prismPYP: Power-spectrum and image domain learning for self-supervised micrograph evaluation.

Area of Science:

• Structural Biology
• Biophysics
• Computational Biology

Background:

• High-throughput cryo-electron microscopy (cryo-EM) demands efficient micrograph quality assessment.
• Manual curation is time-consuming and prone to errors with large datasets.
• Existing deep learning methods require extensive annotated data, limiting generalizability.

Purpose of the Study:

• To develop a scalable, generalizable, and automated method for cryo-EM micrograph quality assessment.
• To overcome limitations of manual curation and supervised deep learning approaches.
• To enable label-free feature learning for improved image quality classification.

Main Methods:

• Developed prismPYP, a self-supervised, data-driven framework utilizing domain-specific image augmentations.
• Employed label-free feature learning on micrographs and their power spectra.
• Utilized learned low-dimensional representations for feature-based image clustering.

Main Results:

• Identified distinct and consistent indicators of micrograph quality through clustering.
• Achieved high-resolution structure determination using fewer particles from prismPYP-selected images compared to manual curation.
• Demonstrated generalization across diverse experimental conditions, hardware, and cryo-EM techniques (single-particle and time-resolved).

Conclusions:

• prismPYP provides a computationally efficient and interpretable solution for rapid cryo-EM micrograph quality assessment.
• The framework enhances scalability and generalizability in structural biology.
• Enables faster and more reliable high-resolution structure determination from cryo-EM data.