Three Seattle research institutions are launching AI BioDesign, a roughly $95 million initiative to use artificial intelligence to design proteins and genes that do not exist in nature. The Allen Institute, the University of Washington and Fred Hutch Cancer Center are splitting the funding: $46.1 million to the Allen Institute, $43.8 million to UW, and $4.7 million to Fred Hutch.
The money comes from the Fund for Science and Technology, created through the estate of Microsoft co-founder Paul Allen.
Who is running it
The lead scientific directors are David Baker — director of UW Medicine’s Institute for Protein Design, an HHMI investigator, and co-recipient of the 2024 Nobel Prize in Chemistry for computational protein design — and Jay Shendure, a genome scientist who directs UW Medicine’s Brotman Baty Institute and the Seattle Hub for Synthetic Biology.
Sanjay Srivatsan at Fred Hutch and Sud Pinglay serve as principal investigators; Rui Costa is president and CEO of the Allen Institute, and Marc Malandro is chief programs officer at the Fund for Science and Technology.
The contrarian bet on model architecture
The most interesting decision here is technical, and it runs against the industry’s prevailing direction.
Most AI development is pursuing ever-larger foundation models — single systems trained on vast general corpora, expected to generalise across tasks. AI BioDesign is deliberately building multiple modular, task-specific models, each aimed at a tractable biological problem.
The reasoning is about data rather than compute. Language models work because the internet supplied enormous quantities of text essentially free. Biology has no equivalent: the data that matters must be generated experimentally, at cost, one assay at a time. Given that, generating a large purpose-built dataset for a well-defined design problem and training a focused model on it is a better use of resources than assembling heterogeneous data into a general model.
It also produces something checkable. A model designed to solve a specific problem can be evaluated against whether the designed protein actually works in a laboratory.
Design, build, measure, learn
The initiative is organised around design-build-measure-learn cycles paired with multiplex experiments that test many designs simultaneously.
That coupling is the point Baker emphasises: “For the first time, the speed of AI is beginning to match the experimental power of synthetic biology. That changes the question from ‘what has nature already made?’ to ‘what else is possible?’”
The bottleneck in computational design has rarely been generating candidates — models can propose more sequences than anyone can test. It has been measurement. High-throughput assays that evaluate thousands of designs at once are what turn each round of prediction into training data for the next, and the accelerator is built around closing that loop rather than around the modelling alone.
Shendure frames the underlying premise: “Biology is code that builds: DNA carries digital instructions, and cells turn those instructions into physical systems with extraordinary precision.” Srivatsan puts it more directly: “Biology is ultimately a design challenge.”
What they intend to make
The stated applications range from the plausible to the deliberately speculative: cancer drugs and treatments for neurodegeneration, cells engineered to remove cancer from the body, enzymes that degrade plastics, and biological computers consuming far less power than silicon.
The first two sit closest to existing capability — designed binders and engineered cell therapies are active fields with clinical precedent. Plastic-degrading enzymes exist in nature and have been engineered before, with scale rather than discovery as the obstacle. Biological computing is a genuine moonshot.
The open-science commitment
All models, datasets, assays, reagents and benchmarks will be shared openly with the research community.
The benchmarks may matter most, and are the least glamorous item on the list. Protein design has struggled with evaluation: papers report success against inconsistent criteria, making genuine progress hard to distinguish from favourable framing. Shared benchmarks and reagents let other laboratories reproduce and compare results — the unexciting infrastructure that determines whether a field compounds its knowledge or merely accumulates publications.
Openness also reflects the funder. Philanthropic money can support work with no path to a product, which is precisely where shared infrastructure tends to fall.
Why Seattle, and why now
The location is not incidental. UW’s Institute for Protein Design has spent two decades building the field, and has spun out a series of companies applying computational design to vaccines, cancer therapeutics and biomaterials. The Allen Institute contributes large-scale data generation as an organising discipline — it was built around producing atlases and releasing them rather than around individual laboratory projects, which is unusual and well suited to this problem.
The timing reflects a broader surge of capital into AI-driven biology, much of it commercial and closed. Well-funded private efforts are pursuing protein and biological sequence models with resources that dwarf most academic budgets, and their models and data are generally proprietary.
That contrast sharpens what an open, philanthropically funded accelerator is for. It is not competing on scale of compute. It is producing the experimental datasets, benchmarks and reagents that commercial labs have little incentive to release — the shared substrate the rest of the field builds on.
The honest caveat
Designing a protein that folds as intended is now routine. Designing one that performs a useful function inside a living system, safely and durably, remains hard, and computational design has produced far more successful molecules in a test tube than approved medicines. Five years and $95 million buys serious data generation and model development — it does not buy a drug. The output to judge this on will be the datasets, benchmarks and models, and how much other laboratories manage to do with them. R&D news, not investment advice.