A scalable method for discovering significant subnetworks

Study of biological networks is an essential first step to understand the complex functions they govern in different organisms. The topology of interactions that define how biological networks operate is often determined through high-throughput experiments. Noisy nature of high-throughput experiments, however, can result in multiple alternative network topologies that explain this data equally well. One key step to resolve the differences is to identify the subnetworks which appear significantly more frequently in a biological network data set than expected.

We present a method named SiS (Significant Subnetworks) to find subnetworks with the largest probability to appear in a collection of biological networks. We define these subnetworks as the most probable subnetworks. SiS summarizes the interactions in the given collection of networks in a special template network. It uses the template network to guide the search for most probable subnetworks. It computes the lower and upper bound scores on how good the potential solutions are (i.e., the number of input networks that contain the subnetwork). As the search continues, it tightens the bound dynamically and prunes a massive number of unpromising solutions in that process.

Experiments on comprehensive data sets depict that the most probable subnetworks found by SiS in a large collection of networks are also very frequent as well. In metabolic network data set, we found that subnetworks in eukaryote are more conserved than those of prokaryote. SiS also scales well to large data sets and subnetworks and runs orders of magnitude faster than an existing method, MULE. Depending on the size of the subnetwork in the same data set, the running time of SiS ranges from a few seconds to minutes; MULE, on the other hand, runs either for hours or does not even finish in days. In human transcription regulatory network data set, SiS finds a large backbone subnetwork that appears frequently regardless of diverse cell types.