Non-linear archetypal analysis of single-cell RNA-seq data by deep autoencoders

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables cell type discovery and analysis of cellular heterogeneity.
• Clustering methods, while useful, cannot capture continuous cellular states or shared gene expression programs (GEPs).

Purpose of the Study:

• To introduce scAAnet, an autoencoder for single-cell non-linear archetypal analysis.
• To identify GEPs and infer their activity across cells, overcoming limitations of existing clustering techniques.

Main Methods:

• Developed scAAnet, incorporating a count distribution-based loss for sparse, overdispersed scRNA-seq data.
• Implemented an archetypal constraint within the scAAnet loss function.
• Validated performance against existing archetypal analysis methods via simulations.

Main Results:

• scAAnet demonstrated superior performance in simulations compared to existing archetypal analysis methods.
• Successfully extracted biologically meaningful GEPs from diverse scRNA-seq datasets (pancreatic islets, lung IPF, prefrontal cortex).

Conclusions:

• scAAnet effectively identifies shared gene expression programs and continuous cellular states from scRNA-seq data.
• The method provides a powerful new tool for analyzing cellular heterogeneity and gene expression dynamics.