A comprehensive survey and benchmark of deep learning-based methods for atomic model building from cryo-electron...

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Single-particle cryo-electron microscopy (cryo-EM) generates 3D density maps for atomic model construction.
• Automated methods are crucial for efficiently building these atomic models.
• Deep learning (DL) has emerged as a powerful tool for this task.

Purpose of the Study:

• To comprehensively survey and evaluate deep learning methods for automated protein model building from cryo-EM data.
• To compare DL-based approaches against traditional physics-based methods.
• To assess the impact of integrating AlphaFold sequence-to-structure predictions.

Main Methods:

• Categorized DL methods into direct (density map only) and indirect (integrating AlphaFold) approaches.
• Refined existing metrics for precise evaluation.
• Benchmarked representative DL methods against physics-based methods using 50 cryo-EM density maps across various resolutions.

Main Results:

• Deep learning methods generally outperform traditional physics-based approaches for protein model building.
• Integrating AlphaFold predictions significantly improved model completeness and accuracy.
• The effectiveness of AlphaFold integration is dependent on sequence information availability and training data.

Conclusions:

• Deep learning offers superior performance in automated protein model building from cryo-EM data.
• AlphaFold integration represents a significant advancement, enhancing model quality.
• Future research should address limitations related to data dependency for broader applicability.