Model Compression for Faster Structural Separation of Macromolecules Captured by Cellular Electron Cryo-Tomography

Area of Science:

• Structural biology
• Computational biology
• Biophysics

Background:

• Electron cryo-tomography (ECT) provides 3D visualization of macromolecular structures within cells.
• Convolutional neural networks (CNNs) are used for classifying millions of macromolecules from ECT data.
• Increasing ECT data volume necessitates faster and more efficient macromolecule classification.

Purpose of the Study:

• To develop a method for compressing CNN models used in ECT data analysis.
• To reduce the computational cost and prediction time for macromolecule classification.
• To maintain high classification accuracy after model compression.

Main Methods:

• Model compression using knowledge distillation.
• Training a complex 'teacher' network to generate soft labels.
• Training simplified 'student' networks using soft labels from the teacher network.

Main Results:

• Compressed models significantly reduced the number of parameters.
• Substantial reduction in time cost for macromolecule classification.
• Maintained classification accuracy comparable to the original complex models.

Conclusions:

• Knowledge distillation is an effective method for compressing CNN models for ECT data.
• Compressed models enable efficient and accurate classification of large-scale macromolecule datasets.
• This approach addresses the computational challenges posed by the growing volume of ECT data.