Reliable and Efficient Agentic Protein Design with Toolset Harnessing

Computational workflows for designing proteins efficiently and reliably often involve coordinating various tools, data formats, and analysis stages, which can frequently be disjointed. We present Chaperone, an adaptable and streamlined framework that facilitates easy integration of the RFD3 model for sequence design with downstream tools for predicting protein structures and assessing designs, including ProteinMPNN, AlphaFold3, and over 50 other protein design tools.

Chaperone's programmable agent-friendly interface, coupled with parallel processing on high-performance computing systems, serves as a harness to facilitate rapid iterations and scalable exploration of intricate protein design workflows. In assessments of nine biomolecular design challenges, utilizing Chaperone led to an 86% reduction in total token consumption compared to initial agents, and a 64% decrease in GPU expenditures, while simultaneously raising the overall computational design success rate.

We showcase Chaperone's usefulness for extensive exploration of design spaces through a two-week endeavor to design small-molecule binders for cortisol, achieving nanomolar binding affinities without prior knowledge. Furthermore, we highlight the adaptability of Chaperone across diverse design challenges, such as constructing a symmetric 12-subunit self-assembly nanoparticle and a linker that enables interactions between C2- and D3-symmetric protein complexes to form crystalline assemblies.

These findings underscore how agent harnesses constructed around standardized tool interfaces can significantly enhance the efficiency and success of intricate, long-term agentic design workflows in the field of biology. Chaperone is made freely accessible to the community on GitHub: https://github.com/RosettaCommons/chaperone.