A general chatbot drove the specialist protein tools itself — and two outside labs confirmed the molecules stick

Anthropic published results on 18 August from a protein design campaign in which Claude was not the designer of molecules so much as the operator of the machinery. It was given 15 protein targets and asked to produce 30 binders for each by driving the publicly available specialist design and co-folding models the field already uses. Two outside companies, Adaptyv Bio and Twist Bioscience, then made the designs and tested them in the lab. From 1,320 designs, 354 confirmed binders came back, covering 14 of the 15 targets, with hit rates of 35.1 percent in one configuration and 22.6 percent in another against an industry norm of 10 to 15 percent. Against one target, RBX1, Claude hit 40 percent where entrants in a public competition managed 3.7 percent. It failed cleanly on two targets, producing nothing at all against maltose-binding protein from 90 attempts. In a separate test, the model read raw NMR and mass spectrometry files and reached the lab's own conclusions, calling a sample 96.4 percent pure where the lab said 96.33. These are Anthropic's own results, not peer-reviewed work, though the wet-lab validation was done by named third parties.

✓ Verified · 2 sources

Related stories