Complex hierarchical structures analysis in single-cell data with Poincaré deep manifold transformation.

Area of Science:

• Computational Biology
• Genomics
• Data Science

Background:

• Single-cell RNA sequencing (scRNA-seq) provides high-resolution cellular data.
• Current dimensionality reduction methods struggle with complex cell trajectories and data distortion.
• Effective visualization and interpretation of scRNA-seq data remain challenging.

Purpose of the Study:

• To introduce Poincaré deep manifold transformation (PoincaréDMT) for improved scRNA-seq data analysis.
• To address limitations in structure preservation and distortion in existing embedding methods.
• To provide a unified framework for trajectory inference, batch correction, and marker gene selection.

Main Methods:

• Mapping high-dimensional scRNA-seq data to a hyperbolic Poincaré disk.
• Utilizing graph Laplacian for global structure preservation and a structure module with data augmentation for local correction.
• Integrating a batch graph to mitigate batch effects during network training.
• Employing Shapley additive explanations for marker gene identification.

Main Results:

• PoincaréDMT effectively preserves both global and local data structures in scRNA-seq datasets.
• Demonstrated superior performance in trajectory inference, pseudotime inference, and batch correction compared to existing methods.
• Successfully identified important marker genes crucial for cellular differentiation processes.

Conclusions:

• PoincaréDMT offers a robust and unified framework for critical scRNA-seq analysis tasks.
• The hyperbolic embedding approach significantly enhances the interpretation of complex cellular dynamics.
• Validated through extensive experiments on simulated and real-world scRNA-seq data, showcasing its efficacy.