DNA fragments jumping between cells & AI co-scientists speed drug discovery - News (May 21, 2026)

Let’s start with that unexpected genetics story. Researchers at UT Southwestern report evidence that large fragments of genomic DNA can move directly from one human cell to another. In their observations, DNA-containing structures formed after damage or division errors, then traveled through brief cell-to-cell connections. Some of that DNA appeared to reach the nucleus of a neighboring cell, become active, and even persist across multiple rounds of cell division. One striking example: Y-chromosome fragments moving from male cells to female cells, with male-specific genes turning on. If this holds up broadly, it complicates the old assumption that neighboring cells’ genomes evolve entirely independently—and it could reshape how scientists think about cancer cells changing after stresses like chemotherapy or radiation.

Staying in biomedical research, two new “AI co-scientist” systems reported in Nature are getting attention for how they organize AI into teams of specialized agents. Google DeepMind’s system was tested on drug repurposing for acute myeloid leukaemia, generating candidate medicines within hours. Researchers selected a handful to test, and several showed promising early effects in cultured cells. A second system from the non-profit FutureHouse, called Robin, tackled dry age-related macular degeneration—coordinating literature review, experimental planning, and analysis—and flagged the glaucoma drug ripasudil as a potential candidate, along with suggested follow-up assays. The key point: these multi-agent workflows may compress parts of early-stage discovery from weeks or months into hours or days, though the researchers and outside experts are clear that early lab signals often fail under tougher validation.

Another AI-for-science headline: Google researchers, working with Harvard’s Michael Brenner, described a system called Empirical Research Assistance—ERA—that can generate and refine scientific software. This focuses on problems where you can score performance numerically, like prediction and modeling tasks. Instead of writing code once and hoping for the best, ERA tries many variations quickly, tests them, and iterates—sometimes producing models that outperform strong human baselines. In demos, the team says it built COVID-19 hospitalization models that beat leading CDC models, improved ways to integrate single-cell RNA sequencing data, and sped up zebrafish neuron-activity modeling that would normally take weeks or months. The big takeaway is practical: if tools like this become reliable, researchers may spend less time wrestling with code and more time deciding which scientific questions are worth asking next.

On the cancer front, UCLA researchers reported a preclinical CAR-T approach aimed at one of the toughest targets: glioblastoma. Their strategy used CAR-T cells that go after a glioblastoma-linked marker, IL-13Rα2, but with an added twist—these cells also release immune-activating signals designed to pull more of the body’s immune system into the tumor. In mouse models with intact immune systems, that broader activation helped control tumors even when not all cancer cells carried the exact target the CAR-T was designed for. Because one of the immune signals, IL-12, can drive dangerous inflammation, the team also tested safety-minded design changes that appeared to reduce toxicity while keeping anti-tumor effects in mice. This is still preclinical, but it points to a potential path for solid tumors that have historically resisted CAR-T therapies.

Now to trust and authenticity online. Google says its SynthID watermarking has labeled a huge amount of AI-generated media—images, video, and audio—and the bigger news is that it’s expanding beyond Google’s own models. Partners are expected to adopt it across more AI tools, and Google is also pushing the C2PA metadata standard so platforms can attach clearer “provenance” information about how media was created or edited. Google’s pitch is straightforward: watermarking only becomes truly useful if multiple major providers participate. The limitation, of course, is that open models and unwatermarked content will still exist—so this is about improving the odds of detection at scale, not guaranteeing it every time.

In a related AI milestone claim, OpenAI says a new general-purpose reasoning model produced an original proof that disproves a discrete geometry conjecture dating back to 1946, associated with Paul Erdős. This announcement lands differently because OpenAI took heat previously after a high-profile claim about solving Erdős problems didn’t hold up the way people initially thought. This time, OpenAI pointed to supportive remarks from established mathematicians, including some who criticized the earlier episode. The broader significance—if the proof withstands community scrutiny—is that it suggests AI systems may be getting better at sustained, multi-step reasoning in areas where correctness is unforgiving and can be checked.

Zooming out to national AI strategy: Singapore signed separate agreements with OpenAI and Google to accelerate AI use across public services, healthcare, education, and business. OpenAI says it will invest more than three hundred million Singapore dollars and set up an Applied AI Lab in Singapore—its first outside the United States—with plans to hire hundreds over the coming years. Google’s partnership emphasizes training and research collaboration, including work tied to healthcare. The signal here is that Singapore is positioning itself as a high-talent, globally connected place to build and test AI systems—while also pushing for safer deployment.

Turning to geopolitics, President Donald Trump said he plans to speak with Taiwan’s president Lai Ching-te about a potential US arms sale. That would break decades of diplomatic practice—since Washington switched formal recognition to Beijing in 1979, US and Taiwanese leaders have generally avoided direct talks at the top level even while the US supports Taiwan’s self-defense. China condemned the idea and is reportedly linking broader US-China engagement to whether an arms package goes forward. The stakes are high because even symbolic moves—like a direct call—can shift perceptions of deterrence and raise the temperature across the Taiwan Strait.

And in Europe, NATO officials expect the US to announce it will reduce the military capabilities and forces it makes available to NATO in a crisis or wartime scenario. This doesn’t necessarily mean an immediate cut to the roughly seventy-six thousand US troops currently stationed across NATO territory, but it could still matter in a major contingency—because it’s the surge capacity, logistics, and material support that shape real readiness. The development adds pressure on European allies to fill gaps faster, especially if US priorities continue moving toward an “America First” posture.

Finally, a reminder that AI’s social impact isn’t only about science and geopolitics—it’s also about everyday trust online. UC San Diego researchers report that modern large language models can pass a classic three-party Turing test when they’re prompted to adopt a specific human persona in live chats. In their experiments, one model was judged to be the human most of the time, and another performed roughly on par with real human chat partners. Without persona prompting, the ‘human’ ratings dropped noticeably. The implication is less about whether machines are “intelligent,” and more about how easily they can appear convincingly human—raising obvious risks for fraud, manipulation, and social engineering.