Biological interaction networks are conserved at the module level

Abstract

Background: Orthologous genes are highly conserved between closely related species and biological systems often utilize the same genes across different organisms. However, while sequence similarity often implies functional similarity, interaction data is not well conserved even for proteins with high sequence similarity. Several recent studies comparing high throughput data including expression, protein-protein, protein-DNA, and genetic interactions between close species show conservation at a much lower rate than expected.

Results: In this work we collected comprehensive high-throughput interaction datasets for four model organisms (S. cerevisiae, S. pombe, C. elegans, and D. melanogaster) and carried out systematic analyses in order to explain the apparent lower conservation of interaction data when compared to the conservation of sequence data. We first showed that several previously proposed hypotheses only provide a limited explanation for such lower conservation rates. We combined all interaction evidences into an integrated network for each species and identified functional modules from these integrated networks. We then demonstrate that interactions that are part of functional modules are conserved at much higher rates than previous reports in the literature, while interactions that connect between distinct functional modules are conserved at lower rates.

Conclusions: We show that conservation is maintained between species, but mainly at the module level. Our results indicate that interactions within modules are much more likely to be conserved than interactions between proteins in different modules. This provides a network based explanation to the observed conservation rates that can also help explain why so many biological processes are well conserved despite the lower levels of conservation for the interactions of proteins participating in these processes.Accompanying website: http://www.sb.cs.cmu.edu/CrossSP.

Figure 1

Overview of the modules identification procedure. For each species, available co-expression, PPI, GI, and sequence data were extracted and converted into networks. For PPI and GI the networks representation is straightforward. For co-expression, sequence, and GO we computed a similarity score between genes and used a cutoff to construct a network. Expression, PPI, positive GI, and sequence were combined to create a joint weighted network where the weight is a function of the number of edges connecting two genes. Next, the MCL algorithm was applied on the combined network to identify modules for each species separately. See Methods and Supplementary Methods for details.

Figure 2

Edge conservation across species. (a) Types of conservation. We denote one species as the query species (species A, left) and the other as the reference species (B, right). Shaded groups of nodes represent modules. Nodes connected by a grey line between the species represent orthologous genes. The bold black edge in the upper module of both species is a within-module conservation edge. The purple edge connecting the two modules of species A is a between-modules conserved edge. The blue edge (upper module of species A) is an extended-module conserved edge as both proteins connected by this edge are in the same module in species B. (b) Conservation of the integrated network across all pairwise comparisons. Orange bars and blue bars represent within and between conservation rates respectively. Gray bars represent conservation statistics for random modules with error bars showing the standard deviation for 1000 random runs.

Figure 3

Differences between WMI and BMI conservation rates across all pairwise comparison. Green bars and red bars represent conservation statistic for real and random modules respectively. The bars represent the difference between WMI and BMI conservation rates (darker green and red) and the difference between extended WMI and extended BMI conservation rates (brighter green and red). The species are indicated on the vertical axis as follows (c-S.cerevisiae, p-S.pombe, e-C.elegans, f-D.melanogaster). For most data types the improvement for the real networks is very large. In contrast, for random networks the within module edges are usually less conserved when compared to the overall conservation indicating that the within module conservation bias is even stronger.

Figure 4

(a) Module matching. Green, yellow, blue, and grey nodes correspond to modules in S. cerevisiae, S. pombe, C. elegans, and D. melanogaster respectively. The size of a node corresponds to the number of genes in the module. The width of an edge connecting two nodes reflects the p-value of the reciprocal match between two modules, when more significant matches correspond to wider edges. (see Additional File 13 for complete listing). (b-d) Examples for matched modules across S. cerevisiae, S. pombe, C. elegans, and D. melanogaster. Each row contains modules that significantly overlap based on orthology for all pairwise comparison. The examples are marked in a red circle in Figure 4a. The nodes are colored with the same color scheme of 4a. The edges are colored based on the interaction type (see legend - note that GI edges refer to both positive GI and negative GI edges), and multiple edges between two nodes are allowed. For clarity, only genes that have orthologs in at least one of the other modules are shown. See text for details on the matched modules.

Figure 5

Module conservation is analogous to sequence conservation. For sequences (left) coding regions are usually much more conserved than the genome as a whole. Similarly, in the network setting, modules are more conserved than the entire network. In addition, coding regions can often tolerate synonymous mutations that change the DNA sequence itself but do not alter the protein product. Similarly, modules may be able to tolerate loss of specific interactions as long as the two interacting orthologs remain in the same module (often through redundant interactions or interactions with other module members).