LLMs

Anthropic's Claude Experiment Falls Short of Drug Discovery Claims

Anthropic and research firm Citrini tested an unreleased Claude model on scorpion antivenom design. The AI generated protein candidates, but independent lab tests found none viable.

Last verified:

Claude’s Scorpion Antivenom Experiment Yields No Viable Drug Candidates

According to The Verge, Anthropic and investment research firm Citrini Research conducted a high-profile experiment to test whether an unreleased version of Claude could design novel proteins for scorpion antivenom. Despite generating over 320,000 protein candidates in 64 hours, independent laboratory validation found none of the designs functionally viable—a significant gap between AI-assisted protein generation and real-world drug efficacy.

How the Collaboration Began

Citrini Research initiated the partnership after encountering Claude’s biosecurity safeguards while attempting to research drug candidates. According to The Verge, the firm’s author “spam called anyone in my contact list who worked at an AI lab” before securing Anthropic’s participation. Anthropic agreed to disable its standard restrictions and provide access to a more advanced unreleased model, conditional on the team working from Anthropic’s offices. Anthropic also covered the substantial computational overhead—nearly $50,000 in compute and token expenses.

The timing was strategic: Anthropic was weeks away from a widely anticipated initial public offering, creating incentives to demonstrate AI’s potential in drug development and life sciences, two sectors the company is actively pursuing.

The Experiment’s Technical Approach

The team selected scorpion antivenom as their target, given the treatment’s scarcity and production complexity—traditional methods involve injecting venom into horses to generate antibodies. Using a 35-page prompt, the researchers deployed approximately 300 AI subagents to identify suitable protein sequences.

According to The Verge’s account, Claude generated nearly 320,000 candidates, then filtered the pool to 180 designs after removing sequences resembling known toxins or existing proteins. Through an iterative process with Claude, the team added 31 additional designs, reaching a final count of 211 protein sequences. The entire computational run consumed 64 hours.

Laboratory Validation Reveals the Gap

Citrini sent the AI-generated designs to two independent laboratories for validation: Swiss biotech firm Adaptyv Bio and US-based Twist Biosciences. This testing phase, aimed at confirming whether the proteins could actually bind to scorpion venom toxins—a critical prerequisite for antivenom efficacy—cost an additional $150,000.

The results were negative. According to The Verge, experts confirmed that none of the AI-generated protein candidates showed binding activity against the target venom. This marked a clear boundary between AI-assisted molecular hypothesis generation and functional drug discovery.

Why This Matters

The experiment highlights a recurring gap in AI drug-discovery narratives: the distance between pattern-matching at scale and biological validation. While large language models excel at synthesizing existing scientific knowledge and generating plausible hypotheses, they lack ground-truth feedback from wet-lab biology. Citrini and Anthropic’s candid reporting of failure—rare in the AI industry—offers a calibrated data point for teams evaluating whether LLM-driven protein design can accelerate early-stage drug pipelines. The $200,000 cost and negative result suggest that computational protein generation, even with unreleased models and custom guardrail removal, remains a tool for narrowing search space rather than a substitute for experimental validation.

Frequently Asked Questions

Did Claude successfully design a scorpion antivenom?

No. While Claude generated 320,000 protein candidates, independent laboratory testing by Adaptyv Bio and Twist Biosciences found none were functionally viable for binding to scorpion venom toxins.

Who conducted this experiment and why?

Anthropic partnered with Citrini Research, an investment research firm, to test whether an unreleased Claude model could accelerate drug development—an area Anthropic is actively pursuing ahead of its IPO.

How much did this experiment cost?

Approximately $200,000 in total: nearly $50,000 in compute and token costs, plus $150,000 in laboratory testing fees at Adaptyv Bio and Twist Biosciences.

#Claude #drug-discovery #protein-design #AI-limits #Anthropic