MONGKIE¶

Import a GBM-altered network¶

You will import a GBM-altered network from 2 CSV files for nodes and edges,

• Select File ‣ Import ‣ Comma-Separated Values (CSV), now a wizard window will guide you to remaining steps.

• Choose tcga_gbm_vertices.csv, and set ID column to name, Label column to hgnc_symbol, then click the Next button.

• Choose tcga_gbm_edges.csv, and set Source column to from, Target column to to, then click the Finish button.

• A report dialog finally shows the summary of the imported graph, including number of nodes and edges, type of graph, issues occurred during the importing process etc. Click the OK button.

• The imported network looks like below:

Data-to-Visual mapping¶

Data-to-Visual mapping allows you to map data attributes of nodes or edges to various visual styles.

• (Optional) Before proceeding to further steps, change the default font for nodes to Sony Sketch EF that was used in figures in the Case Study.

  Note

  To use the Sony Sketch EF font, you should download and install it on your system. It is available at here.

  • To open the Display Options panel, click the button at the bottom-right of a network display window.
  • Select the Nodes tab, and click the Font name button, and choose a font name, style, and size, then click the OK button.
  • To close the Display Options panel, click the button at the top-right of the panel.

  

• To map the alteration frequencies of genes to the sizes of nodes,

  1. Select the Visual Mapping window at the top-left of main application.
  2. Select the Nodes tab button.
  3. Select the icon.
  4. In the drop-down list, choose the Freq data attribute to map.
  5. Click the Spline... button to open the Spline Editor.
  6. Choose a pre-defined scale function.
  7. Adjust the transform scale to make more frequently altered genes be magnified.
  8. Close the Spline Editor.
  9. Click the Apply button.

  

• Now the network looks like the following:

• Now you will map a different shape to the linker nodes using discrete Data-to-Visual mapping.

  1. Select the Discrete tab at the bottom of Visual Mapping window.
  2. In the drop-down list, choose the Type data attribute of Nodes to map.
  3. Expand the LINKER, now various editors for visual styles will be shown.
  4. In the Shape filed, click the ... button to show up the node shape chooser.
  5. Choose the Diamond shape, and click the OK button.

  

• Now the network looks like the following:

• You can map expression correlations to the thicknesses of edges.

  1. Select the Continuous tab at the bottom of Visual Mapping window.
  2. Select the Edges tab button.
  3. Select the icon.
  4. In the drop-down list, choose the i.weight data attribute to map.
  5. Set the Max value to 10.
  6. Click the Spline... button to open the Spline Editor.
  7. Restore the scale function to the linear function.
  8. Click the Apply button.

  

• The final network is shown below:

Network clustering¶

You imported the GBM-altered network into MONGKIE, now you will apply the MCL Network clustering algorithm to identify network modules representing genes with topological proximity and correlated expression.

Hint

In the edge table, you can see that the i.weight column contains expression correlations.

MONGKIE internally assigns the values of i.weight column in a edge table to the weights of edges for clustering. Therfore, to give weights to edges in your network, you should set the name of the column to i.weight.

• In the Clustering window at the top-left of main application

  1. Choose the MCL algorithm in the drop-down list.
  2. Click the button to start the algorithm.
  3. After a little, the identified clusters will be listed.

  

• You can define these clusters as group nodes on the network.

  1. Select top 5 largest clusters by clicking Cluster 1 and Cluster 5 holding down the Shift key.
  2. Right-click on the selection will show up a pop-up menu. Click the Group menu item.

  

• Now you see that group nodes on the network are too large or a lot of overlapping with each other. You can use the force-directed algorithm to lay out them with more optimized size and position.

  1. In the Layout window at the bottom-left of main application, choose the Force Directed algorithm.
  2. Click the bottom to start the algorithm.
  3. Set Spring Coefficient to the minimu value 4.0E-5 by dragging the slider tick to the left-most.
  4. When being satisfied with the result, you can stop the running of algorithm by clicking the button.

  

• The final network looks like the following:

Make a new network from a cluster¶

In this section, you will create new networks from previously identified clusters. A new network consists of nodes and edges only in the cluster it derived from.

• Right-click on the group node for cluster 5 on the network, then a pop-up menu for the group node will be shown up.

