Adaptive differentiable grids for cryo-electron tomography reconstruction and denoising

Area of Science:

• Structural Biology
• Biophysics
• Computational Biology

Background:

• Cryo-electron tomography (cryo-ET) is crucial for 3D structural biology.
• Reconstruction in cryo-ET faces challenges like missing wedge data, low signal-to-noise ratio, and motion artifacts.

Purpose of the Study:

• To introduce an adaptive learning-based representation for improved cryo-ET density field reconstruction.
• To address limitations of existing cryo-ET reconstruction methods.

Main Methods:

• Proposed an octree-based adaptive learning representation for the sample's density field.
• Utilized a differentiable image formation model with regularization terms (total variation, boundary consistency, cross-nodes non-local constraint) for optimization.
• Reconstructed tomograms by interpolating the learned density grid.

Main Results:

• The adaptive representation effectively handles missing wedge data in cryo-ET.
• Demonstrated significant improvement in the signal-to-noise ratio of reconstructed tomograms.
• Achieved higher reconstruction quality compared to state-of-the-art methods with reduced computational time.

Conclusions:

• The proposed adaptive learning-based method offers a robust solution for cryo-ET reconstruction.
• This approach enhances the resolution and quality of 3D structural data.
• The method shows promise for advancing structural biology research.