Knowledge Transfer from Macro-world to Micro-world: Enhancing 3D Cryo-ET Classification through Fine-Tuning...

Area of Science:

• Computational Biology
• Structural Biology
• Machine Learning

Background:

• Deep learning models excel in natural-world tasks due to large datasets and transferable pre-trained models.
• Applying natural-domain models to Cryo-Electron Tomography (Cryo-ET) remains underexplored.
• 3D Cryo-ET data can be conceptualized as evolving video frames.

Purpose of the Study:

• To enhance Cryo-ET subtomogram classification using 3D models pre-trained on large-scale video datasets.
• To investigate the transferability of video-based models to Cryo-ET data.
• To reduce training costs and improve feature extraction in Cryo-ET analysis.

Main Methods:

• Utilized 3D models pre-trained on large-scale video datasets for Cryo-ET subtomogram classification.
• Conducted experiments on both simulated and real Cryo-ET datasets.
• Implemented the approach using the aitom repository (https://github.com/xulabs/aitom).

Main Results:

• Video initialization significantly improved Cryo-ET classification accuracy.
• Substantially reduced training costs compared to traditional methods.
• Enhanced subtomogram feature extraction capabilities.
• Observed positive effects in medical 3D classification tasks.

Conclusions:

• Cross-domain knowledge transfer from video models is effective for Cryo-ET.
• Video initialization offers a promising approach for advancing biological and medical 3D data analysis.
• This method improves efficiency and accuracy in Cryo-ET classification.