• Click the New Graph menu item. You will see a new network created in a new tab.
• Repeat above steps on the cluster 3. Two networks derived from Cluster 5 and Cluster 3 are shown below:

Import expression log2FCs in normal vs. tumor¶

You will import gene expression log2FCs between normal vs. tumor condition in all TCGA GBM patients. and map them to the colors of nodes.

• Select the Network tab of previously created network from Cluster 3 in the original network.

• Select File ‣ Import ‣ Attributes from CSV File.

• In the Import Attributes dialog,

  1. Be sure Attributes for: Node to be checked.
  2. Choose a CSV file to import: tcga_gbm_exp_log2fc.csv
  3. Check The first line in a CSV file contains column names
  4. Key Column in Annotation File: Gene_Symbol
  5. Key Attribute in Network: hgnc_symbol
  6. Click the OK button.

  

• Now you can see a newly added column log2FC in the node table as shown in below:

• In the Visual Mapping (Nodes) window,

  1. Click icon.
  2. Choose log2FC in the drop-down list.
  3. Click to show up a palette chooser.
  4. Select (Red-Yellow-Blue).
  5. Click to map Red to the maximum and Blue to the minimum value.
  6. Check to show a list of node names with their mapped colors ranked by values.
  7. Click the Apply button.

  

• The final network of Cluster 3 is shown below:

• Repeat above steps on the network of Cluster 5.

Tip

How to add more gradation points and colors.

• You can add any number of gradation points by clicking desired positions on the color bar.
• Once you have a gradation point, you can edit its color by double-clicking the point.
• You can also move the points by dragging them.

Functional annotation of a cluster¶

In this section, you will perform a GO Over-representation analysis for genes in Cluster 3 and visually annotate the cluster with representative functions via the Visual Editor UI.

• Make sure you select the Network tab of Cluster 3.

• Select the Window ‣ Enrichment Analysis.

• Before going to remaining steps, you need to import symbol2uniprot_all.csv into the node table. The reason is that our in-house GO over-representation analysis tool expects IDs of gene set to be UniProt Accessions.

  1. Select File ‣ Import ‣ Attributes from CSV File as described in previous section.
  2. In the Import Attributes dialog,

  • Choose a CSV file to import: symbol2uniprot_all.csv
  • Key Column in Annotation File: HGNC.symbol
  • Key Attribute in Network: hgnc_symbol

  

• In the Enrichment Analysis window,

  1. Choose Gene Ontology in the drop-down list.
  2. Set Gene ID column to UniProt.Ac.
  3. Set the multiple testing Correction method to Bonferroni.
  4. Click the Run button.

  

• After a little, the result is shown in the Enrichment window at the bottom of main application.

  • You can see that GO terms of cell cycle checkpoint and DNA damage checkpoint are listed in first and second rows respectively.
  • Select the cell cycle checkpoint GO term, the click icon to see the details about the term.

  

• Next you will visually annotate the group node of Cluster 3 with two representative functions.

  Caution

  Now be sure to switch to the original network tab named tcga_gbm_edges.

  • Click the group node of cluster 3 in the original GBM-altered network.
  • In the Editor window,
  • Click the ... icon in Name property to edit a group name.
  • Enter DNA damage response and Cell Cycle.
  • Set the Font property to Droid Serif 80 Bold or what you want. (You can open the font chooser by clicking ... button in the Font property)

  

• The result is shown below:

Import expression profiles for 4 GBM subtypes¶

Now overlay average gene expressions in GBM patients of 4 subtypes onto the GBM-altered network.

• Select the original GBM-altered network tab named tcga_gbm_edges.
• Select Window ‣ Heat Map.

• In the just opened Heat Map window, click the Load heatmap data... label.

• In the Import dialog,

  1. Choose tcga_gbm_exp_subtypes.csv file to import.
  2. Make sure that Key column to map is hgnc_symbol.
  3. Be sure to check First line in the file contains header names.
  4. Click the OK button.

  

• Now your screen looks like the following:

• In the bottom panel of application,
  • Drag the slider tick to Mesenchymal to see average expressions in patients of mesenchymal subtype.
  • You can switch subtypes automatically to see animated pictures by clicking button.

Save as a image, CSV files, GraphML, and a VLG (Visualized Graph) file¶

• You can lauch export dialogs from File ‣ Export ‣ Graph Files... or Image Files....

