CASSPER is a semantic segmentation-based particle picking algorithm for single-particle cryo-electron microscopy.

Area of Science:

• Structural biology
• Biophysics
• Cryo-electron microscopy

Background:

• Automating particle identification is crucial for high-resolution structure determination in single-particle cryo-electron microscopy (cryo-EM).
• Manual particle picking is a significant bottleneck in cryo-EM data processing pipelines.
• Developing automated methods is essential for increasing throughput and efficiency in structural biology.

Purpose of the Study:

• To introduce CASSPER, a generalized deep learning tool for automated detection and isolation of protein particles in transmission microscope images.
• To overcome the limitations of manual particle picking in cryo-EM.
• To enable on-the-fly particle picking for complete data processing automation.

Main Methods:

• CASSPER utilizes semantic segmentation for pixel-level classification of particles, ice, and impurities.
• The tool is trained on visually curated datasets to distinguish between different components in micrographs.
• Contrast Limited Adaptive Histogram Equalization (CLAHE) is integrated to improve particle detection under varying ice conditions.

Main Results:

• CASSPER effectively performs automated detection and isolation of protein particles, eliminating the need for manual intervention.
• The deep learning model demonstrates high-fidelity particle detection even with variable ice thickness and contrast.
• The generalized CASSPER model shows high efficiency on previously unseen datasets.

Conclusions:

• CASSPER significantly advances automated structure determination in cryo-EM by providing a robust particle picking solution.
• The tool's ability to work on unseen data and potential for on-the-fly processing paves the way for fully automated cryo-EM workflows.
• CASSPER offers a generalized approach applicable to various cryo-EM datasets, accelerating biological macromolecule structure analysis.