Submitted toGEM x Adaptyv: RBX1 Binder Design Competition
Description
- Introduction and Target Rationale RBX1 is an attractive but challenging target for binder design because it combines a disordered N-terminal region with a structured C-terminal RING domain. This makes conventional pocket-based docking insufficient and instead calls for a structure-aware, surface-oriented de novo design strategy. To address this problem, we developed a multi-branch computational pipeline with unified downstream screening and ranking.
- Overall Workflow Overview Our design strategy consists of four parallel computational pipelines—BoltzGen, an RFdiffusion3-based pipeline, an RFdiffusion–ProteinMPNN–AF2 pipeline, and BindCraft—followed by a unified multi-stage screening and ranking procedure. The parallel generation stage was intended to broaden exploration of the RBX1 binder design space by leveraging different backbone generation, sequence design, and refinement strategies. This diversified generation scheme was followed by an initial candidate filtering step based on ipSAE_min, which enabled consistent early prioritization across workflows. Selected candidates were then subjected to AF3-based complex evaluation for deeper structural assessment of binder folding, target engagement, and interface plausibility. Final candidates were ranked across all workflows under a common prioritization framework and reduced to a non-redundant, diverse set for submission.
- Initial Candidate Filtering by ipSAE_min Candidates produced from all four pipelines were pooled and subjected to an initial filtering step based on ipSAE_min. This stage served as a shared early evaluation layer across otherwise distinct design branches. By applying a common structural quality criterion at this point, we were able to reduce the candidate pool while preserving the strongest designs from each workflow. This also improved comparability across branches and enabled a more unified downstream evaluation strategy. The purpose of this filtering step was not to determine final winners directly, but rather to remove low-priority candidates before more computationally intensive complex-level assessment. In this way, ipSAE_min functioned as an efficient funnel for triaging designs emerging from diverse generation routes. 4.Cross-Workflow Ranking and Final Selection Following AF3-based complex evaluation, candidates were ranked under a unified cross-workflow framework. Rather than ranking each branch independently, we compared designs from all four pipelines in a shared prioritization scheme. This allowed us to select the most promising candidates globally rather than overcommitting to any single workflow. Final ranking emphasized several principles: structural confidence, predicted interface quality, consistency of target engagement, and diversity across the submission set. Redundant candidates were removed to avoid overrepresentation of closely related designs, and diversity-aware filtering was used to preserve breadth in both sequence and putative binding mode. The final output was a ranked set of top RBX1 binder candidates selected for submission. 5.Summary In summary, we developed a four-branch computational workflow for de novo RBX1 binder design that integrates generative design, sequence assignment, structure-based filtering, AF3-based complex evaluation, and unified ranking. By combining multiple parallel design strategies with a common downstream screening framework, our approach aims to balance novelty, structural plausibility, and diversity in the final submission set. This workflow provides a rational and scalable strategy for identifying candidate binders for RBX1 in the GEM x Adaptyv competition setting.