Tpgen: a language model for stable protein design with a specific topology structure

Area of Science:

• Protein engineering
• Computational biology
• Artificial intelligence in protein design

Background:

• Natural proteins represent a limited fraction of possible protein sequences.
• Designing artificial proteins offers broader possibilities but requires specific methods.
• Small proteins exhibit advantageous properties like stability and ease of modification.

Purpose of the Study:

• To develop an AI model for generating stable small protein sequences with desired topologies.
• To leverage advanced machine learning for exploring the protein sequence space.

Main Methods:

• A neural network model, TopoProGenerator (TPGen), was developed using a transformer decoder architecture.
• TPGen was trained on sequences up to 65 amino acids, incorporating reinforcement and adversarial learning for fine-tuning.
• A stability predictive model was trained on over 200,000 protein sequences.

Main Results:

• TopoProGenerator successfully designed stable small protein sequences.
• The generated sequences folded into specified topology structures.
• The model demonstrates capability in targeted protein design.

Conclusions:

• TopoProGenerator (TPGen) can generate protein sequences with predictable folding into desired topologies.
• The proposed pretraining and fine-tuning methodologies offer a versatile framework for diverse protein design tasks.