Scalable and unbiased sequence-informed embedding of single-cell ATAC-seq data with CellSpace

Area of Science:

• Genomics
• Computational Biology
• Single-cell analysis

Background:

• Standard single-cell ATAC sequencing (scATAC-seq) analysis represents cells as sparse vectors, overlooking valuable genomic sequence information at accessible DNA regions.
• This limitation hinders the comprehensive understanding of cellular heterogeneity and regulatory mechanisms.

Purpose of the Study:

• To introduce CellSpace, an efficient and scalable algorithm for scATAC-seq that integrates genomic sequence information.
• To address the limitations of existing methods by learning a joint embedding space for DNA k-mers and cells.

Main Methods:

• CellSpace employs a sequence-informed embedding approach, mapping DNA k-mers and single cells into a shared latent space.
• The algorithm learns representations that capture the underlying genomic sequence features associated with cell accessibility profiles.

Main Results:

• CellSpace successfully identifies meaningful latent structures within scATAC-seq data, including distinct cell subpopulations and developmental trajectories.
• The method enables the scoring of transcription factor activities in individual cells by assessing their proximity to embedded binding motifs.
• CellSpace demonstrates robust mitigation of batch effects across diverse datasets, even those processed with different peak atlases.

Conclusions:

• CellSpace offers a powerful, sequence-aware approach for analyzing scATAC-seq data, enhancing the interpretation of cellular states and regulatory elements.
• This algorithm facilitates the integration and analysis of large-scale scATAC-seq compendia, advancing single-cell genomics research.