Refinement of cryo-EM 3D maps with a self-supervised denoising model: crefDenoiser

Area of Science:

• Structural Biology
• Biophysics
• Computational Biology

Background:

• Cryo-electron microscopy (cryo-EM) is crucial for macromolecular structure determination.
• High-resolution 3D protein density maps are often degraded by noise, limiting structural accuracy.
• Existing methods struggle to effectively remove noise without compromising map quality.

Purpose of the Study:

• Introduce crefDenoiser, a novel neural network for denoising 3D cryo-EM maps.
• Enhance signal quality and improve the visibility of structural features in cryo-EM data.
• Develop a self-supervised denoising approach that does not require clean ground-truth maps.

Main Methods:

• Developed crefDenoiser, a self-supervised neural network model.
• Trained the model using real noisy protein half-maps from the Electron Microscopy Data Bank.
• Optimized the model to target a theoretical noise-free map during training.

Main Results:

• crefDenoiser successfully amplifies signal across diverse protein maps.
• The model outperforms a classic denoising method and a network-based sharpening model.
• Improved visibility of protein domains, secondary structures, and restoration of modest high-resolution features.

Conclusions:

• crefDenoiser effectively denoises 3D cryo-EM maps, enhancing structural detail.
• The self-supervised approach offers a robust alternative for improving cryo-EM map quality.
• This method aids in more accurate macromolecular structure representation without introducing bias.