The boxed warning arrived on November 14, 2025—one of those regulatory announcements that tends to rattle through the biotech industry long after the initial press release fades. Sarepta Therapeutics' Elevidys, a gene therapy for Duchenne muscular dystrophy, had been linked to patient deaths and liver toxicity severe enough to prompt the FDA's most serious cautionary language. For the families who'd pinned their hopes on the treatment, the news was devastating. For the broader gene therapy field, it was something else: a reminder that getting therapeutic genes into cells is only half the problem.
The harder part, it turns out, might be controlling what those genes do once they arrive.
This question of expression control—how strongly a therapeutic gene fires, where it fires, when it shuts off—has emerged as perhaps the defining challenge for an industry that's grown from speculative science to a $36.5 billion market in 2025. Astute Analytica projects that figure could balloon to $183.1 billion by 2035, assuming the field can navigate a gauntlet of technical and safety hurdles. Much of the industry's energy has focused on delivery: engineering viral capsids that slip past immune defenses, designing lipid nanoparticles that release their cargo at just the right moment.
But a quieter revolution has been brewing upstream, in the regulatory DNA sequences that function as the molecular equivalent of volume knobs and on/off switches. These sequences—promoters, enhancers, the whole alphabet soup of regulatory elements—determine whether a therapeutic gene whispers or shouts, whether it activates in heart muscle or liver tissue, whether it sustains expression for years or burns out in weeks.
Get it wrong, and the consequences range from merely ineffective to actively dangerous.
The Tonic Signaling Problem
Consider CAR-T cell therapy, where immune cells are reprogrammed to hunt cancer. The treatment has produced some genuinely remarkable remissions in blood cancers, the kind of results that seemed impossible a decade ago. But early CAR-T constructs often used promoters borrowed from viruses—sequences that had evolved to drive maximum protein production during viral replication. These very strong promoters did their job almost too well, plastering CAR receptors across the engineered cells' surfaces.
The result? A phenomenon called tonic signaling, where the therapeutic cells effectively exhaust themselves before they ever reach the tumor. Research published in 2021 demonstrated that switching to milder promoters—variants like EFS or MND-class sequences—could dial down that cytokine release while maintaining the cells' tumor-killing capacity. It was a rare instance of less being more.
The spatial control problem shows up differently in AAV gene therapies. Take Duchenne muscular dystrophy, where the goal is delivering a truncated dystrophin gene to skeletal muscle. Promoters like MHCK7 have been engineered to concentrate expression in the target tissues: skeletal muscle, cardiac muscle, the diaphragm. Minimize expression everywhere else, the thinking goes, and you minimize the immune responses that can derail treatment. A 2020 JAMA Neurology study of an AAV micro-dystrophin therapy illustrated this tissue-specific approach in pediatric DMD patients, though the field is still learning which patterns of expression the immune system will tolerate.
For years, the industry has made do with a limited toolkit of natural promoters, most of them borrowed from viruses or housekeeping genes that cells use for basic maintenance. These sequences work, after a fashion. But they evolved for viral replication or cellular upkeep, not for the delicate balance required in therapeutic applications. As safety signals accumulate and regulators sharpen their scrutiny, that genetic hand-me-down approach looks increasingly inadequate.
Enter the AI Sequence Designers
Which brings us to a cohort of startups now applying artificial intelligence to design regulatory DNA from scratch. Origin Bio, a San Francisco outfit that emerged from Y Combinator's Winter 2026 batch, offers a case study in this nascent field. The company's four-person team, led by CEO Yash Rathod and CTO Malhar Bhide—both recent graduates of UIUC's computer science program—describes its mission as building "AI for safer cell and gene therapies." It's the kind of pitch that sounds almost comically broad until you dig into the technical details.
On October 8, 2025, Origin unveiled Axis, a multifunctional AI model designed to both generate novel regulatory DNA sequences and predict their activity. According to the company's announcement, Axis notched a 6.7 percent average improvement over DeepMind's AlphaGenome on regulatory activity prediction tasks. When prompted to design sequences with high transcription factor affinity, the model demonstrated up to nine-fold enrichment in the relevant binding motifs. The training data came from ENCODE V4's candidate cis-regulatory elements—a sprawling dataset covering more than 2.3 million regulatory elements across 1,888 human cell types.
