CryoPromptSeg: prompt-guided segmentation with integrated denoising for cryo-EM particle picking.

Area of Science:

• Structural biology
• Biophysics
• Computational biology

Background:

• Cryo-electron microscopy single particle analysis (Cryo-EM SPA) is crucial for determining biomacromolecular structures.
• Automated particle picking is essential for high-resolution 3D reconstruction but faces challenges with noise and efficiency.
• Existing methods, including the Segment Anything Model (SAM), have limitations in fully addressing Cryo-EM specific noise and semantic information loss.

Purpose of the Study:

• To develop an advanced automated particle picking method for Cryo-EM SPA.
• To improve the accuracy and efficiency of particle identification in noisy Cryo-EM images.
• To enhance the preservation of particle structural information during denoising.

Main Methods:

• CryoPromptSeg utilizes prompt-guided SAM with domain adaptation for precise cryo-EM particle segmentation.
• A parallel multi-task framework jointly trains a semantically enhanced image denoiser and an automatic prompt generator.
• Particle semantic information from the prompt generator is integrated into the denoiser to improve noise suppression while preserving particle features.

Main Results:

• CryoPromptSeg demonstrates superior performance in particle picking accuracy compared to existing methods.
• The integrated denoising component achieves higher image quality by preserving distinguishable particle structures.
• A user-friendly online prediction platform was developed to facilitate practical application.

Conclusions:

• CryoPromptSeg offers a novel and effective solution for the automation of particle picking in Cryo-EM SPA.
• The method successfully addresses challenges related to image noise and semantic information in Cryo-EM data.
• This advancement has the potential to accelerate high-resolution structure determination of biomacromolecules.