Designing proteins with reduced T-cell epitopes through policy optimization.

Area of Science:

• Computational biology
• Immunoinformatics
• Protein engineering

Background:

• Deep generative models are advancing protein design for therapeutics.
• Clinical success hinges on immune system compatibility, involving proteasomal cleavage, peptide elution, and Major Histocompatibility Complex (MHC) Class I binding.
• Previous models often analyzed these immune response steps in isolation or for limited MHC alleles.

Purpose of the Study:

• To develop integrated computational predictors for MHC Class I pathway interactions.
• To incorporate uncertainty estimation using evidential deep learning.
• To generate protein candidates optimized for immune compatibility across diverse MHC alleles.

Main Methods:

• Fine-tuning a protein language model on human proteins.
• Utilizing group relative policy optimization (GRPO) to reduce MHC Class I epitopes.
• Implementing a curriculum learning framework that progressively increases complexity (masked epitopes and MHC alleles).

Main Results:

• Development of unified uncertainty estimates for cleavage, elution, and binding predictions.
• Successful alignment of a protein language model to minimize MHC Class I epitopes.
• Demonstration of a strategy for generating immune-compatible protein candidates.

Conclusions:

• The developed predictors offer a unified approach to modeling immune response in the MHC Class I pathway.
• The GRPO and curriculum learning strategy effectively reduces immunogenic epitopes across diverse MHC alleles.
• This work advances the design of safer and more effective protein therapeutics by accounting for predictive uncertainty.