# Interpreting Single-Cell UMAP Plots: What the Distances and Densities Really Mean ## Key Takeaways - UMAP embeddings prioritize preserving local neighborhood structure over global distances, meaning spatial proximity in the plot does not directly correlate with transcriptional similarity or biological distance between cells. - The visual density of clusters in a UMAP plot is influenced by the `min_dist` parameter and local data structure, not solely by cell abundance; actual cell counts from the count matrix are required for accurate proportion estimations. - Continuous gradients or trajectories observed in UMAP plots are not definitive evidence of biological differentiation; dedicated trajectory inference methods operating on high-dimensional data are necessary for robust lineage analysis. - UMAP is a stochastic algorithm sensitive to hyperparameters like `n_neighbors` and `min_dist`, necessitating the use of fixed random seeds and comprehensive documentation of all parameters for reproducibility and robust interpretation. - Biological interpretation of UMAP clusters must be rigorously validated through differential gene expression analysis to identify statistically significant marker genes, rather than relying solely on visual separation or grouping. - Claims derived from UMAP plots should pass a four-gate validation framework: reproducible cluster existence in the underlying data, robust marker gene support, quantitative proportion verification, and biological plausibility supported by independent evidence. --- Uniform Manifold Approximation and Projection (UMAP) is a dimensionality reduction technique widely used in single-cell RNA sequencing (scRNA-seq) analysis to visualize high-dimensional transcriptomic data in two dimensions. The practical problem is that researchers frequently interpret spatial proximity in UMAP plots as evidence of cellular relatedness, and interpret density as evidence of cell abundance, when neither interpretation is mathematically guaranteed. This article explains what UMAP embeddings actually preserve, what they distort, and how to build an interpretation workflow that prevents false biological conclusions. The guidance applies to researchers, laboratory professionals, and biology students who use scRNA-seq, single-nucleus RNA-seq, and related single-cell assays in their work. ## What UMAP Does and Does Not Preserve UMAP is a manifold learning technique that constructs a high-dimensional graph representation of the data and then projects that graph into a low-dimensional space. The algorithm attempts to preserve both local neighborhood structure and some global structure, but the tradeoffs between these goals are controlled by hyperparameters that the user selects. Understanding these hyperparameters is essential before any biological interpretation begins. ### The Mathematical Basis of UMAP Embeddings UMAP operates in two phases. First, it builds a fuzzy topological representation of the high-dimensional data by connecting each cell to its nearest neighbors based on a distance metric. Second, it optimizes a low-dimensional representation that minimizes the difference between the high-dimensional fuzzy graph and a low-dimensional fuzzy graph. The optimization process balances two forces: attractive forces that pull connected points together and repulsive forces that push disconnected points apart. The key consequence is that UMAP preserves local structure preferentially over global structure. Cells that are close in the original high-dimensional space tend to remain close in the embedding, but cells that are far apart in the original space may appear close, far, or anywhere in between depending on the hyperparameters and the data distribution. This means that the absence of distance between two clusters in a UMAP plot does not prove that the underlying cell states are transcriptionally similar. The [Bioconductor project](https://bioconductor.org/) provides official documentation for many single-cell analysis packages that implement UMAP and related visualization methods. These packages include detailed workflow guidance that emphasizes the importance of understanding the algorithm's parameters before interpreting the output. ### Hyperparameters That Change the Visual Output The most influential hyperparameters in UMAP are n_neighbors and min_dist. The n_neighbors parameter controls how many neighboring points are considered when constructing the high-dimensional graph. A small value, such as 5, focuses on very local structure and tends to produce more fragmented clusters. A large value, such as 50, considers broader neighborhoods and tends to produce more connected, larger clusters that may merge distinct cell states. The min_dist parameter controls how tightly points are allowed to pack together in the low-dimensional embedding. A small min_dist value, such as 0.01, produces dense, well-separated clusters. A large min_dist value, such as 0.5, produces more diffuse clusters with less separation. Neither setting is biologically correct or incorrect, but each produces a different visual impression of the same underlying data. The [Galaxy Training Network](https://training.galaxyproject.org/) offers accessible tutorials on single-cell analysis workflows that include practical guidance on parameter selection and the interpretation of embedding outputs. These tutorials emphasize that parameter choices should be documented and reported alongside any biological conclusions. ### Why Distances in UMAP Are Not Biological Distances The distances between points in a UMAP plot are not proportional to transcriptional distances in the original high-dimensional space. UMAP does not preserve global distances, and the scale of the axes in a UMAP plot has no biological meaning. A cluster that appears isolated in the embedding may be transcriptionally similar to another cluster that appears distant, and two clusters that appear adjacent may be transcriptionally distinct. This distortion occurs because UMAP optimizes for a balance between local and global structure that is determined by the hyperparameters. The algorithm does not attempt to preserve the actual distances between all pairs of cells. Instead, it attempts to preserve the connectivity structure of the neighborhood graph. This is a fundamentally different objective, and it means that the visual arrangement of clusters should be