Data-driven regularization lowers the size barrier of cryo-EM structure determination

Area of Science:

• Structural biology
• Biophysics
• Computational biology

Background:

• Electron cryo-microscopy (cryo-EM) is crucial for determining macromolecular structures.
• Image alignment of noisy particle images is a key limitation, especially for smaller molecules with weak signals.
• Current alignment methods struggle with low signal-to-noise ratios, limiting cryo-EM's applicability.

Purpose of the Study:

• To improve image alignment in cryo-EM using deep learning.
• To overcome size limitations imposed by alignment errors in cryo-EM.
• To develop a novel regularization technique for enhanced cryo-EM reconstructions.

Main Methods:

• Trained a denoising convolutional neural network on cryo-EM data from the Electron Microscopy Data Bank (EMDB).
• Applied the trained neural network as Blush regularization, an alternative to traditional smoothness priors.
• Evaluated reconstruction quality against existing algorithms, particularly for low signal-to-noise data.

Main Results:

• Blush regularization significantly improved cryo-EM image alignment and reconstruction quality.
• The method outperformed existing algorithms, especially in challenging low signal-to-noise conditions.
• Successfully reconstructed a previously intractable 40 kDa protein-nucleic acid complex.

Conclusions:

• Deep learning-based denoising neural networks enhance cryo-EM structure determination.
• Blush regularization expands the scope of cryo-EM to smaller and lower-signal biological macromolecules.
• This advancement broadens the applicability of cryo-EM across diverse biological research areas.