Enrichment Score (ES)

The ES is calculated using a running sum statistic that walks down the entire ranked gene list. For each gene encountered, the statistic increases if the gene is a member of the target set and decreases if it is not. The increment is weighted by the gene's correlation with the phenotype, ensuring that genes with stronger differential expression have a greater impact on the score. The final ES is the maximum deviation from zero encountered during this walk, capturing the point where the set is most non-randomly distributed.

GSEA offers two primary modes for calculating the ES, controlled by the exponent p applied to the correlation weights:

• Weighted (p=1, default): The contribution of each gene to the running sum is proportional to its correlation with the phenotype. This is the classic Kolmogorov-Smirnov-like statistic that emphasizes genes with strong differential signals.
• Unweighted (p=0): All genes in the set contribute equally. This reduces to a standard Kolmogorov-Smirnov statistic and is more sensitive to sets where many genes show a small, coordinated shift rather than a few large changes.

The sign of the ES indicates the end of the ranked list where the gene set is concentrated:

• A positive ES means the set is enriched at the top of the ranked list, among genes upregulated in the phenotype of interest.
• A negative ES indicates enrichment at the bottom, among downregulated genes. A score near zero suggests the gene set's members are randomly scattered throughout the ranking, showing no association with the phenotypic distinction.

The leading-edge subset is the core group of genes within the gene set that drives the enrichment signal. These are the genes that appear in the ranked list before the point where the running sum reaches its maximum deviation. Identifying this subset is crucial for biological interpretation, as it isolates the specific molecular players responsible for the pathway's differential activity, filtering out noise from non-contributing members of a larger gene set.

The raw ES is not directly interpretable as a p-value. Its statistical significance is assessed by comparing it to a null distribution generated through phenotype permutation. The sample labels are randomly shuffled, the entire GSEA calculation is repeated thousands of times, and a nominal p-value is derived from the fraction of permutations where the random ES exceeds the observed ES. This non-parametric approach preserves the complex correlation structure of the gene expression data.

To enable fair comparison across gene sets of different sizes, the ES is normalized to produce the Normalized Enrichment Score (NES). The ES is divided by the mean of all positive (or negative) ES values from the permutation runs for gene sets of the same size. This correction accounts for the fact that larger gene sets naturally tend to have smaller maximum deviations, allowing the analyst to rank all tested pathways by their corrected statistical significance.

The sequential calculation at the heart of ES computation. The algorithm walks down a ranked gene list from top to bottom, incrementing a running sum when encountering a gene in the target set and decrementing it otherwise. The magnitude of increment is weighted by the gene's correlation with the phenotype, while the decrement is proportional to the number of genes not in the set. This creates a random walk-like trajectory that reveals whether the gene set clusters at the extremes of the ranked list.

The ES corrected for gene set size and multiple hypothesis testing. Because larger gene sets naturally accumulate higher running sum deviations, raw ES values are not directly comparable across sets of different sizes. NES is calculated by dividing each ES by the mean of all ES values obtained from permutations of the same gene set size. This normalization enables cross-gene-set comparisons and accounts for dataset-specific correlation structures.

The core group of genes within an enriched set that contributes most significantly to the enrichment signal. These genes appear at the extreme ends of the ranked list—before the point of maximum deviation in the running sum. Identifying the leading-edge subset is critical for:

• Mechanistic interpretation: pinpointing which specific genes drive the pathway signal
• Enrichment Map construction: defining overlap between related gene sets
• Drug target prioritization: focusing on the most statistically impactful members of a pathway

A resampling strategy for estimating the null distribution of the ES. Rather than permuting gene labels—which destroys the correlation structure of expression data—phenotype permutation randomly shuffles sample class labels while preserving the gene-gene correlation matrix. For each permutation, the entire GSEA procedure is repeated, generating a distribution of ES values under the null hypothesis. The observed ES is then compared to this empirical null to calculate a nominal p-value. This approach maintains the biological dependencies inherent in the dataset.

The expected proportion of false positives among all gene sets called significant. When testing thousands of gene sets simultaneously, traditional p-value thresholds produce many false positives. FDR correction—typically via the Benjamini-Hochberg procedure—adjusts for this multiplicity. In GSEA output, an FDR q-value of 0.05 means that 5% of gene sets with equal or lower q-values are expected to be false discoveries. FDR is the primary statistical filter for reporting robust enrichment results.

The curated library of annotated gene sets against which ES is calculated. The Molecular Signatures Database (MSigDB) organizes sets into collections:

• Hallmark (H): 50 refined biological states with minimal redundancy
• C2 (Curated): KEGG, Reactome, and other pathway databases
• C5 (GO): Gene Ontology biological process, molecular function, cellular component
• C6 (Oncogenic): Signatures of cellular perturbations
• C7 (Immunologic): Immune cell states and responses The choice of collection directly impacts the biological interpretability of ES results.