The announcement arrived with little fanfare. A four-person company in San Francisco, barely a year old, claimed its AI platform had bested Google DeepMind's genomics model at one of gene therapy's thorniest design problems: engineering the regulatory DNA sequences that control when, where, and how loudly therapeutic genes speak inside the human body.
Bold? Perhaps. Implausible? The benchmarks suggest otherwise—though in a field littered with computational promises that never quite survive the lab bench, caution is warranted.
Origin Bio unveiled its Axis model last October, reporting an average 6.7% improvement over DeepMind's AlphaGenome across a battery of prediction tasks: chromatin accessibility, transcription factor binding, and a handful of other regulatory signals that determine whether a gene therapy will work, fail quietly, or trigger something worse. This March, the Y Combinator-backed startup followed with Switch, a public repository of 10,000 AI-designed regulatory sequences optimized for three disease-relevant cell lines—neuroblastoma, liver cancer, and leukemia. Each sequence arrives annotated with predicted activity scores from the Broad Institute's Malinois model, transcription factor motifs, and quality-control metrics.
The company's wager is straightforward, if ambitious. Regulatory DNA—the promoters and enhancers that function as molecular switches and volume dials for gene expression—remains one of the slipperiest design challenges in cell and gene therapy. Get the sequence wrong and a treatment can summon an immune assault, damage the wrong tissues, or simply do nothing. Get it right and you can dial therapeutic gene expression to a Goldilocks zone: enough to work, not so much that it harms.
It's the kind of precision the field has chased for years, mostly through trial and error. Now a handful of startups and academic labs believe machine learning can accelerate the hunt.
A Sector Maturing, With Complications
The cell and gene therapy sector logged 1,905 ongoing clinical trials globally in the first half of 2025, according to the Alliance for Regenerative Medicine. North America led with 844 trials; Asia-Pacific and Europe followed with 750 and 453, respectively. Oncology dominated—64% of the 80 gene therapy trials launched in Q2 alone targeted cancer.
Yet the industry is entering what ARM describes as a "disciplined, sustainable growth phase." Translation: the early euphoria has cooled. Investment deal flow captured 18% of total biotech therapeutic deal value recently, ARM reported this past January, but quarter-over-quarter dealmaking slowed. ASGCT and Citeline tallied 90 announced deals in Q1 2025, down 20% from the prior quarter. One new approval each across gene, cell, and RNA categories marked Q1—modest, steady, unspectacular.
Regulatory winds, at least, are shifting in developers' favor. The FDA signaled in January 2026 that it would "increase flexibility on requirements for cell and gene therapies to advance innovation," and the following month dropped its default two-study standard to a single pivotal study for novel approvals. Europe's EMA unified quality and nonclinical guidelines for investigational advanced therapy medicinal products took effect last July, streamlining pathways across member states.
Against this backdrop, AI has begun threading into nearly every corner of the development stack. DeepMind published AlphaGenome in Nature on January 28, 2026—a long-context model trained on 5,930 human and 1,128 mouse genomic signals that predicts regulatory activity across sequences stretching up to 1 megabase. The model interprets genetic variants, maps enhancer-gene interactions, and helps prioritize candidate regulatory elements. It's available via a noncommercial API.
Academic labs and startups are racing to commercialize similar capabilities. The Broad Institute, Jackson Laboratory, and Yale published CODA and Malinois in Nature back in October 2024, demonstrating machine-learning-guided design of synthetic cis-regulatory elements validated by massively parallel reporter assays. ENCODE V4—which cataloged 2.37 million regulatory DNA elements as of January 2026—provides the foundational training data for many of these models.
The question isn't whether AI will reshape regulatory DNA design. It's whether the tools can deliver what they promise when they leave the server and enter the messy reality of living cells.
