CryoSIP: unleashing protein high-resolution Cryo-EM via semantic-instance collaborative picking.

Area of Science:

• Structural Biology
• Biophysics
• Microscopy

Background:

• Accurate particle picking is crucial for high-resolution 3D reconstruction in cryo-electron microscopy (cryo-EM).
• Low signal-to-noise ratios (SNRs) and weak contrast in cryo-EM images present challenges, leading to low detection rates and high false positives.
• Existing methods struggle to balance sensitivity and specificity in particle identification.

Purpose of the Study:

• To develop an advanced particle picking framework for cryo-EM single-particle analysis.
• To address limitations of current methods, specifically low detection rates and high false positives in low SNR conditions.
• To improve the quality and resolution of 3D cryo-EM reconstructions.

Main Methods:

• Developed a novel semantic-instance collaborative picking framework integrating a multi-frequency adaptive U-Net and Segment Anything Model (SAM).
• Utilized global-context semantic modeling and multi-scale feature fusion within the U-Net for precise particle localization.
• Enhanced collaborative optimization between SAM and U-Net to refine semantic priors and improve instance mask generation.

Main Results:

• The proposed framework significantly reduced false positives compared to state-of-the-art tools.
• Achieved high recall rates in particle picking, demonstrating improved sensitivity.
• 3D reconstructions generated using this method showed enhanced density map resolution.

Conclusions:

• The semantic-instance collaborative picking framework offers a robust solution for precise particle identification in cryo-EM.
• This approach effectively mitigates challenges posed by low SNR and weak contrast.
• The method has the potential to advance atomic-resolution structural analysis using cryo-EM.