The two deaths happened quickly. Non-ambulatory Duchenne muscular dystrophy patients, their bodies already ravaged by progressive muscle degeneration, received Sarepta Therapeutics' Elevidys—a $3.2 million gene therapy meant to give them hope. Instead, acute liver failure. A temporary shipment pause followed. Then, in July 2025, came the FDA's reckoning: a boxed warning that would restrict the therapy's use to ambulatory patients four years and older, effectively shrinking its addressable population to a fraction of those it was designed to help.
The tragedy laid bare something the gene therapy industry has quietly struggled with since its earliest triumphs. The problem isn't always the therapeutic gene itself—the biological instruction to produce dystrophin or engineer cancer-fighting T cells. It's the genetic switches that control when those genes turn on, how loudly they express, and in which tissues. Promoters, enhancers, the entire regulatory DNA apparatus: when they misfire, patients die. Or worse, they survive with complications that haunt them—and the therapy's commercial prospects—for years.
And the FDA has noticed. The agency added a boxed warning on T-cell malignancies to all BCMA- and CD19-directed CAR-T therapies in January 2024. Elevidys got its turn in July 2025. These aren't statistical footnotes buried in supplementary materials. They're label changes that fundamentally alter who can receive a treatment, who will pay for it, and whether investors will back the next one.
Now a new cohort of computational biology startups is placing a bet that feels almost obvious in hindsight. The next leap in gene therapy safety, they argue, won't come solely from better viral delivery vehicles or optimized payloads. It will come from redesigning the DNA sequences that govern expression—the control layer determining whether a micro-dystrophin construct floods the liver or concentrates in skeletal muscle, whether a CAR burns out through chronic signaling or maintains function over years.
It's a wager on regulatory elements, the genetic dimmer switches most researchers have treated as interchangeable parts borrowed from the nearest available library. Perhaps that was always going to be a problem.
The Industry Recalibrates
The gene therapy sector entered 2026 in what the Alliance for Regenerative Medicine diplomatically termed a phase of "disciplined, sustainable growth." Translation: first-generation products revealed hard truths about pricing, durability, and safety that no amount of venture capital could smooth over.
Investment in cell and gene therapies reached roughly $11.1 billion in 2025 across equity and venture sources, according to ARM's January briefing. Market analysts remain optimistic, projecting gene therapy revenues to balloon from $6.6 billion in 2025 to $19.3 billion by 2034, per IMARC forecasts. But pipeline dynamics tell a different story—one that's considerably less tidy.
The third quarter of 2025 saw new trial initiations across gene, cell, and RNA therapies decline to the lowest quarterly figure in over a year based on ASGCT-Citeline data. Roughly 1,905 trials are ongoing globally now, with North America accounting for 844. The slowdown isn't for lack of ambition. The sector is digesting lessons from products like Bluebird Bio's Skysona and Zynteglo, which carry $3 million and $2.8 million list prices respectively but have struggled mightily to gain commercial traction. Bluebird's acquisition by Carlyle and SK Capital in February 2025 underscored just how acute that strain has become.
Yet the FDA has shown signs of pragmatism. In January 2026, the Center for Biologics Evaluation and Research announced increased flexibility around certain chemistry, manufacturing, and controls requirements—a signal that the agency is open to platform approaches demonstrating reproducible safety and potency without reinventing assays for every single construct. That flexibility matters considerably more when you're iterating on regulatory DNA sequences rather than starting from scratch with each new viral serotype or payload.
The Expression Control Dilemma
Gene therapies deliver potent biological instructions. But instructions require switches—promoters that determine baseline transcription, enhancers that amplify it in specific cell types or states, silencers that shut it down where it doesn't belong. Get the combination wrong and a therapy designed to restore dystrophin in muscle might trigger catastrophic immune responses in the liver. Or a CAR-T construct might stimulate continuously, driving T cells toward exhaustion before they ever clear the tumor.
