Integration of biological networks and gene expression data using Cytoscape

Obtain network data

This section describes three ways to import network data into Cytoscape.

The first method is to query protein interaction databases such as cPath³³ with a list of genes of interest. cPath queries the IntAct¹⁵ and MINT¹⁷ databases. This is an appropriate method for users who are interested in assessing the connections between genes with significant experimental responses, in well-studied organisms such as S. cerevisiae or Homo sapiens. Cytoscape users can interrogate additional protein interaction databases via MiMi³⁴ or BioNetBuilder³², plugins that are similar in application to cPath. In each of these cases, following the steps in this workflow section yields a network that contains known and putative functional associations between the genes of interest.

The second method is to build a text-mining association network using the Agilent Literature Search plugin²¹. This method is most appropriate for those who are working in organisms that are not well represented in protein interaction databases, or want to restrict the network to associations observed only in specific contexts. For Literature Search the user builds a set of queries by entering terms, such as gene names, and contexts, such as an organism or disease name. The query set is submitted to selected search engines, for example PubMed or OMIM. The resulting documents are fetched, parsed into sentences, and analyzed for known interaction terms, like ‘binding’ or ‘activate’. Agilent Literature Search uses a lexicon set for defining gene names (concepts) and aliases, drawn from Entrez Gene, and interaction terms (verbs) of interest. An association is extracted for every sentence containing at least two concepts and one verb. Associations are then converted into interactions with corresponding sentences and source hyperlinks, and added to a Cytoscape network. Interaction network and text-mining association network data are complementary: protein interaction databases contain experimentally determined interactions; whereas text mining association networks contain more general association types and offer an alternative network source where interaction data are limited.

The third method described for importing network data in to Cytoscape is to import a network file, such as a SIF (Simple Interaction Format) file. The SIF file format is detailed in Box 1. These files are straightforward for a user to create with a standard text editor. SIF file import is the most appropriate method for users who want to focus their analysis on network data identified in advance, such as those who are interested in the impact of the experimental conditions on sets of specific interactions or pathways.

In addition to the import methods described in this protocol, Cytoscape users can also import pathways from repositories such as KEGG⁵⁸, Reactome⁵⁹ via the PSI-MI, BioPAX, or SBML data exchange formats (as reviewed in ref. 60), although such pathway data contain non-pairwise interactions between molecules. For instance, there is a single interaction between multiple substrates and products in a biochemical reaction, which must be mapped to pairwise interactions in Cytoscape.

After importing interaction data, the user may optionally filter the resulting network to reduce the network size by selecting just the types of network information of interest. For instance, the CABIN plugin¹² enables the user to merge network data from multiple experimental sources and select the interactions observed multiple times, which are likely more reliable than those observed once.

Explore network and generate layout

This section introduces the user to basic aspects of Cytoscape operation, including network navigation and layout. Apart from basic operation, basic filter features are available to identify genes as important by virtue of their high number of connections in the network⁶¹. Essential genes are also associated with nodes, which occupy central positions in large interaction networks, particularly those frequently found to connect other nodes together⁶². This section uses the yFiles layout algorithm, although many other layout algorithms exist, some of which support more specialized operations.

Annotate with attribute and expression data

Expression profiles can provide powerful insights into cellular state and dynamics when integrated with network data. For example, if two nodes consistently show similar changes in expression levels, or they consistently show changes in opposite directions, then one gene might regulate the production of the other. Also, a group of connected nodes characterized by large fold changes may represent a signal-transduction cascade or protein complex that is repressed or induced under the experimental condition⁶³. This section shows how to apply mRNA expression data to a network as an example, but the steps outlined apply to any form of molecular profile data.