• You can save the final visualization as one of supported image formats:

  • Use SVG (Scalable Vector Graphics) or EPS (Encapsulated PostScript) for vector graphics.
  • Use PNG, JPG, GIF, or BMP formats for bitmap images.

  

• You can export the graph to:

  1. A GraphML file that contains all graph elements in one formatted file, or

  

  2. Two CSV files, each stores list of nodes and edges in a tabular format respectively.

  

  3. A VLG (Visual Graph) file that stores all visualization-wide information as well as graph data.

  

Tip

CSV files for nodes or edges can be exported in the Data Table window.

Visual representation of biological entities and interactions¶

Graphical representations of biological networks includes how to intuitively visualize a set of connected nodes (or vertice) corresponding to biological entities, including genes, gene products (protein, transcript factor, miRNA, etc.), small molecules (compound, metabolite etc.), protein family and complex, and their links (or edges), such as physical or genetic interactions, regulatory events (transcriptional and translational activation or inhibition, phosphorylation, etc.), co-expression, shared protein domain, complex formation, trans-location and other biochemical reactions.

An example of visualization of biological networks¶

This illustrates how MONGKIE visually represents biological entities and relationships between them, by visualizing an example network, Glucocorticoid receptor regulatory network, which is a signaling pathway curated by the NCI-Nature PID (Schaefer et al., 2009).

MONGKIE provides the sophisticated data models for visualization of biological networks with advanced graph drawing techniques, and therefore can represent different types of biological entities and interactions between them with out-of-the-box visual styles, shown in Fig. 4.1. Both nodes and edges differ in their style according to their biological meaning. The style of nodes - e.g. label, font, shape, color, size and icon image - shows the type and state of biological components, and edges linking a relation participant with the information about the role also differ in their style - e.g. shape or thickness or color of lines, shape or color of arrows as well as label and font.

Pathway visualization¶

MONGKIE provides a built-in software module for pathway visualization, which supports visual analytics and exploratory studies of metabolic or signaling pathways in an interactive fashion tightly integrated with the human integrated pathway database, hiPathDB (Yu et al., 2012).

Like other types of biological interactions that can be modeled as regular binary-graphs, where interactions are between exactly two interactors, biological pathways also can be represented as a graph consisting of nodes and edges. Although these simple models have been shown to yield biological insights, biological pathways can be complex multi-modal or hyper-graphs (Saraiya et al., 2005), in which an edge could connect an arbitrary number of nodes or might connect a node to another edge. Therefore basic graph representations are incapable of modeling more complicated and interwoven biochemical reactions that involve more than two sub-components, e.g. the formation of protein complexes, and interactions that controlled by external controllers. In order to visually capture such complex biochemical events, MONGKIE presents an optimized pathway visualization based on the sophisticated visualization model taking into consideration these important domain-specific knowledges.

Visual models for the pathway visualization¶

A-1 Hierarchical modeling of protein complex assembly (orange border) can be flattened by splitting nodes representing complexes into their own individual members. A-2 Protein complexes (red nodes). A-3 Protein family (F icon). A-4 Dimer (D icon). A-5 Sub-cellular location (red icon). B-1 Multiple edges (gray and blue). B-2 Cellular states of nodes, mode of interactions of edges (activation, inhibition, phospholyration, etc.) B-3 Self interactions. B-4 Hyper-edges connect nodes to another edges (yellow), for example, one molecule (node) might prevent some other interaction (edge) from occurring.

Fig. 4.2 illustrates how diverse types of complex components and their relationships in biological pathways are visually represented in MONGKIE. A node representing a biological molecule in a pathway visualization may be either a protein, family, complex, dimer, enzyme, other small molecules (e.g. compound, metabolites), or especially a super-node which is a hierarchically decomposed composite node representing adjacency and inclusion relationships, e.g. the hierarchical modeling of protein complex assembly can be flattened by splitting nodes representing complexes into their own individual members (see Fig. 4.2 A). An edge in a pathway visualization represents a relationship or some form of interaction between nodes. The interaction could be one of many types: complex formation, activation, inhibition, aggregation, trans-location, catalysis, chemical modification, etc. In most cases, single-line connections are insufficient to capture the whole range of information contained in a biological pathway, because biological entities are often linked by more than one type of relationship. In such cases, multi-edge networks offer the possibility to link two entities by multiple edges, in which every edge having a different meaning. Also, hyper-edge connects a node to another edge, e.g. an inhibitory interaction (edge) actually indicates a biological process by which one molecule (node) might prevent some other interaction (edge) from occurring (see Fig. 4.2 B).

