LSMMD-MA: scaling multimodal data integration for single-cell genomics data analysis

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell omics data integration is crucial for unifying insights from diverse genomic assays.
• Current multimodal computational methods struggle to scale to large single-cell datasets (millions of cells).

Purpose of the Study:

• To develop a scalable computational method for modality matching in large-scale single-cell omics data.
• To enable effective multimodal data integration for biological and clinical discovery.

Main Methods:

• We introduce LSMMD-MA, a large-scale Python implementation of the MMD-MA method.
• The optimization problem is reformulated using linear algebra and solved with KeOps, a CUDA framework for symbolic matrix computation.
• This approach enables efficient processing of large datasets.

Main Results:

• LSMMD-MA demonstrates scalability to one million cells per modality, a two-orders-of-magnitude improvement over existing methods.
• The method facilitates robust multimodal data integration for large single-cell datasets.
• Successful integration of diverse single-cell omics data types.

Conclusions:

• LSMMD-MA overcomes computational limitations in single-cell multimodal data integration.
• The method unlocks the potential for deeper biological and clinical insights from large-scale omics studies.
• LSMMD-MA is publicly available for the research community.