Analyze network features

This section explains how to run analysis methods that identify putative functional or structural modules within the network that may highlight protein complexes active under a profiled experimental condition as complex topologies of interaction networks make this difficult to do by eye. The jActiveModules plugin⁴² automates this analysis and identifies connected sections of the network in which the nodes have significant P-values. This indicates a group of nodes that may be coregulated, suggesting a module whose activity is influenced by the experimental context of the expression data. A complex is a module in which the macromolecules involved form a structure to execute some function called a molecular machine. Densely connected regions in protein interaction networks tend to correspond to protein complexes. The MCODE plugin⁴³ identifies putative complexes by finding regions of significant local density. At this stage, the user could also identify over-represented network motifs with the Metabolica or NetMatch plugins³⁹, although these alternatives are not covered here.

Detect enriched gene functions

The function of interesting gene sets or network regions can be summarized by finding significantly enriched functional annotation terms. This may be used to support module or complex predictions. This section outlines steps required to identify enriched GO processes with the BiNGO plugin⁷. Users can optionally use the GOlorize plugin³⁶ to refine the network layout according to selected GO classes.

Cytoscape can be used in many additional visualization and analysis workflows, including published protocols on the reverse engineering of regulatory⁶⁴ and metabolic⁶⁵ networks, and can be extended using Java programming to implement new features and analysis methods. Over 40 excellent plugins contributed by many software development groups provide examples. While these topics are not covered here, Cytoscape has a community of software developers and users who can answer questions about potential new Cytoscape uses, as detailed at http://www.cytoscape.org.

EQUIPMENT

• Internet access

EQUIPMENT SETUP

Obtain network data

1. Start Cytoscape. For webstart users, go to http://www.cytoscape.org/nature.protocols/ and click on the indicated link to launch Cytoscape. This will automatically download Cytoscape to the local computer and start it. For users with Cytoscape already installed locally, navigate to the Cytoscape home directory and execute cytoscape.bat (Windows users) or cytoscape.sh (Linux and Mac OS X users).

2. After Cytoscape has started, you should see a Cytoscape Desktop window, as shown in Figure 1. (a) The main Cytoscape canvas displays network data and is initially blank. (b) The toolbar, at the top of the Cytoscape desktop, contains the command buttons. A description of each command button is shown when the mouse pointer hovers over it. (c) The Control Panel, at the left, displays the Network tree viewer. This lists the available networks by name and number of nodes and edges. (d) The Network Overview Pane, at the lower left, shows the current network in white with the displayed portion overlaid in blue. (e) The Data Panel, at the lower right, can be used to display node, edge and network attribute data.

3. Identify a list of genes of interest. Typically, these will be genes with a pronounced response to your experimental conditions of interest. This protocol describes three options for importing network data related to these genes into Cytoscape: querying interaction databases (option A), building an association network through text mining (option B) and loading your own network data from a text file (option C). Continue with one or more of these steps based on which is most appropriate for your use case. Users wishing to follow this protocol as a tutorial should download the Supplementary Data 2 file (galFiltered.sif) and continue with Step 6. The steps in this protocol are most effective with a network of at least 250 interactions. To obtain such a network, start with a set of at least 25 genes and add more genes or import more interactions as needed to obtain a network of the desired size.

Explore the network and generate a layout

• 4

  Reorganize the network by going to the Layout menu and selecting yFiles → Organic. This algorithm implements a variant of the force-directed paradigm⁶⁷ in which nodes are modeled as objects with mutually repulsive forces, edges induce attractive forces between the nodes they connect and nodes are placed such that the sum of these forces is minimized. This layout has the effect of exposing structure inherent in the network. In particular, it facilitates identification of clusters of tightly connected nodes, which suggest functional modules, and ‘hub nodes’, which are involved in many interactions and often represent functionally important proteins. Note that Cytoscape provides a wide variety of alternatives for layout, including hierarchical layouts, circular layouts and attribute-based layouts. Experiment with several of these.

• 5

  To generate more space for the network canvas, ‘float’ the Control Panel and Data Panel by clicking the Float Window control, the icon at the top right corner of each panel in the title bar. This detaches the panels as separate windows. Resize the network canvas by dragging its lower right corner.

