Cancer Discov
AI ties tumor mutations to immunotherapy response
Clinical takeaway: Applying AI to tumor genomes may extend the value of sequencing beyond single-gene matches, predicting checkpoint inhibitor response from the broader mutation pattern. The tool is research-stage and needs prospective validation before clinical use.
Genetic sequencing is routine in cancer care, but only about 8% of patients are matched to an FDA-approved targeted therapy on the basis of genetics. When that occurs, it's usually a single-gene marker, so most of the mutation information in a tumor goes unused. This study examined whether a model trained across many tumors could turn the full mutation profile into a useful predictor of treatment response.
Dubbed MutationProjector, the model offers a compact representation of the tumor's biological state. Across independent bladder, lung, and melanoma cohorts, it matched or exceeded current methods for predicting immunotherapy response and surfaced both known and unexpected biomarkers of outcomes.
Patients flagged as likely responders to anti-PD1/PDL1 therapy had meaningfully longer survival than those flagged as nonresponders in all three cohorts, with the strongest separation in melanoma and the most consistent signal in lung cancer. Performance matched or exceeded PD-L1 expression, tumor mutation burden, microsatellite instability, KRAS status, and standard machine learning baselines.
Looking at which features drove the predictions surfaced both familiar and less-recognized signals. KMT2D mutation tracked with sensitivity. Co-alterations of SMARCA4-STK11, KEAP1-STK11, and KRAS-STK11 tracked with resistance, even though none of these genes alone were more common in nonresponders.
The model also showed results in chemotherapy, separating cisplatin responders from non-responders in a small bladder cancer cohort.
MutationProjector integrates tumor genetic alterations from 30,328 tumors across 10 solid cancer types with eight molecular interaction networks. The model was pretrained on alterations in 468 cancer-associated genes, plus tumor mutation burden, aneuploidy, and mutational signatures, then fine-tuned on smaller treatment-annotated cohorts. Validation cohorts comprised 130 bladder, 229 lung, and 144 melanoma patients on checkpoint inhibitors, and 42 bladder patients on cisplatin.
The model offers a way to read combinatorial mutation patterns the current biomarker framework misses. Roughly 94% of the gene pairs the model weighted highly would not have surfaced through standard co-mutation analysis. That matters most for patients whose tumors lack a single actionable marker but carry informative combinations.
The authors flag several directions for the model, including expansion to cancer types not in the current set such as pancreatic, prostate, and sarcoma, and integration of imaging, transcriptomics, and electronic health record data into pretraining. Liquid biopsy of circulating tumor DNA for early detection is noted as a potential future application.
"Our results suggest that tumor genome foundation models may help extend the clinical value of sequencing beyond a handful of well-known genes," said Trey Ideker, PhD, professor of medicine at the University of California San Diego School of Medicine and director of the Big Data Institute at the University of Oxford. "This could support a more comprehensive and biologically grounded approach to precision oncology."
Source: Kong J. Cancer Discov. 2026 May 27. A foundation model of cancer genotype enables precise predictions of therapeutic response