The main idea was to design a binder that locks a naturally disordered region of RBX1 into a β-sheet conformation. The target was extracted from PDB 1LDJ, where this region is already captured in a β-strand by a native binding partner. Two methods released in 2026 were used. β-pairing RFdiffusion was built for this kind of problem, generating binder backbones that form direct hydrogen-bonded β-sheet contacts with the target strand. Lengths from 80–200 aa were sampled and hotspot anchors were placed on both the disordered loop and the ordered RING domain. In practice, the RING domain didn't contribute much, while the loop was captured consistently and with clean geometry. ProteinMPNN was used for sequence design, designs were validated with ColabFold and Boltz-2, and the best were redesigned with ProteinMPNN — which massively improved confidence metrics. Protein Complexa — very freshly released — was used in three modes: Best-of-N, beam search, and hydrogen bond-guided, of which beam search and hydrogen bond-guided were the most successful. Only 1,000 designs were generated per mode, each run taking 5–6 hours on a standard gaming PC. AlphaFold and ESMFold were used as in-model reward signals, and the top designs were folded separately with Boltz-2. The best were then redesigned with ProteinMPNN, which again massively improved the confidence metrics. It is a pleasure to use — very well coded, as one would expect. Final filtering was based on Boltz-2 and ESMFold: monomer pLDDT ≥ 73, pTM ≥ 0.70, and self-consistency RMSD < 2.0 Å, retaining only binders that fold stably in the absence of the target. iPSAE was used as the primary interface metric, as conventional PAE-based scores are unreliable when part of the target is disordered in solution. Final selection yielded 70 RFdiffusion and 30 Complexa binders.