# Visualizing Single-Cell Proteomics Data: Best Practices for t-SNE, UMAP, and Heatmaps

Single-cell proteomics generates high-dimensional datasets that require deliberate dimensionality reduction and visualization before biological interpretation becomes reliable. This article provides a practical framework for selecting between t-SNE, UMAP, and heatmap approaches, tuning their parameters, and avoiding common interpretive errors, with examples drawn from CyTOF and other single-cell protein measurement platforms.

## The Core Problem: High-Dimensional Data Exceeds Human Visual Capacity

A single CyTOF experiment can measure 30 to 50 protein markers simultaneously across hundreds of thousands of individual cells. Mass spectrometry-based single-cell proteomics platforms similarly generate measurements for dozens to hundreds of proteins per cell. The resulting data matrix, with cells as rows and protein measurements as columns, cannot be directly visualized because human perception operates in two or three dimensions.

Dimensionality reduction methods address this gap by projecting high-dimensional measurements into a low-dimensional space that preserves meaningful structure. The two most widely used methods for single-cell proteomics are t-distributed stochastic neighbor embedding (t-SNE) and uniform manifold approximation and projection (UMAP). Heatmaps provide a complementary view that preserves the original measurement scale while organizing cells and proteins into interpretable blocks.

The choice of visualization method and its parameters directly affects which biological patterns become visible and which remain hidden. A poorly tuned t-SNE can separate a homogeneous cell population into artificial clusters, while an over-clustered UMAP can obscure meaningful continuous transitions between cell states. Understanding the mathematical behavior of each method is therefore a prerequisite for correct biological interpretation.

## At a Glance: Method Selection for Single-Cell Proteomics

| Visualization Method | Best Use Case | Key Parameters | Primary Limitation |
| --- | --- | --- | --- |
| t-SNE | Exploring local cell-cell relationships, identifying discrete cell populations | Perplexity (5 to 50), learning rate, number of iterations | Global structure is not preserved, cluster distances are not meaningful |
| UMAP | Visualizing both local and global structure, comparing multiple samples | n_neighbors (5 to 50), min_dist (0.01 to 0.5), spread | Parameter sensitivity can create artificial separation or merging |
| Heatmap | Examining protein expression patterns across defined cell populations | Row and column ordering, scaling method, color range | Requires pre-defined cell groupings, does not show single-cell relationships |
| Joint t-SNE or UMAP | Integrating multimodal data such as protein and RNA measurements | Modality weighting, per-modality preprocessing | Requires matched multimodal measurements from the same cells |

## Core Principles of Dimensionality Reduction for Protein Data

### The Geometry of Protein Measurement Space

Single-cell proteomics data occupy a high-dimensional space where each axis represents the measured abundance of one protein. Cells with similar protein expression profiles sit close together in this space, while cells with different functional states sit farther apart. The goal of dimensionality reduction is to create a two-dimensional map that preserves these distances as faithfully as possible.

The mathematical challenge is that no two-dimensional projection can perfectly preserve all pairwise distances from a 30-dimensional space. Every method makes tradeoffs about which distances to prioritize. t-SNE prioritizes local structure, ensuring that cells that are very similar in high-dimensional space appear close together in the projection. UMAP attempts to preserve both local and global structure, making it more suitable for visualizing relationships between distant cell populations. These differences matter for protein data because cell types often form a hierarchy, with closely related subtypes nested within broader lineages.

### Why Protein Data Differ from RNA Data

Single-cell proteomics datasets have properties that distinguish them from single-cell RNA sequencing data. Protein measurements are often less sparse than RNA measurements, because proteins are more stable than transcripts and are detected across a wider dynamic range. However, protein data frequently suffer from batch effects, where measurements from different experimental runs are not directly comparable due to instrument drift or reagent lot variation.

The choice of dimensionality reduction method should account for these properties. Methods that assume a particular noise distribution may perform poorly on protein data with different error characteristics. The generalization of t-SNE and UMAP to multimodal omics data, where protein and RNA measurements from the same cells are jointly embedded, requires careful weighting of each modality to prevent one data type from dominating the visualization. This joint approach can produce embeddings that better agree with known cell types and that harmonize RNA and protein velocity landscapes, as demonstrated in the j-SNE and j-UMAP methods applied to eight datasets ([PubMed: A generalization of t-SNE and UMAP to single-cell multimodal omics](https://pubmed.ncbi.nlm.nih.gov/33941244)).

### The Role of Preprocessing Before Visualization

Dimensionality reduction operates on the output of preprocessing steps, and poor preprocessing cannot be corrected by parameter tuning. For CyTOF data, standard preprocessing includes arcsinh transformation to stabilize variance, compensation to correct for signal spillover between channels, and normalization to align measurements across batches. For mass spectrometry-based single-cell proteomics, preprocessing includes peptide identification, protein quantification, and missing value handling.

