MutationProjector AI Model Links Tumor Mutations to Predict Cancer Therapy Response

Researchers at UC San Diego introduced MutationProjector, a foundation-model AI system trained on more than 30,000 tumor genomes from ten solid-cancer types. The model converts thousands of mutations into a low-dimensional embedding that predicts response to immunotherapy, chemotherapy, and metastasis risk, and classifies cancer subtypes.

AI Foundation Model Maps Tumor Mutations to Predict Cancer Therapy ResponseAI Model Connects Tumor Mutations to Predictive Treatment Outcomes

Genomic sequencing is now routine in oncology, yet clinicians rely on a limited set of validated biomarkers. Only about 8 percent of patients receive an FDA-approved therapy matched to their tumor genetics, even though an average tumor harbors roughly 11 distinct alterations, leaving most sequencing data underused.

AI Foundation Model Maps Tumor Mutations to Predict Cancer Therapy Response

The training set includes >30,000 genomes across ten solid cancers. First author JungHo Kong, PhD, fine-tuned the model on as few as 94 immunotherapy-response samples, achieving performance superior to competing approaches on independent cohorts. The model highlighted combinatorial biomarkers (KRAS-STK11, STK11-KEAP1), single-gene signals (KMT2A, SMARCA4), and distinguished HPV-positive from HPV-negative tumors as well as basal versus luminal subtypes in bladder and breast cancer.

AI Foundation Model Maps Tumor Mutations to Predict Cancer Therapy Response

By modeling pathway-level mutational patterns, MutationProjector overcomes the single-gene biomarker bottleneck, offering clinicians actionable predictions for therapy selection, supporting molecular tumor-board deliberations, and revealing novel targets for drug development. Its interpretability also builds clinician trust in AI-driven recommendations.

AI Model Connects Tumor Mutations to Predictive Treatment Outcomes

Ideker, also director of the Big Data Institute at Oxford, described MutationProjector as a general-purpose model that learns from large genomic datasets to improve treatment prediction. He noted that sequencing provides abundant information but few decisions use it, shifting focus from isolated gene markers to pathway-level characterization. The team will expand the model to more cancers, add multi-omic data, and embed it in molecular tumor-board workflows.

AI Foundation Model Maps Tumor Mutations to Predict Cancer Therapy Response

The source material reports consistent performance; no contradictory findings appear. Prospective clinical validation remains an unmet need.

AI Model Connects Tumor Mutations to Predictive Treatment Outcomes

• “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.” — Trey Ideker, PhD

Scripps Institution of Oceanography

• “Right now, increasingly physicians will order gene sequencing panels, and in a couple of years it’ll be the whole genome,” — Trey Ideker, PhD

AI Foundation Model Maps Tumor Mutations to Predict Cancer Therapy Response

• “If I want to learn Italian but I only have one book in Italian, I’m going to learn a lot first from English and Spanish, where I have much more data,” — Trey Ideker, PhD

AI Foundation Model Maps Tumor Mutations to Predict Cancer Therapy Response

• “Those mutations individually may not tell you very much,” — Trey Ideker, PhD

AI Foundation Model Maps Tumor Mutations to Predict Cancer Therapy Response

The team is expanding the training corpus to roughly 300,000 genomes, adding pancreatic, prostate and sarcoma datasets, and incorporating multimodal inputs such as transcriptomics, imaging and electronic health records. Long-term goals include routine use in molecular tumor boards and guiding adaptive clinical-trial designs.

AI Foundation Model Maps Tumor Mutations to Predict Cancer Therapy Response