• 6

  Select any node by clicking it with the left mouse button. Notice that the selected node turns yellow. Select several nodes by holding down and dragging the left mouse button to define a rectangular selection region.

• 7

  Notice that the Data Panel lists the identifiers of all nodes selected. This browser can also display node attributes. Nodes obtained from cPath have attributes including ‘FULL_NAME’, a descriptive protein name. Nodes from the Agilent Literature Search plugin have attributes including ‘NbrConnections’, the degree of the node.

• 8

  Select one of these attributes for display by clicking on the Select Attributes button, the leftmost button of the Data Panel toolbar. This will display the list of available attributes, with selected attributes highlighted. Select or deselect an attribute by clicking on its name. Exit by clicking the right mouse button.

• 9

  Select one or more edges by dragging a rectangular region across them. Edges obtained from cPath have attributes including ‘EXPERIMENTAL_SYSTEM_NAME’, the type of experimental evidence supporting the interaction. Edges obtained from the Agilent Literature Search plugin have attributes including NbrSentences, the number of distinct sentences that support the association.

• 10

  View these attributes by clicking on the Edge Attribute Browser tab at the bottom of the Data Panel and selecting attributes for display as described in Step 8.

• 11

  Zoom into the network using the Zoom In button on the toolbar, depicted as a magnifying glass with a plus sign (‘+’). Notice that in the Control Panel, the blue box in the Network Overview window shrinks to depict the portion of the network displayed. Move this box to navigate the network. Other controls include the Zoom Out button (to the left of Zoom In) and the Zoom Selected Region button (to the right of Zoom In). Alternatively, zooming can be accomplished with a right-click drag in the main Cytoscape canvas.

• 12

  Identify the ID name of any node in the network, such as ‘YPL248C’ for the sample network in galFiltered.sif. Search for that node ID by going to the Search: field at the right of the Cytoscape toolbar and typing the name. After each character is typed, a drop-down menu displays the list of node IDs matching the portion of the name entered. This field can be configured to search any attribute (not just ID) by clicking the ‘Configure search options’ button to the right.

• 13

  Cytoscape stores user data as a session file, containing all network, layout and attribute data in the program’s memory. Save the current session by clicking on the Save Current Session As… button on the toolbar (appearing as a floppy disk) or by going to Save or Save Asyunder the File menu. Once a session is saved, it can be reloaded with the Open option under the File menu. A sample session file is provided under Supplementary Data 4, SampleData.cys, containing a SIF file imported in Step 3C and an association network for the same genes from Step 3B.

Annotate with attribute and expression data

• 14

  Optional step: Additional node or edge attribute data can be integrated with your network by following the steps in Box 3 (and see Supplementary Data 5).

• 15

  The expression data should be formatted as described in Box 1. For users wishing to follow this protocol as a tutorial, the Supplementary Data section includes a sample expression data file (Supplementary Data 1, galExpData.pvals); this file can be viewed with a standard text editor to reveal the file format. If the identifiers in the first column of your expression file match the node names in the Cytoscape network exactly in a case-sensitive match, then skip directly to Step 16 (this applies to users of the sample data). Otherwise, follow the procedure described in Box 3 to load an attribute file that describes the correspondence between the gene/protein ID names in the expression data file and the ID names of nodes in your network. Box 4 provides further details of one method to build such an attribute file.

Load expression data

• 16

  Load the expression data file by going to the File menu, then Import and then Attribute/Expression Matrix. Select your expression data file in the dialog box. If an ID-mapping attribute was loaded in Step 15, then specify this attribute for expression data import by using the control labeled Assign values to nodes using…. For the sample data, specify the Supplementary Data file galExpData.pvals (Supplementary Data 1) and assign values to nodes using the attribute ID. Click on Import.

