Learning interpretable cellular and gene signature embeddings from single-cell transcriptomic data.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) has transformed transcriptomics.
• Analyzing large-scale scRNA-seq data is challenging due to batch effects and limitations in current computational methods' scalability, interpretability, and transferability.

Purpose of the Study:

• To introduce single-cell Embedded Topic Model (scETM), a novel computational framework for large-scale integrative analysis of scRNA-seq data.
• To address the challenges of batch effects, interpretability, and scalability in scRNA-seq data integration.

Main Methods:

• scETM utilizes a transferable neural-network-based encoder and an interpretable linear decoder based on matrix tri-factorization.
• It simultaneously learns cell type mixtures, gene embeddings, topic embeddings, and batch-effect intercepts from multiple scRNA-seq datasets.
• The model is designed to be scalable to over 10^6 cells.

Main Results:

• scETM demonstrates remarkable cross-tissue and cross-species zero-shot transfer-learning capabilities.
• Gene set enrichment analysis reveals that scETM-learned topics are significantly enriched in biologically meaningful and disease-related pathways.
• The model allows direct learning of associations between pathways and topics by incorporating known gene sets.

Conclusions:

• scETM offers a scalable and interpretable solution for integrating multiple scRNA-seq datasets.
• The framework facilitates robust transfer learning across different biological contexts.
• scETM enhances biological discovery by linking gene expression patterns to known biological pathways and disease relevance.