An AI model has found a new ‘CRISPR-like’ biological system. Here’s what it means for science

Anthropic announced today that its Claude artificial intelligence (AI) system identified a previously unknown biological system in the DNA of viruses that infect bacteria.

The press release is the first from Anthropic’s new life sciences laboratory. It highlights some similarities to CRISPR, the natural bacterial defence system that scientists now use as a gene-editing technology.

It is too soon to say that Claude – or the team of human researchers involved – has discovered a “new CRISPR”. But the result does show us something about how AI is beginning to work in science, with more autonomy than simply analysing data or answering the specific questions of a researcher.

What did Claude find?

The new system is called array-associated reverse transcriptases (ART). It contains three key elements:

• an enzyme known as a reverse transcriptase, which copies RNA into DNA

• a partner gene which appears next to the enzyme

• a long array of repeating DNA sequences.

The enzyme itself was already known. What is new is the recognition that it may form part of this larger system.

CRISPR has a similar structure, which is the basis for the comparison. Anthropic’s laboratory experiments also found the ART array behaves similarly to CRISPR in some ways.

However, CRISPR’s key feature is that it can be programmed to target and edit specific DNA sequences. It is not known whether the same is true of ART.

At this point in time, we can say that ART is CRISPR-like in its architecture, but there is no evidence that it is CRISPR-like in its function.

How did Claude find this ART system?

What was found is perhaps less interesting than how it was found.

The important advance is in autonomy. Starting from an overall research brief, Claude AI agents searched through more than 200,000 reverse transcriptases, identified 3,500 candidate systems, and selected 20 for detailed analysis.

During this process, one AI agent followed an unexpected observation. It noticed the repeat array pattern, compared it with known systems, and recognised a potentially interesting result. It searched the existing literature, and then flagged ART for human researchers.

None of the individual techniques involved are fundamentally new. There are well-established tools to find genes related to each other, to cluster sequences, to detect motifs and repeats in sequences, to examine which genes occur together, and to calculate how RNA molecules fold into three-dimensional structures.

A purpose-built conventional bioinformatics pipeline could probably have detected the same pattern. The study does not provide a comparison that would tell us whether ART would have been an obvious result of such a pipeline or a difficult signal to identify among thousands of candidates, though the press release claims this type of analysis can take researchers “weeks to months of work”.

The analysis was not beyond the capabilities of existing methods; what makes this work important is something else. It is that Claude was given a fairly broad scientific objective and then chose which analyses to perform and which leads to follow, and that it recognised one result as interesting enough for experimental validation.

One missing piece of information is that we do not know how many of the 3,500 candidate systems, or of the selected 20, represent genuinely novel or interesting biology. Finding one intriguing result after exploring thousands of possibilities is not the same as showing that the system can reliably recognise discoveries.

The future of AI in science

If the approach can be shown to reliably make discoveries, it could become an important application of AI in biology. Modern biological datasets contain more sequences than researchers can examine individually.

AI agents could become another tool for researchers to use to manage existing tools. They could conduct large exploratory searches, connect observations with the scientific literature, generate hypotheses, and prioritise candidates for scientists to test.

Whether ART itself proves important remains an open question, but the value of the study is that it suggests a potentially important shift in how AI could contribute to science. AI may increasingly help decide where existing tools should be pointed and which questions are worth asking.

Claude needed human researchers to carry out experiments to confirm its suggestions, but that might not always be the case. There are ongoing research efforts to build robot-driven biological laboratories.

Posting on X about the discovery, Anthropic chief executive Dario Amodei wrote

eventually it may even be possible for Claude itself to safely perform the experiments by autonomously controlling lab equipment.

These developments may fundamentally transform biological research. They will also raise questions around how the technology can be kept safe.