Whether those benchmark improvements translate to therapeutic benefit remains, of course, an open question. Computational biology has a rich history of promising in-silico results that wither under experimental scrutiny.
Origin's scientific advisory board suggests the founders understand the complexity involved. Manolis Kellis from MIT and the Broad Institute brings computational genomics expertise. Nicole Paulk contributes knowledge of AAV gene therapy's practical challenges. Rashid Bashir, affiliated with UIUC and the CZ Biohub Chicago, adds bioengineering perspective. It's a roster designed to bridge the gap between computational predictions and clinical reality, though that gap remains formidable.
The company's pitch centers on rebalancing efficacy and safety. For CAR-T therapies, Origin proposes dialing down tonic signaling with milder promoters. For AAV therapies, the focus shifts to tissue-specific sequences that reduce off-target expression and the immunogenicity that often follows. Origin claims to be assembling "the largest proprietary dataset of synthetic regulatory sequences across diverse cell-states," though the company has been circumspect about the dataset's actual scope or composition.
To validate its AI-generated designs, Origin employs Malinois, a convolutional neural network published in Nature in 2024 by researchers at the Broad Institute. Malinois predicts cell-type-specific regulatory activity from massively parallel reporter assay data, establishing a two-stage validation cycle: in-silico design followed by computational prediction via an independent model. Whether that computational validation correlates with wet-lab results, and whether wet-lab results predict clinical outcomes, are questions the company will need to answer in the years ahead.
The Landscape Is More Crowded Than It Appears
Origin isn't operating in a vacuum, though the specific niche of AI-designed promoters and enhancers for therapeutic applications remains relatively unpopulated. AskBio's Synpromics unit has been developing data-driven synthetic promoters through its PromPT bioinformatics engine for some time now, producing tissue-specific and tunable regulatory elements for AAV gene therapy. The approach predates the current AI frenzy but demonstrates that demand for engineered expression control existed well before transformers and foundation models became buzzwords.
Senti Biosciences has pursued promoter discovery alongside what it calls "smart sensor" gene circuits—systems that enable state-specific control in cell therapies. It's a different angle on the same fundamental challenge: making therapeutic genes responsive to cellular context.
Then there's Tune Therapeutics, which announced a $175 million Series B on January 12, 2025, and takes an entirely different approach. Rather than redesigning DNA sequences, Tune's TEMPO platform uses catalytically dead CRISPR machinery to modulate gene expression without cutting DNA—epigenome editing instead of sequence editing. The company's hepatitis B program has entered clinical trials, offering a real-world test of whether modulating expression via epigenetic marks can match or exceed the results from traditional gene therapy constructs.
Epic Bio, previously known as Epicrispr Biotechnologies, similarly employs compact dCas-based modulators for epigenetic modulation. The company has muscle disease programs in its pipeline and presented at the 2025 ASGCT annual meeting.
These approaches sit adjacent to the delivery optimization layer, where companies like Dyno Therapeutics, Voyager Therapeutics, and Capsida are engineering AAV capsids using AI and directed evolution. Dyno launched three breakthrough capsid vectors—targeting eye, muscle, and CNS tissues—at the May 14, 2025 ASGCT meeting. Capsida secured an AbbVie opt-in for its first genetic medicine program on January 7, 2025, adding to existing partnerships with Eli Lilly and CRISPR Therapeutics.
The distinction matters more than it might initially appear. Capsid engineering determines which cells a therapy reaches. Regulatory DNA engineering determines what happens after arrival. Both are necessary; neither alone is sufficient.
Regulatory Tailwinds and Market Urgency
The urgency animating this innovation stems from converging market opportunity and regulatory pressure. Four gene therapies received FDA approval in 2025 alone: Encelto for macular telangiectasia type 2 in March, prademagene zamikeracel for recessive dystrophic epidermolysis bullosa in April, and Waskyra for Wiskott-Aldrich syndrome in December, alongside obecabtagene autoleucel's late-2024 approval for relapsed/refractory B-cell acute lymphoblastic leukemia.
