Claude Discovers a Novel Enzyme System: AI-Driven Science Arrives
On September 23, 2026, Anthropic published results showing Claude autonomously discovered ART — a new enzyme system with CRISPR-like properties — after 21 hours of analysis by 950 agents using 210 million tokens. This lesson shows what AI-driven hypothesis generation looks like and how to apply the same workflow pattern in your own work.
Claude found something no one had noticed before. In spring 2026, Anthropic's new life sciences lab gave Claude a single high-level prompt: search a massive database of DNA sequences for interesting new examples of reverse transcriptases (RTs — enzymes that copy RNA into DNA). 21 hours later, after 950 agents processed 210 million tokens, one agent spotted a repeating DNA pattern next to an unusual RT gene and wrote: "[The DNA next to the RT] is spectacular: I can see by eye a tandem repeat array … that's a CRISPR-like … repeat array?!"
The discovery is called Array-associated Reverse Transcriptases (ART). Claude gathered over 200,000 RTs from the database, narrowed them to 3,500 candidate systems, then to the 20 most compelling — a process that takes expert scientists weeks to months. ART is found in bacteriophages (viruses that infect bacteria) and has three parts: the RT itself, a partner gene, and a long array of evenly spaced DNA repeat sequences. That repeat layout resembles a CRISPR array. CRISPR is already the foundation of gene-editing medicine. Feng Zhang, a CRISPR pioneer at MIT and the Broad Institute, called this result "genuinely intriguing and merits further investigation."
What made this work was the workflow, not just the model. Anthropic's team gave Claude a high-level goal, then let agents run autonomously — reading literature, reproducing known results to check methods, generating short candidate reports, and critically eliminating weak candidates at each stage. The team used Claude Science and Claude Code, tools available to any researcher today. Their role was the initial prompt and the lab experiments. The rest was Claude. Because Claude generates hypotheses prolifically, the team's evaluations became a feedback signal: what they learned about which candidates were worth testing went back into the next prompt, teaching Claude to mimic their scientific taste.
You don't need a molecular biology lab to use this pattern. The core workflow — survey a domain, generate candidate reports, eliminate weak candidates, feed your evaluation criteria back into the prompt — works wherever you have a large unstructured dataset and a hypothesis worth testing. Legal precedent mining, financial document analysis, literature reviews, competitive intelligence: the pattern transfers. Define the family of things you're searching for. Ask Claude to reproduce known results first to verify its method. Request short reports that propose a function and describe the evidence. Evaluate what made the useful candidates different from the noise — then give that criteria back to Claude for the next run.
What you can do now. Claude can run large-scale hypothesis generation across datasets — in biology and in any field with lots of unstructured information. The ART discovery shows three practices that transfer directly: (1) give Claude a high-level search goal with a clear family to survey and ask it to flag what falls outside known patterns, (2) ask for short candidate reports that state a hypothesis and the evidence behind it — not just a list, and (3) feed your evaluation of which candidates were useful back into the next prompt. The loop between your judgment and Claude's breadth is where the value lives.