Novel cancer subtyping method guided by tumor-normal sample in latent space of transcriptomic variational autoencoder.

Tumorigenesis is a microevolutionary process in which heterogeneous tumor cells adapt within a complex microenvironment. Current tumor omics analyses often focus exclusively on tumor samples, overlooking the valuable insights that normal tissues can provide. To address this gap, we introduce VaDTN (Variational Autoencoder-Derived Tumor-to-Normal), a pan-cancer framework that integrates transcriptomic data from both tumor and normal samples into a unified latent space. By measuring each tumor's "distance" from a normal reference within this latent space, VaDTN reveals subtle molecular shifts linked to tumor evolution and heterogeneity. We applied VaDTN to six representative cancers (SKCM, BRCA, LIHC, LUSC, STAD, and PAAD), identifying distinct subtypes characterized by unique transcriptional profiles. Notably, four cancer types (SKCM, BRCA, LIHC, and STAD) displayed significant survival stratification based on these subtype groupings, underscoring the clinical relevance of the distance-based approach. This reference-centered perspective thus provides a refined lens for dissecting intra-tumor diversity and guiding potential precision oncology strategies.