Discriminator-Guided Inverse Folding for Multi-Property Protein Design

Area of Science:

• Protein engineering
• Computational biology
• Biochemistry

Background:

• Designing proteins with multiple desired physicochemical properties is crucial for real-world applications.
• Structure-based de novo protein design is a leading paradigm, but joint multi-property optimization remains a challenge.
• Current methods struggle with multi-property optimization due to limited datasets with multiple property annotations.

Purpose of the Study:

• To develop a novel framework for simultaneous multi-property optimization in de novo protein design.
• To overcome the limitations of existing inverse folding methods that require multi-property annotated datasets.
• To enable the design of proteins with improved and simultaneously optimized properties like thermostability and solubility.

Main Methods:

• Introduced Discriminator-Guided Inverse Folding (DGIF), a framework guiding inverse folding models.
• Utilized an auxiliary discriminator module integrating multiple single-property predictors.
• Adjusted internal history states of the inverse folding model via the discriminator.

Main Results:

• DGIF achieved substantial improvements in protein thermostability and solubility.
• The framework successfully generated protein sequences optimized for multiple properties simultaneously.
• Designed proteins demonstrated a significant shift towards the Pareto front, indicating optimal trade-offs.

Conclusions:

• DGIF effectively enables multi-property optimization in structure-based protein design without multi-property datasets.
• The developed framework significantly enhances key protein traits and facilitates joint optimization.
• Experimental validation confirms DGIF's efficacy for designing proteins with tailored, multiple characteristics.