BiCLUM: Bilateral contrastive learning for unpaired single-cell multi-omics integration

Area of Science:

• Single-cell multi-omics
• Computational biology
• Genomics

Background:

• Single-cell multi-omics data integration is crucial for understanding molecular interplay.
• Existing methods struggle with unpaired datasets and limited cross-modal insights.
• Challenges include unknown cell correspondences and cell-type-specific activity in non-RNA modalities.

Purpose of the Study:

• To develop a robust method for integrating unpaired single-cell multi-omics data.
• To simultaneously align cell-level and feature-level information across modalities.
• To improve visualization, quantitative benchmarks, and biological interpretation of integrated data.

Main Methods:

• BiCLUM (Bilateral Contrastive Learning for Unpaired single-cell Multi-omics integration) framework.
• Transformation of one modality into another's data space using genomic knowledge.
• Bilateral contrastive learning with cell-level and feature-level losses for embedding generation.

Main Results:

• BiCLUM outperforms existing methods in visualization and quantitative benchmarks on RNA+ATAC and RNA+protein datasets.
• Preserves biologically relevant regulatory relationships between chromatin accessibility and gene expression.
• Facilitates downstream analyses including transcription factor activity inference and cell-cell interaction mapping.

Conclusions:

• BiCLUM offers a robust and interpretable framework for effective cross-modal alignment.
• Successfully retains the underlying regulatory and functional landscape across single-cell modalities.
• Enables deeper biological insights from integrated unpaired single-cell multi-omics data.