A de novo RBX1 binder design workflow was implemented with BindCraft and focused on the structured RING-domain region of RBX1.

Multiple RBX1 target representations were evaluated to improve sequence recoverability and interface quality, including trimmed constructs centered on the structured core and preserving the zinc-coordinated architecture. Early full-length runs produced trajectory-level backbones but showed weak downstream acceptance after sequence redesign, indicating that target representation was a critical determinant of successful binder recovery. Subsequent runs therefore prioritized structure-guided trimming and interface-focused sampling.

Design generation was performed with AlphaFold2-based hallucination and ProteinMPNN sequence optimization within the BindCraft workflow. Sampling was carried out across multiple binder length ranges, and target-side hotspot guidance was introduced in later rounds to bias design toward recurrent, structurally plausible surface patches identified from earlier successful trajectories. Candidate designs were filtered using the standard multistage BindCraft criteria, including complex confidence, interface quality, structural confidence, and post-redesign validation. This produced a set of de novo binders that were not template-copied but arose from iterative structure-conditioned generation and ranking.

The central hypothesis was that RBX1 is more tractable as a binder-design target when the search is concentrated on the zinc-stabilized structured core and when disordered or weakly constrained regions are excluded from the target model. A second hypothesis was that not all apparently strong backbone trajectories are sequence-recoverable; therefore, downstream redesign robustness and interface reproducibility are more informative than trajectory aesthetics alone. The workflow was thus deliberately optimized not only for predicted binding, but also for design consistency under sequence resampling.

To support experimental relevance, candidate prioritization was extended beyond nominal binding scores by explicitly considering binder developability. This was assessed through PRIB with attention to properties that affect practical progression after computational selection: structural compactness, sequence recoverability after redesign, avoidance of excessively unstable or overly exposed interfaces, reduced indication of frustrated packing or unsatisfied interfacial interactions, and preference for candidates whose binding mode remained credible across redesign and re-prediction steps. In this context, developability was treated as a selection criterion rather than a post hoc consideration. The objective was to avoid candidates that are only computationally attractive in a single trajectory, and instead prioritize designs with a higher likelihood of remaining manufacturable, stable, and experimentally testable.

Final ranking therefore reflected a combined view of interface plausibility, structural confidence, redesign robustness, and developability-informed selection. This was intended to increase the probability that submitted sequences would not only score well computationally, but also represent realistic starting points for downstream biochemical validation.

BindCraft: https://github.com/martinpacesa/BindCraft