Bioinformatics Advance Access originally published online on April 23, 2008
Bioinformatics 2008 24(12):1442-1447; doi:10.1093/bioinformatics/btn200
A global pathway crosstalk network
Computational Biology, GlaxoSmithKline R&D, 709 Swedeland Road, UMW2230, King of Prussia, PA 19406, USA
*To whom correspondence should be addressed.
 |
ABSTRACT |
|---|
 TOP ABSTRACT 1 INTRODUCTION2 METHODS3 RESULTS4 DISCUSSIONACKNOWLEDGEMENTSREFERENCES |
|---|
Motivation: Given the complex nature of biological systems, pathways often need to function in a coordinated fashion in order to produce appropriate physiological responses to both internal and external stimuli. Therefore, understanding the interaction and crosstalk between pathways is important for understanding the function of both cells and more complex systems.
Results: We have developed a computational approach to detect crosstalk among pathways based on protein interactions between the pathway components. We built a global mammalian pathway crosstalk network that includes 580 pathways (covering 4753 genes) with 1815 edges between pathways. This crosstalk network follows a power-law distribution: P(k) 
k–
, = 1.45, where P(k) is the number of pathways with k neighbors, thus pathway interactions may exhibit the same scale-free phenomenon that has been documented for protein interaction networks. We further used this network to understand colorectal cancer progression to metastasis based on transcriptomic data.
Contact: yong.2.li{at}gsk.com
Supplementary information: Supplementary data are available at Bioinformatics online.
|
1 INTRODUCTION |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 DISCUSSIONACKNOWLEDGEMENTSREFERENCES |
|---|
High-throughput biological experiments that interrogate many genes simultaneously have generated unprecedented amounts of data. These data in turn have stimulated the development of a variety of data analysis methods which aim at understanding biological phenomena at a systems level. Genome-wide DNA microarray experiments sometimes can generate lists of hundreds or thousands of genes whose expression significantly changed between experiment and control conditions. One way to quickly extract the underlying biology from such large gene lists is to identify pathways that are significantly changed in those lists. This information is particularly useful since genes often act together in the form of pathways to perform certain biological functions. Two commonly used methods have been developed to address this need. In one approach, the hypergeometric distribution was used to assess the statistical significance of pathway enrichment (Al Shahrour et al., 2004). In the second approach, a non-parametric method called Gene Set Enrichment Analysis was developed to determine if a pathway is significantly changed in a microarray experiment (Mootha et al., 2003; Subramanian et al., 2005).
In addition to gene-expression profiles, protein–protein interactions also provide valuable information regarding how genes carry out their biological functions. Collectively, protein interactions form a global interaction network which could shed light on the overall organization and interaction among functionalities that are essential to cell survival and growth. Genome-wide protein–protein interaction screenings based on yeast two-hybrid technique have been performed in budding yeast (Uetz et al., 2000), fruit fly (Giot et al., 2003), worm (Li et al., 2004), human (Stelzl et al., 2005) and other organisms. On the other hand, literature-based curation and data-mining processes have also generated substantial amounts of protein interaction data (Rajagopalan and Agarwal, 2005; Ramani et al., 2005). Network-based methods have been used to analyze these interaction data and gain insights into the mechanism by which biological systems operate (Barabasi and Oltvai, 2004; Xia et al., 2004). For example, interaction networks characterized so far have scale-free and small-world properties. In a scale-free network, the number of nodes with k neighbors, denoted as P(k), follows a power-law distribution: P(k) k–, where is the degree exponent. In other words, many nodes have relatively few neighbors whereas exponentially smaller numbers of nodes, also called hubs, have many neighbors. It has been shown that, in yeast, hubs are much more likely to be essential genes compared to non-hub genes (Jeong et al., 2001). Proteins sharing similar functions can form modules or clusters in interaction networks (Han et al., 2004; Rives and Galitski, 2003).
