The work addresses a gap that has shaped the trajectory of genome engineering for more than a decade. Since the advent of programmable nucleases, researchers have been able to direct double-strand breaks to almost any chosen sequence, and cellular repair machinery can sometimes stitch in a new DNA cassette at the break site. But that approach leans on the cell’s own repair pathways, which are error-prone, unpredictable and often disabling to the very sequences scientists want to insert. Broken DNA is dangerous DNA, and cells treat integration events as injuries to be patched rather than as opportunities for precise reconstruction.

Prime editing, first described in 2019, took a fundamentally different route. Rather than cutting both strands of the DNA double helix, a prime editor pairs a Cas9 nickase — an engineered enzyme that cuts only one strand — with an engineered reverse transcriptase. The editing instructions are carried on a prime editing guide RNA, or pegRNA, which both locates the target site and encodes the new genetic information the reverse transcriptase should write into the nicked strand. Because the process avoids double-strand breaks and does not require an additional donor DNA template supplied in bulk, prime editing has proven remarkably clean for small substitutions, insertions and deletions.

Where prime editing has historically faltered, however, is scale. The reverse transcriptase copies a sequence encoded within the pegRNA itself, and practical constraints on RNA length, delivery and synthesis efficiency have limited the size of the DNA payload that a single prime-editing event can install. For applications in which a functional gene, a large regulatory element or a multi-kilobase cassette must be placed at a defined genomic address, the technology’s ceiling has been a persistent frustration. Complementary systems — including CRISPR-associated transposases and integrase-based platforms — can move larger cargoes, but they typically bring their own constraints on target-site selection, orientation and cargo compatibility.

The new study, led by researchers working at the interface of protein engineering and genome technology, tackles the size problem by rethinking how the donor DNA participates in the reaction. In the reported strategy, termed donor-directed annealing, the genetic cargo is carried on a separate donor molecule rather than being encoded within the pegRNA. The prime editor still performs its characteristic task of opening the target site and synthesizing an exposed stretch of new DNA on the nicked strand, but that newly synthesized sequence is designed to serve as a molecular landing pad. Once exposed, it is complementary to sequences at the end of the donor fragment, and the two single-stranded regions find each other and anneal, drawing the donor cargo into the editing site.

The elegance of the design lies in what happens next. Cellular DNA repair enzymes process the annealed intermediate, ligating the donor fragment into the genome through the natural resolution of the flap-like structure that the prime editor has created. Because the specificity of the event is dictated by sequence complementarity between the editor-generated overhang and the donor terminus, the cell is never asked to recognize a double-strand break or to improvise an end-joining reaction. The authors report that this mechanism allows fragments substantially larger than the payloads accessible to conventional prime editing to be incorporated at defined loci, with precision determined largely by the programmed overlap rather than by stochastic cellular repair.

From a biochemical standpoint, donor-directed annealing converts what has been an intramolecular copying reaction into a hybridization-guided assembly step. Conventional prime editing is, in essence, a controlled form of DNA synthesis: the pegRNA templates every base that the reverse transcriptase installs. The new method retains that templated synthesis for a short anchoring sequence but delegates the bulk of the payload to a separate donor, which can be produced synthetically or by standard cloning at lengths far beyond what a pegRNA can encode. The trade-off is that the donor and the pegRNA must be co-delivered and their sequences coordinated, but the payoff is a system in which cargo size is no longer bound to the physical limits of the guide RNA.

The practical implications extend across both research and therapeutic arenas. In basic biology, the ability to drop large regulatory modules, reporter constructs or engineered gene circuits into precise genomic contexts would simplify experiments that currently require laborious screening of random integration events. In medicine, many inherited disorders are caused by mutations in genes that are too large, too structurally complex or too mutationally diverse to be addressed base by base. Delivering a corrected copy of a gene, or a functional cDNA, into its native locus under the control of endogenous regulatory elements — rather than scattering it randomly through the genome as viral vector gene therapy does — remains the aspirational gold standard, and integration strategies of this kind are among the most credible paths toward it.

The reported system also speaks to a recurring theme in the genome-editing literature: the value of avoiding double-strand breaks altogether. Studies across multiple cell types have associated double-strand-break-based editing with p53 activation, chromosomal rearrangements and large unintended deletions, concerns that are particularly acute for ex vivo cell therapies and in vivo applications alike. By building integration on a nicking enzyme and sequence-programmed annealing rather than on blunt-ended break repair, the approach aligns with the field’s broader movement toward editing chemistries that leave the genome’s integrity machinery largely undisturbed.

