AI models are taking on complex scientific tasks with uneven results. An Anthropic agent successfully designed binding proteins in a closed-loop experiment, while a separate effort to write a machine-learning paper ended in rejection by academic reviewers.
Anthropic published two experiments in which Claude models directed an end-to-end protein design campaign against 15 targets over 48 hours. Using open-source tools such as RFdiffusion3 and ESMFold, the Claude Mythos Preview model achieved a 26.7 per cent hit rate. Human involvement remained limited to approving requests, resolving infrastructure problems and ordering the resulting designs from laboratory providers.[1]
The agent outperformed human baselines on specific targets, particularly against RBX1, where its best design bound at 3.9 nanomolar compared to 45 nanomolar for the winner of a human contest. The model reached a 40 per cent hit rate on that target, while human entrants achieved 3.7 per cent. Extending the runtime to 24 hours per target increased the overall hit rate to 35.1 per cent, though the system failed outright on the maltose-binding protein.[1]
A separate effort to have an agent produce novel research resulted in rejection. Researchers at Princeton University gave a system six days and a 3,000-dollar budget to answer unpublished questions from two NeurIPS 2026 papers. The original authors evaluated the generated conference submissions and rejected both. This outcome demonstrates that directing existing simulation tools requires a different capability than independently formulating a conceptual scientific contribution.[2]
