When to Trust a FlowSOM Cluster: Interpreting Unsupervised Clustering in Flow Cytometry

FlowSOM hands you twenty metaclusters and a tidy star-chart tree. Cluster 14 has a phenotype label and a clean position on your UMAP. The temptation is to treat it as a discovered cell population and put it in the figure. But an unsupervised algorithm will always return clusters—whether or not they correspond to anything biological. The work is deciding which ones to trust.

This is a decision framework for reading FlowSOM output: the factors that separate a real population from an artifact, the conditions under which a cluster earns its place in your analysis, and the conditions under which it needs more work first. FlowSOM is a strong tool for surfacing populations manual gating would miss—but every cluster is a hypothesis until you check it.

Factors that decide whether a cluster is real

Four things separate a trustworthy metacluster from a convenient one:

• Back-gating coherence. When you project the cluster’s events back onto the original bivariate plots (and onto a UMAP or tSNE map), do they land in a tight, sensible region—or scatter across unrelated areas?
• Marker definition. Is the cluster defined by a clear on/off difference in a few markers, or only by small shifts that could be noise or spreading error?
• Biological plausibility. Does the phenotype correspond to a population that should exist in this sample, with a marker combination that makes immunological sense?
• Stability across settings. Does the cluster survive when you change the metacluster count, or does it appear and disappear depending on where you set the target number?

Path A — Trust the cluster (with the usual caveats)

A metacluster is ready to use when it back-gates to a coherent region, is defined by distinct markers rather than marginal shifts, matches a biologically expected phenotype, and stays put when you nudge the cluster count up or down. This is the case FlowSOM is built for: it found a population that manual gating could have missed, and the population checks out under inspection.

Even here, treat the algorithm’s phenotype label as a starting description, not a verified identity. Confirm it back-gates the way the label implies. Overlaying metaclusters onto a dimensionality-reduction map is the standard coherence check—the interpretation of those maps is covered in interpreting tSNE and UMAP.

Path B — Don’t trust it yet

Hold off when any of these are true:

• The cluster splits on a technical axis. If a cluster separates by a scatter parameter, a time-related artifact, or a fluorochrome known to spread, it may be tracking instrument behavior rather than biology.
• It only appears at one cluster count. A population that exists at 25 metaclusters but vanishes at 15 and 30 is likely an over-clustering fragment, not a stable subset.
• Back-gating scatters. If the events land in several unrelated places on the original plots, the cluster is grouping cells by something other than a shared real phenotype.
• The phenotype label is low-confidence. When the top discriminating markers are close in power, the auto-generated label is a guess. Cytomaton marks these with a confidence indicator and a “not biologically validated—verify by back-gating” caveat for exactly this reason.

The fix for most Path B clusters is the same: intentionally over-cluster, then merge. Running more metaclusters than you expect populations, and checking each for marker unimodality before merging related ones, is a recognized way to avoid both over- and under-splitting.

The batch-effect trap

One failure mode deserves its own mention because it’s easy to miss: clusters that separate samples by acquisition day rather than by biology. If a metacluster is almost entirely cells from one batch, it may be encoding an instrument or staining-day difference. This is a real risk when you cluster across samples run on different days—the same instrument-drift problem tracked in flow cytometry QC across instrument runs. Check the per-sample composition of any cluster before interpreting it as a population.

Summary: a quick decision table

FlowSOM works best as a discovery layer over a clean gating strategy, not a replacement for one. It can surface a rare subset you would never have drawn a gate for—see detecting rare cell populations—but the cluster still has to survive back-gating before it’s a finding. Note that implementations differ: Cytomaton’s clustering uses a self-organizing-map implementation that may assign cells slightly differently from the original R package, so cluster boundaries aren’t identical across software.

Cytomaton runs FlowSOM-style clustering on gated populations with adjustable resolution, auto-generated phenotype summaries, and a back-gating bridge to verify each cluster against the original plots. The method’s foundations are in Van Gassen et al., “FlowSOM: Using self-organizing maps for visualization and interpretation of cytometry data” (Cytometry Part A, 2015).