Two teenagers, both confined to wheelchairs by Duchenne muscular dystrophy, both already losing ground to the disease. And both dead from acute liver failure after receiving Elevidys, Sarepta's gene therapy that was supposed to help them.
The FDA's response, when it came on November 14, 2024, was blunt: a Boxed Warning for acute serious liver injury, the agency's starkest advisory. The therapy would now be restricted to ambulatory patients four and older—essentially, the kids who could still walk. The ones who needed it most could no longer get it.
That regulatory action wasn't just about one drug. It crystallized something the gene therapy industry has been reluctant to say out loud: the safety problems aren't anomalies anymore. They're structural. And as these treatments move beyond ultra-rare diseases into broader patient populations, the question is whether the field can engineer its way out of trouble before regulators do it for them.
The core issue is almost embarrassingly simple. Most gene therapies work too well in the wrong places.
Momentum, Then Friction
On paper, 2025 looked like gene therapy's year. Four new approvals globally in the third quarter alone, spanning gene, cell, and RNA modalities, per the Citeline quarterly landscape report. Real progress, the kind you can point to at investor meetings.
Look closer, though. Trial initiations that same quarter slowed to 125. Deal activity in Q1 dropped 20% from the prior quarter to just 90 transactions. The sector is growing, yes, but it's also bumping up against limits.
Elevidys wasn't an isolated incident, either. Sarepta temporarily paused U.S. shipments in July 2025 while investigating safety signals. Academic groups flagged dorsal root ganglion toxicity when AAV vectors were delivered into cerebrospinal fluid—in animal models, sure, but concerning enough that papers started circulating. A systematic review published in March cataloged safety trends across ocular AAV programs, detailing which immunosuppression protocols worked and which routes of delivery carried the highest risks.
The FDA, for its part, has been staffing up. The agency reorganized its Office of Therapeutic Products back in 2023 to handle the surge in cell and gene therapy applications, and under PDUFA VII it's been hiring aggressively. That's not a sign of enthusiasm. That's a sign regulators know what's coming.
The therapies themselves work, often spectacularly. That's not the problem. The problem is where they work. High-dose AAV vectors carrying therapeutic genes don't just home to the liver, heart, or nervous system where they're needed. They also express in tissues where they're decidedly not needed—triggering immune responses, organ toxicity, or both. Dose escalation studies hit ceilings. Patients who might benefit get excluded from trials because their liver enzymes are already elevated, or because they have pre-existing antibodies to the viral vector.
It's a design flaw baked into the therapies themselves.
The Expression Problem
Most AAV gene therapies use what's called constitutive expression—promoters like CMV or CAG that tell cells to produce the therapeutic protein constantly, everywhere the vector ends up. If the vector biodistributes to the liver, the liver starts churning out protein. Heart? Same thing. Brain? Sure, why not.
That indiscriminate expression creates what researchers politely call "antigen burden." Less politely: you're asking the immune system to tolerate a foreign protein showing up in multiple organs at once, often at high levels. Sometimes the immune system says no. Sometimes organs say no. Sometimes both.
Tissue-specific promoters have existed for years—MHCK7 for muscle and cardiac tissue in Duchenne programs, for instance. But designing these elements has been painstaking work. You pick a known promoter from the literature, test it in cell lines, move to animals, iterate. Maybe it works. Maybe it doesn't express strongly enough. Maybe it's too large to fit in the AAV capsid alongside your therapeutic gene. The process can stretch months or years for a single indication.
Meanwhile, the field's ambitions are expanding. Deloitte's October industry analysis noted that sector leaders were describing a shift from "surviving" to "thriving," with R&D pivoting toward broader indications beyond oncology and rare disease. Translation: more patients, more tissues, more complexity. To scale safely, the industry needs a faster, more reliable way to program where a therapy expresses, how much, and when.
Which brings us to AI. Or more precisely, to the sudden realization that machine learning might be able to do in weeks what took human scientists months.
