Anthropic says its Claude AI model has designed 17 protein binders targeting dangerous scorpion neurotoxins during a largely autonomous research campaign that took just 64 hours. The experiment generated more than 318,000 computational protein designs, with 211 candidates eventually tested by external laboratories.
The findings, published by Anthropic on October 20, 2026, highlight how AI could accelerate early-stage drug discovery, particularly for neglected diseases that receive limited research funding.
However, the results do not mean scientists have developed a working scorpion antivenom. The 17 successful candidates met laboratory criteria for binding to individual toxins, but researchers have not yet demonstrated that these proteins neutralize venom in cells or protect animals from scorpion stings.
The project was conducted in collaboration with investment research firm Citrini Research, with Adaptyv Bio and Twist Bioscience carrying out laboratory testing.
Key takeaways
- 64 hours: The main AI-driven protein-design campaign took approximately 64 hours.
- 318,484 designs: Claude generated hundreds of thousands of computational protein candidates.
- 211 candidates tested: External laboratories evaluated 211 selected designs for toxin binding.
- 17 successful binders: Seventeen candidates met the study’s binding criteria, each targeting one specific neurotoxin.
- Four toxins covered: The successful candidates targeted four of the five neurotoxins selected for the experiment.
- No proven antivenom yet: Researchers have not established that the proteins neutralize venom in biological systems or work as treatments.
What did Anthropic’s Claude AI accomplish?
Anthropic’s experiment explored whether an AI model could help researchers design small proteins that attach to scorpion neurotoxins. These proteins, known as minibinders, are designed to recognize and attach to particular molecular targets.
The research team focused on five neurotoxins associated with scorpion species found in Mexico, Brazil, North Africa and the Middle East. These toxins can interfere with sodium channels in nerve cells, potentially triggering severe physiological reactions.
Rather than relying exclusively on conventional laboratory discovery methods, the researchers gave Claude a detailed prompt describing the targets, available computational tools and rules for evaluating candidate proteins.
Claude then worked through protein-design and structure-prediction workflows, selecting candidates for further assessment. Researchers guided the process through four follow-up prompts and supervised the overall work.
The model generated 318,484 designs and selected 180 candidates for testing. The research team subsequently added 31 more designs, bringing the total sent for laboratory evaluation to 211.
The experiment therefore demonstrated more than the ability to generate protein sequences. It tested whether an AI system could help coordinate a research workflow involving target analysis, candidate generation, computational screening and preparation for experimental validation.
By the numbers: Claude’s scorpion-toxin research
| Research metric | Reported result |
|---|---|
| Main design campaign | 64 hours |
| Computational protein designs generated | 318,484 |
| Designs selected by Claude for testing | 180 |
| Additional designs added by researchers | 31 |
| Total candidates tested | 211 |
| Candidates meeting binder criteria | 17 |
| Distinct neurotoxins with successful binders | 4 |
| Experimental testing partners | 2 |
| Computational design tools used | 12 |
Source: Anthropic’s research report.
The figures illustrate the difference between computational discovery and experimental success. Although Claude generated more than 318,000 designs, only a small selection advanced to physical testing, and 17 met the researchers’ criteria for binding.
Which scorpion toxins did the AI-designed proteins target?
The researchers selected five neurotoxins associated with four geographic regions. Their aim was to develop binders that could attach to the sites through which these toxins interact with sodium channels.
The five targets were Cn2, Ts1, AaH I, AaH II and LqhIII.
The experiment produced successful binders against four of these targets, but none of the candidates met the study’s criteria for binding to Ts1.
| Neurotoxin | Geographic association | Successful binders |
|---|---|---|
| Cn2 | Mexico | 7 |
| AaH II | North Africa | 4 |
| AaH I | North Africa | 4 |
| LqhIII | Middle East | 2 |
| Ts1 | Brazil | 0 |
| Total | Four toxins with successful binders | 17 |
Source: Anthropic’s published findings and associated research dataset.
Cn2 produced the largest number of successful candidates, accounting for seven of the 17 binders. Four candidates each met the criteria for AaH I and AaH II, while two were successful against LqhIII.
Importantly, these results do not demonstrate broad protection against all scorpion species. Each successful candidate met the criteria for binding to an individual toxin, and the team did not establish that any design could neutralize multiple toxins effectively.
How did Claude design the protein binders?
The campaign combined AI-assisted reasoning with existing computational biology tools. Claude was not simply asked to invent a protein and declare it successful. Its proposed designs went through several stages before researchers selected candidates for laboratory testing.
1. Selecting targets and identifying binding sites
Researchers identified five neurotoxins that can be dangerous to humans and used existing structural information to guide the design process.
Claude searched resources including the Protein Data Bank and the AlphaFold database to identify suitable toxin structures. It then selected specific amino acid positions, known as hotspots, that the proposed binders were intended to recognize.
The objective was to target regions involved in the toxins’ interactions with sodium channels.
2. Generating protein candidates
Claude used a collection of 12 design tools to generate and evaluate candidate proteins.
