De Novo RBX-1 Binder Design via Boltz-2 Optimization with Templated Backbone Amplification

Team: Jonah Kallenbach (Member of Technical Staff, Anthropic) Contact: jonahkallenbach@gmail.com

Background

Analysis of natural binding partners (glomulin, UBC12, UBE2D2, ARIH1; PDB 4F52, 4P5O, 6TTU, 7B5L) showed a conserved epitope centered on residues Ile44, Arg46, Trp87, Arg91, Pro95, and Leu96. Glomulin achieves 40 nM affinity via a 30-residue interface on this surface — the largest among known partners — motivating a strategy of wrapping this epitope with a similarly extensive designed interface.

Method

Design. Binders were generated de novo by direct optimization against Boltz-2 [2] using the Mosaic pipeline [1]. The primary configuration targeted the RING domain (residues 39–108) with 120-residue binders, using PDB 4P5O chain B as a target-side structural template. The loss combined binder–target contact, inverse-folding sequence recovery (ProteinMPNN [3], weight 5.0), PAE, ipTM, and pLDDT terms. Cysteine was excluded from the binder alphabet to avoid interference with the target's zinc coordination.

After identifying a strong initial hit, I generated a second round of designs using that hit's AF3-predicted backbone as a binder-side structural template, constraining new sequences to the validated fold while diversifying surface chemistry.

All code for design, validation, and ranking was written and orchestrated using Claude Code and other internal Anthropic tools.

Filtering. Candidates were cross-validated with AlphaFold 2-multimer (interface loss) and monomer pLDDT (fold stability). Approximately 50 designs were evaluated on the AlphaFold 3 Server [5], which served as a more trusted ranking signal. Protenix [4] was calibrated against the AF3-tested set and used to rank remaining untested designs.

Ranking. The final top 10 was chosen by AF3 ipTM with a bonus for AF2-multimer agreement and a penalty for monomer pLDDT < 0.80. A fold-diversity constraint capped designs sharing the templated backbone at five of ten slots; the remaining five are independent hallucinations with distinct folds.

In-Silico Results

Approximately 1,000 designs were generated across configurations. Of 50 tested on AF3, 30 achieved ipTM ≥ 0.70. The templated-backbone configuration achieved a 70% AF3 hit rate (18/25 tested). The top 10 submitted designs span AF3 ipTM 0.82–0.91, with monomer pLDDT 0.78–0.98. The rank-1 design is the only candidate with replicated AF3 validation (ipTM 0.83–0.86 across runs) and AF2-multimer consensus (interface loss −0.29). All submitted designs are cysteine-free and 120 residues.

References

[1] Escalante Bio. Mosaic: Direct optimization for protein design. https://github.com/escalante-bio/mosaic

[2] Passaro, S., Corso, G., Wohlwend, J., et al. Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. bioRxiv (2025). doi:10.1101/2025.06.14.659707

[3] Dauparas, J., Anishchenko, I., Bennett, N., et al. Robust deep learning-based protein sequence design using ProteinMPNN. Science 378, 49–56 (2022). doi:10.1126/science.add2187

[4] ByteDance AML AI4Science Team. Protenix: Advancing Structure Prediction Through a Comprehensive AlphaFold 3 Reproduction. bioRxiv (2025). doi:10.1101/2025.01.08.631967

[5] Abramson, J., Adler, J., Dunger, J., et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024). doi:10.1038/s41586-024-07487-w