In this study, we performed de novo binder design against RBX1 in the 1U6G structure. We first examined the structure of RBX1 and noted the presence of intrinsically disordered regions, which could introduce additional conformational uncertainty and interfere with interface recognition. Based on this observation, we divided the design task into two parallel tracks: one using full-length RBX1 as the design target, and the other using RBX1 with the disordered region removed. This setup was intended to cover both the complete structural context and a more stable core structural context, thereby broadening the design space explored.

For candidate generation, we used BoltzGen in both design tracks, generating approximately 100,000 sequences for each target formulation. We then selected an initial set of 200 candidates from each batch for further screening. This step was not based solely on a single score cutoff. Instead, we manually reviewed the candidates and further refined the set based on their overall sequence characteristics. In particular, we prioritized candidates with clearly distinct sequence patterns in order to prevent the final set from collapsing into a single design family, while removing sequences that were obviously redundant or differed only by minor local variations. We also aimed to retain representative candidates from both design tracks, rather than merging all sequences into a single ranking by score, so that the final set would cover designs derived from different target formulations.

To further improve novelty, we used MMseqs2 to compare the candidate sequences against the UniRef50 database and removed sequences with greater than 70% sequence identity to known proteins. After this filtering process, we retained 100 sequences as the final submission set. Overall, this workflow was designed to balance binding feasibility, diversity within the candidate pool, and novelty relative to known protein sequences.