High-throughput methods for measuring transcript abundance, like SAGE or microarrays, are widely used for determining differences in gene expression between different tissue types, dignities (normal/malignant) or time points. Further analysis of such data frequently aims at the identification of gene interaction networks that form the causal basis for the observed properties of the systems under examination. To this end, it is usually not sufficient to rely on the measured gene expression levels alone; rather, additional biological knowledge has to be taken into account in order to generate useful hypotheses about the molecular mechanism leading to the realization of a certain phenotype. We present a method that combines gene expression data with biological expert knowledge on molecular interaction networks, as described by the TRANSPATH database on signal transduction, to predict additional--and not necessarily differentially expressed--genes or gene products which might participate in processes specific for either of the examined tissues or conditions. In a first step, significance values for over-expression in tissue/condition A or B are assigned to all genes in the expression data set. Genes with a significance value exceeding a certain threshold are used as starting points for the reconstruction of a graph with signaling components as nodes and signaling events as edges. In a subsequent graph traversal process, again starting from the previously identified differentially expressed genes, all encountered nodes 'inherit' all their starting nodes' significance values. In a final step, the graph is visualized, the nodes being colored according to a weighted average of their inherited significance values. Each node's, or sub-network's, predominant color, ranging from green (significant for tissue/condition A) over yellow (not significant for either tissue/condition) to red (significant for tissue/condition B), thus gives an immediate visual clue on which molecules--differentially expressed or not--may play pivotal roles in the tissues or conditions under examination. The described method has been implemented in Java as a client/server application and a web interface called DEEP (Differential Expression Effector Prediction). The client, which features an easy-to-use graphical interface, can freely be downloaded from the following URL: http://deep.bioinf.med.uni-goettingen.de.