Here we introduce the design principles and implementation details employed by our visualization methods. We first present an overview of the visualization workflow we propose. We then provide details about each of its components.
3.1 Design Overview
Researchers analyze their data in the following workflow:
- import a model of a canonical pathway representation either by loading a signaling pathway image and preprocessing it to help the system infer the structure (, lower left) or by specifying the model explicitly by placing proteins and interactions on an empty canvas (, lower right);
Fig. 2 Structuring protein interactions around familiar canonical pathways provides intuitive visualizations. A canonical signaling pathway representation (top) can be imported into the system in two ways: on the lower left the pathway image itself is loaded (more ...)
- load one or more quantitative datasets ();
Fig. 3 Loading and visualizing quantitative proteomic data. (a) Selected-ion chromatogram of a peptide belonging to the FceRI gamma protein. (b) Line in a file containing experimental data ready to be loaded. (c) Software display of a heatmap representing the (more ...)
- automatically extract proteins and interactions from protein interaction databases such as HPRD and build a network around the pathway model specified in step 1 and the quantitative data from step 2;
- represent the network graphically using a novel canonical pathway-oriented layout ();
Fig. 4 Proteins and interactions from HPRD (small fonts) that are connected to the canonical pathway model are: (left) integrated directly into the signaling pathway image with one protein selected and its interactions highlighted; (right) structured around (more ...)
- explore and analyze the network guided by interesting features noted in the experimental data; investigate the network at interaction level using a Focus+Context technique; analyze how known information blends with the new experimental results using such features as clustering of quantitative proteomic data, filtering, highlighting, and information on demand ();
- derive insights or generate new hypotheses, design and run new experiments, and restart from step 2.
3.2 Pathway Model Specification
Our solution requires the user to specify, using a simple interface, the canonical pathway representation of the signaling pathway under investigation. This can be done either by putting proteins and interactions on an empty canvas or by using a pathway image that is preprocessed to help the system extract the pathway structure; the preprocessing entails drawing single, continuous strokes over or around each pathway element – proteins, interactions and other entities. These strokes aid the software in identifying image features () as detailed below.
If the stroke endpoints are far apart compared to the stroke length, the image feature is probably an interaction and the endpoints are matched against protein positions to find which proteins are involved. The interaction strokes snap to image features in a manner similar to a lasso tool. This is done in order to obtain the correct image region that the interaction is covering, for reasons described in Section 3.7.
If the stroke endpoints are close relative to the total length of the stroke, the feature-detection algorithm decides to classify the feature as a protein. It then computes an average color for the area enclosed by the stroke and removes all points dissimilar to it. In most cases this leaves only the image shape selected. The protein position on the canvas can be inferred through this computation.
If a selection is unsatisfactory the user can cancel it and try again – depending on the previous selection, the algorithm will attempt to correct the image-processing parameters for the second try. For instance, if the area selected by the user is much larger than that returned by the algorithm, the color-similarity threshold is increased.
Once the graphical model is complete, either by pathway processing or by pathway drawing, the placed proteins need to be linked to protein identifiers in the protein interaction database. The user chooses the correct protein by searching the database for keywords using a dedicated dialog box. In our test cases this process took between 15 and 30 minutes for medium pathways such as those in the figures, but these times vary with image complexity and user training.
3.3 Interaction Data
In our experimental prototype we use the HPRD protein interaction database. HPRD is a protein interaction and metadata source based on manual literature search. The database information is stored and loaded as flat files.
We have also experimented with the STRING interaction database, version 7.0. STRING searches multiple sources for evidence of protein-pair interactions: database occurrence (HPRD, KEGG, REACTOME), genomic context, coexpression, high-throughput experiments, and the literature. A score is computed for each source and aggregated into a number that quantifies the likelihood that a protein pair interacts. Due to STRING's unsupervized automatic parsing and computation, it has greater naming redundancy.
The network exploration paradigms defined here could be used with any protein interaction database. One of the main challenges in supporting a protein interaction database is providing access to useful metadata from other databases. This is due to the inherent difficulty of translating protein identifiers across independent protein databases.
