Doublet Detection In Single-Cell RNA-seq

If you have ever run a single-cell RNA-seq experiment and found a cluster co expressing marker genes from two different cell types, you have seen a doublet. A doublet is a single droplet or well that captured two (or more) cells instead of one. Doublets inflate cell counts, create spurious clusters, and distort differential expression results. This guide explains the core concepts, decision points, workflow, quality checks, common mistakes, and limits of doublet detection. It is written for bioinformaticians, bench scientists, and data analysts who process or plan to process single-cell RNA-seq data. Use this guide to design a robust doublet detection strategy that fits your experimental design and computational resources.

Doublet detection is not optional. Even the most careful pipetting yields a doublet rate of roughly 1 to 8 percent, and that rate rises with higher cell loading concentrations. The Galaxy Training Network provides hands on tutorials for quality control steps, including doublet removal, that you can adapt to your own data. In practice, a systematic approach that combines library complexity filtering, marker gene evaluation, and dedicated doublet detection software will save you from downstream misinterpretation. The Bioconductor project offers several well documented packages for this purpose.

At a Glance

Core Concepts

A doublet originates during cell capture. In droplet based platforms (10x Genomics, Drop seq), two cells can enter the same droplet. In plate based methods, a well may receive two cells. The result is a single “cell” whose transcriptome is a mixture of two distinct cellular profiles. Doublets come in two flavors. Heterotypic doublets arise from two different cell types and are relatively easy to spot because they co express markers from both lineages. Homotypic doublets come from two cells of the same type and are nearly impossible to detect computationally because their combined profile looks like a normal single cell of that type.

Doublet rates scale with the number of cells loaded. Loading 10,000 cells instead of 5,000 increases the expected doublet proportion, and the manufacturer’s guidelines often provide an estimate. The EMBL EBI Training materials on single cell analysis emphasize that you should always account for the expected doublet rate when interpreting clusters. Because doublets are technical artifacts, they do not represent biological heterogeneity. Removing them improves the reproducibility of downstream analyses, especially clustering and marker gene identification.

Decision Criteria

Not every doublet detection method works for every experiment. Here are the key factors to consider before choosing an algorithm.

Species and sample mixing. If you have pooled cells from two different species (e.g., human and mouse), doublets appear as cells where a large fraction of reads map to both genomes. The NCBI Sequence Read Archive hosts many datasets that demonstrate this “species mixing” approach, which is the most straightforward detection strategy. For same species, same individual samples, you need computational methods that simulate doublets from the data itself.

Genotype information availability. When you multiplex multiple donors in one lane, genotype based demultiplexing can simultaneously assign donor identity and flag doublets. Tools such as demuxlet or souporcell use SNP information from the sequencing data. The Semi parametric empirical Bayes method for multiplet detection extends a similar idea to single nucleus ATAC seq. If you have genotype data, these methods are highly accurate. Without it, you must rely on expression based methods.

Computational resources. Scrublet and DoubletFinder run quickly on 10,000 20,000 cells. scDblFinder, available through Bioconductor, scales to hundreds of thousands of cells but uses more memory. Decide based on your dataset size and available RAM.

Expected doublet rate and heterogeneity. Datasets with many rare cell types are more vulnerable to doublet misclassification because a rare cell type may be mistaken for a doublet. In such cases, use a method that provides a doublet score rather than a hard threshold, and manually inspect borderline cells.

Practical Workflow

The following workflow assumes you have already performed basic quality control (removing low quality cells with few genes and high mitochondrial content). Doublet detection should occur before normalization and clustering, but after filtering out empty droplets.

Step 1: Prepare the Count Matrix

Input a raw UMI count matrix (cells x genes). Many pipelines expect filtered or unfiltered matrices. If you use 10x Cell Ranger, use the raw matrix or the filtered matrix after basic QC. The SCALPEL pipeline for spatial transcriptomics demonstrates similar preprocessing steps that can be adapted to scRNA seq.

Step 2: Run Doublet Detection

Choose a method based on your criteria described above. For most users, scDblFinder (R Bioconductor) or scrublet (Python) are solid defaults. Example using scDblFinder:

library(scDblFinder)
sce <- scDblFinder(sce)  # adds column "scDblFinder.class"

This function simulates artificial doublets from the data and compares each real cell to the simulated ones. It returns a classification and a doublet score. The Consensus Pituitary Atlas used similar methods to annotate cell types, and their approach can guide your own thresholding.

Step 3: Evaluate Doublet Calls

Do not blindly remove all cells flagged as doublets. Instead, visualize the doublet score distribution. Look for a bimodal pattern where most cells are low and a small peak is high. Use the expected doublet rate as a guide: if you loaded 6,000 cells, expect roughly 4 percent doublets. Adjust the threshold if necessary.

Examine the top clusters that are enriched for doublets. Are they expressing markers from two distinct cell types? For example, a cluster that co expresses T cell markers (CD3E) and B cell markers (CD19) is almost certainly composed of doublets. The fastdemux method for SNP based demultiplexing similarly evaluates doublet clusters by genetic mixing.

