Analysis
Clinical Flow
SuperCellCyto: compression needs a cell-count audit
A 2024 research method groups similar cytometry events to reduce computational load, but preserving a cell map and checking local rare populations remain essential.
High-dimensional cytometry can produce millions of events. A clustering workflow that is practical for one sample may become slow or memory intensive across a study. Randomly keeping a fraction of the cells makes the calculation smaller, but it may also remove an uncommon population before the analyst has a chance to inspect it. SuperCellCyto offers a different research strategy: aggregate similar events into smaller units while retaining a map back to the original cells.

Putri and colleagues described this R package in Genome Biology in 2024. Their evaluation used six public cytometry datasets and several downstream analytical tasks. It is useful evidence about computational analysis; it is not validation of a clinical test or a guarantee that a particular rare population will survive every compression setting.
What is compressed?
After cytometry preprocessing and transformation, SuperCellCyto performs principal-component analysis and constructs a nearest-neighbour graph. Closely related cells are grouped into supercells within each sample. For each marker, the supercell is represented by a summary of the expression values in its member cells. The package also returns a map from each original cell to its assigned supercell.
The parameter gamma controls the nominal ratio between the numbers of cells and supercells. A larger gamma creates fewer aggregated units, reducing the data passed to subsequent analyses. The trade-off is that a supercell becomes more likely to contain biologically different cell types. The authors used gamma 20 as a default, not as a universally validated setting for every panel or clinical question.
An important practical detail is that supercells are not guaranteed to contain equal numbers of cells. If one supercell represents 10 events and another represents 40, treating them as equal units can distort an abundance calculation. The map back to individual cells is therefore part of the result, not disposable bookkeeping.
What the examples show
For the Samusik_all dataset, the authors report reducing 841,644 single cells to 42,082 supercells at gamma 20, approximately a twentyfold reduction. They compared supercell composition and clustering with existing annotations. Average purity was above 0.9 across the examined compression settings for two datasets, but purity and adjusted Rand index were calculated only for the cells with available annotations: 39% of Levine_32dim and 61% of Samusik_all. These numbers describe the evaluated subset, not every event.
The paper also examined a healthy human bone-marrow dataset containing more than eight million B cells. At gamma 20, the researchers obtained 415,711 supercells and reported identifying all ten B-cell subsets in that analysis, including 162 plasma cells, or 0.002% of the individual cells. In a comparison using an equally sized random sample of individual cells, they reported finding seven of the ten subsets. This is an informative example, not an estimate of universal rare-event sensitivity. Tissue, gating, panel design, preprocessing, clustering parameters and the biological definition of a population all matter.
Return to cells before comparing proportions
The authors recommend expanding annotated supercells back to the individual-cell level before calculating proportions for differential-abundance testing. In their melanoma dataset, different supercells represented different numbers of cells, and the statistical results varied between supercell-level and cell-level abundance analyses.
That recommendation addresses a counting problem, not every statistical problem. Cells from one specimen are not automatically independent patients. The relevant biological replication, sample structure and uncertainty must still be respected when comparing groups. Analysts should retain the sample identifiers, original event counts, supercell membership, annotation method and filtering decisions needed to reconstruct the calculation.
A practical evaluation before use
A laboratory or research group considering the method could compare several gamma settings on the same representative data. For each setting, inspect the number of original events recovered for populations that matter, their marker distributions, cluster composition and agreement with a defined reference workflow. Check whether changing preprocessing, graph parameters or the sample mix changes those findings. Report the analysis and hardware when discussing runtime: a reduction observed in one benchmark is not a performance promise for every system.
Clinical rare-event work requires additional caution. This publication does not establish a limit of detection, limit of quantification, clinical sensitivity or multicentre reproducibility for measurable residual disease. It also does not show that a single compressed view can replace examination of original events or local assay validation. A useful exploratory shortcut becomes trustworthy for a defined application only after the complete local workflow has been assessed for that purpose.
Daily Learning Pill: Compression changes the analytical units. Preserve the cell-to-supercell map, count the original cells when estimating abundance, and test the populations your actual question depends on.
Related CFCM analysis: AI-assisted rare-event interpretation in B-ALL.
Primary evidence: Putri GH, Howitt G, Marsh-Wakefield F, Ashhurst TM, Phipson B. SuperCellCyto: enabling efficient analysis of large scale cytometry datasets. Genome Biology 25, 89 (2024). DOI 10.1186/s13059-024-03229-3.