Dimensionality Reduction and Clustering in Flow Cytometry: A Practical Guide

Written by: Eric Haas, MEng, July 17, 2026

Summary

• Dimensionality reduction is a visualization tool; clustering is a grouping tool.
• Use DR to see patterns across samples. Use clustering to identify populations.
• Common DR algorithms: UMAP, opt-SNE, tSNE, PCA.
• Common clustering algorithms: FlowSOM, CITRUS, ClusterX.

What is high-dimensional cytometry?

High-dimensional cytometry is a single-cell proteomics method that enables the simultaneous measurement of 25 to 60 proteins per cell. The two most common technologies that enable this method are spectral and mass cytometry. Many technologies enable 40 to 50+ markers in a given assay, yielding high-dimensional datasets with many cells and many features. In contrast, traditional flow cytometry assays generally measure 15 or fewer markers per assay.

What's challenging about analyzing high-dimensional cytometry data?

The complex nature of high-dimensional cytometry data poses challenges for traditional analysis strategies used in lower-dimensional assays. In a traditional flow cytometry dataset with 10 markers, most populations can be easily accounted for using manual gating techniques. As the dimensionality increases, the granularity of information that can be gleaned from the data increases as well.

Of course, manual gating plays a role in high-dimensional data analysis and is often used for

• Cleanup and preprocessing
• Parent population identification (lymphocytes, live cells, singlets)
• Upstream analysis tasks

Often, manual gating is also used in long-form gating strategies to identify multiple populations of interest using well-defined criteria; however, this approach quickly becomes tedious, laborious, and difficult to manage. The other issue with manual gating of high-parameter cytometry data is that, even with the most complex manual gating strategies, the information available across all markers and cell types may not be captured by gating alone.

This is where dimensionality reduction and clustering can be helpful.

How are dimensionality reduction and clustering similar? What is the difference between them?

While many people often confuse dimensionality reduction (DR) with clustering, assuming they all fall under the “clustering” umbrella, this is simply a common misconception. While these tools are often used together in the same cytometry data analysis workflow, DR and clustering are distinct unsupervised machine learning (ML) algorithms.

Dimensionality reduction and clustering are similar in this context in that these tools can be used to connect datasets in their high-dimensional space and extract information from multiple dimensions into something more interpretable and usable for the human brain.

Both DR and clustering are useful tools for high-dimensional data analysis, not just for cytometry data. These tools are used in other biological modalities and in other types of data analysis more broadly, and each serves a unique role. DR allows for a more surface-level understanding of your data's high-dimensional space. Clustering enables deeper classification of a dataset's high-dimensional space and provides an understanding of the data's actual composition, not just the surface-level view that DR can provide.

As shown in the illustration below, DR can project cells onto a two-dimensional layout, and clustering groups cells into populations that can then be visualized on the DR layout. The difference between dimensionality reduction and clustering is that dimensionality reduction is useful for visualization, whereas clustering groups cells into populations.

Figure 1. Concepts of dimensionality reduction and clustering, with simulated data.

What is dimensionality reduction?

Dimensionality reduction is a set of algorithms that project high-dimensional cytometry data into two or three dimensions for visualization.

DR algorithms take input dimensions, i.e., the markers you select to feed into it, compute a distance/similarity metric between each pair of events in the high-dimensional space, and then project the high-dimensional distances into TWO (sometimes three) dimensions. In essence, DR takes the many-parameter space and converts it into an inferred, lower-dimensional space that is easier for the human brain to visualize and digest.

This matters because we humans can think in two dimensions fairly intuitively. In effect, DR makes a seemingly abstract feature space more concrete for our minds to work with by condensing the multi-feature space into typically two features, which can then inform us about the nature and structure of the multi-feature latent space. This is very useful for visualizing general differences across time, space, conditions, etc., among samples in a dataset and can help inform further investigation of the data.

How is dimensionality reduction used in flow cytometry?

Figure 2. Example use of DR to visualize trends in CAR T cell immunometabolism across different compartments (Peripheral Blood, Bone Marrow, Cerebrospinal Fluid) at the same time point. The colored contours represent the corresponding compartment, and the gray contours in each plot represent the other compartments analyzed. This figure was created from a UMAP DR run in OMIQ and published in Cancer Research. [1]

Dimensionality reduction is often used in flow cytometry data analysis to assess trends across comparison groups by examining high-dimensional shifts that are hard (often impossible) to assess from biaxial plots alone. With flow cytometry panels spanning dozens of markers, traditional visualization methods fail to capture the heterogeneity and complexity of the protein interactions at play. DR enables visual identification of complex protein interactions.

For example, in the dimensionality-reduction plots above, CAR T-cell cytometry data from three different compartments are compared visually. This DR run was performed using twelve metabolic and functional markers as the input to the algorithm. UMAP was the DR algorithm used here, which took the input data from the twelve markers and projected the 12-dimensional space into the two dimensions shown. This enables a visually interpretable assessment of the differences between these three groups in all 12 dimensions.

