Introduction

Gene set enrichment analysis is widely used to interpret differential expression results, yet one implementation choice is often treated as routine rather than consequential: the definition of the background gene universe. We investigate whether small, realistic changes in background universe specification materially alter downstream enrichment conclusions. Using publicly available transcriptomic datasets with binary group comparisons, we evaluate multiple commonly used background definitions, including all annotated genes, all detected genes, expression-filtered genes, and low-expression-pruned genes. For each dataset, we perform differential expression analysis followed by over-representation analysis across Hallmark, KEGG, Reactome, and Gene Ontology gene sets. We quantify instability using five metrics: change in the number of significant pathways, Top-N overlap, rank correlation, significance flip rate, and effect-size variation.

Motivation

In over-representation analysis, enrichment p-values are explicitly computed relative to a universe of genes considered eligible for selection. Yet this universe is defined inconsistently across studies. Some analyses use all annotated genes in the organism; others use all genes present on a platform; others restrict the background to genes detected above some expression threshold or to genes surviving low-count filtering. These alternatives are individually defensible, but they are not equivalent. Small changes in universe construction alter both the denominator of enrichment tests and the effective null model against which pathway over-representation is assessed.

Methods

We performed a multi-dataset sensitivity analysis of pathway enrichment under alternative background gene universe definitions. For each transcriptomic dataset, we conducted a standard differential expression workflow, generated a list of significant genes, and then repeated ORA using multiple plausible background universes while holding all other analytical steps fixed. We compared the resulting enrichment outputs using instability metrics designed to capture both discrete and continuous changes in interpretation.

We evaluated multiple realistic universe definitions:

• U1: All annotated genes
• U2: Platform-measured genes
• U3: All detected genes
• U4: Expression-filtered genes
• U5: Low-expression-pruned genes
• U6: Annotation-intersected genes

We performed over-representation analysis using commonly used pathway collections:

• MSigDB Hallmark
• KEGG
• Reactome
• Gene Ontology Biological Process

We quantified enrichment instability using five primary metrics:

1. Significant pathway count change
2. Top-N overlap
3. Rank correlation
4. Significance flip rate
5. Effect-size variation

Results

Across datasets, enrichment conclusions are consistently sensitive to background definition, with the strongest instability concentrated in small gene sets, near-threshold pathways, and analyses with modest differential signal. In representative comparisons, pathways that are significant under one plausible background definition frequently lose significance under another, despite unchanged expression data and identical differential expression procedures. The effect is more pronounced for over-representation analysis than for rank-based methods.

Top-ranked pathways are only partially stable under realistic universe perturbations. In many dataset–database combinations, the top 10 pathways overlapped only partially when the universe moved from one realistic definition to another. Overlap improved for larger Top-N sets, but this reflects the expected dilution that occurs as more pathways are included.

The strongest instability was concentrated among pathways near the significance boundary. These pathways frequently crossed the adjusted p-value threshold when the universe changed, despite no change to the gene-level differential expression input. By contrast, pathways with very strong enrichment tended to remain significant across universe definitions, and pathways with very weak evidence typically remained non-significant.

Discussion

These findings do not imply that enrichment analysis is unreliable in general; rather, they show that background universe specification is an under-reported analytical degree of freedom with measurable impact on biological interpretation. Strong signals often remain strong. Instead, the problem is selective instability in the very region where interpretation is most contestable.

A practical recommendation follows directly. Enrichment workflows should include a minimal sensitivity analysis over plausible universe definitions, especially when the DE gene list is small, pathway significance is borderline, conclusions rely on a few small pathways, or ORA is the primary interpretive tool.

Limitations

This study has several limitations. First, our conclusions are strongest for ORA-style analyses and may not generalize equally to all enrichment frameworks. Second, the magnitude of instability depends on dataset composition, filtering choices, and pathway database structure. Third, this work does not resolve which universe definition is “correct” in all settings. Our point is not that one choice is universally best, but that plausible choices are not interchangeable.

Conclusion

Gene set enrichment analysis remains a useful interpretive tool, but its conclusions are more sensitive to background universe specification than is often acknowledged. Small, realistic changes in universe definition can alter pathway significance, ranking, and interpretive emphasis, particularly for near-threshold and small gene sets. At minimum, enrichment studies should report universe definition explicitly and assess whether key pathway claims persist under plausible alternatives.