Improved AlphaFold modeling with implicit experimental information.

Area of Science:

• Structural biology
• Computational biology
• Biophysics

Background:

• Machine learning (ML) models like AlphaFold and RoseTTAFold predict protein structures with high accuracy.
• However, these ML models often yield low-confidence predictions in certain protein regions.
• Experimental data, such as density maps, can provide crucial information for refining these models.

Purpose of the Study:

• To enhance the accuracy of protein structure prediction by integrating experimental data with ML models.
• To investigate the synergistic effects of combining ML predictions with experimental information.
• To develop an automated procedure for interpreting experimental maps using ML.

Main Methods:

• An iterative computational procedure was developed to rebuild AlphaFold models using experimental density maps.
• Rebuilt models were used as templates for subsequent AlphaFold predictions.
• The method was integrated into an automated pipeline for map interpretation.

Main Results:

• Integrating experimental density maps significantly improved the accuracy of AlphaFold protein models.
• The combined approach yielded better results than simple rebuilding guided solely by experimental data.
• The developed procedure effectively aids in the interpretation of crystallographic and cryo-EM maps.

Conclusions:

• Combining ML-based protein modeling with experimental data offers a synergistic advantage for improving structural accuracy.
• This integrated approach enhances the reliability of protein models, particularly in regions of low confidence.
• The automated procedure facilitates the interpretation of experimental structural data, advancing structural biology.