According to Anthropic, the model generated 318,484 designs while filtering out candidates that resembled toxins or known proteins. This step was intended to identify novel proteins suitable for further investigation.
Generating a candidate, however, does not establish that it will fold correctly, bind to its target or behave as intended in a biological system.
3. Screening predicted structures
The research process also included computational structure prediction and screening. Claude evaluated 553,052 design-toxin pairs, using predicted structures to assess whether candidate proteins might form suitable complexes with their targets.
The model applied quality thresholds and selected a smaller set of candidates for physical testing.
The full computational campaign used approximately 10,000 GPU-hours, according to Anthropic. The reported 64 hours refers to the elapsed time of the main design campaign, not the total computing resources consumed.
4. Testing candidates in external laboratories
The final stage involved physically producing the selected proteins and measuring whether they bound to their intended toxins.
Anthropic worked with Adaptyv Bio and Twist Bioscience to conduct the experiments. Their results provided a real-world check on the computational predictions.
This validation step is crucial because predicted protein structures and binding scores do not always translate into successful experimental results.
How did independent laboratories verify the results?
Adaptyv Bio and Twist Bioscience tested the 211 candidates using different experimental approaches to produce the proteins.
Adaptyv Bio used a cell-free protein-expression system, while Twist Bioscience expressed candidates in mammalian cells as fusions with an antibody component.
Both laboratories used surface plasmon resonance, or SPR, to assess binding. This technique measures changes associated with molecules attaching to and detaching from one another, allowing researchers to estimate binding affinity.
Anthropic applied strict criteria before classifying a candidate as a successful binder. The design had to show evidence of binding in the required experimental replicates, with measured dissociation constants of 1 micromolar or tighter. It also had to avoid a signal against an unrelated human protein used to screen for unwanted binding.
The results were encouraging but selective: 17 of the 211 tested candidates met the stated criteria.
Among the 60 candidates designed against individual toxins, 11 qualified as binders. The complete set also included candidates intended to target groups of related toxins or all five toxins simultaneously.
None of the multi-toxin designs met the study’s criteria for binding to more than one target. This is an important limitation because a broadly effective antivenom would ideally protect against several dangerous toxins without requiring a separate treatment for each one.
Why the research matters for scorpion antivenom development
Scorpion stings are a significant public-health problem in parts of the world where access to medical care is limited. Anthropic cites estimates of at least 1.2 million stings and more than 3,250 deaths annually.
Existing antivenoms are commonly produced by immunizing animals, such as horses, with venom and subsequently collecting and purifying antibodies. This process requires biological production infrastructure and can take months. Available treatments may also have limited coverage against particular toxins.
AI-designed proteins could eventually offer an alternative approach to developing treatments for venomous bites and stings.
A precisely designed binder might attach to a toxin and prevent it from interacting with its biological target. If researchers can establish that the protein reliably neutralizes the toxin, it could become a starting point for a future therapeutic.
Small, engineered proteins could also potentially be optimized for stability, manufacturing and distribution. Those properties could matter in regions where antivenom access is limited.
However, these are potential future benefits rather than demonstrated outcomes of the current study. The researchers have not shown that their 17 binders can safely treat scorpion envenomation in humans.
The work is best understood as an early step in a longer drug-development process.
Why the 17 successful binders are not yet an antivenom
The distinction between binding and neutralization is central to interpreting the findings.
A protein can attach to a toxin without necessarily preventing the toxin from causing harm. Even when a binder recognizes the right region, researchers must establish that the interaction is strong enough and biologically effective under relevant conditions.
Anthropic said the candidates were tested against recombinant toxins produced in laboratories. The research team had not yet confirmed that the strongest designs bind to toxins purified directly from venom, neutralize toxin activity in cells or protect living animals.
Several additional stages are therefore required before any candidate could be considered a potential treatment.
What researchers still need to establish
1. Confirm binding to natural venom toxins
The team needs to determine whether the strongest candidates recognize the relevant toxins in material derived from actual scorpion venom, not just laboratory-produced versions of the target proteins.
2. Test whether the binders neutralize the toxins
Cell-based experiments, including measurements of sodium-channel activity, could help establish whether a binder prevents a toxin from disrupting normal nerve-cell function.
3. Evaluate efficacy in animals
Appropriate preclinical studies would be needed to assess whether a candidate can protect against toxin exposure in a living organism. Binding measurements alone cannot establish this outcome.
4. Assess safety and drug properties
Researchers would need to investigate unwanted immune responses, toxicity, stability, dosage and other characteristics required for a potential medicine.
5. Establish manufacturing and clinical readiness
Even a promising candidate would require further development, manufacturing validation and clinical evaluation before it could be considered for use in people.
These stages can take substantial time and resources. AI may accelerate candidate discovery, but it does not eliminate the need for biological testing or clinical evidence.
Could AI accelerate drug discovery for neglected diseases?
The experiment addresses a major challenge in pharmaceutical research: finding promising candidates can be expensive, slow and technically demanding.
