Hypergeometric Distribution

Unlike the binomial distribution, the hypergeometric model assumes sampling without replacement. Each gene drawn from the universe is not returned before the next draw. This is critical because a gene list for enrichment analysis contains unique identifiers—a gene cannot appear twice. The probability of selecting a gene belonging to a pathway changes after each draw, as the composition of the remaining population shifts. This dependency accurately reflects the discrete, finite nature of the genome background.

The distribution is fully defined by four parameters derived from the experimental context:

• N: Total number of genes in the background universe (e.g., all protein-coding genes).
• K: Number of genes in the universe that belong to the pathway of interest (successes in population).
• n: Number of genes in the user's differentially expressed (DE) list (sample size).
• k: The observed overlap—the number of DE genes that are also in the pathway.

The probability mass function calculates P(X = k), the exact chance of that overlap occurring randomly.

The core formula calculates the exact probability of observing exactly k pathway genes in a DE list of size n:

P(X = k) = [ C(K, k) * C(N-K, n-k) ] / C(N, n)

Where C(a, b) is the binomial coefficient "a choose b". The numerator multiplies the ways to choose the pathway genes by the ways to choose the non-pathway genes. The denominator represents all possible ways to draw the DE list from the universe. This combinatorial logic directly models the null hypothesis of random selection.

ORA uses the upper cumulative distribution function (CDF) to test for enrichment, not the PMF of a single point. The p-value is the probability of observing k or more pathway genes:

P(X ≥ k) = 1 - P(X ≤ k-1)

This one-tailed test asks: "What is the chance that random sampling would produce an overlap at least this extreme?" A small p-value (typically < 0.05 after correction) rejects the null hypothesis, suggesting the pathway is significantly over-represented. Depletion testing uses the lower tail P(X ≤ k).

For a 2×2 contingency table classifying genes by DE status (yes/no) and pathway membership (yes/no), the hypergeometric distribution is mathematically identical to Fisher's exact test. The test evaluates the null hypothesis of independence between the two classifications. This equivalence is why ORA tools often report "Fisher's exact p-value"—it is the same underlying combinatorial calculation applied to the overlap count. The test is exact because it computes the p-value directly from the distribution without asymptotic approximations.

The choice of N (the background gene universe) is the most critical parameter and a common source of bias. Using the entire genome as background inflates significance for pathways with genes detectable by the assay platform. Best practice dictates using only the expressed or measurable gene set as the background—genes that had a non-zero chance of being detected as DE. Mismatched backgrounds violate the assumption that every gene had an equal probability of being sampled, leading to inflated Type I error rates and spurious enrichments.

A statistical significance test used in the analysis of contingency tables, which is mathematically equivalent to the hypergeometric test for 2x2 tables. In Over-Representation Analysis, it calculates the exact probability of observing a specific overlap between a gene list and a pathway. Unlike chi-squared tests, it is valid for all sample sizes, making it the standard for small or sparse gene sets where asymptotic approximations fail.

A discrete probability distribution that models the number of successes in a fixed number of independent trials with a constant probability. The hypergeometric distribution converges to the binomial when the population size is large relative to the sample size. In pathway analysis, the binomial test is sometimes used as an approximation, but it incorrectly assumes sampling with replacement, leading to slightly liberal p-values compared to the exact hypergeometric model.

A set of statistical adjustments applied when testing thousands of pathways simultaneously. Since each ORA test produces a p-value, the probability of false positives inflates dramatically. Key methods include:

• Bonferroni correction: Controls the Family-Wise Error Rate by dividing alpha by the number of tests.
• Benjamini-Hochberg procedure: Controls the False Discovery Rate, offering a less conservative balance for exploratory biology.

An asymptotic alternative to the hypergeometric test that evaluates the independence between two categorical variables. In enrichment analysis, the Pearson chi-squared test compares observed gene counts to expected counts under the null hypothesis. It is computationally faster than the exact test but relies on large-sample approximations. It is generally considered unreliable when any expected cell frequency falls below 5, making the hypergeometric distribution the preferred exact method.

A complementary approach to ORA that uses a Kolmogorov-Smirnov-like running sum statistic rather than a discrete cutoff. Unlike the hypergeometric test, which requires an arbitrary threshold to define a gene list, GSEA ranks all genes by differential expression and evaluates whether a pathway's genes cluster at the extremes. This eliminates the information loss inherent in the binary classification step of Over-Representation Analysis.

The foundational data structure for the hypergeometric test in ORA, organizing genes into a 2x2 matrix:

• a: Differentially expressed genes in the pathway
• b: Differentially expressed genes not in the pathway
• c: Non-differential genes in the pathway
• d: Non-differential genes not in the pathway The probability of observing value 'a' or greater is calculated using the hypergeometric probability mass function with parameters N (total genes), K (pathway genes), and n (differential genes).