Weill Cornell gets $1 million to build AI tools for catching treatment-resistant prostate cancer early
Prostate cancer is, in most cases, a disease that responds to treatment. Hormone therapies, surgery, and radiation together cure or control the majority of cases diagnosed each year. But a subset of prostate cancers do not follow that pattern. They are resistant to the standard arsenal, evolve rapidly to evade newer therapies, and are often not identified as dangerous until they have already progressed.
The clinical problem is not just that these cancers are hard to treat - it is that they are hard to identify early, when more options remain available. They can look, under the microscope, similar to cancers that will respond to conventional therapy. By the time their resistance becomes apparent through treatment failure, time and options have been lost.
The Prostate Cancer Foundation has awarded Dr. Ekta Khurana at Weill Cornell Medicine a $1 million Challenge Award to address this identification problem. The two-year grant funds a collaboration with Memorial Sloan Kettering Cancer Center to build artificial intelligence tools that can detect treatment-resistant subtypes from data that exists at the time of initial diagnosis.
Reading the cancer's future from its present
The core technical goal of the project is prediction: given a biopsy and the associated molecular data available at diagnosis, can an AI system identify which tumors are on a trajectory toward treatment resistance, before resistance has been demonstrated clinically?
The approach uses two data streams. The first is pathology slides - the microscopic images of tissue that pathologists examine to grade and characterize tumors. These images contain spatial and structural information about the tumor architecture that has historically been interpreted by human experts applying established grading systems. AI systems can, in principle, identify patterns in these images that human graders do not consistently capture.
The second data stream is gene activity patterns - transcriptomic data that measures which genes are being expressed at what levels in the tumor. Different prostate cancer subtypes have distinct molecular signatures that can predict clinical behavior. Some of these signatures are already associated with resistance to specific therapies. Combining pathological and molecular information in a single model could produce predictions more accurate than either source alone.
An interdisciplinary team
The project brings together computational biomedicine researchers, pathologists, and physician-scientists. This combination is not incidental. Building a tool that clinicians will actually use requires that it produce outputs that integrate into clinical decision-making - that it speaks the language of diagnosis rather than just the language of machine learning. Pathologists need to understand what the system is responding to in the slide; oncologists need predictions formatted in terms of treatment decisions.
The Memorial Sloan Kettering collaboration adds clinical depth. MSK's patient volume and data infrastructure make it one of the premier environments for training and validating cancer AI models in the United States. Access to well-annotated clinical outcomes - which patients responded to which treatments, and which did not - is essential for training a system designed to predict exactly those outcomes.
The end goal: matching patients to trials
The immediate application the team is targeting is patient selection for clinical trials of experimental therapies. Early-phase trials for novel prostate cancer treatments need patients who are likely to have the specific tumor subtype the therapy is designed to address. Selecting the wrong patients wastes the trial's limited slots, dilutes the signal, and can cause a genuinely effective therapy to fail a trial because it was tested in an unresponsive population.
An AI tool that accurately identifies treatment-resistant subtypes early would allow clinicians to direct patients toward trials of therapies designed for their specific cancer biology, rather than routing them through standard treatments that are unlikely to work. This is a narrower and more immediately achievable goal than curing treatment-resistant prostate cancer - but it is a meaningful step toward better outcomes for the patients most likely to be failed by current standard care.
Longer term, if the predictive model performs well enough in clinical validation, it could be incorporated into routine diagnostic workflows. That would require regulatory review, prospective clinical trial data, and considerable implementation work - but the Challenge Award is explicitly designed to fund the foundational research that makes those later steps possible.
What remains uncertain
The project is in its early stages, and the proof of concept for using combined pathological and molecular AI in this specific clinical application has not yet been demonstrated. The team needs to show that their model can distinguish treatment-resistant subtypes with sufficient accuracy and specificity to be clinically useful - a higher bar than demonstrating statistical significance in a research dataset.
The two-year funding window is relatively short for a project of this scope. Challenge Awards are designed to accelerate progress on defined problems, not to fund open-ended research programs. The timeline creates pressure to produce a validated, interpretable model quickly - which is either a useful constraint or an optimistic one, depending on how the data cooperate.