Visualization of Aligned Biological Networks: A Survey

Transcription

1 Visualization of Aligned Biological Networks: A Survey Steffen Brasch Department of Mathematics and Computer Science University of Greifswald Greifswald, Germany srudnick@uni-greifswald.de Lars Linsen School of Engineering and Science Jacobs University Bremen, Germany l.linsen@jacobs-university.de Georg Fuellen Department of Mathematics and Computer Science University of Greifswald Greifswald, Germany fuellen@uni-greifswald.de Abstract In modern biology, major efforts are undertaken to understand diseases, aging, evolution, and many other aspects of life. Biological networks play a key role in this research. Examples of such networks are protein-protein interaction (PPI) networks, gene regulatory networks, and metabolic pathways. Intuitive and comprehensible visualizations can significantly support the understanding and the analysis of such biological networks. Therefore, much research is conducted in the areas of bio-informatics and information visualization dealing with the problem of visualizing networks. In the course of evolution biological networks changed gradually. Therefore, the conserved core parts of the networks of different species can be aligned by matching the corresponding instances in the species, such as orthologous proteins or genes. Analyzing these alignments is of high relevance, as they convey significant information on protein function and organismal phenotype. For analysis purposes, the alignments have to be made transparent. Visualization methods are needed to display alignments in conjunction with the individual network structures. This is no standard network visualization problem, as we have to deal with more than one network and inter-network relationships. Several approaches to visualize aligned networks exist. In this survey we present and discuss these different approaches and report on their advantages and drawbacks. We draw conclusions on the applicability of the various approaches. I. INTRODUCTION In many biological processes, proteins play a key role. They are involved in biological regulation, development, growth, locomotion, metabolism, and reproduction. Therefore, their analysis is of high importance in biology. Due to their chemical structure, proteins are able to interact with each other, triggering biological processes. For example, signals from the exterior of a cell are mediated to the inside of the cell by protein-protein interaction (PPI) of the signaling proteins. Such processes are also involved in diseases like cancer. Therefore PPIs are fundamental to life, and their investigation yields insight into the evolution of animals [1] and into biochemical function [2]. For each species its proteins and their interactions form a PPI network. The PPI networks of different species are related, since they evolved from a common ancestor whose PPI network can be viewed as their common ancestral network. Learning more about the evolution of PPI networks helps us to understand the networks themselves. PPI networks can be aligned by finding proteins with the same common ancestor, so-called orthologs [3], [4]. Such an alignment shows similarities (and dissimilarities) between different species. For example, the interaction network between key regulators of stem cell pluripotency is believed to be shared by mouse and human, while there are differences in the signaling network that controls the key regulators [5]. More biological background on proteins, the different types of biological networks, and network alignment is given in Section II. For research on biological networks graphical display of such networks is important as humans have the capability of understanding complex structures on a global and local scale and deriving interesting properties using visual tools. Using standard graph layouts for the alignment of several networks has its limitations. The biologist is interested in seeing the interaction of the proteins within one species, but also the alignment based on the ortholog proteins between the species. For visualizing network alignments different approaches exist. Scientists who do research on network alignment developed these approaches in the need to show their results in publications and also to gain insight into their results. We review the different visualization approaches in Section III, discuss them in Section IV, and draw conclusions in Section V. II. NETWORK ALIGNMENT IN BIOLOGY In biology we are not only faced with PPI networks but with many other kinds of biological networks including regulatory networks that involve DNA-protein interaction and metabolic networks that include small metabolites. These networks are also related by evolution and can be aligned. Analysis of all kinds of networks will gain importance, in particular in biomedicine. After all, complex diseases must be tackled nowadays; cancer, arteriosclerosis and dementia are all multifactorial. They all have their cause in the interplay of a multitude of factors, many of which correspond to networks gone out of order. A. Protein-Protein Interaction Networks Protein-protein interactions are important for many biological phenomena such as signaling, transcriptional regulation and multi-enzyme complexes and they are explained by molecular adhesive forces between parts of the proteins (called

