Claude Designs Protein Binders to Counter Scorpion Neurotoxins

Scorpions sting about 1.2 million people each year, and conventional antivenom is often expensive and in short supply. In a science blog post, Anthropic disclosed that its large model Claude has taken an early step toward finding novel protein binders for scorpion neurotoxins. If validated, this advance could open a low-cost, scalable new path for antivenom therapies.

The Antivenom Dilemma and Where AI Fits In

Around 1.2 million people worldwide are stung by scorpions each year, a substantial share of them in regions with scarce medical resources. Existing antivenom relies on extraction through animal immunization, which involves long production cycles, high costs, and limited coverage of specific toxins. When supplies run short, clinicians are often left with few options.

Anthropic noted in its blog that antivenom is "often expensive or in short supply." This is precisely where AI has a reasonable entry point: if a model can design small proteins that bind to toxins, it could potentially bypass the production bottlenecks of traditional antibodies. Claude's work here is an early exploration that has not yet entered clinical stages, but it points toward a clear pain point.

Antivenom (anti-snake venom/anti-scorpion venom serum) is produced by injecting trace amounts of toxin into large animals (usually horses or sheep), waiting for their immune systems to generate antibodies, and then extracting and purifying antibody fragments from the blood. This process involves animal rearing, immunization cycles, blood collection, and purification, typically taking months, with high costs and capacity limited by the number of animals. More critically, animal antibodies are heterologous proteins to the human body and may trigger serum sickness or allergic reactions. In addition, a single antivenom usually neutralizes only the toxins of a specific species or closely related species; given the thousands of scorpion species worldwide, coverage is extremely limited. These structural shortcomings are exactly what computational design of small protein binders seeks to circumvent—if humanized proteins that bind toxins with high affinity can be designed directly, the capacity bottleneck of animal immunization can be eliminated.

What Claude Did: From Sequence to Binder

According to the blog post, Claude was used to design protein binders targeting scorpion neurotoxins. Specifically, the model needs to understand the structural features of toxin proteins and generate protein sequences that can bind to them stably. Such tasks place high demands on the model's spatial reasoning and biochemical knowledge.

Anthropic frames this as an "early step," meaning the current results are closer to a proof of concept than a mature therapy. The blog does not disclose binding affinity data, experimental validation details, or hit rates—precisely the key metrics for judging the value of such work. Without this information, outsiders can only view it as a directional signal rather than a reproducible result.

Protein binders are small protein molecules that can bind to a target protein (here, scorpion neurotoxins) with high affinity and specificity. Compared with traditional antibodies (~150 kDa), small binders such as antibody mimetics, affibodies, or designed ankyrin repeat proteins typically have molecular weights of only a few to a dozen kDa, offering better tissue penetration and lower manufacturing costs. The core challenge in designing such molecules is that they must simultaneously satisfy shape complementarity, electrostatic complementarity, and hydrophobic matching with the target surface while maintaining their own folding stability. Computational protein design usually relies on physics-based modeling tools such as Rosetta or diffusion models such as RFdiffusion; having a general-purpose large language model directly generate binder sequences is a newer approach that treats protein sequences as a "language" to be modeled.

What It Means for Large Models to Enter the Protein Design Arena

Claude's attempt is not an isolated event. In recent years, a variety of computational tools have emerged in protein design, but most rely on specialized physics-based modeling or diffusion models. General-purpose large language models directly participating in toxin binder design remains a relatively new exploration.

Its potential advantage lies in cross-task generalization: if the same model can handle different toxin targets, it could significantly compress early-stage R&D time. The risks are equally obvious—language model outputs require rigorous wet-lab validation, and any "plausible-looking" sequence may fail in the lab. Anthropic's choice to publicly share this stage of work reads more like a research signal than an announcement of productized capability.

Wet-lab validation is an indispensable downstream step in the computational protein design pipeline. Model-generated sequences are only hypotheses; they must be gene-synthesized, recombinantly expressed, purified, and then tested for binding affinity (e.g., KD values) using techniques such as surface plasmon resonance (SPR) or biolayer interferometry (BLI) to determine whether the design succeeded. The hit rate of current computational design tools (the proportion of generated sequences that actually fold and bind the target) is typically far below 100%, and because language models lack explicit modeling of three-dimensional structure, the "plausibility" of their outputs warrants even greater caution. Therefore, without experimental data from Anthropic, outsiders cannot assess the actual success rate of this work—a key piece of missing information for judging its maturity.

How Far From an Actual Therapy

From design to therapy, there are multiple hurdles: binding activity testing, stability assessment, animal studies, and clinical trials. The regulatory pathway for antivenom is itself somewhat unusual, and new types of binders would need to demonstrate efficacy in real envenomation scenarios.

Anthropic's blog does not provide a timeline. For readers interested in AI for Science, the value of this work lies in demonstrating a possibility: large models are not limited to text tasks but can also tackle design problems at the molecular level. Whether it can truly alleviate antivenom supply gaps will require more experimental data to answer.

Source Note

This article is based on a summary of Anthropic's science blog post published on X, with the original tweet linking to its official blog. Because the summary does not include experimental data or methodological details, this article takes a cautious stance on judging the maturity of the results.

AI for Science refers to the application of artificial intelligence methods to basic scientific research, covering directions such as protein structure prediction (e.g., AlphaFold), molecule generation, materials discovery, and optimization of experimental design. In the antivenom field, AI's potential contributions are not limited to designing binders; it can also be used for toxin diversity analysis, epitope prediction, and cross-reactivity assessment of antivenom. However, going from computational design to a drug candidate typically requires lead optimization, preclinical toxicology, pharmacokinetic studies, and multi-phase clinical trials, often measured in years. For diseases such as scorpion envenomation that mainly affect low-income regions, insufficient commercial incentives are also a constraining factor; even if the technical path is opened, accessibility still requires policy and funding support.

One-line Summary

Anthropic says Claude designed protein binders for scorpion neurotoxins, an early proof-of-concept step.

Paragraph Summary

Anthropic disclosed in a science blog post that its large model Claude was used to design protein binders targeting scorpion neurotoxins, positioned as an early proof of concept. About 1.2 million people worldwide are stung by scorpions each year, and traditional antivenom relies on animal immunization, with high costs, long cycles, and frequent shortages. If AI can design small proteins that bind toxins, it could potentially bypass this production bottleneck. However, the blog does not disclose binding affinity, experimental validation details, or hit rates, so outsiders can only view it as a directional signal. General-purpose large language models directly participating in protein design remains a relatively new exploration, and their outputs require rigorous wet-lab validation; an actual therapy is still separated by binding activity testing, animal studies, and clinical trials.