We developed a fully computational pipeline to design de novo protein binders targeting the GLMN inhibitory interface of RBX1 as part of the GEM x Adaptyv RBX1 Binder Design Competition. Our approach combines structure-guided generative modeling using BoltzGen, reinforcement learning (RL)-enhanced weights (Escalante Bio) to improve structural consistency, and consensus-based validation using AlphaFold3 (AF3) and Boltz-2.

RBX1 is a RING E3 ligase whose interaction with GLMN is mediated by a defined surface within the RING domain. Based on structural analysis, we targeted a bipartite interface spanning residues 43–58 and 86–98. To minimize artifacts from flexible regions while preserving structural context, we used a trimmed construct (residues 34–108) including all three Zn²⁺ ions. This retains the full structured RING domain while removing the disordered N-terminus that biases generative models.

Binder design was performed using BoltzGen (protein-anything protocol) with sequence lengths of 65–100 amino acids and explicit hotspot constraints. To improve structural consistency and interface fidelity, we incorporated reinforcement learning (RL) fine-tuned weights (Escalante Bio), which significantly increased the yield of structurally consistent designs.

We conducted two parallel large-scale campaigns: one using default BoltzGen weights and one using RL-enhanced weights, each scaled to thousands of generated designs. Outputs were merged, deduplicated, and re-filtered using BoltzGen’s internal pipeline. The top 100 designs were selected based on internal ranking.

These 100 designs were further validated using two independent structure prediction models: AlphaFold3 (AF3) and Boltz-2, using the full RBX1 sequence (residues 1–108 with Zn²⁺ ions) to approximate experimental conditions. Designs were ranked using a consensus scheme prioritizing AF3 iPSAE, followed by Boltz-2 iPSAE, AF3 iPTM, and Boltz-2 iPTM. Candidates were grouped into confidence tiers based on cross-model agreement.

Final designs were visually inspected to confirm binding to the GLMN interface, correct hotspot engagement, and absence of steric clashes with Zn²⁺ ions. Top-ranked designs consistently achieved high-confidence interface predictions (AF3 iPSAE > 0.8, iPTM > 0.9) with strong agreement across models.

An interactive dashboard containing all ranked designs, structural metrics, and visualizations is available at: https://carlos-laborda.github.io/rbx1-binder-dashboard/