TomoCPT: a generalizable model for 3D particle detection and localization in cryo-electron tomograms.

Area of Science:

• Structural Biology
• Biophysics
• Computational Biology

Background:

• Cryo-electron tomography (cryo-ET) is crucial for visualizing macromolecular complexes in situ.
• Accurate particle identification in cryo-tomograms is a major bottleneck for downstream analysis like subtomogram averaging.
• Existing methods for particle detection in cryo-ET have limitations in speed and accuracy.

Purpose of the Study:

• To develop an advanced computational tool for efficient and accurate particle detection in cryo-tomograms.
• To address the limitations of current particle identification methods in cryo-electron tomography.
• To improve the throughput and resolution of cryo-ET data processing.

Main Methods:

• Development of tomoCPT (Tomogram Centroid Prediction Tool), a transformer-based deep learning model.
• Reformulation of particle detection as a centroid-prediction task using Gaussian labels.
• Utilizing the SwinUNETR architecture for enhanced performance in particle identification.

Main Results:

• TomoCPT demonstrated superior performance over conventional binary labeling and template matching methods.
• Achieved high-resolution reconstructions for apoferritin (3.0 Å), SARS-CoV-2 spike proteins (18.3 Å), and rubisco (8.0 Å).
• Showcased effective generalization to novel particle types via zero-shot inference and significant improvement with fine-tuning.

Conclusions:

• TomoCPT offers a practical and efficient solution for particle detection in cryo-electron tomography.
• The tool effectively handles complex scenarios, including densely packed molecules and membrane-bound proteins.
• Its command-line implementation and minimal data requirements make it suitable for high-throughput cryo-ET workflows.