InteractionFetcher dynamically retrieves remote biological information for selected nodes in the current network within Cytoscape. The plug-in requests biological data via the XML-RPC protocol 18 from a remote server, which retrieves the requested information from an SQL database and passes it back to the plug-in. The plug-in then adds the retrieved information to the current network as additional nodes, edges, and/or attributes. Currently implemented data types include: protein/gene synonyms, orthologs, sequences (gene/protein/upstream), and interactions/associations. Some of this information can be obtained via integrated queries. For example, retrieved gene/protein synonym information may be used to increase the number of molecular interactions that are found. Currently-available interaction-data sets include HPRD 67, BIND 89, DIP 1011, and several other predicted interaction and co-expression data sets 192021. Many options are available, including the ability to do cross-species queries, using ortholog information from Homologene 22 among species including H. sapiens, M. musculus, S. cerevisiae, C. elegans, and D. melanogaster. For example, if two proteins in H. sapiens have not been observed to interact, but both of their orthologs in S. cerevisiae are known to interact, then an inferred interaction (also known as an interolog) can be added to the network. Moreover, the tool allows for easy viewing of the source database's web page or linked PubMed abstract(s) describing each fetched interaction. Because the source code for both the client and server of this plug-in are available, we hope that the capabilities of plug-ins such as these can be expanded by other researchers to include, for example, experimental data (such as mRNA expression levels), metabolic information, or functional annotations. Cytoscape, the InteractionFetcher and related plug-ins, plus all server-side software are open-source and may be obtained at our laboratory web site 23 or at the Cytoscape web site 17.

CytoTalk enables a Cytoscape user to dynamically interact with and manipulate the current network in a Cytoscape window from an external process. This plug-in runs an internal XML-RPC 18 server that enables the currently-displayed network and its various attributes to be manipulated from an external client that is XML-RPC-capable. Example clients may include Perl and Python scripts, scripts written in the R statistical language 24, UNIX shell scripts, C or C++ programs, or Java processes. It moreover expands the developmental possibilities of Cytoscape plug-in developers by allowing other plug-ins to be written in these languages. The external process may be run on the same machine as Cytoscape, or anywhere else on an accessible network. The open-source CytoTalk and Cytoscape software as well as example CytoTalk clients in Perl, Python, and R may be obtained at our laboratory web site 23 or at the Cytoscape web site 17.

For our study, we utilized expression data derived from 28 liver biopsies collected by 13 from 11 HCV-positive liver transplant patients, between 1 and 24 months post-transplant. Since roughly 50% of HCV+ liver transplant patients become re-infected during the two years after receiving their new livers 25, these biopsies provide a unique model for tracking the changes in gene expression during HCV infection 13. To compare gene-expression patterns in liver tissue before and after infection with HCV, 13 collected 28 post-transplant liver biopsies, plus pre-transplant control biopsies, from 11 HCV+ liver-transplant patients. Liver biopsies were obtained at intervals of 3 to 6 months, between 1 and 24 months post-transplant 13. These samples contain a mixture of cell types including hepatocytes, hepatic stellate cells, Kupffer cells (liver macrophages), in addition to various blood cells 13. mRNA expression ratios of about 7000 genes were measured relative to a common reference pool of pre-transplant biopsies. Using Rosetta Resolver(R) software 26, the data were normalized and transformed to log₁₀ ratios, and p-values were computed for expression difference from the reference pool. The measurements showed a high degree of patient-to-patient variation. Most of the genes (5968) were significantly expressed (p <10^-7) in at least one of the 28 samples. The research of 13 involved genomic-expression data derived from human subjects.

We sought to generate a network of molecular pathways that are active (either induced or repressed) in HCV-infected human liver cells. The effects of HCV infection are likely to be complex, and the presence of contaminating blood cells and mixtures of various cell types in the biopsy samples will add further complexity. In order to emphasize the network interface between viral molecules and human molecules, we initiated network construction with a small "seed" network of interactions among HCV-encoded molecules and between HCV-encoded and host-encoded proteins. Interaction data were curated from review articles (2728, and references therein). The seed network also included the JAK-STAT interferon-response pathway that is known to play a role in the response to HCV infection 29. This set comprised 106 interactions between 86 macromolecules (proteins and the viral RNA). The proteins were, when possible, cross-referenced to RefSeq protein identifiers 3031. Figure 1 shows the seed network visualized using Cytoscape 1617. This network is available for exploration and analysis via Cytoscape at our laboratory web site 23.

The seed network was expanded to a full "scaffold" network using the 5968 genes implicated by the genomic-expression data and large-scale molecular-interaction data sets in public databases by searching for interactions among the 5968 expressed genes and the molecules in the seed network. To automate the construction of the scaffold network, we implemented a Cytoscape plug-in, InteractionFetcher, for dynamic retrieval of molecular interactions and binding partners via the Internet. InteractionFetcher rapidly adds interactions among molecules of interest in a network. In addition, it may be used to iteratively expand a network through "in silico pull-down" of molecules that are currently not present in the network but are known to interact with molecules that are present. Using this plug-in, we were able to integrate as many as 15,000 interactions among the proteins implicated by the HCV expression-data set and seed-network proteins (among the available interaction data sets, which include HPRD 67, BIND 89, DIP 1011, PreBIND 3233, and several other predicted and co-expression data sets; see Methods). However, for this paper, we restricted our search to individually curated human-only protein interactions from HPRD and BIND, resulting in a scaffold network of 4,592 unique interactions among 1,950 molecules (Figure 2). This network is available for exploration and analysis with Cytoscape at our laboratory web site 23, which also allows for easy viewing of additional information provided by InteractionFetcher, such as each interaction's source database web page and PubMed abstract identifier(s).

