Anthropic has demonstrated how artificial intelligence could help accelerate research into treatments for scorpion venom, using its Claude Science research system to design small proteins that bind to dangerous neurotoxins. In a study published on October 20, 2026, the company reported that Claude generated 318,484 candidate protein designs and helped select 211 for laboratory testing. Of those tested, 17 met the researchers’ criteria for binding to specific scorpion toxins.
The experiment targeted five neurotoxins from four scorpion species found in Mexico, Brazil, North Africa and the Middle East. Two independent research organisations, Adaptyv Bio and Twist Bioscience, tested the candidate proteins. Although the results offer a promising starting point for developing new antivenom treatments, the researchers have not yet demonstrated that the designed proteins neutralise venom in cells or protect animals. The project therefore represents an advance in early-stage drug discovery, rather than a finished treatment for scorpion stings.
Key Takeaways
- Claude designed more than 318,000 protein candidates: The system worked largely autonomously for 64 hours, with researchers providing four follow-up instructions.
- 211 designs underwent laboratory testing: Researchers identified 17 proteins that met the study’s binding criteria.
- The strongest candidates targeted Cn2: Two designed proteins bound tightly to a neurotoxin from the Mexican scorpion Centruroides noxius.
- The research involved two independent laboratories: Adaptyv Bio and Twist Bioscience experimentally tested the designs.
- No universal antivenom was created: None of the tested designs met the criteria for binding to multiple toxins.
- The next step is biological validation: Researchers still need to test whether the proteins neutralise toxins and can be developed into safe, effective treatments.
What Is Claude Science, and How Did It Design the Proteins?
Claude Science is Anthropic’s AI research environment, designed to help scientists perform multistep investigations using scientific databases, computational tools and specialised software.
For this project, Anthropic used an internal research model operating within Claude Science. The system was given an initial prompt defining the target toxins, available protein-design tools and requirements for evaluating candidate proteins.
Claude then searched scientific databases for relevant toxin structures, selected target regions and generated protein designs intended to attach to those regions. It used computational tools to predict how candidate proteins might fold and interact with their targets before filtering the results for laboratory testing.
The system generated 318,484 designs and selected 180 candidates. Researchers subsequently added 31 more, bringing the total submitted for experimental testing to 211.
Anthropic reported that the computational campaign used approximately 10,000 GPU-hours. Claude worked autonomously for an initial 24 hours, after which researchers supplied additional instructions to guide the search, particularly toward proteins that might recognise multiple toxins.
The overall design campaign took 64 hours. Laboratory production and testing took longer, highlighting an important distinction between generating a promising molecular design and demonstrating that it works in a biological system.
Why Scorpion Venom Is a Difficult Medical Problem
Scorpion stings are a significant health problem in several regions, particularly where access to emergency medical care is limited. Anthropic cited estimates of at least 1.2 million stings and more than 3,250 deaths annually.
Certain scorpion venoms contain neurotoxins that interfere with sodium channels, proteins involved in transmitting electrical signals through nerve cells. Disrupting these channels can trigger abnormal nerve activity and severe physiological responses, including dangerous cardiovascular complications.
Existing antivenoms are commonly produced by immunising animals, such as horses, with venom and then extracting antibodies from their blood. These products can be effective, but manufacturing is complex, and treatments may have limited coverage against particular species or toxins.
The economics also present a challenge. Scorpion envenomation disproportionately affects populations in regions where healthcare resources may be constrained, limiting commercial incentives to develop new treatments.
AI-assisted protein design could potentially help researchers develop smaller, more targeted molecules that are easier to manufacture. However, affordability, stability, delivery and clinical effectiveness would all need to be demonstrated before such proteins could become viable treatments.
What Are Protein Binders?
A protein binder is a molecule designed to attach to a particular target. Antibodies are familiar examples of proteins that bind to specific molecules, including toxins and disease-related proteins.
In this study, Claude designed small proteins known as minibinders. Unlike conventional antibodies, which are produced by the immune system and typically have a more complex structure, these candidates were designed computationally from scratch.
The objective was to create proteins that attach to a region of each neurotoxin involved in interacting with sodium channels. If a binder can occupy the relevant region strongly enough, it may prevent the toxin from interacting with its normal target.
That outcome, however, cannot be assumed simply because a protein binds to a toxin. The binding location, strength, stability and behaviour in a biological environment all matter.
The researchers therefore tested whether the designed molecules attached to their intended targets and whether they also showed unwanted binding to an unrelated human protein used as a control.
What the Laboratory Tests Found
The study targeted five neurotoxins: Cn2, Ts1, AaH I, AaH II and LqhIII. These represent toxins from four scorpion species and different toxin families.
The candidate proteins were produced and tested by Adaptyv Bio and Twist Bioscience using surface plasmon resonance, a laboratory technique that measures interactions between molecules.
Anthropic applied strict criteria. A design generally had to show consistent binding in both laboratories, achieve a dissociation constant of 1 micromolar or tighter, and avoid unwanted binding to the control protein.
The results were as follows:
| Research result | Outcome |
|---|---|
| Computational protein designs generated | 318,484 |
| Initial candidates selected by Claude | 180 |
| Additional candidates added by researchers | 31 |
| Total designs tested in laboratories | 211 |
| Designs meeting the binder criteria | 17 |
| Binders targeting Cn2 | 7 |
| Binders targeting AaH II | 4 |
| Binders targeting AaH I | 4 |
| Binders targeting LqhIII | 2 |
| Binders targeting Ts1 | 0 |
| Designs meeting criteria for multiple toxins | 0 |
Source: Anthropic’s technical report and publicly released experimental dataset.
