Denoising Autoencoder Trained on Simulation-Derived Structures for Noise Reduction in Chromatin Scanning Transmission...

Area of Science:

• Structural Biology
• Genomics
• Computational Biology

Background:

• Scanning transmission electron microscopy tomography with ChromEM staining (ChromSTEM) enables 3D genome organization studies.
• Understanding chromatin fiber structure is crucial for gene regulation and DNA accessibility.

Purpose of the Study:

• To develop a computational method for enhancing ChromSTEM images to achieve nucleosome-level resolution.
• To investigate chromatin folding motifs and their impact on DNA accessibility.

Main Methods:

• Development of a denoising autoencoder (DAE) using convolutional neural networks and molecular dynamics simulations.
• Training the DAE on synthetic chromatin fiber images generated by the 1-cylinder per nucleosome (1CPN) model.
• Postprocessing experimental ChromSTEM images to remove noise and enhance structural features.

Main Results:

• The DAE effectively removes noise from high-angle annular dark field (HAADF) STEM images, outperforming other denoising algorithms.
• The DAE resolves nucleosome-level structural features, including α-tetrahedron tetranucleosome motifs.
• No evidence for the proposed 30 nm fiber structure was found, suggesting alternative higher-order chromatin organization.

Conclusions:

• The DAE provides high-resolution images of chromatin, enabling the visualization of single nucleosomes and organized domains.
• This method reveals folding motifs that modulate DNA accessibility, offering new insights into genome regulation.
• The findings challenge the long-standing 30 nm fiber model and propose a new perspective on chromatin higher-order structure.