Version Compatibility
Reference examples tested with: CATALYST 1.26+, diffcyt 1.22+, FlowSOM 2.10+, flowCore 2.14+, flowWorkspace 4.14+, flowStats 4.14+, edgeR 4.0+, limma 3.58+, ggplot2 3.5+; Python (partial alt) flowkit 1.1+.
Before using code patterns, verify installed versions match. If versions differ:
- R:
packageVersion('<pkg>') then ?function_name to verify parameters
- Python:
pip show <package> then help(module.function) to check signatures
If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt rather than retrying. Each stage defers depth to its component skill.
Flow Cytometry Pipeline
"Process my cytometry data from FCS to differential populations" -> read raw -> compensate/unmix -> transform -> QC -> remove doublets -> cluster (or gate) -> annotate -> test DA/DS, with the sample as the unit of inference.
- R:
flowCore + CATALYST::prepData/cluster/runDR + diffcyt::diffcyt()
The Single Most Important Modern Insight -- A Pipeline Is a Chain of Irreversible Decisions, and the Unit of Inference Is the Sample
Each early choice silently gates the validity of the final test: reading raw (not log-linearized), compensating BEFORE transforming, removing margin events before density QC, assigning type-vs-state markers correctly, and removing doublets before clustering. None of these is recoverable downstream - a doublet clustered as a "double-positive," a state marker used for clustering, or an uncompensated channel becomes a false population that the differential test then "confirms." The second critical thread is that the SAMPLE/subject, not the cell, is the experimental unit: diffcyt aggregates cells to per-sample-per-cluster counts (DA) and medians (DS) before testing, so biological replication (>= 2-3 per group) is mandatory and a per-cell test is invalid. Two normalization layers sit at different points in the pipeline - EQ-bead drift correction on raw counts at the very front (CyTOF), and CytoNorm cross-batch harmonization on transformed data before the analytical clustering (its internal FlowSOM clustering is part of the batch model, not the analysis) - and conflating them is a classic error.
Decision Tree: Which Path
| Situation | Path | Why |
|---|
| Conventional fluorescence flow | compensate ($SPILLOVER/flowStats) -> logicle -> ... | optical spillover; logicle handles negatives |
| Spectral cytometer (Aurora/ID7000) | UNMIX (not compensate) -> arcsinh ~150 | overdetermined system; fluorescence-scale |
| Mass cytometry (CyTOF) | EQ-bead normalize (raw) -> arcsinh cofactor 5 -> compCytof if needed | metals barely spill (~1-4%); drift correction first |
| High-dim discovery, no prior gates | cluster (FlowSOM via CATALYST) | scales; finds unexpected populations |
| Well-defined populations / rare events (MRD) | hierarchical gating (openCyto) | interpretable; clustering fails for ultra-rare |
| Multi-batch / multi-day | anchor sample per batch -> CytoNorm (normalize transformed data before analytical clustering) | model batch in the design for inference |
Pipeline Overview
FCS -> compensate/unmix -> transform -> QC (margins, time, dead) -> doublets
-> [ cluster (FlowSOM) | gate (openCyto) ] -> annotate -> diffcyt DA/DS -> report
EQ-bead drift normalization (CyTOF) runs on raw counts BEFORE everything; CytoNorm runs on transformed data and its normalized output feeds the cluster/gate step.
1. Panel, Metadata, and Load
Goal: Define the type/state panel and sample metadata, then load FCS.
Approach: Panel marker_class drives everything downstream (type clusters, state is tested); metadata keys samples to condition/subject. See flow-cytometry/fcs-handling.
library(CATALYST); library(diffcyt); library(flowCore); library(ggplot2)
panel <- data.frame(
fcs_colname = c('FSC-A','SSC-A','CD45','CD3','CD4','CD8','CD19','CD14','Ki67','IFNg'),
antigen = c('FSC','SSC','CD45','CD3','CD4','CD8','CD19','CD14','Ki67','IFNg'),
marker_class = c
md data.framefile_name list.files pattern
sample_id paste0
condition each
patient_id paste0
fs read.flowSetfile.path mdfile_name transformation truncate_max_range
2. Compensate / Unmix, then Transform
Goal: Remove spillover on linear data, then variance-stabilize.
