Antibodies really are hard to model computationally, even non-ai approaches like Rosetta require antibody-specific methods. CDR3 loops don’t have a lot of secondary structure, they don’t evolve in the same way that regular proteins do, and they often assume a different conformation on binding