The warning came on a Friday. November 14, 2025—late afternoon, when most biotech investors were already mentally checked out for the weekend. The FDA's boxed warning on Sarepta's ELEVIDYS landed with the kind of bureaucratic precision that can crater a stock before Monday's opening bell. Acute liver failure. Fatal cases. A gene therapy for Duchenne muscular dystrophy, suddenly restricted to a narrower slice of desperately ill children.
It wasn't the sector's first stumble. Four months prior, the FDA had opened an investigation into multiple deaths tied to Sarepta's AAV-based therapies. And if you rewind to April 2024, regulators had already slapped a class-wide boxed warning on CAR-T therapies after reports surfaced of secondary T-cell malignancies. A pattern, in other words.
Gene therapy's safety crisis isn't theoretical anymore. It's happening at industrial scale. Roughly 3,200 clinical trials are underway as of the third quarter of 2025. The market, valued at $11.4 billion this year, is projected to hit $58.9 billion by 2034. These aren't fringe experiments—they're becoming standard of care for diseases that once had no cure. Which means the sector can no longer treat toxicity as an unfortunate edge case, the price of doing business with cutting-edge science.
The therapies work, often brilliantly. That's not the question. The question is whether developers can control where, when, and how much a therapeutic gene gets expressed once it's inside the body. Get that calculus wrong—even slightly—and you flood the liver with a protein it doesn't need, trigger cascading immune responses, or worse.
The Control Problem (Which Shouldn't Still Be a Problem)
Here's the uncomfortable truth: the issue is surprisingly fundamental.
Most gene therapies use viral vectors, typically adeno-associated viruses, to smuggle a therapeutic gene into cells. Once inside, that gene needs a promoter—a stretch of DNA that essentially tells the cell's transcription machinery to flip the switch and start cranking out protein. But the promoters used today? Borrowed from viruses or adapted from natural human sequences, often with minimal tuning. They're blunt instruments dressed up as precision tools.
A strong promoter might drive high expression everywhere, risking toxicity in tissues you never intended to treat. A weak one might not produce enough therapeutic benefit to justify the million-dollar price tag. Tissue-specific promoters offer a middle path, but they leak. They activate in the wrong cell types. They fail to account for disease-state differences—the fact that a tumor cell's regulatory landscape looks nothing like its healthy counterpart.
Take Sarepta's ELEVIDYS. The therapy uses a synthetic promoter called MHCK7, engineered specifically to drive expression in skeletal and cardiac muscle. The design was meant to minimize off-target effects, to keep the therapy localized where the disease actually lives. Yet the hepatotoxicity seen in patients suggests systemic exposure issues—likely some mix of vector tropism, dose escalation, and immune response—that even careful promoter design couldn't solve.
Still. Better regulatory DNA could have helped. Researchers have shown that embedding microRNA target sites (like miR-122 for liver de-targeting) can suppress expression by orders of magnitude in unwanted tissues. The trouble is, designing these elements has been more art than science, more educated guesswork than engineering.
AI Enters the Design Stack
Origin, a San Francisco startup emerging from Y Combinator's Winter 2026 batch, is betting that machine learning can finally turn regulatory DNA design into a predictable engineering discipline.
Co-founders Yash Rathod and Malhar Bhide—who cut their teeth at NVIDIA, Berkeley, and UPenn—have built what they're calling Axis, "the first multifunctional DNA model capable of both design and prediction." The pitch sounds almost too clean: give the model a desired expression profile (strong in tumor cells, silent in healthy tissue; active in astrocytes, not neurons), and it generates synthetic promoters and enhancers tailored to those specifications.
Origin's claims are bold, perhaps more so than the founders fully appreciate. The company says Axis outperforms DeepMind's AlphaGenome, a long-context model released in mid-2025 that predicts regulatory effects from DNA sequences up to 1 megabase long. AlphaGenome set a new benchmark for non-coding variant interpretation and has been widely adopted in research labs via a free API—the kind of tool that becomes infrastructure overnight. If Origin's internal benchmarks hold—and it's worth emphasizing these are vendor claims, not yet peer-reviewed—it would suggest the startup has assembled proprietary training data that gives it a genuine edge.
The company says it's building "millions" of experimentally validated datapoints. Screening synthetic sequences across multiple cell lines and tissue contexts. That's the bet, really—not just a better algorithm, but a better feedback loop between prediction and reality.
Building the Dataset (Where the Real Work Happens)
That experimental backbone matters. Possibly more than the model architecture itself.
