MuTATE: an interpretable multi-endpoint machine learning framework for automated molecular subtyping in cancer.

Effective and interpretable molecular subtyping is critical for cancer risk stratification and treatment, yet existing methods face key limitations. Traditional models cannot jointly model multiple clinical endpoints, limiting prognostic utility, while machine learning (ML) approaches often lack transparency. We developed MuTATE, an automated, interpretable decision-tree framework powered by ML that improves subtyping accuracy and enables multi-endpoint risk stratification. MuTATE was evaluated using 18,400 simulations and 682 patient biopsies from three TCGA cancers: lower-grade glioma (LGG), endometrial carcinoma (EC), and gastric adenocarcinoma (GA). Compared to established clinical models, MuTATE improved accuracy, interpretability, and biomarker discovery, and reclassified risk groups. In LGG, MuTATE reassigned 13% of "low-risk" IDH-1p19q cases into higher-risk subtypes, and 19% of "high-risk" IDH wild-type cases were reassigned to higher-risk categories. In GA, MuTATE refined the "intermediate-risk" genomically stable group into a higher-risk ARID1A wild-type subtype. In EC, 72% of "intermediate-risk" MSI/MLH1 cases were reassigned to the highest-risk category. These findings demonstrate MuTATE's potential to reduce diagnostic bias, improve risk stratification, and support scalable integration of multi-endpoint ML into precision oncology workflows.