Completion of partial structures using Patterson maps with the CrysFormer machine-learning model.

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Protein structure determination is crucial in structural biology.
• Deep machine learning (ML) models are increasingly used but often omit experimental data.
• Integrating experimental measurements with ML is a key challenge.

Purpose of the Study:

• To develop a hybrid approach combining X-ray crystallography and ML for protein structure determination.
• To improve electron density map prediction and atomic model refinement.
• To address limitations of ML models that do not incorporate experimental diffraction data.

Main Methods:

• Training a hybrid 3D vision transformer and convolutional network.
• Utilizing Patterson maps from crystallographic data and partial structure templates from AlphaFold.
• Predicting electron-density maps and post-processing into atomic models via crystallographic refinement.

Main Results:

• Demonstrated effectiveness on small protein fragments.
• Successfully improved crystallographic structure factor phases.
• Enhanced completion of missing regions in partial structure templates.
• Improved agreement between predicted electron-density maps and ground-truth atomic structures.

Conclusions:

• The hybrid ML-crystallography method offers a powerful new paradigm for protein structure determination.
• This approach effectively integrates experimental data with advanced computational techniques.
• The method shows promise for refining and completing atomic models in structural biology.