3.4 Experimental Data
The quantitative proteomic data is loaded as XML or flat files upon pathway creation and can contain multiple quantitative data points as well as protein identifiers and other metadata. For graphical representation, the quantitative proteomic data are transformed into a colored heatmap representation () indicating fold changes of a given peptide across different experimental conditions (time course of receptor activation or comparison between wild type and mutant cells). The following color-coding is used: blue – decrease of proteomic quantity, yellow – increase of proteomic quantity, black – no change.
If multiple experimental files are loaded, as in a comparison between wild type and mutant cells, special types of heatmaps are computed for each pair of experiments to reflect changes between experiments: yellow then indicates a major change between the two experiments, while black corresponds to no change. A single protein can have multiple heatmaps, one for each assigned peptide. The heatmap icon appears in two places: displayed in the expanded network exploration upper plane, attached to proteins revealed in the experiment (), and in a dedicated panel on the right () containing all peptides discovered in an experiment.
For multiple quantitative data sets, the heatmap experimental data panel on the right () is configured to contain tabs not only for each separate experimental data-set but also for changes observed between pairs of data-sets. For instance, in a phosphoproteomic receptor activation timecourse experiment involving wild type and cells lacking critical signaling proteins, the heatmap tab contains one tab dedicated to timecourse phosphopeptide heatmaps in the wild type cell, another tab for the mutated cell, and a third tab displaying the fold change of individual phosphopeptides observed between the two cell types through the receptor activation timecourse. This feature can be particularly useful in knockout-type experiments since the differences in behavior between a normal and a mutated cell become evident immediately.
The experimental data panel is kept visible at all times so that researchers can use it to explore the new quantitative data systematically. The items in the experimental data panel can be used to start the exploration by linking directly to Focus+Context representation.
Using experimental data to guide exploration was also discussed in [2
]. Our work differs both in the way we present the information to the user and in the emphasis we put on comparative analysis of multiple experiments. Such analysis can also be performed with their system, but we believe the small multiple approach would overload the display if used with dense networks and large quantities of experimental data. Their parallel coordinates view was also not extended for both multiple time-points and multiple experiments.
3.5 Network Generation
From the user-provided pathway skeleton, the software constructs a protein-protein interaction network by loading proteins and interactions from the HPRD database. The network is grown iteratively in a breadth-first manner: first, proteins interacting directly with the canonical signaling pathway model are imported, and then in subsequent steps, proteins interacting with those added in the previous iteration are extracted from HPRD and included. Finally, interactions among all proteins are loaded.
The number of levels to grow the network and optional filters used to exclude proteins from the build process are specified by the user. However, growing the pathway from the user-specified proteins alone may leave experimental proteins outside the network. To ensure inclusion of all experimental proteins in the final visualization, we also grow the network from the experimental proteins themselves. This solution increases the chances of linking the experimental proteins to the pathway since two networks are grown simultaneously toward each other.
3.6 Computing Protein Positions
While the canonical pathway proteins have user-provided predefined positions, our prototype must compute where to put the proteins extracted from the interaction database. These proteins are placed depending on their distance, in terms of number of interactions, from each of the pathway proteins. If protein P is interacting directly with protein A and is three interactions away from protein B, it is placed on the line segment between A and B, closer to A. The distances are not necessarily directly proportional to the path lengths: they can be weighted so that direct connections are much shorter then longer interaction paths.
Essentially the nodes are placed at a path-length weighted barycenter of the pathway nodes. Barycenter positioning was also used in [7
] to place new nodes in relation to already existing ones in the context of evolving graph drawings. This algorithm produces positions close to those computed by a traditional spring layout algorithm, since a node is dragged by the edge springs to a similar location.
This methodology leads to identical positions for some proteins, however, and a force-directed approach based on [8
] is used to perturb the layout and remove overlaps; a simple linear grid approach is used to improve the performance of the layout algorithm by using vicinities to reduce the number of comparisons needed to compute forces on protein-nodes. We also apply a force to keep the nodes close to their initial position computed by barycentric placement.
The sizes of nodes are taken into consideration when computing repulsive forces. The aspect ratio of nodes in relation to the force vectors can also be taken into account so that forces are applied anisotropically. This leads to slightly longer run times but minimizes overlap, especially in augmented pathway images where some nodes can be much larger in one direction.
As a special case, positions cannot be computed for proteins linked only to the experimental data and not to the known pathway. These are placed in the lower right side of the display, yielding a cluster of proteins that are not known to be connected to the pathway (, lower right).