As with any genome-engineering advance, several questions will shape how the technique matures. The efficiency of integration across different genomic loci, cell types and species will need systematic mapping; cargo lengths will have practical ceilings set by delivery vehicles rather than by chemistry; and off-target activity — a concern for any nuclease-fusion system — will require careful characterization at both the sequence and chromosomal level. The study’s authors report encouraging precision at the sites they examined, and the strategy’s dependence on designed sequence complementarity offers a built-in specificity checkpoint that many integration methods lack. Independent replication and optimization in therapeutically relevant primary cells will be the next milestones.

What the work illustrates most clearly is how quickly the conceptual boundaries of genome editing continue to move. In barely a decade, the field has progressed from cutting DNA at chosen addresses, to rewriting individual letters of the genetic code, to contemplating the programmed installation of whole functional modules at will. Donor-directed annealing extends prime editing’s core strengths — precision, minimized DNA damage and programmability — into a size regime that those strengths had not previously reached. If the method’s efficiency and reliability hold up as it is tested more broadly, large-fragment insertion could shift from a heroic, low-yield exercise to a routine operation in the genome engineer’s toolkit, with consequences for drug discovery, synthetic biology and, ultimately, the treatment of diseases that small edits alone cannot fix.

Subject of Research: Precise integration of large DNA fragments into genomic target sites using prime editing with donor-directed annealing

Article Title: Precise genomic integration of large DNA fragments by donor-directed annealing using prime editing

Article References: Jung, H., Jeong, B., Kim, Y.-W., Jung, C., Lee, S., Uhm, H., Kim, H., Oh, Y. E., Park, Y., Lee, Y., Kang, M., Im, H. W., Kim, D., Lee, S., Kim, Y., Choi, K., & Bae, S. (2026). Precise genomic integration of large DNA fragments by donor-directed annealing using prime editing. Nature Biotechnology. https://doi.org/10.1038/s41587-026-03301-2

Image Credits: AI Generated

DOI: 10.1038/s41587-026-03301-2

Keywords: prime editing, genome editing, CRISPR, gene therapy, DNA integration, pegRNA, reverse transcriptase, Cas9 nickase, large DNA fragments, donor-directed annealing, biotechnology, genetic engineering

For decades, mRNA technology was relegated to the outskirts of mainstream medicine. Although recognized in principle for its potential to deliver coded instructions for proteins into a patient’s cells, it needed years of trial and error to mature. Then came the extraordinary acceleration offered by COVID-19 vaccine development, where mRNA-based vaccines from firms like Moderna and BioNTech/Pfizer demonstrated that these treatments could indeed be developed and deployed in record time. But what Larry Ellison is suggesting goes far beyond the principle that mRNA can be used to mount immune responses. He envisions a future in which we create an mRNA therapy specifically for each patient’s cancer profile—meaning that no two people’s vaccines need be exactly alike. You wouldn’t just have a “generic” immunization against, say, a subtype of breast cancer or lung cancer. Instead, medical labs, assisted by AI software, would map the precise mutations or surface markers in a patient’s tumor cells, then create a unique mRNA blueprint that instructs that individual’s immune system to identify and target the malignant cells. If you imagine multiple patients, each with a different set of tumor mutations and immunological nuances, the idea is that thousands or even millions of unique mRNA sequences could be generated and tested or, at the very least, validated in silico within days. The AI part is crucial because the scale of computations needed to design such tailored vaccines is mind-boggling.

What sets Ellison’s statement apart is not merely the mention of AI in medicine, for that is no longer revolutionary. Instead, it’s the bold claim that the entire pipeline—from diagnosing a patient’s tumor signature, to figuring out the relevant immunological targets, to coding an mRNA therapy, to physically manufacturing it—could be done in under two days. Whether that is 48 hours from the moment a patient’s blood or tumor sample is taken, or from the time the physician presses “go” on a software platform, is unclear. Yet even the very idea of compressing the vaccine design cycle to two days marks a quantum leap from the norm. Typically, it can take weeks or months just to finalize the design of a novel therapeutic, let alone test it for safety or efficacy. So the notion here is that specialized AI software, presumably fed by colossal data sets, will automatically generate a new mRNA sequence that instructs the patient’s cells on what cancer-related proteins to target. The advanced robots or “lights-out” manufacturing lines, as some call them, then deposit the materials into a microfluidic system that produces small, personalized batches of vaccine. The entire process is so frictionless, so automated, that it can happen in hours, not weeks.

