March 15, 2019  |  FlowSOM, viSNE  |  By  |  0 Comments

Beginner’s guide to FlowSOM: Profiling the innate immune response to viral infection

FlowSOM detailChikungunya virus is transmitted to humans in tropical regions through the bite of a mosquito carrying the virus. Pediatric populations are most at risk for developing acute illness of fever, rash, and joint pain, often followed chronic, debilitating arthritis. With no available vaccine or effective anti-viral treatment, a comprehensive immunophenotyping study was performed by Michlmayr et. al. on PBMCs of 42 children during acute and convalescent phases of infection, to better understand the innate anti-viral immune response.

Using 35-parameter CyTOF data available on ImmPort (SDY1288), we applied Cytobank’s machine learning tools viSNE followed by FlowSOM to simplify visualization, then cluster cells and identify which clusters were altered during acute infection. FlowSOM is a powerful algorithm that builds self-organizing maps to provide an overview of marker expression on all cells and reveal cell subsets that could be overlooked with manual gating. It can be implemented at many points in your analysis workflow: prior to, after, or even instead of manual gating. Here, we demonstrate how FlowSOM can be used along with viSNE and manual gating to accelerate the process of biological discovery. Cytobank makes advanced machine learning tools like these easily accessible to scientists, regardless of computational biology or programming experience.


Analysis steps

1. Manual gating

If you know you have specific populations that are important to you, start here. We performed manual gating of major populations as selected by Michlmayr et. al. This can serve as a benchmark for FlowSOM performance.

2. Dimensionality reduction with viSNE

We ran viSNE to a) visualize a high-level overview of the varying cell phenotypes present in our samples, and b) inform parameter settings for FlowSOM. viSNE revealed about nine distinct islands, representing populations or clusters of cells, similar to the viSNE maps generated by Michlmayr et. al.

a) Visualization of manually gated populations on viSNE map

To quickly see how well viSNE separated known manually gated populations, we overlaid manual gates on the viSNE map (Fig. 1). Manually-gated populations separated well by islands, except: 1) a portion of CD8+ T cells and NKT cells, which have similar effector functions and surface markers, overlapped by viSNE, and 2) cells gated as CD11c+ myeloid dendritic cells (mDC) separated into two viSNE islands. As the CyTOF panel focused heavily on monocyte and dendritic cell markers to refine and identify nuances in subpopulations of these cells previously known to be involved in the innate response to Chikungunya, the viSNE clearly distinguished these subpopulations of mDCs.

Manually gated populations overlaid on viSNE map
Fig. 1. Manual gating of major lymphocyte subsets overlaid on viSNE. viSNE was run using all FCS files and all marker channels. A representative subject is shown at acute (left) and convalescent (right) timepoints.

b) viSNE map informs FlowSOM parameters

We used the viSNE map to inform the setup of a FlowSOM analysis for identification of cell clusters and metaclusters. A good starting point for the metacluster number can be taken from the viSNE map, based on the number of islands. It is recommended to err on the side of too many rather than too few clusters, as clusters can be combined more easily if needed, but not subdivided. We suggest repeating FlowSOM multiple times with different random seeds to ensure reproducibility.


3. Run FlowSOM to automatically group cells by clusters and metaclusters

FlowSOM performs four steps behind-the-scenes:  

  1. Prepare data: Read in FCS files, compensate, transform, concatenate, and scale
  2. Build the self-organizing map (SOM) to generate an overview of the data
  3. Build a minimum spanning tree (MST) by connecting the nodes/clusters of the SOM
  4. Metacluster the related nodes/clusters of the SOM

FlowSOM results:

FlowSOM performs all these steps very quickly. In as little as a couple minutes, FlowSOM generates several simple yet information-rich visualizations and spreadsheets to download, and adds metacluster and cluster channels to your cell events for downstream analysis in a new Cytobank experiment.

Assessment FlowSOM metacluster assignments

We assessed how well FlowSOM metaclustered the data as compared to manual gating of major cell populations, by a) using FlowSOM’s built-in pie chart visualizations within the minimal spanning tree or grid output, and b) mapping FlowSOM metaclusters onto viSNE to see how the metaclusters compare to separation of islands.

a) FlowSOM Minimum Spanning Tree

The results file output of FlowSOM in Cytobank includes MSTs composed of the user-selected number of clusters and metaclusters. Each cluster contains a pie chart, where each pie slice indicates the proportion of that pie in a selected manually gated population. Each pie has a background halo color indicating its metacluster assignment. Finally, the distance of clusters from each other in the tree shows their similarity to each other. We see that the manually gated populations generally fall within a single metacluster (Fig. 2), indicating good agreement between FlowSOM’s metaclustering and manual gating.

The MSTs show striking differences in metacluster 2 during acute infection, which expanded and contained cell types gated as non-classical and intermediate monocytes that were not present in the convalescent sample (Fig. 2, top). An alternative view is a grid of clusters and metaclusters (Fig. 2, bottom), which can help visualization by spreading out overlapping elements.

Fig 2

FlowSOM MST and grid
Fig. 2. FlowSOM results for one representative subject during acute (left) and convalescent (right) timepoints as MST (top) and grid (bottom). FlowSOM was performed using 225 clusters, each represented by one pie chart, and 9 metaclusters, denoted by background shading. Most manually gated populations appear within a single metacluster, indicating similar outcomes between manual gating and FlowSOM clustering. Major differences observed in metacluster 2 between acute and convalescent samples are outlined in blue.

b) FlowSOM metaclusters map to distinct viSNE islands consistent with manual gating

A second method to visually assess FlowSOM performance is to map metaclusters back onto viSNE. Here, the metaclusters clearly separated by viSNE islands, with the exception of metaclusters 5 (purple), 7 (pink), and a small portion of metacluster 1 (blue), which appeared in two separate islands (Fig. 3).

