Anthropic’s Claude Phage Finding—and the New Pace of Research

AI is beginning to change the pace of scientific research. Anthropic's new phage study offers a glimpse: Claude searched a huge biological dataset, followed an anomaly and brought scientists a question worth testing. That kind of search consumes scarce expert time. AI makes it possible to run many such searches, compare the leads and focus lab work where it has a chance to reveal something new. The larger question is what becomes possible when search, analysis and experiment planning accelerate together. [1][1]

From prediction to research loops

The progression has been fast. In 2020, a model screened compounds and picked out halicin as a possible antibiotic; researchers tested it in bacteria and mice. In 2021, AlphaFold 2 made accurate protein structures far easier to predict. AlphaFold 3 later extended that approach to interactions involving proteins, DNA, RNA and small molecules. These were powerful tools for defined tasks. They changed what researchers could examine, but the scientist still decided what to investigate next. [3–5][3–5]

Then the tools began to participate in the next decision. A materials-science A-Lab linked software planning to robotic experiments and used each result to choose the next synthesis attempt. A later correction narrowed some of its success claims, but the important shift remained: prediction and experiment were connected in a repeating loop. [6,7][6,7]

In 2026, Google's Co-Scientist generated and refined biomedical hypotheses, including drug candidates tested in leukaemia cells. Robin, a separate research system, proposed candidates for dry age-related macular degeneration, analysed results from human cell experiments and suggested a follow-up mechanism to investigate. Human scientists still ran the experiments. The AI systems were doing more than retrieving papers: they were helping decide what to test and what the results might mean. [8,9][8,9]

Where experiments happen entirely on computers, the loop can go further. An AI Scientist system generated ideas, wrote code, ran experiments and drafted papers in machine-learning research; one paper passed review at a conference workshop. This is a glimpse of a full digital research cycle. Biology adds physical experiments, but that makes the link between AI and lab work more valuable, not less. [10][10]

In a few years, AI has moved from predicting an answer to helping choose the next question, analysing the result and starting another round. That is the acceleration to watch. [3–10][3–10]

What Claude added

Anthropic gave Claude a broad search across related protein sequences. Over 21.5 hours and 949 agent sessions, one session noticed a repeated DNA pattern beside a phage gene for reverse transcriptase, an enzyme that copies RNA into DNA. The gene and nearby non-coding DNA were already known. The combination of a repeat array and a neighbouring partner gene had not been described. Human follow-up detected short RNAs from one array. The system's function is still open. [1,11][1,11]

What matters for AI-assisted research is how the lead emerged. The agent moved beyond a list of proteins, inspected the surrounding DNA, recognized an unusual arrangement and made a case for testing it. Scientists could then focus their lab work on that lead. Ten repeat searches did not find the array again, so this is not yet a reliable discovery rate. It is a concrete example of an AI system expanding the surface of biology a team can explore. [1][1]

What further acceleration could make possible

Imagine applying that pattern to many research questions at once. Agents could compare enormous genomic, molecular and clinical datasets, look for results that do not fit the prevailing explanation, and propose the experiment most likely to distinguish competing ideas. Researchers would spend less time finding the first lead and more time deciding which leads deserve scarce lab capacity. [1,8,9][1,8,9]

The next step is to connect those agents to repeatable experiments. In areas with well-defined assays, a system could propose a test, receive the data, update its model and choose the next test. A-Lab and Robin show pieces of that loop already working. Better models could improve each decision; better laboratory interfaces could shorten the wait between decisions. The gains would come from more useful cycles, not simply more generated text. [6–9][6–9]

For longevity research, this could change how we work across fragmented evidence. An AI-assisted team might compare aging signals across tissues, species and interventions, identify where findings disagree, and design experiments that resolve the disagreement. For a biotechnology company or research service, the advantage may lie in linking data, experimental capacity and scientific judgment into a continuously improving workflow. [1,8,9][1,8,9]

In his 23 September X post about Anthropic's ART preprint, Anthropic CEO Dario Amodei argues that biological AI could improve as rapidly as AI for mathematics. His more practical point is about throughput: AI may send many more promising candidates into research even when the time needed for any one clinical trial does not change. That is a forecast, but it captures something important. Faster discovery need not mean every step becomes faster. It can mean exploring more avenues in parallel and sending stronger candidates into the slow steps. [1,2][1,2]

The new research bottleneck

The opportunity is large precisely because so much of science still runs in a narrow sequence: find an idea, secure time and tools, run one experiment, interpret it, then decide what comes next. AI can widen the search and help plan the next round. To turn that into progress, research organisations will need experiments, data systems and independent checks that can keep up. AstaBench still finds gaps in agents' ability to manage complete research workflows; Anthropic's missed reruns make the same point in this case. [1,12][1,12]

The right measure is not how many hypotheses an agent produces. It is how many findings survive testing, how quickly they do so and what they cost to establish. If those numbers improve, AI will let scientists investigate questions that are currently too broad, connected or expensive to pursue. That is the deeper promise: not just a faster version of today's research, but a larger map of biology within reach. [1,8–10][1,8–10]