Active learning to classify macromolecular structures in situ for less supervision in cryo-electron tomography

Area of Science:

• Structural biology
• Bioimaging
• Computational biology

Background:

• Cryo-electron tomography (cryo-ET) enables near-native visualization of macromolecular structures within cells.
• Challenges in cryo-ET include complex structures and imaging limitations, hindering systematic analysis.
• Deep learning for subtomogram classification requires extensive, labor-intensive manual annotations.

Purpose of the Study:

• To develop an efficient method for subtomogram classification in cryo-electron tomography (cryo-ET).
• To reduce the burden of manual data labeling for deep learning models in cryo-ET.
• To improve the accuracy and efficiency of macromolecular structure identification using cryo-ET data.

Main Methods:

• Proposed a Hybrid Active Learning (HAL) framework for subtomogram selection.
• Employed uncertainty sampling to identify informative subtomograms.
• Integrated a discriminator for unbiased sampling and subset sampling to enhance diversity and efficiency.

Main Results:

• Achieved comparable testing performance with less than 30% of labeled subtomograms.
• Demonstrated an average accuracy drop of only 3% compared to fully supervised methods.
• Showcased the effectiveness of HAL on both simulated and real cryo-ET data.

Conclusions:

• The HAL framework significantly reduces labeling requirements for cryo-ET subtomogram classification.
• HAL offers a promising solution for analyzing complex macromolecular structures with limited labeled data.
• This approach enhances the practical utility of cryo-ET in structural biology research.