Three deaths.
That's the grim arithmetic that finally forced regulators to narrow the use of Elevidys, Sarepta Therapeutics' $3.2 million gene therapy for Duchenne muscular dystrophy. By November 2025, the FDA had slapped a boxed warning on the treatment and restricted it to a fraction of the patients who might have benefited—only those still able to walk, age four and up. The problem wasn't the therapeutic gene itself. It was that the gene wouldn't shut up.
Fatal acute liver failure, traced to runaway expression of the treatment's transgene in tissues that were never supposed to receive it. At the wrong dose. In the wrong place.
A month earlier, the agency had issued blanket boxed warnings across every approved CAR-T therapy targeting BCMA and CD19—the engineered immune cells that have revolutionized blood cancer treatment. The concern: secondary T-cell malignancies cropping up in patients years after infusion. Different products, different diseases. But investigators kept circling back to the same culprit: the DNA switches controlling therapeutic gene expression were essentially on-off toggles borrowed from viruses, blasting at full volume regardless of context.
Efficacy, it turns out, has a dark side. Tonic signaling. Off-target activation. Cytokine storms that can kill in hours. And in rare cases, the kind of cellular chaos that tips healthy cells toward cancer.
The gene therapy sector—valued at roughly $25 billion and barreling toward $118 billion by 2034—now finds itself confronting what insiders are starting to call a fundamental design flaw. These therapies work. Sometimes spectacularly. But the regulatory elements controlling when and where therapeutic genes express? Those were never built for precision. They were cobbled together from whatever worked in the lab.
Beyond the Hype Cycle
Gene and cell therapies have shed their moonshot status. More than a dozen CAR-T and AAV-based products sit on pharmacy shelves. Novartis extended Zolgensma—its $2.1 million treatment for spinal muscular atrophy—with an intrathecal formulation last November, opening the door to older patients. Intellia Therapeutics is prepping a biologics license application for its in vivo CRISPR program targeting hereditary angioedema. Eli Lilly bet roughly $1.3 billion in 2025 to acquire Verve Therapeutics and its cardiovascular base-editing platform.
Wall Street loves the narrative. But as the field scales, safety signals are getting harder to dismiss.
High-dose AAV9 vectors—the delivery vehicles for many gene therapies—trigger hepatocellular injury and damage to the delicate endothelial cells lining liver sinuses. Ocular gene therapies show elevated rates of uveitis when delivered directly into the eye. And in CAR-T, the choice of promoter doesn't just dial efficacy up or down. It fundamentally shapes toxicity.
Data published in early 2021 demonstrated that weaker promoters—MND or EFS variants—could reduce CAR surface density, dampen the tonic signaling that exhausts T cells, and preserve antitumor activity. Compared to the EF1α workhorses most companies still use, these alternatives looked safer. Yet clinical adoption has been patchy at best, perhaps because switching promoters means re-validating manufacturing processes and potentially triggering new regulatory reviews.
The industry's reflex response has centered on delivery engineering. Companies like Dyno Therapeutics, Voyager Therapeutics with its TRACER platform, and 4D Molecular Therapeutics have wielded artificial intelligence to evolve AAV capsids—the protein shells that package therapeutic DNA—with better targeting and higher transduction efficiency. The pitch: if you can get more gene into the right cells, you can dose lower. Ring Therapeutics is developing anellovirus-based vectors that claim redosability and immune tolerance, addressing one of AAV's persistent weaknesses.
But the cargo's regulatory architecture? The promoters, enhancers, and microRNA target sites that dictate when, where, and how strongly a therapeutic gene fires? That's been treated as plumbing. Functional, yes. But not exactly innovation territory.
The Reckoning
Three forces are now converging to drag regulatory DNA design from academic curiosity into industrial necessity.
