AI experiment fails to create viable scorpion antivenom drug

The ambitious drug discovery test

Citrini Research, an investment firm, recently attempted to see if artificial intelligence could revolutionize the way we develop new medications. They partnered with Anthropic to use an unreleased version of Claude to design a new type of scorpion antivenom. The goal was to see if the model could create a drug from scratch rather than just assisting in existing processes.

The team chose scorpion antivenom because it is a costly and difficult treatment to produce. Traditional methods often involve the complex process of injecting venom into horses to develop a counter-agent. By using AI, the researchers hoped to design specific proteins that could bind to and neutralize the toxins found in scorpion venom.

To get started, the team had to bypass Claude's standard biosecurity guardrails, which initially blocked them from researching certain drug candidates. They eventually gained access to a more advanced model and were provided with a 35-page prompt. Anthropic also covered the compute and token costs for the project, which totaled just under $50,000.

How the AI generated candidates

The researchers deployed approximately 300 subagents to work with Claude to identify suitable protein designs. The AI was tasked with scanning for molecules that would stick to the specific toxins the team wanted to target. Over a period of 64 hours, the model produced nearly 320,000 different candidates.

Anthropic reported that the team then whittled these down to 180 designs by discarding any that looked too much like known toxins or existing proteins. The researchers then worked alongside the AI to add another 31 designs, bringing the total to 211. This automated discovery process was intended to find a 'hit' that could serve as a foundation for a medicine.

To verify the AI's work, Citrini sent the designs to two outside laboratories: Adaptyv Bio and Twist Biosciences. These tests were designed to see if the proteins could actually bind to their intended targets. This phase of the experiment cost the team roughly $150,000 to complete.

Promising binders but no medicine

The laboratory results showed that some of the proteins passed the initial test. These proteins collectively bound to four of the five specific scorpion toxins the team had set out to target. However, none of these proteins were able to bind to more than a single toxin at a time.

While the results were technically successful in terms of binding, experts say they are far from a finished product. Binding to a toxin is considered a relatively simple task in biotechnology. The real challenge lies in ensuring that a protein can actually neutralize a toxin safely within a living body without causing side effects.

Citrini acknowledged that no experiment had yet tested whether these proteins actually stopped the toxins from working. Because the proteins had not been tested for safety, stability, or behavior in a biological system, the project did not result in a viable drug. Researchers described the findings as being 'far from a medicine and hyped.'

The limits of AI in biology

The experiment highlights a significant gap between identifying a molecule and creating a treatment. Professors from the Technical University of Denmark noted that translating these binders into actual medicine is a long and challenging process. They argued that while the autonomous discovery process was novel, the biological results are still in the very early stages.

Some experts also pointed out that AI might not be the only way to achieve these results. They noted that existing methods, such as high-throughput screening, can often find many hits at manageable cost without using AI. Designing proteins that bind to targets is described by some as a problem that is already 'kind of' solved.

Despite the lack of a new drug, Citrini framed the experiment as a success for understanding investment opportunities. They argued that the project provided a glimpse into the future of AI-enabled drug discovery. However, they clarified that they are an investment research firm and not a biotech company capable of conducting the long-term clinical trials required for medicine.

What comes next for AI drugs

The results of this experiment suggest that AI is currently better at the 'easy-ish' parts of drug discovery. It can quickly generate thousands of candidates and filter them down based on specific criteria. This ability to accelerate the initial discovery phase could eventually save time and money for pharmaceutical companies.

The next major hurdle for AI in this field is the 'translation' phase. This involves taking a promising molecule and proving it can work safely in humans. This requires extensive wet lab testing, animal studies, and clinical trials, which are much more complex than the initial protein design.

Moving forward, the industry will likely focus on how AI can help navigate these complex biological hurdles. While Claude did not invent a new antivenom this week, the experiment provided a data point on how much work remains before AI can independently produce life-saving treatments.

The cost of discovery

The financial scale of this experiment provides a clear picture of the costs involved in AI-driven research. Between the compute costs and the external laboratory testing, the project required significant funding. This underscores the reality that even with powerful AI, physical laboratory validation remains a major expense.

Citrini Research noted that while the AI portion of the work was essentially 'free' for them, the underlying costs were substantial. The $150,000 spent on external labs highlights that the physical world still requires human-led infrastructure to verify digital predictions. This creates a bottleneck for purely autonomous drug discovery.

