UC San Diego AI Foundation Model Predicts Cancer Treatment Response

Only about 8% of cancer patients whose tumors are sequenced end up matched to an FDA-approved therapy on the basis of genetics. That number reflects not a failure of sequencing technology – a standard clinical panel already finds roughly 11 distinct genetic alterations in the average tumor – but a fundamental shortage of tools for interpreting what most of those mutations mean for treatment choice.

Researchers at the University of California San Diego published a paper Monday in Cancer Discovery describing an AI foundation model called MutationProjector that reads a tumor’s full mutation landscape and predicts treatment response with accuracy matching or exceeding existing methods across several independent patient cohorts. The second-order implication gets less attention than the performance numbers: if a system like this holds up at scale, it changes the purpose of a sequencing panel from a narrow flag-list into a comprehensive treatment-strategy input.

The 92% Gap in Genomic Treatment Matching

Genetic testing after a cancer diagnosis has become routine for good reason. The published study in Cancer Discovery notes that DNA sequencing panels flagging alterations in cancer-associated genes are relatively fast and low-cost, with a strong track record where validated biomarkers exist. The problem is that validated biomarkers currently cover a narrow slice of what a panel actually finds.

“Genetic sequencing is already routine in cancer care, but we still struggle to fully interpret the many mutations found in a patient’s tumor,” said Trey Ideker, PhD, professor of medicine at UC San Diego School of Medicine and director of the Big Data Institute at the University of Oxford, who led the research. Clinical treatment decisions based on genetics today still hinge mostly on a single gene, whether BRCA1, EGFR, or KRAS, while the other 10 alterations in the same report sit unused.

Two problems make those unused mutations hard to act on. First, many individual mutations are rare – appearing in only a tiny fraction of patients – which makes them nearly impossible to study through conventional clinical trials. Second, one mutation’s effect on drug response often depends on what other mutations are present in the same tumor. Standard biomarker testing checks each gene in isolation and misses interaction effects entirely.

How the Model Encodes a Tumor’s Full Mutation Landscape

From Mutation Lists to Biological State

A standard biomarker approach scans for a known alteration, confirms its presence, and maps it to a drug. The UC San Diego team’s system works differently. It takes the full set of genetic alterations identified in a tumor and compresses them into what the researchers describe as a “compact representation” of the tumor’s biological state – a summary of which molecular pathways are likely disrupted, rather than a catalogue of individual mutations.

That representation is what the system uses to make treatment predictions. Instead of asking “does this tumor carry KRAS G12C?” as a binary flag, it asks what the combined mutation profile implies about how the tumor’s cellular machinery is functioning. The shift from mutation list to pathway state is the core architectural choice behind the model’s design, and it is what allows the system to produce an explanation alongside each prediction rather than a raw probability score.

The model belongs to a class called foundation models (AI systems pretrained on large, broad datasets and then adapted for specific downstream tasks with relatively small labeled samples). Foundation models underpin large language models in general AI applications; the UC San Diego version runs on tumor genomes rather than text, but the conceptual logic is the same: expose the model to enough examples across enough variation that it can detect patterns no single institution could find in isolation. A 2026 Nature Reviews Cancer perspective on machine learning and genomics notes that as large clinicogenomic datasets mature, such tools have a real opportunity to extract more information from sequencing data and generate therapeutic hypotheses for patients who currently have none.

Why Rare Mutations Become Detectable at Scale

Training on more than 30,000 tumor profiles across 10 solid cancer types gives the model exposure to mutations that no single institution sees often enough to study in a conventional clinical setting. “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,” said JungHo Kong, PhD, the study’s first author and a postdoctoral researcher in the Department of Medicine at UC San Diego School of Medicine.

The model integrates biological network data – information about how genes and proteins interact – alongside raw mutation profiles. That integration connects mutations to known biological pathways, giving the predictions interpretable biological grounding rather than a pure statistical correlation drawn from the training data.

Bladder Cancer, Lung Cancer, Melanoma: Validation Across Independent Cohorts

The research team validated the AI tool across several independent patient cohorts – populations whose data were held out from training and tested separately, the standard check for whether a model’s patterns hold beyond the data it saw during training. Across cancers including bladder cancer, lung cancer, and melanoma, the tool matched or exceeded existing methods for predicting response to both immunotherapy and chemotherapy.

Peer-reviewed publication in Cancer Discovery, a journal of the American Association for Cancer Research, placed the results through external evaluation before they were made public. The key performance findings from the published paper:

• The tool matched or exceeded state-of-the-art prediction accuracy in every independent cohort evaluated.
• Immunotherapy response predictions held across multiple cancer types, not just one tumor category.
• Chemotherapy response predictions also held, suggesting the approach generalizes across different treatment mechanisms rather than optimizing for a single drug class.
• Multi-cohort validation separated the results from single-dataset overfitting, a common failure mode in earlier machine-learning oncology research.

