Partially shared multi-modal embedding learns holistic representation of cell state.

Area of Science:

• Computational biology
• Single-cell multi-omics analysis
• Bioinformatics

Background:

• Single-cell technologies generate diverse data types simultaneously.
• Current integration methods often obscure modality-specific contributions.
• Need for methods that retain and distinguish shared and modality-specific information.

Purpose of the Study:

• Introduce APOLLO, a computational framework for integrating multi-modal single-cell data.
• Enable learning of partial information sharing between different data types.
• Provide a more interpretable and holistic view of cell states.

Main Methods:

• Developed an Autoencoder with a Partially Overlapping Latent space learned through Latent Optimization (APOLLO).
• Tested on simulated data and four real-world single-cell datasets (SHARE-seq, CITE-seq, multiplexed imaging).

Main Results:

• APOLLO successfully integrates diverse single-cell data modalities.
• Enables prediction of missing data, such as unmeasured protein stains.
• Allows disentangling modality or cellular compartment contributions to specific phenotypes.

Conclusions:

• APOLLO offers an efficient approach to multi-modal single-cell data integration.
• Retains and distinguishes shared and modality-specific information for enhanced interpretability.
• Facilitates a holistic understanding of cell states and phenotypes.