We designed 100 protein binders targeting RBX-1 using RFdiffusion for backbone generation and SolubleMPNN for sequence design, guided by an active learning framework that efficiently explores the design parameter space across multiple structural conformations and binding strategies.

Active learning for parameter space exploration:

• Binder design is formulated as optimization over 5 parameters: binder length (40-150aa), RFdiffusion noise scale (0-0.5), MPNN sampling temperature (0.05-0.2), target PDB conformation (10 structures), and hotspot steering subset
• Each parameter combination produces candidates scored by a composite fitness function combining Rosetta ddG, interface area, E2 hotspot contacts, and H-bond counts
• A Gaussian Process (GP) with Matérn-5/2 kernel models the fitness landscape as a surrogate: after Cycle 1 (Latin hypercube sampling, 3-5 diverse configs), the GP infers expected fitness and uncertainty across the parameter space
• Upper Confidence Bound (UCB) acquisition selects the next configuration: UCB = μ(x) + β·σ(x), balancing exploitation (high predicted fitness μ) with exploration (high uncertainty σ). β decays across cycles (2.0 × 0.8^cycle) to progressively shift from broad search to local refinement
• 3 cycles per target: Cycle 1 explores broadly, Cycles 2-3 exploit GP-identified promising regions. This navigates ~500 possible configurations per target while evaluating only ~15-20

Structural context (informing hotspot selection and target preparation):

• E2 recruitment hotspots identified via alanine scanning mutagenesis: W87A and R91A abolish CDC34~Ub binding and activation respectively (PMC3366790); the E2-binding cleft spans Loop 1 (I44) to the α2-helix (W87, K89, R91, P95, L96)
• Glomulin (PDB 4F52), a natural E2 blocker, enters from the IDR-RING junction (res 34-48) and spans to the E2 helix — this informed our Junction+RING protocol
• NMR ensemble (PDB 2LGV, 20 models) shows the IDR folds back toward the E2 face in some conformations, enabling binders to bridge both regions
• The IDR-RING junction (res 30-39) is cysteine-free, favorable for cell-free expression

Multi-conformation design:

• 10 RBX-1 conformations spanning inactive (1LDJ), neddylated-active (3DPL), CAND1-inhibited (1U6G), CRL4-bound (2HYE), Glomulin-blocked (4F52), E2-docked (4P5O), binary (3RTR), and apo NMR states (2LGV m0/m5/m10)
• Designing across conformations produces binders robust to RBX-1's conformational flexibility in solution
• RING domain RMSD across all structures: 0.5-0.9Å at the E2 face, confirming the target surface is structurally conserved

Expression optimization:

• SolubleMPNN for all sequence design (93% expression rate in PO-102 EGFR competition vs 0% for standard ProteinMPNN)
• Amino acid bias (C:-0.8) reduces cysteine incorporation for cell-free expression compatibility (no disulfide machinery)
• ESM-2 log-likelihood filtering removes sequences with low naturalness scores

Scoring and ranking:

• Rosetta InterfaceAnalyzer (ddG via constrained FastRelax + pack_separated, dSASA), Boltz-2 structure prediction (ipTM, iPAE), BioPython E2 hotspot contacts, ESM-2 log-likelihood
• Composite ranking: binding energy (25%) + E2 contacts (40%) + expression likelihood (25%) + structural confidence (10%)