Doublet Detection

Methods like DoubletFinder and scDblFinder generate artificial doublets by randomly combining transcriptional profiles from the observed dataset. A classifier is then trained to distinguish these synthetic doublets from real singlets based on gene expression features.

• Mechanism: Creates a training set of pseudo-doublets in silico
• Key Assumption: Artificial doublets recapitulate the transcriptional signatures of true doublets
• Strength: Does not require prior knowledge of cell types or marker genes
• Limitation: Performance degrades when true doublets are rare or when cell types are highly transcriptionally similar

Algorithms such as Scrublet operate on the principle that doublets exhibit a hybrid transcriptomic identity, co-expressing marker genes from two distinct cell types. By constructing a nearest-neighbor graph and analyzing the local density of simulated doublet neighbors, these methods score each barcode for its doublet probability.

• Mechanism: Evaluates whether a cell's transcriptome is better explained as a mixture of two neighboring cell states
• Key Metric: Doublet score derived from the ratio of simulated doublet neighbors to singlet neighbors
• Advantage: Computationally efficient and scalable to large datasets
• Caveat: Requires sufficient cell type separation in the embedding space to detect hybrid states

When pooling samples from individuals with distinct genotypes, tools like Souporcell and Demuxlet leverage single-nucleotide polymorphisms (SNPs) to identify doublets. A barcode containing alleles from two different donors is unambiguously flagged as a cross-sample doublet.

• Mechanism: Variant allele frequency analysis at known SNP positions
• Gold Standard: Genetic methods provide the highest specificity for cross-sample doublets
• Key Distinction: Cannot detect doublets formed from two cells of the same donor (same-sample doublets)
• Requirement: Requires genotyping data or sufficient read depth at polymorphic loci

Cell Hashing uses oligonucleotide-conjugated antibodies against ubiquitous surface proteins to label cells from different samples with distinct barcodes. After pooling and sequencing, the relative abundance of hashtag oligos (HTOs) per barcode reveals doublets as those with high counts for two distinct tags.

• Mechanism: Multiplexed sample tagging with lipid- or antibody-conjugated barcodes
• Readout: HTO count matrices analyzed for barcode purity
• Advantage: Simultaneously enables sample demultiplexing and doublet detection
• Limitation: Adds cost and a sample processing step; cannot detect same-sample doublets

Emerging approaches like Solo use variational autoencoders to learn a latent representation of single-cell transcriptomes and directly model the probability that a given barcode is a doublet. These methods can capture non-linear doublet signatures that linear classifiers miss.

• Mechanism: Unsupervised or semi-supervised neural networks trained on raw count data
• Key Innovation: Learns doublet features directly from data without explicit simulation
• Strength: Can model complex, non-linear doublet transcriptomes
• Consideration: Requires GPU access and careful hyperparameter tuning for optimal performance

Doublet detection must be distinguished from ambient RNA contamination. While doublets represent true multi-cell capture events, ambient RNA arises from lysed cells releasing mRNA into the suspension. Tools like SoupX and CellBender model and subtract this background, preventing ambient profiles from being misclassified as doublets.

• Distinction: Doublets = intact multi-cell capture; Ambient = extracellular mRNA contamination
• Correction Strategy: Estimate the ambient RNA fraction using empty droplets as a background model
• Integration: Ambient RNA removal should precede doublet detection for accurate scoring
• Impact: Failure to correct ambient RNA inflates doublet scores and reduces detection specificity

The initial filtering step that removes low-quality cells before doublet detection is applied. QC metrics identify empty droplets, dying cells, and debris.

• Key metrics: Total UMI counts, number of genes detected, and mitochondrial read fraction
• Cells with high mitochondrial content (>20%) typically indicate membrane damage
• Doublet detection is performed after basic QC filtering but before normalization
• QC thresholds must be dataset-specific; no universal cutoffs exist

Cell-free mRNA molecules present in the cell suspension that contaminate droplets, producing background counts that can mimic doublet signatures.

• Ambient RNA originates from lysed or damaged cells in the sample
• Algorithms like SoupX and CellBender estimate and remove this contamination
• High ambient RNA can inflate doublet scores in tools like Scrublet
• Distinguishing true doublets from ambient contamination requires careful thresholding

A sparse numerical matrix where rows represent genes and columns represent cell barcodes, serving as the primary input for doublet detection algorithms.

• Stored in formats like AnnData (Python) or Seurat objects (R)
• Doublet detectors operate on raw or normalized count matrices
• The sparsity of scRNA-seq data (>90% zeros) complicates doublet identification
• Matrix dimensions directly impact computational runtime of detection methods

A multiplexing technique that labels cells from different samples with distinct oligonucleotide-tagged antibodies, enabling sample pooling and computational demultiplexing.

• Provides an orthogonal method for identifying doublets formed from different samples
• Hashtag oligonucleotide (HTO) counts reveal mixed-signal barcodes
• Cannot detect doublets from the same sample (same hashtag)
• Often used alongside computational methods like DoubletFinder for comprehensive detection

Non-biological systematic variation introduced by technical factors that can confound doublet detection when analyzing multiple samples together.

• Doublet rates vary between batches due to differences in cell concentration
• Running doublet detection separately per sample is recommended
• Batch effects can cause cells from one batch to appear as artificial doublets in another
• Integration methods like Harmony or scVI are applied after doublet removal

Statistical comparison of gene expression between cell groups that is directly compromised by undetected doublets in the dataset.

• Doublets create artificial intermediate transcriptomes that dilute cluster purity
• False-positive marker genes can emerge from doublet-derived pseudo-clusters
• Removing doublets before differential testing is critical for reproducible results
• Doublet-derived clusters often show co-expression of mutually exclusive lineage markers