Overview
Dimensionality reduction maps high-dimensional data, where each sample has thousands or tens of thousands of measured features, into a low-dimensional space, typically two or three dimensions, that preserves important structure. This serves two primary purposes: visualization, allowing researchers to see patterns, clusters, and outliers by eye; and denoising, removing irrelevant variation that degrades downstream analysis. Biological data, from gene expression matrices to mass spectra, is inherently high-dimensional and often redundant, making dimensionality reduction an essential preprocessing step.
Methods
Principal component analysis (PCA) finds orthogonal axes of maximum variance through linear combinations of the original features. It is fast, deterministic, and frequently employed for initial exploration. t-distributed stochastic neighbor embedding (t-SNE) constructs a probability distribution over pairwise similarities in high-dimensional space and minimizes the Kullback-Leibler divergence to a low-dimensional map, excelling at revealing local structure and clusters. Uniform Manifold Approximation and Projection (UMAP) builds on similar principles but better preserves global structure and is significantly faster. Autoencoders use neural networks to learn non-linear embeddings and can capture complex manifolds. Each method has hyperparameters, perplexity for t-SNE, n_neighbors for UMAP, that strongly influence the resulting visualization.
Practical Protocol
A practical dimensionality reduction workflow for single-cell RNA-seq data begins with a genes-by-cells count matrix. After quality filtering, normalization, and log-transformation, the top 2,000 highly variable genes are selected. PCA is applied first to reduce the matrix from roughly 20,000 dimensions to 50 principal components, capturing dominant variation while removing technical noise. The researcher inspects the variance explained using a scree plot and typically retains the top 20–30 components for downstream analysis. These PCA-reduced coordinates serve as input for UMAP. The key hyperparameters are set: n_neighbors = 15, min_dist = 0.1, and n_components = 2. The UMAP embedding is plotted, with each cell colored by known cell type markers, clustering assignment, or experimental condition. Interpreting the plot involves identifying distinct cell populations as isolated point clouds, continuous trajectories suggesting differentiation pathways, and outlier cells that may represent doublets or rare populations. t-SNE can be used as an alternative for finer local structure, though UMAP better preserves global relationships. For example, in a study of peripheral blood mononuclear cells, this pipeline revealed 22 distinct immune cell types from a single 10x Genomics run, including rare dendritic cell subsets comprising less than 1% of the total population. In cancer research, dimensionality reduction of single-cell tumor data has uncovered distinct subclonal populations with different proliferative capacities and drug sensitivities, guiding personalized treatment strategies.
Applications
Dimensionality reduction is standard practice for exploring DNA microarrays and gene expression data, visualizing cell populations in flow cytometry, and inspecting quality in proteomics and mass spectrometry experiments. Single-cell RNA-seq analysis pipelines routinely use PCA for initial denoising followed by UMAP for visualization, enabling the discovery of novel cell types and states from transcriptomic data.
