MutationProjector AI Predicts Cancer Therapy Response from Tumor Mutations
Published in May 2026, the study shows a pretrained model produces compact, interpretable tumor embeddings that match or exceed current prediction methods while remaining at a research and validation stage.
Overview
- The University of California San Diego team published the MutationProjector study in May 2026 after training the foundation model on genomic profiles from more than 30,000 tumors spanning 10 solid cancer types.
- The model was validated across multiple independent patient cohorts, including bladder, lung and melanoma, and matched or outperformed existing methods for predicting response to common immunotherapy and chemotherapy.
- MutationProjector converts complex sets of co‑occurring mutations into compact, interpretable representations that link genotypes to disrupted pathways and treatment outcomes.
- The study identified both known and unexpected candidate biomarkers, including KMT2D linked to immunotherapy sensitivity and combined SMARCA4 and STK11 alterations associated with immunotherapy resistance.
- Authors stress the work is at the research stage and plan broader validation and expansion to more tumor types and data sources such as international genomes, imaging, transcriptomics, electronic health records and liquid biopsies to test clinical utility and help more patients benefit from routine sequencing.