AI biology lab Anthropic’s first discovery questioned – scientists see his unpublished work in it

Last week, Anthropic presented findings from its biolabs. The company uses artificial intelligence agents to discover promising new enzymes. Some scientists were quick to express doubts about the achievement. An expert at the University of Copenhagen said his team shared their research results with the Anthropic AI model, and that the neural network’s touted “discovery” of new enzymes coincided with years of work.

Image source: Human

Computational biologist Mario Rodríguez Mestre says he and his colleagues have been studying these enzymes and related molecules (what Anthropic calls ART – array-associated reverse transcriptase) for the past four years. Their findings have not yet been published. Over the past three years, they have regularly used Anthropic’s artificial intelligence models to write code, prepare manuscripts, and perform other tasks. Meister believes that in this way, they allowed Anthropic to obtain key results of enzyme research.

Anthropic denies the allegations: “We are not aware of any previously published work describing our recently discovered ART system. Nor has Claude received any training on customizing transcripts, nor does our molecular biology team have access to them.” Anthropic said Claude’s key contribution was the identification of a set of RNA and protein molecules associated with reverse transcriptase in the system.

But Meester said he and his colleagues made the discovery more than a year ago. The question, according to Meester, is whether Anthropic’s artificial intelligence actually arrived at these results in a logical way, or whether it was guided in part by its own work. “My concern is that this information is used to train future versions of the model. – said Mestre. — I think it’s important to ask this question. “

Anthropic makes biological research a core part of its mission. “We believe that this type of application will develop rapidly in the coming years, and that we will be able to make significant progress in the near future in the fight against diseases that have plagued humans for thousands of years,” ——Dario Amodai, director of the Human Research Center, said in a speech at the United Nations.

Human researchers said they instructed artificial intelligence agents to scour a database of billions of genetic sequences, but gave no precise clues about what they were using. These libraries contain DNA collected from the ocean, soil and other habitats. Scientists asked artificial intelligence to find the gene that controls the production of reverse transcriptase. These enzymes read the sequence of an RNA molecule and create a corresponding DNA molecule with the same sequence.

Bacteria use certain reverse transcriptase enzymes to detect invading viruses. Some viruses use reverse transcriptase as a weapon against other viruses. Human researchers provided the artificial intelligence with five examples of reverse transcriptase. AI agents wrote their own software to mine the database for similar genes and vet promising candidates.

These reagents identified genes that may contain reverse transcriptase in bacteria infected with viruses called macrophages. Each of these sequences is located near a group of genes that encode RNA molecules that may help the enzyme carry out its function. Anthropic researchers call these enzymes microarray-associated reverse transcriptases (ARTs). They claim that AI is independently developed “The first step in biological discovery.”

Meester disputed that claim. “These are the same systems we’ve been studying for years.”he said. These enzymes, he said, were the same ones he described to Anthropic’s artificial intelligence. He said he first discovered these enzymes in giant phages in 2022 while searching for new reverse transcriptases. He briefly served as a consultant to ReNegade Therapeutics, which filed a patent application in 2023 for several reverse transcriptase enzymes, including those discovered by Dr. Mestre. He is listed as a co-author on the patent application.

In 2023, a year after Dr. Mester began his research, Anthropic released Cloud. Like many other scientists, he found it useful in solving a wide variety of problems. Mester acknowledges use of artificial intelligence “It completely changed his attitude toward science.” The AI ​​actively participated in his research, sharing drafts of his papers and manuscripts with the neural network. He and his colleagues also downloaded more information and used Cloud to perform various administrative tasks.

Meister’s words were corroborated by other scientists. Bioengineer Seth Shipman said he is familiar with Meister’s research. “We’ve been discussing this with him for years, – he said. — Many of us are aware of these systems. ” Anthropic announced the discovery of the enzyme, causing serious concern among scientists. “A lot of our colleagues ask, ‘Oh, are they learning my skills? Shipman added. — We can’t help but think about it. “

Meester acknowledged that it’s impossible to say with certainty that Anthropic’s AI findings are based entirely on his research. “I mean, coincidences happenhe said. — Everyone in this field is doing more or less the same thing. But we’ve known about these replays for over a year. ” He recalled that AI companies have also been accused of training models on the work of artists and writers without copyright permission. Mestre also announced that he would end all projects using Cloud.

Mester’s comments echoed another scientific conflict. OpenAI announced that it has used artificial intelligence to solve an important previously unproven mathematical problem. But mathematician Tristan Buckmaster, who uses OpenAI’s artificial intelligence agents extensively in his work, suspects that the AI ​​borrowed ideas from his work. “Is it ethical to use customer data to try to intercept their customers?”, he wondered on social media. OpenAI denies that its models were trained on Buckmaster’s development.