DeepEMhancer: a deep learning solution for cryo-EM volume post-processing.

Area of Science:

• Structural Biology
• Biophysics
• Computational Biology

Background:

• Cryo-electron microscopy (cryo-EM) maps are crucial for protein structure determination.
• High-frequency data loss in cryo-EM maps necessitates post-processing for improved interpretability.
• Current post-processing methods, like global B-factor correction, fail to address local variations in map quality.

Purpose of the Study:

• To develop an automated deep learning approach for cryo-EM map post-processing.
• To enhance the interpretability and detail of cryo-EM maps.
• To overcome limitations of existing global B-factor correction methods.

Main Methods:

• Developed DeepEMhancer, a deep learning model for cryo-EM map enhancement.
• Trained the model on pairs of experimental cryo-EM maps and their sharpened counterparts derived from atomic models.
• Implemented masking-like and sharpening-like operations within a single deep learning framework.

Main Results:

• DeepEMhancer successfully reduced noise levels in cryo-EM maps.
• The method generated more detailed versions of experimental cryo-EM maps.
• Evaluated on 20 diverse experimental maps, demonstrating consistent performance.
• Showcased improved visualization of the SARS-CoV-2 RNA polymerase structure.

Conclusions:

• DeepEMhancer offers an effective, automated solution for cryo-EM map post-processing.
• The deep learning approach enhances structural detail and interpretability beyond traditional methods.
• This tool has significant implications for advancing protein structure modeling using cryo-EM data.