34  How to read an enrichment result

An enrichment result is an association between an input feature set and a knowledge source. Read it from the input and the null model outward, rather than from the term names alone.

34.1 Start with the tested question

  • ORA asks whether a thresholded set contains more members of a term than expected under the chosen universe.
  • GSEA asks whether members of a term accumulate toward one end of a complete ranking.
  • compareCluster repeats the selected enrichment function for several groups; the groups are comparable only when their input and background definitions are compatible.
  • Network-aware methods add topology or feature weights to the evidence model and should not be interpreted as ordinary ORA or GSEA.

The result class records which route was used. Keep the original enrichResult, gseaResult, or compareClusterResult object with the analysis.

34.2 Check the input and background

Before interpreting a term, check:

  1. identifier type and mapping rate;
  2. the selected gene vector or the complete ranked vector;
  3. the universe used for ORA;
  4. the gene-set source, organism, release, and filtering rules;
  5. the number of tested terms and the multiple-testing method.

A small p-value cannot correct an incorrect identifier namespace or an unsuitable universe. For GSEA, inspect the ranking direction and the score type before assigning biological direction.

34.3 Read significance together with effect and coverage

For ORA, inspect Count, GeneRatio, BgRatio, pvalue, and p.adjust together. A term with a small adjusted p-value but a small overlap may be sensitive to the background and annotation coverage.

For GSEA, inspect ES or NES, p.adjust, setSize, and the core_enrichment or leading-edge genes. A positive NES indicates accumulation toward the top of the supplied ranking; it does not by itself establish activation or causality.

For comparisons, inspect the cluster label, term identity, effect measure, and adjusted significance together. A term appearing in only one group may reflect biology, input size, annotation coverage, or differences in the tested background.

34.4 Reduce redundancy before summarizing themes

GO and other hierarchical resources often return related terms. Use pairwise_termsim() and simplify(), or bayes_enrich() when joint term–gene coverage is the target, before turning a long table into a short list of themes. Keep the original object and the reduction parameters so each retained term can be traced back to the full result.

34.5 Follow the evidence to genes and networks

Use core_enrichment, geneID, or geneInCategory() to identify the genes behind a term. Compare these genes with fold changes, expression patterns, PPI context, and independent knowledge. A term name is a label for an annotation set, not a mechanistic conclusion.

34.6 Choose the next view

  • Use a dot or bar plot for ranked term summaries.
  • Use gseaplot2() or hplot() to inspect a GSEA signal along the ranking.
  • Use cnetplot() for shared genes among a small set of terms.
  • Use emapplot() or semantic similarity to organize overlapping terms.
  • Use compareCluster plots when the question is about differences among groups.

Continue to evidence-guided interpretation only after these checks are complete. For a report that preserves the analysis record, see reproducible reporting.

34.7 Common interpretation errors

  • Treating an adjusted p-value as an effect size.
  • Treating a term name as a causal mechanism.
  • Comparing results made with different universes without noting the difference.
  • Interpreting a readable symbol table as if it changed the statistical test.
  • Reporting only the top terms without the database, release, identifier type, and parameters.