All theses different types of nodes and edges in the pathway visualization are visually represented with their distinct visual styles and additional informations, including their sub-cellular location, cellular state - e.g. activated, inhibited, phospholyrated, etc. Each component of pathways have informations about the originating data source, and hyper-links to the corresponding web pages is presented at the table view to link relevant databases - eg. NCBI, PubChem, etc. The force-directed layout algorithm (Frick et al., 1999) is optimized for the virtually automatic placement of components in the pathways.

Visual editor UI and Data-to-Visual mapping¶

Every component - nodes, edges, and groups - of the visualization model has a set of visual properties, including label text, text font, stroke, shape, size, color, line shape and width etc. These visual styles of them can be fully customized individually through the integrated Visual Editor UI which allows the user to edit them in any way the user desires with numerous predefined palettes. Actual UIs for editing visual properties of nodes, edges, and groups are shown in Fig. 4.3.

Visual Editor UIs¶

Visual editor UIs for (A) nodes, (B) edges, and (C) groups with their members.

Each component in the network can have associated data attributes possibly describing visual properties of them. In addition to selective and manual editing of visual styles, MONKGIE also provides a very useful way to automatically set visual aspects of components based on their data attributes. This continuous or discrete Data-to-Visual attribute mapping allows researchers to synoptically view multiple types of data in a network context (see Fig. 4.4). Data-to-Visual mapping lets the user, for example, load omics data from various high-throughput experiments, e.g. expression profiles, and visually project them into the network by automatically transforming data to some graphical attributes, e.g. color, size, visibility etc.

Continuous and discrete Data-to-Visual Mapping¶

Using an example network that are modeling cross-talks between biological pathways, (A) the number of genes in each pathway was continuously mapped to visual properties (size and color) of the corresponding pathway node. (B) data sources (KEGG or BioCarta) of each pathway were discretely mapped to node colors.

Exploring network¶

In order to substantially facilitate network navigation and information extraction, MONGKIE provides sophisticated options to explore networks in highly interactive ways, including searching, filtering, grouping, manual or automatic node selection, highlight, dragging, zoom and panning the display, overview of the complete network, and lastly data table view displaying attributes of nodes and edges in a tabular format.

MONGKIE provides easy-to-use search functions for the loaded network. One can enter any keyword or regular expression to search all data attributes held in the nodes and edges. The matching nodes or edges immediately highlighted in the visualization; and selection in the visualization propagates to the selection of the relevant rows in the data table scrolling to them, and vice versa. Furthermore, the network can be filtered down to interactions meeting a given filtering constraint according to their attributes. For instance, the network may be filtered to show only proteins occurring in particular locations, thus reducing the network complexity and restricting the one’s attention only to interactions within a given sub-cellular location, and this could greatly improve the visual perception of complex biological networks.

Given the large and complex networks, one common approach to interpret and visualize such networks is trying to display the complete network on the screen and providing functionalities to zoom, pan, and overview of the network for exploration. Like many other network visualization tools, MONGKIE provides these basic techniques for network navigation too. However, as the size and complexity of interactions grow, it is increasingly impossible to understand underlying structures and extract biological insights from such huge networks just using those basic navigation techniques. Another improved strategy is to dissect the complete network into smaller sub-networks that are manageable, biologically significant regions that can be understandable (Gehlenborg et al., 2010). These sub-networks are typically defined as, for example, sets of proteins that are occurring at the same sub-cellular location, or that belong to similar functional GO terms, or that are members of a densely connected cluster identified through the established network clustering methods. The resultant sub-networks are typically of a size that is more amenable to visualization and analysis.

In order to specifically support of these processes of dissection, MONGKIE provides a variety of ways to define groups of functionally or topologically related nodes, including enrichment analysis for functional modules (See Over-representation analysis), clustering analysis for topological clusters (See Network clustering), grouping by manual selection, or automatic partitioning of nodes according to their attributes. The defined groups or sub-networks are visually represented with distinct styles and importantly laid out separately from other parts in a way that automatically attracts each other nodes in a group while repelling other groups (See Fig. 5.1), and this geometric separation is essential to focus only on particular groups without being disturbed by unnecessary or noisy interactions. Additionally, one can create a new visualization of each sub-network, then analysis it independently from original one from which it derived.