A useful method of integrated analysis of expression data within an interaction network is the ActivePaths algorithm 5. This method identifies contiguous pathways or subnetworks that are active (induced or repressed relative to randomly selected subnetworks) in subsets of the expression data. We applied this algorithm, which is available as a Cytoscape plug-in, to the scaffold network. Due to the high order (number of vertices) and size (number of edges) of the scaffold network, it was necessary to iteratively apply the algorithm, as suggested by the developers, to obtain increasingly smaller active subnetworks until they contained fewer than 100 nodes. The resulting four active subnetworks contained between 40 and 121 interactions. Because there were overlaps among these four highest-scoring active subnetworks, we combined them into a single fully connected active subnetwork.

Additional analyses were performed by selecting scaffold subnetworks that are significantly active and/or co-regulated in temporal subsets of the microarray data. Because the scaffold network is not differentiated with regard to tissues, cell types, or cellular state, and the biopsy samples from which the expression data were derived likewise contain mixtures of cell types and other contaminants, the information in the active scaffold network does not, by itself, answer the questions we are addressing. To increase our chance of identifying pathways that might be modulated in response to HCV infection, we performed a differential analysis of the scaffold network, to identify subnetworks that become active more than eight months after transplant. This choice of cut-off was made to nearly-evenly divide the expression data into two halves (those from biopsies prior to, and after, eight months post-transplant), and by performing a differential analysis we can hope to subtract out some of the effects of the transplant and post-transplant immune response signals from those of HCV reinfection and progression.

We used the R statistical environment 24 to perform this analysis. Because R is external to Cytoscape, we developed a plug-in, called CytoTalk, that enables a user of R (or a wide variety of other environments or languages; see Methods) to interactively query and modify Cytoscape networks, thereby greatly expanding the analytical capabilities available to users of Cytoscape. We used R with CytoTalk to select the proteins and interactions implicated by specific statistical queries on the expression data. This enabled us to extract a subnetwork of genes that were significantly induced or repressed with |log₁₀(ratio)| > 0.4 in biopsies obtained more than 8 months after transplant. The proteins encoded by these genes form a "late-active" network. We similarly extracted an "early-active" network encoded by genes that were active in biopsies obtained earlier than 8 months after transplant. We compared these two networks and identified an "only-late-active" subnetwork that was not active prior to eight months, but was active afterward. The expectation is that this "only-late-active" subnetwork will contain pathways from the "late-active" network that are activated in response to Hepatitis C virus re-infection, while pathways from the "early-active" network that may contain pathways activated as a result of the transplant are removed.

In Figure 3, we have integrated the seed network, the composite active-paths network, and the "only-late-active" network into one network. This network is available for exploration and analysis in Cytoscape at our laboratory web site 23. Genes that were induced on average after 8 months following transplant are indicated with a red colour. Genes that were repressed are green. We have highlighted the nodes and edges of the composite active-paths subnetwork in "bold". The network in Figure 3 is significantly over-represented with genes of several biological processes, as annotated by the Gene Ontology Consortium Database 3435; using the BioDataServer tool in Cytoscape, and computed in R via CytoTalk, using the Bonferroni-corrected hypergeometric distribution. Among these include blood coagulation (p = 10^-11), immune response (p = 10^-7), proteolysis and peptidolysis (p = 10^-5), lipid transport (p = 10^-3), and complement activation (p = 10^-2). In addition, nearly the entire JAK-STAT interferon-response signalling pathway is activated in this network.

The visualization in Figure 3 enables one to identify these pathways and see whether they are "turning on" (red) or "turning off" (green) in the expression data. For example, the blood coagulation pathway is active in the expression data, although (as is to be expected with large and complex pathways) not coherently induced or repressed. The interferon-response pathway and genes activated by ISGF are clearly induced, probably due in part to the immune response to viral infection and partly in response to standard treatment of HCV-positive patients with interferon-alpha. Also, genes encoding the Toll-like receptors TLR1 and 2, as well as the downstream signalling pathway connecting them, through MYD88, to the interferon-response pathway appear to be repressed. TLRs 1 and 2 are known viral detection receptors; it is known that TLR2 detects HCV 36. The interleukin receptor IL1R1, upstream of MYD88, is also repressed along with other IL receptors, whereas IL1A and B are induced. Additionally, we see that many apoptosis-related genes encoding TNF, TNF receptors, and TNF-signalling factors, are activated, whereas growth factors (IGF and connected pathways), and cell cycle and translation-related pathways (e.g. CDKN and connected pathways) are repressed. Ignoring the observed responses that are likely due to by-products of the biopsy process (e.g. the blood coagulation pathway), the active pathways observed are jointly consistent with a large-scale response of complex molecular pathways to viral infection: hepatic cell reproduction is repressed and programmed cell death is induced.

Finally we note that a visual inspection of the network suggests that many of the proteins that bind directly to HCV-encoded molecules (i.e., are their first neighbours in the network) appear on average to be down-regulated relative to the rest of the network. Statistical analysis of the data supports this suggestion. As computed via CytoTalk and R, about 80% of the first neighbours of the viral RNA and proteins are down-regulated in the network of Figure 3, compared to 35% of the remaining genes in the network (p = 0.0029). This finding suggests two non-exclusive possibilities: genes encoding HCV-neighbour proteins are targets of host regulatory mechanisms counteracting viral replication; or they are targets of virus-encoded regulatory mechanisms that sabotage anti-viral defences.