The champagne was still flowing from December 2023's FDA approval of Casgevy—the first CRISPR therapy to reach American patients—when regulators delivered a reality check four months later. In April 2024, the agency slapped a class-wide boxed warning on CAR-T therapies treating blood cancers, citing cases of secondary T-cell malignancies that would require lifelong patient monitoring.
That whiplash captured something essential about the $11 billion gene therapy sector: these treatments work, sometimes miraculously. They also carry risks that even the most exhaustive trials are still discovering.
And the problem extends well beyond CAR-T. High-dose AAV gene therapies have triggered liver toxicity and muscle damage in clinical programs. Trials targeting the central nervous system face an exquisite dosing challenge with genes like MECP2, where too much expression proves as harmful as too little. Even as investment surged to $15.2 billion in 2024—a 30% jump year-over-year—researchers confront a question that's both fundamental and vexing: once you've delivered a therapeutic gene into a patient, how do you control where it turns on, when, and how loudly?
Money, Approvals, and Mounting Signals
The sector has hit an inflection point, though not quite the one investors anticipated. Alliance for Regenerative Medicine data shows $11.1 billion flowing into cell and gene therapy across 216 deals in 2025, funding roughly 2,000 active clinical trials. The FDA approved Lenmeldy last March, a $4.25 million gene therapy for metachromatic leukodystrophy that immediately claimed the title of world's most expensive drug. Seven months later, the Centers for Medicare & Medicaid Services launched its Cell & Gene Therapy Access Model, enrolling 33 states in outcomes-based payment agreements for sickle cell therapies.
But beneath the deal flow and regulatory milestones, safety signals keep accumulating. Most gene therapies rely on viral promoters—CMV, CAG, CBA—that deliver therapeutic genes indiscriminately, like turning on every light switch in a house when you only need the kitchen lamp. Reviews of CNS gene therapy trials found that between 50% and 76% use these "always-on" switches, despite the brain's staggering cellular diversity demanding far more precision. Dose-escalation studies have repeatedly hit toxicity ceilings before reaching therapeutic efficacy, particularly in AAV programs targeting neuromuscular diseases.
The industry's response has scattered across delivery engineering, payload optimization, and manufacturing improvements. ARM's recent white papers emphasize AI, automation, and industrialization as critical to scaling production. Fair enough. Yet a parallel conversation has emerged, quieter but potentially more consequential, around transcriptional control—the regulatory DNA sequences that determine gene expression patterns.
The Noncoding Genome's Hidden Instructions
What scientists once dismissed as "junk DNA" actually houses the instructions for where and when genes activate. Enhancers and promoters—regulatory elements that can sit thousands of base pairs from their target genes—orchestrate this control through transcription factor binding sites and chromatin architecture. Natural regulatory elements evolved over millions of years to achieve remarkable cell-type specificity. A neuron-specific enhancer might drive 100-fold higher expression in inhibitory neurons than in neighboring astrocytes, for instance.
Gene therapies rarely capture this precision. Most viral vectors carry compact, strong promoters chosen for packaging constraints and predictable potency, not biological specificity. The consequences show up in clinical data, sometimes starkly. CNS programs delivering constitutively active transgenes face off-target expression in non-neuronal cells, contributing to immune responses and dose-limiting toxicities. Oncology CAR-T trials struggle with what's called "on-target, off-tumor" toxicity—engineered cells attacking healthy tissue that happens to express the target antigen.
Consider MECP2 gene therapy, which illustrates the dosage sensitivity problem with particular clarity. The gene causes Rett syndrome when mutated, but it operates within such a narrow therapeutic window that duplication causes a distinct neurological disorder. Trials employ regulatory elements to modulate expression, yet designing those elements has historically relied on trial-and-error testing of naturally occurring sequences. A synthetic promoter library published in Nature Communications in 2024 demonstrated dynamic ranges of 50-100x across 6,144 compact promoters, proving the design space is vast. And largely unexplored.
Enter the Machines
Origin Bio emerged from Y Combinator's Winter 2026 batch with a pitch that cuts straight to the problem: use AI to design the DNA "switches and dials" controlling therapeutic gene expression. Co-founders Yash Rathod and Malhar Bhide assembled advisors including MIT's Manolis Kellis, a regulatory genomics pioneer, and Nicole Paulk, whose AAV gene therapy work spans neurological diseases and oncology applications.
The San Francisco startup's Axis model, announced last October, combines generative and predictive capabilities for regulatory DNA design. Trained on ENCODE V4 cis-regulatory elements, Axis claims a 6.7% average improvement over DeepMind's AlphaGenome on regulatory binding prediction benchmarks. Under high-affinity design prompts, the model enriched targeted transcription factor motifs up to ninefold. The company emphasizes its proprietary experimental dataset of synthetic regulatory elements validated across cell and tissue types—what they call the "largest" collection of its kind. Independent benchmarking remains pending.
Origin's timing coincides with broader maturation in genomic AI. DeepMind released AlphaGenome in January 2026, a model processing up to one megabase of DNA sequence to predict regulatory effects across splicing, chromatin accessibility, and gene expression. The architecture builds on Enformer, the 2021 transformer model that established 200-kilobase context windows as feasible for regulatory prediction. Research groups have begun applying reinforcement learning and diffusion models to cis-regulatory sequence design, optimizing for cell-type specificity and constraint satisfaction.
