PB-GPT: An innovative GPT-based model for protein backbone generation.

Area of Science:

• Computational biology
• Protein engineering
• Artificial intelligence in medicine

Background:

• Protein structure and function are intrinsically linked to their backbones.
• Designing protein backbones is crucial for engineering proteins with specific functions.
• Advanced computational methods enable protein modification for disease treatment and other applications.

Purpose of the Study:

• To develop a method for unconditionally generating protein backbones.
• To explore the potential of large language models in protein structure design.

Main Methods:

• Converted protein backbone structures into a coded language using codebook quantization and compression dictionaries.
• Proposed and developed a GPT-based protein backbone generation model named PB-GPT.
• Trained and evaluated PB-GPT on public and small protein datasets to assess generalization performance.

Main Results:

• PB-GPT demonstrated the capability to unconditionally generate elaborate and highly realistic protein backbones.
• The generated protein backbones exhibited structural patterns similar to those found in natural proteins.
• The model showed strong generalization performance across different datasets.

Conclusions:

• Large language models, like PB-GPT, hold significant potential for advancing protein structure design.
• The developed method offers a novel approach to computational protein engineering.
• This work paves the way for designing novel proteins with tailored functions for various applications.