• 17

  Check the list of node attributes in the Node Attribute Browser. There should be two new sets of attributes per experiment, named ‘Xexp’ and ‘Xsig’ where ‘X’ is the name of your experiment. The expression values will be contained in ‘Xexp’. If your expression data set contains P-values (see Fig. 3 and Box 1), they will be contained in ‘Xsig’; otherwise, ‘Xsig’ will be blank. Users of the sample data should see attributes for experiments ‘gal1RG’, ‘gal4RG’ and ‘gal80R’: expression fold changes in ‘gal1RGexp’, ‘gal4RGexp’ and ‘gal80Rexp’; and P-values in ‘gal1RGsig’, ‘gal4RGsig’ and ‘gal80Rsig’. Verify the expression data import by selecting attributes from one or more experiments for display and selecting one or more nodes on the Cytoscape canvas.

Setting visual node properties

• 18

  The Cytoscape VizMapper controls the mapping of node and edge attributes to visual properties. Open the VizMapper by navigating to View in the Cytoscape menu and then Open VizMapper. This will bring up the VizMapper in the Control Panel, as shown in Figure 4a. Note that if you have not done so already, you can expand the width of the Control Panel by ‘floating’ it into a separate window, as described in Step 5, and dragging on the resize control at the lower right corner of the window. The VizMapper contains a two-column list, showing the available visual properties at the left and their current settings at the right. Expanding the window may give a better view of the settings.

Color nodes with red-to-green gradient

• 19

  The first step to color your nodes according to expression values is to locate the Node Color property, which is grayed-out initially. Double click on it.

• 20

  Next to Node Color, you will see a message reading ‘Please Select an Attribute!’. Click on this message. This will bring up a list of node attributes. Scroll down to your expression data value. For sample data users, this will be ‘gal4Rexp’.

• 21

  Underneath Node Color, locate the entry for Mapping Type and click in the right column, directly underneath your expression data attribute. A pull-down menu will appear. Select Continuous Mapping.

• 22

  Underneath Mapping Type, you will see the label Graphical View, located next to an empty rectangle, as shown in Figure 4a. Click on the rectangle. This will bring up the window Gradient Editor for Node Color, which shows a bar corresponding to the range of your expression data attribute, with the Min and Max values displayed at the two ends.

• 23

  In the Gradient Editor for Node Color window, click the Add button twice. You will see the range bar for your expression data attribute divided in half, with the lower half colored black and the upper half colored white. See Figure 4b.

• 24

  Double-click the white downward-facing triangle in the center of the range to bring up a color selection window and select Green. You should now see your range bar divided in half, with the left half colored black and the right half showing a color gradient from green to white. See Figure 4c.

• 25

  Slide your green triangle toward the left end of the range bar, leaving a short black portion at the left end of the range bar. See Figure 4d. On the Cytoscape canvas, you should see some nodes colored black (those with the lowest expression values), most nodes colored in various shades of darker and lighter green and some nodes colored white (those with the highest expression values).

• 26

  Click the Add button one more time. Locate one of the downward-facing triangles at the right end of the bar and slide it to the center of the range. See Figure 4e. There should now be triangles at left, center and right.

• 27

  Double-click the downward-facing triangle at the right end of the bar to bring up a color selection window. Select Red. This will create a gradient of white to red at the right side of the range. See Figure 4f. On the Cytoscape canvas, you should see nodes with higher expression values colored pink or red, with darker colors indicating higher expression values.

• 28

  Slide the red triangle a short distance to the left. Double-click on the white right-facing triangle at the right end of the bar and select Blue. See Figure 4g.

• 29

  On your Cytoscape canvas (Fig. 4h), you should now see nodes colored according to a red–green continuum: (i) Nodes with expression values at the lower end of the range are colored more green, those with expression values at the higher end of the range are colored more red and those in the middle of the range are colored more white. (ii) Lighter colors indicate expression data values near the middle of the range, whereas darker colors indicate values near the ends of the range. (iii) Notice that as you slide the triangles in the range bar, the colors of nodes on the Cytoscape canvas change accordingly. (iv) Close the Gradient Editor window.