We want to take a step beyond identifying lists of significantly changed pathways to further understand how they act together to account for the observed phenotypes. One effective biological approach to identifying pathway interaction is through genetic screenings where synthetic lethality of two mutations often indicates interaction between two pathways where those two mutations reside separately (Tong et al., 2004). Such genetic interactions have also been used to link sets of densely connected proteins in a protein network (Kelley and Ideker, 2005). We have taken a computational approach to explore pathway interaction by systematically combining both pathway data and interaction network data. Our approach is based on the assumption that two pathways are likely to interact with or influence each other (crosstalk) if significantly more protein interactions are detected between these two pathways than expected by chance. We will show here how we built a pathway network based on crosstalk, characterized its network architecture, and applied it to help us further understand colorectal cancer progression to metastasis based on gene-expression profiling data.
|
2 METHODS |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 DISCUSSIONACKNOWLEDGEMENTSREFERENCES |
|---|
2.1 Pathway data
We collected pathways from HumanCyc and BioCarta. We also treated gene sets from Gene Ontology (GO) Biological Process (BP) and Cellular Component (CC) as pathways in our analysis since these gene sets represent group of genes of related function. We reasoned that the GO CC gene sets can add unique value to the pathway crosstalk network since they indicate where the interaction is likely to take place. After excluding pathways with less than 5 genes or more than 100 genes, we ended up with 1697 pathways with following source distribution: 1318 GO, 281 BioCarta and 98 HumanCyc. These pathways cover a total of 5978 human genes.
2.2 Protein interaction data
We collected human protein interaction data from the following public sources: HPRD (2004–10), Reactome (2006–03), MINT (2006–03), BIND (2006–03), and Ramani et al. (2005). Interactions from MINT and Ramani et al. also included inferred interactions from model organisms based on protein orthology. After removing self interactions, we ended up with 60 336 interactions which cover 9685 human genes. Many methodologies have been developed to clean up interaction data generated by high-throughput experiments or computational inference (Chua et al., 2006; Lage et al., 2007; Ramani et al., 2005; Stelzl et al., 2005). We chose to simply use the downloaded interactions ‘as is’ because the majority of them (90%) are derived from either manual curation (HPRD, Reactome, BIND,and the curated human interactions from MINT) or pre-filtered (Ramani et al., 2005). Furthermore, our randomization procedure in 2.3.5 can help eliminate the impact of potential false-positives in our interaction data.
2.3 Construction of pathway crosstalk network
- Generate a set of pathways for crosstalk analysis by removing pathways containing less than 5 genes or more than 100 genes. Pathways with too many genes might be too generic and pathways with too few genes may not have sufficient biological content. These size cutoffs were set up arbitrarily.
- Evaluate gene overlap between any given pair of pathways by performing Fisher Exact test (Al Shahrour et al., 2004). Raw P-values were adjusted by false discovery rate (FDR) Benjamini-Hochberg (BH) procedure (Benjamini and Hochberg, 1995) to account for multiple hypothesis testing. Any pathway pairs with adjusted P-value <0.05 were considered to overlap significantly and were removed during the pruning step (see below).
- Count number of protein interactions between any two pathways. For each pathway pair, first remove genes common to both pathways, then count all protein interactions between these two pathways.
- Estimate background distribution of protein interaction count of each pathway pair. Each pathway was randomized as follows. Go through all genes in a given pathway. If a gene does not have any interactions, skip it. If a gene has interactions, first count the number of genes it interacts with, then randomly draw a gene from the protein interaction dataset which interacts with the same or similar number of genes, and replace the original pathway gene with this newly selected gene. Once both pathways were randomized, Step 3 is performed to count the number of interactions between them. This randomization step is repeated 1000 times.
- Perform one-sided Fisher Exact test on all pathway pairs using the 2 x 2 contingency table that include the following numbers: n, N-n, r, R-r where n denotes the interaction count between original pathways, N denotes the number of total interaction counts of all pathway pairs, r denotes the average of interaction counts between the pair of corresponding randomized pathways after 1000 rounds of randomizations and R denotes the average of total interaction counts of all randomized pathway pairs after 1000 rounds of randomizations. The null hypothesis is that the ratio of true interactions between two pathways to all interactions (n/N) is the same as the ratio of random interactions to all random interactions (r/R). In our analysis, we only focused on pathway pairs where n/N is significantly higher than r/R. Fisher exact test P-values were adjusted using FDR BH procedure (Benjamini and Hochberg, 1995) to account for multiple hypothesis testing. We also used a second approach to assess the significance of interaction count between two pathways. Specifically, we ranked the count of true interactions against the random interaction counts from 1000 permutations performed on the same pathway pair. Empirical P-value was calculated by counting the number of permutations in which the random interaction count is higher than or equal to the true interaction count, then dividing that number by the total number of permutations. Fisher exact test P-values and empirical P-values were very well correlated (data not shown).