The results show that the system could produce experimentally confirmed binders against several toxins. However, the 17 successful candidates were not distributed evenly across all targets, and none met the study’s criteria for binding multiple toxins.
That limitation is important because a broad-spectrum antivenom would ideally neutralise several dangerous toxins without requiring a separate treatment for each one.
Why the Cn2 Protein Designs Are Particularly Interesting
Among the strongest candidates were Cn2-B1 and Cn2-B2, designed to bind Cn2, a neurotoxin from the Mexican scorpion Centruroides noxius.
Both candidates showed binding affinities in the single-digit nanomolar range in the reported tests. A lower dissociation constant indicates stronger binding, making these results encouraging for further investigation.
The researchers also performed competition experiments involving an antibody known as LR, which has previously been reported to neutralise Cn2 in mice.
The experiments suggested that Cn2-B1 and Cn2-B2 bind at or near the same region of the toxin as the LR antibody. That region is involved in the toxin’s interaction with sodium channels.
This finding provides a reason to investigate whether the designed proteins might block the toxin’s activity. But it remains an inference, not a demonstrated therapeutic effect.
Anthropic explicitly stated that it had not tested the strongest binders for neutralisation in a cellular environment or for protective effects in live animals.
Why This Is Not Yet an Antivenom
The study demonstrates that AI can help design proteins that bind to selected scorpion toxins. Developing an antivenom requires several additional steps.
First, researchers need to confirm that the binders recognise toxins in their naturally occurring form, including toxins purified directly from venom. Recombinant proteins used in laboratory experiments may not reproduce every property of naturally sourced material.
Second, scientists must establish whether the binders actually prevent toxins from disrupting sodium channels. Cellular experiments, including patch-clamp recordings, can help test this question.
Third, promising candidates need to undergo animal studies to determine whether they protect against venom exposure, followed by appropriate safety assessments and clinical development.
Researchers must also consider how the proteins behave in the body, how they can be manufactured consistently, and whether they remain stable under the storage and transport conditions required in affected regions.
A molecule that binds strongly in a laboratory may still fail at one or more of these stages.
What the Study Means for AI Drug Discovery
The experiment illustrates how AI systems can automate parts of a complex scientific workflow. Instead of asking a model only to summarise published research, the team gave Claude access to tools for molecular design, structure prediction, candidate filtering and data analysis.
This approach allowed the system to explore a very large design space and identify candidates for physical testing in a relatively short computational period.
The work also demonstrates the importance of external validation. The candidates were not accepted simply because the model predicted that they would bind. They were synthesised and tested in two laboratories using experimental measurements.
Anthropic has released all 211 tested designs and their associated data through a public research dataset. This could help other researchers reproduce the results, examine unsuccessful candidates and develop improved approaches.
Nevertheless, the experiment does not show that AI can independently create a finished drug. It shows that AI can contribute to the discovery stage, while laboratory work and subsequent biological testing remain essential.
The Bigger Picture
Scorpion envenomation is an example of a medical problem that can receive limited research investment despite causing serious harm. AI-assisted protein design could help reduce the time and specialist effort required to generate candidate treatments for such neglected conditions.
The wider significance lies in the combination of computational design and experimental validation. Claude generated hundreds of thousands of possibilities, but only laboratory testing established which candidates met the study’s binding requirements. The next challenge is to determine whether those candidates can produce meaningful protection in biological systems.
If future studies demonstrate effective neutralisation, safety and practical manufacturing, AI-designed proteins could become useful components of new antivenom treatments. For now, the results are an early scientific milestone rather than evidence that patients can be treated with a new Claude-designed medicine.
FAQs
1. What did Claude Science achieve in the scorpion venom study?
Claude helped generate 318,484 protein designs and select candidates for testing. Of 211 designs tested by two research laboratories, 17 met the study’s criteria for binding to specific scorpion neurotoxins.
2. Has Anthropic created a new scorpion antivenom?
No. The study confirmed that some designed proteins bind to toxins, but it did not establish that they neutralise venom in cells or protect animals. Further research is required before any treatment could be considered for clinical use.
3. Which scorpion toxins did the researchers target?
The project targeted Cn2, Ts1, AaH I, AaH II and LqhIII, representing five neurotoxins from four scorpion species.
4. Why are the Cn2-B1 and Cn2-B2 proteins important?
They showed binding affinities in the single-digit nanomolar range and appeared to bind near the same toxin region as an antibody previously reported to neutralise Cn2 in mice. Their ability to neutralise the toxin has not yet been demonstrated.
5. Why does the study matter for AI drug discovery?
It demonstrates that an AI research system can help generate and filter large numbers of molecular designs, with experimental testing used to validate the results. This could accelerate early-stage research, although biological validation and clinical development remain necessary.
Looking Ahead
The next stage of the research will focus on whether the strongest binders can neutralise their target toxins under biological conditions. For the leading Cn2 candidates, researchers would need to test interactions with naturally sourced toxin, evaluate effects on sodium-channel activity and determine whether the proteins provide protection in appropriate animal studies.
The broader lesson is that AI can make the search for candidate medicines faster, but it cannot remove the need to prove that a molecule works safely in the body. If researchers can translate these early binding results into effective and affordable treatments, the approach could eventually support new therapies for scorpion stings and other neglected diseases.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
