UNSUPERVISED DOMAIN ALIGNMENT BASED OPEN SET STRUCTURAL RECOGNITION OF MACROMOLECULES CAPTURED BY CRYO-ELECTRON...

Area of Science:

• Structural Biology
• Biophysics
• Computational Biology

Background:

• Cellular cryo-Electron Tomography (cryo-ET) visualizes cellular macromolecules in 3D.
• Subtomogram classification is crucial for identifying these structures.
• Current deep learning methods struggle with unknown macromolecular classes.

Purpose of the Study:

• To develop an open set learning method for subtomogram classification.
• To enable recognition of previously unknown macromolecular structures.
• To enhance the power of automatic subtomogram classification in cryo-ET.

Main Methods:

• Proposed Margin-based Loss for Unsupervised Domain Alignment (MLUDA) for open set recognition.
• Applied MLUDA to cross-domain classification problems with limited shared categories.
• Validated performance on public and medical imaging datasets.

Main Results:

• MLUDA demonstrated strong performance in cross-domain open-set classification.
• The method successfully recognized unknown macromolecular structural classes.
• Experiments confirmed the practical importance and effectiveness of MLUDA.

Conclusions:

• MLUDA advances automatic subtomogram classification by enabling open set recognition.
• The method is valuable for analyzing complex cellular structures in cryo-ET.
• MLUDA has significant practical applications in structural biology and medical imaging.