Cryo-shift: reducing domain shift in cryo-electron subtomograms with unsupervised domain adaptation and randomization

Area of Science:

• Structural biology
• Computational biology
• Microscopy

Background:

• Cryo-electron tomography (cryo-ET) visualizes subcellular structures at near-atomic resolution.
• Subtomogram classification is crucial for analyzing macromolecular structures within cells.
• Supervised deep learning excels at classification but requires extensive annotated data, which is often scarce.

Purpose of the Study:

• To develop an unsupervised domain adaptation and randomization framework for cross-domain subtomogram classification.
• To address the performance limitations of deep learning models trained on simulated data when applied to experimental cryo-ET data.
• To enable accurate subtomogram classification without the need for labeled experimental data.

Main Methods:

• Implemented Cryo-Shift, an unsupervised framework utilizing multi-adversarial domain adaptation.
• Integrated a network-driven domain randomization procedure with 'warp' modules to enhance generalization.
• Focused on reducing the domain shift between simulated and experimental cryo-ET data features.

Main Results:

• Cryo-Shift achieved superior performance in cross-domain subtomogram classification compared to existing methods.
• The framework effectively reduces the domain shift between simulated and experimental data.
• No labeled experimental data was required for training, demonstrating the power of unsupervised learning.

Conclusions:

• Cryo-Shift offers a robust and efficient solution for subtomogram classification in cryo-ET.
• The unsupervised approach significantly broadens the applicability of deep learning in structural biology.
• This method advances the analysis of macromolecular structures from cryo-ET data.