Multi-Track De Novo Binder Design for RBX1

Team: Jozef Fülöp Affiliation: UCT Prague Contact: fulopj@vscht.cz

Background

RBX1 (108 aa; UniProt P62877) comprises an intrinsically disordered N-terminal segment (residues 1-39) and a zinc-coordinated RING-H2 domain (residues 40-108). We addressed this mixed architecture using three computational design tracks and a shared scoring pipeline.

Method

We generated 1,748 unique binder candidates across three tracks. BoltzGen v0.3.1 in sequence-only co-folding mode produced 1,329 designs by modeling the binder together with full-length RBX1 [1]. In this track, binders of 40-70 aa showed the highest internal pass rate (22%). A separate hotspot-guided campaign focused on the 2LGV RING domain and E2-binding groove yielded no passing designs (0/11,600). BindCraft generated 366 designs against the RBX1 RING domain (PDB 2LGV; residues 40-108) using AF2-based hallucination with E2-groove hotspot constraints [2]. ProteinMPNN/SolMPNN redesign of selected BoltzGen backbones produced 53 additional sequences while preserving interface geometry.

Scoring and ranking

All candidates were scored with ESM2 pseudo-log-likelihood and Boltz-2 co-fold ipTM [3]. Boltz-2 coverage was 92%, and ESM2 coverage was 100%. A subset was cross-checked with OpenFold3 (24 designs) and AF2-Multimer ipTM (366 BindCraft designs). Final ranking was driven primarily by Boltz-2 ipTM. When duplicate Boltz-2 predictions were present, the lower score was retained as a conservative estimate. This prioritization strategy is consistent with recent binder-design benchmarking and filtering studies that emphasize confidence-based ranking, especially interface-focused metrics [4,5].

Selection

We selected the top 120 designs, applied diversity clustering (edit-distance threshold 0.2; maximum four designs per cluster), and retained 105. MMseqs2 novelty filtering against UniRef50, using both full-length sequences and 30-residue sliding windows at ≤75% identity, reduced this set to 102 passing candidates. The top 100 were submitted.

In silico results

All 100 submitted designs had Boltz-2 ipTM > 0.92 (mean 0.934; max 0.965). Lengths ranged from 40 to 120 aa. The final set comprised 58 BoltzGen designs, 40 BindCraft designs, and 2 MPNN redesigns. No top-ranked submitted design contained an unpaired cysteine. Across 300 predicted complexes, three recurring binding modes were observed: IDP+RING bridging (46%), IDP-dominant (44%), and RING-dominant (10%). The co-folding track was the main source of designs spanning both the disordered N-terminus and the structured RING region.

References

[1] Stark H, Faltings F, Choi M, et al. BoltzGen: Toward Universal Binder Design. bioRxiv. 2025. https://doi.org/10.1101/2025.11.20.689494
[2] Pačesa M, Nickel L, Schellhaas C, et al. One-shot design of functional protein binders with BindCraft. Nature. 2025;646(8084):483-492. https://doi.org/10.1038/s41586-025-09429-6
[3] Passaro S, Corso G, Wohlwend J, et al. Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. bioRxiv. 2025. https://doi.org/10.1101/2025.06.14.659707
[4] Overath MD, Rygaard AM, Jacobsen CS, et al. Predicting Experimental Success in De Novo Binder Design: A Meta-Analysis of 3,766 Experimentally Characterised Binders. bioRxiv. 2025. https://doi.org/10.1101/2025.08.14.670059
[5] Ren M, Sun J, Guan J, et al. PXDesign: Fast, Modular, and Accurate De Novo Design of Protein Binders. bioRxiv. 2025. https://doi.org/10.1101/2025.08.15.670450