Motivation: In the recent years, technical developments enabled the facilitated measurements of biological high-throughput data. This results in a qualitative and a quantitative improvement of the gene- rated data and offers the potential to understand complex biological systems in more detail. Perturbation experiments, for example using RNA interference, are an easy and fast way to screen genes in a high-content, high-throughput manner and thereby, to elucidate their gene function. The inference of signal transduction networks from this data, however, is a challenging task. One of the problems is the exponentially increasing number of possible network topologies with an increasing number of nodes. Here, we formulate the problem of net- work inference as a linear optimization program which can be solved efficiently even for large-scale problems.
Results: Based on simulated data for networks of different sizes we show that our method outperforms a recently published approach, especially when applied to large-scale problems. Using our approach, we achieve increased sensitivity and specificity values and a signi- ficant reduction in computation time in comparison to the other approach. Furthermore, we show that our method can deal with noisy and missing data and that prior knowledge can be easily integrated and thus, improves results. Finally, we use a real data set studying ErbB signaling to reconstruct the underlying network topology. Based on the gene interactions as given in the STRING database we achieve an accuracy much better than random guessing. We were able to reconstruct several already known interactions, as well as identify potential new ones.
Availability: The R source code of the method can be downloaded from http://tudresden.de/med/lpmodel