We know that mRNA vaccines are agile in principle—once you have a certain packaging technology, like lipid nanoparticles, the only change you need is the specific code in the RNA. But we also know that bridging from a conceptual framework to a standard medical procedure involves an enormous array of challenges. Biopharmaceutical regulation, for instance, typically requires any new therapy to go through a rigorous clinical trial process, ensuring it is both safe and effective. So, does Ellison’s scenario foresee a streamlined or even partially automated regulatory structure that can handle a mass of new, personalized therapies? Are we about to see advanced computational models and in vitro microfluidic tests that can all but guarantee the safety of such a vaccine before it is administered to the patient? We might imagine advanced AI systems simulating immunological responses in silicon with such fidelity that real-world trials become less arduous. But as of now, we do not have that level of official acceptance for preclinical computational evidence. If we are heading this direction, it would mean the entire regulatory system, from the FDA to the EMA and all other jurisdictions, would have to evolve to accommodate near-real-time generation of immunotherapies. Some might see that as pure fantasy; others see it as the inevitable future.

Yet there’s more to “people not understanding what this means” than just the timeline for design or regulatory complexities. The statement implies that if you can design a custom mRNA vaccine in two days, you’re basically bringing Moore’s Law–style iteration to the fight against cancer. You might vaccinate a patient with a certain design, evaluate the immune response in real-time, gather data about which mutated peptides or antigens elicited the best T-cell infiltration. Then you tweak the design, re-run it, and generate the next batch. This iterative cycle of “design-test-redesign” might occur at breakneck speed. The synergy between AI’s algorithmic power and the swift manufacturing pipeline merges to create a personalized, dynamic therapy that evolves with the tumor. Suppose the tumor acquires new mutations or reverts to a new strategy to evade the immune system; in principle, you could spool up a fresh vaccine code to block the new malignant variant. This near-term future, if realized, transforms cancer management from a static “Here’s your chemotherapy or targeted therapy regimen, hope it works” approach to an adaptive “We’ll chase the cancer and keep updating your therapy as if we’re rolling out software patches.” That’s radical—like turning the entire fight against cancer into a constant arms race at the molecular level.

One might also wonder about the role of Oracle here. Ellison’s company is known primarily for database systems, enterprise software, and cloud services, but in the last few years, it has pivoted somewhat to focus on health data and analytics. Conceivably, Oracle might be the data platform that integrates all the genomic and clinical records. The combination of patient data, advanced analytics, and AI could indeed allow for that dynamic synergy. That Ellison himself is heralding this future might be read as a sign that Oracle sees a big opportunity in health-care data management for personalized medicine—one in which the cost of storing and processing large-scale genomic data is trivial compared to the potential advantages in patient care.

Of course, the public reaction to the idea of AI designing personalized mRNA therapies may be complicated by concerns about data privacy, algorithmic biases, or errors that slip through an automated pipeline. We need not only to trust AI to design a therapy but also to trust that the code it generates is robust enough not to harm the patient. The fiasco scenario would be an AI that incorrectly identifies a normal protein as a target, leading the vaccine to trigger an autoimmunity crisis. This is where advanced AI verification and interpretability become crucial. Additionally, the system must ensure that data used to train these models covers the huge genetic diversity of human populations, because a solution that works for one set of genotypes may not work for another. If the AI is solely trained on the data from large medical centers in North America or Western Europe, we risk ignoring the particular genetic variants in, for instance, sub-Saharan Africa or East Asia, leading to suboptimal or unsafe designs in those populations. Hence, to fully realize Ellison’s vision, we must push for global data-sharing, or at least a set of robust, widely representative training sets that can handle the entire diversity of the human genome.

The mention of “making them robotically in 48 hours” also underscores the larger trend that manufacturing is becoming more agile, smaller-scale, and automated. If you have fully robotic labs that can do everything from mixing reagents to packaging the final product, you might indeed pump out custom vaccine vials for a single patient. But that also implies an infrastructural shift. Are these production lines likely to exist in major medical centers, or could they be deployed in smaller labs across the world? The logistics behind shipping raw reagents, guaranteeing sterility, controlling for quality assurance, delivering final products, and training staff to operate such advanced robotics could be daunting. For countries that have underdeveloped health-care systems, the gap might become even more glaring. Possibly, though, the availability of advanced robotics might eventually reduce costs so that remote areas can “print” these therapeutics locally. Or, these specialized manufacturing sites remain in large advanced hubs, and the final products get shipped or flown to the patient. One can see the complexities branching out in every direction.

