https://drive.google.com/file/d/1QWCedePQbzfSXvcJ_03tA4Kuyxuqkt0W/view?usp=share_link

Hypothesis Our central hypothesis is that effective binders to the RBX1 N-terminal region, especially residues 27–33, are not isolated one-off solutions but members of reproducible structural families. These families should repeatedly realize the same β-augmentation-like interface geometry across different H/E grammar patterns and length blueprints. In addition to recognizing the core epitope, successful binders should also maintain compatible entry geometry toward the upstream disordered segment (residues 19–26) while remaining sterically compatible with the adjacent rigid RING domain. Therefore, local binding discovery and body-compatibility evaluation should be decoupled, and selection should be performed at the family level rather than on single backbones.

Workflow

1. Grammar-guided backbone generation We used RFdiffusion on an RBX1 N-terminal reference fragment with predefined 5- or 6-module H/E grammar patterns and multiple length blueprints. Generation was performed under strict length freezing and low-noise settings, with hotspot guidance centered on key target residues. The rigid RING body was intentionally excluded at this stage so that diffusion focused on local binding geometry rather than full-body interference.
2. Initial structural filtering and family clustering Generated backbones were filtered for mixed α/β character, the presence of a target-facing β element, and sufficient β-core pairing. Surviving designs were clustered with Foldseek and grouped into families based on reproducible fold class, β-pairing pattern, and interface character across different blueprints. Family definition relied on preserving hard invariants while allowing limited variation in softer geometric features.
3. Family-preservation gate and fixed-body compatibility audit Candidates that preserved family-level structural semantics were then tested against an external fixed-body gate. In this step, the IDR core was aligned to a canonical frame, a rigid RBX1 body reference was inserted, and a small hinge/translation ensemble was scanned. Designs were evaluated for steric clashes, binder–body distance, β-pairing continuity over residues 27–33, contacts toward residues 19–26, and relative exposure metrics. The body was used only as an external auditor, not as a generative input.
4. Sequence realization with SolubleMPNN For each gated backbone, sequences were generated using SolubleMPNN, including both full-backbone realizations and partial redesigns around the entry region near residues 19–26. This produced multiple sequence variants per backbone and improved the chance of obtaining soluble, well-folded single-domain binders.
5. Staged AF2 validation and final ranking Sequence candidates were first screened as monomers, then evaluated in cropped primary complexes. Final selection integrated AlphaFold2-based confidence with PyRosetta-based structural assessment, emphasizing interface self-consistency, family-level robustness, and packing/geometry quality, while penalizing failure modes such as self-dimerization and β-core-to-helix drift even when raw AF2 confidence appeared high.