← Về thư mục
Flow Cytometry Pipeline Usage Guide
Overview
This workflow processes flow cytometry data from raw FCS files through compensation, transformation, clustering or gating, and differential analysis.
Prerequisites
BiocManager::install(c('CATALYST', 'flowCore', 'diffcyt'))
install.packages(c('FlowSOM', 'uwot'))
Quick Start
Tell your AI agent what you want to do:
- "Run the cytometry pipeline on my FCS files"
- "Cluster my CyTOF data and find differential populations"
- "Analyze my immunophenotyping experiment"
Example Prompts
Basic Analysis
"I have FCS files from a CyTOF experiment, run the full pipeline"
"Cluster my flow cytometry data with FlowSOM"
Differential Analysis
"Compare immune populations between treatment and control groups"
"Find differentially abundant cell types in my flow data"
Customization
"Use manual gating instead of clustering for my data"
"Run differential state analysis on my T cell markers"
When to Use This Pipeline
- Multi-color flow cytometry panels
- CyTOF/mass cytometry experiments
- Immunophenotyping studies
- Treatment response analysis
- Biomarker discovery
- FCS files - One per sample
- Panel file - Channel to marker mapping with marker classes
- Sample metadata - Conditions, patient IDs, batches
fcs_colname,antigen,marker_class
FSC-A,FSC,none
SSC-A,SSC,none
CD45,CD45,type
CD3,CD3,type
CD4,CD4,type
Ki67,Ki67,state
IFNg,IFNg,state
Marker classes:
- type: Lineage markers for clustering/phenotyping
- state: Functional markers for differential state analysis
- none: Scatter or non-analyzed channels
Pipeline Steps
1. Data Loading
- Read FCS files into R
- Validate panel and metadata
2. Compensation
- Apply spillover matrix
- Essential for fluorescence-based flow
- Not needed for CyTOF
- Arcsinh transformation (standard)
- Cofactor: 5 for CyTOF, 150 for flow
- Normalizes distributions
4. QC
- Check cell counts per sample
- Visualize expression distributions
- Identify outlier samples
5. Clustering or Gating
- Clustering: FlowSOM, Phenograph (unsupervised)
- Gating: Manual population identification
6. Dimensionality Reduction
- UMAP for visualization
- Based on lineage markers
7. Differential Analysis
- DA: Are population proportions different?
- DS: Is marker expression different within populations?
Analysis Approaches
Approach 1: Unsupervised Clustering
Best for: High-dimensional panels, discovery
- Cluster on lineage markers
- Annotate clusters based on expression
- Test differential abundance
Approach 2: Manual Gating
Best for: Well-defined populations, lower dimensions
- Gate populations hierarchically
- Export gated populations
- Compare frequencies
Cofactor Guidelines
| Platform |
Typical Cofactor |
| Flow cytometry |
150-500 |
| CyTOF |
5 |
| Spectral flow |
~150 |
Statistical Considerations
Differential Abundance
- Need biological replicates (n >= 3)
- Uses edgeR (count-based)
- Reports fold change in frequency
Differential State
- Tests marker expression within clusters
- Uses limma
- Adjust for multiple testing (clusters x markers)
Common Issues
Compensation problems
- Check single-stain controls
- Verify spillover matrix
- Look for negative values
Poor clustering
- Adjust number of clusters (K)
- Check if markers are appropriate
- Try different resolution
No significant results
- Check sample size (n >= 3)
- Verify biological effect expected
- Adjust thresholds carefully
Output Files
| File |
Description |
| cytometry_analysis.rds |
Complete analysis object |
| da_results.csv |
Differential abundance results |
| umap_clusters.png |
Cluster visualization (runDR embedding) |
| heatmap.png |
Cluster x marker expression heatmap |
| da_heatmap.png |
Differential abundance heatmap |
| abundances.png |
Population frequencies per sample |
Tips
- Unit of inference: the sample/subject, not the cell, is the experimental unit - diffcyt aggregates cells to per-sample-per-cluster counts/medians before testing; a per-cell test is invalid.
- Order is irreversible: compensate before transform; remove margin events before density-based QC; remove doublets before clustering. None of these is fixable downstream.
- Two normalization layers: EQ-bead drift correction runs first on raw counts (CyTOF); CytoNorm cross-batch harmonization runs on transformed data before the analytical clustering. Do not conflate them.
- Panel file: Essential for correct channel-to-marker mapping; marker_class drives everything (type clusters, state is tested).
- Cofactor: Use 5 for CyTOF, ~150 for conventional/spectral fluorescence.
- Batch effects: model batch in the diffcyt design; reserve normalization for visualization. Fully confounded batch (batch == condition) cannot be rescued.
- Embeddings: UMAP/tSNE are for visualization only - never gate on them or read biology off blob distances.
- Replicates: Minimum 2-3 biological replicates per condition for valid statistics.
References
- CATALYST (CyTOF workflow): Nowicka et al., F1000Research 6:748, doi:10.12688/f1000research.11622.4
- diffcyt: doi:10.1038/s42003-019-0415-5
- FlowSOM: doi:10.1002/cyto.a.22625