UMAP (Uniform Manifold Approximation and Projection) has become the default method for visualizing single-cell data, turning datasets with thousands of gene measurements per cell into the colorful two-dimensional scatter plots that fill genomics papers today. Built on a mathematical framework rooted in topology and geometry, UMAP compresses high-dimensional information so that cells with similar gene expression profiles land near each other in a flat plot you can actually look at.1arXiv. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction Its speed, visual clarity, and ability to preserve the broad shape of the data explain why it displaced earlier methods in single-cell genomics, but the plots it produces are easier to over-interpret than most users realize.
Why Single-Cell Data Needs to Be Compressed
A typical single-cell RNA sequencing experiment measures expression levels for tens of thousands of genes in each individual cell, and modern experiments can profile hundreds of thousands or even millions of cells at once. You cannot plot a point in 20,000-dimensional space and make sense of it visually. Dimension reduction methods exist to solve this problem: they find a lower-dimensional representation, usually two dimensions, that preserves some meaningful structure from the original high-dimensional space. The goal is a plot where cells that behave similarly in the full gene-expression space appear close together on screen, letting researchers spot cell types, developmental states, and other biologically interesting groupings.
Before UMAP, most single-cell labs relied on principal component analysis (PCA) as a first pass and then t-SNE for visualization. PCA is fast and linear, which makes it useful as a preprocessing step, but its two-dimensional plots tend to smear distinct cell populations together. t-SNE was better at pulling apart clusters but came with its own frustrations: slow runtimes, difficulty preserving the arrangement of clusters relative to one another, and sensitivity to its settings. UMAP arrived as an alternative that addressed many of these pain points.
How UMAP Overtook t-SNE
A head-to-head comparison across multiple single-cell datasets found that UMAP offered the fastest runtimes, the highest reproducibility, and the most meaningful organization of cell clusters among six dimensionality reduction tools tested.2Nature Biotechnology. Dimensionality reduction for visualizing single-cell data using UMAP That combination drove rapid adoption. Where t-SNE often shuffled the relative positions of clusters randomly from run to run, UMAP preserved more of the global arrangement: if two cell types are transcriptionally related, their clusters tend to sit closer together on a UMAP plot than on a t-SNE plot. This matters when you want to see not just which groups exist, but how they relate to each other.
UMAP also handles large datasets more gracefully. Its computational cost scales better than t-SNE’s, making it practical for experiments that profile hundreds of thousands of cells without needing special hardware. Analyses of brain tissue, for instance, showed that UMAP generated what researchers described as the cleanest visual separation between cell-type clusters while maintaining global data structure.3PubMed Central. Visualizing and interpreting single-cell gene expression datasets with Similarity Weighted Nonnegative Embedding That visual clarity, combined with speed, made UMAP the standard in workflows built around Scanpy, Seurat, and other popular single-cell analysis packages.
The global-structure advantage is real but has limits. A review of advances in single-cell RNA sequencing analysis noted that UMAP better preserves global structure than t-SNE, which aids in identifying rare cell types and analyzing developmental trajectories.4Briefings in Bioinformatics. Advances and challenges in single-cell RNA sequencing data analysis: a comprehensive review Still, “better than t-SNE” is not the same as “accurate,” and the distinction matters once you start drawing biological conclusions from the shape of the plot.
What UMAP Plots Actually Show
The most common misreading of a UMAP plot is treating distances and cluster sizes as though they are quantitatively meaningful. A cluster that looks large on a UMAP plot is not necessarily more transcriptionally diverse than a cluster that looks compact. Both UMAP and t-SNE neglect the local density of data points in the original space, which often results in visualizations where densely populated groups of cells get stretched out and given more visual space than their actual transcriptional diversity warrants.5PubMed Central. Assessing Single-Cell Transcriptomic Variability through Density-Preserving Data Visualization A rare cell type with tight expression profiles can appear as a small dot, while a large, homogeneous population fans out into a sprawling island simply because there are more cells to push apart.
Distances between clusters on a UMAP plot are also unreliable as a measure of biological similarity. Two clusters sitting far apart are not necessarily more different from each other than two clusters sitting close together. UMAP prioritizes local neighborhood relationships (keeping nearby cells together) at the expense of long-range distances, so the gaps between clusters can shift dramatically depending on parameter settings and random initialization. The relative positions of well-separated groups carry some information, but the exact distances do not.
