从读取结构到编写结构：AIntibody挑战赛与AI抗体设计

AlphaFold2 settled a question the field had argued over for decades: Could a protein’s structure be predicted from its sequence? In the years since, structure prediction has become basic infrastructure for biology. A harder question remains: Can that same structural reasoning carry over from reading a structure to writing one? Can a model not only predict what an antibody looks like, but actually design a better one? AIntibody, run in 2025 and now reported in Nature Biotechnology (doi: 10.1038/s41587-026-03238-6), put that leap from prediction to design to the test. Participating teams tackled a series of antibody-design challenges computationally. In all, 29 organizations submitted 511 candidate antibody sequences targeting the receptor-binding domain (RBD) of the SARS-CoV-2 spike protein. …