Biomedical Engineering Reference
In-Depth Information
presuppose that concentration varies continuously and deterministically.
However, these assumptions do not hold when we try to model systems that
involve just a few molecules in the cell or when the timing of cellular events
fl uctuates.
Another way to model is to use stochastic models of gene regulation
(Arkin
et al.,
1998; Gillespie, 1977; McAdams and Arkin, 1997). This type
of modeling results in a closer approximation of describing a gene regula-
tory network but requires a detailed knowledge of the reaction mechanism
which is often unavailable. Moreover, stochastic models are costly due to
the number of simulations required. On a larger time- or population-scale,
stochastic effects may level out and deterministic models can form a good
approximation of the system.
One of the major issues in all modeling techniques for genomic and pro-
teomic data is that the number of nodes (genes, proteins) usually exceeds
the number of experiments, leaving the model poorly determined. The mod-
els require ample time and effort to build and often end up having narrow
applicability because they only look at a portion of the system.
Another modeling paradigm is to use literature data to infer networks
(using Pathway Studio, Ariadne Genomics Inc., and PathwayArchitect
TM
,
Stratagene Inc.). The problem with using just literature and text mining
tools is that interactions discovered can be specifi c to a cell type and not
take place in other cells; also, interactions that may be indirect can be
highlighted. This occurs because the context of the whole cellular net-
work was not taken into account during the experiment. In addition, we
sometimes encounter confl icting results in the literature about proteins
or genes that up regulate another target in one publication but opposite
results in another publication. Thus, literature data cannot be taken liter-
ally, even with excellent text mining software. These models are useful for
putting the results of an experiment in perspective to what is known, but
are not able to discover unknown functions and relationships between
genes.
There is a substantial benefi t from using microarray data to reverse engi-
neer networks and pathways from the data themselves. The three most
common methods are graphical Gaussian models, relevance networks and
Bayesian networks.
The relevance network (Butte and Kohane, 2000, 2003; Pathway Architect,
Stratagene) is based on pair-wise association scores (Pearson correlation).
This method is very appealing due to its low computational cost but under-
performs compared to the other two methods (Werhli
et al.
, 2006). Since
the inference of interaction between nodes is done pair-wise, as opposed to
the context of the whole system, this model is not particularly powerful in
distinguishing between direct and indirect interactions between the nodes.
