Title: [GEM x Adaptyv: RBX1 Binder Design Competition] Team: [1]Harshit Singh,Phd Scholar , Shiv Nadar University, Delhi-NCR, India, [2] Rajeev Kumar Singh,Professor,Shiv Nadar University, Delhi-NCR, [3] Rohan Gorantla, Novartis Biomedical Research, Basel, Switzerland

Contact : hs494@snu.edu.in,rajeev.kumar@snu.edu.in, gorantlarohan@gmail.com

Description: We present a fully computational de novo pipeline for designing protein binders against RBX1 (RING Box Protein 1), a critical component of Cullin-RING E3 ubiquitin ligase (CRL) complexes. Our approach combines RFdiffusion for backbone generation conditioned on the E2-ubiquitin binding interface, followed by ProteinMPNN for sequence design. Candidate binders were ranked by ProteinMPNN log-likelihood scores and sequence diversity. All submitted sequences are de novo with verified edit distance >25% to UniRef50.

1. Target Analysis RBX1 is a 108-amino acid RING-H2 finger protein that recruits E2 ubiquitin-conjugating enzymes to Cullin-RING ligase complexes, controlling degradation of ~20% of cellular proteins. The protein comprises two structurally distinct regions: an intrinsically disordered N-terminal region (residues 1–26) and a structured C-terminal RING-H2 domain (residues 27–108) coordinating three zinc ions via a cross-brace Cys/His motif (C42, C45, H48, C53, C56, C68, C75, H77). Binder design was targeted exclusively to the structured RING-H2 domain. The E2-binding interface was identified from the Cul1-Rbx1-Skp1 co-crystal structure (PDB: 1LDJ, Zheng et al. 2002) and confirmed via computational surface exposure analysis using Biopython NeighborSearch. Zinc-coordinating residues were explicitly excluded from hotspot selection to preserve the structural integrity of the fold. Six hotspot residues were selected targeting the E2-enzyme contact surface: W33, K89, W35, R91, R86, W87 — a combination of hydrophobic anchor residues (W33, W35, W87) and electrostatically important contacts (R86, K89, R91).
2. Computational Design Pipeline 2.1 Structure Preparation The full-length RBX1 structure was obtained from the AlphaFold Database (UniProt: P62877, AlphaFold DB v6) via the EBI API. The structure was cleaned using Biopython PDBParser, retaining only chain A and standard amino acid residues. The experimental NMR structure (PDB: 2LGV) was used as a cross-reference for RING domain conformation validation. 2.2 Backbone Generation — RFdiffusion Binder backbones were generated using RFdiffusion (Watson et al. 2023) with the Complex_base checkpoint. The model was conditioned on the RING-H2 domain (residues 27–108) with six hotspot residues specifying the target interface. Binder length was sampled in the range 70–100 amino acids. The diffusion process achieved consistently low motif RMSDs (0.10–0.17 Å), indicating high-confidence placement of the binder backbone relative to the target interface. RFdiffusion parameters: contigmap.contigs=[A27-108/0 70-100] ppi.hotspot_res=[A33,A89,A35,A91,A86,A87] 2.3 Sequence Design — ProteinMPNN Amino acid sequences were designed for each backbone using ProteinMPNN (Dauparas et al. 2022). Chain A (RBX1) was held fixed during design; only chain B (the binder) was redesigned. Multiple sequences were generated per backbone at sampling temperature 0.1 to produce stable, low-energy sequences. Sequences were ranked by ProteinMPNN log-likelihood score (lower = better). Near-duplicate sequences (>95% identity) were removed to maximize submission diversity.
3. Filtering and Ranking Criteria Criterion Threshold Rationale ProteinMPNN score < median of all designs Sequence-structure compatibility Sequence length <= 250 AA Competition rule compliance Edit distance (UniRef50) >= 25% Competition rule — de novo requirement Sequence identity < 95% to other submissions Maximize submission diversity Standard amino acids ACDEFGHIKLMNPQRSTVWY only Experimental synthesis feasibility Final ranking was determined by ProteinMPNN log-likelihood score (primary), with sequence diversity used as a tiebreaker to ensure the submission covers distinct regions of sequence space. Rank 1 corresponds to the design with the best (lowest) ProteinMPNN score.
4. Novelty and Design Rationale All submitted sequences are strictly de novo — no motif scaffolding, template grafting, or lead optimization was performed. The design strategy is distinguished by three features: (1) Interface-focused hotspot selection grounded in crystallographic data from the Cul1-Rbx1-Skp1-F box complex rather than generic surface exposure metrics; (2) Explicit zinc-coordination avoidance — the eight zinc-coordinating residues of the RING-H2 domain were excluded from the hotspot set to prevent designs that would disrupt the structural fold; and (3) Therapeutic relevance — targeting the E2-binding surface directly disrupts CRL catalytic activity rather than merely occupying a peripheral surface, maximizing the potential functional impact of a successful binder.