Single Cell Pseudobulk Analysis
Single cell pseudobulk analysis is a statistical framework that aggregates single cell expression data by sample and cell type to perform differential expression using bulk RNA sequencing tools, preserving biological replication while avoiding the pseudo-replication pitfall of treating each cell as independent. This guide is for researchers who have completed single cell clustering and annotation and now need to identify genes that change between conditions across cell types. Use this guide when you have at least three biological samples per condition and need robust, interpretable differential expression results.
The method works by summing counts across all cells of a given cell type within a biological sample, then applying standard differential expression tools such as DESeq2 or edgeR. Bioconductor provides validated packages for both single cell preprocessing and pseudobulk workflows. Galaxy Training Network offers step by step tutorials that demonstrate how to move from a single cell count matrix to aggregated pseudobulk counts.
At a Glance
| Aspect | Description | Key Considerations |
|---|---|---|
| Purpose | Identify differentially expressed genes per cell type between conditions | Requires biological replicates, not technical replicates |
| Input | Per cell count matrix with cell type annotations and sample IDs | Minimum 3 samples per condition, recommended 5 or more |
| Aggregation | Sum counts per gene per sample per cell type | Also possible to use mean or median, but sum preserves count distribution |
| Tools | DESeq2, edgeR, limma voom, Muscat | All available via Bioconductor |
| Output | Gene level fold changes, p values, adjusted p values per cell type | Correct for multiple testing across cell types |
| Cell count filter | Exclude sample cell type combinations with fewer than 10 to 50 cells | Low counts cause inflated false positives |
Core Concepts and Decision Points
The central idea of pseudobulk analysis is to transform single cell data into a form that respects experimental design. Each cell is not an independent biological replicate, cells from the same sample share technical and biological variation. Aggregating by sample recovers the correct unit of replication. This principle is explained in detail in the NCBI Bookshelf chapter on RNA sequencing experimental design, which emphasizes the need to account for multi level structure.
Aggregation Strategy
Two main aggregation strategies exist: sum counts or use the mean expression. Summing is standard because it retains count based distributions required by DESeq2 and edgeR. Mean expression can be used with limma if transformed properly, but sum is simpler and more widely adopted. A third option is to use median or trimmed mean, but these are less common.
Cell Type Definition
You must have high confidence cell type labels. Using suboptimal clustering may mix distinct populations and obscure real signals. EMBL EBI Training recommends iterating between clustering and marker validation before proceeding to pseudobulk analysis.
Statistical Test Choice
DESeq2 and edgeR model counts with negative binomial distributions and handle small sample sizes well. limma voom is appropriate when counts are transformed to log CPM and sample sizes are larger. Muscat (multi sample multi condition) is a dedicated package for pseudobulk that wraps these methods and adds multilevel modeling.
When to Use Pseudobulk
Choose pseudobulk when you have clearly separated cell types and at least three samples per condition. Use mixed models or cell level analysis only when you cannot aggregate due to single cell resolution requirements (e.g., trajectory inference). Recent studies such as Fibroblast State Remodelling use pseudobulk to validate cell state specific expression changes across aging samples.
Workflow Implementation
The following steps outline a practical pseudobulk workflow. Each step is grounded in tools available from Bioconductor and validated by Galaxy Training Network materials.
Step 1: Prepare the Single Cell Object
Ensure your Seurat or SingleCellExperiment object contains two essential metadata columns: sample ID and cell type. Filter out doublets, debris, and low quality cells. Cell filtering decisions should be made before aggregation. Automatic extraction of mesenchymal stromal cells demonstrates that careful cell type identification improves pseudobulk signal.
Step 2: Aggregate Counts
For each sample and each cell type, sum the raw UMI or read counts across all cells. This creates a pseudobulk count matrix with rows as genes and columns as sample cell type combinations. The Bioconductor package Muscat provides the aggregateData function for this task.
Step 3: Filter Low Expression
Remove genes with very low counts across most samples. A common filter is to retain genes with at least 10 counts in a minimum number of samples. Also filter out sample cell type combinations with fewer than a minimum number of cells (e.g., 10 to 50 cells). Single cell pseudobulk analyses reveal hidden mitochondrial expression imbalance shows how low cell counts can amplify technical noise.
Step 4: Normalize and Model
Import the count matrix into DESeq2 or edgeR. Include sample level covariates such as sex, age, or batch. If you have multiple cell types, consider running separate models per cell type or using a model with interaction terms. The Development and Validation of a Six Gene Signature uses DESeq2 on pseudobulk counts to identify myeloid dysregulation signatures.
Step 5: Multiple Testing Correction
Adjust p values within each cell type (e.g., Benjamini Hochberg). If you test many cell types, consider an additional global correction. Report effect sizes and confidence intervals, not only p values.
