Bispecific Fusion Binder Strategy The central design hypothesis was that linking two single-domain binders against non-overlapping RBX1 epitopes into one construct would yield tighter engagement through avidity than either domain alone. This approach was motivated by the RBX1 PDB ensemble, where endogenous partners—cullins, E2 enzymes, and NEDD8 machinery—simultaneously contact multiple distinct surface regions, indicating the target naturally supports multi-mode recognition. Epitope patches were identified by extracting contact fingerprints from all available RBX1 complex structures at 4.5 Å (heavy-atom) and 8.0 Å (Cα) cutoffs with six-class chemistry annotation. Spatial clustering partitioned the RBX1 surface into discrete regions, and each designed binder was assigned to the patch it engaged. Compatible fusion pairs were required to have non-overlapping footprints. Constructs were assembled with flexible GS linkers (10–20 aa); optimal N-to-C orientation was determined by termini distance analysis, with steric clash filtering removing incompatible pairs. Interface contacts for each partner were frozen during ProteinMPNN linker redesign and enforced as distance constraints during structure prediction. Pairs were ranked by a compound score: per-domain ipSAE was mapped to P(bind) via a logistic function and pairs ranked by P(A)×P(B). A key limitation is that Boltz-2 confidence metrics degrade with construct length; full-length fusion refolds showed ipSAE reductions of ~0.35 relative to single-domain predictions, so fusion rankings are treated as relative rather than absolute.

Binder Generation and Screening Four generators were benchmarked on the same target. BoltzGen (Boltzmann generator, backbone and sequence co-generated) was the primary source of production candidates; 53% of 90–130 aa outputs were antibody frameworks identified by FR1 motif scanning and excluded. Proteina-Complexa (flow-matching with AF2-guided beam search, n_branch=4, beam_width=5) produced strong interfaces (median ipSAE ~0.70) but initially poor monomer folds (pLDDT 0.56); increasing the pLDDT reward weight to 0.8 and applying ProteinMPNN redesign improved pass rates to 7/50, with a top ipSAE of 0.89. RFdiffusion2 generated backbone-only outputs conditioned on RBX1 hotspot patches, with ProteinMPNN for inverse folding. BindCraft was run for comparison only. All designs passed a four-gate screen using Boltz-2 structure prediction (200 diffusion steps, 3 recycling): monomer pLDDT ≥0.70, complex iPTM ≥0.50 and iPAE ≤0.45, ipSAE ≥0.70, and target-dependency confirmation. The iPAE threshold was relaxed from 0.35 to 0.45 after identifying a size-dependent bias that incorrectly rejected 47% of viable candidates on this small target.

Sequence Optimization and Lead Refinement ProteinMPNN was used for inverse folding and iterative refinement. Soft interface bias (bias_by_res, strength=2.0) had little effect—only 1/32 biased positions adopted the target residue at T=0.15—and hard-substitution of natural partner residues at interface positions reduced mean ipSAE from 0.92 to 0.55, confirming that backbone geometry rather than residue identity governs interface quality. ProteinMPNN's practical value was scaffold stabilization: designs with <20% interface mutations had a 93% screen pass rate versus 59% for those with 30–40% interface mutations. Top designs entered an iterative refinement loop: ProteinMPNN mutagenesis (T=0.1), Boltz-2 refolding, rescoring, with the refolded structure—not the generator backbone—feeding back as input for the next round. Four of five test trajectories reached ipSAE ≥0.90 within 2–3 rounds (best: 0.75→0.93, gains of 0.05–0.08 per round). The protocol was counterproductive above ipSAE 0.85: 0/14 designs improved at that threshold (mean delta −0.11), placing the effective optimization window at 0.80–0.85. A confound worth noting: 62% of initial refinement parents were antibody frameworks. After restricting to de novo scaffolds only, the graduation rate fell from 13% to 3%

Structures and Methods PDF: https://drive.google.com/drive/folders/173RgENHHzwcOjbI4iDPX5RcxxiYh0JhP?usp=drive_link