2 domains) which in turn can be tracked down to the atomar level. Interaction networks evolve by the loss and gain of nodes (proteins, metabolites,...) and edges (interactions, regulations,...) connecting nodes. Biologists assume that the complex networks interconnecting the components of an organism such as a human being are, like all of life, the result of a more or less gradual evolutionary process of descent with modification. Emergence of biological complexity is nevertheless poorly understood, and a deeper understanding is of utmost importance. Molecular imaging technologies for interaction detection are currently in a rapid phase of development. Using FRET (Förster resonance energy transfer, also known as Fluorescence resonance energy transfer [6]) technology, close physical interaction of two proteins is detected by tagging them with two separate molecules that cause an emission if they are brought together. More recently, BiFC (Bimolecular fluorescence complemention [7], [8]) uses two fragments of a fluorescent molecule such as GFP (green fluorescent protein) that are flagging an interaction if their carriers are coming close enough so that the fluorescent molecule is assembled from its fragments. Using fluorescent proteins of different color, interactions between different carriers can be traced at the same time. Current BiFC research is aimed at designing proteins that disassemble easily, so that the temporal dynamics of an interaction network can be traced. Other techniques developed for the detection of multiple protein interactions are based on antibodies, such as Visual Immunoprecipitation Assays [9] and Multi-Epitope-Ligand Cartography (MELC) [10]. Assume that we have followed the evolution of our proteins and their network up to the common ancestor of human and mouse. After speciation, entities evolve differently in the two lineages of descent that yield the two different species. As described the proteins related by speciation are called orthologs. The inherited interactions and the networks themselves are called orthologous. Conservation of PPIs across species has been observed in many cases [11], [12], [13], for example for the transcriptional network which regulates the development of the heart. It is regulated by PPIs, some of which have been conserved at least since the last common ancestor of fly and human [14]. By comparing the networks of many species, their ancestral core network can be estimated. Network growth can be modeled using specific probability distributions for the gain and loss of nodes and links. In a recent strand of research several groups have begun to systematically compare interaction networks between organisms [15] and orthologous subnetworks are inferred. In particular, the PathBlast tool [16] can detect common descent of linear network substructures by aligning pathways of prespecified length between two networks, matching interacting proteins that are similar according to BLAST (Basic local alignment search tool [17]). More recently, PathBlast has been extended to align more than two networks simultaneously, and to non-linear substructures [18]. Yet another network alignment approach called Local Graph Aligner was developed based on a spin model [19]. It is described in the context of network motif search, and comes with a detailed theory on which statistical significance of matches can be based. Proteins are treated as anonymous entities, only topologies are aligned. The alignment is based on maximizing the overlap of edges ( link overlap ). Most recently, the overlap of orthologous proteins ( node overlap ) can been considered as well [20]. B. Gene Regulatory Networks and Metabolic Pathways Gene regulatory networks and metabolic pathways are posing visualization challenges similar to PPIs. The main difference in the case of gene regulatory networks is that the edges are directed. The biological meaning of the direction is that a protein (transcription factor) regulates the expression of another protein (target gene) but not necessarily vice versa. In the case of metabolic pathways the main difference is that several classes of nodes have to be distinguished, in particular proteins (enzymes) and metabolites. III. LAYOUT OF ALIGNMENTS In general there exist two main approaches of visualizing networks, in a node link diagram or in a matrix representation. In a node link diagram the nodes of the networks are represented by geometrical primitives such as points, circles, or squares and the edges between the nodes are drawn as lines or curves connecting the geometric primitives representing the nodes. In the matrix representation a matrix represents the network, where the entry (i, j) is 0 if there is no edge connecting i and j and 1 if there is an edge. Also an approach to combine both the node link and the matrix representation exists [21]. Most people and also most biologists are more used to the node link representation of a network and therefore this is usually used for layouting biological networks and their alignments. For node link diagrams there are some widely used criteria for the quality of the diagram such as the following: All nodes should be clearly separated, nodes connected by an edge should be placed nearby each other to prevent edges that are too long, edge crossings should be minimized, and the available space should be used in an optimal way. An alignment of networks is not simply a network and therefore some special requirements for its layout should be defined. A. Special Layout Requirements To decide how to produce a layout or to decide which layout is good and which is bad we have to discuss the requirements which such a layout of alignments should meet. For this discussion we have to ask the biologists, what do they want to see, what information is inside the alignments. Each alignment consists of different biological networks which are of interest by themselves. This means for the visualization that the structure of the single networks should be obvious and easily identified and the networks should be

