Patients receiving Sarepta Therapeutics' muscle gene therapies—treatments designed to fight muscular dystrophy—began showing signs of acute liver failure. By July, the FDA had launched an investigation. Come December, the agency slapped new safety warnings on the drug and narrowed its approved use.
The irony was cruel: the therapeutic gene itself worked as intended. What proved lethal was where it activated. Instead of confining expression to muscle tissue, the gene switched on in the liver, with fatal consequences.
For an industry projecting growth from roughly $27.02 billion in 2025 to $232.22 billion by 2035, the Sarepta episode landed like a wake-up call. Gene therapy's promise has always been straightforward enough—fix broken genes, cure disease. But delivering DNA into cells is only half the battle. The harder part, it turns out, is controlling what happens once it arrives.
That control resides in regulatory DNA: the promoters and enhancers that act as molecular dimmer switches, determining when, where, and how loudly a gene expresses. Scientists have borrowed these sequences from nature or tinkered with them manually for decades. Now a new cohort of startups and academic labs is trying something different—using machine learning to design them from scratch.
Whether the approach works could determine not just which companies thrive, but whether gene therapy can scale beyond rare diseases without killing people in the process.
Toxicity Patterns the Industry Didn't Fully Anticipate
By early 2026, the Alliance for Regenerative Medicine was pointing toward the field's trajectory entering a more disciplined, sustainable growth phase. Industry insiders understood the translation: the euphoria had cooled. Clinical trial activity remained strong—125 new studies launched in the third quarter of 2025 alone—but regulators and companies were confronting safety signals that early studies hadn't fully predicted.
Sarepta wasn't an outlier. The FDA placed clinical holds on Intellia Therapeutics' CRISPR programs following liver toxicity reports, lifting them only after the company submitted additional safety data. These incidents followed subtler warning signs that had accumulated over years. CAR-T cancer therapies caused patient exhaustion when promoters drove excessive receptor expression. Viral gene therapy vectors triggered immune responses when therapeutic proteins showed up in tissues where they shouldn't have been.
The common denominator? Imprecise gene expression. Traditional promoters—DNA sequences like EF1α or CMV that researchers pull from well-studied viral genomes—drive strong, widespread activity. Great for laboratory experiments. Problematic when you're trying to cure a human being.
Studies published between 2020 and 2024 demonstrated that powerful promoters in CAR-T constructs increased what researchers call "tonic signaling," effectively exhausting the modified T cells before they could fight cancer. Cardiac gene therapy teams shifted toward tissue-specific promoters, adding microRNA target sites—molecular "off switches" that silence genes in liver cells or neurons—to prevent expression in the wrong places.
The market dynamics are unforgiving. That projected $232.22 billion valuation by 2035 assumes therapies can be manufactured safely at commercial scale. Without better expression control, the math doesn't work.
The Limited Menu of Natural Switches
Promoters and enhancers sound simple in theory—short stretches of DNA, typically hundreds to thousands of base pairs long, that recruit the cellular machinery needed to activate genes. Their behavior, though, is maddeningly context-dependent. A promoter that performs beautifully in mouse muscle might wreak havoc in human liver. An enhancer fine-tuned for one disease might trigger unintended pathways in another.
For years, scientists worked from a constrained toolkit. They raided viral genomes for potent promoters, borrowed tissue-specific sequences from genes biology had already optimized, or hand-engineered variants through trial and error. Synpromics, a Scottish biotech acquired by Asklepios BioPharmaceutical in 2019, built a business around synthetic, tissue-selective promoters. The company's PromPT platform offered tunable switches—an improvement over one-size-fits-all options, certainly, but still limited by what human designers could imagine.
Duchenne muscular dystrophy illustrates the challenge. Micro-dystrophin gene therapies rely on a promoter called MHCK7, a muscle-biased sequence with cardiac enhancement properties. Clinical studies from 2020 onward validated the basic approach. The promoter concentrated expression in skeletal and cardiac muscle, as intended.
