All-Atom Protein Sequence Design Based on Geometric Deep Learning.

Area of Science:

• Protein engineering
• Computational biology
• Structural biology

Background:

• Designing novel proteins with specific functions requires precise sequence generation for given backbones.
• Accurate prediction of side chain conformations is crucial for determining the final 3D structure and function.

Purpose of the Study:

• Introduce GeoSeqBuilder, a deep learning framework for de novo protein sequence design.
• Integrate sequence generation with side chain conformation prediction for complete all-atom structure generation.
• Validate GeoSeqBuilder's performance against leading methods and experimental data.

Main Methods:

• GeoSeqBuilder utilizes spatial geometric features from protein backbones.
• Incorporates three-body interactions of neighboring residues in its deep learning model.
• Predicts both sequence and side chain conformations for designed proteins.

Main Results:

• Achieved a native residue type recovery rate of 51.6%, comparable to state-of-the-art methods.
• Designed sequences for thioredoxin and a three-helical bundle resulted in soluble, stable proteins with solved crystal structures matching designs.
• Redesigned glutathione peroxidase 4 showed improved enzyme activity in 3 out of 5 designs.

Conclusions:

• GeoSeqBuilder is a powerful tool for designing novel protein sequences for predefined structures with atomic detail.
• Demonstrates a high experimental success rate, facilitating the creation of new functional proteins.
• The framework offers low sequence similarity to originals and significantly altered hydrophobic cores, enabling novel protein designs.