An AI ran a whole protein-design campaign on its own

Anthropic published a research result this week where an AI model ran almost the whole protein-design campaign itself.

The goal was small proteins that bind scorpion-sting toxins, the starting point for a possible future antivenom. The model drove 12 existing design tools on its own, generated candidates, folded them against the target, and chose which ones to send to the lab.

The scale it worked at: 318,484 designs generated, 553,052 design-and-target pairs folded and screened. From that it picked 211 to make and test. 17 met the binding criteria. Start to finished list took about 64 hours and roughly 10,000 hours of graphics-processor time, with the model working on its own for the first 24.

The team is careful about what this is not. It's not an antivenom. Binding was measured only against lab-made versions of the toxins, not venom itself, and nothing was tested for whether it actually neutralizes a sting in cells or animals.

Here's the arc. Two years ago the Nobel in chemistry went to computational protein design, the idea that you can design a working protein on a computer. That was about the method existing. This is about the method running itself: an agent that plans the campaign, spends the compute, reads its own results and narrows more than 500,000 candidates to a shortlist a human can test. They open-sourced all 211 designs.

The design step is becoming something you can point at a problem and leave running.

Read the primary source →