What UMAP plots are genuinely good at is showing which cells belong together. Benchmark studies using metrics like k-nearest-neighbor accuracy found that UMAP and t-SNE consistently scored above 90% for placing cells near others of the same type, far outperforming linear methods like PCA.6PLOS Computational Biology. The art of seeing the elephant in the room: 2D embeddings of single-cell data do make sense Even though UMAP does not perfectly preserve high-dimensional nearest neighbors (recall scores stayed below 40% across methods), the neighbors it does place nearby tend to be from the same cell type. In other words, UMAP is reliable for telling you “these cells are alike,” less reliable for telling you “these groups are this far apart.”
The Two Settings That Change Everything
UMAP has two key hyperparameters that shape the resulting plot: n_neighbors and min_dist. The first controls how many nearby cells UMAP considers when building its picture of local structure. A small value (say, 5 or 10) makes the algorithm focus on very fine-grained neighborhoods, often producing tighter, more fragmented clusters. A larger value (50 or more) smooths things out, emphasizing broader patterns at the expense of local detail. The second parameter, min_dist, controls how tightly UMAP is allowed to pack points together in the final plot. A value near zero produces dense, compact clusters with clear separation; larger values spread points more evenly, producing a more diffuse plot.
These settings are not cosmetic. A statistical method called scDEED was developed specifically to detect “dubious” cell embeddings, spots on a UMAP plot where cells are placed misleadingly. The researchers found that different combinations of n_neighbors and min_dist can produce dramatically different layouts from the same data, and they built tools to optimize these parameters by minimizing the number of dubiously placed cells.7Nature Communications. Statistical method scDEED for detecting dubious 2D single-cell embeddings and optimizing t-SNE and UMAP hyperparameters Most users leave both parameters at their defaults (typically n_neighbors = 15 and min_dist = 0.1 in popular packages), which works reasonably well for many datasets but is far from optimal for all of them.
If you are using UMAP for anything beyond a quick visual sanity check, experimenting with these two parameters is worth the effort. A plot that looks like two merged clusters at n_neighbors = 15 might resolve into clearly distinct groups at n_neighbors = 5, or vice versa. Neither version is “right” in an absolute sense; they reflect different scales of the data’s structure.
Where Trajectories Break Down
Single-cell biology is not only about identifying discrete cell types. Many experiments aim to capture cells in the process of transitioning from one state to another: stem cells differentiating, immune cells activating, tumors evolving. These continuous processes show up as trajectories in high-dimensional space, and researchers often want to see those trajectories in their UMAP plots. Here, UMAP’s track record is more mixed.
While UMAP can capture well-ordered biological processes like embryonic development or pancreatic differentiation reasonably well, less coherent systems may have their trajectory structures obscured in the low-dimensional representation.8PubMed Central. Pseudotime graph diffusion for post hoc visualization of inferred single-cell trajectories When trajectory inference is performed in a higher-dimensional space (typically PCA space), the two-dimensional UMAP plot may fail to reflect the inferred structure because compressing so many dimensions into two inevitably loses information. An analysis of nonnegative embedding methods noted that UMAP, while good at capturing global structure in datasets with distinct clusters, still tends to distort single-cell gene expression trajectories.9PubMed Central. Visualizing and interpreting single-cell gene expression datasets with Similarity Weighted Nonnegative Embedding
The practical consequence is that a branching trajectory visible in pseudotime analysis may appear as a smooth continuum, a broken stream, or even separate islands on a UMAP plot depending on the complexity of the underlying biology and the parameter settings. Researchers working with trajectory data increasingly treat UMAP as a rough guide and rely on dedicated trajectory inference tools (like RNA velocity, diffusion pseudotime, or graph-based methods) for their actual quantitative conclusions, using the UMAP plot only to display those results in a visually accessible way.
Handling Batch Effects
One of the trickiest aspects of multi-sample single-cell experiments is batch effects: technical variation introduced by processing samples on different days, with different reagents, or on different sequencing runs. If you generate a UMAP plot from uncorrected data, cells may cluster by batch rather than by biology, with the same cell type splitting into separate islands just because the samples were prepared a week apart.
