Probabilistic harmonization and annotation of single-cell transcriptomics data with deep generative models.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• The increasing volume of single-cell transcriptomics data necessitates robust methods for integration and cell type annotation.
• Comparing gene expression across datasets and automatically labeling cell types remain significant challenges in scRNA-seq analysis.

Purpose of the Study:

• To present scVI, a probabilistic method for joint representation and analysis of scRNA-seq data, accounting for noise.
• To introduce scANVI, a semi-supervised extension of scVI for leveraging existing cell state annotations.
• To evaluate the performance of scVI and scANVI against state-of-the-art methods for data integration and cell annotation.

Main Methods:

• Development and application of scVI, a fully probabilistic model for joint analysis of multiple scRNA-seq datasets.
• Introduction of scANVI, a semi-supervised variant of scVI that utilizes prior cell state annotations.
• Comparative analysis of scVI and scANVI with existing methods on accuracy, scalability, and adaptability.

Main Results:

• scVI provides an effective probabilistic approach for integrating scRNA-seq data and analyzing gene expression.
• scANVI demonstrates strong performance in cell state annotation by leveraging existing labels.
• Both scVI and scANVI outperform current methods in accuracy, scalability, and adaptability, offering a unified generative model for downstream tasks.

Conclusions:

• scVI and scANVI offer powerful, scalable, and accurate solutions for single-cell data integration and cell type annotation.
• These methods provide a unified probabilistic framework, simplifying downstream analyses like differential expression.
• scvi-tools makes scVI and scANVI easily accessible for the research community.