The death came in March 2025, buried in a routine safety disclosure. A patient receiving Sarepta's Elevidys—a gene therapy for Duchenne muscular dystrophy delivered via adeno-associated virus—succumbed to acute liver failure. Same year, the FDA issued boxed warnings across every approved CAR-T therapy for secondary T-cell malignancies. Industry veterans saw the pattern immediately.
Gene therapy has a control problem.
Once you inject a therapeutic gene into a patient, you can't fully dictate where it activates, how strongly it expresses, or when—if ever—it shuts down. The consequences cascade: liver toxicity from unchecked expression, T-cell exhaustion from overstimulated CAR constructs, viral vectors transfecting entirely wrong tissues. Market watchers project the cell and gene therapy sector will balloon from $25 billion today to $118.6 billion by 2034. That nearly fivefold leap assumes one big thing: developers figure out how to control what happens after the vector leaves the syringe.
What if the answer isn't engineering better viral shells or more selective cell modifications? What if it's rewriting the instruction manual itself—the regulatory DNA that tells genes when to switch on?
The Momentum Trap
As of Q3 2025, more than 3,200 cell, gene, and RNA therapy trials are underway globally, according to Citeline and the American Society of Gene & Cell Therapy. That's 125 new studies launched in a single quarter. Since 2004, 76 cell and gene therapies have reached patients worldwide; 106 products now carry regulatory approval somewhere, and 91 more sit under review, per the International Society for Cell & Gene Therapy's latest pipeline report.
Impressive numbers. Misleading ones, too.
In January and April 2024, the FDA mandated classwide boxed warnings across all CD19- and BCMA-directed CAR-T therapies after reports of secondary T-cell cancers emerged. By October 2025, another warning landed—this time for immune effector cell-associated enterocolitis affecting Carvykti. A 2025 study in Molecular Therapy documented dorsal root ganglion toxicity in nonhuman primates and mice following AAV delivery into the central nervous system. Biomarkers like phosphorylated neurofilament heavy chain, the researchers noted, correlated with neural injury in ways that earlier safety models had missed.
The Sarepta fatality underscored what preclinical data had whispered for years: AAV vectors paired with ubiquitous promoters—regulatory sequences that activate genes everywhere—can push transgene expression past the liver's breaking point. Promoters like CMV and CAG, borrowed decades ago from viral genomes because they worked robustly in lab dishes, offer raw power but zero tissue selectivity. They're the equivalent of wiring a house with one light switch that controls every room at full brightness. Gene therapy needs dimmer switches with room selectors.
CMS attempted to ease access bottlenecks last July by launching its Cell & Gene Therapy Access Model, enrolling 33 states, Washington D.C., and Puerto Rico in outcomes-based reimbursement agreements for sickle cell gene therapies. But payment innovation only stretches so far when the underlying products carry safety flags that narrow dosing windows and shrink eligible patient pools.
Where Capsids Can't Go
Gene therapies typically bundle three components: the therapeutic transgene (the gene doing the work), the delivery vehicle (an AAV capsid or lentiviral vector), and the regulatory DNA governing when and where that gene fires up. The field has poured billions into capsid optimization. Companies like Dyno Therapeutics and Capsida Biotherapeutics have built entire platforms around AI-designed AAV variants aimed at improving tissue tropism—the tendency of a vector to home in on specific organs or cell types.
Here's the thing: capsid engineering only narrows biodistribution. It doesn't stop the transgene from expressing in every single cell the vector manages to reach.
That's the promoter's job. These stretches of noncoding DNA—sometimes just a few hundred base pairs—act as molecular switches, recruiting transcription factors that either activate or silence gene expression. A muscle-specific promoter like MHCK7, used in Sarepta's Duchenne programs, drives strong expression in cardiac and skeletal muscle while staying mostly quiet in liver and kidney. Retinal promoters light up in photoreceptors or retinal ganglion cells but not in surrounding support cells.
The catalog of naturally occurring tissue-specific promoters is thin, though, and most haven't been tuned for therapeutic use. Clinical and preclinical data reveal what happens when promoter choice goes wrong. A 2021 review in Cell Reports Medicine found that strong constitutive promoters like EF1α in CAR-T constructs can spike CAR density on T-cell surfaces, triggering tonic signaling, cytokine storms, and premature immune cell burnout. Switching to milder promoters—MND or EFS—reduced cytokine release and improved CAR-T persistence in several studies.
In AAV gene therapy, ex vivo work on human brain tissue showed that common AAV2 and AAV9 vectors exhibit broad, indiscriminate transduction across cell types. Meaning: if you want precision, the capsid alone won't deliver it. The promoter has to shoulder that burden.
