Finetuning ESM3 with Contrastive Preference Optimization for Antigen-Specific Antibody Design

Our data is derived from this study. Using contrastive preference optimization, a variant of direct preference optimization, we can finetune ESM3 to more effectively redesign the CDR3 region of Trastuzumab bound to HER2. We find that sequences generated by the finetuned model greatly surpass those of existing protein language foundation models in terms of plausibility and edit distances from ground truth high affinity sequences. Furthermore, after folding and docking the generated antibody structures, we find that our finetuned model generates unique sequences with binding affinities comparable to those of the high affinity sequences from the training dataset.