A unified model for interpretable latent embedding of multi-sample, multi-condition single-cell data.

Area of Science:

• Computational biology
• Single-cell genomics
• Machine learning

Background:

• Single-cell analysis across multiple samples necessitates modeling cell state continua and variability sources.
• Integrating multi-sample, multi-condition single-cell data presents significant computational challenges.

Purpose of the Study:

• To introduce GEDI, a generative model for identifying and attributing latent space variations in multi-sample single-cell datasets.
• To enable advanced analyses including cross-sample cell state mapping, differential gene expression, and sample characteristic prediction.

Main Methods:

• Developed GEDI, a generative model for latent space variation identification.
• Applied GEDI to multi-sample, multi-condition single-cell datasets.
• Incorporated gene-level prior knowledge for pathway and regulatory network inference.
• Extended the model to dual-measurement modalities.

Main Results:

• GEDI achieves state-of-the-art cross-sample cell state mapping.
• Enables cluster-free differential gene expression analysis along cell state continua.
• Facilitates machine learning-based prediction of sample characteristics.
• Successfully models alternative splicing and mRNA stability from dual measurements.

Conclusions:

• GEDI provides a robust framework for analyzing complex single-cell data.
• The model enhances understanding of biological variability and cell state dynamics.
• GEDI offers novel capabilities for multi-modal single-cell data integration and interpretation.