In this work, I will present SparseGMM, a method for gene regulatory network learning and inference, which is novel in its ability to combine co-expression patterns with a graph-based Bayesian framework to infer the complex molecular structure of cancer tissue from transcriptomic data. SparseGMM tackles an important limitation by allowing probabilistic assignments of target genes to modules, which allowed us to calculate an entropy value for each target gene, which represents uncertainty in gene assignment. We further showed that gene entropy reflects molecular crosstalk between major cancer pathways, as well as the heterogeneity of cancer tissue. We applied our method to data from healthy and hepatocellular (HCC) tissue, and captured key biological pathways that are specific to each tissue type. Then, we identified shared pathways between normal liver tissue and liver cancer and their corresponding regulators. Finally, we used single cell RNAseq data to evaluate immune function modules discovered from bulk data. For example, the combined analysis of both tissue types highlights the importance of cholesterol biosynthesis for tumor growth and confirms important regulators, such as SREBF2. In examining high entropy genes, we show that they are more common in cancer than in normal samples. We also show that sparse GMM is able not only to infer key regulators, but also to identify key multifunctional components shared by critical cancer pathways, such as p53 and estrogen signaling.