Analysis of Regulatory and Interaction Networks from Clusters of Co-expressed Genes - Clustering Challenges in Biological Network - page 61

relative cost of running live animal experiments. Therefore, an insufficient number
of perturbed states have been obtained for one to infer comprehensive binary or
Boolean networks from the available data. Due to the fact that the construction of
mammalian transcriptional networks is data-constrained as opposed to studies in
yeast, most of the interactions must comprise of predicted links obtained through
various transcription factor binding site prediction algorithms such as CONSITE,
FOOTER, AlignACE, and CORG [55-58]. The problem with utilizing these algo-
rithms for the prediction of transcription factor binding sites lies primarily in the
large number of false positives and false negatives. Work has been previously done
on ways to improve the prediction based on the concept of phylogenetic footprint-
ing in which false positives are eliminated by looking at evolutionarily conserved
segments in the promoter region. However, despite these attempts, the algorithms
are still relatively inaccurate [59]. Due to the relative inaccuracies of the transcrip-
tion factor prediction methods, gene expression data and clustering were used to
identify the subnet that was active under the experimental conditions and to per-
form gross organization of genes into clusters before network construction. What
one hopes to obtain from the construction of the network is the identification of
possible biological systems affected via corticosteroid administration as well as
the identification of possible points of intervention which allow for the control of
specific biological processes independently. Therefore, one may be able to ascer-
tain whether or not it is possible to affect the inflammatory response of the liver
without triggering currently concurrent responses in metabolism or the immune
response.
The transcriptional network was constructed using transcription factor binding
site prediction via CORG. CORG was selected over other tools such as CONSITE
because of a built in facility for extracting promoter regions of a specific gene as
well as automatic phylogenetic footprinting which allows for the analysis of the
network characteristics as the false positive links are pruned. The network gener-
ated via CORG is a standard bi-partite network because it finds the feed-forward
interactions of transcription factors and their respective genes. However, as men-
tioned before, it can be transformed into a standard DAG if the output genes can
be associated with a transcription factor. Given the large number of uncertainties
in the network construction step due to the use of transcription factor prediction,
the analysis will focus upon the global properties of the links rather than the bi-
ological significance of each link. The network is given in Fig. 3.4, and is a
representation of a bi-partite network where the nodes in blue are the associated
transcription factors and the nodes in green represent the final genes. We have
treated the transcription factors as separate from the set of genes, and have con-
ducted the graph analyses based on the outgoing connectivity of the transcription