Doublet detection

Version Compatibility

Reference examples tested with: flowCore 2.14+, CATALYST 1.26+, ggplot2 3.5+.

Before using code patterns, verify installed versions match. If versions differ:

• R: packageVersion('<pkg>') then ?function_name to verify parameters

If code throws an error, introspect the installed package and adapt rather than retrying.

Doublet Detection

"Remove doublets from my cytometry data" -> Discriminate single cells from aggregates using pulse geometry (flow) or ion-cloud parameters (CyTOF), before any clustering or quantification.

• R (flow/spectral): flowCore gate on the FSC-A vs FSC-H diagonal (+ FSC-W/SSC-W)
• R (mass/CyTOF): gate on DNA intercalator + Gaussian/Event_length parameters

The Single Most Important Modern Insight -- Doublets Are Caught by Area-vs-Height Non-Proportionality, and Scatter Gating Is Necessary but Not Sufficient

A doublet has roughly double the pulse AREA of a singlet but NOT double the Height, and a longer Width/transit time - so singlets fall on a tight FSC-A vs FSC-H diagonal and doublets deflect above it. A 1D area histogram therefore does NOT remove doublets; the discriminating signal is the Area-Height relationship (plus Width). This matters because an unremoved doublet of a CD3+ and a CD19+ cell reads as an artifactual CD3+CD19+ "double-positive," and clustering will faithfully (and wrongly) carve it out as a real population. Crucially, scatter gating is necessary but NOT sufficient: heterotypic conjugates (e.g. a CD3+CD14+ T:monocyte) survive standard FSC-A/H gates and present as genuine double-positives whose lineage-marker levels look COMPARABLE to true single-positives - the tell is an ELEVATED shared marker (e.g. CD45) and a high bright-field aspect ratio, so the definitive resolver is imaging flow cytometry, not a lineage-intensity check (Stadinski 2020 Cytometry A 97:1102). On CyTOF there is no scatter at all - doublets are removed by ion-cloud Gaussian parameters and DNA intercalator content (Bagwell 2020 Cytometry A 97:184).

Method Taxonomy

Method	Instrument	Principle	Caveat
FSC-A vs FSC-H	flow/spectral	singlets on the A-H diagonal	the standard; the discriminator is non-proportionality, not area
FSC-W / SSC-W	flow/spectral	doublets have longer pulse Width	complementary to A-vs-H
DNA intercalator (Ir191/193)	CyTOF	doublets show ~2N+ DNA	also separates cells from beads/debris
Gaussian params + Event_length	CyTOF	ion-cloud fit residual/length flags fusions	catches fusions DNA alone misses (Bagwell 2020)
imaging cytometry	imaging flow	bright-field aspect ratio	the only clean resolver of heterotypic conjugates

Note: cytometry doublet removal is GATING-based. DoubletFinder/Scrublet/scDblFinder are scRNA-seq DROPLET methods (they simulate artificial doublets) - limited transfer, because cytometry has direct physical doublet signals.

FSC-A vs FSC-H Singlet Gating (flow/spectral)

Goal: Keep events on the singlet diagonal.

Approach: A polygon along the A=H diagonal (preferred over a rectangle, which keeps off-diagonal doublets); visualize with the gate overlaid.

library(flowCore); library(ggcyto)

# matrix dimnames preserve 'FSC-A'/'FSC-H'; data.frame() would mangle them to FSC.A
singlet <- polygonGate(filterId = 'singlets', .gate = matrix(
  c(20000, 10000, 250000, 200000, 250000, 260000, 20000, 40000), ncol = 2, byrow = TRUE,
  dimnames = list(NULL, c('FSC-A', 'FSC-H'))))
singlets <- Subset(fs, singlet)
autoplot(fs[[1]], 'FSC-A', 'FSC-H') + ggcyto::geom_gate(singlet)

CyTOF Doublet Removal

Goal: Keep intercalator-positive single ion clouds.

Approach: Gate DNA intercalator (nucleated, ~2N) and Event_length/Gaussian residual; CATALYST exposes these as channels in the SCE.

library(CATALYST)
# prepData moves Time/Event_length to int_colData by default - keep them in the assay with FACS=TRUE
sce <- prepData(fs, panel, md, transform = TRUE, cofactor = 5, FACS = TRUE)
e <- assay(sce, 'exprs')

dna <- e['DNA1', ]                                   # intercalator-positive = nucleated single cells
keep <- dna > quantile(dna, 0.05) & dna < quantile(dna, 0.95)
if ('Event_length' %in% rownames(sce))               # retained by FACS=TRUE (now on the arcsinh scale)
  keep <- keep & e['Event_length', ] <= quantile(e['Event_length', ], 0.99)   # quantile-relative, so scale is fine
sce_singlets <- sce[, keep]

Per-Method Failure Modes

1D area gate leaves doublets

Trigger: gating only FSC-A. Mechanism: doublets overlap singlets in area. Symptom: double-positive clusters persist. Fix: gate the FSC-A vs FSC-H diagonal (+ Width).

Heterotypic conjugate survives scatter gating

Trigger: a surprising double-positive between two single-positive clusters. Mechanism: T:monocyte conjugate is scatter-normal, lineage markers comparable to singlets. Symptom: "novel" DP population with an elevated shared marker (e.g. CD45). Fix: treat as suspected doublet; check the shared-marker signal; confirm/resolve by imaging flow (bright-field aspect ratio) when load-bearing.

CyTOF "doublet gate" using scatter

Trigger: porting flow logic to CyTOF. Mechanism: no FSC/SSC exists. Symptom: no scatter channels. Fix: use DNA + Gaussian/Event_length.

Quantitative Thresholds

Threshold	Source	Rationale
expected doublet rate ~1-5% (PBMC), higher in tissue	community	flag samples far above as prep issues - not a removal cutoff
Gaussian + DNA gating improves CV (3.45 -> ~2.04)	Bagwell 2020 Cytometry A 97:184	combined DNA + Gaussian over baseline (Gaussian alone ~2.41)

Note: a fixed "95th-percentile residual" cutoff is arbitrary; prefer a visual diagonal gate or the instrument's Gaussian parameters over an unjustified quantile.

Common Errors

Error / symptom	Cause	Solution
double-positive cluster that "shouldn't" exist	residual heterotypic doublets	check for an elevated shared marker (CD45); confirm by imaging flow
no FSC/SSC channels (CyTOF)	mass data has no scatter	use DNA/Gaussian/Event_length
over-removal of large cells	rectangle gate clips real large singlets	use a diagonal polygon, not a box

References

• Stadinski 2020 Cytometry A 97(11):1102-1104 — heterotypic doublets survive scatter gating.
• Bagwell 2020 Cytometry A 97(2):184-198 — automated CyTOF cleanup via Gaussian/Event_length.
• Finck 2013 Cytometry A 83(5):483-494 — CyTOF DNA/event parameters in normalization context.

Related Skills

Workflow order (CyTOF): EQ-bead drift normalization (raw, FIRST) -> cytometry-qc -> doublet-detection -> clustering -> CytoNorm cross-batch (LAST)

• cytometry-qc - Run first: flow-rate/signal/margin cleaning
• bead-normalization - CyTOF drift correction after doublet removal
• fcs-handling - Load FCS files
• gating-analysis - Where singlet discrimination sits in the hierarchy
• clustering-phenotyping - Downstream analysis after doublet removal
• single-cell/doublet-detection - Droplet scRNA-seq doublet methods (different principle)