Claude discovers array-associated reverse transcriptases (ART)

TL;DR

Anthropic announced that Claude, its large language model, autonomously discovered a novel enzyme system—array‑associated reverse transcriptases (ART)—characterized by a CRISPR‑like repeat array adjacent to a reverse transcriptase in bacteriophages, showcasing AI’s ability to generate and prioritize biological hypotheses at scale.

Background: AI‑augmented life‑sciences research

Anthropic created a dedicated life‑sciences research group to explore whether general AI models can systematize and accelerate fundamental biology discoveries. The team trains Claude on biological literature, equips it with tools for large‑scale DNA database mining, and integrates its outputs with wet‑lab validation performed by human scientists.

Discovery workflow

  • Prompt and dataset: Researchers gave Claude a high‑level prompt to search a massive DNA sequence database for interesting reverse transcriptases (RTs). No detailed hypotheses were supplied.
  • Agent scale: Approximately 950 Claude agents processed the task over 21 hours, consuming roughly 210 million tokens.
  • Candidate triage: Agents gathered >200 k RT sequences, identified 3.5 k novel candidate systems, and distilled them to 20 high‑confidence candidates, a process that would take weeks to months for a human expert.
  • Pattern detection: One agent flagged a tandem repeat array next to an atypical RT, describing it as “spectacular” and “CRISPR‑like.”
  • Report generation: The agent produced a human‑readable report summarizing repeat count, spacing, comparative analysis, and literature search, then submitted it for expert review.

The ART system

  • Components: ART consists of three parts—(1) a reverse transcriptase gene found in a jumbo phage, (2) a neighboring accessory protein of unknown function, and (3) a long array of evenly spaced, non‑coding DNA repeats.
  • CRISPR‑like features: The repeat array resembles CRISPR arrays that store programmable RNA guides. Initial experiments show the ART array is transcribed into distinct short RNAs, hinting at a potentially programmable mechanism.
  • Novelty: While the underlying RT had been cataloged previously, Claude was the first to recognize the co‑occurring repeat array and accessory gene as a coherent, previously uncharacterized system.

Expert commentary

“This is an exciting example of how AI agents can contribute to biological discovery. The identification of RNA‑repeat arrays associated with reverse transcriptases is genuinely intriguing and merits further investigation.” — Feng Zhang, MIT and Broad Institute.

Significance and implications

  • Proof of concept: The ART discovery demonstrates that large language models can autonomously generate biologically meaningful hypotheses, prioritize candidates, and guide experimental validation without detailed human direction.
  • Accelerated genome mining: Claude reduced a multi‑month manual curation process to a single day of distributed AI computation, suggesting a scalable pathway for uncovering new enzymatic systems across vast genomic repositories.
  • Potential biotechnological tools: If ART functions analogously to CRISPR—providing programmable DNA manipulation—it could become a new platform for gene editing, synthetic biology, or diagnostics.
  • Open science: Anthropic released a pre‑print detailing the ART system and invites the broader community to explore its function, fostering collaborative validation.

How the lab operates

  • Computational pipeline: Claude reads literature, reproduces known results, surveys protein families, flags anomalous genomic neighborhoods, and drafts concise reports.
  • Human‑in‑the‑loop: Scientists review Claude’s reports, conduct wet‑lab expression and biochemical assays, and interpret experimental data with Claude’s assistance.
  • Iterative improvement: Feedback from experimental outcomes refines Claude’s prompting and instruction set, teaching the model to emulate expert scientific judgment.

Future directions

  • Functional characterization: Ongoing experiments aim to determine whether ART mediates DNA cutting, copying, or other programmable activities.
  • Broadening scope: Anthropic plans to apply the same AI‑driven workflow to other unexplored protein families, expanding the discovery pipeline beyond RTs.
  • Collaboration invitation: The lab seeks external partners to test ART and to co‑develop AI‑augmented research programs in genomics and related fields.

Resources

Sources