treated as a schematic representation instead of a quantitative map. The [EMBL-EBI Training](https://www.ebi.ac.uk/training) portal provides structured learning pathways for bioinformatics that cover dimensionality reduction and its limitations. These training materials explicitly warn against overinterpreting distances in low-dimensional embeddings and recommend complementary quantitative analyses. ## At a Glance: UMAP Interpretation Decision Table The following table summarizes common interpretation scenarios and the appropriate analytical response. Use this table as a quick reference when evaluating UMAP plots in your own analyses. | Observation in UMAP Plot | Common Misinterpretation | Recommended Analytical Response | | --- | --- | --- | | Two clusters appear adjacent or overlapping | Cells are transcriptionally similar or related | Verify with differential expression analysis and marker gene assessment before concluding similarity | | A cluster appears dense and compact | Cells are highly abundant or represent a dominant cell type | Check actual cell counts in the cluster and compare with proportion calculations from the count matrix | | A cluster appears isolated and distant from others | Cells are transcriptionally unique or unrelated | Confirm with pairwise distance calculations in the original high-dimensional space or with cluster stability analysis | | Cells form a continuous gradient or trajectory | Cells represent a differentiation continuum | Test with pseudotime or trajectory inference methods that use the high-dimensional data, not the UMAP coordinates | | Two separate UMAP runs produce different layouts | The data are inconsistent or the analysis is unreliable | Recognize that UMAP is stochastic and that hyperparameter choices affect the output, then document parameters and use a fixed random seed | ## Core Principles for Cautious UMAP Interpretation The interpretation of UMAP plots requires a disciplined approach that treats the embedding as a visualization tool instead of as a quantitative result. The following principles should guide every analysis. ### Principle 1: Cluster Identity Requires Marker Gene Validation A cluster in a UMAP plot is a visual grouping of cells that the algorithm placed near each other in the low-dimensional space. This grouping is not proof that the cells share a common identity. The only reliable way to assign biological identity to a cluster is to examine the expression of known marker genes within that cluster and compare it with the expression in other clusters. For example, a study of infectious keratitis used single-cell transcriptome sequencing on corneal tissues from patients with Acanthamoeba keratitis, fungal keratitis, and bacterial keratitis. The UMAP plot demonstrated 13 major cell clusters in infectious keratitis, and the researchers used the embedding to visualize the cell subsets and their proportions within different infectious keratitis conditions. The biological conclusions about neutrophil proportions and CXCL pathway signaling were based on quantitative analyses of the expression data, not on the visual arrangement of the UMAP plot alone. This study is reported in [EBioMedicine](https://pubmed.ncbi.nlm.nih.gov/40222104) and illustrates the correct workflow of using UMAP for visualization while relying on quantitative methods for biological inference. ### Principle 2: Density Does Not Equal Abundance The density of points in a UMAP plot is influenced by the min_dist parameter and by the local structure of the data. A region of the plot that appears densely packed may contain many cells, but it may also contain cells that are transcriptionally similar to each other and therefore placed close together by the algorithm. Conversely, a sparse region may contain few cells, or it may contain cells that are transcriptionally diverse and therefore spread apart. To determine the actual abundance of a cell type, count the number of cells assigned to each cluster in the count matrix and calculate proportions. Do not estimate proportions from the visual density of the UMAP plot. The [nf-core documentation](https://nf-co.re/docs) provides standards for reproducible single-cell analysis pipelines that include explicit steps for calculating and reporting cell proportions from the data instead of from visual inspection. ### Principle 3: Trajectories Require Dedicated Inference Methods A continuous gradient of cells in a UMAP plot is often interpreted as evidence of a differentiation trajectory or a developmental continuum. This interpretation is risky because UMAP can produce continuous-looking arrangements of cells that are not connected by any biological process. The algorithm places cells based on transcriptional similarity, and a gradient in the embedding may simply reflect a gradual change in gene expression that is not related to differentiation. Trajectory inference methods such as pseudotime analysis use the high-dimensional expression data to construct lineage relationships. These methods should be applied to the original data, not to the UMAP coordinates. The [Bioconductor project](https://bioconductor.org/) hosts several trajectory inference packages with documentation that explains the assumptions and limitations of each method. ### Principle 4: UMAP Is Stochastic and Parameter-Dependent UMAP is a stochastic algorithm, which means that running it multiple times on the same data can produce different embeddings. The differences are usually minor when the data have clear cluster structure, but they can be substantial when the data are noisy or when the clusters are poorly separated. The hyperparameters n_neighbors and min_dist also change the output, sometimes dramatically. For reproducible analyses, set a fixed random seed and document all hyperparameters in the analysis report. The [The Carpentries Lessons](https://carpentries.org/lessons) include training on reproducible research practices that emphasize the importance of documenting computational parameters and using version control for analysis code. ## Practical Workflow for UMAP Generation and Interpretation The following workflow describes the steps required to generate a UMAP plot that supports cautious interpretation. Each step includes specific decisions that affect the final output. ### Step 1: Quality Control and Data Filtering Before any dimensionality reduction, the single-cell