The choice of preprocessing directly affects visualization outcomes. Failure to normalize across batches can create artificial clusters that separate cells by experimental run instead of by biological state. Failure to handle missing values appropriately can distort distances between cells, because cells with missing measurements may appear more similar to each other than they truly are. The sensitivity analysis framework for t-SNE embeddings demonstrates that the stability of the visual layout depends on the input data, and that uncertainty in the input propagates through the embedding to create positional uncertainty in the projected points ([Frontiers in Bioinformatics: Visualizing stability](https://doi.org/10.3389/fbinf.2025.1719516)).

## Practical Workflow for Generating t-SNE and UMAP Plots

### Step 1: Define the Biological Question

Before generating any visualization, specify the biological question that the plot must answer. Are you trying to identify novel cell populations? Are you comparing protein expression between treatment groups? Are you examining the continuity of a differentiation trajectory? The answer determines which visualization method and parameters are appropriate.

For discrete population identification, t-SNE with moderate perplexity often works well because it emphasizes local structure and separates distinct cell types. For continuous processes such as activation or differentiation, UMAP with a larger n_neighbors value may better preserve the gradual transitions between cell states. For patient stratification based on functional proteomics, the visualization must preserve differences between clinically meaningful groups, as demonstrated in the analysis of CAR-T cell products where t-SNE and UMAP revealed distinct cytokine profiles between nonresponders and responders ([PubMed: Advanced Cell Mapping Visualizations](https://pubmed.ncbi.nlm.nih.gov/32108428)).

### Step 2: Preprocess and Normalize the Data

Apply the appropriate transformation for your data type. For CyTOF data, arcsinh transformation with a cofactor between 5 and 10 is standard practice. For mass spectrometry data, log2 transformation is common. Normalize across batches using methods appropriate to your platform, and document all preprocessing steps for reproducibility.

Check for batch effects before proceeding to dimensionality reduction. Visualize the data colored by batch using an initial UMAP or principal component analysis. If cells cluster by batch instead of by biological condition, apply batch correction before generating final visualizations.

### Step 3: Select the Dimensionality Reduction Method

Choose between t-SNE and UMAP based on the biological question and the structure of the data. For datasets where local relationships are most important, t-SNE is appropriate. For datasets where global relationships between distant populations matter, UMAP is preferred.

Consider whether the data are spatially resolved. Standard t-SNE and UMAP were not tailored for spatially resolved profiling data, and methods such as SpaSNE that integrate both spatial and molecular information can achieve more accurate and meaningful visualization that better elucidates the underlying spatial and molecular data structures ([PubMed: Dimensionality reduction for spatially resolved profiling data](https://pubmed.ncbi.nlm.nih.gov/39960663)). If your protein data include spatial coordinates, use a spatial-aware method instead of a standard embedding.

### Step 4: Tune Parameters Systematically

Do not rely on default parameters without examination. For t-SNE, the perplexity parameter controls the balance between local and global structure. Low perplexity values emphasize very local neighborhoods, while high values incorporate more distant cells into each point's neighborhood. For UMAP, the n_neighbors parameter serves a similar role, and the min_dist parameter controls how tightly points are packed in the embedding.

Run the embedding across a range of parameter values and compare the results. A stable embedding, where the major cell populations remain separated and the relationships between populations remain consistent across parameter values, provides more confidence than an embedding that changes dramatically with small parameter changes. The sensitivity analysis framework for t-SNE can reveal which input features are most influential in shaping the embedding and identify regions of structural instability ([Frontiers in Bioinformatics: Visualizing stability](https://doi.org/10.3389/fbinf.2025.1719516)).

### Step 5: Validate the Embedding

A dimensionality reduction plot is only useful if it accurately represents the underlying data. Validate the embedding by checking that known cell populations appear as expected. If you have a reference marker panel, verify that cells expressing known lineage markers cluster together.

Use clustering algorithms to define cell populations in the embedding, then examine the protein expression profiles of each cluster using heatmaps or violin plots. The clusters should have distinct, interpretable protein expression patterns. If clusters lack clear protein markers, the embedding may be driven by technical artifacts instead of biological signal.

### Step 6: Generate Publication-Quality Figures

For final figures, ensure that the visualization is reproducible by recording all parameters and preprocessing steps. Color cells by biologically meaningful variables such as cell type, treatment condition, or protein expression level. Include axis labels that indicate the dimensionality reduction method and parameters used.

For heatmaps, order rows and columns to reveal structure. Hierarchical clustering with appropriate distance metrics and linkage methods is standard. Scale the color range to the data distribution, and avoid color scales that obscure low-abundance proteins or saturate at high abundance.

## Parameter Tuning: A Detailed Examination

### t-SNE Perplexity

The perplexity parameter in t-SNE determines the effective number of neighbors considered for each point. Low perplexity values, such as 5 to 10, create embeddings that emphasize very local structure and can fragment continuous populations into multiple clusters. High perplexity values, such as 50 to 100, create embeddings that incorporate more global structure but can merge distinct populations that are close together.

