Bayesian Networks in the Absence of Temporal Information - Bayesian Networks in R: With Applications in Systems Biology - page 54

stimulus meant to activate the desired paths, nine other data sets subject to different
stimulatory cues and inhibitory interventions are used to elucidate the direction of
the causal relationships in the network. Such data are often called interventional ,be-
cause the values of specific variables in the model are set by an external intervention
of the investigator.
Overall, the ten data sets contain 5,400 observations; in addition to the 11
signaling levels analyzed above, the protein which is activated or inhibited ( INT )
is recorded for each sample.
> isachs = read.table("sachs.interventional.txt",
+ header = TRUE, colClasses = "factor")
One intuitive way to model these data sets with a single, encompassing model is to
include the intervention INT in the network and to make all variables depend on it.
This can be achieved with a whitelist containing all possible arcs from INT to
the other nodes, thus forcing such arcs to be present in the learned network structure.
> wh = matrix(c(rep("INT", 11), names(isachs)[1:11]),
+ ncol = 2)
> bn.wh = tabu(isachs, whitelist = wh, score = "bde",
+ iss = 10, tabu = 50)
Using tabu search instead of hill-climbing improves the stability of the score-based
search; once a locally optimum network is found, tabu search performs an addi-
tional 50 iterations (as specified by the tabu argument) to ensure that no other (and
potentially better) local optimum is found.
We can also let the structure learning algorithm decide which arcs connecting
INT to the other nodes should be included in the network. To this end, we can
use the tiers2blacklist function to blacklist all arcs toward INT , thus en-
suring that only outgoing arcs will be included in the network. In the general case,
tiers2blacklist builds a blacklist such that all arcs going from a node in a
particular element of the nodes argument to a node in one of the previous ele-
ments are blacklisted.
> tiers = list("INT", names(isachs)[1:11])
> bl = tiers2blacklist(nodes = tiers)
> bn.tiers = tabu(isachs, blacklist = bl,
+ score = "bde", iss = 10, tabu = 50)
The networks learned with these two approaches are shown in Fig. 2.10 .Some
of the structural features detected in Sachs et al. ( 2005 ) are present in both bn.wh
and bn.tiers . For example, the interplay between plcg , PIP2 ,and PIP3 and
between PKC , P38 ,and pjnk are both correctly modeled. The lack of any direct
intervention on PIP2 is also correctly modeled in bn.tiers . The most notice-
able feature missing from both networks is the pathway linking praf to pakt473
through pmek and p44.42 .