Claude Scorpion Binders Remap AI Drug Discovery

Animal-immunization antivenoms are the patient-facing products when supply exists; engineered antibodies and nanobodies are the research affinity bar, with mouse protection data on a few reported molecules. On October 20, 2026, Anthropic's research post put Claude on that same map with a supervised, largely autonomous minibinder campaign against five toxins, run with Citrini Research and wet-lab partners Adaptyv Bio and Twist Bioscience.

318,484 designs over 64 hours

Claude Science started from a binder prompt Anthropic had used in an August study, then revised with Citrini for scorpion targets. Four follow-ups steered the run. The model picked structures from the Protein Data Bank and AlphaFold database, chose five hotspots per toxin, set quality thresholds, weighted twelve design tools, and built single-toxin plus multitoxin strategies without a human rewriting every loop.

It filtered designs that looked like toxins or known proteins, co-folded and screened 553,052 design-toxin pairs, and returned 180 sequences for synthesis. Anthropic later added 31 more with Claude's help. Footnote two on the research post puts campaign compute on the order of 10,000 GPU hours. Separately, it says Anthropic has made these models about 4x faster on average, and that a simplified prompt using a single agent and a single GPU over 24 hours reaches comparable in silico scores.

Targets spanned Mexico (Cn2), Brazil (Ts1), North Africa (AaH I and AaH II), and the Middle East (LqhIII), aimed at channel-binding sites on sodium-channel toxins. Pairwise sequence identity between toxins ran only 26% to 62%, which made the multitoxin ask deliberately hard.

Under Anthropic's supervision, Claude analyzed the research data, generated the figures, and drafted the final report for this campaign.

1.2 million stings each year

Anthropic's post cites more than 3,250 deaths, mostly among children where care is scarce. New treatments are expensive to make and unlikely to earn development costs back, so envenomation sits in the neglected-tropical-disease bucket.

Standard antivenoms still come from milking venom, immunizing horses, and purifying antibodies. That pipeline can take months. Coverage is usually narrow across toxin panels. Supply is thin and costly in the regions that need it most.

De novo binders are attractive on paper because they can be designed against a chosen epitope and tuned for stability and potency before anyone scales fermentation. Anthropic points to earlier snake-neurotoxin de novo work that protected mice, and says it knows of no prior de novo scorpion-neurotoxin binder papers.

Anthropic has not confirmed neutralization in a cellular environment, and it does not claim to have produced an antivenom from this run.

Sub-1 nM antibody affinity bar

Anthropic treats antibodies and nanobodies as the bar de novo designs must clear. Three such molecules already report affinities below one nanomolar against Cn2 or AaH II, with mouse protection data attached.

Claude's strongest hits, Cn2-B1 and Cn2-B2, against the Mexican toxin Centruroides noxius Cn2, reached single-digit nanomolar KD at both Adaptyv Bio and Twist Bioscience. Competition experiments suggest they attach on or near the same epitope as antibody LR, which covers the sodium-channel latch and has been reported to neutralize Cn2 in mice. Anthropic has not run those two designs in cells or animals, and it stops short of calling them neutralizing agents.

Designing credible binders against five toxins in parallel, in a matter of days, is a great showcase of the AI-driven paradigm shift we are seeing in therapeutic discovery.

That line is from Timothy P. Jenkins (Technical University of Denmark), one of two outside readers Anthropic quotes. Susana Vázquez Torres (CNIO, Madrid) and Jenkins both said they felt this work was a promising first step. Anthropic's post then notes that if binders can be confirmed in the lab within weeks, discovery for envenomation and other neglected diseases could, in theory, speed up. Neither endorsement substitutes for in vivo efficacy.

Antibodies still hold the only complete affinity-plus-protection story in this comparison. Claude's best designs look epitope-plausible and tight on recombinant toxin, not clinical.

17 binders from 211 designs

Adaptyv Bio expressed designs in a cell-free bacterial system. Twist Bioscience expressed them in human kidney cells as Fc fusions. Both read binding by surface plasmon resonance. Anthropic required a hit in every replicate at both labs (LqhIII only at Adaptyv), every KD at or below 1 µM, and no signal against CLEC12A, an unrelated human protein used as a stickiness control.

Every design that cleared those rules bound a single toxin: seven against Cn2, four against AaH II, four against AaH I, two against LqhIII. None of the multitoxin shortlists met criteria for more than one toxin. Among the sixty single-toxin designs, eleven passed, an 18% rate Anthropic says matches typical protein-design campaigns, including its earlier binder work.

The shortlist mix was sixty single-toxin designs (twelve per toxin), thirty β-family (Ts1 and Cn2), thirty α-family (AaH II, AaH I, LqhIII), and ninety-one against all five. Claude emailed sequences to Twist in messages marked as written by Claude. Anthropic is releasing all sequences, predicted structures, scores, and experimental traces.

The dual-CRO filter is doing the scientific gatekeeping. It correctly refuses to promote a universal cocktail molecule from this run.

Neutralization and patient access gap

Recombinant SPR hits shorten the early discovery loop. They still leave the middle of the pipeline untouched.

Anthropic lists the two biggest limits plainly. No design was tested for neutralization in vitro. None was tested for efficacy in vivo. Binders saw recombinant toxins, not venom-purified protein. Advancing Cn2-B1 would still require venom binding, patch-clamp proof that the toxin cannot open the channel, mouse protection, then safety, stability, cost, and distribution work with clinicians, funders, and policymakers.

None of the actors in this release closes that path. The empty slot is still venom-purified binding, patch-clamp neutralization, and mouse protection for a design such as Cn2-B1 before anyone can talk about a stocked product.