DNA fragments jumping between cells & AI co-scientists speeding drug ideas - News (May 20, 2026)

In biomedical science, a striking new finding from UT Southwestern suggests human cells may be less genetically isolated than we assumed. Researchers report that large chunks of genomic DNA can move directly from one cell to another through brief cell-to-cell connections, reach the nucleus, and even integrate into the recipient genome. In live imaging, they saw Y-chromosome fragments travel from male cells into female cells, where male-specific genes then became active. If this holds up broadly, it could change how scientists think about genomic evolution inside tissues—and potentially how cancers pick up major chromosomal changes after stresses like chemotherapy or radiation.

Staying with science, two separate teams reported “AI co-scientist” systems in Nature that use multiple specialized AI agents to move faster through early-stage discovery work—like brainstorming hypotheses, scanning literature, proposing experiments, and analyzing results. Google DeepMind’s system was tested on drug repurposing ideas for acute myeloid leukaemia, generating candidate medicines within hours. Human researchers picked a handful to test, and several showed promising early effects in cultured cells. A second system, Robin from the nonprofit FutureHouse, tackled dry age-related macular degeneration and flagged an existing glaucoma drug, ripasudil, as a potential lead—along with suggested follow-up assays. The big takeaway is speed: these workflows may compress parts of discovery from weeks or months into hours or days. The caution is just as important: early cell results are not the same as a validated therapy, and many leads fail when testing gets tougher.

Another Nature paper points to a different bottleneck AI might loosen: scientific coding. A Google-led group, co-led by Harvard’s Michael Brenner, unveiled a system called Empirical Research Assistance, or ERA, that can automatically generate and refine research software for tasks where you can score performance numerically. Instead of writing one version of a program and tweaking it by hand, the system rapidly explores many variations and keeps what works. In demos, it produced COVID-19 hospitalization models that outperformed well-known baselines, improved ways to combine single-cell biology datasets, and sped up neural-activity modeling that normally drags on for weeks. If tools like this generalize, they could shift researchers away from endless tuning and toward higher-level questions—what to test next, and what results actually mean.

In China, startups are pushing brain–computer interfaces toward real-world products, combining neural sensors with large language models to interpret brain signals more accurately. One company, NeuroXess, reported early trial results where an implanted system helped a man with a spinal-cord injury move a cursor and control home devices. It also claims it demonstrated real-time Mandarin decoding in a person with epilepsy. The Chinese government is backing the field with explicit goals for breakthroughs by 2027 and multiple world-class BCI firms by the end of the decade. The opportunity here is obvious—restoring movement or communication—but so are the concerns: neural data is deeply personal, and as AI learns from users’ signals, questions about consent, long-term storage, and secondary use get more urgent.

In the AI industry itself, a very human story—egos, money, and governance—played out in court. The California trial between Elon Musk and OpenAI chief Sam Altman ended with a verdict that largely favored OpenAI, with Musk described as losing on a technicality. Beyond the personalities, the signal to the market is that aggressive competition and profit-seeking in AI are increasingly being treated as standard business behavior, not a betrayal of early “for humanity” messaging. It may also clear some air for fundraising and longer-term corporate plans at OpenAI, while leaving bigger governance questions unresolved—and possibly further denting public trust by reinforcing the sense that the industry is steered by a small circle of rivals.

On the fight against synthetic media confusion, Google says its SynthID watermarking system has now labeled an enormous amount of AI-generated content—across images, video, and audio—and it’s expanding beyond Google’s own tools. The notable change is the partner list: Google says OpenAI plans to add SynthID to its image generation, Nvidia to its Cosmos models, with others like Kakao and ElevenLabs also adopting it. Google is also pairing watermarking with the C2PA metadata standard, so content can carry richer “where did this come from” labels through editing and sharing. None of this makes detection perfect—unwatermarked open models still exist—but wider adoption could make it significantly easier for ordinary users and platforms to flag AI media without turning it into a daily guessing game.

Singapore is leaning harder into becoming a global testbed for AI deployment. The country signed separate AI agreements with Google and OpenAI aimed at accelerating use across public services, healthcare, education, and business. OpenAI is set to establish an Applied AI Lab in Singapore—its first outside the US—alongside a major investment and hiring plans. Google’s partnership focuses on training and research collaboration, including healthcare-oriented work. For Singapore, the strategic value is clear: attract talent, run pilots at national scale, and position itself as a neutral hub for building and validating AI systems that can be exported globally.

Turning to security and geopolitics, Ukraine has released imagery of what it says is its first domestically developed glide bomb, now through trials and ready for combat use. Developed under the government-backed Brave1 initiative in roughly a year and a half, it’s designed to strike targets dozens of kilometers behind the front line. Ukraine has highlighted integration on a Su-24 so far, with talk of eventual compatibility with aircraft like F-16s after certification. The significance is practical: as air defenses near the front stay dense and supplies of Western precision weapons face uncertainty, a homegrown standoff weapon gives Ukraine more control over scale, timing, and targeting flexibility.

In Europe, NATO officials expect the United States to announce it will reduce the military capabilities and forces it makes available to NATO in a crisis or wartime scenario, as part of a broader shift in priorities. Reports suggest this doesn’t immediately mean fewer US troops stationed in NATO countries day to day, but it could still translate into less material support when it matters most—reinforcement, logistics, and the rapid surge of equipment. For European allies, it’s another nudge toward filling gaps faster, not just spending more, as deterrence planning depends on what can show up quickly in a real emergency.

And finally, in consumer tech, Apple is expected to add new privacy controls to a revamped Siri in iOS 27, including options to automatically delete Siri conversation history after a set period. The idea is to give users clearer control over how long sensitive AI interactions stick around, while Apple continues to market on-device processing and tightly managed cloud handling as a differentiator. The trade-off is familiar: less retained history can mean less personalization and weaker continuity. But as people increasingly share personal, legal, or health-related details with assistants, default data minimization is becoming a competitive feature—not just a nice-to-have.