• 30

  By default, nodes are colored pink. Consequently, a pink node color could imply an expression data value somewhat higher than the middle of the range, or it could imply that there were no expression data for the node. To remove this ambiguity, change the default node color to gray, as follows: (i) Under the VizMapper, locate the Defaults panel, showing the default node and edge rendering. (ii) Click on either node in this panel. This will bring up a window showing the default node attribute settings. (iii) In this window, click on the button for Node Fill Color and select Gray in the color selection window. Click Apply.

• 31

  The nodes on the canvas should now be colored according to expression levels, which can suggest patterns of coregulation and modular structure. For example, examine the network for connected regions in which the nodes are predominantly colored red or green (see Fig. 1). Such regions are suggestive of protein complexes or signaling pathways whose activities are being modulated in the expression experiment. Several utilities are available to locate such regions automatically within large networks, including the jActiveModules plugin (Step 33) and Cytoscape Filters (not described here—see Cytoscape manual at http://www.cytoscape.org).

• 32

  Supplementary data users: Change the node label from the default ID to the common gene name attribute as follows. (i) Locate the Node Label entry in the VizMapper. (ii) Click on the field opposite Node Label, which is set to ‘ID’ by default. Select ‘GeneName’ from the menu. (iii) Zoom in to the network. You should now see nodes labeled by their common gene names.

Analyze network features

• 33

  Functional modules can be identified as highly connected network regions with similar responses across multiple experimental conditions. These patterns can be detected visually, but in practice the high level of connectivity in interaction networks makes visual detection difficult. Putative modules can be identified with the jActiveModules plugin as follows. Further details are available in the original publication⁴². Under the Plugins menu, select jActiveModules. This will bring up a jActiveModules section on the Control Panel, listing the available expression experiments.

• 34

  Select all of your expression experiments by clicking on them with the left mouse button (one can run jActiveModules on only one expression experiment or on a subset of the experiments as well).

• 35

  In the General Parameters panel, keep the default values, as these are effective for most initial analyses. (a) ‘Number of modules’ indicates the number of putative modules that will be reported. (b) ‘Adjust score for size’ corrects for the fact that a larger putative module is more likely to contain nodes with significant P-values by random chance. (c) Regional scoring affects how the score of a given module is calculated. Instead of scoring only those nodes within the module, the neighboring nodes of the module are also included. This aids in the identification of active modules in networks that contain nodes with many neighbors (i.e., hubs). Consider a transcription factor with many targets: even if the transcription factor is not active, it is likely that some of its targets will be expressed simply due to random chance. Without regional scoring, this subset of targets can be selected as an active module, even though this arrangement is not unexpected. Regional scoring prevents this problem by forcing all targets to be scored simultaneously.

• 36

  In the Strategy panel, the user can select the search strategy used to identify high-scoring modules. (a) Search: By this strategy, local (greedy) searches are initiated from single nodes in the network. (b) Search Depth: At each step in the greedy search, this parameter determines how close (as determined by the shortest path) a node must be to the current active module to be considered for inclusion. (c) Max Depth: This parameter determines how close a node must be to the initial seed node to be considered for inclusion. (d) Search from Selected Nodes: By default, a separate search is initiated for each node in the network. Using this option, searches are initiated only from those nodes selected by the user. (e) Anneal: In this strategy, all active modules in the network are discovered simultaneously using the method of simulated annealing as previously described⁴². The Annealing Parameters define the simulated annealing schedule. Annealing Extensions enable various modifications to the simulated annealing search procedure. For a detailed description of all parameters, click the Help button in the jActiveModules main panel.

• 37

  Click the Find Modules button.

• 38

  Shortly, the Results panel will appear, as illustrated in Figure 5. This window contains a table in which each row represents one putative module, listed according to the number of nodes, and an associated Z-score. Z-scores greater than 3.0 are generally considered significant. The table contains one column per expression experiment; for these columns, any row in which the cell is filled indicates an active module that showed a significant response under the conditions of the experiment. Clicking on any row causes the corresponding nodes to be selected in the Cytoscape canvas. To view these nodes as a separate network, click on the Create Network button. Lay out this new network using Layout → yFiles → Organic from within the new view.

