Biologists at Anthropic announced on Wednesday that they had used artificial intelligence to identify genes in viruses that produce enzymes that work in a new way, the first discovery emanating from a research lab the company opened this spring.
“We have chosen biology and medicine as the primary way that we believe the benefits from A.I. can come about,” said Eric Kauderer-Abrams, the head of life sciences at Anthropic.
Dr. Kauderer-Abrams and his colleagues described their work in a technical report posted on the company’s website. The results have not yet been submitted to a scientific journal for publication.
Dario Amodei, the chief executive of Anthropic, saw great potential in the finding. “Today we announced the Claude-led discovery of a molecular machine that we suspect could represent a new gene editing mechanism,” he wrote on X on Wednesday.
Outside experts said that such claims jumped far ahead of the evidence the team has gathered so far.
“It’s not yet a breakthrough,” said Philip Kranzusch, a microbiologist at Harvard Medical School. “It’s a wrinkle on what we know in the field.”
Anthropic’s foray into biology comes amid intense controversy about the safety of A.I. Earlier this month, Anthropic reported potentially dangerous uses of its models, including queries that could potentially assist people in building biological weapons.
Two days after that announcement, Dr. Amodei called for slowing development so that humanity can enjoy the benefits of A.I. without risking catastrophe.
Dr. Amodei, who has a Ph.D. in biophysics, has long argued that one of those benefits will be new treatments for diseases. “The first step is showing that AI can first help with, and then drive, biological discoveries,” he wrote on X.
In hopes of safely demonstrating that, Anthropic opened up a biology lab earlier this year, as first reported last week by Reuters.
“Our team is still relatively small, but we are growing very rapidly,” said Dr. Kauderer-Abrams. “This is one of the most significant investment areas for the company.”
Anthropic is far from the first organization to use artificial intelligence to probe biology. A number of academic labs and start-ups are training A.I. models to discover hidden molecular machines, design new proteins or even predict how entire cells will respond to drugs.
Dr. Kauderer-Adams and his colleagues are using Anthropic’s own products, such as the large language model Mythos and Claude Code, which writes computer programs on demand.
In the new report, Anthropic’s biologists describe their search for enzymes, called reverse transcriptases. These enzymes can read the sequence in an RNA molecule and produce a matching piece of DNA.
Viruses and bacteria make reverse transcriptases for many purposes, and scientists have found them to be powerful tools. One type of Covid test, for example, uses reverse transcriptases to detect the coronavirus. Scientists have also harnessed the enzymes to develop new ways to edit genes.
As promising as these enzymes are, however, scientists have barely begun to document their diversity in the natural world. Dr. Kauderer-Adams and his colleagues created A.I. agents to search through databases of viral and bacterial genes in search of new reverse transcriptases.
The agents wrote their own programs to examine genes that serve as blueprints for 1.9 billion kinds of proteins. They searched for molecular patterns that were similar to known reverse transcriptases, along with the proteins and RNA molecules that help the enzymes do their job.
The search yielded a list of 17 possible reverse transcriptases, but 14 turned out to be errors. In some cases, they were not new. In other cases, they were not reverse transcriptases at all. But the remaining three held up to scrutiny.
The researchers claim that one of the new kinds works very differently from previously discovered ones.
The enzymes, which the scientists call array-associated reverse transcriptases, are carried by viruses. They are produced along with a set of RNA molecules that appear to work with them in some way.
But the experiment described in the technical report mostly involved analyzing DNA sequences and estimating the shapes of enzymes and RNA. There was relatively little wet-lab work with real molecules.
Dr. Kranzusch said that the Anthropic scientists have yet to do a series of basic experiments to figure out if array-associated reverse transcriptases actually work the way the researchers suggest.
They could, for example, delete the gene for the enzyme from viruses and see if they can still infect bacteria.
“This would be the first stage in dissertation research,” Dr. Kranzusch said. “They’ve published a really exciting beginning, but it’s the beginning of a story, not a complete unit that we typically see in the field.”
Nicholas Perry, an Anthropic biologist, said that he and his colleagues are now going “to try to figure out what exactly these systems are doing in the natural world.”
For Dr. Kauderer-Abrams, one of the most revealing insights was not about enzymes, but about the capacity that A.I. has to perform science. “There’s been a lot of debate about whether A.I. can have creativity, or is it only good at just executing on instructions,” he said.
Dr. Kauderer-Abrams saw evidence of scientific creativity in the transcripts that the agents generated as they carried out their assignments.
When an agent inspected array-associated reverse transcriptases, it initially determined that they were not something worth examining. But then it flagged the RNA molecules that accompanied the proteins, which were peculiar enough to warrant a second look.
“You can see the agent say, ‘Huh, that’s interesting,’” Dr. Kauderer-Abrams said.
Discussion about this post