Volumetric macromolecule identification in cryo-electron tomograms using capsule networks.

Area of Science:

• Structural biology
• Computational biology
• Biophysics

Background:

• Automated macromolecule identification in cellular cryo-electron tomography (CET) is challenging due to data limitations and complex structures.
• Current deep learning methods primarily use Convolutional Neural Networks (CNNs), limiting performance.
• Identifying diverse macromolecules is a time-consuming manual process.

Purpose of the Study:

• To develop an automated tool for macromolecule identification using a capsule-based architecture.
• To improve the accuracy and efficiency of macromolecule identification in submolecular resolution.
• To address the limitations of existing CNN-based approaches.

Main Methods:

• A novel capsule-based architecture, 3D-UCaps, was developed.
• The architecture comprises a feature extractor, a 3D Capsule Network (CapsNet) encoder, and a 3D CNN decoder.
• The method processes sub-tomogram voxel intensities to identify and localize macromolecules.

Main Results:

• 3D-UCaps achieved an F1-score above 70% on test data, outperforming 3D-UNet (60%).
• In multi-class identification of experimental data, 3D-UCaps reached a 91% F1-score compared to 3D-UNet's 64%.
• The capsule network encoder contributed to higher precision and improved performance with increased network depth.

Conclusions:

• 3D-UCaps effectively performs macromolecule identification and localization, competing with and often surpassing CNN architectures.
• The capsule layers' ability to extract molecular orientation offers potentially more interpretable data representations.
• The developed method demonstrates significant potential for advancing structural biology analysis.