However, none of these complexities seem to deter Ellison’s optimism. His statement, if it truly captures the direction that Oracle and other tech titans are heading, illuminates the scale of ambition. We are at the point that the synergy among big data, machine learning, genomic science, and advanced biotechnology can yield leaps forward that might have felt unattainable a decade ago. People who dismiss these claims might say, “It’s hype; 48 hours is a marketing slogan.” But there is also a strong possibility that we are seeing the early signals of a disruptive approach. We might see a pilot program in the next few years where a small subset of cancer patients with a specific tumor type receive AI-designed mRNA vaccines. Early results might be uncertain, but the iterative process of improvement will refine both the AI’s accuracy and the manufacturing pipeline. If, after a few cycles, the outcomes show improved survival or fewer side effects than conventional chemo or immunotherapy, the impetus to expand the pilot becomes immense.

At a conceptual level, it’s reminiscent of how, in the late 1990s, only a handful of visionaries could fathom how the Internet might transform commerce and communication globally. Now, with personalized mRNA vaccines designed by AI, we might witness a transformation in health care so profound that it shifts from diagnosing diseases to systematically customizing a cure for each person. The possible benefits for cancer treatment alone are staggering, but we can extrapolate to other maladies—infectious diseases, autoimmune disorders, or even certain forms of degenerative conditions. In principle, once you master the puzzle of coding instructions into cells, you can do it for nearly any protein-based therapy. Moreover, the dynamic, iterative approach might open pathways to “always current” therapies that adapt to a pathogen’s or tumor’s mutations in near real-time, effectively curtailing the race that disease processes typically run uncontested.

There will be ethical ramifications, too. Not only who pays for such technology, but who gets it. Does this become something available solely to the wealthy who can afford custom immunization? If the process truly scales and is driven by mostly robotic labor, maybe the cost can drop dramatically. The dream scenario is that once the pipeline is standardized, the marginal cost of generating each new vaccine is minimal, so you can produce it cheaply for millions of people. But this dream depends on large-scale adoption, supportive regulation, robust oversight, and indeed a shift in how we conceive of health care, from broad-spectrum mass-market therapies to individually tailored ones.

All in all, Ellison’s remarks carry the power to astonish because they cut to the heart of what might be the greatest aspiration of modern medicine: the capacity to defeat, or at least substantially tame, cancer. Many experts already foresee a day when we treat cancer as a manageable chronic condition, thanks to advanced immunotherapies. The arrival of AI-driven, mRNA-based solutions speeds that timeline in ways that can be jarring to those used to the plodding pace of medical research. At the same time, one must temper the euphoria with caution, bearing in mind the regulatory labyrinth, the reliability of AI’s predictive capabilities, and the sheer engineering complexity of mass customization in biotech. Realizing these aims will require visionary leadership, huge investments, and perhaps a decade or more to refine the pipeline to the point that it is widely deployed. Nonetheless, Ellison’s statement signals that major players in the technology sphere intend to push vigorously in that direction.

Whatever shape it ultimately takes, the possibility that AI will design an mRNA vaccine for each patient’s unique cancer signature, then have it robotically produced in under two days, is a scenario that redefines the boundaries of what we believed was possible in health care. It also reframes the role of large data management corporations like Oracle, showing that the interplay of data, AI, cloud computing, robotics, and pharmaceutical science is rapidly converging. It may be that we look back in a few years and marvel at how quickly personalized medicine advanced once these technologies converged. Or we might find that the hype outstripped reality, that regulatory constraints and real-world complexities led to a more modest revolution. The only certainty is that the conversation has changed. The pronouncements of Larry Ellison have become a rallying cry for an era in which custom vaccines—once an almost utopian idea—are to be viewed not as a remote possibility but as an impending milestone. And it underscores the sense of astonishment and perhaps the sense of hope: if this truly works, we might say farewell to the notion that cancer is unstoppable, and greet an era in which therapy is swiftly shaped to each patient’s genome, delivered by precise robots, and iterated at near-lightning speed. That is indeed enough to leave one speechless.