Dyno Launches Open-Source Agentic Protein Design Suite at GTC 2026
During NVIDIA GTC 2026, Dyno Therapeutics introduced Psi-Phi, a suite of agentic AI models, filtering tools, and APIs designed to link computational protein design more directly with experimental results. The platform combines generative models that create protein candidates with evaluation layers trained on laboratory data, and is intended to make protein binder design more consistent and scalable across research and drug development workflows.
The stated objective of Psi-Phi is to address a persistent gap in molecular design workflows: the disconnect between computational optimization and experimental validation. The system was developed in collaboration with NVIDIA, with training and inference accelerated using DGX Cloud and newer components such as the Transformer Engine.
Composition
Psi-1 is an open-weight AI model for protein binder design, built to handle complex multimeric targets. The model prioritizes structural diversity over narrow computational scoring, which Dyno says improves the odds that generated designs will validate experimentally. It was trained on NVIDIA DGX Cloud using Hopper GPUs.
Alongside the model, Dyno released a synthetic domain interface (SDI) dataset that was used during training. The company says this dataset was included to increase diversity in binder design, since many existing models are trained on a relatively narrow set of known protein structures.
The platform also includes Dyno Phi, a set of predictive filters trained and continuously updated with experimental data. These filters are used to estimate whether a generated protein sequence is likely to hold up in wet-lab testing. This is meant to address a common problem in protein design, where molecules that look strong by computational measures do not always work in real experiments.
Together, Psi-1 and Dyno Phi form a workflow where design and evaluation are tightly coupled, allowing users to iteratively generate, assess, and refine protein candidates under different constraints such as confidence in success or structural diversity. This setup introduces elements of an agentic system, where generation, feedback, and selection are combined into a continuous process.
According to the company, Dyno Phi can be used with Dyno’s own model or paired with other generative protein design models to help prioritize candidates with a better chance of experimental success.
Dyno also provides API access, so users can generate and assess protein binders without setting up their own computational infrastructure. Through these APIs, the platform can connect to outside tools such as OpenFold3 through NVIDIA BioNeMo services, supporting structure-guided design and probabilistic estimates of binding performance.
Dyno also released Psi-Phi Claude Code Skills, a plug-in that lets users access Dyno’s protein design tools directly inside Claude Code. They run the same GPU-powered models and filtering systems in real time, so scientists can design and refine proteins through a chat interface without setting up their own infrastructure.
About Dyno Therapeutics
Founded in 2018 and based in Watertown, MA, Dyno Therapeutics is a biotech company working at the intersection of artificial intelligence and genetic medicine. Its core focus is on improving how genetic therapies are delivered into the body via engineered AAV capsids, particularly to tissues such as the eye, muscle, and central nervous system. Dyno’s portfolio includes multiple vectors tailored for different targets, with reported gains over standard capsids like AAV9 and AAV2 in both delivery efficiency and tissue selectivity.
Dyno’s approach combines large-scale in vivo experimentation with machine learning, using techniques such as DNA multiplexing to generate and analyze biological data at scale. These datasets are used to model how genetic sequences behave in real biological systems, helping guide the design of more effective delivery vectors and therapies.
In 2024, Dyno Therapeutics announced its first AI collaboration with NVIDIA to scale its lab-in-the-loop sequence design workflows, followed by a second Roche collaboration to develop next-generation neurological AAV vectors that included $50 million upfront and potential milestones above $1 billion. Under that deal, one of Dyno’s AAV vectors was recently licensed by Roche.
For a broader view, a separate analysis explores how emerging therapeutic formats such as gene therapies, RNA drugs, and cell therapies are shaping recent deals, approvals, and pipeline value, with delivery technologies like engineered capsids playing a central role in how these treatments reach target tissues.
Topic: AI in Bio
