- Dimensionality reduction is an important computational step in single-cell RNA sequencing (scRNA-seq) and other high-dimensional biological data analyses. Modern transcriptomic experiments measure the expression of thousands of genes across individual cells, producing datasets with many variables. Although these measurements contain valuable biological information, analysing and visualizing thousands of dimensions simultaneously is difficult. Dimensionality reduction transforms high-dimensional data into a smaller number of informative dimensions while attempting to preserve important patterns and relationships. Principal component analysis (PCA), uniform manifold approximation and projection (UMAP), and t-distributed stochastic neighbour embedding (t-SNE) are three widely used methods for exploring cellular diversity, identifying cell populations and visualizing complex gene-expression patterns.
- In single-cell RNA sequencing, the expression matrix generally contains cells as observations and genes as variables. Each cell is represented by a vector of gene-expression measurements, and each gene contributes a dimension to the dataset. For example, a dataset containing 5,000 cells and 20,000 genes has 5,000 observations described by 20,000 features. Many genes show little variation across cells, while others contain redundant or correlated information. Technical variation, sequencing depth and measurement noise can also influence the data. Dimensionality reduction helps summarize these complex expression profiles into a smaller set of features that captures important sources of variation.
- Principal component analysis (PCA) is a linear dimensionality-reduction technique that transforms correlated variables into a smaller number of uncorrelated variables called principal components (PCs). Each principal component is a weighted combination of the original variables and represents a direction of variation in the dataset. The first principal component explains the greatest possible amount of variation, while the second explains the greatest remaining amount subject to being orthogonal to the first. Subsequent components capture progressively smaller proportions of the total variation.
- PCA begins with a suitable expression matrix, usually after quality control, normalization and selection of highly variable genes. The expression values are often centred by subtracting the mean expression of each gene across cells. Depending on the analysis, genes may also be scaled to comparable variances. PCA then identifies directions that maximize the variance in the transformed data. These directions are calculated from the covariance or correlation structure of the variables, commonly using eigenvalue decomposition or singular value decomposition (SVD).
- The mathematical representation of PCA can be expressed as:
- PCₖ = w₁ₖX₁ + w₂ₖX₂ + … + wₚₖXₚ
- Here, PCₖ represents the k-th principal component, X₁ to Xₚ represent the original gene-expression variables, and w₁ₖ to wₚₖ represent the weights assigned to those variables in that component. Each principal component summarizes a particular pattern of variation across the original genes.
- The amount of variation explained by each principal component is determined by its associated eigenvalue. Components with larger eigenvalues account for more variation in the dataset. The proportion of variance explained by a principal component can be calculated as:
- Explained variance ratio = Eigenvalue of the component / Sum of all eigenvalues
- This measure helps researchers decide how many principal components to retain. For example, the first 30 components may capture a substantial proportion of the meaningful variation in a dataset, although the appropriate number depends on the biological system and experimental design. Scree plots, cumulative variance plots and downstream clustering performance can help guide this decision.
- In single-cell analysis, PCA is commonly used as an initial dimensionality-reduction step before neighbourhood graph construction and nonlinear visualization. It reduces the number of features from thousands of genes to a smaller set of principal components. These components are then used to calculate similarities between cells and identify groups with related expression profiles. PCA can reveal major sources of variation associated with cell types, developmental stages, disease conditions or experimental treatments. However, because PCA is linear, it may not fully capture complex nonlinear relationships in gene-expression data.
- Uniform manifold approximation and projection (UMAP) is a nonlinear dimensionality-reduction method widely used for visualizing high-dimensional biological datasets. It aims to preserve important local relationships between observations while also retaining aspects of the broader structure of the data. UMAP constructs a representation of the neighbourhood relationships among cells in the original high-dimensional space and then optimizes a low-dimensional arrangement that reflects these relationships.
- UMAP generally begins by identifying the nearest neighbours of each cell using a distance metric such as Euclidean distance or cosine distance. These neighbourhood relationships are converted into a weighted graph representing the similarity between cells. UMAP then constructs a low-dimensional graph, usually in two or three dimensions, and optimizes the positions of cells so that the low-dimensional representation preserves important relationships from the original data. The resulting visualization often shows groups of cells with similar gene-expression profiles located close together.
- UMAP is particularly useful for visualizing complex cellular populations and identifying potential subpopulations. In single-cell RNA-seq studies, it can reveal clusters associated with immune-cell types, developmental stages, tissue-specific populations or disease-related cellular states. It is also useful for examining continuous biological processes, such as cell differentiation and developmental trajectories. However, the distances and sizes of clusters in a UMAP plot should not automatically be interpreted as exact measures of biological similarity, abundance or developmental time.
- The appearance of a UMAP visualization depends on several parameters. The number of nearest neighbours controls how many neighbouring cells influence the local structure. Smaller values emphasize local relationships and may produce more fragmented clusters, while larger values incorporate broader neighbourhood information. The min_dist parameter controls how closely points can be packed together in the low-dimensional representation. A smaller value often produces more compact clusters, whereas a larger value generally creates more spread-out visualizations. The choice of distance metric, number of dimensions and random initialization can also influence the result.
- t-distributed stochastic neighbour embedding (t-SNE) is another nonlinear dimensionality-reduction technique used to visualize high-dimensional biological data. It is designed primarily to preserve local neighbourhood relationships by ensuring that observations that are similar in the original space remain close together in the low-dimensional representation. t-SNE converts similarities between observations into probability distributions and then attempts to minimize the difference between the high-dimensional and low-dimensional similarity distributions.