MONGKIE is developed specially with these grouping functionalities in mind as one of main development goals so that it can be used to help dissect large interactions, thus reducing the overall complexity, and focusing on smaller but biologically interesting parts to gain biological insights without being overwhelmed by complexity and noises in the network.

Graph layouts¶

Graph layout algorithms are used to place graph nodes and edges in various geometric distribution for the clarity and readability of networks such that the number of edges crossing minimized and that the layout represents the overall structure of the network legibly.

A variety of state-of-the-art layout algorithms are implemented in MONGKIE, including Circular, Grid, Fruchterman-Reingold (Thomas et al., 1991), Radial Tree (Book et al., 2001), and Force-Directed (Frick et al., 19999) layouts, both for efficiency and quality (see Fig. 4.5).

Graph Layouts¶

A variety of state-of-the-art layout algorithms are implemented in MONGKIE, including Circular, Grid, Fruchterman-Reingold, Radial Tree, and Force-Directed.

Like most of other typical network visualization softwares, we try a force-directed layout first because this layout can usually well organize most biological networks based on the non-deterministic algorithm that lets forces between nodes influence the position of the node in the network. All nodes exert repulsive force on the others whereas connected nodes are attracted to each other. After several iterations in which the positions are adjusted according to the calculated force, the layout stabilizes, keeping edge-crossings to a minimum (Herman et al., 2000 and Heer et al. 2005). It also visually animate the process for laying out the network so that one can watch nodes in the network incrementally being placed in optimum positions and can terminate the algorithm when a good layout is obtained. However, this layout quickly becomes inadequate if the size and complexity of network are too larger to handle and interact. For such cases, MONGKIE provides an opportunity to go without animations or to choose other faster but simpler one - e.g. circular, grid etc.

UI for controlling layouts¶

Each layout algorithm can be easily started, canceled, and customized through the unified layout control UI shown in Fig. 4.6. Highly configurable layout algorithms also allow the user to change layout settings in real-time, and therefore dramatically increase user feedback and experience. For instance, settings of the force-directed layout, including gravity, spring and forces, can be configured and immediately applied even while the algorithm is running.

In addition to automatic layout algorithms, MONGKIE offers another way to interactively change the layout of the network by manually dragging each node or user-defined groups into any positions, and this is very useful in fine-tuning the automatic layout or emphasizing important nodes or biologically significant regions in the network by geometrically separating them from other parts.

Network clustering¶

Fundamentally network-based approaches in systems biology are based on the hypothesis that biological entities rarely act alone in the cell, instead they interact spatiotemporally with others, forming modules, in order to perform specific cellular functions (Hartwell et al., 1999 and Alberts et al., 1998). Graph-based clustering algorithms that are mainly developed in graph theory and recently computational systems biology have been successfully applied to the study of detection of protein complexes (Krogan et al., 2006) or families (Enright et al., 2002), or identification of functions of uncharacterized proteins (Bader et al.,2002), or extraction of co-expressed clusters from co-expression networks (Lee et al., 2004), etc., and shown to obtain good performances for extracting such modules from a variety of biological interactions.

These so-called ‘network clustering’ methods are also known to be less susceptible to inherent false-positives and more accurately predict modules made up of functionally relate nodes rather than conventional iterative pair-wise clustering, where individual relationship - e.g. sequence similarity, co-expression - between two biological entities is investigated without considering structural patterns in their interactions with neighbors (Freeman et al., 2007).

MONGKIE currently incorporates two popular structure-based network clustering algorithms, including MCODE (Molecular COmplex DEtection algorithm), MCL (Markov CLustering algorithm), and these make it easy to find densely inter-connected, thus functionally related nodes in biological interactions, exploit both local and also global structural patterns, and visually map them onto the network. MCODE is a graph theoretic clustering algorithm for finding molecular complex in large protein interaction networks (Bader and Hogue, 2003). The MCODE plugin, which is implemented by porting from the pre-existing plug-in in Cytoscape (Shannon et al., 2003), identifies clusters by finding regions of significant local density. MCL is a fast and scalable unsupervised clustering algorithm for graphs based on simulation of the flow in the graph (Van Dongen, 2000). Because MCL is a robust state-of-the-art general purpose clustering algorithm for large graphs, it can be applied to any complex biological networks, e.g. protein functional relationship network to look for candidate cancer driver mutations and relevant functional modules they belong to (Wu et al., 2010).