What can be determined from a dimensionality reduction visualization?

From this plot, it is clear that PB, BM, and CSF (day 28) each occupy distinct positions in the high-dimensional space; that is, the expression of and interactions among the 12 input markers differ among the three groups. We can expand on this to say that these groups must have differences in their metabolic and/or functional activity. Dimensionality-reduction visualizations can reveal the overall trend in differences among the input parameters.

What we cannot say from the above figure is which proteins are differentially expressed, the extent of their expression, which proteins are expressed in a coordinated fashion, or even how different the protein expression levels are between groups. All we can say with confidence from this plot is that there are clearly differences in some way in or across the 12 input proteins between the three groups.

This brings us to a few key points about dimensionality reduction:

1. Through the reduction of features/compression of data, some feature information may be lost in the process, so it is critical not to take the results of a DR algorithm to be the ground truth of the dataset.
2. DR algorithms do not group events/cells in the way gating (or clustering) does. After the run of a DR algorithm, we get two (or more) features that represent the multi-dimensional input space, yet we do not get an actual grouping of this data in any meaningful way for statistical testing.
3.  DR algorithms are inherently visualization tools in the cytometry context.
4. Depending on the algorithm and run features, just because two events/cells fall in proximity in the reduced space does not inherently mean that they are the nearest neighbors in high-dimensional space, so gating in the dimensionality reduction embedding is not advised.
5. DR is useful for answering the question of “Are there noticeable patterns or differences between the samples or groups in a given dataset?”
6. DR is also an effective tool for quick quality control (QC) of a dataset. For example, if samples of the same type and condition collected over different batches occupy completely different regions in the DR embedding, there are likely batch effects affecting the data.

What is clustering?

Clustering is a set of algorithms that automatically group cells into populations based on marker expression similarity across high-dimensional space.

There are several clustering algorithms, and they all work in conceptually similar ways. As with DR, a clustering algorithm takes a selected set of parameters or features as input, calculates distances between events in the multi-parameter space, and, based on the similarity or multi-parameter differences between events, groups them into distinct populations.

How is clustering used in flow cytometry analysis?

Clustering, as employed in flow cytometry data analysis, is used to automatically group cells as specific populations. To get a better practical understanding of clustering, it is helpful to compare it to manual gating, the canonical flow cytometry analysis method.

In manual gating, thresholds are set manually for each parameter to define positive or negative expression (thus separating distinct populations). In clustering, an algorithm computes cell similarity using distance metrics in high-dimensional space and then groups similar cells, separating dissimilar cells into other groups. These groups are typically referred to as clusters or metaclusters, depending on the method and terminology used.

What can be determined from clustering?

From a clustering run, assuming it was performed properly and accurately, many things can be determined from the resulting clusters, as the amount of information that can be learned about a dataset through clustering is immense.

The clusters produced by a clustering algorithm can be analyzed in the same way as traditional, manually gated populations, including, but not limited to, abundance statistics and marker expression in each population. This makes clustering useful for understanding population-level dynamics in a dataset and for discovering rare or unknown populations, which can then be confirmed via manual gating. Clustering can also be helpful for understanding the multi-parameter expression profiles of previously known populations; for instance, T cells can be classified into various memory subsets and then clustered to determine functionally distinct characteristics of these populations in greater depth.

This brings us to a few key points about clustering in flow cytometry:

1.  Clustering provides an efficient method of cell grouping that has the potential to capture heterogeneity often missed through manual gating alone.
2.  Clustering may lead to spurious populations, depending on whether the data is over- or under-clustered, meaning some algorithms may group cells into more complex groupings than could be reasonably expected, or they may group cells into fewer populations than reasonably expected and thus confound some populations of cells.
3.  Using these algorithms with caution and discernment is critical, as the number of clusters identified for a given dataset can vary depending on the method and inputs used. Most clustering algorithms have built-in or commonly used best practices for determining the number of clusters in the data.
4. Clustering goes beyond data visualization and allows for true analysis of the data, and is often used in conjunction with DR to visualize the distribution of clusters
5. Clustering is useful for answering the question of “What populations are present in a dataset, and are there populations which I cannot resolve through manual gating?”

In flow cytometry analysis, when should dimensionality reduction be used and when should clustering be used?

In general, dimensionality reduction should be used when an appealing visualization of high-dimensional data is desired, i.e., to visually compare differences between groups in a dataset. Clustering should be used when manual population identification is insufficient for capturing the full heterogeneity of the data or when novel populations are of interest. Oftentimes, DR and clustering are used together in high-parameter workflows. DR is used to visualize data, and clusters can often be shown in DR space to indicate where they sit in the DR embedding, which is helpful for determining population dynamics (for example, both are used for certain trajectory algorithms).

The following decision tree may help determine when to perform DR or clustering.