For single-cell proteomics data with tens of thousands of cells, perplexity values between 15 and 50 are commonly used. The optimal value depends on the density of the data and the size of the smallest population of interest. If a rare cell population constitutes less than 1 percent of the total cells, a perplexity value that is too high may fail to resolve it.

The stability of t-SNE embeddings varies across the parameter space. Some regions of the embedding may be highly sensitive to small changes in perplexity, while others remain stable. The sensitivity analysis framework can identify these regions of structural instability, allowing you to focus interpretive attention on stable features ([Frontiers in Bioinformatics: Visualizing stability](https://doi.org/10.3389/fbinf.2025.1719516)).

### UMAP n_neighbors and min_dist

The n_neighbors parameter in UMAP controls the size of the local neighborhood used to estimate the manifold structure. Small values, such as 5 to 10, emphasize local structure and can create many small clusters. Large values, such as 50 to 100, incorporate more global structure and produce smoother embeddings with fewer, larger clusters.

The min_dist parameter controls the minimum distance between points in the embedding. Small values, such as 0.01, pack points tightly together and create dense clusters. Large values, such as 0.5, spread points out and create more diffuse clusters. The choice of min_dist affects the visual appearance of the embedding but has less effect on the overall structure than n_neighbors.

For single-cell proteomics data, n_neighbors values between 10 and 30 and min_dist values between 0.1 and 0.5 are common starting points. As with t-SNE, run the embedding across a range of parameters and compare the results to identify stable features.

### Learning Rate and Iterations

The learning rate controls the step size during optimization of the embedding. Values that are too low can cause the optimization to get stuck in poor local optima, while values that are too high can cause the embedding to become unstable. For t-SNE, learning rates between 200 and 1000 are common. For UMAP, the learning rate is typically set by the implementation and requires less adjustment.

The number of iterations controls how long the optimization runs. Too few iterations can produce an embedding that has not converged, while too many iterations waste computational resources. For t-SNE, 1000 to 5000 iterations are typically sufficient. For UMAP, the default number of epochs is usually adequate, but increasing the number of epochs can improve convergence for large datasets.

### Random Seed and Reproducibility

Both t-SNE and UMAP use stochastic optimization, meaning that different runs can produce different embeddings even with the same parameters and data. Set a random seed before running the embedding to ensure that results are reproducible. Document the random seed in your analysis code and in the methods section of any publication.

The variability between runs provides information about the stability of the embedding. Run the embedding multiple times with different random seeds and compare the results. If the major cell populations remain consistent across runs, the embedding is stable. If the embedding changes dramatically, the data may not have a clear low-dimensional structure, or the parameters may be inappropriate.

## Heatmaps for Single-Cell Proteomics

### When to Use Heatmaps Instead of or Alongside t-SNE and UMAP

Heatmaps provide a different view of single-cell proteomics data than t-SNE or UMAP. While dimensionality reduction plots show the relationships between individual cells, heatmaps show the protein expression values directly. This makes heatmaps useful for examining the specific proteins that define each cell population and for comparing expression patterns across conditions.

Heatmaps require that cells be grouped into populations before visualization. These groupings can come from clustering the dimensionality reduction embedding, from manual gating, or from known biological classifications. The heatmap then displays the average or median expression of each protein within each group, or the expression of individual cells organized by group.

For single-cell functional proteomics, heatmaps can reveal the polyfunctional subsets of cells that co-secrete many proteins from live single cells. The visualization of these subsets can correlate with patient response to therapy, as demonstrated in the analysis of CAR-T cell products where distinct cytokine profiles between nonresponders and responders were identified ([PubMed: Advanced Cell Mapping Visualizations](https://pubmed.ncbi.nlm.nih.gov/32108428)).

### Heatmap Construction and Scaling

The construction of a heatmap requires several decisions that affect its interpretability. The choice of scaling method determines how protein expression values are displayed. Z-score scaling, where each protein's expression is centered and scaled to unit variance, is common because it allows comparison of proteins with different abundance ranges. However, z-score scaling can exaggerate the importance of low-abundance proteins with small variance.

The choice of color scale affects which patterns are visible. Sequential color scales, such as white to blue or white to red, are appropriate for showing increasing expression. Diverging color scales, such as blue to white to red, are appropriate for showing expression relative to a reference value. The color range should be set to the data distribution, avoiding saturation that hides differences at the extremes.

The ordering of rows and columns determines which patterns are visible. Hierarchical clustering with Euclidean distance and Ward's linkage is a common choice. The resulting dendrogram shows the relationships between proteins and between cell populations, allowing you to identify groups of co-expressed proteins and groups of cells with similar expression profiles.

### Heatmap Limitations

Heatmaps have important limitations for single-cell proteomics. They do not show the relationships between individual cells, because cells are aggregated into groups. This can obscure heterogeneity within groups and hide rare cell populations. Heatmaps also require that the grouping of cells is correct, and errors in clustering or gating propagate into the heatmap display.

