Biological context networks: a mosaic view of the interactome

John Rachlin^1,2, Dikla Dotan Cohen², Charles Cantor^3,4,5 & Simon Kasif^2,3,6

Article highlights

Synopsis

The machinery of the cell involves a complex network of biochemical interactions. The full complement of known interactions (the so-called interactome) is not a static construct. Rather, specific interactions are activated or deactivated as part of cell-specific processes involving metabolism, cell cycle regulation, signal transduction, and hundreds of other characteristic processes. In this paper, we present a parsimonious network model (a 'biological context network') in which the nodes of the network, representing proteins, are deployed in one of several biological contexts, affecting whether or not connecting edges are active. The specification or the biological program provided by the model only articulates which biological contexts are associated with each protein and whether two proteins, each active in a particular context, interact. We study two types of context specification, GO biological processes or GO protein localization, but the formalism naturally extends to modeling contexts in the form of environmental stimuli, pathological conditions, tissues, or even organisms. The model elucidates interactions that are either conserved or variable from one context to another.

Graphs and their variants are the foundation for modeling complex biological systems. Graph topology reveals the basic properties of connectivity, robustness, modularity, hierarchical structure, and other properties, enabling identification of protein complexes or functional modules (Segal et al, 2003; Spirin and Mirny, 2003), and serves to aid whole-genome functional annotation efforts (Marcotte et al, 1999; Letovsky and Kasif, 2003). Biological networks are also of commercial interest as an aid to drug target discovery (Gardner et al, 2003) or for predicting toxic side effects, and are at the heart of pharmaceutical initiatives focused on integrating and mining pathway data sets (Hood et al, 2004).

Protein interaction networks are often obtained by high-throughput detection assays (Rual et al, 2005) or inferred from literature surveys (Mishra et al, 2006). As a result, they represent a high-level integrated summary of a large number of interactions inferred from many biological contexts. However, representing the interactome as a static biological network is akin to a long-exposure photograph that can mask context-specific patterns of activation across multiple processes, cellular locations, and time. Conclusions drawn from the full network's topology may be compromised by these inherent limitations. A central goal of systems biology research is to elucidate the underlying patterns of interaction in an effort to obtain more realistic and predictive models of the cell (Ideker et al, 2001). This has prompted the development of a broad range of graphical representations coupled with mathematical equations intended to model cellular dynamics. By contrast, protein–protein interaction networks are typically represented as a standard undirected graph where vertices correspond to individual proteins and edges connect pairs of interacting proteins. Biological context networks provide an intermediate-level formalism, in which we label proteins with contextual information about the protein and activate protein interactions as specified by the succinct biological program associated with the network. In its simplest form, the program activates an edge whenever two interacting proteins are in a shared contextual state, and otherwise assumes that the interaction has been inactivated. The biological context network model enables one to view the interactome as a mosaic of overlapping sub-networks each associated with specific contexts or conditions and to further characterize changes in topology from one context to another. For example, in Figure 1, we show the context-specific sub-networks in the local neighborhood of the protein Sec13, highlighting its association with both the nuclear pore complex and the endoplasmic reticulum (Enninga et al, 2003).

Figure 1

The local context networks for Sec13 with respect to two of its current GO biological process annotations—GO:0006888, nuclear pore organization and biogenesis, and GO:0006999, ER to Golgi transport—highlighting Sec13's association with both the nuclear pore complex and the ER. Sec13 is an example of a protein whose interacting partners vary from one process context to another. We characterize such proteins as 'interactively promiscuous.' The shuttling of Sec13 between the nucleus and the cytoplasm is believed to play a cross-functional regulatory role (Enninga et al, 2003).

Full figure and legend (260K)Figures & Tables index

It has been widely observed that a broad range of social, technological, and biological networks are scale-free, characterized by a power-law degree distribution where a few 'hub' proteins have many interacting partners, whereas most proteins have very few (Barabasi and Oltvai, 2004). Furthermore, high-degree 'hubs' in protein–protein interaction networks are more likely to be essential for the viability of the organism. In this paper, we provide some evidence that a power-law distribution, while clearly evident in the aggregate experimental protein–protein interaction data, is plausibly an artifact of the aggregation of interactions across multiple process-specific contexts. This observation suggests that paths connecting disparate protein pairs may be substantially impacted by intervening contextual differences. We show, for example, that aggregating approximately 100 small (<14 edges) GO biological process 'leaf-term' sub-networks (i.e., sub-networks corresponding to the most highly specific GO terms) is sufficient to reconstitute a scale-free network (R²=0.88).

Context-specific sub-networks derived using the biological context network model provide the basis for characterizing proteins with respect to several context-specific measures. These include the following

• Context degree: The degree of a node (number of interacting partners), considering only those partners that share at least one context (annotation).
• Context mutual information: Measures the degree to which the annotations of neighboring proteins are correlated.
• Interactive promiscuity: Measures the variability of annotations (contexts) among a protein's interacting partners in an effort to identify those proteins likely to play a cross-contextual 'linking' role.

The context mutual information and interactive promiscuity measures are based on patterns of annotation among a protein's interacting partners, as shown, for example, in Figure 5.

Figure 5

(A) DIP yeast sub-network of all 991 essential proteins. (B) Essential proteins having context degree >1. Node coloring is according to the degree of the protein in the full DIP network: 1–4 neighbors (blue), 5–9 neighbors (green), 10–14 neighbors (yellow), 15+ neighbors (red). Many of the essential proteins aggregate into clusters of essential protein complexes that are typically related to cell-cycle regulation and mRNA processing. As a result of the network's improved specificity, context degree is a better predictor for knockout lethality, although applicable only to annotated nodes.

Full figure and legend (179K)Figures & Tables index

Interestingly, we find that the top-ranked proteins with respect to each of these context-specific measures are highly enriched in essential proteins and these measures provide a significantly improved predictor for knockout lethality than the static measure of degree computed from the original 'context-free' network. For example, more than 70% of the top 2% of proteins ranked by either context mutual information or interactive promiscuity are essential, compared to just 53% of the top 2% of proteins ranked by degree.

The biological context network formalism provides insight into the statistical topological characteristics of the network within specific contexts, including hubs, dense-sub-graphs, connectivity, and centrality, but quantified with respect to particular contexts. This formalism also suggests an explanation for the emergence of scale-free properties in protein–protein interaction networks, and offers a measure for the interactive promiscuity of a protein, highlighting those proteins that are either intrinsically multi-functional or are a subunit of a multi-functional complex. Future applications of context-specific functional modules and networks include the modeling of cross-context connectivity, directional and stochastic models of context networks, and the context-specific effects of perturbations on biological function, and enabling improved selection of drug targets by way of more reliable models of toxicity.

MORE ARTICLES LIKE THIS

These links to content published by NPG are automatically generated.