Deaths were reported in 2025 related to ELEVIDYS—a promising adeno-associated virus gene therapy for Duchenne muscular dystrophy—with subsequent FDA actions following reports of fatal outcomes. By November, the FDA had added a boxed warning to the drug and revised its indication to limit who could receive it. For a field still haunted by the death of Jesse Gelsinger in 1999, it was a grim reminder: getting therapeutic genes into the body is the easy part. Controlling what happens next is where things get dangerous.
The problem, researchers will tell you, isn't the delivery vehicle. AAV vectors are remarkably good at their job—sometimes too good. The problem is the genetic payload itself, specifically the regulatory sequences that determine when, where, and how vigorously a therapeutic gene switches on. Borrow a promoter from a virus or scavenge one from the human genome, and you're flying blind. Expression might surge in the liver when you need it in muscle. Or it might barely register at all.
Now a small group of startups and academic labs think they have a better way: let machine learning write the instructions.
Origin Bio, a four-person team that emerged from Y Combinator's Winter 2026 cohort, is perhaps the most audacious bet in this space. In October 2025, they launched Axis, a model designed to both generate new regulatory DNA sequences and predict how they'll behave inside cells. Five months later, they released Switch—an open repository of 10,000 AI-designed regulatory elements, each tagged with predicted activity levels across neuroblastoma, liver, and blood cell lines. The pitch? Stop guessing. Start designing.
Whether it works remains an open question. But the timing is hard to ignore.
A Field Growing Fast, Struggling to Stay Safe
Gene therapy, for all its stumbles, isn't short on momentum. The Alliance for Regenerative Medicine counted five to six U.S. approvals expected in 2025 alone. Market forecasts—caveat emptor on those—routinely project double-digit growth through the 2030s. Astute Analytica, in a March 2026 report, projected the cell and gene therapy market would balloon from $36.5 billion in 2025 to $183.1 billion by 2035, a compound annual growth rate just north of 17%, though such figures are subject to varying methodologies.
But scratch beneath the optimism and you find unease. An August 2025 review in the Orphanet Journal of Rare Diseases catalogued a litany of AAV trial complications: elevated liver enzymes, neurotoxicity signals, reported deaths. The immune response to the viral shell gets most of the attention, but that's only part of it. The real danger, increasingly, appears to be what happens after the vector delivers its cargo. If the therapeutic gene expresses too strongly, or in the wrong tissues entirely, patients pay the price.
The standard workaround has been tissue-specific promoters. MHCK7, for instance, tilts expression toward skeletal and cardiac muscle—useful in Duchenne trials using micro-dystrophin constructs. But these sequences weren't built for therapy. They evolved for other purposes. Some are too bulky to fit inside AAV's notoriously cramped packaging limit. Others leak into off-target tissues. Still others drive expression at levels that fall short of therapeutic benefit, or overshoot into toxicity.
Which brings us back to machine learning.
Three Convergences
Three things have aligned to make this moment possible—things that have advanced significantly in recent years.
First, the data deluge. High-throughput assays like massively parallel reporter assays (MPRA) and STARR-seq can now test tens of thousands of DNA sequences simultaneously, measuring how strongly each one drives gene expression across different cell types. Origin Bio's Switch repository exemplifies the shift: 10,000 AI-generated sequences, each annotated with transcription factor binding sites, quality control metrics, even predicted 3D structures. The dataset alone would have been unthinkable a decade back.
Second, the models themselves. Foundation models trained on DNA sequences—DeepMind's AlphaGenome, unveiled in June 2025, is probably the best known—can now predict regulatory activity from sequence alone with something approaching reliability. Origin Bio claims Axis outperforms AlphaGenome on certain benchmarks, though they're coy about which ones. Smaller players have been at this longer than you might think. AskBio acquired Synpromics back in 2019 and has been quietly assembling synthetic promoter libraries. MeiraGTx, in a 2023 poster series, claimed their synthetic neuronal promoters beat the widely used synapsin (hSyn) promoter on both strength and specificity.
Third, a regulatory tailwind. In January 2026, the FDA announced a "flexible approach" to chemistry, manufacturing, and controls for cell and gene therapies—bureaucratese for lowering barriers in early development. Six weeks later came a draft framework for individualized therapies targeting ultra-rare diseases, explicitly encouraging master protocols and genome-specific interventions. Translation: the agency wants innovation, provided you can prove safety through the right assays.
Whether "the right assays" will include AI-designed regulatory elements is still up for debate.
