Multi-batch single-cell comparative atlas construction by deep learning disentanglement.

Area of Science:

• Single-cell genomics
• Computational biology
• Systems biology

Background:

• Single-cell RNA-seq and ATAC-seq generate cell state atlases crucial for studying genetic and drug perturbations.
• Comparative analysis of these atlases reveals cell state and trajectory changes.
• Multi-batch experiments are common but introduce technical distortions, complicating data comparison.

Purpose of the Study:

• To develop a computational model that disentangles technical batch effects from biological variations in single-cell data.
• To improve the accuracy of cell type discovery and biological interpretation in multi-batch single-cell experiments.

Main Methods:

• Proposed CODAL (Cell-state disentanglement via Autoencoder-based Latent-space decomposition), a variational autoencoder model.
• Utilized mutual information regularization to separate technical and biological factors.
• Applied the model to simulated datasets and embryonic development atlases with gene knockouts.

Main Results:

• CODAL successfully identified cell types despite batch confounds in simulated and real-world data.
• The model improved the representation of both RNA-seq and ATAC-seq data modalities.
• CODAL generated interpretable modules of biological variation and allowed generalization of other models.

Conclusions:

• CODAL effectively addresses batch effects in single-cell multi-omics data analysis.
• The model enhances the reliability of cell state and trajectory analyses in perturbation studies.
• CODAL provides a robust framework for multi-batch single-cell data integration and interpretation.