In 2024, a Nature Communications study screened 6,144 short synthetic promoters using massively parallel reporter assays—MPRAs, in the shorthand—demonstrating tunable dynamic ranges of 50- to 100-fold. Impressive, but still operating in a simplified system. Single-cell MPRAs, published in Nature Genetics the year before, pushed the technique further by enabling cell-type-specific readouts in mixed tissues. Suddenly you could test thousands of regulatory sequences in a single experiment and actually know which ones worked in, say, neurons versus astrocytes.
These tools generate the kind of high-resolution feedback loops that train better models. They also generate humility—turns out predicting gene expression is harder than anyone thought.
Origin isn't the first company to marry AI and experimental throughput in gene therapy. Dyno Therapeutics has used machine learning to design novel AAV capsids—the protein shells that determine which tissues a vector can actually infect—and inked deals with Roche and Novartis worth over $1 billion in potential milestone payments. Voyager Therapeutics licensed its TRACER capsids to Novartis in 2024 for CNS programs. Capsida Biotherapeutics has similar partnerships with AbbVie, Lilly, and CRISPR Therapeutics.
But those companies focus on delivery—getting the package to the right address. Origin is tackling the payload: the regulatory logic that governs gene expression once the vector reaches its target. It's the difference between building a better delivery truck and designing a smarter thermostat.
A Crowded but Specialized Field
Origin isn't working in a vacuum, though it sometimes sounds like it.
AskBio, now a Bayer subsidiary, acquired Synpromics back in 2019 to gain access to PromPT, a synthetic promoter platform built on bioinformatics and early machine learning. Academic tools like DeepSTARR (which designs enhancers) and Malinois/CODA (which optimizes cell-type-specific regulatory elements) have shown proof of concept in peer-reviewed studies. Enhanc3D Genomics, which raised a Series A in early 2025, uses 3D chromatin mapping to identify enhancer-promoter interactions. And DeepMind's Enformer, a transformer model published in 2021, laid much of the conceptual groundwork for predicting gene expression from sequence.
What sets Origin apart—if its thesis actually proves out—is the combination. Generative design, not just prediction. Disease-state conditioning: tumor versus normal, inflamed versus quiescent. And a feedback loop grounded in proprietary MPRA data that doesn't exist anywhere else, or at least not in this density.
The startup is targeting cancer and CNS disorders first, indications where precise expression control could reduce toxicity without sacrificing efficacy. Smart choices, therapeutically. Also the hardest problems in the field. Origin is assembling a scientific advisory board that includes Manolis Kellis from MIT's computational biology group, Nicole Paulk (gene therapy at UCSF), and Rashid Bashir (bioengineering at UIUC). That lineup signals ambitions beyond a single product—they're building a platform, or at least trying to.
The Shifting Landscape
The regulatory and commercial environment is changing in ways that favor platforms like Origin's, assuming they can execute.
The FDA's draft guidance on potency assurance, released in December 2023, emphasizes risk-based, multi-assay validation strategies—exactly the kind of rigor that AI-designed, MPRA-validated regulatory elements can provide. Documentation that shows you understand not just what a therapy does, but why. The CMS Cell and Gene Therapy Access Model, launched in January 2025, ties Medicaid reimbursement for sickle cell gene therapies to real-world outcomes. Translation: payers are getting smarter, and they'll potentially reward developers who can demonstrate safer, more predictable expression profiles.
At the same time, the Alliance for Regenerative Medicine notes that the sector is entering a phase of "disciplined, sustainable growth." Which is consultant-speak for: investors want translational rigor now, not just scientific novelty. The era of funding anything with "CRISPR" or "AAV" in the pitch deck is over. Maybe it needed to be.
Origin's bet is that as the gene therapy market scales—125 new trials started in Q3 2025 alone—the bottleneck won't be delivering a gene anymore. It'll be controlling what happens next. Whether Axis proves genuinely superior to AlphaGenome or other emergent models remains an open question. Independent benchmarking will tell the story, as it always does. But the underlying problem? Undeniable. And the FDA's warnings have made it impossible to ignore.
Gene therapy needs better switches and dials. Smarter thermostats for biology. The question now is who builds them first—and whether Origin's early bet on regulatory DNA turns out to be the right problem to solve, or just one piece of a much larger puzzle. For the children waiting on therapies like ELEVIDYS, and the investors who've poured billions into this sector, the answer can't come soon enough.