UMAP itself does not correct batch effects. Instead, it is used as the visualization layer after batch integration methods have aligned the data. Tools like scDML, Harmony, scVI, and others operate upstream, adjusting the gene expression matrix or the shared embedding so that biological variation is preserved while technical variation is removed. UMAP then takes the corrected data and generates a plot where the researcher can visually confirm that cell types from different batches now overlap as expected.10Nature Communications. Batch alignment of single-cell transcriptomics data using deep metric learning This makes UMAP both the quality-control tool and the presentation tool for batch correction: if the corrected UMAP plot still shows batch-driven separation, the integration step needs adjustment.
Scaling to Millions of Cells
As single-cell experiments have grown from thousands to millions of cells, the computational cost of every analysis step has become a practical constraint. Standard CPU-based UMAP implementations can handle tens of thousands of cells in minutes, but datasets in the hundreds of thousands or millions push runtimes into hours. GPU-accelerated workflows have emerged as the solution.
A benchmarking study using a dataset of 1.3 million mouse brain cells found that the complete single-cell analysis pipeline, including UMAP, took over four hours on CPU but dropped to roughly 16 minutes on a single GPU and just over six minutes using eight GPUs in parallel.11bioRxiv. Accelerating single-cell genomic analysis with GPUs The UMAP step itself was one of the steps that benefited most from GPU acceleration. Frameworks like rapids-singlecell, which integrates with the widely used scverse ecosystem and operates on the standard AnnData structure, report speedups of up to several hundred-fold compared to optimized CPU baselines across standard single-cell workflows including dimensionality reduction.12arXiv. GPU-accelerated single-cell analysis at scale with rapids-singlecell Dedicated pipelines like ScaleSC have packaged these GPU-accelerated components into end-to-end solutions for large-scale single-cell data processing.13PubMed Central. ScaleSC: a superfast and scalable single-cell RNA-seq data analysis pipeline powered by GPU
For most researchers without access to dedicated GPU servers, the practical ceiling for CPU-based UMAP is roughly 200,000 to 500,000 cells before runtimes become frustrating. Cloud computing platforms that offer GPU instances have lowered the barrier, but the added cost and setup complexity mean that many labs still subsample their data before running UMAP on the full dataset.
Multi-Modal and Spatial Applications
Single-cell biology increasingly involves more than just RNA. Technologies like CITE-seq measure both gene expression and surface protein levels in the same cell, while spatial transcriptomics methods capture gene expression along with each cell’s physical location in a tissue. UMAP plays a role in each of these modalities, but the integration challenges multiply when you try to combine them.
Methods like MaxFuse tackle this by iteratively co-embedding data from different modalities, smoothing, and matching cells across datasets even when the features linking them are weak, such as integrating spatial proteomic data with single-cell sequencing data to consolidate proteomic, transcriptomic, and epigenomic information at single-cell resolution on the same tissue section.14Nature Biotechnology. Integration of spatial and single-cell data across modalities with weakly linked features In melanoma research, a multimodal integration toolkit was used to co-anchor and align transcriptomes, epitomes (protein-level CITE-seq), and morphomes (spatial protein imaging) into a shared dimensional space, with UMAP serving as the visualization layer for the combined embedding.15Cell Reports. Longitudinal multimodal single-cell analysis identifies immune-striving tumor microenvironments and therapeutic resistance in melanoma
In these multi-modal settings, UMAP is rarely the step where the hard analytical work happens. It sits downstream of the integration algorithm, providing a human-readable summary of what the integration produced. But its visual output is often the first thing collaborators and reviewers see, which gives it outsized influence on how the results are perceived.
Measuring Whether a UMAP Plot Is Actually Good
For years, UMAP quality was judged mostly by eye: does the plot look like it separates known cell types? Does it match expectations from the biology? This is surprisingly unreliable. A plot that looks clean can still misplace individual cells, and a plot that looks messy might actually be faithfully representing a noisy biological system.
Quantitative metrics have become more common. The silhouette score measures how compact and well-separated clusters are, ranging from -1 (poor separation) to 1 (clean clusters). Trajectory correlation quantifies how well the embedding dimensions track pseudotime values inferred from the data. A newer composite metric called TAES averages the silhouette score and trajectory correlation to balance evaluation of both discrete clusters and continuous developmental structures.16Scientific Reports. A comparative study of manifold learning methods for scRNA-seq with a trajectory-aware metric These metrics allow researchers to compare UMAP runs with different parameter settings, or to compare UMAP against alternative methods, using numbers rather than visual impressions.
