Three of Seattle’s top scientific institutions are launching a nearly $95 million research initiative that will generate data and train AI models to design proteins and genes that don’t exist in nature — sharing the results freely to help others develop new medicines and materials.
The initiative, called AI BioDesign, brings together the Allen Institute, the University of Washington and Fred Hutch Cancer Center, with funding from the Fund for Science and Technology (FFST), created by the estate of Microsoft co-founder Paul Allen.
AI BioDesign is led by David Baker, the UW biochemist who won the 2024 Nobel Prize in Chemistry for using computers to design new proteins, and Jay Shendure, a leading genome scientist at the UW and the Allen Institute.
The plan is to “hijack a lot of the machinery that evolution provided us” — the cellular assembly line that turns DNA into proteins — to design and measure millions of novel biological molecules, Shendure said in an interview in advance of the announcement.
That will help AI models learn the rules of biological design from a huge set of examples, instead of inferring them from the relatively limited number that nature has produced.
The field, Shendure said, is “putting too much emphasis on taking the cranks that we have and just running with them, as opposed to building the right cranks.”
The goal is to make designing biology more like ordering a part: a molecule that latches onto a cancer cell, for example, or a genetic switch that fires only inside brain cells and nowhere else.
Potential outcomes could include everything from new therapies for disease, to proteins that dissolve plastic in the environment, to cells that travel through the body in a programmed way, said Sanjay Srivatsan, a Fred Hutch assistant professor who leads the cancer center’s work on the initiative, in a video released with the announcement.
“For the first time, the speed of AI is beginning to match the experimental power of synthetic biology,” Baker said in a statement. “That changes the question from ‘what has nature already made?’ to ‘what else is possible, and how can we test it?'”
Where the money goes
The Fund for Science and Technology is providing $94.6 million for AI BioDesign over five years. The foundation launched publicly last year with a mandate to direct a large share of Allen’s fortune into bioscience, environmental and AI research.
The funding from FFST is allocated as $46.1 million to the Allen Institute, $43.8 million to the UW and $4.7 million to Fred Hutch, according to an Allen Institute spokesperson.
The initiative had 62 people as of mid-August, including some new hires and others redirected from existing projects at the three institutions. The UW accounts for 41 of them, the Allen Institute 13, and Fred Hutch eight. AI BioDesign is expected to continue growing over time.
“AI BioDesign is exactly the kind of ambitious, collaborative science FFST was created to support,” said Marc Malandro, the foundation’s chief programs officer and co-lead, in a statement. He joined FFST in May after nearly a decade at the Chan Zuckerberg Initiative, most recently as chief operating officer of CZI and the Chan Zuckerberg Biohub Network.
Malandro and Chief Financial and Operations Officer Liz Carey have been leading FFST on an interim basis since founding CEO Lynda Stuart stepped down in May.
Inside the lab
On a recent tour of the AI BioDesign lab, research associate Jack Boylan pulled up results from a run he’d done on their new DNA sequencer that morning — on free kits donated by a neighboring biotech company, a year past their expiration date.
“We decided, let’s give it a roll,” he said. It worked fine.
The sequencer is what makes the whole approach possible. It reads all of the millions of DNA sequences in a single tube at once and reports which ones performed. One recent experiment ran 6 million distinct sequences through it at once.
“The scale comes not from robotics, but from parallelizing inside the test tube,” said Jesse Gray, executive director of strategy and platform for AI BioDesign and the Seattle Hub for Synthetic Biology, and a former Harvard Medical School geneticist.
The lab, at Dexter Yard in Seattle’s South Lake Union neighborhood, a short walk from the Allen Institute’s headquarters, is organized into teams of five or six people, each working on a different design problem.
A separate four-person team of machine-learning specialists takes the incoming results and works with the bench teams to decide which experiments come next — the ones that will teach the models the most. Each round is judged on how much the models improved.
The Allen Institute calls projects like this “accelerators,” a term Rui Costa, the institute’s president and CEO, traced back to Paul Allen himself. The word came up in early planning sessions, Costa said. Allen wanted to “exponentially accelerate the field.”
Other accelerators at Dexter Yard include the Seattle Hub for Synthetic Biology, the Allen Institute’s collaboration with the Chan Zuckerberg Initiative and the UW, which Shendure also leads; and Cell Science, which works on engineering cells to assemble themselves into tissues.
The Allen Institute for AI (Ai2), the separate Seattle research organization also founded by Paul Allen, is involved informally rather than as a funded partner, Costa said.
Its robotics team has been talking with AI BioDesign about scaling up the protein work, and the two expect to collaborate on models and on tools that generate research hypotheses.
Why give it away
The decision to focus on open science also came from Allen, Costa said in an interview this week. “He was so visionary in the early 2000s: radically open science to exponentially impact and change fields, not to compete.”
That raises a question the initiative will face as soon as it produces anything valuable: what happens if a company builds a lucrative drug on data given away free? In traditional science, Costa said, being beaten to a discovery counts as a loss. Here it’s the goal.
“We would be so lucky if many companies would be taking this data and changing the world for good,” he said.
At the same time, Costa left open the possibility of the three principal institutions spinning out their own startups, nonprofits, or other initiatives from the work done by AI BioDesign.
Betting against the field
AI BioDesign’s approach runs against much of the current thinking in the field. Costa said most efforts to apply AI to biology are chasing a single general model that could answer questions about how any cell works. AI BioDesign is betting on the opposite: narrow models built for specific design problems, trained on data generated for that purpose.
“This project is a clear bet on a different way of doing things,” Costa said.
The people running the initiative are careful not to oversell. Gray said it remains an open question as to whether their approach beats the alternatives. “The jury’s still out,” he said.
Shendure put it plainly: “It’s never as easy as you think it’s going to be,” he said.
Costa said AI BioDesign needs to show real progress within 18 to 24 months — ideally even sooner — and expand to researchers around the world within five years.