Self-supervised learning for generalizable particle picking in cryo-EM micrographs.

Area of Science:

• Structural Biology
• Biophysics
• Computational Biology

Background:

• Manual annotation of cryo-electron microscopy (cryo-EM) data is labor-intensive and time-consuming.
• Developing automated methods for particle picking is crucial for efficient cryo-EM data analysis.
• Existing methods often require large annotated datasets for training supervised models.

Purpose of the Study:

• To introduce cryo-electron microscopy masked autoencoder (cryo-EMMAE), a novel self-supervised learning method for cryo-EM data analysis.
• To demonstrate the effectiveness of cryo-EMMAE in particle picking and 3D reconstruction.
• To reduce the reliance on manual annotation in cryo-EM workflows.

Main Methods:

• Implemented a masked autoencoder (MAE) architecture for self-supervised learning on cryo-EM images.
• Utilized the latent representation space of the MAE for clustering particle pixels.
• Evaluated cryo-EMMAE performance on diverse EMPIAR datasets for particle picking and 3D reconstruction.

Main Results:

• cryo-EMMAE achieved superior generalization capabilities compared to state-of-the-art supervised methods.
• The method demonstrated consistent performance across different training datasets and is data-efficient, converging with as few as five micrographs.
• Superior 3D reconstruction performance was observed for both single-particle datasets and cell extracts.

Conclusions:

• Self-supervised learning, exemplified by cryo-EMMAE, offers a powerful alternative for cryo-EM image analysis.
• cryo-EMMAE enhances efficiency and reduces costs in structural biology research.
• The method shows significant potential for advancing automated cryo-EM data processing.