CancerSubtyper: A Web-Based Deep Learning Platform for Cancer Subtyping Through DNA Methylation Data

Abstract

Cancer subtyping plays a critical role in understanding tumor heterogeneity, predicting patient outcomes, and guiding personalized therapies. While DNA methylation data offers an informative molecular source for subtyping, leveraging its signals across cohorts remains challenging due to high dimensionality, batch effects, and lack of standardized tools. In this thesis, we present CancerSubtyper, a web-based deep learning platform that enables both supervised classification and semi-supervised discovery of cancer subtypes using methylation data. The platform incorporates two models: BCtypeFinder, designed for subtype prediction with domain adaptation, and CancerSubminer, a flexible model for subtype discovery or refinement with potential applicability across different cancer types. Users can upload labeled and unlabeled datasets, select models, and visualize results through various interactive plots including UMAP projections, boxplots, and Kaplan–Meier survival curves. We evaluate the platform using The Cancer Genome Atlas (TCGA) breast cancer cohort, demonstrating distinct cancer subtypes, effective batch correction, and clinical relevance via survival analysis. CancerSubtyper provides a reproducible, user-oriented platform that streamlines cancer subtype analysis using DNA methylation data. It bridges a methodological gap by enabling both classification and discovery tasks, supports batch correction across datasets, and facilitates result interpretation through interactive visualizations. Designed for accessibility, the platform requires no programming expertise, making it a practical tool for researchers and clinicians to explore cancer subtypes and contribute to personalized care.