3 Fig. 1. The side-by-side approach with matched nodes on one horizontal line and therefore without additional edges showing aligned PPI networks of fly, yeast, and rat [22]. Fig. 2. The side-by-side approach with matched nodes on nearly the same relative positions but not exactly on the same horizontal line. Example showing aligned PPI networks of yeast, worm, and fly [18]. clearly distinguishable. Nevertheless the alignment relationship, that is which nodes and edges in one network correspond to which nodes and edges in the other networks, is very important and has to be shown clearly as well. Biologists are interested in the core of an alignment. The core of an alignment is the common subnetwork of each of the single networks, where the aligned nodes are simply equated. This gives insight into the evolution of the networks because the core of an alignment is the same in all species and therefore it is the network that is kept during evolution. B. Layout Approaches Each approach to lay out aligned networks deals with the problem to lay out several networks and additional internetwork relationships, given by the alignment itself. Basically aligned networks can be understood as a large network with two types of edges, the edges within the networks (e.g. interactions) and the edges representing the alignment relation. Both could be laid out with standard algorithms and the two types of edges could be separated by color. This approach would not be satisfactory because the single networks would be completely mixed in one network and neither the structure of the single networks nor the matching of nodes and edges could be seen here. In the literature today better approaches are used. 1) Side-by-Side: An intuitive approach is to visualize the single networks side by side and to draw edges connecting the aligned nodes. The separation of the single networks in the visualization helps because these single networks are valuable in themselves. The additional edges connecting the aligned nodes however decrease the readability of the visualization because of the large amount of such edges. One approach to avoid this problem is to draw aligned nodes on a single horizontal line. This allows to ignore the additional edges in the drawing, which makes the layout much clearer. Various examples can be found in the literature [16], [22], [23] and a layout constructed with this approach is shown in Figure 1. An approach that is slightly different from side-by-side in Figure 1 is to draw the matched nodes on nearly, but not exactly, the same relative position in the networks, i. e., not on a single horizontal line [15], [18], [25]. Again the additional edges showing the alignment relationship have to be drawn. An example is shown in Figure 2. Fig. 3. Side-by-side visualization of protein networks retrieved from STRING [24] and aligned manually, centered around the insulin receptor (INSR, top; human, rat and chimp data) and around a protein involved in pluripotency and cancer (MYC, bottom; human, zebrafish and rat data). Renderings were prepared by our own software (under development). If the networks are very similar, side-by-side with nodes on the same positions allows easy by-eye matching of orthologs also without additional edges (Figure 3, top). However, the larger the difference, the more difficult it is to match orthologs (Figure 3, bottom); processing of the visual information becomes difficult because the pattern of network similarity between the left the center and the right panel vanishes. 2) All-in-One: Another approach is to draw all the aligned networks in one node link diagram where the aligned nodes of the different networks are drawn as one node. Labels show to which networks the nodes belong. A problem occurs with the edges because one edge in the visualization represents n potential edges in the n single networks, one edge might exist in one network or not and the edge drawn in the visualization must show to which networks it belongs and to which it does not. This is a large amount of information to be attached to the edges. Examples for this kind of visualization can be found in the literature [26], [27]. Figure 4 shows examples using this approach. A more theoretical access to this approach uses the metagraph concept [28]. Here metanodes are defined each consisting of aligned nodes and metaedges connecting the metanodes. These metaedges represent edges of the single networks. The idea is a dynamic layout where the metaedge might be substituted by the single edges that are represented by it or where the metanode might be restricted to just the nodes of some of the single networks. By navigating and interacting different layouts of the network might be produced. A toy

