Context-aware geometric deep learning for protein sequence design.

Area of Science:

• Computational biology
• Protein engineering
• Deep learning applications

Background:

• Protein design and engineering are rapidly advancing with deep learning.
• Existing models struggle to incorporate non-protein molecules into the design process.
• There is a need for protein design tools that account for diverse molecular contexts.

Purpose of the Study:

• To introduce a novel deep learning method for protein sequence prediction.
• To enable protein design that accounts for restraints from non-protein molecular environments.
• To enhance the versatility of protein engineering pipelines.

Main Methods:

• A deep learning approach utilizing a geometric transformer.
• Input based solely on atomic coordinates and element names.
• Prediction of protein sequences from backbone scaffolds within specific molecular environments.

Main Results:

• The method successfully predicts protein sequences from backbone scaffolds.
• Generated enzymes exhibited high thermostability and catalytic activity.
• High success rates were achieved in producing functional enzymes.

Conclusions:

• The developed deep learning approach effectively designs protein sequences.
• The method's ability to consider molecular environments broadens protein design capabilities.
• This approach is expected to significantly improve the creation of proteins with desired functions.