High-throughput cryo-ET structural pattern mining by unsupervised deep iterative subtomogram clustering.

Area of Science:

• Structural biology
• Biophysics
• Computational biology

Background:

• Cryoelectron tomography (cryo-ET) visualizes cellular structures in situ.
• Current methods for analyzing cryo-ET data are limited by throughput and reliance on templates or manual labels.

Purpose of the Study:

• To develop a high-throughput, template- and label-free deep learning approach for automated macromolecular structure detection in cryo-ET data.
• To overcome limitations of existing computer-assisted structure sorting methods.

Main Methods:

• Introduced Deep Iterative Subtomogram Clustering Approach (DISCA), a novel deep learning algorithm.
• DISCA learns and models 3D structural features and their distributions for automated detection.
• Employed an unsupervised learning strategy, eliminating the need for templates and manual labels.

Main Results:

• DISCA demonstrated high-throughput capability in analyzing cryo-ET datasets.
• The method successfully detected diverse macromolecular structures across a range of molecular sizes.
• Evaluation on five experimental datasets confirmed the effectiveness of the unsupervised deep learning approach.

Conclusions:

• Unsupervised deep learning, as implemented in DISCA, offers a powerful solution for analyzing heterogeneous macromolecular structures.
• This approach enables systematic and unbiased recognition of macromolecular complexes within their native cellular contexts.
• DISCA paves the way for advancing in situ structural biology and understanding cellular organization.