Improved protein structure refinement guided by deep learning based accuracy estimation.

Area of Science:

• Computational Biology
• Structural Biology
• Artificial Intelligence in Biochemistry

Background:

• Protein structure modeling is crucial for understanding biological function.
• Assessing and improving the accuracy of protein models is a significant challenge.
• Existing methods for predicting protein model accuracy have limitations.

Purpose of the Study:

• To develop a deep learning framework (DeepAccNet) for accurate per-residue accuracy estimation in protein models.
• To utilize these accuracy predictions to guide protein structure refinement using the Rosetta protocol.
• To provide a tool for assessing the quality of both predicted and experimentally determined protein structures.

Main Methods:

• Development of DeepAccNet, a deep learning network employing 3D and 2D convolutions to analyze local atomic environments and global contexts.
• Application of DeepAccNet to estimate per-residue accuracy and residue-residue distance signed error in protein models.
• Integration of DeepAccNet's accuracy predictions into the Rosetta protein structure refinement protocol.

Main Results:

• DeepAccNet outperforms existing methods in predicting the accuracy of protein structure models.
• Accuracy predictions correlate with experimental resolution for X-ray and cryo-EM structures.
• Incorporating DeepAccNet predictions into Rosetta refinement significantly improved the accuracy of resulting protein models.

Conclusions:

• DeepAccNet is a powerful tool for assessing protein model accuracy and identifying regions prone to error.
• The framework is applicable to both computationally predicted and experimentally determined protein structures.
• Deep learning-based accuracy prediction can enhance protein structure refinement by improving the search for optimal molecular conformations.