Elevidys itself illustrates what's at stake. Clinical and preclinical work on Duchenne therapies has documented how using a muscle-biased promoter like MHCK7, paired with a cardiac enhancer, can concentrate micro-dystrophin expression in skeletal and cardiac muscle while reducing off-tissue exposure. Studies published between 2017 and 2025 showed improved functional outcomes. Yet even with tissue-specific regulatory elements, the therapy's high AAV dose and the particular vulnerability of non-ambulatory patients to hepatotoxicity led to those fatal adverse events.
CAR-T therapies face a different variant of the same problem. Tonic signaling—when CARs signal even without antigen engagement—can drive chronic activation, cytokine release, and T-cell exhaustion. Recent literature suggests promoter choice matters: milder promoters like MND or EFS can reduce CAR surface density, lowering tonic signaling compared to strong EF1α backbones. It's fundamentally a tuning problem, and historically the dial has been coarse.
Then there's immunogenicity. Incorporating microRNA target sites into AAV vectors—specifically miR-142-3p, enriched in antigen-presenting cells—can reduce transgene expression in those cells, blunting the immune system's ability to recognize and attack the therapy. Studies from 2013 through 2025 validated the approach across multiple constructs. The FDA notes that over 90 percent of humans carry preexisting antibodies to some AAV serotypes. Designing around immune surveillance has become non-negotiable.
The Computational Turn
Origin Bio, which emerged from Y Combinator's Winter 2026 cohort, announced its Axis platform in October 2025. Axis is what the company describes as a multifunctional model—one that predicts regulatory element activity from DNA sequences, generates novel regulatory DNA guided by transcription factor-binding "prompts," and claims to unify multiple design tasks in a single framework. The company reported a 6.7 percent average improvement over DeepMind's AlphaGenome on regulatory element activity prediction benchmarks, and up to 9× enrichment of targeted transcription factor motifs when using high-affinity prompts versus low-affinity ones.
Those are company claims, it's worth noting. Independent peer-reviewed validation had not surfaced publicly as of early 2026.
Origin's founders—CEO Yash Rathod and CTO Malhar Bhide, both with University of Illinois computer science backgrounds—are building what they call a proprietary dataset of synthetic regulatory sequences tested across diverse cell states. The advisory roster includes Dr. Manolis Kellis of MIT and the Broad Institute, Dr. Nicole Paulk (an AAV expert and CEO of Siren Bio), and Dr. Rashid Bashir of UIUC and the CZ Biohub Chicago. A March 2026 blog post benchmarked optimizer algorithms for regulatory DNA modeling, noting that the MuonW optimizer reached target validation perplexity with roughly 37 percent fewer FLOPs than AdamH—an efficiency play that matters considerably when scaling generative design loops.
Asimov launched its AAV Edge suite in September 2024, offering AI-designed tissue-specific promoters, sequence optimization for in vivo expression, and tools to silence genes of interest during AAV production—reducing manufacturing toxicity. The platform positions regulatory DNA design alongside capsid selection and production-cell engineering in what amounts to a vertically integrated workflow.
The broader competitive landscape includes Dyno Therapeutics, which in January 2025 saw Roche exercise an option for a next-generation AAV capsid targeting neurological indications. Voyager Therapeutics has licensed its TRACER capsid platform to Novartis multiple times in 2024 and 2025. Capsida Biotherapeutics notched a $40 million option exercise from AbbVie for a CNS program in late 2024.
These are delivery plays—capsid AI to get payloads to the right tissues. But they leave the expression control problem largely unaddressed unless paired with smarter regulatory elements. Getting to the right tissue is half the battle; knowing how much to express once you're there is the other half.
Elsewhere, companies like Senti Bio are embedding logic gates into CAR-NK constructs (CD33/FLT3 OR with NOT gates), with preliminary complete remissions reported in relapsed/refractory AML as of AACR data presented in April 2025. Outpace Bio introduced an OP1 promoter and modular safety switches at SITC in November 2024, designed to reduce CAR-T exhaustion and localize cytokine secretion. These are circuit-level interventions—using regulatory DNA to add programmability and fail-safes directly into the therapeutic construct.
