Dino

Dino models and normalizes gene-level single-cell RNA-seq counts by fitting a flexible negative-binomial mixture to adjust full gene expression distributions for library size effects and high zero abundance, including UMI data.

Key Features:

• Negative-binomial mixture modeling: Fits a flexible negative-binomial mixture model to gene expression counts.
• Distribution-level normalization: Normalizes entire gene-specific expression distributions rather than only adjusting mean expression for library size.
• Library size variance correction: Addresses non-constant variance and distributional distortions across different library sizes.
• Zero-count accommodation: Handles high proportions of zero counts typical in scRNA-seq datasets.
• Robustness to shallow sequencing and heterogeneity: Maintains performance under shallow sequencing depths and sample heterogeneity.
• Mitigation of technical artifacts: Reduces technical and experimental artifacts observed even when unique molecular identifiers (UMIs) are employed.
• Improved downstream performance: Increases statistical power and reduces false discoveries in downstream analyses.
• Implementation: Provided as an R package.
• Validation: Demonstrated effectiveness using simulated experiments and case-study datasets.

Scientific Applications:

• scRNA-seq preprocessing: Normalization of single-cell RNA-seq data to provide distributionally consistent inputs for downstream analyses.
• Datasets with shallow sequencing or high zeros: Normalization for datasets with shallow sequencing depth, heterogeneous samples, or varying proportions of zero counts.
• Improving downstream inference: Enhancing power and reliability of downstream analyses by reducing distributional artifacts and false discoveries.

Methodology:

Fits a flexible negative-binomial mixture model to gene-level scRNA-seq counts to normalize entire gene expression distributions across library sizes and was validated using simulated experiments and case-study datasets.