N-Gene and T-Gene Deregulation Networks: A data-driven causal framework for the analysis of gene interventions in cancer

BackgroundThe construction of realistic gene regulatory networks is currently hindered by the complex, often redundant, combinatorial and multi-layered nature of gene interdependencies, the limited availability of functional annotations, and the intrinsic shortcomings of co-expression-based approaches. Here, we introduce Gene Deregulation Networks (GDNs), a new structure in which linked genes indicate that an expression deregulation in one is likely to propagate to the other. GDNs are inferred directly from gene expression data using a probabilistic theory of causation, without requiring prior functional annotation or other domain knowledge. MethodsUsing highly specific tissue markers--N-genes specifically expressed in normal tissues and T-genes specifically expressed in tumors--we construct separate GDNs for normal and tumor tissues. Data are obtained from TCGA RNA-Seq profiles of bulk tumor and normal tissue samples across multiple cancer localizations. GDN links are preliminarily identified via a statistical test of causal sufficiency between gene expression deregulation events, and then reassessed as spurious or redundant using additional causal criteria. A simple scheme based on the GDNs is implemented in order to describe the evolution dynamics. ResultsThe T-GDN in prostate adenocarcinoma comprises 6138 genes and 102362 directed edges (0.27% of possible connections). Genes with low expression deregulation frequency ( 0.4) show high in-degrees, suggesting they are convergence points of deregulation cascades. EPHA10 (frequency 0.74, in-degree 182) and ENSG00000275479 (out-degree 209) represent these extremes. The N-GDN contains 1097 genes and 4984 edges, featuring compressed nodes reflecting the multi-step nature of somatic evolution. Analysis of tumor samples reveals that early-stage tumors rely primarily on spontaneous gene activations, while advanced tumors exhibit extensive cascade-driven deregulation. Preliminary simulations show qualitative agreement with experiments on gene knockdown in tumor cellular lines. ConclusionsGDNs provide a robust framework to understand cancer progression through deregulation cascades. The separation into N and T networks, connected by NT-genes, i.e. genes common to both networks, offers a systematic basis for modeling carcinogenesis and the possible outcomes of therapeutic interventions.