Dictionary learning for integrative, multimodal and scalable single-cell analysis.

Area of Science:

• Single-cell multiomics
• Computational biology
• Immunology

Background:

• Single-cell sequencing is crucial for biological research.
• Current reference datasets primarily use gene expression data.
• Limitations exist in annotating datasets lacking gene expression information.

Purpose of the Study:

• To develop a novel method for integrating single-cell datasets across different molecular modalities.
• To enable the annotation of diverse single-cell data types using a unified reference.
• To enhance the comparability of single-cell experiments measuring various molecular features.

Main Methods:

• Introduced 'bridge integration' utilizing multiomic datasets as a molecular bridge.
• Developed a dictionary-based approach to reconstruct and transform unimodal datasets into a shared space.
• Combined dictionary learning with sketching techniques for improved computational scalability.

Main Results:

• Successfully integrated transcriptomic data with single-cell chromatin accessibility, histone modifications, DNA methylation, and protein levels.
• Demonstrated accurate integration across diverse molecular modalities.
• Harmonized 8.6 million human immune cell profiles from sequencing and mass cytometry experiments.

Conclusions:

• Bridge integration broadens the utility of single-cell reference datasets.
• Facilitates cross-modal comparisons in single-cell biology.
• The method is implemented in the Seurat toolkit (v5).