Integrative Learning of Disentangled Representations from Single-Cell RNA-Sequencing Datasets

Area of Science:

• Computational biology
• Single-cell genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) reveals cellular heterogeneity.
• Existing batch correction methods often require matching features or paired samples.
• Analyzing diverse scRNA-seq datasets with nonmatching features remains challenging.

Purpose of the Study:

• To present shared-private Variational Inference via Product of Experts with Supervision (spVIPES), a probabilistic framework.
• To decompose unpaired scRNA-seq datasets with nonmatching features into shared and private components.
• To enable accurate cell-type identification across datasets lacking matching features.

Main Methods:

• Developed a probabilistic latent variable model separating dataset-specific (private) from conserved (shared) cellular features.
• Implemented supervised and unsupervised variants of spVIPES.
• Utilized optimal transport for cell correspondence in the unsupervised variant.

Main Results:

• spVIPES effectively disentangles dataset-specific and conserved cellular features.
• Outperformed state-of-the-art methods in batch correction.
• Achieved more accurate cell-type identification across datasets with nonmatching features.

Conclusions:

• spVIPES provides a robust framework for analyzing diverse scRNA-seq data.
• The method successfully addresses limitations of existing batch correction techniques.
• spVIPES enhances cross-dataset cell-type identification, particularly for unpaired data with nonmatching features.