Single-cell data integration across weakly linked modalities.

Area of Science:

• Computational Biology
• Bioinformatics
• Data Science

Background:

• Single-cell technologies generate multimodal data, often with weak correlations between modalities.
• Integrating these weakly linked datasets presents computational challenges in modeling cell relationships and learning representations.

Purpose of the Study:

• To present MMIHCL, a novel deep learning framework for multimodal single-cell data integration.
• To address challenges in modeling cell relationships and learning meaningful cell representations.

Main Methods:

• MMIHCL utilizes an optimized adaptive k-nearest neighbor graph to model cell pair-wise relationships.
• The framework employs hypergraph contrastive learning to capture high-order graph information for cell representation.

Main Results:

• MMIHCL achieves high-quality integration across diverse weakly linked datasets.
• It maintains accuracy in strongly linked scenarios and enables accurate cross-modality feature prediction.
• MMIHCL demonstrates utility in disease classification and drug target discovery.

Conclusions:

• MMIHCL offers a robust solution for multimodal single-cell data integration, particularly for weakly linked data.
• The framework enhances downstream applications by providing optimized joint embeddings and enabling explicit cell matching.
• A Python implementation is publicly available for broader research use.