Models Meet Molecules
DeepMind made waves in June 2025 with AlphaGenome, demonstrating that long-context language models could predict multi-omic regulatory readouts and potentially guide the design of synthetic regulatory DNA for cell-type-specific expression. A November 2024 Nature paper on "Malinois," a machine-guided system for designing cis-regulatory elements, showed that ML predictions outperformed classical selection methods in high-throughput tests across multiple cell lines. Academic groups started dropping arXiv preprints on reinforcement learning models for generating cell-type-specific promoters and enhancers.
The technical momentum was clear. The commercial rationale, maybe clearer. Regulators wanted safer constructs after events like the Elevidys warning. Payers wanted better outcomes—CMS had launched its Cell and Gene Therapy Access Model in early 2025, initially for sickle cell disease but signaling a broader shift toward tying reimbursement to measurable safety and durability. Developers, boxed in by dose-limiting toxicities, needed constructs that concentrated therapeutic effect where it mattered and went quiet everywhere else.
Not exactly a small ask. But suddenly, computationally feasible.
Four People, One Model
Origin.bio is a four-person startup in Y Combinator's Winter 2026 batch. Their pitch: AI can make the regulatory DNA that controls gene expression as programmable as code. Design enhancers and promoters that act as "switches and dials," tuning therapeutic gene output precisely in target cell states while staying silent in off-target tissues.
Co-founders Yash Rathod (CEO, with a computer science and reinforcement learning background from the University of Illinois Urbana-Champaign) and Malhar Bhide (CTO, ML research, previously at YC alum Automorphic) launched their first model, Axis, on October 8, 2025. They claim it's the first model that both generates regulatory elements and predicts their function—a unified design-and-validation loop, rather than separate tools for each step.
The numbers Origin reports: a 6.7% average gain over AlphaGenome on regulatory binding prediction benchmarks, and up to 9× motif enrichment when using high-affinity prompts. The model trains on ENCODE v4 cis-regulatory elements and is building what the company describes as a proprietary wet-lab dataset of "millions" of experimentally validated regulatory DNA sequences across cell and tissue types.
The pitch, distilled: tailor expression to disease states in cancer and CNS disorders, where off-tissue expression can mean neurotoxicity, immune flares, or simply wasted therapeutic effect. Origin's advisory board includes Manolis Kellis from MIT and the Broad Institute (regulatory genomics), Nicole Paulk from UCSF (AAV gene therapy and founder of Siren Bio), and Rashid Bashir (Dean of Engineering at UIUC). In their October blog post introducing Axis, Origin argued that unified DNA models offer a scalable path to optimizing both safety and efficacy—citing examples like tuning promoter strength in CAR-T therapies to reduce tonic signaling, and using tissue-specific promoters in AAV to limit off-tissue exposure.
It's an ambitious claim for a four-person company. Then again, the field is full of ambitious claims. The question is always execution.
Others Placing Bets
Origin isn't alone in rethinking how gene therapies express their payloads. AskBio, which acquired synthetic promoter platform Synpromics back in 2019, continues to engineer cell- and tissue-selective promoters for AAV programs. Chromatin Bioscience announced a collaboration with Johnson & Johnson in January 2025 around synthetic promoter design. Pacira Biosciences presented data at the ASGCT 2025 meeting on PCRX-201, an AAV program using an inflammation-responsive promoter to conditionally express IL-1Ra in osteoarthritis joints—gene expression that activates only when inflammation signals appear.
Adjacent approaches tackle the same core problem from different angles. Senti Bio's off-the-shelf CAR-NK candidate SENTI-202 uses logic gates—CD33 OR FLT3, NOT EMCN—to selectively target AML cells, with clinical activity updates presented at the American Society of Hematology meeting in late 2025. A2 Bio's A2B395, a logic-gated CAR-T targeting EGFR-positive tumors with HLA loss, entered trials with data at ASCO.
Strand Therapeutics raised $153 million in August 2025 for programmable mRNA circuits that express IL-12 only in tumor microenvironments and use circRNA circuits to de-target liver expression—essentially, RNA-level control of where proteins get made. Dyno Therapeutics continues AI-engineered AAV capsid design, trying to improve tropism and reduce immune risk at the vector level rather than the payload level.
