Network construction.

Networks are generally represented within visualization applications as annotated graphs. Biological networks are represented as graphs, and defined by a set of vertices (often referred to as nodes) and edges, plus additional annotations that describe the properties of these objects. This model allows the use of many algorithms devoted to graph analysis. Networks can be obtained in several ways, generally beginning with a list of genes of interest (often derived from experimental data) and an interaction file defining global interactions from a larger set. From this point networks can be manually constructed—with the user assigning positions and node description characteristics, they can be automatically generated (for example, random networks), or could be formed by querying public databases (e.g., protein–protein interaction or pathway databases) to assign further node descriptions. To aid data interchange, several community standards, e.g., PSI-MI [1], BioPAX [2], SBML [3], GML, CML, CellML [4], and ASCII TAB delimited file have been defined. Some commonly used databases of interactions include Intact [5], BioGRID [6], HPRD [7], I2D [8], and iRefindex [9]. Some of these have been integrated under the IMEx consortium [10], and are accessible via the PSICQUIC standard [11]. The features of each of these data formats and sources vary, being extremely dependent on their individual purpose. For example, the C. elegans interactions in I2D ver. 1.95 (7,060 proteins/nodes and 57,224 interactions/edges) can be represented as a 3MB TAB delimited file, containing only graph structure, protein names, and interaction sources. The same data, represented in PSI-MI format, is structured to define many other features, such as database cross-references and experiment documentation, and requires 56MB. The time to load these different formats in NAViGaTOR is 10 and 54 seconds, respectively, as complex formats require more system resources within the visualization program. File formats must be carefully selected to serve the needs and competencies of the researcher.

Depending on the workflow and features of a program, it may be possible to merge networks and annotations from multiple sources, while considering context and confidence of individual interactions. Thus, tools supporting data interchange formats such as PSI-MI enable this loose program integration.

Due to the ambiguity of terms and many-to-many relationships to IDs, it is important to note that the accuracy and reproducibility of a representation is entirely dependent on the level of care with which its data was documented. Researchers should always document the source of their data, which nomenclature databases (and their versions) were used (UniProt release 2012_07, mirDIP v2.0, DrugBank v2.5, etc.), and what ID mapping methods have been used in instances where different sources were merged.

Layout.

Network layouts determine the positions of individual nodes in a coordinate space. There are several basic requirements that a good layout should satisfy. It should provide a biologically meaningful representation of the network while allowing “easy” interpretation by the viewer. Edges between nodes of interest should be easy to follow. This can be achieved by efficient use of screen space and minimization of edge crossings, and by optimal use of transparency and other visualization methods. Layouts can be created by manually or algorithmically placing nodes, or by some combination of these two approaches. For algorithmic layouts, scalability can be important, as biological graphs can often grow to encompass tens of thousands of nodes and millions of edges. Several effective algorithms, such as Graph Drawing with Intelligent Placement (GRIP) [12], have been proposed to address this requirement, as well as algorithms that take advantage of multi-core CPUs or GPUs. Even so, large biological graphs are sometimes impossible to lay out coherently even with the best algorithms, resulting in “hairball” layouts.

Visualizing node and edge attributes.

The annotations contained in a network can be represented using different visual attributes, such as color, size, and shape. Multiple combinations of attributes to visual feature mapping exist, but some are more suitable and easier to interpret. One has to consider continuous versus discrete attributes, dynamic ranges, and number of features and their values that need to be differentiated. The type of annotation and its visual attribute must also be appropriate: node size can effectively represent a numeric value, but might poorly represent different classes of object. For example, expression can be effectively represented by color, but may also be represented as discrete values with node shape. In both examples, using transparency can further improve interpretability of important parts in the network.

Analysis.

Systematic analysis of protein–protein interactions (PPIs) and other biological networks can uncover biologically relevant information. PPI networks have a strong structure-function relationship [13], which can be used to help interpret integrated data sets [14]–[16]. Beyond the different data that can be integrated to help visualize highly important nodes in the graph, the analysis can also involve graph features like node degree, node centrality, shortest paths, cliques, cycles in directed graphs, etc.

Diverse data sources and visualizations can offer different perspectives on the same biological data. There are many different metrics for evaluating a graph's aesthetics, such as path bend and edge crossing minimization [17]. However, these metrics often give different results when presented with the same graph, which leaves room for subjective decisions about individual layout tasks. Below, we explore alternative visualizations of the relation between aging and cancer to highlight some of the features discussed above.

Network construction.

First, we converted 65 human aging genes (downloaded from GenAge build 15, http://genomics.senescence.info/genes/) into 112 UniProt IDs and 475 human cancer genes (from the Sanger Cancer Gene Census, http://www.sanger.ac.uk/genetics/CGP/Census/, accessed Feb. 2012) into 857 UniProt IDs (Table S1) using the I2D web interface (http://ophid.utoronto.ca/ophidv2.201/). The mapping between different database IDs is never one-to-one, and this causes the differences in numbers between the genes and the UniProt IDs list. Next, we uploaded this list of 969 IDs into NAViGaTOR and retrieved known, publicly available human PPIs from I2D 1.95 (this step retrieves all connections that include at least one gene from the starting list). Next, we used the NAViGaTOR search function to select only those nodes corresponding to the cancer or aging gene lists (e.g., query proteins), and deleted all other nodes (we only wanted to consider connections between aging and cancer genes in our analysis). At this point, the network comprised 219 nodes and 1,195 edges (data available as a NAViGaTOR XML file in Dataset S1); the default visualization for this network is the infamous “hairball”. Further layout and analysis steps are needed to highlight the interesting features of this network. An interactive combination of analysis and visualization implements a visual data mining workflow that enables the best integration of user's needs and expertise.