- The main objective of t-SNE is to preserve local structures rather than accurately represent all global distances. Cells with similar gene-expression profiles are therefore often placed close together, producing visually distinct groups. This makes t-SNE useful for exploring complex cell populations, identifying subclusters and examining cellular heterogeneity. However, distances between separate clusters, their relative sizes and the gaps between them may not reflect their true biological relationships.
- One important parameter in t-SNE is perplexity, which influences the effective number of neighbouring observations considered when calculating similarities. Lower perplexity values emphasize smaller neighbourhoods, while higher values incorporate broader relationships. The appropriate value depends on dataset size and structure. Other important settings include the learning rate, number of iterations, initialization method and distance metric. Different parameter settings can produce different visualizations, so researchers should assess whether the observed patterns are stable and biologically meaningful.
- PCA, UMAP and t-SNE differ in their mathematical principles and intended uses. PCA is a linear method that identifies major directions of variation and produces components that can be examined quantitatively. UMAP and t-SNE are nonlinear methods that generate low-dimensional representations designed primarily to preserve neighbourhood relationships. PCA is commonly used for initial data compression and noise reduction, while UMAP and t-SNE are frequently used for visual exploration of cellular populations.
- In a typical single-cell RNA-seq workflow, PCA is performed after quality control, normalization, variance stabilization and highly variable gene selection. The first several principal components are then used to construct a cell-neighbourhood graph. UMAP or t-SNE can be applied to this graph or to the selected principal components to produce two-dimensional or three-dimensional visualizations. Clustering algorithms such as Leiden and Louvain can also use the neighbourhood graph to identify groups of cells with similar expression profiles. Thus, visualization and clustering are related but distinct processes: a two-dimensional plot is not itself a definitive cell-type classification.
- Several software packages support dimensionality reduction in single-cell transcriptomics. Seurat provides PCA through RunPCA, UMAP through RunUMAP and t-SNE through RunTSNE. Scanpy provides corresponding functions such as sc.tl.pca, sc.tl.umap and sc.tl.tsne. These frameworks also support neighbourhood graph construction, clustering, marker-gene identification and visualization. Other computational libraries, including scikit-learn, provide general implementations of PCA and t-SNE, while UMAP is available through the umap-learn package.
- Dimensionality reduction is also used to investigate biological trajectories and cellular differentiation. For example, cells collected at different developmental stages may show gradual changes in gene expression rather than forming completely separate populations. PCA can reveal major axes of transcriptional variation, while UMAP and t-SNE can help visualize the distribution of cells along complex expression patterns. However, neither UMAP nor t-SNE directly establishes developmental direction or pseudotime. Dedicated trajectory-inference methods such as Monocle3, Slingshot and PAGA are used to investigate potential developmental relationships, while RNA velocity methods such as scVelo can provide additional information about the direction of transcriptional change.
- An important limitation of dimensionality reduction is that the resulting visualization depends on the quality of the original data and the preprocessing methods used. Poor-quality cells, doublets, ambient RNA contamination, batch effects and inappropriate normalization can create misleading patterns. Differences in the selected genes, scaling procedures, distance metrics and algorithm parameters can also affect the resulting representation. Researchers should therefore assess the stability of their findings, compare alternative parameter settings and interpret visual patterns alongside gene-expression markers and biological knowledge.
- Batch effects require particular attention when datasets from multiple samples, laboratories or experimental conditions are combined. Technical differences may cause cells to separate according to sample origin rather than biological identity. Batch-aware integration approaches, such as Harmony, Seurat integration, fastMNN and scVI, can help align comparable populations before dimensionality reduction. However, integration should be performed carefully to avoid removing genuine biological differences, particularly when experimental conditions are expected to produce substantial changes in gene expression.
- In single-nucleus RNA sequencing (snRNA-seq), PCA, UMAP and t-SNE are also used to explore nuclear expression profiles. However, nuclear transcriptomes differ from whole-cell transcriptomes because they contain a higher relative proportion of intronic and unspliced RNA and may show different expression distributions. Appropriate quality control, normalization, feature selection and interpretation are therefore important when applying these methods to nuclear data.
- In animal genomics and breeding research, dimensionality reduction can help investigate cellular diversity and gene-expression differences in tissues associated with economically important traits. Single-cell studies of muscle, mammary gland, ovary, testis, liver and immune tissues can use PCA, UMAP and t-SNE to explore cell populations associated with growth, muscle development, milk production, fertility, metabolism and disease resistance. For example, UMAP may reveal distinct immune-cell populations in an animal tissue, while PCA can identify major patterns of gene-expression variation across developmental stages or experimental conditions. Integrating these analyses with genome-wide association studies, quantitative trait loci, genomic variation and epigenomic information can help connect cellular expression patterns with the genetic basis of complex traits.
- Dimensionality reduction with PCA, UMAP and t-SNE is therefore fundamental to exploring high-dimensional single-cell transcriptomic data. PCA provides a linear summary of major expression patterns, while UMAP and t-SNE generate nonlinear visualizations that help reveal local cellular relationships and population structure. These methods make complex biological datasets easier to explore and interpret, but their results must be evaluated in the context of data quality, preprocessing, parameter selection and biological evidence. When used appropriately, they provide a foundation for cell clustering, cell-type identification and the investigation of cellular heterogeneity in bioinformatics.