I tried two approaches. For the first approach, I first ran RFdiffusion3 -> ProteinMPNN -> initial_guess pipeline to quickly screen multiple hotspots, then proceed with scaling up using the best performing one. Then I did partial diffusion on the best structures according to InitialGuess, then ran MPNN -> initial_guess, select the best structures again to further refine the sequences using MPNN. Lastly, I screen using both initial_guess and AlphaFold3, selecting the best sequences with the highest ipSAE scores from either of them (50 from AlphaFold3, 19 further unique ones from initial_guess) along with other filters.

For the second approach, I used BindCraft targeting the same hotspot, using either the default as well as hardtarget settings, and using relaxed filters. I submitted all 4 of the sequences generated based on the default setting, and filled up the remaining ones (27) with top-ranking sequences based on the hardtarget setting.