Identifying differentially regulated subnetworks from phosphoproteomic data

Abstract

Background: Various high throughput methods are available for detecting regulations at the level of transcription, translation or posttranslation (e.g. phosphorylation). Integrating these data with protein networks should make it possible to identify subnetworks that are significantly regulated. Furthermore, such integration can support identification of regulated entities from often noisy high throughput data. In particular, processing mass spectrometry-based phosphoproteomic data in this manner may expose signal transduction pathways and, in the case of experiments with drug-treated cells, reveal the drug's mode of action.

Results: Here, we introduce SubExtractor, an algorithm that combines phosphoproteomic data with protein network information from STRING to identify differentially regulated subnetworks and individual proteins. The method is based on a Bayesian probabilistic model combined with a genetic algorithm and rigorous significance testing. The Bayesian model accounts for information about both differential regulation and network topology. The method was tested with artificial data and subsequently applied to a comprehensive phosphoproteomics study investigating the mode of action of sorafenib, a small molecule kinase inhibitor.

Conclusions: SubExtractor reliably identifies differentially regulated subnetworks from phosphoproteomic data by integrating protein networks. The method can also be applied to gene or protein expression data.

Figure 1

Workflow of the subnetwork extraction. First, single and combined z-scores are calculated from the phosphoproteomics data set and subsequently mapped on an interaction network (orange nodes). Proteins that do not occur in the interaction network are stored in a separate list (violet node). For the genetic algorithm (GA) procedure the network is encoded into a binary vector, where 1 codes for the associated node being active (i.e. part of a regulated subnetwork) and 0 inactive. The GA runs for a defined number of generations (exemplarily, the two-point crossover step in combination with a single-point mutation is depicted), and the strongest individual of the final generation encodes for the globally best achievable solution (here, this would be a subnetwork containing six nodes and a single-node network). Finally, the global rank (GR) significance test is performed on both extracted subnetworks and single nodes (or-more generally-single-node subnetworks) resulting in a set of significantly regulated subnetworks (only one in the depicted example).

Figure 2

SubExtracor's performance on artificial data. Ten artificial data sets were generated to assess the prediction quality of SubExtractor. The top figures (2a and 2b) show the performance for varying σ_z values and a fixed α of 1.0. The figures at the bottom (2c and 2d) depict the mean accuracy for varying α values ranging from 0.01 to 10 and a fixed σ_z of 5.0. Nodes sampled with the background distribution (σ = 1) are the negatives, those coming from the distribution with σ = 5 are the positives. The FN rate is defined as , the FP rate as The overall prediction accuracy is . Error bars display the standard error of the mean over the ten generated data sets.

Figure 3

Example of subnetwork extraction for one artificial data set. The top left area shows the network of 31 nodes that have been sampled from the normal distribution with μ = 0 and σ = 5, thus being the regulated ones in the artificial data set containing 1000 nodes in total. The remaining three areas show networks reconstructed by the proposed algorithm using different values of the parameter α. The colouring represents the level of regulation, where down-regulated nodes are coloured blue, up-regulated ones red and non-regulated nodes white (the darker the colour the stronger the regulation). The differences between the original and the reconstructed subnetworks are highlighted by green ellipses.

Figure 4

Subnetwork extraction for sorafenib mode of action study. The largest two resulting subnetworks are shown (blue nodes are down-regulated, red ones up-regulated) Proteins in the orange circles belong to the MAPK pathway, which is known to be affected by sorafenib. The green rectangle depicts the part of the largest subnetwork that belongs to the mTOR pathway, has not previously been reported to be affected by sorafenib. The network on the right hand side shows important strength of the algorithm, i.e. that subnetworks are also reconstructed if the centre node (i.e. the hub) is not detected to be regulated.