Doublet Detection Exercise

Here, we will explore how to identify and remove doublets (two cells captured in one droplet) from single-cell RNA-seq data using computational methods.

Dataset

We will work with the 10x Genomics 10k PBMCs from a Healthy Donor (v3 chemistry). The higher cell count increases doublet probability (~8% expected doublet rate).

• Download link: 10k PBMCs, 3’ v3.1
• Download the “Feature / cell matrix (filtered)”

Part 1: Initial Data Processing

Step 1: Load and explore the dataset

• Import the 10k PBMC dataset
• Check initial cell and gene counts
• Note: With ~10,000 cells loaded, expect 600-800 doublets

Step 2: Standard preprocessing

• Calculate QC metrics (% mitochondrial, nGenes, nUMIs)
• Create QC plots:
• Scatter plot: nUMIs vs nGenes (color by % mitochondrial)
• Histogram of nUMIs per cell

Step 3: Process without doublet removal

• Filter cells using standard thresholds
• Important: the doublet detection methods in Part 2 expect raw counts, because they simulate artificial doublets from raw UMIs and do their own internal normalization. Before you normalize, keep a copy of the raw counts (e.g. adata.layers["counts"] = adata.X.copy()), and run the Part 2 detection step on those raw counts. Store the resulting scores now, but remove the cells only later, in Part 4.
• Normalize, find variable genes, scale
• Run PCA → UMAP → clustering
• Save this object as “pre_doublet_removal”

Part 2: Doublet Detection

Run these on the raw counts you preserved in Step 3 (not on normalized data).

For Seurat users:

Install scDblFinder and use it according to the package documentation. It operates on the raw counts assay directly.

For scanpy users:

Use scrublet, which is now part of core scanpy:

• sc.pp.scrublet(adata, expected_doublet_rate=0.08) (the older scanpy.external.pp.scrublet is deprecated)
• Run this while adata.X holds raw counts (i.e. before normalizing).
• It adds two columns to adata.obs: doublet_score (continuous) and predicted_doublet (boolean).

Note: in real multi-sample data, doublet detection should be run separately per sample/capture (e.g. via the batch_key argument), since doublets can only form within a single droplet reaction. Our single-sample dataset does not require this.

Step 4: Analyze doublet predictions

• Add doublet scores/classifications to your object
• Create visualizations:
• UMAP colored by doublet score (continuous)
• UMAP colored by doublet classification (binary)
• Violin plot of nUMIs split by singlet/doublet
• Histogram of doublet scores with threshold line

Part 3: Biological Validation

Step 1: Check for heterotypic doublets

Look for cells expressing mutually exclusive markers:

• T cell + B cell markers: CD3D + CD79A
• T cell + Monocyte markers: CD3D + CD14
• B cell + NK markers: CD79A + NKG7

Create feature plots:

• Plot these marker pairs
• Overlay predicted doublets

Step 2: Examine clustering patterns

• Which clusters have the highest proportion of predicted doublets?
• Do any clusters consist mainly of doublets?
• Calculate doublet percentage per cluster

Part 4: Remove Doublets and Re-analyze

Step 1: Filter doublets

• Remove cells classified as doublets
• Document how many cells were removed

Step 2: Re-process cleaned data

• Re-run the full pipeline (normalize → HVG → PCA → UMAP → cluster)
• Save this object as “post_doublet_removal”

Step 3: Compare before and after

Create side-by-side comparisons:

1. Number of cells retained
2. UMAP plots (before/after)
3. Number of clusters
4. Marker gene expression clarity

Part 5: Critical Evaluation

Discuss with a neighbor:

1. What percentage of cells were identified as doublets?
2. Were doublets randomly distributed or concentrated in specific areas?
3. Did doublet removal reveal any new cell populations?
4. Did any clusters disappear after doublet removal?
5. How did doublet scores correlate with QC metrics (nUMIs, nGenes)?