Gene interactions form complex networks that shape disease susceptibility, prognosis, and therapeutic response. Although bulk transcriptomic datasets represent a clinically scalable and widely available resource for studying these interactions, the application of Graph Neural Networks (GNNs) to bulk transcriptomics remains constrained by limited methodological guidance, particularly for constructing biologically meaningful gene interaction graphs. Here, we present REGEN (REconstruction of GEne Networks), a dynamic graph-based framework that simultaneously learns latent gene interaction networks directly from bulk transcriptomic profiles and predicts patient vital status. Across seven cancer types from The Cancer Genome Atlas (TCGA), REGEN outperformed baseline approaches in five datasets while generating robust and interpretable network structures. By systematically evaluating strategies for initializing gene–gene adjacency matrices, we establish practical guidelines for applying GNNs to clinically relevant bulk transcriptomic data. Importantly, analysis of the learned kidney cancer networks identified cancer-associated pathways and candidate biomarkers linked to disease biology and patient outcome, supporting the translational utility of the approach. Together, these findings provide a principled framework for dynamic gene network learning from bulk transcriptomics and highlight the potential of GNN-based models to support biomarker discovery, patient stratification, and future precision oncology applications without requiring predefined interaction priors.