First, regulators are walking a tightrope. They want to encourage innovation—the FDA's January 11, 2026 announcement of a "flexible approach" to chemistry, manufacturing, and controls for cell and gene therapies signaled willingness to accept iterative improvements, including tweaks to promoter cassettes. Draft guidance on Platform Technology Designation hints that sponsors with validated synthetic promoter libraries could apply data across multiple programs, reducing the nonclinical grind that slows development.
But that flexibility comes with strings attached. Tighter safety scrutiny. More boxed warnings. Restricted labels. The message is clear: move fast, but get the safety engineering right the first time.
Second, payers are skeptical. A 2025 McKesson survey found that two-thirds of patients view cell and gene therapies as "too experimental." Insurers cite gaps in durability data and evidence that justifies seven-figure price tags. Outcomes-linked payment models hinge on real-world safety, and expression control datasets tied to pharmacodynamic biomarkers could make or break coverage decisions—and the rebate structures that follow.
Third, the science caught up.
Academic labs published proof-of-concept studies in Nature in December 2023 and August 2024, demonstrating machine-guided design of cell-type-specific enhancers validated in fruit fly neurons and human cells. By January 2026, a DNA-Diffusion generative framework in Nature Genetics reported compact, 200-base-pair regulatory elements screened via STARR-seq across nearly 6,000 sequences. DeepMind's AlphaGenome—launched mid-2025 and published in Nature early this year—brought megabase-length sequence modeling to the table, predicting not just gene expression but chromatin state, splicing, and three-dimensional genome contacts.
Foundational models like BPNet, which decodes transcription factor binding syntax, and Borzoi (along with its faster cousin Flashzoi), which predicts RNA-seq coverage from long DNA stretches, created an algorithmic substrate ready for translation into products. Academic groups had figured out how to read the genome's regulatory code. The question was whether biotech could write it.
The Builders
Origin Bio doesn't look like the kind of company that would take on a multibillion-dollar design flaw. Four people. YC Winter 2026. Founded in 2025, based in San Francisco. But CEO Yash Rathod and CTO Malhar Bhide think the sector's been asking the wrong question.
It's not about prediction, Rathod argues. Plenty of models can tell you what a given DNA sequence will do. The bottleneck is generation—designing sequences from scratch that do exactly what you need, in the cells you want, at the expression levels you specify.
Origin's Axis model, unveiled October 8, 2025, claims to be the first system that both generates regulatory DNA and predicts its function. The company reports a 6.7 percent improvement over DeepMind's AlphaGenome on activity prediction and ninefold enrichment of transcription factor binding motifs under high-affinity design prompts. Sequences show the expected hallmarks—HNF4A motifs in liver-derived HepG2 cells, for instance—and more than 72 percent are novel by standard genomic alignment.
Those are company-reported benchmarks, not peer-reviewed data. But the advisory board signals ambition beyond computational biology: Manolis Kellis at MIT and the Broad Institute, Nicole Paulk at Siren Biotech (an AAV specialist), and Rashid Bashir at the University of Illinois and the Chan Zuckerberg Biohub. Origin has acknowledged using Broad's independent "Malinois" model for external validation, a nod to the scrutiny that awaits any startup claiming algorithmic breakthroughs in a field littered with overhyped tools.
They're not alone. Asimov unveiled its AAV Edge platform with AI-designed tissue-specific promoters reporting more than 200-fold selectivity between target and off-target tissues, alongside optimization tools and stable producer cell lines to cut manufacturing costs. AskBio, now a Bayer subsidiary, leverages Synpromics' synthetic promoter technology for bioinformatics-driven tissue control and inducible switches. Senti Biosciences is engineering logic-gated circuits to restrict CAR-T activation to tumor microenvironments—CEO Timothy Lu laid out the SPECIFIC pipeline at last year's Synthetic Biology: Engineering, Evolution & Design conference. MeiraGTx disclosed at the 2024 European Society of Gene and Cell Therapy meeting that it had screened a 240,000-member promoter library via massively parallel reporter assays, hunting for short, potent, tissue-specific sequences.
