong background, or ID mismatch can all hide real signal).
• Don't compare raw ES across gene sets as effect sizes; NES still depends on the null, gene-set collection and submitted ranking. Neither proves causality.
• Report Monte Carlo resolution and method; nominal p-values at the simulation floor do not establish arbitrarily tiny probabilities. More permutations do not fix confounding, selection bias, or an invalid null.