An empirical Bayes method for differential expression analysis of single cells with deep generative models

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) is crucial for cell subpopulation characterization.
• Technical variations in scRNA-seq data obscure biological signals, necessitating advanced analytical methods.
• Deep generative models are used for scRNA-seq data analysis, including dimensionality reduction and batch correction.

Purpose of the Study:

• To develop a novel Bayesian approach for differential expression (DE) analysis in scRNA-seq data.
• To leverage the uncertainty inherent in deep generative models for improved DE detection.
• To introduce control for effect size and false discovery rate (FDR) in DE analysis.

Main Methods:

• Developed lvm-DE, a generic Bayesian framework for DE predictions from deep generative models.
• Applied the lvm-DE framework to existing models like scVI and scSphere.
• Incorporated uncertainty quantification from deep generative models into the DE analysis pipeline.

Main Results:

• The lvm-DE approach demonstrates superior performance in estimating log fold change in gene expression.
• It accurately detects differentially expressed genes between cell subpopulations.
• Outperforms existing state-of-the-art methods in DE analysis for scRNA-seq data.

Conclusions:

• lvm-DE provides a robust and flexible Bayesian framework for DE analysis in scRNA-seq.
• The method effectively utilizes deep generative model uncertainty for enhanced DE detection.
• Offers improved accuracy and control over FDR in identifying gene expression differences.