Principal Component Analysis (PCA)

PCA sequentially identifies orthogonal directions—principal components (PCs)—that capture the maximum remaining variance in the data. PC1 explains the largest proportion of total variance, PC2 explains the second largest while being uncorrelated with PC1, and so on. In RNA-seq data, this means the first few components often separate samples by the strongest biological signal, such as disease versus control, while later components capture subtle or technical variation.

A typical RNA-seq dataset contains expression measurements for 20,000+ genes across dozens of samples. PCA compresses this high-dimensional space into a manageable number of components—typically 2 to 10—that retain the majority of the information. This allows researchers to:

• Visualize sample relationships in 2D or 3D scatter plots
• Reduce noise by discarding low-variance components
• Use PCs as input features for downstream clustering or classification algorithms

Each principal component is a linear combination of the original gene expression variables. The weights, called loadings, indicate how much each gene contributes to a given component. Genes with large absolute loadings on PC1 are the primary drivers of the largest variance source. This linearity makes PCA interpretable—you can trace a component back to specific genes—but also means PCA cannot capture non-linear relationships or complex gene-gene interactions.

In genomics, PCA serves as a critical quality control diagnostic. When samples cluster by technical covariates—such as sequencing date, processing technician, or library preparation batch—rather than by biological condition, it reveals the presence of batch effects. This visual evidence prompts researchers to apply correction methods like ComBat-seq or limma's removeBatchEffect before proceeding with differential expression analysis.

PCA is mathematically computed through eigendecomposition of the covariance matrix or, more commonly in genomics, through Singular Value Decomposition (SVD) of the mean-centered data matrix. SVD factorizes the expression matrix into three matrices: left singular vectors (sample scores), singular values (variance explained), and right singular vectors (gene loadings). This decomposition is the computational engine behind PCA visualizations in tools like DESeq2's plotPCA and scikit-learn's PCA.

PCA assumes that variance corresponds to meaningful signal and that the directions of maximum variance are the most informative. Key limitations include:

• Sensitivity to scaling: Genes with larger expression magnitudes dominate unless data are standardized
• Linearity constraint: Cannot model curved manifolds or non-linear relationships
• Interpretability: PCs are mathematical constructs, not directly measurable biological entities
• Outlier sensitivity: Extreme values can distort principal component directions For non-linear patterns, alternatives like t-SNE or UMAP are often preferred for visualization, though they lack PCA's interpretable loadings.

t-Distributed Stochastic Neighbor Embedding is a non-linear dimensionality reduction technique that excels at preserving local structure. Unlike PCA, which prioritizes global variance, t-SNE models pairwise similarities as conditional probabilities and minimizes the Kullback-Leibler divergence between high-dimensional and low-dimensional distributions. In single-cell genomics, t-SNE is often run on the top 30-50 principal components rather than raw data, using PCA as a denoising pre-processing step. The heavy-tailed Student-t distribution in the low-dimensional space prevents the crowding problem that plagues earlier methods like SNE.

Uniform Manifold Approximation and Projection constructs a fuzzy topological representation of high-dimensional data and optimizes a low-dimensional embedding to preserve both local and global structure. UMAP has largely replaced t-SNE in single-cell analysis because it better preserves global relationships between clusters and runs significantly faster. Like t-SNE, UMAP is typically applied to PCA-reduced data. The algorithm assumes data lies on a Riemannian manifold and uses fuzzy simplicial set constructions to model connectivity, making it more faithful to the underlying data topology than t-SNE's probabilistic approach.

A diagnostic visualization that displays the eigenvalue or percentage of variance explained by each principal component in descending order. The 'elbow' in the curve indicates the point where additional components contribute diminishing returns. In genomics, a sharp drop after PC1-PC2 followed by a gradual decline suggests the first few components capture dominant biological or batch signals, while the rest represent noise. The cumulative variance plot variant shows the running total of explained variance, helping researchers decide how many components to retain for downstream analysis—typically targeting 80-90% cumulative variance.

The linear algebra engine underlying PCA. SVD factorizes the centered data matrix X into three matrices: X = UΣVᵀ, where U contains left singular vectors (sample scores), Σ holds singular values (proportional to eigenvalues), and Vᵀ contains right singular vectors (gene loadings). The principal components are the columns of UΣ, and the squared singular values divided by (n-1) give the variance explained. SVD is preferred over eigendecomposition of the covariance matrix for numerical stability, especially when the number of features (genes) far exceeds the number of samples—the classic p >> n problem in genomics.