scds

scds identifies and scores doublets in single-cell RNA sequencing (scRNA-seq) data to computationally annotate multiplet artifacts that can confound downstream analyses.

Key Features:

• Co-expression based doublet scoring (cxds): Uses binarized gene expression and a binomial model to assess co-expression of gene pairs and provides interpretable annotations.
• Binary classification based doublet scoring (bcds): Uses a binary classification framework that leverages artificial doublets to distinguish doublets from authentic single-cell profiles.
• Binarized gene expression handling: Operates on presence/absence (binarized) gene expression values for co-expression analysis.
• Binomial model for gene-pair co-expression: Applies a binomial statistical model to evaluate significance of co-expression between gene pairs.
• Artificial doublet integration: Incorporates artificial doublets within the classification framework to separate multiplet signatures from singlets.
• Benchmarking on experimentally annotated datasets: Evaluated across four datasets with known experimental doublet annotations, demonstrating competitive accuracy and reduced computational cost.
• Computational efficiency and scalability: Achieves reduced computational cost and can process datasets with thousands of cells in seconds.
• Method performance variability: Reports that performance varies between methods and datasets, indicating no single universally superior approach.

Scientific Applications:

• Doublet detection in scRNA-seq: Annotates and scores putative doublets to reduce artifacts in single-cell transcriptomic analyses.
• Method benchmarking: Enables comparative evaluation of doublet detection approaches using datasets with experimental doublet annotations.
• Scalable preprocessing of large datasets: Provides rapid doublet annotation for large scRNA-seq datasets containing thousands of cells.

Methodology:

cxds leverages binarized gene expression and a binomial model to assess gene-pair co-expression, while bcds employs a binary classification framework that uses artificial doublets to distinguish doublets from authentic single-cell data.