The common thread running through all of this? Computational design is moving from capsids into payloads and regulatory cassettes. Reviews published over the past year emphasize promoter choice and "detargeting" strategies—like inserting microRNA target sites (miR-122 for liver, for example) to suppress expression in off-target tissues. A methods report demonstrated miR-122 target-site detargeting to enhance cardiac specificity in an AAV-CRISPR editing vector. Tenaya Therapeutics presented compact chimeric cardiac promoters at ASGCT 2024, designed to fit within AAV payload constraints while maintaining strong, heart-specific expression.
These aren't academic exercises. They're responses to regulatory pressure that's only getting sharper.
What Regulators Want
The FDA's February 2026 Grand Rounds session on "AAV-Mediated Gene Therapy: Advances, Immune Challenges, and Future Directions" highlighted ongoing research into computational capsid engineering, sex-based immunology differences, and product characterization for safety. All areas where expression-level control matters.
The Elevidys Boxed Warning likely won't be the last regulatory action tightening gene therapy's risk-benefit calculus. As AAV programs advance into larger, more diverse patient populations—neuromuscular, cardiac, CNS indications—dose-limiting toxicities will surface more frequently. Developers will face a narrowing choice: accept restricted labels and smaller addressable markets, or redesign constructs to reduce off-tissue expression and antigen load.
Regulatory science is trying to keep up. The FDA's Office of Therapeutic Products, still ramping under PDUFA VII, will likely issue more guidance on expression cassette design and nonclinical characterization as construct complexity increases. ICH S12, the 2023 guideline on nonclinical biodistribution for gene therapy products, sets a baseline but doesn't address adaptive or inducible systems in detail yet. EMA reflection papers on product design modifications flag promoter and enhancer changes as comparability considerations—regulators are paying attention, even if formal frameworks lag behind the innovation.
Payer dynamics may accelerate adoption of safer designs faster than regulations do. CMS's Gene Therapy Access Model, which began rolling out in January 2025 and expanded to 33 states plus DC and Puerto Rico, ties reimbursement to outcomes. If a therapy causes liver toxicity or fails to show durability, payers won't pay—or won't pay as much. Constructs that demonstrate predictable, cell-state-specific expression and reduced off-target effects will have an edge in outcomes-based contracts. Deloitte's analysis of CGT financing models noted the sector's shift toward value-based arrangements, creating further incentive to design for safety and efficacy simultaneously rather than sequentially.
The Market, and The Gap
For founders, the opportunity is clear, if daunting. Gene therapy market projections remain sizable—Custom Market Insights estimated $25.2 billion in 2025, with growth projected to $118.6 billion by 2034—but the winners will be those who solve the expression problem, not just the delivery problem. Investors are already getting more selective. That 20% drop in deal volume in Q1 2025 suggests capital is sorting. Programs that can demonstrate AI-driven, validated expression control—backed by proprietary datasets and advisor networks spanning genomics, AAV biology, and regulatory affairs—will stand out.
The technical pieces are converging faster than anyone expected a few years ago. Machine learning models can now generate and evaluate regulatory DNA sequences. Wet-lab platforms can validate thousands of sequences in parallel using massively parallel reporter assays. Regulatory agencies are staffed and attentive, perhaps more than developers would like. Payers are demanding proof, not promises.
What's missing is execution. Companies that can close the loop from model output to clinical construct to patient benefit. Origin.bio is one early bet on that execution. Others will follow—likely, are already forming in stealth mode or pivoting existing platforms.
The question isn't whether AI will reshape how gene therapies are designed at the regulatory element level. It's whether that reshaping happens fast enough to outpace the accumulating safety signals and regulatory restrictions. The Elevidys warning was a wake-up call. The field's response will determine whether gene therapy scales into a broad platform or remains constrained to niches where risk-benefit calculations still pencil out.
Two teenagers died. The FDA moved. Now the question is whether the industry can move faster—not just with better vectors or higher doses, but with smarter expression control that keeps the therapy where it belongs and silent where it doesn't. Because the alternative is narrower labels, smaller markets, and more Boxed Warnings.
And perhaps, eventually, a realization that the problem was never whether we could deliver genes. It was whether we could control them once they arrived.
