Interpretable dimensionality reduction of single cell transcriptome data with deep generative models.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) enables cell type discovery, state identification, lineage tracing, and spatial organization reconstruction.
• Interpreting scRNA-seq data is challenging due to high dimensionality, with existing methods often failing to reveal clustering structures or preserve global relationships.

Purpose of the Study:

• To introduce scvis, a robust statistical model for capturing and visualizing low-dimensional structures in scRNA-seq data.
• To address limitations of existing dimension reduction techniques that struggle with preserving both local and global data topology.

Main Methods:

• Development of scvis, a probabilistic parametric model designed for scRNA-seq data.
• Utilizing simulation studies to evaluate the preservation of local and global neighbor structures.
• Application of scvis to four diverse scRNA-seq datasets.

Main Results:

• scvis effectively captures and visualizes low-dimensional structures in scRNA-seq data.
• Learned representations preserve both local and global neighbor structures, outperforming existing methods.
• The model demonstrates robustness to varying numbers of data points and includes a probabilistic mapping function for new data integration.

Conclusions:

• scvis provides interpretable two-dimensional representations of high-dimensional scRNA-seq data.
• The method enhances the discovery of cellular heterogeneity and relationships within complex biological systems.
• scvis offers a powerful tool for analyzing scRNA-seq datasets, facilitating biological insights.