Template Learning: Deep learning with domain randomization for particle picking in cryo-electron tomography.

Area of Science:

• Structural Biology
• Biophysics
• Computational Biology

Background:

• Cryo-electron tomography (cryo-ET) visualizes biomolecules in near-native states.
• Particle picking in cryo-ET is challenging, often relying on template matching or supervised deep learning.
• Supervised deep learning requires extensive annotated datasets, limiting its practical application.

Purpose of the Study:

• To develop an automated particle picking method for cryo-ET that reduces reliance on annotated data.
• To combine the accuracy of deep learning with the convenience of template-based training.
• To improve the precision and efficiency of particle detection in cryo-ET data analysis.

Main Methods:

• Introduced Template Learning, a technique utilizing deep learning with domain randomization for synthetic dataset generation.
• Modeled molecular crowding, structural variability, and data acquisition variations in synthetic data.
• Trained models using automated synthetic datasets, with optional fine-tuning on experimental data.

Main Results:

• Models trained with Template Learning outperformed those trained solely on annotations.
• Template Learning demonstrated higher precision and more uniform orientation detection than traditional template matching.
• The method is particularly effective for small, non-spherical particles.

Conclusions:

• Template Learning offers an efficient and accurate solution for particle picking in cryo-ET.
• The automated synthetic data generation significantly reduces the labor associated with training deep learning models.
• The open-source, parallelized software facilitates broader adoption in structural biology.