FlowSOM overlay on viSNE
Fig. 3. Top: FlowSOM metaclusters overlaid on viSNE map.
Bottom: CHIKV E2 expression in greyscale overlaid on viSNE map.

Comparison of metaclusters to selected manual gates revealed a significant degree of congruence. Most notably, metacluster 2 (Fig. 3, orange) completely overlapped with CD14+ monocytes (Fig. 1, orange), which contained the majority of Chikungunya-infected cells (Fig. 3, bottom). Metacluster 5 (Fig. 3, purple) overlapped with CD56+ NK (Fig. 1, pink) as well as NKT cells (Fig. 1, lime). Metacluster 3 (Fig. 3, green) overlapped well with CD123+ pDCs (Fig. 1, brown).

In one case, FlowSOM metaclusters did not separate manually gated populations: both CD4+ and CD8+ T cells were assigned to metacluster 7 (Fig. 3, pink), a reflection of the panel designed to investigate the innate response, particularly monocytes. If T cells were of specific interest in this study, the inclusion of additional T cell markers would likely have allowed their separation by FlowSOM by giving more weight to differences between T cell populations. To ensure that FlowSOM could separate CD4+ and CD8+ T cells, we provided more weight to CD4 and CD8 markers by running FlowSOM on only CD3+ cells, using only those markers. Indeed, CD4+ and CD8+ cells then separated by metacluster. If this separation was important to us when considering all PBMCs, we could gradually increase the markers selected for inclusion by FlowSOM.

On the other hand, FlowSOM enhanced sensitivity to further separate manually gated CD19+ B cells. Metaclusters 1 (Fig. 3, blue) and 4 (Fig. 3, red) mapped to two separate viSNE islands that were both manually gated as CD19+ B cells. Further gating on more specific B cell subsets may identify how they were separated by viSNE. Overall, FlowSOM’s unbiased identfication of metaclusters coincided well with gates defined using prior immunological knowledge as well as unbiased separation by viSNE.

Findings: FlowSOM highlights populations of interest in viral infection

Both the MST or grid visualizations as well as the overlay of metaclusters onto viSNE gave clear messages about the immune response during acute infection:

  1. A shift in the subpopulations that make up metacluster 2, (Fig. 2, circled clusters)
  2. A significant expansion of metacluster 2 (Fig. 3, orange)
  3. Chikungunya-infected cells belonged to metaclusters 1 and 2

With clear directions on which metaclusters are altered during viral infection, we chose to hone further analysis to these metaclusters. Particularly since we observed that some CD19+ B cells in metacluster 1, but not metacluster 4, were CHIKV E2+, we wanted to investigate these metaclusters further, to provide insight into viral tropism in the human host. To visualize the differences between FlowSOM metaclusters 1 and 4, we opened the resulting FlowSOM experiment that contained the FCS files with additional FlowSOM cluster and metacluster channels in Cytobank, then overlaid histograms of selected B cell surface markers. Metacluster 1, which contains virus-infected cells, had higher expression of CD20, CXCR5, CCR6, and CD1c than metacluster 4 (Fig. 4). You can find guidance for how to perform statistical comparisons of metaclusters here.

Acute vs. Convalescent
Fig. 4. Comparison of FlowSOM metaclusters 1 and 4 by expression of B cell surface markers. (Click to zoom)

The overall concordance between manual gating and FlowSOM metaclustering seen in the MST and viSNE map demonstrates that FlowSOM is an effective unbiased tool to explore cell populations in the absence of manual gating, which can get increasingly overwhelming with large panels such as this one. The use of FlowSOM easily refines future research focus to metaclusters 1 and 2 to delve into understanding the innate anti-viral response, without the need to do extensive manual gating or statistical comparisons of individual populations thought to be important.

FlowSOM directs future analysis

If only a subset of samples were available for initial analysis while sample collection was ongoing, we could apply new samples to the SOM created on this initial data set, providing reproducible analysis over time. This is a major benefit of FlowSOM for prospective and interim studies, like clinical trials with long-term follow-up, whether for vaccines, immunotherapies, or drug trials.

The powerful visualization tools are a central strength of FlowSOM, however, are only a small piece of FlowSOM results provided Cytobank. The complete FlowSOM output empowers you to visually and statistically compare study samples. Results also include:

  • MSTs with star charts to visually compare median marker expression levels in each cluster between samples
  • MSTs colored by median marker expression within each cluster to quickly visualize cluster phenotypes
  • Grids of clusters with pies and star charts colored by metacluster
  • Spreadsheets with marker expression values for each cluster and metacluster
  • Spreadsheets of abundance of each cluster and metacluster
  • Records of all your run settings, to ensure reproducibility


To get started with Cytobank, sign up for a free 30-day trial. For more detailed documentation on how to perform and interpret your FlowSOM analysis, see our support site and don’t hesitate to contact our support team with questions. You also check out our video tutorial and webinar with FlowSOM developer Sofie Van Gassen. If you are looking for assistance on a large or complex project, get in touch with our Scientific Services team!



Michlmayr D, Pak TR, Rahman AH, Amir ED, Kim EY, et al. Comprehensive innate immune profiling of chikungunya virus infection in pediatric cases. Mol Syst Biol. 2018 Aug 27;14(8):e7862. PubMed PMID: 30150281; PubMed Central PMCID: PMC6110311.

Van Gassen S, Callebaut B, Van Helden MJ, Lambrecht BN, Demeester P, et al. FlowSOM: Using self-organizing maps for visualization and interpretation of cytometry data. Cytometry A. 2015 Jul;87(7):636-45. PubMed PMID: 25573116.