To Do

Clearly separate the network topology inference from its predictive ability

The analytical pipeline is now:

1. identify a network topology, either using the regrnet or the bayesian network inference approach

2. bring predictive ability to the network model either by fitting local regression models or by estimating CPTs

3. make predictions with the new network model built at step 2.

Create proper R classes for networks built at step 1 and 2.

Have a black and a white lists parameters

Black/white edge lists would allow users to choose to force some edges to be absent/present in the inferred network

Let users skip the bootstrapping step in the ensemble network inference

Always choose the best n solutions and do not compare then statistically by using bootstrap. -> Done

Let users export networks+statistics into cytoscape

See Rcytoscape

Implementation of MCC

Here is an implementation of the MCC extracted from the metaDE package:

cal.MCC<-function(dt1,dt2,l1,l2) {
l1<-unclass(factor(l1))
l2<-unclass(factor(l2))
K<-nlevels(l1)
n1<-table(factor(l1,levels=unique(l1)))
n2<-table(factor(l2,levels=unique(l2)))
# get means for each group each gene
ind1<-diag(rep(1,length(n1)))[rep(1:nlevels(l1),n1),] 
ind2<-diag(rep(1,length(n2)))[rep(1:nlevels(l2),n2),]
xk.<-dt1%*%ind1%*%diag(1/n1)
yk.<-dt2%*%ind2%*%diag(1/n2)
x..<-rowMeans(xk.)
y..<-rowMeans(yk.)
sxk.yk.<-rowSums(xk.*yk.)
num<-1/K*sxk.yk.-x..*y..

sumx2<-dt1^2%*%ind1
sumy2<-dt2^2%*%ind2
vx<-1/K*rowSums((sumx2-xk.^2)%*%diag(1/(n1-1)))-x..^2
vy<-1/K*rowSums((sumy2-yk.^2)%*%diag(1/(n2-1)))-y..^2
den<-sqrt(vx*vy)
r<-num/den
return(r)
}

Include a class to predict

Discussing with Amira, we concluded that it would be of interest to include a variable class taht could represent the survival of a patient or any other variable of interest, so the network could be used to predict it (Markov blanket, multivariate models with the causal gene -- parents -- and a combination of univariate classifiers -- effects or children).