Property guidance for protein sequence generative models with ProteinGuide

Area of Science:

• Computational biology
• Protein engineering
• Machine learning

Background:

• Existing protein generative models lack a framework for conditioning on auxiliary data without retraining.
• This limits the adaptability and application of these models in protein engineering.

Purpose of the Study:

• Introduce ProteinGuide, a novel method for 'on-the-fly' conditioning of protein generative models.
• Provide a unifying statistical framework for diverse protein generative model classes.
• Demonstrate the utility of ProteinGuide in designing proteins with specific properties and enhancing biological functions.

Main Methods:

• ProteinGuide utilizes a unifying statistical framework applicable to various generative models (e.g., ESM3, ProteinMPNN, MultiFlow).
• The method allows for 'on-the-fly' conditioning on auxiliary information, such as experimental data.
• Integration with wet-lab experiments for iterative design and validation.

Main Results:

• Successfully designed proteins with user-specified properties like enhanced stability and activity.
• Optimized proteins for two conflicting desired properties simultaneously.
• Developed an adenine base editor with significantly higher in vivo editing efficiency compared to directed evolution.

Conclusions:

• ProteinGuide offers a flexible and efficient framework for conditioning protein generative models.
• The method accelerates protein design and optimization for various applications.
• ProteinGuide demonstrates superior performance in improving protein function, exemplified by the enhanced base editor.