The 80-amino-acid miniproteins were designed using Boltzgen. We ran four parallel jobs, each generating 10,000 designs with a compute budget of 2,000. The final 2,000 designs were ranked based on their pAE and iPTM scores. Of the 40,000 total designs, only 173 met our thresholds (pAE < 10 and iPTM > 0.7). These sequences were then refolded using Boltz2, AF3, and Chai-1. We performed an additional round of filtering based on the ipSAE score from the Boltz2 predictions (ipSAE > 0.8) and on the RMSDs between the Boltz2/AF3 and Boltz2/Chai-1 predicted structures, reducing the pool to 14 candidates.

These 14 candidates include a mixture of topologies such as beta-sandwich, helical-bundle, and Top7-like folds. To further filter these candidates, we performed 500-ns folding simulations using MELD. We hypothesized that designs passing both AI-based and physics-based filters would represent high-confidence candidates with greater likelihood of experimental success. This step narrowed the set to five candidates.

Because these five candidates contained cysteine, we attempted to further optimize their sequences using SolubleMPNN, generating 1,000 new sequences per candidate while preserving the binder sequence within the target hotspot. The top 100 redesigned sequences for each candidate were refolded with Boltz2, AF3, and Chai-1 to identify the best candidates. The final list of candidates corresponds to sequences with the highest ipSAE values in both the ProteinBase website and our local predictions.