SCIM

SCIM recovers correspondences between single cells profiled with different omics technologies to enable integrated multi-omics analyses.

Key Features:

• Universal matching across technologies: Integrates datasets from diverse single-cell profiling technologies without requiring pairwise cell-level correspondence.
• Technology-invariant latent space: Constructs a low-dimensional latent space invariant to technological differences using an auto-encoder framework complemented by an adversarial objective.
• Bipartite matching on latent representations: Performs bipartite matching operating on low-dimensional latent representations to pair cells across technologies.

Scientific Applications:

• Biological and clinical insights: Enables discovery of biologically or clinically meaningful observations by unifying complementary measurements from multiple technologies.
• Simulated cellular branching process: Recovers cell-to-cell matches that reflect pseudotime in a simulated branching process.
• Melanoma tumor sample (scRNA–CyTOF): Demonstrated 93% cell-matching accuracy between single-cell RNA sequencing (scRNA) and CyTOF measurements.
• Human bone marrow sample: Demonstrated 84% cell-matching accuracy in matching across technologies.

Methodology:

Assumes a shared low-dimensional cellular structure across technologies; constructs a technology-invariant latent space via an auto-encoder with an adversarial objective; applies bipartite matching on the resulting low-dimensional latent representations.