CellBender

CellBender removes background ambient RNA and technical noise from droplet-based single-cell assays (scRNA-seq, snRNA-seq, CITE-seq) by applying an unsupervised deep generative model to produce denoised gene expression counts for improved downstream analysis.

Key Features:

• Deep generative model: A model structured to reflect the phenomenology of background noise generation in droplet-based assays.
• Unsupervised droplet classification: Distinguishes cell-containing and cell-free droplets without prior labeling.
• Background noise modeling: Learns the profile of ambient RNA and off-target gene expression across droplets.
• End-to-end denoising: Retrieves noise-free quantification of gene expression through an end-to-end process.
• Supported modalities: Explicit support for droplet-based scRNA-seq, snRNA-seq, and CITE-seq data.
• Performance on simulations: Performance approaches the theoretically optimal denoising limit on simulated datasets.
• Scalability and robustness: Implemented to scale to large droplet-based datasets while maintaining robust denoising.
• Biological insight: The learned background profile can provide insight into degraded or uncaptured cell types.
• Reduction of technical artifacts: Mitigates batch effects and spurious differential gene expression signals caused by ambient/background counts.

Scientific Applications:

• Ambient RNA removal: Removing ambient and technical background from droplet-based scRNA-seq, snRNA-seq, and CITE-seq datasets.
• Improved differential expression: Reducing spurious differential gene expression results caused by background counts.
• Batch-effect mitigation: Decreasing batch effects arising from inconsistent ambient contamination across experiments.
• Data alignment across modalities: Enhancing alignment of droplet-based single-cell data with established gene expression patterns across tissues, protocols, and modalities.
• Detection of degraded/uncaptured cells: Using background profiles to infer presence of degraded or uncaptured cell types.

Methodology:

An unsupervised deep generative model structured to reflect background noise generation is trained to distinguish cell-containing from cell-free droplets, learn a background noise profile, and output denoised gene expression counts via an end-to-end process.