September 25, 2026 | AI4Science, Global Healthcare and Cross-Border Capital
On September 23, Anthropic reported that Claude had identified a previously uncharacterized enzyme system in public DNA sequence data. The team named it array-associated reverse transcriptases, or ART. Roughly 950 AI agents — programs that carry out multi-step tasks on their own — ran the search in 21 hours and consumed 210 million tokens, the unit by which models measure the text they process. The numbers make a good headline. The better question is whether AI can repeatedly turn large biological datasets into hypotheses that are testable and that survive the lab.
My read: this is a meaningful result at the discovery and early-validation stage. It does not show that ART is a new gene-editing tool, and it does not show that AI has solved the hard part, which is experimental biology.
According to Anthropic, its agents assembled more than 200,000 reverse transcriptases, identified about 3,500 candidate systems, and selected 20 for deeper analysis. One agent noticed, in phage sequence data, an unusual reverse transcriptase sitting next to a regular array of DNA repeats. Researchers then found a partner gene nearby. Those three elements together form the system now called ART.
The reverse transcriptase itself had appeared in earlier work. The reported contribution is recognizing its association with the repeat array and the partner gene. Anthropic says preliminary experiments show the array is expressed as a set of distinct short RNAs, and that all laboratory work was carried out by human scientists. The findings are described in a preprint — a paper that has not been peer reviewed — and the system’s biological function has not been established.
The repeat pattern is reminiscent of one feature of CRISPR, the gene-editing technology derived from a bacterial immune mechanism. Resemblance is a reason to investigate, not evidence that two systems behave alike. There is no evidence today that ART can be programmed for gene editing, or that it does the “cut, copy, paste” some headlines imply. The accurate description right now is narrower: an interesting, partially characterized biological system whose natural function is still unknown.
Public sequence data and the scientific literature long ago outgrew what any one team can read closely. The bottleneck is deciding which anomaly is worth an experiment. The workflow Anthropic describes has agents survey a protein family, read prior work, reproduce known results to calibrate the method, investigate genomic neighbors, discard weak candidates, and produce a report for human review. Researchers decide what to test, and they run the experiments.
What this changes is the economics of attention. A useful agent does not have to make the final discovery on its own. It can help a scientist search a larger space, articulate why a candidate looks anomalous, and kill unpromising leads before scarce bench time is spent on them. But the metric is not agent count or token spend. It is what share of hypotheses turn out to be reproducible, informative, and worth the cost of testing.
There is also an organizational point. Anthropic built a life-sciences team and its own Bay Area laboratory to connect computational proposals to experiments. That choice matters: scientific software earns trust when its conclusions can be challenged by measurement, not only by another model’s answer. The company says its researchers work under BSL-1 and BSL-2 conditions and do not handle human pathogens.
For AI-for-science platform founders. An impressive demo is an opening, not a moat. The questions I would ask: how were candidates filtered? How many reached the lab? How many failed? Are the results reproducible? What does each validated finding cost? A system that produces ten times more suggestions without improving experimental quality has only moved the bottleneck downstream.
For pharma and medical device teams. The opportunity is broader than finding new molecular systems. Similar methods can organize literature and trial evidence, surface anomalous patterns in testing or manufacturing data, and put specific questions in front of scientists and engineers. Those applications need clean data rights, traceable evidence, and domain experts who can overrule a confident but thin hypothesis. ART itself is not a milestone for any pharma or device company outside the reported study.
For investors. I separate the value into three layers: a better discovery process; a proprietary, repeatable experimental feedback loop; and an eventual product or IP position. Only the first is clearly visible today. ART’s practical use and its commercial economics are both unestablished, and it is early to value the whole chain as if it were proven.
Nearly two decades of healthcare investing taught me to watch one particular moment: when a scientific story moves from possibility to evidence. There are three distinct claims here. Claude helped identify a system researchers had not fully characterized. Preliminary experiments support part of the proposed structure. A useful, programmable biological tool may eventually emerge. The first two are reported results. The third is still a research question.
The news also brings to mind two kinds of companies we have invested in: an innovative drug developer, and a medical device company doing R&D, manufacturing and CDMO — contract development and manufacturing — work. Their potential AI applications differ. A pharma team can use AI to surface scientific or clinical questions for expert review. A device team can turn test records and process knowledge into hypotheses that engineering work can check. These are envisioned applications. They do not indicate that any company has adopted ART, derived revenue from it, or that any confidential project information is involved.
I will be watching three things: whether scientists can establish ART’s natural function; whether there is a functional link between its short RNAs and the enzyme; and whether comparable AI-led surveys produce independent, reproducible findings in other datasets. The first two concern this specific system. The third concerns the larger platform opportunity.
This result is a company’s account of its own early research, currently a preprint, without substantial independent replication. An unusual pattern can be biologically interesting and still never become a useful technology. Large-scale agent searches consume compute and expert review time of their own, and the economics have to be measured against conventional research methods. On the other hand, demanding a market-ready tool immediately misses the real value here: extending the range of what scientists can notice and test.
The balanced conclusion is simple. AI appears to be getting better at helping start biological discoveries. The test is whether better questions lead, again and again, to experiments that are faster, more reliable, and cheaper.
In your field, what do you think AI will change first — asking the questions, or validating the answers? I’d welcome your thoughts.
Scientific findings are drawn from Anthropic’s early report. Commercial applications and investment judgments are the author’s own views. This article is for industry discussion and is not personalized investment, legal or medical advice.
I am a cross-border investor with twenty years of investing experience. I studied at Tsinghua PBC School of Finance and at Stanford Graduate School of Business.
My investing work spans China and the United States, across medical devices, biopharmaceuticals, consumer electronics, innovative retail chains, online education and fintech, and covers equity investment, M&A and restructuring, and cross-border industrial partnerships. Healthcare has always been where I have invested the most, and where I most want to build for the long term.
Past projects include medical device and healthcare companies in China as well as a U.S. drug program that has reached Phase 3. Those experiences keep bringing me back to the same question: how does a genuinely valuable innovation travel from technology and product to real market demand, and how does it attract capital matched to its stage?
My focus on healthcare is also personal. Health experiences in my own family have made it vivid to me what good medicines, devices and care mean for a person and for the people around them. That shapes how I choose projects — I look for innovation that can ease the burden of disease, improve daily life, and bring better care to more people.
Today I pay particular attention to where healthcare, artificial intelligence and capital meet: how AI helps us understand science and improve the efficiency of research and industry; how capital supports valuable technology through validation and into the market; and how cross-border collaboration lets innovation reach more people.
Chunyu Capital AI is where I record what I see on the ground, how I think about investing, and what I am still learning. From Silicon Valley, connecting China and the world, I hope to work alongside researchers, founders and fellow investors on innovation that improves human health and quality of life.