4 Fig. 4. Examples for the all-in-one approach where labels show which nodes of the single networks are represented by one node, and the type and/or color of the edges represents the edge properties. Both examples show an alignment of PPI networks of yeast and fly [26], [27]. Fig. 5. A toy example of the idea for a more general all-in-one approach using the metagraph concept with different possible views of the aligned networks [28]. a) shows only the metanodes without any edges and b) - d) show projections to the single networks example showing this idea is given in Figure 5. IV. DISCUSSION Many approaches for visualizing aligned networks exist and are used in the literature. Roughly these approaches can be classified into two types, side-by-side and all-in-one. The advantage of both types of approaches is that these are just concepts and for the concrete layout computation traditional algorithms can be used, maybe with some constraints. A. Side-by-Side The approach, with orthologous nodes drawn on one horizontal line, forces the layout of the single networks to be longish in the vertical direction because no other than the orthologous nodes must lie on one horizontal line. This is no problem for small networks like in Figure 1 but it causes bad layouts for larger networks because of the loss of freedom for the layout computations. An advantage is obviously the smaller number of edges which also causes less edge crossings. The matched nodes can be observed easily and also the visualization of paralogous nodes is no problem. Nevertheless it is hard to recognize the specific structure of the alignment or of the core because it is hard to observe conserved edges in the networks. The best layout for single networks with the side-by-side approach is achieved by the approach with free positions for the nodes but with the matched nodes roughly on one relative position. A problem occurring is the huge amount of additional edges as can be seen in Figure 2 but they are obviously needed as seen in Figure 3 bottom and therefore also the number of edge crossings increases. Additionally identifying the matched nodes is harder with this layout than with the other side-byside approaches. Nevertheless this approach is able to handle also larger networks and like all side-by-side approaches it handles paralogous nodes. B. All-in-One The approach of drawing all aligned networks in one graph allows the user to easily see the core of the networks. But two major problems exist, namely, how to draw the edges, and how to draw paralogs. For the edges the question is how to encode the type of the edge. For two networks we have three possibilities: an edge may exist in both networks, only in the first, or only in the second network. For three networks we already have seven cases. The problem with the paralogs arises because only one node per network is aligned with at most one node in the other network. Just spelling out the different paralogous nodes in this one node in the visualization is not satisfactory because these nodes are also connected by individual edges. The metagraph concept gives the possibility to account for this subnetwork of paralogs problem. Another problem occurs if we have more than two networks in one alignment, because each node then stands for more than two nodes in the original networks and therefore has got more than two labels. A possible solution is again the concept of metanodes consisting of orthologs and subnetworks of paralogs. These potentially very large metanodes consisting of several small networks in the metagraph will also be a problem for good visualizations. Nevertheless there is merit in the approach using the metagraph concept enabling user interaction with the network and filtering of certain information. Because of the complex structure of aligned networks filtering might give deeper insight. V. CONCLUSION In biology and biomedical research, networks are very important and they have to be visualized to get an intuitive understanding. In addition to just exploring single networks the interest in aligned networks giving information on evolution increases. Such network alignments are new structures consisting of several single networks and an inter-network matching of nodes and edges. For exploring these alignments visualization is very helpful and therefore several visualization approaches have already been developed. In this paper we have given an overview of the different approaches and discussed their advantages and disadvantages with respect to some quality criteria which we consider natural. The simple all-in-one approach must be extended to be able to visualize also subnetworks of paralogs, maybe using the concept of metagraphs, to be able to fulfill all basic needs. But this has not been done up to now and several problems have to be solved beforehand. Nevertheless the advantage of