• 39

  Repeat the Find Modules step a few times. Depending on the parameter settings, jActiveModules relies on random sampling and is not guaranteed to return the same result with each execution. Consequently, it is worthwhile to run jActiveModules repeatedly to ensure that the approach is converging, that is, the identified modules are reproducible across runs. One method for validating module predictions is to determine if the nodes in the putative module are enriched for any GO biological processes, since it is expected that all genes in a module are involved in the same biological process. See Step 43.

• 40

  Highly connected network regions can indicate protein complexes. As with the differentially expressed modules in the previous step, complexes can in principle be detected visually, but in practice this is difficult due to complex topologies of interaction networks. Putative complexes can be automatically identified using the MCODE plugin, as follows. Further details are available in the original MCODE publication⁴³. Note that MCODE analysis does not require an expression data set, and thus can alternatively be performed immediately after Step 13. At this point, the Cytoscape session might contain several networks. Complexes are easier to detect in larger networks. Move to your largest interaction network by selecting the Network control in the Control Panel and then mouse-click on this network.

• 41

  Run MCODE by navigating to the Plugins menu, then MCODE → Start MCODE and then clicking the Analyze button in the Control Panel. After a few seconds, the results will appear in the Results Panel, as shown in Figure 6.

• 42

  The results present a diagram of each putative complex and provide details including a score, the number of nodes and edges, and a diagram of the complex. Significant predictions are typically those with scores of greater than 2.0, and at least four nodes. Clicking on any row will highlight the corresponding nodes in the network canvas. The Cytoscape canvas shows the network according to the MCODE visual style, in which red nodes belong to high-scoring predicted complexes and black nodes do not. Squares and circles signify nodes in predicted complexes: squares indicate the ‘seed’ of the prediction, which are usually the nodes in the complex with the highest connectivity. Diamonds indicate nodes that are not part of a predicted complex. Note that the visual style can be changed back to its original settings by going to the VizMapper in the Control Panel and selecting a visual style using the Current Visual Style control. One method to validate a prediction of a complex is to determine if the nodes of the putative complex are enriched for any GO biological processes. The next section outlines the steps required.

Detect enriched gene functions

• 43

  At this point, you should have identified one or more sets of nodes that may be functionally related, such as nodes in an expression module (jActiveModules) or a putative complex (MCODE). We will use the BiNGO plugin to assess if the nodes in a set are enriched for any biological processes recorded in the Gene Ontology database⁶⁸. Further details are available in the original BINGO publication⁷. Select a group of nodes with a putative functional association. For instance, select a putative module from the jActiveModules results (Results Panel).

• 44

  Under the Plugins menu, select BiNGO 2.0. This will bring up the BiNGO Settings window.

• 45

  In the Cluster name: field, choose a name (arbitrary) of your selected node cluster. In the field labeled Select organism/ annotation: select the organism (Saccharomyces cerevisiae for the sample data). The default values shown in the other menu options are appropriate for most cases. For more information on any parameter, click the Help button and follow the links to the BiNGO User Guide.

• 46

  Click on the Start BiNGO button.

• 47

  After a few minutes, the BiNGO Output window should appear listing the P-value (corrected for multiple testing) for each enriched GO class and the corresponding nodes from the network. On the Cytoscape desktop, a new network should appear as shown in Figure 7. In this network, each node represents a class in GO, and an edge between two nodes indicates that one class is the parent of the other. The rendering of each node indicates any significant enrichment for the corresponding GO class: (a) The size of each node is proportional to the number of proteins from the input set that belong to that class. (b) Nodes are color-coded according to P-values reporting the significance of enrichment of that term in the input set. Darker colors indicate a more significant degree of enrichment.

