Beyond current boundaries: Integrating deep learning and AlphaFold for enhanced protein structure prediction from...

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Atomic model construction from cryo-electron microscopy (cryo-EM) maps is essential but challenging.
• Current deep learning tools struggle with low-resolution cryo-EM maps (beyond 4 Å).

Purpose of the Study:

• To develop an advanced computational framework, DeepTracer-LowResEnhance, for improving atomic model accuracy from low-resolution cryo-EM maps.
• To enhance map interpretability and modeling precision in structural biology.

Main Methods:

• Integration of AlphaFold structural predictions with a deep-learning-based map refinement strategy.
• Tailored neural network-driven refinement process for low-resolution cryo-EM data.
• Testing on a diverse dataset of 37 protein cryo-EM maps (2.5–8.4 Å resolution).

Main Results:

• DeepTracer-LowResEnhance achieved a 3.53x average TM-score improvement over baseline DeepTracer.
• Significant performance gains observed in 95.5% of tested low-resolution datasets.
• Demonstrated superior capability in generating detailed atomic models compared to traditional sharpening methods.

Conclusions:

• DeepTracer-LowResEnhance effectively enhances low-resolution cryo-EM map interpretability and atomic model accuracy.
• The framework pushes the boundaries of computational structural biology for challenging datasets.
• This tool offers a significant advancement for researchers working with low-resolution cryo-EM data.