An LLM agent that operates Rosetta, the dominant physics-based protein design software, can match specialist models and human experts on protein design benchmarks.
Researchers introduced Agent Rosetta, a system that wraps the Rosetta software suite inside a structured environment a large language model can navigate to design proteins. Unlike most recent protein AI, which leans on deep learning trained on known sequences, Agent Rosetta handles non-canonical amino acids - building blocks outside the standard 20 - where those ML methods break down. The system iteratively refines designs against user-defined goals, combining LLM reasoning with Rosetta's physics-based generality. One critical finding: prompt engineering alone was insufficient to generate valid Rosetta actions; getting the surrounding environment right proved to be the essential ingredient.
That last point deserves more attention than the benchmark numbers. The prevailing assumption in AI-for-science circles is that a smart enough prompt can coax any LLM into operating complex tools - this work is direct evidence against it. The non-canonical capability matters separately, because some of the most promising drug targets require unusual chemistry that standard amino acids can't provide, a gap that ML models trained on natural proteins simply don't cover. If the environment-design lesson generalizes, it has implications for every lab trying to retrofit an AI layer onto legacy scientific software.
This is a research paper, not a shipped product, and "comparable performance" on curated benchmarks is a long way from displacing the bench scientists who built those baselines in the first place.