In this section, we highlight some ofTVNViewer’s visualizations available for dynamic network analysis. In each case, we use the yeast or breast cancer data to show how an analyst would use TVNViewer to discover patterns and information in the recovered set of networks. After the demonstrations, we will discuss the results of using TVNViewer for dynamic network analysis.
One-level network circle view
An important challenge in dynamic network analysis is the recognition of subtle changes in the network topology over time. In the one-level network circle view, the analyst sees all the genes in the dataset aligned in a circle layout. The genes are represented by circles (nodes) and the connections between genes are represented by edges (lines between nodes). The genes areclustered to minimize the number of edges going across the circle, keeping most edges local to tight clusters of genes around the edge of the circle. Genes are colored by this clustering; details are provided in the online documentation describing how this is done. Also, the analyst can use the tree view to view the sorting tree of how the nodes were clustered.
In the one-level network circle view, the analyst can step through the sequence of networks in real time to explore the rewiring of the gene networks. We demonstrate this feature in Figure 2, where we show a subnetwork of genes at 24 time points from a large network derived from yeast gene expression data. The top graph in Figure 2 represents the gene network at Time 1, and all nodes are labeled by the names of the genes they represent. To enhance the figure’s readability, we have utilized TVNViewer’s option to remove gene name labels in the graphs representing the other time points. The 24 time points in this figure represent two cell cycles where the first occurs between time point 1 and 12 and the second occurs between time point 13 and 24. The one-level gene network view in TVNViewer makes the appearance and disappearance of edges in the network readily accessible to the analyst, without the awkward integration or customization required by other network visualization tools. The analyst can quickly identify that this particular network is active in the beginning of each cell cycle which corresponds to the G1 phase of the cell cycle.
Two-level network with GO annotations
Often, there are more genes in the network than can be visualized by using circle view. In this case, it is more helpful to group similar genes by function (i.e. gene ontology (GO) groups) and then visualize the interactions amongst the groups. TVNViewer provides a two-level network view specifically designed to allow high level exploration of the network at the group level, while still being able to zoom in to explore individual gene interactions. Consider analyzing a T4 malignant breast cancer cell network with 5440 genes (nodes), generated using Treegl [15]. A two-level network view using second level GO biological process groups is shown in Figure 3A. One can zoom in on a specific group, such as “necrosis” (Figure 3B), revealing the genes associated with that group. The analyst can zoom even further by selecting a particular gene to reveal its specific interactions. For example Figure 3C shows that the TUBB gene (tubulin beta) interacts with genes from many groups, most notably the signaling process and biological adhesion groups. This makes sense since TUBB encodes proteins that are important to GTP binding and GTPase activityin addition to its involvement in the structure of the cytoskeleton. Thus, the two-level view provides the analyst with both a high level perspective of the networks while simultaneously allowing him to focus on particular genes.
Directed graphs
TVNViewer can be used to visualize both directed and undirected graphs. Directed graphs are valuable if an analyst is interested in cases where the direction of the edge is significant, such as in a regulatory cascade. The initial layout of the graph is not changed in the case of directed graphs for the circle and force views. However, as the analyst hovers over different genes, TVNViewer will highlight all of the gene’s in-edges in red, out-edges in green, and bidirectional edges in cyan. If an analyst is interested in one particular gene or gene group, he can select that particular node and TVNViewer will isolate that node and show only the genes connected to it. For example, in Figure 4A, we have selected MIG1 in the yeast dataset; all the edges connected to it are highlighted in red indicating that they are in-edges, implying that they may have a regulatory relationship with MIG1. However, in Figure 4B, the selected node INO4 has only out-degree nodes since the edges connected to it are green. This suggests that these genes may be regulated by INO4. These regulatory relationships may change across time or space, and the analyst can use TVNViewer to trace these relationships using directional information.
Stack view
While the circle layouts allow analysts to understand how gene networks rewire over time or space, the stack view visualization is better fit for exploring how specific interactions between genes or gene ontology groups change over time. For instance, we would like to be able to view how the biological functions of the network change over time, such as over the course of a cell cycle. This can be done by grouping the genes by their GO functional group to visualize with the stack view (Figure 5). In this view, the out-degree of each GO category is stacked, one on top of the other. Thus, the variation in individual GO categories is clear, and the overall variation in out-degree is emphasized. This visualization clearly shows that the overall network is active during G1 phase and we observe that genes in the GO categories: ATP binding, electron transport chains, and phospholipase C activity are especially active. This is expected as these are all functions involved in cellular respiration, which is the signature activity of the G1 phase of the cell cycle. By hovering over the GO category in the stack; both the GO category and its degree at the given time point is displayed.
The analyst can select specific GO categories of interest by listing them using the filter box, or by simply selecting them on the plot. Additionally, if the analyst is interested in specific genes, he can go past the group level and generate stack plots of genes of interest. Although it is relatively simple to implement selection and filtering functions in a visualization, the impact provided by these features is substantial. By allowing analysts to rapidly and simply subset their data while highlighting items of interest, we allow analysts to play “what if” scenarios, which may combine a number of highlights or filters. These visualization features, comparable to dynamic queries, drastically lower the cost of exploring and experimenting with the data and evaluating the outcome of varying queries in comparison to database queries or other approaches [32]. In Figure 6, we use filters on the stack view to show recurring GO groups including electron carrier activity, alcohol dehydrogenase (NADH) activity, and various enzymatic processes. Figure 7 shows that these groups are active between time points 1–8 and 14–19. The timing is consistent with G1-phase which occurs at the beginning of each cell cycle. This observation is expected biologically; we expect that the cell is growing during G1, and thus cellular respiration, which requires electron carrier activity, and NADH activity, and other enzymatic activities are occurring.