Network Clustering¶

This demonstrates the procedure of identification network clusters, organization them onto the network with distinct visual schemes, and how nodes in the same cluster are laid out coherently.

One can define resultant network clusters (or modules) as groups, therefore they, as described in Exploring network, can be visually organized and laid out onto the network with distinct visual schemes, as well as displayed in a tabular format. As shown in Fig. 5.1, each cluster and its members are visualized with a distinct color and shape according to their cluster membership, also laid out using the optimized force-directed layout algorithm that automatically attracts each member in a cluster while repelling other clusters. This helps users to visually interpret the coherence of clusters in the context of the network, that cannot be easily obtained by simply examining lists of clusters or their membership. Cluster nodes in the network can then be manipulated just like other general nodes for any exploratory purposes, e.g. one can select, then drag them in order to place in desired positions. One also can create a new visualization for a sub-network made up of nodes and edges in a cluster.

These features of network clustering and grouping in MONGKIE can facilitate to analyze or visualize the large data set of biological interactions in a more modular way that can provide biological insight into both local and also global structures in networks between biologically related nodes - e.g. a pathway affected by mutated genes, proteins belonging to a same complex or family, or cross-talk among biological pathways etc.

Gene expression overlay¶

Gene expression data obtained by micro-array experiments or RNA-seq techniques can provide powerful insights into underlying cellular states and dynamics when they are well integrated into the context of biological networks. Therefore, overlaying expression profile data onto a visualized network is an essential way in identifying a set of genes or proteins that share a related pattern of expression under a particular condition - eg. co-regulated gene sets and their interactions in a certain disease sample - or capture dynamic changes of their expression levels over a range of time points or different conditions (Gehlenborg et al., 2010). For these analyses, we require tools that allow users to visually represent expression profiles of the nodes in the network according to their expression level, thus better perceive the dynamic mechanism of a underlying biological system being guided by visualized expression patterns or changes.

MONGKIE provides powerful functionalities for visual analyses of high-throughput omics data in the context of networks, in particular for the gene expression data analysis of time series or multiple conditions. It supports dynamic visual representations, including a color gradient, size, and label of relevant nodes, that are easy to separate and interpret independently in order to depict the corresponding expression profile ratio of gene or protein nodes. Once expression profiles are imported as data attributes of nodes using the CSV file format, then a heat map visualization, which is used in a wide range of tools for the process of gene expression visualization (Gehlenborg et al., 2010) appears to display those expression data as a ratio-based graphical matrix, as well as incorporated within the graphical representation of nodes in the network (see Fig. 5.2).

An important challenge for expression data analysis is to interpret gene expression data produced from more than one condition, for example, time series experiments, or multiple perturbation studies. Therefore, it is necessary to consider all time points or conditions in order to detect temporal patterns and their changes in gene expression profiles whose values vary over time or different condition. This requires a selective or sequential visualization of multiple expression levels in the network context. MONGKIE allows users to incorporate such dynamics of gene expression profiles into the loaded network visualization by offering a way to change the visual mapping - e.g. color, size - of nodes to reflect the expression levels of a particular time point or condition according to the user selection.

Overlaying of multiple expression profiles¶

This demonstrates the procedure of overlaying of gene expression data from multiple experiments onto genes in the network, and the UI for capturing dynamic changes of their expression levels over a range of different experiments.

Fig. 5.2 illustrates an example of this process, where gene expression profiles from multiple micro-array experiments in six cell lines were loaded into the heat map display, and also mapped on corresponding nodes in the network. A expression level of a particular experiment can be navigated using a sliding bar UI on the bottom of the window, and also the navigation process can be animated, as introduced in some tools (Hu et al., 2013 and Kincaid et al., 2008), by automatically switch to visualization of the next experiment with a predefined time interval. This is well suited to investigating by eye changes of expression levels within a group of interesting genes - e.g. genes that share a same functional term, members of a network cluster, or deferentially expressed genes - over given experiments. Furthermore, by arranging visualization windows in a grid (see User Interface), one can in parallel compare multiple visualizations of the same network, where each visualizes for its navigating experiment - known as the ‘small multiples’ approach (Gehlenborg et al., 2010).