The Ecosystem Taking Shape
Origin Bio isn't working in a vacuum. A loose ecosystem has quietly assembled itself.
Dyno Therapeutics, for one, is tackling the flip side of the problem—engineering better AAV capsids using machine learning to improve tissue targeting and immune evasion. In January 2025, Roche exercised an option on one of Dyno's AI-designed capsids for neurological gene therapy. By May, Dyno had rolled out three new capsid variants at the American Society of Gene & Cell Therapy annual meeting, each tailored for eye, muscle, or central nervous system applications. Recent preprints from early 2026 on reinforcement learning-guided capsid design suggest the pace is quickening.
Avista Therapeutics expanded its ARTEMIS platform in February 2026, weaving machine learning together with structural dynamics to predict how vectors will behave once inside the body. Academic labs have contributed foundational pieces: the CODA system, described in Nature in October 2024 by groups at the Broad Institute, JAX, and Yale, uses machine learning to guide the design of cell-type-specific cis-regulatory elements with validated in vivo activity.
Practical applications are starting to surface, if you know where to look. A 2025 Nature Communications paper detailed high-throughput discovery of erythroid enhancers—compact, potent sequences for gene therapy vectors aimed at blood disorders. Gene Therapy published work on cardiac-specific promoter evaluation using high-throughput screening for AAV applications. In March 2026, Earli, a cancer diagnostics outfit, posted a job listing for scientists to design tumor-activated synthetic promoters using AI and machine learning. Niche, but telling.
Not everyone is focused solely on expression control. Senti Bio and Laverock Therapeutics are building logic-gated circuits—molecular programs that make therapeutic decisions based on what state a cell is in. Omega Therapeutics uses epigenetic control, modulating gene expression without inserting new DNA at all. But the common thread is the same: precision matters, and machine learning might be the only way to get there.
What Comes Next
The immediate payoff, if it materializes, is risk reduction. Gene therapies currently carry list prices in the millions—ELEVIDYS is priced at approximately $3.2 million as of Q1 2026 for a one-time infusion. Payers are skittish. The CMS Cell and Gene Therapy Access Model, which began rolling out to states in January 2025, ties reimbursement directly to outcomes. If AI-designed regulatory elements can cut off-target toxicity, therapies become safer. Trials become cheaper. The regulatory path clears.
The longer-term vision is more ambitious, perhaps implausibly so. The FDA's draft framework for individualized therapies, released in February 2026, cracks open the door to bespoke genetic interventions tailored not just to disease but to a patient's specific genotype and cellular landscape. Picture a gene therapy where the promoter is custom-designed for a patient's muscle cell transcription factor profile, driving expression only where needed and shutting off everywhere else.
That's the bet Origin Bio's founders appear to be making. Yash Rathod, who won the 2022 OpenCV AI Research Competition, and Malhar Bhide, a YC alum from the Summer 2023 batch, claim Axis can generate regulatory DNA "prompted" by transcription factor binding preferences—enabling cell-type-specific, dose-tunable expression.
Whether that holds up under clinical scrutiny is another matter entirely. Origin Bio has assembled an impressive roster of advisors: Manolis Kellis at MIT and the Broad Institute, Nicole Paulk at Siren Bio (formerly on the scientific advisory boards of Dyno, Astellas Gene Therapies, and Metagenomi), and Rashid Bashir, Dean of the Grainger College of Engineering at UIUC. But the company is still pre-clinical. Their internal benchmarking suggests efficiency gains—one March 2026 post claimed their MuonW optimizer reached target validation perplexity with 37% fewer floating-point operations than a standard optimizer—but the real test will be wet lab validation. And eventually, patients.
The Trust Problem
The field's biggest obstacle may not be technical. It's trust.
After the ELEVIDYS deaths, regulators and patients alike are wary. The ICH Cell and Gene Therapy Development Considerations Recommendation Paper, endorsed in November 2025, emphasizes potency assays, transduction efficiency, transgene expression characterization, and rigorous dose-response studies. AI-designed regulatory elements will need to clear those same bars, and do so transparently. No black boxes.
If they do? The payoff could be profound. Gene therapy has always been a field defined by its promise and shadowed by its failures. Machine learning won't eliminate risk—nothing will. But it offers something the field has lacked from the beginning: a systematic way to engineer safety in from the start, rather than retrofit it after the damage is done.
The deaths in 2025 were a tragedy. They were also, perhaps more than anyone wanted, a turning point. The question now is what gets built on the other side of it.
