Protein sequence design with a learned potential

Area of Science:

• Computational biology
• Protein engineering
• Artificial intelligence in science

Background:

• Protein sequence design is crucial for protein engineering.
• Current methods rely on energy functions, requiring significant human input.
• Automating this process can accelerate protein design.

Purpose of the Study:

• To investigate a deep neural network's capability for automating protein sequence design.
• To assess the model's ability to generalize to novel protein topologies.
• To demonstrate an entirely learned method for protein sequence design.

Main Methods:

• Training a deep neural network model directly on crystal structure data.
• Using the trained model to predict sequences for given protein backbones.
• Evaluating designs using experimental stability and high-resolution crystal structures.

Main Results:

• The deep neural network model successfully designed novel protein sequences.
• The model generalized to native protein topologies not seen during training.
• Experimentally stable designs were produced, with crystal structures matching in silico predictions.

Conclusions:

• A deep neural network can effectively automate protein sequence design.
• The learned method shows strong generalizability to new protein scaffolds.
• This approach offers a tractable, data-driven alternative to traditional methods.