- All pathway pairs with adjusted Fisher exact test P-value <0.05 were pulled together to construct a network in which a node is a pathway and an edge represents crosstalk between two pathways. Two types of ‘redundant’ edges caused by gene redundancy among some pathways were removed during the following pruning steps to clean up the network.
- (a) All edges between two pathways with significant gene overlap identified in Step 2 were considered as not informative and thus removed from the network. Note that it is our intent to discover crosstalk among different biological activities. Pathway pairs where both pathways significantly overlap with each other in terms of gene members represent similar biology and thus were excluded from the network.(b) Two overlapping pathways may both interact with the same pathway. In this case, the two edges were considered redundant and one of them was removed. For example, pathway A has four neighbors: pathway B, C, D and E. If at least 75% of genes in B are also in C and B has less number of genes than C, then the edge between A and B was removed. Repeat this process until no more edges between pathway A and its neighbors could be removed. This step is especially necessary for pathways derived from GO where all children gene sets are complete subsets of their parents. If both a child and a parent GO pathway interact with pathway A, the edge between pathway A and the child GO pathway will be removed during this step of pruning.
2.4 Network clustering
Pathways in pathway crosstalk network (PCN) were clustered based on their shortest path profiles (Rives and Galitski, 2003). In brief, the shortest path (SP) between any given pair of pathways was computed using a standard graph-based procedure. Distance between two pathways (d) is defined as the length of the SP between them. A simple transformation (1/d) was used to generate a similarity score for all pathway pairs which is in the range of 0 and 1. The 1/d between a pathway and itself is set to 1. Hierarchical agglomerative average-linkage clustering was then performed using Spotfire DecisionSite 8.1 (Spotfire Inc, Somerville, MA). Clustering result was viewed in a heatmap symmetrical along the diagonal line (self-associations). Darker color indicates two pathways are closer to each other in terms of SP length.
2.5 Enrichment analysis of pathway clusters
The pathways in the hierarchical clustering shown in Figure 2 were extracted and are listed in sequential order (clustering ordering) in Supplementary Table 10. Thus, pathways next to each other in numerical order (in Supplementary Table 10) fall into the same cluster unless they are located close to a cluster boundary. A few of these clusters were identified based on the dendrogram on the heat map display (Fig. 2). The pathways in these clusters were determined based on clustering ordering (Supplementary Table 10) and their genes were pooled together. This gene list is analyzed against GO categories to determine which categories are significantly enriched (Al Shahrour et al., 2004). Clusters with predominant underlying biology (in terms of the top enriched GO categories, see Supplementary Table 11 for details) were identified and labeled in Figure 2.
2.6 Microarray data analysis using PCN
Provenzani et al. (2006) identified 2018 probesets that are significantly changed in a cell model for colorectal cancer metastasis. These probesets were downloaded from Gene Expression Omnibus, a public gene-expression database, and mapped to individual genes. The gene list was then analyzed against the 580 pathways in PCN to determine which pathways are enriched in the gene list (Al Shahrour et al., 2004). Pathways that are significantly enriched (adjusted P-value < 0.05) were analyzed to uncover potential modules (Rajagopalan and Agarwal, 2005). A module here is a subnetwork where enriched pathways lie close to each other and has a high score based on their enrichment P-values. In order to use this algorithm, we used PCN as the underlying network and the input is a list of enriched pathways with enrichment P-values. Modules with more than two pathways were shown in Figure 4.
|
3 RESULTS |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 DISCUSSIONACKNOWLEDGEMENTSREFERENCES |
|---|
We collected pathways from BioCarta, HumanCyc, and gene sets from GO BP category and CC category. We ended up with 1697 pathways after removing pathways with too many genes (>100) or too few genes (<5). We also collected 60 336 unique interactions from the following public sources and publications: HPRD (Peri et al., 2004), MINT (Zanzoni et al., 2002), Reactome (Joshi-Tope et al., 2005), BIND (Bader et al., 2003) and Ramani et al. (2005). Interaction types include protein–protein interaction, transcriptional regulation, phosphorylation, co-participation in the same biochemical reaction and co-membership in the same protein complex. All interactions are treated equally during pathway crosstalk analysis.
