In the long human effort to read the language written into living things, a new kind of reader has entered the library. Anthropic's AI agents — nearly a thousand of them, working autonomously through the night — surfaced a strange repeating pattern in the genomes of giant viruses, one that echoes the architecture of bacterial immune memory. The finding is preliminary and its meaning unresolved, but it marks a moment in which machine intelligence begins to participate, not merely assist, in the act of biological discovery. What the pattern does, and whether it can be made useful, now falls to h
Anthropic's AI discovers CRISPR-like DNA patterns in giant viruses
Biology is hard. It requires actually running experiments.
Mark
So 950 AI agents spent over 21 hours searching databases and found these repeating DNA patterns. That's a lot of computational power aimed at one discovery. Why does that matter?
Mimi
Because genomic databases contain billions of sequences. A human researcher might spend months or years looking through them manually. The AI agents can explore that space much faster, and they can do it in a self-directed way—they're not just following a rigid script, they're discussing findings with each other and deciding what to investigate next.
Luke
But let's be precise about what they actually found. These are repeating DNA sequences in giant viruses that structurally resemble CRISPR arrays. That's the observation. The function is completely unknown.
Mimi
Right. And that's why Kauderer-Abrams called it a promising lead rather than a breakthrough. The structural similarity is interesting, but it doesn't tell us what these sequences do.
Mark
So the next step would be to actually test them in a lab?
Mimi
Yes. Someone has to design experiments to figure out whether these viral sequences have any immune function at all, whether they work with any known enzymes, whether they could be engineered into tools. That's the hard part.
Luke
And that work hasn't started yet. This is a preprint, not peer-reviewed. We don't even know if other researchers will be able to reproduce the pattern-finding or if they'll agree on what the patterns mean.
Mark
Does that undermine the finding?
Luke
Not undermine it, but it contextualizes it. This is an interesting observation from an AI system. Whether it becomes a real discovery depends on what happens in the wet lab and what the scientific community makes of it.
Mimi
The broader point is that AI can be very good at finding anomalies in large datasets. But biology is fundamentally experimental. You can't know what something does just by looking at its structure.
O Pulso
- Nearly 950 AI agents spent over 21 hours autonomously navigating vast genomic databases, a scale and speed of search no human team could replicate.
- The agents surfaced repeating DNA sequences in giant viruses that structurally mirror bacterial CRISPR arrays — one of the most powerful molecular tools ever discovered — raising the possibility of an entirely unknown viral immune mechanism.
- The discovery remains unverified and unpeer-reviewed, with no known DNA-cutting enzyme yet identified to work alongside the sequences, leaving their function an open question.
- Anthropic's biology chief acknowledged the hard boundary AI cannot cross: understanding what these patterns actually do requires physical experimentation that only human scientists can conduct.
- The finding positions AI not as a replacement for biological research but as a systematic scout — capable of mining petabytes of data for leads that humans must then pursue in the lab.
In the long human effort to read the language written into living things, a new kind of reader has entered the library. Anthropic's AI agents — nearly a thousand of them, working autonomously through the night — surfaced a strange repeating pattern in the genomes of giant viruses, one that echoes the architecture of bacterial immune memory. The finding is preliminary and its meaning unresolved, but it marks a moment in which machine intelligence begins to participate, not merely assist, in the act of biological discovery. What the pattern does, and whether it can be made useful, now falls to human hands and human experiments to determine.
Anthropic, the San Francisco AI company, has opened a biology laboratory where human scientists and AI systems work together — and this week revealed what may be its first meaningful discovery. Nearly 950 autonomous AI agents spent more than 21 hours moving through enormous genomic databases, examining billions of proteins in search of sequences that might interact with reverse transcriptases. In the genome of one giant virus, the agents noticed the same short stretch of DNA letters appearing again and again. They flagged it, and found similar patterns in other viral genomes as well. The finding was posted as a preprint on September 23 and has not yet been peer reviewed.
What caught researchers' attention is the structural resemblance these repeating sequences bear to bacterial CRISPR arrays. In bacteria, CRISPR systems store molecular memories of past invaders — repeated DNA segments interspersed with captured fragments of viral genetic material — and use them to guide enzymes that destroy returning threats. Scientists have since turned this mechanism into one of biology's most transformative tools for gene editing.
Whether the viral sequences discovered by Anthropic's agents do anything similar remains unknown. No companion DNA-cutting enzyme has been identified, and the functional role of the patterns is entirely uncharacterized. Eric Kauderer-Abrams, who leads life sciences at Anthropic, called the finding a promising lead while being candid about what lies ahead: the hard, physical work of wet-lab experimentation that AI cannot perform. The agents can find patterns in data at extraordinary scale, but determining what those patterns mean — and whether they hold practical value — still depends entirely on human scientists working at the bench.
Anthropic, the artificial intelligence company based in San Francisco, has opened a biology laboratory where human scientists and AI systems work side by side. This week, the company revealed what may be the first significant discovery to emerge from that partnership: a strange repeating pattern of DNA found in the genomes of several giant viruses, a pattern that bears structural resemblance to the immune systems bacteria use to defend themselves against invaders.
The discovery came about through an unusual process. Nearly 950 AI agents—autonomous systems built on large language models—spent more than 21 hours moving through vast databases of DNA sequences, examining billions of proteins in a self-directed manner. The agents were tasked with finding proteins that might interact with reverse transcriptases, enzymes that copy RNA into DNA. While investigating the genetic material surrounding a reverse transcriptase gene in one giant virus, the agents noticed the same short sequence of DNA letters appearing repeatedly. They flagged this as worth investigating further, and in doing so, identified similar patterns in the genomes of other viruses as well. The finding was posted online on September 23 on the alphaXiv preprint platform and has not yet undergone peer review.
The repeated DNA sequences the agents discovered bear a structural likeness to arrays found in bacterial CRISPR systems. In bacteria, CRISPR systems typically contain multiple chunks of repeated DNA separated by segments copied from viruses or other genetic invaders. When a threat arrives, the bacterium manufactures RNA from those copied segments, which then guides an enzyme to cut up the invader's DNA, destroying it. It is a form of molecular immune memory. Scientists have harnessed this bacterial mechanism into one of the most powerful tools in modern biology, enabling researchers to edit genes with unprecedented precision.
But the viral sequences Anthropic's team discovered remain largely a mystery. There is little evidence that these repeating patterns perform functions similar to their bacterial counterparts. No known DNA-cutting enzyme has been found to work alongside them. Eric Kauderer-Abrams, who heads life sciences research at Anthropic, described the finding as a promising lead but acknowledged the difficulty of what comes next: fully characterizing the sequences, understanding what they actually do, and determining whether they could be developed into useful tools.
The discovery offers a window into how AI might be used to systematically search through enormous volumes of genomic data in search of novel molecular mechanisms. Kauderer-Abrams noted that many of biology's most important breakthroughs have come from identifying unexpected phenomena in microbes, and the goal of Anthropic's new lab is to scale and systematize that kind of exploratory work. Yet he was clear about the limits of what AI alone can accomplish. Biology, he said, requires actually conducting experiments in the physical world. The agents can spot patterns in data, but humans must still design and run the wet-lab work needed to understand what those patterns mean and whether they have practical value. For the foreseeable future, the human contribution to biological research remains essential.
Citações Notáveis
Our goal was to see if we could systematize and scale up that kind of research—finding weird things in microbes.— Eric Kauderer-Abrams, head of life sciences at Anthropic
This is a promising lead. What's hard is to then go and completely characterize it, understand its function, and develop it into an interesting tool.— Eric Kauderer-Abrams
Quer a matéria completa?
Leia o original em Nature ↗