Commonly used Dimensionality Reduction Tools

Table 1. Examples of commonly used dimensionality reduction algorithms: a. PCA, b. UMAP, c. OptSNE

Commonly used Clustering Tools

Table 2. Examples of commonly used clustering algorithms: a. FlowSOM, b. CITRUS, c. ClusterX

Example workflow in OMIQ

First and foremost, before starting a workflow, it is helpful to have a plan for how the data will be analyzed and which tools will be used. It is important, especially for high-parameter cytometry data, to develop an analysis plan, ideally in advance of the experiment. If this plan is in place in advance, implementing it is relatively straightforward. If no plan is made in advance, a simple workflow heuristic can be implemented for most discovery datasets. Most discovery workflows use both dimensionality reduction and clustering.

To put it simply, most workflows utilize the following steps:

1. Compensation/Unmixing: For traditional and spectral flow cytometry data, the first step of analysis preprocessing should be compensation or spectral unmixing, respectively.
2. Scaling: Enables accurate visualization and data properties, i.e., making cytometry data values more linear, which improves ML algorithm performance.
3. Data clean-up: Another critical pre-processing step is data clean-up, which often consists of running QC algorithms like PeacoQC or flowAI, which remove run-based artifacts from the data and clean up gating.
4. Population gating and Subsampling: The next typical step is to perform manual gating to identify the populations of interest in the dataset, which often serve as the starting point for running more advanced algorithms such as DR and clustering. After the parent population is identified, subsampling is used to select a subset for downstream analysis.
5. Dimensionality reduction: This needs to be performed as a child task stemming from the subsampling task selecting the population of interest. In OMIQ, the samples to be analyzed are selected (note: if a sample is not selected at this step, the DR algorithm will not be applied to it, and it will not be available for analysis). Next, the features/markers that will be input to the algorithm are selected. Typically, in OMIQ, the default algorithm settings work for most datasets; however, it is advisable to experiment with the settings for each dataset to find those that optimize the outcome.
6. Clustering: Clustering can be performed before or after DR; however, it is often easier to follow the workflow logic to perform DR upstream of clustering. A similar setup is found in a clustering task to DR tasks in OMIQ. First, the samples to be used must be selected. The input parameters are selected; often, the same parameters used for DR are a reasonable choice for clustering, which can be copied and pasted from one task to the other. Again, the default settings in OMIQ tend to work well for most datasets. In both DR and clustering, random seed settings are critical to keep consistent when running the same algorithm repeatedly.
7. Using Clustering Results: In order to use clustering results in OMIQ, the clusters must be set as a categorical filter in a gating task downstream from the clustering run. These filters can then be used downstream for further analysis.
8. Downstream analysis: Typically, the next steps are to create a clustered heatmap or special x-axis plots, showing all samples and all clusters, with the median expression of the input parameters, and gain insight into what the clusters are expressing. Another downstream task is to export statistics to obtain cluster abundances for each sample for comparative analysis. Comparative analysis can also be performed between two groups using EdgeR in OMIQ, which compares differential abundance between the groups for the selected filters.
9. Figure creation: An important part of the analysis process is to create figures illustrating the analysis performed. In this sort of workflow, a common plot is to create a DR embedding with clusters selected as filter overlays, so the distribution of clusters in the reduced, high-dimensional space can be visualized. Another commonly used plot type is a bar or box plots showing the abundance of clusters within a group or sample. Boxplots in OMIQ are also useful for illustrating expression data of a filter.

OMIQ Workflow Video Walkthrough

To demonstrate this workflow, let’s explore an example dataset from a CyTOF run, using the Maxpar Direct Immune Profiling Assay™ (MDIPA), a broad immunophenotyping assay. This dataset consists of six samples: three donors compared across two conditions.

Follow along with the video below for a full walkthrough of the example workflow using UMAP dimensionality reduction and FlowSOM clustering, with elbow-method meta-clustering to find the optimal number of clusters in these data. The walkthrough also shows how to use the results of these algorithms to extract information from the data.

Why is OMIQ good for flow cytometry data analysis?

OMIQ has 30+ natively integrated algorithms, with several features that enable advanced high-dimensional cytometry analysis, making it a more favorable platform for such data. Among these features are virtual concatenation, customizable metadata handling, cloud-based performance, and traceable analysis workflows.

Performing flow cytometry data analysis in OMIQ uses a cloud-hosted, task-based, logically ordered workflow that offers many advantages over older cloud-based services and local analysis software. DR and clustering algorithms are included as selectable tasks with easy access to modify settings and no need to configure plugins. Because these algorithms are selectable tasks native to an OMIQ workflow, the traceability of the analysis steps performed is much cleaner in OMIQ than in other platforms, where the analysis steps can get lost in the sample window pane, which happens readily when many tasks are running in parallel.

Try OMIQ today, with a free 30-day trial, no installation, and the algorithms in this article are all available in the workflow from day one.