data must pass quality control. This includes filtering out low-quality cells based on metrics such as the number of detected genes, the total number of counts, and the percentage of mitochondrial reads. The specific thresholds depend on the tissue type and the experimental protocol, and they should be determined from the data distribution instead of applied arbitrarily. The [Galaxy Training Network](https://training.galaxyproject.org/) provides tutorials on single-cell quality control that demonstrate how to evaluate these metrics and set appropriate thresholds. Quality control is a critical step because low-quality cells can create spurious clusters in the UMAP embedding that do not represent real biological states. ### Step 2: Normalization and Feature Selection After quality control, the count data must be normalized to account for differences in sequencing depth between cells. Several normalization methods are available, and the choice of method affects the downstream analysis. After normalization, highly variable genes are selected for the dimensionality reduction step. The number of highly variable genes is a user choice that affects the resolution of the embedding. The [Bioconductor project](https://bioconductor.org/) provides documentation for normalization and feature selection methods that are widely used in single-cell analysis. These methods include detailed explanations of the assumptions behind each approach. ### Step 3: Dimensionality Reduction with PCA Before UMAP, the data are typically reduced to a set of principal components using PCA. The number of principal components retained is a user choice that affects the UMAP output. Retaining too few components discards biological signal, while retaining too many components introduces noise. The choice is often made by examining the elbow in the variance explained plot or by using more formal statistical approaches. The [EMBL-EBI Training](https://www.ebi.ac.uk/training) materials cover PCA and its role in single-cell analysis workflows. The number of principal components should be documented and reported alongside the UMAP parameters. ### Step 4: UMAP Parameter Selection Select the n_neighbors and min_dist parameters based on the goals of the visualization. For exploratory analysis, a range of parameter values should be tested to understand how the embedding changes. For final figures, the parameters should be chosen to match the biological question. For example, if the goal is to identify rare cell populations, a smaller n_neighbors value may be appropriate. If the goal is to visualize broad cell type relationships, a larger value may be better. The [nf-core documentation](https://nf-co.re/docs) describes standard pipeline configurations for single-cell analysis that include default UMAP parameters. These defaults are reasonable starting points, but they should be adjusted based on the data and the biological question. ### Step 5: Cluster Identification and Marker Validation After generating the UMAP embedding, clustering is performed on the PCA-reduced data, not on the UMAP coordinates. The clustering algorithm groups cells based on transcriptional similarity, and the UMAP plot is used to visualize the resulting clusters. Each cluster should be annotated based on marker gene expression, and the annotation should be validated using differential expression analysis. A study of intraepithelial lymphocytes in chickens challenged with Salmonella Enteritidis used single-cell RNA sequencing to identify clusters of progenitor T cells and innate-like cytotoxic T cells that expanded in response to infection. The UMAP embedding was used to visualize these clusters, but the biological conclusions were based on quantitative analysis of the expression data. This study is reported in [Frontiers in Immunology](https://doi.org/10.3389/fimmu.2026.1842283) and demonstrates the correct workflow of using UMAP for visualization while relying on quantitative methods for biological inference. ### Step 6: Documentation and Reporting Every UMAP plot in a publication or report should include the following information: the software version, the random seed, the n_neighbors value, the min_dist value, the number of principal components used, and the normalization method. This documentation allows other researchers to reproduce the embedding and to assess the robustness of the visual conclusions. The [The Carpentries Lessons](https://carpentries.org/lessons) provide training on reproducible research practices that emphasize the importance of documenting computational parameters. The [nf-core documentation](https://nf-co.re/docs) also emphasizes reproducibility as a core principle of community pipeline development. ## Options and Tradeoffs in UMAP Visualization Several alternatives to UMAP exist for visualizing single-cell data, and each has distinct properties that affect interpretation. ### t-SNE as an Alternative t-distributed stochastic neighbor embedding (t-SNE) is another popular dimensionality reduction method. t-SNE focuses almost exclusively on preserving local structure and is known to produce embeddings where the global arrangement of clusters is not meaningful. t-SNE also has a perplexity parameter that strongly influences the output. For these reasons, t-SNE plots should be interpreted with even more caution than UMAP plots regarding global distances. The [PLoS Computational Biology](https://doi.org/10.1371/journal.pcbi.1014418) article on the CAdir clustering algorithm notes that embeddings produced by UMAP or t-SNE can distort the true cluster structure and are known to produce radically different embeddings depending on the chosen hyperparameters. This observation applies to both methods and reinforces the need for quantitative validation of visual impressions. ### Correspondence Analysis-Based Approaches The CAdir algorithm uses correspondence analysis to cluster cells and genes jointly, and it provides diagnostic plots that help assess clustering quality. This approach avoids some of the interpretability problems of UMAP because it does not rely on a low-dimensional embedding for cluster validation. The algorithm can infer the number of clusters in the data and provides cluster-specific genes, which supports biological interpretation. The CAdir software is available from the project repository described in the [PLoS Computational