Flow Cytometry Batch Effect Correction: When CytoNorm Normalization Helps and When It Hides Biology

The UMAP came back with six islands, and they line up with acquisition dates, not treatment arms. Before you run CytoNorm or any other flow cytometry batch effect correction on that dataset, answer one question: if the correction works perfectly, will the difference you are trying to measure still be there? This post walks the decision in the order it should be made — confirm the batch effect, check the study design, then pick a correction (or decide not to correct).

Step 1: Confirm it is a batch effect before you normalize

An island on a UMAP is a symptom, not a diagnosis. Batch effects come from acquisition-day differences — instrument drift, antibody cocktail age, a different operator — and they have a recognisable signature. Technical shifts are abrupt and uncorrelated with the biology: every sample acquired that day moves together, in the same channels, by about the same amount. Biological variation is usually gradual and correlated with the condition.

The only clean way to tell the two apart is a bridge sample (also called an anchor or reference sample): aliquots of one cryopreserved PBMC or whole-blood lot, stained and acquired in every batch. Because the bridge is biologically identical across batches, any shift in its marker distributions is technical by construction.

Pair the bridge check with the instrument record. If your daily bead QC and Levey-Jennings tracking shows the same day drifting, you have found the cause. If the beads are flat and only one fluorochrome moved, suspect the reagent — tandem dye degradation in PE-Cy7 and APC-Cy7 produces exactly this single-channel, single-batch pattern, and it also changes the spillover values.

A batch compensated with a stale matrix shows diagonal smear in the affected channel pair. Recompute the matrix from that day’s single-stain controls (the compensation calculator shows what over- and under-compensated data look like) before you consider normalization. A spline cannot fix a spillover error.

Step 2: Check whether batch is confounded with condition

This is the question that decides whether normalization helps or hides biology. Lay out a table of batches against conditions and look for empty cells.

• Balanced: each batch contains samples from every condition (or at least each condition appears in more than one batch). Correction can separate technical from biological variation.
• Confounded: all treated samples ran on Monday and all controls on Thursday. No algorithm can tell a Monday effect from a treatment effect, because in this dataset they are the same variable.

If you are still designing the study, this is where most of the value is. Randomise samples across acquisition days, fix PMT voltages to target MFIs (the voltration and peak-2 method sets them), use one antibody cocktail lot for the whole study, and put the bridge sample in every batch. Normalization is a repair; design is prevention.

Path A: You ran a bridge sample in every batch — use CytoNorm

CytoNorm (Van Gassen et al., Cytometry Part A, 2020) is built for exactly this design. It learns the batch effect from the control samples, not from your study samples:

1. The aggregated control data are clustered with FlowSOM, so the correction is fitted per cell population rather than to the whole sample. A CD4 T cell and a monocyte can shift by different amounts in the same channel.
2. For each cluster and each marker, it computes quantiles of the control distribution in every batch.
3. A goal distribution is set — by default the mean of each quantile across batches, or one batch chosen as the reference.
4. A monotone spline maps each batch’s quantiles onto the goal, and that spline is applied to every study sample in the batch.

The paper states the assumptions plainly. The controls must “encompass the whole range of expression values” and contain cell types similar to the study samples, because a spline extrapolates poorly outside the range it was trained on. Batch effects must also be small enough that clustering is not driven by them.

The original validation was on a mass cytometry dataset (whole blood from a pregnancy cohort, 33 markers). The authors say the method applies to fluorescence flow data too, but a mass cytometry dataset is not the same as a conventional flow dataset with day-to-day compensation differences. Treat your first run as something to verify, not something to trust. CytoNorm is available as an R package and as a FlowJo plugin.

Path B: No bridge sample, but conditions are spread across batches

When there is no technical replicate, a reference-free method can still work — provided the design is balanced. cyCombine (Pedersen et al., Nature Communications, 2022) clusters cells with a self-organising map (default 8 × 8, so 64 clusters) and applies ComBat within each cluster. It needs no technical replicates. It can take a biological cofactor such as treatment so that difference is protected, and it requires that at least one condition from each batch also appears in at least one other batch.

That last requirement is the whole game. A reference-free method estimates the batch effect from the study samples themselves. Without overlapping conditions there is nothing to estimate it from except the biology. Recent CytoNorm releases also document a mode that runs without dedicated controls; the same design constraint applies.

Path C: Batch and condition are confounded — do not normalize

If every treated sample shares a batch, any normalization that aligns the batches also aligns the conditions. You will get a clean UMAP with no islands and no treatment effect, and nothing in the output will tell you the effect was removed.

The defensible options are, in order of preference:

• Re-acquire a subset of samples across batches, if material remains, to break the confound.
• Restrict the analysis to manually gated frequencies, with gates placed per batch on that batch’s own FMO controls. Report the confound as a limitation, and do not compare MFIs across batches.
• Drop the cross-batch comparison and report within-batch contrasts only.

Path D: The shift is small and your readout is a gated frequency

Normalization matters most for analyses that compare raw intensity across samples: clustering, t-SNE and UMAP embeddings, and MFI endpoints. A percent-positive readout from a gate that sits in a clear trough between populations is far less sensitive. The failure case is a fixed gate on a shifted population: when staining or instrument variation moves a population by half a decade or more, a gate drawn on Day 1 cuts through it on Day 4. If your populations are well separated and the bridge sample shifts are small relative to the gap, review the gate placement batch by batch and skip the spline.

After correcting: prove you did not erase the biology

• Bridge overlap: after normalization, the bridge sample’s distributions should overlay across batches. If they do not, the fit failed.
• Known-biology check: pick one difference you expect before you look — a marker that should differ between conditions on prior evidence — and confirm it survives with roughly the same magnitude.
• Cluster re-check: re-run clustering on the normalized data and check that the islands are gone because batches merged, not because real subsets merged. The FlowSOM cluster trust checks apply here unchanged.

Summary: which path you are on

Most of the time the decision is made at Step 1, by laying the bridge sample’s per-batch medians side by side. Cytomaton’s batch template workflow applies one gating template across up to 100 FCS files per job and exports % parent and MFI per parameter for every sample to a single CSV, which is that table. Cytomaton does not run CytoNorm or other normalization itself; it gives you the numbers you need to choose a path, and the corrected files can come back in afterwards.