Unsupervised manifold alignment for single-cell multi-omics data.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Integrating diverse single-cell genomic measurements is crucial for advancing genome biology.
• Aligning datasets from different single-cell experiments is challenging due to a lack of shared features or cell correspondences.
• Unsupervised algorithms are needed to create shared latent spaces for cross-modal single-cell data integration.

Purpose of the Study:

• To evaluate and enhance the Maximum Mean Discrepancy-based Manifold Alignment (MMD-MA) algorithm for large-scale single-cell data integration.
• To adapt MMD-MA for practical application in analyzing thousands of single cells across different genomic modalities.
• To improve the computational efficiency and usability of MMD-MA for the scientific community.

Main Methods:

• Applied MMD-MA to large-scale single-cell datasets measuring transcriptome and chromatin accessibility (~2000 cells).
• Extended the MMD-MA implementation to leverage Graphics Processing Units (GPUs) for accelerated computation.
• Introduced automated parameter selection to reduce the hyperparameter search space and simplify method usage.

Main Results:

• Demonstrated accurate alignment of large-scale single-cell experiments using the enhanced MMD-MA.
• Showcased the scalability of MMD-MA on datasets with thousands of cells through GPU acceleration.
• Validated the effectiveness of automated parameter selection in optimizing alignment performance.

Conclusions:

• The GPU-accelerated MMD-MA effectively integrates large-scale single-cell multi-omic datasets.
• This approach enables robust cross-modal analysis of transcriptome and chromatin accessibility data.
• The enhanced MMD-MA provides a scalable and practical tool for single-cell genomics research.