Over-representation analysis¶

A group of interesting genes in the biological network, e.g. a module of potentially related genes has been found by the Network clustering, may be investigated to find biological pathways or other functional categories like GO (Gene Ontology Consortium, 2004) terms, where they are significantly over-represented. This approach, so-called ‘enrichment analysis’, is widely used in order to study the gene set for their over-representation in certain annotation classes that are usually related to biological functions of those genes (Huang et al., 2009).

GO over-representation analysis¶

This demonstrates the procedure of enrichment analysis, and visual annotation of relevant nodes or regions (e.g. clusters) in the network with significantly over-represented GO terms.

MONGKIE provides a pipeline for this analysis, shown in Fig. 5.3, where researchers perform a statistical test for enrichment or depletion using the GO categories, in order to identify over-represented functional terms with statistically significance from the selected set of interesting genes in the certain network region. The result of analysis is displayed in table views with both a list of resultant functional terms and also statistics of the enrichment analysis, including number of enriched terms, number of population genes, number of query genes, and detailed information about each annotation term. In the result table, each row of resultant term has a background color that is mapped to its corresponding p-value - that was calculated using the hyper-geometric testing, and might be adjusted by one of following multiple testing correction methods: Bonferroni, Bonferroni-Holm, Benjamini-Hochberg, or Benjamini-Yekutieli - and the set of resultant term can be reduced based on user’s cut-offs and ranked by their statistical p-values.

Another optimized way to show the result of enriched functional terms in the hierarchical structure, e.g. Gene Ontology, is to display them in a tree table. In the tree table view, intermediate terms for tree hierarchy are automatically included, and users can expand or collapse any sub-tree in the view. Therefore it allows the user to interactively investigate terms according to specific biological complexity within the hierarchical structure as well as their significance.

Once a list of resultant term is displayed in a tabular format, users can select significant terms of interest, then visualize the group of relevant nodes or regions annotated with those terms in the network context by mapping distinguishable visual aspects to them. This visual mapping of functional terms in the context of the network has the distinct advantage of allowing the user to quickly identify by eye both biological functions of certain parts and also higher-order interactions between those parts in the network that would not be obvious without this type of visual representation.

hiPathDB¶

With heterogeneous biological pathway data sets in the diversity of potential data formats available, the integration of pathway resources has become an important issue in utilizing these resources systematically and efficiently. In order to utilize these resources to interpret and analyze genomic data using network-based analysis methods, the information stored in dispersed data repositories needs to be linked and combined in efficient ways, and strongly required to unify heterogeneous interactions in different pathway data sources into the one general network model.

Here, MONGKIE provides a built-in software module for highly interactive pathway visualization and exploration, tightly integrated with hiPathDB (Yu et al., 2012). hiPathDB is an integrated pathway database that combines curated human pathway databases of NCI-Nature PID (Schaefer et al., 2009), Reactome (Croft et al., 2010), BioCarta (Nishimura, 2001) and KEGG (Kanehisa and Goto, 2000). hiPathDB provides two different types of integration. One is the pathway-level integration - shown in Figure 3 of Yu et al., 2012 - which is conceptually a simple collection of individual pathways, was achieved by devising an elaborate model that takes distinct features of four databases into account, and subsequently reformatting all pathways in accordance with our model - shown in Figure 2 of Yu et al., 2012 - while maintaining molecular details of signaling processes. Another is the entity-level integration - shown in Figure 4 of Yu et al., 2012 - creates a single unified pathway, super-pathway, that encompasses all pathways by merging common components. Even though the detailed molecular-level information such as complex formation or post-translational modifications tends to be lost in the entity-level integration, such integration makes it possible to investigate signaling network over the entire pathways and allows identification of pathway cross-talks. Therefore, the unified super-pathway achieved by entity level integration facilitates the network-based analysis and navigation from the perspective of biological pathways.

Interaction Manager¶

In addition to biological pathways, there are other types of biological interactions that are well modeled by simple binary, pair-wise graphs, such as PPIs, TF-target, miRNA-target and genetic interactions, and they are also essential for the network analysis of genomic data. MONGKIE provides a elegant way to use these types of network as background networks for the exploratory network-based analysis and visualization, through the Interaction Manager UI, shown in Fig. 6.1 A.

Interaction Manager¶

(A) Interaction Manager UI. (B) demonstrates the procedure of dynamic network construction starting with a small part of genes of interest (orange nodes).

