CrysFormer: Protein structure determination via Patterson maps, deep learning, and partial structure attention.

Area of Science:

• Structural biology
• Computational chemistry
• Biophysics

Background:

• Determining atomic-level protein structure is crucial but challenging.
• Existing methods primarily use sequence data and known templates.
• Prior knowledge from X-ray crystallography and residue conformations is underutilized.

Purpose of the Study:

• To develop the first transformer-based model for protein structure calculation using experimental crystallographic data.
• To directly compute electron density maps, bypassing the crystallographic phase problem.

Main Methods:

• Proposed CrysFormer, a transformer model integrating experimental crystallographic data (Patterson maps) and partial structure information.
• Utilized Patterson maps derived directly from X-ray crystallography data.
• Trained and tested on synthetic datasets of peptide fragments in crystalline forms.

Main Results:

• CrysFormer accurately predicts electron density maps from crystallographic data.
• Achieved precise predictions on datasets with varying complexity (2 and 15 residues per unit cell).
• Generated accurate atomic models using established crystallographic refinement programs.

Conclusions:

• CrysFormer represents a significant advancement in protein structure determination.
• The model effectively leverages experimental data to overcome limitations of sequence-only approaches.
• Enables more accurate and efficient generation of atomic models for proteins.