Anthropic publishes Claude protein design study: 310,000 designs generated in 64 hours, 17 validated experimentally

Claude designs 310,000 proteins in 64 hours, 17 pass experimental validation

Machine Heart Download the client | Exclusive early access October 21, 2026 07:05:47 From Beijing Updated October 21, 2026 11:27:45

Just moments ago, Anthropic announced a biology study: in 64 hours, Claude generated 318,484 protein designs in an attempt to find candidate molecules capable of binding scorpion venom neurotoxins.

Anthropic says that about 1.2 million people worldwide are stung by scorpions each year, and antivenom serum is often expensive or in short supply. Using Claude to design proteins that can bind scorpion venom neurotoxins may represent an early step on the path toward exploring new treatments.

Data shows that the research team screened 211 candidate proteins from the 318,484 generated protein designs and handed them to laboratories for testing. Ultimately, 17 met the preset binding criteria, and these molecules may provide leads for the development of new anti-scorpion-venom drugs.

Of course, the current experiments only validated their ability to bind the target toxins; they have not yet proven that the proteins can neutralize the toxins or provide protection in living organisms. But what is truly noteworthy about this study is that Claude not only participated in protein design but also assisted in multiple steps including screening candidate solutions. From generating molecules to laboratory validation, AI is beginning to take part in a more complete scientific research workflow...

More than 1.2 million scorpion stings occur worldwide each year, causing more than 3,250 deaths. Traditional antivenom serum relies on animal immunization for production, which is costly and covers only a limited range of toxins, making scorpion venom poisoning a public health problem that has long lacked sufficient research and development investment.

Anthropic collaborated with investment research firm Citrini Research, hoping to use AI to design from scratch small proteins capable of binding scorpion venom neurotoxins, namely minibinders.

Simply put, some neurotoxins in scorpion venom bind to sodium ion channels on nerve cell membranes, interfering with nerve signal transmission and, in severe cases, potentially causing heart failure, pulmonary edema, and other consequences.

The researchers' idea was that if sufficiently precise proteins could be designed to preemptively bind the sites where toxins interact with sodium ion channels, it might be possible to prevent the toxins from taking effect.

This experiment selected 5 scorpion venom neurotoxins from Mexico, Brazil, North Africa, and the Middle East: Cn2, Ts1, AaH I, AaH II, and LqhIII.

Claude analyzed the protein structures and key binding sites of the five scorpion venom neurotoxins and designed candidate binding proteins accordingly. The upper part shows the three-dimensional structure of the toxins, and the lower part marks the potential binding regions selected by Claude.

Claude generated 318,484 candidate protein designs and ultimately selected 180. With Claude's help, the research team then added 31 more, for a total of 211 candidate proteins, which were handed to two laboratories, Adaptyv Bio and Twist Bioscience, for validation.

Ultimately, 17 designs met the research team's protein binding criteria.

Specifically, 7 of them were able to bind Cn2, 4 bound AaH I, 4 bound AaH II, and another 2 bound LqhIII. The designs targeting Ts1 failed to pass screening.

Claude's candidate proteins designed against the five scorpion venom neurotoxins and the experimental screening results. Among them, Cn2 yielded 7 effective binding proteins, AaH I and AaH II each yielded 4, LqhIII yielded 2, and Ts1 yielded no effective binding proteins.

The most noteworthy findings in the study are two candidate proteins targeting the Cn2 toxin, Cn2-B1 and Cn2-B2.

Both showed strong binding ability in experiments, with dissociation constants KD reaching the single-digit nanomolar level. A lower KD generally means the molecule binds more tightly to the target protein.

The research team also conducted competitive binding experiments.

They found that these two proteins may bind at the same or a nearby site as a known neutralizing antibody called LR. Previous studies have already shown that LR can neutralize the Cn2 toxin in mouse experiments.

Predicted binding structures of Claude-designed candidate proteins Cn2-B1 and Cn2-B2 with the scorpion venom neurotoxin Cn2. Both target the key region where the toxin binds to ion channels, compared with the binding sites of the known antibodies LR and RU1.

This means that the two candidate proteins designed by Claude may have value for further research.

However, it should be noted that this experiment validated the binding ability of the proteins to the toxins, but did not actually prove that these proteins can neutralize scorpion venom.

The researchers have not yet conducted cell-level toxin neutralization tests, nor have they validated protective effects in animals.

Another interesting result is that Claude encountered difficulties when designing proteins targeting multiple toxins simultaneously.

The research team had originally hoped to find universal proteins capable of binding multiple scorpion venoms, in order to reduce the number of molecules that future antivenom formulations would need to combine.

