Anthropic's Claude AI discovers novel viral defense system

Anthropic announced that its Claude AI system has identified a previously unknown biological system in viruses that infect bacteria, marking the first major announcement from the company's new life sciences laboratory.

The discovery, reported September 24, centers on array-associated reverse transcriptases (ART), a structure that shares architectural similarities with CRISPR, the bacterial defense mechanism that revolutionized gene editing. The system consists of three components: a reverse transcriptase enzyme that copies RNA into DNA, an adjacent partner gene, and a long array of repeating DNA sequences.

While the individual enzyme was already known to science, Claude identified how these elements may function together as a coordinated system—a pattern that had escaped human researchers.

How the AI worked autonomously

What distinguishes this work is not simply what was found, but how Claude found it. The AI agents operated with substantial independence, searching through more than 200,000 reverse transcriptases, identifying 3,500 candidate systems, and selecting 20 for detailed analysis.

During this process, one agent noticed an unexpected repeat array pattern, compared it against known biological systems, searched existing scientific literature, and flagged ART as potentially significant—all without explicit human direction at each step.

According to Dimitri Perrin, Head of School for Computer Science at Queensland University of Technology, who analyzed the announcement for The Conversation, the individual analytical techniques Claude used are well-established in bioinformatics. A conventional computational pipeline could likely have detected the same pattern.

The advance lies in Claude's ability to receive a broad research objective, then autonomously decide which analyses to perform, which leads to pursue, and which results warranted experimental validation by human researchers.

Why it matters

This represents a potential shift in how AI contributes to scientific discovery. Rather than simply processing data or answering specific researcher queries, AI agents may increasingly help determine which questions are worth asking in the first place. Modern biological datasets contain more sequences than researchers can individually examine—AI systems that can conduct exploratory searches, connect observations with existing literature, and prioritize candidates could become essential research tools.

The practical significance of ART itself remains uncertain. Perrin notes that while the system is architecturally similar to CRISPR, there is no evidence yet that it shares CRISPR's key functional capability: programmable targeting and editing of specific DNA sequences. Laboratory experiments confirmed some behavioral similarities, but calling ART a "new CRISPR" would be premature.

One critical unknown is how many of Claude's 3,500 candidate systems represent genuinely novel biology. Finding one interesting result among thousands of possibilities differs from demonstrating reliable discovery capability.

Anthropic CEO Dario Amodei suggested on X that Claude might eventually autonomously control laboratory equipment to perform its own experiments, connecting to ongoing efforts to build robot-driven biological laboratories. Such developments would fundamentally transform biological research while raising new questions about safety and oversight.

The details were first reported by The Conversation, with analysis by Dimitri Perrin of Queensland University of Technology.