My general idea was to generate a large number of designs then score them using an MSA-augmented PLM like MSA Pairformer or Profluent E1. My hypothesis was using the evolutionary context existing natural binders for the NiV-G protein could help judge the designed binders better. Potentially, in a longer time horizon project, these models could be finetuned on specific protein-binder complex grading. Given they have latent contact point prediction abilities, this could work well. I think my idea is interesting but I was too short on time to really give it a good try.