Citrini and Anthropic's Attempt at AI-Driven Antivenom Falls Short

The Journey into AI Drug Development

In a bold experiment to explore artificial intelligence in drug discovery, Citrini Research partnered with Anthropic to investigate whether Claude, a cutting-edge AI model, could aid in creating scorpion antivenom. Scorpion stings can pose severe health risks, making effective antivenom treatments critically important.

Facing Unique Challenges

Creating antivenom is a complex and expensive process often involving the extraction of venom from horses. Citrini, admitting its own lack of expertise in biology, set out to tackle this thorny issue using AI. Their journey was documented in a blog post titled, “We Wanted to Use Claude to Create a New Drug. Here’s What Happened.”

The Collaboration with Anthropic

Citrini initially encountered obstacles due to biosecurity concerns that limited their use of Claude for drug research. Committed to advancing their project, they enlisted Anthropic's help. The two organizations worked together, with Anthropic agreeing to provide access to a more advanced version of Claude under specific conditions.

Financially, this collaboration required significant investment, with computing costs approaching $50,000. Citrini equipped Claude with a structured 35-page prompt, allowing it to generate a plethora of protein designs intended to bind to toxins found in scorpion venom.

AI Generates Protein Candidates

Over the span of approximately 64 hours, Claude produced around 320,000 protein candidates. After sifting through the results and eliminating those that shared similarities with existing proteins or known toxins, the team narrowed it down to 211 promising designs. To determine their effectiveness, Citrini sent these designs to external labs, including Adaptyv Bio and Twist Biosciences, at an additional cost of about $150,000.

Encounters with Reality

Of the 211 protein candidates, 17 exhibited the ability to bind to four of the five targeted scorpion toxins. However, none were able to bind to more than one toxin at a time, which is a critical limitation. Moreover, Citrini acknowledged that demonstrating binding does not guarantee that these proteins could neutralize the toxins, raising concerns about their potential as a therapeutic solution.

Expert Opinions on the Findings

Experts weighed in on the experiment, expressing that while the findings hold some technical promise, they are far from a fully developed medication. Professor Andreas Hougaard Laustsen-Kiel noted that while discovering good binders is relatively straightforward, converting these findings into actual medicines is a much greater challenge.

Melisa Benard Valle and Anne Ljungars from the Technical University of Denmark echoed this sentiment, emphasizing the necessity for further evaluations to understand the proteins' safety, effectiveness, and how they interact within the body. They commented that translating these binders into functional treatments remains an arduous process.

Looking Forward

Despite not achieving the desired outcome of creating a new drug, Citrini framed the experiment as a significant step toward understanding AI's potential in pharmaceutical development. The firm argued that lessons learned could pave the way for investment opportunities in the field of AI-assisted drug discovery.

Final Thoughts

The initiative undertaken by Citrini and Anthropic highlights both the potential and the pitfalls of utilizing AI in drug development. While the journey of transforming AI-generated solutions into effective medical treatments is fraught with complexities, it also points to a future where AI could play a transformative role in how we approach drug discovery.