Then there's the longevity problem. CMV and CAG promoters, stalwarts of early gene therapy, are prone to methylation-driven silencing and immune-mediated shutdown in living organisms. A 2013 study in Human Gene Therapy documented rapid decline in liver transgene expression within weeks when driven by CMV, while endogenous liver promoters maintained stable output. If your therapeutic gene's promoter powers down six months after dosing, you haven't cured anything. You've installed a ticking clock.
MicroRNA-based "detargeting" offers one workaround. Inserting miR-142-3p target sites into a transgene cassette lets developers suppress expression in antigen-presenting cells, dialing down immunogenicity. Liver expression can be muted via miR-122 target sites. Clever, sure. But these are patches. The real issue? Most therapeutic constructs still lean on a handful of natural promoters never designed for precision medicine.
When Prediction Met Design
DeepMind's AlphaGenome, unveiled in June 2025 and published in Nature this past January, marked a turning point. The long-context, multitask AI model can predict gene regulation and variant effects across up to one megabase of DNA sequence—roughly the length of a small bacterial genome. By early 2026, DeepMind reported about 3,000 researchers using it, with API calls topping one million daily.
AlphaGenome wasn't purpose-built for therapy design. It was built to decode how regulatory DNA works. But prediction is step one. Engineering comes next.
Origin.bio took that next step. The Y Combinator Winter 2026 startup, based in San Francisco, announced Axis on October 8, 2025—a multifunctional Transformer model that both generates regulatory DNA and predicts its function. According to Origin's internal benchmarks, Axis delivers a 6.7 percent average improvement over AlphaGenome on binding-activity prediction tasks. More critically, it can generate promoters and enhancers guided by transcription factor motifs and cell-state prompts, achieving up to ninefold enrichment of desired TF binding sites.
CEO Yash Rathod and CTO Malhar Bhide describe their mission in disarmingly simple terms: building "DNA switches and dials" to control therapeutic gene expression. They're targeting oncology and central nervous system indications where safety hinges on avoiding off-target expression in healthy tissue. The company's advisory roster includes Manolis Kellis of MIT and the Broad Institute—whose work on regulatory genomics and insulator elements carries weight—and Nicole Paulk, an AAV gene therapy specialist with a track record of publications on vector delivery and targeting efficiency.
Origin claims it's assembling the largest proprietary dataset of experimentally validated regulatory sequences across diverse cell states. That's not just marketing fluff. Public datasets like ENCODE offer millions of regulatory annotations, but they skew heavily toward immortalized cell lines and bulk tissue samples. Therapeutically relevant data—primary human neurons in diseased states, or tumor-infiltrating T cells—remain sparse. High-quality, disease-context-specific training data could become the competitive moat that matters.
Origin isn't the only player. Asimov unveiled its AI-powered AAV Edge platform in 2024, integrating machine-designed tissue-specific promoters with claims of over 200-fold on/off dynamic range in preclinical work. AskBio, via its 2019 acquisition of Synpromics, operates the PromPT machine learning engine to design customizable, tissue-selective, even inducible promoters for AAV payloads. Chromatin Bioscience's chromatinLENS platform focuses on small, cell-selective synthetic promoters that can also be tuned to remain silent in manufacturing producer cells—boosting vector yields by preventing payload toxicity during production. Annogen's SuRE (Survey of Regulatory Elements) platform screens millions of genomic fragments to fish out promoters with desired specificity and recently partnered with VectorY to develop CNS-specific regulatory elements.
Even diagnostics companies are edging into the space. Illumina launched PromoterAI in January 2025, a deep learning algorithm for interpreting promoter variants in rare disease diagnostics. While not a design tool per se, its arrival signals how AI models trained on regulatory sequences are proliferating across adjacent markets.
The Leap From Prediction to Synthesis
The jump from forecasting promoter activity to generating novel regulatory DNA involves generative modeling techniques borrowed from text and image synthesis. Ctrl-DNA, a reinforcement learning framework described in recent arXiv preprints, enables controllable, cell-type-specific regulatory sequence generation. GPro, an open-source toolkit, provides end-to-end pipelines for promoter design. These methods cycle through rounds of in silico generation, followed by experimental validation using massively parallel reporter assays (MPRAs) or in vivo barcoded promoter libraries.
MPRAs—which test thousands of synthetic promoters simultaneously by linking each sequence to a unique barcode—have become the proving ground for AI-designed regulatory DNA. A 2025 study in Genome Biology that harmonized six MPRA and STARR-seq datasets revealed something sobering, though: assay and lab effects introduce significant variability. Cross-assay reproducibility is low. A promoter that shines in one MPRA might fail in another, let alone in an animal model. That's why Origin and competitors emphasize proprietary, experimentally validated datasets. Public benchmarks alone won't cut it for therapeutic applications.
A 2026 arXiv preprint on robustness gaps in regulatory sequence machine learning models flagged another risk: these models often stumble under biological and technical distribution shifts. Calibration breaks down when moving from one cell type to another or from one experimental platform to the next. The authors argued for incorporating structural priors and uncertainty-aware predictions—essentially, building models that know when they don't know. Humility in silico, if you will.
