GIRAF description for RBX-1 design is available via: https://drive.google.com/file/d/1nr6WwyzQg4r0ajO6sb4MKiFM1atJ9HQB/view?usp=share_link

Our RBX-1 submission was generated with GIRAF, our in-development hallucination pipeline for de novo binder design. We used the RBX-1 sequence together with three Zn ions as input and designed both VHH and Fv binders from fixed antibody frameworks, optimizing only the CDRs.

Starting from 5 VHH scaffolds and 5 Fv scaffolds, GIRAF generated 300 designs per scaffold, giving 3,000 initial candidates in total. Designs first passed a Protenix-v1 filter requiring iPTM > 0.85 and framework-aligned complex-to-monomer CDR RMSDs below 1.0 A. The Protenix-v1-passing set was then refolded with AF3, and we kept only designs with AF3 iPTM > 0.80 and Protenix-versus-AF3 DockQ > 0.49.

We then reviewed the AF3-passing pool using developability metrics, including AbNatiV nativeness, CamSol solubility, NanoMelt-predicted melting temperature, sequence-liability filters, and PyRosetta-derived interface metrics. From this filtered pool, we selected a final panel of 15 VHHs and 15 Fvs. The final Fv submissions were formatted as scFvs by linking the VH and VL chains with a (G4S)4 linker.