Earlier attempts to apply machine learning to cancer treatment prediction have sometimes produced strong numbers on a training dataset and poor performance when tested elsewhere. Building the validation around independent cohorts was deliberate, and the consistency of results across cancer types and treatment classes is among the more credible aspects of the published work.

Two Biomarker Findings the Training Data Did Not Obviously Predict

Beyond the core accuracy results, the model identified associations that were not part of the researchers’ initial hypotheses – and at least one sits in tension with how current clinical sequencing guidance is constructed.

Mutations in the gene KMT2D were flagged as a predictor of sensitivity to immunotherapy. KMT2D encodes a histone methyltransferase involved in gene regulation. Prior research had associated KMT2 family mutations with immune checkpoint inhibitor response in pan-cancer analyses, but the finding had not been incorporated into standard clinical sequencing guidance. The model’s identification of KMT2D sensitivity, surfaced from patterns across the training population, points toward a candidate biomarker for prospective clinical study.

Our results suggest that tumor genome foundation models may help extend the clinical value of sequencing beyond a handful of well-known genes. This could support a more comprehensive and biologically grounded approach to precision oncology.

Trey Ideker made that comment in the UC San Diego release accompanying the paper’s publication on May 26.

The second finding is more operationally immediate. The joint presence of mutations in both SMARCA4 and STK11 emerged as a predictor of immunotherapy resistance. Prior published research had already linked SMARCA4 mutations to poor immunotherapy outcomes in lung cancer and STK11 mutations to resistance in non-small-cell lung cancer. What the model added was evidence that when both appear together in the same tumor, the resistance signal is stronger and more consistent than either gene alone. That co-occurrence pattern is precisely what single-biomarker testing misses: SMARCA4 mutant tumors have been documented as having poor immunotherapy response even at high tumor mutational burden, and the co-mutation picture with STK11 deepens that signal considerably.

Interpretability as the Clinical Bridge

A prediction model that cannot explain its reasoning has limited use in the clinic. Oncologists need to understand which biological features drove a prediction in order to weigh it against other patient factors and make a defensible treatment decision. That interpretability requirement was built into the foundation model’s design from the start, rather than added as a post-hoc visualization layer.

Because the tool maps mutation combinations to biological pathways rather than learning purely statistical correlations, the pathway disruptions it identifies can be read back as a form of explanation. A prediction of chemotherapy resistance, for instance, comes attached to information about which molecular pathways the model weighted most heavily – information a clinician or researcher can evaluate against existing biological knowledge, rather than being asked to trust an opaque score.

That design also shapes how the system can contribute to biomarker refinement over time. When a novel association surfaces, as with KMT2D, the pathway-level framing gives researchers a biologically grounded starting point for hypothesis generation, shortening the distance between a computational observation and a testable clinical question. A 2026 BJC Reports analysis on AI and precision oncology flags dataset diversity as a parallel challenge: training populations that underrepresent certain ethnic and geographic groups risk missing population-specific variations in tumor mutational burden and drug metabolism that affect therapy response – a limitation the UC San Diego team will need to address as the model expands.

Pancreatic Cancer, Prostate Cancer, Liquid Biopsies: The Expansion Roadmap

The published study covers 10 solid cancer types. Among the expansion targets explicitly named by the research team are pancreatic cancer, prostate cancer, and sarcomas – all with large unmet needs in treatment stratification and all representing tumor profiles the current training set did not include.

Beyond adding cancer types, the team wants to incorporate other data modalities. Medical imaging, transcriptomics (gene expression analysis, distinct from mutation detection), and electronic health records all carry treatment-relevant information that standard sequencing panels do not capture. Liquid biopsies – tests that detect circulating tumor DNA (ctDNA, fragments of tumor genetic material shed into the bloodstream) from a blood draw rather than a tissue sample – were specifically called out as a future application, particularly for early cancer detection before a solid tumor is large enough to biopsy directly.

The study listed co-authors from Lunit Incorporated, a Seoul-based AI company focused on oncology applications, alongside the UC San Diego team. That collaboration points toward the kind of multi-institutional data-sharing that expansion will require: access to large international genomic datasets of sufficient quality to train on rarer cancer types and data modalities that are harder to standardize across different health systems.

If that data access materializes and the model holds its performance in new tumor contexts, the current 8% treatment-match rate faces structural pressure from a direction the field has not previously had the tools to pursue. If the gains prove specific to the cancer types and treatment classes in the current training set, the tool joins a growing list of promising oncology AI research that still awaits the harder test of broad clinical deployment – where dataset diversity gaps and regulatory standards will matter as much as benchmark accuracy numbers do.