Positive-unlabeled convolutional neural networks for particle picking in cryo-electron micrographs

Area of Science:

• Structural Biology
• Biophysics
• Computational Biology

Background:

• Cryo-electron microscopy (cryo-EM) is crucial for protein structure determination.
• Manual particle picking in cryo-EM is time-consuming and labor-intensive.
• Existing computational methods often yield false positives and require manual postprocessing, especially for difficult particle shapes.

Purpose of the Study:

• To develop an efficient and accurate automated particle-picking pipeline for cryo-EM.
• To overcome limitations of manual picking and current computational approaches.
• To improve the identification of both conventional and unusually shaped protein particles.

Main Methods:

• Development of Topaz, a particle-picking pipeline utilizing neural networks.
• Training models with a positive-unlabeled learning method, requiring few labeled particles and no labeled negatives.
• Application of the pipeline to diverse cryo-EM datasets, including challenging cases.

Main Results:

• Topaz significantly outperforms conventional methods in retrieving true particles.
• Achieved low false-positive rates, even for small, non-globular, and asymmetric particles.
• Generated more representative particle sets without the need for post hoc curation.

Conclusions:

• Topaz offers an efficient and accurate solution for particle picking in cryo-EM.
• The positive-unlabeled learning framework enables robust model training.
• Topaz is a free, open-source tool that streamlines cryo-EM data analysis.