The designs crossed from computation into the laboratory
Anthropic says Claude designed 30 candidate binders for each target by choosing binding sites, orchestrating specialized open-source tools, optimizing candidates, and screening for properties such as novelty, solubility, and likely expression.
Adaptyv Bio and Twist Bioscience then produced and tested the designs. Confirmed binding against 14 of 15 reported targets makes the experiment more consequential than a benchmark based only on model answers.
A strong binder is still an early step
Anthropic reports overall hit rates above the typical range it derived from current protein-design campaigns and says some designs matched or exceeded previously published affinity. The campaign also failed against maltose-binding protein and had limited results on another difficult target.
A binder is not a safe, effective therapy. Researchers still need characterization, delivery, toxicology, manufacturing, clinical evidence, and regulatory review. The result accelerates one bottleneck; it does not erase the rest of drug development.
The access problem is part of the science
The campaign used large GPU allocations, specialized tools, internet and data connectors, and models not all available to the public. That combination raises questions about reproducibility and whether independent researchers can obtain comparable capability.
Anthropic also describes autonomous biological research as dual-use and restricts some of its strongest capabilities. A credible access program must let legitimate scientists work while recording use, controlling dangerous operations, and supporting independent evaluation rather than asking the public to choose between unrestricted release and corporate secrecy.
Go to the source
Read the evidence behind this analysis. External links open in a new tab.
Anthropic — How Claude is accelerating protein design and analytical chemistry