We generated a global PCN by integrating both pathway and interaction data mentioned above (Section 2). We represented PCN as a graph in which nodes are pathways and edges represent crosstalk between pathways. The PCN contains 580 nodes and 1815 edges (Supplementary Table 1), covering 4753 genes. The PCN can be broken down to 12 connected components. Most of the pathways in PCN (552 out of 580) reside in the largest component, which was used in our analysis throughout this article. The remaining 11 small components contain 28 pathways and 18 crosstalk edges. Key network topology measurements of the PCN (Fig. 1a and b) suggest that it is a scale-free, small-world network (Barabasi and Oltvai, 2004). Note that the PCN might have an intrinsic hierarchical structure (Barabasi and Oltvai, 2004) as the average clustering coefficient of node decreases when the node degree increases (Fig. 1c).
|
We performed network clustering on PCN based on SP profiles (Rives and Galitski, 2003). It is well recognized at least in protein interaction networks that genes of similar functions tend to cluster together (Rives and Galitski, 2003; Tanay et al., 2004). As shown in Figure 2, a few relatively discrete clusters are evident from visual inspection of the heatmap display of the hierarchical clustering of the PCN. Some of them have a clear boundary around them as pathways in those clusters mainly interact with each other within the same cluster, whereas others have a fuzzy boundary since pathways in those clusters also interact with pathways from outside the clusters. To uncover the underlying biology of those clusters, pathways from each cluster were collected based on their clustering ordering and enrichment analysis was performed (Section 2). Clusters with a clear predominant underlying biology are shown in Figure 2. The first cluster in the top left corner in Figure 2 contains pathways mainly involved in mitochondrial functions (Supplementary Fig. 1a). The second cluster (Supplementary Fig. 1b) contains pathways involved in RNA production, splicing, cleavage, processing, capping and transport. Next to this cluster is a cluster (Supplementary Fig. 1c) responsible for cell-cycle regulation and biology of nuclear chromosomes such as DNA replication and recombination, DNA damage repair and responses and telomere maintenance. At the right bottom corner in Figure 2 lies a large cluster with a fuzzy boundary. It contains pathways involved in a wide variety of cellular signaling events. Many of those pathways are involved in inflammation, immune response, and B/T-cell development and a portion of it is shown in Supplementary Figure 1d. The clustering of PCN is certainly not unique, and the graphics shown are illustrative exemplars that provide insight into the biology of the PCN.
|
We used Cytoscape (Shannon et al., 2003) to visualize the PCN. We used the yFiles Organic Layout option as it is designed to reveal the clustered structure of a graph (Cytoscape User Manual). Indeed, with no manual manipulation, the layout of the whole network readily reveals a few relatively isolated clusters (Fig. 3). When pathways were colored based on their cluster membership shown in Figure 2, it is evident that pathways in clusters in Figure 3 also share similar biological functions. For example, the ‘mitochondrion’ cluster is relatively isolated from the rest of PCN and contains pathways mainly involved in mitochondrial functions. Not surprisingly, the ‘cell-signaling’ cluster is the biggest cluster which mainly contains a diverse array of cell-signaling pathways. It is located in the center of PCN and connected to all other clusters.