As the field evolves, the goal will be to find ways to make these expensive laboratory tests more efficient. For now, the collaboration between Anthropic and Citrini serves as a high-profile example of the current state of the art: AI can find the needle in the haystack, but humans still have to build the medicine.

Scientific limitations

The experiment also served as a reality check for the hype surrounding AI in the life sciences. Many observers were quick to point out that the results were not a breakthrough in medicine. Instead, it was a demonstration of an autonomous discovery process that reached a known scientific milestone.

The fact that the AI-generated proteins only bound to single toxins is a key limitation. A practical antivenom would ideally need to be more versatile or more potent. The current output shows that while the AI understands the 'binding' instruction, it hasn't yet mastered the complexity of multi-target neutralization.

Ultimately, the experiment confirms that AI is a powerful tool for generating hypotheses and candidates. It can process vast amounts of data to find potential leads, but the transition from a digital design to a safe, effective treatment remains a deeply complex biological challenge that AI cannot yet solve alone.

The role of Anthropic

Anthropic's involvement in this project is notable as the company prepares for its upcoming IPO. By partnering with Citrini, Anthropic is positioning itself as a key player in the life sciences and drug development sectors. They provided the advanced model and the necessary compute power to facilitate the experiment.

The fact that Anthropic agreed to drop its biosecurity guardrails for this specific project shows the level of cooperation required for high-level research. It also highlights the need for specialized environments, as the team was required to work from Anthropic's offices to access the unreleased model. This suggests that high-end AI research still requires a high degree of human oversight and physical infrastructure.

While the experiment did not produce a drug, it served as a high-profile test of Claude's capabilities in a complex scientific domain. It demonstrated the model's ability to handle long, complex prompts and manage multiple subagents to achieve a specific research goal. For Anthropic, this serves as a proof of concept for the model's utility in sophisticated, multi-step scientific workflows.

Investment implications

For the investors who read Citrini's reports, the experiment was intended to highlight emerging opportunities. The firm focused on the companies involved in the testing side of drug development, such as Charles River Laboratories. This suggests that the real value may currently lie in the infrastructure that validates AI designs.

The report also serves as a reminder of the distinction between research and production. Citrini's refusal to pursue the results further because they are not a biotech startup illustrates the current boundaries of the industry. AI is currently a tool for discovery, and the path to commercialization still requires traditional pharmaceutical expertise.

As AI continues to integrate into the life sciences, the focus will likely shift toward refining these discovery pipelines. The goal is to move from identifying 'hits' to identifying 'leads' that have a higher probability of success in clinical trials. This experiment marks a small but significant step in mapping out that journey.

The path to medicine

The journey from a protein design to a vial of medicine is long and fraught with failure. This experiment only completed the very first step of that journey. It showed that Claude can navigate the initial search space of protein chemistry to find molecules that interact with a target.

To turn these successful binders into a medicine, researchers would need to perform extensive toxicity tests. They would need to see how the proteins behave in different tissues and whether they cause unintended reactions. These are the types of biological questions that current AI models are not yet equipped to answer autonomously.

The experiment serves as a grounded look at the future of AI in science. It moves the conversation away from 'AI will cure everything' toward 'AI can help us find the right starting point.' By identifying the 'easy-ish' parts of the process, AI allows scientists to focus their human expertise on the harder challenges of medicine.

Summary of findings

In summary, the collaboration between Citrini Research and Anthropic successfully used Claude to generate a massive number of protein candidates for scorpion antivenom. While the AI was able to identify several proteins that could bind to the toxins, these results do not constitute a functional drug. The experiment highlights the current capabilities of AI in the discovery phase of biotechnology.

Experts agree that while the autonomous discovery process is impressive, the biological hurdles remain significant. The transition from a binder to a medicine involves safety, stability, and efficacy tests that AI cannot currently perform. The experiment serves as a reminder that AI is a powerful assistant, not a replacement for the complex work of drug development.

The results provide a clear roadmap of what is possible today and what remains for the future. AI can handle the heavy lifting of initial screening and candidate generation. However, the final leap into human medicine still requires the rigorous, multi-year validation process of traditional science.

Technical hurdles

One of the primary technical hurdles identified in the experiment was the specificity of the AI's designs. While the proteins bound to the toxins, they were not versatile enough to bind to multiple toxins simultaneously. This limitation is a common challenge in protein engineering, where achieving high affinity for one target often comes at the expense of others.

