Over-Representation Analysis (ORA)

ORA fundamentally relies on the hypergeometric distribution (or Fisher's exact test) to calculate enrichment probability. The test models the likelihood of observing k or more differentially expressed genes (DEGs) in a pathway of size n, given a total background of N genes, of which K are DEGs. It answers the question: 'Given my total list of genes, is the overlap with this specific pathway greater than expected by random chance?' The resulting p-value quantifies this statistical significance without requiring the original continuous expression data.

Unlike rank-based methods, ORA requires a dichotomized gene list as input. A researcher must apply an arbitrary threshold (e.g., p-value < 0.05, |log2 fold change| > 1) to a differential expression result to create a binary 'interesting' vs. 'not interesting' classification. This simplification is both a strength (simplicity) and a critical weakness, as information about the magnitude and direction of expression changes is discarded. The final enrichment result is highly sensitive to this chosen cutoff.

A core assumption of the hypergeometric test is that each gene has an equal and independent probability of being selected. In biological reality, this is often violated due to co-regulation and pathway crosstalk. Furthermore, the choice of the background gene set (e.g., all protein-coding genes vs. only expressed genes) dramatically impacts results. Using an inappropriate background can lead to inflated significance and spurious findings, making background selection a critical parameter for reproducible ORA.

ORA tests hundreds or thousands of gene sets simultaneously, which inflates the risk of false positives. A multiple hypothesis testing correction is mandatory. The False Discovery Rate (FDR), often estimated via the Benjamini-Hochberg procedure, is the standard metric. An FDR-adjusted p-value of 0.05 implies that 5% of the results deemed significant are expected to be false positives. Reporting unadjusted p-values in ORA is a common statistical error.

ORA's primary limitation is its 'threshold-free' rival, Gene Set Enrichment Analysis (GSEA). By requiring a binary gene list, ORA ignores the trend of coordinated but subtle changes that don't pass a hard cutoff. A pathway with many genes showing a 1.4-fold change (p=0.06) would be missed by ORA but detected by GSEA's running sum statistic. ORA is best suited for validating a small number of top hits, not for discovering subtle, coordinated pathway-level shifts.