Yet safety concerns continue to cast long shadows. The Elevidys hepatotoxicity cases prompted not just the boxed warning but shipment holds and resumptions throughout mid-2025, creating uncertainty for patients and clinicians alike. Regulators are attempting to thread the needle between maintaining oversight and not strangling innovation. On January 11, 2026, the FDA announced a more flexible approach to chemistry, manufacturing, and controls oversight for cell and gene therapies, emphasizing risk-based flexibilities while maintaining quality standards. A STAT analysis later that month noted the potential to accelerate approvals while warning of systemic risks requiring careful monitoring.
The Alliance for Regenerative Medicine cited the FDA's CMC flexibility in a February 2026 brief as grounds for optimism about pipeline progress. An FDA-ASGCT workshop on pediatric cell and gene therapy clinical trials, scheduled for April 9, 2026 and posted in early March, signals continued regulatory attention to trial design and ethics questions that are particularly fraught when patients are children with life-limiting conditions.
The underlying science, meanwhile, continues to advance. DeepMind's AlphaGenome, announced in August 2025, provided what the company described as a unified long-context DNA model for noncoding regulatory interpretation. Illumina unveiled PromoterAI in June 2025, an algorithm designed to decipher pathogenic promoter variants that may explain up to 6 percent of rare disease genetic causes—a non-trivial fraction when you consider how many rare disease patients lack definitive molecular diagnoses.
Academic preprints from 2025 through early 2026 have detailed methods like Ctrl-DNA for cell-type-specific regulatory sequence design with off-target constraints, and the Central Dogma Transformer II, which links DNA and RNA through cross-attention mechanisms. Whether any of these academic tools transition to commercial applications remains uncertain, though the intellectual ferment is undeniable.
A Deloitte survey in January 2026 found that 62 percent of life sciences executives expect cell, gene, and RNA-based therapies to drive near-term revenue growth, with increased AI adoption anticipated through the year. It's the kind of finding that could reflect genuine insight or simply executives following the prevailing winds. Probably some of both.
The Manufacturing Inflection
The convergence of AI-designed regulatory DNA with manufacturing scale-up points toward a potential inflection, though the timing remains uncertain. Bristol Myers Squibb's $380 million capacity reservation with Cellares in August 2024 for automated CAR-T manufacturing illustrates the industrialization already underway in this space. As gene therapy production becomes more standardized—more factory, less artisanal craft—the pressure to optimize not just manufacturing processes but the therapeutic constructs themselves intensifies.
After all, there's limited value in perfecting the manufacturing process for a construct that triggers liver toxicity.
Origin and its peers face a validation gauntlet that has humbled more than a few computational biology ventures. Predictions must survive in vitro testing. In vitro results must translate to in vivo models. Animal data must predict clinical outcomes. The field has seen promising computational approaches collapse at each of those transitions, often spectacularly.
Yet the alternative—continuing to rely on a handful of legacy promoters optimized by evolution for purposes entirely unrelated to human therapeutics—looks increasingly untenable as safety signals accumulate and regulatory expectations evolve. The Chan Zuckerberg Initiative's December 2025 announcement of a major AI-biology initiative centered on building a "model of the cell" reflects broader institutional bets on multi-modal biological foundation models. The regulatory DNA layer sits squarely within that ambition, representing the code that translates genomic information into cellular behavior.
For gene therapy developers, the question has shifted from whether to engineer expression control to how. The answer will likely arrive from multiple directions simultaneously: synthetic promoters, epigenome editors, AI-designed regulatory elements each finding their niche depending on therapeutic context, target tissue, and manufacturing constraints.
Origin's ultimate contribution will be measured not by benchmark improvements over AlphaGenome—those are table stakes—but by whether its sequences survive the journey from computational design to patient benefit. In a field where expression control can mean the difference between cure and catastrophe, where the gap between breakthrough and breakdown can hinge on how strongly a gene fires in liver versus muscle tissue, that journey is worth watching closely.
The volume knobs matter more than we thought.