Approach: Conventional flow compensates (matrix before transform); CyTOF skips fluorescence compensation and uses cofactor 5; spectral unmixes then uses ~150. See flow-cytometry/compensation-transformation.
spill <- spillover(fs[[1]]); spill <- spill[[which(!vapply(spill, is.null, logical(1)))[1]]]
fs_comp <- compensate(fs, spill)
COFACTOR <- 150
sce <- prepData(fs_comp, panel, md, transform = TRUE, cofactor = COFACTOR, FACS = TRUE)
3. QC (order matters)
Goal: Remove margin/boundary events and time anomalies before any density step.
Approach: Margins first, then time-based cleaning; on CyTOF, EQ-bead drift correction happens upstream on raw counts. See flow-cytometry/cytometry-qc and flow-cytometry/bead-normalization.
plotExprs(sce, color_by = 'condition'); pbMDS(sce, color_by = 'condition')
4. Remove Doublets
Goal: Drop aggregates before clustering so they don't form phantom double-positives.
Approach: Flow uses the FSC-A vs FSC-H diagonal; CyTOF uses DNA intercalator + Gaussian/Event_length. See flow-cytometry/doublet-detection.
e <- assay(sce, 'exprs')
if (all(c('DNA1','Event_length') %in% rownames(sce))) {
keep <- e['DNA1', ] > quantile(e['DNA1', ], 0.05) &
e['Event_length', ] <= quantile(e['Event_length', ], 0.99)
sce <- sce[, keep]
}
5. Cluster (FlowSOM) or Gate
Goal: Define populations by unsupervised clustering on TYPE markers (discovery) or hierarchical gating (defined/rare).
Approach: cluster() wraps FlowSOM+ConsensusClusterPlus; over-provision the grid, set a seed. See flow-cytometry/clustering-phenotyping (clustering) and flow-cytometry/gating-analysis (gating).
sce <- cluster(sce, features = 'type', xdim = 10, ydim = 10, maxK = 20, seed = 42)
6. Annotate and Visualize Structure
Goal: Label metaclusters from marker medians; embed for display only.
Approach: Median heatmap drives annotation; UMAP colors by cluster but is never used to define or quantify populations.
plotExprHeatmap(sce, features = 'type', by = 'cluster_id', k = 'meta20', scale = 'last')
sce <- runDR(sce, dr = 'UMAP', features = 'type', cells = 2000)
plotDR(sce, dr = 'UMAP', color_by = 'meta20')
7. Differential Abundance and State
Goal: Test which populations change in frequency (DA) or state-marker expression (DS) between conditions.
Approach: The diffcyt() wrapper aggregates to the sample level; results live in res$res. See flow-cytometry/differential-analysis.
design <- createDesignMatrix(ei(sce), cols_design = 'condition')
contrast <- createContrast(c(0, 1))
res_DA <- diffcyt(sce, clustering_to_use = 'meta20', analysis_type = 'DA',
method_DA = 'diffcyt-DA-edgeR', design = design, contrast = contrast)
res_DS <- diffcyt(sce, clustering_to_use = 'meta20', analysis_type = 'DS',
method_DS = 'diffcyt-DS-limma', design = design, contrast = contrast)
da <- as.data.frame(SummarizedExperiment::rowData(res_DAres
8. Visualize Results and Export
Goal: Summarize significant populations and persist results.
Approach: Pass the inner result object (res$res) to plotting; export tables and the SCE.
plotDiffHeatmap(sce, res_DA$res, all = TRUE, fdr = 0.05)
plotAbundances(sce, k = 'meta20', by = 'cluster_id', group_by = 'condition')
write.csv(da, 'da_results.csv', row.names = FALSE); saveRDS(sce, 'cytometry_analysis.rds')
Paired / Repeated-Measures Variant
Goal: Account for within-subject correlation (pre/post on the same donor).
Approach: Use a GLMM with a random effect for subject (NOT voom, which is fixed-effects only).
formula <- createFormula(ei(sce), cols_fixed = 'condition', cols_random = 'patient_id')
res_DA <- diffcyt(sce, clustering_to_use = 'meta20', analysis_type = 'DA',
method_DA = 'diffcyt-DA-GLMM', formula = formula, contrast = createContrast(c(0, 1)))
Manual Gating Path (alternative to clustering)
Goal: Define populations by a reproducible hierarchy when they are well-defined or rare.
