Doublet Detection

These methods generate artificial doublets by randomly combining transcriptional profiles from the dataset itself. A classifier is then trained to distinguish these synthetic artifacts from real singletons.

• Mechanism: Creates a training set of pseudo-doublets by averaging count vectors of random cell pairs.
• Key Algorithms: Scrublet, DoubletFinder, scDblFinder.
• Advantage: Does not require prior knowledge of cell-type markers.
• Limitation: Performance degrades if the dataset lacks sufficient heterotypic cell types to make doublets distinguishable.

These methods identify doublets by detecting cells that simultaneously express mutually exclusive lineage markers. A cell co-expressing T-cell and B-cell receptors is flagged as a heterotypic doublet.

• Mechanism: Compares a cell's gene expression against a curated list of non-overlapping cell-type signatures.
• Key Algorithms: DoubletDecon, Chord.
• Advantage: High biological interpretability and precision for well-annotated tissues.
• Limitation: Cannot detect homotypic doublets (two cells of the same type) and requires a high-quality reference of cell-type markers.

This approach leverages natural genetic variation between donors. By profiling single nucleotide polymorphisms (SNPs), a cell containing more than two alleles at multiple loci is identified as a doublet formed from genetically distinct individuals.

• Mechanism: Uses scRNA-seq reads to call genotypes and identifies cells with an excess of heterozygous variants.
• Key Algorithms: Demuxlet, Souporcell, Vireo.
• Advantage: Near-ground-truth accuracy for heterotypic doublets from different donors.
• Limitation: Ineffective for genetically identical samples (e.g., inbred mice) and requires deep sequencing coverage.

A physical multiplexing strategy where cells from different samples are labeled with oligonucleotide-conjugated antibodies (hashtags) before pooling. Computational demultiplexing reveals that a cell with multiple hashtags is a cross-sample doublet.

• Mechanism: Counts unique hashtag oligos per cell barcode; multiple dominant tags indicate a multiplet.
• Key Algorithms: HTODemux, GMM-Demux.
• Advantage: Highly accurate and also enables cost-efficient sample pooling.
• Limitation: Adds a wet-lab step and reagent cost; cannot detect doublets formed within the same original sample.

Neural networks learn a probabilistic latent representation of single-cell data. Doublets are identified as outliers with low likelihood under the model or by using a variational autoencoder to score the probability of a cell being a multiplet.

• Mechanism: Trains a VAE on the count matrix; doublets often exhibit high reconstruction error or map to ambiguous latent space regions.
• Key Algorithms: Solo, scVI doublet mode.
• Advantage: Captures complex, non-linear gene expression patterns and scales well to massive datasets.
• Limitation: Computationally intensive and requires careful hyperparameter tuning to avoid overfitting.

A pre-computational filtering strategy that uses cell loading concentration statistics. By applying Poisson statistics to the expected cell-to-droplet encapsulation rate, the theoretical multiplet rate is estimated and high-risk droplets are filtered.

• Mechanism: Relies on the Poisson distribution to estimate the multiplet rate based on the number of recovered cells relative to the number of droplets.
• Key Tools: 10x Genomics Cell Ranger force-cells parameter.
• Advantage: Zero computational overhead and provides an immediate physical estimate.
• Limitation: Cannot identify which specific droplets are doublets; only provides a global expected rate.