Summary
A bioRxiv preprint introduces Chaperone, a framework for coordinating protein-design tools. Its authors report reduced token use and GPU costs across nine design tasks, alongside results from several protein-design campaigns.
A bioRxiv preprint posted on October 20, 2026, introduces Chaperone, a framework for coordinating software used in computational protein design. The authors report that across nine biomolecular design tasks, Chaperone reduced token use by 86% and GPU costs by 64% relative to baseline agents, while increasing the computational design pass rate. The work is a preprint, and its abstract reports aggregate results rather than task-by-task figures.
A coordination layer for protein-design software
Protein-design projects often combine several programs: one may generate a protein structure, another propose its amino-acid sequence, and others predict or assess the resulting design. Chaperone is intended to connect these stages through a programmable interface designed for use by AI agents. The authors say it integrates RFdiffusion3 (RFD3) with sequence-design, structure-prediction and evaluation tools, including ProteinMPNN and AlphaFold3, as well as more than 50 other tools.
The framework also supports parallel execution on high-performance computing systems. In practical terms, that lets an agent coordinate multiple computational steps and run parts of a workflow concurrently, rather than treating each tool as a separate manual task. The authors present this combination of standardized tool access and parallel execution as a way to make longer, more complex design workflows easier to run and iterate.
Reported results across design tasks
In the evaluation of nine biomolecular design tasks, the authors compared Chaperone with baseline agents. They report 86% lower total token use and 64% lower GPU costs, along with an increase in the overall computational design pass rate. The abstract does not give the size of that pass-rate increase or results for each task, so the reported reductions describe the tested evaluation rather than a general performance guarantee.
The preprint also describes a two-week design campaign for small-molecule binders. The authors report that it produced low-nanomolar binders to cortisol without experimental optimization during the campaign, which they describe as “zero shot.” Other examples include a symmetric, 12-subunit self-assembly nanoparticle and a designed linker connecting C2- and D3-symmetric protein complexes into crystalline assemblies.
These examples illustrate the intended value of a workflow harness: coordinating tools can help researchers explore more design candidates and carry out diverse computational tasks with less overhead. The authors say Chaperone is freely available through the RosettaCommons GitHub repository.