We designed a ranked portfolio of de novo minibinders targeting RBX1 using a structure-guided and developability-aware workflow.

Our hypothesis was that the best competition candidates would not come from ranking by a single structural score alone, but from combining target-aware interface design with sequence-level risk control. RBX1 contains a structured C-terminal RING domain that mediates key functional interactions, so we focused on candidates predicted to engage structured target surfaces rather than relying on flexible or poorly defined binding modes.

The workflow started from de novo binder generation and scaffold diversification, followed by structure-based complex evaluation using Boltz-2-derived interface confidence metrics. We prioritized candidates with stronger interface plausibility, cleaner target engagement, and better overall complex quality. Rather than selecting only the top-scoring near-duplicates, we treated the submission as a portfolio problem and retained multiple scaffold families and interaction patterns.

To reduce experimental risk, we added an orthogonal sequence-level triage layer using proprietary developability models (Astra). We prioritized candidates with stronger predicted solubility, favorable thermostability profiles, and lower indicators of disorder or aggregation propensity - filtering out likely experimental failures before submission. The final set is designed to maximize the fraction of candidates that express, fold, and produce interpretable binding signal.