Anthropic’s Claude Discovers Potential New Enzyme System
In a statement released on 23rd September, Anthropic announced that its advanced language model, Claude, has identified an enzymatic system exhibiting “CRISPR-like” properties, referring to the Nobel Prize-winning gene-editing tool. According to Anthropic, approximately 950 instances of Claude operating simultaneously were able to locate genetic sequences of interest within a span of 21.5 hours. Fyodor Urnov, a genetic editing expert from the University of California, Berkeley, and director of therapeutic research and development at the Innovative Genomics Institute (IGI), praised Anthropic for making their discovery public, although he clarified that the IGI, which collaborates with Anthropic, did not contribute to this particular finding.
However, not all scientists share the same enthusiasm regarding this discovery, which emerges from a newly formed research group within Anthropic. Le Cong, a professor at Stanford University with a focus on integrating AI into genomic engineering research, remarked, “Experiments are still ongoing. The press release has already been published.”
Continued Exploration of the Discovery
Anthropic has clarified that this finding is merely the beginning of its research journey. “We do not yet fully understand the function of this system, but only a handful of known systems share its characteristics, all of which are capable of cutting, copying, and pasting DNA,” the company stated in a post on X (formerly Twitter). A physical experiment was included in a technical report that is yet to undergo peer review. There remains considerable work to ascertain the significance of this discovery, and it is still uncertain whether this enzymatic system can be utilised as a gene-editing tool, and if so, whether it would be effective.
“Imagine we are at Santa Monica Beach, trying to sift through all the sand to find a diamond. The AI has discovered something shiny, and now we must return to the lab to determine: Is it crystal? Is it a diamond?” Cong explained.
Claude’s Search for Reverse Transcriptases
Claude made this discovery following a request from Anthropic researchers to search vast genomic databases for “new and interesting examples” of reverse transcriptases, proteins that have the capability to convert RNA back into DNA. This process occurs in the opposite direction to the usual sequence in human cells, where DNA is transcribed into RNA. Nevertheless, various organisms employ reverse transcriptases for diverse biological functions, including the insertion of DNA segments.
Initially, the agents of Claude identified over 200,000 potential reverse transcriptases, which were subsequently narrowed down to several thousand candidates that appeared novel. Further filtering led to the identification of an “unusual” family of these proteins associated with a long region of DNA sequences that resemble CRISPR. Anthropic has dubbed this system ART, an acronym for array-associated reverse transcriptases. The system is observed in giant phages, large viruses that infect bacteria.
Scientific Perspectives on the Discovery
One AI agent noted, “I can visually identify a matrix of tandem repeats… Is that a CRISPR-like repeat matrix?” The agent also indicated that this system could be a retron, as per the technical report. Both CRISPR and retrons are components of bacterial defence systems, albeit functioning in distinct manners. Retrons can also be harnessed for gene editing; however, they are not as versatile as CRISPR. Therefore, it is unsurprising that Anthropic highlighted the comparison to CRISPR prominently in their announcement.
Seth Shipman, a research associate at the Gladstone Institutes, does not believe that Anthropic has discovered a new CRISPR system but acknowledges the interest surrounding the finding. “What is novel is how they found it, not what it is,” he stated. His lab has utilised retrons to develop gene-editing systems, indicating that such retrons could be employed similarly.
The Challenge of Attribution in AI Discoveries
Identifying new reverse transcriptases typically requires months of manual searches through genomic databases, making Anthropic’s achievement of accomplishing this in a single day remarkable. However, Shipman cautions against attributing the discovery entirely to an AI model: “We should be careful to assert that Claude discovered something autonomously, as there are scientists involved in the study.”
In fact, this particular reverse transcriptase had already been identified by Jason Gill, a microbiologist at Texas A&M University, and his colleagues in a 2021 paper on giant phages. What may be new is that Claude detected surrounding repetitions suggesting that the enzyme might be part of a CRISPR-like system.
“These models excel at pattern recognition, better than a person observing with their own eyes. A person could accomplish all this, but they would need to have a hypothesis in mind,” Gill commented. He added that ART does not appear to have any “obvious relationship” with any known CRISPR system, and it is up to Anthropic to demonstrate that it indeed possesses gene-editing activity.
Future Implications and Considerations
Cong acknowledged the merit of scientists in formulating the ideal problem for Claude to tackle, given its pattern recognition capabilities. While the model may be applicable for similar projects involving large datasets, that does not imply that Claude will suddenly expedite all types of discoveries in the life sciences.
The scientific community is also unable to assess the nature of the information used to train this model, clouding the findings at a time when leading journals advocate for transparency and the sharing of code to enhance reproducibility. This has led one researcher, who has been studying these enzymes, to suspect that the model may have been trained on their unpublished work shared with the public version of Claude.
The Role of Human Verification in Scientific Discovery
Even if Claude were to identify the next major breakthrough in gene editing, it would still require human scientists to validate that hypothesis in the laboratory. In a post on X regarding the discovery, Anthropic’s CEO, Dario Amodei, reflected, “In time, it might be possible for Claude itself to conduct experiments safely by autonomously controlling laboratory equipment.” Should this become feasible, it would be in a distant future. This notion raises numerous questions regarding safety and oversight, and Amodei quickly clarified that Anthropic’s laboratories operate at the lowest biosafety levels, meaning they do not contain anything that poses a significant risk to humans.
Advancements in this field do not occur swiftly. It took 25 years from the initial discovery of CRISPR sequences in bacteria in 1987 until Jennifer Doudna and Emmanuelle Charpentier demonstrated in 2012 that it could be a programmable tool for cutting DNA. Although CRISPR is now ubiquitous in laboratories and is being tested in numerous clinical trials aimed at treating cardiovascular conditions, cancers, autoimmune diseases, and rare disorders, only one drug using this technology was approved at the end of 2023. This illustrates the lengthy path of scientific progress and the challenges in transforming promising basic scientific findings into viable medications.
Exploring the Limitations of AI Discoveries
The limitations of Anthropic’s discovery also raise a more fundamental question: what made CRISPR revolutionary was the identification of an entirely unknown biological system by scientists. It remains to be seen whether any AI model can uncover something useful without precisely knowing what it is searching for. “If you had an AI trained solely on knowledge prior to the discovery of CRISPR and it was still able to uncover it, that would be a far more suitable scenario for testing an AI scientist,” concludes Cong.