Home
This Title All WIREs
WIREs RSS Feed
How to cite this WIREs title:
WIREs Syst Biol Med
Impact Factor: 3.542

Protein–protein interaction networks and subnetworks in the biology of disease

Full article on Wiley Online Library:   HTML PDF

Can't access this content? Tell your librarian.

Abstract The main goal of systems medicine is to provide predictive models of the patho‐physiology of complex diseases as well as define healthy states. The reason is clear—we hope accurate models will ultimately lead to more specific and sensitive markers of disease that will help clinicians better stratify their patient populations and optimize treatment plans. In addition, we expect that these models will define novel targets for combating disease. However, for many complex diseases, particularly at the clinical level, it is becoming increasingly clear that one or a few genomic variations alone (e.g., simple models) cannot adequately explain the multiple phenotypes related to disease states, or the variable risks that attend disease progression. We suggest that models that account for the activities of many interacting proteins will explain a wider range of variability inherent in these phenotypes. These models, which encompass protein interaction networks dysregulated for specific diseases and specific patient sub‐populations, will be constructed by integrating protein interaction data with multiple types of other relevant cellular information. Protein interaction databases are thus playing an increasingly important role in systems biology approaches to the study of disease. They present us with a static, but highly functional view of the cellular state, and thus give us a better understanding of not only the normal phenotype, but also the overall disease phenotype at the level of the whole organism when certain interactions become dysregulated. WIREs Syst Biol Med 2011 3 357–367 DOI: 10.1002/wsbm.121 This article is categorized under: Biological Mechanisms > Regulatory Biology

This WIREs title offers downloadable PowerPoint presentations of figures for non-profit, educational use, provided the content is not modified and full credit is given to the author and publication.

Download a PowerPoint presentation of all images


Network‐based disease modeling approach. Beginning with a model of disease, e.g., human tissues, cell culture, or relevant animal models, one may assay for significant changes between disease and control (e.g., mutations or differentially expressed genes or proteins or SNPs, etc.). The result of any one of these assays (or, theoretically, more than one) is used to ‘seed’ a computational search of the PPI for candidate subnetworks discriminative of the disease (see Box 1 for an example). The approach is motivated by the hypothesis that gene products with a role in disease tend to cluster in the interactome. Further computational modeling can be employed to assess the classification power of the candidate subnetworks to discriminate control from disease, and more importantly provide a basis for validation by perturbation analysis (e.g., siRNA screening) to drive validation of disease biomarkers for clinical utility.

[ Normal View | Magnified View ]

A sample subnetwork of the human PPI network with a state function indicative of liver metastasis in human colorectal cancer. This subnetwork was identified by the CRANE algorithm on the GSE6988 dataset obtained from Gene Expression Omnibus (GEO). The topology of the network that connects the proteins in this subnetwork is shown on the left panel. The mRNA expression profiles of the subnetwork proteins in metastatic and non‐metastatic samples are shown on the right panel. For this subnetwork, the state function LLLLLH (in the order of rows of the gene expression matrix, where L and H, respectively, indicate low and high expression) indicates metastasis, i.e., a sample is likely to be metastatic if the first five genes exhibit low expression, but Osteopontin shows high expression. The overall combinatorial coordinate dysregulation of this subnetwork is 0.72. (Reprinted with permission from Ref 69. Copyright 2010 Springer).

[ Normal View | Magnified View ]

Additive versus combinatorial coordinate dysregulation. A hypothetical example illustrating the difference between additive and combinatorial formulations for subnetwork dysregulation. A hypothetical subnetwork S of the human PPI network is shown in (a). Genes are shown as nodes, interactions between their products are shown as edges. In (b) and (c), each row shows a coding gene's expression level in control and phenotype samples. (b) and (c) each display different hypothetical scenarios for gene expression data. The last row shows subnetwork activity, which is computed as the average of the expression of these four genes in each sample.68 Please refer to Figure 2 for description of subnetwork activity and additive coordinate dysregulation.68 In contrast to additive coordinate dysregulation, combinatorial coordinate dysregulation is defined in terms of how much the expression state of a subnetwork can discriminate control and phenotype samples.69 Here, the state of a subnetwork refers to the combination of expression levels of all genes in the subnetwork. In (b) and (c), the state of the subnetwork in each sample is given by the first four entries of the column corresponding to that sample (e.g., in (b), the subnetwork has expression state MHML in sample N2, whereas it has state HHMM in sample P2, where H, M, and L, respectively, denote high expression, moderate expression, and low expression). In the case shown in (b), subnetwork activity perfectly discriminates control and phenotype, so the subnetwork is considered dysregulated according to the additive formulation of subnetwork dysregulation. On the other hand, in the case shown in (c), neither the expression of individual genes in S, nor the subnetwork activity of S can discriminate control and phenotype. However, combination of the expression states of the genes in S can perfectly discriminate between control and phenotype (either G1 and G3 or G2 and G4 are expressed in normal samples, whereas in phenotype samples, either G1 and G2 or G3 and G4 are expressed). This example demontsrates the potential power of combinatorial approaches in discovering dysregulated subnetworks, beyond what can be discovered by additive approaches.

[ Normal View | Magnified View ]

Univariate versus multivariate assessment of subnetwork dysregulation. A hypothetical example illustrating the difference between univariate and multivariate approaches to identifying dysregulated subnetworks. A hypothetical subnetwork S of the human PPI network is shown in (a). Genes are shown as nodes; interactions between their products are shown as edges. In (b) and (c), each row shows a coding gene's expression level in control and phenotype samples. (b) and (c) each display different hypothetical scenarios for gene expression data. The last row shows subnetwork activity, which is computed as the average of the expression of these five genes in each sample.68 Dark red shows high expression, light gray shows moderate expression, and light green shows low expression. The dysregulation (differential expression) of a gene (or a subnetwork) is measured in terms of how much its expression profile (or activity) can discriminate phenotype and control. Ideker et al.60 define subnetwork dysregulation as the aggregate significance of the dysregulation of each gene, normalized by the number of genes in the subnetwork. We consider this a univariate approach as the dysregulation of each gene is assessed separately and then the results are combined to assess the dysregulation of the subnetwork. On the contrary, Chuang et al.68 define the dysregulation of the subnetwork as the mutual information between phenotype and subnetwork activity, i.e., how much the average expression of the genes in the subnetwork can discriminate phenotype and control. We consider this a multivariate approach as the dysregulation of all genes in the subnetwork are assessed together to compute the dysregulation of the subnetwork. In (b), all genes exhibit maximum dysregulation, as each of them can perfectly discriminate phenotype and control. Consequently, the univariate approach can correctly identify this subnetwork as a dysregulated subnetwork, as all genes in the subnetwork are dysregulated. On the other hand, in (c), each gene exhibits moderate individual dysregulation, so the subnetwork would not be considered a dysregulated subnetwork by the univariate approach. However, in this case, the subnetwork activity can perfectly discriminate phenotype and control, thereby capturing the coordinate dysregulation of the genes in this subnetwork. This example demonstrates the potential of multivariate approaches in discovering dysregulated subnetworks, beyond what can be discovered by a univariate approach.

[ Normal View | Magnified View ]

Browse by Topic

Biological Mechanisms > Regulatory Biology

Access to this WIREs title is by subscription only.

Recommend to Your
Librarian Now!

The latest WIREs articles in your inbox

Sign Up for Article Alerts