Unleashing the Power of Prometheus to Improve Cancer Treatments
Bolstered by a generous donation from alumnus Dominic Orr, computational biologist Matt Thomson has developed an audacious framework to leverage AI in the design of immunotherapies.
Dominic Orr (MS '76, PhD '82, DAA '10) has always followed his curiosity. As a student at Caltech, he moved from studying theoretical physics to neurobiology. Upon graduation, early work experience at Hewlett-Packard launched a career in the computer networking industry where he became a renowned expert in the field, guiding several leading technology companies—including Aruba Networks where he served as president and CEO—in bringing innovative products and technologies to market.
"I dove headfirst into the high-tech startup world, but after several decades, my interest in basic research rekindled," says Orr, who is now retired. "Through involvement in Caltech chairs' councils, I started to fund smaller, interdisciplinary projects that had the potential to be high-risk, high-reward research."
Recently, Orr has been spending time in the lab of Matt Thomson, professor of computational biology, learning more about work that uses machine learning to design new cancer treatments. He met Thomson during a chair's council visit soon after a close relative was diagnosed with a rare type of cancer.
"I had started studying the disease and investing in related research at the Dana-Farber Cancer Institute that was using single-cell multiomics technology plus machine learning to assess treatment options," Orr explains. "When I met Matt, I found the connection between his computational techniques and what I was doing on the East Coast very synergistic."
Machine Learning Meets Immunotherapy
The problem that many cancer patients face is most therapies are not effective across all recipients. Thomson, who is a Heritage Medical Research Institute Investigator, has long been interested in cancer and wanted a way to predict how a patient would respond to a given treatment. In particular, he was interested in assessing immunotherapies, which harness a person's own immune system to fight cancer and have shown great promise in certain subsets of patients and tumor types.
"When we can get the immune system to identify and attack a tumor, you can get these unbelievable outcomes where even late-stage cancers can be put into remission. But they don't work for everyone," Thomson says. "It's a good situation for predictive modeling that can help us develop new versions of immunotherapies that work for everybody, or immunotherapies that work for different subsets of people, so we have enough of them to cover every case."
Thomson began pursuing this goal by designing Morpheus, a machine learning framework that uses molecular imaging data collected from primary human tumors and automatically designs interventions. He and his team have used the technology to predict immunotherapy targets in melanoma, breast cancer, and colorectal cancer tumors that were then validated experimentally in the lab. Now, with the help of a generous gift from Orr, Thomson is developing Prometheus, a foundational model that aims to extend Morpheus's abilities across all tumor types and patients.
Revolutionizing Cancer Immunotherapy
"I feel like what Matt's doing is going to revolutionize the field of cancer immunotherapy," Orr says. "This is the time to apply resources to accelerate things, to make discoveries in a timespan that was unthinkable even four or five years ago."
Prometheus is a scaled-up version of Morpheus that functions like a "virtual tissue," enabling large scale simulation of how a tumor will respond to the knockdown or overexpression of thousands of different genes alone or in combination. "By doing these simulations, we've got new targets that are predicted to derive durable responses in hard-to-treat tumors like pancreatic cancer, which is the dream I wanted to achieve," says Thomson. "I lost my dad to pancreatic cancer, and one of the things I wanted to do was try to make a dent in these really bad diseases. We've got new targets coming out of this model and I can't believe it."
Like any machine learning model, Thomson says Prometheus can be updated as therapeutic strategies are tested in mouse models or human patients, potentially leading to a world in which AI can design therapies for individual patients based on their tumor-specific needs.
"We're starting to see that we might be able to crank this model and really get wheels turning more quickly toward a future where we have thousands of new cancer drugs in a relatively short period of time," says Thomson, noting the name Prometheus was inspired by the unbelievable power of AI, which he is now bringing to therapeutics.
The Visionary and the Researcher
Thomson emphasizes that without Orr's dedication to the project, he and his team simply would not have been able to do this work, as the cost of training AI models has become extremely expensive. It is also challenging for researchers to obtain grants before they can show evidence of success in their work.
"Dom's a visionary. No government agency would have funded this project, but now with his support, we can show results and get follow-on funding," Thomson says.
According to Orr, his willingness to take risks on innovation comes from his time as a graduate student.
"I owe Caltech a lot for training me to be not afraid of fear, and be able to face the unknown to push things forward," he says. "Supporting creative and collaborative research at the Institute is my very enjoyable way of returning something to where I came from, where I learned to be who I am."