Single-Sample GSEA (ssGSEA)

ssGSEA ranks genes within a single sample by absolute expression and compares the empirical cumulative distribution functions (ECDFs) of genes inside and outside the gene set. This produces a Normalized Enrichment Score (NES) per sample, independent of other samples in the cohort. Unlike traditional GSEA, which requires phenotype labels to rank genes by differential expression, ssGSEA uses the sample's own expression profile as the ranking metric.

• Input: A single sample's expression vector and a gene set
• Output: One enrichment score per sample-gene set pair
• Key distinction: No phenotype permutation required; null distributions are built from gene-level randomization

The core algorithm walks down a gene list ranked by expression values within a single sample. It computes two empirical cumulative distribution functions: one for genes in the target gene set and one for genes outside it. The enrichment score is the sum of differences between these ECDFs, weighted by the rank position.

• Genes are ranked from highest to lowest expression
• A running sum statistic increases when encountering a gene set member and decreases for non-members
• The maximum deviation from zero becomes the raw enrichment score
• This approach captures coordinated upregulation or downregulation of pathway members within a single sample

ssGSEA eliminates the requirement for predefined class labels (e.g., tumor vs. normal). This makes it ideal for exploratory analyses where phenotypes are unknown, heterogeneous, or continuous. Each sample becomes its own analytical unit, enabling pathway-level comparisons across large cohorts without binary grouping.

• Use cases: Tumor subtype discovery, drug response prediction, single-cell pathway analysis
• Advantage: Detects pathway activity gradients rather than just on/off states
• Integration: Resulting sample-by-pathway matrices can feed into clustering algorithms or machine learning classifiers

ssGSEA and GSVA are both single-sample methods but differ in their statistical frameworks. ssGSEA uses a rank-based ECDF approach, while GSVA employs a non-parametric kernel estimation of the cumulative density function. GSVA applies a Kolmogorov-Smirnov-like random walk, whereas ssGSEA uses a weighted running sum.

• ssGSEA: Weighted ECDF difference, sensitive to the extremes of the ranked list
• GSVA: Kernel-smoothed CDF estimation, more robust to outliers
• Both produce: Sample-level pathway enrichment scores suitable for downstream modeling
• Selection criteria: ssGSEA for detecting strong pathway activation; GSVA for noisy or sparse data

The sample-by-pathway matrix generated by ssGSEA serves as a feature matrix for predictive modeling. Each column represents a biological pathway, and each row represents a sample. This transforms high-dimensional gene expression data (~20,000 genes) into a biologically interpretable feature space (~50–5,000 pathways).

• Classification: Train models to predict drug response or disease subtype using pathway scores as features
• Clustering: Apply hierarchical clustering or t-SNE to pathway enrichment profiles for patient stratification
• Survival analysis: Use pathway scores as continuous covariates in Cox proportional hazards models
• Biomarker discovery: Identify pathways whose ssGSEA scores significantly differ between outcome groups

ssGSEA identifies the leading-edge subset—the core genes driving the enrichment signal—within each individual sample. This reveals which specific pathway members are most responsible for the observed activity, enabling fine-grained mechanistic interpretation. Unlike traditional GSEA where the leading edge is defined across a phenotype contrast, ssGSEA pinpoints sample-specific pathway drivers.

• Output: A binary mask indicating which genes in the set contribute to the maximum enrichment score
• Application: Identifying patient-specific therapeutic targets within activated oncogenic pathways
• Visualization: Heatmaps of leading-edge membership across samples reveal heterogeneity in pathway utilization

The foundational method from which ssGSEA was derived. Unlike the single-sample variant, classical GSEA requires a phenotype-defined contrast between two biological states (e.g., tumor vs. normal). It computes an Enrichment Score (ES) by walking down a gene list ranked by differential expression, detecting whether a priori defined gene sets cluster at the extremes. The significance is assessed via phenotype permutation, preserving gene-gene correlations within the dataset.

ssGSEA produces raw Enrichment Scores that are not directly comparable across gene sets of different sizes. The Normalized Enrichment Score (NES) corrects for this by accounting for gene set size and the correlation structure of the dataset. In ssGSEA, normalization is achieved by dividing the raw ES by the range of scores obtained from gene-level permutation, producing a score bounded between -1 and 1 that represents the degree of absolute enrichment.

Within an ssGSEA result, the leading-edge subset identifies the core genes driving the enrichment signal. These are the genes that appear at the extreme ends of the ranked expression profile and contribute most to the running sum statistic. For single-sample interpretation, the leading edge reveals which specific members of a pathway are most activated or suppressed in that individual, enabling mechanistic hypothesis generation about the molecular drivers of a patient's disease state.