Biology](https://doi.org/10.1371/journal.pcbi.1014418) article. ### Spatial Transcriptomics Integration When spatial transcriptomics data are available, the spatial coordinates of cells can be used to validate interpretations derived from UMAP embeddings. A study of non-small cell lung cancer integrated single-cell RNA-seq with spatial transcriptomics to confirm the expression of a 10-gene signature in tumor versus adjacent tissues. The spatial data provided an independent validation of the cell type assignments derived from the single-cell analysis. This study is reported in [Translational Oncology](https://doi.org/10.1016/j.tranon.2026.102853) and illustrates the value of integrating multiple data modalities. ## Observations and Measurements That Support Interpretation The interpretation of UMAP plots should be supported by quantitative measurements that are independent of the embedding. The following measurements are essential. ### Cell Counts and Proportions For each cluster in the UMAP plot, count the number of cells and calculate the proportion of the total. These proportions should be compared across conditions to identify changes in cell type abundance. For example, the infectious keratitis study found that the neutrophil proportion in the Acanthamoeba keratitis group was markedly reduced compared with the non-Acanthamoeba group. This conclusion was based on quantitative cell proportion analysis, not on the visual density of the UMAP plot. ### Differential Expression Statistics For each cluster, perform differential expression analysis to identify marker genes. The statistical significance and effect size of the expression differences should be reported. A cluster that appears distinct in the UMAP plot but has no statistically significant marker genes may be an artifact of the embedding instead of a real biological state. ### Cluster Stability Metrics Cluster stability measures how consistently cells are assigned to the same cluster across multiple runs of the clustering algorithm or across subsamples of the data. Unstable clusters should be interpreted with caution. The CAdir algorithm provides diagnostic plots that help assess the quality of cluster assignments, as described in the [PLoS Computational Biology](https://doi.org/10.1371/journal.pcbi.1014418) article. ### Pseudotime and Trajectory Statistics When trajectory inference is performed, the pseudotime values should be calculated from the high-dimensional data using dedicated methods. The resulting trajectories should be validated by examining the expression of known marker genes along the trajectory. The UMAP plot can be used to visualize the trajectory, but the trajectory itself should not be inferred from the visual arrangement of cells in the embedding. ## Records and Documentation for Reproducible UMAP Analysis Maintaining detailed records of the analysis is essential for reproducibility and for defending biological conclusions. The following records should be kept for every single-cell analysis project. ### Analysis Log The analysis log should record every step of the computational workflow, including software versions, parameter values, and the dates of each analysis run. This log should be maintained in a version-controlled repository. The [The Carpentries Lessons](https://carpentries.org/lessons) provide training on using Git for version control, and the [nf-core documentation](https://nf-co.re/docs) describes standards for reproducible pipeline configuration. ### Parameter Documentation For each UMAP plot, record the n_neighbors value, the min_dist value, the random seed, the number of principal components, and the normalization method. This information should be included in the figure legend or in a supplementary table. Without this documentation, the embedding cannot be reproduced and the visual conclusions cannot be assessed. ### Cluster Annotation Records For each cluster, record the marker genes used for annotation, the differential expression statistics, and the evidence supporting the biological interpretation. This record should be updated as new analyses are performed or as new biological knowledge becomes available. ### Data Availability The raw count matrix, the metadata, and the analysis code should be deposited in a public repository to allow other researchers to reproduce the analysis. The [NCBI](https://www.ncbi.nlm.nih.gov/) provides databases for depositing sequencing data, and the [EMBL-EBI Training](https://www.ebi.ac.uk/training) portal provides guidance on data sharing best practices. ## Common Failure Patterns in UMAP Interpretation The following failure patterns are frequently observed in published analyses and in laboratory discussions. Recognizing these patterns can prevent incorrect biological conclusions. ### Failure Pattern 1: Treating UMAP Distances as Quantitative Researchers sometimes measure distances between clusters in a UMAP plot and use these measurements to make claims about the degree of similarity between cell types. This practice is invalid because UMAP does not preserve global distances. The distance between two clusters in the embedding is influenced by the hyperparameters and by the local structure of the data, and it has no direct biological interpretation. ### Failure Pattern 2: Inferring Lineage Relationships from Visual Gradients A continuous gradient of cells in a UMAP plot is sometimes interpreted as evidence of a differentiation trajectory. This interpretation is risky because UMAP can produce gradients that do not correspond to any biological process. Trajectory inference should be performed using dedicated methods that operate on the high-dimensional data. ### Failure Pattern 3: Estimating Proportions from Visual Density Researchers sometimes estimate the relative abundance of cell types by looking at the density of points in a UMAP plot. This practice is invalid because the density of points is influenced by the min_dist parameter and by the local structure of the data. Cell proportions should be calculated from the count matrix. ### Failure Pattern 4: Ignoring Parameter Sensitivity Researchers sometimes generate a single UMAP plot with default parameters and treat the resulting layout as the definitive representation of the data. This practice ignores the sensitivity of UMAP to hyperparameter choices. A robust analysis should include a parameter sensitivity assessment that demonstrates that