Citrini and Anthropic Explore AI in Scorpion Antivenom Research

In a recent initiative to explore the applications of artificial intelligence in drug discovery, Citrini Research partnered with Anthropic. Their ambitious goal was to determine whether Claude, a cutting-edge AI model, could aid in the development of scorpion antivenom.

Understanding the Challenge

Scorpion stings can be deadly, underscoring the urgent need for effective antivenom treatments. However, creating these drugs is challenging and costly, often involving complex processes such as extracting venom from horses. Citrini, while acknowledging its limited experience in biological sciences, decided to tackle this issue using AI, as detailed in their blog post titled, “We Wanted to Use Claude to Create a New Drug. Here’s What Happened.”

A Partnership Forms

Initially, Citrini faced difficulties when trying to use Claude due to built-in biosecurity guardrails. In an effort to progress, they reached out to Anthropic, which agreed to provide access to a more advanced version of Claude, with preliminary support requiring them to work on-site.

The collaboration was costly, with the computational expenses nearing $50,000. With a structured 35-page prompt, they unleashed Claude to develop a variety of protein designs aimed at binding to scorpion toxins.

Exploring Protein Candidates

Over the course of approximately 64 hours, Claude generated nearly 320,000 protein candidates. Following a detailed assessment, the team refined these down to 211 viable designs after eliminating those that resembled existing proteins or known toxins.

To test the efficacy of these protein designs, Citrini sent samples to external laboratories, including Swiss biotech firm Adaptyv Bio and US-based Twist Biosciences. This phase required a budget of about $150,000 and aimed to check if the proteins could effectively bind to the targeted toxins.

Mixed Results

Out of the assessed proteins, 17 successfully bound to four of the five scorpion toxins they aimed to target. However, a significant limitation was observed: none of the proteins could bind to more than one toxin at a time. Citrini acknowledged that while these results were a step forward, they were still far from achieving a functional antivenom.

Experts express that binding to toxins does not translate into the ability to neutralize those toxins, making it vital to conduct further experiments to determine safety and effectiveness. Andreas Hougaard Laustsen-Kiel, a professor at the Technical University of Denmark, remarked on the ease of discovering good binders but pointed out that crafting these findings into an actual medicine represents a much larger hurdle.

The Path Forward

Citrini’s foray into drug discovery showcases the possibilities but also highlights the limitations of using AI for pharmaceutical development. While the team achieved the fundamental step of binding discovery, the road to creating viable medical treatments remains long and complex.

Many experts, including Melisa Benard Valle and Anne Ljungars from the same institution, emphasized the need for rigorous evaluations of the proteins' performance within biological systems to understand their safety and functionality.

Despite falling short of their initial goal, Citrini framed the experiment as a valuable insight into the potential of AI in drug discovery. They suggested that there are significant investment opportunities in this emerging field as AI continues to evolve.

Concluding Thoughts

This venture into AI-driven drug discovery illustrates both the progress and the challenges faced in the pharmaceutical industry. While Citrini may not have produced a breakthrough antivenom, their experience serves as a reminder that the journey from AI-generated concepts to effective medicines requires careful navigation through complex biological questions. The future of AI in drug development holds promise, yet it also demands substantial research and investment to overcome existing hurdles.