Anthropic’s model and investment firm Citrini Research designed proteins against scorpion toxins in 64 hours; 17 of the 211 proteins tested bound a toxin

Anthropic’s model and investment firm Citrini Research designed proteins against scorpion toxins in 64 hours; 17 of the 211 proteins tested bound a toxin

On October 20, Anthropic published a technical report and open data from its project with Citrini.

Scorpions sting an estimated 1,2 million people each year, and more than 3 250 die. Dangerous scorpion toxins bind to sodium channels in nerve cells, disrupting nerve signaling and, in severe cases, causing heart failure. Antivenom made from horse antibodies targets only a limited set of toxins. A protein that occupied a toxin’s channel-binding site could prevent it from binding to the channel.

Citrini’s five-person team, including three members without scientific training, set out to test AI’s capabilities in biology for themselves and allocated 200 000 dollars. They needed a task that depended on molecular design and could be tested quickly in vitro. Antivenom development fit those requirements. David Baker’s group had already designed proteins against snake toxins that protected mice, so Citrini turned to scorpions.

Anthropic and Citrini gave Claude one prompt and four follow-up instructions. The humans selected five toxins and a channel-binding site on each. The model used 12 design programs, generated 318 484 protein candidates, and selected 180; the humans added another 31 with its help.

Two laboratories, Adaptyv Bio and Twist Bioscience, tested whether the proteins bound to laboratory-produced toxins. The authors defined a binder as a protein that bound to the toxin in every measurement at every laboratory that tested it and did not bind to one unrelated protein. They set these criteria after receiving the data. There were 17 such binders: seven against Cn2 from a Mexican scorpion, ten against three other toxins, and none against the Brazilian toxin Ts1.

To investigate where the seven proteins bound to Cn2, the researchers tested their competition with LR, an antibody fragment that had protected mice against venom. When LR was already bound to the toxin, six of the seven proteins lost 87–95% of their binding signal. The LR binding site overlaps the toxin’s channel-binding site, and the authors concluded that the proteins occupy the same site or an adjacent one. They expect Cn2-B1, the strongest binder of the seven, to neutralize Cn2 as LR does. The authors use antibodies as a benchmark: LR’s binding affinity for Cn2 is 0,11 nM (a lower value means stronger binding), while the two laboratories measured values of no more than 3,1 and 1,7 nM for Cn2-B1.

Of the 60 proteins each designed to target a single toxin, 11 (18%) qualified as binders. The authors describe this proportion as typical for such design campaigns, including the August campaign in which Claude designed 1 320 proteins, each for a single target. Citrini added the goal of designing one molecule against all five toxins as a “moonshot.” Such a molecule would simplify a cocktail of several molecules, but although the toxins share a common shape, their pairwise amino acid sequence identity is only 26–62%. The remaining six of the 17 binders came from the 151 proteins designed to target multiple toxins, although each bound only one toxin.

The list of 180 proteins was ready after 64 hours; laboratory measurements took about two months. The authors want to test whether a designed protein binds to toxin obtained from venom, blocks the toxin’s effects in cells, and protects mice. Anascorp, the only scorpion antivenom approved in the United States, underwent testing under US regulatory requirements for about 12 years. Susana Vázquez Torres, an author of the snake-toxin study and an external expert, said:

Very strong binders can now be produced surprisingly quickly, and binding itself is no longer the main bottleneck.

Report coauthor Nathaniel Horwitz of Hunterbrook believes that specialists using existing software could have done the work without Claude, but would have needed more time and money.