Biological context networks: a mosaic view of the interactome

Abstract

Network models are a fundamental tool for the visualization and analysis of molecular interactions occurring in biological systems. While broadly illuminating the molecular machinery of the cell, graphical representations of protein interaction networks mask complex patterns of interaction that depend on temporal, spatial, or condition-specific contexts. In this paper, we introduce a novel graph construct called a biological context network that explicitly captures these changing patterns of interaction from one biological context to another. We consider known gene ontology biological process and cellular component annotations as a proxy for context, and show that aggregating small process-specific protein interaction sub-networks leads to the emergence of observed scale-free properties. The biological context model also provides the basis for characterizing proteins in terms of several context-specific measures, including 'interactive promiscuity,' which identifies proteins whose interacting partners vary from one context to another. We show that such context-sensitive measures are significantly better predictors of knockout lethality than node degree, reaching better than 70% accuracy among the top scoring proteins.

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).

Figure 2

A simulation of the aggregation of random (Erdös-Rényi) graphs showing rapid reconstitution of a scale-free degree distribution. The simulation involved a set of random labels (contexts), L, distributed across N=1000 nodes, subject to the additional condition that the number of labels per node follows a power-law distribution. Displayed results are the average of 100 trials, showing degree distribution after aggregating L=1, 25, 50, 75, and 100 labels. In our simulation, we assume that two nodes sharing a given context have a fixed probability of interaction (P=0.05), thus any context-specific sub-network is a random (Erdös-Rényi) graph.

Figure 3

(A) Location of GO biological process leaf-term sub-networks in EDGE-NODE space, showing the variability in the size of the resulting projections. (B) The resulting sub-networks reveal a broad range of irregular distributions. Singleton nodes are excluded for clarity. Node color coding is by degree: 1–4 neighbors (blue), 5–9 neighbors (green), 10–14 neighbors (yellow), 15+ neighbors (red).

Figure 4

Neighborhood-annotation matrices for three interactively promiscuous examples: (A) Cdc6, (B) Spt5, and (C) Exo84. Column headers include all neighbors having at least one GO biological process annotation. Rows correspond to particular GO annotations associated with these neighboring proteins. Red boxes indicate that the neighbor protein has the annotation explicitly or a more specific annotation in the GO biological process ontology.

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.

Figure 6

(A) Percent of proteins (N=4741) that are essential (knockout lethal) for the top-ranked proteins (1–20%), according to various measures including degree, context degree, context mutual information, and interactive promiscuity. The highest ranked proteins (top 1%) using context-verified degree contain the highest proportion of lethal nodes, but this measure is surpassed by the mutual-information-based measure when we include the top 2%. All three measures outperform degree as a predictor of lethality, although, as we encompass larger numbers of proteins, the differences are less pronounced. (B) Approximate measure cutoffs for corresponding rank levels.