Analysis of temporally dependent gene-gene interactions across the yeastcell cycle
Budding yeast (Saccharomyces cerevisiae) serves as an excellent model for dynamic network learning because the molecular mechanisms of the cell cycle control system are well known [33]. Budding yeast follows the eukaryotic cell cycle, which is divided into 4 distinct phases [34]. The first is G1-phase (gap 1), which is the interval between mitosis and DNA synthesis where the cell is actively growing. This is followed by S-phase (synthesis) during which DNA replication occurs. The cell continues to grow during G2 (gap 2) and then divides in the M or mitosis phase. For the purpose of this study, we group the G2 and M phase and refer to it as G2M.
Studying the yeast cell cycle is a fitting scenario for utilizing TVNViewer as both an exploratory tool and a method of validation.We first generate a series of networks across time from yeast gene expression data using TV-DBN [12]. Then we select subnetworks that are active during certain cell cycle phases and observe their temporal activity as it relates to their function. For example, Figure 7 shows a network with genes that were found to be active during the G2M-phase. Here, we observe functional groups that are clearly relevant to M-phase such as chromosome segregation, mitotic spindle elongation, and telomere maintenance. In addition, we observe GO groups like DNA repair, recombinational repair, and response to DNA damage stimulus which are indicative of G2-phase. One of the major checkpoints occurs in G2 phase, whereby cells are arrested in response to damaged or unreplicated DNA [34]. Thus, we can conclude that these functions are aligned with what we expect from genes that are active in G2M.
An important characteristic of cell cycle data is that it is repetitive. Thus, we should observe recurring patterns in the time-varying networks. Figure 8 shows a set of genes, active in S-phase. The colored layers of plots clearly indicate that the interactions between the genes repeat over the two cell cycles; the first cell cycle occurs between time points 1–12 and the second during time points 13–24. If we take the same subnetwork shown in Figure 8 and annotate the genes using GO functional groups, we can observe which groups are active over the time series (Figure 9). Similar to Figure 10, the colored layers show the GO groups repeat across the two cell cycles. The GO terms listed are also relevant to S-phase as they indicate the presence of genes involved in DNA binding, helicase activity and ATP binding.
From this preliminary overview of the functional significance of the genes provided by TVNViewer, we can then focus on particular genes and investigate supporting biological literature that can both confirm and explain why these genes interact. For instance, the gene HMI1 was found to be a DNA helicase and experimental results indicated that it localized in the mitochondria and was required for the maintenance of the functional mitochondrial genome [35]. The unwinding activity of the helicase requires ATP hydrolysis and has a 3′ to 5′ polarity [36]. Another gene in the subnetwork is YNL208W. While not much is known about the function ofYNL208W, the protein was detected in purified mitochondria [37]. Interestingly, experimental evidence places both HMI1 and YNL208W at the same cellular location, supporting the prediction by our network that these genes interact.
Studying developmental processes such as the yeast cell cycle requires the integration of temporal and functional information. By using TVNViewer, we identify the recurring patterns of the gene subnetworks in S-phase.We also find that the functional roles of the genes in the network are consistent with the timing of network activity. This analysis canguide the exploration of biological literature to link the gene-gene interactions and formulate a summarizing regulatory mechanism.
Exploring the progression and reversal of breast cancer
Using TVNViewer, we also investigate the progression and reversion of breast cancer cells using dynamic network analysis. Functional analysis of 3D culture models of breast cancer has led to a deeper understanding of the effect of a cell’s microenvironment on tumorgenesis and metastasis [30]. It was found that micro-environmental factors and signaling pathways have a dramatic influence on the growth dynamics and malignancy of the cells [38, 39]. Furthermore, treatment with inhibitors of various signaling molecules causes reversion of T4 cells into morphologically-normal-looking cells (T4R cells). Our objective is to analyze the functional differences amongst the different cell states.
We first used Treegl [15] to reverse engineer gene networks for each cell state. As shown in Figure 11, compared to S1 cells, T4 cells display increased activities in cell proliferation and locomotion, both of which are indicative of cancer. Furthermore, we see that that the T4 network exhibits significantly more interaction with the extracellular matrix and other components related to the cell membrane such as the vesicle (Figure 12). This is expected since it has been found that a cell’s interaction with its microenvironment affects tumorgenicity and metastasis [40]. Finally, one can see that the T4 network also displays increased signal transducer activity (Figure 10). Signal transducers and activators of transcription, especially those associated with cytokine and growth factor activity have been implicated in tumorigenesis [41].
As we can readily observe from the figures, the T4R cells are different from the S1 and T4 cells, but are also distinct from each other. The MMP-T4R network is very sparse and thus has few interactions. Notably,cell proliferation and other indicators of cancer are absent in MMP-T4R cells. On the other hand, the PI3K-MAPKK-T4R cells still display considerable cell proliferation and interaction with the extracellular matrix. PI3K-MAPKK –T4R cells also exhibit more activity such as tetrapyrole binding, demethylase activity and carbohydrate binding, all of which are absent in the other cell states. Collectively, these data suggest that although T4 cells can be morphologically reverted back to the normal-looking T4R cells, the underlying molecular mechanisms in the reverted cells are different from those in either S1 or T4 cells and from one another.