Mosaic integration and knowledge transfer of single-cell multimodal data with MIDAS.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Integrating single-cell data from multiple omics technologies is crucial for understanding cellular heterogeneity.
• Mosaic integration, where datasets share limited modalities, presents significant challenges in modality alignment and batch effect removal.

Purpose of the Study:

• To present a novel deep probabilistic framework, MIDAS (Mosaic Integration and Knowledge Transfer), for the integration of single-cell multimodal data.
• To address challenges in mosaic integration, including dimensionality reduction, imputation, and batch correction.

Main Methods:

• Developed a deep probabilistic framework utilizing self-supervised modality alignment and information-theoretic latent disentanglement.
• Implemented MIDAS for simultaneous dimensionality reduction, imputation, and batch correction of mosaic single-cell data.
• Constructed a single-cell trimodal atlas of human peripheral blood mononuclear cells for knowledge transfer.

Main Results:

• MIDAS demonstrated superior performance compared to 19 other methods in trimodal and mosaic integration tasks.
• Validated the reliability and versatility of MIDAS through applications in mosaic integration, pseudotime analysis, and cross-tissue knowledge transfer.
• Successfully enabled flexible and accurate knowledge transfer using tailored transfer learning and reciprocal reference mapping schemes.

Conclusions:

• MIDAS provides a robust and versatile solution for the mosaic integration of single-cell multimodal data.
• The framework facilitates accurate knowledge transfer from atlases to new datasets, advancing single-cell data analysis.
• MIDAS offers significant improvements in handling complex multimodal single-cell datasets and their integration.