SCIM: universal single-cell matching with unpaired feature sets

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Single-cell technologies generate vast amounts of data, offering unprecedented biological insights.
• Integrating multi-technology single-cell measurements is crucial for unifying perspectives and discovering meaningful observations.
• Current methods struggle with cell correspondences due to cell consumption by profiling technologies and dataset size.

Purpose of the Study:

• To develop a scalable algorithm for recovering cell correspondences across multiple single-cell profiling technologies.
• To enable the integration of multi-modal single-cell datasets by accurately matching cells.
• To facilitate the unification of biological insights from diverse single-cell measurements.

Main Methods:

• Proposed Single-Cell data Integration via Matching (SCIM), a scalable approach using an autoencoder with an adversarial objective.
• Constructed a technology-invariant latent space assuming a common underlying cell structure.
• Employed a bipartite matching scheme on latent representations to pair cells across technologies.

Main Results:

• SCIM accurately reflects pseudotime in simulated cellular branching processes.
• Achieved 90% cell-matching accuracy when integrating scRNA and CyTOF data from a melanoma tumor sample.
• Demonstrated 78% cell-matching accuracy for a human bone marrow sample integrating scRNA and CyTOF data.

Conclusions:

• SCIM provides a scalable and accurate method for integrating multi-technology single-cell data.
• The ability to match sibling cells across datasets enhances biological and clinical discovery.
• SCIM facilitates deeper understanding by unifying perspectives from diverse single-cell profiling technologies.