Anthropic’s Claude has demonstrated that artificial intelligence can help design candidate proteins for a real medical problem, but a high-profile experiment involving scorpion venom has exposed the considerable distance between generating promising molecules and creating an actual medicine. In collaboration with investment research firm Citrini Research, Anthropic tested whether an AI system could help a team with limited drug-development experience design proteins that might eventually form the basis of a new scorpion antivenom. The experiment produced hundreds of thousands of computational designs and dozens of proteins that bound to their intended targets in laboratory tests. However, none has been shown to neutralize scorpion venom, protect an animal from its effects or qualify as a working treatment.
The project began with an ambitious question: could a relatively small team use AI to tackle a neglected medical problem that conventional pharmaceutical companies have limited financial incentive to pursue? Scorpion stings affect an estimated 1.2 million people or more annually, and thousands of people die from scorpion envenomation each year, with children and communities with limited access to medical care particularly vulnerable. Existing antivenoms can be expensive to produce, may target only certain toxins and can be difficult to obtain in the regions where they are needed most. Citrini and Anthropic wanted to explore whether computational protein design could help create a more flexible approach to treatment.
The work focused on protein minibinders, small engineered proteins designed to attach to specific biological targets. In principle, a minibinder could be designed to attach to a critical region of a scorpion neurotoxin and prevent it from interacting with the proteins responsible for transmitting electrical signals in nerve cells. If that interaction could be blocked effectively, the molecule might help prevent the dangerous effects of the venom. But this is only a proposed mechanism: a protein that attaches to a toxin is not necessarily capable of stopping the toxin from causing harm.
Anthropic used an internal research version of Claude to conduct the design campaign over approximately 64 hours, with human researchers directing the work through an initial prompt and four follow-up instructions. The model coordinated multiple specialist computational tools, evaluated potential structures and generated an enormous pool of candidates. Across the campaign, Claude produced 318,484 protein designs and screened hundreds of thousands of predicted protein-toxin combinations before selecting candidates for physical testing. Human researchers contributed additional designs, bringing the final set submitted for laboratory evaluation to 211.
The experiment reportedly cost Citrini approximately $200,000, including roughly $50,000 in computing expenses and nearly $150,000 in laboratory work, according to coverage of the project. That figure illustrates an important distinction in AI-assisted drug discovery: generating ideas computationally may become faster and cheaper, but confirming whether those ideas work requires physical materials, specialized equipment and scientific expertise. The overall economic picture is also incomplete if it excludes the value of Anthropic's staff time, access to its internal model and other support provided to the project.
The candidates were sent to Adaptyv Bio and Twist Bioscience, two organizations that performed laboratory testing. Researchers evaluated whether the engineered proteins physically bound to selected scorpion neurotoxins, using methods that measure the interaction between molecules. To qualify as a binder under the project's criteria, a candidate generally had to demonstrate sufficiently strong and repeatable binding while avoiding unwanted interaction with an unrelated control protein. Seventeen of the 211 designs met the stated criteria against at least one target toxin.
Although 17 successful binders may sound like a promising result, the details are crucial. The targets included five neurotoxins associated with scorpion species or regions in Mexico, Brazil, North Africa and the Middle East. The successful designs bound to individual toxins, but none met the project's criteria for binding to multiple toxins. That fell short of the ambition to create a broadly useful antivenom that could address several dangerous scorpion toxins with fewer engineered proteins. A treatment with broad coverage could be especially valuable in settings where identifying the precise scorpion species is difficult or where access to multiple specialized treatments is limited.
More importantly, the researchers did not demonstrate that any of the 17 proteins could neutralize venom. The experiment established that certain engineered molecules could recognize and bind to recombinant versions of their intended targets under laboratory conditions. It did not establish that the molecules could prevent those toxins from disrupting sodium channels in nerve cells, reduce the severity of poisoning in an animal or safely treat a human patient. Anthropic explicitly acknowledged that the work had not yet established neutralization in a cellular environment or efficacy in living animals.
This gap between molecular binding and therapeutic effectiveness is the central lesson of the project. A protein can attach tightly to a toxin without blocking the biological interaction that causes illness. Even when it attaches at a promising location, researchers must determine whether the molecule remains stable, reaches the relevant tissues, works at a practical dose and avoids harmful effects of its own. A potential medicine must also be manufactured consistently, stored safely, administered appropriately and tested through increasingly demanding stages of development. A computationally impressive design is therefore a starting point for drug discovery, not a substitute for it.