The Safety Imperative
The push to engineer better regulatory DNA is driven by a catalog of well-documented failures. Adeno-associated virus (AAV) vectors—the workhorse delivery vehicle for many gene therapies—can cause dose-limiting hepatotoxicity, dorsal root ganglion lesions, and off-target expression in tissues where therapeutic genes have no business being. A 2025 review in the Royal Society of Chemistry documented cases of DRG toxicity following intra-CSF AAV delivery. A fatal event involving a blood-brain-barrier-crossing AAV capsid the same year sent shockwaves through the field, though details remain sparse in the public domain.
Promoter and enhancer design offers a way to mitigate some of these risks. The FDA-approved Duchenne therapy Elevidys uses the MHCK7 promoter paired with a cardiac enhancer to bias expression toward skeletal and cardiac muscle, reducing systemic spillover. Other groups are layering in microRNA target sites to silence transgenes in specific cell types, or engineering synthetic inverted terminal repeats to attenuate toxicity.
But designing these sequences by hand is slow, empirical, and often suboptimal. Researchers typically borrow promoters from one biological context and hope they'll work in another—a strategy that leaves efficacy and safety on the table. Sometimes it works. Sometimes it doesn't. There's rarely a good way to predict which outcome you'll get until you've already spent months in the lab.
The Team and the Tool
Origin Bio's founders, Yash Rathod and Malhar Bhide—both computer science graduates from the University of Illinois Urbana-Champaign—are building what they describe as "the largest proprietary dataset of synthetic regulatory sequences across diverse cell states." Their scientific advisors bring heavyweight credentials: Manolis Kellis at MIT and the Broad Institute (a leader in regulatory genomics), Nicole Paulk at Siren Bio and UCSF (AAV gene therapy), and Rashid Bashir at UIUC and the Chan Zuckerberg Biohub Chicago.
Axis, the company's flagship model, is a multifunctional transformer that both generates new regulatory DNA and predicts its functional activity. Trained on the ENCODE V4 registry, it reports benchmark gains over AlphaGenome on tasks spanning chromatin accessibility and transcription factor binding. According to Origin's October blog post, the model can enrich target transcription factor motifs up to 9× via prompt-guided generation—essentially, you tell it which transcription factors you want to activate or repress, and it designs DNA sequences to match.
In March, the company published internal benchmarking showing that the Muon optimizer reached target validation perplexity with roughly 37% fewer FLOPs than the Adam optimizer—a technical detail that hints at Origin's focus on computational efficiency for genomic sequence modeling. For those who don't speak machine learning, that translates to: their approach is faster and cheaper to run at scale.
Switch, the public dataset released this past March, is Origin's first community-facing product. The 10,000 sequences span proximal enhancer-like elements designed for three cell lines: SK-N-SH (neuroblastoma), HepG2 (liver cancer), and K562 (leukemia). Each entry includes predicted activity scores from Malinois, transcription factor motif annotations, sequence quality-control metrics, and 3D double-stranded DNA structures modeled by Protenix. The sequences are graded to titrate gene expression levels—a feature the company positions as a tool to systematically map dose-response curves for therapeutic genes like IL-10 in tumor-infiltrating lymphocytes.
It's a bid to gain traction through open access—a strategy that mirrors DeepMind's noncommercial API release but with a narrower, therapy-focused scope.
The Competition
Origin Bio is hardly alone in this space, though its claim to outperform DeepMind has drawn attention. Annogen, based in Europe, runs a SuRE-based promoter and enhancer screening platform with what it calls "billions of measurements" and counts Bayer, BASF, Orchard Therapeutics, Novo Nordisk, uniQure, MeiraGTx, and Pfizer among its customers. Annogen announced an R&D collaboration with JURA on variational synthesis for cell-type-specific regulatory DNA sometime in 2025.
AskBio acquired Synpromics back in 2019 and has since maintained a synthetic promoter platform. The company's ASGCT agenda included a session on "AI use for promoter and transgene design," and its materials describe data-driven luciferase benchmarking against conventional promoters like CBA. Chromatin Bioscience offers synthetic promoter design services for cell and gene therapies, positioning itself as a partner for developers looking to optimize expression profiles.
