Cryo-EM image alignment: From pair-wise to joint with deep unsupervised difference learning.

Area of Science:

• Structural Biology
• Biophysics
• Computational Imaging

Background:

• Cryo-electron microscopy (cryo-EM) single-particle analysis is crucial for visualizing biomacromolecules at high resolution.
• Accurate image alignment is fundamental for precise distance calculations in cryo-EM but is challenging due to high noise and low signal-to-noise ratios.

Purpose of the Study:

• To introduce a novel deep unsupervised difference learning (UDL) strategy for pair-wise image alignment in cryo-EM.
• To develop a joint UDL (JUDL) variant that leverages dataset-wide similarity for enhanced alignment precision.

Main Methods:

• Implementation of a deep unsupervised difference learning (UDL) strategy with a pseudo-label guided learning network.
• Development of a joint UDL (JUDL) method incorporating global dataset similarity information.
• Application and assessment on both real-world and synthetic cryo-EM single-particle image datasets.

Main Results:

• The proposed JUDL method achieves more accurate image alignment results compared to existing methods.
• The unsupervised framework effectively handles noisy and low signal-to-noise ratio cryo-EM images.
• The method demonstrates high efficiency by utilizing GPU acceleration.

Conclusions:

• The novel unsupervised joint alignment method significantly improves precision in cryo-EM image processing.
• JUDL offers a robust and efficient solution for a critical step in single-particle analysis.
• The source code is publicly available to facilitate academic research.