Gene expression data have become an instrumental resource in describing the molecular state associated with various cellular phenotypes and responses to environmental perturbations. The utility of expression profiling has been demonstrated in partitioning clinical states, predicting the class of unknown samples and in assigning putative functional roles to previously uncharacterized genes based on profile similarity. However, gene expression profiling has had only limited success in identifying therapeutic targets. This is partly due to the fact that current methods based on fold-change focus only on single genes in isolation, and thus cannot convey causal information. In this paper, we present a technique for analysis of expression data in a graph-theoretic framework that relies on associations between genes. We describe the global organization of these networks and biological correlates of their structure. We go on to present a novel technique for the molecular characterization of disparate cellular states that adds a new dimension to the fold-based methods and conclude with an example application to a human medulloblastoma dataset. We have shown that expression networks generated from large model-organism expression datasets are scale-free and that the average clustering coefficient of these networks is several orders of magnitude higher than would be expected for similarly sized scale-free networks, suggesting an inherent hierarchical modularity similar to that previously identified in other biological networks. Furthermore, we have shown that these properties are robust with respect to the parameters of network construction. We have demonstrated an enrichment of genes having lethal knockout phenotypes in the high-degree (i.e. hub) nodes in networks generated from aggregate condition datasets; using process-focused Saccharomyces cerivisiae datasets we have demonstrated additional high-degree enrichments of condition-specific genes encoding proteins known to be involved in or important for the processes interrogated by the microarrays. These results demonstrate the utility of network analysis applied to expression data in identifying genes that are regulated in a state-specific manner. We concluded by showing that a sample application to a human clinical dataset prominently identified a known therapeutic target. Software implementing the methods for network generation presented in this paper is available for academic use by request from the authors in the form of compiled linux binary executables.