Approach: Build a GatingSet on transformed data; recompute after adding gates. See flow-cytometry/gating-analysis.
library(flowWorkspace)
tl <- estimateLogicle(fs_comp[[1]], colnames(spill))
gs <- GatingSet(transform(fs_comp, tl))
recompute(gs); gs_pop_get_stats(gs, type = 'count')
Python Alternative (FlowKit) -- partial
Goal: Read, compensate, and gate in Python where an R pipeline is not an option.
Approach: FlowKit covers IO/compensation/GatingML; there is NO Python equivalent for diffcyt DA/DS, so the differential step stays in R (or bridge via readfcs -> AnnData -> scanpy for clustering only).
import flowkit as fk
sample = fk.Sample('sample.fcs')
sample.apply_compensation(sample.metadata['spillover'])
df = sample.as_dataframe(source='comp')
Per-Stage Failure Modes
Per-cell pseudoreplication
Trigger: testing across all cells. Mechanism: cells are not independent replicates. Symptom: p ~ 1e-40 from few subjects. Fix: diffcyt aggregates to sample level; require >= 2-3 replicates/group.
Clustering on state markers
Trigger: activation/phospho markers in features. Mechanism: state contaminates lineage identity. Symptom: activated/resting splits of one type. Fix: cluster on type; test state in DS.
Doublets / wrong cofactor / uncompensated input
Trigger: skipping doublet removal, cofactor 5 on fluorescence, or clustering raw data. Mechanism: phantom double-positives, compressed dim markers, spillover-dominated distances. Symptom: non-reproducible "novel" populations. Fix: remove doublets first; cofactor 5 (CyTOF) / 150 (fluorescence); compensate+transform before clustering.
Batch cleaned instead of modeled
Trigger: CytoNorm-ing then testing naively, or batch confounded with condition. Mechanism: over-correction / non-identifiability. Symptom: attenuated or fabricated effects. Fix: model batch in the design; if batch == condition, no rescue.
Quantitative Thresholds
| Threshold | Source | Rationale |
|---|
| arcsinh cofactor 5 (CyTOF) / ~150 (fluorescence) | Nowicka 2017 F1000Res 6:748 | matches platform noise scale |
| >= 2-3 biological replicates per group | Weber 2019 Commun Biol 2:183 | minimum for a valid DA/DS error term |
| > ~10K cells per sample | community | stable per-sample cluster frequencies |
| 10-30 metaclusters typical (maxK=20 default) | Weber & Robinson 2016 Cytometry A 89:1084 | over-provision then merge |
| BH FDR across clusters (and clusters x markers for DS) | diffcyt | many simultaneous tests |
Common Errors
| Error / symptom | Cause | Solution |
|---|
testDA_edgeR(sce, ...) not found / wrong | fabricated signature | use the diffcyt() wrapper; results in res$res |
compensate() errors / silent NULL | spillover(ff) returns a 3-slot list; the matrix is often under SPILL/$SPILLOVER, not [[1]] | select the first non-null slot, not positional [[1]] |
| empty DS results | state markers not flagged | set marker_class='state' in the panel |
| paired design ignored | used fixed-effect method | diffcyt-DA-GLMM with a random effect |
References
- Weber 2019 Commun Biol 2:183 — diffcyt DA/DS framework.
- Nowicka 2017 F1000Research 6:748 — CATALYST CyTOF workflow; type/state, cofactor 5.
- Weber & Robinson 2016 Cytometry A 89(12):1084-1096 — FlowSOM clustering benchmark.
- Van Gassen 2020 Cytometry A 97(3):268-278 — CytoNorm cross-batch normalization.
- Hurlbert 1984 Ecol Monogr 54(2):187-211 — pseudoreplication (sample is the unit).
Related Skills
- flow-cytometry/fcs-handling - Read FCS and map channels
- flow-cytometry/compensation-transformation - Compensate/unmix and transform
- flow-cytometry/cytometry-qc - Time/margin/dead-cell QC
- flow-cytometry/doublet-detection - Singlet discrimination
- flow-cytometry/bead-normalization - EQ-bead drift and CytoNorm batch correction
- flow-cytometry/gating-analysis - Hierarchical/automated gating path
- flow-cytometry/clustering-phenotyping - FlowSOM clustering and annotation
- flow-cytometry/differential-analysis - diffcyt DA/DS testing
- single-cell/clustering - Related graph-clustering for scRNA-seq
