Mapping cancer dynamics from normal tissue to malignancy using N- and T-gene expression markers.

Background: Carcinogenesis involves two major phases: somatic evolution of normal tissue toward tumor formation, followed by tumor progression toward malignancy. Standard differential expression analysis cannot assign genes specifically to either phase. We introduce N- and T-genes with exclusive expression intervals for normal tissue and tumors, respectively, as markers that could potentially map these dynamics. Methods: Using TCGA RNA-Seq data from prostate adenocarcinoma (PRAD), lung adenocarcinoma (LUAD) and liver hepatocellular carcinoma (LIHC), we identify N- and T-genes through statistically significant expression intervals exclusive to normal or tumor samples, respectively. We discretize expression into three states ( Results: Across different cancer types, normal and tumor samples occupy two well-separated attractors in gene-expression space, giving rise to large sets of N- and T-genes. The number of active N-genes decreases continuously as samples move away from the normal attractor, whereas the number of active T-genes increases as samples progress towards the tumor attractor. Large blocks of N-genes are observed, suggesting coordinated multi-gene deactivation events. The combination of staging through the number of active genes with the molecular taxonomy based on panel genes provides a complete description of the somatic evolution of a normal tissue and the progression towards malignancy of tumors. Conclusion: The N- and T-gene framework provides a natural decomposition of carcinogenesis into somatic evolution (loss of N-gene activity) and tumor progression (gain of T-gene activity). Active N- and T-gene counts can be interpreted as tally marks of somatic evolution and tumor progression, respectively, while complete gene panels enable a taxonomy of samples and helps identifying the biological programs active in each sample. The framework is general and applicable across cancer types, although quantitative results depend on the underlying dataset.