|
We next looked at pathways that have high degrees in the PCN. The key feature of a scale-free network is that it has a few nodes (hubs) connecting many other nodes while majority of nodes have low degree of connection (Barabasi and Oltvai, 2004). We ranked pathways by their degrees and the top 20 pathways were shown (Supplementary Table 2). Most of them are signaling pathways. The number one hub pathway is ‘Small GTPase Mediated Signal Transduction’ from GO. This is not surprising, given that small GTPases are involved in regulating a wide variety of cellular processes such as growth, cellular differentiation, cell movement and lipid vesicle transport (Macara et al., 1996). In our analysis, this pathway interacts with 47 other pathways involved in cytoskeleton regulation, cytokine and chemokine mediated signaling, mitogen-activated protein kinase (MAPK) signaling, cell migration and growth, etc. It has been shown that removal of hubs in protein interaction networks could cause dramatic changes in the overall network structure such as break-down of the whole network into smaller pieces (Han et al., 2004). Removal of top 10 hub pathways, about 2% of all pathways, from PCN removed about 22% of all edges. However, it did not break down the network since it still contains 530 connected pathways. It did, nonetheless, decrease the overall connectivity since mean path length increased from 4.65 to 5.01.
We ranked pathway crosstalks by their adjusted P-values (Section 2) and top 20 are shown in Supplementary Table 3 (See Supplementary Table 1 for a complete list of crosstalks). Many crosstalks in PCN recapitulated known biology such as interaction between ribosome and translation regulation, mTOR signaling and ribosome, mRNA cleavage and RNA splicing, fatty-acid oxidation and aerobic respiration, cytoskeleton and cell migration, DNA damage repair and cell cycle regulation, etc. Interesting crosstalks can also be found in PCN. For example, crosstalk between DNA mismatch repair and meiosis was identified based on interactions between proteins in meiosis such as RAD50/51 and ATM and proteins involved in mismatch repair such as MLH1 and MSH2/6. We also looked at crosstalk between pathways from different categories such as GO BP and CC category (Supplementary Table 4). These crosstalks may suggest where a particular process takes place or the location of an influencing signal.
We looked for interesting topological structures such as cliques in PCN. A clique is a fully connected subgraph. We have identified 602 size-3 and 30 size-4 cliques (Supplementary Tables 5 and 6). No size-5 cliques were present in the current version of PCN. We are interested in cliques because they represent a group of pathways that intensively interact with each other, thus implying a tight functional coupling among them. Note that in our analysis, pathways in each clique do not overlap much with each other in terms of gene members (Supplementary Table 7), and yet they all interact with each other. Not surprisingly, many of them are related to signal transduction. We also noted that there are only 47 unique pathways in all 30 size-4 cliques, suggesting most of those cliques overlap with each other in PCN. Indeed, 40 out of 47 pathways form a single, dense cluster with an average clustering coefficient = 0.7 and mean shortest path length = 2.6. This high connectivity could explain why removal of top 10 hubs did not break down the network as 8 out of those 10 hubs are present in this dense cluster. We speculate that cliques may represent key information integration sites in a global cellular network where different pathways intensively interact with each other to integrate and regulate information flow.
In principle, PCN could represent a new framework upon which biological data is integrated and analyzed. To this end, we applied the PCN to analyze microarray data from a study on colorectal cancer progression to metastasis (Provenzani et al., 2006). In this study, 1587 genes were found to have significantly changed in the polysomal compartment in a cell model of colorectal cancer metastasis. We identified 71 pathways that are enriched in those genes (Section 2). To further understand the underlying biology of those pathways, we mapped them onto PCN and focused on those that can form compact modules in PCN by applying a subnetwork detection algorithm (Rajagopalan and Agarwal, 2005). When pathways lie close to each other in PCN to form a compact module, they might functionally influence each other. We therefore reasoned that if we could identify pathways that not only are significantly changed in terms of gene expression but also influence each other, we could have higher confidence that those pathways are indeed biologically relevant. We found 2 pathway modules which contain 19 pathways (Section 2, see Supplementary Table 8 and 9 for details). As shown in Figure 4, one module contains six pathways and is clearly involved in cell-cycle regulation. The other module has 13 pathways where 4 of them can be readily associated with cell migration: actin-filament-based process, cell migration, cell to cell adhesion signaling, and mCalpain and friends in cell motility. In addition, key players of four signaling pathways from this module, namely hepatocyte growth factor receptor (also known as c-Met), EGFR, VEGF and CXCR4, have also been implicated in cancer metastasis (Christensen et al., 2005; Mercurio et al., 2005; Wang et al., 2006; Zhang et al., 2006). A pathway concerning telomeres is also present in this module. Interestingly, telomerase, aside from being involved in maintenance of telomere length, may play a role in tumor metastasis (Bagheri et al., 2006). Further study of crosstalk among these pathways could potentially shed new light on metastasis of colorectal cancer.