For large numbers of cells, heatmaps can become unwieldy. Displaying individual cells as rows in a heatmap requires substantial computational resources and produces figures that are difficult to interpret. Aggregating cells into groups loses single-cell resolution but produces more interpretable figures.

## Common Failure Patterns and How to Avoid Them

### Artificial Clustering from Overly High Perplexity or n_neighbors

A common failure pattern is the appearance of artificial clusters that do not correspond to real biological populations. This can occur when the perplexity or n_neighbors parameter is too high, causing the embedding to merge distinct populations or to split continuous populations. The resulting clusters may have overlapping protein expression profiles and lack clear markers.

To avoid this failure, run the embedding across a range of parameter values and compare the results. If clusters appear and disappear with small parameter changes, they are likely artifacts. Validate clusters by examining their protein expression profiles and by checking that they correspond to known cell types.

### Batch Effects Creating False Separation

Batch effects can create artificial separation between cells from different experimental runs. This appears as clusters that correspond to batch instead of to biological condition. The problem is particularly acute in single-cell proteomics, where instrument drift and reagent variation are common.

To avoid this failure, check for batch effects before dimensionality reduction. Visualize the data colored by batch and look for separation. Apply batch correction methods appropriate to your platform, and document the correction in your analysis. After correction, re-check that cells from different batches are mixed within biological populations.

### Over-interpretation of Distances in t-SNE

A common interpretive error is treating distances between clusters in t-SNE as meaningful. t-SNE does not preserve global distances, and the distance between two distant clusters in the embedding has no quantitative meaning. Only the local structure, where nearby points are similar, is reliable.

To avoid this error, use UMAP when global relationships matter. UMAP preserves more global structure than t-SNE, making distances between distant populations more interpretable. Alternatively, use a separate visualization method, such as principal component analysis, to examine global relationships.

### Ignoring Uncertainty in the Embedding

Dimensionality reduction embeddings are deterministic functions of the input data, and uncertainty in the input propagates to uncertainty in the embedding. Cells with noisy measurements may be placed in unstable positions, and the visual layout may change with small changes in the input data.

To avoid this failure, assess the stability of the embedding. Run the embedding multiple times with different random seeds and compare the results. Use sensitivity analysis to identify which input features are most influential in shaping the embedding and which regions are structurally unstable ([Frontiers in Bioinformatics: Visualizing stability](https://doi.org/10.3389/fbinf.2025.1719516)). Report the stability of the embedding alongside the visualization.

### Using Default Parameters Without Examination

Default parameters are convenient but may not be appropriate for your data. The optimal parameters depend on the number of cells, the number of proteins, the density of the data, and the biological question. Using default parameters without examination can produce embeddings that obscure important structure or create artifacts.

To avoid this failure, treat parameter selection as part of the analysis. Run the embedding across a range of parameter values, compare the results, and document the parameter choices. Report the parameters in the methods section of any publication.

## Records and Measurements for Reproducible Visualization

### What to Record for Each Visualization

Reproducible visualization requires detailed records of every decision that affects the output. For each t-SNE or UMAP plot, record the input data version, the preprocessing steps, the transformation applied, the normalization method, the dimensionality reduction method, all parameters, the random seed, and the software version.

For heatmaps, record the scaling method, the distance metric, the linkage method, the row and column ordering, the color scale, and the grouping of cells. Record the software and version used to generate the figure.

### Version Control for Data and Code

Version control is essential for reproducible visualization. Store the raw data, the preprocessing code, the dimensionality reduction code, and the figure generation code in a version control system. Tag each version with a descriptive commit message that records the changes made.

The [Carpentries Lessons](https://carpentries.org/lessons) provide foundational training in version control with Git, which is applicable to bioinformatics workflows. The [Galaxy Training Network](https://training.galaxyproject.org/) offers accessible workflow training that emphasizes reproducibility, and the [nf-core Documentation](https://nf-co.re/docs) describes community standards for reproducible pipeline configuration.

### Documentation of Parameter Choices

Document the rationale for each parameter choice. For example, record why a particular perplexity value was selected for t-SNE, or why a particular n_neighbors value was selected for UMAP. This documentation helps other researchers understand the analysis and helps you reproduce the results later.

The [Bioconductor](https://bioconductor.org/) project provides official documentation for many R packages used in single-cell analysis, including package vignettes that describe recommended workflows and parameter choices. The [EMBL-EBI Training](https://www.ebi.ac.uk/training) portal offers learning pathways for bioinformatics data analysis that include practical guidance on visualization.

## Quality Controls for Visualization Outputs

### Checking for Convergence

Both t-SNE and UMAP use iterative optimization, and the quality of the final embedding depends on convergence. Check that the optimization has converged by examining the loss function or by comparing the embedding across different numbers of iterations. An embedding that changes substantially with more iterations has not converged.

### Checking for Parameter Sensitivity

A robust embedding should not change dramatically with small parameter changes. Run the embedding with slightly different parameter values and compare the results. If the major cell populations remain consistent, the embedding is robust. If the embedding changes dramatically, the parameters may be inappropriate or the data may not have a clear low-dimensional structure.

