SCUDDO: an unsupervised clustering algorithm for single-cell Hi-C maps using diagonal diffusion operators

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• High-throughput chromatin conformation capture techniques like Hi-C reveal 3D genome organization.
• Single-cell Hi-C (scHi-C) provides cell-specific interaction maps but suffers from high sparsity, complicating analysis.
• Existing clustering algorithms struggle with scHi-C data sparsity, limiting cell type identification.

Purpose of the Study:

• To introduce a novel unsupervised algorithm, SCUDDO, for embedding and clustering sparse single-cell Hi-C maps.
• To improve the accuracy of cell type identification from scHi-C data.

Main Methods:

• Developed Single-cell Clustering Using Diagonal Diffusion Operators (SCUDDO), an unsupervised algorithm.
• Evaluated SCUDDO on four challenging scHi-C datasets.
• Assessed algorithm performance using Adjusted Rand Index (ARI) and varying data sparsity.

Main Results:

• SCUDDO significantly outperforms current algorithms, improving ARI by over 0.2 on difficult-to-cluster scHi-C datasets.
• SCUDDO maintains superior performance even with reduced intrachromosomal maps or limited contact data.
• The algorithm accurately captures latent features for cell type labeling without prior knowledge.

Conclusions:

• SCUDDO offers a robust solution for clustering sparse single-cell Hi-C data.
• The method enhances the reliability of cell type classification from chromatin conformation data.
• SCUDDO is publicly available and facilitates advanced analysis of 3D genome organization in individual cells.