Anthropic identified two of its strongest candidates, Cn2-B1 and Cn2-B2, as minibinders targeting Cn2, a neurotoxin associated with the Mexican scorpion Centruroides noxius. Both demonstrated binding in laboratory tests, and competition experiments suggested that they may attach at or near a site used by an antibody previously reported to neutralize the toxin in mice. That is a potentially useful clue because interfering with the toxin's ability to interact with its biological target could provide a route toward neutralization. However, the company has not established that these new minibinders actually block the toxin's effects in cells or protect living animals.
The next steps would require additional experimental work. Researchers would first need to verify that promising candidates bind to toxin obtained from actual venom, rather than only to laboratory-produced versions. They would then need to test whether the proteins prevent the toxins from disrupting sodium-channel function, followed by animal studies to assess protective effects and safety. Even successful results would not immediately produce a treatment ready for patients. Further research would be necessary to establish dosage, stability, manufacturing methods, potential immune reactions and performance across relevant toxin variants.
The project also provides a useful test of claims that AI could allow people without extensive scientific training to carry out sophisticated research. Citrini's team approached the challenge from an investment-research background rather than as an established pharmaceutical drug-development group. Claude helped coordinate the process from defining the computational problem to preparing designs for external laboratory testing. That suggests AI may lower certain barriers to conducting early-stage research by helping users navigate specialized tools and organize complex workflows.
Nevertheless, the experiment was not simply a case of an ordinary user asking a chatbot to invent a medicine. It relied on an internal research model, Anthropic's scientific support, specialized computational tools and external laboratories capable of synthesizing and testing the proteins. The team also benefited from human direction and decisions about which biological targets to pursue. AI helped accelerate parts of the process, but the results depended on an extensive scientific and technical infrastructure beyond the model itself.
The distinction matters for the biotechnology industry, where AI companies and investors increasingly emphasize the potential to compress drug discovery timelines. AI can help researchers search enormous design spaces, propose molecular structures, analyze scientific literature and prioritize candidates for testing. These capabilities could reduce the time spent on some early research tasks. But laboratory experiments remain necessary to determine whether predictions correspond to biological reality, and later development stages introduce additional hurdles that computational models cannot resolve simply by producing more candidates.
The scorpion experiment also highlights why the quality of a benchmark matters. If success is defined as producing a molecule that binds to a target, Claude's results offer evidence of progress in protein design. If success is defined as creating a drug that neutralizes venom and can eventually be administered safely to patients, the experiment has not yet met that standard. Both descriptions concern the same work, but they measure very different achievements. Confusing them can lead to exaggerated expectations about how close AI systems are to replacing the wider process of drug development.
Anthropic has released all 211 designs and associated experimental data so other researchers can study the results and attempt to improve on them. Making the data available creates an opportunity for independent scientists to compare computational predictions with measured binding, investigate unsuccessful candidates and identify approaches that may be worth further testing. It also makes the work more useful as a scientific contribution, even without an approved treatment emerging from the initial campaign.
For Anthropic, the project demonstrates both the promise and the limitations of expanding Claude into scientific research. The model helped generate a large number of candidate molecules in a relatively short period and identified proteins that performed as intended in a specific binding assay. But the most important medical question remains unanswered: can any of those molecules neutralize dangerous toxins safely and reliably? Until that is demonstrated, the work should be regarded as an early-stage research achievement rather than the creation of an antivenom.
The larger implication is that AI may change how quickly scientists generate and prioritize drug candidates without eliminating the need for painstaking biological validation. In some cases, the abundance of computational designs could make laboratory testing, experimental capacity and careful interpretation more important bottlenecks. Producing more candidates is valuable only if researchers can determine which ones work, understand why they work and turn them into practical therapies.
Claude's scorpion experiment is therefore neither proof that AI has solved drug discovery nor evidence that AI-assisted biotechnology is a dead end. It is a concrete demonstration of what current tools can accomplish—and where their capabilities stop. The system helped design proteins that bound to scorpion neurotoxins, but no working antivenom emerged from the test. The decisive breakthrough will come only if further experiments show that those promising molecular interactions can be translated into a safe, effective and affordable treatment for people at risk of scorpion stings.