Biomedical Engineering Reference
In-Depth Information
Graphical Gaussian models use the assumption that the data follows a
multivariate Gaussian distribution to identify conditional independence
between nodes. Though they are able to separate direct from indirect inter-
actions, the resulting graphs are undirected and therefore not predictive
(Shafer and Strimmer, 2005a, 2005b).
Bayesian networks are graphical models for probabilistic relationships
among a set of variables. Due to their statistical nature, Bayesian models
readily handle missing data, and can be used to learn causal relationships
and predict the consequences of interventions (Heckerman, 1996). Their
key advantages are their relative robustness to unobserved data, as they can
represent complex non-linear stochastic relationships, and their probabilis-
tic nature which can accommodate noise that is inherent to biological data.
With Bayesian models, one can use already known facts as priors and learn
the rest as they compute the posterior probability of an event occurring.
They have been used to learn and infer the prior probability of a patient
having one or many diseases given his present, absent or unknown symp-
toms (Shwe and Cooper, 1991). They have also been used to infer genetic
sub-networks in
Saccharomyces cerevisiae
from genome-wide expression
profi les of the yeast cells subjected to various treatments and perturbations
(Pe'er et al., 2001), regulatory pathways in
E. coli
(Ong
et al.,
2002) and gene
interactions in leukemia (Djebbari and Quackenbush, 2008). Wang
et al.
(2006) created a gene network reconstitution tool to combine microarray
datasets from different sources. This method theoretically ensures the deri-
vation of the best model with regard to all the datasets presented.
8.4.5 Meta-analysis
Once a list of signifi cant genes is obtained, it is advisable to assess your
data to what is known from the literature. A good avenue to pursue is an
enrichment analysis as described by Subramanian
et al.
(2005). If you have
access to ontology data for your genes (such as cellular function, location,
pathways implicated within, etc.), you can:
use a Fisher's exact test (Hosack
et al.,
2003) to examine in your list of
signifi cant genes, whether you have more than what is expected at ran-
dom of one of those categories;
assess if genes in a particular network are regulated in a signifi cant
manner;
detect signifi cant associations with molecular annotation (e.g., enrich-
ment analysis of gene ontology categories or sequences in promoter
region).
