Dimensionality Reduction in Biology: UMAP, t-SNE, PCA, and When To Use Each

Dimensionality reduction compresses thousands of gene expression measurements into a two-dimensional plot a person can actually read, and uniform manifold approximation and projection (UMAP) has become the standard way to do that in single-cell biology.

Whether the clusters a UMAP plot shows are biologically meaningful is a separate question from whether the plot looks clean, and principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) answer that question in different ways. Each method preserves a different piece of the original high-dimensional structure, and misreading which piece survives is what turns a clean-looking plot into a misleading one.

What dimensionality reduction is for in biological data

Dimensionality reduction takes measurements across thousands of genes, proteins, or other biological features and compresses them into a handful of dimensions that can be plotted and interpreted. The goal is to filter noise and highlight the structure that actually distinguishes cell types, disease states, or experimental conditions, without requiring a researcher to visually parse thousands of columns by hand.

This compression is never lossless. Reducing high-dimensional data to two dimensions inevitably distorts some property of the original space, whether that property is pairwise distance, local neighborhood composition, or overall variance. That trade-off is examined in more detail later in this article. Understanding what each method actually preserves matters more than simply running the default pipeline. These same methods sit within the broader supervised and unsupervised learning toolkit that most life scientists now apply to biological data.

Most single-cell pipelines do not use one method in isolation. A typical single-cell RNA sequencing (scRNA-seq) workflow reduces the data to a smaller set of principal components first, identifies cell populations by clustering that reduced representation, then applies a nonlinear method such as t-SNE or UMAP for a final two-dimensional visualization. Each step in that chain has its own assumptions, and problems introduced early, such as an unrepresentative PCA step, can propagate into the final plot.

PCA: A linear baseline for genomic data

PCA is the oldest and simplest dimensionality reduction method used in biology, and it finds the directions of maximum variance in the data before projecting points onto those directions to produce a genuinely linear summary of the dataset. Because it is a fixed mathematical transformation rather than an iterative optimization, PCA output is deterministic, fast to compute, and directly interpretable in terms of which original genes or features drive each component.

That same linearity is PCA's central limitation for biological data. Cell populations that differ in a nonlinear way, such as branching differentiation trajectories or subtle transcriptional gradients, often overlap heavily in a two-dimensional PCA plot even when they are clearly distinguishable in the full gene expression space. A dimensionality reduction benchmark study across simulated and real scRNA-seq datasets found that PCA's clustering performance degraded faster than nonlinear methods as the number of true cell types increased, though PCA remained among the fastest and most stable methods with respect to changes in cell number.

PCA still earns a place in most modern pipelines, though usually not as the final visualization. It serves as the standard preprocessing step before t-SNE or UMAP, reducing noise and computational burden before the nonlinear methods run, and it remains the most transparent option when a reviewer or collaborator needs to know exactly which measured features are driving an observed pattern.

t-SNE: Strong single-cell clusters and the crowding problem

t-SNE is a nonlinear dimensionality reduction method that builds a probability distribution over pairs of points based on their similarity in high-dimensional space, then optimizes a low-dimensional layout so that similar points stay close together. This makes it particularly effective at revealing tight, well-separated clusters when a dataset genuinely contains discrete subpopulations, which is why it became a widely used default visualization once large scRNA-seq datasets became routine.

The method's known weakness is what its original authors termed the crowding problem: as the number of points grows large, the two-dimensional space available to represent moderately distant points becomes disproportionately small relative to the space needed for very close points, and clusters can blow up or bleed into one another. A theoretical analysis of t-SNE's optimization dynamics has since formalized why this happens, tracing the algorithm's early exaggeration phase to a form of spectral clustering and using that connection to explain both t-SNE's strong empirical performance on clustered data and its sensitivity to tuning parameters such as perplexity.

Perplexity, which roughly controls how many neighbors each point considers when building its local similarity distribution, has an outsized effect on the resulting plot. A perplexity set too low can fragment a single population into spurious subclusters, while one set too high can smear distinct populations together, so reporting the perplexity value used is a basic transparency step still missing from many published figures.

UMAP: How it works and the hyperparameters that matter in biology

UMAP models the data as a fuzzy topological structure built from each point's nearest neighbors, then finds a low-dimensional layout that preserves that structure as closely as possible. Two hyperparameters drive most of what a UMAP plot ends up looking like: the number of neighbors considered for each point, and the minimum distance allowed between points in the embedding, which controls how tightly packed the final clusters appear.

