Anthropic has reported a closed-loop protein-design campaign in which Claude generated 318,484 candidate minibinders in 64 hours for five scorpion neurotoxins. Of 211 designs evaluated in laboratory testing, 17 met predefined binding criteria. The results are early-stage evidence of toxin binding, not evidence of a working antivenom, but they demonstrate how generative AI can compress the design phase of a biotech research program while retaining experimental validation as the decisive step.
The campaign, conducted with Citrini Research, targeted the cell-binding portions of five toxins: Cn2, Ts1, AaH I, AaH II and LqhIII. Anthropic details the workflow, data release and experimental results in its official report on Claude-designed scorpion toxin binders. The work involved two contract research organizations, Adaptyv Bio and Twist Bioscience, which tested candidates using distinct expression approaches.
From a large design set to lab-confirmed binders
The project used Claude as part of an iterative design process rather than treating model output as a final biological result. The model generated a large candidate pool, from which 180 top designs were selected. Researchers later added 31 further candidates, bringing the laboratory-tested total to 211.
Each candidate faced a deliberately demanding definition of success. It had to bind its intended recombinant toxin in testing at both laboratories, show a dissociation constant, or KD, of 1 micromolar or better, and show no off-target binding to CLEC12A. A lower KD indicates tighter binding under the relevant assay conditions. Seventeen designs cleared those criteria for individual toxins.
| Campaign stage | Result | What it establishes |
|---|---|---|
| AI design run | 318,484 designs generated in 64 hours | High-throughput candidate generation |
| Laboratory evaluation | 211 designs tested | Experimental assessment across two CROs |
| Predefined binding hits | 17 binders | Binding to individual recombinant toxins under the stated criteria |
This funnel matters because protein-design systems can produce many plausible candidates, while biology determines which candidates are useful. Anthropic has released the full set of 211 designs with their experimental data, as well as the technical report. That level of disclosure gives researchers an opportunity to inspect the relationship between computational selection and measured outcomes rather than evaluating the project only through a headline result.
What the strongest Cn2 results show
Two Cn2-targeting designs, Cn2-B1 and Cn2-B2, achieved binding in the single-digit nanomolar KD range. Anthropic also reports that they competed for the same epitope as a known toxin-neutralizing antibody. That is a meaningful experimental signal because it suggests the binders interact with a biologically relevant region of the toxin.
However, this distinction is essential: binding is not neutralization. The reported assays used recombinant toxins. They do not establish that the designs neutralize toxin effects in cells or in living organisms, and Anthropic does not present them as an antivenom. Further functional experiments, safety work and regulatory review would be required before any therapeutic application could be considered.
Why the workflow is relevant beyond antivenom research
The immediate result concerns scorpion toxins, but the business and research significance is the workflow. It joins large-scale model generation with external experimental testing, then evaluates results against explicit thresholds. For biotech teams, this can shift AI from a general research assistant to a way of expanding the set of molecules that reach a lab.
Potential practical advantages include:
- Faster early-stage exploration, because teams can screen a far larger design space before commissioning experiments.
- More focused lab spending, because synthesis and assay capacity can be directed toward prioritized candidates rather than a manually assembled list.
- Accessible specialist capacity, because CRO partners can provide expression and testing capabilities that many research organizations do not maintain internally.
- Clearer learning loops, when computational designs and assay results are captured together for subsequent analysis.
None of those advantages removes the cost, time or uncertainty of wet-lab validation. The 17 successful binders emerged from 211 tested designs, illustrating that model-generated candidates still need carefully designed assays, negative controls and independent confirmation. The use of two expression modalities, cell-free and Fc-tagged mammalian expression, also highlights a practical point: results can depend on how a protein is produced and measured.
For smaller biotech companies and lab partners, the more realistic opportunity is not to replace discovery science with a chatbot. It is to use AI-assisted design to increase the quality and breadth of hypotheses entering an existing experimental process. Teams still need target selection, assay design, biosafety practices, data interpretation and decisions about which findings merit additional investment.
The next tests matter more than the design count
Anthropic frames the campaign as an initial step toward potential therapeutic development. The next scientific questions are therefore functional. Can the binders neutralize the toxins in relevant cellular assays? Do they retain appropriate behavior outside the recombinant binding setup? Can promising candidates be developed safely and consistently?
The project also concerns individual toxin binders, not a single binder that covers all five toxins. That distinction matters for any eventual antivenom strategy, where breadth of coverage and real-world neutralization would be central development questions. The campaign's strongest contribution so far is a documented example of AI-assisted molecular design reaching lab-backed binding results at substantial scale.
For biotech teams considering AI-assisted R&D, the value lies in connecting model output to a repeatable experimental workflow, not simply producing more candidates. Scalevise can help assess where AI fits into research operations, select practical integration paths and turn disconnected tools into an implementation plan through AI consultancy for practical adoption. Request a consultation to map an AI workflow that supports your team's real validation process.
Frequently Asked Questions
What did Claude design in Anthropic's scorpion toxin project?
Claude generated candidate minibinder proteins intended to bind five scorpion neurotoxins: Cn2, Ts1, AaH I, AaH II and LqhIII.
How many Claude-generated designs were tested in the lab?
Two CROs tested 211 designs. Anthropic reports that 17 met its criteria for binding an individual toxin in both laboratories, with no off-target CLEC12A binding.
Do the reported binders prove that Claude created an antivenom?
No. The results show binding to recombinant toxins. They do not establish toxin neutralization in cells or in vivo, and Anthropic does not claim to have created an antivenom.
Why were two laboratories involved in the testing?
Adaptyv Bio and Twist Bioscience tested designs through cell-free and Fc-tagged mammalian expression modalities. A successful candidate had to meet the stated binding criteria in both labs.
What does the project mean for biotech teams?
It shows a potential workflow in which AI generates and prioritizes many protein candidates, while laboratory testing determines which candidates have measurable promise. The approach still requires rigorous assays and further development work.
Conclusion
Anthropic's scorpion toxin campaign provides a concrete, lab-backed example of Claude supporting high-throughput protein design. Generating 318,484 candidates quickly is notable, but the 17 validated binders and the public experimental data are the more important result. The findings support further research into the selected proteins while making clear that binding evidence is only an early step toward any therapeutic use.