### Checking for Batch Effects

After generating the embedding, color the points by batch and examine whether cells from different batches are mixed within biological populations. If cells separate by batch, the preprocessing or batch correction was insufficient. Re-examine the preprocessing steps and apply additional correction if needed.

### Checking for Known Biological Markers

The most important quality control is biological validation. Check that known cell populations appear as expected in the embedding. If you have a reference marker panel, verify that cells expressing known lineage markers cluster together. If known populations are not resolved, the embedding may be driven by technical artifacts or the parameters may be inappropriate.

## Limitations of Dimensionality Reduction for Single-Cell Proteomics

### Loss of Quantitative Information

Dimensionality reduction plots show the relationships between cells but do not show the actual protein expression values. A cell's position in the embedding is a nonlinear function of all measured proteins, and the contribution of individual proteins to the position is not directly visible. To examine specific protein expression, use heatmaps or other visualization methods that preserve the measurement scale.

### Sensitivity to Preprocessing Choices

The embedding is sensitive to preprocessing choices, including transformation, normalization, and batch correction. Different preprocessing choices can produce different embeddings, and there is no universally correct preprocessing pipeline. Document all preprocessing steps and consider how they affect the visualization.

### Inability to Represent All Structure

No two-dimensional embedding can perfectly represent all structure in a high-dimensional dataset. Some biological relationships will be distorted or hidden, regardless of the method and parameters used. The sensitivity analysis framework for t-SNE can reveal which features are most influential in shaping the embedding and identify regions of structural instability ([Frontiers in Bioinformatics: Visualizing stability](https://doi.org/10.3389/fbinf.2025.1719516)), but it cannot recover structure that is lost in the projection.

### Computational Requirements

t-SNE and UMAP are computationally intensive, particularly for large datasets. The computational cost increases with the number of cells and the number of proteins. For very large datasets, consider using approximate methods or subsampling. The [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/) provides data resources and analysis services that may be useful for managing large datasets.

## Professional Escalation Criteria

### When to Seek Expert Assistance

Some visualization problems require expert assistance. Seek help from a bioinformatics specialist or a statistician when the embedding is highly unstable across parameter values, when batch effects cannot be corrected with standard methods, when the biological interpretation of the embedding is unclear, or when the visualization is intended to support regulatory or clinical decisions.

The [EMBL-EBI Training](https://www.ebi.ac.uk/training) portal offers learning pathways that can help you build the skills needed to address these problems independently. The [Bioconductor](https://bioconductor.org/) support forum provides access to experts who can help with specific analysis questions.

### When to Question the Underlying Data

A poor visualization may indicate problems with the underlying data instead of with the visualization method. Question the data quality when the embedding shows no clear structure, when known cell populations are not resolved, or when the embedding is dominated by technical artifacts. Re-examine the preprocessing steps, check for batch effects, and verify the quality of the protein measurements.

### When to Use Alternative Methods

Standard t-SNE and UMAP may not be appropriate for all single-cell proteomics data. For spatially resolved data, use spatial-aware methods such as SpaSNE that integrate spatial and molecular information ([PubMed: Dimensionality reduction for spatially resolved profiling data](https://pubmed.ncbi.nlm.nih.gov/39960663)). For multimodal data, use joint embedding methods that learn the relative contribution of each modality ([PubMed: A generalization of t-SNE and UMAP to single-cell multimodal omics](https://pubmed.ncbi.nlm.nih.gov/33941244)). For data with high uncertainty, use methods that propagate uncertainty into the visualization ([Frontiers in Bioinformatics: Visualizing stability](https://doi.org/10.3389/fbinf.2025.1719516)).

## A Decision Framework for Matching Visualization Choice to Data Structure and Question

Selecting between t-SNE, UMAP, and heatmaps is not a matter of preference but a function of the biological question, the data structure, and the downstream analytical steps. This section provides a structured decision framework that connects data characteristics to visualization choices, along with a record system for documenting visualization decisions and a troubleshooting method for diagnosing poor embeddings.

### The Three-Question Screening Test

Before generating any plot, answer three questions about the dataset and the analytical goal. The answers determine which visualization method is appropriate and which parameters deserve the most attention.

**Question 1: What is the primary analytical goal?**

The goal determines whether local or global structure matters more. If the goal is to identify discrete cell populations, such as immune cell subtypes in a CyTOF panel, t-SNE with moderate perplexity emphasizes local neighborhoods and separates distinct types. If the goal is to examine relationships between distant populations or to compare multiple samples, UMAP preserves more global structure and supports cross-sample comparison. If the goal is to examine the specific proteins that define each population, a heatmap with hierarchical clustering provides direct access to expression values.

