Reconstructing 3D chromosome structures from single-cell Hi-C data with SO(3)-equivariant graph neural networks

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• The 3D spatial arrangement of chromosomes influences cellular functions like gene expression.
• Single-cell Hi-C (ScHi-C) captures chromosomal contacts but yields sparse data.
• Reconstructing accurate 3D genome structures from sparse ScHi-C data is computationally challenging.

Purpose of the Study:

• To develop a novel machine learning method for reconstructing single-cell 3D chromosome structures.
• To overcome limitations of traditional methods in handling sparse ScHi-C data.
• To improve the accuracy and robustness of 3D genome structure reconstruction.

Main Methods:

• Utilized a novel SO(3)-equivariant graph neural network (HiCEGNN).
• Applied HiCEGNN to reconstruct 3D chromosome structures from single-cell Hi-C data.
• Compared HiCEGNN performance against traditional optimization and existing deep learning methods.

Main Results:

• HiCEGNN demonstrated superior performance in reconstructing 3D chromosome structures.
• The method consistently outperformed traditional and deep learning approaches.
• HiCEGNN showed robustness across diverse cell types, resolutions, and noise levels.

Conclusions:

• HiCEGNN provides a powerful and robust machine learning solution for 3D genome structure reconstruction from ScHi-C data.
• This advancement facilitates a deeper understanding of the relationship between genome organization and cellular function.
• The method offers improved accuracy for analyzing chromosome conformation in single cells.