Meanwhile, companies like Senti Biosciences and Strand Therapeutics are embedding logic-gated circuits into allogeneic CAR-NK and programmable mRNA constructs, respectively. Senti presented promoter optimization work at the Society for Immunotherapy of Cancer in December 2023 and is advancing clinical programs with constructs that sense and respond to tumor microenvironments. Strand is developing systemically delivered mRNA circuits with liver-avoidance logic; an IND for a systemic construct was anticipated in the first half of 2026, at least as of early 2025.
Tune Therapeutics raised more than $175 million in Series B financing in January 2025 to advance its TEMPO epigenome-editing platform, which reprograms gene expression without altering DNA sequence. The company's lead candidate, TUNE-401 for hepatitis B, is in clinical trials, and the platform has demonstrated durable epigenetic repression in nonhuman primates.
On the capsid side—where the delivery vehicle itself is the target of optimization—Dyno Therapeutics and Capsida are using AI to evolve AAV variants with improved tissue tropism and reduced toxicity. Roche exercised an option with Dyno in January 2025 for a next-generation AAV vector for neurological gene therapy. Capsida triggered a $40 million milestone payment from AbbVie the same month after demonstrating brain-targeting AAV capsids in nonhuman primates with reduced liver and DRG exposure. Both companies are collaborating with NVIDIA's BioNeMo platform to accelerate in silico capsid design.
The convergence of AI-driven capsid evolution and regulatory DNA design is beginning to reshape how the field thinks about vector engineering. Where early gene therapies relied on off-the-shelf promoters and naturally derived AAV serotypes, developers are now assembling modular platforms with computationally optimized components at every layer.
It's an industrial mindset creeping into what was, until recently, a largely artisanal discipline.
The Proof Is in the Assay
Whether Axis's benchmarks translate into real-world therapeutic impact remains to be proven. Massively parallel reporter assays—the gold standard for validating synthetic regulatory elements—can vary widely across protocols and cell contexts, as a 2025 Genome Biology study documented. Single-nucleotide-resolution MPRA methods are advancing, but harmonization across labs is still a work in progress.
The regulatory DNA design space remains fragmented. Wet-lab screening platforms, academic models, and startup tools are all competing for developer attention, each with slightly different strengths and blind spots. Origin Bio's public dataset is a bid to cut through that fragmentation by inviting scrutiny—and adoption.
Regulatory considerations loom, too. The EU AI Act, which entered into force in August 2024, will impose staged compliance requirements starting in August 2026 for most provisions and August 2027 for high-risk AI systems embedded in regulated products. Discovery-focused tools like Axis and AlphaGenome likely face lighter near-term burdens than AI embedded in medical devices, but the regulatory landscape for AI in drug development is still taking shape. The FDA's emerging interest in platform technology designations and advanced manufacturing pathways could favor modular regulatory DNA libraries if they can demonstrate consistent performance across multiple therapeutic programs.
ARM's sector snapshot highlighted platform designations as a potential accelerant for reusable components, and Origin Bio's graded sequence library could fit that bill—if it scales beyond the initial three cell lines.
Can a Four-Person Team Sustain Momentum?
For now, the field is watching to see whether a four-person startup can sustain its momentum against DeepMind, academic consortia, and established commercial platforms. The March Switch release is a proof point, not a product. But in a sector where a well-chosen promoter can mean the difference between a therapy that works and one that harms, even incremental gains in predictive accuracy matter.
The broader question is whether AI can compress the iterative cycles that have historically made regulatory DNA design so painfully slow. If tools like Axis can reliably generate sequences that pass experimental validation on the first or second try, they'll shave months—maybe years—off preclinical timelines. If they can't, they'll join the long list of genomics models that looked impressive on benchmarks but struggled to generalize beyond their training distributions.
Origin Bio is making its dataset public to invite that scrutiny. Whether the community validates or refutes the claims will determine whether this is a turning point or a footnote.
In gene therapy, as in most of biology, the map is not the territory. A model can predict all it wants. The real test comes when you put the DNA into cells and see what happens. Origin Bio is betting it's cracked the code—or at least gotten closer than anyone else. The experiments to prove it, one way or another, are just beginning.
