Deep Learning for Validating and Estimating Resolution of Cryo-Electron Microscopy Density Maps ^†.

Area of Science:

• Structural Biology
• Biophysics
• Computational Biology

Background:

• Cryo-electron microscopy (cryo-EM) is a leading technique for determining atomic protein structures from 3D density maps.
• Accurate resolution assessment of cryo-EM maps is crucial but challenging due to limitations of the Fourier Shell Correlation (FSC) method.
• Current FSC resolution measures can be subjective and prone to manipulation, leading to scientific debate.

Purpose of the Study:

• To develop and evaluate supervised deep learning methods for objective resolution validation of 3D cryo-EM density maps.
• To compare the performance of different deep learning architectures, including DNN, 3D CNN, and U-Net, for resolution classification.
• To assess the potential of deep learning to improve the reliability of cryo-EM resolution evaluation.

Main Methods:

• Trained deep artificial neural network (DNN) and 3D convolutional neural network (3D CNN) models on simulated cryo-EM density maps for global resolution classification (high, medium, low).
• Applied trained DNN and 3D CNN models to experimental cryo-EM maps to compare their classifications with author-published resolution values.
• Developed and evaluated a 3D U-Net model for voxel-wise local resolution classification into ten classes on experimental maps.

Main Results:

• Preliminary DNN and 3D CNN models achieved high accuracy (92.73% and 99.75%) on simulated data.
• DNN and 3D CNN models showed moderate agreement (60.0% and 56.7%) with published resolutions on experimental maps.
• The 3D U-Net model achieved high accuracy (88.3% and 94.7%) for local resolution classification on experimental maps.

Conclusions:

• Supervised deep learning methods show promise for objectively validating cryo-EM map resolutions.
• Deep learning models can provide a more reliable and less subjective alternative to traditional FSC methods.
• Further development of deep learning approaches, particularly for local resolution analysis, can enhance the accuracy of cryo-EM structural assessments.