A Transfer Learning-Based Classification Model for Particle Pruning in Cryo-Electron Microscopy.

Area of Science:

• Structural Biology
• Biophysics
• Computational Biology

Background:

• Cryo-electron microscopy (cryo-EM) single-particle analysis demands numerous particle projections for macromolecular complex structural determination.
• Automated particle picking in cryo-EM is often hindered by low signal-to-noise ratios and artifacts, leading to high false-positive rates and reduced structural confidence.

Purpose of the Study:

• To introduce PickerOptimizer (PO), a novel transfer learning-based neural network designed to refine particle selection in cryo-EM.
• To enhance the accuracy and reliability of automated particle picking algorithms by reducing false positives.

Main Methods:

• Developed PO, a convolutional neural network leveraging transfer learning from public classification datasets for cryo-EM applications.
• Implemented a multiloss strategy to optimize network parameters effectively.
• Created the first cryo-EM image classification dataset using EMPIAR entries to mitigate domain shift during pretraining.

Main Results:

• Achieved accuracy and F1 scores exceeding 95% on 14 public cryo-EM datasets.
• Demonstrated superior or comparable performance to existing particle pruning strategies in three case studies involving challenging particle selections.
• Validated PO's effectiveness in improving structural determination confidence through rigorous testing.

Conclusions:

• PickerOptimizer (PO) offers a robust solution for particle pruning in cryo-EM, significantly improving the accuracy of automated particle picking.
• The integration of transfer learning and a custom cryo-EM dataset enhances PO's performance, addressing key challenges in cryo-EM data processing.
• PO serves as a valuable complementary tool for researchers aiming to achieve higher confidence in cryo-EM structure determination.