**Question 2: What is the expected structure of the data?**

The expected structure influences parameter selection. Discrete populations with clear boundaries respond well to lower perplexity or n_neighbors values that emphasize local separation. Continuous processes, such as differentiation trajectories or activation gradients, require higher values that preserve gradual transitions. For spatially resolved proteomics data, standard t-SNE and UMAP are not tailored to integrate spatial and molecular information, and methods such as SpaSNE that incorporate both dimensions can produce more accurate visualization ([PubMed: Dimensionality reduction for spatially resolved profiling data](https://pubmed.ncbi.nlm.nih.gov/39960663)).

**Question 3: What is the data quality profile?**

Data quality determines whether the visualization can be trusted. High missingness, strong batch effects, or high technical noise require additional preprocessing before visualization and may require methods that propagate uncertainty into the embedding. The sensitivity analysis framework for t-SNE demonstrates that uncertainty in input data propagates through the embedding to create positional uncertainty in projected points, and this uncertainty should be assessed before interpreting visual patterns ([Frontiers in Bioinformatics: Visualizing stability](https://doi.org/10.3389/fbinf.2025.1719516)).

### A Decision Matrix for Method Selection

The following matrix connects data characteristics to recommended visualization approaches. This matrix complements the At a Glance table by adding the data structure and quality dimensions that parameter selection depends on.

| Data Characteristic | Recommended Method | Primary Parameter Focus | Validation Priority |
| --- | --- | --- | --- |
| Discrete populations, low noise | t-SNE with perplexity 15 to 30 | Perplexity, learning rate | Cluster marker specificity |
| Continuous transitions, high cell count | UMAP with n_neighbors 20 to 50 | n_neighbors, min_dist | Trajectory continuity |
| Multimodal data (protein plus RNA) | Joint embedding (j-SNE or j-UMAP) | Modality weighting | Cross-modality agreement |
| Spatially resolved protein data | SpaSNE or spatial-aware method | Spatial weight, molecular weight | Spatial coherence |
| High missingness or batch effects | UMAP after batch correction | Preprocessing, n_neighbors | Batch mixing within populations |
| Defined populations for marker inspection | Heatmap with hierarchical clustering | Scaling, row and column order | Cluster expression profiles |

The joint embedding approach deserves specific attention for multimodal experiments. The j-SNE and j-UMAP methods automatically learn the relative contribution of each modality to a concise representation of cellular identity, promoting discriminative features while suppressing noise. Applied to eight datasets, these methods produced unified embeddings that better agreed with known cell types and harmonized RNA and protein velocity landscapes ([PubMed: A generalization of t-SNE and UMAP to single-cell multimodal omics](https://pubmed.ncbi.nlm.nih.gov/33941244)). If your experiment measures both protein and RNA from the same cells, use a joint embedding method instead of separate embeddings for each modality.

### The Parameter Sweep Protocol

A systematic parameter sweep provides more information than a single embedding with default parameters. The protocol below produces a parameter sensitivity record that supports both interpretation and publication.

**Step 1: Define the parameter range.**

For t-SNE, test perplexity values of 5, 10, 15, 30, 50, and 100. For UMAP, test n_neighbors values of 5, 10, 15, 30, 50, and 100, and min_dist values of 0.01, 0.1, 0.3, and 0.5. These ranges cover the commonly used parameter space for single-cell data.

**Step 2: Run the embedding across the full grid.**

Generate an embedding for each parameter combination. For each embedding, record the random seed, the software version, and the computational time. Store all embeddings in a single object or directory for comparison.

**Step 3: Score each embedding for stability and interpretability.**

For each embedding, record three scores. First, the cluster stability score, which measures whether the major cell populations remain separated across parameter values. Second, the marker specificity score, which measures whether known lineage markers are enriched in the expected clusters. Third, the batch mixing score, which measures whether cells from different experimental runs are mixed within biological populations.

**Step 4: Select the parameter set with the best combined scores.**

The optimal parameter set balances cluster resolution with stability. If the embedding changes dramatically across the parameter range, the data may not have a clear low-dimensional structure, or the preprocessing may be inadequate. The sensitivity analysis framework can identify which input features are most influential in shaping the embedding and which regions are structurally unstable ([Frontiers in Bioinformatics: Visualizing stability](https://doi.org/10.3389/fbinf.2025.1719516)).

### A Record System for Visualization Decisions

Reproducible visualization requires a structured record of every decision that affects the output. The following record template captures the information needed to reproduce any plot and to diagnose problems when they arise.

**Dataset record.** Record the data file name and version, the number of cells, the number of protein markers, the platform used for measurement, and the date of data export.

**Preprocessing record.** Record the transformation applied (arcsinh with cofactor, log2, or other), the normalization method, the batch correction method if used, and the missing value handling strategy.

**Embedding record.** Record the dimensionality reduction method, all parameters, the random seed, the software package and version, and the computational time.

**Validation record.** Record the results of the marker specificity check, the batch mixing check, and the parameter sensitivity check. Note any regions of the embedding that are structurally unstable.