Visualizing attributes.

(1) Using shape to distinguish node types. First, we defined two subsets in NAViGaTOR: aging proteins and cancer proteins. We used the set operation tool to create a third subset, defined as the intersection of the first two subsets, i.e., proteins known to be involved both in aging and in cancer. A different node shape has been manually assigned to each subset (Figure 1a and Figure S1). The group of shared genes comprised ATM, BRCA1, BUB1B, CEBPA, ERCC2, ERCC4, MSH2, TP53, and CHEK2. (2) Using node color to distinguish protein function. Using the I2D plugin, we downloaded the Gene Ontology (GO) [18] terms, PubMed references, and gene synonyms associated with each node. We decided to color the nodes according to the GO category to which they belong (Figure 1a), and to show the gene names of the nodes shared between the aging and cancer lists. We decided to display gene names because they are more interpretable by humans than database IDs, but any text feature associated to the nodes can be displayed as a label. Importantly, gene or protein IDs are best used to interrogate external databases, as they eliminate the ambiguity of name matching.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Network visualization of the query genes and their involvement in other tumor types.

a) Network built on aging and cancer genes. Labeled nodes belong to both gene lists. Square nodes represent cancer genes while diamonds represent aging genes. b) Deregulation of the network genes in lung (represented by down arrows) and ovarian cancer (represented by up arrows). The height and width of the nodes are proportional to the number of studies where the genes are deregulated. Node transparency corresponds to overall number of studies where the gene is deregulated. c) Network integrating chemical compounds targeting the query genes. Hexagonal nodes represent drugs. The names of the drugs interacting with the shared genes are shown. C: cancer genes, A: aging genes, D: drugs. C1, A1: genes interacting with shared ones, C2, A2: genes not interacting with the shared ones. D1: drugs targeting only aging genes, D2: drugs targeting both aging and cancer genes, D3: drugs targeting only cancer genes. Node colors represent GO categories as per legend. Edges are colored to differentiate inter- and intra-group interactions.

https://doi.org/10.1371/journal.pcbi.1002833.g001

Pathway enrichment analysis of the proteins in the network using the NAViGaTOR plugin showed that the top two pathways for the complete network as well as for the shared genes are DNA Repair (p<2.83⁻⁰⁹ and p<5.65⁻⁰⁶) and Cell Cycle (p<6.15⁻⁰⁹ and p<8.35⁻⁰⁵) (Figure S2). It has been described that the choice between cell survival and cell death represents the trade-off between cancer and aging, triggered by damage to the DNA [19]. These pathways are highly interconnected [20]; and thus, it is not surprising to identify them to be most represented in our cross-talk genes.

Layout.

We used a separated circular layout for the nodes shared between the cancer and aging datasets while the other nodes are distributed in circular layouts, including only cancer genes (C) or only aging genes (A) directly interacting with the shared ones (A1, C1) or not (A2, C2). The circular layout lets us appreciate the number of nodes for each category and the distribution of the visual features in the subset. As shown in Figure S3, these views can be further enhanced by sorting the circular view using GO, thus further highlighting the diversity or enrichment of individual GO functions.

Network construction.

We downloaded genes up- or down-regulated in non-small cell lung or ovarian cancer from CDIP, the Cancer Data Integration Portal (http://ophid.utoronto.ca/cdip/), a collection of gene expression data from published studies.

Visualizing attributes.

To represent up- and down-regulated genes we used node shape. We also related node height to the number of times the gene is aberrantly expressed in lung cancer studies and node width to the number of ovarian cancer studies (Figure 1b and Figure S4). Transparency shows the total number of times the gene was deregulated in the two cancers. Edge color was used to differentiate inter- and intra-group interactions, and node color corresponds to GO biological function.

Interpretation.

It is now apparent that the shared genes are more frequently aberrantly expressed in ovarian cancer than in non-small cell lung cancer. This implies the involvement of the DNA repair pathways in the two tumor types: they play a central role in ovarian carcinogenesis (all of the high penetrance ovarian cancer susceptibility genes identified so far play a role in DNA repair) [21], while in lung cancer they share causality with signaling pathways (especially those that promote growth) [22].

Network nodes and edges may represent other information besides physical PPIs, e.g., genetic interactions [23], metabolic reactions [24], gene regulations [25], and microRNA–target [26] and drug–target associations [27]. The next paragraph discusses an example. A second one is shown in Figure S5.

Network construction.

We downloaded lists of drugs that target the aging and cancer genes in our network from CTD, the Comparative Toxicogenomics Database (http://ctdbase.org/) [28]. We uploaded these drug–protein interactions into a new subset in NAViGaTOR.

Layout.

The nodes representing the drugs are distributed in sets close to the nodes category they interact with. For drugs interacting with the shared genes, we used a linear layout to show node names more clearly (Figure 1c and Figure S6).

Interpretation.

It is interesting to note that 99 compounds (23% of all compounds) target genes from both the cancer and aging subsets. The most connected node in this subset is pirinixic acid, a peroxisome proliferator activated receptor α (PPARα) agonist. PPARα and the genes under its control play a role in the evolution of oxidative stress excesses observed in aging, the activation of the receptor being involved in the restoration of the cellular redox balance [29]. Moreover, it has been demonstrated that several genes involved in cell cycle control and DNA repair pathways are induced upon activation of PPARα [30].