A Genome-Wide Causal Network Reveals Conserved Cross-Module Regulatory Architecture in Human Cancer

Objective: Classical causal discovery methods fail beyond a few hundred genes, preventing genome-wide causal reconstruction of gene regulatory networks. We apply low-rank differentiable causal discovery to DepMap expression data across 1,208 cancer cell lines, scaling to d = 18,435 genes via W = UV^T factorization (r = 64, a 144× parameter reduction), to construct the first genome-scale directed regulatory network in human cancer. Results: The method recovers 28,247 high-confidence directed edges (predictive r = 0.912 for held-out CRISPR dependencies). External validation against TRRUST v2 yields 94/94 (100%) directional consistency; top hub genes are enriched in COSMIC Cancer Gene Census (38 of 50, p < 10^−16) and STRING (14.2×, p < 10^−300). Structurally, 11 of 13 top hub genes function as Date hubs (cross-module connectors); GTEx analysis across 8 normal tissues found Date hub dominance in 6 of 8 organs. Eighty-seven percent of discovered edges lack measurable CRISPR co-dependency signal, suggesting the network captures regulatory relationships largely independent of genetic essentiality screens. Twenty of 22 top hubs (91%) are druggable. TCGA BRCA survival analysis identifies EZH2 as associated with overall survival after FDR correction (q = 0.036). The complete edge list and validation tables are provided as Supplementary Material.