AI systems can help researchers analyze scientific literature, explore molecular structures, generate candidates and prioritize experiments. In principle, these capabilities can reduce the amount of manual work required to identify molecules worth testing.
This could be particularly useful for neglected diseases, where commercial incentives may be insufficient to support extensive traditional drug-development programmes.
Scorpion envenomation is one example. Treatments are needed in affected regions, but the economics of developing and distributing new medicines can make investment difficult.
The project also highlights a distinction between computational progress and the pace of biological experiments. Claude’s design work took 64 hours, but producing and testing proteins still required external laboratories.
In other words, faster design does not automatically mean a complete medicine can be developed in days. Experimental capacity, reproducibility, safety assessment, manufacturing and clinical trials remain essential.
The most significant long-term impact may come from combining AI systems with laboratories capable of testing more candidates efficiently and producing reliable experimental data.
What this means for Anthropic and AI-driven biotechnology
Anthropic’s experiment extends the company’s work beyond conventional software, writing and reasoning tasks into computational biology.
The study suggests that an AI model can coordinate multiple steps in a protein-design workflow and help produce candidates that survive external laboratory testing. The fact that the research involved a relatively short design campaign makes the result notable.
Still, the experiment does not establish that AI can independently develop complete medicines. Human researchers selected the research problem, defined the initial task, supervised the work and collaborated with specialist laboratories. The campaign also used substantial computing resources and existing protein-design tools.
The findings should therefore be viewed as evidence of AI-assisted research capability rather than proof that scientists can be removed from drug development.
Anthropic has made all 211 designs and their associated data available through its scorpion-toxin binder dataset. This gives other researchers an opportunity to examine the results, compare candidates and investigate why some designs worked while others did not.
The dataset may also help the scientific community study the limits of AI-designed proteins, including the difficulty of developing binders that recognize several different toxins.
What happens next?
The next phase is likely to focus on the strongest individual candidates, especially Cn2-B1 and Cn2-B2, which Anthropic reported as having binding affinities in the single-digit nanomolar range against the Cn2 toxin.
The researchers said these candidates appear to target a region that overlaps with, or is close to, the binding site of an antibody previously reported to neutralize Cn2 in mice. That provides a reason to investigate whether the new proteins could also neutralize the toxin.
However, this remains a hypothesis rather than a demonstrated therapeutic effect. The team has not yet confirmed the candidates’ activity in cellular experiments or live animals.
Further work will need to establish whether the proteins block toxin activity, remain stable under relevant conditions and can be developed into safe and affordable treatments.
If these hurdles can be overcome, AI-designed binders could become useful starting points for future antivenom research. If not, the experiment will still provide valuable evidence about how computational design performs when tested against real biological targets.
Frequently asked questions
1. What did Claude AI design in the scorpion-toxin experiment?
Claude helped design small proteins called minibinders that attach to selected scorpion neurotoxins. Of the 211 candidates tested in laboratories, 17 met Anthropic’s criteria for binding to an individual toxin.
2. How long did Claude take to design the protein candidates?
Anthropic reported that the main design campaign took 64 hours. The workflow generated 318,484 computational designs before researchers selected candidates for experimental testing.
3. Has Claude developed a working scorpion antivenom?
No. The experiment demonstrated laboratory binding, not a proven treatment. Researchers still need to establish whether the proteins neutralize toxins in cells and protect living animals before considering further therapeutic development.
4. Which scorpion toxins did the successful binders target?
The 17 successful candidates targeted four toxins: Cn2, AaH I, AaH II and LqhIII. No candidate met the study’s criteria for binding to Ts1.
5. Which laboratories tested the AI-designed proteins?
Adaptyv Bio and Twist Bioscience produced and tested the candidates. They used complementary laboratory methods and surface plasmon resonance to evaluate binding.
6. Why is AI-designed protein research important?
AI could help scientists explore more molecular candidates and prioritize promising designs for experiments. This may speed up early-stage discovery, including for neglected diseases where research funding and treatment options are limited.
7. What are the biggest limitations of the experiment?
The successful candidates have not yet been shown to neutralize venom in cellular tests or protect animals. The study also did not produce a binder that met the criteria for targeting multiple toxins, a potentially important feature of a broadly effective antivenom.
Featured image direction
Image concept: A scientifically grounded editorial illustration showing an AI-assisted protein-design workflow alongside a stylized scorpion neurotoxin and a folded protein minibinder. Use a clean biotechnology research aesthetic with molecular structures, a subtle computational interface and a laboratory-testing visual. Avoid suggesting that a finished antivenom or approved treatment already exists.
Image-generation prompt: Create a professional 16:9 editorial illustration for a technology and biotechnology news article. Show a stylized scorpion silhouette on one side, a detailed but illustrative protein neurotoxin structure in the center, and a small engineered protein binder approaching its molecular target. Include subtle AI-computing and laboratory-research motifs in the background. Use realistic scientific visualization, polished newsroom art direction, crisp molecular details and restrained blue, teal and neutral tones. No text, no logos, no human medical claims, no syringe implying an approved treatment, and no depiction of a completed antivenom.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