Still, early results hint at real promise. Asimov's 200-fold dynamic range claim, if it holds up in pivotal studies, would represent a stepwise leap over natural promoters. Tissue-specific promoters used in current AAV therapies—MHCK7 in Duchenne programs, hSyn in neurological applications—typically achieve ten- to fiftyfold selectivity. Orders-of-magnitude improvements could enable lower vector doses, trimming both cost and toxicity.
Manufacturing presents another frontier. Chromatin Bioscience's approach of designing promoters that stay dormant in HEK293 or other producer cells tackles a longstanding bottleneck: if the therapeutic transgene is toxic to the cells making the vector, yields crater. Artificial microRNA strategies during AAV production have shown improved titers by silencing payload expression during manufacturing, then flipping it on post-transduction in the patient. Pairing AI-designed promoters with microRNA-based control layers could streamline both upstream production and downstream safety.
The Battleground Ahead
As capsid engineering matures and delivery becomes more predictable, the expression control layer is poised to become the next competitive front. Dyno Therapeutics' recent option exercise by Roche and Capsida's ongoing collaboration with AbbVie demonstrate Big Pharma's appetite for next-generation AAV vectors. But a perfectly targeted capsid paired with a promiscuous promoter still delivers the therapeutic gene to off-target cells within the intended tissue. Precision demands both components working in lockstep.
Regulatory agencies are adjusting in real time. The FDA eliminated Risk Evaluation and Mitigation Strategies (REMS) for CAR-T therapies in June 2025, citing increased provider experience, but kept long-term safety monitoring requirements intact—typically fifteen years for gene and cell therapies. Boxed warnings now blanket entire CAR-T classes. If synthetic promoters demonstrably cut off-target expression and associated toxicities, they could shift from competitive edge to regulatory expectation. The agency may soon ask: why wouldn't you use a tissue-specific promoter when one's available?
Reimbursement dynamics will amplify this shift. The CMS Cell & Gene Therapy Access Model ties Medicaid payments for sickle cell gene therapies to patient outcomes, with 33 states and Puerto Rico participating as of July 2025. Payers increasingly favor constructs that achieve efficacy at lower vector or cell doses—lower doses translate to lower manufacturing costs and reduced toxicity risk. A gene therapy with a tightly controlled, tissue-specific promoter that works at one-tenth the dose of a comparator will have an easier path through payer negotiations. Perhaps much easier.
Europe presents different headwinds. New EU joint clinical assessment rules, effective January 2025, emphasize randomized controlled trial data and could disadvantage rare disease cell and gene therapies that rely on single-arm studies. Advocacy groups have warned that the shift de-emphasizes real-world evidence precisely when gene therapies for ultra-rare conditions are entering the market. If a synthetic promoter improves safety enough to expand eligible patient populations, it might ease the trial design burden by making larger studies feasible.
The data moat question looms. AlphaGenome is open source; DeepMind released API and model code in January 2026. That democratizes access to state-of-the-art prediction tools but doesn't provide the experimental validation layer. Origin's claim of the "largest proprietary dataset" of synthetic regulatory elements matters if that dataset includes primary human cells in disease-relevant states, in vivo validation, and manufacturing compatibility data. Asimov, AskBio, Chromatin Bioscience, and Annogen are all building similar proprietary libraries. The winners in this space will be those who close the loop fastest between AI design, wet-lab validation, and clinical translation.
Watch for convergence. Just as diffusion models are now being applied to AAV capsid generation (e.g., AAVDiff), generative AI techniques will increasingly span both delivery and payload control. A platform that designs matched capsid-promoter pairs optimized for a specific indication and cell type could command significant value. Tune Therapeutics' $175 million raise in 2025 for epigenetic gene regulation—its TEMPO platform represses PCSK9 in nonhuman primates without editing DNA—hints at adjacent modalities that could complement or compete with synthetic promoters.
What remains to be seen is whether AI-designed regulatory DNA can overcome the reproducibility and generalization hurdles that tripped up earlier waves of computational biology. Company benchmarks need independent replication. Models trained on immortalized cell lines need to perform in primary human tissue. Promoters that work in mouse models need to translate to primates and humans—a notoriously treacherous leap in gene therapy, one that's humbled more than a few well-funded programs.
But if the technology delivers? It won't just patch a safety problem. It will redefine what's possible in genetic medicine.
The difference between a therapy that cures and one that harms often comes down to a few hundred base pairs of regulatory DNA. Getting those base pairs right might finally let the $25 billion cell and gene therapy market become the $100 billion industry it's racing toward. Whether AI-designed switches and dials can carry that weight is the question keeping a growing number of gene therapy executives awake at night—and the one investors will be betting on for the next decade.
