scHiCEmbed: Bin-Specific Embeddings of Single-Cell Hi-C Data Using Graph Auto-Encoders.

Area of Science:

• Genomics
• Computational Biology
• Structural Biology

Background:

• Single-cell Hi-C data is often sparse, limiting high-resolution 3D genome structure analysis.
• Learning latent representations from sparse data offers a novel approach to extract valuable information.
• Existing methods struggle with the inherent sparsity and cell-to-cell variability in single-cell Hi-C.

Purpose of the Study:

• To develop an unsupervised computational method, scHiCEmbed, for learning bin-specific embeddings from sparse single-cell Hi-C data.
• To apply scHiCEmbed for 3D genome structure reconstruction and topologically associating domain (TAD) detection.
• To investigate chromatin dynamics and cell-to-cell variability in TAD organization.

Main Methods:

• scHiCEmbed utilizes graph auto-encoders to embed genomic bins into a higher-dimensional latent space.
• Input data includes raw or imputed single-cell Hi-C matrices.
• Constrained hierarchical clustering with S_Dbw is employed for TAD detection on the learned latent matrix.

Main Results:

• Reconstructed 3D genome structures reveal chromatin expansion during the cell cycle.
• Detected TADs exhibit significant cell-to-cell variability and differ from bulk Hi-C data.
• The method successfully identifies potential TADs and characterizes dynamic chromatin structures.

Conclusions:

• scHiCEmbed effectively learns from sparse single-cell Hi-C data, enabling robust 3D genome structure and TAD analysis.
• The findings highlight significant heterogeneity in TAD organization across individual cells.
• The computational system provides insights into cell cycle-dependent chromatin dynamics and TAD variability.