Integration of single cell multiomics data by deep transfer hypergraph neural network.

Area of Science:

• Single-cell biology
• Computational biology
• Machine learning

Background:

• Multi-omics characterization of single cells is crucial for understanding gene regulatory dynamics.
• Integrating multimodal single-cell data remains challenging due to accuracy and biological variation issues.

Purpose of the Study:

• To present scHyper, a novel deep transfer model for integrating paired and unpaired single-cell multimodal data.
• To address limitations in existing methods regarding accuracy and retention of modality-specific biological variation.

Main Methods:

• scHyper utilizes a low-code, data-efficient deep transfer learning approach.
• The model learns a low-dimensional representation and aligns covariance matrices of measured modalities.
• Benchmarked against diverse multimodal datasets, including large-scale atlas-level data.

Main Results:

• scHyper achieves high accuracy in integrating multimodal single-cell data.
• Demonstrates efficient performance with low memory and computational time on large datasets.
• Successfully sheds light on regulatory relationships between different omics types.

Conclusions:

• scHyper is a versatile and robust tool for single-cell data integration.
• Enables accurate cell-type label transfer from multimodal single-cell datasets.
• Facilitates deeper understanding of gene regulatory states across millions of cells.