Claude designed protein binders against scorpion neurotoxins

What it claims

Anthropic ran an internal Claude research model in Claude Science to design de novo protein minibinders against five scorpion neurotoxins, aiming each design at the toxin's sodium-channel-binding site. Of 211 designs lab-tested, 17 met the binding criteria, each against a single toxin. The two strongest, Cn2-B1 and Cn2-B2, bound the Mexican scorpion toxin Cn2 at single-digit nanomolar affinity, and competition experiments suggest they bind on or near the same site as LR, an antibody reported to neutralize Cn2 in mice. The project started from a question posed by Citrini Research: whether AI drug discovery could let motivated nonexperts tackle a real health problem. Scorpion envenomation is a neglected tropical disease, with 1.2 million or more stings and over 3,250 deaths a year, mostly children.

How it works

The five targets span four regions: Cn2 (Mexico), Ts1 (Brazil), AaH I and AaH II (North Africa), and LqhIII (Middle East). They share one fold but are only 26% to 62% identical in sequence, chosen as a hard test of multitoxin design. Claude picked toxin structures from the PDB and AlphaFold database, selected five hotspot amino-acid positions per toxin, set its own quality thresholds, chose how much to use each of 12 design tools, and developed the multitoxin strategies. It generated 318,484 designs, filtered out any resembling a known protein or toxin, and co-folded 553,052 design-toxin pairs. The campaign used on the order of 10,000 GPU hours. The 211 total breaks down as 60 single-toxin (12 each), 30 β-family, 30 α-family, and 91 against all five.

Evidence

Both CROs measured binding by surface plasmon resonance. Adaptyv Bio used a cell-free bacterial expression system; Twist Bioscience expressed the designs in human kidney cells as Fc fusions. The binder criteria were strict: a hit at both labs, binding in every replicate, every KD at 1 µM or tighter, and no signal against the unrelated human protein CLEC12A (an off-target "stickiness" control). Among the 60 single-toxin designs, 11 were binders, an 18% rate the authors call consistent with typical campaigns including their own prior work. The 17 binders: seven against Cn2, four against AaH II, four against AaH I, two against LqhIII.

Limits

No multitoxin design bound more than one toxin; the authors list possible causes including prompting, tool limits, and the biological feasibility of the goal itself. Results show binding only against recombinant toxins grown from genes, not toxin purified from venom. Nothing was tested for neutralization in vitro (cells) or efficacy in vivo (mice), so Anthropic makes no antivenom claim. Advancing Cn2-B1 would require confirming it binds venom-purified toxin, patch-clamp functional tests, and mouse protection studies, plus safety, stability, and affordability. Two outside experts, Susana Vázquez Torres and Timothy P. Jenkins, called the work a promising first step; that is their stated opinion, not a measured result. All 211 designs, predicted structures, scores, and binding data are released alongside a technical report.