Initial Candidate Generation We utilized RFdiffusion, ProteinMPNN, and our in-house model Ribona (to be release) as base methods to design de novo binder candidates against RBX1. Two backbone generation modes were employed across these tools: unconditioned de novo design and hotspot-conditioned design with hotspots (<6) defined over the structured non-disordered regions of RBX1 and distant to the zinc coordination sites. We retain only the structured C-terminal RING-H2 domain, excluding the intrinsically disordered N-terminal region of RBX1, as the target binding interface context. The crystal structure of the Cul1-Rbx1-Skp1-F box Skp2 SCF ubiquitin ligase complex (PDB: 1LDJ, resolution 3.00 Å; Zheng et al., Nature 2002) was sed as a template. This represents a high-resolution experimental characterization of RBX1 within its native CRL catalytic core, capturing the RBX1 RING-H2 domain in an intermolecular beta-sheet interaction with Cullin-1 and coordinating three zinc ions essential for domain stability, and make it a well-validated scaffold.

Optimizations All top ranking candidates generated from the part 1 were passed as inputs into two independent optimization pipelines. The first used our in-house framework SCALE (to be release) in which structure-quality-guided sequence mutations were proposed and accepted based on predicted structural quality metrics across successive refinement rounds. The second pipeline employed ProteinHunter with default Boltz settings for structure refolding, combined with LigandMPNN-based sequence redesign. ProteinHunter was applied in two modes: first, as a downstream optimization pipeline receiving good candidates from the part 1; and second, as a standalone generative pipeline in which all designs were initialized from scratch (X). In both modes, only candidates with an ipTM score above 0.65 were retained for downstream consideration.

Candidates Filtering and Selection The top 600 candidates ranked by weighted ipTM score across all methods and pipelines were aggregated and subjected to AlphaFold3 complex structure prediction, with zinc ions explicitly included in the prediction inputs to maintain RING domain coordination geometry. For structural quality assessment, the ipSAE (interaction prediction Score from Aligned Errors) metric was computed, as it restricts the TM-score calculation to residue pairs with good predicted aligned error values at the interface and adjusts the d0 normalization parameter accordingly, making it robust to the presence of disordered regions distal to the binding interface. Final candidate ranking was performed through particular consideration of ipSAE, ipTM, pTM, and QDock scores. Rosetta was additionally employed to evaluate interface biophysical properties including solvent-accessible surface area (SASA) and binding free energy (ΔG), for complementary thermodynamic assessment of interface quality. Top-ranked designs underwent manual inspection of predicted binding pose and interface geometry for likely structural stability, yielding the final submitted candidate set.