the biological conclusions are stable across a range of parameter values. ### Failure Pattern 5: Overinterpreting Cluster Separation Researchers sometimes interpret well-separated clusters in a UMAP plot as evidence of distinct cell types. While separation can indicate transcriptional differences, it can also result from technical artifacts or from the choice of hyperparameters. Cluster identity should always be validated with marker gene expression and differential expression analysis. ### Failure Pattern 6: Using UMAP for Quantitative Comparisons Researchers sometimes use the coordinates of cells in a UMAP plot as input for downstream quantitative analyses, such as calculating distances between cells or comparing the positions of cells across conditions. This practice is invalid because the UMAP coordinates are not quantitative representations of the data. Downstream analyses should use the original high-dimensional data or the PCA-reduced representation. ## Limitations of UMAP and Complementary Approaches UMAP is a powerful visualization tool, but it has inherent limitations that cannot be overcome by parameter tuning alone. Understanding these limitations is essential for designing analyses that produce reliable biological conclusions. ### Loss of Quantitative Information UMAP reduces the data to two dimensions, which necessarily discards information. The two-dimensional representation cannot capture all of the structure in the high-dimensional data, and the visual impression of the embedding may not reflect the true relationships between cells. Quantitative analyses should always be performed on the high-dimensional data. ### Sensitivity to Data Quality UMAP is sensitive to the quality of the input data. Low-quality cells, batch effects, and technical artifacts can create spurious structure in the embedding. Quality control and data integration are essential preprocessing steps that affect the reliability of the UMAP output. ### Lack of Statistical Inference UMAP does not provide statistical measures of uncertainty. The embedding is a deterministic function of the input data and the hyperparameters, but it does not come with confidence intervals or significance tests. Biological conclusions should be supported by statistical analyses that are independent of the embedding. ### Complementary Approaches Several complementary approaches can support the interpretation of UMAP plots. Clustering algorithms that provide diagnostic plots, such as CAdir, can help assess the quality of cluster assignments. Trajectory inference methods can identify lineage relationships. Spatial transcriptomics can validate cell type assignments. Integration of multiple approaches provides stronger evidence than any single method. The [Bioconductor project](https://bioconductor.org/) hosts a wide range of packages for these complementary analyses, and the [Galaxy Training Network](https://training.galaxyproject.org/) provides tutorials that demonstrate how to combine these approaches in a complete workflow. ## Quality Controls and Validation Steps The following quality controls should be applied at each stage of the single-cell analysis workflow to ensure that the UMAP plot supports reliable interpretation. ### Quality Control Before Dimensionality Reduction Before generating a UMAP plot, verify that the quality control filtering was appropriate. Examine the distribution of quality metrics and confirm that the filtering thresholds were chosen based on the data. Low-quality cells that pass the filtering can create spurious clusters in the embedding. ### Validation of Cluster Markers For each cluster in the UMAP plot, validate the marker genes using differential expression analysis. The marker genes should be expressed in a substantial fraction of the cells in the cluster and should show statistically significant upregulation compared with other clusters. A cluster without validated markers should be treated as provisional. ### Parameter Sensitivity Assessment Generate UMAP plots with a range of n_neighbors and min_dist values and compare the resulting layouts. If the biological conclusions change substantially across parameter values, the conclusions are not robust and should be treated with caution. The parameter sensitivity assessment should be documented in the analysis report. ### Independent Validation When possible, validate the biological conclusions using independent data or methods. Spatial transcriptomics can confirm the spatial distribution of cell types. Flow cytometry can confirm the proportions of cell populations. Functional experiments can test the predicted roles of specific cell types. The [Frontiers in Cell and Developmental Biology](https://doi.org/10.3389/fcell.2026.1798845) study on immune-related adverse events used machine learning models and experimental validation to confirm the biological significance of the single-cell findings. ## Safety and Regulatory Context for Single-Cell Research Single-cell research involving human or animal tissues is subject to ethical and regulatory requirements that vary by jurisdiction. Researchers should ensure that their studies have the necessary approvals from institutional review boards or animal care committees before collecting samples. ### Data Privacy and Sharing Single-cell data derived from human subjects may contain sensitive information. Data sharing should comply with applicable privacy regulations and with the terms of the informed consent obtained from participants. The [NCBI](https://www.ncbi.nlm.nih.gov/) provides controlled-access databases for human data that require appropriate authorization. ### Reproducibility Standards Funding agencies and journals increasingly require that single-cell analyses be reproducible. This requirement includes depositing the raw data, the processed data, and the analysis code in public repositories. The [nf-core documentation](https://nf-co.re/docs) describes standards for reproducible pipeline development, and the [EMBL-EBI Training](https://www.ebi.ac.uk/training) portal provides guidance on data management and sharing. ### Professional Escalation Criteria When the interpretation of a UMAP plot has significant consequences, such as a clinical decision or a regulatory submission, the analysis should be reviewed by a qualified bioinformatician or biostatistician. The following situations warrant professional escalation: - The biological conclusions depend on the visual arrangement of clusters in a UMAP plot without quantitative validation - The UMAP embedding changes substantially across parameter values and the conclusions are not robust - The cluster annotations are based on incomplete marker gene evidence - The analysis is part of a regulatory submission or a clinical decision - The data are derived from human subjects and the conclusions have health implications ## A Decision Framework for Cluster-Level UMAP Claims The visual properties of a UMAP embedding should never be the final evidence for a biological claim. This section provides a structured decision framework that forces explicit justification for each interpretation step, a record system for tracking the evidence behind every cluster-level claim, and a troubleshooting method for embeddings that produce unexpected or unstable layouts. The framework is designed to be applied before any biological conclusion is written into a manuscript, report, or presentation. ### The Four-Gate Claim Validation Framework Every biological claim derived from a UMAP plot should pass through four sequential gates before it is accepted. Each gate requires a specific type of evidence that is independent of the embedding itself. If a claim fails at any gate, it must be revised or abandoned. **Gate 1: Cluster Existence.** The first gate asks whether the cluster is a reproducible grouping of cells instead of an artifact of the embedding. To pass this gate, the cluster must be recovered by the clustering algorithm applied to the PCA-reduced data, not to the UMAP coordinates. The cluster should also appear in multiple UMAP runs with different random seeds. The [PLoS Computational Biology](https://doi.org/10.1371/journal.pcbi.1014418) article on the CAdir algorithm notes that embeddings can distort the true cluster structure and produce radically different layouts depending on hyperparameters. A cluster that appears in one embedding but disappears when the random seed changes fails this gate. **Gate 2: Marker Gene Support.** The second gate asks whether the cells in the cluster share a coherent transcriptional identity. To pass this gate, the cluster must show statistically significant upregulation of known marker genes compared with other clusters. The differential expression analysis must be performed on the count matrix, not on the UMAP coordinates. A cluster that appears visually distinct but has no significant marker genes fails this gate and should be treated as a provisional grouping. **Gate 3: Quantitative Proportion Verification.** The third gate asks whether the apparent density or size of the cluster in the UMAP plot matches the actual cell counts. To pass this gate, the number of cells in the cluster must be counted from the metadata and compared with the visual impression. A cluster that appears large and dense but contains only a small fraction of the total cells fails this gate. The infectious keratitis study in [EBioMedicine](https://pubmed.ncbi.nlm.nih.gov/40222104) demonstrated this principle by calculating cell subsets and proportions within different keratitis conditions instead of relying on visual density. **Gate 4: Biological Plausibility.** The fourth gate asks whether the proposed biological interpretation is consistent with existing knowledge and with other data modalities. To pass this gate, the interpretation should be supported by independent evidence such as spatial transcriptomics, flow cytometry, or functional experiments. A claim that passes the first three gates but contradicts established biology should be re-examined instead of accepted. ### Applying the Framework to Common Claim Types The four gates apply differently depending on the type of claim being made. The following claim types are the most common in single-cell research. **Cell Type Identification Claims.** A claim that a cluster represents a specific cell type must pass all four gates. The marker gene evidence in Gate 2 is the most critical for this claim type. The chicken intraepithelial lymphocyte study in [Frontiers in Immunology](https://doi.org/10.3389/fimmu.2026.1842283) identified progenitor T cells and innate-like cytotoxic T cells using single-cell RNA sequencing. The biological conclusions about these cell types were based on quantitative expression analysis, not on the visual arrangement of the UMAP embedding. **Abundance Comparison Claims.** A claim that one condition has more of a particular cell type than another condition must pass Gate 3. The proportions must be calculated from the count matrix and compared using appropriate statistical tests. The visual density of the UMAP plot is not an acceptable substitute for this calculation. The [Frontiers in Cell and Developmental Biology](https://doi.org/10.3389/fcell.2026.1798845) study on immune-related adverse events used immune infiltration signatures and machine learning models to predict outcomes, demonstrating that quantitative approaches provide stronger evidence than visual inspection. **Trajectory and Lineage Claims.** A claim that cells form a differentiation continuum must pass Gate 4 with dedicated trajectory inference methods applied to the high-dimensional data. The UMAP gradient is not evidence of lineage. The porcine kidney study in [Frontiers in Pharmacology](https://doi.org/10.3389/fphar.2026.1841271) used pseudotime inference and PT-state stratification to examine injury-associated proximal tubule states, showing that trajectory claims require dedicated computational methods. **Cell-Cell Communication Claims.** A claim about signaling between cell types must be supported by dedicated communication analysis tools that operate on the expression data. The UMAP proximity of two clusters does not demonstrate that the cells communicate. The [Translational Oncology](https://doi.org/10.1016/j.tranon.2026.102853) study on non-small cell lung cancer used cell-cell communication analysis to highlight MIF pathway signaling in the tumor microenvironment, with the analysis based on expression data instead of embedding geometry. ### The Cluster Evidence Record System A structured record system prevents the loss of evidence and makes the reasoning behind each claim auditable. The following record format should be maintained for every cluster that appears in a publication or report. **Cluster Evidence Record Fields.