On September 23, Anthropic announced that its large language model, Claude, identified an enzyme system with properties similar to Crispr, the gene-editing tool that won a Nobel Prize. According to Anthropic, approximately 950 Claude agents worked simultaneously for 21.5 hours to find genetic sequences. Fyodor Urnov, a gene-editing expert at the University of California, Berkeley, praised Anthropic for sharing their discovery, although he noted that the Innovative Genomics Institute, where he works, did not participate in this specific finding.
Some scientists expressed skepticism about the significance of the discovery, which is the first from a research group formed by Anthropic earlier this year. Le Cong, a Stanford University professor, remarked that while the announcement is premature, the experiments are still pending. Anthropic clarified that the finding is not conclusive, stating, "We don’t yet understand what this system does," and noted that only a few known systems share its features.
Claude's search was prompted to identify "interesting new examples" of reverse transcriptases, proteins that convert RNA into DNA. The model initially identified over 200,000 potential reverse transcriptases and narrowed it down to a family that resembles Crispr, which Anthropic has named ART (array-associated reverse transcriptases). This system is found in jumbo phages, large viruses that infect bacteria.
Seth Shipman from the Gladstone Institutes acknowledged the interest in the discovery but emphasized that the novelty lies in the method of identification rather than the finding itself. He cautioned against attributing the discovery solely to AI, as human scientists were involved in the study. Jason Gill, a microbiologist at Texas A&M University, noted that the same reverse transcriptase had been previously identified in a 2021 study, but Claude's identification of surrounding repeats suggests a possible connection to a Crispr-like system.
Cong highlighted the importance of selecting appropriate problems for AI models to tackle, suggesting that Claude's capabilities could be beneficial for large-data projects, though it does not imply that AI will revolutionize life sciences discoveries. Concerns regarding the transparency of the model's training data were raised, as the scientific community cannot evaluate the information used, which contrasts with current expectations for reproducibility in research.
Anthropic's CEO, Dario Amodei, speculated on the future potential of Claude to autonomously conduct experiments, though he acknowledged that such advancements would require careful consideration of safety and oversight. The timeline for significant advancements in gene editing remains lengthy, as evidenced by the 25 years it took to develop Crispr into a practical tool.
The limitations of Anthropic's finding raise questions about the ability of AI to discover useful biological systems without prior knowledge. Cong remarked that an AI trained only on pre-Crispr knowledge discovering Crispr would be a more compelling test of AI's capabilities in scientific discovery.