Automated modelling of signal transduction networks

Abstract

Background: Intracellular signal transduction is achieved by networks of proteins and small molecules that transmit information from the cell surface to the nucleus, where they ultimately effect transcriptional changes. Understanding the mechanisms cells use to accomplish this important process requires a detailed molecular description of the networks involved.

Results: We have developed a computational approach for generating static models of signal transduction networks which utilizes protein-interaction maps generated from large-scale two-hybrid screens and expression profiles from DNA microarrays. Networks are determined entirely by integrating protein-protein interaction data with microarray expression data, without prior knowledge of any pathway intermediates. In effect, this is equivalent to extracting subnetworks of the protein interaction dataset whose members have the most correlated expression profiles.

Conclusion: We show that our technique accurately reconstructs MAP Kinase signaling networks in Saccharomyces cerevisiae. This approach should enhance our ability to model signaling networks and to discover new components of known networks. More generally, it provides a method for synthesizing molecular data, either individual transcript abundance measurements or pairwise protein interactions, into higher level structures, such as pathways and networks.

Figure 1

MAPK signal transduction pathways in yeast. Membrane proteins are depicted in blue, transcription factors in red, and intermediate proteins in green. Figure adapted from [6].

Figure 2

Histogram of the number of proteins with a given number of protein-protein interactions. Interaction data obtained by high-throughput two-hybrid assays [2-4]. The highly interacting proteins in red were removed from the interaction dataset (see text for details).

Figure 3

Histogram of pathways with a given coclustering score for experimental and randomized protein-protein interactions. Histogram of the number of pathways with a given coclustering score for experimental and randomized protein-protein interactions. Shown here is the tail of the distribution with the highest coclustering scores. The paths were drawn with a depth-first search algorithm [43] from membrane to DNA-binding proteins. It is evident that at high coclustering scores, pathways from the experimentally observed interactions (blue) outnumber those generated from randomized interactions (red – an average of three separate randomizations). The total number of paths for experimental and randomized interaction data (averaged) were within 5% of each other.

Figure 4

Network models produced by NetSearch. Pathways predicted by NetSearch for (A, B) pheromone response, (C) cell wall integrity, and (D) filamentation pathways, with the starting membrane protein for path drawing (blue), intermediate proteins (green) and transcription factor (red). In each case, the fifteen highest ranked paths between common endpoints were combined to form the signaling network. For the cell wall integrity pathway, the sensor proteins that initiate signal transduction Wsc/1/2/3p and Mid2p did not have any productive interactions. For this pathway, we began our searches at Rho1p and searched for a path length of seven. The size of each vertex is proportional to the sum of the scores of the paths in which it was included. Network graphs were produced with PAJEK graph drawing software [44], http://vlado.fmf.uni-lj.si/pub/networks/pajek.