From Theory to Practice
Academic work has kept pace, though translating it to the clinic remains another matter. In May 2025, a preprint described constrained reinforcement learning for generating cell-type-specific regulatory sequences. A March 2026 paper demonstrated diffusion transformers for regulatory DNA generation with reduced memorization—a nod to the overfitting risks when training models on finite MPRA datasets. Another January 2026 preprint tackled robustness under distribution shifts, acknowledging that regulatory models trained on in vitro data don't always transfer cleanly to in vivo contexts. Which is to say: predictions made in a dish don't always hold up in a mouse, let alone a human.
The foundation for these efforts was laid by tools like Enformer, a 2021 Nature Methods transformer that integrated 100 to 200 kilobase genomic context and improved prediction of gene expression and chromatin profiles. DeepMind's AlphaGenome, launched in June 2025 with a Nature paper following in January 2026, extended the "Alpha" line to non-coding genome prediction and proposed applications in synthetic regulatory design. These predictors provide the evaluators—ways to score candidate regulatory sequences without exhaustive wet-lab testing—that make generative design feasible in the first place.
The clinical payoff, though, remains largely speculative. No approved gene or cell therapy has yet publicly disclosed that its regulatory elements were designed via generative AI. The regulatory DNA strategies that have reached patients—MHCK7 in Elevidys, miR-142 detargeting in research-stage AAV constructs, promoter tuning in CAR-Ts—were largely hand-engineered or selected from known motifs.
Axis, AAV Edge, and similar platforms are tools for what comes next. Therapies entering development in 2026 and beyond, where the design-build-test cycle for regulatory elements might compress from years to months. Whether that compression actually happens is the $19 billion question.
What Comes Next
The FDA's January 2026 flexibility announcement and the ongoing rollout of CMS's Cell and Gene Therapy Access Model—now active in 33 states, Washington D.C., and Puerto Rico for sickle cell therapies—create something of a window for smarter therapeutic design. Outcomes-based contracts, like Bluebird's Zynteglo agreement with up to 80 percent rebates if durability targets aren't met, mean manufacturers have real financial incentives to derisk expression-related failures. If a therapy's regulatory DNA can demonstrably reduce off-tissue exposure, lower immunogenicity, or improve durability, that becomes a value proposition for payers and regulators alike.
The competitive intensity in capsid engineering—Dyno, Voyager, Capsida, 4D Molecular Therapeutics—shows no signs of abating. But capsids are only half the equation. A perfectly targeted delivery vehicle still needs to know how much transgene to express, in which cell states, and for how long. That's the layer regulatory DNA controls, and it's becoming a parallel differentiation axis.
Less serotype swapping, more circuit-level precision. That's the pitch, anyway.
In vivo gene editing programs targeting cardiovascular indications like PCSK9 and LPA or autoimmune diseases are moving toward what industry insiders are calling "one-shot functional cures." Their success will hinge on precise delivery and precise expression control—promoters that activate only in hepatocytes, enhancers that respond to disease state, silencers that shut down in off-target tissues. The same logic applies to emerging non-viral delivery platforms using lipid nanoparticles or virus-like particles, which bypass some AAV constraints but still require regulatory elements to govern transgene behavior.
ARM's January 2026 briefing characterized the sector as entering a phase of "disciplined, sustainable growth," adapting to first-generation product lessons around pricing, durability, and safety. If that discipline translates into therapies with tighter therapeutic indices—less toxicity, more consistent efficacy—it will be because the field stopped assuming regulatory elements could be borrowed from the nearest available promoter library and started engineering them as carefully as the payloads they control.
The AI tools exist now. The regulatory tailwinds are forming. The clinical need, underscored by boxed warnings and label restrictions that reshape entire commercial strategies overnight, is undeniable. Whether generative regulatory DNA becomes standard practice or remains a boutique optimization step will depend on how quickly the first AI-designed elements move through validation, into constructs, and ultimately into patients who need them.
The race isn't to make gene therapies work—they already do, often spectacularly. It's to make them safe enough, reliable enough, and targetable enough that a $3 million therapy doesn't come with a sudden indication restriction six months post-launch.
That's the control problem. And for the first time, the tools to solve it may be more than theoretical. Whether they're sufficient is another question entirely.
