Scientists at University of California San Diego have developed a new approach to destroying cancer stem cells – hard-to-find cells that help cancers spread, come back after treatment and resist therapy. The new approach, which the researchers tested in colon cancer, leveraged artificial intelligence (AI) to identify treatments that can reprogram cancer stem cells, ultimately triggering them to self-destruct. Because it only targets cancer cells without affecting surrounding tissues, the approach could be a safer and more precise alternative to current therapeutic approaches. The results are published in Cell Reports Medicine.
"Cancer stem cells are like shapeshifters," said Pradipta Ghosh, M.D., senior author of the study and professor of medicine and cellular & molecular medicine at UC San Diego School of Medicine. "They play hide-and-seek inside tumors. Just when you think you've spotted them, they disappear or change their identity. It's like trying to hold on to a wet bar of soap in the shower."
To outsmart these elusive cells, the team built a machine learning tool, called CANDiT (Cancer Associated Nodes for Differentiation Targeting), that can identify new treatment targets for a specific tumor based on its unique genetics. The tool works by starting with a single key gene, one that healthy cells need to grow but that is missing in aggressive cancers. From there, the tool identifies a network of genes related to the initial gene, suggesting treatment targets that can leverage this biochemical network to revert the cells to a healthier state.
By starting with CDX2, a significant gene in colon cancer, the researchers used CANDiT to rapidly scan the entire human genome in more than 4,600 unique human tumors, reflecting the genetic diversity typical of large, multi-center clinical trials. Their approach identified an unexpected new treatment target: a protein called PRKAB1, which helps cells respond to stress. By using an existing drug that activates this protein, the researchers were able to restore function of the CDX2 gene in colon cancer stem cells.
After treatment, the cancer stem cells began to behave more like normal healthy cells, but this isn't all that happened.
"What surprised us most was that after we reprogrammed the cancer stem cells to behave like normal cells, they chose to self-destruct instead," said first author Saptarshi Sinha, Ph.D., interim director of the Center for Precision Computational Systems Network (PreCSN) , part of the Institute for Network Medicine (iNetMed) at UC San Diego School of Medicine.
"It was as if they couldn't live without their cancerous identity."
To demonstrate the clinical potential of this approach, the researchers were able to leverage UC San Diego's HUMANOID™ Center , also part of (iNetMed), to successfully test the drug in patient-derived organoids — tiny, lab-grown replicas of human tumors.
These organoids faithfully preserve the structure, behavior and biology of real cancers, allowing researchers to safely and effectively test treatments in human tissues. Organoid experiments can streamline the process of bringing treatments to clinical trials, as many therapies that succeed in animal models ultimately fail in humans.
"It's like doing clinical trials in a dish, which collapses timelines from years to months," said Ghosh, who is also director of the HUMANOID™ Center. "We used a complete suite of cell analysis platforms at the Agilent Center of Excellence to measure not just whether a drug works, but how precisely and safely it works, before it ever reaches a patient."
To explore the potential real-world impact of the treatment and identify who would benefit most from it, the researchers also developed a gene signature — a measurable pattern of gene activation — that can be used to predict how well someone might respond to this kind of therapy. Using advanced computer simulations of clinical trials, they tested this signature across 10 independent patient groups totaling more than 2,100 people, mirroring the diversity of large Phase 3 clinical trials. They found that using the drug to restore CDX2 in colon cancers could cut the risk of recurrence and death by up to 50%.
"This was heartwarming, but not surprising," said Sinha. "For decades, the Holy Grail of cancer has been its stem cells — resilient, elusive and beyond our ability to identify or track them. They are able to outsmart every form of treatment, even the most advanced immunotherapies. To be able to track and selectively kill them brings us closer to rewriting the rules of cancer treatment."
The researchers are now building on their momentum in collaboration with researchers across campus. This includes chemistry professor Jerry Yang, Ph.D., who has designed a more potent version of the compound with the goal of advancing it into clinical trials, and professor of surgery and UC San Diego Health surgical oncologist Michael Bouvet, M.D., who is leading efforts to deploy CANDiT across multiple tumor types, including pancreatic, esophageal, gastric, biliary, and others.
"It truly takes a village to get it right, and we're fortunate to have the kind of partnerships that allow us to stay nimble yet impactful," added Ghosh.
The team is also diving deeper into the question posed by their results: what made the cancer stem cells spontaneously die? Cracking that code could unlock an entirely new arsenal of therapies.
"This isn't just about colon cancer," said Ghosh. "CANDiT is an end-to-end human roadmap — we can apply it to any tumor, find the right targets, and finally take aim at the cells that have been the hardest to define, track or treat. By constantly anchoring small-scale organoid insights to Phase 3–sized human diversity in the clinic, we can build discoveries that are rigorous, reproducible and scalable, all without losing sight of the essentials of human disease. The potential of this approach to transform clinical medicine is not just immense — it's inevitable."
Link to full study: https://www.cell.com/cell-reports-medicine/fulltext/S2666-3791(25)00494-X .
Additional co-authors of the study include Joshua Alcantara, Kevin Perry, Vanessa Castillo, Annelies Ondersma, Satarupa Banerjee, Ella McLaren, Celia R. Espinoza, Eleadah Vidales, Courtney Tindle, Adel Adel and Siamak Amirfakhri at UC San Diego School of Medicine, Sahar Taheri at UC San Diego Jacobs School of Engineering and Joseph R. Sawires at UC San Diego School of Physical Sciences.
The study was supported, in part, by the National Institutes of Health (grants UG3TR002968, UH3 TR002968, R01-CA238042, R01-CA100768, R01-CA16091, R01-AI155696, R01- AI141630 and UG3TR003355) and UC San Diego's Institute for Network Medicine.
Declarations: The authors declare no conflicts of interest.