Tuning ProteinMPNN to reduce protein visibility via MHC Class I through direct preference optimization

ProteinMPNN is widely used in protein design workflows due to its ability to identify amino acid sequences that fold into specific 3D protein structures. In our work, we adjust ProteinMPNN to design proteins for a given 3D protein structure with reduced immune-visibility to cytotoxic T lymphocytes that recognize proteins via the MHC-I pathway. To achieve this, we developed a novel framework that integrates direct preference optimization (DPO)-a tuning method originally designed for large language models-with MHC-I peptide presentation predictions. This approach fosters the generation of designs with fewer MHC-I epitopes while preserving the protein's original structure. Our results demonstrate that DPO effectively reduces MHC-I visibility without compromising the structural integrity of the proteins.