** Each record should contain the cluster identifier, the number of cells, the proportion of total cells, the marker genes used for annotation, the differential expression statistics, the clustering algorithm and parameters, the UMAP parameters used for visualization, and the date of the analysis. The record should also note which of the four gates the cluster has passed and which gates remain unverified. **Version Control for Records.** The records should be maintained in a version-controlled repository alongside the analysis code. The [The Carpentries Lessons](https://carpentries.org/lessons) provide training on using Git for version control, and the [nf-core documentation](https://nf-co.re/docs) describes standards for reproducible pipeline configuration. Version control ensures that the evidence record can be traced back to the exact code and parameters that produced it. **Record Review Schedule.** The evidence records should be reviewed whenever the analysis is updated, when new biological knowledge becomes available, or when a manuscript is revised. A cluster that passed all four gates in an initial analysis may fail a gate after the data are updated or after the clustering parameters are changed. ### Troubleshooting Unexpected UMAP Layouts When a UMAP embedding produces an unexpected layout, the cause is usually one of several identifiable problems. The following troubleshooting method addresses the most common causes. **Problem 1: Batch Effects Creating Artificial Separation.** If cells from the same biological condition separate into distinct clusters, batch effects are the likely cause. The solution is to apply data integration methods before dimensionality reduction. After integration, the UMAP plot should be examined to confirm that cells from different batches are mixed within clusters. The [Bioconductor project](https://bioconductor.org/) hosts several integration methods with documentation that explains their assumptions and limitations. **Problem 2: Excessive Noise Producing Fragmented Clusters.** If the embedding shows many small, fragmented clusters with no clear structure, the data may contain excessive noise. The solution is to increase the number of principal components retained, increase the n_neighbors value, or apply additional filtering to remove low-quality cells. The [Galaxy Training Network](https://training.galaxyproject.org/) provides tutorials on single-cell quality control that demonstrate how to evaluate these metrics. **Problem 3: Overly Compact Clusters Hiding Internal Structure.** If the embedding shows very dense, compact clusters with no visible internal structure, the min_dist parameter may be too small. The solution is to increase min_dist to allow cells within a cluster to spread apart, revealing potential substructure. A range of min_dist values should be tested to determine whether the internal structure is biologically meaningful. **Problem 4: Parameter Sensitivity Producing Unstable Layouts.** If the embedding changes substantially across a range of n_neighbors values, the data may not have clear cluster structure. The solution is to perform a parameter sensitivity assessment and document the results. If the biological conclusions change across parameter values, the conclusions are not robust and should be treated with caution. **Problem 5: Rare Populations Hidden by Dominant Cell Types.** If rare cell populations are not visible in the embedding, the dominant cell types may be overwhelming the structure. The solution is to use a smaller n_neighbors value to focus on local structure, or to subset the data to remove dominant populations and re-run the embedding. The [EMBL-EBI Training](https://www.ebi.ac.uk/training) portal provides learning pathways that cover strategies for identifying rare populations in single-cell data. ### Escalation Criteria for Problematic Embeddings Some embedding problems cannot be resolved by parameter tuning alone and require professional escalation. The following situations warrant consultation with a bioinformatician or biostatistician: - The embedding produces radically different layouts across a wide range of parameter values and no stable structure emerges - The clustering results cannot be validated by any marker genes despite extensive testing - The biological conclusions depend on subtle visual features that cannot be reproduced across multiple runs - The analysis is part of a regulatory submission or a clinical decision where the consequences of error are high - The data contain severe batch effects that cannot be resolved by standard integration methods The [NCBI](https://www.ncbi.nlm.nih.gov/) provides access to databases and analysis services that can support more advanced troubleshooting, and the [EMBL-EBI Training](https://www.ebi.ac.uk/training) portal offers structured learning pathways for researchers who need to develop deeper expertise in single-cell analysis. ## Frequently Asked Questions ### Why do cells that appear close together in a UMAP plot sometimes have very different gene expression profiles? UMAP preserves local neighborhood structure preferentially over global structure, but the local neighborhoods are defined in the high-dimensional space using a distance metric. Cells that are placed close together in the embedding are connected in the high-dimensional neighborhood graph, but the degree of transcriptional similarity can vary. The visual proximity in the UMAP plot does not guarantee that the cells are transcriptionally identical or even highly similar. Always validate the relationship between cells using quantitative measures such as pairwise distances in the high-dimensional space or differential expression analysis. ### Can I use the distance between clusters in a UMAP plot to measure how different two cell types are? No. UMAP does not preserve global distances, and the scale of the axes in a UMAP plot has no biological meaning. The distance between two clusters in the embedding is influenced by the hyperparameters and by the local structure of the data. To measure the transcriptional difference between two cell types, use quantitative methods such as differential expression analysis or distance calculations in the original high-dimensional space. ### How should I choose the n_neighbors and min_dist parameters for my UMAP plot? The choice of parameters depends on the biological question. For exploratory