But in the end, each of the 17 proteins that passed screening was confirmed to bind only one toxin, and none of the designs met the criterion of binding multiple toxins simultaneously.

Looking at the design of single-toxin-binding proteins, of the 60 candidate designs specifically targeting a single toxin, 11 passed screening, a success rate of about 18%.

Anthropic said this level is broadly consistent with the performance of typical protein design research.

Although there is still a clear gap before a universal antivenom, this study at least demonstrates that Claude can propose a large number of candidate designs in a relatively short time and screen out proteins that genuinely have experimental binding ability.

Behind the 64 hours, how much work did Claude actually complete?

Traditional protein design requires researchers to use multiple tools to complete tasks such as structural analysis, candidate molecule generation, binding prediction, and screening. This time, Anthropic tried to have Claude organize these steps autonomously.

The research team first specified the target toxins, available tools, and screening rules, after which Claude independently retrieved protein structures, selected design tools, set screening criteria, and continuously adjusted its design plans.

Throughout the process, Claude invoked 12 protein design tools, generated 318,484 candidate designs, and also performed structure predictions for more than 550,000 candidate protein-toxin combinations.

After receiving the initial prompt, Claude ran autonomously and continuously for 24 hours. The researchers then added 4 more prompts to guide it to further optimize the designs. The entire computational design phase lasted 64 hours, and in the end Claude screened out 180 candidate proteins.

Even the task of sending protein sequences to the laboratory was completed by Claude according to the researchers' requirements.

This means Claude is already capable of taking on a large amount of the planning, analysis, and tool coordination work involved in the scientific research process, and of continuously adjusting the research plan based on intermediate results.

However, humans were still involved in the experiments. The researchers were responsible for setting goals, adding prompts, and supplementing designs, and the subsequent protein synthesis and validation were also carried out by professional laboratories. In addition, the 64 hours covered only the computational design phase, which consumed about 10,000 GPU hours.

This is also not the first time Claude has attempted protein design. In August this year, Anthropic published a study in which Claude carried out designs for 15 protein targets, and 14 of them successfully yielded experimentally validated binding proteins.

Compared with the previous study, this experiment further demonstrates Claude's ability to coordinate complex scientific research workflows in specific biomedical problems.

But there is still a lot of work to do before a true antivenom. Currently, Anthropic has publicly released all 211 candidate designs and the related experimental data for other researchers to continue validating.

From protein design to drug development,

If Anthropic is demonstrating AI's ability to design biological molecules, then another biological AI company, Radical Numerics, has pushed the story into the human treatment setting.

Just now, Radical Numerics released the genomic large model Omnii-1 and announced that a phage designed by its AI has for the first time been approved for compassionate use in a patient with a drug-resistant bacterial infection. The company said the relevant phage has passed laboratory bactericidal tests and has been authorized for use in patient treatment.

Phages are a type of virus that can infect bacteria, and under specific conditions they can be used in treatment research targeting bacterial infections, especially in scenarios such as drug-resistant bacterial infections where conventional treatment options are limited.

At the same time, OpenAI announced the acquisition of biomedical AI company Orchestra Bio and plans to integrate its team and technology into its life sciences platform Rosalind.

Unlike Anthropic, which focuses on demonstrating AI's protein design capabilities, this OpenAI acquisition focuses on another complex part of the drug development process: the coordination and management of scientific research projects.

In April this year, OpenAI already launched GPT-Rosalind, specifically for life sciences research, to help researchers handle complex problems in fields such as biology, drug development, and translational medicine. Rosalind Workbench further connects scientific tools and data to help researchers analyze results and plan experiments.

The acquired Orchestra Bio mainly helps biomedical teams coordinate cross-departmental R&D work, track experimental progress, and identify dependencies among different R&D tasks.

For example, during drug development, the failure of a manufacturing batch may affect the start time of subsequent clinical trials. Orchestra's Agent can help track affected tasks, collect relevant information, and compare different response options for researchers and project leaders to make decisions.

For OpenAI, this is also a step toward extending AI from scientific problem analysis further into the entire drug development process.

From Anthropic designing candidate proteins against scorpion venom, to Radical Numerics using AI-designed biological therapeutic candidates in a patient, to OpenAI attempting to coordinate the drug development process, the explorations of these three companies are covering different stages of biomedical R&D.

The next step is how to guide these designs and discoveries smoothly through the long validation process and ultimately turn them into safe, effective treatment options that can truly help patients.

After all, finding potentially effective molecules in a computer is only the beginning of new drug development.

"Special statement: The above content (including videos, pictures, or audio if any) is uploaded and published by users of the Dafeng Hao, a self-media platform under Phoenix Net. This platform only provides information storage space services.

Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."