Yet even these carefully selected switches couldn't prevent the off-target liver expression that, at high doses or in vulnerable patients, led to the deaths that forced the FDA to intervene in 2025.
What the field needs isn't just incrementally better versions of existing switches. It's the capacity to design entirely new ones, optimized for therapeutic contexts that evolution never encountered.
When Prediction Became Generation
Machine learning entered regulatory genomics gradually at first. Researchers spent years training neural networks to predict which DNA sequences might function as enhancers or promoters, using massive datasets from initiatives like the ENCODE project. Models such as Enformer, published in 2021, achieved correlations around 0.81 with experimental measurements of promoter activity.
Progress, sure. Revolutionary? Not quite.
The shift came when teams realized they could flip the models around—using them to generate sequences rather than merely evaluate ones that already existed. In October 2024, researchers from the Broad Institute, Yale, and Jackson Laboratory published work on Malinois, a deep convolutional network that both predicted cell-type-specific regulatory activity and guided the design of new elements. When tested in massively parallel reporter assays—high-throughput experiments measuring thousands of sequences simultaneously—the Malinois-designed sequences outperformed naive selections for specificity in cancer cell lines.
Then DeepMind entered the picture. In January 2026, the Alphabet subsidiary released AlphaGenome, a unified model predicting regulatory effects across megabase-scale genomic contexts. The accompanying Nature publication and code release sent ripples through both academic and industry circles. AlphaGenome didn't design sequences itself, but its prediction accuracy established a new standard.
That same month, the ENCODE project released an expanded registry of candidate regulatory elements, incorporating data from multiple experimental platforms across various tissue types. The datasets are massive—millions of sequences with measured activity—and expanding. Research groups at the Broad Institute have characterized millions of synthetic promoters, work aimed at decoding what they call the "grammar" of gene regulation. Some of that research appeared in Nature publications accepted in late December 2025.
For startups eyeing this space, the landscape presents both opportunity and risk. The data is increasingly public. The models trend toward open-source. The competitive advantage lies not in hoarding information, but in execution.
The Dataset Moat Strategy
Origin Bio emerged from this environment in 2025, founded by Yash Rathod and Malhar Bhide, both computer science graduates from the University of Illinois Urbana-Champaign. Rathod had won first prize in the 2022 OpenCV AI Research Competition; Bhide brought machine learning experience from Wadhwani AI and had co-founded Automorphic, a company that went through Y Combinator's Summer 2023 batch. Both published disease modeling research in Scientific Reports before pivoting to regulatory DNA.
The San Francisco-based duo joined Y Combinator's Winter 2026 cohort with a straightforward thesis: gene therapy's safety crisis stems from an inability to precisely control gene expression, and machine learning can solve it—provided you build your own experimental dataset, not just rely on public sources.
Origin's advisor roster signals ambition. Manolis Kellis from MIT and the Broad Institute, a leading figure in regulatory genomics. Nicole Paulk from UCSF, founder of Siren Biotechnology and an expert in AAV vector manufacturing. Rashid Bashir, dean of UIUC's Grainger College of Engineering.
In October 2025, Origin announced Axis, described as a multifunctional model that both generates regulatory DNA and predicts function. The company claimed Axis outperformed DeepMind's AlphaGenome by 6.7% on average for predicting regulatory element binding activity. High-affinity sequences designed by the model showed up to 9x enrichment of targeted transcription factor motifs compared to baseline approaches.
Notably, Origin reported that training the model on sequence generation improved its prediction accuracy—suggesting that generative and discriminative tasks in this domain reinforce each other. The model was trained on ENCODE data and benchmarked using independent tools, including the Malinois model. But the company's messaging emphasized that public datasets are merely a foundation. Origin is building what it characterizes as a proprietary library of experimentally validated sequences, initially targeting cancer and central nervous system applications.
