LLM‑driven Agent Rosetta can command the Rosetta physics engine to design proteins.
The research team built a language‑model agent that talks to a structured Rosetta environment, issuing actions instead of raw text prompts. In tests the agent iteratively refined sequences to meet user‑specified goals. On standard amino‑acid designs it performed on par with dedicated machine‑learning models and human experts. When asked to incorporate non‑canonical residues—something most ML pipelines cannot handle—Agent Rosetta achieved comparable success, while pure prompt‑only approaches failed to produce any useful Rosetta commands.
This matters because it proves that coupling LLM reasoning with a well‑designed software interface can extend the reach of AI into niche scientific domains. It shows a path beyond niche, pre‑trained models toward generalist agents that leverage existing physics‑based tools, potentially accelerating discovery in biochemistry and synthetic biology.
The takeaway: a carefully built environment lets LLMs become practical operators of specialist software, delivering results that match or exceed narrow AI solutions while keeping the flexibility of a generalist agent.