ProtSeqGen: a novel deep learning model for protein sequence design

Area of Science:

• Computational biology
• Protein engineering
• Artificial intelligence in drug discovery

Background:

• The protein inverse folding problem is crucial for de novo protein design.
• Current methods struggle to balance accuracy, stability, efficiency, and generalizability.

Purpose of the Study:

• To introduce ProtSeqGen, a deep learning model addressing limitations in protein inverse folding.
• To achieve accurate and generalizable sequence design for specified protein structures.

Main Methods:

• Developed a multi-stage, graph-based deep learning framework (ProtSeqGen).
• Encoded protein structures as local geometric graphs.
• Utilized message-passing neural networks and multi-layer perceptrons for residue interaction modeling and amino acid prediction.

Main Results:

• ProtSeqGen demonstrated superior sequence recovery over state-of-the-art methods on standard and challenging benchmarks.
• Successfully generated accurate, designable sequences for nine diverse protein structures.
• Exhibited remarkable generalization capabilities across different protein folds.

Conclusions:

• ProtSeqGen offers a robust and scalable solution for the protein inverse folding problem.
• The model significantly advances de novo protein design with high structural precision.
• Paves the way for designing novel proteins with tailored functions.