Designing proteins with reduced T-cell epitopes through policy optimization

Deep generative models for protein structure and sequence are increasingly used to design proteins with therapeutic and industrial applications, but the clinical success of designed therapeutics ultimately depends on their compatibility with the human immune system. Immune response is triggered by a cascade of molecular events, including proteasomal cleavage, peptide elution, and binding to the highly polymorphic Major Histocompatibility Complex (MHC) Class I molecules, yet prior approaches have generally modeled these processes in isolation or restricted attention to a limited set of alleles. In this work, we develop predictors for cleavage, elution, and binding affinity in the MHC Class I pathway, incorporating evidential deep learning to provide unified uncertainty estimates. We first perform supervised fine-tuning of a protein language model on human proteins, and then align the model through group relative policy optimization (GRPO) to reduce MHC Class I epitopes under a curriculum learning framework, in which the curriculum progressively increases the number of masked predicted epitopes and the number of alleles considered during training. This strategy enables the generation of protein candidates that are optimized for immune compatibility across diverse MHC Class I alleles while accounting for predictive uncertainty in modeling the immune response.