Integrating Various Data Sources for Improved Quality in Reverse Engineering of Gene Regulatory Networks

Introduction

Biological systems are intrinsically complex, still robust and at the same time able to quickly adapt to new situations. To understand, describe and model a wide range of biological systems −involving genes, proteins, metabolites and ecological food webs− networks have served as the unifying language (Barabasi et al. 2004). This description has often revealed a complex network topology. In the case of Gene Regulatory Networks (GRNs), some features are the existence of key genes regulating multiple processes (“hubs”), feed-back motifs and modularity enhancing the system robustness (Milo et al. 2002; Barabasi et al. 2004). Furthermore, the dynamical systems seem to be tuned to enable a stable system by keeping hubs repressed, but still flexible by utilizing, e.g., incoherent feed-back loops (Gustafsson et al. In press b, Ma’ayan et al. 2008). In addition to the architectural complications, we know that gene regulation is a non-linear process including combinatorial control, saturation and stochasticity. These pieces give raise to an extremely challenging modelling problem, which becomes even more complicated by the size of the genome.

Further, the experimental advancements in the last decades have resulted in a vast amount of large-scale data sets available through public databases. To infer a large-scale GRN it is of uttermost importance to take as much as possible of these data into account. Particularly informative for understanding genome-wide gene regulation is the interaction map between Transcription Factors (TFs) and their DNA binding regions. This information may give direct structural properties of the regulatory possibilities, e.g., the presence of a binding element upstream of gene of A for a TF which gene B codes for induces an enhanced possibility for regulation of gene A by gene B.

Other types of structural information may come from sequence based predictions, e.g., prediction of putative regulations from the TF binding sites (TFBS) and from common biological knowledge. The latter can be incorporated in a variety of ways, which may come from annotation knowledge or more “unclean” knowledge as text-mining. Annotation knowledge may be the collection of detailed knowledge from previous experiments, while text-mining may be a possibility to include the plethora of published biological papers in databases. On a more detailed causal level there is also a large number of time-series expression data sets for mRNA levels (see, e.g., Omnibus at Entrez (PubMed 2007) for collections at a unified format). However, although all these experiments are present on a large-scale, they are all typically several orders of magnitudes smaller than the number of presumptive regulators. Hence, all data at hand should be taken in consideration to overcome the indefiniteness of the reverse engineering problem. The greatest challenge in GRN inference to tackle is that the number of genes vastly exceeds the number of experiments, making it a tough statistical question. We should therefore strive to avoid introducing more entities in the model. Consequently, we project gene regulation onto the space of genes only, despite the fact that gene regulation is carried out from the interactions of mRNA molecules, proteins and metabolites (Brazhnik et al. 2002; Ptashne et al. 2002). Indeed, the obtained GRN is then an effective network of gene-to-gene interactions, where these interactions cannot be interpreted as biochemical reactions.