The trick this run is named for: AlphaFold2 re-predicts the binder-target complex, but its structure module is initialized from the designed backbone coordinates instead of from scratch. A real interface tends to relax back onto roughly the same geometry; a fake one tends to drift away from it, which shows up as high inter-chain PAE even though the model was handed the answer as a starting point — a much harder test than an unconditioned prediction Reference/Bennett 2023. Cut at pae_interaction < 10 Å (confidence in binder-target relative orientation) and plddt_binder > 0.80 (the binder has to fold with confidence on its own, not just sit passively next to the target). Applied to all 6,000 MPNN sequences; the best-scoring 24 are what this campaign actually ordered — everything else in the 6,000 stops here. Built on Reference/AlphaFold2 2021.
Stage G of PD-L1/Campaign/PDL1-R1.