The time required to execute this protocol is most strongly related to the time required to download the sample data and software over the Internet. With a 1.5 Mbs ADSL connection under favorable operating conditions, it takes approximately 9 min to download Cytoscape under Java Web Start, 2.5 min to download a network of ~1,000 nodes and ~2,000 interactions from cPath and 8 min to download and process 1,000 articles with the Agilent Literature Search plugin. On a Macintosh Powerbook G4 with 768 MB of memory, a network of ~3,000 nodes and ~7,000 edges requires approximately 1.50 min for loading, 2.25 min for analysis with jActiveModules using one expression experiment and 0.75 min for analysis with MCODE. An experienced user can execute the full protocol described within 30 min.

Limited network data

The effectiveness of network analysis depends strongly on the amount of network data available. Species such as yeast, fly, Escherichia coli and human have tens of thousands of measured interactions available. For other species, the methods outlined here will likely become more effective over time as the amount of interaction data increases. Even in these cases, a gene association network can be constructed from published abstracts, the usefulness of which depends on the number of available publications, not the number of protein interaction assays. Another option when protein interaction measurements are lacking is to perform network analysis on a closely related organism for which more interactions have been measured (or for which a greater volume of publications are available). Protein identifiers can be mapped via orthology relationships, that is, as calculated by services such as PhyloFacts⁶⁹ or Roundup⁷⁰.

Mismatched gene identifiers

The most common problems with Cytoscape arise from mismatched gene identifiers across different input data files. This problem occurs most often when integrating attribute or expression data with an existing network. In Cytoscape, the identifiers in attribute or expression data files should match exactly to node names in the network in a case-sensitive match. If there is a case-sensitive match but data are still not loaded correctly, the following steps can help isolate the source of the problem:

• (1)

  Under the Node or Edge Attribute Browser, select the new attribute from the list of available attributes by clicking on the leftmost button in the attribute browser toolbar (Select Attributes; square button with horizontal gray central stripe). If the name of the attribute does not appear, then the input file was not correctly loaded. Verify the format of your input file against the sample files provided in Supplementary Data or in the sampleData subdirectory within the Cytoscape home directory.

• (2)

  If the name of the attribute appears, then select several nodes (or edges) in the Cytoscape canvas and check their attribute values in the Node/Edge Attribute Browser against the expected values. If any node or edge has a blank attribute value, then the identifier of the node or edge could not be matched against the input file. Ensure that the input attribute file contains exact matches to the nodes in the network.

• (3)

  If attempting to load expression data using an ID-mapping identifier, first verify that the probeset attribute file was loaded correctly using the steps above. If the nodes have the probeset identifiers assigned correctly but have blank values for expression results, there is probably a mismatch between the probeset identifier in the probeset attribute file and the expression data file.

Multiple values for a node attribute

Sometimes a node attribute is a list of multiple values rather than just a single value (e.g., multiple gene function terms per node). However, in most other cases, Cytoscape assumes that a node or edge will be assigned only one value per attribute. In these cases, if a node or edge attribute file contains two values for the same node or edge, such as two probesets for one gene, the latter will be retained and the former overwritten. Consequently, if you discover that the expression data loaded for some node(s) are not as expected, verify that multiple attribute entries have not been specified.

Out-of-memory errors

As is true with many programs, Cytoscape can fail when the computer runs out of memory. This problem is manifest in a myriad of ways. Typically in these cases, Cytoscape will display an error message such as a Java Null Pointer Exception. In other cases, no error will be reported, but a pop-up window may fail to appear on the screen or a network may fail to appear in the Cytoscape canvas. If Cytoscape exhibits such behavior when analyzing a very large network, or running alongside many other applications on the same computer, insufficient memory may be the cause. The problem may be addressed by freeing memory: deleting extraneous networks, closing other applications that are running on the same computer or saving the session and rebooting the computer. If sufficient memory is available, Cytoscape may need to be set to use it. Further instructions on this are available at http://www.cytoscape.org/cgi-bin/moin.cgi/How_to_increase_memory_for_Cytoscape.

