A University of California San Diego team published a new AI foundation-style model, MutationProjector, in Cancer Discovery that compresses mutational patterns into a low-dimensional representation designed to predict treatment response. The model was trained on tumor genomics from more than 30,000 patients across 10 solid cancer types using Project GENIE and The Cancer Genome Atlas. Researchers said MutationProjector learns co-dependencies between mutated genes and can recover masked mutation statuses as a test of technical accuracy. In evaluation cohorts, the model matched or exceeded performance of existing prediction approaches for chemotherapy and immunotherapy response. The approach addresses a recurring clinical problem: most patients have multiple alterations on sequencing panels, yet only a minority yield actionable links to approved targeted therapies based on single-gene decision rules.