sCIN: a contrastive learning framework for single-cell multi-omics data integration.

Area of Science:

• Single-cell multi-omics integration
• Computational biology
• Genomics and transcriptomics

Background:

• Single-cell omics technologies like scRNA-seq and scATAC-seq have advanced cellular heterogeneity studies.
• Integrating multi-omics data is challenging due to distributional discrepancies and distinct feature spaces.

Purpose of the Study:

• To present a novel framework, single-cell Contrastive INtegration (sCIN), for integrating diverse single-cell omics modalities.
• To enable the combination of different omics data types into a shared latent space, overcoming technical biases.

Main Methods:

• Developed sCIN, a framework utilizing modality-specific encoders and contrastive learning.
• Ensured rigorous prevention of data leakage between training and testing sets.
• Evaluated on paired (scRNA-ATAC, 10X PBMC, CITE-seq) and unpaired (gene expression, chromatin accessibility) datasets.

Main Results:

• sCIN demonstrated superior performance compared to state-of-the-art models (scGLUE, scBridge, sciCAN, Con-AAE, Harmony, MOFA+) on multiple integration and clustering metrics.
• Effective integration was shown on both paired and unpaired datasets, including simulated unpaired data.
• The framework successfully preserved biological meaning during multimodal integration.

Conclusions:

• sCIN provides a robust solution for integrating single-cell omics modalities.
• The framework effectively addresses technical biases and preserves biological insights in both paired and unpaired multi-omics data.
• sCIN advances the understanding of cellular heterogeneity and regulatory mechanisms through reliable data integration.