Failure to display very large networks

If a loaded network fails to display in the Cytoscape canvas, it may be that the loaded network is very large. By default, Cytoscape displays (or ‘creates views’) for networks of up to 10,000 nodes. If the user attempts to load a network of over 10,000 nodes, the network will be loaded into memory but by default no view will be created. In such cases, the network is listed in the control panel in red. At this point, the user can create a view by right-clicking on the network name in the Control Panel and selecting Create View. Before doing so, the user should bear in mind that visual analysis of very large networks is not usually productive, and that it may be more expedient to issue a more modest query or subdivide the input file. Also, the user can filter to select nodes and edges of interest in a loaded network even if it does not have a view and these can be extracted to create a smaller, more manageable network that can be reasonably viewed.

Network collection and layout

We have obtained the following average numbers of interactions for networks constructed based on the following methods and representative species. Numbers represent (mean ± s.d.) interactions per gene sampled from ten randomly chosen genes in July 2007. cPATH: yeast 90.0 ± 194.1, human 63.1 ± 81.6, mouse 3.3 ± 5.5, rat 0.5 ± 1.1. Agilent Literature Search: yeast 18.9 ± 22.7, human 178.8 ± 134.0, mouse 152.5 ± 143.1, rat 60.3 ± 92.1. If the network sizes obtained are much smaller than expected, see TROUBLESHOOTING. Successful import of the sample SIF file and application of the yFiles Organic layout algorithm produces a Cytoscape view similar to that shown in Figure 1.

In the near future, Cytoscape and its plugins will expand to make network loading easier. For instance, the cPath³³ plugin currently serves protein interaction data from recent releases of the IntAct¹⁵ and MINT¹⁷ public data repositories. The cPath site will grow to include a larger number of databases that share data in the BioPAX or PSI-MI⁷¹ standard formats, providing a convenient point of access for public biological interaction and pathway information.

Analysis of network features

Execution of jActiveModules on the sample network with the expression condition gal4RG produces five modules (as specified by the default parameters), all having significant Z-scores > 3.0. The top module (shown in Fig. 5) contains the genes encoding the known galactose enzymes (GAL1, 7, 10), the galactose transporter (GAL2) and the galactose pathway transcriptional regulators (GAL4, 80, 3). Gal4, the central transcriptional activator of the pathway, is connected to other GAL genes through protein–DNA interactions (yellow), which are strongly differentially expressed since the GAL4 gene is deleted in this experiment. In this way, expression values superimposed on a protein network in the Cytoscape canvas can identify expression-activated pathways or protein complexes.

Execution of MCODE on the sample network produces 46 protein interaction clusters with scores ranging from 1 to 5.5, with the five top scores >2.5. Figure 6 displays one such cluster containing five PEX genes that encode components of a protein complex associated with the peroxisomal membrane. This set of proteins has been detected as a putative complex by MCODE based on its high level of enrichment for protein–protein interactions (seven observed interactions out of ten possible among five proteins).

The key regulatory “control points” of a biological system may not always be genes with the most pronounced expression ratios; more often, the most differentially expressed genes are in fact regulated by the controller, which lies upstream. An effective approach to finding upstream agents is to identify groups of genes that exhibit large measured changes under the experimental conditions and then to examine the genes adjacent to those in the network. When doing so, bear in mind that existing network data may be incomplete, and so one should not expect all connections between the causative agent(s) and the affected genes to be present in the network.

Evidence suggests that the activity of a complex may be controlled by restricting the production of one component, even if most are produced ubiquitously⁷². Thus, if one gene in the putative complex shows decreased expression levels, the complex may be less active under the given experimental conditions (although it is known that mRNA–protein levels do not always correlate). Similarly, signal-transduction pathways are often branched with multiple alternatives, yet all that is necessary for signaling is that the proteins in some branch are available and activated. The expression levels on each node suggest which alternatives may be produced under the given experimental conditions. In cases where the signaling activity can be measured, gene knockdown methods such as RNA interference can be used to verify these hypotheses experimentally.