A comprehensive benchmark of single-cell Hi-C embedding tools.

Area of Science:

• Genomics and Computational Biology
• Epigenetics and Genome Architecture

Background:

• Single-cell Hi-C (scHi-C) analysis relies on embedding to capture genome architecture heterogeneity.
• Existing embedding tools vary in performance across different scHi-C datasets and biological questions.

Purpose of the Study:

• To benchmark thirteen scHi-C embedding tools, including the novel convolutional neural network model Va3DE.
• To evaluate the impact of preprocessing options and data representation on embedding performance.
• To identify optimal embedding strategies for different biological scales (compartment vs. loop) and resolutions.

Main Methods:

• Developed a software framework to decouple preprocessing options for benchmarking.
• Evaluated thirteen embedding tools on ten diverse scHi-C datasets.
• Assessed performance based on capturing long-range (compartment) and short-range (loop) genomic contacts.

Main Results:

• No single embedding tool performed best across all datasets under default settings.
• Data representation and preprocessing significantly impact embedding performance.
• Deep learning methods demonstrated superior versatility in handling sparsity and different resolutions compared to random-walk and IDF methods.
• Diagonal integration shows promise for distinguishing similar cell subpopulations.

Conclusions:

• The choice of embedding method and preprocessing is critical for accurate scHi-C analysis.
• Deep learning approaches offer a more robust and versatile solution for capturing genome architecture heterogeneity at various scales.
• Further research into appropriate priors and integration strategies is warranted for advancing scHi-C embedding.