JungHo Kong and his team at UC San Diego have built an artificial intelligence model that does something oncologists have long struggled to do: translate a tumor's sprawling genetic blueprint into a clear prediction of which treatments will actually work. The model, called MutationProjector, was trained on genomic data from more than 30,000 tumors across 10 solid cancer types, and it marks a meaningful step toward making genetic testing deliver on its promise in cancer care.
Genetic sequencing has become routine in cancer diagnosis, yet doctors still face a stubborn problem: patients' tumors typically carry hundreds of mutations, most of them rare and individually poorly understood. Current treatment decisions rely on only a handful of validated genetic biomarkers—so much so that only about 8% of cancer cases are successfully matched to an FDA-approved therapy based on genetics alone. The rest remain a puzzle, leaving many patients with limited guidance on which drugs might work best for them.
MutationProjector takes a radically different approach. Rather than hunting for a few known mutations, it analyzes the broader combination of genetic alterations present in a tumor, then generates a compact representation of the tumor's biological state. This allows the model to identify which molecular pathways are disrupted and, by extension, which treatments are most likely to be effective. Trey Ideker, a professor of medicine at UC San Diego School of Medicine and director of the Big Data Institute at Oxford University, explains the core insight: "Genetic sequencing is already routine in cancer care, but we still struggle to fully interpret the many mutations found in a patient's tumor. Our goal with MutationProjector was to build a general-purpose model that can learn from tens of thousands of tumor genomes and turn those mutation patterns into more precise predictions about treatment response."
The validation came through real-world testing. Across independent patient cohorts with bladder cancer, lung cancer, and melanoma, MutationProjector matched or exceeded existing methods for predicting response to immunotherapy and chemotherapy. Crucially, the model also uncovered both known and unexpected biomarkers associated with treatment outcomes—discoveries that could reshape how genetic testing is currently done.
What sets MutationProjector apart is not just its predictive power but its transparency. The researchers built it to explain why it makes each prediction, a feature essential in precision oncology where clinicians must understand how a patient's tumor genetics connect to treatment decisions. Kong, the first author of the study, notes that "by pretraining on a large collection of tumors and integrating molecular network knowledge, MutationProjector can detect patterns that would be easy to miss with conventional biomarker approaches. That gives us a way to move from long lists of mutations toward a more functional understanding of the tumor."
The team published their findings in Cancer Discovery and is already looking ahead. They plan to expand MutationProjector to additional cancer types and integrate new data sources—international cancer genome datasets, imaging, transcriptomics, and electronic health records. Ideker sees the broader potential: tumor genome foundation models may help extend the clinical value of sequencing far beyond the small set of well-known genes that currently dominate treatment decisions, offering a more comprehensive strategy for matching patients to the therapies most likely to help them.