The line between capsid innovation and regulatory DNA design is blurring. PackGene, an AAV manufacturing vendor, published a preprint in January 2026 describing a spatial control strategy that layers liver-detargeted capsid variants, microRNA regulation, and intramuscular delivery—essentially stacking specificity mechanisms to decouple efficacy from hepatotoxicity. Clinical data back the logic: studies in CAR-T show promoter-driven expression levels modulate tonic signaling and cytokine release, while AAV trials for ocular indications demonstrate that compact human mini-promoters reduce retinal cytotoxicity versus the viral CMV or CBA promoters inherited from early research vectors.
The Trade
The regulatory landscape is settling into a bargain. Tighter safety labeling—boxed warnings, narrowed indications—in exchange for process flexibility. Platform designations. Advanced manufacturing support. Risk-based CMC reviews.
Sponsors who can demonstrate that rationally designed promoters and microRNA detargeting elements reduce off-tissue transgene exposure stand to collect regulatory credit and, perhaps more importantly, payer confidence. The FDA's guidance calling for five- to 15-year long-term follow-up for gene therapies creates an incentive structure that rewards frontloading safety engineering. Integrating vectors—those that splice into the genome—trigger longer surveillance, amplifying the value of non-integrating designs or those with precise expression control that minimize genotoxicity risk.
Three immediate opportunities are crystallizing.
First, oncology CAR-Ts will face mounting pressure to replace strong constitutive promoters with cell-state-responsive or tumor-microenvironment-gated elements. The goal: preserve the tumor-killing punch while mitigating tonic signaling and the specter of secondary malignancies.
Second, AAV programs targeting liver, muscle, and the central nervous system will adopt tissue-specific promoters and microRNA switches to narrow biodistribution and dial down dose-dependent toxicity. This becomes critical as next-generation capsids from Dyno and Voyager—engineered to breach the blood-brain barrier—enter human trials. The paradox is real: higher transduction efficiency multiplies the risks tied to uncontrolled expression.
Third, in vivo CRISPR and base-editing therapies, already under heightened genotoxicity scrutiny, will embed regulatable promoters to pulse editor expression transiently and then silence. Shorter editing windows mean fewer off-target cuts.
Show Your Work
Commercial traction will hinge on validation rigor. STARR-seq, lentiviral massively parallel reporter assays, in vivo transgenic readouts—these remain the gold standard. Academic meta-analyses published in early 2026 caution against assay-specific biases and call for standardized pipelines. Companies touting large proprietary datasets must demonstrate cross-assay concordance and, ideally, perturbation data in primary human cells and animal models. Otherwise, pharma partners and regulators will remain skeptical.
The Alliance for Regenerative Medicine's January 2026 State of the Industry report framed the sector as entering a "disciplined, sustainable growth phase." Translation: the capital bonanza is over. Investors want capital efficiency and manufacturing modernization. Deloitte's 2025 advanced therapies outlook projects an 18 percent compound annual growth rate through 2033, but that trajectory depends on scaling beyond boutique, patient-by-patient models and generating evidence that persuades payers. Evaluate's 2026 preview anticipates a wave of mergers and acquisitions as Big Pharma fills looming patent cliffs with differentiated gene therapy assets.
For founders, the calculus is straightforward. Regulatory DNA is no longer negotiable infrastructure—it's a design axis that can differentiate products on safety, underpin platform strategies, and compress timelines by catching problems in silico rather than in late-stage clinical trials. The infrastructure is there: massive MPRA datasets, generative models, interpretable attribution tools, and a regulatory environment cautiously receptive to AI-guided design.
What's missing is execution. Translating in silico sequences into clinically validated, reproducibly manufactured DNA parts that regulators and payers will trust.
Elevidys delivered a $3.2 million lesson in the cost of imprecision. The industry now has a choice: absorb that lesson proactively, or wait for the next boxed warning to arrive. Given the trajectory of regulatory scrutiny and the commercial stakes, waiting may not be an option.
