950 AI agents discovered unknown system in bacterial virus

The company Anthropic has used around 950 of its AI agents, autonomously working versions of the chatbot Claude, to search a huge database of DNA from bacteriophages. Bacteriophages are viruses that attack bacteria.

The task was to find so-called reverse transcriptases: enzymes that work as a kind of copy machine and transfer information from RNA to DNA. Afterward, the agents were to look at the DNA around each enzyme to see if there was anything unusual that could point to an entirely new biological system.

The agents worked for 21 hours and used 210 million tokens, the amount of text an AI model processes along the way. They found more than 200,000 of the enzymes and identified around 3,500 possible new systems. The list was finally narrowed down to 20 strong candidates, which were described in reports that researchers could read and assess.

Found an unusual pattern

One of the agents discovered a repeated DNA pattern next to one of the enzymes. It was exactly the kind of strange repetition that originally put researchers on the trail of the gene-editing tool CRISPR. The enzyme itself was not new, but the system around it appeared to be.

Researchers tested the finding in the lab and confirmed that the pattern actually produced small RNA molecules. Anthropic has named the system ART. It is still unknown what the system is actually used for in nature.

Humans still run the lab

Anthropic CEO Dario Amodei says that Claude did most of the actual discovery work, while humans decided where to search and carried out the lab experiments themselves. He compares the development to how AI has gone from being able to solve high school math in 2023 to being able to tackle difficult, open research questions today.

Amodei has previously said that AI can help cure most diseases within 5-10 years. But even though AI can speed up the search process, lab experiments and clinical trials still take place at the pace the physical world allows. The key question will therefore be whether AI can keep proposing better experiments while researchers wait for results from the slower lab experiments.