Step 4: Remove Doublets

Apply your finalized threshold. Keep in mind that removing doublets will reduce your cell count, but that reduction is necessary for accurate biology. Many researchers remove between 2 and 8 percent of cells. After removal, re run normalization and clustering.

Step 5: Verify

Compare the cluster composition before and after doublet removal. You should see that clusters with mixed markers collapse or disappear. Typical clusters (e.g., CD4+ T cells) should become more distinct. Report your doublet filtering step in any methods section. The MICA/MICB mediated immune escape study used rigorous doublet detection to ensure that single cell transcriptomic mapping reflected true biology rather than artifacts.

Quality Checks

After removal, doublet contamination can still hide in the data. Perform these checks.

Marker co expression plots. For each cluster, plot the top two cluster specific markers. If you see a subpopulation expressing both markers, consider revisiting your doublet threshold.

Percent doublet score per cluster. Even after removal, you may have clusters with an average doublet score higher than expected. This sometimes indicates that the cluster contains doublets that the algorithm missed. Consider a more stringent threshold or a different method.

Cross species or cross donor mixing. If you have combined samples from different donors or species, check that each cluster contains reads from only one donor or species. Mixed clusters suggest doublets. The NCBI Bookshelf provides reference material on interpreting such mixing.

Common Mistakes

Using default parameters without checking. Every dataset has a different doublet composition. Default parameters may overfilter or underfilter. Always inspect the score distribution.

Removing too many cells. Aggressive doublet removal can eliminate rare but real cell types. Rare cell types are often small and may resemble doublets because their expression pattern is unlike the majority. Cross reference with expected doublet rate.

Running doublet detection after clustering. Some methods (e.g., DoubletFinder) are designed to be run before clustering. Running them after can create circular arguments where clusters artifically define doublet properties.

Ignoring batch effects. Doublet rates can vary across batches. If you have multiple samples, run doublet detection per sample or batch, not on the merged object.

Assuming homotypic doublets are removed. No expression based method can reliably remove doublets from two cells of the same type. Their transcriptomes are indistinguishable from a normal cell. This is a fundamental limit.

Limits of Interpretation

Doublet detection is probabilistic, not deterministic. Even the best algorithms produce false positives and false negatives. You must accept some uncertainty.

Homotypic doublets are invisible. If your doublet rate is 5 percent and half are homotypic, you will likely miss 2.5 percent of doublets. This can affect clustering of highly abundant cell types, but the impact on rare types is minimal because homotypic doublets of rare types are themselves rare.

Rare cell types are at risk. Doublet detection algorithms may flag rare cell types as doublets because their expression profile looks like an outlier. Always check whether a putative doublet cluster expresses canonical markers of a known rare type.

Thresholding is arbitrary. Using a fixed score cutoff (e.g., top 5 percent) assumes that doublets are always the highest scoring. This assumption holds reasonably well but not perfectly. Consider using a doublet score as a weight in downstream analyses instead of a hard filter.

Method comparisons vary by dataset. The performance of any algorithm depends on cell type diversity, sequencing depth, and number of cells. The fastdemux preprint notes that SNP based methods outperform expression based ones when genotype data are available.

Frequently Asked Questions

What is the difference between a doublet and a multiplet? Doublet and multiplet are often used interchangeably. Technically a doublet contains two cells, a multiplet contains two or more. Most software treats all multiplets as doublets for simplicity.

Should I run doublet detection before or after ambient RNA removal? Run doublet detection after ambient RNA removal. Ambient RNA can make single cells look like mixtures of cell types. Tools like SoupX or CellBender remove ambient contamination first, then you can more accurately detect true doublets.

Can I use doublet detection for non droplet platforms? Yes. Plate based methods (e.g., Smart seq2) also generate doublets. Computational methods like scDblFinder work with any count matrix. However, the expected doublet rate in plate based methods is often lower (less than 1 percent) because cells are manually sorted.

How do I choose between DoubletFinder, scrublet, and scDblFinder? All three are reliable. Scrublet and DoubletFinder are simpler and work well for small datasets. scDblFinder offers more flexibility, integrates with Bioconductor, and is recommended for larger or complex datasets. Check documentation for specific input requirements.

References and Further Reading

• NCBI Bookshelf. Single Cell RNA Sequencing general reference for data preprocessing.
• EMBL EBI Training. Single cell RNA seq analysis step by step training materials.
• Galaxy Training Network. Single cell quality control and doublet detection practical workflows.
• Bioconductor. scDblFinder documentation official package reference.
• NCBI Sequence Read Archive. scRNA seq datasets with known doublets for benchmarking.
• MICA/MICB immune escape study. Single cell transcriptomic mapping example of rigorous doublet removal.
• Consensus Pituitary Atlas. Annotation and quality control using doublet detection.
• Semi parametric empirical Bayes method. Multiplet detection in snATAC seq advanced method.
• fastdemux preprint. SNP based demultiplexing for population genomics.
• SCALPEL pipeline. Spatial transcriptomics preprocessing relevant QC steps.
• Plant scRNA seq review. Why location matters includes doublet discussion.

Related Articles

• Protein Synthesis
• Incomplete Dominance Gene
• Cell Membrane Function Biology
• Dna Structure
• Protein Structure