The field is finally generating the data infrastructure these models require. Massively parallel reporter assays—MPRAs—now routinely test thousands of synthetic sequences simultaneously, quantifying enhancer activity with statistical rigor. In vivo MPRAs delivered via AAV have screened regulatory elements directly in neurons, closing the gap between cell-culture predictions and tissue-level performance. Single-cell chromatin accessibility atlases mapping open chromatin across human fetal development and adult tissues provide training data that captures cell-state-specific regulatory grammar.
Clinical Validation Takes Shape
Encoded Therapeutics has traveled furthest along the clinical path. The company's ETX101 harnesses regulatory elements to upregulate SCN1A specifically in inhibitory neurons, treating Dravet syndrome, a severe childhood epilepsy. Phase 1/2 interim data presented last December showed a 78% median seizure reduction at the third dose level, with manageable tolerability. The FDA granted Breakthrough Therapy and RMAT designations; Encoded announced plans for a pivotal trial in 2026. The company licensed its regulatory element platform to Lilly and Prevail in a 2024 deal, validating the technology's broader applicability—at least in Big Pharma's estimation.
AskBio took a different route, acquiring Synpromics in 2019 to internalize synthetic promoter design. The PromPT platform uses bioinformatics and machine learning to engineer tissue-specific promoters—liver, muscle, CNS—along with inducible and safety-switch features. The approach reflects industrial-scale promoter engineering, with libraries optimized for manufacturing reproducibility and regulatory filing. Less moonshot, more manufacturing discipline.
ArsenalBio and Outpace Bio represent cell therapy's embrace of transcriptional logic. Arsenal's engineered T cells use logic gates and synthetic circuits to restrict cytotoxic payloads to tumor microenvironments, combining endogenous promoter rewiring with machine-learning-enabled screening. Outpace designs AI-guided proteins that program CAR-T function and persistence, essentially adding software layers to cellular hardware.
Then there's Capsida Biotherapeutics, demonstrating that expression control complements delivery engineering rather than replacing it. The company's engineered AAV capsids achieve 100-fold CNS enrichment with liver detargeting in non-human primates. AbbVie exercised a $40 million option on the technology last January. Pairing cell-selective capsids with tissue-specific promoters could multiply specificity gains—a capsid delivers cargo to the right organ, a promoter activates it in the right cell type. Simple in theory. Devilish in execution.
The Regulatory Path Forward
The FDA is laying groundwork for AI-designed therapeutics, though perhaps more tentatively than the sector would prefer. Draft guidance issued in January 2025 proposes a framework for assessing AI model credibility in drug and biologic submissions, requiring documentation of training data, validation methods, and performance metrics for each "context of use." Origin Bio's eventual regulatory path will hinge on demonstrating that Axis-designed sequences perform as predicted across relevant cell types and disease contexts—likely requiring extensive MPRA validation and potentially in vivo confirmation before filing an investigational new drug application.
The broader regulatory environment is shifting to accommodate the sector's complexities, albeit slowly. The FDA announced increased CMC flexibility for cell and gene therapies in January 2026, recognizing that rigid manufacturing specifications can throttle innovation. CMS's outcomes-based payment model for sickle cell gene therapies, launched with 33 participating states, establishes a template for managing the reimbursement challenge as more multi-million-dollar therapies reach market. Whether it's an adequate template remains hotly debated.
McKinsey projects that AI can accelerate cell and gene therapy R&D and operations end-to-end—from patient selection and safety risk prediction to digitized CMC and manufacturing optimization. The value creation potential is substantial: faster design cycles, higher success rates, lower cost of goods. But realizing that potential requires integrating computational predictions with high-throughput experimental validation. NVIDIA's BioNeMo platform, adopted by Illumina, IQVIA, Arc Institute, and others, is lowering barriers to training bio-foundation models. Dyno Therapeutics uses the infrastructure for AI-guided AAV capsid design; similar approaches are proliferating across the therapeutic development stack.
The Convergence Ahead
The technical trajectory points toward convergence, though on what timeline remains anyone's guess. Larger, multi-modal foundation models spanning DNA, RNA, protein, and cellular phenotypes will combine with richer experimental datasets—MPRAs, single-cell atlases, high-throughput screening—and improved assay-in-the-loop systems. That convergence could compress regulatory DNA design cycles from years to months, particularly for CNS and oncology programs where cell-type specificity is paramount.
ARM's 2026 outlook anticipates "cautious optimism" as the industry awaits multiple Phase 3 readouts and regulatory decisions. The Alliance expects the field to move from scientific promise to operational and commercial maturation, with manufacturing scale and cost-effectiveness becoming competitive differentiators. AI-designed regulatory elements fit that narrative neatly—they address a manufacturing and safety bottleneck simultaneously, potentially reducing required vector doses while widening therapeutic windows.
Whether Origin Bio's Axis model proves transformative or merely incremental will depend on factors the company hasn't yet published: independent benchmarking against state-of-the-art regulatory models, wet-lab validation throughput, assay choices that balance speed with physiological relevance, and transferability to clinical vectors carrying therapeutic cargo. The dataset claims—size, diversity, noise handling—matter enormously. So does how the designed sequences generalize across cell states and from reporter assays to endogenous genomic contexts.
Perhaps the more fundamental question is how quickly the industry adopts precise transcriptional control as standard practice. Every gene therapy program makes tradeoffs between potency, specificity, and safety. The tools to optimize that tradeoff are maturing rapidly, backed by serious capital and serious science. The clinical and financial stakes—billions in investment, millions per treatment, patients with no other options—ensure they'll be tested rigorously.
The champagne moments will keep coming. So will the boxed warnings. What happens in between is where the real work unfolds.