**Figure record.** Record the color scheme, the axis labels, the point size, and any annotations added to the figure.

Store these records in a structured format, such as a spreadsheet or a YAML file, alongside the analysis code. The [Carpentries Lessons](https://carpentries.org/lessons) provide foundational training in version control with Git that supports this record keeping. The [Galaxy Training Network](https://training.galaxyproject.org/) offers workflow training that emphasizes reproducibility through structured analysis records.

### Troubleshooting Poor Embeddings

When an embedding fails to reveal expected structure or shows artifacts, work through the following diagnostic sequence. This sequence separates data problems from parameter problems and prevents wasted effort on parameter tuning when the data are the issue.

**Diagnostic 1: Check the preprocessing.**

Examine the distribution of each protein marker after transformation. Markers with extreme outliers or bimodal distributions may distort the embedding. Check for batch effects by coloring an initial UMAP by batch. If cells separate by batch, apply batch correction before proceeding.

**Diagnostic 2: Check the parameter sensitivity.**

Run the embedding with the parameter sweep protocol described above. If the major populations remain consistent across parameter values, the embedding is stable. If the embedding changes dramatically, the parameters are inappropriate or the data lack clear structure.

**Diagnostic 3: Check the marker specificity.**

Color the embedding by each known lineage marker. Expected populations should show enriched marker expression. If known populations are not resolved, the embedding may be driven by technical artifacts. The sensitivity analysis framework can reveal which input features are most influential in shaping the embedding, helping you identify whether the visualization is driven by biological signal or technical noise ([Frontiers in Bioinformatics: Visualizing stability](https://doi.org/10.3389/fbinf.2025.1719516)).

**Diagnostic 4: Check the cluster interpretability.**

Use a clustering algorithm to define populations in the embedding, then examine the protein expression profiles of each cluster. Clusters should have distinct, interpretable protein patterns. If clusters lack clear markers, the embedding may be separating cells by technical factors instead of biological state.

**Diagnostic 5: Consider alternative methods.**

If standard t-SNE and UMAP fail to reveal meaningful structure, consider whether the data require a specialized method. For spatially resolved data, use SpaSNE or another spatial-aware method ([PubMed: Dimensionality reduction for spatially resolved profiling data](https://pubmed.ncbi.nlm.nih.gov/39960663)). For multimodal data, use joint embedding methods that learn modality weights ([PubMed: A generalization of t-SNE and UMAP to single-cell multimodal omics](https://pubmed.ncbi.nlm.nih.gov/33941244)). For data with high uncertainty, use methods that propagate uncertainty into the visualization ([Frontiers in Bioinformatics: Visualizing stability](https://doi.org/10.3389/fbinf.2025.1719516)).

### Integrating Visualization with Downstream Analysis

The visualization choice affects downstream analytical steps, and the decision framework should account for this connection. Clustering results derived from the embedding depend on the parameter choices, and trajectory analysis depends on the continuity preserved by the embedding.

For clustering, the embedding parameters determine the resolution of the resulting clusters. Lower perplexity or n_neighbors values produce more clusters, while higher values produce fewer, larger clusters. The clustering resolution should match the biological question. If you are looking for rare populations, use lower values. If you are looking for broad lineages, use higher values.

For trajectory analysis, the embedding must preserve continuous transitions. UMAP with higher n_neighbors values preserves more global structure and supports trajectory inference better than t-SNE. The joint embedding methods j-SNE and j-UMAP harmonize RNA and protein velocity landscapes, supporting trajectory analysis across modalities ([PubMed: A generalization of t-SNE and UMAP to single-cell multimodal omics](https://pubmed.ncbi.nlm.nih.gov/33941244)).

For patient stratification, the visualization must preserve differences between clinically meaningful groups. The analysis of CAR-T cell products demonstrated that t-SNE and UMAP can reveal distinct cytokine profiles between nonresponders and responders, supporting biomarker discovery and treatment development ([PubMed: Advanced Cell Mapping Visualizations](https://pubmed.ncbi.nlm.nih.gov/32108428)). When the visualization supports clinical decisions, document all parameter choices and validation results thoroughly.

### Professional Escalation Criteria for Visualization Problems

Some visualization problems require expert assistance. Seek help from a bioinformatics specialist or statistician when the embedding is highly unstable across the parameter sweep, when batch effects cannot be corrected with standard methods, when the biological interpretation of the embedding is unclear after completing the diagnostic sequence, or when the visualization supports regulatory or clinical decisions.

The [EMBL-EBI Training](https://www.ebi.ac.uk/training) portal offers learning pathways that build the skills needed to address these problems independently. The [Bioconductor](https://bioconductor.org/) support forum provides access to experts who can help with specific analysis questions. The [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/) provides data resources and analysis services that may support large-scale visualization projects.

When escalating, provide the complete visualization record, including the dataset record, preprocessing record, embedding record, validation record, and figure record. This documentation allows the expert to reproduce the analysis and diagnose the problem efficiently.

