Quantifying common and distinct information in single-cell multimodal data with Tilted Canonical Correlation...

Area of Science:

• Single-cell biology
• Computational biology
• Bioinformatics

Background:

• Multimodal single-cell technologies provide comprehensive cell characterization by profiling multiple data types per cell.
• Current dimension-reduction methods often combine information across modalities, focusing on the union of data.
• A need exists to disentangle shared and modality-specific information for deeper biological insights.

Purpose of the Study:

• To develop a novel method, Tilted Canonical Correlation Analysis (Tilted-CCA), for decomposing multimodal single-cell data.
• To separate and quantify shared ('intersection') and unique ('distinct') information between paired modalities.
• To enable new downstream analyses for optimizing experimental design and biological interpretation.

Main Methods:

• Tilted Canonical Correlation Analysis (Tilted-CCA) was developed to decompose paired multimodal datasets into three distinct lower-dimensional embeddings.
• The method identifies shared geometric relations (intersection) and modality-specific relations (distinct).
• Tilted-CCA was applied to CITE-seq (RNA + surface antibodies) and 10x (RNA + chromatin accessibility) datasets.

Main Results:

• Tilted-CCA successfully visualized and quantified cross-modal information in single-cell datasets.
• The method demonstrated utility in designing optimal antibody panels for transcriptome complementarity.
• Tilted-CCA aided in identifying developmental genes and distinguishing cell types in developmental datasets.

Conclusions:

• Tilted-CCA offers a novel framework for analyzing multimodal single-cell data by separating shared and unique information.
• This approach provides enhanced visualization and quantification of inter-modal relationships.
• Tilted-CCA facilitates targeted downstream applications in experimental design and biological discovery.