Researchers at Fred Hutch Cancer Center are testing whether a collaborative AI research platform can accelerate the pace of cancer research leading to faster diagnoses and more precise, targeted therapies (especially for rare types of cancers) while safeguarding patient privacy.
A year ago, Fred Hutch and three other comprehensive cancer centers — Dana-Farber Cancer Institute, Memorial Sloan Kettering Cancer Center, and Johns Hopkins University — launched the Cancer AI Alliance to help researchers find hidden patterns and connections in clinical data across these four participating institutions while still maintaining each center’s patient privacy and data security standards.
The collaborative research platform provides researchers with the infrastructure to build and test AI-driven computer models that aim to accurately predict how a patient’s cancer is likely to progress, what treatments are likely to work under which conditions, and how the cancer might adapt to evade those treatments.
The models — trained on comprehensive and diverse de-identified clinical data at CAIA’s member institutions — will help researchers make better sense of the complex molecular interactions underlying tumor biology, disease progression, response to therapy and resistance to treatment, especially for rare cancers.
Each center is now putting the platform to the test with eight pilot projects aimed at understanding how AI and machine learning (ML) could be used to better understand cancer and ultimately improve clinical practice.
Fred Hutch researchers are leading two projects:
When cancer becomes metastatic and spreads to other parts of the body, it often spreads to the bones, which leads to fractures, painful spinal cord compressions, and other skeletal problems so severe they require treatment, including hospitalization.
“This is a very common problem, up to half of all metastatic patients will develop bone metastasis at some point,” said Clemens Grassberger, PhD, an associate professor in Fred Hutch’s Radiation Oncology Division, who focuses on mathematical modeling of cancer and its response to treatment. “All of these bone metastases are at risk to cause skeletal problems that lead to a lower quality of life, more costs, and generally bad outcomes for the patient.”
Clinical studies led by Erin Gillespie, MD MPH, show promising results for knocking out high-risk bone metastases, or mets, with radiation before they can cause trouble, but there are so many patients, often with multiple bone mets each, that it’s not feasible to treat them all.
“There are too many patients and too many bone mets,” Grassberger said. “It would mean treating half of all advanced cancer patients.”
Clinicians need a way to identify which patients have the highest risk and would benefit the most from early radiation.
“We need to triage,” Grassberger said. “It's a classic prediction problem.”
He is building a computer model on the CAIA platform.
“We will answer these questions on the Fred Hutch data where we can play around with these models and then when we’ve homed in on one model structure, we’ll send this to the other centers,” Grassberger said.
He’ll send the model — but not the data that trained it — to the other centers, which can then each train the model on their own de-identified clinical data. Individual patient data remains securely behind each cancer center's firewalls. The results of how well the model did at each center will then be merged in a central location to create an updated model, which is then sent back to each cancer center for further refinement.
Each center participating in the study will train the model at the edge node with roughly 70% of its de-identified data, following which point the model weights and measures are sent back to improve the model. Once the model can make accurate predictions based on data it has already seen, the updated model will be sent back to the edge node to allow researchers to test it on the remaining 30% that the model has never seen, a process called model validation.
Ultimately, they want to train the model to accurately triage new patients who come through the door based on all it has learned from previous patients at CAIA’s participating institutions.
This approach, called federated learning, fosters collaboration across institutional lines without compromising patient privacy.
Federated learning represents new technology, but it’s an old idea in medical research: Doctors write journal articles based on what they’ve learned from their own patients, but they only publish the results, not the specific details that could identify individual patients.
“In principle, the federated learning idea is what we've always done in medicine,” Grassberger said. “We all know that this is possible. What was missing was the infrastructure.”
He said he remains cautious because he’s seen many approaches fail in the last 15 years, but he is encouraged by the platform, which can be used to build many models aimed at specific questions about cancer pursuant to IRB-approved protocols.
“If the infrastructure works, the model-building, training and validation is relatively simple,” Grassberger said.
He hopes to have a model up and running in the next few months.
“Our timeline is very ambitious,” Grassberger said.
The other project Fred Hutch researchers are road-testing with CAIA aims to address a key shortcoming of clinical trials: The patients who typically qualify for them are different in significant ways from the overall population of people with that disease.
“A lot of our evidence that we get in oncology is from clinical trials, but clinical trials are also limited,” said Simone Dekker, MD, PhD, who is in the Hematology-Oncology Fellowship Program at Fred Hutch. “It's often relatively healthy patients with not many comorbidities who live close to a cancer center and are not very old.”
The data available across CAIA's member institutions are more representative of the real-world population of cancer patients in academic centers.
