Anthropic’s Claude used free, open-source protein-design software to create 1320 candidates, with 354 binding successfully across 14 of 15 targets tested.
Anthropic’s Claude autonomously designed 1320 protein binders using only open-source, freely downloadable design software, with Anthropic paying no software licensing fees. The technical report says AlphaFold 3 weights and two other packages were excluded because of restrictive licences.
Given 16 protein targets, cloud compute and a roughly 16,000-word protocol, Claude researched each target, selected binding sites, ran the software and produced 30 ranked designs per target. Two independent laboratories then synthesised and tested the candidates.
Of the 1,320 designs, 354 bound their targets, producing a 27% hit rate, while 49% of Claude’s top-ranked designs succeeded. At least one design worked for 14 of the 15 targets with readable measurements. Against RBX1, Claude produced 28 binders from 90 designs, compared with nine from 245 entries in a human open design competition. Its strongest design bound roughly 10 times more tightly than the competition winner in side-by-side testing.
However, the experiment also exposed the next bottleneck: physical synthesis and experimental screening. Adaptyv Bio tested all 1,320 designs, while Twist Bioscience tested 1,260 independently.
“Experimental screening remains the only way to learn which targets a campaign has succeeded on and the affinity of the designed binders.”
Anthropic has published the prompts and data on Hugging Face for others to rerun. The findings suggest open-source tools can lower the computational expertise barrier, but wet-lab capacity remains scarce.
The study demonstrated binding, not biological function, and does not establish that Claude outperformed human experts.