Imported and managed interaction data sources can be utilized for various exploratory and analytical purposes, such as network generation from a scratch by search query, dynamic network exploration, expansion and filtering of interactions by their sources. For example, there are cases when the size of a interaction network is so huge that it is impossible to handle and visualize the complete network at once. It is recommended for such situations to follow the approach of a classical top-down exploration, in which rather than display the entire network in one display, initially start with a small part of nodes in the network, e.g. deferentially expressed genes, and then iteratively build a larger network by allowing the user to successively expand particular nodes of interest with their further interactions and neighbors.

The procedure for this strategy of top-down exploration provided by the Interaction Manager is shown in Fig. 6.1 B. It allows the user to expand interactions through the context-sensitive right-click menus on a selected nodes of interest (orange nodes), as well as just to connect existing nodes using the interaction dataset from a data source by selecting the check-box of that source in the interaction manager UI. Later, the user may delete neighbors out of the interest to reduce complexities of the network such that he/she focuses on the region of interest. Furthermore, each newly added node resulting from a expansion action is placed in an appropriate position with an animation, well incorporated into the force-directed layout algorithm (Frick et al., 1999, Heer et al., 2005), and this allows the user to easily preserve the so-called mental map (Misue et al., 1995) during exploration of the network.

Users can import their own binary interaction dataset into the interaction manager from files using standard formats - e.g. GraphML and CSV files. Additionally, visual styles of each interaction source can be fully customized through the Visual Style editor, and they will be persisted, the latter means that customized visual properties of edges from the interaction source as well as imported interaction dataset themselves is locally stored and will be available for styling edges on the next run of the application.

Plug-in architecture¶

MONGKIE is a java-based application built on top of the NetBeans Rich Client Platform that supports plug-in programming architecture, thus it is easy to implement various new plug-ins with additional functionalities. An overview of its modular architecture is given in Fig. 9.1.

Schematic overview of the plug-in architecture implemented in MONGKIE¶

Based on its extensible architecture, MONGKIE provides core APIs (Application Programming Interfaces), SPIs (Service Provider Interfaces), and UI widgets for the base functionalities, such as graph visualization, network analysis, data integration, and many out-of-the-box supports that enable to build your own plug-ins onto the platform.

For example, The software module for Graph layouts provides well-defined APIs, SPIs and UI components that can be utilized by plug-in developers (See Fig. 9.2 A). Therefore, if you want to add a new layout algorithm into the MONGKIE, you only need to implement the logic of the layout algorithm without having to worry about other things like UI components, program states, window management, event handling, and data persistence and so on. All of these fundamental features for developing plug-ins are provided out of the box. This approach can allow a great deal of flexibility in the building various improvements of existing modules as well as the introducing of new functionalities or tools.

Development and Management of Plug-ins¶

(A) an example of developing a layout plug-in implementing a new algorithm. (B) UI for plug-ins management.

MONGKIE also provides a GUI (Graphical User Interface) shown in Fig. 9.2 B, Plug-in Manager, in order to facilitate the management of different plug-ins. Hence, users can install, update, remove, activate, or deactivate individual plug-ins through the integrated UI, this allows the customization of the application functionalities according to their needs.

RESTFul Web service API¶

The multi-tiered system is applied for abstracting remote data sources and separating control logics from the data and presentation layers to improve data integrity and accessibility. We put retrieval logics for the remote data at the middle-tier, which enables to control database connections and provide unified and maintainable data access. RESTFul Web Service APIs, which are at the middle-tier, are implemented using the JAX-RS technology - the Java API for RESTful (Representational State Transfer) Web Services. Based on these web APIs, MONGKIE provides the functionality that allow users to access our integrated data sources and services from outside of the platform in a programmatic fashion through any REST clients, including Java Applet, Python or Ruby scripts (See Fig. 9.1).

Overview of the 3-tier system implemented in hiPathDB (Yu et al., 2012)¶

(1) A REST client written for pathway visualization (blue colors), (2) RESTFul Web service at the middle tier (orange colors), (3) Relational database backend.

As an example, the implementation and usage of the hiPathDB (Yu et al., 2012) RESTful Web Service API is shown in Fig. 9.3. hiPathDB APIs allow researchers to retrieve data from hiPathDB database by offering methods to get pathways given their names, or get member genes given a list of pathways, and some others.