UMAP earned an early reputation for preserving global structure better than t-SNE, but that claim has not held up cleanly. A correction to the original comparison showed that UMAP's apparent advantage came from the default initialization used in each software implementation rather than from the underlying algorithm, and that t-SNE performs just as well at preserving global structure once it is initialized the same informative way. The practical takeaway for researchers is to treat initialization as a parameter worth checking rather than assuming a modern default automatically produces the most faithful plot, a discipline that matters just as much when tuning tree-based methods like random forests for other biological classification tasks.

Because UMAP constructs a genuine low-dimensional embedding rather than a purely visualization-oriented layout, it has found use well beyond scRNA-seq, including in population genetics. A UMAP review in population genetics recommends using it as a complement to PCA for surfacing fine-scale population structure, rather than as a wholesale replacement, since the two methods answer related but different questions about the same genotype data.

Comparing PCA, t-SNE, and UMAP and avoiding misread clusters

PCA, t-SNE, and UMAP each win on different metrics, and no single method dominates across all of them for biological data. The 10-method benchmark referenced earlier found that t-SNE achieved the strongest overall accuracy on real scRNA-seq datasets, but at a higher computing cost, while UMAP showed the greatest stability across simulation conditions with moderate accuracy and a lower computational footprint than t-SNE. Which trade-off matters most depends on dataset size and whether the goal is a quick exploratory plot or a figure meant to support a specific biological claim, a decision that sits within the broader AI and data science landscape now shaping most single-cell workflows.

The deeper risk is treating any two-dimensional embedding as a faithful map of the underlying biology. An analysis of single-cell embeddings found that common two-dimensional plots introduce significant distortion of local and global relationships relative to the original high-dimensional data, and argued that this makes such plots a weak basis for quantitative conclusions about cluster separation or cell population relationships. That finding has itself been contested: a subsequent reanalysis using neighborhood metrics rather than raw distance correlation alone found that t-SNE and UMAP do preserve meaningful local structure and correctly separate known cell types, even though they distort absolute distances. Researchers should treat the debate as a reminder that any single evaluation metric tells only part of the story, and that a plot's apparent clusters deserve validation against independent evidence such as marker gene expression before being treated as biological fact.

Building Reproducible qPCR Workflows From Experimental Design to Data Analysis

This guide walks through a practical six-step workflow covering experimental design, sample preparation, primer validation, reference gene selection, and data analysis to help generate reproducible qPCR data.
View How To Guide

Advertisement

Table 1: Practical characteristics of PCA, t-SNE, and UMAP for biological data visualization.

Several checks separate a defensible dimensionality reduction figure from one that overstates its evidence:

• Report the key hyperparameters used, including t-SNE perplexity and UMAP's number of neighbors and minimum distance, since different defaults across software packages can change the apparent structure substantially.
• Run the embedding more than once with a fixed random seed and confirm that clusters of interest reproduce, rather than relying on a single stochastic run.
• Use an informative initialization, such as a PCA-based starting layout, for both t-SNE and UMAP rather than accepting whichever default a package ships with.
• Validate any biologically important cluster against an independent line of evidence, such as marker gene expression or a supervised classifier, rather than relying on visual separation alone.
• Avoid drawing conclusions about the relative distance between distinct clusters, since none of these three methods reliably preserves quantitative distance across the full plot.

A short decision framework helps translate these characteristics into a concrete workflow:

1. Start with PCA on the full dataset to reduce noise and confirm that a small number of components captures most of the meaningful variance.
2. Choose t-SNE when the immediate goal is a clean visual separation of a moderate number of discrete populations and computational cost is not a constraint.
3. Choose UMAP when the dataset is large, computational speed matters, or the embedding needs to support downstream tasks such as clustering rather than visualization alone.
4. Set and report hyperparameters deliberately rather than accepting software defaults, and use PCA-based initialization for both nonlinear methods.
5. Treat the resulting two-dimensional plot as a hypothesis-generating tool, and confirm any cluster-level claim with a method that operates in the original high-dimensional space.

Choosing dimensionality reduction methods with confidence

Dimensionality reduction in biology will always trade some information for a plot a person can actually look at, and PCA, t-SNE, and UMAP each make that trade differently. PCA offers speed and transparency for linear structure, t-SNE offers strong local cluster separation at a real computational cost, and UMAP offers a practical balance of speed and stability that depends more on initialization and hyperparameter choices than its early reputation suggested.

None of these methods should be the last word on a biological claim. Reporting parameters transparently, checking reproducibility across runs, and validating clusters against independent evidence turn a visualization from a pretty picture into a defensible piece of scientific reasoning, which is the standard every dimensionality reduction figure in a published paper should meet.

This content includes text that has been created with the assistance of generative AI and has undergone editorial review before publishing. Technology Networks' AI policy can be found here.