5 this approach is the large freedom for the layout of the single networks which allows for better layouts and it allows user interaction and filtering. In contrast to the all-in-one approach, the side-by-side approach is generally more useful, but it places more restrictions on the layout of the single networks, producing worse layouts, or it leaves more freedom for these layouts at the cost of more and longer edges and more edge crossings and, in particular, at the cost of a less readable alignment. The research in aligning networks goes on and therefore in the future larger alignments of more species will be calculated and will have to be visualized. Thus, in the near future new approaches have to be developed, since the existing approaches do not scale well to large networks and will not be able to provide sufficiently good visualizations. This is a new and interesting challenge for the information visualization community. REFERENCES [1] E. H. Davidson and D. H. Erwin, Gene regulatory networks and the evolution of animal body plans. Science, vol. 311, no. 5762, pp , February [Online]. Available: [2] R. Sharan, I. Ulitsky, and R. Shamir, Network-based prediction of protein function, Mol Syst Biol, vol. 3, March [Online]. Available: [3] W. M. Fitch, Distinguishing homologous from analogous proteins, Systematic Zoology, vol. 19, no. 2, pp , June [4], Homology: a personal view on some of the problems, Trends in Genetics, vol. 16, no. 5, pp , May [5] M. Boiani and H. R. Scholer, Regulatory networks in embryo-derived pluripotent stem cells, Nat Rev Mol Cell Biol, vol. 6, pp , November [6] E. A. Jares-Erijman and T. M. Jovin, Imaging molecular interactions in living cells by FRET microscopy, Current Opinion in Chemical Biology, vol. 10, no. 5, pp , October [7] C.-D. Hu and T. K. Kerppola, Direct Visualization of Protein Interactions in Living Cells Using Bimolecular Fluorescence Complementation Analysis. Cold Spring Harbor Laboratory Press, [8] T. K. Kerppola, Design and implementation of bimolecular fluorescence complementation (BiFC) assays for the visualization of protein interactions in living cells, Nat. Protocols, vol. 1, no. 3, pp , October [9] P. Niethammer, I. Kronja, S. Kandels-Lewis, S. Rybina, P. Bastiaens, and E. Karsenti, Discrete states of a protein interaction network govern interphase and mitotic microtubule dynamics, PLoS Biology, vol. 5, [10] W. Schubert, B. Bonnekoh, A. J. Pommer, L. Philipsen, R. Böckelmann, Y. Malykh, H. Gollnick, M. Friedenberger, M. Bode, and A. W. M. Dress, Analyzing proteome topology and function by automated multidimensional fluorescence microscopy, Nature Biotechnology, vol. 24, pp , October [11] P. Bork, L. Jensen, C. von Mering, A. Ramani, I. Lee, and E. Marcotte, Protein interaction networks from yeast to human. Curr Opin Struct Biol, vol. 14, no. 3, pp , [12] A. Ramani and E. Marcotte, Exploiting the co-evolution of interacting proteins to discover interaction specificity. J Mol Biol, vol. 327, no. 1, pp , [13] L. Matthews, P. Vaglio, J. Reboul, H. Ge, B. Davis, J. Garrels, S. Vincent, and M. Vidal, Identification of potential interaction networks using sequence-based searches for conserved protein-protein interactions or interologs. Genome Res, vol. 11, no. 12, pp , [14] R. Cripps and E. Olson, Control of cardiac development by an evolutionarily conserved transcriptional network. Dev Biol, vol. 246, no. 1, pp , [15] R. Sharan and T. Ideker, Modeling cellular machinery through biological network comparison, Nature Biotechnology, vol. 24, no. 4, pp , April [Online]. Available: [16] B. P. Kelley, R. Sharan, R. M. Karp, T. Sittler, D. E. Root, B. R. Stockwell, and T. Ideker, Conserved pathways within bacteria and yeast as revealed by global protein network alignment. Proc Natl Acad Sci U S A, vol. 100, no. 20, pp , September [Online]. Available: [17] S. Altschul, T. Madden, A. Schaffer, J. Zhang, Z. Zhang, W. Miller, and D. Lipman, Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucleic Acids Res, vol. 25, no. 17, pp , [18] R. Sharan, S. Suthram, R. Kelley, T. Kuhn, S. McCuine, P. Uetz, T. Sittler, R. Karp, and T. Ideker, Conserved patterns of protein interaction in multiple species. Proc Natl Acad Sci U S A, vol. 102, no. 6, pp , [19] J. Berg and M. Lässig, Local graph alignment and motif search in biological networks. Proc Natl Acad Sci U S A, vol. 101, no. 41, pp , [20], Cross-species analysis of biological networks by bayesian alignment. Proc Natl Acad Sci U S A, vol. 103, no. 29, pp , July [Online]. Available: [21] N. Henry and J.-D. Fekete, MatrixExplorer: a dual-representation system to explore social networks. IEEE Trans. Vis. Comput. Graph., vol. 12, no. 5, pp , [22] M. Koyutürk, Y. Kim, S. Subramaniam, W. Szpankowski, and A. Grama, Detecting conserved interaction patterns in biological networks. J Comput Biol, vol. 13, no. 7, pp , September [Online]. Available: [23] Z. Liang, M. Xu, M. Teng, and L. Niu, Comparison of protein interaction networks reveals species conservation and divergence. BMC Bioinformatics, vol. 7, pp. 457+, October [24] C. von Mering, L. J. Jensen, M. Kuhn, S. Chaffron, T. Doerks, B. Krüger, B. Snel, and P. Bork, STRING 7 recent developments in the integration and prediction of protein interactions. Nucleic Acids Res, vol. 35, no. Database issue, January [25] K. Tan, T. Shlomi, H. Feizi, T. Ideker, and R. Sharan, Transcriptional regulation of protein complexes within and across species, PNAS, vol. 104, no. 4, pp , January [Online]. Available: [26] S. Bandyopadhyay, R. Sharan, and T. Ideker, Systematic identification of functional orthologs based on protein network comparison, Genome Res., vol. 16, no. 3, pp , March [Online]. Available: [27] E. Hirsh and R. Sharan, Identification of conserved protein complexes based on a model of protein network evolution. Bioinformatics, vol. 23, no. 2, January [28] Z. Hu, J. Mellor, J. Wu, M. Kanehisa, J. M. Stuart, and C. Delisi, Towards zoomable multidimensional maps of the cell, Nature Biotechnology, vol. 25, no. 5, pp , May 2007.