Anthropic's Claude AI successfully designs functional proteins

Anthropic has demonstrated that its Claude AI model can autonomously design functional protein binders using open-source software. In a study involving 1,320 protein candidates, the model achieved a 27% success rate, with 354 candidates successfully binding to their targets. This success rate significantly exceeds the typical industry standard of 10-15%.

The designs were validated through independent laboratory testing by Adaptyv Bio and Twist Bioscience. Claude successfully produced binders for 14 of the 15 targets tested. In one instance, Claude's top design bound approximately 10 times more tightly than the winner of a human open design competition.

Beyond design, the model demonstrated the ability to process raw scientific data, such as NMR and LC-MS, with high accuracy and speed. Due to the dual-use nature of protein design—which can be applied to both medicine and harmful biosynthesis—Anthropic is developing a vetting process and an access program for scientists rather than releasing the capability as a standard feature.