|
|
4 DISCUSSION |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 DISCUSSIONACKNOWLEDGEMENTSREFERENCES |
|---|
We present here a novel computational approach to identifying interactions (crosstalk) among pathways by integrating biological pathways and protein interaction data. Many expected pathway crosstalks have been identified by this approach. When two pathways have few protein interactions between them and yet still functionally influence each other, they will be missed by our current approach. In addition, it is possible that the current PCN can be further improved by cleaning up the interaction data using a variety of methodologies (Section 2.2). The resulting PCN connects together many pathways from a wide variety of biological processes. Clusters representing broad functional category are formed within PCN. Valuable insights into the coordination of different biological activities might be gained by analyzing (1) how pathways influence each other within a cluster and (2) how different clusters ‘crosstalk’ with each other through pathways located at the interface between clusters. We showed that PCN can be utilized to analyze genome-wide expression profiling data. One additional advantage of this approach is that, in certain cases when pathways have only marginal enrichment P-values, they may still collectively carry a stronger signal if they can form a compact module in PCN. An analogous situation would be one that is associated with some microarray experiments where a few genes are only marginally changed in terms of their expression level. However, if those genes all belong to the same pathway, then the change at pathway level is more likely to be real.
The completion of the sequencing of human genome and the development of new high-throughput screening technologies greatly enabled systems biology approach in biomedical research. The ultimate goal of systems biology is to understand biological phenomena through analysis of interactions of all cellular and biochemical components within a cell or organism (Liu, 2005). The pathway crosstalk network approach could potentially help us move closer to achieving that goal.
|
ACKNOWLEDGEMENTS |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 DISCUSSIONACKNOWLEDGEMENTSREFERENCES |
|---|
We thank Liwen Liu for providing pathway datasets, Michael Lutz and David Searls for their encouragement, support and comments on the manuscript.
Conflict of Interest: none declared.
|
FOOTNOTES |
|---|
Associate Editor: Limsoon Wong
Received on July 11, 2007; revised on April 3, 2008; accepted on April 21, 2008
|
REFERENCES |
|---|
|
TOPABSTRACT1 INTRODUCTION2 METHODS3 RESULTS4 DISCUSSIONACKNOWLEDGEMENTSREFERENCES |
|---|
Al Shahrour F, et al. FatiGO: a web tool for finding significant associations of Gene Ontology terms with groups of genes. Bioinformatics (2004) 20:578–580.
Bader GD, et al. BIND: the Biomolecular Interaction Network Database. Nucleic Acids Res (2003) 31:248–250.
Bagheri S, et al. Genes and pathways downstream of telomerase in melanoma metastasis. Proc. Natl Acad. Sci. USA (2006) 103:11306–11311.
Barabasi AL, Oltvai ZN. Network biology: understanding the cell's functional organization. Nat. Rev. Genet (2004) 5:101–113.[CrossRef][Web of Science][Medline]
Benjamini Y, Hochberg Y. Controlling the false discovery rate: a practical and powerful approach to multiple testing. J. Roy. Stat. Soc. B (1995) 57:289–300.
Christensen JG, et al. c-Met as a target for human cancer and characterization of inhibitors for therapeutic intervention. Cancer Lett (2005) 225:1–26.[CrossRef][Web of Science][Medline]
Chua HN, et al. Exploiting indirect neighbours and topological weight to predict protein function from protein-protein interactions. Bioinformatics (2006) 22:1623–1630.
Giot L, et al. A protein interaction map of Drosophila melanogaster. Science (2003) 302:1727–1736.
Han JD, et al. Evidence for dynamically organized modularity in the yeast protein-protein interaction network. Nature (2004) 430:88–93.[CrossRef][Medline]
Jeong H, et al. Lethality and centrality in protein networks. Nature (2001) 411:41–42.[CrossRef][Medline]
Joshi-Tope G, et al. Reactome: a knowledgebase of biological pathways. Nucleic Acids Res (2005) 33:D428–D432.
