AI-Designed AAV Capsids Are Solving Gene Therapy's Delivery Problem

$15 million upfront. That's what Astellas paid for a license to use a virus shell designed by artificial intelligence. Not for a drug candidate, not for a platform. For a single adeno-associated virus (AAV) capsid engineered to deliver genetic payloads to skeletal muscle.

The Problem That Caps AI Is Solving

Gene therapy has a delivery problem. The AAV capsids that carry therapeutic genes into cells were discovered in nature, not designed for medicine. AAV9, the most commonly used serotype, requires high systemic doses to reach therapeutic levels in muscle tissue. Those high doses trigger liver toxicity, drive manufacturing costs, and limit the patient populations that can be treated.

The numbers tell the story. Two gene therapy trials (clinical identifiers NCT03368742 and NCT03199469) were placed on clinical hold by the FDA due to severe liver-related adverse events linked to high-dose AAV administration. The underlying capsid was the problem, not the therapeutic payload.

The field has known this for years. Directed evolution and rational design produced incremental improvements. A capsid here, a tropism shift there. But the sequence space of AAV capsid proteins is vast: a 7-amino-acid insertion in the VP1 protein generates 1.28 billion possible variants. Traditional screening methods cannot navigate that space. Machine learning can.

Dyno's CapsidMap: Biology's Search Engine

Dyno's approach differs from earlier capsid engineering efforts in one fundamental way: it does not start from a natural capsid and mutate outward. Instead, its platform builds sequence-function maps from billions of in vivo measurements gathered from multiplexed NHP experiments. It learns which sequence motifs correlate with which delivery properties and then generates candidates that optimize across all of them simultaneously. Transduction efficiency, tissue tropism, immune evasion, manufacturability. The platform handles them in one pass.

"We're achieving safer, more effective therapeutic delivery to skeletal and cardiac muscle at low doses that were out of reach just a few years ago," Adrian Veres, M.D., Ph.D., Chief Scientific Officer and Cofounder, said in the company's April 2026 announcement.

The results bear that out. Dyno-bn8, licensed by Astellas for skeletal muscle delivery, achieved 74% skeletal myofiber transduction and 16% cardiomyocyte transduction at 5.2×10¹² vg/kg in NHPs. The same capsid transduced 50% of satellite cells at 1.6×10¹³ vg/kg, opening the door to gene editing applications in muscle stem cells. It engaged a receptor conserved across NHPs and humans, supporting translatability, and was compatible with existing AAV9 manufacturing processes. The field has learned to value that after a decade of seeing promising preclinical capsids fail to scale.

Two Deals That Validate the Model

The company has now licensed AI-designed capsids to two of the largest gene therapy developers in the world. Roche exercised an option for a CNS capsid in January 2025, paying an undisclosed license fee and taking responsibility for preclinical and clinical development. Astellas followed in April 2026 with a $15 million upfront payment for a muscle-targeted capsid. Clinical development, regulatory, and commercial milestone payments plus royalties apply on top.

The significance of these deals extends beyond the company. An AI-designed biological sequence has now passed the diligence process of two major pharmaceutical companies and resulted in binding financial commitments. The technology readiness level has shifted from "interesting platform" to "licensed asset." A protein shell 26 nanometers in diameter.

This is the second material proof point for AI-designed biologics as a category. The first came from protein therapeutics, where AI-designed antibodies and enzymes entered clinical development. Dyno's capsids are different: they do not treat disease directly. They enable the delivery of other treatments. The licensing model is also distinct. It does not develop its own gene therapies. It designs capsids and licenses them, collecting upfront fees, milestones, and royalties while partners bear the clinical and commercial risk.

Beyond Muscle: The Full Portfolio Strategy

It has built a portfolio of capsids targeting multiple tissue types rather than betting on a single indication. At the 2026 ASGCT Annual Meeting in May, the company launched two new capsids (Dyno-9zh for CNS delivery and Dyno-n96 for muscular delivery) alongside updated in vivo results for Dyno-yp2 (CNS) and Dyno-bn8 (muscle).

The strategy is deliberate. Gene therapy developers working on neurological disorders need capsids that cross the blood-brain barrier efficiently. Developers working on muscular dystrophies need capsids that transduce skeletal muscle at low doses. Developers working on ocular indications need capsids that reach the retina. Each tissue is a separate delivery problem, and Dyno is systematically solving them one by one.

A 2024 Nature Communications paper from the company's scientific cofounders demonstrated the methodology: a "systematic multi-trait protein optimization paradigm" that combines production fitness models with functional fitness models to design capsids that perform across multiple desired traits simultaneously. The approach, called Fit4Function, starts from a library that uniformly samples the 7-mer sequence space, 1.28 billion variants. It filters through production fitness first, then functional screens in vitro and in vivo, then machine learning model building, then multi-function library design.

Meanwhile, researchers at San Diego State University published a preprint on bioRxiv in May 2026 demonstrating a computational method to design larger AAV capsids at the T=3 icosahedral architecture, 45 nm in diameter versus the native 26 nm, with a potential cargo capacity of 35 kilobases versus the native 4.7 kb. If validated experimentally, this would remove the single most persistent limitation of AAV vectors: their inability to carry large genes like dystrophin or CRISPR systems. The method uses AI-based protein folding to design VP3 trimers that assemble into the larger architecture.

What the Market Is Pricing

Dyno has raised $100 million in venture funding and secured collaborations with Roche, Astellas, Sarepta, and NVIDIA. The company's total disclosed deal value exceeds $3 billion in combined milestone potential, though realized revenue to date is limited to upfront payments and option exercise fees.

The revenue profile matters for how you value this company. Dyno does not sell drugs. It sells delivery. That means its revenue is capped by the number of licensing deals it can sign and the number of capsids it can develop per partner. But it also means its downside is limited. No clinical trial risk. No manufacturing scale-up for its own products. No pricing and reimbursement battles. The risk sits with the partners.

The Road to Human Data

The gap between NHP data and human outcomes has been the graveyard of many promising capsids. AAV capsids that work in cynomolgus monkeys do not always work in humans. The immune system, the receptor landscape, and the tissue architecture differ across species. Dyno's strategy of engaging conserved receptors and testing across multiple NHP lineages is designed to narrow that gap, but it has not been closed yet.

No AI-designed AAV capsid has yet delivered a gene therapy to a human patient in a clinical trial. The Astellas-licensed capsid will need to go through IND-enabling studies, Phase 1 safety trials, and eventually proof-of-concept efficacy studies before the thesis is validated in humans. That timeline is measured in years, not months.

What happens when capsids can be designed for any tissue?

Development scenarios