A study that directly challenged whether 2D embeddings carry meaningful information found that UMAP and t-SNE consistently outperformed PCA on k-nearest-neighbor accuracy (above 90% versus below 62%), silhouette coefficient, and adjusted mutual information between clusters and known cell-type labels.17PLOS Computational Biology. The art of seeing the elephant in the room: 2D embeddings of single-cell data do make sense The results push back against a skeptical view that had gained traction, namely that 2D embeddings are so lossy they are effectively meaningless. They are lossy, but they carry more reliable local structure than critics suggested.
Interactive Tools for Exploring UMAP Plots
A static UMAP figure in a paper is a snapshot. Researchers increasingly share interactive versions that let collaborators hover over cells, color the plot by different genes or metadata, and zoom into regions of interest. ShinyCell, for example, generates lightweight web apps from single-cell datasets that display two UMAP plots side by side, one colored by cluster identity and the other by expression of a gene the user selects.18Bioinformatics. ShinyCell: simple and sharable visualization of single-cell gene expression data This lets a user identify a cell population expressing a particular gene in one view and immediately check its cluster identity in the other.
Platforms like BIOMEX extend this to non-bioinformaticians, offering interactive workflows for data mining of bulk and single-cell omics datasets without requiring programming skills.19Nucleic Acids Research. BIOMEX: an interactive workflow for (single cell) omics data interpretation and visualization The trend across the field is toward making UMAP plots less like finished figures and more like interactive dashboards, which helps address some of the interpretation pitfalls: when you can recolor the same plot by different variables, you are less likely to be misled by a single static view.
Alternatives and Where the Field Is Heading
UMAP is dominant but not unchallenged. An evaluation framework applied to popular dimension reduction algorithms confirmed that both t-SNE and UMAP are highly sensitive to parameter and preprocessing choices and do not perform well on global structure metrics. The study found that a newer method called PaCMAP performed well in comparison across multiple evaluation criteria.20Communications Biology. Towards a comprehensive evaluation of dimension reduction methods for transcriptomic data visualization PaCMAP works by balancing preservation of local, mid-range, and global structure explicitly, rather than relying primarily on local neighborhoods the way UMAP does.
Density-preserving variants of UMAP and t-SNE have also been developed to address the problem of misleading cluster sizes. These modified algorithms adjust the embedding so that areas of high density in the original space remain visually dense in the plot, giving a more honest representation of transcriptional diversity.21PubMed Central. Assessing Single-Cell Transcriptomic Variability through Density-Preserving Data Visualization Approaches that combine PCA with UMAP for denoising applications have also shown promise, using the PCA-UMAP manifold as the basis for diffusion-based imputation of dropout events in single-cell data.22bioRxiv. Diffusion on PCA-UMAP manifold captures a well-balance of local, global, and continuum structure to denoise single-cell RNA sequencing data
Supervised and semi-supervised approaches represent another direction. Methods like scArches use transfer learning to map new query datasets onto existing reference atlases, enabling cell-type annotation by projecting cells into a shared space trained on curated data.23Nature Biotechnology. Mapping single-cell data to reference atlases by transfer learning In these workflows, UMAP often serves as the final visualization step after the heavy lifting of alignment and annotation is done by the underlying model. Meanwhile, deterministic methods like LDA have been proposed as alternatives that avoid UMAP’s stochastic nature entirely, producing identical results each time they run and remaining stable even with low cell counts.24PubMed Central. Supervised dimensionality reduction for exploration of single-cell data by HSS-LDA These methods trade some of UMAP’s visual separation quality for guaranteed reproducibility, which matters in clinical or regulatory settings where a result needs to be exactly replicated.
None of these alternatives has displaced UMAP from its central role, and it is unlikely any single method will. The more probable future is one where researchers routinely generate multiple types of embeddings and compare them, treating UMAP as one lens among several rather than the definitive picture of their data. The plots are powerful and useful, but they are summaries, and the best practice is to never let a two-dimensional picture be the sole basis for a biological conclusion drawn from a 20,000-dimensional dataset.