## Frequently Asked Questions

### What is the difference between t-SNE and UMAP for single-cell proteomics data?

t-SNE emphasizes local structure and is effective for separating discrete cell populations, but it does not preserve global distances between distant clusters. UMAP preserves both local and global structure, making it more suitable for visualizing relationships between distant cell populations and for comparing multiple samples. For single-cell proteomics data, UMAP is often preferred when the biological question involves relationships between cell types, while t-SNE is useful for resolving closely related subtypes.

### How do I choose the perplexity value for t-SNE?

The perplexity value controls the effective number of neighbors considered for each point. Low values between 5 and 15 emphasize very local structure and can fragment continuous populations. High values between 50 and 100 incorporate more global structure but can merge distinct populations. For single-cell proteomics data with tens of thousands of cells, values between 15 and 50 are common. Run the embedding across a range of values and compare the results to identify stable features.

### How do I choose the n_neighbors and min_dist values for UMAP?

The n_neighbors parameter controls the size of the local neighborhood used to estimate the manifold structure. Small values between 5 and 10 emphasize local structure and create many small clusters. Large values between 50 and 100 produce smoother embeddings with fewer, larger clusters. The min_dist parameter controls the minimum distance between points in the embedding, with small values packing points tightly and large values spreading them out. Values of 10 to 30 for n_neighbors and 0.1 to 0.5 for min_dist are common starting points.

### Why do my t-SNE and UMAP plots change every time I run them?

Both t-SNE and UMAP use stochastic optimization, meaning that different runs can produce different embeddings even with the same parameters and data. Set a random seed before running the embedding to ensure reproducibility. Run the embedding multiple times with different random seeds to assess stability. If the major cell populations remain consistent across runs, the embedding is stable.

### How do I know if my clusters are real or artifacts?

Validate clusters by examining their protein expression profiles and by checking that they correspond to known cell types. Run the embedding across a range of parameter values and compare the results. If clusters appear and disappear with small parameter changes, they are likely artifacts. Use clustering algorithms to define cell populations in the embedding, then examine the protein expression profiles of each cluster using heatmaps or violin plots.

### What should I do if my data separate by batch instead of by biological condition?

Batch effects can create artificial separation between cells from different experimental runs. Check for batch effects before dimensionality reduction by visualizing the data colored by batch. Apply batch correction methods appropriate to your platform, and document the correction in your analysis. After correction, re-check that cells from different batches are mixed within biological populations.

### Can I use heatmaps instead of t-SNE or UMAP for single-cell proteomics?

Heatmaps provide a different view of the data than dimensionality reduction plots. While t-SNE and UMAP show the relationships between individual cells, heatmaps show the protein expression values directly. Heatmaps require that cells be grouped into populations before visualization. Use heatmaps to examine the specific proteins that define each cell population and to compare expression patterns across conditions, alongside dimensionality reduction plots for single-cell relationships.

### How do I make my visualizations reproducible?

Record the input data version, all preprocessing steps, the transformation applied, the normalization method, the dimensionality reduction method, all parameters, the random seed, and the software version. Store the raw data, the preprocessing code, the dimensionality reduction code, and the figure generation code in a version control system. Document the rationale for each parameter choice. The [Carpentries Lessons](https://carpentries.org/lessons) provide foundational training in version control, and the [Galaxy Training Network](https://training.galaxyproject.org/) offers accessible workflow training that emphasizes reproducibility.

## 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)
- [Single-Cell vs Single-Nucleus RNA Sequencing: Choosing the Right Approach](/knowledge/bioinformatics/single-cell-vs-single-nucleus-rna-sequencing-choosing-the-right-approach)
- [Benchmarking Atlas-Level Data Integration in Single-Cell Genomics: Methods and Best Practices](/knowledge/bioinformatics/benchmarking-atlas-level-data-integration-in-single-cell-genomics-methods-and-best-practices)
- [Multi-Omics Data Integration: A Comparative Framework for Choosing the Right Method](/knowledge/bioinformatics/multi-omics-data-integration-a-comparative-framework-for-choosing-the-right-method)
- [Spatial Proteomics Method of the Year: What It Means for Your Research](/knowledge/bioinformatics/spatial-proteomics-method-of-the-year-what-it-means-for-your-research)

## 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.
- [A generalization of t-SNE and UMAP to single-cell multimodal omics.](https://pubmed.ncbi.nlm.nih.gov/33941244). Genome biology, 2021.
- [Advanced Cell Mapping Visualizations for Single Cell Functional Proteomics Enabling Patient Stratification.](https://pubmed.ncbi.nlm.nih.gov/32108428). Proteomics, 2020.
- [Dimensionality reduction for visualizing spatially resolved profiling data using SpaSNE.](https://pubmed.ncbi.nlm.nih.gov/39960663). GigaScience, 2025.
- [AI and Machine Learning for Proteomics-Driven Drug Discovery: Methods, Tools, and Best Practices.](https://doi.org/10.3390/cimb48050532). 2026.
- [Visualizing stability: a sensitivity analysis framework for t-SNE embeddings.](https://doi.org/10.3389/fbinf.2025.1719516). 2025.

> This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.