Claude discovers a novel enzyme system

Open original

Anthropic's new life sciences lab used Claude to search through massive DNA databases and autonomously discovered a previously unknown enzyme system (called ART) that resembles CRISPR in structure and function. The discovery took 21 hours of computational search across 950 parallel Claude agents, finding something human researchers had missed despite having the underlying enzyme already identified.

  • Claude found a reverse transcriptase paired with repeat DNA arrays and an unknown accessory protein—a combination only seen in a handful of other programmable DNA-manipulation systems like CRISPR
  • The search process narrowed 200,000+ reverse transcriptases down to 3,500 candidates, then 20 finalists, work that would typically take a human scientist weeks to months
  • Anthropic built an in-house lab where Claude generates hypotheses at scale while human scientists handle wet lab experiments and final validation, creating a new human-AI research workflow

Anthropic built a lab and research team to test whether Claude can systematize biological discovery. They gave the AI a straightforward task: search a massive database of DNA sequences for interesting reverse transcriptase genes. Over 21 hours, roughly 950 Claude agents processed 210 million tokens and found something notable—a repeating pattern of DNA sequences next to an unusual reverse transcriptase gene. Lab testing confirmed this marks a previously unknown enzyme system in bacteriophages that the team calls array-associated reverse transcriptases (ART). The pattern resembles CRISPR's structure, which matters because CRISPR and a handful of other programmable DNA systems all share similar architectural features. Feng Zhang, a CRISPR pioneer at MIT, endorsed the finding as "genuinely intriguing" and worth investigating further. The team's workflow shows how they're using Claude as a research partner rather than a tool. A scientist gives Claude a high-level prompt. Claude then reads the literature, reproduces existing results to verify its methods, searches for uncharacterized proteins, and generates hundreds or thousands of candidate reports with proposed functions and supporting evidence. Humans filter these down—most get eliminated at review. When something survives scrutiny, the lab team physically tests it, expressing proteins in bacteria and characterizing them biochemically and structurally, with Claude helping interpret the data. The whole process happens in Claude's standard interfaces plus some custom coordination software for running multiple sessions in parallel. What makes this different from traditional computational biology is scale and speed. A human expert would spend weeks or months doing the kind of genomic analysis Claude completed in 21 hours. The team culled over 200,000 reverse transcriptases down to 3,500 candidate systems, then 20 compelling ones that got full reports. They're treating Claude's prolific hypothesis generation as data itself—studying which proposals they choose to test versus discard, then feeding those patterns back into Claude's instructions to refine its scientific judgment. The lab handles all wet work themselves, sticking to low biosafety-level experiments with non-pathogenic organisms.

Ask about this article