“The nice thing about this study is that all patients that ever walked into the hospital or the clinic are included,” Dekker said. “And then with real-world data, you can answer questions about patients who are often excluded from clinical trials, and you can find answers that you otherwise couldn't.”
But there’s a catch.
“If I, as a physician-scientist, want to ask a research question, then I need a strong background in statistics, such as what statistical test is appropriate for my question and the data? And I also need to know how to perform these tasks, for example how to write computer code, generate informative figures, and understand how to interpret the data and produce the results. That’s a multi-step process that requires a lot of time and expertise that very often physicians don’t have,” Dekker said. “Even for biostatisticians, it still takes a lot of time, expertise and multidisciplinary collaborations.”
She and Steve Salerno, PhD, a postdoctoral researcher in biostatistics at Fred Hutch, are collaborating with Ai2, a Seattle-based nonprofit research institute founded by the late Paul Allen in 2014 to simplify and accelerate that process.
They have adapted a tool developed by Ai2 called Asta DataVoyager that uses AI/Large Language Models to quickly translate plain-language research questions into the behind-the-scenes computer code and statistical machinery needed to analyze big datasets.
DataVoyager then translates its analysis back into clearly cited, explainable answers.
“It understands what I'm asking,” Dekker said. “It knows what statistical tests to run and then it can write the code, do the analysis and give you the results with interpretation and caveats. It's done in minutes, which is very exciting. It can really help accelerate research and discovery.”
Dekker and Salerno will use DataVoyager to analyze de-identified data from about 30,000 patients with non-small cell lung cancer.
Salerno works with Vice President and Chief Data Officer Jeff Leek, PhD, who holds the J. Orin Edson Foundation Endowed Chair. Salerno said DataVoyager basically builds a research team of AI agents to tackle each research question.
The first AI agent interacts with the human researcher, translating simple language into an action plan.
“It will take that plan and give it to another AI that will write you code, and then it takes that code and gives it to another AI that will execute that code,” Salerno said. “And then another AI will take the results from that code and interpret it. And you have this little research team that's kind of working in a circle to be able to produce results and everything is documented. There's a notebook that this will produce where you see every step it took, every decision it made, all the code that it wrote, all the variables that it used, how it did this. And so, as a statistician, it's sort of the best research assistant you can have because you have everything that happened all documented right in front of you in minutes.”
But Salerno and Dekker aren’t going to take the AI’s word for it.
“The first thing of course with all new tools, new technology is you need to validate it,” Dekker said.
“We created questions in a range of different diseases, but for this project specifically on non-small cell lung cancer, we created questions ranging from very easy questions to more complex questions,” she said.
An easy question, for example, might be to ask the average age of the patients in the dataset.
“Steve and I are going to do the analysis ourselves and then have DataVoyager do the same analysis and we're going to see how well it performs, where we can improve it, why it is not working well, and then improve the model,” Dekker said.
In October, the CAIA team presented a video showing DataVoyager in action at the annual Madrona IA Summit in Seattle, the same conference where CAIA was officially launched a year ago.
The video shows a prompt entered in the dialog box asking DataVoyager if there are differences in overall survival from start of treatment between two drugs that each target a different kind of genetic mutation in non-small cell lung cancer.
"Using traditional methods would have taken several months of work," Dekker said.
But within seconds, DataVoyager works through the problem and scrolls a detailed answer to the question, including an X-Y graph showing a Kaplan-Meier Survival Curve, a common statistical tool in medical research, comparing the two drugs over almost 140 months.
A human then enters this prompt: “That KM plot looks good, but can you adjust the X-axis so that we only show the first 80 months of the data?”
And voilà! The plot is adjusted for 80 months.
“We were able to produce that plot for the first time using all four cancer centers’ data for the summit, which I thought was really inspiring,” Salerno said. “It was a cool experience being in that room and watching the first results from four cancer centers coming out of this tool.”
CAIA receives financial and technical support from technology industry leaders Amazon Web Services (AWS), Deloitte, Ai2 (Allen Institute for AI), Google, Microsoft, NVIDIA, and Slalom.
John Higgins, a staff writer at Fred Hutch Cancer Center, was an education reporter at The Seattle Times and the Akron Beacon Journal. He was a Knight Science Journalism Fellow at MIT, where he studied the emerging science of teaching. Reach him at jhiggin2@fredhutch.org or @jhigginswriter.bsky.social.
Are you interested in reprinting or republishing this story? Be our guest! We want to help connect people with the information they need. We just ask that you link back to the original article, preserve the author’s byline and refrain from making edits that alter the original context. Questions? Email us at communications@fredhutch.org
Every dollar counts. Please support lifesaving research today.
For the Media