In 2010, Euan Ashley, a geneticist and cardiologist at Stanford University in California, led the first clinical analysis of a human genome, which took his team of 31 scientists nine months to complete.¹.

This week, while unpacking after a vacation, Ashley asked the artificial intelligence tool Claude, developed by Anthropic in San Francisco, California, to examine her own genome to the same standard.

The analysis lasted 30 minutes and correctly identified a risk allele for Alzheimer’s disease and genetic variants that affect drug metabolism (Ashley had analyzed her genome in 2012 but did not publish the results). “There is no world where this isn’t absolutely extraordinary,” Ashley wrote in a LinkedIn post.

On June 30, Anthropic unveiled a platform called Claude Science, designed firmly with biological research in mind. The tool joins general-purpose AI tools for departmental science created by technology companies and academic labs. Others include offerings from OpenAI in San Francisco and Co-Scientist from Google DeepMind in Mountain View, California. Another is an open source tool called Biomni, developed by academic researchers and described yesterday in Science². And there are many others, researchers say.

“Work that normally took me hours now takes me minutes. I can really dedicate my time to the science that a human being needs,” says co-author Yuanhao Qu, co-founder and president of Phylo, a startup in south San Francisco, California.

What are they and how do scientists use them?

Sometimes called “AI scientists,” these tools build on the large language models that power chatbots, helping scientists with tasks such as literature reviews, data analysis, figure generation, and manuscript preparation. They are a form of agent AI, in which requests are broken down into steps that often involve hiring external software systems.

These scientific agents differ from more specialized research tools, such as the AlphaFold protein structure prediction model, but can employ custom models. For example, Gabriele Corso, co-founder and CEO of London-based firm Boltz, and his team commissioned a Claude agent to design an antibody that recognized two therapeutic targets, using the company’s open-source artificial intelligence tools for protein folding prediction and design.

AI ‘scientists’ joined these research teams – here’s what happened

The AI ​​results aligned with the protein designers’ intuitions; Corso’s team did not validate the designs experimentally, but other antibodies made with AI agents have, he says. Boltz’s tools are among dozens of specialized software systems that Claude Science and other AI scientists can interact with.

Clare Bryant, an immunologist at the University of Cambridge, UK, was an early adopter of Co-Scientist, which mines scientific literature and other sources to generate scientific hypotheses. Bryant, who was researching immune responses to zoonotic pathogens, provided the tool with a grant application and more data.

Some of the ideas he generated were not feasible, but others were in his lab’s wheelhouse. His team is now testing a Co-Scientist idea, introducing specific mutations into an innate immune protein and seeing how they affect influenza infection. Bryant says he could have eventually come up with the experiment on his own, but it could have taken two years.

“You feel like you’re talking to an oracle,” says Gary Peltz, a Stanford biomedical scientist who used Co-Scientist to identify existing drugs that could treat an organoid model of a disease called liver fibrosis.³.

How should scientists decide which tools to use?

Many scientists already use AI tools like Claude to generate presentation slides and email drafts. But Ashu Singhal, president and co-founder of cloud platform Benchling in San Francisco, estimates that fewer than 20% of labs have fully integrated AI scientists into their research. “It’s really important that people actually try these things, rather than just relying on what’s being shared in the headlines,” he says.

Singhal recommends that researchers test various tools to determine which ones are suitable for which tasks. Hypothesis-generating AIs, like Co-Scientist, could help during the early stages of a project. Later, tools like Claude Science and Biomni could perform specific tasks, such as analyzing genomic data.

Corso recommends that researchers start with small tasks, the outcome of which can be easily verified. “Worst case, you have to remake them,” he says.

How can researchers trust ‘AI scientists’?