This algorithm is relatively fast, interactive, and achieves the desired results without the complexities of more powerful constraint-based techniques such as [5
]. The layouts in took around 2 minutes to compute. We also experimented with simulated annealing methods. These, however, were much slower and did not improve the layouts significantly due to the high network density. Some parameters inherent to force-directed methods still require user adjustment.
3.7 Augmenting a Pathway Image with Dynamic Data
The case of specifying a pathway image and integrating dynamic information seamlessly into the already existing representation is more complicated than assembling a completely new visualization. Simply drawing the database extracted elements on top of the pathway image has several disadvantages, as shown in the cutout of . In contrast, our method creates the illusion that the proteins and interactions drawn dynamically are part of the pathway image (, left).
The following specialized operations are used to create the illusion that the HPRD proteins and interactions are part of the pathway image. The shapes and locations of proteins and interactions in the image are computed in the image preprocessing step. They are then used in the layout stage to minimize overlap (dynamically loaded proteins tend to move to empty image areas). Finally, they are copied from the image and redrawn as masks on top of the final network. This technique ensures that the pathway model stays on top of the dynamic network and gives the illusion that the canonical pathway representation and the dynamic network coexist and interact (, left).
3.8 Exploring the network
In our design the interaction network can be explored at two levels simultaneously: at a global level, where the signaling pathway and other high-level structures are evident, and at a local level, where only one protein and its neighbors appear in detail as the researcher jumps from protein to protein in the network. The two types of visualization coexist as two parallel planes, the local one gliding above the global one (). With these complementary views of the pathway space, the user explores the network in the detailed space that is rich in focused protein information while maintaining an overview of the explored area and orienting the expanded exploration to his or her location within the global view.
Exploration is done in a plane that hovers above the global view and shows in detail only one protein and its interactors. Initial access to the exploration plane can be obtained by double-clicking proteins in the global-view, in the experimental lists, or in a list of all proteins present in the visualization. While in exploration view, clicking one of the interactors shifts the center of the view to this selected protein, a change performed through smooth animation to maintain context understanding. Standard zooming and panning using mouse controls are also available, but testing has found them less favored by users. Proteins in the exploration plane are arranged so as to mimic their placement in the global layer while satisfying aesthetic criteria such as minimum distances between proteins or interaction overlap (, left). The effect is achieved by applying a simulated annealing [4
] algorithm that attempts to maximize layout similarities while ensuring a pleasing drawing. The area allocated to the exploration view is computed dynamically on the basis of the number of proteins to be displayed. A view that places the main protein in the center and its interactors circularly around it is also provided.
Fig. 5 Exploration plane versus zoom-and-pan. (left) The network is explored in a separate plane showing only one protein and its interactors. Selecting an interactor changes the view of that particular protein via a smooth animation. This interaction network (more ...)
Clicking a protein in the exploration view highlights it and its neighbors in the lower plane, making it easier for the user to establish a correspondence between the two.
3.9 Visualization prototype
A compact set of features were added to allow our researchers to operate on the network data and pose visual queries. For instance, selectors and the ability to adjust appearance allow the researcher to highlight interesting aspects of the visualization. In the right panel of , a user has selected various groups or classes of proteins and attached to them special visual attributes such as shape and color a technique often used in stylized signaling pathway representations. The method described in [13
] is used to highlight interactions of one or more selected proteins; interaction highlighting can also be restricted to interactions occurring only between selected proteins (, right).
Fig. 6 Improving network readability through filtering. (left) unfiltered and un-highlighted, (right) filtered by keeping only the pathway proteins and those proteins that connect the pathway proteins (proteins that are not on interaction-paths linking any two (more ...)
Easily extensible filters allow a researcher to remove proteins deemed uninteresting. One potentially useful filter with significant effects keeps only proteins that connect a set of user selected proteins. As an example, shows how a heavily cluttered network was filtered to keep only pathway proteins and those proteins known to connect them through interactions. These filters are crucial since protein interaction networks often contain thousands of proteins and interactions, making comprehension and interaction tedious.
3.10 Implementation Details
The prototype application was written in C++. The G3D 6.7 graphics library was used for 3D graphics and rendering and the Qt 4.3 library for user interface elements. The HPRD database can be downloaded as flat files together with the application.