Another hurdle is the lack of biological context in the AI's output. Claude can predict how a protein might interact with a toxin in a vacuum, but it cannot yet predict how that interaction will occur in a complex biological environment. Factors like pH levels, temperature, and the presence of other proteins can all drastically change how a drug behaves in the body.

The experiment also highlighted the need for high-quality data and precise prompting. The 35-page prompt was necessary to guide the AI through the specific requirements of the task. As AI becomes more integrated into science, the ability to provide these high-quality instructions will be a critical skill for researchers and developers alike.

The role of wet labs

The necessity of external laboratories like Adaptyv Bio and Twist Biosciences underscores the importance of 'wet labs' in the AI era. Even with the most advanced models, physical testing is required to validate digital predictions. This creates a hybrid model of drug discovery where AI handles the 'dry' research and labs handle the 'wet' validation.

The cost of this physical validation—roughly $150,000 for this specific test—remains a significant barrier to entry. It means that while AI can make discovery cheaper, the overall process of bringing a drug to market remains expensive. This is why the infrastructure for testing AI-generated molecules is becoming a key area of interest for investors.

In the future, we may see more integrated systems where AI models are directly connected to automated laboratory robots. This could potentially speed up the feedback loop between digital design and physical testing. However, for now, the human-led laboratory remains the ultimate gatekeeper for turning AI ideas into reality.

Conclusion on AI potential

The Citrini and Anthropic experiment provides a sober, realistic look at the current state of AI in drug discovery. It avoids the hype of 'instant cures' and instead shows the incremental progress being made in protein design. The AI successfully narrowed down hundreds of thousands of possibilities to a manageable number of candidates.

This ability to filter the 'noise' of chemical space is a major win for science. It allows researchers to focus their time and resources on the most promising candidates rather than testing every possible combination. This efficiency is where AI will likely provide the most immediate value to the pharmaceutical industry.

Ultimately, the experiment shows that AI is a powerful tool for exploring the unknown. It can help us find the starting points for new medicines, but the journey from those points to a finished drug is still a long and difficult road. The results are promising, but as the experts noted, the excitement should be tempered by the reality of the work still to come.

Future of AI discovery

Looking ahead, the success of the 'autonomous discovery process' will likely be the focus of future research. The goal is to refine these AI systems so they can handle more complex tasks, such as predicting the safety and stability of the proteins they design. This would move the AI further along the path toward creating viable medicines.

We can also expect to see more collaborations between AI labs and biotech firms. As models like Claude become more capable, they will be integrated into the standard workflows of drug development. This will likely start with 'easy-ish' tasks like binder discovery before moving into more complex areas of pharmacology.

The experiment also highlights the importance of biosecurity and responsible AI development. The fact that the team had to navigate guardrails shows that the industry is actively thinking about the risks associated with AI in biology. Balancing these safety concerns with the potential for life-saving innovation will be a key theme in the coming years.

The investment landscape

For those following the investment side of AI, this story highlights a shift toward the 'picks and shovels' of the AI revolution. Rather than just betting on the AI models themselves, there is growing interest in the companies that provide the testing and validation infrastructure for AI-generated discoveries.

Citrini's focus on companies like Charles River Laboratories suggests that the physical reality of science remains a bottleneck. As AI generates more candidates, the demand for high-throughput screening and animal research will likely increase. This creates a stable investment opportunity in the infrastructure that supports the AI boom.

The experiment also serves as a case study for how investment firms are exploring the boundaries of technology. By attempting to push AI to its limits, they can better understand the timeline for commercial viability. This helps investors make more informed decisions about where the real value lies in the rapidly evolving AI landscape.

Final thoughts on limitations

The most important takeaway from this experiment is the distinction between a 'hit' and a 'drug.' A hit is a molecule that shows the desired interaction in a controlled test. A drug is a safe, effective, and stable treatment that can be administered to humans. AI has made significant strides in finding hits, but the path to drugs is still dominated by complex biology.

The fact that none of the proteins bound to more than one toxin is a specific technical hurdle that the AI has yet to overcome. Solving this would require a deeper understanding of the multi-functional nature of venom. It is a reminder that biology is messy and that AI models still need to learn the nuances of complex systems.

However, the fact that the AI could generate 320,000 candidates in 64 hours is a testament to its power. It can explore a vast amount of chemical space in a fraction of the time it would take a human. This capability will continue to