Kelley R, Ideker T. Systematic interpretation of genetic interactions using protein networks. Nat. Biotechnol (2005) 23:561–566.[CrossRef][Web of Science][Medline]
Lage K, et al. A human phenome-interactome network of protein complexes implicated in genetic disorders. Nat. Biotechnol (2007) 25:309–316.[CrossRef][Web of Science][Medline]
Li S, et al. A map of the interactome network of the metazoan C. elegans. Science (2004) 303:540–543.
Liu ET. Systems biology, integrative biology, predictive biology. Cell (2005) 121:505–506.[Medline]
Macara IG, et al. The Ras superfamily of GTPases. FASEB J (1996) 10:625–630.[Abstract]
Mercurio AM, et al. Non-angiogenic functions of VEGF in breast cancer. J. Mammary Gland Biol. Neoplasia (2005) 10:283–290.[CrossRef][Web of Science][Medline]
Mootha VK, et al. PGC-1alpha-responsive genes involved in oxidative phosphorylation are coordinately downregulated in human diabetes. Nat. Genet (2003) 34:267–273.[CrossRef][Web of Science][Medline]
Peri S, et al. Human protein reference database as a discovery resource for proteomics. Nucleic Acids Res (2004) 32:D497–D501.
Provenzani A, et al. Global alterations in mRNA polysomal recruitment in a cell model of colorectal cancer progression to metastasis. Carcinogenesis (2006) 27:1323–1333.
Rajagopalan D, Agarwal P. Inferring pathways from gene lists using a literature-derived network of biological relationships. Bioinformatics (2005) 21:788–793.
Ramani AK, et al. Consolidating the set of known human protein-protein interactions in preparation for large-scale mapping of the human interactome. Genome Biol (2005) 6:R40.[CrossRef][Medline]
Rives AW, Galitski T. Modular organization of cellular networks. Proc. Natl Acad. Sci. USA (2003) 100:1128–1133.
Shannon P, et al. Cytoscape: a software environment for integrated models of biomolecular interaction networks. Genome Res (2003) 13:2498–2504.
Stelzl U, et al. A human protein-protein interaction network: a resource for annotating the proteome. Cell (2005) 122:957–968.[CrossRef][Web of Science][Medline]
Subramanian A, et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proc. Natl Acad. Sci. USA (2005) 102:15545–15550.
Tanay A, et al. Revealing modularity and organization in the yeast molecular network by integrated analysis of highly heterogeneous genomewide data. Proc. Natl Acad. Sci. USA (2004) 101:2981–2986.
Tong AH, et al. Global mapping of the yeast genetic interaction network. Science (2004) 303:808–813.
Uetz P, et al. A comprehensive analysis of protein-protein interactions in Saccharomyces cerevisiae. Nature (2000) 403:623–627.[CrossRef][Medline]
Wang J, et al. The pivotal role of CXCL12 (SDF-1)/CXCR4 axis in bone metastasis. Cancer Metastasis Rev (2006) 25:573–587.[CrossRef][Web of Science][Medline]
Xia Y, et al. Analyzing cellular biochemistry in terms of molecular networks. Annu. Rev. Biochem (2004) 73:1051–1087.[CrossRef][Web of Science][Medline]
Zanzoni A, et al. MINT: a Molecular INTeraction database. FEBS Lett (2002) 513:135–140.[CrossRef][Web of Science][Medline]
Zhang W, et al. Novel approaches to treatment of advanced colorectal cancer with anti-EGFR monoclonal antibodies. Ann. Med (2006) 38:545–551.[CrossRef][Web of Science][Medline]
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  G. Bindea, B. Mlecnik, H. Hackl, P. Charoentong, M. Tosolini, A. Kirilovsky, W.-H. Fridman, F. Pages, Z. Trajanoski, and J. Galon ClueGO: a Cytoscape plug-in to decipher functionally grouped gene ontology and pathway annotation networks Bioinformatics, April 15, 2009; 25(8): 1091 - 1093. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||