generateNetwork: Random networks and data sampling

Description

1. Random network generation; 2. sampling of data from a given network topology

Usage

sampleRndNetwork(Sgenes, scaleFree=TRUE, gamma=2.5, maxOutDegree=length(Sgenes), maxInDegree=length(Sgenes), trans.close=TRUE, DAG=FALSE)
sampleData(Phi, m, prob=NULL, uninformative=0, type="binary", replicates=4, typeI.err=0.05, typeII.err=0.2, alpha=sample(seq(0.1,0.9,by=0.1),ncol(Phi),replace=TRUE), beta=sample(5:50,ncol(Phi),replace=TRUE), lambda=matrix(sample(seq(0.01,0.49,by=0.01),ncol(Phi)*2,replace=TRUE),ncol=2), meansH1=rep(0.5, ncol(Phi)), meansH0=rep(-0.5, ncol(Phi)), sdsH1=sample(seq(0.1,1,by=0.1),ncol(Phi),replace=TRUE), sdsH0=sample(seq(0.1,1,by=0.1),ncol(Phi),replace=TRUE))

Arguments

Sgenes

character vector of S-genes

scaleFree

should the network topology be scale free?

gamma

for scale free networks: out-degrees of nodes are sampled from $\frac{1}{Z} * (0:maxOutDegree)^{-\gamma}$, where Z is a normalization factor

maxOutDegree

maximal out-degree of nodes

maxInDegree

maximal in-degree of nodes prior to transitive closure

trans.close

Should the transitive closure of the graph be returned? Default: TRUE

DAG

Should only DAGs be sampled? Default: FALSE

Phi

adjacency matrix

number of E-genes to sample

prob

probability for each S-gene to get an E-gene attached

uninformative

additional number of uninformative E-genes, i.e. E-genes carrying no information about the nested structure

type

"binary" = binary data; "density" = log 'p-value' densities sampled from beta-uniform mixture model; "lodds" = log odds sampled from two normal distributions

replicates

number of replicate measurements to simulate for binary data

typeI.err

simulated type I error for binary data

typeII.err

simulated type II error for binary data

alpha

parameter for $Beta(\alpha,1)$ distribution: one parameter per S-gene

beta

parameter for $Beta(1,\beta)$ distribution: one parameter per S-gene

lambda

mixing coefficients for beta-uniform mixture model of the form: $\lambda_1 + \lambda_2*Beta(\alpha,1) + \lambda_3*Beta(1,\beta)$. There is a vector of 3 mixing coefficients per model and one model per S-gene.

meansH1

normal distribution means of log odds ratios under the hypothesis of expecting an effect: one mean per S-gene

meansH0

normal distribution means of log odds ratios under the null hypothesis: one mean per S-gene

sdsH1

normal distribution standard deviations of log odds values under the hypothesis of expecting an effect: one sd per S-gene

sdsH0

normal distribution standard deviations of log odds values under the null hypothesis: one sd per S-gene

Value

For sampleRndNetwork an adjacency matrix, for sampleData a data matrix, for sampleData.BN a data matrix and a linking of effects to signals.

Details

Random networks are generated as follows: For each S-gene $S_{k}$ we randomly choose the number $o$ of outgoing edges between 0 and maxOutDegree. This is either done uniform randomly or, if scale free networks are created, according to a power law distribution specified by gamma. We then select $o$ S-genes having at most maxInDegree ingoing edge and connected $S_{k}$ to them.

The function sampleData samples data from a given network topology as follows: We first attach E-genes to S-genes according to the probabilities prob (default: uniform). We then simulate knock-downs of the individual S-genes. For those E-genes, where no effects are expected, values are sampled from a null distribution, otherwise from an alternative distribution. In the simplest case we only sample binary data, where 1 indicates an effect an 0 no effect. Alternatively, we can sample log "p-value" densities according to a beta-uniform mixture model, where the null distribution is uniform and the alternative a mixture of two beta distributions. A third possibility is to sample log odds ratios, where alternative and null distribution are both normal.

Examples

Run this code

	Phi = sampleRndNetwork(paste("S",1:5,sep=""))
	D = sampleData(Phi, 100, type="density")$D	
  if(require(Rgraphviz)){
	  plot(as(transitive.reduction(Phi),"graphNEL"), main="original graph")
	  x11()
	  plot.nem(nem(D, control=set.default.parameters(unique(colnames(D)), type="CONTmLLBayes")), transitiveReduction=TRUE, SCC=FALSE, main    ="inferred graph")
  }