analysis, test a range of parameter values and compare the resulting layouts. For identifying rare cell populations, a smaller n_neighbors value may be appropriate. For visualizing broad cell type relationships, a larger value may be better. The min_dist parameter controls the density of the embedding, with smaller values producing more compact clusters. Document the chosen parameters and include them in the figure legend. ### Why does my UMAP plot look different every time I run it? UMAP is a stochastic algorithm, which means that it uses random initialization and random sampling during the optimization process. Running UMAP multiple times on the same data can produce different embeddings. To make the analysis reproducible, set a fixed random seed and document it in the analysis report. The differences between runs are usually minor when the data have clear cluster structure, but they can be substantial when the data are noisy. ### How do I know if a cluster in my UMAP plot represents a real cell type? A cluster should be validated using marker gene expression and differential expression analysis. Identify marker genes that are significantly upregulated in the cluster compared with other clusters, and verify that these markers are consistent with known biology. A cluster without validated markers should be treated as provisional. Cluster stability analysis can also help determine whether the cluster is consistently reproduced across multiple runs of the clustering algorithm. ### Can I infer cell lineage relationships from the arrangement of cells in a UMAP plot? No. A continuous gradient of cells in a UMAP plot does not prove that the cells are connected by a differentiation process. UMAP can produce gradients that reflect gradual changes in gene expression that are not related to lineage. Trajectory inference methods such as pseudotime analysis should be used to construct lineage relationships, and these methods should be applied to the high-dimensional data instead of to the UMAP coordinates. ### What should I report when I publish a UMAP plot? Report the software version, the random seed, the n_neighbors value, the min_dist value, the number of principal components used, and the normalization method. This information allows other researchers to reproduce the embedding and to assess the robustness of the visual conclusions. The [nf-core documentation](https://nf-co.re/docs) provides standards for reproducible pipeline configuration that can guide this reporting. ### How do I handle batch effects when generating UMAP plots? Batch effects can create spurious structure in UMAP embeddings. Data integration methods should be applied to remove batch effects before dimensionality reduction. The choice of integration method depends on the experimental design and the severity of the batch effects. After integration, the UMAP plot should be examined to confirm that cells from different batches are mixed within clusters instead of separated by batch. The [Bioconductor project](https://bioconductor.org/) hosts several integration methods with documentation that explains their assumptions and limitations. ## Related Bioinformatics Guides - [Spatial Proteomics vs. Single-Cell Proteomics: Choosing the Right Approach](/knowledge/bioinformatics/spatial-proteomics-vs-single-cell-proteomics-choosing-the-right-approach) - [Spatial Transcriptomics vs. Single-Cell RNA Sequencing: Which Approach Fits Your Research?](/knowledge/bioinformatics/spatial-transcriptomics-vs-single-cell-rna-sequencing-which-approach-fits-your-research) - [Single-Cell Genomics: From Concept to Application](/knowledge/bioinformatics/single-cell-genomics-from-concept-to-application) - [Single-Cell Annotation: A Workflow for Cell Type Identification](/knowledge/bioinformatics/single-cell-annotation-a-workflow-for-cell-type-identification) - [Single-Cell Isolation Techniques: A Practical Comparison](/knowledge/bioinformatics/single-cell-isolation-techniques-a-practical-comparison) ## Related Clinical & Scientific Guides * [A Practical Guide to Detecting Antimicrobial Resistance Genes in Shotgun Metagenomic Data](/knowledge/bioinformatics/a-practical-guide-to-detecting-antimicrobial-resistance-genes-in-shotgun-metagenomic-data) * [Computational Immunology: Modeling the Immune System](/knowledge/bioinformatics/computational-immunology-modeling-the-immune-system) * [How to Set Hard Filters for Germline Variant Calling: A Practical Guide to GATK Best Practices](/knowledge/bioinformatics/how-to-set-hard-filters-for-germline-variant-calling-a-practical-guide-to-gatk-best-practices) ## References and Further Reading - [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/). National Center for Biotechnology Information. - [EMBL-EBI Training](https://www.ebi.ac.uk/training). European Bioinformatics Institute. - [Bioconductor](https://bioconductor.org/). Bioconductor Project. - [Galaxy Training Network](https://training.galaxyproject.org/). Galaxy Project. - [nf-core Documentation](https://nf-co.re/docs). nf-core. - [The Carpentries Lessons](https://carpentries.org/lessons). The Carpentries. - [Bifunctional chemokine-nanobody fusion protein enhances neutrophil recruitment to impede Acanthamoeba immune evasion.](https://pubmed.ncbi.nlm.nih.gov/40222104). EBioMedicine, 2025. - [Single-cell transcriptomic profiling reveals innate-like cytotoxic intraepithelial lymphocyte expansion during <,i>,Salmonella<,/i>, Enteritidis infection in chickens.](https://doi.org/10.3389/fimmu.2026.1842283). 2026. - [Interferon-primed immune landscapes predict immune-related adverse events during immune checkpoint inhibitor therapy.](https://doi.org/10.3389/fcell.2026.1798845). 2026. - [Integrative single-cell and spatial transcriptomics analysis reveals a baicalein-responsive 10-gene signature for non-small cell lung cancer.](https://doi.org/10.1016/j.tranon.2026.102853). 2026. - [Integrated transcriptomics identifies ER stress-associated apoptosis in post-resuscitation AKI and supports early Dl-3-n-butylphthalide-associated renoprotection in a porcine TCA model.](https://doi.org/10.3389/fphar.2026.1841271). 2026. - [CAdir: Joint clustering of cells and genes for single-cell transcriptomics with visualization-driven cluster quality assessment](https://doi.org/10.1371/journal.pcbi.1014418). PLoS Computational Biology, 2026. > This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.