Which ‘AI scientist’ suits your lab? A guide for the perplexed

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. 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 complete1. This week, while unpacking after a holiday, Ashley asked the AI tool Claude, developed by Anthropic in San Francisco, California, to examine his own genome to the same standard. The analysis took 30 minutes and correctly identified an Alzheimer’s disease risk allele and gene variants affecting drug metabolism (Ashley had analysed his genome in 2012 but did not publish the results). “There is no world in which this is not utterly remarkable,” Ashley wrote in a LinkedIn post. On 30 June, Anthropic unveiled a platform called Claude Science, designed with biology research firmly in mind. The tool joins a department’s worth of general purpose AI tools for science created by technology firms and academic laboratories. 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 usually takes me hours now takes minutes. I can really spend my time on the science that needs a human,” says co-author Yuanhao Qu, the co-founder and president of Phylo, a start-up firm in South San Francisco, California. These scientific agents are distinct from more specialized research tools, such as the AlphaFold protein-structure-prediction model, but they can employ bespoke models. For example, Gabriele Corso, co-founder and chief executive of the London-based firm Boltz, and his team tasked a Claude agent to design an antibody that recognized two therapeutic targets, using the company’s open-source AI tools for protein-folding prediction and design. The AI’s outputs aligned with the protein designers’ intuitions; Corso's team did not validate the designs experimentally, but other antibodies made with AI agents have been, he says. Boltz’s tools are among the 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 the scientific literature and other sources to come up with scientific hypotheses. Bryant, who was investigating immune responses to zoonotic pathogens, provided the tool with a grant application and further data. Some of the ideas that it generated weren’t doable, but others were right in her lab’s wheelhouse. Her team is now testing an idea from Co-Scientist, introducing specific mutations into an innate-immune protein and seeing how they impact influenza infection. Bryant says she might have eventually come up with the experiment on her own, but it could have taken two years. “You feel like you’re talking to an oracle,” says Gary Peltz, a biomedical scientist at Stanford, who used Co-Scientist to identify existing drugs that could treat an organoid model of a disease called liver fibrosis3. How should scientists decide which tools to use? Many scientists already use AI tools such as Claude to generate presentation slides and draft e-mails. But Ashu Singhal, president and co-founder of the cloud platform Benchling in San Francisco, estimates that less than 20% of labs have fully embedded AI scientists into their research. “It’s really important that people actually try these things out, rather than simply trusting what gets shared in headlines,” he says. Singhal recommends that researchers trial several tools to work out which ones are suitable for which tasks. Hypothesis-generating AIs, such as Co-Scientist, might help during the earliest stages of a project. Later on, tools such as Claude Science and Biomni could carry out specific tasks, such as genomic data analysis. Corso recommends that researchers start with small tasks, the output of which can be verified easily. “Worst case, you have to do them over,” he says. How can researchers trust 'AI scientists'? Enjoying our latest content? Log in or create an account to continue Access the most recent journalism from Nature's award-winning team Explore the latest features & opinion covering groundbreaking research