Deep Learning to Predict Protein Backbone Structure from High-Resolution Cryo-EM Density Maps.

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Cryo-electron microscopy (cryo-EM) enables high-resolution protein structure determination.
• Predicting protein backbone traces from cryo-EM density maps is challenging at resolutions worse than 2.5 Å.
• Existing methods struggle with incomplete or noisy density maps.

Purpose of the Study:

• To develop a deep learning model for accurate prediction of protein backbone traces and Cα atoms from cryo-EM data.
• To improve the automation and accuracy of protein structure modeling from experimental cryo-EM maps.
• To provide a robust method for analyzing cryo-EM density maps across a range of resolutions.

Main Methods:

• A novel cascaded convolutional neural network (C-CNN) architecture was developed for semantic segmentation of cryo-EM density maps.
• The C-CNN model predicts secondary structure elements, backbone structure, and Cα atom positions.
• Specialized algorithms for path walking, helix refinement, and quality assessment-based sequence mapping were integrated.

Main Results:

• The C-CNN model accurately predicted a mean of 88.9% of Cα atoms within 3 Å of the protein backbone.
• This significantly outperformed state-of-the-art methods, achieving higher Cα atom prediction completeness.
• The model demonstrated a low average root-mean-square deviation (RMSD) of 1.24 Å on experimental density maps.

Conclusions:

• The cascaded-CNN deep learning model offers a significant advancement in automated protein structure prediction from cryo-EM data.
• This method enhances the ability to determine protein structures from lower-resolution and less pristine experimental maps.
• The developed approach improves the efficiency and accuracy of structural biology research using cryo-EM.