Claude identified 17 proteins that bind to scorpion venom, covering 4 of the 5 toxins targeted by the team, but none of them bound to more than one type of venom.
The key point is that binding to venom does not mean the protein can neutralize its effects or actually be used to treat people. The experiments did not yet test safety, stability, behavior in the body, or the ability to neutralize the venom. This is only the beginning of the research process and remains far from a ready-to-use antivenom.
What This Experiment Actually Proved
Citrini Research and Anthropic gave Claude a 35-page prompt and used 300 subagents to search for proteins that might bind to scorpion venom. The process took 64 hours and generated nearly 320,000 candidates before narrowing them down to 211.
The selected candidates were then sent to Adaptyv Bio and Twist Biosciences for testing in external laboratories. Seventeen proteins bound to toxins from 4 of the 5 targeted venom types. The results show that AI can help accelerate the search for starting points, but they do not yet represent a drug or a ready-to-use product.
Why Developing a Scorpion Antivenom Is a Worthwhile Challenge
Scorpion antivenom is expensive and difficult to produce because the process often involves injecting horses with venom to generate antibodies. This approach makes drug development both time- and resource-intensive.
The interesting question is how much AI can accelerate the earliest stages, particularly the design of proteins that bind to toxins in scorpion venom. If Claude can identify candidates more quickly, researchers may have a starting point for further drug development.
Where Claude Fits Into AI-Driven Drug Development
In this experiment, Claude served as a tool for searching for and designing proteins that might bind to toxins in scorpion venom. It did not create a finished antivenom. The results therefore represent only a starting point for further laboratory testing.
Anthropic is pushing the potential of AI in drug development and biological science by giving the team access to an unreleased version of Claude to search for protein candidates. The process generated nearly 320,000 candidates within 64 hours before narrowing them down to 211.
However, proteins that bind to toxins still require extensive testing for safety, stability, and behavior in the body. Claude is therefore better suited to the role of a research assistant than an automated drug factory.
From the Original Experiment to the Model Used in Testing
Before the project began, the Citrini team encountered biosecurity guardrail restrictions, so they contacted Anthropic and eventually worked directly with its research team. This experiment used an unreleased model, with those restrictions lifted under the condition that the work be conducted from Anthropic’s office.
| Factor | Restrictions Before the Project | This Experiment |
|---|---|---|
| Biosecurity | Restrictions applied during drug candidate research | Anthropic agreed to lift the restrictions |
| Model access | Unable to access the desired model | Access to an advanced unreleased model |
| Collaboration | The team worked independently | Collaborated with Anthropic and worked from its office |
What Claude Was Able to Do in a Real Laboratory Setting
-
Process a 35-page prompt defining the task of searching for proteins that bind to toxins in scorpion venom
-
Use approximately 300 subagents to assist with the search, generating nearly 320,000 candidates before narrowing them down to 180
-
Work with the team to add 31 more protein designs, bringing the total to 211
-
When tested by external laboratories, 17 proteins were found to bind to the toxins, but they were not yet ready-to-use antivenoms
How This Compares With Drug Development
The experiment reached the stage of identifying proteins that bind to toxins, but it remains far from an antivenom that can be used to treat people.
| Factor | Experimental Results | Ready-to-Use Medicine |
|---|---|---|
| Binding to toxins | Achieved partially | Must work effectively |
| Neutralizing the venom | Not yet tested | Must be proven |
| Safety and stability | Unknown | Must pass validation |
| Behavior in the body | Unknown | Must be supported by data |
| Drug development | Still far from becoming a drug | Must be proven safe and effective for treatment |
Claude’s main strength is helping researchers identify starting molecules more quickly. Turning those molecules into real medicines remains a long way off.
Strengths and Limitations of This Experiment
Claude helped organize the search systematically, generate a large number of candidates, and identify proteins that bound to 4 of the 5 targeted venom types. However, these results are still only the beginning of drug development.
Pros
- +Systematic protein search
- +Found proteins that bound to 4 of the 5 targeted venom types
Cons
- −No protein bound to more than one type of venom
- −It has not yet been tested whether they can neutralize the venom
- −Still far from a usable medicine
The Hidden Costs Behind Using AI
Anthropic’s compute and token costs came to nearly $50,000, even though the team used Claude to search through a large number of protein designs.
The team then had to send the protein designs to two external laboratories for testing, at a cost of approximately $150,000, to determine whether the proteins actually bound to the target toxins.
This case shows that AI can accelerate the search process, but it does not replace biological experiments. The major expenses still come from verifying whether molecules actually work, are safe, and can be developed into medicines.
What This Experiment Should Change About Our Understanding
The novelty of this work may not lie in its biological results, but in its automated discovery process, which allows AI to explore a large number of possibilities more quickly.
When we see news about a molecule that appears promising, we should distinguish how far it still is from becoming a safe and practical medicine. Binding to a toxin is not the same as neutralizing it or actually treating patients.