By early March 2026, the team was publishing technical blog posts on computational efficiency, reporting that a specialized optimizer reached target validation metrics with roughly 37% fewer floating-point operations than standard methods. One post emphasized that DNA modeling differs fundamentally from natural language processing—DNA has only four tokens, follows a non-Zipfian distribution, and responds poorly to techniques borrowed blindly from large language models.
Origin's pitch to pharmaceutical partners is direct: commission tissue-specific, tunable switches for your therapeutic gene, reducing off-target expression and toxicity risk. The company lists four team members as of March 2026. Glasswing Ventures includes Origin in its portfolio, though no funding amounts have been disclosed publicly.
A Crowded but Not Winner-Take-All Field
The regulatory DNA design space is filling up, but multiple approaches can coexist. Dyno Therapeutics focuses on AI-designed AAV capsids—the viral protein shells that determine which organs a gene therapy reaches—with partnerships announced between 2020 and 2021 spanning Novartis, Roche, Sarepta, and Astellas. Senti Biosciences builds logic-gated gene circuits that integrate multiple biological inputs, presenting promoter work at conferences like AACR and ASGCT in recent years.
Academic labs continue releasing open models. Genentech published regLM for realistic regulatory element design in late 2025. Reinforcement learning approaches like Ctrl-DNA appeared in May 2025. Frameworks addressing distribution shifts between different cell types and experimental assays emerged in January 2026.
Each strategy has trade-offs. Capsid engineering gets genes to the right organ but doesn't prevent expression in off-target cells within that organ. Logic circuits add programmability at the cost of construct complexity, potentially complicating manufacturing and regulatory approval. Open models democratize access but lack the proprietary datasets that might differentiate performance in therapeutically relevant contexts.
The immediate commercial catalyst is regulatory pressure. FDA presentations and guidance documents from late 2025 and early 2026 emphasize post-approval safety monitoring for cell and gene therapies. Companies will need to demonstrate not just efficacy but safety across diverse patient populations. Expression control—achieved through optimized promoters, enhancers, and molecular "off switches"—offers a concrete lever.
Clinical validation will separate hype from substance. Synthetic regulatory elements need to prove themselves in human trials, not just in dish-based assays or mouse studies. That timeline stretches across years. Companies that initiated partnerships around 2020 are only now potentially seeing early clinical data. Origin, Senti, and other recent entrants are at the beginning of that arc.
The most interesting question may not be which platform wins, but how competitive dynamics evolve as datasets grow and models improve. AlphaGenome's January 2026 release as open code suggests basic prediction capability could commoditize quickly. If Origin's claim about generative training improving prediction accuracy holds up under peer review, other groups will likely adopt similar architectures.
The proprietary experimental datasets Origin is building could matter—assuming the company validates them rigorously and achieves enough volume to cover therapeutically relevant cell types and disease contexts. That's a substantial assumption.
Chastened but Not Broken
For gene therapy developers, the sales pitch is compelling: safer therapies, smoother regulatory paths, reduced late-stage clinical risk. The industry absorbed hard lessons throughout 2025 about what happens when expression control fails. Machine learning won't solve every gene therapy problem—manufacturing challenges, immune responses, delivery obstacles all remain. But for the specific challenge of designing DNA switches that activate in the right place, at the right time, and at the right intensity, computational approaches might be the only viable path forward.
The cell and gene therapy sector entered 2026 sobered but functional. Trial activity remained robust. Regulatory approvals continued, including gene therapies for rare diseases like recessive dystrophic epidermolysis bullosa and macular telangiectasia. Manufacturing partnerships continued scaling. The underlying science works.
Whether the industry can engineer enough precision to match its ambition remains an open question. AI-designed regulatory elements won't solve that challenge alone. But they might be a necessary component of therapies that are both transformative and safe—assuming the models deliver on their promise, the datasets prove relevant, and clinical validation follows years from now.
The deaths from Sarepta's therapy weren't a crisis that ended the field. They were a reminder that getting DNA into cells was always the easier part. Controlling what it does once it arrives—that's the work that remains.