Quality Checks
After running the analysis, perform several quality checks to ensure robustness.
Library Size Balance
Verify that pseudobulk library sizes correlate with the number of cells aggregated. Large discrepancies may indicate dropout issues or sample contamination. The NCBI Bookshelf chapter on RNA seq quality assessment recommends plotting library size versus cell count.
Principal Component Analysis
Plot PCA of pseudobulk samples colored by condition and cell type. Samples should cluster primarily by condition within each cell type if the biological signal is strong. Batch effects should be assessed. Single Cell RNA Sequencing of Murine Liver uses PCA on pseudobulk data to confirm circadian clustering.
Replicate Reproducibility
Check that biological replicates within the same condition show similar expression profiles. Calculate pairwise correlations and look for outlier samples.
Marker Gene Validation
Confirm that known marker genes for each cell type appear in the top differentially expressed genes. If they do not, your cell type annotations may require revision. Drought and salinity stress remodel Asian rice used pseudobulk to validate cell type specific stress responses.
Common Mistakes
Aggregating Before Filtering Cells
Doublets, dead cells, or multiplets contribute noise to pseudobulk counts. Always filter the single cell data thoroughly before aggregation. This mistake is one of the most frequently noted in single cell best practices from EMBL EBI Training.
Ignoring Low Cell Counts
When a sample has very few cells of a given type, the pseudobulk counts are unreliable. Excluding those sample cell type combinations reduces false positives and stabilizes variance estimates.
Treating All Cells as Independent Replicates
Running differential expression directly on single cell counts without aggregation inflates sample size artificially and leads to severe type I error inflation. Pseudobulk corrects this by collapsing to the biological replicate level.
Overlooking Batch Effects
If samples were processed in different batches, batch effects can distort pseudobulk results. Include batch as a covariate in your model. The Galaxy Training Network recommends checking batch confounding before the differential expression step.
Limits and Uncertainty
Pseudobulk analysis loses information about cell to cell heterogeneity within a sample. You cannot detect subpopulations within a cell type or study gene cell coupling. The method assumes that cells within a sample cell type combination are homogenous, which may not hold.
Statistical power depends on the number of biological samples, not the number of cells. Adding more cells per sample does not increase the effective sample size for testing condition effects. For small sample sizes (e.g., 2 per condition), pseudobulk methods are unreliable and may produce inflated or deflated p values.
The choice of aggregation method can influence results. Summing counts assumes that all cells contribute equally, which ignores differences in cell size or capture efficiency. Normalization within the single cell step (e.g., SCTransform) does not directly transfer to pseudobulk, you should use raw counts for aggregation and let the bulk tool normalize.
Finally, pseudobulk cannot address technical variation introduced during library preparation of single cell data, such as ambient RNA or batch specific dropout. These issues persist in the aggregated counts. EMBL EBI Training notes that ambient RNA correction (e.g., SoupX) should be applied before aggregation.
Frequently Asked Questions
How many cells per sample cell type are needed for pseudobulk?
There is no strict cutoff, but at least 10 to 50 cells per sample is recommended. Fewer cells yield unstable counts and high variance. Many published studies use a minimum of 30 cells.
Can I use pseudobulk for single nucleus data?
Yes. Single nucleus data can be aggregated the same way, though you should account for lower mRNA content. The same aggregation and differential expression tools apply.
Should I normalize single cell counts before aggregation?
No. Use raw UMI or read counts for aggregation. Normalization is performed by the downstream bulk differential expression tool (e.g., DESeq2). Any prior normalization would distort the count distribution.
How do I handle multiple cell types simultaneously?
You have two options. Run a separate differential expression model for each cell type, or use a single linear model that includes cell type as a factor and tests interaction terms. The latter requires care to avoid overparameterization.
References and Further Reading
- NCBI Bookshelf for fundamental RNA seq experimental design and quality control.
- EMBL EBI Training for single cell analysis best practices and batch effect handling.
- Galaxy Training Network for step by step pseudobulk workflows with DESeq2.
- Bioconductor for the Muscat package and comprehensive documentation on aggregation and differential testing.
- NCBI Sequence Read Archive for finding public single cell datasets to practice pseudobulk analysis.
- Fibroblast State Remodelling and Stromal Depletion as a real world pseudobulk application in aging skin.
- Automatic extraction of mesenchymal stromal cells for cell type identification and its impact on pseudobulk quality.
- Development and Validation of a Six Gene Signature for pseudobulk applied to myeloid dysregulation in hepatocellular carcinoma.
- Single Cell RNA Sequencing of Murine Liver for circadian rhythm analysis using pseudobulk.
- Single cell and pseudobulk analyses reveal hidden mitochondrial expression imbalance for a cautionary example of low cell count artifacts.