NEW YORK – The University of California, San Francisco (UCSF) Helen Diller Family Comprehensive Cancer Center and the Stanford Cancer Institute are joining forces to advance artificial intelligence and personalized medicine research in oncology.
The partnered institutions will use a $100 million grant from the Weill Family Foundation to launch Weill Cancer Hub West, which will house four cancer research projects that will begin in the next few months. Along with the $100 million grant, both institutions have committed to raising $50 million in support of the 10-year program, bringing the total funding for the Cancer Hub to $200 million.
In March, the Weill Family Foundation also gave $50 million to support the creation of a similar collaborative research hub, Weill Cancer Hub East, among Weill Cornell Medicine, Princeton University, Rockefeller University, and the Ludwig Institute for Cancer Research. The Weill Cancer Hub East's research will focus on advancing cancer immunotherapies.
One of the four funded efforts at the West Coast hub is Project Impact, in which researchers will build AI systems that combine clinical and biological data to facilitate personalized cancer treatments. The other programs, Project Vital, Project Promise, and Project Feast, are focused respectively on developing CRISPR-based cell therapies; using AI and cellular engineering to address treatment resistance in solid tumors; and exploring diet and drug interventions that reduce cancer development and progression.
The two co-principal investigators for Project Impact, Julian Hong, an associate professor in the department of radiation oncology at UCSF, and Sylvia Plevritis, professor of biomedical data science and radiology at Stanford, emphasized the importance of building AI systems that work across different healthcare institutions.
"What we're trying to do is harmonize our solution between two institutions, which is important," Plevritis said. "You can easily make AI systems that are not reproducible, and those type of systems make you question how accurate they really are. But when you start developing systems that work at two different institutions, it does increase your confidence that these are really robust systems."
Hong similarly highlighted the importance of breaking down data silos by creating technologies that work across institutions.
Hong and Plevritis will try to achieve this within Project Impact by bringing together electronic health record data from both Stanford and UCSF and building an AI-enabled model that clinicians can use to better predict the trajectory of their patients' disease. In the first year of the program, Plevritis said, the team will focus on combining the disparate EHR information and building a vector database that will harmonize clinical, biological, and genomic datasets from the two institutions, which will allow for the development of a large-language model. In total, the model will be built on data from hundreds of thousands of patients treated over the past decade, Hong said.
The Project Impact team hopes the resulting model will have research and clinical applications by improving understanding of how cancer progresses, which clinicians can then use to better treat their patients. Hong said the team will look for shared biological or genomic characteristics of cancer progression across all cancer types, along with the characteristics of individual tumor types.
"As we create this new database that is represented in this new AI language with vector embeddings, we will actually retrieve more information out of the data than we have been able to so far," she explained.
And the eventual goal, Hong said, is to build a model that can comb through large amounts of data, organize and characterize it, and finally represent it in a usable format at the point of care. "For a given patient, [it could be used] to come up with a personalized treatment plan," Hong said. "Being able to anticipate [disease] changes ahead of time and optimize how well we can take care of each individual person is important."
Plevritis added that the team will create a database of natural language prompts, which will allow clinicians to ask questions and the LLM to provide answers based on its database, much like ChatGPT. They will work together with clinicians and researchers to design prompts that will be most useful for patient care or research purposes, such as questions about treatment toxicity and how patients with certain biological or genomic characteristics fared on certain treatments. "Right now, we have our databases available for research clinicians, but they're not set up in a way to have natural language prompts," Plevritis said.
Within Project Impact, the database Hong and Plevritis' team is building will include all cancer types, but researchers will initially focus on gastrointestinal and genitourinary cancers, such as liver, colon, and prostate cancers, and consider patients' outcomes at different stages of disease. "We know cancer is a very dynamic disease and often physicians don't have a full understanding of the likely course of the disease for a given patient," Plevritis said. "With this new way of representing the data, we'll have a better ability to predict likely future events for a given patient. It's our intent to understand in a new way, across decades of data, what the likely trajectories are for cancer patients."
For patients with early-stage cancers, the resulting model could help inform whether they can be treated with surgery alone, if they're at low risk of recurrence, or for those with high-risk biomarkers, if they need additional treatment like an adjuvant chemo or immunotherapy. For those with advanced cancers, treatments can become more complicated due to resistance and metastases, which also presents an opportunity for predictive AI tools.
"For disease that has metastasized to other locations from its initial site, we're looking at a variety of treatments and maybe combinations of treatments to try to eradicate this more disseminated disease," she said. "But we need to take into account the toxicity and off-target effects. So, we need to balance the treatment with the effects of that treatment. The question is always: How aggressive should you go with the treatment? Is the patient likely to respond or is there a better shot on goal?"
Project Impact is part of a groundswell of recently launched AI-based cancer research, in which researchers are trying to develop machine-learning tools and algorithms that better predict who will get sick and how to best treat them. Another large collaborative research project that began in June, funded by the Advanced Research Projects Agency for Health (ARPA-H) of the US Department of Health and Human Services, aims to build a clinical AI tool that can guide oncologists' treatment decisions. ArteraAI, Pangea Biomed, Io9, and researchers at the Icahn School of Medicine at Mount Sinai and Memorial Sloan Kettering Cancer Center are also conducting AI-driven precision medicine research.
But just because these efforts are under way, it doesn't mean doctors will trust AI tools or adopt them. The Project Impact team also aims to understand how clinicians will use the AI model and incorporate it into their practice. For example, researchers will study the "human-computer interaction" and get feedback from clinicians who use the model, Hong said.
"In the AI world, one of the more important questions that you should be asking is how do folks interact with it? How does it affect our behaviors as clinicians?" Hong said. "Interacting with computers and applying AI to decision-making processes, we've already seen that it changes how we all behave. Being cognizant and understanding of all of those means we can better design computer interfaces."