EvoAntigen: An Evolving Agent for Antigen Design

Computational antigen design requires efficient search over large sequence spaces under limited experimental budgets. Protein language models (PLMs) provide useful sequence priors, but their preferences can be misaligned with experimentally measured antigen phenotypes. We introduce EvoAntigen, an LLM-powered agentic workflow that adapts candidate-generation strategies through iterative feedback from retrospective experimental measurements. EvoAntigen maintains complementary exploration and exploitation generators as executable programs and revises them using previously evaluated candidates. Candidate generation is further guided by predicted epitope regions together with auxiliary sequence- and structure-based assessments. We evaluate EvoAntigen retrospectively on 30 viral deep mutational scanning (DMS) tasks spanning binding, expression, cell entry, viral growth, and stability. Under matched candidate budgets, EvoAntigen significantly improves the mean quality of generated candidates relative to a static PLM baseline, while improvements in Top-1, Top-5, and Top-10 performance are positive but not statistically significant after multiple-testing correction. These results indicate that the clearest benefit of adaptive policy revision is a more consistently high-quality candidate set rather than uniformly better recovery of the single best mutation. Case studies further illustrate that feedback can redirect search toward high-fitness candidates when the PLM prior is poorly aligned with the experimental landscape.